diff --git a/parse/train/9xC2tWEwBD/9xC2tWEwBD.md b/parse/train/9xC2tWEwBD/9xC2tWEwBD.md new file mode 100644 index 0000000000000000000000000000000000000000..b8d8edfdabc996c15d307befa012b2d1448c96a1 --- /dev/null +++ b/parse/train/9xC2tWEwBD/9xC2tWEwBD.md @@ -0,0 +1,327 @@ +# A PANDA? NO, IT’S A SLOTH: SLOWDOWN ATTACKS ON ADAPTIVE MULTI-EXIT NEURAL NETWORK INFERENCE + +Sanghyun Hong∗, Yigitcan Kaya ˇ ∗, Ionut,-Vlad Modoranu†, Tudor Dumitras, + +University of Maryland, College Park, USA †Alexandru Ioan Cuza University, Ias,i, Romania + +shhong@cs.umd.edu, yigitcan@cs.umd.edu, modoranu.ionut.vlad@hotmail.com, tudor@umd.edu + +# ABSTRACT + +Recent increases in the computational demands of deep neural networks (DNNs), combined with the observation that most input samples require only simple models, have sparked interest in input-adaptive multi-exit architectures, such as MSDNets or Shallow-Deep Networks. These architectures enable faster inferences and could bring DNNs to low-power devices, e.g., in the Internet of Things (IoT). However, it is unknown if the computational savings provided by this approach are robust against adversarial pressure. In particular, an adversary may aim to slowdown adaptive DNNs by increasing their average inference time—a threat analogous to the denial-of-service attacks from the Internet. In this paper, we conduct a systematic evaluation of this threat by experimenting with three generic multi-exit DNNs (based on VGG16, MobileNet, and ResNet56) and a custom multi-exit architecture, on two popular image classification benchmarks (CIFAR-10 and Tiny ImageNet). To this end, we show that adversarial example-crafting techniques can be modified to cause slowdown, and we propose a metric for comparing their impact on different architectures. We show that a slowdown attack reduces the efficacy of multi-exit DNNs by $9 0 { - } 1 0 0 \%$ , and it amplifies the latency by $1 . 5 – 5 \times$ in a typical IoT deployment. We also show that it is possible to craft universal, reusable perturbations and that the attack can be effective in realistic black-box scenarios, where the attacker has limited knowledge about the victim. Finally, we show that adversarial training provides limited protection against slowdowns. These results suggest that further research is needed for defending multi-exit architectures against this emerging threat. Our code is available at https://github.com/sanghyun-hong/deepsloth. + +# 1 INTRODUCTION + +The inference-time computational demands of deep neural networks (DNNs) are increasing, owing to the “going deeper" (Szegedy et al., 2015) strategy for improving accuracy: as a DNN gets deeper, it progressively gains the ability to learn higher-level, complex representations. This strategy has enabled breakthroughs in many tasks, such as image classification (Krizhevsky et al., 2012) or speech recognition (Hinton et al., 2012), at the price of costly inferences. For instance, with $4 \times$ more inference cost, a 56-layer ResNet (He et al., 2016) improved the Top-1 accuracy on ImageNet by $19 \%$ over the 8-layer AlexNet. This trend continued with the 57-layer state-of-the-art EfficientNet (Tan & Le, 2019): it improved the accuracy by $10 \%$ over ResNet, with $9 \times$ costlier inferences. + +The accuracy improvements stem from the fact that the deeper networks fix the mistakes of the shallow ones (Huang et al., 2018). This implies that some samples, which are already correctly classified by shallow networks, do not necessitate the extra complexity. This observation has motivated research on input-adaptive mechanisms, in particular, multi-exit architectures (Teerapittayanon et al., 2016; Huang et al., 2018; Kaya et al., 2019; Hu et al., 2020). Multi-exit architectures save computation by making input-specific decisions about bypassing the remaining layers, once the model becomes confident, and are orthogonal to techniques that achieve savings by permanently modifying the model (Li et al., 2016; Banner et al., 2018; Han et al., 2015; Taylor et al., 2018). Figure 1 illustrates how a multi-exit model (Kaya et al., 2019), based on a standard VGG-16 architecture, correctly classifies a selection of test images from ‘Tiny ImageNet’ before the final layer. We see that more typical samples, which have more supporting examples in the training set, require less depth and, therefore, less computation. + +It is unknown if the computational savings provided by multi-exit architectures are robust against adversarial pressure. Prior research showed that DNNs are vulnerable to a wide range of attacks, which involve imperceptible input perturbations (Szegedy et al., 2014; Goodfellow et al., 2015; Papernot et al., 2016; Hu et al., 2020). Considering that a multi-exit model, on the worst-case input, does not provide any computational savings, we ask: Can the savings from multi-exit models be maliciously negated by input perturbations? As some natural inputs do require the full depth of the model, it may be possible to craft adversarial examples that delay the correct decision; it is unclear, however, how many inputs can be delayed with imperceptible perturbations. Furthermore, it is unknown if universal versions of these adversarial examples exist, if the examples transfer across multi-exit architectures and datasets, or if existing defenses (e.g. adversarial training) are effective against slowdown attacks. + +![](images/746d8efde4dde3f7f81419a6f62cb228c7561c4c925bbdc393202621421b32de.jpg) +Figure 1: Simple to complex inputs. Some Tiny ImageNet images a VGG-16 model can correctly classify if computation stops at the $1 ^ { s t }$ , $5 ^ { t h }$ , and $1 4 ^ { t h }$ layers. + +Threat Model. We consider a new threat against DNNs, analogous to the denial-of-service $( D o S )$ attacks that have been plaguing the Internet for decades. By imperceptibly perturbing the input to trigger this worst-case, the adversary aims to slow down the inferences and increase the cost of using the DNN. This is an important threat for many practical applications, which impose strict limits on the responsiveness and resource usage of DNN models (e.g. in the Internet-of-Things (Taylor et al., 2018)), because the adversary could push the victim outside these limits. For example, against a commercial image classification system, such as Clarifai.com, a slowdown attack might waste valuable computational resources. Against a model partitioning scheme, such as Big-Little (De Coninck et al., 2015), it might introduce network latency by forcing excessive transmissions between local and remote models. A slowdown attack aims to force the victim to do more work than the adversary, e.g. by amplifying the latency needed to process the sample or by crafting reusable perturbations. The adversary may have to achieve this with incomplete information about the multi-exit architecture targeted, the training data used by the victim or the classification task (see discussion in Appendix A). + +Our Contributions. To our best knowledge, we conduct the first study of the robustness of multi-exit architectures against adversarial slowdowns. To this end, we find that examples crafted by prior evasion attacks (Madry et al., 2017; Hu et al., 2020) fail to bypass the victim model’s early exits, and we show that an adversary can adapt such attacks to the goal of model slowdown by modifying its objective function. We call the resulting attack DeepSloth. We also propose an efficacy metric for comparing slowdowns across different multi-exit architectures. We experiment with three generic multi-exit DNNs (based on VGG16, ResNet56 and MobileNet) (Kaya et al., 2019) and a speciallydesigned multi-exit architecture, MSDNets (Huang et al., 2018), on two popular image classification benchmarks (CIFAR-10 and Tiny ImageNet). We find that DeepSloth reduces the efficacy of multiexit DNNs by $9 0 { - } 1 0 0 \%$ , i.e., the perturbations render nearly all early exits ineffective. In a scenario typical for IoT deployments, where the model is partitioned between edge devices and the cloud, our attack amplifies the latency by $1 . 5 – 5 \times$ , negating the benefits of model partitioning. We also show that it is possible to craft a universal DeepSloth perturbation, which can slow down the model on either all or a class of inputs. While more constrained, this attack still reduces the efficacy by $5- 4 5 \%$ . Further, we observe that DeepSloth can be effective in some black-box scenarios, where the attacker has limited knowledge about the victim. Finally, we show that a standard defense against adversarial samples—adversarial training—is inadequate against slowdowns. Our results suggest that further research will be required for protecting multi-exit architectures against this emerging security threat. + +# 2 RELATED WORK + +Adversarial Examples and Defenses. Prior work on adversarial examples has shown that DNNs are vulnerable to test-time input perturbations (Szegedy et al., 2014; Goodfellow et al., 2015; Papernot et al., 2017; Carlini & Wagner, 2017; Madry et al., 2018). An adversary who wants to maximize a model’s error on specific test-time samples can introduce human-imperceptible perturbations to these samples. Moreover, an adversary can also exploit a surrogate model for launching the attack and still hurt an unknown victim (Athalye et al., 2018; Tramèr et al., 2017b; Inkawhich et al., 2019). This transferability leads to adversarial examples in more practical black-box scenarios. Although many defenses (Kurakin et al., 2016; Xu et al., 2017; Song et al., 2018; Liao et al., 2018; Lecuyer et al., 2019) have been proposed against this threat, adversarial training (AT) has become the frontrunner (Madry et al., 2018). In Sec 5, we evaluate the vulnerability of multi-exit DNNs to adversarial slowdowns in white-box and black-box scenarios. In Sec 6, we show that standard AT and its simple adaptation to our perturbations are not sufficient for preventing slowdown attacks. + +Efficient Input-Adaptive Inference. Recent input-adaptive DNN architectures have brought two seemingly distant goals closer: achieving both high predictive quality and computational efficiency. There are two types of input-adaptive DNNs: adaptive neural networks (AdNNs) and multi-exit architectures. During the inference, AdNNs (Wang et al., 2018; Figurnov et al., 2017) dynamically skip a certain part of the model to reduce the number of computations. This mechanism can be used only for ResNet-based architectures as they facilitate skipping within a network. On the other hand, multi-exit architectures (Teerapittayanon et al., 2016; Huang et al., 2018; Kaya et al., 2019) introduce multiple side branches—or early-exits—to a model. During the inference on an input sample, these models can preemptively stop the computation altogether once the stopping criteria are met at one of the branches. Kaya et al. (2019) have also identified that standard, non-adaptive DNNs are susceptible to overthinking, i.e., their inability to stop computation leads to inefficient inferences on many inputs. + +Haque et al. (2020) presented attacks specifically designed for reducing the energy-efficiency of AdNNs by using adversarial input perturbations. However, our work studies a new threat model that an adversary causes slowdowns on multi-exit architectures. By imperceptibly perturbing the inputs, our attacker can (i) introduce network latency to an infrastructure that utilizes multi-exit architectures and (ii) waste the victim’s computational resources. To quantify this vulnerability, we define a new metric to measure the impact of adversarial input perturbation on different multi-exit architectures (Sec 3). In Sec 5, we also study practical attack scenarios and the transferability of adversarial input perturbations crafted by our attacker. Moreover, we discuss the potential defense mechanisms against this vulnerability, by proposing a simple adaptation of adversarial training (Sec 6). To the best of our knowledge, our work is the first systematic study of this new vulnerability. + +Model Partitioning. Model partitioning has been proposed to bring DNNs to resource-constrained devices (De Coninck et al., 2015; Taylor et al., 2018). These schemes split a multi-exit model into sequential components and deploy them in separate endpoints, e.g., a small, local on-device part and a large, cloud-based part. For bringing DNNs to the Internet of Things (IoT), partitioning is instrumental as it reduces the transmissions between endpoints, a major bottleneck. In Sec 5.1, on a partitioning scenario, we show that our attack can force excessive transmissions. + +# 3 EXPERIMENTAL SETUP + +Datasets. We use two datasets: CIFAR-10 (Krizhevsky et al., 2009) and Tiny-ImageNet (Tiny). For testing the cross-domain transferability of our attacks, we use the CIFAR-100 dataset. + +Architectures and Hyper-parameters. To demonstrate that the vulnerability to adversarial slowdowns is common among multi-exit architectures, we experiment on two recent techniques: ShallowDeep Networks (SDNs) (Kaya et al., 2019) and MSDNets (Huang et al., 2018). These architectures were designed for different purposes: SDNs are generic and can convert any DNN into a multi-exit model, and MSDNets are custom designed for efficiency. We evaluate an MSDNet architecture (6 exits) and three SDN architectures, based on VGG-16 (Simonyan & Zisserman, 2014) (14 exits), ResNet-56 (He et al., 2016) (27 exits), and MobileNet (Howard et al., 2017) (14 exits). + +Metrics. We define the early-exit capability (EEC) curve of a multi-exit model to indicate the fraction of the test samples that exit early at a specific fraction of the model’s full inference cost. + +Figure 2 shows the EEC curves of our SDNs on Tiny ImageNet, assuming that the computation stops when there is a correct classification at an exit point. For example, VGG-16-based SDN model can correctly classify ${ \sim } 5 0 \%$ of the samples using ${ \sim } 5 0 \%$ of its full cost. Note that this stopping criterion is impractical; in Sec 4, we will discuss the practical ones. + +We define the early-exit efficacy, or efficacy in short, to quantify a model’s ability of utilizing its exit points. The efficacy of a multi-exit model is the area under its EEC curve, estimated via the trapezoidal rule. An ideal efficacy for a model is close to 1, when most of the input samples the computation stops very early; models that do not use their early exits have 0 efficacy. A model with low efficacy generally exhibits a higher latency; in a partitioned model, the low efficacy will cause more input transmissions to the cloud, and the latency is further amplified by the network round trips. A multi-exit model’s efficacy and accuracy are dictated by its stopping criteria, which we discuss in the next section. As for the classification performance, we report the Top-1 accuracy on the test data. + +![](images/2bf06ae00279a3cedeb142c999740cbd1b60e5613928b8a30f86d951383600a5.jpg) +tinyimagenet-vgg16bn(ACC 59.0) resnet56(ACC 53.5) mobilenet(ACC 59.5) +Figure 2: The EEC curves. Each curve shows the fraction of test samples a model classifies using a certain fraction of its full inference cost. ‘EFCY’ is short for the model’s efficacy. + +# 4 ATTACKING THE MULTI-EXIT ARCHITECTURES + +Setting. We consider the supervised classification setting with standard feedforward DNN architectures. A DNN model consists of $N$ blocks, or layers, that process the input sample, $x \in \mathbb { R } ^ { d }$ , from beginning to end and produce a classification. A classification, $F ( x , \theta ) \in \mathbf { \mathbb { R } } ^ { m }$ , is the predicted probability distribution of $x$ belonging to each label $y \in M = \{ 1 , . . . , m \}$ . Here, $\theta$ denotes the tunable parameters, or the weights, of the model. The parameters are learned on a training set $\mathcal { D }$ that contains multiple $( x _ { i } , y _ { i } )$ pairs; where $y _ { i }$ is the ground-truth label of the training sample $x _ { i }$ . We use $\theta _ { i }$ to denote the parameters at and before the $i ^ { t h }$ block; i.e., $\theta _ { i } \subset \theta _ { i + 1 }$ and $\theta _ { N } = \theta$ . Once a model is trained, its performance is then tested on a set of unseen samples, $s$ . + +Multi-Exit Architectures. A multi-exit model contains $K$ exit points—internal classifiers—attached to a model’s hidden blocks. We use $F _ { i }$ to denote the $i ^ { t h }$ exit point, which is attached to the $j ^ { t h }$ block. Using the output of the $j ^ { t h } \left( j < N \right)$ block on $x$ , $F _ { i }$ produces an internal classification, i.e., $F _ { i } ( x , \theta _ { j } )$ , which we simply denote as $F _ { i } ( x )$ . In our experiments, we set $K = N$ for SDNs, i.e., one internal classifier at each block and $K = 6$ for MSDNets. Given $F _ { i } ( x )$ , a multi-exit model uses deterministic criteria to decide between forwarding $x$ to compute $F _ { i + 1 } ( x )$ and stopping for taking the early-exit at this block. Bypassing early-exits decreases a network’s efficacy as each additional block increases the inference cost. Note that multi-exit models process each sample individually, not in batches. + +Practical Stopping Criteria. Ideally, a multi-exit model stops when it reaches a correct classification at an exit point, i.e., $\begin{array} { r } { \operatorname * { a r g m a x } _ { j \in M } F _ { i } ^ { ( \bar { j } ) } ( x ) = \hat { y } _ { i } = y ; y } \end{array}$ is the ground-truth label. However, for unseen samples, this is impractical as $y$ is unknown. The prior work has proposed two simple strategies to judge whether ${ \hat { y } } _ { i } = y$ : $F _ { i } ( x )$ ’s entropy (Teerapittayanon et al., 2016; Huang et al., 2018) or its confidence (Kaya et al., 2019). Our attack (see Sec 4.3) leverages the fact that a uniform $F _ { i } ( x )$ has both the highest entropy and the lowest confidence. For generality, we experiment with both confidence-based—SDNs—and entropy-based—MSDNets—strategies. + +A strategy selects confidence, or entropy, thresholds, $T _ { i }$ , that determine whether the model should take the $\overline { { i } } ^ { t h }$ exit for an input sample. Conservative $T _ { i }$ ’s lead to fewer early exits and the opposite hurts the accuracy as the estimate of whether ${ \hat { y } } _ { i } = y$ becomes unreliable. As utility is a major practical concern, we set $T _ { i }$ ’s for balancing between efficiency and accuracy. On a holdout set, we set the thresholds to maximize a model’s efficacy while keeping its relative accuracy drop (RAD) over its maximum accuracy within $5 \%$ and $15 \%$ . We refer to these two settings as RAD $123 \%$ and RAD ${ < } 1 5 \%$ Table 2 (first segment) shows how accuracy and efficacy change in each setting. + +# 4.1 THREAT MODEL + +We consider an adversary who aims to decrease the early-exit efficacy of a victim model. The attacker crafts an imperceptible adversarial perturbation, $v \in \mathbb { R } ^ { d }$ that, when added to a test-time sample $x \in S$ , prevents the model from taking early-exits. + +Adversary’s Capabilities. The attacker is able to modify the victim’s test-time samples to apply the perturbations, e.g., by compromising a camera that collects the data for inference. To ensure the imperceptibility, we focus on $\ell _ { \infty }$ norm bounded perturbations as they (i) are well are studied; (ii) have successful defenses (Madry et al., 2018); (iii) have prior extension to multi-exit models (Hu et al., 2020); and (iv) are usually the most efficient to craft. We show results on $\ell _ { 2 }$ and $\ell _ { 1 }$ perturbations in Appendix C. In line with the prior work, we bound the perturbations as follows: for CIFAR-10, $\lvert | v \rvert | _ { \infty } \leq \epsilon = 0 . 0 3$ (Madry et al., 2017), $| | v | | _ { 1 } \leq 8$ (Tramèr & Boneh, 2019) and $| | v | | _ { 2 } \leq 0 . 3 5$ (Chen et al., 2017); for Tiny ImageNet, $| | v | | _ { \infty } \leq \epsilon = 0 . 0 3$ (Yang et al., 2019), $| | \boldsymbol { v } | | _ { 1 } \le \mathrm { \dot { 1 6 } }$ and $| | v | | _ { 2 } \leq 0 . 6$ + +Adversary’s Knowledge. To assess the security vulnerability of multi-exit architectures, we study white-box scenarios, i.e., the attacker knows all the details of the victim model, including its $\mathcal { D }$ and $\theta$ . Further, in Sec 5.2, we study more practical black-box scenarios, i.e., the attacker crafts $v$ on a surrogate model and applies it to an unknown victim model. + +Adversary’s Goals. We consider three DeepSloth variants, (i) the standard, (ii) the universal and (iii) the class-universal. The adversary, in (i) crafts a different $v$ for each $x \in S$ ; in (ii) crafts a single $v$ for all $x \in S$ ; in (iii) crafts a single $v$ for a target class $i \in M$ . Further, although the adversary does not explicitly target it; we observe that DeepSloth usually hurts the accuracy. By modifying the objective function we describe in Sec 4.3, we also experiment with DeepSloth variants that can explicitly preserve or hurt the accuracy, in addition to causing slowdowns. + +# 4.2 STANDARD ADVERSARIAL ATTACKS DO NOT CAUSE DELAYS + +To motivate DeepSloth, we first evaluate whether previous adversarial attacks have any effect on the efficacy of multi-exit models. These attacks add imperceptible perturbations to a victim’s test-time samples to force misclassifications. We experiment with the standard PGD attack (Madry et al., 2017); PGD-avg and PGD-max variants against multi-exit models (Hu et al., 2020) and the Universal Adversarial Perturbation (UAP) attack that crafts a single perturbation for all test samples (MoosaviDezfooli et al., 2017). Table 1 summarizes our findings that these attacks, although they hurt the accuracy, fail to cause any meaningful decrease in efficacy. In many cases, we observe that the attacks actually increase the efficacy. These experiments help us to identify the critical elements of the objective function of an attack that decreases the efficacy. + +Table 1: Impact of existing evasion attacks on efficacy. Each entry shows a model’s efficacy (left) and accuracy (right) when subjected to the respective attack. The multi-exit models are trained on CIFAR-10 and use RAD ${ < } 5 \%$ as their early-exit strategy. + +
NETWORKNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)UAP
VGG-160.77 /89%0.79 / 29%0.85 /10%0.81/27%0.71/68%
REsNET-560.52 / 87%0.55 / 12%0.82/1%0.70/ 6%0.55 / 44%
MOBILENET0.83/87%0.85 /14%0.93/ 3%0.89 / 12%0.77 / 60%
+ +# 4.3 THE DEEPSLOTH ATTACK + +The Layer-Wise Objective Function. Figure 3 shows that the attacks that only optimize for the final output, e.g., PGD or UAP, do not perturb the model’s earlier layer representations. This does not bypass the early-exits, which makes these attacks ineffective for decreasing the efficacy. Therefore, we modify the objective functions of adversarial example-crafting algorithms to incorporate the outputs of all $F _ { i } | i < K$ . For crafting $\ell _ { \infty }$ , $\ell _ { 2 }$ and $\ell _ { 1 }$ -bounded perturbations, we adapt the PGD (Madry et al., 2017), the DDN (Rony et al., 2019) and the SLIDE algorithms (Tramèr & Boneh, 2019), respectively. Next, we describe how we modify the PGD algorithm—we modify the others similarly: + +$$ +v ^ { t + 1 } = \prod _ { | | v | | _ { \infty } < \epsilon } \left( v ^ { t } + \alpha \operatorname { s g n } \left( { \nabla } _ { v } \sum _ { x \in D ^ { \prime } } \sum _ { 0 < i < K } { \mathcal { L } } \left( F _ { i } \left( x + v \right) , \bar { y } \right) \right) \right) +$$ + +Here, $t$ is the current attack iteration; $\alpha$ is the step size; $\Pi$ is the projection operator that enforces $| | v | | _ { \infty } < \epsilon$ and $\mathcal { L }$ is the cross-entropy loss function. The selection of $\mathcal { D } ^ { \prime }$ determines the type of the attack. For the standard variant: ${ \mathcal { D } } ^ { \prime } = \{ x \}$ , i.e., a single test-time sample. For the universal variant: $\mathcal { D } ^ { \prime } = \mathcal { D }$ , i.e., the whole training set. For the class-universal variant against the target class $i \in M$ : $\mathcal { D } ^ { \prime } = \{ ( x , y ) \in \mathcal { D } | y = i \}$ , i.e., the training set samples from the $i ^ { t h }$ class. Finally, $\bar { y }$ is the target label distribution our objective pushes $F _ { i } ( x )$ towards. Next, we explain how we select $\bar { y }$ . + +Pushing $F _ { i } ( x )$ Towards a Uniform Distribution. Despite including all $F _ { i }$ , attacks such as PGDavg and PGD-max $\mathrm { H u }$ et al., 2020) still fail to decrease efficacy. How these attacks select $\bar { y }$ reflects their goal of causing misclassifications and, therefore, they trigger errors in early-exits, i.e., $\mathrm { a r g m a x } _ { j \in M } \bar { F } _ { i } ^ { ( j ) } ( x ) = \bar { y } \bar { \ne y }$ . However, as the early-exits still have high confidence, or low entropy, the model still stops its computation early. We select $\bar { y }$ as a uniform distribution over the class labels, i.e., $\bar { y } ^ { ( i ) } = 1 / m$ . This ensures that $( x + v )$ bypasses common stopping criteria as a uniform $F _ { i } ( x )$ has both the lowest confidence and the highest entropy. + +# 5 EMPIRICAL EVALUATION + +Here, we present the results for $\ell _ { \infty }$ DeepSloth against two SDNs—VGG-16 and MobileNet-based— and against the MSDNets. In the Appendix, we report the hyperparameters; the $\ell _ { 1 }$ and $\ell _ { 2 }$ attacks; the results on ResNet-56-based SDNs; the cost of the attacks; and some perturbed samples. Overall, we observe that $\ell _ { \infty }$ -bounded perturbations are more effective for slowdowns. The optimization challenges might explain this, as $\ell _ { 1 }$ and $\ell _ { 2 }$ attacks are usually harder to optimize (Carlini & Wagner, 2017; Tramèr & Boneh, 2019). Unlike objectives for misclassifications, the objective for slowdowns involves multiple loss terms and optimizes over all the output logits. + +# 5.1 WHITE-BOX SCENARIOS + +Perturbations Eliminate Early-Exits. Table 2 (second segment) shows that the victim models have $\sim 0$ efficacy on the samples perturbed by DeepSloth. Across the board, the attack makes the early-exit completely ineffective and force the victim models to forward all input samples till the end. Further, DeepSloth also drops the victim’s accuracy by $7 5 - 9 9 \%$ , comparable to the PGD attack. These results give an answer to our main research question: the multi-exit mechanisms are vulnerable and their benefits can be maliciously offset by adversarial input perturbations. In particular, as SDN modification mitigates overthinking in standard, non-adaptive DNNs (Kaya et al., 2019), DeepSloth also leads SDN-based models to overthink on almost all samples by forcing extra computations. + +Note that crafting a single perturbation requires multiple back-propagations through the model and more floating points operations $( F L O P s )$ than the forward pass. The high cost of crafting, relative to the computational damage to the victim, might make this vulnerability unattractive for the adversary. In the next sections, we highlight scenarios where this vulnerability might lead to practical exploitation. First, we show that in an IoT-like scenarios, the input transmission is a major bottleneck and DeepSloth can exploit it. Second, we evaluate universal DeepSloth attacks that enable the adversary to craft the perturbation only once and reuse it on multiple inputs. + +Attacking an IoT Scenario. Many IoT scenarios, e.g., health monitoring for elderly (Park et al., 2017), require collecting data from edge devices and making low-latency inferences on this data. However, complex deep learning models are impractical for low-power edge devices, such as an Arduino, that are common in the IoT scenarios (Chen & Ran, 2019). For example, on standard hardware, an average inference takes MSDNet model on Tiny ImageNet 35M FLOPs and ${ \sim } 1 0 \mathrm { m s }$ . + +A potential solution is sending the inputs from the edge to a cloud model, which then returns the prediction. Even in our optimistic estimate with a nearby AWS EC2 instance, this back-and-forth introduces ${ \sim } 1 1 \mathrm { m s }$ latency per inference. Model partitioning alleviates this bottleneck by splitting a multi-exit model into two; deploying the small first part at the edge and the large second part at the cloud (De Coninck et al., 2015). The edge part sends an input only when its prediction does not meet the stopping criteria. For example, the first early-exit of MSDNets sends only $5 \%$ and $67 \%$ of all test samples, on CIFAR-10 and Tiny ImageNet, respectively. This leads to a lower average latency per inference, i.e., from 11ms down to $0 . 5 \mathrm { m s }$ and $7 . 4 \mathrm { m s }$ , respectively. + +Table 2: The effectiveness of $\ell _ { \infty }$ DeepSloth. ‘ $\mathrm { 2 A D { < } } 5 , 1 5 \%$ ’ columns list the results in each early-exit setting. Each entry includes the model’s efficacy (left) and accuracy (right). The class-universal attack’s results are an average of 10 classes. ‘TI’: Tiny ImageNet and $\mathbf { \dot { C } } 1 0 ^ { \circ }$ : CIFAR-10. + +
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 / 85%0.89 / 85%0.77 /88%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 /57%0.51 / 52%0.42 /57%0.59 /51%
DEEPSLOTH
C100.06 / 17%0.06 / 17%0.01 /13%0.04 / 16%0.01 /12%0.06 /16%
TI0.06 /7%0.06 /7%0.00 /2%0.01/2%0.02 / 6%0.04 / 6%
UNIVERSAL DEEPSLOTH
C100.85 / 65%0.85 /65%0.62 / 65%0.86 / 60%0.73 / 61%0.90 / 59%
TI0.58 / 46%0.81 /41%0.31 / 47%0.44 / 44%0.33 / 47%0.51 / 43%
CLASS-UNIVERSAL DEEPSLOTH
C100.82/32%0.82 /32%0.47 /35%0.78 /33%0.60 /30%0.85 /27%
TI0.41 / 21%0.71 /17%0.20 /28%0.33 /27%0.21 /27%0.38 /25%
+ +The adversary we study uses DeepSloth perturbations to force the edge part to send all the input samples to the cloud. For the victim, we deploy MSDNet models that we split into two parts at their first exit point. Targeting the first part with DeepSloth forces it to send $96 \%$ and $9 9 . 9 7 \%$ of all test samples to the second part. This increases average inference latency to ${ \sim } 1 1 \mathrm { m s }$ and invalidates the benefits of model partitioning. In this scenario, perturbing each sample takes ${ \sim } 2 \mathrm { m s }$ on a Tesla V-100 GPU, i.e., the time adversary spends is amplified by $1 . 5 – 5 \times$ as the victim’s latency increase. + +Reusable Universal Perturbations. The universal attacks, although limited, are a practical as the adversary can reuse the same perturbation indefinitely to cause minor slowdowns. Table 2 (third segment) shows that they decrease the efficacy by $3- 2 1 \%$ and the accuracy by $1 5 \mathrm { - } 2 5 \%$ , over the baselines. Having a less conservative early-exit strategy, e.g., RAD ${ < } 1 5 \%$ , increases the resilience to the attack at the cost of accuracy. Further, MSDNets are fairly resilient with only $3- 9 \%$ efficacy drop; whereas SDNs are more vulnerable with $12 \mathrm { - } 2 1 \%$ drop. The attack is also slightly more effective on the more complex task, Tiny ImageNet, as the early-exits become easier to bypass. Using random noise as a baseline, i.e., $v \sim U ^ { d } ( - \epsilon , \epsilon )$ , we find that at most it decreases the efficacy by ${ \sim } 3 \%$ . + +In the universal attack, we observe a phenomenon: it pushes the samples towards a small subset of all classes. For example, ${ \sim } 1 7 \%$ of the perturbed samples are classified into the ’bird’ class of CIFAR-10; up from ${ \sim } 1 0 \%$ for the clean samples. Considering certain classes are distant in the feature space, e.g., ’truck’ and ’bird’; we expect the class-universal variant to be more effective. The results in Table 2 (fourth segment) confirm our intuition. We see that this attack decreases the baseline efficacy by $8 - 5 0 \%$ and the accuracy by $5 0 \mathrm { - } 6 5 \%$ . We report the average results across multiple classes; however, we observe that certain classes are slightly more vulnerable to this attack. + +Feature Visualization of DeepSloth. In Figure 3, to shed light on how DeepSloth differs from prior attacks, e.g., PGD and PGD-avg, we visualize a model’s hidden block (layer) features on the original and perturbed test-time samples. We observe that in an earlier block (left panel), DeepSloth seems to disrupt the original features slightly more than the PGD attacks. Leaving earlier representations intact prevents PGDs from bypassing the early-exits. The behaviors of the attacks diverge in the middle blocks (middle panel). Here, DeepSloth features remain closer to the original features than prior attacks. The significant disruption of prior attacks leads to high-confidence misclassifications and fails to bypass early-exits. In the later block (right panel), we see that the divergent behavior persists. + +Preserving or Hurting the Accuracy with DeepSloth. Here, we aim to answer whether DeepSloth can be applied when the adversary explicitly aims to cause or prevent misclassifications, while still causing slowdowns. Our main threat model has no explicit goal regarding misclassifications that hurt the user of the model, i.e., who consumes the output of the model. Whereas, slowdowns additionally hurt the executor or the owner of the model through the computations and latency increased at the cloud providers. In some ML-in-the-cloud scenarios, where these two are different actors, the adversary might aim to target only the executor or both the executor and the user. To this end, we modify our objective function to push $F _ { i } ( x )$ towards a slightly non-uniform distribution, favoring either the ground truth label for preventing misclassifications or a wrong label for causing them. We test this idea on our VGG-16-based SDN model on CIFAR-10 in RAD ${ < } 5 \%$ setting. We see that DeepSloth for preserving the accuracy leads to $81 \%$ accuracy with 0.02 efficacy and DeepSloth for hurting the accuracy leads to $4 \%$ accuracy with 0.01 efficacy—the original DeepSloth led to $13 \%$ accuracy with 0.01 efficacy. These results show the flexibility of DeepSloth and how it could be modified depending on the attacker’s goals. + +![](images/eb4db376fcea7527d2b8a56c5918847e8a4e1a6ae201be0033e35f93b1112bd1.jpg) +Figure 3: Visualising features against attacks using UMAP. VGG-16’s 3rd (left), 8th (middle), and 14th (right) hidden block features on CIFAR-10’s ’dog’ class (Best viewed in color, zoomed in). + +5.2 TOWARDS A BLACK-BOX ATTACK: TRANSFERABILITY OF DEEPSLOTH + +Transferability of adversarial examples imply that they can still hurt a model that they were not crafted on (Tramèr et al., 2017a; Liu et al., 2017). Even though white-box attacks are important to expose the vulnerability, black-box attacks, by requiring fewer assumptions, are more practical. Here, on four distinct scenarios, we investigate whether DeepSloth is transferable. Based on the scenario’s constraints, we (i) train a surrogate model; (ii) craft the DeepSloth samples on it; and (iii) use these samples on the victim. We run these experiments on CIFAR-10 in the RAD ${ < } 5 \%$ . + +Cross-Architecture. First, we relax the assumption that the attacker knows the victim architecture. We evaluate the transferability between a VGG-16-based SDN an an MSDNet—all trained using the same $\mathcal { D }$ . We find that the samples crafted on the MSDNet can slowdown the SDN: reducing its efficacy to 0.63 (from 0.77) and accuracy to $78 \%$ (from $8 8 \%$ ). Interestingly, the opposite seems not to be the case: on the samples crafted against the SDN, the MSDNet still has 0.87 efficacy (from 0.89) and $73 \%$ accuracy (from $8 5 \%$ ). This hints that DeepSloth transfers if the adversary uses an effective multi-exit models as the surrogate. + +Limited Training Set Knowledge. Second, we relax the assumption that the attacker knows the victim’s training set, $\mathcal { D }$ . Here, the attacker only knows a random portion of $\mathcal { D }$ , i.e., $10 \%$ , $2 5 \%$ , and $50 \%$ . We use VGG-16 architecture for both the surrogate and victim models. In the $10 \%$ , $2 5 \%$ and $50 \%$ settings, respectively, the attacks reduce the victim’s efficacy to 0.66, 0.5, 0.45 and 0.43 (from 0.77); its accuracy to $81 \%$ , $73 \%$ , $72 \%$ and $74 \%$ (from $8 8 \%$ ). Overall, the more limited the adversary’s $\mathcal { D }$ is, the less generalization ability the surrogate has and the less transferable the attacks are. + +Cross-Domain. Third, we relax the assumption that the attacker exactly knows the victim’s task. Here, the attacker uses $\mathcal { D } _ { \boldsymbol { \mathcal { J } } }$ to train the surrogate, different from the victim’s $\mathcal { D }$ altogether. We use a VGG-16 on CIFAR-100 as the surrogate and attack a VGG-16-based victim model on CIFAR-10. This transfer attack reduces the victim’s efficacy to 0.63 (from 0.77) and its accuracy to $83 \%$ (from $8 8 \%$ ). We see that the cross-domain attack might be more effective than the limited $\mathcal { D }$ scenarios. This makes DeepSloth particularly dangerous as the attacker, without knowing the victim’s $\mathcal { D }$ , can collect a similar dataset and still slowdown the victim. We hypothesize the transferability of earlier layer features in CNNs (Yosinski et al., 2014) enables the perturbations attack to transfer from one domain to another, as long as they are similar enough. + +Cross-Mechnanism. Finally, we test the scenario where the victim uses a completely different mechanism than a multi-exit architecture to implement input adaptiveness, i.e., SkipNet (Wang et al., 2018). A SkipNet, a modified residual network, selectively skips convolutional blocks based on the activations of the previous layer and, therefore, does not include any internal classifiers. We use a pre-trained SkipNet on CIFAR-10 that reduces the average computation for each input sample by ${ \sim } 5 0 \%$ over an equivalent ResNet and achieves ${ \sim } 9 4 \%$ accuracy. We then feed DeepSloth samples crafted on a MSDNet to this SkipNet, which reduces its average computational saving to ${ \sim } 3 2 \%$ $3 6 \%$ less effective) and its accuracy to $37 \%$ . This result suggests that the two different mechanisms have more in common than previously known and might share the vulnerability. We believe that understanding the underlying mechanisms through which adaptive models save computation is an important research question for future work. + +# 6 STANDARD ADVERSARIAL TRAINING IS NOT A COUNTERMEASURE + +In this section, we examine whether a defender can adapt a standard countermeasure against adversarial perturbations, adversarial training (AT) (Madry et al., 2018), to mitigate our attack. AT decreases a model’s sensitivity to perturbations that significantly change the model’s outputs. While this scheme is effective against adversarial examples that aim to trigger misclassifications; it is unclear whether using our DeepSloth samples for AT can also robustify a multi-exit model against slowdown attacks. + +To evaluate, we train our multi-exit models as follows. We first take a base network—VGG-16—and train it on CIFAR-10 on PGD-10 adversarial examples. We then convert the resulting model into a multi-exit architecture, using the modification from (Kaya et al., 2019). During this conversion, we adversarially train individual exit points using PGD-10, PGD-10 (avg.), PGD-10 (max.), and DeepSloth; similar to (Hu et al., 2020). Finally, we measure the efficacy and accuracy of the trained models against PGD-20, PGD-20 (avg.), PGD-20 (max.), and DeepSloth, on CIFAR-10’s test-set. + +Table 3: Evaluating adversarial training against slowdown attacks. Each entry includes the model’s efficacy score (left) and accuracy (right). Results are on CIFAR-10, in the RAD ${ < } 5 \%$ setting. + +
ADV. TRAININGNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)DEEPSLOTH
UNDEFENDED0.77 /89%0.79/29%0.85 /10%0.81/27%0.01 / 13%
PGD-100.61/ 72%0.55 /38%0.64 /23%0.58 /29%0.33 / 70%
PGD-10 (AVG.)0.53 / 72%0.47 / 36%0.47 /35%0.47 /35%0.32 / 70%
PGD-10 (MAX.)0.57 /72%0.51/37%0.54 / 30%0.52 /34%0.32 / 70%
OURS0.74/72%0.71/38%0.82 /14%0.77 /21%0.44/ 67%
OURS + PGD-100.61/73%0.55 /38%0.63 /23%0.58 /28%0.33 / 70%
+ +Our results in Table 3 verify that AT provides resilience against all PGD attacks. Besides, AT provides some resilience to our attack: DeepSloth reduces the efficacy to ${ \sim } 0 . 3 2$ on robust models vs. 0.01 on the undefended one. However, we identify a trade-off between the robustness and efficiency of multi-exits. Compared to the undefended model, on clean samples, we see that robust models have lower efficacy—0.77 vs. $0 . 5 3 \sim 0 . 6 1$ . We observe that the model trained only with our DeepSloth samples (Ours) can recover the efficacy on both the clean and our DeepSloth samples, but this model loses its robustness against PGD attacks. Moreover, when we train a model on both our DeepSloth samples and PGD-10 $\mathrm { \ O u r s } + \mathrm { P G D } { - } 1 0$ ), the trained model suffers from low efficacy. Our results imply that a defender may require an out-of-the-box defense, such as flagging the users whose queries bypass the early-exits more often than clean samples for which the multi-exit network was calibrated. + +# 7 CONCLUSIONS + +This work exposes the vulnerability of input-adaptive inference mechanisms against adversarial slowdowns. As a vehicle for exploring this vulnerability systematically, we propose DeepSloth, an attack that introduces imperceptible adversarial perturbations to test-time inputs for offsetting the computational benefits of multi-exit inference mechanisms. We show that a white-box attack, which perturbs each sample individually, eliminates any computational savings these mechanisms provide. We also show that it is possible to craft universal slowdown perturbations, which can be reused, and transferable samples, in a black-box setting. Moreover, adversarial training, a standard countermeasure for adversarial perturbations, is not effective against DeepSloth. Our analysis suggests that slowdown attacks are a realistic, yet under-appreciated, threat against adaptive models. + +# ACKNOWLEDGMENT + +We thank the anonymous reviewers for their feedback. This research was partially supported by the Department of Defense. + +# REFERENCES + +Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp. 274–283, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http: //proceedings.mlr.press/v80/athalye18a.html. + +Ron Banner, Itay Hubara, Elad Hoffer, and Daniel Soudry. Scalable methods for 8-bit training of neural networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett (eds.), Advances in Neural Information Processing Systems 31, pp. 5145–5153. Curran Associates, Inc., 2018. URL http://papers.nips.cc/paper/ 7761-scalable-methods-for-8-bit-training-of-neural-networks.pdf. + +N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (SP), pp. 39–57, 2017. + +Jiasi Chen and Xukan Ran. Deep learning with edge computing: A review. Proceedings of the IEEE, 107(8): 1655–1674, 2019. + +Pin-Yu Chen, Yash Sharma, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh. Ead: elastic-net attacks to deep neural networks via adversarial examples. arXiv preprint arXiv:1709.04114, 2017. + +Elias De Coninck, Tim Verbelen, Bert Vankeirsbilck, Steven Bohez, Pieter Simoens, Piet Demeester, and Bart Dhoedt. Distributed neural networks for internet of things: The big-little approach. In International Internet of Things Summit, pp. 484–492. Springer, 2015. + +Michael Figurnov, Maxwell D. Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov. Spatially Adaptive Computation Time for Residual Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. + +Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. URL http://arxiv.org/abs/1412. 6572. + +Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. + +Mirazul Haque, Anki Chauhan, Cong Liu, and Wei Yang. Ilfo: Adversarial attack on adaptive neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. + +K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016. + +Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6): 82–97, 2012. + +Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, and Qun Liu. DynaBERT: Dynamic BERT with Adaptive Width and Depth. In Advances in Neural Information Processing Systems, 2020. + +Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications, 2017. + +C. Hu, W. Bao, D. Wang, and F. Liu. Dynamic adaptive dnn surgery for inference acceleration on the edge. In IEEE INFOCOM 2019 - IEEE Conference on Computer Communications, pp. 1423–1431, 2019. doi: 10.1109/INFOCOM.2019.8737614. + +Ting-Kuei Hu, Tianlong Chen, Haotao Wang, and Zhangyang Wang. Triple wins: Boosting accuracy, robustness and efficiency together by enabling input-adaptive inference. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id $\equiv$ rJgzzJHtDB. + +Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Weinberger. Multiscale dense networks for resource efficient image classification. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $\equiv$ Hk2aImxAb. + +Nathan Inkawhich, Wei Wen, Hai (Helen) Li, and Yiran Chen. Feature space perturbations yield more transferable adversarial examples. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. + +Weiwen Jiang, Edwin H.-M. Sha, Xinyi Zhang, Lei Yang, Qingfeng Zhuge, Yiyu Shi, and Jingtong Hu. Achieving super-linear speedup across multi-fpga for real-time dnn inference. ACM Trans. Embed. Comput. Syst., 18(5s), October 2019. ISSN 1539-9087. doi: 10.1145/3358192. URL https://doi.org/10.1145/3358192. + +Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. In Proceedings of the TwentySecond International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS’17, pp. 615–629, New York, NY, USA, 2017. Association for Computing Machinery. ISBN 9781450344654. doi: 10.1145/3037697.3037698. URL https://doi.org/10.1145/3037697. 3037698. + +Yigitcan Kaya, Sanghyun Hong, and Tudor Dumitra¸s. Shallow-Deep Networks: Understanding and mitigating ˘ network overthinking. In Proceedings of the 2019 International Conference on Machine Learning (ICML), Long Beach, CA, Jun 2019. + +Alex Krizhevsky, Geoffrey Hinton, et al. Learning Multiple Layers of Features from Tiny Images. 2009. + +Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. + +Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236, 2016. + +M. Lecuyer, V. Atlidakis, R. Geambasu, D. Hsu, and S. Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE Symposium on Security and Privacy (SP), pp. 656–672, 2019. + +En Li, Zhi Zhou, and Xu Chen. Edge intelligence: On-demand deep learning model co-inference with deviceedge synergy. In Proceedings of the 2018 Workshop on Mobile Edge Communications, MECOMM’18, pp. 31–36, New York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450359061. doi: 10.1145/3229556.3229562. URL https://doi.org/10.1145/3229556.3229562. + +Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. CoRR, abs/1608.08710, 2016. URL http://arxiv.org/abs/1608.08710. + +Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. + +Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. URL https://openreview. net/forum?id $\equiv$ Sys6GJqxl. + +Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. + +Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=rJzIBfZAb. + +Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. + +N. Papernot, P. McDaniel, S. Jha, M. Fredrikson, Z. B. Celik, and A. Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy (EuroS P), pp. 372–387, 2016. + +Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pp. 506–519, 2017. + +Se Jin Park, Murali Subramaniyam, Seoung Eun Kim, Seunghee Hong, Joo Hyeong Lee, Chan Min Jo, and Youngseob Seo. Development of the elderly healthcare monitoring system with iot. In Advances in Human Factors and Ergonomics in Healthcare, pp. 309–315. Springer, 2017. + +Jérôme Rony, Luiz G Hafemann, Luiz S Oliveira, Ismail Ben Ayed, Robert Sabourin, and Eric Granger. Decoupling direction and norm for efficient gradient-based l2 adversarial attacks and defenses. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4322–4330, 2019. + +Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition, 2014. + +Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ rJUYGxbCW. + +C. Szegedy, Wei Liu, Yangqing Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9, 2015. + +Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. URL http://arxiv.org/abs/1312.6199. + +Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 6105–6114, Long Beach, California, USA, 09–15 Jun 2019. PMLR. URL http://proceedings.mlr.press/v97/tan19a. html. + +Ben Taylor, Vicent Sanz Marco, Willy Wolff, Yehia Elkhatib, and Zheng Wang. Adaptive deep learning model selection on embedded systems. ACM SIGPLAN Notices, 53(6):31–43, 2018. + +S. Teerapittayanon, B. McDanel, and H. T. Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pp. 2464–2469, 2016. + +Tiny. Tiny ImageNet Visual Recognition Challenge. http://tiny-imagenet.herokuapp.com/. Accessed: 2020-09-28. + +Florian Tramèr and Dan Boneh. Adversarial training and robustness for multiple perturbations. In Advances in Neural Information Processing Systems, pp. 5866–5876, 2019. + +Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017a. + +Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017b. + +Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E. Gonzalez. SkipNet: Learning Dynamic Routing in Convolutional Networks. In The European Conference on Computer Vision (ECCV), September 2018. + +Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. arXiv preprint arXiv:1704.01155, 2017. + +Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. arXiv preprint arXiv:1905.11971, 2019. + +Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? In Advances in neural information processing systems, pp. 3320–3328, 2014. + +Li Zhou, Hao Wen, Radu Teodorescu, and David H.C. Du. Distributing deep neural networks with containerized partitions at the edge. In 2nd USENIX Workshop on Hot Topics in Edge Computing (HotEdge 19), Renton, WA, July 2019. USENIX Association. URL https://www.usenix.org/conference/hotedge19/ presentation/zhou. + +Wangchunshu Zhou, Canwen Xu, Tao Ge, Julian McAuley, Ke Xu, and Furu Wei. BERT Loses Patience: Fast and Robust Inference with Early Exit. In Advances in Neural Information Processing Systems, 2020. + +# A MOTIVATING EXAMPLES + +Here, we discuss two exemplary scenarios where an adversary can exploit the slowdown attacks. + +• (Case 1) Attacks on cloud-based IoT applications. In most cases, cloud-based IoT applications, such as Apple Siri, Google Now, or Microsoft Cortana, run their DNN inferences in the cloud. This cloud-only approach puts all the computational burden on cloud servers and increases the communications between the servers and IoT devices. In consequence, recent work (Kang et al., 2017; Li et al., 2018; Zhou et al., 2019) utilizes multi-exit architectures for bringing computationally expensive models, e.g. language models (Zhou et al., 2020; Hou et al., 2020), in the cloud to IoT (or mobile) devices. They split a multi-exit model into two partitions and deploy each of them to a server and IoT devices, respectively. Under this scheme, the cloud server only takes care of complex inputs that the shallow partition cannot correctly classify at the edge. As a result, one can reduce the computations in the cloud and decrease communications between the cloud and edge. + +On the other hand, our adversary, by applying human-imperceptible perturbations, can convert simple inputs into complex inputs. These adversarial inputs will bypass early-exits and, as a result, reduce (or even offset) the computational and communication savings provided by prior work. + +Here, a defender may deploy DoS defenses such as firewalls or rate-limiting. In this setting, the attacker may not cause DoS because defenses keep the communications between the server and IoT devices under a certain-level. Nevertheless, the attacker still increases: (i) the computations at the edge (by making inputs skip early-exits) and (ii) the number of samples that cloud servers process. Recall that a VGG-16 SDN model classifies $90 \%$ of clean CIFAR-10 instances correctly at the first exit. If the adversarial examples crafted by the attacker bypass only the first exit, one can easily increase the computations on IoT devices and make them send requests to the cloud. + +• (Case 2) Attacks on real-time DNN inference for resource- and time-constrained scenarios. Recent work on the real-time systems (Hu et al., 2019; Jiang et al., 2019) harnesses multi-exit architectures and model partitioning as a solution to optimize real-time DNN inference for resourceand time-constrained scenarios. Hu et al. (2019) showed a real-world prototype of an optimal model partitioning, which is based on a self-driving car video dataset, can improve latency and throughput of partitioned models on the cloud and edge by $6 . 5 – 1 4 \times$ , respectively. + +However, the prior work does not consider the danger of slowdown attacks; our threat model has not been discussed before in the literature. Our results in Sec 5 suggest that slowdown can be induced adversarially, potentially violating real-time guarantees. For example, our attacker can force partitioned models on the cloud and edge to use maximal computations for inference. Further, the same adversarial examples also require the inference results from the model running on the cloud, which potentially increases the response time of the edge devices by $1 . 5 – 5 \times$ . Our work showed that multi-exit architectures should be used with caution in real-time systems. + +# B HYPERPARAMETERS + +In our experiments, we use the following hyperparameters to craft adversarial perturbations. + +$\ell _ { \infty }$ -based DeepSloth. We find that $\ell _ { \infty }$ -based DeepSloth does not require careful tuning. For the standard attack, we set the total number of iterations to 30 and the step size to $\alpha = 0 . 0 0 2$ . For the modified attacks for hurting or preserving the accuracy, we set the total number of iterations to 75 and the step size to $\alpha = 0 . 0 0 1$ . We compute the standard perturbations using the entire $1 0 \mathrm { k }$ test-set samples in CIFAR-10 and Tiny Imagenet. For the universal variants, we set the total number of iterations to 12 and reduce the initial step size of $\alpha = 0 . 0 0 5$ by a factor of 10 every 4 iterations. To compute a universal perturbation, we use randomly chosen 250 (CIFAR-10) and 200 (Tiny Imagenet) training samples. + +$\ell _ { 2 }$ -based DeepSloth. For both the standard and universal attacks, we set the total number of iterations to 550 and the step size $\gamma$ to 0.1. Our initial perturbation has the $\ell _ { 2 }$ -norm of 1.0. Here, we use the same number of samples for crafting the standard and universal perturbations as the $\ell _ { \infty }$ -based attacks. + +$\ell _ { 1 }$ -based DeepSloth. For our standard $\ell _ { 1 }$ -based DeepSloth, we set the total number of iterations to 250, the step size $\alpha$ to 0.5, and the gradient sparsity to 99. For the universal variants, we reduce the total number of iterations to 100 and set the gradient sparsity to 90. Other hyperparameters remain the same. We use the same number of samples as the $\ell _ { \infty }$ -based attacks, to craft the perturbations. + +# C EMPIRICAL EVALUATION OF $\ell _ { 1 }$ AND $\ell _ { 2 }$ DEEPSLOTH + +Table 4 and Table 5 shows the effectiveness of $\ell _ { 1 }$ -based and $\ell _ { 2 }$ -based DeepSloth attacks, respectively. + +Table 4: The effectiveness of $\ell _ { 1 }$ DeepSloth. ‘ $R \mathrm { A D } { < } 5 , 1 5 \%$ ’ columns list the results in each early-exit setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal attack’s results are an average of 10 classes. ‘TI’ is Tiny Imagenet and $\mathbf { \dot { C } } 1 0 ^ { \circ }$ is CIFAR-10. + +
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 /89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 / 57%0.51 /52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.36 / 51%0.35 / 51%0.12 /36%0.34 /45%0.18 / 41%0.49 / 53%
TI0.23 /37%0.51 / 40%0.08 /22%0.15 / 25%0.08 /33%0.19 /35%
UNIVERSAL DEEPSLOTH
C100.89 /83%0.89 /83%0.75 /85%0.88 / 75%0.82 /85%0.92 /77%
TI0.64 / 55%0.83/50%0.38 /57%0.51 / 52%0.41 / 57%0.59 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88/ 73%0.88 / 73%0.69 / 78%0.86 /67%0.76 /74%0.89 / 65%
TI0.64 /54%0.83 /49%0.39 / 59%0.50 / 58%0.41 / 60%0.58 / 53%
+ +Table 5: The effectiveness of $\ell _ { 2 }$ DeepSloth. $\cdot \mathrm { R A D } { < } 5 , 1 5 \%$ ’ columns list the results in each early-exit setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal attack’s results are an average of 10 classes. ‘TI’ is Tiny Imagenet and $\mathbf { \dot { C } } 1 0 ^ { \circ }$ is CIFAR-10. + +
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 / 89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 / 50%0.39 / 57%0.51 / 52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.52 / 64%0.52 /64%0.22 /60%0.45 / 62%0.23 /46%0.48 / 55%
TI0.24 /42%0.52 / 44%0.13 /35%0.21 /36%0.12 /38%0.25 /40%
UNIVERSAL DEEPSLOTH
C100.89 /81%0.89 /81%0.75 /87%0.88 / 76%0.81/84%0.92 /76%
TI0.63 / 54%0.82 /48%0.38 /56%0.51 / 52%0.41 / 56%0.58 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88 /73%0.88 / 73%0.71 /81%0.86 / 70%0.76 /76%0.89 /66%
TI0.64 / 53%0.83 /49%0.38 /57%0.50 / 57%0.41 /58%0.58 / 53%
+ +Our results show that the $\ell _ { 1 } \cdot$ - and $\ell _ { 2 }$ -based attacks are less effective than the $\ell _ { \infty }$ -based attacks. In contrast to the $\ell _ { \infty }$ -based attacks that eliminate the efficacy of victim multi-exit models, the $\ell _ { 1 }$ - and $\ell _ { 2 }$ -based attacks reduce the efficacy of the same models by $0 . 2 4 { \sim } 0 . 6 5$ . Besides, the accuracy drops caused by $\ell _ { 1 } \cdot$ - and $\ell _ { 2 }$ -based attacks are in $6 \sim 2 1 \%$ , smaller than that of $\ell _ { \infty }$ -based DeepSloth $( 7 5 \sim 9 9 \% )$ . Moreover, we see that the universal variants of $\ell _ { 1 }$ - and $\ell _ { 2 }$ -based attacks can barely reduce the efficacy of multi-exit models—they decrease the efficacy up to 0.08 and the accuracy by $12 \%$ . + +# D EMPIRICAL EVALUATION OF DEEPSLOTH ON RESNET56 + +Table 6 shows the the effectiveness of our DeepSloth attacks on ResNet56-base models. + +Table 6: The effectiveness of DeepSloth on the ResNet-based models. $\cdot \mathrm { R A D } { < } 5 , 1 5 \%$ ’ columns list the results in each early-exit setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal attack’s results are an average of 10 classes. + +
NETWORKRESNET(lo)RESNET (l1)RESNET(2)
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.52 /87%0.69 /80%0.52/87%0.69 / 80%0.51 /87%0.69 / 80%
TI0.25 / 51%0.39 /46%0.25 /51%0.39 /46%0.25 /51%0.39 / 46%
DEEPSLOTH
C100.00 /19%0.01 /19%0.05 /43%0.18 /47%0.06 /45%0.17 /48%
TI0.00 /7%0.01 / 7%0.04 /27%0.10 /28%0.05 /34%0.13 /35%
UNIVERSAL DEEPSLOTH
C100.35 / 63%0.59 /60%0.49 /84%0.68 /75%0.48 /85%0.67 /76%
TI0.25 / 25%0.34 / 37%0.25 / 51%0.39 / 46%0.25 / 51%0.38 /46%
CLASS-UNIVERSAL DEEPSLOTH
C100.23 / 33%0.48 / 29%0.39 / 70%0.60 / 61%0.39 / 71%0.60 / 61%
TI0.11 /21%0.23 /18%0.23 / 51%0.36 /46%0.23 / 50%0.36 /46%
+ +Our results show that ResNet56-based models are vulnerable to all the $\ell _ { \infty }$ , $\ell _ { 2 }$ , and $\ell _ { 1 }$ -based DeepSloth attacks. Using our $\ell _ { \infty }$ -based DeepSloth, the attacker can reduce the efficacy of the victim models to $0 . 0 0 { \sim } 0 . 0 1$ and the accuracy by $3 9 \sim 6 8 \%$ . Besides, the $\ell _ { 2 }$ , and $\ell _ { 1 }$ -based attacks also decrease the efficacy to $0 . 0 4 { \sim } 0 . 1 8$ and the accuracy by $1 1 { \sim } 4 4 \%$ . Compared to the results on MSDNet, VGG16, and MobileNet in Table 4 and 5, the same attacks are more effective. The universal variants decrease the efficacy up to 0.21 and the accuracy up to $24 \%$ . In particular, the $\ell _ { 2 }$ , and $\ell _ { 1 }$ -based attacks (on CIFAR-10 models) are effective than the same attacks on MSDNet, VGG16, and MobileNet models. + +# E COST OF CRAFTING DEEPSLOTH SAMPLES + +In Table 7, we compare the cost of DeepSloth with other attack algorithms on a VGG16-based CIFAR10 model—executed on a single Nvidia Tesla-V100 GPU. For the universal DeepSloth, we measure the execution time for crafting a perturbation using one batch (250 samples) of the training set. For the other attacks, we measure the time for perturbing the whole test set of CIFAR-10. Our DeepSloth takes roughly the same time as the PGD and PGD-avg attacks and significantly less time than the PGD-max attack. Our + +
ATTACKSTIME (SEC.)
PGD-2038
PGD-20 (AVG.)48
PGD-20 (MAX.)475
DEEPSLOTH44
UNIVERSAL DEEPSLOTH2
+ +Table 7: Time it takes to craft attacks. +universal DeepSloth takes only 2 seconds ( $1 0 \mathrm { x }$ faster than DeepSloth) as it only uses 250 samples. + +# F ADVERSARIAL EXAMPLES FROM STANDARD ATTACKS AND DEEPSLOTH + +In Figure 4, we visualize the adversarial examples from the PGD, UAP and our DeepSloth attacks. + +![](images/0bdcaea5dbd66848101c989036785e7979050485b14922c9d380c77e29ee961e.jpg) +Figure 4: Adversarial examples from the standard and our DeepSloth attacks. The leftmost column shows the clean images. In the next four columns, we show adversarial examples from PGD, PGD (avg.), PGD (max.), and UAP attacks, respectively. The last four columns include adversarial examples from the three variants of DeepSloth. Each row corresponds to each sample, and the last row contains the average $\ell _ { \mathrm { i n f } }$ -norm of the perturbations over the eight samples in each attack. \ No newline at end of file diff --git a/parse/train/9xC2tWEwBD/9xC2tWEwBD_content_list.json b/parse/train/9xC2tWEwBD/9xC2tWEwBD_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ff58435cd44a57d97963679c0e069ca223f25ae9 --- /dev/null +++ b/parse/train/9xC2tWEwBD/9xC2tWEwBD_content_list.json @@ -0,0 +1,1816 @@ +[ + { + "type": "text", + "text": "A PANDA? NO, IT’S A SLOTH: SLOWDOWN ATTACKS ON ADAPTIVE MULTI-EXIT NEURAL NETWORK INFERENCE ", + "text_level": 1, + "bbox": [ + 176, + 98, + 797, + 145 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanghyun Hong∗, Yigitcan Kaya ˇ ∗, Ionut,-Vlad Modoranu†, Tudor Dumitras, ", + "bbox": [ + 186, + 167, + 704, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Maryland, College Park, USA †Alexandru Ioan Cuza University, Ias,i, Romania ", + "bbox": [ + 186, + 185, + 501, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shhong@cs.umd.edu, yigitcan@cs.umd.edu, modoranu.ionut.vlad@hotmail.com, tudor@umd.edu ", + "bbox": [ + 184, + 217, + 589, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 280, + 544, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent increases in the computational demands of deep neural networks (DNNs), combined with the observation that most input samples require only simple models, have sparked interest in input-adaptive multi-exit architectures, such as MSDNets or Shallow-Deep Networks. These architectures enable faster inferences and could bring DNNs to low-power devices, e.g., in the Internet of Things (IoT). However, it is unknown if the computational savings provided by this approach are robust against adversarial pressure. In particular, an adversary may aim to slowdown adaptive DNNs by increasing their average inference time—a threat analogous to the denial-of-service attacks from the Internet. In this paper, we conduct a systematic evaluation of this threat by experimenting with three generic multi-exit DNNs (based on VGG16, MobileNet, and ResNet56) and a custom multi-exit architecture, on two popular image classification benchmarks (CIFAR-10 and Tiny ImageNet). To this end, we show that adversarial example-crafting techniques can be modified to cause slowdown, and we propose a metric for comparing their impact on different architectures. We show that a slowdown attack reduces the efficacy of multi-exit DNNs by $9 0 { - } 1 0 0 \\%$ , and it amplifies the latency by $1 . 5 – 5 \\times$ in a typical IoT deployment. We also show that it is possible to craft universal, reusable perturbations and that the attack can be effective in realistic black-box scenarios, where the attacker has limited knowledge about the victim. Finally, we show that adversarial training provides limited protection against slowdowns. These results suggest that further research is needed for defending multi-exit architectures against this emerging threat. Our code is available at https://github.com/sanghyun-hong/deepsloth. ", + "bbox": [ + 232, + 310, + 766, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 655, + 336, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The inference-time computational demands of deep neural networks (DNNs) are increasing, owing to the “going deeper\" (Szegedy et al., 2015) strategy for improving accuracy: as a DNN gets deeper, it progressively gains the ability to learn higher-level, complex representations. This strategy has enabled breakthroughs in many tasks, such as image classification (Krizhevsky et al., 2012) or speech recognition (Hinton et al., 2012), at the price of costly inferences. For instance, with $4 \\times$ more inference cost, a 56-layer ResNet (He et al., 2016) improved the Top-1 accuracy on ImageNet by $19 \\%$ over the 8-layer AlexNet. This trend continued with the 57-layer state-of-the-art EfficientNet (Tan & Le, 2019): it improved the accuracy by $10 \\%$ over ResNet, with $9 \\times$ costlier inferences. ", + "bbox": [ + 174, + 686, + 825, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The accuracy improvements stem from the fact that the deeper networks fix the mistakes of the shallow ones (Huang et al., 2018). This implies that some samples, which are already correctly classified by shallow networks, do not necessitate the extra complexity. This observation has motivated research on input-adaptive mechanisms, in particular, multi-exit architectures (Teerapittayanon et al., 2016; Huang et al., 2018; Kaya et al., 2019; Hu et al., 2020). Multi-exit architectures save computation by making input-specific decisions about bypassing the remaining layers, once the model becomes confident, and are orthogonal to techniques that achieve savings by permanently modifying the model (Li et al., 2016; Banner et al., 2018; Han et al., 2015; Taylor et al., 2018). Figure 1 illustrates how a multi-exit model (Kaya et al., 2019), based on a standard VGG-16 architecture, correctly classifies a selection of test images from ‘Tiny ImageNet’ before the final layer. We see that more typical samples, which have more supporting examples in the training set, require less depth and, therefore, less computation. ", + "bbox": [ + 174, + 804, + 825, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 825, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is unknown if the computational savings provided by multi-exit architectures are robust against adversarial pressure. Prior research showed that DNNs are vulnerable to a wide range of attacks, which involve imperceptible input perturbations (Szegedy et al., 2014; Goodfellow et al., 2015; Papernot et al., 2016; Hu et al., 2020). Considering that a multi-exit model, on the worst-case input, does not provide any computational savings, we ask: Can the savings from multi-exit models be maliciously negated by input perturbations? As some natural inputs do require the full depth of the model, it may be possible to craft adversarial examples that delay the correct decision; it is unclear, however, how many inputs can be delayed with imperceptible perturbations. Furthermore, it is unknown if universal versions of these adversarial examples exist, if the examples transfer across multi-exit architectures and datasets, or if existing defenses (e.g. adversarial training) are effective against slowdown attacks. ", + "bbox": [ + 174, + 181, + 562, + 373 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/746d8efde4dde3f7f81419a6f62cb228c7561c4c925bbdc393202621421b32de.jpg", + "image_caption": [ + "Figure 1: Simple to complex inputs. Some Tiny ImageNet images a VGG-16 model can correctly classify if computation stops at the $1 ^ { s t }$ , $5 ^ { t h }$ , and $1 4 ^ { t h }$ layers. " + ], + "image_footnote": [], + "bbox": [ + 588, + 181, + 810, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 375, + 823, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Threat Model. We consider a new threat against DNNs, analogous to the denial-of-service $( D o S )$ attacks that have been plaguing the Internet for decades. By imperceptibly perturbing the input to trigger this worst-case, the adversary aims to slow down the inferences and increase the cost of using the DNN. This is an important threat for many practical applications, which impose strict limits on the responsiveness and resource usage of DNN models (e.g. in the Internet-of-Things (Taylor et al., 2018)), because the adversary could push the victim outside these limits. For example, against a commercial image classification system, such as Clarifai.com, a slowdown attack might waste valuable computational resources. Against a model partitioning scheme, such as Big-Little (De Coninck et al., 2015), it might introduce network latency by forcing excessive transmissions between local and remote models. A slowdown attack aims to force the victim to do more work than the adversary, e.g. by amplifying the latency needed to process the sample or by crafting reusable perturbations. The adversary may have to achieve this with incomplete information about the multi-exit architecture targeted, the training data used by the victim or the classification task (see discussion in Appendix A). ", + "bbox": [ + 173, + 410, + 826, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Contributions. To our best knowledge, we conduct the first study of the robustness of multi-exit architectures against adversarial slowdowns. To this end, we find that examples crafted by prior evasion attacks (Madry et al., 2017; Hu et al., 2020) fail to bypass the victim model’s early exits, and we show that an adversary can adapt such attacks to the goal of model slowdown by modifying its objective function. We call the resulting attack DeepSloth. We also propose an efficacy metric for comparing slowdowns across different multi-exit architectures. We experiment with three generic multi-exit DNNs (based on VGG16, ResNet56 and MobileNet) (Kaya et al., 2019) and a speciallydesigned multi-exit architecture, MSDNets (Huang et al., 2018), on two popular image classification benchmarks (CIFAR-10 and Tiny ImageNet). We find that DeepSloth reduces the efficacy of multiexit DNNs by $9 0 { - } 1 0 0 \\%$ , i.e., the perturbations render nearly all early exits ineffective. In a scenario typical for IoT deployments, where the model is partitioned between edge devices and the cloud, our attack amplifies the latency by $1 . 5 – 5 \\times$ , negating the benefits of model partitioning. We also show that it is possible to craft a universal DeepSloth perturbation, which can slow down the model on either all or a class of inputs. While more constrained, this attack still reduces the efficacy by $5- 4 5 \\%$ . Further, we observe that DeepSloth can be effective in some black-box scenarios, where the attacker has limited knowledge about the victim. Finally, we show that a standard defense against adversarial samples—adversarial training—is inadequate against slowdowns. Our results suggest that further research will be required for protecting multi-exit architectures against this emerging security threat. ", + "bbox": [ + 173, + 599, + 826, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 102, + 343, + 117 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Examples and Defenses. Prior work on adversarial examples has shown that DNNs are vulnerable to test-time input perturbations (Szegedy et al., 2014; Goodfellow et al., 2015; Papernot et al., 2017; Carlini & Wagner, 2017; Madry et al., 2018). An adversary who wants to maximize a model’s error on specific test-time samples can introduce human-imperceptible perturbations to these samples. Moreover, an adversary can also exploit a surrogate model for launching the attack and still hurt an unknown victim (Athalye et al., 2018; Tramèr et al., 2017b; Inkawhich et al., 2019). This transferability leads to adversarial examples in more practical black-box scenarios. Although many defenses (Kurakin et al., 2016; Xu et al., 2017; Song et al., 2018; Liao et al., 2018; Lecuyer et al., 2019) have been proposed against this threat, adversarial training (AT) has become the frontrunner (Madry et al., 2018). In Sec 5, we evaluate the vulnerability of multi-exit DNNs to adversarial slowdowns in white-box and black-box scenarios. In Sec 6, we show that standard AT and its simple adaptation to our perturbations are not sufficient for preventing slowdown attacks. ", + "bbox": [ + 174, + 135, + 825, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Input-Adaptive Inference. Recent input-adaptive DNN architectures have brought two seemingly distant goals closer: achieving both high predictive quality and computational efficiency. There are two types of input-adaptive DNNs: adaptive neural networks (AdNNs) and multi-exit architectures. During the inference, AdNNs (Wang et al., 2018; Figurnov et al., 2017) dynamically skip a certain part of the model to reduce the number of computations. This mechanism can be used only for ResNet-based architectures as they facilitate skipping within a network. On the other hand, multi-exit architectures (Teerapittayanon et al., 2016; Huang et al., 2018; Kaya et al., 2019) introduce multiple side branches—or early-exits—to a model. During the inference on an input sample, these models can preemptively stop the computation altogether once the stopping criteria are met at one of the branches. Kaya et al. (2019) have also identified that standard, non-adaptive DNNs are susceptible to overthinking, i.e., their inability to stop computation leads to inefficient inferences on many inputs. ", + "bbox": [ + 174, + 309, + 825, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Haque et al. (2020) presented attacks specifically designed for reducing the energy-efficiency of AdNNs by using adversarial input perturbations. However, our work studies a new threat model that an adversary causes slowdowns on multi-exit architectures. By imperceptibly perturbing the inputs, our attacker can (i) introduce network latency to an infrastructure that utilizes multi-exit architectures and (ii) waste the victim’s computational resources. To quantify this vulnerability, we define a new metric to measure the impact of adversarial input perturbation on different multi-exit architectures (Sec 3). In Sec 5, we also study practical attack scenarios and the transferability of adversarial input perturbations crafted by our attacker. Moreover, we discuss the potential defense mechanisms against this vulnerability, by proposing a simple adaptation of adversarial training (Sec 6). To the best of our knowledge, our work is the first systematic study of this new vulnerability. ", + "bbox": [ + 174, + 468, + 825, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Partitioning. Model partitioning has been proposed to bring DNNs to resource-constrained devices (De Coninck et al., 2015; Taylor et al., 2018). These schemes split a multi-exit model into sequential components and deploy them in separate endpoints, e.g., a small, local on-device part and a large, cloud-based part. For bringing DNNs to the Internet of Things (IoT), partitioning is instrumental as it reduces the transmissions between endpoints, a major bottleneck. In Sec 5.1, on a partitioning scenario, we show that our attack can force excessive transmissions. ", + "bbox": [ + 174, + 616, + 825, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 176, + 720, + 398, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets. We use two datasets: CIFAR-10 (Krizhevsky et al., 2009) and Tiny-ImageNet (Tiny). For testing the cross-domain transferability of our attacks, we use the CIFAR-100 dataset. ", + "bbox": [ + 173, + 753, + 823, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architectures and Hyper-parameters. To demonstrate that the vulnerability to adversarial slowdowns is common among multi-exit architectures, we experiment on two recent techniques: ShallowDeep Networks (SDNs) (Kaya et al., 2019) and MSDNets (Huang et al., 2018). These architectures were designed for different purposes: SDNs are generic and can convert any DNN into a multi-exit model, and MSDNets are custom designed for efficiency. We evaluate an MSDNet architecture (6 exits) and three SDN architectures, based on VGG-16 (Simonyan & Zisserman, 2014) (14 exits), ResNet-56 (He et al., 2016) (27 exits), and MobileNet (Howard et al., 2017) (14 exits). ", + "bbox": [ + 174, + 790, + 825, + 887 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Metrics. We define the early-exit capability (EEC) curve of a multi-exit model to indicate the fraction of the test samples that exit early at a specific fraction of the model’s full inference cost. ", + "bbox": [ + 174, + 895, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows the EEC curves of our SDNs on Tiny ImageNet, assuming that the computation stops when there is a correct classification at an exit point. For example, VGG-16-based SDN model can correctly classify ${ \\sim } 5 0 \\%$ of the samples using ${ \\sim } 5 0 \\%$ of its full cost. Note that this stopping criterion is impractical; in Sec 4, we will discuss the practical ones. ", + "bbox": [ + 173, + 103, + 825, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the early-exit efficacy, or efficacy in short, to quantify a model’s ability of utilizing its exit points. The efficacy of a multi-exit model is the area under its EEC curve, estimated via the trapezoidal rule. An ideal efficacy for a model is close to 1, when most of the input samples the computation stops very early; models that do not use their early exits have 0 efficacy. A model with low efficacy generally exhibits a higher latency; in a partitioned model, the low efficacy will cause more input transmissions to the cloud, and the latency is further amplified by the network round trips. A multi-exit model’s efficacy and accuracy are dictated by its stopping criteria, which we discuss in the next section. As for the classification performance, we report the Top-1 accuracy on the test data. ", + "bbox": [ + 174, + 166, + 496, + 401 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2bf06ae00279a3cedeb142c999740cbd1b60e5613928b8a30f86d951383600a5.jpg", + "image_caption": [ + "tinyimagenet-vgg16bn(ACC 59.0) resnet56(ACC 53.5) mobilenet(ACC 59.5) ", + "Figure 2: The EEC curves. Each curve shows the fraction of test samples a model classifies using a certain fraction of its full inference cost. ‘EFCY’ is short for the model’s efficacy. " + ], + "image_footnote": [], + "bbox": [ + 511, + 172, + 807, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 ATTACKING THE MULTI-EXIT ARCHITECTURES ", + "text_level": 1, + "bbox": [ + 174, + 438, + 602, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setting. We consider the supervised classification setting with standard feedforward DNN architectures. A DNN model consists of $N$ blocks, or layers, that process the input sample, $x \\in \\mathbb { R } ^ { d }$ , from beginning to end and produce a classification. A classification, $F ( x , \\theta ) \\in \\mathbf { \\mathbb { R } } ^ { m }$ , is the predicted probability distribution of $x$ belonging to each label $y \\in M = \\{ 1 , . . . , m \\}$ . Here, $\\theta$ denotes the tunable parameters, or the weights, of the model. The parameters are learned on a training set $\\mathcal { D }$ that contains multiple $( x _ { i } , y _ { i } )$ pairs; where $y _ { i }$ is the ground-truth label of the training sample $x _ { i }$ . We use $\\theta _ { i }$ to denote the parameters at and before the $i ^ { t h }$ block; i.e., $\\theta _ { i } \\subset \\theta _ { i + 1 }$ and $\\theta _ { N } = \\theta$ . Once a model is trained, its performance is then tested on a set of unseen samples, $s$ . ", + "bbox": [ + 173, + 479, + 826, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Exit Architectures. A multi-exit model contains $K$ exit points—internal classifiers—attached to a model’s hidden blocks. We use $F _ { i }$ to denote the $i ^ { t h }$ exit point, which is attached to the $j ^ { t h }$ block. Using the output of the $j ^ { t h } \\left( j < N \\right)$ block on $x$ , $F _ { i }$ produces an internal classification, i.e., $F _ { i } ( x , \\theta _ { j } )$ , which we simply denote as $F _ { i } ( x )$ . In our experiments, we set $K = N$ for SDNs, i.e., one internal classifier at each block and $K = 6$ for MSDNets. Given $F _ { i } ( x )$ , a multi-exit model uses deterministic criteria to decide between forwarding $x$ to compute $F _ { i + 1 } ( x )$ and stopping for taking the early-exit at this block. Bypassing early-exits decreases a network’s efficacy as each additional block increases the inference cost. Note that multi-exit models process each sample individually, not in batches. ", + "bbox": [ + 173, + 599, + 825, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Practical Stopping Criteria. Ideally, a multi-exit model stops when it reaches a correct classification at an exit point, i.e., $\\begin{array} { r } { \\operatorname * { a r g m a x } _ { j \\in M } F _ { i } ^ { ( \\bar { j } ) } ( x ) = \\hat { y } _ { i } = y ; y } \\end{array}$ is the ground-truth label. However, for unseen samples, this is impractical as $y$ is unknown. The prior work has proposed two simple strategies to judge whether ${ \\hat { y } } _ { i } = y$ : $F _ { i } ( x )$ ’s entropy (Teerapittayanon et al., 2016; Huang et al., 2018) or its confidence (Kaya et al., 2019). Our attack (see Sec 4.3) leverages the fact that a uniform $F _ { i } ( x )$ has both the highest entropy and the lowest confidence. For generality, we experiment with both confidence-based—SDNs—and entropy-based—MSDNets—strategies. ", + "bbox": [ + 173, + 718, + 825, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A strategy selects confidence, or entropy, thresholds, $T _ { i }$ , that determine whether the model should take the $\\overline { { i } } ^ { t h }$ exit for an input sample. Conservative $T _ { i }$ ’s lead to fewer early exits and the opposite hurts the accuracy as the estimate of whether ${ \\hat { y } } _ { i } = y$ becomes unreliable. As utility is a major practical concern, we set $T _ { i }$ ’s for balancing between efficiency and accuracy. On a holdout set, we set the thresholds to maximize a model’s efficacy while keeping its relative accuracy drop (RAD) over its maximum accuracy within $5 \\%$ and $15 \\%$ . We refer to these two settings as RAD $123 \\%$ and RAD ${ < } 1 5 \\%$ Table 2 (first segment) shows how accuracy and efficacy change in each setting. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 THREAT MODEL ", + "text_level": 1, + "bbox": [ + 176, + 103, + 328, + 117 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider an adversary who aims to decrease the early-exit efficacy of a victim model. The attacker crafts an imperceptible adversarial perturbation, $v \\in \\mathbb { R } ^ { d }$ that, when added to a test-time sample $x \\in S$ , prevents the model from taking early-exits. ", + "bbox": [ + 174, + 132, + 825, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversary’s Capabilities. The attacker is able to modify the victim’s test-time samples to apply the perturbations, e.g., by compromising a camera that collects the data for inference. To ensure the imperceptibility, we focus on $\\ell _ { \\infty }$ norm bounded perturbations as they (i) are well are studied; (ii) have successful defenses (Madry et al., 2018); (iii) have prior extension to multi-exit models (Hu et al., 2020); and (iv) are usually the most efficient to craft. We show results on $\\ell _ { 2 }$ and $\\ell _ { 1 }$ perturbations in Appendix C. In line with the prior work, we bound the perturbations as follows: for CIFAR-10, $\\lvert | v \\rvert | _ { \\infty } \\leq \\epsilon = 0 . 0 3$ (Madry et al., 2017), $| | v | | _ { 1 } \\leq 8$ (Tramèr & Boneh, 2019) and $| | v | | _ { 2 } \\leq 0 . 3 5$ (Chen et al., 2017); for Tiny ImageNet, $| | v | | _ { \\infty } \\leq \\epsilon = 0 . 0 3$ (Yang et al., 2019), $| | \\boldsymbol { v } | | _ { 1 } \\le \\mathrm { \\dot { 1 6 } }$ and $| | v | | _ { 2 } \\leq 0 . 6$ ", + "bbox": [ + 174, + 181, + 825, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversary’s Knowledge. To assess the security vulnerability of multi-exit architectures, we study white-box scenarios, i.e., the attacker knows all the details of the victim model, including its $\\mathcal { D }$ and $\\theta$ . Further, in Sec 5.2, we study more practical black-box scenarios, i.e., the attacker crafts $v$ on a surrogate model and applies it to an unknown victim model. ", + "bbox": [ + 173, + 301, + 825, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversary’s Goals. We consider three DeepSloth variants, (i) the standard, (ii) the universal and (iii) the class-universal. The adversary, in (i) crafts a different $v$ for each $x \\in S$ ; in (ii) crafts a single $v$ for all $x \\in S$ ; in (iii) crafts a single $v$ for a target class $i \\in M$ . Further, although the adversary does not explicitly target it; we observe that DeepSloth usually hurts the accuracy. By modifying the objective function we describe in Sec 4.3, we also experiment with DeepSloth variants that can explicitly preserve or hurt the accuracy, in addition to causing slowdowns. ", + "bbox": [ + 173, + 364, + 825, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 STANDARD ADVERSARIAL ATTACKS DO NOT CAUSE DELAYS ", + "text_level": 1, + "bbox": [ + 176, + 472, + 642, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To motivate DeepSloth, we first evaluate whether previous adversarial attacks have any effect on the efficacy of multi-exit models. These attacks add imperceptible perturbations to a victim’s test-time samples to force misclassifications. We experiment with the standard PGD attack (Madry et al., 2017); PGD-avg and PGD-max variants against multi-exit models (Hu et al., 2020) and the Universal Adversarial Perturbation (UAP) attack that crafts a single perturbation for all test samples (MoosaviDezfooli et al., 2017). Table 1 summarizes our findings that these attacks, although they hurt the accuracy, fail to cause any meaningful decrease in efficacy. In many cases, we observe that the attacks actually increase the efficacy. These experiments help us to identify the critical elements of the objective function of an attack that decreases the efficacy. ", + "bbox": [ + 173, + 500, + 826, + 626 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/82e0346a0d05698abb99c5ca3e4c2c3416ecb56f0610de4ca3f806a7f479f01b.jpg", + "table_caption": [ + "Table 1: Impact of existing evasion attacks on efficacy. Each entry shows a model’s efficacy (left) and accuracy (right) when subjected to the respective attack. The multi-exit models are trained on CIFAR-10 and use RAD ${ < } 5 \\%$ as their early-exit strategy. " + ], + "table_footnote": [], + "table_body": "
NETWORKNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)UAP
VGG-160.77 /89%0.79 / 29%0.85 /10%0.81/27%0.71/68%
REsNET-560.52 / 87%0.55 / 12%0.82/1%0.70/ 6%0.55 / 44%
MOBILENET0.83/87%0.85 /14%0.93/ 3%0.89 / 12%0.77 / 60%
", + "bbox": [ + 173, + 695, + 833, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3 THE DEEPSLOTH ATTACK ", + "text_level": 1, + "bbox": [ + 174, + 796, + 395, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Layer-Wise Objective Function. Figure 3 shows that the attacks that only optimize for the final output, e.g., PGD or UAP, do not perturb the model’s earlier layer representations. This does not bypass the early-exits, which makes these attacks ineffective for decreasing the efficacy. Therefore, we modify the objective functions of adversarial example-crafting algorithms to incorporate the outputs of all $F _ { i } | i < K$ . For crafting $\\ell _ { \\infty }$ , $\\ell _ { 2 }$ and $\\ell _ { 1 }$ -bounded perturbations, we adapt the PGD (Madry et al., 2017), the DDN (Rony et al., 2019) and the SLIDE algorithms (Tramèr & Boneh, 2019), respectively. Next, we describe how we modify the PGD algorithm—we modify the others similarly: ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/86efb71bee2692f9ee3c0c9c37aeec2609a9c74c0132e9fd6eb8385d4f735434.jpg", + "text": "$$\nv ^ { t + 1 } = \\prod _ { | | v | | _ { \\infty } < \\epsilon } \\left( v ^ { t } + \\alpha \\operatorname { s g n } \\left( { \\nabla } _ { v } \\sum _ { x \\in D ^ { \\prime } } \\sum _ { 0 < i < K } { \\mathcal { L } } \\left( F _ { i } \\left( x + v \\right) , \\bar { y } \\right) \\right) \\right)\n$$", + "text_format": "latex", + "bbox": [ + 256, + 113, + 741, + 157 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $t$ is the current attack iteration; $\\alpha$ is the step size; $\\Pi$ is the projection operator that enforces $| | v | | _ { \\infty } < \\epsilon$ and $\\mathcal { L }$ is the cross-entropy loss function. The selection of $\\mathcal { D } ^ { \\prime }$ determines the type of the attack. For the standard variant: ${ \\mathcal { D } } ^ { \\prime } = \\{ x \\}$ , i.e., a single test-time sample. For the universal variant: $\\mathcal { D } ^ { \\prime } = \\mathcal { D }$ , i.e., the whole training set. For the class-universal variant against the target class $i \\in M$ : $\\mathcal { D } ^ { \\prime } = \\{ ( x , y ) \\in \\mathcal { D } | y = i \\}$ , i.e., the training set samples from the $i ^ { t h }$ class. Finally, $\\bar { y }$ is the target label distribution our objective pushes $F _ { i } ( x )$ towards. Next, we explain how we select $\\bar { y }$ . ", + "bbox": [ + 173, + 166, + 826, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pushing $F _ { i } ( x )$ Towards a Uniform Distribution. Despite including all $F _ { i }$ , attacks such as PGDavg and PGD-max $\\mathrm { H u }$ et al., 2020) still fail to decrease efficacy. How these attacks select $\\bar { y }$ reflects their goal of causing misclassifications and, therefore, they trigger errors in early-exits, i.e., $\\mathrm { a r g m a x } _ { j \\in M } \\bar { F } _ { i } ^ { ( j ) } ( x ) = \\bar { y } \\bar { \\ne y }$ . However, as the early-exits still have high confidence, or low entropy, the model still stops its computation early. We select $\\bar { y }$ as a uniform distribution over the class labels, i.e., $\\bar { y } ^ { ( i ) } = 1 / m$ . This ensures that $( x + v )$ bypasses common stopping criteria as a uniform $F _ { i } ( x )$ has both the lowest confidence and the highest entropy. ", + "bbox": [ + 173, + 257, + 826, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 EMPIRICAL EVALUATION ", + "text_level": 1, + "bbox": [ + 176, + 381, + 415, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we present the results for $\\ell _ { \\infty }$ DeepSloth against two SDNs—VGG-16 and MobileNet-based— and against the MSDNets. In the Appendix, we report the hyperparameters; the $\\ell _ { 1 }$ and $\\ell _ { 2 }$ attacks; the results on ResNet-56-based SDNs; the cost of the attacks; and some perturbed samples. Overall, we observe that $\\ell _ { \\infty }$ -bounded perturbations are more effective for slowdowns. The optimization challenges might explain this, as $\\ell _ { 1 }$ and $\\ell _ { 2 }$ attacks are usually harder to optimize (Carlini & Wagner, 2017; Tramèr & Boneh, 2019). Unlike objectives for misclassifications, the objective for slowdowns involves multiple loss terms and optimizes over all the output logits. ", + "bbox": [ + 174, + 411, + 826, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 WHITE-BOX SCENARIOS ", + "text_level": 1, + "bbox": [ + 174, + 526, + 387, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perturbations Eliminate Early-Exits. Table 2 (second segment) shows that the victim models have $\\sim 0$ efficacy on the samples perturbed by DeepSloth. Across the board, the attack makes the early-exit completely ineffective and force the victim models to forward all input samples till the end. Further, DeepSloth also drops the victim’s accuracy by $7 5 - 9 9 \\%$ , comparable to the PGD attack. These results give an answer to our main research question: the multi-exit mechanisms are vulnerable and their benefits can be maliciously offset by adversarial input perturbations. In particular, as SDN modification mitigates overthinking in standard, non-adaptive DNNs (Kaya et al., 2019), DeepSloth also leads SDN-based models to overthink on almost all samples by forcing extra computations. ", + "bbox": [ + 173, + 553, + 825, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that crafting a single perturbation requires multiple back-propagations through the model and more floating points operations $( F L O P s )$ than the forward pass. The high cost of crafting, relative to the computational damage to the victim, might make this vulnerability unattractive for the adversary. In the next sections, we highlight scenarios where this vulnerability might lead to practical exploitation. First, we show that in an IoT-like scenarios, the input transmission is a major bottleneck and DeepSloth can exploit it. Second, we evaluate universal DeepSloth attacks that enable the adversary to craft the perturbation only once and reuse it on multiple inputs. ", + "bbox": [ + 174, + 671, + 825, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attacking an IoT Scenario. Many IoT scenarios, e.g., health monitoring for elderly (Park et al., 2017), require collecting data from edge devices and making low-latency inferences on this data. However, complex deep learning models are impractical for low-power edge devices, such as an Arduino, that are common in the IoT scenarios (Chen & Ran, 2019). For example, on standard hardware, an average inference takes MSDNet model on Tiny ImageNet 35M FLOPs and ${ \\sim } 1 0 \\mathrm { m s }$ . ", + "bbox": [ + 174, + 776, + 825, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A potential solution is sending the inputs from the edge to a cloud model, which then returns the prediction. Even in our optimistic estimate with a nearby AWS EC2 instance, this back-and-forth introduces ${ \\sim } 1 1 \\mathrm { m s }$ latency per inference. Model partitioning alleviates this bottleneck by splitting a multi-exit model into two; deploying the small first part at the edge and the large second part at the cloud (De Coninck et al., 2015). The edge part sends an input only when its prediction does not meet the stopping criteria. For example, the first early-exit of MSDNets sends only $5 \\%$ and $67 \\%$ of all test samples, on CIFAR-10 and Tiny ImageNet, respectively. This leads to a lower average latency per inference, i.e., from 11ms down to $0 . 5 \\mathrm { m s }$ and $7 . 4 \\mathrm { m s }$ , respectively. ", + "bbox": [ + 174, + 853, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/15f95d76d72cb36008898a8f228ac096fa222a80b0bd4f1332610847e2a19f34.jpg", + "table_caption": [ + "Table 2: The effectiveness of $\\ell _ { \\infty }$ DeepSloth. ‘ $\\mathrm { 2 A D { < } } 5 , 1 5 \\%$ ’ columns list the results in each early-exit setting. Each entry includes the model’s efficacy (left) and accuracy (right). The class-universal attack’s results are an average of 10 classes. ‘TI’: Tiny ImageNet and $\\mathbf { \\dot { C } } 1 0 ^ { \\circ }$ : CIFAR-10. " + ], + "table_footnote": [], + "table_body": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 / 85%0.89 / 85%0.77 /88%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 /57%0.51 / 52%0.42 /57%0.59 /51%
DEEPSLOTH
C100.06 / 17%0.06 / 17%0.01 /13%0.04 / 16%0.01 /12%0.06 /16%
TI0.06 /7%0.06 /7%0.00 /2%0.01/2%0.02 / 6%0.04 / 6%
UNIVERSAL DEEPSLOTH
C100.85 / 65%0.85 /65%0.62 / 65%0.86 / 60%0.73 / 61%0.90 / 59%
TI0.58 / 46%0.81 /41%0.31 / 47%0.44 / 44%0.33 / 47%0.51 / 43%
CLASS-UNIVERSAL DEEPSLOTH
C100.82/32%0.82 /32%0.47 /35%0.78 /33%0.60 /30%0.85 /27%
TI0.41 / 21%0.71 /17%0.20 /28%0.33 /27%0.21 /27%0.38 /25%
", + "bbox": [ + 173, + 152, + 823, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 412, + 825, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The adversary we study uses DeepSloth perturbations to force the edge part to send all the input samples to the cloud. For the victim, we deploy MSDNet models that we split into two parts at their first exit point. Targeting the first part with DeepSloth forces it to send $96 \\%$ and $9 9 . 9 7 \\%$ of all test samples to the second part. This increases average inference latency to ${ \\sim } 1 1 \\mathrm { m s }$ and invalidates the benefits of model partitioning. In this scenario, perturbing each sample takes ${ \\sim } 2 \\mathrm { m s }$ on a Tesla V-100 GPU, i.e., the time adversary spends is amplified by $1 . 5 – 5 \\times$ as the victim’s latency increase. ", + "bbox": [ + 174, + 460, + 825, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reusable Universal Perturbations. The universal attacks, although limited, are a practical as the adversary can reuse the same perturbation indefinitely to cause minor slowdowns. Table 2 (third segment) shows that they decrease the efficacy by $3- 2 1 \\%$ and the accuracy by $1 5 \\mathrm { - } 2 5 \\%$ , over the baselines. Having a less conservative early-exit strategy, e.g., RAD ${ < } 1 5 \\%$ , increases the resilience to the attack at the cost of accuracy. Further, MSDNets are fairly resilient with only $3- 9 \\%$ efficacy drop; whereas SDNs are more vulnerable with $12 \\mathrm { - } 2 1 \\%$ drop. The attack is also slightly more effective on the more complex task, Tiny ImageNet, as the early-exits become easier to bypass. Using random noise as a baseline, i.e., $v \\sim U ^ { d } ( - \\epsilon , \\epsilon )$ , we find that at most it decreases the efficacy by ${ \\sim } 3 \\%$ . ", + "bbox": [ + 174, + 553, + 825, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the universal attack, we observe a phenomenon: it pushes the samples towards a small subset of all classes. For example, ${ \\sim } 1 7 \\%$ of the perturbed samples are classified into the ’bird’ class of CIFAR-10; up from ${ \\sim } 1 0 \\%$ for the clean samples. Considering certain classes are distant in the feature space, e.g., ’truck’ and ’bird’; we expect the class-universal variant to be more effective. The results in Table 2 (fourth segment) confirm our intuition. We see that this attack decreases the baseline efficacy by $8 - 5 0 \\%$ and the accuracy by $5 0 \\mathrm { - } 6 5 \\%$ . We report the average results across multiple classes; however, we observe that certain classes are slightly more vulnerable to this attack. ", + "bbox": [ + 173, + 671, + 825, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Visualization of DeepSloth. In Figure 3, to shed light on how DeepSloth differs from prior attacks, e.g., PGD and PGD-avg, we visualize a model’s hidden block (layer) features on the original and perturbed test-time samples. We observe that in an earlier block (left panel), DeepSloth seems to disrupt the original features slightly more than the PGD attacks. Leaving earlier representations intact prevents PGDs from bypassing the early-exits. The behaviors of the attacks diverge in the middle blocks (middle panel). Here, DeepSloth features remain closer to the original features than prior attacks. The significant disruption of prior attacks leads to high-confidence misclassifications and fails to bypass early-exits. In the later block (right panel), we see that the divergent behavior persists. ", + "bbox": [ + 174, + 776, + 825, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Preserving or Hurting the Accuracy with DeepSloth. Here, we aim to answer whether DeepSloth can be applied when the adversary explicitly aims to cause or prevent misclassifications, while still causing slowdowns. Our main threat model has no explicit goal regarding misclassifications that hurt the user of the model, i.e., who consumes the output of the model. Whereas, slowdowns additionally hurt the executor or the owner of the model through the computations and latency increased at the cloud providers. In some ML-in-the-cloud scenarios, where these two are different actors, the adversary might aim to target only the executor or both the executor and the user. To this end, we modify our objective function to push $F _ { i } ( x )$ towards a slightly non-uniform distribution, favoring either the ground truth label for preventing misclassifications or a wrong label for causing them. We test this idea on our VGG-16-based SDN model on CIFAR-10 in RAD ${ < } 5 \\%$ setting. We see that DeepSloth for preserving the accuracy leads to $81 \\%$ accuracy with 0.02 efficacy and DeepSloth for hurting the accuracy leads to $4 \\%$ accuracy with 0.01 efficacy—the original DeepSloth led to $13 \\%$ accuracy with 0.01 efficacy. These results show the flexibility of DeepSloth and how it could be modified depending on the attacker’s goals. ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb4db376fcea7527d2b8a56c5918847e8a4e1a6ae201be0033e35f93b1112bd1.jpg", + "image_caption": [ + "Figure 3: Visualising features against attacks using UMAP. VGG-16’s 3rd (left), 8th (middle), and 14th (right) hidden block features on CIFAR-10’s ’dog’ class (Best viewed in color, zoomed in). " + ], + "image_footnote": [], + "bbox": [ + 178, + 102, + 820, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 281, + 825, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 TOWARDS A BLACK-BOX ATTACK: TRANSFERABILITY OF DEEPSLOTH ", + "bbox": [ + 176, + 478, + 705, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferability of adversarial examples imply that they can still hurt a model that they were not crafted on (Tramèr et al., 2017a; Liu et al., 2017). Even though white-box attacks are important to expose the vulnerability, black-box attacks, by requiring fewer assumptions, are more practical. Here, on four distinct scenarios, we investigate whether DeepSloth is transferable. Based on the scenario’s constraints, we (i) train a surrogate model; (ii) craft the DeepSloth samples on it; and (iii) use these samples on the victim. We run these experiments on CIFAR-10 in the RAD ${ < } 5 \\%$ . ", + "bbox": [ + 174, + 510, + 825, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Architecture. First, we relax the assumption that the attacker knows the victim architecture. We evaluate the transferability between a VGG-16-based SDN an an MSDNet—all trained using the same $\\mathcal { D }$ . We find that the samples crafted on the MSDNet can slowdown the SDN: reducing its efficacy to 0.63 (from 0.77) and accuracy to $78 \\%$ (from $8 8 \\%$ ). Interestingly, the opposite seems not to be the case: on the samples crafted against the SDN, the MSDNet still has 0.87 efficacy (from 0.89) and $73 \\%$ accuracy (from $8 5 \\%$ ). This hints that DeepSloth transfers if the adversary uses an effective multi-exit models as the surrogate. ", + "bbox": [ + 174, + 602, + 825, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limited Training Set Knowledge. Second, we relax the assumption that the attacker knows the victim’s training set, $\\mathcal { D }$ . Here, the attacker only knows a random portion of $\\mathcal { D }$ , i.e., $10 \\%$ , $2 5 \\%$ , and $50 \\%$ . We use VGG-16 architecture for both the surrogate and victim models. In the $10 \\%$ , $2 5 \\%$ and $50 \\%$ settings, respectively, the attacks reduce the victim’s efficacy to 0.66, 0.5, 0.45 and 0.43 (from 0.77); its accuracy to $81 \\%$ , $73 \\%$ , $72 \\%$ and $74 \\%$ (from $8 8 \\%$ ). Overall, the more limited the adversary’s $\\mathcal { D }$ is, the less generalization ability the surrogate has and the less transferable the attacks are. ", + "bbox": [ + 174, + 707, + 825, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Domain. Third, we relax the assumption that the attacker exactly knows the victim’s task. Here, the attacker uses $\\mathcal { D } _ { \\boldsymbol { \\mathcal { J } } }$ to train the surrogate, different from the victim’s $\\mathcal { D }$ altogether. We use a VGG-16 on CIFAR-100 as the surrogate and attack a VGG-16-based victim model on CIFAR-10. This transfer attack reduces the victim’s efficacy to 0.63 (from 0.77) and its accuracy to $83 \\%$ (from $8 8 \\%$ ). We see that the cross-domain attack might be more effective than the limited $\\mathcal { D }$ scenarios. This makes DeepSloth particularly dangerous as the attacker, without knowing the victim’s $\\mathcal { D }$ , can collect a similar dataset and still slowdown the victim. We hypothesize the transferability of earlier layer features in CNNs (Yosinski et al., 2014) enables the perturbations attack to transfer from one domain to another, as long as they are similar enough. ", + "bbox": [ + 174, + 799, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Mechnanism. Finally, we test the scenario where the victim uses a completely different mechanism than a multi-exit architecture to implement input adaptiveness, i.e., SkipNet (Wang et al., 2018). A SkipNet, a modified residual network, selectively skips convolutional blocks based on the activations of the previous layer and, therefore, does not include any internal classifiers. We use a pre-trained SkipNet on CIFAR-10 that reduces the average computation for each input sample by ${ \\sim } 5 0 \\%$ over an equivalent ResNet and achieves ${ \\sim } 9 4 \\%$ accuracy. We then feed DeepSloth samples crafted on a MSDNet to this SkipNet, which reduces its average computational saving to ${ \\sim } 3 2 \\%$ $3 6 \\%$ less effective) and its accuracy to $37 \\%$ . This result suggests that the two different mechanisms have more in common than previously known and might share the vulnerability. We believe that understanding the underlying mechanisms through which adaptive models save computation is an important research question for future work. ", + "bbox": [ + 174, + 103, + 825, + 256 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 STANDARD ADVERSARIAL TRAINING IS NOT A COUNTERMEASURE ", + "text_level": 1, + "bbox": [ + 174, + 277, + 767, + 295 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we examine whether a defender can adapt a standard countermeasure against adversarial perturbations, adversarial training (AT) (Madry et al., 2018), to mitigate our attack. AT decreases a model’s sensitivity to perturbations that significantly change the model’s outputs. While this scheme is effective against adversarial examples that aim to trigger misclassifications; it is unclear whether using our DeepSloth samples for AT can also robustify a multi-exit model against slowdown attacks. ", + "bbox": [ + 174, + 310, + 825, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate, we train our multi-exit models as follows. We first take a base network—VGG-16—and train it on CIFAR-10 on PGD-10 adversarial examples. We then convert the resulting model into a multi-exit architecture, using the modification from (Kaya et al., 2019). During this conversion, we adversarially train individual exit points using PGD-10, PGD-10 (avg.), PGD-10 (max.), and DeepSloth; similar to (Hu et al., 2020). Finally, we measure the efficacy and accuracy of the trained models against PGD-20, PGD-20 (avg.), PGD-20 (max.), and DeepSloth, on CIFAR-10’s test-set. ", + "bbox": [ + 174, + 387, + 825, + 472 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7552a3da79778e2f12ef0c7b85c99a13a72681f7575d595cbdb355fabe60ff74.jpg", + "table_caption": [ + "Table 3: Evaluating adversarial training against slowdown attacks. Each entry includes the model’s efficacy score (left) and accuracy (right). Results are on CIFAR-10, in the RAD ${ < } 5 \\%$ setting. " + ], + "table_footnote": [], + "table_body": "
ADV. TRAININGNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)DEEPSLOTH
UNDEFENDED0.77 /89%0.79/29%0.85 /10%0.81/27%0.01 / 13%
PGD-100.61/ 72%0.55 /38%0.64 /23%0.58 /29%0.33 / 70%
PGD-10 (AVG.)0.53 / 72%0.47 / 36%0.47 /35%0.47 /35%0.32 / 70%
PGD-10 (MAX.)0.57 /72%0.51/37%0.54 / 30%0.52 /34%0.32 / 70%
OURS0.74/72%0.71/38%0.82 /14%0.77 /21%0.44/ 67%
OURS + PGD-100.61/73%0.55 /38%0.63 /23%0.58 /28%0.33 / 70%
", + "bbox": [ + 173, + 525, + 823, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our results in Table 3 verify that AT provides resilience against all PGD attacks. Besides, AT provides some resilience to our attack: DeepSloth reduces the efficacy to ${ \\sim } 0 . 3 2$ on robust models vs. 0.01 on the undefended one. However, we identify a trade-off between the robustness and efficiency of multi-exits. Compared to the undefended model, on clean samples, we see that robust models have lower efficacy—0.77 vs. $0 . 5 3 \\sim 0 . 6 1$ . We observe that the model trained only with our DeepSloth samples (Ours) can recover the efficacy on both the clean and our DeepSloth samples, but this model loses its robustness against PGD attacks. Moreover, when we train a model on both our DeepSloth samples and PGD-10 $\\mathrm { \\ O u r s } + \\mathrm { P G D } { - } 1 0$ ), the trained model suffers from low efficacy. Our results imply that a defender may require an out-of-the-box defense, such as flagging the users whose queries bypass the early-exits more often than clean samples for which the multi-exit network was calibrated. ", + "bbox": [ + 174, + 646, + 825, + 786 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7 CONCLUSIONS ", + "text_level": 1, + "bbox": [ + 176, + 808, + 328, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work exposes the vulnerability of input-adaptive inference mechanisms against adversarial slowdowns. As a vehicle for exploring this vulnerability systematically, we propose DeepSloth, an attack that introduces imperceptible adversarial perturbations to test-time inputs for offsetting the computational benefits of multi-exit inference mechanisms. We show that a white-box attack, which perturbs each sample individually, eliminates any computational savings these mechanisms provide. We also show that it is possible to craft universal slowdown perturbations, which can be reused, and transferable samples, in a black-box setting. Moreover, adversarial training, a standard countermeasure for adversarial perturbations, is not effective against DeepSloth. Our analysis suggests that slowdown attacks are a realistic, yet under-appreciated, threat against adaptive models. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENT ", + "text_level": 1, + "bbox": [ + 176, + 167, + 346, + 183 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers for their feedback. This research was partially supported by the Department of Defense. ", + "bbox": [ + 176, + 198, + 825, + 226 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 247, + 285, + 262 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp. 274–283, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http: //proceedings.mlr.press/v80/athalye18a.html. ", + "bbox": [ + 174, + 268, + 826, + 332 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ron Banner, Itay Hubara, Elad Hoffer, and Daniel Soudry. Scalable methods for 8-bit training of neural networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett (eds.), Advances in Neural Information Processing Systems 31, pp. 5145–5153. Curran Associates, Inc., 2018. URL http://papers.nips.cc/paper/ 7761-scalable-methods-for-8-bit-training-of-neural-networks.pdf. ", + "bbox": [ + 173, + 342, + 825, + 406 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy (SP), pp. 39–57, 2017. ", + "bbox": [ + 171, + 416, + 823, + 443 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jiasi Chen and Xukan Ran. Deep learning with edge computing: A review. Proceedings of the IEEE, 107(8): 1655–1674, 2019. ", + "bbox": [ + 173, + 452, + 823, + 478 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pin-Yu Chen, Yash Sharma, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh. Ead: elastic-net attacks to deep neural networks via adversarial examples. arXiv preprint arXiv:1709.04114, 2017. ", + "bbox": [ + 173, + 487, + 825, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Elias De Coninck, Tim Verbelen, Bert Vankeirsbilck, Steven Bohez, Pieter Simoens, Piet Demeester, and Bart Dhoedt. Distributed neural networks for internet of things: The big-little approach. In International Internet of Things Summit, pp. 484–492. Springer, 2015. ", + "bbox": [ + 173, + 522, + 825, + 563 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Michael Figurnov, Maxwell D. Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov. Spatially Adaptive Computation Time for Residual Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. ", + "bbox": [ + 173, + 570, + 823, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. URL http://arxiv.org/abs/1412. 6572. ", + "bbox": [ + 173, + 619, + 825, + 659 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. ", + "bbox": [ + 173, + 667, + 825, + 695 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mirazul Haque, Anki Chauhan, Cong Liu, and Wei Yang. Ilfo: Adversarial attack on adaptive neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. ", + "bbox": [ + 174, + 704, + 823, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016. ", + "bbox": [ + 171, + 752, + 823, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6): 82–97, 2012. ", + "bbox": [ + 174, + 787, + 825, + 839 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, and Qun Liu. DynaBERT: Dynamic BERT with Adaptive Width and Depth. In Advances in Neural Information Processing Systems, 2020. ", + "bbox": [ + 171, + 848, + 823, + 876 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications, 2017. ", + "bbox": [ + 176, + 885, + 823, + 922 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C. Hu, W. Bao, D. Wang, and F. Liu. Dynamic adaptive dnn surgery for inference acceleration on the edge. In IEEE INFOCOM 2019 - IEEE Conference on Computer Communications, pp. 1423–1431, 2019. doi: 10.1109/INFOCOM.2019.8737614. ", + "bbox": [ + 176, + 104, + 823, + 142 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ting-Kuei Hu, Tianlong Chen, Haotao Wang, and Zhangyang Wang. Triple wins: Boosting accuracy, robustness and efficiency together by enabling input-adaptive inference. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id $\\equiv$ rJgzzJHtDB. ", + "bbox": [ + 176, + 152, + 821, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Weinberger. Multiscale dense networks for resource efficient image classification. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $\\equiv$ Hk2aImxAb. ", + "bbox": [ + 174, + 200, + 823, + 239 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Nathan Inkawhich, Wei Wen, Hai (Helen) Li, and Yiran Chen. Feature space perturbations yield more transferable adversarial examples. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. ", + "bbox": [ + 173, + 248, + 826, + 286 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Weiwen Jiang, Edwin H.-M. Sha, Xinyi Zhang, Lei Yang, Qingfeng Zhuge, Yiyu Shi, and Jingtong Hu. Achieving super-linear speedup across multi-fpga for real-time dnn inference. ACM Trans. Embed. Comput. Syst., 18(5s), October 2019. ISSN 1539-9087. doi: 10.1145/3358192. URL https://doi.org/10.1145/3358192. ", + "bbox": [ + 174, + 296, + 826, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. In Proceedings of the TwentySecond International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS’17, pp. 615–629, New York, NY, USA, 2017. Association for Computing Machinery. ISBN 9781450344654. doi: 10.1145/3037697.3037698. URL https://doi.org/10.1145/3037697. 3037698. ", + "bbox": [ + 174, + 344, + 826, + 421 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yigitcan Kaya, Sanghyun Hong, and Tudor Dumitra¸s. Shallow-Deep Networks: Understanding and mitigating ˘ network overthinking. In Proceedings of the 2019 International Conference on Machine Learning (ICML), Long Beach, CA, Jun 2019. ", + "bbox": [ + 176, + 430, + 823, + 469 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning Multiple Layers of Features from Tiny Images. 2009. ", + "bbox": [ + 171, + 478, + 790, + 493 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. ", + "bbox": [ + 171, + 502, + 823, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236, 2016. ", + "bbox": [ + 174, + 537, + 823, + 564 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "M. Lecuyer, V. Atlidakis, R. Geambasu, D. Hsu, and S. Jana. Certified robustness to adversarial examples with differential privacy. In 2019 IEEE Symposium on Security and Privacy (SP), pp. 656–672, 2019. ", + "bbox": [ + 173, + 573, + 823, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "En Li, Zhi Zhou, and Xu Chen. Edge intelligence: On-demand deep learning model co-inference with deviceedge synergy. In Proceedings of the 2018 Workshop on Mobile Edge Communications, MECOMM’18, pp. 31–36, New York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450359061. doi: 10.1145/3229556.3229562. URL https://doi.org/10.1145/3229556.3229562. ", + "bbox": [ + 174, + 608, + 826, + 660 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. CoRR, abs/1608.08710, 2016. URL http://arxiv.org/abs/1608.08710. ", + "bbox": [ + 169, + 669, + 823, + 696 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. ", + "bbox": [ + 173, + 704, + 823, + 744 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. URL https://openreview. net/forum?id $\\equiv$ Sys6GJqxl. ", + "bbox": [ + 174, + 752, + 826, + 804 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. ", + "bbox": [ + 169, + 814, + 823, + 840 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=rJzIBfZAb. ", + "bbox": [ + 174, + 849, + 826, + 888 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. ", + "bbox": [ + 173, + 897, + 826, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "N. Papernot, P. McDaniel, S. Jha, M. Fredrikson, Z. B. Celik, and A. Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy (EuroS P), pp. 372–387, 2016. ", + "bbox": [ + 173, + 103, + 825, + 143 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pp. 506–519, 2017. ", + "bbox": [ + 176, + 154, + 821, + 194 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Se Jin Park, Murali Subramaniyam, Seoung Eun Kim, Seunghee Hong, Joo Hyeong Lee, Chan Min Jo, and Youngseob Seo. Development of the elderly healthcare monitoring system with iot. In Advances in Human Factors and Ergonomics in Healthcare, pp. 309–315. Springer, 2017. ", + "bbox": [ + 176, + 204, + 823, + 243 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jérôme Rony, Luiz G Hafemann, Luiz S Oliveira, Ismail Ben Ayed, Robert Sabourin, and Eric Granger. Decoupling direction and norm for efficient gradient-based l2 adversarial attacks and defenses. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4322–4330, 2019. ", + "bbox": [ + 174, + 252, + 823, + 292 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition, 2014. ", + "bbox": [ + 173, + 303, + 825, + 330 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ rJUYGxbCW. ", + "bbox": [ + 174, + 339, + 823, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C. Szegedy, Wei Liu, Yangqing Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9, 2015. ", + "bbox": [ + 174, + 390, + 823, + 429 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. URL http://arxiv.org/abs/1312.6199. ", + "bbox": [ + 174, + 439, + 823, + 478 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 6105–6114, Long Beach, California, USA, 09–15 Jun 2019. PMLR. URL http://proceedings.mlr.press/v97/tan19a. html. ", + "bbox": [ + 173, + 488, + 826, + 553 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ben Taylor, Vicent Sanz Marco, Willy Wolff, Yehia Elkhatib, and Zheng Wang. Adaptive deep learning model selection on embedded systems. ACM SIGPLAN Notices, 53(6):31–43, 2018. ", + "bbox": [ + 173, + 564, + 823, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "S. Teerapittayanon, B. McDanel, and H. T. Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pp. 2464–2469, 2016. ", + "bbox": [ + 174, + 601, + 823, + 627 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tiny. Tiny ImageNet Visual Recognition Challenge. http://tiny-imagenet.herokuapp.com/. Accessed: 2020-09-28. ", + "bbox": [ + 174, + 637, + 823, + 665 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Florian Tramèr and Dan Boneh. Adversarial training and robustness for multiple perturbations. In Advances in Neural Information Processing Systems, pp. 5866–5876, 2019. ", + "bbox": [ + 173, + 674, + 823, + 702 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017a. ", + "bbox": [ + 173, + 712, + 823, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017b. ", + "bbox": [ + 173, + 750, + 823, + 776 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E. Gonzalez. SkipNet: Learning Dynamic Routing in Convolutional Networks. In The European Conference on Computer Vision (ECCV), September 2018. ", + "bbox": [ + 173, + 786, + 825, + 813 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. arXiv preprint arXiv:1704.01155, 2017. ", + "bbox": [ + 173, + 823, + 823, + 849 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. arXiv preprint arXiv:1905.11971, 2019. ", + "bbox": [ + 173, + 859, + 823, + 887 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? In Advances in neural information processing systems, pp. 3320–3328, 2014. ", + "bbox": [ + 171, + 897, + 825, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Li Zhou, Hao Wen, Radu Teodorescu, and David H.C. Du. Distributing deep neural networks with containerized partitions at the edge. In 2nd USENIX Workshop on Hot Topics in Edge Computing (HotEdge 19), Renton, WA, July 2019. USENIX Association. URL https://www.usenix.org/conference/hotedge19/ presentation/zhou. ", + "bbox": [ + 174, + 103, + 825, + 155 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Wangchunshu Zhou, Canwen Xu, Tao Ge, Julian McAuley, Ke Xu, and Furu Wei. BERT Loses Patience: Fast and Robust Inference with Early Exit. In Advances in Neural Information Processing Systems, 2020. ", + "bbox": [ + 174, + 165, + 823, + 190 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A MOTIVATING EXAMPLES ", + "text_level": 1, + "bbox": [ + 178, + 102, + 416, + 118 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Here, we discuss two exemplary scenarios where an adversary can exploit the slowdown attacks. ", + "bbox": [ + 171, + 132, + 803, + 147 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "• (Case 1) Attacks on cloud-based IoT applications. In most cases, cloud-based IoT applications, such as Apple Siri, Google Now, or Microsoft Cortana, run their DNN inferences in the cloud. This cloud-only approach puts all the computational burden on cloud servers and increases the communications between the servers and IoT devices. In consequence, recent work (Kang et al., 2017; Li et al., 2018; Zhou et al., 2019) utilizes multi-exit architectures for bringing computationally expensive models, e.g. language models (Zhou et al., 2020; Hou et al., 2020), in the cloud to IoT (or mobile) devices. They split a multi-exit model into two partitions and deploy each of them to a server and IoT devices, respectively. Under this scheme, the cloud server only takes care of complex inputs that the shallow partition cannot correctly classify at the edge. As a result, one can reduce the computations in the cloud and decrease communications between the cloud and edge. ", + "bbox": [ + 174, + 159, + 826, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "On the other hand, our adversary, by applying human-imperceptible perturbations, can convert simple inputs into complex inputs. These adversarial inputs will bypass early-exits and, as a result, reduce (or even offset) the computational and communication savings provided by prior work. ", + "bbox": [ + 183, + 299, + 825, + 340 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Here, a defender may deploy DoS defenses such as firewalls or rate-limiting. In this setting, the attacker may not cause DoS because defenses keep the communications between the server and IoT devices under a certain-level. Nevertheless, the attacker still increases: (i) the computations at the edge (by making inputs skip early-exits) and (ii) the number of samples that cloud servers process. Recall that a VGG-16 SDN model classifies $90 \\%$ of clean CIFAR-10 instances correctly at the first exit. If the adversarial examples crafted by the attacker bypass only the first exit, one can easily increase the computations on IoT devices and make them send requests to the cloud. ", + "bbox": [ + 184, + 343, + 825, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "• (Case 2) Attacks on real-time DNN inference for resource- and time-constrained scenarios. Recent work on the real-time systems (Hu et al., 2019; Jiang et al., 2019) harnesses multi-exit architectures and model partitioning as a solution to optimize real-time DNN inference for resourceand time-constrained scenarios. Hu et al. (2019) showed a real-world prototype of an optimal model partitioning, which is based on a self-driving car video dataset, can improve latency and throughput of partitioned models on the cloud and edge by $6 . 5 – 1 4 \\times$ , respectively. ", + "bbox": [ + 174, + 444, + 826, + 527 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "However, the prior work does not consider the danger of slowdown attacks; our threat model has not been discussed before in the literature. Our results in Sec 5 suggest that slowdown can be induced adversarially, potentially violating real-time guarantees. For example, our attacker can force partitioned models on the cloud and edge to use maximal computations for inference. Further, the same adversarial examples also require the inference results from the model running on the cloud, which potentially increases the response time of the edge devices by $1 . 5 – 5 \\times$ . Our work showed that multi-exit architectures should be used with caution in real-time systems. ", + "bbox": [ + 179, + 529, + 825, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 176, + 645, + 379, + 661 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In our experiments, we use the following hyperparameters to craft adversarial perturbations. ", + "bbox": [ + 171, + 676, + 774, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "$\\ell _ { \\infty }$ -based DeepSloth. We find that $\\ell _ { \\infty }$ -based DeepSloth does not require careful tuning. For the standard attack, we set the total number of iterations to 30 and the step size to $\\alpha = 0 . 0 0 2$ . For the modified attacks for hurting or preserving the accuracy, we set the total number of iterations to 75 and the step size to $\\alpha = 0 . 0 0 1$ . We compute the standard perturbations using the entire $1 0 \\mathrm { k }$ test-set samples in CIFAR-10 and Tiny Imagenet. For the universal variants, we set the total number of iterations to 12 and reduce the initial step size of $\\alpha = 0 . 0 0 5$ by a factor of 10 every 4 iterations. To compute a universal perturbation, we use randomly chosen 250 (CIFAR-10) and 200 (Tiny Imagenet) training samples. ", + "bbox": [ + 174, + 698, + 825, + 810 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "$\\ell _ { 2 }$ -based DeepSloth. For both the standard and universal attacks, we set the total number of iterations to 550 and the step size $\\gamma$ to 0.1. Our initial perturbation has the $\\ell _ { 2 }$ -norm of 1.0. Here, we use the same number of samples for crafting the standard and universal perturbations as the $\\ell _ { \\infty }$ -based attacks. ", + "bbox": [ + 176, + 818, + 823, + 859 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "$\\ell _ { 1 }$ -based DeepSloth. For our standard $\\ell _ { 1 }$ -based DeepSloth, we set the total number of iterations to 250, the step size $\\alpha$ to 0.5, and the gradient sparsity to 99. For the universal variants, we reduce the total number of iterations to 100 and set the gradient sparsity to 90. Other hyperparameters remain the same. We use the same number of samples as the $\\ell _ { \\infty }$ -based attacks, to craft the perturbations. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C EMPIRICAL EVALUATION OF $\\ell _ { 1 }$ AND $\\ell _ { 2 }$ DEEPSLOTH ", + "text_level": 1, + "bbox": [ + 174, + 102, + 635, + 118 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Table 4 and Table 5 shows the effectiveness of $\\ell _ { 1 }$ -based and $\\ell _ { 2 }$ -based DeepSloth attacks, respectively. ", + "bbox": [ + 173, + 140, + 825, + 155 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/6b9f5d0671919db5ee08c6a0cfa8dee4ae1967265bbd425b6363d71e75af115c.jpg", + "table_caption": [ + "Table 4: The effectiveness of $\\ell _ { 1 }$ DeepSloth. ‘ $R \\mathrm { A D } { < } 5 , 1 5 \\%$ ’ columns list the results in each early-exit setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal attack’s results are an average of 10 classes. ‘TI’ is Tiny Imagenet and $\\mathbf { \\dot { C } } 1 0 ^ { \\circ }$ is CIFAR-10. " + ], + "table_footnote": [], + "table_body": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 /89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 / 57%0.51 /52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.36 / 51%0.35 / 51%0.12 /36%0.34 /45%0.18 / 41%0.49 / 53%
TI0.23 /37%0.51 / 40%0.08 /22%0.15 / 25%0.08 /33%0.19 /35%
UNIVERSAL DEEPSLOTH
C100.89 /83%0.89 /83%0.75 /85%0.88 / 75%0.82 /85%0.92 /77%
TI0.64 / 55%0.83/50%0.38 /57%0.51 / 52%0.41 / 57%0.59 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88/ 73%0.88 / 73%0.69 / 78%0.86 /67%0.76 /74%0.89 / 65%
TI0.64 /54%0.83 /49%0.39 / 59%0.50 / 58%0.41 / 60%0.58 / 53%
", + "bbox": [ + 173, + 233, + 823, + 467 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/e75aec7697dc259a45f3e17c19b3d6b5b0ffa1634f7b75a2b00810815ceb6fae.jpg", + "table_caption": [ + "Table 5: The effectiveness of $\\ell _ { 2 }$ DeepSloth. $\\cdot \\mathrm { R A D } { < } 5 , 1 5 \\%$ ’ columns list the results in each early-exit setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal attack’s results are an average of 10 classes. ‘TI’ is Tiny Imagenet and $\\mathbf { \\dot { C } } 1 0 ^ { \\circ }$ is CIFAR-10. " + ], + "table_footnote": [], + "table_body": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 / 89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 / 50%0.39 / 57%0.51 / 52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.52 / 64%0.52 /64%0.22 /60%0.45 / 62%0.23 /46%0.48 / 55%
TI0.24 /42%0.52 / 44%0.13 /35%0.21 /36%0.12 /38%0.25 /40%
UNIVERSAL DEEPSLOTH
C100.89 /81%0.89 /81%0.75 /87%0.88 / 76%0.81/84%0.92 /76%
TI0.63 / 54%0.82 /48%0.38 /56%0.51 / 52%0.41 / 56%0.58 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88 /73%0.88 / 73%0.71 /81%0.86 / 70%0.76 /76%0.89 /66%
TI0.64 / 53%0.83 /49%0.38 /57%0.50 / 57%0.41 /58%0.58 / 53%
", + "bbox": [ + 173, + 570, + 823, + 804 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Our results show that the $\\ell _ { 1 } \\cdot$ - and $\\ell _ { 2 }$ -based attacks are less effective than the $\\ell _ { \\infty }$ -based attacks. In contrast to the $\\ell _ { \\infty }$ -based attacks that eliminate the efficacy of victim multi-exit models, the $\\ell _ { 1 }$ - and $\\ell _ { 2 }$ -based attacks reduce the efficacy of the same models by $0 . 2 4 { \\sim } 0 . 6 5$ . Besides, the accuracy drops caused by $\\ell _ { 1 } \\cdot$ - and $\\ell _ { 2 }$ -based attacks are in $6 \\sim 2 1 \\%$ , smaller than that of $\\ell _ { \\infty }$ -based DeepSloth $( 7 5 \\sim 9 9 \\% )$ . Moreover, we see that the universal variants of $\\ell _ { 1 }$ - and $\\ell _ { 2 }$ -based attacks can barely reduce the efficacy of multi-exit models—they decrease the efficacy up to 0.08 and the accuracy by $12 \\%$ . ", + "bbox": [ + 173, + 839, + 826, + 924 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D EMPIRICAL EVALUATION OF DEEPSLOTH ON RESNET56 ", + "text_level": 1, + "bbox": [ + 174, + 102, + 679, + 118 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Table 6 shows the the effectiveness of our DeepSloth attacks on ResNet56-base models. ", + "bbox": [ + 176, + 142, + 746, + 156 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/5db15f9dd474b600c6dace3769bf9c0f1f48b737c914f14f9a26cce1d52ad611.jpg", + "table_caption": [ + "Table 6: The effectiveness of DeepSloth on the ResNet-based models. $\\cdot \\mathrm { R A D } { < } 5 , 1 5 \\%$ ’ columns list the results in each early-exit setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal attack’s results are an average of 10 classes. " + ], + "table_footnote": [], + "table_body": "
NETWORKRESNET(lo)RESNET (l1)RESNET(2)
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.52 /87%0.69 /80%0.52/87%0.69 / 80%0.51 /87%0.69 / 80%
TI0.25 / 51%0.39 /46%0.25 /51%0.39 /46%0.25 /51%0.39 / 46%
DEEPSLOTH
C100.00 /19%0.01 /19%0.05 /43%0.18 /47%0.06 /45%0.17 /48%
TI0.00 /7%0.01 / 7%0.04 /27%0.10 /28%0.05 /34%0.13 /35%
UNIVERSAL DEEPSLOTH
C100.35 / 63%0.59 /60%0.49 /84%0.68 /75%0.48 /85%0.67 /76%
TI0.25 / 25%0.34 / 37%0.25 / 51%0.39 / 46%0.25 / 51%0.38 /46%
CLASS-UNIVERSAL DEEPSLOTH
C100.23 / 33%0.48 / 29%0.39 / 70%0.60 / 61%0.39 / 71%0.60 / 61%
TI0.11 /21%0.23 /18%0.23 / 51%0.36 /46%0.23 / 50%0.36 /46%
", + "bbox": [ + 173, + 237, + 823, + 472 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Our results show that ResNet56-based models are vulnerable to all the $\\ell _ { \\infty }$ , $\\ell _ { 2 }$ , and $\\ell _ { 1 }$ -based DeepSloth attacks. Using our $\\ell _ { \\infty }$ -based DeepSloth, the attacker can reduce the efficacy of the victim models to $0 . 0 0 { \\sim } 0 . 0 1$ and the accuracy by $3 9 \\sim 6 8 \\%$ . Besides, the $\\ell _ { 2 }$ , and $\\ell _ { 1 }$ -based attacks also decrease the efficacy to $0 . 0 4 { \\sim } 0 . 1 8$ and the accuracy by $1 1 { \\sim } 4 4 \\%$ . Compared to the results on MSDNet, VGG16, and MobileNet in Table 4 and 5, the same attacks are more effective. The universal variants decrease the efficacy up to 0.21 and the accuracy up to $24 \\%$ . In particular, the $\\ell _ { 2 }$ , and $\\ell _ { 1 }$ -based attacks (on CIFAR-10 models) are effective than the same attacks on MSDNet, VGG16, and MobileNet models. ", + "bbox": [ + 173, + 508, + 826, + 607 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "E COST OF CRAFTING DEEPSLOTH SAMPLES ", + "text_level": 1, + "bbox": [ + 174, + 641, + 566, + 657 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "In Table 7, we compare the cost of DeepSloth with other attack algorithms on a VGG16-based CIFAR10 model—executed on a single Nvidia Tesla-V100 GPU. For the universal DeepSloth, we measure the execution time for crafting a perturbation using one batch (250 samples) of the training set. For the other attacks, we measure the time for perturbing the whole test set of CIFAR-10. Our DeepSloth takes roughly the same time as the PGD and PGD-avg attacks and significantly less time than the PGD-max attack. Our ", + "bbox": [ + 174, + 681, + 529, + 825 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/f4af52396b56906f7cf4187b9615e7a86dffd1aa63bb7f7499d79e474d5187d9.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7: Time it takes to craft attacks. ", + "universal DeepSloth takes only 2 seconds ( $1 0 \\mathrm { x }$ faster than DeepSloth) as it only uses 250 samples. " + ], + "table_body": "
ATTACKSTIME (SEC.)
PGD-2038
PGD-20 (AVG.)48
PGD-20 (MAX.)475
DEEPSLOTH44
UNIVERSAL DEEPSLOTH2
", + "bbox": [ + 545, + 683, + 823, + 791 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "F ADVERSARIAL EXAMPLES FROM STANDARD ATTACKS AND DEEPSLOTH ", + "text_level": 1, + "bbox": [ + 173, + 869, + 808, + 886 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "In Figure 4, we visualize the adversarial examples from the PGD, UAP and our DeepSloth attacks. ", + "bbox": [ + 171, + 909, + 815, + 924 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/0bdcaea5dbd66848101c989036785e7979050485b14922c9d380c77e29ee961e.jpg", + "image_caption": [ + "Figure 4: Adversarial examples from the standard and our DeepSloth attacks. The leftmost column shows the clean images. In the next four columns, we show adversarial examples from PGD, PGD (avg.), PGD (max.), and UAP attacks, respectively. The last four columns include adversarial examples from the three variants of DeepSloth. Each row corresponds to each sample, and the last row contains the average $\\ell _ { \\mathrm { i n f } }$ -norm of the perturbations over the eight samples in each attack. " + ], + "image_footnote": [], + "bbox": [ + 173, + 98, + 812, + 583 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/parse/train/9xC2tWEwBD/9xC2tWEwBD_middle.json b/parse/train/9xC2tWEwBD/9xC2tWEwBD_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..3fc3fcd410558d3982fb92235466d8a58e716e35 --- /dev/null +++ b/parse/train/9xC2tWEwBD/9xC2tWEwBD_middle.json @@ -0,0 +1,47807 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 488, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 447, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 447, + 95 + ], + "score": 1.0, + "content": "A PANDA? NO, IT’S A SLOTH: SLOWDOWN ATTACKS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 490, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 490, + 116 + ], + "score": 1.0, + "content": "ON ADAPTIVE MULTI-EXIT NEURAL NETWORK INFERENCE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 133, + 431, + 145 + ], + "lines": [ + { + "bbox": [ + 111, + 131, + 433, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 131, + 433, + 149 + ], + "score": 1.0, + "content": "Sanghyun Hong∗, Yigitcan Kaya ˇ ∗, Ionut,-Vlad Modoranu†, Tudor Dumitras,", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 114, + 147, + 307, + 169 + ], + "lines": [ + { + "bbox": [ + 113, + 146, + 290, + 159 + ], + "spans": [ + { + "bbox": [ + 113, + 146, + 290, + 159 + ], + "score": 1.0, + "content": "University of Maryland, College Park, USA", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 157, + 308, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 308, + 171 + ], + "score": 1.0, + "content": "†Alexandru Ioan Cuza University, Ias,i, Romania", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 113, + 172, + 361, + 193 + ], + "lines": [ + { + "bbox": [ + 113, + 172, + 323, + 183 + ], + "spans": [ + { + "bbox": [ + 113, + 172, + 323, + 183 + ], + "score": 1.0, + "content": "shhong@cs.umd.edu, yigitcan@cs.umd.edu,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 183, + 362, + 194 + ], + "spans": [ + { + "bbox": [ + 111, + 183, + 362, + 194 + ], + "score": 1.0, + "content": "modoranu.ionut.vlad@hotmail.com, tudor@umd.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 278, + 222, + 333, + 234 + ], + "lines": [ + { + "bbox": [ + 276, + 222, + 335, + 235 + ], + "spans": [ + { + "bbox": [ + 276, + 222, + 335, + 235 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 142, + 246, + 469, + 499 + ], + "lines": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "score": 1.0, + "content": "Recent increases in the computational demands of deep neural networks (DNNs),", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 258, + 470, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 258, + 470, + 270 + ], + "score": 1.0, + "content": "combined with the observation that most input samples require only simple models,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 269, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 269, + 470, + 280 + ], + "score": 1.0, + "content": "have sparked interest in input-adaptive multi-exit architectures, such as MSDNets", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 280, + 470, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 291 + ], + "score": 1.0, + "content": "or Shallow-Deep Networks. These architectures enable faster inferences and could", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 470, + 304 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 304 + ], + "score": 1.0, + "content": "bring DNNs to low-power devices, e.g., in the Internet of Things (IoT). However,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 301, + 470, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 301, + 470, + 315 + ], + "score": 1.0, + "content": "it is unknown if the computational savings provided by this approach are robust", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 313, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 469, + 325 + ], + "score": 1.0, + "content": "against adversarial pressure. In particular, an adversary may aim to slowdown", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "score": 1.0, + "content": "adaptive DNNs by increasing their average inference time—a threat analogous", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 334, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 470, + 347 + ], + "score": 1.0, + "content": "to the denial-of-service attacks from the Internet. In this paper, we conduct a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 345, + 470, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 470, + 358 + ], + "score": 1.0, + "content": "systematic evaluation of this threat by experimenting with three generic multi-exit", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 357, + 470, + 368 + ], + "spans": [ + { + "bbox": [ + 141, + 357, + 470, + 368 + ], + "score": 1.0, + "content": "DNNs (based on VGG16, MobileNet, and ResNet56) and a custom multi-exit", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 469, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 469, + 380 + ], + "score": 1.0, + "content": "architecture, on two popular image classification benchmarks (CIFAR-10 and Tiny", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 378, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 469, + 392 + ], + "score": 1.0, + "content": "ImageNet). To this end, we show that adversarial example-crafting techniques", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 388, + 469, + 403 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 469, + 403 + ], + "score": 1.0, + "content": "can be modified to cause slowdown, and we propose a metric for comparing", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 401, + 470, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 401, + 470, + 412 + ], + "score": 1.0, + "content": "their impact on different architectures. We show that a slowdown attack reduces", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 411, + 469, + 424 + ], + "spans": [ + { + "bbox": [ + 141, + 411, + 293, + 424 + ], + "score": 1.0, + "content": "the efficacy of multi-exit DNNs by", + "type": "text" + }, + { + "bbox": [ + 293, + 411, + 333, + 422 + ], + "score": 0.88, + "content": "9 0 { - } 1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 411, + 469, + 424 + ], + "score": 1.0, + "content": ", and it amplifies the latency by", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 422, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 142, + 422, + 174, + 433 + ], + "score": 0.88, + "content": "1 . 5 – 5 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 422, + 470, + 434 + ], + "score": 1.0, + "content": "in a typical IoT deployment. We also show that it is possible to craft", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 434, + 469, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 434, + 469, + 444 + ], + "score": 1.0, + "content": "universal, reusable perturbations and that the attack can be effective in realistic", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 444, + 470, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 444, + 470, + 456 + ], + "score": 1.0, + "content": "black-box scenarios, where the attacker has limited knowledge about the victim.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 455, + 470, + 467 + ], + "spans": [ + { + "bbox": [ + 141, + 455, + 470, + 467 + ], + "score": 1.0, + "content": "Finally, we show that adversarial training provides limited protection against", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 466, + 469, + 478 + ], + "spans": [ + { + "bbox": [ + 141, + 466, + 469, + 478 + ], + "score": 1.0, + "content": "slowdowns. These results suggest that further research is needed for defending", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 477, + 470, + 490 + ], + "spans": [ + { + "bbox": [ + 141, + 477, + 470, + 490 + ], + "score": 1.0, + "content": "multi-exit architectures against this emerging threat. Our code is available at", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 489, + 374, + 500 + ], + "spans": [ + { + "bbox": [ + 141, + 489, + 374, + 500 + ], + "score": 1.0, + "content": "https://github.com/sanghyun-hong/deepsloth.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 519, + 206, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 208, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 208, + 535 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "The inference-time computational demands of deep neural networks (DNNs) are increasing, owing to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 554, + 507, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 507, + 569 + ], + "score": 1.0, + "content": "the “going deeper\" (Szegedy et al., 2015) strategy for improving accuracy: as a DNN gets deeper,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "it progressively gains the ability to learn higher-level, complex representations. This strategy has", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "enabled breakthroughs in many tasks, such as image classification (Krizhevsky et al., 2012) or speech", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 465, + 600 + ], + "score": 1.0, + "content": "recognition (Hinton et al., 2012), at the price of costly inferences. For instance, with", + "type": "text" + }, + { + "bbox": [ + 465, + 588, + 480, + 599 + ], + "score": 0.87, + "content": "4 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "more", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 486, + 612 + ], + "score": 1.0, + "content": "inference cost, a 56-layer ResNet (He et al., 2016) improved the Top-1 accuracy on ImageNet by", + "type": "text" + }, + { + "bbox": [ + 486, + 599, + 505, + 609 + ], + "score": 0.85, + "content": "19 \\%", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "over the 8-layer AlexNet. This trend continued with the 57-layer state-of-the-art EfficientNet (Tan &", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 620, + 452, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 264, + 633 + ], + "score": 1.0, + "content": "Le, 2019): it improved the accuracy by", + "type": "text" + }, + { + "bbox": [ + 264, + 621, + 284, + 631 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 620, + 359, + 633 + ], + "score": 1.0, + "content": "over ResNet, with", + "type": "text" + }, + { + "bbox": [ + 359, + 621, + 374, + 631 + ], + "score": 0.86, + "content": "9 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 620, + 452, + 633 + ], + "score": 1.0, + "content": "costlier inferences.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 715 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "score": 1.0, + "content": "The accuracy improvements stem from the fact that the deeper networks fix the mistakes of the shallow", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "ones (Huang et al., 2018). This implies that some samples, which are already correctly classified by", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "shallow networks, do not necessitate the extra complexity. This observation has motivated research", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 672, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 506, + 683 + ], + "score": 1.0, + "content": "on input-adaptive mechanisms, in particular, multi-exit architectures (Teerapittayanon et al., 2016;", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "Huang et al., 2018; Kaya et al., 2019; Hu et al., 2020). Multi-exit architectures save computation", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "by making input-specific decisions about bypassing the remaining layers, once the model becomes", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 704, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 505, + 716 + ], + "score": 1.0, + "content": "confident, and are orthogonal to techniques that achieve savings by permanently modifying the", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 226, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 719, + 227, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 719, + 227, + 734 + ], + "score": 1.0, + "content": "∗Authors contributed equally.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 488, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 447, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 447, + 95 + ], + "score": 1.0, + "content": "A PANDA? NO, IT’S A SLOTH: SLOWDOWN ATTACKS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 490, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 490, + 116 + ], + "score": 1.0, + "content": "ON ADAPTIVE MULTI-EXIT NEURAL NETWORK INFERENCE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 133, + 431, + 145 + ], + "lines": [ + { + "bbox": [ + 111, + 131, + 433, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 131, + 433, + 149 + ], + "score": 1.0, + "content": "Sanghyun Hong∗, Yigitcan Kaya ˇ ∗, Ionut,-Vlad Modoranu†, Tudor Dumitras,", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 111, + 131, + 433, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 114, + 147, + 307, + 169 + ], + "lines": [ + { + "bbox": [ + 113, + 146, + 290, + 159 + ], + "spans": [ + { + "bbox": [ + 113, + 146, + 290, + 159 + ], + "score": 1.0, + "content": "University of Maryland, College Park, USA", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 157, + 308, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 308, + 171 + ], + "score": 1.0, + "content": "†Alexandru Ioan Cuza University, Ias,i, Romania", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 112, + 146, + 308, + 171 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 172, + 361, + 193 + ], + "lines": [ + { + "bbox": [ + 113, + 172, + 323, + 183 + ], + "spans": [ + { + "bbox": [ + 113, + 172, + 323, + 183 + ], + "score": 1.0, + "content": "shhong@cs.umd.edu, yigitcan@cs.umd.edu,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 183, + 362, + 194 + ], + "spans": [ + { + "bbox": [ + 111, + 183, + 362, + 194 + ], + "score": 1.0, + "content": "modoranu.ionut.vlad@hotmail.com, tudor@umd.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 111, + 172, + 362, + 194 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 222, + 333, + 234 + ], + "lines": [ + { + "bbox": [ + 276, + 222, + 335, + 235 + ], + "spans": [ + { + "bbox": [ + 276, + 222, + 335, + 235 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 142, + 246, + 469, + 499 + ], + "lines": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "score": 1.0, + "content": "Recent increases in the computational demands of deep neural networks (DNNs),", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 258, + 470, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 258, + 470, + 270 + ], + "score": 1.0, + "content": "combined with the observation that most input samples require only simple models,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 269, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 269, + 470, + 280 + ], + "score": 1.0, + "content": "have sparked interest in input-adaptive multi-exit architectures, such as MSDNets", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 280, + 470, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 291 + ], + "score": 1.0, + "content": "or Shallow-Deep Networks. These architectures enable faster inferences and could", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 470, + 304 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 304 + ], + "score": 1.0, + "content": "bring DNNs to low-power devices, e.g., in the Internet of Things (IoT). However,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 301, + 470, + 315 + ], + "spans": [ + { + "bbox": [ + 141, + 301, + 470, + 315 + ], + "score": 1.0, + "content": "it is unknown if the computational savings provided by this approach are robust", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 313, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 469, + 325 + ], + "score": 1.0, + "content": "against adversarial pressure. In particular, an adversary may aim to slowdown", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 324, + 469, + 335 + ], + "score": 1.0, + "content": "adaptive DNNs by increasing their average inference time—a threat analogous", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 334, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 470, + 347 + ], + "score": 1.0, + "content": "to the denial-of-service attacks from the Internet. In this paper, we conduct a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 345, + 470, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 470, + 358 + ], + "score": 1.0, + "content": "systematic evaluation of this threat by experimenting with three generic multi-exit", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 357, + 470, + 368 + ], + "spans": [ + { + "bbox": [ + 141, + 357, + 470, + 368 + ], + "score": 1.0, + "content": "DNNs (based on VGG16, MobileNet, and ResNet56) and a custom multi-exit", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 469, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 469, + 380 + ], + "score": 1.0, + "content": "architecture, on two popular image classification benchmarks (CIFAR-10 and Tiny", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 378, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 469, + 392 + ], + "score": 1.0, + "content": "ImageNet). To this end, we show that adversarial example-crafting techniques", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 388, + 469, + 403 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 469, + 403 + ], + "score": 1.0, + "content": "can be modified to cause slowdown, and we propose a metric for comparing", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 401, + 470, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 401, + 470, + 412 + ], + "score": 1.0, + "content": "their impact on different architectures. We show that a slowdown attack reduces", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 411, + 469, + 424 + ], + "spans": [ + { + "bbox": [ + 141, + 411, + 293, + 424 + ], + "score": 1.0, + "content": "the efficacy of multi-exit DNNs by", + "type": "text" + }, + { + "bbox": [ + 293, + 411, + 333, + 422 + ], + "score": 0.88, + "content": "9 0 { - } 1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 411, + 469, + 424 + ], + "score": 1.0, + "content": ", and it amplifies the latency by", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 422, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 142, + 422, + 174, + 433 + ], + "score": 0.88, + "content": "1 . 5 – 5 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 422, + 470, + 434 + ], + "score": 1.0, + "content": "in a typical IoT deployment. We also show that it is possible to craft", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 434, + 469, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 434, + 469, + 444 + ], + "score": 1.0, + "content": "universal, reusable perturbations and that the attack can be effective in realistic", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 444, + 470, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 444, + 470, + 456 + ], + "score": 1.0, + "content": "black-box scenarios, where the attacker has limited knowledge about the victim.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 455, + 470, + 467 + ], + "spans": [ + { + "bbox": [ + 141, + 455, + 470, + 467 + ], + "score": 1.0, + "content": "Finally, we show that adversarial training provides limited protection against", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 466, + 469, + 478 + ], + "spans": [ + { + "bbox": [ + 141, + 466, + 469, + 478 + ], + "score": 1.0, + "content": "slowdowns. These results suggest that further research is needed for defending", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 477, + 470, + 490 + ], + "spans": [ + { + "bbox": [ + 141, + 477, + 470, + 490 + ], + "score": 1.0, + "content": "multi-exit architectures against this emerging threat. Our code is available at", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 489, + 374, + 500 + ], + "spans": [ + { + "bbox": [ + 141, + 489, + 374, + 500 + ], + "score": 1.0, + "content": "https://github.com/sanghyun-hong/deepsloth.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 19, + "bbox_fs": [ + 141, + 246, + 470, + 500 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 519, + 206, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 208, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 208, + 535 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "The inference-time computational demands of deep neural networks (DNNs) are increasing, owing to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 554, + 507, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 507, + 569 + ], + "score": 1.0, + "content": "the “going deeper\" (Szegedy et al., 2015) strategy for improving accuracy: as a DNN gets deeper,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "it progressively gains the ability to learn higher-level, complex representations. This strategy has", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "enabled breakthroughs in many tasks, such as image classification (Krizhevsky et al., 2012) or speech", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 465, + 600 + ], + "score": 1.0, + "content": "recognition (Hinton et al., 2012), at the price of costly inferences. For instance, with", + "type": "text" + }, + { + "bbox": [ + 465, + 588, + 480, + 599 + ], + "score": 0.87, + "content": "4 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "more", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 486, + 612 + ], + "score": 1.0, + "content": "inference cost, a 56-layer ResNet (He et al., 2016) improved the Top-1 accuracy on ImageNet by", + "type": "text" + }, + { + "bbox": [ + 486, + 599, + 505, + 609 + ], + "score": 0.85, + "content": "19 \\%", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "over the 8-layer AlexNet. This trend continued with the 57-layer state-of-the-art EfficientNet (Tan &", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 620, + 452, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 264, + 633 + ], + "score": 1.0, + "content": "Le, 2019): it improved the accuracy by", + "type": "text" + }, + { + "bbox": [ + 264, + 621, + 284, + 631 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 620, + 359, + 633 + ], + "score": 1.0, + "content": "over ResNet, with", + "type": "text" + }, + { + "bbox": [ + 359, + 621, + 374, + 631 + ], + "score": 0.86, + "content": "9 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 620, + 452, + 633 + ], + "score": 1.0, + "content": "costlier inferences.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 543, + 507, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 715 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "score": 1.0, + "content": "The accuracy improvements stem from the fact that the deeper networks fix the mistakes of the shallow", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "ones (Huang et al., 2018). This implies that some samples, which are already correctly classified by", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "shallow networks, do not necessitate the extra complexity. This observation has motivated research", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 672, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 506, + 683 + ], + "score": 1.0, + "content": "on input-adaptive mechanisms, in particular, multi-exit architectures (Teerapittayanon et al., 2016;", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "Huang et al., 2018; Kaya et al., 2019; Hu et al., 2020). Multi-exit architectures save computation", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "by making input-specific decisions about bypassing the remaining layers, once the model becomes", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 704, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 505, + 716 + ], + "score": 1.0, + "content": "confident, and are orthogonal to techniques that achieve savings by permanently modifying the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "model (Li et al., 2016; Banner et al., 2018; Han et al., 2015; Taylor et al., 2018). Figure 1 illustrates", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "how a multi-exit model (Kaya et al., 2019), based on a standard VGG-16 architecture, correctly", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "classifies a selection of test images from ‘Tiny ImageNet’ before the final layer. We see that more", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "typical samples, which have more supporting examples in the training set, require less depth and,", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 219, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 219, + 139 + ], + "score": 1.0, + "content": "therefore, less computation.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 638, + 506, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "model (Li et al., 2016; Banner et al., 2018; Han et al., 2015; Taylor et al., 2018). Figure 1 illustrates", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "how a multi-exit model (Kaya et al., 2019), based on a standard VGG-16 architecture, correctly", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "classifies a selection of test images from ‘Tiny ImageNet’ before the final layer. We see that more", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "typical samples, which have more supporting examples in the training set, require less depth and,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 219, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 219, + 139 + ], + "score": 1.0, + "content": "therefore, less computation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 344, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 344, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 344, + 156 + ], + "score": 1.0, + "content": "It is unknown if the computational savings provided by", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 345, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 345, + 167 + ], + "score": 1.0, + "content": "multi-exit architectures are robust against adversarial pres-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 345, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 345, + 177 + ], + "score": 1.0, + "content": "sure. Prior research showed that DNNs are vulnerable to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 345, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 345, + 188 + ], + "score": 1.0, + "content": "a wide range of attacks, which involve imperceptible input", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 346, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 346, + 200 + ], + "score": 1.0, + "content": "perturbations (Szegedy et al., 2014; Goodfellow et al., 2015;", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 346, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 346, + 210 + ], + "score": 1.0, + "content": "Papernot et al., 2016; Hu et al., 2020). Considering that a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 345, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 345, + 221 + ], + "score": 1.0, + "content": "multi-exit model, on the worst-case input, does not provide", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 221, + 345, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 345, + 232 + ], + "score": 1.0, + "content": "any computational savings, we ask: Can the savings from", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 346, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 346, + 244 + ], + "score": 1.0, + "content": "multi-exit models be maliciously negated by input perturba-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 241, + 345, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 345, + 254 + ], + "score": 1.0, + "content": "tions? As some natural inputs do require the full depth of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 345, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 345, + 265 + ], + "score": 1.0, + "content": "the model, it may be possible to craft adversarial examples", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 345, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 345, + 276 + ], + "score": 1.0, + "content": "that delay the correct decision; it is unclear, however, how", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 346, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 346, + 287 + ], + "score": 1.0, + "content": "many inputs can be delayed with imperceptible perturba-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 345, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 345, + 298 + ], + "score": 1.0, + "content": "tions. Furthermore, it is unknown if universal versions of", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11.5 + }, + { + "type": "image", + "bbox": [ + 360, + 144, + 496, + 237 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 360, + 144, + 496, + 237 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 360, + 144, + 496, + 237 + ], + "spans": [ + { + "bbox": [ + 360, + 144, + 496, + 237 + ], + "score": 0.965, + "type": "image", + "image_path": "746d8efde4dde3f7f81419a6f62cb228c7561c4c925bbdc393202621421b32de.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 360, + 144, + 496, + 190.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 360, + 190.5, + 496, + 237.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 352, + 252, + 506, + 293 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 351, + 252, + 507, + 264 + ], + "spans": [ + { + "bbox": [ + 351, + 252, + 507, + 264 + ], + "score": 1.0, + "content": "Figure 1: Simple to complex inputs.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 351, + 262, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 351, + 262, + 505, + 273 + ], + "score": 1.0, + "content": "Some Tiny ImageNet images a VGG-16", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 351, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 351, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "model can correctly classify if computa-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 350, + 280, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 350, + 280, + 411, + 295 + ], + "score": 1.0, + "content": "tion stops at the", + "type": "text" + }, + { + "bbox": [ + 411, + 282, + 423, + 292 + ], + "score": 0.74, + "content": "1 ^ { s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 280, + 426, + 295 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 426, + 282, + 440, + 292 + ], + "score": 0.74, + "content": "5 ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 280, + 459, + 295 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 459, + 282, + 477, + 292 + ], + "score": 0.88, + "content": "1 4 ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 280, + 506, + 295 + ], + "score": 1.0, + "content": "layers.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + } + ], + "index": 21.0 + }, + { + "type": "text", + "bbox": [ + 107, + 297, + 504, + 319 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 506, + 310 + ], + "score": 1.0, + "content": "these adversarial examples exist, if the examples transfer across multi-exit architectures and datasets,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 307, + 457, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 457, + 320 + ], + "score": 1.0, + "content": "or if existing defenses (e.g. adversarial training) are effective against slowdown attacks.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 325, + 506, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 479, + 339 + ], + "score": 1.0, + "content": "Threat Model. We consider a new threat against DNNs, analogous to the denial-of-service", + "type": "text" + }, + { + "bbox": [ + 479, + 326, + 504, + 337 + ], + "score": 0.68, + "content": "( D o S )", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 335, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 350 + ], + "score": 1.0, + "content": "attacks that have been plaguing the Internet for decades. By imperceptibly perturbing the input to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 345, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 104, + 345, + 506, + 362 + ], + "score": 1.0, + "content": "trigger this worst-case, the adversary aims to slow down the inferences and increase the cost of using", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 358, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 506, + 371 + ], + "score": 1.0, + "content": "the DNN. This is an important threat for many practical applications, which impose strict limits on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 369, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 382 + ], + "score": 1.0, + "content": "the responsiveness and resource usage of DNN models (e.g. in the Internet-of-Things (Taylor et al.,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "2018)), because the adversary could push the victim outside these limits. For example, against a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "commercial image classification system, such as Clarifai.com, a slowdown attack might waste valuable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "computational resources. Against a model partitioning scheme, such as Big-Little (De Coninck et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "2015), it might introduce network latency by forcing excessive transmissions between local and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 422, + 507, + 438 + ], + "spans": [ + { + "bbox": [ + 104, + 422, + 507, + 438 + ], + "score": 1.0, + "content": "remote models. A slowdown attack aims to force the victim to do more work than the adversary, e.g.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "by amplifying the latency needed to process the sample or by crafting reusable perturbations. The", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "adversary may have to achieve this with incomplete information about the multi-exit architecture", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "targeted, the training data used by the victim or the classification task (see discussion in Appendix A).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 506, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "Our Contributions. To our best knowledge, we conduct the first study of the robustness of multi-exit", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 484, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 499 + ], + "score": 1.0, + "content": "architectures against adversarial slowdowns. To this end, we find that examples crafted by prior", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "evasion attacks (Madry et al., 2017; Hu et al., 2020) fail to bypass the victim model’s early exits, and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "we show that an adversary can adapt such attacks to the goal of model slowdown by modifying its", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "objective function. We call the resulting attack DeepSloth. We also propose an efficacy metric for", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "comparing slowdowns across different multi-exit architectures. We experiment with three generic", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 539, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 554 + ], + "score": 1.0, + "content": "multi-exit DNNs (based on VGG16, ResNet56 and MobileNet) (Kaya et al., 2019) and a specially-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "score": 1.0, + "content": "designed multi-exit architecture, MSDNets (Huang et al., 2018), on two popular image classification", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "benchmarks (CIFAR-10 and Tiny ImageNet). We find that DeepSloth reduces the efficacy of multi-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 164, + 586 + ], + "score": 1.0, + "content": "exit DNNs by", + "type": "text" + }, + { + "bbox": [ + 164, + 573, + 203, + 584 + ], + "score": 0.86, + "content": "9 0 { - } 1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 573, + 506, + 586 + ], + "score": 1.0, + "content": ", i.e., the perturbations render nearly all early exits ineffective. In a scenario", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 583, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 583, + 506, + 597 + ], + "score": 1.0, + "content": "typical for IoT deployments, where the model is partitioned between edge devices and the cloud, our", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 232, + 608 + ], + "score": 1.0, + "content": "attack amplifies the latency by", + "type": "text" + }, + { + "bbox": [ + 233, + 595, + 264, + 606 + ], + "score": 0.88, + "content": "1 . 5 – 5 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 595, + 506, + 608 + ], + "score": 1.0, + "content": ", negating the benefits of model partitioning. We also show", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "that it is possible to craft a universal DeepSloth perturbation, which can slow down the model on", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 617, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 473, + 629 + ], + "score": 1.0, + "content": "either all or a class of inputs. While more constrained, this attack still reduces the efficacy by", + "type": "text" + }, + { + "bbox": [ + 474, + 617, + 503, + 628 + ], + "score": 0.87, + "content": "5- 4 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 617, + 506, + 629 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 628, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 506, + 640 + ], + "score": 1.0, + "content": "Further, we observe that DeepSloth can be effective in some black-box scenarios, where the attacker", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "has limited knowledge about the victim. Finally, we show that a standard defense against adversarial", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 651, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 506, + 662 + ], + "score": 1.0, + "content": "samples—adversarial training—is inadequate against slowdowns. Our results suggest that further", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 661, + 507, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 507, + 673 + ], + "score": 1.0, + "content": "research will be required for protecting multi-exit architectures against this emerging security threat.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 48.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 137 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 139 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 344, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 344, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 344, + 156 + ], + "score": 1.0, + "content": "It is unknown if the computational savings provided by", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 345, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 345, + 167 + ], + "score": 1.0, + "content": "multi-exit architectures are robust against adversarial pres-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 345, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 345, + 177 + ], + "score": 1.0, + "content": "sure. Prior research showed that DNNs are vulnerable to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 345, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 345, + 188 + ], + "score": 1.0, + "content": "a wide range of attacks, which involve imperceptible input", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 346, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 346, + 200 + ], + "score": 1.0, + "content": "perturbations (Szegedy et al., 2014; Goodfellow et al., 2015;", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 346, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 346, + 210 + ], + "score": 1.0, + "content": "Papernot et al., 2016; Hu et al., 2020). Considering that a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 345, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 345, + 221 + ], + "score": 1.0, + "content": "multi-exit model, on the worst-case input, does not provide", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 221, + 345, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 345, + 232 + ], + "score": 1.0, + "content": "any computational savings, we ask: Can the savings from", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 346, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 346, + 244 + ], + "score": 1.0, + "content": "multi-exit models be maliciously negated by input perturba-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 241, + 345, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 345, + 254 + ], + "score": 1.0, + "content": "tions? As some natural inputs do require the full depth of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 345, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 345, + 265 + ], + "score": 1.0, + "content": "the model, it may be possible to craft adversarial examples", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 345, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 345, + 276 + ], + "score": 1.0, + "content": "that delay the correct decision; it is unclear, however, how", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 346, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 346, + 287 + ], + "score": 1.0, + "content": "many inputs can be delayed with imperceptible perturba-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 345, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 345, + 298 + ], + "score": 1.0, + "content": "tions. Furthermore, it is unknown if universal versions of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 295, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 506, + 310 + ], + "score": 1.0, + "content": "these adversarial examples exist, if the examples transfer across multi-exit architectures and datasets,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 307, + 457, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 457, + 320 + ], + "score": 1.0, + "content": "or if existing defenses (e.g. adversarial training) are effective against slowdown attacks.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 142, + 346, + 298 + ] + }, + { + "type": "image", + "bbox": [ + 360, + 144, + 496, + 237 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 360, + 144, + 496, + 237 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 360, + 144, + 496, + 237 + ], + "spans": [ + { + "bbox": [ + 360, + 144, + 496, + 237 + ], + "score": 0.965, + "type": "image", + "image_path": "746d8efde4dde3f7f81419a6f62cb228c7561c4c925bbdc393202621421b32de.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 360, + 144, + 496, + 190.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 360, + 190.5, + 496, + 237.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 352, + 252, + 506, + 293 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 351, + 252, + 507, + 264 + ], + "spans": [ + { + "bbox": [ + 351, + 252, + 507, + 264 + ], + "score": 1.0, + "content": "Figure 1: Simple to complex inputs.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 351, + 262, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 351, + 262, + 505, + 273 + ], + "score": 1.0, + "content": "Some Tiny ImageNet images a VGG-16", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 351, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 351, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "model can correctly classify if computa-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 350, + 280, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 350, + 280, + 411, + 295 + ], + "score": 1.0, + "content": "tion stops at the", + "type": "text" + }, + { + "bbox": [ + 411, + 282, + 423, + 292 + ], + "score": 0.74, + "content": "1 ^ { s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 280, + 426, + 295 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 426, + 282, + 440, + 292 + ], + "score": 0.74, + "content": "5 ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 280, + 459, + 295 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 459, + 282, + 477, + 292 + ], + "score": 0.88, + "content": "1 4 ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 280, + 506, + 295 + ], + "score": 1.0, + "content": "layers.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + } + ], + "index": 21.0 + }, + { + "type": "text", + "bbox": [ + 107, + 297, + 504, + 319 + ], + "lines": [], + "index": 25.5, + "bbox_fs": [ + 105, + 295, + 506, + 320 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 325, + 506, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 479, + 339 + ], + "score": 1.0, + "content": "Threat Model. We consider a new threat against DNNs, analogous to the denial-of-service", + "type": "text" + }, + { + "bbox": [ + 479, + 326, + 504, + 337 + ], + "score": 0.68, + "content": "( D o S )", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 335, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 350 + ], + "score": 1.0, + "content": "attacks that have been plaguing the Internet for decades. By imperceptibly perturbing the input to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 345, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 104, + 345, + 506, + 362 + ], + "score": 1.0, + "content": "trigger this worst-case, the adversary aims to slow down the inferences and increase the cost of using", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 358, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 506, + 371 + ], + "score": 1.0, + "content": "the DNN. This is an important threat for many practical applications, which impose strict limits on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 369, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 382 + ], + "score": 1.0, + "content": "the responsiveness and resource usage of DNN models (e.g. in the Internet-of-Things (Taylor et al.,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "2018)), because the adversary could push the victim outside these limits. For example, against a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "commercial image classification system, such as Clarifai.com, a slowdown attack might waste valuable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "computational resources. Against a model partitioning scheme, such as Big-Little (De Coninck et al.,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "2015), it might introduce network latency by forcing excessive transmissions between local and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 422, + 507, + 438 + ], + "spans": [ + { + "bbox": [ + 104, + 422, + 507, + 438 + ], + "score": 1.0, + "content": "remote models. A slowdown attack aims to force the victim to do more work than the adversary, e.g.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "by amplifying the latency needed to process the sample or by crafting reusable perturbations. The", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "adversary may have to achieve this with incomplete information about the multi-exit architecture", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "targeted, the training data used by the victim or the classification task (see discussion in Appendix A).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 325, + 507, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 506, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "Our Contributions. To our best knowledge, we conduct the first study of the robustness of multi-exit", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 484, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 499 + ], + "score": 1.0, + "content": "architectures against adversarial slowdowns. To this end, we find that examples crafted by prior", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "evasion attacks (Madry et al., 2017; Hu et al., 2020) fail to bypass the victim model’s early exits, and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "we show that an adversary can adapt such attacks to the goal of model slowdown by modifying its", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "objective function. We call the resulting attack DeepSloth. We also propose an efficacy metric for", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "comparing slowdowns across different multi-exit architectures. We experiment with three generic", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 539, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 554 + ], + "score": 1.0, + "content": "multi-exit DNNs (based on VGG16, ResNet56 and MobileNet) (Kaya et al., 2019) and a specially-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "score": 1.0, + "content": "designed multi-exit architecture, MSDNets (Huang et al., 2018), on two popular image classification", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "score": 1.0, + "content": "benchmarks (CIFAR-10 and Tiny ImageNet). We find that DeepSloth reduces the efficacy of multi-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 164, + 586 + ], + "score": 1.0, + "content": "exit DNNs by", + "type": "text" + }, + { + "bbox": [ + 164, + 573, + 203, + 584 + ], + "score": 0.86, + "content": "9 0 { - } 1 0 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 573, + 506, + 586 + ], + "score": 1.0, + "content": ", i.e., the perturbations render nearly all early exits ineffective. In a scenario", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 583, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 583, + 506, + 597 + ], + "score": 1.0, + "content": "typical for IoT deployments, where the model is partitioned between edge devices and the cloud, our", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 232, + 608 + ], + "score": 1.0, + "content": "attack amplifies the latency by", + "type": "text" + }, + { + "bbox": [ + 233, + 595, + 264, + 606 + ], + "score": 0.88, + "content": "1 . 5 – 5 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 595, + 506, + 608 + ], + "score": 1.0, + "content": ", negating the benefits of model partitioning. We also show", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "that it is possible to craft a universal DeepSloth perturbation, which can slow down the model on", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 617, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 473, + 629 + ], + "score": 1.0, + "content": "either all or a class of inputs. While more constrained, this attack still reduces the efficacy by", + "type": "text" + }, + { + "bbox": [ + 474, + 617, + 503, + 628 + ], + "score": 0.87, + "content": "5- 4 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 617, + 506, + 629 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 628, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 506, + 640 + ], + "score": 1.0, + "content": "Further, we observe that DeepSloth can be effective in some black-box scenarios, where the attacker", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "has limited knowledge about the victim. Finally, we show that a standard defense against adversarial", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 651, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 506, + 662 + ], + "score": 1.0, + "content": "samples—adversarial training—is inadequate against slowdowns. Our results suggest that further", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 661, + 507, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 507, + 673 + ], + "score": 1.0, + "content": "research will be required for protecting multi-exit architectures against this emerging security threat.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 48.5, + "bbox_fs": [ + 104, + 474, + 507, + 673 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 210, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 213, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 213, + 96 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 107, + 505, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 107, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 107, + 505, + 119 + ], + "score": 1.0, + "content": "Adversarial Examples and Defenses. Prior work on adversarial examples has shown that DNNs are", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 506, + 131 + ], + "score": 1.0, + "content": "vulnerable to test-time input perturbations (Szegedy et al., 2014; Goodfellow et al., 2015; Papernot", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 129, + 504, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 504, + 142 + ], + "score": 1.0, + "content": "et al., 2017; Carlini & Wagner, 2017; Madry et al., 2018). An adversary who wants to maximize", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 140, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 506, + 153 + ], + "score": 1.0, + "content": "a model’s error on specific test-time samples can introduce human-imperceptible perturbations to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 151, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 506, + 164 + ], + "score": 1.0, + "content": "these samples. Moreover, an adversary can also exploit a surrogate model for launching the attack", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "score": 1.0, + "content": "and still hurt an unknown victim (Athalye et al., 2018; Tramèr et al., 2017b; Inkawhich et al., 2019).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "This transferability leads to adversarial examples in more practical black-box scenarios. Although", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 184, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 506, + 198 + ], + "score": 1.0, + "content": "many defenses (Kurakin et al., 2016; Xu et al., 2017; Song et al., 2018; Liao et al., 2018; Lecuyer", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "et al., 2019) have been proposed against this threat, adversarial training (AT) has become the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 206, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 506, + 218 + ], + "score": 1.0, + "content": "frontrunner (Madry et al., 2018). In Sec 5, we evaluate the vulnerability of multi-exit DNNs to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 217, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 228 + ], + "score": 1.0, + "content": "adversarial slowdowns in white-box and black-box scenarios. In Sec 6, we show that standard AT", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 229, + 492, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 492, + 240 + ], + "score": 1.0, + "content": "and its simple adaptation to our perturbations are not sufficient for preventing slowdown attacks.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "Efficient Input-Adaptive Inference. Recent input-adaptive DNN architectures have brought two", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "score": 1.0, + "content": "seemingly distant goals closer: achieving both high predictive quality and computational efficiency.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "There are two types of input-adaptive DNNs: adaptive neural networks (AdNNs) and multi-exit", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "architectures. During the inference, AdNNs (Wang et al., 2018; Figurnov et al., 2017) dynamically", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "skip a certain part of the model to reduce the number of computations. This mechanism can be used", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "only for ResNet-based architectures as they facilitate skipping within a network. On the other hand,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "multi-exit architectures (Teerapittayanon et al., 2016; Huang et al., 2018; Kaya et al., 2019) introduce", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 322, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 334 + ], + "score": 1.0, + "content": "multiple side branches—or early-exits—to a model. During the inference on an input sample, these", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "score": 1.0, + "content": "models can preemptively stop the computation altogether once the stopping criteria are met at one of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "the branches. Kaya et al. (2019) have also identified that standard, non-adaptive DNNs are susceptible", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "to overthinking, i.e., their inability to stop computation leads to inefficient inferences on many inputs.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 384 + ], + "score": 1.0, + "content": "Haque et al. (2020) presented attacks specifically designed for reducing the energy-efficiency of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "score": 1.0, + "content": "AdNNs by using adversarial input perturbations. However, our work studies a new threat model that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "an adversary causes slowdowns on multi-exit architectures. By imperceptibly perturbing the inputs,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "our attacker can (i) introduce network latency to an infrastructure that utilizes multi-exit architectures", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "score": 1.0, + "content": "and (ii) waste the victim’s computational resources. To quantify this vulnerability, we define a new", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 428, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 439 + ], + "score": 1.0, + "content": "metric to measure the impact of adversarial input perturbation on different multi-exit architectures", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "(Sec 3). In Sec 5, we also study practical attack scenarios and the transferability of adversarial input", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "perturbations crafted by our attacker. Moreover, we discuss the potential defense mechanisms against", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "this vulnerability, by proposing a simple adaptation of adversarial training (Sec 6). To the best of our", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 470, + 406, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 406, + 483 + ], + "score": 1.0, + "content": "knowledge, our work is the first systematic study of this new vulnerability.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 488, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Model Partitioning. Model partitioning has been proposed to bring DNNs to resource-constrained", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "devices (De Coninck et al., 2015; Taylor et al., 2018). These schemes split a multi-exit model into", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "sequential components and deploy them in separate endpoints, e.g., a small, local on-device part", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "and a large, cloud-based part. For bringing DNNs to the Internet of Things (IoT), partitioning is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 533, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 506, + 544 + ], + "score": 1.0, + "content": "instrumental as it reduces the transmissions between endpoints, a major bottleneck. In Sec 5.1, on a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 543, + 429, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 429, + 556 + ], + "score": 1.0, + "content": "partitioning scenario, we show that our attack can force excessive transmissions.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "title", + "bbox": [ + 108, + 571, + 244, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 245, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 245, + 585 + ], + "score": 1.0, + "content": "3 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 597, + 504, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "Datasets. We use two datasets: CIFAR-10 (Krizhevsky et al., 2009) and Tiny-ImageNet (Tiny). For", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 608, + 451, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 451, + 621 + ], + "score": 1.0, + "content": "testing the cross-domain transferability of our attacks, we use the CIFAR-100 dataset.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "Architectures and Hyper-parameters. To demonstrate that the vulnerability to adversarial slow-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "downs is common among multi-exit architectures, we experiment on two recent techniques: Shallow-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 648, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 506, + 660 + ], + "score": 1.0, + "content": "Deep Networks (SDNs) (Kaya et al., 2019) and MSDNets (Huang et al., 2018). These architectures", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "were designed for different purposes: SDNs are generic and can convert any DNN into a multi-exit", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "model, and MSDNets are custom designed for efficiency. We evaluate an MSDNet architecture (6", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 681, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 507, + 693 + ], + "score": 1.0, + "content": "exits) and three SDN architectures, based on VGG-16 (Simonyan & Zisserman, 2014) (14 exits),", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 692, + 455, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 455, + 704 + ], + "score": 1.0, + "content": "ResNet-56 (He et al., 2016) (27 exits), and MobileNet (Howard et al., 2017) (14 exits).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Metrics. We define the early-exit capability (EEC) curve of a multi-exit model to indicate the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "fraction of the test samples that exit early at a specific fraction of the model’s full inference cost.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 210, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 213, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 213, + 96 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 107, + 505, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 107, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 107, + 505, + 119 + ], + "score": 1.0, + "content": "Adversarial Examples and Defenses. Prior work on adversarial examples has shown that DNNs are", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 506, + 131 + ], + "score": 1.0, + "content": "vulnerable to test-time input perturbations (Szegedy et al., 2014; Goodfellow et al., 2015; Papernot", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 129, + 504, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 504, + 142 + ], + "score": 1.0, + "content": "et al., 2017; Carlini & Wagner, 2017; Madry et al., 2018). An adversary who wants to maximize", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 140, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 506, + 153 + ], + "score": 1.0, + "content": "a model’s error on specific test-time samples can introduce human-imperceptible perturbations to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 151, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 506, + 164 + ], + "score": 1.0, + "content": "these samples. Moreover, an adversary can also exploit a surrogate model for launching the attack", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "score": 1.0, + "content": "and still hurt an unknown victim (Athalye et al., 2018; Tramèr et al., 2017b; Inkawhich et al., 2019).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "This transferability leads to adversarial examples in more practical black-box scenarios. Although", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 184, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 506, + 198 + ], + "score": 1.0, + "content": "many defenses (Kurakin et al., 2016; Xu et al., 2017; Song et al., 2018; Liao et al., 2018; Lecuyer", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "et al., 2019) have been proposed against this threat, adversarial training (AT) has become the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 206, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 506, + 218 + ], + "score": 1.0, + "content": "frontrunner (Madry et al., 2018). In Sec 5, we evaluate the vulnerability of multi-exit DNNs to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 217, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 228 + ], + "score": 1.0, + "content": "adversarial slowdowns in white-box and black-box scenarios. In Sec 6, we show that standard AT", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 229, + 492, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 492, + 240 + ], + "score": 1.0, + "content": "and its simple adaptation to our perturbations are not sufficient for preventing slowdown attacks.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 107, + 506, + 240 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "Efficient Input-Adaptive Inference. Recent input-adaptive DNN architectures have brought two", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "score": 1.0, + "content": "seemingly distant goals closer: achieving both high predictive quality and computational efficiency.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "There are two types of input-adaptive DNNs: adaptive neural networks (AdNNs) and multi-exit", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "architectures. During the inference, AdNNs (Wang et al., 2018; Figurnov et al., 2017) dynamically", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "skip a certain part of the model to reduce the number of computations. This mechanism can be used", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "only for ResNet-based architectures as they facilitate skipping within a network. On the other hand,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "multi-exit architectures (Teerapittayanon et al., 2016; Huang et al., 2018; Kaya et al., 2019) introduce", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 322, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 334 + ], + "score": 1.0, + "content": "multiple side branches—or early-exits—to a model. During the inference on an input sample, these", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "score": 1.0, + "content": "models can preemptively stop the computation altogether once the stopping criteria are met at one of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "the branches. Kaya et al. (2019) have also identified that standard, non-adaptive DNNs are susceptible", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "to overthinking, i.e., their inability to stop computation leads to inefficient inferences on many inputs.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 245, + 506, + 368 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 384 + ], + "score": 1.0, + "content": "Haque et al. (2020) presented attacks specifically designed for reducing the energy-efficiency of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "score": 1.0, + "content": "AdNNs by using adversarial input perturbations. However, our work studies a new threat model that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "an adversary causes slowdowns on multi-exit architectures. By imperceptibly perturbing the inputs,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "our attacker can (i) introduce network latency to an infrastructure that utilizes multi-exit architectures", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 428 + ], + "score": 1.0, + "content": "and (ii) waste the victim’s computational resources. To quantify this vulnerability, we define a new", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 428, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 439 + ], + "score": 1.0, + "content": "metric to measure the impact of adversarial input perturbation on different multi-exit architectures", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "(Sec 3). In Sec 5, we also study practical attack scenarios and the transferability of adversarial input", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "perturbations crafted by our attacker. Moreover, we discuss the potential defense mechanisms against", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "this vulnerability, by proposing a simple adaptation of adversarial training (Sec 6). To the best of our", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 470, + 406, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 406, + 483 + ], + "score": 1.0, + "content": "knowledge, our work is the first systematic study of this new vulnerability.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 372, + 506, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 488, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Model Partitioning. Model partitioning has been proposed to bring DNNs to resource-constrained", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "devices (De Coninck et al., 2015; Taylor et al., 2018). These schemes split a multi-exit model into", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "sequential components and deploy them in separate endpoints, e.g., a small, local on-device part", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "and a large, cloud-based part. For bringing DNNs to the Internet of Things (IoT), partitioning is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 533, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 506, + 544 + ], + "score": 1.0, + "content": "instrumental as it reduces the transmissions between endpoints, a major bottleneck. In Sec 5.1, on a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 543, + 429, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 429, + 556 + ], + "score": 1.0, + "content": "partitioning scenario, we show that our attack can force excessive transmissions.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 487, + 506, + 556 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 571, + 244, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 245, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 245, + 585 + ], + "score": 1.0, + "content": "3 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 597, + 504, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "Datasets. We use two datasets: CIFAR-10 (Krizhevsky et al., 2009) and Tiny-ImageNet (Tiny). For", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 608, + 451, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 451, + 621 + ], + "score": 1.0, + "content": "testing the cross-domain transferability of our attacks, we use the CIFAR-100 dataset.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 597, + 505, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "Architectures and Hyper-parameters. To demonstrate that the vulnerability to adversarial slow-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "downs is common among multi-exit architectures, we experiment on two recent techniques: Shallow-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 648, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 506, + 660 + ], + "score": 1.0, + "content": "Deep Networks (SDNs) (Kaya et al., 2019) and MSDNets (Huang et al., 2018). These architectures", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "were designed for different purposes: SDNs are generic and can convert any DNN into a multi-exit", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "model, and MSDNets are custom designed for efficiency. We evaluate an MSDNet architecture (6", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 681, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 507, + 693 + ], + "score": 1.0, + "content": "exits) and three SDN architectures, based on VGG-16 (Simonyan & Zisserman, 2014) (14 exits),", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 692, + 455, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 455, + 704 + ], + "score": 1.0, + "content": "ResNet-56 (He et al., 2016) (27 exits), and MobileNet (Howard et al., 2017) (14 exits).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 626, + 507, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Metrics. We define the early-exit capability (EEC) curve of a multi-exit model to indicate the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "fraction of the test samples that exit early at a specific fraction of the model’s full inference cost.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50.5, + "bbox_fs": [ + 105, + 708, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Figure 2 shows the EEC curves of our SDNs on Tiny ImageNet, assuming that the computation stops", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "when there is a correct classification at an exit point. For example, VGG-16-based SDN model can", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 176, + 118 + ], + "score": 1.0, + "content": "correctly classify", + "type": "text" + }, + { + "bbox": [ + 177, + 105, + 204, + 115 + ], + "score": 0.89, + "content": "{ \\sim } 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 105, + 288, + 118 + ], + "score": 1.0, + "content": "of the samples using", + "type": "text" + }, + { + "bbox": [ + 288, + 105, + 316, + 115 + ], + "score": 0.89, + "content": "{ \\sim } 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "of its full cost. Note that this stopping criterion", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 341, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 341, + 128 + ], + "score": 1.0, + "content": "is impractical; in Sec 4, we will discuss the practical ones.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 304, + 318 + ], + "lines": [ + { + "bbox": [ + 106, + 131, + 305, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 305, + 145 + ], + "score": 1.0, + "content": "We define the early-exit efficacy, or efficacy in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 143, + 304, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 304, + 155 + ], + "score": 1.0, + "content": "short, to quantify a model’s ability of utilizing its", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 305, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 305, + 166 + ], + "score": 1.0, + "content": "exit points. The efficacy of a multi-exit model is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 304, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 304, + 177 + ], + "score": 1.0, + "content": "the area under its EEC curve, estimated via the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 177, + 304, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 304, + 187 + ], + "score": 1.0, + "content": "trapezoidal rule. An ideal efficacy for a model is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 304, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 304, + 199 + ], + "score": 1.0, + "content": "close to 1, when most of the input samples the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 199, + 305, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 305, + 210 + ], + "score": 1.0, + "content": "computation stops very early; models that do not", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 305, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 305, + 221 + ], + "score": 1.0, + "content": "use their early exits have 0 efficacy. A model with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 306, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 306, + 232 + ], + "score": 1.0, + "content": "low efficacy generally exhibits a higher latency;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 305, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 305, + 243 + ], + "score": 1.0, + "content": "in a partitioned model, the low efficacy will cause", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 306, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 306, + 253 + ], + "score": 1.0, + "content": "more input transmissions to the cloud, and the la-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 305, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 305, + 264 + ], + "score": 1.0, + "content": "tency is further amplified by the network round", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 306, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 306, + 277 + ], + "score": 1.0, + "content": "trips. A multi-exit model’s efficacy and accuracy", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 306, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 306, + 287 + ], + "score": 1.0, + "content": "are dictated by its stopping criteria, which we dis-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 304, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 304, + 298 + ], + "score": 1.0, + "content": "cuss in the next section. As for the classification", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 297, + 305, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 305, + 309 + ], + "score": 1.0, + "content": "performance, we report the Top-1 accuracy on the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 307, + 145, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 145, + 319 + ], + "score": 1.0, + "content": "test data.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 313, + 137, + 494, + 273 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 322, + 131, + 503, + 138 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 321, + 131, + 504, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 131, + 504, + 139 + ], + "score": 0.982, + "content": "tinyimagenet-vgg16bn(ACC 59.0) resnet56(ACC 53.5) mobilenet(ACC 59.5)", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "image_body", + "bbox": [ + 313, + 137, + 494, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 494, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 494, + 273 + ], + "score": 0.966, + "type": "image", + "image_path": "2bf06ae00279a3cedeb142c999740cbd1b60e5613928b8a30f86d951383600a5.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 313, + 137, + 494, + 149.36363636363637 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 313, + 149.36363636363637, + 494, + 161.72727272727275 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 313, + 161.72727272727275, + 494, + 174.09090909090912 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 313, + 174.09090909090912, + 494, + 186.4545454545455 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 313, + 186.4545454545455, + 494, + 198.81818181818187 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 313, + 198.81818181818187, + 494, + 211.18181818181824 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 313, + 211.18181818181824, + 494, + 223.54545454545462 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 313, + 223.54545454545462, + 494, + 235.909090909091 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 313, + 235.909090909091, + 494, + 248.27272727272737 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 313, + 248.27272727272737, + 494, + 260.63636363636374 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 313, + 260.63636363636374, + 494, + 273.0000000000001 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 312, + 286, + 505, + 330 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 312, + 286, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 286, + 504, + 297 + ], + "score": 1.0, + "content": "Figure 2: The EEC curves. Each curve shows the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "fraction of test samples a model classifies using a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 308, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 312, + 308, + 505, + 318 + ], + "score": 1.0, + "content": "certain fraction of its full inference cost. ‘EFCY’", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 311, + 317, + 443, + 332 + ], + "spans": [ + { + "bbox": [ + 311, + 317, + 443, + 332 + ], + "score": 1.0, + "content": "is short for the model’s efficacy.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 347, + 369, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 369, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 369, + 362 + ], + "score": 1.0, + "content": "4 ATTACKING THE MULTI-EXIT ARCHITECTURES", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 506, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 507, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 507, + 393 + ], + "score": 1.0, + "content": "Setting. We consider the supervised classification setting with standard feedforward DNN archi-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 391, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 104, + 392, + 257, + 403 + ], + "score": 1.0, + "content": "tectures. A DNN model consists of", + "type": "text" + }, + { + "bbox": [ + 257, + 392, + 267, + 402 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 392, + 469, + 403 + ], + "score": 1.0, + "content": "blocks, or layers, that process the input sample,", + "type": "text" + }, + { + "bbox": [ + 469, + 391, + 502, + 402 + ], + "score": 0.91, + "content": "x \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 392, + 506, + 403 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 379, + 414 + ], + "score": 1.0, + "content": "from beginning to end and produce a classification. A classification,", + "type": "text" + }, + { + "bbox": [ + 380, + 402, + 438, + 414 + ], + "score": 0.92, + "content": "F ( x , \\theta ) \\in \\mathbf { \\mathbb { R } } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 402, + 505, + 414 + ], + "score": 1.0, + "content": ", is the predicted", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 104, + 414, + 209, + 426 + ], + "score": 1.0, + "content": "probability distribution of", + "type": "text" + }, + { + "bbox": [ + 209, + 415, + 216, + 424 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 414, + 309, + 426 + ], + "score": 1.0, + "content": "belonging to each label", + "type": "text" + }, + { + "bbox": [ + 309, + 414, + 393, + 426 + ], + "score": 0.94, + "content": "y \\in M = \\{ 1 , . . . , m \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 414, + 420, + 426 + ], + "score": 1.0, + "content": ". Here,", + "type": "text" + }, + { + "bbox": [ + 421, + 415, + 427, + 424 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "denotes the tunable", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 442, + 437 + ], + "score": 1.0, + "content": "parameters, or the weights, of the model. The parameters are learned on a training set", + "type": "text" + }, + { + "bbox": [ + 443, + 425, + 452, + 434 + ], + "score": 0.84, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "that contains", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 144, + 448 + ], + "score": 1.0, + "content": "multiple", + "type": "text" + }, + { + "bbox": [ + 144, + 436, + 174, + 448 + ], + "score": 0.92, + "content": "( x _ { i } , y _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 435, + 228, + 448 + ], + "score": 1.0, + "content": "pairs; where", + "type": "text" + }, + { + "bbox": [ + 229, + 437, + 238, + 447 + ], + "score": 0.84, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 435, + 435, + 448 + ], + "score": 1.0, + "content": "is the ground-truth label of the training sample", + "type": "text" + }, + { + "bbox": [ + 435, + 438, + 445, + 447 + ], + "score": 0.86, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 435, + 483, + 448 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 484, + 436, + 493, + 447 + ], + "score": 0.87, + "content": "\\theta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 272, + 459 + ], + "score": 1.0, + "content": "denote the parameters at and before the", + "type": "text" + }, + { + "bbox": [ + 272, + 446, + 285, + 456 + ], + "score": 0.89, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 446, + 333, + 459 + ], + "score": 1.0, + "content": "block; i.e.,", + "type": "text" + }, + { + "bbox": [ + 334, + 447, + 376, + 458 + ], + "score": 0.92, + "content": "\\theta _ { i } \\subset \\theta _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 446, + 395, + 459 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 396, + 447, + 429, + 457 + ], + "score": 0.91, + "content": "\\theta _ { N } = \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 446, + 506, + 459 + ], + "score": 1.0, + "content": ". Once a model is", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 457, + 381, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 369, + 470 + ], + "score": 1.0, + "content": "trained, its performance is then tested on a set of unseen samples,", + "type": "text" + }, + { + "bbox": [ + 370, + 459, + 377, + 468 + ], + "score": 0.8, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 457, + 381, + 470 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 328, + 487 + ], + "score": 1.0, + "content": "Multi-Exit Architectures. A multi-exit model contains", + "type": "text" + }, + { + "bbox": [ + 329, + 476, + 339, + 485 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "exit points—internal classifiers—attached", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 485, + 507, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 248, + 498 + ], + "score": 1.0, + "content": "to a model’s hidden blocks. We use", + "type": "text" + }, + { + "bbox": [ + 248, + 486, + 259, + 497 + ], + "score": 0.88, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 485, + 313, + 498 + ], + "score": 1.0, + "content": "to denote the", + "type": "text" + }, + { + "bbox": [ + 313, + 486, + 326, + 496 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 485, + 465, + 498 + ], + "score": 1.0, + "content": "exit point, which is attached to the", + "type": "text" + }, + { + "bbox": [ + 465, + 486, + 479, + 497 + ], + "score": 0.89, + "content": "j ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 485, + 507, + 498 + ], + "score": 1.0, + "content": "block.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 495, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 199, + 511 + ], + "score": 1.0, + "content": "Using the output of the", + "type": "text" + }, + { + "bbox": [ + 199, + 497, + 249, + 509 + ], + "score": 0.81, + "content": "j ^ { t h } \\left( j < N \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 495, + 287, + 511 + ], + "score": 1.0, + "content": "block on", + "type": "text" + }, + { + "bbox": [ + 287, + 499, + 294, + 507 + ], + "score": 0.66, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 495, + 298, + 511 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 298, + 498, + 309, + 508 + ], + "score": 0.82, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 495, + 465, + 511 + ], + "score": 1.0, + "content": "produces an internal classification, i.e.,", + "type": "text" + }, + { + "bbox": [ + 465, + 497, + 502, + 509 + ], + "score": 0.92, + "content": "F _ { i } ( x , \\theta _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 495, + 506, + 511 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 219, + 520 + ], + "score": 1.0, + "content": "which we simply denote as", + "type": "text" + }, + { + "bbox": [ + 219, + 509, + 243, + 520 + ], + "score": 0.91, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 508, + 358, + 520 + ], + "score": 1.0, + "content": ". In our experiments, we set", + "type": "text" + }, + { + "bbox": [ + 358, + 509, + 392, + 518 + ], + "score": 0.9, + "content": "K = N", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "for SDNs, i.e., one internal", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 517, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 214, + 532 + ], + "score": 1.0, + "content": "classifier at each block and", + "type": "text" + }, + { + "bbox": [ + 215, + 520, + 244, + 529 + ], + "score": 0.89, + "content": "K = 6", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 517, + 329, + 532 + ], + "score": 1.0, + "content": "for MSDNets. Given", + "type": "text" + }, + { + "bbox": [ + 329, + 519, + 353, + 531 + ], + "score": 0.9, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 517, + 506, + 532 + ], + "score": 1.0, + "content": ", a multi-exit model uses deterministic", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 256, + 542 + ], + "score": 1.0, + "content": "criteria to decide between forwarding", + "type": "text" + }, + { + "bbox": [ + 256, + 532, + 263, + 540 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 530, + 311, + 542 + ], + "score": 1.0, + "content": "to compute", + "type": "text" + }, + { + "bbox": [ + 311, + 531, + 345, + 542 + ], + "score": 0.92, + "content": "F _ { i + 1 } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "and stopping for taking the early-exit at", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 541, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 506, + 553 + ], + "score": 1.0, + "content": "this block. Bypassing early-exits decreases a network’s efficacy as each additional block increases", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 551, + 491, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 491, + 564 + ], + "score": 1.0, + "content": "the inference cost. Note that multi-exit models process each sample individually, not in batches.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 49.5 + }, + { + "type": "text", + "bbox": [ + 106, + 569, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "Practical Stopping Criteria. Ideally, a multi-exit model stops when it reaches a correct classification", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 577, + 508, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 186, + 600 + ], + "score": 1.0, + "content": "at an exit point, i.e.,", + "type": "text" + }, + { + "bbox": [ + 187, + 580, + 319, + 596 + ], + "score": 0.81, + "content": "\\begin{array} { r } { \\operatorname * { a r g m a x } _ { j \\in M } F _ { i } ^ { ( \\bar { j } ) } ( x ) = \\hat { y } _ { i } = y ; y } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 577, + 508, + 600 + ], + "score": 1.0, + "content": "is the ground-truth label. However, for unseen", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 232, + 607 + ], + "score": 1.0, + "content": "samples, this is impractical as", + "type": "text" + }, + { + "bbox": [ + 232, + 597, + 239, + 605 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "is unknown. The prior work has proposed two simple strategies", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 177, + 617 + ], + "score": 1.0, + "content": "to judge whether", + "type": "text" + }, + { + "bbox": [ + 178, + 605, + 207, + 616 + ], + "score": 0.89, + "content": "{ \\hat { y } } _ { i } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 605, + 211, + 617 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 212, + 605, + 236, + 617 + ], + "score": 0.89, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "’s entropy (Teerapittayanon et al., 2016; Huang et al., 2018) or its", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 104, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 480, + 629 + ], + "score": 1.0, + "content": "confidence (Kaya et al., 2019). Our attack (see Sec 4.3) leverages the fact that a uniform", + "type": "text" + }, + { + "bbox": [ + 481, + 616, + 505, + 628 + ], + "score": 0.91, + "content": "F _ { i } ( x )", + "type": "inline_equation" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "has both the highest entropy and the lowest confidence. For generality, we experiment with both", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 637, + 395, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 395, + 651 + ], + "score": 1.0, + "content": "confidence-based—SDNs—and entropy-based—MSDNets—strategies.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 57 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 322, + 667 + ], + "score": 1.0, + "content": "A strategy selects confidence, or entropy, thresholds,", + "type": "text" + }, + { + "bbox": [ + 322, + 655, + 333, + 666 + ], + "score": 0.86, + "content": "T _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 655, + 505, + 667 + ], + "score": 1.0, + "content": ", that determine whether the model should", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 139, + 679 + ], + "score": 1.0, + "content": "take the", + "type": "text" + }, + { + "bbox": [ + 139, + 665, + 152, + 676 + ], + "score": 0.88, + "content": "\\overline { { i } } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 664, + 303, + 679 + ], + "score": 1.0, + "content": "exit for an input sample. Conservative", + "type": "text" + }, + { + "bbox": [ + 304, + 666, + 314, + 677 + ], + "score": 0.87, + "content": "T _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "’s lead to fewer early exits and the opposite hurts", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 269, + 689 + ], + "score": 1.0, + "content": "the accuracy as the estimate of whether", + "type": "text" + }, + { + "bbox": [ + 269, + 677, + 298, + 688 + ], + "score": 0.92, + "content": "{ \\hat { y } } _ { i } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "becomes unreliable. As utility is a major practical", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 172, + 699 + ], + "score": 1.0, + "content": "concern, we set", + "type": "text" + }, + { + "bbox": [ + 173, + 688, + 183, + 699 + ], + "score": 0.86, + "content": "T _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "’s for balancing between efficiency and accuracy. On a holdout set, we set the", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "thresholds to maximize a model’s efficacy while keeping its relative accuracy drop (RAD) over its", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 105, + 710, + 503, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 212, + 722 + ], + "score": 1.0, + "content": "maximum accuracy within", + "type": "text" + }, + { + "bbox": [ + 212, + 710, + 227, + 720 + ], + "score": 0.86, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 710, + 244, + 722 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 245, + 710, + 263, + 720 + ], + "score": 0.88, + "content": "15 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 710, + 416, + 722 + ], + "score": 1.0, + "content": ". We refer to these two settings as RAD", + "type": "text" + }, + { + "bbox": [ + 416, + 710, + 438, + 720 + ], + "score": 0.67, + "content": "123 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 710, + 477, + 722 + ], + "score": 1.0, + "content": "and RAD", + "type": "text" + }, + { + "bbox": [ + 477, + 710, + 503, + 720 + ], + "score": 0.74, + "content": "{ < } 1 5 \\%", + "type": "inline_equation" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 720, + 426, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 426, + 734 + ], + "score": 1.0, + "content": "Table 2 (first segment) shows how accuracy and efficacy change in each setting.", + "type": "text" + } + ], + "index": 67 + } + ], + "index": 64 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Figure 2 shows the EEC curves of our SDNs on Tiny ImageNet, assuming that the computation stops", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "when there is a correct classification at an exit point. For example, VGG-16-based SDN model can", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 176, + 118 + ], + "score": 1.0, + "content": "correctly classify", + "type": "text" + }, + { + "bbox": [ + 177, + 105, + 204, + 115 + ], + "score": 0.89, + "content": "{ \\sim } 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 105, + 288, + 118 + ], + "score": 1.0, + "content": "of the samples using", + "type": "text" + }, + { + "bbox": [ + 288, + 105, + 316, + 115 + ], + "score": 0.89, + "content": "{ \\sim } 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "of its full cost. Note that this stopping criterion", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 341, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 341, + 128 + ], + "score": 1.0, + "content": "is impractical; in Sec 4, we will discuss the practical ones.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 304, + 318 + ], + "lines": [ + { + "bbox": [ + 106, + 131, + 305, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 305, + 145 + ], + "score": 1.0, + "content": "We define the early-exit efficacy, or efficacy in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 143, + 304, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 304, + 155 + ], + "score": 1.0, + "content": "short, to quantify a model’s ability of utilizing its", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 305, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 305, + 166 + ], + "score": 1.0, + "content": "exit points. The efficacy of a multi-exit model is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 304, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 304, + 177 + ], + "score": 1.0, + "content": "the area under its EEC curve, estimated via the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 177, + 304, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 304, + 187 + ], + "score": 1.0, + "content": "trapezoidal rule. An ideal efficacy for a model is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 304, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 304, + 199 + ], + "score": 1.0, + "content": "close to 1, when most of the input samples the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 199, + 305, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 305, + 210 + ], + "score": 1.0, + "content": "computation stops very early; models that do not", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 305, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 305, + 221 + ], + "score": 1.0, + "content": "use their early exits have 0 efficacy. A model with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 306, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 306, + 232 + ], + "score": 1.0, + "content": "low efficacy generally exhibits a higher latency;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 305, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 305, + 243 + ], + "score": 1.0, + "content": "in a partitioned model, the low efficacy will cause", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 306, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 306, + 253 + ], + "score": 1.0, + "content": "more input transmissions to the cloud, and the la-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 305, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 305, + 264 + ], + "score": 1.0, + "content": "tency is further amplified by the network round", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 306, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 306, + 277 + ], + "score": 1.0, + "content": "trips. A multi-exit model’s efficacy and accuracy", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 306, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 306, + 287 + ], + "score": 1.0, + "content": "are dictated by its stopping criteria, which we dis-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 304, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 304, + 298 + ], + "score": 1.0, + "content": "cuss in the next section. As for the classification", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 297, + 305, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 305, + 309 + ], + "score": 1.0, + "content": "performance, we report the Top-1 accuracy on the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 307, + 145, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 145, + 319 + ], + "score": 1.0, + "content": "test data.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 131, + 306, + 319 + ] + }, + { + "type": "image", + "bbox": [ + 313, + 137, + 494, + 273 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 322, + 131, + 503, + 138 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 321, + 131, + 504, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 131, + 504, + 139 + ], + "score": 0.982, + "content": "tinyimagenet-vgg16bn(ACC 59.0) resnet56(ACC 53.5) mobilenet(ACC 59.5)", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "image_body", + "bbox": [ + 313, + 137, + 494, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 494, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 494, + 273 + ], + "score": 0.966, + "type": "image", + "image_path": "2bf06ae00279a3cedeb142c999740cbd1b60e5613928b8a30f86d951383600a5.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 313, + 137, + 494, + 149.36363636363637 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 313, + 149.36363636363637, + 494, + 161.72727272727275 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 313, + 161.72727272727275, + 494, + 174.09090909090912 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 313, + 174.09090909090912, + 494, + 186.4545454545455 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 313, + 186.4545454545455, + 494, + 198.81818181818187 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 313, + 198.81818181818187, + 494, + 211.18181818181824 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 313, + 211.18181818181824, + 494, + 223.54545454545462 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 313, + 223.54545454545462, + 494, + 235.909090909091 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 313, + 235.909090909091, + 494, + 248.27272727272737 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 313, + 248.27272727272737, + 494, + 260.63636363636374 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 313, + 260.63636363636374, + 494, + 273.0000000000001 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 312, + 286, + 505, + 330 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 312, + 286, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 286, + 504, + 297 + ], + "score": 1.0, + "content": "Figure 2: The EEC curves. Each curve shows the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "fraction of test samples a model classifies using a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 308, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 312, + 308, + 505, + 318 + ], + "score": 1.0, + "content": "certain fraction of its full inference cost. ‘EFCY’", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 311, + 317, + 443, + 332 + ], + "spans": [ + { + "bbox": [ + 311, + 317, + 443, + 332 + ], + "score": 1.0, + "content": "is short for the model’s efficacy.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 347, + 369, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 369, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 369, + 362 + ], + "score": 1.0, + "content": "4 ATTACKING THE MULTI-EXIT ARCHITECTURES", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 506, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 507, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 507, + 393 + ], + "score": 1.0, + "content": "Setting. We consider the supervised classification setting with standard feedforward DNN archi-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 391, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 104, + 392, + 257, + 403 + ], + "score": 1.0, + "content": "tectures. A DNN model consists of", + "type": "text" + }, + { + "bbox": [ + 257, + 392, + 267, + 402 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 392, + 469, + 403 + ], + "score": 1.0, + "content": "blocks, or layers, that process the input sample,", + "type": "text" + }, + { + "bbox": [ + 469, + 391, + 502, + 402 + ], + "score": 0.91, + "content": "x \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 392, + 506, + 403 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 379, + 414 + ], + "score": 1.0, + "content": "from beginning to end and produce a classification. A classification,", + "type": "text" + }, + { + "bbox": [ + 380, + 402, + 438, + 414 + ], + "score": 0.92, + "content": "F ( x , \\theta ) \\in \\mathbf { \\mathbb { R } } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 402, + 505, + 414 + ], + "score": 1.0, + "content": ", is the predicted", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 104, + 414, + 209, + 426 + ], + "score": 1.0, + "content": "probability distribution of", + "type": "text" + }, + { + "bbox": [ + 209, + 415, + 216, + 424 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 414, + 309, + 426 + ], + "score": 1.0, + "content": "belonging to each label", + "type": "text" + }, + { + "bbox": [ + 309, + 414, + 393, + 426 + ], + "score": 0.94, + "content": "y \\in M = \\{ 1 , . . . , m \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 414, + 420, + 426 + ], + "score": 1.0, + "content": ". Here,", + "type": "text" + }, + { + "bbox": [ + 421, + 415, + 427, + 424 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "denotes the tunable", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 442, + 437 + ], + "score": 1.0, + "content": "parameters, or the weights, of the model. The parameters are learned on a training set", + "type": "text" + }, + { + "bbox": [ + 443, + 425, + 452, + 434 + ], + "score": 0.84, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "that contains", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 144, + 448 + ], + "score": 1.0, + "content": "multiple", + "type": "text" + }, + { + "bbox": [ + 144, + 436, + 174, + 448 + ], + "score": 0.92, + "content": "( x _ { i } , y _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 435, + 228, + 448 + ], + "score": 1.0, + "content": "pairs; where", + "type": "text" + }, + { + "bbox": [ + 229, + 437, + 238, + 447 + ], + "score": 0.84, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 435, + 435, + 448 + ], + "score": 1.0, + "content": "is the ground-truth label of the training sample", + "type": "text" + }, + { + "bbox": [ + 435, + 438, + 445, + 447 + ], + "score": 0.86, + "content": "x _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 435, + 483, + 448 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 484, + 436, + 493, + 447 + ], + "score": 0.87, + "content": "\\theta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 272, + 459 + ], + "score": 1.0, + "content": "denote the parameters at and before the", + "type": "text" + }, + { + "bbox": [ + 272, + 446, + 285, + 456 + ], + "score": 0.89, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 446, + 333, + 459 + ], + "score": 1.0, + "content": "block; i.e.,", + "type": "text" + }, + { + "bbox": [ + 334, + 447, + 376, + 458 + ], + "score": 0.92, + "content": "\\theta _ { i } \\subset \\theta _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 446, + 395, + 459 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 396, + 447, + 429, + 457 + ], + "score": 0.91, + "content": "\\theta _ { N } = \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 446, + 506, + 459 + ], + "score": 1.0, + "content": ". Once a model is", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 457, + 381, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 369, + 470 + ], + "score": 1.0, + "content": "trained, its performance is then tested on a set of unseen samples,", + "type": "text" + }, + { + "bbox": [ + 370, + 459, + 377, + 468 + ], + "score": 0.8, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 457, + 381, + 470 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5, + "bbox_fs": [ + 104, + 380, + 507, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 328, + 487 + ], + "score": 1.0, + "content": "Multi-Exit Architectures. A multi-exit model contains", + "type": "text" + }, + { + "bbox": [ + 329, + 476, + 339, + 485 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "exit points—internal classifiers—attached", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 485, + 507, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 248, + 498 + ], + "score": 1.0, + "content": "to a model’s hidden blocks. We use", + "type": "text" + }, + { + "bbox": [ + 248, + 486, + 259, + 497 + ], + "score": 0.88, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 485, + 313, + 498 + ], + "score": 1.0, + "content": "to denote the", + "type": "text" + }, + { + "bbox": [ + 313, + 486, + 326, + 496 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 485, + 465, + 498 + ], + "score": 1.0, + "content": "exit point, which is attached to the", + "type": "text" + }, + { + "bbox": [ + 465, + 486, + 479, + 497 + ], + "score": 0.89, + "content": "j ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 485, + 507, + 498 + ], + "score": 1.0, + "content": "block.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 495, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 199, + 511 + ], + "score": 1.0, + "content": "Using the output of the", + "type": "text" + }, + { + "bbox": [ + 199, + 497, + 249, + 509 + ], + "score": 0.81, + "content": "j ^ { t h } \\left( j < N \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 495, + 287, + 511 + ], + "score": 1.0, + "content": "block on", + "type": "text" + }, + { + "bbox": [ + 287, + 499, + 294, + 507 + ], + "score": 0.66, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 495, + 298, + 511 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 298, + 498, + 309, + 508 + ], + "score": 0.82, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 495, + 465, + 511 + ], + "score": 1.0, + "content": "produces an internal classification, i.e.,", + "type": "text" + }, + { + "bbox": [ + 465, + 497, + 502, + 509 + ], + "score": 0.92, + "content": "F _ { i } ( x , \\theta _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 495, + 506, + 511 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 219, + 520 + ], + "score": 1.0, + "content": "which we simply denote as", + "type": "text" + }, + { + "bbox": [ + 219, + 509, + 243, + 520 + ], + "score": 0.91, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 508, + 358, + 520 + ], + "score": 1.0, + "content": ". In our experiments, we set", + "type": "text" + }, + { + "bbox": [ + 358, + 509, + 392, + 518 + ], + "score": 0.9, + "content": "K = N", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "for SDNs, i.e., one internal", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 517, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 214, + 532 + ], + "score": 1.0, + "content": "classifier at each block and", + "type": "text" + }, + { + "bbox": [ + 215, + 520, + 244, + 529 + ], + "score": 0.89, + "content": "K = 6", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 517, + 329, + 532 + ], + "score": 1.0, + "content": "for MSDNets. Given", + "type": "text" + }, + { + "bbox": [ + 329, + 519, + 353, + 531 + ], + "score": 0.9, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 517, + 506, + 532 + ], + "score": 1.0, + "content": ", a multi-exit model uses deterministic", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 256, + 542 + ], + "score": 1.0, + "content": "criteria to decide between forwarding", + "type": "text" + }, + { + "bbox": [ + 256, + 532, + 263, + 540 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 530, + 311, + 542 + ], + "score": 1.0, + "content": "to compute", + "type": "text" + }, + { + "bbox": [ + 311, + 531, + 345, + 542 + ], + "score": 0.92, + "content": "F _ { i + 1 } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "and stopping for taking the early-exit at", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 541, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 506, + 553 + ], + "score": 1.0, + "content": "this block. Bypassing early-exits decreases a network’s efficacy as each additional block increases", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 551, + 491, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 491, + 564 + ], + "score": 1.0, + "content": "the inference cost. Note that multi-exit models process each sample individually, not in batches.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 475, + 507, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 569, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "Practical Stopping Criteria. Ideally, a multi-exit model stops when it reaches a correct classification", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 577, + 508, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 186, + 600 + ], + "score": 1.0, + "content": "at an exit point, i.e.,", + "type": "text" + }, + { + "bbox": [ + 187, + 580, + 319, + 596 + ], + "score": 0.81, + "content": "\\begin{array} { r } { \\operatorname * { a r g m a x } _ { j \\in M } F _ { i } ^ { ( \\bar { j } ) } ( x ) = \\hat { y } _ { i } = y ; y } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 577, + 508, + 600 + ], + "score": 1.0, + "content": "is the ground-truth label. However, for unseen", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 232, + 607 + ], + "score": 1.0, + "content": "samples, this is impractical as", + "type": "text" + }, + { + "bbox": [ + 232, + 597, + 239, + 605 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "is unknown. The prior work has proposed two simple strategies", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 177, + 617 + ], + "score": 1.0, + "content": "to judge whether", + "type": "text" + }, + { + "bbox": [ + 178, + 605, + 207, + 616 + ], + "score": 0.89, + "content": "{ \\hat { y } } _ { i } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 605, + 211, + 617 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 212, + 605, + 236, + 617 + ], + "score": 0.89, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "’s entropy (Teerapittayanon et al., 2016; Huang et al., 2018) or its", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 104, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 104, + 615, + 480, + 629 + ], + "score": 1.0, + "content": "confidence (Kaya et al., 2019). Our attack (see Sec 4.3) leverages the fact that a uniform", + "type": "text" + }, + { + "bbox": [ + 481, + 616, + 505, + 628 + ], + "score": 0.91, + "content": "F _ { i } ( x )", + "type": "inline_equation" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "has both the highest entropy and the lowest confidence. For generality, we experiment with both", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 637, + 395, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 395, + 651 + ], + "score": 1.0, + "content": "confidence-based—SDNs—and entropy-based—MSDNets—strategies.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 57, + "bbox_fs": [ + 104, + 569, + 508, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 322, + 667 + ], + "score": 1.0, + "content": "A strategy selects confidence, or entropy, thresholds,", + "type": "text" + }, + { + "bbox": [ + 322, + 655, + 333, + 666 + ], + "score": 0.86, + "content": "T _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 655, + 505, + 667 + ], + "score": 1.0, + "content": ", that determine whether the model should", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 139, + 679 + ], + "score": 1.0, + "content": "take the", + "type": "text" + }, + { + "bbox": [ + 139, + 665, + 152, + 676 + ], + "score": 0.88, + "content": "\\overline { { i } } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 664, + 303, + 679 + ], + "score": 1.0, + "content": "exit for an input sample. Conservative", + "type": "text" + }, + { + "bbox": [ + 304, + 666, + 314, + 677 + ], + "score": 0.87, + "content": "T _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "’s lead to fewer early exits and the opposite hurts", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 269, + 689 + ], + "score": 1.0, + "content": "the accuracy as the estimate of whether", + "type": "text" + }, + { + "bbox": [ + 269, + 677, + 298, + 688 + ], + "score": 0.92, + "content": "{ \\hat { y } } _ { i } = y", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "becomes unreliable. As utility is a major practical", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 172, + 699 + ], + "score": 1.0, + "content": "concern, we set", + "type": "text" + }, + { + "bbox": [ + 173, + 688, + 183, + 699 + ], + "score": 0.86, + "content": "T _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "’s for balancing between efficiency and accuracy. On a holdout set, we set the", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "thresholds to maximize a model’s efficacy while keeping its relative accuracy drop (RAD) over its", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 105, + 710, + 503, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 212, + 722 + ], + "score": 1.0, + "content": "maximum accuracy within", + "type": "text" + }, + { + "bbox": [ + 212, + 710, + 227, + 720 + ], + "score": 0.86, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 710, + 244, + 722 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 245, + 710, + 263, + 720 + ], + "score": 0.88, + "content": "15 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 710, + 416, + 722 + ], + "score": 1.0, + "content": ". We refer to these two settings as RAD", + "type": "text" + }, + { + "bbox": [ + 416, + 710, + 438, + 720 + ], + "score": 0.67, + "content": "123 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 710, + 477, + 722 + ], + "score": 1.0, + "content": "and RAD", + "type": "text" + }, + { + "bbox": [ + 477, + 710, + 503, + 720 + ], + "score": 0.74, + "content": "{ < } 1 5 \\%", + "type": "inline_equation" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 720, + 426, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 426, + 734 + ], + "score": 1.0, + "content": "Table 2 (first segment) shows how accuracy and efficacy change in each setting.", + "type": "text" + } + ], + "index": 67 + } + ], + "index": 64, + "bbox_fs": [ + 105, + 655, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 201, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 203, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 203, + 95 + ], + "score": 1.0, + "content": "4.1 THREAT MODEL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "We consider an adversary who aims to decrease the early-exit efficacy of a victim model. The attacker", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 307, + 129 + ], + "score": 1.0, + "content": "crafts an imperceptible adversarial perturbation,", + "type": "text" + }, + { + "bbox": [ + 307, + 115, + 340, + 126 + ], + "score": 0.92, + "content": "v \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 115, + 505, + 129 + ], + "score": 1.0, + "content": "that, when added to a test-time sample", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 127, + 310, + 139 + ], + "spans": [ + { + "bbox": [ + 107, + 127, + 132, + 137 + ], + "score": 0.87, + "content": "x \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 127, + 310, + 139 + ], + "score": 1.0, + "content": ", prevents the model from taking early-exits.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 158 + ], + "score": 1.0, + "content": "Adversary’s Capabilities. The attacker is able to modify the victim’s test-time samples to apply", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 155, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 168 + ], + "score": 1.0, + "content": "the perturbations, e.g., by compromising a camera that collects the data for inference. To ensure the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 167, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 221, + 179 + ], + "score": 1.0, + "content": "imperceptibility, we focus on", + "type": "text" + }, + { + "bbox": [ + 222, + 167, + 236, + 177 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 167, + 505, + 179 + ], + "score": 1.0, + "content": "norm bounded perturbations as they (i) are well are studied; (ii) have", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "successful defenses (Madry et al., 2018); (iii) have prior extension to multi-exit models (Hu et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 188, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 408, + 200 + ], + "score": 1.0, + "content": "2020); and (iv) are usually the most efficient to craft. We show results on", + "type": "text" + }, + { + "bbox": [ + 409, + 189, + 419, + 199 + ], + "score": 0.86, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 188, + 437, + 200 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 189, + 447, + 199 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 188, + 506, + 200 + ], + "score": 1.0, + "content": "perturbations", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 199, + 507, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 507, + 212 + ], + "score": 1.0, + "content": "in Appendix C. In line with the prior work, we bound the perturbations as follows: for CIFAR-10,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 210, + 181, + 222 + ], + "score": 0.91, + "content": "\\lvert | v \\rvert | _ { \\infty } \\leq \\epsilon = 0 . 0 3", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 210, + 266, + 223 + ], + "score": 1.0, + "content": "(Madry et al., 2017),", + "type": "text" + }, + { + "bbox": [ + 266, + 210, + 307, + 222 + ], + "score": 0.9, + "content": "| | v | | _ { 1 } \\leq 8", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 210, + 425, + 223 + ], + "score": 1.0, + "content": "(Tramèr & Boneh, 2019) and", + "type": "text" + }, + { + "bbox": [ + 425, + 210, + 478, + 222 + ], + "score": 0.93, + "content": "| | v | | _ { 2 } \\leq 0 . 3 5", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "(Chen", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 503, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 238, + 234 + ], + "score": 1.0, + "content": "et al., 2017); for Tiny ImageNet,", + "type": "text" + }, + { + "bbox": [ + 238, + 222, + 312, + 233 + ], + "score": 0.9, + "content": "| | v | | _ { \\infty } \\leq \\epsilon = 0 . 0 3", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 221, + 392, + 234 + ], + "score": 1.0, + "content": "(Yang et al., 2019),", + "type": "text" + }, + { + "bbox": [ + 392, + 221, + 437, + 233 + ], + "score": 0.91, + "content": "| | \\boldsymbol { v } | | _ { 1 } \\le \\mathrm { \\dot { 1 6 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 221, + 455, + 234 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 456, + 221, + 503, + 233 + ], + "score": 0.92, + "content": "| | v | | _ { 2 } \\leq 0 . 6", + "type": "inline_equation" + } + ], + "index": 11 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "Adversary’s Knowledge. To assess the security vulnerability of multi-exit architectures, we study", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 477, + 262 + ], + "score": 1.0, + "content": "white-box scenarios, i.e., the attacker knows all the details of the victim model, including its", + "type": "text" + }, + { + "bbox": [ + 478, + 250, + 487, + 260 + ], + "score": 0.82, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 259, + 507, + 274 + ], + "spans": [ + { + "bbox": [ + 107, + 261, + 113, + 271 + ], + "score": 0.61, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 259, + 477, + 274 + ], + "score": 1.0, + "content": ". Further, in Sec 5.2, we study more practical black-box scenarios, i.e., the attacker crafts", + "type": "text" + }, + { + "bbox": [ + 478, + 263, + 484, + 271 + ], + "score": 0.77, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 259, + 507, + 274 + ], + "score": 1.0, + "content": "on a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 273, + 350, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 350, + 284 + ], + "score": 1.0, + "content": "surrogate model and applies it to an unknown victim model.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 504, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 504, + 301 + ], + "score": 1.0, + "content": "Adversary’s Goals. We consider three DeepSloth variants, (i) the standard, (ii) the universal and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 301, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 351, + 312 + ], + "score": 1.0, + "content": "(iii) the class-universal. The adversary, in (i) crafts a different", + "type": "text" + }, + { + "bbox": [ + 352, + 302, + 358, + 310 + ], + "score": 0.77, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 301, + 393, + 312 + ], + "score": 1.0, + "content": "for each", + "type": "text" + }, + { + "bbox": [ + 393, + 301, + 419, + 311 + ], + "score": 0.9, + "content": "x \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 301, + 504, + 312 + ], + "score": 1.0, + "content": "; in (ii) crafts a single", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 310, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 107, + 313, + 113, + 321 + ], + "score": 0.71, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 310, + 142, + 325 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 142, + 312, + 169, + 322 + ], + "score": 0.9, + "content": "x \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 310, + 262, + 325 + ], + "score": 1.0, + "content": "; in (iii) crafts a single", + "type": "text" + }, + { + "bbox": [ + 262, + 313, + 269, + 321 + ], + "score": 0.71, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 310, + 340, + 325 + ], + "score": 1.0, + "content": "for a target class", + "type": "text" + }, + { + "bbox": [ + 341, + 312, + 370, + 322 + ], + "score": 0.9, + "content": "i \\in M", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 310, + 506, + 325 + ], + "score": 1.0, + "content": ". Further, although the adversary", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 336 + ], + "score": 1.0, + "content": "does not explicitly target it; we observe that DeepSloth usually hurts the accuracy. By modifying", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 333, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 346 + ], + "score": 1.0, + "content": "the objective function we describe in Sec 4.3, we also experiment with DeepSloth variants that can", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 344, + 405, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 405, + 357 + ], + "score": 1.0, + "content": "explicitly preserve or hurt the accuracy, in addition to causing slowdowns.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 108, + 374, + 393, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 395, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 395, + 387 + ], + "score": 1.0, + "content": "4.2 STANDARD ADVERSARIAL ATTACKS DO NOT CAUSE DELAYS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 506, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 409 + ], + "score": 1.0, + "content": "To motivate DeepSloth, we first evaluate whether previous adversarial attacks have any effect on the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "efficacy of multi-exit models. These attacks add imperceptible perturbations to a victim’s test-time", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 419, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 506, + 431 + ], + "score": 1.0, + "content": "samples to force misclassifications. We experiment with the standard PGD attack (Madry et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "2017); PGD-avg and PGD-max variants against multi-exit models (Hu et al., 2020) and the Universal", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 441, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 506, + 452 + ], + "score": 1.0, + "content": "Adversarial Perturbation (UAP) attack that crafts a single perturbation for all test samples (Moosavi-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "Dezfooli et al., 2017). Table 1 summarizes our findings that these attacks, although they hurt the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 474 + ], + "score": 1.0, + "content": "accuracy, fail to cause any meaningful decrease in efficacy. In many cases, we observe that the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 474, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 506, + 485 + ], + "score": 1.0, + "content": "attacks actually increase the efficacy. These experiments help us to identify the critical elements of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 483, + 352, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 352, + 498 + ], + "score": 1.0, + "content": "the objective function of an attack that decreases the efficacy.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 106, + 551, + 510, + 610 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 511, + 505, + 542 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 510, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 522 + ], + "score": 1.0, + "content": "Table 1: Impact of existing evasion attacks on efficacy. Each entry shows a model’s efficacy (left) and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 521, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 532 + ], + "score": 1.0, + "content": "accuracy (right) when subjected to the respective attack. The multi-exit models are trained on CIFAR-10 and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 531, + 258, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 140, + 543 + ], + "score": 1.0, + "content": "use RAD", + "type": "text" + }, + { + "bbox": [ + 140, + 532, + 160, + 541 + ], + "score": 0.78, + "content": "{ < } 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 531, + 258, + 543 + ], + "score": 1.0, + "content": "as their early-exit strategy.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "table_body", + "bbox": [ + 106, + 551, + 510, + 610 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 551, + 510, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 510, + 610 + ], + "score": 0.979, + "html": "
NETWORKNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)UAP
VGG-160.77 /89%0.79 / 29%0.85 /10%0.81/27%0.71/68%
REsNET-560.52 / 87%0.55 / 12%0.82/1%0.70/ 6%0.55 / 44%
MOBILENET0.83/87%0.85 /14%0.93/ 3%0.89 / 12%0.77 / 60%
", + "type": "table", + "image_path": "82e0346a0d05698abb99c5ca3e4c2c3416ecb56f0610de4ca3f806a7f479f01b.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 106, + 551, + 510, + 570.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 106, + 570.6666666666666, + 510, + 590.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 106, + 590.3333333333333, + 510, + 609.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 107, + 631, + 242, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 631, + 243, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 243, + 644 + ], + "score": 1.0, + "content": "4.3 THE DEEPSLOTH ATTACK", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "The Layer-Wise Objective Function. Figure 3 shows that the attacks that only optimize for the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "final output, e.g., PGD or UAP, do not perturb the model’s earlier layer representations. This does not", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "bypass the early-exits, which makes these attacks ineffective for decreasing the efficacy. Therefore,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "we modify the objective functions of adversarial example-crafting algorithms to incorporate the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 160, + 712 + ], + "score": 1.0, + "content": "outputs of all", + "type": "text" + }, + { + "bbox": [ + 160, + 699, + 200, + 711 + ], + "score": 0.91, + "content": "F _ { i } | i < K", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 698, + 252, + 712 + ], + "score": 1.0, + "content": ". For crafting", + "type": "text" + }, + { + "bbox": [ + 252, + 699, + 266, + 710 + ], + "score": 0.87, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 698, + 269, + 712 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 270, + 699, + 279, + 710 + ], + "score": 0.84, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 698, + 296, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 297, + 699, + 306, + 710 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "-bounded perturbations, we adapt the PGD (Madry", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "et al., 2017), the DDN (Rony et al., 2019) and the SLIDE algorithms (Tramèr & Boneh, 2019),", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "respectively. Next, we describe how we modify the PGD algorithm—we modify the others similarly:", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 201, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 203, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 203, + 95 + ], + "score": 1.0, + "content": "4.1 THREAT MODEL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "We consider an adversary who aims to decrease the early-exit efficacy of a victim model. The attacker", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 115, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 307, + 129 + ], + "score": 1.0, + "content": "crafts an imperceptible adversarial perturbation,", + "type": "text" + }, + { + "bbox": [ + 307, + 115, + 340, + 126 + ], + "score": 0.92, + "content": "v \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 115, + 505, + 129 + ], + "score": 1.0, + "content": "that, when added to a test-time sample", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 127, + 310, + 139 + ], + "spans": [ + { + "bbox": [ + 107, + 127, + 132, + 137 + ], + "score": 0.87, + "content": "x \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 127, + 310, + 139 + ], + "score": 1.0, + "content": ", prevents the model from taking early-exits.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 104, + 506, + 139 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 158 + ], + "score": 1.0, + "content": "Adversary’s Capabilities. The attacker is able to modify the victim’s test-time samples to apply", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 155, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 168 + ], + "score": 1.0, + "content": "the perturbations, e.g., by compromising a camera that collects the data for inference. To ensure the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 167, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 221, + 179 + ], + "score": 1.0, + "content": "imperceptibility, we focus on", + "type": "text" + }, + { + "bbox": [ + 222, + 167, + 236, + 177 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 167, + 505, + 179 + ], + "score": 1.0, + "content": "norm bounded perturbations as they (i) are well are studied; (ii) have", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "successful defenses (Madry et al., 2018); (iii) have prior extension to multi-exit models (Hu et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 188, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 408, + 200 + ], + "score": 1.0, + "content": "2020); and (iv) are usually the most efficient to craft. We show results on", + "type": "text" + }, + { + "bbox": [ + 409, + 189, + 419, + 199 + ], + "score": 0.86, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 188, + 437, + 200 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 189, + 447, + 199 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 188, + 506, + 200 + ], + "score": 1.0, + "content": "perturbations", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 199, + 507, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 507, + 212 + ], + "score": 1.0, + "content": "in Appendix C. In line with the prior work, we bound the perturbations as follows: for CIFAR-10,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 210, + 181, + 222 + ], + "score": 0.91, + "content": "\\lvert | v \\rvert | _ { \\infty } \\leq \\epsilon = 0 . 0 3", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 210, + 266, + 223 + ], + "score": 1.0, + "content": "(Madry et al., 2017),", + "type": "text" + }, + { + "bbox": [ + 266, + 210, + 307, + 222 + ], + "score": 0.9, + "content": "| | v | | _ { 1 } \\leq 8", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 210, + 425, + 223 + ], + "score": 1.0, + "content": "(Tramèr & Boneh, 2019) and", + "type": "text" + }, + { + "bbox": [ + 425, + 210, + 478, + 222 + ], + "score": 0.93, + "content": "| | v | | _ { 2 } \\leq 0 . 3 5", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "(Chen", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 503, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 238, + 234 + ], + "score": 1.0, + "content": "et al., 2017); for Tiny ImageNet,", + "type": "text" + }, + { + "bbox": [ + 238, + 222, + 312, + 233 + ], + "score": 0.9, + "content": "| | v | | _ { \\infty } \\leq \\epsilon = 0 . 0 3", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 221, + 392, + 234 + ], + "score": 1.0, + "content": "(Yang et al., 2019),", + "type": "text" + }, + { + "bbox": [ + 392, + 221, + 437, + 233 + ], + "score": 0.91, + "content": "| | \\boldsymbol { v } | | _ { 1 } \\le \\mathrm { \\dot { 1 6 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 221, + 455, + 234 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 456, + 221, + 503, + 233 + ], + "score": 0.92, + "content": "| | v | | _ { 2 } \\leq 0 . 6", + "type": "inline_equation" + } + ], + "index": 11 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 143, + 507, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "Adversary’s Knowledge. To assess the security vulnerability of multi-exit architectures, we study", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 477, + 262 + ], + "score": 1.0, + "content": "white-box scenarios, i.e., the attacker knows all the details of the victim model, including its", + "type": "text" + }, + { + "bbox": [ + 478, + 250, + 487, + 260 + ], + "score": 0.82, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 259, + 507, + 274 + ], + "spans": [ + { + "bbox": [ + 107, + 261, + 113, + 271 + ], + "score": 0.61, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 259, + 477, + 274 + ], + "score": 1.0, + "content": ". Further, in Sec 5.2, we study more practical black-box scenarios, i.e., the attacker crafts", + "type": "text" + }, + { + "bbox": [ + 478, + 263, + 484, + 271 + ], + "score": 0.77, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 259, + 507, + 274 + ], + "score": 1.0, + "content": "on a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 273, + 350, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 350, + 284 + ], + "score": 1.0, + "content": "surrogate model and applies it to an unknown victim model.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 238, + 507, + 284 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 504, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 504, + 301 + ], + "score": 1.0, + "content": "Adversary’s Goals. We consider three DeepSloth variants, (i) the standard, (ii) the universal and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 301, + 504, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 351, + 312 + ], + "score": 1.0, + "content": "(iii) the class-universal. The adversary, in (i) crafts a different", + "type": "text" + }, + { + "bbox": [ + 352, + 302, + 358, + 310 + ], + "score": 0.77, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 301, + 393, + 312 + ], + "score": 1.0, + "content": "for each", + "type": "text" + }, + { + "bbox": [ + 393, + 301, + 419, + 311 + ], + "score": 0.9, + "content": "x \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 301, + 504, + 312 + ], + "score": 1.0, + "content": "; in (ii) crafts a single", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 310, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 107, + 313, + 113, + 321 + ], + "score": 0.71, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 310, + 142, + 325 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 142, + 312, + 169, + 322 + ], + "score": 0.9, + "content": "x \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 310, + 262, + 325 + ], + "score": 1.0, + "content": "; in (iii) crafts a single", + "type": "text" + }, + { + "bbox": [ + 262, + 313, + 269, + 321 + ], + "score": 0.71, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 310, + 340, + 325 + ], + "score": 1.0, + "content": "for a target class", + "type": "text" + }, + { + "bbox": [ + 341, + 312, + 370, + 322 + ], + "score": 0.9, + "content": "i \\in M", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 310, + 506, + 325 + ], + "score": 1.0, + "content": ". Further, although the adversary", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 336 + ], + "score": 1.0, + "content": "does not explicitly target it; we observe that DeepSloth usually hurts the accuracy. By modifying", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 333, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 346 + ], + "score": 1.0, + "content": "the objective function we describe in Sec 4.3, we also experiment with DeepSloth variants that can", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 344, + 405, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 405, + 357 + ], + "score": 1.0, + "content": "explicitly preserve or hurt the accuracy, in addition to causing slowdowns.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 290, + 506, + 357 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 374, + 393, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 395, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 395, + 387 + ], + "score": 1.0, + "content": "4.2 STANDARD ADVERSARIAL ATTACKS DO NOT CAUSE DELAYS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 506, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 409 + ], + "score": 1.0, + "content": "To motivate DeepSloth, we first evaluate whether previous adversarial attacks have any effect on the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "efficacy of multi-exit models. These attacks add imperceptible perturbations to a victim’s test-time", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 419, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 506, + 431 + ], + "score": 1.0, + "content": "samples to force misclassifications. We experiment with the standard PGD attack (Madry et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "2017); PGD-avg and PGD-max variants against multi-exit models (Hu et al., 2020) and the Universal", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 441, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 506, + 452 + ], + "score": 1.0, + "content": "Adversarial Perturbation (UAP) attack that crafts a single perturbation for all test samples (Moosavi-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "Dezfooli et al., 2017). Table 1 summarizes our findings that these attacks, although they hurt the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 474 + ], + "score": 1.0, + "content": "accuracy, fail to cause any meaningful decrease in efficacy. In many cases, we observe that the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 474, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 506, + 485 + ], + "score": 1.0, + "content": "attacks actually increase the efficacy. These experiments help us to identify the critical elements of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 483, + 352, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 352, + 498 + ], + "score": 1.0, + "content": "the objective function of an attack that decreases the efficacy.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 395, + 506, + 498 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 551, + 510, + 610 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 511, + 505, + 542 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 510, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 522 + ], + "score": 1.0, + "content": "Table 1: Impact of existing evasion attacks on efficacy. Each entry shows a model’s efficacy (left) and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 521, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 532 + ], + "score": 1.0, + "content": "accuracy (right) when subjected to the respective attack. The multi-exit models are trained on CIFAR-10 and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 531, + 258, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 140, + 543 + ], + "score": 1.0, + "content": "use RAD", + "type": "text" + }, + { + "bbox": [ + 140, + 532, + 160, + 541 + ], + "score": 0.78, + "content": "{ < } 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 531, + 258, + 543 + ], + "score": 1.0, + "content": "as their early-exit strategy.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "table_body", + "bbox": [ + 106, + 551, + 510, + 610 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 551, + 510, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 510, + 610 + ], + "score": 0.979, + "html": "
NETWORKNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)UAP
VGG-160.77 /89%0.79 / 29%0.85 /10%0.81/27%0.71/68%
REsNET-560.52 / 87%0.55 / 12%0.82/1%0.70/ 6%0.55 / 44%
MOBILENET0.83/87%0.85 /14%0.93/ 3%0.89 / 12%0.77 / 60%
", + "type": "table", + "image_path": "82e0346a0d05698abb99c5ca3e4c2c3416ecb56f0610de4ca3f806a7f479f01b.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 106, + 551, + 510, + 570.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 106, + 570.6666666666666, + 510, + 590.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 106, + 590.3333333333333, + 510, + 609.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 107, + 631, + 242, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 631, + 243, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 243, + 644 + ], + "score": 1.0, + "content": "4.3 THE DEEPSLOTH ATTACK", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "The Layer-Wise Objective Function. Figure 3 shows that the attacks that only optimize for the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "final output, e.g., PGD or UAP, do not perturb the model’s earlier layer representations. This does not", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "bypass the early-exits, which makes these attacks ineffective for decreasing the efficacy. Therefore,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 700 + ], + "score": 1.0, + "content": "we modify the objective functions of adversarial example-crafting algorithms to incorporate the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 160, + 712 + ], + "score": 1.0, + "content": "outputs of all", + "type": "text" + }, + { + "bbox": [ + 160, + 699, + 200, + 711 + ], + "score": 0.91, + "content": "F _ { i } | i < K", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 698, + 252, + 712 + ], + "score": 1.0, + "content": ". For crafting", + "type": "text" + }, + { + "bbox": [ + 252, + 699, + 266, + 710 + ], + "score": 0.87, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 698, + 269, + 712 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 270, + 699, + 279, + 710 + ], + "score": 0.84, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 698, + 296, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 297, + 699, + 306, + 710 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "-bounded perturbations, we adapt the PGD (Madry", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "et al., 2017), the DDN (Rony et al., 2019) and the SLIDE algorithms (Tramèr & Boneh, 2019),", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "respectively. Next, we describe how we modify the PGD algorithm—we modify the others similarly:", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 654, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "interline_equation", + "bbox": [ + 157, + 90, + 454, + 125 + ], + "lines": [ + { + "bbox": [ + 157, + 90, + 454, + 125 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 454, + 125 + ], + "score": 0.93, + "content": "v ^ { t + 1 } = \\prod _ { | | v | | _ { \\infty } < \\epsilon } \\left( v ^ { t } + \\alpha \\operatorname { s g n } \\left( { \\nabla } _ { v } \\sum _ { x \\in D ^ { \\prime } } \\sum _ { 0 < i < K } { \\mathcal { L } } \\left( F _ { i } \\left( x + v \\right) , \\bar { y } \\right) \\right) \\right)", + "type": "interline_equation", + "image_path": "86efb71bee2692f9ee3c0c9c37aeec2609a9c74c0132e9fd6eb8385d4f735434.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 157, + 90, + 454, + 101.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 157, + 101.66666666666667, + 454, + 113.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 157, + 113.33333333333334, + 454, + 125.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 506, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 131, + 145 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 131, + 133, + 136, + 142 + ], + "score": 0.8, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 131, + 260, + 145 + ], + "score": 1.0, + "content": "is the current attack iteration;", + "type": "text" + }, + { + "bbox": [ + 261, + 135, + 268, + 142 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 131, + 334, + 145 + ], + "score": 1.0, + "content": "is the step size;", + "type": "text" + }, + { + "bbox": [ + 335, + 132, + 344, + 142 + ], + "score": 0.6, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "is the projection operator that enforces", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 150, + 155 + ], + "score": 0.91, + "content": "| | v | | _ { \\infty } < \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 142, + 168, + 156 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 168, + 144, + 176, + 154 + ], + "score": 0.81, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 142, + 384, + 156 + ], + "score": 1.0, + "content": "is the cross-entropy loss function. The selection of", + "type": "text" + }, + { + "bbox": [ + 384, + 144, + 396, + 153 + ], + "score": 0.87, + "content": "\\mathcal { D } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "determines the type of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 153, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 153, + 237, + 168 + ], + "score": 1.0, + "content": "attack. For the standard variant:", + "type": "text" + }, + { + "bbox": [ + 237, + 154, + 277, + 166 + ], + "score": 0.92, + "content": "{ \\mathcal { D } } ^ { \\prime } = \\{ x \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 153, + 507, + 168 + ], + "score": 1.0, + "content": ", i.e., a single test-time sample. For the universal variant:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 163, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 107, + 165, + 140, + 175 + ], + "score": 0.87, + "content": "\\mathcal { D } ^ { \\prime } = \\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 163, + 474, + 178 + ], + "score": 1.0, + "content": ", i.e., the whole training set. For the class-universal variant against the target class", + "type": "text" + }, + { + "bbox": [ + 475, + 165, + 502, + 176 + ], + "score": 0.89, + "content": "i \\in M", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 163, + 507, + 178 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 174, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 213, + 188 + ], + "score": 0.91, + "content": "\\mathcal { D } ^ { \\prime } = \\{ ( x , y ) \\in \\mathcal { D } | y = i \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 174, + 372, + 190 + ], + "score": 1.0, + "content": ", i.e., the training set samples from the", + "type": "text" + }, + { + "bbox": [ + 372, + 176, + 385, + 186 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 174, + 446, + 190 + ], + "score": 1.0, + "content": "class. Finally,", + "type": "text" + }, + { + "bbox": [ + 447, + 177, + 453, + 188 + ], + "score": 0.82, + "content": "\\bar { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 174, + 506, + 190 + ], + "score": 1.0, + "content": "is the target", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 462, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 261, + 200 + ], + "score": 1.0, + "content": "label distribution our objective pushes", + "type": "text" + }, + { + "bbox": [ + 261, + 187, + 285, + 199 + ], + "score": 0.92, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 186, + 452, + 200 + ], + "score": 1.0, + "content": "towards. Next, we explain how we select", + "type": "text" + }, + { + "bbox": [ + 452, + 189, + 458, + 199 + ], + "score": 0.83, + "content": "\\bar { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 186, + 462, + 200 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 506, + 286 + ], + "lines": [ + { + "bbox": [ + 106, + 204, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 144, + 217 + ], + "score": 1.0, + "content": "Pushing", + "type": "text" + }, + { + "bbox": [ + 144, + 205, + 168, + 217 + ], + "score": 0.9, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 204, + 402, + 217 + ], + "score": 1.0, + "content": "Towards a Uniform Distribution. Despite including all", + "type": "text" + }, + { + "bbox": [ + 402, + 205, + 412, + 216 + ], + "score": 0.88, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 204, + 507, + 217 + ], + "score": 1.0, + "content": ", attacks such as PGD-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 214, + 504, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 191, + 230 + ], + "score": 1.0, + "content": "avg and PGD-max", + "type": "text" + }, + { + "bbox": [ + 191, + 216, + 206, + 226 + ], + "score": 0.27, + "content": "\\mathrm { H u }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 214, + 497, + 230 + ], + "score": 1.0, + "content": "et al., 2020) still fail to decrease efficacy. How these attacks select", + "type": "text" + }, + { + "bbox": [ + 497, + 217, + 504, + 227 + ], + "score": 0.77, + "content": "\\bar { y }", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 225, + 507, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 507, + 241 + ], + "score": 1.0, + "content": "reflects their goal of causing misclassifications and, therefore, they trigger errors in early-exits, i.e.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 235, + 510, + 258 + ], + "spans": [ + { + "bbox": [ + 107, + 237, + 229, + 253 + ], + "score": 0.9, + "content": "\\mathrm { a r g m a x } _ { j \\in M } \\bar { F } _ { i } ^ { ( j ) } ( x ) = \\bar { y } \\bar { \\ne y }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 235, + 510, + 258 + ], + "score": 1.0, + "content": ". However, as the early-exits still have high confidence, or low entropy,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 251, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 316, + 264 + ], + "score": 1.0, + "content": "the model still stops its computation early. We select", + "type": "text" + }, + { + "bbox": [ + 316, + 253, + 322, + 263 + ], + "score": 0.83, + "content": "\\bar { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 251, + 506, + 264 + ], + "score": 1.0, + "content": "as a uniform distribution over the class labels,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 260, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 260, + 124, + 277 + ], + "score": 1.0, + "content": "i.e.,", + "type": "text" + }, + { + "bbox": [ + 124, + 262, + 172, + 275 + ], + "score": 0.91, + "content": "\\bar { y } ^ { ( i ) } = 1 / m", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 260, + 248, + 277 + ], + "score": 1.0, + "content": ". This ensures that", + "type": "text" + }, + { + "bbox": [ + 249, + 263, + 280, + 275 + ], + "score": 0.92, + "content": "( x + v )", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 260, + 480, + 277 + ], + "score": 1.0, + "content": "bypasses common stopping criteria as a uniform", + "type": "text" + }, + { + "bbox": [ + 481, + 264, + 505, + 275 + ], + "score": 0.91, + "content": "F _ { i } ( x )", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 272, + 329, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 329, + 287 + ], + "score": 1.0, + "content": "has both the lowest confidence and the highest entropy.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 302, + 254, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 256, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 256, + 317 + ], + "score": 1.0, + "content": "5 EMPIRICAL EVALUATION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 326, + 506, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 232, + 339 + ], + "score": 1.0, + "content": "Here, we present the results for", + "type": "text" + }, + { + "bbox": [ + 232, + 327, + 246, + 338 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "DeepSloth against two SDNs—VGG-16 and MobileNet-based—", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 432, + 351 + ], + "score": 1.0, + "content": "and against the MSDNets. In the Appendix, we report the hyperparameters; the", + "type": "text" + }, + { + "bbox": [ + 432, + 338, + 442, + 349 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 337, + 460, + 351 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 461, + 338, + 471, + 349 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "attacks;", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 349, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 506, + 361 + ], + "score": 1.0, + "content": "the results on ResNet-56-based SDNs; the cost of the attacks; and some perturbed samples. Overall,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 174, + 372 + ], + "score": 1.0, + "content": "we observe that", + "type": "text" + }, + { + "bbox": [ + 174, + 360, + 188, + 371 + ], + "score": 0.9, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "-bounded perturbations are more effective for slowdowns. The optimization", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 369, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 237, + 384 + ], + "score": 1.0, + "content": "challenges might explain this, as", + "type": "text" + }, + { + "bbox": [ + 237, + 371, + 248, + 382 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 369, + 265, + 384 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 266, + 371, + 276, + 382 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 369, + 506, + 384 + ], + "score": 1.0, + "content": "attacks are usually harder to optimize (Carlini & Wagner,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "2017; Tramèr & Boneh, 2019). Unlike objectives for misclassifications, the objective for slowdowns", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 392, + 380, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 380, + 405 + ], + "score": 1.0, + "content": "involves multiple loss terms and optimizes over all the output logits.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 107, + 417, + 237, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 238, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 238, + 430 + ], + "score": 1.0, + "content": "5.1 WHITE-BOX SCENARIOS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 438, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 104, + 438, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 506, + 452 + ], + "score": 1.0, + "content": "Perturbations Eliminate Early-Exits. Table 2 (second segment) shows that the victim models", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 128, + 462 + ], + "score": 1.0, + "content": "have", + "type": "text" + }, + { + "bbox": [ + 128, + 450, + 145, + 460 + ], + "score": 0.86, + "content": "\\sim 0", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "efficacy on the samples perturbed by DeepSloth. Across the board, the attack makes the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "early-exit completely ineffective and force the victim models to forward all input samples till the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 345, + 484 + ], + "score": 1.0, + "content": "end. Further, DeepSloth also drops the victim’s accuracy by", + "type": "text" + }, + { + "bbox": [ + 345, + 472, + 379, + 483 + ], + "score": 0.88, + "content": "7 5 - 9 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 472, + 506, + 484 + ], + "score": 1.0, + "content": ", comparable to the PGD attack.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "These results give an answer to our main research question: the multi-exit mechanisms are vulnerable", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "and their benefits can be maliciously offset by adversarial input perturbations. In particular, as SDN", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "modification mitigates overthinking in standard, non-adaptive DNNs (Kaya et al., 2019), DeepSloth", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 515, + 491, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 491, + 528 + ], + "score": 1.0, + "content": "also leads SDN-based models to overthink on almost all samples by forcing extra computations.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 546 + ], + "score": 1.0, + "content": "Note that crafting a single perturbation requires multiple back-propagations through the model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 543, + 507, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 259, + 557 + ], + "score": 1.0, + "content": "and more floating points operations", + "type": "text" + }, + { + "bbox": [ + 259, + 544, + 295, + 555 + ], + "score": 0.58, + "content": "( F L O P s )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 543, + 507, + 557 + ], + "score": 1.0, + "content": "than the forward pass. The high cost of crafting,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 567 + ], + "score": 1.0, + "content": "relative to the computational damage to the victim, might make this vulnerability unattractive for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "the adversary. In the next sections, we highlight scenarios where this vulnerability might lead to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "practical exploitation. First, we show that in an IoT-like scenarios, the input transmission is a major", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 587, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 504, + 599 + ], + "score": 1.0, + "content": "bottleneck and DeepSloth can exploit it. Second, we evaluate universal DeepSloth attacks that enable", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 599, + 426, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 426, + 611 + ], + "score": 1.0, + "content": "the adversary to craft the perturbation only once and reuse it on multiple inputs.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 507, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 507, + 629 + ], + "score": 1.0, + "content": "Attacking an IoT Scenario. Many IoT scenarios, e.g., health monitoring for elderly (Park et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "2017), require collecting data from edge devices and making low-latency inferences on this data.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "However, complex deep learning models are impractical for low-power edge devices, such as an", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "Arduino, that are common in the IoT scenarios (Chen & Ran, 2019). For example, on standard", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 659, + 501, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 465, + 674 + ], + "score": 1.0, + "content": "hardware, an average inference takes MSDNet model on Tiny ImageNet 35M FLOPs and", + "type": "text" + }, + { + "bbox": [ + 466, + 660, + 496, + 671 + ], + "score": 0.83, + "content": "{ \\sim } 1 0 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 659, + 501, + 674 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "A potential solution is sending the inputs from the edge to a cloud model, which then returns the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "prediction. Even in our optimistic estimate with a nearby AWS EC2 instance, this back-and-forth", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 697, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 150, + 713 + ], + "score": 1.0, + "content": "introduces", + "type": "text" + }, + { + "bbox": [ + 150, + 699, + 182, + 709 + ], + "score": 0.77, + "content": "{ \\sim } 1 1 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 697, + 506, + 713 + ], + "score": 1.0, + "content": "latency per inference. Model partitioning alleviates this bottleneck by splitting a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "multi-exit model into two; deploying the small first part at the edge and the large second part at the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "cloud (De Coninck et al., 2015). The edge part sends an input only when its prediction does not meet", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "interline_equation", + "bbox": [ + 157, + 90, + 454, + 125 + ], + "lines": [ + { + "bbox": [ + 157, + 90, + 454, + 125 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 454, + 125 + ], + "score": 0.93, + "content": "v ^ { t + 1 } = \\prod _ { | | v | | _ { \\infty } < \\epsilon } \\left( v ^ { t } + \\alpha \\operatorname { s g n } \\left( { \\nabla } _ { v } \\sum _ { x \\in D ^ { \\prime } } \\sum _ { 0 < i < K } { \\mathcal { L } } \\left( F _ { i } \\left( x + v \\right) , \\bar { y } \\right) \\right) \\right)", + "type": "interline_equation", + "image_path": "86efb71bee2692f9ee3c0c9c37aeec2609a9c74c0132e9fd6eb8385d4f735434.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 157, + 90, + 454, + 101.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 157, + 101.66666666666667, + 454, + 113.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 157, + 113.33333333333334, + 454, + 125.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 506, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 131, + 145 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 131, + 133, + 136, + 142 + ], + "score": 0.8, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 131, + 260, + 145 + ], + "score": 1.0, + "content": "is the current attack iteration;", + "type": "text" + }, + { + "bbox": [ + 261, + 135, + 268, + 142 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 131, + 334, + 145 + ], + "score": 1.0, + "content": "is the step size;", + "type": "text" + }, + { + "bbox": [ + 335, + 132, + 344, + 142 + ], + "score": 0.6, + "content": "\\Pi", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "is the projection operator that enforces", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 150, + 155 + ], + "score": 0.91, + "content": "| | v | | _ { \\infty } < \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 142, + 168, + 156 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 168, + 144, + 176, + 154 + ], + "score": 0.81, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 142, + 384, + 156 + ], + "score": 1.0, + "content": "is the cross-entropy loss function. The selection of", + "type": "text" + }, + { + "bbox": [ + 384, + 144, + 396, + 153 + ], + "score": 0.87, + "content": "\\mathcal { D } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "determines the type of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 153, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 153, + 237, + 168 + ], + "score": 1.0, + "content": "attack. For the standard variant:", + "type": "text" + }, + { + "bbox": [ + 237, + 154, + 277, + 166 + ], + "score": 0.92, + "content": "{ \\mathcal { D } } ^ { \\prime } = \\{ x \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 153, + 507, + 168 + ], + "score": 1.0, + "content": ", i.e., a single test-time sample. For the universal variant:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 163, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 107, + 165, + 140, + 175 + ], + "score": 0.87, + "content": "\\mathcal { D } ^ { \\prime } = \\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 163, + 474, + 178 + ], + "score": 1.0, + "content": ", i.e., the whole training set. For the class-universal variant against the target class", + "type": "text" + }, + { + "bbox": [ + 475, + 165, + 502, + 176 + ], + "score": 0.89, + "content": "i \\in M", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 163, + 507, + 178 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 174, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 213, + 188 + ], + "score": 0.91, + "content": "\\mathcal { D } ^ { \\prime } = \\{ ( x , y ) \\in \\mathcal { D } | y = i \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 174, + 372, + 190 + ], + "score": 1.0, + "content": ", i.e., the training set samples from the", + "type": "text" + }, + { + "bbox": [ + 372, + 176, + 385, + 186 + ], + "score": 0.88, + "content": "i ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 174, + 446, + 190 + ], + "score": 1.0, + "content": "class. Finally,", + "type": "text" + }, + { + "bbox": [ + 447, + 177, + 453, + 188 + ], + "score": 0.82, + "content": "\\bar { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 174, + 506, + 190 + ], + "score": 1.0, + "content": "is the target", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 462, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 261, + 200 + ], + "score": 1.0, + "content": "label distribution our objective pushes", + "type": "text" + }, + { + "bbox": [ + 261, + 187, + 285, + 199 + ], + "score": 0.92, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 186, + 452, + 200 + ], + "score": 1.0, + "content": "towards. Next, we explain how we select", + "type": "text" + }, + { + "bbox": [ + 452, + 189, + 458, + 199 + ], + "score": 0.83, + "content": "\\bar { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 186, + 462, + 200 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 131, + 507, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 506, + 286 + ], + "lines": [ + { + "bbox": [ + 106, + 204, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 144, + 217 + ], + "score": 1.0, + "content": "Pushing", + "type": "text" + }, + { + "bbox": [ + 144, + 205, + 168, + 217 + ], + "score": 0.9, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 204, + 402, + 217 + ], + "score": 1.0, + "content": "Towards a Uniform Distribution. Despite including all", + "type": "text" + }, + { + "bbox": [ + 402, + 205, + 412, + 216 + ], + "score": 0.88, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 204, + 507, + 217 + ], + "score": 1.0, + "content": ", attacks such as PGD-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 214, + 504, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 191, + 230 + ], + "score": 1.0, + "content": "avg and PGD-max", + "type": "text" + }, + { + "bbox": [ + 191, + 216, + 206, + 226 + ], + "score": 0.27, + "content": "\\mathrm { H u }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 214, + 497, + 230 + ], + "score": 1.0, + "content": "et al., 2020) still fail to decrease efficacy. How these attacks select", + "type": "text" + }, + { + "bbox": [ + 497, + 217, + 504, + 227 + ], + "score": 0.77, + "content": "\\bar { y }", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 225, + 507, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 507, + 241 + ], + "score": 1.0, + "content": "reflects their goal of causing misclassifications and, therefore, they trigger errors in early-exits, i.e.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 235, + 510, + 258 + ], + "spans": [ + { + "bbox": [ + 107, + 237, + 229, + 253 + ], + "score": 0.9, + "content": "\\mathrm { a r g m a x } _ { j \\in M } \\bar { F } _ { i } ^ { ( j ) } ( x ) = \\bar { y } \\bar { \\ne y }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 235, + 510, + 258 + ], + "score": 1.0, + "content": ". However, as the early-exits still have high confidence, or low entropy,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 251, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 316, + 264 + ], + "score": 1.0, + "content": "the model still stops its computation early. We select", + "type": "text" + }, + { + "bbox": [ + 316, + 253, + 322, + 263 + ], + "score": 0.83, + "content": "\\bar { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 251, + 506, + 264 + ], + "score": 1.0, + "content": "as a uniform distribution over the class labels,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 260, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 260, + 124, + 277 + ], + "score": 1.0, + "content": "i.e.,", + "type": "text" + }, + { + "bbox": [ + 124, + 262, + 172, + 275 + ], + "score": 0.91, + "content": "\\bar { y } ^ { ( i ) } = 1 / m", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 260, + 248, + 277 + ], + "score": 1.0, + "content": ". This ensures that", + "type": "text" + }, + { + "bbox": [ + 249, + 263, + 280, + 275 + ], + "score": 0.92, + "content": "( x + v )", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 260, + 480, + 277 + ], + "score": 1.0, + "content": "bypasses common stopping criteria as a uniform", + "type": "text" + }, + { + "bbox": [ + 481, + 264, + 505, + 275 + ], + "score": 0.91, + "content": "F _ { i } ( x )", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 272, + 329, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 329, + 287 + ], + "score": 1.0, + "content": "has both the lowest confidence and the highest entropy.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 104, + 204, + 510, + 287 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 302, + 254, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 256, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 256, + 317 + ], + "score": 1.0, + "content": "5 EMPIRICAL EVALUATION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 326, + 506, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 232, + 339 + ], + "score": 1.0, + "content": "Here, we present the results for", + "type": "text" + }, + { + "bbox": [ + 232, + 327, + 246, + 338 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "DeepSloth against two SDNs—VGG-16 and MobileNet-based—", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 432, + 351 + ], + "score": 1.0, + "content": "and against the MSDNets. In the Appendix, we report the hyperparameters; the", + "type": "text" + }, + { + "bbox": [ + 432, + 338, + 442, + 349 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 337, + 460, + 351 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 461, + 338, + 471, + 349 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "attacks;", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 349, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 506, + 361 + ], + "score": 1.0, + "content": "the results on ResNet-56-based SDNs; the cost of the attacks; and some perturbed samples. Overall,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 174, + 372 + ], + "score": 1.0, + "content": "we observe that", + "type": "text" + }, + { + "bbox": [ + 174, + 360, + 188, + 371 + ], + "score": 0.9, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "-bounded perturbations are more effective for slowdowns. The optimization", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 369, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 237, + 384 + ], + "score": 1.0, + "content": "challenges might explain this, as", + "type": "text" + }, + { + "bbox": [ + 237, + 371, + 248, + 382 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 369, + 265, + 384 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 266, + 371, + 276, + 382 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 369, + 506, + 384 + ], + "score": 1.0, + "content": "attacks are usually harder to optimize (Carlini & Wagner,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "2017; Tramèr & Boneh, 2019). Unlike objectives for misclassifications, the objective for slowdowns", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 392, + 380, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 380, + 405 + ], + "score": 1.0, + "content": "involves multiple loss terms and optimizes over all the output logits.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 326, + 506, + 405 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 417, + 237, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 238, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 238, + 430 + ], + "score": 1.0, + "content": "5.1 WHITE-BOX SCENARIOS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 438, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 104, + 438, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 506, + 452 + ], + "score": 1.0, + "content": "Perturbations Eliminate Early-Exits. Table 2 (second segment) shows that the victim models", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 128, + 462 + ], + "score": 1.0, + "content": "have", + "type": "text" + }, + { + "bbox": [ + 128, + 450, + 145, + 460 + ], + "score": 0.86, + "content": "\\sim 0", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "efficacy on the samples perturbed by DeepSloth. Across the board, the attack makes the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "early-exit completely ineffective and force the victim models to forward all input samples till the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 345, + 484 + ], + "score": 1.0, + "content": "end. Further, DeepSloth also drops the victim’s accuracy by", + "type": "text" + }, + { + "bbox": [ + 345, + 472, + 379, + 483 + ], + "score": 0.88, + "content": "7 5 - 9 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 472, + 506, + 484 + ], + "score": 1.0, + "content": ", comparable to the PGD attack.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "These results give an answer to our main research question: the multi-exit mechanisms are vulnerable", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "and their benefits can be maliciously offset by adversarial input perturbations. In particular, as SDN", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "modification mitigates overthinking in standard, non-adaptive DNNs (Kaya et al., 2019), DeepSloth", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 515, + 491, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 491, + 528 + ], + "score": 1.0, + "content": "also leads SDN-based models to overthink on almost all samples by forcing extra computations.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 438, + 506, + 528 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 546 + ], + "score": 1.0, + "content": "Note that crafting a single perturbation requires multiple back-propagations through the model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 543, + 507, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 259, + 557 + ], + "score": 1.0, + "content": "and more floating points operations", + "type": "text" + }, + { + "bbox": [ + 259, + 544, + 295, + 555 + ], + "score": 0.58, + "content": "( F L O P s )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 543, + 507, + 557 + ], + "score": 1.0, + "content": "than the forward pass. The high cost of crafting,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 567 + ], + "score": 1.0, + "content": "relative to the computational damage to the victim, might make this vulnerability unattractive for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "the adversary. In the next sections, we highlight scenarios where this vulnerability might lead to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "practical exploitation. First, we show that in an IoT-like scenarios, the input transmission is a major", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 587, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 504, + 599 + ], + "score": 1.0, + "content": "bottleneck and DeepSloth can exploit it. Second, we evaluate universal DeepSloth attacks that enable", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 599, + 426, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 426, + 611 + ], + "score": 1.0, + "content": "the adversary to craft the perturbation only once and reuse it on multiple inputs.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 531, + 507, + 611 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 507, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 507, + 629 + ], + "score": 1.0, + "content": "Attacking an IoT Scenario. Many IoT scenarios, e.g., health monitoring for elderly (Park et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "2017), require collecting data from edge devices and making low-latency inferences on this data.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "However, complex deep learning models are impractical for low-power edge devices, such as an", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "Arduino, that are common in the IoT scenarios (Chen & Ran, 2019). For example, on standard", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 659, + 501, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 465, + 674 + ], + "score": 1.0, + "content": "hardware, an average inference takes MSDNet model on Tiny ImageNet 35M FLOPs and", + "type": "text" + }, + { + "bbox": [ + 466, + 660, + 496, + 671 + ], + "score": 0.83, + "content": "{ \\sim } 1 0 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 659, + 501, + 674 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 615, + 507, + 674 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "A potential solution is sending the inputs from the edge to a cloud model, which then returns the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "prediction. Even in our optimistic estimate with a nearby AWS EC2 instance, this back-and-forth", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 697, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 150, + 713 + ], + "score": 1.0, + "content": "introduces", + "type": "text" + }, + { + "bbox": [ + 150, + 699, + 182, + 709 + ], + "score": 0.77, + "content": "{ \\sim } 1 1 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 697, + 506, + 713 + ], + "score": 1.0, + "content": "latency per inference. Model partitioning alleviates this bottleneck by splitting a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "multi-exit model into two; deploying the small first part at the edge and the large second part at the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "cloud (De Coninck et al., 2015). The edge part sends an input only when its prediction does not meet", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 413, + 340 + ], + "score": 1.0, + "content": "the stopping criteria. For example, the first early-exit of MSDNets sends only", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 413, + 327, + 428, + 338 + ], + "score": 0.89, + "content": "5 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 428, + 326, + 445, + 340 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 445, + 327, + 465, + 338 + ], + "score": 0.88, + "content": "67 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 465, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "of all test", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 337, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 351 + ], + "score": 1.0, + "content": "samples, on CIFAR-10 and Tiny ImageNet, respectively. This leads to a lower average latency per", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 348, + 370, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 245, + 361 + ], + "score": 1.0, + "content": "inference, i.e., from 11ms down to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 246, + 349, + 272, + 360 + ], + "score": 0.56, + "content": "0 . 5 \\mathrm { m s }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 272, + 348, + 289, + 361 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 289, + 349, + 315, + 360 + ], + "score": 0.44, + "content": "7 . 4 \\mathrm { m s }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 315, + 348, + 370, + 361 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 677, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 121, + 504, + 305 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 81, + 505, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 79, + 507, + 94 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 217, + 94 + ], + "score": 1.0, + "content": "Table 2: The effectiveness of", + "type": "text" + }, + { + "bbox": [ + 218, + 81, + 230, + 91 + ], + "score": 0.88, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 79, + 282, + 94 + ], + "score": 1.0, + "content": "DeepSloth. ‘", + "type": "text" + }, + { + "bbox": [ + 282, + 81, + 329, + 91 + ], + "score": 0.73, + "content": "\\mathrm { 2 A D { < } } 5 , 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 79, + 507, + 94 + ], + "score": 1.0, + "content": "’ columns list the results in each early-exit setting.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 506, + 102 + ], + "score": 1.0, + "content": "Each entry includes the model’s efficacy (left) and accuracy (right). The class-universal attack’s results are an", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 346, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 279, + 112 + ], + "score": 1.0, + "content": "average of 10 classes. ‘TI’: Tiny ImageNet and", + "type": "text" + }, + { + "bbox": [ + 279, + 101, + 300, + 111 + ], + "score": 0.44, + "content": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 101, + 346, + 112 + ], + "score": 1.0, + "content": ": CIFAR-10.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 106, + 121, + 504, + 305 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 121, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 504, + 305 + ], + "score": 0.983, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 / 85%0.89 / 85%0.77 /88%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 /57%0.51 / 52%0.42 /57%0.59 /51%
DEEPSLOTH
C100.06 / 17%0.06 / 17%0.01 /13%0.04 / 16%0.01 /12%0.06 /16%
TI0.06 /7%0.06 /7%0.00 /2%0.01/2%0.02 / 6%0.04 / 6%
UNIVERSAL DEEPSLOTH
C100.85 / 65%0.85 /65%0.62 / 65%0.86 / 60%0.73 / 61%0.90 / 59%
TI0.58 / 46%0.81 /41%0.31 / 47%0.44 / 44%0.33 / 47%0.51 / 43%
CLASS-UNIVERSAL DEEPSLOTH
C100.82/32%0.82 /32%0.47 /35%0.78 /33%0.60 /30%0.85 /27%
TI0.41 / 21%0.71 /17%0.20 /28%0.33 /27%0.21 /27%0.38 /25%
", + "type": "table", + "image_path": "15f95d76d72cb36008898a8f228ac096fa222a80b0bd4f1332610847e2a19f34.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 106, + 121, + 504, + 182.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 182.33333333333334, + 504, + 243.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 243.66666666666669, + 504, + 305.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 327, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 413, + 340 + ], + "score": 1.0, + "content": "the stopping criteria. For example, the first early-exit of MSDNets sends only", + "type": "text" + }, + { + "bbox": [ + 413, + 327, + 428, + 338 + ], + "score": 0.89, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 326, + 445, + 340 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 445, + 327, + 465, + 338 + ], + "score": 0.88, + "content": "67 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "of all test", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 337, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 351 + ], + "score": 1.0, + "content": "samples, on CIFAR-10 and Tiny ImageNet, respectively. This leads to a lower average latency per", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 348, + 370, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 245, + 361 + ], + "score": 1.0, + "content": "inference, i.e., from 11ms down to", + "type": "text" + }, + { + "bbox": [ + 246, + 349, + 272, + 360 + ], + "score": 0.56, + "content": "0 . 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 348, + 289, + 361 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 289, + 349, + 315, + 360 + ], + "score": 0.44, + "content": "7 . 4 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 348, + 370, + 361 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "The adversary we study uses DeepSloth perturbations to force the edge part to send all the input", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "samples to the cloud. For the victim, we deploy MSDNet models that we split into two parts at their", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 393, + 401 + ], + "score": 1.0, + "content": "first exit point. Targeting the first part with DeepSloth forces it to send", + "type": "text" + }, + { + "bbox": [ + 393, + 388, + 413, + 398 + ], + "score": 0.88, + "content": "96 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 388, + 431, + 401 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 388, + 464, + 398 + ], + "score": 0.9, + "content": "9 9 . 9 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "of all test", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 394, + 411 + ], + "score": 1.0, + "content": "samples to the second part. This increases average inference latency to", + "type": "text" + }, + { + "bbox": [ + 395, + 399, + 426, + 409 + ], + "score": 0.82, + "content": "{ \\sim } 1 1 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "and invalidates the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 410, + 423 + ], + "score": 1.0, + "content": "benefits of model partitioning. In this scenario, perturbing each sample takes", + "type": "text" + }, + { + "bbox": [ + 410, + 410, + 435, + 420 + ], + "score": 0.83, + "content": "{ \\sim } 2 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "on a Tesla V-100", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 420, + 475, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 316, + 433 + ], + "score": 1.0, + "content": "GPU, i.e., the time adversary spends is amplified by", + "type": "text" + }, + { + "bbox": [ + 316, + 421, + 346, + 431 + ], + "score": 0.86, + "content": "1 . 5 – 5 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 420, + 475, + 433 + ], + "score": 1.0, + "content": "as the victim’s latency increase.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "Reusable Universal Perturbations. The universal attacks, although limited, are a practical as the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "adversary can reuse the same perturbation indefinitely to cause minor slowdowns. Table 2 (third", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 313, + 473 + ], + "score": 1.0, + "content": "segment) shows that they decrease the efficacy by", + "type": "text" + }, + { + "bbox": [ + 313, + 460, + 344, + 471 + ], + "score": 0.87, + "content": "3- 2 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 460, + 430, + 473 + ], + "score": 1.0, + "content": "and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 431, + 460, + 465, + 471 + ], + "score": 0.89, + "content": "1 5 \\mathrm { - } 2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 460, + 505, + 473 + ], + "score": 1.0, + "content": ", over the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 371, + 484 + ], + "score": 1.0, + "content": "baselines. Having a less conservative early-exit strategy, e.g., RAD", + "type": "text" + }, + { + "bbox": [ + 371, + 471, + 397, + 482 + ], + "score": 0.63, + "content": "{ < } 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 471, + 506, + 484 + ], + "score": 1.0, + "content": ", increases the resilience to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 480, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 424, + 496 + ], + "score": 1.0, + "content": "the attack at the cost of accuracy. Further, MSDNets are fairly resilient with only", + "type": "text" + }, + { + "bbox": [ + 424, + 482, + 449, + 493 + ], + "score": 0.87, + "content": "3- 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 480, + 506, + 496 + ], + "score": 1.0, + "content": "efficacy drop;", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 270, + 506 + ], + "score": 1.0, + "content": "whereas SDNs are more vulnerable with", + "type": "text" + }, + { + "bbox": [ + 270, + 493, + 304, + 504 + ], + "score": 0.89, + "content": "12 \\mathrm { - } 2 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "drop. The attack is also slightly more effective on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "the more complex task, Tiny ImageNet, as the early-exits become easier to bypass. Using random", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 514, + 483, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 203, + 527 + ], + "score": 1.0, + "content": "noise as a baseline, i.e.,", + "type": "text" + }, + { + "bbox": [ + 203, + 514, + 262, + 527 + ], + "score": 0.93, + "content": "v \\sim U ^ { d } ( - \\epsilon , \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 514, + 457, + 527 + ], + "score": 1.0, + "content": ", we find that at most it decreases the efficacy by", + "type": "text" + }, + { + "bbox": [ + 457, + 515, + 479, + 525 + ], + "score": 0.88, + "content": "{ \\sim } 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 514, + 483, + 527 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 543 + ], + "score": 1.0, + "content": "In the universal attack, we observe a phenomenon: it pushes the samples towards a small subset of all", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 543, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 192, + 554 + ], + "score": 1.0, + "content": "classes. For example,", + "type": "text" + }, + { + "bbox": [ + 192, + 543, + 219, + 554 + ], + "score": 0.89, + "content": "{ \\sim } 1 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 543, + 506, + 554 + ], + "score": 1.0, + "content": "of the perturbed samples are classified into the ’bird’ class of CIFAR-10;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 552, + 507, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 140, + 568 + ], + "score": 1.0, + "content": "up from", + "type": "text" + }, + { + "bbox": [ + 140, + 554, + 167, + 564 + ], + "score": 0.9, + "content": "{ \\sim } 1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 552, + 507, + 568 + ], + "score": 1.0, + "content": "for the clean samples. Considering certain classes are distant in the feature space, e.g.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 565, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 576 + ], + "score": 1.0, + "content": "’truck’ and ’bird’; we expect the class-universal variant to be more effective. The results in Table 2", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "score": 1.0, + "content": "(fourth segment) confirm our intuition. We see that this attack decreases the baseline efficacy by", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 136, + 597 + ], + "score": 0.88, + "content": "8 - 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 586, + 217, + 599 + ], + "score": 1.0, + "content": "and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 218, + 587, + 252, + 597 + ], + "score": 0.88, + "content": "5 0 \\mathrm { - } 6 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 586, + 506, + 599 + ], + "score": 1.0, + "content": ". We report the average results across multiple classes; however,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 598, + 401, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 401, + 609 + ], + "score": 1.0, + "content": "we observe that certain classes are slightly more vulnerable to this attack.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "score": 1.0, + "content": "Feature Visualization of DeepSloth. In Figure 3, to shed light on how DeepSloth differs from prior", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "attacks, e.g., PGD and PGD-avg, we visualize a model’s hidden block (layer) features on the original", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "and perturbed test-time samples. We observe that in an earlier block (left panel), DeepSloth seems to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "disrupt the original features slightly more than the PGD attacks. Leaving earlier representations intact", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "prevents PGDs from bypassing the early-exits. The behaviors of the attacks diverge in the middle", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 682 + ], + "score": 1.0, + "content": "blocks (middle panel). Here, DeepSloth features remain closer to the original features than prior", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 693 + ], + "score": 1.0, + "content": "attacks. The significant disruption of prior attacks leads to high-confidence misclassifications and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 692, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 506, + 704 + ], + "score": 1.0, + "content": "fails to bypass early-exits. In the later block (right panel), we see that the divergent behavior persists.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Preserving or Hurting the Accuracy with DeepSloth. Here, we aim to answer whether DeepSloth", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "can be applied when the adversary explicitly aims to cause or prevent misclassifications, while still", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 121, + 504, + 305 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 81, + 505, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 79, + 507, + 94 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 217, + 94 + ], + "score": 1.0, + "content": "Table 2: The effectiveness of", + "type": "text" + }, + { + "bbox": [ + 218, + 81, + 230, + 91 + ], + "score": 0.88, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 79, + 282, + 94 + ], + "score": 1.0, + "content": "DeepSloth. ‘", + "type": "text" + }, + { + "bbox": [ + 282, + 81, + 329, + 91 + ], + "score": 0.73, + "content": "\\mathrm { 2 A D { < } } 5 , 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 79, + 507, + 94 + ], + "score": 1.0, + "content": "’ columns list the results in each early-exit setting.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 506, + 102 + ], + "score": 1.0, + "content": "Each entry includes the model’s efficacy (left) and accuracy (right). The class-universal attack’s results are an", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 346, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 279, + 112 + ], + "score": 1.0, + "content": "average of 10 classes. ‘TI’: Tiny ImageNet and", + "type": "text" + }, + { + "bbox": [ + 279, + 101, + 300, + 111 + ], + "score": 0.44, + "content": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 101, + 346, + 112 + ], + "score": 1.0, + "content": ": CIFAR-10.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 106, + 121, + 504, + 305 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 121, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 504, + 305 + ], + "score": 0.983, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 / 85%0.89 / 85%0.77 /88%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 /57%0.51 / 52%0.42 /57%0.59 /51%
DEEPSLOTH
C100.06 / 17%0.06 / 17%0.01 /13%0.04 / 16%0.01 /12%0.06 /16%
TI0.06 /7%0.06 /7%0.00 /2%0.01/2%0.02 / 6%0.04 / 6%
UNIVERSAL DEEPSLOTH
C100.85 / 65%0.85 /65%0.62 / 65%0.86 / 60%0.73 / 61%0.90 / 59%
TI0.58 / 46%0.81 /41%0.31 / 47%0.44 / 44%0.33 / 47%0.51 / 43%
CLASS-UNIVERSAL DEEPSLOTH
C100.82/32%0.82 /32%0.47 /35%0.78 /33%0.60 /30%0.85 /27%
TI0.41 / 21%0.71 /17%0.20 /28%0.33 /27%0.21 /27%0.38 /25%
", + "type": "table", + "image_path": "15f95d76d72cb36008898a8f228ac096fa222a80b0bd4f1332610847e2a19f34.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 106, + 121, + 504, + 182.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 182.33333333333334, + 504, + 243.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 243.66666666666669, + 504, + 305.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 327, + 505, + 360 + ], + "lines": [], + "index": 7, + "bbox_fs": [ + 105, + 326, + 506, + 361 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "The adversary we study uses DeepSloth perturbations to force the edge part to send all the input", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "samples to the cloud. For the victim, we deploy MSDNet models that we split into two parts at their", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 393, + 401 + ], + "score": 1.0, + "content": "first exit point. Targeting the first part with DeepSloth forces it to send", + "type": "text" + }, + { + "bbox": [ + 393, + 388, + 413, + 398 + ], + "score": 0.88, + "content": "96 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 388, + 431, + 401 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 388, + 464, + 398 + ], + "score": 0.9, + "content": "9 9 . 9 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "of all test", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 394, + 411 + ], + "score": 1.0, + "content": "samples to the second part. This increases average inference latency to", + "type": "text" + }, + { + "bbox": [ + 395, + 399, + 426, + 409 + ], + "score": 0.82, + "content": "{ \\sim } 1 1 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "and invalidates the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 410, + 423 + ], + "score": 1.0, + "content": "benefits of model partitioning. In this scenario, perturbing each sample takes", + "type": "text" + }, + { + "bbox": [ + 410, + 410, + 435, + 420 + ], + "score": 0.83, + "content": "{ \\sim } 2 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "on a Tesla V-100", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 420, + 475, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 316, + 433 + ], + "score": 1.0, + "content": "GPU, i.e., the time adversary spends is amplified by", + "type": "text" + }, + { + "bbox": [ + 316, + 421, + 346, + 431 + ], + "score": 0.86, + "content": "1 . 5 – 5 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 420, + 475, + 433 + ], + "score": 1.0, + "content": "as the victim’s latency increase.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 365, + 506, + 433 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "Reusable Universal Perturbations. The universal attacks, although limited, are a practical as the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "adversary can reuse the same perturbation indefinitely to cause minor slowdowns. Table 2 (third", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 313, + 473 + ], + "score": 1.0, + "content": "segment) shows that they decrease the efficacy by", + "type": "text" + }, + { + "bbox": [ + 313, + 460, + 344, + 471 + ], + "score": 0.87, + "content": "3- 2 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 460, + 430, + 473 + ], + "score": 1.0, + "content": "and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 431, + 460, + 465, + 471 + ], + "score": 0.89, + "content": "1 5 \\mathrm { - } 2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 460, + 505, + 473 + ], + "score": 1.0, + "content": ", over the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 371, + 484 + ], + "score": 1.0, + "content": "baselines. Having a less conservative early-exit strategy, e.g., RAD", + "type": "text" + }, + { + "bbox": [ + 371, + 471, + 397, + 482 + ], + "score": 0.63, + "content": "{ < } 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 471, + 506, + 484 + ], + "score": 1.0, + "content": ", increases the resilience to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 480, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 424, + 496 + ], + "score": 1.0, + "content": "the attack at the cost of accuracy. Further, MSDNets are fairly resilient with only", + "type": "text" + }, + { + "bbox": [ + 424, + 482, + 449, + 493 + ], + "score": 0.87, + "content": "3- 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 480, + 506, + 496 + ], + "score": 1.0, + "content": "efficacy drop;", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 270, + 506 + ], + "score": 1.0, + "content": "whereas SDNs are more vulnerable with", + "type": "text" + }, + { + "bbox": [ + 270, + 493, + 304, + 504 + ], + "score": 0.89, + "content": "12 \\mathrm { - } 2 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "drop. The attack is also slightly more effective on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "the more complex task, Tiny ImageNet, as the early-exits become easier to bypass. Using random", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 514, + 483, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 203, + 527 + ], + "score": 1.0, + "content": "noise as a baseline, i.e.,", + "type": "text" + }, + { + "bbox": [ + 203, + 514, + 262, + 527 + ], + "score": 0.93, + "content": "v \\sim U ^ { d } ( - \\epsilon , \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 514, + 457, + 527 + ], + "score": 1.0, + "content": ", we find that at most it decreases the efficacy by", + "type": "text" + }, + { + "bbox": [ + 457, + 515, + 479, + 525 + ], + "score": 0.88, + "content": "{ \\sim } 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 514, + 483, + 527 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 438, + 506, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 543 + ], + "score": 1.0, + "content": "In the universal attack, we observe a phenomenon: it pushes the samples towards a small subset of all", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 543, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 192, + 554 + ], + "score": 1.0, + "content": "classes. For example,", + "type": "text" + }, + { + "bbox": [ + 192, + 543, + 219, + 554 + ], + "score": 0.89, + "content": "{ \\sim } 1 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 543, + 506, + 554 + ], + "score": 1.0, + "content": "of the perturbed samples are classified into the ’bird’ class of CIFAR-10;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 552, + 507, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 140, + 568 + ], + "score": 1.0, + "content": "up from", + "type": "text" + }, + { + "bbox": [ + 140, + 554, + 167, + 564 + ], + "score": 0.9, + "content": "{ \\sim } 1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 552, + 507, + 568 + ], + "score": 1.0, + "content": "for the clean samples. Considering certain classes are distant in the feature space, e.g.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 565, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 576 + ], + "score": 1.0, + "content": "’truck’ and ’bird’; we expect the class-universal variant to be more effective. The results in Table 2", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "score": 1.0, + "content": "(fourth segment) confirm our intuition. We see that this attack decreases the baseline efficacy by", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 136, + 597 + ], + "score": 0.88, + "content": "8 - 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 586, + 217, + 599 + ], + "score": 1.0, + "content": "and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 218, + 587, + 252, + 597 + ], + "score": 0.88, + "content": "5 0 \\mathrm { - } 6 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 586, + 506, + 599 + ], + "score": 1.0, + "content": ". We report the average results across multiple classes; however,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 598, + 401, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 401, + 609 + ], + "score": 1.0, + "content": "we observe that certain classes are slightly more vulnerable to this attack.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 531, + 507, + 609 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "score": 1.0, + "content": "Feature Visualization of DeepSloth. In Figure 3, to shed light on how DeepSloth differs from prior", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "attacks, e.g., PGD and PGD-avg, we visualize a model’s hidden block (layer) features on the original", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "and perturbed test-time samples. We observe that in an earlier block (left panel), DeepSloth seems to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "disrupt the original features slightly more than the PGD attacks. Leaving earlier representations intact", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "prevents PGDs from bypassing the early-exits. The behaviors of the attacks diverge in the middle", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 682 + ], + "score": 1.0, + "content": "blocks (middle panel). Here, DeepSloth features remain closer to the original features than prior", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 693 + ], + "score": 1.0, + "content": "attacks. The significant disruption of prior attacks leads to high-confidence misclassifications and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 692, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 506, + 704 + ], + "score": 1.0, + "content": "fails to bypass early-exits. In the later block (right panel), we see that the divergent behavior persists.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 614, + 506, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "Preserving or Hurting the Accuracy with DeepSloth. Here, we aim to answer whether DeepSloth", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "can be applied when the adversary explicitly aims to cause or prevent misclassifications, while still", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "causing slowdowns. Our main threat model has no explicit goal regarding misclassifications that hurt", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 234, + 504, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 504, + 246 + ], + "score": 1.0, + "content": "the user of the model, i.e., who consumes the output of the model. Whereas, slowdowns additionally", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 1.0, + "content": "hurt the executor or the owner of the model through the computations and latency increased at", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 256, + 504, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 256, + 504, + 267 + ], + "score": 1.0, + "content": "the cloud providers. In some ML-in-the-cloud scenarios, where these two are different actors, the", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "score": 1.0, + "content": "adversary might aim to target only the executor or both the executor and the user. To this end, we", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 262, + 291 + ], + "score": 1.0, + "content": "modify our objective function to push", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 263, + 278, + 287, + 289 + ], + "score": 0.93, + "content": "F _ { i } ( x )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 287, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "towards a slightly non-uniform distribution, favoring", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "either the ground truth label for preventing misclassifications or a wrong label for causing them. We", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 398, + 313 + ], + "score": 1.0, + "content": "test this idea on our VGG-16-based SDN model on CIFAR-10 in RAD", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 398, + 300, + 420, + 310 + ], + "score": 0.58, + "content": "{ < } 5 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 420, + 299, + 506, + 313 + ], + "score": 1.0, + "content": "setting. We see that", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 311, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 296, + 324 + ], + "score": 1.0, + "content": "DeepSloth for preserving the accuracy leads to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 297, + 311, + 316, + 321 + ], + "score": 0.86, + "content": "81 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 316, + 311, + 506, + 324 + ], + "score": 1.0, + "content": "accuracy with 0.02 efficacy and DeepSloth for", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 226, + 334 + ], + "score": 1.0, + "content": "hurting the accuracy leads to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 226, + 322, + 241, + 332 + ], + "score": 0.86, + "content": "4 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 241, + 322, + 484, + 334 + ], + "score": 1.0, + "content": "accuracy with 0.01 efficacy—the original DeepSloth led to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 485, + 322, + 505, + 332 + ], + "score": 0.87, + "content": "13 \\%", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "score": 1.0, + "content": "accuracy with 0.01 efficacy. These results show the flexibility of DeepSloth and how it could be", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 343, + 282, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 282, + 357 + ], + "score": 1.0, + "content": "modified depending on the attacker’s goals.", + "type": "text", + "cross_page": true + } + ], + "index": 16 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 81, + 502, + 174 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 81, + 502, + 174 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 81, + 502, + 174 + ], + "spans": [ + { + "bbox": [ + 109, + 81, + 502, + 174 + ], + "score": 0.962, + "type": "image", + "image_path": "eb4db376fcea7527d2b8a56c5918847e8a4e1a6ae201be0033e35f93b1112bd1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 81, + 502, + 112.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 112.0, + 502, + 143.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 143.0, + 502, + 174.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 177, + 504, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "Figure 3: Visualising features against attacks using UMAP. VGG-16’s 3rd (left), 8th (middle), and 14th", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 186, + 436, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 436, + 198 + ], + "score": 1.0, + "content": "(right) hidden block features on CIFAR-10’s ’dog’ class (Best viewed in color, zoomed in).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 223, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "causing slowdowns. Our main threat model has no explicit goal regarding misclassifications that hurt", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 234, + 504, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 504, + 246 + ], + "score": 1.0, + "content": "the user of the model, i.e., who consumes the output of the model. Whereas, slowdowns additionally", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 1.0, + "content": "hurt the executor or the owner of the model through the computations and latency increased at", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 256, + 504, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 256, + 504, + 267 + ], + "score": 1.0, + "content": "the cloud providers. In some ML-in-the-cloud scenarios, where these two are different actors, the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "score": 1.0, + "content": "adversary might aim to target only the executor or both the executor and the user. To this end, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 262, + 291 + ], + "score": 1.0, + "content": "modify our objective function to push", + "type": "text" + }, + { + "bbox": [ + 263, + 278, + 287, + 289 + ], + "score": 0.93, + "content": "F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "towards a slightly non-uniform distribution, favoring", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "either the ground truth label for preventing misclassifications or a wrong label for causing them. We", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 299, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 398, + 313 + ], + "score": 1.0, + "content": "test this idea on our VGG-16-based SDN model on CIFAR-10 in RAD", + "type": "text" + }, + { + "bbox": [ + 398, + 300, + 420, + 310 + ], + "score": 0.58, + "content": "{ < } 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 299, + 506, + 313 + ], + "score": 1.0, + "content": "setting. We see that", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 311, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 296, + 324 + ], + "score": 1.0, + "content": "DeepSloth for preserving the accuracy leads to", + "type": "text" + }, + { + "bbox": [ + 297, + 311, + 316, + 321 + ], + "score": 0.86, + "content": "81 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 311, + 506, + 324 + ], + "score": 1.0, + "content": "accuracy with 0.02 efficacy and DeepSloth for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 226, + 334 + ], + "score": 1.0, + "content": "hurting the accuracy leads to", + "type": "text" + }, + { + "bbox": [ + 226, + 322, + 241, + 332 + ], + "score": 0.86, + "content": "4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 322, + 484, + 334 + ], + "score": 1.0, + "content": "accuracy with 0.01 efficacy—the original DeepSloth led to", + "type": "text" + }, + { + "bbox": [ + 485, + 322, + 505, + 332 + ], + "score": 0.87, + "content": "13 \\%", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "score": 1.0, + "content": "accuracy with 0.01 efficacy. These results show the flexibility of DeepSloth and how it could be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 343, + 282, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 282, + 357 + ], + "score": 1.0, + "content": "modified depending on the attacker’s goals.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 108, + 379, + 432, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 378, + 434, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 434, + 392 + ], + "score": 1.0, + "content": "5.2 TOWARDS A BLACK-BOX ATTACK: TRANSFERABILITY OF DEEPSLOTH", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 404, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 106, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "Transferability of adversarial examples imply that they can still hurt a model that they were not", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "crafted on (Tramèr et al., 2017a; Liu et al., 2017). Even though white-box attacks are important to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 426, + 507, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 507, + 439 + ], + "score": 1.0, + "content": "expose the vulnerability, black-box attacks, by requiring fewer assumptions, are more practical. Here,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "score": 1.0, + "content": "on four distinct scenarios, we investigate whether DeepSloth is transferable. Based on the scenario’s", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "constraints, we (i) train a surrogate model; (ii) craft the DeepSloth samples on it; and (iii) use these", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 459, + 434, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 408, + 471 + ], + "score": 1.0, + "content": "samples on the victim. We run these experiments on CIFAR-10 in the RAD", + "type": "text" + }, + { + "bbox": [ + 408, + 460, + 429, + 470 + ], + "score": 0.84, + "content": "{ < } 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 459, + 434, + 471 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "Cross-Architecture. First, we relax the assumption that the attacker knows the victim architecture.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "We evaluate the transferability between a VGG-16-based SDN an an MSDNet—all trained using", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 144, + 511 + ], + "score": 1.0, + "content": "the same", + "type": "text" + }, + { + "bbox": [ + 144, + 499, + 154, + 509 + ], + "score": 0.68, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 498, + 505, + 511 + ], + "score": 1.0, + "content": ". We find that the samples crafted on the MSDNet can slowdown the SDN: reducing its", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 283, + 523 + ], + "score": 1.0, + "content": "efficacy to 0.63 (from 0.77) and accuracy to", + "type": "text" + }, + { + "bbox": [ + 284, + 510, + 303, + 520 + ], + "score": 0.88, + "content": "78 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 510, + 330, + 523 + ], + "score": 1.0, + "content": "(from", + "type": "text" + }, + { + "bbox": [ + 330, + 510, + 350, + 521 + ], + "score": 0.85, + "content": "8 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "). Interestingly, the opposite seems not", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "to be the case: on the samples crafted against the SDN, the MSDNet still has 0.87 efficacy (from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 531, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 147, + 545 + ], + "score": 1.0, + "content": "0.89) and", + "type": "text" + }, + { + "bbox": [ + 147, + 532, + 168, + 542 + ], + "score": 0.88, + "content": "73 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 531, + 233, + 545 + ], + "score": 1.0, + "content": "accuracy (from", + "type": "text" + }, + { + "bbox": [ + 234, + 532, + 254, + 543 + ], + "score": 0.82, + "content": "8 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 531, + 506, + 545 + ], + "score": 1.0, + "content": "). This hints that DeepSloth transfers if the adversary uses an", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 542, + 284, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 284, + 556 + ], + "score": 1.0, + "content": "effective multi-exit models as the surrogate.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 573 + ], + "score": 1.0, + "content": "Limited Training Set Knowledge. Second, we relax the assumption that the attacker knows the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 191, + 583 + ], + "score": 1.0, + "content": "victim’s training set,", + "type": "text" + }, + { + "bbox": [ + 191, + 572, + 200, + 581 + ], + "score": 0.66, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 571, + 410, + 583 + ], + "score": 1.0, + "content": ". Here, the attacker only knows a random portion of", + "type": "text" + }, + { + "bbox": [ + 411, + 572, + 420, + 581 + ], + "score": 0.78, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 571, + 441, + 583 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 441, + 572, + 461, + 582 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 571, + 464, + 583 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 464, + 572, + 484, + 582 + ], + "score": 0.84, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 571, + 505, + 583 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 126, + 593 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 581, + 444, + 595 + ], + "score": 1.0, + "content": ". We use VGG-16 architecture for both the surrogate and victim models. In the", + "type": "text" + }, + { + "bbox": [ + 444, + 583, + 463, + 593 + ], + "score": 0.84, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 581, + 466, + 595 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 467, + 583, + 487, + 593 + ], + "score": 0.85, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 126, + 604 + ], + "score": 0.83, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "settings, respectively, the attacks reduce the victim’s efficacy to 0.66, 0.5, 0.45 and 0.43 (from", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 190, + 617 + ], + "score": 1.0, + "content": "0.77); its accuracy to", + "type": "text" + }, + { + "bbox": [ + 190, + 604, + 209, + 615 + ], + "score": 0.84, + "content": "81 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 604, + 212, + 617 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 213, + 604, + 232, + 615 + ], + "score": 0.84, + "content": "73 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 604, + 235, + 617 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 235, + 604, + 255, + 615 + ], + "score": 0.86, + "content": "72 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 604, + 272, + 617 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 272, + 604, + 291, + 615 + ], + "score": 0.86, + "content": "74 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 604, + 317, + 617 + ], + "score": 1.0, + "content": "(from", + "type": "text" + }, + { + "bbox": [ + 317, + 604, + 336, + 615 + ], + "score": 0.84, + "content": "8 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "). Overall, the more limited the adversary’s", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 615, + 478, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 116, + 625 + ], + "score": 0.79, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 615, + 478, + 628 + ], + "score": 1.0, + "content": "is, the less generalization ability the surrogate has and the less transferable the attacks are.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "Cross-Domain. Third, we relax the assumption that the attacker exactly knows the victim’s task.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 199, + 656 + ], + "score": 1.0, + "content": "Here, the attacker uses", + "type": "text" + }, + { + "bbox": [ + 199, + 644, + 213, + 656 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { \\boldsymbol { \\mathcal { J } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 643, + 411, + 656 + ], + "score": 1.0, + "content": "to train the surrogate, different from the victim’s", + "type": "text" + }, + { + "bbox": [ + 411, + 644, + 420, + 654 + ], + "score": 0.79, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "altogether. We use a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 655, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 666 + ], + "score": 1.0, + "content": "VGG-16 on CIFAR-100 as the surrogate and attack a VGG-16-based victim model on CIFAR-10.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 664, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 459, + 678 + ], + "score": 1.0, + "content": "This transfer attack reduces the victim’s efficacy to 0.63 (from 0.77) and its accuracy to", + "type": "text" + }, + { + "bbox": [ + 459, + 666, + 479, + 676 + ], + "score": 0.86, + "content": "83 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 664, + 506, + 678 + ], + "score": 1.0, + "content": "(from", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 676, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 127, + 688 + ], + "score": 0.82, + "content": "8 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 676, + 453, + 690 + ], + "score": 1.0, + "content": "). We see that the cross-domain attack might be more effective than the limited", + "type": "text" + }, + { + "bbox": [ + 453, + 677, + 462, + 687 + ], + "score": 0.78, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 676, + 507, + 690 + ], + "score": 1.0, + "content": "scenarios.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 475, + 700 + ], + "score": 1.0, + "content": "This makes DeepSloth particularly dangerous as the attacker, without knowing the victim’s", + "type": "text" + }, + { + "bbox": [ + 475, + 688, + 485, + 698 + ], + "score": 0.79, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 687, + 506, + 700 + ], + "score": 1.0, + "content": ", can", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "collect a similar dataset and still slowdown the victim. We hypothesize the transferability of earlier", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "layer features in CNNs (Yosinski et al., 2014) enables the perturbations attack to transfer from one", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 720, + 323, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 323, + 733 + ], + "score": 1.0, + "content": "domain to another, as long as they are similar enough.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 81, + 502, + 174 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 81, + 502, + 174 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 81, + 502, + 174 + ], + "spans": [ + { + "bbox": [ + 109, + 81, + 502, + 174 + ], + "score": 0.962, + "type": "image", + "image_path": "eb4db376fcea7527d2b8a56c5918847e8a4e1a6ae201be0033e35f93b1112bd1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 81, + 502, + 112.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 112.0, + 502, + 143.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 143.0, + 502, + 174.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 177, + 504, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "Figure 3: Visualising features against attacks using UMAP. VGG-16’s 3rd (left), 8th (middle), and 14th", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 186, + 436, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 436, + 198 + ], + "score": 1.0, + "content": "(right) hidden block features on CIFAR-10’s ’dog’ class (Best viewed in color, zoomed in).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 223, + 505, + 354 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 105, + 223, + 506, + 357 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 379, + 432, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 378, + 434, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 434, + 392 + ], + "score": 1.0, + "content": "5.2 TOWARDS A BLACK-BOX ATTACK: TRANSFERABILITY OF DEEPSLOTH", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 378, + 434, + 392 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 404, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 106, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "Transferability of adversarial examples imply that they can still hurt a model that they were not", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "crafted on (Tramèr et al., 2017a; Liu et al., 2017). Even though white-box attacks are important to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 426, + 507, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 507, + 439 + ], + "score": 1.0, + "content": "expose the vulnerability, black-box attacks, by requiring fewer assumptions, are more practical. Here,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "score": 1.0, + "content": "on four distinct scenarios, we investigate whether DeepSloth is transferable. Based on the scenario’s", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "constraints, we (i) train a surrogate model; (ii) craft the DeepSloth samples on it; and (iii) use these", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 459, + 434, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 408, + 471 + ], + "score": 1.0, + "content": "samples on the victim. We run these experiments on CIFAR-10 in the RAD", + "type": "text" + }, + { + "bbox": [ + 408, + 460, + 429, + 470 + ], + "score": 0.84, + "content": "{ < } 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 459, + 434, + 471 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 405, + 507, + 471 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "Cross-Architecture. First, we relax the assumption that the attacker knows the victim architecture.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "We evaluate the transferability between a VGG-16-based SDN an an MSDNet—all trained using", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 144, + 511 + ], + "score": 1.0, + "content": "the same", + "type": "text" + }, + { + "bbox": [ + 144, + 499, + 154, + 509 + ], + "score": 0.68, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 498, + 505, + 511 + ], + "score": 1.0, + "content": ". We find that the samples crafted on the MSDNet can slowdown the SDN: reducing its", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 283, + 523 + ], + "score": 1.0, + "content": "efficacy to 0.63 (from 0.77) and accuracy to", + "type": "text" + }, + { + "bbox": [ + 284, + 510, + 303, + 520 + ], + "score": 0.88, + "content": "78 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 510, + 330, + 523 + ], + "score": 1.0, + "content": "(from", + "type": "text" + }, + { + "bbox": [ + 330, + 510, + 350, + 521 + ], + "score": 0.85, + "content": "8 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "). Interestingly, the opposite seems not", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "to be the case: on the samples crafted against the SDN, the MSDNet still has 0.87 efficacy (from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 531, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 147, + 545 + ], + "score": 1.0, + "content": "0.89) and", + "type": "text" + }, + { + "bbox": [ + 147, + 532, + 168, + 542 + ], + "score": 0.88, + "content": "73 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 531, + 233, + 545 + ], + "score": 1.0, + "content": "accuracy (from", + "type": "text" + }, + { + "bbox": [ + 234, + 532, + 254, + 543 + ], + "score": 0.82, + "content": "8 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 531, + 506, + 545 + ], + "score": 1.0, + "content": "). This hints that DeepSloth transfers if the adversary uses an", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 542, + 284, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 284, + 556 + ], + "score": 1.0, + "content": "effective multi-exit models as the surrogate.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 104, + 476, + 506, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 573 + ], + "score": 1.0, + "content": "Limited Training Set Knowledge. Second, we relax the assumption that the attacker knows the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 191, + 583 + ], + "score": 1.0, + "content": "victim’s training set,", + "type": "text" + }, + { + "bbox": [ + 191, + 572, + 200, + 581 + ], + "score": 0.66, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 571, + 410, + 583 + ], + "score": 1.0, + "content": ". Here, the attacker only knows a random portion of", + "type": "text" + }, + { + "bbox": [ + 411, + 572, + 420, + 581 + ], + "score": 0.78, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 571, + 441, + 583 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 441, + 572, + 461, + 582 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 571, + 464, + 583 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 464, + 572, + 484, + 582 + ], + "score": 0.84, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 571, + 505, + 583 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 126, + 593 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 581, + 444, + 595 + ], + "score": 1.0, + "content": ". We use VGG-16 architecture for both the surrogate and victim models. In the", + "type": "text" + }, + { + "bbox": [ + 444, + 583, + 463, + 593 + ], + "score": 0.84, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 581, + 466, + 595 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 467, + 583, + 487, + 593 + ], + "score": 0.85, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 126, + 604 + ], + "score": 0.83, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "settings, respectively, the attacks reduce the victim’s efficacy to 0.66, 0.5, 0.45 and 0.43 (from", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 190, + 617 + ], + "score": 1.0, + "content": "0.77); its accuracy to", + "type": "text" + }, + { + "bbox": [ + 190, + 604, + 209, + 615 + ], + "score": 0.84, + "content": "81 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 604, + 212, + 617 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 213, + 604, + 232, + 615 + ], + "score": 0.84, + "content": "73 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 604, + 235, + 617 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 235, + 604, + 255, + 615 + ], + "score": 0.86, + "content": "72 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 604, + 272, + 617 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 272, + 604, + 291, + 615 + ], + "score": 0.86, + "content": "74 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 604, + 317, + 617 + ], + "score": 1.0, + "content": "(from", + "type": "text" + }, + { + "bbox": [ + 317, + 604, + 336, + 615 + ], + "score": 0.84, + "content": "8 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "). Overall, the more limited the adversary’s", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 615, + 478, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 116, + 625 + ], + "score": 0.79, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 615, + 478, + 628 + ], + "score": 1.0, + "content": "is, the less generalization ability the surrogate has and the less transferable the attacks are.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 560, + 506, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "Cross-Domain. Third, we relax the assumption that the attacker exactly knows the victim’s task.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 199, + 656 + ], + "score": 1.0, + "content": "Here, the attacker uses", + "type": "text" + }, + { + "bbox": [ + 199, + 644, + 213, + 656 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { \\boldsymbol { \\mathcal { J } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 643, + 411, + 656 + ], + "score": 1.0, + "content": "to train the surrogate, different from the victim’s", + "type": "text" + }, + { + "bbox": [ + 411, + 644, + 420, + 654 + ], + "score": 0.79, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "altogether. We use a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 655, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 666 + ], + "score": 1.0, + "content": "VGG-16 on CIFAR-100 as the surrogate and attack a VGG-16-based victim model on CIFAR-10.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 664, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 459, + 678 + ], + "score": 1.0, + "content": "This transfer attack reduces the victim’s efficacy to 0.63 (from 0.77) and its accuracy to", + "type": "text" + }, + { + "bbox": [ + 459, + 666, + 479, + 676 + ], + "score": 0.86, + "content": "83 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 664, + 506, + 678 + ], + "score": 1.0, + "content": "(from", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 676, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 127, + 688 + ], + "score": 0.82, + "content": "8 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 676, + 453, + 690 + ], + "score": 1.0, + "content": "). We see that the cross-domain attack might be more effective than the limited", + "type": "text" + }, + { + "bbox": [ + 453, + 677, + 462, + 687 + ], + "score": 0.78, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 676, + 507, + 690 + ], + "score": 1.0, + "content": "scenarios.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 475, + 700 + ], + "score": 1.0, + "content": "This makes DeepSloth particularly dangerous as the attacker, without knowing the victim’s", + "type": "text" + }, + { + "bbox": [ + 475, + 688, + 485, + 698 + ], + "score": 0.79, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 687, + 506, + 700 + ], + "score": 1.0, + "content": ", can", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "collect a similar dataset and still slowdown the victim. We hypothesize the transferability of earlier", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "layer features in CNNs (Yosinski et al., 2014) enables the perturbations attack to transfer from one", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 720, + 323, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 323, + 733 + ], + "score": 1.0, + "content": "domain to another, as long as they are similar enough.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 633, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Cross-Mechnanism. Finally, we test the scenario where the victim uses a completely different", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "mechanism than a multi-exit architecture to implement input adaptiveness, i.e., SkipNet (Wang et al.,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 116 + ], + "score": 1.0, + "content": "2018). A SkipNet, a modified residual network, selectively skips convolutional blocks based on the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 507, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 507, + 129 + ], + "score": 1.0, + "content": "activations of the previous layer and, therefore, does not include any internal classifiers. We use a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "pre-trained SkipNet on CIFAR-10 that reduces the average computation for each input sample by", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 107, + 137, + 134, + 148 + ], + "score": 0.89, + "content": "{ \\sim } 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 137, + 298, + 150 + ], + "score": 1.0, + "content": "over an equivalent ResNet and achieves", + "type": "text" + }, + { + "bbox": [ + 299, + 137, + 326, + 148 + ], + "score": 0.89, + "content": "{ \\sim } 9 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "accuracy. We then feed DeepSloth samples", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 477, + 161 + ], + "score": 1.0, + "content": "crafted on a MSDNet to this SkipNet, which reduces its average computational saving to", + "type": "text" + }, + { + "bbox": [ + 477, + 149, + 505, + 159 + ], + "score": 0.88, + "content": "{ \\sim } 3 2 \\%", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 109, + 172 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 110, + 159, + 129, + 170 + ], + "score": 0.86, + "content": "3 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 159, + 263, + 172 + ], + "score": 1.0, + "content": "less effective) and its accuracy to", + "type": "text" + }, + { + "bbox": [ + 263, + 159, + 283, + 170 + ], + "score": 0.87, + "content": "37 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 159, + 505, + 172 + ], + "score": 1.0, + "content": ". This result suggests that the two different mechanisms", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "have more in common than previously known and might share the vulnerability. We believe that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 194 + ], + "score": 1.0, + "content": "understanding the underlying mechanisms through which adaptive models save computation is an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 285, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 285, + 204 + ], + "score": 1.0, + "content": "important research question for future work.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 107, + 220, + 470, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 472, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 472, + 235 + ], + "score": 1.0, + "content": "6 STANDARD ADVERSARIAL TRAINING IS NOT A COUNTERMEASURE", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "score": 1.0, + "content": "In this section, we examine whether a defender can adapt a standard countermeasure against adversar-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "score": 1.0, + "content": "ial perturbations, adversarial training (AT) (Madry et al., 2018), to mitigate our attack. AT decreases a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "score": 1.0, + "content": "model’s sensitivity to perturbations that significantly change the model’s outputs. While this scheme", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 278, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 293 + ], + "score": 1.0, + "content": "is effective against adversarial examples that aim to trigger misclassifications; it is unclear whether", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "score": 1.0, + "content": "using our DeepSloth samples for AT can also robustify a multi-exit model against slowdown attacks.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 318 + ], + "score": 1.0, + "content": "To evaluate, we train our multi-exit models as follows. We first take a base network—VGG-16—and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "train it on CIFAR-10 on PGD-10 adversarial examples. We then convert the resulting model into", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 328, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 104, + 328, + 507, + 343 + ], + "score": 1.0, + "content": "a multi-exit architecture, using the modification from (Kaya et al., 2019). During this conversion,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 340, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 352 + ], + "score": 1.0, + "content": "we adversarially train individual exit points using PGD-10, PGD-10 (avg.), PGD-10 (max.), and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "DeepSloth; similar to (Hu et al., 2020). Finally, we measure the efficacy and accuracy of the trained", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 362, + 499, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 499, + 375 + ], + "score": 1.0, + "content": "models against PGD-20, PGD-20 (avg.), PGD-20 (max.), and DeepSloth, on CIFAR-10’s test-set.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "table", + "bbox": [ + 106, + 416, + 504, + 507 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 385, + 503, + 408 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Table 3: Evaluating adversarial training against slowdown attacks. Each entry includes the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 451, + 410 + ], + "score": 1.0, + "content": "model’s efficacy score (left) and accuracy (right). Results are on CIFAR-10, in the RAD", + "type": "text" + }, + { + "bbox": [ + 452, + 396, + 473, + 406 + ], + "score": 0.79, + "content": "{ < } 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "setting.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 416, + 504, + 507 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 416, + 504, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 504, + 507 + ], + "score": 0.983, + "html": "
ADV. TRAININGNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)DEEPSLOTH
UNDEFENDED0.77 /89%0.79/29%0.85 /10%0.81/27%0.01 / 13%
PGD-100.61/ 72%0.55 /38%0.64 /23%0.58 /29%0.33 / 70%
PGD-10 (AVG.)0.53 / 72%0.47 / 36%0.47 /35%0.47 /35%0.32 / 70%
PGD-10 (MAX.)0.57 /72%0.51/37%0.54 / 30%0.52 /34%0.32 / 70%
OURS0.74/72%0.71/38%0.82 /14%0.77 /21%0.44/ 67%
OURS + PGD-100.61/73%0.55 /38%0.63 /23%0.58 /28%0.33 / 70%
", + "type": "table", + "image_path": "7552a3da79778e2f12ef0c7b85c99a13a72681f7575d595cbdb355fabe60ff74.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 106, + 416, + 504, + 446.3333333333333 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 106, + 446.3333333333333, + 504, + 476.66666666666663 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 106, + 476.66666666666663, + 504, + 506.99999999999994 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "index": 24.75 + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 623 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "Our results in Table 3 verify that AT provides resilience against all PGD attacks. Besides, AT provides", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 369, + 536 + ], + "score": 1.0, + "content": "some resilience to our attack: DeepSloth reduces the efficacy to", + "type": "text" + }, + { + "bbox": [ + 369, + 524, + 396, + 534 + ], + "score": 0.86, + "content": "{ \\sim } 0 . 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "on robust models vs. 0.01", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "on the undefended one. However, we identify a trade-off between the robustness and efficiency of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "multi-exits. Compared to the undefended model, on clean samples, we see that robust models have", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 209, + 568 + ], + "score": 1.0, + "content": "lower efficacy—0.77 vs.", + "type": "text" + }, + { + "bbox": [ + 210, + 556, + 254, + 567 + ], + "score": 0.85, + "content": "0 . 5 3 \\sim 0 . 6 1", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 556, + 506, + 568 + ], + "score": 1.0, + "content": ". We observe that the model trained only with our DeepSloth", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 568, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 579 + ], + "score": 1.0, + "content": "samples (Ours) can recover the efficacy on both the clean and our DeepSloth samples, but this model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "loses its robustness against PGD attacks. Moreover, when we train a model on both our DeepSloth", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 199, + 602 + ], + "score": 1.0, + "content": "samples and PGD-10", + "type": "text" + }, + { + "bbox": [ + 199, + 590, + 267, + 600 + ], + "score": 0.25, + "content": "\\mathrm { \\ O u r s } + \\mathrm { P G D } { - } 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "), the trained model suffers from low efficacy. Our results", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "score": 1.0, + "content": "imply that a defender may require an out-of-the-box defense, such as flagging the users whose queries", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 612, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 623 + ], + "score": 1.0, + "content": "bypass the early-exits more often than clean samples for which the multi-exit network was calibrated.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 108, + 640, + 201, + 653 + ], + "lines": [ + { + "bbox": [ + 104, + 638, + 203, + 656 + ], + "spans": [ + { + "bbox": [ + 104, + 638, + 203, + 656 + ], + "score": 1.0, + "content": "7 CONCLUSIONS", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "This work exposes the vulnerability of input-adaptive inference mechanisms against adversarial", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "slowdowns. As a vehicle for exploring this vulnerability systematically, we propose DeepSloth,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "an attack that introduces imperceptible adversarial perturbations to test-time inputs for offsetting", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "the computational benefits of multi-exit inference mechanisms. We show that a white-box attack,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "which perturbs each sample individually, eliminates any computational savings these mechanisms", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "provide. We also show that it is possible to craft universal slowdown perturbations, which can be", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Cross-Mechnanism. Finally, we test the scenario where the victim uses a completely different", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "mechanism than a multi-exit architecture to implement input adaptiveness, i.e., SkipNet (Wang et al.,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 116 + ], + "score": 1.0, + "content": "2018). A SkipNet, a modified residual network, selectively skips convolutional blocks based on the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 507, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 507, + 129 + ], + "score": 1.0, + "content": "activations of the previous layer and, therefore, does not include any internal classifiers. We use a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "pre-trained SkipNet on CIFAR-10 that reduces the average computation for each input sample by", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 107, + 137, + 134, + 148 + ], + "score": 0.89, + "content": "{ \\sim } 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 137, + 298, + 150 + ], + "score": 1.0, + "content": "over an equivalent ResNet and achieves", + "type": "text" + }, + { + "bbox": [ + 299, + 137, + 326, + 148 + ], + "score": 0.89, + "content": "{ \\sim } 9 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "accuracy. We then feed DeepSloth samples", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 477, + 161 + ], + "score": 1.0, + "content": "crafted on a MSDNet to this SkipNet, which reduces its average computational saving to", + "type": "text" + }, + { + "bbox": [ + 477, + 149, + 505, + 159 + ], + "score": 0.88, + "content": "{ \\sim } 3 2 \\%", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 109, + 172 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 110, + 159, + 129, + 170 + ], + "score": 0.86, + "content": "3 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 159, + 263, + 172 + ], + "score": 1.0, + "content": "less effective) and its accuracy to", + "type": "text" + }, + { + "bbox": [ + 263, + 159, + 283, + 170 + ], + "score": 0.87, + "content": "37 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 159, + 505, + 172 + ], + "score": 1.0, + "content": ". This result suggests that the two different mechanisms", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "have more in common than previously known and might share the vulnerability. We believe that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 194 + ], + "score": 1.0, + "content": "understanding the underlying mechanisms through which adaptive models save computation is an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 285, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 285, + 204 + ], + "score": 1.0, + "content": "important research question for future work.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 82, + 507, + 204 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 220, + 470, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 472, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 472, + 235 + ], + "score": 1.0, + "content": "6 STANDARD ADVERSARIAL TRAINING IS NOT A COUNTERMEASURE", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "score": 1.0, + "content": "In this section, we examine whether a defender can adapt a standard countermeasure against adversar-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "score": 1.0, + "content": "ial perturbations, adversarial training (AT) (Madry et al., 2018), to mitigate our attack. AT decreases a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "score": 1.0, + "content": "model’s sensitivity to perturbations that significantly change the model’s outputs. While this scheme", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 278, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 293 + ], + "score": 1.0, + "content": "is effective against adversarial examples that aim to trigger misclassifications; it is unclear whether", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "score": 1.0, + "content": "using our DeepSloth samples for AT can also robustify a multi-exit model against slowdown attacks.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 246, + 506, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 318 + ], + "score": 1.0, + "content": "To evaluate, we train our multi-exit models as follows. We first take a base network—VGG-16—and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "train it on CIFAR-10 on PGD-10 adversarial examples. We then convert the resulting model into", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 328, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 104, + 328, + 507, + 343 + ], + "score": 1.0, + "content": "a multi-exit architecture, using the modification from (Kaya et al., 2019). During this conversion,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 340, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 352 + ], + "score": 1.0, + "content": "we adversarially train individual exit points using PGD-10, PGD-10 (avg.), PGD-10 (max.), and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "DeepSloth; similar to (Hu et al., 2020). Finally, we measure the efficacy and accuracy of the trained", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 362, + 499, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 499, + 375 + ], + "score": 1.0, + "content": "models against PGD-20, PGD-20 (avg.), PGD-20 (max.), and DeepSloth, on CIFAR-10’s test-set.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 104, + 307, + 507, + 375 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 416, + 504, + 507 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 385, + 503, + 408 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Table 3: Evaluating adversarial training against slowdown attacks. Each entry includes the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 451, + 410 + ], + "score": 1.0, + "content": "model’s efficacy score (left) and accuracy (right). Results are on CIFAR-10, in the RAD", + "type": "text" + }, + { + "bbox": [ + 452, + 396, + 473, + 406 + ], + "score": 0.79, + "content": "{ < } 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "setting.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 416, + 504, + 507 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 416, + 504, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 504, + 507 + ], + "score": 0.983, + "html": "
ADV. TRAININGNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)DEEPSLOTH
UNDEFENDED0.77 /89%0.79/29%0.85 /10%0.81/27%0.01 / 13%
PGD-100.61/ 72%0.55 /38%0.64 /23%0.58 /29%0.33 / 70%
PGD-10 (AVG.)0.53 / 72%0.47 / 36%0.47 /35%0.47 /35%0.32 / 70%
PGD-10 (MAX.)0.57 /72%0.51/37%0.54 / 30%0.52 /34%0.32 / 70%
OURS0.74/72%0.71/38%0.82 /14%0.77 /21%0.44/ 67%
OURS + PGD-100.61/73%0.55 /38%0.63 /23%0.58 /28%0.33 / 70%
", + "type": "table", + "image_path": "7552a3da79778e2f12ef0c7b85c99a13a72681f7575d595cbdb355fabe60ff74.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 106, + 416, + 504, + 446.3333333333333 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 106, + 446.3333333333333, + 504, + 476.66666666666663 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 106, + 476.66666666666663, + 504, + 506.99999999999994 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "index": 24.75 + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 623 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 524 + ], + "score": 1.0, + "content": "Our results in Table 3 verify that AT provides resilience against all PGD attacks. Besides, AT provides", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 369, + 536 + ], + "score": 1.0, + "content": "some resilience to our attack: DeepSloth reduces the efficacy to", + "type": "text" + }, + { + "bbox": [ + 369, + 524, + 396, + 534 + ], + "score": 0.86, + "content": "{ \\sim } 0 . 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "on robust models vs. 0.01", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "on the undefended one. However, we identify a trade-off between the robustness and efficiency of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "multi-exits. Compared to the undefended model, on clean samples, we see that robust models have", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 209, + 568 + ], + "score": 1.0, + "content": "lower efficacy—0.77 vs.", + "type": "text" + }, + { + "bbox": [ + 210, + 556, + 254, + 567 + ], + "score": 0.85, + "content": "0 . 5 3 \\sim 0 . 6 1", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 556, + 506, + 568 + ], + "score": 1.0, + "content": ". We observe that the model trained only with our DeepSloth", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 568, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 579 + ], + "score": 1.0, + "content": "samples (Ours) can recover the efficacy on both the clean and our DeepSloth samples, but this model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "loses its robustness against PGD attacks. Moreover, when we train a model on both our DeepSloth", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 199, + 602 + ], + "score": 1.0, + "content": "samples and PGD-10", + "type": "text" + }, + { + "bbox": [ + 199, + 590, + 267, + 600 + ], + "score": 0.25, + "content": "\\mathrm { \\ O u r s } + \\mathrm { P G D } { - } 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "), the trained model suffers from low efficacy. Our results", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "score": 1.0, + "content": "imply that a defender may require an out-of-the-box defense, such as flagging the users whose queries", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 612, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 623 + ], + "score": 1.0, + "content": "bypass the early-exits more often than clean samples for which the multi-exit network was calibrated.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 513, + 506, + 623 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 640, + 201, + 653 + ], + "lines": [ + { + "bbox": [ + 104, + 638, + 203, + 656 + ], + "spans": [ + { + "bbox": [ + 104, + 638, + 203, + 656 + ], + "score": 1.0, + "content": "7 CONCLUSIONS", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "This work exposes the vulnerability of input-adaptive inference mechanisms against adversarial", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "slowdowns. As a vehicle for exploring this vulnerability systematically, we propose DeepSloth,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "an attack that introduces imperceptible adversarial perturbations to test-time inputs for offsetting", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "the computational benefits of multi-exit inference mechanisms. We show that a white-box attack,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "which perturbs each sample individually, eliminates any computational savings these mechanisms", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "provide. We also show that it is possible to craft universal slowdown perturbations, which can be", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "reused, and transferable samples, in a black-box setting. Moreover, adversarial training, a standard", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "countermeasure for adversarial perturbations, is not effective against DeepSloth. Our analysis suggests", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 472, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 472, + 117 + ], + "score": 1.0, + "content": "that slowdown attacks are a realistic, yet under-appreciated, threat against adaptive models.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 41.5, + "bbox_fs": [ + 104, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "reused, and transferable samples, in a black-box setting. Moreover, adversarial training, a standard", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "countermeasure for adversarial perturbations, is not effective against DeepSloth. Our analysis suggests", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 472, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 472, + 117 + ], + "score": 1.0, + "content": "that slowdown attacks are a realistic, yet under-appreciated, threat against adaptive models.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 133, + 212, + 145 + ], + "lines": [ + { + "bbox": [ + 107, + 132, + 214, + 147 + ], + "spans": [ + { + "bbox": [ + 107, + 132, + 214, + 147 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 108, + 157, + 505, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 171 + ], + "score": 1.0, + "content": "We thank the anonymous reviewers for their feedback. This research was partially supported by the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 167, + 204, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 204, + 181 + ], + "score": 1.0, + "content": "Department of Defense.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 108, + 196, + 175, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 195, + 177, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 177, + 210 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 213, + 506, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 506, + 225 + ], + "score": 1.0, + "content": "Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security:", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 223, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 116, + 223, + 505, + 234 + ], + "score": 1.0, + "content": "Circumventing defenses to adversarial examples. In Jennifer Dy and Andreas Krause (eds.), Proceedings of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 231, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 114, + 231, + 506, + 246 + ], + "score": 1.0, + "content": "the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 241, + 508, + 257 + ], + "spans": [ + { + "bbox": [ + 114, + 241, + 508, + 257 + ], + "score": 1.0, + "content": "Research, pp. 274–283, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 254, + 352, + 265 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 352, + 265 + ], + "score": 1.0, + "content": "//proceedings.mlr.press/v80/athalye18a.html.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 371, + 283 + ], + "score": 1.0, + "content": "Ron Banner, Itay Hubara, Elad Hoffer, and Daniel Soudry.", + "type": "text" + }, + { + "bbox": [ + 388, + 271, + 506, + 282 + ], + "score": 1.0, + "content": "Scalable methods for 8-bit", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 115, + 281, + 238, + 292 + ], + "score": 1.0, + "content": "training of neural networks.", + "type": "text" + }, + { + "bbox": [ + 252, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "In S. Bengio, H. Wallach, H. Larochelle, K. Grauman,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 291, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 115, + 291, + 506, + 303 + ], + "score": 1.0, + "content": "N. Cesa-Bianchi, and R. Garnett (eds.), Advances in Neural Information Processing Systems", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 114, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "31, pp. 5145–5153. Curran Associates, Inc., 2018. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 311, + 464, + 323 + ], + "spans": [ + { + "bbox": [ + 115, + 311, + 464, + 323 + ], + "score": 1.0, + "content": "7761-scalable-methods-for-8-bit-training-of-neural-networks.pdf.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 105, + 330, + 504, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 339, + 277, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 277, + 351 + ], + "score": 1.0, + "content": "Security and Privacy (SP), pp. 39–57, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "Jiasi Chen and Xukan Ran. Deep learning with edge computing: A review. Proceedings of the IEEE, 107(8):", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 368, + 184, + 379 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 184, + 379 + ], + "score": 1.0, + "content": "1655–1674, 2019.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "Pin-Yu Chen, Yash Sharma, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh. Ead: elastic-net attacks to deep neural", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 396, + 392, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 392, + 408 + ], + "score": 1.0, + "content": "networks via adversarial examples. arXiv preprint arXiv:1709.04114, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "Elias De Coninck, Tim Verbelen, Bert Vankeirsbilck, Steven Bohez, Pieter Simoens, Piet Demeester, and Bart", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 116, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "Dhoedt. Distributed neural networks for internet of things: The big-little approach. In International Internet", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 435, + 290, + 446 + ], + "spans": [ + { + "bbox": [ + 116, + 435, + 290, + 446 + ], + "score": 1.0, + "content": "of Things Summit, pp. 484–492. Springer, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "Michael Figurnov, Maxwell D. Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 117, + 464, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 117, + 464, + 505, + 474 + ], + "score": 1.0, + "content": "Salakhutdinov. Spatially Adaptive Computation Time for Residual Networks. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 473, + 396, + 485 + ], + "spans": [ + { + "bbox": [ + 116, + 473, + 396, + 485 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 491, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 505, + 503 + ], + "score": 1.0, + "content": "Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 502, + 507, + 513 + ], + "spans": [ + { + "bbox": [ + 115, + 502, + 507, + 513 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, 2015. URL http://arxiv.org/abs/1412.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 510, + 144, + 523 + ], + "spans": [ + { + "bbox": [ + 116, + 510, + 144, + 523 + ], + "score": 1.0, + "content": "6572.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 539, + 414, + 551 + ], + "spans": [ + { + "bbox": [ + 115, + 539, + 414, + 551 + ], + "score": 1.0, + "content": "trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 504, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "Mirazul Haque, Anki Chauhan, Cong Liu, and Wei Yang. Ilfo: Adversarial attack on adaptive neural networks.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 568, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 115, + 568, + 506, + 580 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 577, + 140, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 140, + 590 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 105, + 596, + 504, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 606, + 385, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 606, + 385, + 617 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 505, + 665 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 506, + 636 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 117, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 117, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 654, + 166, + 665 + ], + "spans": [ + { + "bbox": [ + 116, + 654, + 166, + 665 + ], + "score": 1.0, + "content": "82–97, 2012.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 105, + 672, + 504, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, and Qun Liu. DynaBERT: Dynamic BERT with", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 683, + 442, + 694 + ], + "spans": [ + { + "bbox": [ + 116, + 683, + 442, + 694 + ], + "score": 1.0, + "content": "Adaptive Width and Depth. In Advances in Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 108, + 701, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 712, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 712, + 505, + 722 + ], + "score": 1.0, + "content": "Andreetto, and Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 721, + 189, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 189, + 732 + ], + "score": 1.0, + "content": "Applications, 2017.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 83, + 506, + 117 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 133, + 212, + 145 + ], + "lines": [ + { + "bbox": [ + 107, + 132, + 214, + 147 + ], + "spans": [ + { + "bbox": [ + 107, + 132, + 214, + 147 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 108, + 157, + 505, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 171 + ], + "score": 1.0, + "content": "We thank the anonymous reviewers for their feedback. This research was partially supported by the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 167, + 204, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 204, + 181 + ], + "score": 1.0, + "content": "Department of Defense.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 155, + 505, + 181 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 196, + 175, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 195, + 177, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 177, + 210 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 213, + 506, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 506, + 225 + ], + "score": 1.0, + "content": "Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security:", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 223, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 116, + 223, + 505, + 234 + ], + "score": 1.0, + "content": "Circumventing defenses to adversarial examples. In Jennifer Dy and Andreas Krause (eds.), Proceedings of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 231, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 114, + 231, + 506, + 246 + ], + "score": 1.0, + "content": "the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 241, + 508, + 257 + ], + "spans": [ + { + "bbox": [ + 114, + 241, + 508, + 257 + ], + "score": 1.0, + "content": "Research, pp. 274–283, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. PMLR. URL http:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 254, + 352, + 265 + ], + "spans": [ + { + "bbox": [ + 116, + 254, + 352, + 265 + ], + "score": 1.0, + "content": "//proceedings.mlr.press/v80/athalye18a.html.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 214, + 508, + 265 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 371, + 283 + ], + "score": 1.0, + "content": "Ron Banner, Itay Hubara, Elad Hoffer, and Daniel Soudry.", + "type": "text" + }, + { + "bbox": [ + 388, + 271, + 506, + 282 + ], + "score": 1.0, + "content": "Scalable methods for 8-bit", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 115, + 281, + 238, + 292 + ], + "score": 1.0, + "content": "training of neural networks.", + "type": "text" + }, + { + "bbox": [ + 252, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "In S. Bengio, H. Wallach, H. Larochelle, K. Grauman,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 291, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 115, + 291, + 506, + 303 + ], + "score": 1.0, + "content": "N. Cesa-Bianchi, and R. Garnett (eds.), Advances in Neural Information Processing Systems", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 114, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "31, pp. 5145–5153. Curran Associates, Inc., 2018. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 311, + 464, + 323 + ], + "spans": [ + { + "bbox": [ + 115, + 311, + 464, + 323 + ], + "score": 1.0, + "content": "7761-scalable-methods-for-8-bit-training-of-neural-networks.pdf.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 270, + 506, + 323 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 330, + 504, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 339, + 277, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 277, + 351 + ], + "score": 1.0, + "content": "Security and Privacy (SP), pp. 39–57, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 329, + 506, + 351 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "Jiasi Chen and Xukan Ran. Deep learning with edge computing: A review. Proceedings of the IEEE, 107(8):", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 368, + 184, + 379 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 184, + 379 + ], + "score": 1.0, + "content": "1655–1674, 2019.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 357, + 506, + 379 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "Pin-Yu Chen, Yash Sharma, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh. Ead: elastic-net attacks to deep neural", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 396, + 392, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 392, + 408 + ], + "score": 1.0, + "content": "networks via adversarial examples. arXiv preprint arXiv:1709.04114, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 385, + 505, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "Elias De Coninck, Tim Verbelen, Bert Vankeirsbilck, Steven Bohez, Pieter Simoens, Piet Demeester, and Bart", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 116, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "Dhoedt. Distributed neural networks for internet of things: The big-little approach. In International Internet", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 435, + 290, + 446 + ], + "spans": [ + { + "bbox": [ + 116, + 435, + 290, + 446 + ], + "score": 1.0, + "content": "of Things Summit, pp. 484–492. Springer, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 414, + 506, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "Michael Figurnov, Maxwell D. Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 117, + 464, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 117, + 464, + 505, + 474 + ], + "score": 1.0, + "content": "Salakhutdinov. Spatially Adaptive Computation Time for Residual Networks. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 473, + 396, + 485 + ], + "spans": [ + { + "bbox": [ + 116, + 473, + 396, + 485 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 453, + 505, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 491, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 505, + 503 + ], + "score": 1.0, + "content": "Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 502, + 507, + 513 + ], + "spans": [ + { + "bbox": [ + 115, + 502, + 507, + 513 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, 2015. URL http://arxiv.org/abs/1412.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 510, + 144, + 523 + ], + "spans": [ + { + "bbox": [ + 116, + 510, + 144, + 523 + ], + "score": 1.0, + "content": "6572.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 491, + 507, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 539, + 414, + 551 + ], + "spans": [ + { + "bbox": [ + 115, + 539, + 414, + 551 + ], + "score": 1.0, + "content": "trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 529, + 506, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 504, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "Mirazul Haque, Anki Chauhan, Cong Liu, and Wei Yang. Ilfo: Adversarial attack on adaptive neural networks.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 568, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 115, + 568, + 506, + 580 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 577, + 140, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 140, + 590 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 557, + 506, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 596, + 504, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 606, + 385, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 606, + 385, + 617 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 596, + 505, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 505, + 665 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 506, + 636 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 117, + 635, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 117, + 635, + 505, + 646 + ], + "score": 1.0, + "content": "Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 654, + 166, + 665 + ], + "spans": [ + { + "bbox": [ + 116, + 654, + 166, + 665 + ], + "score": 1.0, + "content": "82–97, 2012.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 624, + 506, + 665 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 672, + 504, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, and Qun Liu. DynaBERT: Dynamic BERT with", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 683, + 442, + 694 + ], + "spans": [ + { + "bbox": [ + 116, + 683, + 442, + 694 + ], + "score": 1.0, + "content": "Adaptive Width and Depth. In Advances in Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 672, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 701, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 712, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 712, + 505, + 722 + ], + "score": 1.0, + "content": "Andreetto, and Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 721, + 189, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 189, + 732 + ], + "score": 1.0, + "content": "Applications, 2017.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 700, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 83, + 504, + 113 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "C. Hu, W. Bao, D. Wang, and F. Liu. Dynamic adaptive dnn surgery for inference acceleration on the edge.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 105 + ], + "score": 1.0, + "content": "In IEEE INFOCOM 2019 - IEEE Conference on Computer Communications, pp. 1423–1431, 2019. doi:", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 103, + 247, + 113 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 247, + 113 + ], + "score": 1.0, + "content": "10.1109/INFOCOM.2019.8737614.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 121, + 503, + 152 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 506, + 133 + ], + "score": 1.0, + "content": "Ting-Kuei Hu, Tianlong Chen, Haotao Wang, and Zhangyang Wang. Triple wins: Boosting accuracy, robustness", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 114, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "and efficiency together by enabling input-adaptive inference. In International Conference on Learning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 141, + 453, + 153 + ], + "spans": [ + { + "bbox": [ + 116, + 141, + 389, + 153 + ], + "score": 1.0, + "content": "Representations, 2020. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 389, + 143, + 395, + 150 + ], + "score": 0.54, + "content": "\\equiv", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 141, + 453, + 153 + ], + "score": 1.0, + "content": "rJgzzJHtDB.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 159, + 504, + 190 + ], + "lines": [ + { + "bbox": [ + 107, + 160, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 505, + 170 + ], + "score": 1.0, + "content": "Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Weinberger. Multi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 168, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 168, + 505, + 182 + ], + "score": 1.0, + "content": "scale dense networks for resource efficient image classification. In International Conference on Learning", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 179, + 447, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 179, + 389, + 191 + ], + "score": 1.0, + "content": "Representations, 2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 389, + 181, + 395, + 188 + ], + "score": 0.56, + "content": "\\equiv", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 179, + 447, + 191 + ], + "score": 1.0, + "content": "Hk2aImxAb.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 197, + 506, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 209 + ], + "score": 1.0, + "content": "Nathan Inkawhich, Wei Wen, Hai (Helen) Li, and Yiran Chen. Feature space perturbations yield more transferable", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "adversarial examples. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 217, + 140, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 140, + 228 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 506, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "Weiwen Jiang, Edwin H.-M. Sha, Xinyi Zhang, Lei Yang, Qingfeng Zhuge, Yiyu Shi, and Jingtong Hu. Achieving", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 245, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 116, + 245, + 506, + 257 + ], + "score": 1.0, + "content": "super-linear speedup across multi-fpga for real-time dnn inference. ACM Trans. Embed. Comput. Syst., 18(5s),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 117, + 256, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 117, + 256, + 506, + 267 + ], + "score": 1.0, + "content": "October 2019. ISSN 1539-9087. doi: 10.1145/3358192. URL https://doi.org/10.1145/3358192.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 273, + 506, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 272, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 286 + ], + "score": 1.0, + "content": "Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 283, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 114, + 283, + 506, + 296 + ], + "score": 1.0, + "content": "Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. In Proceedings of the Twenty-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 292, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 506, + 307 + ], + "score": 1.0, + "content": "Second International Conference on Architectural Support for Programming Languages and Operating", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 303, + 507, + 316 + ], + "spans": [ + { + "bbox": [ + 114, + 303, + 507, + 316 + ], + "score": 1.0, + "content": "Systems, ASPLOS’17, pp. 615–629, New York, NY, USA, 2017. Association for Computing Machinery.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 313, + 507, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 313, + 507, + 325 + ], + "score": 1.0, + "content": "ISBN 9781450344654. doi: 10.1145/3037697.3037698. URL https://doi.org/10.1145/3037697.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 324, + 159, + 334 + ], + "spans": [ + { + "bbox": [ + 116, + 324, + 159, + 334 + ], + "score": 1.0, + "content": "3037698.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 108, + 341, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "Yigitcan Kaya, Sanghyun Hong, and Tudor Dumitra¸s. Shallow-Deep Networks: Understanding and mitigating ˘", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 352, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 352, + 506, + 363 + ], + "score": 1.0, + "content": "network overthinking. In Proceedings of the 2019 International Conference on Machine Learning (ICML),", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 361, + 219, + 373 + ], + "spans": [ + { + "bbox": [ + 115, + 361, + 219, + 373 + ], + "score": 1.0, + "content": "Long Beach, CA, Jun 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 379, + 484, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 484, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 484, + 392 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning Multiple Layers of Features from Tiny Images. 2009.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 398, + 504, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 407, + 434, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 434, + 420 + ], + "score": 1.0, + "content": "networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. arXiv preprint", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 436, + 209, + 447 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 209, + 447 + ], + "score": 1.0, + "content": "arXiv:1611.01236, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 454, + 504, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 454, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 505, + 465 + ], + "score": 1.0, + "content": "M. Lecuyer, V. Atlidakis, R. Geambasu, D. Hsu, and S. Jana. Certified robustness to adversarial examples with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 465, + 465, + 476 + ], + "spans": [ + { + "bbox": [ + 116, + 465, + 465, + 476 + ], + "score": 1.0, + "content": "differential privacy. In 2019 IEEE Symposium on Security and Privacy (SP), pp. 656–672, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 506, + 523 + ], + "lines": [ + { + "bbox": [ + 104, + 481, + 507, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 481, + 507, + 495 + ], + "score": 1.0, + "content": "En Li, Zhi Zhou, and Xu Chen. Edge intelligence: On-demand deep learning model co-inference with device-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 491, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 506, + 506 + ], + "score": 1.0, + "content": "edge synergy. In Proceedings of the 2018 Workshop on Mobile Edge Communications, MECOMM’18, pp.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 502, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 502, + 506, + 514 + ], + "score": 1.0, + "content": "31–36, New York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450359061. doi:", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 511, + 453, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 511, + 453, + 524 + ], + "score": 1.0, + "content": "10.1145/3229556.3229562. URL https://doi.org/10.1145/3229556.3229562.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 104, + 530, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 506, + 542 + ], + "score": 1.0, + "content": "Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 541, + 419, + 552 + ], + "spans": [ + { + "bbox": [ + 116, + 541, + 419, + 552 + ], + "score": 1.0, + "content": "CoRR, abs/1608.08710, 2016. URL http://arxiv.org/abs/1608.08710.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 558, + 504, + 590 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 506, + 570 + ], + "score": 1.0, + "content": "Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 569, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 569, + 505, + 580 + ], + "score": 1.0, + "content": "attacks using high-level representation guided denoiser. In The IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 579, + 264, + 591 + ], + "spans": [ + { + "bbox": [ + 116, + 579, + 264, + 591 + ], + "score": 1.0, + "content": "Pattern Recognition (CVPR), June 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 506, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 605, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 605, + 506, + 620 + ], + "score": 1.0, + "content": "black-box attacks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 615, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 114, + 615, + 507, + 630 + ], + "score": 1.0, + "content": "April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. URL https://openreview.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 627, + 240, + 639 + ], + "spans": [ + { + "bbox": [ + 115, + 627, + 181, + 639 + ], + "score": 1.0, + "content": "net/forum?id", + "type": "text" + }, + { + "bbox": [ + 182, + 628, + 187, + 636 + ], + "score": 0.52, + "content": "\\equiv", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 627, + 240, + 639 + ], + "score": 1.0, + "content": "Sys6GJqxl.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 104, + 645, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 658 + ], + "score": 1.0, + "content": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 654, + 437, + 666 + ], + "spans": [ + { + "bbox": [ + 115, + 654, + 437, + 666 + ], + "score": 1.0, + "content": "learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 107, + 673, + 506, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 683, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 683, + 506, + 695 + ], + "score": 1.0, + "content": "learning models resistant to adversarial attacks. In International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 693, + 385, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 693, + 385, + 704 + ], + "score": 1.0, + "content": "2018. URL https://openreview.net/forum?id=rJzIBfZAb.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 106, + 711, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "score": 1.0, + "content": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 115, + 721, + 496, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 496, + 733 + ], + "score": 1.0, + "content": "perturbations. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 83, + 504, + 113 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "C. Hu, W. Bao, D. Wang, and F. Liu. Dynamic adaptive dnn surgery for inference acceleration on the edge.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 105 + ], + "score": 1.0, + "content": "In IEEE INFOCOM 2019 - IEEE Conference on Computer Communications, pp. 1423–1431, 2019. doi:", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 103, + 247, + 113 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 247, + 113 + ], + "score": 1.0, + "content": "10.1109/INFOCOM.2019.8737614.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 506, + 113 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 121, + 503, + 152 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 506, + 133 + ], + "score": 1.0, + "content": "Ting-Kuei Hu, Tianlong Chen, Haotao Wang, and Zhangyang Wang. Triple wins: Boosting accuracy, robustness", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 114, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "and efficiency together by enabling input-adaptive inference. In International Conference on Learning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 141, + 453, + 153 + ], + "spans": [ + { + "bbox": [ + 116, + 141, + 389, + 153 + ], + "score": 1.0, + "content": "Representations, 2020. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 389, + 143, + 395, + 150 + ], + "score": 0.54, + "content": "\\equiv", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 141, + 453, + 153 + ], + "score": 1.0, + "content": "rJgzzJHtDB.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 120, + 506, + 153 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 159, + 504, + 190 + ], + "lines": [ + { + "bbox": [ + 107, + 160, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 505, + 170 + ], + "score": 1.0, + "content": "Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Weinberger. Multi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 168, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 168, + 505, + 182 + ], + "score": 1.0, + "content": "scale dense networks for resource efficient image classification. In International Conference on Learning", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 179, + 447, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 179, + 389, + 191 + ], + "score": 1.0, + "content": "Representations, 2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 389, + 181, + 395, + 188 + ], + "score": 0.56, + "content": "\\equiv", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 179, + 447, + 191 + ], + "score": 1.0, + "content": "Hk2aImxAb.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 107, + 160, + 505, + 191 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 197, + 506, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 209 + ], + "score": 1.0, + "content": "Nathan Inkawhich, Wei Wen, Hai (Helen) Li, and Yiran Chen. Feature space perturbations yield more transferable", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "adversarial examples. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 217, + 140, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 140, + 228 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 197, + 505, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 506, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "Weiwen Jiang, Edwin H.-M. Sha, Xinyi Zhang, Lei Yang, Qingfeng Zhuge, Yiyu Shi, and Jingtong Hu. Achieving", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 245, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 116, + 245, + 506, + 257 + ], + "score": 1.0, + "content": "super-linear speedup across multi-fpga for real-time dnn inference. ACM Trans. Embed. Comput. Syst., 18(5s),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 117, + 256, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 117, + 256, + 506, + 267 + ], + "score": 1.0, + "content": "October 2019. ISSN 1539-9087. doi: 10.1145/3358192. URL https://doi.org/10.1145/3358192.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 235, + 506, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 273, + 506, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 272, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 286 + ], + "score": 1.0, + "content": "Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 283, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 114, + 283, + 506, + 296 + ], + "score": 1.0, + "content": "Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. In Proceedings of the Twenty-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 292, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 506, + 307 + ], + "score": 1.0, + "content": "Second International Conference on Architectural Support for Programming Languages and Operating", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 303, + 507, + 316 + ], + "spans": [ + { + "bbox": [ + 114, + 303, + 507, + 316 + ], + "score": 1.0, + "content": "Systems, ASPLOS’17, pp. 615–629, New York, NY, USA, 2017. Association for Computing Machinery.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 313, + 507, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 313, + 507, + 325 + ], + "score": 1.0, + "content": "ISBN 9781450344654. doi: 10.1145/3037697.3037698. URL https://doi.org/10.1145/3037697.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 324, + 159, + 334 + ], + "spans": [ + { + "bbox": [ + 116, + 324, + 159, + 334 + ], + "score": 1.0, + "content": "3037698.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 272, + 507, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 341, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "Yigitcan Kaya, Sanghyun Hong, and Tudor Dumitra¸s. Shallow-Deep Networks: Understanding and mitigating ˘", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 352, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 352, + 506, + 363 + ], + "score": 1.0, + "content": "network overthinking. In Proceedings of the 2019 International Conference on Machine Learning (ICML),", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 361, + 219, + 373 + ], + "spans": [ + { + "bbox": [ + 115, + 361, + 219, + 373 + ], + "score": 1.0, + "content": "Long Beach, CA, Jun 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 340, + 506, + 373 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 379, + 484, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 484, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 484, + 392 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Geoffrey Hinton, et al. Learning Multiple Layers of Features from Tiny Images. 2009.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 379, + 484, + 392 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 398, + 504, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 407, + 434, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 434, + 420 + ], + "score": 1.0, + "content": "networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 398, + 505, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. arXiv preprint", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 436, + 209, + 447 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 209, + 447 + ], + "score": 1.0, + "content": "arXiv:1611.01236, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 425, + 505, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 454, + 504, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 454, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 505, + 465 + ], + "score": 1.0, + "content": "M. Lecuyer, V. Atlidakis, R. Geambasu, D. Hsu, and S. Jana. Certified robustness to adversarial examples with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 465, + 465, + 476 + ], + "spans": [ + { + "bbox": [ + 116, + 465, + 465, + 476 + ], + "score": 1.0, + "content": "differential privacy. In 2019 IEEE Symposium on Security and Privacy (SP), pp. 656–672, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 454, + 505, + 476 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 506, + 523 + ], + "lines": [ + { + "bbox": [ + 104, + 481, + 507, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 481, + 507, + 495 + ], + "score": 1.0, + "content": "En Li, Zhi Zhou, and Xu Chen. Edge intelligence: On-demand deep learning model co-inference with device-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 491, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 506, + 506 + ], + "score": 1.0, + "content": "edge synergy. In Proceedings of the 2018 Workshop on Mobile Edge Communications, MECOMM’18, pp.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 502, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 116, + 502, + 506, + 514 + ], + "score": 1.0, + "content": "31–36, New York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450359061. doi:", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 511, + 453, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 511, + 453, + 524 + ], + "score": 1.0, + "content": "10.1145/3229556.3229562. URL https://doi.org/10.1145/3229556.3229562.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 104, + 481, + 507, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 530, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 506, + 542 + ], + "score": 1.0, + "content": "Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 541, + 419, + 552 + ], + "spans": [ + { + "bbox": [ + 116, + 541, + 419, + 552 + ], + "score": 1.0, + "content": "CoRR, abs/1608.08710, 2016. URL http://arxiv.org/abs/1608.08710.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 531, + 506, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 558, + 504, + 590 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 506, + 570 + ], + "score": 1.0, + "content": "Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 569, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 569, + 505, + 580 + ], + "score": 1.0, + "content": "attacks using high-level representation guided denoiser. In The IEEE Conference on Computer Vision and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 579, + 264, + 591 + ], + "spans": [ + { + "bbox": [ + 116, + 579, + 264, + 591 + ], + "score": 1.0, + "content": "Pattern Recognition (CVPR), June 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 559, + 506, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 506, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 605, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 605, + 506, + 620 + ], + "score": 1.0, + "content": "black-box attacks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 615, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 114, + 615, + 507, + 630 + ], + "score": 1.0, + "content": "April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. URL https://openreview.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 627, + 240, + 639 + ], + "spans": [ + { + "bbox": [ + 115, + 627, + 181, + 639 + ], + "score": 1.0, + "content": "net/forum?id", + "type": "text" + }, + { + "bbox": [ + 182, + 628, + 187, + 636 + ], + "score": 0.52, + "content": "\\equiv", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 627, + 240, + 639 + ], + "score": 1.0, + "content": "Sys6GJqxl.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 596, + 507, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 645, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 658 + ], + "score": 1.0, + "content": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 654, + 437, + 666 + ], + "spans": [ + { + "bbox": [ + 115, + 654, + 437, + 666 + ], + "score": 1.0, + "content": "learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 643, + 506, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 673, + 506, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 683, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 683, + 506, + 695 + ], + "score": 1.0, + "content": "learning models resistant to adversarial attacks. In International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 693, + 385, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 693, + 385, + 704 + ], + "score": 1.0, + "content": "2018. URL https://openreview.net/forum?id=rJzIBfZAb.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47, + "bbox_fs": [ + 106, + 673, + 506, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 711, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "score": 1.0, + "content": "Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 115, + 721, + 496, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 496, + 733 + ], + "score": 1.0, + "content": "perturbations. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5, + "bbox_fs": [ + 106, + 712, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 114 + ], + "lines": [ + { + "bbox": [ + 104, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "N. Papernot, P. McDaniel, S. Jha, M. Fredrikson, Z. B. Celik, and A. Swami. The limitations of deep learning in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "adversarial settings. In 2016 IEEE European Symposium on Security and Privacy (EuroS P), pp. 372–387,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 140, + 114 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 140, + 114 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 122, + 503, + 154 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 504, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 504, + 134 + ], + "score": 1.0, + "content": "Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 133, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 117, + 133, + 505, + 144 + ], + "score": 1.0, + "content": "Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 334, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 334, + 154 + ], + "score": 1.0, + "content": "computer and communications security, pp. 506–519, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 108, + 162, + 504, + 193 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 505, + 173 + ], + "score": 1.0, + "content": "Se Jin Park, Murali Subramaniyam, Seoung Eun Kim, Seunghee Hong, Joo Hyeong Lee, Chan Min Jo, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 172, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 116, + 172, + 505, + 183 + ], + "score": 1.0, + "content": "Youngseob Seo. Development of the elderly healthcare monitoring system with iot. In Advances in Human", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 181, + 367, + 193 + ], + "spans": [ + { + "bbox": [ + 115, + 181, + 367, + 193 + ], + "score": 1.0, + "content": "Factors and Ergonomics in Healthcare, pp. 309–315. Springer, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 200, + 504, + 232 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "Jérôme Rony, Luiz G Hafemann, Luiz S Oliveira, Ismail Ben Ayed, Robert Sabourin, and Eric Granger.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "Decoupling direction and norm for efficient gradient-based l2 adversarial attacks and defenses. In Proceedings", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 221, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 116, + 221, + 451, + 232 + ], + "score": 1.0, + "content": "of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4322–4330, 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 253 + ], + "score": 1.0, + "content": "Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 249, + 141, + 262 + ], + "spans": [ + { + "bbox": [ + 114, + 249, + 141, + 262 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 504, + 301 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 283 + ], + "score": 1.0, + "content": "Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 280, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 506, + 292 + ], + "score": 1.0, + "content": "generative models to understand and defend against adversarial examples. In International Conference on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 289, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 115, + 289, + 424, + 302 + ], + "score": 1.0, + "content": "Learning Representations, 2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 424, + 291, + 430, + 298 + ], + "score": 0.38, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 289, + 482, + 302 + ], + "score": 1.0, + "content": "rJUYGxbCW.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 309, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 506, + 321 + ], + "score": 1.0, + "content": "C. Szegedy, Wei Liu, Yangqing Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Ra-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "binovich. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 329, + 249, + 341 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 249, + 341 + ], + "score": 1.0, + "content": "Recognition (CVPR), pp. 1–9, 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 107, + 349, + 504, + 359 + ], + "spans": [ + { + "bbox": [ + 107, + 349, + 504, + 359 + ], + "score": 1.0, + "content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 115, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 368, + 326, + 380 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 326, + 380 + ], + "score": 1.0, + "content": "2014. URL http://arxiv.org/abs/1312.6199.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 506, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "score": 1.0, + "content": "Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 6105–6114, Long Beach,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 416, + 508, + 430 + ], + "spans": [ + { + "bbox": [ + 114, + 416, + 508, + 430 + ], + "score": 1.0, + "content": "California, USA, 09–15 Jun 2019. PMLR. URL http://proceedings.mlr.press/v97/tan19a.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 427, + 144, + 439 + ], + "spans": [ + { + "bbox": [ + 115, + 427, + 144, + 439 + ], + "score": 1.0, + "content": "html.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 447, + 504, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "Ben Taylor, Vicent Sanz Marco, Willy Wolff, Yehia Elkhatib, and Zheng Wang. Adaptive deep learning model", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 457, + 397, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 457, + 397, + 468 + ], + "score": 1.0, + "content": "selection on embedded systems. ACM SIGPLAN Notices, 53(6):31–43, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 476, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "S. Teerapittayanon, B. McDanel, and H. T. Kung. Branchynet: Fast inference via early exiting from deep neural", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 487, + 492, + 498 + ], + "spans": [ + { + "bbox": [ + 116, + 487, + 492, + 498 + ], + "score": 1.0, + "content": "networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pp. 2464–2469, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "Tiny. Tiny ImageNet Visual Recognition Challenge. http://tiny-imagenet.herokuapp.com/.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 515, + 202, + 527 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 202, + 527 + ], + "score": 1.0, + "content": "Accessed: 2020-09-28.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 504, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "Florian Tramèr and Dan Boneh. Adversarial training and robustness for multiple perturbations. In Advances in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 545, + 344, + 556 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 344, + 556 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 5866–5876, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 504, + 586 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 576 + ], + "score": 1.0, + "content": "Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 574, + 347, + 586 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 347, + 586 + ], + "score": 1.0, + "content": "adversarial examples. arXiv preprint arXiv:1704.03453, 2017a.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 504, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 605 + ], + "score": 1.0, + "content": "Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 604, + 347, + 615 + ], + "spans": [ + { + "bbox": [ + 116, + 604, + 347, + 615 + ], + "score": 1.0, + "content": "adversarial examples. arXiv preprint arXiv:1704.03453, 2017b.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 623, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E. Gonzalez. SkipNet: Learning Dynamic Routing", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 633, + 495, + 645 + ], + "spans": [ + { + "bbox": [ + 116, + 633, + 495, + 645 + ], + "score": 1.0, + "content": "in Convolutional Networks. In The European Conference on Computer Vision (ECCV), September 2018.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 652, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 652, + 505, + 664 + ], + "score": 1.0, + "content": "Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 662, + 302, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 302, + 673 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1704.01155, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 106, + 681, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 681, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 504, + 693 + ], + "score": 1.0, + "content": "Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 692, + 331, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 692, + 331, + 703 + ], + "score": 1.0, + "content": "matrix estimation. arXiv preprint arXiv:1905.11971, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 105, + 711, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "score": 1.0, + "content": "Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 721, + 437, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 437, + 733 + ], + "score": 1.0, + "content": "networks? In Advances in neural information processing systems, pp. 3320–3328, 2014.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 114 + ], + "lines": [ + { + "bbox": [ + 104, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "N. Papernot, P. McDaniel, S. Jha, M. Fredrikson, Z. B. Celik, and A. Swami. The limitations of deep learning in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "adversarial settings. In 2016 IEEE European Symposium on Security and Privacy (EuroS P), pp. 372–387,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 140, + 114 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 140, + 114 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 104, + 82, + 506, + 114 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 122, + 503, + 154 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 504, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 504, + 134 + ], + "score": 1.0, + "content": "Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 133, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 117, + 133, + 505, + 144 + ], + "score": 1.0, + "content": "Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 334, + 154 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 334, + 154 + ], + "score": 1.0, + "content": "computer and communications security, pp. 506–519, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 122, + 505, + 154 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 162, + 504, + 193 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 505, + 173 + ], + "score": 1.0, + "content": "Se Jin Park, Murali Subramaniyam, Seoung Eun Kim, Seunghee Hong, Joo Hyeong Lee, Chan Min Jo, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 172, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 116, + 172, + 505, + 183 + ], + "score": 1.0, + "content": "Youngseob Seo. Development of the elderly healthcare monitoring system with iot. In Advances in Human", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 181, + 367, + 193 + ], + "spans": [ + { + "bbox": [ + 115, + 181, + 367, + 193 + ], + "score": 1.0, + "content": "Factors and Ergonomics in Healthcare, pp. 309–315. Springer, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 162, + 505, + 193 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 200, + 504, + 232 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "Jérôme Rony, Luiz G Hafemann, Luiz S Oliveira, Ismail Ben Ayed, Robert Sabourin, and Eric Granger.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "Decoupling direction and norm for efficient gradient-based l2 adversarial attacks and defenses. In Proceedings", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 221, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 116, + 221, + 451, + 232 + ], + "score": 1.0, + "content": "of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4322–4330, 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 201, + 505, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 253 + ], + "score": 1.0, + "content": "Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 249, + 141, + 262 + ], + "spans": [ + { + "bbox": [ + 114, + 249, + 141, + 262 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 239, + 506, + 262 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 504, + 301 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 283 + ], + "score": 1.0, + "content": "Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 280, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 506, + 292 + ], + "score": 1.0, + "content": "generative models to understand and defend against adversarial examples. In International Conference on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 289, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 115, + 289, + 424, + 302 + ], + "score": 1.0, + "content": "Learning Representations, 2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 424, + 291, + 430, + 298 + ], + "score": 0.38, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 289, + 482, + 302 + ], + "score": 1.0, + "content": "rJUYGxbCW.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 268, + 506, + 302 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 309, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 506, + 321 + ], + "score": 1.0, + "content": "C. Szegedy, Wei Liu, Yangqing Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Ra-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "binovich. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 329, + 249, + 341 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 249, + 341 + ], + "score": 1.0, + "content": "Recognition (CVPR), pp. 1–9, 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 309, + 506, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 107, + 349, + 504, + 359 + ], + "spans": [ + { + "bbox": [ + 107, + 349, + 504, + 359 + ], + "score": 1.0, + "content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 115, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 368, + 326, + 380 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 326, + 380 + ], + "score": 1.0, + "content": "2014. URL http://arxiv.org/abs/1312.6199.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 107, + 349, + 506, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 506, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 399 + ], + "score": 1.0, + "content": "Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 6105–6114, Long Beach,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 416, + 508, + 430 + ], + "spans": [ + { + "bbox": [ + 114, + 416, + 508, + 430 + ], + "score": 1.0, + "content": "California, USA, 09–15 Jun 2019. PMLR. URL http://proceedings.mlr.press/v97/tan19a.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 427, + 144, + 439 + ], + "spans": [ + { + "bbox": [ + 115, + 427, + 144, + 439 + ], + "score": 1.0, + "content": "html.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 388, + 508, + 439 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 447, + 504, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "Ben Taylor, Vicent Sanz Marco, Willy Wolff, Yehia Elkhatib, and Zheng Wang. Adaptive deep learning model", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 457, + 397, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 457, + 397, + 468 + ], + "score": 1.0, + "content": "selection on embedded systems. ACM SIGPLAN Notices, 53(6):31–43, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 447, + 505, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 476, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "S. Teerapittayanon, B. McDanel, and H. T. Kung. Branchynet: Fast inference via early exiting from deep neural", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 487, + 492, + 498 + ], + "spans": [ + { + "bbox": [ + 116, + 487, + 492, + 498 + ], + "score": 1.0, + "content": "networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pp. 2464–2469, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 476, + 505, + 498 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "Tiny. Tiny ImageNet Visual Recognition Challenge. http://tiny-imagenet.herokuapp.com/.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 515, + 202, + 527 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 202, + 527 + ], + "score": 1.0, + "content": "Accessed: 2020-09-28.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 505, + 506, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 504, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "Florian Tramèr and Dan Boneh. Adversarial training and robustness for multiple perturbations. In Advances in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 545, + 344, + 556 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 344, + 556 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 5866–5876, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 533, + 506, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 504, + 586 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 576 + ], + "score": 1.0, + "content": "Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 574, + 347, + 586 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 347, + 586 + ], + "score": 1.0, + "content": "adversarial examples. arXiv preprint arXiv:1704.03453, 2017a.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 564, + 505, + 586 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 504, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 605 + ], + "score": 1.0, + "content": "Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 604, + 347, + 615 + ], + "spans": [ + { + "bbox": [ + 116, + 604, + 347, + 615 + ], + "score": 1.0, + "content": "adversarial examples. arXiv preprint arXiv:1704.03453, 2017b.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 594, + 505, + 615 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 623, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E. Gonzalez. SkipNet: Learning Dynamic Routing", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 633, + 495, + 645 + ], + "spans": [ + { + "bbox": [ + 116, + 633, + 495, + 645 + ], + "score": 1.0, + "content": "in Convolutional Networks. In The European Conference on Computer Vision (ECCV), September 2018.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 622, + 505, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 652, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 652, + 505, + 664 + ], + "score": 1.0, + "content": "Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 662, + 302, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 302, + 673 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1704.01155, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 652, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 681, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 681, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 504, + 693 + ], + "score": 1.0, + "content": "Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 692, + 331, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 692, + 331, + 703 + ], + "score": 1.0, + "content": "matrix estimation. arXiv preprint arXiv:1905.11971, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 106, + 681, + 504, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 711, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 505, + 723 + ], + "score": 1.0, + "content": "Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 721, + 437, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 437, + 733 + ], + "score": 1.0, + "content": "networks? In Advances in neural information processing systems, pp. 3320–3328, 2014.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 106, + 712, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Li Zhou, Hao Wen, Radu Teodorescu, and David H.C. Du. Distributing deep neural networks with containerized", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 117, + 93, + 506, + 105 + ], + "score": 1.0, + "content": "partitions at the edge. In 2nd USENIX Workshop on Hot Topics in Edge Computing (HotEdge 19), Renton, WA,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 506, + 115 + ], + "score": 1.0, + "content": "July 2019. USENIX Association. URL https://www.usenix.org/conference/hotedge19/", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 113, + 213, + 125 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 213, + 125 + ], + "score": 1.0, + "content": "presentation/zhou.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 131, + 504, + 151 + ], + "lines": [ + { + "bbox": [ + 106, + 130, + 506, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 506, + 143 + ], + "score": 1.0, + "content": "Wangchunshu Zhou, Canwen Xu, Tao Ge, Julian McAuley, Ke Xu, and Furu Wei. BERT Loses Patience: Fast", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 141, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 117, + 141, + 479, + 153 + ], + "score": 1.0, + "content": "and Robust Inference with Early Exit. In Advances in Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Li Zhou, Hao Wen, Radu Teodorescu, and David H.C. Du. Distributing deep neural networks with containerized", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 117, + 93, + 506, + 105 + ], + "score": 1.0, + "content": "partitions at the edge. In 2nd USENIX Workshop on Hot Topics in Edge Computing (HotEdge 19), Renton, WA,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 102, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 506, + 115 + ], + "score": 1.0, + "content": "July 2019. USENIX Association. URL https://www.usenix.org/conference/hotedge19/", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 113, + 213, + 125 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 213, + 125 + ], + "score": 1.0, + "content": "presentation/zhou.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 83, + 506, + 125 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 131, + 504, + 151 + ], + "lines": [ + { + "bbox": [ + 106, + 130, + 506, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 506, + 143 + ], + "score": 1.0, + "content": "Wangchunshu Zhou, Canwen Xu, Tao Ge, Julian McAuley, Ke Xu, and Furu Wei. BERT Loses Patience: Fast", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 141, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 117, + 141, + 479, + 153 + ], + "score": 1.0, + "content": "and Robust Inference with Early Exit. In Advances in Neural Information Processing Systems, 2020.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 130, + 506, + 153 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 81, + 255, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 257, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 257, + 96 + ], + "score": 1.0, + "content": "A MOTIVATING EXAMPLES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 105, + 492, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 493, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 493, + 119 + ], + "score": 1.0, + "content": "Here, we discuss two exemplary scenarios where an adversary can exploit the slowdown attacks.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 506, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "• (Case 1) Attacks on cloud-based IoT applications. In most cases, cloud-based IoT applications,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 137, + 507, + 149 + ], + "spans": [ + { + "bbox": [ + 113, + 137, + 507, + 149 + ], + "score": 1.0, + "content": "such as Apple Siri, Google Now, or Microsoft Cortana, run their DNN inferences in the cloud.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 148, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 114, + 148, + 506, + 160 + ], + "score": 1.0, + "content": "This cloud-only approach puts all the computational burden on cloud servers and increases the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 114, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "communications between the servers and IoT devices. In consequence, recent work (Kang et al.,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 169, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 114, + 169, + 506, + 182 + ], + "score": 1.0, + "content": "2017; Li et al., 2018; Zhou et al., 2019) utilizes multi-exit architectures for bringing computationally", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 114, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "expensive models, e.g. language models (Zhou et al., 2020; Hou et al., 2020), in the cloud to IoT", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 191, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 114, + 191, + 506, + 204 + ], + "score": 1.0, + "content": "(or mobile) devices. They split a multi-exit model into two partitions and deploy each of them", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 203, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 203, + 505, + 214 + ], + "score": 1.0, + "content": "to a server and IoT devices, respectively. Under this scheme, the cloud server only takes care of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "complex inputs that the shallow partition cannot correctly classify at the edge. As a result, one can", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 224, + 503, + 237 + ], + "spans": [ + { + "bbox": [ + 113, + 224, + 503, + 237 + ], + "score": 1.0, + "content": "reduce the computations in the cloud and decrease communications between the cloud and edge.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 112, + 237, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 113, + 235, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 113, + 235, + 506, + 251 + ], + "score": 1.0, + "content": "On the other hand, our adversary, by applying human-imperceptible perturbations, can convert", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 248, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 114, + 248, + 506, + 261 + ], + "score": 1.0, + "content": "simple inputs into complex inputs. These adversarial inputs will bypass early-exits and, as a result,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 259, + 492, + 272 + ], + "spans": [ + { + "bbox": [ + 114, + 259, + 492, + 272 + ], + "score": 1.0, + "content": "reduce (or even offset) the computational and communication savings provided by prior work.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 113, + 272, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 114, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 114, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "Here, a defender may deploy DoS defenses such as firewalls or rate-limiting. In this setting, the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 114, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "attacker may not cause DoS because defenses keep the communications between the server and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 113, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 113, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "IoT devices under a certain-level. Nevertheless, the attacker still increases: (i) the computations", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 304, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 114, + 304, + 506, + 317 + ], + "score": 1.0, + "content": "at the edge (by making inputs skip early-exits) and (ii) the number of samples that cloud servers", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 113, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 113, + 315, + 329, + 328 + ], + "score": 1.0, + "content": "process. Recall that a VGG-16 SDN model classifies", + "type": "text" + }, + { + "bbox": [ + 329, + 315, + 349, + 326 + ], + "score": 0.87, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "of clean CIFAR-10 instances correctly", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 114, + 326, + 505, + 339 + ], + "score": 1.0, + "content": "at the first exit. If the adversarial examples crafted by the attacker bypass only the first exit, one", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 337, + 496, + 350 + ], + "spans": [ + { + "bbox": [ + 114, + 337, + 496, + 350 + ], + "score": 1.0, + "content": "can easily increase the computations on IoT devices and make them send requests to the cloud.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 506, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 507, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 507, + 364 + ], + "score": 1.0, + "content": "• (Case 2) Attacks on real-time DNN inference for resource- and time-constrained scenarios.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 114, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "Recent work on the real-time systems (Hu et al., 2019; Jiang et al., 2019) harnesses multi-exit", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 374, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 114, + 374, + 506, + 387 + ], + "score": 1.0, + "content": "architectures and model partitioning as a solution to optimize real-time DNN inference for resource-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 114, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "and time-constrained scenarios. Hu et al. (2019) showed a real-world prototype of an optimal", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 114, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "model partitioning, which is based on a self-driving car video dataset, can improve latency and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 406, + 443, + 419 + ], + "spans": [ + { + "bbox": [ + 114, + 406, + 351, + 419 + ], + "score": 1.0, + "content": "throughput of partitioned models on the cloud and edge by", + "type": "text" + }, + { + "bbox": [ + 351, + 407, + 388, + 418 + ], + "score": 0.88, + "content": "6 . 5 – 1 4 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 406, + 443, + 419 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 110, + 419, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 114, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 114, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "However, the prior work does not consider the danger of slowdown attacks; our threat model has", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 114, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "not been discussed before in the literature. Our results in Sec 5 suggest that slowdown can be", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 114, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "induced adversarially, potentially violating real-time guarantees. For example, our attacker can", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 114, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 114, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "force partitioned models on the cloud and edge to use maximal computations for inference. Further,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 114, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "the same adversarial examples also require the inference results from the model running on the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 114, + 474, + 426, + 486 + ], + "score": 1.0, + "content": "cloud, which potentially increases the response time of the edge devices by", + "type": "text" + }, + { + "bbox": [ + 426, + 474, + 458, + 485 + ], + "score": 0.86, + "content": "1 . 5 – 5 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 474, + 505, + 486 + ], + "score": 1.0, + "content": ". Our work", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 114, + 484, + 459, + 498 + ], + "spans": [ + { + "bbox": [ + 114, + 484, + 459, + 498 + ], + "score": 1.0, + "content": "showed that multi-exit architectures should be used with caution in real-time systems.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 108, + 511, + 232, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 234, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 234, + 525 + ], + "score": 1.0, + "content": "B HYPERPARAMETERS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 105, + 536, + 474, + 547 + ], + "lines": [ + { + "bbox": [ + 104, + 534, + 477, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 534, + 477, + 550 + ], + "score": 1.0, + "content": "In our experiments, we use the following hyperparameters to craft adversarial perturbations.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 107, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 107, + 554, + 120, + 565 + ], + "score": 0.88, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 553, + 253, + 566 + ], + "score": 1.0, + "content": "-based DeepSloth. We find that", + "type": "text" + }, + { + "bbox": [ + 253, + 554, + 266, + 565 + ], + "score": 0.91, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "-based DeepSloth does not require careful tuning. For the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 425, + 577 + ], + "score": 1.0, + "content": "standard attack, we set the total number of iterations to 30 and the step size to", + "type": "text" + }, + { + "bbox": [ + 425, + 565, + 469, + 575 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 2", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 565, + 505, + 577 + ], + "score": 1.0, + "content": ". For the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "score": 1.0, + "content": "modified attacks for hurting or preserving the accuracy, we set the total number of iterations to 75", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 587, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 185, + 599 + ], + "score": 1.0, + "content": "and the step size to", + "type": "text" + }, + { + "bbox": [ + 185, + 587, + 228, + 597 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 587, + 457, + 599 + ], + "score": 1.0, + "content": ". We compute the standard perturbations using the entire", + "type": "text" + }, + { + "bbox": [ + 457, + 587, + 473, + 597 + ], + "score": 0.28, + "content": "1 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 587, + 506, + 599 + ], + "score": 1.0, + "content": "test-set", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 598, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 610 + ], + "score": 1.0, + "content": "samples in CIFAR-10 and Tiny Imagenet. For the universal variants, we set the total number of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 304, + 621 + ], + "score": 1.0, + "content": "iterations to 12 and reduce the initial step size of", + "type": "text" + }, + { + "bbox": [ + 304, + 609, + 348, + 619 + ], + "score": 0.87, + "content": "\\alpha = 0 . 0 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "by a factor of 10 every 4 iterations. To", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "score": 1.0, + "content": "compute a universal perturbation, we use randomly chosen 250 (CIFAR-10) and 200 (Tiny Imagenet)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 631, + 177, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 177, + 643 + ], + "score": 1.0, + "content": "training samples.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 108, + 648, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 116, + 659 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "-based DeepSloth. For both the standard and universal attacks, we set the total number of iterations", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 203, + 672 + ], + "score": 1.0, + "content": "to 550 and the step size", + "type": "text" + }, + { + "bbox": [ + 204, + 661, + 211, + 671 + ], + "score": 0.8, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 659, + 369, + 672 + ], + "score": 1.0, + "content": "to 0.1. Our initial perturbation has the", + "type": "text" + }, + { + "bbox": [ + 369, + 659, + 379, + 670 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "-norm of 1.0. Here, we use the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 434, + 682 + ], + "score": 1.0, + "content": "same number of samples for crafting the standard and universal perturbations as the", + "type": "text" + }, + { + "bbox": [ + 434, + 670, + 448, + 681 + ], + "score": 0.9, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 670, + 506, + 682 + ], + "score": 1.0, + "content": "-based attacks.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 116, + 699 + ], + "score": 0.85, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 687, + 263, + 699 + ], + "score": 1.0, + "content": "-based DeepSloth. For our standard", + "type": "text" + }, + { + "bbox": [ + 263, + 688, + 273, + 699 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 687, + 506, + 699 + ], + "score": 1.0, + "content": "-based DeepSloth, we set the total number of iterations to", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 177, + 711 + ], + "score": 1.0, + "content": "250, the step size", + "type": "text" + }, + { + "bbox": [ + 177, + 700, + 185, + 709 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "to 0.5, and the gradient sparsity to 99. For the universal variants, we reduce the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "total number of iterations to 100 and set the gradient sparsity to 90. Other hyperparameters remain", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 719, + 497, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 320, + 734 + ], + "score": 1.0, + "content": "the same. We use the same number of samples as the", + "type": "text" + }, + { + "bbox": [ + 320, + 721, + 333, + 732 + ], + "score": 0.9, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 719, + 497, + 734 + ], + "score": 1.0, + "content": "-based attacks, to craft the perturbations.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 81, + 255, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 257, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 257, + 96 + ], + "score": 1.0, + "content": "A MOTIVATING EXAMPLES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 105, + 492, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 493, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 493, + 119 + ], + "score": 1.0, + "content": "Here, we discuss two exemplary scenarios where an adversary can exploit the slowdown attacks.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 105, + 493, + 119 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 506, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "• (Case 1) Attacks on cloud-based IoT applications. In most cases, cloud-based IoT applications,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 137, + 507, + 149 + ], + "spans": [ + { + "bbox": [ + 113, + 137, + 507, + 149 + ], + "score": 1.0, + "content": "such as Apple Siri, Google Now, or Microsoft Cortana, run their DNN inferences in the cloud.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 148, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 114, + 148, + 506, + 160 + ], + "score": 1.0, + "content": "This cloud-only approach puts all the computational burden on cloud servers and increases the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 159, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 114, + 159, + 506, + 171 + ], + "score": 1.0, + "content": "communications between the servers and IoT devices. In consequence, recent work (Kang et al.,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 169, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 114, + 169, + 506, + 182 + ], + "score": 1.0, + "content": "2017; Li et al., 2018; Zhou et al., 2019) utilizes multi-exit architectures for bringing computationally", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 114, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "expensive models, e.g. language models (Zhou et al., 2020; Hou et al., 2020), in the cloud to IoT", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 191, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 114, + 191, + 506, + 204 + ], + "score": 1.0, + "content": "(or mobile) devices. They split a multi-exit model into two partitions and deploy each of them", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 203, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 203, + 505, + 214 + ], + "score": 1.0, + "content": "to a server and IoT devices, respectively. Under this scheme, the cloud server only takes care of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "complex inputs that the shallow partition cannot correctly classify at the edge. As a result, one can", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 224, + 503, + 237 + ], + "spans": [ + { + "bbox": [ + 113, + 224, + 503, + 237 + ], + "score": 1.0, + "content": "reduce the computations in the cloud and decrease communications between the cloud and edge.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 125, + 507, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 237, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 113, + 235, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 113, + 235, + 506, + 251 + ], + "score": 1.0, + "content": "On the other hand, our adversary, by applying human-imperceptible perturbations, can convert", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 248, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 114, + 248, + 506, + 261 + ], + "score": 1.0, + "content": "simple inputs into complex inputs. These adversarial inputs will bypass early-exits and, as a result,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 259, + 492, + 272 + ], + "spans": [ + { + "bbox": [ + 114, + 259, + 492, + 272 + ], + "score": 1.0, + "content": "reduce (or even offset) the computational and communication savings provided by prior work.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 113, + 235, + 506, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 272, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 114, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 114, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "Here, a defender may deploy DoS defenses such as firewalls or rate-limiting. In this setting, the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 114, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "attacker may not cause DoS because defenses keep the communications between the server and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 113, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 113, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "IoT devices under a certain-level. Nevertheless, the attacker still increases: (i) the computations", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 304, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 114, + 304, + 506, + 317 + ], + "score": 1.0, + "content": "at the edge (by making inputs skip early-exits) and (ii) the number of samples that cloud servers", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 113, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 113, + 315, + 329, + 328 + ], + "score": 1.0, + "content": "process. Recall that a VGG-16 SDN model classifies", + "type": "text" + }, + { + "bbox": [ + 329, + 315, + 349, + 326 + ], + "score": 0.87, + "content": "90 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "of clean CIFAR-10 instances correctly", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 114, + 326, + 505, + 339 + ], + "score": 1.0, + "content": "at the first exit. If the adversarial examples crafted by the attacker bypass only the first exit, one", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 337, + 496, + 350 + ], + "spans": [ + { + "bbox": [ + 114, + 337, + 496, + 350 + ], + "score": 1.0, + "content": "can easily increase the computations on IoT devices and make them send requests to the cloud.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18, + "bbox_fs": [ + 113, + 271, + 506, + 350 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 506, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 507, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 507, + 364 + ], + "score": 1.0, + "content": "• (Case 2) Attacks on real-time DNN inference for resource- and time-constrained scenarios.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 114, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "Recent work on the real-time systems (Hu et al., 2019; Jiang et al., 2019) harnesses multi-exit", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 374, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 114, + 374, + 506, + 387 + ], + "score": 1.0, + "content": "architectures and model partitioning as a solution to optimize real-time DNN inference for resource-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 114, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "and time-constrained scenarios. Hu et al. (2019) showed a real-world prototype of an optimal", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 114, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "model partitioning, which is based on a self-driving car video dataset, can improve latency and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 406, + 443, + 419 + ], + "spans": [ + { + "bbox": [ + 114, + 406, + 351, + 419 + ], + "score": 1.0, + "content": "throughput of partitioned models on the cloud and edge by", + "type": "text" + }, + { + "bbox": [ + 351, + 407, + 388, + 418 + ], + "score": 0.88, + "content": "6 . 5 – 1 4 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 406, + 443, + 419 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 351, + 507, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 419, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 114, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 114, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "However, the prior work does not consider the danger of slowdown attacks; our threat model has", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 114, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "not been discussed before in the literature. Our results in Sec 5 suggest that slowdown can be", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 114, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "induced adversarially, potentially violating real-time guarantees. For example, our attacker can", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 114, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 114, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "force partitioned models on the cloud and edge to use maximal computations for inference. Further,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 114, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "the same adversarial examples also require the inference results from the model running on the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 114, + 474, + 426, + 486 + ], + "score": 1.0, + "content": "cloud, which potentially increases the response time of the edge devices by", + "type": "text" + }, + { + "bbox": [ + 426, + 474, + 458, + 485 + ], + "score": 0.86, + "content": "1 . 5 – 5 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 474, + 505, + 486 + ], + "score": 1.0, + "content": ". Our work", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 114, + 484, + 459, + 498 + ], + "spans": [ + { + "bbox": [ + 114, + 484, + 459, + 498 + ], + "score": 1.0, + "content": "showed that multi-exit architectures should be used with caution in real-time systems.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31, + "bbox_fs": [ + 114, + 419, + 506, + 498 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 511, + 232, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 511, + 234, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 234, + 525 + ], + "score": 1.0, + "content": "B HYPERPARAMETERS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 105, + 536, + 474, + 547 + ], + "lines": [ + { + "bbox": [ + 104, + 534, + 477, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 534, + 477, + 550 + ], + "score": 1.0, + "content": "In our experiments, we use the following hyperparameters to craft adversarial perturbations.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36, + "bbox_fs": [ + 104, + 534, + 477, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 107, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 107, + 554, + 120, + 565 + ], + "score": 0.88, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 553, + 253, + 566 + ], + "score": 1.0, + "content": "-based DeepSloth. We find that", + "type": "text" + }, + { + "bbox": [ + 253, + 554, + 266, + 565 + ], + "score": 0.91, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "-based DeepSloth does not require careful tuning. For the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 425, + 577 + ], + "score": 1.0, + "content": "standard attack, we set the total number of iterations to 30 and the step size to", + "type": "text" + }, + { + "bbox": [ + 425, + 565, + 469, + 575 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 2", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 565, + 505, + 577 + ], + "score": 1.0, + "content": ". For the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "score": 1.0, + "content": "modified attacks for hurting or preserving the accuracy, we set the total number of iterations to 75", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 587, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 185, + 599 + ], + "score": 1.0, + "content": "and the step size to", + "type": "text" + }, + { + "bbox": [ + 185, + 587, + 228, + 597 + ], + "score": 0.89, + "content": "\\alpha = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 587, + 457, + 599 + ], + "score": 1.0, + "content": ". We compute the standard perturbations using the entire", + "type": "text" + }, + { + "bbox": [ + 457, + 587, + 473, + 597 + ], + "score": 0.28, + "content": "1 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 587, + 506, + 599 + ], + "score": 1.0, + "content": "test-set", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 598, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 610 + ], + "score": 1.0, + "content": "samples in CIFAR-10 and Tiny Imagenet. For the universal variants, we set the total number of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 304, + 621 + ], + "score": 1.0, + "content": "iterations to 12 and reduce the initial step size of", + "type": "text" + }, + { + "bbox": [ + 304, + 609, + 348, + 619 + ], + "score": 0.87, + "content": "\\alpha = 0 . 0 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "by a factor of 10 every 4 iterations. To", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "score": 1.0, + "content": "compute a universal perturbation, we use randomly chosen 250 (CIFAR-10) and 200 (Tiny Imagenet)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 631, + 177, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 177, + 643 + ], + "score": 1.0, + "content": "training samples.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 553, + 506, + 643 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 648, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 116, + 659 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "-based DeepSloth. For both the standard and universal attacks, we set the total number of iterations", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 203, + 672 + ], + "score": 1.0, + "content": "to 550 and the step size", + "type": "text" + }, + { + "bbox": [ + 204, + 661, + 211, + 671 + ], + "score": 0.8, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 659, + 369, + 672 + ], + "score": 1.0, + "content": "to 0.1. Our initial perturbation has the", + "type": "text" + }, + { + "bbox": [ + 369, + 659, + 379, + 670 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "-norm of 1.0. Here, we use the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 670, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 434, + 682 + ], + "score": 1.0, + "content": "same number of samples for crafting the standard and universal perturbations as the", + "type": "text" + }, + { + "bbox": [ + 434, + 670, + 448, + 681 + ], + "score": 0.9, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 670, + 506, + 682 + ], + "score": 1.0, + "content": "-based attacks.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 647, + 506, + 682 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 116, + 699 + ], + "score": 0.85, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 687, + 263, + 699 + ], + "score": 1.0, + "content": "-based DeepSloth. For our standard", + "type": "text" + }, + { + "bbox": [ + 263, + 688, + 273, + 699 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 687, + 506, + 699 + ], + "score": 1.0, + "content": "-based DeepSloth, we set the total number of iterations to", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 177, + 711 + ], + "score": 1.0, + "content": "250, the step size", + "type": "text" + }, + { + "bbox": [ + 177, + 700, + 185, + 709 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "to 0.5, and the gradient sparsity to 99. For the universal variants, we reduce the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "total number of iterations to 100 and set the gradient sparsity to 90. Other hyperparameters remain", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 719, + 497, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 320, + 734 + ], + "score": 1.0, + "content": "the same. We use the same number of samples as the", + "type": "text" + }, + { + "bbox": [ + 320, + 721, + 333, + 732 + ], + "score": 0.9, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 719, + 497, + 734 + ], + "score": 1.0, + "content": "-based attacks, to craft the perturbations.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 687, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 389, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 391, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 274, + 96 + ], + "score": 1.0, + "content": "C EMPIRICAL EVALUATION OF", + "type": "text" + }, + { + "bbox": [ + 274, + 82, + 285, + 94 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 79, + 312, + 96 + ], + "score": 1.0, + "content": "AND", + "type": "text" + }, + { + "bbox": [ + 312, + 81, + 324, + 94 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 79, + 391, + 96 + ], + "score": 1.0, + "content": "DEEPSLOTH", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 505, + 123 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 290, + 125 + ], + "score": 1.0, + "content": "Table 4 and Table 5 shows the effectiveness of", + "type": "text" + }, + { + "bbox": [ + 290, + 112, + 299, + 123 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 109, + 343, + 125 + ], + "score": 1.0, + "content": "-based and", + "type": "text" + }, + { + "bbox": [ + 343, + 112, + 353, + 123 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 109, + 506, + 125 + ], + "score": 1.0, + "content": "-based DeepSloth attacks, respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 106, + 185, + 504, + 370 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 142, + 505, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 223, + 155 + ], + "score": 1.0, + "content": "Table 4: The effectiveness of", + "type": "text" + }, + { + "bbox": [ + 224, + 143, + 233, + 154 + ], + "score": 0.86, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 142, + 287, + 155 + ], + "score": 1.0, + "content": "DeepSloth. ‘", + "type": "text" + }, + { + "bbox": [ + 288, + 143, + 342, + 153 + ], + "score": 0.77, + "content": "R \\mathrm { A D } { < } 5 , 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "’ columns list the results in each early-exit", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 164, + 471, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 390, + 177 + ], + "score": 1.0, + "content": "attack’s results are an average of 10 classes. ‘TI’ is Tiny Imagenet and", + "type": "text" + }, + { + "bbox": [ + 390, + 165, + 413, + 175 + ], + "score": 0.7, + "content": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 164, + 471, + 177 + ], + "score": 1.0, + "content": "is CIFAR-10.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 106, + 185, + 504, + 370 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 185, + 504, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 504, + 370 + ], + "score": 0.983, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 /89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 / 57%0.51 /52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.36 / 51%0.35 / 51%0.12 /36%0.34 /45%0.18 / 41%0.49 / 53%
TI0.23 /37%0.51 / 40%0.08 /22%0.15 / 25%0.08 /33%0.19 /35%
UNIVERSAL DEEPSLOTH
C100.89 /83%0.89 /83%0.75 /85%0.88 / 75%0.82 /85%0.92 /77%
TI0.64 / 55%0.83/50%0.38 /57%0.51 / 52%0.41 / 57%0.59 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88/ 73%0.88 / 73%0.69 / 78%0.86 /67%0.76 /74%0.89 / 65%
TI0.64 /54%0.83 /49%0.39 / 59%0.50 / 58%0.41 / 60%0.58 / 53%
", + "type": "table", + "image_path": "6b9f5d0671919db5ee08c6a0cfa8dee4ae1967265bbd425b6363d71e75af115c.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 106, + 185, + 504, + 246.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 246.66666666666666, + 504, + 308.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 308.3333333333333, + 504, + 370.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 4.5 + }, + { + "type": "table", + "bbox": [ + 106, + 452, + 504, + 637 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 410, + 505, + 444 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 223, + 422 + ], + "score": 1.0, + "content": "Table 5: The effectiveness of", + "type": "text" + }, + { + "bbox": [ + 224, + 410, + 234, + 421 + ], + "score": 0.85, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 410, + 286, + 422 + ], + "score": 1.0, + "content": "DeepSloth.", + "type": "text" + }, + { + "bbox": [ + 286, + 410, + 342, + 421 + ], + "score": 0.83, + "content": "\\cdot \\mathrm { R A D } { < } 5 , 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 410, + 505, + 422 + ], + "score": 1.0, + "content": "’ columns list the results in each early-exit", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 432, + 470, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 389, + 445 + ], + "score": 1.0, + "content": "attack’s results are an average of 10 classes. ‘TI’ is Tiny Imagenet and", + "type": "text" + }, + { + "bbox": [ + 390, + 433, + 413, + 443 + ], + "score": 0.65, + "content": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 432, + 470, + 445 + ], + "score": 1.0, + "content": "is CIFAR-10.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "table_body", + "bbox": [ + 106, + 452, + 504, + 637 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 452, + 504, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 504, + 637 + ], + "score": 0.982, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 / 89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 / 50%0.39 / 57%0.51 / 52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.52 / 64%0.52 /64%0.22 /60%0.45 / 62%0.23 /46%0.48 / 55%
TI0.24 /42%0.52 / 44%0.13 /35%0.21 /36%0.12 /38%0.25 /40%
UNIVERSAL DEEPSLOTH
C100.89 /81%0.89 /81%0.75 /87%0.88 / 76%0.81/84%0.92 /76%
TI0.63 / 54%0.82 /48%0.38 /56%0.51 / 52%0.41 / 56%0.58 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88 /73%0.88 / 73%0.71 /81%0.86 / 70%0.76 /76%0.89 /66%
TI0.64 / 53%0.83 /49%0.38 /57%0.50 / 57%0.41 /58%0.58 / 53%
", + "type": "table", + "image_path": "e75aec7697dc259a45f3e17c19b3d6b5b0ffa1634f7b75a2b00810815ceb6fae.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 106, + 452, + 504, + 513.6666666666666 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 513.6666666666666, + 504, + 575.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 575.3333333333333, + 504, + 636.9999999999999 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 210, + 678 + ], + "score": 1.0, + "content": "Our results show that the", + "type": "text" + }, + { + "bbox": [ + 211, + 666, + 221, + 677 + ], + "score": 0.84, + "content": "\\ell _ { 1 } \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 665, + 243, + 678 + ], + "score": 1.0, + "content": "- and", + "type": "text" + }, + { + "bbox": [ + 243, + 666, + 253, + 677 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 665, + 417, + 678 + ], + "score": 1.0, + "content": "-based attacks are less effective than the", + "type": "text" + }, + { + "bbox": [ + 417, + 666, + 431, + 677 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "-based attacks. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 169, + 689 + ], + "score": 1.0, + "content": "contrast to the", + "type": "text" + }, + { + "bbox": [ + 169, + 677, + 183, + 688 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 677, + 491, + 689 + ], + "score": 1.0, + "content": "-based attacks that eliminate the efficacy of victim multi-exit models, the", + "type": "text" + }, + { + "bbox": [ + 492, + 677, + 502, + 688 + ], + "score": 0.8, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 685, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 123, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 688, + 133, + 699 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 685, + 366, + 702 + ], + "score": 1.0, + "content": "-based attacks reduce the efficacy of the same models by", + "type": "text" + }, + { + "bbox": [ + 366, + 688, + 410, + 698 + ], + "score": 0.88, + "content": "0 . 2 4 { \\sim } 0 . 6 5", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 685, + 506, + 702 + ], + "score": 1.0, + "content": ". Besides, the accuracy", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 174, + 712 + ], + "score": 1.0, + "content": "drops caused by", + "type": "text" + }, + { + "bbox": [ + 174, + 699, + 185, + 710 + ], + "score": 0.84, + "content": "\\ell _ { 1 } \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 699, + 206, + 712 + ], + "score": 1.0, + "content": "- and", + "type": "text" + }, + { + "bbox": [ + 206, + 699, + 216, + 710 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 699, + 300, + 712 + ], + "score": 1.0, + "content": "-based attacks are in", + "type": "text" + }, + { + "bbox": [ + 300, + 699, + 333, + 709 + ], + "score": 0.89, + "content": "6 \\sim 2 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 699, + 417, + 712 + ], + "score": 1.0, + "content": ", smaller than that of", + "type": "text" + }, + { + "bbox": [ + 418, + 699, + 431, + 710 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "-based DeepSloth", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 148, + 721 + ], + "score": 0.86, + "content": "( 7 5 \\sim 9 9 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 709, + 337, + 722 + ], + "score": 1.0, + "content": ". Moreover, we see that the universal variants of", + "type": "text" + }, + { + "bbox": [ + 337, + 710, + 347, + 721 + ], + "score": 0.81, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 709, + 367, + 722 + ], + "score": 1.0, + "content": "- and", + "type": "text" + }, + { + "bbox": [ + 368, + 710, + 378, + 721 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "-based attacks can barely reduce", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 720, + 500, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 475, + 733 + ], + "score": 1.0, + "content": "the efficacy of multi-exit models—they decrease the efficacy up to 0.08 and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 475, + 721, + 495, + 731 + ], + "score": 0.87, + "content": "12 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 720, + 500, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 389, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 391, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 274, + 96 + ], + "score": 1.0, + "content": "C EMPIRICAL EVALUATION OF", + "type": "text" + }, + { + "bbox": [ + 274, + 82, + 285, + 94 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 79, + 312, + 96 + ], + "score": 1.0, + "content": "AND", + "type": "text" + }, + { + "bbox": [ + 312, + 81, + 324, + 94 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 79, + 391, + 96 + ], + "score": 1.0, + "content": "DEEPSLOTH", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 505, + 123 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 290, + 125 + ], + "score": 1.0, + "content": "Table 4 and Table 5 shows the effectiveness of", + "type": "text" + }, + { + "bbox": [ + 290, + 112, + 299, + 123 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 109, + 343, + 125 + ], + "score": 1.0, + "content": "-based and", + "type": "text" + }, + { + "bbox": [ + 343, + 112, + 353, + 123 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 109, + 506, + 125 + ], + "score": 1.0, + "content": "-based DeepSloth attacks, respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 109, + 506, + 125 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 185, + 504, + 370 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 142, + 505, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 223, + 155 + ], + "score": 1.0, + "content": "Table 4: The effectiveness of", + "type": "text" + }, + { + "bbox": [ + 224, + 143, + 233, + 154 + ], + "score": 0.86, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 142, + 287, + 155 + ], + "score": 1.0, + "content": "DeepSloth. ‘", + "type": "text" + }, + { + "bbox": [ + 288, + 143, + 342, + 153 + ], + "score": 0.77, + "content": "R \\mathrm { A D } { < } 5 , 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "’ columns list the results in each early-exit", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 164, + 471, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 390, + 177 + ], + "score": 1.0, + "content": "attack’s results are an average of 10 classes. ‘TI’ is Tiny Imagenet and", + "type": "text" + }, + { + "bbox": [ + 390, + 165, + 413, + 175 + ], + "score": 0.7, + "content": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 164, + 471, + 177 + ], + "score": 1.0, + "content": "is CIFAR-10.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 106, + 185, + 504, + 370 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 185, + 504, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 504, + 370 + ], + "score": 0.983, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 /89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 / 57%0.51 /52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.36 / 51%0.35 / 51%0.12 /36%0.34 /45%0.18 / 41%0.49 / 53%
TI0.23 /37%0.51 / 40%0.08 /22%0.15 / 25%0.08 /33%0.19 /35%
UNIVERSAL DEEPSLOTH
C100.89 /83%0.89 /83%0.75 /85%0.88 / 75%0.82 /85%0.92 /77%
TI0.64 / 55%0.83/50%0.38 /57%0.51 / 52%0.41 / 57%0.59 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88/ 73%0.88 / 73%0.69 / 78%0.86 /67%0.76 /74%0.89 / 65%
TI0.64 /54%0.83 /49%0.39 / 59%0.50 / 58%0.41 / 60%0.58 / 53%
", + "type": "table", + "image_path": "6b9f5d0671919db5ee08c6a0cfa8dee4ae1967265bbd425b6363d71e75af115c.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 106, + 185, + 504, + 246.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 246.66666666666666, + 504, + 308.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 308.3333333333333, + 504, + 370.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 4.5 + }, + { + "type": "table", + "bbox": [ + 106, + 452, + 504, + 637 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 410, + 505, + 444 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 223, + 422 + ], + "score": 1.0, + "content": "Table 5: The effectiveness of", + "type": "text" + }, + { + "bbox": [ + 224, + 410, + 234, + 421 + ], + "score": 0.85, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 410, + 286, + 422 + ], + "score": 1.0, + "content": "DeepSloth.", + "type": "text" + }, + { + "bbox": [ + 286, + 410, + 342, + 421 + ], + "score": 0.83, + "content": "\\cdot \\mathrm { R A D } { < } 5 , 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 410, + 505, + 422 + ], + "score": 1.0, + "content": "’ columns list the results in each early-exit", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "setting. Each entry includes the model’s efficacy score (left) and accuracy (right). The class-universal", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 432, + 470, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 389, + 445 + ], + "score": 1.0, + "content": "attack’s results are an average of 10 classes. ‘TI’ is Tiny Imagenet and", + "type": "text" + }, + { + "bbox": [ + 390, + 433, + 413, + 443 + ], + "score": 0.65, + "content": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 432, + 470, + 445 + ], + "score": 1.0, + "content": "is CIFAR-10.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "table_body", + "bbox": [ + 106, + 452, + 504, + 637 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 452, + 504, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 504, + 637 + ], + "score": 0.982, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 / 89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 / 50%0.39 / 57%0.51 / 52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.52 / 64%0.52 /64%0.22 /60%0.45 / 62%0.23 /46%0.48 / 55%
TI0.24 /42%0.52 / 44%0.13 /35%0.21 /36%0.12 /38%0.25 /40%
UNIVERSAL DEEPSLOTH
C100.89 /81%0.89 /81%0.75 /87%0.88 / 76%0.81/84%0.92 /76%
TI0.63 / 54%0.82 /48%0.38 /56%0.51 / 52%0.41 / 56%0.58 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88 /73%0.88 / 73%0.71 /81%0.86 / 70%0.76 /76%0.89 /66%
TI0.64 / 53%0.83 /49%0.38 /57%0.50 / 57%0.41 /58%0.58 / 53%
", + "type": "table", + "image_path": "e75aec7697dc259a45f3e17c19b3d6b5b0ffa1634f7b75a2b00810815ceb6fae.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 106, + 452, + 504, + 513.6666666666666 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 513.6666666666666, + 504, + 575.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 575.3333333333333, + 504, + 636.9999999999999 + ], + "spans": [], + "index": 13 + } + ] + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 210, + 678 + ], + "score": 1.0, + "content": "Our results show that the", + "type": "text" + }, + { + "bbox": [ + 211, + 666, + 221, + 677 + ], + "score": 0.84, + "content": "\\ell _ { 1 } \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 665, + 243, + 678 + ], + "score": 1.0, + "content": "- and", + "type": "text" + }, + { + "bbox": [ + 243, + 666, + 253, + 677 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 665, + 417, + 678 + ], + "score": 1.0, + "content": "-based attacks are less effective than the", + "type": "text" + }, + { + "bbox": [ + 417, + 666, + 431, + 677 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "-based attacks. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 169, + 689 + ], + "score": 1.0, + "content": "contrast to the", + "type": "text" + }, + { + "bbox": [ + 169, + 677, + 183, + 688 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 677, + 491, + 689 + ], + "score": 1.0, + "content": "-based attacks that eliminate the efficacy of victim multi-exit models, the", + "type": "text" + }, + { + "bbox": [ + 492, + 677, + 502, + 688 + ], + "score": 0.8, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 685, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 123, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 688, + 133, + 699 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 685, + 366, + 702 + ], + "score": 1.0, + "content": "-based attacks reduce the efficacy of the same models by", + "type": "text" + }, + { + "bbox": [ + 366, + 688, + 410, + 698 + ], + "score": 0.88, + "content": "0 . 2 4 { \\sim } 0 . 6 5", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 685, + 506, + 702 + ], + "score": 1.0, + "content": ". Besides, the accuracy", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 174, + 712 + ], + "score": 1.0, + "content": "drops caused by", + "type": "text" + }, + { + "bbox": [ + 174, + 699, + 185, + 710 + ], + "score": 0.84, + "content": "\\ell _ { 1 } \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 699, + 206, + 712 + ], + "score": 1.0, + "content": "- and", + "type": "text" + }, + { + "bbox": [ + 206, + 699, + 216, + 710 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 699, + 300, + 712 + ], + "score": 1.0, + "content": "-based attacks are in", + "type": "text" + }, + { + "bbox": [ + 300, + 699, + 333, + 709 + ], + "score": 0.89, + "content": "6 \\sim 2 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 699, + 417, + 712 + ], + "score": 1.0, + "content": ", smaller than that of", + "type": "text" + }, + { + "bbox": [ + 418, + 699, + 431, + 710 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "-based DeepSloth", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 148, + 721 + ], + "score": 0.86, + "content": "( 7 5 \\sim 9 9 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 709, + 337, + 722 + ], + "score": 1.0, + "content": ". Moreover, we see that the universal variants of", + "type": "text" + }, + { + "bbox": [ + 337, + 710, + 347, + 721 + ], + "score": 0.81, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 709, + 367, + 722 + ], + "score": 1.0, + "content": "- and", + "type": "text" + }, + { + "bbox": [ + 368, + 710, + 378, + 721 + ], + "score": 0.88, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "-based attacks can barely reduce", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 720, + 500, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 475, + 733 + ], + "score": 1.0, + "content": "the efficacy of multi-exit models—they decrease the efficacy up to 0.08 and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 475, + 721, + 495, + 731 + ], + "score": 0.87, + "content": "12 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 720, + 500, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 416, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 417, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 417, + 95 + ], + "score": 1.0, + "content": "D EMPIRICAL EVALUATION OF DEEPSLOTH ON RESNET56", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 113, + 457, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 458, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 458, + 126 + ], + "score": 1.0, + "content": "Table 6 shows the the effectiveness of our DeepSloth attacks on ResNet56-base models.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 106, + 188, + 504, + 374 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 146, + 504, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 146, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 405, + 158 + ], + "score": 1.0, + "content": "Table 6: The effectiveness of DeepSloth on the ResNet-based models.", + "type": "text" + }, + { + "bbox": [ + 406, + 146, + 464, + 157 + ], + "score": 0.83, + "content": "\\cdot \\mathrm { R A D } { < } 5 , 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 146, + 505, + 158 + ], + "score": 1.0, + "content": "’ columns", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "list the results in each early-exit setting. Each entry includes the model’s efficacy score (left) and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 167, + 433, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 433, + 181 + ], + "score": 1.0, + "content": "accuracy (right). The class-universal attack’s results are an average of 10 classes.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 106, + 188, + 504, + 374 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 188, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 504, + 374 + ], + "score": 0.983, + "html": "
NETWORKRESNET(lo)RESNET (l1)RESNET(2)
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.52 /87%0.69 /80%0.52/87%0.69 / 80%0.51 /87%0.69 / 80%
TI0.25 / 51%0.39 /46%0.25 /51%0.39 /46%0.25 /51%0.39 / 46%
DEEPSLOTH
C100.00 /19%0.01 /19%0.05 /43%0.18 /47%0.06 /45%0.17 /48%
TI0.00 /7%0.01 / 7%0.04 /27%0.10 /28%0.05 /34%0.13 /35%
UNIVERSAL DEEPSLOTH
C100.35 / 63%0.59 /60%0.49 /84%0.68 /75%0.48 /85%0.67 /76%
TI0.25 / 25%0.34 / 37%0.25 / 51%0.39 / 46%0.25 / 51%0.38 /46%
CLASS-UNIVERSAL DEEPSLOTH
C100.23 / 33%0.48 / 29%0.39 / 70%0.60 / 61%0.39 / 71%0.60 / 61%
TI0.11 /21%0.23 /18%0.23 / 51%0.36 /46%0.23 / 50%0.36 /46%
", + "type": "table", + "image_path": "5db15f9dd474b600c6dace3769bf9c0f1f48b737c914f14f9a26cce1d52ad611.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 106, + 188, + 504, + 250.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 250.0, + 504, + 312.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 312.0, + 504, + 374.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 506, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 378, + 417 + ], + "score": 1.0, + "content": "Our results show that ResNet56-based models are vulnerable to all the", + "type": "text" + }, + { + "bbox": [ + 379, + 404, + 392, + 415 + ], + "score": 0.73, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 403, + 395, + 417 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 396, + 404, + 406, + 415 + ], + "score": 0.71, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 403, + 425, + 417 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 425, + 404, + 435, + 415 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "-based DeepSloth", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 183, + 427 + ], + "score": 1.0, + "content": "attacks. Using our", + "type": "text" + }, + { + "bbox": [ + 183, + 415, + 197, + 426 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "-based DeepSloth, the attacker can reduce the efficacy of the victim models", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 117, + 438 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 426, + 161, + 436 + ], + "score": 0.89, + "content": "0 . 0 0 { \\sim } 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 425, + 244, + 438 + ], + "score": 1.0, + "content": "and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 244, + 426, + 281, + 436 + ], + "score": 0.9, + "content": "3 9 \\sim 6 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 425, + 336, + 438 + ], + "score": 1.0, + "content": ". Besides, the", + "type": "text" + }, + { + "bbox": [ + 336, + 426, + 347, + 437 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 425, + 367, + 438 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 367, + 426, + 377, + 437 + ], + "score": 0.89, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "-based attacks also decrease the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 436, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 150, + 449 + ], + "score": 1.0, + "content": "efficacy to", + "type": "text" + }, + { + "bbox": [ + 150, + 437, + 194, + 447 + ], + "score": 0.87, + "content": "0 . 0 4 { \\sim } 0 . 1 8", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 436, + 277, + 449 + ], + "score": 1.0, + "content": "and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 278, + 437, + 315, + 447 + ], + "score": 0.89, + "content": "1 1 { \\sim } 4 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 436, + 506, + 449 + ], + "score": 1.0, + "content": ". Compared to the results on MSDNet, VGG16,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "score": 1.0, + "content": "and MobileNet in Table 4 and 5, the same attacks are more effective. The universal variants decrease", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 459, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 294, + 471 + ], + "score": 1.0, + "content": "the efficacy up to 0.21 and the accuracy up to", + "type": "text" + }, + { + "bbox": [ + 294, + 459, + 314, + 469 + ], + "score": 0.87, + "content": "24 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 459, + 388, + 471 + ], + "score": 1.0, + "content": ". In particular, the", + "type": "text" + }, + { + "bbox": [ + 389, + 459, + 399, + 469 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 459, + 420, + 471 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 420, + 459, + 430, + 469 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 459, + 506, + 471 + ], + "score": 1.0, + "content": "-based attacks (on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 470, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 506, + 481 + ], + "score": 1.0, + "content": "CIFAR-10 models) are effective than the same attacks on MSDNet, VGG16, and MobileNet models.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 508, + 347, + 521 + ], + "lines": [ + { + "bbox": [ + 104, + 506, + 348, + 524 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 348, + 524 + ], + "score": 1.0, + "content": "E COST OF CRAFTING DEEPSLOTH SAMPLES", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 540, + 324, + 654 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 324, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 324, + 552 + ], + "score": 1.0, + "content": "In Table 7, we compare the cost of DeepSloth with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 551, + 326, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 326, + 563 + ], + "score": 1.0, + "content": "other attack algorithms on a VGG16-based CIFAR-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 561, + 325, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 325, + 574 + ], + "score": 1.0, + "content": "10 model—executed on a single Nvidia Tesla-V100", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 573, + 325, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 325, + 586 + ], + "score": 1.0, + "content": "GPU. For the universal DeepSloth, we measure the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 583, + 325, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 325, + 597 + ], + "score": 1.0, + "content": "execution time for crafting a perturbation using one", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 595, + 325, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 325, + 607 + ], + "score": 1.0, + "content": "batch (250 samples) of the training set. For the other", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 606, + 325, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 325, + 617 + ], + "score": 1.0, + "content": "attacks, we measure the time for perturbing the whole", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 616, + 325, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 325, + 629 + ], + "score": 1.0, + "content": "test set of CIFAR-10. Our DeepSloth takes roughly", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 628, + 325, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 325, + 639 + ], + "score": 1.0, + "content": "the same time as the PGD and PGD-avg attacks and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 639, + 326, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 326, + 650 + ], + "score": 1.0, + "content": "significantly less time than the PGD-max attack. Our", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5 + }, + { + "type": "table", + "bbox": [ + 334, + 541, + 504, + 627 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 334, + 541, + 504, + 627 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 541, + 504, + 627 + ], + "spans": [ + { + "bbox": [ + 334, + 541, + 504, + 627 + ], + "score": 0.978, + "html": "
ATTACKSTIME (SEC.)
PGD-2038
PGD-20 (AVG.)48
PGD-20 (MAX.)475
DEEPSLOTH44
UNIVERSAL DEEPSLOTH2
", + "type": "table", + "image_path": "f4af52396b56906f7cf4187b9615e7a86dffd1aa63bb7f7499d79e474d5187d9.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 334, + 541, + 504, + 555.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 334, + 555.3333333333334, + 504, + 569.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 334, + 569.6666666666667, + 504, + 584.0000000000001 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 334, + 584.0000000000001, + 504, + 598.3333333333335 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 334, + 598.3333333333335, + 504, + 612.6666666666669 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 334, + 612.6666666666669, + 504, + 627.0000000000002 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 342, + 637, + 495, + 648 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 342, + 637, + 495, + 649 + ], + "spans": [ + { + "bbox": [ + 342, + 637, + 495, + 649 + ], + "score": 1.0, + "content": "Table 7: Time it takes to craft attacks.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 650, + 499, + 661 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 648, + 500, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 278, + 664 + ], + "score": 1.0, + "content": "universal DeepSloth takes only 2 seconds (", + "type": "text" + }, + { + "bbox": [ + 278, + 650, + 294, + 660 + ], + "score": 0.35, + "content": "1 0 \\mathrm { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 648, + 500, + 664 + ], + "score": 1.0, + "content": "faster than DeepSloth) as it only uses 250 samples.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 106, + 689, + 495, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 496, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 496, + 702 + ], + "score": 1.0, + "content": "F ADVERSARIAL EXAMPLES FROM STANDARD ATTACKS AND DEEPSLOTH", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 499, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 500, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 500, + 734 + ], + "score": 1.0, + "content": "In Figure 4, we visualize the adversarial examples from the PGD, UAP and our DeepSloth attacks.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 416, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 417, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 417, + 95 + ], + "score": 1.0, + "content": "D EMPIRICAL EVALUATION OF DEEPSLOTH ON RESNET56", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 113, + 457, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 458, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 458, + 126 + ], + "score": 1.0, + "content": "Table 6 shows the the effectiveness of our DeepSloth attacks on ResNet56-base models.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 112, + 458, + 126 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 188, + 504, + 374 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 146, + 504, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 146, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 405, + 158 + ], + "score": 1.0, + "content": "Table 6: The effectiveness of DeepSloth on the ResNet-based models.", + "type": "text" + }, + { + "bbox": [ + 406, + 146, + 464, + 157 + ], + "score": 0.83, + "content": "\\cdot \\mathrm { R A D } { < } 5 , 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 146, + 505, + 158 + ], + "score": 1.0, + "content": "’ columns", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "list the results in each early-exit setting. Each entry includes the model’s efficacy score (left) and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 167, + 433, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 433, + 181 + ], + "score": 1.0, + "content": "accuracy (right). The class-universal attack’s results are an average of 10 classes.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 106, + 188, + 504, + 374 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 188, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 504, + 374 + ], + "score": 0.983, + "html": "
NETWORKRESNET(lo)RESNET (l1)RESNET(2)
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.52 /87%0.69 /80%0.52/87%0.69 / 80%0.51 /87%0.69 / 80%
TI0.25 / 51%0.39 /46%0.25 /51%0.39 /46%0.25 /51%0.39 / 46%
DEEPSLOTH
C100.00 /19%0.01 /19%0.05 /43%0.18 /47%0.06 /45%0.17 /48%
TI0.00 /7%0.01 / 7%0.04 /27%0.10 /28%0.05 /34%0.13 /35%
UNIVERSAL DEEPSLOTH
C100.35 / 63%0.59 /60%0.49 /84%0.68 /75%0.48 /85%0.67 /76%
TI0.25 / 25%0.34 / 37%0.25 / 51%0.39 / 46%0.25 / 51%0.38 /46%
CLASS-UNIVERSAL DEEPSLOTH
C100.23 / 33%0.48 / 29%0.39 / 70%0.60 / 61%0.39 / 71%0.60 / 61%
TI0.11 /21%0.23 /18%0.23 / 51%0.36 /46%0.23 / 50%0.36 /46%
", + "type": "table", + "image_path": "5db15f9dd474b600c6dace3769bf9c0f1f48b737c914f14f9a26cce1d52ad611.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 106, + 188, + 504, + 250.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 250.0, + 504, + 312.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 312.0, + 504, + 374.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 506, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 378, + 417 + ], + "score": 1.0, + "content": "Our results show that ResNet56-based models are vulnerable to all the", + "type": "text" + }, + { + "bbox": [ + 379, + 404, + 392, + 415 + ], + "score": 0.73, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 403, + 395, + 417 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 396, + 404, + 406, + 415 + ], + "score": 0.71, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 403, + 425, + 417 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 425, + 404, + 435, + 415 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "-based DeepSloth", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 183, + 427 + ], + "score": 1.0, + "content": "attacks. Using our", + "type": "text" + }, + { + "bbox": [ + 183, + 415, + 197, + 426 + ], + "score": 0.89, + "content": "\\ell _ { \\infty }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "-based DeepSloth, the attacker can reduce the efficacy of the victim models", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 117, + 438 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 426, + 161, + 436 + ], + "score": 0.89, + "content": "0 . 0 0 { \\sim } 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 425, + 244, + 438 + ], + "score": 1.0, + "content": "and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 244, + 426, + 281, + 436 + ], + "score": 0.9, + "content": "3 9 \\sim 6 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 425, + 336, + 438 + ], + "score": 1.0, + "content": ". Besides, the", + "type": "text" + }, + { + "bbox": [ + 336, + 426, + 347, + 437 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 425, + 367, + 438 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 367, + 426, + 377, + 437 + ], + "score": 0.89, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "-based attacks also decrease the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 436, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 150, + 449 + ], + "score": 1.0, + "content": "efficacy to", + "type": "text" + }, + { + "bbox": [ + 150, + 437, + 194, + 447 + ], + "score": 0.87, + "content": "0 . 0 4 { \\sim } 0 . 1 8", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 436, + 277, + 449 + ], + "score": 1.0, + "content": "and the accuracy by", + "type": "text" + }, + { + "bbox": [ + 278, + 437, + 315, + 447 + ], + "score": 0.89, + "content": "1 1 { \\sim } 4 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 436, + 506, + 449 + ], + "score": 1.0, + "content": ". Compared to the results on MSDNet, VGG16,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "score": 1.0, + "content": "and MobileNet in Table 4 and 5, the same attacks are more effective. The universal variants decrease", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 459, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 294, + 471 + ], + "score": 1.0, + "content": "the efficacy up to 0.21 and the accuracy up to", + "type": "text" + }, + { + "bbox": [ + 294, + 459, + 314, + 469 + ], + "score": 0.87, + "content": "24 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 459, + 388, + 471 + ], + "score": 1.0, + "content": ". In particular, the", + "type": "text" + }, + { + "bbox": [ + 389, + 459, + 399, + 469 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 459, + 420, + 471 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 420, + 459, + 430, + 469 + ], + "score": 0.88, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 459, + 506, + 471 + ], + "score": 1.0, + "content": "-based attacks (on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 470, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 506, + 481 + ], + "score": 1.0, + "content": "CIFAR-10 models) are effective than the same attacks on MSDNet, VGG16, and MobileNet models.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 403, + 506, + 481 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 508, + 347, + 521 + ], + "lines": [ + { + "bbox": [ + 104, + 506, + 348, + 524 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 348, + 524 + ], + "score": 1.0, + "content": "E COST OF CRAFTING DEEPSLOTH SAMPLES", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 540, + 324, + 654 + ], + "lines": [ + { + "bbox": [ + 106, + 540, + 324, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 324, + 552 + ], + "score": 1.0, + "content": "In Table 7, we compare the cost of DeepSloth with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 551, + 326, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 326, + 563 + ], + "score": 1.0, + "content": "other attack algorithms on a VGG16-based CIFAR-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 561, + 325, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 325, + 574 + ], + "score": 1.0, + "content": "10 model—executed on a single Nvidia Tesla-V100", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 573, + 325, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 325, + 586 + ], + "score": 1.0, + "content": "GPU. For the universal DeepSloth, we measure the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 583, + 325, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 325, + 597 + ], + "score": 1.0, + "content": "execution time for crafting a perturbation using one", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 595, + 325, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 325, + 607 + ], + "score": 1.0, + "content": "batch (250 samples) of the training set. For the other", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 606, + 325, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 325, + 617 + ], + "score": 1.0, + "content": "attacks, we measure the time for perturbing the whole", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 616, + 325, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 325, + 629 + ], + "score": 1.0, + "content": "test set of CIFAR-10. Our DeepSloth takes roughly", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 628, + 325, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 325, + 639 + ], + "score": 1.0, + "content": "the same time as the PGD and PGD-avg attacks and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 639, + 326, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 326, + 650 + ], + "score": 1.0, + "content": "significantly less time than the PGD-max attack. Our", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 540, + 326, + 650 + ] + }, + { + "type": "table", + "bbox": [ + 334, + 541, + 504, + 627 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 334, + 541, + 504, + 627 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 541, + 504, + 627 + ], + "spans": [ + { + "bbox": [ + 334, + 541, + 504, + 627 + ], + "score": 0.978, + "html": "
ATTACKSTIME (SEC.)
PGD-2038
PGD-20 (AVG.)48
PGD-20 (MAX.)475
DEEPSLOTH44
UNIVERSAL DEEPSLOTH2
", + "type": "table", + "image_path": "f4af52396b56906f7cf4187b9615e7a86dffd1aa63bb7f7499d79e474d5187d9.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 334, + 541, + 504, + 555.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 334, + 555.3333333333334, + 504, + 569.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 334, + 569.6666666666667, + 504, + 584.0000000000001 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 334, + 584.0000000000001, + 504, + 598.3333333333335 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 334, + 598.3333333333335, + 504, + 612.6666666666669 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 334, + 612.6666666666669, + 504, + 627.0000000000002 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 342, + 637, + 495, + 648 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 342, + 637, + 495, + 649 + ], + "spans": [ + { + "bbox": [ + 342, + 637, + 495, + 649 + ], + "score": 1.0, + "content": "Table 7: Time it takes to craft attacks.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 650, + 499, + 661 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 648, + 500, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 278, + 664 + ], + "score": 1.0, + "content": "universal DeepSloth takes only 2 seconds (", + "type": "text" + }, + { + "bbox": [ + 278, + 650, + 294, + 660 + ], + "score": 0.35, + "content": "1 0 \\mathrm { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 648, + 500, + 664 + ], + "score": 1.0, + "content": "faster than DeepSloth) as it only uses 250 samples.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 106, + 689, + 495, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 496, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 496, + 702 + ], + "score": 1.0, + "content": "F ADVERSARIAL EXAMPLES FROM STANDARD ATTACKS AND DEEPSLOTH", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 499, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 500, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 500, + 734 + ], + "score": 1.0, + "content": "In Figure 4, we visualize the adversarial examples from the PGD, UAP and our DeepSloth attacks.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 720, + 500, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 78, + 497, + 462 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 78, + 497, + 462 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 78, + 497, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 497, + 462 + ], + "score": 0.926, + "type": "image", + "image_path": "0bdcaea5dbd66848101c989036785e7979050485b14922c9d380c77e29ee961e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 78, + 497, + 206.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 206.0, + 497, + 334.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 334.0, + 497, + 462.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 466, + 505, + 522 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 466, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 506, + 478 + ], + "score": 1.0, + "content": "Figure 4: Adversarial examples from the standard and our DeepSloth attacks. The leftmost", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "column shows the clean images. In the next four columns, we show adversarial examples from PGD,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "PGD (avg.), PGD (max.), and UAP attacks, respectively. The last four columns include adversarial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 499, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 511 + ], + "score": 1.0, + "content": "examples from the three variants of DeepSloth. Each row corresponds to each sample, and the last", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 510, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 207, + 523 + ], + "score": 1.0, + "content": "row contains the average", + "type": "text" + }, + { + "bbox": [ + 208, + 510, + 223, + 522 + ], + "score": 0.89, + "content": "\\ell _ { \\mathrm { i n f } }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 510, + 481, + 523 + ], + "score": 1.0, + "content": "-norm of the perturbations over the eight samples in each attack.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 78, + 497, + 462 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 78, + 497, + 462 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 78, + 497, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 497, + 462 + ], + "score": 0.926, + "type": "image", + "image_path": "0bdcaea5dbd66848101c989036785e7979050485b14922c9d380c77e29ee961e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 78, + 497, + 206.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 206.0, + 497, + 334.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 334.0, + 497, + 462.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 466, + 505, + 522 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 466, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 506, + 478 + ], + "score": 1.0, + "content": "Figure 4: Adversarial examples from the standard and our DeepSloth attacks. The leftmost", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "column shows the clean images. In the next four columns, we show adversarial examples from PGD,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "PGD (avg.), PGD (max.), and UAP attacks, respectively. The last four columns include adversarial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 499, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 511 + ], + "score": 1.0, + "content": "examples from the three variants of DeepSloth. Each row corresponds to each sample, and the last", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 510, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 207, + 523 + ], + "score": 1.0, + "content": "row contains the average", + "type": "text" + }, + { + "bbox": [ + 208, + 510, + 223, + 522 + ], + "score": 0.89, + "content": "\\ell _ { \\mathrm { i n f } }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 510, + 481, + 523 + ], + "score": 1.0, + "content": "-norm of the perturbations over the eight samples in each attack.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/9xC2tWEwBD/9xC2tWEwBD_model.json b/parse/train/9xC2tWEwBD/9xC2tWEwBD_model.json new file mode 100644 index 0000000000000000000000000000000000000000..614e085ed2ad73f366e53af27030934cea060da6 --- /dev/null +++ b/parse/train/9xC2tWEwBD/9xC2tWEwBD_model.json @@ -0,0 +1,22567 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 397, + 686, + 1303, + 686, + 1303, + 1387, + 397, + 1387 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1512, + 1404, + 1512, + 1404, + 1757, + 298, + 1757 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1772, + 1404, + 1772, + 1404, + 1987, + 298, + 1987 + ], + "score": 0.977 + }, + { + "category_id": 0, + "poly": [ + 300, + 218, + 1358, + 218, + 1358, + 320, + 300, + 320 + ], + "score": 0.957 + }, + { + "category_id": 2, + "poly": [ + 333, + 2006, + 628, + 2006, + 628, + 2033, + 333, + 2033 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 302, + 1443, + 573, + 1443, + 573, + 1478, + 302, + 1478 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 773, + 619, + 926, + 619, + 926, + 652, + 773, + 652 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 104, + 300, + 104 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 318, + 370, + 1199, + 370, + 1199, + 405, + 318, + 405 + ], + "score": 0.82 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 856, + 2088, + 856, + 2112, + 840, + 2112 + ], + "score": 0.737 + }, + { + "category_id": 1, + "poly": [ + 317, + 410, + 854, + 410, + 854, + 472, + 317, + 472 + ], + "score": 0.642 + }, + { + "category_id": 1, + "poly": [ + 316, + 480, + 1004, + 480, + 1004, + 537, + 316, + 537 + ], + "score": 0.606 + }, + { + "category_id": 13, + "poly": [ + 816, + 1144, + 927, + 1144, + 927, + 1173, + 816, + 1173 + ], + "score": 0.88, + "latex": "9 0 { - } 1 0 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 395, + 1174, + 484, + 1174, + 484, + 1204, + 395, + 1204 + ], + "score": 0.88, + "latex": "1 . 5 – 5 \\times" + }, + { + "category_id": 13, + "poly": [ + 1293, + 1635, + 1334, + 1635, + 1334, + 1664, + 1293, + 1664 + ], + "score": 0.87, + "latex": "4 \\times" + }, + { + "category_id": 13, + "poly": [ + 736, + 1726, + 790, + 1726, + 790, + 1755, + 736, + 1755 + ], + "score": 0.86, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 999, + 1727, + 1040, + 1727, + 1040, + 1755, + 999, + 1755 + ], + "score": 0.86, + "latex": "9 \\times" + }, + { + "category_id": 13, + "poly": [ + 1351, + 1665, + 1403, + 1665, + 1403, + 1694, + 1351, + 1694 + ], + "score": 0.85, + "latex": "19 \\%" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 218.0, + 1244.0, + 218.0, + 1244.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 275.0, + 1363.0, + 275.0, + 1363.0, + 323.0, + 296.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1999.0, + 633.0, + 1999.0, + 633.0, + 2039.0, + 331.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1440.0, + 579.0, + 1440.0, + 579.0, + 1487.0, + 294.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 618.0, + 932.0, + 618.0, + 932.0, + 655.0, + 769.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 859.0, + 2087.0, + 859.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 684.0, + 1308.0, + 684.0, + 1308.0, + 722.0, + 393.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 717.0, + 1308.0, + 717.0, + 1308.0, + 752.0, + 394.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 748.0, + 1306.0, + 748.0, + 1306.0, + 780.0, + 393.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 779.0, + 1306.0, + 779.0, + 1306.0, + 811.0, + 394.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 807.0, + 1308.0, + 807.0, + 1308.0, + 845.0, + 392.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 838.0, + 1308.0, + 838.0, + 1308.0, + 877.0, + 392.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 872.0, + 1304.0, + 872.0, + 1304.0, + 904.0, + 394.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 901.0, + 1305.0, + 901.0, + 1305.0, + 933.0, + 395.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 930.0, + 1307.0, + 930.0, + 1307.0, + 965.0, + 394.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 960.0, + 1307.0, + 960.0, + 1307.0, + 995.0, + 392.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 992.0, + 1306.0, + 992.0, + 1306.0, + 1024.0, + 394.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1019.0, + 1305.0, + 1019.0, + 1305.0, + 1057.0, + 393.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1051.0, + 1305.0, + 1051.0, + 1305.0, + 1089.0, + 392.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1080.0, + 1305.0, + 1080.0, + 1305.0, + 1120.0, + 392.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1114.0, + 1306.0, + 1114.0, + 1306.0, + 1146.0, + 394.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1142.0, + 815.0, + 1142.0, + 815.0, + 1179.0, + 392.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1142.0, + 1305.0, + 1142.0, + 1305.0, + 1179.0, + 928.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1174.0, + 1306.0, + 1174.0, + 1306.0, + 1207.0, + 485.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1207.0, + 1305.0, + 1207.0, + 1305.0, + 1236.0, + 394.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1234.0, + 1307.0, + 1234.0, + 1307.0, + 1268.0, + 393.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1265.0, + 1306.0, + 1265.0, + 1306.0, + 1299.0, + 393.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1295.0, + 1304.0, + 1295.0, + 1304.0, + 1330.0, + 394.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1326.0, + 1307.0, + 1326.0, + 1307.0, + 1362.0, + 394.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1359.0, + 1040.0, + 1359.0, + 1040.0, + 1391.0, + 394.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1510.0, + 1406.0, + 1510.0, + 1406.0, + 1549.0, + 292.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1541.0, + 1409.0, + 1541.0, + 1409.0, + 1581.0, + 292.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1608.0, + 294.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1404.0, + 1605.0, + 1404.0, + 1638.0, + 294.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1636.0, + 1292.0, + 1636.0, + 1292.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1636.0, + 1405.0, + 1636.0, + 1405.0, + 1669.0, + 1335.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1663.0, + 1350.0, + 1663.0, + 1350.0, + 1700.0, + 293.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1663.0, + 1409.0, + 1663.0, + 1409.0, + 1700.0, + 1404.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1696.0, + 1408.0, + 1696.0, + 1408.0, + 1729.0, + 293.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1724.0, + 735.0, + 1724.0, + 735.0, + 1761.0, + 293.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1724.0, + 998.0, + 1724.0, + 998.0, + 1761.0, + 791.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1724.0, + 1258.0, + 1724.0, + 1258.0, + 1761.0, + 1041.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1774.0, + 1404.0, + 1774.0, + 1404.0, + 1804.0, + 296.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1802.0, + 1404.0, + 1802.0, + 1404.0, + 1839.0, + 293.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1834.0, + 1404.0, + 1834.0, + 1404.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1867.0, + 1406.0, + 1867.0, + 1406.0, + 1898.0, + 296.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 1404.0, + 1893.0, + 1404.0, + 1927.0, + 294.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1924.0, + 1405.0, + 1924.0, + 1405.0, + 1958.0, + 294.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1957.0, + 1404.0, + 1957.0, + 1404.0, + 1991.0, + 294.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 366.0, + 1205.0, + 366.0, + 1205.0, + 416.0, + 309.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 408.0, + 806.0, + 408.0, + 806.0, + 442.0, + 314.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 437.0, + 857.0, + 437.0, + 857.0, + 475.0, + 313.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 479.0, + 898.0, + 479.0, + 898.0, + 509.0, + 314.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 509.0, + 1008.0, + 509.0, + 1008.0, + 540.0, + 311.0, + 540.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1320, + 1406, + 1320, + 1406, + 1869, + 297, + 1869 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 905, + 1406, + 905, + 1406, + 1301, + 297, + 1301 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 400, + 957, + 400, + 957, + 824, + 299, + 824 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1403, + 229, + 1403, + 383, + 297, + 383 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 1000, + 401, + 1379, + 401, + 1379, + 660, + 1000, + 660 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 979, + 702, + 1406, + 702, + 1406, + 816, + 979, + 816 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 299, + 826, + 1402, + 826, + 1402, + 888, + 299, + 888 + ], + "score": 0.944 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.711 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.14 + }, + { + "category_id": 13, + "poly": [ + 648, + 1655, + 735, + 1655, + 735, + 1685, + 648, + 1685 + ], + "score": 0.88, + "latex": "1 . 5 – 5 \\times" + }, + { + "category_id": 13, + "poly": [ + 1276, + 784, + 1325, + 784, + 1325, + 812, + 1276, + 812 + ], + "score": 0.88, + "latex": "1 4 ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 1317, + 1716, + 1398, + 1716, + 1398, + 1745, + 1317, + 1745 + ], + "score": 0.87, + "latex": "5- 4 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 457, + 1594, + 565, + 1594, + 565, + 1624, + 457, + 1624 + ], + "score": 0.86, + "latex": "9 0 { - } 1 0 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1143, + 785, + 1176, + 785, + 1176, + 812, + 1143, + 812 + ], + "score": 0.74, + "latex": "1 ^ { s t }" + }, + { + "category_id": 13, + "poly": [ + 1186, + 784, + 1223, + 784, + 1223, + 813, + 1186, + 813 + ], + "score": 0.74, + "latex": "5 ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 1333, + 907, + 1402, + 907, + 1402, + 937, + 1333, + 937 + ], + "score": 0.68, + "latex": "( D o S )" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 401.0, + 1120.0, + 401.0, + 1120.0, + 436.0, + 1063.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 394.0, + 1243.0, + 394.0, + 1243.0, + 441.0, + 1183.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 397.0, + 1373.0, + 397.0, + 1373.0, + 440.0, + 1293.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 430.0, + 1050.0, + 430.0, + 1050.0, + 561.0, + 1007.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 559.0, + 1037.0, + 559.0, + 1037.0, + 635.0, + 999.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 700.0, + 1409.0, + 700.0, + 1409.0, + 735.0, + 976.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 730.0, + 1404.0, + 730.0, + 1404.0, + 759.0, + 977.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 758.0, + 1408.0, + 758.0, + 1408.0, + 790.0, + 976.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 778.0, + 1142.0, + 778.0, + 1142.0, + 820.0, + 973.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 778.0, + 1185.0, + 778.0, + 1185.0, + 820.0, + 1177.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 778.0, + 1275.0, + 778.0, + 1275.0, + 820.0, + 1224.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 778.0, + 1406.0, + 778.0, + 1406.0, + 820.0, + 1326.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1318.0, + 1403.0, + 1318.0, + 1403.0, + 1350.0, + 297.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1346.0, + 1406.0, + 1346.0, + 1406.0, + 1387.0, + 292.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1380.0, + 1406.0, + 1380.0, + 1406.0, + 1416.0, + 294.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1411.0, + 1406.0, + 1411.0, + 1406.0, + 1446.0, + 294.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1440.0, + 1407.0, + 1440.0, + 1407.0, + 1479.0, + 294.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1470.0, + 1407.0, + 1470.0, + 1407.0, + 1508.0, + 292.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1498.0, + 1408.0, + 1498.0, + 1408.0, + 1540.0, + 292.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1532.0, + 1406.0, + 1532.0, + 1406.0, + 1568.0, + 295.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1560.0, + 1408.0, + 1560.0, + 1408.0, + 1597.0, + 292.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1593.0, + 456.0, + 1593.0, + 456.0, + 1629.0, + 295.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 1593.0, + 1406.0, + 1593.0, + 1406.0, + 1629.0, + 566.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1622.0, + 1407.0, + 1622.0, + 1407.0, + 1660.0, + 291.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1654.0, + 647.0, + 1654.0, + 647.0, + 1690.0, + 294.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1654.0, + 1406.0, + 1654.0, + 1406.0, + 1690.0, + 736.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1683.0, + 1404.0, + 1683.0, + 1404.0, + 1719.0, + 294.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1716.0, + 1316.0, + 1716.0, + 1316.0, + 1748.0, + 295.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1716.0, + 1407.0, + 1716.0, + 1407.0, + 1748.0, + 1399.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1747.0, + 1407.0, + 1747.0, + 1407.0, + 1778.0, + 295.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1776.0, + 1406.0, + 1776.0, + 1406.0, + 1811.0, + 294.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1809.0, + 1407.0, + 1809.0, + 1407.0, + 1840.0, + 296.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1837.0, + 1409.0, + 1837.0, + 1409.0, + 1872.0, + 294.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 903.0, + 1332.0, + 903.0, + 1332.0, + 942.0, + 294.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 903.0, + 1407.0, + 903.0, + 1407.0, + 942.0, + 1403.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 932.0, + 1406.0, + 932.0, + 1406.0, + 973.0, + 292.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 961.0, + 1407.0, + 961.0, + 1407.0, + 1006.0, + 291.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 996.0, + 1406.0, + 996.0, + 1406.0, + 1032.0, + 294.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1027.0, + 1408.0, + 1027.0, + 1408.0, + 1063.0, + 294.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1055.0, + 1408.0, + 1055.0, + 1408.0, + 1094.0, + 294.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1087.0, + 1403.0, + 1087.0, + 1403.0, + 1120.0, + 295.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1117.0, + 1408.0, + 1117.0, + 1408.0, + 1153.0, + 294.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 1407.0, + 1148.0, + 1407.0, + 1184.0, + 295.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1174.0, + 1409.0, + 1174.0, + 1409.0, + 1219.0, + 291.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1210.0, + 1403.0, + 1210.0, + 1403.0, + 1243.0, + 296.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1239.0, + 1404.0, + 1239.0, + 1404.0, + 1275.0, + 295.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1271.0, + 1408.0, + 1271.0, + 1408.0, + 1306.0, + 294.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 396.0, + 958.0, + 396.0, + 958.0, + 436.0, + 295.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 429.0, + 961.0, + 429.0, + 961.0, + 464.0, + 295.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 461.0, + 960.0, + 461.0, + 960.0, + 492.0, + 295.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 491.0, + 960.0, + 491.0, + 960.0, + 523.0, + 294.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 521.0, + 962.0, + 521.0, + 962.0, + 556.0, + 293.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 552.0, + 962.0, + 552.0, + 962.0, + 586.0, + 295.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 583.0, + 960.0, + 583.0, + 960.0, + 615.0, + 296.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 614.0, + 959.0, + 614.0, + 959.0, + 646.0, + 294.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 643.0, + 963.0, + 643.0, + 963.0, + 678.0, + 294.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 672.0, + 961.0, + 672.0, + 961.0, + 707.0, + 295.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 703.0, + 959.0, + 703.0, + 959.0, + 737.0, + 295.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 736.0, + 959.0, + 736.0, + 959.0, + 768.0, + 296.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 764.0, + 962.0, + 764.0, + 962.0, + 799.0, + 295.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 796.0, + 961.0, + 796.0, + 961.0, + 828.0, + 296.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 289.0, + 1407.0, + 289.0, + 1407.0, + 327.0, + 292.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 322.0, + 1408.0, + 322.0, + 1408.0, + 359.0, + 295.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 352.0, + 611.0, + 352.0, + 611.0, + 388.0, + 295.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 822.0, + 1408.0, + 822.0, + 1408.0, + 862.0, + 293.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 854.0, + 1270.0, + 854.0, + 1270.0, + 890.0, + 295.0, + 890.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 298, + 1404, + 298, + 1404, + 665, + 298, + 665 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1740, + 1404, + 1740, + 1404, + 1955, + 299, + 1955 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1033, + 1404, + 1033, + 1404, + 1340, + 298, + 1340 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 683, + 1405, + 683, + 1405, + 1019, + 298, + 1019 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1357, + 1403, + 1357, + 1403, + 1541, + 298, + 1541 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 296, + 1660, + 1400, + 1660, + 1400, + 1722, + 296, + 1722 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1403, + 1971, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 301, + 225, + 586, + 225, + 586, + 261, + 301, + 261 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 301, + 1588, + 679, + 1588, + 679, + 1623, + 301, + 1623 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2112, + 842, + 2112 + ], + "score": 0.584 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.421 + }, + { + "category_id": 15, + "poly": [ + 291.0, + 222.0, + 593.0, + 222.0, + 593.0, + 269.0, + 291.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1587.0, + 681.0, + 1587.0, + 681.0, + 1627.0, + 293.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 299.0, + 1405.0, + 299.0, + 1405.0, + 333.0, + 294.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 328.0, + 1407.0, + 328.0, + 1407.0, + 365.0, + 292.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 361.0, + 1402.0, + 361.0, + 1402.0, + 395.0, + 294.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 390.0, + 1406.0, + 390.0, + 1406.0, + 426.0, + 292.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 422.0, + 1408.0, + 422.0, + 1408.0, + 456.0, + 294.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 452.0, + 1408.0, + 452.0, + 1408.0, + 486.0, + 294.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 483.0, + 1405.0, + 483.0, + 1405.0, + 517.0, + 294.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 512.0, + 1408.0, + 512.0, + 1408.0, + 550.0, + 293.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 542.0, + 1404.0, + 542.0, + 1404.0, + 578.0, + 292.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 574.0, + 1406.0, + 574.0, + 1406.0, + 608.0, + 294.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 603.0, + 1406.0, + 603.0, + 1406.0, + 636.0, + 293.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 637.0, + 1367.0, + 637.0, + 1367.0, + 668.0, + 296.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1740.0, + 1406.0, + 1740.0, + 1406.0, + 1775.0, + 294.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1772.0, + 1408.0, + 1772.0, + 1408.0, + 1807.0, + 294.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1802.0, + 1406.0, + 1802.0, + 1406.0, + 1836.0, + 295.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1832.0, + 1408.0, + 1832.0, + 1408.0, + 1870.0, + 293.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1862.0, + 1405.0, + 1862.0, + 1405.0, + 1897.0, + 294.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 1409.0, + 1893.0, + 1409.0, + 1927.0, + 294.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1923.0, + 1266.0, + 1923.0, + 1266.0, + 1958.0, + 293.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1034.0, + 1406.0, + 1034.0, + 1406.0, + 1069.0, + 294.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1066.0, + 1404.0, + 1066.0, + 1404.0, + 1097.0, + 296.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1095.0, + 1408.0, + 1095.0, + 1408.0, + 1132.0, + 293.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1128.0, + 1405.0, + 1128.0, + 1405.0, + 1160.0, + 294.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1159.0, + 1405.0, + 1159.0, + 1405.0, + 1190.0, + 296.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1189.0, + 1404.0, + 1189.0, + 1404.0, + 1221.0, + 294.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1217.0, + 1406.0, + 1217.0, + 1406.0, + 1253.0, + 294.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1246.0, + 1406.0, + 1246.0, + 1406.0, + 1283.0, + 293.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1277.0, + 1406.0, + 1277.0, + 1406.0, + 1314.0, + 293.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1307.0, + 1129.0, + 1307.0, + 1129.0, + 1344.0, + 293.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 681.0, + 1405.0, + 681.0, + 1405.0, + 718.0, + 293.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 711.0, + 1408.0, + 711.0, + 1408.0, + 752.0, + 292.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 744.0, + 1405.0, + 744.0, + 1405.0, + 779.0, + 295.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 774.0, + 1405.0, + 774.0, + 1405.0, + 809.0, + 293.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 804.0, + 1405.0, + 804.0, + 1405.0, + 839.0, + 295.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 836.0, + 1408.0, + 836.0, + 1408.0, + 871.0, + 295.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 867.0, + 1405.0, + 867.0, + 1405.0, + 902.0, + 295.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 897.0, + 1406.0, + 897.0, + 1406.0, + 930.0, + 293.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 927.0, + 1406.0, + 927.0, + 1406.0, + 961.0, + 295.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 958.0, + 1405.0, + 958.0, + 1405.0, + 992.0, + 295.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 987.0, + 1408.0, + 987.0, + 1408.0, + 1023.0, + 293.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1355.0, + 1405.0, + 1355.0, + 1405.0, + 1392.0, + 293.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1388.0, + 1403.0, + 1388.0, + 1403.0, + 1420.0, + 296.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1417.0, + 1408.0, + 1417.0, + 1408.0, + 1454.0, + 292.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1448.0, + 1405.0, + 1448.0, + 1405.0, + 1485.0, + 293.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1481.0, + 1407.0, + 1481.0, + 1407.0, + 1512.0, + 296.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 1194.0, + 1511.0, + 1194.0, + 1545.0, + 293.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1661.0, + 1405.0, + 1661.0, + 1405.0, + 1696.0, + 295.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1691.0, + 1253.0, + 1691.0, + 1253.0, + 1726.0, + 296.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1969.0, + 1404.0, + 1969.0, + 1404.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 1408.0, + 2001.0, + 1408.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1058, + 1406, + 1058, + 1406, + 1304, + 297, + 1304 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1320, + 1405, + 1320, + 1405, + 1566, + 297, + 1566 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 368, + 846, + 368, + 846, + 886, + 299, + 886 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1582, + 1404, + 1582, + 1404, + 1805, + 297, + 1805 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1405, + 1819, + 1405, + 2035, + 298, + 2035 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1403, + 229, + 1403, + 352, + 297, + 352 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 871, + 382, + 1373, + 382, + 1373, + 759, + 871, + 759 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 867, + 796, + 1405, + 796, + 1405, + 919, + 867, + 919 + ], + "score": 0.965 + }, + { + "category_id": 0, + "poly": [ + 298, + 966, + 1025, + 966, + 1025, + 1003, + 298, + 1003 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.778 + }, + { + "category_id": 4, + "poly": [ + 896, + 365, + 1399, + 365, + 1399, + 384, + 896, + 384 + ], + "score": 0.306 + }, + { + "category_id": 13, + "poly": [ + 861, + 1150, + 1092, + 1150, + 1092, + 1184, + 861, + 1184 + ], + "score": 0.94, + "latex": "y \\in M = \\{ 1 , . . . , m \\}" + }, + { + "category_id": 13, + "poly": [ + 928, + 1242, + 1045, + 1242, + 1045, + 1274, + 928, + 1274 + ], + "score": 0.92, + "latex": "\\theta _ { i } \\subset \\theta _ { i + 1 }" + }, + { + "category_id": 13, + "poly": [ + 1294, + 1383, + 1396, + 1383, + 1396, + 1416, + 1294, + 1416 + ], + "score": 0.92, + "latex": "F _ { i } ( x , \\theta _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 401, + 1212, + 484, + 1212, + 484, + 1245, + 401, + 1245 + ], + "score": 0.92, + "latex": "( x _ { i } , y _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1056, + 1119, + 1218, + 1119, + 1218, + 1152, + 1056, + 1152 + ], + "score": 0.92, + "latex": "F ( x , \\theta ) \\in \\mathbf { \\mathbb { R } } ^ { m }" + }, + { + "category_id": 13, + "poly": [ + 866, + 1475, + 960, + 1475, + 960, + 1507, + 866, + 1507 + ], + "score": 0.92, + "latex": "F _ { i + 1 } ( x )" + }, + { + "category_id": 13, + "poly": [ + 749, + 1882, + 829, + 1882, + 829, + 1913, + 749, + 1913 + ], + "score": 0.92, + "latex": "{ \\hat { y } } _ { i } = y" + }, + { + "category_id": 13, + "poly": [ + 1304, + 1088, + 1396, + 1088, + 1396, + 1119, + 1304, + 1119 + ], + "score": 0.91, + "latex": "x \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1242, + 1194, + 1242, + 1194, + 1272, + 1100, + 1272 + ], + "score": 0.91, + "latex": "\\theta _ { N } = \\theta" + }, + { + "category_id": 13, + "poly": [ + 611, + 1414, + 676, + 1414, + 676, + 1445, + 611, + 1445 + ], + "score": 0.91, + "latex": "F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1337, + 1713, + 1403, + 1713, + 1403, + 1746, + 1337, + 1746 + ], + "score": 0.91, + "latex": "F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 916, + 1443, + 982, + 1443, + 982, + 1475, + 916, + 1475 + ], + "score": 0.9, + "latex": "F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 997, + 1414, + 1089, + 1414, + 1089, + 1440, + 997, + 1440 + ], + "score": 0.9, + "latex": "K = N" + }, + { + "category_id": 13, + "poly": [ + 492, + 292, + 567, + 292, + 567, + 321, + 492, + 321 + ], + "score": 0.89, + "latex": "{ \\sim } 5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 598, + 1445, + 678, + 1445, + 678, + 1471, + 598, + 1471 + ], + "score": 0.89, + "latex": "K = 6" + }, + { + "category_id": 13, + "poly": [ + 802, + 292, + 878, + 292, + 878, + 321, + 802, + 321 + ], + "score": 0.89, + "latex": "{ \\sim } 5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 589, + 1682, + 656, + 1682, + 656, + 1714, + 589, + 1714 + ], + "score": 0.89, + "latex": "F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1293, + 1350, + 1331, + 1350, + 1331, + 1383, + 1293, + 1383 + ], + "score": 0.89, + "latex": "j ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 757, + 1240, + 792, + 1240, + 792, + 1269, + 757, + 1269 + ], + "score": 0.89, + "latex": "i ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 495, + 1683, + 576, + 1683, + 576, + 1713, + 495, + 1713 + ], + "score": 0.89, + "latex": "{ \\hat { y } } _ { i } = y" + }, + { + "category_id": 13, + "poly": [ + 872, + 1350, + 907, + 1350, + 907, + 1379, + 872, + 1379 + ], + "score": 0.88, + "latex": "i ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 388, + 1849, + 424, + 1849, + 424, + 1878, + 388, + 1878 + ], + "score": 0.88, + "latex": "\\overline { { i } } ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 681, + 1973, + 733, + 1973, + 733, + 2002, + 681, + 2002 + ], + "score": 0.88, + "latex": "15 \\%" + }, + { + "category_id": 13, + "poly": [ + 690, + 1352, + 720, + 1352, + 720, + 1382, + 690, + 1382 + ], + "score": 0.88, + "latex": "F _ { i }" + }, + { + "category_id": 13, + "poly": [ + 845, + 1852, + 873, + 1852, + 873, + 1881, + 845, + 1881 + ], + "score": 0.87, + "latex": "T _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1345, + 1212, + 1370, + 1212, + 1370, + 1242, + 1345, + 1242 + ], + "score": 0.87, + "latex": "\\theta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 481, + 1912, + 510, + 1912, + 510, + 1942, + 481, + 1942 + ], + "score": 0.86, + "latex": "T _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1210, + 1217, + 1238, + 1217, + 1238, + 1242, + 1210, + 1242 + ], + "score": 0.86, + "latex": "x _ { i }" + }, + { + "category_id": 13, + "poly": [ + 591, + 1973, + 632, + 1973, + 632, + 2002, + 591, + 2002 + ], + "score": 0.86, + "latex": "5 \\%" + }, + { + "category_id": 13, + "poly": [ + 896, + 1822, + 925, + 1822, + 925, + 1851, + 896, + 1851 + ], + "score": 0.86, + "latex": "T _ { i }" + }, + { + "category_id": 13, + "poly": [ + 637, + 1216, + 663, + 1216, + 663, + 1243, + 637, + 1243 + ], + "score": 0.84, + "latex": "y _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1231, + 1182, + 1257, + 1182, + 1257, + 1208, + 1231, + 1208 + ], + "score": 0.84, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 829, + 1384, + 859, + 1384, + 859, + 1413, + 829, + 1413 + ], + "score": 0.82, + "latex": "F _ { i }" + }, + { + "category_id": 13, + "poly": [ + 520, + 1613, + 888, + 1613, + 888, + 1656, + 520, + 1656 + ], + "score": 0.81, + "latex": "\\begin{array} { r } { \\operatorname * { a r g m a x } _ { j \\in M } F _ { i } ^ { ( \\bar { j } ) } ( x ) = \\hat { y } _ { i } = y ; y } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 914, + 1323, + 943, + 1323, + 943, + 1349, + 914, + 1349 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 646, + 1659, + 665, + 1659, + 665, + 1683, + 646, + 1683 + ], + "score": 0.81, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 555, + 1381, + 693, + 1381, + 693, + 1414, + 555, + 1414 + ], + "score": 0.81, + "latex": "j ^ { t h } \\left( j < N \\right)" + }, + { + "category_id": 13, + "poly": [ + 715, + 1091, + 743, + 1091, + 743, + 1117, + 715, + 1117 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1170, + 1153, + 1187, + 1153, + 1187, + 1178, + 1170, + 1178 + ], + "score": 0.8, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1275, + 1049, + 1275, + 1049, + 1300, + 1028, + 1300 + ], + "score": 0.8, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 582, + 1155, + 601, + 1155, + 601, + 1178, + 582, + 1178 + ], + "score": 0.78, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 713, + 1479, + 733, + 1479, + 733, + 1501, + 713, + 1501 + ], + "score": 0.75, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1326, + 1973, + 1398, + 1973, + 1398, + 2002, + 1326, + 2002 + ], + "score": 0.74, + "latex": "{ < } 1 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1158, + 1974, + 1218, + 1974, + 1218, + 2002, + 1158, + 2002 + ], + "score": 0.67, + "latex": "123 \\%" + }, + { + "category_id": 13, + "poly": [ + 799, + 1388, + 818, + 1388, + 818, + 1410, + 799, + 1410 + ], + "score": 0.66, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 555, + 1380, + 594, + 1380, + 594, + 1415, + 555, + 1415 + ], + "score": 0.47, + "latex": "j ^ { t h }" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 385.0, + 932.0, + 385.0, + 932.0, + 417.0, + 888.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 391.0, + 899.0, + 391.0, + 899.0, + 712.0, + 868.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 394.0, + 1239.0, + 394.0, + 1239.0, + 422.0, + 979.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 417.0, + 1239.0, + 417.0, + 1239.0, + 446.0, + 983.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 449.0, + 930.0, + 449.0, + 930.0, + 475.0, + 893.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 442.0, + 1246.0, + 442.0, + 1246.0, + 470.0, + 980.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 508.0, + 929.0, + 508.0, + 929.0, + 535.0, + 892.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 569.0, + 929.0, + 569.0, + 929.0, + 593.0, + 893.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 626.0, + 929.0, + 626.0, + 929.0, + 653.0, + 892.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 686.0, + 929.0, + 686.0, + 929.0, + 712.0, + 892.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 714.0, + 987.0, + 714.0, + 987.0, + 743.0, + 949.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 714.0, + 1061.0, + 714.0, + 1061.0, + 743.0, + 1020.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 714.0, + 1134.0, + 714.0, + 1134.0, + 742.0, + 1093.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 714.0, + 1206.0, + 714.0, + 1206.0, + 743.0, + 1166.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 714.0, + 1278.0, + 714.0, + 1278.0, + 743.0, + 1238.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 714.0, + 1352.0, + 714.0, + 1352.0, + 743.0, + 1310.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 734.0, + 1313.0, + 734.0, + 1313.0, + 762.0, + 986.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 796.0, + 1402.0, + 796.0, + 1402.0, + 827.0, + 869.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 825.0, + 1405.0, + 825.0, + 1405.0, + 859.0, + 867.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 857.0, + 1404.0, + 857.0, + 1404.0, + 886.0, + 867.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 883.0, + 1231.0, + 883.0, + 1231.0, + 923.0, + 865.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 965.0, + 1027.0, + 965.0, + 1027.0, + 1006.0, + 294.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 364.0, + 1400.0, + 364.0, + 1400.0, + 388.0, + 894.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1057.0, + 1409.0, + 1057.0, + 1409.0, + 1094.0, + 294.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1089.0, + 714.0, + 1089.0, + 714.0, + 1122.0, + 291.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1089.0, + 1303.0, + 1089.0, + 1303.0, + 1122.0, + 744.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1089.0, + 1406.0, + 1089.0, + 1406.0, + 1122.0, + 1397.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1119.0, + 1055.0, + 1119.0, + 1055.0, + 1152.0, + 294.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1119.0, + 1404.0, + 1119.0, + 1404.0, + 1152.0, + 1219.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1150.0, + 581.0, + 1150.0, + 581.0, + 1185.0, + 291.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1150.0, + 860.0, + 1150.0, + 860.0, + 1185.0, + 602.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1150.0, + 1169.0, + 1150.0, + 1169.0, + 1185.0, + 1093.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1150.0, + 1406.0, + 1150.0, + 1406.0, + 1185.0, + 1188.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1180.0, + 1230.0, + 1180.0, + 1230.0, + 1216.0, + 292.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1180.0, + 1406.0, + 1180.0, + 1406.0, + 1216.0, + 1258.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1210.0, + 400.0, + 1210.0, + 400.0, + 1246.0, + 292.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1210.0, + 636.0, + 1210.0, + 636.0, + 1246.0, + 485.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1210.0, + 1209.0, + 1210.0, + 1209.0, + 1246.0, + 664.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1210.0, + 1344.0, + 1210.0, + 1344.0, + 1246.0, + 1239.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1210.0, + 1406.0, + 1210.0, + 1406.0, + 1246.0, + 1371.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1239.0, + 756.0, + 1239.0, + 756.0, + 1276.0, + 292.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1239.0, + 927.0, + 1239.0, + 927.0, + 1276.0, + 793.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1239.0, + 1099.0, + 1239.0, + 1099.0, + 1276.0, + 1046.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 1239.0, + 1407.0, + 1239.0, + 1407.0, + 1276.0, + 1195.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1272.0, + 1027.0, + 1272.0, + 1027.0, + 1306.0, + 295.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1272.0, + 1061.0, + 1272.0, + 1061.0, + 1306.0, + 1050.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1320.0, + 913.0, + 1320.0, + 913.0, + 1354.0, + 295.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1320.0, + 1406.0, + 1320.0, + 1406.0, + 1354.0, + 944.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1348.0, + 689.0, + 1348.0, + 689.0, + 1385.0, + 292.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1348.0, + 871.0, + 1348.0, + 871.0, + 1385.0, + 721.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1348.0, + 1292.0, + 1348.0, + 1292.0, + 1385.0, + 908.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 1348.0, + 1410.0, + 1348.0, + 1410.0, + 1385.0, + 1332.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1377.0, + 554.0, + 1377.0, + 554.0, + 1420.0, + 292.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1377.0, + 798.0, + 1377.0, + 798.0, + 1420.0, + 694.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1377.0, + 828.0, + 1377.0, + 828.0, + 1420.0, + 819.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1377.0, + 1293.0, + 1377.0, + 1293.0, + 1420.0, + 860.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1377.0, + 1408.0, + 1377.0, + 1408.0, + 1420.0, + 1397.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1413.0, + 610.0, + 1413.0, + 610.0, + 1447.0, + 295.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1413.0, + 996.0, + 1413.0, + 996.0, + 1447.0, + 677.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1413.0, + 1406.0, + 1413.0, + 1406.0, + 1447.0, + 1090.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1438.0, + 597.0, + 1438.0, + 597.0, + 1478.0, + 292.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1438.0, + 915.0, + 1438.0, + 915.0, + 1478.0, + 679.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1438.0, + 1406.0, + 1438.0, + 1406.0, + 1478.0, + 983.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1473.0, + 712.0, + 1473.0, + 712.0, + 1507.0, + 295.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1473.0, + 865.0, + 1473.0, + 865.0, + 1507.0, + 734.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1473.0, + 1405.0, + 1473.0, + 1405.0, + 1507.0, + 961.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1504.0, + 1406.0, + 1504.0, + 1406.0, + 1538.0, + 295.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1533.0, + 1364.0, + 1533.0, + 1364.0, + 1568.0, + 294.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 366.0, + 849.0, + 366.0, + 849.0, + 404.0, + 296.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 398.0, + 847.0, + 398.0, + 847.0, + 433.0, + 296.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 430.0, + 849.0, + 430.0, + 849.0, + 462.0, + 294.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 461.0, + 847.0, + 461.0, + 847.0, + 493.0, + 296.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 492.0, + 847.0, + 492.0, + 847.0, + 522.0, + 294.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 522.0, + 847.0, + 522.0, + 847.0, + 554.0, + 297.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 554.0, + 848.0, + 554.0, + 848.0, + 584.0, + 294.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 583.0, + 848.0, + 583.0, + 848.0, + 614.0, + 294.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 613.0, + 851.0, + 613.0, + 851.0, + 646.0, + 294.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 644.0, + 849.0, + 644.0, + 849.0, + 676.0, + 294.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 676.0, + 850.0, + 676.0, + 850.0, + 704.0, + 294.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 704.0, + 848.0, + 704.0, + 848.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 734.0, + 850.0, + 734.0, + 850.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 766.0, + 851.0, + 766.0, + 851.0, + 798.0, + 296.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 797.0, + 847.0, + 797.0, + 847.0, + 828.0, + 295.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 826.0, + 848.0, + 826.0, + 848.0, + 861.0, + 293.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 855.0, + 403.0, + 855.0, + 403.0, + 888.0, + 295.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 1407.0, + 1581.0, + 1407.0, + 1620.0, + 294.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1604.0, + 519.0, + 1604.0, + 519.0, + 1667.0, + 293.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1604.0, + 1413.0, + 1604.0, + 1413.0, + 1667.0, + 889.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1652.0, + 645.0, + 1652.0, + 645.0, + 1687.0, + 295.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1652.0, + 1405.0, + 1652.0, + 1405.0, + 1687.0, + 666.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1681.0, + 494.0, + 1681.0, + 494.0, + 1716.0, + 292.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1681.0, + 588.0, + 1681.0, + 588.0, + 1716.0, + 577.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1681.0, + 1406.0, + 1681.0, + 1406.0, + 1716.0, + 657.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1709.0, + 1336.0, + 1709.0, + 1336.0, + 1749.0, + 291.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 1406.0, + 1743.0, + 1406.0, + 1778.0, + 294.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1770.0, + 1098.0, + 1770.0, + 1098.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1821.0, + 895.0, + 1821.0, + 895.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 926.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1847.0, + 387.0, + 1847.0, + 387.0, + 1888.0, + 292.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 1847.0, + 844.0, + 1847.0, + 844.0, + 1888.0, + 425.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1847.0, + 1405.0, + 1847.0, + 1405.0, + 1888.0, + 874.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 748.0, + 1880.0, + 748.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1916.0, + 830.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1913.0, + 480.0, + 1913.0, + 480.0, + 1944.0, + 295.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1913.0, + 1403.0, + 1913.0, + 1403.0, + 1944.0, + 511.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 590.0, + 1973.0, + 590.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1973.0, + 680.0, + 1973.0, + 680.0, + 2007.0, + 633.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1973.0, + 1157.0, + 1973.0, + 1157.0, + 2007.0, + 734.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1973.0, + 1325.0, + 1973.0, + 1325.0, + 2007.0, + 1219.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1973.0, + 1410.0, + 1973.0, + 1410.0, + 2007.0, + 1399.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 1186.0, + 2000.0, + 1186.0, + 2041.0, + 292.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 260.0, + 1406.0, + 260.0, + 1406.0, + 295.0, + 292.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 292.0, + 491.0, + 292.0, + 491.0, + 328.0, + 293.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 292.0, + 801.0, + 292.0, + 801.0, + 328.0, + 568.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 292.0, + 1405.0, + 292.0, + 1405.0, + 328.0, + 879.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 323.0, + 948.0, + 323.0, + 948.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1101, + 1406, + 1101, + 1406, + 1378, + 297, + 1378 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 401, + 1404, + 401, + 1404, + 650, + 298, + 650 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 804, + 1404, + 804, + 1404, + 989, + 297, + 989 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1405, + 1819, + 1405, + 2034, + 297, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 296, + 1531, + 1417, + 1531, + 1417, + 1696, + 296, + 1696 + ], + "score": 0.979, + "html": "
NETWORKNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)UAP
VGG-160.77 /89%0.79 / 29%0.85 /10%0.81/27%0.71/68%
REsNET-560.52 / 87%0.55 / 12%0.82/1%0.70/ 6%0.55 / 44%
MOBILENET0.83/87%0.85 /14%0.93/ 3%0.89 / 12%0.77 / 60%
" + }, + { + "category_id": 1, + "poly": [ + 297, + 664, + 1403, + 664, + 1403, + 788, + 297, + 788 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 292, + 1403, + 292, + 1403, + 385, + 298, + 385 + ], + "score": 0.973 + }, + { + "category_id": 6, + "poly": [ + 298, + 1420, + 1404, + 1420, + 1404, + 1508, + 298, + 1508 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 299, + 1755, + 674, + 1755, + 674, + 1787, + 299, + 1787 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 300, + 1039, + 1094, + 1039, + 1094, + 1072, + 300, + 1072 + ], + "score": 0.84 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 560, + 229, + 560, + 261, + 300, + 261 + ], + "score": 0.827 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.718 + }, + { + "category_id": 13, + "poly": [ + 1182, + 585, + 1328, + 585, + 1328, + 618, + 1182, + 618 + ], + "score": 0.93, + "latex": "| | v | | _ { 2 } \\leq 0 . 3 5" + }, + { + "category_id": 13, + "poly": [ + 1267, + 616, + 1398, + 616, + 1398, + 649, + 1267, + 649 + ], + "score": 0.92, + "latex": "| | v | | _ { 2 } \\leq 0 . 6" + }, + { + "category_id": 13, + "poly": [ + 855, + 322, + 946, + 322, + 946, + 352, + 855, + 352 + ], + "score": 0.92, + "latex": "v \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1090, + 616, + 1216, + 616, + 1216, + 649, + 1090, + 649 + ], + "score": 0.91, + "latex": "| | \\boldsymbol { v } | | _ { 1 } \\le \\mathrm { \\dot { 1 6 } }" + }, + { + "category_id": 13, + "poly": [ + 446, + 1943, + 556, + 1943, + 556, + 1975, + 446, + 1975 + ], + "score": 0.91, + "latex": "F _ { i } | i < K" + }, + { + "category_id": 13, + "poly": [ + 298, + 585, + 503, + 585, + 503, + 619, + 298, + 619 + ], + "score": 0.91, + "latex": "\\lvert | v \\rvert | _ { \\infty } \\leq \\epsilon = 0 . 0 3" + }, + { + "category_id": 13, + "poly": [ + 396, + 867, + 472, + 867, + 472, + 896, + 396, + 896 + ], + "score": 0.9, + "latex": "x \\in S" + }, + { + "category_id": 13, + "poly": [ + 663, + 617, + 868, + 617, + 868, + 649, + 663, + 649 + ], + "score": 0.9, + "latex": "| | v | | _ { \\infty } \\leq \\epsilon = 0 . 0 3" + }, + { + "category_id": 13, + "poly": [ + 740, + 585, + 853, + 585, + 853, + 618, + 740, + 618 + ], + "score": 0.9, + "latex": "| | v | | _ { 1 } \\leq 8" + }, + { + "category_id": 13, + "poly": [ + 1094, + 837, + 1166, + 837, + 1166, + 865, + 1094, + 865 + ], + "score": 0.9, + "latex": "x \\in S" + }, + { + "category_id": 13, + "poly": [ + 948, + 867, + 1028, + 867, + 1028, + 895, + 948, + 895 + ], + "score": 0.9, + "latex": "i \\in M" + }, + { + "category_id": 13, + "poly": [ + 617, + 465, + 656, + 465, + 656, + 494, + 617, + 494 + ], + "score": 0.89, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 298, + 354, + 369, + 354, + 369, + 382, + 298, + 382 + ], + "score": 0.87, + "latex": "x \\in S" + }, + { + "category_id": 13, + "poly": [ + 825, + 1943, + 852, + 1943, + 852, + 1973, + 825, + 1973 + ], + "score": 0.87, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1216, + 526, + 1244, + 526, + 1244, + 555, + 1216, + 555 + ], + "score": 0.87, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 701, + 1943, + 739, + 1943, + 739, + 1973, + 701, + 1973 + ], + "score": 0.87, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 1137, + 526, + 1164, + 526, + 1164, + 555, + 1137, + 555 + ], + "score": 0.86, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 750, + 1943, + 777, + 1943, + 777, + 1973, + 750, + 1973 + ], + "score": 0.84, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1328, + 697, + 1354, + 697, + 1354, + 723, + 1328, + 723 + ], + "score": 0.82, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 391, + 1478, + 446, + 1478, + 446, + 1503, + 391, + 1503 + ], + "score": 0.78, + "latex": "{ < } 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 978, + 841, + 996, + 841, + 996, + 863, + 978, + 863 + ], + "score": 0.77, + "latex": "v" + }, + { + "category_id": 13, + "poly": [ + 1328, + 732, + 1345, + 732, + 1345, + 753, + 1328, + 753 + ], + "score": 0.77, + "latex": "v" + }, + { + "category_id": 13, + "poly": [ + 730, + 872, + 749, + 872, + 749, + 894, + 730, + 894 + ], + "score": 0.71, + "latex": "v" + }, + { + "category_id": 13, + "poly": [ + 298, + 872, + 315, + 872, + 315, + 894, + 298, + 894 + ], + "score": 0.71, + "latex": "v" + }, + { + "category_id": 13, + "poly": [ + 298, + 726, + 315, + 726, + 315, + 753, + 298, + 753 + ], + "score": 0.61, + "latex": "\\theta" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1419.0, + 1406.0, + 1419.0, + 1406.0, + 1452.0, + 294.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1449.0, + 1406.0, + 1449.0, + 1406.0, + 1478.0, + 293.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1475.0, + 390.0, + 1475.0, + 390.0, + 1511.0, + 293.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1475.0, + 718.0, + 1475.0, + 718.0, + 1511.0, + 447.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1753.0, + 676.0, + 1753.0, + 676.0, + 1790.0, + 295.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1038.0, + 1098.0, + 1038.0, + 1098.0, + 1075.0, + 294.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 566.0, + 226.0, + 566.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1099.0, + 1406.0, + 1099.0, + 1406.0, + 1137.0, + 294.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1133.0, + 1404.0, + 1133.0, + 1404.0, + 1166.0, + 297.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1165.0, + 1407.0, + 1165.0, + 1407.0, + 1198.0, + 295.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1191.0, + 1406.0, + 1191.0, + 1406.0, + 1229.0, + 294.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1225.0, + 1408.0, + 1225.0, + 1408.0, + 1258.0, + 295.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1253.0, + 1406.0, + 1253.0, + 1406.0, + 1291.0, + 294.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1285.0, + 1406.0, + 1285.0, + 1406.0, + 1319.0, + 292.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1317.0, + 1406.0, + 1317.0, + 1406.0, + 1349.0, + 296.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 980.0, + 1344.0, + 980.0, + 1384.0, + 294.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 398.0, + 1405.0, + 398.0, + 1405.0, + 439.0, + 293.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 433.0, + 1407.0, + 433.0, + 1407.0, + 469.0, + 293.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 464.0, + 616.0, + 464.0, + 616.0, + 498.0, + 294.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 464.0, + 1405.0, + 464.0, + 1405.0, + 498.0, + 657.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 493.0, + 1408.0, + 493.0, + 1408.0, + 529.0, + 293.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 524.0, + 1136.0, + 524.0, + 1136.0, + 558.0, + 296.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 524.0, + 1215.0, + 524.0, + 1215.0, + 558.0, + 1165.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 524.0, + 1406.0, + 524.0, + 1406.0, + 558.0, + 1245.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 553.0, + 1409.0, + 553.0, + 1409.0, + 591.0, + 292.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 585.0, + 739.0, + 585.0, + 739.0, + 622.0, + 504.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 585.0, + 1181.0, + 585.0, + 1181.0, + 622.0, + 854.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 585.0, + 1408.0, + 585.0, + 1408.0, + 622.0, + 1329.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 614.0, + 662.0, + 614.0, + 662.0, + 652.0, + 293.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 614.0, + 1089.0, + 614.0, + 1089.0, + 652.0, + 869.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 614.0, + 1266.0, + 614.0, + 1266.0, + 652.0, + 1217.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 614.0, + 1409.0, + 614.0, + 1409.0, + 652.0, + 1399.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 806.0, + 1402.0, + 806.0, + 1402.0, + 838.0, + 295.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 837.0, + 977.0, + 837.0, + 977.0, + 869.0, + 295.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 837.0, + 1093.0, + 837.0, + 1093.0, + 869.0, + 997.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 837.0, + 1402.0, + 837.0, + 1402.0, + 869.0, + 1167.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 862.0, + 297.0, + 862.0, + 297.0, + 904.0, + 291.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 862.0, + 395.0, + 862.0, + 395.0, + 904.0, + 316.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 862.0, + 729.0, + 862.0, + 729.0, + 904.0, + 473.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 862.0, + 947.0, + 862.0, + 947.0, + 904.0, + 750.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 862.0, + 1406.0, + 862.0, + 1406.0, + 904.0, + 1029.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 891.0, + 1405.0, + 891.0, + 1405.0, + 936.0, + 292.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 926.0, + 1405.0, + 926.0, + 1405.0, + 963.0, + 294.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 956.0, + 1126.0, + 956.0, + 1126.0, + 993.0, + 292.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1407.0, + 1849.0, + 1407.0, + 1887.0, + 294.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1407.0, + 1882.0, + 1407.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1914.0, + 1403.0, + 1914.0, + 1403.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 445.0, + 1941.0, + 445.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1941.0, + 700.0, + 1941.0, + 700.0, + 1979.0, + 557.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1941.0, + 749.0, + 1941.0, + 749.0, + 1979.0, + 740.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1941.0, + 824.0, + 1941.0, + 824.0, + 1979.0, + 778.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1979.0, + 853.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2003.0, + 1407.0, + 2003.0, + 1407.0, + 2037.0, + 292.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 663.0, + 1405.0, + 663.0, + 1405.0, + 700.0, + 294.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 696.0, + 1327.0, + 696.0, + 1327.0, + 728.0, + 295.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 696.0, + 1405.0, + 696.0, + 1405.0, + 728.0, + 1355.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 722.0, + 297.0, + 722.0, + 297.0, + 762.0, + 292.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 722.0, + 1327.0, + 722.0, + 1327.0, + 762.0, + 316.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 722.0, + 1409.0, + 722.0, + 1409.0, + 762.0, + 1346.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 759.0, + 973.0, + 759.0, + 973.0, + 789.0, + 293.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1407.0, + 290.0, + 1407.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 854.0, + 322.0, + 854.0, + 359.0, + 294.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 322.0, + 1404.0, + 322.0, + 1404.0, + 359.0, + 947.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 354.0, + 297.0, + 354.0, + 297.0, + 388.0, + 293.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 354.0, + 862.0, + 354.0, + 862.0, + 388.0, + 370.0, + 388.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1219, + 1405, + 1219, + 1405, + 1465, + 297, + 1465 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 908, + 1408, + 908, + 1408, + 1124, + 298, + 1124 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 569, + 1406, + 569, + 1406, + 795, + 297, + 795 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1479, + 1404, + 1479, + 1404, + 1695, + 298, + 1695 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 296, + 367, + 1407, + 367, + 1407, + 554, + 296, + 554 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1711, + 1404, + 1711, + 1404, + 1866, + 298, + 1866 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1402, + 1880, + 1402, + 2035, + 299, + 2035 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 436, + 250, + 1257, + 250, + 1257, + 348, + 436, + 348 + ], + "score": 0.953 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 813, + 74, + 813, + 105, + 298, + 105 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 1161, + 660, + 1161, + 660, + 1191, + 299, + 1191 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 301, + 839, + 707, + 839, + 707, + 874, + 301, + 874 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.779 + }, + { + "category_id": 14, + "poly": [ + 437, + 250, + 1263, + 250, + 1263, + 348, + 437, + 348 + ], + "score": 0.93, + "latex": "v ^ { t + 1 } = \\prod _ { | | v | | _ { \\infty } < \\epsilon } \\left( v ^ { t } + \\alpha \\operatorname { s g n } \\left( { \\nabla } _ { v } \\sum _ { x \\in D ^ { \\prime } } \\sum _ { 0 < i < K } { \\mathcal { L } } \\left( F _ { i } \\left( x + v \\right) , \\bar { y } \\right) \\right) \\right)" + }, + { + "category_id": 13, + "poly": [ + 726, + 522, + 792, + 522, + 792, + 555, + 726, + 555 + ], + "score": 0.92, + "latex": "F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 660, + 429, + 772, + 429, + 772, + 463, + 660, + 463 + ], + "score": 0.92, + "latex": "{ \\mathcal { D } } ^ { \\prime } = \\{ x \\}" + }, + { + "category_id": 13, + "poly": [ + 692, + 733, + 778, + 733, + 778, + 765, + 692, + 765 + ], + "score": 0.92, + "latex": "( x + v )" + }, + { + "category_id": 13, + "poly": [ + 346, + 729, + 480, + 729, + 480, + 765, + 346, + 765 + ], + "score": 0.91, + "latex": "\\bar { y } ^ { ( i ) } = 1 / m" + }, + { + "category_id": 13, + "poly": [ + 1337, + 734, + 1403, + 734, + 1403, + 766, + 1337, + 766 + ], + "score": 0.91, + "latex": "F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 298, + 399, + 418, + 399, + 418, + 433, + 298, + 433 + ], + "score": 0.91, + "latex": "| | v | | _ { \\infty } < \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 297, + 490, + 594, + 490, + 594, + 524, + 297, + 524 + ], + "score": 0.91, + "latex": "\\mathcal { D } ^ { \\prime } = \\{ ( x , y ) \\in \\mathcal { D } | y = i \\}" + }, + { + "category_id": 13, + "poly": [ + 401, + 570, + 467, + 570, + 467, + 603, + 401, + 603 + ], + "score": 0.9, + "latex": "F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 298, + 661, + 637, + 661, + 637, + 705, + 298, + 705 + ], + "score": 0.9, + "latex": "\\mathrm { a r g m a x } _ { j \\in M } \\bar { F } _ { i } ^ { ( j ) } ( x ) = \\bar { y } \\bar { \\ne y }" + }, + { + "category_id": 13, + "poly": [ + 485, + 1001, + 523, + 1001, + 523, + 1031, + 485, + 1031 + ], + "score": 0.9, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 647, + 911, + 684, + 911, + 684, + 940, + 647, + 940 + ], + "score": 0.89, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 1320, + 461, + 1396, + 461, + 1396, + 489, + 1320, + 489 + ], + "score": 0.89, + "latex": "i \\in M" + }, + { + "category_id": 13, + "poly": [ + 1118, + 571, + 1147, + 571, + 1147, + 601, + 1118, + 601 + ], + "score": 0.88, + "latex": "F _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1036, + 489, + 1071, + 489, + 1071, + 518, + 1036, + 518 + ], + "score": 0.88, + "latex": "i ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 960, + 1312, + 1054, + 1312, + 1054, + 1342, + 960, + 1342 + ], + "score": 0.88, + "latex": "7 5 - 9 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 739, + 1032, + 767, + 1032, + 767, + 1062, + 739, + 1062 + ], + "score": 0.88, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 661, + 1031, + 689, + 1031, + 689, + 1062, + 661, + 1062 + ], + "score": 0.88, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 460, + 390, + 460, + 390, + 488, + 298, + 488 + ], + "score": 0.87, + "latex": "\\mathcal { D } ^ { \\prime } = \\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 1281, + 941, + 1309, + 941, + 1309, + 971, + 1281, + 971 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1069, + 400, + 1102, + 400, + 1102, + 427, + 1069, + 427 + ], + "score": 0.87, + "latex": "\\mathcal { D } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 941, + 1230, + 941, + 1230, + 970, + 1202, + 970 + ], + "score": 0.87, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 357, + 1252, + 404, + 1252, + 404, + 1279, + 357, + 1279 + ], + "score": 0.86, + "latex": "\\sim 0" + }, + { + "category_id": 13, + "poly": [ + 1295, + 1835, + 1380, + 1835, + 1380, + 1864, + 1295, + 1864 + ], + "score": 0.83, + "latex": "{ \\sim } 1 0 \\mathrm { m s }" + }, + { + "category_id": 13, + "poly": [ + 879, + 703, + 897, + 703, + 897, + 732, + 879, + 732 + ], + "score": 0.83, + "latex": "\\bar { y }" + }, + { + "category_id": 13, + "poly": [ + 1257, + 526, + 1274, + 526, + 1274, + 554, + 1257, + 554 + ], + "score": 0.83, + "latex": "\\bar { y }" + }, + { + "category_id": 13, + "poly": [ + 1242, + 493, + 1261, + 493, + 1261, + 523, + 1242, + 523 + ], + "score": 0.82, + "latex": "\\bar { y }" + }, + { + "category_id": 13, + "poly": [ + 469, + 401, + 491, + 401, + 491, + 428, + 469, + 428 + ], + "score": 0.81, + "latex": "\\mathcal { L }" + }, + { + "category_id": 13, + "poly": [ + 366, + 372, + 380, + 372, + 380, + 396, + 366, + 396 + ], + "score": 0.8, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1383, + 604, + 1401, + 604, + 1401, + 633, + 1383, + 633 + ], + "score": 0.77, + "latex": "\\bar { y }" + }, + { + "category_id": 13, + "poly": [ + 725, + 375, + 747, + 375, + 747, + 396, + 725, + 396 + ], + "score": 0.77, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 419, + 1943, + 506, + 1943, + 506, + 1972, + 419, + 1972 + ], + "score": 0.77, + "latex": "{ \\sim } 1 1 \\mathrm { m s }" + }, + { + "category_id": 13, + "poly": [ + 931, + 369, + 956, + 369, + 956, + 396, + 931, + 396 + ], + "score": 0.6, + "latex": "\\Pi" + }, + { + "category_id": 13, + "poly": [ + 721, + 1512, + 821, + 1512, + 821, + 1542, + 721, + 1542 + ], + "score": 0.58, + "latex": "( F L O P s )" + }, + { + "category_id": 13, + "poly": [ + 533, + 602, + 573, + 602, + 573, + 630, + 533, + 630 + ], + "score": 0.27, + "latex": "\\mathrm { H u }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1160.0, + 663.0, + 1160.0, + 663.0, + 1195.0, + 294.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 835.0, + 713.0, + 835.0, + 713.0, + 881.0, + 292.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1217.0, + 1406.0, + 1217.0, + 1406.0, + 1256.0, + 291.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1251.0, + 356.0, + 1251.0, + 356.0, + 1284.0, + 294.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1251.0, + 1405.0, + 1251.0, + 1405.0, + 1284.0, + 405.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1282.0, + 1405.0, + 1282.0, + 1405.0, + 1316.0, + 295.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1312.0, + 959.0, + 1312.0, + 959.0, + 1346.0, + 295.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1312.0, + 1408.0, + 1312.0, + 1408.0, + 1346.0, + 1055.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1343.0, + 1405.0, + 1343.0, + 1405.0, + 1377.0, + 295.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1371.0, + 1406.0, + 1371.0, + 1406.0, + 1407.0, + 292.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1403.0, + 1403.0, + 1403.0, + 1403.0, + 1437.0, + 292.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 1364.0, + 1432.0, + 1364.0, + 1468.0, + 294.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 907.0, + 646.0, + 907.0, + 646.0, + 944.0, + 292.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 907.0, + 1408.0, + 907.0, + 1408.0, + 944.0, + 685.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 938.0, + 1201.0, + 938.0, + 1201.0, + 975.0, + 292.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 938.0, + 1280.0, + 938.0, + 1280.0, + 975.0, + 1231.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 938.0, + 1408.0, + 938.0, + 1408.0, + 975.0, + 1310.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 970.0, + 1406.0, + 970.0, + 1406.0, + 1005.0, + 295.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1000.0, + 484.0, + 1000.0, + 484.0, + 1036.0, + 293.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1000.0, + 1405.0, + 1000.0, + 1405.0, + 1036.0, + 524.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1027.0, + 660.0, + 1027.0, + 660.0, + 1068.0, + 292.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1027.0, + 738.0, + 1027.0, + 738.0, + 1068.0, + 690.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1027.0, + 1408.0, + 1027.0, + 1408.0, + 1068.0, + 768.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1059.0, + 1405.0, + 1059.0, + 1405.0, + 1096.0, + 293.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1090.0, + 1058.0, + 1090.0, + 1058.0, + 1127.0, + 293.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 569.0, + 400.0, + 569.0, + 400.0, + 604.0, + 295.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 569.0, + 1117.0, + 569.0, + 1117.0, + 604.0, + 468.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 569.0, + 1409.0, + 569.0, + 1409.0, + 604.0, + 1148.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 597.0, + 532.0, + 597.0, + 532.0, + 640.0, + 292.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 597.0, + 1382.0, + 597.0, + 1382.0, + 640.0, + 574.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 597.0, + 1409.0, + 597.0, + 1409.0, + 640.0, + 1402.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 627.0, + 1409.0, + 627.0, + 1409.0, + 671.0, + 292.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 654.0, + 297.0, + 654.0, + 297.0, + 718.0, + 285.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 654.0, + 1417.0, + 654.0, + 1417.0, + 718.0, + 638.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 699.0, + 878.0, + 699.0, + 878.0, + 734.0, + 295.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 699.0, + 1408.0, + 699.0, + 1408.0, + 734.0, + 898.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 724.0, + 345.0, + 724.0, + 345.0, + 771.0, + 290.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 724.0, + 691.0, + 724.0, + 691.0, + 771.0, + 481.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 724.0, + 1336.0, + 724.0, + 1336.0, + 771.0, + 779.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 724.0, + 1409.0, + 724.0, + 1409.0, + 771.0, + 1404.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 758.0, + 915.0, + 758.0, + 915.0, + 799.0, + 292.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1477.0, + 1406.0, + 1477.0, + 1406.0, + 1517.0, + 292.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1509.0, + 720.0, + 1509.0, + 720.0, + 1549.0, + 292.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1509.0, + 1409.0, + 1509.0, + 1409.0, + 1549.0, + 822.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1541.0, + 1406.0, + 1541.0, + 1406.0, + 1576.0, + 294.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1572.0, + 1406.0, + 1572.0, + 1406.0, + 1607.0, + 294.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1603.0, + 1404.0, + 1603.0, + 1404.0, + 1638.0, + 294.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1631.0, + 1402.0, + 1631.0, + 1402.0, + 1666.0, + 294.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1665.0, + 1186.0, + 1665.0, + 1186.0, + 1699.0, + 296.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 364.0, + 365.0, + 364.0, + 365.0, + 403.0, + 293.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 364.0, + 724.0, + 364.0, + 724.0, + 403.0, + 381.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 364.0, + 930.0, + 364.0, + 930.0, + 403.0, + 748.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 364.0, + 1405.0, + 364.0, + 1405.0, + 403.0, + 957.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 396.0, + 468.0, + 396.0, + 468.0, + 436.0, + 419.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 396.0, + 1068.0, + 396.0, + 1068.0, + 436.0, + 492.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 396.0, + 1407.0, + 396.0, + 1407.0, + 436.0, + 1103.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 425.0, + 659.0, + 425.0, + 659.0, + 467.0, + 291.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 425.0, + 1410.0, + 425.0, + 1410.0, + 467.0, + 773.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 455.0, + 297.0, + 455.0, + 297.0, + 496.0, + 291.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 455.0, + 1319.0, + 455.0, + 1319.0, + 496.0, + 391.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 455.0, + 1410.0, + 455.0, + 1410.0, + 496.0, + 1397.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 485.0, + 296.0, + 485.0, + 296.0, + 528.0, + 293.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 485.0, + 1035.0, + 485.0, + 1035.0, + 528.0, + 595.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 485.0, + 1241.0, + 485.0, + 1241.0, + 528.0, + 1072.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 485.0, + 1408.0, + 485.0, + 1408.0, + 528.0, + 1262.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 519.0, + 725.0, + 519.0, + 725.0, + 558.0, + 293.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 519.0, + 1256.0, + 519.0, + 1256.0, + 558.0, + 793.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 519.0, + 1284.0, + 519.0, + 1284.0, + 558.0, + 1275.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1711.0, + 1409.0, + 1711.0, + 1409.0, + 1749.0, + 294.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1739.0, + 1408.0, + 1739.0, + 1408.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 1406.0, + 1771.0, + 1406.0, + 1810.0, + 292.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1832.0, + 1294.0, + 1832.0, + 1294.0, + 1873.0, + 292.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1381.0, + 1832.0, + 1392.0, + 1832.0, + 1392.0, + 1873.0, + 1381.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1915.0, + 295.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1938.0, + 418.0, + 1938.0, + 418.0, + 1981.0, + 293.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 1938.0, + 1408.0, + 1938.0, + 1408.0, + 1981.0, + 507.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1406.0, + 2000.0, + 1406.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 296, + 337, + 1400, + 337, + 1400, + 849, + 296, + 849 + ], + "score": 0.983, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 / 85%0.89 / 85%0.77 /88%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 /57%0.51 / 52%0.42 /57%0.59 /51%
DEEPSLOTH
C100.06 / 17%0.06 / 17%0.01 /13%0.04 / 16%0.01 /12%0.06 /16%
TI0.06 /7%0.06 /7%0.00 /2%0.01/2%0.02 / 6%0.04 / 6%
UNIVERSAL DEEPSLOTH
C100.85 / 65%0.85 /65%0.62 / 65%0.86 / 60%0.73 / 61%0.90 / 59%
TI0.58 / 46%0.81 /41%0.31 / 47%0.44 / 44%0.33 / 47%0.51 / 43%
CLASS-UNIVERSAL DEEPSLOTH
C100.82/32%0.82 /32%0.47 /35%0.78 /33%0.60 /30%0.85 /27%
TI0.41 / 21%0.71 /17%0.20 /28%0.33 /27%0.21 /27%0.38 /25%
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1217, + 1404, + 1217, + 1404, + 1464, + 298, + 1464 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1478, + 1405, + 1478, + 1405, + 1693, + 297, + 1693 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1709, + 1404, + 1709, + 1404, + 1956, + 298, + 1956 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1016, + 1403, + 1016, + 1403, + 1201, + 298, + 1201 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 909, + 1403, + 909, + 1403, + 1001, + 298, + 1001 + ], + "score": 0.973 + }, + { + "category_id": 6, + "poly": [ + 297, + 226, + 1405, + 226, + 1405, + 311, + 297, + 311 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1400, + 1972, + 1400, + 2034, + 298, + 2034 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 104, + 300, + 104 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.759 + }, + { + "category_id": 13, + "poly": [ + 565, + 1430, + 730, + 1430, + 730, + 1465, + 565, + 1465 + ], + "score": 0.93, + "latex": "v \\sim U ^ { d } ( - \\epsilon , \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 390, + 1540, + 465, + 1540, + 465, + 1569, + 390, + 1569 + ], + "score": 0.9, + "latex": "{ \\sim } 1 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1200, + 1079, + 1290, + 1079, + 1290, + 1108, + 1200, + 1108 + ], + "score": 0.9, + "latex": "9 9 . 9 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 536, + 1510, + 611, + 1510, + 611, + 1539, + 536, + 1539 + ], + "score": 0.89, + "latex": "{ \\sim } 1 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1149, + 910, + 1189, + 910, + 1189, + 939, + 1149, + 939 + ], + "score": 0.89, + "latex": "5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1198, + 1280, + 1293, + 1280, + 1293, + 1309, + 1198, + 1309 + ], + "score": 0.89, + "latex": "1 5 \\mathrm { - } 2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 752, + 1371, + 846, + 1371, + 846, + 1400, + 752, + 1400 + ], + "score": 0.89, + "latex": "12 \\mathrm { - } 2 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 606, + 1631, + 701, + 1631, + 701, + 1660, + 606, + 1660 + ], + "score": 0.88, + "latex": "5 0 \\mathrm { - } 6 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1631, + 378, + 1631, + 378, + 1660, + 297, + 1660 + ], + "score": 0.88, + "latex": "8 - 5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1271, + 1433, + 1332, + 1433, + 1332, + 1461, + 1271, + 1461 + ], + "score": 0.88, + "latex": "{ \\sim } 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 1093, + 1079, + 1149, + 1079, + 1149, + 1108, + 1093, + 1108 + ], + "score": 0.88, + "latex": "96 \\%" + }, + { + "category_id": 13, + "poly": [ + 606, + 227, + 640, + 227, + 640, + 254, + 606, + 254 + ], + "score": 0.88, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 1238, + 910, + 1293, + 910, + 1293, + 939, + 1238, + 939 + ], + "score": 0.88, + "latex": "67 \\%" + }, + { + "category_id": 13, + "poly": [ + 872, + 1280, + 956, + 1280, + 956, + 1309, + 872, + 1309 + ], + "score": 0.87, + "latex": "3- 2 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 1180, + 1341, + 1248, + 1341, + 1248, + 1370, + 1180, + 1370 + ], + "score": 0.87, + "latex": "3- 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 880, + 1171, + 962, + 1171, + 962, + 1199, + 880, + 1199 + ], + "score": 0.86, + "latex": "1 . 5 – 5 \\times" + }, + { + "category_id": 13, + "poly": [ + 1140, + 1141, + 1210, + 1141, + 1210, + 1169, + 1140, + 1169 + ], + "score": 0.83, + "latex": "{ \\sim } 2 \\mathrm { m s }" + }, + { + "category_id": 13, + "poly": [ + 1098, + 1111, + 1185, + 1111, + 1185, + 1138, + 1098, + 1138 + ], + "score": 0.82, + "latex": "{ \\sim } 1 1 \\mathrm { m s }" + }, + { + "category_id": 13, + "poly": [ + 785, + 227, + 915, + 227, + 915, + 254, + 785, + 254 + ], + "score": 0.73, + "latex": "\\mathrm { 2 A D { < } } 5 , 1 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1033, + 1310, + 1105, + 1310, + 1105, + 1339, + 1033, + 1339 + ], + "score": 0.63, + "latex": "{ < } 1 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 684, + 971, + 756, + 971, + 756, + 1000, + 684, + 1000 + ], + "score": 0.56, + "latex": "0 . 5 \\mathrm { m s }" + }, + { + "category_id": 13, + "poly": [ + 776, + 282, + 835, + 282, + 835, + 309, + 776, + 309 + ], + "score": 0.44, + "latex": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 805, + 971, + 876, + 971, + 876, + 1000, + 805, + 1000 + ], + "score": 0.44, + "latex": "7 . 4 \\mathrm { m s }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 220.0, + 605.0, + 220.0, + 605.0, + 262.0, + 291.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 220.0, + 784.0, + 220.0, + 784.0, + 262.0, + 641.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 220.0, + 1409.0, + 220.0, + 1409.0, + 262.0, + 916.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 253.0, + 1406.0, + 253.0, + 1406.0, + 286.0, + 295.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 282.0, + 775.0, + 282.0, + 775.0, + 312.0, + 294.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 282.0, + 963.0, + 282.0, + 963.0, + 312.0, + 836.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1218.0, + 1405.0, + 1218.0, + 1405.0, + 1251.0, + 296.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1249.0, + 1406.0, + 1249.0, + 1406.0, + 1281.0, + 293.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1280.0, + 871.0, + 1280.0, + 871.0, + 1314.0, + 296.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1280.0, + 1197.0, + 1280.0, + 1197.0, + 1314.0, + 957.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1280.0, + 1404.0, + 1280.0, + 1404.0, + 1314.0, + 1294.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1309.0, + 1032.0, + 1309.0, + 1032.0, + 1347.0, + 293.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1309.0, + 1408.0, + 1309.0, + 1408.0, + 1347.0, + 1106.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1336.0, + 1179.0, + 1336.0, + 1179.0, + 1378.0, + 292.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1336.0, + 1408.0, + 1336.0, + 1408.0, + 1378.0, + 1249.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1368.0, + 751.0, + 1368.0, + 751.0, + 1406.0, + 293.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1368.0, + 1404.0, + 1368.0, + 1404.0, + 1406.0, + 847.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1402.0, + 1404.0, + 1402.0, + 1404.0, + 1436.0, + 294.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1430.0, + 564.0, + 1430.0, + 564.0, + 1466.0, + 293.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1430.0, + 1270.0, + 1430.0, + 1270.0, + 1466.0, + 731.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 1430.0, + 1342.0, + 1430.0, + 1342.0, + 1466.0, + 1333.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1477.0, + 1406.0, + 1477.0, + 1406.0, + 1510.0, + 292.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1510.0, + 535.0, + 1510.0, + 535.0, + 1541.0, + 295.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1510.0, + 1407.0, + 1510.0, + 1407.0, + 1541.0, + 612.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1536.0, + 389.0, + 1536.0, + 389.0, + 1578.0, + 292.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1536.0, + 1410.0, + 1536.0, + 1410.0, + 1578.0, + 466.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1571.0, + 1406.0, + 1571.0, + 1406.0, + 1601.0, + 296.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1598.0, + 1406.0, + 1598.0, + 1406.0, + 1637.0, + 292.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1629.0, + 296.0, + 1629.0, + 296.0, + 1665.0, + 292.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1629.0, + 605.0, + 1629.0, + 605.0, + 1665.0, + 379.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1629.0, + 1408.0, + 1629.0, + 1408.0, + 1665.0, + 702.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1663.0, + 1115.0, + 1663.0, + 1115.0, + 1694.0, + 295.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1708.0, + 1405.0, + 1708.0, + 1405.0, + 1746.0, + 293.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 1405.0, + 1741.0, + 1405.0, + 1775.0, + 294.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1772.0, + 1406.0, + 1772.0, + 1406.0, + 1806.0, + 296.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1836.0, + 296.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1833.0, + 1406.0, + 1833.0, + 1406.0, + 1867.0, + 292.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1863.0, + 1406.0, + 1863.0, + 1406.0, + 1896.0, + 294.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1893.0, + 1406.0, + 1893.0, + 1406.0, + 1926.0, + 293.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1924.0, + 1408.0, + 1924.0, + 1408.0, + 1958.0, + 294.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1016.0, + 1407.0, + 1016.0, + 1407.0, + 1053.0, + 293.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1047.0, + 1407.0, + 1047.0, + 1407.0, + 1082.0, + 293.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1078.0, + 1092.0, + 1078.0, + 1092.0, + 1114.0, + 294.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1078.0, + 1199.0, + 1078.0, + 1199.0, + 1114.0, + 1150.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 1078.0, + 1407.0, + 1078.0, + 1407.0, + 1114.0, + 1291.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1108.0, + 1097.0, + 1108.0, + 1097.0, + 1144.0, + 294.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1108.0, + 1405.0, + 1108.0, + 1405.0, + 1144.0, + 1186.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1139.0, + 1139.0, + 1139.0, + 1139.0, + 1175.0, + 294.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1139.0, + 1405.0, + 1139.0, + 1405.0, + 1175.0, + 1211.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1167.0, + 879.0, + 1167.0, + 879.0, + 1205.0, + 293.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1167.0, + 1321.0, + 1167.0, + 1321.0, + 1205.0, + 963.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 908.0, + 1148.0, + 908.0, + 1148.0, + 945.0, + 294.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 908.0, + 1237.0, + 908.0, + 1237.0, + 945.0, + 1190.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 908.0, + 1407.0, + 908.0, + 1407.0, + 945.0, + 1294.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 938.0, + 1405.0, + 938.0, + 1405.0, + 975.0, + 294.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 968.0, + 683.0, + 968.0, + 683.0, + 1005.0, + 295.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 968.0, + 804.0, + 968.0, + 804.0, + 1005.0, + 757.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 968.0, + 1030.0, + 968.0, + 1030.0, + 1005.0, + 877.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1970.0, + 1402.0, + 1970.0, + 1402.0, + 2006.0, + 296.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1759, + 1405, + 1759, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 620, + 1404, + 620, + 1404, + 986, + 298, + 986 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1326, + 1403, + 1326, + 1403, + 1540, + 298, + 1540 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1558, + 1403, + 1558, + 1403, + 1741, + 298, + 1741 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1124, + 1405, + 1124, + 1405, + 1308, + 299, + 1308 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 305, + 227, + 1396, + 227, + 1396, + 484, + 305, + 484 + ], + "score": 0.962 + }, + { + "category_id": 4, + "poly": [ + 296, + 493, + 1402, + 493, + 1402, + 550, + 296, + 550 + ], + "score": 0.948 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.788 + }, + { + "category_id": 1, + "poly": [ + 301, + 1054, + 1200, + 1054, + 1200, + 1086, + 301, + 1086 + ], + "score": 0.464 + }, + { + "category_id": 0, + "poly": [ + 301, + 1054, + 1200, + 1054, + 1200, + 1086, + 301, + 1086 + ], + "score": 0.441 + }, + { + "category_id": 13, + "poly": [ + 731, + 773, + 798, + 773, + 798, + 805, + 731, + 805 + ], + "score": 0.93, + "latex": "F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 555, + 1791, + 593, + 1791, + 593, + 1823, + 555, + 1823 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { \\boldsymbol { \\mathcal { J } } }" + }, + { + "category_id": 13, + "poly": [ + 789, + 1418, + 844, + 1418, + 844, + 1447, + 789, + 1447 + ], + "score": 0.88, + "latex": "78 \\%" + }, + { + "category_id": 13, + "poly": [ + 411, + 1479, + 467, + 1479, + 467, + 1508, + 411, + 1508 + ], + "score": 0.88, + "latex": "73 \\%" + }, + { + "category_id": 13, + "poly": [ + 1348, + 895, + 1403, + 895, + 1403, + 924, + 1348, + 924 + ], + "score": 0.87, + "latex": "13 \\%" + }, + { + "category_id": 13, + "poly": [ + 825, + 865, + 879, + 865, + 879, + 893, + 825, + 893 + ], + "score": 0.86, + "latex": "81 \\%" + }, + { + "category_id": 13, + "poly": [ + 757, + 1680, + 811, + 1680, + 811, + 1709, + 757, + 1709 + ], + "score": 0.86, + "latex": "74 \\%" + }, + { + "category_id": 13, + "poly": [ + 1276, + 1851, + 1331, + 1851, + 1331, + 1880, + 1276, + 1880 + ], + "score": 0.86, + "latex": "83 \\%" + }, + { + "category_id": 13, + "poly": [ + 655, + 1680, + 709, + 1680, + 709, + 1709, + 655, + 1709 + ], + "score": 0.86, + "latex": "72 \\%" + }, + { + "category_id": 13, + "poly": [ + 629, + 895, + 671, + 895, + 671, + 924, + 629, + 924 + ], + "score": 0.86, + "latex": "4 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1619, + 351, + 1619, + 351, + 1648, + 297, + 1648 + ], + "score": 0.86, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 1227, + 1589, + 1281, + 1589, + 1281, + 1618, + 1227, + 1618 + ], + "score": 0.85, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 918, + 1418, + 973, + 1418, + 973, + 1448, + 918, + 1448 + ], + "score": 0.85, + "latex": "8 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1298, + 1620, + 1353, + 1620, + 1353, + 1648, + 1298, + 1648 + ], + "score": 0.85, + "latex": "2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1620, + 1288, + 1620, + 1288, + 1648, + 1235, + 1648 + ], + "score": 0.84, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 1291, + 1589, + 1346, + 1589, + 1346, + 1617, + 1291, + 1617 + ], + "score": 0.84, + "latex": "2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 530, + 1680, + 582, + 1680, + 582, + 1709, + 530, + 1709 + ], + "score": 0.84, + "latex": "81 \\%" + }, + { + "category_id": 13, + "poly": [ + 592, + 1680, + 645, + 1680, + 645, + 1710, + 592, + 1710 + ], + "score": 0.84, + "latex": "73 \\%" + }, + { + "category_id": 13, + "poly": [ + 1135, + 1278, + 1194, + 1278, + 1194, + 1306, + 1135, + 1306 + ], + "score": 0.84, + "latex": "{ < } 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 882, + 1680, + 935, + 1680, + 935, + 1710, + 882, + 1710 + ], + "score": 0.84, + "latex": "8 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1650, + 352, + 1650, + 352, + 1678, + 297, + 1678 + ], + "score": 0.83, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1881, + 355, + 1881, + 355, + 1912, + 297, + 1912 + ], + "score": 0.82, + "latex": "8 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 650, + 1479, + 706, + 1479, + 706, + 1509, + 650, + 1509 + ], + "score": 0.82, + "latex": "8 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1322, + 1913, + 1348, + 1913, + 1348, + 1939, + 1322, + 1939 + ], + "score": 0.79, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 1143, + 1791, + 1169, + 1791, + 1169, + 1818, + 1143, + 1818 + ], + "score": 0.79, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1712, + 323, + 1712, + 323, + 1738, + 297, + 1738 + ], + "score": 0.79, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 1142, + 1590, + 1167, + 1590, + 1167, + 1616, + 1142, + 1616 + ], + "score": 0.78, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 1260, + 1883, + 1286, + 1883, + 1286, + 1909, + 1260, + 1909 + ], + "score": 0.78, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 402, + 1388, + 428, + 1388, + 428, + 1415, + 402, + 1415 + ], + "score": 0.68, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 533, + 1590, + 558, + 1590, + 558, + 1616, + 533, + 1616 + ], + "score": 0.66, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 1107, + 834, + 1168, + 834, + 1168, + 863, + 1107, + 863 + ], + "score": 0.58, + "latex": "{ < } 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 241.0, + 846.0, + 241.0, + 846.0, + 254.0, + 834.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 238.0, + 1343.0, + 238.0, + 1343.0, + 263.0, + 1282.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 248.0, + 1161.0, + 248.0, + 1161.0, + 258.0, + 1151.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 267.0, + 713.0, + 267.0, + 713.0, + 284.0, + 697.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 269.0, + 819.0, + 269.0, + 819.0, + 286.0, + 803.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 280.0, + 1008.0, + 280.0, + 1008.0, + 307.0, + 977.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 255.0, + 1388.0, + 255.0, + 1388.0, + 315.0, + 1282.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 317.0, + 1295.0, + 317.0, + 1295.0, + 328.0, + 1282.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 332.0, + 483.0, + 332.0, + 483.0, + 342.0, + 472.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 333.0, + 498.0, + 333.0, + 498.0, + 348.0, + 488.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 325.0, + 1250.0, + 325.0, + 1250.0, + 344.0, + 1215.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 346.0, + 334.0, + 346.0, + 334.0, + 357.0, + 324.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 344.0, + 1106.0, + 344.0, + 1106.0, + 362.0, + 1090.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 365.0, + 476.0, + 365.0, + 476.0, + 374.0, + 466.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 375.0, + 836.0, + 375.0, + 836.0, + 394.0, + 815.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 389.0, + 875.0, + 389.0, + 875.0, + 403.0, + 867.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 399.0, + 409.0, + 399.0, + 409.0, + 424.0, + 351.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 399.0, + 785.0, + 399.0, + 785.0, + 424.0, + 724.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 403.0, + 877.0, + 403.0, + 877.0, + 447.0, + 851.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 410.0, + 1184.0, + 410.0, + 1184.0, + 421.0, + 1174.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 403.0, + 1270.0, + 403.0, + 1270.0, + 421.0, + 1255.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 417.0, + 455.0, + 417.0, + 455.0, + 477.0, + 349.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 447.0, + 486.0, + 447.0, + 486.0, + 458.0, + 473.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 427.0, + 531.0, + 427.0, + 531.0, + 459.0, + 488.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 426.0, + 714.0, + 426.0, + 714.0, + 471.0, + 704.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 417.0, + 828.0, + 417.0, + 828.0, + 477.0, + 723.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 457.0, + 1174.0, + 457.0, + 1174.0, + 472.0, + 1164.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.75, + 331.5, + 449.75, + 331.5, + 449.75, + 348.5, + 444.75, + 348.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.75, + 338.5, + 937.75, + 338.5, + 937.75, + 369.0, + 889.75, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 346.0, + 1235.0, + 346.0, + 1235.0, + 366.0, + 1163.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 365.5, + 1276.0, + 365.5, + 1276.0, + 395.5, + 1226.0, + 395.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 490.0, + 1403.0, + 490.0, + 1403.0, + 524.0, + 295.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 519.0, + 1212.0, + 519.0, + 1212.0, + 552.0, + 292.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1052.0, + 1207.0, + 1052.0, + 1207.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1760.0, + 1408.0, + 1760.0, + 1408.0, + 1792.0, + 297.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1788.0, + 554.0, + 1788.0, + 554.0, + 1824.0, + 293.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1788.0, + 1142.0, + 1788.0, + 1142.0, + 1824.0, + 594.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1824.0, + 1170.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1852.0, + 296.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1847.0, + 1275.0, + 1847.0, + 1275.0, + 1886.0, + 292.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 1847.0, + 1406.0, + 1847.0, + 1406.0, + 1886.0, + 1332.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1878.0, + 296.0, + 1878.0, + 296.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1878.0, + 1259.0, + 1878.0, + 1259.0, + 1917.0, + 356.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1878.0, + 1410.0, + 1878.0, + 1410.0, + 1917.0, + 1287.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1321.0, + 1911.0, + 1321.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1947.0, + 1349.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 898.0, + 2002.0, + 898.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 621.0, + 1405.0, + 621.0, + 1405.0, + 655.0, + 294.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 650.0, + 1402.0, + 650.0, + 1402.0, + 684.0, + 294.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 680.0, + 1406.0, + 680.0, + 1406.0, + 716.0, + 293.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 713.0, + 1402.0, + 713.0, + 1402.0, + 744.0, + 296.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 742.0, + 1405.0, + 742.0, + 1405.0, + 776.0, + 294.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 770.0, + 730.0, + 770.0, + 730.0, + 810.0, + 292.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 770.0, + 1406.0, + 770.0, + 1406.0, + 810.0, + 799.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 803.0, + 1408.0, + 803.0, + 1408.0, + 841.0, + 293.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 831.0, + 1106.0, + 831.0, + 1106.0, + 870.0, + 292.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 831.0, + 1408.0, + 831.0, + 1408.0, + 870.0, + 1169.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 864.0, + 824.0, + 864.0, + 824.0, + 901.0, + 293.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 864.0, + 1406.0, + 864.0, + 1406.0, + 901.0, + 880.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 895.0, + 628.0, + 895.0, + 628.0, + 929.0, + 294.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 895.0, + 1347.0, + 895.0, + 1347.0, + 929.0, + 672.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 924.0, + 1406.0, + 924.0, + 1406.0, + 962.0, + 293.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 953.0, + 786.0, + 953.0, + 786.0, + 992.0, + 293.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1324.0, + 1408.0, + 1324.0, + 1408.0, + 1361.0, + 296.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1353.0, + 1405.0, + 1353.0, + 1405.0, + 1394.0, + 293.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1386.0, + 401.0, + 1386.0, + 401.0, + 1422.0, + 293.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1386.0, + 1404.0, + 1386.0, + 1404.0, + 1422.0, + 429.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1418.0, + 788.0, + 1418.0, + 788.0, + 1453.0, + 294.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1418.0, + 917.0, + 1418.0, + 917.0, + 1453.0, + 845.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1418.0, + 1405.0, + 1418.0, + 1405.0, + 1453.0, + 974.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1449.0, + 1403.0, + 1449.0, + 1403.0, + 1483.0, + 294.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1476.0, + 410.0, + 1476.0, + 410.0, + 1516.0, + 291.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1476.0, + 649.0, + 1476.0, + 649.0, + 1516.0, + 468.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1476.0, + 1407.0, + 1476.0, + 1407.0, + 1516.0, + 707.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1508.0, + 789.0, + 1508.0, + 789.0, + 1545.0, + 293.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1557.0, + 1405.0, + 1557.0, + 1405.0, + 1593.0, + 294.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1588.0, + 532.0, + 1588.0, + 532.0, + 1622.0, + 293.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1588.0, + 1141.0, + 1588.0, + 1141.0, + 1622.0, + 559.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1588.0, + 1226.0, + 1588.0, + 1226.0, + 1622.0, + 1168.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1588.0, + 1290.0, + 1588.0, + 1290.0, + 1622.0, + 1282.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 1588.0, + 1405.0, + 1588.0, + 1405.0, + 1622.0, + 1347.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1614.0, + 296.0, + 1614.0, + 296.0, + 1655.0, + 292.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1614.0, + 1234.0, + 1614.0, + 1234.0, + 1655.0, + 352.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 1614.0, + 1297.0, + 1614.0, + 1297.0, + 1655.0, + 1289.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1614.0, + 1407.0, + 1614.0, + 1407.0, + 1655.0, + 1354.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1647.0, + 296.0, + 1647.0, + 296.0, + 1684.0, + 293.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1647.0, + 1405.0, + 1647.0, + 1405.0, + 1684.0, + 353.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1679.0, + 529.0, + 1679.0, + 529.0, + 1714.0, + 293.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1679.0, + 591.0, + 1679.0, + 591.0, + 1714.0, + 583.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1679.0, + 654.0, + 1679.0, + 654.0, + 1714.0, + 646.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1679.0, + 756.0, + 1679.0, + 756.0, + 1714.0, + 710.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1679.0, + 881.0, + 1679.0, + 881.0, + 1714.0, + 812.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1679.0, + 1407.0, + 1679.0, + 1407.0, + 1714.0, + 936.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1710.0, + 1330.0, + 1710.0, + 1330.0, + 1746.0, + 324.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1125.0, + 1403.0, + 1125.0, + 1403.0, + 1160.0, + 295.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1153.0, + 1405.0, + 1153.0, + 1405.0, + 1192.0, + 294.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1184.0, + 1410.0, + 1184.0, + 1410.0, + 1222.0, + 292.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1406.0, + 1216.0, + 1406.0, + 1252.0, + 294.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1246.0, + 1406.0, + 1246.0, + 1406.0, + 1281.0, + 293.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1277.0, + 1134.0, + 1277.0, + 1134.0, + 1311.0, + 294.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 1277.0, + 1206.0, + 1277.0, + 1206.0, + 1311.0, + 1195.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1052.0, + 1207.0, + 1052.0, + 1207.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 230, + 1405, + 230, + 1405, + 564, + 298, + 564 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 295, + 1158, + 1400, + 1158, + 1400, + 1410, + 295, + 1410 + ], + "score": 0.983, + "html": "
ADV. TRAININGNO ATTACKPGD-20PGD-20 (AVG.)PGD-20 (MAX.)DEEPSLOTH
UNDEFENDED0.77 /89%0.79/29%0.85 /10%0.81/27%0.01 / 13%
PGD-100.61/ 72%0.55 /38%0.64 /23%0.58 /29%0.33 / 70%
PGD-10 (AVG.)0.53 / 72%0.47 / 36%0.47 /35%0.47 /35%0.32 / 70%
PGD-10 (MAX.)0.57 /72%0.51/37%0.54 / 30%0.52 /34%0.32 / 70%
OURS0.74/72%0.71/38%0.82 /14%0.77 /21%0.44/ 67%
OURS + PGD-100.61/73%0.55 /38%0.63 /23%0.58 /28%0.33 / 70%
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1405, + 1850, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 853, + 1404, + 853, + 1404, + 1039, + 298, + 1039 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 686, + 1404, + 686, + 1404, + 839, + 298, + 839 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1424, + 1404, + 1424, + 1404, + 1731, + 298, + 1731 + ], + "score": 0.975 + }, + { + "category_id": 6, + "poly": [ + 300, + 1070, + 1398, + 1070, + 1398, + 1134, + 300, + 1134 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 298, + 613, + 1306, + 613, + 1306, + 650, + 298, + 650 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 301, + 1779, + 560, + 1779, + 560, + 1814, + 301, + 1814 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.79 + }, + { + "category_id": 13, + "poly": [ + 831, + 383, + 908, + 383, + 908, + 412, + 831, + 412 + ], + "score": 0.89, + "latex": "{ \\sim } 9 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 298, + 383, + 374, + 383, + 374, + 412, + 298, + 412 + ], + "score": 0.89, + "latex": "{ \\sim } 5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1326, + 414, + 1403, + 414, + 1403, + 443, + 1326, + 443 + ], + "score": 0.88, + "latex": "{ \\sim } 3 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 733, + 444, + 787, + 444, + 787, + 473, + 733, + 473 + ], + "score": 0.87, + "latex": "37 \\%" + }, + { + "category_id": 13, + "poly": [ + 306, + 444, + 359, + 444, + 359, + 474, + 306, + 474 + ], + "score": 0.86, + "latex": "3 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1026, + 1457, + 1102, + 1457, + 1102, + 1486, + 1026, + 1486 + ], + "score": 0.86, + "latex": "{ \\sim } 0 . 3 2" + }, + { + "category_id": 13, + "poly": [ + 584, + 1547, + 707, + 1547, + 707, + 1577, + 584, + 1577 + ], + "score": 0.85, + "latex": "0 . 5 3 \\sim 0 . 6 1" + }, + { + "category_id": 13, + "poly": [ + 1256, + 1101, + 1315, + 1101, + 1315, + 1129, + 1256, + 1129 + ], + "score": 0.79, + "latex": "{ < } 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 617, + 1641, + 642, + 1641, + 642, + 1666, + 617, + 1666 + ], + "score": 0.58, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 367, + 1374, + 390, + 1374, + 390, + 1397, + 367, + 1397 + ], + "score": 0.31, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 555, + 1639, + 743, + 1639, + 743, + 1669, + 555, + 1669 + ], + "score": 0.25, + "latex": "\\mathrm { \\ O u r s } + \\mathrm { P G D } { - } 1 0" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1067.0, + 1404.0, + 1067.0, + 1404.0, + 1105.0, + 295.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1096.0, + 1255.0, + 1096.0, + 1255.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1096.0, + 1406.0, + 1096.0, + 1406.0, + 1140.0, + 1316.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 612.0, + 1313.0, + 612.0, + 1313.0, + 653.0, + 294.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1773.0, + 566.0, + 1773.0, + 566.0, + 1824.0, + 291.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 260.0, + 1407.0, + 260.0, + 1407.0, + 296.0, + 292.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 1405.0, + 293.0, + 1405.0, + 324.0, + 296.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 321.0, + 1410.0, + 321.0, + 1410.0, + 359.0, + 293.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1406.0, + 351.0, + 1406.0, + 390.0, + 293.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 383.0, + 830.0, + 383.0, + 830.0, + 418.0, + 375.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 383.0, + 1406.0, + 383.0, + 1406.0, + 418.0, + 909.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 413.0, + 1325.0, + 413.0, + 1325.0, + 448.0, + 293.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 444.0, + 305.0, + 444.0, + 305.0, + 478.0, + 295.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 444.0, + 732.0, + 444.0, + 732.0, + 478.0, + 360.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 444.0, + 1405.0, + 444.0, + 1405.0, + 478.0, + 788.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 475.0, + 1406.0, + 475.0, + 1406.0, + 509.0, + 293.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 503.0, + 1406.0, + 503.0, + 1406.0, + 540.0, + 293.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 536.0, + 794.0, + 536.0, + 794.0, + 567.0, + 294.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1848.0, + 1403.0, + 1848.0, + 1403.0, + 1884.0, + 295.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 1407.0, + 1881.0, + 1407.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1950.0, + 291.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1408.0, + 1941.0, + 1408.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 854.0, + 1404.0, + 854.0, + 1404.0, + 886.0, + 294.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 883.0, + 1404.0, + 883.0, + 1404.0, + 922.0, + 293.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 913.0, + 1409.0, + 913.0, + 1409.0, + 954.0, + 291.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 947.0, + 1406.0, + 947.0, + 1406.0, + 980.0, + 292.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 977.0, + 1405.0, + 977.0, + 1405.0, + 1012.0, + 294.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1007.0, + 1387.0, + 1007.0, + 1387.0, + 1043.0, + 294.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 685.0, + 1408.0, + 685.0, + 1408.0, + 720.0, + 293.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 714.0, + 1406.0, + 714.0, + 1406.0, + 753.0, + 293.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 747.0, + 1404.0, + 747.0, + 1404.0, + 784.0, + 293.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 774.0, + 1404.0, + 774.0, + 1404.0, + 814.0, + 293.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 807.0, + 1407.0, + 807.0, + 1407.0, + 843.0, + 294.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1426.0, + 1404.0, + 1426.0, + 1404.0, + 1458.0, + 296.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1455.0, + 1025.0, + 1455.0, + 1025.0, + 1490.0, + 293.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1455.0, + 1405.0, + 1455.0, + 1405.0, + 1490.0, + 1103.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1486.0, + 1405.0, + 1486.0, + 1405.0, + 1522.0, + 294.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1518.0, + 1405.0, + 1518.0, + 1405.0, + 1550.0, + 296.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1547.0, + 583.0, + 1547.0, + 583.0, + 1580.0, + 292.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1547.0, + 1406.0, + 1547.0, + 1406.0, + 1580.0, + 708.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1578.0, + 1406.0, + 1578.0, + 1406.0, + 1611.0, + 293.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1608.0, + 1405.0, + 1608.0, + 1405.0, + 1644.0, + 294.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1639.0, + 554.0, + 1639.0, + 554.0, + 1673.0, + 293.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1639.0, + 1404.0, + 1639.0, + 1404.0, + 1673.0, + 744.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1667.0, + 1406.0, + 1667.0, + 1406.0, + 1707.0, + 292.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1701.0, + 1406.0, + 1701.0, + 1406.0, + 1731.0, + 294.0, + 1731.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 230, + 1402, + 230, + 1402, + 323, + 298, + 323 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 300, + 438, + 1403, + 438, + 1403, + 498, + 300, + 498 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 299, + 594, + 1406, + 594, + 1406, + 733, + 299, + 733 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 300, + 546, + 488, + 546, + 488, + 578, + 300, + 578 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 301, + 371, + 590, + 371, + 590, + 403, + 301, + 403 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 297, + 755, + 1405, + 755, + 1405, + 895, + 297, + 895 + ], + "score": 0.755 + }, + { + "category_id": 1, + "poly": [ + 294, + 917, + 1402, + 917, + 1402, + 975, + 294, + 975 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 295, + 995, + 1401, + 995, + 1401, + 1053, + 295, + 1053 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 295, + 1074, + 1403, + 1074, + 1403, + 1132, + 295, + 1132 + ], + "score": 0.675 + }, + { + "category_id": 1, + "poly": [ + 292, + 1657, + 1401, + 1657, + 1401, + 1716, + 292, + 1716 + ], + "score": 0.655 + }, + { + "category_id": 1, + "poly": [ + 297, + 1151, + 1404, + 1151, + 1404, + 1240, + 297, + 1240 + ], + "score": 0.653 + }, + { + "category_id": 1, + "poly": [ + 295, + 1471, + 1403, + 1471, + 1403, + 1531, + 295, + 1531 + ], + "score": 0.651 + }, + { + "category_id": 1, + "poly": [ + 297, + 1364, + 1404, + 1364, + 1404, + 1451, + 297, + 1451 + ], + "score": 0.645 + }, + { + "category_id": 1, + "poly": [ + 292, + 1869, + 1400, + 1869, + 1400, + 1928, + 292, + 1928 + ], + "score": 0.634 + }, + { + "category_id": 1, + "poly": [ + 299, + 1550, + 1402, + 1550, + 1402, + 1636, + 299, + 1636 + ], + "score": 0.622 + }, + { + "category_id": 1, + "poly": [ + 297, + 1258, + 1402, + 1258, + 1402, + 1345, + 297, + 1345 + ], + "score": 0.616 + }, + { + "category_id": 1, + "poly": [ + 301, + 1949, + 1402, + 1949, + 1402, + 2033, + 301, + 2033 + ], + "score": 0.612 + }, + { + "category_id": 1, + "poly": [ + 298, + 1735, + 1405, + 1735, + 1405, + 1848, + 298, + 1848 + ], + "score": 0.457 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 544.0, + 492.0, + 544.0, + 492.0, + 584.0, + 296.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 369.0, + 595.0, + 369.0, + 595.0, + 409.0, + 298.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 264.0, + 294.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 1406.0, + 260.0, + 1406.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 1312.0, + 289.0, + 1312.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 433.0, + 1404.0, + 433.0, + 1404.0, + 476.0, + 293.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 465.0, + 568.0, + 465.0, + 568.0, + 503.0, + 295.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 595.0, + 1408.0, + 595.0, + 1408.0, + 626.0, + 297.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 621.0, + 1404.0, + 621.0, + 1404.0, + 652.0, + 324.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 644.0, + 1407.0, + 644.0, + 1407.0, + 686.0, + 319.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 671.0, + 1412.0, + 671.0, + 1412.0, + 715.0, + 318.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 706.0, + 979.0, + 706.0, + 979.0, + 737.0, + 323.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 751.0, + 1032.0, + 751.0, + 1032.0, + 788.0, + 292.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 754.0, + 1406.0, + 754.0, + 1406.0, + 785.0, + 1079.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 781.0, + 662.0, + 781.0, + 662.0, + 812.0, + 321.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 779.0, + 1408.0, + 779.0, + 1408.0, + 816.0, + 701.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 809.0, + 1406.0, + 809.0, + 1406.0, + 843.0, + 321.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 835.0, + 1406.0, + 835.0, + 1406.0, + 871.0, + 319.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 864.0, + 1291.0, + 864.0, + 1291.0, + 898.0, + 321.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 914.0, + 1407.0, + 914.0, + 1407.0, + 952.0, + 293.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 944.0, + 772.0, + 944.0, + 772.0, + 977.0, + 321.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 994.0, + 1407.0, + 994.0, + 1407.0, + 1029.0, + 295.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1024.0, + 513.0, + 1024.0, + 513.0, + 1055.0, + 323.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1072.0, + 1404.0, + 1072.0, + 1404.0, + 1108.0, + 294.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1102.0, + 1089.0, + 1102.0, + 1089.0, + 1134.0, + 322.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1657.0, + 1403.0, + 1657.0, + 1403.0, + 1690.0, + 295.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1685.0, + 1071.0, + 1685.0, + 1071.0, + 1716.0, + 323.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1151.0, + 1406.0, + 1151.0, + 1406.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1179.0, + 1405.0, + 1179.0, + 1405.0, + 1213.0, + 324.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1210.0, + 808.0, + 1210.0, + 808.0, + 1239.0, + 324.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1470.0, + 1408.0, + 1470.0, + 1408.0, + 1507.0, + 294.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1499.0, + 1151.0, + 1499.0, + 1151.0, + 1532.0, + 322.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1365.0, + 1405.0, + 1365.0, + 1405.0, + 1398.0, + 295.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1396.0, + 1411.0, + 1396.0, + 1411.0, + 1425.0, + 322.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1418.0, + 401.0, + 1418.0, + 401.0, + 1454.0, + 324.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1869.0, + 1405.0, + 1869.0, + 1405.0, + 1903.0, + 295.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1899.0, + 1230.0, + 1899.0, + 1230.0, + 1930.0, + 324.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1549.0, + 1407.0, + 1549.0, + 1407.0, + 1584.0, + 294.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1578.0, + 1406.0, + 1578.0, + 1406.0, + 1612.0, + 320.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1603.0, + 390.0, + 1603.0, + 390.0, + 1640.0, + 320.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1259.0, + 1404.0, + 1259.0, + 1404.0, + 1292.0, + 295.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1289.0, + 1404.0, + 1289.0, + 1404.0, + 1318.0, + 325.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1316.0, + 1101.0, + 1316.0, + 1101.0, + 1348.0, + 324.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1946.0, + 1406.0, + 1946.0, + 1406.0, + 1983.0, + 294.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1978.0, + 1404.0, + 1978.0, + 1404.0, + 2007.0, + 324.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2005.0, + 527.0, + 2005.0, + 527.0, + 2036.0, + 321.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1736.0, + 1407.0, + 1736.0, + 1407.0, + 1769.0, + 296.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1764.0, + 1405.0, + 1764.0, + 1405.0, + 1797.0, + 326.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1790.0, + 1408.0, + 1790.0, + 1408.0, + 1827.0, + 321.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1818.0, + 462.0, + 1818.0, + 462.0, + 1849.0, + 323.0, + 1849.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 548, + 1407, + 548, + 1407, + 633, + 296, + 633 + ], + "score": 0.896 + }, + { + "category_id": 1, + "poly": [ + 298, + 654, + 1406, + 654, + 1406, + 741, + 298, + 741 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 299, + 759, + 1407, + 759, + 1407, + 928, + 299, + 928 + ], + "score": 0.881 + }, + { + "category_id": 1, + "poly": [ + 296, + 1551, + 1402, + 1551, + 1402, + 1639, + 296, + 1639 + ], + "score": 0.875 + }, + { + "category_id": 1, + "poly": [ + 299, + 444, + 1401, + 444, + 1401, + 529, + 299, + 529 + ], + "score": 0.872 + }, + { + "category_id": 1, + "poly": [ + 296, + 1262, + 1402, + 1262, + 1402, + 1322, + 296, + 1322 + ], + "score": 0.867 + }, + { + "category_id": 1, + "poly": [ + 295, + 1976, + 1406, + 1976, + 1406, + 2034, + 295, + 2034 + ], + "score": 0.867 + }, + { + "category_id": 1, + "poly": [ + 293, + 1106, + 1401, + 1106, + 1401, + 1165, + 293, + 1165 + ], + "score": 0.862 + }, + { + "category_id": 1, + "poly": [ + 298, + 1658, + 1406, + 1658, + 1406, + 1772, + 298, + 1772 + ], + "score": 0.861 + }, + { + "category_id": 1, + "poly": [ + 298, + 1340, + 1406, + 1340, + 1406, + 1454, + 298, + 1454 + ], + "score": 0.858 + }, + { + "category_id": 1, + "poly": [ + 290, + 1474, + 1402, + 1474, + 1402, + 1534, + 290, + 1534 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 300, + 949, + 1401, + 949, + 1401, + 1036, + 300, + 1036 + ], + "score": 0.854 + }, + { + "category_id": 1, + "poly": [ + 298, + 1184, + 1401, + 1184, + 1401, + 1243, + 298, + 1243 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 300, + 231, + 1401, + 231, + 1401, + 316, + 300, + 316 + ], + "score": 0.848 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 861, + 2088, + 861, + 2113, + 835, + 2113 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 302, + 338, + 1398, + 338, + 1398, + 423, + 302, + 423 + ], + "score": 0.815 + }, + { + "category_id": 1, + "poly": [ + 291, + 1792, + 1402, + 1792, + 1402, + 1850, + 291, + 1850 + ], + "score": 0.814 + }, + { + "category_id": 1, + "poly": [ + 293, + 1054, + 1347, + 1054, + 1347, + 1088, + 293, + 1088 + ], + "score": 0.811 + }, + { + "category_id": 1, + "poly": [ + 298, + 1870, + 1406, + 1870, + 1406, + 1956, + 298, + 1956 + ], + "score": 0.796 + }, + { + "category_id": 13, + "poly": [ + 1082, + 503, + 1099, + 503, + 1099, + 524, + 1082, + 524 + ], + "score": 0.56, + "latex": "\\equiv" + }, + { + "category_id": 13, + "poly": [ + 1082, + 398, + 1098, + 398, + 1098, + 418, + 1082, + 418 + ], + "score": 0.54, + "latex": "\\equiv" + }, + { + "category_id": 13, + "poly": [ + 506, + 1746, + 522, + 1746, + 522, + 1767, + 506, + 1767 + ], + "score": 0.52, + "latex": "\\equiv" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 548.0, + 1404.0, + 548.0, + 1404.0, + 581.0, + 293.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 575.0, + 1405.0, + 575.0, + 1405.0, + 610.0, + 322.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 603.0, + 391.0, + 603.0, + 391.0, + 636.0, + 320.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 653.0, + 1406.0, + 653.0, + 1406.0, + 689.0, + 296.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 683.0, + 1408.0, + 683.0, + 1408.0, + 716.0, + 324.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 713.0, + 1408.0, + 713.0, + 1408.0, + 742.0, + 325.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 757.0, + 1408.0, + 757.0, + 1408.0, + 797.0, + 294.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 787.0, + 1407.0, + 787.0, + 1407.0, + 823.0, + 319.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 813.0, + 1407.0, + 813.0, + 1407.0, + 854.0, + 320.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 843.0, + 1409.0, + 843.0, + 1409.0, + 880.0, + 319.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 871.0, + 1410.0, + 871.0, + 1410.0, + 904.0, + 321.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 901.0, + 444.0, + 901.0, + 444.0, + 929.0, + 323.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1553.0, + 1406.0, + 1553.0, + 1406.0, + 1586.0, + 295.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1583.0, + 1404.0, + 1583.0, + 1404.0, + 1613.0, + 324.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1609.0, + 736.0, + 1609.0, + 736.0, + 1642.0, + 323.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 445.0, + 1405.0, + 445.0, + 1405.0, + 474.0, + 298.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 469.0, + 1405.0, + 469.0, + 1405.0, + 506.0, + 320.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 498.0, + 1081.0, + 498.0, + 1081.0, + 532.0, + 322.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 498.0, + 1243.0, + 498.0, + 1243.0, + 532.0, + 1100.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1263.0, + 1403.0, + 1263.0, + 1403.0, + 1294.0, + 295.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1292.0, + 1294.0, + 1292.0, + 1294.0, + 1323.0, + 324.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1978.0, + 1403.0, + 1978.0, + 1403.0, + 2009.0, + 297.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2004.0, + 1379.0, + 2004.0, + 1379.0, + 2038.0, + 322.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1108.0, + 1403.0, + 1108.0, + 1403.0, + 1139.0, + 297.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1133.0, + 1207.0, + 1133.0, + 1207.0, + 1168.0, + 322.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1695.0, + 296.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1683.0, + 1408.0, + 1683.0, + 1408.0, + 1723.0, + 320.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1711.0, + 1411.0, + 1711.0, + 1411.0, + 1750.0, + 318.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1743.0, + 505.0, + 1743.0, + 505.0, + 1777.0, + 322.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1743.0, + 667.0, + 1743.0, + 667.0, + 1777.0, + 523.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1338.0, + 1409.0, + 1338.0, + 1409.0, + 1377.0, + 291.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1365.0, + 1408.0, + 1365.0, + 1408.0, + 1406.0, + 322.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1396.0, + 1408.0, + 1396.0, + 1408.0, + 1429.0, + 324.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1421.0, + 1260.0, + 1421.0, + 1260.0, + 1457.0, + 321.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1475.0, + 1408.0, + 1475.0, + 1408.0, + 1506.0, + 296.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1503.0, + 1164.0, + 1503.0, + 1164.0, + 1534.0, + 323.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 947.0, + 1406.0, + 947.0, + 1406.0, + 986.0, + 295.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 978.0, + 1407.0, + 978.0, + 1407.0, + 1011.0, + 321.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1005.0, + 609.0, + 1005.0, + 609.0, + 1037.0, + 321.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1182.0, + 1405.0, + 1182.0, + 1405.0, + 1218.0, + 294.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1212.0, + 581.0, + 1212.0, + 581.0, + 1242.0, + 322.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 260.0, + 1408.0, + 260.0, + 1408.0, + 292.0, + 323.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 287.0, + 688.0, + 287.0, + 688.0, + 316.0, + 324.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 335.0, + 1406.0, + 335.0, + 1406.0, + 372.0, + 296.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 364.0, + 1406.0, + 364.0, + 1406.0, + 401.0, + 319.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 394.0, + 1081.0, + 394.0, + 1081.0, + 427.0, + 323.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 394.0, + 1259.0, + 394.0, + 1259.0, + 427.0, + 1099.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1830.0, + 293.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1819.0, + 1214.0, + 1819.0, + 1214.0, + 1852.0, + 320.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1054.0, + 1345.0, + 1054.0, + 1345.0, + 1089.0, + 297.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1870.0, + 1404.0, + 1870.0, + 1404.0, + 1904.0, + 295.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1898.0, + 1408.0, + 1898.0, + 1408.0, + 1933.0, + 321.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1925.0, + 1071.0, + 1925.0, + 1071.0, + 1958.0, + 322.0, + 1958.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.828 + }, + { + "category_id": 1, + "poly": [ + 296, + 1650, + 1402, + 1650, + 1402, + 1709, + 296, + 1709 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 297, + 1568, + 1402, + 1568, + 1402, + 1628, + 297, + 1628 + ], + "score": 0.711 + }, + { + "category_id": 1, + "poly": [ + 296, + 1486, + 1401, + 1486, + 1401, + 1546, + 296, + 1546 + ], + "score": 0.704 + }, + { + "category_id": 1, + "poly": [ + 299, + 859, + 1402, + 859, + 1402, + 946, + 299, + 946 + ], + "score": 0.704 + }, + { + "category_id": 1, + "poly": [ + 299, + 968, + 1402, + 968, + 1402, + 1054, + 299, + 1054 + ], + "score": 0.671 + }, + { + "category_id": 1, + "poly": [ + 296, + 1731, + 1403, + 1731, + 1403, + 1791, + 296, + 1791 + ], + "score": 0.671 + }, + { + "category_id": 1, + "poly": [ + 296, + 230, + 1403, + 230, + 1403, + 317, + 296, + 317 + ], + "score": 0.671 + }, + { + "category_id": 1, + "poly": [ + 298, + 749, + 1401, + 749, + 1401, + 837, + 298, + 837 + ], + "score": 0.634 + }, + { + "category_id": 1, + "poly": [ + 295, + 667, + 1403, + 667, + 1403, + 728, + 295, + 728 + ], + "score": 0.627 + }, + { + "category_id": 1, + "poly": [ + 295, + 1242, + 1402, + 1242, + 1402, + 1300, + 295, + 1300 + ], + "score": 0.625 + }, + { + "category_id": 1, + "poly": [ + 298, + 1405, + 1401, + 1405, + 1401, + 1464, + 298, + 1464 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 296, + 1813, + 1402, + 1813, + 1402, + 1872, + 296, + 1872 + ], + "score": 0.6 + }, + { + "category_id": 1, + "poly": [ + 302, + 450, + 1400, + 450, + 1400, + 537, + 302, + 537 + ], + "score": 0.592 + }, + { + "category_id": 1, + "poly": [ + 299, + 1323, + 1402, + 1323, + 1402, + 1383, + 299, + 1383 + ], + "score": 0.587 + }, + { + "category_id": 1, + "poly": [ + 293, + 1976, + 1403, + 1976, + 1403, + 2035, + 293, + 2035 + ], + "score": 0.583 + }, + { + "category_id": 1, + "poly": [ + 299, + 558, + 1400, + 558, + 1400, + 646, + 299, + 646 + ], + "score": 0.562 + }, + { + "category_id": 1, + "poly": [ + 297, + 1894, + 1400, + 1894, + 1400, + 1954, + 297, + 1954 + ], + "score": 0.554 + }, + { + "category_id": 1, + "poly": [ + 296, + 1076, + 1406, + 1076, + 1406, + 1218, + 296, + 1218 + ], + "score": 0.548 + }, + { + "category_id": 1, + "poly": [ + 300, + 339, + 1398, + 339, + 1398, + 428, + 300, + 428 + ], + "score": 0.544 + }, + { + "category_id": 13, + "poly": [ + 1180, + 811, + 1196, + 811, + 1196, + 830, + 1180, + 830 + ], + "score": 0.38, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1651.0, + 1404.0, + 1651.0, + 1404.0, + 1682.0, + 296.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1678.0, + 966.0, + 1678.0, + 966.0, + 1709.0, + 323.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1569.0, + 1403.0, + 1569.0, + 1403.0, + 1600.0, + 296.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1597.0, + 965.0, + 1597.0, + 965.0, + 1628.0, + 322.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1483.0, + 1406.0, + 1483.0, + 1406.0, + 1521.0, + 292.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1514.0, + 958.0, + 1514.0, + 958.0, + 1547.0, + 320.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 859.0, + 1406.0, + 859.0, + 1406.0, + 892.0, + 295.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 886.0, + 1406.0, + 886.0, + 1406.0, + 920.0, + 322.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 915.0, + 693.0, + 915.0, + 693.0, + 948.0, + 322.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 970.0, + 1402.0, + 970.0, + 1402.0, + 999.0, + 298.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 998.0, + 1408.0, + 998.0, + 1408.0, + 1031.0, + 322.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1023.0, + 908.0, + 1023.0, + 908.0, + 1056.0, + 322.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1728.0, + 1405.0, + 1728.0, + 1405.0, + 1767.0, + 294.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1761.0, + 1376.0, + 1761.0, + 1376.0, + 1792.0, + 323.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 228.0, + 1407.0, + 228.0, + 1407.0, + 267.0, + 291.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 258.0, + 1408.0, + 258.0, + 1408.0, + 292.0, + 320.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 285.0, + 391.0, + 285.0, + 391.0, + 318.0, + 320.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 746.0, + 1406.0, + 746.0, + 1406.0, + 787.0, + 294.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 778.0, + 1407.0, + 778.0, + 1407.0, + 812.0, + 321.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 805.0, + 1179.0, + 805.0, + 1179.0, + 839.0, + 322.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 805.0, + 1339.0, + 805.0, + 1339.0, + 839.0, + 1197.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 665.0, + 1406.0, + 665.0, + 1406.0, + 704.0, + 294.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 694.0, + 392.0, + 694.0, + 392.0, + 729.0, + 319.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1244.0, + 1403.0, + 1244.0, + 1403.0, + 1275.0, + 297.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1270.0, + 1104.0, + 1270.0, + 1104.0, + 1302.0, + 322.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1405.0, + 1408.0, + 1405.0, + 1408.0, + 1439.0, + 296.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1433.0, + 563.0, + 1433.0, + 563.0, + 1464.0, + 322.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1813.0, + 1404.0, + 1813.0, + 1404.0, + 1847.0, + 296.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1841.0, + 839.0, + 1841.0, + 839.0, + 1872.0, + 321.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 450.0, + 1405.0, + 450.0, + 1405.0, + 483.0, + 296.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 478.0, + 1405.0, + 478.0, + 1405.0, + 511.0, + 323.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 505.0, + 1021.0, + 505.0, + 1021.0, + 538.0, + 321.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1324.0, + 1403.0, + 1324.0, + 1403.0, + 1355.0, + 295.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1353.0, + 1368.0, + 1353.0, + 1368.0, + 1384.0, + 323.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1978.0, + 1403.0, + 1978.0, + 1403.0, + 2009.0, + 296.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 2003.0, + 1215.0, + 2003.0, + 1215.0, + 2037.0, + 320.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 559.0, + 1405.0, + 559.0, + 1405.0, + 593.0, + 293.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 584.0, + 1405.0, + 584.0, + 1405.0, + 622.0, + 322.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 616.0, + 1254.0, + 616.0, + 1254.0, + 647.0, + 323.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1894.0, + 1402.0, + 1894.0, + 1402.0, + 1925.0, + 297.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1923.0, + 921.0, + 1923.0, + 921.0, + 1955.0, + 322.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1078.0, + 1404.0, + 1078.0, + 1404.0, + 1109.0, + 296.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1103.0, + 1406.0, + 1103.0, + 1406.0, + 1141.0, + 322.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1133.0, + 1407.0, + 1133.0, + 1407.0, + 1167.0, + 320.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1158.0, + 1412.0, + 1158.0, + 1412.0, + 1195.0, + 319.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1188.0, + 401.0, + 1188.0, + 401.0, + 1222.0, + 321.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 339.0, + 1402.0, + 339.0, + 1402.0, + 373.0, + 295.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 371.0, + 1404.0, + 371.0, + 1404.0, + 400.0, + 325.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 399.0, + 929.0, + 399.0, + 929.0, + 428.0, + 322.0, + 428.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 230, + 1405, + 230, + 1405, + 344, + 298, + 344 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 298, + 364, + 1402, + 364, + 1402, + 422, + 298, + 422 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.834 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.83 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 867.0, + 2084.0, + 867.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 259.0, + 1407.0, + 259.0, + 1407.0, + 292.0, + 325.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 284.0, + 1406.0, + 284.0, + 1406.0, + 320.0, + 321.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 316.0, + 594.0, + 316.0, + 594.0, + 348.0, + 320.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 363.0, + 1407.0, + 363.0, + 1407.0, + 398.0, + 297.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 394.0, + 1333.0, + 394.0, + 1333.0, + 425.0, + 325.0, + 425.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1538, + 1404, + 1538, + 1404, + 1784, + 298, + 1784 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 293, + 294, + 1367, + 294, + 1367, + 326, + 293, + 326 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 308, + 1165, + 1404, + 1165, + 1404, + 1379, + 308, + 1379 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 301, + 1422, + 647, + 1422, + 647, + 1457, + 301, + 1457 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 293, + 1489, + 1317, + 1489, + 1317, + 1522, + 293, + 1522 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 302, + 1801, + 1401, + 1801, + 1401, + 1894, + 302, + 1894 + ], + "score": 0.867 + }, + { + "category_id": 0, + "poly": [ + 303, + 226, + 709, + 226, + 709, + 262, + 303, + 262 + ], + "score": 0.846 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 863, + 2089, + 863, + 2112, + 836, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 1, + "poly": [ + 299, + 978, + 1406, + 978, + 1406, + 1162, + 299, + 1162 + ], + "score": 0.84 + }, + { + "category_id": 1, + "poly": [ + 315, + 756, + 1403, + 756, + 1403, + 969, + 315, + 969 + ], + "score": 0.617 + }, + { + "category_id": 1, + "poly": [ + 313, + 660, + 1403, + 660, + 1403, + 751, + 313, + 751 + ], + "score": 0.429 + }, + { + "category_id": 1, + "poly": [ + 299, + 351, + 1406, + 351, + 1406, + 658, + 299, + 658 + ], + "score": 0.253 + }, + { + "category_id": 13, + "poly": [ + 704, + 1541, + 741, + 1541, + 741, + 1570, + 704, + 1570 + ], + "score": 0.91, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 890, + 2004, + 927, + 2004, + 927, + 2034, + 890, + 2034 + ], + "score": 0.9, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 1207, + 1863, + 1245, + 1863, + 1245, + 1893, + 1207, + 1893 + ], + "score": 0.9, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 516, + 1631, + 636, + 1631, + 636, + 1659, + 516, + 1659 + ], + "score": 0.89, + "latex": "\\alpha = 0 . 0 0 1" + }, + { + "category_id": 13, + "poly": [ + 1182, + 1571, + 1304, + 1571, + 1304, + 1598, + 1182, + 1598 + ], + "score": 0.89, + "latex": "\\alpha = 0 . 0 0 2" + }, + { + "category_id": 13, + "poly": [ + 298, + 1540, + 334, + 1540, + 334, + 1571, + 298, + 1571 + ], + "score": 0.88, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 977, + 1131, + 1078, + 1131, + 1078, + 1162, + 977, + 1162 + ], + "score": 0.88, + "latex": "6 . 5 – 1 4 \\times" + }, + { + "category_id": 13, + "poly": [ + 733, + 1913, + 759, + 1913, + 759, + 1942, + 733, + 1942 + ], + "score": 0.88, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 846, + 1692, + 967, + 1692, + 967, + 1721, + 846, + 1721 + ], + "score": 0.87, + "latex": "\\alpha = 0 . 0 0 5" + }, + { + "category_id": 13, + "poly": [ + 1026, + 1833, + 1054, + 1833, + 1054, + 1863, + 1026, + 1863 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 916, + 877, + 972, + 877, + 972, + 906, + 916, + 906 + ], + "score": 0.87, + "latex": "90 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1802, + 324, + 1802, + 324, + 1832, + 297, + 1832 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1186, + 1318, + 1273, + 1318, + 1273, + 1348, + 1186, + 1348 + ], + "score": 0.86, + "latex": "1 . 5 – 5 \\times" + }, + { + "category_id": 13, + "poly": [ + 297, + 1912, + 323, + 1912, + 323, + 1942, + 297, + 1942 + ], + "score": 0.85, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 567, + 1837, + 588, + 1837, + 588, + 1864, + 567, + 1864 + ], + "score": 0.8, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 494, + 1947, + 516, + 1947, + 516, + 1971, + 494, + 1971 + ], + "score": 0.75, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1271, + 1631, + 1316, + 1631, + 1316, + 1660, + 1271, + 1660 + ], + "score": 0.28, + "latex": "1 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1421.0, + 650.0, + 1421.0, + 650.0, + 1461.0, + 294.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 714.0, + 224.0, + 714.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1537.0, + 703.0, + 1537.0, + 703.0, + 1574.0, + 335.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1574.0, + 742.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1570.0, + 1181.0, + 1570.0, + 1181.0, + 1603.0, + 294.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1570.0, + 1405.0, + 1570.0, + 1405.0, + 1603.0, + 1305.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1601.0, + 1406.0, + 1601.0, + 1406.0, + 1635.0, + 294.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1631.0, + 515.0, + 1631.0, + 515.0, + 1665.0, + 294.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1631.0, + 1270.0, + 1631.0, + 1270.0, + 1665.0, + 637.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 1631.0, + 1406.0, + 1631.0, + 1406.0, + 1665.0, + 1317.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1662.0, + 1406.0, + 1662.0, + 1406.0, + 1696.0, + 294.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1692.0, + 845.0, + 1692.0, + 845.0, + 1726.0, + 294.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1692.0, + 1405.0, + 1692.0, + 1405.0, + 1726.0, + 968.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1721.0, + 1408.0, + 1721.0, + 1408.0, + 1760.0, + 292.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1754.0, + 492.0, + 1754.0, + 492.0, + 1788.0, + 294.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1911.0, + 732.0, + 1911.0, + 732.0, + 1943.0, + 324.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1943.0, + 760.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 493.0, + 1941.0, + 493.0, + 1976.0, + 293.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1941.0, + 1404.0, + 1941.0, + 1404.0, + 1976.0, + 517.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2006.0, + 293.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1999.0, + 889.0, + 1999.0, + 889.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1999.0, + 1382.0, + 1999.0, + 1382.0, + 2039.0, + 928.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 292.0, + 1372.0, + 292.0, + 1372.0, + 331.0, + 296.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1165.0, + 1404.0, + 1165.0, + 1404.0, + 1200.0, + 319.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1196.0, + 1404.0, + 1196.0, + 1404.0, + 1230.0, + 318.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1225.0, + 1405.0, + 1225.0, + 1405.0, + 1261.0, + 318.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1256.0, + 1408.0, + 1256.0, + 1408.0, + 1292.0, + 317.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1287.0, + 1404.0, + 1287.0, + 1404.0, + 1321.0, + 318.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1317.0, + 1185.0, + 1317.0, + 1185.0, + 1352.0, + 318.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1352.0, + 1274.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1347.0, + 1277.0, + 1347.0, + 1277.0, + 1384.0, + 318.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1486.0, + 1325.0, + 1486.0, + 1325.0, + 1529.0, + 291.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1799.0, + 1406.0, + 1799.0, + 1406.0, + 1835.0, + 325.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1833.0, + 566.0, + 1833.0, + 566.0, + 1867.0, + 296.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1833.0, + 1025.0, + 1833.0, + 1025.0, + 1867.0, + 589.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1833.0, + 1404.0, + 1833.0, + 1404.0, + 1867.0, + 1055.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1863.0, + 1206.0, + 1863.0, + 1206.0, + 1896.0, + 296.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 1863.0, + 1406.0, + 1863.0, + 1406.0, + 1896.0, + 1246.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 977.0, + 1409.0, + 977.0, + 1409.0, + 1013.0, + 297.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1008.0, + 1406.0, + 1008.0, + 1406.0, + 1043.0, + 317.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1040.0, + 1408.0, + 1040.0, + 1408.0, + 1075.0, + 318.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1067.0, + 1407.0, + 1067.0, + 1407.0, + 1107.0, + 318.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1100.0, + 1406.0, + 1100.0, + 1406.0, + 1136.0, + 318.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1130.0, + 976.0, + 1130.0, + 976.0, + 1166.0, + 318.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1130.0, + 1232.0, + 1130.0, + 1232.0, + 1166.0, + 1079.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 754.0, + 1405.0, + 754.0, + 1405.0, + 792.0, + 317.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 785.0, + 1405.0, + 785.0, + 1405.0, + 820.0, + 317.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 814.0, + 1406.0, + 814.0, + 1406.0, + 852.0, + 316.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 847.0, + 1406.0, + 847.0, + 1406.0, + 881.0, + 317.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 875.0, + 915.0, + 875.0, + 915.0, + 913.0, + 314.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 913.0, + 973.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 908.0, + 1403.0, + 908.0, + 1403.0, + 942.0, + 317.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 938.0, + 1378.0, + 938.0, + 1378.0, + 973.0, + 318.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 655.0, + 1407.0, + 655.0, + 1407.0, + 698.0, + 316.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 689.0, + 1407.0, + 689.0, + 1407.0, + 725.0, + 318.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 720.0, + 1369.0, + 720.0, + 1369.0, + 757.0, + 318.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 348.0, + 1408.0, + 348.0, + 1408.0, + 385.0, + 296.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 381.0, + 1409.0, + 381.0, + 1409.0, + 414.0, + 316.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 412.0, + 1406.0, + 412.0, + 1406.0, + 447.0, + 318.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 443.0, + 1407.0, + 443.0, + 1407.0, + 475.0, + 318.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 470.0, + 1406.0, + 470.0, + 1406.0, + 508.0, + 317.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 502.0, + 1404.0, + 502.0, + 1404.0, + 538.0, + 319.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 532.0, + 1406.0, + 532.0, + 1406.0, + 568.0, + 317.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 566.0, + 1404.0, + 566.0, + 1404.0, + 597.0, + 318.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 596.0, + 1404.0, + 596.0, + 1404.0, + 628.0, + 319.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 623.0, + 1398.0, + 623.0, + 1398.0, + 661.0, + 316.0, + 661.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 296, + 515, + 1400, + 515, + 1400, + 1030, + 296, + 1030 + ], + "score": 0.983, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 /89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 /50%0.39 / 57%0.51 /52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.36 / 51%0.35 / 51%0.12 /36%0.34 /45%0.18 / 41%0.49 / 53%
TI0.23 /37%0.51 / 40%0.08 /22%0.15 / 25%0.08 /33%0.19 /35%
UNIVERSAL DEEPSLOTH
C100.89 /83%0.89 /83%0.75 /85%0.88 / 75%0.82 /85%0.92 /77%
TI0.64 / 55%0.83/50%0.38 /57%0.51 / 52%0.41 / 57%0.59 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88/ 73%0.88 / 73%0.69 / 78%0.86 /67%0.76 /74%0.89 / 65%
TI0.64 /54%0.83 /49%0.39 / 59%0.50 / 58%0.41 / 60%0.58 / 53%
" + }, + { + "category_id": 5, + "poly": [ + 296, + 1258, + 1400, + 1258, + 1400, + 1771, + 296, + 1771 + ], + "score": 0.982, + "html": "
NETWORKMSDNETVGG16MOBILENET
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.89 /85%0.89 /85%0.77 / 89%0.89 / 79%0.83 /87%0.92 /79%
TI0.64 / 55%0.83 / 50%0.39 / 57%0.51 / 52%0.42 /57%0.59 / 51%
DEEPSLOTH
C100.52 / 64%0.52 /64%0.22 /60%0.45 / 62%0.23 /46%0.48 / 55%
TI0.24 /42%0.52 / 44%0.13 /35%0.21 /36%0.12 /38%0.25 /40%
UNIVERSAL DEEPSLOTH
C100.89 /81%0.89 /81%0.75 /87%0.88 / 76%0.81/84%0.92 /76%
TI0.63 / 54%0.82 /48%0.38 /56%0.51 / 52%0.41 / 56%0.58 / 51%
CLASS-UNIVERSAL DEEPSLOTH
C100.88 /73%0.88 / 73%0.71 /81%0.86 / 70%0.76 /76%0.89 /66%
TI0.64 / 53%0.83 /49%0.38 /57%0.50 / 57%0.41 /58%0.58 / 53%
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1849, + 1406, + 1849, + 1406, + 2035, + 297, + 2035 + ], + "score": 0.973 + }, + { + "category_id": 6, + "poly": [ + 297, + 1139, + 1404, + 1139, + 1404, + 1235, + 297, + 1235 + ], + "score": 0.948 + }, + { + "category_id": 6, + "poly": [ + 296, + 395, + 1405, + 395, + 1405, + 491, + 296, + 491 + ], + "score": 0.934 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 297, + 309, + 1404, + 309, + 1404, + 344, + 297, + 344 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2113, + 835, + 2113 + ], + "score": 0.856 + }, + { + "category_id": 0, + "poly": [ + 299, + 225, + 1083, + 225, + 1083, + 263, + 299, + 263 + ], + "score": 0.599 + }, + { + "category_id": 1, + "poly": [ + 299, + 225, + 1083, + 225, + 1083, + 263, + 299, + 263 + ], + "score": 0.343 + }, + { + "category_id": 6, + "poly": [ + 300, + 395, + 1403, + 395, + 1403, + 492, + 300, + 492 + ], + "score": 0.18 + }, + { + "category_id": 13, + "poly": [ + 1162, + 1943, + 1199, + 1943, + 1199, + 1973, + 1162, + 1973 + ], + "score": 0.89, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 835, + 1943, + 926, + 1943, + 926, + 1972, + 835, + 1972 + ], + "score": 0.89, + "latex": "6 \\sim 2 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 471, + 1882, + 509, + 1882, + 509, + 1912, + 471, + 1912 + ], + "score": 0.89, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 1161, + 1852, + 1199, + 1852, + 1199, + 1881, + 1161, + 1881 + ], + "score": 0.89, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 763, + 228, + 794, + 228, + 794, + 262, + 763, + 262 + ], + "score": 0.88, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 807, + 313, + 833, + 313, + 833, + 342, + 807, + 342 + ], + "score": 0.88, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1018, + 1912, + 1141, + 1912, + 1141, + 1941, + 1018, + 1941 + ], + "score": 0.88, + "latex": "0 . 2 4 { \\sim } 0 . 6 5" + }, + { + "category_id": 13, + "poly": [ + 1023, + 1973, + 1050, + 1973, + 1050, + 2003, + 1023, + 2003 + ], + "score": 0.88, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 574, + 1943, + 601, + 1943, + 601, + 1973, + 574, + 1973 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 677, + 1851, + 703, + 1851, + 703, + 1881, + 677, + 1881 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 955, + 313, + 981, + 313, + 981, + 342, + 955, + 342 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1322, + 2003, + 1375, + 2003, + 1375, + 2032, + 1322, + 2032 + ], + "score": 0.87, + "latex": "12 \\%" + }, + { + "category_id": 13, + "poly": [ + 869, + 227, + 900, + 227, + 900, + 263, + 869, + 263 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 345, + 1912, + 372, + 1912, + 372, + 1942, + 345, + 1942 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1973, + 412, + 1973, + 412, + 2003, + 298, + 2003 + ], + "score": 0.86, + "latex": "( 7 5 \\sim 9 9 \\% )" + }, + { + "category_id": 13, + "poly": [ + 623, + 398, + 649, + 398, + 649, + 428, + 623, + 428 + ], + "score": 0.86, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 623, + 1141, + 650, + 1141, + 650, + 1171, + 623, + 1171 + ], + "score": 0.85, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 587, + 1851, + 616, + 1851, + 616, + 1881, + 587, + 1881 + ], + "score": 0.84, + "latex": "\\ell _ { 1 } \\cdot" + }, + { + "category_id": 13, + "poly": [ + 486, + 1943, + 515, + 1943, + 515, + 1973, + 486, + 1973 + ], + "score": 0.84, + "latex": "\\ell _ { 1 } \\cdot" + }, + { + "category_id": 13, + "poly": [ + 796, + 1140, + 951, + 1140, + 951, + 1170, + 796, + 1170 + ], + "score": 0.83, + "latex": "\\cdot \\mathrm { R A D } { < } 5 , 1 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 938, + 1974, + 965, + 1974, + 965, + 2003, + 938, + 2003 + ], + "score": 0.81, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1367, + 1882, + 1396, + 1882, + 1396, + 1912, + 1367, + 1912 + ], + "score": 0.8, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 800, + 398, + 951, + 398, + 951, + 427, + 800, + 427 + ], + "score": 0.77, + "latex": "R \\mathrm { A D } { < } 5 , 1 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1085, + 460, + 1149, + 460, + 1149, + 487, + 1085, + 487 + ], + "score": 0.7, + "latex": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }" + }, + { + "category_id": 13, + "poly": [ + 1084, + 1203, + 1149, + 1203, + 1149, + 1231, + 1084, + 1231 + ], + "score": 0.65, + "latex": "\\mathbf { \\dot { C } } 1 0 ^ { \\circ }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1139.0, + 622.0, + 1139.0, + 622.0, + 1174.0, + 295.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1139.0, + 795.0, + 1139.0, + 795.0, + 1174.0, + 651.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1139.0, + 1404.0, + 1139.0, + 1404.0, + 1174.0, + 952.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1171.0, + 1405.0, + 1171.0, + 1405.0, + 1205.0, + 295.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1200.0, + 1083.0, + 1200.0, + 1083.0, + 1238.0, + 294.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1200.0, + 1308.0, + 1200.0, + 1308.0, + 1238.0, + 1150.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 396.0, + 622.0, + 396.0, + 622.0, + 431.0, + 294.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 396.0, + 799.0, + 396.0, + 799.0, + 431.0, + 650.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 396.0, + 1405.0, + 396.0, + 1405.0, + 431.0, + 952.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 430.0, + 1403.0, + 430.0, + 1403.0, + 460.0, + 295.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 456.0, + 1084.0, + 456.0, + 1084.0, + 493.0, + 294.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 456.0, + 1309.0, + 456.0, + 1309.0, + 493.0, + 1150.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 221.0, + 762.0, + 221.0, + 762.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 221.0, + 868.0, + 221.0, + 868.0, + 268.0, + 795.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 221.0, + 1088.0, + 221.0, + 1088.0, + 268.0, + 901.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 395.0, + 622.0, + 395.0, + 622.0, + 431.0, + 294.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 395.0, + 799.0, + 395.0, + 799.0, + 431.0, + 650.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 395.0, + 1405.0, + 395.0, + 1405.0, + 431.0, + 952.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 427.0, + 1405.0, + 427.0, + 1405.0, + 462.0, + 295.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 458.0, + 1084.0, + 458.0, + 1084.0, + 493.0, + 296.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 458.0, + 1308.0, + 458.0, + 1308.0, + 493.0, + 1150.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1849.0, + 586.0, + 1849.0, + 586.0, + 1885.0, + 296.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1849.0, + 676.0, + 1849.0, + 676.0, + 1885.0, + 617.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1849.0, + 1160.0, + 1849.0, + 1160.0, + 1885.0, + 704.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1885.0, + 1200.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 470.0, + 1881.0, + 470.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1881.0, + 1366.0, + 1881.0, + 1366.0, + 1916.0, + 510.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1881.0, + 1407.0, + 1881.0, + 1407.0, + 1916.0, + 1397.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1905.0, + 344.0, + 1905.0, + 344.0, + 1950.0, + 292.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1905.0, + 1017.0, + 1905.0, + 1017.0, + 1950.0, + 373.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1905.0, + 1407.0, + 1905.0, + 1407.0, + 1950.0, + 1142.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 485.0, + 1942.0, + 485.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1942.0, + 573.0, + 1942.0, + 573.0, + 1978.0, + 516.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1942.0, + 834.0, + 1942.0, + 834.0, + 1978.0, + 602.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1942.0, + 1161.0, + 1942.0, + 1161.0, + 1978.0, + 927.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1978.0, + 1200.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 297.0, + 1970.0, + 297.0, + 2007.0, + 292.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1970.0, + 937.0, + 1970.0, + 937.0, + 2007.0, + 413.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1970.0, + 1022.0, + 1970.0, + 1022.0, + 2007.0, + 966.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1970.0, + 1404.0, + 1970.0, + 1404.0, + 2007.0, + 1051.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1321.0, + 2002.0, + 1321.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 2002.0, + 1389.0, + 2002.0, + 1389.0, + 2038.0, + 1376.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 304.0, + 806.0, + 304.0, + 806.0, + 349.0, + 292.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 304.0, + 954.0, + 304.0, + 954.0, + 349.0, + 834.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 304.0, + 1408.0, + 304.0, + 1408.0, + 349.0, + 982.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 221.0, + 762.0, + 221.0, + 762.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 221.0, + 868.0, + 221.0, + 868.0, + 268.0, + 795.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 221.0, + 1088.0, + 221.0, + 1088.0, + 268.0, + 901.0, + 268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 295, + 524, + 1400, + 524, + 1400, + 1039, + 295, + 1039 + ], + "score": 0.983, + "html": "
NETWORKRESNET(lo)RESNET (l1)RESNET(2)
SET.RAD<5%RAD<15%RAD<5%RAD<15%RAD<5%RAD<15%
BASELINE (NO ATTACK)
C100.52 /87%0.69 /80%0.52/87%0.69 / 80%0.51 /87%0.69 / 80%
TI0.25 / 51%0.39 /46%0.25 /51%0.39 /46%0.25 /51%0.39 / 46%
DEEPSLOTH
C100.00 /19%0.01 /19%0.05 /43%0.18 /47%0.06 /45%0.17 /48%
TI0.00 /7%0.01 / 7%0.04 /27%0.10 /28%0.05 /34%0.13 /35%
UNIVERSAL DEEPSLOTH
C100.35 / 63%0.59 /60%0.49 /84%0.68 /75%0.48 /85%0.67 /76%
TI0.25 / 25%0.34 / 37%0.25 / 51%0.39 / 46%0.25 / 51%0.38 /46%
CLASS-UNIVERSAL DEEPSLOTH
C100.23 / 33%0.48 / 29%0.39 / 70%0.60 / 61%0.39 / 71%0.60 / 61%
TI0.11 /21%0.23 /18%0.23 / 51%0.36 /46%0.23 / 50%0.36 /46%
" + }, + { + "category_id": 5, + "poly": [ + 930, + 1505, + 1400, + 1505, + 1400, + 1744, + 930, + 1744 + ], + "score": 0.978, + "html": "
ATTACKSTIME (SEC.)
PGD-2038
PGD-20 (AVG.)48
PGD-20 (MAX.)475
DEEPSLOTH44
UNIVERSAL DEEPSLOTH2
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1121, + 1406, + 1121, + 1406, + 1337, + 297, + 1337 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1502, + 902, + 1502, + 902, + 1817, + 299, + 1817 + ], + "score": 0.963 + }, + { + "category_id": 6, + "poly": [ + 296, + 406, + 1402, + 406, + 1402, + 500, + 296, + 500 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 302, + 315, + 1270, + 315, + 1270, + 347, + 302, + 347 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 292, + 2001, + 1387, + 2001, + 1387, + 2035, + 292, + 2035 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.839 + }, + { + "category_id": 0, + "poly": [ + 298, + 1412, + 965, + 1412, + 965, + 1449, + 298, + 1449 + ], + "score": 0.837 + }, + { + "category_id": 0, + "poly": [ + 296, + 1914, + 1377, + 1914, + 1377, + 1951, + 296, + 1951 + ], + "score": 0.799 + }, + { + "category_id": 0, + "poly": [ + 298, + 225, + 1157, + 225, + 1157, + 262, + 298, + 262 + ], + "score": 0.741 + }, + { + "category_id": 7, + "poly": [ + 952, + 1770, + 1377, + 1770, + 1377, + 1801, + 952, + 1801 + ], + "score": 0.53 + }, + { + "category_id": 7, + "poly": [ + 299, + 1808, + 1387, + 1808, + 1387, + 1838, + 299, + 1838 + ], + "score": 0.428 + }, + { + "category_id": 1, + "poly": [ + 299, + 1808, + 1387, + 1808, + 1387, + 1838, + 299, + 1838 + ], + "score": 0.241 + }, + { + "category_id": 1, + "poly": [ + 298, + 225, + 1157, + 225, + 1157, + 262, + 298, + 262 + ], + "score": 0.15 + }, + { + "category_id": 13, + "poly": [ + 679, + 1184, + 782, + 1184, + 782, + 1213, + 679, + 1213 + ], + "score": 0.9, + "latex": "3 9 \\sim 6 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 773, + 1214, + 875, + 1214, + 875, + 1243, + 773, + 1243 + ], + "score": 0.89, + "latex": "1 1 { \\sim } 4 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 326, + 1184, + 448, + 1184, + 448, + 1213, + 326, + 1213 + ], + "score": 0.89, + "latex": "0 . 0 0 { \\sim } 0 . 0 1" + }, + { + "category_id": 13, + "poly": [ + 511, + 1154, + 548, + 1154, + 548, + 1184, + 511, + 1184 + ], + "score": 0.89, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 1021, + 1184, + 1048, + 1184, + 1048, + 1214, + 1021, + 1214 + ], + "score": 0.89, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 1275, + 1195, + 1275, + 1195, + 1305, + 1168, + 1305 + ], + "score": 0.88, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1182, + 1123, + 1209, + 1123, + 1209, + 1153, + 1182, + 1153 + ], + "score": 0.88, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 818, + 1275, + 873, + 1275, + 873, + 1304, + 818, + 1304 + ], + "score": 0.87, + "latex": "24 \\%" + }, + { + "category_id": 13, + "poly": [ + 936, + 1184, + 964, + 1184, + 964, + 1214, + 936, + 1214 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 419, + 1214, + 541, + 1214, + 541, + 1243, + 419, + 1243 + ], + "score": 0.87, + "latex": "0 . 0 4 { \\sim } 0 . 1 8" + }, + { + "category_id": 13, + "poly": [ + 1081, + 1275, + 1109, + 1275, + 1109, + 1305, + 1081, + 1305 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1128, + 406, + 1290, + 406, + 1290, + 437, + 1128, + 437 + ], + "score": 0.83, + "latex": "\\cdot \\mathrm { R A D } { < } 5 , 1 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1053, + 1123, + 1090, + 1123, + 1090, + 1153, + 1053, + 1153 + ], + "score": 0.73, + "latex": "\\ell _ { \\infty }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1123, + 1128, + 1123, + 1128, + 1153, + 1100, + 1153 + ], + "score": 0.71, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 774, + 1808, + 818, + 1808, + 818, + 1835, + 774, + 1835 + ], + "score": 0.35, + "latex": "1 0 \\mathrm { x }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 406.0, + 1127.0, + 406.0, + 1127.0, + 440.0, + 294.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 406.0, + 1404.0, + 406.0, + 1404.0, + 440.0, + 1291.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 436.0, + 1404.0, + 436.0, + 1404.0, + 472.0, + 294.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 466.0, + 1205.0, + 466.0, + 1205.0, + 504.0, + 295.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1408.0, + 968.0, + 1408.0, + 968.0, + 1456.0, + 291.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1915.0, + 1380.0, + 1915.0, + 1380.0, + 1951.0, + 295.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 225.0, + 1160.0, + 225.0, + 1160.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1770.0, + 1375.0, + 1770.0, + 1375.0, + 1803.0, + 950.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1801.0, + 773.0, + 1801.0, + 773.0, + 1845.0, + 294.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1801.0, + 1391.0, + 1801.0, + 1391.0, + 1845.0, + 819.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1120.0, + 1052.0, + 1120.0, + 1052.0, + 1159.0, + 295.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1120.0, + 1099.0, + 1120.0, + 1099.0, + 1159.0, + 1091.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1120.0, + 1181.0, + 1120.0, + 1181.0, + 1159.0, + 1129.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1120.0, + 1406.0, + 1120.0, + 1406.0, + 1159.0, + 1210.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1152.0, + 510.0, + 1152.0, + 510.0, + 1187.0, + 294.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1152.0, + 1406.0, + 1152.0, + 1406.0, + 1187.0, + 549.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1182.0, + 325.0, + 1182.0, + 325.0, + 1217.0, + 294.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1182.0, + 678.0, + 1182.0, + 678.0, + 1217.0, + 449.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1182.0, + 935.0, + 1182.0, + 935.0, + 1217.0, + 783.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1182.0, + 1020.0, + 1182.0, + 1020.0, + 1217.0, + 965.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1182.0, + 1406.0, + 1182.0, + 1406.0, + 1217.0, + 1049.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1213.0, + 418.0, + 1213.0, + 418.0, + 1248.0, + 295.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1213.0, + 772.0, + 1213.0, + 772.0, + 1248.0, + 542.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1213.0, + 1408.0, + 1213.0, + 1408.0, + 1248.0, + 876.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1243.0, + 1406.0, + 1243.0, + 1406.0, + 1280.0, + 294.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1275.0, + 817.0, + 1275.0, + 817.0, + 1309.0, + 295.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1275.0, + 1080.0, + 1275.0, + 1080.0, + 1309.0, + 874.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1275.0, + 1167.0, + 1275.0, + 1167.0, + 1309.0, + 1110.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1275.0, + 1406.0, + 1275.0, + 1406.0, + 1309.0, + 1196.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1307.0, + 1407.0, + 1307.0, + 1407.0, + 1338.0, + 296.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1501.0, + 902.0, + 1501.0, + 902.0, + 1535.0, + 295.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1533.0, + 906.0, + 1533.0, + 906.0, + 1564.0, + 297.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1560.0, + 904.0, + 1560.0, + 904.0, + 1596.0, + 294.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1592.0, + 903.0, + 1592.0, + 903.0, + 1628.0, + 295.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1622.0, + 904.0, + 1622.0, + 904.0, + 1659.0, + 294.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1655.0, + 905.0, + 1655.0, + 905.0, + 1688.0, + 295.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1685.0, + 904.0, + 1685.0, + 904.0, + 1716.0, + 297.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 903.0, + 1712.0, + 903.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1746.0, + 904.0, + 1746.0, + 904.0, + 1777.0, + 296.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1776.0, + 906.0, + 1776.0, + 906.0, + 1807.0, + 296.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1807.0, + 491.0, + 1807.0, + 491.0, + 1820.0, + 465.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1809.0, + 518.0, + 1809.0, + 518.0, + 1819.0, + 502.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1810.0, + 630.0, + 1810.0, + 630.0, + 1818.0, + 622.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1810.0, + 662.0, + 1810.0, + 662.0, + 1818.0, + 650.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1807.0, + 773.0, + 1807.0, + 773.0, + 1820.0, + 764.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1810.0, + 834.0, + 1810.0, + 834.0, + 1818.0, + 824.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1808.0, + 905.0, + 1808.0, + 905.0, + 1821.0, + 891.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 312.0, + 1274.0, + 312.0, + 1274.0, + 350.0, + 295.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2000.0, + 1391.0, + 2000.0, + 1391.0, + 2039.0, + 296.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1801.0, + 773.0, + 1801.0, + 773.0, + 1845.0, + 294.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1801.0, + 1391.0, + 1801.0, + 1391.0, + 1845.0, + 819.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 225.0, + 1160.0, + 225.0, + 1160.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 297, + 219, + 1382, + 219, + 1382, + 1284, + 297, + 1284 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2112, + 835, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 4, + "poly": [ + 296, + 1296, + 1405, + 1296, + 1405, + 1451, + 296, + 1451 + ], + "score": 0.677 + }, + { + "category_id": 6, + "poly": [ + 296, + 1296, + 1405, + 1296, + 1405, + 1451, + 296, + 1451 + ], + "score": 0.198 + }, + { + "category_id": 6, + "poly": [ + 654, + 1199, + 1041, + 1199, + 1041, + 1226, + 654, + 1226 + ], + "score": 0.113 + }, + { + "category_id": 13, + "poly": [ + 578, + 1418, + 621, + 1418, + 621, + 1450, + 578, + 1450 + ], + "score": 0.89, + "latex": "\\ell _ { \\mathrm { i n f } }" + }, + { + "category_id": 13, + "poly": [ + 654, + 1199, + 692, + 1199, + 692, + 1226, + 654, + 1226 + ], + "score": 0.5, + "latex": "\\ell _ { \\mathrm { i n f } }" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 231.0, + 790.0, + 231.0, + 790.0, + 262.0, + 599.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 230.0, + 1241.0, + 230.0, + 1241.0, + 264.0, + 1122.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 271.0, + 401.0, + 271.0, + 401.0, + 303.0, + 303.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 268.0, + 514.0, + 268.0, + 514.0, + 302.0, + 451.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 265.0, + 683.0, + 265.0, + 683.0, + 306.0, + 552.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 268.0, + 832.0, + 268.0, + 832.0, + 305.0, + 694.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 269.0, + 936.0, + 269.0, + 936.0, + 301.0, + 877.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 267.0, + 1234.0, + 267.0, + 1234.0, + 305.0, + 974.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 267.0, + 1388.0, + 267.0, + 1388.0, + 303.0, + 1251.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 824.0, + 554.0, + 824.0, + 554.0, + 842.0, + 546.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 863.0, + 565.0, + 863.0, + 565.0, + 939.0, + 537.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1193.0, + 653.0, + 1193.0, + 653.0, + 1232.0, + 649.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1193.0, + 1043.0, + 1193.0, + 1043.0, + 1232.0, + 693.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1231.0, + 376.0, + 1231.0, + 376.0, + 1272.0, + 327.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1232.0, + 510.0, + 1232.0, + 510.0, + 1269.0, + 453.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1232.0, + 647.0, + 1232.0, + 647.0, + 1269.0, + 589.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1232.0, + 793.0, + 1232.0, + 793.0, + 1267.0, + 736.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1232.0, + 932.0, + 1232.0, + 932.0, + 1267.0, + 875.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1232.0, + 1069.0, + 1232.0, + 1069.0, + 1267.0, + 1012.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1232.0, + 1206.0, + 1232.0, + 1206.0, + 1269.0, + 1149.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1232.0, + 1354.0, + 1232.0, + 1354.0, + 1269.0, + 1297.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1296.0, + 1406.0, + 1296.0, + 1406.0, + 1330.0, + 295.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1327.0, + 1406.0, + 1327.0, + 1406.0, + 1360.0, + 295.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1356.0, + 1406.0, + 1356.0, + 1406.0, + 1392.0, + 294.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1387.0, + 1407.0, + 1387.0, + 1407.0, + 1421.0, + 294.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1417.0, + 577.0, + 1417.0, + 577.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1417.0, + 1337.0, + 1417.0, + 1337.0, + 1455.0, + 622.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1296.0, + 1406.0, + 1296.0, + 1406.0, + 1330.0, + 295.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1327.0, + 1406.0, + 1327.0, + 1406.0, + 1360.0, + 295.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1356.0, + 1406.0, + 1356.0, + 1406.0, + 1392.0, + 294.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1387.0, + 1407.0, + 1387.0, + 1407.0, + 1421.0, + 294.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1417.0, + 577.0, + 1417.0, + 577.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1417.0, + 1337.0, + 1417.0, + 1337.0, + 1455.0, + 622.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1193.0, + 1041.0, + 1193.0, + 1041.0, + 1232.0, + 693.0, + 1232.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/B1al7jg0b/B1al7jg0b.md b/parse/train/B1al7jg0b/B1al7jg0b.md new file mode 100644 index 0000000000000000000000000000000000000000..3d0812a0c88e1a26b04e52839fccc0158b08c740 --- /dev/null +++ b/parse/train/B1al7jg0b/B1al7jg0b.md @@ -0,0 +1,279 @@ +# OVERCOMING CATASTROPHIC INTERFERENCE USING CONCEPTOR-AIDED BACKPROPAGATION + +Xu He, Herbert Jaeger +Department of Computer Science and Electrical Engineering +Jacobs University Bremen +Bremen, 28759, Germany +{x.he,h.jaeger}@jacobs-university.de + +# ABSTRACT + +Catastrophic interference has been a major roadblock in the research of continual learning. Here we propose a variant of the back-propagation algorithm, “conceptor-aided backprop” (CAB), in which gradients are shielded by conceptors against degradation of previously learned tasks. Conceptors have their origin in reservoir computing, where they have been previously shown to overcome catastrophic forgetting. CAB extends these results to deep feedforward networks. On the disjoint and permuted MNIST tasks, CAB outperforms two other methods for coping with catastrophic interference that have recently been proposed. + +# 1 INTRODUCTION + +Agents with general artificial intelligence are supposed to learn and perform well on multiple tasks. Continual learning refers to the scenarios where a machine learning system can retain previously acquired skills while learning new ones. However, when trained on a sequence of tasks, neural networks usually forget about previous tasks after their weights are adjusted for a new task. This notorious problem known as catastrophic interference (CI) (McCloskey & Cohen, 1989; Ratcliff, 1990; French, 1999; Kumaran et al., 2016) poses a serious challenge towards continual learning. + +Many approaches have been proposed to overcome or mitigate the problem of CI in the last three decades (Hinton & Plaut, 1987; French, 1991; Ans & Rousset, 1997; French, 1997; Srivastava et al., 2014). Especially recently, an avalanche of new methods in the deep learning field has brought about dramatic improvements in continual learning in neural networks. Kirkpatrick et al. (2017) introduced a regularization-based method called elastic weight consolidation (EWC), which uses the posterior distribution of parameters for the old tasks as a prior for the new task. They approximated the posterior by a Gaussian distribution with the parameters for old tasks as the mean and the inverse diagonal of the Fisher information matrix as the variance. Lee et al. (2017) introduced two incremental moment matching (IMM) methods called mean-IMM and mode-IMM. Mean-IMM approximates the distribution of parameters for both old and new tasks by a Gaussian distribution, which is estimated by minimizing its KL-divergence from the mixture of two Gaussian posteriors, one for the old task and the other one for the new task. Mode-IMM estimates the mode of this mixture of two Gaussians and uses it as the optimal parameters for both tasks. + +In the field of Reservoir Computing (Jaeger, 2001; Maass et al., 2002), an effective solution to CI using conceptors was proposed by Jaeger (2014) to incrementally train a recurrent neural network to generate spatial-temporal signals. Conceptors are a general-purpose neuro-computational mechanism that can be used in a diversity of neural information processing tasks including temporal pattern classification, one-shot learning, human motion pattern generation, de-noising and signal separation (Jaeger, 2017). In this paper, we adopt and extend the method introduced in Jaeger (2014) and propose a conceptor-aided backpropagation (CAB) algorithm to train feed-forward networks. For each layer of a network, CAB computes a conceptor to characterize the linear subspace spanned by the neural activations in that layer that have appeared in already learned tasks. When the network is trained on a new task, CAB uses the conceptor to adjust the gradients given by backpropagation so that the linear transformation restricted to the characterized subspace will be preserved after the gradient descent procedure. Experiment results of two benchmark tests showed highly competitive performance of CAB. + +The rest of this paper is structured as follows. Section 2 introduces conceptors and their application to incremental learning by ridge regression. Section 3 extends the method to stochastic gradient descent and describes the CAB algorithm. Section 4 compares its performance on the permuted and disjoint MNIST tasks to recent methods that address the same problem. Finally we conclude our paper in Section 5. + +# 2 INCREMENTAL RIDGE REGRESSION BY CONCEPTORS + +This section reviews the basics of conceptor theory and its application to incrementally training linear readouts of recurrent neural networks as used in reservoir computing. A comprehensive treatment can be found in (Jaeger, 2014). + +# 2.1 CONCEPTORS + +![](images/8cc681eef02ddc59f5850e76547b11fae2efd1c7890b07b3a27fa22575869775.jpg) +Figure 1: 3D point clouds (black dots) and their corresponding conceptors, represented by ellipsoids whose axes are the singular vectors of conceptors and the lengths of these axes match the singular values of conceptors. Each edge of the plot boxes range from $- 1$ to $+ 1$ admitted by neural dynamics with a tanh nonlinearity; conceptor ellipsiods lie inside the unit sphere. + +In brief, a matrix conceptor $C$ for some vector-valued random variable $\boldsymbol { x } \in \mathbb { R } ^ { N }$ is defined as a linear transformation that minimizes the following loss function. + +$$ +\mathbb { E } _ { x } [ | | x - C x | | ^ { 2 } ] + \alpha ^ { - 2 } | | C | | _ { \mathrm { f r o } } ^ { 2 } +$$ + +where $\alpha$ is a control parameter called aperture and $| | \cdot | | _ { \mathrm { f r o } }$ is the Frobenius norm. This optimization problem has a closed-form solution + +$$ +C = R ( R + \alpha ^ { - 2 } I ) ^ { - 1 } +$$ + +where $R = \mathbb { E } _ { x } [ x x ^ { \top } ]$ is the $N \times N$ correlation matrix of $x$ , and $I$ is the $N \times N$ identity matrix. This result given in (2) can be understood by studying the singular value decomposition (SVD) of $C$ . If $R = \bar { U } \Sigma U ^ { \top }$ is the SVD of $R$ , then the SVD of $C$ is given as $U S U ^ { \top }$ , where the singular values $s _ { i }$ of $C$ can be written in terms of the singular values $\sigma _ { i }$ of $R$ : $s _ { i } = \sigma _ { i } / ( \sigma _ { i } + \alpha ^ { - 2 } ) \mathbf { \bar { \Omega } } \in [ 0 , 1 )$ . In intuitive terms, $C$ is a soft projection matrix on the linear subspace where the samples of $x$ lie. For a vector $y$ in this subspace, $C$ acts like the identity: $C y \approx y$ , and when some noise $\epsilon$ orthogonal to the subspace is added to $y$ , $C$ de-noises: $C ( y + \epsilon ) \approx y$ . Figure 1 shows the ellipsoids corresponding to three sets of $\mathbb { R } ^ { 3 }$ points. We define the quota $Q ( C )$ of a conceptor to be the mean singular values: $\begin{array} { r } { Q ( C ) : = \frac { 1 } { N } \sum _ { i = 1 } ^ { N } s _ { i } } \end{array}$ . Intuitively, the quota measures the fraction of the total dimensions of the entire vector space that is claimed by $C$ . + +Moreover, logic operations that satisfy most laws of Boolean logic can be defined on matrix conceptors as the following: + +$$ +\begin{array} { c } { { \neg C : = I - C , } } \\ { { C ^ { i } \vee C ^ { j } : = ( R ^ { i } + R ^ { j } ) ( R ^ { i } + R ^ { j } + \alpha ^ { - 2 } I ) ^ { - 1 } } } \\ { { C ^ { i } \wedge C ^ { j } : = \neg ( \neg C ^ { i } \vee \neg C ^ { j } ) } } \end{array} +$$ + +where $\lnot C$ softly projects onto a linear subspace that can be roughly understood as the orthogonal complement of the subspace characterized by $C$ . $C ^ { i } \vee C ^ { j }$ is the conceptor computed from the union of the two sets of sample points from which $C ^ { i }$ and $C ^ { j }$ are computed. It describes a space that is approximately the sum of linear subspaces characterized by $C ^ { i }$ and $C ^ { j }$ , respectively. The definition of $\bar { C } ^ { i } \wedge C ^ { j }$ reflects de Morgan’s law. Figure 2 illustrates the geometry of these operations. + +![](images/ac40caea8a02c56128dcbbb6714517c35c39b998c541267522f8ca3c04099e23.jpg) +Figure 2: Geometry of Boolean operations on 2-dimensional conceptors. The OR (resp. AND) operation gives a conceptor whose ellipsoid approximately is the smallest (largest) ellipsoid enclosing (contained in) the argument conceptor’s ellipsoids. + +# 2.2 INCREMENTAL RIDGE REGRESSION + +This subsection explains how conceptors can be applied to master continual learning in a simple linear model trained on a supervised task by ridge regression. The training is done sequentially on multiple input-to-output mapping tasks. This simplified scenario illustrates the working principle of continual learning with conceptors and will later be used repeatedly as a sub-procedure in the CAB algorithm for training multilayer feed-forward networks. + +Consider a sequence of $m$ incoming tasks indexed by $j$ . We denote the training dataset for the $j$ -th task by $\{ ( x _ { 1 } ^ { j } , y _ { 1 } ^ { j } ) , \cdot \cdot \cdot , ( x _ { n } ^ { j } , y _ { n } ^ { j } ) \}$ , where $\boldsymbol { x } _ { i } ^ { j } \in \mathbb { R } ^ { N }$ are input vectors and $y _ { i } ^ { j } \in \mathbb { R } ^ { M }$ their corresponding target outputs. Whenever the training dataset for a new task is available, the incremental learning method will compute a matrix conceptor $C ^ { j }$ for the input variable of the new task using Equation 2 and update the linear model, resulting in a sequence of linear models $W ^ { 1 } , \ldots W ^ { m }$ such that $W ^ { j }$ solves not only the $j$ -th task but also all previous tasks: for $k \leq j , y ^ { k } \approx W ^ { j } x ^ { k }$ . The conceptor $C ^ { j }$ is a soft projection matrix onto the linear subspace spanned by input patterns from the $j$ -th task. Then, $A ^ { j - 1 } \overset { \cdot } { = } \overset { \cdot } { C } ^ { 1 } \vee \cdots \vee C ^ { j - 1 }$ characterizes the memory space already claimed by the tasks $1 , \ldots , j - 1$ and $F ^ { j } = \neg A ^ { j - 1 }$ , the orthogonal complement of $A ^ { j } - 1$ , represents the memory space still free for the $j$ -th task. Here “memory space” refers to the linear space of input vectors. In detail, this method proceeds in the following way: + +• Initialization (no task trained yet): $W ^ { 0 } = 0 _ { M \times N } , A ^ { 0 } = 0 _ { N \times N }$ • Incremental task learning: For tasks $j = 1 , \ldots , m$ do: + +1. Store the input vectors from the $j$ -th training dataset of size $n$ into a $N \times n$ sized input collection matrix $X ^ { j }$ , and store the output vectors into a $M \times n$ sized output collection matrix $Y ^ { j }$ . + +2. Compute the conceptor for this task by $C ^ { j } ~ = ~ R ^ { j } ( R ^ { j } + \alpha ^ { - 2 } I ) ^ { - 1 }$ , where $R ^ { j } \ =$ $\scriptstyle { \frac { 1 } { n } } X ^ { j ^ { \prime } } X ^ { j ^ { \top } }$ + +3. Train an increment matrix $W _ { i n c } ^ { j }$ (to be added to $W ^ { j - 1 }$ , yielding $W ^ { j }$ ), with the crucial aid of a helper conceptor $F ^ { j }$ : + +(a) $F ^ { j } : = \neg A ^ { j - 1 }$ (comment: this conceptor characterizes the “still disposable” memory space for the $j$ -th task), +(b) $T : = Y ^ { j } - ( W ^ { j - 1 } X ^ { j } )$ (comment: this matrix consists of target values for a linear regression to compute $W _ { i n c . } ^ { j }$ ), +(c) $S : = F ^ { j } X ^ { j }$ (comment: this matrix consists of input arguments for the linear regression), + +(d) $W _ { i n c } ^ { j } \ : = \ : ( ( S S ^ { \top } / n + \lambda ^ { - 2 } I ) ^ { - 1 } S T ^ { \top } / n ) ^ { \top }$ (comment: carry out the regression, regularized by $\lambda ^ { - 2 }$ ), + +$W ^ { j }$ : $W ^ { j } = W ^ { j - 1 } + W _ { i n c } ^ { j }$ + +5. Update $A : A ^ { j } = A ^ { j - 1 } \vee C ^ { j }$ (comment: this is possible due to the associativity of the ∨ operation on conceptors) + +The weight increment $W _ { i n c } ^ { j }$ does not interfere much with the previously learned weights $W ^ { j - 1 }$ because the regularization in step 3(d) constrains the row space of $W _ { i n c } ^ { j }$ to be only the linear subspace spanned by input arguments defined in 3(c), which are inside the kernel of $W ^ { j - 1 }$ due to the projection by $F ^ { j }$ . Intuitively speaking, when learning a new task, this algorithm exploits only the components of input vectors in the still unused space (kernel of $W ^ { j - 1 }$ , characterized by $F ^ { j }$ ) to compensate errors for the new task and leaves the directions in the already used memory space (row space of $W ^ { j - 1 }$ , characterized by $A ^ { j - 1 }$ ) intact. + +# 3 CONCEPTOR-AIDED SGD AND BACK-PROP + +In this section, we first derive a stochastic gradient descent version of the algorithm described in the previous section, then present the procedure of CAB. + +# 3.1 SGD + +In the algorithm introduced in the previous section, $W _ { i n c } ^ { j }$ is computed by ridge regression, which offers a closed-form solution to minimize the following cost function + +$$ +\mathcal { I } ( W _ { i n c } ^ { j } ) : = \mathbb { E } [ | W _ { i n c } ^ { j } s - t | ^ { 2 } ] + \lambda ^ { - 2 } | W _ { i n c } ^ { j } | _ { \mathrm { f r o } } ^ { 2 } +$$ + +where $t = y ^ { j } - W ^ { j - 1 } x ^ { j } , s = F ^ { j } x ^ { j }$ . One can also minimize this cost function by stochastic gradient descent (SGD), which starts from an initial guess of $W _ { i n c } ^ { j }$ and repeatedly performs the following update + +$$ +W _ { i n c } ^ { j } W _ { i n c } ^ { j } - \eta \nabla _ { W _ { i n c } ^ { j } } \mathcal { I } ( W _ { i n c } ^ { j } ) +$$ + +where $\eta$ is the learning rate and the gradient is given by: + +$$ +\nabla _ { W _ { i n c } ^ { j } } \mathcal { I } ( W _ { i n c } ^ { j } ) = 2 \mathbb { E } [ ( W _ { i n c } ^ { j } s - t ) s ^ { \top } ] + 2 \lambda ^ { - 2 } W _ { i n c } ^ { j } +$$ + +Substituting $t$ by $y ^ { j } - W ^ { j - 1 } x ^ { j }$ and $s$ by $F ^ { j } x ^ { j } = ( I - A ^ { j - 1 } ) x ^ { j }$ in (8), we get + +$$ +\begin{array} { r l } & { \nabla _ { W _ { i n c } ^ { j } } \mathcal { I } ( W _ { i n c } ^ { j } ) = 2 \mathbb { E } [ ( W _ { i n c } ^ { j } ( I - A ^ { j - 1 } ) x ^ { j } - y ^ { j } + W ^ { j - 1 } x ^ { j } ) s ^ { \top } ] + 2 \lambda ^ { - 2 } W _ { i n c } ^ { j } } \\ & { \phantom { \frac { 1 } { 1 } } = 2 \mathbb { E } [ ( - W _ { i n c } ^ { j } A ^ { j - 1 } x ^ { j } + ( W ^ { j - 1 } + W _ { i n c } ^ { j } ) x ^ { j } - y ^ { j } ) s ^ { \top } ] + 2 \lambda ^ { - 2 } W _ { i n c } ^ { j } } \end{array} +$$ + +Due to the regularization term in the cost function, as the optimization goes on, eventually $W _ { i n c }$ will null the input components that are not inside the linear subspace characterized by $F ^ { j }$ , hence $W _ { i n c } ^ { j } A ^ { j - 1 } x ^ { j }$ will converge to 0 as the algorithm proceeds. In addition, since $W ^ { j } = W ^ { j - 1 } + W _ { i n c } ^ { j }$ , (10) can be simplified to + +$$ +\nabla _ { { W _ { i n c } ^ { j } } } \mathcal { I } ( W _ { i n c } ^ { j } ) = 2 \mathbb { E } [ ( W ^ { j } x ^ { j } - y ^ { j } ) s ^ { \top } ] + 2 \lambda ^ { - 2 } W _ { i n c } ^ { j } +$$ + +Adding $W ^ { j - 1 }$ to both sides of (7), we obtain the update rule for $W ^ { j }$ : + +$$ +\begin{array} { r } { W ^ { j } W ^ { j } - 2 \eta \mathbb { E } [ e s ^ { \top } ] + 2 \eta \lambda ^ { - 2 } W _ { i n c } ^ { j } } \end{array} +$$ + +where $e : = W ^ { j } x ^ { j } - y ^ { j }$ . In practice, at every iteration, the expected value can be approximated by a mini-batch of size $n _ { B }$ , indexed by $i _ { B }$ : + +$$ +\hat { \mathbb { E } } [ e s ^ { \top } ] = \frac { 1 } { n _ { B } } \sum _ { i _ { B } = 0 } ^ { L } ( W ^ { j } x _ { i _ { B } } ^ { j } - y _ { i _ { B } } ^ { j } ) ( F ^ { j } x _ { i _ { B } } ^ { j } ) ^ { \top } = \frac { 1 } { n _ { B } } \sum _ { i _ { B } = 0 } ^ { L } ( W ^ { j } x _ { i _ { B } } ^ { j } - y _ { i _ { B } } ^ { j } ) x _ { i _ { B } } ^ { j ^ { \top } } F ^ { j ^ { \top } } +$$ + +where the transpose for $F ^ { j }$ can be dropped since it is symmetric. + +If we only train the $j$ −th task without considering the previous tasks, the update rule given by normal SGD is + +$$ +W ^ { j } W ^ { j } - 2 \eta \mathbb { E } [ e x ^ { j \top } ] + 2 \eta \lambda ^ { - 2 } W ^ { j } +$$ + +Comparing this to the update rule in (12), we notice two modifications when a conceptor is adopted to avoid CI: first, the gradient of weights are calculated using the conceptor-projected input vector $s = F ^ { j } x ^ { j }$ instead of the original input vector $x ^ { j }$ ; second, regularization is done on the weight increment $W _ { i n c } ^ { j }$ rather than the final weight $W ^ { j }$ . These two modifications lead to our design of the conceptor-aided algorithm for training multilayer feed-forward networks. + +# 3.2 BACKPROP + +The basic idea of CAB is to guide the gradients of the loss function on every linear component of the network by a matrix conceptor computed from previous tasks during error back-propagation (Rumelhart et al., 1986), repeatedly applying the conceptor-aided SGD technique introduced in the previous section in every layer. + +Consider a feed-forward network with $L + 1$ layers, indexed by $l = 0 , \ldots L$ , such that the 0-th and the $L$ -th layers are the input and output layers respectively. $W ^ { ( l ) }$ represents the linear connections between the $( l - 1 )$ -th and the $l$ -th layer, where we refer to the former as the pre-synaptic layer with respect to $W ^ { ( l ) }$ , and to the latter as the post-synaptic layer. We denote by $N ^ { ( l ) }$ the size of the $l$ -th layer (excluding the bias unit) and $A ^ { ( l ) ^ { j } }$ a conceptor characterizing the memory space in the $l$ -th layer used up by the first $j$ tasks. Let $\sigma ( \cdot )$ be the activation function of the nonlinear neurons and $\theta$ all the parameters of the network to be trained. Then the incremental training method with CAB proceeds as follows: + +• Initialization (no task trained yet): $\forall l = 0 , \ldots , L - 1$ , $A ^ { ( l ) ^ { 0 } } : = 0 _ { ( N ^ { ( l ) } + 1 ) \times ( N ^ { ( l ) } + 1 ) }$ , and randomly initialize $W ^ { ( l + 1 ) ^ { 0 } }$ to be a matrix of size $\boldsymbol { N } ^ { ( l + 1 ) } \times \left( \boldsymbol { N } ^ { ( l ) } + 1 \right)$ . + +• Incremental task learning: For $j = 1 , \ldots , m$ do: + +1. $\forall l = 0 , \ldots , L - 1 , F ^ { ( l ) ^ { j } } = \lnot A ^ { ( l ) ^ { ( j - 1 ) } }$ . (This conceptor characterizes the still disposable vector space in layer l for learning task $j$ ) + +2. Update the network parameters $\theta ^ { ( j - 1 ) }$ obtained after training the first $j - 1$ tasks to $\theta ^ { j }$ by stochastic gradient descent, where the gradients are computed by CAB instead of the classical backprop. Algorithms 1 and 2 detail the forward and backward pass of CAB, respectively. Different from classical backprop, the gradients are guided by a matrix conceptor $F ^ { ( l ) ^ { j } }$ , such that in each layer only the activity in the still disposable memory space will contribute to the gradient. Note that the conceptors remain the same until convergence of the network for task $j$ . + +3. After training on the vectors, indexed by $i _ { B }$ $j$ -th task, run the forward procedure again on a batch of , taken from the $j$ -th training dataset, to collect activations $n _ { B }$ input $h _ { i _ { B } } ^ { ( l ) ^ { j } }$ of each layer into a $N ^ { ( l ) } \times n _ { B }$ sized matrix $H ^ { ( l ) ^ { j } }$ , and set the correlation matrix $\begin{array} { r } { R ^ { ( l ) ^ { j } } = \frac { \dot { 1 } } { n _ { B } } H ^ { ( l ) ^ { j } } ( H ^ { ( l ) ^ { j } } ) ^ { \top } } \end{array}$ . + +4. Compute a conceptor on the $l$ -th layer for the $j$ -th pattern by $C ^ { ( l ) ^ { j } } = R ^ { ( l ) ^ { j } } ( R ^ { ( l ) ^ { j } } +$ $\alpha ^ { - 2 } \hat { I } _ { N ^ { ( l ) } \times N ^ { ( l ) } } ) ^ { - 1 } , \forall l = 0 , \dots , L \bar { - } 1$ . Finding an optimal aperture can be done by a cross-validation search1. + +5. Update the conceptor for already used space in every layer: $A ^ { ( l ) ^ { j } } ~ = ~ A ^ { ( l ) ^ { j } } ~ \vee$ $C ^ { ( l ) ^ { j } } , \forall l = 0 , \dots , L - 1$ . + +Algorithm 1 The forward procedure of conceptor-aided backprop, adapted from the traditional backprop. Input vectors are passed through a feed-forward network to compute the cost function. $\mathcal { L } ( \hat { y } ^ { j } , y ^ { j } )$ denotes the loss for the $j$ -th task, to which a regularizer $\Omega ( \theta _ { i n c } ^ { j } ) = \Omega ( \theta ^ { j } - \theta ^ { j - 1 } ) =$ $| | \theta ^ { j } - \theta ^ { j - 1 } | | _ { \mathrm { f r o } } ^ { 2 }$ is added to obtain the total cost $\mathcal { I }$ , where $\theta$ contains all the weights (biases are considered as weights connected to the bias units). The increment of parameters rather than the parameters themselves are regularized, similar to the conceptor-aided SGD. + +Require: Network depth, $l$ +Require: $W _ { . } ^ { ( l ) ^ { j } } , l \in \bar { \{ 1 , \ldots , L \} }$ , the weight matrices of the network +Require: $x ^ { j }$ , one input vector of the $j$ -th task +Require: $y ^ { j }$ , the target output for $x ^ { j }$ +1: $h ^ { ( 0 ) } = x ^ { j }$ +2: for $l = 1 , \dots L$ do 3: $b ^ { ( l ) } = [ h ^ { ( l - 1 ) \top } , 1 ] ^ { \top }$ , include the bias unit 4: $a ^ { ( l ) } = W ^ { ( l ) ^ { j } } b ^ { ( l ) }$ 5: $h ^ { ( l ) } = \sigma ( a ^ { ( l ) } )$ +6: end for +7: ${ \hat { y } } ^ { j } = h ^ { ( l ) }$ +8: $\mathcal { I } = \mathcal { L } ( \hat { y } ^ { j } , y ^ { j } ) + \lambda \Omega ( \theta _ { i n c } ^ { j } )$ + +Algorithm 2 The backward procedure of conceptor-aided backprop for the $j$ -th task, adapted from the traditional backprop. The gradient $g$ of the loss function $\mathcal { L }$ on the activations $a ^ { ( l ) }$ represents the error for the linear transformation $W ^ { ( l ) ^ { j } }$ between the $( l - 1 )$ -th and the l−th layers. In the standard backprop algorithm, the gradient of $\mathcal { L }$ on $W ^ { ( l ) ^ { j } }$ is computed as an outer product of the post-synaptic errors $g$ and the pre-synaptic activities $\boldsymbol { h } ^ { ( l - 1 ) }$ . This resembles the computation of the gradient in the linear SGD algorithm, which motivates us to apply conceptors in a similar fashion as in the conceptor-aided SGD. Specifically, we project the gradient $\nabla _ { W ^ { ( l ) } } j \mathcal { L }$ by the matrix conceptor F (l−1) that indicates the free memory space on the pre-synaptic layer. + +1: + +$$ +\boldsymbol { g } \gets \nabla _ { \boldsymbol { \hat { y } } } \mathcal { I } = \nabla _ { \boldsymbol { \hat { y } } } \mathcal { L } ( \boldsymbol { \hat { y } } , \boldsymbol { y } ) +$$ + +2: for $l = L , L - 1 , \ldots , 1$ do + +3: Convert the gradient on the layer’s output into a gradient on the pre-nonlinearity activation ( $\odot$ denotes element-wise multiplication): + +$$ +g \nabla _ { a ^ { ( l ) } } \mathcal { I } = g \odot \sigma ^ { \prime } ( a ^ { ( l ) } ) +$$ + +4: Compute the gradient of weights, project it by $F ^ { ( l - 1 ) ^ { j } }$ , and add it to the regularization term on the increment: + +$$ +\begin{array} { l } { { \nabla _ { W ^ { ( l ) ^ { j } } } \mathcal { I } = g \big ( F ^ { ( l - 1 ) ^ { j } } b ^ { ( l - 1 ) } \big ) ^ { \top } + \lambda \nabla _ { W ^ { ( l ) ^ { j } } } \Omega \big ( \theta _ { i n c } ^ { j } \big ) = g b ^ { ( l - 1 ) ^ { \top } } F ^ { ( l - 1 ) ^ { j } } + 2 \lambda W _ { i n c } ^ { ( l ) ^ { j } } } } \\ { { \quad \quad = g b ^ { ( l - 1 ) ^ { \top } } F ^ { ( l - 1 ) ^ { j } } + 2 \lambda \big ( W ^ { ( l ) ^ { j } } - W ^ { ( l ) ^ { j - 1 } } \big ) } } \end{array} +$$ + +5: Propagate the gradients w.r.t. the next lower-level hidden layers activations: + +$$ +g \gets \nabla _ { h ^ { ( l - 1 ) } } \mathcal { I } = W ^ { ( l ) ^ { j } } { } ^ { \top } g +$$ + +6: end for + +![](images/1eda451ad86ac6f785232ef35b6928bbd7acc61b0179a85d29239e094465def7.jpg) +Figure 3: Average performance across already learned permuted MNIST tasks using CAB or EWC + +# 4 EXPERIMENTS + +# 4.1 PERMUTED MNIST EXPERIMENT + +To test the performance of CAB, we evaluated it on the permuted MNIST experiment (Srivastava et al., 2013; Goodfellow et al., 2014; Kirkpatrick et al., 2017; Lee et al., 2017), where a sequence of pattern recognition tasks are created from the MNIST dataset (LeCun et al., 1998). For each task, a random permutation of input image pixels is generated and applied to all images in MNIST to obtain a new shuffled dataset, equally difficult to recognize as the original one, the objective of each task is to recognize these images with shuffled pixels. + +For a proof-of-concept demonstration, we trained a simple but sufficient feed-forward network with [784-100-10] of neurons to classify 10 permuted MNIST datasets. The network has logistic sigmoid neurons in both hidden and output layers, and is trained with mean squared error as the cost function. Vanilla SGD was used in all experiments to optimize the cost function. Learning rate and aperture were set to 0.1 and 4, respectively. For comparison, we also tested EWC on the same task with the same network architecture, based on the implementation by Seff (2017). The parameters chosen for the EWC algorithm were 0.01 for the learning rate and 15 for the weight of the Fisher penalty term. Figure 3 shows the performance of CAB on this task, the average testing accuracy is $9 5 . 2 \%$ after learning all 10 tasks sequentially. Although a fair amount of effort was spent on searching for optimal parameters for EWC, the accuracies shown here might still not reflect its best performance. However, the same experiment with EWC was also conducted in Kemker et al. (2017), where the authors reimplemented EWC on a network with higher capacity (2 hidden layers and 400 ReLU neurons per layer) and the resulting average accuracy after learning 10 tasks sequentially was shown to be around $93 \%$ . + +Since all tasks are generated by permuting the same dataset, the portion of the input space occupied by each of them should have the same size. However, as more tasks are learned, the chance that the space of a new task will overlap with the already used input space increases. Figure 4 shows the singular value spectra and quota of the input and hidden layer conceptors every time after a new task is learned. As the incremental learning proceeds, it becomes less likely for a new task to be in the free space. For example, the second task increases the quota of the input layer memory space by 0.1, whereas the 10th task increases it by only 0.03. However, CAB still manages to make the network learn new tasks based on their input components in the non-overlapping space. + +![](images/53ed68df2db8177a3a69102033ca7c87cdf1a0c0aff09975806a8fe949298c3e.jpg) +(b) Singular value spectra of conceptors $A ^ { ( 1 ) ^ { j } }$ on the hidden layer. +Figure 4: The development of singular value spectra of conceptors for “used-up” space on the input layer and hidden layer during incremental learning of 10 permuted MNIST tasks. Quota of these conceptors are displayed in the legends. + +# 4.2 DISJOINT MNIST EXPERIMENT + +We then applied CAB to categorize the disjoint MNIST datasets into 10 classes (Srivastava et al., 2013; Lee et al., 2017). In this experiment, the original MNIST dataset is divided into two disjoint datasets with the first one consisting of data for the first five digits (0 to 4), and the second one of the remaining five digits (5 to 9). This task requires a network to learn these two datasets one after the other, then examines its performance of classifying the entire MNIST testing images into 10 classes. The current state-of-the-art accuracy on this task, averaged over 10 learning trials, is $9 4 . 1 2 ( \pm 0 . 2 7 ) \%$ , achieved by Lee et al. (2017) using IMM. They also tested EWC on the same task and the average accuracy was $5 2 . 7 2 ( \pm 1 . 3 6 ) \%$ . + +To test our method, we trained a feed-forward network with [784-800-10] neurons. Logistic sigmoid nonlinearities were used in both hidden and output layers, and the network was trained with vanilla SGD to minimize mean squared errors. The aperture $\alpha = 9$ was used for all conceptors on all layers, learning rate $\eta$ and regularization coefficient $\lambda$ were chosen to be 0.1 and 0.005 respectively. The accuracy of CAB on this task, measured by repeating the experiment 10 times, is $9 4 . 9 1 ( \pm 0 . { \dot { 3 } } 0 ) \%$ . It is worth mentioning that the network used by Lee et al. (2017) for testing IMM and EWC had [784-800-800-10] rectified linear units (ReLU), so CAB achieved better performance with fewer layers and neurons. + +# 4.3 COMPUTATIONAL COST + +If a conceptor is computed by ridge regression, the time complexity is $O ( n N ^ { 2 } + N ^ { 3 } )$ when the design matrix is dense, where $n$ is the number of samples and $N$ the number of features. In terms of wall time measures, the time taken to compute a conceptor from the entire MNIST training set (in this case, $n = 5 5 0 0 0$ images and $N = 7 8 4$ pixels, corresponding to the input layer in our networks) is 0.42 seconds of standard notebook CPU time on average. Although we did not implement it in these experiments, incremental online adaptation of conceptors by gradient descent is also possible in principle and would come at a cost of $\dot { O ( N ^ { 2 } ) }$ per update. + +# 5 CONCLUSION + +In this work, we first reviewed the conceptor-based incremental ridge regression algorithm, introduced in section 3.11 of Jaeger (2014) for memory management in recurrent neural networks. Then we derived its stochastic gradient descent version for optimizing the same objective. Finally we designed a conceptor-aided backprop algorithm by applying a conceptor to every linear layer of a feed-forward network. This method uses conceptors to guide gradients of parameters during the backpropagation procedure. As a result, learning a new task interferes only minimally with previously learned tasks, and the amount of already used network capacity can be monitored via the singular value spectra and quota of conceptors. + +In Jaeger (2014), different scenarios for continual learning are investigated in a reservoir computing setting. Two extreme cases are obtained when (i) the involved learning tasks are entirely unrelated to each other, versus (ii) all tasks come from the same parametric family of learning tasks. The two cases differ conspicuously with regards to the geometry of involved conceptors, and with regards to opportunities to re-use previously acquired functionality in subsequent learning episodes. The permuted MNIST task is an example of (i) while the disjoint MNIST task rather is of type (ii). Conceptors provide an analytical tool to discuss the “family relatedness” and enabling/disabling conditions for continual learning in geometrical terms. Ongoing and future research is devoted to a comprehensive mathematical analysis of these phenomena which in our view lie at the heart of understanding continual learning. + +# ACKNOWLEDGMENTS + +The work reported in this article was partly funded through the European H2020 collaborative project NeuRAM3 (grant Nr 687299). + +# REFERENCES + +Bernard Ans and Stephane Rousset. Avoiding catastrophic forgetting by coupling two reverberating ´ neural networks. Comptes Rendus de l’Academie des Sciences-Series III-Sciences de la Vie ´ , 320 (12):989–997, 1997. + +Robert M French. Using semi-distributed representations to overcome catastrophic forgetting in connectionist networks. Proceedings of the 13th Annual Cognitive Science Society Conference, pp. 173178, 1991. + +Robert M French. Pseudo-recurrent connectionist networks: An approach to the ‘sensitivitystability’ dilemma. Connection Science, 9(4):353–380, 1997. + +Robert M French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4):128–135, 1999. + +Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. International Conference on Learning Representations, 2014. + +Geoffrey E Hinton and David C Plaut. Using fast weights to deblur old memories. In Proceedings of the Ninth Annual Conference of the Cognitive Science Society, pp. 177–186. Lawrence Erlbaum Associates, 1987. + +Herbert Jaeger. The echo state approach to analysing and training recurrent neural networks-with an erratum note. German National Research Center for Information Technology GMD Technical Report, 148(34):13, 2001. + +Herbert Jaeger. Controlling recurrent neural networks by conceptors. Jacobs University Technical Reports, (31), 2014. https://arxiv.org/abs/1403.3369. + +Herbert Jaeger. Using conceptors to manage neural long-term memories for temporal patterns. Journal of Machine Learning Research, 18(13):1–43, 2017. URL http://jmlr.org/papers/ v18/15-449.html. + +Ronald Kemker, Angelina Abitino, Marc McClure, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. Computing Research Repository, abs/1708.02072, 2017. http: //arxiv.org/abs/1708.02072. + +James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521, 2017. + +Dharshan Kumaran, Demis Hassabis, and James L McClelland. What learning systems do intelligent agents need? complementary learning systems theory updated. Trends in Cognitive Sciences, 20 (7):512–534, 2016. + +Yann LeCun, Corinna Cortes, and Christopher JC Burges. The MNIST database of handwritten digits. 1998. http://yann.lecun.com/exdb/mnist/. + +Sang-Woo Lee, Jin-Hwa Kim, JungWoo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forgetting by incremental moment matching. Computing Research Repository, abs/1703.08475, 2017. http://arxiv.org/abs/1703.08475. + +Wolfgang Maass, Thomas Natschlager, and Henry Markram. Real-time computing without stable ¨ states: A new framework for neural computation based on perturbations. Neural Computation, 14(11):2531–2560, 2002. + +Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation, 24:109–165, 1989. + +Roger Ratcliff. Connectionist models of recognition memory: Constraints imposed by learning and forgetting functions. Psychological Review, 97(2):285–308, 1990. + +David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by backpropagating errors. Nature, 323:533–535, 1986. +Ari Seff. Implementation of overcoming catastrophic forgetting in neural networks in tensorflow. GitHub Repository, 2017. https://github.com/ariseff/ overcoming-catastrophic. +Rupesh K Srivastava, Jonathan Masci, Sohrob Kazerounian, Faustino Gomez, and Jurgen Schmid- ¨ huber. Compete to compute. In Advances in Neural Information Processing Systems, pp. 2310– 2318, 2013. http://papers.nips.cc/paper/5059-compete-to-compute.pdf. +Vipin Srivastava, Suchitra Sampath, and David J Parker. Overcoming catastrophic interference in connectionist networks using Gram-Schmidt orthogonalization. PloS ONE, 9(9):e105619, 2014. \ No newline at end of file diff --git a/parse/train/B1ewdt9xe/B1ewdt9xe.md b/parse/train/B1ewdt9xe/B1ewdt9xe.md new file mode 100644 index 0000000000000000000000000000000000000000..472493c46d7be210f961a3916f2e90ed7b8ef650 --- /dev/null +++ b/parse/train/B1ewdt9xe/B1ewdt9xe.md @@ -0,0 +1,301 @@ +# DEEP PREDICTIVE CODING NETWORKS FOR VIDEO PREDICTION AND UNSUPERVISED LEARNING + +William Lotter, Gabriel Kreiman & David Cox +Harvard University +Cambridge, MA 02215, USA +{lotter,davidcox}@fas.harvard.edu +gabriel.kreiman@tch.harvard.edu + +# ABSTRACT + +While great strides have been made in using deep learning algorithms to solve supervised learning tasks, the problem of unsupervised learning — leveraging unlabeled examples to learn about the structure of a domain — remains a difficult unsolved challenge. Here, we explore prediction of future frames in a video sequence as an unsupervised learning rule for learning about the structure of the visual world. We describe a predictive neural network (“PredNet”) architecture that is inspired by the concept of “predictive coding” from the neuroscience literature. These networks learn to predict future frames in a video sequence, with each layer in the network making local predictions and only forwarding deviations from those predictions to subsequent network layers. We show that these networks are able to robustly learn to predict the movement of synthetic (rendered) objects, and that in doing so, the networks learn internal representations that are useful for decoding latent object parameters (e.g. pose) that support object recognition with fewer training views. We also show that these networks can scale to complex natural image streams (car-mounted camera videos), capturing key aspects of both egocentric movement and the movement of objects in the visual scene, and the representation learned in this setting is useful for estimating the steering angle. Altogether, these results suggest that prediction represents a powerful framework for unsupervised learning, allowing for implicit learning of object and scene structure. + +# 1 INTRODUCTION + +Many of the most successful current deep learning architectures for vision rely on supervised learning from large sets of labeled training images. While the performance of these networks is undoubtedly impressive, reliance on such large numbers of training examples limits the utility of deep learning in many domains where such datasets are not available. Furthermore, the need for large numbers of labeled examples stands at odds with human visual learning, where one or a few views of an object is often all that is needed to enable robust recognition of that object across a wide range of different views, lightings and contexts. The development of a representation that facilitates such abilities, especially in an unsupervised way, is a largely unsolved problem. + +In addition, while computer vision models are typically trained using static images, in the real world, visual objects are rarely experienced as disjoint snapshots. Instead, the visual world is alive with movement, driven both by self-motion of the viewer and the movement of objects within the scene. Many have suggested that temporal experience with objects as they move and undergo transformations can serve as an important signal for learning about the structure of objects (Foldi ¨ ak, 1991; ´ Softky, 1996; Wiskott & Sejnowski, 2002; George & Hawkins, 2005; Palm, 2012; O’Reilly et al., 2014; Agrawal et al., 2015; Goroshin et al., 2015a; Lotter et al., 2015; Mathieu et al., 2016; Srivastava et al., 2015; Wang & Gupta, 2015; Whitney et al., 2016). For instance, Wiskott and Sejnowski proposed “slow feature analysis” as a framework for exploiting temporal structure in video streams (Wiskott & Sejnowski, 2002). Their approach attempts to build feature representations that extract slowly-varying parameters, such as object identity, from parameters that produce fast changes in the image, such as movement of the object. While approaches that rely on temporal coherence have arguably not yet yielded representations as powerful as those learned by supervised methods, they nonetheless point to the potential of learning useful representations from video (Mohabi et al., 2009; Sun et al., 2014; Goroshin et al., 2015a; Maltoni & Lomonaco, 2015; Wang & Gupta, 2015). + +Here, we explore another potential principle for exploiting video for unsupervised learning: prediction of future image frames (Softky, 1996; Palm, 2012; O’Reilly et al., 2014; Goroshin et al., 2015b; Srivastava et al., 2015; Mathieu et al., 2016; Patraucean et al., 2015; Finn et al., 2016; Vondrick et al., 2016). A key insight here is that in order to be able to predict how the visual world will change over time, an agent must have at least some implicit model of object structure and the possible transformations objects can undergo. To this end, we have designed a neural network architecture, which we informally call a “PredNet,” that attempts to continually predict the appearance of future video frames, using a deep, recurrent convolutional network with both bottom-up and topdown connections. Our work here builds on previous work in next-frame video prediction (Ranzato et al., 2014; Michalski et al., 2014; Srivastava et al., 2015; Mathieu et al., 2016; Lotter et al., 2015; Patraucean et al., 2015; Oh et al., 2015; Finn et al., 2016; Xue et al., 2016; Vondrick et al., 2016; Brabandere et al., 2016), but we take particular inspiration from the concept of “predictive coding” from the neuroscience literature (Rao & Ballard, 1999; Rao & Sejnowski, 2000; Lee & Mumford, 2003; Friston, 2005; Summerfield et al., 2006; Egner et al., 2010; Bastos et al., 2012; Spratling, 2012; Chalasani & Principe, 2013; Clark, 2013; O’Reilly et al., 2014; Kanai et al., 2015). Predictive coding posits that the brain is continually making predictions of incoming sensory stimuli (Rao & Ballard, 1999; Friston, 2005). Top-down (and perhaps lateral) connections convey these predictions, which are compared against actual observations to generate an error signal. The error signal is then propagated back up the hierarchy, eventually leading to an update of the predictions. + +We demonstrate the effectiveness of our model for both synthetic sequences, where we have access to the underlying generative model and can investigate what the model learns, as well as natural videos. Consistent with the idea that prediction requires knowledge of object structure, we find that these networks successfully learn internal representations that are well-suited to subsequent recognition and decoding of latent object parameters (e.g. identity, view, rotation speed, etc.). We also find that our architecture can scale effectively to natural image sequences, by training using car-mounted camera videos. The network is able to successfully learn to predict both the movement of the camera and the movement of objects in the camera’s view. Again supporting the notion of prediction as an unsupervised learning rule, the model’s learned representation in this setting supports decoding of the current steering angle. + +![](images/96a1440b9f8c8bf098c1b691d0dce930f9e50a5bd99be7f869e6ae09b81b7762.jpg) +Figure 1: Predictive Coding Network (PredNet). Left: Illustration of information flow within two layers. Each layer consists of representation neurons $( R _ { l } )$ , which output a layer-specific prediction at each time step $( \hat { A } _ { l } )$ , which is compared against a target $( A _ { l } )$ (Bengio, 2014) to produce an error term $( E _ { l } )$ , which is then propagated laterally and vertically in the network. Right: Module operations for case of video sequences. + +# 2 THE PREDNET MODEL + +The PredNet architecture is diagrammed in Figure 1. The network consists of a series of repeating stacked modules that attempt to make local predictions of the input to the module, which is then subtracted from the actual input and passed along to the next layer. Briefly, each module of the network consists of four basic parts: an input convolutional layer $( A _ { l } )$ , a recurrent representation layer $( R _ { l } )$ , a prediction layer $( \hat { \hat { A } } _ { l } )$ , and an error representation $( E _ { l } )$ . The representation layer, $R _ { l }$ , is a recurrent convolutional network that generates a prediction, $\hat { A } _ { l }$ , of what the layer input, $A _ { l }$ , will be on the next frame. The network takes the difference between $A _ { l }$ and $\hat { A } _ { l }$ and outputs an error representation, $E _ { l }$ , which is split into separate rectified positive and negative error populations. The error, $E _ { l }$ , is then passed forward through a convolutional layer to become the input to the next layer $( A _ { l + 1 } )$ . The recurrent prediction layer $R _ { l }$ receives a copy of the error signal $E _ { l }$ , along with top-down input from the representation layer of the next level of the network $( R _ { l + 1 } )$ . The organization of the network is such that on the first time step of operation, the “right” side of the network ( $A _ { l }$ ’s and $E _ { l }$ ’s) is equivalent to a standard deep convolutional network. Meanwhile, the “left” side of the network (the $R _ { l }$ ’s) is equivalent to a generative deconvolutional network with local recurrence at each stage. The architecture described here is inspired by that originally proposed by (Rao & Ballard, 1999), but is formulated in a modern deep learning framework and trained end-to-end using gradient descent, with a loss function implicitly embedded in the network as the firing rates of the error neurons. Our work also shares motivation with the Deep Predictive Coding Networks of Chalasani & Principe (2013); however, their framework is based upon sparse coding and a linear dynamical system with greedy layer-wise training, whereas ours is rooted in convolutional and recurrent neural networks trained with backprop. + +While the architecture is general with respect to the kinds of data it models, here we focus on image sequence (video) data. Consider a sequence of images, $x _ { t }$ . The target for the lowest layer is set to the the actual sequence itself, i.e. $\bar { A } _ { 0 } ^ { t } = x _ { t } \forall t$ . The targets for higher layers, $A _ { l } ^ { t }$ for $l > 0$ , are computed by a convolution over the error units from the layer below, $E _ { l - 1 } ^ { t }$ , followed by rectified linear unit (ReLU) activation and max-pooling. For the representation neurons, we specifically use convolutional LSTM units (Hochreiter & Schmidhuber, 1997; Shi et al., 2015). In our setting, the $R _ { l } ^ { t }$ hidden state is updated according to $R _ { l } ^ { t - 1 }$ , $E _ { l } ^ { t - 1 }$ , as well as $R _ { l + 1 } ^ { t }$ , which is first spatially upsampled (nearest-neighbor), due to the pooling present in the feedforward path. The predictions, $\hat { A } _ { l } ^ { t }$ are made through a convolution of the $R _ { l } ^ { t }$ stack followed by a ReLU non-linearity. For the lowest layer, $\hat { A } _ { l } ^ { t }$ is also passed through a saturating non-linearity set at the maximum pixel value: $\operatorname { S a t L U } ( x ; p _ { m a x } ) : = \operatorname* { m i n } ( p _ { m a x } , x )$ . Finally, the error response, $E _ { l } ^ { t }$ , is calculated from the difference between $\hat { A } _ { l } ^ { t }$ and $A _ { l } ^ { t }$ and is split into ReLU-activated positive and negative prediction errors, which are concatenated along the feature dimension. As discussed in (Rao & Ballard, 1999), although not explicit in their model, the separate error populations are analogous to the existence of on-center, off-surround and off-center, on-surround neurons early in the visual system. + +The full set of update rules are listed in Equations (1) to (4). The model is trained to minimize the weighted sum of the activity of the error units. Explicitly, the training loss is formalized in Equation 5 with weighting factors by time, $\lambda _ { t }$ , and layer, $\lambda _ { l }$ , and where $n _ { l }$ is the number of units in the lth layer. With error units consisting of subtraction followed by ReLU activation, the loss at each layer is equivalent to an L1 error. Although not explored here, other error unit implementations, potentially even probabilistic or adversarial (Goodfellow et al., 2014), could also be used. + +$$ +\begin{array} { r l } & { A _ { l } ^ { t } = \bigg \{ x _ { t } } & { \mathrm { i f } l = 0 } \\ & { \mathbf { M } _ { \mathrm { A X P o o L } } ( \operatorname { R E L U } ( \operatorname { C o N v } ( E _ { l - 1 } ^ { t } ) ) ) } & { l > 0 } \\ & { \hat { A } _ { l } ^ { t } = \operatorname { R E L U } ( \operatorname { C o N v } ( R _ { l } ^ { t } ) ) } \\ & { E _ { l } ^ { t } = [ \operatorname { R E L U } ( A _ { l } ^ { t } - \hat { A } _ { l } ^ { t } ) ; \operatorname { R E L U } ( \hat { A } _ { l } ^ { t } - A _ { l } ^ { t } ) ] } \\ & { R _ { l } ^ { t } = \operatorname { C o N v L S T M } ( E _ { l } ^ { t - 1 } , R _ { l } ^ { t - 1 } , \operatorname { U P S A M P L E } ( R _ { l + 1 } ^ { t } ) ) } \end{array} +$$ + +$$ +L _ { t r a i n } = \sum _ { t } \lambda _ { t } \sum _ { l } \frac { \lambda _ { l } } { n _ { l } } \sum _ { n _ { l } } E _ { l } ^ { t } +$$ + +# Algorithm 1 Calculation of PredNet states + +
Require: Xt
1:A←xt 2:E,R←0
3:fort=1 to Tdo
4:for l = L to O do
5:if l=L then
6:Rt =CONVLSTM(Et-1,Rt-1)
7:else
8:Rt = CONVLSTM(E𝑡-1,Rt-1,UPSAMPLE(Rt+1))
9:for l= O to L do
10:if l= O then
11:At = SATLU(RELU(CONV(Rδ)))
12:else
13:At = RELU(CoNv(Rt))
14:Et = [RELU(At - At); RELU(At - A)]
15:ifl<L then
16:At+1 = MAXPOOL(CONV(E))
+ +The order in which each unit in the model is updated must also be specified, and our implementation is described in Algorithm 1. Updating of states occurs through two passes: a top-down pass where the $R _ { l } ^ { t }$ states are computed, and then a forward pass to calculate the predictions, errors, and higher level targets. A last detail of note is that $R _ { l }$ and $E _ { l }$ are initialized to zero, which, due to the convolutional nature of the network, means that the initial prediction is spatially uniform. + +# 3 EXPERIMENTS + +# 3.1 RENDERED IMAGE SEQUENCES + +To gain an understanding of the representations learned in the proposed framework, we first trained PredNet models using synthetic images, for which we have access to the underlying generative stimulus model and all latent parameters. We created sequences of rendered faces rotating with two degrees of freedom, along the “pan” (out-of-plane) and “roll” (in-plane) axes. The faces start at a random orientation and rotate at a random constant velocity for a total of 10 frames. A different face was sampled for each sequence. The images were processed to be grayscale, with values normalized between 0 and 1, and 64x64 pixels in size. We used 16K sequences for training and 800 for both validation and testing. + +Predictions generated by a PredNet model are shown in Figure 2. The model is able to accumulate information over time to make accurate predictions of future frames. Since the representation neurons are initialized to zero, the prediction at the first time step is uniform. On the second time step, with no motion information yet, the prediction is a blurry reconstruction of the first time step. After further iterations, the model adapts to the underlying dynamics to generate predictions that closely match the incoming frame. + +For choosing the hyperparameters of the model, we performed a random search and chose the model that had the lowest L1 error in frame prediction averaged over time steps 2-10 on a validation set. Given this selection criteria, the best performing models tended to have a loss solely concentrated at the lowest layer (i.e. $\lambda _ { 0 } = 1$ , $\lambda _ { l > 0 } = 0 $ ), which is the case for the model shown. Using an equal loss at each layer considerably degraded predictions, but enforcing a moderate loss on upper layers that was one magnitude smaller than the lowest layer (i.e. $\lambda _ { 0 } = 1$ , $\lambda _ { l > 0 } = 0 . 1$ ) led to only slightly worse predictions, as illustrated in Figure 9 in the Appendix. In all cases, the time loss weight, $\lambda _ { t }$ , was set to zero for the first time step and then one for all time steps after. As for the remaining hyperparameters, the model shown has 5 layers with 3x3 filter sizes for all convolutions, max-pooling of stride 2, and number of channels per layer, for both $A _ { l }$ and $R _ { l }$ units, of (1, 32, 64, 128, 256). Model weights were optimized using the Adam algorithm (Kingma & Ba, 2014). + +![](images/47e158f8ba016838740a961f995997b73a4e432fd723f2faa4ecb1fae041441c.jpg) +time +Figure 2: PredNet next-frame predictions for sequences of rendered faces rotating with two degrees of freedom. Faces shown were not seen during training. + +Quantitative evaluation of generative models is a difficult, unsolved problem (Theis et al., 2016), but here we report prediction error in terms of meansquared error (MSE) and the Structural Similarity Index Measure (SSIM) (Wang et al., 2004). SSIM is designed to be more correlated with perceptual judgments, and ranges from $- 1$ and 1, with a larger score indicating greater similarity. We compare the PredNet to the trivial solution of copying the last frame, as well as a control model that shares the overall architecture and training scheme of the PredNet, but that sends forward the layer-wise activations $( A _ { l } )$ rather than the errors $( E _ { l } )$ . This model thus takes the form of a more traditional encoder-decoder pair, with a CNN encoder that has lateral skip connections to a convolutional LSTM decoder. The performance of all models on the rotating faces dataset is summarized in Table 1, where the scores were calculated as an average over all predictions after the first frame. We report results for the PredNet model trained with loss only on the lowest layer, denoted as PredNet $L _ { 0 }$ , as well as the model trained with an 0.1 weight on upper layers, denoted as PredNet $L _ { a l l }$ . Both PredNet models outperformed the baselines on both measures, with the $L _ { 0 }$ model slightly outperforming $L _ { a l l }$ , as expected for evaluating the pixel-level predictions. + +Table 1: Evaluation of next-frame predictions on Rotating Faces Dataset (test set). + +
MSESSIM
PredNet Lo0.01520.937
PredNet Lall0.01570.921
CNN-LSTM Enc.-Dec.0.01800.907
Copy Last Frame0.1250.631
+ +Synthetic sequences were chosen as the initial training set in order to better understand what is learned in different layers of the model, specifically with respect to the underlying generative model (Kulkarni et al., 2015). The rotating faces were generated using the FaceGen software package (Singular Inversions, Inc.), which internally generates 3D face meshes by a principal component analysis in “face space”, derived from a corpus of 3D face scans. Thus, the latent parameters of the image sequences used here consist of the initial pan and roll angles, the pan and roll velocities, and the principal component (PC) values, which control the “identity” of the face. To understand the information contained in the trained models, we decoded the latent parameters from the representation neurons $( R _ { l } )$ in different layers, using a ridge regression. The $R _ { l }$ states were taken at the earliest possible informative time steps, which, in the our notation, are the second and third steps, respectively, for the static and dynamic parameters. The regression was trained using $4 K$ sequences with 500 for validation and $1 K$ for testing. For a baseline comparison of the information implicitly embedded in the network architecture, we compare to the decoding accuracies of an untrained network with random initial weights. Note that in this randomly initialized case, we still expect above-chance decoding performance, given past theoretical and empirical work with random networks (Pinto et al., 2009; Jarrett et al., 2009; Saxe et al., 2010). + +Latent variable decoding accuracies of the pan and roll velocities, pan initial angle, and first PC are shown in the left panel of Figure 3. There are several interesting patterns. First, the trained models learn a representation that generally permits a better linear decoding of the underlying latent factors than the randomly initialized model, with the most striking difference in terms of the the pan rotation speed $( \alpha _ { p a n } )$ . Second, the most notable difference between the $L _ { a l l }$ and $L _ { 0 }$ versions occurs with the first principle component, where the model trained with loss on all layers has a higher decoding accuracy than the model trained with loss only on the lowest layer. + +![](images/0abef2797da5fa19f50184da371c6feb85f35b1d2b18d3d978fa3150145a690d.jpg) +Figure 3: Information contained in PredNet representation for rotating faces sequences. Left: Decoding of latent variables using a ridge regression $( \alpha _ { p a n }$ : pan (out-of-frame) angular velocity, $\theta _ { p a n }$ : pan angle, PC-1: first principal component of face, $\alpha _ { r o l l }$ : roll (in-frame) angular velocity). Right: Orientation-invariant classification of static faces. + +The latent variable decoding analysis suggests that the model learns a representation that may generalize well to other tasks for which it was not explicitly trained. To investigate this further, we assessed the models in a classification task from single, static images. We created a dataset of 25 previously unseen FaceGen faces at 7 pan angles, equally spaced between $[ - \frac { \pi } { 2 } , \frac { \pi } { 2 } ]$ , and 8 roll angles, equally spaced between $[ 0 , 2 \pi )$ . There were therefore orientations per identity, which were tested in a cross-validated fashion. A linear SVM to decode face identity was fit on a model’s representation of a random subset of orientations and then tested on the remaining angles. For each size of the SVM training set, ranging from 1-40 orientations per face, 50 different random splits were generated, with results averaged over the splits. + +For the static face classification task, we compare the PredNets to a standard autoencoder and a variant of the Ladder Network (Valpola, 2015; Rasmus et al., 2015). Both models were constructed to have the same number of layers and channel sizes as the PredNets, as well as a similar alternating convolution/max-pooling, then upsampling/convolution scheme. As both networks are autoencoders, they were trained with a reconstruction loss, with a dataset consisting of all of the individual frames from the sequences used to train the PredNets. For the Ladder Network, which is a denoising autoencoder with lateral skip connections, one must also choose a noise parameter, as well as the relative weights of each layer in the total cost. We tested noise levels ranging from 0 to 0.5 in increments of 0.1, with loss weights either evenly distributed across layers, solely concentrated at the pixel layer, or 1 at the bottom layer and 0.1 at upper layers (analogous to the PredNet $L _ { a l l }$ model). Shown is the model that performed best for classification, which consisted of 0.4 noise and only pixel weighting. Lastly, as in our architecture, the Ladder Network has lateral and top-down streams that are combined by a combinator function. Inspired by (Pezeshki et al., 2015), where a learnable MLP improved results, and to be consistent in comparing to the PredNet, we used a purely convolutional combinator. Given the distributed representation in both networks, we decoded from a concatenation of the feature representations at all layers, except the pixel layer. For the PredNets, the representation units were used and features were extracted after processing one input frame. + +Face classification accuracies using the representations learned by the $L _ { 0 }$ and $L _ { a l l }$ PredNets, a standard autoencoder, and a Ladder Network variant are shown in the right panel of Figure 3. Both PredNets compare favorably to the other models at all sizes of the training set, suggesting they learn a representation that is relatively tolerant to object transformations. Similar to the decoding accuracy of the first principle component, the PredNet $L _ { a l l }$ model actually outperformed the $L _ { 0 }$ variant. Altogether, these results suggest that predictive training with the PredNet can be a viable alternative to other models trained with a more traditional reconstructive or denoising loss, and that the relative layer loss weightings $( \lambda _ { l } ^ { } \mathbf { \dot { s } } )$ may be important for the particular task at hand. + +# 3.2 NATURAL IMAGE SEQUENCES + +We next sought to test the PredNet architecture on complex, real-world sequences. As a testbed, we chose car-mounted camera videos, since these videos span across a wide range of settings and are characterized by rich temporal dynamics, including both self-motion of the vehicle and the motion of other objects in the scene (Agrawal et al., 2015). Models were trained using the raw videos from the KITTI dataset (Geiger et al., 2013), which were captured by a roof-mounted camera on a car driving around an urban environment in Germany. Sequences of 10 frames were sampled from the “City”, “Residential”, and “Road” categories, with 57 recording sessions used for training and 4 used for validation. Frames were center-cropped and downsampled to 128x160 pixels. In total, the training set consisted of roughly 41K frames. + +A random hyperparameter search, with model selection based on the validation set, resulted in a 4 layer model with 3x3 convolutions and layer channel sizes of (3, 48, 96, 192). Models were again trained with Adam (Kingma & Ba, 2014) using a loss either solely computed on the lowest layer $( L _ { 0 } )$ or with a weight of 1 on the lowest layer and 0.1 on the upper layers $( L _ { a l l } )$ . Adam parameters were initially set to their default values $\mathbf { \Phi } _ { \mathrm { { ( } } \alpha } = 0 . 0 0 1 $ , $\beta _ { 1 } = 0 . 9$ , $\beta _ { 2 } = 0 . 9 9 9 )$ with the learning rate, $\alpha$ , decreasing by a factor of 10 halfway through training. To assess that the network had indeed learned a robust representation, we tested on the CalTech Pedestrian dataset (Dollar et al., 2009), which´ consists of videos from a dashboard-mounted camera on a vehicle driving around Los Angeles. Testing sequences were made to match the frame rate of the KITTI dataset and again cropped to 128x160 pixels. Quantitative evaluation was performed on the entire CalTech test partition, split into sequences of 10 frames. + +Sample PredNet predictions (for the $L _ { 0 }$ model) on the CalTech Pedestrian dataset are shown in Figure 4, and example videos can be found at https://coxlab.github.io/prednet/. The model is able to make fairly accurate predictions in a wide range of scenarios. In the top sequence of Fig. 4, a car is passing in the opposite direction, and the model, while not perfect, is able to predict its trajectory, as well as fill in the ground it leaves behind. Similarly in Sequence 3, the model is able to predict the motion of a vehicle completing a left turn. Sequences 2 and 5 illustrate that the PredNet can judge its own movement, as it predicts the appearance of shadows and a stationary vehicle as they approach. The model makes reasonable predictions even in difficult scenarios, such as when the camera-mounted vehicle is turning. In Sequence 4, the model predicts the position of a tree, as the vehicle turns onto a road. The turning sequences also further illustrate the model’s ability to “fill-in”, as it is able to extrapolate sky and tree textures as unseen regions come into view. As an additional control, we show a sequence at the bottom of Fig. 4, where the input has been temporally scrambled. In this case, the model generates blurry frames, which mostly just resemble the previous frame. Finally, although the PredNet shown here was trained to predict one frame ahead, it is also possible to predict multiple frames into the future, by feeding back predictions as the inputs and recursively iterating. We explore this in Appendix 5.3. + +Quantitatively, the PredNet models again outperformed the CNN-LSTM EncoderDecoder. To ensure that the difference in performance was not simply because of the choice of hyperparameters, we trained models with four other sets of hyperparameters, which were sampled from the initial random search over the number of layers, fil + +Table 2: Evaluation of Next-Frame Predictions on CalTech Pedestrian Dataset. + +
MSESSIM
PredNet Lo3.13 × 10-30.884
PredNet Lall3.33 ×10-30.875
CNN-LSTMEnc.-Dec.3.67 × 10-30.865
Copy Last Frame7.95 × 10-30.762
+ +ter sizes, and number of filters per layer. For each of the four additional sets, the PredNet $L _ { 0 }$ had the best performance, with an average error reduction of $1 4 . 7 \%$ and $1 4 . 9 \%$ for MSE and SSIM, + +![](images/f2465bddecf762950203060719579f704eebd5b1c06d29b0051a317691866e37.jpg) +time → +Figure 4: PredNet predictions for car-cam videos. The first rows contain ground truth and the second rows contain predictions. The sequence below the red line was temporally scrambled. The model was trained on the KITTI dataset and sequences shown are from the CalTech Pedestrian dataset. + +respectively, compared to the CNN-LSTM Encoder-Decoder. More details, as well as a thorough investigation of systematically simplified models on the continuum between the PredNet and the CNN-LSTM Encoder-Decoder can be found in Appendix 5.1. Briefly, the elementwise subtraction operation in the PredNet seems to be beneficial, and the nonlinearity of positive/negative splitting also adds modest improvements. Finally, while these experiments measure the benefits of each component of our model, we also directly compare against recent work in a similar car-cam setting, by reporting results on a 64x64 pixel, grayscale car-cam dataset released by Brabandere et al. (2016). Our PredNet model outperforms the model by Brabandere et al. (2016) by $2 9 \%$ . Details can be found in Appendix 5.2. Also in Appendix 5.2, we present results for the Human3.6M (Ionescu et al., 2014) dataset, as reported by Finn et al. (2016). Without re-optimizing hyperparameters, our model underperforms the concurrently developed DNA model by Finn et al. (2016), but outperforms the model by Mathieu et al. (2016). + +To test the implicit encoding of latent parameters in the car-cam setting, we used the internal representation in the PredNet to estimate the car’s steering angle (Bojarski et al., 2016; Biasini et al., 2016). We used a dataset released by Comma.ai (Biasini et al., 2016) consisting of 11 videos totaling about 7 hours of mostly highway driving. We first trained networks for next-frame prediction and then fit a linear fully-connected layer on the learned representation to estimate the steering angle, using a MSE loss. We again concatenate the $R _ { l }$ representation at all layers, but first spatially average pool lower layers to match the spatial size of the upper layer, in order to reduce dimensionality. Steering angle estimation results, using the representation on the $1 0 ^ { \mathrm { t h } }$ time step, are shown in Figure 5. Given just 1K labeled training examples, a simple linear readout on the PredNet $L _ { 0 }$ representation explains $7 4 \%$ of the variance in the steering angle and outperforms the CNN-LSTM Enc.-Dec. by $3 5 \%$ . With 25K labeled training examples, the PredNet $L _ { 0 }$ has a MSE (in degrees2) of 2.14. As a point of reference, a CNN model designed to predict the steering angle (Biasini et al., 2016), albeit from a single frame instead of multiple frames, achieve a MSE of ${ \sim } 4$ when trained end-to-end using 396K labeled training examples. Details of this analysis can be found in Appendix 8. Interestingly, in this task, the PredNet $L _ { a l l }$ model actually underperformed the $L _ { 0 }$ model and slightly underperformed the CNN-LSTM Enc.-Dec, again suggesting that the $\lambda _ { l }$ parameter can affect the representation learned, and different values may be preferable in different end tasks. Nonetheless, the readout from the $L _ { a l l }$ model still explained a substantial proportion of the steering angle variance and strongly outperformed the random initial weights. Overall, this analysis again demonstrates that a representation learned through prediction, and particularly with the PredNet model with appropriate hyperparameters, can contain useful information about underlying latent parameters. + +![](images/4eacc964ba30df01ba5dfac8e925e09b97a2e2cd91639d41fa77b3ced7bfe30a.jpg) +Figure 5: Steering angle estimation accuracy on the Comma.ai dataset (Biasini et al., 2016). Left: Example steering angle curve with model estimations for a segment in the test set. Decoding was performed using a fully-connected readout on the PredNet representation trained with 25K labeled training examples. PredNet representation was trained for next-frame prediction on Comma.ai training set. Right: Mean-squared error of steering angle estimation. + +# 4 DISCUSSION + +Above, we have demonstrated a predictive coding inspired architecture that is able to predict future frames in both synthetic and natural image sequences. Importantly, we have shown that learning to predict how an object or scene will move in a future frame confers advantages in decoding latent parameters (such as viewing angle) that give rise to an object’s appearance, and can improve recognition performance. More generally, we argue that prediction can serve as a powerful unsupervised learning signal, since accurately predicting future frames requires at least an implicit model of the objects that make up the scene and how they are allowed to move. Developing a deeper understanding of the nature of the representations learned by the networks, and extending the architecture, by, for instance, allowing sampling, are important future directions. + +# ACKNOWLEDGMENTS + +We would like to thank Rasmus Berg Palm for fruitful discussions and early brainstorming. We would also like to thank the developers of Keras (Chollet, 2016). This work was supported by IARPA (contract D16PC00002), the National Science Foundation (NSF IIS 1409097), and the Center for Brains, Minds and Machines (CBMM, NSF STC award CCF-1231216). + +# REFERENCES + +Pulkit Agrawal, Joao Carreira, and Jitendra Malik. Learning to see by moving. ˜ CoRR, 2015. + +Andre M. Bastos, W. Martin Usrey, Rick A. Adams, George R. Mangun, Pascal Fries, and Karl J. Friston. Canonical microcircuits for predictive coding. Neuron, 2012. + +Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. CoRR, 2015. + +Yoshua Bengio. How auto-encoders could provide credit assignment in deep networks via target propagation. CoRR, 2014. + +Riccardo Biasini, George Hotz, Sam Khalandovsky, Eder Santana, and Niel van der Westhuizen. Comma.ai research, 2016. URL https://github.com/commaai/research. + +Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba. End to end learning for self-driving cars. CoRR, 2016. + +Bert De Brabandere, Xu Jia, Tinne Tuytelaars, and Luc Van Gool. Dynamic filter networks. CoRR, 2016. + +Rakesh Chalasani and Jose C. Principe. Deep predictive coding networks. CoRR, 2013. + +Franc¸ois Chollet. Comma.ai, 2016. URL http://keras.io/. + +Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 2013. + +Piotr Dollar, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark.´ In CVPR, 2009. + +Tobias Egner, Jim M. Monti, and Christopher Summerfield. Expectation and surprise determine neural population responses in the ventral visual stream. J Neurosci, 2010. + +Chelsea Finn, Ian J. Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction through video prediction. CoRR, 2016. + +Peter Foldi ¨ ak. Learning invariance from transformation sequences. ´ Neural Computation, 1991. + +Karl Friston. A theory of cortical responses. Philos Trans R Soc Lond B Biol Sci, 2005. + +Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR), 2013. + +Dileep George and Jeff Hawkins. A hierarchical bayesian model of invariant pattern recognition in the visual cortex. In Proceedings of the International Joint Conference on Neural Networks. IEEE, 2005. + +Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS. 2014. + +Ross Goroshin, Joan Bruna, Jonathan Tompson, David Eigen, and Yann LeCun. Unsupervised learning of spatiotemporally coherent metrics. CoRR, 2015a. + +Ross Goroshin, Michael Mathieu, and Yann LeCun. Learning to linearize under uncertainty. ¨ CoRR, 2015b. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. Neural Computation, 1997. + +Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014. + +Kevin Jarrett, Koray Kavukcuoglu, MarcAurelio Ranzato, and Yann LeCun. What is the best multistage architecture for object recognition? In ICCV. 2009. + +Ryota Kanai, Yutaka Komura, Stewart Shipp, and Karl Friston. Cerebral hierarchies : predictive processing , precision and the pulvinar. Philos Trans R Soc Lond B Biol Sci, 2015. + +Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, 2014. + +Tejas D. Kulkarni, Will Whitney, Pushmeet Kohli, and Joshua B. Tenenbaum. Deep convolutional inverse graphics network. CoRR, 2015. + +Tai Sing Lee and David Mumford. Hierarchical bayesian inference in the visual cortex. J Opt Soc Am A Opt Image Sci Vis, 2003. + +William Lotter, Gabriel Kreiman, and David Cox. Unsupervised learning of visual structure using predictive generative networks. CoRR, 2015. + +Davide Maltoni and Vincenzo Lomonaco. Semi-supervised tuning from temporal coherence. CoRR, 2015. + +Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond ¨ mean square error. ICLR, 2016. + +Vincent Michalski, Roland Memisevic, and Kishore Konda. Modeling deep temporal dependencies with recurrent ”grammar cells”. In NIPS. 2014. + +Hossein Mohabi, Ronan Collobert, and Jason Weston. Deep learning from temporal coherence in video. In ICML. 2009. + +Junhyuk Oh, Xiaoxiao Guo, Honglak Lee, Richard L. Lewis, and Satinder P. Singh. Actionconditional video prediction using deep networks in atari games. CoRR, 2015. + +Randall C. O’Reilly, Dean Wyatte, and John Rohrlich. Learning through time in the thalamocortical loops. CoRR, 2014. + +Rasmus Berg Palm. Prediction as a candidate for learning deep hierarchical models of data. Master’s thesis, Technical University of Denmark, 2012. + +Viorica Patraucean, Ankur Handa, and Roberto Cipolla. Spatio-temporal video autoencoder with differentiable memory. CoRR, 2015. + +Mohammad Pezeshki, Linxi Fan, Philemon Brakel, Aaron C. Courville, and Yoshua Bengio. Deconstructing the ladder network architecture. CoRR, 2015. + +Nicolas Pinto, David Doukhan, James J. DiCarlo, and David D. Cox. A high-throughput screening approach to discovering good forms of biologically inspired visual representation. PLoS Comput Biol, 2009. + +Marc’Aurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit ¨ Chopra. Video (language) modeling: a baseline for generative models of natural videos. CoRR, 2014. + +Rajesh P. N. Rao and Dana H. Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 1999. + +Rajesh P. N. Rao and T. J. Sejnowski. Predictive sequence learning in recurrent neocortical circuits. NIPS, 2000. + +Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, and Tapani Raiko. Semisupervised learning with ladder network. CoRR, 2015. + +Eder Santana and George Hotz. Learning a driving simulator. CoRR, 2016. + +Andrew Saxe, Maneesh Bhand, Zhenghao Chen, Pang Wei Koh, Bipin Suresh, and Andrew Y. Ng. On random weights and unsupervised feature learning. In Workshop: Deep Learning and Unsupervised Feature Learning (NIPS). 2010. + +Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. CoRR, 2015. + +Singular Inversions, Inc. FaceGen. http://facegen.com. + +William R. Softky. Unsupervised pixel-prediction. NIPS, 1996. + +M. W. Spratling. Unsupervised learning of generative and discriminative weights encoding elementary image components in a predictive coding model of cortical function. Neural Computation, 2012. + +Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised learning of video representations using lstms. CoRR, 2015. + +Christopher Summerfield, Tobias Egner, Matthew Greene, Etienne Koechlin, Jennifer Mangels, and Joy Hirsch. Predictive codes for forthcoming perception in the frontal cortex. Science, 314, 2006. + +Lin Sun, Kui Jia, Tsung-Han Chan, Yuqiang Fang, Gang Wang, and Shuicheng Yan. Dl-sfa: Deeplylearned slow feature analysis for action recognition. CVPR, 2014. + +Lucas Theis, Aaron van den Oord, and Matthias Bethge. A note on the evaluation of generative models. ICLR, 2016. + +Harri Valpola. From neural pca to deep unsupervised learning. CoRR, 2015. + +Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. CoRR, 2016. + +Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. CoRR, 2015. + +Zhou Wang, Alan Bovik, Hamid Sheikh, and Eero Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. + +William F. Whitney, Michael Chang, Tejas D. Kulkarni, and Joshua B. Tenenbaum. Understanding visual concepts with continuation learning. CoRR, 2016. + +Laurenz Wiskott and Terrence J. Sejnowski. Learning invariance from transformation sequences. Neural Computation, 2002. + +Tianfan Xue, Jiajun Wu, Katherine L. Bouman, and William T. Freeman. Visual dynamics: Probabilistic future frame synthesis via cross convolutional networks. CoRR, 2016. + +# 5 APPENDIX + +# 5.1 ADDITIONAL CONTROL MODELS + +Table 3 contains results for additional variations of the PredNet and CNN-LSTM Encoder-Decoder evaluated on the CalTech Pedestrian Dataset after being trained on KITTI. We evaluate the models in terms of pixel prediction, thus using the PredNet model trained with loss only on the lowest layer (PredNet $L _ { 0 }$ ) as the base model. In addition to mean-squared error (MSE) and the Structural Similarity Index Measure (SSIM), we include calculations of the Peak Signal-To-Noise Ratio (PSNR). For each model, we evaluate it with the original set of hyperparameters (controlling the number of layers, filter sizes, and number of filters per layer), as well as with the four additional sets of hyperparameters that were randomly sampled from the initial random search (see main text for more details). Below is an explanation of the additional control models: + +• PredNet (no E split): PredNet model except the error responses $( E _ { l } )$ are simply linear $( \hat { A } _ { l } - A _ { l } )$ instead of being split into positive and negative rectifications. + +• CNN-LSTM Enc.-Dec. $2 \mathbf { x } \ A _ { l }$ filts): CNN-LSTM Encoder-Decoder model ( $\mathbf { \delta } _ { \cdot } A _ { l }$ ’s are passed instead of $E _ { l }$ ’s) except the number of filters in $A _ { l }$ is doubled. This controls for the total number of filters in the model compared to the PredNet, since the PredNet has filters to produce $\hat { A } _ { l }$ at each layer, which is integrated into the model’s feedforward response. + +• CNN-LSTM Enc.-Dec. (except pass $E _ { 0 }$ ): CNN-LSTM Encoder-Decoder model except the error is passed at the lowest layer. All remaining layers pass the activations $A _ { l }$ . With training loss taken at only the lowest layer, this variation allows us to determine if the “prediction” subtraction operation in upper layers, which is essentially unconstrained and learnable in the $L _ { 0 }$ case, aids in the model’s performance. + +• CNN-LSTM Enc.-Dec. $+ / -$ split): CNN-LSTM Encoder-Decoder model except the activations $A _ { l }$ are split into positive and negative populations before being passed to other layers in the network. This isolates the effect of the additional nonlinearity introduced by this procedure. + +Table 3: Quantitative evaluation of additional controls for next-frame prediction in CalTech Pedestrian Dataset after training on KITTI. First number indicates score with original hyperparameters. Number in parenthesis indicates score averaged over total of five different hyperparameters. + +
MSE (x 10-3)PSNRSSIM
PredNet3.13 (3.33)25.8 (25.5)0.884 (0.878)
PredNet (no Et split)3.20 (3.37)25.6 (25.4)0.883 (0.878)
CNN-LSTMEnc.-Dec.3.67 (3.91)25.0 (24.6)0.865 (0.856)
CNN-LSTM Enc.-Dec. (2x At filts)3.82 (3.97)24.8 (24.6)0.857 (0.853)
CNN-LSTM Enc.-Dec. (except pass Eo)3.41 (3.61)25.4 (25.1)0.873 (0.866)
CNN-LSTMEnc.-Dec. (+/- split)3.71 (3.84)24.9 (24.7)0.861 (0.857)
Copy Last Frame7.9520.00.762
+ +Equalizing the number of filters in the CNN-LSTM Encoder-Decoder (2x $A _ { l }$ filts) cannot account for its performance difference with the PredNet, and actually leads to overfitting and a decrease in performance. Passing the error at the lowest layer $( E _ { 0 } )$ in the CNN-LSTM Enc.-Dec. improves performance, but still does not match the PredNet, where errors are passed at all layers. Finally, splitting the activations $A _ { l }$ into positive and negative populations in the CNN-LSTM Enc.-Dec. does not help, but the PredNet with linear error activation (“no $E _ { l }$ split”) performs slightly worse than the original split version. Together, these results suggest that the PredNet’s error passing operation can lead to improvements in next-frame prediction performance. + +# 5.2 COMPARING AGAINST OTHER MODELS + +While our main comparison in the text was a control model that isolates the effects of the more unique components in the PredNet, here we directly compare against other published models. We report results on a $6 4 \mathrm { x 6 4 }$ pixel, grayscale car-cam dataset and the Human3.6M dataset (Ionescu et al., 2014) to compare against the two concurrently developed models by Brabandere et al. (2016) + +and Finn et al. (2016), respectively. For both comparisons, we use a model with the same hyperparameters (# of layers, # of filters, etc.) of the PredNet $L _ { 0 }$ model trained on KITTI, but train from scratch on the new datasets. The only modification we make is to train using an L2 loss instead of the effective L1 loss, since both models train with an L2 loss and report results using L2-based metrics (MSE for Brabandere et al. (2016) and PSNR for Finn et al. (2016)). That is, we keep the original PredNet model intact but directly optimize using MSE between actual and predicted frames. We measure next-frame prediction performance after inputting 3 frames and 10 frames, respectively, for the 64x64 car-cam and Human3.6M datasets, to be consistent with the published works. We also include the results using a feedforward multi-scale network, similar to the model of Mathieu et al. (2016), on Human3.6M, as reported by Finn et al. (2016). + +Table 4: Evaluation of Next-Frame Predictions on 64x64 Car-Cam Dataset. MSE (per-pixel) + +
MSE (per-pixel)
DFN (Brabandere et al., 2016)1.71 ×10-3
PredNet1.16 ×10-3
Copy Last Frame3.58 ×10-3
+ +Table 5: Evaluation of Next-Frame Predictions on Human3.6M PSNR + +
DNA (Finn et al., 2016) PredNet FF multi-scale (Mathieu et al.,2016)42.1 38.9 26.7 32.0
+ +On a dataset similar to KITTI, our model outperforms the model proposed by Brabandere et al. (2016). On Human3.6M, our model outperforms a model similar to (Mathieu et al., 2016), but underperforms Finn et al. (2016), although we note we did not perform any hyperparameter optimization. + +# 5.3 MULTIPLE TIME STEP PREDICTION + +![](images/d9c247d44b99433c9cfb95c59a9b7fce77018f62bc8529b66fb02544335c6500.jpg) +Figure 6: Extrapolation sequences generated by feeding PredNet predictions back into model. Left of the orange line: Normal $t + 1$ predictions; Right: Generated by recursively using the predictions as input. First row: Ground truth sequences. Second row: Generated frames of the original model, trained to solely predict $t + 1$ . Third row: Model fine-tuned for extrapolation. + +While the models presented here were originally trained to predict one frame ahead, they can be made to predict multiple frames by treating predictions as actual input and recursively iterating. Examples of this process are shown in Figure 6 for the PredNet $L _ { 0 }$ model. Although the next frame predictions are reasonably accurate, the model naturally breaks down when extrapolating further into the future. This is not surprising since the predictions will unavoidably have different statistics than the natural images for which the model was trained to handle (Bengio et al., 2015). If we additionally train the model to process its own predictions, the model is better able to extrapolate. The third row for every sequence shows the output of the original PredNet fine-tuned for extrapolation. Starting from the trained weights, the model was trained with a loss over 15 time steps, where the actual frame was inputted for the first 10 and then the model’s predictions were used as input to the network for the last 5. For the first 10 time steps, the training loss was calculated on the $E _ { l }$ activations as usual, and for the last 5, it was calculated directly as the mean absolute error with respect to the ground truth frames. Despite eventual blurriness (which might be expected to some extent due to uncertainty), the fine-tuned model captures some key structure in its extrapolations after the tenth time step. For instance, in the first sequence, the model estimates the general shape of an upcoming shadow, despite minimal information in the last seen frame. In the second sequence, the model is able to extrapolate the motion of a car moving to the right. The reader is again encouraged to visit https://coxlab.github.io/prednet/ to view the predictions in video form. Quantitatively, the MSE of the model’s predictions stay well below the trivial solution of copying the last seen frame, as illustrated in Fig 7. The MSE increases fairly linearly from time steps 2-10, even though the model was only trained for up to $t + 5$ prediction. + +![](images/386c934c0c7cd4d4bb7e1f0e1e6977d14b386376dbb66698689a1d26524a0395.jpg) +Figure 7: MSE of PredNet predictions as a function of number of time steps ahead predicted. Model was fine-tuned for up to $t + 5$ prediction. + +# 5.4 ADDITIONAL STEERING ANGLE ANALYSIS + +In Figure 8, we show the steering angle estimation accuracy on the Comma.ai (Biasini et al., 2016) dataset using the representation learned by the PredNet $L _ { 0 }$ model, as a function of the number of frames inputted into the model. The PredNet’s representation at all layers was concatenated (after spatially pooling lower layers to a common spatial resolution) and a fully-connected readout was fit using MSE. For each level of the number of training examples, we average over 10 cross-validation splits. To serve as points of reference, we include results for two static models. The first model is an autoencoder trained on single frame reconstruction with appropriately matching hyperparameters. A fully-connected layer was fit on the autoencoder’s representation to estimate the steering angle in the same fashion as the PredNet. The second model is the default model in the posted Comma.ai code (Biasini et al., 2016), which is a five layer CNN. This model is trained end-to-end to estimate the steering angle given the current frame as input, with a MSE loss. In addition to 25K examples, we trained a version using all of the frames in the Comma dataset (\~396K). For all models, the final weights were chosen at the minimum validation error during training. Given the relatively small number of videos in the dataset compared to the average duration of each video, we used $5 \%$ of each video for validation and testing, chosen as a random continuous chunk, and discarded the 10 frames before and after the chosen segments from the training set. + +![](images/786ee29cf8d282ebc8e20ea4c6a1d601619f455dd6984bab67feeada0044938e.jpg) +Figure 8: Steering angle estimation accuracy as a function of the number of input frames. + +As illustrated in Figure 8, the PredNet’s performance gets better over time, as one might expect, as the model is able to accumulate more information. Interestingly, it performs reasonably well after just one time step, in a regime that is orthogonal to the training procedure of the PredNet where there are no dynamics. Altogether, these results again point to the usefulness of the model in learning underlying latent parameters. + +# 5.5 PREDNET $L _ { a l l }$ NEXT-FRAME PREDICTIONS + +Figures 9 and 10 compare next-frame predictions by the PredNet $L _ { a l l }$ model, trained with a prediction loss on all layers ( $\lambda _ { 0 } = 1$ , $\lambda _ { l > 0 } = 0 . 1 $ ), and the PredNet $L _ { 0 }$ model, trained with a loss only on the lowest layer. At first glance, the difference in predictions seem fairly minor, and indeed, in terms of MSE, the $L _ { a l l }$ model only underperformed the $L _ { 0 }$ version by $3 \%$ and $6 \%$ , respectively, for the rotating faces and CalTech Pedestrian datasets. Upon careful inspection, however, it is apparent that the $L _ { a l l }$ predictions lack some of the finer details of the $L _ { 0 }$ predictions and are more blurry in regions of high variance. For instance, with the rotating faces, the facial features are less defined and with CalTech, details of approaching shadows and cars are less precise. + +![](images/235db1e2a31428ae54787e2a5c30dc4c978774edcf9dd06ef6ac6cecc851ece6.jpg) +time +Figure 9: Next-frame predictions of PredNet $L _ { a l l }$ model on the rotating faces dataset and comparison to $L _ { 0 }$ version. The ”Error ${ \cal L } _ { a l l } { - \cal L } _ { 0 } { } ^ { \cdots }$ visualization shows where the pixel error was smaller for the $L _ { 0 }$ model than the $L _ { a l l }$ model. Green regions correspond to where $L _ { 0 }$ was better and red corresponds to where $L _ { a l l }$ was better. + +![](images/5436ead45eb2f3750a49987e6825e0fe9897d10aa4f0e52240e2823c9aafffd5.jpg) +time +Figure 10: Next-frame predictions of PredNet $L _ { a l l }$ model on the CalTech Pedestrian dataset and comparison to $L _ { 0 }$ version. The ”Error ${ \cal L } _ { a l l } - { \cal L } _ { 0 } { } ^ { }$ visualization shows where the pixel error was smaller for the $L _ { 0 }$ model than the $L _ { a l l }$ model. Green regions correspond to where $L _ { 0 }$ was better and red corresponds to where $L _ { a l l }$ was better. \ No newline at end of file diff --git a/parse/train/B1ewdt9xe/B1ewdt9xe_middle.json b/parse/train/B1ewdt9xe/B1ewdt9xe_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..48dbe13f1f1e6c57c667b61aeef5952c9cccd230 --- /dev/null +++ b/parse/train/B1ewdt9xe/B1ewdt9xe_middle.json @@ -0,0 +1,41808 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 501, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 503, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 503, + 96 + ], + "score": 1.0, + "content": "DEEP PREDICTIVE CODING NETWORKS FOR VIDEO", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 441, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 441, + 118 + ], + "score": 1.0, + "content": "PREDICTION AND UNSUPERVISED LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 313, + 190 + ], + "lines": [ + { + "bbox": [ + 112, + 135, + 314, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 135, + 314, + 147 + ], + "score": 1.0, + "content": "William Lotter, Gabriel Kreiman & David Cox", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 145, + 191, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 145, + 191, + 158 + ], + "score": 1.0, + "content": "Harvard University", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 157, + 233, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 157, + 233, + 168 + ], + "score": 1.0, + "content": "Cambridge, MA 02215, USA", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 168, + 310, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 310, + 181 + ], + "score": 1.0, + "content": "{lotter,davidcox}@fas.harvard.edu", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 179, + 300, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 179, + 300, + 191 + ], + "score": 1.0, + "content": "gabriel.kreiman@tch.harvard.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 276, + 218, + 336, + 234 + ], + "spans": [ + { + "bbox": [ + 276, + 218, + 336, + 234 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 244, + 468, + 464 + ], + "lines": [ + { + "bbox": [ + 142, + 244, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 244, + 469, + 257 + ], + "score": 1.0, + "content": "While great strides have been made in using deep learning algorithms to solve", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 469, + 268 + ], + "score": 1.0, + "content": "supervised learning tasks, the problem of unsupervised learning — leveraging un-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 266, + 469, + 278 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 469, + 278 + ], + "score": 1.0, + "content": "labeled examples to learn about the structure of a domain — remains a difficult", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 278, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 142, + 278, + 470, + 289 + ], + "score": 1.0, + "content": "unsolved challenge. Here, we explore prediction of future frames in a video se-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "score": 1.0, + "content": "quence as an unsupervised learning rule for learning about the structure of the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 300, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 142, + 300, + 469, + 311 + ], + "score": 1.0, + "content": "visual world. We describe a predictive neural network (“PredNet”) architecture", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "that is inspired by the concept of “predictive coding” from the neuroscience lit-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 321, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 469, + 334 + ], + "score": 1.0, + "content": "erature. These networks learn to predict future frames in a video sequence, with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 333, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 142, + 333, + 469, + 345 + ], + "score": 1.0, + "content": "each layer in the network making local predictions and only forwarding deviations", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 343, + 469, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 343, + 469, + 355 + ], + "score": 1.0, + "content": "from those predictions to subsequent network layers. We show that these networks", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 142, + 355, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 355, + 469, + 367 + ], + "score": 1.0, + "content": "are able to robustly learn to predict the movement of synthetic (rendered) objects,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 366, + 469, + 378 + ], + "spans": [ + { + "bbox": [ + 142, + 366, + 469, + 378 + ], + "score": 1.0, + "content": "and that in doing so, the networks learn internal representations that are useful", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 375, + 469, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 375, + 469, + 389 + ], + "score": 1.0, + "content": "for decoding latent object parameters (e.g. pose) that support object recognition", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 469, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 469, + 399 + ], + "score": 1.0, + "content": "with fewer training views. We also show that these networks can scale to com-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 399, + 469, + 410 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 469, + 410 + ], + "score": 1.0, + "content": "plex natural image streams (car-mounted camera videos), capturing key aspects", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 409, + 470, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 409, + 470, + 421 + ], + "score": 1.0, + "content": "of both egocentric movement and the movement of objects in the visual scene,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 420, + 469, + 432 + ], + "spans": [ + { + "bbox": [ + 142, + 420, + 469, + 432 + ], + "score": 1.0, + "content": "and the representation learned in this setting is useful for estimating the steer-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 431, + 470, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 431, + 470, + 444 + ], + "score": 1.0, + "content": "ing angle. Altogether, these results suggest that prediction represents a powerful", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 442, + 469, + 454 + ], + "spans": [ + { + "bbox": [ + 141, + 442, + 469, + 454 + ], + "score": 1.0, + "content": "framework for unsupervised learning, allowing for implicit learning of object and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 455, + 206, + 465 + ], + "spans": [ + { + "bbox": [ + 142, + 455, + 206, + 465 + ], + "score": 1.0, + "content": "scene structure.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 485, + 206, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 208, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 208, + 501 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "Many of the most successful current deep learning architectures for vision rely on supervised learn-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "ing from large sets of labeled training images. While the performance of these networks is un-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "doubtedly impressive, reliance on such large numbers of training examples limits the utility of deep", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "learning in many domains where such datasets are not available. Furthermore, the need for large", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 554, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 566 + ], + "score": 1.0, + "content": "numbers of labeled examples stands at odds with human visual learning, where one or a few views", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 564, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 579 + ], + "score": 1.0, + "content": "of an object is often all that is needed to enable robust recognition of that object across a wide range", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "of different views, lightings and contexts. The development of a representation that facilitates such", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 586, + 406, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 406, + 600 + ], + "score": 1.0, + "content": "abilities, especially in an unsupervised way, is a largely unsolved problem.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "In addition, while computer vision models are typically trained using static images, in the real world,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "visual objects are rarely experienced as disjoint snapshots. Instead, the visual world is alive with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 625, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 625, + 506, + 640 + ], + "score": 1.0, + "content": "movement, driven both by self-motion of the viewer and the movement of objects within the scene.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "Many have suggested that temporal experience with objects as they move and undergo transforma-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "tions can serve as an important signal for learning about the structure of objects (Foldi ¨ ak, 1991; ´", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 658, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 672 + ], + "score": 1.0, + "content": "Softky, 1996; Wiskott & Sejnowski, 2002; George & Hawkins, 2005; Palm, 2012; O’Reilly et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "2014; Agrawal et al., 2015; Goroshin et al., 2015a; Lotter et al., 2015; Mathieu et al., 2016; Srivas-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "tava et al., 2015; Wang & Gupta, 2015; Whitney et al., 2016). For instance, Wiskott and Sejnowski", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "proposed “slow feature analysis” as a framework for exploiting temporal structure in video streams", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 702, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 506, + 716 + ], + "score": 1.0, + "content": "(Wiskott & Sejnowski, 2002). Their approach attempts to build feature representations that extract", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 721, + 455, + 732 + ], + "lines": [ + { + "bbox": [ + 122, + 720, + 457, + 733 + ], + "spans": [ + { + "bbox": [ + 122, + 720, + 457, + 733 + ], + "score": 1.0, + "content": "Code and video examples can be found at: https://coxlab.github.io/prednet/", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 501, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 503, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 503, + 96 + ], + "score": 1.0, + "content": "DEEP PREDICTIVE CODING NETWORKS FOR VIDEO", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 441, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 441, + 118 + ], + "score": 1.0, + "content": "PREDICTION AND UNSUPERVISED LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 113, + 135, + 313, + 190 + ], + "lines": [ + { + "bbox": [ + 112, + 135, + 314, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 135, + 314, + 147 + ], + "score": 1.0, + "content": "William Lotter, Gabriel Kreiman & David Cox", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 145, + 191, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 145, + 191, + 158 + ], + "score": 1.0, + "content": "Harvard University", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 157, + 233, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 157, + 233, + 168 + ], + "score": 1.0, + "content": "Cambridge, MA 02215, USA", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 168, + 310, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 310, + 181 + ], + "score": 1.0, + "content": "{lotter,davidcox}@fas.harvard.edu", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 179, + 300, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 179, + 300, + 191 + ], + "score": 1.0, + "content": "gabriel.kreiman@tch.harvard.edu", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + } + ], + "index": 4, + "bbox_fs": [ + 111, + 135, + 314, + 191 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 219, + 333, + 231 + ], + "lines": [ + { + "bbox": [ + 276, + 218, + 336, + 234 + ], + "spans": [ + { + "bbox": [ + 276, + 218, + 336, + 234 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 244, + 468, + 464 + ], + "lines": [ + { + "bbox": [ + 142, + 244, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 244, + 469, + 257 + ], + "score": 1.0, + "content": "While great strides have been made in using deep learning algorithms to solve", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 469, + 268 + ], + "score": 1.0, + "content": "supervised learning tasks, the problem of unsupervised learning — leveraging un-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 266, + 469, + 278 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 469, + 278 + ], + "score": 1.0, + "content": "labeled examples to learn about the structure of a domain — remains a difficult", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 278, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 142, + 278, + 470, + 289 + ], + "score": 1.0, + "content": "unsolved challenge. Here, we explore prediction of future frames in a video se-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "score": 1.0, + "content": "quence as an unsupervised learning rule for learning about the structure of the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 300, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 142, + 300, + 469, + 311 + ], + "score": 1.0, + "content": "visual world. We describe a predictive neural network (“PredNet”) architecture", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "that is inspired by the concept of “predictive coding” from the neuroscience lit-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 321, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 321, + 469, + 334 + ], + "score": 1.0, + "content": "erature. These networks learn to predict future frames in a video sequence, with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 333, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 142, + 333, + 469, + 345 + ], + "score": 1.0, + "content": "each layer in the network making local predictions and only forwarding deviations", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 343, + 469, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 343, + 469, + 355 + ], + "score": 1.0, + "content": "from those predictions to subsequent network layers. We show that these networks", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 142, + 355, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 355, + 469, + 367 + ], + "score": 1.0, + "content": "are able to robustly learn to predict the movement of synthetic (rendered) objects,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 366, + 469, + 378 + ], + "spans": [ + { + "bbox": [ + 142, + 366, + 469, + 378 + ], + "score": 1.0, + "content": "and that in doing so, the networks learn internal representations that are useful", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 375, + 469, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 375, + 469, + 389 + ], + "score": 1.0, + "content": "for decoding latent object parameters (e.g. pose) that support object recognition", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 469, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 469, + 399 + ], + "score": 1.0, + "content": "with fewer training views. We also show that these networks can scale to com-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 399, + 469, + 410 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 469, + 410 + ], + "score": 1.0, + "content": "plex natural image streams (car-mounted camera videos), capturing key aspects", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 409, + 470, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 409, + 470, + 421 + ], + "score": 1.0, + "content": "of both egocentric movement and the movement of objects in the visual scene,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 420, + 469, + 432 + ], + "spans": [ + { + "bbox": [ + 142, + 420, + 469, + 432 + ], + "score": 1.0, + "content": "and the representation learned in this setting is useful for estimating the steer-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 431, + 470, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 431, + 470, + 444 + ], + "score": 1.0, + "content": "ing angle. Altogether, these results suggest that prediction represents a powerful", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 442, + 469, + 454 + ], + "spans": [ + { + "bbox": [ + 141, + 442, + 469, + 454 + ], + "score": 1.0, + "content": "framework for unsupervised learning, allowing for implicit learning of object and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 455, + 206, + 465 + ], + "spans": [ + { + "bbox": [ + 142, + 455, + 206, + 465 + ], + "score": 1.0, + "content": "scene structure.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 17.5, + "bbox_fs": [ + 141, + 244, + 470, + 465 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 485, + 206, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 208, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 208, + 501 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "Many of the most successful current deep learning architectures for vision rely on supervised learn-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "ing from large sets of labeled training images. While the performance of these networks is un-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "doubtedly impressive, reliance on such large numbers of training examples limits the utility of deep", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "learning in many domains where such datasets are not available. Furthermore, the need for large", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 554, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 566 + ], + "score": 1.0, + "content": "numbers of labeled examples stands at odds with human visual learning, where one or a few views", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 564, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 579 + ], + "score": 1.0, + "content": "of an object is often all that is needed to enable robust recognition of that object across a wide range", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "of different views, lightings and contexts. The development of a representation that facilitates such", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 586, + 406, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 406, + 600 + ], + "score": 1.0, + "content": "abilities, especially in an unsupervised way, is a largely unsolved problem.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 510, + 506, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "In addition, while computer vision models are typically trained using static images, in the real world,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "visual objects are rarely experienced as disjoint snapshots. Instead, the visual world is alive with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 625, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 625, + 506, + 640 + ], + "score": 1.0, + "content": "movement, driven both by self-motion of the viewer and the movement of objects within the scene.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "Many have suggested that temporal experience with objects as they move and undergo transforma-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "tions can serve as an important signal for learning about the structure of objects (Foldi ¨ ak, 1991; ´", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 658, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 672 + ], + "score": 1.0, + "content": "Softky, 1996; Wiskott & Sejnowski, 2002; George & Hawkins, 2005; Palm, 2012; O’Reilly et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "2014; Agrawal et al., 2015; Goroshin et al., 2015a; Lotter et al., 2015; Mathieu et al., 2016; Srivas-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "tava et al., 2015; Wang & Gupta, 2015; Whitney et al., 2016). For instance, Wiskott and Sejnowski", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "proposed “slow feature analysis” as a framework for exploiting temporal structure in video streams", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 702, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 506, + 716 + ], + "score": 1.0, + "content": "(Wiskott & Sejnowski, 2002). Their approach attempts to build feature representations that extract", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "slowly-varying parameters, such as object identity, from parameters that produce fast changes in the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "image, such as movement of the object. While approaches that rely on temporal coherence have", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "arguably not yet yielded representations as powerful as those learned by supervised methods, they", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "nonetheless point to the potential of learning useful representations from video (Mohabi et al., 2009;", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 478, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 478, + 140 + ], + "score": 1.0, + "content": "Sun et al., 2014; Goroshin et al., 2015a; Maltoni & Lomonaco, 2015; Wang & Gupta, 2015).", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 41.5, + "bbox_fs": [ + 104, + 604, + 506, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "slowly-varying parameters, such as object identity, from parameters that produce fast changes in the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "image, such as movement of the object. While approaches that rely on temporal coherence have", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "arguably not yet yielded representations as powerful as those learned by supervised methods, they", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "nonetheless point to the potential of learning useful representations from video (Mohabi et al., 2009;", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 478, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 478, + 140 + ], + "score": 1.0, + "content": "Sun et al., 2014; Goroshin et al., 2015a; Maltoni & Lomonaco, 2015; Wang & Gupta, 2015).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "Here, we explore another potential principle for exploiting video for unsupervised learning: pre-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 504, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 504, + 166 + ], + "score": 1.0, + "content": "diction of future image frames (Softky, 1996; Palm, 2012; O’Reilly et al., 2014; Goroshin et al.,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "2015b; Srivastava et al., 2015; Mathieu et al., 2016; Patraucean et al., 2015; Finn et al., 2016; Von-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "drick et al., 2016). A key insight here is that in order to be able to predict how the visual world", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "will change over time, an agent must have at least some implicit model of object structure and the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "possible transformations objects can undergo. To this end, we have designed a neural network archi-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "tecture, which we informally call a “PredNet,” that attempts to continually predict the appearance", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 504, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 504, + 234 + ], + "score": 1.0, + "content": "of future video frames, using a deep, recurrent convolutional network with both bottom-up and top-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "down connections. Our work here builds on previous work in next-frame video prediction (Ranzato", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 504, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 504, + 253 + ], + "score": 1.0, + "content": "et al., 2014; Michalski et al., 2014; Srivastava et al., 2015; Mathieu et al., 2016; Lotter et al., 2015;", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 264 + ], + "score": 1.0, + "content": "Patraucean et al., 2015; Oh et al., 2015; Finn et al., 2016; Xue et al., 2016; Vondrick et al., 2016;", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 504, + 277 + ], + "score": 1.0, + "content": "Brabandere et al., 2016), but we take particular inspiration from the concept of “predictive coding”", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "score": 1.0, + "content": "from the neuroscience literature (Rao & Ballard, 1999; Rao & Sejnowski, 2000; Lee & Mumford,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "2003; Friston, 2005; Summerfield et al., 2006; Egner et al., 2010; Bastos et al., 2012; Spratling,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "2012; Chalasani & Principe, 2013; Clark, 2013; O’Reilly et al., 2014; Kanai et al., 2015). Predictive", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "coding posits that the brain is continually making predictions of incoming sensory stimuli (Rao &", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 318, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 104, + 318, + 506, + 332 + ], + "score": 1.0, + "content": "Ballard, 1999; Friston, 2005). Top-down (and perhaps lateral) connections convey these predictions,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "which are compared against actual observations to generate an error signal. The error signal is then", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 341, + 446, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 446, + 354 + ], + "score": 1.0, + "content": "propagated back up the hierarchy, eventually leading to an update of the predictions.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "We demonstrate the effectiveness of our model for both synthetic sequences, where we have access", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "score": 1.0, + "content": "to the underlying generative model and can investigate what the model learns, as well as natural", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "videos. Consistent with the idea that prediction requires knowledge of object structure, we find", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "that these networks successfully learn internal representations that are well-suited to subsequent", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "recognition and decoding of latent object parameters (e.g. identity, view, rotation speed, etc.). We", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 410, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 427 + ], + "score": 1.0, + "content": "also find that our architecture can scale effectively to natural image sequences, by training using", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 424, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 435 + ], + "score": 1.0, + "content": "car-mounted camera videos. The network is able to successfully learn to predict both the movement", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "of the camera and the movement of objects in the camera’s view. Again supporting the notion", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 444, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 459 + ], + "score": 1.0, + "content": "of prediction as an unsupervised learning rule, the model’s learned representation in this setting", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 457, + 298, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 298, + 469 + ], + "score": 1.0, + "content": "supports decoding of the current steering angle.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 28.5 + }, + { + "type": "image", + "bbox": [ + 152, + 486, + 452, + 660 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 152, + 486, + 452, + 660 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 152, + 486, + 452, + 660 + ], + "spans": [ + { + "bbox": [ + 152, + 486, + 452, + 660 + ], + "score": 0.973, + "type": "image", + "image_path": "96a1440b9f8c8bf098c1b691d0dce930f9e50a5bd99be7f869e6ae09b81b7762.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 152, + 486, + 452, + 544.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 152, + 544.0, + 452, + 602.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 152, + 602.0, + 452, + 660.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 672, + 506, + 730 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 673, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 505, + 686 + ], + "score": 1.0, + "content": "Figure 1: Predictive Coding Network (PredNet). Left: Illustration of information flow within two", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 683, + 506, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 316, + 697 + ], + "score": 1.0, + "content": "layers. Each layer consists of representation neurons", + "type": "text" + }, + { + "bbox": [ + 316, + 684, + 333, + 695 + ], + "score": 0.8, + "content": "( R _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 683, + 506, + 697 + ], + "score": 1.0, + "content": ", which output a layer-specific prediction at", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 695, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 106, + 697, + 165, + 709 + ], + "score": 1.0, + "content": "each time step", + "type": "text" + }, + { + "bbox": [ + 165, + 695, + 183, + 708 + ], + "score": 0.87, + "content": "( \\hat { A } _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 697, + 326, + 709 + ], + "score": 1.0, + "content": ", which is compared against a target", + "type": "text" + }, + { + "bbox": [ + 326, + 697, + 344, + 708 + ], + "score": 0.82, + "content": "( A _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 697, + 505, + 709 + ], + "score": 1.0, + "content": "(Bengio, 2014) to produce an error term", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 708, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 107, + 708, + 124, + 719 + ], + "score": 0.84, + "content": "( E _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 708, + 505, + 721 + ], + "score": 1.0, + "content": ", which is then propagated laterally and vertically in the network. Right: Module operations for", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 719, + 207, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 207, + 731 + ], + "score": 1.0, + "content": "case of video sequences.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39 + } + ], + "index": 37.0 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 138 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 140 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "Here, we explore another potential principle for exploiting video for unsupervised learning: pre-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 155, + 504, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 504, + 166 + ], + "score": 1.0, + "content": "diction of future image frames (Softky, 1996; Palm, 2012; O’Reilly et al., 2014; Goroshin et al.,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "2015b; Srivastava et al., 2015; Mathieu et al., 2016; Patraucean et al., 2015; Finn et al., 2016; Von-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "drick et al., 2016). A key insight here is that in order to be able to predict how the visual world", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "will change over time, an agent must have at least some implicit model of object structure and the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "possible transformations objects can undergo. To this end, we have designed a neural network archi-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "tecture, which we informally call a “PredNet,” that attempts to continually predict the appearance", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 218, + 504, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 504, + 234 + ], + "score": 1.0, + "content": "of future video frames, using a deep, recurrent convolutional network with both bottom-up and top-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "down connections. Our work here builds on previous work in next-frame video prediction (Ranzato", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 504, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 504, + 253 + ], + "score": 1.0, + "content": "et al., 2014; Michalski et al., 2014; Srivastava et al., 2015; Mathieu et al., 2016; Lotter et al., 2015;", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 253, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 264 + ], + "score": 1.0, + "content": "Patraucean et al., 2015; Oh et al., 2015; Finn et al., 2016; Xue et al., 2016; Vondrick et al., 2016;", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 504, + 277 + ], + "score": 1.0, + "content": "Brabandere et al., 2016), but we take particular inspiration from the concept of “predictive coding”", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "score": 1.0, + "content": "from the neuroscience literature (Rao & Ballard, 1999; Rao & Sejnowski, 2000; Lee & Mumford,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "2003; Friston, 2005; Summerfield et al., 2006; Egner et al., 2010; Bastos et al., 2012; Spratling,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "2012; Chalasani & Principe, 2013; Clark, 2013; O’Reilly et al., 2014; Kanai et al., 2015). Predictive", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "coding posits that the brain is continually making predictions of incoming sensory stimuli (Rao &", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 318, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 104, + 318, + 506, + 332 + ], + "score": 1.0, + "content": "Ballard, 1999; Friston, 2005). Top-down (and perhaps lateral) connections convey these predictions,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "which are compared against actual observations to generate an error signal. The error signal is then", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 341, + 446, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 446, + 354 + ], + "score": 1.0, + "content": "propagated back up the hierarchy, eventually leading to an update of the predictions.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 142, + 506, + 354 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "We demonstrate the effectiveness of our model for both synthetic sequences, where we have access", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "score": 1.0, + "content": "to the underlying generative model and can investigate what the model learns, as well as natural", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "videos. Consistent with the idea that prediction requires knowledge of object structure, we find", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "that these networks successfully learn internal representations that are well-suited to subsequent", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 414 + ], + "score": 1.0, + "content": "recognition and decoding of latent object parameters (e.g. identity, view, rotation speed, etc.). We", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 410, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 427 + ], + "score": 1.0, + "content": "also find that our architecture can scale effectively to natural image sequences, by training using", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 424, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 435 + ], + "score": 1.0, + "content": "car-mounted camera videos. The network is able to successfully learn to predict both the movement", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "of the camera and the movement of objects in the camera’s view. Again supporting the notion", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 444, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 459 + ], + "score": 1.0, + "content": "of prediction as an unsupervised learning rule, the model’s learned representation in this setting", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 457, + 298, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 298, + 469 + ], + "score": 1.0, + "content": "supports decoding of the current steering angle.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 357, + 506, + 469 + ] + }, + { + "type": "image", + "bbox": [ + 152, + 486, + 452, + 660 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 152, + 486, + 452, + 660 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 152, + 486, + 452, + 660 + ], + "spans": [ + { + "bbox": [ + 152, + 486, + 452, + 660 + ], + "score": 0.973, + "type": "image", + "image_path": "96a1440b9f8c8bf098c1b691d0dce930f9e50a5bd99be7f869e6ae09b81b7762.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 152, + 486, + 452, + 544.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 152, + 544.0, + 452, + 602.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 152, + 602.0, + 452, + 660.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 672, + 506, + 730 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 673, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 505, + 686 + ], + "score": 1.0, + "content": "Figure 1: Predictive Coding Network (PredNet). Left: Illustration of information flow within two", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 683, + 506, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 316, + 697 + ], + "score": 1.0, + "content": "layers. Each layer consists of representation neurons", + "type": "text" + }, + { + "bbox": [ + 316, + 684, + 333, + 695 + ], + "score": 0.8, + "content": "( R _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 683, + 506, + 697 + ], + "score": 1.0, + "content": ", which output a layer-specific prediction at", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 695, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 106, + 697, + 165, + 709 + ], + "score": 1.0, + "content": "each time step", + "type": "text" + }, + { + "bbox": [ + 165, + 695, + 183, + 708 + ], + "score": 0.87, + "content": "( \\hat { A } _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 697, + 326, + 709 + ], + "score": 1.0, + "content": ", which is compared against a target", + "type": "text" + }, + { + "bbox": [ + 326, + 697, + 344, + 708 + ], + "score": 0.82, + "content": "( A _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 697, + 505, + 709 + ], + "score": 1.0, + "content": "(Bengio, 2014) to produce an error term", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 708, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 107, + 708, + 124, + 719 + ], + "score": 0.84, + "content": "( E _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 708, + 505, + 721 + ], + "score": 1.0, + "content": ", which is then propagated laterally and vertically in the network. Right: Module operations for", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 719, + 207, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 207, + 731 + ], + "score": 1.0, + "content": "case of video sequences.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39 + } + ], + "index": 37.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 243, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 245, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 245, + 96 + ], + "score": 1.0, + "content": "2 THE PREDNET MODEL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 105, + 505, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 119 + ], + "score": 1.0, + "content": "The PredNet architecture is diagrammed in Figure 1. The network consists of a series of repeating", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 117, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 505, + 128 + ], + "score": 1.0, + "content": "stacked modules that attempt to make local predictions of the input to the module, which is then", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "subtracted from the actual input and passed along to the next layer. Briefly, each module of the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 376, + 151 + ], + "score": 1.0, + "content": "network consists of four basic parts: an input convolutional layer", + "type": "text" + }, + { + "bbox": [ + 376, + 138, + 394, + 150 + ], + "score": 0.85, + "content": "( A _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 138, + 506, + 151 + ], + "score": 1.0, + "content": ", a recurrent representation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 149, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 129, + 164 + ], + "score": 1.0, + "content": "layer", + "type": "text" + }, + { + "bbox": [ + 129, + 151, + 147, + 162 + ], + "score": 0.81, + "content": "( R _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 151, + 224, + 164 + ], + "score": 1.0, + "content": ", a prediction layer", + "type": "text" + }, + { + "bbox": [ + 224, + 149, + 242, + 162 + ], + "score": 0.88, + "content": "( \\hat { \\hat { A } } _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 151, + 356, + 164 + ], + "score": 1.0, + "content": ", and an error representation", + "type": "text" + }, + { + "bbox": [ + 356, + 151, + 374, + 162 + ], + "score": 0.83, + "content": "( E _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 151, + 480, + 164 + ], + "score": 1.0, + "content": ". The representation layer,", + "type": "text" + }, + { + "bbox": [ + 480, + 151, + 492, + 162 + ], + "score": 0.87, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 151, + 505, + 164 + ], + "score": 1.0, + "content": ", is", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 357, + 176 + ], + "score": 1.0, + "content": "a recurrent convolutional network that generates a prediction,", + "type": "text" + }, + { + "bbox": [ + 357, + 162, + 369, + 174 + ], + "score": 0.89, + "content": "\\hat { A } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 164, + 470, + 176 + ], + "score": 1.0, + "content": ", of what the layer input,", + "type": "text" + }, + { + "bbox": [ + 471, + 164, + 483, + 174 + ], + "score": 0.88, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 164, + 505, + 176 + ], + "score": 1.0, + "content": ", will", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 174, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 374, + 189 + ], + "score": 1.0, + "content": "be on the next frame. The network takes the difference between", + "type": "text" + }, + { + "bbox": [ + 374, + 176, + 386, + 187 + ], + "score": 0.87, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 176, + 406, + 189 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 406, + 174, + 418, + 187 + ], + "score": 0.9, + "content": "\\hat { A } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "and outputs an error", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 167, + 199 + ], + "score": 1.0, + "content": "representation,", + "type": "text" + }, + { + "bbox": [ + 168, + 187, + 179, + 198 + ], + "score": 0.86, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 187, + 505, + 199 + ], + "score": 1.0, + "content": ", which is split into separate rectified positive and negative error populations. The", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 130, + 212 + ], + "score": 1.0, + "content": "error,", + "type": "text" + }, + { + "bbox": [ + 131, + 198, + 142, + 209 + ], + "score": 0.87, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 197, + 506, + 212 + ], + "score": 1.0, + "content": ", is then passed forward through a convolutional layer to become the input to the next layer", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 107, + 209, + 135, + 221 + ], + "score": 0.9, + "content": "( A _ { l + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 208, + 259, + 222 + ], + "score": 1.0, + "content": ". The recurrent prediction layer", + "type": "text" + }, + { + "bbox": [ + 259, + 209, + 271, + 220 + ], + "score": 0.87, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 208, + 405, + 222 + ], + "score": 1.0, + "content": "receives a copy of the error signal", + "type": "text" + }, + { + "bbox": [ + 406, + 209, + 417, + 220 + ], + "score": 0.86, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 208, + 505, + 222 + ], + "score": 1.0, + "content": ", along with top-down", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 376, + 232 + ], + "score": 1.0, + "content": "input from the representation layer of the next level of the network", + "type": "text" + }, + { + "bbox": [ + 376, + 220, + 404, + 232 + ], + "score": 0.91, + "content": "( R _ { l + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 219, + 506, + 232 + ], + "score": 1.0, + "content": ". The organization of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 447, + 243 + ], + "score": 1.0, + "content": "network is such that on the first time step of operation, the “right” side of the network (", + "type": "text" + }, + { + "bbox": [ + 447, + 231, + 458, + 242 + ], + "score": 0.85, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 231, + 482, + 243 + ], + "score": 1.0, + "content": "’s and", + "type": "text" + }, + { + "bbox": [ + 482, + 231, + 493, + 242 + ], + "score": 0.86, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "’s)", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "is equivalent to a standard deep convolutional network. Meanwhile, the “left” side of the network", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 124, + 266 + ], + "score": 1.0, + "content": "(the", + "type": "text" + }, + { + "bbox": [ + 125, + 253, + 136, + 264 + ], + "score": 0.85, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "’s) is equivalent to a generative deconvolutional network with local recurrence at each stage.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "The architecture described here is inspired by that originally proposed by (Rao & Ballard, 1999), but", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 274, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 288 + ], + "score": 1.0, + "content": "is formulated in a modern deep learning framework and trained end-to-end using gradient descent,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "with a loss function implicitly embedded in the network as the firing rates of the error neurons. Our", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "work also shares motivation with the Deep Predictive Coding Networks of Chalasani & Principe", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "(2013); however, their framework is based upon sparse coding and a linear dynamical system with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "greedy layer-wise training, whereas ours is rooted in convolutional and recurrent neural networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 328, + 199, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 199, + 343 + ], + "score": 1.0, + "content": "trained with backprop.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 346, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 359 + ], + "score": 1.0, + "content": "While the architecture is general with respect to the kinds of data it models, here we focus on image", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 358, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 336, + 370 + ], + "score": 1.0, + "content": "sequence (video) data. Consider a sequence of images,", + "type": "text" + }, + { + "bbox": [ + 336, + 359, + 347, + 369 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 358, + 506, + 370 + ], + "score": 1.0, + "content": ". The target for the lowest layer is set", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 256, + 381 + ], + "score": 1.0, + "content": "to the the actual sequence itself, i.e.", + "type": "text" + }, + { + "bbox": [ + 257, + 368, + 306, + 380 + ], + "score": 0.93, + "content": "\\bar { A } _ { 0 } ^ { t } = x _ { t } \\forall t", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 368, + 432, + 381 + ], + "score": 1.0, + "content": ". The targets for higher layers,", + "type": "text" + }, + { + "bbox": [ + 433, + 368, + 445, + 380 + ], + "score": 0.89, + "content": "A _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 368, + 461, + 381 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 461, + 369, + 486, + 379 + ], + "score": 0.89, + "content": "l > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 368, + 506, + 381 + ], + "score": 1.0, + "content": ", are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 378, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 392, + 394 + ], + "score": 1.0, + "content": "computed by a convolution over the error units from the layer below,", + "type": "text" + }, + { + "bbox": [ + 392, + 379, + 413, + 392 + ], + "score": 0.92, + "content": "E _ { l - 1 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 378, + 506, + 394 + ], + "score": 1.0, + "content": ", followed by rectified", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "linear unit (ReLU) activation and max-pooling. For the representation neurons, we specifically", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 399, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 415 + ], + "score": 1.0, + "content": "use convolutional LSTM units (Hochreiter & Schmidhuber, 1997; Shi et al., 2015). In our setting,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 103, + 407, + 507, + 429 + ], + "spans": [ + { + "bbox": [ + 103, + 407, + 122, + 429 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 122, + 412, + 134, + 425 + ], + "score": 0.9, + "content": "R _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 407, + 286, + 429 + ], + "score": 1.0, + "content": "hidden state is updated according to", + "type": "text" + }, + { + "bbox": [ + 287, + 411, + 309, + 425 + ], + "score": 0.85, + "content": "R _ { l } ^ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 407, + 314, + 429 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 314, + 411, + 337, + 425 + ], + "score": 0.88, + "content": "E _ { l } ^ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 407, + 385, + 429 + ], + "score": 1.0, + "content": ", as well as", + "type": "text" + }, + { + "bbox": [ + 386, + 412, + 407, + 425 + ], + "score": 0.92, + "content": "R _ { l + 1 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 407, + 507, + 429 + ], + "score": 1.0, + "content": ", which is first spatially", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "score": 1.0, + "content": "upsampled (nearest-neighbor), due to the pooling present in the feedforward path. The predictions,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 434, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 119, + 448 + ], + "score": 0.9, + "content": "\\hat { A } _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 434, + 284, + 450 + ], + "score": 1.0, + "content": "are made through a convolution of the", + "type": "text" + }, + { + "bbox": [ + 285, + 435, + 297, + 448 + ], + "score": 0.89, + "content": "R _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 434, + 506, + 450 + ], + "score": 1.0, + "content": "stack followed by a ReLU non-linearity. For the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 160, + 462 + ], + "score": 1.0, + "content": "lowest layer,", + "type": "text" + }, + { + "bbox": [ + 161, + 447, + 173, + 461 + ], + "score": 0.88, + "content": "\\hat { A } _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "is also passed through a saturating non-linearity set at the maximum pixel value:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 458, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 107, + 460, + 245, + 472 + ], + "score": 0.92, + "content": "\\operatorname { S a t L U } ( x ; p _ { m a x } ) : = \\operatorname* { m i n } ( p _ { m a x } , x )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 458, + 358, + 474 + ], + "score": 1.0, + "content": ". Finally, the error response,", + "type": "text" + }, + { + "bbox": [ + 358, + 460, + 370, + 472 + ], + "score": 0.89, + "content": "E _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 458, + 506, + 474 + ], + "score": 1.0, + "content": ", is calculated from the difference", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 142, + 486 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 143, + 472, + 155, + 485 + ], + "score": 0.9, + "content": "\\hat { A } _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 473, + 173, + 486 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 174, + 473, + 186, + 485 + ], + "score": 0.89, + "content": "A _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "and is split into ReLU-activated positive and negative prediction errors, which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "are concatenated along the feature dimension. As discussed in (Rao & Ballard, 1999), although not", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 494, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 509 + ], + "score": 1.0, + "content": "explicit in their model, the separate error populations are analogous to the existence of on-center,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 505, + 410, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 410, + 519 + ], + "score": 1.0, + "content": "off-surround and off-center, on-surround neurons early in the visual system.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 589 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "The full set of update rules are listed in Equations (1) to (4). The model is trained to minimize", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 534, + 504, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 504, + 545 + ], + "score": 1.0, + "content": "the weighted sum of the activity of the error units. Explicitly, the training loss is formalized in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 280, + 558 + ], + "score": 1.0, + "content": "Equation 5 with weighting factors by time,", + "type": "text" + }, + { + "bbox": [ + 280, + 545, + 290, + 556 + ], + "score": 0.87, + "content": "\\lambda _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 545, + 335, + 558 + ], + "score": 1.0, + "content": ", and layer,", + "type": "text" + }, + { + "bbox": [ + 336, + 546, + 346, + 556 + ], + "score": 0.86, + "content": "\\lambda _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 545, + 393, + 558 + ], + "score": 1.0, + "content": ", and where", + "type": "text" + }, + { + "bbox": [ + 393, + 547, + 404, + 556 + ], + "score": 0.83, + "content": "n _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "is the number of units in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "the lth layer. With error units consisting of subtraction followed by ReLU activation, the loss at each", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "score": 1.0, + "content": "layer is equivalent to an L1 error. Although not explored here, other error unit implementations,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 578, + 466, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 466, + 590 + ], + "score": 1.0, + "content": "potentially even probabilistic or adversarial (Goodfellow et al., 2014), could also be used.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 606, + 422, + 685 + ], + "lines": [ + { + "bbox": [ + 203, + 606, + 422, + 685 + ], + "spans": [ + { + "bbox": [ + 203, + 606, + 422, + 685 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { A _ { l } ^ { t } = \\bigg \\{ x _ { t } } & { \\mathrm { i f } l = 0 } \\\\ & { \\mathbf { M } _ { \\mathrm { A X P o o L } } ( \\operatorname { R E L U } ( \\operatorname { C o N v } ( E _ { l - 1 } ^ { t } ) ) ) } & { l > 0 } \\\\ & { \\hat { A } _ { l } ^ { t } = \\operatorname { R E L U } ( \\operatorname { C o N v } ( R _ { l } ^ { t } ) ) } \\\\ & { E _ { l } ^ { t } = [ \\operatorname { R E L U } ( A _ { l } ^ { t } - \\hat { A } _ { l } ^ { t } ) ; \\operatorname { R E L U } ( \\hat { A } _ { l } ^ { t } - A _ { l } ^ { t } ) ] } \\\\ & { R _ { l } ^ { t } = \\operatorname { C o N v L S T M } ( E _ { l } ^ { t - 1 } , R _ { l } ^ { t - 1 } , \\operatorname { U P S A M P L E } ( R _ { l + 1 } ^ { t } ) ) } \\end{array}", + "type": "interline_equation", + "image_path": "d4efd34db2f2780addbfd9d2817ec3f2ce20b4d0c18a15b42c819a20676e5669.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 203, + 606, + 422, + 621.8 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 203, + 621.8, + 422, + 637.5999999999999 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 203, + 637.5999999999999, + 422, + 653.3999999999999 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 203, + 653.3999999999999, + 422, + 669.1999999999998 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 203, + 669.1999999999998, + 422, + 684.9999999999998 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 700, + 315, + 730 + ], + "lines": [ + { + "bbox": [ + 189, + 700, + 315, + 730 + ], + "spans": [ + { + "bbox": [ + 189, + 700, + 315, + 730 + ], + "score": 0.92, + "content": "L _ { t r a i n } = \\sum _ { t } \\lambda _ { t } \\sum _ { l } \\frac { \\lambda _ { l } } { n _ { l } } \\sum _ { n _ { l } } E _ { l } ^ { t }", + "type": "interline_equation", + "image_path": "118394227ec7aaffb1e86b56b4dea6f1278381aa47137052f4ade323fa319af4.jpg" + } + ] + } + ], + "index": 48.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 700, + 315, + 715.0 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 189, + 715.0, + 315, + 730.0 + ], + "spans": [], + "index": 49 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 243, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 245, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 245, + 96 + ], + "score": 1.0, + "content": "2 THE PREDNET MODEL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 105, + 505, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 119 + ], + "score": 1.0, + "content": "The PredNet architecture is diagrammed in Figure 1. The network consists of a series of repeating", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 117, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 505, + 128 + ], + "score": 1.0, + "content": "stacked modules that attempt to make local predictions of the input to the module, which is then", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "subtracted from the actual input and passed along to the next layer. Briefly, each module of the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 376, + 151 + ], + "score": 1.0, + "content": "network consists of four basic parts: an input convolutional layer", + "type": "text" + }, + { + "bbox": [ + 376, + 138, + 394, + 150 + ], + "score": 0.85, + "content": "( A _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 138, + 506, + 151 + ], + "score": 1.0, + "content": ", a recurrent representation", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 149, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 129, + 164 + ], + "score": 1.0, + "content": "layer", + "type": "text" + }, + { + "bbox": [ + 129, + 151, + 147, + 162 + ], + "score": 0.81, + "content": "( R _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 151, + 224, + 164 + ], + "score": 1.0, + "content": ", a prediction layer", + "type": "text" + }, + { + "bbox": [ + 224, + 149, + 242, + 162 + ], + "score": 0.88, + "content": "( \\hat { \\hat { A } } _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 151, + 356, + 164 + ], + "score": 1.0, + "content": ", and an error representation", + "type": "text" + }, + { + "bbox": [ + 356, + 151, + 374, + 162 + ], + "score": 0.83, + "content": "( E _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 151, + 480, + 164 + ], + "score": 1.0, + "content": ". The representation layer,", + "type": "text" + }, + { + "bbox": [ + 480, + 151, + 492, + 162 + ], + "score": 0.87, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 151, + 505, + 164 + ], + "score": 1.0, + "content": ", is", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 357, + 176 + ], + "score": 1.0, + "content": "a recurrent convolutional network that generates a prediction,", + "type": "text" + }, + { + "bbox": [ + 357, + 162, + 369, + 174 + ], + "score": 0.89, + "content": "\\hat { A } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 164, + 470, + 176 + ], + "score": 1.0, + "content": ", of what the layer input,", + "type": "text" + }, + { + "bbox": [ + 471, + 164, + 483, + 174 + ], + "score": 0.88, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 164, + 505, + 176 + ], + "score": 1.0, + "content": ", will", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 174, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 374, + 189 + ], + "score": 1.0, + "content": "be on the next frame. The network takes the difference between", + "type": "text" + }, + { + "bbox": [ + 374, + 176, + 386, + 187 + ], + "score": 0.87, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 176, + 406, + 189 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 406, + 174, + 418, + 187 + ], + "score": 0.9, + "content": "\\hat { A } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "and outputs an error", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 167, + 199 + ], + "score": 1.0, + "content": "representation,", + "type": "text" + }, + { + "bbox": [ + 168, + 187, + 179, + 198 + ], + "score": 0.86, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 187, + 505, + 199 + ], + "score": 1.0, + "content": ", which is split into separate rectified positive and negative error populations. The", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 130, + 212 + ], + "score": 1.0, + "content": "error,", + "type": "text" + }, + { + "bbox": [ + 131, + 198, + 142, + 209 + ], + "score": 0.87, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 197, + 506, + 212 + ], + "score": 1.0, + "content": ", is then passed forward through a convolutional layer to become the input to the next layer", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 107, + 209, + 135, + 221 + ], + "score": 0.9, + "content": "( A _ { l + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 208, + 259, + 222 + ], + "score": 1.0, + "content": ". The recurrent prediction layer", + "type": "text" + }, + { + "bbox": [ + 259, + 209, + 271, + 220 + ], + "score": 0.87, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 208, + 405, + 222 + ], + "score": 1.0, + "content": "receives a copy of the error signal", + "type": "text" + }, + { + "bbox": [ + 406, + 209, + 417, + 220 + ], + "score": 0.86, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 208, + 505, + 222 + ], + "score": 1.0, + "content": ", along with top-down", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 376, + 232 + ], + "score": 1.0, + "content": "input from the representation layer of the next level of the network", + "type": "text" + }, + { + "bbox": [ + 376, + 220, + 404, + 232 + ], + "score": 0.91, + "content": "( R _ { l + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 219, + 506, + 232 + ], + "score": 1.0, + "content": ". The organization of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 447, + 243 + ], + "score": 1.0, + "content": "network is such that on the first time step of operation, the “right” side of the network (", + "type": "text" + }, + { + "bbox": [ + 447, + 231, + 458, + 242 + ], + "score": 0.85, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 231, + 482, + 243 + ], + "score": 1.0, + "content": "’s and", + "type": "text" + }, + { + "bbox": [ + 482, + 231, + 493, + 242 + ], + "score": 0.86, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 231, + 505, + 243 + ], + "score": 1.0, + "content": "’s)", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "is equivalent to a standard deep convolutional network. Meanwhile, the “left” side of the network", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 124, + 266 + ], + "score": 1.0, + "content": "(the", + "type": "text" + }, + { + "bbox": [ + 125, + 253, + 136, + 264 + ], + "score": 0.85, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "’s) is equivalent to a generative deconvolutional network with local recurrence at each stage.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 276 + ], + "score": 1.0, + "content": "The architecture described here is inspired by that originally proposed by (Rao & Ballard, 1999), but", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 274, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 288 + ], + "score": 1.0, + "content": "is formulated in a modern deep learning framework and trained end-to-end using gradient descent,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "with a loss function implicitly embedded in the network as the firing rates of the error neurons. Our", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "work also shares motivation with the Deep Predictive Coding Networks of Chalasani & Principe", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "(2013); however, their framework is based upon sparse coding and a linear dynamical system with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "greedy layer-wise training, whereas ours is rooted in convolutional and recurrent neural networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 328, + 199, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 199, + 343 + ], + "score": 1.0, + "content": "trained with backprop.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 104, + 506, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 346, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 359 + ], + "score": 1.0, + "content": "While the architecture is general with respect to the kinds of data it models, here we focus on image", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 358, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 336, + 370 + ], + "score": 1.0, + "content": "sequence (video) data. Consider a sequence of images,", + "type": "text" + }, + { + "bbox": [ + 336, + 359, + 347, + 369 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 358, + 506, + 370 + ], + "score": 1.0, + "content": ". The target for the lowest layer is set", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 256, + 381 + ], + "score": 1.0, + "content": "to the the actual sequence itself, i.e.", + "type": "text" + }, + { + "bbox": [ + 257, + 368, + 306, + 380 + ], + "score": 0.93, + "content": "\\bar { A } _ { 0 } ^ { t } = x _ { t } \\forall t", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 368, + 432, + 381 + ], + "score": 1.0, + "content": ". The targets for higher layers,", + "type": "text" + }, + { + "bbox": [ + 433, + 368, + 445, + 380 + ], + "score": 0.89, + "content": "A _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 368, + 461, + 381 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 461, + 369, + 486, + 379 + ], + "score": 0.89, + "content": "l > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 368, + 506, + 381 + ], + "score": 1.0, + "content": ", are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 378, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 392, + 394 + ], + "score": 1.0, + "content": "computed by a convolution over the error units from the layer below,", + "type": "text" + }, + { + "bbox": [ + 392, + 379, + 413, + 392 + ], + "score": 0.92, + "content": "E _ { l - 1 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 378, + 506, + 394 + ], + "score": 1.0, + "content": ", followed by rectified", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "linear unit (ReLU) activation and max-pooling. For the representation neurons, we specifically", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 399, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 415 + ], + "score": 1.0, + "content": "use convolutional LSTM units (Hochreiter & Schmidhuber, 1997; Shi et al., 2015). In our setting,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 103, + 407, + 507, + 429 + ], + "spans": [ + { + "bbox": [ + 103, + 407, + 122, + 429 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 122, + 412, + 134, + 425 + ], + "score": 0.9, + "content": "R _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 407, + 286, + 429 + ], + "score": 1.0, + "content": "hidden state is updated according to", + "type": "text" + }, + { + "bbox": [ + 287, + 411, + 309, + 425 + ], + "score": 0.85, + "content": "R _ { l } ^ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 407, + 314, + 429 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 314, + 411, + 337, + 425 + ], + "score": 0.88, + "content": "E _ { l } ^ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 407, + 385, + 429 + ], + "score": 1.0, + "content": ", as well as", + "type": "text" + }, + { + "bbox": [ + 386, + 412, + 407, + 425 + ], + "score": 0.92, + "content": "R _ { l + 1 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 407, + 507, + 429 + ], + "score": 1.0, + "content": ", which is first spatially", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 437 + ], + "score": 1.0, + "content": "upsampled (nearest-neighbor), due to the pooling present in the feedforward path. The predictions,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 434, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 119, + 448 + ], + "score": 0.9, + "content": "\\hat { A } _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 434, + 284, + 450 + ], + "score": 1.0, + "content": "are made through a convolution of the", + "type": "text" + }, + { + "bbox": [ + 285, + 435, + 297, + 448 + ], + "score": 0.89, + "content": "R _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 434, + 506, + 450 + ], + "score": 1.0, + "content": "stack followed by a ReLU non-linearity. For the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 160, + 462 + ], + "score": 1.0, + "content": "lowest layer,", + "type": "text" + }, + { + "bbox": [ + 161, + 447, + 173, + 461 + ], + "score": 0.88, + "content": "\\hat { A } _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "is also passed through a saturating non-linearity set at the maximum pixel value:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 458, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 107, + 460, + 245, + 472 + ], + "score": 0.92, + "content": "\\operatorname { S a t L U } ( x ; p _ { m a x } ) : = \\operatorname* { m i n } ( p _ { m a x } , x )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 458, + 358, + 474 + ], + "score": 1.0, + "content": ". Finally, the error response,", + "type": "text" + }, + { + "bbox": [ + 358, + 460, + 370, + 472 + ], + "score": 0.89, + "content": "E _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 458, + 506, + 474 + ], + "score": 1.0, + "content": ", is calculated from the difference", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 142, + 486 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 143, + 472, + 155, + 485 + ], + "score": 0.9, + "content": "\\hat { A } _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 473, + 173, + 486 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 174, + 473, + 186, + 485 + ], + "score": 0.89, + "content": "A _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "and is split into ReLU-activated positive and negative prediction errors, which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "are concatenated along the feature dimension. As discussed in (Rao & Ballard, 1999), although not", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 494, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 509 + ], + "score": 1.0, + "content": "explicit in their model, the separate error populations are analogous to the existence of on-center,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 505, + 410, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 410, + 519 + ], + "score": 1.0, + "content": "off-surround and off-center, on-surround neurons early in the visual system.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29, + "bbox_fs": [ + 103, + 345, + 507, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 589 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "The full set of update rules are listed in Equations (1) to (4). The model is trained to minimize", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 534, + 504, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 504, + 545 + ], + "score": 1.0, + "content": "the weighted sum of the activity of the error units. Explicitly, the training loss is formalized in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 280, + 558 + ], + "score": 1.0, + "content": "Equation 5 with weighting factors by time,", + "type": "text" + }, + { + "bbox": [ + 280, + 545, + 290, + 556 + ], + "score": 0.87, + "content": "\\lambda _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 545, + 335, + 558 + ], + "score": 1.0, + "content": ", and layer,", + "type": "text" + }, + { + "bbox": [ + 336, + 546, + 346, + 556 + ], + "score": 0.86, + "content": "\\lambda _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 545, + 393, + 558 + ], + "score": 1.0, + "content": ", and where", + "type": "text" + }, + { + "bbox": [ + 393, + 547, + 404, + 556 + ], + "score": 0.83, + "content": "n _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "is the number of units in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "the lth layer. With error units consisting of subtraction followed by ReLU activation, the loss at each", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 579 + ], + "score": 1.0, + "content": "layer is equivalent to an L1 error. Although not explored here, other error unit implementations,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 578, + 466, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 466, + 590 + ], + "score": 1.0, + "content": "potentially even probabilistic or adversarial (Goodfellow et al., 2014), could also be used.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 522, + 506, + 590 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 606, + 422, + 685 + ], + "lines": [ + { + "bbox": [ + 203, + 606, + 422, + 685 + ], + "spans": [ + { + "bbox": [ + 203, + 606, + 422, + 685 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { A _ { l } ^ { t } = \\bigg \\{ x _ { t } } & { \\mathrm { i f } l = 0 } \\\\ & { \\mathbf { M } _ { \\mathrm { A X P o o L } } ( \\operatorname { R E L U } ( \\operatorname { C o N v } ( E _ { l - 1 } ^ { t } ) ) ) } & { l > 0 } \\\\ & { \\hat { A } _ { l } ^ { t } = \\operatorname { R E L U } ( \\operatorname { C o N v } ( R _ { l } ^ { t } ) ) } \\\\ & { E _ { l } ^ { t } = [ \\operatorname { R E L U } ( A _ { l } ^ { t } - \\hat { A } _ { l } ^ { t } ) ; \\operatorname { R E L U } ( \\hat { A } _ { l } ^ { t } - A _ { l } ^ { t } ) ] } \\\\ & { R _ { l } ^ { t } = \\operatorname { C o N v L S T M } ( E _ { l } ^ { t - 1 } , R _ { l } ^ { t - 1 } , \\operatorname { U P S A M P L E } ( R _ { l + 1 } ^ { t } ) ) } \\end{array}", + "type": "interline_equation", + "image_path": "d4efd34db2f2780addbfd9d2817ec3f2ce20b4d0c18a15b42c819a20676e5669.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 203, + 606, + 422, + 621.8 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 203, + 621.8, + 422, + 637.5999999999999 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 203, + 637.5999999999999, + 422, + 653.3999999999999 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 203, + 653.3999999999999, + 422, + 669.1999999999998 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 203, + 669.1999999999998, + 422, + 684.9999999999998 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 700, + 315, + 730 + ], + "lines": [ + { + "bbox": [ + 189, + 700, + 315, + 730 + ], + "spans": [ + { + "bbox": [ + 189, + 700, + 315, + 730 + ], + "score": 0.92, + "content": "L _ { t r a i n } = \\sum _ { t } \\lambda _ { t } \\sum _ { l } \\frac { \\lambda _ { l } } { n _ { l } } \\sum _ { n _ { l } } E _ { l } ^ { t }", + "type": "interline_equation", + "image_path": "118394227ec7aaffb1e86b56b4dea6f1278381aa47137052f4ade323fa319af4.jpg" + } + ] + } + ], + "index": 48.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 700, + 315, + 715.0 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 189, + 715.0, + 315, + 730.0 + ], + "spans": [], + "index": 49 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 82, + 278, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 81, + 280, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 81, + 280, + 95 + ], + "score": 1.0, + "content": "Algorithm 1 Calculation of PredNet states", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 107, + 98, + 506, + 297 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 98, + 506, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 98, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 108, + 98, + 506, + 297 + ], + "score": 0.557, + "html": "
Require: Xt
1:A←xt 2:E,R←0
3:fort=1 to Tdo
4:for l = L to O do
5:if l=L then
6:Rt =CONVLSTM(Et-1,Rt-1)
7:else
8:Rt = CONVLSTM(E𝑡-1,Rt-1,UPSAMPLE(Rt+1))
9:for l= O to L do
10:if l= O then
11:At = SATLU(RELU(CONV(Rδ)))
12:else
13:At = RELU(CoNv(Rt))
14:Et = [RELU(At - At); RELU(At - A)]
15:ifl<L then
16:At+1 = MAXPOOL(CONV(E))
", + "type": "table", + "image_path": "c6086e2374093f12c8dbee3c103704a2878634b0eaf8ad59396dd945a9e15d31.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 107, + 98, + 506, + 164.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 164.33333333333331, + 506, + 230.66666666666663 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 107, + 230.66666666666663, + 506, + 296.99999999999994 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "The order in which each unit in the model is updated must also be specified, and our implementa-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "tion is described in Algorithm 1. Updating of states occurs through two passes: a top-down pass", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 341, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 149, + 355 + ], + "score": 1.0, + "content": "where the", + "type": "text" + }, + { + "bbox": [ + 149, + 342, + 161, + 354 + ], + "score": 0.89, + "content": "R _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 341, + 505, + 355 + ], + "score": 1.0, + "content": "states are computed, and then a forward pass to calculate the predictions, errors, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 299, + 365 + ], + "score": 1.0, + "content": "higher level targets. A last detail of note is that", + "type": "text" + }, + { + "bbox": [ + 299, + 353, + 311, + 364 + ], + "score": 0.88, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 352, + 329, + 365 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 329, + 353, + 341, + 364 + ], + "score": 0.88, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "are initialized to zero, which, due to the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 363, + 464, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 464, + 377 + ], + "score": 1.0, + "content": "convolutional nature of the network, means that the initial prediction is spatially uniform.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 108, + 394, + 200, + 407 + ], + "lines": [ + { + "bbox": [ + 104, + 393, + 202, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 393, + 202, + 410 + ], + "score": 1.0, + "content": "3 EXPERIMENTS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 108, + 421, + 264, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 266, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 266, + 435 + ], + "score": 1.0, + "content": "3.1 RENDERED IMAGE SEQUENCES", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 457 + ], + "score": 1.0, + "content": "To gain an understanding of the representations learned in the proposed framework, we first trained", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "score": 1.0, + "content": "PredNet models using synthetic images, for which we have access to the underlying generative", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "stimulus model and all latent parameters. We created sequences of rendered faces rotating with two", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "degrees of freedom, along the “pan” (out-of-plane) and “roll” (in-plane) axes. The faces start at a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "random orientation and rotate at a random constant velocity for a total of 10 frames. A different face", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "was sampled for each sequence. The images were processed to be grayscale, with values normalized", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "between 0 and 1, and 64x64 pixels in size. We used 16K sequences for training and 800 for both", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 519, + 197, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 197, + 533 + ], + "score": 1.0, + "content": "validation and testing.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 537, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Predictions generated by a PredNet model are shown in Figure 2. The model is able to accumulate", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 547, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 561 + ], + "score": 1.0, + "content": "information over time to make accurate predictions of future frames. Since the representation neu-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 557, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 506, + 573 + ], + "score": 1.0, + "content": "rons are initialized to zero, the prediction at the first time step is uniform. On the second time step,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "with no motion information yet, the prediction is a blurry reconstruction of the first time step. After", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 581, + 504, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 504, + 593 + ], + "score": 1.0, + "content": "further iterations, the model adapts to the underlying dynamics to generate predictions that closely", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 591, + 217, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 217, + 605 + ], + "score": 1.0, + "content": "match the incoming frame.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "For choosing the hyperparameters of the model, we performed a random search and chose the model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "that had the lowest L1 error in frame prediction averaged over time steps 2-10 on a validation set.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 631, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 643 + ], + "score": 1.0, + "content": "Given this selection criteria, the best performing models tended to have a loss solely concentrated at", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 641, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 189, + 655 + ], + "score": 1.0, + "content": "the lowest layer (i.e.", + "type": "text" + }, + { + "bbox": [ + 190, + 642, + 219, + 653 + ], + "score": 0.86, + "content": "\\lambda _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 641, + 223, + 655 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 223, + 642, + 262, + 653 + ], + "score": 0.81, + "content": "\\lambda _ { l > 0 } = 0 ", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 641, + 506, + 655 + ], + "score": 1.0, + "content": "), which is the case for the model shown. Using an equal loss", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "score": 1.0, + "content": "at each layer considerably degraded predictions, but enforcing a moderate loss on upper layers that", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 318, + 676 + ], + "score": 1.0, + "content": "was one magnitude smaller than the lowest layer (i.e.", + "type": "text" + }, + { + "bbox": [ + 318, + 664, + 348, + 675 + ], + "score": 0.86, + "content": "\\lambda _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 663, + 352, + 676 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 352, + 664, + 399, + 675 + ], + "score": 0.85, + "content": "\\lambda _ { l > 0 } = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 663, + 505, + 676 + ], + "score": 1.0, + "content": ") led to only slightly worse", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 674, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 450, + 688 + ], + "score": 1.0, + "content": "predictions, as illustrated in Figure 9 in the Appendix. In all cases, the time loss weight,", + "type": "text" + }, + { + "bbox": [ + 451, + 675, + 461, + 686 + ], + "score": 0.87, + "content": "\\lambda _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 674, + 505, + 688 + ], + "score": 1.0, + "content": ", was set to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 685, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 505, + 699 + ], + "score": 1.0, + "content": "zero for the first time step and then one for all time steps after. As for the remaining hyperparameters,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "the model shown has 5 layers with 3x3 filter sizes for all convolutions, max-pooling of stride 2, and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 707, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 259, + 721 + ], + "score": 1.0, + "content": "number of channels per layer, for both", + "type": "text" + }, + { + "bbox": [ + 259, + 708, + 271, + 718 + ], + "score": 0.89, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 707, + 289, + 721 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 289, + 708, + 300, + 718 + ], + "score": 0.88, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 707, + 505, + 721 + ], + "score": 1.0, + "content": "units, of (1, 32, 64, 128, 256). Model weights were", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 719, + 348, + 731 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 348, + 731 + ], + "score": 1.0, + "content": "optimized using the Adam algorithm (Kingma & Ba, 2014).", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 82, + 278, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 81, + 280, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 81, + 280, + 95 + ], + "score": 1.0, + "content": "Algorithm 1 Calculation of PredNet states", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 107, + 98, + 506, + 297 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 107, + 98, + 506, + 297 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 98, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 108, + 98, + 506, + 297 + ], + "score": 0.557, + "html": "
Require: Xt
1:A←xt 2:E,R←0
3:fort=1 to Tdo
4:for l = L to O do
5:if l=L then
6:Rt =CONVLSTM(Et-1,Rt-1)
7:else
8:Rt = CONVLSTM(E𝑡-1,Rt-1,UPSAMPLE(Rt+1))
9:for l= O to L do
10:if l= O then
11:At = SATLU(RELU(CONV(Rδ)))
12:else
13:At = RELU(CoNv(Rt))
14:Et = [RELU(At - At); RELU(At - A)]
15:ifl<L then
16:At+1 = MAXPOOL(CONV(E))
", + "type": "table", + "image_path": "c6086e2374093f12c8dbee3c103704a2878634b0eaf8ad59396dd945a9e15d31.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 107, + 98, + 506, + 164.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 164.33333333333331, + 506, + 230.66666666666663 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 107, + 230.66666666666663, + 506, + 296.99999999999994 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "The order in which each unit in the model is updated must also be specified, and our implementa-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "tion is described in Algorithm 1. Updating of states occurs through two passes: a top-down pass", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 341, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 149, + 355 + ], + "score": 1.0, + "content": "where the", + "type": "text" + }, + { + "bbox": [ + 149, + 342, + 161, + 354 + ], + "score": 0.89, + "content": "R _ { l } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 341, + 505, + 355 + ], + "score": 1.0, + "content": "states are computed, and then a forward pass to calculate the predictions, errors, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 299, + 365 + ], + "score": 1.0, + "content": "higher level targets. A last detail of note is that", + "type": "text" + }, + { + "bbox": [ + 299, + 353, + 311, + 364 + ], + "score": 0.88, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 352, + 329, + 365 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 329, + 353, + 341, + 364 + ], + "score": 0.88, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "are initialized to zero, which, due to the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 363, + 464, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 464, + 377 + ], + "score": 1.0, + "content": "convolutional nature of the network, means that the initial prediction is spatially uniform.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 320, + 505, + 377 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 394, + 200, + 407 + ], + "lines": [ + { + "bbox": [ + 104, + 393, + 202, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 393, + 202, + 410 + ], + "score": 1.0, + "content": "3 EXPERIMENTS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 108, + 421, + 264, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 266, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 266, + 435 + ], + "score": 1.0, + "content": "3.1 RENDERED IMAGE SEQUENCES", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 457 + ], + "score": 1.0, + "content": "To gain an understanding of the representations learned in the proposed framework, we first trained", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "score": 1.0, + "content": "PredNet models using synthetic images, for which we have access to the underlying generative", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "stimulus model and all latent parameters. We created sequences of rendered faces rotating with two", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "degrees of freedom, along the “pan” (out-of-plane) and “roll” (in-plane) axes. The faces start at a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "random orientation and rotate at a random constant velocity for a total of 10 frames. A different face", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "was sampled for each sequence. The images were processed to be grayscale, with values normalized", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "between 0 and 1, and 64x64 pixels in size. We used 16K sequences for training and 800 for both", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 519, + 197, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 197, + 533 + ], + "score": 1.0, + "content": "validation and testing.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 443, + 506, + 533 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 537, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Predictions generated by a PredNet model are shown in Figure 2. The model is able to accumulate", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 547, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 561 + ], + "score": 1.0, + "content": "information over time to make accurate predictions of future frames. Since the representation neu-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 557, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 506, + 573 + ], + "score": 1.0, + "content": "rons are initialized to zero, the prediction at the first time step is uniform. On the second time step,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "with no motion information yet, the prediction is a blurry reconstruction of the first time step. After", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 581, + 504, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 504, + 593 + ], + "score": 1.0, + "content": "further iterations, the model adapts to the underlying dynamics to generate predictions that closely", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 591, + 217, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 217, + 605 + ], + "score": 1.0, + "content": "match the incoming frame.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 537, + 506, + 605 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "For choosing the hyperparameters of the model, we performed a random search and chose the model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "that had the lowest L1 error in frame prediction averaged over time steps 2-10 on a validation set.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 631, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 643 + ], + "score": 1.0, + "content": "Given this selection criteria, the best performing models tended to have a loss solely concentrated at", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 641, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 189, + 655 + ], + "score": 1.0, + "content": "the lowest layer (i.e.", + "type": "text" + }, + { + "bbox": [ + 190, + 642, + 219, + 653 + ], + "score": 0.86, + "content": "\\lambda _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 641, + 223, + 655 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 223, + 642, + 262, + 653 + ], + "score": 0.81, + "content": "\\lambda _ { l > 0 } = 0 ", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 641, + 506, + 655 + ], + "score": 1.0, + "content": "), which is the case for the model shown. Using an equal loss", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 506, + 665 + ], + "score": 1.0, + "content": "at each layer considerably degraded predictions, but enforcing a moderate loss on upper layers that", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 318, + 676 + ], + "score": 1.0, + "content": "was one magnitude smaller than the lowest layer (i.e.", + "type": "text" + }, + { + "bbox": [ + 318, + 664, + 348, + 675 + ], + "score": 0.86, + "content": "\\lambda _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 663, + 352, + 676 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 352, + 664, + 399, + 675 + ], + "score": 0.85, + "content": "\\lambda _ { l > 0 } = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 663, + 505, + 676 + ], + "score": 1.0, + "content": ") led to only slightly worse", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 674, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 450, + 688 + ], + "score": 1.0, + "content": "predictions, as illustrated in Figure 9 in the Appendix. In all cases, the time loss weight,", + "type": "text" + }, + { + "bbox": [ + 451, + 675, + 461, + 686 + ], + "score": 0.87, + "content": "\\lambda _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 674, + 505, + 688 + ], + "score": 1.0, + "content": ", was set to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 685, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 505, + 699 + ], + "score": 1.0, + "content": "zero for the first time step and then one for all time steps after. As for the remaining hyperparameters,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "the model shown has 5 layers with 3x3 filter sizes for all convolutions, max-pooling of stride 2, and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 707, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 259, + 721 + ], + "score": 1.0, + "content": "number of channels per layer, for both", + "type": "text" + }, + { + "bbox": [ + 259, + 708, + 271, + 718 + ], + "score": 0.89, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 707, + 289, + 721 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 289, + 708, + 300, + 718 + ], + "score": 0.88, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 707, + 505, + 721 + ], + "score": 1.0, + "content": "units, of (1, 32, 64, 128, 256). Model weights were", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 719, + 348, + 731 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 348, + 731 + ], + "score": 1.0, + "content": "optimized using the Adam algorithm (Kingma & Ba, 2014).", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 608, + 506, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 80, + 505, + 302 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 80, + 505, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 505, + 302 + ], + "score": 0.969, + "type": "image", + "image_path": "47e158f8ba016838740a961f995997b73a4e432fd723f2faa4ecb1fae041441c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 80, + 505, + 154.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 154.0, + 505, + 228.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 228.0, + 505, + 302.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 314, + 298, + 348, + 308 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 349, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 337, + 310 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 337, + 299, + 349, + 308 + ], + "score": 0.56, + "content": "", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 309, + 504, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 308, + 504, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 504, + 322 + ], + "score": 1.0, + "content": "Figure 2: PredNet next-frame predictions for sequences of rendered faces rotating with two degrees", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 318, + 331, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 331, + 333 + ], + "score": 1.0, + "content": "of freedom. Faces shown were not seen during training.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 310, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 312, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 312, + 354 + ], + "score": 1.0, + "content": "Quantitative evaluation of generative models is a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 353, + 311, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 311, + 365 + ], + "score": 1.0, + "content": "difficult, unsolved problem (Theis et al., 2016), but", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 364, + 311, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 311, + 376 + ], + "score": 1.0, + "content": "here we report prediction error in terms of mean-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 374, + 311, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 311, + 387 + ], + "score": 1.0, + "content": "squared error (MSE) and the Structural Similarity", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 384, + 312, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 312, + 398 + ], + "score": 1.0, + "content": "Index Measure (SSIM) (Wang et al., 2004). SSIM", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 396, + 311, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 311, + 409 + ], + "score": 1.0, + "content": "is designed to be more correlated with perceptual", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 407, + 312, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 218, + 420 + ], + "score": 1.0, + "content": "judgments, and ranges from", + "type": "text" + }, + { + "bbox": [ + 219, + 408, + 232, + 418 + ], + "score": 0.65, + "content": "- 1", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 407, + 312, + 420 + ], + "score": 1.0, + "content": "and 1, with a larger", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 419, + 311, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 311, + 431 + ], + "score": 1.0, + "content": "score indicating greater similarity. We compare the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 428, + 311, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 311, + 442 + ], + "score": 1.0, + "content": "PredNet to the trivial solution of copying the last", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 319, + 365, + 503, + 429 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 318, + 344, + 504, + 365 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 504, + 356 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 504, + 356 + ], + "score": 1.0, + "content": "Table 1: Evaluation of next-frame predictions", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 354, + 464, + 366 + ], + "spans": [ + { + "bbox": [ + 318, + 354, + 464, + 366 + ], + "score": 1.0, + "content": "on Rotating Faces Dataset (test set).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5 + }, + { + "type": "table_body", + "bbox": [ + 319, + 365, + 503, + 429 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 319, + 365, + 503, + 429 + ], + "spans": [ + { + "bbox": [ + 319, + 365, + 503, + 429 + ], + "score": 0.98, + "html": "
MSESSIM
PredNet Lo0.01520.937
PredNet Lall0.01570.921
CNN-LSTM Enc.-Dec.0.01800.907
Copy Last Frame0.1250.631
", + "type": "table", + "image_path": "52104c4d397a0ebe01aea3333b35ca079c4c81446bf375c1e186c1aa440a9d94.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 319, + 365, + 503, + 377.8 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 319, + 377.8, + 503, + 390.6 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 319, + 390.6, + 503, + 403.40000000000003 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 319, + 403.40000000000003, + 503, + 416.20000000000005 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 319, + 416.20000000000005, + 503, + 429.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + } + ], + "index": 15.25 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "frame, as well as a control model that shares the overall architecture and training scheme of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 452, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 348, + 463 + ], + "score": 1.0, + "content": "PredNet, but that sends forward the layer-wise activations", + "type": "text" + }, + { + "bbox": [ + 348, + 452, + 366, + 463 + ], + "score": 0.81, + "content": "( A _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 452, + 459, + 463 + ], + "score": 1.0, + "content": "rather than the errors", + "type": "text" + }, + { + "bbox": [ + 459, + 452, + 477, + 463 + ], + "score": 0.85, + "content": "( E _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 452, + 505, + 463 + ], + "score": 1.0, + "content": ". This", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 463, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 505, + 474 + ], + "score": 1.0, + "content": "model thus takes the form of a more traditional encoder-decoder pair, with a CNN encoder that has", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "lateral skip connections to a convolutional LSTM decoder. The performance of all models on the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 497 + ], + "score": 1.0, + "content": "rotating faces dataset is summarized in Table 1, where the scores were calculated as an average over", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "all predictions after the first frame. We report results for the PredNet model trained with loss only", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 273, + 518 + ], + "score": 1.0, + "content": "on the lowest layer, denoted as PredNet", + "type": "text" + }, + { + "bbox": [ + 274, + 506, + 286, + 518 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 507, + 505, + 518 + ], + "score": 1.0, + "content": ", as well as the model trained with an 0.1 weight on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 244, + 529 + ], + "score": 1.0, + "content": "upper layers, denoted as PredNet", + "type": "text" + }, + { + "bbox": [ + 244, + 518, + 262, + 528 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 517, + 505, + 529 + ], + "score": 1.0, + "content": ". Both PredNet models outperformed the baselines on both", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 183, + 541 + ], + "score": 1.0, + "content": "measures, with the", + "type": "text" + }, + { + "bbox": [ + 183, + 529, + 196, + 539 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 528, + 317, + 541 + ], + "score": 1.0, + "content": "model slightly outperforming", + "type": "text" + }, + { + "bbox": [ + 318, + 529, + 335, + 539 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 528, + 506, + 541 + ], + "score": 1.0, + "content": ", as expected for evaluating the pixel-level", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 540, + 156, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 156, + 551 + ], + "score": 1.0, + "content": "predictions.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "score": 1.0, + "content": "Synthetic sequences were chosen as the initial training set in order to better understand what is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "learned in different layers of the model, specifically with respect to the underlying generative model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 578, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 504, + 591 + ], + "score": 1.0, + "content": "(Kulkarni et al., 2015). The rotating faces were generated using the FaceGen software package (Sin-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "gular Inversions, Inc.), which internally generates 3D face meshes by a principal component analysis", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 599, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 614 + ], + "score": 1.0, + "content": "in “face space”, derived from a corpus of 3D face scans. Thus, the latent parameters of the image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "sequences used here consist of the initial pan and roll angles, the pan and roll velocities, and the prin-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "cipal component (PC) values, which control the “identity” of the face. To understand the information", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "contained in the trained models, we decoded the latent parameters from the representation neurons", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 107, + 644, + 125, + 655 + ], + "score": 0.82, + "content": "( R _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 644, + 326, + 656 + ], + "score": 1.0, + "content": "in different layers, using a ridge regression. The", + "type": "text" + }, + { + "bbox": [ + 326, + 644, + 339, + 655 + ], + "score": 0.88, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "states were taken at the earliest possible", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "informative time steps, which, in the our notation, are the second and third steps, respectively, for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 390, + 678 + ], + "score": 1.0, + "content": "the static and dynamic parameters. The regression was trained using", + "type": "text" + }, + { + "bbox": [ + 390, + 666, + 406, + 676 + ], + "score": 0.72, + "content": "4 K", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "sequences with 500 for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 167, + 689 + ], + "score": 1.0, + "content": "validation and", + "type": "text" + }, + { + "bbox": [ + 167, + 677, + 182, + 687 + ], + "score": 0.8, + "content": "1 K", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "for testing. For a baseline comparison of the information implicitly embedded", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "in the network architecture, we compare to the decoding accuracies of an untrained network with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 699, + 504, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 710 + ], + "score": 1.0, + "content": "random initial weights. Note that in this randomly initialized case, we still expect above-chance de-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "coding performance, given past theoretical and empirical work with random networks (Pinto et al.,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 720, + 282, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 282, + 733 + ], + "score": 1.0, + "content": "2009; Jarrett et al., 2009; Saxe et al., 2010).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 39.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 80, + 505, + 302 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 80, + 505, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 505, + 302 + ], + "score": 0.969, + "type": "image", + "image_path": "47e158f8ba016838740a961f995997b73a4e432fd723f2faa4ecb1fae041441c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 80, + 505, + 154.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 154.0, + 505, + 228.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 228.0, + 505, + 302.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 314, + 298, + 348, + 308 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 349, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 337, + 310 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 337, + 299, + 349, + 308 + ], + "score": 0.56, + "content": "", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 309, + 504, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 308, + 504, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 504, + 322 + ], + "score": 1.0, + "content": "Figure 2: PredNet next-frame predictions for sequences of rendered faces rotating with two degrees", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 318, + 331, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 331, + 333 + ], + "score": 1.0, + "content": "of freedom. Faces shown were not seen during training.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 310, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 312, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 312, + 354 + ], + "score": 1.0, + "content": "Quantitative evaluation of generative models is a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 353, + 311, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 311, + 365 + ], + "score": 1.0, + "content": "difficult, unsolved problem (Theis et al., 2016), but", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 364, + 311, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 311, + 376 + ], + "score": 1.0, + "content": "here we report prediction error in terms of mean-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 374, + 311, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 311, + 387 + ], + "score": 1.0, + "content": "squared error (MSE) and the Structural Similarity", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 384, + 312, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 312, + 398 + ], + "score": 1.0, + "content": "Index Measure (SSIM) (Wang et al., 2004). SSIM", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 396, + 311, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 311, + 409 + ], + "score": 1.0, + "content": "is designed to be more correlated with perceptual", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 407, + 312, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 218, + 420 + ], + "score": 1.0, + "content": "judgments, and ranges from", + "type": "text" + }, + { + "bbox": [ + 219, + 408, + 232, + 418 + ], + "score": 0.65, + "content": "- 1", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 407, + 312, + 420 + ], + "score": 1.0, + "content": "and 1, with a larger", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 419, + 311, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 311, + 431 + ], + "score": 1.0, + "content": "score indicating greater similarity. We compare the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 428, + 311, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 311, + 442 + ], + "score": 1.0, + "content": "PredNet to the trivial solution of copying the last", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "frame, as well as a control model that shares the overall architecture and training scheme of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 452, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 348, + 463 + ], + "score": 1.0, + "content": "PredNet, but that sends forward the layer-wise activations", + "type": "text" + }, + { + "bbox": [ + 348, + 452, + 366, + 463 + ], + "score": 0.81, + "content": "( A _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 452, + 459, + 463 + ], + "score": 1.0, + "content": "rather than the errors", + "type": "text" + }, + { + "bbox": [ + 459, + 452, + 477, + 463 + ], + "score": 0.85, + "content": "( E _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 452, + 505, + 463 + ], + "score": 1.0, + "content": ". This", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 463, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 505, + 474 + ], + "score": 1.0, + "content": "model thus takes the form of a more traditional encoder-decoder pair, with a CNN encoder that has", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "lateral skip connections to a convolutional LSTM decoder. The performance of all models on the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 497 + ], + "score": 1.0, + "content": "rotating faces dataset is summarized in Table 1, where the scores were calculated as an average over", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "all predictions after the first frame. We report results for the PredNet model trained with loss only", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 273, + 518 + ], + "score": 1.0, + "content": "on the lowest layer, denoted as PredNet", + "type": "text" + }, + { + "bbox": [ + 274, + 506, + 286, + 518 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 507, + 505, + 518 + ], + "score": 1.0, + "content": ", as well as the model trained with an 0.1 weight on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 244, + 529 + ], + "score": 1.0, + "content": "upper layers, denoted as PredNet", + "type": "text" + }, + { + "bbox": [ + 244, + 518, + 262, + 528 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 517, + 505, + 529 + ], + "score": 1.0, + "content": ". Both PredNet models outperformed the baselines on both", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 183, + 541 + ], + "score": 1.0, + "content": "measures, with the", + "type": "text" + }, + { + "bbox": [ + 183, + 529, + 196, + 539 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 528, + 317, + 541 + ], + "score": 1.0, + "content": "model slightly outperforming", + "type": "text" + }, + { + "bbox": [ + 318, + 529, + 335, + 539 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 528, + 506, + 541 + ], + "score": 1.0, + "content": ", as expected for evaluating the pixel-level", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 540, + 156, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 156, + 551 + ], + "score": 1.0, + "content": "predictions.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 342, + 312, + 442 + ] + }, + { + "type": "table", + "bbox": [ + 319, + 365, + 503, + 429 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 318, + 344, + 504, + 365 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 504, + 356 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 504, + 356 + ], + "score": 1.0, + "content": "Table 1: Evaluation of next-frame predictions", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 354, + 464, + 366 + ], + "spans": [ + { + "bbox": [ + 318, + 354, + 464, + 366 + ], + "score": 1.0, + "content": "on Rotating Faces Dataset (test set).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5 + }, + { + "type": "table_body", + "bbox": [ + 319, + 365, + 503, + 429 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 319, + 365, + 503, + 429 + ], + "spans": [ + { + "bbox": [ + 319, + 365, + 503, + 429 + ], + "score": 0.98, + "html": "
MSESSIM
PredNet Lo0.01520.937
PredNet Lall0.01570.921
CNN-LSTM Enc.-Dec.0.01800.907
Copy Last Frame0.1250.631
", + "type": "table", + "image_path": "52104c4d397a0ebe01aea3333b35ca079c4c81446bf375c1e186c1aa440a9d94.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 319, + 365, + 503, + 377.8 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 319, + 377.8, + 503, + 390.6 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 319, + 390.6, + 503, + 403.40000000000003 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 319, + 403.40000000000003, + 503, + 416.20000000000005 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 319, + 416.20000000000005, + 503, + 429.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + } + ], + "index": 15.25 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 550 + ], + "lines": [], + "index": 26.5, + "bbox_fs": [ + 105, + 440, + 506, + 551 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 568 + ], + "score": 1.0, + "content": "Synthetic sequences were chosen as the initial training set in order to better understand what is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "learned in different layers of the model, specifically with respect to the underlying generative model", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 578, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 504, + 591 + ], + "score": 1.0, + "content": "(Kulkarni et al., 2015). The rotating faces were generated using the FaceGen software package (Sin-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "gular Inversions, Inc.), which internally generates 3D face meshes by a principal component analysis", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 599, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 614 + ], + "score": 1.0, + "content": "in “face space”, derived from a corpus of 3D face scans. Thus, the latent parameters of the image", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "sequences used here consist of the initial pan and roll angles, the pan and roll velocities, and the prin-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "cipal component (PC) values, which control the “identity” of the face. To understand the information", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "contained in the trained models, we decoded the latent parameters from the representation neurons", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 107, + 644, + 125, + 655 + ], + "score": 0.82, + "content": "( R _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 644, + 326, + 656 + ], + "score": 1.0, + "content": "in different layers, using a ridge regression. The", + "type": "text" + }, + { + "bbox": [ + 326, + 644, + 339, + 655 + ], + "score": 0.88, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "states were taken at the earliest possible", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "informative time steps, which, in the our notation, are the second and third steps, respectively, for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 390, + 678 + ], + "score": 1.0, + "content": "the static and dynamic parameters. The regression was trained using", + "type": "text" + }, + { + "bbox": [ + 390, + 666, + 406, + 676 + ], + "score": 0.72, + "content": "4 K", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "sequences with 500 for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 167, + 689 + ], + "score": 1.0, + "content": "validation and", + "type": "text" + }, + { + "bbox": [ + 167, + 677, + 182, + 687 + ], + "score": 0.8, + "content": "1 K", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "for testing. For a baseline comparison of the information implicitly embedded", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "in the network architecture, we compare to the decoding accuracies of an untrained network with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 699, + 504, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 710 + ], + "score": 1.0, + "content": "random initial weights. Note that in this randomly initialized case, we still expect above-chance de-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "coding performance, given past theoretical and empirical work with random networks (Pinto et al.,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 720, + 282, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 282, + 733 + ], + "score": 1.0, + "content": "2009; Jarrett et al., 2009; Saxe et al., 2010).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 556, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Latent variable decoding accuracies of the pan and roll velocities, pan initial angle, and first PC are", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "shown in the left panel of Figure 3. There are several interesting patterns. First, the trained models", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "learn a representation that generally permits a better linear decoding of the underlying latent factors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "than the randomly initialized model, with the most striking difference in terms of the the pan rotation", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 132, + 140 + ], + "score": 1.0, + "content": "speed", + "type": "text" + }, + { + "bbox": [ + 133, + 127, + 160, + 138 + ], + "score": 0.91, + "content": "( \\alpha _ { p a n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 126, + 367, + 140 + ], + "score": 1.0, + "content": ". Second, the most notable difference between the", + "type": "text" + }, + { + "bbox": [ + 367, + 127, + 385, + 137 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 126, + 405, + 140 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 127, + 417, + 137 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "versions occurs with", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 151 + ], + "score": 1.0, + "content": "the first principle component, where the model trained with loss on all layers has a higher decoding", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 375, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 375, + 162 + ], + "score": 1.0, + "content": "accuracy than the model trained with loss only on the lowest layer.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 115, + 186, + 497, + 370 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 186, + 497, + 370 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 186, + 497, + 370 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 497, + 370 + ], + "score": 0.972, + "type": "image", + "image_path": "0abef2797da5fa19f50184da371c6feb85f35b1d2b18d3d978fa3150145a690d.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 115, + 186, + 497, + 247.33333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 247.33333333333334, + 497, + 308.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 115, + 308.6666666666667, + 497, + 370.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 375, + 505, + 420 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "score": 1.0, + "content": "Figure 3: Information contained in PredNet representation for rotating faces sequences. Left: De-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 386, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 308, + 400 + ], + "score": 1.0, + "content": "coding of latent variables using a ridge regression", + "type": "text" + }, + { + "bbox": [ + 309, + 388, + 331, + 399 + ], + "score": 0.87, + "content": "( \\alpha _ { p a n }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 386, + 481, + 400 + ], + "score": 1.0, + "content": ": pan (out-of-frame) angular velocity,", + "type": "text" + }, + { + "bbox": [ + 481, + 387, + 501, + 399 + ], + "score": 0.9, + "content": "\\theta _ { p a n }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 386, + 505, + 400 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 315, + 411 + ], + "score": 1.0, + "content": "pan angle, PC-1: first principal component of face,", + "type": "text" + }, + { + "bbox": [ + 315, + 400, + 336, + 409 + ], + "score": 0.89, + "content": "\\alpha _ { r o l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 397, + 505, + 411 + ], + "score": 1.0, + "content": ": roll (in-frame) angular velocity). Right:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 408, + 307, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 307, + 420 + ], + "score": 1.0, + "content": "Orientation-invariant classification of static faces.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + } + ], + "index": 9.75 + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 454 + ], + "score": 1.0, + "content": "The latent variable decoding analysis suggests that the model learns a representation that may gen-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "eralize well to other tasks for which it was not explicitly trained. To investigate this further, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "assessed the models in a classification task from single, static images. We created a dataset of 25", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 468, + 509, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 398, + 492 + ], + "score": 1.0, + "content": "previously unseen FaceGen faces at 7 pan angles, equally spaced between", + "type": "text" + }, + { + "bbox": [ + 399, + 473, + 432, + 487 + ], + "score": 0.93, + "content": "[ - \\frac { \\pi } { 2 } , \\frac { \\pi } { 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 468, + 509, + 492 + ], + "score": 1.0, + "content": ", and 8 roll angles,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 206, + 497 + ], + "score": 1.0, + "content": "equally spaced between", + "type": "text" + }, + { + "bbox": [ + 206, + 484, + 234, + 497 + ], + "score": 0.92, + "content": "[ 0 , 2 \\pi )", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 484, + 329, + 497 + ], + "score": 1.0, + "content": ". There were therefore", + "type": "text" + }, + { + "bbox": [ + 375, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "orientations per identity, which", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "were tested in a cross-validated fashion. A linear SVM to decode face identity was fit on a model’s", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "representation of a random subset of orientations and then tested on the remaining angles. For each", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "size of the SVM training set, ranging from 1-40 orientations per face, 50 different random splits", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 529, + 318, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 318, + 541 + ], + "score": 1.0, + "content": "were generated, with results averaged over the splits.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "For the static face classification task, we compare the PredNets to a standard autoencoder and a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "variant of the Ladder Network (Valpola, 2015; Rasmus et al., 2015). Both models were constructed", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "to have the same number of layers and channel sizes as the PredNets, as well as a similar alternat-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "ing convolution/max-pooling, then upsampling/convolution scheme. As both networks are autoen-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "coders, they were trained with a reconstruction loss, with a dataset consisting of all of the individual", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "frames from the sequences used to train the PredNets. For the Ladder Network, which is a denois-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "ing autoencoder with lateral skip connections, one must also choose a noise parameter, as well as", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "the relative weights of each layer in the total cost. We tested noise levels ranging from 0 to 0.5", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "in increments of 0.1, with loss weights either evenly distributed across layers, solely concentrated", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 642, + 504, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 642, + 485, + 658 + ], + "score": 1.0, + "content": "at the pixel layer, or 1 at the bottom layer and 0.1 at upper layers (analogous to the PredNet", + "type": "text" + }, + { + "bbox": [ + 486, + 644, + 504, + 655 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "model). Shown is the model that performed best for classification, which consisted of 0.4 noise and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "only pixel weighting. Lastly, as in our architecture, the Ladder Network has lateral and top-down", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "streams that are combined by a combinator function. Inspired by (Pezeshki et al., 2015), where a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "learnable MLP improved results, and to be consistent in comparing to the PredNet, we used a purely", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "convolutional combinator. Given the distributed representation in both networks, we decoded from", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "a concatenation of the feature representations at all layers, except the pixel layer. For the PredNets,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 489, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 489, + 734 + ], + "score": 1.0, + "content": "the representation units were used and features were extracted after processing one input frame.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 31 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Latent variable decoding accuracies of the pan and roll velocities, pan initial angle, and first PC are", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "shown in the left panel of Figure 3. There are several interesting patterns. First, the trained models", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "learn a representation that generally permits a better linear decoding of the underlying latent factors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "than the randomly initialized model, with the most striking difference in terms of the the pan rotation", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 132, + 140 + ], + "score": 1.0, + "content": "speed", + "type": "text" + }, + { + "bbox": [ + 133, + 127, + 160, + 138 + ], + "score": 0.91, + "content": "( \\alpha _ { p a n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 126, + 367, + 140 + ], + "score": 1.0, + "content": ". Second, the most notable difference between the", + "type": "text" + }, + { + "bbox": [ + 367, + 127, + 385, + 137 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 126, + 405, + 140 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 127, + 417, + 137 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "versions occurs with", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 151 + ], + "score": 1.0, + "content": "the first principle component, where the model trained with loss on all layers has a higher decoding", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 375, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 375, + 162 + ], + "score": 1.0, + "content": "accuracy than the model trained with loss only on the lowest layer.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 83, + 506, + 162 + ] + }, + { + "type": "image", + "bbox": [ + 115, + 186, + 497, + 370 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 186, + 497, + 370 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 186, + 497, + 370 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 497, + 370 + ], + "score": 0.972, + "type": "image", + "image_path": "0abef2797da5fa19f50184da371c6feb85f35b1d2b18d3d978fa3150145a690d.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 115, + 186, + 497, + 247.33333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 247.33333333333334, + 497, + 308.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 115, + 308.6666666666667, + 497, + 370.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 375, + 505, + 420 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "score": 1.0, + "content": "Figure 3: Information contained in PredNet representation for rotating faces sequences. Left: De-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 386, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 308, + 400 + ], + "score": 1.0, + "content": "coding of latent variables using a ridge regression", + "type": "text" + }, + { + "bbox": [ + 309, + 388, + 331, + 399 + ], + "score": 0.87, + "content": "( \\alpha _ { p a n }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 386, + 481, + 400 + ], + "score": 1.0, + "content": ": pan (out-of-frame) angular velocity,", + "type": "text" + }, + { + "bbox": [ + 481, + 387, + 501, + 399 + ], + "score": 0.9, + "content": "\\theta _ { p a n }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 386, + 505, + 400 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 315, + 411 + ], + "score": 1.0, + "content": "pan angle, PC-1: first principal component of face,", + "type": "text" + }, + { + "bbox": [ + 315, + 400, + 336, + 409 + ], + "score": 0.89, + "content": "\\alpha _ { r o l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 397, + 505, + 411 + ], + "score": 1.0, + "content": ": roll (in-frame) angular velocity). Right:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 408, + 307, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 307, + 420 + ], + "score": 1.0, + "content": "Orientation-invariant classification of static faces.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + } + ], + "index": 9.75 + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 454 + ], + "score": 1.0, + "content": "The latent variable decoding analysis suggests that the model learns a representation that may gen-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "eralize well to other tasks for which it was not explicitly trained. To investigate this further, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "assessed the models in a classification task from single, static images. We created a dataset of 25", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 468, + 509, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 398, + 492 + ], + "score": 1.0, + "content": "previously unseen FaceGen faces at 7 pan angles, equally spaced between", + "type": "text" + }, + { + "bbox": [ + 399, + 473, + 432, + 487 + ], + "score": 0.93, + "content": "[ - \\frac { \\pi } { 2 } , \\frac { \\pi } { 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 468, + 509, + 492 + ], + "score": 1.0, + "content": ", and 8 roll angles,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 206, + 497 + ], + "score": 1.0, + "content": "equally spaced between", + "type": "text" + }, + { + "bbox": [ + 206, + 484, + 234, + 497 + ], + "score": 0.92, + "content": "[ 0 , 2 \\pi )", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 484, + 329, + 497 + ], + "score": 1.0, + "content": ". There were therefore", + "type": "text" + }, + { + "bbox": [ + 375, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "orientations per identity, which", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "were tested in a cross-validated fashion. A linear SVM to decode face identity was fit on a model’s", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "representation of a random subset of orientations and then tested on the remaining angles. For each", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "size of the SVM training set, ranging from 1-40 orientations per face, 50 different random splits", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 529, + 318, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 318, + 541 + ], + "score": 1.0, + "content": "were generated, with results averaged over the splits.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 439, + 509, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "For the static face classification task, we compare the PredNets to a standard autoencoder and a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "variant of the Ladder Network (Valpola, 2015; Rasmus et al., 2015). Both models were constructed", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "to have the same number of layers and channel sizes as the PredNets, as well as a similar alternat-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "ing convolution/max-pooling, then upsampling/convolution scheme. As both networks are autoen-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "coders, they were trained with a reconstruction loss, with a dataset consisting of all of the individual", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "frames from the sequences used to train the PredNets. For the Ladder Network, which is a denois-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "ing autoencoder with lateral skip connections, one must also choose a noise parameter, as well as", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "the relative weights of each layer in the total cost. We tested noise levels ranging from 0 to 0.5", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "in increments of 0.1, with loss weights either evenly distributed across layers, solely concentrated", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 642, + 504, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 642, + 485, + 658 + ], + "score": 1.0, + "content": "at the pixel layer, or 1 at the bottom layer and 0.1 at upper layers (analogous to the PredNet", + "type": "text" + }, + { + "bbox": [ + 486, + 644, + 504, + 655 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "model). Shown is the model that performed best for classification, which consisted of 0.4 noise and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "only pixel weighting. Lastly, as in our architecture, the Ladder Network has lateral and top-down", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "streams that are combined by a combinator function. Inspired by (Pezeshki et al., 2015), where a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "learnable MLP improved results, and to be consistent in comparing to the PredNet, we used a purely", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "convolutional combinator. Given the distributed representation in both networks, we decoded from", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "a concatenation of the feature representations at all layers, except the pixel layer. For the PredNets,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 489, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 489, + 734 + ], + "score": 1.0, + "content": "the representation units were used and features were extracted after processing one input frame.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 545, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 385, + 95 + ], + "score": 1.0, + "content": "Face classification accuracies using the representations learned by the", + "type": "text" + }, + { + "bbox": [ + 385, + 83, + 398, + 93 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 83, + 415, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 416, + 83, + 434, + 93 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "PredNets, a stan-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "dard autoencoder, and a Ladder Network variant are shown in the right panel of Figure 3. Both", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "PredNets compare favorably to the other models at all sizes of the training set, suggesting they learn", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "a representation that is relatively tolerant to object transformations. Similar to the decoding accu-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 504, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 306, + 139 + ], + "score": 1.0, + "content": "racy of the first principle component, the PredNet", + "type": "text" + }, + { + "bbox": [ + 307, + 127, + 325, + 137 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 127, + 459, + 139 + ], + "score": 1.0, + "content": "model actually outperformed the", + "type": "text" + }, + { + "bbox": [ + 459, + 127, + 471, + 137 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 127, + 504, + 139 + ], + "score": 1.0, + "content": "variant.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 504, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 504, + 149 + ], + "score": 1.0, + "content": "Altogether, these results suggest that predictive training with the PredNet can be a viable alternative", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "to other models trained with a more traditional reconstructive or denoising loss, and that the relative", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 160, + 412, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 194, + 172 + ], + "score": 1.0, + "content": "layer loss weightings", + "type": "text" + }, + { + "bbox": [ + 194, + 160, + 216, + 171 + ], + "score": 0.32, + "content": "( \\lambda _ { l } ^ { } \\mathbf { \\dot { s } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 160, + 412, + 172 + ], + "score": 1.0, + "content": "may be important for the particular task at hand.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 108, + 187, + 258, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 261, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 261, + 201 + ], + "score": 1.0, + "content": "3.2 NATURAL IMAGE SEQUENCES", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "We next sought to test the PredNet architecture on complex, real-world sequences. As a testbed, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "chose car-mounted camera videos, since these videos span across a wide range of settings and are", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "characterized by rich temporal dynamics, including both self-motion of the vehicle and the motion", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "of other objects in the scene (Agrawal et al., 2015). Models were trained using the raw videos from", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "the KITTI dataset (Geiger et al., 2013), which were captured by a roof-mounted camera on a car", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "driving around an urban environment in Germany. Sequences of 10 frames were sampled from the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 273, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 506, + 289 + ], + "score": 1.0, + "content": "“City”, “Residential”, and “Road” categories, with 57 recording sessions used for training and 4", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "used for validation. Frames were center-cropped and downsampled to 128x160 pixels. In total, the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 298, + 289, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 289, + 309 + ], + "score": 1.0, + "content": "training set consisted of roughly 41K frames.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "A random hyperparameter search, with model selection based on the validation set, resulted in a 4", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "layer model with 3x3 convolutions and layer channel sizes of (3, 48, 96, 192). Models were again", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "score": 1.0, + "content": "trained with Adam (Kingma & Ba, 2014) using a loss either solely computed on the lowest layer", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 107, + 347, + 126, + 358 + ], + "score": 0.8, + "content": "( L _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 347, + 403, + 361 + ], + "score": 1.0, + "content": "or with a weight of 1 on the lowest layer and 0.1 on the upper layers", + "type": "text" + }, + { + "bbox": [ + 403, + 347, + 427, + 358 + ], + "score": 0.88, + "content": "( L _ { a l l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 347, + 505, + 361 + ], + "score": 1.0, + "content": ". Adam parameters", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 356, + 264, + 372 + ], + "score": 1.0, + "content": "were initially set to their default values", + "type": "text" + }, + { + "bbox": [ + 265, + 358, + 308, + 369 + ], + "score": 0.83, + "content": "\\mathbf { \\Phi } _ { \\mathrm { { ( } } \\alpha } = 0 . 0 0 1 ", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 356, + 312, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 312, + 358, + 349, + 369 + ], + "score": 0.88, + "content": "\\beta _ { 1 } = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 356, + 353, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 353, + 358, + 402, + 369 + ], + "score": 0.88, + "content": "\\beta _ { 2 } = 0 . 9 9 9 )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 356, + 493, + 372 + ], + "score": 1.0, + "content": "with the learning rate,", + "type": "text" + }, + { + "bbox": [ + 494, + 360, + 501, + 368 + ], + "score": 0.66, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 356, + 506, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "decreasing by a factor of 10 halfway through training. To assess that the network had indeed learned", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "score": 1.0, + "content": "a robust representation, we tested on the CalTech Pedestrian dataset (Dollar et al., 2009), which´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "consists of videos from a dashboard-mounted camera on a vehicle driving around Los Angeles.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "Testing sequences were made to match the frame rate of the KITTI dataset and again cropped to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "128x160 pixels. Quantitative evaluation was performed on the entire CalTech test partition, split", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 223, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 223, + 436 + ], + "score": 1.0, + "content": "into sequences of 10 frames.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 259, + 452 + ], + "score": 1.0, + "content": "Sample PredNet predictions (for the", + "type": "text" + }, + { + "bbox": [ + 259, + 441, + 272, + 452 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 441, + 504, + 452 + ], + "score": 1.0, + "content": "model) on the CalTech Pedestrian dataset are shown in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "Figure 4, and example videos can be found at https://coxlab.github.io/prednet/. The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 462, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 476 + ], + "score": 1.0, + "content": "model is able to make fairly accurate predictions in a wide range of scenarios. In the top sequence", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "of Fig. 4, a car is passing in the opposite direction, and the model, while not perfect, is able to predict", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 485, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 497 + ], + "score": 1.0, + "content": "its trajectory, as well as fill in the ground it leaves behind. Similarly in Sequence 3, the model is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "able to predict the motion of a vehicle completing a left turn. Sequences 2 and 5 illustrate that the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "PredNet can judge its own movement, as it predicts the appearance of shadows and a stationary", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "vehicle as they approach. The model makes reasonable predictions even in difficult scenarios, such", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "as when the camera-mounted vehicle is turning. In Sequence 4, the model predicts the position of a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 539, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 553 + ], + "score": 1.0, + "content": "tree, as the vehicle turns onto a road. The turning sequences also further illustrate the model’s ability", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "to “fill-in”, as it is able to extrapolate sky and tree textures as unseen regions come into view. As an", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "additional control, we show a sequence at the bottom of Fig. 4, where the input has been temporally", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "scrambled. In this case, the model generates blurry frames, which mostly just resemble the previous", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "frame. Finally, although the PredNet shown here was trained to predict one frame ahead, it is also", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "possible to predict multiple frames into the future, by feeding back predictions as the inputs and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 605, + 326, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 326, + 617 + ], + "score": 1.0, + "content": "recursively iterating. We explore this in Appendix 5.3.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 281, + 710 + ], + "lines": [ + { + "bbox": [ + 107, + 621, + 282, + 633 + ], + "spans": [ + { + "bbox": [ + 107, + 621, + 282, + 633 + ], + "score": 1.0, + "content": "Quantitatively, the PredNet models again", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 632, + 282, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 282, + 644 + ], + "score": 1.0, + "content": "outperformed the CNN-LSTM Encoder-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 643, + 282, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 282, + 655 + ], + "score": 1.0, + "content": "Decoder. To ensure that the difference in", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 282, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 282, + 666 + ], + "score": 1.0, + "content": "performance was not simply because of the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 666, + 282, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 282, + 678 + ], + "score": 1.0, + "content": "choice of hyperparameters, we trained mod-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 676, + 282, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 282, + 689 + ], + "score": 1.0, + "content": "els with four other sets of hyperparameters,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 687, + 282, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 282, + 700 + ], + "score": 1.0, + "content": "which were sampled from the initial ran-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 282, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 282, + 711 + ], + "score": 1.0, + "content": "dom search over the number of layers, fil-", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 48.5 + }, + { + "type": "table", + "bbox": [ + 289, + 646, + 503, + 707 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 289, + 624, + 504, + 646 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 289, + 623, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 289, + 623, + 505, + 637 + ], + "score": 1.0, + "content": "Table 2: Evaluation of Next-Frame Predictions on", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 290, + 635, + 404, + 647 + ], + "spans": [ + { + "bbox": [ + 290, + 635, + 404, + 647 + ], + "score": 1.0, + "content": "CalTech Pedestrian Dataset.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 53.5 + }, + { + "type": "table_body", + "bbox": [ + 289, + 646, + 503, + 707 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 289, + 646, + 503, + 707 + ], + "spans": [ + { + "bbox": [ + 289, + 646, + 503, + 707 + ], + "score": 0.976, + "html": "
MSESSIM
PredNet Lo3.13 × 10-30.884
PredNet Lall3.33 ×10-30.875
CNN-LSTMEnc.-Dec.3.67 × 10-30.865
Copy Last Frame7.95 × 10-30.762
", + "type": "table", + "image_path": "c30f36c19aa8ec6af0e3d143966f300d0b5a586acb3e306a0d25c2faf3163afe.jpg" + } + ] + } + ], + "index": 57, + "virtual_lines": [ + { + "bbox": [ + 289, + 646, + 503, + 658.2 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 289, + 658.2, + 503, + 670.4000000000001 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 289, + 670.4000000000001, + 503, + 682.6000000000001 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 289, + 682.6000000000001, + 503, + 694.8000000000002 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 289, + 694.8000000000002, + 503, + 707.0000000000002 + ], + "spans": [], + "index": 59 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 710, + 505, + 732 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 473, + 722 + ], + "score": 1.0, + "content": "ter sizes, and number of filters per layer. For each of the four additional sets, the PredNet", + "type": "text" + }, + { + "bbox": [ + 474, + 710, + 486, + 721 + ], + "score": 0.86, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "had", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 342, + 732 + ], + "score": 1.0, + "content": "the best performance, with an average error reduction of", + "type": "text" + }, + { + "bbox": [ + 343, + 720, + 370, + 731 + ], + "score": 0.87, + "content": "1 4 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 720, + 390, + 732 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 390, + 720, + 418, + 731 + ], + "score": 0.87, + "content": "1 4 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "for MSE and SSIM,", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 60.5 + } + ], + "index": 57 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 385, + 95 + ], + "score": 1.0, + "content": "Face classification accuracies using the representations learned by the", + "type": "text" + }, + { + "bbox": [ + 385, + 83, + 398, + 93 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 83, + 415, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 416, + 83, + 434, + 93 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "PredNets, a stan-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "dard autoencoder, and a Ladder Network variant are shown in the right panel of Figure 3. Both", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "PredNets compare favorably to the other models at all sizes of the training set, suggesting they learn", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "a representation that is relatively tolerant to object transformations. Similar to the decoding accu-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 504, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 306, + 139 + ], + "score": 1.0, + "content": "racy of the first principle component, the PredNet", + "type": "text" + }, + { + "bbox": [ + 307, + 127, + 325, + 137 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 127, + 459, + 139 + ], + "score": 1.0, + "content": "model actually outperformed the", + "type": "text" + }, + { + "bbox": [ + 459, + 127, + 471, + 137 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 127, + 504, + 139 + ], + "score": 1.0, + "content": "variant.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 504, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 504, + 149 + ], + "score": 1.0, + "content": "Altogether, these results suggest that predictive training with the PredNet can be a viable alternative", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 160 + ], + "score": 1.0, + "content": "to other models trained with a more traditional reconstructive or denoising loss, and that the relative", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 160, + 412, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 194, + 172 + ], + "score": 1.0, + "content": "layer loss weightings", + "type": "text" + }, + { + "bbox": [ + 194, + 160, + 216, + 171 + ], + "score": 0.32, + "content": "( \\lambda _ { l } ^ { } \\mathbf { \\dot { s } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 160, + 412, + 172 + ], + "score": 1.0, + "content": "may be important for the particular task at hand.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5, + "bbox_fs": [ + 104, + 83, + 505, + 172 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 187, + 258, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 261, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 261, + 201 + ], + "score": 1.0, + "content": "3.2 NATURAL IMAGE SEQUENCES", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "We next sought to test the PredNet architecture on complex, real-world sequences. As a testbed, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "chose car-mounted camera videos, since these videos span across a wide range of settings and are", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "characterized by rich temporal dynamics, including both self-motion of the vehicle and the motion", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "of other objects in the scene (Agrawal et al., 2015). Models were trained using the raw videos from", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "the KITTI dataset (Geiger et al., 2013), which were captured by a roof-mounted camera on a car", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "driving around an urban environment in Germany. Sequences of 10 frames were sampled from the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 273, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 506, + 289 + ], + "score": 1.0, + "content": "“City”, “Residential”, and “Road” categories, with 57 recording sessions used for training and 4", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "used for validation. Frames were center-cropped and downsampled to 128x160 pixels. In total, the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 298, + 289, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 289, + 309 + ], + "score": 1.0, + "content": "training set consisted of roughly 41K frames.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 210, + 506, + 309 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "A random hyperparameter search, with model selection based on the validation set, resulted in a 4", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "layer model with 3x3 convolutions and layer channel sizes of (3, 48, 96, 192). Models were again", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "score": 1.0, + "content": "trained with Adam (Kingma & Ba, 2014) using a loss either solely computed on the lowest layer", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 107, + 347, + 126, + 358 + ], + "score": 0.8, + "content": "( L _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 347, + 403, + 361 + ], + "score": 1.0, + "content": "or with a weight of 1 on the lowest layer and 0.1 on the upper layers", + "type": "text" + }, + { + "bbox": [ + 403, + 347, + 427, + 358 + ], + "score": 0.88, + "content": "( L _ { a l l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 347, + 505, + 361 + ], + "score": 1.0, + "content": ". Adam parameters", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 356, + 264, + 372 + ], + "score": 1.0, + "content": "were initially set to their default values", + "type": "text" + }, + { + "bbox": [ + 265, + 358, + 308, + 369 + ], + "score": 0.83, + "content": "\\mathbf { \\Phi } _ { \\mathrm { { ( } } \\alpha } = 0 . 0 0 1 ", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 356, + 312, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 312, + 358, + 349, + 369 + ], + "score": 0.88, + "content": "\\beta _ { 1 } = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 356, + 353, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 353, + 358, + 402, + 369 + ], + "score": 0.88, + "content": "\\beta _ { 2 } = 0 . 9 9 9 )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 356, + 493, + 372 + ], + "score": 1.0, + "content": "with the learning rate,", + "type": "text" + }, + { + "bbox": [ + 494, + 360, + 501, + 368 + ], + "score": 0.66, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 356, + 506, + 372 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "decreasing by a factor of 10 halfway through training. To assess that the network had indeed learned", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "score": 1.0, + "content": "a robust representation, we tested on the CalTech Pedestrian dataset (Dollar et al., 2009), which´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "consists of videos from a dashboard-mounted camera on a vehicle driving around Los Angeles.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "Testing sequences were made to match the frame rate of the KITTI dataset and again cropped to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "128x160 pixels. Quantitative evaluation was performed on the entire CalTech test partition, split", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 223, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 223, + 436 + ], + "score": 1.0, + "content": "into sequences of 10 frames.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23, + "bbox_fs": [ + 104, + 313, + 506, + 436 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 259, + 452 + ], + "score": 1.0, + "content": "Sample PredNet predictions (for the", + "type": "text" + }, + { + "bbox": [ + 259, + 441, + 272, + 452 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 441, + 504, + 452 + ], + "score": 1.0, + "content": "model) on the CalTech Pedestrian dataset are shown in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "Figure 4, and example videos can be found at https://coxlab.github.io/prednet/. The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 462, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 476 + ], + "score": 1.0, + "content": "model is able to make fairly accurate predictions in a wide range of scenarios. In the top sequence", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "of Fig. 4, a car is passing in the opposite direction, and the model, while not perfect, is able to predict", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 485, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 497 + ], + "score": 1.0, + "content": "its trajectory, as well as fill in the ground it leaves behind. Similarly in Sequence 3, the model is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "able to predict the motion of a vehicle completing a left turn. Sequences 2 and 5 illustrate that the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "PredNet can judge its own movement, as it predicts the appearance of shadows and a stationary", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "vehicle as they approach. The model makes reasonable predictions even in difficult scenarios, such", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "as when the camera-mounted vehicle is turning. In Sequence 4, the model predicts the position of a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 539, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 553 + ], + "score": 1.0, + "content": "tree, as the vehicle turns onto a road. The turning sequences also further illustrate the model’s ability", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "to “fill-in”, as it is able to extrapolate sky and tree textures as unseen regions come into view. As an", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "additional control, we show a sequence at the bottom of Fig. 4, where the input has been temporally", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "scrambled. In this case, the model generates blurry frames, which mostly just resemble the previous", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "frame. Finally, although the PredNet shown here was trained to predict one frame ahead, it is also", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "possible to predict multiple frames into the future, by feeding back predictions as the inputs and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 605, + 326, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 326, + 617 + ], + "score": 1.0, + "content": "recursively iterating. We explore this in Appendix 5.3.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 441, + 506, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 281, + 710 + ], + "lines": [ + { + "bbox": [ + 107, + 621, + 282, + 633 + ], + "spans": [ + { + "bbox": [ + 107, + 621, + 282, + 633 + ], + "score": 1.0, + "content": "Quantitatively, the PredNet models again", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 632, + 282, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 282, + 644 + ], + "score": 1.0, + "content": "outperformed the CNN-LSTM Encoder-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 643, + 282, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 282, + 655 + ], + "score": 1.0, + "content": "Decoder. To ensure that the difference in", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 282, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 282, + 666 + ], + "score": 1.0, + "content": "performance was not simply because of the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 666, + 282, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 282, + 678 + ], + "score": 1.0, + "content": "choice of hyperparameters, we trained mod-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 676, + 282, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 282, + 689 + ], + "score": 1.0, + "content": "els with four other sets of hyperparameters,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 687, + 282, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 282, + 700 + ], + "score": 1.0, + "content": "which were sampled from the initial ran-", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 282, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 282, + 711 + ], + "score": 1.0, + "content": "dom search over the number of layers, fil-", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 621, + 282, + 711 + ] + }, + { + "type": "table", + "bbox": [ + 289, + 646, + 503, + 707 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 289, + 624, + 504, + 646 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 289, + 623, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 289, + 623, + 505, + 637 + ], + "score": 1.0, + "content": "Table 2: Evaluation of Next-Frame Predictions on", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 290, + 635, + 404, + 647 + ], + "spans": [ + { + "bbox": [ + 290, + 635, + 404, + 647 + ], + "score": 1.0, + "content": "CalTech Pedestrian Dataset.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 53.5 + }, + { + "type": "table_body", + "bbox": [ + 289, + 646, + 503, + 707 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 289, + 646, + 503, + 707 + ], + "spans": [ + { + "bbox": [ + 289, + 646, + 503, + 707 + ], + "score": 0.976, + "html": "
MSESSIM
PredNet Lo3.13 × 10-30.884
PredNet Lall3.33 ×10-30.875
CNN-LSTMEnc.-Dec.3.67 × 10-30.865
Copy Last Frame7.95 × 10-30.762
", + "type": "table", + "image_path": "c30f36c19aa8ec6af0e3d143966f300d0b5a586acb3e306a0d25c2faf3163afe.jpg" + } + ] + } + ], + "index": 57, + "virtual_lines": [ + { + "bbox": [ + 289, + 646, + 503, + 658.2 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 289, + 658.2, + 503, + 670.4000000000001 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 289, + 670.4000000000001, + 503, + 682.6000000000001 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 289, + 682.6000000000001, + 503, + 694.8000000000002 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 289, + 694.8000000000002, + 503, + 707.0000000000002 + ], + "spans": [], + "index": 59 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 710, + 505, + 732 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 473, + 722 + ], + "score": 1.0, + "content": "ter sizes, and number of filters per layer. For each of the four additional sets, the PredNet", + "type": "text" + }, + { + "bbox": [ + 474, + 710, + 486, + 721 + ], + "score": 0.86, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "had", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 342, + 732 + ], + "score": 1.0, + "content": "the best performance, with an average error reduction of", + "type": "text" + }, + { + "bbox": [ + 343, + 720, + 370, + 731 + ], + "score": 0.87, + "content": "1 4 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 720, + 390, + 732 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 390, + 720, + 418, + 731 + ], + "score": 0.87, + "content": "1 4 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "for MSE and SSIM,", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 60.5 + } + ], + "index": 57 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 75, + 503, + 549 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 75, + 503, + 549 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 75, + 503, + 549 + ], + "spans": [ + { + "bbox": [ + 108, + 75, + 503, + 549 + ], + "score": 0.968, + "type": "image", + "image_path": "f2465bddecf762950203060719579f704eebd5b1c06d29b0051a317691866e37.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 75, + 503, + 233.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 233.0, + 503, + 391.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 391.0, + 503, + 549.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 308, + 546, + 335, + 554 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 337, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 337, + 556 + ], + "score": 1.0, + "content": "time →", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 107, + 557, + 504, + 591 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 558, + 504, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 504, + 570 + ], + "score": 1.0, + "content": "Figure 4: PredNet predictions for car-cam videos. The first rows contain ground truth and the second", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "rows contain predictions. The sequence below the red line was temporally scrambled. The model", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 579, + 491, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 491, + 591 + ], + "score": 1.0, + "content": "was trained on the KITTI dataset and sequences shown are from the CalTech Pedestrian dataset.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "respectively, compared to the CNN-LSTM Encoder-Decoder. More details, as well as a thorough", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "investigation of systematically simplified models on the continuum between the PredNet and the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "CNN-LSTM Encoder-Decoder can be found in Appendix 5.1. Briefly, the elementwise subtraction", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "score": 1.0, + "content": "operation in the PredNet seems to be beneficial, and the nonlinearity of positive/negative splitting", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "also adds modest improvements. Finally, while these experiments measure the benefits of each com-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "ponent of our model, we also directly compare against recent work in a similar car-cam setting, by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "reporting results on a 64x64 pixel, grayscale car-cam dataset released by Brabandere et al. (2016).", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 416, + 711 + ], + "score": 1.0, + "content": "Our PredNet model outperforms the model by Brabandere et al. (2016) by", + "type": "text" + }, + { + "bbox": [ + 417, + 699, + 436, + 709 + ], + "score": 0.86, + "content": "2 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 699, + 505, + 711 + ], + "score": 1.0, + "content": ". Details can be", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "found in Appendix 5.2. Also in Appendix 5.2, we present results for the Human3.6M (Ionescu", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "et al., 2014) dataset, as reported by Finn et al. (2016). Without re-optimizing hyperparameters, our", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 75, + 503, + 549 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 75, + 503, + 549 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 75, + 503, + 549 + ], + "spans": [ + { + "bbox": [ + 108, + 75, + 503, + 549 + ], + "score": 0.968, + "type": "image", + "image_path": "f2465bddecf762950203060719579f704eebd5b1c06d29b0051a317691866e37.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 75, + 503, + 233.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 233.0, + 503, + 391.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 391.0, + 503, + 549.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 308, + 546, + 335, + 554 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 337, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 337, + 556 + ], + "score": 1.0, + "content": "time →", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 107, + 557, + 504, + 591 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 558, + 504, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 504, + 570 + ], + "score": 1.0, + "content": "Figure 4: PredNet predictions for car-cam videos. The first rows contain ground truth and the second", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "rows contain predictions. The sequence below the red line was temporally scrambled. The model", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 579, + 491, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 491, + 591 + ], + "score": 1.0, + "content": "was trained on the KITTI dataset and sequences shown are from the CalTech Pedestrian dataset.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "respectively, compared to the CNN-LSTM Encoder-Decoder. More details, as well as a thorough", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "investigation of systematically simplified models on the continuum between the PredNet and the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "CNN-LSTM Encoder-Decoder can be found in Appendix 5.1. Briefly, the elementwise subtraction", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "score": 1.0, + "content": "operation in the PredNet seems to be beneficial, and the nonlinearity of positive/negative splitting", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "also adds modest improvements. Finally, while these experiments measure the benefits of each com-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "ponent of our model, we also directly compare against recent work in a similar car-cam setting, by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "reporting results on a 64x64 pixel, grayscale car-cam dataset released by Brabandere et al. (2016).", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 416, + 711 + ], + "score": 1.0, + "content": "Our PredNet model outperforms the model by Brabandere et al. (2016) by", + "type": "text" + }, + { + "bbox": [ + 417, + 699, + 436, + 709 + ], + "score": 0.86, + "content": "2 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 699, + 505, + 711 + ], + "score": 1.0, + "content": ". Details can be", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "found in Appendix 5.2. Also in Appendix 5.2, we present results for the Human3.6M (Ionescu", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "et al., 2014) dataset, as reported by Finn et al. (2016). Without re-optimizing hyperparameters, our", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "model underperforms the concurrently developed DNA model by Finn et al. (2016), but outperforms", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 250, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 250, + 106 + ], + "score": 1.0, + "content": "the model by Mathieu et al. (2016).", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 621, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "model underperforms the concurrently developed DNA model by Finn et al. (2016), but outperforms", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 250, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 250, + 106 + ], + "score": 1.0, + "content": "the model by Mathieu et al. (2016).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 504, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 504, + 123 + ], + "score": 1.0, + "content": "To test the implicit encoding of latent parameters in the car-cam setting, we used the internal rep-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "resentation in the PredNet to estimate the car’s steering angle (Bojarski et al., 2016; Biasini et al.,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "2016). We used a dataset released by Comma.ai (Biasini et al., 2016) consisting of 11 videos total-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "ing about 7 hours of mostly highway driving. We first trained networks for next-frame prediction", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "and then fit a linear fully-connected layer on the learned representation to estimate the steering an-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 308, + 178 + ], + "score": 1.0, + "content": "gle, using a MSE loss. We again concatenate the", + "type": "text" + }, + { + "bbox": [ + 309, + 166, + 320, + 176 + ], + "score": 0.87, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "representation at all layers, but first spatially", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "average pool lower layers to match the spatial size of the upper layer, in order to reduce dimension-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 398, + 199 + ], + "score": 1.0, + "content": "ality. Steering angle estimation results, using the representation on the", + "type": "text" + }, + { + "bbox": [ + 399, + 187, + 416, + 198 + ], + "score": 0.88, + "content": "1 0 ^ { \\mathrm { t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "time step, are shown", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 504, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 491, + 211 + ], + "score": 1.0, + "content": "in Figure 5. Given just 1K labeled training examples, a simple linear readout on the PredNet", + "type": "text" + }, + { + "bbox": [ + 491, + 198, + 504, + 209 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 201, + 222 + ], + "score": 1.0, + "content": "representation explains", + "type": "text" + }, + { + "bbox": [ + 201, + 209, + 221, + 220 + ], + "score": 0.88, + "content": "7 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "of the variance in the steering angle and outperforms the CNN-LSTM", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 104, + 220, + 163, + 234 + ], + "score": 1.0, + "content": "Enc.-Dec. by", + "type": "text" + }, + { + "bbox": [ + 163, + 221, + 183, + 232 + ], + "score": 0.87, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 220, + 389, + 234 + ], + "score": 1.0, + "content": ". With 25K labeled training examples, the PredNet", + "type": "text" + }, + { + "bbox": [ + 389, + 221, + 402, + 232 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "has a MSE (in degrees2)", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "score": 1.0, + "content": "of 2.14. As a point of reference, a CNN model designed to predict the steering angle (Biasini", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 466, + 256 + ], + "score": 1.0, + "content": "et al., 2016), albeit from a single frame instead of multiple frames, achieve a MSE of", + "type": "text" + }, + { + "bbox": [ + 467, + 244, + 479, + 253 + ], + "score": 0.64, + "content": "{ \\sim } 4", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 243, + 506, + 256 + ], + "score": 1.0, + "content": "when", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "trained end-to-end using 396K labeled training examples. Details of this analysis can be found in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 504, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 323, + 279 + ], + "score": 1.0, + "content": "Appendix 8. Interestingly, in this task, the PredNet", + "type": "text" + }, + { + "bbox": [ + 324, + 265, + 342, + 276 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 263, + 491, + 279 + ], + "score": 1.0, + "content": "model actually underperformed the", + "type": "text" + }, + { + "bbox": [ + 491, + 265, + 504, + 276 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 462, + 289 + ], + "score": 1.0, + "content": "model and slightly underperformed the CNN-LSTM Enc.-Dec, again suggesting that the", + "type": "text" + }, + { + "bbox": [ + 463, + 276, + 473, + 287 + ], + "score": 0.86, + "content": "\\lambda _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "param-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 288, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 299 + ], + "score": 1.0, + "content": "eter can affect the representation learned, and different values may be preferable in different end", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 273, + 311 + ], + "score": 1.0, + "content": "tasks. Nonetheless, the readout from the", + "type": "text" + }, + { + "bbox": [ + 273, + 298, + 292, + 309 + ], + "score": 0.91, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "model still explained a substantial proportion of the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "steering angle variance and strongly outperformed the random initial weights. Overall, this anal-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "ysis again demonstrates that a representation learned through prediction, and particularly with the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 330, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 506, + 345 + ], + "score": 1.0, + "content": "PredNet model with appropriate hyperparameters, can contain useful information about underlying", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 342, + 180, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 180, + 355 + ], + "score": 1.0, + "content": "latent parameters.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 12.5 + }, + { + "type": "image", + "bbox": [ + 107, + 371, + 498, + 540 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 371, + 498, + 540 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 371, + 498, + 540 + ], + "spans": [ + { + "bbox": [ + 107, + 371, + 498, + 540 + ], + "score": 0.968, + "type": "image", + "image_path": "4eacc964ba30df01ba5dfac8e925e09b97a2e2cd91639d41fa77b3ced7bfe30a.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 107, + 371, + 498, + 427.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 107, + 427.3333333333333, + 498, + 483.66666666666663 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 107, + 483.66666666666663, + 498, + 540.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 543, + 505, + 598 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "Figure 5: Steering angle estimation accuracy on the Comma.ai dataset (Biasini et al., 2016). Left:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 554, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 566 + ], + "score": 1.0, + "content": "Example steering angle curve with model estimations for a segment in the test set. Decoding was", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "score": 1.0, + "content": "performed using a fully-connected readout on the PredNet representation trained with 25K labeled", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "training examples. PredNet representation was trained for next-frame prediction on Comma.ai train-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 586, + 363, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 363, + 600 + ], + "score": 1.0, + "content": "ing set. Right: Mean-squared error of steering angle estimation.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "index": 27.0 + }, + { + "type": "title", + "bbox": [ + 107, + 609, + 190, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 192, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 192, + 624 + ], + "score": 1.0, + "content": "4 DISCUSSION", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "Above, we have demonstrated a predictive coding inspired architecture that is able to predict future", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "frames in both synthetic and natural image sequences. Importantly, we have shown that learning to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "predict how an object or scene will move in a future frame confers advantages in decoding latent", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 665, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 680 + ], + "score": 1.0, + "content": "parameters (such as viewing angle) that give rise to an object’s appearance, and can improve recog-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "nition performance. More generally, we argue that prediction can serve as a powerful unsupervised", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "learning signal, since accurately predicting future frames requires at least an implicit model of the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "objects that make up the scene and how they are allowed to move. Developing a deeper understand-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "ing of the nature of the representations learned by the networks, and extending the architecture, by,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 363, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 363, + 733 + ], + "score": 1.0, + "content": "for instance, allowing sampling, are important future directions.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 504, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 504, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 504, + 123 + ], + "score": 1.0, + "content": "To test the implicit encoding of latent parameters in the car-cam setting, we used the internal rep-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "resentation in the PredNet to estimate the car’s steering angle (Bojarski et al., 2016; Biasini et al.,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "2016). We used a dataset released by Comma.ai (Biasini et al., 2016) consisting of 11 videos total-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "ing about 7 hours of mostly highway driving. We first trained networks for next-frame prediction", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "and then fit a linear fully-connected layer on the learned representation to estimate the steering an-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 308, + 178 + ], + "score": 1.0, + "content": "gle, using a MSE loss. We again concatenate the", + "type": "text" + }, + { + "bbox": [ + 309, + 166, + 320, + 176 + ], + "score": 0.87, + "content": "R _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "representation at all layers, but first spatially", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "average pool lower layers to match the spatial size of the upper layer, in order to reduce dimension-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 398, + 199 + ], + "score": 1.0, + "content": "ality. Steering angle estimation results, using the representation on the", + "type": "text" + }, + { + "bbox": [ + 399, + 187, + 416, + 198 + ], + "score": 0.88, + "content": "1 0 ^ { \\mathrm { t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "time step, are shown", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 504, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 491, + 211 + ], + "score": 1.0, + "content": "in Figure 5. Given just 1K labeled training examples, a simple linear readout on the PredNet", + "type": "text" + }, + { + "bbox": [ + 491, + 198, + 504, + 209 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 201, + 222 + ], + "score": 1.0, + "content": "representation explains", + "type": "text" + }, + { + "bbox": [ + 201, + 209, + 221, + 220 + ], + "score": 0.88, + "content": "7 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "of the variance in the steering angle and outperforms the CNN-LSTM", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 104, + 220, + 163, + 234 + ], + "score": 1.0, + "content": "Enc.-Dec. by", + "type": "text" + }, + { + "bbox": [ + 163, + 221, + 183, + 232 + ], + "score": 0.87, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 220, + 389, + 234 + ], + "score": 1.0, + "content": ". With 25K labeled training examples, the PredNet", + "type": "text" + }, + { + "bbox": [ + 389, + 221, + 402, + 232 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "has a MSE (in degrees2)", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "score": 1.0, + "content": "of 2.14. As a point of reference, a CNN model designed to predict the steering angle (Biasini", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 466, + 256 + ], + "score": 1.0, + "content": "et al., 2016), albeit from a single frame instead of multiple frames, achieve a MSE of", + "type": "text" + }, + { + "bbox": [ + 467, + 244, + 479, + 253 + ], + "score": 0.64, + "content": "{ \\sim } 4", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 243, + 506, + 256 + ], + "score": 1.0, + "content": "when", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "trained end-to-end using 396K labeled training examples. Details of this analysis can be found in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 504, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 323, + 279 + ], + "score": 1.0, + "content": "Appendix 8. Interestingly, in this task, the PredNet", + "type": "text" + }, + { + "bbox": [ + 324, + 265, + 342, + 276 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 263, + 491, + 279 + ], + "score": 1.0, + "content": "model actually underperformed the", + "type": "text" + }, + { + "bbox": [ + 491, + 265, + 504, + 276 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 462, + 289 + ], + "score": 1.0, + "content": "model and slightly underperformed the CNN-LSTM Enc.-Dec, again suggesting that the", + "type": "text" + }, + { + "bbox": [ + 463, + 276, + 473, + 287 + ], + "score": 0.86, + "content": "\\lambda _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "param-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 288, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 299 + ], + "score": 1.0, + "content": "eter can affect the representation learned, and different values may be preferable in different end", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 273, + 311 + ], + "score": 1.0, + "content": "tasks. Nonetheless, the readout from the", + "type": "text" + }, + { + "bbox": [ + 273, + 298, + 292, + 309 + ], + "score": 0.91, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "model still explained a substantial proportion of the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "steering angle variance and strongly outperformed the random initial weights. Overall, this anal-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "ysis again demonstrates that a representation learned through prediction, and particularly with the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 330, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 506, + 345 + ], + "score": 1.0, + "content": "PredNet model with appropriate hyperparameters, can contain useful information about underlying", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 342, + 180, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 180, + 355 + ], + "score": 1.0, + "content": "latent parameters.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 12.5, + "bbox_fs": [ + 104, + 110, + 506, + 355 + ] + }, + { + "type": "image", + "bbox": [ + 107, + 371, + 498, + 540 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 371, + 498, + 540 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 371, + 498, + 540 + ], + "spans": [ + { + "bbox": [ + 107, + 371, + 498, + 540 + ], + "score": 0.968, + "type": "image", + "image_path": "4eacc964ba30df01ba5dfac8e925e09b97a2e2cd91639d41fa77b3ced7bfe30a.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 107, + 371, + 498, + 427.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 107, + 427.3333333333333, + 498, + 483.66666666666663 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 107, + 483.66666666666663, + 498, + 540.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 543, + 505, + 598 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "Figure 5: Steering angle estimation accuracy on the Comma.ai dataset (Biasini et al., 2016). Left:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 554, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 566 + ], + "score": 1.0, + "content": "Example steering angle curve with model estimations for a segment in the test set. Decoding was", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "score": 1.0, + "content": "performed using a fully-connected readout on the PredNet representation trained with 25K labeled", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "training examples. PredNet representation was trained for next-frame prediction on Comma.ai train-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 586, + 363, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 363, + 600 + ], + "score": 1.0, + "content": "ing set. Right: Mean-squared error of steering angle estimation.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "index": 27.0 + }, + { + "type": "title", + "bbox": [ + 107, + 609, + 190, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 192, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 192, + 624 + ], + "score": 1.0, + "content": "4 DISCUSSION", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "Above, we have demonstrated a predictive coding inspired architecture that is able to predict future", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "frames in both synthetic and natural image sequences. Importantly, we have shown that learning to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "predict how an object or scene will move in a future frame confers advantages in decoding latent", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 665, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 680 + ], + "score": 1.0, + "content": "parameters (such as viewing angle) that give rise to an object’s appearance, and can improve recog-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "nition performance. More generally, we argue that prediction can serve as a powerful unsupervised", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "learning signal, since accurately predicting future frames requires at least an implicit model of the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "objects that make up the scene and how they are allowed to move. Developing a deeper understand-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "ing of the nature of the representations learned by the networks, and extending the architecture, by,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 363, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 363, + 733 + ], + "score": 1.0, + "content": "for instance, allowing sampling, are important future directions.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 633, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 200, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 83, + 200, + 94 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 200, + 94 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 101, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 505, + 114 + ], + "score": 1.0, + "content": "We would like to thank Rasmus Berg Palm for fruitful discussions and early brainstorming. We", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "would also like to thank the developers of Keras (Chollet, 2016). This work was supported by IARPA", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "(contract D16PC00002), the National Science Foundation (NSF IIS 1409097), and the Center for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 397, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 397, + 147 + ], + "score": 1.0, + "content": "Brains, Minds and Machines (CBMM, NSF STC award CCF-1231216).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 162, + 176, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 176, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 176, + 176 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 105, + 180, + 476, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 478, + 194 + ], + "score": 1.0, + "content": "Pulkit Agrawal, Joao Carreira, and Jitendra Malik. Learning to see by moving. ˜ CoRR, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 199, + 503, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 212 + ], + "score": 1.0, + "content": "Andre M. Bastos, W. Martin Usrey, Rick A. Adams, George R. Mangun, Pascal Fries, and Karl J.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 209, + 397, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 397, + 222 + ], + "score": 1.0, + "content": "Friston. Canonical microcircuits for predictive coding. Neuron, 2012.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 504, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 243 + ], + "score": 1.0, + "content": "Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 240, + 342, + 252 + ], + "spans": [ + { + "bbox": [ + 115, + 240, + 342, + 252 + ], + "score": 1.0, + "content": "prediction with recurrent neural networks. CoRR, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 503, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 272 + ], + "score": 1.0, + "content": "Yoshua Bengio. How auto-encoders could provide credit assignment in deep networks via target", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 271, + 224, + 281 + ], + "spans": [ + { + "bbox": [ + 115, + 271, + 224, + 281 + ], + "score": 1.0, + "content": "propagation. CoRR, 2014.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 288, + 504, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "Riccardo Biasini, George Hotz, Sam Khalandovsky, Eder Santana, and Niel van der Westhuizen.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 298, + 461, + 312 + ], + "spans": [ + { + "bbox": [ + 116, + 298, + 461, + 312 + ], + "score": 1.0, + "content": "Comma.ai research, 2016. URL https://github.com/commaai/research.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 506, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 327, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 327, + 506, + 342 + ], + "score": 1.0, + "content": "Goyal, Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 117, + 340, + 406, + 351 + ], + "spans": [ + { + "bbox": [ + 117, + 340, + 406, + 351 + ], + "score": 1.0, + "content": "and Karol Zieba. End to end learning for self-driving cars. CoRR, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "Bert De Brabandere, Xu Jia, Tinne Tuytelaars, and Luc Van Gool. Dynamic filter networks. CoRR,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 367, + 143, + 381 + ], + "spans": [ + { + "bbox": [ + 114, + 367, + 143, + 381 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 109, + 387, + 459, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 386, + 459, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 459, + 401 + ], + "score": 1.0, + "content": "Rakesh Chalasani and Jose C. Principe. Deep predictive coding networks. CoRR, 2013.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 406, + 373, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 374, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 374, + 420 + ], + "score": 1.0, + "content": "Franc¸ois Chollet. Comma.ai, 2016. URL http://keras.io/.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 424, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 504, + 438 + ], + "score": 1.0, + "content": "Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 436, + 270, + 447 + ], + "spans": [ + { + "bbox": [ + 116, + 436, + 270, + 447 + ], + "score": 1.0, + "content": "Behavioral and Brain Sciences, 2013.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 504, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "Piotr Dollar, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark.´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 465, + 182, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 465, + 182, + 477 + ], + "score": 1.0, + "content": "In CVPR, 2009.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 484, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "Tobias Egner, Jim M. Monti, and Christopher Summerfield. Expectation and surprise determine", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 495, + 416, + 506 + ], + "spans": [ + { + "bbox": [ + 115, + 495, + 416, + 506 + ], + "score": 1.0, + "content": "neural population responses in the ventral visual stream. J Neurosci, 2010.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 502, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 527 + ], + "score": 1.0, + "content": "Chelsea Finn, Ian J. Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 524, + 275, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 275, + 537 + ], + "score": 1.0, + "content": "through video prediction. CoRR, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 488, + 555 + ], + "lines": [ + { + "bbox": [ + 104, + 541, + 489, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 541, + 489, + 558 + ], + "score": 1.0, + "content": "Peter Foldi ¨ ak. Learning invariance from transformation sequences. ´ Neural Computation, 1991.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 561, + 458, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 458, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 458, + 574 + ], + "score": 1.0, + "content": "Karl Friston. A theory of cortical responses. Philos Trans R Soc Lond B Biol Sci, 2005.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 503, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "score": 1.0, + "content": "Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 590, + 399, + 604 + ], + "spans": [ + { + "bbox": [ + 116, + 590, + 399, + 604 + ], + "score": 1.0, + "content": "kitti dataset. International Journal of Robotics Research (IJRR), 2013.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 504, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "Dileep George and Jeff Hawkins. A hierarchical bayesian model of invariant pattern recognition", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 621, + 504, + 633 + ], + "spans": [ + { + "bbox": [ + 115, + 621, + 504, + 633 + ], + "score": 1.0, + "content": "in the visual cortex. In Proceedings of the International Joint Conference on Neural Networks.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 632, + 168, + 643 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 168, + 643 + ], + "score": 1.0, + "content": "IEEE, 2005.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 503, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 661, + 447, + 674 + ], + "spans": [ + { + "bbox": [ + 116, + 661, + 447, + 674 + ], + "score": 1.0, + "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS. 2014.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "Ross Goroshin, Joan Bruna, Jonathan Tompson, David Eigen, and Yann LeCun. Unsupervised", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 691, + 362, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 362, + 703 + ], + "score": 1.0, + "content": "learning of spatiotemporally coherent metrics. CoRR, 2015a.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Ross Goroshin, Michael Mathieu, and Yann LeCun. Learning to linearize under uncertainty. ¨ CoRR,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 719, + 147, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 147, + 732 + ], + "score": 1.0, + "content": "2015b.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 200, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 83, + 200, + 94 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 200, + 94 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 101, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 505, + 114 + ], + "score": 1.0, + "content": "We would like to thank Rasmus Berg Palm for fruitful discussions and early brainstorming. We", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "would also like to thank the developers of Keras (Chollet, 2016). This work was supported by IARPA", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "(contract D16PC00002), the National Science Foundation (NSF IIS 1409097), and the Center for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 397, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 397, + 147 + ], + "score": 1.0, + "content": "Brains, Minds and Machines (CBMM, NSF STC award CCF-1231216).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 101, + 505, + 147 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 162, + 176, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 176, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 176, + 176 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 105, + 180, + 476, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 478, + 194 + ], + "score": 1.0, + "content": "Pulkit Agrawal, Joao Carreira, and Jitendra Malik. Learning to see by moving. ˜ CoRR, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 180, + 478, + 194 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 199, + 503, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 212 + ], + "score": 1.0, + "content": "Andre M. Bastos, W. Martin Usrey, Rick A. Adams, George R. Mangun, Pascal Fries, and Karl J.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 209, + 397, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 209, + 397, + 222 + ], + "score": 1.0, + "content": "Friston. Canonical microcircuits for predictive coding. Neuron, 2012.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 199, + 505, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 504, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 243 + ], + "score": 1.0, + "content": "Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 240, + 342, + 252 + ], + "spans": [ + { + "bbox": [ + 115, + 240, + 342, + 252 + ], + "score": 1.0, + "content": "prediction with recurrent neural networks. CoRR, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 227, + 505, + 252 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 503, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 272 + ], + "score": 1.0, + "content": "Yoshua Bengio. How auto-encoders could provide credit assignment in deep networks via target", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 271, + 224, + 281 + ], + "spans": [ + { + "bbox": [ + 115, + 271, + 224, + 281 + ], + "score": 1.0, + "content": "propagation. CoRR, 2014.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 257, + 505, + 281 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 288, + 504, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "Riccardo Biasini, George Hotz, Sam Khalandovsky, Eder Santana, and Niel van der Westhuizen.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 298, + 461, + 312 + ], + "spans": [ + { + "bbox": [ + 116, + 298, + 461, + 312 + ], + "score": 1.0, + "content": "Comma.ai research, 2016. URL https://github.com/commaai/research.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 288, + 505, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 506, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 327, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 115, + 327, + 506, + 342 + ], + "score": 1.0, + "content": "Goyal, Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 117, + 340, + 406, + 351 + ], + "spans": [ + { + "bbox": [ + 117, + 340, + 406, + 351 + ], + "score": 1.0, + "content": "and Karol Zieba. End to end learning for self-driving cars. CoRR, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 316, + 506, + 351 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "Bert De Brabandere, Xu Jia, Tinne Tuytelaars, and Luc Van Gool. Dynamic filter networks. CoRR,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 367, + 143, + 381 + ], + "spans": [ + { + "bbox": [ + 114, + 367, + 143, + 381 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 358, + 505, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 387, + 459, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 386, + 459, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 459, + 401 + ], + "score": 1.0, + "content": "Rakesh Chalasani and Jose C. Principe. Deep predictive coding networks. CoRR, 2013.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 386, + 459, + 401 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 406, + 373, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 374, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 374, + 420 + ], + "score": 1.0, + "content": "Franc¸ois Chollet. Comma.ai, 2016. URL http://keras.io/.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 405, + 374, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 424, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 504, + 438 + ], + "score": 1.0, + "content": "Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 436, + 270, + 447 + ], + "spans": [ + { + "bbox": [ + 116, + 436, + 270, + 447 + ], + "score": 1.0, + "content": "Behavioral and Brain Sciences, 2013.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 423, + 504, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 504, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "Piotr Dollar, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark.´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 465, + 182, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 465, + 182, + 477 + ], + "score": 1.0, + "content": "In CVPR, 2009.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 453, + 504, + 477 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 484, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "Tobias Egner, Jim M. Monti, and Christopher Summerfield. Expectation and surprise determine", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 495, + 416, + 506 + ], + "spans": [ + { + "bbox": [ + 115, + 495, + 416, + 506 + ], + "score": 1.0, + "content": "neural population responses in the ventral visual stream. J Neurosci, 2010.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 483, + 505, + 506 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 502, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 527 + ], + "score": 1.0, + "content": "Chelsea Finn, Ian J. Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 524, + 275, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 275, + 537 + ], + "score": 1.0, + "content": "through video prediction. CoRR, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 513, + 505, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 488, + 555 + ], + "lines": [ + { + "bbox": [ + 104, + 541, + 489, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 541, + 489, + 558 + ], + "score": 1.0, + "content": "Peter Foldi ¨ ak. Learning invariance from transformation sequences. ´ Neural Computation, 1991.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 541, + 489, + 558 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 561, + 458, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 458, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 458, + 574 + ], + "score": 1.0, + "content": "Karl Friston. A theory of cortical responses. Philos Trans R Soc Lond B Biol Sci, 2005.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 561, + 458, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 503, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "score": 1.0, + "content": "Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 590, + 399, + 604 + ], + "spans": [ + { + "bbox": [ + 116, + 590, + 399, + 604 + ], + "score": 1.0, + "content": "kitti dataset. International Journal of Robotics Research (IJRR), 2013.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 580, + 505, + 604 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 504, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "Dileep George and Jeff Hawkins. A hierarchical bayesian model of invariant pattern recognition", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 621, + 504, + 633 + ], + "spans": [ + { + "bbox": [ + 115, + 621, + 504, + 633 + ], + "score": 1.0, + "content": "in the visual cortex. In Proceedings of the International Joint Conference on Neural Networks.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 632, + 168, + 643 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 168, + 643 + ], + "score": 1.0, + "content": "IEEE, 2005.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 610, + 505, + 643 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 503, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 661, + 447, + 674 + ], + "spans": [ + { + "bbox": [ + 116, + 661, + 447, + 674 + ], + "score": 1.0, + "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS. 2014.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 650, + 505, + 674 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "Ross Goroshin, Joan Bruna, Jonathan Tompson, David Eigen, and Yann LeCun. Unsupervised", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 691, + 362, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 362, + 703 + ], + "score": 1.0, + "content": "learning of spatiotemporally coherent metrics. CoRR, 2015a.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 680, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Ross Goroshin, Michael Mathieu, and Yann LeCun. Learning to linearize under uncertainty. ¨ CoRR,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 719, + 147, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 147, + 732 + ], + "score": 1.0, + "content": "2015b.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 495, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 497, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 497, + 96 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. Neural Computation, 1997.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 101, + 503, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 114 + ], + "score": 1.0, + "content": "Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 116, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 124, + 329, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 124, + 329, + 136 + ], + "score": 1.0, + "content": "on Pattern Analysis and Machine Intelligence, 2014.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 105, + 143, + 504, + 167 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "Kevin Jarrett, Koray Kavukcuoglu, MarcAurelio Ranzato, and Yann LeCun. What is the best multi-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 155, + 347, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 347, + 167 + ], + "score": 1.0, + "content": "stage architecture for object recognition? In ICCV. 2009.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 502, + 198 + ], + "lines": [ + { + "bbox": [ + 105, + 175, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 504, + 187 + ], + "score": 1.0, + "content": "Ryota Kanai, Yutaka Komura, Stewart Shipp, and Karl Friston. Cerebral hierarchies : predictive", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 186, + 447, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 447, + 197 + ], + "score": 1.0, + "content": "processing , precision and the pulvinar. Philos Trans R Soc Lond B Biol Sci, 2015.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 205, + 489, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 205, + 489, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 489, + 219 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, 2014.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 502, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 504, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 504, + 238 + ], + "score": 1.0, + "content": "Tejas D. Kulkarni, Will Whitney, Pushmeet Kohli, and Joshua B. Tenenbaum. Deep convolutional", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 236, + 276, + 248 + ], + "spans": [ + { + "bbox": [ + 116, + 236, + 276, + 248 + ], + "score": 1.0, + "content": "inverse graphics network. CoRR, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 255, + 504, + 279 + ], + "lines": [ + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "score": 1.0, + "content": "Tai Sing Lee and David Mumford. Hierarchical bayesian inference in the visual cortex. J Opt Soc", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 267, + 242, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 267, + 242, + 280 + ], + "score": 1.0, + "content": "Am A Opt Image Sci Vis, 2003.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 504, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 302 + ], + "score": 1.0, + "content": "William Lotter, Gabriel Kreiman, and David Cox. Unsupervised learning of visual structure using", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 298, + 298, + 310 + ], + "spans": [ + { + "bbox": [ + 115, + 298, + 298, + 310 + ], + "score": 1.0, + "content": "predictive generative networks. CoRR, 2015.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "Davide Maltoni and Vincenzo Lomonaco. Semi-supervised tuning from temporal coherence. CoRR,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 327, + 143, + 341 + ], + "spans": [ + { + "bbox": [ + 114, + 327, + 143, + 341 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 105, + 348, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond ¨", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 360, + 245, + 371 + ], + "spans": [ + { + "bbox": [ + 115, + 360, + 245, + 371 + ], + "score": 1.0, + "content": "mean square error. ICLR, 2016.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "score": 1.0, + "content": "Vincent Michalski, Roland Memisevic, and Kishore Konda. Modeling deep temporal dependencies", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 391, + 308, + 402 + ], + "spans": [ + { + "bbox": [ + 116, + 391, + 308, + 402 + ], + "score": 1.0, + "content": "with recurrent ”grammar cells”. In NIPS. 2014.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 104, + 408, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 104, + 408, + 506, + 424 + ], + "score": 1.0, + "content": "Hossein Mohabi, Ronan Collobert, and Jason Weston. Deep learning from temporal coherence in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 421, + 209, + 432 + ], + "spans": [ + { + "bbox": [ + 116, + 421, + 209, + 432 + ], + "score": 1.0, + "content": "video. In ICML. 2009.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 503, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 504, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 504, + 453 + ], + "score": 1.0, + "content": "Junhyuk Oh, Xiaoxiao Guo, Honglak Lee, Richard L. Lewis, and Satinder P. Singh. Action-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 452, + 432, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 452, + 432, + 464 + ], + "score": 1.0, + "content": "conditional video prediction using deep networks in atari games. CoRR, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 471, + 503, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "score": 1.0, + "content": "Randall C. O’Reilly, Dean Wyatte, and John Rohrlich. Learning through time in the thalamocortical", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 483, + 199, + 495 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 199, + 495 + ], + "score": 1.0, + "content": "loops. CoRR, 2014.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 503, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "Rasmus Berg Palm. Prediction as a candidate for learning deep hierarchical models of data. Master’s", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 514, + 306, + 525 + ], + "spans": [ + { + "bbox": [ + 116, + 514, + 306, + 525 + ], + "score": 1.0, + "content": "thesis, Technical University of Denmark, 2012.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 504, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "Viorica Patraucean, Ankur Handa, and Roberto Cipolla. Spatio-temporal video autoencoder with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 117, + 545, + 264, + 556 + ], + "spans": [ + { + "bbox": [ + 117, + 545, + 264, + 556 + ], + "score": 1.0, + "content": "differentiable memory. CoRR, 2015.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 504, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 504, + 577 + ], + "score": 1.0, + "content": "Mohammad Pezeshki, Linxi Fan, Philemon Brakel, Aaron C. Courville, and Yoshua Bengio. De-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 576, + 353, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 576, + 353, + 587 + ], + "score": 1.0, + "content": "constructing the ladder network architecture. CoRR, 2015.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 593, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 608 + ], + "score": 1.0, + "content": "Nicolas Pinto, David Doukhan, James J. DiCarlo, and David D. Cox. A high-throughput screening", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 115, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "approach to discovering good forms of biologically inspired visual representation. PLoS Comput", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 616, + 164, + 629 + ], + "spans": [ + { + "bbox": [ + 115, + 616, + 164, + 629 + ], + "score": 1.0, + "content": "Biol, 2009.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 636, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "Marc’Aurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit ¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 647, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 506, + 661 + ], + "score": 1.0, + "content": "Chopra. Video (language) modeling: a baseline for generative models of natural videos. CoRR,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 658, + 142, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 142, + 670 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 104, + 678, + 503, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 504, + 691 + ], + "score": 1.0, + "content": "Rajesh P. N. Rao and Dana H. Ballard. Predictive coding in the visual cortex: a functional interpre-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 690, + 441, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 441, + 701 + ], + "score": 1.0, + "content": "tation of some extra-classical receptive-field effects. Nature Neuroscience, 1999.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 504, + 721 + ], + "score": 1.0, + "content": "Rajesh P. N. Rao and T. J. Sejnowski. Predictive sequence learning in recurrent neocortical circuits.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 720, + 167, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 167, + 732 + ], + "score": 1.0, + "content": "NIPS, 2000.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 495, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 497, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 497, + 96 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. Neural Computation, 1997.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 497, + 96 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 101, + 503, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 114 + ], + "score": 1.0, + "content": "Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 113, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 116, + 113, + 505, + 126 + ], + "score": 1.0, + "content": "datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 124, + 329, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 124, + 329, + 136 + ], + "score": 1.0, + "content": "on Pattern Analysis and Machine Intelligence, 2014.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 102, + 505, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 143, + 504, + 167 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "Kevin Jarrett, Koray Kavukcuoglu, MarcAurelio Ranzato, and Yann LeCun. What is the best multi-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 155, + 347, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 347, + 167 + ], + "score": 1.0, + "content": "stage architecture for object recognition? In ICCV. 2009.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 144, + 505, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 502, + 198 + ], + "lines": [ + { + "bbox": [ + 105, + 175, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 504, + 187 + ], + "score": 1.0, + "content": "Ryota Kanai, Yutaka Komura, Stewart Shipp, and Karl Friston. Cerebral hierarchies : predictive", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 186, + 447, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 447, + 197 + ], + "score": 1.0, + "content": "processing , precision and the pulvinar. Philos Trans R Soc Lond B Biol Sci, 2015.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 175, + 504, + 197 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 205, + 489, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 205, + 489, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 489, + 219 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, 2014.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 205, + 489, + 219 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 502, + 248 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 504, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 504, + 238 + ], + "score": 1.0, + "content": "Tejas D. Kulkarni, Will Whitney, Pushmeet Kohli, and Joshua B. Tenenbaum. Deep convolutional", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 236, + 276, + 248 + ], + "spans": [ + { + "bbox": [ + 116, + 236, + 276, + 248 + ], + "score": 1.0, + "content": "inverse graphics network. CoRR, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 225, + 504, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 255, + 504, + 279 + ], + "lines": [ + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "score": 1.0, + "content": "Tai Sing Lee and David Mumford. Hierarchical bayesian inference in the visual cortex. J Opt Soc", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 267, + 242, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 267, + 242, + 280 + ], + "score": 1.0, + "content": "Am A Opt Image Sci Vis, 2003.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 255, + 505, + 280 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 504, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 302 + ], + "score": 1.0, + "content": "William Lotter, Gabriel Kreiman, and David Cox. Unsupervised learning of visual structure using", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 298, + 298, + 310 + ], + "spans": [ + { + "bbox": [ + 115, + 298, + 298, + 310 + ], + "score": 1.0, + "content": "predictive generative networks. CoRR, 2015.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 285, + 506, + 310 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "Davide Maltoni and Vincenzo Lomonaco. Semi-supervised tuning from temporal coherence. CoRR,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 327, + 143, + 341 + ], + "spans": [ + { + "bbox": [ + 114, + 327, + 143, + 341 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 317, + 505, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 348, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond ¨", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 360, + 245, + 371 + ], + "spans": [ + { + "bbox": [ + 115, + 360, + 245, + 371 + ], + "score": 1.0, + "content": "mean square error. ICLR, 2016.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 348, + 505, + 371 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "score": 1.0, + "content": "Vincent Michalski, Roland Memisevic, and Kishore Konda. Modeling deep temporal dependencies", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 391, + 308, + 402 + ], + "spans": [ + { + "bbox": [ + 116, + 391, + 308, + 402 + ], + "score": 1.0, + "content": "with recurrent ”grammar cells”. In NIPS. 2014.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 378, + 505, + 402 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 104, + 408, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 104, + 408, + 506, + 424 + ], + "score": 1.0, + "content": "Hossein Mohabi, Ronan Collobert, and Jason Weston. Deep learning from temporal coherence in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 421, + 209, + 432 + ], + "spans": [ + { + "bbox": [ + 116, + 421, + 209, + 432 + ], + "score": 1.0, + "content": "video. In ICML. 2009.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 408, + 506, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 503, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 504, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 504, + 453 + ], + "score": 1.0, + "content": "Junhyuk Oh, Xiaoxiao Guo, Honglak Lee, Richard L. Lewis, and Satinder P. Singh. Action-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 452, + 432, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 452, + 432, + 464 + ], + "score": 1.0, + "content": "conditional video prediction using deep networks in atari games. CoRR, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 441, + 504, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 471, + 503, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "score": 1.0, + "content": "Randall C. O’Reilly, Dean Wyatte, and John Rohrlich. Learning through time in the thalamocortical", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 483, + 199, + 495 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 199, + 495 + ], + "score": 1.0, + "content": "loops. CoRR, 2014.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 471, + 505, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 503, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "Rasmus Berg Palm. Prediction as a candidate for learning deep hierarchical models of data. Master’s", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 514, + 306, + 525 + ], + "spans": [ + { + "bbox": [ + 116, + 514, + 306, + 525 + ], + "score": 1.0, + "content": "thesis, Technical University of Denmark, 2012.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 502, + 505, + 525 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 504, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "Viorica Patraucean, Ankur Handa, and Roberto Cipolla. Spatio-temporal video autoencoder with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 117, + 545, + 264, + 556 + ], + "spans": [ + { + "bbox": [ + 117, + 545, + 264, + 556 + ], + "score": 1.0, + "content": "differentiable memory. CoRR, 2015.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 533, + 505, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 504, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 504, + 577 + ], + "score": 1.0, + "content": "Mohammad Pezeshki, Linxi Fan, Philemon Brakel, Aaron C. Courville, and Yoshua Bengio. De-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 576, + 353, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 576, + 353, + 587 + ], + "score": 1.0, + "content": "constructing the ladder network architecture. CoRR, 2015.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 563, + 504, + 587 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 593, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 608 + ], + "score": 1.0, + "content": "Nicolas Pinto, David Doukhan, James J. DiCarlo, and David D. Cox. A high-throughput screening", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 115, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "approach to discovering good forms of biologically inspired visual representation. PLoS Comput", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 616, + 164, + 629 + ], + "spans": [ + { + "bbox": [ + 115, + 616, + 164, + 629 + ], + "score": 1.0, + "content": "Biol, 2009.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 593, + 505, + 629 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 636, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "Marc’Aurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit ¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 647, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 506, + 661 + ], + "score": 1.0, + "content": "Chopra. Video (language) modeling: a baseline for generative models of natural videos. CoRR,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 658, + 142, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 142, + 670 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 636, + 506, + 670 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 678, + 503, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 504, + 691 + ], + "score": 1.0, + "content": "Rajesh P. N. Rao and Dana H. Ballard. Predictive coding in the visual cortex: a functional interpre-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 690, + 441, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 441, + 701 + ], + "score": 1.0, + "content": "tation of some extra-classical receptive-field effects. Nature Neuroscience, 1999.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 677, + 504, + 701 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 504, + 721 + ], + "score": 1.0, + "content": "Rajesh P. N. Rao and T. J. Sejnowski. Predictive sequence learning in recurrent neocortical circuits.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 720, + 167, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 167, + 732 + ], + "score": 1.0, + "content": "NIPS, 2000.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 710, + 504, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, and Tapani Raiko. Semi-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 337, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 337, + 106 + ], + "score": 1.0, + "content": "supervised learning with ladder network. CoRR, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 409, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 409, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 409, + 126 + ], + "score": 1.0, + "content": "Eder Santana and George Hotz. Learning a driving simulator. CoRR, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 109, + 131, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 107, + 131, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 107, + 131, + 505, + 144 + ], + "score": 1.0, + "content": "Andrew Saxe, Maneesh Bhand, Zhenghao Chen, Pang Wei Koh, Bipin Suresh, and Andrew Y.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 116, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "Ng. On random weights and unsupervised feature learning. In Workshop: Deep Learning and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 152, + 303, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 303, + 166 + ], + "score": 1.0, + "content": "Unsupervised Feature Learning (NIPS). 2010.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 182, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 505, + 197 + ], + "score": 1.0, + "content": "Convolutional LSTM network: A machine learning approach for precipitation nowcasting. CoRR,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 142, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 142, + 205 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 361, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 211, + 362, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 362, + 227 + ], + "score": 1.0, + "content": "Singular Inversions, Inc. FaceGen. http://facegen.com.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 231, + 361, + 244 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 362, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 362, + 244 + ], + "score": 1.0, + "content": "William R. Softky. Unsupervised pixel-prediction. NIPS, 1996.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 251, + 504, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 504, + 263 + ], + "score": 1.0, + "content": "M. W. Spratling. Unsupervised learning of generative and discriminative weights encoding elemen-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 115, + 261, + 506, + 275 + ], + "score": 1.0, + "content": "tary image components in a predictive coding model of cortical function. Neural Computation,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 271, + 143, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 271, + 143, + 286 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 105, + 291, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "score": 1.0, + "content": "Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised learning of video", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 303, + 285, + 315 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 285, + 315 + ], + "score": 1.0, + "content": "representations using lstms. CoRR, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 321, + 504, + 344 + ], + "lines": [ + { + "bbox": [ + 106, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "Christopher Summerfield, Tobias Egner, Matthew Greene, Etienne Koechlin, Jennifer Mangels, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "Joy Hirsch. Predictive codes for forthcoming perception in the frontal cortex. Science, 314, 2006.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 105, + 351, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 504, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 504, + 365 + ], + "score": 1.0, + "content": "Lin Sun, Kui Jia, Tsung-Han Chan, Yuqiang Fang, Gang Wang, and Shuicheng Yan. Dl-sfa: Deeply-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 362, + 381, + 375 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 381, + 375 + ], + "score": 1.0, + "content": "learned slow feature analysis for action recognition. CVPR, 2014.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 504, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "score": 1.0, + "content": "Lucas Theis, Aaron van den Oord, and Matthias Bethge. A note on the evaluation of generative", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 392, + 203, + 403 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 203, + 403 + ], + "score": 1.0, + "content": "models. ICLR, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 412, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 414, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 414, + 425 + ], + "score": 1.0, + "content": "Harri Valpola. From neural pca to deep unsupervised learning. CoRR, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 430, + 503, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 443 + ], + "score": 1.0, + "content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 117, + 441, + 170, + 452 + ], + "spans": [ + { + "bbox": [ + 117, + 441, + 170, + 452 + ], + "score": 1.0, + "content": "CoRR, 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 460, + 503, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 471, + 171, + 483 + ], + "spans": [ + { + "bbox": [ + 116, + 471, + 171, + 483 + ], + "score": 1.0, + "content": "CoRR, 2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 105, + 489, + 504, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "Zhou Wang, Alan Bovik, Hamid Sheikh, and Eero Simoncelli. Image quality assessment: From", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 501, + 461, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 461, + 514 + ], + "score": 1.0, + "content": "error visibility to structural similarity. IEEE Transactions on Image Processing, 2004.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 105, + 519, + 504, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "score": 1.0, + "content": "William F. Whitney, Michael Chang, Tejas D. Kulkarni, and Joshua B. Tenenbaum. Understanding", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 530, + 345, + 543 + ], + "spans": [ + { + "bbox": [ + 116, + 530, + 345, + 543 + ], + "score": 1.0, + "content": "visual concepts with continuation learning. CoRR, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 105, + 549, + 503, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "Laurenz Wiskott and Terrence J. Sejnowski. Learning invariance from transformation sequences.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 560, + 228, + 573 + ], + "spans": [ + { + "bbox": [ + 115, + 560, + 228, + 573 + ], + "score": 1.0, + "content": "Neural Computation, 2002.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 579, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 504, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 504, + 592 + ], + "score": 1.0, + "content": "Tianfan Xue, Jiajun Wu, Katherine L. Bouman, and William T. Freeman. Visual dynamics: Proba-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 590, + 428, + 603 + ], + "spans": [ + { + "bbox": [ + 116, + 590, + 428, + 603 + ], + "score": 1.0, + "content": "bilistic future frame synthesis via cross convolutional networks. CoRR, 2016.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Antti Rasmus, Harri Valpola, Mikko Honkala, Mathias Berglund, and Tapani Raiko. Semi-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 337, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 337, + 106 + ], + "score": 1.0, + "content": "supervised learning with ladder network. CoRR, 2015.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 409, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 409, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 409, + 126 + ], + "score": 1.0, + "content": "Eder Santana and George Hotz. Learning a driving simulator. CoRR, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 111, + 409, + 126 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 131, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 107, + 131, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 107, + 131, + 505, + 144 + ], + "score": 1.0, + "content": "Andrew Saxe, Maneesh Bhand, Zhenghao Chen, Pang Wei Koh, Bipin Suresh, and Andrew Y.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 116, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "Ng. On random weights and unsupervised feature learning. In Workshop: Deep Learning and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 152, + 303, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 303, + 166 + ], + "score": 1.0, + "content": "Unsupervised Feature Learning (NIPS). 2010.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 107, + 131, + 506, + 166 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 182, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 505, + 197 + ], + "score": 1.0, + "content": "Convolutional LSTM network: A machine learning approach for precipitation nowcasting. CoRR,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 142, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 142, + 205 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 171, + 505, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 361, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 211, + 362, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 362, + 227 + ], + "score": 1.0, + "content": "Singular Inversions, Inc. FaceGen. http://facegen.com.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 211, + 362, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 231, + 361, + 244 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 362, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 362, + 244 + ], + "score": 1.0, + "content": "William R. Softky. Unsupervised pixel-prediction. NIPS, 1996.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 231, + 362, + 244 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 251, + 504, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 504, + 263 + ], + "score": 1.0, + "content": "M. W. Spratling. Unsupervised learning of generative and discriminative weights encoding elemen-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 115, + 261, + 506, + 275 + ], + "score": 1.0, + "content": "tary image components in a predictive coding model of cortical function. Neural Computation,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 271, + 143, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 271, + 143, + 286 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 251, + 506, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 291, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "score": 1.0, + "content": "Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised learning of video", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 303, + 285, + 315 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 285, + 315 + ], + "score": 1.0, + "content": "representations using lstms. CoRR, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 290, + 505, + 315 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 321, + 504, + 344 + ], + "lines": [ + { + "bbox": [ + 106, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "Christopher Summerfield, Tobias Egner, Matthew Greene, Etienne Koechlin, Jennifer Mangels, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "Joy Hirsch. Predictive codes for forthcoming perception in the frontal cortex. Science, 314, 2006.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 321, + 505, + 345 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 351, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 504, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 504, + 365 + ], + "score": 1.0, + "content": "Lin Sun, Kui Jia, Tsung-Han Chan, Yuqiang Fang, Gang Wang, and Shuicheng Yan. Dl-sfa: Deeply-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 362, + 381, + 375 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 381, + 375 + ], + "score": 1.0, + "content": "learned slow feature analysis for action recognition. CVPR, 2014.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 349, + 504, + 375 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 504, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "score": 1.0, + "content": "Lucas Theis, Aaron van den Oord, and Matthias Bethge. A note on the evaluation of generative", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 392, + 203, + 403 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 203, + 403 + ], + "score": 1.0, + "content": "models. ICLR, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 379, + 505, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 412, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 414, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 414, + 425 + ], + "score": 1.0, + "content": "Harri Valpola. From neural pca to deep unsupervised learning. CoRR, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 410, + 414, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 430, + 503, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 443 + ], + "score": 1.0, + "content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 117, + 441, + 170, + 452 + ], + "spans": [ + { + "bbox": [ + 117, + 441, + 170, + 452 + ], + "score": 1.0, + "content": "CoRR, 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 428, + 505, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 460, + 503, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 471, + 171, + 483 + ], + "spans": [ + { + "bbox": [ + 116, + 471, + 171, + 483 + ], + "score": 1.0, + "content": "CoRR, 2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 460, + 505, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 489, + 504, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "Zhou Wang, Alan Bovik, Hamid Sheikh, and Eero Simoncelli. Image quality assessment: From", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 501, + 461, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 461, + 514 + ], + "score": 1.0, + "content": "error visibility to structural similarity. IEEE Transactions on Image Processing, 2004.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 489, + 505, + 514 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 519, + 504, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "score": 1.0, + "content": "William F. Whitney, Michael Chang, Tejas D. Kulkarni, and Joshua B. Tenenbaum. Understanding", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 530, + 345, + 543 + ], + "spans": [ + { + "bbox": [ + 116, + 530, + 345, + 543 + ], + "score": 1.0, + "content": "visual concepts with continuation learning. CoRR, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 518, + 506, + 543 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 549, + 503, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "Laurenz Wiskott and Terrence J. Sejnowski. Learning invariance from transformation sequences.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 560, + 228, + 573 + ], + "spans": [ + { + "bbox": [ + 115, + 560, + 228, + 573 + ], + "score": 1.0, + "content": "Neural Computation, 2002.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 549, + 505, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 579, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 504, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 504, + 592 + ], + "score": 1.0, + "content": "Tianfan Xue, Jiajun Wu, Katherine L. Bouman, and William T. Freeman. Visual dynamics: Proba-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 590, + 428, + 603 + ], + "spans": [ + { + "bbox": [ + 116, + 590, + 428, + 603 + ], + "score": 1.0, + "content": "bilistic future frame synthesis via cross convolutional networks. CoRR, 2016.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 579, + 504, + 603 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 180, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 182, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 182, + 96 + ], + "score": 1.0, + "content": "5 APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 106, + 271, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 273, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 273, + 119 + ], + "score": 1.0, + "content": "5.1 ADDITIONAL CONTROL MODELS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "Table 3 contains results for additional variations of the PredNet and CNN-LSTM Encoder-Decoder", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "evaluated on the CalTech Pedestrian Dataset after being trained on KITTI. We evaluate the models", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "in terms of pixel prediction, thus using the PredNet model trained with loss only on the lowest layer", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 145, + 172 + ], + "score": 1.0, + "content": "(PredNet", + "type": "text" + }, + { + "bbox": [ + 145, + 160, + 158, + 171 + ], + "score": 0.84, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 159, + 505, + 172 + ], + "score": 1.0, + "content": ") as the base model. In addition to mean-squared error (MSE) and the Structural Sim-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "ilarity Index Measure (SSIM), we include calculations of the Peak Signal-To-Noise Ratio (PSNR).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 195 + ], + "score": 1.0, + "content": "For each model, we evaluate it with the original set of hyperparameters (controlling the number of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "layers, filter sizes, and number of filters per layer), as well as with the four additional sets of hy-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "perparameters that were randomly sampled from the initial random search (see main text for more", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 215, + 372, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 372, + 227 + ], + "score": 1.0, + "content": "details). Below is an explanation of the additional control models:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 134, + 235, + 504, + 259 + ], + "lines": [ + { + "bbox": [ + 133, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 234, + 413, + 248 + ], + "score": 1.0, + "content": "• PredNet (no E split): PredNet model except the error responses", + "type": "text" + }, + { + "bbox": [ + 414, + 236, + 431, + 247 + ], + "score": 0.76, + "content": "( E _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "are simply linear", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 246, + 431, + 262 + ], + "spans": [ + { + "bbox": [ + 142, + 246, + 183, + 260 + ], + "score": 0.9, + "content": "( \\hat { A } _ { l } - A _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 246, + 431, + 262 + ], + "score": 1.0, + "content": "instead of being split into positive and negative rectifications.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 134, + 263, + 503, + 309 + ], + "lines": [ + { + "bbox": [ + 133, + 262, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 133, + 262, + 251, + 276 + ], + "score": 1.0, + "content": "• CNN-LSTM Enc.-Dec.", + "type": "text" + }, + { + "bbox": [ + 252, + 263, + 277, + 274 + ], + "score": 0.45, + "content": "2 \\mathbf { x } \\ A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 262, + 469, + 276 + ], + "score": 1.0, + "content": "filts): CNN-LSTM Encoder-Decoder model (", + "type": "text" + }, + { + "bbox": [ + 469, + 264, + 481, + 274 + ], + "score": 0.83, + "content": "\\mathbf { \\delta } _ { \\cdot } A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 262, + 506, + 276 + ], + "score": 1.0, + "content": "’s are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 141, + 274, + 216, + 286 + ], + "score": 1.0, + "content": "passed instead of", + "type": "text" + }, + { + "bbox": [ + 216, + 274, + 227, + 285 + ], + "score": 0.86, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 274, + 369, + 286 + ], + "score": 1.0, + "content": "’s) except the number of filters in", + "type": "text" + }, + { + "bbox": [ + 369, + 275, + 380, + 285 + ], + "score": 0.87, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "is doubled. This controls for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 142, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "the total number of filters in the model compared to the PredNet, since the PredNet has fil-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 296, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 203, + 311 + ], + "score": 1.0, + "content": "ters to produce", + "type": "text" + }, + { + "bbox": [ + 204, + 296, + 215, + 309 + ], + "score": 0.9, + "content": "\\hat { A } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "at each layer, which is integrated into the model’s feedforward response.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 136, + 312, + 503, + 368 + ], + "lines": [ + { + "bbox": [ + 133, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 312, + 299, + 325 + ], + "score": 1.0, + "content": "• CNN-LSTM Enc.-Dec. (except pass", + "type": "text" + }, + { + "bbox": [ + 299, + 313, + 312, + 324 + ], + "score": 0.86, + "content": "E _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "): CNN-LSTM Encoder-Decoder model except", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 141, + 324, + 465, + 336 + ], + "score": 1.0, + "content": "the error is passed at the lowest layer. All remaining layers pass the activations", + "type": "text" + }, + { + "bbox": [ + 466, + 324, + 477, + 335 + ], + "score": 0.87, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 324, + 505, + 336 + ], + "score": 1.0, + "content": ". With", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 335, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 505, + 346 + ], + "score": 1.0, + "content": "training loss taken at only the lowest layer, this variation allows us to determine if the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 142, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "“prediction” subtraction operation in upper layers, which is essentially unconstrained and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 356, + 375, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 206, + 369 + ], + "score": 1.0, + "content": "learnable in the", + "type": "text" + }, + { + "bbox": [ + 206, + 357, + 219, + 368 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 356, + 375, + 369 + ], + "score": 1.0, + "content": "case, aids in the model’s performance.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 134, + 371, + 504, + 415 + ], + "lines": [ + { + "bbox": [ + 133, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 370, + 247, + 384 + ], + "score": 1.0, + "content": "• CNN-LSTM Enc.-Dec.", + "type": "text" + }, + { + "bbox": [ + 248, + 372, + 261, + 382 + ], + "score": 0.73, + "content": "+ / -", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "split): CNN-LSTM Encoder-Decoder model except the ac-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 381, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 179, + 395 + ], + "score": 1.0, + "content": "tivations", + "type": "text" + }, + { + "bbox": [ + 180, + 383, + 191, + 393 + ], + "score": 0.87, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 381, + 505, + 395 + ], + "score": 1.0, + "content": "are split into positive and negative populations before being passed to other", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 393, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 393, + 505, + 407 + ], + "score": 1.0, + "content": "layers in the network. This isolates the effect of the additional nonlinearity introduced by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 404, + 204, + 417 + ], + "spans": [ + { + "bbox": [ + 142, + 404, + 204, + 417 + ], + "score": 1.0, + "content": "this procedure.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "table", + "bbox": [ + 118, + 457, + 493, + 548 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 420, + 504, + 454 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 419, + 504, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 504, + 433 + ], + "score": 1.0, + "content": "Table 3: Quantitative evaluation of additional controls for next-frame prediction in CalTech Pedes-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 431, + 504, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 504, + 444 + ], + "score": 1.0, + "content": "trian Dataset after training on KITTI. First number indicates score with original hyperparameters.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 474, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 474, + 455 + ], + "score": 1.0, + "content": "Number in parenthesis indicates score averaged over total of five different hyperparameters.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "table_body", + "bbox": [ + 118, + 457, + 493, + 548 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 457, + 493, + 548 + ], + "spans": [ + { + "bbox": [ + 118, + 457, + 493, + 548 + ], + "score": 0.981, + "html": "
MSE (x 10-3)PSNRSSIM
PredNet3.13 (3.33)25.8 (25.5)0.884 (0.878)
PredNet (no Et split)3.20 (3.37)25.6 (25.4)0.883 (0.878)
CNN-LSTMEnc.-Dec.3.67 (3.91)25.0 (24.6)0.865 (0.856)
CNN-LSTM Enc.-Dec. (2x At filts)3.82 (3.97)24.8 (24.6)0.857 (0.853)
CNN-LSTM Enc.-Dec. (except pass Eo)3.41 (3.61)25.4 (25.1)0.873 (0.866)
CNN-LSTMEnc.-Dec. (+/- split)3.71 (3.84)24.9 (24.7)0.861 (0.857)
Copy Last Frame7.9520.00.762
", + "type": "table", + "image_path": "107c35df6bbb37645497e7b53542e2f2d0502b22d0f65ffd8195ba3e604f3321.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 118, + 457, + 493, + 487.3333333333333 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 118, + 487.3333333333333, + 493, + 517.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 118, + 517.6666666666666, + 493, + 548.0 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 407, + 577 + ], + "score": 1.0, + "content": "Equalizing the number of filters in the CNN-LSTM Encoder-Decoder (2x", + "type": "text" + }, + { + "bbox": [ + 407, + 566, + 419, + 576 + ], + "score": 0.75, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "filts) cannot account", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 577, + 504, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 504, + 587 + ], + "score": 1.0, + "content": "for its performance difference with the PredNet, and actually leads to overfitting and a decrease in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 317, + 600 + ], + "score": 1.0, + "content": "performance. Passing the error at the lowest layer", + "type": "text" + }, + { + "bbox": [ + 318, + 587, + 336, + 599 + ], + "score": 0.81, + "content": "( E _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "in the CNN-LSTM Enc.-Dec. improves", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "performance, but still does not match the PredNet, where errors are passed at all layers. Finally,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 200, + 622 + ], + "score": 1.0, + "content": "splitting the activations", + "type": "text" + }, + { + "bbox": [ + 201, + 609, + 212, + 620 + ], + "score": 0.88, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "into positive and negative populations in the CNN-LSTM Enc.-Dec. does", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 335, + 632 + ], + "score": 1.0, + "content": "not help, but the PredNet with linear error activation (“no", + "type": "text" + }, + { + "bbox": [ + 335, + 621, + 346, + 631 + ], + "score": 0.87, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "split”) performs slightly worse than the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "score": 1.0, + "content": "original split version. Together, these results suggest that the PredNet’s error passing operation can", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 642, + 350, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 350, + 655 + ], + "score": 1.0, + "content": "lead to improvements in next-frame prediction performance.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 108, + 667, + 299, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 300, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 300, + 680 + ], + "score": 1.0, + "content": "5.2 COMPARING AGAINST OTHER MODELS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "While our main comparison in the text was a control model that isolates the effects of the more", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "unique components in the PredNet, here we directly compare against other published models. We", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 185, + 721 + ], + "score": 1.0, + "content": "report results on a", + "type": "text" + }, + { + "bbox": [ + 185, + 710, + 212, + 721 + ], + "score": 0.25, + "content": "6 4 \\mathrm { x 6 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 711, + 506, + 721 + ], + "score": 1.0, + "content": "pixel, grayscale car-cam dataset and the Human3.6M dataset (Ionescu", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "et al., 2014) to compare against the two concurrently developed models by Brabandere et al. (2016)", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 180, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 182, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 182, + 96 + ], + "score": 1.0, + "content": "5 APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 106, + 271, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 273, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 273, + 119 + ], + "score": 1.0, + "content": "5.1 ADDITIONAL CONTROL MODELS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "Table 3 contains results for additional variations of the PredNet and CNN-LSTM Encoder-Decoder", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "evaluated on the CalTech Pedestrian Dataset after being trained on KITTI. We evaluate the models", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "in terms of pixel prediction, thus using the PredNet model trained with loss only on the lowest layer", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 145, + 172 + ], + "score": 1.0, + "content": "(PredNet", + "type": "text" + }, + { + "bbox": [ + 145, + 160, + 158, + 171 + ], + "score": 0.84, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 159, + 505, + 172 + ], + "score": 1.0, + "content": ") as the base model. In addition to mean-squared error (MSE) and the Structural Sim-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "ilarity Index Measure (SSIM), we include calculations of the Peak Signal-To-Noise Ratio (PSNR).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 195 + ], + "score": 1.0, + "content": "For each model, we evaluate it with the original set of hyperparameters (controlling the number of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "layers, filter sizes, and number of filters per layer), as well as with the four additional sets of hy-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "perparameters that were randomly sampled from the initial random search (see main text for more", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 215, + 372, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 372, + 227 + ], + "score": 1.0, + "content": "details). Below is an explanation of the additional control models:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 127, + 506, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 134, + 235, + 504, + 259 + ], + "lines": [ + { + "bbox": [ + 133, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 234, + 413, + 248 + ], + "score": 1.0, + "content": "• PredNet (no E split): PredNet model except the error responses", + "type": "text" + }, + { + "bbox": [ + 414, + 236, + 431, + 247 + ], + "score": 0.76, + "content": "( E _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "are simply linear", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 246, + 431, + 262 + ], + "spans": [ + { + "bbox": [ + 142, + 246, + 183, + 260 + ], + "score": 0.9, + "content": "( \\hat { A } _ { l } - A _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 246, + 431, + 262 + ], + "score": 1.0, + "content": "instead of being split into positive and negative rectifications.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 133, + 234, + 505, + 262 + ] + }, + { + "type": "text", + "bbox": [ + 134, + 263, + 503, + 309 + ], + "lines": [ + { + "bbox": [ + 133, + 262, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 133, + 262, + 251, + 276 + ], + "score": 1.0, + "content": "• CNN-LSTM Enc.-Dec.", + "type": "text" + }, + { + "bbox": [ + 252, + 263, + 277, + 274 + ], + "score": 0.45, + "content": "2 \\mathbf { x } \\ A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 262, + 469, + 276 + ], + "score": 1.0, + "content": "filts): CNN-LSTM Encoder-Decoder model (", + "type": "text" + }, + { + "bbox": [ + 469, + 264, + 481, + 274 + ], + "score": 0.83, + "content": "\\mathbf { \\delta } _ { \\cdot } A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 262, + 506, + 276 + ], + "score": 1.0, + "content": "’s are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 141, + 274, + 216, + 286 + ], + "score": 1.0, + "content": "passed instead of", + "type": "text" + }, + { + "bbox": [ + 216, + 274, + 227, + 285 + ], + "score": 0.86, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 274, + 369, + 286 + ], + "score": 1.0, + "content": "’s) except the number of filters in", + "type": "text" + }, + { + "bbox": [ + 369, + 275, + 380, + 285 + ], + "score": 0.87, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "is doubled. This controls for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 142, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "the total number of filters in the model compared to the PredNet, since the PredNet has fil-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 296, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 203, + 311 + ], + "score": 1.0, + "content": "ters to produce", + "type": "text" + }, + { + "bbox": [ + 204, + 296, + 215, + 309 + ], + "score": 0.9, + "content": "\\hat { A } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "at each layer, which is integrated into the model’s feedforward response.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 133, + 262, + 506, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 136, + 312, + 503, + 368 + ], + "lines": [ + { + "bbox": [ + 133, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 312, + 299, + 325 + ], + "score": 1.0, + "content": "• CNN-LSTM Enc.-Dec. (except pass", + "type": "text" + }, + { + "bbox": [ + 299, + 313, + 312, + 324 + ], + "score": 0.86, + "content": "E _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "): CNN-LSTM Encoder-Decoder model except", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 141, + 324, + 465, + 336 + ], + "score": 1.0, + "content": "the error is passed at the lowest layer. All remaining layers pass the activations", + "type": "text" + }, + { + "bbox": [ + 466, + 324, + 477, + 335 + ], + "score": 0.87, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 324, + 505, + 336 + ], + "score": 1.0, + "content": ". With", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 335, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 505, + 346 + ], + "score": 1.0, + "content": "training loss taken at only the lowest layer, this variation allows us to determine if the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 142, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "“prediction” subtraction operation in upper layers, which is essentially unconstrained and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 356, + 375, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 206, + 369 + ], + "score": 1.0, + "content": "learnable in the", + "type": "text" + }, + { + "bbox": [ + 206, + 357, + 219, + 368 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 356, + 375, + 369 + ], + "score": 1.0, + "content": "case, aids in the model’s performance.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 133, + 312, + 505, + 369 + ] + }, + { + "type": "text", + "bbox": [ + 134, + 371, + 504, + 415 + ], + "lines": [ + { + "bbox": [ + 133, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 370, + 247, + 384 + ], + "score": 1.0, + "content": "• CNN-LSTM Enc.-Dec.", + "type": "text" + }, + { + "bbox": [ + 248, + 372, + 261, + 382 + ], + "score": 0.73, + "content": "+ / -", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "split): CNN-LSTM Encoder-Decoder model except the ac-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 381, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 179, + 395 + ], + "score": 1.0, + "content": "tivations", + "type": "text" + }, + { + "bbox": [ + 180, + 383, + 191, + 393 + ], + "score": 0.87, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 381, + 505, + 395 + ], + "score": 1.0, + "content": "are split into positive and negative populations before being passed to other", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 393, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 393, + 505, + 407 + ], + "score": 1.0, + "content": "layers in the network. This isolates the effect of the additional nonlinearity introduced by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 404, + 204, + 417 + ], + "spans": [ + { + "bbox": [ + 142, + 404, + 204, + 417 + ], + "score": 1.0, + "content": "this procedure.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 133, + 370, + 505, + 417 + ] + }, + { + "type": "table", + "bbox": [ + 118, + 457, + 493, + 548 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 420, + 504, + 454 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 419, + 504, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 504, + 433 + ], + "score": 1.0, + "content": "Table 3: Quantitative evaluation of additional controls for next-frame prediction in CalTech Pedes-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 431, + 504, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 504, + 444 + ], + "score": 1.0, + "content": "trian Dataset after training on KITTI. First number indicates score with original hyperparameters.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 474, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 474, + 455 + ], + "score": 1.0, + "content": "Number in parenthesis indicates score averaged over total of five different hyperparameters.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "table_body", + "bbox": [ + 118, + 457, + 493, + 548 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 457, + 493, + 548 + ], + "spans": [ + { + "bbox": [ + 118, + 457, + 493, + 548 + ], + "score": 0.981, + "html": "
MSE (x 10-3)PSNRSSIM
PredNet3.13 (3.33)25.8 (25.5)0.884 (0.878)
PredNet (no Et split)3.20 (3.37)25.6 (25.4)0.883 (0.878)
CNN-LSTMEnc.-Dec.3.67 (3.91)25.0 (24.6)0.865 (0.856)
CNN-LSTM Enc.-Dec. (2x At filts)3.82 (3.97)24.8 (24.6)0.857 (0.853)
CNN-LSTM Enc.-Dec. (except pass Eo)3.41 (3.61)25.4 (25.1)0.873 (0.866)
CNN-LSTMEnc.-Dec. (+/- split)3.71 (3.84)24.9 (24.7)0.861 (0.857)
Copy Last Frame7.9520.00.762
", + "type": "table", + "image_path": "107c35df6bbb37645497e7b53542e2f2d0502b22d0f65ffd8195ba3e604f3321.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 118, + 457, + 493, + 487.3333333333333 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 118, + 487.3333333333333, + 493, + 517.6666666666666 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 118, + 517.6666666666666, + 493, + 548.0 + ], + "spans": [], + "index": 31 + } + ] + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 407, + 577 + ], + "score": 1.0, + "content": "Equalizing the number of filters in the CNN-LSTM Encoder-Decoder (2x", + "type": "text" + }, + { + "bbox": [ + 407, + 566, + 419, + 576 + ], + "score": 0.75, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "filts) cannot account", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 577, + 504, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 504, + 587 + ], + "score": 1.0, + "content": "for its performance difference with the PredNet, and actually leads to overfitting and a decrease in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 317, + 600 + ], + "score": 1.0, + "content": "performance. Passing the error at the lowest layer", + "type": "text" + }, + { + "bbox": [ + 318, + 587, + 336, + 599 + ], + "score": 0.81, + "content": "( E _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "in the CNN-LSTM Enc.-Dec. improves", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "performance, but still does not match the PredNet, where errors are passed at all layers. Finally,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 200, + 622 + ], + "score": 1.0, + "content": "splitting the activations", + "type": "text" + }, + { + "bbox": [ + 201, + 609, + 212, + 620 + ], + "score": 0.88, + "content": "A _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "into positive and negative populations in the CNN-LSTM Enc.-Dec. does", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 335, + 632 + ], + "score": 1.0, + "content": "not help, but the PredNet with linear error activation (“no", + "type": "text" + }, + { + "bbox": [ + 335, + 621, + 346, + 631 + ], + "score": 0.87, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "split”) performs slightly worse than the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "score": 1.0, + "content": "original split version. Together, these results suggest that the PredNet’s error passing operation can", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 642, + 350, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 350, + 655 + ], + "score": 1.0, + "content": "lead to improvements in next-frame prediction performance.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 565, + 505, + 655 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 667, + 299, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 300, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 300, + 680 + ], + "score": 1.0, + "content": "5.2 COMPARING AGAINST OTHER MODELS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "While our main comparison in the text was a control model that isolates the effects of the more", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "unique components in the PredNet, here we directly compare against other published models. We", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 185, + 721 + ], + "score": 1.0, + "content": "report results on a", + "type": "text" + }, + { + "bbox": [ + 185, + 710, + 212, + 721 + ], + "score": 0.25, + "content": "6 4 \\mathrm { x 6 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 711, + 506, + 721 + ], + "score": 1.0, + "content": "pixel, grayscale car-cam dataset and the Human3.6M dataset (Ionescu", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "et al., 2014) to compare against the two concurrently developed models by Brabandere et al. (2016)", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "and Finn et al. (2016), respectively. For both comparisons, we use a model with the same hyperpa-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 328, + 105 + ], + "score": 1.0, + "content": "rameters (# of layers, # of filters, etc.) of the PredNet", + "type": "text" + }, + { + "bbox": [ + 329, + 94, + 341, + 105 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "model trained on KITTI, but train from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 116 + ], + "score": 1.0, + "content": "scratch on the new datasets. The only modification we make is to train using an L2 loss instead", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "of the effective L1 loss, since both models train with an L2 loss and report results using L2-based", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "metrics (MSE for Brabandere et al. (2016) and PSNR for Finn et al. (2016)). That is, we keep the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "original PredNet model intact but directly optimize using MSE between actual and predicted frames.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "We measure next-frame prediction performance after inputting 3 frames and 10 frames, respectively,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "for the 64x64 car-cam and Human3.6M datasets, to be consistent with the published works. We also", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 182 + ], + "score": 1.0, + "content": "include the results using a feedforward multi-scale network, similar to the model of Mathieu et al.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 339, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 339, + 192 + ], + "score": 1.0, + "content": "(2016), on Human3.6M, as reported by Finn et al. (2016).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "table", + "bbox": [ + 106, + 223, + 300, + 275 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 203, + 299, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 203, + 299, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 299, + 215 + ], + "score": 1.0, + "content": "Table 4: Evaluation of Next-Frame Predictions", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 214, + 298, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 221, + 226 + ], + "score": 1.0, + "content": "on 64x64 Car-Cam Dataset.", + "type": "text" + }, + { + "bbox": [ + 229, + 218, + 298, + 229 + ], + "score": 1.0, + "content": "MSE (per-pixel)", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.0 + }, + { + "type": "table_body", + "bbox": [ + 106, + 223, + 300, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 300, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 300, + 275 + ], + "score": 0.953, + "html": "
MSE (per-pixel)
DFN (Brabandere et al., 2016)1.71 ×10-3
PredNet1.16 ×10-3
Copy Last Frame3.58 ×10-3
", + "type": "table", + "image_path": "0859777712d72e7eda2afebb1ef6c2bbf1d95cc70d3349b59b00613c896c12d4.jpg" + } + ] + } + ], + "index": 17.0, + "virtual_lines": [ + { + "bbox": [ + 106, + 223, + 300, + 236.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 236.0, + 300, + 249.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 106, + 249.0, + 300, + 262.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 106, + 262.0, + 300, + 275.0 + ], + "spans": [], + "index": 20 + } + ] + } + ], + "index": 14.0 + }, + { + "type": "table", + "bbox": [ + 325, + 223, + 504, + 277 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 324, + 201, + 504, + 224 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 324, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 324, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "Table 5: Evaluation of Next-Frame Predic-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 212, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 324, + 212, + 412, + 222 + ], + "score": 1.0, + "content": "tions on Human3.6M", + "type": "text" + }, + { + "bbox": [ + 474, + 214, + 504, + 227 + ], + "score": 1.0, + "content": "PSNR", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "table_body", + "bbox": [ + 325, + 223, + 504, + 277 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 325, + 223, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 223, + 504, + 277 + ], + "score": 0.938, + "html": "
DNA (Finn et al., 2016) PredNet FF multi-scale (Mathieu et al.,2016)42.1 38.9 26.7 32.0
", + "type": "table", + "image_path": "af305d25fd31840f4f7d01204db7b851530b42d92119ec96e453c9bf72ee30e6.jpg" + } + ] + } + ], + "index": 18.0, + "virtual_lines": [ + { + "bbox": [ + 325, + 223, + 504, + 236.5 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 325, + 236.5, + 504, + 250.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 325, + 250.0, + 504, + 263.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 325, + 263.5, + 504, + 277.0 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 15.25 + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "On a dataset similar to KITTI, our model outperforms the model proposed by Brabandere et al.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "(2016). On Human3.6M, our model outperforms a model similar to (Mathieu et al., 2016), but", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "score": 1.0, + "content": "underperforms Finn et al. (2016), although we note we did not perform any hyperparameter opti-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 322, + 146, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 146, + 333 + ], + "score": 1.0, + "content": "mization.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 108, + 347, + 282, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 282, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 282, + 359 + ], + "score": 1.0, + "content": "5.3 MULTIPLE TIME STEP PREDICTION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 126, + 371, + 484, + 690 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 371, + 484, + 690 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 371, + 484, + 690 + ], + "spans": [ + { + "bbox": [ + 126, + 371, + 484, + 690 + ], + "score": 0.969, + "type": "image", + "image_path": "d9c247d44b99433c9cfb95c59a9b7fce77018f62bc8529b66fb02544335c6500.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 126, + 371, + 484, + 477.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 126, + 477.3333333333333, + 484, + 583.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 126, + 583.6666666666666, + 484, + 690.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 694, + 505, + 739 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 694, + 506, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 506, + 707 + ], + "score": 1.0, + "content": "Figure 6: Extrapolation sequences generated by feeding PredNet predictions back into model. Left", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 705, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 215, + 717 + ], + "score": 1.0, + "content": "of the orange line: Normal", + "type": "text" + }, + { + "bbox": [ + 216, + 706, + 238, + 716 + ], + "score": 0.89, + "content": "t + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 705, + 505, + 717 + ], + "score": 1.0, + "content": "predictions; Right: Generated by recursively using the predictions", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 717, + 505, + 729 + ], + "spans": [ + { + "bbox": [ + 106, + 717, + 505, + 729 + ], + "score": 1.0, + "content": "as input. First row: Ground truth sequences. Second row: Generated frames of the original model,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 727, + 418, + 740 + ], + "spans": [ + { + "bbox": [ + 105, + 727, + 203, + 740 + ], + "score": 1.0, + "content": "trained to solely predict", + "type": "text" + }, + { + "bbox": [ + 203, + 728, + 225, + 738 + ], + "score": 0.88, + "content": "t + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 727, + 418, + 740 + ], + "score": 1.0, + "content": ". Third row: Model fine-tuned for extrapolation.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + } + ], + "index": 29.75 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "and Finn et al. (2016), respectively. For both comparisons, we use a model with the same hyperpa-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 328, + 105 + ], + "score": 1.0, + "content": "rameters (# of layers, # of filters, etc.) of the PredNet", + "type": "text" + }, + { + "bbox": [ + 329, + 94, + 341, + 105 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "model trained on KITTI, but train from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 116 + ], + "score": 1.0, + "content": "scratch on the new datasets. The only modification we make is to train using an L2 loss instead", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 127 + ], + "score": 1.0, + "content": "of the effective L1 loss, since both models train with an L2 loss and report results using L2-based", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "metrics (MSE for Brabandere et al. (2016) and PSNR for Finn et al. (2016)). That is, we keep the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "original PredNet model intact but directly optimize using MSE between actual and predicted frames.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "We measure next-frame prediction performance after inputting 3 frames and 10 frames, respectively,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "for the 64x64 car-cam and Human3.6M datasets, to be consistent with the published works. We also", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 182 + ], + "score": 1.0, + "content": "include the results using a feedforward multi-scale network, similar to the model of Mathieu et al.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 339, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 339, + 192 + ], + "score": 1.0, + "content": "(2016), on Human3.6M, as reported by Finn et al. (2016).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 81, + 505, + 192 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 223, + 300, + 275 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 203, + 299, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 203, + 299, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 299, + 215 + ], + "score": 1.0, + "content": "Table 4: Evaluation of Next-Frame Predictions", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 214, + 298, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 221, + 226 + ], + "score": 1.0, + "content": "on 64x64 Car-Cam Dataset.", + "type": "text" + }, + { + "bbox": [ + 229, + 218, + 298, + 229 + ], + "score": 1.0, + "content": "MSE (per-pixel)", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.0 + }, + { + "type": "table_body", + "bbox": [ + 106, + 223, + 300, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 300, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 300, + 275 + ], + "score": 0.953, + "html": "
MSE (per-pixel)
DFN (Brabandere et al., 2016)1.71 ×10-3
PredNet1.16 ×10-3
Copy Last Frame3.58 ×10-3
", + "type": "table", + "image_path": "0859777712d72e7eda2afebb1ef6c2bbf1d95cc70d3349b59b00613c896c12d4.jpg" + } + ] + } + ], + "index": 17.0, + "virtual_lines": [ + { + "bbox": [ + 106, + 223, + 300, + 236.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 236.0, + 300, + 249.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 106, + 249.0, + 300, + 262.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 106, + 262.0, + 300, + 275.0 + ], + "spans": [], + "index": 20 + } + ] + } + ], + "index": 14.0 + }, + { + "type": "table", + "bbox": [ + 325, + 223, + 504, + 277 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 324, + 201, + 504, + 224 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 324, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 324, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "Table 5: Evaluation of Next-Frame Predic-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 212, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 324, + 212, + 412, + 222 + ], + "score": 1.0, + "content": "tions on Human3.6M", + "type": "text" + }, + { + "bbox": [ + 474, + 214, + 504, + 227 + ], + "score": 1.0, + "content": "PSNR", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "table_body", + "bbox": [ + 325, + 223, + 504, + 277 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 325, + 223, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 223, + 504, + 277 + ], + "score": 0.938, + "html": "
DNA (Finn et al., 2016) PredNet FF multi-scale (Mathieu et al.,2016)42.1 38.9 26.7 32.0
", + "type": "table", + "image_path": "af305d25fd31840f4f7d01204db7b851530b42d92119ec96e453c9bf72ee30e6.jpg" + } + ] + } + ], + "index": 18.0, + "virtual_lines": [ + { + "bbox": [ + 325, + 223, + 504, + 236.5 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 325, + 236.5, + 504, + 250.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 325, + 250.0, + 504, + 263.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 325, + 263.5, + 504, + 277.0 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 15.25 + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "On a dataset similar to KITTI, our model outperforms the model proposed by Brabandere et al.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "(2016). On Human3.6M, our model outperforms a model similar to (Mathieu et al., 2016), but", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "score": 1.0, + "content": "underperforms Finn et al. (2016), although we note we did not perform any hyperparameter opti-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 322, + 146, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 146, + 333 + ], + "score": 1.0, + "content": "mization.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 287, + 506, + 333 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 347, + 282, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 282, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 282, + 359 + ], + "score": 1.0, + "content": "5.3 MULTIPLE TIME STEP PREDICTION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 126, + 371, + 484, + 690 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 371, + 484, + 690 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 371, + 484, + 690 + ], + "spans": [ + { + "bbox": [ + 126, + 371, + 484, + 690 + ], + "score": 0.969, + "type": "image", + "image_path": "d9c247d44b99433c9cfb95c59a9b7fce77018f62bc8529b66fb02544335c6500.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 126, + 371, + 484, + 477.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 126, + 477.3333333333333, + 484, + 583.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 126, + 583.6666666666666, + 484, + 690.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 694, + 505, + 739 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 694, + 506, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 506, + 707 + ], + "score": 1.0, + "content": "Figure 6: Extrapolation sequences generated by feeding PredNet predictions back into model. Left", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 705, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 215, + 717 + ], + "score": 1.0, + "content": "of the orange line: Normal", + "type": "text" + }, + { + "bbox": [ + 216, + 706, + 238, + 716 + ], + "score": 0.89, + "content": "t + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 705, + 505, + 717 + ], + "score": 1.0, + "content": "predictions; Right: Generated by recursively using the predictions", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 717, + 505, + 729 + ], + "spans": [ + { + "bbox": [ + 106, + 717, + 505, + 729 + ], + "score": 1.0, + "content": "as input. First row: Ground truth sequences. Second row: Generated frames of the original model,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 727, + 418, + 740 + ], + "spans": [ + { + "bbox": [ + 105, + 727, + 203, + 740 + ], + "score": 1.0, + "content": "trained to solely predict", + "type": "text" + }, + { + "bbox": [ + 203, + 728, + 225, + 738 + ], + "score": 0.88, + "content": "t + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 727, + 418, + 740 + ], + "score": 1.0, + "content": ". Third row: Model fine-tuned for extrapolation.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + } + ], + "index": 29.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "While the models presented here were originally trained to predict one frame ahead, they can be", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "made to predict multiple frames by treating predictions as actual input and recursively iterating.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 361, + 118 + ], + "score": 1.0, + "content": "Examples of this process are shown in Figure 6 for the PredNet", + "type": "text" + }, + { + "bbox": [ + 361, + 105, + 373, + 115 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "model. Although the next frame", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "predictions are reasonably accurate, the model naturally breaks down when extrapolating further into", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "the future. This is not surprising since the predictions will unavoidably have different statistics than", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 136, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 104, + 136, + 505, + 151 + ], + "score": 1.0, + "content": "the natural images for which the model was trained to handle (Bengio et al., 2015). If we additionally", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "train the model to process its own predictions, the model is better able to extrapolate. The third row", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 158, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 158, + 506, + 173 + ], + "score": 1.0, + "content": "for every sequence shows the output of the original PredNet fine-tuned for extrapolation. Starting", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 183 + ], + "score": 1.0, + "content": "from the trained weights, the model was trained with a loss over 15 time steps, where the actual", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 505, + 193 + ], + "score": 1.0, + "content": "frame was inputted for the first 10 and then the model’s predictions were used as input to the network", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 434, + 205 + ], + "score": 1.0, + "content": "for the last 5. For the first 10 time steps, the training loss was calculated on the", + "type": "text" + }, + { + "bbox": [ + 435, + 192, + 447, + 203 + ], + "score": 0.88, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "activations as", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "usual, and for the last 5, it was calculated directly as the mean absolute error with respect to the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "ground truth frames. Despite eventual blurriness (which might be expected to some extent due to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "uncertainty), the fine-tuned model captures some key structure in its extrapolations after the tenth", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 234, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 250 + ], + "score": 1.0, + "content": "time step. For instance, in the first sequence, the model estimates the general shape of an upcoming", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "score": 1.0, + "content": "shadow, despite minimal information in the last seen frame. In the second sequence, the model is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "able to extrapolate the motion of a car moving to the right. The reader is again encouraged to visit", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 268, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 104, + 268, + 506, + 283 + ], + "score": 1.0, + "content": "https://coxlab.github.io/prednet/ to view the predictions in video form. Quantitatively,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "the MSE of the model’s predictions stay well below the trivial solution of copying the last seen", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "score": 1.0, + "content": "frame, as illustrated in Fig 7. The MSE increases fairly linearly from time steps 2-10, even though", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 302, + 321, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 253, + 314 + ], + "score": 1.0, + "content": "the model was only trained for up to", + "type": "text" + }, + { + "bbox": [ + 253, + 302, + 276, + 312 + ], + "score": 0.87, + "content": "t + 5", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 302, + 321, + 314 + ], + "score": 1.0, + "content": "prediction.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 151, + 338, + 440, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 151, + 338, + 440, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 338, + 440, + 560 + ], + "spans": [ + { + "bbox": [ + 151, + 338, + 440, + 560 + ], + "score": 0.974, + "type": "image", + "image_path": "386c934c0c7cd4d4bb7e1f0e1e6977d14b386376dbb66698689a1d26524a0395.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 151, + 338, + 440, + 351.05882352941177 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 151, + 351.05882352941177, + 440, + 364.11764705882354 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 151, + 364.11764705882354, + 440, + 377.1764705882353 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 151, + 377.1764705882353, + 440, + 390.2352941176471 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 151, + 390.2352941176471, + 440, + 403.29411764705884 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 151, + 403.29411764705884, + 440, + 416.3529411764706 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 151, + 416.3529411764706, + 440, + 429.4117647058824 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 151, + 429.4117647058824, + 440, + 442.47058823529414 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 151, + 442.47058823529414, + 440, + 455.5294117647059 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 151, + 455.5294117647059, + 440, + 468.5882352941177 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 151, + 468.5882352941177, + 440, + 481.64705882352945 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 151, + 481.64705882352945, + 440, + 494.7058823529412 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 151, + 494.7058823529412, + 440, + 507.764705882353 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 151, + 507.764705882353, + 440, + 520.8235294117648 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 151, + 520.8235294117648, + 440, + 533.8823529411765 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 151, + 533.8823529411765, + 440, + 546.9411764705882 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 151, + 546.9411764705882, + 440, + 559.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 567, + 504, + 589 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 566, + 504, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 504, + 579 + ], + "score": 1.0, + "content": "Figure 7: MSE of PredNet predictions as a function of number of time steps ahead predicted. Model", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 578, + 272, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 204, + 590 + ], + "score": 1.0, + "content": "was fine-tuned for up to", + "type": "text" + }, + { + "bbox": [ + 204, + 578, + 226, + 588 + ], + "score": 0.88, + "content": "t + 5", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 578, + 272, + 590 + ], + "score": 1.0, + "content": "prediction.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "index": 33.75 + }, + { + "type": "title", + "bbox": [ + 107, + 601, + 313, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 314, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 314, + 614 + ], + "score": 1.0, + "content": "5.4 ADDITIONAL STEERING ANGLE ANALYSIS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "In Figure 8, we show the steering angle estimation accuracy on the Comma.ai (Biasini et al., 2016)", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 333, + 646 + ], + "score": 1.0, + "content": "dataset using the representation learned by the PredNet", + "type": "text" + }, + { + "bbox": [ + 333, + 633, + 346, + 644 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "model, as a function of the number of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "frames inputted into the model. The PredNet’s representation at all layers was concatenated (after", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "spatially pooling lower layers to a common spatial resolution) and a fully-connected readout was fit", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "using MSE. For each level of the number of training examples, we average over 10 cross-validation", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "splits. To serve as points of reference, we include results for two static models. The first model is an", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "autoencoder trained on single frame reconstruction with appropriately matching hyperparameters.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "A fully-connected layer was fit on the autoencoder’s representation to estimate the steering angle in", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "the same fashion as the PredNet. The second model is the default model in the posted Comma.ai", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "code (Biasini et al., 2016), which is a five layer CNN. This model is trained end-to-end to estimate", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 45.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "While the models presented here were originally trained to predict one frame ahead, they can be", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "made to predict multiple frames by treating predictions as actual input and recursively iterating.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 361, + 118 + ], + "score": 1.0, + "content": "Examples of this process are shown in Figure 6 for the PredNet", + "type": "text" + }, + { + "bbox": [ + 361, + 105, + 373, + 115 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "model. Although the next frame", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "predictions are reasonably accurate, the model naturally breaks down when extrapolating further into", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "the future. This is not surprising since the predictions will unavoidably have different statistics than", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 136, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 104, + 136, + 505, + 151 + ], + "score": 1.0, + "content": "the natural images for which the model was trained to handle (Bengio et al., 2015). If we additionally", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "train the model to process its own predictions, the model is better able to extrapolate. The third row", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 158, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 158, + 506, + 173 + ], + "score": 1.0, + "content": "for every sequence shows the output of the original PredNet fine-tuned for extrapolation. Starting", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 183 + ], + "score": 1.0, + "content": "from the trained weights, the model was trained with a loss over 15 time steps, where the actual", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 505, + 193 + ], + "score": 1.0, + "content": "frame was inputted for the first 10 and then the model’s predictions were used as input to the network", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 434, + 205 + ], + "score": 1.0, + "content": "for the last 5. For the first 10 time steps, the training loss was calculated on the", + "type": "text" + }, + { + "bbox": [ + 435, + 192, + 447, + 203 + ], + "score": 0.88, + "content": "E _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "activations as", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "usual, and for the last 5, it was calculated directly as the mean absolute error with respect to the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "ground truth frames. Despite eventual blurriness (which might be expected to some extent due to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "uncertainty), the fine-tuned model captures some key structure in its extrapolations after the tenth", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 234, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 250 + ], + "score": 1.0, + "content": "time step. For instance, in the first sequence, the model estimates the general shape of an upcoming", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 260 + ], + "score": 1.0, + "content": "shadow, despite minimal information in the last seen frame. In the second sequence, the model is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "able to extrapolate the motion of a car moving to the right. The reader is again encouraged to visit", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 268, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 104, + 268, + 506, + 283 + ], + "score": 1.0, + "content": "https://coxlab.github.io/prednet/ to view the predictions in video form. Quantitatively,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "the MSE of the model’s predictions stay well below the trivial solution of copying the last seen", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "score": 1.0, + "content": "frame, as illustrated in Fig 7. The MSE increases fairly linearly from time steps 2-10, even though", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 302, + 321, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 253, + 314 + ], + "score": 1.0, + "content": "the model was only trained for up to", + "type": "text" + }, + { + "bbox": [ + 253, + 302, + 276, + 312 + ], + "score": 0.87, + "content": "t + 5", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 302, + 321, + 314 + ], + "score": 1.0, + "content": "prediction.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 83, + 506, + 314 + ] + }, + { + "type": "image", + "bbox": [ + 151, + 338, + 440, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 151, + 338, + 440, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 338, + 440, + 560 + ], + "spans": [ + { + "bbox": [ + 151, + 338, + 440, + 560 + ], + "score": 0.974, + "type": "image", + "image_path": "386c934c0c7cd4d4bb7e1f0e1e6977d14b386376dbb66698689a1d26524a0395.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 151, + 338, + 440, + 351.05882352941177 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 151, + 351.05882352941177, + 440, + 364.11764705882354 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 151, + 364.11764705882354, + 440, + 377.1764705882353 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 151, + 377.1764705882353, + 440, + 390.2352941176471 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 151, + 390.2352941176471, + 440, + 403.29411764705884 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 151, + 403.29411764705884, + 440, + 416.3529411764706 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 151, + 416.3529411764706, + 440, + 429.4117647058824 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 151, + 429.4117647058824, + 440, + 442.47058823529414 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 151, + 442.47058823529414, + 440, + 455.5294117647059 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 151, + 455.5294117647059, + 440, + 468.5882352941177 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 151, + 468.5882352941177, + 440, + 481.64705882352945 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 151, + 481.64705882352945, + 440, + 494.7058823529412 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 151, + 494.7058823529412, + 440, + 507.764705882353 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 151, + 507.764705882353, + 440, + 520.8235294117648 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 151, + 520.8235294117648, + 440, + 533.8823529411765 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 151, + 533.8823529411765, + 440, + 546.9411764705882 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 151, + 546.9411764705882, + 440, + 559.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 567, + 504, + 589 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 566, + 504, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 504, + 579 + ], + "score": 1.0, + "content": "Figure 7: MSE of PredNet predictions as a function of number of time steps ahead predicted. Model", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 578, + 272, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 204, + 590 + ], + "score": 1.0, + "content": "was fine-tuned for up to", + "type": "text" + }, + { + "bbox": [ + 204, + 578, + 226, + 588 + ], + "score": 0.88, + "content": "t + 5", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 578, + 272, + 590 + ], + "score": 1.0, + "content": "prediction.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "index": 33.75 + }, + { + "type": "title", + "bbox": [ + 107, + 601, + 313, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 314, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 314, + 614 + ], + "score": 1.0, + "content": "5.4 ADDITIONAL STEERING ANGLE ANALYSIS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "In Figure 8, we show the steering angle estimation accuracy on the Comma.ai (Biasini et al., 2016)", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 333, + 646 + ], + "score": 1.0, + "content": "dataset using the representation learned by the PredNet", + "type": "text" + }, + { + "bbox": [ + 333, + 633, + 346, + 644 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "model, as a function of the number of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "frames inputted into the model. The PredNet’s representation at all layers was concatenated (after", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "spatially pooling lower layers to a common spatial resolution) and a fully-connected readout was fit", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "using MSE. For each level of the number of training examples, we average over 10 cross-validation", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "splits. To serve as points of reference, we include results for two static models. The first model is an", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "autoencoder trained on single frame reconstruction with appropriately matching hyperparameters.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "A fully-connected layer was fit on the autoencoder’s representation to estimate the steering angle in", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "the same fashion as the PredNet. The second model is the default model in the posted Comma.ai", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "code (Biasini et al., 2016), which is a five layer CNN. This model is trained end-to-end to estimate", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "the steering angle given the current frame as input, with a MSE loss. In addition to 25K examples,", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 366, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 376 + ], + "score": 1.0, + "content": "we trained a version using all of the frames in the Comma dataset (~396K). For all models, the final", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 376, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 388 + ], + "score": 1.0, + "content": "weights were chosen at the minimum validation error during training. Given the relatively small", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 458, + 399 + ], + "score": 1.0, + "content": "number of videos in the dataset compared to the average duration of each video, we used", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 458, + 387, + 473, + 397 + ], + "score": 0.85, + "content": "5 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 473, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "of each", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 398, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 410 + ], + "score": 1.0, + "content": "video for validation and testing, chosen as a random continuous chunk, and discarded the 10 frames", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 408, + 342, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 342, + 421 + ], + "score": 1.0, + "content": "before and after the chosen segments from the training set.", + "type": "text", + "cross_page": true + } + ], + "index": 23 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 621, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 94, + 441, + 318 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 168, + 94, + 441, + 318 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 168, + 94, + 441, + 318 + ], + "spans": [ + { + "bbox": [ + 168, + 94, + 441, + 318 + ], + "score": 0.974, + "type": "image", + "image_path": "786ee29cf8d282ebc8e20ea4c6a1d601619f455dd6984bab67feeada0044938e.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 168, + 94, + 441, + 107.17647058823529 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 168, + 107.17647058823529, + 441, + 120.35294117647058 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 168, + 120.35294117647058, + 441, + 133.52941176470588 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 168, + 133.52941176470588, + 441, + 146.7058823529412 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 168, + 146.7058823529412, + 441, + 159.8823529411765 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 168, + 159.8823529411765, + 441, + 173.0588235294118 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 168, + 173.0588235294118, + 441, + 186.2352941176471 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 168, + 186.2352941176471, + 441, + 199.4117647058824 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 168, + 199.4117647058824, + 441, + 212.5882352941177 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 168, + 212.5882352941177, + 441, + 225.764705882353 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 168, + 225.764705882353, + 441, + 238.94117647058832 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 168, + 238.94117647058832, + 441, + 252.11764705882362 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 168, + 252.11764705882362, + 441, + 265.2941176470589 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 168, + 265.2941176470589, + 441, + 278.4705882352942 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 168, + 278.4705882352942, + 441, + 291.6470588235295 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 168, + 291.6470588235295, + 441, + 304.8235294117648 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 168, + 304.8235294117648, + 441, + 318.0000000000001 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 125, + 331, + 484, + 343 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 329, + 485, + 345 + ], + "spans": [ + { + "bbox": [ + 124, + 329, + 485, + 345 + ], + "score": 1.0, + "content": "Figure 8: Steering angle estimation accuracy as a function of the number of input frames.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "the steering angle given the current frame as input, with a MSE loss. In addition to 25K examples,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 366, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 376 + ], + "score": 1.0, + "content": "we trained a version using all of the frames in the Comma dataset (~396K). For all models, the final", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 376, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 388 + ], + "score": 1.0, + "content": "weights were chosen at the minimum validation error during training. Given the relatively small", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 458, + 399 + ], + "score": 1.0, + "content": "number of videos in the dataset compared to the average duration of each video, we used", + "type": "text" + }, + { + "bbox": [ + 458, + 387, + 473, + 397 + ], + "score": 0.85, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "of each", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 398, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 410 + ], + "score": 1.0, + "content": "video for validation and testing, chosen as a random continuous chunk, and discarded the 10 frames", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 408, + 342, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 342, + 421 + ], + "score": 1.0, + "content": "before and after the chosen segments from the training set.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "As illustrated in Figure 8, the PredNet’s performance gets better over time, as one might expect,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "as the model is able to accumulate more information. Interestingly, it performs reasonably well", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "after just one time step, in a regime that is orthogonal to the training procedure of the PredNet where", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 473 + ], + "score": 1.0, + "content": "there are no dynamics. Altogether, these results again point to the usefulness of the model in learning", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 469, + 225, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 225, + 483 + ], + "score": 1.0, + "content": "underlying latent parameters.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 107, + 494, + 315, + 506 + ], + "lines": [ + { + "bbox": [ + 106, + 494, + 316, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 174, + 507 + ], + "score": 1.0, + "content": "5.5 PREDNET", + "type": "text" + }, + { + "bbox": [ + 175, + 495, + 192, + 506 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 494, + 316, + 507 + ], + "score": 1.0, + "content": "NEXT-FRAME PREDICTIONS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 367, + 528 + ], + "score": 1.0, + "content": "Figures 9 and 10 compare next-frame predictions by the PredNet", + "type": "text" + }, + { + "bbox": [ + 368, + 516, + 386, + 526 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "model, trained with a predic-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 200, + 538 + ], + "score": 1.0, + "content": "tion loss on all layers (", + "type": "text" + }, + { + "bbox": [ + 200, + 527, + 232, + 537 + ], + "score": 0.81, + "content": "\\lambda _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 526, + 237, + 538 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 237, + 527, + 286, + 537 + ], + "score": 0.62, + "content": "\\lambda _ { l > 0 } = 0 . 1 ", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 526, + 362, + 538 + ], + "score": 1.0, + "content": "), and the PredNet", + "type": "text" + }, + { + "bbox": [ + 362, + 527, + 374, + 537 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "model, trained with a loss only", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 538, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 549 + ], + "score": 1.0, + "content": "on the lowest layer. At first glance, the difference in predictions seem fairly minor, and indeed, in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 182, + 560 + ], + "score": 1.0, + "content": "terms of MSE, the", + "type": "text" + }, + { + "bbox": [ + 182, + 548, + 200, + 559 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 548, + 330, + 560 + ], + "score": 1.0, + "content": "model only underperformed the", + "type": "text" + }, + { + "bbox": [ + 330, + 549, + 343, + 559 + ], + "score": 0.89, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 548, + 388, + 560 + ], + "score": 1.0, + "content": "version by", + "type": "text" + }, + { + "bbox": [ + 388, + 548, + 402, + 559 + ], + "score": 0.85, + "content": "3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 548, + 420, + 560 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 420, + 548, + 435, + 559 + ], + "score": 0.86, + "content": "6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 548, + 505, + 560 + ], + "score": 1.0, + "content": ", respectively, for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "the rotating faces and CalTech Pedestrian datasets. Upon careful inspection, however, it is apparent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 139, + 582 + ], + "score": 1.0, + "content": "that the", + "type": "text" + }, + { + "bbox": [ + 139, + 570, + 157, + 581 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 570, + 351, + 582 + ], + "score": 1.0, + "content": "predictions lack some of the finer details of the", + "type": "text" + }, + { + "bbox": [ + 351, + 570, + 363, + 581 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "predictions and are more blurry in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "regions of high variance. For instance, with the rotating faces, the facial features are less defined", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 592, + 410, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 410, + 604 + ], + "score": 1.0, + "content": "and with CalTech, details of approaching shadows and cars are less precise.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 94, + 441, + 318 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 168, + 94, + 441, + 318 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 168, + 94, + 441, + 318 + ], + "spans": [ + { + "bbox": [ + 168, + 94, + 441, + 318 + ], + "score": 0.974, + "type": "image", + "image_path": "786ee29cf8d282ebc8e20ea4c6a1d601619f455dd6984bab67feeada0044938e.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 168, + 94, + 441, + 107.17647058823529 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 168, + 107.17647058823529, + 441, + 120.35294117647058 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 168, + 120.35294117647058, + 441, + 133.52941176470588 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 168, + 133.52941176470588, + 441, + 146.7058823529412 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 168, + 146.7058823529412, + 441, + 159.8823529411765 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 168, + 159.8823529411765, + 441, + 173.0588235294118 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 168, + 173.0588235294118, + 441, + 186.2352941176471 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 168, + 186.2352941176471, + 441, + 199.4117647058824 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 168, + 199.4117647058824, + 441, + 212.5882352941177 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 168, + 212.5882352941177, + 441, + 225.764705882353 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 168, + 225.764705882353, + 441, + 238.94117647058832 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 168, + 238.94117647058832, + 441, + 252.11764705882362 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 168, + 252.11764705882362, + 441, + 265.2941176470589 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 168, + 265.2941176470589, + 441, + 278.4705882352942 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 168, + 278.4705882352942, + 441, + 291.6470588235295 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 168, + 291.6470588235295, + 441, + 304.8235294117648 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 168, + 304.8235294117648, + 441, + 318.0000000000001 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 125, + 331, + 484, + 343 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 329, + 485, + 345 + ], + "spans": [ + { + "bbox": [ + 124, + 329, + 485, + 345 + ], + "score": 1.0, + "content": "Figure 8: Steering angle estimation accuracy as a function of the number of input frames.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 420 + ], + "lines": [], + "index": 20.5, + "bbox_fs": [ + 105, + 354, + 506, + 421 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "As illustrated in Figure 8, the PredNet’s performance gets better over time, as one might expect,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "as the model is able to accumulate more information. Interestingly, it performs reasonably well", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "after just one time step, in a regime that is orthogonal to the training procedure of the PredNet where", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 473 + ], + "score": 1.0, + "content": "there are no dynamics. Altogether, these results again point to the usefulness of the model in learning", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 469, + 225, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 225, + 483 + ], + "score": 1.0, + "content": "underlying latent parameters.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 425, + 506, + 483 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 494, + 315, + 506 + ], + "lines": [ + { + "bbox": [ + 106, + 494, + 316, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 174, + 507 + ], + "score": 1.0, + "content": "5.5 PREDNET", + "type": "text" + }, + { + "bbox": [ + 175, + 495, + 192, + 506 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 494, + 316, + 507 + ], + "score": 1.0, + "content": "NEXT-FRAME PREDICTIONS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 367, + 528 + ], + "score": 1.0, + "content": "Figures 9 and 10 compare next-frame predictions by the PredNet", + "type": "text" + }, + { + "bbox": [ + 368, + 516, + 386, + 526 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "model, trained with a predic-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 200, + 538 + ], + "score": 1.0, + "content": "tion loss on all layers (", + "type": "text" + }, + { + "bbox": [ + 200, + 527, + 232, + 537 + ], + "score": 0.81, + "content": "\\lambda _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 526, + 237, + 538 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 237, + 527, + 286, + 537 + ], + "score": 0.62, + "content": "\\lambda _ { l > 0 } = 0 . 1 ", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 526, + 362, + 538 + ], + "score": 1.0, + "content": "), and the PredNet", + "type": "text" + }, + { + "bbox": [ + 362, + 527, + 374, + 537 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "model, trained with a loss only", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 538, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 549 + ], + "score": 1.0, + "content": "on the lowest layer. At first glance, the difference in predictions seem fairly minor, and indeed, in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 182, + 560 + ], + "score": 1.0, + "content": "terms of MSE, the", + "type": "text" + }, + { + "bbox": [ + 182, + 548, + 200, + 559 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 548, + 330, + 560 + ], + "score": 1.0, + "content": "model only underperformed the", + "type": "text" + }, + { + "bbox": [ + 330, + 549, + 343, + 559 + ], + "score": 0.89, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 548, + 388, + 560 + ], + "score": 1.0, + "content": "version by", + "type": "text" + }, + { + "bbox": [ + 388, + 548, + 402, + 559 + ], + "score": 0.85, + "content": "3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 548, + 420, + 560 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 420, + 548, + 435, + 559 + ], + "score": 0.86, + "content": "6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 548, + 505, + 560 + ], + "score": 1.0, + "content": ", respectively, for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "the rotating faces and CalTech Pedestrian datasets. Upon careful inspection, however, it is apparent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 139, + 582 + ], + "score": 1.0, + "content": "that the", + "type": "text" + }, + { + "bbox": [ + 139, + 570, + 157, + 581 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 570, + 351, + 582 + ], + "score": 1.0, + "content": "predictions lack some of the finer details of the", + "type": "text" + }, + { + "bbox": [ + 351, + 570, + 363, + 581 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "predictions and are more blurry in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 581, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 506, + 594 + ], + "score": 1.0, + "content": "regions of high variance. For instance, with the rotating faces, the facial features are less defined", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 592, + 410, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 410, + 604 + ], + "score": 1.0, + "content": "and with CalTech, details of approaching shadows and cars are less precise.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 515, + 506, + 604 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 213, + 505, + 554 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 213, + 505, + 554 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 213, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 107, + 213, + 505, + 554 + ], + "score": 0.973, + "type": "image", + "image_path": "235db1e2a31428ae54787e2a5c30dc4c978774edcf9dd06ef6ac6cecc851ece6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 213, + 505, + 326.6666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 326.6666666666667, + 505, + 440.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 440.33333333333337, + 505, + 554.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 319, + 550, + 355, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 317, + 549, + 356, + 561 + ], + "spans": [ + { + "bbox": [ + 317, + 549, + 343, + 561 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 343, + 551, + 356, + 559 + ], + "score": 0.5, + "content": "", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 562, + 505, + 606 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 283, + 576 + ], + "score": 1.0, + "content": "Figure 9: Next-frame predictions of PredNet", + "type": "text" + }, + { + "bbox": [ + 284, + 563, + 302, + 573 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "model on the rotating faces dataset and comparison", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 572, + 504, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 116, + 586 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 116, + 573, + 129, + 584 + ], + "score": 0.89, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 572, + 209, + 586 + ], + "score": 1.0, + "content": "version. The ”Error", + "type": "text" + }, + { + "bbox": [ + 209, + 573, + 250, + 584 + ], + "score": 0.89, + "content": "{ \\cal L } _ { a l l } { - \\cal L } _ { 0 } { } ^ { \\cdots }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 572, + 491, + 586 + ], + "score": 1.0, + "content": "visualization shows where the pixel error was smaller for the", + "type": "text" + }, + { + "bbox": [ + 491, + 573, + 504, + 584 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 169, + 597 + ], + "score": 1.0, + "content": "model than the", + "type": "text" + }, + { + "bbox": [ + 169, + 585, + 187, + 595 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 584, + 363, + 597 + ], + "score": 1.0, + "content": "model. Green regions correspond to where", + "type": "text" + }, + { + "bbox": [ + 364, + 585, + 376, + 595 + ], + "score": 0.89, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "was better and red corresponds", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 595, + 208, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 143, + 608 + ], + "score": 1.0, + "content": "to where", + "type": "text" + }, + { + "bbox": [ + 144, + 595, + 162, + 606 + ], + "score": 0.92, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 595, + 208, + 608 + ], + "score": 1.0, + "content": "was better.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + } + ], + "index": 3 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 213, + 505, + 554 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 213, + 505, + 554 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 213, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 107, + 213, + 505, + 554 + ], + "score": 0.973, + "type": "image", + "image_path": "235db1e2a31428ae54787e2a5c30dc4c978774edcf9dd06ef6ac6cecc851ece6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 213, + 505, + 326.6666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 326.6666666666667, + 505, + 440.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 440.33333333333337, + 505, + 554.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 319, + 550, + 355, + 560 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 317, + 549, + 356, + 561 + ], + "spans": [ + { + "bbox": [ + 317, + 549, + 343, + 561 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 343, + 551, + 356, + 559 + ], + "score": 0.5, + "content": "", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 562, + 505, + 606 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 283, + 576 + ], + "score": 1.0, + "content": "Figure 9: Next-frame predictions of PredNet", + "type": "text" + }, + { + "bbox": [ + 284, + 563, + 302, + 573 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "model on the rotating faces dataset and comparison", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 572, + 504, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 116, + 586 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 116, + 573, + 129, + 584 + ], + "score": 0.89, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 572, + 209, + 586 + ], + "score": 1.0, + "content": "version. The ”Error", + "type": "text" + }, + { + "bbox": [ + 209, + 573, + 250, + 584 + ], + "score": 0.89, + "content": "{ \\cal L } _ { a l l } { - \\cal L } _ { 0 } { } ^ { \\cdots }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 572, + 491, + 586 + ], + "score": 1.0, + "content": "visualization shows where the pixel error was smaller for the", + "type": "text" + }, + { + "bbox": [ + 491, + 573, + 504, + 584 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 169, + 597 + ], + "score": 1.0, + "content": "model than the", + "type": "text" + }, + { + "bbox": [ + 169, + 585, + 187, + 595 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 584, + 363, + 597 + ], + "score": 1.0, + "content": "model. Green regions correspond to where", + "type": "text" + }, + { + "bbox": [ + 364, + 585, + 376, + 595 + ], + "score": 0.89, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "was better and red corresponds", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 595, + 208, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 143, + 608 + ], + "score": 1.0, + "content": "to where", + "type": "text" + }, + { + "bbox": [ + 144, + 595, + 162, + 606 + ], + "score": 0.92, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 595, + 208, + 608 + ], + "score": 1.0, + "content": "was better.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + } + ], + "index": 3 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 162, + 504, + 600 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 162, + 504, + 600 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 162, + 504, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 504, + 600 + ], + "score": 0.973, + "type": "image", + "image_path": "5436ead45eb2f3750a49987e6825e0fe9897d10aa4f0e52240e2823c9aafffd5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 162, + 504, + 308.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 308.0, + 504, + 454.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 454.0, + 504, + 600.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 327, + 597, + 363, + 608 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 325, + 596, + 365, + 610 + ], + "spans": [ + { + "bbox": [ + 325, + 596, + 352, + 610 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 352, + 598, + 365, + 607 + ], + "score": 0.47, + "content": "", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 610, + 506, + 655 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 298, + 623 + ], + "score": 1.0, + "content": "Figure 10: Next-frame predictions of PredNet", + "type": "text" + }, + { + "bbox": [ + 298, + 611, + 316, + 622 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "model on the CalTech Pedestrian dataset and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 167, + 635 + ], + "score": 1.0, + "content": "comparison to", + "type": "text" + }, + { + "bbox": [ + 167, + 622, + 180, + 633 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 621, + 267, + 635 + ], + "score": 1.0, + "content": "version. The ”Error", + "type": "text" + }, + { + "bbox": [ + 267, + 622, + 315, + 633 + ], + "score": 0.9, + "content": "{ \\cal L } _ { a l l } - { \\cal L } _ { 0 } { } ^ { }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "visualization shows where the pixel error was", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 166, + 645 + ], + "score": 1.0, + "content": "smaller for the", + "type": "text" + }, + { + "bbox": [ + 167, + 633, + 179, + 644 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 633, + 241, + 645 + ], + "score": 1.0, + "content": "model than the", + "type": "text" + }, + { + "bbox": [ + 241, + 633, + 259, + 644 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 633, + 432, + 645 + ], + "score": 1.0, + "content": "model. Green regions correspond to where", + "type": "text" + }, + { + "bbox": [ + 432, + 633, + 444, + 644 + ], + "score": 0.89, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "was better and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 644, + 274, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 209, + 656 + ], + "score": 1.0, + "content": "red corresponds to where", + "type": "text" + }, + { + "bbox": [ + 209, + 644, + 227, + 655 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 644, + 274, + 656 + ], + "score": 1.0, + "content": "was better.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + } + ], + "index": 3 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2017", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 162, + 504, + 600 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 162, + 504, + 600 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 162, + 504, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 504, + 600 + ], + "score": 0.973, + "type": "image", + "image_path": "5436ead45eb2f3750a49987e6825e0fe9897d10aa4f0e52240e2823c9aafffd5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 162, + 504, + 308.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 308.0, + 504, + 454.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 454.0, + 504, + 600.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 327, + 597, + 363, + 608 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 325, + 596, + 365, + 610 + ], + "spans": [ + { + "bbox": [ + 325, + 596, + 352, + 610 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 352, + 598, + 365, + 607 + ], + "score": 0.47, + "content": "", + "type": "inline_equation" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image_caption", + "bbox": [ + 106, + 610, + 506, + 655 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 298, + 623 + ], + "score": 1.0, + "content": "Figure 10: Next-frame predictions of PredNet", + "type": "text" + }, + { + "bbox": [ + 298, + 611, + 316, + 622 + ], + "score": 0.89, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "model on the CalTech Pedestrian dataset and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 167, + 635 + ], + "score": 1.0, + "content": "comparison to", + "type": "text" + }, + { + "bbox": [ + 167, + 622, + 180, + 633 + ], + "score": 0.88, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 621, + 267, + 635 + ], + "score": 1.0, + "content": "version. The ”Error", + "type": "text" + }, + { + "bbox": [ + 267, + 622, + 315, + 633 + ], + "score": 0.9, + "content": "{ \\cal L } _ { a l l } - { \\cal L } _ { 0 } { } ^ { }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "visualization shows where the pixel error was", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 166, + 645 + ], + "score": 1.0, + "content": "smaller for the", + "type": "text" + }, + { + "bbox": [ + 167, + 633, + 179, + 644 + ], + "score": 0.87, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 633, + 241, + 645 + ], + "score": 1.0, + "content": "model than the", + "type": "text" + }, + { + "bbox": [ + 241, + 633, + 259, + 644 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 633, + 432, + 645 + ], + "score": 1.0, + "content": "model. Green regions correspond to where", + "type": "text" + }, + { + "bbox": [ + 432, + 633, + 444, + 644 + ], + "score": 0.89, + "content": "L _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "was better and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 644, + 274, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 209, + 656 + ], + "score": 1.0, + "content": "red corresponds to where", + "type": "text" + }, + { + "bbox": [ + 209, + 644, + 227, + 655 + ], + "score": 0.9, + "content": "L _ { a l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 644, + 274, + 656 + ], + "score": 1.0, + "content": "was better.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + } + ], + "index": 3 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/B1ewdt9xe/B1ewdt9xe_model.json b/parse/train/B1ewdt9xe/B1ewdt9xe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eec80d6051a3b7071b93be66f9eec678e97f1c10 --- /dev/null +++ b/parse/train/B1ewdt9xe/B1ewdt9xe_model.json @@ -0,0 +1,22295 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 679, + 1302, + 679, + 1302, + 1289, + 398, + 1289 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1419, + 1403, + 1419, + 1403, + 1664, + 298, + 1664 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1678, + 1404, + 1678, + 1404, + 1985, + 298, + 1985 + ], + "score": 0.978 + }, + { + "category_id": 0, + "poly": [ + 299, + 219, + 1394, + 219, + 1394, + 324, + 299, + 324 + ], + "score": 0.966 + }, + { + "category_id": 2, + "poly": [ + 333, + 2005, + 1266, + 2005, + 1266, + 2034, + 333, + 2034 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 302, + 1349, + 573, + 1349, + 573, + 1384, + 302, + 1384 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.881 + }, + { + "category_id": 0, + "poly": [ + 773, + 611, + 927, + 611, + 927, + 644, + 773, + 644 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 857, + 2089, + 857, + 2112, + 841, + 2112 + ], + "score": 0.699 + }, + { + "category_id": 1, + "poly": [ + 314, + 375, + 872, + 375, + 872, + 528, + 314, + 528 + ], + "score": 0.661 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 220.0, + 1399.0, + 220.0, + 1399.0, + 269.0, + 296.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 274.0, + 1226.0, + 274.0, + 1226.0, + 328.0, + 293.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 2001.0, + 1271.0, + 2001.0, + 1271.0, + 2038.0, + 340.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1346.0, + 579.0, + 1346.0, + 579.0, + 1393.0, + 294.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 607.0, + 934.0, + 607.0, + 934.0, + 650.0, + 768.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 861.0, + 2088.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 680.0, + 1305.0, + 680.0, + 1305.0, + 716.0, + 396.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 711.0, + 1304.0, + 711.0, + 1304.0, + 747.0, + 393.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 741.0, + 1305.0, + 741.0, + 1305.0, + 774.0, + 393.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 773.0, + 1306.0, + 773.0, + 1306.0, + 805.0, + 395.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 804.0, + 1306.0, + 804.0, + 1306.0, + 835.0, + 394.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 835.0, + 1305.0, + 835.0, + 1305.0, + 865.0, + 396.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 865.0, + 1305.0, + 865.0, + 1305.0, + 897.0, + 395.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 893.0, + 1305.0, + 893.0, + 1305.0, + 928.0, + 393.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 926.0, + 1305.0, + 926.0, + 1305.0, + 959.0, + 395.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 955.0, + 1305.0, + 955.0, + 1305.0, + 988.0, + 395.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 987.0, + 1304.0, + 987.0, + 1304.0, + 1020.0, + 395.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1017.0, + 1305.0, + 1017.0, + 1305.0, + 1050.0, + 396.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1044.0, + 1305.0, + 1044.0, + 1305.0, + 1082.0, + 392.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1076.0, + 1305.0, + 1076.0, + 1305.0, + 1111.0, + 394.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1109.0, + 1305.0, + 1109.0, + 1305.0, + 1141.0, + 394.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1138.0, + 1306.0, + 1138.0, + 1306.0, + 1172.0, + 394.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1168.0, + 1304.0, + 1168.0, + 1304.0, + 1201.0, + 395.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1198.0, + 1306.0, + 1198.0, + 1306.0, + 1234.0, + 394.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1228.0, + 1305.0, + 1228.0, + 1305.0, + 1263.0, + 393.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1265.0, + 574.0, + 1265.0, + 574.0, + 1293.0, + 395.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1417.0, + 1404.0, + 1417.0, + 1404.0, + 1455.0, + 292.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1449.0, + 1407.0, + 1449.0, + 1407.0, + 1486.0, + 293.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1482.0, + 1405.0, + 1482.0, + 1405.0, + 1515.0, + 294.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1512.0, + 1404.0, + 1512.0, + 1404.0, + 1545.0, + 294.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1540.0, + 1405.0, + 1540.0, + 1405.0, + 1574.0, + 294.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1569.0, + 1405.0, + 1569.0, + 1405.0, + 1609.0, + 292.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1600.0, + 1405.0, + 1600.0, + 1405.0, + 1636.0, + 293.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1630.0, + 1128.0, + 1630.0, + 1128.0, + 1668.0, + 292.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1404.0, + 1679.0, + 1404.0, + 1714.0, + 294.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1709.0, + 1404.0, + 1709.0, + 1404.0, + 1745.0, + 294.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1737.0, + 1408.0, + 1737.0, + 1408.0, + 1778.0, + 291.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1769.0, + 1404.0, + 1769.0, + 1404.0, + 1806.0, + 293.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1801.0, + 1405.0, + 1801.0, + 1405.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1829.0, + 1405.0, + 1829.0, + 1405.0, + 1868.0, + 293.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1861.0, + 1404.0, + 1861.0, + 1404.0, + 1896.0, + 294.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 1404.0, + 1893.0, + 1404.0, + 1928.0, + 294.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1923.0, + 1404.0, + 1923.0, + 1404.0, + 1957.0, + 293.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1952.0, + 1406.0, + 1952.0, + 1406.0, + 1989.0, + 295.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 376.0, + 873.0, + 376.0, + 873.0, + 409.0, + 313.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 405.0, + 532.0, + 405.0, + 532.0, + 441.0, + 312.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 437.0, + 648.0, + 437.0, + 648.0, + 469.0, + 310.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 468.0, + 863.0, + 468.0, + 863.0, + 503.0, + 314.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 499.0, + 835.0, + 499.0, + 835.0, + 532.0, + 312.0, + 532.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 397, + 1405, + 397, + 1405, + 979, + 297, + 979 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 994, + 1404, + 994, + 1404, + 1301, + 297, + 1301 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1402, + 229, + 1402, + 384, + 298, + 384 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 423, + 1350, + 1258, + 1350, + 1258, + 1836, + 423, + 1836 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 294, + 1869, + 1406, + 1869, + 1406, + 2029, + 294, + 2029 + ], + "score": 0.959 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.732 + }, + { + "category_id": 13, + "poly": [ + 461, + 1931, + 510, + 1931, + 510, + 1968, + 461, + 1968 + ], + "score": 0.87, + "latex": "( \\hat { A } _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1968, + 346, + 1968, + 346, + 1999, + 298, + 1999 + ], + "score": 0.84, + "latex": "( E _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 907, + 1937, + 956, + 1937, + 956, + 1968, + 907, + 1968 + ], + "score": 0.82, + "latex": "( A _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 879, + 1902, + 927, + 1902, + 927, + 1933, + 879, + 1933 + ], + "score": 0.8, + "latex": "( R _ { l } )" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1387.0, + 1232.0, + 1387.0, + 1232.0, + 1415.0, + 1160.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1471.0, + 1092.0, + 1471.0, + 1092.0, + 1503.0, + 998.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1535.0, + 924.0, + 1535.0, + 924.0, + 1559.0, + 875.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1555.0, + 929.0, + 1555.0, + 929.0, + 1584.0, + 871.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1545.0, + 1073.0, + 1545.0, + 1073.0, + 1573.0, + 1021.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1595.0, + 605.0, + 1595.0, + 605.0, + 1603.0, + 597.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1585.0, + 1241.0, + 1585.0, + 1241.0, + 1617.0, + 1149.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1611.0, + 1079.0, + 1611.0, + 1079.0, + 1645.0, + 1013.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1638.0, + 968.0, + 1638.0, + 968.0, + 1669.0, + 831.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1637.0, + 1240.0, + 1637.0, + 1240.0, + 1666.0, + 1155.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1654.0, + 1072.0, + 1654.0, + 1072.0, + 1682.0, + 1019.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1680.0, + 1226.0, + 1680.0, + 1226.0, + 1711.0, + 1173.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1707.0, + 1073.0, + 1707.0, + 1073.0, + 1734.0, + 1021.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1730.0, + 747.0, + 1730.0, + 747.0, + 1764.0, + 712.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1784.0, + 1080.0, + 1784.0, + 1080.0, + 1816.0, + 1019.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1431.0, + 525.0, + 1431.0, + 525.0, + 1477.5, + 462.0, + 1477.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1426.5, + 635.0, + 1426.5, + 635.0, + 1483.0, + 572.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.25, + 1487.5, + 754.25, + 1487.5, + 754.25, + 1530.5, + 702.25, + 1530.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1542.0, + 634.0, + 1542.0, + 634.0, + 1586.0, + 574.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1675.0, + 506.0, + 1675.0, + 506.0, + 1712.5, + 480.0, + 1712.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1675.5, + 616.0, + 1675.5, + 616.0, + 1711.5, + 591.0, + 1711.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.25, + 1783.5, + 619.25, + 1783.5, + 619.25, + 1820.5, + 587.25, + 1820.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1870.0, + 1404.0, + 1870.0, + 1404.0, + 1906.0, + 295.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1899.0, + 878.0, + 1899.0, + 878.0, + 1938.0, + 293.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1899.0, + 1406.0, + 1899.0, + 1406.0, + 1938.0, + 928.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1937.0, + 460.0, + 1937.0, + 460.0, + 1971.0, + 296.0, + 1971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1937.0, + 906.0, + 1937.0, + 906.0, + 1971.0, + 511.0, + 1971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1937.0, + 1403.0, + 1937.0, + 1403.0, + 1971.0, + 957.0, + 1971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1967.0, + 297.0, + 1967.0, + 297.0, + 2003.0, + 294.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1967.0, + 1404.0, + 1967.0, + 1404.0, + 2003.0, + 347.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1999.0, + 575.0, + 1999.0, + 575.0, + 2033.0, + 295.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 397.0, + 1403.0, + 397.0, + 1403.0, + 435.0, + 294.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 431.0, + 1402.0, + 431.0, + 1402.0, + 462.0, + 296.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 460.0, + 1403.0, + 460.0, + 1403.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 490.0, + 1405.0, + 490.0, + 1405.0, + 525.0, + 295.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 521.0, + 1405.0, + 521.0, + 1405.0, + 555.0, + 294.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 551.0, + 1403.0, + 551.0, + 1403.0, + 585.0, + 294.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 580.0, + 1405.0, + 580.0, + 1405.0, + 618.0, + 292.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 608.0, + 1402.0, + 608.0, + 1402.0, + 650.0, + 292.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 642.0, + 1403.0, + 642.0, + 1403.0, + 675.0, + 295.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 673.0, + 1402.0, + 673.0, + 1402.0, + 704.0, + 295.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 704.0, + 1403.0, + 704.0, + 1403.0, + 736.0, + 295.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 732.0, + 1402.0, + 732.0, + 1402.0, + 770.0, + 292.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 762.0, + 1403.0, + 762.0, + 1403.0, + 798.0, + 294.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 793.0, + 1405.0, + 793.0, + 1405.0, + 831.0, + 294.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 825.0, + 1405.0, + 825.0, + 1405.0, + 860.0, + 295.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 857.0, + 1403.0, + 857.0, + 1403.0, + 889.0, + 296.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 884.0, + 1406.0, + 884.0, + 1406.0, + 924.0, + 291.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 917.0, + 1405.0, + 917.0, + 1405.0, + 953.0, + 295.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 948.0, + 1239.0, + 948.0, + 1239.0, + 984.0, + 292.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 994.0, + 1405.0, + 994.0, + 1405.0, + 1028.0, + 296.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1025.0, + 1405.0, + 1025.0, + 1405.0, + 1057.0, + 295.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1055.0, + 1406.0, + 1055.0, + 1406.0, + 1090.0, + 295.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1085.0, + 1406.0, + 1085.0, + 1406.0, + 1121.0, + 295.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1116.0, + 1406.0, + 1116.0, + 1406.0, + 1150.0, + 292.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1140.0, + 1406.0, + 1140.0, + 1406.0, + 1187.0, + 292.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1179.0, + 1405.0, + 1179.0, + 1405.0, + 1211.0, + 296.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1242.0, + 294.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1234.0, + 1406.0, + 1234.0, + 1406.0, + 1276.0, + 292.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1270.0, + 830.0, + 1270.0, + 830.0, + 1304.0, + 293.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 263.0, + 1404.0, + 263.0, + 1404.0, + 296.0, + 297.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 1406.0, + 289.0, + 1406.0, + 330.0, + 293.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 1406.0, + 320.0, + 1406.0, + 357.0, + 293.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1328.0, + 352.0, + 1328.0, + 389.0, + 294.0, + 389.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 963, + 1404, + 963, + 1404, + 1437, + 298, + 1437 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1453, + 1404, + 1453, + 1404, + 1638, + 298, + 1638 + ], + "score": 0.978 + }, + { + "category_id": 8, + "poly": [ + 528, + 1947, + 874, + 1947, + 874, + 2027, + 528, + 2027 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 297, + 292, + 1405, + 292, + 1405, + 948, + 297, + 948 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 569, + 1687, + 1170, + 1687, + 1170, + 1763, + 569, + 1763 + ], + "score": 0.923 + }, + { + "category_id": 8, + "poly": [ + 569, + 1861, + 1169, + 1861, + 1169, + 1901, + 569, + 1901 + ], + "score": 0.916 + }, + { + "category_id": 8, + "poly": [ + 571, + 1770, + 864, + 1770, + 864, + 1808, + 571, + 1808 + ], + "score": 0.912 + }, + { + "category_id": 8, + "poly": [ + 571, + 1816, + 1050, + 1816, + 1050, + 1855, + 571, + 1855 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 300, + 225, + 677, + 225, + 677, + 261, + 300, + 261 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1967, + 1401, + 1967, + 1401, + 1997, + 1365, + 1997 + ], + "score": 0.868 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1868, + 1401, + 1868, + 1401, + 1897, + 1366, + 1897 + ], + "score": 0.864 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1776, + 1400, + 1776, + 1400, + 1805, + 1366, + 1805 + ], + "score": 0.86 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1709, + 1400, + 1709, + 1400, + 1739, + 1366, + 1739 + ], + "score": 0.858 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1823, + 1400, + 1823, + 1400, + 1852, + 1366, + 1852 + ], + "score": 0.849 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.679 + }, + { + "category_id": 8, + "poly": [ + 1079, + 1689, + 1169, + 1689, + 1169, + 1718, + 1079, + 1718 + ], + "score": 0.332 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.313 + }, + { + "category_id": 13, + "poly": [ + 714, + 1024, + 852, + 1024, + 852, + 1058, + 714, + 1058 + ], + "score": 0.93, + "latex": "\\bar { A } _ { 0 } ^ { t } = x _ { t } \\forall t" + }, + { + "category_id": 14, + "poly": [ + 564, + 1684, + 1173, + 1684, + 1173, + 1904, + 564, + 1904 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } & { A _ { l } ^ { t } = \\bigg \\{ x _ { t } } & { \\mathrm { i f } l = 0 } \\\\ & { \\mathbf { M } _ { \\mathrm { A X P o o L } } ( \\operatorname { R E L U } ( \\operatorname { C o N v } ( E _ { l - 1 } ^ { t } ) ) ) } & { l > 0 } \\\\ & { \\hat { A } _ { l } ^ { t } = \\operatorname { R E L U } ( \\operatorname { C o N v } ( R _ { l } ^ { t } ) ) } \\\\ & { E _ { l } ^ { t } = [ \\operatorname { R E L U } ( A _ { l } ^ { t } - \\hat { A } _ { l } ^ { t } ) ; \\operatorname { R E L U } ( \\hat { A } _ { l } ^ { t } - A _ { l } ^ { t } ) ] } \\\\ & { R _ { l } ^ { t } = \\operatorname { C o N v L S T M } ( E _ { l } ^ { t - 1 } , R _ { l } ^ { t - 1 } , \\operatorname { U P S A M P L E } ( R _ { l + 1 } ^ { t } ) ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1280, + 682, + 1280, + 682, + 1313, + 298, + 1313 + ], + "score": 0.92, + "latex": "\\operatorname { S a t L U } ( x ; p _ { m a x } ) : = \\operatorname* { m i n } ( p _ { m a x } , x )" + }, + { + "category_id": 14, + "poly": [ + 527, + 1945, + 875, + 1945, + 875, + 2030, + 527, + 2030 + ], + "score": 0.92, + "latex": "L _ { t r a i n } = \\sum _ { t } \\lambda _ { t } \\sum _ { l } \\frac { \\lambda _ { l } } { n _ { l } } \\sum _ { n _ { l } } E _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1090, + 1055, + 1149, + 1055, + 1149, + 1090, + 1090, + 1090 + ], + "score": 0.92, + "latex": "E _ { l - 1 } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1073, + 1146, + 1132, + 1146, + 1132, + 1183, + 1073, + 1183 + ], + "score": 0.92, + "latex": "R _ { l + 1 } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1046, + 613, + 1123, + 613, + 1123, + 645, + 1046, + 645 + ], + "score": 0.91, + "latex": "( R _ { l + 1 } )" + }, + { + "category_id": 13, + "poly": [ + 1129, + 486, + 1162, + 486, + 1162, + 521, + 1129, + 521 + ], + "score": 0.9, + "latex": "\\hat { A } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1207, + 331, + 1207, + 331, + 1246, + 297, + 1246 + ], + "score": 0.9, + "latex": "\\hat { A } _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 582, + 375, + 582, + 375, + 614, + 298, + 614 + ], + "score": 0.9, + "latex": "( A _ { l + 1 } )" + }, + { + "category_id": 13, + "poly": [ + 340, + 1146, + 374, + 1146, + 374, + 1181, + 340, + 1181 + ], + "score": 0.9, + "latex": "R _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 398, + 1312, + 431, + 1312, + 431, + 1349, + 398, + 1349 + ], + "score": 0.9, + "latex": "\\hat { A } _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1283, + 1025, + 1351, + 1025, + 1351, + 1054, + 1283, + 1054 + ], + "score": 0.89, + "latex": "l > 0" + }, + { + "category_id": 13, + "poly": [ + 1203, + 1024, + 1237, + 1024, + 1237, + 1058, + 1203, + 1058 + ], + "score": 0.89, + "latex": "A _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 996, + 1280, + 1030, + 1280, + 1030, + 1313, + 996, + 1313 + ], + "score": 0.89, + "latex": "E _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 484, + 1315, + 518, + 1315, + 518, + 1349, + 484, + 1349 + ], + "score": 0.89, + "latex": "A _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 792, + 1211, + 826, + 1211, + 826, + 1246, + 792, + 1246 + ], + "score": 0.89, + "latex": "R _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 994, + 452, + 1026, + 452, + 1026, + 486, + 994, + 486 + ], + "score": 0.89, + "latex": "\\hat { A } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 448, + 1243, + 481, + 1243, + 481, + 1281, + 448, + 1281 + ], + "score": 0.88, + "latex": "\\hat { A } _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 874, + 1143, + 937, + 1143, + 937, + 1181, + 874, + 1181 + ], + "score": 0.88, + "latex": "E _ { l } ^ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 624, + 415, + 673, + 415, + 673, + 452, + 624, + 452 + ], + "score": 0.88, + "latex": "( \\hat { \\hat { A } } _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 1309, + 456, + 1342, + 456, + 1342, + 486, + 1309, + 486 + ], + "score": 0.88, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 722, + 583, + 754, + 583, + 754, + 612, + 722, + 612 + ], + "score": 0.87, + "latex": "R _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1041, + 490, + 1074, + 490, + 1074, + 521, + 1041, + 521 + ], + "score": 0.87, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 779, + 1516, + 808, + 1516, + 808, + 1546, + 779, + 1546 + ], + "score": 0.87, + "latex": "\\lambda _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1335, + 421, + 1367, + 421, + 1367, + 451, + 1335, + 451 + ], + "score": 0.87, + "latex": "R _ { l }" + }, + { + "category_id": 13, + "poly": [ + 364, + 552, + 395, + 552, + 395, + 582, + 364, + 582 + ], + "score": 0.87, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 467, + 521, + 499, + 521, + 499, + 552, + 467, + 552 + ], + "score": 0.86, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1128, + 583, + 1159, + 583, + 1159, + 612, + 1128, + 612 + ], + "score": 0.86, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1341, + 643, + 1372, + 643, + 1372, + 673, + 1341, + 673 + ], + "score": 0.86, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 934, + 1517, + 962, + 1517, + 962, + 1546, + 934, + 1546 + ], + "score": 0.86, + "latex": "\\lambda _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1243, + 643, + 1273, + 643, + 1273, + 673, + 1243, + 673 + ], + "score": 0.85, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 348, + 705, + 380, + 705, + 380, + 734, + 348, + 734 + ], + "score": 0.85, + "latex": "R _ { l }" + }, + { + "category_id": 13, + "poly": [ + 798, + 1143, + 860, + 1143, + 860, + 1181, + 798, + 1181 + ], + "score": 0.85, + "latex": "R _ { l } ^ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1047, + 386, + 1096, + 386, + 1096, + 417, + 1047, + 417 + ], + "score": 0.85, + "latex": "( A _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 935, + 999, + 964, + 999, + 964, + 1025, + 935, + 1025 + ], + "score": 0.84, + "latex": "x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1094, + 1520, + 1123, + 1520, + 1123, + 1546, + 1094, + 1546 + ], + "score": 0.83, + "latex": "n _ { l }" + }, + { + "category_id": 13, + "poly": [ + 991, + 421, + 1039, + 421, + 1039, + 451, + 991, + 451 + ], + "score": 0.83, + "latex": "( E _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 361, + 421, + 410, + 421, + 410, + 452, + 361, + 452 + ], + "score": 0.81, + "latex": "( R _ { l } )" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 221.0, + 681.0, + 221.0, + 681.0, + 268.0, + 291.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 960.0, + 1404.0, + 960.0, + 1404.0, + 999.0, + 295.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 995.0, + 934.0, + 995.0, + 934.0, + 1030.0, + 293.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 995.0, + 1406.0, + 995.0, + 1406.0, + 1030.0, + 965.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1023.0, + 713.0, + 1023.0, + 713.0, + 1059.0, + 292.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1023.0, + 1202.0, + 1023.0, + 1202.0, + 1059.0, + 853.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1023.0, + 1282.0, + 1023.0, + 1282.0, + 1059.0, + 1238.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1023.0, + 1406.0, + 1023.0, + 1406.0, + 1059.0, + 1352.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1051.0, + 1089.0, + 1051.0, + 1089.0, + 1095.0, + 292.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1051.0, + 1408.0, + 1051.0, + 1408.0, + 1095.0, + 1150.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1086.0, + 1404.0, + 1086.0, + 1404.0, + 1122.0, + 294.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1110.0, + 1405.0, + 1110.0, + 1405.0, + 1154.0, + 292.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1131.0, + 339.0, + 1131.0, + 339.0, + 1194.0, + 287.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1131.0, + 797.0, + 1131.0, + 797.0, + 1194.0, + 375.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1131.0, + 873.0, + 1131.0, + 873.0, + 1194.0, + 861.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1131.0, + 1072.0, + 1131.0, + 1072.0, + 1194.0, + 938.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1131.0, + 1410.0, + 1131.0, + 1410.0, + 1194.0, + 1133.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1174.0, + 1406.0, + 1174.0, + 1406.0, + 1214.0, + 293.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1206.0, + 296.0, + 1206.0, + 296.0, + 1250.0, + 293.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1206.0, + 791.0, + 1206.0, + 791.0, + 1250.0, + 332.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1206.0, + 1406.0, + 1206.0, + 1406.0, + 1250.0, + 827.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1246.0, + 447.0, + 1246.0, + 447.0, + 1285.0, + 293.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1246.0, + 1404.0, + 1246.0, + 1404.0, + 1285.0, + 482.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1274.0, + 297.0, + 1274.0, + 297.0, + 1319.0, + 291.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1274.0, + 995.0, + 1274.0, + 995.0, + 1319.0, + 683.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1274.0, + 1408.0, + 1274.0, + 1408.0, + 1319.0, + 1031.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1314.0, + 397.0, + 1314.0, + 397.0, + 1350.0, + 294.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1314.0, + 483.0, + 1314.0, + 483.0, + 1350.0, + 432.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1314.0, + 1404.0, + 1314.0, + 1404.0, + 1350.0, + 519.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1346.0, + 1405.0, + 1346.0, + 1405.0, + 1382.0, + 294.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1373.0, + 1406.0, + 1373.0, + 1406.0, + 1415.0, + 292.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1405.0, + 1141.0, + 1405.0, + 1141.0, + 1443.0, + 295.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1451.0, + 1406.0, + 1451.0, + 1406.0, + 1490.0, + 295.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1484.0, + 1402.0, + 1484.0, + 1402.0, + 1516.0, + 296.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1515.0, + 778.0, + 1515.0, + 778.0, + 1551.0, + 293.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1515.0, + 933.0, + 1515.0, + 933.0, + 1551.0, + 809.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1515.0, + 1093.0, + 1515.0, + 1093.0, + 1551.0, + 963.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1515.0, + 1405.0, + 1515.0, + 1405.0, + 1551.0, + 1124.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1543.0, + 1405.0, + 1543.0, + 1405.0, + 1580.0, + 293.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1579.0, + 1405.0, + 1579.0, + 1405.0, + 1610.0, + 294.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1607.0, + 1296.0, + 1607.0, + 1296.0, + 1640.0, + 293.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 289.0, + 1406.0, + 289.0, + 1406.0, + 333.0, + 294.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 325.0, + 1403.0, + 325.0, + 1403.0, + 356.0, + 295.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 390.0, + 295.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 386.0, + 1046.0, + 386.0, + 1046.0, + 421.0, + 294.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 386.0, + 1406.0, + 386.0, + 1406.0, + 421.0, + 1097.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 421.0, + 360.0, + 421.0, + 360.0, + 456.0, + 295.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 421.0, + 623.0, + 421.0, + 623.0, + 456.0, + 411.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 421.0, + 990.0, + 421.0, + 990.0, + 456.0, + 674.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 421.0, + 1334.0, + 421.0, + 1334.0, + 456.0, + 1040.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 421.0, + 1405.0, + 421.0, + 1405.0, + 456.0, + 1368.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 456.0, + 993.0, + 456.0, + 993.0, + 491.0, + 292.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 456.0, + 1308.0, + 456.0, + 1308.0, + 491.0, + 1027.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 456.0, + 1405.0, + 456.0, + 1405.0, + 491.0, + 1343.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 491.0, + 1040.0, + 491.0, + 1040.0, + 526.0, + 294.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 491.0, + 1128.0, + 491.0, + 1128.0, + 526.0, + 1075.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 526.0, + 1163.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 522.0, + 466.0, + 522.0, + 466.0, + 555.0, + 292.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 522.0, + 1405.0, + 522.0, + 1405.0, + 555.0, + 500.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 549.0, + 363.0, + 549.0, + 363.0, + 589.0, + 292.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 549.0, + 1406.0, + 549.0, + 1406.0, + 589.0, + 396.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 580.0, + 297.0, + 580.0, + 297.0, + 619.0, + 294.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 580.0, + 721.0, + 580.0, + 721.0, + 619.0, + 376.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 580.0, + 1127.0, + 580.0, + 1127.0, + 619.0, + 755.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 580.0, + 1405.0, + 580.0, + 1405.0, + 619.0, + 1160.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 611.0, + 1045.0, + 611.0, + 1045.0, + 647.0, + 292.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 611.0, + 1406.0, + 611.0, + 1406.0, + 647.0, + 1124.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 642.0, + 1242.0, + 642.0, + 1242.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 642.0, + 1340.0, + 642.0, + 1340.0, + 677.0, + 1274.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 642.0, + 1405.0, + 642.0, + 1405.0, + 677.0, + 1373.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 673.0, + 1405.0, + 673.0, + 1405.0, + 706.0, + 292.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 701.0, + 347.0, + 701.0, + 347.0, + 741.0, + 292.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 701.0, + 1406.0, + 701.0, + 1406.0, + 741.0, + 381.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 732.0, + 1405.0, + 732.0, + 1405.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 763.0, + 1405.0, + 763.0, + 1405.0, + 800.0, + 292.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 793.0, + 1405.0, + 793.0, + 1405.0, + 829.0, + 295.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 825.0, + 1405.0, + 825.0, + 1405.0, + 860.0, + 295.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 854.0, + 1406.0, + 854.0, + 1406.0, + 892.0, + 294.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 1405.0, + 887.0, + 1405.0, + 922.0, + 295.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 913.0, + 553.0, + 913.0, + 553.0, + 953.0, + 294.0, + 953.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1692, + 1404, + 1692, + 1404, + 2029, + 298, + 2029 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1232, + 1404, + 1232, + 1404, + 1478, + 298, + 1478 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1492, + 1403, + 1492, + 1403, + 1677, + 298, + 1677 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 889, + 1403, + 889, + 1403, + 1044, + 298, + 1044 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 300, + 1096, + 557, + 1096, + 557, + 1132, + 300, + 1132 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 300, + 1171, + 735, + 1171, + 735, + 1203, + 300, + 1203 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 303, + 229, + 774, + 229, + 774, + 261, + 303, + 261 + ], + "score": 0.844 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.76 + }, + { + "category_id": 5, + "poly": [ + 301, + 274, + 1406, + 274, + 1406, + 827, + 301, + 827 + ], + "score": 0.557, + "html": "
Require: Xt
1:A←xt 2:E,R←0
3:fort=1 to Tdo
4:for l = L to O do
5:if l=L then
6:Rt =CONVLSTM(Et-1,Rt-1)
7:else
8:Rt = CONVLSTM(E𝑡-1,Rt-1,UPSAMPLE(Rt+1))
9:for l= O to L do
10:if l= O then
11:At = SATLU(RELU(CONV(Rδ)))
12:else
13:At = RELU(CoNv(Rt))
14:Et = [RELU(At - At); RELU(At - A)]
15:ifl<L then
16:At+1 = MAXPOOL(CONV(E))
" + }, + { + "category_id": 0, + "poly": [ + 299, + 274, + 442, + 274, + 442, + 303, + 299, + 303 + ], + "score": 0.407 + }, + { + "category_id": 8, + "poly": [ + 301, + 274, + 1406, + 274, + 1406, + 827, + 301, + 827 + ], + "score": 0.136 + }, + { + "category_id": 13, + "poly": [ + 1212, + 551, + 1332, + 551, + 1332, + 591, + 1212, + 591 + ], + "score": 0.93, + "latex": "\\hat { A } _ { l } ^ { t } , A _ { l } ^ { t } , E _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 415, + 951, + 449, + 951, + 449, + 985, + 415, + 985 + ], + "score": 0.89, + "latex": "R _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 722, + 1968, + 754, + 1968, + 754, + 1997, + 722, + 1997 + ], + "score": 0.89, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1298, + 393, + 1332, + 393, + 1332, + 428, + 1298, + 428 + ], + "score": 0.88, + "latex": "R _ { l } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 427, + 557, + 491, + 557, + 491, + 585, + 427, + 585 + ], + "score": 0.88, + "latex": "l = 0" + }, + { + "category_id": 13, + "poly": [ + 804, + 1967, + 836, + 1967, + 836, + 1997, + 804, + 1997 + ], + "score": 0.88, + "latex": "R _ { l }" + }, + { + "category_id": 13, + "poly": [ + 832, + 983, + 864, + 983, + 864, + 1012, + 832, + 1012 + ], + "score": 0.88, + "latex": "R _ { l }" + }, + { + "category_id": 13, + "poly": [ + 916, + 983, + 948, + 983, + 948, + 1012, + 916, + 1012 + ], + "score": 0.88, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1253, + 1877, + 1282, + 1877, + 1282, + 1906, + 1253, + 1906 + ], + "score": 0.87, + "latex": "\\lambda _ { t }" + }, + { + "category_id": 13, + "poly": [ + 886, + 1845, + 969, + 1845, + 969, + 1875, + 886, + 1875 + ], + "score": 0.86, + "latex": "\\lambda _ { 0 } = 1" + }, + { + "category_id": 13, + "poly": [ + 528, + 1784, + 611, + 1784, + 611, + 1815, + 528, + 1815 + ], + "score": 0.86, + "latex": "\\lambda _ { 0 } = 1" + }, + { + "category_id": 13, + "poly": [ + 979, + 1845, + 1109, + 1845, + 1109, + 1876, + 979, + 1876 + ], + "score": 0.85, + "latex": "\\lambda _ { l > 0 } = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 451, + 587, + 515, + 587, + 515, + 614, + 451, + 614 + ], + "score": 0.82, + "latex": "l = 0" + }, + { + "category_id": 13, + "poly": [ + 622, + 1784, + 730, + 1784, + 730, + 1815, + 622, + 1815 + ], + "score": 0.81, + "latex": "\\lambda _ { l > 0 } = 0 " + }, + { + "category_id": 13, + "poly": [ + 386, + 366, + 450, + 366, + 450, + 391, + 386, + 391 + ], + "score": 0.78, + "latex": "t = 1" + }, + { + "category_id": 13, + "poly": [ + 451, + 425, + 520, + 425, + 520, + 452, + 451, + 452 + ], + "score": 0.78, + "latex": "l = L" + }, + { + "category_id": 13, + "poly": [ + 427, + 394, + 496, + 394, + 496, + 422, + 427, + 422 + ], + "score": 0.77, + "latex": "l = L" + }, + { + "category_id": 14, + "poly": [ + 465, + 449, + 850, + 449, + 850, + 491, + 465, + 491 + ], + "score": 0.71, + "latex": "R _ { L } ^ { t } = \\mathrm { C o N v L S T M } ( E _ { L } ^ { t - 1 } , R _ { L } ^ { t - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 523, + 558, + 547, + 558, + 547, + 584, + 523, + 584 + ], + "score": 0.7, + "latex": "L" + }, + { + "category_id": 14, + "poly": [ + 467, + 509, + 1068, + 509, + 1068, + 552, + 467, + 552 + ], + "score": 0.62, + "latex": "\\begin{array} { r } { \\boldsymbol { R _ { l } ^ { t } } = \\mathrm { C o N v L S T M } ( E _ { l } ^ { t - 1 } , R _ { l } ^ { t - 1 } , \\mathrm { U P S A M P L E } ( R _ { l + 1 } ^ { t } ) ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 466, + 614, + 865, + 614, + 865, + 652, + 466, + 652 + ], + "score": 0.58, + "latex": "\\hat { A } _ { 0 } ^ { t } = \\mathrm { S A T L U } ( \\mathrm { R E L U } ( \\mathrm { C O N V } ( R _ { 0 } ^ { t } ) ) )" + }, + { + "category_id": 13, + "poly": [ + 483, + 365, + 507, + 365, + 507, + 391, + 483, + 391 + ], + "score": 0.57, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 415, + 279, + 444, + 279, + 444, + 303, + 415, + 303 + ], + "score": 0.55, + "latex": "x _ { t }" + }, + { + "category_id": 14, + "poly": [ + 425, + 648, + 914, + 648, + 914, + 819, + 425, + 819 + ], + "score": 0.46, + "latex": "\\begin{array} { r l } & { \\mathbf { e l s e } } \\\\ & { \\qquad \\hat { A } _ { l } ^ { t } = \\mathrm { R E L U } ( \\mathrm { C O N V } ( R _ { l } ^ { t } ) ) } \\\\ & { E _ { l } ^ { t } = [ \\mathrm { R E L U } ( A _ { l } ^ { t } - \\hat { A } _ { l } ^ { t } ) ; \\mathrm { R E L U } ( \\hat { A } _ { l } ^ { t } - A _ { t } ^ { l } ) ] } \\\\ & { \\mathbf { i f } l < L \\mathbf { t h e n } } \\\\ & { \\qquad A _ { l + 1 } ^ { t } = \\mathrm { M A X P O O L } ( \\mathrm { C O N V } ( E _ { t } ^ { l } ) ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1101, + 560, + 1123, + 560, + 1123, + 585, + 1101, + 585 + ], + "score": 0.43, + "latex": "\\triangleright" + }, + { + "category_id": 13, + "poly": [ + 345, + 303, + 450, + 303, + 450, + 332, + 345, + 332 + ], + "score": 0.28, + "latex": "A _ { 0 } ^ { t } x _ { t }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1093.0, + 562.0, + 1093.0, + 562.0, + 1139.0, + 291.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1167.0, + 741.0, + 1167.0, + 741.0, + 1209.0, + 293.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 226.0, + 779.0, + 226.0, + 779.0, + 266.0, + 299.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 267.0, + 414.0, + 267.0, + 414.0, + 313.0, + 294.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 267.0, + 451.0, + 267.0, + 451.0, + 313.0, + 445.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1690.0, + 1405.0, + 1690.0, + 1405.0, + 1726.0, + 293.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1723.0, + 1404.0, + 1723.0, + 1404.0, + 1757.0, + 294.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1754.0, + 1406.0, + 1754.0, + 1406.0, + 1788.0, + 294.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1782.0, + 527.0, + 1782.0, + 527.0, + 1821.0, + 293.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1782.0, + 621.0, + 1782.0, + 621.0, + 1821.0, + 612.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1782.0, + 1406.0, + 1782.0, + 1406.0, + 1821.0, + 731.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1815.0, + 1406.0, + 1815.0, + 1406.0, + 1849.0, + 293.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1844.0, + 885.0, + 1844.0, + 885.0, + 1879.0, + 293.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1844.0, + 978.0, + 1844.0, + 978.0, + 1879.0, + 970.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1844.0, + 1405.0, + 1844.0, + 1405.0, + 1879.0, + 1110.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1874.0, + 1252.0, + 1874.0, + 1252.0, + 1913.0, + 293.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1874.0, + 1405.0, + 1874.0, + 1405.0, + 1913.0, + 1283.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1904.0, + 1404.0, + 1904.0, + 1404.0, + 1942.0, + 292.0, + 1942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1935.0, + 1404.0, + 1935.0, + 1404.0, + 1971.0, + 293.0, + 1971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1965.0, + 721.0, + 1965.0, + 721.0, + 2003.0, + 293.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1965.0, + 803.0, + 1965.0, + 803.0, + 2003.0, + 755.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1965.0, + 1405.0, + 1965.0, + 1405.0, + 2003.0, + 837.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1998.0, + 969.0, + 1998.0, + 969.0, + 2031.0, + 294.0, + 2031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1232.0, + 1406.0, + 1232.0, + 1406.0, + 1270.0, + 294.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1260.0, + 1405.0, + 1260.0, + 1405.0, + 1300.0, + 293.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1294.0, + 1406.0, + 1294.0, + 1406.0, + 1327.0, + 294.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1325.0, + 1408.0, + 1325.0, + 1408.0, + 1359.0, + 296.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1355.0, + 1405.0, + 1355.0, + 1405.0, + 1389.0, + 294.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1385.0, + 1406.0, + 1385.0, + 1406.0, + 1421.0, + 292.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1416.0, + 1404.0, + 1416.0, + 1404.0, + 1450.0, + 294.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1444.0, + 548.0, + 1444.0, + 548.0, + 1483.0, + 295.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1494.0, + 1404.0, + 1494.0, + 1404.0, + 1526.0, + 294.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1521.0, + 1405.0, + 1521.0, + 1405.0, + 1559.0, + 293.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1549.0, + 1407.0, + 1549.0, + 1407.0, + 1593.0, + 291.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1586.0, + 1404.0, + 1586.0, + 1404.0, + 1618.0, + 296.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1614.0, + 1402.0, + 1614.0, + 1402.0, + 1649.0, + 294.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1643.0, + 603.0, + 1643.0, + 603.0, + 1681.0, + 293.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 891.0, + 1403.0, + 891.0, + 1403.0, + 924.0, + 297.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 919.0, + 1404.0, + 919.0, + 1404.0, + 956.0, + 294.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 949.0, + 414.0, + 949.0, + 414.0, + 988.0, + 293.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 949.0, + 1405.0, + 949.0, + 1405.0, + 988.0, + 450.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 980.0, + 831.0, + 980.0, + 831.0, + 1016.0, + 293.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 980.0, + 915.0, + 980.0, + 915.0, + 1016.0, + 865.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 980.0, + 1405.0, + 980.0, + 1405.0, + 1016.0, + 949.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1010.0, + 1290.0, + 1010.0, + 1290.0, + 1049.0, + 293.0, + 1049.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1547, + 1404, + 1547, + 1404, + 2033, + 298, + 2033 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 888, + 1016, + 1398, + 1016, + 1398, + 1194, + 888, + 1194 + ], + "score": 0.98, + "html": "
MSESSIM
PredNet Lo0.01520.937
PredNet Lall0.01570.921
CNN-LSTM Enc.-Dec.0.01800.907
Copy Last Frame0.1250.631
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1227, + 1404, + 1227, + 1404, + 1530, + 298, + 1530 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 293, + 224, + 1403, + 224, + 1403, + 841, + 293, + 841 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 299, + 952, + 862, + 952, + 862, + 1226, + 299, + 1226 + ], + "score": 0.965 + }, + { + "category_id": 6, + "poly": [ + 885, + 956, + 1401, + 956, + 1401, + 1015, + 885, + 1015 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 4, + "poly": [ + 297, + 860, + 1400, + 860, + 1400, + 922, + 297, + 922 + ], + "score": 0.827 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.741 + }, + { + "category_id": 4, + "poly": [ + 874, + 830, + 969, + 830, + 969, + 857, + 874, + 857 + ], + "score": 0.636 + }, + { + "category_id": 1, + "poly": [ + 297, + 860, + 1400, + 860, + 1400, + 922, + 297, + 922 + ], + "score": 0.097 + }, + { + "category_id": 13, + "poly": [ + 884, + 1470, + 933, + 1470, + 933, + 1499, + 884, + 1499 + ], + "score": 0.89, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 680, + 1439, + 729, + 1439, + 729, + 1469, + 680, + 1469 + ], + "score": 0.89, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 511, + 1470, + 546, + 1470, + 546, + 1499, + 511, + 1499 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 762, + 1408, + 797, + 1408, + 797, + 1439, + 762, + 1439 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 908, + 1791, + 942, + 1791, + 942, + 1820, + 908, + 1820 + ], + "score": 0.88, + "latex": "R _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1277, + 1256, + 1326, + 1256, + 1326, + 1287, + 1277, + 1287 + ], + "score": 0.85, + "latex": "( E _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1790, + 348, + 1790, + 348, + 1821, + 298, + 1821 + ], + "score": 0.82, + "latex": "( R _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 969, + 1256, + 1018, + 1256, + 1018, + 1287, + 969, + 1287 + ], + "score": 0.81, + "latex": "( A _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 465, + 1882, + 508, + 1882, + 508, + 1910, + 465, + 1910 + ], + "score": 0.8, + "latex": "1 K" + }, + { + "category_id": 13, + "poly": [ + 1086, + 1851, + 1130, + 1851, + 1130, + 1879, + 1086, + 1879 + ], + "score": 0.72, + "latex": "4 K" + }, + { + "category_id": 13, + "poly": [ + 609, + 1135, + 647, + 1135, + 647, + 1163, + 609, + 1163 + ], + "score": 0.65, + "latex": "- 1" + }, + { + "category_id": 13, + "poly": [ + 938, + 831, + 972, + 831, + 972, + 856, + 938, + 856 + ], + "score": 0.56, + "latex": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 259.0, + 409.0, + 259.0, + 409.0, + 292.0, + 329.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 357.0, + 409.0, + 357.0, + 409.0, + 394.0, + 291.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 454.0, + 412.0, + 454.0, + 412.0, + 493.0, + 326.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 556.0, + 411.0, + 556.0, + 411.0, + 592.0, + 292.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 658.0, + 411.0, + 658.0, + 411.0, + 696.0, + 325.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 759.0, + 409.0, + 759.0, + 409.0, + 795.0, + 292.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 830.0, + 902.0, + 830.0, + 902.0, + 844.0, + 874.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 955.0, + 1402.0, + 955.0, + 1402.0, + 989.0, + 883.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 986.0, + 1289.0, + 986.0, + 1289.0, + 1018.0, + 885.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 856.0, + 1402.0, + 856.0, + 1402.0, + 895.0, + 295.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 886.0, + 920.0, + 886.0, + 920.0, + 927.0, + 294.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 827.0, + 937.0, + 827.0, + 937.0, + 863.0, + 872.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1580.0, + 294.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1612.0, + 292.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1608.0, + 1402.0, + 1608.0, + 1402.0, + 1642.0, + 296.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1637.0, + 1406.0, + 1637.0, + 1406.0, + 1674.0, + 292.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1665.0, + 1405.0, + 1665.0, + 1405.0, + 1706.0, + 292.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1697.0, + 1404.0, + 1697.0, + 1404.0, + 1734.0, + 293.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1729.0, + 1404.0, + 1729.0, + 1404.0, + 1762.0, + 294.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1759.0, + 1405.0, + 1759.0, + 1405.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1790.0, + 907.0, + 1790.0, + 907.0, + 1824.0, + 349.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 1790.0, + 1404.0, + 1790.0, + 1404.0, + 1824.0, + 943.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 1406.0, + 1817.0, + 1406.0, + 1857.0, + 292.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1850.0, + 1085.0, + 1850.0, + 1085.0, + 1886.0, + 292.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1886.0, + 1131.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1882.0, + 464.0, + 1882.0, + 464.0, + 1916.0, + 296.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1916.0, + 509.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1402.0, + 1944.0, + 1402.0, + 1974.0, + 296.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 784.0, + 2001.0, + 784.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1224.0, + 1406.0, + 1224.0, + 1406.0, + 1260.0, + 293.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1256.0, + 968.0, + 1256.0, + 968.0, + 1288.0, + 296.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1256.0, + 1276.0, + 1256.0, + 1276.0, + 1288.0, + 1019.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1288.0, + 1327.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1287.0, + 1404.0, + 1287.0, + 1404.0, + 1318.0, + 296.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1318.0, + 1405.0, + 1318.0, + 1405.0, + 1350.0, + 296.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1350.0, + 1405.0, + 1350.0, + 1405.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1376.0, + 1405.0, + 1376.0, + 1405.0, + 1413.0, + 293.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1409.0, + 761.0, + 1409.0, + 761.0, + 1440.0, + 296.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1409.0, + 1405.0, + 1409.0, + 1405.0, + 1440.0, + 798.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1438.0, + 679.0, + 1438.0, + 679.0, + 1472.0, + 294.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1438.0, + 1405.0, + 1438.0, + 1405.0, + 1472.0, + 730.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1468.0, + 510.0, + 1468.0, + 510.0, + 1503.0, + 292.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1468.0, + 883.0, + 1468.0, + 883.0, + 1503.0, + 547.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1468.0, + 1406.0, + 1468.0, + 1406.0, + 1503.0, + 934.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1501.0, + 434.0, + 1501.0, + 434.0, + 1533.0, + 294.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 951.0, + 867.0, + 951.0, + 867.0, + 984.0, + 296.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 981.0, + 865.0, + 981.0, + 865.0, + 1014.0, + 296.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1013.0, + 865.0, + 1013.0, + 865.0, + 1046.0, + 295.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1041.0, + 866.0, + 1041.0, + 866.0, + 1076.0, + 293.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1069.0, + 867.0, + 1069.0, + 867.0, + 1108.0, + 293.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1102.0, + 866.0, + 1102.0, + 866.0, + 1138.0, + 293.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1133.0, + 608.0, + 1133.0, + 608.0, + 1169.0, + 292.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1133.0, + 867.0, + 1133.0, + 867.0, + 1169.0, + 648.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1165.0, + 865.0, + 1165.0, + 865.0, + 1198.0, + 294.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1191.0, + 866.0, + 1191.0, + 866.0, + 1230.0, + 293.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 856.0, + 1402.0, + 856.0, + 1402.0, + 895.0, + 295.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 886.0, + 920.0, + 886.0, + 920.0, + 927.0, + 294.0, + 927.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1515, + 1403, + 1515, + 1403, + 2035, + 297, + 2035 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1404, + 229, + 1404, + 446, + 297, + 446 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1224, + 1404, + 1224, + 1404, + 1500, + 298, + 1500 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 322, + 519, + 1383, + 519, + 1383, + 1028, + 322, + 1028 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 296, + 1044, + 1404, + 1044, + 1404, + 1167, + 296, + 1167 + ], + "score": 0.967 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.78 + }, + { + "category_id": 13, + "poly": [ + 1109, + 1316, + 1201, + 1316, + 1201, + 1353, + 1109, + 1353 + ], + "score": 0.93, + "latex": "[ - \\frac { \\pi } { 2 } , \\frac { \\pi } { 2 } ]" + }, + { + "category_id": 13, + "poly": [ + 574, + 1347, + 650, + 1347, + 650, + 1381, + 574, + 1381 + ], + "score": 0.92, + "latex": "[ 0 , 2 \\pi )" + }, + { + "category_id": 13, + "poly": [ + 370, + 354, + 446, + 354, + 446, + 386, + 370, + 386 + ], + "score": 0.91, + "latex": "( \\alpha _ { p a n } )" + }, + { + "category_id": 13, + "poly": [ + 1022, + 353, + 1072, + 353, + 1072, + 383, + 1022, + 383 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 1338, + 1076, + 1392, + 1076, + 1392, + 1109, + 1338, + 1109 + ], + "score": 0.9, + "latex": "\\theta _ { p a n }" + }, + { + "category_id": 13, + "poly": [ + 1350, + 1790, + 1401, + 1790, + 1401, + 1820, + 1350, + 1820 + ], + "score": 0.89, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 915, + 1348, + 1042, + 1348, + 1042, + 1376, + 915, + 1376 + ], + "score": 0.89, + "latex": "7 \\cdot 8 = 5 6" + }, + { + "category_id": 13, + "poly": [ + 877, + 1112, + 935, + 1112, + 935, + 1137, + 877, + 1137 + ], + "score": 0.89, + "latex": "\\alpha _ { r o l l }" + }, + { + "category_id": 13, + "poly": [ + 1126, + 353, + 1161, + 353, + 1161, + 383, + 1126, + 383 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 859, + 1078, + 920, + 1078, + 920, + 1109, + 859, + 1109 + ], + "score": 0.87, + "latex": "( \\alpha _ { p a n }" + }, + { + "category_id": 13, + "poly": [ + 1113, + 556, + 1154, + 556, + 1154, + 580, + 1113, + 580 + ], + "score": 0.32, + "latex": "\\mathbf { \\delta L } _ { a l l }" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 538.0, + 397.0, + 538.0, + 397.0, + 565.0, + 357.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 538.0, + 638.0, + 538.0, + 638.0, + 564.0, + 596.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 564.0, + 710.0, + 564.0, + 710.0, + 576.0, + 683.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 533.0, + 952.0, + 533.0, + 952.0, + 562.0, + 917.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 545.0, + 1112.0, + 545.0, + 1112.0, + 588.0, + 1002.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 545.0, + 1162.0, + 545.0, + 1162.0, + 588.0, + 1155.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 580.0, + 395.0, + 580.0, + 395.0, + 608.0, + 346.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 582.0, + 634.0, + 582.0, + 634.0, + 607.0, + 586.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 571.0, + 1152.0, + 571.0, + 1152.0, + 616.0, + 1004.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 599.0, + 926.0, + 599.0, + 926.0, + 919.0, + 884.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 605.0, + 950.0, + 605.0, + 950.0, + 632.0, + 918.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 605.0, + 1293.0, + 605.0, + 1293.0, + 638.0, + 1002.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 612.0, + 666.0, + 612.0, + 666.0, + 634.0, + 657.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 624.0, + 395.0, + 624.0, + 395.0, + 651.0, + 357.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 624.0, + 632.0, + 624.0, + 632.0, + 651.0, + 595.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 633.0, + 1161.0, + 633.0, + 1161.0, + 664.0, + 1006.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 651.0, + 355.0, + 651.0, + 355.0, + 880.0, + 317.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 654.0, + 653.0, + 654.0, + 653.0, + 688.0, + 641.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 661.0, + 1217.0, + 661.0, + 1217.0, + 691.0, + 1005.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 665.0, + 395.0, + 665.0, + 395.0, + 694.0, + 346.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 666.0, + 634.0, + 666.0, + 634.0, + 695.0, + 584.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 675.0, + 950.0, + 675.0, + 950.0, + 702.0, + 918.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 710.0, + 395.0, + 710.0, + 395.0, + 735.0, + 357.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 720.0, + 487.0, + 720.0, + 487.0, + 729.0, + 473.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 716.0, + 574.0, + 716.0, + 574.0, + 728.0, + 511.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 710.0, + 636.0, + 710.0, + 636.0, + 737.0, + 595.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 729.0, + 405.0, + 729.0, + 405.0, + 752.0, + 386.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 733.0, + 445.0, + 733.0, + 445.0, + 750.0, + 430.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 729.0, + 491.0, + 729.0, + 491.0, + 752.0, + 472.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 729.0, + 534.0, + 729.0, + 534.0, + 752.0, + 514.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 731.0, + 578.0, + 731.0, + 578.0, + 752.0, + 557.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 729.0, + 644.0, + 729.0, + 644.0, + 752.0, + 625.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 731.0, + 687.0, + 731.0, + 687.0, + 751.0, + 669.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 729.0, + 732.0, + 729.0, + 732.0, + 752.0, + 711.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 729.0, + 774.0, + 729.0, + 774.0, + 752.0, + 753.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 731.0, + 816.0, + 731.0, + 816.0, + 752.0, + 797.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 746.0, + 949.0, + 746.0, + 949.0, + 770.0, + 919.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 787.0, + 397.0, + 787.0, + 397.0, + 817.0, + 357.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 770.0, + 515.0, + 770.0, + 515.0, + 802.0, + 447.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 789.0, + 638.0, + 789.0, + 638.0, + 815.0, + 596.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 772.0, + 752.0, + 772.0, + 752.0, + 807.0, + 690.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 802.0, + 573.0, + 802.0, + 573.0, + 833.0, + 450.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 816.0, + 729.0, + 816.0, + 729.0, + 825.0, + 715.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 807.0, + 814.0, + 807.0, + 814.0, + 828.0, + 730.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 815.0, + 950.0, + 815.0, + 950.0, + 842.0, + 918.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 833.0, + 394.0, + 833.0, + 394.0, + 858.0, + 347.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 832.0, + 450.0, + 832.0, + 450.0, + 847.0, + 412.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 823.0, + 565.0, + 823.0, + 565.0, + 853.0, + 451.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 828.0, + 634.0, + 828.0, + 634.0, + 861.0, + 586.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 846.0, + 578.0, + 846.0, + 578.0, + 872.0, + 406.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 875.0, + 395.0, + 875.0, + 395.0, + 902.0, + 357.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 888.0, + 488.0, + 888.0, + 488.0, + 904.0, + 473.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 879.0, + 540.0, + 879.0, + 540.0, + 904.0, + 509.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 875.0, + 634.0, + 875.0, + 634.0, + 901.0, + 595.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 879.0, + 666.0, + 879.0, + 666.0, + 901.0, + 654.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 886.0, + 949.0, + 886.0, + 949.0, + 909.0, + 919.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 916.0, + 395.0, + 916.0, + 395.0, + 946.0, + 346.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 919.0, + 632.0, + 919.0, + 632.0, + 944.0, + 586.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 960.0, + 395.0, + 960.0, + 395.0, + 988.0, + 357.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 960.0, + 634.0, + 960.0, + 634.0, + 986.0, + 596.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 960.0, + 946.0, + 960.0, + 946.0, + 976.0, + 931.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 982.0, + 406.0, + 982.0, + 406.0, + 1003.0, + 386.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 982.0, + 449.0, + 982.0, + 449.0, + 1002.0, + 428.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 979.0, + 495.0, + 979.0, + 495.0, + 1004.0, + 469.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 980.0, + 537.0, + 980.0, + 537.0, + 1003.0, + 514.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 982.0, + 577.0, + 982.0, + 577.0, + 1003.0, + 557.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 982.0, + 644.0, + 982.0, + 644.0, + 1002.0, + 625.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 982.0, + 688.0, + 982.0, + 688.0, + 1002.0, + 669.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 979.0, + 733.0, + 979.0, + 733.0, + 1004.0, + 709.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 979.0, + 776.0, + 979.0, + 776.0, + 1004.0, + 751.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 983.0, + 816.0, + 983.0, + 816.0, + 1002.0, + 797.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 978.0, + 1016.0, + 978.0, + 1016.0, + 1001.0, + 998.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 975.0, + 1123.0, + 975.0, + 1123.0, + 1003.0, + 1100.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 975.0, + 1235.0, + 975.0, + 1235.0, + 1003.0, + 1200.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 976.0, + 1340.0, + 976.0, + 1340.0, + 1003.0, + 1305.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 998.0, + 514.0, + 998.0, + 514.0, + 1033.0, + 450.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1001.0, + 1340.0, + 1001.0, + 1340.0, + 1030.0, + 990.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 519.0, + 745.0, + 519.0, + 745.0, + 557.0, + 694.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 525.0, + 511.0, + 525.0, + 511.0, + 553.5, + 452.0, + 553.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 564.0, + 532.0, + 564.0, + 532.0, + 577.0, + 509.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 576.0, + 509.0, + 576.0, + 509.0, + 595.0, + 464.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 592.5, + 683.0, + 592.5, + 683.0, + 597.0, + 653.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.75, + 896.5, + 470.75, + 896.5, + 470.75, + 921.5, + 419.75, + 921.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 901.0, + 514.0, + 901.0, + 514.0, + 910.0, + 493.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.75, + 903.0, + 562.75, + 903.0, + 562.75, + 923.5, + 511.75, + 923.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 918.5, + 443.0, + 918.5, + 443.0, + 940.5, + 376.0, + 940.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.25, + 945.5, + 659.25, + 945.5, + 659.25, + 949.5, + 625.25, + 949.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 1404.0, + 1044.0, + 1404.0, + 1079.0, + 294.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1073.0, + 858.0, + 1073.0, + 858.0, + 1112.0, + 293.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1073.0, + 1337.0, + 1073.0, + 1337.0, + 1112.0, + 921.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1073.0, + 1404.0, + 1073.0, + 1404.0, + 1112.0, + 1393.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1105.0, + 876.0, + 1105.0, + 876.0, + 1142.0, + 294.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1105.0, + 1404.0, + 1105.0, + 1404.0, + 1142.0, + 936.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1136.0, + 855.0, + 1136.0, + 855.0, + 1169.0, + 295.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1514.0, + 1405.0, + 1514.0, + 1405.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1545.0, + 1405.0, + 1545.0, + 1405.0, + 1580.0, + 293.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1612.0, + 292.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1607.0, + 1405.0, + 1607.0, + 1405.0, + 1643.0, + 293.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1637.0, + 1405.0, + 1637.0, + 1405.0, + 1672.0, + 292.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1666.0, + 1405.0, + 1666.0, + 1405.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1699.0, + 1406.0, + 1699.0, + 1406.0, + 1734.0, + 292.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1730.0, + 1405.0, + 1730.0, + 1405.0, + 1762.0, + 295.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1759.0, + 1407.0, + 1759.0, + 1407.0, + 1796.0, + 293.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1785.0, + 1349.0, + 1785.0, + 1349.0, + 1829.0, + 291.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1785.0, + 1405.0, + 1785.0, + 1405.0, + 1829.0, + 1402.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1820.0, + 1405.0, + 1820.0, + 1405.0, + 1855.0, + 292.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1408.0, + 1881.0, + 1408.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1909.0, + 1404.0, + 1909.0, + 1404.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1361.0, + 2001.0, + 1361.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 299.0, + 294.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 1404.0, + 292.0, + 1404.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 320.0, + 1404.0, + 320.0, + 1404.0, + 356.0, + 294.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 369.0, + 351.0, + 369.0, + 390.0, + 294.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 351.0, + 1021.0, + 351.0, + 1021.0, + 390.0, + 447.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 351.0, + 1125.0, + 351.0, + 1125.0, + 390.0, + 1073.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 390.0, + 1162.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 378.0, + 1405.0, + 378.0, + 1405.0, + 421.0, + 292.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 412.0, + 1042.0, + 412.0, + 1042.0, + 451.0, + 292.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1222.0, + 1404.0, + 1222.0, + 1404.0, + 1262.0, + 295.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1289.0, + 296.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1288.0, + 1404.0, + 1288.0, + 1404.0, + 1320.0, + 294.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1302.0, + 1108.0, + 1302.0, + 1108.0, + 1369.0, + 292.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1302.0, + 1414.0, + 1302.0, + 1414.0, + 1369.0, + 1202.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1346.0, + 573.0, + 1346.0, + 573.0, + 1383.0, + 294.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1346.0, + 914.0, + 1346.0, + 914.0, + 1383.0, + 651.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1346.0, + 1405.0, + 1346.0, + 1405.0, + 1383.0, + 1043.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1378.0, + 1405.0, + 1378.0, + 1405.0, + 1410.0, + 296.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1408.0, + 1406.0, + 1408.0, + 1406.0, + 1442.0, + 292.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1438.0, + 1405.0, + 1438.0, + 1405.0, + 1474.0, + 294.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1470.0, + 885.0, + 1470.0, + 885.0, + 1503.0, + 296.0, + 1503.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1226, + 1404, + 1226, + 1404, + 1711, + 298, + 1711 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 582, + 1404, + 582, + 1404, + 858, + 298, + 858 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 873, + 1405, + 873, + 1405, + 1209, + 297, + 1209 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1404, + 229, + 1404, + 475, + 297, + 475 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 805, + 1795, + 1398, + 1795, + 1398, + 1966, + 805, + 1966 + ], + "score": 0.976, + "html": "
MSESSIM
PredNet Lo3.13 × 10-30.884
PredNet Lall3.33 ×10-30.875
CNN-LSTMEnc.-Dec.3.67 × 10-30.865
Copy Last Frame7.95 × 10-30.762
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1729, + 783, + 1729, + 783, + 1974, + 299, + 1974 + ], + "score": 0.964 + }, + { + "category_id": 6, + "poly": [ + 804, + 1736, + 1401, + 1736, + 1401, + 1795, + 804, + 1795 + ], + "score": 0.948 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 302, + 522, + 718, + 522, + 718, + 553, + 302, + 553 + ], + "score": 0.881 + }, + { + "category_id": 7, + "poly": [ + 297, + 1974, + 1404, + 1974, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.683 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.297 + }, + { + "category_id": 13, + "poly": [ + 853, + 353, + 903, + 353, + 903, + 383, + 853, + 383 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 1156, + 232, + 1207, + 232, + 1207, + 261, + 1156, + 261 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 1276, + 354, + 1310, + 354, + 1310, + 383, + 1276, + 383 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1072, + 232, + 1107, + 232, + 1107, + 261, + 1072, + 261 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 869, + 996, + 972, + 996, + 972, + 1027, + 869, + 1027 + ], + "score": 0.88, + "latex": "\\beta _ { 1 } = 0 . 9" + }, + { + "category_id": 13, + "poly": [ + 721, + 1227, + 756, + 1227, + 756, + 1256, + 721, + 1256 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 982, + 996, + 1119, + 996, + 1119, + 1027, + 982, + 1027 + ], + "score": 0.88, + "latex": "\\beta _ { 2 } = 0 . 9 9 9 )" + }, + { + "category_id": 13, + "poly": [ + 1121, + 966, + 1187, + 966, + 1187, + 997, + 1121, + 997 + ], + "score": 0.88, + "latex": "( L _ { a l l } )" + }, + { + "category_id": 13, + "poly": [ + 1086, + 2002, + 1162, + 2002, + 1162, + 2033, + 1086, + 2033 + ], + "score": 0.87, + "latex": "1 4 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 953, + 2002, + 1030, + 2002, + 1030, + 2033, + 953, + 2033 + ], + "score": 0.87, + "latex": "1 4 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1317, + 1973, + 1352, + 1973, + 1352, + 2003, + 1317, + 2003 + ], + "score": 0.86, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 737, + 996, + 858, + 996, + 858, + 1026, + 737, + 1026 + ], + "score": 0.83, + "latex": "\\mathbf { \\Phi } _ { \\mathrm { { ( } } \\alpha } = 0 . 0 0 1 " + }, + { + "category_id": 13, + "poly": [ + 298, + 966, + 350, + 966, + 350, + 997, + 298, + 997 + ], + "score": 0.8, + "latex": "( L _ { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1002, + 1394, + 1002, + 1394, + 1024, + 1373, + 1024 + ], + "score": 0.66, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 547, + 445, + 574, + 445, + 574, + 474, + 547, + 474 + ], + "score": 0.57, + "latex": "\\lambda _ { l }" + }, + { + "category_id": 13, + "poly": [ + 540, + 445, + 602, + 445, + 602, + 475, + 540, + 475 + ], + "score": 0.32, + "latex": "( \\lambda _ { l } ^ { } \\mathbf { \\dot { s } } )" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 1732.0, + 1405.0, + 1732.0, + 1405.0, + 1770.0, + 805.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1764.0, + 1124.0, + 1764.0, + 1124.0, + 1798.0, + 806.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 517.0, + 725.0, + 517.0, + 725.0, + 559.0, + 293.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1316.0, + 1972.0, + 1316.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2006.0, + 1353.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 952.0, + 2002.0, + 952.0, + 2035.0, + 294.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 2002.0, + 1085.0, + 2002.0, + 1085.0, + 2035.0, + 1031.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2035.0, + 1163.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1227.0, + 720.0, + 1227.0, + 720.0, + 1258.0, + 296.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1227.0, + 1402.0, + 1227.0, + 1402.0, + 1258.0, + 757.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1256.0, + 1404.0, + 1256.0, + 1404.0, + 1291.0, + 294.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1284.0, + 1406.0, + 1284.0, + 1406.0, + 1323.0, + 292.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1352.0, + 294.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1348.0, + 1406.0, + 1348.0, + 1406.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1377.0, + 1405.0, + 1377.0, + 1405.0, + 1411.0, + 293.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1405.0, + 1405.0, + 1405.0, + 1405.0, + 1444.0, + 292.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1439.0, + 1405.0, + 1439.0, + 1405.0, + 1473.0, + 294.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1467.0, + 1406.0, + 1467.0, + 1406.0, + 1503.0, + 292.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1499.0, + 1405.0, + 1499.0, + 1405.0, + 1537.0, + 292.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1528.0, + 1406.0, + 1528.0, + 1406.0, + 1564.0, + 293.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 1404.0, + 1561.0, + 1404.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1589.0, + 1404.0, + 1589.0, + 1404.0, + 1626.0, + 293.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1621.0, + 1405.0, + 1621.0, + 1405.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1651.0, + 1406.0, + 1651.0, + 1406.0, + 1687.0, + 292.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1683.0, + 906.0, + 1683.0, + 906.0, + 1715.0, + 292.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 584.0, + 1404.0, + 584.0, + 1404.0, + 617.0, + 297.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 612.0, + 1405.0, + 612.0, + 1405.0, + 650.0, + 293.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 643.0, + 1405.0, + 643.0, + 1405.0, + 680.0, + 294.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 673.0, + 1405.0, + 673.0, + 1405.0, + 708.0, + 293.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 704.0, + 1406.0, + 704.0, + 1406.0, + 741.0, + 294.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 733.0, + 1405.0, + 733.0, + 1405.0, + 771.0, + 293.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 760.0, + 1408.0, + 760.0, + 1408.0, + 804.0, + 292.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 796.0, + 1405.0, + 796.0, + 1405.0, + 832.0, + 294.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 828.0, + 805.0, + 828.0, + 805.0, + 861.0, + 296.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 871.0, + 1406.0, + 871.0, + 1406.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 903.0, + 1405.0, + 903.0, + 1405.0, + 939.0, + 292.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 932.0, + 1405.0, + 932.0, + 1405.0, + 970.0, + 292.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 964.0, + 297.0, + 964.0, + 297.0, + 1003.0, + 294.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 964.0, + 1120.0, + 964.0, + 1120.0, + 1003.0, + 351.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 964.0, + 1405.0, + 964.0, + 1405.0, + 1003.0, + 1188.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 991.0, + 736.0, + 991.0, + 736.0, + 1034.0, + 291.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 991.0, + 868.0, + 991.0, + 868.0, + 1034.0, + 859.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 991.0, + 981.0, + 991.0, + 981.0, + 1034.0, + 973.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 991.0, + 1372.0, + 991.0, + 1372.0, + 1034.0, + 1120.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 991.0, + 1406.0, + 991.0, + 1406.0, + 1034.0, + 1395.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1060.0, + 295.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1058.0, + 1402.0, + 1058.0, + 1402.0, + 1089.0, + 295.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 1403.0, + 1087.0, + 1403.0, + 1122.0, + 294.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1114.0, + 1406.0, + 1114.0, + 1406.0, + 1154.0, + 292.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1405.0, + 1146.0, + 1405.0, + 1183.0, + 294.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 622.0, + 1178.0, + 622.0, + 1213.0, + 294.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 1071.0, + 231.0, + 1071.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 231.0, + 1155.0, + 231.0, + 1155.0, + 264.0, + 1108.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 264.0, + 1208.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 290.0, + 1404.0, + 290.0, + 1404.0, + 327.0, + 292.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 321.0, + 1404.0, + 321.0, + 1404.0, + 358.0, + 291.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 353.0, + 852.0, + 353.0, + 852.0, + 387.0, + 292.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 353.0, + 1275.0, + 353.0, + 1275.0, + 387.0, + 904.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 353.0, + 1402.0, + 353.0, + 1402.0, + 387.0, + 1311.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 382.0, + 1402.0, + 382.0, + 1402.0, + 416.0, + 295.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 416.0, + 1404.0, + 416.0, + 1404.0, + 446.0, + 295.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 445.0, + 539.0, + 445.0, + 539.0, + 478.0, + 296.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 445.0, + 1146.0, + 445.0, + 1146.0, + 478.0, + 603.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1727.0, + 785.0, + 1727.0, + 785.0, + 1761.0, + 298.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1757.0, + 786.0, + 1757.0, + 786.0, + 1791.0, + 294.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 785.0, + 1788.0, + 785.0, + 1820.0, + 294.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 786.0, + 1820.0, + 786.0, + 1852.0, + 293.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1850.0, + 786.0, + 1850.0, + 786.0, + 1884.0, + 295.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1878.0, + 786.0, + 1878.0, + 786.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1909.0, + 786.0, + 1909.0, + 786.0, + 1945.0, + 296.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1942.0, + 786.0, + 1942.0, + 786.0, + 1975.0, + 297.0, + 1975.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1727, + 1404, + 1727, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 300, + 211, + 1398, + 211, + 1398, + 1527, + 300, + 1527 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 298, + 1549, + 1401, + 1549, + 1401, + 1643, + 298, + 1643 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.787 + }, + { + "category_id": 4, + "poly": [ + 858, + 1518, + 932, + 1518, + 932, + 1541, + 858, + 1541 + ], + "score": 0.516 + }, + { + "category_id": 13, + "poly": [ + 1159, + 1943, + 1213, + 1943, + 1213, + 1972, + 1159, + 1972 + ], + "score": 0.86, + "latex": "2 9 \\%" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 257.0, + 391.0, + 257.0, + 391.0, + 278.0, + 337.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 331.0, + 391.0, + 331.0, + 391.0, + 350.0, + 317.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 415.0, + 394.0, + 415.0, + 394.0, + 438.0, + 332.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 453.0, + 394.0, + 453.0, + 394.0, + 472.0, + 377.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 484.0, + 397.0, + 484.0, + 397.0, + 512.0, + 304.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 550.0, + 1071.0, + 550.0, + 1071.0, + 589.0, + 897.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 556.0, + 1169.0, + 556.0, + 1169.0, + 577.0, + 1107.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 577.0, + 395.0, + 577.0, + 395.0, + 602.0, + 332.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 617.0, + 395.0, + 617.0, + 395.0, + 637.0, + 377.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 648.0, + 395.0, + 648.0, + 395.0, + 677.0, + 304.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 741.0, + 392.0, + 741.0, + 392.0, + 763.0, + 334.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 781.0, + 394.0, + 781.0, + 394.0, + 794.0, + 379.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 811.0, + 394.0, + 811.0, + 394.0, + 837.0, + 307.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 900.0, + 395.0, + 900.0, + 395.0, + 925.0, + 332.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 937.0, + 398.0, + 937.0, + 398.0, + 963.0, + 376.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 972.0, + 394.0, + 972.0, + 394.0, + 997.0, + 308.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1058.0, + 397.0, + 1058.0, + 397.0, + 1083.0, + 332.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1062.0, + 545.0, + 1062.0, + 545.0, + 1077.0, + 497.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1099.0, + 395.0, + 1099.0, + 395.0, + 1117.0, + 377.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 1132.0, + 392.0, + 1132.0, + 392.0, + 1155.0, + 310.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1222.0, + 394.0, + 1222.0, + 394.0, + 1276.0, + 332.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1248.0, + 758.0, + 1248.0, + 758.0, + 1263.0, + 729.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1205.0, + 1322.0, + 1205.0, + 1322.0, + 1314.0, + 779.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1295.0, + 391.0, + 1295.0, + 391.0, + 1319.0, + 320.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1382.0, + 395.0, + 1382.0, + 395.0, + 1410.0, + 298.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1348.0, + 1411.0, + 1348.0, + 1411.0, + 1532.0, + 755.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1407.0, + 611.0, + 1407.0, + 611.0, + 1487.0, + 485.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1422.0, + 397.0, + 1422.0, + 397.0, + 1447.0, + 374.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1458.0, + 392.0, + 1458.0, + 392.0, + 1483.0, + 305.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 558.0, + 853.0, + 558.0, + 853.0, + 574.0, + 827.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 553.0, + 1260.0, + 553.0, + 1260.0, + 580.0, + 1202.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.25, + 1062.0, + 638.25, + 1062.0, + 638.25, + 1077.0, + 601.25, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1550.0, + 1402.0, + 1550.0, + 1402.0, + 1584.0, + 296.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1580.0, + 1405.0, + 1580.0, + 1405.0, + 1614.0, + 294.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1611.0, + 1366.0, + 1611.0, + 1366.0, + 1644.0, + 294.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1514.0, + 937.0, + 1514.0, + 937.0, + 1547.0, + 854.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1727.0, + 1404.0, + 1727.0, + 1404.0, + 1765.0, + 294.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1761.0, + 1405.0, + 1761.0, + 1405.0, + 1793.0, + 295.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 1405.0, + 1788.0, + 1405.0, + 1825.0, + 294.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 1405.0, + 1819.0, + 1405.0, + 1859.0, + 294.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1404.0, + 1849.0, + 1404.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1919.0, + 291.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1404.0, + 1911.0, + 1404.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1158.0, + 1944.0, + 1158.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 1944.0, + 1404.0, + 1944.0, + 1404.0, + 1976.0, + 1214.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2040.0, + 292.0, + 2040.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 308, + 1404, + 308, + 1404, + 981, + 296, + 981 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1758, + 1404, + 1758, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 3, + "poly": [ + 299, + 1033, + 1386, + 1033, + 1386, + 1501, + 299, + 1501 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 297, + 1509, + 1404, + 1509, + 1404, + 1663, + 297, + 1663 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1400, + 229, + 1400, + 292, + 297, + 292 + ], + "score": 0.945 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 299, + 1693, + 529, + 1693, + 529, + 1727, + 299, + 1727 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.792 + }, + { + "category_id": 13, + "poly": [ + 761, + 830, + 812, + 830, + 812, + 859, + 761, + 859 + ], + "score": 0.91, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 900, + 738, + 950, + 738, + 950, + 768, + 900, + 768 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 561, + 582, + 615, + 582, + 615, + 613, + 561, + 613 + ], + "score": 0.88, + "latex": "7 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1109, + 520, + 1158, + 520, + 1158, + 550, + 1109, + 550 + ], + "score": 0.88, + "latex": "1 0 ^ { \\mathrm { t h } }" + }, + { + "category_id": 13, + "poly": [ + 1083, + 616, + 1119, + 616, + 1119, + 646, + 1083, + 646 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1366, + 552, + 1400, + 552, + 1400, + 582, + 1366, + 582 + ], + "score": 0.87, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 455, + 614, + 509, + 614, + 509, + 646, + 455, + 646 + ], + "score": 0.87, + "latex": "3 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 859, + 462, + 891, + 462, + 891, + 491, + 859, + 491 + ], + "score": 0.87, + "latex": "R _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1366, + 738, + 1400, + 738, + 1400, + 769, + 1366, + 769 + ], + "score": 0.87, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1287, + 769, + 1315, + 769, + 1315, + 798, + 1287, + 798 + ], + "score": 0.86, + "latex": "\\lambda _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1298, + 678, + 1332, + 678, + 1332, + 705, + 1298, + 705 + ], + "score": 0.64, + "latex": "{ \\sim } 4" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1053.0, + 365.0, + 1053.0, + 365.0, + 1064.0, + 355.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1044.0, + 871.0, + 1044.0, + 871.0, + 1072.0, + 723.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1035.0, + 946.0, + 1035.0, + 946.0, + 1059.0, + 915.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1045.0, + 1151.0, + 1045.0, + 1151.0, + 1077.0, + 952.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1032.0, + 1320.0, + 1032.0, + 1320.0, + 1085.0, + 1199.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1062.0, + 850.0, + 1062.0, + 850.0, + 1101.0, + 721.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1068.0, + 924.0, + 1068.0, + 924.0, + 1421.0, + 877.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1077.0, + 1390.0, + 1077.0, + 1390.0, + 1101.0, + 1204.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1096.0, + 368.0, + 1096.0, + 368.0, + 1111.0, + 353.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1093.0, + 1357.0, + 1093.0, + 1357.0, + 1124.0, + 1201.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1103.0, + 944.0, + 1103.0, + 944.0, + 1121.0, + 918.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1107.0, + 337.0, + 1107.0, + 337.0, + 1412.0, + 301.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1138.0, + 366.0, + 1138.0, + 366.0, + 1155.0, + 352.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1169.0, + 943.0, + 1169.0, + 943.0, + 1186.0, + 928.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1183.0, + 366.0, + 1183.0, + 366.0, + 1200.0, + 350.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1226.0, + 369.0, + 1226.0, + 369.0, + 1247.0, + 348.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1234.0, + 944.0, + 1234.0, + 944.0, + 1252.0, + 928.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1273.0, + 368.0, + 1273.0, + 368.0, + 1291.0, + 352.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1303.0, + 944.0, + 1303.0, + 944.0, + 1316.0, + 928.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1314.0, + 369.0, + 1314.0, + 369.0, + 1338.0, + 340.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1358.0, + 369.0, + 1358.0, + 369.0, + 1382.0, + 340.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1372.0, + 589.0, + 1372.0, + 589.0, + 1404.0, + 532.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1367.0, + 943.0, + 1367.0, + 943.0, + 1382.0, + 928.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1406.0, + 368.0, + 1406.0, + 368.0, + 1424.0, + 343.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1446.0, + 1029.0, + 1446.0, + 1029.0, + 1474.0, + 993.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1446.0, + 1149.0, + 1446.0, + 1149.0, + 1475.0, + 1115.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 1446.0, + 1274.0, + 1446.0, + 1274.0, + 1474.0, + 1232.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1475.0, + 1294.0, + 1475.0, + 1294.0, + 1507.0, + 959.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.25, + 1089.0, + 857.25, + 1089.0, + 857.25, + 1122.0, + 721.25, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1508.0, + 1404.0, + 1508.0, + 1404.0, + 1545.0, + 295.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1541.0, + 1404.0, + 1541.0, + 1404.0, + 1574.0, + 296.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1569.0, + 1406.0, + 1569.0, + 1406.0, + 1606.0, + 292.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1601.0, + 1404.0, + 1601.0, + 1404.0, + 1636.0, + 294.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1630.0, + 1010.0, + 1630.0, + 1010.0, + 1668.0, + 292.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1688.0, + 535.0, + 1688.0, + 535.0, + 1734.0, + 292.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 306.0, + 1402.0, + 306.0, + 1402.0, + 344.0, + 294.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 338.0, + 1405.0, + 338.0, + 1405.0, + 374.0, + 293.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 366.0, + 1404.0, + 366.0, + 1404.0, + 403.0, + 293.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 399.0, + 1404.0, + 399.0, + 1404.0, + 435.0, + 295.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 428.0, + 1406.0, + 428.0, + 1406.0, + 465.0, + 293.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 459.0, + 858.0, + 459.0, + 858.0, + 497.0, + 293.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 459.0, + 1405.0, + 459.0, + 1405.0, + 497.0, + 892.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 491.0, + 1404.0, + 491.0, + 1404.0, + 526.0, + 294.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 520.0, + 1108.0, + 520.0, + 1108.0, + 555.0, + 293.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 520.0, + 1404.0, + 520.0, + 1404.0, + 555.0, + 1159.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 552.0, + 1365.0, + 552.0, + 1365.0, + 587.0, + 294.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 552.0, + 1404.0, + 552.0, + 1404.0, + 587.0, + 1401.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 582.0, + 560.0, + 582.0, + 560.0, + 617.0, + 293.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 582.0, + 1405.0, + 582.0, + 1405.0, + 617.0, + 616.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 613.0, + 454.0, + 613.0, + 454.0, + 650.0, + 291.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 613.0, + 1082.0, + 613.0, + 1082.0, + 650.0, + 510.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 613.0, + 1405.0, + 613.0, + 1405.0, + 650.0, + 1120.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 643.0, + 1405.0, + 643.0, + 1405.0, + 681.0, + 293.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 676.0, + 1297.0, + 676.0, + 1297.0, + 712.0, + 294.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 676.0, + 1406.0, + 676.0, + 1406.0, + 712.0, + 1333.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 707.0, + 1405.0, + 707.0, + 1405.0, + 742.0, + 294.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 733.0, + 899.0, + 733.0, + 899.0, + 775.0, + 293.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 733.0, + 1365.0, + 733.0, + 1365.0, + 775.0, + 951.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 733.0, + 1405.0, + 733.0, + 1405.0, + 775.0, + 1401.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 768.0, + 1286.0, + 768.0, + 1286.0, + 803.0, + 294.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 768.0, + 1404.0, + 768.0, + 1404.0, + 803.0, + 1316.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 800.0, + 1404.0, + 800.0, + 1404.0, + 831.0, + 295.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 829.0, + 760.0, + 829.0, + 760.0, + 864.0, + 295.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 829.0, + 1405.0, + 829.0, + 1405.0, + 864.0, + 813.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 858.0, + 1404.0, + 858.0, + 1404.0, + 892.0, + 293.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 890.0, + 1404.0, + 890.0, + 1404.0, + 924.0, + 294.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 917.0, + 1406.0, + 917.0, + 1406.0, + 959.0, + 291.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 951.0, + 500.0, + 951.0, + 500.0, + 987.0, + 294.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1760.0, + 1405.0, + 1760.0, + 1405.0, + 1793.0, + 297.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1789.0, + 1406.0, + 1789.0, + 1406.0, + 1826.0, + 294.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1891.0, + 292.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1918.0, + 293.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1913.0, + 1404.0, + 1913.0, + 1404.0, + 1946.0, + 296.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1402.0, + 1944.0, + 1402.0, + 1977.0, + 296.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1406.0, + 1969.0, + 1406.0, + 2012.0, + 292.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1011.0, + 2000.0, + 1011.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1402.0, + 230.0, + 1402.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 696.0, + 261.0, + 696.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 283, + 1403, + 283, + 1403, + 406, + 299, + 406 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 296, + 1128, + 1038, + 1128, + 1038, + 1162, + 296, + 1162 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 296, + 1180, + 1402, + 1180, + 1402, + 1244, + 296, + 1244 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 295, + 995, + 1400, + 995, + 1400, + 1058, + 295, + 1058 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 305, + 1077, + 1275, + 1077, + 1275, + 1111, + 305, + 1111 + ], + "score": 0.863 + }, + { + "category_id": 1, + "poly": [ + 298, + 1427, + 1397, + 1427, + 1397, + 1490, + 298, + 1490 + ], + "score": 0.863 + }, + { + "category_id": 1, + "poly": [ + 293, + 502, + 1324, + 502, + 1324, + 536, + 293, + 536 + ], + "score": 0.856 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1399, + 1971, + 1399, + 2033, + 298, + 2033 + ], + "score": 0.852 + }, + { + "category_id": 1, + "poly": [ + 298, + 1261, + 1401, + 1261, + 1401, + 1325, + 298, + 1325 + ], + "score": 0.848 + }, + { + "category_id": 0, + "poly": [ + 299, + 452, + 489, + 452, + 489, + 486, + 299, + 486 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 297, + 1345, + 1401, + 1345, + 1401, + 1408, + 297, + 1408 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 299, + 1508, + 1356, + 1508, + 1356, + 1543, + 299, + 1543 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 298, + 1806, + 1398, + 1806, + 1398, + 1871, + 298, + 1871 + ], + "score": 0.829 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 299, + 800, + 1401, + 800, + 1401, + 865, + 299, + 865 + ], + "score": 0.82 + }, + { + "category_id": 1, + "poly": [ + 299, + 1889, + 1398, + 1889, + 1398, + 1953, + 299, + 1953 + ], + "score": 0.81 + }, + { + "category_id": 1, + "poly": [ + 296, + 554, + 1398, + 554, + 1398, + 618, + 296, + 618 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 298, + 636, + 1401, + 636, + 1401, + 699, + 298, + 699 + ], + "score": 0.798 + }, + { + "category_id": 1, + "poly": [ + 299, + 718, + 1399, + 718, + 1399, + 782, + 299, + 782 + ], + "score": 0.775 + }, + { + "category_id": 1, + "poly": [ + 295, + 1560, + 1273, + 1560, + 1273, + 1594, + 295, + 1594 + ], + "score": 0.764 + }, + { + "category_id": 1, + "poly": [ + 297, + 882, + 1406, + 882, + 1406, + 976, + 297, + 976 + ], + "score": 0.755 + }, + { + "category_id": 1, + "poly": [ + 298, + 1611, + 1399, + 1611, + 1399, + 1677, + 298, + 1677 + ], + "score": 0.751 + }, + { + "category_id": 0, + "poly": [ + 301, + 232, + 558, + 232, + 558, + 261, + 301, + 261 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 298, + 1694, + 1402, + 1694, + 1402, + 1788, + 298, + 1788 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 301, + 232, + 558, + 232, + 558, + 261, + 301, + 261 + ], + "score": 0.117 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 453.0, + 491.0, + 453.0, + 491.0, + 489.0, + 297.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 233.0, + 557.0, + 233.0, + 557.0, + 263.0, + 299.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 283.0, + 1405.0, + 283.0, + 1405.0, + 319.0, + 294.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 314.0, + 1404.0, + 314.0, + 1404.0, + 350.0, + 294.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 346.0, + 1405.0, + 346.0, + 1405.0, + 378.0, + 295.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 374.0, + 1103.0, + 374.0, + 1103.0, + 411.0, + 293.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1126.0, + 1041.0, + 1126.0, + 1041.0, + 1169.0, + 293.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1177.0, + 1402.0, + 1177.0, + 1402.0, + 1217.0, + 295.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1212.0, + 751.0, + 1212.0, + 751.0, + 1244.0, + 324.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 995.0, + 1404.0, + 995.0, + 1404.0, + 1030.0, + 294.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1020.0, + 399.0, + 1020.0, + 399.0, + 1061.0, + 318.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1074.0, + 1277.0, + 1074.0, + 1277.0, + 1115.0, + 297.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1425.0, + 1403.0, + 1425.0, + 1403.0, + 1465.0, + 294.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1457.0, + 765.0, + 1457.0, + 765.0, + 1493.0, + 324.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 501.0, + 1328.0, + 501.0, + 1328.0, + 540.0, + 294.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1968.0, + 1404.0, + 1968.0, + 1404.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1998.0, + 411.0, + 1998.0, + 411.0, + 2036.0, + 321.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1261.0, + 1402.0, + 1261.0, + 1402.0, + 1297.0, + 294.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1294.0, + 506.0, + 1294.0, + 506.0, + 1326.0, + 324.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1343.0, + 1403.0, + 1343.0, + 1403.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1377.0, + 1157.0, + 1377.0, + 1157.0, + 1408.0, + 322.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1504.0, + 1361.0, + 1504.0, + 1361.0, + 1550.0, + 291.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1806.0, + 1403.0, + 1806.0, + 1403.0, + 1844.0, + 293.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1837.0, + 1242.0, + 1837.0, + 1242.0, + 1873.0, + 324.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 800.0, + 1403.0, + 800.0, + 1403.0, + 836.0, + 295.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 830.0, + 1283.0, + 830.0, + 1283.0, + 867.0, + 324.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1889.0, + 1403.0, + 1889.0, + 1403.0, + 1925.0, + 295.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1921.0, + 1007.0, + 1921.0, + 1007.0, + 1953.0, + 324.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 554.0, + 1404.0, + 554.0, + 1404.0, + 590.0, + 296.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 583.0, + 1105.0, + 583.0, + 1105.0, + 619.0, + 322.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 633.0, + 1405.0, + 633.0, + 1405.0, + 675.0, + 293.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 668.0, + 952.0, + 668.0, + 952.0, + 701.0, + 321.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 714.0, + 1405.0, + 714.0, + 1405.0, + 756.0, + 294.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 753.0, + 624.0, + 753.0, + 624.0, + 783.0, + 320.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1560.0, + 1274.0, + 1560.0, + 1274.0, + 1596.0, + 296.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 880.0, + 1403.0, + 880.0, + 1403.0, + 918.0, + 294.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 910.0, + 1406.0, + 910.0, + 1406.0, + 952.0, + 321.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 946.0, + 1130.0, + 946.0, + 1130.0, + 976.0, + 325.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1612.0, + 1404.0, + 1612.0, + 1404.0, + 1648.0, + 296.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1641.0, + 1111.0, + 1641.0, + 1111.0, + 1678.0, + 323.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1695.0, + 1403.0, + 1695.0, + 1403.0, + 1729.0, + 296.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1726.0, + 1401.0, + 1726.0, + 1401.0, + 1760.0, + 322.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1757.0, + 469.0, + 1757.0, + 469.0, + 1788.0, + 323.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 233.0, + 557.0, + 233.0, + 557.0, + 263.0, + 299.0, + 263.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 292, + 228, + 1377, + 228, + 1377, + 264, + 292, + 264 + ], + "score": 0.775 + }, + { + "category_id": 1, + "poly": [ + 298, + 1481, + 1400, + 1481, + 1400, + 1547, + 298, + 1547 + ], + "score": 0.757 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1398, + 1971, + 1398, + 2033, + 297, + 2033 + ], + "score": 0.756 + }, + { + "category_id": 1, + "poly": [ + 295, + 1396, + 1399, + 1396, + 1399, + 1461, + 295, + 1461 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 296, + 1567, + 1402, + 1567, + 1402, + 1632, + 296, + 1632 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 296, + 711, + 1402, + 711, + 1402, + 776, + 296, + 776 + ], + "score": 0.731 + }, + { + "category_id": 1, + "poly": [ + 297, + 1652, + 1403, + 1652, + 1403, + 1747, + 297, + 1747 + ], + "score": 0.726 + }, + { + "category_id": 1, + "poly": [ + 299, + 797, + 1402, + 797, + 1402, + 862, + 299, + 862 + ], + "score": 0.721 + }, + { + "category_id": 1, + "poly": [ + 296, + 882, + 1405, + 882, + 1405, + 947, + 296, + 947 + ], + "score": 0.718 + }, + { + "category_id": 1, + "poly": [ + 297, + 1310, + 1399, + 1310, + 1399, + 1376, + 297, + 1376 + ], + "score": 0.715 + }, + { + "category_id": 1, + "poly": [ + 293, + 968, + 1402, + 968, + 1402, + 1033, + 293, + 1033 + ], + "score": 0.712 + }, + { + "category_id": 1, + "poly": [ + 298, + 571, + 1361, + 571, + 1361, + 605, + 298, + 605 + ], + "score": 0.711 + }, + { + "category_id": 1, + "poly": [ + 299, + 1768, + 1401, + 1768, + 1401, + 1863, + 299, + 1863 + ], + "score": 0.71 + }, + { + "category_id": 1, + "poly": [ + 297, + 625, + 1396, + 625, + 1396, + 691, + 297, + 691 + ], + "score": 0.698 + }, + { + "category_id": 1, + "poly": [ + 295, + 1053, + 1402, + 1053, + 1402, + 1119, + 295, + 1119 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 296, + 1224, + 1398, + 1224, + 1398, + 1291, + 296, + 1291 + ], + "score": 0.685 + }, + { + "category_id": 1, + "poly": [ + 289, + 1885, + 1399, + 1885, + 1399, + 1949, + 289, + 1949 + ], + "score": 0.682 + }, + { + "category_id": 1, + "poly": [ + 296, + 1139, + 1402, + 1139, + 1402, + 1203, + 296, + 1203 + ], + "score": 0.673 + }, + { + "category_id": 1, + "poly": [ + 299, + 485, + 1396, + 485, + 1396, + 550, + 299, + 550 + ], + "score": 0.668 + }, + { + "category_id": 1, + "poly": [ + 293, + 399, + 1401, + 399, + 1401, + 465, + 293, + 465 + ], + "score": 0.595 + }, + { + "category_id": 1, + "poly": [ + 301, + 283, + 1398, + 283, + 1398, + 379, + 301, + 379 + ], + "score": 0.56 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 1381.0, + 226.0, + 1381.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1482.0, + 1404.0, + 1482.0, + 1404.0, + 1518.0, + 297.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1514.0, + 736.0, + 1514.0, + 736.0, + 1546.0, + 326.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1973.0, + 1400.0, + 1973.0, + 1400.0, + 2005.0, + 296.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 466.0, + 2002.0, + 466.0, + 2035.0, + 322.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1396.0, + 1405.0, + 1396.0, + 1405.0, + 1432.0, + 295.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1428.0, + 851.0, + 1428.0, + 851.0, + 1461.0, + 324.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1566.0, + 1402.0, + 1566.0, + 1402.0, + 1605.0, + 293.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1600.0, + 982.0, + 1600.0, + 982.0, + 1632.0, + 321.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 710.0, + 1403.0, + 710.0, + 1403.0, + 746.0, + 296.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 742.0, + 674.0, + 742.0, + 674.0, + 778.0, + 321.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1648.0, + 1404.0, + 1648.0, + 1404.0, + 1691.0, + 292.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1721.0, + 321.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1713.0, + 456.0, + 1713.0, + 456.0, + 1748.0, + 322.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 792.0, + 1407.0, + 792.0, + 1407.0, + 839.0, + 294.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 830.0, + 829.0, + 830.0, + 829.0, + 862.0, + 322.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 881.0, + 1403.0, + 881.0, + 1403.0, + 920.0, + 293.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 911.0, + 398.0, + 911.0, + 398.0, + 949.0, + 319.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1309.0, + 1405.0, + 1309.0, + 1405.0, + 1349.0, + 293.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1342.0, + 553.0, + 1342.0, + 553.0, + 1377.0, + 322.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 967.0, + 1404.0, + 967.0, + 1404.0, + 1006.0, + 293.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1002.0, + 682.0, + 1002.0, + 682.0, + 1033.0, + 322.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 570.0, + 1361.0, + 570.0, + 1361.0, + 609.0, + 295.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1769.0, + 1405.0, + 1769.0, + 1405.0, + 1803.0, + 295.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1798.0, + 1406.0, + 1798.0, + 1406.0, + 1837.0, + 320.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1829.0, + 397.0, + 1829.0, + 397.0, + 1863.0, + 322.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 626.0, + 1401.0, + 626.0, + 1401.0, + 662.0, + 296.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 658.0, + 768.0, + 658.0, + 768.0, + 690.0, + 324.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1051.0, + 1403.0, + 1051.0, + 1403.0, + 1091.0, + 295.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1087.0, + 857.0, + 1087.0, + 857.0, + 1117.0, + 324.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1225.0, + 1402.0, + 1225.0, + 1402.0, + 1261.0, + 295.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1258.0, + 1201.0, + 1258.0, + 1201.0, + 1291.0, + 322.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1402.0, + 1882.0, + 1402.0, + 1922.0, + 295.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1917.0, + 1225.0, + 1917.0, + 1225.0, + 1949.0, + 322.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1135.0, + 1406.0, + 1135.0, + 1406.0, + 1179.0, + 291.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1171.0, + 581.0, + 1171.0, + 581.0, + 1202.0, + 323.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 487.0, + 1402.0, + 487.0, + 1402.0, + 521.0, + 294.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 519.0, + 1244.0, + 519.0, + 1244.0, + 548.0, + 321.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 400.0, + 1403.0, + 400.0, + 1403.0, + 436.0, + 296.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 432.0, + 965.0, + 432.0, + 965.0, + 466.0, + 322.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 284.0, + 1403.0, + 284.0, + 1403.0, + 319.0, + 297.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 316.0, + 1404.0, + 316.0, + 1404.0, + 351.0, + 323.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 346.0, + 914.0, + 346.0, + 914.0, + 380.0, + 322.0, + 380.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 697, + 1403, + 697, + 1403, + 789, + 297, + 789 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 296, + 1195, + 1399, + 1195, + 1399, + 1258, + 296, + 1258 + ], + "score": 0.911 + }, + { + "category_id": 1, + "poly": [ + 294, + 810, + 1403, + 810, + 1403, + 874, + 294, + 874 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 295, + 1278, + 1399, + 1278, + 1399, + 1342, + 295, + 1342 + ], + "score": 0.907 + }, + { + "category_id": 1, + "poly": [ + 297, + 892, + 1402, + 892, + 1402, + 958, + 297, + 958 + ], + "score": 0.904 + }, + { + "category_id": 1, + "poly": [ + 297, + 592, + 1003, + 592, + 1003, + 626, + 297, + 626 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 297, + 1059, + 1400, + 1059, + 1400, + 1123, + 297, + 1123 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 292, + 1361, + 1402, + 1361, + 1402, + 1425, + 292, + 1425 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 294, + 228, + 1400, + 228, + 1400, + 294, + 294, + 294 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 298, + 1142, + 1147, + 1142, + 1147, + 1176, + 298, + 1176 + ], + "score": 0.896 + }, + { + "category_id": 1, + "poly": [ + 297, + 644, + 1005, + 644, + 1005, + 678, + 297, + 678 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 294, + 976, + 1401, + 976, + 1401, + 1041, + 294, + 1041 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 295, + 1610, + 1400, + 1610, + 1400, + 1674, + 295, + 1674 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 292, + 1443, + 1401, + 1443, + 1401, + 1508, + 292, + 1508 + ], + "score": 0.874 + }, + { + "category_id": 1, + "poly": [ + 298, + 477, + 1405, + 477, + 1405, + 572, + 298, + 572 + ], + "score": 0.871 + }, + { + "category_id": 1, + "poly": [ + 304, + 364, + 1403, + 364, + 1403, + 460, + 304, + 460 + ], + "score": 0.87 + }, + { + "category_id": 1, + "poly": [ + 294, + 1526, + 1399, + 1526, + 1399, + 1591, + 294, + 1591 + ], + "score": 0.868 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 297, + 311, + 1138, + 311, + 1138, + 347, + 297, + 347 + ], + "score": 0.544 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 699.0, + 1401.0, + 699.0, + 1401.0, + 732.0, + 295.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 727.0, + 1407.0, + 727.0, + 1407.0, + 765.0, + 321.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 755.0, + 398.0, + 755.0, + 398.0, + 795.0, + 320.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1191.0, + 1404.0, + 1191.0, + 1404.0, + 1233.0, + 292.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1225.0, + 474.0, + 1225.0, + 474.0, + 1258.0, + 325.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 806.0, + 1405.0, + 806.0, + 1405.0, + 848.0, + 292.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 843.0, + 792.0, + 843.0, + 792.0, + 875.0, + 322.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1278.0, + 1403.0, + 1278.0, + 1403.0, + 1314.0, + 294.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1309.0, + 476.0, + 1309.0, + 476.0, + 1342.0, + 324.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 893.0, + 1404.0, + 893.0, + 1404.0, + 929.0, + 296.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 925.0, + 1404.0, + 925.0, + 1404.0, + 961.0, + 322.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 587.0, + 1007.0, + 587.0, + 1007.0, + 633.0, + 293.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1055.0, + 1405.0, + 1055.0, + 1405.0, + 1096.0, + 292.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1091.0, + 565.0, + 1091.0, + 565.0, + 1122.0, + 322.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1360.0, + 1404.0, + 1360.0, + 1404.0, + 1397.0, + 294.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1392.0, + 1283.0, + 1392.0, + 1283.0, + 1428.0, + 321.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 297.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 937.0, + 261.0, + 937.0, + 295.0, + 323.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1139.0, + 1151.0, + 1139.0, + 1151.0, + 1182.0, + 293.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 643.0, + 1007.0, + 643.0, + 1007.0, + 680.0, + 297.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 972.0, + 1402.0, + 972.0, + 1402.0, + 1015.0, + 292.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1007.0, + 1061.0, + 1007.0, + 1061.0, + 1043.0, + 322.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1610.0, + 1402.0, + 1610.0, + 1402.0, + 1646.0, + 295.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1641.0, + 1189.0, + 1641.0, + 1189.0, + 1677.0, + 323.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1439.0, + 1406.0, + 1439.0, + 1406.0, + 1483.0, + 294.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1474.0, + 961.0, + 1474.0, + 961.0, + 1509.0, + 323.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 477.0, + 1403.0, + 477.0, + 1403.0, + 515.0, + 295.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 548.0, + 321.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 538.0, + 395.0, + 538.0, + 395.0, + 572.0, + 320.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 365.0, + 1403.0, + 365.0, + 1403.0, + 400.0, + 299.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 397.0, + 1406.0, + 397.0, + 1406.0, + 432.0, + 323.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 424.0, + 842.0, + 424.0, + 842.0, + 463.0, + 324.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1525.0, + 1403.0, + 1525.0, + 1403.0, + 1564.0, + 294.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1557.0, + 636.0, + 1557.0, + 636.0, + 1593.0, + 322.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 310.0, + 1137.0, + 310.0, + 1137.0, + 350.0, + 295.0, + 350.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 328, + 1272, + 1372, + 1272, + 1372, + 1524, + 328, + 1524 + ], + "score": 0.981, + "html": "
MSE (x 10-3)PSNRSSIM
PredNet3.13 (3.33)25.8 (25.5)0.884 (0.878)
PredNet (no Et split)3.20 (3.37)25.6 (25.4)0.883 (0.878)
CNN-LSTMEnc.-Dec.3.67 (3.91)25.0 (24.6)0.865 (0.856)
CNN-LSTM Enc.-Dec. (2x At filts)3.82 (3.97)24.8 (24.6)0.857 (0.853)
CNN-LSTM Enc.-Dec. (except pass Eo)3.41 (3.61)25.4 (25.1)0.873 (0.866)
CNN-LSTMEnc.-Dec. (+/- split)3.71 (3.84)24.9 (24.7)0.861 (0.857)
Copy Last Frame7.9520.00.762
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1570, + 1404, + 1570, + 1404, + 1816, + 298, + 1816 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 352, + 1405, + 352, + 1405, + 629, + 298, + 629 + ], + "score": 0.971 + }, + { + "category_id": 6, + "poly": [ + 299, + 1168, + 1400, + 1168, + 1400, + 1262, + 299, + 1262 + ], + "score": 0.938 + }, + { + "category_id": 0, + "poly": [ + 299, + 226, + 500, + 226, + 500, + 261, + 299, + 261 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 300, + 1853, + 832, + 1853, + 832, + 1885, + 300, + 1885 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 104, + 299, + 104 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 301, + 296, + 755, + 296, + 755, + 328, + 301, + 328 + ], + "score": 0.848 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.842 + }, + { + "category_id": 1, + "poly": [ + 374, + 731, + 1399, + 731, + 1399, + 860, + 374, + 860 + ], + "score": 0.57 + }, + { + "category_id": 1, + "poly": [ + 378, + 868, + 1398, + 868, + 1398, + 1023, + 378, + 1023 + ], + "score": 0.558 + }, + { + "category_id": 1, + "poly": [ + 373, + 655, + 1402, + 655, + 1402, + 722, + 373, + 722 + ], + "score": 0.509 + }, + { + "category_id": 1, + "poly": [ + 374, + 1032, + 1402, + 1032, + 1402, + 1155, + 374, + 1155 + ], + "score": 0.425 + }, + { + "category_id": 13, + "poly": [ + 397, + 685, + 511, + 685, + 511, + 723, + 397, + 723 + ], + "score": 0.9, + "latex": "( \\hat { A } _ { l } - A _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 567, + 823, + 599, + 823, + 599, + 859, + 567, + 859 + ], + "score": 0.9, + "latex": "\\hat { A } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 559, + 1694, + 590, + 1694, + 590, + 1724, + 559, + 1724 + ], + "score": 0.88, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 574, + 993, + 609, + 993, + 609, + 1023, + 574, + 1023 + ], + "score": 0.87, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1026, + 764, + 1058, + 764, + 1058, + 793, + 1026, + 793 + ], + "score": 0.87, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 500, + 1065, + 533, + 1065, + 533, + 1094, + 500, + 1094 + ], + "score": 0.87, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 932, + 1725, + 963, + 1725, + 963, + 1754, + 932, + 1754 + ], + "score": 0.87, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1295, + 901, + 1327, + 901, + 1327, + 931, + 1295, + 931 + ], + "score": 0.87, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 602, + 763, + 633, + 763, + 633, + 793, + 602, + 793 + ], + "score": 0.86, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 833, + 870, + 869, + 870, + 869, + 901, + 833, + 901 + ], + "score": 0.86, + "latex": "E _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 405, + 446, + 440, + 446, + 440, + 476, + 405, + 476 + ], + "score": 0.84, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1305, + 734, + 1337, + 734, + 1337, + 763, + 1305, + 763 + ], + "score": 0.83, + "latex": "\\mathbf { \\delta } _ { \\cdot } A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 884, + 1633, + 935, + 1633, + 935, + 1664, + 884, + 1664 + ], + "score": 0.81, + "latex": "( E _ { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 1150, + 656, + 1198, + 656, + 1198, + 687, + 1150, + 687 + ], + "score": 0.76, + "latex": "( E _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 1132, + 1573, + 1165, + 1573, + 1165, + 1602, + 1132, + 1602 + ], + "score": 0.75, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 689, + 1034, + 727, + 1034, + 727, + 1063, + 689, + 1063 + ], + "score": 0.73, + "latex": "+ / -" + }, + { + "category_id": 13, + "poly": [ + 489, + 1339, + 522, + 1339, + 522, + 1368, + 489, + 1368 + ], + "score": 0.62, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 759, + 1430, + 794, + 1430, + 794, + 1459, + 759, + 1459 + ], + "score": 0.57, + "latex": "E _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 624, + 1461, + 662, + 1461, + 662, + 1488, + 624, + 1488 + ], + "score": 0.52, + "latex": "+ / -" + }, + { + "category_id": 13, + "poly": [ + 700, + 732, + 772, + 732, + 772, + 763, + 700, + 763 + ], + "score": 0.45, + "latex": "2 \\mathbf { x } \\ A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 739, + 733, + 771, + 733, + 771, + 763, + 739, + 763 + ], + "score": 0.41, + "latex": "A _ { l }" + }, + { + "category_id": 13, + "poly": [ + 516, + 1973, + 591, + 1973, + 591, + 2003, + 516, + 2003 + ], + "score": 0.25, + "latex": "6 4 \\mathrm { x 6 4 }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1166.0, + 1402.0, + 1166.0, + 1402.0, + 1203.0, + 294.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1198.0, + 1402.0, + 1198.0, + 1402.0, + 1236.0, + 294.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1228.0, + 1319.0, + 1228.0, + 1319.0, + 1265.0, + 294.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 221.0, + 507.0, + 221.0, + 507.0, + 268.0, + 291.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1853.0, + 836.0, + 1853.0, + 836.0, + 1889.0, + 294.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 761.0, + 291.0, + 761.0, + 333.0, + 294.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2086.0, + 868.0, + 2086.0, + 868.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1975.0, + 515.0, + 1975.0, + 515.0, + 2005.0, + 293.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1975.0, + 1406.0, + 1975.0, + 1406.0, + 2005.0, + 592.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1404.0, + 2004.0, + 1404.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1572.0, + 1131.0, + 1572.0, + 1131.0, + 1605.0, + 294.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1572.0, + 1404.0, + 1572.0, + 1404.0, + 1605.0, + 1166.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1603.0, + 1401.0, + 1603.0, + 1401.0, + 1633.0, + 296.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1631.0, + 883.0, + 1631.0, + 883.0, + 1669.0, + 293.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1631.0, + 1405.0, + 1631.0, + 1405.0, + 1669.0, + 936.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1662.0, + 1404.0, + 1662.0, + 1404.0, + 1699.0, + 293.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1694.0, + 558.0, + 1694.0, + 558.0, + 1728.0, + 293.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1694.0, + 1404.0, + 1694.0, + 1404.0, + 1728.0, + 591.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1724.0, + 931.0, + 1724.0, + 931.0, + 1758.0, + 294.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1724.0, + 1404.0, + 1724.0, + 1404.0, + 1758.0, + 964.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1754.0, + 1405.0, + 1754.0, + 1405.0, + 1792.0, + 293.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1784.0, + 974.0, + 1784.0, + 974.0, + 1820.0, + 293.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 353.0, + 1405.0, + 353.0, + 1405.0, + 386.0, + 297.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 383.0, + 1406.0, + 383.0, + 1406.0, + 420.0, + 295.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 416.0, + 1405.0, + 416.0, + 1405.0, + 449.0, + 296.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 443.0, + 404.0, + 443.0, + 404.0, + 480.0, + 295.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 443.0, + 1405.0, + 443.0, + 1405.0, + 480.0, + 441.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 475.0, + 1403.0, + 475.0, + 1403.0, + 511.0, + 293.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 502.0, + 1406.0, + 502.0, + 1406.0, + 542.0, + 292.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 1403.0, + 535.0, + 1403.0, + 573.0, + 292.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 566.0, + 1406.0, + 566.0, + 1406.0, + 602.0, + 292.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 598.0, + 1035.0, + 598.0, + 1035.0, + 631.0, + 296.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 729.0, + 699.0, + 729.0, + 699.0, + 768.0, + 372.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 729.0, + 1304.0, + 729.0, + 1304.0, + 768.0, + 773.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 729.0, + 1406.0, + 729.0, + 1406.0, + 768.0, + 1338.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 762.0, + 601.0, + 762.0, + 601.0, + 795.0, + 393.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 762.0, + 1025.0, + 762.0, + 1025.0, + 795.0, + 634.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 762.0, + 1406.0, + 762.0, + 1406.0, + 795.0, + 1059.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 794.0, + 1404.0, + 794.0, + 1404.0, + 828.0, + 395.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 825.0, + 566.0, + 825.0, + 566.0, + 866.0, + 393.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 825.0, + 1404.0, + 825.0, + 1404.0, + 866.0, + 600.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 869.0, + 832.0, + 869.0, + 832.0, + 905.0, + 372.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 869.0, + 1405.0, + 869.0, + 1405.0, + 905.0, + 870.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 900.0, + 1294.0, + 900.0, + 1294.0, + 934.0, + 392.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 900.0, + 1405.0, + 900.0, + 1405.0, + 934.0, + 1328.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 932.0, + 1404.0, + 932.0, + 1404.0, + 963.0, + 393.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 961.0, + 1405.0, + 961.0, + 1405.0, + 997.0, + 397.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 991.0, + 573.0, + 991.0, + 573.0, + 1027.0, + 392.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 991.0, + 1043.0, + 991.0, + 1043.0, + 1027.0, + 610.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 651.0, + 1149.0, + 651.0, + 1149.0, + 691.0, + 370.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 651.0, + 1404.0, + 651.0, + 1404.0, + 691.0, + 1199.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 686.0, + 1198.0, + 686.0, + 1198.0, + 728.0, + 512.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1030.0, + 688.0, + 1030.0, + 688.0, + 1069.0, + 371.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1030.0, + 1405.0, + 1030.0, + 1405.0, + 1069.0, + 728.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1061.0, + 499.0, + 1061.0, + 499.0, + 1098.0, + 393.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1061.0, + 1405.0, + 1061.0, + 1405.0, + 1098.0, + 534.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1092.0, + 1404.0, + 1092.0, + 1404.0, + 1131.0, + 393.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1124.0, + 569.0, + 1124.0, + 569.0, + 1160.0, + 395.0, + 1160.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 229, + 1405, + 229, + 1405, + 536, + 297, + 536 + ], + "score": 0.98 + }, + { + "category_id": 3, + "poly": [ + 352, + 1032, + 1347, + 1032, + 1347, + 1919, + 352, + 1919 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 802, + 1405, + 802, + 1405, + 924, + 297, + 924 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 296, + 1930, + 1404, + 1930, + 1404, + 2054, + 296, + 2054 + ], + "score": 0.958 + }, + { + "category_id": 5, + "poly": [ + 297, + 622, + 834, + 622, + 834, + 765, + 297, + 765 + ], + "score": 0.953, + "html": "
MSE (per-pixel)
DFN (Brabandere et al., 2016)1.71 ×10-3
PredNet1.16 ×10-3
Copy Last Frame3.58 ×10-3
" + }, + { + "category_id": 5, + "poly": [ + 904, + 621, + 1401, + 621, + 1401, + 771, + 904, + 771 + ], + "score": 0.938, + "html": "
DNA (Finn et al., 2016) PredNet FF multi-scale (Mathieu et al.,2016)42.1 38.9 26.7 32.0
" + }, + { + "category_id": 6, + "poly": [ + 900, + 559, + 1401, + 559, + 1401, + 624, + 900, + 624 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 300, + 964, + 784, + 964, + 784, + 995, + 300, + 995 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 6, + "poly": [ + 296, + 565, + 833, + 565, + 833, + 627, + 296, + 627 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.863 + }, + { + "category_id": 13, + "poly": [ + 600, + 1962, + 662, + 1962, + 662, + 1990, + 600, + 1990 + ], + "score": 0.89, + "latex": "t + 1" + }, + { + "category_id": 13, + "poly": [ + 566, + 2023, + 627, + 2023, + 627, + 2051, + 566, + 2051 + ], + "score": 0.88, + "latex": "t + 1" + }, + { + "category_id": 13, + "poly": [ + 914, + 262, + 948, + 262, + 948, + 292, + 914, + 292 + ], + "score": 0.87, + "latex": "L _ { 0 }" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1024.0, + 755.0, + 1024.0, + 755.0, + 1061.0, + 558.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1035.0, + 1199.0, + 1035.0, + 1199.0, + 1068.0, + 1027.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1093.0, + 427.0, + 1093.0, + 427.0, + 1127.0, + 350.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1159.0, + 435.0, + 1159.0, + 435.0, + 1213.0, + 364.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1191.0, + 431.0, + 1191.0, + 431.0, + 1234.0, + 348.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1261.0, + 429.0, + 1261.0, + 429.0, + 1300.0, + 361.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1292.0, + 429.0, + 1292.0, + 429.0, + 1327.0, + 348.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1369.0, + 430.0, + 1369.0, + 430.0, + 1408.0, + 349.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1439.0, + 434.0, + 1439.0, + 434.0, + 1492.0, + 362.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1470.0, + 432.0, + 1470.0, + 432.0, + 1513.0, + 348.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1536.0, + 430.0, + 1536.0, + 430.0, + 1582.0, + 359.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1567.0, + 430.0, + 1567.0, + 430.0, + 1606.0, + 348.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1642.0, + 429.0, + 1642.0, + 429.0, + 1681.0, + 350.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1718.0, + 429.0, + 1718.0, + 429.0, + 1763.0, + 364.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1747.0, + 430.0, + 1747.0, + 430.0, + 1787.0, + 350.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1823.0, + 425.0, + 1823.0, + 425.0, + 1855.0, + 364.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1849.0, + 427.0, + 1849.0, + 427.0, + 1883.0, + 350.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1897.0, + 784.0, + 1897.0, + 784.0, + 1927.0, + 634.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1894.0, + 973.0, + 1894.0, + 973.0, + 1925.0, + 909.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 1892.0, + 1086.0, + 1892.0, + 1086.0, + 1924.0, + 1022.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1892.0, + 1200.0, + 1892.0, + 1200.0, + 1924.0, + 1131.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1890.0, + 1314.0, + 1890.0, + 1314.0, + 1926.0, + 1240.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1929.0, + 1406.0, + 1929.0, + 1406.0, + 1965.0, + 293.0, + 1965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1959.0, + 599.0, + 1959.0, + 599.0, + 1994.0, + 294.0, + 1994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1959.0, + 1405.0, + 1959.0, + 1405.0, + 1994.0, + 663.0, + 1994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1992.0, + 1405.0, + 1992.0, + 1405.0, + 2025.0, + 295.0, + 2025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2021.0, + 565.0, + 2021.0, + 565.0, + 2057.0, + 294.0, + 2057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 2021.0, + 1163.0, + 2021.0, + 1163.0, + 2057.0, + 628.0, + 2057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 559.0, + 1403.0, + 559.0, + 1403.0, + 590.0, + 901.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 589.0, + 1146.0, + 589.0, + 1146.0, + 619.0, + 902.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 597.0, + 1400.0, + 597.0, + 1400.0, + 631.0, + 1319.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 962.0, + 786.0, + 962.0, + 786.0, + 998.0, + 295.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 564.0, + 831.0, + 564.0, + 831.0, + 598.0, + 295.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 595.0, + 616.0, + 595.0, + 616.0, + 629.0, + 296.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 606.0, + 829.0, + 606.0, + 829.0, + 637.0, + 637.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 913.0, + 262.0, + 913.0, + 293.0, + 295.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 262.0, + 1403.0, + 262.0, + 1403.0, + 293.0, + 949.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 1403.0, + 292.0, + 1403.0, + 324.0, + 295.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 320.0, + 1403.0, + 320.0, + 1403.0, + 355.0, + 295.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 381.0, + 1403.0, + 381.0, + 1403.0, + 417.0, + 295.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 412.0, + 1403.0, + 412.0, + 1403.0, + 450.0, + 294.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 441.0, + 1403.0, + 441.0, + 1403.0, + 478.0, + 294.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 475.0, + 1403.0, + 475.0, + 1403.0, + 507.0, + 295.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 505.0, + 942.0, + 505.0, + 942.0, + 536.0, + 296.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 836.0, + 294.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 833.0, + 1403.0, + 833.0, + 1403.0, + 865.0, + 296.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 862.0, + 1406.0, + 862.0, + 1406.0, + 901.0, + 292.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 895.0, + 408.0, + 895.0, + 408.0, + 925.0, + 295.0, + 925.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 230, + 1405, + 230, + 1405, + 871, + 297, + 871 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1728, + 1404, + 1728, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.981 + }, + { + "category_id": 3, + "poly": [ + 422, + 940, + 1224, + 940, + 1224, + 1557, + 422, + 1557 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 295, + 1575, + 1400, + 1575, + 1400, + 1637, + 295, + 1637 + ], + "score": 0.94 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 814, + 76, + 814, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.85 + }, + { + "category_id": 0, + "poly": [ + 298, + 1671, + 872, + 1671, + 872, + 1702, + 298, + 1702 + ], + "score": 0.817 + }, + { + "category_id": 13, + "poly": [ + 568, + 1607, + 629, + 1607, + 629, + 1635, + 568, + 1635 + ], + "score": 0.88, + "latex": "t + 5" + }, + { + "category_id": 13, + "poly": [ + 1004, + 293, + 1038, + 293, + 1038, + 322, + 1004, + 322 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1209, + 536, + 1242, + 536, + 1242, + 566, + 1209, + 566 + ], + "score": 0.88, + "latex": "E _ { l }" + }, + { + "category_id": 13, + "poly": [ + 705, + 841, + 768, + 841, + 768, + 869, + 705, + 869 + ], + "score": 0.87, + "latex": "t + 5" + }, + { + "category_id": 13, + "poly": [ + 927, + 1761, + 962, + 1761, + 962, + 1790, + 927, + 1790 + ], + "score": 0.87, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 742, + 992, + 800, + 992, + 800, + 1018, + 742, + 1018 + ], + "score": 0.38, + "latex": "\\mathbf { \\delta } t + 5" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 936.0, + 528.0, + 936.0, + 528.0, + 966.0, + 454.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 959.0, + 902.0, + 959.0, + 902.0, + 993.0, + 636.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 988.0, + 741.0, + 988.0, + 741.0, + 1019.0, + 635.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 988.0, + 939.0, + 988.0, + 939.0, + 1019.0, + 801.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1010.0, + 530.0, + 1010.0, + 530.0, + 1045.0, + 463.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1080.0, + 462.0, + 1080.0, + 462.0, + 1361.0, + 420.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1087.0, + 522.0, + 1087.0, + 522.0, + 1121.0, + 452.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1164.0, + 527.0, + 1164.0, + 527.0, + 1197.0, + 461.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1243.0, + 523.0, + 1243.0, + 523.0, + 1274.0, + 453.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1285.0, + 986.0, + 1285.0, + 986.0, + 1294.0, + 976.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1320.0, + 523.0, + 1320.0, + 523.0, + 1351.0, + 464.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1334.0, + 832.0, + 1334.0, + 832.0, + 1343.0, + 823.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1397.0, + 527.0, + 1397.0, + 527.0, + 1429.0, + 454.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1477.0, + 520.0, + 1477.0, + 520.0, + 1503.0, + 495.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 1494.0, + 606.0, + 1494.0, + 606.0, + 1502.0, + 595.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1495.0, + 1135.0, + 1495.0, + 1135.0, + 1502.0, + 1127.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1490.0, + 1213.0, + 1490.0, + 1213.0, + 1503.0, + 1200.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1503.0, + 534.0, + 1503.0, + 534.0, + 1527.0, + 515.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1503.0, + 611.0, + 1503.0, + 611.0, + 1527.0, + 590.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1503.0, + 686.0, + 1503.0, + 686.0, + 1528.0, + 666.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1502.0, + 765.0, + 1502.0, + 765.0, + 1529.0, + 740.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1502.0, + 840.0, + 1502.0, + 840.0, + 1529.0, + 817.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1501.0, + 917.0, + 1501.0, + 917.0, + 1529.0, + 891.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1502.0, + 989.0, + 1502.0, + 989.0, + 1527.0, + 969.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1503.0, + 1066.0, + 1503.0, + 1066.0, + 1527.0, + 1045.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1501.0, + 1143.0, + 1501.0, + 1143.0, + 1529.0, + 1120.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1501.0, + 1224.0, + 1501.0, + 1224.0, + 1530.0, + 1190.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1525.0, + 991.0, + 1525.0, + 991.0, + 1563.0, + 741.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1279.5, + 613.0, + 1279.5, + 613.0, + 1286.5, + 591.0, + 1286.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1352.0, + 557.0, + 1352.0, + 557.0, + 1363.0, + 512.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1573.0, + 1402.0, + 1573.0, + 1402.0, + 1609.0, + 295.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1607.0, + 567.0, + 1607.0, + 567.0, + 1639.0, + 296.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1607.0, + 757.0, + 1607.0, + 757.0, + 1639.0, + 630.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1669.0, + 874.0, + 1669.0, + 874.0, + 1706.0, + 294.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 231.0, + 1403.0, + 231.0, + 1403.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 258.0, + 1405.0, + 258.0, + 1405.0, + 301.0, + 291.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1003.0, + 291.0, + 1003.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 328.0, + 1039.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 323.0, + 1403.0, + 323.0, + 1403.0, + 355.0, + 295.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1406.0, + 352.0, + 1406.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 379.0, + 1405.0, + 379.0, + 1405.0, + 421.0, + 291.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 439.0, + 1406.0, + 439.0, + 1406.0, + 482.0, + 291.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 472.0, + 1406.0, + 472.0, + 1406.0, + 509.0, + 292.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 505.0, + 1403.0, + 505.0, + 1403.0, + 538.0, + 295.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 532.0, + 1208.0, + 532.0, + 1208.0, + 570.0, + 292.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 532.0, + 1406.0, + 532.0, + 1406.0, + 570.0, + 1243.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 565.0, + 1405.0, + 565.0, + 1405.0, + 601.0, + 295.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 594.0, + 1407.0, + 594.0, + 1407.0, + 632.0, + 294.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 626.0, + 1405.0, + 626.0, + 1405.0, + 662.0, + 295.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 652.0, + 1406.0, + 652.0, + 1406.0, + 697.0, + 292.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 687.0, + 1405.0, + 687.0, + 1405.0, + 723.0, + 294.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 718.0, + 1406.0, + 718.0, + 1406.0, + 754.0, + 294.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 745.0, + 1406.0, + 745.0, + 1406.0, + 787.0, + 291.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 778.0, + 1405.0, + 778.0, + 1405.0, + 814.0, + 295.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 810.0, + 1402.0, + 810.0, + 1402.0, + 843.0, + 296.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 841.0, + 704.0, + 841.0, + 704.0, + 873.0, + 296.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 841.0, + 894.0, + 841.0, + 894.0, + 873.0, + 769.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1726.0, + 1405.0, + 1726.0, + 1405.0, + 1762.0, + 294.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1759.0, + 926.0, + 1759.0, + 926.0, + 1795.0, + 295.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1759.0, + 1407.0, + 1759.0, + 1407.0, + 1795.0, + 963.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1788.0, + 1406.0, + 1788.0, + 1406.0, + 1824.0, + 295.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1821.0, + 1406.0, + 1821.0, + 1406.0, + 1853.0, + 293.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1886.0, + 295.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2005.0, + 296.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1431, + 1404, + 1431, + 1404, + 1678, + 297, + 1678 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 984, + 1404, + 984, + 1404, + 1169, + 298, + 1169 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1184, + 1403, + 1184, + 1403, + 1337, + 299, + 1337 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 468, + 263, + 1227, + 263, + 1227, + 885, + 468, + 885 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 349, + 920, + 1345, + 920, + 1345, + 953, + 349, + 953 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 299, + 1374, + 875, + 1374, + 875, + 1407, + 299, + 1407 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.85 + }, + { + "category_id": 13, + "poly": [ + 507, + 1524, + 557, + 1524, + 557, + 1554, + 507, + 1554 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 1023, + 1434, + 1073, + 1434, + 1073, + 1463, + 1023, + 1463 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 388, + 1586, + 438, + 1586, + 438, + 1615, + 388, + 1615 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 487, + 1377, + 536, + 1377, + 536, + 1406, + 487, + 1406 + ], + "score": 0.89, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 918, + 1525, + 953, + 1525, + 953, + 1555, + 918, + 1555 + ], + "score": 0.89, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1007, + 1464, + 1041, + 1464, + 1041, + 1493, + 1007, + 1493 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 976, + 1585, + 1011, + 1585, + 1011, + 1615, + 976, + 1615 + ], + "score": 0.87, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1169, + 1524, + 1210, + 1524, + 1210, + 1553, + 1169, + 1553 + ], + "score": 0.86, + "latex": "6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1274, + 1077, + 1315, + 1077, + 1315, + 1105, + 1274, + 1105 + ], + "score": 0.85, + "latex": "5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1079, + 1524, + 1119, + 1524, + 1119, + 1553, + 1079, + 1553 + ], + "score": 0.85, + "latex": "3 \\%" + }, + { + "category_id": 13, + "poly": [ + 557, + 1464, + 646, + 1464, + 646, + 1494, + 557, + 1494 + ], + "score": 0.81, + "latex": "\\lambda _ { 0 } = 1" + }, + { + "category_id": 13, + "poly": [ + 660, + 1464, + 797, + 1464, + 797, + 1494, + 660, + 1494 + ], + "score": 0.62, + "latex": "\\lambda _ { l > 0 } = 0 . 1 " + }, + { + "category_id": 13, + "poly": [ + 959, + 334, + 983, + 334, + 983, + 352, + 959, + 352 + ], + "score": 0.47, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 960, + 410, + 993, + 410, + 993, + 428, + 960, + 428 + ], + "score": 0.36, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 959, + 360, + 983, + 360, + 983, + 377, + 959, + 377 + ], + "score": 0.33, + "latex": "L _ { 0 }" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 264.0, + 522.0, + 264.0, + 522.0, + 288.0, + 504.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 318.0, + 541.0, + 318.0, + 541.0, + 337.0, + 522.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 325.0, + 958.0, + 325.0, + 958.0, + 356.0, + 870.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 325.0, + 1022.0, + 325.0, + 1022.0, + 356.0, + 984.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 337.0, + 511.0, + 337.0, + 511.0, + 756.0, + 458.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 341.0, + 525.0, + 341.0, + 525.0, + 366.0, + 502.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 350.0, + 958.0, + 350.0, + 958.0, + 382.0, + 868.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 350.0, + 1024.0, + 350.0, + 1024.0, + 382.0, + 984.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 360.0, + 878.0, + 360.0, + 878.0, + 372.0, + 864.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 375.0, + 1033.0, + 375.0, + 1033.0, + 406.0, + 870.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 406.0, + 543.0, + 406.0, + 543.0, + 440.0, + 520.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 411.0, + 858.0, + 411.0, + 858.0, + 424.0, + 845.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 401.0, + 959.0, + 401.0, + 959.0, + 433.0, + 869.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 401.0, + 1043.0, + 401.0, + 1043.0, + 433.0, + 994.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 421.0, + 520.0, + 421.0, + 520.0, + 439.0, + 505.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 445.0, + 546.0, + 445.0, + 546.0, + 453.0, + 539.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 424.0, + 1052.0, + 424.0, + 1052.0, + 457.0, + 872.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 462.0, + 858.0, + 462.0, + 858.0, + 475.0, + 846.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 454.0, + 1060.0, + 454.0, + 1060.0, + 481.0, + 874.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 485.0, + 859.0, + 485.0, + 859.0, + 502.0, + 844.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 477.0, + 1047.0, + 477.0, + 1047.0, + 507.0, + 873.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 495.0, + 522.0, + 495.0, + 522.0, + 518.0, + 504.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 539.0, + 538.0, + 539.0, + 538.0, + 558.0, + 523.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 555.0, + 542.0, + 555.0, + 542.0, + 568.0, + 529.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 574.0, + 520.0, + 574.0, + 520.0, + 594.0, + 504.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 644.0, + 537.0, + 644.0, + 537.0, + 655.0, + 523.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 652.0, + 521.0, + 652.0, + 521.0, + 670.0, + 505.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 726.0, + 523.0, + 726.0, + 523.0, + 749.0, + 504.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 782.0, + 730.0, + 782.0, + 730.0, + 792.0, + 721.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 792.0, + 829.0, + 792.0, + 829.0, + 801.0, + 820.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 795.0, + 927.0, + 795.0, + 927.0, + 806.0, + 915.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 805.0, + 521.0, + 805.0, + 521.0, + 825.0, + 505.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 834.0, + 536.0, + 834.0, + 536.0, + 848.0, + 525.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 828.0, + 639.0, + 828.0, + 639.0, + 852.0, + 620.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 828.0, + 736.0, + 828.0, + 736.0, + 852.0, + 717.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 829.0, + 835.0, + 829.0, + 835.0, + 852.0, + 816.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 828.0, + 932.0, + 828.0, + 932.0, + 853.0, + 913.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 829.0, + 1029.0, + 829.0, + 1029.0, + 852.0, + 1011.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 829.0, + 1225.0, + 829.0, + 1225.0, + 852.0, + 1206.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 851.0, + 999.0, + 851.0, + 999.0, + 886.0, + 748.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 916.0, + 1349.0, + 916.0, + 1349.0, + 959.0, + 347.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1374.0, + 486.0, + 1374.0, + 486.0, + 1410.0, + 295.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1374.0, + 879.0, + 1374.0, + 879.0, + 1410.0, + 537.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1433.0, + 1022.0, + 1433.0, + 1022.0, + 1467.0, + 296.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1433.0, + 1404.0, + 1433.0, + 1404.0, + 1467.0, + 1074.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1463.0, + 556.0, + 1463.0, + 556.0, + 1497.0, + 295.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1463.0, + 659.0, + 1463.0, + 659.0, + 1497.0, + 647.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1463.0, + 1006.0, + 1463.0, + 1006.0, + 1497.0, + 798.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1463.0, + 1404.0, + 1463.0, + 1404.0, + 1497.0, + 1042.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1495.0, + 1404.0, + 1495.0, + 1404.0, + 1526.0, + 296.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1524.0, + 506.0, + 1524.0, + 506.0, + 1558.0, + 295.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1524.0, + 917.0, + 1524.0, + 917.0, + 1558.0, + 558.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1524.0, + 1078.0, + 1524.0, + 1078.0, + 1558.0, + 954.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1524.0, + 1168.0, + 1524.0, + 1168.0, + 1558.0, + 1120.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1524.0, + 1405.0, + 1524.0, + 1405.0, + 1558.0, + 1211.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1553.0, + 1406.0, + 1553.0, + 1406.0, + 1589.0, + 294.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1585.0, + 387.0, + 1585.0, + 387.0, + 1618.0, + 295.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1585.0, + 975.0, + 1585.0, + 975.0, + 1618.0, + 439.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1585.0, + 1405.0, + 1585.0, + 1405.0, + 1618.0, + 1012.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1616.0, + 1406.0, + 1616.0, + 1406.0, + 1650.0, + 295.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1646.0, + 1141.0, + 1646.0, + 1141.0, + 1680.0, + 295.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 984.0, + 1405.0, + 984.0, + 1405.0, + 1019.0, + 294.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1017.0, + 1405.0, + 1017.0, + 1405.0, + 1046.0, + 294.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1045.0, + 1406.0, + 1045.0, + 1406.0, + 1079.0, + 294.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1075.0, + 1273.0, + 1075.0, + 1273.0, + 1110.0, + 293.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1075.0, + 1405.0, + 1075.0, + 1405.0, + 1110.0, + 1316.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1106.0, + 1406.0, + 1106.0, + 1406.0, + 1141.0, + 294.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1136.0, + 950.0, + 1136.0, + 950.0, + 1172.0, + 295.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1183.0, + 1404.0, + 1183.0, + 1404.0, + 1218.0, + 295.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1212.0, + 1404.0, + 1212.0, + 1404.0, + 1249.0, + 294.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1245.0, + 1407.0, + 1245.0, + 1407.0, + 1282.0, + 294.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1271.0, + 1405.0, + 1271.0, + 1405.0, + 1315.0, + 292.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1305.0, + 627.0, + 1305.0, + 627.0, + 1342.0, + 295.0, + 1342.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 298, + 592, + 1404, + 592, + 1404, + 1541, + 298, + 1541 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 296, + 1562, + 1405, + 1562, + 1405, + 1686, + 296, + 1686 + ], + "score": 0.948 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2112, + 836, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 4, + "poly": [ + 887, + 1529, + 987, + 1529, + 987, + 1556, + 887, + 1556 + ], + "score": 0.511 + }, + { + "category_id": 13, + "poly": [ + 400, + 1655, + 450, + 1655, + 450, + 1685, + 400, + 1685 + ], + "score": 0.92, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 789, + 1564, + 839, + 1564, + 839, + 1593, + 789, + 1593 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 471, + 1625, + 521, + 1625, + 521, + 1654, + 471, + 1654 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 582, + 1593, + 697, + 1593, + 697, + 1624, + 582, + 1624 + ], + "score": 0.89, + "latex": "{ \\cal L } _ { a l l } { - \\cal L } _ { 0 } { } ^ { \\cdots }" + }, + { + "category_id": 13, + "poly": [ + 324, + 1594, + 359, + 1594, + 359, + 1624, + 324, + 1624 + ], + "score": 0.89, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1012, + 1625, + 1047, + 1625, + 1047, + 1654, + 1012, + 1654 + ], + "score": 0.89, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1366, + 1594, + 1401, + 1594, + 1401, + 1624, + 1366, + 1624 + ], + "score": 0.87, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 418, + 1338, + 466, + 1338, + 466, + 1369, + 418, + 1369 + ], + "score": 0.81, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 372, + 983, + 469, + 983, + 469, + 1016, + 372, + 1016 + ], + "score": 0.81, + "latex": "L _ { a l l } - L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 370, + 1445, + 467, + 1445, + 467, + 1479, + 370, + 1479 + ], + "score": 0.79, + "latex": "L _ { a l l } - L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 432, + 1218, + 464, + 1218, + 464, + 1250, + 432, + 1250 + ], + "score": 0.75, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 420, + 874, + 468, + 874, + 468, + 907, + 420, + 907 + ], + "score": 0.74, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 434, + 755, + 466, + 755, + 466, + 787, + 434, + 787 + ], + "score": 0.6, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 955, + 1531, + 991, + 1531, + 991, + 1554, + 955, + 1554 + ], + "score": 0.5, + "latex": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 634.0, + 477.0, + 634.0, + 477.0, + 672.0, + 385.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 739.0, + 433.0, + 739.0, + 433.0, + 798.0, + 323.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 739.0, + 478.0, + 739.0, + 478.0, + 798.0, + 467.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 970.0, + 371.0, + 970.0, + 371.0, + 1027.0, + 295.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 970.0, + 479.0, + 970.0, + 479.0, + 1027.0, + 470.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1094.0, + 478.0, + 1094.0, + 478.0, + 1137.0, + 380.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1203.0, + 431.0, + 1203.0, + 431.0, + 1260.0, + 320.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1203.0, + 476.0, + 1203.0, + 476.0, + 1260.0, + 465.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1323.0, + 417.0, + 1323.0, + 417.0, + 1381.0, + 306.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1323.0, + 477.0, + 1323.0, + 477.0, + 1381.0, + 467.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1434.0, + 369.0, + 1434.0, + 369.0, + 1489.0, + 292.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1434.0, + 477.0, + 1434.0, + 477.0, + 1489.0, + 468.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1527.0, + 953.0, + 1527.0, + 953.0, + 1547.0, + 885.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.75, + 866.0, + 474.75, + 866.0, + 474.75, + 911.5, + 308.75, + 911.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1560.0, + 788.0, + 1560.0, + 788.0, + 1600.0, + 293.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1600.0, + 840.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 323.0, + 1591.0, + 323.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1591.0, + 581.0, + 1591.0, + 581.0, + 1628.0, + 360.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1591.0, + 1365.0, + 1591.0, + 1365.0, + 1628.0, + 698.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1623.0, + 470.0, + 1623.0, + 470.0, + 1659.0, + 294.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1623.0, + 1011.0, + 1623.0, + 1011.0, + 1659.0, + 522.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1623.0, + 1405.0, + 1623.0, + 1405.0, + 1659.0, + 1048.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1653.0, + 399.0, + 1653.0, + 399.0, + 1689.0, + 294.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1653.0, + 579.0, + 1653.0, + 579.0, + 1689.0, + 451.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1525.0, + 954.0, + 1525.0, + 954.0, + 1561.0, + 883.0, + 1561.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 296, + 451, + 1401, + 451, + 1401, + 1669, + 296, + 1669 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 296, + 1697, + 1406, + 1697, + 1406, + 1822, + 296, + 1822 + ], + "score": 0.957 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.85 + }, + { + "category_id": 4, + "poly": [ + 909, + 1661, + 1011, + 1661, + 1011, + 1689, + 909, + 1689 + ], + "score": 0.363 + }, + { + "category_id": 13, + "poly": [ + 583, + 1791, + 633, + 1791, + 633, + 1820, + 583, + 1820 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 743, + 1729, + 875, + 1729, + 875, + 1760, + 743, + 1760 + ], + "score": 0.9, + "latex": "{ \\cal L } _ { a l l } - { \\cal L } _ { 0 } { } ^ { }" + }, + { + "category_id": 13, + "poly": [ + 671, + 1761, + 721, + 1761, + 721, + 1790, + 671, + 1790 + ], + "score": 0.9, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 830, + 1699, + 880, + 1699, + 880, + 1729, + 830, + 1729 + ], + "score": 0.89, + "latex": "L _ { a l l }" + }, + { + "category_id": 13, + "poly": [ + 1201, + 1761, + 1235, + 1761, + 1235, + 1790, + 1201, + 1790 + ], + "score": 0.89, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 466, + 1730, + 500, + 1730, + 500, + 1759, + 466, + 1759 + ], + "score": 0.88, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 464, + 1761, + 498, + 1761, + 498, + 1790, + 464, + 1790 + ], + "score": 0.87, + "latex": "L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 368, + 1560, + 464, + 1560, + 464, + 1592, + 368, + 1592 + ], + "score": 0.83, + "latex": "L _ { a l l } - L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 368, + 957, + 464, + 957, + 464, + 991, + 368, + 991 + ], + "score": 0.57, + "latex": "L _ { a l l } - L _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 979, + 1663, + 1014, + 1663, + 1014, + 1687, + 979, + 1687 + ], + "score": 0.47, + "latex": "" + }, + { + "category_id": 13, + "poly": [ + 415, + 1418, + 462, + 1418, + 462, + 1449, + 415, + 1449 + ], + "score": 0.47, + "latex": "L _ { a l l }" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 507.0, + 469.0, + 507.0, + 469.0, + 549.0, + 376.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 653.0, + 475.0, + 653.0, + 475.0, + 709.0, + 323.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 802.0, + 473.0, + 802.0, + 473.0, + 865.0, + 302.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 947.0, + 367.0, + 947.0, + 367.0, + 1001.0, + 288.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 947.0, + 473.0, + 947.0, + 473.0, + 1001.0, + 465.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1105.0, + 472.0, + 1105.0, + 472.0, + 1152.0, + 374.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1256.0, + 474.0, + 1256.0, + 474.0, + 1309.0, + 325.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1402.0, + 414.0, + 1402.0, + 414.0, + 1460.0, + 304.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1402.0, + 474.0, + 1402.0, + 474.0, + 1460.0, + 463.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1548.0, + 367.0, + 1548.0, + 367.0, + 1601.0, + 290.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1548.0, + 474.0, + 1548.0, + 474.0, + 1601.0, + 465.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1699.0, + 829.0, + 1699.0, + 829.0, + 1731.0, + 295.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1699.0, + 1404.0, + 1699.0, + 1404.0, + 1731.0, + 881.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1726.0, + 465.0, + 1726.0, + 465.0, + 1764.0, + 293.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 1726.0, + 742.0, + 1726.0, + 742.0, + 1764.0, + 501.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1726.0, + 1407.0, + 1726.0, + 1407.0, + 1764.0, + 876.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1761.0, + 463.0, + 1761.0, + 463.0, + 1793.0, + 294.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1761.0, + 670.0, + 1761.0, + 670.0, + 1793.0, + 499.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1761.0, + 1200.0, + 1761.0, + 1200.0, + 1793.0, + 722.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1761.0, + 1404.0, + 1761.0, + 1404.0, + 1793.0, + 1236.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1791.0, + 582.0, + 1791.0, + 582.0, + 1824.0, + 295.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1791.0, + 762.0, + 1791.0, + 762.0, + 1824.0, + 634.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1656.0, + 978.0, + 1656.0, + 978.0, + 1696.0, + 905.0, + 1696.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/B1xm3RVtwB/B1xm3RVtwB.md b/parse/train/B1xm3RVtwB/B1xm3RVtwB.md new file mode 100644 index 0000000000000000000000000000000000000000..c22428bb2352f69b59bf5a717f6c2df1ff9dd550 --- /dev/null +++ b/parse/train/B1xm3RVtwB/B1xm3RVtwB.md @@ -0,0 +1,279 @@ +# SIMPLIFIED ACTION DECODER FOR DEEP MULTI-AGENT REINFORCEMENT LEARNING + +Hengyuan Hu, Jakob N Foerster Facebook AI Research, CA, USA {hengyuan,jnf}@fb.com + +# ABSTRACT + +In recent years we have seen fast progress on a number of benchmark problems in AI, with modern methods achieving near or super human performance in Go, Poker and Dota. One common aspect of all of these challenges is that they are by design adversarial or, technically speaking, zero-sum. In contrast to these settings, success in the real world commonly requires humans to collaborate and communicate with others, in settings that are, at least partially, cooperative. In the last year, the card game Hanabi has been established as a new benchmark environment for AI to fill this gap. In particular, Hanabi is interesting to humans since it is entirely focused on theory of mind, i.e., the ability to effectively reason over the intentions, beliefs and point of view of other agents when observing their actions. Learning to be informative when observed by others is an interesting challenge for Reinforcement Learning (RL): Fundamentally, RL requires agents to explore in order to discover good policies. However, when done naively, this randomness will inherently make their actions less informative to others during training. We present a new deep multi-agent RL method, the Simplified Action Decoder (SAD), which resolves this contradiction exploiting the centralized training phase. During training SAD allows other agents to not only observe the (exploratory) action chosen, but agents instead also observe the greedy action of their team mates. By combining this simple intuition with best practices for multi-agent learning, SAD establishes a new SOTA for learning methods for 2-5 players on the self-play part of the Hanabi challenge. Our ablations show the contributions of SAD compared with the best practice components. All of our code and trained agents are available at https://github.com/facebookresearch/Hanabi_SAD. + +# 1 INTRODUCTION + +Humans are highly social creatures and spend vast amounts of time coordinating, collaborating and communicating with others. In contrast to these, at least partially, cooperative settings most progress on AI in games has been in zero-sum games where agents compete against each other, typically rendering communication futile. This includes examples such as Go (Silver et al., 2016; 2017; 2018), poker (Brown & Sandholm, 2017; Moravcík et al., 2017; Brown & Sandholm, 2019) ˇ and chess (Campbell et al., 2002). + +This narrow focus is unfortunate, since communication and coordination require unique abilities. In order to enable smooth and efficient social interactions of groups of people, it is commonly required to reason over the intents, points of views and beliefs of other agents from observing their actions. For example, a driver can reasonably infer that if a truck in front of them is slowing down when approaching an intersection, then there is likely an obstacle ahead. Furthermore, humans are both able to interpret the actions of others and can act in a way that is informative when their actions are being observed by others, capabilities that are commonly called theory of Mind (ToM), (Baker et al., 2017). Importantly, in order to carry out this kind of reasoning, an agent needs to consider why a given action is taken and what this decision indicates about the state of the world. Simply observing what other agents are doing is not sufficient. + +While these abilities are particularly relevant in partially observable, fully cooperative multi-agent settings, ToM reasoning clearly matters in a variety of real world scenarios. For example, autonomous cars will likely need to understand the point of view, intents and beliefs of other traffic participants in order to deal with highly interactive settings such as 4-way crossing or dense traffic in cities. + +Hanabi is a fully cooperative, partially-observable card game that has recently been proposed as a new benchmark challenge problem for AI research (Bard et al., 2019) to fill the gap around ToM. In Hanabi, players need to find conventions that allow them to effectively exchange information from their local observations through their actions, taking advantage of the fact that actions are observed by all team mates. + +Most prior state-of-the-art agents for Hanabi were developed using handcrafted algorithms, which beat off-the-shelf deep multi-agent RL methods by a large margin. This makes intuitive sense: Beyond the “standard” multi-agent challenges of credit assignment, nonstationarity and joint exploration, learning an informative policy presents an additional fundamentally new conflict. On the one hand, an RL agent needs to explore in order to discover good policies through trial and error. On the other hand, when carried out naively, this exploration will add noise to the policy of the agent during the training process, making their actions strictly less informative to their team mates. + +One possible solution to this is to explore in the space of deterministic partial policies, rather than actions, and sample these policies from a distribution that conditions on a common knowledge Bayesian belief. This is successfully carried out in the Bayesian Action Decoder (BAD) (Foerster et al., 2019), the only previous Deep RL method to achieve a state-of-the-art in Hanabi. While this is a notable accomplishment, it comes at the cost of simplicity and generality. For a start, BAD requires an explicit common knowledge Bayesian belief to be tracked, which not only adds computational burden due to the required sampling steps, but also uses expert knowledge regarding the game dynamics. Furthermore, BAD, as presented, is trained using actor-critic methods which are sample inefficient and suffer from local optima. In order to get around this, BAD uses population based training, further increasing the number of samples required. Lastly, BAD’s explicit reliance on common knowledge limits the generality of the method. + +In this paper we propose the Simplified Action Decoder (SAD), a method that achieves a similar goal to BAD, but addresses all of the issues mentioned above. At the core of SAD is a different approach towards resolving the conflict between exploration and being interpretable, which, like BAD, relies on the centralized training with decentralized control (CT/DC) regime. Under CT/DC information can be exchanged freely amongst all agents during centralized training, as long as the final policies are compatible with decentralized execution. + +The key insight is that during training we do not have to chose between being informative, by taking greedy actions, and exploring, by taking random actions. To be informative, the greedy actions do not need to be executed by the environment, but only need to be observed by the team mates. Thus in SAD each agent takes two different actions at each time step: One greedy action, which is not presented to the environment but observed by the team mates at the next time step as an additional input, and the “standard” (exploratory) action that gets executed by the environment and is observed by the team mates as part of the environment dynamics. Importantly, during greedy execution the observed environment action can be used instead of centralized information for the additional input, since now the agent has stopped exploring. + +Furthermore, to ensure that these greedy actions and observations get decoded into a meaningful representation, we can optionally train an auxiliary task that predicts key hidden game properties from the action-observation trajectories. While we note that this idea is in principle compatible with any kind of model-free deep RL method with minimal modifications to the core algorithm, we use a distributed version of recurrent DQN in order to improve sample efficiency, account for partial observability and reduce the risk of local optima. We also train a joint-action Q-function that consists of the sum of per-agent Q-values to allow for off-policy learning in this multi-agent setting using Value Decomposition Networks (VDN) (Sunehag et al., 2017). + +Using SAD we establish a new SOTA for learning methods for 2-5 players in Hanabi, with a method that not only requires less expert knowledge and compute, but is also more general than previous approaches. In order to ensure that our results can be easily verified and extended, we also evaluate our method on a proof-of-principle matrix game and open-source our training code and agents. Beyond enabling more research into the self-play aspect of Hanabi, we believe these resources will provide a much needed starting point for the ad-hoc teamwork part of the Hanabi challenge. + +# 2 RELATED WORK + +Our work relates closely to research on emergent communication protocols using deep multi-agent RL, as first undertaken by Sukhbaatar et al. (2016) and Foerster et al. (2016) . There has been a large number of follow-up papers in this area, so listing all relevant work is beyond the scope and we refer the reader to Nguyen et al. (2018), a recent survey on deep multi-agent RL. One major difference to our work is that the environments considered typically contain a cheap-talk channel, which can be modeled as a continuous variable during the course of training. This allows agents to, for example, use differentiation across the communication channel in order to learn protocols. In contrast, in our setting agents have to communicate through the observable environment actions themselves, requiring fundamentally different methods. + +Furthermore, our work is an example of cooperative multi-agent learning in partially observable settings under centralized training and decentralized control. There have been a large number of papers in this space, with seminal work including MADDPG (Lowe et al., 2017) and COMA (Foerster et al., 2018a), both of which are actor-critic methods that employ a centralized critic with decentralized actors. Again, we refer the reader to Nguyen et al. (2018) for a more comprehensive survey. + +Until 2018, work on Hanabi had been focused on hand-coded methods and heuristics. Some relevant examples include SmartBot (O’Dwyer, 2019) and the so-called “hat-coding” strategies, as implemented by WTFWThat (Wu, 2018). These strategies use the information theoretic ideas that allow each hint to reveal information to all other agents at the same time. While they do not perform well for 2-player Hanabi due to the smaller action space, they get near perfect scores for 3-5 players. + +In contrast, so far learning methods have seen limited success on Hanabi. Bard et al. (2019) undertake a systematic evaluation of current Deep RL methods for 2-5 players in two different regimes and open-source the Hanabi-Learning-Environment (HLE) to foster research on the game. They evaluate a feed-forward version of DQN trained on 100 million samples and a recurrent actor-critic agent with population based training using 20 billion samples. Notably, while both agents achieve near $0 \%$ win rate for 3-5 players in Hanabim, at a high level their DQN agent is a good starting point for our work. However, since the authors did not propose any specific method of accounting for the issues introduced by $\epsilon$ -greedy exploration in a ToM task, they resorted to setting $\epsilon$ to zero after a short burn-in phase. The only state-of-the-art in Hanabi established by an RL agent is from Foerster et al. (2019) which we refer to in more detail in Section 1 and Section 4. Recently there have also been attempts to train agents that are robust to different team-mates (Canaan et al., 2019) and even to extend to human-AI collaboration (Liang et al., 2019). For a more comprehensive review on previous results on Hanabi we refer the reader to Bard et al. (2019). + +Poker is another partially observable multi-agent setting, although it is fundamentally different due to the game being zero-sum. Recent success in Poker has extensively benefited from search (Brown et al.). Examples of using search in Hanabi include Goodman (2019). + +# 3 BACKGROUND + +# 3.1 SETTING + +In this paper we assume a Dec-POMDP (Oliehoek, 2012), in which $N$ agents interact in a partially observable environment. At each time step agent $a \in 1 . . N$ obtains an observation, $o _ { t } ^ { a } = { \bar { O } } ( s _ { t } , a )$ , where $s _ { t } \in S$ is the Markov state of the system and $O ( s _ { t } , a )$ is the deterministic observation function. Since we are interested in ToM, in our setting the observation function includes the last action of the acting agent, which is observed by all other agents at the next time step. We note that actions are commonly observable not only in board games but also in some real world multi-agent settings, such as autonomous driving. + +For simplicity, we restrict ourselves to turn based settings, in which at each time step only the acting agents takes an action, $u _ { t } ^ { a }$ , which is sampled from their policy, $u ^ { a } \sim \pi _ { \theta } ^ { a } ( u ^ { a } | \tau ^ { a } )$ , while all other agents take a no-op action. Here $\tau ^ { a }$ is the action-observation history of agent $a$ , $\tau ^ { a } = \{ o _ { 0 } ^ { a } , u _ { 0 } ^ { a } , r _ { 1 } , . . r _ { T } , o _ { T } ^ { a } \}$ , $T$ is the length of the episode and $\theta$ are the weights of a function approximator that represents the policy, in our case recurrent neural networks, such as LSTMs (Hochreiter & Schmidhuber, 1997). + +We further use $\tau _ { t }$ to describe the state-action sequence, $\tau = \{ s _ { 0 } , \mathbf { u _ { 0 } } , r _ { 1 } , . . r _ { T } , s _ { T } \}$ , where $\mathbf { u _ { t } }$ is the joint action of all agents. + +As is typical in cooperative multi-agent RL, the goal of the agents is to maximize the total expected return, $J _ { \theta } = \mathbb { E } _ { \tau \sim P ( \tau \mid \theta ) } R _ { 0 } ( \tau )$ , where $R _ { 0 } ( \tau )$ is the return of the trajectory (in general $R _ { t } \bar { ( } \tau ) ~ =$ $\sum _ { t ^ { \prime } \geq t } \gamma ^ { t ^ { \prime } - t } r _ { t ^ { \prime } } )$ and $\gamma$ is an optional discount factor. We have also assumed that agents are sharing parameters, $\theta$ , as is common in cooperative MARL. + +# 3.2 DISTRIBUTED RECURRENT DQN AND AUXILIARY TASKS + +In Q-learning the agent approximates the expected return for a given state action-pair, $s , u$ , assuming that the agent acts greedily with respect to the Q-function for all future time steps, $Q ( s , u ) =$ $\mathbb { E } _ { \tau \sim P ( \tau | s , u ) } R _ { t } ( \tau )$ , where $\tau = \{ s _ { t } , u _ { t } , r _ { t + 1 } , . . . , s _ { T } \}$ , $u _ { t } = u$ and $u _ { t ^ { \prime } } = \arg \operatorname* { m a x } _ { u ^ { \prime } } Q ( s _ { t ^ { \prime } } , u ^ { \prime } ) , \forall t ^ { \prime } >$ $t$ . A common exploration scheme is $\epsilon$ -greedy, in which the agent takes a random action with probability $\epsilon$ and acts greedily otherwise. Importantly, the Q-function can be trained efficiently using the Bellman equation: $Q ( s , u ) = \mathbb { E } _ { s ^ { \prime } } [ r _ { t + 1 } + \gamma \operatorname* { m a x } _ { u ^ { \prime } } Q ( s ^ { \prime } , u ^ { \prime } ) ]$ , where for simplicity we have assumed a deterministic reward. In Deep Q-Learning (DQN) (Mnih et al., 2015) the Q-function is parameterized by a deep neural network and trained with transitions sampled from experience replay. + +In our work we also incorporate other best practice components of the last few years, including doubleDQN (van Hasselt et al., 2015), dueling network architecture (Wang et al., 2015) and prioritized replay (Schaul et al., 2015). We also employ a distributed training architecture similar to the one proposed by Horgan et al. (2018) where a number of different actors with their own exploration rates collect experiences in parallel and feed them into a central replay buffer. Since our setting is partially observable the natural choice for the function approximator is a recurrent neural network. A combination of these techniques was first explored by Kapturowski et al. (2019) in single agent environments such as Atari and DMLab-30. + +Another common best-practice in RL are auxiliary tasks Mirowski et al. (2016); Jaderberg et al. (2016), in which the agent produces extra output-heads that are trained on supervised tasks and optimized alongside the RL loss. + +3.3 CENTRALISED TRAINING, DECENTRALIZED EXECUTION AND JOINT Q-FUNCTIONS + +The most straight forward application of Q-learning to multi-agent settings is Independent Q-Learning (IQL) (Tan, 1993) in which each agent keeps an independent estimate of the expected return, treating all other agents as part of the environment. One challenge with IQL is that the exploratory behavior of other agents is not corrected for via the max operator in the bootstrap. Notably, IQL does typically not take any advantage of centralized training with decentralised control (CT/DC), a paradigm under which information can be exchanged freely amongst agents during the training phase as long as the policies rely only on local observations during execution. + +There are various approaches for learning joint-Q-functions in the CT/DC regime. For example, Value-Decomposition-Networks (VDN) (Sunehag et al., 2017) represent the joint-Q-function as a sum of per-agent contributions and QMIX (Rashid et al., 2018) learns a non-linear but monotonic combination of these contributions. + +# 4 METHOD + +# 4.1 THEORY OF MIND AND BAYESIAN REASONING + +At the very core of interpreting the actions of another agent, and ToM in general, is Bayesian reasoning. Fundamentally, asking what a given action by another agent implies about the state of the world requires understanding of why this action was taken. To illustrate this, we start out with an agent that has a given belief about the state-action history of the world, $\tau _ { t }$ , given her own action-observation history $\tau _ { t } ^ { a }$ : $B ( \tau _ { t } ) = P ( \tau _ { t } | \tau _ { t } ^ { a } )$ . + +Next the agent observes the action $u _ { t } ^ { a ^ { \prime } }$ of her team mate, $a ^ { \prime }$ , and carries out a Bayesian update: + +$$ +\begin{array} { r l } & { P ( \tau _ { t } | \tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \prime } } ) = \frac { P ( u _ { t } ^ { a ^ { \prime } } | \tau _ { t } ) P ( \tau _ { t } | \tau _ { t } ^ { a } ) } { \sum _ { \tau _ { t } ^ { \prime } } P ( u _ { t } ^ { a ^ { \prime } } | \tau _ { t } ^ { \prime } ) P ( \tau _ { t } ^ { \prime } | \tau _ { t } ^ { a } ) } } \\ & { \qquad = \frac { \pi ^ { a ^ { \prime } } ( u _ { t } ^ { a ^ { \prime } } | O ( a ^ { \prime } , \tau _ { t } ) ) B ( \tau _ { t } ) } { \sum _ { \tau _ { t } ^ { \prime } } \pi ^ { a ^ { \prime } } ( u _ { t } ^ { a ^ { \prime } } | O ( a ^ { \prime } , \tau _ { t } ^ { \prime } ) ) B ( \tau _ { t } ^ { \prime } ) } , } \end{array} +$$ + +where, with a slight abuse of notation, we have used (and will keep using) $O ( a ^ { \prime } , \tau _ { t } )$ for the actionobservation history, $\tau _ { t } ^ { a ^ { \prime } }$ , that results from applying the observation function for agent $a ^ { \prime }$ to $\tau _ { t }$ at each time step. Note that for non-deterministic observation functions we would have to marginalize over $P ( \tau _ { t } ^ { a ^ { \prime } } | \bar { \tau _ { t } } )$ . + +Clearly, since agents have access to the policy of their teammate during centralised training, we could in principle evaluate this explicit Bayesian belief. However, beyond the practical difficulty of computing this explicit belief, when it is used as an input to the policy it will lead to prohibitively costly higher order beliefs. The typical workout for this is a public belief over private features which only conditions on common knowledge and can therefore be calculated by all agents individually, we refer to Moravcík et al. (2017); Nayyar et al. (2013); Foerster et al. (2018b) for more details. ˇ + +Instead, in this work we rely on RNNs to learn implicit representations of the sufficient statistics over the distribution of the Markov state given the action-observation histories, noting that they are unlikely to recover exact beliefs due to the issues mentioned above. + +# 4.2 EXPLORATION AND BELIEFS + +Next we illustrate the impact of exploration on the beliefs, which we will do in the explicit (exact) case, since it serves as an upper bound on the accuracy of the implicit beliefs. Since we are looking at fully-cooperative settings we assume that the optimal policy of the agent is deterministic and any randomness is due to exploration. Given that we are focused on value based methods we furthermore assume an $\epsilon$ -greedy exploration scheme, noting that the same analysis can be extended to other methods. Under this exploration scheme $\pi ^ { a ^ { \prime } } ( u _ { t } ^ { a ^ { \prime } } | O ( a ^ { \prime } , \tau _ { t } ) )$ ) becomes: + +$$ +\pi ^ { a ^ { \prime } } ( u _ { t } ^ { a ^ { \prime } } | O ( a ^ { \prime } , \tau _ { t } ) ) = ( 1 - \epsilon ) \mathbf { I } ( u ^ { * } ( \tau _ { t } ) , u _ { t } ^ { a ^ { \prime } } ) + \epsilon / | U | , +$$ + +where we have used $\boldsymbol { u } ^ { * } ( \tau _ { t } )$ to indicate the greedy action of the agent $a ^ { \prime }$ , $\begin{array} { r l } { u ^ { * } ( \tau _ { t } ) } & { { } = } \end{array}$ arg $\operatorname* { m a x } _ { u } Q ^ { a ^ { \prime } } ( u , O ( a ^ { \prime } , \tau _ { t } ) )$ and $\mathbf { I }$ is the indicator function. + +While the first part corresponds to a filtering operator, in which the indicator function only attributes finite probability to those histories that are consistent with the action taken under greedy execution, the exploration term adds a fixed (history independent) probability, which effectively ‘blurs’ the posterior: + +$$ +\begin{array} { r l r } { { P ( \tau _ { t } | \tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \prime } } ) = \frac { ( ( 1 - \epsilon ) \mathbf { I } ( u ^ { * } ( \tau _ { t } ) , u _ { t } ^ { a ^ { \prime } } ) + \epsilon / | U | ) B ( \tau _ { t } ) } { \sum _ { \tau ^ { \prime } } ( ( 1 - \epsilon ) \mathbf { I } ( u ^ { * } ( \tau ^ { \prime } ) , u _ { t } ^ { a ^ { \prime } } ) + \epsilon / | U | ) B ( \tau ^ { \prime } ) } } } \\ & { } & { = \frac { ( ( 1 - \epsilon ) \mathbf { I } ( u ^ { * } ( \tau _ { t } ) , u _ { t } ^ { a ^ { \prime } } ) + \epsilon / | U | ) B ( \tau _ { t } ) } { \epsilon / | U | + \sum _ { \tau ^ { \prime } } ( ( 1 - \epsilon ) \mathbf { I } ( u ^ { * } ( \tau ^ { \prime } ) , u _ { t } ^ { a ^ { \prime } } ) ) B ( \tau ^ { \prime } ) } } \\ & { } & { = \frac { B ( \tau _ { t } ) } { 1 + | U | \sum _ { s ^ { \prime } } ( ( 1 / \epsilon - 1 ) \mathbf { I } ( u ^ { * } ( \tau ^ { \prime } ) , u _ { t } ^ { a ^ { \prime } } ) B ( \tau ^ { \prime } ) } } \\ & { } & { + \frac { ( ( 1 - \epsilon ) \mathbf { I } ( u ^ { * } ( \tau _ { t } ) , u _ { t } ^ { a ^ { \prime } } ) ) B ( \tau _ { t } ) } { \epsilon / | U | + \sum _ { \tau ^ { \prime } } ( ( 1 - \epsilon ) \mathbf { I } ( u ^ { * } ( \tau ^ { \prime } ) , u _ { t } ^ { a ^ { \prime } } ) ) B ( \tau ^ { \prime } ) } . } \end{array} +$$ + +We find that the posterior includes an additional term of the form $B ( \tau _ { t } )$ which carries over an unfiltered density over the trajectories from the prior. We further confirm that in the limit of $\epsilon = 1$ , the posterior collapses to the prior, $P ( \tau _ { t } | \tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \prime } } ) = B ( \tau _ { t } )$ . This can be particularly worrisome in the context of our training setup, whereby different agents run different, and potentially high,  throughout the course of training. It fundamentally makes the beliefs obtained less informative. + +While not making the above argument explicitly, the Bayesian Action Decoder (BAD) (Foerster et al., 2019), resolves this issue by shifting exploration to the level of deterministic partial policies, rather than action-level, and tracking an approximate Bayesian belief. As outlined in Section 1 this comes at a huge cost in the complexity of the method, the computation requirements and in the loss of generality of the method. + +# 4.3 SIMPLIFIED ACTION DECODING + +In this paper we take a drastically simpler and different approach towards the issue. We note that the ‘blurring’, which makes decoding of an action challenging, is entirely due to the $\epsilon$ -greedy exploration term. Furthermore, in order for another agent to do an implicit Bayesian update over an action taken, it is not required that this action is executed by the environment. Indeed, if we assume that other agents can observe the greedy action, $u ^ { * }$ , at every time step and condition their belief update on this, the terms depending on $\epsilon$ disappear from the Bayesian update: + +$$ +P ( \tau _ { t } | \tau _ { t } ^ { a } , u ^ { * } ) = \frac { \mathbf { I } ( u ^ { * } ( \tau _ { t } ) , u ^ { * } ) \big ) B ( \tau _ { t } ) } { \sum _ { \tau ^ { \prime } } \mathbf { I } ( u ^ { * } ( \tau ^ { \prime } ) , u ^ { * } ) \big ) B ( \tau ^ { \prime } ) } +$$ + +Therefore, to have our cake and eat it, in the Simplified Action Decoder (SAD) the acting agent is allowed to ‘take’ two actions at any given time step during training. The first action, $u ^ { a }$ , is the standard environment action, which gets executed as usual and is observed by all agents through the observation function at the next time step, as mentioned in Section 3. The second action, $u ^ { * }$ , is the greedy action of the active agent. This action does not get executed by the environment but instead is presented as an additional input to the other agents at the next time step, taking advantage of the centralized training regime during which information can be exchanged freely. + +Clearly we are not allowed to pass around extra information during decentralized control, but luckily this is not needed. Since we set $\epsilon$ to 0 at test time we can simply use the, now greedy, environment action obtained from the observation function as our greedy-action input. + +While this is most straight forward in settings where the last action is observed by other agents directly, in principle SAD can also be extended to settings where it is indirectly observed by all agents through the environment dynamics. In these cases we can replace the greedy-action side-channel with a learned inverse model that recovers the action from the observation history during execution. + +Furthermore, to encourage the agent to meaningfully decode the information contained in the greedy action, we can optionally add an auxiliary task to the training process, such as predicting unobserved information from their observation history . + +While this idea is compatible with any deep RL algorithm with minimal modifications, we use a recurrent version of DQN with distributed training, dueling networks and prioritized replay. We also learn a joint Q-function using VDN in order to address the challenges of multi-agent off-policy learning, please see Section 3 for details on all of these standard methods. + +# 5 EXPERIMENTS + +# 5.1 MATRIX GAME + +We first verify the effectiveness of SAD in the two step, two player matrix game from Foerster et al. (2019), which replicates the communication through action challenge of Hanabi in a highly simplified setting. In this fully cooperative game each player obtains a privately observed ‘card’, which is drawn iid from two options (1,2). + +After observing her card, the first player takes one of three possible discrete actions (1, 2, 3). Crucially, the second player observes both her own private card and the team mate’s action before acting herself, which establishes the opportunity to communicate. The payout is a function of both the two private cards and the two actions taken by both agents, as shown in Figure 1. + +Importantly, there are some obvious strategies that do not “Tiny Hanabi”require any communication. For example, if both player learn to play the 2nd action, the payout is always 8 points, independent of the cards dealt. However, if the players do ● Each player knows their learn to communicate it is possible to achieve 10 points for every pair of cards dealt. + +# 5.2 HANABI + +● Player 2 observes Player Hanabi is a fully cooperative card game in which all play1’s action and acts ers work together to complete piles of cards referred to as second.fireworks. Each card has a rank, 1 to 5, and a color, G / B $/ \textbf { W } / \textbf { Y } / \textbf { R }$ . Each firework (one per color) starts with a 1 and is finished once the 5 has been added. There are three 1s, one 5 and two of all other ranks for each of the colors, adding up to a total of 50 cards in the deck. The twist in Hanabi is that while players can observe the cards held by their team mates, they cannot observe their own cards and thus need to exchange information with each other in order to understand what cards can be played. There are two main means for doing so: First of all, players can take grounded hint actions, in which they reveal the subset of a team mate’s hand that matches a specific rank or color. An example hint is “Your third and fifth card are 1s”. These hint actions cost scarce information tokens, which can be replenished by discarding a card, an action that both removes the card from the game and makes it visible to all players. + +![](images/25f404762b879bc27cf6c88933eae05b823c51b2b16509fef1e770d1b412b811.jpg) +Figure 1: Illustration of the matrix game from Foerster et al. (2019) + +Finally players can also choose to play a card. If this card is the next card for the firework of the corresponding color, it is added to the firework and the team scores one point. Otherwise the card is removed from the game, the identity is made public, and the team loses one of the 3 life tokens. If the team runs out of life tokens before the end of the game, all points collected so far are lost and the game finishes immediately. These rules result in a maximum score of $5 \times 5 = 2 5$ points in any game, which corresponds to all five fireworks being completed with five cards per firework. + +To ensure reproducibility and comparability of our results we use the Hanabi Learning Environment (HLE) (Bard et al., 2019) for all experimentation. For further details regarding Hanabi and the self-play part of the Hanabi challenge please see Bard et al. (2019). + +# 5.3 ARCHITECTURE AND COMPUTATION REQUIREMENTS + +We borrow some ideas and insights from prior distributed Q-learning methods while bring extensions to MARL as well as innovations to improve throughput and efficiency. Following Horgan et al. (2018) and Kapturowski et al. (2019), we use a distributed prioritized replay buffer shared by $N$ asynchronous actors and a centralized trainer that samples mini-batches from the replay buffer to update the model. In each actor thread, we run $K$ environments sequentially and batch their observations together. The observation batch is then fed into an actor that utilize a GPU to compute a batch of actions. All asynchronous actors share one GPU and the trainer uses another GPU for gradient computation and model updates. This is different from prior works which run single actor and single environment in each thread on a CPU. Our method enables us to run a very large number of simulations with moderate computation resources. In all Hanabi experiments, we run $N = 8 0$ actor threads with $K = 8 0$ environments in each thread on single machine with 40 CPU cores and 2 GPUs. Without this architectural improvement, it may require at least a few hundred CPU cores to run 6400 Hanabi environments, in which case neural network agents and simulations have to be distributed across multiple machines, greatly reducing the reproducibility and accessibility of such research. Please refer to Appendix A for implementation details and hyper-parameters. + +# 6 RESULTS + +# 6.1 MATRIX GAME + +As we can see in Figure 2, even in our simple matrix game the greedy action input makes a drastic difference. With an average reward of around 9.5 points, tabular IQL does well in this task, matching the $B A D$ results from Foerster et al. (2019). However, just by adding the greedy action as an additional input, we obtain an average performance of $9 . 9 7 \pm 0 . 0 2$ . Results are averaged over 100 seeds, and shading is s.e.m. The code is available here: www.bit.ly/2mBJLyk. + +![](images/5ad1d82f24e9ee7d32e4c518b21671940bf756b1552a474256a05432182b3238.jpg) +Figure 2: Results for the matrix game. + +# 6.2 HANABI + +As shown in Table 1, our findings from the matrix game are for the most part confirmed on the challenging Hanabi benchmark. To illustrate the contributions of the different components, we compare average scores and win rates across 13 independent training runs of SAD and three different options: IQL is simply the recurrent DQN agent with parameter sharing, VDN is the same agent but also learns a joint Q-function and finally SAD & AuxTask is the SAD agent with the auxiliary task. + +While we find that SAD significantly outperforms our baselines (IQL and VDN) for 2, 4 and 5 players in terms of average score and/or win rate, there is no significant difference for 3 players, where VDN matches the performance of SAD. + +Interestingly, the auxiliary task only significantly helps the 2-player performance, where it substantially boosts the average score and win rate. In contrast, it drastically hurts performance for 3-5 players, which opens an interesting avenue for future work. + +For completeness we have included training curves showing average scores and s.e.m. across all training runs for all numbers of players for our methods and ablations in Appendix B. We find that for 5 players the auxiliary task drastically reduces the variance of SAD and intermittently leads to higher performance during training but ultimately results in lower final performance. We can also clearly see that despite 72 hours of training and billions of samples consumed, the performance has not plateaued for 3-5 players, pointing to an obvious avenue for further improvements. + +The original numbers in the Hanabi challenge and BAD used population based training (Jaderberg et al., 2018), effectively reporting maximum performance across a large number of different runs. Therefore, for reproducibility purposes, we report evaluations of the best model from our various training runs for each method in Table 2. + +As shown, under this reporting we establish a new SOTA for learning methods on the self-play part of the Hanabi challenge for 2-5 players, with the most drastic improvements being achieved for 3-5 players. In particular, we beat both the ACHA agent from Bard et al. (2019) and the BAD agent on average score, even though both of them used population based training and require more compute. We note that while we follow the counting convention proposed by the challenge paper, BAD was optimized for a different counting scheme, in which agents keep their scores when they run out of lives. This may explain the higher win rate $( 5 8 . 6 \% )$ of BAD combined with a relatively low mean score, which is exceeded even by our baseline methods. Once again, only the performance for 2-player is significantly improved by the auxiliary task and the 3-player setting is an outlier in the sense that SAD does not improve the best performance compared to VDN. + +# 7 CONCLUSION AND FUTURE WORK + +In this paper we presented the Simplified Action Decoder (SAD), a novel deep multi-agent RL algorithm that allows agents to learn communication protocols in settings where no cheap-talk channel is available. On the challenging benchmark Hanabi our work substantially improves the SOTA for an RL method for all numbers of players. For two players SAD establishes a new high-score across any method. Furthermore we accomplish all of this with a method that is both simpler and requires less compute than previous advances. While these are encouraging steps, there is clearly more work to do. In particular, there remains a large performance gap between the numbers achieved by SAD and the known performance of hat-coding strategies (Wu, 2018) for 3-5 players. One possible reason is that SAD does not undertake any explicit exploration in the space of possible conventions. Another promising route for future work is to integrate search with RL, since this has produced SOTA results in a number of different domains including Poker, Go and backgammon. + +Table 1: Mean performance of our methods and baselines on Hanabi. We take the final models of 13 independent runs, i.e. 13 models per algorithm per player setting. Each model is evaluated on 100K games. Mean and s.e.m over the mean scores of the 13 models are shown in the table. The second row of each section is the win rate. + +
Agent 2 Players 3 Players 4 Players 5 Players
IQL (Baseline)23.77 ± 0.04 43.88 ± 1.21 %23.02 ± 0.10 26.16 ± 2.01 %21.99 ± 0.09 10.15 ± 0.86 %20.60 ± 0.11 2.27 ± 0.32 %
VDN23.83 ± 0.0323.71 ± 0.0623.03 ± 0.1521.18 ± 0.12
(Baseline) SAD44.97 ± 1.28 % 23.87 ± 0.0341.16 ± 1.27 % 23.69 ± 0.0523.57 ± 2.20 % 23.27 ± 0.162.26 ± 0.32 % 22.06 ± 0.23
SAD AuxTask47.90 ± 1.10 % 24.02 ± 0.0141.12 ± 1.10 % 23.56 ± 0.0729.38 ± 2.63 % 22.78 ± 0.107.22 ± 1.29 % 21.47 ± 0.08
+ +Table 2: Comparison between the previous SOTA learning methods and ours. We take the best model of 13 runs for each of our methods and baselines. Each model is evaluated on 100K games with different seeds. Mean and s.e.m over the 100K games are shown in the table. The s.e.m. is less than 0.01 for most models. Bold numbers are the best results achieved with learning algorithms. The second row of each section is the win rate. + +
Agent2 Players 3 Players 4 Players 5 Players
Rainbow (Bard et al., 2019)20.64 ± 0.03 2.5%18.71 ±0.01 0.2%18.00 ± 0.17 0%15.26 ± 0.18 0%
ACHA (Bard et al., 2019)22.73 ± 0.12 15.1%20.24 ± 0.15 1.1%21.57 ± 0.12 2.4%16.80 ± 0.13 0%
BAD (Foerster et al.,2019)23.92 ± 0.01 58.56%==1
IQL (Baseline) 50.47%23.97 ± 0.01 40.25%23.69 ± 0.01 19.39%22.76 ± 0.01 4.93%21.29 ± 0.01
VDN (Baseline)23.96 ± 0.01 50.27%23.99 ± 0.01 50.37%23.79 ± 0.00 38.86%21.80 ± 0.01 4.98%
SAD24.01 ± 0.01 52.39%23.93 ± 0.01 48.05%23.81 ± 0.01 41.45%23.01 ± 0.01 13.93%
SAD& AuxTask24.08 ± 0.01 56.09%23.81 ± 0.01 49.74%23.47 ± 0.01 33.87%22.25 ± 0.01 7.33%
+ +# REFERENCES + +Chris L Baker, Julian Jara-Ettinger, Rebecca Saxe, and Joshua B Tenenbaum. Rational quantitative attribution of beliefs, desires and percepts in human mentalizing. Nature Human Behaviour, 1(4): 0064, 2017. + +Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad, Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The Hanabi Challenge: A New Frontier for AI Research. arXiv:1902.00506 [cs, stat], February 2019. URL http://arxiv. org/abs/1902.00506. arXiv: 1902.00506. + +Noam Brown and Tuomas Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top professionals. Science, pp. eaao1733, 2017. + +Noam Brown and Tuomas Sandholm. Superhuman AI for multiplayer poker. Science, pp. eaay2400, 2019. + +Noam Brown, Tuomas Sandholm, and Brandon Amos. Depth-Limited Solving for ImperfectInformation Games. pp. 14. + +Murray Campbell, A Joseph Hoane Jr, and Feng-hsiung Hsu. Deep Blue. Artificial intelligence, 134 (1-2):57–83, 2002. + +Rodrigo Canaan, Julian Togelius, Andy Nealen, and Stefan Menzel. Diverse agents for ad-hoc cooperation in hanabi. arXiv preprint arXiv:1907.03840, 2019. + +Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems, pp. 2137–2145, 2016. + +Jakob Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson, Matthew Botvinick, and Michael Bowling. Bayesian action decoder for deep multi-agent reinforcement learning. In International Conference on Machine Learning, pp. 1942–1951, 2019. + +Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018a. + +Jakob N. Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson, Matthew Botvinick, and Michael Bowling. Bayesian Action Decoder for Deep Multi-Agent Reinforcement Learning. arXiv:1811.01458 [cs], November 2018b. URL http://arxiv. org/abs/1811.01458. arXiv: 1811.01458. + +James Goodman. Re-determinizing information set monte carlo tree search in hanabi. arXiv preprint arXiv:1902.06075, 2019. + +Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735– 1780, November 1997. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL http://dx. doi.org/10.1162/neco.1997.9.8.1735. + +Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt, and David Silver. Distributed prioritized experience replay. CoRR, abs/1803.00933, 2018. URL http://arxiv.org/abs/1803.00933. + +Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016. + +Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castaneda, Charles Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel. Human-level performance in first-person multiplayer games with population-based deep reinforcement learning. arXiv:1807.01281 [cs, stat], July 2018. doi: 10.1126/science.aau6249. URL http://arxiv.org/abs/1807.01281. arXiv: 1807.01281. + +Steven Kapturowski, Georg Ostrovski, John Quan, and Will Dabney. RECURRENT EXPERIENCE REPLAY IN DISTRIBUTED REINFORCEMENT LEARNING. pp. 1–19, 2019. + +Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. + +Claire Liang, Julia Proft, Erik Andersen, and Ross A Knepper. Implicit communication of actionable information in human-ai teams. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pp. 95. ACM, 2019. + +Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems, pp. 6379–6390, 2017. + +Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in complex environments. arXiv preprint arXiv:1611.03673, 2016. + +Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 02 2015. URL http://dx.doi.org/10.1038/nature14236. + +Matej Moravcík, Martin Schmid, Neil Burch, Viliam Lis ˇ y, Dustin Morrill, Nolan Bard, Trevor \` Davis, Kevin Waugh, Michael Johanson, and Michael Bowling. Deepstack: Expert-level artificial intelligence in heads-up no-limit poker. Science, 356(6337):508–513, 2017. + +Ashutosh Nayyar, Aditya Mahajan, and Demosthenis Teneketzis. Decentralized stochastic control with partial history sharing: A common information approach. IEEE Transactions on Automatic Control, 58(7):1644–1658, 2013. + +Thanh Thi Nguyen, Ngoc Duy Nguyen, and Saeid Nahavandi. Deep reinforcement learning for multiagent systems: a review of challenges, solutions and applications. arXiv preprint arXiv:1812.11794, 2018. + +Arthur O’Dwyer. Hanabi. https://github.com/Quuxplusone/Hanabi, 2019. + +Frans A Oliehoek. Decentralized pomdps. In Reinforcement Learning, pp. 471–503. Springer, 2012. + +Tabish Rashid, Mikayel Samvelyan, Christian Schröder de Witt, Gregory Farquhar, Jakob N. Foerster, and Shimon Whiteson. QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning. CoRR, abs/1803.11485, 2018. URL http://arxiv.org/abs/ 1803.11485. + +Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized Experience Replay. arXiv e-prints, art. arXiv:1511.05952, Nov 2015. + +David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. + +David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. Nature, 550(7676):354, 2017. + +David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419): 1140–1144, 2018. + +Sainbayar Sukhbaatar, arthur szlam, and Rob Fergus. Learning Multiagent Communication with Backpropagation. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp. 2244–2252. Curran Associates, Inc., 2016. URL http://papers.nips.cc/paper/ 6398-learning-multiagent-communication-with-backpropagation.pdf. + +Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinícius Flores Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z. Leibo, Karl Tuyls, and Thore Graepel. Value-decomposition networks for cooperative multi-agent learning. CoRR, abs/1706.05296, 2017. URL http://arxiv.org/abs/1706.05296. + +Richard S. Sutton. Learning to predict by the methods of temporal differences. Mach. Learn., 3(1):9–44, August 1988. ISSN 0885-6125. doi: 10.1023/A:1022633531479. URL https: //doi.org/10.1023/A:1022633531479. + +Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pp. 330–337, 1993. + +Hado van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double qlearning. CoRR, abs/1509.06461, 2015. URL http://arxiv.org/abs/1509.06461. + +Ziyu Wang, Nando de Freitas, and Marc Lanctot. Dueling network architectures for deep reinforcement learning. CoRR, abs/1511.06581, 2015. URL http://arxiv.org/abs/1511.06581. + +Jeff Wu. Hanabi simulation in rust. https://github.com/WuTheFWasThat/hanabi.rs, 2018. + +# A NETWORK ARCHITECTURE AND HYPER-PAMAMETERS FOR HANABI + +Our Hanabi agent uses dueling network architecture (Wang et al., 2015). The main body of the network consists of 1 fully connected layer of 512 units and 2 LSTM (Hochreiter & Schmidhuber, 1997) layers of 512 units, followed by two output heads for value and advantages respectively. The same network configuration is used across all Hanabi experiments. We take the default featurization of HLE and replace the card knowledge section with the V0-Belief proposed by Foerster et al. (2019). The maximum length of an episode is capped at 80 steps and the entire episode is stored in the replay buffer as one training sample. This avoids the “slate hidden states” problem as described in Kapturowski et al. (2019) because we can simply initialize the hidden states of LSTM as zero during training. For exploration and experience prioritization, we follow the simple strategy as in Horgan et al. (2018) and Kapturowski et al. (2019). Each actor executes an $\epsilon _ { i }$ -greedy policy where $\epsilon _ { i } = \bar { \epsilon } ^ { 1 + \frac { 1 } { N - 1 } \alpha }$ for $i \in \{ 0 , . . . , N - 1 \}$ but with a smaller $\epsilon = 0 . 1$ and $\alpha = 7$ . For simplicity, all players of a game use the same epsilon. The per time-step priority $\delta _ { t }$ is the TD error and per episode priority is computed following $\delta _ { e } = \eta \operatorname* { m a x } _ { t } \delta _ { i } + ( 1 - \eta ) \hat { \delta }$ where $\eta = 0 . 9$ . Priority exponent is set to 0.9 and importance sampling exponent is set to 0.6. We use $n$ -step return (Sutton, 1988) and double Q-learning (van Hasselt et al., 2015) for target computation during training. The discount factor $\gamma$ is set to 0.999. The network is updated using Adam optimizer (Kingma & Ba, 2014) with learning rate $l r = 6 . 2 5 \times 1 0 ^ { - 5 }$ and $\epsilon = \dot { 1 . 5 } \times 1 0 ^ { - 5 }$ . Trainer sends its network weights to all actors every 10 updates and target network is synchronized with online network every 2500 updates. These hyper-parameters are fixed across all experiments. + +In the baseline, we use Independent Q-Learning where each player estimates the Q value and selects action independently at each time-step. Note that all players need to operate on the observations in order to update their recurrent hidden states while only the current player has non-trivial legal moves and other players can only select ‘pass’. Each player then writes its own version of the episode into the prioritized replay buffer and they are sampled independently during training. The prioritized replay buffer contains $2 ^ { 1 7 } ( 1 3 1 0 7 2 )$ episodes. We warm up the replay buffer with 10,000 episodes before training starts. Batch size during training is 128 for games of different numbers of players. + +As mentioned in Section 4, the SAD agent is built on top of joint Q-function where the Q value is the sum of the individual Q value of all players given their own actions. One episode produces only one training sample with an extra dimension for the number of players. The replay buffer size is reduced to $2 ^ { 1 6 }$ for 2-player and 3-player games and $2 ^ { 1 5 }$ for 4-player and 5-player games. The batch sizes for 2-, 3-, 4-, 5-players are 64, 43, 32, 26 respectively to account for the fact that each sample contains more data. + +Auxiliary task can be added to the agent to help it decode the greedy action more effectively. In Hanabi, the natural choice is the predict the card of player’s own hand. In our experiments, the auxiliary task is to predict the status of a card, which can be playable, discardable, or unknown. The loss is the average cross entropy loss per card and is simply added to the TD-error of reinforcement learning during training. + +# B LEARNING CURVES FOR HANABI + +![](images/ccf9077cc9f7d726e6a7f0ea3a2fa7941948d5e7bedb2d3bea373709b6945a93.jpg) +Figure 3 shows learning curves of different algorithms averaged over 13 seeds per algorithm per player setting. Shading is error of the mean. +Figure 3: Learning Curves \ No newline at end of file diff --git a/parse/train/B1xm3RVtwB/B1xm3RVtwB_content_list.json b/parse/train/B1xm3RVtwB/B1xm3RVtwB_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44f3deba2475d21e5e2059b81ecaf5795f2b3f1b --- /dev/null +++ b/parse/train/B1xm3RVtwB/B1xm3RVtwB_content_list.json @@ -0,0 +1,1551 @@ +[ + { + "type": "text", + "text": "SIMPLIFIED ACTION DECODER FOR DEEP MULTI-AGENT REINFORCEMENT LEARNING ", + "text_level": 1, + "bbox": [ + 176, + 98, + 781, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hengyuan Hu, Jakob N Foerster Facebook AI Research, CA, USA {hengyuan,jnf}@fb.com ", + "bbox": [ + 184, + 169, + 410, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 250, + 544, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years we have seen fast progress on a number of benchmark problems in AI, with modern methods achieving near or super human performance in Go, Poker and Dota. One common aspect of all of these challenges is that they are by design adversarial or, technically speaking, zero-sum. In contrast to these settings, success in the real world commonly requires humans to collaborate and communicate with others, in settings that are, at least partially, cooperative. In the last year, the card game Hanabi has been established as a new benchmark environment for AI to fill this gap. In particular, Hanabi is interesting to humans since it is entirely focused on theory of mind, i.e., the ability to effectively reason over the intentions, beliefs and point of view of other agents when observing their actions. Learning to be informative when observed by others is an interesting challenge for Reinforcement Learning (RL): Fundamentally, RL requires agents to explore in order to discover good policies. However, when done naively, this randomness will inherently make their actions less informative to others during training. We present a new deep multi-agent RL method, the Simplified Action Decoder (SAD), which resolves this contradiction exploiting the centralized training phase. During training SAD allows other agents to not only observe the (exploratory) action chosen, but agents instead also observe the greedy action of their team mates. By combining this simple intuition with best practices for multi-agent learning, SAD establishes a new SOTA for learning methods for 2-5 players on the self-play part of the Hanabi challenge. Our ablations show the contributions of SAD compared with the best practice components. All of our code and trained agents are available at https://github.com/facebookresearch/Hanabi_SAD. ", + "bbox": [ + 232, + 280, + 764, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 627, + 336, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans are highly social creatures and spend vast amounts of time coordinating, collaborating and communicating with others. In contrast to these, at least partially, cooperative settings most progress on AI in games has been in zero-sum games where agents compete against each other, typically rendering communication futile. This includes examples such as Go (Silver et al., 2016; 2017; 2018), poker (Brown & Sandholm, 2017; Moravcík et al., 2017; Brown & Sandholm, 2019) ˇ and chess (Campbell et al., 2002). ", + "bbox": [ + 174, + 659, + 825, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This narrow focus is unfortunate, since communication and coordination require unique abilities. In order to enable smooth and efficient social interactions of groups of people, it is commonly required to reason over the intents, points of views and beliefs of other agents from observing their actions. For example, a driver can reasonably infer that if a truck in front of them is slowing down when approaching an intersection, then there is likely an obstacle ahead. Furthermore, humans are both able to interpret the actions of others and can act in a way that is informative when their actions are being observed by others, capabilities that are commonly called theory of Mind (ToM), (Baker et al., 2017). Importantly, in order to carry out this kind of reasoning, an agent needs to consider why a given action is taken and what this decision indicates about the state of the world. Simply observing what other agents are doing is not sufficient. ", + "bbox": [ + 174, + 750, + 825, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While these abilities are particularly relevant in partially observable, fully cooperative multi-agent settings, ToM reasoning clearly matters in a variety of real world scenarios. For example, autonomous cars will likely need to understand the point of view, intents and beliefs of other traffic participants in order to deal with highly interactive settings such as 4-way crossing or dense traffic in cities. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 821, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hanabi is a fully cooperative, partially-observable card game that has recently been proposed as a new benchmark challenge problem for AI research (Bard et al., 2019) to fill the gap around ToM. In Hanabi, players need to find conventions that allow them to effectively exchange information from their local observations through their actions, taking advantage of the fact that actions are observed by all team mates. ", + "bbox": [ + 174, + 138, + 823, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most prior state-of-the-art agents for Hanabi were developed using handcrafted algorithms, which beat off-the-shelf deep multi-agent RL methods by a large margin. This makes intuitive sense: Beyond the “standard” multi-agent challenges of credit assignment, nonstationarity and joint exploration, learning an informative policy presents an additional fundamentally new conflict. On the one hand, an RL agent needs to explore in order to discover good policies through trial and error. On the other hand, when carried out naively, this exploration will add noise to the policy of the agent during the training process, making their actions strictly less informative to their team mates. ", + "bbox": [ + 174, + 215, + 825, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One possible solution to this is to explore in the space of deterministic partial policies, rather than actions, and sample these policies from a distribution that conditions on a common knowledge Bayesian belief. This is successfully carried out in the Bayesian Action Decoder (BAD) (Foerster et al., 2019), the only previous Deep RL method to achieve a state-of-the-art in Hanabi. While this is a notable accomplishment, it comes at the cost of simplicity and generality. For a start, BAD requires an explicit common knowledge Bayesian belief to be tracked, which not only adds computational burden due to the required sampling steps, but also uses expert knowledge regarding the game dynamics. Furthermore, BAD, as presented, is trained using actor-critic methods which are sample inefficient and suffer from local optima. In order to get around this, BAD uses population based training, further increasing the number of samples required. Lastly, BAD’s explicit reliance on common knowledge limits the generality of the method. ", + "bbox": [ + 174, + 320, + 825, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper we propose the Simplified Action Decoder (SAD), a method that achieves a similar goal to BAD, but addresses all of the issues mentioned above. At the core of SAD is a different approach towards resolving the conflict between exploration and being interpretable, which, like BAD, relies on the centralized training with decentralized control (CT/DC) regime. Under CT/DC information can be exchanged freely amongst all agents during centralized training, as long as the final policies are compatible with decentralized execution. ", + "bbox": [ + 174, + 479, + 825, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key insight is that during training we do not have to chose between being informative, by taking greedy actions, and exploring, by taking random actions. To be informative, the greedy actions do not need to be executed by the environment, but only need to be observed by the team mates. Thus in SAD each agent takes two different actions at each time step: One greedy action, which is not presented to the environment but observed by the team mates at the next time step as an additional input, and the “standard” (exploratory) action that gets executed by the environment and is observed by the team mates as part of the environment dynamics. Importantly, during greedy execution the observed environment action can be used instead of centralized information for the additional input, since now the agent has stopped exploring. ", + "bbox": [ + 174, + 570, + 825, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, to ensure that these greedy actions and observations get decoded into a meaningful representation, we can optionally train an auxiliary task that predicts key hidden game properties from the action-observation trajectories. While we note that this idea is in principle compatible with any kind of model-free deep RL method with minimal modifications to the core algorithm, we use a distributed version of recurrent DQN in order to improve sample efficiency, account for partial observability and reduce the risk of local optima. We also train a joint-action Q-function that consists of the sum of per-agent Q-values to allow for off-policy learning in this multi-agent setting using Value Decomposition Networks (VDN) (Sunehag et al., 2017). ", + "bbox": [ + 174, + 702, + 825, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using SAD we establish a new SOTA for learning methods for 2-5 players in Hanabi, with a method that not only requires less expert knowledge and compute, but is also more general than previous approaches. In order to ensure that our results can be easily verified and extended, we also evaluate our method on a proof-of-principle matrix game and open-source our training code and agents. Beyond enabling more research into the self-play aspect of Hanabi, we believe these resources will provide a much needed starting point for the ad-hoc teamwork part of the Hanabi challenge. ", + "bbox": [ + 174, + 820, + 825, + 905 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 102, + 344, + 117 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work relates closely to research on emergent communication protocols using deep multi-agent RL, as first undertaken by Sukhbaatar et al. (2016) and Foerster et al. (2016) . There has been a large number of follow-up papers in this area, so listing all relevant work is beyond the scope and we refer the reader to Nguyen et al. (2018), a recent survey on deep multi-agent RL. One major difference to our work is that the environments considered typically contain a cheap-talk channel, which can be modeled as a continuous variable during the course of training. This allows agents to, for example, use differentiation across the communication channel in order to learn protocols. In contrast, in our setting agents have to communicate through the observable environment actions themselves, requiring fundamentally different methods. ", + "bbox": [ + 174, + 138, + 825, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, our work is an example of cooperative multi-agent learning in partially observable settings under centralized training and decentralized control. There have been a large number of papers in this space, with seminal work including MADDPG (Lowe et al., 2017) and COMA (Foerster et al., 2018a), both of which are actor-critic methods that employ a centralized critic with decentralized actors. Again, we refer the reader to Nguyen et al. (2018) for a more comprehensive survey. ", + "bbox": [ + 174, + 271, + 825, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Until 2018, work on Hanabi had been focused on hand-coded methods and heuristics. Some relevant examples include SmartBot (O’Dwyer, 2019) and the so-called “hat-coding” strategies, as implemented by WTFWThat (Wu, 2018). These strategies use the information theoretic ideas that allow each hint to reveal information to all other agents at the same time. While they do not perform well for 2-player Hanabi due to the smaller action space, they get near perfect scores for 3-5 players. ", + "bbox": [ + 174, + 347, + 825, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, so far learning methods have seen limited success on Hanabi. Bard et al. (2019) undertake a systematic evaluation of current Deep RL methods for 2-5 players in two different regimes and open-source the Hanabi-Learning-Environment (HLE) to foster research on the game. They evaluate a feed-forward version of DQN trained on 100 million samples and a recurrent actor-critic agent with population based training using 20 billion samples. Notably, while both agents achieve near $0 \\%$ win rate for 3-5 players in Hanabim, at a high level their DQN agent is a good starting point for our work. However, since the authors did not propose any specific method of accounting for the issues introduced by $\\epsilon$ -greedy exploration in a ToM task, they resorted to setting $\\epsilon$ to zero after a short burn-in phase. The only state-of-the-art in Hanabi established by an RL agent is from Foerster et al. (2019) which we refer to in more detail in Section 1 and Section 4. Recently there have also been attempts to train agents that are robust to different team-mates (Canaan et al., 2019) and even to extend to human-AI collaboration (Liang et al., 2019). For a more comprehensive review on previous results on Hanabi we refer the reader to Bard et al. (2019). ", + "bbox": [ + 174, + 424, + 825, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Poker is another partially observable multi-agent setting, although it is fundamentally different due to the game being zero-sum. Recent success in Poker has extensively benefited from search (Brown et al.). Examples of using search in Hanabi include Goodman (2019). ", + "bbox": [ + 174, + 612, + 825, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 BACKGROUND ", + "text_level": 1, + "bbox": [ + 176, + 683, + 326, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 SETTING ", + "text_level": 1, + "bbox": [ + 174, + 719, + 276, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper we assume a Dec-POMDP (Oliehoek, 2012), in which $N$ agents interact in a partially observable environment. At each time step agent $a \\in 1 . . N$ obtains an observation, $o _ { t } ^ { a } = { \\bar { O } } ( s _ { t } , a )$ , where $s _ { t } \\in S$ is the Markov state of the system and $O ( s _ { t } , a )$ is the deterministic observation function. Since we are interested in ToM, in our setting the observation function includes the last action of the acting agent, which is observed by all other agents at the next time step. We note that actions are commonly observable not only in board games but also in some real world multi-agent settings, such as autonomous driving. ", + "bbox": [ + 174, + 748, + 825, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For simplicity, we restrict ourselves to turn based settings, in which at each time step only the acting agents takes an action, $u _ { t } ^ { a }$ , which is sampled from their policy, $u ^ { a } \\sim \\pi _ { \\theta } ^ { a } ( u ^ { a } | \\tau ^ { a } )$ , while all other agents take a no-op action. Here $\\tau ^ { a }$ is the action-observation history of agent $a$ , $\\tau ^ { a } = \\{ o _ { 0 } ^ { a } , u _ { 0 } ^ { a } , r _ { 1 } , . . r _ { T } , o _ { T } ^ { a } \\}$ , $T$ is the length of the episode and $\\theta$ are the weights of a function approximator that represents the policy, in our case recurrent neural networks, such as LSTMs (Hochreiter & Schmidhuber, 1997). ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We further use $\\tau _ { t }$ to describe the state-action sequence, $\\tau = \\{ s _ { 0 } , \\mathbf { u _ { 0 } } , r _ { 1 } , . . r _ { T } , s _ { T } \\}$ , where $\\mathbf { u _ { t } }$ is the joint action of all agents. ", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As is typical in cooperative multi-agent RL, the goal of the agents is to maximize the total expected return, $J _ { \\theta } = \\mathbb { E } _ { \\tau \\sim P ( \\tau \\mid \\theta ) } R _ { 0 } ( \\tau )$ , where $R _ { 0 } ( \\tau )$ is the return of the trajectory (in general $R _ { t } \\bar { ( } \\tau ) ~ =$ $\\sum _ { t ^ { \\prime } \\geq t } \\gamma ^ { t ^ { \\prime } - t } r _ { t ^ { \\prime } } )$ and $\\gamma$ is an optional discount factor. We have also assumed that agents are sharing parameters, $\\theta$ , as is common in cooperative MARL. ", + "bbox": [ + 174, + 138, + 825, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 DISTRIBUTED RECURRENT DQN AND AUXILIARY TASKS ", + "text_level": 1, + "bbox": [ + 173, + 227, + 612, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Q-learning the agent approximates the expected return for a given state action-pair, $s , u$ , assuming that the agent acts greedily with respect to the Q-function for all future time steps, $Q ( s , u ) =$ $\\mathbb { E } _ { \\tau \\sim P ( \\tau | s , u ) } R _ { t } ( \\tau )$ , where $\\tau = \\{ s _ { t } , u _ { t } , r _ { t + 1 } , . . . , s _ { T } \\}$ , $u _ { t } = u$ and $u _ { t ^ { \\prime } } = \\arg \\operatorname* { m a x } _ { u ^ { \\prime } } Q ( s _ { t ^ { \\prime } } , u ^ { \\prime } ) , \\forall t ^ { \\prime } >$ $t$ . A common exploration scheme is $\\epsilon$ -greedy, in which the agent takes a random action with probability $\\epsilon$ and acts greedily otherwise. Importantly, the Q-function can be trained efficiently using the Bellman equation: $Q ( s , u ) = \\mathbb { E } _ { s ^ { \\prime } } [ r _ { t + 1 } + \\gamma \\operatorname* { m a x } _ { u ^ { \\prime } } Q ( s ^ { \\prime } , u ^ { \\prime } ) ]$ , where for simplicity we have assumed a deterministic reward. In Deep Q-Learning (DQN) (Mnih et al., 2015) the Q-function is parameterized by a deep neural network and trained with transitions sampled from experience replay. ", + "bbox": [ + 174, + 257, + 825, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our work we also incorporate other best practice components of the last few years, including doubleDQN (van Hasselt et al., 2015), dueling network architecture (Wang et al., 2015) and prioritized replay (Schaul et al., 2015). We also employ a distributed training architecture similar to the one proposed by Horgan et al. (2018) where a number of different actors with their own exploration rates collect experiences in parallel and feed them into a central replay buffer. Since our setting is partially observable the natural choice for the function approximator is a recurrent neural network. A combination of these techniques was first explored by Kapturowski et al. (2019) in single agent environments such as Atari and DMLab-30. ", + "bbox": [ + 174, + 376, + 825, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another common best-practice in RL are auxiliary tasks Mirowski et al. (2016); Jaderberg et al. (2016), in which the agent produces extra output-heads that are trained on supervised tasks and optimized alongside the RL loss. ", + "bbox": [ + 174, + 494, + 825, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 CENTRALISED TRAINING, DECENTRALIZED EXECUTION AND JOINT Q-FUNCTIONS ", + "bbox": [ + 176, + 564, + 790, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The most straight forward application of Q-learning to multi-agent settings is Independent Q-Learning (IQL) (Tan, 1993) in which each agent keeps an independent estimate of the expected return, treating all other agents as part of the environment. One challenge with IQL is that the exploratory behavior of other agents is not corrected for via the max operator in the bootstrap. Notably, IQL does typically not take any advantage of centralized training with decentralised control (CT/DC), a paradigm under which information can be exchanged freely amongst agents during the training phase as long as the policies rely only on local observations during execution. ", + "bbox": [ + 173, + 594, + 825, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are various approaches for learning joint-Q-functions in the CT/DC regime. For example, Value-Decomposition-Networks (VDN) (Sunehag et al., 2017) represent the joint-Q-function as a sum of per-agent contributions and QMIX (Rashid et al., 2018) learns a non-linear but monotonic combination of these contributions. ", + "bbox": [ + 174, + 699, + 825, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 METHOD ", + "text_level": 1, + "bbox": [ + 174, + 785, + 281, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 THEORY OF MIND AND BAYESIAN REASONING ", + "text_level": 1, + "bbox": [ + 174, + 823, + 540, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the very core of interpreting the actions of another agent, and ToM in general, is Bayesian reasoning. Fundamentally, asking what a given action by another agent implies about the state of the world requires understanding of why this action was taken. To illustrate this, we start out with an agent that has a given belief about the state-action history of the world, $\\tau _ { t }$ , given her own action-observation history $\\tau _ { t } ^ { a }$ : $B ( \\tau _ { t } ) = P ( \\tau _ { t } | \\tau _ { t } ^ { a } )$ . ", + "bbox": [ + 174, + 853, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next the agent observes the action $u _ { t } ^ { a ^ { \\prime } }$ of her team mate, $a ^ { \\prime }$ , and carries out a Bayesian update: ", + "bbox": [ + 169, + 102, + 792, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a2a28603de33f151bf7cd0c953e46f71ea22ac1ecdfe82cad7f1eacf3ca3f110.jpg", + "text": "$$\n\\begin{array} { r l } & { P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = \\frac { P ( u _ { t } ^ { a ^ { \\prime } } | \\tau _ { t } ) P ( \\tau _ { t } | \\tau _ { t } ^ { a } ) } { \\sum _ { \\tau _ { t } ^ { \\prime } } P ( u _ { t } ^ { a ^ { \\prime } } | \\tau _ { t } ^ { \\prime } ) P ( \\tau _ { t } ^ { \\prime } | \\tau _ { t } ^ { a } ) } } \\\\ & { \\qquad = \\frac { \\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) ) B ( \\tau _ { t } ) } { \\sum _ { \\tau _ { t } ^ { \\prime } } \\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ^ { \\prime } ) ) B ( \\tau _ { t } ^ { \\prime } ) } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 341, + 126, + 655, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, with a slight abuse of notation, we have used (and will keep using) $O ( a ^ { \\prime } , \\tau _ { t } )$ for the actionobservation history, $\\tau _ { t } ^ { a ^ { \\prime } }$ , that results from applying the observation function for agent $a ^ { \\prime }$ to $\\tau _ { t }$ at each time step. Note that for non-deterministic observation functions we would have to marginalize over $P ( \\tau _ { t } ^ { a ^ { \\prime } } | \\bar { \\tau _ { t } } )$ . ", + "bbox": [ + 173, + 222, + 826, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clearly, since agents have access to the policy of their teammate during centralised training, we could in principle evaluate this explicit Bayesian belief. However, beyond the practical difficulty of computing this explicit belief, when it is used as an input to the policy it will lead to prohibitively costly higher order beliefs. The typical workout for this is a public belief over private features which only conditions on common knowledge and can therefore be calculated by all agents individually, we refer to Moravcík et al. (2017); Nayyar et al. (2013); Foerster et al. (2018b) for more details. ˇ ", + "bbox": [ + 173, + 289, + 825, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead, in this work we rely on RNNs to learn implicit representations of the sufficient statistics over the distribution of the Markov state given the action-observation histories, noting that they are unlikely to recover exact beliefs due to the issues mentioned above. ", + "bbox": [ + 176, + 380, + 821, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 EXPLORATION AND BELIEFS ", + "text_level": 1, + "bbox": [ + 176, + 440, + 413, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next we illustrate the impact of exploration on the beliefs, which we will do in the explicit (exact) case, since it serves as an upper bound on the accuracy of the implicit beliefs. Since we are looking at fully-cooperative settings we assume that the optimal policy of the agent is deterministic and any randomness is due to exploration. Given that we are focused on value based methods we furthermore assume an $\\epsilon$ -greedy exploration scheme, noting that the same analysis can be extended to other methods. Under this exploration scheme $\\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) )$ ) becomes: ", + "bbox": [ + 173, + 465, + 825, + 554 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0a52e16234105ddd8c60fd2ee098383204ea93e1db137247763f8bfdcacb7de2.jpg", + "text": "$$\n\\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) ) = ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ,\n$$", + "text_format": "latex", + "bbox": [ + 325, + 560, + 671, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we have used $\\boldsymbol { u } ^ { * } ( \\tau _ { t } )$ to indicate the greedy action of the agent $a ^ { \\prime }$ , $\\begin{array} { r l } { u ^ { * } ( \\tau _ { t } ) } & { { } = } \\end{array}$ arg $\\operatorname* { m a x } _ { u } Q ^ { a ^ { \\prime } } ( u , O ( a ^ { \\prime } , \\tau _ { t } ) )$ and $\\mathbf { I }$ is the indicator function. ", + "bbox": [ + 173, + 588, + 823, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the first part corresponds to a filtering operator, in which the indicator function only attributes finite probability to those histories that are consistent with the action taken under greedy execution, the exploration term adds a fixed (history independent) probability, which effectively ‘blurs’ the posterior: ", + "bbox": [ + 174, + 626, + 825, + 683 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/8c13d7bf6f8e98e87f500ef80bef560a22101e5bea382a8ce2a57da50625424b.jpg", + "text": "$$\n\\begin{array} { r l r } { { P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau _ { t } ) } { \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau ^ { \\prime } ) } } } \\\\ & { } & { = \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau _ { t } ) } { \\epsilon / | U | + \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau ^ { \\prime } ) } } \\\\ & { } & { = \\frac { B ( \\tau _ { t } ) } { 1 + | U | \\sum _ { s ^ { \\prime } } ( ( 1 / \\epsilon - 1 ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) B ( \\tau ^ { \\prime } ) } } \\\\ & { } & { + \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau _ { t } ) } { \\epsilon / | U | + \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau ^ { \\prime } ) } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 295, + 708, + 699, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We find that the posterior includes an additional term of the form $B ( \\tau _ { t } )$ which carries over an unfiltered density over the trajectories from the prior. We further confirm that in the limit of $\\epsilon = 1$ , the posterior collapses to the prior, $P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = B ( \\tau _ { t } )$ . This can be particularly worrisome in the context of our training setup, whereby different agents run different, and potentially high, \u000f throughout the course of training. It fundamentally makes the beliefs obtained less informative. ", + "bbox": [ + 173, + 878, + 825, + 925 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While not making the above argument explicitly, the Bayesian Action Decoder (BAD) (Foerster et al., 2019), resolves this issue by shifting exploration to the level of deterministic partial policies, rather than action-level, and tracking an approximate Bayesian belief. As outlined in Section 1 this comes at a huge cost in the complexity of the method, the computation requirements and in the loss of generality of the method. ", + "bbox": [ + 174, + 138, + 825, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 SIMPLIFIED ACTION DECODING ", + "text_level": 1, + "bbox": [ + 176, + 227, + 436, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper we take a drastically simpler and different approach towards the issue. We note that the ‘blurring’, which makes decoding of an action challenging, is entirely due to the $\\epsilon$ -greedy exploration term. Furthermore, in order for another agent to do an implicit Bayesian update over an action taken, it is not required that this action is executed by the environment. Indeed, if we assume that other agents can observe the greedy action, $u ^ { * }$ , at every time step and condition their belief update on this, the terms depending on $\\epsilon$ disappear from the Bayesian update: ", + "bbox": [ + 173, + 253, + 825, + 338 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/62bf5170694637a8ab7015c0a7a0e4aeece9279c0b2d28977289192fc59283c5.jpg", + "text": "$$\nP ( \\tau _ { t } | \\tau _ { t } ^ { a } , u ^ { * } ) = \\frac { \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u ^ { * } ) \\big ) B ( \\tau _ { t } ) } { \\sum _ { \\tau ^ { \\prime } } \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u ^ { * } ) \\big ) B ( \\tau ^ { \\prime } ) }\n$$", + "text_format": "latex", + "bbox": [ + 357, + 347, + 640, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, to have our cake and eat it, in the Simplified Action Decoder (SAD) the acting agent is allowed to ‘take’ two actions at any given time step during training. The first action, $u ^ { a }$ , is the standard environment action, which gets executed as usual and is observed by all agents through the observation function at the next time step, as mentioned in Section 3. The second action, $u ^ { * }$ , is the greedy action of the active agent. This action does not get executed by the environment but instead is presented as an additional input to the other agents at the next time step, taking advantage of the centralized training regime during which information can be exchanged freely. ", + "bbox": [ + 173, + 398, + 825, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clearly we are not allowed to pass around extra information during decentralized control, but luckily this is not needed. Since we set $\\epsilon$ to 0 at test time we can simply use the, now greedy, environment action obtained from the observation function as our greedy-action input. ", + "bbox": [ + 176, + 503, + 821, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While this is most straight forward in settings where the last action is observed by other agents directly, in principle SAD can also be extended to settings where it is indirectly observed by all agents through the environment dynamics. In these cases we can replace the greedy-action side-channel with a learned inverse model that recovers the action from the observation history during execution. ", + "bbox": [ + 174, + 553, + 825, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, to encourage the agent to meaningfully decode the information contained in the greedy action, we can optionally add an auxiliary task to the training process, such as predicting unobserved information from their observation history . ", + "bbox": [ + 176, + 614, + 823, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While this idea is compatible with any deep RL algorithm with minimal modifications, we use a recurrent version of DQN with distributed training, dueling networks and prioritized replay. We also learn a joint Q-function using VDN in order to address the challenges of multi-agent off-policy learning, please see Section 3 for details on all of these standard methods. ", + "bbox": [ + 174, + 664, + 825, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 742, + 326, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 MATRIX GAME ", + "text_level": 1, + "bbox": [ + 174, + 775, + 321, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first verify the effectiveness of SAD in the two step, two player matrix game from Foerster et al. (2019), which replicates the communication through action challenge of Hanabi in a highly simplified setting. In this fully cooperative game each player obtains a privately observed ‘card’, which is drawn iid from two options (1,2). ", + "bbox": [ + 173, + 801, + 825, + 858 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After observing her card, the first player takes one of three possible discrete actions (1, 2, 3). Crucially, the second player observes both her own private card and the team mate’s action before acting herself, which establishes the opportunity to communicate. The payout is a function of both the two private cards and the two actions taken by both agents, as shown in Figure 1. ", + "bbox": [ + 174, + 864, + 825, + 921 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importantly, there are some obvious strategies that do not “Tiny Hanabi”require any communication. For example, if both player learn to play the 2nd action, the payout is always 8 points, independent of the cards dealt. However, if the players do ● Each player knows their learn to communicate it is possible to achieve 10 points for every pair of cards dealt. ", + "bbox": [ + 174, + 103, + 549, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 HANABI ", + "text_level": 1, + "bbox": [ + 174, + 204, + 272, + 218 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "● Player 2 observes Player Hanabi is a fully cooperative card game in which all play1’s action and acts ers work together to complete piles of cards referred to as second.fireworks. Each card has a rank, 1 to 5, and a color, G / B $/ \\textbf { W } / \\textbf { Y } / \\textbf { R }$ . Each firework (one per color) starts with a 1 and is finished once the 5 has been added. There are three 1s, one 5 and two of all other ranks for each of the colors, adding up to a total of 50 cards in the deck. The twist in Hanabi is that while players can observe the cards held by their team mates, they cannot observe their own cards and thus need to exchange information with each other in order to understand what cards can be played. There are two main means for doing so: First of all, players can take grounded hint actions, in which they reveal the subset of a team mate’s hand that matches a specific rank or color. An example hint is “Your third and fifth card are 1s”. These hint actions cost scarce information tokens, which can be replenished by discarding a card, an action that both removes the card from the game and makes it visible to all players. ", + "bbox": [ + 174, + 231, + 549, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25f404762b879bc27cf6c88933eae05b823c51b2b16509fef1e770d1b412b811.jpg", + "image_caption": [ + "Figure 1: Illustration of the matrix game from Foerster et al. (2019) " + ], + "image_footnote": [], + "bbox": [ + 563, + 109, + 818, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 356, + 825, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally players can also choose to play a card. If this card is the next card for the firework of the corresponding color, it is added to the firework and the team scores one point. Otherwise the card is removed from the game, the identity is made public, and the team loses one of the 3 life tokens. If the team runs out of life tokens before the end of the game, all points collected so far are lost and the game finishes immediately. These rules result in a maximum score of $5 \\times 5 = 2 5$ points in any game, which corresponds to all five fireworks being completed with five cards per firework. ", + "bbox": [ + 174, + 445, + 825, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure reproducibility and comparability of our results we use the Hanabi Learning Environment (HLE) (Bard et al., 2019) for all experimentation. For further details regarding Hanabi and the self-play part of the Hanabi challenge please see Bard et al. (2019). ", + "bbox": [ + 174, + 536, + 825, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 ARCHITECTURE AND COMPUTATION REQUIREMENTS ", + "text_level": 1, + "bbox": [ + 174, + 595, + 583, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We borrow some ideas and insights from prior distributed Q-learning methods while bring extensions to MARL as well as innovations to improve throughput and efficiency. Following Horgan et al. (2018) and Kapturowski et al. (2019), we use a distributed prioritized replay buffer shared by $N$ asynchronous actors and a centralized trainer that samples mini-batches from the replay buffer to update the model. In each actor thread, we run $K$ environments sequentially and batch their observations together. The observation batch is then fed into an actor that utilize a GPU to compute a batch of actions. All asynchronous actors share one GPU and the trainer uses another GPU for gradient computation and model updates. This is different from prior works which run single actor and single environment in each thread on a CPU. Our method enables us to run a very large number of simulations with moderate computation resources. In all Hanabi experiments, we run $N = 8 0$ actor threads with $K = 8 0$ environments in each thread on single machine with 40 CPU cores and 2 GPUs. Without this architectural improvement, it may require at least a few hundred CPU cores to run 6400 Hanabi environments, in which case neural network agents and simulations have to be distributed across multiple machines, greatly reducing the reproducibility and accessibility of such research. Please refer to Appendix A for implementation details and hyper-parameters. ", + "bbox": [ + 174, + 621, + 825, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 281, + 117 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1 MATRIX GAME ", + "text_level": 1, + "bbox": [ + 174, + 133, + 320, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As we can see in Figure 2, even in our simple matrix game the greedy action input makes a drastic difference. With an average reward of around 9.5 points, tabular IQL does well in this task, matching the $B A D$ results from Foerster et al. (2019). However, just by adding the greedy action as an additional input, we obtain an average performance of $9 . 9 7 \\pm 0 . 0 2$ . Results are averaged over 100 seeds, and shading is s.e.m. The code is available here: www.bit.ly/2mBJLyk. ", + "bbox": [ + 174, + 161, + 517, + 300 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ad1d82f24e9ee7d32e4c518b21671940bf756b1552a474256a05432182b3238.jpg", + "image_caption": [ + "Figure 2: Results for the matrix game. " + ], + "image_footnote": [], + "bbox": [ + 534, + 103, + 800, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2 HANABI ", + "text_level": 1, + "bbox": [ + 174, + 319, + 272, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 1, our findings from the matrix game are for the most part confirmed on the challenging Hanabi benchmark. To illustrate the contributions of the different components, we compare average scores and win rates across 13 independent training runs of SAD and three different options: IQL is simply the recurrent DQN agent with parameter sharing, VDN is the same agent but also learns a joint Q-function and finally SAD & AuxTask is the SAD agent with the auxiliary task. ", + "bbox": [ + 174, + 345, + 825, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While we find that SAD significantly outperforms our baselines (IQL and VDN) for 2, 4 and 5 players in terms of average score and/or win rate, there is no significant difference for 3 players, where VDN matches the performance of SAD. ", + "bbox": [ + 174, + 422, + 825, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interestingly, the auxiliary task only significantly helps the 2-player performance, where it substantially boosts the average score and win rate. In contrast, it drastically hurts performance for 3-5 players, which opens an interesting avenue for future work. ", + "bbox": [ + 174, + 472, + 825, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For completeness we have included training curves showing average scores and s.e.m. across all training runs for all numbers of players for our methods and ablations in Appendix B. We find that for 5 players the auxiliary task drastically reduces the variance of SAD and intermittently leads to higher performance during training but ultimately results in lower final performance. We can also clearly see that despite 72 hours of training and billions of samples consumed, the performance has not plateaued for 3-5 players, pointing to an obvious avenue for further improvements. ", + "bbox": [ + 174, + 520, + 825, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The original numbers in the Hanabi challenge and BAD used population based training (Jaderberg et al., 2018), effectively reporting maximum performance across a large number of different runs. Therefore, for reproducibility purposes, we report evaluations of the best model from our various training runs for each method in Table 2. ", + "bbox": [ + 174, + 611, + 825, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown, under this reporting we establish a new SOTA for learning methods on the self-play part of the Hanabi challenge for 2-5 players, with the most drastic improvements being achieved for 3-5 players. In particular, we beat both the ACHA agent from Bard et al. (2019) and the BAD agent on average score, even though both of them used population based training and require more compute. We note that while we follow the counting convention proposed by the challenge paper, BAD was optimized for a different counting scheme, in which agents keep their scores when they run out of lives. This may explain the higher win rate $( 5 8 . 6 \\% )$ of BAD combined with a relatively low mean score, which is exceeded even by our baseline methods. Once again, only the performance for 2-player is significantly improved by the auxiliary task and the 3-player setting is an outlier in the sense that SAD does not improve the best performance compared to VDN. ", + "bbox": [ + 174, + 674, + 825, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7 CONCLUSION AND FUTURE WORK ", + "text_level": 1, + "bbox": [ + 174, + 835, + 495, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper we presented the Simplified Action Decoder (SAD), a novel deep multi-agent RL algorithm that allows agents to learn communication protocols in settings where no cheap-talk channel is available. On the challenging benchmark Hanabi our work substantially improves the SOTA for an RL method for all numbers of players. For two players SAD establishes a new high-score across any method. Furthermore we accomplish all of this with a method that is both simpler and requires less compute than previous advances. While these are encouraging steps, there is clearly more work to do. In particular, there remains a large performance gap between the numbers achieved by SAD and the known performance of hat-coding strategies (Wu, 2018) for 3-5 players. One possible reason is that SAD does not undertake any explicit exploration in the space of possible conventions. Another promising route for future work is to integrate search with RL, since this has produced SOTA results in a number of different domains including Poker, Go and backgammon. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/06f1e8f230e055bf72fef14c157586d32d70dba51aea45b44487b3a4cfae18fc.jpg", + "table_caption": [ + "Table 1: Mean performance of our methods and baselines on Hanabi. We take the final models of 13 independent runs, i.e. 13 models per algorithm per player setting. Each model is evaluated on 100K games. Mean and s.e.m over the mean scores of the 13 models are shown in the table. The second row of each section is the win rate. " + ], + "table_footnote": [], + "table_body": "
Agent 2 Players 3 Players 4 Players 5 Players
IQL (Baseline)23.77 ± 0.04 43.88 ± 1.21 %23.02 ± 0.10 26.16 ± 2.01 %21.99 ± 0.09 10.15 ± 0.86 %20.60 ± 0.11 2.27 ± 0.32 %
VDN23.83 ± 0.0323.71 ± 0.0623.03 ± 0.1521.18 ± 0.12
(Baseline) SAD44.97 ± 1.28 % 23.87 ± 0.0341.16 ± 1.27 % 23.69 ± 0.0523.57 ± 2.20 % 23.27 ± 0.162.26 ± 0.32 % 22.06 ± 0.23
SAD AuxTask47.90 ± 1.10 % 24.02 ± 0.0141.12 ± 1.10 % 23.56 ± 0.0729.38 ± 2.63 % 22.78 ± 0.107.22 ± 1.29 % 21.47 ± 0.08
", + "bbox": [ + 210, + 103, + 787, + 262 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/20d58361e17764c14fcd32f70066f1ef6e07903eb74e90302ea7c9ff64241dd4.jpg", + "table_caption": [ + "Table 2: Comparison between the previous SOTA learning methods and ours. We take the best model of 13 runs for each of our methods and baselines. Each model is evaluated on 100K games with different seeds. Mean and s.e.m over the 100K games are shown in the table. The s.e.m. is less than 0.01 for most models. Bold numbers are the best results achieved with learning algorithms. The second row of each section is the win rate. " + ], + "table_footnote": [], + "table_body": "
Agent2 Players 3 Players 4 Players 5 Players
Rainbow (Bard et al., 2019)20.64 ± 0.03 2.5%18.71 ±0.01 0.2%18.00 ± 0.17 0%15.26 ± 0.18 0%
ACHA (Bard et al., 2019)22.73 ± 0.12 15.1%20.24 ± 0.15 1.1%21.57 ± 0.12 2.4%16.80 ± 0.13 0%
BAD (Foerster et al.,2019)23.92 ± 0.01 58.56%==1
IQL (Baseline) 50.47%23.97 ± 0.01 40.25%23.69 ± 0.01 19.39%22.76 ± 0.01 4.93%21.29 ± 0.01
VDN (Baseline)23.96 ± 0.01 50.27%23.99 ± 0.01 50.37%23.79 ± 0.00 38.86%21.80 ± 0.01 4.98%
SAD24.01 ± 0.01 52.39%23.93 ± 0.01 48.05%23.81 ± 0.01 41.45%23.01 ± 0.01 13.93%
SAD& AuxTask24.08 ± 0.01 56.09%23.81 ± 0.01 49.74%23.47 ± 0.01 33.87%22.25 ± 0.01 7.33%
", + "bbox": [ + 205, + 337, + 792, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 700, + 825, + 799 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chris L Baker, Julian Jara-Ettinger, Rebecca Saxe, and Joshua B Tenenbaum. Rational quantitative attribution of beliefs, desires and percepts in human mentalizing. Nature Human Behaviour, 1(4): 0064, 2017. ", + "bbox": [ + 174, + 126, + 825, + 167 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad, Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The Hanabi Challenge: A New Frontier for AI Research. arXiv:1902.00506 [cs, stat], February 2019. URL http://arxiv. org/abs/1902.00506. arXiv: 1902.00506. ", + "bbox": [ + 174, + 175, + 826, + 246 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Noam Brown and Tuomas Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top professionals. Science, pp. eaao1733, 2017. ", + "bbox": [ + 173, + 253, + 825, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Noam Brown and Tuomas Sandholm. Superhuman AI for multiplayer poker. Science, pp. eaay2400, 2019. ", + "bbox": [ + 173, + 290, + 825, + 319 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Noam Brown, Tuomas Sandholm, and Brandon Amos. Depth-Limited Solving for ImperfectInformation Games. pp. 14. ", + "bbox": [ + 173, + 327, + 825, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Murray Campbell, A Joseph Hoane Jr, and Feng-hsiung Hsu. Deep Blue. Artificial intelligence, 134 (1-2):57–83, 2002. ", + "bbox": [ + 173, + 363, + 825, + 392 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rodrigo Canaan, Julian Togelius, Andy Nealen, and Stefan Menzel. Diverse agents for ad-hoc cooperation in hanabi. arXiv preprint arXiv:1907.03840, 2019. ", + "bbox": [ + 171, + 400, + 823, + 429 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. In Advances in Neural Information Processing Systems, pp. 2137–2145, 2016. ", + "bbox": [ + 173, + 435, + 825, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jakob Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson, Matthew Botvinick, and Michael Bowling. Bayesian action decoder for deep multi-agent reinforcement learning. In International Conference on Machine Learning, pp. 1942–1951, 2019. ", + "bbox": [ + 173, + 487, + 825, + 530 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018a. ", + "bbox": [ + 173, + 537, + 825, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jakob N. Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson, Matthew Botvinick, and Michael Bowling. Bayesian Action Decoder for Deep Multi-Agent Reinforcement Learning. arXiv:1811.01458 [cs], November 2018b. URL http://arxiv. org/abs/1811.01458. arXiv: 1811.01458. ", + "bbox": [ + 173, + 588, + 826, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "James Goodman. Re-determinizing information set monte carlo tree search in hanabi. arXiv preprint arXiv:1902.06075, 2019. ", + "bbox": [ + 169, + 651, + 823, + 680 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735– 1780, November 1997. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL http://dx. doi.org/10.1162/neco.1997.9.8.1735. ", + "bbox": [ + 173, + 689, + 826, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt, and David Silver. Distributed prioritized experience replay. CoRR, abs/1803.00933, 2018. URL http://arxiv.org/abs/1803.00933. ", + "bbox": [ + 173, + 738, + 825, + 781 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016. ", + "bbox": [ + 174, + 789, + 825, + 832 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castaneda, Charles Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel. Human-level performance in first-person multiplayer games with population-based deep reinforcement learning. arXiv:1807.01281 [cs, stat], July 2018. doi: 10.1126/science.aau6249. URL http://arxiv.org/abs/1807.01281. arXiv: 1807.01281. ", + "bbox": [ + 174, + 840, + 826, + 922 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Steven Kapturowski, Georg Ostrovski, John Quan, and Will Dabney. RECURRENT EXPERIENCE REPLAY IN DISTRIBUTED REINFORCEMENT LEARNING. pp. 1–19, 2019. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ", + "bbox": [ + 171, + 143, + 823, + 172 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Claire Liang, Julia Proft, Erik Andersen, and Ross A Knepper. Implicit communication of actionable information in human-ai teams. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pp. 95. ACM, 2019. ", + "bbox": [ + 173, + 185, + 825, + 227 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems, pp. 6379–6390, 2017. ", + "bbox": [ + 174, + 239, + 823, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in complex environments. arXiv preprint arXiv:1611.03673, 2016. ", + "bbox": [ + 174, + 294, + 825, + 337 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 02 2015. URL http://dx.doi.org/10.1038/nature14236. ", + "bbox": [ + 174, + 348, + 826, + 420 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Matej Moravcík, Martin Schmid, Neil Burch, Viliam Lis ˇ y, Dustin Morrill, Nolan Bard, Trevor \\` Davis, Kevin Waugh, Michael Johanson, and Michael Bowling. Deepstack: Expert-level artificial intelligence in heads-up no-limit poker. Science, 356(6337):508–513, 2017. ", + "bbox": [ + 174, + 431, + 825, + 473 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ashutosh Nayyar, Aditya Mahajan, and Demosthenis Teneketzis. Decentralized stochastic control with partial history sharing: A common information approach. IEEE Transactions on Automatic Control, 58(7):1644–1658, 2013. ", + "bbox": [ + 174, + 486, + 825, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Thanh Thi Nguyen, Ngoc Duy Nguyen, and Saeid Nahavandi. Deep reinforcement learning for multiagent systems: a review of challenges, solutions and applications. arXiv preprint arXiv:1812.11794, 2018. ", + "bbox": [ + 174, + 540, + 825, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Arthur O’Dwyer. Hanabi. https://github.com/Quuxplusone/Hanabi, 2019. ", + "bbox": [ + 171, + 595, + 758, + 611 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Frans A Oliehoek. Decentralized pomdps. In Reinforcement Learning, pp. 471–503. Springer, 2012. ", + "bbox": [ + 173, + 622, + 823, + 637 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tabish Rashid, Mikayel Samvelyan, Christian Schröder de Witt, Gregory Farquhar, Jakob N. Foerster, and Shimon Whiteson. QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning. CoRR, abs/1803.11485, 2018. URL http://arxiv.org/abs/ 1803.11485. ", + "bbox": [ + 174, + 648, + 826, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized Experience Replay. arXiv e-prints, art. arXiv:1511.05952, Nov 2015. ", + "bbox": [ + 171, + 717, + 825, + 746 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. ", + "bbox": [ + 174, + 758, + 825, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without human knowledge. Nature, 550(7676):354, 2017. ", + "bbox": [ + 176, + 813, + 825, + 856 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419): 1140–1144, 2018. ", + "bbox": [ + 174, + 867, + 825, + 922 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sainbayar Sukhbaatar, arthur szlam, and Rob Fergus. Learning Multiagent Communication with Backpropagation. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp. 2244–2252. Curran Associates, Inc., 2016. URL http://papers.nips.cc/paper/ 6398-learning-multiagent-communication-with-backpropagation.pdf. ", + "bbox": [ + 174, + 103, + 828, + 174 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinícius Flores Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z. Leibo, Karl Tuyls, and Thore Graepel. Value-decomposition networks for cooperative multi-agent learning. CoRR, abs/1706.05296, 2017. URL http://arxiv.org/abs/1706.05296. ", + "bbox": [ + 174, + 183, + 826, + 238 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Richard S. Sutton. Learning to predict by the methods of temporal differences. Mach. Learn., 3(1):9–44, August 1988. ISSN 0885-6125. doi: 10.1023/A:1022633531479. URL https: //doi.org/10.1023/A:1022633531479. ", + "bbox": [ + 173, + 247, + 828, + 290 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the tenth international conference on machine learning, pp. 330–337, 1993. ", + "bbox": [ + 173, + 299, + 823, + 329 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hado van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double qlearning. CoRR, abs/1509.06461, 2015. URL http://arxiv.org/abs/1509.06461. ", + "bbox": [ + 173, + 337, + 825, + 366 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ziyu Wang, Nando de Freitas, and Marc Lanctot. Dueling network architectures for deep reinforcement learning. CoRR, abs/1511.06581, 2015. URL http://arxiv.org/abs/1511.06581. ", + "bbox": [ + 174, + 376, + 825, + 404 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jeff Wu. Hanabi simulation in rust. https://github.com/WuTheFWasThat/hanabi.rs, 2018. ", + "bbox": [ + 169, + 414, + 826, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A NETWORK ARCHITECTURE AND HYPER-PAMAMETERS FOR HANABI ", + "text_level": 1, + "bbox": [ + 173, + 102, + 777, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our Hanabi agent uses dueling network architecture (Wang et al., 2015). The main body of the network consists of 1 fully connected layer of 512 units and 2 LSTM (Hochreiter & Schmidhuber, 1997) layers of 512 units, followed by two output heads for value and advantages respectively. The same network configuration is used across all Hanabi experiments. We take the default featurization of HLE and replace the card knowledge section with the V0-Belief proposed by Foerster et al. (2019). The maximum length of an episode is capped at 80 steps and the entire episode is stored in the replay buffer as one training sample. This avoids the “slate hidden states” problem as described in Kapturowski et al. (2019) because we can simply initialize the hidden states of LSTM as zero during training. For exploration and experience prioritization, we follow the simple strategy as in Horgan et al. (2018) and Kapturowski et al. (2019). Each actor executes an $\\epsilon _ { i }$ -greedy policy where $\\epsilon _ { i } = \\bar { \\epsilon } ^ { 1 + \\frac { 1 } { N - 1 } \\alpha }$ for $i \\in \\{ 0 , . . . , N - 1 \\}$ but with a smaller $\\epsilon = 0 . 1$ and $\\alpha = 7$ . For simplicity, all players of a game use the same epsilon. The per time-step priority $\\delta _ { t }$ is the TD error and per episode priority is computed following $\\delta _ { e } = \\eta \\operatorname* { m a x } _ { t } \\delta _ { i } + ( 1 - \\eta ) \\hat { \\delta }$ where $\\eta = 0 . 9$ . Priority exponent is set to 0.9 and importance sampling exponent is set to 0.6. We use $n$ -step return (Sutton, 1988) and double Q-learning (van Hasselt et al., 2015) for target computation during training. The discount factor $\\gamma$ is set to 0.999. The network is updated using Adam optimizer (Kingma & Ba, 2014) with learning rate $l r = 6 . 2 5 \\times 1 0 ^ { - 5 }$ and $\\epsilon = \\dot { 1 . 5 } \\times 1 0 ^ { - 5 }$ . Trainer sends its network weights to all actors every 10 updates and target network is synchronized with online network every 2500 updates. These hyper-parameters are fixed across all experiments. ", + "bbox": [ + 173, + 133, + 826, + 402 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the baseline, we use Independent Q-Learning where each player estimates the Q value and selects action independently at each time-step. Note that all players need to operate on the observations in order to update their recurrent hidden states while only the current player has non-trivial legal moves and other players can only select ‘pass’. Each player then writes its own version of the episode into the prioritized replay buffer and they are sampled independently during training. The prioritized replay buffer contains $2 ^ { 1 7 } ( 1 3 1 0 7 2 )$ episodes. We warm up the replay buffer with 10,000 episodes before training starts. Batch size during training is 128 for games of different numbers of players. ", + "bbox": [ + 174, + 409, + 825, + 507 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As mentioned in Section 4, the SAD agent is built on top of joint Q-function where the Q value is the sum of the individual Q value of all players given their own actions. One episode produces only one training sample with an extra dimension for the number of players. The replay buffer size is reduced to $2 ^ { 1 6 }$ for 2-player and 3-player games and $2 ^ { 1 5 }$ for 4-player and 5-player games. The batch sizes for 2-, 3-, 4-, 5-players are 64, 43, 32, 26 respectively to account for the fact that each sample contains more data. ", + "bbox": [ + 174, + 513, + 825, + 597 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Auxiliary task can be added to the agent to help it decode the greedy action more effectively. In Hanabi, the natural choice is the predict the card of player’s own hand. In our experiments, the auxiliary task is to predict the status of a card, which can be playable, discardable, or unknown. The loss is the average cross entropy loss per card and is simply added to the TD-error of reinforcement learning during training. ", + "bbox": [ + 174, + 604, + 825, + 674 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B LEARNING CURVES FOR HANABI ", + "text_level": 1, + "bbox": [ + 176, + 102, + 483, + 118 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ccf9077cc9f7d726e6a7f0ea3a2fa7941948d5e7bedb2d3bea373709b6945a93.jpg", + "image_caption": [ + "Figure 3 shows learning curves of different algorithms averaged over 13 seeds per algorithm per player setting. Shading is error of the mean. ", + "Figure 3: Learning Curves " + ], + "image_footnote": [], + "bbox": [ + 179, + 176, + 812, + 619 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/parse/train/B1xm3RVtwB/B1xm3RVtwB_middle.json b/parse/train/B1xm3RVtwB/B1xm3RVtwB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..235c295178ccbcbd87b8c306d60a21b16eadbe3b --- /dev/null +++ b/parse/train/B1xm3RVtwB/B1xm3RVtwB_middle.json @@ -0,0 +1,33735 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 478, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 369, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 369, + 98 + ], + "score": 1.0, + "content": "SIMPLIFIED ACTION DECODER FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 479, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 479, + 118 + ], + "score": 1.0, + "content": "DEEP MULTI-AGENT REINFORCEMENT LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 251, + 168 + ], + "lines": [ + { + "bbox": [ + 112, + 135, + 253, + 146 + ], + "spans": [ + { + "bbox": [ + 112, + 135, + 253, + 146 + ], + "score": 1.0, + "content": "Hengyuan Hu, Jakob N Foerster", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 249, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 249, + 158 + ], + "score": 1.0, + "content": "Facebook AI Research, CA, USA", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 158, + 240, + 169 + ], + "spans": [ + { + "bbox": [ + 114, + 158, + 240, + 169 + ], + "score": 1.0, + "content": "{hengyuan,jnf}@fb.com", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 278, + 198, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "spans": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 142, + 222, + 468, + 475 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 235 + ], + "score": 1.0, + "content": "In recent years we have seen fast progress on a number of benchmark problems in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "score": 1.0, + "content": "AI, with modern methods achieving near or super human performance in Go, Poker", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 244, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 244, + 469, + 258 + ], + "score": 1.0, + "content": "and Dota. One common aspect of all of these challenges is that they are by design", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 469, + 268 + ], + "score": 1.0, + "content": "adversarial or, technically speaking, zero-sum. In contrast to these settings, success", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "in the real world commonly requires humans to collaborate and communicate with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 278, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 142, + 278, + 470, + 290 + ], + "score": 1.0, + "content": "others, in settings that are, at least partially, cooperative. In the last year, the card", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "score": 1.0, + "content": "game Hanabi has been established as a new benchmark environment for AI to fill", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 312 + ], + "score": 1.0, + "content": "this gap. In particular, Hanabi is interesting to humans since it is entirely focused", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "score": 1.0, + "content": "on theory of mind, i.e., the ability to effectively reason over the intentions, beliefs", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 322, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 142, + 322, + 469, + 333 + ], + "score": 1.0, + "content": "and point of view of other agents when observing their actions. Learning to be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 332, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 469, + 345 + ], + "score": 1.0, + "content": "informative when observed by others is an interesting challenge for Reinforcement", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 470, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 356 + ], + "score": 1.0, + "content": "Learning (RL): Fundamentally, RL requires agents to explore in order to discover", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 354, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 140, + 354, + 470, + 367 + ], + "score": 1.0, + "content": "good policies. However, when done naively, this randomness will inherently make", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 364, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 469, + 379 + ], + "score": 1.0, + "content": "their actions less informative to others during training. We present a new deep", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 376, + 469, + 388 + ], + "spans": [ + { + "bbox": [ + 142, + 376, + 469, + 388 + ], + "score": 1.0, + "content": "multi-agent RL method, the Simplified Action Decoder (SAD), which resolves", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 470, + 401 + ], + "score": 1.0, + "content": "this contradiction exploiting the centralized training phase. During training SAD", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 398, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 142, + 398, + 470, + 411 + ], + "score": 1.0, + "content": "allows other agents to not only observe the (exploratory) action chosen, but agents", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 410, + 469, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 469, + 421 + ], + "score": 1.0, + "content": "instead also observe the greedy action of their team mates. By combining this", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 420, + 470, + 433 + ], + "spans": [ + { + "bbox": [ + 141, + 420, + 470, + 433 + ], + "score": 1.0, + "content": "simple intuition with best practices for multi-agent learning, SAD establishes", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 432, + 469, + 443 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 443 + ], + "score": 1.0, + "content": "a new SOTA for learning methods for 2-5 players on the self-play part of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 442, + 469, + 453 + ], + "spans": [ + { + "bbox": [ + 141, + 442, + 469, + 453 + ], + "score": 1.0, + "content": "Hanabi challenge. Our ablations show the contributions of SAD compared with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 453, + 470, + 465 + ], + "spans": [ + { + "bbox": [ + 142, + 453, + 470, + 465 + ], + "score": 1.0, + "content": "the best practice components. All of our code and trained agents are available at", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 464, + 423, + 476 + ], + "spans": [ + { + "bbox": [ + 141, + 464, + 423, + 476 + ], + "score": 1.0, + "content": "https://github.com/facebookresearch/Hanabi_SAD.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 206, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 208, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 208, + 513 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 536 + ], + "score": 1.0, + "content": "Humans are highly social creatures and spend vast amounts of time coordinating, collaborating", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "and communicating with others. In contrast to these, at least partially, cooperative settings most", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 544, + 507, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 507, + 558 + ], + "score": 1.0, + "content": "progress on AI in games has been in zero-sum games where agents compete against each other,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 555, + 507, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 507, + 568 + ], + "score": 1.0, + "content": "typically rendering communication futile. This includes examples such as Go (Silver et al., 2016;", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 566, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 578 + ], + "score": 1.0, + "content": "2017; 2018), poker (Brown & Sandholm, 2017; Moravcík et al., 2017; Brown & Sandholm, 2019) ˇ", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 577, + 245, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 245, + 589 + ], + "score": 1.0, + "content": "and chess (Campbell et al., 2002).", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "This narrow focus is unfortunate, since communication and coordination require unique abilities. In", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "order to enable smooth and efficient social interactions of groups of people, it is commonly required", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 617, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 506, + 628 + ], + "score": 1.0, + "content": "to reason over the intents, points of views and beliefs of other agents from observing their actions.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "For example, a driver can reasonably infer that if a truck in front of them is slowing down when", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "score": 1.0, + "content": "approaching an intersection, then there is likely an obstacle ahead. Furthermore, humans are both", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "able to interpret the actions of others and can act in a way that is informative when their actions are", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "score": 1.0, + "content": "being observed by others, capabilities that are commonly called theory of Mind (ToM), (Baker et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "2017). Importantly, in order to carry out this kind of reasoning, an agent needs to consider why a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "given action is taken and what this decision indicates about the state of the world. Simply observing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 693, + 284, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 284, + 704 + ], + "score": 1.0, + "content": "what other agents are doing is not sufficient.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "While these abilities are particularly relevant in partially observable, fully cooperative multi-agent", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "settings, ToM reasoning clearly matters in a variety of real world scenarios. For example, autonomous", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 478, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 369, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 77, + 369, + 98 + ], + "score": 1.0, + "content": "SIMPLIFIED ACTION DECODER FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 479, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 479, + 118 + ], + "score": 1.0, + "content": "DEEP MULTI-AGENT REINFORCEMENT LEARNING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 251, + 168 + ], + "lines": [ + { + "bbox": [ + 112, + 135, + 253, + 146 + ], + "spans": [ + { + "bbox": [ + 112, + 135, + 253, + 146 + ], + "score": 1.0, + "content": "Hengyuan Hu, Jakob N Foerster", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 249, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 249, + 158 + ], + "score": 1.0, + "content": "Facebook AI Research, CA, USA", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 158, + 240, + 169 + ], + "spans": [ + { + "bbox": [ + 114, + 158, + 240, + 169 + ], + "score": 1.0, + "content": "{hengyuan,jnf}@fb.com", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 112, + 135, + 253, + 169 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 198, + 333, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "spans": [ + { + "bbox": [ + 276, + 196, + 336, + 212 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 142, + 222, + 468, + 475 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 235 + ], + "score": 1.0, + "content": "In recent years we have seen fast progress on a number of benchmark problems in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "score": 1.0, + "content": "AI, with modern methods achieving near or super human performance in Go, Poker", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 244, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 244, + 469, + 258 + ], + "score": 1.0, + "content": "and Dota. One common aspect of all of these challenges is that they are by design", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 469, + 268 + ], + "score": 1.0, + "content": "adversarial or, technically speaking, zero-sum. In contrast to these settings, success", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "in the real world commonly requires humans to collaborate and communicate with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 278, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 142, + 278, + 470, + 290 + ], + "score": 1.0, + "content": "others, in settings that are, at least partially, cooperative. In the last year, the card", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "score": 1.0, + "content": "game Hanabi has been established as a new benchmark environment for AI to fill", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 312 + ], + "score": 1.0, + "content": "this gap. In particular, Hanabi is interesting to humans since it is entirely focused", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 469, + 323 + ], + "score": 1.0, + "content": "on theory of mind, i.e., the ability to effectively reason over the intentions, beliefs", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 322, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 142, + 322, + 469, + 333 + ], + "score": 1.0, + "content": "and point of view of other agents when observing their actions. Learning to be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 332, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 469, + 345 + ], + "score": 1.0, + "content": "informative when observed by others is an interesting challenge for Reinforcement", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 470, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 356 + ], + "score": 1.0, + "content": "Learning (RL): Fundamentally, RL requires agents to explore in order to discover", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 140, + 354, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 140, + 354, + 470, + 367 + ], + "score": 1.0, + "content": "good policies. However, when done naively, this randomness will inherently make", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 364, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 469, + 379 + ], + "score": 1.0, + "content": "their actions less informative to others during training. We present a new deep", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 376, + 469, + 388 + ], + "spans": [ + { + "bbox": [ + 142, + 376, + 469, + 388 + ], + "score": 1.0, + "content": "multi-agent RL method, the Simplified Action Decoder (SAD), which resolves", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 387, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 470, + 401 + ], + "score": 1.0, + "content": "this contradiction exploiting the centralized training phase. During training SAD", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 398, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 142, + 398, + 470, + 411 + ], + "score": 1.0, + "content": "allows other agents to not only observe the (exploratory) action chosen, but agents", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 410, + 469, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 469, + 421 + ], + "score": 1.0, + "content": "instead also observe the greedy action of their team mates. By combining this", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 420, + 470, + 433 + ], + "spans": [ + { + "bbox": [ + 141, + 420, + 470, + 433 + ], + "score": 1.0, + "content": "simple intuition with best practices for multi-agent learning, SAD establishes", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 432, + 469, + 443 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 443 + ], + "score": 1.0, + "content": "a new SOTA for learning methods for 2-5 players on the self-play part of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 442, + 469, + 453 + ], + "spans": [ + { + "bbox": [ + 141, + 442, + 469, + 453 + ], + "score": 1.0, + "content": "Hanabi challenge. Our ablations show the contributions of SAD compared with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 453, + 470, + 465 + ], + "spans": [ + { + "bbox": [ + 142, + 453, + 470, + 465 + ], + "score": 1.0, + "content": "the best practice components. All of our code and trained agents are available at", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 464, + 423, + 476 + ], + "spans": [ + { + "bbox": [ + 141, + 464, + 423, + 476 + ], + "score": 1.0, + "content": "https://github.com/facebookresearch/Hanabi_SAD.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 17, + "bbox_fs": [ + 140, + 222, + 470, + 476 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 206, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 208, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 208, + 513 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 536 + ], + "score": 1.0, + "content": "Humans are highly social creatures and spend vast amounts of time coordinating, collaborating", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "and communicating with others. In contrast to these, at least partially, cooperative settings most", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 544, + 507, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 507, + 558 + ], + "score": 1.0, + "content": "progress on AI in games has been in zero-sum games where agents compete against each other,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 555, + 507, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 507, + 568 + ], + "score": 1.0, + "content": "typically rendering communication futile. This includes examples such as Go (Silver et al., 2016;", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 566, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 578 + ], + "score": 1.0, + "content": "2017; 2018), poker (Brown & Sandholm, 2017; Moravcík et al., 2017; Brown & Sandholm, 2019) ˇ", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 577, + 245, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 245, + 589 + ], + "score": 1.0, + "content": "and chess (Campbell et al., 2002).", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 520, + 507, + 589 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "This narrow focus is unfortunate, since communication and coordination require unique abilities. In", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "order to enable smooth and efficient social interactions of groups of people, it is commonly required", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 617, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 506, + 628 + ], + "score": 1.0, + "content": "to reason over the intents, points of views and beliefs of other agents from observing their actions.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "For example, a driver can reasonably infer that if a truck in front of them is slowing down when", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 650 + ], + "score": 1.0, + "content": "approaching an intersection, then there is likely an obstacle ahead. Furthermore, humans are both", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "able to interpret the actions of others and can act in a way that is informative when their actions are", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "score": 1.0, + "content": "being observed by others, capabilities that are commonly called theory of Mind (ToM), (Baker et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "2017). Importantly, in order to carry out this kind of reasoning, an agent needs to consider why a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "given action is taken and what this decision indicates about the state of the world. Simply observing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 693, + 284, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 284, + 704 + ], + "score": 1.0, + "content": "what other agents are doing is not sufficient.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 594, + 506, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "While these abilities are particularly relevant in partially observable, fully cooperative multi-agent", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "settings, ToM reasoning clearly matters in a variety of real world scenarios. For example, autonomous", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "cars will likely need to understand the point of view, intents and beliefs of other traffic participants in", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 477, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 477, + 106 + ], + "score": 1.0, + "content": "order to deal with highly interactive settings such as 4-way crossing or dense traffic in cities.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 46.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "cars will likely need to understand the point of view, intents and beliefs of other traffic participants in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 477, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 477, + 106 + ], + "score": 1.0, + "content": "order to deal with highly interactive settings such as 4-way crossing or dense traffic in cities.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "Hanabi is a fully cooperative, partially-observable card game that has recently been proposed as a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 134 + ], + "score": 1.0, + "content": "new benchmark challenge problem for AI research (Bard et al., 2019) to fill the gap around ToM. In", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "score": 1.0, + "content": "Hanabi, players need to find conventions that allow them to effectively exchange information from", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "their local observations through their actions, taking advantage of the fact that actions are observed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 181, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 181, + 167 + ], + "score": 1.0, + "content": "by all team mates.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "Most prior state-of-the-art agents for Hanabi were developed using handcrafted algorithms, which", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "beat off-the-shelf deep multi-agent RL methods by a large margin. This makes intuitive sense: Beyond", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "the “standard” multi-agent challenges of credit assignment, nonstationarity and joint exploration,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 205, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 506, + 216 + ], + "score": 1.0, + "content": "learning an informative policy presents an additional fundamentally new conflict. On the one hand,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "score": 1.0, + "content": "an RL agent needs to explore in order to discover good policies through trial and error. On the other", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "hand, when carried out naively, this exploration will add noise to the policy of the agent during the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 237, + 436, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 436, + 250 + ], + "score": 1.0, + "content": "training process, making their actions strictly less informative to their team mates.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 254, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "One possible solution to this is to explore in the space of deterministic partial policies, rather than", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "actions, and sample these policies from a distribution that conditions on a common knowledge", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "Bayesian belief. This is successfully carried out in the Bayesian Action Decoder (BAD) (Foerster", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 287, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 506, + 299 + ], + "score": 1.0, + "content": "et al., 2019), the only previous Deep RL method to achieve a state-of-the-art in Hanabi. While this is a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "notable accomplishment, it comes at the cost of simplicity and generality. For a start, BAD requires an", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 309, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 506, + 321 + ], + "score": 1.0, + "content": "explicit common knowledge Bayesian belief to be tracked, which not only adds computational burden", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "score": 1.0, + "content": "due to the required sampling steps, but also uses expert knowledge regarding the game dynamics.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 343 + ], + "score": 1.0, + "content": "Furthermore, BAD, as presented, is trained using actor-critic methods which are sample inefficient", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 343, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 353 + ], + "score": 1.0, + "content": "and suffer from local optima. In order to get around this, BAD uses population based training, further", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "score": 1.0, + "content": "increasing the number of samples required. Lastly, BAD’s explicit reliance on common knowledge", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 363, + 249, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 249, + 376 + ], + "score": 1.0, + "content": "limits the generality of the method.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "In this paper we propose the Simplified Action Decoder (SAD), a method that achieves a similar goal", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 404 + ], + "score": 1.0, + "content": "to BAD, but addresses all of the issues mentioned above. At the core of SAD is a different approach", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 403, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 414 + ], + "score": 1.0, + "content": "towards resolving the conflict between exploration and being interpretable, which, like BAD, relies", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "on the centralized training with decentralized control (CT/DC) regime. Under CT/DC information", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "can be exchanged freely amongst all agents during centralized training, as long as the final policies", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 435, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 288, + 448 + ], + "score": 1.0, + "content": "are compatible with decentralized execution.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "The key insight is that during training we do not have to chose between being informative, by taking", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 462, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 476 + ], + "score": 1.0, + "content": "greedy actions, and exploring, by taking random actions. To be informative, the greedy actions do", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "not need to be executed by the environment, but only need to be observed by the team mates. Thus", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "in SAD each agent takes two different actions at each time step: One greedy action, which is not", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "presented to the environment but observed by the team mates at the next time step as an additional", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "input, and the “standard” (exploratory) action that gets executed by the environment and is observed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "by the team mates as part of the environment dynamics. Importantly, during greedy execution the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "observed environment action can be used instead of centralized information for the additional input,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 539, + 279, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 279, + 554 + ], + "score": 1.0, + "content": "since now the agent has stopped exploring.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "Furthermore, to ensure that these greedy actions and observations get decoded into a meaningful", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 568, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 580 + ], + "score": 1.0, + "content": "representation, we can optionally train an auxiliary task that predicts key hidden game properties", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "from the action-observation trajectories. While we note that this idea is in principle compatible with", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 589, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 603 + ], + "score": 1.0, + "content": "any kind of model-free deep RL method with minimal modifications to the core algorithm, we use", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "a distributed version of recurrent DQN in order to improve sample efficiency, account for partial", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 611, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 625 + ], + "score": 1.0, + "content": "observability and reduce the risk of local optima. We also train a joint-action Q-function that consists", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "of the sum of per-agent Q-values to allow for off-policy learning in this multi-agent setting using", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 633, + 359, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 359, + 646 + ], + "score": 1.0, + "content": "Value Decomposition Networks (VDN) (Sunehag et al., 2017).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 505, + 717 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "Using SAD we establish a new SOTA for learning methods for 2-5 players in Hanabi, with a method", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 660, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 674 + ], + "score": 1.0, + "content": "that not only requires less expert knowledge and compute, but is also more general than previous", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "score": 1.0, + "content": "approaches. In order to ensure that our results can be easily verified and extended, we also evaluate", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 683, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 506, + 696 + ], + "score": 1.0, + "content": "our method on a proof-of-principle matrix game and open-source our training code and agents.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 694, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 505, + 707 + ], + "score": 1.0, + "content": "Beyond enabling more research into the self-play aspect of Hanabi, we believe these resources will", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 704, + 475, + 719 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 475, + 719 + ], + "score": 1.0, + "content": "provide a much needed starting point for the ad-hoc teamwork part of the Hanabi challenge.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 506, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "Hanabi is a fully cooperative, partially-observable card game that has recently been proposed as a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 134 + ], + "score": 1.0, + "content": "new benchmark challenge problem for AI research (Bard et al., 2019) to fill the gap around ToM. In", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "score": 1.0, + "content": "Hanabi, players need to find conventions that allow them to effectively exchange information from", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "their local observations through their actions, taking advantage of the fact that actions are observed", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 181, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 181, + 167 + ], + "score": 1.0, + "content": "by all team mates.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 110, + 506, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "Most prior state-of-the-art agents for Hanabi were developed using handcrafted algorithms, which", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "beat off-the-shelf deep multi-agent RL methods by a large margin. This makes intuitive sense: Beyond", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "the “standard” multi-agent challenges of credit assignment, nonstationarity and joint exploration,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 205, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 506, + 216 + ], + "score": 1.0, + "content": "learning an informative policy presents an additional fundamentally new conflict. On the one hand,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 506, + 228 + ], + "score": 1.0, + "content": "an RL agent needs to explore in order to discover good policies through trial and error. On the other", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "hand, when carried out naively, this exploration will add noise to the policy of the agent during the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 237, + 436, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 436, + 250 + ], + "score": 1.0, + "content": "training process, making their actions strictly less informative to their team mates.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 171, + 506, + 250 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 254, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "One possible solution to this is to explore in the space of deterministic partial policies, rather than", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "actions, and sample these policies from a distribution that conditions on a common knowledge", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "Bayesian belief. This is successfully carried out in the Bayesian Action Decoder (BAD) (Foerster", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 287, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 506, + 299 + ], + "score": 1.0, + "content": "et al., 2019), the only previous Deep RL method to achieve a state-of-the-art in Hanabi. While this is a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "notable accomplishment, it comes at the cost of simplicity and generality. For a start, BAD requires an", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 309, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 506, + 321 + ], + "score": 1.0, + "content": "explicit common knowledge Bayesian belief to be tracked, which not only adds computational burden", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "score": 1.0, + "content": "due to the required sampling steps, but also uses expert knowledge regarding the game dynamics.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 331, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 343 + ], + "score": 1.0, + "content": "Furthermore, BAD, as presented, is trained using actor-critic methods which are sample inefficient", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 343, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 353 + ], + "score": 1.0, + "content": "and suffer from local optima. In order to get around this, BAD uses population based training, further", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "score": 1.0, + "content": "increasing the number of samples required. Lastly, BAD’s explicit reliance on common knowledge", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 363, + 249, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 249, + 376 + ], + "score": 1.0, + "content": "limits the generality of the method.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 254, + 506, + 376 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "In this paper we propose the Simplified Action Decoder (SAD), a method that achieves a similar goal", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 404 + ], + "score": 1.0, + "content": "to BAD, but addresses all of the issues mentioned above. At the core of SAD is a different approach", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 403, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 414 + ], + "score": 1.0, + "content": "towards resolving the conflict between exploration and being interpretable, which, like BAD, relies", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "on the centralized training with decentralized control (CT/DC) regime. Under CT/DC information", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "can be exchanged freely amongst all agents during centralized training, as long as the final policies", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 435, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 288, + 448 + ], + "score": 1.0, + "content": "are compatible with decentralized execution.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 380, + 505, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "The key insight is that during training we do not have to chose between being informative, by taking", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 462, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 476 + ], + "score": 1.0, + "content": "greedy actions, and exploring, by taking random actions. To be informative, the greedy actions do", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "not need to be executed by the environment, but only need to be observed by the team mates. Thus", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "in SAD each agent takes two different actions at each time step: One greedy action, which is not", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "presented to the environment but observed by the team mates at the next time step as an additional", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "input, and the “standard” (exploratory) action that gets executed by the environment and is observed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "by the team mates as part of the environment dynamics. Importantly, during greedy execution the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "observed environment action can be used instead of centralized information for the additional input,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 539, + 279, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 279, + 554 + ], + "score": 1.0, + "content": "since now the agent has stopped exploring.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 451, + 506, + 554 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "Furthermore, to ensure that these greedy actions and observations get decoded into a meaningful", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 568, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 580 + ], + "score": 1.0, + "content": "representation, we can optionally train an auxiliary task that predicts key hidden game properties", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "from the action-observation trajectories. While we note that this idea is in principle compatible with", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 589, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 603 + ], + "score": 1.0, + "content": "any kind of model-free deep RL method with minimal modifications to the core algorithm, we use", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "a distributed version of recurrent DQN in order to improve sample efficiency, account for partial", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 611, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 625 + ], + "score": 1.0, + "content": "observability and reduce the risk of local optima. We also train a joint-action Q-function that consists", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "of the sum of per-agent Q-values to allow for off-policy learning in this multi-agent setting using", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 633, + 359, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 359, + 646 + ], + "score": 1.0, + "content": "Value Decomposition Networks (VDN) (Sunehag et al., 2017).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 556, + 506, + 646 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 505, + 717 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "Using SAD we establish a new SOTA for learning methods for 2-5 players in Hanabi, with a method", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 660, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 674 + ], + "score": 1.0, + "content": "that not only requires less expert knowledge and compute, but is also more general than previous", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 685 + ], + "score": 1.0, + "content": "approaches. In order to ensure that our results can be easily verified and extended, we also evaluate", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 683, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 506, + 696 + ], + "score": 1.0, + "content": "our method on a proof-of-principle matrix game and open-source our training code and agents.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 694, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 505, + 707 + ], + "score": 1.0, + "content": "Beyond enabling more research into the self-play aspect of Hanabi, we believe these resources will", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 704, + 475, + 719 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 475, + 719 + ], + "score": 1.0, + "content": "provide a much needed starting point for the ad-hoc teamwork part of the Hanabi challenge.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5, + "bbox_fs": [ + 105, + 650, + 506, + 719 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 211, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 213, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 213, + 96 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "Our work relates closely to research on emergent communication protocols using deep multi-agent", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 122, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 506, + 133 + ], + "score": 1.0, + "content": "RL, as first undertaken by Sukhbaatar et al. (2016) and Foerster et al. (2016) . There has been a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "large number of follow-up papers in this area, so listing all relevant work is beyond the scope and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "we refer the reader to Nguyen et al. (2018), a recent survey on deep multi-agent RL. One major", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 153, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 167 + ], + "score": 1.0, + "content": "difference to our work is that the environments considered typically contain a cheap-talk channel,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "which can be modeled as a continuous variable during the course of training. This allows agents", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 177, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 506, + 189 + ], + "score": 1.0, + "content": "to, for example, use differentiation across the communication channel in order to learn protocols.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "In contrast, in our setting agents have to communicate through the observable environment actions", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 329, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 329, + 210 + ], + "score": 1.0, + "content": "themselves, requiring fundamentally different methods.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "Furthermore, our work is an example of cooperative multi-agent learning in partially observable", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "settings under centralized training and decentralized control. There have been a large number of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "papers in this space, with seminal work including MADDPG (Lowe et al., 2017) and COMA (Foerster", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "et al., 2018a), both of which are actor-critic methods that employ a centralized critic with decentralized", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 258, + 474, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 474, + 272 + ], + "score": 1.0, + "content": "actors. Again, we refer the reader to Nguyen et al. (2018) for a more comprehensive survey.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 275, + 505, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "Until 2018, work on Hanabi had been focused on hand-coded methods and heuristics. Some", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 285, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 300 + ], + "score": 1.0, + "content": "relevant examples include SmartBot (O’Dwyer, 2019) and the so-called “hat-coding” strategies, as", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 298, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 310 + ], + "score": 1.0, + "content": "implemented by WTFWThat (Wu, 2018). These strategies use the information theoretic ideas that", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "allow each hint to reveal information to all other agents at the same time. While they do not perform", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "well for 2-player Hanabi due to the smaller action space, they get near perfect scores for 3-5 players.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 348 + ], + "score": 1.0, + "content": "In contrast, so far learning methods have seen limited success on Hanabi. Bard et al. (2019) undertake", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "a systematic evaluation of current Deep RL methods for 2-5 players in two different regimes and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "open-source the Hanabi-Learning-Environment (HLE) to foster research on the game. They evaluate", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "score": 1.0, + "content": "a feed-forward version of DQN trained on 100 million samples and a recurrent actor-critic agent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "score": 1.0, + "content": "with population based training using 20 billion samples. Notably, while both agents achieve near", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 122, + 402 + ], + "score": 0.84, + "content": "0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "win rate for 3-5 players in Hanabim, at a high level their DQN agent is a good starting point", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "for our work. However, since the authors did not propose any specific method of accounting for the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 192, + 426 + ], + "score": 1.0, + "content": "issues introduced by", + "type": "text" + }, + { + "bbox": [ + 193, + 415, + 198, + 423 + ], + "score": 0.62, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 413, + 438, + 426 + ], + "score": 1.0, + "content": "-greedy exploration in a ToM task, they resorted to setting", + "type": "text" + }, + { + "bbox": [ + 438, + 415, + 444, + 424 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "to zero after a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "short burn-in phase. The only state-of-the-art in Hanabi established by an RL agent is from Foerster", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "et al. (2019) which we refer to in more detail in Section 1 and Section 4. Recently there have also", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "been attempts to train agents that are robust to different team-mates (Canaan et al., 2019) and even to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "extend to human-AI collaboration (Liang et al., 2019). For a more comprehensive review on previous", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 469, + 340, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 340, + 479 + ], + "score": 1.0, + "content": "results on Hanabi we refer the reader to Bard et al. (2019).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "Poker is another partially observable multi-agent setting, although it is fundamentally different due to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "the game being zero-sum. Recent success in Poker has extensively benefited from search (Brown", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 386, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 386, + 520 + ], + "score": 1.0, + "content": "et al.). Examples of using search in Hanabi include Goodman (2019).", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 108, + 541, + 200, + 554 + ], + "lines": [ + { + "bbox": [ + 104, + 540, + 201, + 557 + ], + "spans": [ + { + "bbox": [ + 104, + 540, + 201, + 557 + ], + "score": 1.0, + "content": "3 BACKGROUND", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 107, + 570, + 169, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 171, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 171, + 583 + ], + "score": 1.0, + "content": "3.1 SETTING", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 380, + 606 + ], + "score": 1.0, + "content": "In this paper we assume a Dec-POMDP (Oliehoek, 2012), in which", + "type": "text" + }, + { + "bbox": [ + 380, + 595, + 390, + 604 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "agents interact in a partially", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 603, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 308, + 618 + ], + "score": 1.0, + "content": "observable environment. At each time step agent", + "type": "text" + }, + { + "bbox": [ + 308, + 605, + 347, + 615 + ], + "score": 0.89, + "content": "a \\in 1 . . N", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 603, + 444, + 618 + ], + "score": 1.0, + "content": "obtains an observation,", + "type": "text" + }, + { + "bbox": [ + 445, + 605, + 502, + 617 + ], + "score": 0.92, + "content": "o _ { t } ^ { a } = { \\bar { O } } ( s _ { t } , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 603, + 507, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 132, + 628 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 616, + 161, + 627 + ], + "score": 0.92, + "content": "s _ { t } \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 616, + 309, + 628 + ], + "score": 1.0, + "content": "is the Markov state of the system and", + "type": "text" + }, + { + "bbox": [ + 309, + 616, + 344, + 628 + ], + "score": 0.92, + "content": "O ( s _ { t } , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "is the deterministic observation function.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "Since we are interested in ToM, in our setting the observation function includes the last action of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "acting agent, which is observed by all other agents at the next time step. We note that actions are", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "commonly observable not only in board games but also in some real world multi-agent settings, such", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 659, + 202, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 202, + 673 + ], + "score": 1.0, + "content": "as autonomous driving.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "For simplicity, we restrict ourselves to turn based settings, in which at each time step only the acting", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 195, + 702 + ], + "score": 1.0, + "content": "agents takes an action,", + "type": "text" + }, + { + "bbox": [ + 195, + 688, + 207, + 700 + ], + "score": 0.88, + "content": "u _ { t } ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 687, + 349, + 702 + ], + "score": 1.0, + "content": ", which is sampled from their policy,", + "type": "text" + }, + { + "bbox": [ + 349, + 687, + 416, + 700 + ], + "score": 0.92, + "content": "u ^ { a } \\sim \\pi _ { \\theta } ^ { a } ( u ^ { a } | \\tau ^ { a } )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 687, + 506, + 702 + ], + "score": 1.0, + "content": ", while all other agents", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 697, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 208, + 713 + ], + "score": 1.0, + "content": "take a no-op action. Here", + "type": "text" + }, + { + "bbox": [ + 208, + 699, + 219, + 709 + ], + "score": 0.88, + "content": "\\tau ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 697, + 383, + 713 + ], + "score": 1.0, + "content": "is the action-observation history of agent", + "type": "text" + }, + { + "bbox": [ + 384, + 702, + 389, + 709 + ], + "score": 0.59, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 697, + 394, + 713 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 394, + 699, + 502, + 711 + ], + "score": 0.89, + "content": "\\tau ^ { a } = \\{ o _ { 0 } ^ { a } , u _ { 0 } ^ { a } , r _ { 1 } , . . r _ { T } , o _ { T } ^ { a } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 697, + 506, + 713 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 107, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 115, + 720 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 708, + 245, + 723 + ], + "score": 1.0, + "content": "is the length of the episode and", + "type": "text" + }, + { + "bbox": [ + 246, + 710, + 252, + 720 + ], + "score": 0.81, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "are the weights of a function approximator that represents the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "policy, in our case recurrent neural networks, such as LSTMs (Hochreiter & Schmidhuber, 1997).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 211, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 213, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 213, + 96 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "Our work relates closely to research on emergent communication protocols using deep multi-agent", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 122, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 506, + 133 + ], + "score": 1.0, + "content": "RL, as first undertaken by Sukhbaatar et al. (2016) and Foerster et al. (2016) . There has been a", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "large number of follow-up papers in this area, so listing all relevant work is beyond the scope and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "we refer the reader to Nguyen et al. (2018), a recent survey on deep multi-agent RL. One major", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 153, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 167 + ], + "score": 1.0, + "content": "difference to our work is that the environments considered typically contain a cheap-talk channel,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "which can be modeled as a continuous variable during the course of training. This allows agents", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 177, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 506, + 189 + ], + "score": 1.0, + "content": "to, for example, use differentiation across the communication channel in order to learn protocols.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "In contrast, in our setting agents have to communicate through the observable environment actions", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 329, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 329, + 210 + ], + "score": 1.0, + "content": "themselves, requiring fundamentally different methods.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 110, + 506, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "Furthermore, our work is an example of cooperative multi-agent learning in partially observable", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "settings under centralized training and decentralized control. There have been a large number of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 250 + ], + "score": 1.0, + "content": "papers in this space, with seminal work including MADDPG (Lowe et al., 2017) and COMA (Foerster", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "et al., 2018a), both of which are actor-critic methods that employ a centralized critic with decentralized", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 258, + 474, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 474, + 272 + ], + "score": 1.0, + "content": "actors. Again, we refer the reader to Nguyen et al. (2018) for a more comprehensive survey.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 214, + 506, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 275, + 505, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "Until 2018, work on Hanabi had been focused on hand-coded methods and heuristics. Some", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 285, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 300 + ], + "score": 1.0, + "content": "relevant examples include SmartBot (O’Dwyer, 2019) and the so-called “hat-coding” strategies, as", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 298, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 310 + ], + "score": 1.0, + "content": "implemented by WTFWThat (Wu, 2018). These strategies use the information theoretic ideas that", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "allow each hint to reveal information to all other agents at the same time. While they do not perform", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "well for 2-player Hanabi due to the smaller action space, they get near perfect scores for 3-5 players.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 275, + 506, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 348 + ], + "score": 1.0, + "content": "In contrast, so far learning methods have seen limited success on Hanabi. Bard et al. (2019) undertake", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 360 + ], + "score": 1.0, + "content": "a systematic evaluation of current Deep RL methods for 2-5 players in two different regimes and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "open-source the Hanabi-Learning-Environment (HLE) to foster research on the game. They evaluate", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "score": 1.0, + "content": "a feed-forward version of DQN trained on 100 million samples and a recurrent actor-critic agent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "score": 1.0, + "content": "with population based training using 20 billion samples. Notably, while both agents achieve near", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 122, + 402 + ], + "score": 0.84, + "content": "0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "win rate for 3-5 players in Hanabim, at a high level their DQN agent is a good starting point", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "for our work. However, since the authors did not propose any specific method of accounting for the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 192, + 426 + ], + "score": 1.0, + "content": "issues introduced by", + "type": "text" + }, + { + "bbox": [ + 193, + 415, + 198, + 423 + ], + "score": 0.62, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 413, + 438, + 426 + ], + "score": 1.0, + "content": "-greedy exploration in a ToM task, they resorted to setting", + "type": "text" + }, + { + "bbox": [ + 438, + 415, + 444, + 424 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "to zero after a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "short burn-in phase. The only state-of-the-art in Hanabi established by an RL agent is from Foerster", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "et al. (2019) which we refer to in more detail in Section 1 and Section 4. Recently there have also", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "been attempts to train agents that are robust to different team-mates (Canaan et al., 2019) and even to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "extend to human-AI collaboration (Liang et al., 2019). For a more comprehensive review on previous", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 469, + 340, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 340, + 479 + ], + "score": 1.0, + "content": "results on Hanabi we refer the reader to Bard et al. (2019).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 337, + 506, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "Poker is another partially observable multi-agent setting, although it is fundamentally different due to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "the game being zero-sum. Recent success in Poker has extensively benefited from search (Brown", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 386, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 386, + 520 + ], + "score": 1.0, + "content": "et al.). Examples of using search in Hanabi include Goodman (2019).", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 484, + 506, + 520 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 541, + 200, + 554 + ], + "lines": [ + { + "bbox": [ + 104, + 540, + 201, + 557 + ], + "spans": [ + { + "bbox": [ + 104, + 540, + 201, + 557 + ], + "score": 1.0, + "content": "3 BACKGROUND", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 107, + 570, + 169, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 171, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 171, + 583 + ], + "score": 1.0, + "content": "3.1 SETTING", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 380, + 606 + ], + "score": 1.0, + "content": "In this paper we assume a Dec-POMDP (Oliehoek, 2012), in which", + "type": "text" + }, + { + "bbox": [ + 380, + 595, + 390, + 604 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "agents interact in a partially", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 603, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 308, + 618 + ], + "score": 1.0, + "content": "observable environment. At each time step agent", + "type": "text" + }, + { + "bbox": [ + 308, + 605, + 347, + 615 + ], + "score": 0.89, + "content": "a \\in 1 . . N", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 603, + 444, + 618 + ], + "score": 1.0, + "content": "obtains an observation,", + "type": "text" + }, + { + "bbox": [ + 445, + 605, + 502, + 617 + ], + "score": 0.92, + "content": "o _ { t } ^ { a } = { \\bar { O } } ( s _ { t } , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 603, + 507, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 132, + 628 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 616, + 161, + 627 + ], + "score": 0.92, + "content": "s _ { t } \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 616, + 309, + 628 + ], + "score": 1.0, + "content": "is the Markov state of the system and", + "type": "text" + }, + { + "bbox": [ + 309, + 616, + 344, + 628 + ], + "score": 0.92, + "content": "O ( s _ { t } , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "is the deterministic observation function.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "Since we are interested in ToM, in our setting the observation function includes the last action of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "acting agent, which is observed by all other agents at the next time step. We note that actions are", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "commonly observable not only in board games but also in some real world multi-agent settings, such", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 659, + 202, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 202, + 673 + ], + "score": 1.0, + "content": "as autonomous driving.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 594, + 507, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "For simplicity, we restrict ourselves to turn based settings, in which at each time step only the acting", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 195, + 702 + ], + "score": 1.0, + "content": "agents takes an action,", + "type": "text" + }, + { + "bbox": [ + 195, + 688, + 207, + 700 + ], + "score": 0.88, + "content": "u _ { t } ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 687, + 349, + 702 + ], + "score": 1.0, + "content": ", which is sampled from their policy,", + "type": "text" + }, + { + "bbox": [ + 349, + 687, + 416, + 700 + ], + "score": 0.92, + "content": "u ^ { a } \\sim \\pi _ { \\theta } ^ { a } ( u ^ { a } | \\tau ^ { a } )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 687, + 506, + 702 + ], + "score": 1.0, + "content": ", while all other agents", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 697, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 208, + 713 + ], + "score": 1.0, + "content": "take a no-op action. Here", + "type": "text" + }, + { + "bbox": [ + 208, + 699, + 219, + 709 + ], + "score": 0.88, + "content": "\\tau ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 697, + 383, + 713 + ], + "score": 1.0, + "content": "is the action-observation history of agent", + "type": "text" + }, + { + "bbox": [ + 384, + 702, + 389, + 709 + ], + "score": 0.59, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 697, + 394, + 713 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 394, + 699, + 502, + 711 + ], + "score": 0.89, + "content": "\\tau ^ { a } = \\{ o _ { 0 } ^ { a } , u _ { 0 } ^ { a } , r _ { 1 } , . . r _ { T } , o _ { T } ^ { a } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 697, + 506, + 713 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 107, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 115, + 720 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 708, + 245, + 723 + ], + "score": 1.0, + "content": "is the length of the episode and", + "type": "text" + }, + { + "bbox": [ + 246, + 710, + 252, + 720 + ], + "score": 0.81, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "are the weights of a function approximator that represents the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "policy, in our case recurrent neural networks, such as LSTMs (Hochreiter & Schmidhuber, 1997).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 676, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 168, + 96 + ], + "score": 1.0, + "content": "We further use", + "type": "text" + }, + { + "bbox": [ + 168, + 84, + 177, + 93 + ], + "score": 0.83, + "content": "\\tau _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 81, + 332, + 96 + ], + "score": 1.0, + "content": "to describe the state-action sequence,", + "type": "text" + }, + { + "bbox": [ + 332, + 82, + 437, + 95 + ], + "score": 0.92, + "content": "\\tau = \\{ s _ { 0 } , \\mathbf { u _ { 0 } } , r _ { 1 } , . . r _ { T } , s _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 81, + 468, + 96 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 468, + 84, + 479, + 93 + ], + "score": 0.85, + "content": "\\mathbf { u _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 208, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 208, + 106 + ], + "score": 1.0, + "content": "joint action of all agents.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "As is typical in cooperative multi-agent RL, the goal of the agents is to maximize the total expected", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 137, + 135 + ], + "score": 1.0, + "content": "return,", + "type": "text" + }, + { + "bbox": [ + 137, + 122, + 231, + 135 + ], + "score": 0.93, + "content": "J _ { \\theta } = \\mathbb { E } _ { \\tau \\sim P ( \\tau \\mid \\theta ) } R _ { 0 } ( \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 121, + 264, + 135 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 265, + 122, + 291, + 134 + ], + "score": 0.94, + "content": "R _ { 0 } ( \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 121, + 466, + 135 + ], + "score": 1.0, + "content": "is the return of the trajectory (in general", + "type": "text" + }, + { + "bbox": [ + 466, + 121, + 505, + 134 + ], + "score": 0.91, + "content": "R _ { t } \\bar { ( } \\tau ) ~ =", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 507, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 134, + 171, + 149 + ], + "score": 0.92, + "content": "\\sum _ { t ^ { \\prime } \\geq t } \\gamma ^ { t ^ { \\prime } - t } r _ { t ^ { \\prime } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 132, + 188, + 151 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 189, + 137, + 196, + 147 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 132, + 507, + 151 + ], + "score": 1.0, + "content": "is an optional discount factor. We have also assumed that agents are sharing", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 316, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 155, + 160 + ], + "score": 1.0, + "content": "parameters,", + "type": "text" + }, + { + "bbox": [ + 155, + 149, + 161, + 158 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 147, + 316, + 160 + ], + "score": 1.0, + "content": ", as is common in cooperative MARL.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 106, + 180, + 375, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 377, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 377, + 194 + ], + "score": 1.0, + "content": "3.2 DISTRIBUTED RECURRENT DQN AND AUXILIARY TASKS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 446, + 218 + ], + "score": 1.0, + "content": "In Q-learning the agent approximates the expected return for a given state action-pair,", + "type": "text" + }, + { + "bbox": [ + 446, + 207, + 462, + 216 + ], + "score": 0.82, + "content": "s , u", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 204, + 506, + 218 + ], + "score": 1.0, + "content": ", assuming", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 459, + 228 + ], + "score": 1.0, + "content": "that the agent acts greedily with respect to the Q-function for all future time steps,", + "type": "text" + }, + { + "bbox": [ + 460, + 216, + 505, + 227 + ], + "score": 0.89, + "content": "Q ( s , u ) =", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 225, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 180, + 240 + ], + "score": 0.91, + "content": "\\mathbb { E } _ { \\tau \\sim P ( \\tau | s , u ) } R _ { t } ( \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 225, + 209, + 240 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 210, + 227, + 317, + 239 + ], + "score": 0.88, + "content": "\\tau = \\{ s _ { t } , u _ { t } , r _ { t + 1 } , . . . , s _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 225, + 320, + 240 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 320, + 228, + 350, + 238 + ], + "score": 0.85, + "content": "u _ { t } = u", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 225, + 367, + 240 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 227, + 505, + 239 + ], + "score": 0.7, + "content": "u _ { t ^ { \\prime } } = \\arg \\operatorname* { m a x } _ { u ^ { \\prime } } Q ( s _ { t ^ { \\prime } } , u ^ { \\prime } ) , \\forall t ^ { \\prime } >", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 107, + 240, + 111, + 248 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 237, + 264, + 250 + ], + "score": 1.0, + "content": ". A common exploration scheme is", + "type": "text" + }, + { + "bbox": [ + 265, + 240, + 270, + 248 + ], + "score": 0.63, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "-greedy, in which the agent takes a random action with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 154, + 261 + ], + "score": 1.0, + "content": "probability", + "type": "text" + }, + { + "bbox": [ + 154, + 250, + 160, + 258 + ], + "score": 0.71, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "and acts greedily otherwise. Importantly, the Q-function can be trained efficiently", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 259, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 219, + 273 + ], + "score": 1.0, + "content": "using the Bellman equation:", + "type": "text" + }, + { + "bbox": [ + 220, + 259, + 385, + 271 + ], + "score": 0.91, + "content": "Q ( s , u ) = \\mathbb { E } _ { s ^ { \\prime } } [ r _ { t + 1 } + \\gamma \\operatorname* { m a x } _ { u ^ { \\prime } } Q ( s ^ { \\prime } , u ^ { \\prime } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 259, + 505, + 273 + ], + "score": 1.0, + "content": ", where for simplicity we have", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 271, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 506, + 283 + ], + "score": 1.0, + "content": "assumed a deterministic reward. In Deep Q-Learning (DQN) (Mnih et al., 2015) the Q-function is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 281, + 507, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 507, + 295 + ], + "score": 1.0, + "content": "parameterized by a deep neural network and trained with transitions sampled from experience replay.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 386 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 311 + ], + "score": 1.0, + "content": "In our work we also incorporate other best practice components of the last few years, including double-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "DQN (van Hasselt et al., 2015), dueling network architecture (Wang et al., 2015) and prioritized", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "replay (Schaul et al., 2015). We also employ a distributed training architecture similar to the one", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "proposed by Horgan et al. (2018) where a number of different actors with their own exploration", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "rates collect experiences in parallel and feed them into a central replay buffer. Since our setting is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "score": 1.0, + "content": "partially observable the natural choice for the function approximator is a recurrent neural network.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 378 + ], + "score": 1.0, + "content": "A combination of these techniques was first explored by Kapturowski et al. (2019) in single agent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 376, + 284, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 284, + 385 + ], + "score": 1.0, + "content": "environments such as Atari and DMLab-30.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "Another common best-practice in RL are auxiliary tasks Mirowski et al. (2016); Jaderberg et al.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "(2016), in which the agent produces extra output-heads that are trained on supervised tasks and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 414, + 241, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 241, + 426 + ], + "score": 1.0, + "content": "optimized alongside the RL loss.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 447, + 484, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 484, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 484, + 460 + ], + "score": 1.0, + "content": "3.3 CENTRALISED TRAINING, DECENTRALIZED EXECUTION AND JOINT Q-FUNCTIONS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 471, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 485 + ], + "score": 1.0, + "content": "The most straight forward application of Q-learning to multi-agent settings is Independent Q-Learning", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "score": 1.0, + "content": "(IQL) (Tan, 1993) in which each agent keeps an independent estimate of the expected return, treating", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "all other agents as part of the environment. One challenge with IQL is that the exploratory behavior", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 503, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 503, + 505, + 518 + ], + "score": 1.0, + "content": "of other agents is not corrected for via the max operator in the bootstrap. Notably, IQL does typically", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "not take any advantage of centralized training with decentralised control (CT/DC), a paradigm under", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "which information can be exchanged freely amongst agents during the training phase as long as the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 538, + 337, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 337, + 550 + ], + "score": 1.0, + "content": "policies rely only on local observations during execution.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 507, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 507, + 568 + ], + "score": 1.0, + "content": "There are various approaches for learning joint-Q-functions in the CT/DC regime. For example,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 506, + 577 + ], + "score": 1.0, + "content": "Value-Decomposition-Networks (VDN) (Sunehag et al., 2017) represent the joint-Q-function as a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 588 + ], + "score": 1.0, + "content": "sum of per-agent contributions and QMIX (Rashid et al., 2018) learns a non-linear but monotonic", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 250, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 250, + 599 + ], + "score": 1.0, + "content": "combination of these contributions.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 172, + 635 + ], + "lines": [ + { + "bbox": [ + 104, + 621, + 174, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 621, + 174, + 638 + ], + "score": 1.0, + "content": "4 METHOD", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 652, + 331, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 652, + 333, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 333, + 665 + ], + "score": 1.0, + "content": "4.1 THEORY OF MIND AND BAYESIAN REASONING", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "At the very core of interpreting the actions of another agent, and ToM in general, is Bayesian", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "reasoning. Fundamentally, asking what a given action by another agent implies about the state", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "of the world requires understanding of why this action was taken. To illustrate this, we start out", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 431, + 722 + ], + "score": 1.0, + "content": "with an agent that has a given belief about the state-action history of the world,", + "type": "text" + }, + { + "bbox": [ + 432, + 711, + 441, + 721 + ], + "score": 0.81, + "content": "\\tau _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ", given her own", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 308, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 213, + 734 + ], + "score": 1.0, + "content": "action-observation history", + "type": "text" + }, + { + "bbox": [ + 213, + 721, + 224, + 732 + ], + "score": 0.73, + "content": "\\tau _ { t } ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 720, + 226, + 734 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 227, + 720, + 304, + 733 + ], + "score": 0.91, + "content": "B ( \\tau _ { t } ) = P ( \\tau _ { t } | \\tau _ { t } ^ { a } )", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 720, + 308, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 168, + 96 + ], + "score": 1.0, + "content": "We further use", + "type": "text" + }, + { + "bbox": [ + 168, + 84, + 177, + 93 + ], + "score": 0.83, + "content": "\\tau _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 81, + 332, + 96 + ], + "score": 1.0, + "content": "to describe the state-action sequence,", + "type": "text" + }, + { + "bbox": [ + 332, + 82, + 437, + 95 + ], + "score": 0.92, + "content": "\\tau = \\{ s _ { 0 } , \\mathbf { u _ { 0 } } , r _ { 1 } , . . r _ { T } , s _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 81, + 468, + 96 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 468, + 84, + 479, + 93 + ], + "score": 0.85, + "content": "\\mathbf { u _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 208, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 208, + 106 + ], + "score": 1.0, + "content": "joint action of all agents.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "As is typical in cooperative multi-agent RL, the goal of the agents is to maximize the total expected", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 137, + 135 + ], + "score": 1.0, + "content": "return,", + "type": "text" + }, + { + "bbox": [ + 137, + 122, + 231, + 135 + ], + "score": 0.93, + "content": "J _ { \\theta } = \\mathbb { E } _ { \\tau \\sim P ( \\tau \\mid \\theta ) } R _ { 0 } ( \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 121, + 264, + 135 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 265, + 122, + 291, + 134 + ], + "score": 0.94, + "content": "R _ { 0 } ( \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 121, + 466, + 135 + ], + "score": 1.0, + "content": "is the return of the trajectory (in general", + "type": "text" + }, + { + "bbox": [ + 466, + 121, + 505, + 134 + ], + "score": 0.91, + "content": "R _ { t } \\bar { ( } \\tau ) ~ =", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 507, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 134, + 171, + 149 + ], + "score": 0.92, + "content": "\\sum _ { t ^ { \\prime } \\geq t } \\gamma ^ { t ^ { \\prime } - t } r _ { t ^ { \\prime } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 132, + 188, + 151 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 189, + 137, + 196, + 147 + ], + "score": 0.82, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 132, + 507, + 151 + ], + "score": 1.0, + "content": "is an optional discount factor. We have also assumed that agents are sharing", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 316, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 155, + 160 + ], + "score": 1.0, + "content": "parameters,", + "type": "text" + }, + { + "bbox": [ + 155, + 149, + 161, + 158 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 147, + 316, + 160 + ], + "score": 1.0, + "content": ", as is common in cooperative MARL.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 110, + 507, + 160 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 180, + 375, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 377, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 377, + 194 + ], + "score": 1.0, + "content": "3.2 DISTRIBUTED RECURRENT DQN AND AUXILIARY TASKS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 446, + 218 + ], + "score": 1.0, + "content": "In Q-learning the agent approximates the expected return for a given state action-pair,", + "type": "text" + }, + { + "bbox": [ + 446, + 207, + 462, + 216 + ], + "score": 0.82, + "content": "s , u", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 204, + 506, + 218 + ], + "score": 1.0, + "content": ", assuming", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 459, + 228 + ], + "score": 1.0, + "content": "that the agent acts greedily with respect to the Q-function for all future time steps,", + "type": "text" + }, + { + "bbox": [ + 460, + 216, + 505, + 227 + ], + "score": 0.89, + "content": "Q ( s , u ) =", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 225, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 180, + 240 + ], + "score": 0.91, + "content": "\\mathbb { E } _ { \\tau \\sim P ( \\tau | s , u ) } R _ { t } ( \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 225, + 209, + 240 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 210, + 227, + 317, + 239 + ], + "score": 0.88, + "content": "\\tau = \\{ s _ { t } , u _ { t } , r _ { t + 1 } , . . . , s _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 225, + 320, + 240 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 320, + 228, + 350, + 238 + ], + "score": 0.85, + "content": "u _ { t } = u", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 225, + 367, + 240 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 227, + 505, + 239 + ], + "score": 0.7, + "content": "u _ { t ^ { \\prime } } = \\arg \\operatorname* { m a x } _ { u ^ { \\prime } } Q ( s _ { t ^ { \\prime } } , u ^ { \\prime } ) , \\forall t ^ { \\prime } >", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 107, + 240, + 111, + 248 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 237, + 264, + 250 + ], + "score": 1.0, + "content": ". A common exploration scheme is", + "type": "text" + }, + { + "bbox": [ + 265, + 240, + 270, + 248 + ], + "score": 0.63, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "-greedy, in which the agent takes a random action with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 154, + 261 + ], + "score": 1.0, + "content": "probability", + "type": "text" + }, + { + "bbox": [ + 154, + 250, + 160, + 258 + ], + "score": 0.71, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "and acts greedily otherwise. Importantly, the Q-function can be trained efficiently", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 259, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 219, + 273 + ], + "score": 1.0, + "content": "using the Bellman equation:", + "type": "text" + }, + { + "bbox": [ + 220, + 259, + 385, + 271 + ], + "score": 0.91, + "content": "Q ( s , u ) = \\mathbb { E } _ { s ^ { \\prime } } [ r _ { t + 1 } + \\gamma \\operatorname* { m a x } _ { u ^ { \\prime } } Q ( s ^ { \\prime } , u ^ { \\prime } ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 259, + 505, + 273 + ], + "score": 1.0, + "content": ", where for simplicity we have", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 271, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 506, + 283 + ], + "score": 1.0, + "content": "assumed a deterministic reward. In Deep Q-Learning (DQN) (Mnih et al., 2015) the Q-function is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 281, + 507, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 507, + 295 + ], + "score": 1.0, + "content": "parameterized by a deep neural network and trained with transitions sampled from experience replay.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 204, + 507, + 295 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 386 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 311 + ], + "score": 1.0, + "content": "In our work we also incorporate other best practice components of the last few years, including double-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "DQN (van Hasselt et al., 2015), dueling network architecture (Wang et al., 2015) and prioritized", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "replay (Schaul et al., 2015). We also employ a distributed training architecture similar to the one", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "proposed by Horgan et al. (2018) where a number of different actors with their own exploration", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "rates collect experiences in parallel and feed them into a central replay buffer. Since our setting is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 366 + ], + "score": 1.0, + "content": "partially observable the natural choice for the function approximator is a recurrent neural network.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 378 + ], + "score": 1.0, + "content": "A combination of these techniques was first explored by Kapturowski et al. (2019) in single agent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 376, + 284, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 284, + 385 + ], + "score": 1.0, + "content": "environments such as Atari and DMLab-30.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 299, + 506, + 385 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "Another common best-practice in RL are auxiliary tasks Mirowski et al. (2016); Jaderberg et al.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "(2016), in which the agent produces extra output-heads that are trained on supervised tasks and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 414, + 241, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 241, + 426 + ], + "score": 1.0, + "content": "optimized alongside the RL loss.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 391, + 506, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 447, + 484, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 484, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 484, + 460 + ], + "score": 1.0, + "content": "3.3 CENTRALISED TRAINING, DECENTRALIZED EXECUTION AND JOINT Q-FUNCTIONS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 446, + 484, + 460 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 471, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 485 + ], + "score": 1.0, + "content": "The most straight forward application of Q-learning to multi-agent settings is Independent Q-Learning", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "score": 1.0, + "content": "(IQL) (Tan, 1993) in which each agent keeps an independent estimate of the expected return, treating", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "all other agents as part of the environment. One challenge with IQL is that the exploratory behavior", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 503, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 503, + 505, + 518 + ], + "score": 1.0, + "content": "of other agents is not corrected for via the max operator in the bootstrap. Notably, IQL does typically", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "not take any advantage of centralized training with decentralised control (CT/DC), a paradigm under", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "which information can be exchanged freely amongst agents during the training phase as long as the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 538, + 337, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 337, + 550 + ], + "score": 1.0, + "content": "policies rely only on local observations during execution.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 470, + 505, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 507, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 507, + 568 + ], + "score": 1.0, + "content": "There are various approaches for learning joint-Q-functions in the CT/DC regime. For example,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 506, + 577 + ], + "score": 1.0, + "content": "Value-Decomposition-Networks (VDN) (Sunehag et al., 2017) represent the joint-Q-function as a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 588 + ], + "score": 1.0, + "content": "sum of per-agent contributions and QMIX (Rashid et al., 2018) learns a non-linear but monotonic", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 250, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 250, + 599 + ], + "score": 1.0, + "content": "combination of these contributions.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 553, + 507, + 599 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 172, + 635 + ], + "lines": [ + { + "bbox": [ + 104, + 621, + 174, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 621, + 174, + 638 + ], + "score": 1.0, + "content": "4 METHOD", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 652, + 331, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 652, + 333, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 333, + 665 + ], + "score": 1.0, + "content": "4.1 THEORY OF MIND AND BAYESIAN REASONING", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "At the very core of interpreting the actions of another agent, and ToM in general, is Bayesian", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "reasoning. Fundamentally, asking what a given action by another agent implies about the state", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "of the world requires understanding of why this action was taken. To illustrate this, we start out", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 431, + 722 + ], + "score": 1.0, + "content": "with an agent that has a given belief about the state-action history of the world,", + "type": "text" + }, + { + "bbox": [ + 432, + 711, + 441, + 721 + ], + "score": 0.81, + "content": "\\tau _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ", given her own", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 308, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 213, + 734 + ], + "score": 1.0, + "content": "action-observation history", + "type": "text" + }, + { + "bbox": [ + 213, + 721, + 224, + 732 + ], + "score": 0.73, + "content": "\\tau _ { t } ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 720, + 226, + 734 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 227, + 720, + 304, + 733 + ], + "score": 0.91, + "content": "B ( \\tau _ { t } ) = P ( \\tau _ { t } | \\tau _ { t } ^ { a } )", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 720, + 308, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 677, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 485, + 95 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 486, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 245, + 97 + ], + "score": 1.0, + "content": "Next the agent observes the action", + "type": "text" + }, + { + "bbox": [ + 246, + 81, + 259, + 95 + ], + "score": 0.92, + "content": "u _ { t } ^ { a ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 80, + 334, + 97 + ], + "score": 1.0, + "content": "of her team mate,", + "type": "text" + }, + { + "bbox": [ + 334, + 83, + 343, + 92 + ], + "score": 0.84, + "content": "a ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 80, + 486, + 97 + ], + "score": 1.0, + "content": ", and carries out a Bayesian update:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 100, + 401, + 166 + ], + "lines": [ + { + "bbox": [ + 209, + 100, + 401, + 166 + ], + "spans": [ + { + "bbox": [ + 209, + 100, + 401, + 166 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = \\frac { P ( u _ { t } ^ { a ^ { \\prime } } | \\tau _ { t } ) P ( \\tau _ { t } | \\tau _ { t } ^ { a } ) } { \\sum _ { \\tau _ { t } ^ { \\prime } } P ( u _ { t } ^ { a ^ { \\prime } } | \\tau _ { t } ^ { \\prime } ) P ( \\tau _ { t } ^ { \\prime } | \\tau _ { t } ^ { a } ) } } \\\\ & { \\qquad = \\frac { \\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) ) B ( \\tau _ { t } ) } { \\sum _ { \\tau _ { t } ^ { \\prime } } \\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ^ { \\prime } ) ) B ( \\tau _ { t } ^ { \\prime } ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "a2a28603de33f151bf7cd0c953e46f71ea22ac1ecdfe82cad7f1eacf3ca3f110.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 209, + 100, + 401, + 113.2 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 209, + 113.2, + 401, + 126.4 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 209, + 126.4, + 401, + 139.6 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 209, + 139.6, + 401, + 152.79999999999998 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 209, + 152.79999999999998, + 401, + 165.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 176, + 506, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 408, + 189 + ], + "score": 1.0, + "content": "where, with a slight abuse of notation, we have used (and will keep using)", + "type": "text" + }, + { + "bbox": [ + 408, + 176, + 445, + 189 + ], + "score": 0.93, + "content": "O ( a ^ { \\prime } , \\tau _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "for the action-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 187, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 186, + 203 + ], + "score": 1.0, + "content": "observation history,", + "type": "text" + }, + { + "bbox": [ + 186, + 187, + 200, + 201 + ], + "score": 0.9, + "content": "\\tau _ { t } ^ { a ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 188, + 444, + 203 + ], + "score": 1.0, + "content": ", that results from applying the observation function for agent", + "type": "text" + }, + { + "bbox": [ + 444, + 189, + 453, + 200 + ], + "score": 0.84, + "content": "a ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 188, + 464, + 203 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 464, + 191, + 474, + 200 + ], + "score": 0.83, + "content": "\\tau _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 188, + 506, + 203 + ], + "score": 1.0, + "content": "at each", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "time step. Note that for non-deterministic observation functions we would have to marginalize over", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 210, + 152, + 226 + ], + "spans": [ + { + "bbox": [ + 107, + 211, + 147, + 225 + ], + "score": 0.91, + "content": "P ( \\tau _ { t } ^ { a ^ { \\prime } } | \\bar { \\tau _ { t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 210, + 152, + 226 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 229, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 229, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 506, + 243 + ], + "score": 1.0, + "content": "Clearly, since agents have access to the policy of their teammate during centralised training, we", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "could in principle evaluate this explicit Bayesian belief. However, beyond the practical difficulty of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "computing this explicit belief, when it is used as an input to the policy it will lead to prohibitively", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "costly higher order beliefs. The typical workout for this is a public belief over private features which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 273, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 506, + 286 + ], + "score": 1.0, + "content": "only conditions on common knowledge and can therefore be calculated by all agents individually, we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 285, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 479, + 297 + ], + "score": 1.0, + "content": "refer to Moravcík et al. (2017); Nayyar et al. (2013); Foerster et al. (2018b) for more details. ˇ", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 108, + 301, + 503, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "Instead, in this work we rely on RNNs to learn implicit representations of the sufficient statistics", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "over the distribution of the Markov state given the action-observation histories, noting that they are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 324, + 377, + 335 + ], + "spans": [ + { + "bbox": [ + 107, + 324, + 377, + 335 + ], + "score": 1.0, + "content": "unlikely to recover exact beliefs due to the issues mentioned above.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 349, + 253, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 254, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 254, + 361 + ], + "score": 1.0, + "content": "4.2 EXPLORATION AND BELIEFS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 369, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "score": 1.0, + "content": "Next we illustrate the impact of exploration on the beliefs, which we will do in the explicit (exact)", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "case, since it serves as an upper bound on the accuracy of the implicit beliefs. Since we are looking", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "at fully-cooperative settings we assume that the optimal policy of the agent is deterministic and any", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "randomness is due to exploration. Given that we are focused on value based methods we furthermore", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 152, + 426 + ], + "score": 1.0, + "content": "assume an", + "type": "text" + }, + { + "bbox": [ + 153, + 417, + 158, + 424 + ], + "score": 0.57, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "-greedy exploration scheme, noting that the same analysis can be extended to other", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 425, + 386, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 270, + 439 + ], + "score": 1.0, + "content": "methods. Under this exploration scheme", + "type": "text" + }, + { + "bbox": [ + 271, + 425, + 343, + 439 + ], + "score": 0.93, + "content": "\\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 426, + 386, + 439 + ], + "score": 1.0, + "content": ") becomes:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 444, + 411, + 461 + ], + "lines": [ + { + "bbox": [ + 199, + 444, + 411, + 461 + ], + "spans": [ + { + "bbox": [ + 199, + 444, + 411, + 461 + ], + "score": 0.9, + "content": "\\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) ) = ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ,", + "type": "interline_equation", + "image_path": "0a52e16234105ddd8c60fd2ee098383204ea93e1db137247763f8bfdcacb7de2.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 199, + 444, + 411, + 461 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 504, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 206, + 480 + ], + "score": 1.0, + "content": "where we have used", + "type": "text" + }, + { + "bbox": [ + 207, + 466, + 234, + 479 + ], + "score": 0.92, + "content": "\\boldsymbol { u } ^ { * } ( \\tau _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 464, + 440, + 480 + ], + "score": 1.0, + "content": "to indicate the greedy action of the agent", + "type": "text" + }, + { + "bbox": [ + 440, + 466, + 449, + 477 + ], + "score": 0.66, + "content": "a ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 464, + 458, + 480 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 458, + 466, + 505, + 479 + ], + "score": 0.76, + "content": "\\begin{array} { r l } { u ^ { * } ( \\tau _ { t } ) } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 478, + 343, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 122, + 493 + ], + "score": 1.0, + "content": "arg", + "type": "text" + }, + { + "bbox": [ + 123, + 478, + 217, + 492 + ], + "score": 0.87, + "content": "\\operatorname* { m a x } _ { u } Q ^ { a ^ { \\prime } } ( u , O ( a ^ { \\prime } , \\tau _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 478, + 236, + 493 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 480, + 242, + 489 + ], + "score": 0.37, + "content": "\\mathbf { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 478, + 343, + 493 + ], + "score": 1.0, + "content": "is the indicator function.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 496, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "While the first part corresponds to a filtering operator, in which the indicator function only attributes", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "finite probability to those histories that are consistent with the action taken under greedy execution,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "the exploration term adds a fixed (history independent) probability, which effectively ‘blurs’ the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 530, + 147, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 147, + 541 + ], + "score": 1.0, + "content": "posterior:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "interline_equation", + "bbox": [ + 181, + 561, + 428, + 687 + ], + "lines": [ + { + "bbox": [ + 181, + 561, + 428, + 687 + ], + "spans": [ + { + "bbox": [ + 181, + 561, + 428, + 687 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { { P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau _ { t } ) } { \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau ^ { \\prime } ) } } } \\\\ & { } & { = \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau _ { t } ) } { \\epsilon / | U | + \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau ^ { \\prime } ) } } \\\\ & { } & { = \\frac { B ( \\tau _ { t } ) } { 1 + | U | \\sum _ { s ^ { \\prime } } ( ( 1 / \\epsilon - 1 ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) B ( \\tau ^ { \\prime } ) } } \\\\ & { } & { + \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau _ { t } ) } { \\epsilon / | U | + \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau ^ { \\prime } ) } . } \\end{array}", + "type": "interline_equation", + "image_path": "8c13d7bf6f8e98e87f500ef80bef560a22101e5bea382a8ce2a57da50625424b.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 181, + 561, + 428, + 603.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 181, + 603.0, + 428, + 645.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 181, + 645.0, + 428, + 687.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 696, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 106, + 696, + 385, + 709 + ], + "score": 1.0, + "content": "We find that the posterior includes an additional term of the form", + "type": "text" + }, + { + "bbox": [ + 385, + 697, + 410, + 709 + ], + "score": 0.92, + "content": "B ( \\tau _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "which carries over an", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 708, + 506, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 479, + 720 + ], + "score": 1.0, + "content": "unfiltered density over the trajectories from the prior. We further confirm that in the limit of", + "type": "text" + }, + { + "bbox": [ + 479, + 709, + 502, + 718 + ], + "score": 0.89, + "content": "\\epsilon = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 708, + 506, + 720 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 251, + 734 + ], + "score": 1.0, + "content": "the posterior collapses to the prior,", + "type": "text" + }, + { + "bbox": [ + 251, + 719, + 344, + 733 + ], + "score": 0.93, + "content": "P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = B ( \\tau _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 720, + 505, + 734 + ], + "score": 1.0, + "content": ". This can be particularly worrisome in", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 485, + 95 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 486, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 245, + 97 + ], + "score": 1.0, + "content": "Next the agent observes the action", + "type": "text" + }, + { + "bbox": [ + 246, + 81, + 259, + 95 + ], + "score": 0.92, + "content": "u _ { t } ^ { a ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 80, + 334, + 97 + ], + "score": 1.0, + "content": "of her team mate,", + "type": "text" + }, + { + "bbox": [ + 334, + 83, + 343, + 92 + ], + "score": 0.84, + "content": "a ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 80, + 486, + 97 + ], + "score": 1.0, + "content": ", and carries out a Bayesian update:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 104, + 80, + 486, + 97 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 100, + 401, + 166 + ], + "lines": [ + { + "bbox": [ + 209, + 100, + 401, + 166 + ], + "spans": [ + { + "bbox": [ + 209, + 100, + 401, + 166 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = \\frac { P ( u _ { t } ^ { a ^ { \\prime } } | \\tau _ { t } ) P ( \\tau _ { t } | \\tau _ { t } ^ { a } ) } { \\sum _ { \\tau _ { t } ^ { \\prime } } P ( u _ { t } ^ { a ^ { \\prime } } | \\tau _ { t } ^ { \\prime } ) P ( \\tau _ { t } ^ { \\prime } | \\tau _ { t } ^ { a } ) } } \\\\ & { \\qquad = \\frac { \\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) ) B ( \\tau _ { t } ) } { \\sum _ { \\tau _ { t } ^ { \\prime } } \\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ^ { \\prime } ) ) B ( \\tau _ { t } ^ { \\prime } ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "a2a28603de33f151bf7cd0c953e46f71ea22ac1ecdfe82cad7f1eacf3ca3f110.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 209, + 100, + 401, + 113.2 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 209, + 113.2, + 401, + 126.4 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 209, + 126.4, + 401, + 139.6 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 209, + 139.6, + 401, + 152.79999999999998 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 209, + 152.79999999999998, + 401, + 165.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 176, + 506, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 408, + 189 + ], + "score": 1.0, + "content": "where, with a slight abuse of notation, we have used (and will keep using)", + "type": "text" + }, + { + "bbox": [ + 408, + 176, + 445, + 189 + ], + "score": 0.93, + "content": "O ( a ^ { \\prime } , \\tau _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "for the action-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 187, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 186, + 203 + ], + "score": 1.0, + "content": "observation history,", + "type": "text" + }, + { + "bbox": [ + 186, + 187, + 200, + 201 + ], + "score": 0.9, + "content": "\\tau _ { t } ^ { a ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 188, + 444, + 203 + ], + "score": 1.0, + "content": ", that results from applying the observation function for agent", + "type": "text" + }, + { + "bbox": [ + 444, + 189, + 453, + 200 + ], + "score": 0.84, + "content": "a ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 188, + 464, + 203 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 464, + 191, + 474, + 200 + ], + "score": 0.83, + "content": "\\tau _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 188, + 506, + 203 + ], + "score": 1.0, + "content": "at each", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "time step. Note that for non-deterministic observation functions we would have to marginalize over", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 210, + 152, + 226 + ], + "spans": [ + { + "bbox": [ + 107, + 211, + 147, + 225 + ], + "score": 0.91, + "content": "P ( \\tau _ { t } ^ { a ^ { \\prime } } | \\bar { \\tau _ { t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 210, + 152, + 226 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 176, + 506, + 226 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 229, + 505, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 229, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 506, + 243 + ], + "score": 1.0, + "content": "Clearly, since agents have access to the policy of their teammate during centralised training, we", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "could in principle evaluate this explicit Bayesian belief. However, beyond the practical difficulty of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "computing this explicit belief, when it is used as an input to the policy it will lead to prohibitively", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "costly higher order beliefs. The typical workout for this is a public belief over private features which", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 273, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 506, + 286 + ], + "score": 1.0, + "content": "only conditions on common knowledge and can therefore be calculated by all agents individually, we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 285, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 479, + 297 + ], + "score": 1.0, + "content": "refer to Moravcík et al. (2017); Nayyar et al. (2013); Foerster et al. (2018b) for more details. ˇ", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 229, + 506, + 297 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 301, + 503, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "Instead, in this work we rely on RNNs to learn implicit representations of the sufficient statistics", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "over the distribution of the Markov state given the action-observation histories, noting that they are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 324, + 377, + 335 + ], + "spans": [ + { + "bbox": [ + 107, + 324, + 377, + 335 + ], + "score": 1.0, + "content": "unlikely to recover exact beliefs due to the issues mentioned above.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 301, + 505, + 335 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 349, + 253, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 254, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 254, + 361 + ], + "score": 1.0, + "content": "4.2 EXPLORATION AND BELIEFS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 369, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "score": 1.0, + "content": "Next we illustrate the impact of exploration on the beliefs, which we will do in the explicit (exact)", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "case, since it serves as an upper bound on the accuracy of the implicit beliefs. Since we are looking", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "at fully-cooperative settings we assume that the optimal policy of the agent is deterministic and any", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "randomness is due to exploration. Given that we are focused on value based methods we furthermore", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 152, + 426 + ], + "score": 1.0, + "content": "assume an", + "type": "text" + }, + { + "bbox": [ + 153, + 417, + 158, + 424 + ], + "score": 0.57, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "-greedy exploration scheme, noting that the same analysis can be extended to other", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 425, + 386, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 270, + 439 + ], + "score": 1.0, + "content": "methods. Under this exploration scheme", + "type": "text" + }, + { + "bbox": [ + 271, + 425, + 343, + 439 + ], + "score": 0.93, + "content": "\\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 426, + 386, + 439 + ], + "score": 1.0, + "content": ") becomes:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 369, + 506, + 439 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 444, + 411, + 461 + ], + "lines": [ + { + "bbox": [ + 199, + 444, + 411, + 461 + ], + "spans": [ + { + "bbox": [ + 199, + 444, + 411, + 461 + ], + "score": 0.9, + "content": "\\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) ) = ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ,", + "type": "interline_equation", + "image_path": "0a52e16234105ddd8c60fd2ee098383204ea93e1db137247763f8bfdcacb7de2.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 199, + 444, + 411, + 461 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 504, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 206, + 480 + ], + "score": 1.0, + "content": "where we have used", + "type": "text" + }, + { + "bbox": [ + 207, + 466, + 234, + 479 + ], + "score": 0.92, + "content": "\\boldsymbol { u } ^ { * } ( \\tau _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 464, + 440, + 480 + ], + "score": 1.0, + "content": "to indicate the greedy action of the agent", + "type": "text" + }, + { + "bbox": [ + 440, + 466, + 449, + 477 + ], + "score": 0.66, + "content": "a ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 464, + 458, + 480 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 458, + 466, + 505, + 479 + ], + "score": 0.76, + "content": "\\begin{array} { r l } { u ^ { * } ( \\tau _ { t } ) } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 478, + 343, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 122, + 493 + ], + "score": 1.0, + "content": "arg", + "type": "text" + }, + { + "bbox": [ + 123, + 478, + 217, + 492 + ], + "score": 0.87, + "content": "\\operatorname* { m a x } _ { u } Q ^ { a ^ { \\prime } } ( u , O ( a ^ { \\prime } , \\tau _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 478, + 236, + 493 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 480, + 242, + 489 + ], + "score": 0.37, + "content": "\\mathbf { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 478, + 343, + 493 + ], + "score": 1.0, + "content": "is the indicator function.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 464, + 505, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 496, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "While the first part corresponds to a filtering operator, in which the indicator function only attributes", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "finite probability to those histories that are consistent with the action taken under greedy execution,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "the exploration term adds a fixed (history independent) probability, which effectively ‘blurs’ the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 530, + 147, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 147, + 541 + ], + "score": 1.0, + "content": "posterior:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 496, + 506, + 541 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 181, + 561, + 428, + 687 + ], + "lines": [ + { + "bbox": [ + 181, + 561, + 428, + 687 + ], + "spans": [ + { + "bbox": [ + 181, + 561, + 428, + 687 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { { P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau _ { t } ) } { \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau ^ { \\prime } ) } } } \\\\ & { } & { = \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau _ { t } ) } { \\epsilon / | U | + \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau ^ { \\prime } ) } } \\\\ & { } & { = \\frac { B ( \\tau _ { t } ) } { 1 + | U | \\sum _ { s ^ { \\prime } } ( ( 1 / \\epsilon - 1 ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) B ( \\tau ^ { \\prime } ) } } \\\\ & { } & { + \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau _ { t } ) } { \\epsilon / | U | + \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau ^ { \\prime } ) } . } \\end{array}", + "type": "interline_equation", + "image_path": "8c13d7bf6f8e98e87f500ef80bef560a22101e5bea382a8ce2a57da50625424b.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 181, + 561, + 428, + 603.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 181, + 603.0, + 428, + 645.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 181, + 645.0, + 428, + 687.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 696, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 106, + 696, + 385, + 709 + ], + "score": 1.0, + "content": "We find that the posterior includes an additional term of the form", + "type": "text" + }, + { + "bbox": [ + 385, + 697, + 410, + 709 + ], + "score": 0.92, + "content": "B ( \\tau _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "which carries over an", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 708, + 506, + 720 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 479, + 720 + ], + "score": 1.0, + "content": "unfiltered density over the trajectories from the prior. We further confirm that in the limit of", + "type": "text" + }, + { + "bbox": [ + 479, + 709, + 502, + 718 + ], + "score": 0.89, + "content": "\\epsilon = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 708, + 506, + 720 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 251, + 734 + ], + "score": 1.0, + "content": "the posterior collapses to the prior,", + "type": "text" + }, + { + "bbox": [ + 251, + 719, + 344, + 733 + ], + "score": 0.93, + "content": "P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = B ( \\tau _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 720, + 505, + 734 + ], + "score": 1.0, + "content": ". This can be particularly worrisome in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "the context of our training setup, whereby different agents run different, and potentially high, \u000f", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 489, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 489, + 106 + ], + "score": 1.0, + "content": "throughout the course of training. It fundamentally makes the beliefs obtained less informative.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 696, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "the context of our training setup, whereby different agents run different, and potentially high, \u000f", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 489, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 489, + 106 + ], + "score": 1.0, + "content": "throughout the course of training. It fundamentally makes the beliefs obtained less informative.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "While not making the above argument explicitly, the Bayesian Action Decoder (BAD) (Foerster et al.,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 134 + ], + "score": 1.0, + "content": "2019), resolves this issue by shifting exploration to the level of deterministic partial policies, rather", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "than action-level, and tracking an approximate Bayesian belief. As outlined in Section 1 this comes", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "at a huge cost in the complexity of the method, the computation requirements and in the loss of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 209, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 209, + 167 + ], + "score": 1.0, + "content": "generality of the method.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 108, + 180, + 267, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 268, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 268, + 193 + ], + "score": 1.0, + "content": "4.3 SIMPLIFIED ACTION DECODING", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 201, + 505, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "In this paper we take a drastically simpler and different approach towards the issue. We note that the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 421, + 225 + ], + "score": 1.0, + "content": "‘blurring’, which makes decoding of an action challenging, is entirely due to the", + "type": "text" + }, + { + "bbox": [ + 422, + 215, + 427, + 223 + ], + "score": 0.66, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "-greedy exploration", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "term. Furthermore, in order for another agent to do an implicit Bayesian update over an action taken,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "score": 1.0, + "content": "it is not required that this action is executed by the environment. Indeed, if we assume that other", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 256, + 258 + ], + "score": 1.0, + "content": "agents can observe the greedy action,", + "type": "text" + }, + { + "bbox": [ + 256, + 246, + 267, + 256 + ], + "score": 0.86, + "content": "u ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 245, + 506, + 258 + ], + "score": 1.0, + "content": ", at every time step and condition their belief update on this,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 257, + 356, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 202, + 270 + ], + "score": 1.0, + "content": "the terms depending on", + "type": "text" + }, + { + "bbox": [ + 202, + 259, + 208, + 267 + ], + "score": 0.68, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 257, + 356, + 270 + ], + "score": 1.0, + "content": "disappear from the Bayesian update:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 275, + 392, + 305 + ], + "lines": [ + { + "bbox": [ + 219, + 275, + 392, + 305 + ], + "spans": [ + { + "bbox": [ + 219, + 275, + 392, + 305 + ], + "score": 0.94, + "content": "P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u ^ { * } ) = \\frac { \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u ^ { * } ) \\big ) B ( \\tau _ { t } ) } { \\sum _ { \\tau ^ { \\prime } } \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u ^ { * } ) \\big ) B ( \\tau ^ { \\prime } ) }", + "type": "interline_equation", + "image_path": "62bf5170694637a8ab7015c0a7a0e4aeece9279c0b2d28977289192fc59283c5.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 275, + 392, + 290.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 219, + 290.0, + 392, + 305.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "Therefore, to have our cake and eat it, in the Simplified Action Decoder (SAD) the acting agent", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 327, + 506, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 464, + 341 + ], + "score": 1.0, + "content": "is allowed to ‘take’ two actions at any given time step during training. The first action,", + "type": "text" + }, + { + "bbox": [ + 465, + 329, + 476, + 338 + ], + "score": 0.85, + "content": "u ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 327, + 506, + 341 + ], + "score": 1.0, + "content": ", is the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "standard environment action, which gets executed as usual and is observed by all agents through the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 350, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 465, + 361 + ], + "score": 1.0, + "content": "observation function at the next time step, as mentioned in Section 3. The second action,", + "type": "text" + }, + { + "bbox": [ + 466, + 350, + 477, + 360 + ], + "score": 0.85, + "content": "u ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 350, + 505, + 361 + ], + "score": 1.0, + "content": ", is the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "score": 1.0, + "content": "greedy action of the active agent. This action does not get executed by the environment but instead", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "is presented as an additional input to the other agents at the next time step, taking advantage of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 382, + 422, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 422, + 396 + ], + "score": 1.0, + "content": "centralized training regime during which information can be exchanged freely.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 108, + 399, + 503, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "Clearly we are not allowed to pass around extra information during decentralized control, but luckily", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 235, + 423 + ], + "score": 1.0, + "content": "this is not needed. Since we set", + "type": "text" + }, + { + "bbox": [ + 235, + 412, + 241, + 420 + ], + "score": 0.71, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "to 0 at test time we can simply use the, now greedy, environment", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 421, + 400, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 400, + 434 + ], + "score": 1.0, + "content": "action obtained from the observation function as our greedy-action input.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "While this is most straight forward in settings where the last action is observed by other agents", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "score": 1.0, + "content": "directly, in principle SAD can also be extended to settings where it is indirectly observed by all agents", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "through the environment dynamics. In these cases we can replace the greedy-action side-channel", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "with a learned inverse model that recovers the action from the observation history during execution.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 487, + 504, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Furthermore, to encourage the agent to meaningfully decode the information contained in the greedy", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "action, we can optionally add an auxiliary task to the training process, such as predicting unobserved", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 509, + 283, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 283, + 523 + ], + "score": 1.0, + "content": "information from their observation history .", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 506, + 539 + ], + "score": 1.0, + "content": "While this idea is compatible with any deep RL algorithm with minimal modifications, we use a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "recurrent version of DQN with distributed training, dueling networks and prioritized replay. We", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "score": 1.0, + "content": "also learn a joint Q-function using VDN in order to address the challenges of multi-agent off-policy", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 560, + 403, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 403, + 572 + ], + "score": 1.0, + "content": "learning, please see Section 3 for details on all of these standard methods.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 588, + 200, + 601 + ], + "lines": [ + { + "bbox": [ + 104, + 587, + 202, + 603 + ], + "spans": [ + { + "bbox": [ + 104, + 587, + 202, + 603 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 107, + 614, + 197, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 198, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 198, + 627 + ], + "score": 1.0, + "content": "5.1 MATRIX GAME", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 507, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 507, + 648 + ], + "score": 1.0, + "content": "We first verify the effectiveness of SAD in the two step, two player matrix game from Foerster et al.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "score": 1.0, + "content": "(2019), which replicates the communication through action challenge of Hanabi in a highly simplified", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "setting. In this fully cooperative game each player obtains a privately observed ‘card’, which is drawn", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 668, + 214, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 214, + 682 + ], + "score": 1.0, + "content": "iid from two options (1,2).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 685, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 106, + 686, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 506, + 698 + ], + "score": 1.0, + "content": "After observing her card, the first player takes one of three possible discrete actions (1, 2, 3). Crucially,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 696, + 505, + 708 + ], + "spans": [ + { + "bbox": [ + 106, + 696, + 505, + 708 + ], + "score": 1.0, + "content": "the second player observes both her own private card and the team mate’s action before acting herself,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 707, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 505, + 720 + ], + "score": 1.0, + "content": "which establishes the opportunity to communicate. The payout is a function of both the two private", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 719, + 385, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 385, + 730 + ], + "score": 1.0, + "content": "cards and the two actions taken by both agents, as shown in Figure 1.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "While not making the above argument explicitly, the Bayesian Action Decoder (BAD) (Foerster et al.,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 134 + ], + "score": 1.0, + "content": "2019), resolves this issue by shifting exploration to the level of deterministic partial policies, rather", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "than action-level, and tracking an approximate Bayesian belief. As outlined in Section 1 this comes", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "at a huge cost in the complexity of the method, the computation requirements and in the loss of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 209, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 209, + 167 + ], + "score": 1.0, + "content": "generality of the method.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 110, + 506, + 167 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 180, + 267, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 268, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 268, + 193 + ], + "score": 1.0, + "content": "4.3 SIMPLIFIED ACTION DECODING", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 201, + 505, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "In this paper we take a drastically simpler and different approach towards the issue. We note that the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 104, + 212, + 421, + 225 + ], + "score": 1.0, + "content": "‘blurring’, which makes decoding of an action challenging, is entirely due to the", + "type": "text" + }, + { + "bbox": [ + 422, + 215, + 427, + 223 + ], + "score": 0.66, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "-greedy exploration", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "term. Furthermore, in order for another agent to do an implicit Bayesian update over an action taken,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "score": 1.0, + "content": "it is not required that this action is executed by the environment. Indeed, if we assume that other", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 256, + 258 + ], + "score": 1.0, + "content": "agents can observe the greedy action,", + "type": "text" + }, + { + "bbox": [ + 256, + 246, + 267, + 256 + ], + "score": 0.86, + "content": "u ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 245, + 506, + 258 + ], + "score": 1.0, + "content": ", at every time step and condition their belief update on this,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 257, + 356, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 202, + 270 + ], + "score": 1.0, + "content": "the terms depending on", + "type": "text" + }, + { + "bbox": [ + 202, + 259, + 208, + 267 + ], + "score": 0.68, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 257, + 356, + 270 + ], + "score": 1.0, + "content": "disappear from the Bayesian update:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 201, + 506, + 270 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 275, + 392, + 305 + ], + "lines": [ + { + "bbox": [ + 219, + 275, + 392, + 305 + ], + "spans": [ + { + "bbox": [ + 219, + 275, + 392, + 305 + ], + "score": 0.94, + "content": "P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u ^ { * } ) = \\frac { \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u ^ { * } ) \\big ) B ( \\tau _ { t } ) } { \\sum _ { \\tau ^ { \\prime } } \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u ^ { * } ) \\big ) B ( \\tau ^ { \\prime } ) }", + "type": "interline_equation", + "image_path": "62bf5170694637a8ab7015c0a7a0e4aeece9279c0b2d28977289192fc59283c5.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 275, + 392, + 290.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 219, + 290.0, + 392, + 305.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "Therefore, to have our cake and eat it, in the Simplified Action Decoder (SAD) the acting agent", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 327, + 506, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 464, + 341 + ], + "score": 1.0, + "content": "is allowed to ‘take’ two actions at any given time step during training. The first action,", + "type": "text" + }, + { + "bbox": [ + 465, + 329, + 476, + 338 + ], + "score": 0.85, + "content": "u ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 327, + 506, + 341 + ], + "score": 1.0, + "content": ", is the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "standard environment action, which gets executed as usual and is observed by all agents through the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 350, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 465, + 361 + ], + "score": 1.0, + "content": "observation function at the next time step, as mentioned in Section 3. The second action,", + "type": "text" + }, + { + "bbox": [ + 466, + 350, + 477, + 360 + ], + "score": 0.85, + "content": "u ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 350, + 505, + 361 + ], + "score": 1.0, + "content": ", is the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "score": 1.0, + "content": "greedy action of the active agent. This action does not get executed by the environment but instead", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "is presented as an additional input to the other agents at the next time step, taking advantage of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 382, + 422, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 422, + 396 + ], + "score": 1.0, + "content": "centralized training regime during which information can be exchanged freely.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 315, + 506, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 399, + 503, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "Clearly we are not allowed to pass around extra information during decentralized control, but luckily", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 235, + 423 + ], + "score": 1.0, + "content": "this is not needed. Since we set", + "type": "text" + }, + { + "bbox": [ + 235, + 412, + 241, + 420 + ], + "score": 0.71, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "to 0 at test time we can simply use the, now greedy, environment", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 421, + 400, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 400, + 434 + ], + "score": 1.0, + "content": "action obtained from the observation function as our greedy-action input.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 398, + 505, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "While this is most straight forward in settings where the last action is observed by other agents", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 461 + ], + "score": 1.0, + "content": "directly, in principle SAD can also be extended to settings where it is indirectly observed by all agents", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "through the environment dynamics. In these cases we can replace the greedy-action side-channel", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "with a learned inverse model that recovers the action from the observation history during execution.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 437, + 505, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 487, + 504, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Furthermore, to encourage the agent to meaningfully decode the information contained in the greedy", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "action, we can optionally add an auxiliary task to the training process, such as predicting unobserved", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 509, + 283, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 283, + 523 + ], + "score": 1.0, + "content": "information from their observation history .", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 487, + 505, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 506, + 539 + ], + "score": 1.0, + "content": "While this idea is compatible with any deep RL algorithm with minimal modifications, we use a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "recurrent version of DQN with distributed training, dueling networks and prioritized replay. We", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 562 + ], + "score": 1.0, + "content": "also learn a joint Q-function using VDN in order to address the challenges of multi-agent off-policy", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 560, + 403, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 403, + 572 + ], + "score": 1.0, + "content": "learning, please see Section 3 for details on all of these standard methods.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 525, + 506, + 572 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 588, + 200, + 601 + ], + "lines": [ + { + "bbox": [ + 104, + 587, + 202, + 603 + ], + "spans": [ + { + "bbox": [ + 104, + 587, + 202, + 603 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 107, + 614, + 197, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 198, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 198, + 627 + ], + "score": 1.0, + "content": "5.1 MATRIX GAME", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 505, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 507, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 507, + 648 + ], + "score": 1.0, + "content": "We first verify the effectiveness of SAD in the two step, two player matrix game from Foerster et al.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "score": 1.0, + "content": "(2019), which replicates the communication through action challenge of Hanabi in a highly simplified", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "setting. In this fully cooperative game each player obtains a privately observed ‘card’, which is drawn", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 668, + 214, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 214, + 682 + ], + "score": 1.0, + "content": "iid from two options (1,2).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 635, + 507, + 682 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 685, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 106, + 686, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 506, + 698 + ], + "score": 1.0, + "content": "After observing her card, the first player takes one of three possible discrete actions (1, 2, 3). Crucially,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 696, + 505, + 708 + ], + "spans": [ + { + "bbox": [ + 106, + 696, + 505, + 708 + ], + "score": 1.0, + "content": "the second player observes both her own private card and the team mate’s action before acting herself,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 707, + 505, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 505, + 720 + ], + "score": 1.0, + "content": "which establishes the opportunity to communicate. The payout is a function of both the two private", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 719, + 385, + 730 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 385, + 730 + ], + "score": 1.0, + "content": "cards and the two actions taken by both agents, as shown in Figure 1.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 686, + 506, + 730 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 336, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 337, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 337, + 95 + ], + "score": 1.0, + "content": "Importantly, there are some obvious strategies that do not", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 337, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 337, + 106 + ], + "score": 1.0, + "content": "“Tiny Hanabi”require any communication. For example, if both player", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 337, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 337, + 116 + ], + "score": 1.0, + "content": "learn to play the 2nd action, the payout is always 8 points,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 336, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 336, + 127 + ], + "score": 1.0, + "content": "independent of the cards dealt. However, if the players do", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 337, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 337, + 138 + ], + "score": 1.0, + "content": "● Each player knows their learn to communicate it is possible to achieve 10 points", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 222, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 222, + 149 + ], + "score": 1.0, + "content": "for every pair of cards dealt.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 162, + 167, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 161, + 169, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 169, + 175 + ], + "score": 1.0, + "content": "5.2 HANABI", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 336, + 281 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 338, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 338, + 195 + ], + "score": 1.0, + "content": "● Player 2 observes Player Hanabi is a fully cooperative card game in which all play-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 194, + 337, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 337, + 205 + ], + "score": 1.0, + "content": "1’s action and acts ers work together to complete piles of cards referred to as", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 337, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 337, + 216 + ], + "score": 1.0, + "content": "second.fireworks. Each card has a rank, 1 to 5, and a color, G / B", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 216, + 337, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 153, + 226 + ], + "score": 0.5, + "content": "/ \\textbf { W } / \\textbf { Y } / \\textbf { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 216, + 337, + 227 + ], + "score": 1.0, + "content": ". Each firework (one per color) starts with a 1", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 337, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 337, + 238 + ], + "score": 1.0, + "content": "and is finished once the 5 has been added. There are three", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 237, + 338, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 338, + 249 + ], + "score": 1.0, + "content": "1s, one 5 and two of all other ranks for each of the colors,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 248, + 337, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 337, + 260 + ], + "score": 1.0, + "content": "adding up to a total of 50 cards in the deck. The twist in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 337, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 337, + 271 + ], + "score": 1.0, + "content": "Hanabi is that while players can observe the cards held by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 270, + 337, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 337, + 282 + ], + "score": 1.0, + "content": "their team mates, they cannot observe their own cards and", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 345, + 87, + 501, + 243 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 345, + 87, + 501, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 345, + 87, + 501, + 243 + ], + "spans": [ + { + "bbox": [ + 345, + 87, + 501, + 243 + ], + "score": 0.963, + "type": "image", + "image_path": "25f404762b879bc27cf6c88933eae05b823c51b2b16509fef1e770d1b412b811.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 345, + 87, + 501, + 100.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 345, + 100.0, + 501, + 113.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 345, + 113.0, + 501, + 126.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 345, + 126.0, + 501, + 139.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 345, + 139.0, + 501, + 152.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 345, + 152.0, + 501, + 165.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 345, + 165.0, + 501, + 178.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 345, + 178.0, + 501, + 191.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 345, + 191.0, + 501, + 204.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 345, + 204.0, + 501, + 217.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 345, + 217.0, + 501, + 230.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 345, + 230.0, + 501, + 243.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 343, + 258, + 504, + 278 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 343, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 343, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "Figure 1: Illustration of the matrix game", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 344, + 268, + 441, + 278 + ], + "spans": [ + { + "bbox": [ + 344, + 268, + 441, + 278 + ], + "score": 1.0, + "content": "from Foerster et al. (2019)", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + } + ], + "index": 25.0 + }, + { + "type": "text", + "bbox": [ + 106, + 282, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "thus need to exchange information with each other in order to understand what cards can be played.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 291, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 305 + ], + "score": 1.0, + "content": "There are two main means for doing so: First of all, players can take grounded hint actions, in which", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "score": 1.0, + "content": "they reveal the subset of a team mate’s hand that matches a specific rank or color. An example hint is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 104, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "“Your third and fifth card are 1s”. These hint actions cost scarce information tokens, which can be", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 326, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 337 + ], + "score": 1.0, + "content": "replenished by discarding a card, an action that both removes the card from the game and makes it", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 335, + 192, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 192, + 349 + ], + "score": 1.0, + "content": "visible to all players.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 353, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "Finally players can also choose to play a card. If this card is the next card for the firework of the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 364, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 376 + ], + "score": 1.0, + "content": "corresponding color, it is added to the firework and the team scores one point. Otherwise the card is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "removed from the game, the identity is made public, and the team loses one of the 3 life tokens. If the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 387, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 504, + 398 + ], + "score": 1.0, + "content": "team runs out of life tokens before the end of the game, all points collected so far are lost and the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 397, + 507, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 380, + 411 + ], + "score": 1.0, + "content": "game finishes immediately. These rules result in a maximum score of", + "type": "text" + }, + { + "bbox": [ + 380, + 397, + 426, + 408 + ], + "score": 0.91, + "content": "5 \\times 5 = 2 5", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 397, + 507, + 411 + ], + "score": 1.0, + "content": "points in any game,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 408, + 446, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 446, + 421 + ], + "score": 1.0, + "content": "which corresponds to all five fireworks being completed with five cards per firework.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 505, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "To ensure reproducibility and comparability of our results we use the Hanabi Learning Environment", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "(HLE) (Bard et al., 2019) for all experimentation. For further details regarding Hanabi and the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 446, + 377, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 377, + 460 + ], + "score": 1.0, + "content": "self-play part of the Hanabi challenge please see Bard et al. (2019).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "title", + "bbox": [ + 107, + 472, + 357, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 359, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 359, + 485 + ], + "score": 1.0, + "content": "5.3 ARCHITECTURE AND COMPUTATION REQUIREMENTS", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "We borrow some ideas and insights from prior distributed Q-learning methods while bring extensions", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "score": 1.0, + "content": "to MARL as well as innovations to improve throughput and efficiency. Following Horgan et al.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 514, + 504, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 493, + 527 + ], + "score": 1.0, + "content": "(2018) and Kapturowski et al. (2019), we use a distributed prioritized replay buffer shared by", + "type": "text" + }, + { + "bbox": [ + 494, + 515, + 504, + 524 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 526, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 538 + ], + "score": 1.0, + "content": "asynchronous actors and a centralized trainer that samples mini-batches from the replay buffer", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 318, + 549 + ], + "score": 1.0, + "content": "to update the model. In each actor thread, we run", + "type": "text" + }, + { + "bbox": [ + 318, + 537, + 329, + 546 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "environments sequentially and batch their", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "observations together. The observation batch is then fed into an actor that utilize a GPU to compute", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "a batch of actions. All asynchronous actors share one GPU and the trainer uses another GPU for", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "gradient computation and model updates. This is different from prior works which run single actor", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "score": 1.0, + "content": "and single environment in each thread on a CPU. Our method enables us to run a very large number", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 591, + 504, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 469, + 604 + ], + "score": 1.0, + "content": "of simulations with moderate computation resources. In all Hanabi experiments, we run", + "type": "text" + }, + { + "bbox": [ + 469, + 591, + 504, + 601 + ], + "score": 0.89, + "content": "N = 8 0", + "type": "inline_equation" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 181, + 614 + ], + "score": 1.0, + "content": "actor threads with", + "type": "text" + }, + { + "bbox": [ + 182, + 602, + 216, + 612 + ], + "score": 0.9, + "content": "K = 8 0", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 603, + 505, + 614 + ], + "score": 1.0, + "content": "environments in each thread on single machine with 40 CPU cores and", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "2 GPUs. Without this architectural improvement, it may require at least a few hundred CPU cores", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 624, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 504, + 636 + ], + "score": 1.0, + "content": "to run 6400 Hanabi environments, in which case neural network agents and simulations have to be", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "distributed across multiple machines, greatly reducing the reproducibility and accessibility of such", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 646, + 456, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 456, + 659 + ], + "score": 1.0, + "content": "research. Please refer to Appendix A for implementation details and hyper-parameters.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 53 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 336, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 337, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 337, + 95 + ], + "score": 1.0, + "content": "Importantly, there are some obvious strategies that do not", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 337, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 337, + 106 + ], + "score": 1.0, + "content": "“Tiny Hanabi”require any communication. For example, if both player", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 337, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 337, + 116 + ], + "score": 1.0, + "content": "learn to play the 2nd action, the payout is always 8 points,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 336, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 336, + 127 + ], + "score": 1.0, + "content": "independent of the cards dealt. However, if the players do", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 337, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 337, + 138 + ], + "score": 1.0, + "content": "● Each player knows their learn to communicate it is possible to achieve 10 points", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 222, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 222, + 149 + ], + "score": 1.0, + "content": "for every pair of cards dealt.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 82, + 337, + 149 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 162, + 167, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 161, + 169, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 169, + 175 + ], + "score": 1.0, + "content": "5.2 HANABI", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 336, + 281 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 338, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 338, + 195 + ], + "score": 1.0, + "content": "● Player 2 observes Player Hanabi is a fully cooperative card game in which all play-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 194, + 337, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 337, + 205 + ], + "score": 1.0, + "content": "1’s action and acts ers work together to complete piles of cards referred to as", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 205, + 337, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 337, + 216 + ], + "score": 1.0, + "content": "second.fireworks. Each card has a rank, 1 to 5, and a color, G / B", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 216, + 337, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 153, + 226 + ], + "score": 0.5, + "content": "/ \\textbf { W } / \\textbf { Y } / \\textbf { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 216, + 337, + 227 + ], + "score": 1.0, + "content": ". Each firework (one per color) starts with a 1", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 337, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 337, + 238 + ], + "score": 1.0, + "content": "and is finished once the 5 has been added. There are three", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 237, + 338, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 338, + 249 + ], + "score": 1.0, + "content": "1s, one 5 and two of all other ranks for each of the colors,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 248, + 337, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 337, + 260 + ], + "score": 1.0, + "content": "adding up to a total of 50 cards in the deck. The twist in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 337, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 337, + 271 + ], + "score": 1.0, + "content": "Hanabi is that while players can observe the cards held by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 270, + 337, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 337, + 282 + ], + "score": 1.0, + "content": "their team mates, they cannot observe their own cards and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "thus need to exchange information with each other in order to understand what cards can be played.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 291, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 305 + ], + "score": 1.0, + "content": "There are two main means for doing so: First of all, players can take grounded hint actions, in which", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 315 + ], + "score": 1.0, + "content": "they reveal the subset of a team mate’s hand that matches a specific rank or color. An example hint is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 104, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "“Your third and fifth card are 1s”. These hint actions cost scarce information tokens, which can be", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 326, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 337 + ], + "score": 1.0, + "content": "replenished by discarding a card, an action that both removes the card from the game and makes it", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 335, + 192, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 192, + 349 + ], + "score": 1.0, + "content": "visible to all players.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 182, + 338, + 282 + ] + }, + { + "type": "image", + "bbox": [ + 345, + 87, + 501, + 243 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 345, + 87, + 501, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 345, + 87, + 501, + 243 + ], + "spans": [ + { + "bbox": [ + 345, + 87, + 501, + 243 + ], + "score": 0.963, + "type": "image", + "image_path": "25f404762b879bc27cf6c88933eae05b823c51b2b16509fef1e770d1b412b811.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 345, + 87, + 501, + 100.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 345, + 100.0, + 501, + 113.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 345, + 113.0, + 501, + 126.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 345, + 126.0, + 501, + 139.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 345, + 139.0, + 501, + 152.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 345, + 152.0, + 501, + 165.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 345, + 165.0, + 501, + 178.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 345, + 178.0, + 501, + 191.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 345, + 191.0, + 501, + 204.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 345, + 204.0, + 501, + 217.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 345, + 217.0, + 501, + 230.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 345, + 230.0, + 501, + 243.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 343, + 258, + 504, + 278 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 343, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 343, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "Figure 1: Illustration of the matrix game", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 344, + 268, + 441, + 278 + ], + "spans": [ + { + "bbox": [ + 344, + 268, + 441, + 278 + ], + "score": 1.0, + "content": "from Foerster et al. (2019)", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + } + ], + "index": 25.0 + }, + { + "type": "text", + "bbox": [ + 106, + 282, + 505, + 348 + ], + "lines": [], + "index": 32.5, + "bbox_fs": [ + 104, + 281, + 506, + 349 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 353, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "Finally players can also choose to play a card. If this card is the next card for the firework of the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 364, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 376 + ], + "score": 1.0, + "content": "corresponding color, it is added to the firework and the team scores one point. Otherwise the card is", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "removed from the game, the identity is made public, and the team loses one of the 3 life tokens. If the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 387, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 504, + 398 + ], + "score": 1.0, + "content": "team runs out of life tokens before the end of the game, all points collected so far are lost and the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 397, + 507, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 380, + 411 + ], + "score": 1.0, + "content": "game finishes immediately. These rules result in a maximum score of", + "type": "text" + }, + { + "bbox": [ + 380, + 397, + 426, + 408 + ], + "score": 0.91, + "content": "5 \\times 5 = 2 5", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 397, + 507, + 411 + ], + "score": 1.0, + "content": "points in any game,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 408, + 446, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 446, + 421 + ], + "score": 1.0, + "content": "which corresponds to all five fireworks being completed with five cards per firework.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 353, + 507, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 505, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "To ensure reproducibility and comparability of our results we use the Hanabi Learning Environment", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "(HLE) (Bard et al., 2019) for all experimentation. For further details regarding Hanabi and the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 446, + 377, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 377, + 460 + ], + "score": 1.0, + "content": "self-play part of the Hanabi challenge please see Bard et al. (2019).", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 424, + 505, + 460 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 472, + 357, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 359, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 359, + 485 + ], + "score": 1.0, + "content": "5.3 ARCHITECTURE AND COMPUTATION REQUIREMENTS", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "We borrow some ideas and insights from prior distributed Q-learning methods while bring extensions", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "score": 1.0, + "content": "to MARL as well as innovations to improve throughput and efficiency. Following Horgan et al.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 514, + 504, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 493, + 527 + ], + "score": 1.0, + "content": "(2018) and Kapturowski et al. (2019), we use a distributed prioritized replay buffer shared by", + "type": "text" + }, + { + "bbox": [ + 494, + 515, + 504, + 524 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 526, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 506, + 538 + ], + "score": 1.0, + "content": "asynchronous actors and a centralized trainer that samples mini-batches from the replay buffer", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 318, + 549 + ], + "score": 1.0, + "content": "to update the model. In each actor thread, we run", + "type": "text" + }, + { + "bbox": [ + 318, + 537, + 329, + 546 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "environments sequentially and batch their", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "observations together. The observation batch is then fed into an actor that utilize a GPU to compute", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "a batch of actions. All asynchronous actors share one GPU and the trainer uses another GPU for", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "gradient computation and model updates. This is different from prior works which run single actor", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "score": 1.0, + "content": "and single environment in each thread on a CPU. Our method enables us to run a very large number", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 591, + 504, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 469, + 604 + ], + "score": 1.0, + "content": "of simulations with moderate computation resources. In all Hanabi experiments, we run", + "type": "text" + }, + { + "bbox": [ + 469, + 591, + 504, + 601 + ], + "score": 0.89, + "content": "N = 8 0", + "type": "inline_equation" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 181, + 614 + ], + "score": 1.0, + "content": "actor threads with", + "type": "text" + }, + { + "bbox": [ + 182, + 602, + 216, + 612 + ], + "score": 0.9, + "content": "K = 8 0", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 603, + 505, + 614 + ], + "score": 1.0, + "content": "environments in each thread on single machine with 40 CPU cores and", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "2 GPUs. Without this architectural improvement, it may require at least a few hundred CPU cores", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 624, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 504, + 636 + ], + "score": 1.0, + "content": "to run 6400 Hanabi environments, in which case neural network agents and simulations have to be", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "distributed across multiple machines, greatly reducing the reproducibility and accessibility of such", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 646, + 456, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 456, + 659 + ], + "score": 1.0, + "content": "research. Please refer to Appendix A for implementation details and hyper-parameters.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 53, + "bbox_fs": [ + 105, + 492, + 506, + 659 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 172, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 174, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 174, + 96 + ], + "score": 1.0, + "content": "6 RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 106, + 196, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 198, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 198, + 120 + ], + "score": 1.0, + "content": "6.1 MATRIX GAME", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 128, + 317, + 238 + ], + "lines": [ + { + "bbox": [ + 106, + 128, + 318, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 318, + 140 + ], + "score": 1.0, + "content": "As we can see in Figure 2, even in our simple ma-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 139, + 317, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 317, + 150 + ], + "score": 1.0, + "content": "trix game the greedy action input makes a drastic", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 150, + 317, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 317, + 162 + ], + "score": 1.0, + "content": "difference. With an average reward of around 9.5", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 160, + 318, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 318, + 174 + ], + "score": 1.0, + "content": "points, tabular IQL does well in this task, matching", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 319, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 121, + 183 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 122, + 172, + 143, + 182 + ], + "score": 0.52, + "content": "B A D", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 171, + 319, + 183 + ], + "score": 1.0, + "content": "results from Foerster et al. (2019). How-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 318, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 318, + 194 + ], + "score": 1.0, + "content": "ever, just by adding the greedy action as an addi-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 193, + 318, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 318, + 206 + ], + "score": 1.0, + "content": "tional input, we obtain an average performance of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 318, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 156, + 216 + ], + "score": 0.89, + "content": "9 . 9 7 \\pm 0 . 0 2", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 204, + 318, + 217 + ], + "score": 1.0, + "content": ". Results are averaged over 100 seeds,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 216, + 318, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 318, + 227 + ], + "score": 1.0, + "content": "and shading is s.e.m. The code is available here:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 219, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 219, + 238 + ], + "score": 1.0, + "content": "www.bit.ly/2mBJLyk.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 6.5 + }, + { + "type": "image", + "bbox": [ + 327, + 82, + 490, + 240 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 327, + 82, + 490, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 327, + 82, + 490, + 240 + ], + "spans": [ + { + "bbox": [ + 327, + 82, + 490, + 240 + ], + "score": 0.969, + "type": "image", + "image_path": "5ad1d82f24e9ee7d32e4c518b21671940bf756b1552a474256a05432182b3238.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 327, + 82, + 490, + 95.16666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 327, + 95.16666666666667, + 490, + 108.33333333333334 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 327, + 108.33333333333334, + 490, + 121.50000000000001 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 327, + 121.50000000000001, + 490, + 134.66666666666669 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 327, + 134.66666666666669, + 490, + 147.83333333333334 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 327, + 147.83333333333334, + 490, + 161.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 327, + 161.0, + 490, + 174.16666666666666 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 327, + 174.16666666666666, + 490, + 187.33333333333331 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 327, + 187.33333333333331, + 490, + 200.49999999999997 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 327, + 200.49999999999997, + 490, + 213.66666666666663 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 327, + 213.66666666666663, + 490, + 226.8333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 327, + 226.8333333333333, + 490, + 239.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 343, + 250, + 485, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 341, + 249, + 487, + 263 + ], + "spans": [ + { + "bbox": [ + 341, + 249, + 487, + 263 + ], + "score": 1.0, + "content": "Figure 2: Results for the matrix game.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + } + ], + "index": 21.25 + }, + { + "type": "title", + "bbox": [ + 107, + 253, + 167, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 168, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 168, + 265 + ], + "score": 1.0, + "content": "6.2 HANABI", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 274, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "As shown in Table 1, our findings from the matrix game are for the most part confirmed on the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 284, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 298 + ], + "score": 1.0, + "content": "challenging Hanabi benchmark. To illustrate the contributions of the different components, we", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "score": 1.0, + "content": "compare average scores and win rates across 13 independent training runs of SAD and three different", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "options: IQL is simply the recurrent DQN agent with parameter sharing, VDN is the same agent but", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 318, + 501, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 501, + 330 + ], + "score": 1.0, + "content": "also learns a joint Q-function and finally SAD & AuxTask is the SAD agent with the auxiliary task.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 335, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "score": 1.0, + "content": "While we find that SAD significantly outperforms our baselines (IQL and VDN) for 2, 4 and 5 players", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "in terms of average score and/or win rate, there is no significant difference for 3 players, where VDN", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 357, + 245, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 245, + 369 + ], + "score": 1.0, + "content": "matches the performance of SAD.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "score": 1.0, + "content": "Interestingly, the auxiliary task only significantly helps the 2-player performance, where it substan-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "tially boosts the average score and win rate. In contrast, it drastically hurts performance for 3-5", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 396, + 345, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 345, + 408 + ], + "score": 1.0, + "content": "players, which opens an interesting avenue for future work.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 412, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "For completeness we have included training curves showing average scores and s.e.m. across all", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 424, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 435 + ], + "score": 1.0, + "content": "training runs for all numbers of players for our methods and ablations in Appendix B. We find that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "for 5 players the auxiliary task drastically reduces the variance of SAD and intermittently leads to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "higher performance during training but ultimately results in lower final performance. We can also", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "score": 1.0, + "content": "clearly see that despite 72 hours of training and billions of samples consumed, the performance has", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 468, + 453, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 453, + 480 + ], + "score": 1.0, + "content": "not plateaued for 3-5 players, pointing to an obvious avenue for further improvements.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 505, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "The original numbers in the Hanabi challenge and BAD used population based training (Jaderberg", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 496, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 507 + ], + "score": 1.0, + "content": "et al., 2018), effectively reporting maximum performance across a large number of different runs.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "Therefore, for reproducibility purposes, we report evaluations of the best model from our various", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 518, + 272, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 272, + 529 + ], + "score": 1.0, + "content": "training runs for each method in Table 2.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 107, + 534, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 548 + ], + "score": 1.0, + "content": "As shown, under this reporting we establish a new SOTA for learning methods on the self-play", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 545, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 506, + 558 + ], + "score": 1.0, + "content": "part of the Hanabi challenge for 2-5 players, with the most drastic improvements being achieved", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "for 3-5 players. In particular, we beat both the ACHA agent from Bard et al. (2019) and the BAD", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "score": 1.0, + "content": "agent on average score, even though both of them used population based training and require more", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 577, + 507, + 592 + ], + "spans": [ + { + "bbox": [ + 104, + 577, + 507, + 592 + ], + "score": 1.0, + "content": "compute. We note that while we follow the counting convention proposed by the challenge paper,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "BAD was optimized for a different counting scheme, in which agents keep their scores when they run", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 308, + 613 + ], + "score": 1.0, + "content": "out of lives. This may explain the higher win rate", + "type": "text" + }, + { + "bbox": [ + 309, + 600, + 342, + 611 + ], + "score": 0.87, + "content": "( 5 8 . 6 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "of BAD combined with a relatively low", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "mean score, which is exceeded even by our baseline methods. Once again, only the performance for", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "2-player is significantly improved by the auxiliary task and the 3-player setting is an outlier in the", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 633, + 407, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 407, + 645 + ], + "score": 1.0, + "content": "sense that SAD does not improve the best performance compared to VDN.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 51.5 + }, + { + "type": "title", + "bbox": [ + 107, + 662, + 303, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 660, + 304, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 304, + 676 + ], + "score": 1.0, + "content": "7 CONCLUSION AND FUTURE WORK", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 57 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "In this paper we presented the Simplified Action Decoder (SAD), a novel deep multi-agent RL", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "algorithm that allows agents to learn communication protocols in settings where no cheap-talk", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "channel is available. On the challenging benchmark Hanabi our work substantially improves the", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "SOTA for an RL method for all numbers of players. For two players SAD establishes a new high-score", + "type": "text" + } + ], + "index": 61 + } + ], + "index": 59.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 172, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 174, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 174, + 96 + ], + "score": 1.0, + "content": "6 RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 106, + 196, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 198, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 198, + 120 + ], + "score": 1.0, + "content": "6.1 MATRIX GAME", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 128, + 317, + 238 + ], + "lines": [ + { + "bbox": [ + 106, + 128, + 318, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 318, + 140 + ], + "score": 1.0, + "content": "As we can see in Figure 2, even in our simple ma-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 139, + 317, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 317, + 150 + ], + "score": 1.0, + "content": "trix game the greedy action input makes a drastic", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 150, + 317, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 317, + 162 + ], + "score": 1.0, + "content": "difference. With an average reward of around 9.5", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 160, + 318, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 318, + 174 + ], + "score": 1.0, + "content": "points, tabular IQL does well in this task, matching", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 319, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 121, + 183 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 122, + 172, + 143, + 182 + ], + "score": 0.52, + "content": "B A D", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 171, + 319, + 183 + ], + "score": 1.0, + "content": "results from Foerster et al. (2019). How-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 318, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 318, + 194 + ], + "score": 1.0, + "content": "ever, just by adding the greedy action as an addi-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 193, + 318, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 318, + 206 + ], + "score": 1.0, + "content": "tional input, we obtain an average performance of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 318, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 156, + 216 + ], + "score": 0.89, + "content": "9 . 9 7 \\pm 0 . 0 2", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 204, + 318, + 217 + ], + "score": 1.0, + "content": ". Results are averaged over 100 seeds,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 216, + 318, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 318, + 227 + ], + "score": 1.0, + "content": "and shading is s.e.m. The code is available here:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 219, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 219, + 238 + ], + "score": 1.0, + "content": "www.bit.ly/2mBJLyk.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 128, + 319, + 238 + ] + }, + { + "type": "image", + "bbox": [ + 327, + 82, + 490, + 240 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 327, + 82, + 490, + 240 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 327, + 82, + 490, + 240 + ], + "spans": [ + { + "bbox": [ + 327, + 82, + 490, + 240 + ], + "score": 0.969, + "type": "image", + "image_path": "5ad1d82f24e9ee7d32e4c518b21671940bf756b1552a474256a05432182b3238.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 327, + 82, + 490, + 95.16666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 327, + 95.16666666666667, + 490, + 108.33333333333334 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 327, + 108.33333333333334, + 490, + 121.50000000000001 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 327, + 121.50000000000001, + 490, + 134.66666666666669 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 327, + 134.66666666666669, + 490, + 147.83333333333334 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 327, + 147.83333333333334, + 490, + 161.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 327, + 161.0, + 490, + 174.16666666666666 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 327, + 174.16666666666666, + 490, + 187.33333333333331 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 327, + 187.33333333333331, + 490, + 200.49999999999997 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 327, + 200.49999999999997, + 490, + 213.66666666666663 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 327, + 213.66666666666663, + 490, + 226.8333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 327, + 226.8333333333333, + 490, + 239.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 343, + 250, + 485, + 261 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 341, + 249, + 487, + 263 + ], + "spans": [ + { + "bbox": [ + 341, + 249, + 487, + 263 + ], + "score": 1.0, + "content": "Figure 2: Results for the matrix game.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + } + ], + "index": 21.25 + }, + { + "type": "title", + "bbox": [ + 107, + 253, + 167, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 168, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 168, + 265 + ], + "score": 1.0, + "content": "6.2 HANABI", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 274, + 505, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "As shown in Table 1, our findings from the matrix game are for the most part confirmed on the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 284, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 298 + ], + "score": 1.0, + "content": "challenging Hanabi benchmark. To illustrate the contributions of the different components, we", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 506, + 308 + ], + "score": 1.0, + "content": "compare average scores and win rates across 13 independent training runs of SAD and three different", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "options: IQL is simply the recurrent DQN agent with parameter sharing, VDN is the same agent but", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 318, + 501, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 501, + 330 + ], + "score": 1.0, + "content": "also learns a joint Q-function and finally SAD & AuxTask is the SAD agent with the auxiliary task.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 273, + 506, + 330 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 335, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "score": 1.0, + "content": "While we find that SAD significantly outperforms our baselines (IQL and VDN) for 2, 4 and 5 players", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "in terms of average score and/or win rate, there is no significant difference for 3 players, where VDN", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 357, + 245, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 245, + 369 + ], + "score": 1.0, + "content": "matches the performance of SAD.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 334, + 505, + 369 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "score": 1.0, + "content": "Interestingly, the auxiliary task only significantly helps the 2-player performance, where it substan-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "tially boosts the average score and win rate. In contrast, it drastically hurts performance for 3-5", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 396, + 345, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 345, + 408 + ], + "score": 1.0, + "content": "players, which opens an interesting avenue for future work.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 373, + 506, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 412, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "For completeness we have included training curves showing average scores and s.e.m. across all", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 424, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 435 + ], + "score": 1.0, + "content": "training runs for all numbers of players for our methods and ablations in Appendix B. We find that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "for 5 players the auxiliary task drastically reduces the variance of SAD and intermittently leads to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "higher performance during training but ultimately results in lower final performance. We can also", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "score": 1.0, + "content": "clearly see that despite 72 hours of training and billions of samples consumed, the performance has", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 468, + 453, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 453, + 480 + ], + "score": 1.0, + "content": "not plateaued for 3-5 players, pointing to an obvious avenue for further improvements.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 412, + 505, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 505, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "The original numbers in the Hanabi challenge and BAD used population based training (Jaderberg", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 496, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 507 + ], + "score": 1.0, + "content": "et al., 2018), effectively reporting maximum performance across a large number of different runs.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "Therefore, for reproducibility purposes, we report evaluations of the best model from our various", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 518, + 272, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 272, + 529 + ], + "score": 1.0, + "content": "training runs for each method in Table 2.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 483, + 506, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 534, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 548 + ], + "score": 1.0, + "content": "As shown, under this reporting we establish a new SOTA for learning methods on the self-play", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 545, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 506, + 558 + ], + "score": 1.0, + "content": "part of the Hanabi challenge for 2-5 players, with the most drastic improvements being achieved", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "for 3-5 players. In particular, we beat both the ACHA agent from Bard et al. (2019) and the BAD", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "score": 1.0, + "content": "agent on average score, even though both of them used population based training and require more", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 577, + 507, + 592 + ], + "spans": [ + { + "bbox": [ + 104, + 577, + 507, + 592 + ], + "score": 1.0, + "content": "compute. We note that while we follow the counting convention proposed by the challenge paper,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "BAD was optimized for a different counting scheme, in which agents keep their scores when they run", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 308, + 613 + ], + "score": 1.0, + "content": "out of lives. This may explain the higher win rate", + "type": "text" + }, + { + "bbox": [ + 309, + 600, + 342, + 611 + ], + "score": 0.87, + "content": "( 5 8 . 6 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "of BAD combined with a relatively low", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "mean score, which is exceeded even by our baseline methods. Once again, only the performance for", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "2-player is significantly improved by the auxiliary task and the 3-player setting is an outlier in the", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 633, + 407, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 407, + 645 + ], + "score": 1.0, + "content": "sense that SAD does not improve the best performance compared to VDN.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 51.5, + "bbox_fs": [ + 104, + 533, + 507, + 645 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 662, + 303, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 660, + 304, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 304, + 676 + ], + "score": 1.0, + "content": "7 CONCLUSION AND FUTURE WORK", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 57 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "In this paper we presented the Simplified Action Decoder (SAD), a novel deep multi-agent RL", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "algorithm that allows agents to learn communication protocols in settings where no cheap-talk", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "channel is available. On the challenging benchmark Hanabi our work substantially improves the", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "SOTA for an RL method for all numbers of players. For two players SAD establishes a new high-score", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "across any method. Furthermore we accomplish all of this with a method that is both simpler and", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 568, + 504, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 504, + 579 + ], + "score": 1.0, + "content": "requires less compute than previous advances. While these are encouraging steps, there is clearly", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "more work to do. In particular, there remains a large performance gap between the numbers achieved", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "by SAD and the known performance of hat-coding strategies (Wu, 2018) for 3-5 players. One possible", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "reason is that SAD does not undertake any explicit exploration in the space of possible conventions.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "Another promising route for future work is to integrate search with RL, since this has produced SOTA", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 621, + 426, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 426, + 635 + ], + "score": 1.0, + "content": "results in a number of different domains including Poker, Go and backgammon.", + "type": "text", + "cross_page": true + } + ], + "index": 19 + } + ], + "index": 59.5, + "bbox_fs": [ + 105, + 687, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 129, + 82, + 482, + 208 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 129, + 82, + 482, + 208 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 82, + 482, + 208 + ], + "spans": [ + { + "bbox": [ + 129, + 82, + 482, + 208 + ], + "score": 0.983, + "html": "
Agent 2 Players 3 Players 4 Players 5 Players
IQL (Baseline)23.77 ± 0.04 43.88 ± 1.21 %23.02 ± 0.10 26.16 ± 2.01 %21.99 ± 0.09 10.15 ± 0.86 %20.60 ± 0.11 2.27 ± 0.32 %
VDN23.83 ± 0.0323.71 ± 0.0623.03 ± 0.1521.18 ± 0.12
(Baseline) SAD44.97 ± 1.28 % 23.87 ± 0.0341.16 ± 1.27 % 23.69 ± 0.0523.57 ± 2.20 % 23.27 ± 0.162.26 ± 0.32 % 22.06 ± 0.23
SAD AuxTask47.90 ± 1.10 % 24.02 ± 0.0141.12 ± 1.10 % 23.56 ± 0.0729.38 ± 2.63 % 22.78 ± 0.107.22 ± 1.29 % 21.47 ± 0.08
", + "type": "table", + "image_path": "06f1e8f230e055bf72fef14c157586d32d70dba51aea45b44487b3a4cfae18fc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 82, + 482, + 124.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 124.0, + 482, + 166.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 166.0, + 482, + 208.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 219, + 505, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 231 + ], + "score": 1.0, + "content": "Table 1: Mean performance of our methods and baselines on Hanabi. We take the final models of 13 independent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 230, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 240 + ], + "score": 1.0, + "content": "runs, i.e. 13 models per algorithm per player setting. Each model is evaluated on 100K games. Mean and s.e.m", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 240, + 500, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 500, + 250 + ], + "score": 1.0, + "content": "over the mean scores of the 13 models are shown in the table. The second row of each section is the win rate.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 126, + 267, + 485, + 478 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 126, + 267, + 485, + 478 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 126, + 267, + 485, + 478 + ], + "spans": [ + { + "bbox": [ + 126, + 267, + 485, + 478 + ], + "score": 0.985, + "html": "
Agent2 Players 3 Players 4 Players 5 Players
Rainbow (Bard et al., 2019)20.64 ± 0.03 2.5%18.71 ±0.01 0.2%18.00 ± 0.17 0%15.26 ± 0.18 0%
ACHA (Bard et al., 2019)22.73 ± 0.12 15.1%20.24 ± 0.15 1.1%21.57 ± 0.12 2.4%16.80 ± 0.13 0%
BAD (Foerster et al.,2019)23.92 ± 0.01 58.56%==1
IQL (Baseline) 50.47%23.97 ± 0.01 40.25%23.69 ± 0.01 19.39%22.76 ± 0.01 4.93%21.29 ± 0.01
VDN (Baseline)23.96 ± 0.01 50.27%23.99 ± 0.01 50.37%23.79 ± 0.00 38.86%21.80 ± 0.01 4.98%
SAD24.01 ± 0.01 52.39%23.93 ± 0.01 48.05%23.81 ± 0.01 41.45%23.01 ± 0.01 13.93%
SAD& AuxTask24.08 ± 0.01 56.09%23.81 ± 0.01 49.74%23.47 ± 0.01 33.87%22.25 ± 0.01 7.33%
", + "type": "table", + "image_path": "20d58361e17764c14fcd32f70066f1ef6e07903eb74e90302ea7c9ff64241dd4.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 126, + 267, + 485, + 337.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 126, + 337.3333333333333, + 485, + 407.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 126, + 407.66666666666663, + 485, + 477.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 488, + 505, + 530 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "Table 2: Comparison between the previous SOTA learning methods and ours. We take the best model of 13", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "runs for each of our methods and baselines. Each model is evaluated on 100K games with different seeds. Mean", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 509, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 505, + 520 + ], + "score": 1.0, + "content": "and s.e.m over the 100K games are shown in the table. The s.e.m. is less than 0.01 for most models. Bold", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 519, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 530 + ], + "score": 1.0, + "content": "numbers are the best results achieved with learning algorithms. The second row of each section is the win rate.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + } + ], + "index": 8.75 + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "across any method. Furthermore we accomplish all of this with a method that is both simpler and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 568, + 504, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 504, + 579 + ], + "score": 1.0, + "content": "requires less compute than previous advances. While these are encouraging steps, there is clearly", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "more work to do. In particular, there remains a large performance gap between the numbers achieved", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "by SAD and the known performance of hat-coding strategies (Wu, 2018) for 3-5 players. One possible", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "reason is that SAD does not undertake any explicit exploration in the space of possible conventions.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "Another promising route for future work is to integrate search with RL, since this has produced SOTA", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 621, + 426, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 426, + 635 + ], + "score": 1.0, + "content": "results in a number of different domains including Poker, Go and backgammon.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 129, + 82, + 482, + 208 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 129, + 82, + 482, + 208 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 82, + 482, + 208 + ], + "spans": [ + { + "bbox": [ + 129, + 82, + 482, + 208 + ], + "score": 0.983, + "html": "
Agent 2 Players 3 Players 4 Players 5 Players
IQL (Baseline)23.77 ± 0.04 43.88 ± 1.21 %23.02 ± 0.10 26.16 ± 2.01 %21.99 ± 0.09 10.15 ± 0.86 %20.60 ± 0.11 2.27 ± 0.32 %
VDN23.83 ± 0.0323.71 ± 0.0623.03 ± 0.1521.18 ± 0.12
(Baseline) SAD44.97 ± 1.28 % 23.87 ± 0.0341.16 ± 1.27 % 23.69 ± 0.0523.57 ± 2.20 % 23.27 ± 0.162.26 ± 0.32 % 22.06 ± 0.23
SAD AuxTask47.90 ± 1.10 % 24.02 ± 0.0141.12 ± 1.10 % 23.56 ± 0.0729.38 ± 2.63 % 22.78 ± 0.107.22 ± 1.29 % 21.47 ± 0.08
", + "type": "table", + "image_path": "06f1e8f230e055bf72fef14c157586d32d70dba51aea45b44487b3a4cfae18fc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 82, + 482, + 124.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 124.0, + 482, + 166.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 166.0, + 482, + 208.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 219, + 505, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 231 + ], + "score": 1.0, + "content": "Table 1: Mean performance of our methods and baselines on Hanabi. We take the final models of 13 independent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 230, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 240 + ], + "score": 1.0, + "content": "runs, i.e. 13 models per algorithm per player setting. Each model is evaluated on 100K games. Mean and s.e.m", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 240, + 500, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 500, + 250 + ], + "score": 1.0, + "content": "over the mean scores of the 13 models are shown in the table. The second row of each section is the win rate.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 126, + 267, + 485, + 478 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 126, + 267, + 485, + 478 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 126, + 267, + 485, + 478 + ], + "spans": [ + { + "bbox": [ + 126, + 267, + 485, + 478 + ], + "score": 0.985, + "html": "
Agent2 Players 3 Players 4 Players 5 Players
Rainbow (Bard et al., 2019)20.64 ± 0.03 2.5%18.71 ±0.01 0.2%18.00 ± 0.17 0%15.26 ± 0.18 0%
ACHA (Bard et al., 2019)22.73 ± 0.12 15.1%20.24 ± 0.15 1.1%21.57 ± 0.12 2.4%16.80 ± 0.13 0%
BAD (Foerster et al.,2019)23.92 ± 0.01 58.56%==1
IQL (Baseline) 50.47%23.97 ± 0.01 40.25%23.69 ± 0.01 19.39%22.76 ± 0.01 4.93%21.29 ± 0.01
VDN (Baseline)23.96 ± 0.01 50.27%23.99 ± 0.01 50.37%23.79 ± 0.00 38.86%21.80 ± 0.01 4.98%
SAD24.01 ± 0.01 52.39%23.93 ± 0.01 48.05%23.81 ± 0.01 41.45%23.01 ± 0.01 13.93%
SAD& AuxTask24.08 ± 0.01 56.09%23.81 ± 0.01 49.74%23.47 ± 0.01 33.87%22.25 ± 0.01 7.33%
", + "type": "table", + "image_path": "20d58361e17764c14fcd32f70066f1ef6e07903eb74e90302ea7c9ff64241dd4.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 126, + 267, + 485, + 337.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 126, + 337.3333333333333, + 485, + 407.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 126, + 407.66666666666663, + 485, + 477.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 488, + 505, + 530 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "Table 2: Comparison between the previous SOTA learning methods and ours. We take the best model of 13", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "runs for each of our methods and baselines. Each model is evaluated on 100K games with different seeds. Mean", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 509, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 505, + 520 + ], + "score": 1.0, + "content": "and s.e.m over the 100K games are shown in the table. The s.e.m. is less than 0.01 for most models. Bold", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 519, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 530 + ], + "score": 1.0, + "content": "numbers are the best results achieved with learning algorithms. The second row of each section is the win rate.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + } + ], + "index": 8.75 + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 633 + ], + "lines": [], + "index": 16, + "bbox_fs": [ + 105, + 556, + 506, + 635 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Chris L Baker, Julian Jara-Ettinger, Rebecca Saxe, and Joshua B Tenenbaum. Rational quantitative", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 116, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "attribution of beliefs, desires and percepts in human mentalizing. Nature Human Behaviour, 1(4):", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 167, + 133 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 167, + 133 + ], + "score": 1.0, + "content": "0064, 2017.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 139, + 506, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 151, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 151, + 506, + 163 + ], + "score": 1.0, + "content": "Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 115, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The Hanabi Challenge: A New", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 171, + 508, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 508, + 186 + ], + "score": 1.0, + "content": "Frontier for AI Research. arXiv:1902.00506 [cs, stat], February 2019. URL http://arxiv.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 183, + 310, + 195 + ], + "spans": [ + { + "bbox": [ + 115, + 183, + 310, + 195 + ], + "score": 1.0, + "content": "org/abs/1902.00506. arXiv: 1902.00506.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 201, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 200, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 215 + ], + "score": 1.0, + "content": "Noam Brown and Tuomas Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 213, + 309, + 225 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 309, + 225 + ], + "score": 1.0, + "content": "top professionals. Science, pp. eaao1733, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 506, + 244 + ], + "score": 1.0, + "content": "Noam Brown and Tuomas Sandholm. Superhuman AI for multiplayer poker. Science, pp. eaay2400,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 114, + 240, + 143, + 254 + ], + "spans": [ + { + "bbox": [ + 114, + 240, + 143, + 254 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 505, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 506, + 272 + ], + "score": 1.0, + "content": "Noam Brown, Tuomas Sandholm, and Brandon Amos. Depth-Limited Solving for Imperfect-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 270, + 233, + 283 + ], + "spans": [ + { + "bbox": [ + 116, + 270, + 233, + 283 + ], + "score": 1.0, + "content": "Information Games. pp. 14.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 505, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "Murray Campbell, A Joseph Hoane Jr, and Feng-hsiung Hsu. Deep Blue. Artificial intelligence, 134", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 299, + 194, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 299, + 194, + 311 + ], + "score": 1.0, + "content": "(1-2):57–83, 2002.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 105, + 317, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "Rodrigo Canaan, Julian Togelius, Andy Nealen, and Stefan Menzel. Diverse agents for ad-hoc", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 329, + 371, + 340 + ], + "spans": [ + { + "bbox": [ + 116, + 329, + 371, + 340 + ], + "score": 1.0, + "content": "cooperation in hanabi. arXiv preprint arXiv:1907.03840, 2019.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 359 + ], + "score": 1.0, + "content": "Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "communicate with deep multi-agent reinforcement learning. In Advances in Neural Information", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 368, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 288, + 380 + ], + "score": 1.0, + "content": "Processing Systems, pp. 2137–2145, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "Jakob Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson, Matthew", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "Botvinick, and Michael Bowling. Bayesian action decoder for deep multi-agent reinforcement", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 408, + 450, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 450, + 421 + ], + "score": 1.0, + "content": "learning. In International Conference on Machine Learning, pp. 1942–1951, 2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 426, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 104, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 436, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 506, + 450 + ], + "score": 1.0, + "content": "Counterfactual multi-agent policy gradients. In Thirty-Second AAAI Conference on Artificial", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 448, + 198, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 198, + 461 + ], + "score": 1.0, + "content": "Intelligence, 2018a.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 104, + 465, + 507, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 507, + 479 + ], + "score": 1.0, + "content": "Jakob N. Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 115, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "Matthew Botvinick, and Michael Bowling. Bayesian Action Decoder for Deep Multi-Agent", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 488, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 507, + 501 + ], + "score": 1.0, + "content": "Reinforcement Learning. arXiv:1811.01458 [cs], November 2018b. URL http://arxiv.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 500, + 311, + 510 + ], + "spans": [ + { + "bbox": [ + 116, + 500, + 311, + 510 + ], + "score": 1.0, + "content": "org/abs/1811.01458. arXiv: 1811.01458.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 104, + 516, + 504, + 539 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 530 + ], + "score": 1.0, + "content": "James Goodman. Re-determinizing information set monte carlo tree search in hanabi. arXiv preprint", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 528, + 219, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 528, + 219, + 538 + ], + "score": 1.0, + "content": "arXiv:1902.06075, 2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 506, + 579 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 507, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 507, + 559 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735–", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 555, + 509, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 509, + 570 + ], + "score": 1.0, + "content": "1780, November 1997. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL http://dx.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 568, + 325, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 568, + 325, + 580 + ], + "score": 1.0, + "content": "doi.org/10.1162/neco.1997.9.8.1735.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 585, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 597, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 116, + 597, + 506, + 609 + ], + "score": 1.0, + "content": "and David Silver. Distributed prioritized experience replay. CoRR, abs/1803.00933, 2018. URL", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 117, + 609, + 307, + 619 + ], + "spans": [ + { + "bbox": [ + 117, + 609, + 307, + 619 + ], + "score": 1.0, + "content": "http://arxiv.org/abs/1803.00933.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 505, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 639 + ], + "score": 1.0, + "content": "Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 116, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 648, + 254, + 659 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 254, + 659 + ], + "score": 1.0, + "content": "preprint arXiv:1611.05397, 2016.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 506, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 507, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 507, + 678 + ], + "score": 1.0, + "content": "Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Cas-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 115, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "taneda, Charles Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 115, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 114, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 114, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "Thore Graepel. Human-level performance in first-person multiplayer games with population-based", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 709, + 507, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 507, + 722 + ], + "score": 1.0, + "content": "deep reinforcement learning. arXiv:1807.01281 [cs, stat], July 2018. doi: 10.1126/science.aau6249.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 117, + 721, + 411, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 411, + 732 + ], + "score": 1.0, + "content": "URL http://arxiv.org/abs/1807.01281. arXiv: 1807.01281.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Chris L Baker, Julian Jara-Ettinger, Rebecca Saxe, and Joshua B Tenenbaum. Rational quantitative", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 116, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "attribution of beliefs, desires and percepts in human mentalizing. Nature Human Behaviour, 1(4):", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 167, + 133 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 167, + 133 + ], + "score": 1.0, + "content": "0064, 2017.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 100, + 506, + 133 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 139, + 506, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 151, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 151, + 506, + 163 + ], + "score": 1.0, + "content": "Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 115, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The Hanabi Challenge: A New", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 171, + 508, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 508, + 186 + ], + "score": 1.0, + "content": "Frontier for AI Research. arXiv:1902.00506 [cs, stat], February 2019. URL http://arxiv.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 183, + 310, + 195 + ], + "spans": [ + { + "bbox": [ + 115, + 183, + 310, + 195 + ], + "score": 1.0, + "content": "org/abs/1902.00506. arXiv: 1902.00506.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 140, + 508, + 195 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 201, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 200, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 215 + ], + "score": 1.0, + "content": "Noam Brown and Tuomas Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 213, + 309, + 225 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 309, + 225 + ], + "score": 1.0, + "content": "top professionals. Science, pp. eaao1733, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 200, + 505, + 225 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 506, + 244 + ], + "score": 1.0, + "content": "Noam Brown and Tuomas Sandholm. Superhuman AI for multiplayer poker. Science, pp. eaay2400,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 114, + 240, + 143, + 254 + ], + "spans": [ + { + "bbox": [ + 114, + 240, + 143, + 254 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 228, + 506, + 254 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 259, + 505, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 506, + 272 + ], + "score": 1.0, + "content": "Noam Brown, Tuomas Sandholm, and Brandon Amos. Depth-Limited Solving for Imperfect-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 270, + 233, + 283 + ], + "spans": [ + { + "bbox": [ + 116, + 270, + 233, + 283 + ], + "score": 1.0, + "content": "Information Games. pp. 14.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 259, + 506, + 283 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 505, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "Murray Campbell, A Joseph Hoane Jr, and Feng-hsiung Hsu. Deep Blue. Artificial intelligence, 134", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 299, + 194, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 299, + 194, + 311 + ], + "score": 1.0, + "content": "(1-2):57–83, 2002.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 288, + 505, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 317, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "Rodrigo Canaan, Julian Togelius, Andy Nealen, and Stefan Menzel. Diverse agents for ad-hoc", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 329, + 371, + 340 + ], + "spans": [ + { + "bbox": [ + 116, + 329, + 371, + 340 + ], + "score": 1.0, + "content": "cooperation in hanabi. arXiv preprint arXiv:1907.03840, 2019.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 316, + 505, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 359 + ], + "score": 1.0, + "content": "Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "communicate with deep multi-agent reinforcement learning. In Advances in Neural Information", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 368, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 288, + 380 + ], + "score": 1.0, + "content": "Processing Systems, pp. 2137–2145, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 345, + 505, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 386, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 399 + ], + "score": 1.0, + "content": "Jakob Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson, Matthew", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "Botvinick, and Michael Bowling. Bayesian action decoder for deep multi-agent reinforcement", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 408, + 450, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 408, + 450, + 421 + ], + "score": 1.0, + "content": "learning. In International Conference on Machine Learning, pp. 1942–1951, 2019.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 386, + 505, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 426, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 104, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 436, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 506, + 450 + ], + "score": 1.0, + "content": "Counterfactual multi-agent policy gradients. In Thirty-Second AAAI Conference on Artificial", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 448, + 198, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 198, + 461 + ], + "score": 1.0, + "content": "Intelligence, 2018a.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 425, + 506, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 104, + 465, + 507, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 507, + 479 + ], + "score": 1.0, + "content": "Jakob N. Foerster, Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 115, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "Matthew Botvinick, and Michael Bowling. Bayesian Action Decoder for Deep Multi-Agent", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 488, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 507, + 501 + ], + "score": 1.0, + "content": "Reinforcement Learning. arXiv:1811.01458 [cs], November 2018b. URL http://arxiv.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 500, + 311, + 510 + ], + "spans": [ + { + "bbox": [ + 116, + 500, + 311, + 510 + ], + "score": 1.0, + "content": "org/abs/1811.01458. arXiv: 1811.01458.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 465, + 507, + 510 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 516, + 504, + 539 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 530 + ], + "score": 1.0, + "content": "James Goodman. Re-determinizing information set monte carlo tree search in hanabi. arXiv preprint", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 528, + 219, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 528, + 219, + 538 + ], + "score": 1.0, + "content": "arXiv:1902.06075, 2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 515, + 505, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 506, + 579 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 507, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 507, + 559 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735–", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 555, + 509, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 509, + 570 + ], + "score": 1.0, + "content": "1780, November 1997. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL http://dx.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 568, + 325, + 580 + ], + "spans": [ + { + "bbox": [ + 116, + 568, + 325, + 580 + ], + "score": 1.0, + "content": "doi.org/10.1162/neco.1997.9.8.1735.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 545, + 509, + 580 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 585, + 505, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado van Hasselt,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 597, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 116, + 597, + 506, + 609 + ], + "score": 1.0, + "content": "and David Silver. Distributed prioritized experience replay. CoRR, abs/1803.00933, 2018. URL", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 117, + 609, + 307, + 619 + ], + "spans": [ + { + "bbox": [ + 117, + 609, + 307, + 619 + ], + "score": 1.0, + "content": "http://arxiv.org/abs/1803.00933.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 585, + 506, + 619 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 505, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 505, + 639 + ], + "score": 1.0, + "content": "Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 116, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 648, + 254, + 659 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 254, + 659 + ], + "score": 1.0, + "content": "preprint arXiv:1611.05397, 2016.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 625, + 505, + 659 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 506, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 507, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 507, + 678 + ], + "score": 1.0, + "content": "Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Cas-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 115, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "taneda, Charles Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 115, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 114, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 114, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "Thore Graepel. Human-level performance in first-person multiplayer games with population-based", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 709, + 507, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 507, + 722 + ], + "score": 1.0, + "content": "deep reinforcement learning. arXiv:1807.01281 [cs, stat], July 2018. doi: 10.1126/science.aau6249.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 117, + 721, + 411, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 411, + 732 + ], + "score": 1.0, + "content": "URL http://arxiv.org/abs/1807.01281. arXiv: 1807.01281.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 665, + 507, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Steven Kapturowski, Georg Ostrovski, John Quan, and Will Dabney. RECURRENT EXPERIENCE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 445, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 445, + 107 + ], + "score": 1.0, + "content": "REPLAY IN DISTRIBUTED REINFORCEMENT LEARNING. pp. 1–19, 2019.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 114, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 126, + 214, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 214, + 136 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 147, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 106, + 147, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 505, + 159 + ], + "score": 1.0, + "content": "Claire Liang, Julia Proft, Erik Andersen, and Ross A Knepper. Implicit communication of actionable", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 116, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "information in human-ai teams. In Proceedings of the 2019 CHI Conference on Human Factors in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 169, + 284, + 181 + ], + "spans": [ + { + "bbox": [ + 116, + 169, + 284, + 181 + ], + "score": 1.0, + "content": "Computing Systems, pp. 95. ACM, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 190, + 504, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 202, + 504, + 213 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 504, + 213 + ], + "score": 1.0, + "content": "actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 212, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 116, + 212, + 289, + 224 + ], + "score": 1.0, + "content": "Processing Systems, pp. 6379–6390, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 233, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 256, + 375, + 267 + ], + "spans": [ + { + "bbox": [ + 116, + 256, + 375, + 267 + ], + "score": 1.0, + "content": "complex environments. arXiv preprint arXiv:1611.03673, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 506, + 333 + ], + "lines": [ + { + "bbox": [ + 107, + 277, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 107, + 277, + 506, + 289 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Belle-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 287, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 115, + 287, + 506, + 302 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 298, + 507, + 312 + ], + "spans": [ + { + "bbox": [ + 115, + 298, + 507, + 312 + ], + "score": 1.0, + "content": "Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 309, + 507, + 323 + ], + "spans": [ + { + "bbox": [ + 115, + 309, + 507, + 323 + ], + "score": 1.0, + "content": "Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Na-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 320, + 507, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 320, + 507, + 334 + ], + "score": 1.0, + "content": "ture, 518(7540):529–533, 02 2015. URL http://dx.doi.org/10.1038/nature14236.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "Matej Moravcík, Martin Schmid, Neil Burch, Viliam Lis ˇ y, Dustin Morrill, Nolan Bard, Trevor `", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 116, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "Davis, Kevin Waugh, Michael Johanson, and Michael Bowling. Deepstack: Expert-level artificial", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 117, + 364, + 421, + 376 + ], + "spans": [ + { + "bbox": [ + 117, + 364, + 421, + 376 + ], + "score": 1.0, + "content": "intelligence in heads-up no-limit poker. Science, 356(6337):508–513, 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "Ashutosh Nayyar, Aditya Mahajan, and Demosthenis Teneketzis. Decentralized stochastic control", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 116, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "with partial history sharing: A common information approach. IEEE Transactions on Automatic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 407, + 251, + 419 + ], + "spans": [ + { + "bbox": [ + 117, + 407, + 251, + 419 + ], + "score": 1.0, + "content": "Control, 58(7):1644–1658, 2013.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 428, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 507, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 507, + 442 + ], + "score": 1.0, + "content": "Thanh Thi Nguyen, Ngoc Duy Nguyen, and Saeid Nahavandi. Deep reinforcement learning for multi-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 115, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "agent systems: a review of challenges, solutions and applications. arXiv preprint arXiv:1812.11794,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 450, + 142, + 462 + ], + "spans": [ + { + "bbox": [ + 115, + 450, + 142, + 462 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 105, + 472, + 464, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 465, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 465, + 485 + ], + "score": 1.0, + "content": "Arthur O’Dwyer. Hanabi. https://github.com/Quuxplusone/Hanabi, 2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 493, + 504, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "Frans A Oliehoek. Decentralized pomdps. In Reinforcement Learning, pp. 471–503. Springer, 2012.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 506, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 507, + 527 + ], + "score": 1.0, + "content": "Tabish Rashid, Mikayel Samvelyan, Christian Schröder de Witt, Gregory Farquhar, Jakob N. Foerster,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 526, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 116, + 526, + 506, + 538 + ], + "score": 1.0, + "content": "and Shimon Whiteson. QMIX: monotonic value function factorisation for deep multi-agent", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 537, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 115, + 537, + 506, + 549 + ], + "score": 1.0, + "content": "reinforcement learning. CoRR, abs/1803.11485, 2018. URL http://arxiv.org/abs/", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 548, + 181, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 181, + 559 + ], + "score": 1.0, + "content": "1803.11485.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 105, + 568, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 582 + ], + "score": 1.0, + "content": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized Experience Replay. arXiv", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 580, + 290, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 580, + 290, + 591 + ], + "score": 1.0, + "content": "e-prints, art. arXiv:1511.05952, Nov 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 601, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 611, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 115, + 611, + 506, + 626 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 623, + 493, + 636 + ], + "spans": [ + { + "bbox": [ + 117, + 623, + 493, + 636 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 644, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 506, + 658 + ], + "score": 1.0, + "content": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 115, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 117, + 667, + 317, + 679 + ], + "spans": [ + { + "bbox": [ + 117, + 667, + 317, + 679 + ], + "score": 1.0, + "content": "human knowledge. Nature, 550(7676):354, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 507, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 507, + 700 + ], + "score": 1.0, + "content": "David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. A general reinforcement", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419):", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 721, + 190, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 190, + 731 + ], + "score": 1.0, + "content": "1140–1144, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Steven Kapturowski, Georg Ostrovski, John Quan, and Will Dabney. RECURRENT EXPERIENCE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 445, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 445, + 107 + ], + "score": 1.0, + "content": "REPLAY IN DISTRIBUTED REINFORCEMENT LEARNING. pp. 1–19, 2019.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 107 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 114, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "score": 1.0, + "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 126, + 214, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 214, + 136 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 113, + 506, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 147, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 106, + 147, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 505, + 159 + ], + "score": 1.0, + "content": "Claire Liang, Julia Proft, Erik Andersen, and Ross A Knepper. Implicit communication of actionable", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 116, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "information in human-ai teams. In Proceedings of the 2019 CHI Conference on Human Factors in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 169, + 284, + 181 + ], + "spans": [ + { + "bbox": [ + 116, + 169, + 284, + 181 + ], + "score": 1.0, + "content": "Computing Systems, pp. 95. ACM, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 147, + 505, + 181 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 190, + 504, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 202, + 504, + 213 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 504, + 213 + ], + "score": 1.0, + "content": "actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 212, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 116, + 212, + 289, + 224 + ], + "score": 1.0, + "content": "Processing Systems, pp. 6379–6390, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 190, + 505, + 224 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 233, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 256, + 375, + 267 + ], + "spans": [ + { + "bbox": [ + 116, + 256, + 375, + 267 + ], + "score": 1.0, + "content": "complex environments. arXiv preprint arXiv:1611.03673, 2016.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 232, + 506, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 506, + 333 + ], + "lines": [ + { + "bbox": [ + 107, + 277, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 107, + 277, + 506, + 289 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Belle-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 287, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 115, + 287, + 506, + 302 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 298, + 507, + 312 + ], + "spans": [ + { + "bbox": [ + 115, + 298, + 507, + 312 + ], + "score": 1.0, + "content": "Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 309, + 507, + 323 + ], + "spans": [ + { + "bbox": [ + 115, + 309, + 507, + 323 + ], + "score": 1.0, + "content": "Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Na-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 320, + 507, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 320, + 507, + 334 + ], + "score": 1.0, + "content": "ture, 518(7540):529–533, 02 2015. URL http://dx.doi.org/10.1038/nature14236.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 107, + 277, + 507, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "Matej Moravcík, Martin Schmid, Neil Burch, Viliam Lis ˇ y, Dustin Morrill, Nolan Bard, Trevor `", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 116, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "Davis, Kevin Waugh, Michael Johanson, and Michael Bowling. Deepstack: Expert-level artificial", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 117, + 364, + 421, + 376 + ], + "spans": [ + { + "bbox": [ + 117, + 364, + 421, + 376 + ], + "score": 1.0, + "content": "intelligence in heads-up no-limit poker. Science, 356(6337):508–513, 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 342, + 505, + 376 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "Ashutosh Nayyar, Aditya Mahajan, and Demosthenis Teneketzis. Decentralized stochastic control", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 116, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "with partial history sharing: A common information approach. IEEE Transactions on Automatic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 407, + 251, + 419 + ], + "spans": [ + { + "bbox": [ + 117, + 407, + 251, + 419 + ], + "score": 1.0, + "content": "Control, 58(7):1644–1658, 2013.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 385, + 505, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 428, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 428, + 507, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 507, + 442 + ], + "score": 1.0, + "content": "Thanh Thi Nguyen, Ngoc Duy Nguyen, and Saeid Nahavandi. Deep reinforcement learning for multi-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 115, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "agent systems: a review of challenges, solutions and applications. arXiv preprint arXiv:1812.11794,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 450, + 142, + 462 + ], + "spans": [ + { + "bbox": [ + 115, + 450, + 142, + 462 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 428, + 507, + 462 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 472, + 464, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 465, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 465, + 485 + ], + "score": 1.0, + "content": "Arthur O’Dwyer. Hanabi. https://github.com/Quuxplusone/Hanabi, 2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 471, + 465, + 485 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 493, + 504, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "Frans A Oliehoek. Decentralized pomdps. In Reinforcement Learning, pp. 471–503. Springer, 2012.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 492, + 506, + 506 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 506, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 507, + 527 + ], + "score": 1.0, + "content": "Tabish Rashid, Mikayel Samvelyan, Christian Schröder de Witt, Gregory Farquhar, Jakob N. Foerster,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 526, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 116, + 526, + 506, + 538 + ], + "score": 1.0, + "content": "and Shimon Whiteson. QMIX: monotonic value function factorisation for deep multi-agent", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 537, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 115, + 537, + 506, + 549 + ], + "score": 1.0, + "content": "reinforcement learning. CoRR, abs/1803.11485, 2018. URL http://arxiv.org/abs/", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 548, + 181, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 181, + 559 + ], + "score": 1.0, + "content": "1803.11485.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 513, + 507, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 568, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 582 + ], + "score": 1.0, + "content": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized Experience Replay. arXiv", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 580, + 290, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 580, + 290, + 591 + ], + "score": 1.0, + "content": "e-prints, art. arXiv:1511.05952, Nov 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 567, + 505, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 601, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 611, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 115, + 611, + 506, + 626 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 623, + 493, + 636 + ], + "spans": [ + { + "bbox": [ + 117, + 623, + 493, + 636 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 600, + 506, + 636 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 644, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 506, + 658 + ], + "score": 1.0, + "content": "David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 115, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of go without", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 117, + 667, + 317, + 679 + ], + "spans": [ + { + "bbox": [ + 117, + 667, + 317, + 679 + ], + "score": 1.0, + "content": "human knowledge. Nature, 550(7676):354, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 643, + 506, + 679 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 507, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 507, + 700 + ], + "score": 1.0, + "content": "David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. A general reinforcement", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419):", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 721, + 190, + 731 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 190, + 731 + ], + "score": 1.0, + "content": "1140–1144, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 686, + 507, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 507, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 507, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 507, + 95 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, arthur szlam, and Rob Fergus. Learning Multiagent Communica-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 507, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 507, + 107 + ], + "score": 1.0, + "content": "tion with Backpropagation. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 102, + 507, + 119 + ], + "spans": [ + { + "bbox": [ + 114, + 102, + 507, + 119 + ], + "score": 1.0, + "content": "and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "2244–2252. Curran Associates, Inc., 2016. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 127, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 117, + 127, + 501, + 138 + ], + "score": 1.0, + "content": "6398-learning-multiagent-communication-with-backpropagation.pdf.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 506, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinícius Flores Zambaldi,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 507, + 168 + ], + "score": 1.0, + "content": "Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z. Leibo, Karl Tuyls, and Thore Graepel.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 167, + 508, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 508, + 180 + ], + "score": 1.0, + "content": "Value-decomposition networks for cooperative multi-agent learning. CoRR, abs/1706.05296, 2017.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 177, + 330, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 177, + 330, + 191 + ], + "score": 1.0, + "content": "URL http://arxiv.org/abs/1706.05296.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 507, + 230 + ], + "lines": [ + { + "bbox": [ + 105, + 195, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 507, + 210 + ], + "score": 1.0, + "content": "Richard S. Sutton. Learning to predict by the methods of temporal differences. Mach. Learn.,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 207, + 508, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 508, + 221 + ], + "score": 1.0, + "content": "3(1):9–44, August 1988. ISSN 0885-6125. doi: 10.1023/A:1022633531479. URL https:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 219, + 319, + 231 + ], + "spans": [ + { + "bbox": [ + 116, + 219, + 319, + 231 + ], + "score": 1.0, + "content": "//doi.org/10.1023/A:1022633531479.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 251 + ], + "score": 1.0, + "content": "Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 249, + 431, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 431, + 261 + ], + "score": 1.0, + "content": "of the tenth international conference on machine learning, pp. 330–337, 1993.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 505, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 507, + 282 + ], + "score": 1.0, + "content": "Hado van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 279, + 491, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 279, + 491, + 291 + ], + "score": 1.0, + "content": "learning. CoRR, abs/1509.06461, 2015. URL http://arxiv.org/abs/1509.06461.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 506, + 310 + ], + "score": 1.0, + "content": "Ziyu Wang, Nando de Freitas, and Marc Lanctot. Dueling network architectures for deep reinforce-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 308, + 507, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 507, + 321 + ], + "score": 1.0, + "content": "ment learning. CoRR, abs/1511.06581, 2015. URL http://arxiv.org/abs/1511.06581.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 104, + 328, + 506, + 350 + ], + "lines": [ + { + "bbox": [ + 104, + 325, + 507, + 341 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 507, + 341 + ], + "score": 1.0, + "content": "Jeff Wu. Hanabi simulation in rust. https://github.com/WuTheFWasThat/hanabi.rs,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 338, + 142, + 350 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 142, + 350 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 507, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 507, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 507, + 95 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, arthur szlam, and Rob Fergus. Learning Multiagent Communica-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 507, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 507, + 107 + ], + "score": 1.0, + "content": "tion with Backpropagation. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 102, + 507, + 119 + ], + "spans": [ + { + "bbox": [ + 114, + 102, + 507, + 119 + ], + "score": 1.0, + "content": "and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "2244–2252. Curran Associates, Inc., 2016. URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 127, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 117, + 127, + 501, + 138 + ], + "score": 1.0, + "content": "6398-learning-multiagent-communication-with-backpropagation.pdf.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 82, + 507, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 506, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinícius Flores Zambaldi,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 507, + 168 + ], + "score": 1.0, + "content": "Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z. Leibo, Karl Tuyls, and Thore Graepel.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 167, + 508, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 508, + 180 + ], + "score": 1.0, + "content": "Value-decomposition networks for cooperative multi-agent learning. CoRR, abs/1706.05296, 2017.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 177, + 330, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 177, + 330, + 191 + ], + "score": 1.0, + "content": "URL http://arxiv.org/abs/1706.05296.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 145, + 508, + 191 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 507, + 230 + ], + "lines": [ + { + "bbox": [ + 105, + 195, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 507, + 210 + ], + "score": 1.0, + "content": "Richard S. Sutton. Learning to predict by the methods of temporal differences. Mach. Learn.,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 207, + 508, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 508, + 221 + ], + "score": 1.0, + "content": "3(1):9–44, August 1988. ISSN 0885-6125. doi: 10.1023/A:1022633531479. URL https:", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 219, + 319, + 231 + ], + "spans": [ + { + "bbox": [ + 116, + 219, + 319, + 231 + ], + "score": 1.0, + "content": "//doi.org/10.1023/A:1022633531479.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 195, + 508, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 251 + ], + "score": 1.0, + "content": "Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 249, + 431, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 431, + 261 + ], + "score": 1.0, + "content": "of the tenth international conference on machine learning, pp. 330–337, 1993.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 237, + 505, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 505, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 507, + 282 + ], + "score": 1.0, + "content": "Hado van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 279, + 491, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 279, + 491, + 291 + ], + "score": 1.0, + "content": "learning. CoRR, abs/1509.06461, 2015. URL http://arxiv.org/abs/1509.06461.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 266, + 507, + 291 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 506, + 310 + ], + "score": 1.0, + "content": "Ziyu Wang, Nando de Freitas, and Marc Lanctot. Dueling network architectures for deep reinforce-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 308, + 507, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 507, + 321 + ], + "score": 1.0, + "content": "ment learning. CoRR, abs/1511.06581, 2015. URL http://arxiv.org/abs/1511.06581.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 298, + 507, + 321 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 328, + 506, + 350 + ], + "lines": [ + { + "bbox": [ + 104, + 325, + 507, + 341 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 507, + 341 + ], + "score": 1.0, + "content": "Jeff Wu. Hanabi simulation in rust. https://github.com/WuTheFWasThat/hanabi.rs,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 338, + 142, + 350 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 142, + 350 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 325, + 507, + 350 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 476, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 478, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 478, + 95 + ], + "score": 1.0, + "content": "A NETWORK ARCHITECTURE AND HYPER-PAMAMETERS FOR HANABI", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 506, + 319 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "score": 1.0, + "content": "Our Hanabi agent uses dueling network architecture (Wang et al., 2015). The main body of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 117, + 507, + 130 + ], + "spans": [ + { + "bbox": [ + 104, + 117, + 507, + 130 + ], + "score": 1.0, + "content": "network consists of 1 fully connected layer of 512 units and 2 LSTM (Hochreiter & Schmidhuber,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "1997) layers of 512 units, followed by two output heads for value and advantages respectively. The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "same network configuration is used across all Hanabi experiments. We take the default featurization", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 150, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 506, + 163 + ], + "score": 1.0, + "content": "of HLE and replace the card knowledge section with the V0-Belief proposed by Foerster et al. (2019).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "score": 1.0, + "content": "The maximum length of an episode is capped at 80 steps and the entire episode is stored in the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "score": 1.0, + "content": "replay buffer as one training sample. This avoids the “slate hidden states” problem as described", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "in Kapturowski et al. (2019) because we can simply initialize the hidden states of LSTM as zero", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 208 + ], + "score": 1.0, + "content": "during training. For exploration and experience prioritization, we follow the simple strategy as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 412, + 219 + ], + "score": 1.0, + "content": "in Horgan et al. (2018) and Kapturowski et al. (2019). Each actor executes an", + "type": "text" + }, + { + "bbox": [ + 412, + 207, + 421, + 216 + ], + "score": 0.84, + "content": "\\epsilon _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 204, + 506, + 219 + ], + "score": 1.0, + "content": "-greedy policy where", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 214, + 507, + 232 + ], + "spans": [ + { + "bbox": [ + 107, + 216, + 167, + 229 + ], + "score": 0.92, + "content": "\\epsilon _ { i } = \\bar { \\epsilon } ^ { 1 + \\frac { 1 } { N - 1 } \\alpha }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 214, + 184, + 232 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 185, + 217, + 262, + 230 + ], + "score": 0.92, + "content": "i \\in \\{ 0 , . . . , N - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 214, + 342, + 232 + ], + "score": 1.0, + "content": "but with a smaller", + "type": "text" + }, + { + "bbox": [ + 342, + 218, + 376, + 228 + ], + "score": 0.89, + "content": "\\epsilon = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 214, + 395, + 232 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 395, + 218, + 423, + 228 + ], + "score": 0.89, + "content": "\\alpha = 7", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 214, + 507, + 232 + ], + "score": 1.0, + "content": ". For simplicity, all", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 229, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 369, + 243 + ], + "score": 1.0, + "content": "players of a game use the same epsilon. The per time-step priority", + "type": "text" + }, + { + "bbox": [ + 370, + 230, + 379, + 240 + ], + "score": 0.87, + "content": "\\delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 229, + 506, + 243 + ], + "score": 1.0, + "content": "is the TD error and per episode", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 233, + 255 + ], + "score": 1.0, + "content": "priority is computed following", + "type": "text" + }, + { + "bbox": [ + 234, + 240, + 343, + 254 + ], + "score": 0.9, + "content": "\\delta _ { e } = \\eta \\operatorname* { m a x } _ { t } \\delta _ { i } + ( 1 - \\eta ) \\hat { \\delta }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 241, + 371, + 255 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 371, + 242, + 404, + 254 + ], + "score": 0.89, + "content": "\\eta = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 241, + 506, + 255 + ], + "score": 1.0, + "content": ". Priority exponent is set", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 369, + 266 + ], + "score": 1.0, + "content": "to 0.9 and importance sampling exponent is set to 0.6. We use", + "type": "text" + }, + { + "bbox": [ + 369, + 255, + 376, + 263 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "-step return (Sutton, 1988) and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "double Q-learning (van Hasselt et al., 2015) for target computation during training. The discount", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 132, + 288 + ], + "score": 1.0, + "content": "factor", + "type": "text" + }, + { + "bbox": [ + 132, + 276, + 140, + 286 + ], + "score": 0.83, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "is set to 0.999. The network is updated using Adam optimizer (Kingma & Ba, 2014) with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 284, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 159, + 299 + ], + "score": 1.0, + "content": "learning rate", + "type": "text" + }, + { + "bbox": [ + 159, + 285, + 232, + 296 + ], + "score": 0.91, + "content": "l r = 6 . 2 5 \\times 1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 284, + 250, + 299 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 251, + 285, + 314, + 296 + ], + "score": 0.92, + "content": "\\epsilon = \\dot { 1 . 5 } \\times 1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 284, + 506, + 299 + ], + "score": 1.0, + "content": ". Trainer sends its network weights to all actors", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "every 10 updates and target network is synchronized with online network every 2500 updates. These", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 309, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 309, + 320 + ], + "score": 1.0, + "content": "hyper-parameters are fixed across all experiments.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "In the baseline, we use Independent Q-Learning where each player estimates the Q value and selects", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "action independently at each time-step. Note that all players need to operate on the observations in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "order to update their recurrent hidden states while only the current player has non-trivial legal moves", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "and other players can only select ‘pass’. Each player then writes its own version of the episode into", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "the prioritized replay buffer and they are sampled independently during training. The prioritized", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 378, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 198, + 393 + ], + "score": 1.0, + "content": "replay buffer contains", + "type": "text" + }, + { + "bbox": [ + 198, + 379, + 249, + 391 + ], + "score": 0.83, + "content": "2 ^ { 1 7 } ( 1 3 1 0 7 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 378, + 506, + 393 + ], + "score": 1.0, + "content": "episodes. We warm up the replay buffer with 10,000 episodes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 389, + 498, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 498, + 403 + ], + "score": 1.0, + "content": "before training starts. Batch size during training is 128 for games of different numbers of players.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "As mentioned in Section 4, the SAD agent is built on top of joint Q-function where the Q value is the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "score": 1.0, + "content": "sum of the individual Q value of all players given their own actions. One episode produces only one", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "training sample with an extra dimension for the number of players. The replay buffer size is reduced", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 116, + 453 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 439, + 131, + 450 + ], + "score": 0.83, + "content": "2 ^ { 1 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 439, + 279, + 453 + ], + "score": 1.0, + "content": "for 2-player and 3-player games and", + "type": "text" + }, + { + "bbox": [ + 279, + 439, + 293, + 450 + ], + "score": 0.83, + "content": "2 ^ { 1 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "for 4-player and 5-player games. The batch sizes for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "score": 1.0, + "content": "2-, 3-, 4-, 5-players are 64, 43, 32, 26 respectively to account for the fact that each sample contains", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 462, + 151, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 151, + 474 + ], + "score": 1.0, + "content": "more data.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Auxiliary task can be added to the agent to help it decode the greedy action more effectively. In", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "Hanabi, the natural choice is the predict the card of player’s own hand. In our experiments, the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "auxiliary task is to predict the status of a card, which can be playable, discardable, or unknown. The", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 511, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 526 + ], + "score": 1.0, + "content": "loss is the average cross entropy loss per card and is simply added to the TD-error of reinforcement", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 522, + 206, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 206, + 537 + ], + "score": 1.0, + "content": "learning during training.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 476, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 478, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 478, + 95 + ], + "score": 1.0, + "content": "A NETWORK ARCHITECTURE AND HYPER-PAMAMETERS FOR HANABI", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 506, + 319 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "score": 1.0, + "content": "Our Hanabi agent uses dueling network architecture (Wang et al., 2015). The main body of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 117, + 507, + 130 + ], + "spans": [ + { + "bbox": [ + 104, + 117, + 507, + 130 + ], + "score": 1.0, + "content": "network consists of 1 fully connected layer of 512 units and 2 LSTM (Hochreiter & Schmidhuber,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "1997) layers of 512 units, followed by two output heads for value and advantages respectively. The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "same network configuration is used across all Hanabi experiments. We take the default featurization", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 150, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 506, + 163 + ], + "score": 1.0, + "content": "of HLE and replace the card knowledge section with the V0-Belief proposed by Foerster et al. (2019).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "score": 1.0, + "content": "The maximum length of an episode is capped at 80 steps and the entire episode is stored in the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "score": 1.0, + "content": "replay buffer as one training sample. This avoids the “slate hidden states” problem as described", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "in Kapturowski et al. (2019) because we can simply initialize the hidden states of LSTM as zero", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 208 + ], + "score": 1.0, + "content": "during training. For exploration and experience prioritization, we follow the simple strategy as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 412, + 219 + ], + "score": 1.0, + "content": "in Horgan et al. (2018) and Kapturowski et al. (2019). Each actor executes an", + "type": "text" + }, + { + "bbox": [ + 412, + 207, + 421, + 216 + ], + "score": 0.84, + "content": "\\epsilon _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 204, + 506, + 219 + ], + "score": 1.0, + "content": "-greedy policy where", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 214, + 507, + 232 + ], + "spans": [ + { + "bbox": [ + 107, + 216, + 167, + 229 + ], + "score": 0.92, + "content": "\\epsilon _ { i } = \\bar { \\epsilon } ^ { 1 + \\frac { 1 } { N - 1 } \\alpha }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 214, + 184, + 232 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 185, + 217, + 262, + 230 + ], + "score": 0.92, + "content": "i \\in \\{ 0 , . . . , N - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 214, + 342, + 232 + ], + "score": 1.0, + "content": "but with a smaller", + "type": "text" + }, + { + "bbox": [ + 342, + 218, + 376, + 228 + ], + "score": 0.89, + "content": "\\epsilon = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 214, + 395, + 232 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 395, + 218, + 423, + 228 + ], + "score": 0.89, + "content": "\\alpha = 7", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 214, + 507, + 232 + ], + "score": 1.0, + "content": ". For simplicity, all", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 229, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 369, + 243 + ], + "score": 1.0, + "content": "players of a game use the same epsilon. The per time-step priority", + "type": "text" + }, + { + "bbox": [ + 370, + 230, + 379, + 240 + ], + "score": 0.87, + "content": "\\delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 229, + 506, + 243 + ], + "score": 1.0, + "content": "is the TD error and per episode", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 233, + 255 + ], + "score": 1.0, + "content": "priority is computed following", + "type": "text" + }, + { + "bbox": [ + 234, + 240, + 343, + 254 + ], + "score": 0.9, + "content": "\\delta _ { e } = \\eta \\operatorname* { m a x } _ { t } \\delta _ { i } + ( 1 - \\eta ) \\hat { \\delta }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 241, + 371, + 255 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 371, + 242, + 404, + 254 + ], + "score": 0.89, + "content": "\\eta = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 241, + 506, + 255 + ], + "score": 1.0, + "content": ". Priority exponent is set", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 369, + 266 + ], + "score": 1.0, + "content": "to 0.9 and importance sampling exponent is set to 0.6. We use", + "type": "text" + }, + { + "bbox": [ + 369, + 255, + 376, + 263 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "-step return (Sutton, 1988) and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "double Q-learning (van Hasselt et al., 2015) for target computation during training. The discount", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 132, + 288 + ], + "score": 1.0, + "content": "factor", + "type": "text" + }, + { + "bbox": [ + 132, + 276, + 140, + 286 + ], + "score": 0.83, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "is set to 0.999. The network is updated using Adam optimizer (Kingma & Ba, 2014) with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 284, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 159, + 299 + ], + "score": 1.0, + "content": "learning rate", + "type": "text" + }, + { + "bbox": [ + 159, + 285, + 232, + 296 + ], + "score": 0.91, + "content": "l r = 6 . 2 5 \\times 1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 284, + 250, + 299 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 251, + 285, + 314, + 296 + ], + "score": 0.92, + "content": "\\epsilon = \\dot { 1 . 5 } \\times 1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 284, + 506, + 299 + ], + "score": 1.0, + "content": ". Trainer sends its network weights to all actors", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "every 10 updates and target network is synchronized with online network every 2500 updates. These", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 309, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 309, + 320 + ], + "score": 1.0, + "content": "hyper-parameters are fixed across all experiments.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 106, + 507, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "In the baseline, we use Independent Q-Learning where each player estimates the Q value and selects", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "action independently at each time-step. Note that all players need to operate on the observations in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "order to update their recurrent hidden states while only the current player has non-trivial legal moves", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "and other players can only select ‘pass’. Each player then writes its own version of the episode into", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 381 + ], + "score": 1.0, + "content": "the prioritized replay buffer and they are sampled independently during training. The prioritized", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 378, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 198, + 393 + ], + "score": 1.0, + "content": "replay buffer contains", + "type": "text" + }, + { + "bbox": [ + 198, + 379, + 249, + 391 + ], + "score": 0.83, + "content": "2 ^ { 1 7 } ( 1 3 1 0 7 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 378, + 506, + 393 + ], + "score": 1.0, + "content": "episodes. We warm up the replay buffer with 10,000 episodes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 389, + 498, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 498, + 403 + ], + "score": 1.0, + "content": "before training starts. Batch size during training is 128 for games of different numbers of players.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 325, + 506, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "As mentioned in Section 4, the SAD agent is built on top of joint Q-function where the Q value is the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "score": 1.0, + "content": "sum of the individual Q value of all players given their own actions. One episode produces only one", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "training sample with an extra dimension for the number of players. The replay buffer size is reduced", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 116, + 453 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 439, + 131, + 450 + ], + "score": 0.83, + "content": "2 ^ { 1 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 439, + 279, + 453 + ], + "score": 1.0, + "content": "for 2-player and 3-player games and", + "type": "text" + }, + { + "bbox": [ + 279, + 439, + 293, + 450 + ], + "score": 0.83, + "content": "2 ^ { 1 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "for 4-player and 5-player games. The batch sizes for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "score": 1.0, + "content": "2-, 3-, 4-, 5-players are 64, 43, 32, 26 respectively to account for the fact that each sample contains", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 462, + 151, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 151, + 474 + ], + "score": 1.0, + "content": "more data.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 407, + 506, + 474 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Auxiliary task can be added to the agent to help it decode the greedy action more effectively. In", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "Hanabi, the natural choice is the predict the card of player’s own hand. In our experiments, the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "auxiliary task is to predict the status of a card, which can be playable, discardable, or unknown. The", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 511, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 526 + ], + "score": 1.0, + "content": "loss is the average cross entropy loss per card and is simply added to the TD-error of reinforcement", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 522, + 206, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 206, + 537 + ], + "score": 1.0, + "content": "learning during training.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 479, + 506, + 537 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 296, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 299, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 299, + 96 + ], + "score": 1.0, + "content": "B LEARNING CURVES FOR HANABI", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 140, + 497, + 491 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 107, + 106, + 504, + 128 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 105, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 120 + ], + "score": 1.0, + "content": "Figure 3 shows learning curves of different algorithms averaged over 13 seeds per algorithm per", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 284, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 284, + 130 + ], + "score": 1.0, + "content": "player setting. Shading is error of the mean.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "image_body", + "bbox": [ + 110, + 140, + 497, + 491 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 140, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 110, + 140, + 497, + 491 + ], + "score": 0.975, + "type": "image", + "image_path": "ccf9077cc9f7d726e6a7f0ea3a2fa7941948d5e7bedb2d3bea373709b6945a93.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 110, + 140, + 497, + 257.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 110, + 257.0, + 497, + 374.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 374.0, + 497, + 491.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 254, + 501, + 356, + 513 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 254, + 500, + 357, + 515 + ], + "spans": [ + { + "bbox": [ + 254, + 500, + 357, + 515 + ], + "score": 1.0, + "content": "Figure 3: Learning Curves", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 4 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 296, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 299, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 299, + 96 + ], + "score": 1.0, + "content": "B LEARNING CURVES FOR HANABI", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 140, + 497, + 491 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 107, + 106, + 504, + 128 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 105, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 120 + ], + "score": 1.0, + "content": "Figure 3 shows learning curves of different algorithms averaged over 13 seeds per algorithm per", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 284, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 284, + 130 + ], + "score": 1.0, + "content": "player setting. Shading is error of the mean.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "image_body", + "bbox": [ + 110, + 140, + 497, + 491 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 140, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 110, + 140, + 497, + 491 + ], + "score": 0.975, + "type": "image", + "image_path": "ccf9077cc9f7d726e6a7f0ea3a2fa7941948d5e7bedb2d3bea373709b6945a93.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 110, + 140, + 497, + 257.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 110, + 257.0, + 497, + 374.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 374.0, + 497, + 491.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 254, + 501, + 356, + 513 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 254, + 500, + 357, + 515 + ], + "spans": [ + { + "bbox": [ + 254, + 500, + 357, + 515 + ], + "score": 1.0, + "content": "Figure 3: Learning Curves", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 4 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/B1xm3RVtwB/B1xm3RVtwB_model.json b/parse/train/B1xm3RVtwB/B1xm3RVtwB_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ae033d4f1f492362345e58e203ea9b37428a94cc --- /dev/null +++ b/parse/train/B1xm3RVtwB/B1xm3RVtwB_model.json @@ -0,0 +1,16059 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 397, + 618, + 1302, + 618, + 1302, + 1320, + 397, + 1320 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 299, + 1650, + 1404, + 1650, + 1404, + 1957, + 299, + 1957 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1452, + 1404, + 1452, + 1404, + 1635, + 299, + 1635 + ], + "score": 0.978 + }, + { + "category_id": 0, + "poly": [ + 301, + 219, + 1330, + 219, + 1330, + 324, + 301, + 324 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1399, + 1972, + 1399, + 2034, + 298, + 2034 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 314, + 374, + 699, + 374, + 699, + 469, + 314, + 469 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 302, + 1381, + 573, + 1381, + 573, + 1416, + 302, + 1416 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 773, + 550, + 927, + 550, + 927, + 583, + 773, + 583 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.724 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 216.0, + 1026.0, + 216.0, + 1026.0, + 273.0, + 296.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 275.0, + 1333.0, + 275.0, + 1333.0, + 328.0, + 296.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1378.0, + 579.0, + 1378.0, + 579.0, + 1425.0, + 294.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 546.0, + 934.0, + 546.0, + 934.0, + 589.0, + 768.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 618.0, + 1305.0, + 618.0, + 1305.0, + 655.0, + 392.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 650.0, + 1306.0, + 650.0, + 1306.0, + 686.0, + 393.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 678.0, + 1304.0, + 678.0, + 1304.0, + 717.0, + 392.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 711.0, + 1305.0, + 711.0, + 1305.0, + 746.0, + 394.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 741.0, + 1307.0, + 741.0, + 1307.0, + 777.0, + 393.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 774.0, + 1306.0, + 774.0, + 1306.0, + 806.0, + 395.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 804.0, + 1306.0, + 804.0, + 1306.0, + 834.0, + 393.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 833.0, + 1306.0, + 833.0, + 1306.0, + 869.0, + 394.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 862.0, + 1305.0, + 862.0, + 1305.0, + 898.0, + 393.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 895.0, + 1305.0, + 895.0, + 1305.0, + 927.0, + 395.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 924.0, + 1305.0, + 924.0, + 1305.0, + 959.0, + 394.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 954.0, + 1306.0, + 954.0, + 1306.0, + 990.0, + 393.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 985.0, + 1307.0, + 985.0, + 1307.0, + 1021.0, + 391.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1013.0, + 1305.0, + 1013.0, + 1305.0, + 1053.0, + 392.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1047.0, + 1305.0, + 1047.0, + 1305.0, + 1080.0, + 395.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1076.0, + 1307.0, + 1076.0, + 1307.0, + 1114.0, + 393.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1108.0, + 1307.0, + 1108.0, + 1307.0, + 1143.0, + 395.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1139.0, + 1305.0, + 1139.0, + 1305.0, + 1172.0, + 394.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1167.0, + 1306.0, + 1167.0, + 1306.0, + 1203.0, + 392.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1200.0, + 1305.0, + 1200.0, + 1305.0, + 1232.0, + 393.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1229.0, + 1305.0, + 1229.0, + 1305.0, + 1261.0, + 394.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1259.0, + 1306.0, + 1259.0, + 1306.0, + 1294.0, + 395.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1291.0, + 1175.0, + 1291.0, + 1175.0, + 1323.0, + 394.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1652.0, + 1404.0, + 1652.0, + 1404.0, + 1684.0, + 295.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 1405.0, + 1681.0, + 1405.0, + 1717.0, + 294.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1714.0, + 1406.0, + 1714.0, + 1406.0, + 1746.0, + 295.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1744.0, + 1403.0, + 1744.0, + 1403.0, + 1776.0, + 295.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1773.0, + 1406.0, + 1773.0, + 1406.0, + 1807.0, + 293.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1802.0, + 1406.0, + 1802.0, + 1406.0, + 1839.0, + 293.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1834.0, + 1408.0, + 1834.0, + 1408.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1862.0, + 1408.0, + 1862.0, + 1408.0, + 1900.0, + 293.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1894.0, + 1406.0, + 1894.0, + 1406.0, + 1933.0, + 293.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1926.0, + 791.0, + 1926.0, + 791.0, + 1957.0, + 295.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1447.0, + 1405.0, + 1447.0, + 1405.0, + 1490.0, + 293.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1482.0, + 1406.0, + 1482.0, + 1406.0, + 1518.0, + 293.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1513.0, + 1409.0, + 1513.0, + 1409.0, + 1550.0, + 292.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1544.0, + 1409.0, + 1544.0, + 1409.0, + 1578.0, + 293.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1573.0, + 1408.0, + 1573.0, + 1408.0, + 1608.0, + 294.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1604.0, + 681.0, + 1604.0, + 681.0, + 1638.0, + 294.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1970.0, + 1403.0, + 1970.0, + 1403.0, + 2006.0, + 296.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2038.0, + 296.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 377.0, + 703.0, + 377.0, + 703.0, + 408.0, + 312.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 406.0, + 693.0, + 406.0, + 693.0, + 440.0, + 312.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 440.0, + 669.0, + 440.0, + 669.0, + 471.0, + 317.0, + 471.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1257, + 1404, + 1257, + 1404, + 1532, + 298, + 1532 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 706, + 1405, + 706, + 1405, + 1041, + 298, + 1041 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1547, + 1404, + 1547, + 1404, + 1792, + 298, + 1792 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1807, + 1404, + 1807, + 1404, + 1992, + 298, + 1992 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1057, + 1403, + 1057, + 1403, + 1241, + 299, + 1241 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 308, + 1402, + 308, + 1402, + 459, + 298, + 459 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 476, + 1405, + 476, + 1405, + 691, + 298, + 691 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1399, + 229, + 1399, + 293, + 294, + 293 + ], + "score": 0.955 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.712 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.098 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1255.0, + 1405.0, + 1255.0, + 1405.0, + 1294.0, + 293.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1286.0, + 1406.0, + 1286.0, + 1406.0, + 1324.0, + 292.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1318.0, + 1406.0, + 1318.0, + 1406.0, + 1354.0, + 293.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1348.0, + 1405.0, + 1348.0, + 1405.0, + 1383.0, + 293.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1379.0, + 1406.0, + 1379.0, + 1406.0, + 1414.0, + 292.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1411.0, + 1404.0, + 1411.0, + 1404.0, + 1444.0, + 296.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1440.0, + 1405.0, + 1440.0, + 1405.0, + 1477.0, + 294.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1469.0, + 1408.0, + 1469.0, + 1408.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1499.0, + 777.0, + 1499.0, + 777.0, + 1539.0, + 293.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 707.0, + 1405.0, + 707.0, + 1405.0, + 741.0, + 296.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 738.0, + 1403.0, + 738.0, + 1403.0, + 772.0, + 295.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 766.0, + 1406.0, + 766.0, + 1406.0, + 802.0, + 293.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 1407.0, + 798.0, + 1407.0, + 833.0, + 295.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 828.0, + 1406.0, + 828.0, + 1406.0, + 864.0, + 293.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 859.0, + 1406.0, + 859.0, + 1406.0, + 893.0, + 295.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 888.0, + 1407.0, + 888.0, + 1407.0, + 924.0, + 293.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 922.0, + 1405.0, + 922.0, + 1405.0, + 953.0, + 295.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 953.0, + 1405.0, + 953.0, + 1405.0, + 983.0, + 296.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 980.0, + 1403.0, + 980.0, + 1403.0, + 1017.0, + 293.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1011.0, + 693.0, + 1011.0, + 693.0, + 1045.0, + 295.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1545.0, + 1404.0, + 1545.0, + 1404.0, + 1583.0, + 292.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1580.0, + 1405.0, + 1580.0, + 1405.0, + 1613.0, + 293.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1607.0, + 1404.0, + 1607.0, + 1404.0, + 1643.0, + 293.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1638.0, + 1406.0, + 1638.0, + 1406.0, + 1675.0, + 293.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1668.0, + 1404.0, + 1668.0, + 1404.0, + 1704.0, + 292.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1737.0, + 292.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1728.0, + 1405.0, + 1728.0, + 1405.0, + 1769.0, + 292.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1759.0, + 998.0, + 1759.0, + 998.0, + 1795.0, + 293.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1808.0, + 1404.0, + 1808.0, + 1404.0, + 1840.0, + 297.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1836.0, + 1405.0, + 1836.0, + 1405.0, + 1874.0, + 293.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1868.0, + 1406.0, + 1868.0, + 1406.0, + 1903.0, + 293.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1899.0, + 1408.0, + 1899.0, + 1408.0, + 1935.0, + 293.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1929.0, + 1405.0, + 1929.0, + 1405.0, + 1964.0, + 294.0, + 1964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1958.0, + 1322.0, + 1958.0, + 1322.0, + 1999.0, + 292.0, + 1999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1058.0, + 1405.0, + 1058.0, + 1405.0, + 1093.0, + 294.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1088.0, + 1405.0, + 1088.0, + 1405.0, + 1124.0, + 294.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1120.0, + 1404.0, + 1120.0, + 1404.0, + 1152.0, + 295.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1149.0, + 1404.0, + 1149.0, + 1404.0, + 1185.0, + 295.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1178.0, + 1404.0, + 1178.0, + 1404.0, + 1214.0, + 295.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1211.0, + 800.0, + 1211.0, + 800.0, + 1245.0, + 294.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 308.0, + 1407.0, + 308.0, + 1407.0, + 344.0, + 294.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 336.0, + 1406.0, + 336.0, + 1406.0, + 374.0, + 293.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 365.0, + 1406.0, + 365.0, + 1406.0, + 404.0, + 293.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1406.0, + 397.0, + 1406.0, + 436.0, + 293.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 428.0, + 505.0, + 428.0, + 505.0, + 465.0, + 295.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 475.0, + 1405.0, + 475.0, + 1405.0, + 510.0, + 293.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 506.0, + 1405.0, + 506.0, + 1405.0, + 543.0, + 293.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 535.0, + 1407.0, + 535.0, + 1407.0, + 572.0, + 293.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 570.0, + 1407.0, + 570.0, + 1407.0, + 602.0, + 293.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 597.0, + 1407.0, + 597.0, + 1407.0, + 635.0, + 293.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 630.0, + 1403.0, + 630.0, + 1403.0, + 665.0, + 295.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 661.0, + 1212.0, + 661.0, + 1212.0, + 695.0, + 296.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1327.0, + 261.0, + 1327.0, + 297.0, + 295.0, + 297.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 935, + 1405, + 935, + 1405, + 1332, + 298, + 1332 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 307, + 1405, + 307, + 1405, + 583, + 298, + 583 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1649, + 1405, + 1649, + 1405, + 1866, + 298, + 1866 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1404, + 1881, + 1404, + 2034, + 299, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 766, + 1404, + 766, + 1404, + 921, + 299, + 921 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 598, + 1403, + 598, + 1403, + 752, + 299, + 752 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1348, + 1403, + 1348, + 1403, + 1441, + 299, + 1441 + ], + "score": 0.97 + }, + { + "category_id": 0, + "poly": [ + 300, + 1504, + 557, + 1504, + 557, + 1541, + 300, + 1541 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 300, + 225, + 587, + 225, + 587, + 261, + 300, + 261 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 299, + 1584, + 472, + 1584, + 472, + 1617, + 299, + 1617 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.592 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.425 + }, + { + "category_id": 13, + "poly": [ + 1237, + 1681, + 1396, + 1681, + 1396, + 1715, + 1237, + 1715 + ], + "score": 0.92, + "latex": "o _ { t } ^ { a } = { \\bar { O } } ( s _ { t } , a )" + }, + { + "category_id": 13, + "poly": [ + 861, + 1713, + 956, + 1713, + 956, + 1746, + 861, + 1746 + ], + "score": 0.92, + "latex": "O ( s _ { t } , a )" + }, + { + "category_id": 13, + "poly": [ + 972, + 1911, + 1157, + 1911, + 1157, + 1946, + 972, + 1946 + ], + "score": 0.92, + "latex": "u ^ { a } \\sim \\pi _ { \\theta } ^ { a } ( u ^ { a } | \\tau ^ { a } )" + }, + { + "category_id": 13, + "poly": [ + 369, + 1713, + 448, + 1713, + 448, + 1743, + 369, + 1743 + ], + "score": 0.92, + "latex": "s _ { t } \\in S" + }, + { + "category_id": 13, + "poly": [ + 1096, + 1943, + 1397, + 1943, + 1397, + 1976, + 1096, + 1976 + ], + "score": 0.89, + "latex": "\\tau ^ { a } = \\{ o _ { 0 } ^ { a } , u _ { 0 } ^ { a } , r _ { 1 } , . . r _ { T } , o _ { T } ^ { a } \\}" + }, + { + "category_id": 13, + "poly": [ + 857, + 1683, + 964, + 1683, + 964, + 1710, + 857, + 1710 + ], + "score": 0.89, + "latex": "a \\in 1 . . N" + }, + { + "category_id": 13, + "poly": [ + 544, + 1913, + 576, + 1913, + 576, + 1945, + 544, + 1945 + ], + "score": 0.88, + "latex": "u _ { t } ^ { a }" + }, + { + "category_id": 13, + "poly": [ + 580, + 1944, + 611, + 1944, + 611, + 1970, + 580, + 1970 + ], + "score": 0.88, + "latex": "\\tau ^ { a }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1089, + 339, + 1089, + 339, + 1117, + 297, + 1117 + ], + "score": 0.84, + "latex": "0 \\%" + }, + { + "category_id": 13, + "poly": [ + 298, + 1974, + 321, + 1974, + 321, + 2000, + 298, + 2000 + ], + "score": 0.81, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 684, + 1974, + 700, + 1974, + 700, + 2000, + 684, + 2000 + ], + "score": 0.81, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1057, + 1653, + 1085, + 1653, + 1085, + 1679, + 1057, + 1679 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1219, + 1155, + 1236, + 1155, + 1236, + 1178, + 1219, + 1178 + ], + "score": 0.67, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 537, + 1155, + 552, + 1155, + 552, + 1177, + 537, + 1177 + ], + "score": 0.62, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1067, + 1950, + 1083, + 1950, + 1083, + 1970, + 1067, + 1970 + ], + "score": 0.59, + "latex": "a" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1501.0, + 561.0, + 1501.0, + 561.0, + 1548.0, + 291.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 222.0, + 594.0, + 222.0, + 594.0, + 269.0, + 291.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1582.0, + 475.0, + 1582.0, + 475.0, + 1621.0, + 294.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 937.0, + 1403.0, + 937.0, + 1403.0, + 969.0, + 293.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 967.0, + 1405.0, + 967.0, + 1405.0, + 1001.0, + 292.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 996.0, + 1403.0, + 996.0, + 1403.0, + 1033.0, + 295.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1025.0, + 1405.0, + 1025.0, + 1405.0, + 1064.0, + 292.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1057.0, + 1406.0, + 1057.0, + 1406.0, + 1093.0, + 293.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1086.0, + 296.0, + 1086.0, + 296.0, + 1124.0, + 292.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1086.0, + 1405.0, + 1086.0, + 1405.0, + 1124.0, + 340.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1117.0, + 1405.0, + 1117.0, + 1405.0, + 1156.0, + 292.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1149.0, + 536.0, + 1149.0, + 536.0, + 1185.0, + 293.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1149.0, + 1218.0, + 1149.0, + 1218.0, + 1185.0, + 553.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1149.0, + 1407.0, + 1149.0, + 1407.0, + 1185.0, + 1237.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1177.0, + 1407.0, + 1177.0, + 1407.0, + 1216.0, + 293.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1210.0, + 1405.0, + 1210.0, + 1405.0, + 1246.0, + 293.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1240.0, + 1405.0, + 1240.0, + 1405.0, + 1276.0, + 293.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1269.0, + 1405.0, + 1269.0, + 1405.0, + 1308.0, + 293.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1303.0, + 947.0, + 1303.0, + 947.0, + 1332.0, + 296.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 308.0, + 1406.0, + 308.0, + 1406.0, + 344.0, + 296.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 339.0, + 1406.0, + 339.0, + 1406.0, + 372.0, + 295.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 368.0, + 1406.0, + 368.0, + 1406.0, + 405.0, + 295.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 399.0, + 1405.0, + 399.0, + 1405.0, + 436.0, + 293.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 427.0, + 1408.0, + 427.0, + 1408.0, + 466.0, + 293.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 458.0, + 1406.0, + 458.0, + 1406.0, + 497.0, + 293.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 492.0, + 1407.0, + 492.0, + 1407.0, + 525.0, + 296.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 523.0, + 1405.0, + 523.0, + 1405.0, + 556.0, + 296.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 553.0, + 914.0, + 553.0, + 914.0, + 586.0, + 296.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1651.0, + 1056.0, + 1651.0, + 1056.0, + 1686.0, + 295.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1651.0, + 1405.0, + 1651.0, + 1405.0, + 1686.0, + 1086.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1677.0, + 856.0, + 1677.0, + 856.0, + 1718.0, + 292.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1677.0, + 1236.0, + 1677.0, + 1236.0, + 1718.0, + 965.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1677.0, + 1409.0, + 1677.0, + 1409.0, + 1718.0, + 1397.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 368.0, + 1712.0, + 368.0, + 1746.0, + 295.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1712.0, + 860.0, + 1712.0, + 860.0, + 1746.0, + 449.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1712.0, + 1408.0, + 1712.0, + 1408.0, + 1746.0, + 957.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1744.0, + 1403.0, + 1744.0, + 1403.0, + 1775.0, + 297.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1774.0, + 1405.0, + 1774.0, + 1405.0, + 1808.0, + 296.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1802.0, + 1406.0, + 1802.0, + 1406.0, + 1839.0, + 292.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1833.0, + 562.0, + 1833.0, + 562.0, + 1872.0, + 294.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1878.0, + 1404.0, + 1878.0, + 1404.0, + 1918.0, + 293.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 543.0, + 1909.0, + 543.0, + 1950.0, + 294.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1909.0, + 971.0, + 1909.0, + 971.0, + 1950.0, + 577.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1950.0, + 1158.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1938.0, + 579.0, + 1938.0, + 579.0, + 1981.0, + 292.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1938.0, + 1066.0, + 1938.0, + 1066.0, + 1981.0, + 612.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1938.0, + 1095.0, + 1938.0, + 1095.0, + 1981.0, + 1084.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1938.0, + 1408.0, + 1938.0, + 1408.0, + 1981.0, + 1398.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 297.0, + 1969.0, + 297.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1969.0, + 683.0, + 1969.0, + 683.0, + 2009.0, + 322.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1969.0, + 1406.0, + 1969.0, + 1406.0, + 2009.0, + 701.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1408.0, + 2001.0, + 1408.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 766.0, + 1404.0, + 766.0, + 1404.0, + 801.0, + 295.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 794.0, + 1406.0, + 794.0, + 1406.0, + 834.0, + 293.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 829.0, + 1405.0, + 829.0, + 1405.0, + 862.0, + 295.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 858.0, + 1404.0, + 858.0, + 1404.0, + 893.0, + 294.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 889.0, + 1408.0, + 889.0, + 1408.0, + 926.0, + 294.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 596.0, + 1404.0, + 596.0, + 1404.0, + 634.0, + 294.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 630.0, + 1404.0, + 630.0, + 1404.0, + 663.0, + 297.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 658.0, + 1408.0, + 658.0, + 1408.0, + 695.0, + 292.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 688.0, + 1405.0, + 688.0, + 1405.0, + 725.0, + 294.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 719.0, + 1319.0, + 719.0, + 1319.0, + 758.0, + 293.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1346.0, + 1407.0, + 1346.0, + 1407.0, + 1384.0, + 293.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1379.0, + 1404.0, + 1379.0, + 1404.0, + 1413.0, + 294.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1409.0, + 1074.0, + 1409.0, + 1074.0, + 1446.0, + 294.0, + 1446.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 569, + 1404, + 569, + 1404, + 815, + 298, + 815 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 829, + 1405, + 829, + 1405, + 1074, + 298, + 1074 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1309, + 1404, + 1309, + 1404, + 1525, + 297, + 1525 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1402, + 1880, + 1402, + 2036, + 299, + 2036 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1539, + 1404, + 1539, + 1404, + 1663, + 298, + 1663 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 308, + 1403, + 308, + 1403, + 443, + 298, + 443 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1090, + 1405, + 1090, + 1405, + 1182, + 299, + 1182 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 295, + 229, + 1402, + 229, + 1402, + 293, + 295, + 293 + ], + "score": 0.948 + }, + { + "category_id": 0, + "poly": [ + 299, + 1730, + 480, + 1730, + 480, + 1766, + 299, + 1766 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 299, + 1813, + 922, + 1813, + 922, + 1845, + 299, + 1845 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 297, + 502, + 1043, + 502, + 1043, + 534, + 297, + 534 + ], + "score": 0.787 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 300, + 1243, + 1345, + 1243, + 1345, + 1276, + 300, + 1276 + ], + "score": 0.576 + }, + { + "category_id": 0, + "poly": [ + 300, + 1243, + 1345, + 1243, + 1345, + 1276, + 300, + 1276 + ], + "score": 0.363 + }, + { + "category_id": 1, + "poly": [ + 297, + 502, + 1043, + 502, + 1043, + 534, + 297, + 534 + ], + "score": 0.111 + }, + { + "category_id": 13, + "poly": [ + 737, + 339, + 811, + 339, + 811, + 373, + 737, + 373 + ], + "score": 0.94, + "latex": "R _ { 0 } ( \\tau )" + }, + { + "category_id": 13, + "poly": [ + 383, + 339, + 644, + 339, + 644, + 375, + 383, + 375 + ], + "score": 0.93, + "latex": "J _ { \\theta } = \\mathbb { E } _ { \\tau \\sim P ( \\tau \\mid \\theta ) } R _ { 0 } ( \\tau )" + }, + { + "category_id": 13, + "poly": [ + 924, + 230, + 1214, + 230, + 1214, + 264, + 924, + 264 + ], + "score": 0.92, + "latex": "\\tau = \\{ s _ { 0 } , \\mathbf { u _ { 0 } } , r _ { 1 } , . . r _ { T } , s _ { T } \\}" + }, + { + "category_id": 13, + "poly": [ + 297, + 373, + 475, + 373, + 475, + 415, + 297, + 415 + ], + "score": 0.92, + "latex": "\\sum _ { t ^ { \\prime } \\geq t } \\gamma ^ { t ^ { \\prime } - t } r _ { t ^ { \\prime } } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 632, + 500, + 632, + 500, + 668, + 297, + 668 + ], + "score": 0.91, + "latex": "\\mathbb { E } _ { \\tau \\sim P ( \\tau | s , u ) } R _ { t } ( \\tau )" + }, + { + "category_id": 13, + "poly": [ + 1296, + 338, + 1403, + 338, + 1403, + 373, + 1296, + 373 + ], + "score": 0.91, + "latex": "R _ { t } \\bar { ( } \\tau ) ~ =" + }, + { + "category_id": 13, + "poly": [ + 631, + 2002, + 845, + 2002, + 845, + 2037, + 631, + 2037 + ], + "score": 0.91, + "latex": "B ( \\tau _ { t } ) = P ( \\tau _ { t } | \\tau _ { t } ^ { a } )" + }, + { + "category_id": 13, + "poly": [ + 612, + 722, + 1072, + 722, + 1072, + 755, + 612, + 755 + ], + "score": 0.91, + "latex": "Q ( s , u ) = \\mathbb { E } _ { s ^ { \\prime } } [ r _ { t + 1 } + \\gamma \\operatorname* { m a x } _ { u ^ { \\prime } } Q ( s ^ { \\prime } , u ^ { \\prime } ) ]" + }, + { + "category_id": 13, + "poly": [ + 1278, + 601, + 1403, + 601, + 1403, + 632, + 1278, + 632 + ], + "score": 0.89, + "latex": "Q ( s , u ) =" + }, + { + "category_id": 13, + "poly": [ + 584, + 632, + 882, + 632, + 882, + 665, + 584, + 665 + ], + "score": 0.88, + "latex": "\\tau = \\{ s _ { t } , u _ { t } , r _ { t + 1 } , . . . , s _ { T } \\}" + }, + { + "category_id": 13, + "poly": [ + 891, + 634, + 974, + 634, + 974, + 663, + 891, + 663 + ], + "score": 0.85, + "latex": "u _ { t } = u" + }, + { + "category_id": 13, + "poly": [ + 1301, + 234, + 1333, + 234, + 1333, + 261, + 1301, + 261 + ], + "score": 0.85, + "latex": "\\mathbf { u _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 468, + 236, + 494, + 236, + 494, + 261, + 468, + 261 + ], + "score": 0.83, + "latex": "\\tau _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1240, + 577, + 1284, + 577, + 1284, + 601, + 1240, + 601 + ], + "score": 0.82, + "latex": "s , u" + }, + { + "category_id": 13, + "poly": [ + 525, + 381, + 546, + 381, + 546, + 410, + 525, + 410 + ], + "score": 0.82, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 1200, + 1976, + 1225, + 1976, + 1225, + 2003, + 1200, + 2003 + ], + "score": 0.81, + "latex": "\\tau _ { t }" + }, + { + "category_id": 13, + "poly": [ + 433, + 414, + 449, + 414, + 449, + 439, + 433, + 439 + ], + "score": 0.79, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 593, + 2004, + 624, + 2004, + 624, + 2036, + 593, + 2036 + ], + "score": 0.73, + "latex": "\\tau _ { t } ^ { a }" + }, + { + "category_id": 13, + "poly": [ + 430, + 697, + 446, + 697, + 446, + 719, + 430, + 719 + ], + "score": 0.71, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 298, + 668, + 310, + 668, + 310, + 689, + 298, + 689 + ], + "score": 0.71, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1022, + 632, + 1404, + 632, + 1404, + 665, + 1022, + 665 + ], + "score": 0.7, + "latex": "u _ { t ^ { \\prime } } = \\arg \\operatorname* { m a x } _ { u ^ { \\prime } } Q ( s _ { t ^ { \\prime } } , u ^ { \\prime } ) , \\forall t ^ { \\prime } >" + }, + { + "category_id": 13, + "poly": [ + 737, + 669, + 752, + 669, + 752, + 689, + 737, + 689 + ], + "score": 0.63, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1727.0, + 485.0, + 1727.0, + 485.0, + 1773.0, + 291.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1813.0, + 925.0, + 1813.0, + 925.0, + 1849.0, + 294.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 499.0, + 1048.0, + 499.0, + 1048.0, + 540.0, + 293.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 1347.0, + 1241.0, + 1347.0, + 1280.0, + 294.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 567.0, + 1239.0, + 567.0, + 1239.0, + 607.0, + 292.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 567.0, + 1406.0, + 567.0, + 1406.0, + 607.0, + 1285.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 601.0, + 1277.0, + 601.0, + 1277.0, + 634.0, + 294.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 627.0, + 296.0, + 627.0, + 296.0, + 669.0, + 289.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 627.0, + 583.0, + 627.0, + 583.0, + 669.0, + 501.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 627.0, + 890.0, + 627.0, + 890.0, + 669.0, + 883.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 627.0, + 1021.0, + 627.0, + 1021.0, + 669.0, + 975.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 660.0, + 297.0, + 660.0, + 297.0, + 696.0, + 292.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 660.0, + 736.0, + 660.0, + 736.0, + 696.0, + 311.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 696.0, + 753.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 691.0, + 429.0, + 691.0, + 429.0, + 727.0, + 292.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 691.0, + 1405.0, + 691.0, + 1405.0, + 727.0, + 447.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 722.0, + 611.0, + 722.0, + 611.0, + 759.0, + 293.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 722.0, + 1405.0, + 722.0, + 1405.0, + 759.0, + 1073.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 753.0, + 1406.0, + 753.0, + 1406.0, + 787.0, + 294.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 781.0, + 1410.0, + 781.0, + 1410.0, + 822.0, + 293.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 831.0, + 1407.0, + 831.0, + 1407.0, + 864.0, + 295.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 859.0, + 1405.0, + 859.0, + 1405.0, + 895.0, + 293.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 892.0, + 1406.0, + 892.0, + 1406.0, + 928.0, + 292.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 922.0, + 1406.0, + 922.0, + 1406.0, + 956.0, + 292.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 950.0, + 1405.0, + 950.0, + 1405.0, + 987.0, + 292.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 981.0, + 1408.0, + 981.0, + 1408.0, + 1019.0, + 293.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1013.0, + 1407.0, + 1013.0, + 1407.0, + 1050.0, + 292.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1045.0, + 789.0, + 1045.0, + 789.0, + 1072.0, + 294.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1306.0, + 1405.0, + 1306.0, + 1405.0, + 1349.0, + 292.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1338.0, + 1405.0, + 1338.0, + 1405.0, + 1379.0, + 292.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1371.0, + 1405.0, + 1371.0, + 1405.0, + 1406.0, + 295.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1398.0, + 1405.0, + 1398.0, + 1405.0, + 1439.0, + 291.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1433.0, + 1405.0, + 1433.0, + 1405.0, + 1468.0, + 294.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1461.0, + 1404.0, + 1461.0, + 1404.0, + 1498.0, + 292.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1495.0, + 937.0, + 1495.0, + 937.0, + 1529.0, + 295.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1915.0, + 295.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 1404.0, + 1913.0, + 1404.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1199.0, + 1971.0, + 1199.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 1226.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 592.0, + 2002.0, + 592.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 2002.0, + 630.0, + 2002.0, + 630.0, + 2039.0, + 625.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 2002.0, + 856.0, + 2002.0, + 856.0, + 2039.0, + 846.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1538.0, + 1410.0, + 1538.0, + 1410.0, + 1578.0, + 292.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1569.0, + 1406.0, + 1569.0, + 1406.0, + 1605.0, + 295.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1603.0, + 1404.0, + 1603.0, + 1404.0, + 1636.0, + 294.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 695.0, + 1633.0, + 695.0, + 1665.0, + 294.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 306.0, + 1405.0, + 306.0, + 1405.0, + 344.0, + 293.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 338.0, + 382.0, + 338.0, + 382.0, + 376.0, + 292.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 338.0, + 736.0, + 338.0, + 736.0, + 376.0, + 645.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 338.0, + 1295.0, + 338.0, + 1295.0, + 376.0, + 812.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 368.0, + 296.0, + 368.0, + 296.0, + 420.0, + 288.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 368.0, + 524.0, + 368.0, + 524.0, + 420.0, + 476.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 368.0, + 1409.0, + 368.0, + 1409.0, + 420.0, + 547.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 411.0, + 432.0, + 411.0, + 432.0, + 446.0, + 292.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 411.0, + 878.0, + 411.0, + 878.0, + 446.0, + 450.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1088.0, + 1407.0, + 1088.0, + 1407.0, + 1124.0, + 294.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1118.0, + 1406.0, + 1118.0, + 1406.0, + 1156.0, + 293.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1151.0, + 670.0, + 1151.0, + 670.0, + 1185.0, + 293.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 225.0, + 467.0, + 225.0, + 467.0, + 269.0, + 292.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 225.0, + 923.0, + 225.0, + 923.0, + 269.0, + 495.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 225.0, + 1300.0, + 225.0, + 1300.0, + 269.0, + 1215.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 225.0, + 1407.0, + 225.0, + 1407.0, + 269.0, + 1334.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 579.0, + 261.0, + 579.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 1347.0, + 1241.0, + 1347.0, + 1280.0, + 294.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 499.0, + 1048.0, + 499.0, + 1048.0, + 540.0, + 293.0, + 540.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 637, + 1405, + 637, + 1405, + 824, + 297, + 824 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1027, + 1405, + 1027, + 1405, + 1220, + 297, + 1220 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 296, + 489, + 1407, + 489, + 1407, + 624, + 296, + 624 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1379, + 1405, + 1379, + 1405, + 1503, + 298, + 1503 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1936, + 1404, + 1936, + 1404, + 2038, + 297, + 2038 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 295, + 1296, + 1401, + 1296, + 1401, + 1366, + 295, + 1366 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 555, + 1238, + 1144, + 1238, + 1144, + 1283, + 555, + 1283 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 300, + 970, + 704, + 970, + 704, + 1003, + 300, + 1003 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 290, + 225, + 1348, + 225, + 1348, + 264, + 290, + 264 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 301, + 837, + 1399, + 837, + 1399, + 931, + 301, + 931 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1247, + 1400, + 1247, + 1400, + 1277, + 1366, + 1277 + ], + "score": 0.89 + }, + { + "category_id": 8, + "poly": [ + 506, + 1563, + 1181, + 1563, + 1181, + 1647, + 506, + 1647 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1591, + 1401, + 1591, + 1401, + 1620, + 1366, + 1620 + ], + "score": 0.877 + }, + { + "category_id": 9, + "poly": [ + 1365, + 398, + 1400, + 398, + 1400, + 428, + 1365, + 428 + ], + "score": 0.875 + }, + { + "category_id": 9, + "poly": [ + 1365, + 308, + 1400, + 308, + 1400, + 337, + 1365, + 337 + ], + "score": 0.874 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1679, + 1400, + 1679, + 1400, + 1708, + 1366, + 1708 + ], + "score": 0.869 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1760, + 1400, + 1760, + 1400, + 1789, + 1366, + 1789 + ], + "score": 0.868 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1847, + 1400, + 1847, + 1400, + 1876, + 1366, + 1876 + ], + "score": 0.864 + }, + { + "category_id": 8, + "poly": [ + 659, + 1694, + 1187, + 1694, + 1187, + 1733, + 659, + 1733 + ], + "score": 0.767 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2112, + 841, + 2112 + ], + "score": 0.735 + }, + { + "category_id": 8, + "poly": [ + 664, + 1775, + 1189, + 1775, + 1189, + 1814, + 664, + 1814 + ], + "score": 0.729 + }, + { + "category_id": 8, + "poly": [ + 583, + 280, + 1116, + 280, + 1116, + 458, + 583, + 458 + ], + "score": 0.67 + }, + { + "category_id": 8, + "poly": [ + 664, + 1820, + 1188, + 1820, + 1188, + 1905, + 664, + 1905 + ], + "score": 0.631 + }, + { + "category_id": 8, + "poly": [ + 740, + 372, + 1116, + 372, + 1116, + 456, + 740, + 456 + ], + "score": 0.526 + }, + { + "category_id": 8, + "poly": [ + 659, + 1653, + 1165, + 1653, + 1165, + 1692, + 659, + 1692 + ], + "score": 0.441 + }, + { + "category_id": 8, + "poly": [ + 581, + 282, + 1050, + 282, + 1050, + 366, + 581, + 366 + ], + "score": 0.43 + }, + { + "category_id": 8, + "poly": [ + 664, + 1653, + 1183, + 1653, + 1183, + 1734, + 664, + 1734 + ], + "score": 0.236 + }, + { + "category_id": 8, + "poly": [ + 663, + 1820, + 1188, + 1820, + 1188, + 1905, + 663, + 1905 + ], + "score": 0.22 + }, + { + "category_id": 8, + "poly": [ + 666, + 1739, + 1189, + 1739, + 1189, + 1814, + 666, + 1814 + ], + "score": 0.134 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2112, + 841, + 2112 + ], + "score": 0.104 + }, + { + "category_id": 14, + "poly": [ + 581, + 279, + 1116, + 279, + 1116, + 463, + 581, + 463 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = \\frac { P ( u _ { t } ^ { a ^ { \\prime } } | \\tau _ { t } ) P ( \\tau _ { t } | \\tau _ { t } ^ { a } ) } { \\sum _ { \\tau _ { t } ^ { \\prime } } P ( u _ { t } ^ { a ^ { \\prime } } | \\tau _ { t } ^ { \\prime } ) P ( \\tau _ { t } ^ { \\prime } | \\tau _ { t } ^ { a } ) } } \\\\ & { \\qquad = \\frac { \\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) ) B ( \\tau _ { t } ) } { \\sum _ { \\tau _ { t } ^ { \\prime } } \\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ^ { \\prime } ) ) B ( \\tau _ { t } ^ { \\prime } ) } , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 505, + 1559, + 1191, + 1559, + 1191, + 1909, + 505, + 1909 + ], + "score": 0.94, + "latex": "\\begin{array} { r l r } { { P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau _ { t } ) } { \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau ^ { \\prime } ) } } } \\\\ & { } & { = \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ) B ( \\tau _ { t } ) } { \\epsilon / | U | + \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau ^ { \\prime } ) } } \\\\ & { } & { = \\frac { B ( \\tau _ { t } ) } { 1 + | U | \\sum _ { s ^ { \\prime } } ( ( 1 / \\epsilon - 1 ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) B ( \\tau ^ { \\prime } ) } } \\\\ & { } & { + \\frac { ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau _ { t } ) } { \\epsilon / | U | + \\sum _ { \\tau ^ { \\prime } } ( ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u _ { t } ^ { a ^ { \\prime } } ) ) B ( \\tau ^ { \\prime } ) } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 699, + 1999, + 956, + 1999, + 956, + 2037, + 699, + 2037 + ], + "score": 0.93, + "latex": "P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u _ { t } ^ { a ^ { \\prime } } ) = B ( \\tau _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 753, + 1181, + 954, + 1181, + 954, + 1221, + 753, + 1221 + ], + "score": 0.93, + "latex": "\\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) )" + }, + { + "category_id": 13, + "poly": [ + 1136, + 491, + 1238, + 491, + 1238, + 525, + 1136, + 525 + ], + "score": 0.93, + "latex": "O ( a ^ { \\prime } , \\tau _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1072, + 1938, + 1140, + 1938, + 1140, + 1971, + 1072, + 1971 + ], + "score": 0.92, + "latex": "B ( \\tau _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 684, + 226, + 722, + 226, + 722, + 264, + 684, + 264 + ], + "score": 0.92, + "latex": "u _ { t } ^ { a ^ { \\prime } }" + }, + { + "category_id": 13, + "poly": [ + 575, + 1297, + 651, + 1297, + 651, + 1331, + 575, + 1331 + ], + "score": 0.92, + "latex": "\\boldsymbol { u } ^ { * } ( \\tau _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 587, + 409, + 587, + 409, + 627, + 298, + 627 + ], + "score": 0.91, + "latex": "P ( \\tau _ { t } ^ { a ^ { \\prime } } | \\bar { \\tau _ { t } } )" + }, + { + "category_id": 13, + "poly": [ + 519, + 522, + 557, + 522, + 557, + 561, + 519, + 561 + ], + "score": 0.9, + "latex": "\\tau _ { t } ^ { a ^ { \\prime } }" + }, + { + "category_id": 14, + "poly": [ + 555, + 1235, + 1142, + 1235, + 1142, + 1281, + 555, + 1281 + ], + "score": 0.9, + "latex": "\\pi ^ { a ^ { \\prime } } ( u _ { t } ^ { a ^ { \\prime } } | O ( a ^ { \\prime } , \\tau _ { t } ) ) = ( 1 - \\epsilon ) \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u _ { t } ^ { a ^ { \\prime } } ) + \\epsilon / | U | ," + }, + { + "category_id": 13, + "poly": [ + 1332, + 1970, + 1396, + 1970, + 1396, + 1997, + 1332, + 1997 + ], + "score": 0.89, + "latex": "\\epsilon = 1" + }, + { + "category_id": 13, + "poly": [ + 342, + 1328, + 605, + 1328, + 605, + 1368, + 342, + 1368 + ], + "score": 0.87, + "latex": "\\operatorname* { m a x } _ { u } Q ^ { a ^ { \\prime } } ( u , O ( a ^ { \\prime } , \\tau _ { t } ) )" + }, + { + "category_id": 13, + "poly": [ + 1235, + 527, + 1260, + 527, + 1260, + 556, + 1235, + 556 + ], + "score": 0.84, + "latex": "a ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 929, + 231, + 953, + 231, + 953, + 258, + 929, + 258 + ], + "score": 0.84, + "latex": "a ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1291, + 531, + 1317, + 531, + 1317, + 558, + 1291, + 558 + ], + "score": 0.83, + "latex": "\\tau _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1274, + 1295, + 1403, + 1295, + 1403, + 1332, + 1274, + 1332 + ], + "score": 0.76, + "latex": "\\begin{array} { r l } { u ^ { * } ( \\tau _ { t } ) } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1224, + 1297, + 1249, + 1297, + 1249, + 1326, + 1224, + 1326 + ], + "score": 0.66, + "latex": "a ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 425, + 1159, + 440, + 1159, + 440, + 1179, + 425, + 1179 + ], + "score": 0.57, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 657, + 1334, + 674, + 1334, + 674, + 1361, + 657, + 1361 + ], + "score": 0.37, + "latex": "\\mathbf { I }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 971.0, + 707.0, + 971.0, + 707.0, + 1004.0, + 296.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 638.0, + 1406.0, + 638.0, + 1406.0, + 675.0, + 295.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 670.0, + 1406.0, + 670.0, + 1406.0, + 706.0, + 295.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 700.0, + 1405.0, + 700.0, + 1405.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 766.0, + 295.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 761.0, + 1406.0, + 761.0, + 1406.0, + 797.0, + 295.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 793.0, + 1333.0, + 793.0, + 1333.0, + 825.0, + 295.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1027.0, + 1406.0, + 1027.0, + 1406.0, + 1065.0, + 294.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1057.0, + 1406.0, + 1057.0, + 1406.0, + 1099.0, + 292.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1089.0, + 1405.0, + 1089.0, + 1405.0, + 1127.0, + 294.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1122.0, + 1405.0, + 1122.0, + 1405.0, + 1155.0, + 295.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1151.0, + 424.0, + 1151.0, + 424.0, + 1186.0, + 292.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1151.0, + 1407.0, + 1151.0, + 1407.0, + 1186.0, + 441.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1185.0, + 752.0, + 1185.0, + 752.0, + 1222.0, + 295.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1185.0, + 1074.0, + 1185.0, + 1074.0, + 1222.0, + 955.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 489.0, + 1135.0, + 489.0, + 1135.0, + 527.0, + 295.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 489.0, + 1408.0, + 489.0, + 1408.0, + 527.0, + 1239.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 523.0, + 518.0, + 523.0, + 518.0, + 564.0, + 294.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 523.0, + 1234.0, + 523.0, + 1234.0, + 564.0, + 558.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 523.0, + 1290.0, + 523.0, + 1290.0, + 564.0, + 1261.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 523.0, + 1407.0, + 523.0, + 1407.0, + 564.0, + 1318.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 555.0, + 1405.0, + 555.0, + 1405.0, + 592.0, + 295.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 586.0, + 297.0, + 586.0, + 297.0, + 630.0, + 293.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 586.0, + 424.0, + 586.0, + 424.0, + 630.0, + 410.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1378.0, + 1405.0, + 1378.0, + 1405.0, + 1415.0, + 292.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1409.0, + 1407.0, + 1409.0, + 1407.0, + 1445.0, + 293.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1439.0, + 1405.0, + 1439.0, + 1405.0, + 1477.0, + 292.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1474.0, + 410.0, + 1474.0, + 410.0, + 1505.0, + 292.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1936.0, + 1071.0, + 1936.0, + 1071.0, + 1972.0, + 295.0, + 1972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1936.0, + 1404.0, + 1936.0, + 1404.0, + 1972.0, + 1141.0, + 1972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1968.0, + 1331.0, + 1968.0, + 1331.0, + 2002.0, + 295.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1968.0, + 1407.0, + 1968.0, + 1407.0, + 2002.0, + 1397.0, + 2002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 698.0, + 2001.0, + 698.0, + 2040.0, + 295.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2040.0, + 957.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1291.0, + 574.0, + 1291.0, + 574.0, + 1335.0, + 293.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1291.0, + 1223.0, + 1291.0, + 1223.0, + 1335.0, + 652.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 1291.0, + 1273.0, + 1291.0, + 1273.0, + 1335.0, + 1250.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1330.0, + 341.0, + 1330.0, + 341.0, + 1371.0, + 295.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1330.0, + 656.0, + 1330.0, + 656.0, + 1371.0, + 606.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1330.0, + 954.0, + 1330.0, + 954.0, + 1371.0, + 675.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 223.0, + 683.0, + 223.0, + 683.0, + 271.0, + 291.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 223.0, + 928.0, + 223.0, + 928.0, + 271.0, + 723.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 223.0, + 1352.0, + 223.0, + 1352.0, + 271.0, + 954.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 838.0, + 1404.0, + 838.0, + 1404.0, + 872.0, + 295.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 868.0, + 1403.0, + 868.0, + 1403.0, + 902.0, + 296.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 901.0, + 1048.0, + 901.0, + 1048.0, + 931.0, + 298.0, + 931.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 878, + 1404, + 878, + 1404, + 1096, + 297, + 1096 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 559, + 1405, + 559, + 1405, + 747, + 297, + 747 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 307, + 1405, + 307, + 1405, + 461, + 298, + 461 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1463, + 1404, + 1463, + 1404, + 1587, + 298, + 1587 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1217, + 1403, + 1217, + 1403, + 1341, + 298, + 1341 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1905, + 1404, + 1905, + 1404, + 2028, + 299, + 2028 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1766, + 1404, + 1766, + 1404, + 1890, + 297, + 1890 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 301, + 1109, + 1398, + 1109, + 1398, + 1204, + 301, + 1204 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 301, + 1355, + 1400, + 1355, + 1400, + 1449, + 301, + 1449 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 613, + 765, + 1086, + 765, + 1086, + 849, + 613, + 849 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 293, + 229, + 1401, + 229, + 1401, + 293, + 293, + 293 + ], + "score": 0.954 + }, + { + "category_id": 0, + "poly": [ + 299, + 1706, + 548, + 1706, + 548, + 1739, + 299, + 1739 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 300, + 502, + 742, + 502, + 742, + 534, + 300, + 534 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 300, + 1636, + 557, + 1636, + 557, + 1672, + 300, + 1672 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 104, + 300, + 104 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1366, + 790, + 1400, + 790, + 1400, + 820, + 1366, + 820 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.78 + }, + { + "category_id": 14, + "poly": [ + 611, + 764, + 1089, + 764, + 1089, + 848, + 611, + 848 + ], + "score": 0.94, + "latex": "P ( \\tau _ { t } | \\tau _ { t } ^ { a } , u ^ { * } ) = \\frac { \\mathbf { I } ( u ^ { * } ( \\tau _ { t } ) , u ^ { * } ) \\big ) B ( \\tau _ { t } ) } { \\sum _ { \\tau ^ { \\prime } } \\mathbf { I } ( u ^ { * } ( \\tau ^ { \\prime } ) , u ^ { * } ) \\big ) B ( \\tau ^ { \\prime } ) }" + }, + { + "category_id": 13, + "poly": [ + 713, + 685, + 744, + 685, + 744, + 712, + 713, + 712 + ], + "score": 0.86, + "latex": "u ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1295, + 974, + 1326, + 974, + 1326, + 1001, + 1295, + 1001 + ], + "score": 0.85, + "latex": "u ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1292, + 914, + 1324, + 914, + 1324, + 940, + 1292, + 940 + ], + "score": 0.85, + "latex": "u ^ { a }" + }, + { + "category_id": 13, + "poly": [ + 655, + 1147, + 671, + 1147, + 671, + 1169, + 655, + 1169 + ], + "score": 0.71, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 563, + 721, + 579, + 721, + 579, + 742, + 563, + 742 + ], + "score": 0.68, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1173, + 598, + 1188, + 598, + 1188, + 620, + 1173, + 620 + ], + "score": 0.66, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1704.0, + 551.0, + 1704.0, + 551.0, + 1743.0, + 294.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 502.0, + 747.0, + 502.0, + 747.0, + 538.0, + 295.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1632.0, + 562.0, + 1632.0, + 562.0, + 1677.0, + 291.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 876.0, + 1405.0, + 876.0, + 1405.0, + 916.0, + 294.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 910.0, + 1291.0, + 910.0, + 1291.0, + 948.0, + 292.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 910.0, + 1406.0, + 910.0, + 1406.0, + 948.0, + 1325.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 941.0, + 1404.0, + 941.0, + 1404.0, + 976.0, + 295.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 973.0, + 1294.0, + 973.0, + 1294.0, + 1004.0, + 295.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 973.0, + 1404.0, + 973.0, + 1404.0, + 1004.0, + 1327.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1001.0, + 1406.0, + 1001.0, + 1406.0, + 1039.0, + 294.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1034.0, + 1404.0, + 1034.0, + 1404.0, + 1069.0, + 294.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1062.0, + 1174.0, + 1062.0, + 1174.0, + 1101.0, + 294.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 560.0, + 1405.0, + 560.0, + 1405.0, + 596.0, + 294.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 590.0, + 1172.0, + 590.0, + 1172.0, + 627.0, + 291.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 590.0, + 1405.0, + 590.0, + 1405.0, + 627.0, + 1189.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 621.0, + 1408.0, + 621.0, + 1408.0, + 658.0, + 292.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 654.0, + 1405.0, + 654.0, + 1405.0, + 686.0, + 295.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 683.0, + 712.0, + 683.0, + 712.0, + 719.0, + 294.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 683.0, + 1408.0, + 683.0, + 1408.0, + 719.0, + 745.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 715.0, + 562.0, + 715.0, + 562.0, + 751.0, + 295.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 715.0, + 991.0, + 715.0, + 991.0, + 751.0, + 580.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 307.0, + 1408.0, + 307.0, + 1408.0, + 344.0, + 295.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 335.0, + 1406.0, + 335.0, + 1406.0, + 373.0, + 293.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 367.0, + 1406.0, + 367.0, + 1406.0, + 404.0, + 293.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 399.0, + 1407.0, + 399.0, + 1407.0, + 436.0, + 295.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 429.0, + 581.0, + 429.0, + 581.0, + 465.0, + 293.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1461.0, + 1408.0, + 1461.0, + 1408.0, + 1499.0, + 295.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1496.0, + 1404.0, + 1496.0, + 1404.0, + 1528.0, + 293.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1524.0, + 1405.0, + 1524.0, + 1405.0, + 1562.0, + 292.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1558.0, + 1120.0, + 1558.0, + 1120.0, + 1590.0, + 294.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1215.0, + 1404.0, + 1215.0, + 1404.0, + 1253.0, + 293.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1250.0, + 1405.0, + 1250.0, + 1405.0, + 1283.0, + 296.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1279.0, + 1404.0, + 1279.0, + 1404.0, + 1314.0, + 294.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1308.0, + 1403.0, + 1308.0, + 1403.0, + 1344.0, + 293.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1906.0, + 1406.0, + 1906.0, + 1406.0, + 1939.0, + 295.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1936.0, + 1405.0, + 1936.0, + 1405.0, + 1969.0, + 295.0, + 1969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1966.0, + 1405.0, + 1966.0, + 1405.0, + 2000.0, + 294.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1998.0, + 1070.0, + 1998.0, + 1070.0, + 2030.0, + 295.0, + 2030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1765.0, + 1409.0, + 1765.0, + 1409.0, + 1802.0, + 292.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1796.0, + 1406.0, + 1796.0, + 1406.0, + 1833.0, + 294.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1828.0, + 1405.0, + 1828.0, + 1405.0, + 1864.0, + 294.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1857.0, + 597.0, + 1857.0, + 597.0, + 1895.0, + 294.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1108.0, + 1403.0, + 1108.0, + 1403.0, + 1147.0, + 293.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1141.0, + 654.0, + 1141.0, + 654.0, + 1177.0, + 295.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1141.0, + 1404.0, + 1141.0, + 1404.0, + 1177.0, + 672.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1170.0, + 1113.0, + 1170.0, + 1113.0, + 1208.0, + 293.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1355.0, + 1404.0, + 1355.0, + 1404.0, + 1392.0, + 295.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1386.0, + 1404.0, + 1386.0, + 1404.0, + 1423.0, + 293.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1414.0, + 788.0, + 1414.0, + 788.0, + 1453.0, + 295.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1359.0, + 261.0, + 1359.0, + 297.0, + 296.0, + 297.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1369, + 1404, + 1369, + 1404, + 1827, + 298, + 1827 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 230, + 936, + 230, + 936, + 413, + 299, + 413 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 982, + 1404, + 982, + 1404, + 1165, + 298, + 1165 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 510, + 936, + 510, + 936, + 783, + 298, + 783 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 785, + 1404, + 785, + 1404, + 967, + 297, + 967 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 299, + 1181, + 1403, + 1181, + 1403, + 1273, + 299, + 1273 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 959, + 243, + 1394, + 243, + 1394, + 676, + 959, + 676 + ], + "score": 0.963 + }, + { + "category_id": 4, + "poly": [ + 955, + 717, + 1402, + 717, + 1402, + 774, + 955, + 774 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 299, + 1313, + 994, + 1313, + 994, + 1343, + 299, + 1343 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 299, + 452, + 464, + 452, + 464, + 483, + 299, + 483 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2110, + 842, + 2110 + ], + "score": 0.74 + }, + { + "category_id": 13, + "poly": [ + 1057, + 1105, + 1186, + 1105, + 1186, + 1134, + 1057, + 1134 + ], + "score": 0.91, + "latex": "5 \\times 5 = 2 5" + }, + { + "category_id": 13, + "poly": [ + 506, + 1674, + 600, + 1674, + 600, + 1702, + 506, + 1702 + ], + "score": 0.9, + "latex": "K = 8 0" + }, + { + "category_id": 13, + "poly": [ + 1305, + 1643, + 1402, + 1643, + 1402, + 1672, + 1305, + 1672 + ], + "score": 0.89, + "latex": "N = 8 0" + }, + { + "category_id": 13, + "poly": [ + 885, + 1492, + 914, + 1492, + 914, + 1518, + 885, + 1518 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1431, + 1401, + 1431, + 1401, + 1457, + 1373, + 1457 + ], + "score": 0.79, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 293, + 600, + 426, + 600, + 426, + 630, + 293, + 630 + ], + "score": 0.5, + "latex": "/ \\textbf { W } / \\textbf { Y } / \\textbf { R }" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 243.0, + 1326.0, + 243.0, + 1326.0, + 276.0, + 1136.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 267.0, + 1174.0, + 267.0, + 1174.0, + 297.0, + 1111.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 268.0, + 1350.0, + 268.0, + 1350.0, + 295.0, + 1289.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 294.0, + 1198.0, + 294.0, + 1198.0, + 322.0, + 1085.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 322.0, + 1106.0, + 322.0, + 1106.0, + 339.0, + 1090.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 319.0, + 1153.0, + 319.0, + 1153.0, + 342.0, + 1132.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 319.0, + 1197.0, + 319.0, + 1197.0, + 341.0, + 1177.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 355.0, + 1385.0, + 355.0, + 1385.0, + 389.0, + 1345.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 356.0, + 1044.0, + 356.0, + 1044.0, + 471.0, + 1015.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 357.0, + 1116.0, + 357.0, + 1116.0, + 387.0, + 1079.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 357.0, + 1288.0, + 357.0, + 1288.0, + 388.0, + 1261.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 357.0, + 1333.0, + 357.0, + 1333.0, + 387.0, + 1306.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 359.0, + 1154.0, + 359.0, + 1154.0, + 387.0, + 1130.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 359.0, + 1199.0, + 359.0, + 1199.0, + 387.0, + 1176.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 363.0, + 1063.0, + 363.0, + 1063.0, + 381.0, + 1045.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 387.0, + 1011.0, + 387.0, + 1011.0, + 444.0, + 988.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 409.0, + 1062.0, + 409.0, + 1062.0, + 427.0, + 1046.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 404.0, + 1108.0, + 404.0, + 1108.0, + 430.0, + 1084.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 404.0, + 1153.0, + 404.0, + 1153.0, + 431.0, + 1131.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 403.0, + 1201.0, + 403.0, + 1201.0, + 433.0, + 1175.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 402.0, + 1288.0, + 402.0, + 1288.0, + 433.0, + 1260.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 404.0, + 1330.0, + 404.0, + 1330.0, + 431.0, + 1308.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 404.0, + 1376.0, + 404.0, + 1376.0, + 431.0, + 1352.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 418.0, + 987.0, + 418.0, + 987.0, + 578.0, + 958.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 447.0, + 1117.0, + 447.0, + 1117.0, + 478.0, + 1079.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 447.0, + 1288.0, + 447.0, + 1288.0, + 478.0, + 1261.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 447.0, + 1383.0, + 447.0, + 1383.0, + 479.0, + 1347.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 448.0, + 1156.0, + 448.0, + 1156.0, + 478.0, + 1129.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 448.0, + 1199.0, + 448.0, + 1199.0, + 476.0, + 1176.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 449.0, + 1331.0, + 449.0, + 1331.0, + 476.0, + 1308.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 454.0, + 1062.0, + 454.0, + 1062.0, + 471.0, + 1046.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 535.0, + 1109.0, + 535.0, + 1109.0, + 564.0, + 1085.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 535.0, + 1155.0, + 535.0, + 1155.0, + 562.0, + 1131.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 533.0, + 1205.0, + 533.0, + 1205.0, + 565.0, + 1169.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 532.0, + 1294.0, + 532.0, + 1294.0, + 567.0, + 1254.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 535.0, + 1330.0, + 535.0, + 1330.0, + 563.0, + 1308.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 535.0, + 1376.0, + 535.0, + 1376.0, + 564.0, + 1353.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 562.0, + 1011.0, + 562.0, + 1011.0, + 621.0, + 988.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 582.0, + 1110.0, + 582.0, + 1110.0, + 609.0, + 1085.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 581.0, + 1154.0, + 581.0, + 1154.0, + 608.0, + 1131.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 581.0, + 1201.0, + 581.0, + 1201.0, + 612.0, + 1175.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 580.0, + 1288.0, + 580.0, + 1288.0, + 610.0, + 1261.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 581.0, + 1331.0, + 581.0, + 1331.0, + 609.0, + 1307.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 581.0, + 1376.0, + 581.0, + 1376.0, + 610.0, + 1353.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 627.0, + 1109.0, + 627.0, + 1109.0, + 654.0, + 1086.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 627.0, + 1155.0, + 627.0, + 1155.0, + 654.0, + 1130.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 627.0, + 1200.0, + 627.0, + 1200.0, + 654.0, + 1177.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 623.0, + 1294.0, + 623.0, + 1294.0, + 657.0, + 1254.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 625.0, + 1331.0, + 625.0, + 1331.0, + 654.0, + 1308.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 626.0, + 1376.0, + 626.0, + 1376.0, + 654.0, + 1353.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 714.0, + 1404.0, + 714.0, + 1404.0, + 750.0, + 955.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 745.0, + 1227.0, + 745.0, + 1227.0, + 774.0, + 956.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1309.0, + 999.0, + 1309.0, + 999.0, + 1349.0, + 293.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 448.0, + 470.0, + 448.0, + 470.0, + 487.0, + 294.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 859.0, + 2087.0, + 859.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1368.0, + 1405.0, + 1368.0, + 1405.0, + 1404.0, + 293.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1397.0, + 1408.0, + 1397.0, + 1408.0, + 1435.0, + 292.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1430.0, + 1372.0, + 1430.0, + 1372.0, + 1465.0, + 294.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1462.0, + 1406.0, + 1462.0, + 1406.0, + 1495.0, + 293.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1491.0, + 884.0, + 1491.0, + 884.0, + 1525.0, + 293.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1491.0, + 1406.0, + 1491.0, + 1406.0, + 1525.0, + 915.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1521.0, + 1406.0, + 1521.0, + 1406.0, + 1557.0, + 293.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1551.0, + 1405.0, + 1551.0, + 1405.0, + 1586.0, + 293.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1582.0, + 1407.0, + 1582.0, + 1407.0, + 1619.0, + 292.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1613.0, + 1406.0, + 1613.0, + 1406.0, + 1648.0, + 294.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1644.0, + 1304.0, + 1644.0, + 1304.0, + 1679.0, + 294.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1644.0, + 1406.0, + 1644.0, + 1406.0, + 1679.0, + 1403.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1675.0, + 505.0, + 1675.0, + 505.0, + 1706.0, + 294.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1675.0, + 1405.0, + 1675.0, + 1405.0, + 1706.0, + 601.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1704.0, + 1406.0, + 1704.0, + 1406.0, + 1739.0, + 293.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1736.0, + 1402.0, + 1736.0, + 1402.0, + 1767.0, + 294.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1765.0, + 1405.0, + 1765.0, + 1405.0, + 1800.0, + 294.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1796.0, + 1267.0, + 1796.0, + 1267.0, + 1831.0, + 294.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 938.0, + 230.0, + 938.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 938.0, + 262.0, + 938.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 293.0, + 938.0, + 293.0, + 938.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 322.0, + 936.0, + 322.0, + 936.0, + 353.0, + 295.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 350.0, + 937.0, + 350.0, + 937.0, + 386.0, + 293.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 383.0, + 617.0, + 383.0, + 617.0, + 415.0, + 294.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 981.0, + 1405.0, + 981.0, + 1405.0, + 1017.0, + 294.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1012.0, + 1406.0, + 1012.0, + 1406.0, + 1047.0, + 293.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 1405.0, + 1044.0, + 1405.0, + 1076.0, + 294.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1075.0, + 1402.0, + 1075.0, + 1402.0, + 1106.0, + 296.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1104.0, + 1056.0, + 1104.0, + 1056.0, + 1142.0, + 292.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1104.0, + 1410.0, + 1104.0, + 1410.0, + 1142.0, + 1187.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1134.0, + 1240.0, + 1134.0, + 1240.0, + 1170.0, + 294.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 507.0, + 940.0, + 507.0, + 940.0, + 544.0, + 295.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 541.0, + 937.0, + 541.0, + 937.0, + 571.0, + 296.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 571.0, + 937.0, + 571.0, + 937.0, + 600.0, + 295.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 601.0, + 938.0, + 601.0, + 938.0, + 631.0, + 427.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 632.0, + 937.0, + 632.0, + 937.0, + 662.0, + 296.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 661.0, + 939.0, + 661.0, + 939.0, + 693.0, + 293.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 691.0, + 937.0, + 691.0, + 937.0, + 724.0, + 295.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 720.0, + 937.0, + 720.0, + 937.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 752.0, + 938.0, + 752.0, + 938.0, + 785.0, + 296.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 783.0, + 1405.0, + 783.0, + 1405.0, + 814.0, + 296.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 811.0, + 1406.0, + 811.0, + 1406.0, + 848.0, + 292.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 846.0, + 1405.0, + 846.0, + 1405.0, + 877.0, + 295.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 871.0, + 1406.0, + 871.0, + 1406.0, + 910.0, + 290.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 906.0, + 1406.0, + 906.0, + 1406.0, + 938.0, + 295.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 931.0, + 534.0, + 931.0, + 534.0, + 972.0, + 293.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1180.0, + 1405.0, + 1180.0, + 1405.0, + 1217.0, + 293.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1211.0, + 1404.0, + 1211.0, + 1404.0, + 1247.0, + 294.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 1048.0, + 1241.0, + 1048.0, + 1279.0, + 294.0, + 1279.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1484, + 1404, + 1484, + 1404, + 1791, + 298, + 1791 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 357, + 882, + 357, + 882, + 663, + 299, + 663 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1147, + 1403, + 1147, + 1403, + 1331, + 298, + 1331 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1347, + 1404, + 1347, + 1404, + 1468, + 299, + 1468 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1039, + 1403, + 1039, + 1403, + 1131, + 299, + 1131 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 763, + 1403, + 763, + 1403, + 916, + 299, + 916 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 931, + 1403, + 931, + 1403, + 1023, + 299, + 1023 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 911, + 230, + 1362, + 230, + 1362, + 669, + 911, + 669 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 954, + 697, + 1348, + 697, + 1348, + 727, + 954, + 727 + ], + "score": 0.921 + }, + { + "category_id": 0, + "poly": [ + 299, + 1839, + 842, + 1839, + 842, + 1875, + 299, + 1875 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 299, + 226, + 480, + 226, + 480, + 261, + 299, + 261 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 299, + 297, + 546, + 297, + 546, + 329, + 299, + 329 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 298, + 704, + 464, + 704, + 464, + 734, + 298, + 734 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.782 + }, + { + "category_id": 13, + "poly": [ + 297, + 570, + 436, + 570, + 436, + 601, + 297, + 601 + ], + "score": 0.89, + "latex": "9 . 9 7 \\pm 0 . 0 2" + }, + { + "category_id": 13, + "poly": [ + 859, + 1669, + 951, + 1669, + 951, + 1699, + 859, + 1699 + ], + "score": 0.87, + "latex": "( 5 8 . 6 \\% )" + }, + { + "category_id": 13, + "poly": [ + 339, + 480, + 399, + 480, + 399, + 507, + 339, + 507 + ], + "score": 0.52, + "latex": "B A D" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 257.0, + 966.0, + 257.0, + 966.0, + 292.0, + 901.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 324.0, + 967.0, + 324.0, + 967.0, + 360.0, + 913.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 391.0, + 970.0, + 391.0, + 970.0, + 427.0, + 913.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 457.0, + 972.0, + 457.0, + 972.0, + 493.0, + 913.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 505.0, + 1285.0, + 505.0, + 1285.0, + 541.0, + 1232.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 525.0, + 970.0, + 525.0, + 970.0, + 559.0, + 913.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 533.0, + 1292.0, + 533.0, + 1292.0, + 567.0, + 1233.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 563.0, + 1341.0, + 563.0, + 1341.0, + 596.0, + 1235.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 592.0, + 966.0, + 592.0, + 966.0, + 623.0, + 912.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 615.0, + 997.0, + 615.0, + 997.0, + 644.0, + 973.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 612.0, + 1369.0, + 612.0, + 1369.0, + 647.0, + 1018.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 640.0, + 1198.0, + 640.0, + 1198.0, + 676.0, + 1119.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 692.0, + 1354.0, + 692.0, + 1354.0, + 732.0, + 949.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1834.0, + 847.0, + 1834.0, + 847.0, + 1880.0, + 292.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 221.0, + 485.0, + 221.0, + 485.0, + 269.0, + 292.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 294.0, + 551.0, + 294.0, + 551.0, + 334.0, + 294.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 702.0, + 468.0, + 702.0, + 468.0, + 738.0, + 295.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1481.0, + 1404.0, + 1481.0, + 1404.0, + 1523.0, + 293.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1514.0, + 1406.0, + 1514.0, + 1406.0, + 1551.0, + 291.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1546.0, + 1405.0, + 1546.0, + 1405.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1575.0, + 1404.0, + 1575.0, + 1404.0, + 1615.0, + 293.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1604.0, + 1409.0, + 1604.0, + 1409.0, + 1646.0, + 291.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1636.0, + 1406.0, + 1636.0, + 1406.0, + 1674.0, + 292.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1667.0, + 858.0, + 1667.0, + 858.0, + 1704.0, + 292.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1667.0, + 1406.0, + 1667.0, + 1406.0, + 1704.0, + 952.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1699.0, + 1406.0, + 1699.0, + 1406.0, + 1734.0, + 292.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1730.0, + 1404.0, + 1730.0, + 1404.0, + 1762.0, + 296.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1131.0, + 1760.0, + 1131.0, + 1793.0, + 292.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 356.0, + 885.0, + 356.0, + 885.0, + 391.0, + 296.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 387.0, + 882.0, + 387.0, + 882.0, + 419.0, + 294.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 417.0, + 882.0, + 417.0, + 882.0, + 451.0, + 296.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 447.0, + 884.0, + 447.0, + 884.0, + 484.0, + 293.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 477.0, + 338.0, + 477.0, + 338.0, + 511.0, + 293.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 477.0, + 887.0, + 477.0, + 887.0, + 511.0, + 400.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 510.0, + 886.0, + 510.0, + 886.0, + 541.0, + 294.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 537.0, + 886.0, + 537.0, + 886.0, + 573.0, + 292.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 568.0, + 296.0, + 568.0, + 296.0, + 604.0, + 293.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 568.0, + 885.0, + 568.0, + 885.0, + 604.0, + 437.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 600.0, + 885.0, + 600.0, + 885.0, + 633.0, + 295.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 632.0, + 609.0, + 632.0, + 609.0, + 663.0, + 295.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1145.0, + 1405.0, + 1145.0, + 1405.0, + 1182.0, + 293.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1179.0, + 1405.0, + 1179.0, + 1405.0, + 1211.0, + 294.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1206.0, + 1405.0, + 1206.0, + 1405.0, + 1244.0, + 292.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1239.0, + 1405.0, + 1239.0, + 1405.0, + 1275.0, + 294.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1271.0, + 1404.0, + 1271.0, + 1404.0, + 1302.0, + 296.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1261.0, + 1300.0, + 1261.0, + 1335.0, + 294.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1383.0, + 293.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1379.0, + 1408.0, + 1379.0, + 1408.0, + 1411.0, + 294.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1407.0, + 1405.0, + 1407.0, + 1405.0, + 1443.0, + 294.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1440.0, + 756.0, + 1440.0, + 756.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1037.0, + 1408.0, + 1037.0, + 1408.0, + 1078.0, + 293.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1068.0, + 1404.0, + 1068.0, + 1404.0, + 1105.0, + 294.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1101.0, + 961.0, + 1101.0, + 961.0, + 1135.0, + 293.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1407.0, + 1911.0, + 1407.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1406.0, + 2000.0, + 1406.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 760.0, + 1405.0, + 760.0, + 1405.0, + 799.0, + 293.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 791.0, + 1405.0, + 791.0, + 1405.0, + 829.0, + 293.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 822.0, + 1407.0, + 822.0, + 1407.0, + 858.0, + 293.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 853.0, + 1407.0, + 853.0, + 1407.0, + 891.0, + 293.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 886.0, + 1392.0, + 886.0, + 1392.0, + 919.0, + 295.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 928.0, + 1404.0, + 928.0, + 1404.0, + 967.0, + 293.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 961.0, + 1405.0, + 961.0, + 1405.0, + 996.0, + 293.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 993.0, + 682.0, + 993.0, + 682.0, + 1026.0, + 295.0, + 1026.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 350, + 743, + 1349, + 743, + 1349, + 1328, + 350, + 1328 + ], + "score": 0.985, + "html": "
Agent2 Players 3 Players 4 Players 5 Players
Rainbow (Bard et al., 2019)20.64 ± 0.03 2.5%18.71 ±0.01 0.2%18.00 ± 0.17 0%15.26 ± 0.18 0%
ACHA (Bard et al., 2019)22.73 ± 0.12 15.1%20.24 ± 0.15 1.1%21.57 ± 0.12 2.4%16.80 ± 0.13 0%
BAD (Foerster et al.,2019)23.92 ± 0.01 58.56%==1
IQL (Baseline) 50.47%23.97 ± 0.01 40.25%23.69 ± 0.01 19.39%22.76 ± 0.01 4.93%21.29 ± 0.01
VDN (Baseline)23.96 ± 0.01 50.27%23.99 ± 0.01 50.37%23.79 ± 0.00 38.86%21.80 ± 0.01 4.98%
SAD24.01 ± 0.01 52.39%23.93 ± 0.01 48.05%23.81 ± 0.01 41.45%23.01 ± 0.01 13.93%
SAD& AuxTask24.08 ± 0.01 56.09%23.81 ± 0.01 49.74%23.47 ± 0.01 33.87%22.25 ± 0.01 7.33%
" + }, + { + "category_id": 5, + "poly": [ + 360, + 228, + 1341, + 228, + 1341, + 580, + 360, + 580 + ], + "score": 0.983, + "html": "
Agent 2 Players 3 Players 4 Players 5 Players
IQL (Baseline)23.77 ± 0.04 43.88 ± 1.21 %23.02 ± 0.10 26.16 ± 2.01 %21.99 ± 0.09 10.15 ± 0.86 %20.60 ± 0.11 2.27 ± 0.32 %
VDN23.83 ± 0.0323.71 ± 0.0623.03 ± 0.1521.18 ± 0.12
(Baseline) SAD44.97 ± 1.28 % 23.87 ± 0.0341.16 ± 1.27 % 23.69 ± 0.0523.57 ± 2.20 % 23.27 ± 0.162.26 ± 0.32 % 22.06 ± 0.23
SAD AuxTask47.90 ± 1.10 % 24.02 ± 0.0141.12 ± 1.10 % 23.56 ± 0.0729.38 ± 2.63 % 22.78 ± 0.107.22 ± 1.29 % 21.47 ± 0.08
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1543, + 1405, + 1543, + 1405, + 1761, + 297, + 1761 + ], + "score": 0.98 + }, + { + "category_id": 6, + "poly": [ + 298, + 611, + 1403, + 611, + 1403, + 697, + 298, + 697 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.901 + }, + { + "category_id": 6, + "poly": [ + 298, + 1358, + 1404, + 1358, + 1404, + 1473, + 298, + 1473 + ], + "score": 0.807 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.79 + }, + { + "category_id": 1, + "poly": [ + 298, + 1358, + 1404, + 1358, + 1404, + 1473, + 298, + 1473 + ], + "score": 0.143 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 609.0, + 1405.0, + 609.0, + 1405.0, + 642.0, + 293.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 640.0, + 1405.0, + 640.0, + 1405.0, + 669.0, + 294.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 668.0, + 1390.0, + 668.0, + 1390.0, + 697.0, + 296.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1355.0, + 1406.0, + 1355.0, + 1406.0, + 1394.0, + 293.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1420.0, + 294.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1416.0, + 1404.0, + 1416.0, + 1404.0, + 1445.0, + 296.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1444.0, + 1404.0, + 1444.0, + 1404.0, + 1473.0, + 296.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1545.0, + 1405.0, + 1545.0, + 1405.0, + 1579.0, + 294.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1579.0, + 1402.0, + 1579.0, + 1402.0, + 1610.0, + 295.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1606.0, + 1405.0, + 1606.0, + 1405.0, + 1641.0, + 292.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1637.0, + 1405.0, + 1637.0, + 1405.0, + 1672.0, + 295.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1666.0, + 1407.0, + 1666.0, + 1407.0, + 1703.0, + 292.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1733.0, + 294.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1727.0, + 1186.0, + 1727.0, + 1186.0, + 1764.0, + 292.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1355.0, + 1406.0, + 1355.0, + 1406.0, + 1394.0, + 293.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1420.0, + 294.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1416.0, + 1404.0, + 1416.0, + 1404.0, + 1445.0, + 296.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1444.0, + 1404.0, + 1444.0, + 1404.0, + 1473.0, + 296.0, + 1473.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.887 + }, + { + "category_id": 1, + "poly": [ + 298, + 278, + 1405, + 278, + 1405, + 370, + 298, + 370 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 295, + 640, + 1404, + 640, + 1404, + 705, + 295, + 705 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 295, + 560, + 1403, + 560, + 1403, + 624, + 295, + 624 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 296, + 720, + 1404, + 720, + 1404, + 786, + 296, + 786 + ], + "score": 0.795 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.781 + }, + { + "category_id": 1, + "poly": [ + 297, + 1295, + 1408, + 1295, + 1408, + 1419, + 297, + 1419 + ], + "score": 0.739 + }, + { + "category_id": 1, + "poly": [ + 296, + 1627, + 1405, + 1627, + 1405, + 1722, + 296, + 1722 + ], + "score": 0.735 + }, + { + "category_id": 1, + "poly": [ + 294, + 881, + 1402, + 881, + 1402, + 946, + 294, + 946 + ], + "score": 0.734 + }, + { + "category_id": 1, + "poly": [ + 296, + 1184, + 1405, + 1184, + 1405, + 1278, + 296, + 1278 + ], + "score": 0.733 + }, + { + "category_id": 1, + "poly": [ + 298, + 388, + 1406, + 388, + 1406, + 542, + 298, + 542 + ], + "score": 0.733 + }, + { + "category_id": 1, + "poly": [ + 297, + 801, + 1405, + 801, + 1405, + 865, + 297, + 865 + ], + "score": 0.729 + }, + { + "category_id": 1, + "poly": [ + 299, + 1738, + 1404, + 1738, + 1404, + 1832, + 299, + 1832 + ], + "score": 0.714 + }, + { + "category_id": 1, + "poly": [ + 295, + 1517, + 1406, + 1517, + 1406, + 1610, + 295, + 1610 + ], + "score": 0.681 + }, + { + "category_id": 1, + "poly": [ + 289, + 1436, + 1402, + 1436, + 1402, + 1499, + 289, + 1499 + ], + "score": 0.678 + }, + { + "category_id": 1, + "poly": [ + 297, + 1073, + 1403, + 1073, + 1403, + 1168, + 297, + 1168 + ], + "score": 0.644 + }, + { + "category_id": 1, + "poly": [ + 299, + 1850, + 1407, + 1850, + 1407, + 2033, + 299, + 2033 + ], + "score": 0.52 + }, + { + "category_id": 1, + "poly": [ + 297, + 961, + 1403, + 961, + 1403, + 1057, + 297, + 1057 + ], + "score": 0.149 + }, + { + "category_id": 1, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.096 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 278.0, + 1405.0, + 278.0, + 1405.0, + 312.0, + 296.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 308.0, + 1406.0, + 308.0, + 1406.0, + 342.0, + 323.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 337.0, + 466.0, + 337.0, + 466.0, + 372.0, + 322.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 635.0, + 1407.0, + 635.0, + 1407.0, + 679.0, + 292.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 669.0, + 399.0, + 669.0, + 399.0, + 707.0, + 319.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 556.0, + 1405.0, + 556.0, + 1405.0, + 598.0, + 293.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 592.0, + 861.0, + 592.0, + 861.0, + 625.0, + 322.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 721.0, + 1407.0, + 721.0, + 1407.0, + 757.0, + 295.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 750.0, + 649.0, + 750.0, + 649.0, + 787.0, + 323.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1293.0, + 1409.0, + 1293.0, + 1409.0, + 1333.0, + 291.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1325.0, + 1406.0, + 1325.0, + 1406.0, + 1361.0, + 321.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1357.0, + 1411.0, + 1357.0, + 1411.0, + 1392.0, + 321.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1389.0, + 864.0, + 1389.0, + 864.0, + 1419.0, + 323.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1627.0, + 1407.0, + 1627.0, + 1407.0, + 1661.0, + 296.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1660.0, + 1407.0, + 1660.0, + 1407.0, + 1694.0, + 324.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1692.0, + 854.0, + 1692.0, + 854.0, + 1722.0, + 325.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 880.0, + 1404.0, + 880.0, + 1404.0, + 919.0, + 294.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 914.0, + 1033.0, + 914.0, + 1033.0, + 947.0, + 323.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1181.0, + 1408.0, + 1181.0, + 1408.0, + 1221.0, + 291.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1213.0, + 1407.0, + 1213.0, + 1407.0, + 1252.0, + 322.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1246.0, + 550.0, + 1246.0, + 550.0, + 1281.0, + 322.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 390.0, + 1404.0, + 390.0, + 1404.0, + 423.0, + 295.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 420.0, + 1406.0, + 420.0, + 1406.0, + 453.0, + 322.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 448.0, + 1406.0, + 448.0, + 1406.0, + 488.0, + 320.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 476.0, + 1412.0, + 476.0, + 1412.0, + 518.0, + 321.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 511.0, + 863.0, + 511.0, + 863.0, + 544.0, + 322.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 801.0, + 1403.0, + 801.0, + 1403.0, + 837.0, + 295.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 832.0, + 541.0, + 832.0, + 541.0, + 866.0, + 321.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1738.0, + 1404.0, + 1738.0, + 1404.0, + 1775.0, + 294.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1772.0, + 1404.0, + 1772.0, + 1404.0, + 1806.0, + 324.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1802.0, + 707.0, + 1802.0, + 707.0, + 1833.0, + 320.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1516.0, + 1409.0, + 1516.0, + 1409.0, + 1554.0, + 295.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1544.0, + 1414.0, + 1544.0, + 1414.0, + 1584.0, + 321.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1579.0, + 905.0, + 1579.0, + 905.0, + 1613.0, + 324.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1432.0, + 1404.0, + 1432.0, + 1404.0, + 1474.0, + 293.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1468.0, + 610.0, + 1468.0, + 610.0, + 1497.0, + 321.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1073.0, + 1404.0, + 1073.0, + 1404.0, + 1111.0, + 293.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1102.0, + 1405.0, + 1102.0, + 1405.0, + 1141.0, + 321.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1134.0, + 1252.0, + 1134.0, + 1252.0, + 1170.0, + 321.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1848.0, + 1409.0, + 1848.0, + 1409.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1881.0, + 1407.0, + 1881.0, + 1407.0, + 1916.0, + 322.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1947.0, + 321.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1939.0, + 1407.0, + 1939.0, + 1407.0, + 1981.0, + 319.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1972.0, + 1409.0, + 1972.0, + 1409.0, + 2006.0, + 321.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 2003.0, + 1142.0, + 2003.0, + 1142.0, + 2035.0, + 325.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 960.0, + 1405.0, + 960.0, + 1405.0, + 999.0, + 292.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 994.0, + 1405.0, + 994.0, + 1405.0, + 1029.0, + 324.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1024.0, + 802.0, + 1024.0, + 802.0, + 1058.0, + 322.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 950, + 1405, + 950, + 1405, + 1044, + 298, + 1044 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 298, + 1070, + 1405, + 1070, + 1405, + 1164, + 298, + 1164 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 298, + 769, + 1406, + 769, + 1406, + 925, + 298, + 925 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 298, + 1911, + 1405, + 1911, + 1405, + 2033, + 298, + 2033 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 298, + 1670, + 1405, + 1670, + 1405, + 1764, + 298, + 1764 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 298, + 1429, + 1406, + 1429, + 1406, + 1554, + 298, + 1554 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 299, + 1191, + 1405, + 1191, + 1405, + 1284, + 299, + 1284 + ], + "score": 0.913 + }, + { + "category_id": 1, + "poly": [ + 298, + 649, + 1405, + 649, + 1405, + 743, + 298, + 743 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 294, + 1580, + 1403, + 1580, + 1403, + 1643, + 294, + 1643 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 297, + 409, + 1405, + 409, + 1405, + 502, + 297, + 502 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1402, + 229, + 1402, + 294, + 294, + 294 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 817, + 75, + 817, + 105, + 300, + 105 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 302, + 1791, + 1404, + 1791, + 1404, + 1884, + 302, + 1884 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 292, + 319, + 1401, + 319, + 1401, + 381, + 292, + 381 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 299, + 529, + 1400, + 529, + 1400, + 623, + 299, + 623 + ], + "score": 0.871 + }, + { + "category_id": 1, + "poly": [ + 292, + 1312, + 1290, + 1312, + 1290, + 1345, + 292, + 1345 + ], + "score": 0.861 + }, + { + "category_id": 1, + "poly": [ + 297, + 1370, + 1401, + 1370, + 1401, + 1404, + 297, + 1404 + ], + "score": 0.842 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 861, + 2088, + 861, + 2113, + 835, + 2113 + ], + "score": 0.806 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 951.0, + 1405.0, + 951.0, + 1405.0, + 985.0, + 296.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 981.0, + 1403.0, + 981.0, + 1403.0, + 1015.0, + 323.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1013.0, + 1172.0, + 1013.0, + 1172.0, + 1047.0, + 325.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1071.0, + 1405.0, + 1071.0, + 1405.0, + 1105.0, + 296.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1102.0, + 1405.0, + 1102.0, + 1405.0, + 1136.0, + 323.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1132.0, + 698.0, + 1132.0, + 698.0, + 1166.0, + 325.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 771.0, + 1408.0, + 771.0, + 1408.0, + 804.0, + 298.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 798.0, + 1408.0, + 798.0, + 1408.0, + 839.0, + 321.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 828.0, + 1409.0, + 828.0, + 1409.0, + 868.0, + 322.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 859.0, + 1409.0, + 859.0, + 1409.0, + 898.0, + 322.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 891.0, + 1411.0, + 891.0, + 1411.0, + 928.0, + 322.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1907.0, + 1410.0, + 1907.0, + 1410.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1975.0, + 321.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1972.0, + 1408.0, + 1972.0, + 1408.0, + 2008.0, + 321.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2003.0, + 530.0, + 2003.0, + 530.0, + 2033.0, + 322.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1668.0, + 1408.0, + 1668.0, + 1408.0, + 1707.0, + 293.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1699.0, + 1406.0, + 1699.0, + 1406.0, + 1739.0, + 320.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1733.0, + 1371.0, + 1733.0, + 1371.0, + 1767.0, + 325.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1427.0, + 1409.0, + 1427.0, + 1409.0, + 1466.0, + 293.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1462.0, + 1406.0, + 1462.0, + 1406.0, + 1495.0, + 324.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1493.0, + 1406.0, + 1493.0, + 1406.0, + 1525.0, + 322.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1524.0, + 505.0, + 1524.0, + 505.0, + 1554.0, + 322.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1190.0, + 1409.0, + 1190.0, + 1409.0, + 1228.0, + 295.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1222.0, + 1408.0, + 1222.0, + 1408.0, + 1257.0, + 321.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1251.0, + 397.0, + 1251.0, + 397.0, + 1285.0, + 322.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 647.0, + 1407.0, + 647.0, + 1407.0, + 685.0, + 293.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 679.0, + 1405.0, + 679.0, + 1405.0, + 715.0, + 322.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 713.0, + 1044.0, + 713.0, + 1044.0, + 744.0, + 323.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1577.0, + 1404.0, + 1577.0, + 1404.0, + 1617.0, + 294.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1612.0, + 807.0, + 1612.0, + 807.0, + 1642.0, + 322.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 410.0, + 1405.0, + 410.0, + 1405.0, + 443.0, + 297.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 439.0, + 1405.0, + 439.0, + 1405.0, + 475.0, + 323.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 470.0, + 789.0, + 470.0, + 789.0, + 504.0, + 324.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 1238.0, + 261.0, + 1238.0, + 298.0, + 323.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1787.0, + 1408.0, + 1787.0, + 1408.0, + 1829.0, + 295.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1819.0, + 1405.0, + 1819.0, + 1405.0, + 1858.0, + 321.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1853.0, + 882.0, + 1853.0, + 882.0, + 1887.0, + 325.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 315.0, + 1406.0, + 315.0, + 1406.0, + 358.0, + 292.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 350.0, + 597.0, + 350.0, + 597.0, + 380.0, + 322.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 529.0, + 1405.0, + 529.0, + 1405.0, + 566.0, + 294.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 563.0, + 1402.0, + 563.0, + 1402.0, + 593.0, + 324.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 591.0, + 803.0, + 591.0, + 803.0, + 624.0, + 323.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1310.0, + 1294.0, + 1310.0, + 1294.0, + 1349.0, + 295.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1369.0, + 1406.0, + 1369.0, + 1406.0, + 1408.0, + 295.0, + 1408.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 105, + 300, + 105 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 291, + 912, + 1407, + 912, + 1407, + 973, + 291, + 973 + ], + "score": 0.798 + }, + { + "category_id": 1, + "poly": [ + 299, + 230, + 1409, + 230, + 1409, + 384, + 299, + 384 + ], + "score": 0.773 + }, + { + "category_id": 1, + "poly": [ + 298, + 828, + 1404, + 828, + 1404, + 891, + 298, + 891 + ], + "score": 0.712 + }, + { + "category_id": 1, + "poly": [ + 296, + 661, + 1400, + 661, + 1400, + 726, + 296, + 726 + ], + "score": 0.637 + }, + { + "category_id": 1, + "poly": [ + 297, + 744, + 1404, + 744, + 1404, + 808, + 297, + 808 + ], + "score": 0.632 + }, + { + "category_id": 1, + "poly": [ + 296, + 547, + 1411, + 547, + 1411, + 641, + 296, + 641 + ], + "score": 0.629 + }, + { + "category_id": 1, + "poly": [ + 298, + 403, + 1407, + 403, + 1407, + 527, + 298, + 527 + ], + "score": 0.619 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 905.0, + 1411.0, + 905.0, + 1411.0, + 949.0, + 290.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 939.0, + 397.0, + 939.0, + 397.0, + 974.0, + 321.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1409.0, + 229.0, + 1409.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 257.0, + 1409.0, + 257.0, + 1409.0, + 298.0, + 321.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 286.0, + 1410.0, + 286.0, + 1410.0, + 331.0, + 318.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 356.0, + 322.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 353.0, + 1392.0, + 353.0, + 1392.0, + 386.0, + 325.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 828.0, + 1408.0, + 828.0, + 1408.0, + 863.0, + 296.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 858.0, + 1410.0, + 858.0, + 1410.0, + 893.0, + 322.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 659.0, + 1404.0, + 659.0, + 1404.0, + 699.0, + 294.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 692.0, + 1199.0, + 692.0, + 1199.0, + 726.0, + 322.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 739.0, + 1410.0, + 739.0, + 1410.0, + 785.0, + 292.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 775.0, + 1366.0, + 775.0, + 1366.0, + 811.0, + 324.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 544.0, + 1410.0, + 544.0, + 1410.0, + 586.0, + 293.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 576.0, + 1412.0, + 576.0, + 1412.0, + 616.0, + 321.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 610.0, + 888.0, + 610.0, + 888.0, + 644.0, + 323.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 403.0, + 1408.0, + 403.0, + 1408.0, + 439.0, + 293.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 436.0, + 1409.0, + 436.0, + 1409.0, + 469.0, + 322.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 464.0, + 1412.0, + 464.0, + 1412.0, + 500.0, + 321.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 494.0, + 918.0, + 494.0, + 918.0, + 532.0, + 322.0, + 532.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 295, + 1406, + 295, + 1406, + 887, + 297, + 887 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 902, + 1404, + 902, + 1404, + 1118, + 298, + 1118 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1131, + 1403, + 1131, + 1403, + 1316, + 298, + 1316 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1331, + 1403, + 1331, + 1403, + 1486, + 299, + 1486 + ], + "score": 0.972 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.838 + }, + { + "category_id": 0, + "poly": [ + 296, + 226, + 1324, + 226, + 1324, + 262, + 296, + 262 + ], + "score": 0.764 + }, + { + "category_id": 1, + "poly": [ + 296, + 226, + 1324, + 226, + 1324, + 262, + 296, + 262 + ], + "score": 0.135 + }, + { + "category_id": 13, + "poly": [ + 514, + 605, + 730, + 605, + 730, + 641, + 514, + 641 + ], + "score": 0.92, + "latex": "i \\in \\{ 0 , . . . , N - 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 698, + 793, + 874, + 793, + 874, + 824, + 698, + 824 + ], + "score": 0.92, + "latex": "\\epsilon = \\dot { 1 . 5 } \\times 1 0 ^ { - 5 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 600, + 466, + 600, + 466, + 638, + 298, + 638 + ], + "score": 0.92, + "latex": "\\epsilon _ { i } = \\bar { \\epsilon } ^ { 1 + \\frac { 1 } { N - 1 } \\alpha }" + }, + { + "category_id": 13, + "poly": [ + 444, + 793, + 646, + 793, + 646, + 824, + 444, + 824 + ], + "score": 0.91, + "latex": "l r = 6 . 2 5 \\times 1 0 ^ { - 5 }" + }, + { + "category_id": 13, + "poly": [ + 650, + 668, + 953, + 668, + 953, + 706, + 650, + 706 + ], + "score": 0.9, + "latex": "\\delta _ { e } = \\eta \\operatorname* { m a x } _ { t } \\delta _ { i } + ( 1 - \\eta ) \\hat { \\delta }" + }, + { + "category_id": 13, + "poly": [ + 1099, + 607, + 1177, + 607, + 1177, + 636, + 1099, + 636 + ], + "score": 0.89, + "latex": "\\alpha = 7" + }, + { + "category_id": 13, + "poly": [ + 1032, + 673, + 1124, + 673, + 1124, + 706, + 1032, + 706 + ], + "score": 0.89, + "latex": "\\eta = 0 . 9" + }, + { + "category_id": 13, + "poly": [ + 951, + 607, + 1045, + 607, + 1045, + 636, + 951, + 636 + ], + "score": 0.89, + "latex": "\\epsilon = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 1028, + 639, + 1053, + 639, + 1053, + 668, + 1028, + 668 + ], + "score": 0.87, + "latex": "\\delta _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1146, + 576, + 1170, + 576, + 1170, + 602, + 1146, + 602 + ], + "score": 0.84, + "latex": "\\epsilon _ { i }" + }, + { + "category_id": 13, + "poly": [ + 369, + 769, + 390, + 769, + 390, + 797, + 369, + 797 + ], + "score": 0.83, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 325, + 1222, + 366, + 1222, + 366, + 1251, + 325, + 1251 + ], + "score": 0.83, + "latex": "2 ^ { 1 6 }" + }, + { + "category_id": 13, + "poly": [ + 777, + 1222, + 816, + 1222, + 816, + 1252, + 777, + 1252 + ], + "score": 0.83, + "latex": "2 ^ { 1 5 }" + }, + { + "category_id": 13, + "poly": [ + 551, + 1053, + 694, + 1053, + 694, + 1088, + 551, + 1088 + ], + "score": 0.83, + "latex": "2 ^ { 1 7 } ( 1 3 1 0 7 2 )" + }, + { + "category_id": 13, + "poly": [ + 1027, + 710, + 1047, + 710, + 1047, + 731, + 1027, + 731 + ], + "score": 0.74, + "latex": "n" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 867.0, + 2084.0, + 867.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 1329.0, + 224.0, + 1329.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 296.0, + 1406.0, + 296.0, + 1406.0, + 331.0, + 295.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 325.0, + 1409.0, + 325.0, + 1409.0, + 362.0, + 291.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 357.0, + 1404.0, + 357.0, + 1404.0, + 391.0, + 294.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 386.0, + 1406.0, + 386.0, + 1406.0, + 422.0, + 294.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 418.0, + 1408.0, + 418.0, + 1408.0, + 454.0, + 294.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 449.0, + 1406.0, + 449.0, + 1406.0, + 484.0, + 295.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 478.0, + 1406.0, + 478.0, + 1406.0, + 513.0, + 292.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 511.0, + 1403.0, + 511.0, + 1403.0, + 543.0, + 295.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 536.0, + 1407.0, + 536.0, + 1407.0, + 578.0, + 292.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 567.0, + 1145.0, + 567.0, + 1145.0, + 609.0, + 292.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 567.0, + 1406.0, + 567.0, + 1406.0, + 609.0, + 1171.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 597.0, + 297.0, + 597.0, + 297.0, + 647.0, + 287.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 597.0, + 513.0, + 597.0, + 513.0, + 647.0, + 467.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 597.0, + 950.0, + 597.0, + 950.0, + 647.0, + 731.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 597.0, + 1098.0, + 597.0, + 1098.0, + 647.0, + 1046.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 597.0, + 1409.0, + 597.0, + 1409.0, + 647.0, + 1178.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 637.0, + 1027.0, + 637.0, + 1027.0, + 675.0, + 292.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 637.0, + 1406.0, + 637.0, + 1406.0, + 675.0, + 1054.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 672.0, + 649.0, + 672.0, + 649.0, + 709.0, + 294.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 672.0, + 1031.0, + 672.0, + 1031.0, + 709.0, + 954.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 672.0, + 1407.0, + 672.0, + 1407.0, + 709.0, + 1125.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 702.0, + 1026.0, + 702.0, + 1026.0, + 741.0, + 292.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 702.0, + 1407.0, + 702.0, + 1407.0, + 741.0, + 1048.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 733.0, + 1407.0, + 733.0, + 1407.0, + 769.0, + 295.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 764.0, + 368.0, + 764.0, + 368.0, + 800.0, + 294.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 764.0, + 1404.0, + 764.0, + 1404.0, + 800.0, + 391.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 789.0, + 443.0, + 789.0, + 443.0, + 831.0, + 291.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 789.0, + 697.0, + 789.0, + 697.0, + 831.0, + 647.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 789.0, + 1407.0, + 789.0, + 1407.0, + 831.0, + 875.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 825.0, + 1404.0, + 825.0, + 1404.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 856.0, + 859.0, + 856.0, + 859.0, + 891.0, + 292.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 903.0, + 1405.0, + 903.0, + 1405.0, + 937.0, + 294.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 933.0, + 1405.0, + 933.0, + 1405.0, + 968.0, + 294.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 961.0, + 1408.0, + 961.0, + 1408.0, + 1000.0, + 292.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 994.0, + 1404.0, + 994.0, + 1404.0, + 1030.0, + 293.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1025.0, + 1405.0, + 1025.0, + 1405.0, + 1059.0, + 294.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1051.0, + 550.0, + 1051.0, + 550.0, + 1093.0, + 292.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1051.0, + 1408.0, + 1051.0, + 1408.0, + 1093.0, + 695.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1083.0, + 1384.0, + 1083.0, + 1384.0, + 1122.0, + 292.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1133.0, + 1404.0, + 1133.0, + 1404.0, + 1165.0, + 296.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1162.0, + 1405.0, + 1162.0, + 1405.0, + 1198.0, + 293.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1193.0, + 1405.0, + 1193.0, + 1405.0, + 1227.0, + 293.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1220.0, + 324.0, + 1220.0, + 324.0, + 1259.0, + 292.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1220.0, + 776.0, + 1220.0, + 776.0, + 1259.0, + 367.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1220.0, + 1407.0, + 1220.0, + 1407.0, + 1259.0, + 817.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1251.0, + 1405.0, + 1251.0, + 1405.0, + 1290.0, + 292.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1285.0, + 422.0, + 1285.0, + 422.0, + 1318.0, + 293.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1331.0, + 1404.0, + 1331.0, + 1404.0, + 1365.0, + 297.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1361.0, + 1404.0, + 1361.0, + 1404.0, + 1397.0, + 294.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1391.0, + 1404.0, + 1391.0, + 1404.0, + 1427.0, + 294.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1420.0, + 1408.0, + 1420.0, + 1408.0, + 1462.0, + 293.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1451.0, + 573.0, + 1451.0, + 573.0, + 1494.0, + 293.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 1329.0, + 224.0, + 1329.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 308, + 390, + 1381, + 390, + 1381, + 1366, + 308, + 1366 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 708, + 1394, + 991, + 1394, + 991, + 1427, + 708, + 1427 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2112, + 835, + 2112 + ], + "score": 0.847 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 824, + 226, + 824, + 262, + 300, + 262 + ], + "score": 0.835 + }, + { + "category_id": 4, + "poly": [ + 298, + 295, + 1402, + 295, + 1402, + 358, + 298, + 358 + ], + "score": 0.648 + }, + { + "category_id": 1, + "poly": [ + 298, + 295, + 1402, + 295, + 1402, + 358, + 298, + 358 + ], + "score": 0.225 + }, + { + "category_id": 15, + "poly": [ + 312.0, + 419.0, + 346.0, + 419.0, + 346.0, + 448.0, + 312.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 419.0, + 896.0, + 419.0, + 896.0, + 448.0, + 862.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 488.0, + 347.0, + 488.0, + 347.0, + 521.0, + 309.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 489.0, + 897.0, + 489.0, + 897.0, + 523.0, + 860.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 560.0, + 347.0, + 560.0, + 347.0, + 595.0, + 309.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 562.0, + 896.0, + 562.0, + 896.0, + 592.0, + 862.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 632.0, + 346.0, + 632.0, + 346.0, + 667.0, + 311.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 633.0, + 897.0, + 633.0, + 897.0, + 667.0, + 859.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 704.0, + 348.0, + 704.0, + 348.0, + 736.0, + 311.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 692.0, + 787.0, + 692.0, + 787.0, + 745.0, + 733.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 705.0, + 898.0, + 705.0, + 898.0, + 737.0, + 860.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 692.0, + 1338.0, + 692.0, + 1338.0, + 745.0, + 1283.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 731.0, + 786.0, + 731.0, + 786.0, + 766.0, + 733.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 731.0, + 1337.0, + 731.0, + 1337.0, + 766.0, + 1282.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 756.0, + 825.0, + 756.0, + 825.0, + 787.0, + 735.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 756.0, + 1374.0, + 756.0, + 1374.0, + 787.0, + 1286.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 788.0, + 430.0, + 788.0, + 430.0, + 819.0, + 388.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 788.0, + 497.0, + 788.0, + 497.0, + 819.0, + 454.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 789.0, + 561.0, + 789.0, + 561.0, + 817.0, + 522.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 789.0, + 628.0, + 789.0, + 628.0, + 817.0, + 588.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 788.0, + 697.0, + 788.0, + 697.0, + 818.0, + 653.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 787.0, + 763.0, + 787.0, + 763.0, + 819.0, + 717.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 788.0, + 829.0, + 788.0, + 829.0, + 819.0, + 783.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 772.0, + 919.0, + 772.0, + 919.0, + 823.0, + 855.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 787.0, + 989.0, + 787.0, + 989.0, + 819.0, + 945.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 788.0, + 1064.0, + 788.0, + 1064.0, + 819.0, + 1022.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 788.0, + 1137.0, + 788.0, + 1137.0, + 816.0, + 1096.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 788.0, + 1214.0, + 788.0, + 1214.0, + 818.0, + 1170.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 787.0, + 1288.0, + 787.0, + 1288.0, + 819.0, + 1244.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 786.0, + 1363.0, + 786.0, + 1363.0, + 822.0, + 1316.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 805.0, + 644.0, + 805.0, + 644.0, + 836.0, + 527.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 805.0, + 1196.0, + 805.0, + 1196.0, + 836.0, + 1078.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 843.0, + 644.0, + 843.0, + 644.0, + 883.0, + 508.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 843.0, + 1194.0, + 843.0, + 1194.0, + 883.0, + 1057.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 882.0, + 895.0, + 882.0, + 895.0, + 909.0, + 862.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 910.0, + 346.0, + 910.0, + 346.0, + 939.0, + 312.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 960.0, + 897.0, + 960.0, + 897.0, + 987.0, + 862.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 980.0, + 347.0, + 980.0, + 347.0, + 1012.0, + 309.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1037.0, + 896.0, + 1037.0, + 896.0, + 1065.0, + 862.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1053.0, + 346.0, + 1053.0, + 346.0, + 1083.0, + 312.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1123.0, + 346.0, + 1123.0, + 346.0, + 1157.0, + 311.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1112.0, + 898.0, + 1112.0, + 898.0, + 1145.0, + 860.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1194.0, + 348.0, + 1194.0, + 348.0, + 1227.0, + 309.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1182.0, + 786.0, + 1182.0, + 786.0, + 1233.0, + 733.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1190.0, + 898.0, + 1190.0, + 898.0, + 1223.0, + 860.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1184.0, + 1330.0, + 1184.0, + 1330.0, + 1237.0, + 1277.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1223.0, + 786.0, + 1223.0, + 786.0, + 1257.0, + 733.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1224.0, + 1328.0, + 1224.0, + 1328.0, + 1259.0, + 1276.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1244.0, + 825.0, + 1244.0, + 825.0, + 1279.0, + 735.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1247.0, + 1367.0, + 1247.0, + 1367.0, + 1279.0, + 1277.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1278.0, + 439.0, + 1278.0, + 439.0, + 1309.0, + 399.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1278.0, + 516.0, + 1278.0, + 516.0, + 1309.0, + 474.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1279.0, + 590.0, + 1279.0, + 590.0, + 1307.0, + 550.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1278.0, + 667.0, + 1278.0, + 667.0, + 1309.0, + 625.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1278.0, + 743.0, + 1278.0, + 743.0, + 1309.0, + 700.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1278.0, + 819.0, + 1278.0, + 819.0, + 1310.0, + 776.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1266.0, + 915.0, + 1266.0, + 915.0, + 1312.0, + 857.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1280.0, + 991.0, + 1280.0, + 991.0, + 1308.0, + 952.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1282.0, + 1069.0, + 1282.0, + 1069.0, + 1306.0, + 1031.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1279.0, + 1151.0, + 1279.0, + 1151.0, + 1307.0, + 1111.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1278.0, + 1232.0, + 1278.0, + 1232.0, + 1309.0, + 1189.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 1278.0, + 1311.0, + 1278.0, + 1311.0, + 1310.0, + 1270.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 1282.0, + 1388.0, + 1282.0, + 1388.0, + 1308.0, + 1350.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1294.0, + 647.0, + 1294.0, + 647.0, + 1328.0, + 527.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1294.0, + 1190.0, + 1294.0, + 1190.0, + 1328.0, + 1073.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1333.0, + 645.0, + 1333.0, + 645.0, + 1376.0, + 506.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1329.0, + 1195.0, + 1329.0, + 1195.0, + 1377.0, + 1054.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 778.5, + 351.0, + 778.5, + 351.0, + 817.0, + 322.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.25, + 1269.5, + 351.25, + 1269.5, + 351.25, + 1308.5, + 320.25, + 1308.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1390.0, + 994.0, + 1390.0, + 994.0, + 1432.0, + 706.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 831.0, + 224.0, + 831.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 292.0, + 1406.0, + 292.0, + 1406.0, + 335.0, + 292.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 327.0, + 791.0, + 327.0, + 791.0, + 362.0, + 293.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 292.0, + 1406.0, + 292.0, + 1406.0, + 335.0, + 292.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 327.0, + 791.0, + 327.0, + 791.0, + 362.0, + 293.0, + 362.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/BkM27IxR-/BkM27IxR-.md b/parse/train/BkM27IxR-/BkM27IxR-.md new file mode 100644 index 0000000000000000000000000000000000000000..0e734d484f48140615a1bf3e5155f5d2e8bec3cd --- /dev/null +++ b/parse/train/BkM27IxR-/BkM27IxR-.md @@ -0,0 +1,274 @@ +# LEARNING TO OPTIMIZE NEURAL NETS + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Learning to Optimize (Li & Malik, 2016) is a recently proposed framework for learning optimization algorithms using reinforcement learning. In this paper, we explore learning an optimization algorithm for training shallow neural nets. Such high-dimensional stochastic optimization problems present interesting challenges for existing reinforcement learning algorithms. We develop an extension that is suited to learning optimization algorithms in this setting and demonstrate that the learned optimization algorithm consistently outperforms other known optimization algorithms even on unseen tasks and is robust to changes in stochasticity of gradients and the neural net architecture. More specifically, we show that an optimization algorithm trained with the proposed method on the problem of training a neural net on MNIST generalizes to the problems of training neural nets on the Toronto Faces Dataset, CIFAR-10 and CIFAR-100. + +# 1 INTRODUCTION + +Machine learning is centred on the philosophy that learning patterns automatically from data is generally better than meticulously crafting rules by hand. This data-driven approach has delivered: today, machine learning techniques can be found in a wide range of application areas, both in AI and beyond. Yet, there is one domain that has conspicuously been left untouched by machine learning: the design of tools that power machine learning itself. + +One of the most widely used tools in machine learning is optimization algorithms. We have grown accustomed to seeing an optimization algorithm as a black box that takes in a model that we design and the data that we collect and outputs the optimal model parameters. The optimization algorithm itself largely stays static: its design is reserved for human experts, who must toil through many rounds of theoretical analysis and empirical validation to devise a better optimization algorithm. Given this state of affairs, perhaps it is time for us to start practicing what we preach and learn how to learn. + +Recently, Li & Malik (2016) and Andrychowicz et al. (2016) introduced two different frameworks for learning optimization algorithms. Whereas Andrychowicz et al. (2016) focuses on learning an optimization algorithm for training models on a particular task, Li & Malik (2016) sets a more ambitious objective of learning an optimization algorithm for training models that is task-independent. We study the latter paradigm in this paper and develop a method for learning an optimization algorithm for high-dimensional stochastic optimization problems, like the problem of training shallow neural nets. + +Under the “Learning to Optimize” framework proposed by Li & Malik (2016), the problem of learning an optimization algorithm is formulated as a reinforcement learning problem. We consider the general structure of an unconstrained continuous optimization algorithm, as shown in Algorithm 1. In each iteration, the algorithm takes a step $\Delta x$ and uses it to update the current iterate $\boldsymbol { x } ^ { ( i ) }$ . In hand-engineered optimization algorithms, $\Delta x$ is computed using some fixed formula $\phi$ that depends on the objective function, the current iterate and past iterates. Often, it is simply a function of the current and past gradients. + +Different choices of $\phi$ yield different optimization algorithms and so each optimization algorithm is essentially characterized by its update formula $\phi$ . Hence, by learning $\phi$ , we can learn an optimization algorithm. Li & Malik (2016) observed that an optimization algorithm can be viewed as a Markov decision process (MDP), where the state includes the current iterate, the action is the step vector $\Delta x$ + +Require: Objective function $f$ $\bar { x ^ { ( 0 ) } } \gets$ random point in the domain of $f$ for $i = 1 , 2 , \dots { \bf d }$ o $\Delta x \gets \acute { \phi } ( f , \{ x ^ { ( 0 ) } , \ldots , x ^ { ( i - 1 ) } \} )$ if stopping condition is met then return x(i−1) end if $x ^ { ( i ) } \gets x ^ { ( i - 1 ) } + \Delta x$ end for + +and the policy is the update formula $\phi$ . Hence, the problem of learning $\phi$ simply reduces to a policy search problem. + +In this paper, we build on the method proposed in (Li & Malik, 2016) and develop an extension that is suited to learning optimization algorithms for high-dimensional stochastic problems. We use it to learn an optimization algorithm for training shallow neural nets and show that it outperforms popular hand-engineered optimization algorithms like ADAM (Kingma & Ba, 2014), AdaGrad (Duchi et al., 2011) and RMSprop (Tieleman & Hinton, 2012) and an optimization algorithm learned using the supervised learning method proposed in (Andrychowicz et al., 2016). Furthermore, we demonstrate that our optimization algorithm learned from the experience of training on MNIST generalizes to training on other datasets that have very dissimilar statistics, like the Toronto Faces Dataset, CIFAR10 and CIFAR-100. + +# 2 RELATED WORK + +The line of work on learning optimization algorithms is fairly recent. Li & Malik (2016) and Andrychowicz et al. (2016) were the first to propose learning general optimization algorithms. Li & Malik (2016) explored learning task-independent optimization algorithms and used reinforcement learning to learn the optimization algorithm, while Andrychowicz et al. (2016) investigated learning task-dependent optimization algorithms and used supervised learning. + +In the special case where objective functions that the optimization algorithm is trained on are loss functions for training other models, these methods can be used for “learning to learn” or “metalearning”. While these terms have appeared from time to time in the literature (Baxter et al., 1995; Vilalta & Drissi, 2002; Brazdil et al., 2008; Thrun & Pratt, 2012), they have been used by different authors to refer to disparate methods with different purposes. These methods all share the objective of learning some form of meta-knowledge about learning, but differ in the type of meta-knowledge they aim to learn. We can divide the various methods into the following three categories. + +# 2.1 LEARNING WHAT TO LEARN + +Methods in this category Thrun & Pratt (2012) aim to learn what parameter values of the base-level learner are useful across a family of related tasks. The meta-knowledge captures commonalities shared by tasks in the family, which enables learning on a new task from the family to be performed more quickly. Most early methods fall into this category; this line of work has blossomed into an area that has later become known as transfer learning and multi-task learning. + +# 2.2 LEARNING WHICH MODEL TO LEARN + +Methods in this category Brazdil et al. (2008) aim to learn which base-level learner achieves the best performance on a task. The meta-knowledge captures correlations between different tasks and the performance of different base-level learners on those tasks. One challenge under this setting is to decide on a parameterization of the space of base-level learners that is both rich enough to be capable of representing disparate base-level learners and compact enough to permit tractable search over this space. Brazdil et al. (2003) proposes a nonparametric representation and stores examples of different base-level learners in a database, whereas Schmidhuber (2004) proposes representing baselevel learners as general-purpose programs. The former has limited representation power, while the latter makes search and learning in the space of base-level learners intractable. Hochreiter et al. (2001) views the (online) training procedure of any base-learner as a black box function that maps a sequence of training examples to a sequence of predictions and models it as a recurrent neural net. Under this formulation, meta-training reduces to training the recurrent net, and the base-level learner is encoded in the memory state of the recurrent net. + +Hyperparameter optimization can be seen as another example of methods in this category. The space of base-level learners to search over is parameterized by a predefined set of hyperparameters. Unlike the methods above, multiple trials with different hyperparameter settings on the same task are permitted, and so generalization across tasks is not required. The discovered hyperparameters are generally specific to the task at hand and hyperparameter optimization must be rerun for new tasks. Various kinds of methods have been proposed, such those based on Bayesian optimization (Hutter et al., 2011; Bergstra et al., 2011; Snoek et al., 2012; Swersky et al., 2013; Feurer et al., 2015), random search (Bergstra & Bengio, 2012) and gradient-based optimization (Bengio, 2000; Domke, 2012; Maclaurin et al., 2015). + +# 2.3 LEARNING HOW TO LEARN + +Methods in this category aim to learn a good algorithm for training a base-level learner. Unlike methods in the previous categories, the goal is not to learn about the outcome of learning, but rather the process of learning. The meta-knowledge captures commonalities in the behaviours of learning algorithms that achieve good performance. The base-level learner and the task are given by the user, so the learned algorithm must generalize across base-level learners and tasks. Since learning in most cases is equivalent to optimizing some objective function, learning a learning algorithm often reduces to learning an optimization algorithm. This problem was explored in (Li & Malik, 2016) and (Andrychowicz et al., 2016). Closely related is (Bengio et al., 1991), which learns a Hebblike synaptic learning rule that does not depend on the objective function, which does not allow for generalization to different objective functions. + +Various work has explored learning how to adjust the hyperparameters of hand-engineered optimization algorithms, like the step size (Hansen, 2016; Daniel et al., 2016; Fu et al., 2016) or the damping factor in the Levenberg-Marquardt algorithm (Ruvolo et al., 2009). Related to this line of work is stochastic meta-descent (Bray et al., 2004), which derives a rule for adjusting the step size analytically. A different line of work (Gregor & LeCun, 2010; Sprechmann et al., 2013) parameterizes intermediate operands of special-purpose solvers for a class of optimization problems that arise in sparse coding and learns them using supervised learning. + +# 3 LEARNING TO OPTIMIZE + +# 3.1 SETTING + +In the “Learning to Optimize” framework, we are given a set of training objective functions $f _ { 1 } , \ldots , f _ { n }$ drawn from some distribution $\mathcal { F }$ . An optimization algorithm $\mathcal { P }$ takes an objective function $f$ and an initial iterate $x ^ { ( 0 ) }$ as input and produces a sequence of iterates $x ^ { ( 1 ) } , \ldots , x ^ { ( T ) }$ , where $x ^ { ( T ) }$ is the solution found by the optimizer. We are also given a distribution $\mathcal { D }$ that generates the initial iterate $x ^ { ( 0 ) }$ and a meta-loss $\mathcal { L }$ , which takes an objective function $f$ and a sequence of iterates $x ^ { ( 1 ) } , \ldots , x ^ { ( T ) }$ produced by an optimization algorithm as input and outputs a scalar that measures the quality of the iterates. The goal is to learn an optimization algorithm ${ \mathcal { P } } ^ { * }$ such that $\mathbb { E } _ { f \sim \mathcal { F } , x ^ { ( 0 ) } \sim \mathcal { D } } \left[ \mathcal { L } ( f , \mathcal { P } ^ { * } ( f , x ^ { ( 0 ) } ) ) \right]$ is minimized. The meta-loss is chosen to penalize optimization algorithms that exhibit behaviours we find undesirable, like slow convergence or excessive oscillations. Assuming we would like to learn an algoritha good choice of meta-loss would then simply be regret and can be interpreted as the area under the $\textstyle \sum _ { i = 1 } ^ { T } f ( x ^ { ( i ) } )$ zes the objective function it is given,, which is equivalent to cumulativective values over time. + +The objective functions $f _ { 1 } , \ldots , f _ { n }$ may correspond to loss functions for training base-level learners, in which case the algorithm that learns the optimization algorithm can be viewed as a meta-learner. In this setting, each objective function is the loss function for training a particular base-learner on a particular task, and so the set of training objective functions can be loss functions for training a base-learner or a family of base-learners on different tasks. At test time, the learned optimization algorithm is evaluated on unseen objective functions, which correspond to loss functions for training base-learners on new tasks, which may be completely unrelated to tasks used for training the optimization algorithm. Therefore, the learned optimization algorithm must not learn anything about the tasks used for training. Instead, the goal is to learn an optimization algorithm that can exploit the geometric structure of the error surface induced by the base-learners. For example, if the base-level model is a neural net with ReLU activation units, the optimization algorithm should hopefully learn to leverage the piecewise linearity of the model. Hence, there is a clear division of responsibilities between the meta-learner and base-learners. The knowledge learned at the meta-level should be pertinent for all tasks, whereas the knowledge learned at the base-level should be task-specific. The meta-learner should therefore generalize across tasks, whereas the base-learner should generalize across instances. + +# 3.2 RL PRELIMINARIES + +The goal of reinforcement learning is to learn to interact with an environment in a way that minimizes cumulative costs that are expected to be incurred over time. The environment is formalized as a partially observable Markov decision process $( \mathrm { P O M D P } ) ^ { 1 }$ , which is defined by the tuple $( S , \mathcal { O } , \mathcal { A } , p _ { i } , p , p _ { o } , c , T )$ , where $\mathcal { S } \subseteq \mathbb { R } ^ { D }$ is the set of states, $\mathcal { O } \subseteq \mathbb { R } ^ { D ^ { \prime } }$ is the set of observations, $\mathcal { A } \subseteq \mathbb { R } ^ { d }$ is the set of actions, $p _ { i } \left( s _ { 0 } \right)$ is the probability density over initial states $s _ { 0 }$ , $p \left( { { s } _ { t + 1 } } \left| { { s } _ { t } } , { { a } _ { t } } \right. \right)$ is the probability density over the subsequent state $s _ { t + 1 }$ given the current state $s _ { t }$ and action $a _ { t }$ , $p _ { o } \left( o _ { t } \bar { | } _ { s _ { t } } \right)$ is the probability density over the current observation $o _ { t }$ given the current state $s _ { t }$ , $c : { \mathcal { S } } \mathbb { R }$ is a function that assigns a cost to each state and $T$ is the time horizon. Often, the probability densities $p$ and $p _ { o }$ are unknown and not given to the learning algorithm. + +A policy $\pi \left( \boldsymbol { a } _ { t } | \boldsymbol { o } _ { t } , t \right)$ is a conditional probability density over actions $a _ { t }$ given the current observation $o _ { t }$ and time step $t$ . When a policy is independent of $t$ , it is known as a stationary policy. The goal of the reinforcement learning algorithm is to learn a policy $\pi ^ { * }$ that minimizes the total expected cost over time. More precisely, + +$$ +\pi ^ { * } = \arg \operatorname* { m i n } _ { \pi } \mathbb { E } _ { s _ { 0 } , a _ { 0 } , s _ { 1 } , \dots , s _ { T } } \left[ \sum _ { t = 0 } ^ { T } c ( s _ { t } ) \right] , +$$ + +where the expectation is taken with respect to the joint distribution over the sequence of states and actions, often referred to as a trajectory, which has the density + +$$ +\begin{array} { c } { { \displaystyle q \big ( s _ { 0 } , a _ { 0 } , s _ { 1 } , \ldots , s _ { T } \big ) = \int _ { o _ { 0 } , \ldots , o _ { T } } p _ { i } \left( s _ { 0 } \right) p _ { o } \left( o _ { 0 } \big | s _ { 0 } \right) } } \\ { { { \displaystyle \prod _ { t = 0 } ^ { T - 1 } \pi \left( \left. a _ { t } \right| o _ { t } , t \right) p \left( \left. s _ { t + 1 } \right| s _ { t } , a _ { t } \right) p _ { o } \left( \left. o _ { t + 1 } \right| s _ { t + 1 } \right) . } } } \end{array} +$$ + +To make learning tractable, $\pi$ is often constrained to lie in a parameterized family. A common assumption is that $\pi \left( \left. a _ { t } \right| o _ { t } , t \right) = \mathcal { N } \left( \mu ^ { \pi } ( o _ { t } ) , \Sigma ^ { \pi } ( o _ { t } ) \right)$ , where $\textstyle { \mathcal { N } } ( { \boldsymbol { \mu } } , { \boldsymbol { \Sigma } } )$ denotes the density of a Gaussian with mean $\mu$ and covariance $\Sigma$ . The functions $\mu ^ { \pi } ( \cdot )$ and possibly $\Sigma ^ { \pi } ( \cdot )$ are modelled using function approximators, whose parameters are learned. + +# 3.3 FORMULATION + +In our setting, the state $s _ { t }$ consists of the current iterate $x ^ { ( t ) }$ and features $\Phi ( \cdot )$ that depend on the history of iterates $x ^ { ( 1 ) } , \ldots , x ^ { ( t ) }$ , (noisy) gradients $\nabla \hat { f } ( x ^ { ( 1 ) } ) , \ldots , \nabla \hat { f } ( x ^ { ( t ) } )$ and (noisy) objective values ${ \hat { f } } ( x ^ { ( 1 ) } ) , \ldots , { \hat { f } } ( x ^ { ( t ) } )$ . The action $a _ { t }$ is the step $\Delta x$ that will be used to update the iterate. The observation $o _ { t }$ excludes $x ^ { ( t ) }$ and consists of features $\Psi ( \cdot )$ that depend on the iterates, gradient and objective values from recent iterations, and the previous memory state of the learned optimization algorithm, which takes the form of a recurrent neural net. This memory state can be viewed as a statistic of the previous observations that is learned jointly with the policy. + +Under this formulation, the initial probability density $p _ { i }$ captures how the initial iterate, gradient and objective value tend to be distributed. The transition probability density $p$ captures the how the gradient and objective value are likely to change given the step that is taken currently; in other words, it encodes the local geometry of the training objective functions. Assuming the goal is to learn an optimization algorithm that minimizes the objective function, the cost $c$ of a state $\boldsymbol { s } _ { t } = \left( \boldsymbol { x } ^ { ( t ) } , \Phi \left( \cdot \right) \right) ^ { T }$ is simply the true objective value $f ( \boldsymbol { x } ^ { ( t ) } )$ . + +Any particular policy $\pi \left( \boldsymbol { a } _ { t } \left| \boldsymbol { o } _ { t } , t \right. \right)$ , which generates $a _ { t } ~ = ~ \Delta x$ at every time step, corresponds to a particular (noisy) update formula $\phi$ , and therefore a particular (noisy) optimization algorithm. Therefore, learning an optimization algorithm simply reduces to searching for the optimal policy. + +The mean of the policy is modelled as a recurrent neural net fragment that corresponds to a single time step, which takes the observation features $\Psi ( \cdot )$ and the previous memory state as input and outputs the step to take. + +# 3.4 GUIDED POLICY SEARCH + +The reinforcement learning method we use is guided policy search (GPS) (Levine et al., 2015), which is a policy search method designed for searching over large classes of expressive non-linear policies in continuous state and action spaces. It maintains two policies, $\psi$ and $\pi$ , where the former lies in a time-varying linear policy class in which the optimal policy can found in closed form, and the latter lies in a stationary non-linear policy class in which policy optimization is challenging. In each iteration, it performs policy optimization on $\psi$ , and uses the resulting policy as supervision to train $\pi$ . + +More precisely, GPS solves the following constrained optimization problem: + +$$ +\operatorname* { m i n } _ { \theta , \eta } \mathbb { E } _ { \psi } \left[ \sum _ { t = 0 } ^ { T } c ( s _ { t } ) \right] \mathrm { ~ s . t . ~ } \psi \left( \left. a _ { t } \right| s _ { t } , t ; \eta \right) = \pi \left( \left. a _ { t } \right| s _ { t } ; \theta \right) \forall a _ { t } , s _ { t } , t +$$ + +where $\eta$ and $\theta$ denote the parameters of $\psi$ and $\pi$ respectively, $\mathbb { E } _ { \rho } \left[ \cdot \right]$ denotes the expectation taken with respect to the trajectory induced by a policy $\rho$ and $\begin{array} { r } { \pi \left( \left. a _ { t } \right| s _ { t } ; \theta \right) : = \int _ { o _ { t } } \pi \left( \left. a _ { t } \right| o _ { t } ; \theta \right) p _ { o } \left( \left. o _ { t } \right| s _ { t } \right) ^ { 2 } } \end{array}$ . + +Since there are an infinite number of equality constraints, the problem is relaxed by enforcing equality on the mean actions taken by $\psi$ and $\pi$ at every time step3. So, the problem becomes: + +$$ +\operatorname* { m i n } _ { \theta , \eta } \mathbb { E } _ { \psi } [ \sum _ { t = 0 } ^ { T } c ( s _ { t } ) ] \mathrm { ~ s . t . ~ } \mathbb { E } _ { \psi } [ a _ { t } ] = \mathbb { E } _ { \psi } [ \mathbb { E } _ { \pi } [ a _ { t } | s _ { t } ] ] \forall t +$$ + +This problem is solved using Bregman ADMM (Wang & Banerjee, 2014), which performs the following updates in each iteration: + +$$ +\begin{array} { r l } & { \eta \gets \arg \underset { \eta } { \operatorname* { m i n } } \sum _ { t = 0 } ^ { T } \mathbb { E } _ { \psi } [ c ( s _ { t } ) - \lambda _ { t } ^ { T } a _ { t } ] + \nu _ { t } D _ { t } ( \eta , \theta ) } \\ & { \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad } \\ & { \theta \gets \arg \underset { \theta } { \operatorname* { m i n } } \sum _ { t = 0 } ^ { T } \lambda _ { t } ^ { T } \mathbb { E } _ { \psi } [ \mathbb { E } _ { \pi } [ a _ { t } | s _ { t } ] ] + \nu _ { t } D _ { t } ( \theta , \eta ) } \\ & { \quad \quad \quad \quad \quad \quad \quad \lambda _ { t } \gets \lambda _ { t } + \alpha \nu _ { t } ( \mathbb { E } _ { \psi } [ \mathbb { E } _ { \pi } [ a _ { t } | s _ { t } ] ] - \mathbb { E } _ { \psi } [ a _ { t } ] ) \ \forall t , } \\ & { \mathrm { ~ } \mathrm { ~ } : \mathrm { r e } \quad D _ { t } ( \theta , \eta ) \quad : = \quad \mathbb { E } _ { \psi } [ D _ { K L } ( \pi ( a _ { t } | s _ { t } ; \theta ) ) ] \ \psi ( a _ { t } | s _ { t } , t ; \eta ) ) ] \quad \mathrm { ~ a n d ~ } \quad D _ { t } ( \eta , \theta ) } \\ & { [ D _ { K L } ( \psi ( a _ { t } | s _ { t } , t ; \eta ) ] ) | \ \pi ( a _ { t } | s _ { t } ; \theta ) ) | . } \end{array} +$$ + +The algorithm assumes that $\psi \left( \left. a _ { t } \right| s _ { t } , t ; \eta \right) = \mathcal { N } \left( K _ { t } s _ { t } + k _ { t } , G _ { t } \right)$ , where $\boldsymbol { \eta } : = \left( K _ { t } , k _ { t } , G _ { t } \right) _ { t = 1 } ^ { T }$ and $\pi \left( a _ { t } | o _ { t } ; \theta \right) = \mathcal { N } \left( \mu _ { \omega } ^ { \pi } ( o _ { t } ) , \Sigma ^ { \pi } \right)$ , where $\boldsymbol { \theta } : = ( \omega , \Sigma ^ { \pi } )$ and $\mu _ { \omega } ^ { \pi } ( \cdot )$ can be an arbitrary function that is typically modelled using a nonlinear function approximator like a neural net. + +At the start of each iteration, the algorithm constructs a model of the transition probability density $\tilde { p } \left( \left. s _ { t + 1 } \right| s _ { t } , a _ { t } , t ; \zeta \right) = \mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )$ , where $\boldsymbol { \zeta } : = \left( A _ { t } , B _ { t } , c _ { t } , F _ { t } \right) _ { t = 1 } ^ { T }$ is fitted to samples of $s _ { t }$ drawn from the trajectory induced by $\psi$ , which essentially amounts to a local linearization of the true transition probability $p \left( \left. s _ { t + 1 } \right| s _ { t } , a _ { t } , t \right)$ . We will use $\mathbb { E } _ { \widetilde { \psi } } \left[ \cdot \right]$ to denote expectation taken with respect to the trajectory induced by $\psi$ under the modelled transition probability $\tilde { p }$ . Additionally, the algorithm fits local quadratic approximations to $c ( s _ { t } )$ around samples of $s _ { t }$ drawn from the trajectory induced by $\psi$ so that $\begin{array} { r } { c ( s _ { t } ) \approx \dot { \tilde { c } } \dot { ( s _ { t } ) } : = \frac { 1 } { 2 } s _ { t } ^ { T } C _ { t } s _ { t } + \dot { d _ { t } ^ { T } } s _ { t } + h _ { t } } \end{array}$ for $s _ { t }$ ’s that are near the samples. + +With these assumptions, the subproblem that needs to be solved to update $\eta ~ = ~ ( K _ { t } , k _ { t } , G _ { t } ) _ { t = 1 } ^ { T }$ becomes: + +$$ +\begin{array} { r l r } & { \underset { \eta } { \operatorname* { m i n } } \displaystyle \sum _ { t = 0 } ^ { T } \mathbb { E } _ { \boldsymbol { \tilde { \psi } } } [ \boldsymbol { \tilde { c } } ( s _ { t } ) - \lambda _ { t } ^ { T } \boldsymbol { a } _ { t } ] + \nu _ { t } D _ { t } ( \eta , \theta ) } & \\ & { \mathrm { s . t . } \displaystyle \sum _ { t = 0 } ^ { T } \mathbb { E } _ { \boldsymbol { \tilde { \psi } } } [ D _ { K L } ( \boldsymbol { \psi } ( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } , t ; \eta ) | \Big | \psi ( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } , t ; \eta ^ { \prime } ) ) ] \leq \epsilon , } \end{array} +$$ + +where $\eta ^ { \prime }$ denotes the old $\eta$ from the previous iteration. Because $\tilde { p }$ and $\tilde { c }$ are only valid locally around the trajectory induced by $\psi$ , the constraint is added to limit the amount by which $\eta$ is updated. It turns out that the unconstrained problem can be solved in closed form using a dynamic programming algorithm known as linear-quadratic-Gaussian (LQG) regulator in time linear in the time horizon $T$ and cubic in the dimensionality of the state space $D$ . The constrained problem is solved using dual gradient descent, which uses LQG as a subroutine to solve for the primal variables in each iteration and increments the dual variable on the constraint until it is satisfied. + +Updating $\theta$ is straightforward, since expectations taken with respect to the trajectory induced by $\pi$ are always conditioned on $s _ { t }$ and all outer expectations over $s _ { t }$ are taken with respect to the trajectory induced by $\psi$ . Therefore, $\pi$ is essentially decoupled from the transition probability $p \left( \left. s _ { t + 1 } \right| s _ { t } , a _ { t } , t \right)$ and so its parameters can be updated without affecting the distribution of $s _ { t }$ ’s. The subproblem that needs to be solved to update $\theta$ therefore amounts to a standard supervised learning problem. + +Since $\psi \left( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } , t ; \boldsymbol { \eta } \right)$ and $\pi \left( \boldsymbol { a } _ { t } | \boldsymbol { s } _ { t } ; \boldsymbol { \theta } \right)$ are Gaussian, $D _ { t } \left( { \theta , \eta } \right)$ can be computed analytically. More concretely, if we assume $\Sigma ^ { \pi }$ to be fixed for simplicity, the subproblem that is solved for updating $\theta = \left( \omega , \Sigma ^ { \pi } \right)$ is: + +$$ +\begin{array} { l } { \displaystyle \operatorname* { m i n } _ { \theta } \mathbb { E } _ { \psi } [ \sum _ { t = 0 } ^ { T } \lambda _ { t } ^ { T } \mu _ { \omega } ^ { \pi } ( o _ { t } ) + \frac { \nu _ { t } } { 2 } ( \operatorname { t r } ( G _ { t } ^ { - 1 } \Sigma ^ { \pi } ) - \log | \Sigma ^ { \pi } | ) } \\ { \displaystyle + \frac { \nu _ { t } } { 2 } ( \mu _ { \omega } ^ { \pi } ( o _ { t } ) - \mathbb { E } _ { \psi } [ a _ { t } | s _ { t } , t ] ) ^ { T } G _ { t } ^ { - 1 } ( \mu _ { \omega } ^ { \pi } ( o _ { t } ) - \mathbb { E } _ { \psi } [ a _ { t } | s _ { t } , t ] ) ] } \end{array} +$$ + +Note that the last term is the squared Mahalanobis distance between the mean actions of $\psi$ and $\pi$ at time step $t$ , which is intuitive as we would like to encourage $\pi$ to match $\psi$ . + +# 3.5 CONVOLUTIONAL GPS + +The problem of learning high-dimensional optimization algorithms presents challenges for reinforcement learning algorithms due to high dimensionality of the state and action spaces. For example, in the case of GPS, because the running time of LQG is cubic in dimensionality of the state space, performing policy search even in the simple class of linear-Gaussian policies would be prohibitively expensive when the dimensionality of the optimization problem is high. + +Fortunately, many high-dimensional optimization problems have underlying structure that can be exploited. For example, the parameters of neural nets are equivalent up to permutation among certain coordinates. More concretely, for fully connected neural nets, the dimensions of a hidden layer and the corresponding weights can be permuted arbitrarily without changing the function they compute. Because permuting the dimensions of two adjacent layers can permute the weight matrix arbitrarily, an optimization algorithm should be invariant to permutations of the rows and columns of a weight matrix. A reasonable prior to impose is that the algorithm should behave in the same manner on all coordinates that correspond to entries in the same matrix. That is, if the values of two coordinates in all current and past gradients and iterates are identical, then the step vector produced by the algorithm should have identical values in these two coordinates. We will refer to the set of coordinates on which permutation invariance is enforced as a coordinate group. For the purposes of learning an optimization algorithm for neural nets, a natural choice would be to make each coordinate group correspond to a weight matrix or a bias vector. Hence, the total number of coordinate groups is twice the number of layers, which is usually fairly small. + +![](images/c32356121b006ac72da2adddd464187f503567b08623692799a8cd7bffe7f36e.jpg) +Figure 1: Comparison of the various hand-engineered and learned algorithms on training neural nets with 48 input and hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-batches of size 64. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. + +In the case of GPS, we impose this prior on both $\psi$ and $\pi$ . For the purposes of updating $\eta$ , we first impose a block-diagonal structure on the parameters $A _ { t } , B _ { t }$ and $F _ { t }$ of the fitted transition probability density $\tilde { p } \left( \left. s _ { t + 1 } \right| s _ { t } , a _ { t } , t ; \zeta \right) = \mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )$ , so that for each coordinate in the optimization problem, the dimensions of $s _ { t + 1 }$ that correspond to the coordinate only depend on the dimensions of $s _ { t }$ and $a _ { t }$ that correspond to the same coordinate. As a result, $\tilde { p } \big ( s _ { t + 1 } \big | s _ { t } , a _ { t } , t ; \zeta \big )$ decomposes into multiple independent probability densities $\tilde { p } ^ { j } \left( s _ { t + 1 } ^ { j } \Big | s _ { t } ^ { j } , a _ { t } ^ { j } , t ; \zeta ^ { j } \right)$ , one for each coordinate $j$ . Similarly, we also impose a block-diagonal structure on $C _ { t }$ for fitting $\tilde { c } ( s _ { t } )$ and on the parameter matrix of the fitted model for $\pi \left( a _ { t } | \boldsymbol s _ { t } \bar { ; \boldsymbol \theta } \right)$ . Under these assumptions, $K _ { t }$ and $G _ { t }$ are guaranteed to be block-diagonal as well. Hence, the Bregman divergence penalty term, $D \left( \eta , \theta \right)$ decomposes into a sum of Bregman divergence terms, one for each coordinate. + +We then further constrain dual variables $\lambda _ { t }$ , sub-vectors of parameter vectors and sub-matrices of parameter matrices corresponding to each coordinate group to be identical across the group. Additionally, we replace the weight $\nu _ { t }$ on $D \left( \eta , \theta \right)$ with an individual weight on each Bregman divergence term for each coordinate group. The problem then decomposes into multiple independent subproblems, one for each coordinate group. Because the dimensionality of the state subspace corresponding to each coordinate is constant, LQG can be executed on each subproblem much more efficiently. + +Similarly, for $\pi$ , we choose a $\mu _ { \omega } ^ { \pi } ( \cdot )$ that shares parameters across different coordinates in the same group. We also impose a block-diagonal structure on $\Sigma ^ { \pi }$ and constrain the appropriate sub-matrices to share their entries. + +# 3.6 FEATURES + +We describe the features $\Phi ( \cdot )$ and $\Psi ( \cdot )$ at time step $t$ , which define the state $s _ { t }$ and observation $o _ { t }$ respectively. + +Because of the stochasticity of gradients and objective values, the state features $\Phi ( \cdot )$ are defined in terms of summary statistics of the history of iterates $\left\{ x ^ { ( i ) } \right\} _ { i = 0 } ^ { t }$ , gradients $\left\{ \nabla \hat { f } ( x ^ { ( i ) } ) \right\} _ { i = 0 } ^ { t }$ and objective values $\left\{ \hat { f } ( \boldsymbol x ^ { ( i ) } ) \right\} _ { i = 0 } ^ { t }$ We define the following statistics, which we will refer to as the average recent iterate, gradient and objective value respectively: + +$$ +\begin{array} { r } { \overline { { \boldsymbol { x } ^ { ( i ) } } } : = \frac { 1 } { \operatorname* { m i n } ( i + 1 , 3 ) } \sum _ { j = \operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \boldsymbol { x } ^ { ( j ) } } \end{array} +$$ + +![](images/3fea8480e1dc0f7dbc3c3a6dc8993cac0f6ffab3eae77436f1e3e961bbffec59.jpg) +Figure 2: Comparison of the various hand-engineered and learned algorithms on training neural nets with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with minibatches of size 64. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. + +$$ +\begin{array} { r l } & { \bullet \ \overline { { \nabla \hat { f } ( x ^ { ( i ) } ) } } : = \frac { 1 } { \operatorname* { m i n } ( i + 1 , 3 ) } \sum _ { j = \operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } { \nabla \hat { f } ( x ^ { ( j ) } ) } } \\ & { \bullet \ \overline { { \hat { f } ( x ^ { ( i ) } ) } } : = \frac { 1 } { \operatorname* { m i n } ( i + 1 , 3 ) } \sum _ { j = \operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \hat { f } ( x ^ { ( j ) } ) } \end{array} +$$ + +The state features $\Phi ( \cdot )$ consist of the relative change in the average recent objective value, the average recent gradient normalized by the magnitude of the a previous average recent gradient and a previous change in average recent iterate relative to the current change in average recent iterate: + +$$ +\begin{array} { r l } & { \bullet \{ ( \hat { f } ( x ^ { ( t - 5 i ) } ) - \widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) ) / \widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) \} _ { i = 0 } ^ { 2 4 } } \\ & { \bullet \{ \overline { { \nabla \hat { f } ( x ^ { ( t - 5 i ) } ) } } / ( | \overline { { \nabla \hat { f } ( x ^ { ( m a x ( t - 5 ( i + 1 ) , t \mathrm { m o d } 5 ) ) } ) } } | + 1 ) \} _ { i = 0 } ^ { 2 5 } } \\ & \bullet \{ \frac { | \frac { \overline { { x ^ { ( \mathrm { m a x } ( t - 5 ( i + 1 ) , t \mathrm { m o d } 5 + 5 ) ) } } } - \overline { { x ^ { ( \mathrm { m a x } ( t - 5 ( i + 2 ) , t \mathrm { m o d } 5 ) ) } } } } { | \overline { { x ^ { ( t - 5 i ) } } } - \overline { { x ^ { ( t - 5 ( i + 1 ) ) } } } | + 0 . 1 } \} _ { i = 0 } ^ { 2 4 } \} _ { i = 0 } ^ { 2 } } \end{array} +$$ + +Note that all operations are applied element-wise. Also, whenever a feature becomes undefined (i.e.: when the time step index becomes negative), it is replaced with the all-zeros vector. + +Unlike state features, which are only used when training the optimization algorithm, observation features $\Psi ( \cdot )$ are used both during training and at test time. Consequently, we use noisier observation features that can be computed more efficiently and require less memory overhead. The observation features consist of the following: + +$$ +\begin{array} { r l } { \bullet } & { \left( \hat { f } ( x ^ { ( t ) } ) - \hat { f } ( x ^ { ( t - 1 ) } ) \right) / \hat { f } ( x ^ { ( t - 1 ) } ) } \\ & { \bullet \ \nabla \hat { f } ( x ^ { ( t ) } ) / \left( \left| \nabla \hat { f } ( x ^ { ( \operatorname* { m a x } ( t - 1 , 0 ) ) } ) \right| + 1 \right) } \\ & { \bullet \ \frac { \left| x ^ { ( \operatorname* { m a x } ( t - 1 , 1 ) ) } - x ^ { ( \operatorname* { m a x } ( t - 2 , 0 ) ) } \right| } { \left| x ^ { ( t ) } - x ^ { ( t - 1 ) } \right| + 0 . 1 } } \end{array} +$$ + +# 4 EXPERIMENTS + +For clarity, we will refer to training of the optimization algorithm as “meta-training” to differentiate it from base-level training, which will simply be referred to as “training”. + +We meta-trained an optimization algorithm on a single objective function, which corresponds to the problem of training a two-layer neural net with 48 input units, 48 hidden units and 10 output units on a randomly projected and normalized version of the MNIST training set with dimensionality 48 and unit variance in each dimension. We modelled the optimization algorithm using an recurrent neural net with a single layer of 128 LSTM (Hochreiter & Schmidhuber, 1997) cells. We used a time horizon of 400 iterations and a mini-batch size of 64 for computing stochastic gradients and objective values. We evaluate the optimization algorithm on its ability to generalize to unseen objective functions, which correspond to the problems of training neural nets on different tasks/datasets. + +![](images/dba70d81f44e9b66aeedbf35fb705247b4161ee34db83bed3b4ba0993a37fd89.jpg) +Figure 3: Comparison of the various hand-engineered and learned algorithms on training neural nets with 48 input and hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-batches of size 10. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. + +![](images/afba0d91683764d4cd4161719336a9f48fc43fabd99a2b8440e62626e6cabcb3.jpg) +Figure 4: Comparison of the various hand-engineered and learned algorithms on training neural nets with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with minibatches of size 10. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. + +We evaluate the learned optimization algorithm on three datasets, the Toronto Faces Dataset (TFD), CIFAR-10 and CIFAR-100. These datasets are chosen for their very different characteristics from MNIST and each other: TFD contains 3300 grayscale images that have relatively little variation and has seven different categories, whereas CIFAR-100 contains 50,000 colour images that have varied appearance and has 100 different categories. + +All algorithms are tuned on the training objective function. For hand-engineered algorithms, this entails choosing the best hyperparameters; for learned algorithms, this entails meta-training on the objective function. We compare to the seven hand-engineered algorithms: stochastic gradient descent, momentum, conjugate gradient, L-BFGS, ADAM, AdaGrad and RMSprop. In addition, we compare to an optimization algorithm meta-trained using the method described in (Andrychowicz et al., 2016) on the same training objective function (training two-layer neural net on randomly projected and normalized MNIST) under the same setting (a time horizon of 400 iterations and a mini-batch size of 64). + +First, we examine the performance of various optimization algorithms on similar objective functions. The optimization problems under consideration are those for training neural nets that have the same number of input and hidden units (48 and 48) as those used during meta-training. The number of output units varies with the number of categories in each dataset. We use the same mini-batch size as that used during meta-training. As shown in Figure 1, the optimization algorithm meta-trained using our method (which we will refer to as Predicted Step Descent) consistently descends to the optimum the fastest across all datasets. On the other hand, other algorithms are not as consistent and the relative ranking of other algorithms varies by dataset. This suggests that Predicted Step Descent has learned to be robust to variations in the data distributions, despite being trained on only one objective function, which is associated with a very specific data distribution that characterizes MNIST. It is also interesting to note that while the algorithm meta-trained using (Andrychowicz et al., 2016) (which we will refer to as L2LBGDBGD) performs well on CIFAR, it is unable to reach the optimum on TFD. + +![](images/2226d92233c78e2cc8398fe5882a7f24065a5392b547b417b88d986cc9309d20.jpg) +Figure 5: Comparison of the various hand-engineered and learned algorithms on training neural nets with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 for 800 iterations with mini-batches of size 64. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. + +Next, we change the architecture of the neural nets and see if Predicted Step Descent generalizes to the new architecture. We increase the number of input units to 100 and the number of hidden units to 200, so that the number of parameters is roughly increased by a factor of 8. As shown in Figure 2, Predicted Step Descent consistently outperforms other algorithms on each dataset, despite having not been trained to optimize neural nets of this architecture. Interestingly, while it exhibited a bit of oscillation initially on TFD and CIFAR-10, it quickly recovered and overtook other algorithms, which is reminiscent of the phenomenon reported in (Li & Malik, 2016) for low-dimensional optimization problems. This suggests that it has learned to detect when it is performing poorly and knows how to change tack accordingly. L2LBGDBGD experienced difficulties on TFD and CIFAR10 as well, but slowly diverged. + +We now investigate how robust Predicted Step Descent is to stochasticity of the gradients. To this end, we take a look at its performance when we reduce the mini-batch size from 64 to 10 on both the original architecture with 48 input and hidden units and the enlarged architecture with 100 input units and 200 hidden units. As shown in Figure 3, on the original architecture, Predicted Step Descent still outperforms all other algorithms and is able to handle the increased stochasticity fairly well. In contrast, conjugate gradient and L2LBGDBGD had some difficulty handling the increased stochasticity on TFD and to a lesser extent, on CIFAR-10. In the former case, both diverged; in the latter case, both were progressing slowly towards the optimum. + +On the enlarged architecture, Predicted Step Descent experienced some significant oscillations on TFD and CIFAR-10, but still managed to achieve a much better objective value than all the other algorithms. Many hand-engineered algorithms also experienced much greater oscillations than previously, suggesting that the optimization problems are inherently harder. L2LBGDBGD diverged fairly quickly on these two datasets. + +Finally, we try doubling the number of iterations. As shown in Figure 5, despite being trained over a time horizon of 400 iterations, Predicted Step Descent behaves reasonably beyond the number of iterations it is trained for. + +# 5 CONCLUSION + +In this paper, we presented a new method for learning optimization algorithms for high-dimensional stochastic problems. We applied the method to learning an optimization algorithm for training shallow neural nets. We showed that the algorithm learned using our method on the problem of training a neural net on MNIST generalizes to the problems of training neural nets on unrelated tasks/datasets like the Toronto Faces Dataset, CIFAR-10 and CIFAR-100. We also demonstrated that the learned optimization algorithm is robust to changes in the stochasticity of gradients and the neural net architecture. + +# REFERENCES + +Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, and Nando de Freitas. Learning to learn by gradient descent by gradient descent. arXiv preprint arXiv:1606.04474, 2016. + +Jonathan Baxter, Rich Caruana, Tom Mitchell, Lorien Y Pratt, Daniel L Silver, and Sebastian Thrun. NIPS 1995 workshop on learning to learn: Knowledge consolidation and transfer in inductive systems. https://web.archive.org/web/20000618135816/http://www. cs.cmu.edu/afs/cs.cmu.edu/user/caruana/pub/transfer.html, 1995. Accessed: 2015-12-05. + +Y Bengio, S Bengio, and J Cloutier. Learning a synaptic learning rule. In Neural Networks, 1991., IJCNN-91-Seattle International Joint Conference on, volume 2, pp. 969–vol. IEEE, 1991. + +Yoshua Bengio. Gradient-based optimization of hyperparameters. Neural computation, 12(8):1889– 1900, 2000. + +James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. The Journal of Machine Learning Research, 13(1):281–305, 2012. + +James S Bergstra, Remi Bardenet, Yoshua Bengio, and Bal ´ azs K ´ egl. Algorithms for hyper-parameter ´ optimization. In Advances in Neural Information Processing Systems, pp. 2546–2554, 2011. + +M Bray, E Koller-Meier, P Muller, L Van Gool, and NN Schraudolph. 3D hand tracking by rapid stochastic gradient descent using a skinning model. In Visual Media Production, 2004.(CVMP). 1st European Conference on, pp. 59–68. IET, 2004. + +Pavel Brazdil, Christophe Giraud Carrier, Carlos Soares, and Ricardo Vilalta. Metalearning: applications to data mining. Springer Science & Business Media, 2008. + +Pavel B Brazdil, Carlos Soares, and Joaquim Pinto Da Costa. Ranking learning algorithms: Using ibl and meta-learning on accuracy and time results. Machine Learning, 50(3):251–277, 2003. + +Christian Daniel, Jonathan Taylor, and Sebastian Nowozin. Learning step size controllers for robust neural network training. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. + +Justin Domke. Generic methods for optimization-based modeling. In AISTATS, volume 22, pp. 318–326, 2012. + +John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011. + +Matthias Feurer, Jost Tobias Springenberg, and Frank Hutter. Initializing bayesian hyperparameter optimization via meta-learning. In AAAI, pp. 1128–1135, 2015. + +Jie Fu, Zichuan Lin, Miao Liu, Nicholas Leonard, Jiashi Feng, and Tat-Seng Chua. Deep q-networks for accelerating the training of deep neural networks. arXiv preprint arXiv:1606.01467, 2016. + +Karol Gregor and Yann LeCun. Learning fast approximations of sparse coding. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pp. 399–406, 2010. + +Samantha Hansen. Using deep q-learning to control optimization hyperparameters. arXiv preprint arXiv:1602.04062, 2016. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. + +Sepp Hochreiter, A Steven Younger, and Peter R Conwell. Learning to learn using gradient descent. In International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001. + +Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization, pp. 507–523. Springer, 2011. + +Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. + +Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. arXiv preprint arXiv:1504.00702, 2015. + +Ke Li and Jitendra Malik. Learning to optimize. CoRR, abs/1606.01885, 2016. + +Dougal Maclaurin, David Duvenaud, and Ryan P Adams. Gradient-based hyperparameter optimization through reversible learning. arXiv preprint arXiv:1502.03492, 2015. + +Paul L Ruvolo, Ian Fasel, and Javier R Movellan. Optimization on a budget: A reinforcement learning approach. In Advances in Neural Information Processing Systems, pp. 1385–1392, 2009. + +Jurgen Schmidhuber. Optimal ordered problem solver. ¨ Machine Learning, 54(3):211–254, 2004. + +Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems, pp. 2951–2959, 2012. + +Pablo Sprechmann, Roee Litman, Tal Ben Yakar, Alexander M Bronstein, and Guillermo Sapiro. Supervised sparse analysis and synthesis operators. In Advances in Neural Information Processing Systems, pp. 908–916, 2013. + +Kevin Swersky, Jasper Snoek, and Ryan P Adams. Multi-task bayesian optimization. In Advances in neural information processing systems, pp. 2004–2012, 2013. + +Sebastian Thrun and Lorien Pratt. Learning to learn. Springer Science & Business Media, 2012. + +Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2), 2012. + +Ricardo Vilalta and Youssef Drissi. A perspective view and survey of meta-learning. Artificial Intelligence Review, 18(2):77–95, 2002. + +Huahua Wang and Arindam Banerjee. Bregman alternating direction method of multipliers. CoRR, abs/1306.3203, 2014. \ No newline at end of file diff --git a/parse/train/BkM27IxR-/BkM27IxR-_content_list.json b/parse/train/BkM27IxR-/BkM27IxR-_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4c9fa0d46d7cf93f97b278c1b967f1901bc10649 --- /dev/null +++ b/parse/train/BkM27IxR-/BkM27IxR-_content_list.json @@ -0,0 +1,1469 @@ +[ + { + "type": "text", + "text": "LEARNING TO OPTIMIZE NEURAL NETS ", + "text_level": 1, + "bbox": [ + 176, + 98, + 655, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 145, + 398, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 210, + 544, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning to Optimize (Li & Malik, 2016) is a recently proposed framework for learning optimization algorithms using reinforcement learning. In this paper, we explore learning an optimization algorithm for training shallow neural nets. Such high-dimensional stochastic optimization problems present interesting challenges for existing reinforcement learning algorithms. We develop an extension that is suited to learning optimization algorithms in this setting and demonstrate that the learned optimization algorithm consistently outperforms other known optimization algorithms even on unseen tasks and is robust to changes in stochasticity of gradients and the neural net architecture. More specifically, we show that an optimization algorithm trained with the proposed method on the problem of training a neural net on MNIST generalizes to the problems of training neural nets on the Toronto Faces Dataset, CIFAR-10 and CIFAR-100. ", + "bbox": [ + 233, + 243, + 764, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 443, + 336, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning is centred on the philosophy that learning patterns automatically from data is generally better than meticulously crafting rules by hand. This data-driven approach has delivered: today, machine learning techniques can be found in a wide range of application areas, both in AI and beyond. Yet, there is one domain that has conspicuously been left untouched by machine learning: the design of tools that power machine learning itself. ", + "bbox": [ + 174, + 476, + 823, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the most widely used tools in machine learning is optimization algorithms. We have grown accustomed to seeing an optimization algorithm as a black box that takes in a model that we design and the data that we collect and outputs the optimal model parameters. The optimization algorithm itself largely stays static: its design is reserved for human experts, who must toil through many rounds of theoretical analysis and empirical validation to devise a better optimization algorithm. Given this state of affairs, perhaps it is time for us to start practicing what we preach and learn how to learn. ", + "bbox": [ + 174, + 553, + 825, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Li & Malik (2016) and Andrychowicz et al. (2016) introduced two different frameworks for learning optimization algorithms. Whereas Andrychowicz et al. (2016) focuses on learning an optimization algorithm for training models on a particular task, Li & Malik (2016) sets a more ambitious objective of learning an optimization algorithm for training models that is task-independent. We study the latter paradigm in this paper and develop a method for learning an optimization algorithm for high-dimensional stochastic optimization problems, like the problem of training shallow neural nets. ", + "bbox": [ + 174, + 657, + 825, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Under the “Learning to Optimize” framework proposed by Li & Malik (2016), the problem of learning an optimization algorithm is formulated as a reinforcement learning problem. We consider the general structure of an unconstrained continuous optimization algorithm, as shown in Algorithm 1. In each iteration, the algorithm takes a step $\\Delta x$ and uses it to update the current iterate $\\boldsymbol { x } ^ { ( i ) }$ . In hand-engineered optimization algorithms, $\\Delta x$ is computed using some fixed formula $\\phi$ that depends on the objective function, the current iterate and past iterates. Often, it is simply a function of the current and past gradients. ", + "bbox": [ + 174, + 762, + 825, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Different choices of $\\phi$ yield different optimization algorithms and so each optimization algorithm is essentially characterized by its update formula $\\phi$ . Hence, by learning $\\phi$ , we can learn an optimization algorithm. Li & Malik (2016) observed that an optimization algorithm can be viewed as a Markov decision process (MDP), where the state includes the current iterate, the action is the step vector $\\Delta x$ ", + "bbox": [ + 176, + 867, + 823, + 922 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Require: Objective function $f$ $\\bar { x ^ { ( 0 ) } } \\gets$ random point in the domain of $f$ for $i = 1 , 2 , \\dots { \\bf d }$ o $\\Delta x \\gets \\acute { \\phi } ( f , \\{ x ^ { ( 0 ) } , \\ldots , x ^ { ( i - 1 ) } \\} )$ if stopping condition is met then return x(i−1) end if $x ^ { ( i ) } \\gets x ^ { ( i - 1 ) } + \\Delta x$ end for ", + "bbox": [ + 174, + 122, + 457, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and the policy is the update formula $\\phi$ . Hence, the problem of learning $\\phi$ simply reduces to a policy search problem. ", + "bbox": [ + 173, + 280, + 823, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we build on the method proposed in (Li & Malik, 2016) and develop an extension that is suited to learning optimization algorithms for high-dimensional stochastic problems. We use it to learn an optimization algorithm for training shallow neural nets and show that it outperforms popular hand-engineered optimization algorithms like ADAM (Kingma & Ba, 2014), AdaGrad (Duchi et al., 2011) and RMSprop (Tieleman & Hinton, 2012) and an optimization algorithm learned using the supervised learning method proposed in (Andrychowicz et al., 2016). Furthermore, we demonstrate that our optimization algorithm learned from the experience of training on MNIST generalizes to training on other datasets that have very dissimilar statistics, like the Toronto Faces Dataset, CIFAR10 and CIFAR-100. ", + "bbox": [ + 173, + 315, + 825, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 462, + 344, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The line of work on learning optimization algorithms is fairly recent. Li & Malik (2016) and Andrychowicz et al. (2016) were the first to propose learning general optimization algorithms. Li & Malik (2016) explored learning task-independent optimization algorithms and used reinforcement learning to learn the optimization algorithm, while Andrychowicz et al. (2016) investigated learning task-dependent optimization algorithms and used supervised learning. ", + "bbox": [ + 174, + 494, + 825, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the special case where objective functions that the optimization algorithm is trained on are loss functions for training other models, these methods can be used for “learning to learn” or “metalearning”. While these terms have appeared from time to time in the literature (Baxter et al., 1995; Vilalta & Drissi, 2002; Brazdil et al., 2008; Thrun & Pratt, 2012), they have been used by different authors to refer to disparate methods with different purposes. These methods all share the objective of learning some form of meta-knowledge about learning, but differ in the type of meta-knowledge they aim to learn. We can divide the various methods into the following three categories. ", + "bbox": [ + 174, + 570, + 825, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 LEARNING WHAT TO LEARN ", + "text_level": 1, + "bbox": [ + 176, + 685, + 413, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods in this category Thrun & Pratt (2012) aim to learn what parameter values of the base-level learner are useful across a family of related tasks. The meta-knowledge captures commonalities shared by tasks in the family, which enables learning on a new task from the family to be performed more quickly. Most early methods fall into this category; this line of work has blossomed into an area that has later become known as transfer learning and multi-task learning. ", + "bbox": [ + 174, + 712, + 825, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 LEARNING WHICH MODEL TO LEARN ", + "text_level": 1, + "bbox": [ + 176, + 799, + 478, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods in this category Brazdil et al. (2008) aim to learn which base-level learner achieves the best performance on a task. The meta-knowledge captures correlations between different tasks and the performance of different base-level learners on those tasks. One challenge under this setting is to decide on a parameterization of the space of base-level learners that is both rich enough to be capable of representing disparate base-level learners and compact enough to permit tractable search over this space. Brazdil et al. (2003) proposes a nonparametric representation and stores examples of different base-level learners in a database, whereas Schmidhuber (2004) proposes representing baselevel learners as general-purpose programs. The former has limited representation power, while the latter makes search and learning in the space of base-level learners intractable. Hochreiter et al. (2001) views the (online) training procedure of any base-learner as a black box function that maps a sequence of training examples to a sequence of predictions and models it as a recurrent neural net. Under this formulation, meta-training reduces to training the recurrent net, and the base-level learner is encoded in the memory state of the recurrent net. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperparameter optimization can be seen as another example of methods in this category. The space of base-level learners to search over is parameterized by a predefined set of hyperparameters. Unlike the methods above, multiple trials with different hyperparameter settings on the same task are permitted, and so generalization across tasks is not required. The discovered hyperparameters are generally specific to the task at hand and hyperparameter optimization must be rerun for new tasks. Various kinds of methods have been proposed, such those based on Bayesian optimization (Hutter et al., 2011; Bergstra et al., 2011; Snoek et al., 2012; Swersky et al., 2013; Feurer et al., 2015), random search (Bergstra & Bengio, 2012) and gradient-based optimization (Bengio, 2000; Domke, 2012; Maclaurin et al., 2015). ", + "bbox": [ + 174, + 194, + 825, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 LEARNING HOW TO LEARN ", + "text_level": 1, + "bbox": [ + 176, + 337, + 405, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Methods in this category aim to learn a good algorithm for training a base-level learner. Unlike methods in the previous categories, the goal is not to learn about the outcome of learning, but rather the process of learning. The meta-knowledge captures commonalities in the behaviours of learning algorithms that achieve good performance. The base-level learner and the task are given by the user, so the learned algorithm must generalize across base-level learners and tasks. Since learning in most cases is equivalent to optimizing some objective function, learning a learning algorithm often reduces to learning an optimization algorithm. This problem was explored in (Li & Malik, 2016) and (Andrychowicz et al., 2016). Closely related is (Bengio et al., 1991), which learns a Hebblike synaptic learning rule that does not depend on the objective function, which does not allow for generalization to different objective functions. ", + "bbox": [ + 173, + 363, + 825, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Various work has explored learning how to adjust the hyperparameters of hand-engineered optimization algorithms, like the step size (Hansen, 2016; Daniel et al., 2016; Fu et al., 2016) or the damping factor in the Levenberg-Marquardt algorithm (Ruvolo et al., 2009). Related to this line of work is stochastic meta-descent (Bray et al., 2004), which derives a rule for adjusting the step size analytically. A different line of work (Gregor & LeCun, 2010; Sprechmann et al., 2013) parameterizes intermediate operands of special-purpose solvers for a class of optimization problems that arise in sparse coding and learns them using supervised learning. ", + "bbox": [ + 174, + 508, + 825, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 LEARNING TO OPTIMIZE ", + "text_level": 1, + "bbox": [ + 176, + 627, + 411, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 SETTING ", + "text_level": 1, + "bbox": [ + 174, + 659, + 276, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the “Learning to Optimize” framework, we are given a set of training objective functions $f _ { 1 } , \\ldots , f _ { n }$ drawn from some distribution $\\mathcal { F }$ . An optimization algorithm $\\mathcal { P }$ takes an objective function $f$ and an initial iterate $x ^ { ( 0 ) }$ as input and produces a sequence of iterates $x ^ { ( 1 ) } , \\ldots , x ^ { ( T ) }$ , where $x ^ { ( T ) }$ is the solution found by the optimizer. We are also given a distribution $\\mathcal { D }$ that generates the initial iterate $x ^ { ( 0 ) }$ and a meta-loss $\\mathcal { L }$ , which takes an objective function $f$ and a sequence of iterates $x ^ { ( 1 ) } , \\ldots , x ^ { ( T ) }$ produced by an optimization algorithm as input and outputs a scalar that measures the quality of the iterates. The goal is to learn an optimization algorithm ${ \\mathcal { P } } ^ { * }$ such that $\\mathbb { E } _ { f \\sim \\mathcal { F } , x ^ { ( 0 ) } \\sim \\mathcal { D } } \\left[ \\mathcal { L } ( f , \\mathcal { P } ^ { * } ( f , x ^ { ( 0 ) } ) ) \\right]$ is minimized. The meta-loss is chosen to penalize optimization algorithms that exhibit behaviours we find undesirable, like slow convergence or excessive oscillations. Assuming we would like to learn an algoritha good choice of meta-loss would then simply be regret and can be interpreted as the area under the $\\textstyle \\sum _ { i = 1 } ^ { T } f ( x ^ { ( i ) } )$ zes the objective function it is given,, which is equivalent to cumulativective values over time. ", + "bbox": [ + 174, + 684, + 825, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective functions $f _ { 1 } , \\ldots , f _ { n }$ may correspond to loss functions for training base-level learners, in which case the algorithm that learns the optimization algorithm can be viewed as a meta-learner. In this setting, each objective function is the loss function for training a particular base-learner on a particular task, and so the set of training objective functions can be loss functions for training a base-learner or a family of base-learners on different tasks. At test time, the learned optimization algorithm is evaluated on unseen objective functions, which correspond to loss functions for training base-learners on new tasks, which may be completely unrelated to tasks used for training the optimization algorithm. Therefore, the learned optimization algorithm must not learn anything about the tasks used for training. Instead, the goal is to learn an optimization algorithm that can exploit the geometric structure of the error surface induced by the base-learners. For example, if the base-level model is a neural net with ReLU activation units, the optimization algorithm should hopefully learn to leverage the piecewise linearity of the model. Hence, there is a clear division of responsibilities between the meta-learner and base-learners. The knowledge learned at the meta-level should be pertinent for all tasks, whereas the knowledge learned at the base-level should be task-specific. The meta-learner should therefore generalize across tasks, whereas the base-learner should generalize across instances. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 825, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 RL PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 176, + 286, + 352, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of reinforcement learning is to learn to interact with an environment in a way that minimizes cumulative costs that are expected to be incurred over time. The environment is formalized as a partially observable Markov decision process $( \\mathrm { P O M D P } ) ^ { 1 }$ , which is defined by the tuple $( S , \\mathcal { O } , \\mathcal { A } , p _ { i } , p , p _ { o } , c , T )$ , where $\\mathcal { S } \\subseteq \\mathbb { R } ^ { D }$ is the set of states, $\\mathcal { O } \\subseteq \\mathbb { R } ^ { D ^ { \\prime } }$ is the set of observations, $\\mathcal { A } \\subseteq \\mathbb { R } ^ { d }$ is the set of actions, $p _ { i } \\left( s _ { 0 } \\right)$ is the probability density over initial states $s _ { 0 }$ , $p \\left( { { s } _ { t + 1 } } \\left| { { s } _ { t } } , { { a } _ { t } } \\right. \\right)$ is the probability density over the subsequent state $s _ { t + 1 }$ given the current state $s _ { t }$ and action $a _ { t }$ , $p _ { o } \\left( o _ { t } \\bar { | } _ { s _ { t } } \\right)$ is the probability density over the current observation $o _ { t }$ given the current state $s _ { t }$ , $c : { \\mathcal { S } } \\mathbb { R }$ is a function that assigns a cost to each state and $T$ is the time horizon. Often, the probability densities $p$ and $p _ { o }$ are unknown and not given to the learning algorithm. ", + "bbox": [ + 173, + 313, + 825, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A policy $\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { o } _ { t } , t \\right)$ is a conditional probability density over actions $a _ { t }$ given the current observation $o _ { t }$ and time step $t$ . When a policy is independent of $t$ , it is known as a stationary policy. The goal of the reinforcement learning algorithm is to learn a policy $\\pi ^ { * }$ that minimizes the total expected cost over time. More precisely, ", + "bbox": [ + 174, + 446, + 825, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/b919d595be0d1a432a00d2fbbfe2eab638d4374ebd57fafd68cfd8d40797150b.jpg", + "text": "$$\n\\pi ^ { * } = \\arg \\operatorname* { m i n } _ { \\pi } \\mathbb { E } _ { s _ { 0 } , a _ { 0 } , s _ { 1 } , \\dots , s _ { T } } \\left[ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 359, + 510, + 637, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the expectation is taken with respect to the joint distribution over the sequence of states and actions, often referred to as a trajectory, which has the density ", + "bbox": [ + 173, + 558, + 821, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/51df13704307a3794a4cd1301cf4a3226c41c612fa6a34d8eca973a93a49d9ae.jpg", + "text": "$$\n\\begin{array} { c } { { \\displaystyle q \\big ( s _ { 0 } , a _ { 0 } , s _ { 1 } , \\ldots , s _ { T } \\big ) = \\int _ { o _ { 0 } , \\ldots , o _ { T } } p _ { i } \\left( s _ { 0 } \\right) p _ { o } \\left( o _ { 0 } \\big | s _ { 0 } \\right) } } \\\\ { { { \\displaystyle \\prod _ { t = 0 } ^ { T - 1 } \\pi \\left( \\left. a _ { t } \\right| o _ { t } , t \\right) p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } \\right) p _ { o } \\left( \\left. o _ { t + 1 } \\right| s _ { t + 1 } \\right) . } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 308, + 592, + 689, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make learning tractable, $\\pi$ is often constrained to lie in a parameterized family. A common assumption is that $\\pi \\left( \\left. a _ { t } \\right| o _ { t } , t \\right) = \\mathcal { N } \\left( \\mu ^ { \\pi } ( o _ { t } ) , \\Sigma ^ { \\pi } ( o _ { t } ) \\right)$ , where $\\textstyle { \\mathcal { N } } ( { \\boldsymbol { \\mu } } , { \\boldsymbol { \\Sigma } } )$ denotes the density of a Gaussian with mean $\\mu$ and covariance $\\Sigma$ . The functions $\\mu ^ { \\pi } ( \\cdot )$ and possibly $\\Sigma ^ { \\pi } ( \\cdot )$ are modelled using function approximators, whose parameters are learned. ", + "bbox": [ + 176, + 684, + 821, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 FORMULATION ", + "text_level": 1, + "bbox": [ + 174, + 756, + 318, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our setting, the state $s _ { t }$ consists of the current iterate $x ^ { ( t ) }$ and features $\\Phi ( \\cdot )$ that depend on the history of iterates $x ^ { ( 1 ) } , \\ldots , x ^ { ( t ) }$ , (noisy) gradients $\\nabla \\hat { f } ( x ^ { ( 1 ) } ) , \\ldots , \\nabla \\hat { f } ( x ^ { ( t ) } )$ and (noisy) objective values ${ \\hat { f } } ( x ^ { ( 1 ) } ) , \\ldots , { \\hat { f } } ( x ^ { ( t ) } )$ . The action $a _ { t }$ is the step $\\Delta x$ that will be used to update the iterate. The observation $o _ { t }$ excludes $x ^ { ( t ) }$ and consists of features $\\Psi ( \\cdot )$ that depend on the iterates, gradient and objective values from recent iterations, and the previous memory state of the learned optimization algorithm, which takes the form of a recurrent neural net. This memory state can be viewed as a statistic of the previous observations that is learned jointly with the policy. ", + "bbox": [ + 173, + 781, + 825, + 887 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under this formulation, the initial probability density $p _ { i }$ captures how the initial iterate, gradient and objective value tend to be distributed. The transition probability density $p$ captures the how the gradient and objective value are likely to change given the step that is taken currently; in other words, it encodes the local geometry of the training objective functions. Assuming the goal is to learn an optimization algorithm that minimizes the objective function, the cost $c$ of a state $\\boldsymbol { s } _ { t } = \\left( \\boldsymbol { x } ^ { ( t ) } , \\Phi \\left( \\cdot \\right) \\right) ^ { T }$ is simply the true objective value $f ( \\boldsymbol { x } ^ { ( t ) } )$ . ", + "bbox": [ + 173, + 103, + 825, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Any particular policy $\\pi \\left( \\boldsymbol { a } _ { t } \\left| \\boldsymbol { o } _ { t } , t \\right. \\right)$ , which generates $a _ { t } ~ = ~ \\Delta x$ at every time step, corresponds to a particular (noisy) update formula $\\phi$ , and therefore a particular (noisy) optimization algorithm. Therefore, learning an optimization algorithm simply reduces to searching for the optimal policy. ", + "bbox": [ + 174, + 200, + 821, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mean of the policy is modelled as a recurrent neural net fragment that corresponds to a single time step, which takes the observation features $\\Psi ( \\cdot )$ and the previous memory state as input and outputs the step to take. ", + "bbox": [ + 174, + 250, + 825, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 GUIDED POLICY SEARCH ", + "text_level": 1, + "bbox": [ + 174, + 310, + 392, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reinforcement learning method we use is guided policy search (GPS) (Levine et al., 2015), which is a policy search method designed for searching over large classes of expressive non-linear policies in continuous state and action spaces. It maintains two policies, $\\psi$ and $\\pi$ , where the former lies in a time-varying linear policy class in which the optimal policy can found in closed form, and the latter lies in a stationary non-linear policy class in which policy optimization is challenging. In each iteration, it performs policy optimization on $\\psi$ , and uses the resulting policy as supervision to train $\\pi$ . ", + "bbox": [ + 173, + 337, + 826, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More precisely, GPS solves the following constrained optimization problem: ", + "bbox": [ + 173, + 439, + 676, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/f5be126419b4a29a54757648e664beaddb30c079479aa1e9291425a258d92b7d.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta , \\eta } \\mathbb { E } _ { \\psi } \\left[ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) \\right] \\mathrm { ~ s . t . ~ } \\psi \\left( \\left. a _ { t } \\right| s _ { t } , t ; \\eta \\right) = \\pi \\left( \\left. a _ { t } \\right| s _ { t } ; \\theta \\right) \\forall a _ { t } , s _ { t } , t\n$$", + "text_format": "latex", + "bbox": [ + 302, + 462, + 696, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta$ and $\\theta$ denote the parameters of $\\psi$ and $\\pi$ respectively, $\\mathbb { E } _ { \\rho } \\left[ \\cdot \\right]$ denotes the expectation taken with respect to the trajectory induced by a policy $\\rho$ and $\\begin{array} { r } { \\pi \\left( \\left. a _ { t } \\right| s _ { t } ; \\theta \\right) : = \\int _ { o _ { t } } \\pi \\left( \\left. a _ { t } \\right| o _ { t } ; \\theta \\right) p _ { o } \\left( \\left. o _ { t } \\right| s _ { t } \\right) ^ { 2 } } \\end{array}$ . ", + "bbox": [ + 173, + 510, + 823, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since there are an infinite number of equality constraints, the problem is relaxed by enforcing equality on the mean actions taken by $\\psi$ and $\\pi$ at every time step3. So, the problem becomes: ", + "bbox": [ + 173, + 547, + 823, + 578 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/407d95c56a995ddd518677a9f9c837a0fe2a62d6a02d1ebce4c22c60b61b0cfe.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\theta , \\eta } \\mathbb { E } _ { \\psi } [ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) ] \\mathrm { ~ s . t . ~ } \\mathbb { E } _ { \\psi } [ a _ { t } ] = \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] \\forall t\n$$", + "text_format": "latex", + "bbox": [ + 333, + 585, + 665, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This problem is solved using Bregman ADMM (Wang & Banerjee, 2014), which performs the following updates in each iteration: ", + "bbox": [ + 173, + 638, + 823, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/98b4e8aee640645dca59af14914cf556ee40ccd35b91af03094a1e5f58cc3027.jpg", + "text": "$$\n\\begin{array} { r l } & { \\eta \\gets \\arg \\underset { \\eta } { \\operatorname* { m i n } } \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\psi } [ c ( s _ { t } ) - \\lambda _ { t } ^ { T } a _ { t } ] + \\nu _ { t } D _ { t } ( \\eta , \\theta ) } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad } \\\\ & { \\theta \\gets \\arg \\underset { \\theta } { \\operatorname* { m i n } } \\sum _ { t = 0 } ^ { T } \\lambda _ { t } ^ { T } \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] + \\nu _ { t } D _ { t } ( \\theta , \\eta ) } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\lambda _ { t } \\gets \\lambda _ { t } + \\alpha \\nu _ { t } ( \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] - \\mathbb { E } _ { \\psi } [ a _ { t } ] ) \\ \\forall t , } \\\\ & { \\mathrm { ~ } \\mathrm { ~ } : \\mathrm { r e } \\quad D _ { t } ( \\theta , \\eta ) \\quad : = \\quad \\mathbb { E } _ { \\psi } [ D _ { K L } ( \\pi ( a _ { t } | s _ { t } ; \\theta ) ) ] \\ \\psi ( a _ { t } | s _ { t } , t ; \\eta ) ) ] \\quad \\mathrm { ~ a n d ~ } \\quad D _ { t } ( \\eta , \\theta ) } \\\\ & { [ D _ { K L } ( \\psi ( a _ { t } | s _ { t } , t ; \\eta ) ] ) | \\ \\pi ( a _ { t } | s _ { t } ; \\theta ) ) | . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 169, + 672, + 825, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The algorithm assumes that $\\psi \\left( \\left. a _ { t } \\right| s _ { t } , t ; \\eta \\right) = \\mathcal { N } \\left( K _ { t } s _ { t } + k _ { t } , G _ { t } \\right)$ , where $\\boldsymbol { \\eta } : = \\left( K _ { t } , k _ { t } , G _ { t } \\right) _ { t = 1 } ^ { T }$ and $\\pi \\left( a _ { t } | o _ { t } ; \\theta \\right) = \\mathcal { N } \\left( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) , \\Sigma ^ { \\pi } \\right)$ , where $\\boldsymbol { \\theta } : = ( \\omega , \\Sigma ^ { \\pi } )$ and $\\mu _ { \\omega } ^ { \\pi } ( \\cdot )$ can be an arbitrary function that is typically modelled using a nonlinear function approximator like a neural net. ", + "bbox": [ + 173, + 815, + 825, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the start of each iteration, the algorithm constructs a model of the transition probability density $\\tilde { p } \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t ; \\zeta \\right) = \\mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )$ , where $\\boldsymbol { \\zeta } : = \\left( A _ { t } , B _ { t } , c _ { t } , F _ { t } \\right) _ { t = 1 } ^ { T }$ is fitted to samples of $s _ { t }$ drawn from the trajectory induced by $\\psi$ , which essentially amounts to a local linearization of the true transition probability $p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t \\right)$ . We will use $\\mathbb { E } _ { \\widetilde { \\psi } } \\left[ \\cdot \\right]$ to denote expectation taken with respect to the trajectory induced by $\\psi$ under the modelled transition probability $\\tilde { p }$ . Additionally, the algorithm fits local quadratic approximations to $c ( s _ { t } )$ around samples of $s _ { t }$ drawn from the trajectory induced by $\\psi$ so that $\\begin{array} { r } { c ( s _ { t } ) \\approx \\dot { \\tilde { c } } \\dot { ( s _ { t } ) } : = \\frac { 1 } { 2 } s _ { t } ^ { T } C _ { t } s _ { t } + \\dot { d _ { t } ^ { T } } s _ { t } + h _ { t } } \\end{array}$ for $s _ { t }$ ’s that are near the samples. ", + "bbox": [ + 173, + 102, + 825, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With these assumptions, the subproblem that needs to be solved to update $\\eta ~ = ~ ( K _ { t } , k _ { t } , G _ { t } ) _ { t = 1 } ^ { T }$ becomes: ", + "bbox": [ + 173, + 215, + 820, + 244 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/da17dbd224435326b9983c40ce40ffd1486ed56e1c09a13c8b435c27ba5a8491.jpg", + "text": "$$\n\\begin{array} { r l r } & { \\underset { \\eta } { \\operatorname* { m i n } } \\displaystyle \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\boldsymbol { \\tilde { \\psi } } } [ \\boldsymbol { \\tilde { c } } ( s _ { t } ) - \\lambda _ { t } ^ { T } \\boldsymbol { a } _ { t } ] + \\nu _ { t } D _ { t } ( \\eta , \\theta ) } & \\\\ & { \\mathrm { s . t . } \\displaystyle \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\boldsymbol { \\tilde { \\psi } } } [ D _ { K L } ( \\boldsymbol { \\psi } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\eta ) | \\Big | \\psi ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\eta ^ { \\prime } ) ) ] \\leq \\epsilon , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 321, + 248, + 676, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\eta ^ { \\prime }$ denotes the old $\\eta$ from the previous iteration. Because $\\tilde { p }$ and $\\tilde { c }$ are only valid locally around the trajectory induced by $\\psi$ , the constraint is added to limit the amount by which $\\eta$ is updated. It turns out that the unconstrained problem can be solved in closed form using a dynamic programming algorithm known as linear-quadratic-Gaussian (LQG) regulator in time linear in the time horizon $T$ and cubic in the dimensionality of the state space $D$ . The constrained problem is solved using dual gradient descent, which uses LQG as a subroutine to solve for the primal variables in each iteration and increments the dual variable on the constraint until it is satisfied. ", + "bbox": [ + 173, + 334, + 826, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Updating $\\theta$ is straightforward, since expectations taken with respect to the trajectory induced by $\\pi$ are always conditioned on $s _ { t }$ and all outer expectations over $s _ { t }$ are taken with respect to the trajectory induced by $\\psi$ . Therefore, $\\pi$ is essentially decoupled from the transition probability $p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t \\right)$ and so its parameters can be updated without affecting the distribution of $s _ { t }$ ’s. The subproblem that needs to be solved to update $\\theta$ therefore amounts to a standard supervised learning problem. ", + "bbox": [ + 174, + 439, + 825, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since $\\psi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\boldsymbol { \\eta } \\right)$ and $\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ; \\boldsymbol { \\theta } \\right)$ are Gaussian, $D _ { t } \\left( { \\theta , \\eta } \\right)$ can be computed analytically. More concretely, if we assume $\\Sigma ^ { \\pi }$ to be fixed for simplicity, the subproblem that is solved for updating $\\theta = \\left( \\omega , \\Sigma ^ { \\pi } \\right)$ is: ", + "bbox": [ + 174, + 515, + 825, + 559 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/ec360d7a6f7432925bf8f970733b046a6e04fac406cb257daeefeab826336ea9.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\operatorname* { m i n } _ { \\theta } \\mathbb { E } _ { \\psi } [ \\sum _ { t = 0 } ^ { T } \\lambda _ { t } ^ { T } \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) + \\frac { \\nu _ { t } } { 2 } ( \\operatorname { t r } ( G _ { t } ^ { - 1 } \\Sigma ^ { \\pi } ) - \\log | \\Sigma ^ { \\pi } | ) } \\\\ { \\displaystyle + \\frac { \\nu _ { t } } { 2 } ( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) - \\mathbb { E } _ { \\psi } [ a _ { t } | s _ { t } , t ] ) ^ { T } G _ { t } ^ { - 1 } ( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) - \\mathbb { E } _ { \\psi } [ a _ { t } | s _ { t } , t ] ) ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 259, + 583, + 736, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the last term is the squared Mahalanobis distance between the mean actions of $\\psi$ and $\\pi$ at time step $t$ , which is intuitive as we would like to encourage $\\pi$ to match $\\psi$ . ", + "bbox": [ + 173, + 676, + 823, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5 CONVOLUTIONAL GPS ", + "text_level": 1, + "bbox": [ + 174, + 723, + 375, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The problem of learning high-dimensional optimization algorithms presents challenges for reinforcement learning algorithms due to high dimensionality of the state and action spaces. For example, in the case of GPS, because the running time of LQG is cubic in dimensionality of the state space, performing policy search even in the simple class of linear-Gaussian policies would be prohibitively expensive when the dimensionality of the optimization problem is high. ", + "bbox": [ + 173, + 750, + 825, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fortunately, many high-dimensional optimization problems have underlying structure that can be exploited. For example, the parameters of neural nets are equivalent up to permutation among certain coordinates. More concretely, for fully connected neural nets, the dimensions of a hidden layer and the corresponding weights can be permuted arbitrarily without changing the function they compute. Because permuting the dimensions of two adjacent layers can permute the weight matrix arbitrarily, an optimization algorithm should be invariant to permutations of the rows and columns of a weight matrix. A reasonable prior to impose is that the algorithm should behave in the same manner on all coordinates that correspond to entries in the same matrix. That is, if the values of two coordinates in all current and past gradients and iterates are identical, then the step vector produced by the algorithm should have identical values in these two coordinates. We will refer to the set of coordinates on which permutation invariance is enforced as a coordinate group. For the purposes of learning an optimization algorithm for neural nets, a natural choice would be to make each coordinate group correspond to a weight matrix or a bias vector. Hence, the total number of coordinate groups is twice the number of layers, which is usually fairly small. ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c32356121b006ac72da2adddd464187f503567b08623692799a8cd7bffe7f36e.jpg", + "image_caption": [ + "Figure 1: Comparison of the various hand-engineered and learned algorithms on training neural nets with 48 input and hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-batches of size 64. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. " + ], + "image_footnote": [], + "bbox": [ + 210, + 122, + 782, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 334, + 825, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the case of GPS, we impose this prior on both $\\psi$ and $\\pi$ . For the purposes of updating $\\eta$ , we first impose a block-diagonal structure on the parameters $A _ { t } , B _ { t }$ and $F _ { t }$ of the fitted transition probability density $\\tilde { p } \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t ; \\zeta \\right) = \\mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )$ , so that for each coordinate in the optimization problem, the dimensions of $s _ { t + 1 }$ that correspond to the coordinate only depend on the dimensions of $s _ { t }$ and $a _ { t }$ that correspond to the same coordinate. As a result, $\\tilde { p } \\big ( s _ { t + 1 } \\big | s _ { t } , a _ { t } , t ; \\zeta \\big )$ decomposes into multiple independent probability densities $\\tilde { p } ^ { j } \\left( s _ { t + 1 } ^ { j } \\Big | s _ { t } ^ { j } , a _ { t } ^ { j } , t ; \\zeta ^ { j } \\right)$ , one for each coordinate $j$ . Similarly, we also impose a block-diagonal structure on $C _ { t }$ for fitting $\\tilde { c } ( s _ { t } )$ and on the parameter matrix of the fitted model for $\\pi \\left( a _ { t } | \\boldsymbol s _ { t } \\bar { ; \\boldsymbol \\theta } \\right)$ . Under these assumptions, $K _ { t }$ and $G _ { t }$ are guaranteed to be block-diagonal as well. Hence, the Bregman divergence penalty term, $D \\left( \\eta , \\theta \\right)$ decomposes into a sum of Bregman divergence terms, one for each coordinate. ", + "bbox": [ + 173, + 439, + 825, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then further constrain dual variables $\\lambda _ { t }$ , sub-vectors of parameter vectors and sub-matrices of parameter matrices corresponding to each coordinate group to be identical across the group. Additionally, we replace the weight $\\nu _ { t }$ on $D \\left( \\eta , \\theta \\right)$ with an individual weight on each Bregman divergence term for each coordinate group. The problem then decomposes into multiple independent subproblems, one for each coordinate group. Because the dimensionality of the state subspace corresponding to each coordinate is constant, LQG can be executed on each subproblem much more efficiently. ", + "bbox": [ + 173, + 595, + 825, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similarly, for $\\pi$ , we choose a $\\mu _ { \\omega } ^ { \\pi } ( \\cdot )$ that shares parameters across different coordinates in the same group. We also impose a block-diagonal structure on $\\Sigma ^ { \\pi }$ and constrain the appropriate sub-matrices to share their entries. ", + "bbox": [ + 174, + 685, + 825, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.6 FEATURES ", + "text_level": 1, + "bbox": [ + 174, + 748, + 287, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We describe the features $\\Phi ( \\cdot )$ and $\\Psi ( \\cdot )$ at time step $t$ , which define the state $s _ { t }$ and observation $o _ { t }$ respectively. ", + "bbox": [ + 174, + 775, + 825, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Because of the stochasticity of gradients and objective values, the state features $\\Phi ( \\cdot )$ are defined in terms of summary statistics of the history of iterates $\\left\\{ x ^ { ( i ) } \\right\\} _ { i = 0 } ^ { t }$ , gradients $\\left\\{ \\nabla \\hat { f } ( x ^ { ( i ) } ) \\right\\} _ { i = 0 } ^ { t }$ and objective values $\\left\\{ \\hat { f } ( \\boldsymbol x ^ { ( i ) } ) \\right\\} _ { i = 0 } ^ { t }$ We define the following statistics, which we will refer to as the average recent iterate, gradient and objective value respectively: ", + "bbox": [ + 173, + 810, + 826, + 893 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/127a52a319bf4db3aba86fdbf80c5271cf7ab40491a1981f2627a709932c8dc8.jpg", + "text": "$$\n\\begin{array} { r } { \\overline { { \\boldsymbol { x } ^ { ( i ) } } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \\boldsymbol { x } ^ { ( j ) } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 228, + 907, + 465, + 928 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fea8480e1dc0f7dbc3c3a6dc8993cac0f6ffab3eae77436f1e3e961bbffec59.jpg", + "image_caption": [ + "Figure 2: Comparison of the various hand-engineered and learned algorithms on training neural nets with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with minibatches of size 64. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. " + ], + "image_footnote": [], + "bbox": [ + 209, + 121, + 782, + 237 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/7a6761589b90ee0c7200cf58fa051acd5f2b63ec2d32a67a2f1eb5d8569dab37.jpg", + "text": "$$\n\\begin{array} { r l } & { \\bullet \\ \\overline { { \\nabla \\hat { f } ( x ^ { ( i ) } ) } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } { \\nabla \\hat { f } ( x ^ { ( j ) } ) } } \\\\ & { \\bullet \\ \\overline { { \\hat { f } ( x ^ { ( i ) } ) } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \\hat { f } ( x ^ { ( j ) } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 212, + 328, + 532, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The state features $\\Phi ( \\cdot )$ consist of the relative change in the average recent objective value, the average recent gradient normalized by the magnitude of the a previous average recent gradient and a previous change in average recent iterate relative to the current change in average recent iterate: ", + "bbox": [ + 174, + 387, + 823, + 430 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/1c2af21e865a7d14970ed22ee4629d67c1f8d6b7a64a8e247bf2ff52147bdb2e.jpg", + "text": "$$\n\\begin{array} { r l } & { \\bullet \\{ ( \\hat { f } ( x ^ { ( t - 5 i ) } ) - \\widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) ) / \\widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) \\} _ { i = 0 } ^ { 2 4 } } \\\\ & { \\bullet \\{ \\overline { { \\nabla \\hat { f } ( x ^ { ( t - 5 i ) } ) } } / ( | \\overline { { \\nabla \\hat { f } ( x ^ { ( m a x ( t - 5 ( i + 1 ) , t \\mathrm { m o d } 5 ) ) } ) } } | + 1 ) \\} _ { i = 0 } ^ { 2 5 } } \\\\ & \\bullet \\{ \\frac { | \\frac { \\overline { { x ^ { ( \\mathrm { m a x } ( t - 5 ( i + 1 ) , t \\mathrm { m o d } 5 + 5 ) ) } } } - \\overline { { x ^ { ( \\mathrm { m a x } ( t - 5 ( i + 2 ) , t \\mathrm { m o d } 5 ) ) } } } } { | \\overline { { x ^ { ( t - 5 i ) } } } - \\overline { { x ^ { ( t - 5 ( i + 1 ) ) } } } | + 0 . 1 } \\} _ { i = 0 } ^ { 2 4 } \\} _ { i = 0 } ^ { 2 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 214, + 440, + 584, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that all operations are applied element-wise. Also, whenever a feature becomes undefined (i.e.: when the time step index becomes negative), it is replaced with the all-zeros vector. ", + "bbox": [ + 176, + 545, + 821, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike state features, which are only used when training the optimization algorithm, observation features $\\Psi ( \\cdot )$ are used both during training and at test time. Consequently, we use noisier observation features that can be computed more efficiently and require less memory overhead. The observation features consist of the following: ", + "bbox": [ + 174, + 580, + 825, + 637 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/1f917eed8b0253b0c9e42f56b3a726b04c793a3b7caf3b66cb9ae38147b4b29a.jpg", + "text": "$$\n\\begin{array} { r l } { \\bullet } & { \\left( \\hat { f } ( x ^ { ( t ) } ) - \\hat { f } ( x ^ { ( t - 1 ) } ) \\right) / \\hat { f } ( x ^ { ( t - 1 ) } ) } \\\\ & { \\bullet \\ \\nabla \\hat { f } ( x ^ { ( t ) } ) / \\left( \\left| \\nabla \\hat { f } ( x ^ { ( \\operatorname* { m a x } ( t - 1 , 0 ) ) } ) \\right| + 1 \\right) } \\\\ & { \\bullet \\ \\frac { \\left| x ^ { ( \\operatorname* { m a x } ( t - 1 , 1 ) ) } - x ^ { ( \\operatorname* { m a x } ( t - 2 , 0 ) ) } \\right| } { \\left| x ^ { ( t ) } - x ^ { ( t - 1 ) } \\right| + 0 . 1 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 212, + 647, + 470, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 746, + 328, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For clarity, we will refer to training of the optimization algorithm as “meta-training” to differentiate it from base-level training, which will simply be referred to as “training”. ", + "bbox": [ + 173, + 776, + 821, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We meta-trained an optimization algorithm on a single objective function, which corresponds to the problem of training a two-layer neural net with 48 input units, 48 hidden units and 10 output units on a randomly projected and normalized version of the MNIST training set with dimensionality 48 and unit variance in each dimension. We modelled the optimization algorithm using an recurrent neural net with a single layer of 128 LSTM (Hochreiter & Schmidhuber, 1997) cells. We used a time horizon of 400 iterations and a mini-batch size of 64 for computing stochastic gradients and objective values. We evaluate the optimization algorithm on its ability to generalize to unseen objective functions, which correspond to the problems of training neural nets on different tasks/datasets. ", + "bbox": [ + 173, + 811, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dba70d81f44e9b66aeedbf35fb705247b4161ee34db83bed3b4ba0993a37fd89.jpg", + "image_caption": [ + "Figure 3: Comparison of the various hand-engineered and learned algorithms on training neural nets with 48 input and hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-batches of size 10. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. " + ], + "image_footnote": [], + "bbox": [ + 210, + 122, + 782, + 237 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/afba0d91683764d4cd4161719336a9f48fc43fabd99a2b8440e62626e6cabcb3.jpg", + "image_caption": [ + "Figure 4: Comparison of the various hand-engineered and learned algorithms on training neural nets with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with minibatches of size 10. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. " + ], + "image_footnote": [], + "bbox": [ + 209, + 352, + 782, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the learned optimization algorithm on three datasets, the Toronto Faces Dataset (TFD), CIFAR-10 and CIFAR-100. These datasets are chosen for their very different characteristics from MNIST and each other: TFD contains 3300 grayscale images that have relatively little variation and has seven different categories, whereas CIFAR-100 contains 50,000 colour images that have varied appearance and has 100 different categories. ", + "bbox": [ + 174, + 575, + 825, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "All algorithms are tuned on the training objective function. For hand-engineered algorithms, this entails choosing the best hyperparameters; for learned algorithms, this entails meta-training on the objective function. We compare to the seven hand-engineered algorithms: stochastic gradient descent, momentum, conjugate gradient, L-BFGS, ADAM, AdaGrad and RMSprop. In addition, we compare to an optimization algorithm meta-trained using the method described in (Andrychowicz et al., 2016) on the same training objective function (training two-layer neural net on randomly projected and normalized MNIST) under the same setting (a time horizon of 400 iterations and a mini-batch size of 64). ", + "bbox": [ + 174, + 652, + 825, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First, we examine the performance of various optimization algorithms on similar objective functions. The optimization problems under consideration are those for training neural nets that have the same number of input and hidden units (48 and 48) as those used during meta-training. The number of output units varies with the number of categories in each dataset. We use the same mini-batch size as that used during meta-training. As shown in Figure 1, the optimization algorithm meta-trained using our method (which we will refer to as Predicted Step Descent) consistently descends to the optimum the fastest across all datasets. On the other hand, other algorithms are not as consistent and the relative ranking of other algorithms varies by dataset. This suggests that Predicted Step Descent has learned to be robust to variations in the data distributions, despite being trained on only one objective function, which is associated with a very specific data distribution that characterizes MNIST. It is also interesting to note that while the algorithm meta-trained using (Andrychowicz et al., 2016) (which we will refer to as L2LBGDBGD) performs well on CIFAR, it is unable to reach the optimum on TFD. ", + "bbox": [ + 174, + 770, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2226d92233c78e2cc8398fe5882a7f24065a5392b547b417b88d986cc9309d20.jpg", + "image_caption": [ + "Figure 5: Comparison of the various hand-engineered and learned algorithms on training neural nets with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 for 800 iterations with mini-batches of size 64. The vertical axis is the true objective value and the horizontal axis represents the iteration. Best viewed in colour. " + ], + "image_footnote": [], + "bbox": [ + 209, + 122, + 781, + 237 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 340, + 823, + 368 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Next, we change the architecture of the neural nets and see if Predicted Step Descent generalizes to the new architecture. We increase the number of input units to 100 and the number of hidden units to 200, so that the number of parameters is roughly increased by a factor of 8. As shown in Figure 2, Predicted Step Descent consistently outperforms other algorithms on each dataset, despite having not been trained to optimize neural nets of this architecture. Interestingly, while it exhibited a bit of oscillation initially on TFD and CIFAR-10, it quickly recovered and overtook other algorithms, which is reminiscent of the phenomenon reported in (Li & Malik, 2016) for low-dimensional optimization problems. This suggests that it has learned to detect when it is performing poorly and knows how to change tack accordingly. L2LBGDBGD experienced difficulties on TFD and CIFAR10 as well, but slowly diverged. ", + "bbox": [ + 174, + 376, + 825, + 515 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We now investigate how robust Predicted Step Descent is to stochasticity of the gradients. To this end, we take a look at its performance when we reduce the mini-batch size from 64 to 10 on both the original architecture with 48 input and hidden units and the enlarged architecture with 100 input units and 200 hidden units. As shown in Figure 3, on the original architecture, Predicted Step Descent still outperforms all other algorithms and is able to handle the increased stochasticity fairly well. In contrast, conjugate gradient and L2LBGDBGD had some difficulty handling the increased stochasticity on TFD and to a lesser extent, on CIFAR-10. In the former case, both diverged; in the latter case, both were progressing slowly towards the optimum. ", + "bbox": [ + 174, + 521, + 825, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "On the enlarged architecture, Predicted Step Descent experienced some significant oscillations on TFD and CIFAR-10, but still managed to achieve a much better objective value than all the other algorithms. Many hand-engineered algorithms also experienced much greater oscillations than previously, suggesting that the optimization problems are inherently harder. L2LBGDBGD diverged fairly quickly on these two datasets. ", + "bbox": [ + 174, + 640, + 823, + 709 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, we try doubling the number of iterations. As shown in Figure 5, despite being trained over a time horizon of 400 iterations, Predicted Step Descent behaves reasonably beyond the number of iterations it is trained for. ", + "bbox": [ + 176, + 717, + 821, + 758 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 789, + 318, + 804 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this paper, we presented a new method for learning optimization algorithms for high-dimensional stochastic problems. We applied the method to learning an optimization algorithm for training shallow neural nets. We showed that the algorithm learned using our method on the problem of training a neural net on MNIST generalizes to the problems of training neural nets on unrelated tasks/datasets like the Toronto Faces Dataset, CIFAR-10 and CIFAR-100. We also demonstrated that the learned optimization algorithm is robust to changes in the stochasticity of gradients and the neural net architecture. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, and Nando de Freitas. Learning to learn by gradient descent by gradient descent. arXiv preprint arXiv:1606.04474, 2016. ", + "bbox": [ + 174, + 126, + 825, + 167 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jonathan Baxter, Rich Caruana, Tom Mitchell, Lorien Y Pratt, Daniel L Silver, and Sebastian Thrun. NIPS 1995 workshop on learning to learn: Knowledge consolidation and transfer in inductive systems. https://web.archive.org/web/20000618135816/http://www. cs.cmu.edu/afs/cs.cmu.edu/user/caruana/pub/transfer.html, 1995. Accessed: 2015-12-05. ", + "bbox": [ + 173, + 178, + 825, + 247 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Y Bengio, S Bengio, and J Cloutier. Learning a synaptic learning rule. In Neural Networks, 1991., IJCNN-91-Seattle International Joint Conference on, volume 2, pp. 969–vol. IEEE, 1991. ", + "bbox": [ + 176, + 257, + 820, + 286 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yoshua Bengio. Gradient-based optimization of hyperparameters. Neural computation, 12(8):1889– 1900, 2000. ", + "bbox": [ + 173, + 295, + 821, + 324 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. The Journal of Machine Learning Research, 13(1):281–305, 2012. ", + "bbox": [ + 173, + 333, + 821, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "James S Bergstra, Remi Bardenet, Yoshua Bengio, and Bal ´ azs K ´ egl. Algorithms for hyper-parameter ´ optimization. In Advances in Neural Information Processing Systems, pp. 2546–2554, 2011. ", + "bbox": [ + 171, + 371, + 821, + 401 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "M Bray, E Koller-Meier, P Muller, L Van Gool, and NN Schraudolph. 3D hand tracking by rapid stochastic gradient descent using a skinning model. In Visual Media Production, 2004.(CVMP). 1st European Conference on, pp. 59–68. IET, 2004. ", + "bbox": [ + 173, + 410, + 823, + 453 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pavel Brazdil, Christophe Giraud Carrier, Carlos Soares, and Ricardo Vilalta. Metalearning: applications to data mining. Springer Science & Business Media, 2008. ", + "bbox": [ + 174, + 462, + 821, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pavel B Brazdil, Carlos Soares, and Joaquim Pinto Da Costa. Ranking learning algorithms: Using ibl and meta-learning on accuracy and time results. Machine Learning, 50(3):251–277, 2003. ", + "bbox": [ + 173, + 500, + 823, + 530 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Christian Daniel, Jonathan Taylor, and Sebastian Nowozin. Learning step size controllers for robust neural network training. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. ", + "bbox": [ + 174, + 537, + 823, + 568 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Justin Domke. Generic methods for optimization-based modeling. In AISTATS, volume 22, pp. 318–326, 2012. ", + "bbox": [ + 173, + 575, + 823, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011. ", + "bbox": [ + 173, + 613, + 823, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Matthias Feurer, Jost Tobias Springenberg, and Frank Hutter. Initializing bayesian hyperparameter optimization via meta-learning. In AAAI, pp. 1128–1135, 2015. ", + "bbox": [ + 173, + 651, + 823, + 683 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jie Fu, Zichuan Lin, Miao Liu, Nicholas Leonard, Jiashi Feng, and Tat-Seng Chua. Deep q-networks for accelerating the training of deep neural networks. arXiv preprint arXiv:1606.01467, 2016. ", + "bbox": [ + 171, + 690, + 823, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Karol Gregor and Yann LeCun. Learning fast approximations of sparse coding. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pp. 399–406, 2010. ", + "bbox": [ + 173, + 728, + 825, + 758 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Samantha Hansen. Using deep q-learning to control optimization hyperparameters. arXiv preprint arXiv:1602.04062, 2016. ", + "bbox": [ + 173, + 767, + 823, + 795 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 173, + 805, + 823, + 833 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sepp Hochreiter, A Steven Younger, and Peter R Conwell. Learning to learn using gradient descent. In International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001. ", + "bbox": [ + 173, + 843, + 823, + 872 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization, pp. 507–523. Springer, 2011. ", + "bbox": [ + 174, + 882, + 823, + 922 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. arXiv preprint arXiv:1504.00702, 2015. ", + "bbox": [ + 174, + 140, + 823, + 170 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ke Li and Jitendra Malik. Learning to optimize. CoRR, abs/1606.01885, 2016. ", + "bbox": [ + 174, + 178, + 694, + 194 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dougal Maclaurin, David Duvenaud, and Ryan P Adams. Gradient-based hyperparameter optimization through reversible learning. arXiv preprint arXiv:1502.03492, 2015. ", + "bbox": [ + 174, + 202, + 821, + 232 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Paul L Ruvolo, Ian Fasel, and Javier R Movellan. Optimization on a budget: A reinforcement learning approach. In Advances in Neural Information Processing Systems, pp. 1385–1392, 2009. ", + "bbox": [ + 173, + 239, + 823, + 270 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jurgen Schmidhuber. Optimal ordered problem solver. ¨ Machine Learning, 54(3):211–254, 2004. ", + "bbox": [ + 173, + 277, + 813, + 295 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems, pp. 2951–2959, 2012. ", + "bbox": [ + 169, + 301, + 823, + 332 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pablo Sprechmann, Roee Litman, Tal Ben Yakar, Alexander M Bronstein, and Guillermo Sapiro. Supervised sparse analysis and synthesis operators. In Advances in Neural Information Processing Systems, pp. 908–916, 2013. ", + "bbox": [ + 174, + 339, + 825, + 382 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Kevin Swersky, Jasper Snoek, and Ryan P Adams. Multi-task bayesian optimization. In Advances in neural information processing systems, pp. 2004–2012, 2013. ", + "bbox": [ + 171, + 391, + 823, + 421 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sebastian Thrun and Lorien Pratt. Learning to learn. Springer Science & Business Media, 2012. ", + "bbox": [ + 171, + 429, + 808, + 445 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2), 2012. ", + "bbox": [ + 171, + 453, + 823, + 483 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ricardo Vilalta and Youssef Drissi. A perspective view and survey of meta-learning. Artificial Intelligence Review, 18(2):77–95, 2002. ", + "bbox": [ + 169, + 491, + 823, + 520 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Huahua Wang and Arindam Banerjee. Bregman alternating direction method of multipliers. CoRR, abs/1306.3203, 2014. ", + "bbox": [ + 169, + 529, + 825, + 558 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/BkM27IxR-/BkM27IxR-_middle.json b/parse/train/BkM27IxR-/BkM27IxR-_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..4c4383b57c2bf776aadd61e59e954e7fc1f24477 --- /dev/null +++ b/parse/train/BkM27IxR-/BkM27IxR-_middle.json @@ -0,0 +1,34031 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 401, + 96 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 403, + 98 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 403, + 98 + ], + "score": 1.0, + "content": "LEARNING TO OPTIMIZE NEURAL NETS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 179 + ], + "lines": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "spans": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 193, + 468, + 325 + ], + "lines": [ + { + "bbox": [ + 141, + 193, + 469, + 205 + ], + "spans": [ + { + "bbox": [ + 141, + 193, + 469, + 205 + ], + "score": 1.0, + "content": "Learning to Optimize (Li & Malik, 2016) is a recently proposed framework for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 205, + 470, + 218 + ], + "spans": [ + { + "bbox": [ + 141, + 205, + 470, + 218 + ], + "score": 1.0, + "content": "learning optimization algorithms using reinforcement learning. In this paper, we", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 216, + 469, + 227 + ], + "spans": [ + { + "bbox": [ + 141, + 216, + 469, + 227 + ], + "score": 1.0, + "content": "explore learning an optimization algorithm for training shallow neural nets. Such", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 226, + 469, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 226, + 469, + 239 + ], + "score": 1.0, + "content": "high-dimensional stochastic optimization problems present interesting challenges", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 237, + 470, + 250 + ], + "spans": [ + { + "bbox": [ + 141, + 237, + 470, + 250 + ], + "score": 1.0, + "content": "for existing reinforcement learning algorithms. We develop an extension that is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 248, + 470, + 261 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 470, + 261 + ], + "score": 1.0, + "content": "suited to learning optimization algorithms in this setting and demonstrate that the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 259, + 469, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 469, + 272 + ], + "score": 1.0, + "content": "learned optimization algorithm consistently outperforms other known optimiza-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 271, + 470, + 283 + ], + "spans": [ + { + "bbox": [ + 141, + 271, + 470, + 283 + ], + "score": 1.0, + "content": "tion algorithms even on unseen tasks and is robust to changes in stochasticity of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 281, + 470, + 295 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 470, + 295 + ], + "score": 1.0, + "content": "gradients and the neural net architecture. More specifically, we show that an op-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 291, + 470, + 306 + ], + "spans": [ + { + "bbox": [ + 141, + 291, + 470, + 306 + ], + "score": 1.0, + "content": "timization algorithm trained with the proposed method on the problem of training", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "score": 1.0, + "content": "a neural net on MNIST generalizes to the problems of training neural nets on the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 314, + 349, + 325 + ], + "spans": [ + { + "bbox": [ + 142, + 314, + 349, + 325 + ], + "score": 1.0, + "content": "Toronto Faces Dataset, CIFAR-10 and CIFAR-100.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 351, + 206, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 208, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 208, + 367 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 377, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 377, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 392 + ], + "score": 1.0, + "content": "Machine learning is centred on the philosophy that learning patterns automatically from data is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "generally better than meticulously crafting rules by hand. This data-driven approach has delivered:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "today, machine learning techniques can be found in a wide range of application areas, both in AI and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 409, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 506, + 425 + ], + "score": 1.0, + "content": "beyond. Yet, there is one domain that has conspicuously been left untouched by machine learning:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 421, + 323, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 323, + 434 + ], + "score": 1.0, + "content": "the design of tools that power machine learning itself.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "One of the most widely used tools in machine learning is optimization algorithms. We have grown", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 450, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 504, + 462 + ], + "score": 1.0, + "content": "accustomed to seeing an optimization algorithm as a black box that takes in a model that we design", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 461, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 472 + ], + "score": 1.0, + "content": "and the data that we collect and outputs the optimal model parameters. The optimization algorithm", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 470, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 485 + ], + "score": 1.0, + "content": "itself largely stays static: its design is reserved for human experts, who must toil through many", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "rounds of theoretical analysis and empirical validation to devise a better optimization algorithm.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "Given this state of affairs, perhaps it is time for us to start practicing what we preach and learn how", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 503, + 143, + 517 + ], + "spans": [ + { + "bbox": [ + 104, + 503, + 143, + 517 + ], + "score": 1.0, + "content": "to learn.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 533 + ], + "score": 1.0, + "content": "Recently, Li & Malik (2016) and Andrychowicz et al. (2016) introduced two different frameworks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "for learning optimization algorithms. Whereas Andrychowicz et al. (2016) focuses on learning an", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "optimization algorithm for training models on a particular task, Li & Malik (2016) sets a more am-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 555, + 504, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 504, + 566 + ], + "score": 1.0, + "content": "bitious objective of learning an optimization algorithm for training models that is task-independent.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 564, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 579 + ], + "score": 1.0, + "content": "We study the latter paradigm in this paper and develop a method for learning an optimization algo-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "score": 1.0, + "content": "rithm for high-dimensional stochastic optimization problems, like the problem of training shallow", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 588, + 154, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 154, + 598 + ], + "score": 1.0, + "content": "neural nets.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "Under the “Learning to Optimize” framework proposed by Li & Malik (2016), the problem of learn-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "ing an optimization algorithm is formulated as a reinforcement learning problem. We consider the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "general structure of an unconstrained continuous optimization algorithm, as shown in Algorithm 1.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 288, + 650 + ], + "score": 1.0, + "content": "In each iteration, the algorithm takes a step", + "type": "text" + }, + { + "bbox": [ + 289, + 638, + 304, + 648 + ], + "score": 0.85, + "content": "\\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 637, + 470, + 650 + ], + "score": 1.0, + "content": "and uses it to update the current iterate", + "type": "text" + }, + { + "bbox": [ + 470, + 637, + 487, + 648 + ], + "score": 0.91, + "content": "\\boldsymbol { x } ^ { ( i ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 637, + 506, + 650 + ], + "score": 1.0, + "content": ". In", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 275, + 661 + ], + "score": 1.0, + "content": "hand-engineered optimization algorithms,", + "type": "text" + }, + { + "bbox": [ + 275, + 649, + 290, + 659 + ], + "score": 0.87, + "content": "\\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 649, + 444, + 661 + ], + "score": 1.0, + "content": "is computed using some fixed formula", + "type": "text" + }, + { + "bbox": [ + 444, + 649, + 452, + 660 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "that depends", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "on the objective function, the current iterate and past iterates. Often, it is simply a function of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 672, + 214, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 214, + 683 + ], + "score": 1.0, + "content": "current and past gradients.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 188, + 700 + ], + "score": 1.0, + "content": "Different choices of", + "type": "text" + }, + { + "bbox": [ + 188, + 688, + 196, + 699 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "yield different optimization algorithms and so each optimization algorithm is", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 291, + 712 + ], + "score": 1.0, + "content": "essentially characterized by its update formula", + "type": "text" + }, + { + "bbox": [ + 291, + 701, + 298, + 710 + ], + "score": 0.86, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 698, + 379, + 712 + ], + "score": 1.0, + "content": ". Hence, by learning", + "type": "text" + }, + { + "bbox": [ + 379, + 700, + 386, + 710 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 698, + 506, + 712 + ], + "score": 1.0, + "content": ", we can learn an optimization", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "algorithm. Li & Malik (2016) observed that an optimization algorithm can be viewed as a Markov", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 721, + 504, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 722, + 489, + 732 + ], + "score": 1.0, + "content": "decision process (MDP), where the state includes the current iterate, the action is the step vector", + "type": "text" + }, + { + "bbox": [ + 489, + 721, + 504, + 730 + ], + "score": 0.84, + "content": "\\Delta x", + "type": "inline_equation" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 401, + 96 + ], + "lines": [ + { + "bbox": [ + 104, + 78, + 403, + 98 + ], + "spans": [ + { + "bbox": [ + 104, + 78, + 403, + 98 + ], + "score": 1.0, + "content": "LEARNING TO OPTIMIZE NEURAL NETS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 111, + 115, + 245, + 138 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 179 + ], + "lines": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "spans": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 193, + 468, + 325 + ], + "lines": [ + { + "bbox": [ + 141, + 193, + 469, + 205 + ], + "spans": [ + { + "bbox": [ + 141, + 193, + 469, + 205 + ], + "score": 1.0, + "content": "Learning to Optimize (Li & Malik, 2016) is a recently proposed framework for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 205, + 470, + 218 + ], + "spans": [ + { + "bbox": [ + 141, + 205, + 470, + 218 + ], + "score": 1.0, + "content": "learning optimization algorithms using reinforcement learning. In this paper, we", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 216, + 469, + 227 + ], + "spans": [ + { + "bbox": [ + 141, + 216, + 469, + 227 + ], + "score": 1.0, + "content": "explore learning an optimization algorithm for training shallow neural nets. Such", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 226, + 469, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 226, + 469, + 239 + ], + "score": 1.0, + "content": "high-dimensional stochastic optimization problems present interesting challenges", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 237, + 470, + 250 + ], + "spans": [ + { + "bbox": [ + 141, + 237, + 470, + 250 + ], + "score": 1.0, + "content": "for existing reinforcement learning algorithms. We develop an extension that is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 248, + 470, + 261 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 470, + 261 + ], + "score": 1.0, + "content": "suited to learning optimization algorithms in this setting and demonstrate that the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 259, + 469, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 469, + 272 + ], + "score": 1.0, + "content": "learned optimization algorithm consistently outperforms other known optimiza-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 271, + 470, + 283 + ], + "spans": [ + { + "bbox": [ + 141, + 271, + 470, + 283 + ], + "score": 1.0, + "content": "tion algorithms even on unseen tasks and is robust to changes in stochasticity of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 281, + 470, + 295 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 470, + 295 + ], + "score": 1.0, + "content": "gradients and the neural net architecture. More specifically, we show that an op-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 291, + 470, + 306 + ], + "spans": [ + { + "bbox": [ + 141, + 291, + 470, + 306 + ], + "score": 1.0, + "content": "timization algorithm trained with the proposed method on the problem of training", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "score": 1.0, + "content": "a neural net on MNIST generalizes to the problems of training neural nets on the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 314, + 349, + 325 + ], + "spans": [ + { + "bbox": [ + 142, + 314, + 349, + 325 + ], + "score": 1.0, + "content": "Toronto Faces Dataset, CIFAR-10 and CIFAR-100.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9.5, + "bbox_fs": [ + 141, + 193, + 470, + 325 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 351, + 206, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 350, + 208, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 208, + 367 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 377, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 377, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 392 + ], + "score": 1.0, + "content": "Machine learning is centred on the philosophy that learning patterns automatically from data is", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "generally better than meticulously crafting rules by hand. This data-driven approach has delivered:", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "today, machine learning techniques can be found in a wide range of application areas, both in AI and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 409, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 506, + 425 + ], + "score": 1.0, + "content": "beyond. Yet, there is one domain that has conspicuously been left untouched by machine learning:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 421, + 323, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 323, + 434 + ], + "score": 1.0, + "content": "the design of tools that power machine learning itself.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 377, + 506, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "One of the most widely used tools in machine learning is optimization algorithms. We have grown", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 450, + 504, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 504, + 462 + ], + "score": 1.0, + "content": "accustomed to seeing an optimization algorithm as a black box that takes in a model that we design", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 461, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 472 + ], + "score": 1.0, + "content": "and the data that we collect and outputs the optimal model parameters. The optimization algorithm", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 470, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 485 + ], + "score": 1.0, + "content": "itself largely stays static: its design is reserved for human experts, who must toil through many", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "rounds of theoretical analysis and empirical validation to devise a better optimization algorithm.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "Given this state of affairs, perhaps it is time for us to start practicing what we preach and learn how", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 503, + 143, + 517 + ], + "spans": [ + { + "bbox": [ + 104, + 503, + 143, + 517 + ], + "score": 1.0, + "content": "to learn.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 438, + 505, + 517 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 533 + ], + "score": 1.0, + "content": "Recently, Li & Malik (2016) and Andrychowicz et al. (2016) introduced two different frameworks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "for learning optimization algorithms. Whereas Andrychowicz et al. (2016) focuses on learning an", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "optimization algorithm for training models on a particular task, Li & Malik (2016) sets a more am-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 555, + 504, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 504, + 566 + ], + "score": 1.0, + "content": "bitious objective of learning an optimization algorithm for training models that is task-independent.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 564, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 579 + ], + "score": 1.0, + "content": "We study the latter paradigm in this paper and develop a method for learning an optimization algo-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "score": 1.0, + "content": "rithm for high-dimensional stochastic optimization problems, like the problem of training shallow", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 588, + 154, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 154, + 598 + ], + "score": 1.0, + "content": "neural nets.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 522, + 506, + 598 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "Under the “Learning to Optimize” framework proposed by Li & Malik (2016), the problem of learn-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "ing an optimization algorithm is formulated as a reinforcement learning problem. We consider the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "general structure of an unconstrained continuous optimization algorithm, as shown in Algorithm 1.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 288, + 650 + ], + "score": 1.0, + "content": "In each iteration, the algorithm takes a step", + "type": "text" + }, + { + "bbox": [ + 289, + 638, + 304, + 648 + ], + "score": 0.85, + "content": "\\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 637, + 470, + 650 + ], + "score": 1.0, + "content": "and uses it to update the current iterate", + "type": "text" + }, + { + "bbox": [ + 470, + 637, + 487, + 648 + ], + "score": 0.91, + "content": "\\boldsymbol { x } ^ { ( i ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 637, + 506, + 650 + ], + "score": 1.0, + "content": ". In", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 275, + 661 + ], + "score": 1.0, + "content": "hand-engineered optimization algorithms,", + "type": "text" + }, + { + "bbox": [ + 275, + 649, + 290, + 659 + ], + "score": 0.87, + "content": "\\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 649, + 444, + 661 + ], + "score": 1.0, + "content": "is computed using some fixed formula", + "type": "text" + }, + { + "bbox": [ + 444, + 649, + 452, + 660 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "that depends", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "on the objective function, the current iterate and past iterates. Often, it is simply a function of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 672, + 214, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 214, + 683 + ], + "score": 1.0, + "content": "current and past gradients.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 603, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 188, + 700 + ], + "score": 1.0, + "content": "Different choices of", + "type": "text" + }, + { + "bbox": [ + 188, + 688, + 196, + 699 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "yield different optimization algorithms and so each optimization algorithm is", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 291, + 712 + ], + "score": 1.0, + "content": "essentially characterized by its update formula", + "type": "text" + }, + { + "bbox": [ + 291, + 701, + 298, + 710 + ], + "score": 0.86, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 698, + 379, + 712 + ], + "score": 1.0, + "content": ". Hence, by learning", + "type": "text" + }, + { + "bbox": [ + 379, + 700, + 386, + 710 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 698, + 506, + 712 + ], + "score": 1.0, + "content": ", we can learn an optimization", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "algorithm. Li & Malik (2016) observed that an optimization algorithm can be viewed as a Markov", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 107, + 721, + 504, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 722, + 489, + 732 + ], + "score": 1.0, + "content": "decision process (MDP), where the state includes the current iterate, the action is the step vector", + "type": "text" + }, + { + "bbox": [ + 489, + 721, + 504, + 730 + ], + "score": 0.84, + "content": "\\Delta x", + "type": "inline_equation" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 686, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 97, + 280, + 200 + ], + "lines": [ + { + "bbox": [ + 106, + 97, + 233, + 109 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 225, + 109 + ], + "score": 1.0, + "content": "Require: Objective function", + "type": "text" + }, + { + "bbox": [ + 226, + 98, + 233, + 109 + ], + "score": 0.65, + "content": "f", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 107, + 279, + 122 + ], + "spans": [ + { + "bbox": [ + 117, + 108, + 147, + 119 + ], + "score": 0.88, + "content": "\\bar { x ^ { ( 0 ) } } \\gets", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 107, + 272, + 122 + ], + "score": 1.0, + "content": "random point in the domain of", + "type": "text" + }, + { + "bbox": [ + 272, + 110, + 279, + 121 + ], + "score": 0.72, + "content": "f", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 120, + 195, + 132 + ], + "spans": [ + { + "bbox": [ + 115, + 120, + 131, + 132 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 131, + 121, + 189, + 132 + ], + "score": 0.49, + "content": "i = 1 , 2 , \\dots { \\bf d }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 120, + 195, + 132 + ], + "score": 1.0, + "content": "o", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 130, + 260, + 144 + ], + "spans": [ + { + "bbox": [ + 131, + 130, + 260, + 144 + ], + "score": 0.74, + "content": "\\Delta x \\gets \\acute { \\phi } ( f , \\{ x ^ { ( 0 ) } , \\ldots , x ^ { ( i - 1 ) } \\} )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 142, + 264, + 155 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 264, + 155 + ], + "score": 1.0, + "content": "if stopping condition is met then", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 144, + 153, + 204, + 165 + ], + "spans": [ + { + "bbox": [ + 144, + 153, + 204, + 165 + ], + "score": 1.0, + "content": "return x(i−1)", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 165, + 160, + 176 + ], + "spans": [ + { + "bbox": [ + 129, + 165, + 160, + 176 + ], + "score": 1.0, + "content": "end if", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 175, + 216, + 187 + ], + "spans": [ + { + "bbox": [ + 131, + 175, + 216, + 187 + ], + "score": 0.32, + "content": "x ^ { ( i ) } \\gets x ^ { ( i - 1 ) } + \\Delta x", + "type": "inline_equation", + "image_path": "f3101b7278e6482d518fcf99960473adf1a4c6ebf4607953f74df2d02368c299.jpg" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 150, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 150, + 199 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 504, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 252, + 236 + ], + "score": 1.0, + "content": "and the policy is the update formula", + "type": "text" + }, + { + "bbox": [ + 252, + 223, + 259, + 234 + ], + "score": 0.82, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 221, + 389, + 236 + ], + "score": 1.0, + "content": ". Hence, the problem of learning", + "type": "text" + }, + { + "bbox": [ + 390, + 223, + 397, + 234 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 221, + 505, + 236 + ], + "score": 1.0, + "content": "simply reduces to a policy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 233, + 173, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 173, + 245 + ], + "score": 1.0, + "content": "search problem.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "score": 1.0, + "content": "In this paper, we build on the method proposed in (Li & Malik, 2016) and develop an extension that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "is suited to learning optimization algorithms for high-dimensional stochastic problems. We use it to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "learn an optimization algorithm for training shallow neural nets and show that it outperforms popular", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "hand-engineered optimization algorithms like ADAM (Kingma & Ba, 2014), AdaGrad (Duchi et al.,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 293, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 307 + ], + "score": 1.0, + "content": "2011) and RMSprop (Tieleman & Hinton, 2012) and an optimization algorithm learned using the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "score": 1.0, + "content": "supervised learning method proposed in (Andrychowicz et al., 2016). Furthermore, we demonstrate", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "that our optimization algorithm learned from the experience of training on MNIST generalizes to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 327, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 506, + 339 + ], + "score": 1.0, + "content": "training on other datasets that have very dissimilar statistics, like the Toronto Faces Dataset, CIFAR-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 338, + 187, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 187, + 350 + ], + "score": 1.0, + "content": "10 and CIFAR-100.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 366, + 211, + 379 + ], + "lines": [ + { + "bbox": [ + 104, + 365, + 213, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 365, + 213, + 381 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "The line of work on learning optimization algorithms is fairly recent. Li & Malik (2016) and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "Andrychowicz et al. (2016) were the first to propose learning general optimization algorithms. Li &", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "Malik (2016) explored learning task-independent optimization algorithms and used reinforcement", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 423, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 439 + ], + "score": 1.0, + "content": "learning to learn the optimization algorithm, while Andrychowicz et al. (2016) investigated learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 434, + 388, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 388, + 450 + ], + "score": 1.0, + "content": "task-dependent optimization algorithms and used supervised learning.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "In the special case where objective functions that the optimization algorithm is trained on are loss", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "functions for training other models, these methods can be used for “learning to learn” or “meta-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 474, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 504, + 486 + ], + "score": 1.0, + "content": "learning”. While these terms have appeared from time to time in the literature (Baxter et al., 1995;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "Vilalta & Drissi, 2002; Brazdil et al., 2008; Thrun & Pratt, 2012), they have been used by different", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "authors to refer to disparate methods with different purposes. These methods all share the objective", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "of learning some form of meta-knowledge about learning, but differ in the type of meta-knowledge", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 518, + 462, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 462, + 531 + ], + "score": 1.0, + "content": "they aim to learn. We can divide the various methods into the following three categories.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 108, + 543, + 253, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 255, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 255, + 556 + ], + "score": 1.0, + "content": "2.1 LEARNING WHAT TO LEARN", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Methods in this category Thrun & Pratt (2012) aim to learn what parameter values of the base-level", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "learner are useful across a family of related tasks. The meta-knowledge captures commonalities", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "score": 1.0, + "content": "shared by tasks in the family, which enables learning on a new task from the family to be performed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 596, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 611 + ], + "score": 1.0, + "content": "more quickly. Most early methods fall into this category; this line of work has blossomed into an", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 607, + 417, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 417, + 622 + ], + "score": 1.0, + "content": "area that has later become known as transfer learning and multi-task learning.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 633, + 293, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 294, + 647 + ], + "score": 1.0, + "content": "2.2 LEARNING WHICH MODEL TO LEARN", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "Methods in this category Brazdil et al. (2008) aim to learn which base-level learner achieves the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "best performance on a task. The meta-knowledge captures correlations between different tasks and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "the performance of different base-level learners on those tasks. One challenge under this setting is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "to decide on a parameterization of the space of base-level learners that is both rich enough to be", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "capable of representing disparate base-level learners and compact enough to permit tractable search", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "over this space. Brazdil et al. (2003) proposes a nonparametric representation and stores examples of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "different base-level learners in a database, whereas Schmidhuber (2004) proposes representing base-", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 97, + 280, + 200 + ], + "lines": [ + { + "bbox": [ + 106, + 97, + 233, + 109 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 225, + 109 + ], + "score": 1.0, + "content": "Require: Objective function", + "type": "text" + }, + { + "bbox": [ + 226, + 98, + 233, + 109 + ], + "score": 0.65, + "content": "f", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 107, + 279, + 122 + ], + "spans": [ + { + "bbox": [ + 117, + 108, + 147, + 119 + ], + "score": 0.88, + "content": "\\bar { x ^ { ( 0 ) } } \\gets", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 107, + 272, + 122 + ], + "score": 1.0, + "content": "random point in the domain of", + "type": "text" + }, + { + "bbox": [ + 272, + 110, + 279, + 121 + ], + "score": 0.72, + "content": "f", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 120, + 195, + 132 + ], + "spans": [ + { + "bbox": [ + 115, + 120, + 131, + 132 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 131, + 121, + 189, + 132 + ], + "score": 0.49, + "content": "i = 1 , 2 , \\dots { \\bf d }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 120, + 195, + 132 + ], + "score": 1.0, + "content": "o", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 130, + 260, + 144 + ], + "spans": [ + { + "bbox": [ + 131, + 130, + 260, + 144 + ], + "score": 0.74, + "content": "\\Delta x \\gets \\acute { \\phi } ( f , \\{ x ^ { ( 0 ) } , \\ldots , x ^ { ( i - 1 ) } \\} )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 142, + 264, + 155 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 264, + 155 + ], + "score": 1.0, + "content": "if stopping condition is met then", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 144, + 153, + 204, + 165 + ], + "spans": [ + { + "bbox": [ + 144, + 153, + 204, + 165 + ], + "score": 1.0, + "content": "return x(i−1)", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 165, + 160, + 176 + ], + "spans": [ + { + "bbox": [ + 129, + 165, + 160, + 176 + ], + "score": 1.0, + "content": "end if", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 175, + 216, + 187 + ], + "spans": [ + { + "bbox": [ + 131, + 175, + 216, + 187 + ], + "score": 0.32, + "content": "x ^ { ( i ) } \\gets x ^ { ( i - 1 ) } + \\Delta x", + "type": "inline_equation", + "image_path": "f3101b7278e6482d518fcf99960473adf1a4c6ebf4607953f74df2d02368c299.jpg" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 150, + 199 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 150, + 199 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 97, + 279, + 199 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 504, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 252, + 236 + ], + "score": 1.0, + "content": "and the policy is the update formula", + "type": "text" + }, + { + "bbox": [ + 252, + 223, + 259, + 234 + ], + "score": 0.82, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 221, + 389, + 236 + ], + "score": 1.0, + "content": ". Hence, the problem of learning", + "type": "text" + }, + { + "bbox": [ + 390, + 223, + 397, + 234 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 221, + 505, + 236 + ], + "score": 1.0, + "content": "simply reduces to a policy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 233, + 173, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 173, + 245 + ], + "score": 1.0, + "content": "search problem.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 221, + 505, + 245 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 263 + ], + "score": 1.0, + "content": "In this paper, we build on the method proposed in (Li & Malik, 2016) and develop an extension that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "is suited to learning optimization algorithms for high-dimensional stochastic problems. We use it to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "learn an optimization algorithm for training shallow neural nets and show that it outperforms popular", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "hand-engineered optimization algorithms like ADAM (Kingma & Ba, 2014), AdaGrad (Duchi et al.,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 293, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 307 + ], + "score": 1.0, + "content": "2011) and RMSprop (Tieleman & Hinton, 2012) and an optimization algorithm learned using the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "score": 1.0, + "content": "supervised learning method proposed in (Andrychowicz et al., 2016). Furthermore, we demonstrate", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "that our optimization algorithm learned from the experience of training on MNIST generalizes to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 327, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 506, + 339 + ], + "score": 1.0, + "content": "training on other datasets that have very dissimilar statistics, like the Toronto Faces Dataset, CIFAR-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 338, + 187, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 187, + 350 + ], + "score": 1.0, + "content": "10 and CIFAR-100.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 250, + 506, + 350 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 366, + 211, + 379 + ], + "lines": [ + { + "bbox": [ + 104, + 365, + 213, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 365, + 213, + 381 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "The line of work on learning optimization algorithms is fairly recent. Li & Malik (2016) and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "Andrychowicz et al. (2016) were the first to propose learning general optimization algorithms. Li &", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "Malik (2016) explored learning task-independent optimization algorithms and used reinforcement", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 423, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 439 + ], + "score": 1.0, + "content": "learning to learn the optimization algorithm, while Andrychowicz et al. (2016) investigated learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 434, + 388, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 388, + 450 + ], + "score": 1.0, + "content": "task-dependent optimization algorithms and used supervised learning.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 390, + 506, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "In the special case where objective functions that the optimization algorithm is trained on are loss", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "functions for training other models, these methods can be used for “learning to learn” or “meta-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 474, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 504, + 486 + ], + "score": 1.0, + "content": "learning”. While these terms have appeared from time to time in the literature (Baxter et al., 1995;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "Vilalta & Drissi, 2002; Brazdil et al., 2008; Thrun & Pratt, 2012), they have been used by different", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "authors to refer to disparate methods with different purposes. These methods all share the objective", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "of learning some form of meta-knowledge about learning, but differ in the type of meta-knowledge", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 518, + 462, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 462, + 531 + ], + "score": 1.0, + "content": "they aim to learn. We can divide the various methods into the following three categories.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 452, + 506, + 531 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 543, + 253, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 255, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 255, + 556 + ], + "score": 1.0, + "content": "2.1 LEARNING WHAT TO LEARN", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Methods in this category Thrun & Pratt (2012) aim to learn what parameter values of the base-level", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "learner are useful across a family of related tasks. The meta-knowledge captures commonalities", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "score": 1.0, + "content": "shared by tasks in the family, which enables learning on a new task from the family to be performed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 596, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 611 + ], + "score": 1.0, + "content": "more quickly. Most early methods fall into this category; this line of work has blossomed into an", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 607, + 417, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 417, + 622 + ], + "score": 1.0, + "content": "area that has later become known as transfer learning and multi-task learning.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 565, + 506, + 622 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 633, + 293, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 294, + 647 + ], + "score": 1.0, + "content": "2.2 LEARNING WHICH MODEL TO LEARN", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "Methods in this category Brazdil et al. (2008) aim to learn which base-level learner achieves the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "best performance on a task. The meta-knowledge captures correlations between different tasks and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "the performance of different base-level learners on those tasks. One challenge under this setting is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "to decide on a parameterization of the space of base-level learners that is both rich enough to be", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "capable of representing disparate base-level learners and compact enough to permit tractable search", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "over this space. Brazdil et al. (2003) proposes a nonparametric representation and stores examples of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "different base-level learners in a database, whereas Schmidhuber (2004) proposes representing base-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "level learners as general-purpose programs. The former has limited representation power, while the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "latter makes search and learning in the space of base-level learners intractable. Hochreiter et al.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "(2001) views the (online) training procedure of any base-learner as a black box function that maps", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "a sequence of training examples to a sequence of predictions and models it as a recurrent neural", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 138 + ], + "score": 1.0, + "content": "net. Under this formulation, meta-training reduces to training the recurrent net, and the base-level", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 344, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 344, + 149 + ], + "score": 1.0, + "content": "learner is encoded in the memory state of the recurrent net.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 654, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "level learners as general-purpose programs. The former has limited representation power, while the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "latter makes search and learning in the space of base-level learners intractable. Hochreiter et al.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "(2001) views the (online) training procedure of any base-learner as a black box function that maps", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "a sequence of training examples to a sequence of predictions and models it as a recurrent neural", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 138 + ], + "score": 1.0, + "content": "net. Under this formulation, meta-training reduces to training the recurrent net, and the base-level", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 344, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 344, + 149 + ], + "score": 1.0, + "content": "learner is encoded in the memory state of the recurrent net.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "Hyperparameter optimization can be seen as another example of methods in this category. The", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 179 + ], + "score": 1.0, + "content": "space of base-level learners to search over is parameterized by a predefined set of hyperparameters.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "Unlike the methods above, multiple trials with different hyperparameter settings on the same task are", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "permitted, and so generalization across tasks is not required. The discovered hyperparameters are", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "generally specific to the task at hand and hyperparameter optimization must be rerun for new tasks.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 208, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 506, + 222 + ], + "score": 1.0, + "content": "Various kinds of methods have been proposed, such those based on Bayesian optimization (Hutter", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "et al., 2011; Bergstra et al., 2011; Snoek et al., 2012; Swersky et al., 2013; Feurer et al., 2015),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "score": 1.0, + "content": "random search (Bergstra & Bengio, 2012) and gradient-based optimization (Bengio, 2000; Domke,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 227, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 227, + 254 + ], + "score": 1.0, + "content": "2012; Maclaurin et al., 2015).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 108, + 267, + 248, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 250, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 250, + 279 + ], + "score": 1.0, + "content": "2.3 LEARNING HOW TO LEARN", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 300 + ], + "score": 1.0, + "content": "Methods in this category aim to learn a good algorithm for training a base-level learner. Unlike", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "methods in the previous categories, the goal is not to learn about the outcome of learning, but rather", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 308, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 324 + ], + "score": 1.0, + "content": "the process of learning. The meta-knowledge captures commonalities in the behaviours of learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 334 + ], + "score": 1.0, + "content": "algorithms that achieve good performance. The base-level learner and the task are given by the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 104, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "user, so the learned algorithm must generalize across base-level learners and tasks. Since learning in", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "most cases is equivalent to optimizing some objective function, learning a learning algorithm often", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "reduces to learning an optimization algorithm. This problem was explored in (Li & Malik, 2016)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "and (Andrychowicz et al., 2016). Closely related is (Bengio et al., 1991), which learns a Hebb-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "like synaptic learning rule that does not depend on the objective function, which does not allow for", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 387, + 293, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 293, + 399 + ], + "score": 1.0, + "content": "generalization to different objective functions.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 403, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "Various work has explored learning how to adjust the hyperparameters of hand-engineered optimiza-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 413, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 505, + 428 + ], + "score": 1.0, + "content": "tion algorithms, like the step size (Hansen, 2016; Daniel et al., 2016; Fu et al., 2016) or the damping", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "factor in the Levenberg-Marquardt algorithm (Ruvolo et al., 2009). Related to this line of work is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "stochastic meta-descent (Bray et al., 2004), which derives a rule for adjusting the step size analyt-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 461 + ], + "score": 1.0, + "content": "ically. A different line of work (Gregor & LeCun, 2010; Sprechmann et al., 2013) parameterizes", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "intermediate operands of special-purpose solvers for a class of optimization problems that arise in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 469, + 336, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 336, + 483 + ], + "score": 1.0, + "content": "sparse coding and learns them using supervised learning.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 252, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 254, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 254, + 511 + ], + "score": 1.0, + "content": "3 LEARNING TO OPTIMIZE", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "title", + "bbox": [ + 107, + 522, + 169, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 521, + 171, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 171, + 535 + ], + "score": 1.0, + "content": "3.1 SETTING", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "In the “Learning to Optimize” framework, we are given a set of training objective functions", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 554, + 504, + 567 + ], + "spans": [ + { + "bbox": [ + 107, + 554, + 149, + 566 + ], + "score": 0.91, + "content": "f _ { 1 } , \\ldots , f _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 554, + 273, + 567 + ], + "score": 1.0, + "content": "drawn from some distribution", + "type": "text" + }, + { + "bbox": [ + 273, + 555, + 282, + 564 + ], + "score": 0.81, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 554, + 397, + 567 + ], + "score": 1.0, + "content": ". An optimization algorithm", + "type": "text" + }, + { + "bbox": [ + 397, + 555, + 406, + 564 + ], + "score": 0.79, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 554, + 504, + 567 + ], + "score": 1.0, + "content": "takes an objective func-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 564, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 104, + 564, + 124, + 579 + ], + "score": 1.0, + "content": "tion", + "type": "text" + }, + { + "bbox": [ + 125, + 567, + 132, + 577 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 564, + 217, + 579 + ], + "score": 1.0, + "content": "and an initial iterate", + "type": "text" + }, + { + "bbox": [ + 217, + 565, + 234, + 576 + ], + "score": 0.9, + "content": "x ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 564, + 416, + 579 + ], + "score": 1.0, + "content": "as input and produces a sequence of iterates", + "type": "text" + }, + { + "bbox": [ + 416, + 565, + 474, + 577 + ], + "score": 0.93, + "content": "x ^ { ( 1 ) } , \\ldots , x ^ { ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 564, + 506, + 579 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 575, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 126, + 588 + ], + "score": 0.89, + "content": "x ^ { ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 575, + 434, + 592 + ], + "score": 1.0, + "content": "is the solution found by the optimizer. We are also given a distribution", + "type": "text" + }, + { + "bbox": [ + 434, + 578, + 443, + 588 + ], + "score": 0.81, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 575, + 506, + 592 + ], + "score": 1.0, + "content": "that generates", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 588, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 177, + 604 + ], + "score": 1.0, + "content": "the initial iterate", + "type": "text" + }, + { + "bbox": [ + 177, + 589, + 195, + 600 + ], + "score": 0.9, + "content": "x ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 588, + 264, + 604 + ], + "score": 1.0, + "content": "and a meta-loss", + "type": "text" + }, + { + "bbox": [ + 265, + 590, + 273, + 600 + ], + "score": 0.76, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 588, + 419, + 604 + ], + "score": 1.0, + "content": ", which takes an objective function", + "type": "text" + }, + { + "bbox": [ + 419, + 590, + 426, + 602 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 588, + 506, + 604 + ], + "score": 1.0, + "content": "and a sequence of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 598, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 104, + 598, + 139, + 615 + ], + "score": 1.0, + "content": "iterates", + "type": "text" + }, + { + "bbox": [ + 139, + 601, + 197, + 613 + ], + "score": 0.92, + "content": "x ^ { ( 1 ) } , \\ldots , x ^ { ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 598, + 506, + 615 + ], + "score": 1.0, + "content": "produced by an optimization algorithm as input and outputs a scalar that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 450, + 626 + ], + "score": 1.0, + "content": "measures the quality of the iterates. The goal is to learn an optimization algorithm", + "type": "text" + }, + { + "bbox": [ + 450, + 613, + 464, + 623 + ], + "score": 0.86, + "content": "{ \\mathcal { P } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 622, + 507, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 241, + 638 + ], + "score": 0.85, + "content": "\\mathbb { E } _ { f \\sim \\mathcal { F } , x ^ { ( 0 ) } \\sim \\mathcal { D } } \\left[ \\mathcal { L } ( f , \\mathcal { P } ^ { * } ( f , x ^ { ( 0 ) } ) ) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 622, + 507, + 639 + ], + "score": 1.0, + "content": "is minimized. The meta-loss is chosen to penalize optimization", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "algorithms that exhibit behaviours we find undesirable, like slow convergence or excessive oscilla-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 101, + 646, + 510, + 684 + ], + "spans": [ + { + "bbox": [ + 101, + 646, + 308, + 684 + ], + "score": 1.0, + "content": "tions. Assuming we would like to learn an algoritha good choice of meta-loss would then simply be regret and can be interpreted as the area under the", + "type": "text" + }, + { + "bbox": [ + 308, + 658, + 363, + 672 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { i = 1 } ^ { T } f ( x ^ { ( i ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 646, + 510, + 684 + ], + "score": 1.0, + "content": "zes the objective function it is given,, which is equivalent to cumulativective values over time.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 202, + 702 + ], + "score": 1.0, + "content": "The objective functions", + "type": "text" + }, + { + "bbox": [ + 202, + 688, + 245, + 699 + ], + "score": 0.92, + "content": "f _ { 1 } , \\ldots , f _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "may correspond to loss functions for training base-level learners,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "in which case the algorithm that learns the optimization algorithm can be viewed as a meta-learner.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "In this setting, each objective function is the loss function for training a particular base-learner on", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "score": 1.0, + "content": "a particular task, and so the set of training objective functions can be loss functions for training a", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 148 + ], + "lines": [], + "index": 2.5, + "bbox_fs": [ + 105, + 82, + 506, + 149 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 155, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 167 + ], + "score": 1.0, + "content": "Hyperparameter optimization can be seen as another example of methods in this category. The", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 179 + ], + "score": 1.0, + "content": "space of base-level learners to search over is parameterized by a predefined set of hyperparameters.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "Unlike the methods above, multiple trials with different hyperparameter settings on the same task are", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "permitted, and so generalization across tasks is not required. The discovered hyperparameters are", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "generally specific to the task at hand and hyperparameter optimization must be rerun for new tasks.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 208, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 506, + 222 + ], + "score": 1.0, + "content": "Various kinds of methods have been proposed, such those based on Bayesian optimization (Hutter", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "et al., 2011; Bergstra et al., 2011; Snoek et al., 2012; Swersky et al., 2013; Feurer et al., 2015),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "score": 1.0, + "content": "random search (Bergstra & Bengio, 2012) and gradient-based optimization (Bengio, 2000; Domke,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 227, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 227, + 254 + ], + "score": 1.0, + "content": "2012; Maclaurin et al., 2015).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 155, + 506, + 254 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 267, + 248, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 250, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 250, + 279 + ], + "score": 1.0, + "content": "2.3 LEARNING HOW TO LEARN", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 300 + ], + "score": 1.0, + "content": "Methods in this category aim to learn a good algorithm for training a base-level learner. Unlike", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "methods in the previous categories, the goal is not to learn about the outcome of learning, but rather", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 308, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 324 + ], + "score": 1.0, + "content": "the process of learning. The meta-knowledge captures commonalities in the behaviours of learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 334 + ], + "score": 1.0, + "content": "algorithms that achieve good performance. The base-level learner and the task are given by the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 104, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "user, so the learned algorithm must generalize across base-level learners and tasks. Since learning in", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "most cases is equivalent to optimizing some objective function, learning a learning algorithm often", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "reduces to learning an optimization algorithm. This problem was explored in (Li & Malik, 2016)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "and (Andrychowicz et al., 2016). Closely related is (Bengio et al., 1991), which learns a Hebb-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "like synaptic learning rule that does not depend on the objective function, which does not allow for", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 387, + 293, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 293, + 399 + ], + "score": 1.0, + "content": "generalization to different objective functions.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 288, + 506, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 403, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "Various work has explored learning how to adjust the hyperparameters of hand-engineered optimiza-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 413, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 505, + 428 + ], + "score": 1.0, + "content": "tion algorithms, like the step size (Hansen, 2016; Daniel et al., 2016; Fu et al., 2016) or the damping", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "factor in the Levenberg-Marquardt algorithm (Ruvolo et al., 2009). Related to this line of work is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "stochastic meta-descent (Bray et al., 2004), which derives a rule for adjusting the step size analyt-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 461 + ], + "score": 1.0, + "content": "ically. A different line of work (Gregor & LeCun, 2010; Sprechmann et al., 2013) parameterizes", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "intermediate operands of special-purpose solvers for a class of optimization problems that arise in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 469, + 336, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 336, + 483 + ], + "score": 1.0, + "content": "sparse coding and learns them using supervised learning.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 403, + 506, + 483 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 252, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 254, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 254, + 511 + ], + "score": 1.0, + "content": "3 LEARNING TO OPTIMIZE", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "title", + "bbox": [ + 107, + 522, + 169, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 521, + 171, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 171, + 535 + ], + "score": 1.0, + "content": "3.1 SETTING", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "In the “Learning to Optimize” framework, we are given a set of training objective functions", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 554, + 504, + 567 + ], + "spans": [ + { + "bbox": [ + 107, + 554, + 149, + 566 + ], + "score": 0.91, + "content": "f _ { 1 } , \\ldots , f _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 554, + 273, + 567 + ], + "score": 1.0, + "content": "drawn from some distribution", + "type": "text" + }, + { + "bbox": [ + 273, + 555, + 282, + 564 + ], + "score": 0.81, + "content": "\\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 554, + 397, + 567 + ], + "score": 1.0, + "content": ". An optimization algorithm", + "type": "text" + }, + { + "bbox": [ + 397, + 555, + 406, + 564 + ], + "score": 0.79, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 554, + 504, + 567 + ], + "score": 1.0, + "content": "takes an objective func-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 564, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 104, + 564, + 124, + 579 + ], + "score": 1.0, + "content": "tion", + "type": "text" + }, + { + "bbox": [ + 125, + 567, + 132, + 577 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 564, + 217, + 579 + ], + "score": 1.0, + "content": "and an initial iterate", + "type": "text" + }, + { + "bbox": [ + 217, + 565, + 234, + 576 + ], + "score": 0.9, + "content": "x ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 564, + 416, + 579 + ], + "score": 1.0, + "content": "as input and produces a sequence of iterates", + "type": "text" + }, + { + "bbox": [ + 416, + 565, + 474, + 577 + ], + "score": 0.93, + "content": "x ^ { ( 1 ) } , \\ldots , x ^ { ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 564, + 506, + 579 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 575, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 126, + 588 + ], + "score": 0.89, + "content": "x ^ { ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 575, + 434, + 592 + ], + "score": 1.0, + "content": "is the solution found by the optimizer. We are also given a distribution", + "type": "text" + }, + { + "bbox": [ + 434, + 578, + 443, + 588 + ], + "score": 0.81, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 575, + 506, + 592 + ], + "score": 1.0, + "content": "that generates", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 588, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 177, + 604 + ], + "score": 1.0, + "content": "the initial iterate", + "type": "text" + }, + { + "bbox": [ + 177, + 589, + 195, + 600 + ], + "score": 0.9, + "content": "x ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 588, + 264, + 604 + ], + "score": 1.0, + "content": "and a meta-loss", + "type": "text" + }, + { + "bbox": [ + 265, + 590, + 273, + 600 + ], + "score": 0.76, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 588, + 419, + 604 + ], + "score": 1.0, + "content": ", which takes an objective function", + "type": "text" + }, + { + "bbox": [ + 419, + 590, + 426, + 602 + ], + "score": 0.84, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 588, + 506, + 604 + ], + "score": 1.0, + "content": "and a sequence of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 598, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 104, + 598, + 139, + 615 + ], + "score": 1.0, + "content": "iterates", + "type": "text" + }, + { + "bbox": [ + 139, + 601, + 197, + 613 + ], + "score": 0.92, + "content": "x ^ { ( 1 ) } , \\ldots , x ^ { ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 598, + 506, + 615 + ], + "score": 1.0, + "content": "produced by an optimization algorithm as input and outputs a scalar that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 450, + 626 + ], + "score": 1.0, + "content": "measures the quality of the iterates. The goal is to learn an optimization algorithm", + "type": "text" + }, + { + "bbox": [ + 450, + 613, + 464, + 623 + ], + "score": 0.86, + "content": "{ \\mathcal { P } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 622, + 507, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 241, + 638 + ], + "score": 0.85, + "content": "\\mathbb { E } _ { f \\sim \\mathcal { F } , x ^ { ( 0 ) } \\sim \\mathcal { D } } \\left[ \\mathcal { L } ( f , \\mathcal { P } ^ { * } ( f , x ^ { ( 0 ) } ) ) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 622, + 507, + 639 + ], + "score": 1.0, + "content": "is minimized. The meta-loss is chosen to penalize optimization", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "algorithms that exhibit behaviours we find undesirable, like slow convergence or excessive oscilla-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 101, + 646, + 510, + 684 + ], + "spans": [ + { + "bbox": [ + 101, + 646, + 308, + 684 + ], + "score": 1.0, + "content": "tions. Assuming we would like to learn an algoritha good choice of meta-loss would then simply be regret and can be interpreted as the area under the", + "type": "text" + }, + { + "bbox": [ + 308, + 658, + 363, + 672 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { i = 1 } ^ { T } f ( x ^ { ( i ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 646, + 510, + 684 + ], + "score": 1.0, + "content": "zes the objective function it is given,, which is equivalent to cumulativective values over time.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 39.5, + "bbox_fs": [ + 101, + 542, + 510, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 202, + 702 + ], + "score": 1.0, + "content": "The objective functions", + "type": "text" + }, + { + "bbox": [ + 202, + 688, + 245, + 699 + ], + "score": 0.92, + "content": "f _ { 1 } , \\ldots , f _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "may correspond to loss functions for training base-level learners,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "in which case the algorithm that learns the optimization algorithm can be viewed as a meta-learner.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "In this setting, each objective function is the loss function for training a particular base-learner on", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 734 + ], + "score": 1.0, + "content": "a particular task, and so the set of training objective functions can be loss functions for training a", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "base-learner or a family of base-learners on different tasks. At test time, the learned optimization", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "algorithm is evaluated on unseen objective functions, which correspond to loss functions for training", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 505, + 118 + ], + "score": 1.0, + "content": "base-learners on new tasks, which may be completely unrelated to tasks used for training the op-", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "timization algorithm. Therefore, the learned optimization algorithm must not learn anything about", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "the tasks used for training. Instead, the goal is to learn an optimization algorithm that can exploit the", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "geometric structure of the error surface induced by the base-learners. For example, if the base-level", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "model is a neural net with ReLU activation units, the optimization algorithm should hopefully learn", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "to leverage the piecewise linearity of the model. Hence, there is a clear division of responsibilities", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 182 + ], + "score": 1.0, + "content": "between the meta-learner and base-learners. The knowledge learned at the meta-level should be", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 193 + ], + "score": 1.0, + "content": "pertinent for all tasks, whereas the knowledge learned at the base-level should be task-specific. The", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 1.0, + "content": "meta-learner should therefore generalize across tasks, whereas the base-learner should generalize", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 176, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 176, + 215 + ], + "score": 1.0, + "content": "across instances.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 686, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "base-learner or a family of base-learners on different tasks. At test time, the learned optimization", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "algorithm is evaluated on unseen objective functions, which correspond to loss functions for training", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 505, + 118 + ], + "score": 1.0, + "content": "base-learners on new tasks, which may be completely unrelated to tasks used for training the op-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "timization algorithm. Therefore, the learned optimization algorithm must not learn anything about", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "the tasks used for training. Instead, the goal is to learn an optimization algorithm that can exploit the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "geometric structure of the error surface induced by the base-learners. For example, if the base-level", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "model is a neural net with ReLU activation units, the optimization algorithm should hopefully learn", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "to leverage the piecewise linearity of the model. Hence, there is a clear division of responsibilities", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 182 + ], + "score": 1.0, + "content": "between the meta-learner and base-learners. The knowledge learned at the meta-level should be", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 193 + ], + "score": 1.0, + "content": "pertinent for all tasks, whereas the knowledge learned at the base-level should be task-specific. The", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 1.0, + "content": "meta-learner should therefore generalize across tasks, whereas the base-learner should generalize", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 176, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 176, + 215 + ], + "score": 1.0, + "content": "across instances.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 108, + 227, + 216, + 239 + ], + "lines": [ + { + "bbox": [ + 106, + 227, + 217, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 217, + 240 + ], + "score": 1.0, + "content": "3.2 RL PRELIMINARIES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 248, + 505, + 350 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "The goal of reinforcement learning is to learn to interact with an environment in a way that min-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 272 + ], + "score": 1.0, + "content": "imizes cumulative costs that are expected to be incurred over time. The environment is formal-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 333, + 283 + ], + "score": 1.0, + "content": "ized as a partially observable Markov decision process", + "type": "text" + }, + { + "bbox": [ + 333, + 270, + 379, + 281 + ], + "score": 0.29, + "content": "( \\mathrm { P O M D P } ) ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 271, + 505, + 283 + ], + "score": 1.0, + "content": ", which is defined by the tuple", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 281, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 107, + 282, + 204, + 295 + ], + "score": 0.88, + "content": "( S , \\mathcal { O } , \\mathcal { A } , p _ { i } , p , p _ { o } , c , T )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 281, + 235, + 296 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 236, + 282, + 274, + 294 + ], + "score": 0.92, + "content": "\\mathcal { S } \\subseteq \\mathbb { R } ^ { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 281, + 354, + 296 + ], + "score": 1.0, + "content": "is the set of states,", + "type": "text" + }, + { + "bbox": [ + 355, + 281, + 396, + 294 + ], + "score": 0.93, + "content": "\\mathcal { O } \\subseteq \\mathbb { R } ^ { D ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 281, + 506, + 296 + ], + "score": 1.0, + "content": "is the set of observations,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 292, + 504, + 307 + ], + "spans": [ + { + "bbox": [ + 107, + 294, + 141, + 305 + ], + "score": 0.89, + "content": "\\mathcal { A } \\subseteq \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 292, + 225, + 307 + ], + "score": 1.0, + "content": "is the set of actions,", + "type": "text" + }, + { + "bbox": [ + 226, + 294, + 253, + 306 + ], + "score": 0.87, + "content": "p _ { i } \\left( s _ { 0 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 292, + 428, + 307 + ], + "score": 1.0, + "content": "is the probability density over initial states", + "type": "text" + }, + { + "bbox": [ + 429, + 295, + 439, + 305 + ], + "score": 0.56, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 292, + 443, + 307 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 443, + 294, + 504, + 306 + ], + "score": 0.89, + "content": "p \\left( { { s } _ { t + 1 } } \\left| { { s } _ { t } } , { { a } _ { t } } \\right. \\right)", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 304, + 507, + 319 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 317, + 319 + ], + "score": 1.0, + "content": "is the probability density over the subsequent state", + "type": "text" + }, + { + "bbox": [ + 318, + 307, + 337, + 317 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 304, + 434, + 319 + ], + "score": 1.0, + "content": "given the current state", + "type": "text" + }, + { + "bbox": [ + 434, + 307, + 443, + 316 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 304, + 491, + 319 + ], + "score": 1.0, + "content": "and action", + "type": "text" + }, + { + "bbox": [ + 491, + 308, + 501, + 316 + ], + "score": 0.8, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 304, + 507, + 319 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 315, + 507, + 330 + ], + "spans": [ + { + "bbox": [ + 107, + 316, + 149, + 327 + ], + "score": 0.9, + "content": "p _ { o } \\left( o _ { t } \\bar { | } _ { s _ { t } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 315, + 381, + 330 + ], + "score": 1.0, + "content": "is the probability density over the current observation", + "type": "text" + }, + { + "bbox": [ + 381, + 317, + 391, + 327 + ], + "score": 0.84, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 315, + 491, + 330 + ], + "score": 1.0, + "content": "given the current state", + "type": "text" + }, + { + "bbox": [ + 492, + 318, + 501, + 327 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 315, + 507, + 330 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 107, + 327, + 157, + 337 + ], + "score": 0.87, + "content": "c : { \\mathcal { S } } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 327, + 365, + 339 + ], + "score": 1.0, + "content": "is a function that assigns a cost to each state and", + "type": "text" + }, + { + "bbox": [ + 365, + 327, + 374, + 337 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "is the time horizon. Often, the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 338, + 440, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 190, + 350 + ], + "score": 1.0, + "content": "probability densities", + "type": "text" + }, + { + "bbox": [ + 190, + 339, + 197, + 349 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 338, + 214, + 350 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 214, + 339, + 225, + 349 + ], + "score": 0.85, + "content": "p _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 338, + 440, + 350 + ], + "score": 1.0, + "content": "are unknown and not given to the learning algorithm.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 142, + 367 + ], + "score": 1.0, + "content": "A policy", + "type": "text" + }, + { + "bbox": [ + 142, + 354, + 189, + 367 + ], + "score": 0.93, + "content": "\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { o } _ { t } , t \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 354, + 378, + 367 + ], + "score": 1.0, + "content": "is a conditional probability density over actions", + "type": "text" + }, + { + "bbox": [ + 378, + 357, + 388, + 366 + ], + "score": 0.85, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "given the current observation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 366, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 107, + 368, + 116, + 377 + ], + "score": 0.85, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 366, + 172, + 378 + ], + "score": 1.0, + "content": "and time step", + "type": "text" + }, + { + "bbox": [ + 173, + 368, + 178, + 376 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 366, + 313, + 378 + ], + "score": 1.0, + "content": ". When a policy is independent of", + "type": "text" + }, + { + "bbox": [ + 313, + 367, + 318, + 376 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 366, + 506, + 378 + ], + "score": 1.0, + "content": ", it is known as a stationary policy. The goal of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 336, + 389 + ], + "score": 1.0, + "content": "the reinforcement learning algorithm is to learn a policy", + "type": "text" + }, + { + "bbox": [ + 336, + 377, + 348, + 387 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "that minimizes the total expected cost", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 388, + 214, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 214, + 400 + ], + "score": 1.0, + "content": "over time. More precisely,", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 404, + 390, + 439 + ], + "lines": [ + { + "bbox": [ + 220, + 404, + 390, + 439 + ], + "spans": [ + { + "bbox": [ + 220, + 404, + 390, + 439 + ], + "score": 0.94, + "content": "\\pi ^ { * } = \\arg \\operatorname* { m i n } _ { \\pi } \\mathbb { E } _ { s _ { 0 } , a _ { 0 } , s _ { 1 } , \\dots , s _ { T } } \\left[ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) \\right] ,", + "type": "interline_equation", + "image_path": "b919d595be0d1a432a00d2fbbfe2eab638d4374ebd57fafd68cfd8d40797150b.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 220, + 404, + 390, + 421.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 220, + 421.5, + 390, + 439.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 503, + 465 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "where the expectation is taken with respect to the joint distribution over the sequence of states and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 453, + 355, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 355, + 466 + ], + "score": 1.0, + "content": "actions, often referred to as a trajectory, which has the density", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 469, + 422, + 533 + ], + "lines": [ + { + "bbox": [ + 189, + 469, + 422, + 533 + ], + "spans": [ + { + "bbox": [ + 189, + 469, + 422, + 533 + ], + "score": 0.93, + "content": "\\begin{array} { c } { { \\displaystyle q \\big ( s _ { 0 } , a _ { 0 } , s _ { 1 } , \\ldots , s _ { T } \\big ) = \\int _ { o _ { 0 } , \\ldots , o _ { T } } p _ { i } \\left( s _ { 0 } \\right) p _ { o } \\left( o _ { 0 } \\big | s _ { 0 } \\right) } } \\\\ { { { \\displaystyle \\prod _ { t = 0 } ^ { T - 1 } \\pi \\left( \\left. a _ { t } \\right| o _ { t } , t \\right) p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } \\right) p _ { o } \\left( \\left. o _ { t + 1 } \\right| s _ { t + 1 } \\right) . } } } \\end{array}", + "type": "interline_equation", + "image_path": "51df13704307a3794a4cd1301cf4a3226c41c612fa6a34d8eca973a93a49d9ae.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 469, + 422, + 485.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 189, + 485.0, + 422, + 501.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 189, + 501.0, + 422, + 517.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 189, + 517.0, + 422, + 533.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 542, + 503, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 222, + 554 + ], + "score": 1.0, + "content": "To make learning tractable,", + "type": "text" + }, + { + "bbox": [ + 223, + 545, + 231, + 552 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "is often constrained to lie in a parameterized family. A common", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 185, + 565 + ], + "score": 1.0, + "content": "assumption is that", + "type": "text" + }, + { + "bbox": [ + 185, + 553, + 330, + 565 + ], + "score": 0.92, + "content": "\\pi \\left( \\left. a _ { t } \\right| o _ { t } , t \\right) = \\mathcal { N } \\left( \\mu ^ { \\pi } ( o _ { t } ) , \\Sigma ^ { \\pi } ( o _ { t } ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 553, + 364, + 565 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 364, + 553, + 400, + 565 + ], + "score": 0.93, + "content": "\\textstyle { \\mathcal { N } } ( { \\boldsymbol { \\mu } } , { \\boldsymbol { \\Sigma } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "denotes the density of a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 193, + 577 + ], + "score": 1.0, + "content": "Gaussian with mean", + "type": "text" + }, + { + "bbox": [ + 193, + 567, + 200, + 576 + ], + "score": 0.8, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 564, + 267, + 577 + ], + "score": 1.0, + "content": "and covariance", + "type": "text" + }, + { + "bbox": [ + 267, + 565, + 275, + 574 + ], + "score": 0.77, + "content": "\\Sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 564, + 342, + 577 + ], + "score": 1.0, + "content": ". The functions", + "type": "text" + }, + { + "bbox": [ + 343, + 564, + 366, + 576 + ], + "score": 0.91, + "content": "\\mu ^ { \\pi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 564, + 423, + 577 + ], + "score": 1.0, + "content": "and possibly", + "type": "text" + }, + { + "bbox": [ + 423, + 564, + 447, + 576 + ], + "score": 0.91, + "content": "\\Sigma ^ { \\pi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "are modelled", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 575, + 352, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 352, + 587 + ], + "score": 1.0, + "content": "using function approximators, whose parameters are learned.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 107, + 599, + 195, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 197, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 197, + 613 + ], + "score": 1.0, + "content": "3.3 FORMULATION", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 619, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 203, + 634 + ], + "score": 1.0, + "content": "In our setting, the state", + "type": "text" + }, + { + "bbox": [ + 203, + 623, + 213, + 632 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 618, + 335, + 634 + ], + "score": 1.0, + "content": "consists of the current iterate", + "type": "text" + }, + { + "bbox": [ + 336, + 619, + 352, + 630 + ], + "score": 0.9, + "content": "x ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 618, + 406, + 634 + ], + "score": 1.0, + "content": "and features", + "type": "text" + }, + { + "bbox": [ + 406, + 620, + 425, + 632 + ], + "score": 0.9, + "content": "\\Phi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 618, + 506, + 634 + ], + "score": 1.0, + "content": "that depend on the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 630, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 104, + 630, + 182, + 648 + ], + "score": 1.0, + "content": "history of iterates", + "type": "text" + }, + { + "bbox": [ + 182, + 632, + 237, + 645 + ], + "score": 0.91, + "content": "x ^ { ( 1 ) } , \\ldots , x ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 630, + 315, + 648 + ], + "score": 1.0, + "content": ", (noisy) gradients", + "type": "text" + }, + { + "bbox": [ + 315, + 631, + 414, + 646 + ], + "score": 0.91, + "content": "\\nabla \\hat { f } ( x ^ { ( 1 ) } ) , \\ldots , \\nabla \\hat { f } ( x ^ { ( t ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 630, + 506, + 648 + ], + "score": 1.0, + "content": "and (noisy) objective", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 644, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 134, + 660 + ], + "score": 1.0, + "content": "values", + "type": "text" + }, + { + "bbox": [ + 135, + 645, + 217, + 659 + ], + "score": 0.92, + "content": "{ \\hat { f } } ( x ^ { ( 1 ) } ) , \\ldots , { \\hat { f } } ( x ^ { ( t ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 644, + 265, + 660 + ], + "score": 1.0, + "content": ". The action", + "type": "text" + }, + { + "bbox": [ + 266, + 648, + 276, + 658 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 644, + 319, + 660 + ], + "score": 1.0, + "content": "is the step", + "type": "text" + }, + { + "bbox": [ + 319, + 647, + 334, + 657 + ], + "score": 0.84, + "content": "\\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 644, + 506, + 660 + ], + "score": 1.0, + "content": "that will be used to update the iterate. The", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 656, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 155, + 672 + ], + "score": 1.0, + "content": "observation", + "type": "text" + }, + { + "bbox": [ + 155, + 660, + 165, + 670 + ], + "score": 0.82, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 656, + 204, + 672 + ], + "score": 1.0, + "content": "excludes", + "type": "text" + }, + { + "bbox": [ + 204, + 658, + 221, + 669 + ], + "score": 0.86, + "content": "x ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 656, + 320, + 672 + ], + "score": 1.0, + "content": "and consists of features", + "type": "text" + }, + { + "bbox": [ + 320, + 659, + 339, + 671 + ], + "score": 0.91, + "content": "\\Psi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 656, + 506, + 672 + ], + "score": 1.0, + "content": "that depend on the iterates, gradient and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "objective values from recent iterations, and the previous memory state of the learned optimization", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "algorithm, which takes the form of a recurrent neural net. This memory state can be viewed as a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 691, + 405, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 405, + 705 + ], + "score": 1.0, + "content": "statistic of the previous observations that is learned jointly with the policy.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 711, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "1What is described is an undiscounted finite-horizon POMDP with continuous state, observation and action", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 721, + 136, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 136, + 734 + ], + "score": 1.0, + "content": "spaces.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "4", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 214 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 82, + 505, + 215 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 227, + 216, + 239 + ], + "lines": [ + { + "bbox": [ + 106, + 227, + 217, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 217, + 240 + ], + "score": 1.0, + "content": "3.2 RL PRELIMINARIES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 248, + 505, + 350 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "The goal of reinforcement learning is to learn to interact with an environment in a way that min-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 272 + ], + "score": 1.0, + "content": "imizes cumulative costs that are expected to be incurred over time. The environment is formal-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 333, + 283 + ], + "score": 1.0, + "content": "ized as a partially observable Markov decision process", + "type": "text" + }, + { + "bbox": [ + 333, + 270, + 379, + 281 + ], + "score": 0.29, + "content": "( \\mathrm { P O M D P } ) ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 271, + 505, + 283 + ], + "score": 1.0, + "content": ", which is defined by the tuple", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 281, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 107, + 282, + 204, + 295 + ], + "score": 0.88, + "content": "( S , \\mathcal { O } , \\mathcal { A } , p _ { i } , p , p _ { o } , c , T )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 281, + 235, + 296 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 236, + 282, + 274, + 294 + ], + "score": 0.92, + "content": "\\mathcal { S } \\subseteq \\mathbb { R } ^ { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 281, + 354, + 296 + ], + "score": 1.0, + "content": "is the set of states,", + "type": "text" + }, + { + "bbox": [ + 355, + 281, + 396, + 294 + ], + "score": 0.93, + "content": "\\mathcal { O } \\subseteq \\mathbb { R } ^ { D ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 281, + 506, + 296 + ], + "score": 1.0, + "content": "is the set of observations,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 292, + 504, + 307 + ], + "spans": [ + { + "bbox": [ + 107, + 294, + 141, + 305 + ], + "score": 0.89, + "content": "\\mathcal { A } \\subseteq \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 292, + 225, + 307 + ], + "score": 1.0, + "content": "is the set of actions,", + "type": "text" + }, + { + "bbox": [ + 226, + 294, + 253, + 306 + ], + "score": 0.87, + "content": "p _ { i } \\left( s _ { 0 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 292, + 428, + 307 + ], + "score": 1.0, + "content": "is the probability density over initial states", + "type": "text" + }, + { + "bbox": [ + 429, + 295, + 439, + 305 + ], + "score": 0.56, + "content": "s _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 292, + 443, + 307 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 443, + 294, + 504, + 306 + ], + "score": 0.89, + "content": "p \\left( { { s } _ { t + 1 } } \\left| { { s } _ { t } } , { { a } _ { t } } \\right. \\right)", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 304, + 507, + 319 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 317, + 319 + ], + "score": 1.0, + "content": "is the probability density over the subsequent state", + "type": "text" + }, + { + "bbox": [ + 318, + 307, + 337, + 317 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 304, + 434, + 319 + ], + "score": 1.0, + "content": "given the current state", + "type": "text" + }, + { + "bbox": [ + 434, + 307, + 443, + 316 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 304, + 491, + 319 + ], + "score": 1.0, + "content": "and action", + "type": "text" + }, + { + "bbox": [ + 491, + 308, + 501, + 316 + ], + "score": 0.8, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 304, + 507, + 319 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 315, + 507, + 330 + ], + "spans": [ + { + "bbox": [ + 107, + 316, + 149, + 327 + ], + "score": 0.9, + "content": "p _ { o } \\left( o _ { t } \\bar { | } _ { s _ { t } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 315, + 381, + 330 + ], + "score": 1.0, + "content": "is the probability density over the current observation", + "type": "text" + }, + { + "bbox": [ + 381, + 317, + 391, + 327 + ], + "score": 0.84, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 315, + 491, + 330 + ], + "score": 1.0, + "content": "given the current state", + "type": "text" + }, + { + "bbox": [ + 492, + 318, + 501, + 327 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 315, + 507, + 330 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 107, + 327, + 157, + 337 + ], + "score": 0.87, + "content": "c : { \\mathcal { S } } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 327, + 365, + 339 + ], + "score": 1.0, + "content": "is a function that assigns a cost to each state and", + "type": "text" + }, + { + "bbox": [ + 365, + 327, + 374, + 337 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "is the time horizon. Often, the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 338, + 440, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 190, + 350 + ], + "score": 1.0, + "content": "probability densities", + "type": "text" + }, + { + "bbox": [ + 190, + 339, + 197, + 349 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 338, + 214, + 350 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 214, + 339, + 225, + 349 + ], + "score": 0.85, + "content": "p _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 338, + 440, + 350 + ], + "score": 1.0, + "content": "are unknown and not given to the learning algorithm.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 248, + 507, + 350 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 142, + 367 + ], + "score": 1.0, + "content": "A policy", + "type": "text" + }, + { + "bbox": [ + 142, + 354, + 189, + 367 + ], + "score": 0.93, + "content": "\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { o } _ { t } , t \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 354, + 378, + 367 + ], + "score": 1.0, + "content": "is a conditional probability density over actions", + "type": "text" + }, + { + "bbox": [ + 378, + 357, + 388, + 366 + ], + "score": 0.85, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "given the current observation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 366, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 107, + 368, + 116, + 377 + ], + "score": 0.85, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 366, + 172, + 378 + ], + "score": 1.0, + "content": "and time step", + "type": "text" + }, + { + "bbox": [ + 173, + 368, + 178, + 376 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 366, + 313, + 378 + ], + "score": 1.0, + "content": ". When a policy is independent of", + "type": "text" + }, + { + "bbox": [ + 313, + 367, + 318, + 376 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 366, + 506, + 378 + ], + "score": 1.0, + "content": ", it is known as a stationary policy. The goal of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 336, + 389 + ], + "score": 1.0, + "content": "the reinforcement learning algorithm is to learn a policy", + "type": "text" + }, + { + "bbox": [ + 336, + 377, + 348, + 387 + ], + "score": 0.86, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "that minimizes the total expected cost", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 388, + 214, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 214, + 400 + ], + "score": 1.0, + "content": "over time. More precisely,", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 354, + 506, + 400 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 404, + 390, + 439 + ], + "lines": [ + { + "bbox": [ + 220, + 404, + 390, + 439 + ], + "spans": [ + { + "bbox": [ + 220, + 404, + 390, + 439 + ], + "score": 0.94, + "content": "\\pi ^ { * } = \\arg \\operatorname* { m i n } _ { \\pi } \\mathbb { E } _ { s _ { 0 } , a _ { 0 } , s _ { 1 } , \\dots , s _ { T } } \\left[ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) \\right] ,", + "type": "interline_equation", + "image_path": "b919d595be0d1a432a00d2fbbfe2eab638d4374ebd57fafd68cfd8d40797150b.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 220, + 404, + 390, + 421.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 220, + 421.5, + 390, + 439.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 503, + 465 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "where the expectation is taken with respect to the joint distribution over the sequence of states and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 453, + 355, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 355, + 466 + ], + "score": 1.0, + "content": "actions, often referred to as a trajectory, which has the density", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 442, + 505, + 466 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 469, + 422, + 533 + ], + "lines": [ + { + "bbox": [ + 189, + 469, + 422, + 533 + ], + "spans": [ + { + "bbox": [ + 189, + 469, + 422, + 533 + ], + "score": 0.93, + "content": "\\begin{array} { c } { { \\displaystyle q \\big ( s _ { 0 } , a _ { 0 } , s _ { 1 } , \\ldots , s _ { T } \\big ) = \\int _ { o _ { 0 } , \\ldots , o _ { T } } p _ { i } \\left( s _ { 0 } \\right) p _ { o } \\left( o _ { 0 } \\big | s _ { 0 } \\right) } } \\\\ { { { \\displaystyle \\prod _ { t = 0 } ^ { T - 1 } \\pi \\left( \\left. a _ { t } \\right| o _ { t } , t \\right) p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } \\right) p _ { o } \\left( \\left. o _ { t + 1 } \\right| s _ { t + 1 } \\right) . } } } \\end{array}", + "type": "interline_equation", + "image_path": "51df13704307a3794a4cd1301cf4a3226c41c612fa6a34d8eca973a93a49d9ae.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 469, + 422, + 485.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 189, + 485.0, + 422, + 501.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 189, + 501.0, + 422, + 517.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 189, + 517.0, + 422, + 533.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 542, + 503, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 222, + 554 + ], + "score": 1.0, + "content": "To make learning tractable,", + "type": "text" + }, + { + "bbox": [ + 223, + 545, + 231, + 552 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "is often constrained to lie in a parameterized family. A common", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 185, + 565 + ], + "score": 1.0, + "content": "assumption is that", + "type": "text" + }, + { + "bbox": [ + 185, + 553, + 330, + 565 + ], + "score": 0.92, + "content": "\\pi \\left( \\left. a _ { t } \\right| o _ { t } , t \\right) = \\mathcal { N } \\left( \\mu ^ { \\pi } ( o _ { t } ) , \\Sigma ^ { \\pi } ( o _ { t } ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 553, + 364, + 565 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 364, + 553, + 400, + 565 + ], + "score": 0.93, + "content": "\\textstyle { \\mathcal { N } } ( { \\boldsymbol { \\mu } } , { \\boldsymbol { \\Sigma } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "denotes the density of a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 193, + 577 + ], + "score": 1.0, + "content": "Gaussian with mean", + "type": "text" + }, + { + "bbox": [ + 193, + 567, + 200, + 576 + ], + "score": 0.8, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 564, + 267, + 577 + ], + "score": 1.0, + "content": "and covariance", + "type": "text" + }, + { + "bbox": [ + 267, + 565, + 275, + 574 + ], + "score": 0.77, + "content": "\\Sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 564, + 342, + 577 + ], + "score": 1.0, + "content": ". The functions", + "type": "text" + }, + { + "bbox": [ + 343, + 564, + 366, + 576 + ], + "score": 0.91, + "content": "\\mu ^ { \\pi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 564, + 423, + 577 + ], + "score": 1.0, + "content": "and possibly", + "type": "text" + }, + { + "bbox": [ + 423, + 564, + 447, + 576 + ], + "score": 0.91, + "content": "\\Sigma ^ { \\pi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "are modelled", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 575, + 352, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 352, + 587 + ], + "score": 1.0, + "content": "using function approximators, whose parameters are learned.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 542, + 505, + 587 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 599, + 195, + 611 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 197, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 197, + 613 + ], + "score": 1.0, + "content": "3.3 FORMULATION", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 619, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 203, + 634 + ], + "score": 1.0, + "content": "In our setting, the state", + "type": "text" + }, + { + "bbox": [ + 203, + 623, + 213, + 632 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 618, + 335, + 634 + ], + "score": 1.0, + "content": "consists of the current iterate", + "type": "text" + }, + { + "bbox": [ + 336, + 619, + 352, + 630 + ], + "score": 0.9, + "content": "x ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 618, + 406, + 634 + ], + "score": 1.0, + "content": "and features", + "type": "text" + }, + { + "bbox": [ + 406, + 620, + 425, + 632 + ], + "score": 0.9, + "content": "\\Phi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 618, + 506, + 634 + ], + "score": 1.0, + "content": "that depend on the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 630, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 104, + 630, + 182, + 648 + ], + "score": 1.0, + "content": "history of iterates", + "type": "text" + }, + { + "bbox": [ + 182, + 632, + 237, + 645 + ], + "score": 0.91, + "content": "x ^ { ( 1 ) } , \\ldots , x ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 630, + 315, + 648 + ], + "score": 1.0, + "content": ", (noisy) gradients", + "type": "text" + }, + { + "bbox": [ + 315, + 631, + 414, + 646 + ], + "score": 0.91, + "content": "\\nabla \\hat { f } ( x ^ { ( 1 ) } ) , \\ldots , \\nabla \\hat { f } ( x ^ { ( t ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 630, + 506, + 648 + ], + "score": 1.0, + "content": "and (noisy) objective", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 644, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 134, + 660 + ], + "score": 1.0, + "content": "values", + "type": "text" + }, + { + "bbox": [ + 135, + 645, + 217, + 659 + ], + "score": 0.92, + "content": "{ \\hat { f } } ( x ^ { ( 1 ) } ) , \\ldots , { \\hat { f } } ( x ^ { ( t ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 644, + 265, + 660 + ], + "score": 1.0, + "content": ". The action", + "type": "text" + }, + { + "bbox": [ + 266, + 648, + 276, + 658 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 644, + 319, + 660 + ], + "score": 1.0, + "content": "is the step", + "type": "text" + }, + { + "bbox": [ + 319, + 647, + 334, + 657 + ], + "score": 0.84, + "content": "\\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 644, + 506, + 660 + ], + "score": 1.0, + "content": "that will be used to update the iterate. The", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 656, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 155, + 672 + ], + "score": 1.0, + "content": "observation", + "type": "text" + }, + { + "bbox": [ + 155, + 660, + 165, + 670 + ], + "score": 0.82, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 656, + 204, + 672 + ], + "score": 1.0, + "content": "excludes", + "type": "text" + }, + { + "bbox": [ + 204, + 658, + 221, + 669 + ], + "score": 0.86, + "content": "x ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 656, + 320, + 672 + ], + "score": 1.0, + "content": "and consists of features", + "type": "text" + }, + { + "bbox": [ + 320, + 659, + 339, + 671 + ], + "score": 0.91, + "content": "\\Psi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 656, + 506, + 672 + ], + "score": 1.0, + "content": "that depend on the iterates, gradient and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "objective values from recent iterations, and the previous memory state of the learned optimization", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "algorithm, which takes the form of a recurrent neural net. This memory state can be viewed as a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 691, + 405, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 405, + 705 + ], + "score": 1.0, + "content": "statistic of the previous observations that is learned jointly with the policy.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42, + "bbox_fs": [ + 104, + 618, + 506, + 705 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 155 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 327, + 95 + ], + "score": 1.0, + "content": "Under this formulation, the initial probability density", + "type": "text" + }, + { + "bbox": [ + 327, + 84, + 337, + 94 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "captures how the initial iterate, gradient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 412, + 106 + ], + "score": 1.0, + "content": "and objective value tend to be distributed. The transition probability density", + "type": "text" + }, + { + "bbox": [ + 412, + 96, + 419, + 105 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "captures the how the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "gradient and objective value are likely to change given the step that is taken currently; in other words,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "it encodes the local geometry of the training objective functions. Assuming the goal is to learn an", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 126, + 504, + 144 + ], + "spans": [ + { + "bbox": [ + 104, + 128, + 383, + 144 + ], + "score": 1.0, + "content": "optimization algorithm that minimizes the objective function, the cost", + "type": "text" + }, + { + "bbox": [ + 383, + 132, + 389, + 140 + ], + "score": 0.72, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 128, + 426, + 144 + ], + "score": 1.0, + "content": "of a state", + "type": "text" + }, + { + "bbox": [ + 427, + 126, + 504, + 143 + ], + "score": 0.93, + "content": "\\boldsymbol { s } _ { t } = \\left( \\boldsymbol { x } ^ { ( t ) } , \\Phi \\left( \\cdot \\right) \\right) ^ { T }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 141, + 273, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 240, + 155 + ], + "score": 1.0, + "content": "is simply the true objective value", + "type": "text" + }, + { + "bbox": [ + 241, + 141, + 270, + 155 + ], + "score": 0.93, + "content": "f ( \\boldsymbol { x } ^ { ( t ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 142, + 273, + 155 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 159, + 503, + 194 + ], + "lines": [ + { + "bbox": [ + 106, + 159, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 197, + 173 + ], + "score": 1.0, + "content": "Any particular policy", + "type": "text" + }, + { + "bbox": [ + 198, + 159, + 244, + 172 + ], + "score": 0.93, + "content": "\\pi \\left( \\boldsymbol { a } _ { t } \\left| \\boldsymbol { o } _ { t } , t \\right. \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 159, + 319, + 173 + ], + "score": 1.0, + "content": ", which generates", + "type": "text" + }, + { + "bbox": [ + 319, + 160, + 361, + 171 + ], + "score": 0.91, + "content": "a _ { t } ~ = ~ \\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 159, + 505, + 173 + ], + "score": 1.0, + "content": "at every time step, corresponds to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 504, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 254, + 183 + ], + "score": 1.0, + "content": "a particular (noisy) update formula", + "type": "text" + }, + { + "bbox": [ + 254, + 172, + 262, + 182 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 171, + 504, + 183 + ], + "score": 1.0, + "content": ", and therefore a particular (noisy) optimization algorithm.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 181, + 496, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 496, + 195 + ], + "score": 1.0, + "content": "Therefore, learning an optimization algorithm simply reduces to searching for the optimal policy.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 198, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "The mean of the policy is modelled as a recurrent neural net fragment that corresponds to a single", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 301, + 222 + ], + "score": 1.0, + "content": "time step, which takes the observation features", + "type": "text" + }, + { + "bbox": [ + 302, + 210, + 321, + 222 + ], + "score": 0.92, + "content": "\\Psi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "and the previous memory state as input and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 203, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 203, + 233 + ], + "score": 1.0, + "content": "outputs the step to take.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 107, + 246, + 240, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 242, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 242, + 259 + ], + "score": 1.0, + "content": "3.4 GUIDED POLICY SEARCH", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 506, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "The reinforcement learning method we use is guided policy search (GPS) (Levine et al., 2015),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 278, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 506, + 292 + ], + "score": 1.0, + "content": "which is a policy search method designed for searching over large classes of expressive non-linear", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 396, + 302 + ], + "score": 1.0, + "content": "policies in continuous state and action spaces. It maintains two policies,", + "type": "text" + }, + { + "bbox": [ + 396, + 290, + 405, + 301 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 290, + 423, + 302 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 423, + 291, + 430, + 300 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 290, + 505, + 302 + ], + "score": 1.0, + "content": ", where the former", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 302, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 313 + ], + "score": 1.0, + "content": "lies in a time-varying linear policy class in which the optimal policy can found in closed form, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "the latter lies in a stationary non-linear policy class in which policy optimization is challenging. In", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 306, + 336 + ], + "score": 1.0, + "content": "each iteration, it performs policy optimization on", + "type": "text" + }, + { + "bbox": [ + 306, + 323, + 314, + 334 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 321, + 506, + 336 + ], + "score": 1.0, + "content": ", and uses the resulting policy as supervision to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 333, + 140, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 127, + 346 + ], + "score": 1.0, + "content": "train", + "type": "text" + }, + { + "bbox": [ + 127, + 336, + 135, + 343 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 333, + 140, + 346 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 414, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 415, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 415, + 365 + ], + "score": 1.0, + "content": "More precisely, GPS solves the following constrained optimization problem:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 366, + 426, + 398 + ], + "lines": [ + { + "bbox": [ + 185, + 366, + 426, + 398 + ], + "spans": [ + { + "bbox": [ + 185, + 366, + 426, + 398 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { \\theta , \\eta } \\mathbb { E } _ { \\psi } \\left[ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) \\right] \\mathrm { ~ s . t . ~ } \\psi \\left( \\left. a _ { t } \\right| s _ { t } , t ; \\eta \\right) = \\pi \\left( \\left. a _ { t } \\right| s _ { t } ; \\theta \\right) \\forall a _ { t } , s _ { t } , t", + "type": "interline_equation", + "image_path": "f5be126419b4a29a54757648e664beaddb30c079479aa1e9291425a258d92b7d.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 366, + 426, + 382.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 185, + 382.0, + 426, + 398.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 404, + 504, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 133, + 418 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 406, + 141, + 416 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 403, + 159, + 418 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 160, + 405, + 166, + 414 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 403, + 270, + 418 + ], + "score": 1.0, + "content": "denote the parameters of", + "type": "text" + }, + { + "bbox": [ + 270, + 405, + 279, + 416 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 403, + 297, + 418 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 298, + 406, + 305, + 414 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 403, + 359, + 418 + ], + "score": 1.0, + "content": "respectively,", + "type": "text" + }, + { + "bbox": [ + 359, + 404, + 382, + 416 + ], + "score": 0.9, + "content": "\\mathbb { E } _ { \\rho } \\left[ \\cdot \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 403, + 506, + 418 + ], + "score": 1.0, + "content": "denotes the expectation taken", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 415, + 504, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 303, + 430 + ], + "score": 1.0, + "content": "with respect to the trajectory induced by a policy", + "type": "text" + }, + { + "bbox": [ + 303, + 419, + 310, + 428 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 415, + 327, + 430 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 416, + 501, + 430 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\pi \\left( \\left. a _ { t } \\right| s _ { t } ; \\theta \\right) : = \\int _ { o _ { t } } \\pi \\left( \\left. a _ { t } \\right| o _ { t } ; \\theta \\right) p _ { o } \\left( \\left. o _ { t } \\right| s _ { t } \\right) ^ { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 415, + 504, + 430 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 504, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "Since there are an infinite number of equality constraints, the problem is relaxed by enforcing equal-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 445, + 459, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 237, + 460 + ], + "score": 1.0, + "content": "ity on the mean actions taken by", + "type": "text" + }, + { + "bbox": [ + 238, + 447, + 246, + 458 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 445, + 264, + 460 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 264, + 448, + 271, + 456 + ], + "score": 0.81, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 445, + 459, + 460 + ], + "score": 1.0, + "content": "at every time step3. So, the problem becomes:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 464, + 407, + 496 + ], + "lines": [ + { + "bbox": [ + 204, + 464, + 407, + 496 + ], + "spans": [ + { + "bbox": [ + 204, + 464, + 407, + 496 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { \\theta , \\eta } \\mathbb { E } _ { \\psi } [ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) ] \\mathrm { ~ s . t . ~ } \\mathbb { E } _ { \\psi } [ a _ { t } ] = \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] \\forall t", + "type": "interline_equation", + "image_path": "407d95c56a995ddd518677a9f9c837a0fe2a62d6a02d1ebce4c22c60b61b0cfe.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 464, + 407, + 480.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 204, + 480.0, + 407, + 496.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 504, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "This problem is solved using Bregman ADMM (Wang & Banerjee, 2014), which performs the fol-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 517, + 239, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 239, + 529 + ], + "score": 1.0, + "content": "lowing updates in each iteration:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 104, + 533, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 122, + 533, + 485, + 642 + ], + "spans": [ + { + "bbox": [ + 122, + 533, + 485, + 642 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\eta \\gets \\arg \\underset { \\eta } { \\operatorname* { m i n } } \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\psi } [ c ( s _ { t } ) - \\lambda _ { t } ^ { T } a _ { t } ] + \\nu _ { t } D _ { t } ( \\eta , \\theta ) } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad } \\\\ & { \\theta \\gets \\arg \\underset { \\theta } { \\operatorname* { m i n } } \\sum _ { t = 0 } ^ { T } \\lambda _ { t } ^ { T } \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] + \\nu _ { t } D _ { t } ( \\theta , \\eta ) } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\lambda _ { t } \\gets \\lambda _ { t } + \\alpha \\nu _ { t } ( \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] - \\mathbb { E } _ { \\psi } [ a _ { t } ] ) \\ \\forall t , } \\\\ & { \\mathrm { ~ } \\mathrm { ~ } : \\mathrm { r e } \\quad D _ { t } ( \\theta , \\eta ) \\quad : = \\quad \\mathbb { E } _ { \\psi } [ D _ { K L } ( \\pi ( a _ { t } | s _ { t } ; \\theta ) ) ] \\ \\psi ( a _ { t } | s _ { t } , t ; \\eta ) ) ] \\quad \\mathrm { ~ a n d ~ } \\quad D _ { t } ( \\eta , \\theta ) } \\\\ & { [ D _ { K L } ( \\psi ( a _ { t } | s _ { t } , t ; \\eta ) ] ) | \\ \\pi ( a _ { t } | s _ { t } ; \\theta ) ) | . } \\end{array}", + "type": "interline_equation", + "image_path": "98b4e8aee640645dca59af14914cf556ee40ccd35b91af03094a1e5f58cc3027.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 104, + 533, + 505, + 569.3333333333334 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 104, + 569.3333333333334, + 505, + 605.6666666666667 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 104, + 605.6666666666667, + 505, + 642.0000000000001 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 104, + 645, + 507, + 661 + ], + "spans": [ + { + "bbox": [ + 104, + 645, + 220, + 661 + ], + "score": 1.0, + "content": "The algorithm assumes that", + "type": "text" + }, + { + "bbox": [ + 221, + 646, + 370, + 659 + ], + "score": 0.91, + "content": "\\psi \\left( \\left. a _ { t } \\right| s _ { t } , t ; \\eta \\right) = \\mathcal { N } \\left( K _ { t } s _ { t } + k _ { t } , G _ { t } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 645, + 401, + 661 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 401, + 645, + 486, + 660 + ], + "score": 0.92, + "content": "\\boldsymbol { \\eta } : = \\left( K _ { t } , k _ { t } , G _ { t } \\right) _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 645, + 507, + 661 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 657, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 234, + 670 + ], + "score": 0.91, + "content": "\\pi \\left( a _ { t } | o _ { t } ; \\theta \\right) = \\mathcal { N } \\left( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) , \\Sigma ^ { \\pi } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 657, + 266, + 671 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 266, + 659, + 319, + 670 + ], + "score": 0.92, + "content": "\\boldsymbol { \\theta } : = ( \\omega , \\Sigma ^ { \\pi } )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 657, + 338, + 671 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 339, + 659, + 362, + 670 + ], + "score": 0.89, + "content": "\\mu _ { \\omega } ^ { \\pi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 657, + 506, + 671 + ], + "score": 1.0, + "content": "can be an arbitrary function that is", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 668, + 416, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 416, + 681 + ], + "score": 1.0, + "content": "typically modelled using a nonlinear function approximator like a neural net.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 690, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 689, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 118, + 689, + 335, + 704 + ], + "score": 1.0, + "content": "2In practice, the explicit form of the observation probability", + "type": "text" + }, + { + "bbox": [ + 336, + 693, + 346, + 701 + ], + "score": 0.85, + "content": "p _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 689, + 506, + 704 + ], + "score": 1.0, + "content": "is usually not known or the integral may be", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 379, + 713 + ], + "score": 1.0, + "content": "intractable to compute. So, a linear Gaussian model is fitted to samples of", + "type": "text" + }, + { + "bbox": [ + 379, + 703, + 388, + 710 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 700, + 405, + 713 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 702, + 414, + 710 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "and used in place of the", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 710, + 233, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 122, + 723 + ], + "score": 1.0, + "content": "true", + "type": "text" + }, + { + "bbox": [ + 123, + 711, + 168, + 721 + ], + "score": 0.91, + "content": "\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ; \\boldsymbol { \\theta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 710, + 233, + 723 + ], + "score": 1.0, + "content": "where necessary.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 719, + 479, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 466, + 734 + ], + "score": 1.0, + "content": "3Though the Bregman divergence penalty is applied to the original probability distributions over", + "type": "text" + }, + { + "bbox": [ + 467, + 723, + 475, + 731 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 719, + 479, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 155 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 327, + 95 + ], + "score": 1.0, + "content": "Under this formulation, the initial probability density", + "type": "text" + }, + { + "bbox": [ + 327, + 84, + 337, + 94 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "captures how the initial iterate, gradient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 412, + 106 + ], + "score": 1.0, + "content": "and objective value tend to be distributed. The transition probability density", + "type": "text" + }, + { + "bbox": [ + 412, + 96, + 419, + 105 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "captures the how the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "gradient and objective value are likely to change given the step that is taken currently; in other words,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "it encodes the local geometry of the training objective functions. Assuming the goal is to learn an", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 126, + 504, + 144 + ], + "spans": [ + { + "bbox": [ + 104, + 128, + 383, + 144 + ], + "score": 1.0, + "content": "optimization algorithm that minimizes the objective function, the cost", + "type": "text" + }, + { + "bbox": [ + 383, + 132, + 389, + 140 + ], + "score": 0.72, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 128, + 426, + 144 + ], + "score": 1.0, + "content": "of a state", + "type": "text" + }, + { + "bbox": [ + 427, + 126, + 504, + 143 + ], + "score": 0.93, + "content": "\\boldsymbol { s } _ { t } = \\left( \\boldsymbol { x } ^ { ( t ) } , \\Phi \\left( \\cdot \\right) \\right) ^ { T }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 141, + 273, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 240, + 155 + ], + "score": 1.0, + "content": "is simply the true objective value", + "type": "text" + }, + { + "bbox": [ + 241, + 141, + 270, + 155 + ], + "score": 0.93, + "content": "f ( \\boldsymbol { x } ^ { ( t ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 142, + 273, + 155 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 82, + 506, + 155 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 159, + 503, + 194 + ], + "lines": [ + { + "bbox": [ + 106, + 159, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 197, + 173 + ], + "score": 1.0, + "content": "Any particular policy", + "type": "text" + }, + { + "bbox": [ + 198, + 159, + 244, + 172 + ], + "score": 0.93, + "content": "\\pi \\left( \\boldsymbol { a } _ { t } \\left| \\boldsymbol { o } _ { t } , t \\right. \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 159, + 319, + 173 + ], + "score": 1.0, + "content": ", which generates", + "type": "text" + }, + { + "bbox": [ + 319, + 160, + 361, + 171 + ], + "score": 0.91, + "content": "a _ { t } ~ = ~ \\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 159, + 505, + 173 + ], + "score": 1.0, + "content": "at every time step, corresponds to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 504, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 254, + 183 + ], + "score": 1.0, + "content": "a particular (noisy) update formula", + "type": "text" + }, + { + "bbox": [ + 254, + 172, + 262, + 182 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 171, + 504, + 183 + ], + "score": 1.0, + "content": ", and therefore a particular (noisy) optimization algorithm.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 181, + 496, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 496, + 195 + ], + "score": 1.0, + "content": "Therefore, learning an optimization algorithm simply reduces to searching for the optimal policy.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 159, + 505, + 195 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 198, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "The mean of the policy is modelled as a recurrent neural net fragment that corresponds to a single", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 301, + 222 + ], + "score": 1.0, + "content": "time step, which takes the observation features", + "type": "text" + }, + { + "bbox": [ + 302, + 210, + 321, + 222 + ], + "score": 0.92, + "content": "\\Psi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "and the previous memory state as input and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 203, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 203, + 233 + ], + "score": 1.0, + "content": "outputs the step to take.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 198, + 505, + 233 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 246, + 240, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 245, + 242, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 242, + 259 + ], + "score": 1.0, + "content": "3.4 GUIDED POLICY SEARCH", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 506, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "The reinforcement learning method we use is guided policy search (GPS) (Levine et al., 2015),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 278, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 506, + 292 + ], + "score": 1.0, + "content": "which is a policy search method designed for searching over large classes of expressive non-linear", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 396, + 302 + ], + "score": 1.0, + "content": "policies in continuous state and action spaces. It maintains two policies,", + "type": "text" + }, + { + "bbox": [ + 396, + 290, + 405, + 301 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 290, + 423, + 302 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 423, + 291, + 430, + 300 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 290, + 505, + 302 + ], + "score": 1.0, + "content": ", where the former", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 302, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 313 + ], + "score": 1.0, + "content": "lies in a time-varying linear policy class in which the optimal policy can found in closed form, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "the latter lies in a stationary non-linear policy class in which policy optimization is challenging. In", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 306, + 336 + ], + "score": 1.0, + "content": "each iteration, it performs policy optimization on", + "type": "text" + }, + { + "bbox": [ + 306, + 323, + 314, + 334 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 321, + 506, + 336 + ], + "score": 1.0, + "content": ", and uses the resulting policy as supervision to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 333, + 140, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 127, + 346 + ], + "score": 1.0, + "content": "train", + "type": "text" + }, + { + "bbox": [ + 127, + 336, + 135, + 343 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 333, + 140, + 346 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 267, + 506, + 346 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 414, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 415, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 415, + 365 + ], + "score": 1.0, + "content": "More precisely, GPS solves the following constrained optimization problem:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 347, + 415, + 365 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 366, + 426, + 398 + ], + "lines": [ + { + "bbox": [ + 185, + 366, + 426, + 398 + ], + "spans": [ + { + "bbox": [ + 185, + 366, + 426, + 398 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { \\theta , \\eta } \\mathbb { E } _ { \\psi } \\left[ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) \\right] \\mathrm { ~ s . t . ~ } \\psi \\left( \\left. a _ { t } \\right| s _ { t } , t ; \\eta \\right) = \\pi \\left( \\left. a _ { t } \\right| s _ { t } ; \\theta \\right) \\forall a _ { t } , s _ { t } , t", + "type": "interline_equation", + "image_path": "f5be126419b4a29a54757648e664beaddb30c079479aa1e9291425a258d92b7d.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 366, + 426, + 382.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 185, + 382.0, + 426, + 398.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 404, + 504, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 133, + 418 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 406, + 141, + 416 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 403, + 159, + 418 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 160, + 405, + 166, + 414 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 403, + 270, + 418 + ], + "score": 1.0, + "content": "denote the parameters of", + "type": "text" + }, + { + "bbox": [ + 270, + 405, + 279, + 416 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 403, + 297, + 418 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 298, + 406, + 305, + 414 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 403, + 359, + 418 + ], + "score": 1.0, + "content": "respectively,", + "type": "text" + }, + { + "bbox": [ + 359, + 404, + 382, + 416 + ], + "score": 0.9, + "content": "\\mathbb { E } _ { \\rho } \\left[ \\cdot \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 403, + 506, + 418 + ], + "score": 1.0, + "content": "denotes the expectation taken", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 415, + 504, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 303, + 430 + ], + "score": 1.0, + "content": "with respect to the trajectory induced by a policy", + "type": "text" + }, + { + "bbox": [ + 303, + 419, + 310, + 428 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 415, + 327, + 430 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 416, + 501, + 430 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\pi \\left( \\left. a _ { t } \\right| s _ { t } ; \\theta \\right) : = \\int _ { o _ { t } } \\pi \\left( \\left. a _ { t } \\right| o _ { t } ; \\theta \\right) p _ { o } \\left( \\left. o _ { t } \\right| s _ { t } \\right) ^ { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 415, + 504, + 430 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 403, + 506, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 504, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "Since there are an infinite number of equality constraints, the problem is relaxed by enforcing equal-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 445, + 459, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 237, + 460 + ], + "score": 1.0, + "content": "ity on the mean actions taken by", + "type": "text" + }, + { + "bbox": [ + 238, + 447, + 246, + 458 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 445, + 264, + 460 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 264, + 448, + 271, + 456 + ], + "score": 0.81, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 445, + 459, + 460 + ], + "score": 1.0, + "content": "at every time step3. So, the problem becomes:", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 435, + 505, + 460 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 464, + 407, + 496 + ], + "lines": [ + { + "bbox": [ + 204, + 464, + 407, + 496 + ], + "spans": [ + { + "bbox": [ + 204, + 464, + 407, + 496 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { \\theta , \\eta } \\mathbb { E } _ { \\psi } [ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) ] \\mathrm { ~ s . t . ~ } \\mathbb { E } _ { \\psi } [ a _ { t } ] = \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] \\forall t", + "type": "interline_equation", + "image_path": "407d95c56a995ddd518677a9f9c837a0fe2a62d6a02d1ebce4c22c60b61b0cfe.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 464, + 407, + 480.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 204, + 480.0, + 407, + 496.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 504, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "This problem is solved using Bregman ADMM (Wang & Banerjee, 2014), which performs the fol-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 517, + 239, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 239, + 529 + ], + "score": 1.0, + "content": "lowing updates in each iteration:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 506, + 505, + 529 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 104, + 533, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 122, + 533, + 485, + 642 + ], + "spans": [ + { + "bbox": [ + 122, + 533, + 485, + 642 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\eta \\gets \\arg \\underset { \\eta } { \\operatorname* { m i n } } \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\psi } [ c ( s _ { t } ) - \\lambda _ { t } ^ { T } a _ { t } ] + \\nu _ { t } D _ { t } ( \\eta , \\theta ) } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad } \\\\ & { \\theta \\gets \\arg \\underset { \\theta } { \\operatorname* { m i n } } \\sum _ { t = 0 } ^ { T } \\lambda _ { t } ^ { T } \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] + \\nu _ { t } D _ { t } ( \\theta , \\eta ) } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\lambda _ { t } \\gets \\lambda _ { t } + \\alpha \\nu _ { t } ( \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] - \\mathbb { E } _ { \\psi } [ a _ { t } ] ) \\ \\forall t , } \\\\ & { \\mathrm { ~ } \\mathrm { ~ } : \\mathrm { r e } \\quad D _ { t } ( \\theta , \\eta ) \\quad : = \\quad \\mathbb { E } _ { \\psi } [ D _ { K L } ( \\pi ( a _ { t } | s _ { t } ; \\theta ) ) ] \\ \\psi ( a _ { t } | s _ { t } , t ; \\eta ) ) ] \\quad \\mathrm { ~ a n d ~ } \\quad D _ { t } ( \\eta , \\theta ) } \\\\ & { [ D _ { K L } ( \\psi ( a _ { t } | s _ { t } , t ; \\eta ) ] ) | \\ \\pi ( a _ { t } | s _ { t } ; \\theta ) ) | . } \\end{array}", + "type": "interline_equation", + "image_path": "98b4e8aee640645dca59af14914cf556ee40ccd35b91af03094a1e5f58cc3027.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 104, + 533, + 505, + 569.3333333333334 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 104, + 569.3333333333334, + 505, + 605.6666666666667 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 104, + 605.6666666666667, + 505, + 642.0000000000001 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 646, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 104, + 645, + 507, + 661 + ], + "spans": [ + { + "bbox": [ + 104, + 645, + 220, + 661 + ], + "score": 1.0, + "content": "The algorithm assumes that", + "type": "text" + }, + { + "bbox": [ + 221, + 646, + 370, + 659 + ], + "score": 0.91, + "content": "\\psi \\left( \\left. a _ { t } \\right| s _ { t } , t ; \\eta \\right) = \\mathcal { N } \\left( K _ { t } s _ { t } + k _ { t } , G _ { t } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 645, + 401, + 661 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 401, + 645, + 486, + 660 + ], + "score": 0.92, + "content": "\\boldsymbol { \\eta } : = \\left( K _ { t } , k _ { t } , G _ { t } \\right) _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 645, + 507, + 661 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 657, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 234, + 670 + ], + "score": 0.91, + "content": "\\pi \\left( a _ { t } | o _ { t } ; \\theta \\right) = \\mathcal { N } \\left( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) , \\Sigma ^ { \\pi } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 657, + 266, + 671 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 266, + 659, + 319, + 670 + ], + "score": 0.92, + "content": "\\boldsymbol { \\theta } : = ( \\omega , \\Sigma ^ { \\pi } )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 657, + 338, + 671 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 339, + 659, + 362, + 670 + ], + "score": 0.89, + "content": "\\mu _ { \\omega } ^ { \\pi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 657, + 506, + 671 + ], + "score": 1.0, + "content": "can be an arbitrary function that is", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 668, + 416, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 416, + 681 + ], + "score": 1.0, + "content": "typically modelled using a nonlinear function approximator like a neural net.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 104, + 645, + 507, + 681 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "At the start of each iteration, the algorithm constructs a model of the transition probability density", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 90, + 508, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 300, + 108 + ], + "score": 0.91, + "content": "\\tilde { p } \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t ; \\zeta \\right) = \\mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 90, + 331, + 113 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 331, + 93, + 427, + 108 + ], + "score": 0.93, + "content": "\\boldsymbol { \\zeta } : = \\left( A _ { t } , B _ { t } , c _ { t } , F _ { t } \\right) _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 90, + 508, + 113 + ], + "score": 1.0, + "content": "is fitted to samples", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 117, + 120 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 109, + 127, + 118 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 105, + 281, + 120 + ], + "score": 1.0, + "content": "drawn from the trajectory induced by", + "type": "text" + }, + { + "bbox": [ + 281, + 107, + 289, + 118 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 105, + 506, + 120 + ], + "score": 1.0, + "content": ", which essentially amounts to a local linearization of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 225, + 131 + ], + "score": 1.0, + "content": "the true transition probability", + "type": "text" + }, + { + "bbox": [ + 226, + 118, + 294, + 129 + ], + "score": 0.91, + "content": "p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 117, + 348, + 131 + ], + "score": 1.0, + "content": ". We will use", + "type": "text" + }, + { + "bbox": [ + 348, + 117, + 372, + 131 + ], + "score": 0.92, + "content": "\\mathbb { E } _ { \\widetilde { \\psi } } \\left[ \\cdot \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 117, + 506, + 131 + ], + "score": 1.0, + "content": "to denote expectation taken with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 129, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 104, + 129, + 249, + 142 + ], + "score": 1.0, + "content": "respect to the trajectory induced by", + "type": "text" + }, + { + "bbox": [ + 249, + 131, + 258, + 141 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 129, + 424, + 142 + ], + "score": 1.0, + "content": "under the modelled transition probability", + "type": "text" + }, + { + "bbox": [ + 425, + 130, + 431, + 141 + ], + "score": 0.8, + "content": "\\tilde { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 129, + 506, + 142 + ], + "score": 1.0, + "content": ". Additionally, the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 139, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 295, + 155 + ], + "score": 1.0, + "content": "algorithm fits local quadratic approximations to", + "type": "text" + }, + { + "bbox": [ + 296, + 141, + 317, + 152 + ], + "score": 0.86, + "content": "c ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 139, + 392, + 155 + ], + "score": 1.0, + "content": "around samples of", + "type": "text" + }, + { + "bbox": [ + 392, + 142, + 401, + 152 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 139, + 505, + 155 + ], + "score": 1.0, + "content": "drawn from the trajectory", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 151, + 487, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 153, + 165 + ], + "score": 1.0, + "content": "induced by", + "type": "text" + }, + { + "bbox": [ + 153, + 152, + 161, + 163 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 151, + 191, + 165 + ], + "score": 1.0, + "content": "so that", + "type": "text" + }, + { + "bbox": [ + 192, + 151, + 351, + 165 + ], + "score": 0.93, + "content": "\\begin{array} { r } { c ( s _ { t } ) \\approx \\dot { \\tilde { c } } \\dot { ( s _ { t } ) } : = \\frac { 1 } { 2 } s _ { t } ^ { T } C _ { t } s _ { t } + \\dot { d _ { t } ^ { T } } s _ { t } + h _ { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 151, + 366, + 165 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 366, + 153, + 375, + 163 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 151, + 487, + 165 + ], + "score": 1.0, + "content": "’s that are near the samples.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 171, + 502, + 194 + ], + "lines": [ + { + "bbox": [ + 104, + 168, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 104, + 168, + 417, + 187 + ], + "score": 1.0, + "content": "With these assumptions, the subproblem that needs to be solved to update", + "type": "text" + }, + { + "bbox": [ + 417, + 169, + 504, + 184 + ], + "score": 0.92, + "content": "\\eta ~ = ~ ( K _ { t } , k _ { t } , G _ { t } ) _ { t = 1 } ^ { T }", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 181, + 147, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 147, + 194 + ], + "score": 1.0, + "content": "becomes:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 197, + 197, + 414, + 261 + ], + "lines": [ + { + "bbox": [ + 197, + 197, + 414, + 261 + ], + "spans": [ + { + "bbox": [ + 197, + 197, + 414, + 261 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } & { \\underset { \\eta } { \\operatorname* { m i n } } \\displaystyle \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\boldsymbol { \\tilde { \\psi } } } [ \\boldsymbol { \\tilde { c } } ( s _ { t } ) - \\lambda _ { t } ^ { T } \\boldsymbol { a } _ { t } ] + \\nu _ { t } D _ { t } ( \\eta , \\theta ) } & \\\\ & { \\mathrm { s . t . } \\displaystyle \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\boldsymbol { \\tilde { \\psi } } } [ D _ { K L } ( \\boldsymbol { \\psi } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\eta ) | \\Big | \\psi ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\eta ^ { \\prime } ) ) ] \\leq \\epsilon , } \\end{array}", + "type": "interline_equation", + "image_path": "da17dbd224435326b9983c40ce40ffd1486ed56e1c09a13c8b435c27ba5a8491.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 197, + 197, + 414, + 213.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 197, + 213.0, + 414, + 229.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 197, + 229.0, + 414, + 245.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 197, + 245.0, + 414, + 261.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 506, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 133, + 278 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 266, + 142, + 277 + ], + "score": 0.87, + "content": "\\eta ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 266, + 205, + 278 + ], + "score": 1.0, + "content": "denotes the old", + "type": "text" + }, + { + "bbox": [ + 205, + 267, + 212, + 277 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 266, + 358, + 278 + ], + "score": 1.0, + "content": "from the previous iteration. Because", + "type": "text" + }, + { + "bbox": [ + 359, + 267, + 366, + 277 + ], + "score": 0.85, + "content": "\\tilde { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 266, + 383, + 278 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 383, + 267, + 389, + 276 + ], + "score": 0.65, + "content": "\\tilde { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "are only valid locally around", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 210, + 289 + ], + "score": 1.0, + "content": "the trajectory induced by", + "type": "text" + }, + { + "bbox": [ + 211, + 278, + 219, + 288 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 276, + 439, + 289 + ], + "score": 1.0, + "content": ", the constraint is added to limit the amount by which", + "type": "text" + }, + { + "bbox": [ + 439, + 279, + 446, + 289 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "is updated. It", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 287, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 301 + ], + "score": 1.0, + "content": "turns out that the unconstrained problem can be solved in closed form using a dynamic programming", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 299, + 504, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 495, + 311 + ], + "score": 1.0, + "content": "algorithm known as linear-quadratic-Gaussian (LQG) regulator in time linear in the time horizon", + "type": "text" + }, + { + "bbox": [ + 496, + 300, + 504, + 309 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 306, + 321 + ], + "score": 1.0, + "content": "and cubic in the dimensionality of the state space", + "type": "text" + }, + { + "bbox": [ + 307, + 311, + 316, + 320 + ], + "score": 0.8, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 309, + 505, + 321 + ], + "score": 1.0, + "content": ". The constrained problem is solved using dual", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "gradient descent, which uses LQG as a subroutine to solve for the primal variables in each iteration", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 332, + 383, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 383, + 343 + ], + "score": 1.0, + "content": "and increments the dual variable on the constraint until it is satisfied.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 504, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 146, + 361 + ], + "score": 1.0, + "content": "Updating", + "type": "text" + }, + { + "bbox": [ + 146, + 349, + 153, + 359 + ], + "score": 0.75, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 349, + 496, + 361 + ], + "score": 1.0, + "content": "is straightforward, since expectations taken with respect to the trajectory induced by", + "type": "text" + }, + { + "bbox": [ + 497, + 351, + 504, + 358 + ], + "score": 0.67, + "content": "\\pi", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 358, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 210, + 373 + ], + "score": 1.0, + "content": "are always conditioned on", + "type": "text" + }, + { + "bbox": [ + 211, + 361, + 221, + 371 + ], + "score": 0.86, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 358, + 343, + 373 + ], + "score": 1.0, + "content": "and all outer expectations over", + "type": "text" + }, + { + "bbox": [ + 343, + 362, + 352, + 371 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 358, + 505, + 373 + ], + "score": 1.0, + "content": "are taken with respect to the trajectory", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 369, + 504, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 152, + 384 + ], + "score": 1.0, + "content": "induced by", + "type": "text" + }, + { + "bbox": [ + 153, + 371, + 160, + 382 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 369, + 208, + 384 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 209, + 372, + 216, + 380 + ], + "score": 0.8, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 369, + 435, + 384 + ], + "score": 1.0, + "content": "is essentially decoupled from the transition probability", + "type": "text" + }, + { + "bbox": [ + 435, + 371, + 504, + 383 + ], + "score": 0.91, + "content": "p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t \\right)", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 399, + 393 + ], + "score": 1.0, + "content": "and so its parameters can be updated without affecting the distribution of", + "type": "text" + }, + { + "bbox": [ + 400, + 383, + 409, + 393 + ], + "score": 0.82, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "’s. The subproblem that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 393, + 474, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 221, + 405 + ], + "score": 1.0, + "content": "needs to be solved to update", + "type": "text" + }, + { + "bbox": [ + 222, + 393, + 228, + 402 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 393, + 474, + 405 + ], + "score": 1.0, + "content": "therefore amounts to a standard supervised learning problem.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 132, + 422 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 132, + 409, + 189, + 421 + ], + "score": 0.92, + "content": "\\psi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\boldsymbol { \\eta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 408, + 209, + 422 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 209, + 409, + 257, + 421 + ], + "score": 0.92, + "content": "\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ; \\boldsymbol { \\theta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 408, + 317, + 422 + ], + "score": 1.0, + "content": "are Gaussian,", + "type": "text" + }, + { + "bbox": [ + 317, + 409, + 354, + 421 + ], + "score": 0.92, + "content": "D _ { t } \\left( { \\theta , \\eta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "can be computed analytically. More", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 209, + 433 + ], + "score": 1.0, + "content": "concretely, if we assume", + "type": "text" + }, + { + "bbox": [ + 209, + 421, + 223, + 430 + ], + "score": 0.87, + "content": "\\Sigma ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "to be fixed for simplicity, the subproblem that is solved for updating", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 431, + 171, + 444 + ], + "spans": [ + { + "bbox": [ + 107, + 431, + 157, + 443 + ], + "score": 0.93, + "content": "\\theta = \\left( \\omega , \\Sigma ^ { \\pi } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 431, + 171, + 444 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 462, + 451, + 525 + ], + "lines": [ + { + "bbox": [ + 159, + 462, + 451, + 525 + ], + "spans": [ + { + "bbox": [ + 159, + 462, + 451, + 525 + ], + "score": 0.92, + "content": "\\begin{array} { l } { \\displaystyle \\operatorname* { m i n } _ { \\theta } \\mathbb { E } _ { \\psi } [ \\sum _ { t = 0 } ^ { T } \\lambda _ { t } ^ { T } \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) + \\frac { \\nu _ { t } } { 2 } ( \\operatorname { t r } ( G _ { t } ^ { - 1 } \\Sigma ^ { \\pi } ) - \\log | \\Sigma ^ { \\pi } | ) } \\\\ { \\displaystyle + \\frac { \\nu _ { t } } { 2 } ( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) - \\mathbb { E } _ { \\psi } [ a _ { t } | s _ { t } , t ] ) ^ { T } G _ { t } ^ { - 1 } ( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) - \\mathbb { E } _ { \\psi } [ a _ { t } | s _ { t } , t ] ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "ec360d7a6f7432925bf8f970733b046a6e04fac406cb257daeefeab826336ea9.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 159, + 462, + 451, + 483.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 159, + 483.0, + 451, + 504.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 159, + 504.0, + 451, + 525.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 504, + 559 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 460, + 548 + ], + "score": 1.0, + "content": "Note that the last term is the squared Mahalanobis distance between the mean actions of", + "type": "text" + }, + { + "bbox": [ + 461, + 537, + 469, + 548 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 537, + 487, + 548 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 487, + 538, + 494, + 546 + ], + "score": 0.78, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 537, + 505, + 548 + ], + "score": 1.0, + "content": "at", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 547, + 406, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 145, + 560 + ], + "score": 1.0, + "content": "time step", + "type": "text" + }, + { + "bbox": [ + 145, + 549, + 150, + 558 + ], + "score": 0.64, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 547, + 348, + 560 + ], + "score": 1.0, + "content": ", which is intuitive as we would like to encourage", + "type": "text" + }, + { + "bbox": [ + 348, + 550, + 356, + 558 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 547, + 394, + 560 + ], + "score": 1.0, + "content": "to match", + "type": "text" + }, + { + "bbox": [ + 394, + 548, + 402, + 559 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 547, + 406, + 560 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 107, + 573, + 230, + 585 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 231, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 231, + 586 + ], + "score": 1.0, + "content": "3.5 CONVOLUTIONAL GPS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 504, + 606 + ], + "score": 1.0, + "content": "The problem of learning high-dimensional optimization algorithms presents challenges for rein-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "forcement learning algorithms due to high dimensionality of the state and action spaces. For exam-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "ple, in the case of GPS, because the running time of LQG is cubic in dimensionality of the state", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 626, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 505, + 641 + ], + "score": 1.0, + "content": "space, performing policy search even in the simple class of linear-Gaussian policies would be pro-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 639, + 434, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 434, + 650 + ], + "score": 1.0, + "content": "hibitively expensive when the dimensionality of the optimization problem is high.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "Fortunately, many high-dimensional optimization problems have underlying structure that can be", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 665, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 680 + ], + "score": 1.0, + "content": "exploited. For example, the parameters of neural nets are equivalent up to permutation among", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "certain coordinates. More concretely, for fully connected neural nets, the dimensions of a hidden", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "layer and the corresponding weights can be permuted arbitrarily without changing the function they", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "compute. Because permuting the dimensions of two adjacent layers can permute the weight matrix", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "arbitrarily, an optimization algorithm should be invariant to permutations of the rows and columns", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "of a weight matrix. A reasonable prior to impose is that the algorithm should behave in the same", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 761 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "At the start of each iteration, the algorithm constructs a model of the transition probability density", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 90, + 508, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 300, + 108 + ], + "score": 0.91, + "content": "\\tilde { p } \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t ; \\zeta \\right) = \\mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 90, + 331, + 113 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 331, + 93, + 427, + 108 + ], + "score": 0.93, + "content": "\\boldsymbol { \\zeta } : = \\left( A _ { t } , B _ { t } , c _ { t } , F _ { t } \\right) _ { t = 1 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 90, + 508, + 113 + ], + "score": 1.0, + "content": "is fitted to samples", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 117, + 120 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 109, + 127, + 118 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 105, + 281, + 120 + ], + "score": 1.0, + "content": "drawn from the trajectory induced by", + "type": "text" + }, + { + "bbox": [ + 281, + 107, + 289, + 118 + ], + "score": 0.83, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 105, + 506, + 120 + ], + "score": 1.0, + "content": ", which essentially amounts to a local linearization of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 225, + 131 + ], + "score": 1.0, + "content": "the true transition probability", + "type": "text" + }, + { + "bbox": [ + 226, + 118, + 294, + 129 + ], + "score": 0.91, + "content": "p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 117, + 348, + 131 + ], + "score": 1.0, + "content": ". We will use", + "type": "text" + }, + { + "bbox": [ + 348, + 117, + 372, + 131 + ], + "score": 0.92, + "content": "\\mathbb { E } _ { \\widetilde { \\psi } } \\left[ \\cdot \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 117, + 506, + 131 + ], + "score": 1.0, + "content": "to denote expectation taken with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 129, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 104, + 129, + 249, + 142 + ], + "score": 1.0, + "content": "respect to the trajectory induced by", + "type": "text" + }, + { + "bbox": [ + 249, + 131, + 258, + 141 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 129, + 424, + 142 + ], + "score": 1.0, + "content": "under the modelled transition probability", + "type": "text" + }, + { + "bbox": [ + 425, + 130, + 431, + 141 + ], + "score": 0.8, + "content": "\\tilde { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 129, + 506, + 142 + ], + "score": 1.0, + "content": ". Additionally, the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 139, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 295, + 155 + ], + "score": 1.0, + "content": "algorithm fits local quadratic approximations to", + "type": "text" + }, + { + "bbox": [ + 296, + 141, + 317, + 152 + ], + "score": 0.86, + "content": "c ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 139, + 392, + 155 + ], + "score": 1.0, + "content": "around samples of", + "type": "text" + }, + { + "bbox": [ + 392, + 142, + 401, + 152 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 139, + 505, + 155 + ], + "score": 1.0, + "content": "drawn from the trajectory", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 151, + 487, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 153, + 165 + ], + "score": 1.0, + "content": "induced by", + "type": "text" + }, + { + "bbox": [ + 153, + 152, + 161, + 163 + ], + "score": 0.85, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 151, + 191, + 165 + ], + "score": 1.0, + "content": "so that", + "type": "text" + }, + { + "bbox": [ + 192, + 151, + 351, + 165 + ], + "score": 0.93, + "content": "\\begin{array} { r } { c ( s _ { t } ) \\approx \\dot { \\tilde { c } } \\dot { ( s _ { t } ) } : = \\frac { 1 } { 2 } s _ { t } ^ { T } C _ { t } s _ { t } + \\dot { d _ { t } ^ { T } } s _ { t } + h _ { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 151, + 366, + 165 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 366, + 153, + 375, + 163 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 151, + 487, + 165 + ], + "score": 1.0, + "content": "’s that are near the samples.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 81, + 508, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 171, + 502, + 194 + ], + "lines": [ + { + "bbox": [ + 104, + 168, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 104, + 168, + 417, + 187 + ], + "score": 1.0, + "content": "With these assumptions, the subproblem that needs to be solved to update", + "type": "text" + }, + { + "bbox": [ + 417, + 169, + 504, + 184 + ], + "score": 0.92, + "content": "\\eta ~ = ~ ( K _ { t } , k _ { t } , G _ { t } ) _ { t = 1 } ^ { T }", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 181, + 147, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 147, + 194 + ], + "score": 1.0, + "content": "becomes:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 104, + 168, + 504, + 194 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 197, + 197, + 414, + 261 + ], + "lines": [ + { + "bbox": [ + 197, + 197, + 414, + 261 + ], + "spans": [ + { + "bbox": [ + 197, + 197, + 414, + 261 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } & { \\underset { \\eta } { \\operatorname* { m i n } } \\displaystyle \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\boldsymbol { \\tilde { \\psi } } } [ \\boldsymbol { \\tilde { c } } ( s _ { t } ) - \\lambda _ { t } ^ { T } \\boldsymbol { a } _ { t } ] + \\nu _ { t } D _ { t } ( \\eta , \\theta ) } & \\\\ & { \\mathrm { s . t . } \\displaystyle \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\boldsymbol { \\tilde { \\psi } } } [ D _ { K L } ( \\boldsymbol { \\psi } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\eta ) | \\Big | \\psi ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\eta ^ { \\prime } ) ) ] \\leq \\epsilon , } \\end{array}", + "type": "interline_equation", + "image_path": "da17dbd224435326b9983c40ce40ffd1486ed56e1c09a13c8b435c27ba5a8491.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 197, + 197, + 414, + 213.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 197, + 213.0, + 414, + 229.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 197, + 229.0, + 414, + 245.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 197, + 245.0, + 414, + 261.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 506, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 133, + 278 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 266, + 142, + 277 + ], + "score": 0.87, + "content": "\\eta ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 266, + 205, + 278 + ], + "score": 1.0, + "content": "denotes the old", + "type": "text" + }, + { + "bbox": [ + 205, + 267, + 212, + 277 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 266, + 358, + 278 + ], + "score": 1.0, + "content": "from the previous iteration. Because", + "type": "text" + }, + { + "bbox": [ + 359, + 267, + 366, + 277 + ], + "score": 0.85, + "content": "\\tilde { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 266, + 383, + 278 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 383, + 267, + 389, + 276 + ], + "score": 0.65, + "content": "\\tilde { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "are only valid locally around", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 210, + 289 + ], + "score": 1.0, + "content": "the trajectory induced by", + "type": "text" + }, + { + "bbox": [ + 211, + 278, + 219, + 288 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 276, + 439, + 289 + ], + "score": 1.0, + "content": ", the constraint is added to limit the amount by which", + "type": "text" + }, + { + "bbox": [ + 439, + 279, + 446, + 289 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "is updated. It", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 287, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 301 + ], + "score": 1.0, + "content": "turns out that the unconstrained problem can be solved in closed form using a dynamic programming", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 299, + 504, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 495, + 311 + ], + "score": 1.0, + "content": "algorithm known as linear-quadratic-Gaussian (LQG) regulator in time linear in the time horizon", + "type": "text" + }, + { + "bbox": [ + 496, + 300, + 504, + 309 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 306, + 321 + ], + "score": 1.0, + "content": "and cubic in the dimensionality of the state space", + "type": "text" + }, + { + "bbox": [ + 307, + 311, + 316, + 320 + ], + "score": 0.8, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 309, + 505, + 321 + ], + "score": 1.0, + "content": ". The constrained problem is solved using dual", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "gradient descent, which uses LQG as a subroutine to solve for the primal variables in each iteration", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 332, + 383, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 383, + 343 + ], + "score": 1.0, + "content": "and increments the dual variable on the constraint until it is satisfied.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 266, + 505, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 504, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 146, + 361 + ], + "score": 1.0, + "content": "Updating", + "type": "text" + }, + { + "bbox": [ + 146, + 349, + 153, + 359 + ], + "score": 0.75, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 349, + 496, + 361 + ], + "score": 1.0, + "content": "is straightforward, since expectations taken with respect to the trajectory induced by", + "type": "text" + }, + { + "bbox": [ + 497, + 351, + 504, + 358 + ], + "score": 0.67, + "content": "\\pi", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 358, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 210, + 373 + ], + "score": 1.0, + "content": "are always conditioned on", + "type": "text" + }, + { + "bbox": [ + 211, + 361, + 221, + 371 + ], + "score": 0.86, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 358, + 343, + 373 + ], + "score": 1.0, + "content": "and all outer expectations over", + "type": "text" + }, + { + "bbox": [ + 343, + 362, + 352, + 371 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 358, + 505, + 373 + ], + "score": 1.0, + "content": "are taken with respect to the trajectory", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 369, + 504, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 152, + 384 + ], + "score": 1.0, + "content": "induced by", + "type": "text" + }, + { + "bbox": [ + 153, + 371, + 160, + 382 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 369, + 208, + 384 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 209, + 372, + 216, + 380 + ], + "score": 0.8, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 369, + 435, + 384 + ], + "score": 1.0, + "content": "is essentially decoupled from the transition probability", + "type": "text" + }, + { + "bbox": [ + 435, + 371, + 504, + 383 + ], + "score": 0.91, + "content": "p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t \\right)", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 399, + 393 + ], + "score": 1.0, + "content": "and so its parameters can be updated without affecting the distribution of", + "type": "text" + }, + { + "bbox": [ + 400, + 383, + 409, + 393 + ], + "score": 0.82, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "’s. The subproblem that", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 393, + 474, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 221, + 405 + ], + "score": 1.0, + "content": "needs to be solved to update", + "type": "text" + }, + { + "bbox": [ + 222, + 393, + 228, + 402 + ], + "score": 0.8, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 393, + 474, + 405 + ], + "score": 1.0, + "content": "therefore amounts to a standard supervised learning problem.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 349, + 505, + 405 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 132, + 422 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 132, + 409, + 189, + 421 + ], + "score": 0.92, + "content": "\\psi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\boldsymbol { \\eta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 408, + 209, + 422 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 209, + 409, + 257, + 421 + ], + "score": 0.92, + "content": "\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ; \\boldsymbol { \\theta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 408, + 317, + 422 + ], + "score": 1.0, + "content": "are Gaussian,", + "type": "text" + }, + { + "bbox": [ + 317, + 409, + 354, + 421 + ], + "score": 0.92, + "content": "D _ { t } \\left( { \\theta , \\eta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "can be computed analytically. More", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 209, + 433 + ], + "score": 1.0, + "content": "concretely, if we assume", + "type": "text" + }, + { + "bbox": [ + 209, + 421, + 223, + 430 + ], + "score": 0.87, + "content": "\\Sigma ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "to be fixed for simplicity, the subproblem that is solved for updating", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 431, + 171, + 444 + ], + "spans": [ + { + "bbox": [ + 107, + 431, + 157, + 443 + ], + "score": 0.93, + "content": "\\theta = \\left( \\omega , \\Sigma ^ { \\pi } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 431, + 171, + 444 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 408, + 506, + 444 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 462, + 451, + 525 + ], + "lines": [ + { + "bbox": [ + 159, + 462, + 451, + 525 + ], + "spans": [ + { + "bbox": [ + 159, + 462, + 451, + 525 + ], + "score": 0.92, + "content": "\\begin{array} { l } { \\displaystyle \\operatorname* { m i n } _ { \\theta } \\mathbb { E } _ { \\psi } [ \\sum _ { t = 0 } ^ { T } \\lambda _ { t } ^ { T } \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) + \\frac { \\nu _ { t } } { 2 } ( \\operatorname { t r } ( G _ { t } ^ { - 1 } \\Sigma ^ { \\pi } ) - \\log | \\Sigma ^ { \\pi } | ) } \\\\ { \\displaystyle + \\frac { \\nu _ { t } } { 2 } ( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) - \\mathbb { E } _ { \\psi } [ a _ { t } | s _ { t } , t ] ) ^ { T } G _ { t } ^ { - 1 } ( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) - \\mathbb { E } _ { \\psi } [ a _ { t } | s _ { t } , t ] ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "ec360d7a6f7432925bf8f970733b046a6e04fac406cb257daeefeab826336ea9.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 159, + 462, + 451, + 483.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 159, + 483.0, + 451, + 504.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 159, + 504.0, + 451, + 525.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 504, + 559 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 460, + 548 + ], + "score": 1.0, + "content": "Note that the last term is the squared Mahalanobis distance between the mean actions of", + "type": "text" + }, + { + "bbox": [ + 461, + 537, + 469, + 548 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 537, + 487, + 548 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 487, + 538, + 494, + 546 + ], + "score": 0.78, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 537, + 505, + 548 + ], + "score": 1.0, + "content": "at", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 547, + 406, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 145, + 560 + ], + "score": 1.0, + "content": "time step", + "type": "text" + }, + { + "bbox": [ + 145, + 549, + 150, + 558 + ], + "score": 0.64, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 547, + 348, + 560 + ], + "score": 1.0, + "content": ", which is intuitive as we would like to encourage", + "type": "text" + }, + { + "bbox": [ + 348, + 550, + 356, + 558 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 547, + 394, + 560 + ], + "score": 1.0, + "content": "to match", + "type": "text" + }, + { + "bbox": [ + 394, + 548, + 402, + 559 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 547, + 406, + 560 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 537, + 505, + 560 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 573, + 230, + 585 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 231, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 231, + 586 + ], + "score": 1.0, + "content": "3.5 CONVOLUTIONAL GPS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 504, + 606 + ], + "score": 1.0, + "content": "The problem of learning high-dimensional optimization algorithms presents challenges for rein-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "forcement learning algorithms due to high dimensionality of the state and action spaces. For exam-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "ple, in the case of GPS, because the running time of LQG is cubic in dimensionality of the state", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 626, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 505, + 641 + ], + "score": 1.0, + "content": "space, performing policy search even in the simple class of linear-Gaussian policies would be pro-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 639, + 434, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 434, + 650 + ], + "score": 1.0, + "content": "hibitively expensive when the dimensionality of the optimization problem is high.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 104, + 594, + 505, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "Fortunately, many high-dimensional optimization problems have underlying structure that can be", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 665, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 680 + ], + "score": 1.0, + "content": "exploited. For example, the parameters of neural nets are equivalent up to permutation among", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "certain coordinates. More concretely, for fully connected neural nets, the dimensions of a hidden", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "layer and the corresponding weights can be permuted arbitrarily without changing the function they", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "compute. Because permuting the dimensions of two adjacent layers can permute the weight matrix", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "arbitrarily, an optimization algorithm should be invariant to permutations of the rows and columns", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "of a weight matrix. A reasonable prior to impose is that the algorithm should behave in the same", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "score": 1.0, + "content": "manner on all coordinates that correspond to entries in the same matrix. That is, if the values", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 277, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 506, + 289 + ], + "score": 1.0, + "content": "of two coordinates in all current and past gradients and iterates are identical, then the step vector", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "produced by the algorithm should have identical values in these two coordinates. We will refer to", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "the set of coordinates on which permutation invariance is enforced as a coordinate group. For the", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "purposes of learning an optimization algorithm for neural nets, a natural choice would be to make", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "each coordinate group correspond to a weight matrix or a bias vector. Hence, the total number of", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 332, + 418, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 418, + 344 + ], + "score": 1.0, + "content": "coordinate groups is twice the number of layers, which is usually fairly small.", + "type": "text", + "cross_page": true + } + ], + "index": 13 + } + ], + "index": 42, + "bbox_fs": [ + 104, + 654, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 97, + 479, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 97, + 479, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 97, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 129, + 97, + 479, + 188 + ], + "score": 0.967, + "type": "image", + "image_path": "c32356121b006ac72da2adddd464187f503567b08623692799a8cd7bffe7f36e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 97, + 479, + 127.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 127.33333333333333, + 479, + 157.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 157.66666666666666, + 479, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 506, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "Figure 1: Comparison of the various hand-engineered and learned algorithms on training neural nets", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 209, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 219 + ], + "score": 1.0, + "content": "with 48 input and hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-batches of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "size 64. The vertical axis is the true objective value and the horizontal axis represents the iteration.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 229, + 198, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 198, + 243 + ], + "score": 1.0, + "content": "Best viewed in colour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 505, + 343 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 277 + ], + "score": 1.0, + "content": "manner on all coordinates that correspond to entries in the same matrix. That is, if the values", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 277, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 506, + 289 + ], + "score": 1.0, + "content": "of two coordinates in all current and past gradients and iterates are identical, then the step vector", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "produced by the algorithm should have identical values in these two coordinates. We will refer to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "the set of coordinates on which permutation invariance is enforced as a coordinate group. For the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "purposes of learning an optimization algorithm for neural nets, a natural choice would be to make", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "each coordinate group correspond to a weight matrix or a bias vector. Hence, the total number of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 332, + 418, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 418, + 344 + ], + "score": 1.0, + "content": "coordinate groups is twice the number of layers, which is usually fairly small.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 305, + 361 + ], + "score": 1.0, + "content": "In the case of GPS, we impose this prior on both", + "type": "text" + }, + { + "bbox": [ + 305, + 349, + 313, + 360 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 347, + 331, + 361 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 332, + 350, + 339, + 358 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 347, + 462, + 361 + ], + "score": 1.0, + "content": ". For the purposes of updating", + "type": "text" + }, + { + "bbox": [ + 462, + 351, + 469, + 360 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 347, + 505, + 361 + ], + "score": 1.0, + "content": ", we first", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 324, + 372 + ], + "score": 1.0, + "content": "impose a block-diagonal structure on the parameters", + "type": "text" + }, + { + "bbox": [ + 324, + 360, + 351, + 370 + ], + "score": 0.92, + "content": "A _ { t } , B _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 360, + 371, + 372 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 371, + 360, + 382, + 370 + ], + "score": 0.88, + "content": "F _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "of the fitted transition proba-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 162, + 384 + ], + "score": 1.0, + "content": "bility density", + "type": "text" + }, + { + "bbox": [ + 163, + 370, + 362, + 383 + ], + "score": 0.9, + "content": "\\tilde { p } \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t ; \\zeta \\right) = \\mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 370, + 505, + 384 + ], + "score": 1.0, + "content": ", so that for each coordinate in the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 271, + 394 + ], + "score": 1.0, + "content": "optimization problem, the dimensions of", + "type": "text" + }, + { + "bbox": [ + 272, + 384, + 290, + 393 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "that correspond to the coordinate only depend on the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 392, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 167, + 405 + ], + "score": 1.0, + "content": "dimensions of", + "type": "text" + }, + { + "bbox": [ + 167, + 394, + 176, + 403 + ], + "score": 0.87, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 392, + 196, + 405 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 196, + 394, + 206, + 403 + ], + "score": 0.86, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 392, + 425, + 405 + ], + "score": 1.0, + "content": "that correspond to the same coordinate. As a result,", + "type": "text" + }, + { + "bbox": [ + 426, + 392, + 504, + 404 + ], + "score": 0.89, + "content": "\\tilde { p } \\big ( s _ { t + 1 } \\big | s _ { t } , a _ { t } , t ; \\zeta \\big )", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 403, + 507, + 424 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 353, + 424 + ], + "score": 1.0, + "content": "decomposes into multiple independent probability densities", + "type": "text" + }, + { + "bbox": [ + 353, + 403, + 446, + 423 + ], + "score": 0.93, + "content": "\\tilde { p } ^ { j } \\left( s _ { t + 1 } ^ { j } \\Big | s _ { t } ^ { j } , a _ { t } ^ { j } , t ; \\zeta ^ { j } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 403, + 507, + 424 + ], + "score": 1.0, + "content": ", one for each", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 422, + 504, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 151, + 434 + ], + "score": 1.0, + "content": "coordinate", + "type": "text" + }, + { + "bbox": [ + 152, + 423, + 158, + 434 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 422, + 395, + 434 + ], + "score": 1.0, + "content": ". Similarly, we also impose a block-diagonal structure on", + "type": "text" + }, + { + "bbox": [ + 396, + 423, + 407, + 433 + ], + "score": 0.86, + "content": "C _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 422, + 451, + 434 + ], + "score": 1.0, + "content": "for fitting", + "type": "text" + }, + { + "bbox": [ + 452, + 422, + 473, + 434 + ], + "score": 0.9, + "content": "\\tilde { c } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 422, + 504, + 434 + ], + "score": 1.0, + "content": "and on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 285, + 446 + ], + "score": 1.0, + "content": "the parameter matrix of the fitted model for", + "type": "text" + }, + { + "bbox": [ + 285, + 433, + 333, + 445 + ], + "score": 0.92, + "content": "\\pi \\left( a _ { t } | \\boldsymbol s _ { t } \\bar { ; \\boldsymbol \\theta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 432, + 444, + 446 + ], + "score": 1.0, + "content": ". Under these assumptions,", + "type": "text" + }, + { + "bbox": [ + 444, + 434, + 457, + 444 + ], + "score": 0.89, + "content": "K _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 432, + 476, + 446 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 476, + 434, + 488, + 444 + ], + "score": 0.89, + "content": "G _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 444, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 470, + 457 + ], + "score": 1.0, + "content": "guaranteed to be block-diagonal as well. Hence, the Bregman divergence penalty term,", + "type": "text" + }, + { + "bbox": [ + 470, + 444, + 504, + 456 + ], + "score": 0.92, + "content": "D \\left( \\eta , \\theta \\right)", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 455, + 423, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 423, + 467 + ], + "score": 1.0, + "content": "decomposes into a sum of Bregman divergence terms, one for each coordinate.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 272, + 484 + ], + "score": 1.0, + "content": "We then further constrain dual variables", + "type": "text" + }, + { + "bbox": [ + 272, + 473, + 282, + 483 + ], + "score": 0.88, + "content": "\\lambda _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 471, + 506, + 484 + ], + "score": 1.0, + "content": ", sub-vectors of parameter vectors and sub-matrices of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "parameter matrices corresponding to each coordinate group to be identical across the group. Addi-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 228, + 507 + ], + "score": 1.0, + "content": "tionally, we replace the weight", + "type": "text" + }, + { + "bbox": [ + 229, + 495, + 239, + 505 + ], + "score": 0.84, + "content": "\\nu _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 494, + 251, + 507 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 252, + 494, + 285, + 506 + ], + "score": 0.91, + "content": "D \\left( \\eta , \\theta \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "with an individual weight on each Bregman divergence", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "term for each coordinate group. The problem then decomposes into multiple independent subprob-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "lems, one for each coordinate group. Because the dimensionality of the state subspace corresponding", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 527, + 492, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 492, + 539 + ], + "score": 1.0, + "content": "to each coordinate is constant, LQG can be executed on each subproblem much more efficiently.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 162, + 556 + ], + "score": 1.0, + "content": "Similarly, for", + "type": "text" + }, + { + "bbox": [ + 163, + 546, + 170, + 554 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 543, + 226, + 556 + ], + "score": 1.0, + "content": ", we choose a", + "type": "text" + }, + { + "bbox": [ + 226, + 543, + 249, + 556 + ], + "score": 0.92, + "content": "\\mu _ { \\omega } ^ { \\pi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "that shares parameters across different coordinates in the same", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 318, + 568 + ], + "score": 1.0, + "content": "group. We also impose a block-diagonal structure on", + "type": "text" + }, + { + "bbox": [ + 319, + 555, + 332, + 565 + ], + "score": 0.87, + "content": "\\Sigma ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "and constrain the appropriate sub-matrices", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 566, + 192, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 192, + 577 + ], + "score": 1.0, + "content": "to share their entries.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 107, + 593, + 176, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 177, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 177, + 606 + ], + "score": 1.0, + "content": "3.6 FEATURES", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 504, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 208, + 629 + ], + "score": 1.0, + "content": "We describe the features", + "type": "text" + }, + { + "bbox": [ + 208, + 615, + 227, + 627 + ], + "score": 0.91, + "content": "\\Phi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 612, + 246, + 629 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 246, + 615, + 265, + 627 + ], + "score": 0.91, + "content": "\\Psi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 612, + 317, + 629 + ], + "score": 1.0, + "content": "at time step", + "type": "text" + }, + { + "bbox": [ + 317, + 617, + 322, + 625 + ], + "score": 0.68, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 612, + 416, + 629 + ], + "score": 1.0, + "content": ", which define the state", + "type": "text" + }, + { + "bbox": [ + 417, + 617, + 426, + 626 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 612, + 494, + 629 + ], + "score": 1.0, + "content": "and observation", + "type": "text" + }, + { + "bbox": [ + 495, + 617, + 504, + 626 + ], + "score": 0.81, + "content": "o _ { t }", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 625, + 160, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 160, + 640 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 642, + 506, + 708 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 436, + 656 + ], + "score": 1.0, + "content": "Because of the stochasticity of gradients and objective values, the state features", + "type": "text" + }, + { + "bbox": [ + 437, + 643, + 455, + 655 + ], + "score": 0.9, + "content": "\\Phi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "are defined", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 102, + 654, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 102, + 655, + 335, + 677 + ], + "score": 1.0, + "content": "in terms of summary statistics of the history of iterates", + "type": "text" + }, + { + "bbox": [ + 335, + 656, + 376, + 673 + ], + "score": 0.93, + "content": "\\left\\{ x ^ { ( i ) } \\right\\} _ { i = 0 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 655, + 421, + 677 + ], + "score": 1.0, + "content": ", gradients", + "type": "text" + }, + { + "bbox": [ + 421, + 654, + 486, + 677 + ], + "score": 0.92, + "content": "\\left\\{ \\nabla \\hat { f } ( x ^ { ( i ) } ) \\right\\} _ { i = 0 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 675, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 175, + 696 + ], + "score": 1.0, + "content": "objective values", + "type": "text" + }, + { + "bbox": [ + 175, + 675, + 232, + 698 + ], + "score": 0.93, + "content": "\\left\\{ \\hat { f } ( \\boldsymbol x ^ { ( i ) } ) \\right\\} _ { i = 0 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 678, + 506, + 694 + ], + "score": 1.0, + "content": "We define the following statistics, which we will refer to as the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 694, + 365, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 694, + 365, + 710 + ], + "score": 1.0, + "content": "average recent iterate, gradient and objective value respectively:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 719, + 285, + 735 + ], + "lines": [ + { + "bbox": [ + 140, + 719, + 285, + 735 + ], + "spans": [ + { + "bbox": [ + 140, + 719, + 285, + 735 + ], + "score": 0.83, + "content": "\\begin{array} { r } { \\overline { { \\boldsymbol { x } ^ { ( i ) } } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \\boldsymbol { x } ^ { ( j ) } } \\end{array}", + "type": "interline_equation", + "image_path": "127a52a319bf4db3aba86fdbf80c5271cf7ab40491a1981f2627a709932c8dc8.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 140, + 719, + 285, + 735 + ], + "spans": [], + "index": 40 + } + ] + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 97, + 479, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 97, + 479, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 97, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 129, + 97, + 479, + 188 + ], + "score": 0.967, + "type": "image", + "image_path": "c32356121b006ac72da2adddd464187f503567b08623692799a8cd7bffe7f36e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 97, + 479, + 127.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 127.33333333333333, + 479, + 157.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 157.66666666666666, + 479, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 506, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "Figure 1: Comparison of the various hand-engineered and learned algorithms on training neural nets", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 209, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 219 + ], + "score": 1.0, + "content": "with 48 input and hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-batches of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "size 64. The vertical axis is the true objective value and the horizontal axis represents the iteration.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 229, + 198, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 198, + 243 + ], + "score": 1.0, + "content": "Best viewed in colour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 505, + 343 + ], + "lines": [], + "index": 10, + "bbox_fs": [ + 105, + 266, + 506, + 344 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 348, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 305, + 361 + ], + "score": 1.0, + "content": "In the case of GPS, we impose this prior on both", + "type": "text" + }, + { + "bbox": [ + 305, + 349, + 313, + 360 + ], + "score": 0.86, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 347, + 331, + 361 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 332, + 350, + 339, + 358 + ], + "score": 0.76, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 347, + 462, + 361 + ], + "score": 1.0, + "content": ". For the purposes of updating", + "type": "text" + }, + { + "bbox": [ + 462, + 351, + 469, + 360 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 347, + 505, + 361 + ], + "score": 1.0, + "content": ", we first", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 324, + 372 + ], + "score": 1.0, + "content": "impose a block-diagonal structure on the parameters", + "type": "text" + }, + { + "bbox": [ + 324, + 360, + 351, + 370 + ], + "score": 0.92, + "content": "A _ { t } , B _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 360, + 371, + 372 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 371, + 360, + 382, + 370 + ], + "score": 0.88, + "content": "F _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "of the fitted transition proba-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 162, + 384 + ], + "score": 1.0, + "content": "bility density", + "type": "text" + }, + { + "bbox": [ + 163, + 370, + 362, + 383 + ], + "score": 0.9, + "content": "\\tilde { p } \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t ; \\zeta \\right) = \\mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 370, + 505, + 384 + ], + "score": 1.0, + "content": ", so that for each coordinate in the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 271, + 394 + ], + "score": 1.0, + "content": "optimization problem, the dimensions of", + "type": "text" + }, + { + "bbox": [ + 272, + 384, + 290, + 393 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "that correspond to the coordinate only depend on the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 392, + 504, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 167, + 405 + ], + "score": 1.0, + "content": "dimensions of", + "type": "text" + }, + { + "bbox": [ + 167, + 394, + 176, + 403 + ], + "score": 0.87, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 392, + 196, + 405 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 196, + 394, + 206, + 403 + ], + "score": 0.86, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 392, + 425, + 405 + ], + "score": 1.0, + "content": "that correspond to the same coordinate. As a result,", + "type": "text" + }, + { + "bbox": [ + 426, + 392, + 504, + 404 + ], + "score": 0.89, + "content": "\\tilde { p } \\big ( s _ { t + 1 } \\big | s _ { t } , a _ { t } , t ; \\zeta \\big )", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 403, + 507, + 424 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 353, + 424 + ], + "score": 1.0, + "content": "decomposes into multiple independent probability densities", + "type": "text" + }, + { + "bbox": [ + 353, + 403, + 446, + 423 + ], + "score": 0.93, + "content": "\\tilde { p } ^ { j } \\left( s _ { t + 1 } ^ { j } \\Big | s _ { t } ^ { j } , a _ { t } ^ { j } , t ; \\zeta ^ { j } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 403, + 507, + 424 + ], + "score": 1.0, + "content": ", one for each", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 422, + 504, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 151, + 434 + ], + "score": 1.0, + "content": "coordinate", + "type": "text" + }, + { + "bbox": [ + 152, + 423, + 158, + 434 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 422, + 395, + 434 + ], + "score": 1.0, + "content": ". Similarly, we also impose a block-diagonal structure on", + "type": "text" + }, + { + "bbox": [ + 396, + 423, + 407, + 433 + ], + "score": 0.86, + "content": "C _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 422, + 451, + 434 + ], + "score": 1.0, + "content": "for fitting", + "type": "text" + }, + { + "bbox": [ + 452, + 422, + 473, + 434 + ], + "score": 0.9, + "content": "\\tilde { c } ( s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 422, + 504, + 434 + ], + "score": 1.0, + "content": "and on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 285, + 446 + ], + "score": 1.0, + "content": "the parameter matrix of the fitted model for", + "type": "text" + }, + { + "bbox": [ + 285, + 433, + 333, + 445 + ], + "score": 0.92, + "content": "\\pi \\left( a _ { t } | \\boldsymbol s _ { t } \\bar { ; \\boldsymbol \\theta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 432, + 444, + 446 + ], + "score": 1.0, + "content": ". Under these assumptions,", + "type": "text" + }, + { + "bbox": [ + 444, + 434, + 457, + 444 + ], + "score": 0.89, + "content": "K _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 432, + 476, + 446 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 476, + 434, + 488, + 444 + ], + "score": 0.89, + "content": "G _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 444, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 470, + 457 + ], + "score": 1.0, + "content": "guaranteed to be block-diagonal as well. Hence, the Bregman divergence penalty term,", + "type": "text" + }, + { + "bbox": [ + 470, + 444, + 504, + 456 + ], + "score": 0.92, + "content": "D \\left( \\eta , \\theta \\right)", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 455, + 423, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 423, + 467 + ], + "score": 1.0, + "content": "decomposes into a sum of Bregman divergence terms, one for each coordinate.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 347, + 507, + 467 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 272, + 484 + ], + "score": 1.0, + "content": "We then further constrain dual variables", + "type": "text" + }, + { + "bbox": [ + 272, + 473, + 282, + 483 + ], + "score": 0.88, + "content": "\\lambda _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 471, + 506, + 484 + ], + "score": 1.0, + "content": ", sub-vectors of parameter vectors and sub-matrices of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "parameter matrices corresponding to each coordinate group to be identical across the group. Addi-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 228, + 507 + ], + "score": 1.0, + "content": "tionally, we replace the weight", + "type": "text" + }, + { + "bbox": [ + 229, + 495, + 239, + 505 + ], + "score": 0.84, + "content": "\\nu _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 494, + 251, + 507 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 252, + 494, + 285, + 506 + ], + "score": 0.91, + "content": "D \\left( \\eta , \\theta \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "with an individual weight on each Bregman divergence", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "term for each coordinate group. The problem then decomposes into multiple independent subprob-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "lems, one for each coordinate group. Because the dimensionality of the state subspace corresponding", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 527, + 492, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 492, + 539 + ], + "score": 1.0, + "content": "to each coordinate is constant, LQG can be executed on each subproblem much more efficiently.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 471, + 506, + 539 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 162, + 556 + ], + "score": 1.0, + "content": "Similarly, for", + "type": "text" + }, + { + "bbox": [ + 163, + 546, + 170, + 554 + ], + "score": 0.74, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 543, + 226, + 556 + ], + "score": 1.0, + "content": ", we choose a", + "type": "text" + }, + { + "bbox": [ + 226, + 543, + 249, + 556 + ], + "score": 0.92, + "content": "\\mu _ { \\omega } ^ { \\pi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "that shares parameters across different coordinates in the same", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 318, + 568 + ], + "score": 1.0, + "content": "group. We also impose a block-diagonal structure on", + "type": "text" + }, + { + "bbox": [ + 319, + 555, + 332, + 565 + ], + "score": 0.87, + "content": "\\Sigma ^ { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "and constrain the appropriate sub-matrices", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 566, + 192, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 192, + 577 + ], + "score": 1.0, + "content": "to share their entries.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 543, + 505, + 577 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 593, + 176, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 177, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 177, + 606 + ], + "score": 1.0, + "content": "3.6 FEATURES", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 504, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 208, + 629 + ], + "score": 1.0, + "content": "We describe the features", + "type": "text" + }, + { + "bbox": [ + 208, + 615, + 227, + 627 + ], + "score": 0.91, + "content": "\\Phi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 612, + 246, + 629 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 246, + 615, + 265, + 627 + ], + "score": 0.91, + "content": "\\Psi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 612, + 317, + 629 + ], + "score": 1.0, + "content": "at time step", + "type": "text" + }, + { + "bbox": [ + 317, + 617, + 322, + 625 + ], + "score": 0.68, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 612, + 416, + 629 + ], + "score": 1.0, + "content": ", which define the state", + "type": "text" + }, + { + "bbox": [ + 417, + 617, + 426, + 626 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 612, + 494, + 629 + ], + "score": 1.0, + "content": "and observation", + "type": "text" + }, + { + "bbox": [ + 495, + 617, + 504, + 626 + ], + "score": 0.81, + "content": "o _ { t }", + "type": "inline_equation" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 625, + 160, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 160, + 640 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 612, + 504, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 642, + 506, + 708 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 436, + 656 + ], + "score": 1.0, + "content": "Because of the stochasticity of gradients and objective values, the state features", + "type": "text" + }, + { + "bbox": [ + 437, + 643, + 455, + 655 + ], + "score": 0.9, + "content": "\\Phi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "are defined", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 102, + 654, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 102, + 655, + 335, + 677 + ], + "score": 1.0, + "content": "in terms of summary statistics of the history of iterates", + "type": "text" + }, + { + "bbox": [ + 335, + 656, + 376, + 673 + ], + "score": 0.93, + "content": "\\left\\{ x ^ { ( i ) } \\right\\} _ { i = 0 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 655, + 421, + 677 + ], + "score": 1.0, + "content": ", gradients", + "type": "text" + }, + { + "bbox": [ + 421, + 654, + 486, + 677 + ], + "score": 0.92, + "content": "\\left\\{ \\nabla \\hat { f } ( x ^ { ( i ) } ) \\right\\} _ { i = 0 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 675, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 175, + 696 + ], + "score": 1.0, + "content": "objective values", + "type": "text" + }, + { + "bbox": [ + 175, + 675, + 232, + 698 + ], + "score": 0.93, + "content": "\\left\\{ \\hat { f } ( \\boldsymbol x ^ { ( i ) } ) \\right\\} _ { i = 0 } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 678, + 506, + 694 + ], + "score": 1.0, + "content": "We define the following statistics, which we will refer to as the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 694, + 365, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 694, + 365, + 710 + ], + "score": 1.0, + "content": "average recent iterate, gradient and objective value respectively:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 102, + 642, + 506, + 710 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 719, + 285, + 735 + ], + "lines": [ + { + "bbox": [ + 140, + 719, + 285, + 735 + ], + "spans": [ + { + "bbox": [ + 140, + 719, + 285, + 735 + ], + "score": 0.83, + "content": "\\begin{array} { r } { \\overline { { \\boldsymbol { x } ^ { ( i ) } } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \\boldsymbol { x } ^ { ( j ) } } \\end{array}", + "type": "interline_equation", + "image_path": "127a52a319bf4db3aba86fdbf80c5271cf7ab40491a1981f2627a709932c8dc8.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 140, + 719, + 285, + 735 + ], + "spans": [], + "index": 40 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 96, + 479, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 96, + 479, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 96, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 128, + 96, + 479, + 188 + ], + "score": 0.971, + "type": "image", + "image_path": "3fea8480e1dc0f7dbc3c3a6dc8993cac0f6ffab3eae77436f1e3e961bbffec59.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 96, + 479, + 126.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 126.66666666666667, + 479, + 157.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 157.33333333333334, + 479, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 506, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 197, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 209 + ], + "score": 1.0, + "content": "Figure 2: Comparison of the various hand-engineered and learned algorithms on training neural nets", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "batches of size 64. The vertical axis is the true objective value and the horizontal axis represents the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 236, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 236, + 242 + ], + "score": 1.0, + "content": "iteration. Best viewed in colour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 260, + 326, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 326, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 326, + 299 + ], + "score": 0.81, + "content": "\\begin{array} { r l } & { \\bullet \\ \\overline { { \\nabla \\hat { f } ( x ^ { ( i ) } ) } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } { \\nabla \\hat { f } ( x ^ { ( j ) } ) } } \\\\ & { \\bullet \\ \\overline { { \\hat { f } ( x ^ { ( i ) } ) } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \\hat { f } ( x ^ { ( j ) } ) } \\end{array}", + "type": "interline_equation", + "image_path": "7a6761589b90ee0c7200cf58fa051acd5f2b63ec2d32a67a2f1eb5d8569dab37.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 130, + 260, + 326, + 273.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 130, + 273.0, + 326, + 286.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 130, + 286.0, + 326, + 299.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 504, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 183, + 320 + ], + "score": 1.0, + "content": "The state features", + "type": "text" + }, + { + "bbox": [ + 184, + 307, + 203, + 319 + ], + "score": 0.91, + "content": "\\Phi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 306, + 506, + 320 + ], + "score": 1.0, + "content": "consist of the relative change in the average recent objective value, the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "average recent gradient normalized by the magnitude of the a previous average recent gradient and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 330, + 497, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 497, + 342 + ], + "score": 1.0, + "content": "a previous change in average recent iterate relative to the current change in average recent iterate:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 131, + 349, + 358, + 426 + ], + "lines": [ + { + "bbox": [ + 131, + 349, + 358, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 358, + 426 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\bullet \\{ ( \\hat { f } ( x ^ { ( t - 5 i ) } ) - \\widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) ) / \\widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) \\} _ { i = 0 } ^ { 2 4 } } \\\\ & { \\bullet \\{ \\overline { { \\nabla \\hat { f } ( x ^ { ( t - 5 i ) } ) } } / ( | \\overline { { \\nabla \\hat { f } ( x ^ { ( m a x ( t - 5 ( i + 1 ) , t \\mathrm { m o d } 5 ) ) } ) } } | + 1 ) \\} _ { i = 0 } ^ { 2 5 } } \\\\ & \\bullet \\{ \\frac { | \\frac { \\overline { { x ^ { ( \\mathrm { m a x } ( t - 5 ( i + 1 ) , t \\mathrm { m o d } 5 + 5 ) ) } } } - \\overline { { x ^ { ( \\mathrm { m a x } ( t - 5 ( i + 2 ) , t \\mathrm { m o d } 5 ) ) } } } } { | \\overline { { x ^ { ( t - 5 i ) } } } - \\overline { { x ^ { ( t - 5 ( i + 1 ) ) } } } | + 0 . 1 } \\} _ { i = 0 } ^ { 2 4 } \\} _ { i = 0 } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "1c2af21e865a7d14970ed22ee4629d67c1f8d6b7a64a8e247bf2ff52147bdb2e.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 131, + 349, + 358, + 361.8333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 131, + 361.8333333333333, + 358, + 374.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 131, + 374.66666666666663, + 358, + 387.49999999999994 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 131, + 387.49999999999994, + 358, + 400.33333333333326 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 131, + 400.33333333333326, + 358, + 413.1666666666666 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 131, + 413.1666666666666, + 358, + 425.9999999999999 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 432, + 503, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 432, + 504, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 504, + 444 + ], + "score": 1.0, + "content": "Note that all operations are applied element-wise. Also, whenever a feature becomes undefined (i.e.:", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 444, + 442, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 442, + 455 + ], + "score": 1.0, + "content": "when the time step index becomes negative), it is replaced with the all-zeros vector.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 473 + ], + "score": 1.0, + "content": "Unlike state features, which are only used when training the optimization algorithm, observation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 140, + 484 + ], + "score": 1.0, + "content": "features", + "type": "text" + }, + { + "bbox": [ + 141, + 471, + 159, + 483 + ], + "score": 0.91, + "content": "\\Psi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "are used both during training and at test time. Consequently, we use noisier observation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "features that can be computed more efficiently and require less memory overhead. The observation", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 492, + 240, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 240, + 507 + ], + "score": 1.0, + "content": "features consist of the following:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 513, + 288, + 579 + ], + "lines": [ + { + "bbox": [ + 130, + 513, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 288, + 579 + ], + "score": 0.91, + "content": "\\begin{array} { r l } { \\bullet } & { \\left( \\hat { f } ( x ^ { ( t ) } ) - \\hat { f } ( x ^ { ( t - 1 ) } ) \\right) / \\hat { f } ( x ^ { ( t - 1 ) } ) } \\\\ & { \\bullet \\ \\nabla \\hat { f } ( x ^ { ( t ) } ) / \\left( \\left| \\nabla \\hat { f } ( x ^ { ( \\operatorname* { m a x } ( t - 1 , 0 ) ) } ) \\right| + 1 \\right) } \\\\ & { \\bullet \\ \\frac { \\left| x ^ { ( \\operatorname* { m a x } ( t - 1 , 1 ) ) } - x ^ { ( \\operatorname* { m a x } ( t - 2 , 0 ) ) } \\right| } { \\left| x ^ { ( t ) } - x ^ { ( t - 1 ) } \\right| + 0 . 1 } } \\end{array}", + "type": "interline_equation", + "image_path": "1f917eed8b0253b0c9e42f56b3a726b04c793a3b7caf3b66cb9ae38147b4b29a.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 130, + 513, + 288, + 526.2 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 130, + 526.2, + 288, + 539.4000000000001 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 130, + 539.4000000000001, + 288, + 552.6000000000001 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 130, + 552.6000000000001, + 288, + 565.8000000000002 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 130, + 565.8000000000002, + 288, + 579.0000000000002 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 591, + 201, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 201, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 201, + 605 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 615, + 503, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 614, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 629 + ], + "score": 1.0, + "content": "For clarity, we will refer to training of the optimization algorithm as “meta-training” to differentiate", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 626, + 400, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 400, + 640 + ], + "score": 1.0, + "content": "it from base-level training, which will simply be referred to as “training”.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "We meta-trained an optimization algorithm on a single objective function, which corresponds to the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "problem of training a two-layer neural net with 48 input units, 48 hidden units and 10 output units", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "on a randomly projected and normalized version of the MNIST training set with dimensionality 48", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "and unit variance in each dimension. We modelled the optimization algorithm using an recurrent", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "neural net with a single layer of 128 LSTM (Hochreiter & Schmidhuber, 1997) cells. We used a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "time horizon of 400 iterations and a mini-batch size of 64 for computing stochastic gradients and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "objective values. We evaluate the optimization algorithm on its ability to generalize to unseen objec-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "tive functions, which correspond to the problems of training neural nets on different tasks/datasets.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 96, + 479, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 96, + 479, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 96, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 128, + 96, + 479, + 188 + ], + "score": 0.971, + "type": "image", + "image_path": "3fea8480e1dc0f7dbc3c3a6dc8993cac0f6ffab3eae77436f1e3e961bbffec59.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 96, + 479, + 126.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 126.66666666666667, + 479, + 157.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 157.33333333333334, + 479, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 506, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 197, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 209 + ], + "score": 1.0, + "content": "Figure 2: Comparison of the various hand-engineered and learned algorithms on training neural nets", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "batches of size 64. The vertical axis is the true objective value and the horizontal axis represents the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 236, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 236, + 242 + ], + "score": 1.0, + "content": "iteration. Best viewed in colour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 260, + 326, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 326, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 326, + 299 + ], + "score": 0.81, + "content": "\\begin{array} { r l } & { \\bullet \\ \\overline { { \\nabla \\hat { f } ( x ^ { ( i ) } ) } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } { \\nabla \\hat { f } ( x ^ { ( j ) } ) } } \\\\ & { \\bullet \\ \\overline { { \\hat { f } ( x ^ { ( i ) } ) } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \\hat { f } ( x ^ { ( j ) } ) } \\end{array}", + "type": "interline_equation", + "image_path": "7a6761589b90ee0c7200cf58fa051acd5f2b63ec2d32a67a2f1eb5d8569dab37.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 130, + 260, + 326, + 273.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 130, + 273.0, + 326, + 286.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 130, + 286.0, + 326, + 299.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 504, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 183, + 320 + ], + "score": 1.0, + "content": "The state features", + "type": "text" + }, + { + "bbox": [ + 184, + 307, + 203, + 319 + ], + "score": 0.91, + "content": "\\Phi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 306, + 506, + 320 + ], + "score": 1.0, + "content": "consist of the relative change in the average recent objective value, the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "average recent gradient normalized by the magnitude of the a previous average recent gradient and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 330, + 497, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 497, + 342 + ], + "score": 1.0, + "content": "a previous change in average recent iterate relative to the current change in average recent iterate:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 306, + 506, + 342 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 131, + 349, + 358, + 426 + ], + "lines": [ + { + "bbox": [ + 131, + 349, + 358, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 358, + 426 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\bullet \\{ ( \\hat { f } ( x ^ { ( t - 5 i ) } ) - \\widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) ) / \\widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) \\} _ { i = 0 } ^ { 2 4 } } \\\\ & { \\bullet \\{ \\overline { { \\nabla \\hat { f } ( x ^ { ( t - 5 i ) } ) } } / ( | \\overline { { \\nabla \\hat { f } ( x ^ { ( m a x ( t - 5 ( i + 1 ) , t \\mathrm { m o d } 5 ) ) } ) } } | + 1 ) \\} _ { i = 0 } ^ { 2 5 } } \\\\ & \\bullet \\{ \\frac { | \\frac { \\overline { { x ^ { ( \\mathrm { m a x } ( t - 5 ( i + 1 ) , t \\mathrm { m o d } 5 + 5 ) ) } } } - \\overline { { x ^ { ( \\mathrm { m a x } ( t - 5 ( i + 2 ) , t \\mathrm { m o d } 5 ) ) } } } } { | \\overline { { x ^ { ( t - 5 i ) } } } - \\overline { { x ^ { ( t - 5 ( i + 1 ) ) } } } | + 0 . 1 } \\} _ { i = 0 } ^ { 2 4 } \\} _ { i = 0 } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "1c2af21e865a7d14970ed22ee4629d67c1f8d6b7a64a8e247bf2ff52147bdb2e.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 131, + 349, + 358, + 361.8333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 131, + 361.8333333333333, + 358, + 374.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 131, + 374.66666666666663, + 358, + 387.49999999999994 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 131, + 387.49999999999994, + 358, + 400.33333333333326 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 131, + 400.33333333333326, + 358, + 413.1666666666666 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 131, + 413.1666666666666, + 358, + 425.9999999999999 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 432, + 503, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 432, + 504, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 504, + 444 + ], + "score": 1.0, + "content": "Note that all operations are applied element-wise. Also, whenever a feature becomes undefined (i.e.:", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 444, + 442, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 442, + 455 + ], + "score": 1.0, + "content": "when the time step index becomes negative), it is replaced with the all-zeros vector.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 432, + 504, + 455 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 473 + ], + "score": 1.0, + "content": "Unlike state features, which are only used when training the optimization algorithm, observation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 140, + 484 + ], + "score": 1.0, + "content": "features", + "type": "text" + }, + { + "bbox": [ + 141, + 471, + 159, + 483 + ], + "score": 0.91, + "content": "\\Psi ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "are used both during training and at test time. Consequently, we use noisier observation", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "features that can be computed more efficiently and require less memory overhead. The observation", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 492, + 240, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 240, + 507 + ], + "score": 1.0, + "content": "features consist of the following:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 459, + 506, + 507 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 513, + 288, + 579 + ], + "lines": [ + { + "bbox": [ + 130, + 513, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 288, + 579 + ], + "score": 0.91, + "content": "\\begin{array} { r l } { \\bullet } & { \\left( \\hat { f } ( x ^ { ( t ) } ) - \\hat { f } ( x ^ { ( t - 1 ) } ) \\right) / \\hat { f } ( x ^ { ( t - 1 ) } ) } \\\\ & { \\bullet \\ \\nabla \\hat { f } ( x ^ { ( t ) } ) / \\left( \\left| \\nabla \\hat { f } ( x ^ { ( \\operatorname* { m a x } ( t - 1 , 0 ) ) } ) \\right| + 1 \\right) } \\\\ & { \\bullet \\ \\frac { \\left| x ^ { ( \\operatorname* { m a x } ( t - 1 , 1 ) ) } - x ^ { ( \\operatorname* { m a x } ( t - 2 , 0 ) ) } \\right| } { \\left| x ^ { ( t ) } - x ^ { ( t - 1 ) } \\right| + 0 . 1 } } \\end{array}", + "type": "interline_equation", + "image_path": "1f917eed8b0253b0c9e42f56b3a726b04c793a3b7caf3b66cb9ae38147b4b29a.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 130, + 513, + 288, + 526.2 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 130, + 526.2, + 288, + 539.4000000000001 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 130, + 539.4000000000001, + 288, + 552.6000000000001 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 130, + 552.6000000000001, + 288, + 565.8000000000002 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 130, + 565.8000000000002, + 288, + 579.0000000000002 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 591, + 201, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 201, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 201, + 605 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 615, + 503, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 614, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 629 + ], + "score": 1.0, + "content": "For clarity, we will refer to training of the optimization algorithm as “meta-training” to differentiate", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 626, + 400, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 400, + 640 + ], + "score": 1.0, + "content": "it from base-level training, which will simply be referred to as “training”.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 614, + 505, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "We meta-trained an optimization algorithm on a single objective function, which corresponds to the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "problem of training a two-layer neural net with 48 input units, 48 hidden units and 10 output units", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "on a randomly projected and normalized version of the MNIST training set with dimensionality 48", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "and unit variance in each dimension. We modelled the optimization algorithm using an recurrent", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "neural net with a single layer of 128 LSTM (Hochreiter & Schmidhuber, 1997) cells. We used a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "time horizon of 400 iterations and a mini-batch size of 64 for computing stochastic gradients and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "objective values. We evaluate the optimization algorithm on its ability to generalize to unseen objec-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "tive functions, which correspond to the problems of training neural nets on different tasks/datasets.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 644, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 97, + 479, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 97, + 479, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 97, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 129, + 97, + 479, + 188 + ], + "score": 0.969, + "type": "image", + "image_path": "dba70d81f44e9b66aeedbf35fb705247b4161ee34db83bed3b4ba0993a37fd89.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 97, + 479, + 127.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 127.33333333333333, + 479, + 157.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 157.66666666666666, + 479, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 506, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "score": 1.0, + "content": "Figure 3: Comparison of the various hand-engineered and learned algorithms on training neural nets", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 208, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 506, + 220 + ], + "score": 1.0, + "content": "with 48 input and hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-batches of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "size 10. The vertical axis is the true objective value and the horizontal axis represents the iteration.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 199, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 199, + 242 + ], + "score": 1.0, + "content": "Best viewed in colour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 128, + 279, + 479, + 371 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 279, + 479, + 371 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 128, + 279, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 128, + 279, + 479, + 371 + ], + "score": 0.972, + "type": "image", + "image_path": "afba0d91683764d4cd4161719336a9f48fc43fabd99a2b8440e62626e6cabcb3.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 128, + 279, + 479, + 309.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 128, + 309.6666666666667, + 479, + 340.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 128, + 340.33333333333337, + 479, + 371.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 380, + 505, + 424 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "score": 1.0, + "content": "Figure 4: Comparison of the various hand-engineered and learned algorithms on training neural nets", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "batches of size 10. The vertical axis is the true objective value and the horizontal axis represents the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 413, + 236, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 236, + 425 + ], + "score": 1.0, + "content": "iteration. Best viewed in colour.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + } + ], + "index": 9.75 + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "score": 1.0, + "content": "We evaluate the learned optimization algorithm on three datasets, the Toronto Faces Dataset (TFD),", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "CIFAR-10 and CIFAR-100. These datasets are chosen for their very different characteristics from", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "MNIST and each other: TFD contains 3300 grayscale images that have relatively little variation and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "has seven different categories, whereas CIFAR-100 contains 50,000 colour images that have varied", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 285, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 285, + 514 + ], + "score": 1.0, + "content": "appearance and has 100 different categories.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "All algorithms are tuned on the training objective function. For hand-engineered algorithms, this", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "entails choosing the best hyperparameters; for learned algorithms, this entails meta-training on the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 540, + 504, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 504, + 552 + ], + "score": 1.0, + "content": "objective function. We compare to the seven hand-engineered algorithms: stochastic gradient de-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "scent, momentum, conjugate gradient, L-BFGS, ADAM, AdaGrad and RMSprop. In addition, we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "compare to an optimization algorithm meta-trained using the method described in (Andrychowicz", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "et al., 2016) on the same training objective function (training two-layer neural net on randomly", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "projected and normalized MNIST) under the same setting (a time horizon of 400 iterations and a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 595, + 199, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 199, + 605 + ], + "score": 1.0, + "content": "mini-batch size of 64).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "First, we examine the performance of various optimization algorithms on similar objective functions.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "The optimization problems under consideration are those for training neural nets that have the same", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "number of input and hidden units (48 and 48) as those used during meta-training. The number of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "output units varies with the number of categories in each dataset. We use the same mini-batch size", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "as that used during meta-training. As shown in Figure 1, the optimization algorithm meta-trained", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "using our method (which we will refer to as Predicted Step Descent) consistently descends to the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "optimum the fastest across all datasets. On the other hand, other algorithms are not as consistent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "and the relative ranking of other algorithms varies by dataset. This suggests that Predicted Step", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Descent has learned to be robust to variations in the data distributions, despite being trained on only", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "one objective function, which is associated with a very specific data distribution that characterizes", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "MNIST. It is also interesting to note that while the algorithm meta-trained using (Andrychowicz", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 97, + 479, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 97, + 479, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 97, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 129, + 97, + 479, + 188 + ], + "score": 0.969, + "type": "image", + "image_path": "dba70d81f44e9b66aeedbf35fb705247b4161ee34db83bed3b4ba0993a37fd89.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 97, + 479, + 127.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 127.33333333333333, + 479, + 157.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 157.66666666666666, + 479, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 506, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 210 + ], + "score": 1.0, + "content": "Figure 3: Comparison of the various hand-engineered and learned algorithms on training neural nets", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 208, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 506, + 220 + ], + "score": 1.0, + "content": "with 48 input and hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-batches of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "size 10. The vertical axis is the true objective value and the horizontal axis represents the iteration.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 199, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 199, + 242 + ], + "score": 1.0, + "content": "Best viewed in colour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 128, + 279, + 479, + 371 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 279, + 479, + 371 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 128, + 279, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 128, + 279, + 479, + 371 + ], + "score": 0.972, + "type": "image", + "image_path": "afba0d91683764d4cd4161719336a9f48fc43fabd99a2b8440e62626e6cabcb3.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 128, + 279, + 479, + 309.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 128, + 309.6666666666667, + 479, + 340.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 128, + 340.33333333333337, + 479, + 371.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 380, + 505, + 424 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "score": 1.0, + "content": "Figure 4: Comparison of the various hand-engineered and learned algorithms on training neural nets", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 with mini-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "batches of size 10. The vertical axis is the true objective value and the horizontal axis represents the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 413, + 236, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 236, + 425 + ], + "score": 1.0, + "content": "iteration. Best viewed in colour.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + } + ], + "index": 9.75 + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "score": 1.0, + "content": "We evaluate the learned optimization algorithm on three datasets, the Toronto Faces Dataset (TFD),", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "CIFAR-10 and CIFAR-100. These datasets are chosen for their very different characteristics from", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "MNIST and each other: TFD contains 3300 grayscale images that have relatively little variation and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "has seven different categories, whereas CIFAR-100 contains 50,000 colour images that have varied", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 285, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 285, + 514 + ], + "score": 1.0, + "content": "appearance and has 100 different categories.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 455, + 505, + 514 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "All algorithms are tuned on the training objective function. For hand-engineered algorithms, this", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "entails choosing the best hyperparameters; for learned algorithms, this entails meta-training on the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 540, + 504, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 504, + 552 + ], + "score": 1.0, + "content": "objective function. We compare to the seven hand-engineered algorithms: stochastic gradient de-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "scent, momentum, conjugate gradient, L-BFGS, ADAM, AdaGrad and RMSprop. In addition, we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "compare to an optimization algorithm meta-trained using the method described in (Andrychowicz", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "et al., 2016) on the same training objective function (training two-layer neural net on randomly", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "projected and normalized MNIST) under the same setting (a time horizon of 400 iterations and a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 595, + 199, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 199, + 605 + ], + "score": 1.0, + "content": "mini-batch size of 64).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 517, + 506, + 605 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "First, we examine the performance of various optimization algorithms on similar objective functions.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "The optimization problems under consideration are those for training neural nets that have the same", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "number of input and hidden units (48 and 48) as those used during meta-training. The number of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "output units varies with the number of categories in each dataset. We use the same mini-batch size", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "as that used during meta-training. As shown in Figure 1, the optimization algorithm meta-trained", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "using our method (which we will refer to as Predicted Step Descent) consistently descends to the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "optimum the fastest across all datasets. On the other hand, other algorithms are not as consistent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 686, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 702 + ], + "score": 1.0, + "content": "and the relative ranking of other algorithms varies by dataset. This suggests that Predicted Step", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Descent has learned to be robust to variations in the data distributions, despite being trained on only", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "one objective function, which is associated with a very specific data distribution that characterizes", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "MNIST. It is also interesting to note that while the algorithm meta-trained using (Andrychowicz", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "score": 1.0, + "content": "et al., 2016) (which we will refer to as L2LBGDBGD) performs well on CIFAR, it is unable to", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 281, + 220, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 220, + 293 + ], + "score": 1.0, + "content": "reach the optimum on TFD.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 612, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 97, + 478, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 97, + 478, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 97, + 478, + 188 + ], + "spans": [ + { + "bbox": [ + 128, + 97, + 478, + 188 + ], + "score": 0.965, + "type": "image", + "image_path": "2226d92233c78e2cc8398fe5882a7f24065a5392b547b417b88d986cc9309d20.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 97, + 478, + 127.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 127.33333333333333, + 478, + 157.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 157.66666666666666, + 478, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 506, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "score": 1.0, + "content": "Figure 5: Comparison of the various hand-engineered and learned algorithms on training neural", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 208, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 219 + ], + "score": 1.0, + "content": "nets with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "800 iterations with mini-batches of size 64. The vertical axis is the true objective value and the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 356, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 356, + 242 + ], + "score": 1.0, + "content": "horizontal axis represents the iteration. Best viewed in colour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 504, + 292 + ], + "lines": [ + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "score": 1.0, + "content": "et al., 2016) (which we will refer to as L2LBGDBGD) performs well on CIFAR, it is unable to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 281, + 220, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 220, + 293 + ], + "score": 1.0, + "content": "reach the optimum on TFD.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "Next, we change the architecture of the neural nets and see if Predicted Step Descent generalizes to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 309, + 504, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 504, + 320 + ], + "score": 1.0, + "content": "the new architecture. We increase the number of input units to 100 and the number of hidden units", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 318, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 318, + 506, + 334 + ], + "score": 1.0, + "content": "to 200, so that the number of parameters is roughly increased by a factor of 8. As shown in Figure 2,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 329, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 506, + 345 + ], + "score": 1.0, + "content": "Predicted Step Descent consistently outperforms other algorithms on each dataset, despite having", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "score": 1.0, + "content": "not been trained to optimize neural nets of this architecture. Interestingly, while it exhibited a bit", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 352, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 366 + ], + "score": 1.0, + "content": "of oscillation initially on TFD and CIFAR-10, it quickly recovered and overtook other algorithms,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 362, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 378 + ], + "score": 1.0, + "content": "which is reminiscent of the phenomenon reported in (Li & Malik, 2016) for low-dimensional op-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "timization problems. This suggests that it has learned to detect when it is performing poorly and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "knows how to change tack accordingly. L2LBGDBGD experienced difficulties on TFD and CIFAR-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 397, + 235, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 235, + 410 + ], + "score": 1.0, + "content": "10 as well, but slowly diverged.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 413, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "We now investigate how robust Predicted Step Descent is to stochasticity of the gradients. To this", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 425, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 504, + 435 + ], + "score": 1.0, + "content": "end, we take a look at its performance when we reduce the mini-batch size from 64 to 10 on both", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "the original architecture with 48 input and hidden units and the enlarged architecture with 100 input", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 445, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 460 + ], + "score": 1.0, + "content": "units and 200 hidden units. As shown in Figure 3, on the original architecture, Predicted Step", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "Descent still outperforms all other algorithms and is able to handle the increased stochasticity fairly", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "well. In contrast, conjugate gradient and L2LBGDBGD had some difficulty handling the increased", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "stochasticity on TFD and to a lesser extent, on CIFAR-10. In the former case, both diverged; in the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 491, + 360, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 360, + 503 + ], + "score": 1.0, + "content": "latter case, both were progressing slowly towards the optimum.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 507, + 504, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "On the enlarged architecture, Predicted Step Descent experienced some significant oscillations on", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "TFD and CIFAR-10, but still managed to achieve a much better objective value than all the other", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "algorithms. Many hand-engineered algorithms also experienced much greater oscillations than pre-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 540, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 506, + 553 + ], + "score": 1.0, + "content": "viously, suggesting that the optimization problems are inherently harder. L2LBGDBGD diverged", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 552, + 252, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 252, + 564 + ], + "score": 1.0, + "content": "fairly quickly on these two datasets.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 568, + 503, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "score": 1.0, + "content": "Finally, we try doubling the number of iterations. As shown in Figure 5, despite being trained over", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "a time horizon of 400 iterations, Predicted Step Descent behaves reasonably beyond the number of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 589, + 209, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 209, + 601 + ], + "score": 1.0, + "content": "iterations it is trained for.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "title", + "bbox": [ + 108, + 625, + 195, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 197, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 197, + 641 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "In this paper, we presented a new method for learning optimization algorithms for high-dimensional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "stochastic problems. We applied the method to learning an optimization algorithm for training", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "shallow neural nets. We showed that the algorithm learned using our method on the problem of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "training a neural net on MNIST generalizes to the problems of training neural nets on unrelated", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "tasks/datasets like the Toronto Faces Dataset, CIFAR-10 and CIFAR-100. We also demonstrated", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "that the learned optimization algorithm is robust to changes in the stochasticity of gradients and the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 720, + 201, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 201, + 732 + ], + "score": 1.0, + "content": "neural net architecture.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 97, + 478, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 97, + 478, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 97, + 478, + 188 + ], + "spans": [ + { + "bbox": [ + 128, + 97, + 478, + 188 + ], + "score": 0.965, + "type": "image", + "image_path": "2226d92233c78e2cc8398fe5882a7f24065a5392b547b417b88d986cc9309d20.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 128, + 97, + 478, + 127.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 128, + 127.33333333333333, + 478, + 157.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 157.66666666666666, + 478, + 188.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 197, + 506, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 209 + ], + "score": 1.0, + "content": "Figure 5: Comparison of the various hand-engineered and learned algorithms on training neural", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 208, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 219 + ], + "score": 1.0, + "content": "nets with 100 input units and 200 hidden units on (a) TFD, (b) CIFAR-10 and (c) CIFAR-100 for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "800 iterations with mini-batches of size 64. The vertical axis is the true objective value and the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 356, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 356, + 242 + ], + "score": 1.0, + "content": "horizontal axis represents the iteration. Best viewed in colour.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 504, + 292 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 269, + 506, + 293 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "Next, we change the architecture of the neural nets and see if Predicted Step Descent generalizes to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 309, + 504, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 504, + 320 + ], + "score": 1.0, + "content": "the new architecture. We increase the number of input units to 100 and the number of hidden units", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 318, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 318, + 506, + 334 + ], + "score": 1.0, + "content": "to 200, so that the number of parameters is roughly increased by a factor of 8. As shown in Figure 2,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 329, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 506, + 345 + ], + "score": 1.0, + "content": "Predicted Step Descent consistently outperforms other algorithms on each dataset, despite having", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "score": 1.0, + "content": "not been trained to optimize neural nets of this architecture. Interestingly, while it exhibited a bit", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 352, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 366 + ], + "score": 1.0, + "content": "of oscillation initially on TFD and CIFAR-10, it quickly recovered and overtook other algorithms,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 362, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 378 + ], + "score": 1.0, + "content": "which is reminiscent of the phenomenon reported in (Li & Malik, 2016) for low-dimensional op-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "timization problems. This suggests that it has learned to detect when it is performing poorly and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "knows how to change tack accordingly. L2LBGDBGD experienced difficulties on TFD and CIFAR-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 397, + 235, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 235, + 410 + ], + "score": 1.0, + "content": "10 as well, but slowly diverged.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 298, + 506, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 413, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "We now investigate how robust Predicted Step Descent is to stochasticity of the gradients. To this", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 425, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 504, + 435 + ], + "score": 1.0, + "content": "end, we take a look at its performance when we reduce the mini-batch size from 64 to 10 on both", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "the original architecture with 48 input and hidden units and the enlarged architecture with 100 input", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 445, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 460 + ], + "score": 1.0, + "content": "units and 200 hidden units. As shown in Figure 3, on the original architecture, Predicted Step", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "Descent still outperforms all other algorithms and is able to handle the increased stochasticity fairly", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "well. In contrast, conjugate gradient and L2LBGDBGD had some difficulty handling the increased", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "stochasticity on TFD and to a lesser extent, on CIFAR-10. In the former case, both diverged; in the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 491, + 360, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 360, + 503 + ], + "score": 1.0, + "content": "latter case, both were progressing slowly towards the optimum.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 414, + 506, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 507, + 504, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "On the enlarged architecture, Predicted Step Descent experienced some significant oscillations on", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "TFD and CIFAR-10, but still managed to achieve a much better objective value than all the other", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "algorithms. Many hand-engineered algorithms also experienced much greater oscillations than pre-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 540, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 506, + 553 + ], + "score": 1.0, + "content": "viously, suggesting that the optimization problems are inherently harder. L2LBGDBGD diverged", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 552, + 252, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 252, + 564 + ], + "score": 1.0, + "content": "fairly quickly on these two datasets.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 506, + 506, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 568, + 503, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 581 + ], + "score": 1.0, + "content": "Finally, we try doubling the number of iterations. As shown in Figure 5, despite being trained over", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "a time horizon of 400 iterations, Predicted Step Descent behaves reasonably beyond the number of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 589, + 209, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 209, + 601 + ], + "score": 1.0, + "content": "iterations it is trained for.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 567, + 505, + 601 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 625, + 195, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 197, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 197, + 641 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "In this paper, we presented a new method for learning optimization algorithms for high-dimensional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "stochastic problems. We applied the method to learning an optimization algorithm for training", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "shallow neural nets. We showed that the algorithm learned using our method on the problem of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "training a neural net on MNIST generalizes to the problems of training neural nets on unrelated", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "tasks/datasets like the Toronto Faces Dataset, CIFAR-10 and CIFAR-100. We also demonstrated", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "that the learned optimization algorithm is robust to changes in the stochasticity of gradients and the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 720, + 201, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 201, + 732 + ], + "score": 1.0, + "content": "neural net architecture.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 654, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "and Nando de Freitas. Learning to learn by gradient descent by gradient descent. arXiv preprint", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 219, + 133 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 219, + 133 + ], + "score": 1.0, + "content": "arXiv:1606.04474, 2016.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 141, + 505, + 196 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "Jonathan Baxter, Rich Caruana, Tom Mitchell, Lorien Y Pratt, Daniel L Silver, and Sebastian Thrun.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 152, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 505, + 164 + ], + "score": 1.0, + "content": "NIPS 1995 workshop on learning to learn: Knowledge consolidation and transfer in induc-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 504, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 504, + 176 + ], + "score": 1.0, + "content": "tive systems. https://web.archive.org/web/20000618135816/http://www.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 174, + 504, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 504, + 186 + ], + "score": 1.0, + "content": "cs.cmu.edu/afs/cs.cmu.edu/user/caruana/pub/transfer.html, 1995. Ac-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 185, + 200, + 196 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 200, + 196 + ], + "score": 1.0, + "content": "cessed: 2015-12-05.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 108, + 204, + 502, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Y Bengio, S Bengio, and J Cloutier. Learning a synaptic learning rule. In Neural Networks, 1991.,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 215, + 477, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 215, + 477, + 228 + ], + "score": 1.0, + "content": "IJCNN-91-Seattle International Joint Conference on, volume 2, pp. 969–vol. IEEE, 1991.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 503, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 233, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 505, + 248 + ], + "score": 1.0, + "content": "Yoshua Bengio. Gradient-based optimization of hyperparameters. Neural computation, 12(8):1889–", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 245, + 167, + 257 + ], + "spans": [ + { + "bbox": [ + 117, + 245, + 167, + 257 + ], + "score": 1.0, + "content": "1900, 2000.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 503, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. The Journal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 276, + 333, + 288 + ], + "spans": [ + { + "bbox": [ + 116, + 276, + 333, + 288 + ], + "score": 1.0, + "content": "of Machine Learning Research, 13(1):281–305, 2012.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 105, + 294, + 503, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "score": 1.0, + "content": "James S Bergstra, Remi Bardenet, Yoshua Bengio, and Bal ´ azs K ´ egl. Algorithms for hyper-parameter ´", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 306, + 488, + 319 + ], + "spans": [ + { + "bbox": [ + 117, + 306, + 488, + 319 + ], + "score": 1.0, + "content": "optimization. In Advances in Neural Information Processing Systems, pp. 2546–2554, 2011.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 325, + 504, + 359 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "M Bray, E Koller-Meier, P Muller, L Van Gool, and NN Schraudolph. 3D hand tracking by rapid", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 116, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "stochastic gradient descent using a skinning model. In Visual Media Production, 2004.(CVMP).", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 347, + 325, + 360 + ], + "spans": [ + { + "bbox": [ + 116, + 347, + 325, + 360 + ], + "score": 1.0, + "content": "1st European Conference on, pp. 59–68. IET, 2004.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 503, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 504, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 504, + 379 + ], + "score": 1.0, + "content": "Pavel Brazdil, Christophe Giraud Carrier, Carlos Soares, and Ricardo Vilalta. Metalearning: appli-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 378, + 386, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 386, + 389 + ], + "score": 1.0, + "content": "cations to data mining. Springer Science & Business Media, 2008.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 504, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 411 + ], + "score": 1.0, + "content": "Pavel B Brazdil, Carlos Soares, and Joaquim Pinto Da Costa. Ranking learning algorithms: Using", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 407, + 491, + 420 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 491, + 420 + ], + "score": 1.0, + "content": "ibl and meta-learning on accuracy and time results. Machine Learning, 50(3):251–277, 2003.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 504, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "Christian Daniel, Jonathan Taylor, and Sebastian Nowozin. Learning step size controllers for robust", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 438, + 463, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 463, + 450 + ], + "score": 1.0, + "content": "neural network training. In Thirtieth AAAI Conference on Artificial Intelligence, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 504, + 479 + ], + "lines": [ + { + "bbox": [ + 104, + 454, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 104, + 454, + 506, + 471 + ], + "score": 1.0, + "content": "Justin Domke. Generic methods for optimization-based modeling. In AISTATS, volume 22, pp.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 468, + 182, + 480 + ], + "spans": [ + { + "bbox": [ + 116, + 468, + 182, + 480 + ], + "score": 1.0, + "content": "318–326, 2012.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 498, + 483, + 510 + ], + "spans": [ + { + "bbox": [ + 116, + 498, + 483, + 510 + ], + "score": 1.0, + "content": "stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 504, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 531 + ], + "score": 1.0, + "content": "Matthias Feurer, Jost Tobias Springenberg, and Frank Hutter. Initializing bayesian hyperparameter", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 529, + 372, + 541 + ], + "spans": [ + { + "bbox": [ + 116, + 529, + 372, + 541 + ], + "score": 1.0, + "content": "optimization via meta-learning. In AAAI, pp. 1128–1135, 2015.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 105, + 547, + 504, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "Jie Fu, Zichuan Lin, Miao Liu, Nicholas Leonard, Jiashi Feng, and Tat-Seng Chua. Deep q-networks", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 558, + 493, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 493, + 571 + ], + "score": 1.0, + "content": "for accelerating the training of deep neural networks. arXiv preprint arXiv:1606.01467, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "Karol Gregor and Yann LeCun. Learning fast approximations of sparse coding. In Proceedings of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 588, + 471, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 588, + 471, + 601 + ], + "score": 1.0, + "content": "the 27th International Conference on Machine Learning (ICML-10), pp. 399–406, 2010.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 608, + 504, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 621 + ], + "score": 1.0, + "content": "Samantha Hansen. Using deep q-learning to control optimization hyperparameters. arXiv preprint", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 619, + 219, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 219, + 630 + ], + "score": 1.0, + "content": "arXiv:1602.04062, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 638, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 649, + 191, + 660 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 191, + 660 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 504, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 504, + 681 + ], + "score": 1.0, + "content": "Sepp Hochreiter, A Steven Younger, and Peter R Conwell. Learning to learn using gradient descent.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 678, + 462, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 462, + 691 + ], + "score": 1.0, + "content": "In International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 114, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "general algorithm configuration. In Learning and Intelligent Optimization, pp. 507–523. Springer,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "and Nando de Freitas. Learning to learn by gradient descent by gradient descent. arXiv preprint", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 219, + 133 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 219, + 133 + ], + "score": 1.0, + "content": "arXiv:1606.04474, 2016.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 99, + 505, + 133 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 141, + 505, + 196 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "Jonathan Baxter, Rich Caruana, Tom Mitchell, Lorien Y Pratt, Daniel L Silver, and Sebastian Thrun.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 152, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 152, + 505, + 164 + ], + "score": 1.0, + "content": "NIPS 1995 workshop on learning to learn: Knowledge consolidation and transfer in induc-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 504, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 504, + 176 + ], + "score": 1.0, + "content": "tive systems. https://web.archive.org/web/20000618135816/http://www.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 174, + 504, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 504, + 186 + ], + "score": 1.0, + "content": "cs.cmu.edu/afs/cs.cmu.edu/user/caruana/pub/transfer.html, 1995. Ac-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 185, + 200, + 196 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 200, + 196 + ], + "score": 1.0, + "content": "cessed: 2015-12-05.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 142, + 505, + 196 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 204, + 502, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Y Bengio, S Bengio, and J Cloutier. Learning a synaptic learning rule. In Neural Networks, 1991.,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 215, + 477, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 215, + 477, + 228 + ], + "score": 1.0, + "content": "IJCNN-91-Seattle International Joint Conference on, volume 2, pp. 969–vol. IEEE, 1991.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 204, + 505, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 503, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 233, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 505, + 248 + ], + "score": 1.0, + "content": "Yoshua Bengio. Gradient-based optimization of hyperparameters. Neural computation, 12(8):1889–", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 245, + 167, + 257 + ], + "spans": [ + { + "bbox": [ + 117, + 245, + 167, + 257 + ], + "score": 1.0, + "content": "1900, 2000.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 233, + 505, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 503, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. The Journal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 276, + 333, + 288 + ], + "spans": [ + { + "bbox": [ + 116, + 276, + 333, + 288 + ], + "score": 1.0, + "content": "of Machine Learning Research, 13(1):281–305, 2012.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 264, + 505, + 288 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 294, + 503, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "score": 1.0, + "content": "James S Bergstra, Remi Bardenet, Yoshua Bengio, and Bal ´ azs K ´ egl. Algorithms for hyper-parameter ´", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 306, + 488, + 319 + ], + "spans": [ + { + "bbox": [ + 117, + 306, + 488, + 319 + ], + "score": 1.0, + "content": "optimization. In Advances in Neural Information Processing Systems, pp. 2546–2554, 2011.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 294, + 505, + 319 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 325, + 504, + 359 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "M Bray, E Koller-Meier, P Muller, L Van Gool, and NN Schraudolph. 3D hand tracking by rapid", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 116, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "stochastic gradient descent using a skinning model. In Visual Media Production, 2004.(CVMP).", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 347, + 325, + 360 + ], + "spans": [ + { + "bbox": [ + 116, + 347, + 325, + 360 + ], + "score": 1.0, + "content": "1st European Conference on, pp. 59–68. IET, 2004.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 324, + 505, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 503, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 504, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 504, + 379 + ], + "score": 1.0, + "content": "Pavel Brazdil, Christophe Giraud Carrier, Carlos Soares, and Ricardo Vilalta. Metalearning: appli-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 378, + 386, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 386, + 389 + ], + "score": 1.0, + "content": "cations to data mining. Springer Science & Business Media, 2008.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 366, + 504, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 504, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 411 + ], + "score": 1.0, + "content": "Pavel B Brazdil, Carlos Soares, and Joaquim Pinto Da Costa. Ranking learning algorithms: Using", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 407, + 491, + 420 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 491, + 420 + ], + "score": 1.0, + "content": "ibl and meta-learning on accuracy and time results. Machine Learning, 50(3):251–277, 2003.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 394, + 506, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 504, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "Christian Daniel, Jonathan Taylor, and Sebastian Nowozin. Learning step size controllers for robust", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 438, + 463, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 463, + 450 + ], + "score": 1.0, + "content": "neural network training. In Thirtieth AAAI Conference on Artificial Intelligence, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 427, + 505, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 504, + 479 + ], + "lines": [ + { + "bbox": [ + 104, + 454, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 104, + 454, + 506, + 471 + ], + "score": 1.0, + "content": "Justin Domke. Generic methods for optimization-based modeling. In AISTATS, volume 22, pp.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 468, + 182, + 480 + ], + "spans": [ + { + "bbox": [ + 116, + 468, + 182, + 480 + ], + "score": 1.0, + "content": "318–326, 2012.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 454, + 506, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 498, + 483, + 510 + ], + "spans": [ + { + "bbox": [ + 116, + 498, + 483, + 510 + ], + "score": 1.0, + "content": "stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 487, + 505, + 510 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 516, + 504, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 531 + ], + "score": 1.0, + "content": "Matthias Feurer, Jost Tobias Springenberg, and Frank Hutter. Initializing bayesian hyperparameter", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 529, + 372, + 541 + ], + "spans": [ + { + "bbox": [ + 116, + 529, + 372, + 541 + ], + "score": 1.0, + "content": "optimization via meta-learning. In AAAI, pp. 1128–1135, 2015.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 515, + 506, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 547, + 504, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "Jie Fu, Zichuan Lin, Miao Liu, Nicholas Leonard, Jiashi Feng, and Tat-Seng Chua. Deep q-networks", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 558, + 493, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 493, + 571 + ], + "score": 1.0, + "content": "for accelerating the training of deep neural networks. arXiv preprint arXiv:1606.01467, 2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 547, + 505, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "Karol Gregor and Yann LeCun. Learning fast approximations of sparse coding. In Proceedings of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 588, + 471, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 588, + 471, + 601 + ], + "score": 1.0, + "content": "the 27th International Conference on Machine Learning (ICML-10), pp. 399–406, 2010.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 577, + 506, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 608, + 504, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 621 + ], + "score": 1.0, + "content": "Samantha Hansen. Using deep q-learning to control optimization hyperparameters. arXiv preprint", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 619, + 219, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 219, + 630 + ], + "score": 1.0, + "content": "arXiv:1602.04062, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 606, + 506, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 638, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 649, + 191, + 660 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 191, + 660 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 637, + 505, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 504, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 504, + 681 + ], + "score": 1.0, + "content": "Sepp Hochreiter, A Steven Younger, and Peter R Conwell. Learning to learn using gradient descent.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 678, + 462, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 462, + 691 + ], + "score": 1.0, + "content": "In International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 668, + 504, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Frank Hutter, Holger H Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 114, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "general algorithm configuration. In Learning and Intelligent Optimization, pp. 507–523. Springer,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 142, + 732 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 214, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 214, + 105 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuo-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 342, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 342, + 136 + ], + "score": 1.0, + "content": "motor policies. arXiv preprint arXiv:1504.00702, 2015.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 425, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 424, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 424, + 155 + ], + "score": 1.0, + "content": "Ke Li and Jitendra Malik. Learning to optimize. CoRR, abs/1606.01885, 2016.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 160, + 503, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 161, + 504, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 504, + 174 + ], + "score": 1.0, + "content": "Dougal Maclaurin, David Duvenaud, and Ryan P Adams. Gradient-based hyperparameter optimiza-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 172, + 411, + 185 + ], + "spans": [ + { + "bbox": [ + 116, + 172, + 411, + 185 + ], + "score": 1.0, + "content": "tion through reversible learning. arXiv preprint arXiv:1502.03492, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 190, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "Paul L Ruvolo, Ian Fasel, and Javier R Movellan. Optimization on a budget: A reinforcement", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 202, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 505, + 215 + ], + "score": 1.0, + "content": "learning approach. In Advances in Neural Information Processing Systems, pp. 1385–1392, 2009.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 220, + 498, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 496, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 496, + 234 + ], + "score": 1.0, + "content": "Jurgen Schmidhuber. Optimal ordered problem solver. ¨ Machine Learning, 54(3):211–254, 2004.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 104, + 239, + 504, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical bayesian optimization of machine", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 115, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "learning algorithms. In Advances in neural information processing systems, pp. 2951–2959, 2012.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "score": 1.0, + "content": "Pablo Sprechmann, Roee Litman, Tal Ben Yakar, Alexander M Bronstein, and Guillermo Sapiro.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 116, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "Supervised sparse analysis and synthesis operators. In Advances in Neural Information Process-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 292, + 248, + 303 + ], + "spans": [ + { + "bbox": [ + 116, + 292, + 248, + 303 + ], + "score": 1.0, + "content": "ing Systems, pp. 908–916, 2013.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 105, + 310, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Kevin Swersky, Jasper Snoek, and Ryan P Adams. Multi-task bayesian optimization. In Advances", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 321, + 375, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 375, + 334 + ], + "score": 1.0, + "content": "in neural information processing systems, pp. 2004–2012, 2013.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 105, + 340, + 495, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 495, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 495, + 353 + ], + "score": 1.0, + "content": "Sebastian Thrun and Lorien Pratt. Learning to learn. Springer Science & Business Media, 2012.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 105, + 359, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 374 + ], + "score": 1.0, + "content": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 116, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2), 2012.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 104, + 389, + 504, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "Ricardo Vilalta and Youssef Drissi. A perspective view and survey of meta-learning. Artificial", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 399, + 279, + 412 + ], + "spans": [ + { + "bbox": [ + 115, + 399, + 279, + 412 + ], + "score": 1.0, + "content": "Intelligence Review, 18(2):77–95, 2002.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 104, + 419, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "score": 1.0, + "content": "Huahua Wang and Arindam Banerjee. Bregman alternating direction method of multipliers. CoRR,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 430, + 205, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 205, + 441 + ], + "score": 1.0, + "content": "abs/1306.3203, 2014.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 214, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 214, + 105 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 80, + 506, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuo-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 342, + 136 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 342, + 136 + ], + "score": 1.0, + "content": "motor policies. arXiv preprint arXiv:1504.00702, 2015.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 505, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 425, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 424, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 424, + 155 + ], + "score": 1.0, + "content": "Ke Li and Jitendra Malik. Learning to optimize. CoRR, abs/1606.01885, 2016.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 141, + 424, + 155 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 160, + 503, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 161, + 504, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 504, + 174 + ], + "score": 1.0, + "content": "Dougal Maclaurin, David Duvenaud, and Ryan P Adams. Gradient-based hyperparameter optimiza-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 172, + 411, + 185 + ], + "spans": [ + { + "bbox": [ + 116, + 172, + 411, + 185 + ], + "score": 1.0, + "content": "tion through reversible learning. arXiv preprint arXiv:1502.03492, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 161, + 504, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 190, + 504, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "Paul L Ruvolo, Ian Fasel, and Javier R Movellan. Optimization on a budget: A reinforcement", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 202, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 505, + 215 + ], + "score": 1.0, + "content": "learning approach. In Advances in Neural Information Processing Systems, pp. 1385–1392, 2009.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 190, + 505, + 215 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 220, + 498, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 496, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 496, + 234 + ], + "score": 1.0, + "content": "Jurgen Schmidhuber. Optimal ordered problem solver. ¨ Machine Learning, 54(3):211–254, 2004.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 219, + 496, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 239, + 504, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical bayesian optimization of machine", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 115, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "learning algorithms. In Advances in neural information processing systems, pp. 2951–2959, 2012.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 239, + 505, + 263 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "score": 1.0, + "content": "Pablo Sprechmann, Roee Litman, Tal Ben Yakar, Alexander M Bronstein, and Guillermo Sapiro.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 116, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "Supervised sparse analysis and synthesis operators. In Advances in Neural Information Process-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 292, + 248, + 303 + ], + "spans": [ + { + "bbox": [ + 116, + 292, + 248, + 303 + ], + "score": 1.0, + "content": "ing Systems, pp. 908–916, 2013.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 268, + 505, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 310, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Kevin Swersky, Jasper Snoek, and Ryan P Adams. Multi-task bayesian optimization. In Advances", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 321, + 375, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 375, + 334 + ], + "score": 1.0, + "content": "in neural information processing systems, pp. 2004–2012, 2013.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 311, + 505, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 340, + 495, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 495, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 495, + 353 + ], + "score": 1.0, + "content": "Sebastian Thrun and Lorien Pratt. Learning to learn. Springer Science & Business Media, 2012.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 339, + 495, + 353 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 359, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 374 + ], + "score": 1.0, + "content": "Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 116, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "average of its recent magnitude. COURSERA: Neural networks for machine learning, 4(2), 2012.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 358, + 505, + 383 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 389, + 504, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "Ricardo Vilalta and Youssef Drissi. A perspective view and survey of meta-learning. Artificial", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 399, + 279, + 412 + ], + "spans": [ + { + "bbox": [ + 115, + 399, + 279, + 412 + ], + "score": 1.0, + "content": "Intelligence Review, 18(2):77–95, 2002.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 389, + 505, + 412 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 419, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "score": 1.0, + "content": "Huahua Wang and Arindam Banerjee. Bregman alternating direction method of multipliers. CoRR,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 430, + 205, + 441 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 205, + 441 + ], + "score": 1.0, + "content": "abs/1306.3203, 2014.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 418, + 506, + 441 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/BkM27IxR-/BkM27IxR-_model.json b/parse/train/BkM27IxR-/BkM27IxR-_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fd88f01ff2d196bd14e8f8a731affbd0e312a633 --- /dev/null +++ b/parse/train/BkM27IxR-/BkM27IxR-_model.json @@ -0,0 +1,20057 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 537, + 1302, + 537, + 1302, + 904, + 398, + 904 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1678, + 1403, + 1678, + 1403, + 1896, + 298, + 1896 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1448, + 1403, + 1448, + 1403, + 1662, + 298, + 1662 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1219, + 1403, + 1219, + 1403, + 1432, + 298, + 1432 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1049, + 1402, + 1049, + 1402, + 1203, + 298, + 1203 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1911, + 1401, + 1911, + 1401, + 2033, + 300, + 2033 + ], + "score": 0.973 + }, + { + "category_id": 0, + "poly": [ + 301, + 219, + 1115, + 219, + 1115, + 269, + 301, + 269 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 313, + 322, + 680, + 322, + 680, + 383, + 313, + 383 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 302, + 976, + 573, + 976, + 573, + 1011, + 302, + 1011 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 773, + 464, + 927, + 464, + 927, + 498, + 773, + 498 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.629 + }, + { + "category_id": 13, + "poly": [ + 1308, + 1770, + 1353, + 1770, + 1353, + 1802, + 1308, + 1802 + ], + "score": 0.91, + "latex": "\\boldsymbol { x } ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 766, + 1805, + 808, + 1805, + 808, + 1832, + 766, + 1832 + ], + "score": 0.87, + "latex": "\\Delta x" + }, + { + "category_id": 13, + "poly": [ + 811, + 1948, + 829, + 1948, + 829, + 1974, + 811, + 1974 + ], + "score": 0.86, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1805, + 1257, + 1805, + 1257, + 1835, + 1236, + 1835 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 524, + 1913, + 545, + 1913, + 545, + 1944, + 524, + 1944 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 803, + 1774, + 846, + 1774, + 846, + 1801, + 803, + 1801 + ], + "score": 0.85, + "latex": "\\Delta x" + }, + { + "category_id": 13, + "poly": [ + 1360, + 2004, + 1401, + 2004, + 1401, + 2030, + 1360, + 2030 + ], + "score": 0.84, + "latex": "\\Delta x" + }, + { + "category_id": 13, + "poly": [ + 1055, + 1946, + 1073, + 1946, + 1073, + 1974, + 1055, + 1974 + ], + "score": 0.83, + "latex": "\\phi" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 218.0, + 1121.0, + 218.0, + 1121.0, + 274.0, + 291.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 973.0, + 579.0, + 973.0, + 579.0, + 1020.0, + 294.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 463.0, + 932.0, + 463.0, + 932.0, + 501.0, + 770.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 538.0, + 1305.0, + 538.0, + 1305.0, + 572.0, + 393.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 570.0, + 1307.0, + 570.0, + 1307.0, + 606.0, + 393.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 601.0, + 1305.0, + 601.0, + 1305.0, + 632.0, + 394.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 629.0, + 1305.0, + 629.0, + 1305.0, + 666.0, + 393.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 659.0, + 1307.0, + 659.0, + 1307.0, + 695.0, + 392.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 691.0, + 1306.0, + 691.0, + 1306.0, + 727.0, + 393.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 722.0, + 1305.0, + 722.0, + 1305.0, + 756.0, + 393.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 753.0, + 1307.0, + 753.0, + 1307.0, + 788.0, + 393.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 781.0, + 1306.0, + 781.0, + 1306.0, + 820.0, + 393.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 810.0, + 1306.0, + 810.0, + 1306.0, + 850.0, + 392.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 843.0, + 1305.0, + 843.0, + 1305.0, + 878.0, + 392.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 874.0, + 970.0, + 874.0, + 970.0, + 905.0, + 395.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1675.0, + 1403.0, + 1675.0, + 1403.0, + 1715.0, + 293.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 1405.0, + 1711.0, + 1405.0, + 1744.0, + 293.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1740.0, + 1406.0, + 1740.0, + 1406.0, + 1775.0, + 292.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1772.0, + 802.0, + 1772.0, + 802.0, + 1808.0, + 292.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1772.0, + 1307.0, + 1772.0, + 1307.0, + 1808.0, + 847.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1772.0, + 1406.0, + 1772.0, + 1406.0, + 1808.0, + 1354.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 765.0, + 1803.0, + 765.0, + 1838.0, + 294.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1803.0, + 1235.0, + 1803.0, + 1235.0, + 1838.0, + 809.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1838.0, + 1258.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 1405.0, + 1835.0, + 1405.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1867.0, + 595.0, + 1867.0, + 595.0, + 1899.0, + 293.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1451.0, + 1403.0, + 1451.0, + 1403.0, + 1481.0, + 296.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1480.0, + 1405.0, + 1480.0, + 1405.0, + 1515.0, + 293.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1510.0, + 1405.0, + 1510.0, + 1405.0, + 1546.0, + 292.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1543.0, + 1402.0, + 1543.0, + 1402.0, + 1573.0, + 296.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1568.0, + 1403.0, + 1568.0, + 1403.0, + 1609.0, + 293.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1599.0, + 1407.0, + 1599.0, + 1407.0, + 1638.0, + 292.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1635.0, + 430.0, + 1635.0, + 430.0, + 1663.0, + 296.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1219.0, + 1405.0, + 1219.0, + 1405.0, + 1253.0, + 296.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1250.0, + 1402.0, + 1250.0, + 1402.0, + 1284.0, + 293.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1283.0, + 1403.0, + 1283.0, + 1403.0, + 1313.0, + 296.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1307.0, + 1404.0, + 1307.0, + 1404.0, + 1349.0, + 292.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1340.0, + 1403.0, + 1340.0, + 1403.0, + 1376.0, + 293.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1371.0, + 1404.0, + 1371.0, + 1404.0, + 1407.0, + 293.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1399.0, + 398.0, + 1399.0, + 398.0, + 1437.0, + 291.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1048.0, + 1407.0, + 1048.0, + 1407.0, + 1089.0, + 293.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1081.0, + 1404.0, + 1081.0, + 1404.0, + 1115.0, + 293.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1111.0, + 1403.0, + 1111.0, + 1403.0, + 1145.0, + 296.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1137.0, + 1406.0, + 1137.0, + 1406.0, + 1183.0, + 292.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1171.0, + 898.0, + 1171.0, + 898.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1908.0, + 523.0, + 1908.0, + 523.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1947.0, + 546.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 810.0, + 1941.0, + 810.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1941.0, + 1054.0, + 1941.0, + 1054.0, + 1978.0, + 830.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 1074.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 2006.0, + 1359.0, + 2006.0, + 1359.0, + 2034.0, + 298.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 322.0, + 560.0, + 322.0, + 560.0, + 355.0, + 315.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 352.0, + 682.0, + 352.0, + 682.0, + 385.0, + 311.0, + 385.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 695, + 1405, + 695, + 1405, + 971, + 297, + 971 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1257, + 1404, + 1257, + 1404, + 1472, + 298, + 1472 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1568, + 1403, + 1568, + 1403, + 1723, + 299, + 1723 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1089, + 1403, + 1089, + 1403, + 1241, + 299, + 1241 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 296, + 618, + 1400, + 618, + 1400, + 680, + 296, + 680 + ], + "score": 0.949 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 301, + 1017, + 587, + 1017, + 587, + 1053, + 301, + 1053 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 301, + 1761, + 816, + 1761, + 816, + 1793, + 301, + 1793 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 301, + 1510, + 705, + 1510, + 705, + 1542, + 301, + 1542 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.706 + }, + { + "category_id": 1, + "poly": [ + 298, + 270, + 779, + 270, + 779, + 557, + 298, + 557 + ], + "score": 0.64 + }, + { + "category_id": 6, + "poly": [ + 297, + 228, + 947, + 228, + 947, + 263, + 297, + 263 + ], + "score": 0.492 + }, + { + "category_id": 0, + "poly": [ + 297, + 228, + 947, + 228, + 947, + 263, + 297, + 263 + ], + "score": 0.265 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.147 + }, + { + "category_id": 13, + "poly": [ + 325, + 302, + 411, + 302, + 411, + 333, + 325, + 333 + ], + "score": 0.88, + "latex": "\\bar { x ^ { ( 0 ) } } \\gets" + }, + { + "category_id": 13, + "poly": [ + 1084, + 621, + 1105, + 621, + 1105, + 652, + 1084, + 652 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 701, + 621, + 721, + 621, + 721, + 651, + 701, + 651 + ], + "score": 0.82, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 364, + 363, + 723, + 363, + 723, + 400, + 364, + 400 + ], + "score": 0.74, + "latex": "\\Delta x \\gets \\acute { \\phi } ( f , \\{ x ^ { ( 0 ) } , \\ldots , x ^ { ( i - 1 ) } \\} )" + }, + { + "category_id": 13, + "poly": [ + 758, + 307, + 777, + 307, + 777, + 338, + 758, + 338 + ], + "score": 0.72, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 628, + 274, + 648, + 274, + 648, + 305, + 628, + 305 + ], + "score": 0.65, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 491, + 427, + 563, + 427, + 563, + 458, + 491, + 458 + ], + "score": 0.55, + "latex": "x ^ { ( i - 1 ) }" + }, + { + "category_id": 13, + "poly": [ + 366, + 337, + 526, + 337, + 526, + 367, + 366, + 367 + ], + "score": 0.49, + "latex": "i = 1 , 2 , \\dots { \\bf d }" + }, + { + "category_id": 14, + "poly": [ + 366, + 487, + 600, + 487, + 600, + 521, + 366, + 521 + ], + "score": 0.32, + "latex": "x ^ { ( i ) } \\gets x ^ { ( i - 1 ) } + \\Delta x" + }, + { + "category_id": 13, + "poly": [ + 366, + 487, + 600, + 487, + 600, + 520, + 366, + 520 + ], + "score": 0.28, + "latex": "x ^ { ( i ) } \\gets x ^ { ( i - 1 ) } + \\Delta x" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1014.0, + 594.0, + 1014.0, + 594.0, + 1061.0, + 291.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1758.0, + 819.0, + 1758.0, + 819.0, + 1798.0, + 294.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1510.0, + 709.0, + 1510.0, + 709.0, + 1546.0, + 295.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 223.0, + 950.0, + 223.0, + 950.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 223.0, + 950.0, + 223.0, + 950.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 696.0, + 1406.0, + 696.0, + 1406.0, + 732.0, + 294.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 727.0, + 1403.0, + 727.0, + 1403.0, + 760.0, + 295.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 757.0, + 1407.0, + 757.0, + 1407.0, + 794.0, + 294.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 786.0, + 1405.0, + 786.0, + 1405.0, + 823.0, + 294.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 816.0, + 1405.0, + 816.0, + 1405.0, + 854.0, + 294.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 847.0, + 1403.0, + 847.0, + 1403.0, + 884.0, + 294.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 877.0, + 1403.0, + 877.0, + 1403.0, + 914.0, + 294.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 909.0, + 1406.0, + 909.0, + 1406.0, + 944.0, + 292.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 940.0, + 521.0, + 940.0, + 521.0, + 973.0, + 297.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1256.0, + 1404.0, + 1256.0, + 1404.0, + 1291.0, + 293.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1288.0, + 1404.0, + 1288.0, + 1404.0, + 1323.0, + 293.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1318.0, + 1402.0, + 1318.0, + 1402.0, + 1352.0, + 294.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1347.0, + 1406.0, + 1347.0, + 1406.0, + 1384.0, + 295.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1378.0, + 1405.0, + 1378.0, + 1405.0, + 1414.0, + 293.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1411.0, + 1404.0, + 1411.0, + 1404.0, + 1446.0, + 294.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1440.0, + 1284.0, + 1440.0, + 1284.0, + 1476.0, + 293.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1570.0, + 1404.0, + 1570.0, + 1404.0, + 1603.0, + 295.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1600.0, + 1404.0, + 1600.0, + 1404.0, + 1634.0, + 297.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1628.0, + 1405.0, + 1628.0, + 1405.0, + 1667.0, + 294.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1657.0, + 1407.0, + 1657.0, + 1407.0, + 1699.0, + 293.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1688.0, + 1161.0, + 1688.0, + 1161.0, + 1729.0, + 293.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1855.0, + 293.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1945.0, + 293.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1977.0, + 292.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1086.0, + 1405.0, + 1086.0, + 1405.0, + 1124.0, + 294.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1119.0, + 1405.0, + 1119.0, + 1405.0, + 1156.0, + 294.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1149.0, + 1407.0, + 1149.0, + 1407.0, + 1186.0, + 294.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1175.0, + 1405.0, + 1175.0, + 1405.0, + 1220.0, + 292.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1207.0, + 1078.0, + 1207.0, + 1078.0, + 1250.0, + 292.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 615.0, + 700.0, + 615.0, + 700.0, + 657.0, + 294.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 615.0, + 1083.0, + 615.0, + 1083.0, + 657.0, + 722.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 615.0, + 1404.0, + 615.0, + 1404.0, + 657.0, + 1106.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 649.0, + 481.0, + 649.0, + 481.0, + 683.0, + 292.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 272.0, + 627.0, + 272.0, + 627.0, + 305.0, + 297.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 298.0, + 324.0, + 298.0, + 324.0, + 341.0, + 320.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 298.0, + 757.0, + 298.0, + 757.0, + 341.0, + 412.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 298.0, + 782.0, + 298.0, + 782.0, + 341.0, + 778.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 335.0, + 365.0, + 335.0, + 365.0, + 368.0, + 321.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 335.0, + 543.0, + 335.0, + 543.0, + 368.0, + 527.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 356.0, + 363.0, + 356.0, + 363.0, + 403.0, + 360.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 356.0, + 730.0, + 356.0, + 730.0, + 403.0, + 724.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 395.0, + 736.0, + 395.0, + 736.0, + 431.0, + 363.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 460.0, + 445.0, + 460.0, + 445.0, + 491.0, + 360.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 479.0, + 365.0, + 479.0, + 365.0, + 528.0, + 357.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 479.0, + 606.0, + 479.0, + 606.0, + 528.0, + 601.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 518.0, + 419.0, + 518.0, + 419.0, + 555.0, + 322.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.25, + 425.0, + 569.25, + 425.0, + 569.25, + 461.0, + 402.25, + 461.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1508, + 1404, + 1508, + 1404, + 1896, + 298, + 1896 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 800, + 1404, + 800, + 1404, + 1106, + 297, + 1106 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 429, + 1404, + 429, + 1404, + 705, + 298, + 705 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1122, + 1403, + 1122, + 1403, + 1336, + 298, + 1336 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1402, + 229, + 1402, + 413, + 299, + 413 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 301, + 743, + 690, + 743, + 690, + 774, + 301, + 774 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 302, + 1381, + 700, + 1381, + 700, + 1417, + 302, + 1417 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 299, + 1451, + 472, + 1451, + 472, + 1483, + 299, + 1483 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 856, + 76, + 856, + 104, + 298, + 104 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2112, + 842, + 2112 + ], + "score": 0.603 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2112, + 842, + 2112 + ], + "score": 0.382 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 855, + 76, + 855, + 104, + 300, + 104 + ], + "score": 0.106 + }, + { + "category_id": 13, + "poly": [ + 1157, + 1570, + 1317, + 1570, + 1317, + 1605, + 1157, + 1605 + ], + "score": 0.93, + "latex": "x ^ { ( 1 ) } , \\ldots , x ^ { ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 857, + 1829, + 1009, + 1829, + 1009, + 1869, + 857, + 1869 + ], + "score": 0.93, + "latex": "\\textstyle \\sum _ { i = 1 } ^ { T } f ( x ^ { ( i ) } )" + }, + { + "category_id": 13, + "poly": [ + 563, + 1912, + 681, + 1912, + 681, + 1944, + 563, + 1944 + ], + "score": 0.92, + "latex": "f _ { 1 } , \\ldots , f _ { n }" + }, + { + "category_id": 13, + "poly": [ + 388, + 1670, + 548, + 1670, + 548, + 1705, + 388, + 1705 + ], + "score": 0.92, + "latex": "x ^ { ( 1 ) } , \\ldots , x ^ { ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1541, + 416, + 1541, + 416, + 1573, + 298, + 1573 + ], + "score": 0.91, + "latex": "f _ { 1 } , \\ldots , f _ { n }" + }, + { + "category_id": 13, + "poly": [ + 494, + 1637, + 542, + 1637, + 542, + 1668, + 494, + 1668 + ], + "score": 0.9, + "latex": "x ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 604, + 1570, + 652, + 1570, + 652, + 1602, + 604, + 1602 + ], + "score": 0.9, + "latex": "x ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1604, + 350, + 1604, + 350, + 1634, + 297, + 1634 + ], + "score": 0.89, + "latex": "x ^ { ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 1252, + 1704, + 1289, + 1704, + 1289, + 1732, + 1252, + 1732 + ], + "score": 0.86, + "latex": "{ \\mathcal { P } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1734, + 671, + 1734, + 671, + 1774, + 297, + 1774 + ], + "score": 0.85, + "latex": "\\mathbb { E } _ { f \\sim \\mathcal { F } , x ^ { ( 0 ) } \\sim \\mathcal { D } } \\left[ \\mathcal { L } ( f , \\mathcal { P } ^ { * } ( f , x ^ { ( 0 ) } ) ) \\right]" + }, + { + "category_id": 13, + "poly": [ + 1165, + 1641, + 1186, + 1641, + 1186, + 1673, + 1165, + 1673 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 348, + 1575, + 369, + 1575, + 369, + 1605, + 348, + 1605 + ], + "score": 0.84, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1207, + 1608, + 1233, + 1608, + 1233, + 1634, + 1207, + 1634 + ], + "score": 0.81, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 760, + 1542, + 785, + 1542, + 785, + 1568, + 760, + 1568 + ], + "score": 0.81, + "latex": "\\mathcal { F }" + }, + { + "category_id": 13, + "poly": [ + 1104, + 1542, + 1129, + 1542, + 1129, + 1569, + 1104, + 1569 + ], + "score": 0.79, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 737, + 1641, + 759, + 1641, + 759, + 1669, + 737, + 1669 + ], + "score": 0.76, + "latex": "\\mathcal { L }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 740.0, + 696.0, + 740.0, + 696.0, + 777.0, + 293.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1379.0, + 706.0, + 1379.0, + 706.0, + 1422.0, + 292.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1448.0, + 476.0, + 1448.0, + 476.0, + 1488.0, + 294.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1506.0, + 1404.0, + 1506.0, + 1404.0, + 1544.0, + 292.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1541.0, + 759.0, + 1541.0, + 759.0, + 1576.0, + 417.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1541.0, + 1103.0, + 1541.0, + 1103.0, + 1576.0, + 786.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1541.0, + 1402.0, + 1541.0, + 1402.0, + 1576.0, + 1130.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1567.0, + 347.0, + 1567.0, + 347.0, + 1609.0, + 291.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1567.0, + 603.0, + 1567.0, + 603.0, + 1609.0, + 370.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1567.0, + 1156.0, + 1567.0, + 1156.0, + 1609.0, + 653.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1567.0, + 1408.0, + 1567.0, + 1408.0, + 1609.0, + 1318.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1599.0, + 296.0, + 1599.0, + 296.0, + 1646.0, + 292.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1599.0, + 1206.0, + 1599.0, + 1206.0, + 1646.0, + 351.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1599.0, + 1408.0, + 1599.0, + 1408.0, + 1646.0, + 1234.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1635.0, + 493.0, + 1635.0, + 493.0, + 1678.0, + 293.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1635.0, + 736.0, + 1635.0, + 736.0, + 1678.0, + 543.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1635.0, + 1164.0, + 1635.0, + 1164.0, + 1678.0, + 760.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1635.0, + 1408.0, + 1635.0, + 1408.0, + 1678.0, + 1187.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1662.0, + 387.0, + 1662.0, + 387.0, + 1711.0, + 291.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1662.0, + 1408.0, + 1662.0, + 1408.0, + 1711.0, + 549.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1703.0, + 1251.0, + 1703.0, + 1251.0, + 1739.0, + 294.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1703.0, + 1406.0, + 1703.0, + 1406.0, + 1739.0, + 1290.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1730.0, + 296.0, + 1730.0, + 296.0, + 1776.0, + 289.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1730.0, + 1409.0, + 1730.0, + 1409.0, + 1776.0, + 672.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1767.0, + 1404.0, + 1767.0, + 1404.0, + 1802.0, + 294.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 281.0, + 1797.0, + 856.0, + 1797.0, + 856.0, + 1900.0, + 281.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1797.0, + 1418.0, + 1797.0, + 1418.0, + 1900.0, + 1010.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 801.0, + 1406.0, + 801.0, + 1406.0, + 836.0, + 294.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 831.0, + 1406.0, + 831.0, + 1406.0, + 867.0, + 294.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 858.0, + 1405.0, + 858.0, + 1405.0, + 901.0, + 292.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 891.0, + 1404.0, + 891.0, + 1404.0, + 929.0, + 294.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 920.0, + 1406.0, + 920.0, + 1406.0, + 961.0, + 291.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 954.0, + 1406.0, + 954.0, + 1406.0, + 990.0, + 294.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 981.0, + 1406.0, + 981.0, + 1406.0, + 1020.0, + 292.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1013.0, + 1404.0, + 1013.0, + 1404.0, + 1049.0, + 292.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 1406.0, + 1044.0, + 1406.0, + 1080.0, + 294.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1076.0, + 814.0, + 1076.0, + 814.0, + 1110.0, + 293.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 431.0, + 1404.0, + 431.0, + 1404.0, + 464.0, + 296.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 457.0, + 1405.0, + 457.0, + 1405.0, + 499.0, + 292.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 490.0, + 1406.0, + 490.0, + 1406.0, + 527.0, + 296.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 521.0, + 1405.0, + 521.0, + 1405.0, + 558.0, + 294.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 551.0, + 1405.0, + 551.0, + 1405.0, + 587.0, + 293.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 580.0, + 1406.0, + 580.0, + 1406.0, + 618.0, + 295.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 611.0, + 1404.0, + 611.0, + 1404.0, + 648.0, + 294.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 640.0, + 1406.0, + 640.0, + 1406.0, + 679.0, + 292.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 674.0, + 632.0, + 674.0, + 632.0, + 707.0, + 296.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1122.0, + 1403.0, + 1122.0, + 1403.0, + 1157.0, + 296.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1148.0, + 1405.0, + 1148.0, + 1405.0, + 1191.0, + 292.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1184.0, + 1405.0, + 1184.0, + 1405.0, + 1218.0, + 293.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1213.0, + 1404.0, + 1213.0, + 1404.0, + 1249.0, + 293.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1240.0, + 1407.0, + 1240.0, + 1407.0, + 1281.0, + 292.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1273.0, + 1407.0, + 1273.0, + 1407.0, + 1310.0, + 292.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1304.0, + 935.0, + 1304.0, + 935.0, + 1343.0, + 293.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 230.0, + 1406.0, + 230.0, + 1406.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 263.0, + 1404.0, + 263.0, + 1404.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 1406.0, + 291.0, + 1406.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 322.0, + 1406.0, + 322.0, + 1406.0, + 358.0, + 293.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1406.0, + 351.0, + 1406.0, + 386.0, + 293.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 382.0, + 956.0, + 382.0, + 956.0, + 416.0, + 294.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1907.0, + 562.0, + 1907.0, + 562.0, + 1950.0, + 292.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1907.0, + 1408.0, + 1907.0, + 1408.0, + 1950.0, + 682.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1403.0, + 1942.0, + 1403.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2003.0, + 1408.0, + 2003.0, + 1408.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 229, + 1405, + 229, + 1405, + 595, + 297, + 595 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 689, + 1404, + 689, + 1404, + 973, + 297, + 973 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1720, + 1405, + 1720, + 1405, + 1955, + 297, + 1955 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 301, + 1506, + 1399, + 1506, + 1399, + 1631, + 301, + 1631 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 985, + 1404, + 985, + 1404, + 1110, + 298, + 1110 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 528, + 1306, + 1168, + 1306, + 1168, + 1478, + 528, + 1478 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 296, + 1230, + 1399, + 1230, + 1399, + 1292, + 296, + 1292 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 611, + 1125, + 1085, + 1125, + 1085, + 1219, + 611, + 1219 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 298, + 1977, + 1406, + 1977, + 1406, + 2034, + 298, + 2034 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 300, + 633, + 600, + 633, + 600, + 665, + 300, + 665 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 299, + 1666, + 544, + 1666, + 544, + 1699, + 299, + 1699 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2111, + 841, + 2111 + ], + "score": 0.775 + }, + { + "category_id": 14, + "poly": [ + 612, + 1123, + 1084, + 1123, + 1084, + 1220, + 612, + 1220 + ], + "score": 0.94, + "latex": "\\pi ^ { * } = \\arg \\operatorname* { m i n } _ { \\pi } \\mathbb { E } _ { s _ { 0 } , a _ { 0 } , s _ { 1 } , \\dots , s _ { T } } \\left[ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) \\right] ," + }, + { + "category_id": 13, + "poly": [ + 1013, + 1537, + 1113, + 1537, + 1113, + 1570, + 1013, + 1570 + ], + "score": 0.93, + "latex": "\\textstyle { \\mathcal { N } } ( { \\boldsymbol { \\mu } } , { \\boldsymbol { \\Sigma } } )" + }, + { + "category_id": 14, + "poly": [ + 526, + 1304, + 1174, + 1304, + 1174, + 1482, + 526, + 1482 + ], + "score": 0.93, + "latex": "\\begin{array} { c } { { \\displaystyle q \\big ( s _ { 0 } , a _ { 0 } , s _ { 1 } , \\ldots , s _ { T } \\big ) = \\int _ { o _ { 0 } , \\ldots , o _ { T } } p _ { i } \\left( s _ { 0 } \\right) p _ { o } \\left( o _ { 0 } \\big | s _ { 0 } \\right) } } \\\\ { { { \\displaystyle \\prod _ { t = 0 } ^ { T - 1 } \\pi \\left( \\left. a _ { t } \\right| o _ { t } , t \\right) p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } \\right) p _ { o } \\left( \\left. o _ { t + 1 } \\right| s _ { t + 1 } \\right) . } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 987, + 783, + 1102, + 783, + 1102, + 818, + 987, + 818 + ], + "score": 0.93, + "latex": "\\mathcal { O } \\subseteq \\mathbb { R } ^ { D ^ { \\prime } }" + }, + { + "category_id": 13, + "poly": [ + 397, + 985, + 527, + 985, + 527, + 1020, + 397, + 1020 + ], + "score": 0.93, + "latex": "\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { o } _ { t } , t \\right)" + }, + { + "category_id": 13, + "poly": [ + 375, + 1792, + 603, + 1792, + 603, + 1831, + 375, + 1831 + ], + "score": 0.92, + "latex": "{ \\hat { f } } ( x ^ { ( 1 ) } ) , \\ldots , { \\hat { f } } ( x ^ { ( t ) } )" + }, + { + "category_id": 13, + "poly": [ + 656, + 784, + 762, + 784, + 762, + 817, + 656, + 817 + ], + "score": 0.92, + "latex": "\\mathcal { S } \\subseteq \\mathbb { R } ^ { D }" + }, + { + "category_id": 13, + "poly": [ + 516, + 1537, + 918, + 1537, + 918, + 1571, + 516, + 1571 + ], + "score": 0.92, + "latex": "\\pi \\left( \\left. a _ { t } \\right| o _ { t } , t \\right) = \\mathcal { N } \\left( \\mu ^ { \\pi } ( o _ { t } ) , \\Sigma ^ { \\pi } ( o _ { t } ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 890, + 1831, + 944, + 1831, + 944, + 1865, + 890, + 1865 + ], + "score": 0.91, + "latex": "\\Psi ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1177, + 1568, + 1244, + 1568, + 1244, + 1602, + 1177, + 1602 + ], + "score": 0.91, + "latex": "\\Sigma ^ { \\pi } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 953, + 1569, + 1017, + 1569, + 1017, + 1602, + 953, + 1602 + ], + "score": 0.91, + "latex": "\\mu ^ { \\pi } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 876, + 1755, + 1151, + 1755, + 1151, + 1795, + 876, + 1795 + ], + "score": 0.91, + "latex": "\\nabla \\hat { f } ( x ^ { ( 1 ) } ) , \\ldots , \\nabla \\hat { f } ( x ^ { ( t ) } )" + }, + { + "category_id": 13, + "poly": [ + 508, + 1757, + 660, + 1757, + 660, + 1793, + 508, + 1793 + ], + "score": 0.91, + "latex": "x ^ { ( 1 ) } , \\ldots , x ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 1130, + 1724, + 1183, + 1724, + 1183, + 1757, + 1130, + 1757 + ], + "score": 0.9, + "latex": "\\Phi ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 934, + 1721, + 979, + 1721, + 979, + 1752, + 934, + 1752 + ], + "score": 0.9, + "latex": "x ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 298, + 878, + 414, + 878, + 414, + 911, + 298, + 911 + ], + "score": 0.9, + "latex": "p _ { o } \\left( o _ { t } \\bar { | } _ { s _ { t } } \\right)" + }, + { + "category_id": 13, + "poly": [ + 298, + 818, + 394, + 818, + 394, + 848, + 298, + 848 + ], + "score": 0.89, + "latex": "\\mathcal { A } \\subseteq \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1232, + 818, + 1400, + 818, + 1400, + 851, + 1232, + 851 + ], + "score": 0.89, + "latex": "p \\left( { { s } _ { t + 1 } } \\left| { { s } _ { t } } , { { a } _ { t } } \\right. \\right)" + }, + { + "category_id": 13, + "poly": [ + 884, + 854, + 938, + 854, + 938, + 881, + 884, + 881 + ], + "score": 0.89, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 786, + 567, + 786, + 567, + 820, + 298, + 820 + ], + "score": 0.88, + "latex": "( S , \\mathcal { O } , \\mathcal { A } , p _ { i } , p , p _ { o } , c , T )" + }, + { + "category_id": 13, + "poly": [ + 628, + 819, + 705, + 819, + 705, + 851, + 628, + 851 + ], + "score": 0.87, + "latex": "p _ { i } \\left( s _ { 0 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 298, + 911, + 438, + 911, + 438, + 938, + 298, + 938 + ], + "score": 0.87, + "latex": "c : { \\mathcal { S } } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 936, + 1049, + 968, + 1049, + 968, + 1075, + 936, + 1075 + ], + "score": 0.86, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 569, + 1830, + 615, + 1830, + 615, + 1860, + 569, + 1860 + ], + "score": 0.86, + "latex": "x ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 597, + 944, + 627, + 944, + 627, + 972, + 597, + 972 + ], + "score": 0.85, + "latex": "p _ { o }" + }, + { + "category_id": 13, + "poly": [ + 566, + 1731, + 592, + 1731, + 592, + 1756, + 566, + 1756 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1207, + 855, + 1233, + 855, + 1233, + 879, + 1207, + 879 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1051, + 992, + 1078, + 992, + 1078, + 1017, + 1051, + 1017 + ], + "score": 0.85, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1023, + 323, + 1023, + 323, + 1048, + 298, + 1048 + ], + "score": 0.85, + "latex": "o _ { t }" + }, + { + "category_id": 13, + "poly": [ + 739, + 1801, + 767, + 1801, + 767, + 1828, + 739, + 1828 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 888, + 1798, + 930, + 1798, + 930, + 1826, + 888, + 1826 + ], + "score": 0.84, + "latex": "\\Delta x" + }, + { + "category_id": 13, + "poly": [ + 1061, + 883, + 1088, + 883, + 1088, + 909, + 1061, + 909 + ], + "score": 0.84, + "latex": "o _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1367, + 885, + 1393, + 885, + 1393, + 909, + 1367, + 909 + ], + "score": 0.83, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 433, + 1836, + 461, + 1836, + 461, + 1862, + 433, + 1862 + ], + "score": 0.82, + "latex": "o _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1366, + 856, + 1393, + 856, + 1393, + 879, + 1366, + 879 + ], + "score": 0.8, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1015, + 911, + 1039, + 911, + 1039, + 937, + 1015, + 937 + ], + "score": 0.8, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 529, + 944, + 548, + 944, + 548, + 972, + 529, + 972 + ], + "score": 0.8, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 538, + 1575, + 558, + 1575, + 558, + 1600, + 538, + 1600 + ], + "score": 0.8, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 743, + 1572, + 766, + 1572, + 766, + 1596, + 743, + 1596 + ], + "score": 0.77, + "latex": "\\Sigma" + }, + { + "category_id": 13, + "poly": [ + 620, + 1514, + 642, + 1514, + 642, + 1535, + 620, + 1535 + ], + "score": 0.77, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 871, + 1020, + 884, + 1020, + 884, + 1045, + 871, + 1045 + ], + "score": 0.74, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 481, + 1023, + 495, + 1023, + 495, + 1045, + 481, + 1045 + ], + "score": 0.72, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1192, + 822, + 1220, + 822, + 1220, + 849, + 1192, + 849 + ], + "score": 0.56, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 927, + 751, + 1053, + 751, + 1053, + 782, + 927, + 782 + ], + "score": 0.29, + "latex": "( \\mathrm { P O M D P } ) ^ { 1 }" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2010.0, + 330.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2004.0, + 379.0, + 2004.0, + 379.0, + 2041.0, + 292.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 633.0, + 604.0, + 633.0, + 604.0, + 669.0, + 295.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1662.0, + 549.0, + 1662.0, + 549.0, + 1703.0, + 292.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 264.0, + 294.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 288.0, + 1405.0, + 288.0, + 1405.0, + 330.0, + 292.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 357.0, + 294.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 384.0, + 1405.0, + 384.0, + 1405.0, + 417.0, + 294.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 413.0, + 1403.0, + 413.0, + 1403.0, + 448.0, + 294.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 442.0, + 1405.0, + 442.0, + 1405.0, + 479.0, + 294.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 476.0, + 1403.0, + 476.0, + 1403.0, + 507.0, + 296.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 504.0, + 1405.0, + 504.0, + 1405.0, + 538.0, + 294.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 533.0, + 1405.0, + 533.0, + 1405.0, + 570.0, + 292.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 566.0, + 490.0, + 566.0, + 490.0, + 599.0, + 294.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 689.0, + 1404.0, + 689.0, + 1404.0, + 725.0, + 294.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 723.0, + 1404.0, + 723.0, + 1404.0, + 757.0, + 295.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 753.0, + 926.0, + 753.0, + 926.0, + 787.0, + 295.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 753.0, + 1404.0, + 753.0, + 1404.0, + 787.0, + 1054.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 783.0, + 297.0, + 783.0, + 297.0, + 824.0, + 292.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 783.0, + 655.0, + 783.0, + 655.0, + 824.0, + 568.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 783.0, + 986.0, + 783.0, + 986.0, + 824.0, + 763.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 783.0, + 1406.0, + 783.0, + 1406.0, + 824.0, + 1103.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 813.0, + 297.0, + 813.0, + 297.0, + 855.0, + 292.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 813.0, + 627.0, + 813.0, + 627.0, + 855.0, + 395.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 813.0, + 1191.0, + 813.0, + 1191.0, + 855.0, + 706.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 813.0, + 1231.0, + 813.0, + 1231.0, + 855.0, + 1221.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 813.0, + 1405.0, + 813.0, + 1405.0, + 855.0, + 1401.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 847.0, + 883.0, + 847.0, + 883.0, + 887.0, + 291.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 847.0, + 1206.0, + 847.0, + 1206.0, + 887.0, + 939.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 847.0, + 1365.0, + 847.0, + 1365.0, + 887.0, + 1234.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 847.0, + 1409.0, + 847.0, + 1409.0, + 887.0, + 1394.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 877.0, + 297.0, + 877.0, + 297.0, + 917.0, + 290.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 877.0, + 1060.0, + 877.0, + 1060.0, + 917.0, + 415.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 877.0, + 1366.0, + 877.0, + 1366.0, + 917.0, + 1089.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 877.0, + 1409.0, + 877.0, + 1409.0, + 917.0, + 1394.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 910.0, + 297.0, + 910.0, + 297.0, + 944.0, + 294.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 910.0, + 1014.0, + 910.0, + 1014.0, + 944.0, + 439.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 910.0, + 1405.0, + 910.0, + 1405.0, + 944.0, + 1040.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 940.0, + 528.0, + 940.0, + 528.0, + 974.0, + 294.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 940.0, + 596.0, + 940.0, + 596.0, + 974.0, + 549.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 940.0, + 1223.0, + 940.0, + 1223.0, + 974.0, + 628.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1717.0, + 565.0, + 1717.0, + 565.0, + 1762.0, + 292.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1717.0, + 933.0, + 1717.0, + 933.0, + 1762.0, + 593.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1717.0, + 1129.0, + 1717.0, + 1129.0, + 1762.0, + 980.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1717.0, + 1407.0, + 1717.0, + 1407.0, + 1762.0, + 1184.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1751.0, + 507.0, + 1751.0, + 507.0, + 1800.0, + 291.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1751.0, + 875.0, + 1751.0, + 875.0, + 1800.0, + 661.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1751.0, + 1408.0, + 1751.0, + 1408.0, + 1800.0, + 1152.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 374.0, + 1790.0, + 374.0, + 1835.0, + 294.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1790.0, + 738.0, + 1790.0, + 738.0, + 1835.0, + 604.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1790.0, + 887.0, + 1790.0, + 887.0, + 1835.0, + 768.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1790.0, + 1406.0, + 1790.0, + 1406.0, + 1835.0, + 931.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1824.0, + 432.0, + 1824.0, + 432.0, + 1868.0, + 292.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1824.0, + 568.0, + 1824.0, + 568.0, + 1868.0, + 462.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1824.0, + 889.0, + 1824.0, + 889.0, + 1868.0, + 616.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1824.0, + 1407.0, + 1824.0, + 1407.0, + 1868.0, + 945.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1860.0, + 1403.0, + 1860.0, + 1403.0, + 1896.0, + 295.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1891.0, + 1408.0, + 1891.0, + 1408.0, + 1928.0, + 295.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1920.0, + 1127.0, + 1920.0, + 1127.0, + 1959.0, + 292.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1508.0, + 619.0, + 1508.0, + 619.0, + 1540.0, + 297.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1508.0, + 1403.0, + 1508.0, + 1403.0, + 1540.0, + 643.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1537.0, + 515.0, + 1537.0, + 515.0, + 1571.0, + 293.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1537.0, + 1012.0, + 1537.0, + 1012.0, + 1571.0, + 919.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1571.0, + 1114.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1567.0, + 537.0, + 1567.0, + 537.0, + 1604.0, + 295.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1567.0, + 742.0, + 1567.0, + 742.0, + 1604.0, + 559.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1567.0, + 952.0, + 1567.0, + 952.0, + 1604.0, + 767.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1567.0, + 1176.0, + 1567.0, + 1176.0, + 1604.0, + 1018.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 1567.0, + 1405.0, + 1567.0, + 1405.0, + 1604.0, + 1245.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1599.0, + 978.0, + 1599.0, + 978.0, + 1633.0, + 296.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 985.0, + 396.0, + 985.0, + 396.0, + 1022.0, + 294.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 985.0, + 1050.0, + 985.0, + 1050.0, + 1022.0, + 528.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 985.0, + 1405.0, + 985.0, + 1405.0, + 1022.0, + 1079.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1017.0, + 297.0, + 1017.0, + 297.0, + 1051.0, + 292.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1017.0, + 480.0, + 1017.0, + 480.0, + 1051.0, + 324.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1017.0, + 870.0, + 1017.0, + 870.0, + 1051.0, + 496.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1017.0, + 1407.0, + 1017.0, + 1407.0, + 1051.0, + 885.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1049.0, + 935.0, + 1049.0, + 935.0, + 1081.0, + 296.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1049.0, + 1405.0, + 1049.0, + 1405.0, + 1081.0, + 969.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1079.0, + 596.0, + 1079.0, + 596.0, + 1113.0, + 296.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1228.0, + 1403.0, + 1228.0, + 1403.0, + 1264.0, + 296.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1261.0, + 988.0, + 1261.0, + 988.0, + 1296.0, + 296.0, + 1296.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 743, + 1406, + 743, + 1406, + 959, + 297, + 959 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1405, + 228, + 1405, + 432, + 297, + 432 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 552, + 1405, + 552, + 1405, + 646, + 298, + 646 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 444, + 1399, + 444, + 1399, + 539, + 299, + 539 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 580, + 1487, + 1119, + 1487, + 1119, + 1699, + 580, + 1699 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 297, + 1407, + 1402, + 1407, + 1402, + 1469, + 297, + 1469 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 568, + 1290, + 1131, + 1290, + 1131, + 1378, + 568, + 1378 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 515, + 1021, + 1183, + 1021, + 1183, + 1108, + 515, + 1108 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 289, + 1709, + 1404, + 1709, + 1404, + 1780, + 289, + 1780 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 296, + 1123, + 1401, + 1123, + 1401, + 1194, + 296, + 1194 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 297, + 1208, + 1401, + 1208, + 1401, + 1274, + 297, + 1274 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 299, + 685, + 669, + 685, + 669, + 718, + 299, + 718 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 296, + 969, + 1151, + 969, + 1151, + 1004, + 296, + 1004 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 297, + 1796, + 1404, + 1796, + 1404, + 1893, + 297, + 1893 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 297, + 1918, + 1404, + 1918, + 1404, + 2036, + 297, + 2036 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.761 + }, + { + "category_id": 2, + "poly": [ + 304, + 1918, + 1399, + 1918, + 1399, + 2003, + 304, + 2003 + ], + "score": 0.173 + }, + { + "category_id": 14, + "poly": [ + 515, + 1019, + 1184, + 1019, + 1184, + 1108, + 515, + 1108 + ], + "score": 0.93, + "latex": "\\operatorname* { m i n } _ { \\theta , \\eta } \\mathbb { E } _ { \\psi } \\left[ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) \\right] \\mathrm { ~ s . t . ~ } \\psi \\left( \\left. a _ { t } \\right| s _ { t } , t ; \\eta \\right) = \\pi \\left( \\left. a _ { t } \\right| s _ { t } ; \\theta \\right) \\forall a _ { t } , s _ { t } , t" + }, + { + "category_id": 13, + "poly": [ + 1187, + 352, + 1400, + 352, + 1400, + 398, + 1187, + 398 + ], + "score": 0.93, + "latex": "\\boldsymbol { s } _ { t } = \\left( \\boldsymbol { x } ^ { ( t ) } , \\Phi \\left( \\cdot \\right) \\right) ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 550, + 444, + 680, + 444, + 680, + 479, + 550, + 479 + ], + "score": 0.93, + "latex": "\\pi \\left( \\boldsymbol { a } _ { t } \\left| \\boldsymbol { o } _ { t } , t \\right. \\right)" + }, + { + "category_id": 13, + "poly": [ + 670, + 394, + 752, + 394, + 752, + 432, + 670, + 432 + ], + "score": 0.93, + "latex": "f ( \\boldsymbol { x } ^ { ( t ) } )" + }, + { + "category_id": 14, + "poly": [ + 567, + 1289, + 1133, + 1289, + 1133, + 1378, + 567, + 1378 + ], + "score": 0.93, + "latex": "\\operatorname* { m i n } _ { \\theta , \\eta } \\mathbb { E } _ { \\psi } [ \\sum _ { t = 0 } ^ { T } c ( s _ { t } ) ] \\mathrm { ~ s . t . ~ } \\mathbb { E } _ { \\psi } [ a _ { t } ] = \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] \\forall t" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1793, + 1352, + 1793, + 1352, + 1834, + 1115, + 1834 + ], + "score": 0.92, + "latex": "\\boldsymbol { \\eta } : = \\left( K _ { t } , k _ { t } , G _ { t } \\right) _ { t = 1 } ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 740, + 1831, + 888, + 1831, + 888, + 1863, + 740, + 1863 + ], + "score": 0.92, + "latex": "\\boldsymbol { \\theta } : = ( \\omega , \\Sigma ^ { \\pi } )" + }, + { + "category_id": 13, + "poly": [ + 839, + 584, + 893, + 584, + 893, + 618, + 839, + 618 + ], + "score": 0.92, + "latex": "\\Psi ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 342, + 1975, + 467, + 1975, + 467, + 2005, + 342, + 2005 + ], + "score": 0.91, + "latex": "\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ; \\boldsymbol { \\theta } \\right)" + }, + { + "category_id": 13, + "poly": [ + 614, + 1797, + 1028, + 1797, + 1028, + 1831, + 614, + 1831 + ], + "score": 0.91, + "latex": "\\psi \\left( \\left. a _ { t } \\right| s _ { t } , t ; \\eta \\right) = \\mathcal { N } \\left( K _ { t } s _ { t } + k _ { t } , G _ { t } \\right)" + }, + { + "category_id": 13, + "poly": [ + 888, + 447, + 1003, + 447, + 1003, + 476, + 888, + 476 + ], + "score": 0.91, + "latex": "a _ { t } ~ = ~ \\Delta x" + }, + { + "category_id": 13, + "poly": [ + 912, + 1156, + 1392, + 1156, + 1392, + 1197, + 912, + 1197 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\pi \\left( \\left. a _ { t } \\right| s _ { t } ; \\theta \\right) : = \\int _ { o _ { t } } \\pi \\left( \\left. a _ { t } \\right| o _ { t } ; \\theta \\right) p _ { o } \\left( \\left. o _ { t } \\right| s _ { t } \\right) ^ { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1829, + 650, + 1829, + 650, + 1863, + 297, + 1863 + ], + "score": 0.91, + "latex": "\\pi \\left( a _ { t } | o _ { t } ; \\theta \\right) = \\mathcal { N } \\left( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) , \\Sigma ^ { \\pi } \\right)" + }, + { + "category_id": 14, + "poly": [ + 341, + 1483, + 1348, + 1483, + 1348, + 1785, + 341, + 1785 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } & { \\eta \\gets \\arg \\underset { \\eta } { \\operatorname* { m i n } } \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\psi } [ c ( s _ { t } ) - \\lambda _ { t } ^ { T } a _ { t } ] + \\nu _ { t } D _ { t } ( \\eta , \\theta ) } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad } \\\\ & { \\theta \\gets \\arg \\underset { \\theta } { \\operatorname* { m i n } } \\sum _ { t = 0 } ^ { T } \\lambda _ { t } ^ { T } \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] + \\nu _ { t } D _ { t } ( \\theta , \\eta ) } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\lambda _ { t } \\gets \\lambda _ { t } + \\alpha \\nu _ { t } ( \\mathbb { E } _ { \\psi } [ \\mathbb { E } _ { \\pi } [ a _ { t } | s _ { t } ] ] - \\mathbb { E } _ { \\psi } [ a _ { t } ] ) \\ \\forall t , } \\\\ & { \\mathrm { ~ } \\mathrm { ~ } : \\mathrm { r e } \\quad D _ { t } ( \\theta , \\eta ) \\quad : = \\quad \\mathbb { E } _ { \\psi } [ D _ { K L } ( \\pi ( a _ { t } | s _ { t } ; \\theta ) ) ] \\ \\psi ( a _ { t } | s _ { t } , t ; \\eta ) ) ] \\quad \\mathrm { ~ a n d ~ } \\quad D _ { t } ( \\eta , \\theta ) } \\\\ & { [ D _ { K L } ( \\psi ( a _ { t } | s _ { t } , t ; \\eta ) ] ) | \\ \\pi ( a _ { t } | s _ { t } ; \\theta ) ) | . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 999, + 1124, + 1062, + 1124, + 1062, + 1158, + 999, + 1158 + ], + "score": 0.9, + "latex": "\\mathbb { E } _ { \\rho } \\left[ \\cdot \\right]" + }, + { + "category_id": 13, + "poly": [ + 942, + 1832, + 1006, + 1832, + 1006, + 1863, + 942, + 1863 + ], + "score": 0.89, + "latex": "\\mu _ { \\omega } ^ { \\pi } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 752, + 1126, + 775, + 1126, + 775, + 1156, + 752, + 1156 + ], + "score": 0.86, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 662, + 1243, + 685, + 1243, + 685, + 1274, + 662, + 1274 + ], + "score": 0.85, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 1102, + 807, + 1125, + 807, + 1125, + 837, + 1102, + 837 + ], + "score": 0.85, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 934, + 1925, + 962, + 1925, + 962, + 1949, + 934, + 1949 + ], + "score": 0.85, + "latex": "p _ { o }" + }, + { + "category_id": 13, + "poly": [ + 852, + 898, + 873, + 898, + 873, + 929, + 852, + 929 + ], + "score": 0.85, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 708, + 478, + 728, + 478, + 728, + 508, + 708, + 508 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 911, + 236, + 938, + 236, + 938, + 262, + 911, + 262 + ], + "score": 0.84, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1298, + 2011, + 1322, + 2011, + 1322, + 2032, + 1298, + 2032 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1055, + 1953, + 1078, + 1953, + 1078, + 1974, + 1055, + 1974 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1952, + 1151, + 1952, + 1151, + 1974, + 1127, + 1974 + ], + "score": 0.83, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 844, + 1164, + 863, + 1164, + 863, + 1191, + 844, + 1191 + ], + "score": 0.81, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 735, + 1246, + 755, + 1246, + 755, + 1269, + 735, + 1269 + ], + "score": 0.81, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 373, + 1130, + 392, + 1130, + 392, + 1157, + 373, + 1157 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 445, + 1126, + 463, + 1126, + 463, + 1152, + 445, + 1152 + ], + "score": 0.78, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1147, + 267, + 1166, + 267, + 1166, + 294, + 1147, + 294 + ], + "score": 0.77, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1176, + 810, + 1195, + 810, + 1195, + 834, + 1176, + 834 + ], + "score": 0.77, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 828, + 1130, + 849, + 1130, + 849, + 1152, + 828, + 1152 + ], + "score": 0.76, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 355, + 934, + 375, + 934, + 375, + 955, + 355, + 955 + ], + "score": 0.75, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1065, + 368, + 1082, + 368, + 1082, + 389, + 1065, + 389 + ], + "score": 0.72, + "latex": "c" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 681.0, + 673.0, + 681.0, + 673.0, + 722.0, + 293.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1914.0, + 933.0, + 1914.0, + 933.0, + 1957.0, + 329.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1914.0, + 1406.0, + 1914.0, + 1406.0, + 1957.0, + 963.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1947.0, + 1054.0, + 1947.0, + 1054.0, + 1981.0, + 295.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1947.0, + 1126.0, + 1947.0, + 1126.0, + 1981.0, + 1079.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1947.0, + 1404.0, + 1947.0, + 1404.0, + 1981.0, + 1152.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 341.0, + 1973.0, + 341.0, + 2010.0, + 294.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1973.0, + 648.0, + 1973.0, + 648.0, + 2010.0, + 468.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1999.0, + 1297.0, + 1999.0, + 1297.0, + 2041.0, + 329.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 1999.0, + 1333.0, + 1999.0, + 1333.0, + 2041.0, + 1323.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1914.0, + 933.0, + 1914.0, + 933.0, + 1952.0, + 329.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1914.0, + 1407.0, + 1914.0, + 1407.0, + 1952.0, + 963.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1949.0, + 1054.0, + 1949.0, + 1054.0, + 1978.0, + 300.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1949.0, + 1126.0, + 1949.0, + 1126.0, + 1978.0, + 1079.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1949.0, + 1403.0, + 1949.0, + 1403.0, + 1978.0, + 1152.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1975.0, + 341.0, + 1975.0, + 341.0, + 2007.0, + 299.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1975.0, + 646.0, + 1975.0, + 646.0, + 2007.0, + 468.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 742.0, + 1404.0, + 742.0, + 1404.0, + 779.0, + 295.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 774.0, + 1407.0, + 774.0, + 1407.0, + 813.0, + 295.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 808.0, + 1101.0, + 808.0, + 1101.0, + 839.0, + 296.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 808.0, + 1175.0, + 808.0, + 1175.0, + 839.0, + 1126.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 808.0, + 1404.0, + 808.0, + 1404.0, + 839.0, + 1196.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 839.0, + 1403.0, + 839.0, + 1403.0, + 870.0, + 296.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 867.0, + 1404.0, + 867.0, + 1404.0, + 902.0, + 295.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 894.0, + 851.0, + 894.0, + 851.0, + 935.0, + 292.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 894.0, + 1407.0, + 894.0, + 1407.0, + 935.0, + 874.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 926.0, + 354.0, + 926.0, + 354.0, + 962.0, + 293.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 926.0, + 390.0, + 926.0, + 390.0, + 962.0, + 376.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 910.0, + 228.0, + 910.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 266.0, + 939.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 1146.0, + 262.0, + 1146.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 296.0, + 1167.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 1403.0, + 293.0, + 1403.0, + 326.0, + 296.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 323.0, + 1405.0, + 323.0, + 1405.0, + 356.0, + 295.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 357.0, + 1064.0, + 357.0, + 1064.0, + 400.0, + 291.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 357.0, + 1186.0, + 357.0, + 1186.0, + 400.0, + 1083.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 395.0, + 669.0, + 395.0, + 669.0, + 433.0, + 291.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 395.0, + 761.0, + 395.0, + 761.0, + 433.0, + 753.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 550.0, + 1403.0, + 550.0, + 1403.0, + 588.0, + 295.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 584.0, + 838.0, + 584.0, + 838.0, + 618.0, + 296.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 584.0, + 1405.0, + 584.0, + 1405.0, + 618.0, + 894.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 615.0, + 566.0, + 615.0, + 566.0, + 649.0, + 294.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 444.0, + 549.0, + 444.0, + 549.0, + 482.0, + 295.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 444.0, + 887.0, + 444.0, + 887.0, + 482.0, + 681.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 444.0, + 1405.0, + 444.0, + 1405.0, + 482.0, + 1004.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 477.0, + 707.0, + 477.0, + 707.0, + 511.0, + 293.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 477.0, + 1401.0, + 477.0, + 1401.0, + 511.0, + 729.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 504.0, + 1379.0, + 504.0, + 1379.0, + 544.0, + 295.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1408.0, + 1403.0, + 1408.0, + 1403.0, + 1443.0, + 295.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1438.0, + 664.0, + 1438.0, + 664.0, + 1470.0, + 296.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1710.0, + 340.0, + 1710.0, + 340.0, + 1749.0, + 295.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1349.0, + 1718.0, + 1400.0, + 1718.0, + 1400.0, + 1745.0, + 1349.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1743.0, + 340.0, + 1743.0, + 340.0, + 1781.0, + 295.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1121.0, + 372.0, + 1121.0, + 372.0, + 1163.0, + 295.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1121.0, + 444.0, + 1121.0, + 444.0, + 1163.0, + 393.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1121.0, + 751.0, + 1121.0, + 751.0, + 1163.0, + 464.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1121.0, + 827.0, + 1121.0, + 827.0, + 1163.0, + 776.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 1121.0, + 998.0, + 1121.0, + 998.0, + 1163.0, + 850.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1121.0, + 1406.0, + 1121.0, + 1406.0, + 1163.0, + 1063.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1155.0, + 843.0, + 1155.0, + 843.0, + 1196.0, + 295.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1155.0, + 911.0, + 1155.0, + 911.0, + 1196.0, + 864.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1155.0, + 1401.0, + 1155.0, + 1401.0, + 1196.0, + 1393.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1209.0, + 1403.0, + 1209.0, + 1403.0, + 1245.0, + 296.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1238.0, + 661.0, + 1238.0, + 661.0, + 1280.0, + 292.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1238.0, + 734.0, + 1238.0, + 734.0, + 1280.0, + 686.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1238.0, + 1275.0, + 1238.0, + 1275.0, + 1280.0, + 756.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 964.0, + 1154.0, + 964.0, + 1154.0, + 1014.0, + 292.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1792.0, + 613.0, + 1792.0, + 613.0, + 1838.0, + 291.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1792.0, + 1114.0, + 1792.0, + 1114.0, + 1838.0, + 1029.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1792.0, + 1410.0, + 1792.0, + 1410.0, + 1838.0, + 1353.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1827.0, + 739.0, + 1827.0, + 739.0, + 1865.0, + 651.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1827.0, + 941.0, + 1827.0, + 941.0, + 1865.0, + 889.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1827.0, + 1406.0, + 1827.0, + 1406.0, + 1865.0, + 1007.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1858.0, + 1156.0, + 1858.0, + 1156.0, + 1894.0, + 293.0, + 1894.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 227, + 1405, + 227, + 1405, + 459, + 296, + 459 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1650, + 1403, + 1650, + 1403, + 1806, + 297, + 1806 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1403, + 1819, + 1403, + 2035, + 297, + 2035 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 737, + 1407, + 737, + 1407, + 955, + 297, + 955 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 968, + 1403, + 968, + 1403, + 1124, + 298, + 1124 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1136, + 1403, + 1136, + 1403, + 1233, + 299, + 1233 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 445, + 1286, + 1250, + 1286, + 1250, + 1458, + 445, + 1458 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 295, + 1490, + 1401, + 1490, + 1401, + 1555, + 295, + 1555 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 547, + 551, + 1151, + 551, + 1151, + 722, + 547, + 722 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 475, + 1397, + 475, + 1397, + 541, + 296, + 541 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 299, + 1592, + 641, + 1592, + 641, + 1625, + 299, + 1625 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.794 + }, + { + "category_id": 13, + "poly": [ + 534, + 420, + 975, + 420, + 975, + 459, + 534, + 459 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { c ( s _ { t } ) \\approx \\dot { \\tilde { c } } \\dot { ( s _ { t } ) } : = \\frac { 1 } { 2 } s _ { t } ^ { T } C _ { t } s _ { t } + \\dot { d _ { t } ^ { T } } s _ { t } + h _ { t } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1199, + 437, + 1199, + 437, + 1233, + 298, + 1233 + ], + "score": 0.93, + "latex": "\\theta = \\left( \\omega , \\Sigma ^ { \\pi } \\right)" + }, + { + "category_id": 13, + "poly": [ + 921, + 261, + 1187, + 261, + 1187, + 301, + 921, + 301 + ], + "score": 0.93, + "latex": "\\boldsymbol { \\zeta } : = \\left( A _ { t } , B _ { t } , c _ { t } , F _ { t } \\right) _ { t = 1 } ^ { T }" + }, + { + "category_id": 14, + "poly": [ + 549, + 548, + 1151, + 548, + 1151, + 725, + 549, + 725 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } & { \\underset { \\eta } { \\operatorname* { m i n } } \\displaystyle \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\boldsymbol { \\tilde { \\psi } } } [ \\boldsymbol { \\tilde { c } } ( s _ { t } ) - \\lambda _ { t } ^ { T } \\boldsymbol { a } _ { t } ] + \\nu _ { t } D _ { t } ( \\eta , \\theta ) } & \\\\ & { \\mathrm { s . t . } \\displaystyle \\sum _ { t = 0 } ^ { T } \\mathbb { E } _ { \\boldsymbol { \\tilde { \\psi } } } [ D _ { K L } ( \\boldsymbol { \\psi } ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\eta ) | \\Big | \\psi ( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\eta ^ { \\prime } ) ) ] \\leq \\epsilon , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 882, + 1138, + 984, + 1138, + 984, + 1171, + 882, + 1171 + ], + "score": 0.92, + "latex": "D _ { t } \\left( { \\theta , \\eta } \\right)" + }, + { + "category_id": 13, + "poly": [ + 368, + 1137, + 527, + 1137, + 527, + 1172, + 368, + 1172 + ], + "score": 0.92, + "latex": "\\psi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } , t ; \\boldsymbol { \\eta } \\right)" + }, + { + "category_id": 13, + "poly": [ + 968, + 327, + 1035, + 327, + 1035, + 366, + 968, + 366 + ], + "score": 0.92, + "latex": "\\mathbb { E } _ { \\widetilde { \\psi } } \\left[ \\cdot \\right]" + }, + { + "category_id": 13, + "poly": [ + 582, + 1137, + 716, + 1137, + 716, + 1171, + 582, + 1171 + ], + "score": 0.92, + "latex": "\\pi \\left( \\boldsymbol { a } _ { t } | \\boldsymbol { s } _ { t } ; \\boldsymbol { \\theta } \\right)" + }, + { + "category_id": 14, + "poly": [ + 444, + 1285, + 1255, + 1285, + 1255, + 1461, + 444, + 1461 + ], + "score": 0.92, + "latex": "\\begin{array} { l } { \\displaystyle \\operatorname* { m i n } _ { \\theta } \\mathbb { E } _ { \\psi } [ \\sum _ { t = 0 } ^ { T } \\lambda _ { t } ^ { T } \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) + \\frac { \\nu _ { t } } { 2 } ( \\operatorname { t r } ( G _ { t } ^ { - 1 } \\Sigma ^ { \\pi } ) - \\log | \\Sigma ^ { \\pi } | ) } \\\\ { \\displaystyle + \\frac { \\nu _ { t } } { 2 } ( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) - \\mathbb { E } _ { \\psi } [ a _ { t } | s _ { t } , t ] ) ^ { T } G _ { t } ^ { - 1 } ( \\mu _ { \\omega } ^ { \\pi } ( o _ { t } ) - \\mathbb { E } _ { \\psi } [ a _ { t } | s _ { t } , t ] ) ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1161, + 472, + 1401, + 472, + 1401, + 513, + 1161, + 513 + ], + "score": 0.92, + "latex": "\\eta ~ = ~ ( K _ { t } , k _ { t } , G _ { t } ) _ { t = 1 } ^ { T }" + }, + { + "category_id": 13, + "poly": [ + 628, + 328, + 818, + 328, + 818, + 361, + 628, + 361 + ], + "score": 0.91, + "latex": "p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t \\right)" + }, + { + "category_id": 13, + "poly": [ + 1210, + 1031, + 1400, + 1031, + 1400, + 1064, + 1210, + 1064 + ], + "score": 0.91, + "latex": "p \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t \\right)" + }, + { + "category_id": 13, + "poly": [ + 296, + 265, + 836, + 265, + 836, + 300, + 296, + 300 + ], + "score": 0.91, + "latex": "\\tilde { p } \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t ; \\zeta \\right) = \\mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 583, + 1171, + 620, + 1171, + 620, + 1196, + 583, + 1196 + ], + "score": 0.87, + "latex": "\\Sigma ^ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 371, + 740, + 397, + 740, + 397, + 772, + 371, + 772 + ], + "score": 0.87, + "latex": "\\eta ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1281, + 1493, + 1303, + 1493, + 1303, + 1523, + 1281, + 1523 + ], + "score": 0.86, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 823, + 392, + 881, + 392, + 881, + 424, + 823, + 424 + ], + "score": 0.86, + "latex": "c ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 694, + 364, + 717, + 364, + 717, + 393, + 694, + 393 + ], + "score": 0.86, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 587, + 1005, + 614, + 1005, + 614, + 1031, + 587, + 1031 + ], + "score": 0.86, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1096, + 1524, + 1118, + 1524, + 1118, + 1555, + 1096, + 1555 + ], + "score": 0.86, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 1091, + 396, + 1116, + 396, + 1116, + 423, + 1091, + 423 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 998, + 742, + 1017, + 742, + 1017, + 771, + 998, + 771 + ], + "score": 0.85, + "latex": "\\tilde { p }" + }, + { + "category_id": 13, + "poly": [ + 426, + 424, + 449, + 424, + 449, + 454, + 426, + 454 + ], + "score": 0.85, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 328, + 303, + 354, + 303, + 354, + 328, + 328, + 328 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 587, + 773, + 609, + 773, + 609, + 802, + 587, + 802 + ], + "score": 0.84, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 954, + 1006, + 979, + 1006, + 979, + 1031, + 954, + 1031 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1019, + 426, + 1044, + 426, + 1044, + 453, + 1019, + 453 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 425, + 1032, + 447, + 1032, + 447, + 1063, + 425, + 1063 + ], + "score": 0.84, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 783, + 299, + 804, + 299, + 804, + 328, + 783, + 328 + ], + "score": 0.83, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 1112, + 1064, + 1138, + 1064, + 1138, + 1092, + 1112, + 1092 + ], + "score": 0.82, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1378, + 836, + 1401, + 836, + 1401, + 859, + 1378, + 859 + ], + "score": 0.81, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1181, + 363, + 1199, + 363, + 1199, + 394, + 1181, + 394 + ], + "score": 0.8, + "latex": "\\tilde { p }" + }, + { + "category_id": 13, + "poly": [ + 572, + 744, + 590, + 744, + 590, + 771, + 572, + 771 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 617, + 1093, + 634, + 1093, + 634, + 1119, + 617, + 1119 + ], + "score": 0.8, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 853, + 864, + 878, + 864, + 878, + 889, + 853, + 889 + ], + "score": 0.8, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1221, + 775, + 1241, + 775, + 1241, + 803, + 1221, + 803 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 581, + 1036, + 602, + 1036, + 602, + 1058, + 581, + 1058 + ], + "score": 0.8, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1354, + 1497, + 1374, + 1497, + 1374, + 1519, + 1354, + 1519 + ], + "score": 0.78, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 969, + 1528, + 990, + 1528, + 990, + 1550, + 969, + 1550 + ], + "score": 0.76, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 408, + 971, + 426, + 971, + 426, + 998, + 408, + 998 + ], + "score": 0.75, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1381, + 976, + 1401, + 976, + 1401, + 997, + 1381, + 997 + ], + "score": 0.67, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1065, + 742, + 1081, + 742, + 1081, + 768, + 1065, + 768 + ], + "score": 0.65, + "latex": "\\tilde { c }" + }, + { + "category_id": 13, + "poly": [ + 405, + 1526, + 418, + 1526, + 418, + 1551, + 405, + 1551 + ], + "score": 0.64, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1588.0, + 643.0, + 1588.0, + 643.0, + 1630.0, + 293.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1405.0, + 225.0, + 1405.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 250.0, + 295.0, + 250.0, + 295.0, + 314.0, + 285.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 250.0, + 920.0, + 250.0, + 920.0, + 314.0, + 837.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 250.0, + 1413.0, + 250.0, + 1413.0, + 314.0, + 1188.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 294.0, + 327.0, + 294.0, + 327.0, + 334.0, + 293.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 294.0, + 782.0, + 294.0, + 782.0, + 334.0, + 355.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 294.0, + 1407.0, + 294.0, + 1407.0, + 334.0, + 805.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 326.0, + 627.0, + 326.0, + 627.0, + 366.0, + 294.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 326.0, + 967.0, + 326.0, + 967.0, + 366.0, + 819.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 326.0, + 1406.0, + 326.0, + 1406.0, + 366.0, + 1036.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 359.0, + 693.0, + 359.0, + 693.0, + 397.0, + 291.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 359.0, + 1180.0, + 359.0, + 1180.0, + 397.0, + 718.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 359.0, + 1406.0, + 359.0, + 1406.0, + 397.0, + 1200.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 388.0, + 822.0, + 388.0, + 822.0, + 431.0, + 294.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 388.0, + 1090.0, + 388.0, + 1090.0, + 431.0, + 882.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 388.0, + 1405.0, + 388.0, + 1405.0, + 431.0, + 1117.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 421.0, + 425.0, + 421.0, + 425.0, + 461.0, + 293.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 421.0, + 533.0, + 421.0, + 533.0, + 461.0, + 450.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 421.0, + 1018.0, + 421.0, + 1018.0, + 461.0, + 976.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 421.0, + 1353.0, + 421.0, + 1353.0, + 461.0, + 1045.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1651.0, + 1400.0, + 1651.0, + 1400.0, + 1684.0, + 296.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1679.0, + 1403.0, + 1679.0, + 1403.0, + 1719.0, + 292.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1740.0, + 1405.0, + 1740.0, + 1405.0, + 1781.0, + 291.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1775.0, + 1206.0, + 1775.0, + 1206.0, + 1808.0, + 295.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 1407.0, + 1817.0, + 1407.0, + 1857.0, + 292.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1890.0, + 294.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1949.0, + 291.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 739.0, + 370.0, + 739.0, + 370.0, + 774.0, + 297.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 739.0, + 571.0, + 739.0, + 571.0, + 774.0, + 398.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 739.0, + 997.0, + 739.0, + 997.0, + 774.0, + 591.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 739.0, + 1064.0, + 739.0, + 1064.0, + 774.0, + 1018.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 739.0, + 1404.0, + 739.0, + 1404.0, + 774.0, + 1082.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 769.0, + 586.0, + 769.0, + 586.0, + 805.0, + 294.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 769.0, + 1220.0, + 769.0, + 1220.0, + 805.0, + 610.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 769.0, + 1405.0, + 769.0, + 1405.0, + 805.0, + 1242.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 837.0, + 294.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 831.0, + 1377.0, + 831.0, + 1377.0, + 865.0, + 295.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 861.0, + 852.0, + 861.0, + 852.0, + 894.0, + 294.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 861.0, + 1405.0, + 861.0, + 1405.0, + 894.0, + 879.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 893.0, + 1405.0, + 893.0, + 1405.0, + 925.0, + 294.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 924.0, + 1064.0, + 924.0, + 1064.0, + 953.0, + 295.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 970.0, + 407.0, + 970.0, + 407.0, + 1003.0, + 297.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 970.0, + 1380.0, + 970.0, + 1380.0, + 1003.0, + 427.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 997.0, + 586.0, + 997.0, + 586.0, + 1038.0, + 293.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 997.0, + 953.0, + 997.0, + 953.0, + 1038.0, + 615.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 997.0, + 1405.0, + 997.0, + 1405.0, + 1038.0, + 980.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1027.0, + 424.0, + 1027.0, + 424.0, + 1069.0, + 293.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1027.0, + 580.0, + 1027.0, + 580.0, + 1069.0, + 448.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1027.0, + 1209.0, + 1027.0, + 1209.0, + 1069.0, + 603.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1027.0, + 1404.0, + 1027.0, + 1404.0, + 1069.0, + 1401.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1061.0, + 1111.0, + 1061.0, + 1111.0, + 1094.0, + 296.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1061.0, + 1405.0, + 1061.0, + 1405.0, + 1094.0, + 1139.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1093.0, + 616.0, + 1093.0, + 616.0, + 1126.0, + 296.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1093.0, + 1319.0, + 1093.0, + 1319.0, + 1126.0, + 635.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1135.0, + 367.0, + 1135.0, + 367.0, + 1173.0, + 294.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1135.0, + 581.0, + 1135.0, + 581.0, + 1173.0, + 528.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1135.0, + 881.0, + 1135.0, + 881.0, + 1173.0, + 717.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 1135.0, + 1407.0, + 1135.0, + 1407.0, + 1173.0, + 985.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1165.0, + 582.0, + 1165.0, + 582.0, + 1205.0, + 292.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1165.0, + 1404.0, + 1165.0, + 1404.0, + 1205.0, + 621.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1198.0, + 297.0, + 1198.0, + 297.0, + 1234.0, + 293.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1198.0, + 475.0, + 1198.0, + 475.0, + 1234.0, + 438.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1492.0, + 1280.0, + 1492.0, + 1280.0, + 1524.0, + 295.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 1492.0, + 1353.0, + 1492.0, + 1353.0, + 1524.0, + 1304.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1375.0, + 1492.0, + 1405.0, + 1492.0, + 1405.0, + 1524.0, + 1375.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1521.0, + 404.0, + 1521.0, + 404.0, + 1557.0, + 295.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1521.0, + 968.0, + 1521.0, + 968.0, + 1557.0, + 419.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1521.0, + 1095.0, + 1521.0, + 1095.0, + 1557.0, + 991.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1521.0, + 1128.0, + 1521.0, + 1128.0, + 1557.0, + 1119.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 469.0, + 1160.0, + 469.0, + 1160.0, + 522.0, + 290.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 504.0, + 411.0, + 504.0, + 411.0, + 541.0, + 294.0, + 541.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 968, + 1404, + 968, + 1404, + 1297, + 296, + 1297 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 738, + 1405, + 738, + 1405, + 955, + 298, + 955 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1312, + 1405, + 1312, + 1405, + 1496, + 297, + 1496 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 296, + 1785, + 1406, + 1785, + 1406, + 1967, + 296, + 1967 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1510, + 1404, + 1510, + 1404, + 1604, + 299, + 1604 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 361, + 270, + 1332, + 270, + 1332, + 523, + 361, + 523 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 296, + 549, + 1407, + 549, + 1407, + 671, + 296, + 671 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 299, + 1707, + 1403, + 1707, + 1403, + 1773, + 299, + 1773 + ], + "score": 0.937 + }, + { + "category_id": 8, + "poly": [ + 372, + 1996, + 791, + 1996, + 791, + 2044, + 372, + 2044 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 299, + 1649, + 491, + 1649, + 491, + 1680, + 299, + 1680 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.693 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.294 + }, + { + "category_id": 13, + "poly": [ + 982, + 1121, + 1239, + 1121, + 1239, + 1176, + 982, + 1176 + ], + "score": 0.93, + "latex": "\\tilde { p } ^ { j } \\left( s _ { t + 1 } ^ { j } \\Big | s _ { t } ^ { j } , a _ { t } ^ { j } , t ; \\zeta ^ { j } \\right)" + }, + { + "category_id": 13, + "poly": [ + 932, + 1824, + 1046, + 1824, + 1046, + 1872, + 932, + 1872 + ], + "score": 0.93, + "latex": "\\left\\{ x ^ { ( i ) } \\right\\} _ { i = 0 } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 488, + 1876, + 645, + 1876, + 645, + 1940, + 488, + 1940 + ], + "score": 0.93, + "latex": "\\left\\{ \\hat { f } ( \\boldsymbol x ^ { ( i ) } ) \\right\\} _ { i = 0 } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 1307, + 1235, + 1400, + 1235, + 1400, + 1268, + 1307, + 1268 + ], + "score": 0.92, + "latex": "D \\left( \\eta , \\theta \\right)" + }, + { + "category_id": 13, + "poly": [ + 793, + 1204, + 926, + 1204, + 926, + 1238, + 793, + 1238 + ], + "score": 0.92, + "latex": "\\pi \\left( a _ { t } | \\boldsymbol s _ { t } \\bar { ; \\boldsymbol \\theta } \\right)" + }, + { + "category_id": 13, + "poly": [ + 1171, + 1819, + 1351, + 1819, + 1351, + 1882, + 1171, + 1882 + ], + "score": 0.92, + "latex": "\\left\\{ \\nabla \\hat { f } ( x ^ { ( i ) } ) \\right\\} _ { i = 0 } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 630, + 1511, + 694, + 1511, + 694, + 1545, + 630, + 1545 + ], + "score": 0.92, + "latex": "\\mu _ { \\omega } ^ { \\pi } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 901, + 1001, + 977, + 1001, + 977, + 1030, + 901, + 1030 + ], + "score": 0.92, + "latex": "A _ { t } , B _ { t }" + }, + { + "category_id": 13, + "poly": [ + 686, + 1709, + 738, + 1709, + 738, + 1743, + 686, + 1743 + ], + "score": 0.91, + "latex": "\\Psi ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 700, + 1373, + 793, + 1373, + 793, + 1406, + 700, + 1406 + ], + "score": 0.91, + "latex": "D \\left( \\eta , \\theta \\right)" + }, + { + "category_id": 13, + "poly": [ + 580, + 1709, + 631, + 1709, + 631, + 1743, + 580, + 1743 + ], + "score": 0.91, + "latex": "\\Phi ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1787, + 1266, + 1787, + 1266, + 1820, + 1214, + 1820 + ], + "score": 0.9, + "latex": "\\Phi ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1256, + 1174, + 1314, + 1174, + 1314, + 1207, + 1256, + 1207 + ], + "score": 0.9, + "latex": "\\tilde { c } ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 453, + 1030, + 1008, + 1030, + 1008, + 1064, + 453, + 1064 + ], + "score": 0.9, + "latex": "\\tilde { p } \\left( \\left. s _ { t + 1 } \\right| s _ { t } , a _ { t } , t ; \\zeta \\right) = \\mathcal { N } ( A _ { t } s _ { t } + B _ { t } a _ { t } + c _ { t } , F _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1207, + 1271, + 1207, + 1271, + 1235, + 1236, + 1235 + ], + "score": 0.89, + "latex": "K _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1184, + 1090, + 1400, + 1090, + 1400, + 1124, + 1184, + 1124 + ], + "score": 0.89, + "latex": "\\tilde { p } \\big ( s _ { t + 1 } \\big | s _ { t } , a _ { t } , t ; \\zeta \\big )" + }, + { + "category_id": 13, + "poly": [ + 1324, + 1206, + 1358, + 1206, + 1358, + 1234, + 1324, + 1234 + ], + "score": 0.89, + "latex": "G _ { t }" + }, + { + "category_id": 13, + "poly": [ + 756, + 1067, + 808, + 1067, + 808, + 1093, + 756, + 1093 + ], + "score": 0.89, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 1033, + 1001, + 1063, + 1001, + 1063, + 1030, + 1033, + 1030 + ], + "score": 0.88, + "latex": "F _ { t }" + }, + { + "category_id": 13, + "poly": [ + 757, + 1314, + 786, + 1314, + 786, + 1343, + 757, + 1343 + ], + "score": 0.88, + "latex": "\\lambda _ { t }" + }, + { + "category_id": 13, + "poly": [ + 887, + 1543, + 924, + 1543, + 924, + 1570, + 887, + 1570 + ], + "score": 0.87, + "latex": "\\Sigma ^ { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 465, + 1095, + 491, + 1095, + 491, + 1122, + 465, + 1122 + ], + "score": 0.87, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 546, + 1095, + 574, + 1095, + 574, + 1122, + 546, + 1122 + ], + "score": 0.86, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1177, + 1132, + 1177, + 1132, + 1205, + 1100, + 1205 + ], + "score": 0.86, + "latex": "C _ { t }" + }, + { + "category_id": 13, + "poly": [ + 849, + 971, + 871, + 971, + 871, + 1001, + 849, + 1001 + ], + "score": 0.86, + "latex": "\\psi" + }, + { + "category_id": 13, + "poly": [ + 637, + 1376, + 664, + 1376, + 664, + 1404, + 637, + 1404 + ], + "score": 0.84, + "latex": "\\nu _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1159, + 1715, + 1185, + 1715, + 1185, + 1740, + 1159, + 1740 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 14, + "poly": [ + 391, + 1998, + 793, + 1998, + 793, + 2043, + 391, + 2043 + ], + "score": 0.83, + "latex": "\\begin{array} { r } { \\overline { { \\boldsymbol { x } ^ { ( i ) } } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \\boldsymbol { x } ^ { ( j ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1375, + 1716, + 1400, + 1716, + 1400, + 1740, + 1375, + 1740 + ], + "score": 0.81, + "latex": "o _ { t }" + }, + { + "category_id": 13, + "poly": [ + 423, + 1177, + 439, + 1177, + 439, + 1206, + 423, + 1206 + ], + "score": 0.8, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 923, + 974, + 943, + 974, + 943, + 997, + 923, + 997 + ], + "score": 0.76, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1285, + 977, + 1303, + 977, + 1303, + 1001, + 1285, + 1001 + ], + "score": 0.76, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 453, + 1518, + 473, + 1518, + 473, + 1539, + 453, + 1539 + ], + "score": 0.74, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 882, + 1714, + 895, + 1714, + 895, + 1737, + 882, + 1737 + ], + "score": 0.68, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 280.0, + 618.0, + 280.0, + 618.0, + 294.0, + 552.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 280.0, + 961.0, + 280.0, + 961.0, + 294.0, + 896.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 280.0, + 1305.0, + 280.0, + 1305.0, + 294.0, + 1239.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 291.0, + 595.0, + 291.0, + 595.0, + 302.0, + 553.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 288.0, + 940.0, + 288.0, + 940.0, + 303.0, + 896.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 288.0, + 1284.0, + 288.0, + 1284.0, + 303.0, + 1239.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 298.0, + 624.0, + 298.0, + 624.0, + 313.0, + 552.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 298.0, + 968.0, + 298.0, + 968.0, + 313.0, + 896.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 298.0, + 1312.0, + 298.0, + 1312.0, + 313.0, + 1239.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 307.0, + 583.0, + 307.0, + 583.0, + 321.0, + 552.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 307.0, + 927.0, + 307.0, + 927.0, + 323.0, + 896.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 307.0, + 1272.0, + 307.0, + 1272.0, + 323.0, + 1239.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 317.0, + 588.0, + 317.0, + 588.0, + 331.0, + 551.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 317.0, + 931.0, + 317.0, + 931.0, + 331.0, + 896.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 317.0, + 1276.0, + 317.0, + 1276.0, + 331.0, + 1238.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 326.0, + 580.0, + 326.0, + 580.0, + 341.0, + 551.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 326.0, + 923.0, + 326.0, + 923.0, + 341.0, + 896.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 326.0, + 1267.0, + 326.0, + 1267.0, + 339.0, + 1236.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 333.0, + 590.0, + 333.0, + 590.0, + 352.0, + 552.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 333.0, + 934.0, + 333.0, + 934.0, + 352.0, + 895.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 333.0, + 1278.0, + 333.0, + 1278.0, + 352.0, + 1239.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 346.0, + 600.0, + 346.0, + 600.0, + 360.0, + 552.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 346.0, + 944.0, + 346.0, + 944.0, + 360.0, + 896.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 346.0, + 1060.0, + 346.0, + 1060.0, + 394.0, + 1048.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 346.0, + 1288.0, + 346.0, + 1288.0, + 360.0, + 1239.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 357.0, + 635.0, + 357.0, + 635.0, + 368.0, + 553.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 357.0, + 979.0, + 357.0, + 979.0, + 368.0, + 898.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 357.0, + 1322.0, + 357.0, + 1322.0, + 368.0, + 1240.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 399.0, + 726.0, + 399.0, + 726.0, + 408.0, + 717.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 422.0, + 384.0, + 422.0, + 384.0, + 432.0, + 373.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 462.0, + 420.0, + 462.0, + 420.0, + 472.0, + 410.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 459.0, + 455.0, + 459.0, + 455.0, + 473.0, + 440.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 460.0, + 487.0, + 460.0, + 487.0, + 473.0, + 472.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 459.0, + 521.0, + 459.0, + 521.0, + 473.0, + 503.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 460.0, + 552.0, + 460.0, + 552.0, + 473.0, + 535.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 460.0, + 584.0, + 460.0, + 584.0, + 473.0, + 568.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 460.0, + 616.0, + 460.0, + 616.0, + 473.0, + 599.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 460.0, + 767.0, + 460.0, + 767.0, + 473.0, + 753.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 460.0, + 800.0, + 460.0, + 800.0, + 473.0, + 783.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 459.0, + 831.0, + 459.0, + 831.0, + 473.0, + 815.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 459.0, + 863.0, + 459.0, + 863.0, + 473.0, + 847.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 460.0, + 897.0, + 460.0, + 897.0, + 473.0, + 879.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 460.0, + 928.0, + 460.0, + 928.0, + 473.0, + 911.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 460.0, + 960.0, + 460.0, + 960.0, + 473.0, + 942.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 460.0, + 1111.0, + 460.0, + 1111.0, + 473.0, + 1096.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 460.0, + 1143.0, + 460.0, + 1143.0, + 473.0, + 1127.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 460.0, + 1176.0, + 460.0, + 1176.0, + 473.0, + 1159.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 460.0, + 1208.0, + 460.0, + 1208.0, + 473.0, + 1191.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 460.0, + 1239.0, + 460.0, + 1239.0, + 473.0, + 1222.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 460.0, + 1272.0, + 460.0, + 1272.0, + 473.0, + 1255.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 460.0, + 1303.0, + 460.0, + 1303.0, + 473.0, + 1287.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 467.0, + 526.0, + 467.0, + 526.0, + 481.0, + 495.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 467.0, + 870.0, + 467.0, + 870.0, + 481.0, + 839.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 467.0, + 1214.0, + 467.0, + 1214.0, + 481.0, + 1182.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 494.0, + 527.0, + 494.0, + 527.0, + 527.0, + 485.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 492.0, + 871.0, + 492.0, + 871.0, + 528.0, + 830.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 494.0, + 1214.0, + 494.0, + 1214.0, + 528.0, + 1174.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 548.0, + 1405.0, + 548.0, + 1405.0, + 585.0, + 294.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 581.0, + 1404.0, + 581.0, + 1404.0, + 611.0, + 296.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 609.0, + 1404.0, + 609.0, + 1404.0, + 646.0, + 294.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 638.0, + 551.0, + 638.0, + 551.0, + 675.0, + 293.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1645.0, + 493.0, + 1645.0, + 493.0, + 1684.0, + 294.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 966.0, + 848.0, + 966.0, + 848.0, + 1005.0, + 293.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 966.0, + 922.0, + 966.0, + 922.0, + 1005.0, + 872.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 966.0, + 1284.0, + 966.0, + 1284.0, + 1005.0, + 944.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 966.0, + 1405.0, + 966.0, + 1405.0, + 1005.0, + 1304.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1000.0, + 900.0, + 1000.0, + 900.0, + 1034.0, + 295.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1000.0, + 1032.0, + 1000.0, + 1032.0, + 1034.0, + 978.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1000.0, + 1404.0, + 1000.0, + 1404.0, + 1034.0, + 1064.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1029.0, + 452.0, + 1029.0, + 452.0, + 1067.0, + 294.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1067.0, + 1009.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1059.0, + 755.0, + 1059.0, + 755.0, + 1095.0, + 293.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1059.0, + 1405.0, + 1059.0, + 1405.0, + 1095.0, + 809.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1089.0, + 464.0, + 1089.0, + 464.0, + 1126.0, + 295.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1089.0, + 545.0, + 1089.0, + 545.0, + 1126.0, + 492.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1089.0, + 1183.0, + 1089.0, + 1183.0, + 1126.0, + 575.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1122.0, + 981.0, + 1122.0, + 981.0, + 1179.0, + 289.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1122.0, + 1410.0, + 1122.0, + 1410.0, + 1179.0, + 1240.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1174.0, + 422.0, + 1174.0, + 422.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1174.0, + 1099.0, + 1174.0, + 1099.0, + 1208.0, + 440.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1174.0, + 1255.0, + 1174.0, + 1255.0, + 1208.0, + 1133.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1174.0, + 1402.0, + 1174.0, + 1402.0, + 1208.0, + 1315.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1202.0, + 792.0, + 1202.0, + 792.0, + 1241.0, + 293.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1202.0, + 1235.0, + 1202.0, + 1235.0, + 1241.0, + 927.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1202.0, + 1323.0, + 1202.0, + 1323.0, + 1241.0, + 1272.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1202.0, + 1405.0, + 1202.0, + 1405.0, + 1241.0, + 1359.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1234.0, + 1306.0, + 1234.0, + 1306.0, + 1271.0, + 295.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1234.0, + 1404.0, + 1234.0, + 1404.0, + 1271.0, + 1401.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1264.0, + 1176.0, + 1264.0, + 1176.0, + 1298.0, + 296.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 741.0, + 1403.0, + 741.0, + 1403.0, + 772.0, + 295.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 770.0, + 1406.0, + 770.0, + 1406.0, + 804.0, + 295.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 801.0, + 1403.0, + 801.0, + 1403.0, + 835.0, + 295.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 833.0, + 1403.0, + 833.0, + 1403.0, + 864.0, + 296.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 863.0, + 1405.0, + 863.0, + 1405.0, + 895.0, + 293.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 891.0, + 1407.0, + 891.0, + 1407.0, + 926.0, + 293.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 923.0, + 1163.0, + 923.0, + 1163.0, + 958.0, + 295.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1311.0, + 756.0, + 1311.0, + 756.0, + 1347.0, + 296.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1311.0, + 1406.0, + 1311.0, + 1406.0, + 1347.0, + 787.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1342.0, + 1403.0, + 1342.0, + 1403.0, + 1377.0, + 292.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1374.0, + 636.0, + 1374.0, + 636.0, + 1409.0, + 295.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1374.0, + 699.0, + 1374.0, + 699.0, + 1409.0, + 665.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1409.0, + 794.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1404.0, + 1403.0, + 1404.0, + 1403.0, + 1440.0, + 294.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1433.0, + 1405.0, + 1433.0, + 1405.0, + 1471.0, + 292.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1466.0, + 1368.0, + 1466.0, + 1368.0, + 1498.0, + 295.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1784.0, + 1213.0, + 1784.0, + 1213.0, + 1823.0, + 296.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1784.0, + 1406.0, + 1784.0, + 1406.0, + 1823.0, + 1267.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1822.0, + 931.0, + 1822.0, + 931.0, + 1882.0, + 285.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1822.0, + 1170.0, + 1822.0, + 1170.0, + 1882.0, + 1047.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1831.0, + 1403.0, + 1831.0, + 1403.0, + 1868.0, + 1354.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1876.0, + 487.0, + 1876.0, + 487.0, + 1936.0, + 293.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1886.0, + 1406.0, + 1886.0, + 1406.0, + 1929.0, + 659.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1930.0, + 1014.0, + 1930.0, + 1014.0, + 1973.0, + 295.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1510.0, + 452.0, + 1510.0, + 452.0, + 1547.0, + 294.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1510.0, + 629.0, + 1510.0, + 629.0, + 1547.0, + 474.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1510.0, + 1405.0, + 1510.0, + 1405.0, + 1547.0, + 695.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1541.0, + 886.0, + 1541.0, + 886.0, + 1578.0, + 293.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1541.0, + 1405.0, + 1541.0, + 1405.0, + 1578.0, + 925.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1574.0, + 536.0, + 1574.0, + 536.0, + 1604.0, + 295.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1702.0, + 579.0, + 1702.0, + 579.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1702.0, + 685.0, + 1702.0, + 685.0, + 1748.0, + 632.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1702.0, + 881.0, + 1702.0, + 881.0, + 1748.0, + 739.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1702.0, + 1158.0, + 1702.0, + 1158.0, + 1748.0, + 896.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1702.0, + 1374.0, + 1702.0, + 1374.0, + 1748.0, + 1186.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1702.0, + 1404.0, + 1702.0, + 1404.0, + 1748.0, + 1401.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1737.0, + 445.0, + 1737.0, + 445.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1787, + 1404, + 1787, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1278, + 1405, + 1278, + 1405, + 1404, + 298, + 1404 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 853, + 1402, + 853, + 1402, + 948, + 298, + 948 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 356, + 269, + 1331, + 269, + 1331, + 524, + 356, + 524 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 295, + 548, + 1407, + 548, + 1407, + 672, + 295, + 672 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 368, + 722, + 906, + 722, + 906, + 830, + 368, + 830 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 296, + 1711, + 1399, + 1711, + 1399, + 1775, + 296, + 1775 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 300, + 1202, + 1399, + 1202, + 1399, + 1265, + 300, + 1265 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 298, + 1642, + 559, + 1642, + 559, + 1677, + 298, + 1677 + ], + "score": 0.919 + }, + { + "category_id": 8, + "poly": [ + 368, + 1427, + 798, + 1427, + 798, + 1609, + 368, + 1609 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.909 + }, + { + "category_id": 8, + "poly": [ + 369, + 971, + 997, + 971, + 997, + 1180, + 369, + 1180 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.805 + }, + { + "category_id": 8, + "poly": [ + 367, + 971, + 997, + 971, + 997, + 1180, + 367, + 1180 + ], + "score": 0.485 + }, + { + "category_id": 13, + "poly": [ + 512, + 854, + 564, + 854, + 564, + 888, + 512, + 888 + ], + "score": 0.91, + "latex": "\\Phi ( \\cdot )" + }, + { + "category_id": 14, + "poly": [ + 362, + 1426, + 801, + 1426, + 801, + 1611, + 362, + 1611 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } { \\bullet } & { \\left( \\hat { f } ( x ^ { ( t ) } ) - \\hat { f } ( x ^ { ( t - 1 ) } ) \\right) / \\hat { f } ( x ^ { ( t - 1 ) } ) } \\\\ & { \\bullet \\ \\nabla \\hat { f } ( x ^ { ( t ) } ) / \\left( \\left| \\nabla \\hat { f } ( x ^ { ( \\operatorname* { m a x } ( t - 1 , 0 ) ) } ) \\right| + 1 \\right) } \\\\ & { \\bullet \\ \\frac { \\left| x ^ { ( \\operatorname* { m a x } ( t - 1 , 1 ) ) } - x ^ { ( \\operatorname* { m a x } ( t - 2 , 0 ) ) } \\right| } { \\left| x ^ { ( t ) } - x ^ { ( t - 1 ) } \\right| + 0 . 1 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 392, + 1311, + 444, + 1311, + 444, + 1344, + 392, + 1344 + ], + "score": 0.91, + "latex": "\\Psi ( \\cdot )" + }, + { + "category_id": 14, + "poly": [ + 365, + 971, + 996, + 971, + 996, + 1184, + 365, + 1184 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { \\bullet \\{ ( \\hat { f } ( x ^ { ( t - 5 i ) } ) - \\widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) ) / \\widehat { f } ( x ^ { ( t - 5 ( i + 1 ) ) } ) \\} _ { i = 0 } ^ { 2 4 } } \\\\ & { \\bullet \\{ \\overline { { \\nabla \\hat { f } ( x ^ { ( t - 5 i ) } ) } } / ( | \\overline { { \\nabla \\hat { f } ( x ^ { ( m a x ( t - 5 ( i + 1 ) , t \\mathrm { m o d } 5 ) ) } ) } } | + 1 ) \\} _ { i = 0 } ^ { 2 5 } } \\\\ & \\bullet \\{ \\frac { | \\frac { \\overline { { x ^ { ( \\mathrm { m a x } ( t - 5 ( i + 1 ) , t \\mathrm { m o d } 5 + 5 ) ) } } } - \\overline { { x ^ { ( \\mathrm { m a x } ( t - 5 ( i + 2 ) , t \\mathrm { m o d } 5 ) ) } } } } { | \\overline { { x ^ { ( t - 5 i ) } } } - \\overline { { x ^ { ( t - 5 ( i + 1 ) ) } } } | + 0 . 1 } \\} _ { i = 0 } ^ { 2 4 } \\} _ { i = 0 } ^ { 2 } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 362, + 723, + 907, + 723, + 907, + 831, + 362, + 831 + ], + "score": 0.81, + "latex": "\\begin{array} { r l } & { \\bullet \\ \\overline { { \\nabla \\hat { f } ( x ^ { ( i ) } ) } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } { \\nabla \\hat { f } ( x ^ { ( j ) } ) } } \\\\ & { \\bullet \\ \\overline { { \\hat { f } ( x ^ { ( i ) } ) } } : = \\frac { 1 } { \\operatorname* { m i n } ( i + 1 , 3 ) } \\sum _ { j = \\operatorname* { m a x } ( i - 2 , 0 ) } ^ { i } \\hat { f } ( x ^ { ( j ) } ) } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 279.0, + 617.0, + 279.0, + 617.0, + 293.0, + 552.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 279.0, + 961.0, + 279.0, + 961.0, + 293.0, + 895.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 279.0, + 1306.0, + 279.0, + 1306.0, + 293.0, + 1239.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 289.0, + 380.0, + 289.0, + 380.0, + 298.0, + 371.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 290.0, + 596.0, + 290.0, + 596.0, + 301.0, + 553.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 290.0, + 940.0, + 290.0, + 940.0, + 301.0, + 896.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 286.0, + 1072.0, + 286.0, + 1072.0, + 298.0, + 1056.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 290.0, + 1284.0, + 290.0, + 1284.0, + 301.0, + 1240.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 297.0, + 624.0, + 297.0, + 624.0, + 311.0, + 552.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 297.0, + 967.0, + 297.0, + 967.0, + 311.0, + 894.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 298.0, + 1312.0, + 298.0, + 1312.0, + 312.0, + 1239.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 308.0, + 588.0, + 308.0, + 588.0, + 331.0, + 551.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 307.0, + 932.0, + 307.0, + 932.0, + 333.0, + 894.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 311.0, + 1069.0, + 311.0, + 1069.0, + 320.0, + 1057.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 307.0, + 1276.0, + 307.0, + 1276.0, + 332.0, + 1239.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 331.0, + 382.0, + 331.0, + 382.0, + 341.0, + 371.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 327.0, + 590.0, + 327.0, + 590.0, + 352.0, + 551.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 335.0, + 726.0, + 335.0, + 726.0, + 349.0, + 712.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 327.0, + 935.0, + 327.0, + 935.0, + 352.0, + 894.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 335.0, + 1067.0, + 335.0, + 1067.0, + 345.0, + 1057.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 327.0, + 1278.0, + 327.0, + 1278.0, + 352.0, + 1238.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 345.0, + 601.0, + 345.0, + 601.0, + 360.0, + 551.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 339.0, + 733.0, + 339.0, + 733.0, + 401.0, + 698.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 345.0, + 944.0, + 345.0, + 944.0, + 360.0, + 895.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 345.0, + 1288.0, + 345.0, + 1288.0, + 360.0, + 1239.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 355.0, + 636.0, + 355.0, + 636.0, + 370.0, + 552.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 355.0, + 980.0, + 355.0, + 980.0, + 370.0, + 896.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 355.0, + 1324.0, + 355.0, + 1324.0, + 370.0, + 1239.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 369.0, + 555.0, + 369.0, + 555.0, + 377.0, + 546.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 413.0, + 641.0, + 413.0, + 641.0, + 425.0, + 625.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 410.0, + 1071.0, + 410.0, + 1071.0, + 422.0, + 1056.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 419.0, + 727.0, + 419.0, + 727.0, + 429.0, + 716.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 462.0, + 420.0, + 462.0, + 420.0, + 472.0, + 411.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 462.0, + 453.0, + 462.0, + 453.0, + 471.0, + 441.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 462.0, + 485.0, + 462.0, + 485.0, + 472.0, + 474.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 461.0, + 520.0, + 461.0, + 520.0, + 473.0, + 503.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 461.0, + 551.0, + 461.0, + 551.0, + 473.0, + 536.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 461.0, + 584.0, + 461.0, + 584.0, + 473.0, + 567.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 461.0, + 616.0, + 461.0, + 616.0, + 473.0, + 599.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 462.0, + 764.0, + 462.0, + 764.0, + 472.0, + 754.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 461.0, + 800.0, + 461.0, + 800.0, + 473.0, + 783.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 461.0, + 831.0, + 461.0, + 831.0, + 473.0, + 816.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 461.0, + 863.0, + 461.0, + 863.0, + 473.0, + 847.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 461.0, + 895.0, + 461.0, + 895.0, + 473.0, + 879.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 461.0, + 927.0, + 461.0, + 927.0, + 473.0, + 912.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 462.0, + 958.0, + 462.0, + 958.0, + 471.0, + 944.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 461.0, + 1110.0, + 461.0, + 1110.0, + 473.0, + 1095.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 461.0, + 1144.0, + 461.0, + 1144.0, + 473.0, + 1127.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 461.0, + 1176.0, + 461.0, + 1176.0, + 473.0, + 1159.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 461.0, + 1207.0, + 461.0, + 1207.0, + 473.0, + 1190.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 461.0, + 1239.0, + 461.0, + 1239.0, + 473.0, + 1223.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 461.0, + 1270.0, + 461.0, + 1270.0, + 473.0, + 1256.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 461.0, + 1303.0, + 461.0, + 1303.0, + 473.0, + 1287.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 467.0, + 527.0, + 467.0, + 527.0, + 482.0, + 494.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 467.0, + 870.0, + 467.0, + 870.0, + 482.0, + 839.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 467.0, + 1214.0, + 467.0, + 1214.0, + 482.0, + 1183.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 495.0, + 525.0, + 495.0, + 525.0, + 525.0, + 487.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 492.0, + 870.0, + 492.0, + 870.0, + 526.0, + 829.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 493.0, + 1215.0, + 493.0, + 1215.0, + 527.0, + 1174.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 369.0, + 535.0, + 369.0, + 535.0, + 392.5, + 487.0, + 392.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 380.0, + 775.0, + 380.0, + 775.0, + 397.0, + 742.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 400.5, + 828.0, + 400.5, + 828.0, + 421.5, + 754.0, + 421.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.75, + 397.0, + 873.75, + 397.0, + 873.75, + 419.5, + 763.75, + 419.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.25, + 394.0, + 854.25, + 394.0, + 854.25, + 405.5, + 827.25, + 405.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.75, + 423.5, + 639.75, + 423.5, + 639.75, + 440.5, + 480.75, + 440.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.75, + 425.5, + 928.75, + 425.5, + 928.75, + 439.5, + 880.75, + 439.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.25, + 434.0, + 938.25, + 434.0, + 938.25, + 442.5, + 928.25, + 442.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 548.0, + 1406.0, + 548.0, + 1406.0, + 583.0, + 293.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 580.0, + 1404.0, + 580.0, + 1404.0, + 612.0, + 294.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 610.0, + 1405.0, + 610.0, + 1405.0, + 645.0, + 293.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 639.0, + 658.0, + 639.0, + 658.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1642.0, + 561.0, + 1642.0, + 561.0, + 1682.0, + 293.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1790.0, + 1404.0, + 1790.0, + 1404.0, + 1824.0, + 297.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1822.0, + 1404.0, + 1822.0, + 1404.0, + 1854.0, + 294.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1852.0, + 1404.0, + 1852.0, + 1404.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1406.0, + 1912.0, + 1406.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1276.0, + 1405.0, + 1276.0, + 1405.0, + 1316.0, + 293.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1308.0, + 391.0, + 1308.0, + 391.0, + 1346.0, + 292.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1308.0, + 1406.0, + 1308.0, + 1406.0, + 1346.0, + 445.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1340.0, + 1403.0, + 1340.0, + 1403.0, + 1375.0, + 293.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1367.0, + 669.0, + 1367.0, + 669.0, + 1409.0, + 292.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 852.0, + 511.0, + 852.0, + 511.0, + 891.0, + 293.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 852.0, + 1406.0, + 852.0, + 1406.0, + 891.0, + 565.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 885.0, + 1404.0, + 885.0, + 1404.0, + 919.0, + 293.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 918.0, + 1381.0, + 918.0, + 1381.0, + 952.0, + 294.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1707.0, + 1405.0, + 1707.0, + 1405.0, + 1749.0, + 292.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1739.0, + 1113.0, + 1739.0, + 1113.0, + 1779.0, + 292.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1202.0, + 1402.0, + 1202.0, + 1402.0, + 1234.0, + 296.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1234.0, + 1228.0, + 1234.0, + 1228.0, + 1265.0, + 297.0, + 1265.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1697, + 1405, + 1697, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1438, + 1404, + 1438, + 1404, + 1682, + 298, + 1682 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1268, + 1403, + 1268, + 1403, + 1423, + 298, + 1423 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 357, + 775, + 1331, + 775, + 1331, + 1032, + 357, + 1032 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 361, + 270, + 1331, + 270, + 1331, + 523, + 361, + 523 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 296, + 548, + 1407, + 548, + 1407, + 672, + 296, + 672 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 297, + 1057, + 1404, + 1057, + 1404, + 1180, + 297, + 1180 + ], + "score": 0.964 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.793 + }, + { + "category_id": 15, + "poly": [ + 551.0, + 787.0, + 624.0, + 787.0, + 624.0, + 821.0, + 551.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 805.0, + 725.0, + 805.0, + 725.0, + 814.0, + 713.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 799.0, + 785.0, + 799.0, + 785.0, + 811.0, + 770.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 790.0, + 968.0, + 790.0, + 968.0, + 821.0, + 895.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 795.0, + 1067.0, + 795.0, + 1067.0, + 804.0, + 1058.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 790.0, + 1312.0, + 790.0, + 1312.0, + 820.0, + 1239.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 815.0, + 584.0, + 815.0, + 584.0, + 830.0, + 551.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 815.0, + 928.0, + 815.0, + 928.0, + 830.0, + 894.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 815.0, + 1271.0, + 815.0, + 1271.0, + 830.0, + 1239.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 825.0, + 588.0, + 825.0, + 588.0, + 840.0, + 551.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 825.0, + 931.0, + 825.0, + 931.0, + 840.0, + 894.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 824.0, + 1276.0, + 824.0, + 1276.0, + 840.0, + 1239.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 839.0, + 381.0, + 839.0, + 381.0, + 905.0, + 358.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 834.0, + 600.0, + 834.0, + 600.0, + 868.0, + 551.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 847.0, + 725.0, + 847.0, + 725.0, + 856.0, + 714.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 834.0, + 945.0, + 834.0, + 945.0, + 868.0, + 894.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 842.0, + 1069.0, + 842.0, + 1069.0, + 906.0, + 1042.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 833.0, + 1288.0, + 833.0, + 1288.0, + 868.0, + 1239.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 863.0, + 636.0, + 863.0, + 636.0, + 878.0, + 552.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 863.0, + 979.0, + 863.0, + 979.0, + 878.0, + 895.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 870.0, + 1068.0, + 870.0, + 1068.0, + 879.0, + 1057.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 863.0, + 1324.0, + 863.0, + 1324.0, + 878.0, + 1239.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 917.0, + 1071.0, + 917.0, + 1071.0, + 929.0, + 1055.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 929.0, + 724.0, + 929.0, + 724.0, + 935.0, + 719.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 955.0, + 512.0, + 955.0, + 512.0, + 969.0, + 496.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 968.0, + 422.0, + 968.0, + 422.0, + 980.0, + 408.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 966.0, + 457.0, + 966.0, + 457.0, + 983.0, + 437.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 968.0, + 487.0, + 968.0, + 487.0, + 982.0, + 473.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 972.0, + 518.0, + 972.0, + 518.0, + 980.0, + 506.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 966.0, + 553.0, + 966.0, + 553.0, + 983.0, + 534.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 968.0, + 583.0, + 968.0, + 583.0, + 982.0, + 568.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 968.0, + 615.0, + 968.0, + 615.0, + 982.0, + 599.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 968.0, + 767.0, + 968.0, + 767.0, + 982.0, + 752.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 968.0, + 799.0, + 968.0, + 799.0, + 982.0, + 782.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 968.0, + 832.0, + 968.0, + 832.0, + 982.0, + 816.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 964.0, + 874.0, + 964.0, + 874.0, + 993.0, + 835.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 968.0, + 897.0, + 968.0, + 897.0, + 982.0, + 879.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 968.0, + 928.0, + 968.0, + 928.0, + 982.0, + 911.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 968.0, + 959.0, + 968.0, + 959.0, + 982.0, + 942.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 968.0, + 1110.0, + 968.0, + 1110.0, + 982.0, + 1096.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 968.0, + 1144.0, + 968.0, + 1144.0, + 982.0, + 1127.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 968.0, + 1175.0, + 968.0, + 1175.0, + 982.0, + 1158.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 964.0, + 1218.0, + 964.0, + 1218.0, + 993.0, + 1180.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 968.0, + 1240.0, + 968.0, + 1240.0, + 982.0, + 1223.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 968.0, + 1272.0, + 968.0, + 1272.0, + 982.0, + 1255.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 968.0, + 1303.0, + 968.0, + 1303.0, + 980.0, + 1287.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 978.0, + 511.0, + 978.0, + 511.0, + 986.0, + 497.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1003.0, + 525.0, + 1003.0, + 525.0, + 1034.0, + 487.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 999.0, + 871.0, + 999.0, + 871.0, + 1035.0, + 828.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1001.0, + 1215.0, + 1001.0, + 1215.0, + 1036.0, + 1173.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 872.5, + 986.0, + 872.5, + 986.0, + 915.5, + 753.0, + 915.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 854.5, + 768.0, + 854.5, + 768.0, + 866.5, + 758.0, + 866.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 871.0, + 1191.0, + 871.0, + 1191.0, + 885.5, + 1150.0, + 885.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.25, + 893.0, + 642.25, + 893.0, + 642.25, + 929.5, + 460.25, + 929.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 280.0, + 618.0, + 280.0, + 618.0, + 294.0, + 552.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 280.0, + 961.0, + 280.0, + 961.0, + 294.0, + 895.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 280.0, + 1305.0, + 280.0, + 1305.0, + 294.0, + 1239.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 288.0, + 941.0, + 288.0, + 941.0, + 303.0, + 895.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 288.0, + 1071.0, + 288.0, + 1071.0, + 297.0, + 1058.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 286.0, + 1285.0, + 286.0, + 1285.0, + 304.0, + 1237.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 298.0, + 624.0, + 298.0, + 624.0, + 313.0, + 552.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 301.0, + 727.0, + 301.0, + 727.0, + 310.0, + 717.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 297.0, + 967.0, + 297.0, + 967.0, + 312.0, + 894.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 298.0, + 1312.0, + 298.0, + 1312.0, + 313.0, + 1239.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 307.0, + 588.0, + 307.0, + 588.0, + 332.0, + 550.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 307.0, + 932.0, + 307.0, + 932.0, + 331.0, + 895.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 307.0, + 1276.0, + 307.0, + 1276.0, + 331.0, + 1238.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 330.0, + 447.0, + 330.0, + 447.0, + 339.0, + 438.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 326.0, + 600.0, + 326.0, + 600.0, + 360.0, + 550.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 333.0, + 733.0, + 333.0, + 733.0, + 401.0, + 702.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 326.0, + 944.0, + 326.0, + 944.0, + 360.0, + 895.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 345.0, + 1061.0, + 345.0, + 1061.0, + 395.0, + 1047.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 326.0, + 1288.0, + 326.0, + 1288.0, + 360.0, + 1238.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 357.0, + 634.0, + 357.0, + 634.0, + 368.0, + 554.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 357.0, + 979.0, + 357.0, + 979.0, + 368.0, + 897.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 359.0, + 1070.0, + 359.0, + 1070.0, + 368.0, + 1061.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 353.0, + 1324.0, + 353.0, + 1324.0, + 370.0, + 1238.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 393.0, + 759.0, + 393.0, + 759.0, + 403.0, + 751.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 394.0, + 1071.0, + 394.0, + 1071.0, + 403.0, + 1062.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 400.0, + 728.0, + 400.0, + 728.0, + 409.0, + 717.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 410.0, + 428.0, + 410.0, + 428.0, + 419.0, + 418.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 424.0, + 383.0, + 424.0, + 383.0, + 433.0, + 373.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 434.0, + 639.0, + 434.0, + 639.0, + 442.0, + 616.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 433.0, + 728.0, + 433.0, + 728.0, + 442.0, + 719.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 430.0, + 939.0, + 430.0, + 939.0, + 441.0, + 911.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 435.0, + 963.0, + 435.0, + 963.0, + 443.0, + 950.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 436.0, + 982.0, + 436.0, + 982.0, + 444.0, + 964.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 432.0, + 1069.0, + 432.0, + 1069.0, + 437.0, + 1063.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 460.0, + 422.0, + 460.0, + 422.0, + 473.0, + 408.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 460.0, + 455.0, + 460.0, + 455.0, + 473.0, + 440.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 460.0, + 487.0, + 460.0, + 487.0, + 473.0, + 470.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 456.0, + 529.0, + 456.0, + 529.0, + 485.0, + 492.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 460.0, + 553.0, + 460.0, + 553.0, + 473.0, + 535.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 460.0, + 584.0, + 460.0, + 584.0, + 473.0, + 567.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 460.0, + 615.0, + 460.0, + 615.0, + 473.0, + 600.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 462.0, + 765.0, + 462.0, + 765.0, + 472.0, + 755.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 460.0, + 799.0, + 460.0, + 799.0, + 473.0, + 782.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 460.0, + 832.0, + 460.0, + 832.0, + 473.0, + 816.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 460.0, + 865.0, + 460.0, + 865.0, + 473.0, + 846.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 460.0, + 896.0, + 460.0, + 896.0, + 473.0, + 879.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 460.0, + 927.0, + 460.0, + 927.0, + 473.0, + 911.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 460.0, + 960.0, + 460.0, + 960.0, + 473.0, + 942.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 462.0, + 1109.0, + 462.0, + 1109.0, + 472.0, + 1098.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 460.0, + 1145.0, + 460.0, + 1145.0, + 473.0, + 1127.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 460.0, + 1176.0, + 460.0, + 1176.0, + 473.0, + 1158.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 460.0, + 1208.0, + 460.0, + 1208.0, + 473.0, + 1190.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 460.0, + 1239.0, + 460.0, + 1239.0, + 473.0, + 1223.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 460.0, + 1272.0, + 460.0, + 1272.0, + 473.0, + 1254.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 460.0, + 1303.0, + 460.0, + 1303.0, + 473.0, + 1286.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 467.0, + 871.0, + 467.0, + 871.0, + 481.0, + 838.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 467.0, + 1214.0, + 467.0, + 1214.0, + 481.0, + 1183.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 494.0, + 527.0, + 494.0, + 527.0, + 527.0, + 485.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 492.0, + 871.0, + 492.0, + 871.0, + 527.0, + 829.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 494.0, + 1214.0, + 494.0, + 1214.0, + 528.0, + 1174.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 288.5, + 596.0, + 288.5, + 596.0, + 302.5, + 551.0, + 302.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 403.0, + 876.0, + 403.0, + 876.0, + 429.0, + 770.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 548.0, + 1407.0, + 548.0, + 1407.0, + 584.0, + 294.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 580.0, + 1407.0, + 580.0, + 1407.0, + 612.0, + 295.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 610.0, + 1404.0, + 610.0, + 1404.0, + 645.0, + 293.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 639.0, + 553.0, + 639.0, + 553.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1058.0, + 1402.0, + 1058.0, + 1402.0, + 1091.0, + 295.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1087.0, + 1404.0, + 1087.0, + 1404.0, + 1122.0, + 292.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1116.0, + 1404.0, + 1116.0, + 1404.0, + 1155.0, + 294.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1149.0, + 658.0, + 1149.0, + 658.0, + 1182.0, + 294.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1700.0, + 1403.0, + 1700.0, + 1403.0, + 1731.0, + 295.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1764.0, + 296.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1761.0, + 1405.0, + 1761.0, + 1405.0, + 1792.0, + 296.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1790.0, + 1405.0, + 1790.0, + 1405.0, + 1823.0, + 293.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1856.0, + 295.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1882.0, + 1406.0, + 1882.0, + 1406.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1950.0, + 292.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1403.0, + 1942.0, + 1403.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1437.0, + 1404.0, + 1437.0, + 1404.0, + 1473.0, + 295.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1468.0, + 1404.0, + 1468.0, + 1404.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1500.0, + 1402.0, + 1500.0, + 1402.0, + 1534.0, + 296.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1528.0, + 1406.0, + 1528.0, + 1406.0, + 1566.0, + 292.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1595.0, + 293.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1591.0, + 1404.0, + 1591.0, + 1404.0, + 1624.0, + 293.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1620.0, + 1407.0, + 1620.0, + 1407.0, + 1656.0, + 292.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1653.0, + 553.0, + 1653.0, + 553.0, + 1683.0, + 296.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1266.0, + 1405.0, + 1266.0, + 1405.0, + 1306.0, + 294.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 1404.0, + 1298.0, + 1404.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1331.0, + 1404.0, + 1331.0, + 1404.0, + 1364.0, + 294.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1359.0, + 1405.0, + 1359.0, + 1405.0, + 1396.0, + 293.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1391.0, + 793.0, + 1391.0, + 793.0, + 1428.0, + 293.0, + 1428.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 828, + 1404, + 828, + 1404, + 1135, + 298, + 1135 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1149, + 1404, + 1149, + 1404, + 1395, + 298, + 1395 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1410, + 1402, + 1410, + 1402, + 1563, + 299, + 1563 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 296, + 549, + 1406, + 549, + 1406, + 671, + 296, + 671 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 357, + 270, + 1330, + 270, + 1330, + 523, + 357, + 523 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 301, + 1579, + 1399, + 1579, + 1399, + 1671, + 301, + 1671 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 299, + 750, + 1400, + 750, + 1400, + 813, + 299, + 813 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 300, + 1737, + 543, + 1737, + 543, + 1772, + 300, + 1772 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 550.0, + 1404.0, + 550.0, + 1404.0, + 582.0, + 295.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 579.0, + 1403.0, + 579.0, + 1403.0, + 611.0, + 294.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 610.0, + 1404.0, + 610.0, + 1404.0, + 645.0, + 294.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 641.0, + 990.0, + 641.0, + 990.0, + 674.0, + 294.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 288.0, + 381.0, + 288.0, + 381.0, + 297.0, + 370.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 280.0, + 618.0, + 280.0, + 618.0, + 303.0, + 552.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 280.0, + 961.0, + 280.0, + 961.0, + 303.0, + 895.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 285.0, + 1070.0, + 285.0, + 1070.0, + 297.0, + 1056.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 280.0, + 1305.0, + 280.0, + 1305.0, + 303.0, + 1239.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 298.0, + 624.0, + 298.0, + 624.0, + 313.0, + 552.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 298.0, + 969.0, + 298.0, + 969.0, + 313.0, + 895.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 298.0, + 1312.0, + 298.0, + 1312.0, + 313.0, + 1238.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 307.0, + 583.0, + 307.0, + 583.0, + 321.0, + 552.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 307.0, + 927.0, + 307.0, + 927.0, + 321.0, + 895.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 312.0, + 1069.0, + 312.0, + 1069.0, + 320.0, + 1058.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 307.0, + 1272.0, + 307.0, + 1272.0, + 323.0, + 1238.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 317.0, + 589.0, + 317.0, + 589.0, + 331.0, + 551.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 317.0, + 932.0, + 317.0, + 932.0, + 331.0, + 895.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 316.0, + 1275.0, + 316.0, + 1275.0, + 331.0, + 1238.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 329.0, + 386.0, + 329.0, + 386.0, + 401.0, + 355.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 326.0, + 601.0, + 326.0, + 601.0, + 360.0, + 550.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 336.0, + 726.0, + 336.0, + 726.0, + 348.0, + 712.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 326.0, + 945.0, + 326.0, + 945.0, + 360.0, + 894.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 336.0, + 1068.0, + 336.0, + 1068.0, + 346.0, + 1057.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 326.0, + 1288.0, + 326.0, + 1288.0, + 360.0, + 1238.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 353.0, + 638.0, + 353.0, + 638.0, + 370.0, + 552.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 356.0, + 980.0, + 356.0, + 980.0, + 370.0, + 896.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 356.0, + 1323.0, + 356.0, + 1323.0, + 370.0, + 1239.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 394.0, + 788.0, + 394.0, + 788.0, + 403.0, + 774.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 402.0, + 452.0, + 402.0, + 452.0, + 414.0, + 428.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 411.0, + 484.0, + 411.0, + 484.0, + 419.0, + 463.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 411.0, + 1072.0, + 411.0, + 1072.0, + 420.0, + 1057.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 416.0, + 382.0, + 416.0, + 382.0, + 425.0, + 374.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 425.0, + 497.0, + 425.0, + 497.0, + 437.0, + 482.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 417.0, + 729.0, + 417.0, + 729.0, + 430.0, + 715.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 421.0, + 962.0, + 421.0, + 962.0, + 432.0, + 935.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 435.0, + 1072.0, + 435.0, + 1072.0, + 443.0, + 1060.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 460.0, + 424.0, + 460.0, + 424.0, + 473.0, + 406.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 460.0, + 456.0, + 460.0, + 456.0, + 473.0, + 439.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 460.0, + 487.0, + 460.0, + 487.0, + 473.0, + 471.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 460.0, + 520.0, + 460.0, + 520.0, + 473.0, + 500.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 460.0, + 551.0, + 460.0, + 551.0, + 473.0, + 535.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 460.0, + 584.0, + 460.0, + 584.0, + 473.0, + 566.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 460.0, + 615.0, + 460.0, + 615.0, + 473.0, + 599.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 460.0, + 769.0, + 460.0, + 769.0, + 473.0, + 751.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 460.0, + 800.0, + 460.0, + 800.0, + 473.0, + 782.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 460.0, + 831.0, + 460.0, + 831.0, + 473.0, + 816.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 460.0, + 864.0, + 460.0, + 864.0, + 473.0, + 845.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 460.0, + 895.0, + 460.0, + 895.0, + 473.0, + 878.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 460.0, + 928.0, + 460.0, + 928.0, + 473.0, + 909.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 460.0, + 960.0, + 460.0, + 960.0, + 473.0, + 943.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 460.0, + 1112.0, + 460.0, + 1112.0, + 473.0, + 1095.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 460.0, + 1144.0, + 460.0, + 1144.0, + 473.0, + 1125.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 460.0, + 1175.0, + 460.0, + 1175.0, + 473.0, + 1160.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 460.0, + 1208.0, + 460.0, + 1208.0, + 473.0, + 1190.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 460.0, + 1239.0, + 460.0, + 1239.0, + 473.0, + 1222.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 460.0, + 1272.0, + 460.0, + 1272.0, + 473.0, + 1255.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 460.0, + 1304.0, + 460.0, + 1304.0, + 473.0, + 1286.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 467.0, + 526.0, + 467.0, + 526.0, + 481.0, + 495.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 467.0, + 870.0, + 467.0, + 870.0, + 481.0, + 839.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 467.0, + 1215.0, + 467.0, + 1215.0, + 481.0, + 1182.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 495.0, + 525.0, + 495.0, + 525.0, + 526.0, + 488.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 492.0, + 870.0, + 492.0, + 870.0, + 527.0, + 829.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 492.0, + 1215.0, + 492.0, + 1215.0, + 528.0, + 1174.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 397.0, + 414.0, + 397.0, + 414.0, + 414.0, + 392.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.25, + 407.0, + 767.25, + 407.0, + 767.25, + 417.0, + 754.25, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 410.5, + 789.0, + 410.5, + 789.0, + 418.5, + 775.0, + 418.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.75, + 434.5, + 526.75, + 434.5, + 526.75, + 450.5, + 477.75, + 450.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.75, + 425.5, + 816.75, + 425.5, + 816.75, + 436.0, + 802.75, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1733.0, + 549.0, + 1733.0, + 549.0, + 1781.0, + 292.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 828.0, + 1406.0, + 828.0, + 1406.0, + 865.0, + 293.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 859.0, + 1401.0, + 859.0, + 1401.0, + 891.0, + 296.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 886.0, + 1406.0, + 886.0, + 1406.0, + 928.0, + 291.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 915.0, + 1408.0, + 915.0, + 1408.0, + 959.0, + 291.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 951.0, + 1406.0, + 951.0, + 1406.0, + 985.0, + 292.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 980.0, + 1406.0, + 980.0, + 1406.0, + 1017.0, + 293.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1007.0, + 1405.0, + 1007.0, + 1405.0, + 1050.0, + 292.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1043.0, + 1406.0, + 1043.0, + 1406.0, + 1078.0, + 294.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1072.0, + 1405.0, + 1072.0, + 1405.0, + 1106.0, + 292.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1104.0, + 654.0, + 1104.0, + 654.0, + 1139.0, + 296.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1151.0, + 1405.0, + 1151.0, + 1405.0, + 1184.0, + 296.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1181.0, + 1402.0, + 1181.0, + 1402.0, + 1211.0, + 296.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1211.0, + 1406.0, + 1211.0, + 1406.0, + 1246.0, + 293.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1278.0, + 292.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1272.0, + 1404.0, + 1272.0, + 1404.0, + 1307.0, + 293.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1303.0, + 1404.0, + 1303.0, + 1404.0, + 1337.0, + 294.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1333.0, + 1404.0, + 1333.0, + 1404.0, + 1367.0, + 294.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1365.0, + 1002.0, + 1365.0, + 1002.0, + 1398.0, + 294.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1819.0, + 1407.0, + 1819.0, + 1407.0, + 1858.0, + 293.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1848.0, + 1407.0, + 1848.0, + 1407.0, + 1887.0, + 292.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 561.0, + 2001.0, + 561.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1407.0, + 1406.0, + 1407.0, + 1406.0, + 1446.0, + 293.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1440.0, + 1404.0, + 1440.0, + 1404.0, + 1473.0, + 297.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1471.0, + 1406.0, + 1471.0, + 1406.0, + 1506.0, + 294.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1501.0, + 1406.0, + 1501.0, + 1406.0, + 1538.0, + 295.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1534.0, + 701.0, + 1534.0, + 701.0, + 1567.0, + 295.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1615.0, + 293.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1609.0, + 1404.0, + 1609.0, + 1404.0, + 1643.0, + 295.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1638.0, + 582.0, + 1638.0, + 582.0, + 1672.0, + 296.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 748.0, + 1406.0, + 748.0, + 1406.0, + 788.0, + 293.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 783.0, + 612.0, + 783.0, + 612.0, + 815.0, + 294.0, + 815.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.824 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.818 + }, + { + "category_id": 1, + "poly": [ + 296, + 1689, + 1401, + 1689, + 1401, + 1752, + 296, + 1752 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 295, + 1773, + 1401, + 1773, + 1401, + 1836, + 295, + 1836 + ], + "score": 0.731 + }, + { + "category_id": 1, + "poly": [ + 295, + 736, + 1399, + 736, + 1399, + 800, + 295, + 800 + ], + "score": 0.704 + }, + { + "category_id": 1, + "poly": [ + 295, + 1605, + 1403, + 1605, + 1403, + 1670, + 295, + 1670 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 294, + 1520, + 1402, + 1520, + 1402, + 1587, + 294, + 1587 + ], + "score": 0.691 + }, + { + "category_id": 1, + "poly": [ + 296, + 1857, + 1401, + 1857, + 1401, + 1921, + 296, + 1921 + ], + "score": 0.684 + }, + { + "category_id": 1, + "poly": [ + 297, + 1101, + 1402, + 1101, + 1402, + 1167, + 297, + 1167 + ], + "score": 0.684 + }, + { + "category_id": 1, + "poly": [ + 293, + 819, + 1399, + 819, + 1399, + 885, + 293, + 885 + ], + "score": 0.681 + }, + { + "category_id": 1, + "poly": [ + 296, + 651, + 1399, + 651, + 1399, + 716, + 296, + 716 + ], + "score": 0.681 + }, + { + "category_id": 1, + "poly": [ + 300, + 568, + 1397, + 568, + 1397, + 632, + 300, + 632 + ], + "score": 0.676 + }, + { + "category_id": 1, + "poly": [ + 298, + 279, + 1405, + 279, + 1405, + 371, + 298, + 371 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 298, + 1017, + 1399, + 1017, + 1399, + 1082, + 298, + 1082 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 296, + 1268, + 1402, + 1268, + 1402, + 1333, + 296, + 1333 + ], + "score": 0.653 + }, + { + "category_id": 1, + "poly": [ + 297, + 903, + 1400, + 903, + 1400, + 998, + 297, + 998 + ], + "score": 0.643 + }, + { + "category_id": 1, + "poly": [ + 295, + 1436, + 1401, + 1436, + 1401, + 1503, + 295, + 1503 + ], + "score": 0.627 + }, + { + "category_id": 1, + "poly": [ + 296, + 1352, + 1401, + 1352, + 1401, + 1419, + 296, + 1419 + ], + "score": 0.627 + }, + { + "category_id": 1, + "poly": [ + 299, + 1185, + 1402, + 1185, + 1402, + 1251, + 299, + 1251 + ], + "score": 0.613 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1402, + 1942, + 1402, + 2033, + 299, + 2033 + ], + "score": 0.588 + }, + { + "category_id": 1, + "poly": [ + 296, + 393, + 1404, + 393, + 1404, + 547, + 296, + 547 + ], + "score": 0.584 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1686.0, + 1406.0, + 1686.0, + 1406.0, + 1727.0, + 294.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1721.0, + 610.0, + 1721.0, + 610.0, + 1751.0, + 322.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1811.0, + 294.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1803.0, + 533.0, + 1803.0, + 533.0, + 1836.0, + 324.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 735.0, + 1405.0, + 735.0, + 1405.0, + 774.0, + 294.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 767.0, + 926.0, + 767.0, + 926.0, + 801.0, + 324.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1604.0, + 1406.0, + 1604.0, + 1406.0, + 1643.0, + 295.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1635.0, + 1311.0, + 1635.0, + 1311.0, + 1670.0, + 322.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1521.0, + 1404.0, + 1521.0, + 1404.0, + 1557.0, + 296.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1552.0, + 1371.0, + 1552.0, + 1371.0, + 1587.0, + 322.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1856.0, + 1402.0, + 1856.0, + 1402.0, + 1892.0, + 296.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1886.0, + 1284.0, + 1886.0, + 1284.0, + 1922.0, + 321.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1096.0, + 1407.0, + 1096.0, + 1407.0, + 1143.0, + 292.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1133.0, + 1364.0, + 1133.0, + 1364.0, + 1169.0, + 324.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 818.0, + 1405.0, + 818.0, + 1405.0, + 857.0, + 293.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 851.0, + 1356.0, + 851.0, + 1356.0, + 887.0, + 325.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 649.0, + 1403.0, + 649.0, + 1403.0, + 689.0, + 295.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 682.0, + 464.0, + 682.0, + 464.0, + 716.0, + 325.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 568.0, + 1403.0, + 568.0, + 1403.0, + 604.0, + 296.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 599.0, + 1326.0, + 599.0, + 1326.0, + 635.0, + 324.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 277.0, + 1405.0, + 277.0, + 1405.0, + 314.0, + 295.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 309.0, + 1405.0, + 309.0, + 1405.0, + 345.0, + 322.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 340.0, + 611.0, + 340.0, + 611.0, + 372.0, + 322.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1019.0, + 1401.0, + 1019.0, + 1401.0, + 1055.0, + 296.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1050.0, + 1074.0, + 1050.0, + 1074.0, + 1083.0, + 323.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1262.0, + 1406.0, + 1262.0, + 1406.0, + 1311.0, + 290.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1300.0, + 506.0, + 1300.0, + 506.0, + 1334.0, + 323.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 902.0, + 1405.0, + 902.0, + 1405.0, + 941.0, + 293.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 936.0, + 1404.0, + 936.0, + 1404.0, + 970.0, + 323.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 964.0, + 904.0, + 964.0, + 904.0, + 1002.0, + 323.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1433.0, + 1406.0, + 1433.0, + 1406.0, + 1477.0, + 292.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1470.0, + 1036.0, + 1470.0, + 1036.0, + 1503.0, + 323.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1353.0, + 1405.0, + 1353.0, + 1405.0, + 1389.0, + 295.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1386.0, + 1342.0, + 1386.0, + 1342.0, + 1419.0, + 323.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1188.0, + 1404.0, + 1188.0, + 1404.0, + 1221.0, + 297.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1218.0, + 1288.0, + 1218.0, + 1288.0, + 1252.0, + 320.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1970.0, + 1406.0, + 1970.0, + 1406.0, + 2011.0, + 319.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2001.0, + 397.0, + 2001.0, + 397.0, + 2034.0, + 322.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 395.0, + 1404.0, + 395.0, + 1404.0, + 428.0, + 295.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 424.0, + 1404.0, + 424.0, + 1404.0, + 457.0, + 324.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 457.0, + 1401.0, + 457.0, + 1401.0, + 490.0, + 324.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 486.0, + 1402.0, + 486.0, + 1402.0, + 519.0, + 323.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 516.0, + 557.0, + 516.0, + 557.0, + 547.0, + 320.0, + 547.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 106, + 298, + 106 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.828 + }, + { + "category_id": 1, + "poly": [ + 290, + 1164, + 1404, + 1164, + 1404, + 1230, + 290, + 1230 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 291, + 1081, + 1402, + 1081, + 1402, + 1146, + 291, + 1146 + ], + "score": 0.601 + }, + { + "category_id": 1, + "poly": [ + 293, + 998, + 1400, + 998, + 1400, + 1064, + 293, + 1064 + ], + "score": 0.573 + }, + { + "category_id": 1, + "poly": [ + 293, + 945, + 1377, + 945, + 1377, + 982, + 293, + 982 + ], + "score": 0.524 + }, + { + "category_id": 1, + "poly": [ + 299, + 748, + 1403, + 748, + 1403, + 844, + 299, + 844 + ], + "score": 0.473 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1402, + 229, + 1402, + 293, + 296, + 293 + ], + "score": 0.467 + }, + { + "category_id": 1, + "poly": [ + 294, + 862, + 1402, + 862, + 1402, + 928, + 294, + 928 + ], + "score": 0.466 + }, + { + "category_id": 1, + "poly": [ + 298, + 447, + 1399, + 447, + 1399, + 513, + 298, + 513 + ], + "score": 0.453 + }, + { + "category_id": 1, + "poly": [ + 291, + 665, + 1400, + 665, + 1400, + 732, + 291, + 732 + ], + "score": 0.452 + }, + { + "category_id": 1, + "poly": [ + 299, + 311, + 1400, + 311, + 1400, + 377, + 299, + 377 + ], + "score": 0.426 + }, + { + "category_id": 1, + "poly": [ + 297, + 530, + 1401, + 530, + 1401, + 596, + 297, + 596 + ], + "score": 0.356 + }, + { + "category_id": 1, + "poly": [ + 296, + 613, + 1384, + 613, + 1384, + 650, + 296, + 650 + ], + "score": 0.34 + }, + { + "category_id": 1, + "poly": [ + 299, + 394, + 1181, + 394, + 1181, + 430, + 299, + 430 + ], + "score": 0.18 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 298.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1163.0, + 1406.0, + 1163.0, + 1406.0, + 1200.0, + 294.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1196.0, + 570.0, + 1196.0, + 570.0, + 1227.0, + 322.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1081.0, + 1405.0, + 1081.0, + 1405.0, + 1117.0, + 295.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1111.0, + 775.0, + 1111.0, + 775.0, + 1147.0, + 322.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 995.0, + 1405.0, + 995.0, + 1405.0, + 1039.0, + 293.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1030.0, + 1403.0, + 1030.0, + 1403.0, + 1066.0, + 323.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 944.0, + 1376.0, + 944.0, + 1376.0, + 983.0, + 295.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 747.0, + 1403.0, + 747.0, + 1403.0, + 786.0, + 293.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 780.0, + 1404.0, + 780.0, + 1404.0, + 814.0, + 324.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 813.0, + 690.0, + 813.0, + 690.0, + 844.0, + 323.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 1406.0, + 224.0, + 1406.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 596.0, + 261.0, + 596.0, + 292.0, + 323.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 864.0, + 1405.0, + 864.0, + 1405.0, + 900.0, + 296.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 894.0, + 1043.0, + 894.0, + 1043.0, + 929.0, + 321.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 448.0, + 1401.0, + 448.0, + 1401.0, + 485.0, + 294.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 479.0, + 1142.0, + 479.0, + 1142.0, + 515.0, + 323.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 666.0, + 1405.0, + 666.0, + 1405.0, + 702.0, + 295.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 697.0, + 1403.0, + 697.0, + 1403.0, + 732.0, + 322.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 310.0, + 1404.0, + 310.0, + 1404.0, + 350.0, + 294.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 344.0, + 952.0, + 344.0, + 952.0, + 378.0, + 322.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 529.0, + 1405.0, + 529.0, + 1405.0, + 567.0, + 293.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 562.0, + 1403.0, + 562.0, + 1403.0, + 598.0, + 324.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 610.0, + 1378.0, + 610.0, + 1378.0, + 651.0, + 293.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 393.0, + 1180.0, + 393.0, + 1180.0, + 432.0, + 294.0, + 432.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/ByGq7hRqKX/ByGq7hRqKX.md b/parse/train/ByGq7hRqKX/ByGq7hRqKX.md new file mode 100644 index 0000000000000000000000000000000000000000..63c3943fcf54f259909d8a44e3e12e24b1696107 --- /dev/null +++ b/parse/train/ByGq7hRqKX/ByGq7hRqKX.md @@ -0,0 +1,308 @@ +# CROSS-TASK KNOWLEDGE TRANSFER FOR VISUALLY-GROUNDED NAVIGATION + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Recent efforts on training visual navigation agents conditioned on language using deep reinforcement learning have been successful in learning policies for two different tasks: learning to follow navigational instructions and embodied question answering. In this paper, we aim to learn a multitask model capable of jointly learning both tasks, and transferring knowledge of words and their grounding in visual objects across tasks. The proposed model uses a novel Dual-Attention unit to disentangle the knowledge of words in the textual representations and visual objects in the visual representations, and align them with each other. This disentangled task-invariant alignment of representations facilitates grounding and knowledge transfer across both tasks. We show that the proposed model outperforms a range of baselines on both tasks in simulated 3D environments. We also show that this disentanglement of representations makes our model modular, interpretable, and allows for transfer to instructions containing new words by leveraging object detectors. + +# 1 INTRODUCTION + +Deep reinforcement learning has been shown to be capable of achieving super-human performance in playing games such as Atari 2600 (Mnih et al., 2013) and Go (Silver et al., 2016). Following the success of deep reinforcement learning in 3D Games such as Doom (Lample & Chaplot, 2017; Dosovitskiy & Koltun, 2017) and DeepmindLab (Mnih et al., 2016), there has been increased interest in using deep reinforcement learning for training embodied agents, which interact with a 3D environment by receiving first-person views of the environment and taking navigational actions. The simplest navigational agents learn a particular behaviour such as collecting or avoiding particular objects (Kempka et al., 2016; Jaderberg et al., 2016; Mirowski et al., 2016) or playing deathmatches (Lample & Chaplot, 2017; Dosovitskiy & Koltun, 2017). Subsequently, there have been efforts on training navigational agents whose behaviour is conditioned on a target specified using images (Zhu et al., 2017) or coordinates (Gupta et al., 2017a; Savva et al., 2017). More recently, there has been much interest in training agents conditioned on language as it offers several advantages over using images or coordinates. + +Firstly, the compositionality of language allows generalization to new tasks without additional learning. Prior work (Oh et al., 2017; Hermann et al., 2017; Chaplot et al., 2017) has trained navigational agents to follow instructions and shown zero-shot generalization to new instructions which contain unseen composition of words seen in the training instructions. Secondly, language is also a convenient means for humans to communicate with autonomous agents. Language not only allows instruction but also interaction. Gordon et al. (2018) and Das et al. (2017) train agents to answer questions by navigating in the environment to gather the required information. + +These multimodal tasks involve several challenges, such as perception from raw pixels, grounding of words in the instruction or question to visual objects and attributes, reasoning to perform relational tasks, fine-grained navigation in 3D environments with continuous state space, and learning to answer questions. Training a multi-task model can also facilitate knowledge transfer between the tasks and allow the model to generalize to scenarios which were not possible with single tasks. For example, if an agent learns to follow the instruction ‘Go to the red pillar’ and answer the question ‘What color is the torch?’, then it should also be able to follow the instruction ‘Go to the red torch’ and answer the question ‘What color is the pillar?’ without any additional training. + +In this paper, we aim to train a multi-task navigation model to follow instructions and answer questions. To test the generalization of multi-task models, we define cross-task knowledge transfer as an evaluation criteria, evaluating zero-shot learning on instructions and questions consisting of unseen composition of words in both tasks. In order to achieve cross-task knowledge transfer, words in the input space of both tasks need to be aligned with each other and with the answer space while they are being grounded to visual objects and attributes. Prior models fail to achieve this as they are designed for a single task. We propose a novel dual-attention model involving sequential Gated- and Spatial-Attention operations to perform explicit task-invariant alignment between the image representation channels and the words in the input and answer space. We create datasets and simulation scenarios for testing cross-task knowledge transfer in the Doom environment (Kempka et al., 2016) and show that the proposed model outperforms a range of baselines on both tasks. Additionally, we demonstrate that the modularity of our model allows easy addition of new objects and attributes to a trained model. We plan to open-source the implementation of our proposed model as well as the datasets and simulation environments. + +![](images/c73220668e2566d5356bf6cee3722954a8f4847bd33438358c45fbb39f48d227.jpg) +Figure 1: An example of first-person view in the 3D Doom environment with sample instructions and questions. The test set consists of unseen instructions and questions. The dataset evaluates a model for cross-task knowledge transfer between Semantic Goal Navigation (SGN) and Embodied Question Answering (EQA). + +# 2 RELATED WORK + +This paper is motivated by a series of works on learning to follow navigation instructions (Oh et al., 2017; Hermann et al., 2017; Chaplot et al., 2017; Wu et al., 2018; Yu et al., 2018a) and learning to answer questions by navigating around the environment (Das et al., 2017; Gordon et al., 2018). Among methods learning from instructions in 3D environments, Oh et al. (2017) introduced a hierarchical RL model for learning sequences of instructions by learning skills to solve subtasks. Chaplot et al. (2017) introduced a gated-attention model for multimodal fusion of textual and visual representations using multiplicative interactions, whereas Hermann et al. (2017) introduced auxiliary tasks such as temporal autoencoding and language prediction to improve sample efficiency for this task. Yu et al. (2018a) proposed guided feature transformation which involves transformation of visual representations using latent sentence embeddings computed from the language input. + +Among models for embodied question answering, Das et al. (2017) introduced a hierarchical model consisting of 4 modules, each for processing images, encoding questions, navigation, and questionanswering, each of which is pretrained with supervised or imitation learning, followed by fine-tuning of the navigation model using reinforcement learning. Gordon et al. (2018) introduced the task of Interactive Question Answering which involves interacting with objects in the environment with non-navigational actions for answering questions. They proposed Hierarchical Interactive Memory Network (HIMN), which allows temporal abstraction using a factorized set of controllers. + +All of the above methods are designed for a single task, following navigational instructions or answering questions, whereas we aim to train a single model for both tasks. Yu et al. (2018b) introduced a model for interactive language acquisition by training on both Visual Question Answering and following instructions in a 2D grid world environment. We aim to tackle multimodal multitask learning in challenging 3D environments. Partial observability results in the requirement of learning to navigate for answering the questions, turning visual question answering to embodied question answering. 3D environments also allow us to test on interesting and more challenging instructions based on relative size of the objects, in addition to colors and types. + +In addition to the above, there is a huge body of work on multimodal learning in static settings which do not involve navigation or reinforcement learning. Some relevant works which use attention mechanisms similar to the ones used in our proposed model include Perez et al. (2017); Fukui et al. (2016); Xu & Saenko (2016); Hudson & Manning (2018); Gupta et al. (2017b) for Visual Question Answering and Zhao et al. (2018) for grounding audio to vision. + +![](images/a49406c1969d1fd52bfa11b26df5a13669079ff90e34764405f0a9c3943df268.jpg) +Figure 2: Overview of our proposed architecture, described in detail in Section 4. + +# 3 PROBLEM FORMULATION + +Consider an autonomous agent interacting with an episodic environment as shown in Figure 1. In the beginning of each episode, the agent receives a textual input $T$ specifying the task that it needs to achieve. For example, $T$ could be an instruction describing the target object or a question querying about some visual detail of objects in the environment. At each time step $t$ , the agent observes a state $s _ { t } = ( I _ { t } , T )$ where $I _ { t }$ is the first-person (egocentric) view of the environment, and takes an action $a _ { t }$ , which could be a navigational action or an answer action. The agent’s objective is to learn a policy $\pi ( a _ { t } | s _ { t } )$ which leads to successful completion of the task specified by the textual input $T$ . + +Tasks. We focus on the multi-task learning of two visually-grounded language navigation tasks: In Embodied Question Answering $( E Q A )$ , the agent is given a question (“What color is the torch?”), and it must navigate around the 3D environment to explore the environment and gather information to answer the question (“red”). In Semantic Goal Navigation (SGN), the agent is given a language instruction ( $^ { 6 6 } \mathrm { G o }$ to the red torch”) to navigate to a goal location. + +Environments. We adapt the ViZDoom (Kempka et al., 2016)-based language grounding environment proposed by Chaplot et al. (2017) for visually-grounded multitask learning. It consists of a single room with 5 objects. The objects are randomized in each episode based on the textual input. We use two difficulty settings for the Doom domain: Easy: The agent is spawned at a fixed location. The candidate objects are spawned at five fixed locations along a single horizontal line in the field of view of the agent. Hard: The candidate objects and the agent are spawned at random locations and the objects may or may not be in the agents field of view in the initial configuration. The agent must explore the map to view all objects. + +Datasets. We use the set of instructions from Chaplot et al. (2017) and create a dataset for questions using the same set of objects and attributes. We define cross-task knowledge transfer as an evaluation criteria for testing generalization of multi-task models. We create train-test splits for both instructions and questions datasets to explicitly test a multitask model’s ability to transfer the knowledge of words across different tasks. Each instruction in the test set contains a word that is never seen in any instruction in the training set but is seen in some questions in the training set. Similarly, each question in the test set contains a word never seen in any training set question. Figure 1 illustrates the train-test split of instructions and questions used in our experiments in the Doom domain. Note that for the EQA trainset, unseen words can be present in the answer. + +The agent can take 4 actions: 3 navigational actions (forward, left, right) and 1 answer action. When the agent takes the answer action, the answer with the maximum probability in the output answer distribution is used. Other details such as the train-test splits are deferred to the Appendix. We also report results on an additional environment based on House3D (Wu et al., 2018) in the Appendix. + +# 4 PROPOSED METHOD + +In this section, we detail our proposed architecture (illustrated in Figure 2). At the start of each episode, the agent receives a textual input $T$ (an instruction or a question) specifying the task that it needs to achieve. At each time step, the agent observes an egocentric image $I _ { t }$ which is passed through a convolutional neural network (LeCun et al., 1995) with ReLU activations (Glorot et al., 2011) to produce the image representation $x _ { I } = f ( I _ { t } ; \theta _ { \mathrm { c o n v } } ) \in \mathbb { R } ^ { V \times H \times W }$ , where $\theta _ { \mathrm { c o n v } }$ denotes the parameters of the convolutional network, $V$ is the number of feature maps in the convolutional network output which is equal to the vocabulary size, and $H$ and $W$ are the height and width of each feature map. We use two representations for the textual input $T$ : (1) the bag-of-words representation denoted by $x _ { \mathrm { B o W } } ~ \in ~ \mathbb { R } ^ { V }$ and (2) a sentence representation $x _ { \mathrm { s e n t } } = f ( \breve { T } ; \theta _ { \mathrm { s e n t } } ) \in \mathbb { R } ^ { \breve { V } }$ , which is computed by passing the words in $T$ through a Gated Recurrent Unit (GRU) (Cho et al., 2014) network followed by a linear layer. Here, $\theta _ { \mathrm { s e n t } }$ denotes the parameters of the GRU network and the linear layer with ReLU activations. Next, the Dual-Attention unit $f _ { \mathrm { D A } }$ combines the image representation with the text representations to get the complete state representation $x _ { S }$ and answer prediction $x _ { \mathrm { A n s } }$ : + +$$ +x _ { \mathrm { S } } , x _ { \mathrm { A n s } } = f _ { \mathrm { D A } } ( x _ { I } , x _ { \mathrm { B o W } } , x _ { \mathrm { s e n t } } ) +$$ + +Finally, $x _ { S }$ and $x _ { \mathrm { A n s } }$ , along with a time step embedding and a task indicator variable (for whether the task is SGN or EQA), are passed to the policy module to produce an action. + +# 4.1 DUAL-ATTENTION UNIT + +The Dual-Attention unit uses two types of attention mechanisms, Gated-Attention $f _ { \mathrm { G A } }$ and SpatialAttention $f _ { \mathrm { S A } }$ , to align representations in different modalities and tasks. + +Gated-Attention. The Gated-Attention unit (Figure 3) was proposed in (Chaplot et al., 2017) for multimodal fusion. Intuitively, a GA unit attends to the different channels in the image representation based on the text representation. For example, if the textual input is the instruction ‘Go to the red pillar’, then the GA unit can learn to attend to channels which detect red things and pillars. More specifically, the GA unit takes as input a 3-dimensional tensor image representation $y _ { I } \in \mathring { \mathbb { R } } ^ { d \times H \times W }$ and a text representation $\bar { y _ { T } } \in \mathbb { R } ^ { d }$ , and outputs a 3-dimensional tensor ${ \boldsymbol { z } } \in \mathbb { R } ^ { d \times H \times W }$ . Note that the dimension of $y _ { T }$ is equal to the number of feature maps and the size of the first dimension of $y _ { I }$ . In the Gated-Attention unit, each element of $y _ { T }$ is expanded to a $H \times W$ matrix, resulting in a 3-dimensional tensor $M _ { y _ { T } } \in \mathbb { R } ^ { d \times H \times W }$ , whose $( i , j , k ) ^ { t \bar { h } }$ element is given by $M _ { y _ { T } } [ i , j , k ] = \bar { y } _ { T } [ i ]$ . This matrix is multiplied element-wise with the image representation: $z = f _ { \mathrm { G A } } ( y _ { I } , y _ { T } ) = M _ { y _ { T } } \odot y _ { I }$ , where $\odot$ denotes the Hadamard product (Horn, 1990). + +![](images/888f192624493aa788242aabe8918602f25c59af6c7afb47da6321d3f27acf6d.jpg) +Figure 3: Gated-Attention unit $f _ { \mathrm { G A } }$ + +Spatial-Attention. We propose a Spatial-Attention unit (Figure 4) which is analogous to the Gated-Attention unit except that it attends to different pixels in the image representation rather than the channels. For example, if the textual input is the question ‘Which object is blue in color?’, then we would like to spatially attend to the parts of the image which contain a blue object in order recognize the type of the blue object. The Spatial-Attention unit takes as input a 3-dimensional tensor image representation $y _ { I } \in \dot { \mathbb { R } ^ { d \times H \times W } }$ and a 2-dimensional spatial attention map $\mathit { y } _ { S } \in \mathbb { R } ^ { H \times W }$ , and outputs a tensor $\mathit { \check { z } } \in \mathbb { R } ^ { d \times H \times W }$ . Note that the height and width of the spatial attention map is equal to the height and width of the image representation. In the spatialattention unit, each element of the spatial attention map is expanded to a $d$ dimensional vector. This again results in a 3-dimensional tensor $M _ { y _ { S } } \in \mathbb { R } ^ { d \times \hat { H } \times W }$ , whose $( i , j , k ) ^ { t h }$ element is given by: $\bar { M } _ { y s } [ i , j , k ] = y _ { S } [ j , k ]$ . Just like in the Gated-Attention unit, this matrix is multiplied element-wise with the image representation: $z = f _ { \mathrm { S A } } ( y _ { I } , y _ { S } ) = M _ { y _ { S } } \odot y _ { I }$ . Similar spatial attention mechanisms have been used for Visual Question Answering (Fukui et al., 2016; Xu & Saenko, 2016; Hudson & Manning, 2018; Gupta et al., 2017b) and grounding audio in vision (Zhao et al., 2018). + +![](images/e7a785f4ad73bbe78d600384b1b5f88adcfc6098742f7d2bb686856c9e974c72.jpg) +Figure 4: Spatial-Attention unit $f _ { \mathrm { S A } }$ + +Dual-Attention. We now describe the operations in the Dual-Attention unit shown in Figure 5, as well as motivate the intuitions behind each operation. Given $x _ { I }$ , $x _ { \mathrm { B o W } }$ , and $x _ { \mathrm { s e n t } }$ , the Dual-Attention unit first computes a Gated-Attention over $x _ { I }$ using $x _ { \mathrm { B o W } }$ : + +$$ +x _ { \mathrm { G A l } } = f _ { \mathrm { G A } } ( x _ { I } , x _ { \mathrm { B o W } } ) \in \mathbb { R } ^ { V \times H \times W } . +$$ + +Intuitively, this first Gated-Attention unit associates each word in the vocabulary with a feature map in the image representation. A particular feature map is activated if and only if the corresponding word occurs in the textual input. In other words, the feature maps in the convolutional output learns to detect different objects and attributes, and words in the textual input specify which objects and attributes are relevant to the current task. The Gated-Attention using BoW representation attends to feature maps detecting corresponding objects and attributes, and masks all other feature maps. We use bag-of-words representation for the first GA unit as it explicitly aligns the words in textual input irrespective of whether it is a question or an instruction. Note that bag-of-words representation has been used previously in models trained for learning to follow instructions (Hermann et al., 2017). + +![](images/a7e5f5b2f4bacd0b9ce717a508e7d79925d868fa21a8c187b4d6a12dd64fc350.jpg) +Figure 5: Architecture of the Dual-Attention unit with example intermediate representations and operations. + +Next, the output of the Gated-Attention unit $x _ { \mathrm { G A 1 } }$ is converted to a spatial attention map by summing over all channels followed by a softmax over $H \times W$ elements: + +$$ +x _ { \mathrm { s p a t } } = \sigma \left( \sum _ { i } ^ { V } { x _ { \mathrm { G A l } } [ i , : , : ] } \right) \in \mathbb { R } ^ { H \times W } +$$ + +where the softmax $\sigma ( z ) _ { j } = \exp ( z _ { j } ) / \sum _ { j } \exp ( z _ { j } )$ ensures that the attention map is normalized. Summation of $x _ { \mathrm { G A l } }$ along the depth dimension gives a spatial attention map which has high activations at spatial locations where relevant objects or attributes are detected. ReLU activations in the convolutional feature maps makes all elements positive, ensuring that the summation aggregates the activations of relevant feature maps. + +$x _ { \mathrm { s p a t } }$ and $x _ { I }$ are then passed through a Spatial-Attention unit: + +$$ +x _ { \mathrm { S A } } = f _ { \mathrm { S A } } ( x _ { I } , x _ { \mathrm { s p a t } } ) \in \mathbb { R } ^ { V \times H \times W } +$$ + +The Spatial-Attention unit outputs all attributes present at the locations where relevant objects and attributes are detected. This is especially helpful for question answering, where a single GatedAttention may not be sufficient. For example, if the textual input is ‘Which color is the pillar?’, then the model needs to attend not only to feature maps detecting pillars (done by the Gated-Attention), but also to other attributes at the spatial locations where pillars are seen in order to predict their color. Note that a single Gated-Attention is sufficient for instruction following, as shown in (Chaplot et al., 2017). For example, if the textual input is ‘Go to the green pillar’, the first Gated-Attention unit can learn to attend to feature maps detecting green objects and pillar, and learn a navigation policy based on the spatial locations of the feature map activations. + +$x _ { \mathrm { S A } }$ is then passed through another Gated-Attention unit with the sentence-level text representation: + +$$ +x _ { \mathrm { G A 2 } } = f _ { \mathrm { G A } } ( x _ { \mathrm { S A } } , x _ { \mathrm { s e n t } } ) \in \mathbb { R } ^ { V \times H \times W } +$$ + +This second Gated-Attention unit enables the model to attend to different types of attributes based on the question. For instance, if the question is asking about the color (‘Which color is the pillar?’), then the model needs to attend to the feature maps corresponding to colors; or if the question is asking about the object type (‘Which object is green in color?’), then the model needs to attend to the feature maps corresponding to object types. The sentence embedding $x _ { \mathrm { s e n t } }$ can learn to attend to multiple channels based on the textual input and mask the rest. + +Next, the output is transformed to answer prediction by again doing a summation and softmax but this time summing over the height and width instead of the channels: + +$$ +x _ { \mathrm { A n s } } = \sigma \left( \sum _ { j , k } ^ { H , W } x _ { \mathrm { G A 2 } } [ : , j , k ] \right) \in \mathbb { R } ^ { V } +$$ + +Summation of $x _ { \mathrm { G A } 2 }$ along each feature map aggregates the activations for relevant attributes spatially. Again, ReLU activations for sentence embedding ensure aggregation of activations for each attribute or word. The answer space is identical to the textual input space $\mathbb { R } ^ { V }$ . + +Finally, the Dual-Attention unit $f _ { \mathrm { D A } }$ outputs the answer prediction $x _ { \mathrm { A n s } }$ and the flattened spatial attention map $x _ { \mathrm { S } } = \mathrm { v e c } ( x _ { \mathrm { s p a t } } )$ , where $\mathrm { v e c } \bar { ( . ) }$ denotes the flattening operation. + +Policy Module. The policy module takes as input the state representation $x _ { S }$ from the DualAttention unit, a time step embedding $t$ , and a task indicator variable $I$ (for whether the task is SGN or EQA). The inputs are concatenated then passed through a linear layer, then a recurrent GRU layer, then linear layers to estimate the policy function $\pi ( a _ { t } \mid I _ { t } , T )$ and the value function $V ( I _ { t } , T )$ . + +All above operations are differentiable, making the entire architecture trainable end-to-end. Note that all attention mechanisms in the Dual-Attention unit only modulate the input image representation, i.e., mask or amplify specific feature maps or pixels. This ensures that there is an explicit alignment between the words in the textual input, the feature maps in the image representation, and the words in answer space. This forces the convolutional network to encode all the information required with respect to a certain word in the corresponding output channel. This explicit taskinvariant alignment between convolutional feature maps and words in the input and answer space facilitates grounding and allows for cross-task knowledge transfer. As shown in the results later, this also makes our model modular and allows easy addition of objects and attributes to a trained model. + +# 4.2 OPTIMIZATION + +The entire model is trained to predict both navigational actions and answers jointly. The policy is trained using Proximal Policy Optimization (PPO) (Schulman et al., 2017). For training the answer predictions, we use a supervised cross-entropy loss. Both types of losses have common parameters as the answer prediction is essentially an intermediate representation for the policy. + +Auxiliary Task. As mentioned earlier, the feature maps in the convolutional output are expected to detect different objects and attributes. Consequently, we add a spatial auxiliary task to detect the object or attribute in the convolutional output channels corresponding to the word in the bag-of-words representation. A prior work (Gupta et al., 2017b) also explored the use of attribute and object recognition as an auxiliary task for Visual Question Answering. Rather than doing fine-grained object detection, we keep the size of the auxiliary predictions the same as the convolutional output to avoid increase in number of parameters, and maintain the explicit alignment on the convolutional feature maps with the words. Consequently, auxiliary labels are $( V \times H \times W )$ -dimensional tensors, where each of the $V$ channels correspond to a word in the vocabulary, and each element in a channel is 1 if the corresponding object or attribute is present in the current frame spatially. Figure 6 shows examples of auxiliary task labels for the channel corresponding to the word ‘red’. The auxiliary tasks are also trained with cross-entropy loss. + +![](images/fa7a1639c57b9e7bccb73053f54f5d1fe41e28c1028ec51acbe83fd8097150c0.jpg) +Figure 6: Example auxiliary task labels for the red channel. + +# 5 EXPERIMENTS & RESULTS + +Jointly learning semantic goal navigation and embodied question answering essentially involves a fusion of verbal and visual modalities. While prior methods are designed for a single task, we adapt several baselines for our environment and tasks by using their multimodal fusion techniques. We use two naive baselines, Image only and Text only; two baselines based on prior semantic goal navigation models, Concat (used by Hermann et al. (2017); Misra et al. (2017)) and Gated-Attention (GA) (Chaplot et al., 2017); and two baselines based on Question Answering models, FiLM (Perez et al., 2017) and PACMAN (Das et al., 2017). For fair comparison, we replace the proposed DualAttention unit with multimodal fusion techniques in the baselines and keep everything else identical to the proposed model. We provide more implementation details of all baselines in the Appendix. + +# 5.1 RESULTS + +We train all models for 10 million frames in the Easy setting and 50 million frames in the Hard setting. We use a $+ 1$ reward for reaching the correct object in SGN episodes and predicting the correct answer in EQA episodes. We use a small negative reward of -0.001 per time step to encourage shorter paths to target and answering questions as soon as possible. We also use distance-based reward shaping for SGN episodes, where the agent receives a small reward proportional to decrease in distance to the target. In the next subsection we evaluate the performance of the proposed model without the reward shaping. SGN episodes end when the agent reaches any object, and EQA episodes when the agent predicts any answer. All episodes have a maximum length of 210 time steps. We train all models with and without the auxiliary tasks using identical reward functions. + +![](images/0f763d40f3cf3cc304935d3ef380ea01fe08bd89fadf0f7ce22ecb962077a7f5.jpg) +Figure 7: Training accuracy of all models trained with auxiliary tasks for Easy (left) and Hard (right). + +Table 1: Accuracy of all models on SGN & EQA test sets for both Easy & Hard difficulties. + +
EasyHard
No AuxAuxNo AuxAux
ModelSGNEQASGNEQASGNEQASGNEQA
Text only0.20.330.20.330.20.330.20.33
Image only0.200.090.210.080.160.080.150.08
Concat0.330.210.310.190.20.260.390.22
GA0.270.180.350.240.180.110.220.24
FiLM0.240.110.340.120.120.030.250.15
PACMAN0.260.120.330.100.290.330.110.27
Dual-Attention0.860.530.960.580.860.380.820.59
+ +All models are trained jointly for both the tasks and tested on each task separately. In Figure 7, we show the training performance curves for all models trained with Auxiliary tasks in both Easy and Hard settings. In Table 1, we report the test performance of all models on both SGN and EQA for both Easy and Hard settings. During training, the Dual-Attention model learns faster as compared to the baselines in the Easy setting while achieving higher final performance in the Hard setting (see Figure 7). More interestingly, as shown in Table 1, the Dual-Attention model achieves considerably higher accuracy on the test set for SGN and EQA in both the difficulty settings when trained with or without auxiliary tasks. These results confirm the hypothesis that prior models, which are designed for a single task, lack the ability align the words in both the tasks and transfer knowledge across tasks. Lower accuracy on EQA for most models (see Table 1) indicates that EQA is more challenging than SGN as it involves alignment between not only input textual and visual representations but also with the answer space. As expected, using spatial auxiliary tasks lead to better performance for all models Visualization of the attention maps and intermediate representations in the model indicate that the textual and visual representations are aligned as expected (see Appendix for visualizations)1. + +# 5.2 ABLATION TESTS + +We perform a series of ablation tests in order to analyze the contribution of each component in the Dual-Attention unit: without Spatial-Attention (w/o SA), without the first Gated-Attention with $x _ { B o W }$ (w/o GA1), and without the second Gated-Attention with $x _ { \mathrm { s e n t } }$ (w/o GA2). We also try removing the task indicator variable (w/o Indicator Variable), removing reward shaping (w/o Reward Shaping), and training the proposed model on a single task, SGN or EQA (DA Single-Task). + +Figure 8 shows the training performance curves for the Dual-Attention model along with all ablation models in the Easy Setting. In Table 2, we report the test set performance of all ablation models. The results indicate that SA and GA1 contribute the most to the performance of the Dual-Attention model. GA2 is critical for performance on EQA but not SGN (see Table 2). This is expected as GA2 is designed to attend to different objects and attributes based on the question and is used mainly for answer prediction. It is not critical for SGN as the spatial attention map consists of locations of relevant objects, which is sufficient for navigating to the correct object. Reward shaping and indicator variable help with learning speed (see Figure 8), but have little effect on the final performance (see Table 2). Dual-Attention models trained only on single tasks work well on SGN especially with auxiliary tasks. This is because the auxiliary task for single task models includes object detection labels corresponding to the words in the test set. This highlights a key advantage of the proposed model. Due to its modular and interpretable design, the model can used for transferring the policy to new objects and attributes without fine-tuning as discussed in the following subsection. + +![](images/69a4073a5b891f27a0c6d0e7e4b8f3eb9a94e900eda6ec575b749ee8e403a1d8.jpg) +Figure 8: Training accuracy of proposed Dual-Attention model with all ablation models trained without (left) and with (right) auxiliary tasks for the Easy environment. + +Table 2: Accuracy of all the ablation models trained with and without Auxiliary tasks on SGN and EQA test sets for the Doom Easy environment. + +
ModelNo Aux SGN EQAAux SGN
w/o SAEQA
w/o GA10.20 0.16 0.250.20 0.15 0.16 0.38
w/o GA20.14 0.80 0.330.97 0.15
w/o Task Indicator0.79 0.470.96 0.56
w/o Reward Shaping0.82 0.490.93 0.51
DA Single-Task0.63 0.310.91 0.34
DA Multi-Task0.86 0.530.96 0.58
+ +Table 3: The performance of a trained policy appended with object detectors on instructions containing unseen words (‘red’ and ‘pillar’). + +
InstructionEasyHard
Go to the pillar1.000.71
Go to the red object0.990.89
Go to the tall/short pillar0.990.68
Go to the <known_color>pillar.1.000.79
Go to the red <known_object>1.000.93
Go to the largest/smallest red object0.950.69
Go to the tall/short red pillar0.990.88
Go to the red pillar0.990.82
+ +# 5.3 EXTENSION: TRANSFER TO NEW WORDS + +Consider a scenario of SGN where the agent is trained to follow instructions of certain objects and attributes. Suppose that the user wants the agent to follow instructions about a new object such as ‘pillar’ or a new attribute such the color ‘red’ which are never seen in any training instruction. Prior SGN models are shown to perform well to unseen combination of object-attribute pairs (Chaplot et al., 2017), but they do not generalize well to instructions containing a new word. The model retrained only on new instructions will lead to catastrophic forgetting of previous instructions. + +In contrast, our model can be used for transfer to new words by training an object detector for each new word and appending it to the image representation $x _ { I }$ . In order to test this, we train a single-task SGN model using the proposed architecture on the training set for instructions. We use auxiliary tasks but only for words in the vocabulary of the instructions training set. After training the policy, we would like the agent to follow instructions containing test words ‘red’ and ‘pillar’, which the agent has never seen or received any supervision about how this attribute or object looks visually. For transferring the policy, we assume access to two object detectors which would give object detections for ‘red’ and ‘pillar’ separately. We resize the object detections to the size of a feature map in the image representation $( H \times W )$ and append them as channels to the image representation. We also append the words ‘red’ and ‘pillar’ to the bag-of-words representations in the same order such that they are aligned with the appended feature maps. We randomly initialize the embeddings of the new words for computing the sentence embedding. The results in Table 3 show that this policy generalizes well to different types of instructions with unseen words. This suggests that a trained policy can be scaled to more objects provided the complexity of navigation remains consistent. + +# 6 CONCLUSION + +We proposed a Dual-Attention model for visually-grounded multitask learning which uses Gatedand Spatial-Attention to disentangle attributes in feature representations and align them with the answer space. We show that the proposed model is able to transfer the knowledge of words across tasks and outperforms the baselines on both Semantic Goal Navigation and Embodied Question Answering by a considerable margin. We showed that disentangled and interpretable representations make our model modular and allows for easy addition of new objects or attributes to a trained model. In future, the model can potentially be extended to transferring knowledge across different domains by using modular interpretable representations of objects which are domain-invariant. + +# REFERENCES + +Devendra Singh Chaplot, Kanthashree Mysore Sathyendra, Rama Kumar Pasumarthi, Dheeraj Rajagopal, and Ruslan Salakhutdinov. Gated-attention architectures for task-oriented language grounding. arXiv preprint arXiv:1706.07230, 2017. + +Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties ¨ of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. + +Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. arXiv preprint arXiv:1711.11543, 2017. + +Alexey Dosovitskiy and Vladlen Koltun. Learning to act by predicting the future. In ICLR, 2017. + +Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. + +Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp. 315–323, 2011. + +Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4089–4098, 2018. + +Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. arXiv preprint arXiv:1702.03920, 3, 2017a. + +Tanmay Gupta, Kevin J Shih, Saurabh Singh, Derek Hoiem, Kevin J Shih, Arun Mallya, Wei Di, Vignesh Jagadeesh, Robinson Piramuthu, K Shih, et al. Aligned image-word representations improve inductive transfer across vision-language tasks. In ICCV, pp. 4223–4232, 2017b. + +Karl Moritz Hermann, Felix Hill, Simon Green, Fumin Wang, Ryan Faulkner, Hubert Soyer, David Szepesvari, Wojtek Czarnecki, Max Jaderberg, Denis Teplyashin, et al. Grounded language learning in a simulated 3d world. arXiv preprint arXiv:1706.06551, 2017. + +Roger A Horn. The hadamard product. In Proc. Symp. Appl. Math, volume 40, pp. 87–169, 1990. + +Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. arXiv preprint arXiv:1803.03067, 2018. + +Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016. + +Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Viz- ´ doom: A doom-based ai research platform for visual reinforcement learning. arXiv preprint arXiv:1605.02097, 2016. + +Guillaume Lample and Devendra Singh Chaplot. Playing FPS games with deep reinforcement learning. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. + +Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. + +Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in complex environments. arXiv preprint arXiv:1611.03673, 2016. + +Dipendra K Misra, John Langford, and Yoav Artzi. Mapping instructions and visual observations to actions with reinforcement learning. arXiv preprint arXiv:1704.08795, 2017. + +Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. + +Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In ICML, 2016. + +Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning. arXiv preprint arXiv:1706.05064, 2017. + +Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017. + +Manolis Savva, Angel X. Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. MINOS: Multimodal indoor simulator for navigation in complex environments. arXiv:1712.03931, 2017. + +John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. + +David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. + +Yi Wu, Yuxin Wu, Georgia Gkioxari, and Yuandong Tian. Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209, 2018. + +Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question answering. In European Conference on Computer Vision, pp. 451–466. Springer, 2016. + +Haonan Yu, Xiaochen Lian, Haichao Zhang, and Wei Xu. Guided feature transformation (gft): A neural language grounding module for embodied agents. arXiv preprint arXiv:1805.08329, 2018a. + +Haonan Yu, Haichao Zhang, and Wei Xu. Interactive grounded language acquisition and generalization in a 2d world. arXiv preprint arXiv:1802.01433, 2018b. + +Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. arXiv preprint arXiv:1804.03160, 2018. + +Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In Robotics and Automation (ICRA), 2017 IEEE International Conference on, pp. 3357–3364. IEEE, 2017. + +# A VISUALIZATIONS + +![](images/eb82cf3d410040af24e674376db93bf154e95616c9f3df8c13dcc8542653e7ae.jpg) +Figure 9: Visualizations of convolutional output channels. We visualize the convolutional channels corresponding to 7 words (one in each row) for the same frame (shown in the rightmost column). The first column shows the auxiliary task labels for reference. The second column and third column show the output of the corresponding channel for the proposed Dual-Attention model trained without and with auxiliary tasks, respectively. As expected, the Aux model outputs are very close to the auxiliary task labels. The convolutional outputs of the No Aux model show that words and objects/properties in the images have been properly aligned even when the model is not trained with any auxiliary task labels. We do not provide any auxiliary label for words ‘smallest’ and ‘largest’ as they are not properties of an object and require relative comparison of objects. The visualizations in row 5 (corresponding to ‘smallest’) indicate that both models are able to compare the sizes of objects and detect the smallest object in the corresponding output channel even without any aux labels for the smallest object. + +![](images/c09df43f375517b1e7308882219ccaabee4d2f03fc7563c5a1c2705c93dc07be.jpg) +Figure 10: Spatial Attention and Answer Prediction Visualizations. An example EQA episode with the question “Which is the smallest blue object?”. The sentence embedding of the question is shown on the top $( x _ { s e n t } )$ . As expected, the embedding attends to object type words (’torch’, ’pillar’, ’skullkey’, etc.) as the question is asking about an object type (’Which object’). The rows show increasing time steps and columns show the input frame, the input frame overlaid with the spatial attention map, the predicted answer distribution, and the action at each time step. As the agent is turning, the spatial attention attends to small and blue objects. Time steps 1, 2: The model is attending to the yellow skullkey but the probability of the answer is not sufficiently high, likely because the skullkey is not blue. Time step 3: The model cannot see the skullkey anymore so it attends to the armor which is next smallest object. Consequently, the answer prediction also predicts armor, but the policy decides not to answer due to low probability. Time step 4: As the agent turns more, it observes and attends to the blue skullkey. The answer prediction has high probability for skullkey as it’s small and blue and the policy decides to answer the question. + +![](images/6acfbbc3bc1a92d1814e01f282a9c869e31804d2fe115c6a5a2a0783ad26bf0f.jpg) +Figure 11: Architecture of the policy module. + +# B ADDITIONAL EXPERIMENTAL DETAILS + +# B.1 HYPERPARAMETERS AND NETWORK DETAILS + +The input image is rescaled to size $3 \times 1 6 8 \times 3 0 0$ . The convolutional network for processing the image consisted of 3 convolutional layers: conv1 containing $3 2 ~ 8 \mathrm { x } 8$ filters with stride 4, conv2 containing $6 4 ~ 4 \mathrm { x } 4$ filters with stride 2, and conv3 containing $V$ 3x3 filters with stride 2. We use ReLU activations for conv1 and conv2 and sigmoid for conv3, as its output is used as auxiliary task predictions directly. We use word embeddings and GRU of size 32 followed by linear layer of size $V$ to get the sentence-level representation. The policy module uses hidden dimension 128 for the linear and GRU layers (see Figure 11). + +For reinforcement learning, we use Proximal Policy Optimization (PPO) with 8 actors and a time horizon of 128 steps. We use a single batch with 4 PPO epochs. The clipping parameter for PPO is set to 0.2. The discount factor $( \gamma )$ is 0.99. We used Adam optimizer with learning rate $2 . 5 \mathrm { e } { \cdot } 4$ for all experiments. + +# B.2 BASELINE DETAILS + +Image only: Naive baseline of just using the image representation: $x _ { \mathrm { S } } = \sec ( x _ { I } )$ where vec(.) denotes the flattening operation. + +Text only: Naive baseline of just using the textual representations: $x _ { \mathrm { S } } = [ x _ { \mathrm { B o W } } , x _ { \mathrm { s e n t } } ]$ + +Concat: The image and textual representations are concatenated: $x _ { \mathrm { S } } = [ \mathrm { v e c } ( x _ { I } ) , x _ { \mathrm { B o W } } , x _ { \mathrm { s e n t } } ]$ . Note that concatenation is the most common method of combining representations. Hermann et al. (2017) concatenate convolutional image and bag-of-words textual representations for SGN, whereas Misra et al. (2017) use concatenation with sentence-level textual representations. + +Gated-Attention: Adapted from Chaplot et al. (2017), who used Gated-Attention with sentencelevel textual representations for SGN: $x _ { \mathrm { S } } = f _ { \mathrm { G A } } ( x _ { I } , x _ { \mathrm { s e n t } } )$ . + +FiLM: Perez et al. (2017) introduced a general-purpose conditioning method called Feature-wise Linear Modulation (FiLM) for Visual Question Answering. Using FiLM, $x _ { \mathrm { S } } = \gamma ( x _ { \mathrm { s e n t } } ) \odot x _ { I } +$ $\beta ( x _ { \mathrm { s e n t } } )$ where $\gamma ( x _ { \mathrm { s e n t } } )$ and $\beta ( x _ { \mathrm { s e n t } } )$ are learnable projections of the sentence representation. + +PACMAN: Das et al. (2017) presented a hierarchical RL model for EQA. We adapt their method by using the attention mechanism in their QA module, which takes the last 5 frames and the text as input, and computes the similarity of the text with each frame using dot products between image and sentence-level text representations. These similarities are converted into attention weights using softmax, and the attention-weighted image features are concatenated with question embedding and passed through a softmax classifier to predict the answer distribution. For this particular baseline, we use the last 5 frames as input at each time step, unlike the proposed model and all other baselines which use a single frame as input. The attention-weighted image features are used as the state representation. The PACMAN model used a pretrained QA module, but we train this module jointly with the Navigation model for fair comparison with the proposed model. + +For each of the above method except PACMAN, we use a linear layer $f$ with ReLU activations followed by softmax $\sigma$ to get a $V$ -dimensional answer prediction from the state representations: $x _ { \mathrm { A n s } } = \sigma ( \bar { f } ( x _ { \mathrm { S } } ; \theta _ { L i n } ) )$ . $x _ { \mathrm { { S } } }$ and $x _ { A n s }$ are concatenated and passed to the policy module along with the time step and task indicator variable just as in the proposed model. + +![](images/a6315480bb9a073f24a9632c6856b7874a62864e259308fb0863e2161017b386.jpg) +Figure 12: Plot showing the training accuracy of all the models without auxiliary tasks for Doom Easy and Hard environments. + +![](images/634ffdf84813ca11c0e122e1a00e556fb733df50e54d4f6c9664f00480c451dd.jpg) +Figure 13: Plot showing the training accuracy of 3 models across 3 training runs with different seeds with and without auxiliary tasks for Doom Easy environment without any smoothing. + +# C DOOM EXPERIMENT DETAILS + +Additional Results. We show the training accuracy of all models without auxiliary tasks for Doom Easy and Hard in Figure 12. We also show the training accuracy of 3 models across 3 training runs with different seeds with and without auxiliary tasks for Doom Easy environment in Figure 13. + +Dataset. The Doom objects used in our experiments are illustrated in Figure 14. Instructions and questions used for training and evaluation are listed in Tables 4. + +![](images/69d7bb5a3ad8695b8a37ae82bea6b9ec1dbc83f181543eb44b38a763cd846842.jpg) +Figure 14: Objects of various colors and sizes used in the ViZDoom environment. + +Table 4: Instructions and questions for ViZDoom experiments. We used 5 object classes (torch, pillar, keycard, skullkey, armor), 4 colors (red, green, blue, yellow), 2 sizes (tall, short), and 2 superlative sizes (smallest, largest). + +
SGN Instruction Type42 Train Instructions: Not containing‘red'&‘pillar'28 Test Instructions: Containing‘red'or‘pillar'
Go to the (object).Go to the color>object.torch,keycard,skullkey,armor yellow, green, bluepillar red
Go to the size)object. Go to the {color) {object).tall, short blue torch, green torch, green armor,red torch,red skullkey,red pillar,
blue skullkey,blue keycard, yellow keycard, yellow skullkeygreen pillar,red keycard,red armor
Go to the {size) (object). Go to the color) (size)object.short torch,tall torch green tall, blue tall, blue short,tall pillar, short pillar red short, red tall
Go to the size){color)object.green short tall green, tall blue, short blue,short red, tall red
Go to the {color)(size)(object).short green green tall torch, green short torch,red short pillar,red short torch,
blue short torch,blue tall torchred tall pillar, green tall pillar, red tall torch, green short pillar
Go to the size) {color> {object).tall green torch,short green torch, short blue torch,tall blue torchshort red pillar, short red torch, tall red pillar, tall green pillar, tall red torch,short green pillar
Go to the superlative)object. Go to the {superlative> (color) object.largest, smallest smallest yellow, smallest blue, smallest green,largest blue,largest red,smallest red
EQAQuestion Typelargest green,largest yellow 21 Train Questions: Not containing‘blue'&‘torch'8 Test Questions: Containing‘blue'or‘torch'
Which {size)object is {color)in color?What color is the {object)? What color is the {size){object>? Which object is {color) in color?pillar, keycard, skullkey,armor short pillar, tall pillar red,yellow, greentorch short torch, tall torch blue
+ +# D HOUSE3D EXPERIMENTS + +In the House3D domain, we train on one house environment and randomize the colors of each object at the start of each episode. The agent’s spawn location is fixed. We create instructions and questions dataset for this house similar to the Doom domain. The House3D objects used in our experiments are illustrated in Figure 15. Instructions and questions used for training and evaluation are listed in Table 6. + +Each model is trained for 50 million frames jointly on both SGN and EQA, without the auxiliary tasks and using identical reward functions. Similar to Doom, we use a $+ 1$ reward for reaching the correct object in SGN episodes and predicting the correct answer in the EQA episodes. We use a small negative reward of - 0.001 per time step to encourage shorter paths to target and answering the questions as soon as possible. We also use distance based reward shaping for both SGN and EQA episodes, where the agent receives a small reward proportional to decrease in distance to the target. SGN episodes end when the agent reaches any object and EQA episodes when agent predicts any answer. All episodes have a maximum length of 420 time steps. + +Table 5: Accuracy of all the models on the SGN and EQA train and test sets for the House3D Domain. + +
SGNEQA
ModelTrainTestTrainTest
Text only0.630.330.220.23
Image only0.280.010.120.22
Concat0.650.130.310.13
GA0.980.200.920.03
FiLM0.990.370.920.24
PACMAN0.730.200.400.21
Dual-Attention0.990.470.890.29
+ +In Table 5, we report the train and test performance of all the models on both SGN and EQA. The results are similar as in Doom: the Dual-Attention model outperforms the baselines by a considerable margin. + +![](images/af178f0259047b6ec08600ddcade5e3f117345df9c3c00343e94708605dc643a.jpg) +Figure 15: Example first-person views of the House3D environment with sample objects of various colors. + +Table 6: Instructions and questions for House3D experiments. We used 6 object classes (refrigerator, office chair, fish tank, fireplace, bed, sofa) and 4 colors (red, green, blue, yellow). + +
SGNInstruction Type22 Train Instructions: Not containing‘red'&‘bed'11 Test Instructions: Containing‘red'or‘bed'
Go to the (object). Go to the {color)(object).refrigerator,office_chair,fish_tank,fireplace green refrigerator, green office_chair, green fish_tank,green fireplace, green sofa, blue refrigerator, blue office_chair, blue fish_tank,blue fireplace,blue sofa, yellow refrigerator,yellow office_chair, yellow fish_tank,yellow fireplace,yellow sofabed red bed, green bed,blue bed, yellow bed,red refrigerator, red office_chair,red fish_tank, red fireplace,red sofa
EQAGo to the {color) object. Question Typegreen,blue,yellow 7 Train Questions:red 2 Test Questions:
Not containing‘blue'&‘sofa'Containing‘blue'or‘sofa'
What color is the {object)? What object is color> in color?refrigerator,office_chair,fish_tank,fireplace,bed red, green,yellowsofa blue
\ No newline at end of file diff --git a/parse/train/ByGq7hRqKX/ByGq7hRqKX_content_list.json b/parse/train/ByGq7hRqKX/ByGq7hRqKX_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0943875952099ff4b2e303c46a5646cbd711606d --- /dev/null +++ b/parse/train/ByGq7hRqKX/ByGq7hRqKX_content_list.json @@ -0,0 +1,1719 @@ +[ + { + "type": "text", + "text": "CROSS-TASK KNOWLEDGE TRANSFER FOR VISUALLY-GROUNDED NAVIGATION ", + "text_level": 1, + "bbox": [ + 174, + 98, + 692, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 171, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 215, + 544, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent efforts on training visual navigation agents conditioned on language using deep reinforcement learning have been successful in learning policies for two different tasks: learning to follow navigational instructions and embodied question answering. In this paper, we aim to learn a multitask model capable of jointly learning both tasks, and transferring knowledge of words and their grounding in visual objects across tasks. The proposed model uses a novel Dual-Attention unit to disentangle the knowledge of words in the textual representations and visual objects in the visual representations, and align them with each other. This disentangled task-invariant alignment of representations facilitates grounding and knowledge transfer across both tasks. We show that the proposed model outperforms a range of baselines on both tasks in simulated 3D environments. We also show that this disentanglement of representations makes our model modular, interpretable, and allows for transfer to instructions containing new words by leveraging object detectors. ", + "bbox": [ + 233, + 244, + 764, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 460, + 336, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep reinforcement learning has been shown to be capable of achieving super-human performance in playing games such as Atari 2600 (Mnih et al., 2013) and Go (Silver et al., 2016). Following the success of deep reinforcement learning in 3D Games such as Doom (Lample & Chaplot, 2017; Dosovitskiy & Koltun, 2017) and DeepmindLab (Mnih et al., 2016), there has been increased interest in using deep reinforcement learning for training embodied agents, which interact with a 3D environment by receiving first-person views of the environment and taking navigational actions. The simplest navigational agents learn a particular behaviour such as collecting or avoiding particular objects (Kempka et al., 2016; Jaderberg et al., 2016; Mirowski et al., 2016) or playing deathmatches (Lample & Chaplot, 2017; Dosovitskiy & Koltun, 2017). Subsequently, there have been efforts on training navigational agents whose behaviour is conditioned on a target specified using images (Zhu et al., 2017) or coordinates (Gupta et al., 2017a; Savva et al., 2017). More recently, there has been much interest in training agents conditioned on language as it offers several advantages over using images or coordinates. ", + "bbox": [ + 174, + 486, + 825, + 665 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Firstly, the compositionality of language allows generalization to new tasks without additional learning. Prior work (Oh et al., 2017; Hermann et al., 2017; Chaplot et al., 2017) has trained navigational agents to follow instructions and shown zero-shot generalization to new instructions which contain unseen composition of words seen in the training instructions. Secondly, language is also a convenient means for humans to communicate with autonomous agents. Language not only allows instruction but also interaction. Gordon et al. (2018) and Das et al. (2017) train agents to answer questions by navigating in the environment to gather the required information. ", + "bbox": [ + 174, + 672, + 823, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These multimodal tasks involve several challenges, such as perception from raw pixels, grounding of words in the instruction or question to visual objects and attributes, reasoning to perform relational tasks, fine-grained navigation in 3D environments with continuous state space, and learning to answer questions. Training a multi-task model can also facilitate knowledge transfer between the tasks and allow the model to generalize to scenarios which were not possible with single tasks. For example, if an agent learns to follow the instruction ‘Go to the red pillar’ and answer the question ‘What color is the torch?’, then it should also be able to follow the instruction ‘Go to the red torch’ and answer the question ‘What color is the pillar?’ without any additional training. ", + "bbox": [ + 174, + 777, + 825, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to train a multi-task navigation model to follow instructions and answer questions. To test the generalization of multi-task models, we define cross-task knowledge transfer as an evaluation criteria, evaluating zero-shot learning on instructions and questions consisting of unseen composition of words in both tasks. In order to achieve cross-task knowledge transfer, words in the input space of both tasks need to be aligned with each other and with the answer space while they are being grounded to visual objects and attributes. Prior models fail to achieve this as they are designed for a single task. We propose a novel dual-attention model involving sequential Gated- and Spatial-Attention operations to perform explicit task-invariant alignment between the image representation channels and the words in the input and answer space. We create datasets and simulation scenarios for testing cross-task knowledge transfer in the Doom environment (Kempka et al., 2016) and show that the proposed model outperforms a range of baselines on both tasks. Additionally, we demonstrate that the modularity of our model allows easy addition of new objects and attributes to a trained model. We plan to open-source the implementation of our proposed model as well as the datasets and simulation environments. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c73220668e2566d5356bf6cee3722954a8f4847bd33438358c45fbb39f48d227.jpg", + "image_caption": [ + "Figure 1: An example of first-person view in the 3D Doom environment with sample instructions and questions. The test set consists of unseen instructions and questions. The dataset evaluates a model for cross-task knowledge transfer between Semantic Goal Navigation (SGN) and Embodied Question Answering (EQA). " + ], + "image_footnote": [], + "bbox": [ + 174, + 101, + 797, + 188 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 265, + 825, + 433 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 455, + 344, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper is motivated by a series of works on learning to follow navigation instructions (Oh et al., 2017; Hermann et al., 2017; Chaplot et al., 2017; Wu et al., 2018; Yu et al., 2018a) and learning to answer questions by navigating around the environment (Das et al., 2017; Gordon et al., 2018). Among methods learning from instructions in 3D environments, Oh et al. (2017) introduced a hierarchical RL model for learning sequences of instructions by learning skills to solve subtasks. Chaplot et al. (2017) introduced a gated-attention model for multimodal fusion of textual and visual representations using multiplicative interactions, whereas Hermann et al. (2017) introduced auxiliary tasks such as temporal autoencoding and language prediction to improve sample efficiency for this task. Yu et al. (2018a) proposed guided feature transformation which involves transformation of visual representations using latent sentence embeddings computed from the language input. ", + "bbox": [ + 174, + 488, + 825, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among models for embodied question answering, Das et al. (2017) introduced a hierarchical model consisting of 4 modules, each for processing images, encoding questions, navigation, and questionanswering, each of which is pretrained with supervised or imitation learning, followed by fine-tuning of the navigation model using reinforcement learning. Gordon et al. (2018) introduced the task of Interactive Question Answering which involves interacting with objects in the environment with non-navigational actions for answering questions. They proposed Hierarchical Interactive Memory Network (HIMN), which allows temporal abstraction using a factorized set of controllers. ", + "bbox": [ + 174, + 635, + 825, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All of the above methods are designed for a single task, following navigational instructions or answering questions, whereas we aim to train a single model for both tasks. Yu et al. (2018b) introduced a model for interactive language acquisition by training on both Visual Question Answering and following instructions in a 2D grid world environment. We aim to tackle multimodal multitask learning in challenging 3D environments. Partial observability results in the requirement of learning to navigate for answering the questions, turning visual question answering to embodied question answering. 3D environments also allow us to test on interesting and more challenging instructions based on relative size of the objects, in addition to colors and types. ", + "bbox": [ + 174, + 738, + 823, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to the above, there is a huge body of work on multimodal learning in static settings which do not involve navigation or reinforcement learning. Some relevant works which use attention mechanisms similar to the ones used in our proposed model include Perez et al. (2017); Fukui et al. (2016); Xu & Saenko (2016); Hudson & Manning (2018); Gupta et al. (2017b) for Visual Question Answering and Zhao et al. (2018) for grounding audio to vision. ", + "bbox": [ + 176, + 857, + 823, + 928 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a49406c1969d1fd52bfa11b26df5a13669079ff90e34764405f0a9c3943df268.jpg", + "image_caption": [ + "Figure 2: Overview of our proposed architecture, described in detail in Section 4. " + ], + "image_footnote": [], + "bbox": [ + 202, + 104, + 785, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 PROBLEM FORMULATION ", + "text_level": 1, + "bbox": [ + 176, + 241, + 418, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider an autonomous agent interacting with an episodic environment as shown in Figure 1. In the beginning of each episode, the agent receives a textual input $T$ specifying the task that it needs to achieve. For example, $T$ could be an instruction describing the target object or a question querying about some visual detail of objects in the environment. At each time step $t$ , the agent observes a state $s _ { t } = ( I _ { t } , T )$ where $I _ { t }$ is the first-person (egocentric) view of the environment, and takes an action $a _ { t }$ , which could be a navigational action or an answer action. The agent’s objective is to learn a policy $\\pi ( a _ { t } | s _ { t } )$ which leads to successful completion of the task specified by the textual input $T$ . ", + "bbox": [ + 173, + 268, + 825, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tasks. We focus on the multi-task learning of two visually-grounded language navigation tasks: In Embodied Question Answering $( E Q A )$ , the agent is given a question (“What color is the torch?”), and it must navigate around the 3D environment to explore the environment and gather information to answer the question (“red”). In Semantic Goal Navigation (SGN), the agent is given a language instruction ( $^ { 6 6 } \\mathrm { G o }$ to the red torch”) to navigate to a goal location. ", + "bbox": [ + 174, + 372, + 825, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Environments. We adapt the ViZDoom (Kempka et al., 2016)-based language grounding environment proposed by Chaplot et al. (2017) for visually-grounded multitask learning. It consists of a single room with 5 objects. The objects are randomized in each episode based on the textual input. We use two difficulty settings for the Doom domain: Easy: The agent is spawned at a fixed location. The candidate objects are spawned at five fixed locations along a single horizontal line in the field of view of the agent. Hard: The candidate objects and the agent are spawned at random locations and the objects may or may not be in the agents field of view in the initial configuration. The agent must explore the map to view all objects. ", + "bbox": [ + 174, + 449, + 825, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets. We use the set of instructions from Chaplot et al. (2017) and create a dataset for questions using the same set of objects and attributes. We define cross-task knowledge transfer as an evaluation criteria for testing generalization of multi-task models. We create train-test splits for both instructions and questions datasets to explicitly test a multitask model’s ability to transfer the knowledge of words across different tasks. Each instruction in the test set contains a word that is never seen in any instruction in the training set but is seen in some questions in the training set. Similarly, each question in the test set contains a word never seen in any training set question. Figure 1 illustrates the train-test split of instructions and questions used in our experiments in the Doom domain. Note that for the EQA trainset, unseen words can be present in the answer. ", + "bbox": [ + 173, + 566, + 825, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The agent can take 4 actions: 3 navigational actions (forward, left, right) and 1 answer action. When the agent takes the answer action, the answer with the maximum probability in the output answer distribution is used. Other details such as the train-test splits are deferred to the Appendix. We also report results on an additional environment based on House3D (Wu et al., 2018) in the Appendix. ", + "bbox": [ + 174, + 699, + 825, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 PROPOSED METHOD ", + "text_level": 1, + "bbox": [ + 176, + 771, + 377, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we detail our proposed architecture (illustrated in Figure 2). At the start of each episode, the agent receives a textual input $T$ (an instruction or a question) specifying the task that it needs to achieve. At each time step, the agent observes an egocentric image $I _ { t }$ which is passed through a convolutional neural network (LeCun et al., 1995) with ReLU activations (Glorot et al., 2011) to produce the image representation $x _ { I } = f ( I _ { t } ; \\theta _ { \\mathrm { c o n v } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }$ , where $\\theta _ { \\mathrm { c o n v } }$ denotes the parameters of the convolutional network, $V$ is the number of feature maps in the convolutional network output which is equal to the vocabulary size, and $H$ and $W$ are the height and width of each feature map. We use two representations for the textual input $T$ : (1) the bag-of-words representation denoted by $x _ { \\mathrm { B o W } } ~ \\in ~ \\mathbb { R } ^ { V }$ and (2) a sentence representation $x _ { \\mathrm { s e n t } } = f ( \\breve { T } ; \\theta _ { \\mathrm { s e n t } } ) \\in \\mathbb { R } ^ { \\breve { V } }$ , which is computed by passing the words in $T$ through a Gated Recurrent Unit (GRU) (Cho et al., 2014) network followed by a linear layer. Here, $\\theta _ { \\mathrm { s e n t } }$ denotes the parameters of the GRU network and the linear layer with ReLU activations. Next, the Dual-Attention unit $f _ { \\mathrm { D A } }$ combines the image representation with the text representations to get the complete state representation $x _ { S }$ and answer prediction $x _ { \\mathrm { A n s } }$ : ", + "bbox": [ + 173, + 797, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 825, + 172 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/9a477078ce9fc48266be90f6b6df84b63c02c11e6ff0d0820fb6ebcaaa177a0a.jpg", + "text": "$$\nx _ { \\mathrm { S } } , x _ { \\mathrm { A n s } } = f _ { \\mathrm { D A } } ( x _ { I } , x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } )\n$$", + "text_format": "latex", + "bbox": [ + 393, + 171, + 602, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, $x _ { S }$ and $x _ { \\mathrm { A n s } }$ , along with a time step embedding and a task indicator variable (for whether the task is SGN or EQA), are passed to the policy module to produce an action. ", + "bbox": [ + 171, + 190, + 823, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 DUAL-ATTENTION UNIT ", + "text_level": 1, + "bbox": [ + 174, + 236, + 385, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Dual-Attention unit uses two types of attention mechanisms, Gated-Attention $f _ { \\mathrm { G A } }$ and SpatialAttention $f _ { \\mathrm { S A } }$ , to align representations in different modalities and tasks. ", + "bbox": [ + 174, + 262, + 823, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gated-Attention. The Gated-Attention unit (Figure 3) was proposed in (Chaplot et al., 2017) for multimodal fusion. Intuitively, a GA unit attends to the different channels in the image representation based on the text representation. For example, if the textual input is the instruction ‘Go to the red pillar’, then the GA unit can learn to attend to channels which detect red things and pillars. More specifically, the GA unit takes as input a 3-dimensional tensor image representation $y _ { I } \\in \\mathring { \\mathbb { R } } ^ { d \\times H \\times W }$ and a text representation $\\bar { y _ { T } } \\in \\mathbb { R } ^ { d }$ , and outputs a 3-dimensional tensor ${ \\boldsymbol { z } } \\in \\mathbb { R } ^ { d \\times H \\times W }$ . Note that the dimension of $y _ { T }$ is equal to the number of feature maps and the size of the first dimension of $y _ { I }$ . In the Gated-Attention unit, each element of $y _ { T }$ is expanded to a $H \\times W$ matrix, resulting in a 3-dimensional tensor $M _ { y _ { T } } \\in \\mathbb { R } ^ { d \\times H \\times W }$ , whose $( i , j , k ) ^ { t \\bar { h } }$ element is given by $M _ { y _ { T } } [ i , j , k ] = \\bar { y } _ { T } [ i ]$ . This matrix is multiplied element-wise with the image representation: $z = f _ { \\mathrm { G A } } ( y _ { I } , y _ { T } ) = M _ { y _ { T } } \\odot y _ { I }$ , where $\\odot$ denotes the Hadamard product (Horn, 1990). ", + "bbox": [ + 174, + 296, + 549, + 435 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/888f192624493aa788242aabe8918602f25c59af6c7afb47da6321d3f27acf6d.jpg", + "image_caption": [ + "Figure 3: Gated-Attention unit $f _ { \\mathrm { G A } }$ " + ], + "image_footnote": [], + "bbox": [ + 558, + 314, + 808, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 435, + 825, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial-Attention. We propose a Spatial-Attention unit (Figure 4) which is analogous to the Gated-Attention unit except that it attends to different pixels in the image representation rather than the channels. For example, if the textual input is the question ‘Which object is blue in color?’, then we would like to spatially attend to the parts of the image which contain a blue object in order recognize the type of the blue object. The Spatial-Attention unit takes as input a 3-dimensional tensor image representation $y _ { I } \\in \\dot { \\mathbb { R } ^ { d \\times H \\times W } }$ and a 2-dimensional spatial attention map $\\mathit { y } _ { S } \\in \\mathbb { R } ^ { H \\times W }$ , and outputs a tensor $\\mathit { \\check { z } } \\in \\mathbb { R } ^ { d \\times H \\times W }$ . Note that the height and width of the spatial attention map is equal to the height and width of the image representation. In the spatialattention unit, each element of the spatial attention map is expanded to a $d$ dimensional vector. This again results in a 3-dimensional tensor $M _ { y _ { S } } \\in \\mathbb { R } ^ { d \\times \\hat { H } \\times W }$ , whose $( i , j , k ) ^ { t h }$ element is given by: $\\bar { M } _ { y s } [ i , j , k ] = y _ { S } [ j , k ]$ . Just like in the Gated-Attention unit, this matrix is multiplied element-wise with the image representation: $z = f _ { \\mathrm { S A } } ( y _ { I } , y _ { S } ) = M _ { y _ { S } } \\odot y _ { I }$ . Similar spatial attention mechanisms have been used for Visual Question Answering (Fukui et al., 2016; Xu & Saenko, 2016; Hudson & Manning, 2018; Gupta et al., 2017b) and grounding audio in vision (Zhao et al., 2018). ", + "bbox": [ + 174, + 512, + 549, + 650 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e7a785f4ad73bbe78d600384b1b5f88adcfc6098742f7d2bb686856c9e974c72.jpg", + "image_caption": [ + "Figure 4: Spatial-Attention unit $f _ { \\mathrm { S A } }$ " + ], + "image_footnote": [], + "bbox": [ + 557, + 529, + 808, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 650, + 825, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dual-Attention. We now describe the operations in the Dual-Attention unit shown in Figure 5, as well as motivate the intuitions behind each operation. Given $x _ { I }$ , $x _ { \\mathrm { B o W } }$ , and $x _ { \\mathrm { s e n t } }$ , the Dual-Attention unit first computes a Gated-Attention over $x _ { I }$ using $x _ { \\mathrm { B o W } }$ : ", + "bbox": [ + 174, + 768, + 825, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/53dea56cfe2785c0aacf9210d9b50cfff6325aeac3db0508672dd32f789acce9.jpg", + "text": "$$\nx _ { \\mathrm { G A l } } = f _ { \\mathrm { G A } } ( x _ { I } , x _ { \\mathrm { B o W } } ) \\in \\mathbb { R } ^ { V \\times H \\times W } .\n$$", + "text_format": "latex", + "bbox": [ + 375, + 815, + 622, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, this first Gated-Attention unit associates each word in the vocabulary with a feature map in the image representation. A particular feature map is activated if and only if the corresponding word occurs in the textual input. In other words, the feature maps in the convolutional output learns to detect different objects and attributes, and words in the textual input specify which objects and attributes are relevant to the current task. The Gated-Attention using BoW representation attends to feature maps detecting corresponding objects and attributes, and masks all other feature maps. We use bag-of-words representation for the first GA unit as it explicitly aligns the words in textual input irrespective of whether it is a question or an instruction. Note that bag-of-words representation has been used previously in models trained for learning to follow instructions (Hermann et al., 2017). ", + "bbox": [ + 173, + 840, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a7e5f5b2f4bacd0b9ce717a508e7d79925d868fa21a8c187b4d6a12dd64fc350.jpg", + "image_caption": [ + "Figure 5: Architecture of the Dual-Attention unit with example intermediate representations and operations. " + ], + "image_footnote": [], + "bbox": [ + 173, + 98, + 813, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 327, + 821, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, the output of the Gated-Attention unit $x _ { \\mathrm { G A 1 } }$ is converted to a spatial attention map by summing over all channels followed by a softmax over $H \\times W$ elements: ", + "bbox": [ + 171, + 377, + 823, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/5b5dfb779e213cccb5af97bd1b97bd2d98e2f7b15763d52e68a49d10cf9a4c26.jpg", + "text": "$$\nx _ { \\mathrm { s p a t } } = \\sigma \\left( \\sum _ { i } ^ { V } { x _ { \\mathrm { G A l } } [ i , : , : ] } \\right) \\in \\mathbb { R } ^ { H \\times W }\n$$", + "text_format": "latex", + "bbox": [ + 369, + 410, + 627, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the softmax $\\sigma ( z ) _ { j } = \\exp ( z _ { j } ) / \\sum _ { j } \\exp ( z _ { j } )$ ensures that the attention map is normalized. Summation of $x _ { \\mathrm { G A l } }$ along the depth dimension gives a spatial attention map which has high activations at spatial locations where relevant objects or attributes are detected. ReLU activations in the convolutional feature maps makes all elements positive, ensuring that the summation aggregates the activations of relevant feature maps. ", + "bbox": [ + 174, + 458, + 825, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$x _ { \\mathrm { s p a t } }$ and $x _ { I }$ are then passed through a Spatial-Attention unit: ", + "bbox": [ + 173, + 536, + 573, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0a1ce057a4d7916019cef3002f085eefb3d7328dab947a5cf18fd8d11f5597d4.jpg", + "text": "$$\nx _ { \\mathrm { S A } } = f _ { \\mathrm { S A } } ( x _ { I } , x _ { \\mathrm { s p a t } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }\n$$", + "text_format": "latex", + "bbox": [ + 383, + 556, + 612, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Spatial-Attention unit outputs all attributes present at the locations where relevant objects and attributes are detected. This is especially helpful for question answering, where a single GatedAttention may not be sufficient. For example, if the textual input is ‘Which color is the pillar?’, then the model needs to attend not only to feature maps detecting pillars (done by the Gated-Attention), but also to other attributes at the spatial locations where pillars are seen in order to predict their color. Note that a single Gated-Attention is sufficient for instruction following, as shown in (Chaplot et al., 2017). For example, if the textual input is ‘Go to the green pillar’, the first Gated-Attention unit can learn to attend to feature maps detecting green objects and pillar, and learn a navigation policy based on the spatial locations of the feature map activations. ", + "bbox": [ + 173, + 580, + 825, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$x _ { \\mathrm { S A } }$ is then passed through another Gated-Attention unit with the sentence-level text representation: ", + "bbox": [ + 171, + 712, + 823, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/c2f422d394e5aa39fd0b01a9687570166e773a77be9f521371c2444b6252180e.jpg", + "text": "$$\nx _ { \\mathrm { G A 2 } } = f _ { \\mathrm { G A } } ( x _ { \\mathrm { S A } } , x _ { \\mathrm { s e n t } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }\n$$", + "text_format": "latex", + "bbox": [ + 375, + 731, + 620, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This second Gated-Attention unit enables the model to attend to different types of attributes based on the question. For instance, if the question is asking about the color (‘Which color is the pillar?’), then the model needs to attend to the feature maps corresponding to colors; or if the question is asking about the object type (‘Which object is green in color?’), then the model needs to attend to the feature maps corresponding to object types. The sentence embedding $x _ { \\mathrm { s e n t } }$ can learn to attend to multiple channels based on the textual input and mask the rest. ", + "bbox": [ + 173, + 755, + 825, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, the output is transformed to answer prediction by again doing a summation and softmax but this time summing over the height and width instead of the channels: ", + "bbox": [ + 173, + 844, + 823, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0daedaacc108632c304ac35dd2039c7625699fb7bf18eea83f8178c53dfe6988.jpg", + "text": "$$\nx _ { \\mathrm { A n s } } = \\sigma \\left( \\sum _ { j , k } ^ { H , W } x _ { \\mathrm { G A 2 } } [ : , j , k ] \\right) \\in \\mathbb { R } ^ { V }\n$$", + "text_format": "latex", + "bbox": [ + 375, + 878, + 622, + 929 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Summation of $x _ { \\mathrm { G A } 2 }$ along each feature map aggregates the activations for relevant attributes spatially. Again, ReLU activations for sentence embedding ensure aggregation of activations for each attribute or word. The answer space is identical to the textual input space $\\mathbb { R } ^ { V }$ . ", + "bbox": [ + 174, + 103, + 821, + 146 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the Dual-Attention unit $f _ { \\mathrm { D A } }$ outputs the answer prediction $x _ { \\mathrm { A n s } }$ and the flattened spatial attention map $x _ { \\mathrm { S } } = \\mathrm { v e c } ( x _ { \\mathrm { s p a t } } )$ , where $\\mathrm { v e c } \\bar { ( . ) }$ denotes the flattening operation. ", + "bbox": [ + 171, + 152, + 821, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Policy Module. The policy module takes as input the state representation $x _ { S }$ from the DualAttention unit, a time step embedding $t$ , and a task indicator variable $I$ (for whether the task is SGN or EQA). The inputs are concatenated then passed through a linear layer, then a recurrent GRU layer, then linear layers to estimate the policy function $\\pi ( a _ { t } \\mid I _ { t } , T )$ and the value function $V ( I _ { t } , T )$ . ", + "bbox": [ + 174, + 188, + 825, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All above operations are differentiable, making the entire architecture trainable end-to-end. Note that all attention mechanisms in the Dual-Attention unit only modulate the input image representation, i.e., mask or amplify specific feature maps or pixels. This ensures that there is an explicit alignment between the words in the textual input, the feature maps in the image representation, and the words in answer space. This forces the convolutional network to encode all the information required with respect to a certain word in the corresponding output channel. This explicit taskinvariant alignment between convolutional feature maps and words in the input and answer space facilitates grounding and allows for cross-task knowledge transfer. As shown in the results later, this also makes our model modular and allows easy addition of objects and attributes to a trained model. ", + "bbox": [ + 174, + 251, + 825, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 OPTIMIZATION ", + "text_level": 1, + "bbox": [ + 174, + 392, + 318, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The entire model is trained to predict both navigational actions and answers jointly. The policy is trained using Proximal Policy Optimization (PPO) (Schulman et al., 2017). For training the answer predictions, we use a supervised cross-entropy loss. Both types of losses have common parameters as the answer prediction is essentially an intermediate representation for the policy. ", + "bbox": [ + 174, + 412, + 823, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Auxiliary Task. As mentioned earlier, the feature maps in the convolutional output are expected to detect different objects and attributes. Consequently, we add a spatial auxiliary task to detect the object or attribute in the convolutional output channels corresponding to the word in the bag-of-words representation. A prior work (Gupta et al., 2017b) also explored the use of attribute and object recognition as an auxiliary task for Visual Question Answering. Rather than doing fine-grained object detection, we keep the size of the auxiliary predictions the same as the convolutional output to avoid increase in number of parameters, and maintain the explicit alignment on the convolutional feature maps with the words. Consequently, auxiliary labels are $( V \\times H \\times W )$ -dimensional tensors, where each of the $V$ channels correspond to a word in the vocabulary, and each element in a channel is 1 if the corresponding object or attribute is present in the current frame spatially. Figure 6 shows examples of auxiliary task labels for the channel corresponding to the word ‘red’. The auxiliary tasks are also trained with cross-entropy loss. ", + "bbox": [ + 174, + 476, + 614, + 601 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa7a1639c57b9e7bccb73053f54f5d1fe41e28c1028ec51acbe83fd8097150c0.jpg", + "image_caption": [ + "Figure 6: Example auxiliary task labels for the red channel. " + ], + "image_footnote": [], + "bbox": [ + 627, + 479, + 823, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 602, + 825, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 EXPERIMENTS & RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 703, + 428, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Jointly learning semantic goal navigation and embodied question answering essentially involves a fusion of verbal and visual modalities. While prior methods are designed for a single task, we adapt several baselines for our environment and tasks by using their multimodal fusion techniques. We use two naive baselines, Image only and Text only; two baselines based on prior semantic goal navigation models, Concat (used by Hermann et al. (2017); Misra et al. (2017)) and Gated-Attention (GA) (Chaplot et al., 2017); and two baselines based on Question Answering models, FiLM (Perez et al., 2017) and PACMAN (Das et al., 2017). For fair comparison, we replace the proposed DualAttention unit with multimodal fusion techniques in the baselines and keep everything else identical to the proposed model. We provide more implementation details of all baselines in the Appendix. ", + "bbox": [ + 174, + 729, + 825, + 856 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 875, + 277, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train all models for 10 million frames in the Easy setting and 50 million frames in the Hard setting. We use a $+ 1$ reward for reaching the correct object in SGN episodes and predicting the correct answer in EQA episodes. We use a small negative reward of -0.001 per time step to encourage shorter paths to target and answering questions as soon as possible. We also use distance-based reward shaping for SGN episodes, where the agent receives a small reward proportional to decrease in distance to the target. In the next subsection we evaluate the performance of the proposed model without the reward shaping. SGN episodes end when the agent reaches any object, and EQA episodes when the agent predicts any answer. All episodes have a maximum length of 210 time steps. We train all models with and without the auxiliary tasks using identical reward functions. ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f763d40f3cf3cc304935d3ef380ea01fe08bd89fadf0f7ce22ecb962077a7f5.jpg", + "image_caption": [ + "Figure 7: Training accuracy of all models trained with auxiliary tasks for Easy (left) and Hard (right). " + ], + "image_footnote": [], + "bbox": [ + 236, + 98, + 759, + 227 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0780412306b172e3fc51fe31cf166b5d535d27096b5b77bdb329922387195d78.jpg", + "table_caption": [ + "Table 1: Accuracy of all models on SGN & EQA test sets for both Easy & Hard difficulties. " + ], + "table_footnote": [], + "table_body": "
EasyHard
No AuxAuxNo AuxAux
ModelSGNEQASGNEQASGNEQASGNEQA
Text only0.20.330.20.330.20.330.20.33
Image only0.200.090.210.080.160.080.150.08
Concat0.330.210.310.190.20.260.390.22
GA0.270.180.350.240.180.110.220.24
FiLM0.240.110.340.120.120.030.250.15
PACMAN0.260.120.330.100.290.330.110.27
Dual-Attention0.860.530.960.580.860.380.820.59
", + "bbox": [ + 176, + 275, + 823, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 443, + 825, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All models are trained jointly for both the tasks and tested on each task separately. In Figure 7, we show the training performance curves for all models trained with Auxiliary tasks in both Easy and Hard settings. In Table 1, we report the test performance of all models on both SGN and EQA for both Easy and Hard settings. During training, the Dual-Attention model learns faster as compared to the baselines in the Easy setting while achieving higher final performance in the Hard setting (see Figure 7). More interestingly, as shown in Table 1, the Dual-Attention model achieves considerably higher accuracy on the test set for SGN and EQA in both the difficulty settings when trained with or without auxiliary tasks. These results confirm the hypothesis that prior models, which are designed for a single task, lack the ability align the words in both the tasks and transfer knowledge across tasks. Lower accuracy on EQA for most models (see Table 1) indicates that EQA is more challenging than SGN as it involves alignment between not only input textual and visual representations but also with the answer space. As expected, using spatial auxiliary tasks lead to better performance for all models Visualization of the attention maps and intermediate representations in the model indicate that the textual and visual representations are aligned as expected (see Appendix for visualizations)1. ", + "bbox": [ + 173, + 547, + 825, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 ABLATION TESTS ", + "text_level": 1, + "bbox": [ + 176, + 757, + 334, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform a series of ablation tests in order to analyze the contribution of each component in the Dual-Attention unit: without Spatial-Attention (w/o SA), without the first Gated-Attention with $x _ { B o W }$ (w/o GA1), and without the second Gated-Attention with $x _ { \\mathrm { s e n t } }$ (w/o GA2). We also try removing the task indicator variable (w/o Indicator Variable), removing reward shaping (w/o Reward Shaping), and training the proposed model on a single task, SGN or EQA (DA Single-Task). ", + "bbox": [ + 174, + 777, + 825, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 8 shows the training performance curves for the Dual-Attention model along with all ablation models in the Easy Setting. In Table 2, we report the test set performance of all ablation models. The results indicate that SA and GA1 contribute the most to the performance of the Dual-Attention model. GA2 is critical for performance on EQA but not SGN (see Table 2). This is expected as GA2 is designed to attend to different objects and attributes based on the question and is used mainly for answer prediction. It is not critical for SGN as the spatial attention map consists of locations of relevant objects, which is sufficient for navigating to the correct object. Reward shaping and indicator variable help with learning speed (see Figure 8), but have little effect on the final performance (see Table 2). Dual-Attention models trained only on single tasks work well on SGN especially with auxiliary tasks. This is because the auxiliary task for single task models includes object detection labels corresponding to the words in the test set. This highlights a key advantage of the proposed model. Due to its modular and interpretable design, the model can used for transferring the policy to new objects and attributes without fine-tuning as discussed in the following subsection. ", + "bbox": [ + 176, + 853, + 821, + 896 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69a4073a5b891f27a0c6d0e7e4b8f3eb9a94e900eda6ec575b749ee8e403a1d8.jpg", + "image_caption": [ + "Figure 8: Training accuracy of proposed Dual-Attention model with all ablation models trained without (left) and with (right) auxiliary tasks for the Easy environment. " + ], + "image_footnote": [], + "bbox": [ + 238, + 99, + 758, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b49aeed69fcac50cc68bdbb554ea70bb037720c598878d8f4b9ac597c18ecfdf.jpg", + "table_caption": [ + "Table 2: Accuracy of all the ablation models trained with and without Auxiliary tasks on SGN and EQA test sets for the Doom Easy environment. " + ], + "table_footnote": [], + "table_body": "
ModelNo Aux SGN EQAAux SGN
w/o SAEQA
w/o GA10.20 0.16 0.250.20 0.15 0.16 0.38
w/o GA20.14 0.80 0.330.97 0.15
w/o Task Indicator0.79 0.470.96 0.56
w/o Reward Shaping0.82 0.490.93 0.51
DA Single-Task0.63 0.310.91 0.34
DA Multi-Task0.86 0.530.96 0.58
", + "bbox": [ + 173, + 314, + 483, + 443 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5c95bb849656c81c03ec6f0cf06a185b6a09bde23e4d722cc774e16cf0ff6d2f.jpg", + "table_caption": [ + "Table 3: The performance of a trained policy appended with object detectors on instructions containing unseen words (‘red’ and ‘pillar’). " + ], + "table_footnote": [], + "table_body": "
InstructionEasyHard
Go to the pillar1.000.71
Go to the red object0.990.89
Go to the tall/short pillar0.990.68
Go to the <known_color>pillar.1.000.79
Go to the red <known_object>1.000.93
Go to the largest/smallest red object0.950.69
Go to the tall/short red pillar0.990.88
Go to the red pillar0.990.82
", + "bbox": [ + 516, + 314, + 823, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 473, + 825, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 EXTENSION: TRANSFER TO NEW WORDS ", + "text_level": 1, + "bbox": [ + 176, + 631, + 496, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consider a scenario of SGN where the agent is trained to follow instructions of certain objects and attributes. Suppose that the user wants the agent to follow instructions about a new object such as ‘pillar’ or a new attribute such the color ‘red’ which are never seen in any training instruction. Prior SGN models are shown to perform well to unseen combination of object-attribute pairs (Chaplot et al., 2017), but they do not generalize well to instructions containing a new word. The model retrained only on new instructions will lead to catastrophic forgetting of previous instructions. ", + "bbox": [ + 174, + 652, + 825, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, our model can be used for transfer to new words by training an object detector for each new word and appending it to the image representation $x _ { I }$ . In order to test this, we train a single-task SGN model using the proposed architecture on the training set for instructions. We use auxiliary tasks but only for words in the vocabulary of the instructions training set. After training the policy, we would like the agent to follow instructions containing test words ‘red’ and ‘pillar’, which the agent has never seen or received any supervision about how this attribute or object looks visually. For transferring the policy, we assume access to two object detectors which would give object detections for ‘red’ and ‘pillar’ separately. We resize the object detections to the size of a feature map in the image representation $( H \\times W )$ and append them as channels to the image representation. We also append the words ‘red’ and ‘pillar’ to the bag-of-words representations in the same order such that they are aligned with the appended feature maps. We randomly initialize the embeddings of the new words for computing the sentence embedding. The results in Table 3 show that this policy generalizes well to different types of instructions with unseen words. This suggests that a trained policy can be scaled to more objects provided the complexity of navigation remains consistent. ", + "bbox": [ + 174, + 743, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 150, + 318, + 166 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We proposed a Dual-Attention model for visually-grounded multitask learning which uses Gatedand Spatial-Attention to disentangle attributes in feature representations and align them with the answer space. We show that the proposed model is able to transfer the knowledge of words across tasks and outperforms the baselines on both Semantic Goal Navigation and Embodied Question Answering by a considerable margin. We showed that disentangled and interpretable representations make our model modular and allows for easy addition of new objects or attributes to a trained model. In future, the model can potentially be extended to transferring knowledge across different domains by using modular interpretable representations of objects which are domain-invariant. ", + "bbox": [ + 174, + 176, + 825, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 309, + 285, + 324 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Devendra Singh Chaplot, Kanthashree Mysore Sathyendra, Rama Kumar Pasumarthi, Dheeraj Rajagopal, and Ruslan Salakhutdinov. Gated-attention architectures for task-oriented language grounding. arXiv preprint arXiv:1706.07230, 2017. ", + "bbox": [ + 174, + 333, + 825, + 376 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties ¨ of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. ", + "bbox": [ + 174, + 387, + 825, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. arXiv preprint arXiv:1711.11543, 2017. ", + "bbox": [ + 171, + 440, + 823, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alexey Dosovitskiy and Vladlen Koltun. Learning to act by predicting the future. In ICLR, 2017. ", + "bbox": [ + 173, + 481, + 810, + 496 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. ", + "bbox": [ + 176, + 506, + 821, + 549 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp. 315–323, 2011. ", + "bbox": [ + 173, + 559, + 825, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4089–4098, 2018. ", + "bbox": [ + 178, + 613, + 823, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. arXiv preprint arXiv:1702.03920, 3, 2017a. ", + "bbox": [ + 171, + 667, + 823, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tanmay Gupta, Kevin J Shih, Saurabh Singh, Derek Hoiem, Kevin J Shih, Arun Mallya, Wei Di, Vignesh Jagadeesh, Robinson Piramuthu, K Shih, et al. Aligned image-word representations improve inductive transfer across vision-language tasks. In ICCV, pp. 4223–4232, 2017b. ", + "bbox": [ + 174, + 707, + 825, + 751 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Karl Moritz Hermann, Felix Hill, Simon Green, Fumin Wang, Ryan Faulkner, Hubert Soyer, David Szepesvari, Wojtek Czarnecki, Max Jaderberg, Denis Teplyashin, et al. Grounded language learning in a simulated 3d world. arXiv preprint arXiv:1706.06551, 2017. ", + "bbox": [ + 174, + 761, + 825, + 804 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Roger A Horn. The hadamard product. In Proc. Symp. Appl. Math, volume 40, pp. 87–169, 1990. ", + "bbox": [ + 171, + 815, + 812, + 830 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. arXiv preprint arXiv:1803.03067, 2018. ", + "bbox": [ + 171, + 842, + 823, + 869 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016. ", + "bbox": [ + 176, + 882, + 823, + 922 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Viz- ´ doom: A doom-based ai research platform for visual reinforcement learning. arXiv preprint arXiv:1605.02097, 2016. ", + "bbox": [ + 174, + 103, + 821, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Guillaume Lample and Devendra Singh Chaplot. Playing FPS games with deep reinforcement learning. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. ", + "bbox": [ + 173, + 155, + 823, + 184 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. ", + "bbox": [ + 173, + 193, + 823, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in complex environments. arXiv preprint arXiv:1611.03673, 2016. ", + "bbox": [ + 176, + 229, + 821, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dipendra K Misra, John Langford, and Yoav Artzi. Mapping instructions and visual observations to actions with reinforcement learning. arXiv preprint arXiv:1704.08795, 2017. ", + "bbox": [ + 173, + 281, + 821, + 310 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. ", + "bbox": [ + 173, + 319, + 823, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In ICML, 2016. ", + "bbox": [ + 174, + 371, + 825, + 414 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning. arXiv preprint arXiv:1706.05064, 2017. ", + "bbox": [ + 171, + 422, + 821, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017. ", + "bbox": [ + 171, + 460, + 825, + 489 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Manolis Savva, Angel X. Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. MINOS: Multimodal indoor simulator for navigation in complex environments. arXiv:1712.03931, 2017. ", + "bbox": [ + 173, + 497, + 825, + 540 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. ", + "bbox": [ + 173, + 549, + 823, + 579 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. ", + "bbox": [ + 174, + 587, + 823, + 630 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yi Wu, Yuxin Wu, Georgia Gkioxari, and Yuandong Tian. Building generalizable agents with a realistic and rich 3d environment. arXiv preprint arXiv:1801.02209, 2018. ", + "bbox": [ + 169, + 638, + 823, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question answering. In European Conference on Computer Vision, pp. 451–466. Springer, 2016. ", + "bbox": [ + 174, + 676, + 823, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Haonan Yu, Xiaochen Lian, Haichao Zhang, and Wei Xu. Guided feature transformation (gft): A neural language grounding module for embodied agents. arXiv preprint arXiv:1805.08329, 2018a. ", + "bbox": [ + 178, + 728, + 823, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Haonan Yu, Haichao Zhang, and Wei Xu. Interactive grounded language acquisition and generalization in a 2d world. arXiv preprint arXiv:1802.01433, 2018b. ", + "bbox": [ + 171, + 780, + 821, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. arXiv preprint arXiv:1804.03160, 2018. ", + "bbox": [ + 173, + 818, + 823, + 847 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In Robotics and Automation (ICRA), 2017 IEEE International Conference on, pp. 3357–3364. IEEE, 2017. ", + "bbox": [ + 176, + 856, + 825, + 911 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A VISUALIZATIONS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 354, + 117 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/eb82cf3d410040af24e674376db93bf154e95616c9f3df8c13dcc8542653e7ae.jpg", + "image_caption": [ + "Figure 9: Visualizations of convolutional output channels. We visualize the convolutional channels corresponding to 7 words (one in each row) for the same frame (shown in the rightmost column). The first column shows the auxiliary task labels for reference. The second column and third column show the output of the corresponding channel for the proposed Dual-Attention model trained without and with auxiliary tasks, respectively. As expected, the Aux model outputs are very close to the auxiliary task labels. The convolutional outputs of the No Aux model show that words and objects/properties in the images have been properly aligned even when the model is not trained with any auxiliary task labels. We do not provide any auxiliary label for words ‘smallest’ and ‘largest’ as they are not properties of an object and require relative comparison of objects. The visualizations in row 5 (corresponding to ‘smallest’) indicate that both models are able to compare the sizes of objects and detect the smallest object in the corresponding output channel even without any aux labels for the smallest object. " + ], + "image_footnote": [], + "bbox": [ + 187, + 143, + 816, + 654 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c09df43f375517b1e7308882219ccaabee4d2f03fc7563c5a1c2705c93dc07be.jpg", + "image_caption": [ + "Figure 10: Spatial Attention and Answer Prediction Visualizations. An example EQA episode with the question “Which is the smallest blue object?”. The sentence embedding of the question is shown on the top $( x _ { s e n t } )$ . As expected, the embedding attends to object type words (’torch’, ’pillar’, ’skullkey’, etc.) as the question is asking about an object type (’Which object’). The rows show increasing time steps and columns show the input frame, the input frame overlaid with the spatial attention map, the predicted answer distribution, and the action at each time step. As the agent is turning, the spatial attention attends to small and blue objects. Time steps 1, 2: The model is attending to the yellow skullkey but the probability of the answer is not sufficiently high, likely because the skullkey is not blue. Time step 3: The model cannot see the skullkey anymore so it attends to the armor which is next smallest object. Consequently, the answer prediction also predicts armor, but the policy decides not to answer due to low probability. Time step 4: As the agent turns more, it observes and attends to the blue skullkey. The answer prediction has high probability for skullkey as it’s small and blue and the policy decides to answer the question. " + ], + "image_footnote": [], + "bbox": [ + 176, + 286, + 812, + 578 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6acfbbc3bc1a92d1814e01f282a9c869e31804d2fe115c6a5a2a0783ad26bf0f.jpg", + "image_caption": [ + "Figure 11: Architecture of the policy module. " + ], + "image_footnote": [], + "bbox": [ + 210, + 99, + 789, + 186 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B ADDITIONAL EXPERIMENTAL DETAILS ", + "text_level": 1, + "bbox": [ + 174, + 243, + 534, + 260 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.1 HYPERPARAMETERS AND NETWORK DETAILS ", + "text_level": 1, + "bbox": [ + 178, + 279, + 529, + 292 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The input image is rescaled to size $3 \\times 1 6 8 \\times 3 0 0$ . The convolutional network for processing the image consisted of 3 convolutional layers: conv1 containing $3 2 ~ 8 \\mathrm { x } 8$ filters with stride 4, conv2 containing $6 4 ~ 4 \\mathrm { x } 4$ filters with stride 2, and conv3 containing $V$ 3x3 filters with stride 2. We use ReLU activations for conv1 and conv2 and sigmoid for conv3, as its output is used as auxiliary task predictions directly. We use word embeddings and GRU of size 32 followed by linear layer of size $V$ to get the sentence-level representation. The policy module uses hidden dimension 128 for the linear and GRU layers (see Figure 11). ", + "bbox": [ + 174, + 306, + 825, + 405 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For reinforcement learning, we use Proximal Policy Optimization (PPO) with 8 actors and a time horizon of 128 steps. We use a single batch with 4 PPO epochs. The clipping parameter for PPO is set to 0.2. The discount factor $( \\gamma )$ is 0.99. We used Adam optimizer with learning rate $2 . 5 \\mathrm { e } { \\cdot } 4$ for all experiments. ", + "bbox": [ + 174, + 411, + 825, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.2 BASELINE DETAILS", + "text_level": 1, + "bbox": [ + 176, + 489, + 351, + 503 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Image only: Naive baseline of just using the image representation: $x _ { \\mathrm { S } } = \\sec ( x _ { I } )$ where vec(.) denotes the flattening operation. ", + "bbox": [ + 174, + 518, + 823, + 546 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Text only: Naive baseline of just using the textual representations: $x _ { \\mathrm { S } } = [ x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } ]$ ", + "bbox": [ + 176, + 553, + 736, + 568 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Concat: The image and textual representations are concatenated: $x _ { \\mathrm { S } } = [ \\mathrm { v e c } ( x _ { I } ) , x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } ]$ . Note that concatenation is the most common method of combining representations. Hermann et al. (2017) concatenate convolutional image and bag-of-words textual representations for SGN, whereas Misra et al. (2017) use concatenation with sentence-level textual representations. ", + "bbox": [ + 173, + 574, + 825, + 631 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Gated-Attention: Adapted from Chaplot et al. (2017), who used Gated-Attention with sentencelevel textual representations for SGN: $x _ { \\mathrm { S } } = f _ { \\mathrm { G A } } ( x _ { I } , x _ { \\mathrm { s e n t } } )$ . ", + "bbox": [ + 173, + 637, + 821, + 666 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "FiLM: Perez et al. (2017) introduced a general-purpose conditioning method called Feature-wise Linear Modulation (FiLM) for Visual Question Answering. Using FiLM, $x _ { \\mathrm { S } } = \\gamma ( x _ { \\mathrm { s e n t } } ) \\odot x _ { I } +$ $\\beta ( x _ { \\mathrm { s e n t } } )$ where $\\gamma ( x _ { \\mathrm { s e n t } } )$ and $\\beta ( x _ { \\mathrm { s e n t } } )$ are learnable projections of the sentence representation. ", + "bbox": [ + 174, + 672, + 825, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "PACMAN: Das et al. (2017) presented a hierarchical RL model for EQA. We adapt their method by using the attention mechanism in their QA module, which takes the last 5 frames and the text as input, and computes the similarity of the text with each frame using dot products between image and sentence-level text representations. These similarities are converted into attention weights using softmax, and the attention-weighted image features are concatenated with question embedding and passed through a softmax classifier to predict the answer distribution. For this particular baseline, we use the last 5 frames as input at each time step, unlike the proposed model and all other baselines which use a single frame as input. The attention-weighted image features are used as the state representation. The PACMAN model used a pretrained QA module, but we train this module jointly with the Navigation model for fair comparison with the proposed model. ", + "bbox": [ + 173, + 722, + 825, + 861 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For each of the above method except PACMAN, we use a linear layer $f$ with ReLU activations followed by softmax $\\sigma$ to get a $V$ -dimensional answer prediction from the state representations: $x _ { \\mathrm { A n s } } = \\sigma ( \\bar { f } ( x _ { \\mathrm { S } } ; \\theta _ { L i n } ) )$ . $x _ { \\mathrm { { S } } }$ and $x _ { A n s }$ are concatenated and passed to the policy module along with the time step and task indicator variable just as in the proposed model. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a6315480bb9a073f24a9632c6856b7874a62864e259308fb0863e2161017b386.jpg", + "image_caption": [ + "Figure 12: Plot showing the training accuracy of all the models without auxiliary tasks for Doom Easy and Hard environments. " + ], + "image_footnote": [], + "bbox": [ + 238, + 99, + 759, + 228 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/634ffdf84813ca11c0e122e1a00e556fb733df50e54d4f6c9664f00480c451dd.jpg", + "image_caption": [ + "Figure 13: Plot showing the training accuracy of 3 models across 3 training runs with different seeds with and without auxiliary tasks for Doom Easy environment without any smoothing. " + ], + "image_footnote": [], + "bbox": [ + 238, + 277, + 759, + 405 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C DOOM EXPERIMENT DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 465, + 457, + 483 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Additional Results. We show the training accuracy of all models without auxiliary tasks for Doom Easy and Hard in Figure 12. We also show the training accuracy of 3 models across 3 training runs with different seeds with and without auxiliary tasks for Doom Easy environment in Figure 13. ", + "bbox": [ + 174, + 497, + 825, + 540 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Dataset. The Doom objects used in our experiments are illustrated in Figure 14. Instructions and questions used for training and evaluation are listed in Tables 4. ", + "bbox": [ + 173, + 547, + 825, + 575 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/69d7bb5a3ad8695b8a37ae82bea6b9ec1dbc83f181543eb44b38a763cd846842.jpg", + "image_caption": [ + "Figure 14: Objects of various colors and sizes used in the ViZDoom environment. " + ], + "image_footnote": [], + "bbox": [ + 305, + 589, + 686, + 727 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/7057d42e3744461df649f3b29df5a3d9702f5c717d595cdcc3987488ef779a3c.jpg", + "table_caption": [ + "Table 4: Instructions and questions for ViZDoom experiments. We used 5 object classes (torch, pillar, keycard, skullkey, armor), 4 colors (red, green, blue, yellow), 2 sizes (tall, short), and 2 superlative sizes (smallest, largest). " + ], + "table_footnote": [], + "table_body": "
SGN Instruction Type42 Train Instructions: Not containing‘red'&‘pillar'28 Test Instructions: Containing‘red'or‘pillar'
Go to the (object).Go to the color>object.torch,keycard,skullkey,armor yellow, green, bluepillar red
Go to the size)object. Go to the {color) {object).tall, short blue torch, green torch, green armor,red torch,red skullkey,red pillar,
blue skullkey,blue keycard, yellow keycard, yellow skullkeygreen pillar,red keycard,red armor
Go to the {size) (object). Go to the color) (size)object.short torch,tall torch green tall, blue tall, blue short,tall pillar, short pillar red short, red tall
Go to the size){color)object.green short tall green, tall blue, short blue,short red, tall red
Go to the {color)(size)(object).short green green tall torch, green short torch,red short pillar,red short torch,
blue short torch,blue tall torchred tall pillar, green tall pillar, red tall torch, green short pillar
Go to the size) {color> {object).tall green torch,short green torch, short blue torch,tall blue torchshort red pillar, short red torch, tall red pillar, tall green pillar, tall red torch,short green pillar
Go to the superlative)object. Go to the {superlative> (color) object.largest, smallest smallest yellow, smallest blue, smallest green,largest blue,largest red,smallest red
EQAQuestion Typelargest green,largest yellow 21 Train Questions: Not containing‘blue'&‘torch'8 Test Questions: Containing‘blue'or‘torch'
Which {size)object is {color)in color?What color is the {object)? What color is the {size){object>? Which object is {color) in color?pillar, keycard, skullkey,armor short pillar, tall pillar red,yellow, greentorch short torch, tall torch blue
", + "bbox": [ + 191, + 353, + 808, + 720 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D HOUSE3D EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 102, + 421, + 118 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "In the House3D domain, we train on one house environment and randomize the colors of each object at the start of each episode. The agent’s spawn location is fixed. We create instructions and questions dataset for this house similar to the Doom domain. The House3D objects used in our experiments are illustrated in Figure 15. Instructions and questions used for training and evaluation are listed in Table 6. ", + "bbox": [ + 173, + 133, + 825, + 203 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Each model is trained for 50 million frames jointly on both SGN and EQA, without the auxiliary tasks and using identical reward functions. Similar to Doom, we use a $+ 1$ reward for reaching the correct object in SGN episodes and predicting the correct answer in the EQA episodes. We use a small negative reward of - 0.001 per time step to encourage shorter paths to target and answering the questions as soon as possible. We also use distance based reward shaping for both SGN and EQA episodes, where the agent receives a small reward proportional to decrease in distance to the target. SGN episodes end when the agent reaches any object and EQA episodes when agent predicts any answer. All episodes have a maximum length of 420 time steps. ", + "bbox": [ + 174, + 210, + 483, + 417 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/054eaa266cfb697cb6f36a23b30b1a988e4753b5ea56986060bcf5251f1d35fc.jpg", + "table_caption": [ + "Table 5: Accuracy of all the models on the SGN and EQA train and test sets for the House3D Domain. " + ], + "table_footnote": [], + "table_body": "
SGNEQA
ModelTrainTestTrainTest
Text only0.630.330.220.23
Image only0.280.010.120.22
Concat0.650.130.310.13
GA0.980.200.920.03
FiLM0.990.370.920.24
PACMAN0.730.200.400.21
Dual-Attention0.990.470.890.29
", + "bbox": [ + 496, + 265, + 774, + 395 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 419, + 576, + 433 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "In Table 5, we report the train and test performance of all the models on both SGN and EQA. The results are similar as in Doom: the Dual-Attention model outperforms the baselines by a considerable margin. ", + "bbox": [ + 174, + 439, + 826, + 481 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/af178f0259047b6ec08600ddcade5e3f117345df9c3c00343e94708605dc643a.jpg", + "image_caption": [ + "Figure 15: Example first-person views of the House3D environment with sample objects of various colors. " + ], + "image_footnote": [], + "bbox": [ + 184, + 494, + 815, + 604 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/5a2c6c4f7fcdbf5cd5e8da09a70edd31b66260c9c36a6aed86484d9c31826ff4.jpg", + "table_caption": [ + "Table 6: Instructions and questions for House3D experiments. We used 6 object classes (refrigerator, office chair, fish tank, fireplace, bed, sofa) and 4 colors (red, green, blue, yellow). " + ], + "table_footnote": [], + "table_body": "
SGNInstruction Type22 Train Instructions: Not containing‘red'&‘bed'11 Test Instructions: Containing‘red'or‘bed'
Go to the (object). Go to the {color)(object).refrigerator,office_chair,fish_tank,fireplace green refrigerator, green office_chair, green fish_tank,green fireplace, green sofa, blue refrigerator, blue office_chair, blue fish_tank,blue fireplace,blue sofa, yellow refrigerator,yellow office_chair, yellow fish_tank,yellow fireplace,yellow sofabed red bed, green bed,blue bed, yellow bed,red refrigerator, red office_chair,red fish_tank, red fireplace,red sofa
EQAGo to the {color) object. Question Typegreen,blue,yellow 7 Train Questions:red 2 Test Questions:
Not containing‘blue'&‘sofa'Containing‘blue'or‘sofa'
What color is the {object)? What object is color> in color?refrigerator,office_chair,fish_tank,fireplace,bed red, green,yellowsofa blue
", + "bbox": [ + 199, + 694, + 797, + 863 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/parse/train/ByGq7hRqKX/ByGq7hRqKX_middle.json b/parse/train/ByGq7hRqKX/ByGq7hRqKX_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..2197f975eb5304f76c3a6f1e3939270d4429e12e --- /dev/null +++ b/parse/train/ByGq7hRqKX/ByGq7hRqKX_middle.json @@ -0,0 +1,37889 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 424, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 425, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 425, + 97 + ], + "score": 1.0, + "content": "CROSS-TASK KNOWLEDGE TRANSFER FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 375, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 375, + 118 + ], + "score": 1.0, + "content": "VISUALLY-GROUNDED NAVIGATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 171, + 333, + 184 + ], + "lines": [ + { + "bbox": [ + 276, + 170, + 336, + 186 + ], + "spans": [ + { + "bbox": [ + 276, + 170, + 336, + 186 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 194, + 468, + 347 + ], + "lines": [ + { + "bbox": [ + 141, + 193, + 470, + 208 + ], + "spans": [ + { + "bbox": [ + 141, + 193, + 470, + 208 + ], + "score": 1.0, + "content": "Recent efforts on training visual navigation agents conditioned on language using", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 205, + 469, + 217 + ], + "spans": [ + { + "bbox": [ + 141, + 205, + 469, + 217 + ], + "score": 1.0, + "content": "deep reinforcement learning have been successful in learning policies for two dif-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 216, + 469, + 229 + ], + "spans": [ + { + "bbox": [ + 141, + 216, + 469, + 229 + ], + "score": 1.0, + "content": "ferent tasks: learning to follow navigational instructions and embodied question", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 227, + 469, + 240 + ], + "spans": [ + { + "bbox": [ + 141, + 227, + 469, + 240 + ], + "score": 1.0, + "content": "answering. In this paper, we aim to learn a multitask model capable of jointly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 238, + 470, + 250 + ], + "spans": [ + { + "bbox": [ + 141, + 238, + 470, + 250 + ], + "score": 1.0, + "content": "learning both tasks, and transferring knowledge of words and their grounding in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 249, + 470, + 262 + ], + "spans": [ + { + "bbox": [ + 141, + 249, + 470, + 262 + ], + "score": 1.0, + "content": "visual objects across tasks. The proposed model uses a novel Dual-Attention unit", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 260, + 469, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 260, + 469, + 272 + ], + "score": 1.0, + "content": "to disentangle the knowledge of words in the textual representations and visual ob-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 272, + 470, + 282 + ], + "spans": [ + { + "bbox": [ + 141, + 272, + 470, + 282 + ], + "score": 1.0, + "content": "jects in the visual representations, and align them with each other. This disentan-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 282, + 470, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 282, + 470, + 294 + ], + "score": 1.0, + "content": "gled task-invariant alignment of representations facilitates grounding and knowl-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 294, + 470, + 305 + ], + "spans": [ + { + "bbox": [ + 142, + 294, + 470, + 305 + ], + "score": 1.0, + "content": "edge transfer across both tasks. We show that the proposed model outperforms a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 304, + 470, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 470, + 316 + ], + "score": 1.0, + "content": "range of baselines on both tasks in simulated 3D environments. We also show that", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 315, + 469, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 469, + 327 + ], + "score": 1.0, + "content": "this disentanglement of representations makes our model modular, interpretable,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 325, + 470, + 339 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 470, + 339 + ], + "score": 1.0, + "content": "and allows for transfer to instructions containing new words by leveraging object", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 337, + 183, + 348 + ], + "spans": [ + { + "bbox": [ + 141, + 337, + 183, + 348 + ], + "score": 1.0, + "content": "detectors.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 365, + 206, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 208, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 208, + 381 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Deep reinforcement learning has been shown to be capable of achieving super-human performance", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "in playing games such as Atari 2600 (Mnih et al., 2013) and Go (Silver et al., 2016). Following", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "the success of deep reinforcement learning in 3D Games such as Doom (Lample & Chaplot, 2017;", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Dosovitskiy & Koltun, 2017) and DeepmindLab (Mnih et al., 2016), there has been increased in-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "terest in using deep reinforcement learning for training embodied agents, which interact with a 3D", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "environment by receiving first-person views of the environment and taking navigational actions. The", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 450, + 504, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 504, + 464 + ], + "score": 1.0, + "content": "simplest navigational agents learn a particular behaviour such as collecting or avoiding particular ob-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 461, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 104, + 461, + 506, + 475 + ], + "score": 1.0, + "content": "jects (Kempka et al., 2016; Jaderberg et al., 2016; Mirowski et al., 2016) or playing deathmatches", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "score": 1.0, + "content": "(Lample & Chaplot, 2017; Dosovitskiy & Koltun, 2017). Subsequently, there have been efforts on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "training navigational agents whose behaviour is conditioned on a target specified using images (Zhu", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "et al., 2017) or coordinates (Gupta et al., 2017a; Savva et al., 2017). More recently, there has been", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 504, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 520 + ], + "score": 1.0, + "content": "much interest in training agents conditioned on language as it offers several advantages over using", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 516, + 199, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 199, + 529 + ], + "score": 1.0, + "content": "images or coordinates.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 504, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 504, + 546 + ], + "score": 1.0, + "content": "Firstly, the compositionality of language allows generalization to new tasks without additional learn-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "ing. Prior work (Oh et al., 2017; Hermann et al., 2017; Chaplot et al., 2017) has trained navigational", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "agents to follow instructions and shown zero-shot generalization to new instructions which contain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "unseen composition of words seen in the training instructions. Secondly, language is also a con-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "score": 1.0, + "content": "venient means for humans to communicate with autonomous agents. Language not only allows", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "instruction but also interaction. Gordon et al. (2018) and Das et al. (2017) train agents to answer", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 599, + 421, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 421, + 612 + ], + "score": 1.0, + "content": "questions by navigating in the environment to gather the required information.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "These multimodal tasks involve several challenges, such as perception from raw pixels, grounding", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "of words in the instruction or question to visual objects and attributes, reasoning to perform rela-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "tional tasks, fine-grained navigation in 3D environments with continuous state space, and learning", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "to answer questions. Training a multi-task model can also facilitate knowledge transfer between the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "tasks and allow the model to generalize to scenarios which were not possible with single tasks. For", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "example, if an agent learns to follow the instruction ‘Go to the red pillar’ and answer the question", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 681, + 504, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 504, + 694 + ], + "score": 1.0, + "content": "‘What color is the torch?’, then it should also be able to follow the instruction ‘Go to the red torch’", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 691, + 439, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 439, + 707 + ], + "score": 1.0, + "content": "and answer the question ‘What color is the pillar?’ without any additional training.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "In this paper, we aim to train a multi-task navigation model to follow instructions and answer ques-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "tions. To test the generalization of multi-task models, we define cross-task knowledge transfer as an", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 424, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 425, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 425, + 97 + ], + "score": 1.0, + "content": "CROSS-TASK KNOWLEDGE TRANSFER FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 375, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 375, + 118 + ], + "score": 1.0, + "content": "VISUALLY-GROUNDED NAVIGATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 171, + 333, + 184 + ], + "lines": [ + { + "bbox": [ + 276, + 170, + 336, + 186 + ], + "spans": [ + { + "bbox": [ + 276, + 170, + 336, + 186 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 194, + 468, + 347 + ], + "lines": [ + { + "bbox": [ + 141, + 193, + 470, + 208 + ], + "spans": [ + { + "bbox": [ + 141, + 193, + 470, + 208 + ], + "score": 1.0, + "content": "Recent efforts on training visual navigation agents conditioned on language using", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 205, + 469, + 217 + ], + "spans": [ + { + "bbox": [ + 141, + 205, + 469, + 217 + ], + "score": 1.0, + "content": "deep reinforcement learning have been successful in learning policies for two dif-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 216, + 469, + 229 + ], + "spans": [ + { + "bbox": [ + 141, + 216, + 469, + 229 + ], + "score": 1.0, + "content": "ferent tasks: learning to follow navigational instructions and embodied question", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 227, + 469, + 240 + ], + "spans": [ + { + "bbox": [ + 141, + 227, + 469, + 240 + ], + "score": 1.0, + "content": "answering. In this paper, we aim to learn a multitask model capable of jointly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 238, + 470, + 250 + ], + "spans": [ + { + "bbox": [ + 141, + 238, + 470, + 250 + ], + "score": 1.0, + "content": "learning both tasks, and transferring knowledge of words and their grounding in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 249, + 470, + 262 + ], + "spans": [ + { + "bbox": [ + 141, + 249, + 470, + 262 + ], + "score": 1.0, + "content": "visual objects across tasks. The proposed model uses a novel Dual-Attention unit", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 260, + 469, + 272 + ], + "spans": [ + { + "bbox": [ + 141, + 260, + 469, + 272 + ], + "score": 1.0, + "content": "to disentangle the knowledge of words in the textual representations and visual ob-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 272, + 470, + 282 + ], + "spans": [ + { + "bbox": [ + 141, + 272, + 470, + 282 + ], + "score": 1.0, + "content": "jects in the visual representations, and align them with each other. This disentan-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 282, + 470, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 282, + 470, + 294 + ], + "score": 1.0, + "content": "gled task-invariant alignment of representations facilitates grounding and knowl-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 294, + 470, + 305 + ], + "spans": [ + { + "bbox": [ + 142, + 294, + 470, + 305 + ], + "score": 1.0, + "content": "edge transfer across both tasks. We show that the proposed model outperforms a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 304, + 470, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 470, + 316 + ], + "score": 1.0, + "content": "range of baselines on both tasks in simulated 3D environments. We also show that", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 315, + 469, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 469, + 327 + ], + "score": 1.0, + "content": "this disentanglement of representations makes our model modular, interpretable,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 325, + 470, + 339 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 470, + 339 + ], + "score": 1.0, + "content": "and allows for transfer to instructions containing new words by leveraging object", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 337, + 183, + 348 + ], + "spans": [ + { + "bbox": [ + 141, + 337, + 183, + 348 + ], + "score": 1.0, + "content": "detectors.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11.5, + "bbox_fs": [ + 141, + 193, + 470, + 348 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 365, + 206, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 208, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 208, + 381 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Deep reinforcement learning has been shown to be capable of achieving super-human performance", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "in playing games such as Atari 2600 (Mnih et al., 2013) and Go (Silver et al., 2016). Following", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "the success of deep reinforcement learning in 3D Games such as Doom (Lample & Chaplot, 2017;", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "Dosovitskiy & Koltun, 2017) and DeepmindLab (Mnih et al., 2016), there has been increased in-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "terest in using deep reinforcement learning for training embodied agents, which interact with a 3D", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "environment by receiving first-person views of the environment and taking navigational actions. The", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 450, + 504, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 504, + 464 + ], + "score": 1.0, + "content": "simplest navigational agents learn a particular behaviour such as collecting or avoiding particular ob-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 461, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 104, + 461, + 506, + 475 + ], + "score": 1.0, + "content": "jects (Kempka et al., 2016; Jaderberg et al., 2016; Mirowski et al., 2016) or playing deathmatches", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "score": 1.0, + "content": "(Lample & Chaplot, 2017; Dosovitskiy & Koltun, 2017). Subsequently, there have been efforts on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "training navigational agents whose behaviour is conditioned on a target specified using images (Zhu", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "et al., 2017) or coordinates (Gupta et al., 2017a; Savva et al., 2017). More recently, there has been", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 504, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 520 + ], + "score": 1.0, + "content": "much interest in training agents conditioned on language as it offers several advantages over using", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 516, + 199, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 199, + 529 + ], + "score": 1.0, + "content": "images or coordinates.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 384, + 506, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 504, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 504, + 546 + ], + "score": 1.0, + "content": "Firstly, the compositionality of language allows generalization to new tasks without additional learn-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "ing. Prior work (Oh et al., 2017; Hermann et al., 2017; Chaplot et al., 2017) has trained navigational", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 568 + ], + "score": 1.0, + "content": "agents to follow instructions and shown zero-shot generalization to new instructions which contain", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "unseen composition of words seen in the training instructions. Secondly, language is also a con-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "score": 1.0, + "content": "venient means for humans to communicate with autonomous agents. Language not only allows", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "instruction but also interaction. Gordon et al. (2018) and Das et al. (2017) train agents to answer", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 599, + 421, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 421, + 612 + ], + "score": 1.0, + "content": "questions by navigating in the environment to gather the required information.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 533, + 506, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "These multimodal tasks involve several challenges, such as perception from raw pixels, grounding", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "of words in the instruction or question to visual objects and attributes, reasoning to perform rela-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "tional tasks, fine-grained navigation in 3D environments with continuous state space, and learning", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "to answer questions. Training a multi-task model can also facilitate knowledge transfer between the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "tasks and allow the model to generalize to scenarios which were not possible with single tasks. For", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "example, if an agent learns to follow the instruction ‘Go to the red pillar’ and answer the question", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 681, + 504, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 504, + 694 + ], + "score": 1.0, + "content": "‘What color is the torch?’, then it should also be able to follow the instruction ‘Go to the red torch’", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 691, + 439, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 439, + 707 + ], + "score": 1.0, + "content": "and answer the question ‘What color is the pillar?’ without any additional training.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 615, + 505, + 707 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "In this paper, we aim to train a multi-task navigation model to follow instructions and answer ques-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "tions. To test the generalization of multi-task models, we define cross-task knowledge transfer as an", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "evaluation criteria, evaluating zero-shot learning on instructions and questions consisting of unseen", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "score": 1.0, + "content": "composition of words in both tasks. In order to achieve cross-task knowledge transfer, words in", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 233, + 504, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 504, + 245 + ], + "score": 1.0, + "content": "the input space of both tasks need to be aligned with each other and with the answer space while", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "they are being grounded to visual objects and attributes. Prior models fail to achieve this as they are", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "designed for a single task. We propose a novel dual-attention model involving sequential Gated- and", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "Spatial-Attention operations to perform explicit task-invariant alignment between the image repre-", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "sentation channels and the words in the input and answer space. We create datasets and simulation", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "scenarios for testing cross-task knowledge transfer in the Doom environment (Kempka et al., 2016)", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 298, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 312 + ], + "score": 1.0, + "content": "and show that the proposed model outperforms a range of baselines on both tasks. Additionally, we", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "demonstrate that the modularity of our model allows easy addition of new objects and attributes to", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "a trained model. We plan to open-source the implementation of our proposed model as well as the", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 332, + 261, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 261, + 344 + ], + "score": 1.0, + "content": "datasets and simulation environments.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 80, + 488, + 149 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 80, + 488, + 149 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 80, + 488, + 149 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 488, + 149 + ], + "score": 0.645, + "type": "image", + "image_path": "c73220668e2566d5356bf6cee3722954a8f4847bd33438358c45fbb39f48d227.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 80, + 488, + 103.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 103.0, + 488, + 126.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 126.0, + 488, + 149.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 155, + 504, + 185 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 155, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 166 + ], + "score": 1.0, + "content": "Figure 1: An example of first-person view in the 3D Doom environment with sample instructions and ques-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 165, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 176 + ], + "score": 1.0, + "content": "tions. The test set consists of unseen instructions and questions. The dataset evaluates a model for cross-task", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 174, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 492, + 186 + ], + "score": 1.0, + "content": "knowledge transfer between Semantic Goal Navigation (SGN) and Embodied Question Answering (EQA).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 210, + 505, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "evaluation criteria, evaluating zero-shot learning on instructions and questions consisting of unseen", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 234 + ], + "score": 1.0, + "content": "composition of words in both tasks. In order to achieve cross-task knowledge transfer, words in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 233, + 504, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 504, + 245 + ], + "score": 1.0, + "content": "the input space of both tasks need to be aligned with each other and with the answer space while", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "they are being grounded to visual objects and attributes. Prior models fail to achieve this as they are", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "designed for a single task. We propose a novel dual-attention model involving sequential Gated- and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "Spatial-Attention operations to perform explicit task-invariant alignment between the image repre-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "sentation channels and the words in the input and answer space. We create datasets and simulation", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "scenarios for testing cross-task knowledge transfer in the Doom environment (Kempka et al., 2016)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 298, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 312 + ], + "score": 1.0, + "content": "and show that the proposed model outperforms a range of baselines on both tasks. Additionally, we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "demonstrate that the modularity of our model allows easy addition of new objects and attributes to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "a trained model. We plan to open-source the implementation of our proposed model as well as the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 332, + 261, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 261, + 344 + ], + "score": 1.0, + "content": "datasets and simulation environments.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 361, + 211, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 213, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 213, + 376 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 400 + ], + "score": 1.0, + "content": "This paper is motivated by a series of works on learning to follow navigation instructions (Oh et al.,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 396, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 413 + ], + "score": 1.0, + "content": "2017; Hermann et al., 2017; Chaplot et al., 2017; Wu et al., 2018; Yu et al., 2018a) and learning", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "to answer questions by navigating around the environment (Das et al., 2017; Gordon et al., 2018).", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "Among methods learning from instructions in 3D environments, Oh et al. (2017) introduced a hierar-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 430, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 445 + ], + "score": 1.0, + "content": "chical RL model for learning sequences of instructions by learning skills to solve subtasks. Chaplot", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 104, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "et al. (2017) introduced a gated-attention model for multimodal fusion of textual and visual rep-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 467 + ], + "score": 1.0, + "content": "resentations using multiplicative interactions, whereas Hermann et al. (2017) introduced auxiliary", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "tasks such as temporal autoencoding and language prediction to improve sample efficiency for this", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "task. Yu et al. (2018a) proposed guided feature transformation which involves transformation of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 474, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 474, + 500 + ], + "score": 1.0, + "content": "visual representations using latent sentence embeddings computed from the language input.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "Among models for embodied question answering, Das et al. (2017) introduced a hierarchical model", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "consisting of 4 modules, each for processing images, encoding questions, navigation, and question-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 524, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 539 + ], + "score": 1.0, + "content": "answering, each of which is pretrained with supervised or imitation learning, followed by fine-tuning", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "of the navigation model using reinforcement learning. Gordon et al. (2018) introduced the task of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "Interactive Question Answering which involves interacting with objects in the environment with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 557, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 572 + ], + "score": 1.0, + "content": "non-navigational actions for answering questions. They proposed Hierarchical Interactive Memory", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 569, + 466, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 466, + 581 + ], + "score": 1.0, + "content": "Network (HIMN), which allows temporal abstraction using a factorized set of controllers.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 585, + 504, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 599 + ], + "score": 1.0, + "content": "All of the above methods are designed for a single task, following navigational instructions or an-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "swering questions, whereas we aim to train a single model for both tasks. Yu et al. (2018b) intro-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "duced a model for interactive language acquisition by training on both Visual Question Answering", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "score": 1.0, + "content": "and following instructions in a 2D grid world environment. We aim to tackle multimodal multitask", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 629, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 643 + ], + "score": 1.0, + "content": "learning in challenging 3D environments. Partial observability results in the requirement of learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "to navigate for answering the questions, turning visual question answering to embodied question", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 652, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 664 + ], + "score": 1.0, + "content": "answering. 3D environments also allow us to test on interesting and more challenging instructions", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 662, + 378, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 378, + 675 + ], + "score": 1.0, + "content": "based on relative size of the objects, in addition to colors and types.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 108, + 679, + 504, + 735 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 505, + 692 + ], + "score": 1.0, + "content": "In addition to the above, there is a huge body of work on multimodal learning in static settings", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "score": 1.0, + "content": "which do not involve navigation or reinforcement learning. Some relevant works which use attention", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 701, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 505, + 714 + ], + "score": 1.0, + "content": "mechanisms similar to the ones used in our proposed model include Perez et al. (2017); Fukui et al.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 711, + 506, + 726 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 726 + ], + "score": 1.0, + "content": "(2016); Xu & Saenko (2016); Hudson & Manning (2018); Gupta et al. (2017b) for Visual Question", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 723, + 365, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 723, + 365, + 735 + ], + "score": 1.0, + "content": "Answering and Zhao et al. (2018) for grounding audio to vision.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 80, + 488, + 149 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 80, + 488, + 149 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 80, + 488, + 149 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 488, + 149 + ], + "score": 0.645, + "type": "image", + "image_path": "c73220668e2566d5356bf6cee3722954a8f4847bd33438358c45fbb39f48d227.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 80, + 488, + 103.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 103.0, + 488, + 126.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 126.0, + 488, + 149.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 155, + 504, + 185 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 155, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 166 + ], + "score": 1.0, + "content": "Figure 1: An example of first-person view in the 3D Doom environment with sample instructions and ques-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 165, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 176 + ], + "score": 1.0, + "content": "tions. The test set consists of unseen instructions and questions. The dataset evaluates a model for cross-task", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 174, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 492, + 186 + ], + "score": 1.0, + "content": "knowledge transfer between Semantic Goal Navigation (SGN) and Embodied Question Answering (EQA).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 210, + 505, + 343 + ], + "lines": [], + "index": 11.5, + "bbox_fs": [ + 105, + 211, + 506, + 344 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 361, + 211, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 213, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 213, + 376 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 400 + ], + "score": 1.0, + "content": "This paper is motivated by a series of works on learning to follow navigation instructions (Oh et al.,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 396, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 413 + ], + "score": 1.0, + "content": "2017; Hermann et al., 2017; Chaplot et al., 2017; Wu et al., 2018; Yu et al., 2018a) and learning", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "to answer questions by navigating around the environment (Das et al., 2017; Gordon et al., 2018).", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "Among methods learning from instructions in 3D environments, Oh et al. (2017) introduced a hierar-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 430, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 445 + ], + "score": 1.0, + "content": "chical RL model for learning sequences of instructions by learning skills to solve subtasks. Chaplot", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 104, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "et al. (2017) introduced a gated-attention model for multimodal fusion of textual and visual rep-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 467 + ], + "score": 1.0, + "content": "resentations using multiplicative interactions, whereas Hermann et al. (2017) introduced auxiliary", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "tasks such as temporal autoencoding and language prediction to improve sample efficiency for this", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "task. Yu et al. (2018a) proposed guided feature transformation which involves transformation of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 474, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 474, + 500 + ], + "score": 1.0, + "content": "visual representations using latent sentence embeddings computed from the language input.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 388, + 506, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "Among models for embodied question answering, Das et al. (2017) introduced a hierarchical model", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "consisting of 4 modules, each for processing images, encoding questions, navigation, and question-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 524, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 539 + ], + "score": 1.0, + "content": "answering, each of which is pretrained with supervised or imitation learning, followed by fine-tuning", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "of the navigation model using reinforcement learning. Gordon et al. (2018) introduced the task of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "Interactive Question Answering which involves interacting with objects in the environment with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 557, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 572 + ], + "score": 1.0, + "content": "non-navigational actions for answering questions. They proposed Hierarchical Interactive Memory", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 569, + 466, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 466, + 581 + ], + "score": 1.0, + "content": "Network (HIMN), which allows temporal abstraction using a factorized set of controllers.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 503, + 505, + 581 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 585, + 504, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 599 + ], + "score": 1.0, + "content": "All of the above methods are designed for a single task, following navigational instructions or an-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "swering questions, whereas we aim to train a single model for both tasks. Yu et al. (2018b) intro-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "duced a model for interactive language acquisition by training on both Visual Question Answering", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "score": 1.0, + "content": "and following instructions in a 2D grid world environment. We aim to tackle multimodal multitask", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 629, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 643 + ], + "score": 1.0, + "content": "learning in challenging 3D environments. Partial observability results in the requirement of learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "to navigate for answering the questions, turning visual question answering to embodied question", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 652, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 664 + ], + "score": 1.0, + "content": "answering. 3D environments also allow us to test on interesting and more challenging instructions", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 662, + 378, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 378, + 675 + ], + "score": 1.0, + "content": "based on relative size of the objects, in addition to colors and types.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 585, + 506, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 679, + 504, + 735 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 505, + 692 + ], + "score": 1.0, + "content": "In addition to the above, there is a huge body of work on multimodal learning in static settings", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 505, + 702 + ], + "score": 1.0, + "content": "which do not involve navigation or reinforcement learning. Some relevant works which use attention", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 701, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 505, + 714 + ], + "score": 1.0, + "content": "mechanisms similar to the ones used in our proposed model include Perez et al. (2017); Fukui et al.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 711, + 506, + 726 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 726 + ], + "score": 1.0, + "content": "(2016); Xu & Saenko (2016); Hudson & Manning (2018); Gupta et al. (2017b) for Visual Question", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 723, + 365, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 723, + 365, + 735 + ], + "score": 1.0, + "content": "Answering and Zhao et al. (2018) for grounding audio to vision.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 679, + 506, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 83, + 481, + 155 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 83, + 481, + 155 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 83, + 481, + 155 + ], + "spans": [ + { + "bbox": [ + 124, + 83, + 481, + 155 + ], + "score": 0.96, + "type": "image", + "image_path": "a49406c1969d1fd52bfa11b26df5a13669079ff90e34764405f0a9c3943df268.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 83, + 481, + 107.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 107.0, + 481, + 131.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 131.0, + 481, + 155.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 158, + 159, + 453, + 170 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 158, + 454, + 172 + ], + "spans": [ + { + "bbox": [ + 156, + 158, + 454, + 172 + ], + "score": 1.0, + "content": "Figure 2: Overview of our proposed architecture, described in detail in Section 4.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 108, + 191, + 256, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 258, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 258, + 207 + ], + "score": 1.0, + "content": "3 PROBLEM FORMULATION", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 505, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 506, + 226 + ], + "score": 1.0, + "content": "Consider an autonomous agent interacting with an episodic environment as shown in Figure 1. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 224, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 359, + 236 + ], + "score": 1.0, + "content": "the beginning of each episode, the agent receives a textual input", + "type": "text" + }, + { + "bbox": [ + 360, + 224, + 368, + 234 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 224, + 505, + 236 + ], + "score": 1.0, + "content": "specifying the task that it needs to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 233, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 197, + 249 + ], + "score": 1.0, + "content": "achieve. For example,", + "type": "text" + }, + { + "bbox": [ + 198, + 235, + 207, + 245 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 233, + 505, + 249 + ], + "score": 1.0, + "content": "could be an instruction describing the target object or a question querying", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 245, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 410, + 259 + ], + "score": 1.0, + "content": "about some visual detail of objects in the environment. At each time step", + "type": "text" + }, + { + "bbox": [ + 410, + 247, + 415, + 256 + ], + "score": 0.59, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 245, + 506, + 259 + ], + "score": 1.0, + "content": ", the agent observes a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 128, + 269 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 128, + 257, + 181, + 269 + ], + "score": 0.93, + "content": "s _ { t } = ( I _ { t } , T )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 257, + 210, + 269 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 210, + 257, + 219, + 268 + ], + "score": 0.86, + "content": "I _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "is the first-person (egocentric) view of the environment, and takes an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 133, + 280 + ], + "score": 1.0, + "content": "action", + "type": "text" + }, + { + "bbox": [ + 133, + 270, + 143, + 279 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 267, + 505, + 280 + ], + "score": 1.0, + "content": ", which could be a navigational action or an answer action. The agent’s objective is to learn", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 278, + 501, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 140, + 291 + ], + "score": 1.0, + "content": "a policy", + "type": "text" + }, + { + "bbox": [ + 141, + 278, + 175, + 290 + ], + "score": 0.93, + "content": "\\pi ( a _ { t } | s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 279, + 489, + 291 + ], + "score": 1.0, + "content": "which leads to successful completion of the task specified by the textual input", + "type": "text" + }, + { + "bbox": [ + 489, + 279, + 497, + 288 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 279, + 501, + 291 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "Tasks. We focus on the multi-task learning of two visually-grounded language navigation tasks: In", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 235, + 320 + ], + "score": 1.0, + "content": "Embodied Question Answering", + "type": "text" + }, + { + "bbox": [ + 236, + 307, + 262, + 318 + ], + "score": 0.28, + "content": "( E Q A )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 306, + 505, + 320 + ], + "score": 1.0, + "content": ", the agent is given a question (“What color is the torch?”),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "and it must navigate around the 3D environment to explore the environment and gather information", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 327, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 342 + ], + "score": 1.0, + "content": "to answer the question (“red”). In Semantic Goal Navigation (SGN), the agent is given a language", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 339, + 365, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 155, + 352 + ], + "score": 1.0, + "content": "instruction (", + "type": "text" + }, + { + "bbox": [ + 155, + 340, + 173, + 350 + ], + "score": 0.4, + "content": "^ { 6 6 } \\mathrm { G o }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 339, + 365, + 352 + ], + "score": 1.0, + "content": "to the red torch”) to navigate to a goal location.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 356, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Environments. We adapt the ViZDoom (Kempka et al., 2016)-based language grounding environ-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "ment proposed by Chaplot et al. (2017) for visually-grounded multitask learning. It consists of a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 378, + 504, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 504, + 390 + ], + "score": 1.0, + "content": "single room with 5 objects. The objects are randomized in each episode based on the textual input.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "We use two difficulty settings for the Doom domain: Easy: The agent is spawned at a fixed location.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "The candidate objects are spawned at five fixed locations along a single horizontal line in the field", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "of view of the agent. Hard: The candidate objects and the agent are spawned at random locations", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "and the objects may or may not be in the agents field of view in the initial configuration. The agent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 433, + 272, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 272, + 445 + ], + "score": 1.0, + "content": "must explore the map to view all objects.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "Datasets. We use the set of instructions from Chaplot et al. (2017) and create a dataset for questions", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "using the same set of objects and attributes. We define cross-task knowledge transfer as an evaluation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "criteria for testing generalization of multi-task models. We create train-test splits for both instruc-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 482, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 496 + ], + "score": 1.0, + "content": "tions and questions datasets to explicitly test a multitask model’s ability to transfer the knowledge", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "of words across different tasks. Each instruction in the test set contains a word that is never seen in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "any instruction in the training set but is seen in some questions in the training set. Similarly, each", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "question in the test set contains a word never seen in any training set question. Figure 1 illustrates", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "the train-test split of instructions and questions used in our experiments in the Doom domain. Note", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 537, + 384, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 384, + 550 + ], + "score": 1.0, + "content": "that for the EQA trainset, unseen words can be present in the answer.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "The agent can take 4 actions: 3 navigational actions (forward, left, right) and 1 answer action. When", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "the agent takes the answer action, the answer with the maximum probability in the output answer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "distribution is used. Other details such as the train-test splits are deferred to the Appendix. We also", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 497, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 497, + 600 + ], + "score": 1.0, + "content": "report results on an additional environment based on House3D (Wu et al., 2018) in the Appendix.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 108, + 611, + 231, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 611, + 231, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 231, + 626 + ], + "score": 1.0, + "content": "4 PROPOSED METHOD", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "In this section, we detail our proposed architecture (illustrated in Figure 2). At the start of each", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 278, + 656 + ], + "score": 1.0, + "content": "episode, the agent receives a textual input", + "type": "text" + }, + { + "bbox": [ + 279, + 644, + 287, + 654 + ], + "score": 0.77, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "(an instruction or a question) specifying the task that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 427, + 668 + ], + "score": 1.0, + "content": "it needs to achieve. At each time step, the agent observes an egocentric image", + "type": "text" + }, + { + "bbox": [ + 428, + 655, + 437, + 666 + ], + "score": 0.87, + "content": "I _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "which is passed", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "through a convolutional neural network (LeCun et al., 1995) with ReLU activations (Glorot et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 103, + 673, + 507, + 692 + ], + "spans": [ + { + "bbox": [ + 103, + 673, + 285, + 692 + ], + "score": 1.0, + "content": "2011) to produce the image representation", + "type": "text" + }, + { + "bbox": [ + 285, + 676, + 416, + 689 + ], + "score": 0.92, + "content": "x _ { I } = f ( I _ { t } ; \\theta _ { \\mathrm { c o n v } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 673, + 450, + 692 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 450, + 677, + 470, + 688 + ], + "score": 0.9, + "content": "\\theta _ { \\mathrm { c o n v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 673, + 507, + 692 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 289, + 700 + ], + "score": 1.0, + "content": "the parameters of the convolutional network,", + "type": "text" + }, + { + "bbox": [ + 289, + 689, + 298, + 698 + ], + "score": 0.75, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "is the number of feature maps in the convolutional", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 335, + 712 + ], + "score": 1.0, + "content": "network output which is equal to the vocabulary size, and", + "type": "text" + }, + { + "bbox": [ + 335, + 699, + 346, + 709 + ], + "score": 0.84, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 698, + 363, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 363, + 699, + 375, + 709 + ], + "score": 0.73, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "are the height and width of each", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 349, + 722 + ], + "score": 1.0, + "content": "feature map. We use two representations for the textual input", + "type": "text" + }, + { + "bbox": [ + 349, + 710, + 358, + 720 + ], + "score": 0.7, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ": (1) the bag-of-words representation", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 718, + 507, + 734 + ], + "spans": [ + { + "bbox": [ + 104, + 718, + 155, + 734 + ], + "score": 1.0, + "content": "denoted by", + "type": "text" + }, + { + "bbox": [ + 155, + 720, + 209, + 732 + ], + "score": 0.92, + "content": "x _ { \\mathrm { B o W } } ~ \\in ~ \\mathbb { R } ^ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 718, + 352, + 734 + ], + "score": 1.0, + "content": "and (2) a sentence representation", + "type": "text" + }, + { + "bbox": [ + 352, + 720, + 462, + 732 + ], + "score": 0.92, + "content": "x _ { \\mathrm { s e n t } } = f ( \\breve { T } ; \\theta _ { \\mathrm { s e n t } } ) \\in \\mathbb { R } ^ { \\breve { V } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 718, + 507, + 734 + ], + "score": 1.0, + "content": ", which is", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 83, + 481, + 155 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 83, + 481, + 155 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 83, + 481, + 155 + ], + "spans": [ + { + "bbox": [ + 124, + 83, + 481, + 155 + ], + "score": 0.96, + "type": "image", + "image_path": "a49406c1969d1fd52bfa11b26df5a13669079ff90e34764405f0a9c3943df268.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 83, + 481, + 107.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 107.0, + 481, + 131.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 131.0, + 481, + 155.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 158, + 159, + 453, + 170 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 158, + 454, + 172 + ], + "spans": [ + { + "bbox": [ + 156, + 158, + 454, + 172 + ], + "score": 1.0, + "content": "Figure 2: Overview of our proposed architecture, described in detail in Section 4.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 108, + 191, + 256, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 258, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 258, + 207 + ], + "score": 1.0, + "content": "3 PROBLEM FORMULATION", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 505, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 506, + 226 + ], + "score": 1.0, + "content": "Consider an autonomous agent interacting with an episodic environment as shown in Figure 1. In", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 224, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 224, + 359, + 236 + ], + "score": 1.0, + "content": "the beginning of each episode, the agent receives a textual input", + "type": "text" + }, + { + "bbox": [ + 360, + 224, + 368, + 234 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 224, + 505, + 236 + ], + "score": 1.0, + "content": "specifying the task that it needs to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 233, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 197, + 249 + ], + "score": 1.0, + "content": "achieve. For example,", + "type": "text" + }, + { + "bbox": [ + 198, + 235, + 207, + 245 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 233, + 505, + 249 + ], + "score": 1.0, + "content": "could be an instruction describing the target object or a question querying", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 245, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 410, + 259 + ], + "score": 1.0, + "content": "about some visual detail of objects in the environment. At each time step", + "type": "text" + }, + { + "bbox": [ + 410, + 247, + 415, + 256 + ], + "score": 0.59, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 245, + 506, + 259 + ], + "score": 1.0, + "content": ", the agent observes a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 128, + 269 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 128, + 257, + 181, + 269 + ], + "score": 0.93, + "content": "s _ { t } = ( I _ { t } , T )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 257, + 210, + 269 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 210, + 257, + 219, + 268 + ], + "score": 0.86, + "content": "I _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "is the first-person (egocentric) view of the environment, and takes an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 133, + 280 + ], + "score": 1.0, + "content": "action", + "type": "text" + }, + { + "bbox": [ + 133, + 270, + 143, + 279 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 267, + 505, + 280 + ], + "score": 1.0, + "content": ", which could be a navigational action or an answer action. The agent’s objective is to learn", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 278, + 501, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 140, + 291 + ], + "score": 1.0, + "content": "a policy", + "type": "text" + }, + { + "bbox": [ + 141, + 278, + 175, + 290 + ], + "score": 0.93, + "content": "\\pi ( a _ { t } | s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 279, + 489, + 291 + ], + "score": 1.0, + "content": "which leads to successful completion of the task specified by the textual input", + "type": "text" + }, + { + "bbox": [ + 489, + 279, + 497, + 288 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 279, + 501, + 291 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 212, + 506, + 291 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "Tasks. We focus on the multi-task learning of two visually-grounded language navigation tasks: In", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 235, + 320 + ], + "score": 1.0, + "content": "Embodied Question Answering", + "type": "text" + }, + { + "bbox": [ + 236, + 307, + 262, + 318 + ], + "score": 0.28, + "content": "( E Q A )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 306, + 505, + 320 + ], + "score": 1.0, + "content": ", the agent is given a question (“What color is the torch?”),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "and it must navigate around the 3D environment to explore the environment and gather information", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 327, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 342 + ], + "score": 1.0, + "content": "to answer the question (“red”). In Semantic Goal Navigation (SGN), the agent is given a language", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 339, + 365, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 155, + 352 + ], + "score": 1.0, + "content": "instruction (", + "type": "text" + }, + { + "bbox": [ + 155, + 340, + 173, + 350 + ], + "score": 0.4, + "content": "^ { 6 6 } \\mathrm { G o }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 339, + 365, + 352 + ], + "score": 1.0, + "content": "to the red torch”) to navigate to a goal location.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 295, + 505, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 356, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Environments. We adapt the ViZDoom (Kempka et al., 2016)-based language grounding environ-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "ment proposed by Chaplot et al. (2017) for visually-grounded multitask learning. It consists of a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 378, + 504, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 504, + 390 + ], + "score": 1.0, + "content": "single room with 5 objects. The objects are randomized in each episode based on the textual input.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "We use two difficulty settings for the Doom domain: Easy: The agent is spawned at a fixed location.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "The candidate objects are spawned at five fixed locations along a single horizontal line in the field", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "of view of the agent. Hard: The candidate objects and the agent are spawned at random locations", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "and the objects may or may not be in the agents field of view in the initial configuration. The agent", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 433, + 272, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 272, + 445 + ], + "score": 1.0, + "content": "must explore the map to view all objects.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 356, + 506, + 445 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "Datasets. We use the set of instructions from Chaplot et al. (2017) and create a dataset for questions", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "using the same set of objects and attributes. We define cross-task knowledge transfer as an evaluation", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "criteria for testing generalization of multi-task models. We create train-test splits for both instruc-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 482, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 496 + ], + "score": 1.0, + "content": "tions and questions datasets to explicitly test a multitask model’s ability to transfer the knowledge", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "of words across different tasks. Each instruction in the test set contains a word that is never seen in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "any instruction in the training set but is seen in some questions in the training set. Similarly, each", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "question in the test set contains a word never seen in any training set question. Figure 1 illustrates", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "the train-test split of instructions and questions used in our experiments in the Doom domain. Note", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 537, + 384, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 384, + 550 + ], + "score": 1.0, + "content": "that for the EQA trainset, unseen words can be present in the answer.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 450, + 506, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "The agent can take 4 actions: 3 navigational actions (forward, left, right) and 1 answer action. When", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "the agent takes the answer action, the answer with the maximum probability in the output answer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "distribution is used. Other details such as the train-test splits are deferred to the Appendix. We also", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 497, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 497, + 600 + ], + "score": 1.0, + "content": "report results on an additional environment based on House3D (Wu et al., 2018) in the Appendix.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 554, + 506, + 600 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 611, + 231, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 611, + 231, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 231, + 626 + ], + "score": 1.0, + "content": "4 PROPOSED METHOD", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "In this section, we detail our proposed architecture (illustrated in Figure 2). At the start of each", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 644, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 278, + 656 + ], + "score": 1.0, + "content": "episode, the agent receives a textual input", + "type": "text" + }, + { + "bbox": [ + 279, + 644, + 287, + 654 + ], + "score": 0.77, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 644, + 506, + 656 + ], + "score": 1.0, + "content": "(an instruction or a question) specifying the task that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 427, + 668 + ], + "score": 1.0, + "content": "it needs to achieve. At each time step, the agent observes an egocentric image", + "type": "text" + }, + { + "bbox": [ + 428, + 655, + 437, + 666 + ], + "score": 0.87, + "content": "I _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "which is passed", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "through a convolutional neural network (LeCun et al., 1995) with ReLU activations (Glorot et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 103, + 673, + 507, + 692 + ], + "spans": [ + { + "bbox": [ + 103, + 673, + 285, + 692 + ], + "score": 1.0, + "content": "2011) to produce the image representation", + "type": "text" + }, + { + "bbox": [ + 285, + 676, + 416, + 689 + ], + "score": 0.92, + "content": "x _ { I } = f ( I _ { t } ; \\theta _ { \\mathrm { c o n v } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 673, + 450, + 692 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 450, + 677, + 470, + 688 + ], + "score": 0.9, + "content": "\\theta _ { \\mathrm { c o n v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 673, + 507, + 692 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 289, + 700 + ], + "score": 1.0, + "content": "the parameters of the convolutional network,", + "type": "text" + }, + { + "bbox": [ + 289, + 689, + 298, + 698 + ], + "score": 0.75, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "is the number of feature maps in the convolutional", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 335, + 712 + ], + "score": 1.0, + "content": "network output which is equal to the vocabulary size, and", + "type": "text" + }, + { + "bbox": [ + 335, + 699, + 346, + 709 + ], + "score": 0.84, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 698, + 363, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 363, + 699, + 375, + 709 + ], + "score": 0.73, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "are the height and width of each", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 349, + 722 + ], + "score": 1.0, + "content": "feature map. We use two representations for the textual input", + "type": "text" + }, + { + "bbox": [ + 349, + 710, + 358, + 720 + ], + "score": 0.7, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ": (1) the bag-of-words representation", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 718, + 507, + 734 + ], + "spans": [ + { + "bbox": [ + 104, + 718, + 155, + 734 + ], + "score": 1.0, + "content": "denoted by", + "type": "text" + }, + { + "bbox": [ + 155, + 720, + 209, + 732 + ], + "score": 0.92, + "content": "x _ { \\mathrm { B o W } } ~ \\in ~ \\mathbb { R } ^ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 718, + 352, + 734 + ], + "score": 1.0, + "content": "and (2) a sentence representation", + "type": "text" + }, + { + "bbox": [ + 352, + 720, + 462, + 732 + ], + "score": 0.92, + "content": "x _ { \\mathrm { s e n t } } = f ( \\breve { T } ; \\theta _ { \\mathrm { s e n t } } ) \\in \\mathbb { R } ^ { \\breve { V } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 718, + 507, + 734 + ], + "score": 1.0, + "content": ", which is", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 252, + 95 + ], + "score": 1.0, + "content": "computed by passing the words in", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 252, + 83, + 261, + 92 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 261, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "through a Gated Recurrent Unit (GRU) (Cho et al., 2014)", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 283, + 107 + ], + "score": 1.0, + "content": "network followed by a linear layer. Here,", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 283, + 94, + 301, + 105 + ], + "score": 0.9, + "content": "\\theta _ { \\mathrm { s e n t } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 302, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "denotes the parameters of the GRU network and", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 401, + 118 + ], + "score": 1.0, + "content": "the linear layer with ReLU activations. Next, the Dual-Attention unit", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 401, + 105, + 417, + 116 + ], + "score": 0.88, + "content": "f _ { \\mathrm { D A } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 418, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "combines the image", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 443, + 129 + ], + "score": 1.0, + "content": "representation with the text representations to get the complete state representation", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 444, + 117, + 456, + 127 + ], + "score": 0.85, + "content": "x _ { S }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 456, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "and answer", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 173, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 149, + 141 + ], + "score": 1.0, + "content": "prediction", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 149, + 128, + 168, + 138 + ], + "score": 0.87, + "content": "x _ { \\mathrm { A n s } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 168, + 125, + 173, + 141 + ], + "score": 1.0, + "content": ":", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 43, + "bbox_fs": [ + 103, + 632, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 252, + 95 + ], + "score": 1.0, + "content": "computed by passing the words in", + "type": "text" + }, + { + "bbox": [ + 252, + 83, + 261, + 92 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "through a Gated Recurrent Unit (GRU) (Cho et al., 2014)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 283, + 107 + ], + "score": 1.0, + "content": "network followed by a linear layer. Here,", + "type": "text" + }, + { + "bbox": [ + 283, + 94, + 301, + 105 + ], + "score": 0.9, + "content": "\\theta _ { \\mathrm { s e n t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "denotes the parameters of the GRU network and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 401, + 118 + ], + "score": 1.0, + "content": "the linear layer with ReLU activations. Next, the Dual-Attention unit", + "type": "text" + }, + { + "bbox": [ + 401, + 105, + 417, + 116 + ], + "score": 0.88, + "content": "f _ { \\mathrm { D A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 103, + 506, + 118 + ], + "score": 1.0, + "content": "combines the image", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 443, + 129 + ], + "score": 1.0, + "content": "representation with the text representations to get the complete state representation", + "type": "text" + }, + { + "bbox": [ + 444, + 117, + 456, + 127 + ], + "score": 0.85, + "content": "x _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "and answer", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 173, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 149, + 141 + ], + "score": 1.0, + "content": "prediction", + "type": "text" + }, + { + "bbox": [ + 149, + 128, + 168, + 138 + ], + "score": 0.87, + "content": "x _ { \\mathrm { A n s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 125, + 173, + 141 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 136, + 369, + 150 + ], + "lines": [ + { + "bbox": [ + 241, + 136, + 369, + 150 + ], + "spans": [ + { + "bbox": [ + 241, + 136, + 369, + 150 + ], + "score": 0.93, + "content": "x _ { \\mathrm { S } } , x _ { \\mathrm { A n s } } = f _ { \\mathrm { D A } } ( x _ { I } , x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } )", + "type": "interline_equation", + "image_path": "9a477078ce9fc48266be90f6b6df84b63c02c11e6ff0d0820fb6ebcaaa177a0a.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 241, + 136, + 369, + 150 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 151, + 504, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 139, + 164 + ], + "score": 1.0, + "content": "Finally,", + "type": "text" + }, + { + "bbox": [ + 140, + 154, + 153, + 163 + ], + "score": 0.84, + "content": "x _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 151, + 171, + 164 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 172, + 153, + 190, + 163 + ], + "score": 0.9, + "content": "x _ { \\mathrm { A n s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 151, + 505, + 164 + ], + "score": 1.0, + "content": ", along with a time step embedding and a task indicator variable (for whether", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 162, + 425, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 425, + 175 + ], + "score": 1.0, + "content": "the task is SGN or EQA), are passed to the policy module to produce an action.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 187, + 236, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 237, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 237, + 200 + ], + "score": 1.0, + "content": "4.1 DUAL-ATTENTION UNIT", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 504, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 504, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 436, + 220 + ], + "score": 1.0, + "content": "The Dual-Attention unit uses two types of attention mechanisms, Gated-Attention", + "type": "text" + }, + { + "bbox": [ + 437, + 208, + 453, + 219 + ], + "score": 0.89, + "content": "f _ { \\mathrm { G A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 207, + 504, + 220 + ], + "score": 1.0, + "content": "and Spatial-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 219, + 394, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 146, + 231 + ], + "score": 1.0, + "content": "Attention", + "type": "text" + }, + { + "bbox": [ + 147, + 219, + 162, + 231 + ], + "score": 0.87, + "content": "f _ { \\mathrm { S A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 219, + 394, + 231 + ], + "score": 1.0, + "content": ", to align representations in different modalities and tasks.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 336, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 337, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 337, + 248 + ], + "score": 1.0, + "content": "Gated-Attention. The Gated-Attention unit (Figure 3)", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 247, + 337, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 337, + 258 + ], + "score": 1.0, + "content": "was proposed in (Chaplot et al., 2017) for multimodal fu-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 257, + 336, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 336, + 269 + ], + "score": 1.0, + "content": "sion. Intuitively, a GA unit attends to the different chan-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 268, + 336, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 336, + 281 + ], + "score": 1.0, + "content": "nels in the image representation based on the text repre-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 280, + 336, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 336, + 291 + ], + "score": 1.0, + "content": "sentation. For example, if the textual input is the instruc-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 290, + 337, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 337, + 302 + ], + "score": 1.0, + "content": "tion ‘Go to the red pillar’, then the GA unit can learn to at-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 300, + 337, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 337, + 313 + ], + "score": 1.0, + "content": "tend to channels which detect red things and pillars. More", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 313, + 337, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 337, + 324 + ], + "score": 1.0, + "content": "specifically, the GA unit takes as input a 3-dimensional", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 321, + 339, + 336 + ], + "spans": [ + { + "bbox": [ + 104, + 321, + 223, + 336 + ], + "score": 1.0, + "content": "tensor image representation", + "type": "text" + }, + { + "bbox": [ + 223, + 322, + 290, + 335 + ], + "score": 0.92, + "content": "y _ { I } \\in \\mathring { \\mathbb { R } } ^ { d \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 321, + 339, + 336 + ], + "score": 1.0, + "content": "and a text", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 333, + 338, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 165, + 347 + ], + "score": 1.0, + "content": "representation", + "type": "text" + }, + { + "bbox": [ + 165, + 334, + 201, + 345 + ], + "score": 0.91, + "content": "\\bar { y _ { T } } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 333, + 338, + 347 + ], + "score": 1.0, + "content": ", and outputs a 3-dimensional ten-", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 16.0 + }, + { + "type": "image", + "bbox": [ + 342, + 249, + 495, + 315 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 342, + 249, + 495, + 315 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 342, + 249, + 495, + 315 + ], + "spans": [ + { + "bbox": [ + 342, + 249, + 495, + 315 + ], + "score": 0.954, + "type": "image", + "image_path": "888f192624493aa788242aabe8918602f25c59af6c7afb47da6321d3f27acf6d.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 342, + 249, + 495, + 282.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 342, + 282.0, + 495, + 315.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 358, + 325, + 488, + 335 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 357, + 322, + 487, + 337 + ], + "spans": [ + { + "bbox": [ + 357, + 322, + 473, + 337 + ], + "score": 1.0, + "content": "Figure 3: Gated-Attention unit", + "type": "text" + }, + { + "bbox": [ + 473, + 325, + 487, + 335 + ], + "score": 0.29, + "content": "f _ { \\mathrm { G A } }", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 22 + } + ], + "index": 19.75 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 401 + ], + "lines": [ + { + "bbox": [ + 104, + 341, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 104, + 341, + 122, + 359 + ], + "score": 1.0, + "content": "sor", + "type": "text" + }, + { + "bbox": [ + 122, + 345, + 181, + 356 + ], + "score": 0.92, + "content": "{ \\boldsymbol { z } } \\in \\mathbb { R } ^ { d \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 341, + 299, + 359 + ], + "score": 1.0, + "content": ". Note that the dimension of", + "type": "text" + }, + { + "bbox": [ + 299, + 347, + 312, + 357 + ], + "score": 0.85, + "content": "y _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 341, + 506, + 359 + ], + "score": 1.0, + "content": "is equal to the number of feature maps and the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 355, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 225, + 369 + ], + "score": 1.0, + "content": "size of the first dimension of", + "type": "text" + }, + { + "bbox": [ + 225, + 358, + 236, + 367 + ], + "score": 0.82, + "content": "y _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 355, + 423, + 369 + ], + "score": 1.0, + "content": ". In the Gated-Attention unit, each element of", + "type": "text" + }, + { + "bbox": [ + 423, + 358, + 435, + 367 + ], + "score": 0.85, + "content": "y _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 355, + 506, + 369 + ], + "score": 1.0, + "content": "is expanded to a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 365, + 507, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 140, + 378 + ], + "score": 0.91, + "content": "H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 365, + 315, + 381 + ], + "score": 1.0, + "content": "matrix, resulting in a 3-dimensional tensor", + "type": "text" + }, + { + "bbox": [ + 315, + 366, + 389, + 379 + ], + "score": 0.93, + "content": "M _ { y _ { T } } \\in \\mathbb { R } ^ { d \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 365, + 421, + 381 + ], + "score": 1.0, + "content": ", whose", + "type": "text" + }, + { + "bbox": [ + 422, + 367, + 460, + 379 + ], + "score": 0.93, + "content": "( i , j , k ) ^ { t \\bar { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 365, + 507, + 381 + ], + "score": 1.0, + "content": "element is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 377, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 142, + 391 + ], + "score": 1.0, + "content": "given by", + "type": "text" + }, + { + "bbox": [ + 142, + 378, + 224, + 390 + ], + "score": 0.91, + "content": "M _ { y _ { T } } [ i , j , k ] = \\bar { y } _ { T } [ i ]", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 377, + 505, + 391 + ], + "score": 1.0, + "content": ". This matrix is multiplied element-wise with the image representation:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 389, + 451, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 229, + 402 + ], + "score": 0.89, + "content": "z = f _ { \\mathrm { G A } } ( y _ { I } , y _ { T } ) = M _ { y _ { T } } \\odot y _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 389, + 259, + 402 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 259, + 390, + 269, + 399 + ], + "score": 0.84, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 389, + 451, + 402 + ], + "score": 1.0, + "content": "denotes the Hadamard product (Horn, 1990).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 336, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 337, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 337, + 417 + ], + "score": 1.0, + "content": "Spatial-Attention. We propose a Spatial-Attention unit", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 417, + 337, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 337, + 428 + ], + "score": 1.0, + "content": "(Figure 4) which is analogous to the Gated-Attention unit", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 427, + 337, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 337, + 441 + ], + "score": 1.0, + "content": "except that it attends to different pixels in the image repre-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 439, + 336, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 336, + 450 + ], + "score": 1.0, + "content": "sentation rather than the channels. For example, if the tex-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 450, + 336, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 336, + 461 + ], + "score": 1.0, + "content": "tual input is the question ‘Which object is blue in color?’,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 460, + 337, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 337, + 473 + ], + "score": 1.0, + "content": "then we would like to spatially attend to the parts of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 471, + 337, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 337, + 483 + ], + "score": 1.0, + "content": "image which contain a blue object in order recognize", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 483, + 337, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 337, + 495 + ], + "score": 1.0, + "content": "the type of the blue object. The Spatial-Attention unit", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 493, + 337, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 337, + 506 + ], + "score": 1.0, + "content": "takes as input a 3-dimensional tensor image representa-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 501, + 339, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 124, + 518 + ], + "score": 1.0, + "content": "tion", + "type": "text" + }, + { + "bbox": [ + 125, + 504, + 187, + 515 + ], + "score": 0.84, + "content": "y _ { I } \\in \\dot { \\mathbb { R } ^ { d \\times H \\times W } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 501, + 339, + 518 + ], + "score": 1.0, + "content": "and a 2-dimensional spatial attention", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 33.5 + }, + { + "type": "image", + "bbox": [ + 341, + 419, + 495, + 484 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 341, + 419, + 495, + 484 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 341, + 419, + 495, + 484 + ], + "spans": [ + { + "bbox": [ + 341, + 419, + 495, + 484 + ], + "score": 0.959, + "type": "image", + "image_path": "e7a785f4ad73bbe78d600384b1b5f88adcfc6098742f7d2bb686856c9e974c72.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 341, + 419, + 495, + 432.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 341, + 432.0, + 495, + 445.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 341, + 445.0, + 495, + 458.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 341, + 458.0, + 495, + 471.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 341, + 471.0, + 495, + 484.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 357, + 494, + 490, + 505 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 355, + 491, + 489, + 507 + ], + "spans": [ + { + "bbox": [ + 355, + 491, + 475, + 507 + ], + "score": 1.0, + "content": "Figure 4: Spatial-Attention unit", + "type": "text" + }, + { + "bbox": [ + 475, + 494, + 489, + 505 + ], + "score": 0.29, + "content": "f _ { \\mathrm { S A } }", + "type": "inline_equation" + } + ], + "index": 43 + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 103, + 512, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 103, + 512, + 126, + 529 + ], + "score": 1.0, + "content": "map", + "type": "text" + }, + { + "bbox": [ + 126, + 515, + 185, + 527 + ], + "score": 0.87, + "content": "\\mathit { y } _ { S } \\in \\mathbb { R } ^ { H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 512, + 277, + 529 + ], + "score": 1.0, + "content": ", and outputs a tensor", + "type": "text" + }, + { + "bbox": [ + 277, + 515, + 340, + 526 + ], + "score": 0.92, + "content": "\\mathit { \\check { z } } \\in \\mathbb { R } ^ { d \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 512, + 507, + 529 + ], + "score": 1.0, + "content": ". Note that the height and width of the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 525, + 506, + 540 + ], + "score": 1.0, + "content": "spatial attention map is equal to the height and width of the image representation. In the spatial-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 538, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 396, + 549 + ], + "score": 1.0, + "content": "attention unit, each element of the spatial attention map is expanded to a", + "type": "text" + }, + { + "bbox": [ + 396, + 538, + 403, + 547 + ], + "score": 0.73, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 538, + 505, + 549 + ], + "score": 1.0, + "content": "dimensional vector. This", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 545, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 268, + 563 + ], + "score": 1.0, + "content": "again results in a 3-dimensional tensor", + "type": "text" + }, + { + "bbox": [ + 268, + 547, + 344, + 560 + ], + "score": 0.92, + "content": "M _ { y _ { S } } \\in \\mathbb { R } ^ { d \\times \\hat { H } \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 545, + 378, + 563 + ], + "score": 1.0, + "content": ", whose", + "type": "text" + }, + { + "bbox": [ + 379, + 548, + 417, + 561 + ], + "score": 0.93, + "content": "( i , j , k ) ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 545, + 506, + 563 + ], + "score": 1.0, + "content": "element is given by:", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 107, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 107, + 559, + 198, + 572 + ], + "score": 0.92, + "content": "\\bar { M } _ { y s } [ i , j , k ] = y _ { S } [ j , k ]", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 559, + 506, + 573 + ], + "score": 1.0, + "content": ". Just like in the Gated-Attention unit, this matrix is multiplied element-wise", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 569, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 231, + 584 + ], + "score": 1.0, + "content": "with the image representation:", + "type": "text" + }, + { + "bbox": [ + 231, + 570, + 351, + 583 + ], + "score": 0.91, + "content": "z = f _ { \\mathrm { S A } } ( y _ { I } , y _ { S } ) = M _ { y _ { S } } \\odot y _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 569, + 506, + 584 + ], + "score": 1.0, + "content": ". Similar spatial attention mechanisms", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "have been used for Visual Question Answering (Fukui et al., 2016; Xu & Saenko, 2016; Hudson &", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 592, + 455, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 455, + 605 + ], + "score": 1.0, + "content": "Manning, 2018; Gupta et al., 2017b) and grounding audio in vision (Zhao et al., 2018).", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 48.5 + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 622 + ], + "score": 1.0, + "content": "Dual-Attention. We now describe the operations in the Dual-Attention unit shown in Figure 5, as", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 620, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 347, + 633 + ], + "score": 1.0, + "content": "well as motivate the intuitions behind each operation. Given", + "type": "text" + }, + { + "bbox": [ + 348, + 622, + 359, + 631 + ], + "score": 0.6, + "content": "x _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 620, + 362, + 633 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 363, + 622, + 385, + 632 + ], + "score": 0.6, + "content": "x _ { \\mathrm { B o W } }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 620, + 405, + 633 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 405, + 622, + 424, + 632 + ], + "score": 0.88, + "content": "x _ { \\mathrm { s e n t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 620, + 505, + 633 + ], + "score": 1.0, + "content": ", the Dual-Attention", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 631, + 340, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 277, + 644 + ], + "score": 1.0, + "content": "unit first computes a Gated-Attention over", + "type": "text" + }, + { + "bbox": [ + 277, + 633, + 289, + 642 + ], + "score": 0.85, + "content": "x _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 631, + 314, + 644 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 314, + 633, + 336, + 642 + ], + "score": 0.83, + "content": "x _ { \\mathrm { B o W } }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 631, + 340, + 644 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 646, + 381, + 662 + ], + "lines": [ + { + "bbox": [ + 230, + 646, + 381, + 662 + ], + "spans": [ + { + "bbox": [ + 230, + 646, + 381, + 662 + ], + "score": 0.92, + "content": "x _ { \\mathrm { G A l } } = f _ { \\mathrm { G A } } ( x _ { I } , x _ { \\mathrm { B o W } } ) \\in \\mathbb { R } ^ { V \\times H \\times W } .", + "type": "interline_equation", + "image_path": "53dea56cfe2785c0aacf9210d9b50cfff6325aeac3db0508672dd32f789acce9.jpg" + } + ] + } + ], + "index": 56, + "virtual_lines": [ + { + "bbox": [ + 230, + 646, + 381, + 662 + ], + "spans": [], + "index": 56 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 663, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 506, + 680 + ], + "score": 1.0, + "content": "Intuitively, this first Gated-Attention unit associates each word in the vocabulary with a feature map", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "score": 1.0, + "content": "in the image representation. A particular feature map is activated if and only if the corresponding", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "score": 1.0, + "content": "word occurs in the textual input. In other words, the feature maps in the convolutional output learns", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "to detect different objects and attributes, and words in the textual input specify which objects and", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "attributes are relevant to the current task. The Gated-Attention using BoW representation attends to", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "feature maps detecting corresponding objects and attributes, and masks all other feature maps. We", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 59.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 137 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 104, + 83, + 506, + 141 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 136, + 369, + 150 + ], + "lines": [ + { + "bbox": [ + 241, + 136, + 369, + 150 + ], + "spans": [ + { + "bbox": [ + 241, + 136, + 369, + 150 + ], + "score": 0.93, + "content": "x _ { \\mathrm { S } } , x _ { \\mathrm { A n s } } = f _ { \\mathrm { D A } } ( x _ { I } , x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } )", + "type": "interline_equation", + "image_path": "9a477078ce9fc48266be90f6b6df84b63c02c11e6ff0d0820fb6ebcaaa177a0a.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 241, + 136, + 369, + 150 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 151, + 504, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 139, + 164 + ], + "score": 1.0, + "content": "Finally,", + "type": "text" + }, + { + "bbox": [ + 140, + 154, + 153, + 163 + ], + "score": 0.84, + "content": "x _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 151, + 171, + 164 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 172, + 153, + 190, + 163 + ], + "score": 0.9, + "content": "x _ { \\mathrm { A n s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 151, + 505, + 164 + ], + "score": 1.0, + "content": ", along with a time step embedding and a task indicator variable (for whether", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 162, + 425, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 425, + 175 + ], + "score": 1.0, + "content": "the task is SGN or EQA), are passed to the policy module to produce an action.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 151, + 505, + 175 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 187, + 236, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 237, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 237, + 200 + ], + "score": 1.0, + "content": "4.1 DUAL-ATTENTION UNIT", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 504, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 504, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 436, + 220 + ], + "score": 1.0, + "content": "The Dual-Attention unit uses two types of attention mechanisms, Gated-Attention", + "type": "text" + }, + { + "bbox": [ + 437, + 208, + 453, + 219 + ], + "score": 0.89, + "content": "f _ { \\mathrm { G A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 207, + 504, + 220 + ], + "score": 1.0, + "content": "and Spatial-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 219, + 394, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 146, + 231 + ], + "score": 1.0, + "content": "Attention", + "type": "text" + }, + { + "bbox": [ + 147, + 219, + 162, + 231 + ], + "score": 0.87, + "content": "f _ { \\mathrm { S A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 219, + 394, + 231 + ], + "score": 1.0, + "content": ", to align representations in different modalities and tasks.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 207, + 504, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 336, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 337, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 337, + 248 + ], + "score": 1.0, + "content": "Gated-Attention. The Gated-Attention unit (Figure 3)", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 247, + 337, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 337, + 258 + ], + "score": 1.0, + "content": "was proposed in (Chaplot et al., 2017) for multimodal fu-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 257, + 336, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 336, + 269 + ], + "score": 1.0, + "content": "sion. Intuitively, a GA unit attends to the different chan-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 268, + 336, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 336, + 281 + ], + "score": 1.0, + "content": "nels in the image representation based on the text repre-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 280, + 336, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 336, + 291 + ], + "score": 1.0, + "content": "sentation. For example, if the textual input is the instruc-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 290, + 337, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 337, + 302 + ], + "score": 1.0, + "content": "tion ‘Go to the red pillar’, then the GA unit can learn to at-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 300, + 337, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 337, + 313 + ], + "score": 1.0, + "content": "tend to channels which detect red things and pillars. More", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 313, + 337, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 337, + 324 + ], + "score": 1.0, + "content": "specifically, the GA unit takes as input a 3-dimensional", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 321, + 339, + 336 + ], + "spans": [ + { + "bbox": [ + 104, + 321, + 223, + 336 + ], + "score": 1.0, + "content": "tensor image representation", + "type": "text" + }, + { + "bbox": [ + 223, + 322, + 290, + 335 + ], + "score": 0.92, + "content": "y _ { I } \\in \\mathring { \\mathbb { R } } ^ { d \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 321, + 339, + 336 + ], + "score": 1.0, + "content": "and a text", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 333, + 338, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 165, + 347 + ], + "score": 1.0, + "content": "representation", + "type": "text" + }, + { + "bbox": [ + 165, + 334, + 201, + 345 + ], + "score": 0.91, + "content": "\\bar { y _ { T } } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 333, + 338, + 347 + ], + "score": 1.0, + "content": ", and outputs a 3-dimensional ten-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 341, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 104, + 341, + 122, + 359 + ], + "score": 1.0, + "content": "sor", + "type": "text" + }, + { + "bbox": [ + 122, + 345, + 181, + 356 + ], + "score": 0.92, + "content": "{ \\boldsymbol { z } } \\in \\mathbb { R } ^ { d \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 341, + 299, + 359 + ], + "score": 1.0, + "content": ". Note that the dimension of", + "type": "text" + }, + { + "bbox": [ + 299, + 347, + 312, + 357 + ], + "score": 0.85, + "content": "y _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 341, + 506, + 359 + ], + "score": 1.0, + "content": "is equal to the number of feature maps and the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 355, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 225, + 369 + ], + "score": 1.0, + "content": "size of the first dimension of", + "type": "text" + }, + { + "bbox": [ + 225, + 358, + 236, + 367 + ], + "score": 0.82, + "content": "y _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 355, + 423, + 369 + ], + "score": 1.0, + "content": ". In the Gated-Attention unit, each element of", + "type": "text" + }, + { + "bbox": [ + 423, + 358, + 435, + 367 + ], + "score": 0.85, + "content": "y _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 355, + 506, + 369 + ], + "score": 1.0, + "content": "is expanded to a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 365, + 507, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 140, + 378 + ], + "score": 0.91, + "content": "H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 365, + 315, + 381 + ], + "score": 1.0, + "content": "matrix, resulting in a 3-dimensional tensor", + "type": "text" + }, + { + "bbox": [ + 315, + 366, + 389, + 379 + ], + "score": 0.93, + "content": "M _ { y _ { T } } \\in \\mathbb { R } ^ { d \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 365, + 421, + 381 + ], + "score": 1.0, + "content": ", whose", + "type": "text" + }, + { + "bbox": [ + 422, + 367, + 460, + 379 + ], + "score": 0.93, + "content": "( i , j , k ) ^ { t \\bar { h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 365, + 507, + 381 + ], + "score": 1.0, + "content": "element is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 377, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 142, + 391 + ], + "score": 1.0, + "content": "given by", + "type": "text" + }, + { + "bbox": [ + 142, + 378, + 224, + 390 + ], + "score": 0.91, + "content": "M _ { y _ { T } } [ i , j , k ] = \\bar { y } _ { T } [ i ]", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 377, + 505, + 391 + ], + "score": 1.0, + "content": ". This matrix is multiplied element-wise with the image representation:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 389, + 451, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 229, + 402 + ], + "score": 0.89, + "content": "z = f _ { \\mathrm { G A } } ( y _ { I } , y _ { T } ) = M _ { y _ { T } } \\odot y _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 389, + 259, + 402 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 259, + 390, + 269, + 399 + ], + "score": 0.84, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 389, + 451, + 402 + ], + "score": 1.0, + "content": "denotes the Hadamard product (Horn, 1990).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 16.0, + "bbox_fs": [ + 104, + 234, + 339, + 347 + ] + }, + { + "type": "image", + "bbox": [ + 342, + 249, + 495, + 315 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 342, + 249, + 495, + 315 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 342, + 249, + 495, + 315 + ], + "spans": [ + { + "bbox": [ + 342, + 249, + 495, + 315 + ], + "score": 0.954, + "type": "image", + "image_path": "888f192624493aa788242aabe8918602f25c59af6c7afb47da6321d3f27acf6d.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 342, + 249, + 495, + 282.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 342, + 282.0, + 495, + 315.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 358, + 325, + 488, + 335 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 357, + 322, + 487, + 337 + ], + "spans": [ + { + "bbox": [ + 357, + 322, + 473, + 337 + ], + "score": 1.0, + "content": "Figure 3: Gated-Attention unit", + "type": "text" + }, + { + "bbox": [ + 473, + 325, + 487, + 335 + ], + "score": 0.29, + "content": "f _ { \\mathrm { G A } }", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 22 + } + ], + "index": 19.75 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 401 + ], + "lines": [], + "index": 26, + "bbox_fs": [ + 104, + 341, + 507, + 402 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 336, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 406, + 337, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 337, + 417 + ], + "score": 1.0, + "content": "Spatial-Attention. We propose a Spatial-Attention unit", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 417, + 337, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 337, + 428 + ], + "score": 1.0, + "content": "(Figure 4) which is analogous to the Gated-Attention unit", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 427, + 337, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 337, + 441 + ], + "score": 1.0, + "content": "except that it attends to different pixels in the image repre-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 439, + 336, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 336, + 450 + ], + "score": 1.0, + "content": "sentation rather than the channels. For example, if the tex-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 450, + 336, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 336, + 461 + ], + "score": 1.0, + "content": "tual input is the question ‘Which object is blue in color?’,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 460, + 337, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 337, + 473 + ], + "score": 1.0, + "content": "then we would like to spatially attend to the parts of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 471, + 337, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 337, + 483 + ], + "score": 1.0, + "content": "image which contain a blue object in order recognize", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 483, + 337, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 337, + 495 + ], + "score": 1.0, + "content": "the type of the blue object. The Spatial-Attention unit", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 493, + 337, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 337, + 506 + ], + "score": 1.0, + "content": "takes as input a 3-dimensional tensor image representa-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 501, + 339, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 124, + 518 + ], + "score": 1.0, + "content": "tion", + "type": "text" + }, + { + "bbox": [ + 125, + 504, + 187, + 515 + ], + "score": 0.84, + "content": "y _ { I } \\in \\dot { \\mathbb { R } ^ { d \\times H \\times W } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 501, + 339, + 518 + ], + "score": 1.0, + "content": "and a 2-dimensional spatial attention", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 103, + 512, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 103, + 512, + 126, + 529 + ], + "score": 1.0, + "content": "map", + "type": "text" + }, + { + "bbox": [ + 126, + 515, + 185, + 527 + ], + "score": 0.87, + "content": "\\mathit { y } _ { S } \\in \\mathbb { R } ^ { H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 512, + 277, + 529 + ], + "score": 1.0, + "content": ", and outputs a tensor", + "type": "text" + }, + { + "bbox": [ + 277, + 515, + 340, + 526 + ], + "score": 0.92, + "content": "\\mathit { \\check { z } } \\in \\mathbb { R } ^ { d \\times H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 512, + 507, + 529 + ], + "score": 1.0, + "content": ". Note that the height and width of the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 104, + 525, + 506, + 540 + ], + "score": 1.0, + "content": "spatial attention map is equal to the height and width of the image representation. In the spatial-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 538, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 396, + 549 + ], + "score": 1.0, + "content": "attention unit, each element of the spatial attention map is expanded to a", + "type": "text" + }, + { + "bbox": [ + 396, + 538, + 403, + 547 + ], + "score": 0.73, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 538, + 505, + 549 + ], + "score": 1.0, + "content": "dimensional vector. This", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 545, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 268, + 563 + ], + "score": 1.0, + "content": "again results in a 3-dimensional tensor", + "type": "text" + }, + { + "bbox": [ + 268, + 547, + 344, + 560 + ], + "score": 0.92, + "content": "M _ { y _ { S } } \\in \\mathbb { R } ^ { d \\times \\hat { H } \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 545, + 378, + 563 + ], + "score": 1.0, + "content": ", whose", + "type": "text" + }, + { + "bbox": [ + 379, + 548, + 417, + 561 + ], + "score": 0.93, + "content": "( i , j , k ) ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 545, + 506, + 563 + ], + "score": 1.0, + "content": "element is given by:", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 107, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 107, + 559, + 198, + 572 + ], + "score": 0.92, + "content": "\\bar { M } _ { y s } [ i , j , k ] = y _ { S } [ j , k ]", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 559, + 506, + 573 + ], + "score": 1.0, + "content": ". Just like in the Gated-Attention unit, this matrix is multiplied element-wise", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 569, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 231, + 584 + ], + "score": 1.0, + "content": "with the image representation:", + "type": "text" + }, + { + "bbox": [ + 231, + 570, + 351, + 583 + ], + "score": 0.91, + "content": "z = f _ { \\mathrm { S A } } ( y _ { I } , y _ { S } ) = M _ { y _ { S } } \\odot y _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 569, + 506, + 584 + ], + "score": 1.0, + "content": ". Similar spatial attention mechanisms", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "have been used for Visual Question Answering (Fukui et al., 2016; Xu & Saenko, 2016; Hudson &", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 592, + 455, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 455, + 605 + ], + "score": 1.0, + "content": "Manning, 2018; Gupta et al., 2017b) and grounding audio in vision (Zhao et al., 2018).", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 406, + 339, + 518 + ] + }, + { + "type": "image", + "bbox": [ + 341, + 419, + 495, + 484 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 341, + 419, + 495, + 484 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 341, + 419, + 495, + 484 + ], + "spans": [ + { + "bbox": [ + 341, + 419, + 495, + 484 + ], + "score": 0.959, + "type": "image", + "image_path": "e7a785f4ad73bbe78d600384b1b5f88adcfc6098742f7d2bb686856c9e974c72.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 341, + 419, + 495, + 432.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 341, + 432.0, + 495, + 445.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 341, + 445.0, + 495, + 458.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 341, + 458.0, + 495, + 471.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 341, + 471.0, + 495, + 484.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 357, + 494, + 490, + 505 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 355, + 491, + 489, + 507 + ], + "spans": [ + { + "bbox": [ + 355, + 491, + 475, + 507 + ], + "score": 1.0, + "content": "Figure 4: Spatial-Attention unit", + "type": "text" + }, + { + "bbox": [ + 475, + 494, + 489, + 505 + ], + "score": 0.29, + "content": "f _ { \\mathrm { S A } }", + "type": "inline_equation" + } + ], + "index": 43 + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 505, + 604 + ], + "lines": [], + "index": 48.5, + "bbox_fs": [ + 103, + 512, + 507, + 605 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 622 + ], + "score": 1.0, + "content": "Dual-Attention. We now describe the operations in the Dual-Attention unit shown in Figure 5, as", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 620, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 347, + 633 + ], + "score": 1.0, + "content": "well as motivate the intuitions behind each operation. Given", + "type": "text" + }, + { + "bbox": [ + 348, + 622, + 359, + 631 + ], + "score": 0.6, + "content": "x _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 620, + 362, + 633 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 363, + 622, + 385, + 632 + ], + "score": 0.6, + "content": "x _ { \\mathrm { B o W } }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 620, + 405, + 633 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 405, + 622, + 424, + 632 + ], + "score": 0.88, + "content": "x _ { \\mathrm { s e n t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 620, + 505, + 633 + ], + "score": 1.0, + "content": ", the Dual-Attention", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 631, + 340, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 277, + 644 + ], + "score": 1.0, + "content": "unit first computes a Gated-Attention over", + "type": "text" + }, + { + "bbox": [ + 277, + 633, + 289, + 642 + ], + "score": 0.85, + "content": "x _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 631, + 314, + 644 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 314, + 633, + 336, + 642 + ], + "score": 0.83, + "content": "x _ { \\mathrm { B o W } }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 631, + 340, + 644 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54, + "bbox_fs": [ + 105, + 608, + 505, + 644 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 646, + 381, + 662 + ], + "lines": [ + { + "bbox": [ + 230, + 646, + 381, + 662 + ], + "spans": [ + { + "bbox": [ + 230, + 646, + 381, + 662 + ], + "score": 0.92, + "content": "x _ { \\mathrm { G A l } } = f _ { \\mathrm { G A } } ( x _ { I } , x _ { \\mathrm { B o W } } ) \\in \\mathbb { R } ^ { V \\times H \\times W } .", + "type": "interline_equation", + "image_path": "53dea56cfe2785c0aacf9210d9b50cfff6325aeac3db0508672dd32f789acce9.jpg" + } + ] + } + ], + "index": 56, + "virtual_lines": [ + { + "bbox": [ + 230, + 646, + 381, + 662 + ], + "spans": [], + "index": 56 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 663, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 506, + 680 + ], + "score": 1.0, + "content": "Intuitively, this first Gated-Attention unit associates each word in the vocabulary with a feature map", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "score": 1.0, + "content": "in the image representation. A particular feature map is activated if and only if the corresponding", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "score": 1.0, + "content": "word occurs in the textual input. In other words, the feature maps in the convolutional output learns", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "to detect different objects and attributes, and words in the textual input specify which objects and", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "attributes are relevant to the current task. The Gated-Attention using BoW representation attends to", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "feature maps detecting corresponding objects and attributes, and masks all other feature maps. We", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "use bag-of-words representation for the first GA unit as it explicitly aligns the words in textual input", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "irrespective of whether it is a question or an instruction. Note that bag-of-words representation has", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 282, + 497, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 497, + 294 + ], + "score": 1.0, + "content": "been used previously in models trained for learning to follow instructions (Hermann et al., 2017).", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 59.5, + "bbox_fs": [ + 105, + 663, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 78, + 498, + 214 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 78, + 498, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 78, + 498, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 498, + 214 + ], + "score": 0.966, + "type": "image", + "image_path": "a7e5f5b2f4bacd0b9ce717a508e7d79925d868fa21a8c187b4d6a12dd64fc350.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 78, + 498, + 123.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 123.33333333333334, + 498, + 168.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 168.66666666666669, + 498, + 214.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 229, + 502, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 228, + 503, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 503, + 243 + ], + "score": 1.0, + "content": "Figure 5: Architecture of the Dual-Attention unit with example intermediate representations and operations.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 108, + 259, + 503, + 294 + ], + "lines": [ + { + "bbox": [ + 106, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "use bag-of-words representation for the first GA unit as it explicitly aligns the words in textual input", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "irrespective of whether it is a question or an instruction. Note that bag-of-words representation has", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 282, + 497, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 497, + 294 + ], + "score": 1.0, + "content": "been used previously in models trained for learning to follow instructions (Hermann et al., 2017).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 105, + 299, + 504, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 281, + 313 + ], + "score": 1.0, + "content": "Next, the output of the Gated-Attention unit", + "type": "text" + }, + { + "bbox": [ + 281, + 301, + 302, + 311 + ], + "score": 0.87, + "content": "x _ { \\mathrm { G A 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 297, + 505, + 313 + ], + "score": 1.0, + "content": "is converted to a spatial attention map by summing", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 310, + 363, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 288, + 323 + ], + "score": 1.0, + "content": "over all channels followed by a softmax over", + "type": "text" + }, + { + "bbox": [ + 288, + 311, + 321, + 321 + ], + "score": 0.89, + "content": "H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 310, + 363, + 323 + ], + "score": 1.0, + "content": "elements:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 325, + 384, + 360 + ], + "lines": [ + { + "bbox": [ + 226, + 325, + 384, + 360 + ], + "spans": [ + { + "bbox": [ + 226, + 325, + 384, + 360 + ], + "score": 0.96, + "content": "x _ { \\mathrm { s p a t } } = \\sigma \\left( \\sum _ { i } ^ { V } { x _ { \\mathrm { G A l } } [ i , : , : ] } \\right) \\in \\mathbb { R } ^ { H \\times W }", + "type": "interline_equation", + "image_path": "5b5dfb779e213cccb5af97bd1b97bd2d98e2f7b15763d52e68a49d10cf9a4c26.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 226, + 325, + 384, + 342.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 226, + 342.5, + 384, + 360.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 362, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 186, + 378 + ], + "score": 1.0, + "content": "where the softmax", + "type": "text" + }, + { + "bbox": [ + 187, + 363, + 315, + 378 + ], + "score": 0.93, + "content": "\\sigma ( z ) _ { j } = \\exp ( z _ { j } ) / \\sum _ { j } \\exp ( z _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 362, + 505, + 378 + ], + "score": 1.0, + "content": "ensures that the attention map is normalized.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 376, + 504, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 166, + 388 + ], + "score": 1.0, + "content": "Summation of", + "type": "text" + }, + { + "bbox": [ + 166, + 378, + 187, + 387 + ], + "score": 0.87, + "content": "x _ { \\mathrm { G A l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 376, + 504, + 388 + ], + "score": 1.0, + "content": "along the depth dimension gives a spatial attention map which has high activa-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "tions at spatial locations where relevant objects or attributes are detected. ReLU activations in the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "convolutional feature maps makes all elements positive, ensuring that the summation aggregates the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 408, + 253, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 253, + 422 + ], + "score": 1.0, + "content": "activations of relevant feature maps.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 351, + 437 + ], + "lines": [ + { + "bbox": [ + 107, + 424, + 352, + 439 + ], + "spans": [ + { + "bbox": [ + 107, + 428, + 126, + 438 + ], + "score": 0.88, + "content": "x _ { \\mathrm { s p a t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 424, + 143, + 439 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 144, + 428, + 155, + 437 + ], + "score": 0.84, + "content": "x _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 424, + 352, + 439 + ], + "score": 1.0, + "content": "are then passed through a Spatial-Attention unit:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 441, + 375, + 457 + ], + "lines": [ + { + "bbox": [ + 235, + 441, + 375, + 457 + ], + "spans": [ + { + "bbox": [ + 235, + 441, + 375, + 457 + ], + "score": 0.93, + "content": "x _ { \\mathrm { S A } } = f _ { \\mathrm { S A } } ( x _ { I } , x _ { \\mathrm { s p a t } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }", + "type": "interline_equation", + "image_path": "0a1ce057a4d7916019cef3002f085eefb3d7328dab947a5cf18fd8d11f5597d4.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 235, + 441, + 375, + 457 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 460, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 473 + ], + "score": 1.0, + "content": "The Spatial-Attention unit outputs all attributes present at the locations where relevant objects and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 470, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 484 + ], + "score": 1.0, + "content": "attributes are detected. This is especially helpful for question answering, where a single Gated-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 481, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 506, + 495 + ], + "score": 1.0, + "content": "Attention may not be sufficient. For example, if the textual input is ‘Which color is the pillar?’, then", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "the model needs to attend not only to feature maps detecting pillars (done by the Gated-Attention),", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "but also to other attributes at the spatial locations where pillars are seen in order to predict their color.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "score": 1.0, + "content": "Note that a single Gated-Attention is sufficient for instruction following, as shown in (Chaplot et al.,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "2017). For example, if the textual input is ‘Go to the green pillar’, the first Gated-Attention unit can", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "learn to attend to feature maps detecting green objects and pillar, and learn a navigation policy based", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 548, + 324, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 324, + 561 + ], + "score": 1.0, + "content": "on the spatial locations of the feature map activations.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 564, + 504, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 123, + 576 + ], + "score": 0.86, + "content": "x _ { \\mathrm { S A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 563, + 505, + 578 + ], + "score": 1.0, + "content": "is then passed through another Gated-Attention unit with the sentence-level text representation:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 579, + 380, + 595 + ], + "lines": [ + { + "bbox": [ + 230, + 579, + 380, + 595 + ], + "spans": [ + { + "bbox": [ + 230, + 579, + 380, + 595 + ], + "score": 0.93, + "content": "x _ { \\mathrm { G A 2 } } = f _ { \\mathrm { G A } } ( x _ { \\mathrm { S A } } , x _ { \\mathrm { s e n t } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }", + "type": "interline_equation", + "image_path": "c2f422d394e5aa39fd0b01a9687570166e773a77be9f521371c2444b6252180e.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 230, + 579, + 380, + 595 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 505, + 665 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "This second Gated-Attention unit enables the model to attend to different types of attributes based", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "on the question. For instance, if the question is asking about the color (‘Which color is the pillar?’),", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "score": 1.0, + "content": "then the model needs to attend to the feature maps corresponding to colors; or if the question is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "asking about the object type (‘Which object is green in color?’), then the model needs to attend to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 642, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 399, + 655 + ], + "score": 1.0, + "content": "the feature maps corresponding to object types. The sentence embedding", + "type": "text" + }, + { + "bbox": [ + 399, + 644, + 418, + 654 + ], + "score": 0.88, + "content": "x _ { \\mathrm { s e n t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 642, + 506, + 655 + ], + "score": 1.0, + "content": "can learn to attend to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 654, + 358, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 358, + 665 + ], + "score": 1.0, + "content": "multiple channels based on the textual input and mask the rest.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 669, + 504, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "Next, the output is transformed to answer prediction by again doing a summation and softmax but", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 681, + 383, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 383, + 693 + ], + "score": 1.0, + "content": "this time summing over the height and width instead of the channels:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 696, + 381, + 736 + ], + "lines": [ + { + "bbox": [ + 230, + 696, + 381, + 736 + ], + "spans": [ + { + "bbox": [ + 230, + 696, + 381, + 736 + ], + "score": 0.95, + "content": "x _ { \\mathrm { A n s } } = \\sigma \\left( \\sum _ { j , k } ^ { H , W } x _ { \\mathrm { G A 2 } } [ : , j , k ] \\right) \\in \\mathbb { R } ^ { V }", + "type": "interline_equation", + "image_path": "0daedaacc108632c304ac35dd2039c7625699fb7bf18eea83f8178c53dfe6988.jpg" + } + ] + } + ], + "index": 37.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 696, + 381, + 716.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 230, + 716.0, + 381, + 736.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 78, + 498, + 214 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 78, + 498, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 78, + 498, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 498, + 214 + ], + "score": 0.966, + "type": "image", + "image_path": "a7e5f5b2f4bacd0b9ce717a508e7d79925d868fa21a8c187b4d6a12dd64fc350.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 78, + 498, + 123.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 123.33333333333334, + 498, + 168.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 168.66666666666669, + 498, + 214.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 229, + 502, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 228, + 503, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 503, + 243 + ], + "score": 1.0, + "content": "Figure 5: Architecture of the Dual-Attention unit with example intermediate representations and operations.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 108, + 259, + 503, + 294 + ], + "lines": [], + "index": 5, + "bbox_fs": [ + 106, + 260, + 505, + 294 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 105, + 299, + 504, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 281, + 313 + ], + "score": 1.0, + "content": "Next, the output of the Gated-Attention unit", + "type": "text" + }, + { + "bbox": [ + 281, + 301, + 302, + 311 + ], + "score": 0.87, + "content": "x _ { \\mathrm { G A 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 297, + 505, + 313 + ], + "score": 1.0, + "content": "is converted to a spatial attention map by summing", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 310, + 363, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 288, + 323 + ], + "score": 1.0, + "content": "over all channels followed by a softmax over", + "type": "text" + }, + { + "bbox": [ + 288, + 311, + 321, + 321 + ], + "score": 0.89, + "content": "H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 310, + 363, + 323 + ], + "score": 1.0, + "content": "elements:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 297, + 505, + 323 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 325, + 384, + 360 + ], + "lines": [ + { + "bbox": [ + 226, + 325, + 384, + 360 + ], + "spans": [ + { + "bbox": [ + 226, + 325, + 384, + 360 + ], + "score": 0.96, + "content": "x _ { \\mathrm { s p a t } } = \\sigma \\left( \\sum _ { i } ^ { V } { x _ { \\mathrm { G A l } } [ i , : , : ] } \\right) \\in \\mathbb { R } ^ { H \\times W }", + "type": "interline_equation", + "image_path": "5b5dfb779e213cccb5af97bd1b97bd2d98e2f7b15763d52e68a49d10cf9a4c26.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 226, + 325, + 384, + 342.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 226, + 342.5, + 384, + 360.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 362, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 186, + 378 + ], + "score": 1.0, + "content": "where the softmax", + "type": "text" + }, + { + "bbox": [ + 187, + 363, + 315, + 378 + ], + "score": 0.93, + "content": "\\sigma ( z ) _ { j } = \\exp ( z _ { j } ) / \\sum _ { j } \\exp ( z _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 362, + 505, + 378 + ], + "score": 1.0, + "content": "ensures that the attention map is normalized.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 376, + 504, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 166, + 388 + ], + "score": 1.0, + "content": "Summation of", + "type": "text" + }, + { + "bbox": [ + 166, + 378, + 187, + 387 + ], + "score": 0.87, + "content": "x _ { \\mathrm { G A l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 376, + 504, + 388 + ], + "score": 1.0, + "content": "along the depth dimension gives a spatial attention map which has high activa-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "tions at spatial locations where relevant objects or attributes are detected. ReLU activations in the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "convolutional feature maps makes all elements positive, ensuring that the summation aggregates the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 408, + 253, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 253, + 422 + ], + "score": 1.0, + "content": "activations of relevant feature maps.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 362, + 505, + 422 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 351, + 437 + ], + "lines": [ + { + "bbox": [ + 107, + 424, + 352, + 439 + ], + "spans": [ + { + "bbox": [ + 107, + 428, + 126, + 438 + ], + "score": 0.88, + "content": "x _ { \\mathrm { s p a t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 424, + 143, + 439 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 144, + 428, + 155, + 437 + ], + "score": 0.84, + "content": "x _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 424, + 352, + 439 + ], + "score": 1.0, + "content": "are then passed through a Spatial-Attention unit:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 107, + 424, + 352, + 439 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 441, + 375, + 457 + ], + "lines": [ + { + "bbox": [ + 235, + 441, + 375, + 457 + ], + "spans": [ + { + "bbox": [ + 235, + 441, + 375, + 457 + ], + "score": 0.93, + "content": "x _ { \\mathrm { S A } } = f _ { \\mathrm { S A } } ( x _ { I } , x _ { \\mathrm { s p a t } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }", + "type": "interline_equation", + "image_path": "0a1ce057a4d7916019cef3002f085eefb3d7328dab947a5cf18fd8d11f5597d4.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 235, + 441, + 375, + 457 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 460, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 473 + ], + "score": 1.0, + "content": "The Spatial-Attention unit outputs all attributes present at the locations where relevant objects and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 470, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 484 + ], + "score": 1.0, + "content": "attributes are detected. This is especially helpful for question answering, where a single Gated-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 481, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 506, + 495 + ], + "score": 1.0, + "content": "Attention may not be sufficient. For example, if the textual input is ‘Which color is the pillar?’, then", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "the model needs to attend not only to feature maps detecting pillars (done by the Gated-Attention),", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "but also to other attributes at the spatial locations where pillars are seen in order to predict their color.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "score": 1.0, + "content": "Note that a single Gated-Attention is sufficient for instruction following, as shown in (Chaplot et al.,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "2017). For example, if the textual input is ‘Go to the green pillar’, the first Gated-Attention unit can", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "learn to attend to feature maps detecting green objects and pillar, and learn a navigation policy based", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 548, + 324, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 324, + 561 + ], + "score": 1.0, + "content": "on the spatial locations of the feature map activations.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 459, + 506, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 564, + 504, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 123, + 576 + ], + "score": 0.86, + "content": "x _ { \\mathrm { S A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 563, + 505, + 578 + ], + "score": 1.0, + "content": "is then passed through another Gated-Attention unit with the sentence-level text representation:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 563, + 505, + 578 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 579, + 380, + 595 + ], + "lines": [ + { + "bbox": [ + 230, + 579, + 380, + 595 + ], + "spans": [ + { + "bbox": [ + 230, + 579, + 380, + 595 + ], + "score": 0.93, + "content": "x _ { \\mathrm { G A 2 } } = f _ { \\mathrm { G A } } ( x _ { \\mathrm { S A } } , x _ { \\mathrm { s e n t } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }", + "type": "interline_equation", + "image_path": "c2f422d394e5aa39fd0b01a9687570166e773a77be9f521371c2444b6252180e.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 230, + 579, + 380, + 595 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 505, + 665 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "This second Gated-Attention unit enables the model to attend to different types of attributes based", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "on the question. For instance, if the question is asking about the color (‘Which color is the pillar?’),", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "score": 1.0, + "content": "then the model needs to attend to the feature maps corresponding to colors; or if the question is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "asking about the object type (‘Which object is green in color?’), then the model needs to attend to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 642, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 399, + 655 + ], + "score": 1.0, + "content": "the feature maps corresponding to object types. The sentence embedding", + "type": "text" + }, + { + "bbox": [ + 399, + 644, + 418, + 654 + ], + "score": 0.88, + "content": "x _ { \\mathrm { s e n t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 642, + 506, + 655 + ], + "score": 1.0, + "content": "can learn to attend to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 654, + 358, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 358, + 665 + ], + "score": 1.0, + "content": "multiple channels based on the textual input and mask the rest.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 598, + 506, + 665 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 669, + 504, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 682 + ], + "score": 1.0, + "content": "Next, the output is transformed to answer prediction by again doing a summation and softmax but", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 681, + 383, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 383, + 693 + ], + "score": 1.0, + "content": "this time summing over the height and width instead of the channels:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 669, + 505, + 693 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 230, + 696, + 381, + 736 + ], + "lines": [ + { + "bbox": [ + 230, + 696, + 381, + 736 + ], + "spans": [ + { + "bbox": [ + 230, + 696, + 381, + 736 + ], + "score": 0.95, + "content": "x _ { \\mathrm { A n s } } = \\sigma \\left( \\sum _ { j , k } ^ { H , W } x _ { \\mathrm { G A 2 } } [ : , j , k ] \\right) \\in \\mathbb { R } ^ { V }", + "type": "interline_equation", + "image_path": "0daedaacc108632c304ac35dd2039c7625699fb7bf18eea83f8178c53dfe6988.jpg" + } + ] + } + ], + "index": 37.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 696, + 381, + 716.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 230, + 716.0, + 381, + 736.0 + ], + "spans": [], + "index": 38 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 167, + 96 + ], + "score": 1.0, + "content": "Summation of", + "type": "text" + }, + { + "bbox": [ + 167, + 84, + 189, + 94 + ], + "score": 0.86, + "content": "x _ { \\mathrm { G A } 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "along each feature map aggregates the activations for relevant attributes spa-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "tially. Again, ReLU activations for sentence embedding ensure aggregation of activations for each", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 104, + 420, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 400, + 118 + ], + "score": 1.0, + "content": "attribute or word. The answer space is identical to the textual input space", + "type": "text" + }, + { + "bbox": [ + 400, + 104, + 415, + 114 + ], + "score": 0.89, + "content": "\\mathbb { R } ^ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 104, + 420, + 118 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 121, + 503, + 144 + ], + "lines": [ + { + "bbox": [ + 106, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 240, + 134 + ], + "score": 1.0, + "content": "Finally, the Dual-Attention unit", + "type": "text" + }, + { + "bbox": [ + 240, + 122, + 256, + 133 + ], + "score": 0.89, + "content": "f _ { \\mathrm { D A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 121, + 383, + 134 + ], + "score": 1.0, + "content": "outputs the answer prediction", + "type": "text" + }, + { + "bbox": [ + 383, + 123, + 402, + 132 + ], + "score": 0.87, + "content": "x _ { \\mathrm { A n s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "and the flattened spatial", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 414, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 163, + 145 + ], + "score": 1.0, + "content": "attention map", + "type": "text" + }, + { + "bbox": [ + 164, + 132, + 226, + 145 + ], + "score": 0.92, + "content": "x _ { \\mathrm { S } } = \\mathrm { v e c } ( x _ { \\mathrm { s p a t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 132, + 257, + 145 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 257, + 132, + 283, + 144 + ], + "score": 0.3, + "content": "\\mathrm { v e c } \\bar { ( . ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 132, + 414, + 145 + ], + "score": 1.0, + "content": "denotes the flattening operation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 149, + 505, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 425, + 162 + ], + "score": 1.0, + "content": "Policy Module. The policy module takes as input the state representation", + "type": "text" + }, + { + "bbox": [ + 425, + 151, + 437, + 160 + ], + "score": 0.84, + "content": "x _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "from the Dual-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 266, + 173 + ], + "score": 1.0, + "content": "Attention unit, a time step embedding", + "type": "text" + }, + { + "bbox": [ + 266, + 162, + 272, + 171 + ], + "score": 0.65, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 160, + 397, + 173 + ], + "score": 1.0, + "content": ", and a task indicator variable", + "type": "text" + }, + { + "bbox": [ + 397, + 161, + 404, + 170 + ], + "score": 0.74, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 160, + 506, + 173 + ], + "score": 1.0, + "content": "(for whether the task is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "SGN or EQA). The inputs are concatenated then passed through a linear layer, then a recurrent GRU", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 322, + 195 + ], + "score": 1.0, + "content": "layer, then linear layers to estimate the policy function", + "type": "text" + }, + { + "bbox": [ + 323, + 182, + 374, + 194 + ], + "score": 0.93, + "content": "\\pi ( a _ { t } \\mid I _ { t } , T )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 182, + 465, + 195 + ], + "score": 1.0, + "content": "and the value function", + "type": "text" + }, + { + "bbox": [ + 465, + 182, + 501, + 195 + ], + "score": 0.92, + "content": "V ( I _ { t } , T )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 182, + 505, + 195 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 199, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "All above operations are differentiable, making the entire architecture trainable end-to-end. Note", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "score": 1.0, + "content": "that all attention mechanisms in the Dual-Attention unit only modulate the input image represen-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "score": 1.0, + "content": "tation, i.e., mask or amplify specific feature maps or pixels. This ensures that there is an explicit", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "score": 1.0, + "content": "alignment between the words in the textual input, the feature maps in the image representation, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 242, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 256 + ], + "score": 1.0, + "content": "the words in answer space. This forces the convolutional network to encode all the information", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "required with respect to a certain word in the corresponding output channel. This explicit task-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "score": 1.0, + "content": "invariant alignment between convolutional feature maps and words in the input and answer space", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "facilitates grounding and allows for cross-task knowledge transfer. As shown in the results later, this", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "also makes our model modular and allows easy addition of objects and attributes to a trained model.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 107, + 311, + 195, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 197, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 197, + 324 + ], + "score": 1.0, + "content": "4.2 OPTIMIZATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 327, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "The entire model is trained to predict both navigational actions and answers jointly. The policy is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "trained using Proximal Policy Optimization (PPO) (Schulman et al., 2017). For training the answer", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 350, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 362 + ], + "score": 1.0, + "content": "predictions, we use a supervised cross-entropy loss. Both types of losses have common parameters", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 359, + 440, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 440, + 374 + ], + "score": 1.0, + "content": "as the answer prediction is essentially an intermediate representation for the policy.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 377, + 376, + 476 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 376, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 376, + 389 + ], + "score": 1.0, + "content": "Auxiliary Task. As mentioned earlier, the feature maps in the con-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 388, + 376, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 376, + 400 + ], + "score": 1.0, + "content": "volutional output are expected to detect different objects and at-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 398, + 376, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 376, + 411 + ], + "score": 1.0, + "content": "tributes. Consequently, we add a spatial auxiliary task to detect", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 410, + 376, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 376, + 422 + ], + "score": 1.0, + "content": "the object or attribute in the convolutional output channels corre-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 421, + 376, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 376, + 433 + ], + "score": 1.0, + "content": "sponding to the word in the bag-of-words representation. A prior", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 432, + 376, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 376, + 443 + ], + "score": 1.0, + "content": "work (Gupta et al., 2017b) also explored the use of attribute and ob-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 442, + 376, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 376, + 456 + ], + "score": 1.0, + "content": "ject recognition as an auxiliary task for Visual Question Answering.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 453, + 377, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 377, + 466 + ], + "score": 1.0, + "content": "Rather than doing fine-grained object detection, we keep the size of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 465, + 376, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 376, + 477 + ], + "score": 1.0, + "content": "the auxiliary predictions the same as the convolutional output to", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 384, + 380, + 504, + 447 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 384, + 380, + 504, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 384, + 380, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 384, + 380, + 504, + 447 + ], + "score": 0.965, + "type": "image", + "image_path": "fa7a1639c57b9e7bccb73053f54f5d1fe41e28c1028ec51acbe83fd8097150c0.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 384, + 380, + 504, + 413.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 384, + 413.5, + 504, + 447.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 384, + 451, + 504, + 471 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 383, + 451, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 383, + 451, + 505, + 461 + ], + "score": 1.0, + "content": "Figure 6: Example auxiliary task", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 383, + 460, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 383, + 460, + 479, + 471 + ], + "score": 1.0, + "content": "labels for the red channel.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "avoid increase in number of parameters, and maintain the explicit alignment on the convolutional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 361, + 499 + ], + "score": 1.0, + "content": "feature maps with the words. Consequently, auxiliary labels are", + "type": "text" + }, + { + "bbox": [ + 361, + 487, + 419, + 498 + ], + "score": 0.9, + "content": "( V \\times H \\times W )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "-dimensional tensors,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 178, + 510 + ], + "score": 1.0, + "content": "where each of the", + "type": "text" + }, + { + "bbox": [ + 178, + 498, + 188, + 508 + ], + "score": 0.58, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "channels correspond to a word in the vocabulary, and each element in a channel", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "score": 1.0, + "content": "is 1 if the corresponding object or attribute is present in the current frame spatially. Figure 6 shows", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "examples of auxiliary task labels for the channel corresponding to the word ‘red’. The auxiliary", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 529, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 289, + 543 + ], + "score": 1.0, + "content": "tasks are also trained with cross-entropy loss.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + }, + { + "type": "title", + "bbox": [ + 108, + 557, + 262, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 264, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 264, + 571 + ], + "score": 1.0, + "content": "5 EXPERIMENTS & RESULTS", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "score": 1.0, + "content": "Jointly learning semantic goal navigation and embodied question answering essentially involves a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "fusion of verbal and visual modalities. While prior methods are designed for a single task, we adapt", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 601, + 504, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 504, + 613 + ], + "score": 1.0, + "content": "several baselines for our environment and tasks by using their multimodal fusion techniques. We", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "use two naive baselines, Image only and Text only; two baselines based on prior semantic goal nav-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "igation models, Concat (used by Hermann et al. (2017); Misra et al. (2017)) and Gated-Attention", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "(GA) (Chaplot et al., 2017); and two baselines based on Question Answering models, FiLM (Perez", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "et al., 2017) and PACMAN (Das et al., 2017). For fair comparison, we replace the proposed Dual-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "Attention unit with multimodal fusion techniques in the baselines and keep everything else identical", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 667, + 497, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 497, + 678 + ], + "score": 1.0, + "content": "to the proposed model. We provide more implementation details of all baselines in the Appendix.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47 + }, + { + "type": "title", + "bbox": [ + 107, + 693, + 170, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 691, + 171, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 171, + 706 + ], + "score": 1.0, + "content": "5.1 RESULTS", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 52 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "We train all models for 10 million frames in the Easy setting and 50 million frames in the Hard", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 182, + 733 + ], + "score": 1.0, + "content": "setting. We use a", + "type": "text" + }, + { + "bbox": [ + 183, + 721, + 195, + 731 + ], + "score": 0.8, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "reward for reaching the correct object in SGN episodes and predicting the", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 53.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 167, + 96 + ], + "score": 1.0, + "content": "Summation of", + "type": "text" + }, + { + "bbox": [ + 167, + 84, + 189, + 94 + ], + "score": 0.86, + "content": "x _ { \\mathrm { G A } 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "along each feature map aggregates the activations for relevant attributes spa-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "tially. Again, ReLU activations for sentence embedding ensure aggregation of activations for each", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 104, + 420, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 400, + 118 + ], + "score": 1.0, + "content": "attribute or word. The answer space is identical to the textual input space", + "type": "text" + }, + { + "bbox": [ + 400, + 104, + 415, + 114 + ], + "score": 0.89, + "content": "\\mathbb { R } ^ { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 104, + 420, + 118 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 104, + 82, + 505, + 118 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 121, + 503, + 144 + ], + "lines": [ + { + "bbox": [ + 106, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 240, + 134 + ], + "score": 1.0, + "content": "Finally, the Dual-Attention unit", + "type": "text" + }, + { + "bbox": [ + 240, + 122, + 256, + 133 + ], + "score": 0.89, + "content": "f _ { \\mathrm { D A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 121, + 383, + 134 + ], + "score": 1.0, + "content": "outputs the answer prediction", + "type": "text" + }, + { + "bbox": [ + 383, + 123, + 402, + 132 + ], + "score": 0.87, + "content": "x _ { \\mathrm { A n s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "and the flattened spatial", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 414, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 163, + 145 + ], + "score": 1.0, + "content": "attention map", + "type": "text" + }, + { + "bbox": [ + 164, + 132, + 226, + 145 + ], + "score": 0.92, + "content": "x _ { \\mathrm { S } } = \\mathrm { v e c } ( x _ { \\mathrm { s p a t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 132, + 257, + 145 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 257, + 132, + 283, + 144 + ], + "score": 0.3, + "content": "\\mathrm { v e c } \\bar { ( . ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 132, + 414, + 145 + ], + "score": 1.0, + "content": "denotes the flattening operation.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 121, + 505, + 145 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 149, + 505, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 425, + 162 + ], + "score": 1.0, + "content": "Policy Module. The policy module takes as input the state representation", + "type": "text" + }, + { + "bbox": [ + 425, + 151, + 437, + 160 + ], + "score": 0.84, + "content": "x _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "from the Dual-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 266, + 173 + ], + "score": 1.0, + "content": "Attention unit, a time step embedding", + "type": "text" + }, + { + "bbox": [ + 266, + 162, + 272, + 171 + ], + "score": 0.65, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 160, + 397, + 173 + ], + "score": 1.0, + "content": ", and a task indicator variable", + "type": "text" + }, + { + "bbox": [ + 397, + 161, + 404, + 170 + ], + "score": 0.74, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 160, + 506, + 173 + ], + "score": 1.0, + "content": "(for whether the task is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "SGN or EQA). The inputs are concatenated then passed through a linear layer, then a recurrent GRU", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 322, + 195 + ], + "score": 1.0, + "content": "layer, then linear layers to estimate the policy function", + "type": "text" + }, + { + "bbox": [ + 323, + 182, + 374, + 194 + ], + "score": 0.93, + "content": "\\pi ( a _ { t } \\mid I _ { t } , T )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 182, + 465, + 195 + ], + "score": 1.0, + "content": "and the value function", + "type": "text" + }, + { + "bbox": [ + 465, + 182, + 501, + 195 + ], + "score": 0.92, + "content": "V ( I _ { t } , T )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 182, + 505, + 195 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 149, + 506, + 195 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 199, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "All above operations are differentiable, making the entire architecture trainable end-to-end. Note", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 223 + ], + "score": 1.0, + "content": "that all attention mechanisms in the Dual-Attention unit only modulate the input image represen-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "score": 1.0, + "content": "tation, i.e., mask or amplify specific feature maps or pixels. This ensures that there is an explicit", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "score": 1.0, + "content": "alignment between the words in the textual input, the feature maps in the image representation, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 242, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 256 + ], + "score": 1.0, + "content": "the words in answer space. This forces the convolutional network to encode all the information", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "required with respect to a certain word in the corresponding output channel. This explicit task-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 505, + 278 + ], + "score": 1.0, + "content": "invariant alignment between convolutional feature maps and words in the input and answer space", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "facilitates grounding and allows for cross-task knowledge transfer. As shown in the results later, this", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "also makes our model modular and allows easy addition of objects and attributes to a trained model.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 199, + 505, + 299 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 311, + 195, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 197, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 197, + 324 + ], + "score": 1.0, + "content": "4.2 OPTIMIZATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 327, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "The entire model is trained to predict both navigational actions and answers jointly. The policy is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "trained using Proximal Policy Optimization (PPO) (Schulman et al., 2017). For training the answer", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 350, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 506, + 362 + ], + "score": 1.0, + "content": "predictions, we use a supervised cross-entropy loss. Both types of losses have common parameters", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 359, + 440, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 440, + 374 + ], + "score": 1.0, + "content": "as the answer prediction is essentially an intermediate representation for the policy.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 326, + 506, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 377, + 376, + 476 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 376, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 376, + 389 + ], + "score": 1.0, + "content": "Auxiliary Task. As mentioned earlier, the feature maps in the con-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 388, + 376, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 376, + 400 + ], + "score": 1.0, + "content": "volutional output are expected to detect different objects and at-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 398, + 376, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 376, + 411 + ], + "score": 1.0, + "content": "tributes. Consequently, we add a spatial auxiliary task to detect", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 410, + 376, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 376, + 422 + ], + "score": 1.0, + "content": "the object or attribute in the convolutional output channels corre-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 421, + 376, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 376, + 433 + ], + "score": 1.0, + "content": "sponding to the word in the bag-of-words representation. A prior", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 432, + 376, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 376, + 443 + ], + "score": 1.0, + "content": "work (Gupta et al., 2017b) also explored the use of attribute and ob-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 442, + 376, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 376, + 456 + ], + "score": 1.0, + "content": "ject recognition as an auxiliary task for Visual Question Answering.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 453, + 377, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 377, + 466 + ], + "score": 1.0, + "content": "Rather than doing fine-grained object detection, we keep the size of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 465, + 376, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 376, + 477 + ], + "score": 1.0, + "content": "the auxiliary predictions the same as the convolutional output to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "avoid increase in number of parameters, and maintain the explicit alignment on the convolutional", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 361, + 499 + ], + "score": 1.0, + "content": "feature maps with the words. Consequently, auxiliary labels are", + "type": "text" + }, + { + "bbox": [ + 361, + 487, + 419, + 498 + ], + "score": 0.9, + "content": "( V \\times H \\times W )", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "-dimensional tensors,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 178, + 510 + ], + "score": 1.0, + "content": "where each of the", + "type": "text" + }, + { + "bbox": [ + 178, + 498, + 188, + 508 + ], + "score": 0.58, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "channels correspond to a word in the vocabulary, and each element in a channel", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "score": 1.0, + "content": "is 1 if the corresponding object or attribute is present in the current frame spatially. Figure 6 shows", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "examples of auxiliary task labels for the channel corresponding to the word ‘red’. The auxiliary", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 529, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 289, + 543 + ], + "score": 1.0, + "content": "tasks are also trained with cross-entropy loss.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 376, + 377, + 477 + ] + }, + { + "type": "image", + "bbox": [ + 384, + 380, + 504, + 447 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 384, + 380, + 504, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 384, + 380, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 384, + 380, + 504, + 447 + ], + "score": 0.965, + "type": "image", + "image_path": "fa7a1639c57b9e7bccb73053f54f5d1fe41e28c1028ec51acbe83fd8097150c0.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 384, + 380, + 504, + 413.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 384, + 413.5, + 504, + 447.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 384, + 451, + 504, + 471 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 383, + 451, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 383, + 451, + 505, + 461 + ], + "score": 1.0, + "content": "Figure 6: Example auxiliary task", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 383, + 460, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 383, + 460, + 479, + 471 + ], + "score": 1.0, + "content": "labels for the red channel.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 542 + ], + "lines": [], + "index": 38.5, + "bbox_fs": [ + 105, + 476, + 505, + 543 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 557, + 262, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 264, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 264, + 571 + ], + "score": 1.0, + "content": "5 EXPERIMENTS & RESULTS", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 591 + ], + "score": 1.0, + "content": "Jointly learning semantic goal navigation and embodied question answering essentially involves a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "fusion of verbal and visual modalities. While prior methods are designed for a single task, we adapt", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 601, + 504, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 504, + 613 + ], + "score": 1.0, + "content": "several baselines for our environment and tasks by using their multimodal fusion techniques. We", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "use two naive baselines, Image only and Text only; two baselines based on prior semantic goal nav-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "igation models, Concat (used by Hermann et al. (2017); Misra et al. (2017)) and Gated-Attention", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "(GA) (Chaplot et al., 2017); and two baselines based on Question Answering models, FiLM (Perez", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "et al., 2017) and PACMAN (Das et al., 2017). For fair comparison, we replace the proposed Dual-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "Attention unit with multimodal fusion techniques in the baselines and keep everything else identical", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 667, + 497, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 497, + 678 + ], + "score": 1.0, + "content": "to the proposed model. We provide more implementation details of all baselines in the Appendix.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 579, + 506, + 678 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 693, + 170, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 691, + 171, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 171, + 706 + ], + "score": 1.0, + "content": "5.1 RESULTS", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 52 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "We train all models for 10 million frames in the Easy setting and 50 million frames in the Hard", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 182, + 733 + ], + "score": 1.0, + "content": "setting. We use a", + "type": "text" + }, + { + "bbox": [ + 183, + 721, + 195, + 731 + ], + "score": 0.8, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "reward for reaching the correct object in SGN episodes and predicting the", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "correct answer in EQA episodes. We use a small negative reward of -0.001 per time step to encour-", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 363, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 374 + ], + "score": 1.0, + "content": "age shorter paths to target and answering questions as soon as possible. We also use distance-based", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "reward shaping for SGN episodes, where the agent receives a small reward proportional to de-", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "score": 1.0, + "content": "crease in distance to the target. In the next subsection we evaluate the performance of the proposed", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "model without the reward shaping. SGN episodes end when the agent reaches any object, and EQA", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "episodes when the agent predicts any answer. All episodes have a maximum length of 210 time", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 418, + 489, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 489, + 430 + ], + "score": 1.0, + "content": "steps. We train all models with and without the auxiliary tasks using identical reward functions.", + "type": "text", + "cross_page": true + } + ], + "index": 14 + } + ], + "index": 53.5, + "bbox_fs": [ + 106, + 709, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 78, + 465, + 180 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 78, + 465, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 78, + 465, + 180 + ], + "spans": [ + { + "bbox": [ + 145, + 78, + 465, + 180 + ], + "score": 0.963, + "type": "image", + "image_path": "0f763d40f3cf3cc304935d3ef380ea01fe08bd89fadf0f7ce22ecb962077a7f5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 145, + 78, + 465, + 112.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 145, + 112.0, + 465, + 146.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 145, + 146.0, + 465, + 180.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 121, + 183, + 488, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 181, + 491, + 196 + ], + "spans": [ + { + "bbox": [ + 119, + 181, + 491, + 196 + ], + "score": 1.0, + "content": "Figure 7: Training accuracy of all models trained with auxiliary tasks for Easy (left) and Hard (right).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 108, + 218, + 504, + 329 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 134, + 206, + 472, + 217 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 205, + 473, + 218 + ], + "spans": [ + { + "bbox": [ + 137, + 205, + 473, + 218 + ], + "score": 1.0, + "content": "Table 1: Accuracy of all models on SGN & EQA test sets for both Easy & Hard difficulties.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 108, + 218, + 504, + 329 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 218, + 504, + 329 + ], + "spans": [ + { + "bbox": [ + 108, + 218, + 504, + 329 + ], + "score": 0.978, + "html": "
EasyHard
No AuxAuxNo AuxAux
ModelSGNEQASGNEQASGNEQASGNEQA
Text only0.20.330.20.330.20.330.20.33
Image only0.200.090.210.080.160.080.150.08
Concat0.330.210.310.190.20.260.390.22
GA0.270.180.350.240.180.110.220.24
FiLM0.240.110.340.120.120.030.250.15
PACMAN0.260.120.330.100.290.330.110.27
Dual-Attention0.860.530.960.580.860.380.820.59
", + "type": "table", + "image_path": "0780412306b172e3fc51fe31cf166b5d535d27096b5b77bdb329922387195d78.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 218, + 504, + 255.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 255.0, + 504, + 292.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 292.0, + 504, + 329.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "correct answer in EQA episodes. We use a small negative reward of -0.001 per time step to encour-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 363, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 374 + ], + "score": 1.0, + "content": "age shorter paths to target and answering questions as soon as possible. We also use distance-based", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "reward shaping for SGN episodes, where the agent receives a small reward proportional to de-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 396 + ], + "score": 1.0, + "content": "crease in distance to the target. In the next subsection we evaluate the performance of the proposed", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "model without the reward shaping. SGN episodes end when the agent reaches any object, and EQA", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "episodes when the agent predicts any answer. All episodes have a maximum length of 210 time", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 418, + 489, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 489, + 430 + ], + "score": 1.0, + "content": "steps. We train all models with and without the auxiliary tasks using identical reward functions.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "All models are trained jointly for both the tasks and tested on each task separately. In Figure 7, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "show the training performance curves for all models trained with Auxiliary tasks in both Easy and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "Hard settings. In Table 1, we report the test performance of all models on both SGN and EQA for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 467, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 479 + ], + "score": 1.0, + "content": "both Easy and Hard settings. During training, the Dual-Attention model learns faster as compared", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "to the baselines in the Easy setting while achieving higher final performance in the Hard setting (see", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "Figure 7). More interestingly, as shown in Table 1, the Dual-Attention model achieves considerably", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 501, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 512 + ], + "score": 1.0, + "content": "higher accuracy on the test set for SGN and EQA in both the difficulty settings when trained with or", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "without auxiliary tasks. These results confirm the hypothesis that prior models, which are designed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "for a single task, lack the ability align the words in both the tasks and transfer knowledge across", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 531, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 506, + 547 + ], + "score": 1.0, + "content": "tasks. Lower accuracy on EQA for most models (see Table 1) indicates that EQA is more challenging", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 543, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 556 + ], + "score": 1.0, + "content": "than SGN as it involves alignment between not only input textual and visual representations but also", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "with the answer space. As expected, using spatial auxiliary tasks lead to better performance for all", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "models Visualization of the attention maps and intermediate representations in the model indicate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 577, + 504, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 504, + 589 + ], + "score": 1.0, + "content": "that the textual and visual representations are aligned as expected (see Appendix for visualizations)1.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 108, + 600, + 205, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 599, + 207, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 207, + 613 + ], + "score": 1.0, + "content": "5.2 ABLATION TESTS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "We perform a series of ablation tests in order to analyze the contribution of each component in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "the Dual-Attention unit: without Spatial-Attention (w/o SA), without the first Gated-Attention with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 638, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 133, + 650 + ], + "score": 0.83, + "content": "x _ { B o W }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 638, + 378, + 653 + ], + "score": 1.0, + "content": "(w/o GA1), and without the second Gated-Attention with", + "type": "text" + }, + { + "bbox": [ + 378, + 639, + 397, + 650 + ], + "score": 0.87, + "content": "x _ { \\mathrm { s e n t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 638, + 506, + 653 + ], + "score": 1.0, + "content": "(w/o GA2). We also try", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "removing the task indicator variable (w/o Indicator Variable), removing reward shaping (w/o Re-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "ward Shaping), and training the proposed model on a single task, SGN or EQA (DA Single-Task).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 676, + 503, + 710 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "Figure 8 shows the training performance curves for the Dual-Attention model along with all ablation", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "score": 1.0, + "content": "models in the Easy Setting. In Table 2, we report the test set performance of all ablation models.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "The results indicate that SA and GA1 contribute the most to the performance of the Dual-Attention", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 114, + 722, + 487, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 718, + 489, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 718, + 489, + 734 + ], + "score": 1.0, + "content": "1See https://sites.google.com/view/emml for policy execution and visualization videos.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 78, + 465, + 180 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 78, + 465, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 78, + 465, + 180 + ], + "spans": [ + { + "bbox": [ + 145, + 78, + 465, + 180 + ], + "score": 0.963, + "type": "image", + "image_path": "0f763d40f3cf3cc304935d3ef380ea01fe08bd89fadf0f7ce22ecb962077a7f5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 145, + 78, + 465, + 112.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 145, + 112.0, + 465, + 146.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 145, + 146.0, + 465, + 180.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 121, + 183, + 488, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 181, + 491, + 196 + ], + "spans": [ + { + "bbox": [ + 119, + 181, + 491, + 196 + ], + "score": 1.0, + "content": "Figure 7: Training accuracy of all models trained with auxiliary tasks for Easy (left) and Hard (right).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 108, + 218, + 504, + 329 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 134, + 206, + 472, + 217 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 205, + 473, + 218 + ], + "spans": [ + { + "bbox": [ + 137, + 205, + 473, + 218 + ], + "score": 1.0, + "content": "Table 1: Accuracy of all models on SGN & EQA test sets for both Easy & Hard difficulties.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 108, + 218, + 504, + 329 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 218, + 504, + 329 + ], + "spans": [ + { + "bbox": [ + 108, + 218, + 504, + 329 + ], + "score": 0.978, + "html": "
EasyHard
No AuxAuxNo AuxAux
ModelSGNEQASGNEQASGNEQASGNEQA
Text only0.20.330.20.330.20.330.20.33
Image only0.200.090.210.080.160.080.150.08
Concat0.330.210.310.190.20.260.390.22
GA0.270.180.350.240.180.110.220.24
FiLM0.240.110.340.120.120.030.250.15
PACMAN0.260.120.330.100.290.330.110.27
Dual-Attention0.860.530.960.580.860.380.820.59
", + "type": "table", + "image_path": "0780412306b172e3fc51fe31cf166b5d535d27096b5b77bdb329922387195d78.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 218, + 504, + 255.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 255.0, + 504, + 292.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 292.0, + 504, + 329.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 505, + 429 + ], + "lines": [], + "index": 11, + "bbox_fs": [ + 105, + 351, + 505, + 430 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "All models are trained jointly for both the tasks and tested on each task separately. In Figure 7, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "show the training performance curves for all models trained with Auxiliary tasks in both Easy and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "Hard settings. In Table 1, we report the test performance of all models on both SGN and EQA for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 467, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 479 + ], + "score": 1.0, + "content": "both Easy and Hard settings. During training, the Dual-Attention model learns faster as compared", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "to the baselines in the Easy setting while achieving higher final performance in the Hard setting (see", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "Figure 7). More interestingly, as shown in Table 1, the Dual-Attention model achieves considerably", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 501, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 512 + ], + "score": 1.0, + "content": "higher accuracy on the test set for SGN and EQA in both the difficulty settings when trained with or", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "without auxiliary tasks. These results confirm the hypothesis that prior models, which are designed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "for a single task, lack the ability align the words in both the tasks and transfer knowledge across", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 531, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 506, + 547 + ], + "score": 1.0, + "content": "tasks. Lower accuracy on EQA for most models (see Table 1) indicates that EQA is more challenging", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 543, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 556 + ], + "score": 1.0, + "content": "than SGN as it involves alignment between not only input textual and visual representations but also", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "with the answer space. As expected, using spatial auxiliary tasks lead to better performance for all", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "models Visualization of the attention maps and intermediate representations in the model indicate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 577, + 504, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 504, + 589 + ], + "score": 1.0, + "content": "that the textual and visual representations are aligned as expected (see Appendix for visualizations)1.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 434, + 506, + 589 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 600, + 205, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 599, + 207, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 207, + 613 + ], + "score": 1.0, + "content": "5.2 ABLATION TESTS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "We perform a series of ablation tests in order to analyze the contribution of each component in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "the Dual-Attention unit: without Spatial-Attention (w/o SA), without the first Gated-Attention with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 638, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 133, + 650 + ], + "score": 0.83, + "content": "x _ { B o W }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 638, + 378, + 653 + ], + "score": 1.0, + "content": "(w/o GA1), and without the second Gated-Attention with", + "type": "text" + }, + { + "bbox": [ + 378, + 639, + 397, + 650 + ], + "score": 0.87, + "content": "x _ { \\mathrm { s e n t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 638, + 506, + 653 + ], + "score": 1.0, + "content": "(w/o GA2). We also try", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "removing the task indicator variable (w/o Indicator Variable), removing reward shaping (w/o Re-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "ward Shaping), and training the proposed model on a single task, SGN or EQA (DA Single-Task).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 617, + 506, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 676, + 503, + 710 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "Figure 8 shows the training performance curves for the Dual-Attention model along with all ablation", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "score": 1.0, + "content": "models in the Easy Setting. In Table 2, we report the test set performance of all ablation models.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "The results indicate that SA and GA1 contribute the most to the performance of the Dual-Attention", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 376, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 388 + ], + "score": 1.0, + "content": "model. GA2 is critical for performance on EQA but not SGN (see Table 2). This is expected", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "as GA2 is designed to attend to different objects and attributes based on the question and is used", + "type": "text", + "cross_page": true + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "mainly for answer prediction. It is not critical for SGN as the spatial attention map consists of", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 407, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 423 + ], + "score": 1.0, + "content": "locations of relevant objects, which is sufficient for navigating to the correct object. Reward shaping", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 420, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 432 + ], + "score": 1.0, + "content": "and indicator variable help with learning speed (see Figure 8), but have little effect on the final", + "type": "text", + "cross_page": true + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "performance (see Table 2). Dual-Attention models trained only on single tasks work well on SGN", + "type": "text", + "cross_page": true + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "score": 1.0, + "content": "especially with auxiliary tasks. This is because the auxiliary task for single task models includes", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 453, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 465 + ], + "score": 1.0, + "content": "object detection labels corresponding to the words in the test set. This highlights a key advantage of", + "type": "text", + "cross_page": true + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 462, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 478 + ], + "score": 1.0, + "content": "the proposed model. Due to its modular and interpretable design, the model can used for transferring", + "type": "text", + "cross_page": true + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "the policy to new objects and attributes without fine-tuning as discussed in the following subsection.", + "type": "text", + "cross_page": true + } + ], + "index": 34 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 676, + 505, + 711 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 79, + 464, + 180 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 79, + 464, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 79, + 464, + 180 + ], + "spans": [ + { + "bbox": [ + 146, + 79, + 464, + 180 + ], + "score": 0.967, + "type": "image", + "image_path": "69a4073a5b891f27a0c6d0e7e4b8f3eb9a94e900eda6ec575b749ee8e403a1d8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 146, + 79, + 464, + 112.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 146, + 112.66666666666666, + 464, + 146.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 146, + 146.33333333333331, + 464, + 179.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 183, + 504, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "Figure 8: Training accuracy of proposed Dual-Attention model with all ablation models trained without (left)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 194, + 314, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 314, + 205 + ], + "score": 1.0, + "content": "and with (right) auxiliary tasks for the Easy environment.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 106, + 249, + 296, + 351 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 219, + 298, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 218, + 299, + 230 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 299, + 230 + ], + "score": 1.0, + "content": "Table 2: Accuracy of all the ablation models trained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 228, + 299, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 299, + 240 + ], + "score": 1.0, + "content": "with and without Auxiliary tasks on SGN and EQA", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 239, + 257, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 257, + 250 + ], + "score": 1.0, + "content": "test sets for the Doom Easy environment.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 106, + 249, + 296, + 351 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 249, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 296, + 351 + ], + "score": 0.98, + "html": "
ModelNo Aux SGN EQAAux SGN
w/o SAEQA
w/o GA10.20 0.16 0.250.20 0.15 0.16 0.38
w/o GA20.14 0.80 0.330.97 0.15
w/o Task Indicator0.79 0.470.96 0.56
w/o Reward Shaping0.82 0.490.93 0.51
DA Single-Task0.63 0.310.91 0.34
DA Multi-Task0.86 0.530.96 0.58
", + "type": "table", + "image_path": "b49aeed69fcac50cc68bdbb554ea70bb037720c598878d8f4b9ac597c18ecfdf.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 106, + 249, + 296, + 263.57142857142856 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 263.57142857142856, + 296, + 278.1428571428571 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 278.1428571428571, + 296, + 292.71428571428567 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 292.71428571428567, + 296, + 307.2857142857142 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 307.2857142857142, + 296, + 321.8571428571428 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 321.8571428571428, + 296, + 336.42857142857133 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 336.42857142857133, + 296, + 350.9999999999999 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 8.5 + }, + { + "type": "table", + "bbox": [ + 316, + 249, + 504, + 351 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 316, + 219, + 504, + 249 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 315, + 217, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 505, + 231 + ], + "score": 1.0, + "content": "Table 3: The performance of a trained policy ap-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 315, + 228, + 505, + 240 + ], + "score": 1.0, + "content": "pended with object detectors on instructions con-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 464, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 464, + 250 + ], + "score": 1.0, + "content": "taining unseen words (‘red’ and ‘pillar’).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "table_body", + "bbox": [ + 316, + 249, + 504, + 351 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 316, + 249, + 504, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 249, + 504, + 351 + ], + "score": 0.979, + "html": "
InstructionEasyHard
Go to the pillar1.000.71
Go to the red object0.990.89
Go to the tall/short pillar0.990.68
Go to the <known_color>pillar.1.000.79
Go to the red <known_object>1.000.93
Go to the largest/smallest red object0.950.69
Go to the tall/short red pillar0.990.88
Go to the red pillar0.990.82
", + "type": "table", + "image_path": "5c95bb849656c81c03ec6f0cf06a185b6a09bde23e4d722cc774e16cf0ff6d2f.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 316, + 249, + 504, + 263.57142857142856 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 316, + 263.57142857142856, + 504, + 278.1428571428571 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 316, + 278.1428571428571, + 504, + 292.71428571428567 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 316, + 292.71428571428567, + 504, + 307.2857142857142 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 316, + 307.2857142857142, + 504, + 321.8571428571428 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 316, + 321.8571428571428, + 504, + 336.42857142857133 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 316, + 336.42857142857133, + 504, + 350.9999999999999 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 375, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 388 + ], + "score": 1.0, + "content": "model. GA2 is critical for performance on EQA but not SGN (see Table 2). This is expected", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "as GA2 is designed to attend to different objects and attributes based on the question and is used", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "mainly for answer prediction. It is not critical for SGN as the spatial attention map consists of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 407, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 423 + ], + "score": 1.0, + "content": "locations of relevant objects, which is sufficient for navigating to the correct object. Reward shaping", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 420, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 432 + ], + "score": 1.0, + "content": "and indicator variable help with learning speed (see Figure 8), but have little effect on the final", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "performance (see Table 2). Dual-Attention models trained only on single tasks work well on SGN", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 506, + 455 + ], + "score": 1.0, + "content": "especially with auxiliary tasks. This is because the auxiliary task for single task models includes", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 453, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 465 + ], + "score": 1.0, + "content": "object detection labels corresponding to the words in the test set. This highlights a key advantage of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 462, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 478 + ], + "score": 1.0, + "content": "the proposed model. Due to its modular and interpretable design, the model can used for transferring", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "the policy to new objects and attributes without fine-tuning as discussed in the following subsection.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 108, + 500, + 304, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 306, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 306, + 513 + ], + "score": 1.0, + "content": "5.3 EXTENSION: TRANSFER TO NEW WORDS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "Consider a scenario of SGN where the agent is trained to follow instructions of certain objects and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 529, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 540 + ], + "score": 1.0, + "content": "attributes. Suppose that the user wants the agent to follow instructions about a new object such as", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "‘pillar’ or a new attribute such the color ‘red’ which are never seen in any training instruction. Prior", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "SGN models are shown to perform well to unseen combination of object-attribute pairs (Chaplot", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "et al., 2017), but they do not generalize well to instructions containing a new word. The model", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 572, + 484, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 484, + 585 + ], + "score": 1.0, + "content": "retrained only on new instructions will lead to catastrophic forgetting of previous instructions.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 589, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "In contrast, our model can be used for transfer to new words by training an object detector for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 357, + 613 + ], + "score": 1.0, + "content": "each new word and appending it to the image representation", + "type": "text" + }, + { + "bbox": [ + 358, + 602, + 369, + 611 + ], + "score": 0.82, + "content": "x _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 600, + 506, + 613 + ], + "score": 1.0, + "content": ". In order to test this, we train a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "single-task SGN model using the proposed architecture on the training set for instructions. We use", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "auxiliary tasks but only for words in the vocabulary of the instructions training set. After training", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "the policy, we would like the agent to follow instructions containing test words ‘red’ and ‘pillar’,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "which the agent has never seen or received any supervision about how this attribute or object looks", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "visually. For transferring the policy, we assume access to two object detectors which would give", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "object detections for ‘red’ and ‘pillar’ separately. We resize the object detections to the size of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 675, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 104, + 675, + 282, + 691 + ], + "score": 1.0, + "content": "a feature map in the image representation", + "type": "text" + }, + { + "bbox": [ + 283, + 677, + 324, + 688 + ], + "score": 0.89, + "content": "( H \\times W )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 675, + 506, + 691 + ], + "score": 1.0, + "content": "and append them as channels to the image", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "representation. We also append the words ‘red’ and ‘pillar’ to the bag-of-words representations in", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "the same order such that they are aligned with the appended feature maps. We randomly initialize", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "the embeddings of the new words for computing the sentence embedding. The results in Table 3", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "show that this policy generalizes well to different types of instructions with unseen words. This", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 48 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 79, + 464, + 180 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 79, + 464, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 79, + 464, + 180 + ], + "spans": [ + { + "bbox": [ + 146, + 79, + 464, + 180 + ], + "score": 0.967, + "type": "image", + "image_path": "69a4073a5b891f27a0c6d0e7e4b8f3eb9a94e900eda6ec575b749ee8e403a1d8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 146, + 79, + 464, + 112.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 146, + 112.66666666666666, + 464, + 146.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 146, + 146.33333333333331, + 464, + 179.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 183, + 504, + 204 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "Figure 8: Training accuracy of proposed Dual-Attention model with all ablation models trained without (left)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 194, + 314, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 314, + 205 + ], + "score": 1.0, + "content": "and with (right) auxiliary tasks for the Easy environment.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 106, + 249, + 296, + 351 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 219, + 298, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 218, + 299, + 230 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 299, + 230 + ], + "score": 1.0, + "content": "Table 2: Accuracy of all the ablation models trained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 228, + 299, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 299, + 240 + ], + "score": 1.0, + "content": "with and without Auxiliary tasks on SGN and EQA", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 239, + 257, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 257, + 250 + ], + "score": 1.0, + "content": "test sets for the Doom Easy environment.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 106, + 249, + 296, + 351 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 249, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 296, + 351 + ], + "score": 0.98, + "html": "
ModelNo Aux SGN EQAAux SGN
w/o SAEQA
w/o GA10.20 0.16 0.250.20 0.15 0.16 0.38
w/o GA20.14 0.80 0.330.97 0.15
w/o Task Indicator0.79 0.470.96 0.56
w/o Reward Shaping0.82 0.490.93 0.51
DA Single-Task0.63 0.310.91 0.34
DA Multi-Task0.86 0.530.96 0.58
", + "type": "table", + "image_path": "b49aeed69fcac50cc68bdbb554ea70bb037720c598878d8f4b9ac597c18ecfdf.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 106, + 249, + 296, + 263.57142857142856 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 263.57142857142856, + 296, + 278.1428571428571 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 278.1428571428571, + 296, + 292.71428571428567 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 292.71428571428567, + 296, + 307.2857142857142 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 307.2857142857142, + 296, + 321.8571428571428 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 321.8571428571428, + 296, + 336.42857142857133 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 336.42857142857133, + 296, + 350.9999999999999 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 8.5 + }, + { + "type": "table", + "bbox": [ + 316, + 249, + 504, + 351 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 316, + 219, + 504, + 249 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 315, + 217, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 505, + 231 + ], + "score": 1.0, + "content": "Table 3: The performance of a trained policy ap-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 315, + 228, + 505, + 240 + ], + "score": 1.0, + "content": "pended with object detectors on instructions con-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 464, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 464, + 250 + ], + "score": 1.0, + "content": "taining unseen words (‘red’ and ‘pillar’).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "table_body", + "bbox": [ + 316, + 249, + 504, + 351 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 316, + 249, + 504, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 249, + 504, + 351 + ], + "score": 0.979, + "html": "
InstructionEasyHard
Go to the pillar1.000.71
Go to the red object0.990.89
Go to the tall/short pillar0.990.68
Go to the <known_color>pillar.1.000.79
Go to the red <known_object>1.000.93
Go to the largest/smallest red object0.950.69
Go to the tall/short red pillar0.990.88
Go to the red pillar0.990.82
", + "type": "table", + "image_path": "5c95bb849656c81c03ec6f0cf06a185b6a09bde23e4d722cc774e16cf0ff6d2f.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 316, + 249, + 504, + 263.57142857142856 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 316, + 263.57142857142856, + 504, + 278.1428571428571 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 316, + 278.1428571428571, + 504, + 292.71428571428567 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 316, + 292.71428571428567, + 504, + 307.2857142857142 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 316, + 307.2857142857142, + 504, + 321.8571428571428 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 316, + 321.8571428571428, + 504, + 336.42857142857133 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 316, + 336.42857142857133, + 504, + 350.9999999999999 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 375, + 505, + 486 + ], + "lines": [], + "index": 29.5, + "bbox_fs": [ + 105, + 376, + 506, + 487 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 500, + 304, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 306, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 306, + 513 + ], + "score": 1.0, + "content": "5.3 EXTENSION: TRANSFER TO NEW WORDS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "Consider a scenario of SGN where the agent is trained to follow instructions of certain objects and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 529, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 540 + ], + "score": 1.0, + "content": "attributes. Suppose that the user wants the agent to follow instructions about a new object such as", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "‘pillar’ or a new attribute such the color ‘red’ which are never seen in any training instruction. Prior", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "SGN models are shown to perform well to unseen combination of object-attribute pairs (Chaplot", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "et al., 2017), but they do not generalize well to instructions containing a new word. The model", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 572, + 484, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 484, + 585 + ], + "score": 1.0, + "content": "retrained only on new instructions will lead to catastrophic forgetting of previous instructions.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 517, + 505, + 585 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 589, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "In contrast, our model can be used for transfer to new words by training an object detector for", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 357, + 613 + ], + "score": 1.0, + "content": "each new word and appending it to the image representation", + "type": "text" + }, + { + "bbox": [ + 358, + 602, + 369, + 611 + ], + "score": 0.82, + "content": "x _ { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 600, + 506, + 613 + ], + "score": 1.0, + "content": ". In order to test this, we train a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "single-task SGN model using the proposed architecture on the training set for instructions. We use", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 635 + ], + "score": 1.0, + "content": "auxiliary tasks but only for words in the vocabulary of the instructions training set. After training", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "the policy, we would like the agent to follow instructions containing test words ‘red’ and ‘pillar’,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "which the agent has never seen or received any supervision about how this attribute or object looks", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "visually. For transferring the policy, we assume access to two object detectors which would give", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "object detections for ‘red’ and ‘pillar’ separately. We resize the object detections to the size of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 675, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 104, + 675, + 282, + 691 + ], + "score": 1.0, + "content": "a feature map in the image representation", + "type": "text" + }, + { + "bbox": [ + 283, + 677, + 324, + 688 + ], + "score": 0.89, + "content": "( H \\times W )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 675, + 506, + 691 + ], + "score": 1.0, + "content": "and append them as channels to the image", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "representation. We also append the words ‘red’ and ‘pillar’ to the bag-of-words representations in", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "the same order such that they are aligned with the appended feature maps. We randomly initialize", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "the embeddings of the new words for computing the sentence embedding. The results in Table 3", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "show that this policy generalizes well to different types of instructions with unseen words. This", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "suggests that a trained policy can be scaled to more objects provided the complexity of navigation", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 185, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 185, + 106 + ], + "score": 1.0, + "content": "remains consistent.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 48, + "bbox_fs": [ + 104, + 589, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "suggests that a trained policy can be scaled to more objects provided the complexity of navigation", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 185, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 185, + 106 + ], + "score": 1.0, + "content": "remains consistent.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 119, + 195, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 116, + 198, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 198, + 135 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 140, + 505, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 139, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 153 + ], + "score": 1.0, + "content": "We proposed a Dual-Attention model for visually-grounded multitask learning which uses Gated-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "score": 1.0, + "content": "and Spatial-Attention to disentangle attributes in feature representations and align them with the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 162, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 174 + ], + "score": 1.0, + "content": "answer space. We show that the proposed model is able to transfer the knowledge of words across", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "tasks and outperforms the baselines on both Semantic Goal Navigation and Embodied Question", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 183, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 504, + 195 + ], + "score": 1.0, + "content": "Answering by a considerable margin. We showed that disentangled and interpretable representations", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "make our model modular and allows for easy addition of new objects or attributes to a trained model.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 206, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 505, + 218 + ], + "score": 1.0, + "content": "In future, the model can potentially be extended to transferring knowledge across different domains", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 217, + 450, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 450, + 229 + ], + "score": 1.0, + "content": "by using modular interpretable representations of objects which are domain-invariant.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 108, + 245, + 175, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 176, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 176, + 258 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 264, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "Devendra Singh Chaplot, Kanthashree Mysore Sathyendra, Rama Kumar Pasumarthi, Dheeraj Ra-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 114, + 275, + 506, + 290 + ], + "score": 1.0, + "content": "jagopal, and Ruslan Salakhutdinov. Gated-attention architectures for task-oriented language", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 287, + 326, + 298 + ], + "spans": [ + { + "bbox": [ + 116, + 287, + 326, + 298 + ], + "score": 1.0, + "content": "grounding. arXiv preprint arXiv:1706.07230, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties ¨", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 327, + 143, + 341 + ], + "spans": [ + { + "bbox": [ + 114, + 327, + 143, + 341 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 105, + 349, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Em-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 360, + 392, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 360, + 392, + 372 + ], + "score": 1.0, + "content": "bodied question answering. arXiv preprint arXiv:1711.11543, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 496, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 497, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 497, + 395 + ], + "score": 1.0, + "content": "Alexey Dosovitskiy and Vladlen Koltun. Learning to act by predicting the future. In ICLR, 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 401, + 503, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 504, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 504, + 414 + ], + "score": 1.0, + "content": "Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 424, + 254, + 435 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 254, + 435 + ], + "score": 1.0, + "content": "preprint arXiv:1606.01847, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 443, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 454, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 454, + 505, + 469 + ], + "score": 1.0, + "content": "Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 466, + 181, + 478 + ], + "spans": [ + { + "bbox": [ + 116, + 466, + 181, + 478 + ], + "score": 1.0, + "content": "315–323, 2011.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 109, + 486, + 504, + 521 + ], + "lines": [ + { + "bbox": [ + 107, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 107, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 116, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "Farhadi. Iqa: Visual question answering in interactive environments. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 509, + 438, + 523 + ], + "spans": [ + { + "bbox": [ + 116, + 509, + 438, + 523 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pp. 4089–4098, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 105, + 529, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 541, + 473, + 553 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 473, + 553 + ], + "score": 1.0, + "content": "mapping and planning for visual navigation. arXiv preprint arXiv:1702.03920, 3, 2017a.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "Tanmay Gupta, Kevin J Shih, Saurabh Singh, Derek Hoiem, Kevin J Shih, Arun Mallya, Wei Di,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 117, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "Vignesh Jagadeesh, Robinson Piramuthu, K Shih, et al. Aligned image-word representations", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 582, + 477, + 596 + ], + "spans": [ + { + "bbox": [ + 115, + 582, + 477, + 596 + ], + "score": 1.0, + "content": "improve inductive transfer across vision-language tasks. In ICCV, pp. 4223–4232, 2017b.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 603, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "Karl Moritz Hermann, Felix Hill, Simon Green, Fumin Wang, Ryan Faulkner, Hubert Soyer, David", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "Szepesvari, Wojtek Czarnecki, Max Jaderberg, Denis Teplyashin, et al. Grounded language learn-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 626, + 395, + 639 + ], + "spans": [ + { + "bbox": [ + 115, + 626, + 395, + 639 + ], + "score": 1.0, + "content": "ing in a simulated 3d world. arXiv preprint arXiv:1706.06551, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 646, + 497, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 498, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 498, + 660 + ], + "score": 1.0, + "content": "Roger A Horn. The hadamard product. In Proc. Symp. Appl. Math, volume 40, pp. 87–169, 1990.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 105, + 667, + 504, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 680 + ], + "score": 1.0, + "content": "Drew A Hudson and Christopher D Manning. Compositional attention networks for machine rea-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 678, + 311, + 690 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 311, + 690 + ], + "score": 1.0, + "content": "soning. arXiv preprint arXiv:1803.03067, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 721, + 254, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 254, + 732 + ], + "score": 1.0, + "content": "preprint arXiv:1611.05397, 2016.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 119, + 195, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 116, + 198, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 198, + 135 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 140, + 505, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 139, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 153 + ], + "score": 1.0, + "content": "We proposed a Dual-Attention model for visually-grounded multitask learning which uses Gated-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "score": 1.0, + "content": "and Spatial-Attention to disentangle attributes in feature representations and align them with the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 162, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 174 + ], + "score": 1.0, + "content": "answer space. We show that the proposed model is able to transfer the knowledge of words across", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "tasks and outperforms the baselines on both Semantic Goal Navigation and Embodied Question", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 183, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 504, + 195 + ], + "score": 1.0, + "content": "Answering by a considerable margin. We showed that disentangled and interpretable representations", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "make our model modular and allows for easy addition of new objects or attributes to a trained model.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 206, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 505, + 218 + ], + "score": 1.0, + "content": "In future, the model can potentially be extended to transferring knowledge across different domains", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 217, + 450, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 450, + 229 + ], + "score": 1.0, + "content": "by using modular interpretable representations of objects which are domain-invariant.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 139, + 505, + 229 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 245, + 175, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 176, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 176, + 258 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 264, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 277 + ], + "score": 1.0, + "content": "Devendra Singh Chaplot, Kanthashree Mysore Sathyendra, Rama Kumar Pasumarthi, Dheeraj Ra-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 114, + 275, + 506, + 290 + ], + "score": 1.0, + "content": "jagopal, and Ruslan Salakhutdinov. Gated-attention architectures for task-oriented language", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 287, + 326, + 298 + ], + "spans": [ + { + "bbox": [ + 116, + 287, + 326, + 298 + ], + "score": 1.0, + "content": "grounding. arXiv preprint arXiv:1706.07230, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 263, + 506, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties ¨", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 327, + 143, + 341 + ], + "spans": [ + { + "bbox": [ + 114, + 327, + 143, + 341 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 306, + 505, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 349, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Em-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 360, + 392, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 360, + 392, + 372 + ], + "score": 1.0, + "content": "bodied question answering. arXiv preprint arXiv:1711.11543, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 349, + 505, + 372 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 496, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 497, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 497, + 395 + ], + "score": 1.0, + "content": "Alexey Dosovitskiy and Vladlen Koltun. Learning to act by predicting the future. In ICLR, 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 379, + 497, + 395 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 401, + 503, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 504, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 504, + 414 + ], + "score": 1.0, + "content": "Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 424, + 254, + 435 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 254, + 435 + ], + "score": 1.0, + "content": "preprint arXiv:1606.01847, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 402, + 505, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 443, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 454, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 454, + 505, + 469 + ], + "score": 1.0, + "content": "Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 466, + 181, + 478 + ], + "spans": [ + { + "bbox": [ + 116, + 466, + 181, + 478 + ], + "score": 1.0, + "content": "315–323, 2011.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 444, + 505, + 478 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 486, + 504, + 521 + ], + "lines": [ + { + "bbox": [ + 107, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 107, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 116, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "Farhadi. Iqa: Visual question answering in interactive environments. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 509, + 438, + 523 + ], + "spans": [ + { + "bbox": [ + 116, + 509, + 438, + 523 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pp. 4089–4098, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 107, + 487, + 505, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 529, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 541, + 473, + 553 + ], + "spans": [ + { + "bbox": [ + 115, + 541, + 473, + 553 + ], + "score": 1.0, + "content": "mapping and planning for visual navigation. arXiv preprint arXiv:1702.03920, 3, 2017a.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 529, + 505, + 553 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "Tanmay Gupta, Kevin J Shih, Saurabh Singh, Derek Hoiem, Kevin J Shih, Arun Mallya, Wei Di,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 117, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "Vignesh Jagadeesh, Robinson Piramuthu, K Shih, et al. Aligned image-word representations", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 582, + 477, + 596 + ], + "spans": [ + { + "bbox": [ + 115, + 582, + 477, + 596 + ], + "score": 1.0, + "content": "improve inductive transfer across vision-language tasks. In ICCV, pp. 4223–4232, 2017b.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 560, + 505, + 596 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 603, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "Karl Moritz Hermann, Felix Hill, Simon Green, Fumin Wang, Ryan Faulkner, Hubert Soyer, David", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 116, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "Szepesvari, Wojtek Czarnecki, Max Jaderberg, Denis Teplyashin, et al. Grounded language learn-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 626, + 395, + 639 + ], + "spans": [ + { + "bbox": [ + 115, + 626, + 395, + 639 + ], + "score": 1.0, + "content": "ing in a simulated 3d world. arXiv preprint arXiv:1706.06551, 2017.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 603, + 506, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 646, + 497, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 498, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 498, + 660 + ], + "score": 1.0, + "content": "Roger A Horn. The hadamard product. In Proc. Symp. Appl. Math, volume 40, pp. 87–169, 1990.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 644, + 498, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 667, + 504, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 680 + ], + "score": 1.0, + "content": "Drew A Hudson and Christopher D Manning. Compositional attention networks for machine rea-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 678, + 311, + 690 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 311, + 690 + ], + "score": 1.0, + "content": "soning. arXiv preprint arXiv:1803.03067, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 666, + 505, + 690 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 721, + 254, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 254, + 732 + ], + "score": 1.0, + "content": "preprint arXiv:1611.05397, 2016.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Viz- ´", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "doom: A doom-based ai research platform for visual reinforcement learning. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 219, + 115 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 219, + 115 + ], + "score": 1.0, + "content": "arXiv:1605.02097, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Guillaume Lample and Devendra Singh Chaplot. Playing FPS games with deep reinforcement learn-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 394, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 394, + 146 + ], + "score": 1.0, + "content": "ing. In Thirty-First AAAI Conference on Artificial Intelligence, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 164, + 415, + 177 + ], + "spans": [ + { + "bbox": [ + 117, + 164, + 415, + 177 + ], + "score": 1.0, + "content": "The handbook of brain theory and neural networks, 3361(10):1995, 1995.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 108, + 182, + 503, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 375, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 375, + 217 + ], + "score": 1.0, + "content": "complex environments. arXiv preprint arXiv:1611.03673, 2016.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 503, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "Dipendra K Misra, John Langford, and Yoav Artzi. Mapping instructions and visual observations to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 427, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 427, + 247 + ], + "score": 1.0, + "content": "actions with reinforcement learning. arXiv preprint arXiv:1704.08795, 2017.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 504, + 288 + ], + "lines": [ + { + "bbox": [ + 106, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wier-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 115, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "stra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 276, + 214, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 214, + 286 + ], + "score": 1.0, + "content": "arXiv:1312.5602, 2013.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 504, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 504, + 307 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 116, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 317, + 220, + 329 + ], + "spans": [ + { + "bbox": [ + 116, + 317, + 220, + 329 + ], + "score": 1.0, + "content": "learning. In ICML, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 105, + 335, + 503, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 346, + 441, + 358 + ], + "spans": [ + { + "bbox": [ + 115, + 346, + 441, + 358 + ], + "score": 1.0, + "content": "multi-task deep reinforcement learning. arXiv preprint arXiv:1706.05064, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 105, + 365, + 505, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 377 + ], + "score": 1.0, + "content": "Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 376, + 457, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 376, + 457, + 389 + ], + "score": 1.0, + "content": "reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "Manolis Savva, Angel X. Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. MI-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "NOS: Multimodal indoor simulator for navigation in complex environments. arXiv:1712.03931,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 415, + 143, + 430 + ], + "spans": [ + { + "bbox": [ + 115, + 415, + 143, + 430 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 504, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 450 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 447, + 380, + 459 + ], + "spans": [ + { + "bbox": [ + 116, + 447, + 380, + 459 + ], + "score": 1.0, + "content": "optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 504, + 499 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 476, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 115, + 476, + 506, + 491 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 488, + 493, + 500 + ], + "spans": [ + { + "bbox": [ + 116, + 488, + 493, + 500 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 104, + 506, + 504, + 529 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "Yi Wu, Yuxin Wu, Georgia Gkioxari, and Yuandong Tian. Building generalizable agents with a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 517, + 416, + 530 + ], + "spans": [ + { + "bbox": [ + 116, + 517, + 416, + 530 + ], + "score": 1.0, + "content": "realistic and rich 3d environment. arXiv preprint arXiv:1801.02209, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 535, + 504, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 504, + 550 + ], + "score": 1.0, + "content": "Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial atten-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "tion for visual question answering. In European Conference on Computer Vision, pp. 451–466.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 559, + 181, + 571 + ], + "spans": [ + { + "bbox": [ + 116, + 559, + 181, + 571 + ], + "score": 1.0, + "content": "Springer, 2016.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 109, + 577, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "Haonan Yu, Xiaochen Lian, Haichao Zhang, and Wei Xu. Guided feature transformation (gft):", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "A neural language grounding module for embodied agents. arXiv preprint arXiv:1805.08329,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 597, + 146, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 146, + 612 + ], + "score": 1.0, + "content": "2018a.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 618, + 503, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "Haonan Yu, Haichao Zhang, and Wei Xu. Interactive grounded language acquisition and general-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 117, + 630, + 371, + 641 + ], + "spans": [ + { + "bbox": [ + 117, + 630, + 371, + 641 + ], + "score": 1.0, + "content": "ization in a 2d world. arXiv preprint arXiv:1802.01433, 2018b.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 648, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 117, + 659, + 404, + 672 + ], + "spans": [ + { + "bbox": [ + 117, + 659, + 404, + 672 + ], + "score": 1.0, + "content": "Torralba. The sound of pixels. arXiv preprint arXiv:1804.03160, 2018.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 687, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 114, + 687, + 505, + 703 + ], + "score": 1.0, + "content": "Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 700, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 114, + 700, + 505, + 714 + ], + "score": 1.0, + "content": "In Robotics and Automation (ICRA), 2017 IEEE International Conference on, pp. 3357–3364.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 712, + 168, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 712, + 168, + 722 + ], + "score": 1.0, + "content": "IEEE, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Viz- ´", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "doom: A doom-based ai research platform for visual reinforcement learning. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 219, + 115 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 219, + 115 + ], + "score": 1.0, + "content": "arXiv:1605.02097, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Guillaume Lample and Devendra Singh Chaplot. Playing FPS games with deep reinforcement learn-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 394, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 394, + 146 + ], + "score": 1.0, + "content": "ing. In Thirty-First AAAI Conference on Artificial Intelligence, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 122, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 164, + 415, + 177 + ], + "spans": [ + { + "bbox": [ + 117, + 164, + 415, + 177 + ], + "score": 1.0, + "content": "The handbook of brain theory and neural networks, 3361(10):1995, 1995.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 153, + 505, + 177 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 182, + 503, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 375, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 375, + 217 + ], + "score": 1.0, + "content": "complex environments. arXiv preprint arXiv:1611.03673, 2016.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 182, + 505, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 503, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "Dipendra K Misra, John Langford, and Yoav Artzi. Mapping instructions and visual observations to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 427, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 427, + 247 + ], + "score": 1.0, + "content": "actions with reinforcement learning. arXiv preprint arXiv:1704.08795, 2017.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 223, + 505, + 247 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 504, + 288 + ], + "lines": [ + { + "bbox": [ + 106, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wier-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 115, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "stra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 276, + 214, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 214, + 286 + ], + "score": 1.0, + "content": "arXiv:1312.5602, 2013.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 253, + 505, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 504, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 504, + 307 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 116, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 317, + 220, + 329 + ], + "spans": [ + { + "bbox": [ + 116, + 317, + 220, + 329 + ], + "score": 1.0, + "content": "learning. In ICML, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 295, + 505, + 329 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 335, + 503, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 346, + 441, + 358 + ], + "spans": [ + { + "bbox": [ + 115, + 346, + 441, + 358 + ], + "score": 1.0, + "content": "multi-task deep reinforcement learning. arXiv preprint arXiv:1706.05064, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 335, + 505, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 365, + 505, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 377 + ], + "score": 1.0, + "content": "Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 376, + 457, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 376, + 457, + 389 + ], + "score": 1.0, + "content": "reasoning with a general conditioning layer. arXiv preprint arXiv:1709.07871, 2017.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 366, + 505, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "Manolis Savva, Angel X. Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. MI-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "NOS: Multimodal indoor simulator for navigation in complex environments. arXiv:1712.03931,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 415, + 143, + 430 + ], + "spans": [ + { + "bbox": [ + 115, + 415, + 143, + 430 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 395, + 505, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 504, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 450 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 447, + 380, + 459 + ], + "spans": [ + { + "bbox": [ + 116, + 447, + 380, + 459 + ], + "score": 1.0, + "content": "optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 434, + 505, + 459 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 504, + 499 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 476, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 115, + 476, + 506, + 491 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 488, + 493, + 500 + ], + "spans": [ + { + "bbox": [ + 116, + 488, + 493, + 500 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 465, + 506, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 506, + 504, + 529 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "Yi Wu, Yuxin Wu, Georgia Gkioxari, and Yuandong Tian. Building generalizable agents with a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 517, + 416, + 530 + ], + "spans": [ + { + "bbox": [ + 116, + 517, + 416, + 530 + ], + "score": 1.0, + "content": "realistic and rich 3d environment. arXiv preprint arXiv:1801.02209, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 506, + 506, + 530 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 535, + 504, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 504, + 550 + ], + "score": 1.0, + "content": "Huijuan Xu and Kate Saenko. Ask, attend and answer: Exploring question-guided spatial atten-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "tion for visual question answering. In European Conference on Computer Vision, pp. 451–466.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 559, + 181, + 571 + ], + "spans": [ + { + "bbox": [ + 116, + 559, + 181, + 571 + ], + "score": 1.0, + "content": "Springer, 2016.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 535, + 504, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 577, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "Haonan Yu, Xiaochen Lian, Haichao Zhang, and Wei Xu. Guided feature transformation (gft):", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "A neural language grounding module for embodied agents. arXiv preprint arXiv:1805.08329,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 597, + 146, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 597, + 146, + 612 + ], + "score": 1.0, + "content": "2018a.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 576, + 506, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 618, + 503, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "Haonan Yu, Haichao Zhang, and Wei Xu. Interactive grounded language acquisition and general-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 117, + 630, + 371, + 641 + ], + "spans": [ + { + "bbox": [ + 117, + 630, + 371, + 641 + ], + "score": 1.0, + "content": "ization in a 2d world. arXiv preprint arXiv:1802.01433, 2018b.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 618, + 505, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 648, + 504, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 117, + 659, + 404, + 672 + ], + "spans": [ + { + "bbox": [ + 117, + 659, + 404, + 672 + ], + "score": 1.0, + "content": "Torralba. The sound of pixels. arXiv preprint arXiv:1804.03160, 2018.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 648, + 505, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 687, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 114, + 687, + 505, + 703 + ], + "score": 1.0, + "content": "Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 700, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 114, + 700, + 505, + 714 + ], + "score": 1.0, + "content": "In Robotics and Automation (ICRA), 2017 IEEE International Conference on, pp. 3357–3364.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 116, + 712, + 168, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 712, + 168, + 722 + ], + "score": 1.0, + "content": "IEEE, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 678, + 505, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 217, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 218, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 218, + 97 + ], + "score": 1.0, + "content": "A VISUALIZATIONS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 115, + 114, + 500, + 518 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 114, + 500, + 518 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 114, + 500, + 518 + ], + "spans": [ + { + "bbox": [ + 115, + 114, + 500, + 518 + ], + "score": 0.974, + "type": "image", + "image_path": "eb82cf3d410040af24e674376db93bf154e95616c9f3df8c13dcc8542653e7ae.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 115, + 114, + 500, + 248.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 248.66666666666666, + 500, + 383.3333333333333 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 115, + 383.3333333333333, + 500, + 518.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 526, + 505, + 637 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "Figure 9: Visualizations of convolutional output channels. We visualize the convolutional channels corre-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 537, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 547 + ], + "score": 1.0, + "content": "sponding to 7 words (one in each row) for the same frame (shown in the rightmost column). The first column", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "shows the auxiliary task labels for reference. The second column and third column show the output of the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "corresponding channel for the proposed Dual-Attention model trained without and with auxiliary tasks, respec-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "tively. As expected, the Aux model outputs are very close to the auxiliary task labels. The convolutional outputs", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "score": 1.0, + "content": "of the No Aux model show that words and objects/properties in the images have been properly aligned even", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "when the model is not trained with any auxiliary task labels. We do not provide any auxiliary label for words", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "‘smallest’ and ‘largest’ as they are not properties of an object and require relative comparison of objects. The", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 606, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 506, + 618 + ], + "score": 1.0, + "content": "visualizations in row 5 (corresponding to ‘smallest’) indicate that both models are able to compare the sizes of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "objects and detect the smallest object in the corresponding output channel even without any aux labels for the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 626, + 164, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 164, + 638 + ], + "score": 1.0, + "content": "smallest object.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9 + } + ], + "index": 5.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 217, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 218, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 218, + 97 + ], + "score": 1.0, + "content": "A VISUALIZATIONS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 115, + 114, + 500, + 518 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 114, + 500, + 518 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 114, + 500, + 518 + ], + "spans": [ + { + "bbox": [ + 115, + 114, + 500, + 518 + ], + "score": 0.974, + "type": "image", + "image_path": "eb82cf3d410040af24e674376db93bf154e95616c9f3df8c13dcc8542653e7ae.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 115, + 114, + 500, + 248.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 248.66666666666666, + 500, + 383.3333333333333 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 115, + 383.3333333333333, + 500, + 518.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 526, + 505, + 637 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "Figure 9: Visualizations of convolutional output channels. We visualize the convolutional channels corre-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 537, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 547 + ], + "score": 1.0, + "content": "sponding to 7 words (one in each row) for the same frame (shown in the rightmost column). The first column", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "shows the auxiliary task labels for reference. The second column and third column show the output of the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "corresponding channel for the proposed Dual-Attention model trained without and with auxiliary tasks, respec-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "tively. As expected, the Aux model outputs are very close to the auxiliary task labels. The convolutional outputs", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 588 + ], + "score": 1.0, + "content": "of the No Aux model show that words and objects/properties in the images have been properly aligned even", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "when the model is not trained with any auxiliary task labels. We do not provide any auxiliary label for words", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "‘smallest’ and ‘largest’ as they are not properties of an object and require relative comparison of objects. The", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 606, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 506, + 618 + ], + "score": 1.0, + "content": "visualizations in row 5 (corresponding to ‘smallest’) indicate that both models are able to compare the sizes of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "objects and detect the smallest object in the corresponding output channel even without any aux labels for the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 626, + 164, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 164, + 638 + ], + "score": 1.0, + "content": "smallest object.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9 + } + ], + "index": 5.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 227, + 497, + 458 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 227, + 497, + 458 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 227, + 497, + 458 + ], + "spans": [ + { + "bbox": [ + 108, + 227, + 497, + 458 + ], + "score": 0.967, + "type": "image", + "image_path": "c09df43f375517b1e7308882219ccaabee4d2f03fc7563c5a1c2705c93dc07be.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 227, + 497, + 304.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 304.0, + 497, + 381.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 381.0, + 497, + 458.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 466, + 505, + 587 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "Figure 10: Spatial Attention and Answer Prediction Visualizations. An example EQA episode with the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "score": 1.0, + "content": "question “Which is the smallest blue object?”. The sentence embedding of the question is shown on the top", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 485, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 107, + 487, + 133, + 497 + ], + "score": 0.87, + "content": "( x _ { s e n t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 485, + 506, + 500 + ], + "score": 1.0, + "content": ". As expected, the embedding attends to object type words (’torch’, ’pillar’, ’skullkey’, etc.) as the ques-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "tion is asking about an object type (’Which object’). The rows show increasing time steps and columns show", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "the input frame, the input frame overlaid with the spatial attention map, the predicted answer distribution, and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "the action at each time step. As the agent is turning, the spatial attention attends to small and blue objects. Time", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 526, + 504, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 504, + 537 + ], + "score": 1.0, + "content": "steps 1, 2: The model is attending to the yellow skullkey but the probability of the answer is not sufficiently", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 536, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 506, + 547 + ], + "score": 1.0, + "content": "high, likely because the skullkey is not blue. Time step 3: The model cannot see the skullkey anymore so it", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 547, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 558 + ], + "score": 1.0, + "content": "attends to the armor which is next smallest object. Consequently, the answer prediction also predicts armor, but", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 556, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 506, + 567 + ], + "score": 1.0, + "content": "the policy decides not to answer due to low probability. Time step 4: As the agent turns more, it observes and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 566, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 577 + ], + "score": 1.0, + "content": "attends to the blue skullkey. The answer prediction has high probability for skullkey as it’s small and blue and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 577, + 258, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 258, + 588 + ], + "score": 1.0, + "content": "the policy decides to answer the question.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5 + } + ], + "index": 4.75 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 227, + 497, + 458 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 227, + 497, + 458 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 227, + 497, + 458 + ], + "spans": [ + { + "bbox": [ + 108, + 227, + 497, + 458 + ], + "score": 0.967, + "type": "image", + "image_path": "c09df43f375517b1e7308882219ccaabee4d2f03fc7563c5a1c2705c93dc07be.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 227, + 497, + 304.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 304.0, + 497, + 381.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 381.0, + 497, + 458.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 466, + 505, + 587 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "Figure 10: Spatial Attention and Answer Prediction Visualizations. An example EQA episode with the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 489 + ], + "score": 1.0, + "content": "question “Which is the smallest blue object?”. The sentence embedding of the question is shown on the top", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 485, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 107, + 487, + 133, + 497 + ], + "score": 0.87, + "content": "( x _ { s e n t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 485, + 506, + 500 + ], + "score": 1.0, + "content": ". As expected, the embedding attends to object type words (’torch’, ’pillar’, ’skullkey’, etc.) as the ques-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "tion is asking about an object type (’Which object’). The rows show increasing time steps and columns show", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "the input frame, the input frame overlaid with the spatial attention map, the predicted answer distribution, and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "the action at each time step. As the agent is turning, the spatial attention attends to small and blue objects. Time", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 526, + 504, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 504, + 537 + ], + "score": 1.0, + "content": "steps 1, 2: The model is attending to the yellow skullkey but the probability of the answer is not sufficiently", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 536, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 506, + 547 + ], + "score": 1.0, + "content": "high, likely because the skullkey is not blue. Time step 3: The model cannot see the skullkey anymore so it", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 547, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 558 + ], + "score": 1.0, + "content": "attends to the armor which is next smallest object. Consequently, the answer prediction also predicts armor, but", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 556, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 506, + 567 + ], + "score": 1.0, + "content": "the policy decides not to answer due to low probability. Time step 4: As the agent turns more, it observes and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 566, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 577 + ], + "score": 1.0, + "content": "attends to the blue skullkey. The answer prediction has high probability for skullkey as it’s small and blue and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 577, + 258, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 258, + 588 + ], + "score": 1.0, + "content": "the policy decides to answer the question.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5 + } + ], + "index": 4.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 79, + 483, + 148 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 79, + 483, + 148 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 79, + 483, + 148 + ], + "spans": [ + { + "bbox": [ + 129, + 79, + 483, + 148 + ], + "score": 0.966, + "type": "image", + "image_path": "6acfbbc3bc1a92d1814e01f282a9c869e31804d2fe115c6a5a2a0783ad26bf0f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 79, + 483, + 102.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 102.0, + 483, + 125.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 125.0, + 483, + 148.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 222, + 159, + 389, + 170 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 220, + 157, + 390, + 172 + ], + "spans": [ + { + "bbox": [ + 220, + 157, + 390, + 172 + ], + "score": 1.0, + "content": "Figure 11: Architecture of the policy module.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 193, + 327, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 329, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 329, + 208 + ], + "score": 1.0, + "content": "B ADDITIONAL EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 109, + 221, + 324, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 220, + 327, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 327, + 234 + ], + "score": 1.0, + "content": "B.1 HYPERPARAMETERS AND NETWORK DETAILS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 250, + 256 + ], + "score": 1.0, + "content": "The input image is rescaled to size", + "type": "text" + }, + { + "bbox": [ + 251, + 244, + 313, + 254 + ], + "score": 0.9, + "content": "3 \\times 1 6 8 \\times 3 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 243, + 505, + 256 + ], + "score": 1.0, + "content": ". The convolutional network for processing the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 255, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 360, + 266 + ], + "score": 1.0, + "content": "image consisted of 3 convolutional layers: conv1 containing", + "type": "text" + }, + { + "bbox": [ + 360, + 255, + 391, + 265 + ], + "score": 0.67, + "content": "3 2 ~ 8 \\mathrm { x } 8", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 255, + 506, + 266 + ], + "score": 1.0, + "content": "filters with stride 4, conv2", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 266, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 151, + 277 + ], + "score": 1.0, + "content": "containing", + "type": "text" + }, + { + "bbox": [ + 152, + 266, + 182, + 276 + ], + "score": 0.69, + "content": "6 4 ~ 4 \\mathrm { x } 4", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 266, + 357, + 277 + ], + "score": 1.0, + "content": "filters with stride 2, and conv3 containing", + "type": "text" + }, + { + "bbox": [ + 357, + 266, + 367, + 276 + ], + "score": 0.51, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 266, + 506, + 277 + ], + "score": 1.0, + "content": "3x3 filters with stride 2. We use", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "ReLU activations for conv1 and conv2 and sigmoid for conv3, as its output is used as auxiliary task", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "predictions directly. We use word embeddings and GRU of size 32 followed by linear layer of size", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 116, + 308 + ], + "score": 0.73, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "to get the sentence-level representation. The policy module uses hidden dimension 128 for the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 309, + 263, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 263, + 322 + ], + "score": 1.0, + "content": "linear and GRU layers (see Figure 11).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 326, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "For reinforcement learning, we use Proximal Policy Optimization (PPO) with 8 actors and a time", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "horizon of 128 steps. We use a single batch with 4 PPO epochs. The clipping parameter for PPO is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 228, + 361 + ], + "score": 1.0, + "content": "set to 0.2. The discount factor", + "type": "text" + }, + { + "bbox": [ + 228, + 349, + 241, + 360 + ], + "score": 0.81, + "content": "( \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 348, + 450, + 361 + ], + "score": 1.0, + "content": "is 0.99. We used Adam optimizer with learning rate", + "type": "text" + }, + { + "bbox": [ + 451, + 348, + 478, + 359 + ], + "score": 0.39, + "content": "2 . 5 \\mathrm { e } { \\cdot } 4", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "for all", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 360, + 160, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 160, + 371 + ], + "score": 1.0, + "content": "experiments.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 388, + 215, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 387, + 217, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 217, + 401 + ], + "score": 1.0, + "content": "B.2 BASELINE DETAILS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 504, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 388, + 424 + ], + "score": 1.0, + "content": "Image only: Naive baseline of just using the image representation:", + "type": "text" + }, + { + "bbox": [ + 388, + 411, + 448, + 423 + ], + "score": 0.92, + "content": "x _ { \\mathrm { S } } = \\sec ( x _ { I } )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 410, + 504, + 424 + ], + "score": 1.0, + "content": "where vec(.)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 421, + 236, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 236, + 434 + ], + "score": 1.0, + "content": "denotes the flattening operation.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 108, + 438, + 451, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 448, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 375, + 453 + ], + "score": 1.0, + "content": "Text only: Naive baseline of just using the textual representations:", + "type": "text" + }, + { + "bbox": [ + 375, + 438, + 448, + 451 + ], + "score": 0.9, + "content": "x _ { \\mathrm { S } } = [ x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } ]", + "type": "inline_equation" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 369, + 468 + ], + "score": 1.0, + "content": "Concat: The image and textual representations are concatenated:", + "type": "text" + }, + { + "bbox": [ + 370, + 455, + 479, + 468 + ], + "score": 0.89, + "content": "x _ { \\mathrm { S } } = [ \\mathrm { v e c } ( x _ { I } ) , x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 455, + 505, + 468 + ], + "score": 1.0, + "content": ". Note", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "that concatenation is the most common method of combining representations. Hermann et al. (2017)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 478, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 489 + ], + "score": 1.0, + "content": "concatenate convolutional image and bag-of-words textual representations for SGN, whereas Misra", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 488, + 405, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 405, + 500 + ], + "score": 1.0, + "content": "et al. (2017) use concatenation with sentence-level textual representations.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 505, + 503, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 504, + 518 + ], + "score": 1.0, + "content": "Gated-Attention: Adapted from Chaplot et al. (2017), who used Gated-Attention with sentence-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 515, + 343, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 259, + 530 + ], + "score": 1.0, + "content": "level textual representations for SGN:", + "type": "text" + }, + { + "bbox": [ + 260, + 516, + 339, + 528 + ], + "score": 0.92, + "content": "x _ { \\mathrm { S } } = f _ { \\mathrm { G A } } ( x _ { I } , x _ { \\mathrm { s e n t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 515, + 343, + 530 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "FiLM: Perez et al. (2017) introduced a general-purpose conditioning method called Feature-wise", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 410, + 558 + ], + "score": 1.0, + "content": "Linear Modulation (FiLM) for Visual Question Answering. Using FiLM,", + "type": "text" + }, + { + "bbox": [ + 411, + 544, + 505, + 556 + ], + "score": 0.9, + "content": "x _ { \\mathrm { S } } = \\gamma ( x _ { \\mathrm { s e n t } } ) \\odot x _ { I } +", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 555, + 475, + 568 + ], + "spans": [ + { + "bbox": [ + 107, + 555, + 139, + 567 + ], + "score": 0.91, + "content": "\\beta ( x _ { \\mathrm { s e n t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 555, + 167, + 568 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 167, + 555, + 199, + 567 + ], + "score": 0.92, + "content": "\\gamma ( x _ { \\mathrm { s e n t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 555, + 217, + 568 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 218, + 555, + 250, + 567 + ], + "score": 0.92, + "content": "\\beta ( x _ { \\mathrm { s e n t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 555, + 475, + 568 + ], + "score": 1.0, + "content": "are learnable projections of the sentence representation.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "score": 1.0, + "content": "PACMAN: Das et al. (2017) presented a hierarchical RL model for EQA. We adapt their method", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "by using the attention mechanism in their QA module, which takes the last 5 frames and the text", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 592, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 506, + 608 + ], + "score": 1.0, + "content": "as input, and computes the similarity of the text with each frame using dot products between image", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "score": 1.0, + "content": "and sentence-level text representations. These similarities are converted into attention weights using", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "softmax, and the attention-weighted image features are concatenated with question embedding and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 626, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 641 + ], + "score": 1.0, + "content": "passed through a softmax classifier to predict the answer distribution. For this particular baseline,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "we use the last 5 frames as input at each time step, unlike the proposed model and all other baselines", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "which use a single frame as input. The attention-weighted image features are used as the state", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "representation. The PACMAN model used a pretrained QA module, but we train this module jointly", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 670, + 398, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 398, + 683 + ], + "score": 1.0, + "content": "with the Navigation model for fair comparison with the proposed model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 401, + 700 + ], + "score": 1.0, + "content": "For each of the above method except PACMAN, we use a linear layer", + "type": "text" + }, + { + "bbox": [ + 401, + 688, + 408, + 699 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "with ReLU activations", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 194, + 712 + ], + "score": 1.0, + "content": "followed by softmax", + "type": "text" + }, + { + "bbox": [ + 195, + 701, + 202, + 709 + ], + "score": 0.73, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 698, + 239, + 712 + ], + "score": 1.0, + "content": "to get a", + "type": "text" + }, + { + "bbox": [ + 240, + 699, + 248, + 709 + ], + "score": 0.81, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "-dimensional answer prediction from the state representations:", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 198, + 722 + ], + "score": 0.87, + "content": "x _ { \\mathrm { A n s } } = \\sigma ( \\bar { f } ( x _ { \\mathrm { S } } ; \\theta _ { L i n } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 710, + 204, + 723 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 205, + 711, + 216, + 721 + ], + "score": 0.68, + "content": "x _ { \\mathrm { { S } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 710, + 235, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 235, + 712, + 257, + 721 + ], + "score": 0.88, + "content": "x _ { A n s }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "are concatenated and passed to the policy module along with", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 389, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 389, + 733 + ], + "score": 1.0, + "content": "the time step and task indicator variable just as in the proposed model.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 79, + 483, + 148 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 79, + 483, + 148 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 79, + 483, + 148 + ], + "spans": [ + { + "bbox": [ + 129, + 79, + 483, + 148 + ], + "score": 0.966, + "type": "image", + "image_path": "6acfbbc3bc1a92d1814e01f282a9c869e31804d2fe115c6a5a2a0783ad26bf0f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 79, + 483, + 102.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 102.0, + 483, + 125.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 125.0, + 483, + 148.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 222, + 159, + 389, + 170 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 220, + 157, + 390, + 172 + ], + "spans": [ + { + "bbox": [ + 220, + 157, + 390, + 172 + ], + "score": 1.0, + "content": "Figure 11: Architecture of the policy module.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 193, + 327, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 329, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 329, + 208 + ], + "score": 1.0, + "content": "B ADDITIONAL EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 109, + 221, + 324, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 220, + 327, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 327, + 234 + ], + "score": 1.0, + "content": "B.1 HYPERPARAMETERS AND NETWORK DETAILS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 250, + 256 + ], + "score": 1.0, + "content": "The input image is rescaled to size", + "type": "text" + }, + { + "bbox": [ + 251, + 244, + 313, + 254 + ], + "score": 0.9, + "content": "3 \\times 1 6 8 \\times 3 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 243, + 505, + 256 + ], + "score": 1.0, + "content": ". The convolutional network for processing the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 255, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 360, + 266 + ], + "score": 1.0, + "content": "image consisted of 3 convolutional layers: conv1 containing", + "type": "text" + }, + { + "bbox": [ + 360, + 255, + 391, + 265 + ], + "score": 0.67, + "content": "3 2 ~ 8 \\mathrm { x } 8", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 255, + 506, + 266 + ], + "score": 1.0, + "content": "filters with stride 4, conv2", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 266, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 151, + 277 + ], + "score": 1.0, + "content": "containing", + "type": "text" + }, + { + "bbox": [ + 152, + 266, + 182, + 276 + ], + "score": 0.69, + "content": "6 4 ~ 4 \\mathrm { x } 4", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 266, + 357, + 277 + ], + "score": 1.0, + "content": "filters with stride 2, and conv3 containing", + "type": "text" + }, + { + "bbox": [ + 357, + 266, + 367, + 276 + ], + "score": 0.51, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 266, + 506, + 277 + ], + "score": 1.0, + "content": "3x3 filters with stride 2. We use", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "ReLU activations for conv1 and conv2 and sigmoid for conv3, as its output is used as auxiliary task", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "predictions directly. We use word embeddings and GRU of size 32 followed by linear layer of size", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 116, + 308 + ], + "score": 0.73, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "to get the sentence-level representation. The policy module uses hidden dimension 128 for the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 309, + 263, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 263, + 322 + ], + "score": 1.0, + "content": "linear and GRU layers (see Figure 11).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 243, + 506, + 322 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 326, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "For reinforcement learning, we use Proximal Policy Optimization (PPO) with 8 actors and a time", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "horizon of 128 steps. We use a single batch with 4 PPO epochs. The clipping parameter for PPO is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 228, + 361 + ], + "score": 1.0, + "content": "set to 0.2. The discount factor", + "type": "text" + }, + { + "bbox": [ + 228, + 349, + 241, + 360 + ], + "score": 0.81, + "content": "( \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 348, + 450, + 361 + ], + "score": 1.0, + "content": "is 0.99. We used Adam optimizer with learning rate", + "type": "text" + }, + { + "bbox": [ + 451, + 348, + 478, + 359 + ], + "score": 0.39, + "content": "2 . 5 \\mathrm { e } { \\cdot } 4", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "for all", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 360, + 160, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 160, + 371 + ], + "score": 1.0, + "content": "experiments.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 326, + 506, + 371 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 388, + 215, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 387, + 217, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 217, + 401 + ], + "score": 1.0, + "content": "B.2 BASELINE DETAILS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 504, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 388, + 424 + ], + "score": 1.0, + "content": "Image only: Naive baseline of just using the image representation:", + "type": "text" + }, + { + "bbox": [ + 388, + 411, + 448, + 423 + ], + "score": 0.92, + "content": "x _ { \\mathrm { S } } = \\sec ( x _ { I } )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 410, + 504, + 424 + ], + "score": 1.0, + "content": "where vec(.)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 421, + 236, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 236, + 434 + ], + "score": 1.0, + "content": "denotes the flattening operation.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 410, + 504, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 438, + 451, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 448, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 375, + 453 + ], + "score": 1.0, + "content": "Text only: Naive baseline of just using the textual representations:", + "type": "text" + }, + { + "bbox": [ + 375, + 438, + 448, + 451 + ], + "score": 0.9, + "content": "x _ { \\mathrm { S } } = [ x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } ]", + "type": "inline_equation" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 435, + 448, + 453 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 369, + 468 + ], + "score": 1.0, + "content": "Concat: The image and textual representations are concatenated:", + "type": "text" + }, + { + "bbox": [ + 370, + 455, + 479, + 468 + ], + "score": 0.89, + "content": "x _ { \\mathrm { S } } = [ \\mathrm { v e c } ( x _ { I } ) , x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 455, + 505, + 468 + ], + "score": 1.0, + "content": ". Note", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "that concatenation is the most common method of combining representations. Hermann et al. (2017)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 478, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 489 + ], + "score": 1.0, + "content": "concatenate convolutional image and bag-of-words textual representations for SGN, whereas Misra", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 488, + 405, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 405, + 500 + ], + "score": 1.0, + "content": "et al. (2017) use concatenation with sentence-level textual representations.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 455, + 506, + 500 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 505, + 503, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 504, + 518 + ], + "score": 1.0, + "content": "Gated-Attention: Adapted from Chaplot et al. (2017), who used Gated-Attention with sentence-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 515, + 343, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 259, + 530 + ], + "score": 1.0, + "content": "level textual representations for SGN:", + "type": "text" + }, + { + "bbox": [ + 260, + 516, + 339, + 528 + ], + "score": 0.92, + "content": "x _ { \\mathrm { S } } = f _ { \\mathrm { G A } } ( x _ { I } , x _ { \\mathrm { s e n t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 515, + 343, + 530 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 504, + 504, + 530 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "FiLM: Perez et al. (2017) introduced a general-purpose conditioning method called Feature-wise", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 410, + 558 + ], + "score": 1.0, + "content": "Linear Modulation (FiLM) for Visual Question Answering. Using FiLM,", + "type": "text" + }, + { + "bbox": [ + 411, + 544, + 505, + 556 + ], + "score": 0.9, + "content": "x _ { \\mathrm { S } } = \\gamma ( x _ { \\mathrm { s e n t } } ) \\odot x _ { I } +", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 555, + 475, + 568 + ], + "spans": [ + { + "bbox": [ + 107, + 555, + 139, + 567 + ], + "score": 0.91, + "content": "\\beta ( x _ { \\mathrm { s e n t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 555, + 167, + 568 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 167, + 555, + 199, + 567 + ], + "score": 0.92, + "content": "\\gamma ( x _ { \\mathrm { s e n t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 555, + 217, + 568 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 218, + 555, + 250, + 567 + ], + "score": 0.92, + "content": "\\beta ( x _ { \\mathrm { s e n t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 555, + 475, + 568 + ], + "score": 1.0, + "content": "are learnable projections of the sentence representation.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 532, + 505, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "score": 1.0, + "content": "PACMAN: Das et al. (2017) presented a hierarchical RL model for EQA. We adapt their method", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "by using the attention mechanism in their QA module, which takes the last 5 frames and the text", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 592, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 506, + 608 + ], + "score": 1.0, + "content": "as input, and computes the similarity of the text with each frame using dot products between image", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "score": 1.0, + "content": "and sentence-level text representations. These similarities are converted into attention weights using", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "softmax, and the attention-weighted image features are concatenated with question embedding and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 626, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 641 + ], + "score": 1.0, + "content": "passed through a softmax classifier to predict the answer distribution. For this particular baseline,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "we use the last 5 frames as input at each time step, unlike the proposed model and all other baselines", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "which use a single frame as input. The attention-weighted image features are used as the state", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "representation. The PACMAN model used a pretrained QA module, but we train this module jointly", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 670, + 398, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 398, + 683 + ], + "score": 1.0, + "content": "with the Navigation model for fair comparison with the proposed model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 571, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 401, + 700 + ], + "score": 1.0, + "content": "For each of the above method except PACMAN, we use a linear layer", + "type": "text" + }, + { + "bbox": [ + 401, + 688, + 408, + 699 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "with ReLU activations", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 194, + 712 + ], + "score": 1.0, + "content": "followed by softmax", + "type": "text" + }, + { + "bbox": [ + 195, + 701, + 202, + 709 + ], + "score": 0.73, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 698, + 239, + 712 + ], + "score": 1.0, + "content": "to get a", + "type": "text" + }, + { + "bbox": [ + 240, + 699, + 248, + 709 + ], + "score": 0.81, + "content": "V", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "-dimensional answer prediction from the state representations:", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 198, + 722 + ], + "score": 0.87, + "content": "x _ { \\mathrm { A n s } } = \\sigma ( \\bar { f } ( x _ { \\mathrm { S } } ; \\theta _ { L i n } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 710, + 204, + 723 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 205, + 711, + 216, + 721 + ], + "score": 0.68, + "content": "x _ { \\mathrm { { S } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 710, + 235, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 235, + 712, + 257, + 721 + ], + "score": 0.88, + "content": "x _ { A n s }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "are concatenated and passed to the policy module along with", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 389, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 389, + 733 + ], + "score": 1.0, + "content": "the time step and task indicator variable just as in the proposed model.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 79, + 465, + 181 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 79, + 465, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 79, + 465, + 181 + ], + "spans": [ + { + "bbox": [ + 146, + 79, + 465, + 181 + ], + "score": 0.968, + "type": "image", + "image_path": "a6315480bb9a073f24a9632c6856b7874a62864e259308fb0863e2161017b386.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 146, + 79, + 465, + 113.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 146, + 113.0, + 465, + 147.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 146, + 147.0, + 465, + 181.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 188, + 505, + 208 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "Figure 12: Plot showing the training accuracy of all the models without auxiliary tasks for Doom Easy and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 197, + 180, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 180, + 209 + ], + "score": 1.0, + "content": "Hard environments.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 146, + 220, + 465, + 321 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 220, + 465, + 321 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 146, + 220, + 465, + 321 + ], + "spans": [ + { + "bbox": [ + 146, + 220, + 465, + 321 + ], + "score": 0.972, + "type": "image", + "image_path": "634ffdf84813ca11c0e122e1a00e556fb733df50e54d4f6c9664f00480c451dd.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 146, + 220, + 465, + 253.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 146, + 253.66666666666666, + 465, + 287.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 146, + 287.3333333333333, + 465, + 321.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 329, + 504, + 350 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "Figure 13: Plot showing the training accuracy of 3 models across 3 training runs with different seeds with and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 338, + 382, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 382, + 352 + ], + "score": 1.0, + "content": "without auxiliary tasks for Doom Easy environment without any smoothing.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 108, + 369, + 280, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 282, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 282, + 385 + ], + "score": 1.0, + "content": "C DOOM EXPERIMENT DETAILS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "Additional Results. We show the training accuracy of all models without auxiliary tasks for Doom", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Easy and Hard in Figure 12. We also show the training accuracy of 3 models across 3 training runs", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 416, + 488, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 488, + 430 + ], + "score": 1.0, + "content": "with different seeds with and without auxiliary tasks for Doom Easy environment in Figure 13.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "Dataset. The Doom objects used in our experiments are illustrated in Figure 14. Instructions and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 445, + 363, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 363, + 456 + ], + "score": 1.0, + "content": "questions used for training and evaluation are listed in Tables 4.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "image", + "bbox": [ + 187, + 467, + 420, + 576 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 187, + 467, + 420, + 576 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 187, + 467, + 420, + 576 + ], + "spans": [ + { + "bbox": [ + 187, + 467, + 420, + 576 + ], + "score": 0.969, + "type": "image", + "image_path": "69d7bb5a3ad8695b8a37ae82bea6b9ec1dbc83f181543eb44b38a763cd846842.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 467, + 420, + 480.625 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 187, + 480.625, + 420, + 494.25 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 187, + 494.25, + 420, + 507.875 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 187, + 507.875, + 420, + 521.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 187, + 521.5, + 420, + 535.125 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 187, + 535.125, + 420, + 548.75 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 187, + 548.75, + 420, + 562.375 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 187, + 562.375, + 420, + 576.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 156, + 586, + 454, + 598 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 155, + 585, + 455, + 599 + ], + "spans": [ + { + "bbox": [ + 155, + 585, + 455, + 599 + ], + "score": 1.0, + "content": "Figure 14: Objects of various colors and sizes used in the ViZDoom environment.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + } + ], + "index": 21.75 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 79, + 465, + 181 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 79, + 465, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 79, + 465, + 181 + ], + "spans": [ + { + "bbox": [ + 146, + 79, + 465, + 181 + ], + "score": 0.968, + "type": "image", + "image_path": "a6315480bb9a073f24a9632c6856b7874a62864e259308fb0863e2161017b386.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 146, + 79, + 465, + 113.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 146, + 113.0, + 465, + 147.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 146, + 147.0, + 465, + 181.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 188, + 505, + 208 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "Figure 12: Plot showing the training accuracy of all the models without auxiliary tasks for Doom Easy and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 197, + 180, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 180, + 209 + ], + "score": 1.0, + "content": "Hard environments.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 146, + 220, + 465, + 321 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 220, + 465, + 321 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 146, + 220, + 465, + 321 + ], + "spans": [ + { + "bbox": [ + 146, + 220, + 465, + 321 + ], + "score": 0.972, + "type": "image", + "image_path": "634ffdf84813ca11c0e122e1a00e556fb733df50e54d4f6c9664f00480c451dd.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 146, + 220, + 465, + 253.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 146, + 253.66666666666666, + 465, + 287.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 146, + 287.3333333333333, + 465, + 321.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 329, + 504, + 350 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "Figure 13: Plot showing the training accuracy of 3 models across 3 training runs with different seeds with and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 338, + 382, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 382, + 352 + ], + "score": 1.0, + "content": "without auxiliary tasks for Doom Easy environment without any smoothing.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 108, + 369, + 280, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 282, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 282, + 385 + ], + "score": 1.0, + "content": "C DOOM EXPERIMENT DETAILS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "Additional Results. We show the training accuracy of all models without auxiliary tasks for Doom", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Easy and Hard in Figure 12. We also show the training accuracy of 3 models across 3 training runs", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 416, + 488, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 488, + 430 + ], + "score": 1.0, + "content": "with different seeds with and without auxiliary tasks for Doom Easy environment in Figure 13.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 394, + 505, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "Dataset. The Doom objects used in our experiments are illustrated in Figure 14. Instructions and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 445, + 363, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 363, + 456 + ], + "score": 1.0, + "content": "questions used for training and evaluation are listed in Tables 4.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 433, + 505, + 456 + ] + }, + { + "type": "image", + "bbox": [ + 187, + 467, + 420, + 576 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 187, + 467, + 420, + 576 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 187, + 467, + 420, + 576 + ], + "spans": [ + { + "bbox": [ + 187, + 467, + 420, + 576 + ], + "score": 0.969, + "type": "image", + "image_path": "69d7bb5a3ad8695b8a37ae82bea6b9ec1dbc83f181543eb44b38a763cd846842.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 467, + 420, + 480.625 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 187, + 480.625, + 420, + 494.25 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 187, + 494.25, + 420, + 507.875 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 187, + 507.875, + 420, + 521.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 187, + 521.5, + 420, + 535.125 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 187, + 535.125, + 420, + 548.75 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 187, + 548.75, + 420, + 562.375 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 187, + 562.375, + 420, + 576.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 156, + 586, + 454, + 598 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 155, + 585, + 455, + 599 + ], + "spans": [ + { + "bbox": [ + 155, + 585, + 455, + 599 + ], + "score": 1.0, + "content": "Figure 14: Objects of various colors and sizes used in the ViZDoom environment.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + } + ], + "index": 21.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 280, + 495, + 571 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 234, + 504, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 234, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 245 + ], + "score": 1.0, + "content": "Table 4: Instructions and questions for ViZDoom experiments. We used 5 object classes (torch, pillar, keycard,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "skullkey, armor), 4 colors (red, green, blue, yellow), 2 sizes (tall, short), and 2 superlative sizes (smallest,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 253, + 139, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 139, + 267 + ], + "score": 1.0, + "content": "largest).", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 117, + 280, + 495, + 571 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 280, + 495, + 571 + ], + "spans": [ + { + "bbox": [ + 117, + 280, + 495, + 571 + ], + "score": 0.974, + "html": "
SGN Instruction Type42 Train Instructions: Not containing‘red'&‘pillar'28 Test Instructions: Containing‘red'or‘pillar'
Go to the (object).Go to the color>object.torch,keycard,skullkey,armor yellow, green, bluepillar red
Go to the size)object. Go to the {color) {object).tall, short blue torch, green torch, green armor,red torch,red skullkey,red pillar,
blue skullkey,blue keycard, yellow keycard, yellow skullkeygreen pillar,red keycard,red armor
Go to the {size) (object). Go to the color) (size)object.short torch,tall torch green tall, blue tall, blue short,tall pillar, short pillar red short, red tall
Go to the size){color)object.green short tall green, tall blue, short blue,short red, tall red
Go to the {color)(size)(object).short green green tall torch, green short torch,red short pillar,red short torch,
blue short torch,blue tall torchred tall pillar, green tall pillar, red tall torch, green short pillar
Go to the size) {color> {object).tall green torch,short green torch, short blue torch,tall blue torchshort red pillar, short red torch, tall red pillar, tall green pillar, tall red torch,short green pillar
Go to the superlative)object. Go to the {superlative> (color) object.largest, smallest smallest yellow, smallest blue, smallest green,largest blue,largest red,smallest red
EQAQuestion Typelargest green,largest yellow 21 Train Questions: Not containing‘blue'&‘torch'8 Test Questions: Containing‘blue'or‘torch'
Which {size)object is {color)in color?What color is the {object)? What color is the {size){object>? Which object is {color) in color?pillar, keycard, skullkey,armor short pillar, tall pillar red,yellow, greentorch short torch, tall torch blue
", + "type": "table", + "image_path": "7057d42e3744461df649f3b29df5a3d9702f5c717d595cdcc3987488ef779a3c.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 117, + 280, + 495, + 377.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 117, + 377.0, + 495, + 474.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 117, + 474.0, + 495, + 571.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 280, + 495, + 571 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 234, + 504, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 234, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 245 + ], + "score": 1.0, + "content": "Table 4: Instructions and questions for ViZDoom experiments. We used 5 object classes (torch, pillar, keycard,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "skullkey, armor), 4 colors (red, green, blue, yellow), 2 sizes (tall, short), and 2 superlative sizes (smallest,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 253, + 139, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 139, + 267 + ], + "score": 1.0, + "content": "largest).", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 117, + 280, + 495, + 571 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 280, + 495, + 571 + ], + "spans": [ + { + "bbox": [ + 117, + 280, + 495, + 571 + ], + "score": 0.974, + "html": "
SGN Instruction Type42 Train Instructions: Not containing‘red'&‘pillar'28 Test Instructions: Containing‘red'or‘pillar'
Go to the (object).Go to the color>object.torch,keycard,skullkey,armor yellow, green, bluepillar red
Go to the size)object. Go to the {color) {object).tall, short blue torch, green torch, green armor,red torch,red skullkey,red pillar,
blue skullkey,blue keycard, yellow keycard, yellow skullkeygreen pillar,red keycard,red armor
Go to the {size) (object). Go to the color) (size)object.short torch,tall torch green tall, blue tall, blue short,tall pillar, short pillar red short, red tall
Go to the size){color)object.green short tall green, tall blue, short blue,short red, tall red
Go to the {color)(size)(object).short green green tall torch, green short torch,red short pillar,red short torch,
blue short torch,blue tall torchred tall pillar, green tall pillar, red tall torch, green short pillar
Go to the size) {color> {object).tall green torch,short green torch, short blue torch,tall blue torchshort red pillar, short red torch, tall red pillar, tall green pillar, tall red torch,short green pillar
Go to the superlative)object. Go to the {superlative> (color) object.largest, smallest smallest yellow, smallest blue, smallest green,largest blue,largest red,smallest red
EQAQuestion Typelargest green,largest yellow 21 Train Questions: Not containing‘blue'&‘torch'8 Test Questions: Containing‘blue'or‘torch'
Which {size)object is {color)in color?What color is the {object)? What color is the {size){object>? Which object is {color) in color?pillar, keycard, skullkey,armor short pillar, tall pillar red,yellow, greentorch short torch, tall torch blue
", + "type": "table", + "image_path": "7057d42e3744461df649f3b29df5a3d9702f5c717d595cdcc3987488ef779a3c.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 117, + 280, + 495, + 377.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 117, + 377.0, + 495, + 474.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 117, + 474.0, + 495, + 571.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 258, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 259, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 259, + 96 + ], + "score": 1.0, + "content": "D HOUSE3D EXPERIMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 505, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "In the House3D domain, we train on one house environment and randomize the colors of each object", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 504, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 504, + 129 + ], + "score": 1.0, + "content": "at the start of each episode. The agent’s spawn location is fixed. We create instructions and questions", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "score": 1.0, + "content": "dataset for this house similar to the Doom domain. The House3D objects used in our experiments", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 139, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 505, + 151 + ], + "score": 1.0, + "content": "are illustrated in Figure 15. Instructions and questions used for training and evaluation are listed in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 141, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 141, + 163 + ], + "score": 1.0, + "content": "Table 6.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 167, + 297, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 297, + 179 + ], + "score": 1.0, + "content": "Each model is trained for 50 million frames", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 178, + 297, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 297, + 190 + ], + "score": 1.0, + "content": "jointly on both SGN and EQA, without the aux-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 189, + 297, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 297, + 201 + ], + "score": 1.0, + "content": "iliary tasks and using identical reward func-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 252, + 212 + ], + "score": 1.0, + "content": "tions. Similar to Doom, we use a", + "type": "text" + }, + { + "bbox": [ + 252, + 200, + 265, + 210 + ], + "score": 0.82, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 199, + 297, + 212 + ], + "score": 1.0, + "content": "reward", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 210, + 297, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 297, + 222 + ], + "score": 1.0, + "content": "for reaching the correct object in SGN episodes", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 222, + 297, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 297, + 234 + ], + "score": 1.0, + "content": "and predicting the correct answer in the EQA", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 233, + 297, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 297, + 244 + ], + "score": 1.0, + "content": "episodes. We use a small negative reward of -", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 243, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 297, + 255 + ], + "score": 1.0, + "content": "0.001 per time step to encourage shorter paths", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 297, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 297, + 267 + ], + "score": 1.0, + "content": "to target and answering the questions as soon", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 266, + 297, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 297, + 278 + ], + "score": 1.0, + "content": "as possible. We also use distance based re-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 277, + 297, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 297, + 289 + ], + "score": 1.0, + "content": "ward shaping for both SGN and EQA episodes,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 287, + 297, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 297, + 299 + ], + "score": 1.0, + "content": "where the agent receives a small reward propor-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 299, + 297, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 297, + 310 + ], + "score": 1.0, + "content": "tional to decrease in distance to the target. SGN", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 309, + 297, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 297, + 322 + ], + "score": 1.0, + "content": "episodes end when the agent reaches any object", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 297, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 297, + 333 + ], + "score": 1.0, + "content": "and EQA episodes when agent predicts any an-", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 304, + 210, + 474, + 313 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 304, + 180, + 505, + 200 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 505, + 191 + ], + "score": 1.0, + "content": "Table 5: Accuracy of all the models on the SGN and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 190, + 486, + 201 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 486, + 201 + ], + "score": 1.0, + "content": "EQA train and test sets for the House3D Domain.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "table_body", + "bbox": [ + 304, + 210, + 474, + 313 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 474, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 474, + 313 + ], + "score": 0.977, + "html": "
SGNEQA
ModelTrainTestTrainTest
Text only0.630.330.220.23
Image only0.280.010.120.22
Concat0.650.130.310.13
GA0.980.200.920.03
FiLM0.990.370.920.24
PACMAN0.730.200.400.21
Dual-Attention0.990.470.890.29
", + "type": "table", + "image_path": "054eaa266cfb697cb6f36a23b30b1a988e4753b5ea56986060bcf5251f1d35fc.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 304, + 210, + 474, + 222.875 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 304, + 222.875, + 474, + 235.75 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 304, + 235.75, + 474, + 248.625 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 304, + 248.625, + 474, + 261.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 304, + 261.5, + 474, + 274.375 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 304, + 274.375, + 474, + 287.25 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 304, + 287.25, + 474, + 300.125 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 304, + 300.125, + 474, + 313.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 24.0 + }, + { + "type": "text", + "bbox": [ + 106, + 332, + 353, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 354, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 354, + 345 + ], + "score": 1.0, + "content": "swer. All episodes have a maximum length of 420 time steps.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 506, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "score": 1.0, + "content": "In Table 5, we report the train and test performance of all the models on both SGN and EQA. The re-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 359, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 371 + ], + "score": 1.0, + "content": "sults are similar as in Doom: the Dual-Attention model outperforms the baselines by a considerable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 370, + 140, + 384 + ], + "spans": [ + { + "bbox": [ + 104, + 370, + 140, + 384 + ], + "score": 1.0, + "content": "margin.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 113, + 392, + 499, + 479 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 392, + 499, + 479 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 392, + 499, + 479 + ], + "spans": [ + { + "bbox": [ + 113, + 392, + 499, + 479 + ], + "score": 0.972, + "type": "image", + "image_path": "af178f0259047b6ec08600ddcade5e3f117345df9c3c00343e94708605dc643a.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 113, + 392, + 499, + 421.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 113, + 421.0, + 499, + 450.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 113, + 450.0, + 499, + 479.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 112, + 487, + 498, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 486, + 498, + 499 + ], + "spans": [ + { + "bbox": [ + 112, + 486, + 498, + 499 + ], + "score": 1.0, + "content": "Figure 15: Example first-person views of the House3D environment with sample objects of various colors.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + } + ], + "index": 37.0 + }, + { + "type": "table", + "bbox": [ + 122, + 550, + 488, + 684 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 515, + 503, + 536 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 514, + 504, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 504, + 527 + ], + "score": 1.0, + "content": "Table 6: Instructions and questions for House3D experiments. We used 6 object classes (refrigerator, of-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 525, + 394, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 394, + 536 + ], + "score": 1.0, + "content": "fice chair, fish tank, fireplace, bed, sofa) and 4 colors (red, green, blue, yellow).", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "table_body", + "bbox": [ + 122, + 550, + 488, + 684 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 550, + 488, + 684 + ], + "spans": [ + { + "bbox": [ + 122, + 550, + 488, + 684 + ], + "score": 0.983, + "html": "
SGNInstruction Type22 Train Instructions: Not containing‘red'&‘bed'11 Test Instructions: Containing‘red'or‘bed'
Go to the (object). Go to the {color)(object).refrigerator,office_chair,fish_tank,fireplace green refrigerator, green office_chair, green fish_tank,green fireplace, green sofa, blue refrigerator, blue office_chair, blue fish_tank,blue fireplace,blue sofa, yellow refrigerator,yellow office_chair, yellow fish_tank,yellow fireplace,yellow sofabed red bed, green bed,blue bed, yellow bed,red refrigerator, red office_chair,red fish_tank, red fireplace,red sofa
EQAGo to the {color) object. Question Typegreen,blue,yellow 7 Train Questions:red 2 Test Questions:
Not containing‘blue'&‘sofa'Containing‘blue'or‘sofa'
What color is the {object)? What object is color> in color?refrigerator,office_chair,fish_tank,fireplace,bed red, green,yellowsofa blue
", + "type": "table", + "image_path": "5a2c6c4f7fcdbf5cd5e8da09a70edd31b66260c9c36a6aed86484d9c31826ff4.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 122, + 550, + 488, + 594.6666666666666 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 122, + 594.6666666666666, + 488, + 639.3333333333333 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 122, + 639.3333333333333, + 488, + 683.9999999999999 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "index": 40.75 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 258, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 259, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 259, + 96 + ], + "score": 1.0, + "content": "D HOUSE3D EXPERIMENTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 505, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "In the House3D domain, we train on one house environment and randomize the colors of each object", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 504, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 504, + 129 + ], + "score": 1.0, + "content": "at the start of each episode. The agent’s spawn location is fixed. We create instructions and questions", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "score": 1.0, + "content": "dataset for this house similar to the Doom domain. The House3D objects used in our experiments", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 139, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 505, + 151 + ], + "score": 1.0, + "content": "are illustrated in Figure 15. Instructions and questions used for training and evaluation are listed in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 141, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 141, + 163 + ], + "score": 1.0, + "content": "Table 6.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 106, + 506, + 163 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 167, + 297, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 297, + 179 + ], + "score": 1.0, + "content": "Each model is trained for 50 million frames", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 178, + 297, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 297, + 190 + ], + "score": 1.0, + "content": "jointly on both SGN and EQA, without the aux-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 189, + 297, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 297, + 201 + ], + "score": 1.0, + "content": "iliary tasks and using identical reward func-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 252, + 212 + ], + "score": 1.0, + "content": "tions. Similar to Doom, we use a", + "type": "text" + }, + { + "bbox": [ + 252, + 200, + 265, + 210 + ], + "score": 0.82, + "content": "+ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 199, + 297, + 212 + ], + "score": 1.0, + "content": "reward", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 210, + 297, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 297, + 222 + ], + "score": 1.0, + "content": "for reaching the correct object in SGN episodes", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 222, + 297, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 297, + 234 + ], + "score": 1.0, + "content": "and predicting the correct answer in the EQA", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 233, + 297, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 297, + 244 + ], + "score": 1.0, + "content": "episodes. We use a small negative reward of -", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 243, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 297, + 255 + ], + "score": 1.0, + "content": "0.001 per time step to encourage shorter paths", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 297, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 297, + 267 + ], + "score": 1.0, + "content": "to target and answering the questions as soon", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 266, + 297, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 297, + 278 + ], + "score": 1.0, + "content": "as possible. We also use distance based re-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 277, + 297, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 297, + 289 + ], + "score": 1.0, + "content": "ward shaping for both SGN and EQA episodes,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 287, + 297, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 297, + 299 + ], + "score": 1.0, + "content": "where the agent receives a small reward propor-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 299, + 297, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 297, + 310 + ], + "score": 1.0, + "content": "tional to decrease in distance to the target. SGN", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 309, + 297, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 297, + 322 + ], + "score": 1.0, + "content": "episodes end when the agent reaches any object", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 297, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 297, + 333 + ], + "score": 1.0, + "content": "and EQA episodes when agent predicts any an-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 354, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 354, + 345 + ], + "score": 1.0, + "content": "swer. All episodes have a maximum length of 420 time steps.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 167, + 297, + 333 + ] + }, + { + "type": "table", + "bbox": [ + 304, + 210, + 474, + 313 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 304, + 180, + 505, + 200 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 505, + 191 + ], + "score": 1.0, + "content": "Table 5: Accuracy of all the models on the SGN and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 190, + 486, + 201 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 486, + 201 + ], + "score": 1.0, + "content": "EQA train and test sets for the House3D Domain.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "table_body", + "bbox": [ + 304, + 210, + 474, + 313 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 474, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 474, + 313 + ], + "score": 0.977, + "html": "
SGNEQA
ModelTrainTestTrainTest
Text only0.630.330.220.23
Image only0.280.010.120.22
Concat0.650.130.310.13
GA0.980.200.920.03
FiLM0.990.370.920.24
PACMAN0.730.200.400.21
Dual-Attention0.990.470.890.29
", + "type": "table", + "image_path": "054eaa266cfb697cb6f36a23b30b1a988e4753b5ea56986060bcf5251f1d35fc.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 304, + 210, + 474, + 222.875 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 304, + 222.875, + 474, + 235.75 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 304, + 235.75, + 474, + 248.625 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 304, + 248.625, + 474, + 261.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 304, + 261.5, + 474, + 274.375 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 304, + 274.375, + 474, + 287.25 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 304, + 287.25, + 474, + 300.125 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 304, + 300.125, + 474, + 313.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 24.0 + }, + { + "type": "text", + "bbox": [ + 106, + 332, + 353, + 343 + ], + "lines": [], + "index": 31, + "bbox_fs": [ + 105, + 330, + 354, + 345 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 348, + 506, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "score": 1.0, + "content": "In Table 5, we report the train and test performance of all the models on both SGN and EQA. The re-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 359, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 371 + ], + "score": 1.0, + "content": "sults are similar as in Doom: the Dual-Attention model outperforms the baselines by a considerable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 370, + 140, + 384 + ], + "spans": [ + { + "bbox": [ + 104, + 370, + 140, + 384 + ], + "score": 1.0, + "content": "margin.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 347, + 505, + 384 + ] + }, + { + "type": "image", + "bbox": [ + 113, + 392, + 499, + 479 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 392, + 499, + 479 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 392, + 499, + 479 + ], + "spans": [ + { + "bbox": [ + 113, + 392, + 499, + 479 + ], + "score": 0.972, + "type": "image", + "image_path": "af178f0259047b6ec08600ddcade5e3f117345df9c3c00343e94708605dc643a.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 113, + 392, + 499, + 421.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 113, + 421.0, + 499, + 450.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 113, + 450.0, + 499, + 479.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 112, + 487, + 498, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 486, + 498, + 499 + ], + "spans": [ + { + "bbox": [ + 112, + 486, + 498, + 499 + ], + "score": 1.0, + "content": "Figure 15: Example first-person views of the House3D environment with sample objects of various colors.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + } + ], + "index": 37.0 + }, + { + "type": "table", + "bbox": [ + 122, + 550, + 488, + 684 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 515, + 503, + 536 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 514, + 504, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 504, + 527 + ], + "score": 1.0, + "content": "Table 6: Instructions and questions for House3D experiments. We used 6 object classes (refrigerator, of-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 525, + 394, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 394, + 536 + ], + "score": 1.0, + "content": "fice chair, fish tank, fireplace, bed, sofa) and 4 colors (red, green, blue, yellow).", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "table_body", + "bbox": [ + 122, + 550, + 488, + 684 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 550, + 488, + 684 + ], + "spans": [ + { + "bbox": [ + 122, + 550, + 488, + 684 + ], + "score": 0.983, + "html": "
SGNInstruction Type22 Train Instructions: Not containing‘red'&‘bed'11 Test Instructions: Containing‘red'or‘bed'
Go to the (object). Go to the {color)(object).refrigerator,office_chair,fish_tank,fireplace green refrigerator, green office_chair, green fish_tank,green fireplace, green sofa, blue refrigerator, blue office_chair, blue fish_tank,blue fireplace,blue sofa, yellow refrigerator,yellow office_chair, yellow fish_tank,yellow fireplace,yellow sofabed red bed, green bed,blue bed, yellow bed,red refrigerator, red office_chair,red fish_tank, red fireplace,red sofa
EQAGo to the {color) object. Question Typegreen,blue,yellow 7 Train Questions:red 2 Test Questions:
Not containing‘blue'&‘sofa'Containing‘blue'or‘sofa'
What color is the {object)? What object is color> in color?refrigerator,office_chair,fish_tank,fireplace,bed red, green,yellowsofa blue
", + "type": "table", + "image_path": "5a2c6c4f7fcdbf5cd5e8da09a70edd31b66260c9c36a6aed86484d9c31826ff4.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 122, + 550, + 488, + 594.6666666666666 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 122, + 594.6666666666666, + 488, + 639.3333333333333 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 122, + 639.3333333333333, + 488, + 683.9999999999999 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "index": 40.75 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/ByGq7hRqKX/ByGq7hRqKX_model.json b/parse/train/ByGq7hRqKX/ByGq7hRqKX_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6b507da9a5c52d83a9b76dd10157ca3b84f7c79b --- /dev/null +++ b/parse/train/ByGq7hRqKX/ByGq7hRqKX_model.json @@ -0,0 +1,20309 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1070, + 1404, + 1070, + 1404, + 1466, + 298, + 1466 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 398, + 541, + 1302, + 541, + 1302, + 966, + 398, + 966 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1712, + 1403, + 1712, + 1403, + 1957, + 298, + 1957 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1482, + 1402, + 1482, + 1402, + 1697, + 298, + 1697 + ], + "score": 0.98 + }, + { + "category_id": 0, + "poly": [ + 298, + 219, + 1178, + 219, + 1178, + 324, + 298, + 324 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1398, + 1972, + 1398, + 2034, + 299, + 2034 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 313, + 378, + 680, + 378, + 680, + 438, + 313, + 438 + ], + "score": 0.932 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 301, + 1016, + 573, + 1016, + 573, + 1051, + 301, + 1051 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 773, + 477, + 927, + 477, + 927, + 512, + 773, + 512 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.661 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 217.0, + 1181.0, + 217.0, + 1181.0, + 272.0, + 296.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 272.0, + 1043.0, + 272.0, + 1043.0, + 328.0, + 297.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1013.0, + 578.0, + 1013.0, + 578.0, + 1060.0, + 294.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 474.0, + 934.0, + 474.0, + 934.0, + 517.0, + 768.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1069.0, + 1404.0, + 1069.0, + 1404.0, + 1105.0, + 294.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1096.0, + 1406.0, + 1096.0, + 1406.0, + 1140.0, + 292.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1131.0, + 1406.0, + 1131.0, + 1406.0, + 1167.0, + 294.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1161.0, + 1404.0, + 1161.0, + 1404.0, + 1197.0, + 294.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1192.0, + 1405.0, + 1192.0, + 1405.0, + 1228.0, + 293.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1223.0, + 1404.0, + 1223.0, + 1404.0, + 1259.0, + 294.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1252.0, + 1402.0, + 1252.0, + 1402.0, + 1289.0, + 293.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1281.0, + 1406.0, + 1281.0, + 1406.0, + 1320.0, + 291.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1312.0, + 1406.0, + 1312.0, + 1406.0, + 1352.0, + 293.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1343.0, + 1405.0, + 1343.0, + 1405.0, + 1382.0, + 292.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1410.0, + 294.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1401.0, + 1405.0, + 1401.0, + 1405.0, + 1446.0, + 292.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1436.0, + 555.0, + 1436.0, + 555.0, + 1472.0, + 294.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 537.0, + 1307.0, + 537.0, + 1307.0, + 579.0, + 392.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 572.0, + 1305.0, + 572.0, + 1305.0, + 604.0, + 394.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 601.0, + 1305.0, + 601.0, + 1305.0, + 637.0, + 394.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 631.0, + 1304.0, + 631.0, + 1304.0, + 668.0, + 393.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 662.0, + 1306.0, + 662.0, + 1306.0, + 696.0, + 393.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 693.0, + 1307.0, + 693.0, + 1307.0, + 728.0, + 394.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 724.0, + 1305.0, + 724.0, + 1305.0, + 756.0, + 393.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 757.0, + 1306.0, + 757.0, + 1306.0, + 786.0, + 393.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 785.0, + 1306.0, + 785.0, + 1306.0, + 818.0, + 393.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 817.0, + 1307.0, + 817.0, + 1307.0, + 849.0, + 396.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 847.0, + 1306.0, + 847.0, + 1306.0, + 878.0, + 393.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 875.0, + 1305.0, + 875.0, + 1305.0, + 910.0, + 394.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 903.0, + 1306.0, + 903.0, + 1306.0, + 942.0, + 392.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 937.0, + 511.0, + 937.0, + 511.0, + 969.0, + 394.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1710.0, + 1404.0, + 1710.0, + 1404.0, + 1749.0, + 292.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 1404.0, + 1743.0, + 1404.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 1405.0, + 1772.0, + 1405.0, + 1809.0, + 293.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1805.0, + 1404.0, + 1805.0, + 1404.0, + 1838.0, + 293.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1836.0, + 1404.0, + 1836.0, + 1404.0, + 1869.0, + 294.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1864.0, + 1404.0, + 1864.0, + 1404.0, + 1898.0, + 294.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1894.0, + 1402.0, + 1894.0, + 1402.0, + 1928.0, + 297.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1922.0, + 1221.0, + 1922.0, + 1221.0, + 1964.0, + 293.0, + 1964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1482.0, + 1402.0, + 1482.0, + 1402.0, + 1517.0, + 294.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1513.0, + 1404.0, + 1513.0, + 1404.0, + 1548.0, + 294.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1543.0, + 1406.0, + 1543.0, + 1406.0, + 1580.0, + 292.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1573.0, + 1406.0, + 1573.0, + 1406.0, + 1609.0, + 293.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1406.0, + 1605.0, + 1406.0, + 1640.0, + 294.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1633.0, + 1406.0, + 1633.0, + 1406.0, + 1670.0, + 293.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1666.0, + 1171.0, + 1666.0, + 1171.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1403.0, + 1970.0, + 1403.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2000.0, + 1403.0, + 2000.0, + 1403.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 379.0, + 560.0, + 379.0, + 560.0, + 411.0, + 315.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 407.0, + 681.0, + 407.0, + 681.0, + 442.0, + 312.0, + 442.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 586, + 1404, + 586, + 1404, + 953, + 298, + 953 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 1077, + 1404, + 1077, + 1404, + 1384, + 298, + 1384 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1399, + 1403, + 1399, + 1403, + 1613, + 299, + 1613 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1627, + 1402, + 1627, + 1402, + 1873, + 298, + 1873 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 300, + 1888, + 1402, + 1888, + 1402, + 2042, + 300, + 2042 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 301, + 1005, + 587, + 1005, + 587, + 1040, + 301, + 1040 + ], + "score": 0.905 + }, + { + "category_id": 4, + "poly": [ + 296, + 431, + 1401, + 431, + 1401, + 516, + 296, + 516 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.694 + }, + { + "category_id": 3, + "poly": [ + 298, + 224, + 1356, + 224, + 1356, + 415, + 298, + 415 + ], + "score": 0.645 + }, + { + "category_id": 5, + "poly": [ + 298, + 224, + 1356, + 224, + 1356, + 415, + 298, + 415 + ], + "score": 0.418, + "html": "
TaskTrain SetTest Set
SGNInstructions not containing‘red'&‘pillar': ‘Go to the blue object' ‘Go to the torch'Instructions containing ‘red'or‘pillar': ‘Go to the red pillar' ‘Go to the tall red object'
EQA SGN:Go to the tall green objectQuestionsnot containing‘blue'&‘torch': ‘Which object is red in color?' ‘What color is the tall pillar?Questionscontaining‘blue' or‘torch': ‘Which object is blue in color? ‘What color is the torch?'
" + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.259 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1002.0, + 593.0, + 1002.0, + 593.0, + 1046.0, + 292.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 431.0, + 1405.0, + 431.0, + 1405.0, + 463.0, + 295.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 460.0, + 1405.0, + 460.0, + 1405.0, + 489.0, + 296.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 485.0, + 1368.0, + 485.0, + 1368.0, + 519.0, + 292.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 232.0, + 633.0, + 232.0, + 633.0, + 260.0, + 583.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 233.0, + 738.0, + 233.0, + 738.0, + 259.0, + 655.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 232.0, + 1096.0, + 232.0, + 1096.0, + 262.0, + 1022.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 265.0, + 634.0, + 265.0, + 634.0, + 293.0, + 582.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 265.0, + 1001.0, + 265.0, + 1001.0, + 293.0, + 653.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 265.0, + 1340.0, + 265.0, + 1340.0, + 293.0, + 1022.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 298.0, + 351.0, + 298.0, + 351.0, + 335.0, + 319.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 309.0, + 395.0, + 309.0, + 395.0, + 326.0, + 382.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 301.0, + 468.0, + 301.0, + 468.0, + 348.0, + 431.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 320.0, + 533.0, + 320.0, + 533.0, + 338.0, + 510.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 288.0, + 861.0, + 288.0, + 861.0, + 337.0, + 675.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 288.0, + 1247.0, + 288.0, + 1247.0, + 335.0, + 1044.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 334.0, + 634.0, + 334.0, + 634.0, + 362.0, + 581.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 335.0, + 993.0, + 335.0, + 993.0, + 360.0, + 656.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 334.0, + 1332.0, + 334.0, + 1332.0, + 359.0, + 1022.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 355.0, + 922.0, + 355.0, + 922.0, + 383.0, + 675.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 355.0, + 1300.0, + 355.0, + 1300.0, + 383.0, + 1043.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 375.0, + 530.0, + 375.0, + 530.0, + 397.0, + 334.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 378.0, + 912.0, + 378.0, + 912.0, + 404.0, + 675.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 378.0, + 1254.0, + 378.0, + 1254.0, + 403.0, + 1044.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 392.0, + 530.0, + 392.0, + 530.0, + 414.0, + 334.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 588.0, + 1405.0, + 588.0, + 1405.0, + 622.0, + 294.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 620.0, + 1404.0, + 620.0, + 1404.0, + 650.0, + 296.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 649.0, + 1402.0, + 649.0, + 1402.0, + 683.0, + 294.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 677.0, + 1408.0, + 677.0, + 1408.0, + 715.0, + 293.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 710.0, + 1405.0, + 710.0, + 1405.0, + 744.0, + 294.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 740.0, + 1404.0, + 740.0, + 1404.0, + 774.0, + 296.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 771.0, + 1405.0, + 771.0, + 1405.0, + 804.0, + 293.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 800.0, + 1405.0, + 800.0, + 1405.0, + 835.0, + 293.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 828.0, + 1406.0, + 828.0, + 1406.0, + 868.0, + 292.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 861.0, + 1406.0, + 861.0, + 1406.0, + 896.0, + 295.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 891.0, + 1404.0, + 891.0, + 1404.0, + 925.0, + 293.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 923.0, + 725.0, + 923.0, + 725.0, + 956.0, + 295.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1078.0, + 1406.0, + 1078.0, + 1406.0, + 1113.0, + 296.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1102.0, + 1408.0, + 1102.0, + 1408.0, + 1148.0, + 292.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1139.0, + 1405.0, + 1139.0, + 1405.0, + 1175.0, + 294.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1171.0, + 1405.0, + 1171.0, + 1405.0, + 1203.0, + 296.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1196.0, + 1408.0, + 1196.0, + 1408.0, + 1237.0, + 292.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1228.0, + 1404.0, + 1228.0, + 1404.0, + 1269.0, + 291.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1258.0, + 1405.0, + 1258.0, + 1405.0, + 1298.0, + 292.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1292.0, + 1406.0, + 1292.0, + 1406.0, + 1327.0, + 294.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1321.0, + 1408.0, + 1321.0, + 1408.0, + 1356.0, + 294.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1351.0, + 1317.0, + 1351.0, + 1317.0, + 1390.0, + 295.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1398.0, + 1405.0, + 1398.0, + 1405.0, + 1436.0, + 295.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1431.0, + 1403.0, + 1431.0, + 1403.0, + 1466.0, + 295.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1458.0, + 1405.0, + 1458.0, + 1405.0, + 1499.0, + 292.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1491.0, + 1405.0, + 1491.0, + 1405.0, + 1526.0, + 294.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1521.0, + 1405.0, + 1521.0, + 1405.0, + 1555.0, + 294.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1549.0, + 1405.0, + 1549.0, + 1405.0, + 1589.0, + 292.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1582.0, + 1296.0, + 1582.0, + 1296.0, + 1616.0, + 295.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1627.0, + 1404.0, + 1627.0, + 1404.0, + 1664.0, + 294.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1660.0, + 1404.0, + 1660.0, + 1404.0, + 1692.0, + 293.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1687.0, + 1407.0, + 1687.0, + 1407.0, + 1727.0, + 293.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1721.0, + 1406.0, + 1721.0, + 1406.0, + 1755.0, + 296.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1748.0, + 1406.0, + 1748.0, + 1406.0, + 1788.0, + 292.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1781.0, + 1406.0, + 1781.0, + 1406.0, + 1818.0, + 293.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1812.0, + 1404.0, + 1812.0, + 1404.0, + 1846.0, + 294.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1839.0, + 1051.0, + 1839.0, + 1051.0, + 1877.0, + 293.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1887.0, + 1404.0, + 1887.0, + 1404.0, + 1923.0, + 295.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1919.0, + 1403.0, + 1919.0, + 1403.0, + 1952.0, + 296.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1948.0, + 1404.0, + 1948.0, + 1404.0, + 1985.0, + 294.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1977.0, + 1406.0, + 1977.0, + 1406.0, + 2017.0, + 295.0, + 2017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2011.0, + 1014.0, + 2011.0, + 1014.0, + 2044.0, + 296.0, + 2044.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1249, + 1404, + 1249, + 1404, + 1526, + 297, + 1526 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1758, + 1404, + 1758, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 989, + 1404, + 989, + 1404, + 1235, + 298, + 1235 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 592, + 1404, + 592, + 1404, + 806, + 297, + 806 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 821, + 1403, + 821, + 1403, + 975, + 298, + 975 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1541, + 1404, + 1541, + 1404, + 1664, + 298, + 1664 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 347, + 231, + 1338, + 231, + 1338, + 432, + 347, + 432 + ], + "score": 0.96 + }, + { + "category_id": 4, + "poly": [ + 439, + 443, + 1259, + 443, + 1259, + 473, + 439, + 473 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 301, + 1699, + 642, + 1699, + 642, + 1736, + 301, + 1736 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 301, + 533, + 712, + 533, + 712, + 568, + 301, + 568 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.615 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.425 + }, + { + "category_id": 13, + "poly": [ + 392, + 774, + 488, + 774, + 488, + 808, + 392, + 808 + ], + "score": 0.93, + "latex": "\\pi ( a _ { t } | s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 358, + 714, + 503, + 714, + 503, + 748, + 358, + 748 + ], + "score": 0.93, + "latex": "s _ { t } = ( I _ { t } , T )" + }, + { + "category_id": 13, + "poly": [ + 433, + 2001, + 582, + 2001, + 582, + 2035, + 433, + 2035 + ], + "score": 0.92, + "latex": "x _ { \\mathrm { B o W } } ~ \\in ~ \\mathbb { R } ^ { V }" + }, + { + "category_id": 13, + "poly": [ + 793, + 1879, + 1158, + 1879, + 1158, + 1915, + 793, + 1915 + ], + "score": 0.92, + "latex": "x _ { I } = f ( I _ { t } ; \\theta _ { \\mathrm { c o n v } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }" + }, + { + "category_id": 13, + "poly": [ + 980, + 2000, + 1284, + 2000, + 1284, + 2036, + 980, + 2036 + ], + "score": 0.92, + "latex": "x _ { \\mathrm { s e n t } } = f ( \\breve { T } ; \\theta _ { \\mathrm { s e n t } } ) \\in \\mathbb { R } ^ { \\breve { V } }" + }, + { + "category_id": 13, + "poly": [ + 1251, + 1882, + 1307, + 1882, + 1307, + 1912, + 1251, + 1912 + ], + "score": 0.9, + "latex": "\\theta _ { \\mathrm { c o n v } }" + }, + { + "category_id": 13, + "poly": [ + 1189, + 1822, + 1215, + 1822, + 1215, + 1851, + 1189, + 1851 + ], + "score": 0.87, + "latex": "I _ { t }" + }, + { + "category_id": 13, + "poly": [ + 585, + 716, + 611, + 716, + 611, + 745, + 585, + 745 + ], + "score": 0.86, + "latex": "I _ { t }" + }, + { + "category_id": 13, + "poly": [ + 933, + 1944, + 962, + 1944, + 962, + 1970, + 933, + 1970 + ], + "score": 0.84, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 372, + 751, + 398, + 751, + 398, + 775, + 372, + 775 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 775, + 1791, + 799, + 1791, + 799, + 1818, + 775, + 1818 + ], + "score": 0.77, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 550, + 655, + 575, + 655, + 575, + 682, + 550, + 682 + ], + "score": 0.76, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1360, + 776, + 1382, + 776, + 1382, + 802, + 1360, + 802 + ], + "score": 0.76, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1000, + 624, + 1024, + 624, + 1024, + 650, + 1000, + 650 + ], + "score": 0.76, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 804, + 1914, + 830, + 1914, + 830, + 1939, + 804, + 1939 + ], + "score": 0.75, + "latex": "V" + }, + { + "category_id": 13, + "poly": [ + 1011, + 1944, + 1044, + 1944, + 1044, + 1970, + 1011, + 1970 + ], + "score": 0.73, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 972, + 1974, + 995, + 1974, + 995, + 2000, + 972, + 2000 + ], + "score": 0.7, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1141, + 687, + 1155, + 687, + 1155, + 712, + 1141, + 712 + ], + "score": 0.59, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 432, + 946, + 482, + 946, + 482, + 973, + 432, + 973 + ], + "score": 0.4, + "latex": "^ { 6 6 } \\mathrm { G o }" + }, + { + "category_id": 13, + "poly": [ + 656, + 854, + 728, + 854, + 728, + 884, + 656, + 884 + ], + "score": 0.28, + "latex": "( E Q A )" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 228.0, + 430.0, + 228.0, + 430.0, + 252.0, + 379.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 237.0, + 711.0, + 237.0, + 711.0, + 263.0, + 569.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 255.0, + 452.0, + 255.0, + 452.0, + 277.0, + 361.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 252.0, + 945.0, + 252.0, + 945.0, + 273.0, + 927.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 248.0, + 1056.0, + 248.0, + 1056.0, + 278.0, + 959.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 274.0, + 554.0, + 274.0, + 554.0, + 337.0, + 366.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 310.0, + 850.0, + 310.0, + 850.0, + 336.0, + 725.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 282.0, + 1031.0, + 282.0, + 1031.0, + 327.0, + 959.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 297.0, + 1168.0, + 297.0, + 1168.0, + 330.0, + 1110.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 264.0, + 1336.0, + 264.0, + 1336.0, + 321.0, + 1225.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 334.0, + 687.0, + 334.0, + 687.0, + 360.0, + 569.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 328.0, + 1055.0, + 328.0, + 1055.0, + 367.0, + 960.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 323.0, + 1174.0, + 323.0, + 1174.0, + 350.0, + 1103.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 333.0, + 1319.0, + 333.0, + 1319.0, + 362.0, + 1245.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 381.0, + 373.0, + 381.0, + 373.0, + 395.0, + 357.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 387.0, + 452.0, + 387.0, + 452.0, + 400.0, + 435.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 369.0, + 560.0, + 369.0, + 560.0, + 397.0, + 506.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 379.0, + 607.0, + 379.0, + 607.0, + 387.0, + 598.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 364.0, + 651.0, + 364.0, + 651.0, + 401.0, + 613.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 376.0, + 943.0, + 376.0, + 943.0, + 397.0, + 924.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 367.0, + 1045.0, + 367.0, + 1045.0, + 411.0, + 957.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 361.0, + 1334.0, + 361.0, + 1334.0, + 386.0, + 1221.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 395.0, + 667.0, + 395.0, + 667.0, + 417.0, + 592.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 411.0, + 435.0, + 411.0, + 435.0, + 435.0, + 372.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 410.0, + 669.0, + 410.0, + 669.0, + 434.0, + 592.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.25, + 262.0, + 656.25, + 262.0, + 656.25, + 290.5, + 605.25, + 290.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.75, + 288.0, + 946.75, + 288.0, + 946.75, + 316.0, + 918.75, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 312.0, + 656.0, + 312.0, + 656.0, + 341.0, + 605.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.25, + 333.5, + 957.25, + 333.5, + 957.25, + 362.5, + 907.25, + 362.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 441.0, + 1262.0, + 441.0, + 1262.0, + 478.0, + 436.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1699.0, + 644.0, + 1699.0, + 644.0, + 1740.0, + 294.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 529.0, + 717.0, + 529.0, + 717.0, + 575.0, + 292.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1251.0, + 1404.0, + 1251.0, + 1404.0, + 1284.0, + 296.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1283.0, + 1405.0, + 1283.0, + 1405.0, + 1316.0, + 296.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1313.0, + 1405.0, + 1313.0, + 1405.0, + 1346.0, + 296.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1339.0, + 1405.0, + 1339.0, + 1405.0, + 1378.0, + 292.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1372.0, + 1404.0, + 1372.0, + 1404.0, + 1405.0, + 295.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1403.0, + 1405.0, + 1403.0, + 1405.0, + 1439.0, + 292.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1435.0, + 1406.0, + 1435.0, + 1406.0, + 1469.0, + 293.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1466.0, + 1405.0, + 1466.0, + 1405.0, + 1499.0, + 295.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1493.0, + 1067.0, + 1493.0, + 1067.0, + 1530.0, + 295.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1758.0, + 1405.0, + 1758.0, + 1405.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1789.0, + 774.0, + 1789.0, + 774.0, + 1823.0, + 293.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1789.0, + 1406.0, + 1789.0, + 1406.0, + 1823.0, + 800.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1818.0, + 1188.0, + 1818.0, + 1188.0, + 1857.0, + 291.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1818.0, + 1406.0, + 1818.0, + 1406.0, + 1857.0, + 1216.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1886.0, + 295.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1872.0, + 792.0, + 1872.0, + 792.0, + 1923.0, + 288.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1872.0, + 1250.0, + 1872.0, + 1250.0, + 1923.0, + 1159.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1872.0, + 1411.0, + 1872.0, + 1411.0, + 1923.0, + 1308.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 803.0, + 1910.0, + 803.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1946.0, + 831.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 932.0, + 1941.0, + 932.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1941.0, + 1010.0, + 1941.0, + 1010.0, + 1978.0, + 963.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 1045.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 971.0, + 1971.0, + 971.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 996.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1997.0, + 432.0, + 1997.0, + 432.0, + 2040.0, + 291.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1997.0, + 979.0, + 1997.0, + 979.0, + 2040.0, + 583.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1997.0, + 1409.0, + 1997.0, + 1409.0, + 2040.0, + 1285.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 989.0, + 1404.0, + 989.0, + 1404.0, + 1027.0, + 294.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1021.0, + 1407.0, + 1021.0, + 1407.0, + 1056.0, + 292.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1052.0, + 1402.0, + 1052.0, + 1402.0, + 1086.0, + 296.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1083.0, + 1404.0, + 1083.0, + 1404.0, + 1117.0, + 296.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1110.0, + 1406.0, + 1110.0, + 1406.0, + 1148.0, + 292.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1142.0, + 1405.0, + 1142.0, + 1405.0, + 1179.0, + 293.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1174.0, + 1405.0, + 1174.0, + 1405.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1205.0, + 757.0, + 1205.0, + 757.0, + 1238.0, + 294.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 591.0, + 1406.0, + 591.0, + 1406.0, + 629.0, + 295.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 623.0, + 999.0, + 623.0, + 999.0, + 657.0, + 295.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 623.0, + 1405.0, + 623.0, + 1405.0, + 657.0, + 1025.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 649.0, + 549.0, + 649.0, + 549.0, + 693.0, + 292.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 693.0, + 576.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 682.0, + 1140.0, + 682.0, + 1140.0, + 720.0, + 294.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 682.0, + 1408.0, + 682.0, + 1408.0, + 720.0, + 1156.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 715.0, + 357.0, + 715.0, + 357.0, + 749.0, + 295.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 715.0, + 584.0, + 715.0, + 584.0, + 749.0, + 504.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 715.0, + 1405.0, + 715.0, + 1405.0, + 749.0, + 612.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 744.0, + 371.0, + 744.0, + 371.0, + 779.0, + 295.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 744.0, + 1404.0, + 744.0, + 1404.0, + 779.0, + 399.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 775.0, + 391.0, + 775.0, + 391.0, + 809.0, + 295.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 775.0, + 1359.0, + 775.0, + 1359.0, + 809.0, + 489.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1383.0, + 775.0, + 1394.0, + 775.0, + 1394.0, + 809.0, + 1383.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 820.0, + 1405.0, + 820.0, + 1405.0, + 858.0, + 294.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 852.0, + 655.0, + 852.0, + 655.0, + 889.0, + 294.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 852.0, + 1405.0, + 852.0, + 1405.0, + 889.0, + 729.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 882.0, + 1405.0, + 882.0, + 1405.0, + 919.0, + 293.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 911.0, + 1405.0, + 911.0, + 1405.0, + 952.0, + 293.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 943.0, + 431.0, + 943.0, + 431.0, + 980.0, + 294.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 943.0, + 1015.0, + 943.0, + 1015.0, + 980.0, + 483.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1539.0, + 1405.0, + 1539.0, + 1405.0, + 1577.0, + 293.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1571.0, + 1406.0, + 1571.0, + 1406.0, + 1606.0, + 295.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1601.0, + 1404.0, + 1601.0, + 1404.0, + 1636.0, + 293.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1633.0, + 1382.0, + 1633.0, + 1382.0, + 1668.0, + 292.0, + 1668.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1850, + 1404, + 1850, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 655, + 935, + 655, + 935, + 961, + 298, + 961 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1404, + 228, + 1404, + 383, + 297, + 383 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1128, + 935, + 1128, + 935, + 1433, + 298, + 1433 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 1432, + 1405, + 1432, + 1405, + 1679, + 297, + 1679 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 1693, + 1404, + 1693, + 1404, + 1787, + 298, + 1787 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 297, + 961, + 1404, + 961, + 1404, + 1115, + 297, + 1115 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 949, + 1165, + 1377, + 1165, + 1377, + 1346, + 949, + 1346 + ], + "score": 0.959 + }, + { + "category_id": 3, + "poly": [ + 951, + 692, + 1376, + 692, + 1376, + 876, + 951, + 876 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 299, + 579, + 1400, + 579, + 1400, + 641, + 299, + 641 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 294, + 422, + 1400, + 422, + 1400, + 485, + 294, + 485 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 672, + 383, + 1027, + 383, + 1027, + 417, + 672, + 417 + ], + "score": 0.937 + }, + { + "category_id": 8, + "poly": [ + 641, + 1799, + 1055, + 1799, + 1055, + 1840, + 641, + 1840 + ], + "score": 0.932 + }, + { + "category_id": 4, + "poly": [ + 994, + 1373, + 1362, + 1373, + 1362, + 1404, + 994, + 1404 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 520, + 657, + 520, + 657, + 553, + 299, + 553 + ], + "score": 0.905 + }, + { + "category_id": 4, + "poly": [ + 997, + 903, + 1358, + 903, + 1358, + 932, + 997, + 932 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1806, + 1399, + 1806, + 1399, + 1835, + 1366, + 1835 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.772 + }, + { + "category_id": 13, + "poly": [ + 877, + 1019, + 1081, + 1019, + 1081, + 1055, + 877, + 1055 + ], + "score": 0.93, + "latex": "M _ { y _ { T } } \\in \\mathbb { R } ^ { d \\times H \\times W }" + }, + { + "category_id": 13, + "poly": [ + 1053, + 1524, + 1160, + 1524, + 1160, + 1559, + 1053, + 1559 + ], + "score": 0.93, + "latex": "( i , j , k ) ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 1173, + 1020, + 1279, + 1020, + 1279, + 1055, + 1173, + 1055 + ], + "score": 0.93, + "latex": "( i , j , k ) ^ { t \\bar { h } }" + }, + { + "category_id": 14, + "poly": [ + 671, + 380, + 1026, + 380, + 1026, + 417, + 671, + 417 + ], + "score": 0.93, + "latex": "x _ { \\mathrm { S } } , x _ { \\mathrm { A n s } } = f _ { \\mathrm { D A } } ( x _ { I } , x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } )" + }, + { + "category_id": 14, + "poly": [ + 639, + 1797, + 1061, + 1797, + 1061, + 1839, + 639, + 1839 + ], + "score": 0.92, + "latex": "x _ { \\mathrm { G A l } } = f _ { \\mathrm { G A } } ( x _ { I } , x _ { \\mathrm { B o W } } ) \\in \\mathbb { R } ^ { V \\times H \\times W } ." + }, + { + "category_id": 13, + "poly": [ + 747, + 1522, + 957, + 1522, + 957, + 1558, + 747, + 1558 + ], + "score": 0.92, + "latex": "M _ { y _ { S } } \\in \\mathbb { R } ^ { d \\times \\hat { H } \\times W }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1555, + 551, + 1555, + 551, + 1590, + 298, + 1590 + ], + "score": 0.92, + "latex": "\\bar { M } _ { y s } [ i , j , k ] = y _ { S } [ j , k ]" + }, + { + "category_id": 13, + "poly": [ + 621, + 897, + 807, + 897, + 807, + 931, + 621, + 931 + ], + "score": 0.92, + "latex": "y _ { I } \\in \\mathring { \\mathbb { R } } ^ { d \\times H \\times W }" + }, + { + "category_id": 13, + "poly": [ + 772, + 1431, + 945, + 1431, + 945, + 1462, + 772, + 1462 + ], + "score": 0.92, + "latex": "\\mathit { \\check { z } } \\in \\mathbb { R } ^ { d \\times H \\times W }" + }, + { + "category_id": 13, + "poly": [ + 340, + 959, + 505, + 959, + 505, + 989, + 340, + 989 + ], + "score": 0.92, + "latex": "{ \\boldsymbol { z } } \\in \\mathbb { R } ^ { d \\times H \\times W }" + }, + { + "category_id": 13, + "poly": [ + 643, + 1586, + 977, + 1586, + 977, + 1620, + 643, + 1620 + ], + "score": 0.91, + "latex": "z = f _ { \\mathrm { S A } } ( y _ { I } , y _ { S } ) = M _ { y _ { S } } \\odot y _ { I }" + }, + { + "category_id": 13, + "poly": [ + 460, + 929, + 560, + 929, + 560, + 960, + 460, + 960 + ], + "score": 0.91, + "latex": "\\bar { y _ { T } } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 397, + 1051, + 624, + 1051, + 624, + 1084, + 397, + 1084 + ], + "score": 0.91, + "latex": "M _ { y _ { T } } [ i , j , k ] = \\bar { y } _ { T } [ i ]" + }, + { + "category_id": 13, + "poly": [ + 297, + 1021, + 391, + 1021, + 391, + 1051, + 297, + 1051 + ], + "score": 0.91, + "latex": "H \\times W" + }, + { + "category_id": 13, + "poly": [ + 788, + 262, + 838, + 262, + 838, + 293, + 788, + 293 + ], + "score": 0.9, + "latex": "\\theta _ { \\mathrm { s e n t } }" + }, + { + "category_id": 13, + "poly": [ + 478, + 427, + 529, + 427, + 529, + 453, + 478, + 453 + ], + "score": 0.9, + "latex": "x _ { \\mathrm { A n s } }" + }, + { + "category_id": 13, + "poly": [ + 1214, + 580, + 1261, + 580, + 1261, + 611, + 1214, + 611 + ], + "score": 0.89, + "latex": "f _ { \\mathrm { G A } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1085, + 637, + 1085, + 637, + 1117, + 297, + 1117 + ], + "score": 0.89, + "latex": "z = f _ { \\mathrm { G A } } ( y _ { I } , y _ { T } ) = M _ { y _ { T } } \\odot y _ { I }" + }, + { + "category_id": 13, + "poly": [ + 1115, + 292, + 1161, + 292, + 1161, + 324, + 1115, + 324 + ], + "score": 0.88, + "latex": "f _ { \\mathrm { D A } }" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1728, + 1179, + 1728, + 1179, + 1756, + 1127, + 1756 + ], + "score": 0.88, + "latex": "x _ { \\mathrm { s e n t } }" + }, + { + "category_id": 13, + "poly": [ + 352, + 1433, + 515, + 1433, + 515, + 1465, + 352, + 1465 + ], + "score": 0.87, + "latex": "\\mathit { y } _ { S } \\in \\mathbb { R } ^ { H \\times W }" + }, + { + "category_id": 13, + "poly": [ + 416, + 356, + 467, + 356, + 467, + 384, + 416, + 384 + ], + "score": 0.87, + "latex": "x _ { \\mathrm { A n s } }" + }, + { + "category_id": 13, + "poly": [ + 409, + 610, + 451, + 610, + 451, + 642, + 409, + 642 + ], + "score": 0.87, + "latex": "f _ { \\mathrm { S A } }" + }, + { + "category_id": 13, + "poly": [ + 1234, + 327, + 1267, + 327, + 1267, + 353, + 1234, + 353 + ], + "score": 0.85, + "latex": "x _ { S }" + }, + { + "category_id": 13, + "poly": [ + 1176, + 995, + 1210, + 995, + 1210, + 1021, + 1176, + 1021 + ], + "score": 0.85, + "latex": "y _ { T }" + }, + { + "category_id": 13, + "poly": [ + 832, + 966, + 867, + 966, + 867, + 993, + 832, + 993 + ], + "score": 0.85, + "latex": "y _ { T }" + }, + { + "category_id": 13, + "poly": [ + 772, + 1759, + 803, + 1759, + 803, + 1785, + 772, + 1785 + ], + "score": 0.85, + "latex": "x _ { I }" + }, + { + "category_id": 13, + "poly": [ + 348, + 1401, + 520, + 1401, + 520, + 1432, + 348, + 1432 + ], + "score": 0.84, + "latex": "y _ { I } \\in \\dot { \\mathbb { R } ^ { d \\times H \\times W } }" + }, + { + "category_id": 13, + "poly": [ + 389, + 428, + 425, + 428, + 425, + 453, + 389, + 453 + ], + "score": 0.84, + "latex": "x _ { S }" + }, + { + "category_id": 13, + "poly": [ + 722, + 1086, + 748, + 1086, + 748, + 1111, + 722, + 1111 + ], + "score": 0.84, + "latex": "\\odot" + }, + { + "category_id": 13, + "poly": [ + 874, + 1760, + 934, + 1760, + 934, + 1786, + 874, + 1786 + ], + "score": 0.83, + "latex": "x _ { \\mathrm { B o W } }" + }, + { + "category_id": 13, + "poly": [ + 627, + 995, + 656, + 995, + 656, + 1022, + 627, + 1022 + ], + "score": 0.82, + "latex": "y _ { I }" + }, + { + "category_id": 13, + "poly": [ + 701, + 232, + 726, + 232, + 726, + 258, + 701, + 258 + ], + "score": 0.79, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1102, + 1496, + 1121, + 1496, + 1121, + 1522, + 1102, + 1522 + ], + "score": 0.73, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1009, + 1730, + 1070, + 1730, + 1070, + 1756, + 1009, + 1756 + ], + "score": 0.6, + "latex": "x _ { \\mathrm { B o W } }" + }, + { + "category_id": 13, + "poly": [ + 967, + 1730, + 999, + 1730, + 999, + 1755, + 967, + 1755 + ], + "score": 0.6, + "latex": "x _ { I }" + }, + { + "category_id": 13, + "poly": [ + 1321, + 1374, + 1359, + 1374, + 1359, + 1403, + 1321, + 1403 + ], + "score": 0.29, + "latex": "f _ { \\mathrm { S A } }" + }, + { + "category_id": 13, + "poly": [ + 1315, + 903, + 1355, + 903, + 1355, + 932, + 1315, + 932 + ], + "score": 0.29, + "latex": "f _ { \\mathrm { G A } }" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1176.0, + 1204.0, + 1176.0, + 1204.0, + 1193.0, + 1186.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 1253.0, + 1349.0, + 1253.0, + 1349.0, + 1260.0, + 1338.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1290.0, + 1062.0, + 1290.0, + 1062.0, + 1300.0, + 1055.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1292.0, + 1115.0, + 1292.0, + 1115.0, + 1308.0, + 1078.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1318.0, + 1034.0, + 1318.0, + 1034.0, + 1326.0, + 1025.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1236.5, + 1056.0, + 1236.5, + 1056.0, + 1256.0, + 954.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.75, + 1257.5, + 1344.75, + 1257.5, + 1344.75, + 1278.5, + 1227.75, + 1278.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1318.5, + 1165.0, + 1318.5, + 1165.0, + 1337.0, + 1146.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1325.0, + 1084.0, + 1325.0, + 1084.0, + 1344.0, + 952.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 702.0, + 1203.0, + 702.0, + 1203.0, + 716.0, + 1190.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 778.0, + 1346.0, + 778.0, + 1346.0, + 786.0, + 1338.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 818.0, + 1118.0, + 818.0, + 1118.0, + 831.0, + 1081.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 707.0, + 1321.0, + 707.0, + 1321.0, + 764.0, + 1257.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 717.0, + 1076.0, + 717.0, + 1076.0, + 761.0, + 1007.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.25, + 760.5, + 1058.25, + 760.5, + 1058.25, + 782.5, + 957.25, + 782.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 769.5, + 1197.0, + 769.5, + 1197.0, + 833.5, + 1153.0, + 833.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 782.5, + 1346.0, + 782.5, + 1346.0, + 805.5, + 1229.0, + 805.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 790.0, + 1027.0, + 790.0, + 1027.0, + 813.5, + 989.0, + 813.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 827.5, + 1046.0, + 827.5, + 1046.0, + 844.0, + 955.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 845.0, + 1167.0, + 845.0, + 1167.0, + 859.5, + 1146.0, + 859.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1366.0, + 1320.0, + 1366.0, + 1320.0, + 1411.0, + 988.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 516.0, + 661.0, + 516.0, + 661.0, + 557.0, + 293.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 897.0, + 1314.0, + 897.0, + 1314.0, + 938.0, + 992.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 897.0, + 1359.0, + 897.0, + 1359.0, + 938.0, + 1356.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1844.0, + 1406.0, + 1844.0, + 1406.0, + 1889.0, + 292.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1878.0, + 1404.0, + 1878.0, + 1404.0, + 1920.0, + 292.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1913.0, + 1402.0, + 1913.0, + 1402.0, + 1945.0, + 296.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1404.0, + 1944.0, + 1404.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2005.0, + 296.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 652.0, + 938.0, + 652.0, + 938.0, + 691.0, + 294.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 687.0, + 937.0, + 687.0, + 937.0, + 718.0, + 295.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 716.0, + 936.0, + 716.0, + 936.0, + 748.0, + 295.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 747.0, + 935.0, + 747.0, + 935.0, + 781.0, + 295.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 779.0, + 936.0, + 779.0, + 936.0, + 809.0, + 296.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 806.0, + 937.0, + 806.0, + 937.0, + 841.0, + 294.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 836.0, + 937.0, + 836.0, + 937.0, + 871.0, + 294.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 870.0, + 937.0, + 870.0, + 937.0, + 902.0, + 295.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 893.0, + 620.0, + 893.0, + 620.0, + 935.0, + 290.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 893.0, + 942.0, + 893.0, + 942.0, + 935.0, + 808.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 927.0, + 459.0, + 927.0, + 459.0, + 965.0, + 292.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 927.0, + 939.0, + 927.0, + 939.0, + 965.0, + 561.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 700.0, + 231.0, + 700.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 264.0, + 727.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 787.0, + 260.0, + 787.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 260.0, + 1406.0, + 260.0, + 1406.0, + 298.0, + 839.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 288.0, + 1114.0, + 288.0, + 1114.0, + 328.0, + 292.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 288.0, + 1406.0, + 288.0, + 1406.0, + 328.0, + 1162.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 320.0, + 1233.0, + 320.0, + 1233.0, + 360.0, + 291.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 320.0, + 1407.0, + 320.0, + 1407.0, + 360.0, + 1268.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 349.0, + 415.0, + 349.0, + 415.0, + 393.0, + 293.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 349.0, + 481.0, + 349.0, + 481.0, + 393.0, + 468.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1129.0, + 937.0, + 1129.0, + 937.0, + 1161.0, + 295.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1159.0, + 937.0, + 1159.0, + 937.0, + 1191.0, + 296.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1188.0, + 937.0, + 1188.0, + 937.0, + 1226.0, + 293.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1221.0, + 935.0, + 1221.0, + 935.0, + 1251.0, + 295.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1250.0, + 936.0, + 1250.0, + 936.0, + 1282.0, + 294.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1280.0, + 937.0, + 1280.0, + 937.0, + 1315.0, + 294.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1311.0, + 937.0, + 1311.0, + 937.0, + 1344.0, + 295.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1342.0, + 937.0, + 1342.0, + 937.0, + 1376.0, + 295.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1371.0, + 937.0, + 1371.0, + 937.0, + 1406.0, + 294.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1392.0, + 347.0, + 1392.0, + 347.0, + 1441.0, + 289.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1392.0, + 942.0, + 1392.0, + 942.0, + 1441.0, + 521.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1424.0, + 351.0, + 1424.0, + 351.0, + 1471.0, + 288.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1424.0, + 771.0, + 1424.0, + 771.0, + 1471.0, + 516.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 1424.0, + 1410.0, + 1424.0, + 1410.0, + 1471.0, + 946.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1460.0, + 1406.0, + 1460.0, + 1406.0, + 1501.0, + 291.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1495.0, + 1101.0, + 1495.0, + 1101.0, + 1527.0, + 294.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1495.0, + 1405.0, + 1495.0, + 1405.0, + 1527.0, + 1122.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1516.0, + 746.0, + 1516.0, + 746.0, + 1565.0, + 290.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1516.0, + 1052.0, + 1516.0, + 1052.0, + 1565.0, + 958.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1516.0, + 1408.0, + 1516.0, + 1408.0, + 1565.0, + 1161.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1553.0, + 297.0, + 1553.0, + 297.0, + 1594.0, + 294.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1553.0, + 1407.0, + 1553.0, + 1407.0, + 1594.0, + 552.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1581.0, + 642.0, + 1581.0, + 642.0, + 1624.0, + 291.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1581.0, + 1407.0, + 1581.0, + 1407.0, + 1624.0, + 978.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1616.0, + 1405.0, + 1616.0, + 1405.0, + 1649.0, + 294.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1645.0, + 1266.0, + 1645.0, + 1266.0, + 1681.0, + 294.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1691.0, + 1404.0, + 1691.0, + 1404.0, + 1728.0, + 295.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1723.0, + 966.0, + 1723.0, + 966.0, + 1761.0, + 294.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1723.0, + 1008.0, + 1723.0, + 1008.0, + 1761.0, + 1000.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1723.0, + 1126.0, + 1723.0, + 1126.0, + 1761.0, + 1071.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1723.0, + 1405.0, + 1723.0, + 1405.0, + 1761.0, + 1180.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1753.0, + 771.0, + 1753.0, + 771.0, + 1791.0, + 295.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1753.0, + 873.0, + 1753.0, + 873.0, + 1791.0, + 804.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1753.0, + 947.0, + 1753.0, + 947.0, + 1791.0, + 935.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 949.0, + 339.0, + 949.0, + 339.0, + 998.0, + 290.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 949.0, + 831.0, + 949.0, + 831.0, + 998.0, + 506.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 949.0, + 1407.0, + 949.0, + 1407.0, + 998.0, + 868.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 988.0, + 626.0, + 988.0, + 626.0, + 1026.0, + 294.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 988.0, + 1175.0, + 988.0, + 1175.0, + 1026.0, + 657.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 988.0, + 1406.0, + 988.0, + 1406.0, + 1026.0, + 1211.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1014.0, + 296.0, + 1014.0, + 296.0, + 1059.0, + 291.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1014.0, + 876.0, + 1014.0, + 876.0, + 1059.0, + 392.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1014.0, + 1172.0, + 1014.0, + 1172.0, + 1059.0, + 1082.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1014.0, + 1410.0, + 1014.0, + 1410.0, + 1059.0, + 1280.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1049.0, + 396.0, + 1049.0, + 396.0, + 1088.0, + 292.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1049.0, + 1405.0, + 1049.0, + 1405.0, + 1088.0, + 625.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1082.0, + 296.0, + 1082.0, + 296.0, + 1117.0, + 292.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1082.0, + 721.0, + 1082.0, + 721.0, + 1117.0, + 638.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1082.0, + 1254.0, + 1082.0, + 1254.0, + 1117.0, + 749.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 577.0, + 1213.0, + 577.0, + 1213.0, + 613.0, + 297.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 577.0, + 1402.0, + 577.0, + 1402.0, + 613.0, + 1262.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 609.0, + 408.0, + 609.0, + 408.0, + 644.0, + 295.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 609.0, + 1096.0, + 609.0, + 1096.0, + 644.0, + 452.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 422.0, + 388.0, + 422.0, + 388.0, + 457.0, + 296.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 422.0, + 477.0, + 422.0, + 477.0, + 457.0, + 426.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 422.0, + 1405.0, + 422.0, + 1405.0, + 457.0, + 530.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 450.0, + 1181.0, + 450.0, + 1181.0, + 488.0, + 294.0, + 488.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1278, + 1405, + 1278, + 1405, + 1557, + 296, + 1557 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1662, + 1404, + 1662, + 1404, + 1849, + 297, + 1849 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1009, + 1404, + 1009, + 1404, + 1169, + 298, + 1169 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 297, + 219, + 1385, + 219, + 1385, + 595, + 297, + 595 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 639, + 1937, + 1056, + 1937, + 1056, + 2045, + 639, + 2045 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 300, + 722, + 1398, + 722, + 1398, + 817, + 300, + 817 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 630, + 907, + 1066, + 907, + 1066, + 999, + 630, + 999 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 297, + 1861, + 1401, + 1861, + 1401, + 1926, + 297, + 1926 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 294, + 832, + 1400, + 832, + 1400, + 895, + 294, + 895 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 657, + 1227, + 1040, + 1227, + 1040, + 1270, + 657, + 1270 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 643, + 1611, + 1052, + 1611, + 1052, + 1654, + 643, + 1654 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 296, + 1183, + 977, + 1183, + 977, + 1215, + 296, + 1215 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 294, + 1569, + 1400, + 1569, + 1400, + 1602, + 294, + 1602 + ], + "score": 0.92 + }, + { + "category_id": 4, + "poly": [ + 296, + 638, + 1397, + 638, + 1397, + 670, + 296, + 670 + ], + "score": 0.912 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1235, + 1399, + 1235, + 1399, + 1265, + 1366, + 1265 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 104, + 297, + 104 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1620, + 1399, + 1620, + 1399, + 1649, + 1366, + 1649 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 938, + 1400, + 938, + 1400, + 968, + 1366, + 968 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1974, + 1400, + 1974, + 1400, + 2004, + 1365, + 2004 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.779 + }, + { + "category_id": 14, + "poly": [ + 629, + 905, + 1067, + 905, + 1067, + 1001, + 629, + 1001 + ], + "score": 0.96, + "latex": "x _ { \\mathrm { s p a t } } = \\sigma \\left( \\sum _ { i } ^ { V } { x _ { \\mathrm { G A l } } [ i , : , : ] } \\right) \\in \\mathbb { R } ^ { H \\times W }" + }, + { + "category_id": 14, + "poly": [ + 641, + 1934, + 1059, + 1934, + 1059, + 2046, + 641, + 2046 + ], + "score": 0.95, + "latex": "x _ { \\mathrm { A n s } } = \\sigma \\left( \\sum _ { j , k } ^ { H , W } x _ { \\mathrm { G A 2 } } [ : , j , k ] \\right) \\in \\mathbb { R } ^ { V }" + }, + { + "category_id": 14, + "poly": [ + 655, + 1226, + 1042, + 1226, + 1042, + 1270, + 655, + 1270 + ], + "score": 0.93, + "latex": "x _ { \\mathrm { S A } } = f _ { \\mathrm { S A } } ( x _ { I } , x _ { \\mathrm { s p a t } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }" + }, + { + "category_id": 13, + "poly": [ + 520, + 1009, + 876, + 1009, + 876, + 1050, + 520, + 1050 + ], + "score": 0.93, + "latex": "\\sigma ( z ) _ { j } = \\exp ( z _ { j } ) / \\sum _ { j } \\exp ( z _ { j } )" + }, + { + "category_id": 14, + "poly": [ + 641, + 1610, + 1057, + 1610, + 1057, + 1653, + 641, + 1653 + ], + "score": 0.93, + "latex": "x _ { \\mathrm { G A 2 } } = f _ { \\mathrm { G A } } ( x _ { \\mathrm { S A } } , x _ { \\mathrm { s e n t } } ) \\in \\mathbb { R } ^ { V \\times H \\times W }" + }, + { + "category_id": 13, + "poly": [ + 802, + 865, + 893, + 865, + 893, + 892, + 802, + 892 + ], + "score": 0.89, + "latex": "H \\times W" + }, + { + "category_id": 13, + "poly": [ + 1110, + 1789, + 1163, + 1789, + 1163, + 1817, + 1110, + 1817 + ], + "score": 0.88, + "latex": "x _ { \\mathrm { s e n t } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1189, + 350, + 1189, + 350, + 1218, + 298, + 1218 + ], + "score": 0.88, + "latex": "x _ { \\mathrm { s p a t } }" + }, + { + "category_id": 13, + "poly": [ + 463, + 1050, + 520, + 1050, + 520, + 1077, + 463, + 1077 + ], + "score": 0.87, + "latex": "x _ { \\mathrm { G A l } }" + }, + { + "category_id": 13, + "poly": [ + 783, + 837, + 840, + 837, + 840, + 864, + 783, + 864 + ], + "score": 0.87, + "latex": "x _ { \\mathrm { G A 1 } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1575, + 342, + 1575, + 342, + 1602, + 297, + 1602 + ], + "score": 0.86, + "latex": "x _ { \\mathrm { S A } }" + }, + { + "category_id": 13, + "poly": [ + 400, + 1189, + 431, + 1189, + 431, + 1214, + 400, + 1214 + ], + "score": 0.84, + "latex": "x _ { I }" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 283.0, + 885.0, + 283.0, + 885.0, + 319.0, + 784.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 327.0, + 797.0, + 327.0, + 797.0, + 341.0, + 758.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 360.0, + 834.0, + 360.0, + 834.0, + 374.0, + 815.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 333.0, + 1294.0, + 333.0, + 1294.0, + 372.0, + 1231.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 386.0, + 398.0, + 386.0, + 398.0, + 409.0, + 343.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 398.0, + 1178.0, + 398.0, + 1178.0, + 409.0, + 1165.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 437.0, + 679.0, + 437.0, + 679.0, + 455.0, + 652.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 459.0, + 492.0, + 459.0, + 492.0, + 476.0, + 460.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 452.0, + 935.0, + 452.0, + 935.0, + 490.0, + 891.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 482.0, + 542.0, + 482.0, + 542.0, + 519.0, + 436.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 483.0, + 794.0, + 483.0, + 794.0, + 513.0, + 740.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 473.0, + 920.0, + 473.0, + 920.0, + 513.0, + 860.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 517.0, + 414.0, + 517.0, + 414.0, + 542.0, + 344.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 516.0, + 618.0, + 516.0, + 618.0, + 537.0, + 596.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 532.0, + 648.0, + 532.0, + 648.0, + 560.0, + 573.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 542.0, + 1090.0, + 542.0, + 1090.0, + 566.0, + 971.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 562.0, + 1050.0, + 562.0, + 1050.0, + 595.0, + 1011.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 277.0, + 1363.0, + 277.0, + 1363.0, + 318.5, + 1224.0, + 318.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.75, + 251.0, + 1355.75, + 251.0, + 1355.75, + 274.0, + 1244.75, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.75, + 235.5, + 1212.75, + 235.5, + 1212.75, + 252.5, + 1185.75, + 252.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 289.0, + 1014.0, + 289.0, + 1014.0, + 313.0, + 909.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 259.5, + 859.0, + 259.5, + 859.0, + 288.0, + 766.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 329.5, + 538.0, + 329.5, + 538.0, + 354.0, + 440.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 381.0, + 639.0, + 381.0, + 639.0, + 413.0, + 527.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 337.5, + 414.0, + 337.5, + 414.0, + 362.5, + 301.0, + 362.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 366.0, + 618.0, + 366.0, + 618.0, + 378.0, + 608.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.75, + 394.5, + 1119.75, + 394.5, + 1119.75, + 432.5, + 1003.75, + 432.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 355.0, + 409.0, + 355.0, + 409.0, + 378.5, + 306.0, + 378.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 397.5, + 1380.0, + 397.5, + 1380.0, + 423.0, + 1339.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 423.0, + 1162.0, + 423.0, + 1162.0, + 439.0, + 1143.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.25, + 431.5, + 1315.25, + 431.5, + 1315.25, + 456.5, + 1270.25, + 456.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 447.5, + 1148.0, + 447.5, + 1148.0, + 473.0, + 1125.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 459.0, + 871.0, + 459.0, + 871.0, + 485.5, + 825.0, + 485.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 477.5, + 1089.0, + 477.5, + 1089.0, + 494.5, + 1063.0, + 494.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.75, + 493.0, + 1260.75, + 493.0, + 1260.75, + 514.0, + 1215.75, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 519.5, + 960.0, + 519.5, + 960.0, + 545.5, + 910.0, + 545.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 540.0, + 1199.0, + 540.0, + 1199.0, + 561.5, + 1166.0, + 561.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 635.0, + 1399.0, + 635.0, + 1399.0, + 675.0, + 297.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1277.0, + 1405.0, + 1277.0, + 1405.0, + 1315.0, + 296.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1307.0, + 1405.0, + 1307.0, + 1405.0, + 1345.0, + 293.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1337.0, + 1406.0, + 1337.0, + 1406.0, + 1377.0, + 294.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1367.0, + 1405.0, + 1367.0, + 1405.0, + 1407.0, + 293.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1400.0, + 1405.0, + 1400.0, + 1405.0, + 1436.0, + 294.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1428.0, + 1405.0, + 1428.0, + 1405.0, + 1468.0, + 293.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1460.0, + 1405.0, + 1460.0, + 1405.0, + 1499.0, + 293.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1494.0, + 1405.0, + 1494.0, + 1405.0, + 1527.0, + 295.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1523.0, + 900.0, + 1523.0, + 900.0, + 1560.0, + 294.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1662.0, + 1405.0, + 1662.0, + 1405.0, + 1698.0, + 295.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1693.0, + 1404.0, + 1693.0, + 1404.0, + 1730.0, + 294.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1722.0, + 1406.0, + 1722.0, + 1406.0, + 1759.0, + 294.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1756.0, + 1406.0, + 1756.0, + 1406.0, + 1791.0, + 295.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1785.0, + 1109.0, + 1785.0, + 1109.0, + 1821.0, + 295.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1785.0, + 1406.0, + 1785.0, + 1406.0, + 1821.0, + 1164.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1817.0, + 996.0, + 1817.0, + 996.0, + 1849.0, + 296.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1006.0, + 519.0, + 1006.0, + 519.0, + 1050.0, + 293.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1006.0, + 1405.0, + 1006.0, + 1405.0, + 1050.0, + 877.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1046.0, + 462.0, + 1046.0, + 462.0, + 1080.0, + 297.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1046.0, + 1402.0, + 1046.0, + 1402.0, + 1080.0, + 521.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1075.0, + 1404.0, + 1075.0, + 1404.0, + 1109.0, + 296.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1104.0, + 1404.0, + 1104.0, + 1404.0, + 1142.0, + 294.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1135.0, + 703.0, + 1135.0, + 703.0, + 1173.0, + 293.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 724.0, + 1403.0, + 724.0, + 1403.0, + 759.0, + 296.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 756.0, + 1404.0, + 756.0, + 1404.0, + 790.0, + 295.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 785.0, + 1381.0, + 785.0, + 1381.0, + 819.0, + 296.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1860.0, + 1403.0, + 1860.0, + 1403.0, + 1896.0, + 295.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1894.0, + 1066.0, + 1894.0, + 1066.0, + 1926.0, + 295.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 827.0, + 782.0, + 827.0, + 782.0, + 871.0, + 292.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 827.0, + 1405.0, + 827.0, + 1405.0, + 871.0, + 841.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 862.0, + 801.0, + 862.0, + 801.0, + 898.0, + 294.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 862.0, + 1010.0, + 862.0, + 1010.0, + 898.0, + 894.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1180.0, + 297.0, + 1180.0, + 297.0, + 1220.0, + 293.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1180.0, + 399.0, + 1180.0, + 399.0, + 1220.0, + 351.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1180.0, + 979.0, + 1180.0, + 979.0, + 1220.0, + 432.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1566.0, + 296.0, + 1566.0, + 296.0, + 1608.0, + 293.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1566.0, + 1405.0, + 1566.0, + 1405.0, + 1608.0, + 343.0, + 1608.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1608, + 1404, + 1608, + 1404, + 1884, + 298, + 1884 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 553, + 1403, + 553, + 1403, + 829, + 298, + 829 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 911, + 1402, + 911, + 1402, + 1032, + 299, + 1032 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1048, + 1045, + 1048, + 1045, + 1323, + 298, + 1323 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1325, + 1403, + 1325, + 1403, + 1506, + 298, + 1506 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 415, + 1403, + 415, + 1403, + 539, + 299, + 539 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 1068, + 1056, + 1401, + 1056, + 1401, + 1243, + 1068, + 1243 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 299, + 230, + 1399, + 230, + 1399, + 323, + 299, + 323 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 294, + 338, + 1398, + 338, + 1398, + 400, + 294, + 400 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 300, + 1972, + 1402, + 1972, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.938 + }, + { + "category_id": 4, + "poly": [ + 1067, + 1254, + 1401, + 1254, + 1401, + 1311, + 1067, + 1311 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 302, + 1548, + 729, + 1548, + 729, + 1584, + 302, + 1584 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 299, + 864, + 544, + 864, + 544, + 894, + 299, + 894 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 299, + 1926, + 474, + 1926, + 474, + 1956, + 299, + 1956 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.775 + }, + { + "category_id": 13, + "poly": [ + 898, + 507, + 1040, + 507, + 1040, + 541, + 898, + 541 + ], + "score": 0.93, + "latex": "\\pi ( a _ { t } \\mid I _ { t } , T )" + }, + { + "category_id": 13, + "poly": [ + 1293, + 508, + 1393, + 508, + 1393, + 542, + 1293, + 542 + ], + "score": 0.92, + "latex": "V ( I _ { t } , T )" + }, + { + "category_id": 13, + "poly": [ + 456, + 369, + 630, + 369, + 630, + 404, + 456, + 404 + ], + "score": 0.92, + "latex": "x _ { \\mathrm { S } } = \\mathrm { v e c } ( x _ { \\mathrm { s p a t } } )" + }, + { + "category_id": 13, + "poly": [ + 1005, + 1353, + 1164, + 1353, + 1164, + 1386, + 1005, + 1386 + ], + "score": 0.9, + "latex": "( V \\times H \\times W )" + }, + { + "category_id": 13, + "poly": [ + 668, + 339, + 713, + 339, + 713, + 371, + 668, + 371 + ], + "score": 0.89, + "latex": "f _ { \\mathrm { D A } }" + }, + { + "category_id": 13, + "poly": [ + 1113, + 289, + 1154, + 289, + 1154, + 319, + 1113, + 319 + ], + "score": 0.89, + "latex": "\\mathbb { R } ^ { V }" + }, + { + "category_id": 13, + "poly": [ + 1066, + 344, + 1119, + 344, + 1119, + 369, + 1066, + 369 + ], + "score": 0.87, + "latex": "x _ { \\mathrm { A n s } }" + }, + { + "category_id": 13, + "poly": [ + 466, + 236, + 525, + 236, + 525, + 262, + 466, + 262 + ], + "score": 0.86, + "latex": "x _ { \\mathrm { G A } 2 }" + }, + { + "category_id": 13, + "poly": [ + 1182, + 421, + 1215, + 421, + 1215, + 447, + 1182, + 447 + ], + "score": 0.84, + "latex": "x _ { S }" + }, + { + "category_id": 13, + "poly": [ + 509, + 2004, + 543, + 2004, + 543, + 2031, + 509, + 2031 + ], + "score": 0.8, + "latex": "+ 1" + }, + { + "category_id": 13, + "poly": [ + 1104, + 448, + 1124, + 448, + 1124, + 474, + 1104, + 474 + ], + "score": 0.74, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 741, + 450, + 756, + 450, + 756, + 475, + 741, + 475 + ], + "score": 0.65, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 497, + 1384, + 523, + 1384, + 523, + 1412, + 497, + 1412 + ], + "score": 0.58, + "latex": "V" + }, + { + "category_id": 13, + "poly": [ + 757, + 369, + 787, + 369, + 787, + 402, + 757, + 402 + ], + "score": 0.34, + "latex": "( . )" + }, + { + "category_id": 13, + "poly": [ + 716, + 368, + 787, + 368, + 787, + 402, + 716, + 402 + ], + "score": 0.3, + "latex": "\\mathrm { v e c } \\bar { ( . ) }" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1253.0, + 1404.0, + 1253.0, + 1404.0, + 1283.0, + 1065.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1280.0, + 1331.0, + 1280.0, + 1331.0, + 1311.0, + 1066.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1547.0, + 734.0, + 1547.0, + 734.0, + 1587.0, + 293.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 860.0, + 548.0, + 860.0, + 548.0, + 901.0, + 293.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1922.0, + 477.0, + 1922.0, + 477.0, + 1962.0, + 293.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1610.0, + 1406.0, + 1610.0, + 1406.0, + 1643.0, + 294.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1635.0, + 1405.0, + 1635.0, + 1405.0, + 1674.0, + 292.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1670.0, + 1402.0, + 1670.0, + 1402.0, + 1703.0, + 294.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1699.0, + 1404.0, + 1699.0, + 1404.0, + 1736.0, + 293.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1766.0, + 293.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1759.0, + 1405.0, + 1759.0, + 1405.0, + 1797.0, + 293.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1788.0, + 1404.0, + 1788.0, + 1404.0, + 1827.0, + 292.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1823.0, + 1404.0, + 1823.0, + 1404.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1853.0, + 1381.0, + 1853.0, + 1381.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 555.0, + 1405.0, + 555.0, + 1405.0, + 588.0, + 296.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 580.0, + 1405.0, + 580.0, + 1405.0, + 622.0, + 292.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 615.0, + 1403.0, + 615.0, + 1403.0, + 648.0, + 294.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 643.0, + 1405.0, + 643.0, + 1405.0, + 681.0, + 293.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 674.0, + 1405.0, + 674.0, + 1405.0, + 712.0, + 293.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 707.0, + 1404.0, + 707.0, + 1404.0, + 741.0, + 292.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 737.0, + 1405.0, + 737.0, + 1405.0, + 773.0, + 294.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 767.0, + 1405.0, + 767.0, + 1405.0, + 803.0, + 293.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 798.0, + 1403.0, + 798.0, + 1403.0, + 831.0, + 294.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 908.0, + 1407.0, + 908.0, + 1407.0, + 946.0, + 293.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 940.0, + 1407.0, + 940.0, + 1407.0, + 976.0, + 294.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 973.0, + 1406.0, + 973.0, + 1406.0, + 1008.0, + 294.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 999.0, + 1224.0, + 999.0, + 1224.0, + 1039.0, + 293.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1047.0, + 1046.0, + 1047.0, + 1046.0, + 1082.0, + 296.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1078.0, + 1045.0, + 1078.0, + 1045.0, + 1113.0, + 295.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1107.0, + 1047.0, + 1107.0, + 1047.0, + 1144.0, + 295.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1141.0, + 1045.0, + 1141.0, + 1045.0, + 1173.0, + 296.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1171.0, + 1047.0, + 1171.0, + 1047.0, + 1203.0, + 294.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1200.0, + 1046.0, + 1200.0, + 1046.0, + 1233.0, + 295.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1230.0, + 1047.0, + 1230.0, + 1047.0, + 1268.0, + 292.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1260.0, + 1048.0, + 1260.0, + 1048.0, + 1295.0, + 293.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1292.0, + 1047.0, + 1292.0, + 1047.0, + 1327.0, + 295.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1324.0, + 1404.0, + 1324.0, + 1404.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1353.0, + 1004.0, + 1353.0, + 1004.0, + 1388.0, + 293.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1353.0, + 1405.0, + 1353.0, + 1405.0, + 1388.0, + 1165.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1382.0, + 496.0, + 1382.0, + 496.0, + 1418.0, + 293.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1382.0, + 1405.0, + 1382.0, + 1405.0, + 1418.0, + 524.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1414.0, + 1405.0, + 1414.0, + 1405.0, + 1449.0, + 293.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1444.0, + 1403.0, + 1444.0, + 1403.0, + 1479.0, + 293.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1472.0, + 804.0, + 1472.0, + 804.0, + 1510.0, + 295.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 414.0, + 1181.0, + 414.0, + 1181.0, + 450.0, + 293.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 450.0, + 1216.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 445.0, + 740.0, + 445.0, + 740.0, + 482.0, + 294.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 445.0, + 1103.0, + 445.0, + 1103.0, + 482.0, + 757.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 445.0, + 1407.0, + 445.0, + 1407.0, + 482.0, + 1125.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 475.0, + 1405.0, + 475.0, + 1405.0, + 511.0, + 293.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 506.0, + 897.0, + 506.0, + 897.0, + 542.0, + 293.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 506.0, + 1292.0, + 506.0, + 1292.0, + 542.0, + 1041.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 506.0, + 1403.0, + 506.0, + 1403.0, + 542.0, + 1394.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 465.0, + 229.0, + 465.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 267.0, + 526.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 289.0, + 1112.0, + 289.0, + 1112.0, + 328.0, + 291.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 289.0, + 1168.0, + 289.0, + 1168.0, + 328.0, + 1155.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 337.0, + 667.0, + 337.0, + 667.0, + 373.0, + 295.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 337.0, + 1065.0, + 337.0, + 1065.0, + 373.0, + 714.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 337.0, + 1404.0, + 337.0, + 1404.0, + 373.0, + 1120.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 368.0, + 455.0, + 368.0, + 455.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 368.0, + 715.0, + 368.0, + 715.0, + 403.0, + 631.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 368.0, + 1152.0, + 368.0, + 1152.0, + 403.0, + 788.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 508.0, + 2002.0, + 508.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2037.0, + 544.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1207, + 1404, + 1207, + 1404, + 1635, + 297, + 1635 + ], + "score": 0.984 + }, + { + "category_id": 5, + "poly": [ + 300, + 606, + 1400, + 606, + 1400, + 914, + 300, + 914 + ], + "score": 0.978, + "html": "
EasyHard
No AuxAuxNo AuxAux
ModelSGNEQASGNEQASGNEQASGNEQA
Text only0.20.330.20.330.20.330.20.33
Image only0.200.090.210.080.160.080.150.08
Concat0.330.210.310.190.20.260.390.22
GA0.270.180.350.240.180.110.220.24
FiLM0.240.110.340.120.120.030.250.15
PACMAN0.260.120.330.100.290.330.110.27
Dual-Attention0.860.530.960.580.860.380.820.59
" + }, + { + "category_id": 1, + "poly": [ + 298, + 977, + 1403, + 977, + 1403, + 1192, + 298, + 1192 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1713, + 1403, + 1713, + 1403, + 1867, + 298, + 1867 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 404, + 219, + 1292, + 219, + 1292, + 502, + 404, + 502 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 301, + 1880, + 1398, + 1880, + 1398, + 1974, + 301, + 1974 + ], + "score": 0.949 + }, + { + "category_id": 2, + "poly": [ + 318, + 2006, + 1355, + 2006, + 1355, + 2034, + 318, + 2034 + ], + "score": 0.908 + }, + { + "category_id": 6, + "poly": [ + 374, + 574, + 1313, + 574, + 1313, + 603, + 374, + 603 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 300, + 1669, + 572, + 1669, + 572, + 1698, + 300, + 1698 + ], + "score": 0.875 + }, + { + "category_id": 4, + "poly": [ + 337, + 509, + 1358, + 509, + 1358, + 540, + 337, + 540 + ], + "score": 0.837 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.835 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.676 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 855, + 76, + 855, + 104, + 300, + 104 + ], + "score": 0.269 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2111, + 841, + 2111 + ], + "score": 0.209 + }, + { + "category_id": 13, + "poly": [ + 1051, + 1777, + 1104, + 1777, + 1104, + 1806, + 1051, + 1806 + ], + "score": 0.87, + "latex": "x _ { \\mathrm { s e n t } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1778, + 370, + 1778, + 370, + 1806, + 297, + 1806 + ], + "score": 0.83, + "latex": "x _ { B o W }" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 234.0, + 460.0, + 234.0, + 460.0, + 264.0, + 421.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 221.0, + 679.0, + 221.0, + 679.0, + 255.0, + 613.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 234.0, + 905.0, + 234.0, + 905.0, + 264.0, + 864.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 223.0, + 1124.0, + 223.0, + 1124.0, + 253.0, + 1055.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 271.0, + 458.0, + 271.0, + 458.0, + 299.0, + 423.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 272.0, + 903.0, + 272.0, + 903.0, + 299.0, + 867.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 294.0, + 457.0, + 294.0, + 457.0, + 385.0, + 404.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 292.0, + 903.0, + 292.0, + 903.0, + 386.0, + 846.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 343.0, + 458.0, + 343.0, + 458.0, + 370.0, + 423.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 341.0, + 902.0, + 341.0, + 902.0, + 371.0, + 865.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 377.0, + 460.0, + 377.0, + 460.0, + 407.0, + 422.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 379.0, + 902.0, + 379.0, + 902.0, + 405.0, + 867.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 427.0, + 551.0, + 427.0, + 551.0, + 454.0, + 515.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 427.0, + 626.0, + 427.0, + 626.0, + 453.0, + 590.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 427.0, + 703.0, + 427.0, + 703.0, + 453.0, + 666.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 427.0, + 778.0, + 427.0, + 778.0, + 453.0, + 742.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 426.0, + 852.0, + 426.0, + 852.0, + 453.0, + 816.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 431.0, + 984.0, + 431.0, + 984.0, + 450.0, + 967.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 428.0, + 1062.0, + 428.0, + 1062.0, + 452.0, + 1041.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 429.0, + 1138.0, + 429.0, + 1138.0, + 452.0, + 1118.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 430.0, + 1211.0, + 430.0, + 1211.0, + 449.0, + 1195.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 422.0, + 1296.0, + 422.0, + 1296.0, + 472.0, + 1241.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 446.0, + 705.0, + 446.0, + 705.0, + 479.0, + 588.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 441.0, + 843.0, + 441.0, + 843.0, + 467.0, + 803.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 447.0, + 1148.0, + 447.0, + 1148.0, + 478.0, + 1031.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 475.0, + 571.0, + 475.0, + 571.0, + 506.0, + 488.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 478.0, + 704.0, + 478.0, + 704.0, + 504.0, + 612.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 479.0, + 809.0, + 479.0, + 809.0, + 503.0, + 744.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 479.0, + 884.0, + 479.0, + 884.0, + 502.0, + 852.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 478.0, + 974.0, + 478.0, + 974.0, + 503.0, + 927.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 477.0, + 1088.0, + 477.0, + 1088.0, + 503.0, + 1015.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 479.0, + 1242.0, + 479.0, + 1242.0, + 502.0, + 1131.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.75, + 413.5, + 463.75, + 413.5, + 463.75, + 453.0, + 432.75, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.75, + 412.0, + 906.75, + 412.0, + 906.75, + 450.0, + 870.75, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1997.0, + 1359.0, + 1997.0, + 1359.0, + 2040.0, + 330.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 571.0, + 1315.0, + 571.0, + 1315.0, + 608.0, + 382.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1665.0, + 575.0, + 1665.0, + 575.0, + 1704.0, + 295.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 505.0, + 1364.0, + 505.0, + 1364.0, + 545.0, + 333.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1236.0, + 1405.0, + 1236.0, + 1405.0, + 1271.0, + 294.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1267.0, + 1406.0, + 1267.0, + 1406.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 1406.0, + 1298.0, + 1406.0, + 1333.0, + 294.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1330.0, + 1406.0, + 1330.0, + 1406.0, + 1365.0, + 292.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1361.0, + 1404.0, + 1361.0, + 1404.0, + 1393.0, + 295.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1392.0, + 1405.0, + 1392.0, + 1405.0, + 1424.0, + 295.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1418.0, + 1406.0, + 1418.0, + 1406.0, + 1455.0, + 292.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1450.0, + 1405.0, + 1450.0, + 1405.0, + 1484.0, + 294.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1477.0, + 1406.0, + 1477.0, + 1406.0, + 1520.0, + 291.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1511.0, + 1406.0, + 1511.0, + 1406.0, + 1547.0, + 294.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1540.0, + 1407.0, + 1540.0, + 1407.0, + 1580.0, + 294.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1572.0, + 1405.0, + 1572.0, + 1405.0, + 1608.0, + 294.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1603.0, + 1402.0, + 1603.0, + 1402.0, + 1638.0, + 294.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 977.0, + 1405.0, + 977.0, + 1405.0, + 1012.0, + 294.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1009.0, + 1405.0, + 1009.0, + 1405.0, + 1041.0, + 294.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 1403.0, + 1039.0, + 1403.0, + 1073.0, + 294.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1071.0, + 1404.0, + 1071.0, + 1404.0, + 1102.0, + 296.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1098.0, + 1405.0, + 1098.0, + 1405.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1405.0, + 1130.0, + 1405.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1162.0, + 1361.0, + 1162.0, + 1361.0, + 1195.0, + 293.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1715.0, + 1404.0, + 1715.0, + 1404.0, + 1748.0, + 297.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1744.0, + 1404.0, + 1744.0, + 1404.0, + 1777.0, + 296.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1773.0, + 296.0, + 1773.0, + 296.0, + 1814.0, + 292.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1773.0, + 1050.0, + 1773.0, + 1050.0, + 1814.0, + 371.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1773.0, + 1407.0, + 1773.0, + 1407.0, + 1814.0, + 1105.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1840.0, + 294.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1834.0, + 1405.0, + 1834.0, + 1405.0, + 1871.0, + 296.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1402.0, + 1912.0, + 1402.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1940.0, + 1403.0, + 1940.0, + 1403.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1638, + 1404, + 1638, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1044, + 1404, + 1044, + 1404, + 1351, + 298, + 1351 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 296, + 693, + 824, + 693, + 824, + 977, + 296, + 977 + ], + "score": 0.98, + "html": "
ModelNo Aux SGN EQAAux SGN
w/o SAEQA
w/o GA10.20 0.16 0.250.20 0.15 0.16 0.38
w/o GA20.14 0.80 0.330.97 0.15
w/o Task Indicator0.79 0.470.96 0.56
w/o Reward Shaping0.82 0.490.93 0.51
DA Single-Task0.63 0.310.91 0.34
DA Multi-Task0.86 0.530.96 0.58
" + }, + { + "category_id": 5, + "poly": [ + 878, + 693, + 1402, + 693, + 1402, + 976, + 878, + 976 + ], + "score": 0.979, + "html": "
InstructionEasyHard
Go to the pillar1.000.71
Go to the red object0.990.89
Go to the tall/short pillar0.990.68
Go to the <known_color>pillar.1.000.79
Go to the red <known_object>1.000.93
Go to the largest/smallest red object0.950.69
Go to the tall/short red pillar0.990.88
Go to the red pillar0.990.82
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1438, + 1403, + 1438, + 1403, + 1622, + 298, + 1622 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 406, + 221, + 1290, + 221, + 1290, + 502, + 406, + 502 + ], + "score": 0.967 + }, + { + "category_id": 6, + "poly": [ + 296, + 609, + 830, + 609, + 830, + 692, + 296, + 692 + ], + "score": 0.952 + }, + { + "category_id": 6, + "poly": [ + 879, + 610, + 1401, + 610, + 1401, + 693, + 879, + 693 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 301, + 1391, + 845, + 1391, + 845, + 1422, + 301, + 1422 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.831 + }, + { + "category_id": 4, + "poly": [ + 290, + 511, + 1400, + 511, + 1400, + 568, + 290, + 568 + ], + "score": 0.827 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.792 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 855, + 76, + 855, + 104, + 300, + 104 + ], + "score": 0.238 + }, + { + "category_id": 1, + "poly": [ + 290, + 511, + 1400, + 511, + 1400, + 568, + 290, + 568 + ], + "score": 0.105 + }, + { + "category_id": 13, + "poly": [ + 787, + 1882, + 901, + 1882, + 901, + 1912, + 787, + 1912 + ], + "score": 0.89, + "latex": "( H \\times W )" + }, + { + "category_id": 13, + "poly": [ + 995, + 1674, + 1026, + 1674, + 1026, + 1699, + 995, + 1699 + ], + "score": 0.82, + "latex": "x _ { I }" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 234.0, + 460.0, + 234.0, + 460.0, + 264.0, + 422.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 221.0, + 685.0, + 221.0, + 685.0, + 255.0, + 604.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 234.0, + 901.0, + 234.0, + 901.0, + 263.0, + 860.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 219.0, + 1111.0, + 219.0, + 1111.0, + 255.0, + 1059.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 271.0, + 457.0, + 271.0, + 457.0, + 299.0, + 423.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 271.0, + 898.0, + 271.0, + 898.0, + 298.0, + 862.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 295.0, + 457.0, + 295.0, + 457.0, + 384.0, + 404.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 295.0, + 897.0, + 295.0, + 897.0, + 385.0, + 843.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 343.0, + 456.0, + 343.0, + 456.0, + 370.0, + 424.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 344.0, + 897.0, + 344.0, + 897.0, + 369.0, + 862.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 379.0, + 457.0, + 379.0, + 457.0, + 406.0, + 423.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 380.0, + 898.0, + 380.0, + 898.0, + 406.0, + 862.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 427.0, + 550.0, + 427.0, + 550.0, + 454.0, + 514.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 425.0, + 625.0, + 425.0, + 625.0, + 454.0, + 589.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 429.0, + 699.0, + 429.0, + 699.0, + 452.0, + 666.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 428.0, + 775.0, + 428.0, + 775.0, + 453.0, + 738.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 427.0, + 850.0, + 427.0, + 850.0, + 453.0, + 812.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 428.0, + 989.0, + 428.0, + 989.0, + 454.0, + 954.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 427.0, + 1065.0, + 427.0, + 1065.0, + 453.0, + 1029.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 427.0, + 1140.0, + 427.0, + 1140.0, + 453.0, + 1103.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 429.0, + 1214.0, + 429.0, + 1214.0, + 452.0, + 1180.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 427.0, + 1290.0, + 427.0, + 1290.0, + 453.0, + 1253.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 448.0, + 704.0, + 448.0, + 704.0, + 479.0, + 586.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 442.0, + 839.0, + 442.0, + 839.0, + 467.0, + 800.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 447.0, + 1142.0, + 447.0, + 1142.0, + 479.0, + 1027.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 442.0, + 1281.0, + 442.0, + 1281.0, + 468.0, + 1240.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 480.0, + 570.0, + 480.0, + 570.0, + 502.0, + 457.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 479.0, + 676.0, + 479.0, + 676.0, + 504.0, + 602.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 479.0, + 789.0, + 479.0, + 789.0, + 504.0, + 716.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 479.0, + 902.0, + 479.0, + 902.0, + 504.0, + 830.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 481.0, + 1083.0, + 481.0, + 1083.0, + 502.0, + 942.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 474.0, + 1280.0, + 474.0, + 1280.0, + 507.0, + 1123.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 415.5, + 465.0, + 415.5, + 465.0, + 451.5, + 432.0, + 451.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 415.5, + 904.0, + 415.5, + 904.0, + 452.5, + 872.0, + 452.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 608.0, + 833.0, + 608.0, + 833.0, + 640.0, + 297.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 636.0, + 831.0, + 636.0, + 831.0, + 668.0, + 297.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 664.0, + 715.0, + 664.0, + 715.0, + 696.0, + 295.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 605.0, + 1403.0, + 605.0, + 1403.0, + 643.0, + 876.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 636.0, + 1405.0, + 636.0, + 1405.0, + 668.0, + 876.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 665.0, + 1291.0, + 665.0, + 1291.0, + 695.0, + 879.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1390.0, + 850.0, + 1390.0, + 850.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 509.0, + 1403.0, + 509.0, + 1403.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 540.0, + 874.0, + 540.0, + 874.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1638.0, + 1405.0, + 1638.0, + 1405.0, + 1671.0, + 296.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1668.0, + 994.0, + 1668.0, + 994.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1668.0, + 1408.0, + 1668.0, + 1408.0, + 1704.0, + 1027.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1698.0, + 1406.0, + 1698.0, + 1406.0, + 1734.0, + 294.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1726.0, + 1405.0, + 1726.0, + 1405.0, + 1765.0, + 293.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1761.0, + 1404.0, + 1761.0, + 1404.0, + 1793.0, + 296.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1791.0, + 1404.0, + 1791.0, + 1404.0, + 1823.0, + 296.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1857.0, + 294.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1848.0, + 1407.0, + 1848.0, + 1407.0, + 1886.0, + 292.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1877.0, + 786.0, + 1877.0, + 786.0, + 1920.0, + 291.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1877.0, + 1406.0, + 1877.0, + 1406.0, + 1920.0, + 902.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1408.0, + 1910.0, + 1408.0, + 1950.0, + 292.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1404.0, + 1941.0, + 1404.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1045.0, + 1406.0, + 1045.0, + 1406.0, + 1079.0, + 293.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1077.0, + 1406.0, + 1077.0, + 1406.0, + 1110.0, + 293.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1106.0, + 1405.0, + 1106.0, + 1405.0, + 1142.0, + 294.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1131.0, + 1406.0, + 1131.0, + 1406.0, + 1177.0, + 292.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1167.0, + 1406.0, + 1167.0, + 1406.0, + 1201.0, + 293.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1196.0, + 1406.0, + 1196.0, + 1406.0, + 1232.0, + 293.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1227.0, + 1407.0, + 1227.0, + 1407.0, + 1264.0, + 292.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1259.0, + 1406.0, + 1259.0, + 1406.0, + 1294.0, + 294.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1285.0, + 1406.0, + 1285.0, + 1406.0, + 1329.0, + 292.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1320.0, + 1404.0, + 1320.0, + 1404.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1437.0, + 1405.0, + 1437.0, + 1405.0, + 1473.0, + 296.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1470.0, + 1404.0, + 1470.0, + 1404.0, + 1502.0, + 294.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1501.0, + 1404.0, + 1501.0, + 1404.0, + 1533.0, + 297.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1527.0, + 1405.0, + 1527.0, + 1405.0, + 1564.0, + 293.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1559.0, + 1405.0, + 1559.0, + 1405.0, + 1595.0, + 292.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1590.0, + 1345.0, + 1590.0, + 1345.0, + 1626.0, + 292.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 509.0, + 1403.0, + 509.0, + 1403.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 540.0, + 874.0, + 540.0, + 874.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 389, + 1404, + 389, + 1404, + 634, + 298, + 634 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 230, + 1402, + 230, + 1402, + 293, + 297, + 293 + ], + "score": 0.938 + }, + { + "category_id": 0, + "poly": [ + 299, + 331, + 544, + 331, + 544, + 367, + 299, + 367 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 300, + 683, + 488, + 683, + 488, + 715, + 300, + 715 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.774 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.766 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1402, + 1942, + 1402, + 2033, + 300, + 2033 + ], + "score": 0.712 + }, + { + "category_id": 1, + "poly": [ + 297, + 1233, + 1405, + 1233, + 1405, + 1328, + 297, + 1328 + ], + "score": 0.702 + }, + { + "category_id": 1, + "poly": [ + 303, + 1352, + 1400, + 1352, + 1400, + 1448, + 303, + 1448 + ], + "score": 0.695 + }, + { + "category_id": 1, + "poly": [ + 295, + 1059, + 1379, + 1059, + 1379, + 1094, + 295, + 1094 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 292, + 1853, + 1400, + 1853, + 1400, + 1916, + 292, + 1916 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 302, + 1116, + 1398, + 1116, + 1398, + 1210, + 302, + 1210 + ], + "score": 0.686 + }, + { + "category_id": 1, + "poly": [ + 293, + 971, + 1401, + 971, + 1401, + 1035, + 293, + 1035 + ], + "score": 0.685 + }, + { + "category_id": 1, + "poly": [ + 299, + 1558, + 1403, + 1558, + 1403, + 1654, + 299, + 1654 + ], + "score": 0.682 + }, + { + "category_id": 1, + "poly": [ + 293, + 1795, + 1383, + 1795, + 1383, + 1830, + 293, + 1830 + ], + "score": 0.665 + }, + { + "category_id": 1, + "poly": [ + 299, + 735, + 1404, + 735, + 1404, + 828, + 299, + 828 + ], + "score": 0.653 + }, + { + "category_id": 1, + "poly": [ + 299, + 1677, + 1403, + 1677, + 1403, + 1771, + 299, + 1771 + ], + "score": 0.651 + }, + { + "category_id": 1, + "poly": [ + 294, + 1470, + 1402, + 1470, + 1402, + 1535, + 294, + 1535 + ], + "score": 0.648 + }, + { + "category_id": 1, + "poly": [ + 299, + 853, + 1404, + 853, + 1404, + 945, + 299, + 945 + ], + "score": 0.551 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.144 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 324.0, + 550.0, + 324.0, + 550.0, + 376.0, + 292.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 683.0, + 490.0, + 683.0, + 490.0, + 719.0, + 296.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 388.0, + 1404.0, + 388.0, + 1404.0, + 425.0, + 294.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 420.0, + 1404.0, + 420.0, + 1404.0, + 454.0, + 296.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 452.0, + 1405.0, + 452.0, + 1405.0, + 485.0, + 294.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 480.0, + 1404.0, + 480.0, + 1404.0, + 515.0, + 293.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 510.0, + 1402.0, + 510.0, + 1402.0, + 544.0, + 296.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 541.0, + 1404.0, + 541.0, + 1404.0, + 575.0, + 294.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 573.0, + 1405.0, + 573.0, + 1405.0, + 606.0, + 294.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 604.0, + 1252.0, + 604.0, + 1252.0, + 637.0, + 296.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 263.0, + 516.0, + 263.0, + 516.0, + 295.0, + 295.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2007.0, + 324.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 2004.0, + 707.0, + 2004.0, + 707.0, + 2035.0, + 320.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1234.0, + 1405.0, + 1234.0, + 1405.0, + 1268.0, + 295.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1262.0, + 1405.0, + 1262.0, + 1405.0, + 1305.0, + 321.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1296.0, + 505.0, + 1296.0, + 505.0, + 1328.0, + 324.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1353.0, + 1404.0, + 1353.0, + 1404.0, + 1388.0, + 298.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1384.0, + 1405.0, + 1384.0, + 1405.0, + 1418.0, + 323.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1415.0, + 1219.0, + 1415.0, + 1219.0, + 1453.0, + 324.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1055.0, + 1383.0, + 1055.0, + 1383.0, + 1098.0, + 294.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1403.0, + 1851.0, + 1403.0, + 1890.0, + 295.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1885.0, + 865.0, + 1885.0, + 865.0, + 1918.0, + 322.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1117.0, + 1402.0, + 1117.0, + 1402.0, + 1151.0, + 297.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1147.0, + 1403.0, + 1147.0, + 1403.0, + 1184.0, + 322.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1180.0, + 706.0, + 1180.0, + 706.0, + 1211.0, + 320.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 971.0, + 1403.0, + 971.0, + 1403.0, + 1007.0, + 295.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1000.0, + 1089.0, + 1000.0, + 1089.0, + 1036.0, + 324.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1557.0, + 1403.0, + 1557.0, + 1403.0, + 1596.0, + 294.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1591.0, + 1405.0, + 1591.0, + 1405.0, + 1626.0, + 326.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1619.0, + 1325.0, + 1619.0, + 1325.0, + 1657.0, + 322.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 1386.0, + 1790.0, + 1386.0, + 1834.0, + 294.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 732.0, + 1404.0, + 732.0, + 1404.0, + 772.0, + 294.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 765.0, + 1406.0, + 765.0, + 1406.0, + 806.0, + 319.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 798.0, + 907.0, + 798.0, + 907.0, + 828.0, + 323.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1677.0, + 1407.0, + 1677.0, + 1407.0, + 1714.0, + 294.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1709.0, + 1404.0, + 1709.0, + 1404.0, + 1743.0, + 323.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1739.0, + 1099.0, + 1739.0, + 1099.0, + 1776.0, + 320.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1472.0, + 1405.0, + 1472.0, + 1405.0, + 1508.0, + 296.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1503.0, + 1315.0, + 1503.0, + 1315.0, + 1537.0, + 322.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 851.0, + 1404.0, + 851.0, + 1404.0, + 889.0, + 294.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 882.0, + 1404.0, + 882.0, + 1404.0, + 918.0, + 321.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 911.0, + 398.0, + 911.0, + 398.0, + 948.0, + 319.0, + 948.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 293, + 931, + 1399, + 931, + 1399, + 996, + 293, + 996 + ], + "score": 0.736 + }, + { + "category_id": 1, + "poly": [ + 299, + 1294, + 1401, + 1294, + 1401, + 1388, + 299, + 1388 + ], + "score": 0.731 + }, + { + "category_id": 1, + "poly": [ + 304, + 1603, + 1402, + 1603, + 1402, + 1698, + 304, + 1698 + ], + "score": 0.727 + }, + { + "category_id": 1, + "poly": [ + 294, + 1014, + 1403, + 1014, + 1403, + 1079, + 294, + 1079 + ], + "score": 0.727 + }, + { + "category_id": 1, + "poly": [ + 295, + 1210, + 1401, + 1210, + 1401, + 1276, + 295, + 1276 + ], + "score": 0.719 + }, + { + "category_id": 1, + "poly": [ + 299, + 1491, + 1400, + 1491, + 1400, + 1585, + 299, + 1585 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 291, + 1407, + 1401, + 1407, + 1401, + 1472, + 291, + 1472 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 292, + 1718, + 1399, + 1718, + 1399, + 1782, + 292, + 1782 + ], + "score": 0.677 + }, + { + "category_id": 1, + "poly": [ + 295, + 1801, + 1402, + 1801, + 1402, + 1865, + 295, + 1865 + ], + "score": 0.671 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1399, + 229, + 1399, + 323, + 298, + 323 + ], + "score": 0.663 + }, + { + "category_id": 1, + "poly": [ + 297, + 1097, + 1403, + 1097, + 1403, + 1191, + 297, + 1191 + ], + "score": 0.648 + }, + { + "category_id": 1, + "poly": [ + 300, + 1884, + 1403, + 1884, + 1403, + 2008, + 300, + 2008 + ], + "score": 0.62 + }, + { + "category_id": 1, + "poly": [ + 297, + 622, + 1399, + 622, + 1399, + 686, + 297, + 686 + ], + "score": 0.606 + }, + { + "category_id": 1, + "poly": [ + 300, + 508, + 1399, + 508, + 1399, + 602, + 300, + 602 + ], + "score": 0.579 + }, + { + "category_id": 1, + "poly": [ + 297, + 426, + 1400, + 426, + 1400, + 490, + 297, + 490 + ], + "score": 0.567 + }, + { + "category_id": 1, + "poly": [ + 298, + 818, + 1405, + 818, + 1405, + 912, + 298, + 912 + ], + "score": 0.563 + }, + { + "category_id": 1, + "poly": [ + 295, + 343, + 1400, + 343, + 1400, + 407, + 295, + 407 + ], + "score": 0.555 + }, + { + "category_id": 1, + "poly": [ + 296, + 704, + 1401, + 704, + 1401, + 800, + 296, + 800 + ], + "score": 0.467 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 933.0, + 1404.0, + 933.0, + 1404.0, + 969.0, + 295.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 962.0, + 1225.0, + 962.0, + 1225.0, + 997.0, + 322.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1294.0, + 1406.0, + 1294.0, + 1406.0, + 1333.0, + 293.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1324.0, + 1406.0, + 1324.0, + 1406.0, + 1364.0, + 320.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1357.0, + 1372.0, + 1357.0, + 1372.0, + 1391.0, + 324.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1641.0, + 296.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1633.0, + 1406.0, + 1633.0, + 1406.0, + 1672.0, + 320.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1660.0, + 408.0, + 1660.0, + 408.0, + 1701.0, + 320.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1017.0, + 1404.0, + 1017.0, + 1404.0, + 1049.0, + 297.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1045.0, + 1272.0, + 1045.0, + 1272.0, + 1081.0, + 322.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1251.0, + 292.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1243.0, + 1058.0, + 1243.0, + 1058.0, + 1277.0, + 323.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1488.0, + 1401.0, + 1488.0, + 1401.0, + 1528.0, + 293.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1521.0, + 1402.0, + 1521.0, + 1402.0, + 1557.0, + 322.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1553.0, + 503.0, + 1553.0, + 503.0, + 1588.0, + 323.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1406.0, + 1407.0, + 1406.0, + 1407.0, + 1445.0, + 296.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1438.0, + 1156.0, + 1438.0, + 1156.0, + 1474.0, + 323.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1717.0, + 1403.0, + 1717.0, + 1403.0, + 1755.0, + 294.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1750.0, + 1033.0, + 1750.0, + 1033.0, + 1782.0, + 325.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1801.0, + 1405.0, + 1801.0, + 1405.0, + 1837.0, + 295.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1832.0, + 1123.0, + 1832.0, + 1123.0, + 1868.0, + 325.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 260.0, + 1403.0, + 260.0, + 1403.0, + 297.0, + 323.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 292.0, + 611.0, + 292.0, + 611.0, + 321.0, + 323.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1099.0, + 1403.0, + 1099.0, + 1403.0, + 1133.0, + 295.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1129.0, + 1405.0, + 1129.0, + 1405.0, + 1163.0, + 322.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1154.0, + 398.0, + 1154.0, + 398.0, + 1195.0, + 320.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1884.0, + 1405.0, + 1884.0, + 1405.0, + 1920.0, + 296.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1911.0, + 1404.0, + 1911.0, + 1404.0, + 1953.0, + 319.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1945.0, + 1405.0, + 1945.0, + 1405.0, + 1984.0, + 319.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1978.0, + 468.0, + 1978.0, + 468.0, + 2008.0, + 323.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 621.0, + 1403.0, + 621.0, + 1403.0, + 657.0, + 296.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 653.0, + 1187.0, + 653.0, + 1187.0, + 687.0, + 321.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 506.0, + 1403.0, + 506.0, + 1403.0, + 547.0, + 294.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 539.0, + 1404.0, + 539.0, + 1404.0, + 576.0, + 322.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 571.0, + 1043.0, + 571.0, + 1043.0, + 605.0, + 324.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 425.0, + 1404.0, + 425.0, + 1404.0, + 462.0, + 296.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 457.0, + 1153.0, + 457.0, + 1153.0, + 493.0, + 325.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 820.0, + 1402.0, + 820.0, + 1402.0, + 854.0, + 297.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 850.0, + 1403.0, + 850.0, + 1403.0, + 884.0, + 323.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 881.0, + 613.0, + 881.0, + 613.0, + 915.0, + 323.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 339.0, + 1404.0, + 339.0, + 1404.0, + 381.0, + 294.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 374.0, + 1097.0, + 374.0, + 1097.0, + 408.0, + 322.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 704.0, + 1403.0, + 704.0, + 1403.0, + 741.0, + 296.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 736.0, + 1405.0, + 736.0, + 1405.0, + 773.0, + 321.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 768.0, + 596.0, + 768.0, + 596.0, + 797.0, + 322.0, + 797.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 322, + 319, + 1390, + 319, + 1390, + 1441, + 322, + 1441 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 296, + 1462, + 1405, + 1462, + 1405, + 1770, + 296, + 1770 + ], + "score": 0.963 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.847 + }, + { + "category_id": 0, + "poly": [ + 301, + 228, + 603, + 228, + 603, + 261, + 301, + 261 + ], + "score": 0.518 + }, + { + "category_id": 1, + "poly": [ + 301, + 228, + 603, + 228, + 603, + 261, + 301, + 261 + ], + "score": 0.162 + }, + { + "category_id": 15, + "poly": [ + 730.0, + 312.0, + 1027.0, + 312.0, + 1027.0, + 344.0, + 730.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 340.0, + 1026.0, + 340.0, + 1026.0, + 373.0, + 738.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 387.0, + 375.0, + 387.0, + 375.0, + 423.0, + 312.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 390.0, + 550.0, + 390.0, + 550.0, + 423.0, + 449.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 392.0, + 860.0, + 392.0, + 860.0, + 426.0, + 666.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 390.0, + 1088.0, + 390.0, + 1088.0, + 423.0, + 928.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 387.0, + 1310.0, + 387.0, + 1310.0, + 426.0, + 1233.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 459.0, + 362.0, + 459.0, + 362.0, + 532.0, + 329.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 497.0, + 1364.0, + 497.0, + 1364.0, + 513.0, + 1351.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 513.0, + 1227.0, + 513.0, + 1227.0, + 527.0, + 1217.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 508.0, + 1366.0, + 508.0, + 1366.0, + 532.0, + 1347.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 605.0, + 362.0, + 605.0, + 362.0, + 671.0, + 330.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 641.0, + 553.0, + 641.0, + 553.0, + 671.0, + 535.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 657.0, + 1229.0, + 657.0, + 1229.0, + 675.0, + 1215.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 763.0, + 361.0, + 763.0, + 361.0, + 813.0, + 332.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 790.0, + 1362.0, + 790.0, + 1362.0, + 800.0, + 1353.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 881.0, + 363.0, + 881.0, + 363.0, + 974.0, + 329.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 951.0, + 1227.0, + 951.0, + 1227.0, + 961.0, + 1217.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1024.0, + 363.0, + 1024.0, + 363.0, + 1123.0, + 330.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1198.0, + 361.0, + 1198.0, + 361.0, + 1243.0, + 329.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1334.0, + 361.0, + 1334.0, + 361.0, + 1398.0, + 331.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1463.0, + 1403.0, + 1463.0, + 1403.0, + 1495.0, + 295.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1492.0, + 1405.0, + 1492.0, + 1405.0, + 1522.0, + 294.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1519.0, + 1403.0, + 1519.0, + 1403.0, + 1551.0, + 295.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1547.0, + 1405.0, + 1547.0, + 1405.0, + 1579.0, + 294.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1573.0, + 1403.0, + 1573.0, + 1403.0, + 1607.0, + 294.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1601.0, + 1406.0, + 1601.0, + 1406.0, + 1634.0, + 293.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1662.0, + 295.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1657.0, + 1407.0, + 1657.0, + 1407.0, + 1692.0, + 296.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1685.0, + 1406.0, + 1685.0, + 1406.0, + 1717.0, + 295.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1403.0, + 1712.0, + 1403.0, + 1747.0, + 294.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 458.0, + 1740.0, + 458.0, + 1773.0, + 293.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 220.0, + 607.0, + 220.0, + 607.0, + 271.0, + 293.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 220.0, + 607.0, + 220.0, + 607.0, + 271.0, + 293.0, + 271.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 632, + 1382, + 632, + 1382, + 1273, + 301, + 1273 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 295, + 1295, + 1405, + 1295, + 1405, + 1631, + 295, + 1631 + ], + "score": 0.953 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.841 + }, + { + "category_id": 13, + "poly": [ + 298, + 1354, + 372, + 1354, + 372, + 1381, + 298, + 1381 + ], + "score": 0.87, + "latex": "( x _ { s e n t } )" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 673.0, + 688.0, + 673.0, + 688.0, + 700.0, + 493.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 632.0, + 1215.0, + 632.0, + 1215.0, + 696.0, + 848.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 687.0, + 488.0, + 687.0, + 488.0, + 717.0, + 438.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 695.0, + 644.0, + 695.0, + 644.0, + 726.0, + 537.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 699.0, + 744.0, + 699.0, + 744.0, + 708.0, + 732.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 688.0, + 843.0, + 688.0, + 843.0, + 725.0, + 763.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 755.0, + 322.0, + 755.0, + 322.0, + 770.0, + 313.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 742.0, + 518.0, + 742.0, + 518.0, + 777.0, + 488.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 741.0, + 1350.0, + 741.0, + 1350.0, + 777.0, + 1264.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 770.0, + 1111.0, + 770.0, + 1111.0, + 835.0, + 849.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 784.0, + 1217.0, + 784.0, + 1217.0, + 832.0, + 1130.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 820.0, + 327.0, + 820.0, + 327.0, + 844.0, + 309.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 822.0, + 606.0, + 822.0, + 606.0, + 854.0, + 594.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 823.0, + 1356.0, + 823.0, + 1356.0, + 858.0, + 1252.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 853.0, + 691.0, + 853.0, + 691.0, + 878.0, + 674.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 859.0, + 608.0, + 859.0, + 608.0, + 870.0, + 599.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 888.0, + 1213.0, + 888.0, + 1213.0, + 957.0, + 839.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 943.0, + 329.0, + 943.0, + 329.0, + 972.0, + 305.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 941.0, + 1356.0, + 941.0, + 1356.0, + 976.0, + 1252.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 979.0, + 535.0, + 979.0, + 535.0, + 989.0, + 522.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1057.0, + 828.0, + 1057.0, + 828.0, + 1072.0, + 813.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1013.0, + 1211.0, + 1013.0, + 1211.0, + 1080.0, + 846.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1063.0, + 329.0, + 1063.0, + 329.0, + 1094.0, + 306.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1072.0, + 1356.0, + 1072.0, + 1356.0, + 1104.0, + 1254.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1100.0, + 511.0, + 1100.0, + 511.0, + 1116.0, + 496.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1134.0, + 1214.0, + 1134.0, + 1214.0, + 1204.0, + 849.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1191.0, + 327.0, + 1191.0, + 327.0, + 1215.0, + 308.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1187.0, + 1382.0, + 1187.0, + 1382.0, + 1220.0, + 1222.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1223.0, + 485.0, + 1223.0, + 485.0, + 1236.0, + 469.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1235.0, + 597.0, + 1235.0, + 597.0, + 1266.0, + 575.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1240.0, + 817.0, + 1240.0, + 817.0, + 1266.0, + 800.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.75, + 745.0, + 745.75, + 745.0, + 745.75, + 775.0, + 704.75, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 746.5, + 1056.0, + 746.5, + 1056.0, + 777.0, + 1006.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 787.0, + 1148.0, + 787.0, + 1148.0, + 836.0, + 1087.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 858.5, + 465.0, + 858.5, + 465.0, + 868.5, + 451.0, + 868.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1297.0, + 1405.0, + 1297.0, + 1405.0, + 1328.0, + 295.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1321.0, + 1406.0, + 1321.0, + 1406.0, + 1361.0, + 292.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1349.0, + 297.0, + 1349.0, + 297.0, + 1389.0, + 293.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1349.0, + 1406.0, + 1349.0, + 1406.0, + 1389.0, + 373.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1379.0, + 1405.0, + 1379.0, + 1405.0, + 1413.0, + 293.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1407.0, + 1407.0, + 1407.0, + 1407.0, + 1442.0, + 294.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1435.0, + 1406.0, + 1435.0, + 1406.0, + 1469.0, + 294.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1462.0, + 1402.0, + 1462.0, + 1402.0, + 1493.0, + 294.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1491.0, + 1406.0, + 1491.0, + 1406.0, + 1522.0, + 295.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1520.0, + 1406.0, + 1520.0, + 1406.0, + 1551.0, + 294.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1577.0, + 295.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1574.0, + 1405.0, + 1574.0, + 1405.0, + 1605.0, + 295.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1603.0, + 719.0, + 1603.0, + 719.0, + 1634.0, + 295.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1589, + 1404, + 1589, + 1404, + 1896, + 297, + 1896 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 675, + 1404, + 675, + 1404, + 892, + 298, + 892 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 1266, + 1404, + 1266, + 1404, + 1390, + 297, + 1390 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 906, + 1403, + 906, + 1403, + 1030, + 299, + 1030 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 1481, + 1403, + 1481, + 1403, + 1576, + 298, + 1576 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 360, + 221, + 1343, + 221, + 1343, + 412, + 360, + 412 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 297, + 1404, + 1398, + 1404, + 1398, + 1469, + 297, + 1469 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 1142, + 1400, + 1142, + 1400, + 1204, + 298, + 1204 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 300, + 1218, + 1253, + 1218, + 1253, + 1252, + 300, + 1252 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 298, + 538, + 910, + 538, + 910, + 574, + 298, + 574 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.911 + }, + { + "category_id": 4, + "poly": [ + 619, + 442, + 1083, + 442, + 1083, + 473, + 619, + 473 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 300, + 1080, + 599, + 1080, + 599, + 1111, + 300, + 1111 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 305, + 615, + 902, + 615, + 902, + 645, + 305, + 645 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 13, + "poly": [ + 606, + 1543, + 695, + 1543, + 695, + 1577, + 606, + 1577 + ], + "score": 0.92, + "latex": "\\beta ( x _ { \\mathrm { s e n t } } )" + }, + { + "category_id": 13, + "poly": [ + 466, + 1543, + 554, + 1543, + 554, + 1577, + 466, + 1577 + ], + "score": 0.92, + "latex": "\\gamma ( x _ { \\mathrm { s e n t } } )" + }, + { + "category_id": 13, + "poly": [ + 1080, + 1142, + 1246, + 1142, + 1246, + 1176, + 1080, + 1176 + ], + "score": 0.92, + "latex": "x _ { \\mathrm { S } } = \\sec ( x _ { I } )" + }, + { + "category_id": 13, + "poly": [ + 723, + 1436, + 943, + 1436, + 943, + 1469, + 723, + 1469 + ], + "score": 0.92, + "latex": "x _ { \\mathrm { S } } = f _ { \\mathrm { G A } } ( x _ { I } , x _ { \\mathrm { s e n t } } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1543, + 387, + 1543, + 387, + 1577, + 298, + 1577 + ], + "score": 0.91, + "latex": "\\beta ( x _ { \\mathrm { s e n t } } )" + }, + { + "category_id": 13, + "poly": [ + 698, + 678, + 870, + 678, + 870, + 707, + 698, + 707 + ], + "score": 0.9, + "latex": "3 \\times 1 6 8 \\times 3 0 0" + }, + { + "category_id": 13, + "poly": [ + 1043, + 1219, + 1245, + 1219, + 1245, + 1254, + 1043, + 1254 + ], + "score": 0.9, + "latex": "x _ { \\mathrm { S } } = [ x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } ]" + }, + { + "category_id": 13, + "poly": [ + 1142, + 1513, + 1404, + 1513, + 1404, + 1547, + 1142, + 1547 + ], + "score": 0.9, + "latex": "x _ { \\mathrm { S } } = \\gamma ( x _ { \\mathrm { s e n t } } ) \\odot x _ { I } +" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1265, + 1331, + 1265, + 1331, + 1300, + 1028, + 1300 + ], + "score": 0.89, + "latex": "x _ { \\mathrm { S } } = [ \\mathrm { v e c } ( x _ { I } ) , x _ { \\mathrm { B o W } } , x _ { \\mathrm { s e n t } } ]" + }, + { + "category_id": 13, + "poly": [ + 654, + 1978, + 714, + 1978, + 714, + 2004, + 654, + 2004 + ], + "score": 0.88, + "latex": "x _ { A n s }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1972, + 552, + 1972, + 552, + 2006, + 297, + 2006 + ], + "score": 0.87, + "latex": "x _ { \\mathrm { A n s } } = \\sigma ( \\bar { f } ( x _ { \\mathrm { S } } ; \\theta _ { L i n } ) )" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1913, + 1136, + 1913, + 1136, + 1944, + 1115, + 1944 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 635, + 970, + 672, + 970, + 672, + 1001, + 635, + 1001 + ], + "score": 0.81, + "latex": "( \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 667, + 1944, + 691, + 1944, + 691, + 1970, + 667, + 1970 + ], + "score": 0.81, + "latex": "V" + }, + { + "category_id": 13, + "poly": [ + 542, + 1948, + 562, + 1948, + 562, + 1970, + 542, + 1970 + ], + "score": 0.73, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 297, + 830, + 323, + 830, + 323, + 858, + 297, + 858 + ], + "score": 0.73, + "latex": "V" + }, + { + "category_id": 13, + "poly": [ + 423, + 739, + 507, + 739, + 507, + 768, + 423, + 768 + ], + "score": 0.69, + "latex": "6 4 ~ 4 \\mathrm { x } 4" + }, + { + "category_id": 13, + "poly": [ + 570, + 1976, + 602, + 1976, + 602, + 2004, + 570, + 2004 + ], + "score": 0.68, + "latex": "x _ { \\mathrm { { S } } }" + }, + { + "category_id": 13, + "poly": [ + 1001, + 709, + 1087, + 709, + 1087, + 737, + 1001, + 737 + ], + "score": 0.67, + "latex": "3 2 ~ 8 \\mathrm { x } 8" + }, + { + "category_id": 13, + "poly": [ + 994, + 740, + 1021, + 740, + 1021, + 767, + 994, + 767 + ], + "score": 0.51, + "latex": "V" + }, + { + "category_id": 13, + "poly": [ + 1253, + 969, + 1329, + 969, + 1329, + 998, + 1253, + 998 + ], + "score": 0.39, + "latex": "2 . 5 \\mathrm { e } { \\cdot } 4" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 230.0, + 445.0, + 230.0, + 445.0, + 256.0, + 361.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 227.0, + 491.0, + 227.0, + 491.0, + 256.0, + 465.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 227.0, + 949.0, + 227.0, + 949.0, + 266.0, + 801.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 235.0, + 1326.0, + 235.0, + 1326.0, + 260.0, + 1219.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 251.0, + 1289.0, + 251.0, + 1289.0, + 277.0, + 1256.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 267.0, + 429.0, + 267.0, + 429.0, + 313.0, + 358.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 271.0, + 497.0, + 271.0, + 497.0, + 312.0, + 455.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 264.0, + 938.0, + 264.0, + 938.0, + 306.0, + 901.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 272.0, + 1189.0, + 272.0, + 1189.0, + 311.0, + 1104.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 270.0, + 1337.0, + 270.0, + 1337.0, + 311.0, + 1217.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 307.0, + 665.0, + 307.0, + 665.0, + 323.0, + 651.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 297.0, + 756.0, + 297.0, + 756.0, + 336.0, + 671.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 297.0, + 958.0, + 297.0, + 958.0, + 337.0, + 890.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 323.0, + 450.0, + 323.0, + 450.0, + 362.0, + 360.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 317.0, + 636.0, + 317.0, + 636.0, + 361.0, + 520.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 335.0, + 844.0, + 335.0, + 844.0, + 361.0, + 800.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 337.0, + 933.0, + 337.0, + 933.0, + 367.0, + 906.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 335.0, + 1044.0, + 335.0, + 1044.0, + 361.0, + 1000.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 324.0, + 1187.0, + 324.0, + 1187.0, + 362.0, + 1106.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 322.0, + 1321.0, + 322.0, + 1321.0, + 366.0, + 1233.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 362.0, + 1326.0, + 362.0, + 1326.0, + 384.0, + 1222.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 371.0, + 442.0, + 371.0, + 442.0, + 414.0, + 355.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 374.0, + 491.0, + 374.0, + 491.0, + 405.0, + 465.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 376.0, + 1290.0, + 376.0, + 1290.0, + 403.0, + 1258.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.25, + 300.0, + 619.25, + 300.0, + 619.25, + 334.5, + 579.25, + 334.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.25, + 300.5, + 838.25, + 300.5, + 838.25, + 336.5, + 797.25, + 336.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.25, + 302.0, + 1039.25, + 302.0, + 1039.25, + 336.0, + 999.25, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 327.5, + 501.0, + 327.5, + 501.0, + 364.0, + 446.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 914.0, + 535.0, + 914.0, + 579.0, + 292.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 437.0, + 1084.0, + 437.0, + 1084.0, + 478.0, + 613.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1077.0, + 603.0, + 1077.0, + 603.0, + 1115.0, + 294.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 613.0, + 909.0, + 613.0, + 909.0, + 650.0, + 296.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2086.0, + 868.0, + 2086.0, + 868.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1587.0, + 1406.0, + 1587.0, + 1406.0, + 1624.0, + 292.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1620.0, + 1406.0, + 1620.0, + 1406.0, + 1656.0, + 295.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1647.0, + 1408.0, + 1647.0, + 1408.0, + 1690.0, + 291.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1676.0, + 1408.0, + 1676.0, + 1408.0, + 1721.0, + 292.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1711.0, + 1405.0, + 1711.0, + 1405.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1741.0, + 1407.0, + 1741.0, + 1407.0, + 1781.0, + 292.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1773.0, + 1406.0, + 1773.0, + 1406.0, + 1809.0, + 295.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1805.0, + 1405.0, + 1805.0, + 1405.0, + 1837.0, + 296.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1833.0, + 1405.0, + 1833.0, + 1405.0, + 1871.0, + 291.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1863.0, + 1108.0, + 1863.0, + 1108.0, + 1899.0, + 293.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 677.0, + 697.0, + 677.0, + 697.0, + 712.0, + 296.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 677.0, + 1404.0, + 677.0, + 1404.0, + 712.0, + 871.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 709.0, + 1000.0, + 709.0, + 1000.0, + 741.0, + 293.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 709.0, + 1406.0, + 709.0, + 1406.0, + 741.0, + 1088.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 739.0, + 422.0, + 739.0, + 422.0, + 772.0, + 293.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 739.0, + 993.0, + 739.0, + 993.0, + 772.0, + 508.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 739.0, + 1406.0, + 739.0, + 1406.0, + 772.0, + 1022.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 767.0, + 1405.0, + 767.0, + 1405.0, + 805.0, + 293.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 800.0, + 1405.0, + 800.0, + 1405.0, + 834.0, + 294.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 829.0, + 1405.0, + 829.0, + 1405.0, + 864.0, + 324.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 859.0, + 733.0, + 859.0, + 733.0, + 895.0, + 293.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 1114.0, + 1909.0, + 1114.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1946.0, + 1137.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 541.0, + 1940.0, + 541.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 1940.0, + 666.0, + 1940.0, + 666.0, + 1978.0, + 563.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 1940.0, + 1404.0, + 1940.0, + 1404.0, + 1978.0, + 692.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 296.0, + 1973.0, + 296.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1973.0, + 569.0, + 1973.0, + 569.0, + 2009.0, + 553.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1973.0, + 653.0, + 1973.0, + 653.0, + 2009.0, + 603.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2009.0, + 715.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1082.0, + 2002.0, + 1082.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1266.0, + 1027.0, + 1266.0, + 1027.0, + 1302.0, + 295.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 1266.0, + 1405.0, + 1266.0, + 1405.0, + 1302.0, + 1332.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1297.0, + 1406.0, + 1297.0, + 1406.0, + 1331.0, + 292.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1328.0, + 1406.0, + 1328.0, + 1406.0, + 1361.0, + 294.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 1125.0, + 1358.0, + 1125.0, + 1391.0, + 295.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 906.0, + 1407.0, + 906.0, + 1407.0, + 942.0, + 294.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 936.0, + 1407.0, + 936.0, + 1407.0, + 973.0, + 294.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 968.0, + 634.0, + 968.0, + 634.0, + 1004.0, + 293.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 968.0, + 1252.0, + 968.0, + 1252.0, + 1004.0, + 673.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 968.0, + 1405.0, + 968.0, + 1405.0, + 1004.0, + 1330.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1001.0, + 445.0, + 1001.0, + 445.0, + 1032.0, + 293.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1479.0, + 1405.0, + 1479.0, + 1405.0, + 1519.0, + 292.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1509.0, + 1141.0, + 1509.0, + 1141.0, + 1550.0, + 293.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1544.0, + 465.0, + 1544.0, + 465.0, + 1578.0, + 388.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1544.0, + 605.0, + 1544.0, + 605.0, + 1578.0, + 555.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1544.0, + 1320.0, + 1544.0, + 1320.0, + 1578.0, + 696.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1402.0, + 1402.0, + 1402.0, + 1402.0, + 1440.0, + 295.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1433.0, + 722.0, + 1433.0, + 722.0, + 1473.0, + 293.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1433.0, + 955.0, + 1433.0, + 955.0, + 1473.0, + 944.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1140.0, + 1079.0, + 1140.0, + 1079.0, + 1179.0, + 294.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1140.0, + 1402.0, + 1140.0, + 1402.0, + 1179.0, + 1247.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1172.0, + 657.0, + 1172.0, + 657.0, + 1206.0, + 296.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1211.0, + 1042.0, + 1211.0, + 1042.0, + 1260.0, + 292.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 1211.0, + 1257.0, + 1211.0, + 1257.0, + 1260.0, + 1246.0, + 1260.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1097, + 1405, + 1097, + 1405, + 1190, + 298, + 1190 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 406, + 612, + 1293, + 612, + 1293, + 894, + 406, + 894 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 521, + 1299, + 1167, + 1299, + 1167, + 1600, + 521, + 1600 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 407, + 220, + 1294, + 220, + 1294, + 503, + 407, + 503 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 298, + 523, + 1404, + 523, + 1404, + 580, + 298, + 580 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 296, + 1206, + 1404, + 1206, + 1404, + 1269, + 296, + 1269 + ], + "score": 0.948 + }, + { + "category_id": 0, + "poly": [ + 301, + 1027, + 780, + 1027, + 780, + 1064, + 301, + 1064 + ], + "score": 0.914 + }, + { + "category_id": 4, + "poly": [ + 435, + 1630, + 1263, + 1630, + 1263, + 1662, + 435, + 1662 + ], + "score": 0.911 + }, + { + "category_id": 4, + "poly": [ + 294, + 915, + 1402, + 915, + 1402, + 974, + 294, + 974 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.851 + }, + { + "category_id": 15, + "poly": [ + 619.0, + 612.0, + 672.0, + 612.0, + 672.0, + 644.0, + 619.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 612.0, + 1129.0, + 612.0, + 1129.0, + 642.0, + 1051.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 625.0, + 463.0, + 625.0, + 463.0, + 655.0, + 424.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 625.0, + 907.0, + 625.0, + 907.0, + 655.0, + 869.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 662.0, + 461.0, + 662.0, + 461.0, + 690.0, + 426.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 662.0, + 906.0, + 662.0, + 906.0, + 689.0, + 870.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 684.0, + 462.0, + 684.0, + 462.0, + 775.0, + 406.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 683.0, + 905.0, + 683.0, + 905.0, + 774.0, + 852.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 734.0, + 461.0, + 734.0, + 461.0, + 761.0, + 427.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 740.0, + 477.0, + 740.0, + 477.0, + 773.0, + 462.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 734.0, + 905.0, + 734.0, + 905.0, + 760.0, + 871.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 771.0, + 461.0, + 771.0, + 461.0, + 798.0, + 426.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 771.0, + 906.0, + 771.0, + 906.0, + 797.0, + 870.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 816.0, + 553.0, + 816.0, + 553.0, + 846.0, + 515.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 818.0, + 626.0, + 818.0, + 626.0, + 845.0, + 590.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 818.0, + 700.0, + 818.0, + 700.0, + 845.0, + 663.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 818.0, + 774.0, + 818.0, + 774.0, + 845.0, + 738.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 818.0, + 848.0, + 818.0, + 848.0, + 845.0, + 812.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 818.0, + 997.0, + 818.0, + 997.0, + 845.0, + 961.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 818.0, + 1072.0, + 818.0, + 1072.0, + 845.0, + 1035.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 816.0, + 1147.0, + 816.0, + 1147.0, + 846.0, + 1108.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 818.0, + 1220.0, + 818.0, + 1220.0, + 845.0, + 1184.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 811.0, + 1300.0, + 811.0, + 1300.0, + 865.0, + 1235.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 837.0, + 705.0, + 837.0, + 705.0, + 871.0, + 587.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 833.0, + 837.0, + 833.0, + 837.0, + 858.0, + 797.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 837.0, + 1149.0, + 837.0, + 1149.0, + 871.0, + 1031.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 867.0, + 772.0, + 867.0, + 772.0, + 896.0, + 735.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 869.0, + 861.0, + 869.0, + 861.0, + 894.0, + 814.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 870.0, + 1015.0, + 870.0, + 1015.0, + 893.0, + 900.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 665.0, + 517.0, + 665.0, + 517.0, + 675.0, + 489.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 664.5, + 1038.0, + 664.5, + 1038.0, + 708.5, + 891.0, + 708.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.75, + 805.5, + 466.75, + 805.5, + 466.75, + 843.5, + 436.75, + 843.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 806.5, + 911.0, + 806.5, + 911.0, + 843.0, + 881.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1317.0, + 883.0, + 1317.0, + 883.0, + 1417.0, + 706.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1341.0, + 1184.0, + 1341.0, + 1184.0, + 1473.0, + 897.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1406.0, + 670.0, + 1406.0, + 670.0, + 1413.0, + 660.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1418.0, + 672.0, + 1418.0, + 672.0, + 1431.0, + 659.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1454.0, + 738.0, + 1454.0, + 738.0, + 1485.0, + 655.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 1451.0, + 1082.0, + 1451.0, + 1082.0, + 1482.0, + 1015.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1538.0, + 557.0, + 1538.0, + 557.0, + 1572.0, + 531.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1538.0, + 618.0, + 1538.0, + 618.0, + 1570.0, + 592.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1539.0, + 680.0, + 1539.0, + 680.0, + 1571.0, + 652.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1538.0, + 815.0, + 1538.0, + 815.0, + 1567.0, + 796.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1538.0, + 876.0, + 1538.0, + 876.0, + 1568.0, + 858.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1537.0, + 939.0, + 1537.0, + 939.0, + 1569.0, + 918.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1539.0, + 1104.0, + 1539.0, + 1104.0, + 1569.0, + 1063.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1530.0, + 1166.0, + 1530.0, + 1166.0, + 1572.0, + 1116.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1572.0, + 656.0, + 1572.0, + 656.0, + 1601.0, + 559.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1568.0, + 916.0, + 1568.0, + 916.0, + 1601.0, + 818.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 1573.0, + 1151.0, + 1573.0, + 1151.0, + 1598.0, + 1076.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 235.0, + 461.0, + 235.0, + 461.0, + 263.0, + 426.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 220.0, + 681.0, + 220.0, + 681.0, + 254.0, + 618.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 235.0, + 906.0, + 235.0, + 906.0, + 262.0, + 869.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 223.0, + 1127.0, + 223.0, + 1127.0, + 253.0, + 1058.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 271.0, + 461.0, + 271.0, + 461.0, + 299.0, + 426.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 272.0, + 905.0, + 272.0, + 905.0, + 298.0, + 870.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 295.0, + 461.0, + 295.0, + 461.0, + 383.0, + 408.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 294.0, + 905.0, + 294.0, + 905.0, + 384.0, + 851.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 342.0, + 460.0, + 342.0, + 460.0, + 369.0, + 426.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 343.0, + 905.0, + 343.0, + 905.0, + 369.0, + 870.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 378.0, + 461.0, + 378.0, + 461.0, + 405.0, + 426.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 379.0, + 905.0, + 379.0, + 905.0, + 405.0, + 870.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 427.0, + 553.0, + 427.0, + 553.0, + 453.0, + 518.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 426.0, + 629.0, + 426.0, + 629.0, + 452.0, + 592.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 426.0, + 706.0, + 426.0, + 706.0, + 452.0, + 670.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 427.0, + 782.0, + 427.0, + 782.0, + 452.0, + 745.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 426.0, + 856.0, + 426.0, + 856.0, + 452.0, + 820.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 415.0, + 912.0, + 415.0, + 912.0, + 449.0, + 869.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 431.0, + 987.0, + 431.0, + 987.0, + 449.0, + 972.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 428.0, + 1065.0, + 428.0, + 1065.0, + 451.0, + 1045.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 428.0, + 1142.0, + 428.0, + 1142.0, + 451.0, + 1122.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 430.0, + 1215.0, + 430.0, + 1215.0, + 448.0, + 1199.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 422.0, + 1299.0, + 422.0, + 1299.0, + 471.0, + 1245.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 446.0, + 708.0, + 446.0, + 708.0, + 478.0, + 591.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 441.0, + 845.0, + 441.0, + 845.0, + 466.0, + 806.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 446.0, + 1152.0, + 446.0, + 1152.0, + 477.0, + 1035.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 475.0, + 1089.0, + 475.0, + 1089.0, + 504.0, + 477.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 479.0, + 1245.0, + 479.0, + 1245.0, + 501.0, + 1136.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 414.5, + 466.0, + 414.5, + 466.0, + 451.5, + 436.0, + 451.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 522.0, + 1405.0, + 522.0, + 1405.0, + 556.0, + 294.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 549.0, + 502.0, + 549.0, + 502.0, + 582.0, + 293.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1024.0, + 784.0, + 1024.0, + 784.0, + 1070.0, + 294.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1627.0, + 1266.0, + 1627.0, + 1266.0, + 1666.0, + 432.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 913.0, + 1405.0, + 913.0, + 1405.0, + 947.0, + 294.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 941.0, + 1062.0, + 941.0, + 1062.0, + 978.0, + 294.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1096.0, + 1405.0, + 1096.0, + 1405.0, + 1134.0, + 293.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1128.0, + 1403.0, + 1128.0, + 1403.0, + 1164.0, + 293.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1158.0, + 1356.0, + 1158.0, + 1356.0, + 1195.0, + 295.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1203.0, + 1405.0, + 1203.0, + 1405.0, + 1241.0, + 294.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1238.0, + 1010.0, + 1238.0, + 1010.0, + 1268.0, + 295.0, + 1268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 326, + 778, + 1376, + 778, + 1376, + 1588, + 326, + 1588 + ], + "score": 0.974, + "html": "
SGN Instruction Type42 Train Instructions: Not containing‘red'&‘pillar'28 Test Instructions: Containing‘red'or‘pillar'
Go to the (object).Go to the color>object.torch,keycard,skullkey,armor yellow, green, bluepillar red
Go to the size)object. Go to the {color) {object).tall, short blue torch, green torch, green armor,red torch,red skullkey,red pillar,
blue skullkey,blue keycard, yellow keycard, yellow skullkeygreen pillar,red keycard,red armor
Go to the {size) (object). Go to the color) (size)object.short torch,tall torch green tall, blue tall, blue short,tall pillar, short pillar red short, red tall
Go to the size){color)object.green short tall green, tall blue, short blue,short red, tall red
Go to the {color)(size)(object).short green green tall torch, green short torch,red short pillar,red short torch,
blue short torch,blue tall torchred tall pillar, green tall pillar, red tall torch, green short pillar
Go to the size) {color> {object).tall green torch,short green torch, short blue torch,tall blue torchshort red pillar, short red torch, tall red pillar, tall green pillar, tall red torch,short green pillar
Go to the superlative)object. Go to the {superlative> (color) object.largest, smallest smallest yellow, smallest blue, smallest green,largest blue,largest red,smallest red
EQAQuestion Typelargest green,largest yellow 21 Train Questions: Not containing‘blue'&‘torch'8 Test Questions: Containing‘blue'or‘torch'
Which {size)object is {color)in color?What color is the {object)? What color is the {size){object>? Which object is {color) in color?pillar, keycard, skullkey,armor short pillar, tall pillar red,yellow, greentorch short torch, tall torch blue
" + }, + { + "category_id": 6, + "poly": [ + 299, + 651, + 1401, + 651, + 1401, + 736, + 299, + 736 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.515 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.424 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 651.0, + 1403.0, + 651.0, + 1403.0, + 683.0, + 295.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 676.0, + 1405.0, + 676.0, + 1405.0, + 713.0, + 293.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 704.0, + 388.0, + 704.0, + 388.0, + 742.0, + 292.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 341, + 1528, + 1358, + 1528, + 1358, + 1900, + 341, + 1900 + ], + "score": 0.983, + "html": "
SGNInstruction Type22 Train Instructions: Not containing‘red'&‘bed'11 Test Instructions: Containing‘red'or‘bed'
Go to the (object). Go to the {color)(object).refrigerator,office_chair,fish_tank,fireplace green refrigerator, green office_chair, green fish_tank,green fireplace, green sofa, blue refrigerator, blue office_chair, blue fish_tank,blue fireplace,blue sofa, yellow refrigerator,yellow office_chair, yellow fish_tank,yellow fireplace,yellow sofabed red bed, green bed,blue bed, yellow bed,red refrigerator, red office_chair,red fish_tank, red fireplace,red sofa
EQAGo to the {color) object. Question Typegreen,blue,yellow 7 Train Questions:red 2 Test Questions:
Not containing‘blue'&‘sofa'Containing‘blue'or‘sofa'
What color is the {object)? What object is color> in color?refrigerator,office_chair,fish_tank,fireplace,bed red, green,yellowsofa blue
" + }, + { + "category_id": 1, + "poly": [ + 299, + 465, + 823, + 465, + 823, + 921, + 299, + 921 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 845, + 584, + 1319, + 584, + 1319, + 870, + 845, + 870 + ], + "score": 0.977, + "html": "
SGNEQA
ModelTrainTestTrainTest
Text only0.630.330.220.23
Image only0.280.010.120.22
Concat0.650.130.310.13
GA0.980.200.920.03
FiLM0.990.370.920.24
PACMAN0.730.200.400.21
Dual-Attention0.990.470.890.29
" + }, + { + "category_id": 1, + "poly": [ + 297, + 296, + 1404, + 296, + 1404, + 449, + 297, + 449 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 316, + 1091, + 1387, + 1091, + 1387, + 1333, + 316, + 1333 + ], + "score": 0.972 + }, + { + "category_id": 6, + "poly": [ + 845, + 502, + 1403, + 502, + 1403, + 558, + 845, + 558 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 298, + 969, + 1407, + 969, + 1407, + 1061, + 298, + 1061 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 297, + 923, + 981, + 923, + 981, + 953, + 297, + 953 + ], + "score": 0.9 + }, + { + "category_id": 6, + "poly": [ + 299, + 1431, + 1398, + 1431, + 1398, + 1490, + 299, + 1490 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 301, + 226, + 717, + 226, + 717, + 262, + 301, + 262 + ], + "score": 0.895 + }, + { + "category_id": 4, + "poly": [ + 312, + 1353, + 1385, + 1353, + 1385, + 1385, + 312, + 1385 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.861 + }, + { + "category_id": 13, + "poly": [ + 702, + 558, + 737, + 558, + 737, + 585, + 702, + 585 + ], + "score": 0.82, + "latex": "+ 1" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1297.0, + 662.0, + 1297.0, + 662.0, + 1335.0, + 321.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1299.0, + 966.0, + 1299.0, + 966.0, + 1334.0, + 733.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1300.0, + 1376.0, + 1300.0, + 1376.0, + 1333.0, + 1043.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 502.0, + 1403.0, + 502.0, + 1403.0, + 532.0, + 847.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 528.0, + 1350.0, + 528.0, + 1350.0, + 561.0, + 845.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1428.0, + 1402.0, + 1428.0, + 1402.0, + 1465.0, + 294.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1459.0, + 1097.0, + 1459.0, + 1097.0, + 1490.0, + 296.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 722.0, + 224.0, + 722.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1352.0, + 1386.0, + 1352.0, + 1386.0, + 1387.0, + 313.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 464.0, + 825.0, + 464.0, + 825.0, + 498.0, + 295.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 495.0, + 826.0, + 495.0, + 826.0, + 529.0, + 294.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 525.0, + 826.0, + 525.0, + 826.0, + 560.0, + 295.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 555.0, + 701.0, + 555.0, + 701.0, + 590.0, + 295.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 555.0, + 826.0, + 555.0, + 826.0, + 590.0, + 738.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 586.0, + 825.0, + 586.0, + 825.0, + 619.0, + 295.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 617.0, + 826.0, + 617.0, + 826.0, + 652.0, + 295.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 648.0, + 826.0, + 648.0, + 826.0, + 679.0, + 295.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 677.0, + 825.0, + 677.0, + 825.0, + 711.0, + 294.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 710.0, + 826.0, + 710.0, + 826.0, + 742.0, + 294.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 739.0, + 827.0, + 739.0, + 827.0, + 773.0, + 293.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 770.0, + 826.0, + 770.0, + 826.0, + 804.0, + 295.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 799.0, + 825.0, + 799.0, + 825.0, + 833.0, + 296.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 831.0, + 826.0, + 831.0, + 826.0, + 863.0, + 296.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 861.0, + 827.0, + 861.0, + 827.0, + 895.0, + 296.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 890.0, + 825.0, + 890.0, + 825.0, + 927.0, + 295.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 296.0, + 1405.0, + 296.0, + 1405.0, + 329.0, + 295.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 325.0, + 1402.0, + 325.0, + 1402.0, + 361.0, + 294.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 354.0, + 1406.0, + 354.0, + 1406.0, + 394.0, + 295.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 388.0, + 1405.0, + 388.0, + 1405.0, + 421.0, + 295.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 414.0, + 393.0, + 414.0, + 393.0, + 453.0, + 293.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 965.0, + 1405.0, + 965.0, + 1405.0, + 1004.0, + 292.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 998.0, + 1405.0, + 998.0, + 1405.0, + 1033.0, + 293.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1028.0, + 390.0, + 1028.0, + 390.0, + 1068.0, + 291.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 918.0, + 984.0, + 918.0, + 984.0, + 960.0, + 293.0, + 960.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/HXjt-kRBzvu/HXjt-kRBzvu.md b/parse/train/HXjt-kRBzvu/HXjt-kRBzvu.md new file mode 100644 index 0000000000000000000000000000000000000000..5b24585362badffeed744b7a6cde94bea55ca3a8 --- /dev/null +++ b/parse/train/HXjt-kRBzvu/HXjt-kRBzvu.md @@ -0,0 +1,391 @@ +# Bayesian Network Structure Learning using Digital Annealer + +Anonymous Author(s) +Affiliation +Address +email + +# Abstract + +1 Annealing processors, which efficiently solve a quadratic unconstrained binary +2 optimization (QUBO), are a potential breakthrough in improving the accuracy +3 of score-based Bayesian network structure learning. However, currently, the bit +4 capacity of an annealing processor is very limited. To utilize the power of an +5 nealing processors, it is necessary to encode score-based learning problems into +6 QUBO within the upper bound of bits. In this paper, we propose a novel approach +7 with direct encoding of candidate parent sets in the form of Cartesian products. +8 Experimental results on benchmark networks with 27 to 70 variables show that +9 our approach requires lesser bits than the bit capacity of the second-generation +10 Fujitsu digital annealer, a fully coupled annealing processor developed by with +11 semiconductor technology. Moreover, we demonstrate that the digital annealer +12 with our conversion method consistently outperforms the state-of-the-art heuristic +13 algorithms on the benchmark networks. + +# 14 1 Introduction + +15 A Bayesian network is a probabilistic graphical model that represents the structure of a joint probabil +16 ity distribution among random variables in a directed acyclic graph (DAG) [Pearl, 1988]. One class +17 of associated computational problems is learning the structure of a Bayesian network from data. We +18 focus on score-based Bayesian network structure learning for finding the DAG with a maximal score +19 that depends on the data [Cooper and Herskovits, 1992, Cowell, 2001]. +20 The Bayesian network learning problem is NP-hard [Chickering et al., 2004]; therefore, the standard +21 methodology is using heuristic approaches. Many algorithms have been proposed to improve the +22 accuracy and to reduce the running time. A search over the space of orderings [Teyssier and Koller, +23 2005, Scanagatta et al., 2015] is one of the most successful heuristic approaches. +24 Annealing processors may contribute to finding a high-scoring network structure in a realistic +25 timeframe. An annealing processor is expected to be an alternative hardware to von Neumann +26 computers for quadratic unconstrained binary optimization (QUBO) problems. In particular, it is +27 reported that complementary metal oxide semiconductor (CMOS) annealing processors already +28 outperform conventional computers on the speed of solving max-cut problems [Gyoten et al., 2018]. +29 We note that the bit capacity of an annealing processor is currently limited. Therefore, we need an +30 efficient conversion method of Bayesian network structure learning into QUBO within the limited +31 bits. Additionally, it is also important to show the lower bounds of penalty coefficients because the +32 precision for the biases and variable couplers is limited. +33 Annealing processors are classified into the nearest neighbor type and the fully connected type +34 [Yamamoto, 2020]. While the coupling nodes of a nearest neighbor annealing processor is limited to +35 only between adjacent nodes, the coupling exists between arbitrary nodes of a fully coupled annealing +36 processor. Though the scalability of nearest neighbor annealing processors is high, it is necessary to +37 consider the additional bits for minor embedding [Choi, 2008, 2010]. +38 O’Gorman et al. 2014 proposed a method to convert score-based Bayesian network structure learning +39 into QUBO that requires $\mathcal { O } ( n ^ { 2 } )$ bits for $n$ random variables and a maximum parent set size $m = 2$ . +40 They also demonstrated the sufficient lower bounds of penalty coefficients. However, when $m \geq 3$ , +41 the number of necessary auxiliary variables for a quadratization [Boros and Gruber, 2014] is at most +42 $O ( n ( n - 1 ) ^ { \frac { m } { 2 } } )$ . This is a significant disadvantage for the current limited bit capacity of annealing +43 processors. +44 In this study, we propose an efficient conversion method based on the advanced identification of +45 candidate parent sets and their representation in the form of Cartesian products. We also provide a +46 greedy algorithm to decompose the candidate parent sets into the form of Cartesian products and +47 prove the sufficient lower bounds of penalty coefficients. +48 Experimental results on benchmark networks with 27 to 70 variables show that our conversion method +49 reduces the required bits significantly in comparison to the previous work [O’Gorman et al., 2014]. +50 Our approach allows us to utilize the power of the second generation Fujitsu digital annealer, a fully +51 coupled CMOS annealing processor [Aramon et al., 2019]. We demonstrate that the digital annealer +52 consistently outperforms the ordering space search algorithms on the benchmark networks. + +# 53 2 Background + +# 2.1 Score-based Bayesian Network Structure Learning + +55 The goal of score-based Bayesian network structure learning is to find a DAG with maximal score. +56 Given to random variables $\mathscr { X } = ( X _ { i } ) _ { i = 1 } ^ { n }$ and a complete data set of $N$ instances $\mathcal { D } = \{ D _ { 1 } , \cdot \cdot \cdot , D _ { N } \}$ , +57 we optimize the parent set $\Pi _ { i }$ of each random variable, + +$$ +\Pi _ { 1 } ^ { * } , \cdot \cdot \cdot , \Pi _ { n } ^ { * } = \underset { \Pi _ { 1 } , \cdots , \Pi _ { n } \subset { \cal X } } { \arg \operatorname* { m i n } } \sum _ { i = 1 } ^ { n } - \log S ^ { ( i ) } ( \Pi _ { i } \mid { \mathcal { D } } ) , +$$ + +where 58 ${ \mathcal { G } } = ( \mathcal { V } , { \mathcal { E } } ) , \mathcal { V } = \{ 1 , \cdots , n \} , { \mathcal { E } } = \{ ( j , i ) | j , i \in \{ 1 , \cdots , n \} , X _ { j } \in \Pi _ { i } \} ,$ and $S _ { i } : \Pi _ { i } \mathbb { R }$ is a local score function corresponding to 59 $X _ { i }$ . The Bayesian Dirichlet equivalent uniform (BDeu) score 60 [Buntine, 1991] is one of the commonly used scores, + +$$ +S _ { \mathrm { B D e u } } ^ { ( i ) } ( \Pi _ { i } \mid \mathcal { D } ) \equiv \prod _ { j = 1 } ^ { \beta _ { i } } \frac { \Gamma ( \alpha _ { i , j } ) } { \Gamma ( N _ { i , j } + \alpha _ { i , j } ) } \prod _ { k = 1 } ^ { \gamma _ { i } } \frac { \Gamma ( N _ { i , j , k } + \alpha _ { i , j , k } ) } { \Gamma ( \alpha _ { i , j , k } ) } , +$$ + +61 ere $\begin{array} { r } { N = \sum _ { j = 1 } ^ { \beta _ { i } } N _ { i , j } , N _ { i , j } = \sum _ { k = 1 } ^ { \gamma _ { i } } N _ { i , j , k } , \alpha _ { i , j } = \sum _ { k = 1 } ^ { \gamma _ { i } } \alpha _ { i , j , k } , \beta _ { i } } \end{array}$ is the numb f joint states of $\Pi _ { i }$ is the number of states of $X _ { i }$ , $N _ { i , j , k }$ is the number of cases of the parent set $\Pi _ { i }$ 63 and Xi in its k-th state, αi,j,k = αβ γ is the hyperparameter of the Dirichlet function, and $0 < \alpha \in \mathbb { R }$ 64 is called equivalent sample size [Heckerman et al., 1995a]. + +# 65 2.2 Hamiltonian + +66 The Hamiltonian, which is the objective function of an annealing processor, is a quadratic pseudo +67 Boolean function, + +$$ +H ( \pmb { \sigma } ) = \sum _ { i \in \mathcal { V } _ { \mathrm { A P } } } h _ { i } \sigma _ { i } + \sum _ { ( i , j ) \in \mathcal { E } _ { \mathrm { A P } } } J _ { i , j } \sigma _ { i } \sigma _ { j } , +$$ + +68 where ${ \pmb \sigma } = ( { \boldsymbol \sigma } _ { i } ) _ { i = 1 } ^ { | { \boldsymbol \nu } _ { \mathrm { A P } } | } \in \mathbb { B } ^ { | { \boldsymbol \nu } _ { \mathrm { A P } } | }$ , the biases $h _ { i } \in \mathbb { R }$ for all $i \in \mathcal { V } _ { \mathrm { A P } }$ , the couplers $J _ { i , j } \in \mathbb { R }$ for all +69 $( i , j ) \in \mathcal { E } _ { \mathrm { A P } }$ , and the graph $\mathcal { G } _ { \mathrm { A P } } = ( \nu _ { \mathrm { A P } } , \mathcal { E } _ { \mathrm { A P } } )$ . Higher degree problems are reformed into quadratic +70 ones using auxiliary variables. This reformulation is called quadratization. + +71 Definition 1. If a quadratic polynomial function $g ( \pmb { v } , \pmb { h } )$ is a quadratization of a pseudo-Boolean function 72 $f ( v )$ , then $f ( v ) = \operatorname* { m i n } _ { h \in \mathbb { B } ^ { J } } g ( v , h )$ for all $\pmb { v } \in \mathbb { B } ^ { I }$ . + +73 Anthony et al. 2016 proved that every pseudo-Boolean function of $I$ variables and of degree $K$ has +74 a quadratization involving at most $\dot { \mathcal { O } } ( \dot { I } ^ { \frac { K } { 2 } } )$ auxiliary variables. In particular, at most $\check { \mathcal { O } ( 2 ^ { \frac { I } { 2 } } ) }$ when +75 $K = I$ . It is well known that every pseudo-Boolean function can be uniquely represented as a +76 multilinear polynomial in its variables [Boros and Hamme, 2002]. + +# 77 2.3 Basic Conversion of Score-based Bayesian Network Structure Learning + +Using 78 $n ( n - 1 )$ bits to encode the paths into $\pmb { d } = ( ( d _ { j , i } ) _ { 1 \leq j \leq n , j \neq i } ) _ { i = 1 } ^ { n } \in \mathbb { B } ^ { n ( n - 1 ) }$ $( d _ { j , i } = 1$ if 79 $X _ { j }$ is the parent of $X _ { i }$ , $d _ { j , i } = 0$ otherwise) and $\binom { n } { 2 }$ bits to encode the topological orders into 80 $r = ( r _ { i , j } ) _ { 1 \leq i < j \leq n } \in \mathbb { B } ^ { \binom { n } { 2 } } ( r _ { i , j } =$ if the order of $X _ { i }$ is higher than $X _ { j }$ , $r _ { i , j } = 1$ otherwise), it is 81 possible to represent eq. (1) on the Hamiltonian, + +$$ +H _ { \mathrm { t o t a l } } ( d , r ) \equiv \sum _ { i = 1 } ^ { n } H _ { \mathrm { s c o r e } } ^ { ( i ) } ( d , i ) + H _ { \mathrm { c y c l e } } ( d , r ) . +$$ + +82 The states of $d _ { \cdot , i }$ are mapped one-to-one to the states of $\Pi _ { i }$ . Let $\Pi _ { i } = \pi ^ { ( i ) } ( d _ { \cdot , i } )$ for all $1 \leq i \leq n$ . +83 The local score of the Hamiltonian is + +$$ +H _ { \mathrm { s c o r e } } ^ { ( i ) } ( d _ { \cdot , i } ) \equiv - \log S ^ { ( i ) } ( \pi ^ { ( i ) } ( d _ { \cdot , i } ) \mid \mathcal { D } ) + \log S ^ { ( i ) } ( \phi \mid \mathcal { D } ) , +$$ + +84 for all $1 \leq i \leq n$ . The score function has a quadratization involving at most $\mathcal { O } ( n 2 ^ { \frac { n - 1 } { 2 } } )$ auxiliary +85 variables. O’Gorman et al. 2014 added the maximum parent set size constraint to the Hamiltonian. +86 In this case, the number of auxiliary variables is at most $\mathcal { O } ( n ( n - 1 ) ^ { \frac { m } { 2 } } )$ . The cycle constraint of the +87 Hamiltonian consists of the topological order constraint and the consistency constraint, + +$$ +H _ { \mathrm { c y c l e } } ( d , r ) \equiv \sum _ { 1 \leq i < j < k \leq n } \delta _ { 1 } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) + \sum _ { 1 \leq i < j \leq n } \delta _ { 2 } ( d _ { i , j } r _ { i , j } + d _ { j , i } ( 1 - r _ { i , j } ) ) , +$$ + +88 where $R ( r _ { 1 } , r _ { 2 } , r _ { 3 } ) = r _ { 1 } r _ { 2 } ( 1 - r _ { 3 } ) + ( 1 - r _ { 1 } ) ( 1 - r _ { 2 } ) r _ { 3 }$ for all $r _ { 1 } , r _ { 2 } , r _ { 3 } \in \mathbb { B }$ . When the penalty +89 coefficients $0 < \delta _ { 1 } , \delta _ { 2 } \in \mathbb { R }$ are sufficiently large, the DAG constraint is satisfied indirectly through +90 the relationship of the paths $^ d$ and the topological order $\pmb { r }$ . If it holds that + +$$ +\operatorname* { m a x } \{ 0 , \operatorname* { m a x } _ { \substack { 1 \leq j ^ { * } , i ^ { * } \leq n d _ { \cdot , j ^ { * } } \in \mathbb { B } ^ { n - 1 } } } ( H _ { \mathrm { s c o r e } } ^ { ( i ^ { * } ) } ( d _ { \cdot , i ^ { * } } ^ { ( j ^ { * } , i ^ { * } ) } ) - H _ { \mathrm { s c o r e } } ^ { ( i ^ { * } ) } ( d _ { \cdot , i ^ { * } } ) ) \} < \delta _ { 1 } < \frac { \delta _ { 2 } } { n - 2 } , +$$ + +91 then there is no cycle on the paths of the ground state, where ∗ ∗ ∗ ∗ ${ \pmb d } ^ { ( j ^ { \ast } , i ^ { \ast } ) } = ( ( d _ { j , i } ) _ { 1 \leq j \leq n , j \neq i } ^ { ( j ^ { \ast } , i ^ { \ast } ) } ) _ { i = 1 } ^ { n }$ +92 $d _ { j , i } ^ { ( j ^ { * } , i ^ { * } ) } = 0$ if $( j , i ) = ( j ^ { * } , i ^ { * } )$ , $d _ { j , i } ^ { ( j ^ { * } , i ^ { * } ) } = d _ { j , i }$ otherwise. The computational cost to obtain the left +93 side of eq. (7) is at most $\mathcal { O } ( n ^ { m + 1 } )$ . In particular, at most $O ( n ^ { 2 } 2 ^ { n - 2 } )$ when $m = n - 1$ . + +# 94 3 Candidate Parent Set Decomposition + +Parent set identification is a major technique to narrow the search space of structure optimization, based on the relationship between parent sets and local scores under the DAG constraints [de Campos and Ji, 2011, Correia et al., 2020]. The collection of candidate parent sets of a random variable $X _ { i }$ is $\{ W \subseteq { \mathcal { X } } \setminus \{ X _ { i } \} \mid W ^ { \prime } \subset W \Rightarrow S ^ { ( i ) } ( W ^ { \prime } \mid { \mathcal { D } } ) < S ^ { ( i ) } ( W \mid { \mathcal { D } } ) \}$ . To reduce the required bits of the score component of the Hamiltonian, we propose an efficient conversion method with the parent set identification. We directly encode the candidate parent sets instead of using the paths $^ d$ . + +Moreover, we decompose the candidate parent sets $( W _ { h , i } ) _ { h = 0 } ^ { \lambda _ { i } }$ of each random variable into the form of Cartesian products as follows: + +1. Decompose $( W _ { h , i } ) _ { h = 0 } ^ { \lambda _ { i } }$ into $( W _ { h , i } \cap Z _ { i } ) _ { h = 0 } ^ { \lambda _ { i } } , ( W _ { h , i } \cap ( { \mathcal { X } } \setminus Z _ { i } ) ) _ { h = 0 } ^ { \lambda _ { i } } ,$ +2. Remove duplicates in the elements of $( W _ { h , i } \cap Z _ { i } ) _ { h = 0 } ^ { \lambda _ { i } } , ( W _ { h , i } \cap ( { \mathcal { X } } \setminus Z _ { i } ) ) _ { h = 0 } ^ { \lambda _ { i } } ,$ + +$$ +\because ( W _ { h , i } \cap Z _ { i } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } , ( W _ { h , i } \cap ( \mathcal { X } \setminus Z _ { i } ) ) _ { h = 0 } ^ { \lambda _ { 2 , i } } \mathrm { ~ i n ~ } ( U _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 2 , i } } , +$$ + +where 106 $Z _ { i } \subseteq \cup _ { h = 0 } ^ { \lambda _ { i } } W _ { h , i } , W _ { 0 , i } = U _ { 0 , i } = V _ { 0 , i } = \phi , \lambda _ { i } , \lambda _ { 1 , i } , \lambda _ { 2 , i } \in \mathbb { N } \cup \{ 0 \}$ for all $1 \leq i \leq n$ . There 107 is a clear relationship, + +$$ +\begin{array} { r } { \{ W _ { 0 , i } , \cdots , W _ { \lambda _ { i } , i } \} \subseteq \{ U \cup V \mid ( U , V ) \in \{ U _ { 0 , i } , \cdots , U _ { \lambda _ { 1 , i } , i } \} \times \{ V _ { 0 , i } , \cdots , V _ { \lambda _ { 2 , i } , i } \} \} , } \end{array} +$$ + +108 for all $1 \leq i \leq n$ . Here, given that the Hamiltonian is a quadratic pseudo-Boolean function, we can represent the score against 109 fore, it is possible to encod110 $U _ { h , i } \cup V _ { h ^ { \prime } , i }$ by allocating ate parent sets $U _ { h , i } , V _ { h ^ { \prime } , i }$ to two bits on thamiltonian using $( U _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 2 , i } }$ 111 The number of required bits of the score component of the Hamiltonian is ${ \textstyle \sum _ { i = 1 } ^ { n } } ( \lambda _ { 1 , i } + \lambda _ { 2 , i } )$ . + +112 Example 1. An example of the candidate parent sets in the form of Cartesian products as follows: + +$$ +\begin{array} { r l } & { \mathcal { X } = \{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \} , \ Z _ { i } = \{ X _ { 1 } , X _ { 2 } \} , \ \lambda _ { i } = 5 , \ \lambda _ { 1 , i } = 2 , \ \lambda _ { 2 , i } = 1 } \\ & { ( W _ { h , i } ) _ { h = 0 } ^ { \lambda _ { i } } = ( \phi , \{ X _ { 1 } \} , \{ X _ { 1 } , X _ { 2 } \} , \{ X _ { 3 } , X _ { 4 } \} , \{ X _ { 1 } , X _ { 3 } , X _ { 4 } \} , \{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \} ) , } \\ & { ( W _ { h , i } \cap Z _ { i } ) _ { h = 0 } ^ { \lambda _ { i } } = ( \phi , \{ X _ { 1 } \} , \{ X _ { 1 } , X _ { 2 } \} , \phi , \{ X _ { 1 } \} , \{ X _ { 1 } , X _ { 2 } \} ) , } \\ & { ( W _ { h , i } \cap ( \mathcal { X } \setminus Z _ { i } ) ) _ { h = 0 } ^ { \lambda _ { i } } = ( \phi , \phi , \phi , \{ X _ { 3 } , X _ { 4 } \} , \{ X _ { 3 } , X _ { 4 } \} , \{ X _ { 3 } , X _ { 4 } \} ) , } \\ & { ( U _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } = ( \phi , \{ X _ { 1 } \} , \{ X _ { 1 } , X _ { 2 } \} ) , \ ( V _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 2 , i } } = ( \phi , \{ X _ { 3 } , X _ { 4 } \} ) . } \end{array} +$$ + +We optimize 113 $Z _ { i } \subseteq \cup _ { h = 0 } ^ { \lambda _ { i } } W _ { h , i }$ to minimize $\lambda _ { 1 , i } + \lambda _ { 2 , i }$ . However, it is often infeasible to search all 114 elements of the power set $\mathcal { P } ( \cup _ { h = 0 } ^ { \lambda _ { i } } W _ { h , i } )$ . Therefore, we heuristically search 3 $Z _ { i }$ adding elements one by one, as algorithm 1. The computational cost is at most $1 \leq i \leq n$ + +# Algorithm 1 Greedy Candidate Parent Set Decomposition + +1: Input: $( W _ { h , i } ) _ { h = 0 } ^ { \lambda _ { i } }$ Output: $Z$ Initialize: $\lambda \lambda _ { i } , Z ^ { \prime } \phi , Z \phi$ . +2: for $d = 1$ to $| \cup _ { h = 0 } ^ { \lambda _ { i } } W _ { h , i } | - 1$ do +3: for $X$ in $\cup _ { h = 0 } ^ { \lambda _ { i } } W _ { h , i } \setminus Z$ do +4: if $\lambda _ { 1 , i } + \lambda _ { 2 , i } < \lambda$ for $Z _ { i } = Z \cup \{ X \}$ then $\lambda \lambda _ { 1 , i } + \lambda _ { 2 , i } , Z ^ { \prime } Z \cup \{ X \} .$ . +5: if $Z \neq Z ^ { \prime }$ then $Z Z ^ { \prime }$ else break + +115 + +116 Example 2. An example of the bit reduction flow of algorithm 1 is as follows: + +$$ +\begin{array} { r l } & { Z _ { i } = \phi , \lambda _ { 1 , i } = 0 , \lambda _ { 2 , i } = 5 : ( \phi ) \times ( \phi , \{ X _ { 1 } \} , \{ X _ { 1 } , X _ { 2 } \} , \{ X _ { 3 } , X _ { 4 } \} , \{ X _ { 1 } , X _ { 3 } , X _ { 4 } \} , \{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \} , \{ X _ { 1 } , X _ { 3 } , X _ { 3 } \} , } \\ & { Z _ { i } = \{ X _ { 1 } \} , \lambda _ { 1 , i } = 1 , \lambda _ { 2 , i } = 3 : ( \phi , \{ X _ { 1 } \} ) \times ( \phi , \{ X _ { 2 } \} , \{ X _ { 3 } , X _ { 4 } \} , \{ X _ { 2 } , X _ { 3 } , X _ { 4 } \} ) , } \\ & { Z _ { i } = \{ X _ { 1 } , X _ { 2 } \} , \lambda _ { 1 , i } = 2 , \lambda _ { 2 , i } = 1 : ( \phi , \{ X _ { 1 } \} , \{ X _ { 1 } , X _ { 2 } \} ) \times ( \phi , \{ X _ { 3 } , X _ { 4 } \} ) . } \end{array} +$$ + +# 117 4 Efficient Conversion of Score-based Bayesian Network Structure Learning + +We make 118 $( U _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 2 , i } }$ correspond to $( p _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } , ( q _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 2 , i } }$ one-to-one, where $p _ { h , i } , q _ { h ^ { \prime } , i } \in \mathbb { B }$ 119 for all $0 \leq h \leq \lambda _ { 1 , i } , 0 \leq h ^ { \prime } \leq \lambda _ { 2 , i } , 1 \leq i \leq n$ . To identify the parent sets, we use the one-to-one 120 correspondence constraint that $\begin{array} { r } { \sum _ { h = 0 } ^ { \lambda _ { 1 , i } } p _ { h , i } = \sum _ { h = 0 } ^ { \lambda _ { 2 , i } } q _ { h , i } = 1 } \end{array}$ for all $1 \leq i \leq n$ . The Hamiltonian 121 consists of the score component, the one-to-one correspondence constraint, and the cycle constraint, + +$$ +H _ { \mathrm { t o t a l } } ^ { * } ( p , q , r ) \equiv \sum _ { i = 1 } ^ { n } ( H _ { \mathrm { s c o r e } } ^ { * ( i ) } ( p . , i , q . , i ) + H _ { \mathrm { o n e } } ^ { * ( i ) } ( p . , i , q . , i ) ) + H _ { \mathrm { c y c l e } } ^ { * } ( p , q , r ) , +$$ + +where 122 $\pmb { p } = ( ( p _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } ) _ { i = 1 } ^ { n } , \pmb { q } = ( ( q _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 2 , i } } ) _ { i = 1 } ^ { n }$ . Under the one-to correspondence constraint, +123 we can represent the paths among random variables indirectly using $\mathbf { \omega } _ { p , q }$ +124 variables, + +$$ +d _ { j , i } ^ { * } \equiv \sum _ { 1 \leq h \leq \lambda _ { 1 , i } \atop \overline { { X } } _ { j } \in U _ { h , i } } p _ { h , i } + \sum _ { 1 \leq h \leq \lambda _ { 2 , i } \atop \overline { { X } } _ { j } \in V _ { h , i } } q _ { h , i } , +$$ + +![](images/d4f7e8cc2ac82f681ab72206ecf688f4b23e1a397be97e5c0273bc55dfd11cd4.jpg) +Figure 1: An example of bit allocation for our conversion method. $n = 4 , \lambda _ { 1 , 1 } = 3 , \lambda _ { 2 , 1 } = 1 , \lambda _ { 1 , 2 } =$ 1 $\bar { \lambda } _ { 2 , 2 } = 0 , \lambda _ { 1 , 3 } = \bar { 1 } , \lambda _ { 2 , 3 } = 1 , \lambda _ { 1 , 4 } = 0 , \lambda _ { 2 , 4 } = 0 , U _ { 1 , 1 } = \{ 2 , 3 \} , U _ { 2 , 1 } = \{ 3 \} , U _ { 3 , 1 } = \{ 2 \} , V _ { 1 , 1 } = \{ 3 \} , U _ { 2 , 2 } = \{ 3 \} , U _ { 3 , 2 } = \{ 3 \} , U _ { 2 , 2 } = \{ 3 \} ,$ $\{ 4 \} , U _ { 1 , 2 } = \{ 3 \} , U _ { 1 , 3 } = \{ 1 \} , V _ { 1 , 3 } = \{ 4 \}$ . Circle : $\mathbf { \mu } _ { p , q }$ . Square $: \textbf { { r } }$ . Red lines include in the score component of the Hamiltonian, a green line in the one-to-one correspondence constraint, and blue lines in the cycle constraint. + +125 for all $1 \leq j , i \leq n$ . Figure 1 is an example of bit allocation using our conversion method. The 126 number of bits required in our conversion method is $\begin{array} { r } { \sum _ { i = 1 } ^ { n } ( \lambda _ { 1 , i } + \lambda _ { 2 , i } ) + \binom { n } { 2 } } \end{array}$ . Note that we do not directly encode127 $p _ { 0 , i } , q _ { 0 , i }$ on the Hamiltonian. + +128 Score Component. The local score component of the Hamiltonian is + +$$ +H _ { \mathrm { s c o r r e } } ^ { * ( i ) } ( { \pmb p } _ { . , i } , { \pmb q } _ { . , i } ) \equiv \sum _ { h = 1 } ^ { \lambda _ { 1 , i } } s _ { 1 , h , i } p _ { h , i } + \sum _ { h = 1 } ^ { \lambda _ { 2 , i } } s _ { 2 , h , i } q _ { h , i } + \sum _ { h = 1 } ^ { \lambda _ { 1 , i } } \sum _ { h ^ { \prime } = 1 } ^ { \lambda _ { 2 , i } } t _ { h , h ^ { \prime } , i } p _ { h , i } q _ { h ^ { \prime } , i } , +$$ + +129 for all $1 \leq i \leq n$ . We can get these coefficients by solving simultaneous equations under the +130 one-to-one correspondence constraint, $s _ { 1 , h , i } = \_ \log S ^ { ( i ) } ( U _ { h , i } \mid \mathcal { D } ) + \log S ^ { ( i ) } ( \phi \mid \mathcal { D } ) , s _ { 2 , h , i } =$ +131 $\begin{array} { r } { - \log S ^ { ( i ) } ( V _ { h , i } \mid \mathcal D ) + \log S ^ { ( i ) } ( \phi \mid \mathcal D ) , t _ { h , h ^ { \prime } , i } = - \log S ^ { ( i ) } ( U _ { h , i } \cup V _ { h ^ { \prime } , i } \mid \mathcal D ) + \log S ^ { ( i ) } ( U _ { h , i } \mid \mathcal D ) + } \end{array}$ +132 $\log S ^ { ( i ) } ( V _ { h ^ { \prime } , i } \mid \mathcal { D } ) - \log S ^ { ( i ) } ( \phi \mid \mathcal { D } )$ . + +133 One-to-One element from 134 $( U _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \lambda _ { 2 , i } }$ nstraint. We penalize the connection among bits to select each , + +$$ +H _ { \mathrm { o n e } } ^ { * ( i ) } ( p . , . , q . , i ) \equiv \sum _ { 1 \leq h < h ^ { \prime } \leq \lambda _ { 1 , i } } \xi _ { 1 , i } p _ { h , i } p _ { h ^ { \prime } , i } + \sum _ { 1 \leq h < h ^ { \prime } \leq \lambda _ { 2 , i } } \xi _ { 2 , i } q _ { h , i } q _ { h ^ { \prime } , i } , +$$ + +135 136 $1 \leq i \leq n$ penalty coefficient is induced indirectly $0 < \xi _ { 1 , i } , \xi _ { 2 , i } \in \mathbb { R }$ . If $\xi _ { 1 , i } , \xi _ { 2 , i }$ is sufficient large, $\begin{array} { r } { \sum _ { h = 0 } ^ { \lambda _ { 1 , i } } p _ { h , i } = \sum _ { h = 0 } ^ { \lambda _ { 2 , i } } q _ { h , i } = 1 } \end{array}$ + +137 Cycle Constraint. Compared to eq. (6), the cycle constraint of the Hamiltonian is + +$$ +H _ { \mathrm { c y c l e } } ^ { * } ( p , q , r ) \equiv \sum _ { 1 \leq i < j < k \leq n } \delta _ { 1 } ^ { * } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) + \sum _ { 1 \leq i < j \leq n } \delta _ { 2 } ^ { * } ( d _ { i , j } ^ { * } r _ { i , j } + d _ { j , i } ^ { * } ( 1 - r _ { i , j } ) ) , +$$ + +where the penalty coefficients 138 $0 < \delta _ { 1 } ^ { * } , \delta _ { 2 } ^ { * } \in \mathbb { R }$ . By setting $\delta _ { 1 } ^ { * } , \delta _ { 2 } ^ { * }$ appropriately, we can prevent the 139 cycle from occurring. + +141 We demonstrate the sufficient lower bounds of penalty coefficients. The basic idea is that we find the +142 range of penalty coefficients so that the change in return value of the Hamiltonian is negative when +143 the input state changes to the state we desire to induce. +144 One-to-One Correspondence Constraint. We consider to decrease the value of $\scriptstyle \sum _ { h = 1 } ^ { \lambda _ { 1 , i ^ { * } } } p _ { h , i ^ { * } }$ by +145 146 $\begin{array} { r } { p _ { h ^ { * } , i ^ { * } } = 1 , \sum _ { h = 1 } ^ { \lambda _ { 1 , i ^ { * } } } p _ { h , i ^ { * } } > 1 } \end{array}$ , e $H _ { \mathrm { t o t a l } } ^ { * } ( p , q , r ) -$ +$\begin{array} { r } { H _ { \mathrm { t o t a l } } ^ { * } ( p ^ { ( h ^ { * } , i ^ { * } ) } , q , r ) \geq \xi _ { 1 , i ^ { * } } + s _ { 1 , h ^ { * } , i ^ { * } } + \sum _ { h = 1 } ^ { \lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } . } \end{array}$ $\pmb { p } ^ { ( h ^ { * } , i ^ { * } ) } = ( ( p _ { h , i } ^ { ( h ^ { * } , i ^ { * } ) } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } ) _ { i = 1 } ^ { n }$ +147 and p(h∗,i∗)h,i = 0 if (h, i) = (h∗, i∗), p(h∗,h,i otherwise. Considering the case where $\pmb { p }$ and $\pmb q$ +148 are swapped in the above, if $\xi _ { 1 , i } , \xi _ { 2 , i }$ satisfy that + +$$ +\begin{array} { r l r } & { } & { \displaystyle \operatorname* { m a x } _ { 0 \le h \le \lambda _ { 1 , i } } \bigl ( - s _ { 1 , h , i } - \sum _ { h ^ { \prime } = 1 } ^ { \lambda _ { 2 , i } } \operatorname* { m i n } \bigl \{ 0 , t _ { h , h ^ { \prime } , i } \bigr \} \bigr ) \bigr ) < \xi _ { 1 , i } , } \\ & { } & { \displaystyle \operatorname* { m a x } _ { 0 \le h \le \lambda _ { 2 , i } } \bigl ( - s _ { 2 , h , i } - \sum _ { h ^ { \prime } = 1 } ^ { \lambda _ { 1 , i } } \operatorname* { m i n } \bigl \{ 0 , t _ { h ^ { \prime } , h , i } \bigr \} \bigr ) \bigr ) < \xi _ { 2 , i } , } \end{array} +$$ + +149 for all $1 \leq i \leq n$ , then the grand state does not violate the one-to-one correspondence constraint. +150 The computational cost to obtain the left side of eq. (14) and eq. (15) is at most $\mathcal { O } ( \lambda _ { 1 , i } \lambda _ { 2 , i } )$ for all +151 $1 \leq i \leq n$ . +152 Cycle Constraint. We consider four patterns of $( r _ { i ^ { * } , j ^ { * } } , d _ { j ^ { * } , i ^ { * } } ^ { * } , d _ { i ^ { * } , j ^ { * } } ^ { * } )$ violating the consistency +153 constraint. It is assumed that $X _ { j ^ { * } } ~ \in ~ U _ { h ^ { * } , i ^ { * } } , X _ { j ^ { * } } ~ \notin ~ U _ { h ^ { * * } , i ^ { * } } ~ \subset ~ U _ { h ^ { * } , i ^ { * } } , p _ { h ^ { * } , i ^ { * } } ~ = ~ 1 , p _ { h ^ { * * } , i ^ { * } } ~ = ~ 0$ . +154 155 In thewhere $r ^ { ( i ^ { * } , j ^ { * } ) } = ( r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } ) _ { 1 \leq i < j \leq n }$ $( 0 , 1 , 0 )$ t , $H _ { \mathrm { t o t a l } } ^ { * } ( p , q , r ) - H _ { \mathrm { t o t a l } } ^ { * } ( p , q , r ^ { ( i ^ { * } , j ^ { * } ) } ) \geq \delta _ { 2 } ^ { * } - ( n - 2 ) \delta _ { 1 } ^ { * }$ $r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = 1 - r _ { i , j }$ $( i , j ) = ( i ^ { * } , j ^ { * } ) , r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = r _ { i , j }$ +156 otherwise. Similarly, it is possible to consider the case of $( 1 , 0 , 1 )$ . In the case of $( 0 , 1 , 1 )$ , +157 it holds that $\begin{array} { r } { H _ { \mathrm { t o t a l } } ^ { * } ( p , q , r ) - H _ { \mathrm { t o t a l } } ^ { * } ( p ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } , q , r ) \ge \delta _ { 2 } ^ { * } + s _ { 1 , h ^ { * } , i ^ { * } } + \sum _ { h = 1 } ^ { \lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } - \frac { 1 - \lambda _ { 1 } } { \lambda _ { 2 } } | \nabla p _ { h } ( q _ { h } ^ { * } , h ^ { * } , i ^ { * } ) | _ { \mathbb { H } } ( q _ { h } ^ { * } , h ^ { * } , i ^ { * } ) , } \end{array}$ +158 $\begin{array} { r } { s _ { 1 , h ^ { * * } , i ^ { * } } - \sum _ { h = 1 } ^ { \lambda _ { 2 , i } } t _ { h ^ { * * } , h , i ^ { * } } q _ { h , i ^ { * } } } \end{array}$ , where $\pmb { p } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = ( ( p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } ) _ { h = 0 } ^ { \lambda _ { 1 , i } } ) _ { i = 1 } ^ { n }$ h=, and $p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = 0$ +159 if $( h , i ) \ = \ ( h ^ { * } , i ^ { * } ) , p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } \ = \ 1$ p(h∗,h∗∗,i∗)h,i = 1 if (h, i) = (h∗∗, i∗), p(h∗,h,i $p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = p _ { h , i }$ otherwise. Sim +160 ilarly, it is possible to consider the case of $( 1 , 1 , 1 )$ . These results suggest the relationship of +161 $\delta _ { 1 } ^ { * } , \delta _ { 2 } ^ { * }$ to induce the consistency constraint. Here, based on theorem 1, we consider a strategy +162 to repeat picking up one element from $\mathbfit { \Delta } \mathbf { r }$ and switching its value until $H _ { \mathrm { t r a n s } } ( r ) = 0$ . It is as +163 164 sumed that $\begin{array} { r } { T _ { \mathrm { t o t a l } } ^ { * } \big ( p ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } , q , r ^ { ( i ^ { * } , j ^ { * } ) } \big ) \geq \delta _ { 1 } ^ { * } + s _ { 1 , h ^ { * } , i ^ { * } } + \sum _ { h = 1 } ^ { \lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } - s _ { 1 , h ^ { * * } , i ^ { * } } - \sum _ { h = 1 } ^ { \lambda _ { 2 , i } } t _ { h ^ { * * } , h , i ^ { * } } q _ { h , i ^ { * } } . } \end{array}$ $H _ { \mathrm { t r a n s } } ( \mathbf { \bar { r } } ) > H _ { \mathrm { t r a n s } } ( { r ^ { ( i ^ { * } , j ^ { * } ) } } )$ . In the case of $( 1 , 1 , 0 )$ , it holds that $H _ { \mathrm { t o t a l } } ^ { * } ( p , q , r ) \sim$ +165 Similarly, it is possible to consider the case of $( 0 , 0 , 1 )$ . In the case of $( 1 , 0 , 0 )$ or $( 0 , 0 , 0 )$ , it holds +166 that $\bar { H _ { \mathrm { t o t a l } } ^ { * } } ( p , \bar { q } , r ) - \bar { H _ { \mathrm { t o t a l } } ^ { * } } ( p , q , r ^ { ( i ^ { * } , j ^ { * } ) } ) \geq \delta _ { 1 } ^ { * }$ . These results suggest the lower bound of ${ { \delta } _ { 1 } ^ { * } }$ to +167 induce the topological order constraint. Considering the case where $\pmb { p }$ and $\pmb q$ are swapped in the +168 above, if $\delta _ { 1 } ^ { * } , \delta _ { 2 } ^ { * }$ satisfy that + +$$ +\begin{array} { c } { \displaystyle \operatorname* { m a x } _ { 1 \leq i \leq n } \operatorname* { m a x } \{ \eta _ { 1 , i } , \eta _ { 2 , i } \} < \delta _ { 1 } ^ { * } < \displaystyle \frac { \delta _ { 2 } ^ { * } } { n - 2 } , } \\ { \displaystyle \operatorname* { m a x } _ { 1 \leq j \leq n } \displaystyle \operatorname* { m a x } _ { 1 \leq i \leq 1 _ { 1 , i } } \operatorname* { m a x } _ { 0 \leq h ^ { \prime } \leq \lambda _ { 1 , i } } \operatorname* { m a x } _ { 0 \leq h ^ { \prime } \leq \lambda _ { 2 , i } } ( - s _ { 1 , h , i } - t _ { h , h ^ { \prime \prime } , i } + s _ { 1 , h ^ { \prime } , i } + t _ { h ^ { \prime } , h ^ { \prime \prime } , i } ) , } \\ { \displaystyle \eta _ { 1 , i } \equiv \displaystyle \operatorname* { m a x } _ { 1 \leq j \leq n } \displaystyle \frac { \operatorname* { m a x } _ { i } } { X _ { j } \in U _ { h , i } } \operatorname* { m a x } _ { X _ { j } \in U _ { h , i } } \operatorname* { m a x } _ { 0 \leq h ^ { \prime \prime } \leq \lambda _ { 1 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \prime \prime } , h , i } + s _ { 2 , h ^ { \prime } , i } + t _ { h ^ { \prime \prime } , h ^ { \prime \prime } , i } ) , } \\ { \displaystyle \eta _ { 2 , i } \equiv \displaystyle \operatorname* { m a x } _ { 1 \leq j \leq n } \displaystyle \operatorname* { m a x } _ { 0 \leq h ^ { \prime } \leq \lambda _ { 2 , i } } \operatorname* { m a x } _ { 0 \leq h ^ { \prime } \leq \lambda _ { 2 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \prime \prime } , h , i } + s _ { 2 , h ^ { \prime } , i } + t _ { h ^ { \prime \prime } , h ^ { \prime } , i } ) , } \\ { \displaystyle \sum _ { k \geq \ell \leq n } \displaystyle \sum _ { X _ { j } \in V _ { h , i } } \sum _ { j \in V _ { h ^ { \prime } } , i \subset V _ { h , i } } ( \operatorname* { m a x } _ { 0 \leq h ^ { \prime \prime } \leq \lambda _ { 1 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \prime \prime } , h , i } + s _ { 2 , h ^ { \prime } , i } + t _ { h ^ { \prime \prime } , h ^ { \prime \prime } , i } ) , } \end{array} +$$ + +169 for $n \geq 3$ , then the grand state does not violate the cycle constraint under the one-to-one cor +170 respondence constraint. The computational cost to obtain the left side of eq. (16) is at most +171 $\begin{array} { r l } { { \mathcal { O } ( \sum _ { i = 1 } ^ { n } n \lambda _ { 1 , i } \lambda _ { 2 , i } ( \lambda _ { 1 , i } + \lambda _ { 2 , i } ) ) } \quad } & { { } } \end{array}$ . +172 Theorem 1. If it holds that $\begin{array} { r } { H _ { \mathrm { t r a n s } } ( { \pmb r } ) \equiv \sum _ { 1 \leq i < j < k \leq n } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) > 0 } \end{array}$ , then there exists +173 at least one index pair $1 \leq i ^ { * } < j ^ { * } \leq n$ which satisfy $H _ { \mathrm { t r a n s } } ( r ) > H _ { \mathrm { t r a n s } } ( r ^ { ( i ^ { * } , j ^ { * } ) } )$ , where +174 r(i∗,j∗) = (r(i∗,i,j $R ( r _ { 1 } , r _ { 2 } , r _ { 3 } ) = r _ { 1 } r _ { 2 } ( 1 - r _ { 3 } ) + ( 1 - r _ { 1 } ) ( 1 - r _ { 2 } ) r _ { 3 }$ $r ^ { ( i ^ { * } , j ^ { * } ) } = ( r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } ) _ { 1 \leq i < j \leq n }$ $r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = 1 - r _ { i , j } \ i f ( i , j ) = ( i ^ { * } , j ^ { * } )$ $r _ { 1 } , r _ { 2 } , r _ { 3 } \in \mathbb { B }$ ,, $r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = r _ { i , j }$ $r = ( r _ { i , j } ) _ { 1 \leq i < j \leq n } \in \mathbb { B } ^ { \binom { n } { 2 } }$ +176 Proof. It does not lose the generality by considering the case of $( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) = ( 1 , 1 , 0 )$ . Here, it +177 holds that $R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( 1 - r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) + R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( r _ { 1 , 2 } , 1 - r _ { 2 , 3 } , r _ { 1 , 3 } ) + R ( r _ { 1 , 2 } , 1 - r _ { 2 , 3 } , r _ { 1 , 3 } )$ +178 $R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( r _ { 1 , 2 } , r _ { 2 , 3 } , 1 - r _ { 1 , 3 } ) = 3 .$ . Additionally, it holds that $R ( r _ { 1 , 2 } , r _ { 2 , i } , r _ { 1 , i } ) - R ( 1 -$ +179 $r _ { 1 , 2 } , r _ { 2 , i } , r _ { 1 , i } ) + R ( r _ { 2 , 3 } , r _ { 3 , i } , r _ { 2 , i } ) - R ( 1 - r _ { 2 , 3 } , r _ { 3 , i } )$ ${ } _ { 3 } , r _ { 3 , i } , r _ { 2 , i } ) + R ( r _ { 1 , 3 } , r _ { 3 , i } , r _ { 1 , i } ) - R ( 1 - r _ { 1 , 3 } , r _ { 3 , i } , r _ { 1 } )$ ,i) = +180 0 for all $3 < i$ . Therefore, it holds that $H _ { \mathrm { t r a n s } } ( \pmb { r } ) - H _ { \mathrm { t r a n s } } ( \pmb { r } ^ { ( 1 , 2 ) } ) + H _ { \mathrm { t r a n s } } ( \pmb { r } ) - H _ { \mathrm { t r a n s } } ( \pmb { r } ^ { ( 2 , 3 ) } ) +$ +181 $H _ { \mathrm { t r a n s } } ( \pmb { r } ) - H _ { \mathrm { t r a n s } } ( \pmb { r } ^ { ( 1 , 3 ) } ) = 3 .$ . From this result, it holds that $H _ { \mathrm { t r a n s } } ( \pmb { r } ) - H _ { \mathrm { t r a n s } } ( \pmb { r } ^ { ( i ^ { * } , j ^ { * } ) } ) > 0$ for +182 at least one index pair $( i ^ { * } , j ^ { * } ) \in \{ ( 1 , 2 ) , ( 2 , 3 ) , ( 1 , 3 ) \}$ . □ + +Table 1: The benchmark networks from Bayesian network repository. +Algorithm 2 Greedy Candidate Parent Set Identification + +
Namenm∑=1 |IIil n∑=1 βi( -1)Ωi=1X n *
N=100N= 1000N = 10000
insurance273529843538834036
water3256610083165216735
alarm37446509182922729081
barley484841140051813101552
hailfinder5646626561446924277
hepar2706123145348376654782
+ +The average for 10 simulated datasets. + +# 183 6 Experimental Results + +To validate the performance of our approach, we use 10 simulated datasets for each instance size $N = 1 0 0 , 1 0 0 0 , 1 0 0 0 0$ and each benchmark network. The benchmark networks are discrete networks from Bayesian network repository 1. The score function is the BDeu score with $\alpha = 1$ . It is often infeasible to identify exact candidate parent sets by searching the power set $\mathcal { P } ( \mathcal { X } \backslash \{ X _ { i } \} )$ in a realistic timeframe. We use the candidate parent sets from algorithm 2. Note that the candidate parent sets depend on the heuristic search algorithms, but we do not focus on their performance in this study. Table 1 displays the information of benchmark networks. The code to replicate each experiment in this paper is available 2. + +1: Input: $\mathcal { D } , i , m$ Output: $\mathcal { L }$ Initialize: ${ \mathcal { L } } \gets \{ \phi \} , { \mathcal { L } } ^ { \prime } \gets \{ \phi \} , { \mathcal { L } } ^ { \prime \prime } \gets \phi$ +2: for $d = 1$ to $m$ do +3: for $W$ in $\mathcal { L } ^ { \prime }$ do +4: for $X$ in $\chi \setminus \{ X _ { i } \} \setminus W$ do +5: if $S _ { i } ( W ^ { \prime } \mid \mathcal { D } ) < S _ { i } ( W \cup \{ X \} \mid \mathcal { D } )$ for all $W ^ { \prime } \subset W \cup \{ X \} , W ^ { \prime } \in { \mathcal { L } }$ then +6: ${ \mathcal { L } } ^ { \prime \prime } \gets { \mathcal { L } } ^ { \prime \prime } \cup \{ W \cup \{ X \} \}$ . +7: if ${ \mathcal { L } } ^ { \prime \prime } \neq \phi$ then $\mathcal { L } \gets \mathcal { L } \cup \mathcal { L } ^ { \prime \prime } , \mathcal { L } ^ { \prime } \mathcal { L } ^ { \prime \prime } , \mathcal { L } ^ { \prime \prime } \phi$ else break +8: for $W$ in $\mathcal { L }$ do +9: if there exist $W ^ { \prime } \subset W$ that satisfies $S _ { i } ( W \mid \mathcal { D } ) \le S _ { i } ( W ^ { \prime } \mid \mathcal { D } )$ then ${ \mathcal { L } } \gets { \mathcal { L } } \setminus \{ W \}$ . + +# 192 6.1 Number of Required Bits for Score Component + +In comparison to the existing method [O’Gorman et al., 2014], we reduce the number of required bits for the score component by encoding the candidate parent sets directly. While $\textstyle \sum _ { i = 1 } ^ { n } \lambda _ { i }$ candidate parent sets is encoded in our approach, $n ( n - 1 )$ paths plus at most $O ( n ( n - 1 ) ^ { \frac { m } { 2 } } )$ auxiliary variables for $m > 2$ in the existing method. The left side of table 2 shows the reduction rate of the number of required bits for the score component. Moreover, we reduce the number of required bits for the score component to ${ \textstyle \sum _ { i = 1 } ^ { n } } ( \lambda _ { 1 , i } + \lambda _ { 2 , i } )$ by decomposing the candidate parent sets in the form of Cartesian products. The right side of table 2 shows that algorithm 1 reduces the number of required bits for the score component although there is some variation among the networks. + +Table 2: The reduction rate of the number of required bits for score component. + +
Name£ λi/n(n-1)²Σ=1 （1,i+λ2,i）/∑=1 1
N=100N=1000N = 10000N = 100N= 1000N = 10000
insurance0.098730.246771.127420.613670.472850.32476
Water0.000970.001260.004290.726800.705880.44014
alarm0.038140.047380.189380.453320.355370.21617
barley0.001710.002920.014640.767170.755380.54149
hailfinder0.000850.004090.025250.827730.601780.33365
hepar20.000210.000030.000210.496940.633460.31284
+ +\* The average ratio for 10 simulated datasets. + +Table 3: The number of required bits for fully coupled and nearest neighbor annealing processors. + +
Name∑i=1(1,i+ 入2,i）+(2） *∑=1(入1,i + λ2,i)(入1,i +λ2,i + 1) + (2)
N= 100N =1000N = 10000N= 100N= 1000N = 10000
insurance56676716613375902385482
water613648820143417205881
alarm1489147226283676127985169004
barley1247136219686758344624796
hailfinder1659195729672212708480578
hepar24777283639104499169164136939
+ +\* The average ratio for 10 simulated datasets. + +# 6.2 Selection of Annealing Processor + +From the following discussion, the Fujitsu digital annealer is suitable for our approach from the viewpoint of bit capacity. + +Fully Connected Type. To the best of our knowledge, the bit capacity of the Fujitsu digital annealer is the largest in fully coupled annealing processors. The second generation Fujitsu digital annealer can deal with problems on a scale of 8192 bits [Matsubara et al., 2020]. The left side of table 3 shows that it is possible to encode all the logical conversion results for benchmark networks to the circuit of the digital annealer within bit capacity. + +Nearest Neighbor Type. The number of additional bits required for minor embedding depends on the design of the hardware graphs. Oku et al. 2019 proposed a heuristic minor embedding algorithm for the Hitachi CMOS annealing machine [Masanao et al., 2010]. Using this algorithm, the number of required physical spins when embedding a fully connected graph is $I ^ { 2 } + I$ for $I$ variables. The conversion method proposed in this study has $n$ local fully connected graphs on $\mathbf { \Delta } _ { p , q }$ . Therefore, the number of required physical spins must be at least $\begin{array} { r } { \sum _ { i = 1 } ^ { n } ( \lambda _ { 1 , i } + \lambda _ { 2 , i } ) ( \lambda _ { 1 , i } + \lambda _ { 2 , i } + 1 ) + \binom { n } { 2 } } \end{array}$ . From the right side of table 3, it is currently infeasible to encode logical conversion results for at least some networks to the circuit of CMOS annealing machine within its 102400 nodes [Sugie et al., 2021]. As far as we know, the bit capacity of the Hitachi CMOS annealing machine is the largest in nearest neighbor annealing processors. + +# 6.3 Score Maximization + +We demonstrate the performance of Fujitsu digital annealer for score-based Bayesian network structure learning using the conversion results of $N = 1 0 0 0 0$ simulated datasets. The running time for each simulated dataset is 6000 [s]. + +![](images/4b44701b9cc855182e35704f11ec0fbc3d89db610c0991e089d4b455ff631577.jpg) +Figure 2: Results of score maximization by the baseline algorithms. For each simulated dataset and each baseline algorithm, we normalized $\begin{array} { r } { \sum _ { i = 1 } ^ { \bar { n } } ( \log S ^ { ( i ) } ( \Pi _ { i } \vert \mathcal { \bar { D } } ) - \log S ^ { ( i ) } ( \phi \vert \mathcal { D } ) ) } \end{array}$ by dividing it by the corresponding value of the Fujitsu digital annealer. In this experiment, we used the second-generation Fujitsu digital annealer. SA : simulated annealing, OBS : ordering-based search, ASOBS : acyclic selection ordering-based search. + +Baselines. We compare the results obtained by the digital annealer with those of three heuristic algorithms. One algorithm is the simulated annealing algorithm [Heckerman et al., 1995b] with a QUBO same as the one encoded into the digital annealer. Other algorithms are the ordering space search algorithms, i.e., ordering-based search and acyclic selection ordering-based search. For a fair comparison, the running time of the simulated annealing algorithm for each simulated dataset is 6000 [s] and that of the ordering space search algorithms is 6000 [s] plus the running time of algorithm 1. The computing environment is Microsoft Windows 10 Pro, 3.6 GHz Intel Core i9 processor, and 64 GB memory. + +Result. Figure 2 shows that the digital annealer is better than all the baselines for all the simulated datasets from all the benchmark networks. + +# 7 Conclusion + +We proposed a novel approach of converting a score-based Bayesian network structure learning into QUBO. The essence of this approach lies in reducing the number of required bits through the advanced identification of candidate parent sets and their representation as Cartesian products. The Fujitsu digital annealer with our conversion method improved the BDeu score for 27 to 70 variables benchmark networks over existing methods. The bit capacity limitation of annealing processor is being relaxed rapidly 3. Though our approach is still a disadvantage for larger-scale networks, we expect that our proposed algorithms will be effectively applied to larger-scale score-based Bayesian network structure learning in the near future. + +Potential Negative Societal Impacts. The development of annealing processor technology could have an impact on various industry fields. However, the number of companies that have commercialized the API usage of annealing processors is still small. Therefore, there is a concern that the market of annealing processors will not work well and the disparities among stakeholders will be widen. Researchers are required to properly evaluate the value of technology and communicate it to the business side. + +248 References +249 Judea Pearl, editor. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, 1988. Gregory F. Cooper and Edward Herskovits. A bayesian method for the induction of probabilistic networks from data. Journal of Machine Learning, 9(4), 1992. Robert G. Cowell. Conditions under which conditional independence and scoring methods lead to identical selection of bayesian network models. In Jack Breese and Daphne Koller, editors, Proceedings of the 17th Conference on Uncertainty in Artificial Intelligence, pages 91–97. Morgan Kaufmann Publishers, 2001. David M. Chickering, David Heckerman, and Christopher Meek. Large-sample learning of bayesian networks is np-hard. Journal of Machine Learning Research, 20:1287–1330, 2004. Marc Teyssier and Daphne Koller. Ordering-based search: A simple and effective algorithm for learning bayesian networks. In Proceedings of the 21th Conference on Uncertainty in Artificial Intelligence, pages 584–590, 2005. Mauro Scanagatta, Cassio Polpo de Campos, Giorgio Corani, and Marco Zaffalon. Learning bayesian networks with thousands of variables. In Proceedings of the 28th International Conference on Neural Information Processing Systems, pages 1864–1872, 2015. Hidenori Gyoten, Masayuki Hiromoto, and Takashi Sato. Area efficient annealing processor for ising model without random number generator. IEICE Transactions on Information and Systems, E101.D(2):314–323, 2018. Kasho Yamamoto. Research on Annealing Processors for Large-Scale Combinatorial Optimization Problems. PhD thesis, Graduate School of Information Science and Technology Hokkaido University, 2020. Vicky Choi. Minor-embedding in adiabatic quantum computation: I. the parameter setting problem. Quantum Information Processing, 7:193–209, 2008. Vicky Choi. Minor-embedding in adiabatic quantum computation: Ii. minor-universal graph design. Quantum Information Processing, 10:343–352, 2010. Bryan A. O’Gorman, Alejandro Perdomo-Ortiz, Ryan Babbush, Alan Aspuru-Guzik, and Vadim Smelyanskiy. Bayesian network structure learning using quantum annealing. The European Physical Journal Special Topics, 225(1), 2014. +276 Endre Boros and Aritanan Gruber. On quadratization of pseudo-boolean functions. CoRR, abs/1404.6538, 2014. +277 Maliheh Aramon, Gili Rosenberg, Elisabetta Valiante, Toshiyuki Miyazawa, Hirotaka Tamura, and Helmut G. Katzgraber. Physics-inspired optimization for quadratic unconstrained problems using a digital annealer. Frontiers in Physics, 7(48), 2019. Wray Buntine. Theory refinement of bayesian networks. In Proceedings of the 7th Conference on Uncertainty in Artificial Intelligence, pages 52–60, 1991. David Heckerman, Dan Geiger, and David M. Chickering. Learning bayesian networks: The combination of knowledge and statistical data. Journal of Machine Learning, 20(3):197–243, 1995a. +284 Martin Anthony, Endre Boros, Yves Crama, and Aritanan Gruber. Quadratic reformulations of nonlinear binary optimization problems. Mathematical Programming, 162(1):115–144, 2016. Endre Boros and Peter L. Hamme. Pseudo-boolean optimization. Journal of Discrete Applied Mathematics, 123: 155–225, 2002. Cassio P. de Campos and Qiang Ji. Efficient structure learning of bayesian networks using constraints. Journal of Machine Learning Research, 12:663–689, 2011. Alvaro H. C. Correia, James Cussens, and Cassio de Campos. On pruning for score-based bayesian network structure learning. Journal of Machine Learning Research, 108:2709–2718, 2020. Satoshi Matsubara, Motomu Takatsu, Toshiyuki Miyazawa, Takayuki Shibasaki, Yasuhiro Watanabe, Kazuya Takemoto, and Hirotaka Tamura. Digital annealer for high-speed solving of combinatorial optimization problems and its applications. In Proceedings of the 25th Asia and South Pacific Design Automation Conference, pages 667–672, 2020. +296 David Eppstein. Finding large clique minors is hard. Graph Algorithms and Applications, 13(2):197–204, 2009. +Daisuke Oku, Kotaro Terada, Masato Hayashi, Masanao Yamaoka, Shu Tanaka, and Nozomu Togawa. A fully-connected ising model embedding method and its evaluation for cmos annealing machines. IEICE Transactions on Information and Systems, E102-D(9):1696–1706, 2019. +Yamaoka Masanao, Yoshimura Chihiro, Hayashi Masato, Okuyama Takuya, Aoki Hidetaka, and Mizuno Hiroyuki. A 20k-spin ising chip to solve combinatorial optimization problems with cmos annealing. IEEE Journal of Solid-State Circuits, 51(1), 2010. +Yuya Sugie, Yuki Yoshida, Normann Mertig, Takashi Takemoto, Hiroshi Teramoto, Atsuyoshi Nakamura, Ichigaku Takigawa, Shin ichi Minato, Masanao Yamaoka, and Tamiki Komatsuzaki. Minor-embedding heuristics for large-scale annealing processors with sparse hardware graphs of up to 102,400 nodes. Soft Computing, 2021. +Daphne Koller and Nir Friedman, editors. Probabilistic Graphical Models: Principles and Techniques. The MIT Press, Cambridge, Massachusetts, 2009. +David Heckerman, Dan Geiger, and David M. Chickering. Learning bayesian networks: Search methods and experimental results. In Preliminary Papers of the 5th International Workshop on Artificial, pages 112–128, 1995b. + +# Checklist + +1. For all authors... + +(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] +(b) Did you describe the limitations of your work? [Yes] See section 7. +(c) Did you discuss any potential negative societal impacts of your work? [Yes] See section 7. +(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] + +2. If you are including theoretical results... + +(a) Did you state the full set of assumptions of all theoretical results? [Yes] See section 4. +(b) Did you include complete proofs of all theoretical results? [Yes] See section 5. + +3. If you ran experiments... + +(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] in the supplemental material +(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See section 6. +(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] See fig. 2. +(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See section 6.3. + +4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... + +(a) If your work uses existing assets, did you cite the creators? [Yes] in the supplemental material +(b) Did you mention the license of the assets? [N/A] +(c) Did you include any new assets either in the supplemental material or as a URL? [N/A] +(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] +(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] + +If you used crowdsourcing or conducted research with human subjects... + +(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] +(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] +(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] \ No newline at end of file diff --git a/parse/train/HXjt-kRBzvu/HXjt-kRBzvu_content_list.json b/parse/train/HXjt-kRBzvu/HXjt-kRBzvu_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..33736acdbd3b12512c01d1349b13e369207bbe7f --- /dev/null +++ b/parse/train/HXjt-kRBzvu/HXjt-kRBzvu_content_list.json @@ -0,0 +1,1305 @@ +[ + { + "type": "text", + "text": "Bayesian Network Structure Learning using Digital Annealer ", + "text_level": 1, + "bbox": [ + 266, + 122, + 732, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous Author(s) \nAffiliation \nAddress \nemail ", + "bbox": [ + 423, + 226, + 580, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 319, + 535, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Annealing processors, which efficiently solve a quadratic unconstrained binary \n2 optimization (QUBO), are a potential breakthrough in improving the accuracy \n3 of score-based Bayesian network structure learning. However, currently, the bit \n4 capacity of an annealing processor is very limited. To utilize the power of an \n5 nealing processors, it is necessary to encode score-based learning problems into \n6 QUBO within the upper bound of bits. In this paper, we propose a novel approach \n7 with direct encoding of candidate parent sets in the form of Cartesian products. \n8 Experimental results on benchmark networks with 27 to 70 variables show that \n9 our approach requires lesser bits than the bit capacity of the second-generation \n10 Fujitsu digital annealer, a fully coupled annealing processor developed by with \n11 semiconductor technology. Moreover, we demonstrate that the digital annealer \n12 with our conversion method consistently outperforms the state-of-the-art heuristic \n13 algorithms on the benchmark networks. ", + "bbox": [ + 148, + 352, + 766, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "14 1 Introduction ", + "text_level": 1, + "bbox": [ + 148, + 577, + 312, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "15 A Bayesian network is a probabilistic graphical model that represents the structure of a joint probabil \n16 ity distribution among random variables in a directed acyclic graph (DAG) [Pearl, 1988]. One class \n17 of associated computational problems is learning the structure of a Bayesian network from data. We \n18 focus on score-based Bayesian network structure learning for finding the DAG with a maximal score \n19 that depends on the data [Cooper and Herskovits, 1992, Cowell, 2001]. \n20 The Bayesian network learning problem is NP-hard [Chickering et al., 2004]; therefore, the standard \n21 methodology is using heuristic approaches. Many algorithms have been proposed to improve the \n22 accuracy and to reduce the running time. A search over the space of orderings [Teyssier and Koller, \n23 2005, Scanagatta et al., 2015] is one of the most successful heuristic approaches. \n24 Annealing processors may contribute to finding a high-scoring network structure in a realistic \n25 timeframe. An annealing processor is expected to be an alternative hardware to von Neumann \n26 computers for quadratic unconstrained binary optimization (QUBO) problems. In particular, it is \n27 reported that complementary metal oxide semiconductor (CMOS) annealing processors already \n28 outperform conventional computers on the speed of solving max-cut problems [Gyoten et al., 2018]. \n29 We note that the bit capacity of an annealing processor is currently limited. Therefore, we need an \n30 efficient conversion method of Bayesian network structure learning into QUBO within the limited \n31 bits. Additionally, it is also important to show the lower bounds of penalty coefficients because the \n32 precision for the biases and variable couplers is limited. \n33 Annealing processors are classified into the nearest neighbor type and the fully connected type \n34 [Yamamoto, 2020]. While the coupling nodes of a nearest neighbor annealing processor is limited to \n35 only between adjacent nodes, the coupling exists between arbitrary nodes of a fully coupled annealing \n36 processor. Though the scalability of nearest neighbor annealing processors is high, it is necessary to \n37 consider the additional bits for minor embedding [Choi, 2008, 2010]. \n38 O’Gorman et al. 2014 proposed a method to convert score-based Bayesian network structure learning \n39 into QUBO that requires $\\mathcal { O } ( n ^ { 2 } )$ bits for $n$ random variables and a maximum parent set size $m = 2$ . \n40 They also demonstrated the sufficient lower bounds of penalty coefficients. However, when $m \\geq 3$ , \n41 the number of necessary auxiliary variables for a quadratization [Boros and Gruber, 2014] is at most \n42 $O ( n ( n - 1 ) ^ { \\frac { m } { 2 } } )$ . This is a significant disadvantage for the current limited bit capacity of annealing \n43 processors. \n44 In this study, we propose an efficient conversion method based on the advanced identification of \n45 candidate parent sets and their representation in the form of Cartesian products. We also provide a \n46 greedy algorithm to decompose the candidate parent sets into the form of Cartesian products and \n47 prove the sufficient lower bounds of penalty coefficients. \n48 Experimental results on benchmark networks with 27 to 70 variables show that our conversion method \n49 reduces the required bits significantly in comparison to the previous work [O’Gorman et al., 2014]. \n50 Our approach allows us to utilize the power of the second generation Fujitsu digital annealer, a fully \n51 coupled CMOS annealing processor [Aramon et al., 2019]. We demonstrate that the digital annealer \n52 consistently outperforms the ordering space search algorithms on the benchmark networks. ", + "bbox": [ + 147, + 609, + 825, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 693, + 825, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 760, + 825, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 842, + 823, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 90, + 825, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 145, + 172, + 826, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 270, + 825, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 147, + 337, + 826, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "53 2 Background ", + "text_level": 1, + "bbox": [ + 156, + 431, + 308, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Score-based Bayesian Network Structure Learning ", + "text_level": 1, + "bbox": [ + 166, + 462, + 566, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "55 The goal of score-based Bayesian network structure learning is to find a DAG with maximal score. \n56 Given to random variables $\\mathscr { X } = ( X _ { i } ) _ { i = 1 } ^ { n }$ and a complete data set of $N$ instances $\\mathcal { D } = \\{ D _ { 1 } , \\cdot \\cdot \\cdot , D _ { N } \\}$ , \n57 we optimize the parent set $\\Pi _ { i }$ of each random variable, ", + "bbox": [ + 147, + 488, + 828, + 535 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/9885ce80b2d23ce5b895e8fff9505555e3f3c661ed5a89e650eb99b6c8db169e.jpg", + "text": "$$\n\\Pi _ { 1 } ^ { * } , \\cdot \\cdot \\cdot , \\Pi _ { n } ^ { * } = \\underset { \\Pi _ { 1 } , \\cdots , \\Pi _ { n } \\subset { \\cal X } } { \\arg \\operatorname* { m i n } } \\sum _ { i = 1 } ^ { n } - \\log S ^ { ( i ) } ( \\Pi _ { i } \\mid { \\mathcal { D } } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 330, + 536, + 666, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where 58 ${ \\mathcal { G } } = ( \\mathcal { V } , { \\mathcal { E } } ) , \\mathcal { V } = \\{ 1 , \\cdots , n \\} , { \\mathcal { E } } = \\{ ( j , i ) | j , i \\in \\{ 1 , \\cdots , n \\} , X _ { j } \\in \\Pi _ { i } \\} ,$ and $S _ { i } : \\Pi _ { i } \\mathbb { R }$ is a local score function corresponding to 59 $X _ { i }$ . The Bayesian Dirichlet equivalent uniform (BDeu) score 60 [Buntine, 1991] is one of the commonly used scores, ", + "bbox": [ + 147, + 587, + 825, + 633 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/a243805dd06eb5d447eb326ec23a57283f4c6a766c380f56caa35dd89527f897.jpg", + "text": "$$\nS _ { \\mathrm { B D e u } } ^ { ( i ) } ( \\Pi _ { i } \\mid \\mathcal { D } ) \\equiv \\prod _ { j = 1 } ^ { \\beta _ { i } } \\frac { \\Gamma ( \\alpha _ { i , j } ) } { \\Gamma ( N _ { i , j } + \\alpha _ { i , j } ) } \\prod _ { k = 1 } ^ { \\gamma _ { i } } \\frac { \\Gamma ( N _ { i , j , k } + \\alpha _ { i , j , k } ) } { \\Gamma ( \\alpha _ { i , j , k } ) } ,\n$$", + "text_format": "latex", + "bbox": [ + 295, + 636, + 699, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "61 ere $\\begin{array} { r } { N = \\sum _ { j = 1 } ^ { \\beta _ { i } } N _ { i , j } , N _ { i , j } = \\sum _ { k = 1 } ^ { \\gamma _ { i } } N _ { i , j , k } , \\alpha _ { i , j } = \\sum _ { k = 1 } ^ { \\gamma _ { i } } \\alpha _ { i , j , k } , \\beta _ { i } } \\end{array}$ is the numb f joint states of $\\Pi _ { i }$ is the number of states of $X _ { i }$ , $N _ { i , j , k }$ is the number of cases of the parent set $\\Pi _ { i }$ 63 and Xi in its k-th state, αi,j,k = αβ γ is the hyperparameter of the Dirichlet function, and $0 < \\alpha \\in \\mathbb { R }$ 64 is called equivalent sample size [Heckerman et al., 1995a]. ", + "bbox": [ + 147, + 686, + 825, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "65 2.2 Hamiltonian ", + "text_level": 1, + "bbox": [ + 153, + 765, + 302, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "66 The Hamiltonian, which is the objective function of an annealing processor, is a quadratic pseudo \n67 Boolean function, ", + "bbox": [ + 151, + 790, + 825, + 820 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/31597f6745cb0a1832de82ab45c13fd58581ee582bc67ef71d91ef2d73657958.jpg", + "text": "$$\nH ( \\pmb { \\sigma } ) = \\sum _ { i \\in \\mathcal { V } _ { \\mathrm { A P } } } h _ { i } \\sigma _ { i } + \\sum _ { ( i , j ) \\in \\mathcal { E } _ { \\mathrm { A P } } } J _ { i , j } \\sigma _ { i } \\sigma _ { j } ,\n$$", + "text_format": "latex", + "bbox": [ + 361, + 821, + 635, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "68 where ${ \\pmb \\sigma } = ( { \\boldsymbol \\sigma } _ { i } ) _ { i = 1 } ^ { | { \\boldsymbol \\nu } _ { \\mathrm { A P } } | } \\in \\mathbb { B } ^ { | { \\boldsymbol \\nu } _ { \\mathrm { A P } } | }$ , the biases $h _ { i } \\in \\mathbb { R }$ for all $i \\in \\mathcal { V } _ { \\mathrm { A P } }$ , the couplers $J _ { i , j } \\in \\mathbb { R }$ for all \n69 $( i , j ) \\in \\mathcal { E } _ { \\mathrm { A P } }$ , and the graph $\\mathcal { G } _ { \\mathrm { A P } } = ( \\nu _ { \\mathrm { A P } } , \\mathcal { E } _ { \\mathrm { A P } } )$ . Higher degree problems are reformed into quadratic \n70 ones using auxiliary variables. This reformulation is called quadratization. ", + "bbox": [ + 147, + 863, + 825, + 912 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "71 Definition 1. If a quadratic polynomial function $g ( \\pmb { v } , \\pmb { h } )$ is a quadratization of a pseudo-Boolean function 72 $f ( v )$ , then $f ( v ) = \\operatorname* { m i n } _ { h \\in \\mathbb { B } ^ { J } } g ( v , h )$ for all $\\pmb { v } \\in \\mathbb { B } ^ { I }$ . ", + "bbox": [ + 150, + 90, + 825, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "73 Anthony et al. 2016 proved that every pseudo-Boolean function of $I$ variables and of degree $K$ has \n74 a quadratization involving at most $\\dot { \\mathcal { O } } ( \\dot { I } ^ { \\frac { K } { 2 } } )$ auxiliary variables. In particular, at most $\\check { \\mathcal { O } ( 2 ^ { \\frac { I } { 2 } } ) }$ when \n75 $K = I$ . It is well known that every pseudo-Boolean function can be uniquely represented as a \n76 multilinear polynomial in its variables [Boros and Hamme, 2002]. ", + "bbox": [ + 147, + 132, + 825, + 194 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "77 2.3 Basic Conversion of Score-based Bayesian Network Structure Learning ", + "text_level": 1, + "bbox": [ + 151, + 208, + 710, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using 78 $n ( n - 1 )$ bits to encode the paths into $\\pmb { d } = ( ( d _ { j , i } ) _ { 1 \\leq j \\leq n , j \\neq i } ) _ { i = 1 } ^ { n } \\in \\mathbb { B } ^ { n ( n - 1 ) }$ $( d _ { j , i } = 1$ if 79 $X _ { j }$ is the parent of $X _ { i }$ , $d _ { j , i } = 0$ otherwise) and $\\binom { n } { 2 }$ bits to encode the topological orders into 80 $r = ( r _ { i , j } ) _ { 1 \\leq i < j \\leq n } \\in \\mathbb { B } ^ { \\binom { n } { 2 } } ( r _ { i , j } =$ if the order of $X _ { i }$ is higher than $X _ { j }$ , $r _ { i , j } = 1$ otherwise), it is 81 possible to represent eq. (1) on the Hamiltonian, ", + "bbox": [ + 145, + 234, + 825, + 301 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/c0da30d04c8a1fddaacfdffb5b8f06065a58a53e388bc6e3259c4a429cd99b07.jpg", + "text": "$$\nH _ { \\mathrm { t o t a l } } ( d , r ) \\equiv \\sum _ { i = 1 } ^ { n } H _ { \\mathrm { s c o r e } } ^ { ( i ) } ( d , i ) + H _ { \\mathrm { c y c l e } } ( d , r ) .\n$$", + "text_format": "latex", + "bbox": [ + 339, + 305, + 656, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "82 The states of $d _ { \\cdot , i }$ are mapped one-to-one to the states of $\\Pi _ { i }$ . Let $\\Pi _ { i } = \\pi ^ { ( i ) } ( d _ { \\cdot , i } )$ for all $1 \\leq i \\leq n$ . \n83 The local score of the Hamiltonian is ", + "bbox": [ + 148, + 353, + 826, + 385 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/bb09f48b7fec6858384bcbb6cd20d82476963e632bea6c504f451bfb26887634.jpg", + "text": "$$\nH _ { \\mathrm { s c o r e } } ^ { ( i ) } ( d _ { \\cdot , i } ) \\equiv - \\log S ^ { ( i ) } ( \\pi ^ { ( i ) } ( d _ { \\cdot , i } ) \\mid \\mathcal { D } ) + \\log S ^ { ( i ) } ( \\phi \\mid \\mathcal { D } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 302, + 388, + 694, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "84 for all $1 \\leq i \\leq n$ . The score function has a quadratization involving at most $\\mathcal { O } ( n 2 ^ { \\frac { n - 1 } { 2 } } )$ auxiliary \n85 variables. O’Gorman et al. 2014 added the maximum parent set size constraint to the Hamiltonian. \n86 In this case, the number of auxiliary variables is at most $\\mathcal { O } ( n ( n - 1 ) ^ { \\frac { m } { 2 } } )$ . The cycle constraint of the \n87 Hamiltonian consists of the topological order constraint and the consistency constraint, ", + "bbox": [ + 145, + 417, + 830, + 479 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/3e20f28dfe6c797545a4eb730ad0e4487247edad97daff25d4004ea43d97491f.jpg", + "text": "$$\nH _ { \\mathrm { c y c l e } } ( d , r ) \\equiv \\sum _ { 1 \\leq i < j < k \\leq n } \\delta _ { 1 } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) + \\sum _ { 1 \\leq i < j \\leq n } \\delta _ { 2 } ( d _ { i , j } r _ { i , j } + d _ { j , i } ( 1 - r _ { i , j } ) ) ,\n$$", + "text_format": "latex", + "bbox": [ + 212, + 483, + 782, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "88 where $R ( r _ { 1 } , r _ { 2 } , r _ { 3 } ) = r _ { 1 } r _ { 2 } ( 1 - r _ { 3 } ) + ( 1 - r _ { 1 } ) ( 1 - r _ { 2 } ) r _ { 3 }$ for all $r _ { 1 } , r _ { 2 } , r _ { 3 } \\in \\mathbb { B }$ . When the penalty \n89 coefficients $0 < \\delta _ { 1 } , \\delta _ { 2 } \\in \\mathbb { R }$ are sufficiently large, the DAG constraint is satisfied indirectly through \n90 the relationship of the paths $^ d$ and the topological order $\\pmb { r }$ . If it holds that ", + "bbox": [ + 145, + 525, + 826, + 571 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/b8b222fcc0e1e5345675d49bc19d7e6153f96aaed4ffc497120d754cd82a0829.jpg", + "text": "$$\n\\operatorname* { m a x } \\{ 0 , \\operatorname* { m a x } _ { \\substack { 1 \\leq j ^ { * } , i ^ { * } \\leq n d _ { \\cdot , j ^ { * } } \\in \\mathbb { B } ^ { n - 1 } } } ( H _ { \\mathrm { s c o r e } } ^ { ( i ^ { * } ) } ( d _ { \\cdot , i ^ { * } } ^ { ( j ^ { * } , i ^ { * } ) } ) - H _ { \\mathrm { s c o r e } } ^ { ( i ^ { * } ) } ( d _ { \\cdot , i ^ { * } } ) ) \\} < \\delta _ { 1 } < \\frac { \\delta _ { 2 } } { n - 2 } ,\n$$", + "text_format": "latex", + "bbox": [ + 233, + 575, + 763, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "91 then there is no cycle on the paths of the ground state, where ∗ ∗ ∗ ∗ ${ \\pmb d } ^ { ( j ^ { \\ast } , i ^ { \\ast } ) } = ( ( d _ { j , i } ) _ { 1 \\leq j \\leq n , j \\neq i } ^ { ( j ^ { \\ast } , i ^ { \\ast } ) } ) _ { i = 1 } ^ { n }$ \n92 $d _ { j , i } ^ { ( j ^ { * } , i ^ { * } ) } = 0$ if $( j , i ) = ( j ^ { * } , i ^ { * } )$ , $d _ { j , i } ^ { ( j ^ { * } , i ^ { * } ) } = d _ { j , i }$ otherwise. The computational cost to obtain the left \n93 side of eq. (7) is at most $\\mathcal { O } ( n ^ { m + 1 } )$ . In particular, at most $O ( n ^ { 2 } 2 ^ { n - 2 } )$ when $m = n - 1$ . ", + "bbox": [ + 147, + 631, + 825, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "94 3 Candidate Parent Set Decomposition ", + "text_level": 1, + "bbox": [ + 147, + 702, + 514, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parent set identification is a major technique to narrow the search space of structure optimization, based on the relationship between parent sets and local scores under the DAG constraints [de Campos and Ji, 2011, Correia et al., 2020]. The collection of candidate parent sets of a random variable $X _ { i }$ is $\\{ W \\subseteq { \\mathcal { X } } \\setminus \\{ X _ { i } \\} \\mid W ^ { \\prime } \\subset W \\Rightarrow S ^ { ( i ) } ( W ^ { \\prime } \\mid { \\mathcal { D } } ) < S ^ { ( i ) } ( W \\mid { \\mathcal { D } } ) \\}$ . To reduce the required bits of the score component of the Hamiltonian, we propose an efficient conversion method with the parent set identification. We directly encode the candidate parent sets instead of using the paths $^ d$ . ", + "bbox": [ + 173, + 734, + 826, + 827 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, we decompose the candidate parent sets $( W _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { i } }$ of each random variable into the form of Cartesian products as follows: ", + "bbox": [ + 165, + 832, + 825, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Decompose $( W _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { i } }$ into $( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { i } } , ( W _ { h , i } \\cap ( { \\mathcal { X } } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { i } } ,$ \n2. Remove duplicates in the elements of $( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { i } } , ( W _ { h , i } \\cap ( { \\mathcal { X } } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { i } } ,$ ", + "bbox": [ + 210, + 872, + 748, + 914 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/ab3b3cda9ea0934cb1baa7d042cb63f494af0851c4623ac066ff6b9967b3dbd2.jpg", + "text": "$$\n\\because ( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( W _ { h , i } \\cap ( \\mathcal { X } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } \\mathrm { ~ i n ~ } ( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } ,\n$$", + "text_format": "latex", + "bbox": [ + 267, + 87, + 694, + 109 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where 106 $Z _ { i } \\subseteq \\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i } , W _ { 0 , i } = U _ { 0 , i } = V _ { 0 , i } = \\phi , \\lambda _ { i } , \\lambda _ { 1 , i } , \\lambda _ { 2 , i } \\in \\mathbb { N } \\cup \\{ 0 \\}$ for all $1 \\leq i \\leq n$ . There 107 is a clear relationship, ", + "bbox": [ + 142, + 117, + 823, + 150 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/692265dd84ddbbb674f02f50a0e26baf045418a2d9859a4d3b09281bd474988a.jpg", + "text": "$$\n\\begin{array} { r } { \\{ W _ { 0 , i } , \\cdots , W _ { \\lambda _ { i } , i } \\} \\subseteq \\{ U \\cup V \\mid ( U , V ) \\in \\{ U _ { 0 , i } , \\cdots , U _ { \\lambda _ { 1 , i } , i } \\} \\times \\{ V _ { 0 , i } , \\cdots , V _ { \\lambda _ { 2 , i } , i } \\} \\} , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 220, + 155, + 776, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "108 for all $1 \\leq i \\leq n$ . Here, given that the Hamiltonian is a quadratic pseudo-Boolean function, we can represent the score against 109 fore, it is possible to encod110 $U _ { h , i } \\cup V _ { h ^ { \\prime } , i }$ by allocating ate parent sets $U _ { h , i } , V _ { h ^ { \\prime } , i }$ to two bits on thamiltonian using $( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } }$ 111 The number of required bits of the score component of the Hamiltonian is ${ \\textstyle \\sum _ { i = 1 } ^ { n } } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } )$ . ", + "bbox": [ + 140, + 178, + 828, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "112 Example 1. An example of the candidate parent sets in the form of Cartesian products as follows: ", + "bbox": [ + 138, + 246, + 818, + 262 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/177bc226d0abc53ce1d02c4e3167229d1c8c3d1a6f610731ec9795a250229c90.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathcal { X } = \\{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \\} , \\ Z _ { i } = \\{ X _ { 1 } , X _ { 2 } \\} , \\ \\lambda _ { i } = 5 , \\ \\lambda _ { 1 , i } = 2 , \\ \\lambda _ { 2 , i } = 1 } \\\\ & { ( W _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { i } } = ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \\} ) , } \\\\ & { ( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { i } } = ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} , \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} ) , } \\\\ & { ( W _ { h , i } \\cap ( \\mathcal { X } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { i } } = ( \\phi , \\phi , \\phi , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} ) , } \\\\ & { ( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } = ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} ) , \\ ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } = ( \\phi , \\{ X _ { 3 } , X _ { 4 } \\} ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 210, + 265, + 751, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We optimize 113 $Z _ { i } \\subseteq \\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i }$ to minimize $\\lambda _ { 1 , i } + \\lambda _ { 2 , i }$ . However, it is often infeasible to search all 114 elements of the power set $\\mathcal { P } ( \\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i } )$ . Therefore, we heuristically search 3 $Z _ { i }$ adding elements one by one, as algorithm 1. The computational cost is at most $1 \\leq i \\leq n$ ", + "bbox": [ + 142, + 382, + 825, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Greedy Candidate Parent Set Decomposition ", + "text_level": 1, + "bbox": [ + 173, + 445, + 558, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: Input: $( W _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { i } }$ Output: $Z$ Initialize: $\\lambda \\lambda _ { i } , Z ^ { \\prime } \\phi , Z \\phi$ . \n2: for $d = 1$ to $| \\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i } | - 1$ do \n3: for $X$ in $\\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i } \\setminus Z$ do \n4: if $\\lambda _ { 1 , i } + \\lambda _ { 2 , i } < \\lambda$ for $Z _ { i } = Z \\cup \\{ X \\}$ then $\\lambda \\lambda _ { 1 , i } + \\lambda _ { 2 , i } , Z ^ { \\prime } Z \\cup \\{ X \\} .$ . \n5: if $Z \\neq Z ^ { \\prime }$ then $Z Z ^ { \\prime }$ else break ", + "bbox": [ + 178, + 467, + 754, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "115 ", + "bbox": [ + 142, + 558, + 161, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "116 Example 2. An example of the bit reduction flow of algorithm 1 is as follows: ", + "bbox": [ + 142, + 570, + 689, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/edf133bebbed6e9f1548b0b895521a644d3238a24083d9d1c1cef44858e4d6e6.jpg", + "text": "$$\n\\begin{array} { r l } & { Z _ { i } = \\phi , \\lambda _ { 1 , i } = 0 , \\lambda _ { 2 , i } = 5 : ( \\phi ) \\times ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 3 } , X _ { 3 } \\} , } \\\\ & { Z _ { i } = \\{ X _ { 1 } \\} , \\lambda _ { 1 , i } = 1 , \\lambda _ { 2 , i } = 3 : ( \\phi , \\{ X _ { 1 } \\} ) \\times ( \\phi , \\{ X _ { 2 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 2 } , X _ { 3 } , X _ { 4 } \\} ) , } \\\\ & { Z _ { i } = \\{ X _ { 1 } , X _ { 2 } \\} , \\lambda _ { 1 , i } = 2 , \\lambda _ { 2 , i } = 1 : ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} ) \\times ( \\phi , \\{ X _ { 3 } , X _ { 4 } \\} ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 168, + 589, + 816, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "117 4 Efficient Conversion of Score-based Bayesian Network Structure Learning ", + "text_level": 1, + "bbox": [ + 142, + 662, + 825, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We make 118 $( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } }$ correspond to $( p _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( q _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } }$ one-to-one, where $p _ { h , i } , q _ { h ^ { \\prime } , i } \\in \\mathbb { B }$ 119 for all $0 \\leq h \\leq \\lambda _ { 1 , i } , 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 2 , i } , 1 \\leq i \\leq n$ . To identify the parent sets, we use the one-to-one 120 correspondence constraint that $\\begin{array} { r } { \\sum _ { h = 0 } ^ { \\lambda _ { 1 , i } } p _ { h , i } = \\sum _ { h = 0 } ^ { \\lambda _ { 2 , i } } q _ { h , i } = 1 } \\end{array}$ for all $1 \\leq i \\leq n$ . The Hamiltonian 121 consists of the score component, the one-to-one correspondence constraint, and the cycle constraint, ", + "bbox": [ + 142, + 695, + 826, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/059a499284c7d9b86d50bb3dab59523b76c867de03060aaededf191ad90c6143.jpg", + "text": "$$\nH _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) \\equiv \\sum _ { i = 1 } ^ { n } ( H _ { \\mathrm { s c o r e } } ^ { * ( i ) } ( p . , i , q . , i ) + H _ { \\mathrm { o n e } } ^ { * ( i ) } ( p . , i , q . , i ) ) + H _ { \\mathrm { c y c l e } } ^ { * } ( p , q , r ) ,\n$$", + "text_format": "latex", + "bbox": [ + 240, + 765, + 756, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where 122 $\\pmb { p } = ( ( p _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } ) _ { i = 1 } ^ { n } , \\pmb { q } = ( ( q _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } ) _ { i = 1 } ^ { n }$ . Under the one-to correspondence constraint, \n123 we can represent the paths among random variables indirectly using $\\mathbf { \\omega } _ { p , q }$ \n124 variables, ", + "bbox": [ + 140, + 813, + 825, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/6dd36b70ee16072c01d515a9805ef91032a01d1c4e329914aee71ae96585c6b2.jpg", + "text": "$$\nd _ { j , i } ^ { * } \\equiv \\sum _ { 1 \\leq h \\leq \\lambda _ { 1 , i } \\atop \\overline { { X } } _ { j } \\in U _ { h , i } } p _ { h , i } + \\sum _ { 1 \\leq h \\leq \\lambda _ { 2 , i } \\atop \\overline { { X } } _ { j } \\in V _ { h , i } } q _ { h , i } ,\n$$", + "text_format": "latex", + "bbox": [ + 377, + 863, + 619, + 910 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d4f7e8cc2ac82f681ab72206ecf688f4b23e1a397be97e5c0273bc55dfd11cd4.jpg", + "image_caption": [ + "Figure 1: An example of bit allocation for our conversion method. $n = 4 , \\lambda _ { 1 , 1 } = 3 , \\lambda _ { 2 , 1 } = 1 , \\lambda _ { 1 , 2 } =$ 1 $\\bar { \\lambda } _ { 2 , 2 } = 0 , \\lambda _ { 1 , 3 } = \\bar { 1 } , \\lambda _ { 2 , 3 } = 1 , \\lambda _ { 1 , 4 } = 0 , \\lambda _ { 2 , 4 } = 0 , U _ { 1 , 1 } = \\{ 2 , 3 \\} , U _ { 2 , 1 } = \\{ 3 \\} , U _ { 3 , 1 } = \\{ 2 \\} , V _ { 1 , 1 } = \\{ 3 \\} , U _ { 2 , 2 } = \\{ 3 \\} , U _ { 3 , 2 } = \\{ 3 \\} , U _ { 2 , 2 } = \\{ 3 \\} ,$ $\\{ 4 \\} , U _ { 1 , 2 } = \\{ 3 \\} , U _ { 1 , 3 } = \\{ 1 \\} , V _ { 1 , 3 } = \\{ 4 \\}$ . Circle : $\\mathbf { \\mu } _ { p , q }$ . Square $: \\textbf { { r } }$ . Red lines include in the score component of the Hamiltonian, a green line in the one-to-one correspondence constraint, and blue lines in the cycle constraint. " + ], + "image_footnote": [], + "bbox": [ + 202, + 109, + 794, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "125 for all $1 \\leq j , i \\leq n$ . Figure 1 is an example of bit allocation using our conversion method. The 126 number of bits required in our conversion method is $\\begin{array} { r } { \\sum _ { i = 1 } ^ { n } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } ) + \\binom { n } { 2 } } \\end{array}$ . Note that we do not directly encode127 $p _ { 0 , i } , q _ { 0 , i }$ on the Hamiltonian. ", + "bbox": [ + 143, + 518, + 826, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "128 Score Component. The local score component of the Hamiltonian is ", + "bbox": [ + 147, + 570, + 633, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/96ebf3a679e7b0d4a55385a55a460aa1d4d03d17436fd8704348dfc1b0a44867.jpg", + "text": "$$\nH _ { \\mathrm { s c o r r e } } ^ { * ( i ) } ( { \\pmb p } _ { . , i } , { \\pmb q } _ { . , i } ) \\equiv \\sum _ { h = 1 } ^ { \\lambda _ { 1 , i } } s _ { 1 , h , i } p _ { h , i } + \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } s _ { 2 , h , i } q _ { h , i } + \\sum _ { h = 1 } ^ { \\lambda _ { 1 , i } } \\sum _ { h ^ { \\prime } = 1 } ^ { \\lambda _ { 2 , i } } t _ { h , h ^ { \\prime } , i } p _ { h , i } q _ { h ^ { \\prime } , i } ,\n$$", + "text_format": "latex", + "bbox": [ + 246, + 590, + 750, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "129 for all $1 \\leq i \\leq n$ . We can get these coefficients by solving simultaneous equations under the \n130 one-to-one correspondence constraint, $s _ { 1 , h , i } = \\_ \\log S ^ { ( i ) } ( U _ { h , i } \\mid \\mathcal { D } ) + \\log S ^ { ( i ) } ( \\phi \\mid \\mathcal { D } ) , s _ { 2 , h , i } =$ \n131 $\\begin{array} { r } { - \\log S ^ { ( i ) } ( V _ { h , i } \\mid \\mathcal D ) + \\log S ^ { ( i ) } ( \\phi \\mid \\mathcal D ) , t _ { h , h ^ { \\prime } , i } = - \\log S ^ { ( i ) } ( U _ { h , i } \\cup V _ { h ^ { \\prime } , i } \\mid \\mathcal D ) + \\log S ^ { ( i ) } ( U _ { h , i } \\mid \\mathcal D ) + } \\end{array}$ \n132 $\\log S ^ { ( i ) } ( V _ { h ^ { \\prime } , i } \\mid \\mathcal { D } ) - \\log S ^ { ( i ) } ( \\phi \\mid \\mathcal { D } )$ . ", + "bbox": [ + 140, + 638, + 826, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "133 One-to-One element from 134 $( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } }$ nstraint. We penalize the connection among bits to select each , ", + "bbox": [ + 145, + 707, + 826, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a7f05f2001ab49132a03a2f6406d91157f12a44726e37d32f2512341bb23d1b3.jpg", + "text": "$$\nH _ { \\mathrm { o n e } } ^ { * ( i ) } ( p . , . , q . , i ) \\equiv \\sum _ { 1 \\leq h < h ^ { \\prime } \\leq \\lambda _ { 1 , i } } \\xi _ { 1 , i } p _ { h , i } p _ { h ^ { \\prime } , i } + \\sum _ { 1 \\leq h < h ^ { \\prime } \\leq \\lambda _ { 2 , i } } \\xi _ { 2 , i } q _ { h , i } q _ { h ^ { \\prime } , i } ,\n$$", + "text_format": "latex", + "bbox": [ + 261, + 743, + 733, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "135 136 $1 \\leq i \\leq n$ penalty coefficient is induced indirectly $0 < \\xi _ { 1 , i } , \\xi _ { 2 , i } \\in \\mathbb { R }$ . If $\\xi _ { 1 , i } , \\xi _ { 2 , i }$ is sufficient large, $\\begin{array} { r } { \\sum _ { h = 0 } ^ { \\lambda _ { 1 , i } } p _ { h , i } = \\sum _ { h = 0 } ^ { \\lambda _ { 2 , i } } q _ { h , i } = 1 } \\end{array}$ ", + "bbox": [ + 142, + 784, + 828, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "137 Cycle Constraint. Compared to eq. (6), the cycle constraint of the Hamiltonian is ", + "bbox": [ + 137, + 823, + 718, + 839 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/1f487b5e21ae0614865f7830a55d2e685e703b525cc02687348df63ee81a5b93.jpg", + "text": "$$\nH _ { \\mathrm { c y c l e } } ^ { * } ( p , q , r ) \\equiv \\sum _ { 1 \\leq i < j < k \\leq n } \\delta _ { 1 } ^ { * } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) + \\sum _ { 1 \\leq i < j \\leq n } \\delta _ { 2 } ^ { * } ( d _ { i , j } ^ { * } r _ { i , j } + d _ { j , i } ^ { * } ( 1 - r _ { i , j } ) ) ,\n$$", + "text_format": "latex", + "bbox": [ + 191, + 842, + 779, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the penalty coefficients 138 $0 < \\delta _ { 1 } ^ { * } , \\delta _ { 2 } ^ { * } \\in \\mathbb { R }$ . By setting $\\delta _ { 1 } ^ { * } , \\delta _ { 2 } ^ { * }$ appropriately, we can prevent the 139 cycle from occurring. ", + "bbox": [ + 137, + 881, + 828, + 912 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "141 We demonstrate the sufficient lower bounds of penalty coefficients. The basic idea is that we find the \n142 range of penalty coefficients so that the change in return value of the Hamiltonian is negative when \n143 the input state changes to the state we desire to induce. \n144 One-to-One Correspondence Constraint. We consider to decrease the value of $\\scriptstyle \\sum _ { h = 1 } ^ { \\lambda _ { 1 , i ^ { * } } } p _ { h , i ^ { * } }$ by \n145 146 $\\begin{array} { r } { p _ { h ^ { * } , i ^ { * } } = 1 , \\sum _ { h = 1 } ^ { \\lambda _ { 1 , i ^ { * } } } p _ { h , i ^ { * } } > 1 } \\end{array}$ , e $H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) -$ \n$\\begin{array} { r } { H _ { \\mathrm { t o t a l } } ^ { * } ( p ^ { ( h ^ { * } , i ^ { * } ) } , q , r ) \\geq \\xi _ { 1 , i ^ { * } } + s _ { 1 , h ^ { * } , i ^ { * } } + \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } . } \\end{array}$ $\\pmb { p } ^ { ( h ^ { * } , i ^ { * } ) } = ( ( p _ { h , i } ^ { ( h ^ { * } , i ^ { * } ) } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } ) _ { i = 1 } ^ { n }$ \n147 and p(h∗,i∗)h,i = 0 if (h, i) = (h∗, i∗), p(h∗,h,i otherwise. Considering the case where $\\pmb { p }$ and $\\pmb q$ \n148 are swapped in the above, if $\\xi _ { 1 , i } , \\xi _ { 2 , i }$ satisfy that ", + "bbox": [ + 142, + 119, + 826, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 140, + 172, + 825, + 263 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/ff3f4ba76617ec792651b8309ec79b6300bf8cb7f493172eb5077060b8e53503.jpg", + "text": "$$\n\\begin{array} { r l r } & { } & { \\displaystyle \\operatorname* { m a x } _ { 0 \\le h \\le \\lambda _ { 1 , i } } \\bigl ( - s _ { 1 , h , i } - \\sum _ { h ^ { \\prime } = 1 } ^ { \\lambda _ { 2 , i } } \\operatorname* { m i n } \\bigl \\{ 0 , t _ { h , h ^ { \\prime } , i } \\bigr \\} \\bigr ) \\bigr ) < \\xi _ { 1 , i } , } \\\\ & { } & { \\displaystyle \\operatorname* { m a x } _ { 0 \\le h \\le \\lambda _ { 2 , i } } \\bigl ( - s _ { 2 , h , i } - \\sum _ { h ^ { \\prime } = 1 } ^ { \\lambda _ { 1 , i } } \\operatorname* { m i n } \\bigl \\{ 0 , t _ { h ^ { \\prime } , h , i } \\bigr \\} \\bigr ) \\bigr ) < \\xi _ { 2 , i } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 333, + 265, + 661, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "149 for all $1 \\leq i \\leq n$ , then the grand state does not violate the one-to-one correspondence constraint. \n150 The computational cost to obtain the left side of eq. (14) and eq. (15) is at most $\\mathcal { O } ( \\lambda _ { 1 , i } \\lambda _ { 2 , i } )$ for all \n151 $1 \\leq i \\leq n$ . \n152 Cycle Constraint. We consider four patterns of $( r _ { i ^ { * } , j ^ { * } } , d _ { j ^ { * } , i ^ { * } } ^ { * } , d _ { i ^ { * } , j ^ { * } } ^ { * } )$ violating the consistency \n153 constraint. It is assumed that $X _ { j ^ { * } } ~ \\in ~ U _ { h ^ { * } , i ^ { * } } , X _ { j ^ { * } } ~ \\notin ~ U _ { h ^ { * * } , i ^ { * } } ~ \\subset ~ U _ { h ^ { * } , i ^ { * } } , p _ { h ^ { * } , i ^ { * } } ~ = ~ 1 , p _ { h ^ { * * } , i ^ { * } } ~ = ~ 0$ . \n154 155 In thewhere $r ^ { ( i ^ { * } , j ^ { * } ) } = ( r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } ) _ { 1 \\leq i < j \\leq n }$ $( 0 , 1 , 0 )$ t , $H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) - H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ^ { ( i ^ { * } , j ^ { * } ) } ) \\geq \\delta _ { 2 } ^ { * } - ( n - 2 ) \\delta _ { 1 } ^ { * }$ $r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = 1 - r _ { i , j }$ $( i , j ) = ( i ^ { * } , j ^ { * } ) , r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = r _ { i , j }$ \n156 otherwise. Similarly, it is possible to consider the case of $( 1 , 0 , 1 )$ . In the case of $( 0 , 1 , 1 )$ , \n157 it holds that $\\begin{array} { r } { H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) - H _ { \\mathrm { t o t a l } } ^ { * } ( p ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } , q , r ) \\ge \\delta _ { 2 } ^ { * } + s _ { 1 , h ^ { * } , i ^ { * } } + \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } - \\frac { 1 - \\lambda _ { 1 } } { \\lambda _ { 2 } } | \\nabla p _ { h } ( q _ { h } ^ { * } , h ^ { * } , i ^ { * } ) | _ { \\mathbb { H } } ( q _ { h } ^ { * } , h ^ { * } , i ^ { * } ) , } \\end{array}$ \n158 $\\begin{array} { r } { s _ { 1 , h ^ { * * } , i ^ { * } } - \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * * } , h , i ^ { * } } q _ { h , i ^ { * } } } \\end{array}$ , where $\\pmb { p } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = ( ( p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } ) _ { i = 1 } ^ { n }$ h=, and $p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = 0$ \n159 if $( h , i ) \\ = \\ ( h ^ { * } , i ^ { * } ) , p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } \\ = \\ 1$ p(h∗,h∗∗,i∗)h,i = 1 if (h, i) = (h∗∗, i∗), p(h∗,h,i $p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = p _ { h , i }$ otherwise. Sim \n160 ilarly, it is possible to consider the case of $( 1 , 1 , 1 )$ . These results suggest the relationship of \n161 $\\delta _ { 1 } ^ { * } , \\delta _ { 2 } ^ { * }$ to induce the consistency constraint. Here, based on theorem 1, we consider a strategy \n162 to repeat picking up one element from $\\mathbfit { \\Delta } \\mathbf { r }$ and switching its value until $H _ { \\mathrm { t r a n s } } ( r ) = 0$ . It is as \n163 164 sumed that $\\begin{array} { r } { T _ { \\mathrm { t o t a l } } ^ { * } \\big ( p ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } , q , r ^ { ( i ^ { * } , j ^ { * } ) } \\big ) \\geq \\delta _ { 1 } ^ { * } + s _ { 1 , h ^ { * } , i ^ { * } } + \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } - s _ { 1 , h ^ { * * } , i ^ { * } } - \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * * } , h , i ^ { * } } q _ { h , i ^ { * } } . } \\end{array}$ $H _ { \\mathrm { t r a n s } } ( \\mathbf { \\bar { r } } ) > H _ { \\mathrm { t r a n s } } ( { r ^ { ( i ^ { * } , j ^ { * } ) } } )$ . In the case of $( 1 , 1 , 0 )$ , it holds that $H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) \\sim$ \n165 Similarly, it is possible to consider the case of $( 0 , 0 , 1 )$ . In the case of $( 1 , 0 , 0 )$ or $( 0 , 0 , 0 )$ , it holds \n166 that $\\bar { H _ { \\mathrm { t o t a l } } ^ { * } } ( p , \\bar { q } , r ) - \\bar { H _ { \\mathrm { t o t a l } } ^ { * } } ( p , q , r ^ { ( i ^ { * } , j ^ { * } ) } ) \\geq \\delta _ { 1 } ^ { * }$ . These results suggest the lower bound of ${ { \\delta } _ { 1 } ^ { * } }$ to \n167 induce the topological order constraint. Considering the case where $\\pmb { p }$ and $\\pmb q$ are swapped in the \n168 above, if $\\delta _ { 1 } ^ { * } , \\delta _ { 2 } ^ { * }$ satisfy that ", + "bbox": [ + 140, + 356, + 825, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 138, + 406, + 826, + 681 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/f331a970455c12a4a94d5d27fe6e4ec705cbe61b8be399900e2a744f7a5f7faf.jpg", + "text": "$$\n\\begin{array} { c } { \\displaystyle \\operatorname* { m a x } _ { 1 \\leq i \\leq n } \\operatorname* { m a x } \\{ \\eta _ { 1 , i } , \\eta _ { 2 , i } \\} < \\delta _ { 1 } ^ { * } < \\displaystyle \\frac { \\delta _ { 2 } ^ { * } } { n - 2 } , } \\\\ { \\displaystyle \\operatorname* { m a x } _ { 1 \\leq j \\leq n } \\displaystyle \\operatorname* { m a x } _ { 1 \\leq i \\leq 1 _ { 1 , i } } \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 1 , i } } \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 2 , i } } ( - s _ { 1 , h , i } - t _ { h , h ^ { \\prime \\prime } , i } + s _ { 1 , h ^ { \\prime } , i } + t _ { h ^ { \\prime } , h ^ { \\prime \\prime } , i } ) , } \\\\ { \\displaystyle \\eta _ { 1 , i } \\equiv \\displaystyle \\operatorname* { m a x } _ { 1 \\leq j \\leq n } \\displaystyle \\frac { \\operatorname* { m a x } _ { i } } { X _ { j } \\in U _ { h , i } } \\operatorname* { m a x } _ { X _ { j } \\in U _ { h , i } } \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime \\prime } \\leq \\lambda _ { 1 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \\prime \\prime } , h , i } + s _ { 2 , h ^ { \\prime } , i } + t _ { h ^ { \\prime \\prime } , h ^ { \\prime \\prime } , i } ) , } \\\\ { \\displaystyle \\eta _ { 2 , i } \\equiv \\displaystyle \\operatorname* { m a x } _ { 1 \\leq j \\leq n } \\displaystyle \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 2 , i } } \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 2 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \\prime \\prime } , h , i } + s _ { 2 , h ^ { \\prime } , i } + t _ { h ^ { \\prime \\prime } , h ^ { \\prime } , i } ) , } \\\\ { \\displaystyle \\sum _ { k \\geq \\ell \\leq n } \\displaystyle \\sum _ { X _ { j } \\in V _ { h , i } } \\sum _ { j \\in V _ { h ^ { \\prime } } , i \\subset V _ { h , i } } ( \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime \\prime } \\leq \\lambda _ { 1 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \\prime \\prime } , h , i } + s _ { 2 , h ^ { \\prime } , i } + t _ { h ^ { \\prime \\prime } , h ^ { \\prime \\prime } , i } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 200, + 681, + 766, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "169 for $n \\geq 3$ , then the grand state does not violate the cycle constraint under the one-to-one cor \n170 respondence constraint. The computational cost to obtain the left side of eq. (16) is at most \n171 $\\begin{array} { r l } { { \\mathcal { O } ( \\sum _ { i = 1 } ^ { n } n \\lambda _ { 1 , i } \\lambda _ { 2 , i } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } ) ) } \\quad } & { { } } \\end{array}$ . \n172 Theorem 1. If it holds that $\\begin{array} { r } { H _ { \\mathrm { t r a n s } } ( { \\pmb r } ) \\equiv \\sum _ { 1 \\leq i < j < k \\leq n } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) > 0 } \\end{array}$ , then there exists \n173 at least one index pair $1 \\leq i ^ { * } < j ^ { * } \\leq n$ which satisfy $H _ { \\mathrm { t r a n s } } ( r ) > H _ { \\mathrm { t r a n s } } ( r ^ { ( i ^ { * } , j ^ { * } ) } )$ , where \n174 r(i∗,j∗) = (r(i∗,i,j $R ( r _ { 1 } , r _ { 2 } , r _ { 3 } ) = r _ { 1 } r _ { 2 } ( 1 - r _ { 3 } ) + ( 1 - r _ { 1 } ) ( 1 - r _ { 2 } ) r _ { 3 }$ $r ^ { ( i ^ { * } , j ^ { * } ) } = ( r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } ) _ { 1 \\leq i < j \\leq n }$ $r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = 1 - r _ { i , j } \\ i f ( i , j ) = ( i ^ { * } , j ^ { * } )$ $r _ { 1 } , r _ { 2 } , r _ { 3 } \\in \\mathbb { B }$ ,, $r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = r _ { i , j }$ $r = ( r _ { i , j } ) _ { 1 \\leq i < j \\leq n } \\in \\mathbb { B } ^ { \\binom { n } { 2 } }$ \n176 Proof. It does not lose the generality by considering the case of $( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) = ( 1 , 1 , 0 )$ . Here, it \n177 holds that $R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( 1 - r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) + R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( r _ { 1 , 2 } , 1 - r _ { 2 , 3 } , r _ { 1 , 3 } ) + R ( r _ { 1 , 2 } , 1 - r _ { 2 , 3 } , r _ { 1 , 3 } )$ \n178 $R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( r _ { 1 , 2 } , r _ { 2 , 3 } , 1 - r _ { 1 , 3 } ) = 3 .$ . Additionally, it holds that $R ( r _ { 1 , 2 } , r _ { 2 , i } , r _ { 1 , i } ) - R ( 1 -$ \n179 $r _ { 1 , 2 } , r _ { 2 , i } , r _ { 1 , i } ) + R ( r _ { 2 , 3 } , r _ { 3 , i } , r _ { 2 , i } ) - R ( 1 - r _ { 2 , 3 } , r _ { 3 , i } )$ ${ } _ { 3 } , r _ { 3 , i } , r _ { 2 , i } ) + R ( r _ { 1 , 3 } , r _ { 3 , i } , r _ { 1 , i } ) - R ( 1 - r _ { 1 , 3 } , r _ { 3 , i } , r _ { 1 } )$ ,i) = \n180 0 for all $3 < i$ . Therefore, it holds that $H _ { \\mathrm { t r a n s } } ( \\pmb { r } ) - H _ { \\mathrm { t r a n s } } ( \\pmb { r } ^ { ( 1 , 2 ) } ) + H _ { \\mathrm { t r a n s } } ( \\pmb { r } ) - H _ { \\mathrm { t r a n s } } ( \\pmb { r } ^ { ( 2 , 3 ) } ) +$ \n181 $H _ { \\mathrm { t r a n s } } ( \\pmb { r } ) - H _ { \\mathrm { t r a n s } } ( \\pmb { r } ^ { ( 1 , 3 ) } ) = 3 .$ . From this result, it holds that $H _ { \\mathrm { t r a n s } } ( \\pmb { r } ) - H _ { \\mathrm { t r a n s } } ( \\pmb { r } ^ { ( i ^ { * } , j ^ { * } ) } ) > 0$ for \n182 at least one index pair $( i ^ { * } , j ^ { * } ) \\in \\{ ( 1 , 2 ) , ( 2 , 3 ) , ( 1 , 3 ) \\}$ . □ ", + "bbox": [ + 143, + 794, + 825, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 140, + 842, + 825, + 915 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a7c9856d1d949d393dd0eb80d65770a55140d7b9782ff083a2109008e080a453.jpg", + "table_caption": [ + "Table 1: The benchmark networks from Bayesian network repository. ", + "Algorithm 2 Greedy Candidate Parent Set Identification " + ], + "table_footnote": [ + "The average for 10 simulated datasets. " + ], + "table_body": "
Namenm∑=1 |IIil n∑=1 βi( -1)Ωi=1X n *
N=100N= 1000N = 10000
insurance273529843538834036
water3256610083165216735
alarm37446509182922729081
barley484841140051813101552
hailfinder5646626561446924277
hepar2706123145348376654782
", + "bbox": [ + 178, + 112, + 816, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 140, + 284, + 826, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "183 6 Experimental Results ", + "text_level": 1, + "bbox": [ + 148, + 409, + 385, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the performance of our approach, we use 10 simulated datasets for each instance size $N = 1 0 0 , 1 0 0 0 , 1 0 0 0 0$ and each benchmark network. The benchmark networks are discrete networks from Bayesian network repository 1. The score function is the BDeu score with $\\alpha = 1$ . It is often infeasible to identify exact candidate parent sets by searching the power set $\\mathcal { P } ( \\mathcal { X } \\backslash \\{ X _ { i } \\} )$ in a realistic timeframe. We use the candidate parent sets from algorithm 2. Note that the candidate parent sets depend on the heuristic search algorithms, but we do not focus on their performance in this study. Table 1 displays the information of benchmark networks. The code to replicate each experiment in this paper is available 2. ", + "bbox": [ + 161, + 440, + 825, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1: Input: $\\mathcal { D } , i , m$ Output: $\\mathcal { L }$ Initialize: ${ \\mathcal { L } } \\gets \\{ \\phi \\} , { \\mathcal { L } } ^ { \\prime } \\gets \\{ \\phi \\} , { \\mathcal { L } } ^ { \\prime \\prime } \\gets \\phi$ \n2: for $d = 1$ to $m$ do \n3: for $W$ in $\\mathcal { L } ^ { \\prime }$ do \n4: for $X$ in $\\chi \\setminus \\{ X _ { i } \\} \\setminus W$ do \n5: if $S _ { i } ( W ^ { \\prime } \\mid \\mathcal { D } ) < S _ { i } ( W \\cup \\{ X \\} \\mid \\mathcal { D } )$ for all $W ^ { \\prime } \\subset W \\cup \\{ X \\} , W ^ { \\prime } \\in { \\mathcal { L } }$ then \n6: ${ \\mathcal { L } } ^ { \\prime \\prime } \\gets { \\mathcal { L } } ^ { \\prime \\prime } \\cup \\{ W \\cup \\{ X \\} \\}$ . \n7: if ${ \\mathcal { L } } ^ { \\prime \\prime } \\neq \\phi$ then $\\mathcal { L } \\gets \\mathcal { L } \\cup \\mathcal { L } ^ { \\prime \\prime } , \\mathcal { L } ^ { \\prime } \\mathcal { L } ^ { \\prime \\prime } , \\mathcal { L } ^ { \\prime \\prime } \\phi$ else break \n8: for $W$ in $\\mathcal { L }$ do \n9: if there exist $W ^ { \\prime } \\subset W$ that satisfies $S _ { i } ( W \\mid \\mathcal { D } ) \\le S _ { i } ( W ^ { \\prime } \\mid \\mathcal { D } )$ then ${ \\mathcal { L } } \\gets { \\mathcal { L } } \\setminus \\{ W \\}$ . ", + "bbox": [ + 179, + 594, + 769, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "192 6.1 Number of Required Bits for Score Component ", + "text_level": 1, + "bbox": [ + 142, + 756, + 542, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In comparison to the existing method [O’Gorman et al., 2014], we reduce the number of required bits for the score component by encoding the candidate parent sets directly. While $\\textstyle \\sum _ { i = 1 } ^ { n } \\lambda _ { i }$ candidate parent sets is encoded in our approach, $n ( n - 1 )$ paths plus at most $O ( n ( n - 1 ) ^ { \\frac { m } { 2 } } )$ auxiliary variables for $m > 2$ in the existing method. The left side of table 2 shows the reduction rate of the number of required bits for the score component. Moreover, we reduce the number of required bits for the score component to ${ \\textstyle \\sum _ { i = 1 } ^ { n } } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } )$ by decomposing the candidate parent sets in the form of Cartesian products. The right side of table 2 shows that algorithm 1 reduces the number of required bits for the score component although there is some variation among the networks. ", + "bbox": [ + 173, + 782, + 825, + 875 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9bd062d18b3ef55c8e5bb42eb932e0f079596c4f7adfd8ed361ffebe48d4050a.jpg", + "table_caption": [ + "Table 2: The reduction rate of the number of required bits for score component. " + ], + "table_footnote": [ + "\\* The average ratio for 10 simulated datasets. " + ], + "table_body": "
Name£ λi/n(n-1)²Σ=1 （1,i+λ2,i）/∑=1 1
N=100N=1000N = 10000N = 100N= 1000N = 10000
insurance0.098730.246771.127420.613670.472850.32476
Water0.000970.001260.004290.726800.705880.44014
alarm0.038140.047380.189380.453320.355370.21617
barley0.001710.002920.014640.767170.755380.54149
hailfinder0.000850.004090.025250.827730.601780.33365
hepar20.000210.000030.000210.496940.633460.31284
", + "bbox": [ + 187, + 112, + 805, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0e2a828933a4eae3fb9f4dd942541213267e723378a120f59ccbd2c808927665.jpg", + "table_caption": [ + "Table 3: The number of required bits for fully coupled and nearest neighbor annealing processors. " + ], + "table_footnote": [ + "\\* The average ratio for 10 simulated datasets. " + ], + "table_body": "
Name∑i=1(1,i+ 入2,i）+(2） *∑=1(入1,i + λ2,i)(入1,i +λ2,i + 1) + (2)
N= 100N =1000N = 10000N= 100N= 1000N = 10000
insurance56676716613375902385482
water613648820143417205881
alarm1489147226283676127985169004
barley1247136219686758344624796
hailfinder1659195729672212708480578
hepar24777283639104499169164136939
", + "bbox": [ + 174, + 297, + 825, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 166, + 474, + 821, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2 Selection of Annealing Processor ", + "text_level": 1, + "bbox": [ + 171, + 522, + 441, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the following discussion, the Fujitsu digital annealer is suitable for our approach from the viewpoint of bit capacity. ", + "bbox": [ + 176, + 550, + 823, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fully Connected Type. To the best of our knowledge, the bit capacity of the Fujitsu digital annealer is the largest in fully coupled annealing processors. The second generation Fujitsu digital annealer can deal with problems on a scale of 8192 bits [Matsubara et al., 2020]. The left side of table 3 shows that it is possible to encode all the logical conversion results for benchmark networks to the circuit of the digital annealer within bit capacity. ", + "bbox": [ + 173, + 587, + 825, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Nearest Neighbor Type. The number of additional bits required for minor embedding depends on the design of the hardware graphs. Oku et al. 2019 proposed a heuristic minor embedding algorithm for the Hitachi CMOS annealing machine [Masanao et al., 2010]. Using this algorithm, the number of required physical spins when embedding a fully connected graph is $I ^ { 2 } + I$ for $I$ variables. The conversion method proposed in this study has $n$ local fully connected graphs on $\\mathbf { \\Delta } _ { p , q }$ . Therefore, the number of required physical spins must be at least $\\begin{array} { r } { \\sum _ { i = 1 } ^ { n } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } ) ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } + 1 ) + \\binom { n } { 2 } } \\end{array}$ . From the right side of table 3, it is currently infeasible to encode logical conversion results for at least some networks to the circuit of CMOS annealing machine within its 102400 nodes [Sugie et al., 2021]. As far as we know, the bit capacity of the Hitachi CMOS annealing machine is the largest in nearest neighbor annealing processors. ", + "bbox": [ + 173, + 670, + 825, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3 Score Maximization ", + "text_level": 1, + "bbox": [ + 171, + 839, + 352, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate the performance of Fujitsu digital annealer for score-based Bayesian network structure learning using the conversion results of $N = 1 0 0 0 0$ simulated datasets. The running time for each simulated dataset is 6000 [s]. ", + "bbox": [ + 161, + 867, + 823, + 911 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b44701b9cc855182e35704f11ec0fbc3d89db610c0991e089d4b455ff631577.jpg", + "image_caption": [ + "Figure 2: Results of score maximization by the baseline algorithms. For each simulated dataset and each baseline algorithm, we normalized $\\begin{array} { r } { \\sum _ { i = 1 } ^ { \\bar { n } } ( \\log S ^ { ( i ) } ( \\Pi _ { i } \\vert \\mathcal { \\bar { D } } ) - \\log S ^ { ( i ) } ( \\phi \\vert \\mathcal { D } ) ) } \\end{array}$ by dividing it by the corresponding value of the Fujitsu digital annealer. In this experiment, we used the second-generation Fujitsu digital annealer. SA : simulated annealing, OBS : ordering-based search, ASOBS : acyclic selection ordering-based search. " + ], + "image_footnote": [], + "bbox": [ + 176, + 93, + 803, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare the results obtained by the digital annealer with those of three heuristic algorithms. One algorithm is the simulated annealing algorithm [Heckerman et al., 1995b] with a QUBO same as the one encoded into the digital annealer. Other algorithms are the ordering space search algorithms, i.e., ordering-based search and acyclic selection ordering-based search. For a fair comparison, the running time of the simulated annealing algorithm for each simulated dataset is 6000 [s] and that of the ordering space search algorithms is 6000 [s] plus the running time of algorithm 1. The computing environment is Microsoft Windows 10 Pro, 3.6 GHz Intel Core i9 processor, and 64 GB memory. ", + "bbox": [ + 171, + 462, + 825, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Result. Figure 2 shows that the digital annealer is better than all the baselines for all the simulated datasets from all the benchmark networks. ", + "bbox": [ + 173, + 589, + 823, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7 Conclusion ", + "text_level": 1, + "bbox": [ + 171, + 638, + 299, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We proposed a novel approach of converting a score-based Bayesian network structure learning into QUBO. The essence of this approach lies in reducing the number of required bits through the advanced identification of candidate parent sets and their representation as Cartesian products. The Fujitsu digital annealer with our conversion method improved the BDeu score for 27 to 70 variables benchmark networks over existing methods. The bit capacity limitation of annealing processor is being relaxed rapidly 3. Though our approach is still a disadvantage for larger-scale networks, we expect that our proposed algorithms will be effectively applied to larger-scale score-based Bayesian network structure learning in the near future. ", + "bbox": [ + 173, + 671, + 825, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Potential Negative Societal Impacts. The development of annealing processor technology could have an impact on various industry fields. However, the number of companies that have commercialized the API usage of annealing processors is still small. Therefore, there is a concern that the market of annealing processors will not work well and the disparities among stakeholders will be widen. Researchers are required to properly evaluate the value of technology and communicate it to the business side. ", + "bbox": [ + 174, + 799, + 825, + 890 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "248 References \n249 Judea Pearl, editor. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, 1988. Gregory F. Cooper and Edward Herskovits. A bayesian method for the induction of probabilistic networks from data. Journal of Machine Learning, 9(4), 1992. Robert G. Cowell. Conditions under which conditional independence and scoring methods lead to identical selection of bayesian network models. In Jack Breese and Daphne Koller, editors, Proceedings of the 17th Conference on Uncertainty in Artificial Intelligence, pages 91–97. Morgan Kaufmann Publishers, 2001. David M. Chickering, David Heckerman, and Christopher Meek. Large-sample learning of bayesian networks is np-hard. Journal of Machine Learning Research, 20:1287–1330, 2004. Marc Teyssier and Daphne Koller. Ordering-based search: A simple and effective algorithm for learning bayesian networks. In Proceedings of the 21th Conference on Uncertainty in Artificial Intelligence, pages 584–590, 2005. Mauro Scanagatta, Cassio Polpo de Campos, Giorgio Corani, and Marco Zaffalon. Learning bayesian networks with thousands of variables. In Proceedings of the 28th International Conference on Neural Information Processing Systems, pages 1864–1872, 2015. Hidenori Gyoten, Masayuki Hiromoto, and Takashi Sato. Area efficient annealing processor for ising model without random number generator. IEICE Transactions on Information and Systems, E101.D(2):314–323, 2018. Kasho Yamamoto. Research on Annealing Processors for Large-Scale Combinatorial Optimization Problems. PhD thesis, Graduate School of Information Science and Technology Hokkaido University, 2020. Vicky Choi. Minor-embedding in adiabatic quantum computation: I. the parameter setting problem. Quantum Information Processing, 7:193–209, 2008. Vicky Choi. Minor-embedding in adiabatic quantum computation: Ii. minor-universal graph design. Quantum Information Processing, 10:343–352, 2010. Bryan A. O’Gorman, Alejandro Perdomo-Ortiz, Ryan Babbush, Alan Aspuru-Guzik, and Vadim Smelyanskiy. Bayesian network structure learning using quantum annealing. The European Physical Journal Special Topics, 225(1), 2014. \n276 Endre Boros and Aritanan Gruber. On quadratization of pseudo-boolean functions. CoRR, abs/1404.6538, 2014. \n277 Maliheh Aramon, Gili Rosenberg, Elisabetta Valiante, Toshiyuki Miyazawa, Hirotaka Tamura, and Helmut G. Katzgraber. Physics-inspired optimization for quadratic unconstrained problems using a digital annealer. Frontiers in Physics, 7(48), 2019. Wray Buntine. Theory refinement of bayesian networks. In Proceedings of the 7th Conference on Uncertainty in Artificial Intelligence, pages 52–60, 1991. David Heckerman, Dan Geiger, and David M. Chickering. Learning bayesian networks: The combination of knowledge and statistical data. Journal of Machine Learning, 20(3):197–243, 1995a. \n284 Martin Anthony, Endre Boros, Yves Crama, and Aritanan Gruber. Quadratic reformulations of nonlinear binary optimization problems. Mathematical Programming, 162(1):115–144, 2016. Endre Boros and Peter L. Hamme. Pseudo-boolean optimization. Journal of Discrete Applied Mathematics, 123: 155–225, 2002. Cassio P. de Campos and Qiang Ji. Efficient structure learning of bayesian networks using constraints. Journal of Machine Learning Research, 12:663–689, 2011. Alvaro H. C. Correia, James Cussens, and Cassio de Campos. On pruning for score-based bayesian network structure learning. Journal of Machine Learning Research, 108:2709–2718, 2020. Satoshi Matsubara, Motomu Takatsu, Toshiyuki Miyazawa, Takayuki Shibasaki, Yasuhiro Watanabe, Kazuya Takemoto, and Hirotaka Tamura. Digital annealer for high-speed solving of combinatorial optimization problems and its applications. In Proceedings of the 25th Asia and South Pacific Design Automation Conference, pages 667–672, 2020. \n296 David Eppstein. Finding large clique minors is hard. Graph Algorithms and Applications, 13(2):197–204, 2009. \nDaisuke Oku, Kotaro Terada, Masato Hayashi, Masanao Yamaoka, Shu Tanaka, and Nozomu Togawa. A fully-connected ising model embedding method and its evaluation for cmos annealing machines. IEICE Transactions on Information and Systems, E102-D(9):1696–1706, 2019. \nYamaoka Masanao, Yoshimura Chihiro, Hayashi Masato, Okuyama Takuya, Aoki Hidetaka, and Mizuno Hiroyuki. A 20k-spin ising chip to solve combinatorial optimization problems with cmos annealing. IEEE Journal of Solid-State Circuits, 51(1), 2010. \nYuya Sugie, Yuki Yoshida, Normann Mertig, Takashi Takemoto, Hiroshi Teramoto, Atsuyoshi Nakamura, Ichigaku Takigawa, Shin ichi Minato, Masanao Yamaoka, and Tamiki Komatsuzaki. Minor-embedding heuristics for large-scale annealing processors with sparse hardware graphs of up to 102,400 nodes. Soft Computing, 2021. \nDaphne Koller and Nir Friedman, editors. Probabilistic Graphical Models: Principles and Techniques. The MIT Press, Cambridge, Massachusetts, 2009. \nDavid Heckerman, Dan Geiger, and David M. Chickering. Learning bayesian networks: Search methods and experimental results. In Preliminary Papers of the 5th International Workshop on Artificial, pages 112–128, 1995b. ", + "bbox": [ + 145, + 56, + 830, + 912 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 148, + 90, + 826, + 323 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Checklist ", + "text_level": 1, + "bbox": [ + 143, + 347, + 254, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "1. For all authors... ", + "bbox": [ + 214, + 372, + 330, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] See section 7. \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] See section 7. \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ", + "bbox": [ + 240, + 390, + 825, + 462 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "2. If you are including theoretical results... ", + "bbox": [ + 214, + 465, + 467, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] See section 4. \n(b) Did you include complete proofs of all theoretical results? [Yes] See section 5. ", + "bbox": [ + 238, + 483, + 769, + 512 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "3. If you ran experiments... ", + "bbox": [ + 214, + 516, + 377, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] in the supplemental material \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See section 6. \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] See fig. 2. \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See section 6.3. ", + "bbox": [ + 240, + 534, + 825, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ", + "bbox": [ + 210, + 647, + 767, + 660 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(a) If your work uses existing assets, did you cite the creators? [Yes] in the supplemental material \n(b) Did you mention the license of the assets? [N/A] \n(c) Did you include any new assets either in the supplemental material or as a URL? [N/A] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] ", + "bbox": [ + 240, + 665, + 825, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "If you used crowdsourcing or conducted research with human subjects... ", + "bbox": [ + 227, + 767, + 661, + 781 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ", + "bbox": [ + 240, + 785, + 825, + 867 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/parse/train/HXjt-kRBzvu/HXjt-kRBzvu_model.json b/parse/train/HXjt-kRBzvu/HXjt-kRBzvu_model.json new file mode 100644 index 0000000000000000000000000000000000000000..29a870f31cf8d13429258174a94ccf0b2526e846 --- /dev/null +++ b/parse/train/HXjt-kRBzvu/HXjt-kRBzvu_model.json @@ -0,0 +1,22761 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 254, + 777, + 1305, + 777, + 1305, + 1211, + 254, + 1211 + ], + "score": 0.972 + }, + { + "category_id": 0, + "poly": [ + 455, + 271, + 1246, + 271, + 1246, + 382, + 455, + 382 + ], + "score": 0.954 + }, + { + "category_id": 0, + "poly": [ + 787, + 703, + 913, + 703, + 913, + 739, + 787, + 739 + ], + "score": 0.896 + }, + { + "category_id": 1, + "poly": [ + 720, + 499, + 987, + 499, + 987, + 620, + 720, + 620 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 253, + 1270, + 531, + 1270, + 531, + 1308, + 253, + 1308 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 252, + 1855, + 1402, + 1855, + 1402, + 1987, + 252, + 1987 + ], + "score": 0.86 + }, + { + "category_id": 1, + "poly": [ + 252, + 1343, + 1405, + 1343, + 1405, + 1509, + 252, + 1509 + ], + "score": 0.815 + }, + { + "category_id": 1, + "poly": [ + 252, + 1674, + 1404, + 1674, + 1404, + 1840, + 252, + 1840 + ], + "score": 0.775 + }, + { + "category_id": 1, + "poly": [ + 250, + 1525, + 1405, + 1525, + 1405, + 1657, + 250, + 1657 + ], + "score": 0.774 + }, + { + "category_id": 1, + "poly": [ + 292, + 2038, + 1386, + 2038, + 1386, + 2068, + 292, + 2068 + ], + "score": 0.473 + }, + { + "category_id": 2, + "poly": [ + 292, + 2038, + 1386, + 2038, + 1386, + 2068, + 292, + 2068 + ], + "score": 0.429 + }, + { + "category_id": 15, + "poly": [ + 451.0, + 266.0, + 1249.0, + 266.0, + 1249.0, + 332.0, + 451.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 329.0, + 1088.0, + 329.0, + 1088.0, + 383.0, + 612.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 700.0, + 920.0, + 700.0, + 920.0, + 744.0, + 782.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1268.0, + 535.0, + 1268.0, + 535.0, + 1314.0, + 242.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2034.0, + 1389.0, + 2034.0, + 1389.0, + 2071.0, + 294.0, + 2071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 260.0, + 788.0, + 275.0, + 788.0, + 275.0, + 808.0, + 260.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 776.0, + 1306.0, + 776.0, + 1306.0, + 815.0, + 395.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 819.0, + 277.0, + 819.0, + 277.0, + 844.0, + 258.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 810.0, + 1306.0, + 810.0, + 1306.0, + 849.0, + 395.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 257.0, + 852.0, + 278.0, + 852.0, + 278.0, + 881.0, + 257.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 845.0, + 1307.0, + 845.0, + 1307.0, + 882.0, + 395.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 259.0, + 887.0, + 277.0, + 887.0, + 277.0, + 913.0, + 259.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 881.0, + 1307.0, + 881.0, + 1307.0, + 912.0, + 396.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 257.0, + 917.0, + 280.0, + 917.0, + 280.0, + 945.0, + 257.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 912.0, + 1306.0, + 912.0, + 1306.0, + 945.0, + 395.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 257.0, + 950.0, + 280.0, + 950.0, + 280.0, + 979.0, + 257.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 945.0, + 1305.0, + 945.0, + 1305.0, + 979.0, + 397.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 257.0, + 984.0, + 280.0, + 984.0, + 280.0, + 1012.0, + 257.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 979.0, + 1307.0, + 979.0, + 1307.0, + 1012.0, + 395.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 258.0, + 1019.0, + 277.0, + 1019.0, + 277.0, + 1044.0, + 258.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1012.0, + 1306.0, + 1012.0, + 1306.0, + 1045.0, + 395.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 257.0, + 1051.0, + 280.0, + 1051.0, + 280.0, + 1077.0, + 257.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1045.0, + 1305.0, + 1045.0, + 1305.0, + 1079.0, + 393.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1082.0, + 280.0, + 1082.0, + 280.0, + 1113.0, + 250.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1075.0, + 1307.0, + 1075.0, + 1307.0, + 1116.0, + 392.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1116.0, + 277.0, + 1116.0, + 277.0, + 1145.0, + 248.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1112.0, + 1306.0, + 1112.0, + 1306.0, + 1145.0, + 395.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1150.0, + 278.0, + 1150.0, + 278.0, + 1179.0, + 248.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1147.0, + 1303.0, + 1147.0, + 1303.0, + 1176.0, + 397.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1184.0, + 280.0, + 1184.0, + 280.0, + 1211.0, + 248.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1178.0, + 842.0, + 1178.0, + 842.0, + 1210.0, + 396.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 499.0, + 988.0, + 499.0, + 988.0, + 531.0, + 718.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 527.0, + 913.0, + 527.0, + 913.0, + 563.0, + 787.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 556.0, + 902.0, + 556.0, + 902.0, + 593.0, + 797.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 593.0, + 891.0, + 593.0, + 891.0, + 619.0, + 807.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1861.0, + 278.0, + 1861.0, + 278.0, + 1886.0, + 250.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1855.0, + 1406.0, + 1855.0, + 1406.0, + 1892.0, + 293.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1894.0, + 278.0, + 1894.0, + 278.0, + 1920.0, + 249.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1888.0, + 1404.0, + 1888.0, + 1404.0, + 1921.0, + 294.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1927.0, + 278.0, + 1927.0, + 278.0, + 1954.0, + 249.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1922.0, + 1404.0, + 1922.0, + 1404.0, + 1956.0, + 296.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1962.0, + 279.0, + 1962.0, + 279.0, + 1989.0, + 250.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1956.0, + 919.0, + 1956.0, + 919.0, + 1990.0, + 293.0, + 1990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 1351.0, + 278.0, + 1351.0, + 278.0, + 1375.0, + 253.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1341.0, + 1407.0, + 1341.0, + 1407.0, + 1380.0, + 289.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1384.0, + 279.0, + 1384.0, + 279.0, + 1412.0, + 250.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1378.0, + 1407.0, + 1378.0, + 1407.0, + 1414.0, + 292.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1418.0, + 279.0, + 1418.0, + 279.0, + 1444.0, + 250.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1412.0, + 1404.0, + 1412.0, + 1404.0, + 1447.0, + 295.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1450.0, + 279.0, + 1450.0, + 279.0, + 1477.0, + 250.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1441.0, + 1407.0, + 1441.0, + 1407.0, + 1481.0, + 292.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1483.0, + 279.0, + 1483.0, + 279.0, + 1509.0, + 250.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 1088.0, + 1477.0, + 1088.0, + 1512.0, + 293.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1679.0, + 279.0, + 1679.0, + 279.0, + 1707.0, + 249.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1675.0, + 1404.0, + 1675.0, + 1404.0, + 1710.0, + 295.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1713.0, + 279.0, + 1713.0, + 279.0, + 1740.0, + 249.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1706.0, + 1406.0, + 1706.0, + 1406.0, + 1745.0, + 295.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1746.0, + 279.0, + 1746.0, + 279.0, + 1774.0, + 249.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1739.0, + 1406.0, + 1739.0, + 1406.0, + 1776.0, + 292.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1779.0, + 279.0, + 1779.0, + 279.0, + 1808.0, + 249.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 1404.0, + 1772.0, + 1404.0, + 1811.0, + 293.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1812.0, + 280.0, + 1812.0, + 280.0, + 1839.0, + 249.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1808.0, + 1407.0, + 1808.0, + 1407.0, + 1843.0, + 293.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1533.0, + 277.0, + 1533.0, + 277.0, + 1558.0, + 250.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1527.0, + 1405.0, + 1527.0, + 1405.0, + 1561.0, + 294.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1566.0, + 277.0, + 1566.0, + 277.0, + 1591.0, + 248.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 1404.0, + 1561.0, + 1404.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1599.0, + 277.0, + 1599.0, + 277.0, + 1625.0, + 248.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1595.0, + 1408.0, + 1595.0, + 1408.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1632.0, + 277.0, + 1632.0, + 277.0, + 1657.0, + 248.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1626.0, + 1196.0, + 1626.0, + 1196.0, + 1660.0, + 294.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2034.0, + 1389.0, + 2034.0, + 1389.0, + 2071.0, + 294.0, + 2071.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 250, + 1902, + 1404, + 1902, + 1404, + 2009, + 250, + 2009 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 250, + 1293, + 1405, + 1293, + 1405, + 1395, + 250, + 1395 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 250, + 1512, + 1405, + 1512, + 1405, + 1652, + 250, + 1652 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 250, + 744, + 1406, + 744, + 1406, + 913, + 250, + 913 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 250, + 1077, + 1409, + 1077, + 1409, + 1179, + 250, + 1179 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 505, + 1407, + 1194, + 1407, + 1194, + 1501, + 505, + 1501 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 562, + 1189, + 1136, + 1189, + 1136, + 1286, + 562, + 1286 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 250, + 596, + 1405, + 596, + 1405, + 731, + 250, + 731 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 250, + 202, + 1405, + 202, + 1405, + 369, + 250, + 369 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 249, + 383, + 1406, + 383, + 1406, + 583, + 249, + 583 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 259, + 1741, + 1404, + 1741, + 1404, + 1808, + 259, + 1808 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 286, + 1018, + 965, + 1018, + 965, + 1053, + 286, + 1053 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 262, + 1684, + 514, + 1684, + 514, + 1718, + 262, + 1718 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1437, + 1400, + 1437, + 1400, + 1467, + 1366, + 1467 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1213, + 1400, + 1213, + 1400, + 1244, + 1365, + 1244 + ], + "score": 0.877 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1823, + 1400, + 1823, + 1400, + 1853, + 1365, + 1853 + ], + "score": 0.866 + }, + { + "category_id": 0, + "poly": [ + 267, + 952, + 527, + 952, + 527, + 990, + 267, + 990 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2085, + 840, + 2085 + ], + "score": 0.752 + }, + { + "category_id": 8, + "poly": [ + 618, + 1812, + 1082, + 1812, + 1082, + 1891, + 618, + 1891 + ], + "score": 0.745 + }, + { + "category_id": 8, + "poly": [ + 618, + 1812, + 1083, + 1812, + 1083, + 1891, + 618, + 1891 + ], + "score": 0.612 + }, + { + "category_id": 0, + "poly": [ + 266, + 952, + 528, + 952, + 528, + 990, + 266, + 990 + ], + "score": 0.173 + }, + { + "category_id": 2, + "poly": [ + 840, + 2060, + 860, + 2060, + 860, + 2085, + 840, + 2085 + ], + "score": 0.152 + }, + { + "category_id": 13, + "poly": [ + 587, + 1112, + 740, + 1112, + 740, + 1147, + 587, + 1147 + ], + "score": 0.95, + "latex": "\\mathscr { X } = ( X _ { i } ) _ { i = 1 } ^ { n }" + }, + { + "category_id": 14, + "poly": [ + 505, + 1402, + 1190, + 1402, + 1190, + 1502, + 505, + 1502 + ], + "score": 0.94, + "latex": "S _ { \\mathrm { B D e u } } ^ { ( i ) } ( \\Pi _ { i } \\mid \\mathcal { D } ) \\equiv \\prod _ { j = 1 } ^ { \\beta _ { i } } \\frac { \\Gamma ( \\alpha _ { i , j } ) } { \\Gamma ( N _ { i , j } + \\alpha _ { i , j } ) } \\prod _ { k = 1 } ^ { \\gamma _ { i } } \\frac { \\Gamma ( N _ { i , j , k } + \\alpha _ { i , j , k } ) } { \\Gamma ( \\alpha _ { i , j , k } ) } ," + }, + { + "category_id": 14, + "poly": [ + 615, + 1810, + 1081, + 1810, + 1081, + 1892, + 615, + 1892 + ], + "score": 0.94, + "latex": "H ( \\pmb { \\sigma } ) = \\sum _ { i \\in \\mathcal { V } _ { \\mathrm { A P } } } h _ { i } \\sigma _ { i } + \\sum _ { ( i , j ) \\in \\mathcal { E } _ { \\mathrm { A P } } } J _ { i , j } \\sigma _ { i } \\sigma _ { j } ," + }, + { + "category_id": 14, + "poly": [ + 562, + 1183, + 1135, + 1183, + 1135, + 1286, + 562, + 1286 + ], + "score": 0.93, + "latex": "\\Pi _ { 1 } ^ { * } , \\cdot \\cdot \\cdot , \\Pi _ { n } ^ { * } = \\underset { \\Pi _ { 1 } , \\cdots , \\Pi _ { n } \\subset { \\cal X } } { \\arg \\operatorname* { m i n } } \\sum _ { i = 1 } ^ { n } - \\log S ^ { ( i ) } ( \\Pi _ { i } \\mid { \\mathcal { D } } ) ," + }, + { + "category_id": 13, + "poly": [ + 1219, + 1910, + 1323, + 1910, + 1323, + 1944, + 1219, + 1944 + ], + "score": 0.93, + "latex": "J _ { i , j } \\in \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1901, + 654, + 1901, + 654, + 1943, + 373, + 1943 + ], + "score": 0.93, + "latex": "{ \\pmb \\sigma } = ( { \\boldsymbol \\sigma } _ { i } ) _ { i = 1 } ^ { | { \\boldsymbol \\nu } _ { \\mathrm { A P } } | } \\in \\mathbb { B } ^ { | { \\boldsymbol \\nu } _ { \\mathrm { A P } } | }" + }, + { + "category_id": 13, + "poly": [ + 579, + 416, + 655, + 416, + 655, + 451, + 579, + 451 + ], + "score": 0.93, + "latex": "\\mathcal { O } ( n ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 1162, + 1111, + 1397, + 1111, + 1397, + 1146, + 1162, + 1146 + ], + "score": 0.92, + "latex": "\\mathcal { D } = \\{ D _ { 1 } , \\cdot \\cdot \\cdot , D _ { N } \\}" + }, + { + "category_id": 13, + "poly": [ + 559, + 1587, + 706, + 1587, + 706, + 1625, + 559, + 1625 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\alpha _ { i , j , k } = \\frac { \\alpha } { \\beta _ { i } \\gamma _ { i } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1228, + 1297, + 1375, + 1297, + 1375, + 1327, + 1228, + 1327 + ], + "score": 0.92, + "latex": "S _ { i } : \\Pi _ { i } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1910, + 875, + 1910, + 875, + 1940, + 788, + 1940 + ], + "score": 0.92, + "latex": "h _ { i } \\in \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 298, + 514, + 468, + 514, + 468, + 551, + 298, + 551 + ], + "score": 0.92, + "latex": "O ( n ( n - 1 ) ^ { \\frac { m } { 2 } } )" + }, + { + "category_id": 13, + "poly": [ + 601, + 1942, + 819, + 1942, + 819, + 1976, + 601, + 1976 + ], + "score": 0.91, + "latex": "\\mathcal { G } _ { \\mathrm { A P } } = ( \\nu _ { \\mathrm { A P } } , \\mathcal { E } _ { \\mathrm { A P } } )" + }, + { + "category_id": 13, + "poly": [ + 295, + 1942, + 434, + 1942, + 434, + 1976, + 295, + 1976 + ], + "score": 0.91, + "latex": "( i , j ) \\in \\mathcal { E } _ { \\mathrm { A P } }" + }, + { + "category_id": 13, + "poly": [ + 958, + 1910, + 1059, + 1910, + 1059, + 1941, + 958, + 1941 + ], + "score": 0.91, + "latex": "i \\in \\mathcal { V } _ { \\mathrm { A P } }" + }, + { + "category_id": 13, + "poly": [ + 1319, + 451, + 1397, + 451, + 1397, + 481, + 1319, + 481 + ], + "score": 0.9, + "latex": "m \\geq 3" + }, + { + "category_id": 13, + "poly": [ + 1276, + 1586, + 1401, + 1586, + 1401, + 1615, + 1276, + 1615 + ], + "score": 0.9, + "latex": "0 < \\alpha \\in \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1200, + 1552, + 1234, + 1552, + 1234, + 1583, + 1200, + 1583 + ], + "score": 0.89, + "latex": "\\Pi _ { i }" + }, + { + "category_id": 13, + "poly": [ + 344, + 1587, + 378, + 1587, + 378, + 1616, + 344, + 1616 + ], + "score": 0.89, + "latex": "X _ { i }" + }, + { + "category_id": 13, + "poly": [ + 730, + 1332, + 764, + 1332, + 764, + 1360, + 730, + 1360 + ], + "score": 0.88, + "latex": "X _ { i }" + }, + { + "category_id": 13, + "poly": [ + 699, + 1555, + 767, + 1555, + 767, + 1587, + 699, + 1587 + ], + "score": 0.88, + "latex": "N _ { i , j , k }" + }, + { + "category_id": 13, + "poly": [ + 1320, + 418, + 1398, + 418, + 1398, + 446, + 1320, + 446 + ], + "score": 0.88, + "latex": "m = 2" + }, + { + "category_id": 13, + "poly": [ + 370, + 1515, + 1068, + 1515, + 1068, + 1555, + 370, + 1555 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { N = \\sum _ { j = 1 } ^ { \\beta _ { i } } N _ { i , j } , N _ { i , j } = \\sum _ { k = 1 } ^ { \\gamma _ { i } } N _ { i , j , k } , \\alpha _ { i , j } = \\sum _ { k = 1 } ^ { \\gamma _ { i } } \\alpha _ { i , j , k } , \\beta _ { i } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 594, + 1147, + 627, + 1147, + 627, + 1176, + 594, + 1176 + ], + "score": 0.87, + "latex": "\\Pi _ { i }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1295, + 1172, + 1295, + 1172, + 1331, + 371, + 1331 + ], + "score": 0.86, + "latex": "{ \\mathcal { G } } = ( \\mathcal { V } , { \\mathcal { E } } ) , \\mathcal { V } = \\{ 1 , \\cdots , n \\} , { \\mathcal { E } } = \\{ ( j , i ) | j , i \\in \\{ 1 , \\cdots , n \\} , X _ { j } \\in \\Pi _ { i } \\} ," + }, + { + "category_id": 13, + "poly": [ + 1027, + 1115, + 1054, + 1115, + 1054, + 1140, + 1027, + 1140 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1298, + 1554, + 1313, + 1554, + 1313, + 1584, + 1298, + 1584 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 443, + 1587, + 461, + 1587, + 461, + 1613, + 443, + 1613 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 297, + 1552, + 330, + 1552, + 330, + 1583, + 297, + 1583 + ], + "score": 0.8, + "latex": "\\Pi _ { i }" + }, + { + "category_id": 13, + "poly": [ + 341, + 1555, + 367, + 1555, + 367, + 1584, + 341, + 1584 + ], + "score": 0.79, + "latex": "\\gamma _ { i }" + }, + { + "category_id": 13, + "poly": [ + 653, + 1556, + 686, + 1556, + 686, + 1583, + 653, + 1583 + ], + "score": 0.76, + "latex": "X _ { i }" + }, + { + "category_id": 13, + "poly": [ + 746, + 423, + 766, + 423, + 766, + 445, + 746, + 445 + ], + "score": 0.67, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1519, + 1064, + 1519, + 1064, + 1551, + 1035, + 1551 + ], + "score": 0.38, + "latex": "\\beta _ { i }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1012.0, + 970.0, + 1012.0, + 970.0, + 1062.0, + 291.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 1680.0, + 517.0, + 1680.0, + 517.0, + 1723.0, + 252.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 254.0, + 947.0, + 530.0, + 947.0, + 530.0, + 998.0, + 254.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 863.0, + 2058.0, + 863.0, + 2093.0, + 838.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 947.0, + 530.0, + 947.0, + 530.0, + 998.0, + 253.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2093.0, + 838.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1914.0, + 281.0, + 1914.0, + 281.0, + 1940.0, + 250.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1893.0, + 372.0, + 1893.0, + 372.0, + 1955.0, + 285.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1893.0, + 787.0, + 1893.0, + 787.0, + 1955.0, + 655.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1893.0, + 957.0, + 1893.0, + 957.0, + 1955.0, + 876.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1893.0, + 1218.0, + 1893.0, + 1218.0, + 1955.0, + 1060.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 1893.0, + 1407.0, + 1893.0, + 1407.0, + 1955.0, + 1324.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1948.0, + 278.0, + 1948.0, + 278.0, + 1974.0, + 250.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1940.0, + 294.0, + 1940.0, + 294.0, + 1977.0, + 289.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1940.0, + 600.0, + 1940.0, + 600.0, + 1977.0, + 435.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1940.0, + 1407.0, + 1940.0, + 1407.0, + 1977.0, + 820.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1981.0, + 278.0, + 1981.0, + 278.0, + 2007.0, + 250.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1975.0, + 1127.0, + 1975.0, + 1127.0, + 2011.0, + 291.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1294.0, + 370.0, + 1294.0, + 370.0, + 1333.0, + 242.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1294.0, + 1227.0, + 1294.0, + 1227.0, + 1333.0, + 1173.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1294.0, + 1408.0, + 1294.0, + 1408.0, + 1333.0, + 1376.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1329.0, + 729.0, + 1329.0, + 729.0, + 1364.0, + 248.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1329.0, + 1404.0, + 1329.0, + 1404.0, + 1364.0, + 765.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1360.0, + 890.0, + 1360.0, + 890.0, + 1398.0, + 247.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1491.0, + 296.0, + 1491.0, + 296.0, + 1577.0, + 248.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1491.0, + 369.0, + 1491.0, + 369.0, + 1577.0, + 331.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1491.0, + 1199.0, + 1491.0, + 1199.0, + 1577.0, + 1069.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1491.0, + 1424.0, + 1491.0, + 1424.0, + 1577.0, + 1235.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1554.0, + 281.0, + 1554.0, + 281.0, + 1583.0, + 247.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1551.0, + 296.0, + 1551.0, + 296.0, + 1586.0, + 292.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1551.0, + 340.0, + 1551.0, + 340.0, + 1586.0, + 331.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1551.0, + 652.0, + 1551.0, + 652.0, + 1586.0, + 368.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1551.0, + 698.0, + 1551.0, + 698.0, + 1586.0, + 687.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1551.0, + 1199.0, + 1551.0, + 1199.0, + 1586.0, + 768.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1551.0, + 1297.0, + 1551.0, + 1297.0, + 1586.0, + 1235.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1551.0, + 1405.0, + 1551.0, + 1405.0, + 1586.0, + 1314.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1591.0, + 280.0, + 1591.0, + 280.0, + 1616.0, + 248.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1574.0, + 1275.0, + 1574.0, + 1275.0, + 1628.0, + 707.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1574.0, + 1411.0, + 1574.0, + 1411.0, + 1628.0, + 1402.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1624.0, + 281.0, + 1624.0, + 281.0, + 1650.0, + 250.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1616.0, + 953.0, + 1616.0, + 953.0, + 1651.0, + 290.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.25, + 1580.5, + 715.25, + 1580.5, + 715.25, + 1627.5, + 287.25, + 1627.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 751.0, + 278.0, + 751.0, + 278.0, + 779.0, + 248.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 746.0, + 1405.0, + 746.0, + 1405.0, + 781.0, + 294.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 786.0, + 278.0, + 786.0, + 278.0, + 813.0, + 248.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 781.0, + 1408.0, + 781.0, + 1408.0, + 816.0, + 294.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 819.0, + 278.0, + 819.0, + 278.0, + 847.0, + 248.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 814.0, + 1405.0, + 814.0, + 1405.0, + 850.0, + 296.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 852.0, + 278.0, + 852.0, + 278.0, + 880.0, + 247.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 847.0, + 1405.0, + 847.0, + 1405.0, + 882.0, + 296.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 885.0, + 278.0, + 885.0, + 278.0, + 913.0, + 248.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 879.0, + 1311.0, + 879.0, + 1311.0, + 914.0, + 293.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1084.0, + 279.0, + 1084.0, + 279.0, + 1109.0, + 249.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1077.0, + 1408.0, + 1077.0, + 1408.0, + 1113.0, + 292.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1118.0, + 279.0, + 1118.0, + 279.0, + 1143.0, + 249.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1108.0, + 586.0, + 1108.0, + 586.0, + 1149.0, + 292.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1108.0, + 1026.0, + 1108.0, + 1026.0, + 1149.0, + 741.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1108.0, + 1161.0, + 1108.0, + 1161.0, + 1149.0, + 1055.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1108.0, + 1411.0, + 1108.0, + 1411.0, + 1149.0, + 1398.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 1146.0, + 593.0, + 1146.0, + 593.0, + 1181.0, + 247.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1146.0, + 913.0, + 1146.0, + 913.0, + 1181.0, + 628.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 605.0, + 281.0, + 605.0, + 281.0, + 629.0, + 250.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 597.0, + 1407.0, + 597.0, + 1407.0, + 632.0, + 289.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 639.0, + 278.0, + 639.0, + 278.0, + 664.0, + 248.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 634.0, + 1405.0, + 634.0, + 1405.0, + 665.0, + 294.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 672.0, + 278.0, + 672.0, + 278.0, + 698.0, + 248.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 665.0, + 1407.0, + 665.0, + 1407.0, + 701.0, + 291.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 705.0, + 278.0, + 705.0, + 278.0, + 732.0, + 250.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 699.0, + 932.0, + 699.0, + 932.0, + 733.0, + 293.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 209.0, + 278.0, + 209.0, + 278.0, + 237.0, + 248.0, + 237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 241.0, + 293.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 241.0, + 278.0, + 241.0, + 278.0, + 270.0, + 248.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 237.0, + 1405.0, + 237.0, + 1405.0, + 272.0, + 295.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 274.0, + 278.0, + 274.0, + 278.0, + 304.0, + 248.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 267.0, + 1404.0, + 267.0, + 1404.0, + 308.0, + 293.0, + 308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 306.0, + 280.0, + 306.0, + 280.0, + 338.0, + 246.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 299.0, + 1405.0, + 299.0, + 1405.0, + 341.0, + 293.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 343.0, + 278.0, + 343.0, + 278.0, + 369.0, + 248.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 334.0, + 1068.0, + 334.0, + 1068.0, + 372.0, + 293.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 389.0, + 278.0, + 389.0, + 278.0, + 417.0, + 249.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 421.0, + 293.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 423.0, + 278.0, + 423.0, + 278.0, + 451.0, + 247.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 417.0, + 578.0, + 417.0, + 578.0, + 451.0, + 296.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 417.0, + 745.0, + 417.0, + 745.0, + 451.0, + 656.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 417.0, + 1319.0, + 417.0, + 1319.0, + 451.0, + 767.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 417.0, + 1409.0, + 417.0, + 1409.0, + 451.0, + 1399.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 456.0, + 278.0, + 456.0, + 278.0, + 483.0, + 249.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 451.0, + 1318.0, + 451.0, + 1318.0, + 484.0, + 295.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 451.0, + 1408.0, + 451.0, + 1408.0, + 484.0, + 1398.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 488.0, + 278.0, + 488.0, + 278.0, + 516.0, + 247.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 484.0, + 1406.0, + 484.0, + 1406.0, + 518.0, + 296.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 521.0, + 280.0, + 521.0, + 280.0, + 553.0, + 246.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 512.0, + 297.0, + 512.0, + 297.0, + 557.0, + 292.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 512.0, + 1406.0, + 512.0, + 1406.0, + 557.0, + 469.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 554.0, + 280.0, + 554.0, + 280.0, + 585.0, + 246.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 554.0, + 427.0, + 554.0, + 427.0, + 585.0, + 295.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 253.0, + 1751.0, + 279.0, + 1751.0, + 279.0, + 1772.0, + 253.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 1738.0, + 1407.0, + 1738.0, + 1407.0, + 1780.0, + 282.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1775.0, + 503.0, + 1775.0, + 503.0, + 1811.0, + 248.0, + 1811.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1619, + 1406, + 1619, + 1406, + 1821, + 295, + 1821 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 250, + 1389, + 1405, + 1389, + 1405, + 1511, + 250, + 1511 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 251, + 293, + 1405, + 293, + 1405, + 428, + 251, + 428 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 248, + 1157, + 1406, + 1157, + 1406, + 1259, + 248, + 1259 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 580, + 678, + 1118, + 678, + 1118, + 764, + 580, + 764 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 249, + 517, + 1405, + 517, + 1405, + 664, + 249, + 664 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 253, + 779, + 1406, + 779, + 1406, + 848, + 253, + 848 + ], + "score": 0.946 + }, + { + "category_id": 0, + "poly": [ + 252, + 1547, + 875, + 1547, + 875, + 1589, + 252, + 1589 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 256, + 201, + 1405, + 201, + 1405, + 270, + 256, + 270 + ], + "score": 0.94 + }, + { + "category_id": 8, + "poly": [ + 399, + 1275, + 1303, + 1275, + 1303, + 1371, + 399, + 1371 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 281, + 1831, + 1405, + 1831, + 1405, + 1903, + 281, + 1903 + ], + "score": 0.936 + }, + { + "category_id": 8, + "poly": [ + 514, + 860, + 1185, + 860, + 1185, + 904, + 514, + 904 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 358, + 1067, + 1335, + 1067, + 1335, + 1145, + 358, + 1145 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 261, + 459, + 1209, + 459, + 1209, + 496, + 261, + 496 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1365, + 703, + 1401, + 703, + 1401, + 734, + 1365, + 734 + ], + "score": 0.867 + }, + { + "category_id": 9, + "poly": [ + 1365, + 867, + 1400, + 867, + 1400, + 898, + 1365, + 898 + ], + "score": 0.864 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1291, + 1401, + 1291, + 1401, + 1322, + 1365, + 1322 + ], + "score": 0.84 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1079, + 1401, + 1079, + 1401, + 1110, + 1366, + 1110 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 248, + 921, + 1412, + 921, + 1412, + 1056, + 248, + 1056 + ], + "score": 0.814 + }, + { + "category_id": 1, + "poly": [ + 359, + 1920, + 1274, + 1920, + 1274, + 2013, + 359, + 2013 + ], + "score": 0.791 + }, + { + "category_id": 2, + "poly": [ + 841, + 2060, + 860, + 2060, + 860, + 2085, + 841, + 2085 + ], + "score": 0.637 + }, + { + "category_id": 2, + "poly": [ + 840, + 2060, + 859, + 2060, + 859, + 2085, + 840, + 2085 + ], + "score": 0.6 + }, + { + "category_id": 14, + "poly": [ + 579, + 674, + 1119, + 674, + 1119, + 766, + 579, + 766 + ], + "score": 0.95, + "latex": "H _ { \\mathrm { t o t a l } } ( d , r ) \\equiv \\sum _ { i = 1 } ^ { n } H _ { \\mathrm { s c o r e } } ^ { ( i ) } ( d , i ) + H _ { \\mathrm { c y c l e } } ( d , r ) ." + }, + { + "category_id": 13, + "poly": [ + 861, + 1831, + 981, + 1831, + 981, + 1872, + 861, + 1872 + ], + "score": 0.94, + "latex": "( W _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 1025, + 779, + 1201, + 779, + 1201, + 817, + 1025, + 817 + ], + "score": 0.93, + "latex": "\\Pi _ { i } = \\pi ^ { ( i ) } ( d _ { \\cdot , i } )" + }, + { + "category_id": 13, + "poly": [ + 934, + 1474, + 1066, + 1474, + 1066, + 1510, + 934, + 1510 + ], + "score": 0.93, + "latex": "O ( n ^ { 2 } 2 ^ { n - 2 } )" + }, + { + "category_id": 14, + "poly": [ + 363, + 1066, + 1333, + 1066, + 1333, + 1146, + 363, + 1146 + ], + "score": 0.93, + "latex": "H _ { \\mathrm { c y c l e } } ( d , r ) \\equiv \\sum _ { 1 \\leq i < j < k \\leq n } \\delta _ { 1 } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) + \\sum _ { 1 \\leq i < j \\leq n } \\delta _ { 2 } ( d _ { i , j } r _ { i , j } + d _ { j , i } ( 1 - r _ { i , j } ) ) ," + }, + { + "category_id": 13, + "poly": [ + 692, + 324, + 777, + 324, + 777, + 362, + 692, + 362 + ], + "score": 0.93, + "latex": "\\dot { \\mathcal { O } } ( \\dot { I } ^ { \\frac { K } { 2 } } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1429, + 433, + 1429, + 433, + 1478, + 298, + 1478 + ], + "score": 0.92, + "latex": "d _ { j , i } ^ { ( j ^ { * } , i ^ { * } ) } = 0" + }, + { + "category_id": 13, + "poly": [ + 536, + 1923, + 656, + 1923, + 656, + 1964, + 536, + 1964 + ], + "score": 0.92, + "latex": "( W _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 919, + 988, + 1091, + 988, + 1091, + 1024, + 919, + 1024 + ], + "score": 0.92, + "latex": "\\mathcal { O } ( n ( n - 1 ) ^ { \\frac { m } { 2 } } )" + }, + { + "category_id": 13, + "poly": [ + 1172, + 918, + 1292, + 918, + 1292, + 957, + 1172, + 957 + ], + "score": 0.92, + "latex": "\\mathcal { O } ( n 2 ^ { \\frac { n - 1 } { 2 } } )" + }, + { + "category_id": 13, + "poly": [ + 856, + 203, + 943, + 203, + 943, + 236, + 856, + 236 + ], + "score": 0.92, + "latex": "g ( \\pmb { v } , \\pmb { h } )" + }, + { + "category_id": 13, + "poly": [ + 1029, + 1160, + 1185, + 1160, + 1185, + 1192, + 1029, + 1192 + ], + "score": 0.92, + "latex": "r _ { 1 } , r _ { 2 } , r _ { 3 } \\in \\mathbb { B }" + }, + { + "category_id": 13, + "poly": [ + 1030, + 1386, + 1398, + 1386, + 1398, + 1435, + 1030, + 1435 + ], + "score": 0.92, + "latex": "{ \\pmb d } ^ { ( j ^ { \\ast } , i ^ { \\ast } ) } = ( ( d _ { j , i } ) _ { 1 \\leq j \\leq n , j \\neq i } ^ { ( j ^ { \\ast } , i ^ { \\ast } ) } ) _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 1253, + 323, + 1331, + 323, + 1331, + 362, + 1253, + 362 + ], + "score": 0.92, + "latex": "\\check { \\mathcal { O } ( 2 ^ { \\frac { I } { 2 } } ) }" + }, + { + "category_id": 13, + "poly": [ + 374, + 520, + 482, + 520, + 482, + 554, + 374, + 554 + ], + "score": 0.92, + "latex": "n ( n - 1 )" + }, + { + "category_id": 13, + "poly": [ + 397, + 236, + 454, + 236, + 454, + 270, + 397, + 270 + ], + "score": 0.91, + "latex": "f ( v )" + }, + { + "category_id": 13, + "poly": [ + 520, + 236, + 805, + 236, + 805, + 270, + 520, + 270 + ], + "score": 0.91, + "latex": "f ( v ) = \\operatorname* { m i n } _ { h \\in \\mathbb { B } ^ { J } } g ( v , h )" + }, + { + "category_id": 13, + "poly": [ + 374, + 924, + 496, + 924, + 496, + 954, + 374, + 954 + ], + "score": 0.91, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 572, + 1476, + 683, + 1476, + 683, + 1510, + 572, + 1510 + ], + "score": 0.91, + "latex": "\\mathcal { O } ( n ^ { m + 1 } )" + }, + { + "category_id": 13, + "poly": [ + 433, + 1194, + 602, + 1194, + 602, + 1224, + 433, + 1224 + ], + "score": 0.91, + "latex": "0 < \\delta _ { 1 } , \\delta _ { 2 } \\in \\mathbb { R }" + }, + { + "category_id": 14, + "poly": [ + 399, + 1269, + 1298, + 1269, + 1298, + 1371, + 399, + 1371 + ], + "score": 0.9, + "latex": "\\operatorname* { m a x } \\{ 0 , \\operatorname* { m a x } _ { \\substack { 1 \\leq j ^ { * } , i ^ { * } \\leq n d _ { \\cdot , j ^ { * } } \\in \\mathbb { B } ^ { n - 1 } } } ( H _ { \\mathrm { s c o r e } } ^ { ( i ^ { * } ) } ( d _ { \\cdot , i ^ { * } } ^ { ( j ^ { * } , i ^ { * } ) } ) - H _ { \\mathrm { s c o r e } } ^ { ( i ^ { * } ) } ( d _ { \\cdot , i ^ { * } } ) ) \\} < \\delta _ { 1 } < \\frac { \\delta _ { 2 } } { n - 2 } ," + }, + { + "category_id": 13, + "poly": [ + 884, + 235, + 967, + 235, + 967, + 265, + 884, + 265 + ], + "score": 0.9, + "latex": "\\pmb { v } \\in \\mathbb { B } ^ { I }" + }, + { + "category_id": 13, + "poly": [ + 1281, + 783, + 1398, + 783, + 1398, + 813, + 1281, + 813 + ], + "score": 0.9, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 875, + 598, + 909, + 598, + 909, + 628, + 875, + 628 + ], + "score": 0.9, + "latex": "X _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1137, + 1478, + 1265, + 1478, + 1265, + 1506, + 1137, + 1506 + ], + "score": 0.9, + "latex": "m = n - 1" + }, + { + "category_id": 13, + "poly": [ + 448, + 784, + 493, + 784, + 493, + 815, + 448, + 815 + ], + "score": 0.89, + "latex": "d _ { \\cdot , i }" + }, + { + "category_id": 13, + "poly": [ + 1340, + 1688, + 1375, + 1688, + 1375, + 1719, + 1340, + 1719 + ], + "score": 0.89, + "latex": "X _ { i }" + }, + { + "category_id": 14, + "poly": [ + 514, + 858, + 1183, + 858, + 1183, + 903, + 514, + 903 + ], + "score": 0.89, + "latex": "H _ { \\mathrm { s c o r e } } ^ { ( i ) } ( d _ { \\cdot , i } ) \\equiv - \\log S ^ { ( i ) } ( \\pi ^ { ( i ) } ( d _ { \\cdot , i } ) \\mid \\mathcal { D } ) + \\log S ^ { ( i ) } ( \\phi \\mid \\mathcal { D } ) ," + }, + { + "category_id": 13, + "poly": [ + 297, + 554, + 335, + 554, + 335, + 589, + 297, + 589 + ], + "score": 0.89, + "latex": "X _ { j }" + }, + { + "category_id": 13, + "poly": [ + 297, + 362, + 386, + 362, + 386, + 390, + 297, + 390 + ], + "score": 0.89, + "latex": "K = I" + }, + { + "category_id": 13, + "poly": [ + 934, + 783, + 966, + 783, + 966, + 813, + 934, + 813 + ], + "score": 0.87, + "latex": "\\Pi _ { i }" + }, + { + "category_id": 13, + "poly": [ + 579, + 554, + 684, + 554, + 684, + 589, + 579, + 589 + ], + "score": 0.87, + "latex": "d _ { j , i } = 0" + }, + { + "category_id": 13, + "poly": [ + 1124, + 598, + 1218, + 598, + 1218, + 632, + 1124, + 632 + ], + "score": 0.87, + "latex": "r _ { i , j } = 1" + }, + { + "category_id": 13, + "poly": [ + 298, + 1718, + 1007, + 1718, + 1007, + 1756, + 298, + 1756 + ], + "score": 0.87, + "latex": "\\{ W \\subseteq { \\mathcal { X } } \\setminus \\{ X _ { i } \\} \\mid W ^ { \\prime } \\subset W \\Rightarrow S ^ { ( i ) } ( W ^ { \\prime } \\mid { \\mathcal { D } } ) < S ^ { ( i ) } ( W \\mid { \\mathcal { D } } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 826, + 518, + 1257, + 518, + 1257, + 555, + 826, + 555 + ], + "score": 0.85, + "latex": "\\pmb { d } = ( ( d _ { j , i } ) _ { 1 \\leq j \\leq n , j \\neq i } ) _ { i = 1 } ^ { n } \\in \\mathbb { B } ^ { n ( n - 1 ) }" + }, + { + "category_id": 13, + "poly": [ + 297, + 589, + 691, + 589, + 691, + 633, + 297, + 633 + ], + "score": 0.84, + "latex": "r = ( r _ { i , j } ) _ { 1 \\leq i < j \\leq n } \\in \\mathbb { B } ^ { \\binom { n } { 2 } } ( r _ { i , j } =" + }, + { + "category_id": 13, + "poly": [ + 872, + 553, + 910, + 553, + 910, + 591, + 872, + 591 + ], + "score": 0.83, + "latex": "\\binom { n } { 2 }" + }, + { + "category_id": 13, + "poly": [ + 613, + 1227, + 633, + 1227, + 633, + 1253, + 613, + 1253 + ], + "score": 0.81, + "latex": "^ d" + }, + { + "category_id": 13, + "poly": [ + 1270, + 520, + 1375, + 520, + 1375, + 555, + 1270, + 555 + ], + "score": 0.79, + "latex": "( d _ { j , i } = 1" + }, + { + "category_id": 13, + "poly": [ + 375, + 1157, + 947, + 1157, + 947, + 1193, + 375, + 1193 + ], + "score": 0.79, + "latex": "R ( r _ { 1 } , r _ { 2 } , r _ { 3 } ) = r _ { 1 } r _ { 2 } ( 1 - r _ { 3 } ) + ( 1 - r _ { 1 } ) ( 1 - r _ { 2 } ) r _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 1328, + 297, + 1357, + 297, + 1357, + 323, + 1328, + 323 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 649, + 1430, + 810, + 1430, + 810, + 1478, + 649, + 1478 + ], + "score": 0.76, + "latex": "d _ { j , i } ^ { ( j ^ { * } , i ^ { * } ) } = d _ { j , i }" + }, + { + "category_id": 13, + "poly": [ + 920, + 1231, + 938, + 1231, + 938, + 1253, + 920, + 1253 + ], + "score": 0.76, + "latex": "\\pmb { r }" + }, + { + "category_id": 13, + "poly": [ + 711, + 1922, + 1157, + 1922, + 1157, + 1964, + 711, + 1964 + ], + "score": 0.76, + "latex": "( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { i } } , ( W _ { h , i } \\cap ( { \\mathcal { X } } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { i } } ," + }, + { + "category_id": 13, + "poly": [ + 1246, + 1789, + 1266, + 1789, + 1266, + 1815, + 1246, + 1815 + ], + "score": 0.75, + "latex": "^ d" + }, + { + "category_id": 13, + "poly": [ + 820, + 1971, + 1266, + 1971, + 1266, + 2011, + 820, + 2011 + ], + "score": 0.71, + "latex": "( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { i } } , ( W _ { h , i } \\cap ( { \\mathcal { X } } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { i } } ," + }, + { + "category_id": 13, + "poly": [ + 1042, + 297, + 1061, + 297, + 1061, + 323, + 1042, + 323 + ], + "score": 0.71, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 1075, + 598, + 1112, + 598, + 1112, + 632, + 1075, + 632 + ], + "score": 0.7, + "latex": "X _ { j }" + }, + { + "category_id": 13, + "poly": [ + 529, + 554, + 563, + 554, + 563, + 585, + 529, + 585 + ], + "score": 0.69, + "latex": "X _ { i }" + }, + { + "category_id": 13, + "poly": [ + 462, + 1433, + 637, + 1433, + 637, + 1474, + 462, + 1474 + ], + "score": 0.65, + "latex": "( j , i ) = ( j ^ { * } , i ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 611, + 597, + 706, + 597, + 706, + 632, + 611, + 632 + ], + "score": 0.63, + "latex": "( r _ { i , j } = 0" + }, + { + "category_id": 14, + "poly": [ + 710, + 1922, + 1158, + 1922, + 1158, + 1964, + 710, + 1964 + ], + "score": 0.51, + "latex": "( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { i } } , ( W _ { h , i } \\cap ( { \\mathcal { X } } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { i } } ," + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1542.0, + 878.0, + 1542.0, + 878.0, + 1599.0, + 240.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 456.0, + 1211.0, + 456.0, + 1211.0, + 502.0, + 250.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2090.0, + 838.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 861.0, + 2059.0, + 861.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1622.0, + 1407.0, + 1622.0, + 1407.0, + 1655.0, + 295.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1651.0, + 1406.0, + 1651.0, + 1406.0, + 1692.0, + 293.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1688.0, + 1339.0, + 1688.0, + 1339.0, + 1721.0, + 295.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1688.0, + 1404.0, + 1688.0, + 1404.0, + 1721.0, + 1376.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1717.0, + 1404.0, + 1717.0, + 1404.0, + 1758.0, + 1008.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1754.0, + 1407.0, + 1754.0, + 1407.0, + 1791.0, + 293.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1786.0, + 1245.0, + 1786.0, + 1245.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1786.0, + 1277.0, + 1786.0, + 1277.0, + 1824.0, + 1267.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 280.0, + 1376.0, + 1029.0, + 1376.0, + 1029.0, + 1450.0, + 280.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1376.0, + 1421.0, + 1376.0, + 1421.0, + 1450.0, + 1399.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1399.0, + 278.0, + 1399.0, + 278.0, + 1426.0, + 248.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 1440.0, + 281.0, + 1440.0, + 281.0, + 1474.0, + 246.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 1415.0, + 297.0, + 1415.0, + 297.0, + 1496.0, + 282.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 1415.0, + 461.0, + 1415.0, + 461.0, + 1496.0, + 434.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1415.0, + 648.0, + 1415.0, + 648.0, + 1496.0, + 638.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 1415.0, + 1411.0, + 1415.0, + 1411.0, + 1496.0, + 811.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 1481.0, + 280.0, + 1481.0, + 280.0, + 1508.0, + 248.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1465.0, + 571.0, + 1465.0, + 571.0, + 1515.0, + 285.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1465.0, + 933.0, + 1465.0, + 933.0, + 1515.0, + 684.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1465.0, + 1136.0, + 1465.0, + 1136.0, + 1515.0, + 1067.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1465.0, + 1280.0, + 1465.0, + 1280.0, + 1515.0, + 1266.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 243.0, + 293.0, + 1041.0, + 293.0, + 1041.0, + 331.0, + 243.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 293.0, + 1327.0, + 293.0, + 1327.0, + 331.0, + 1062.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1358.0, + 293.0, + 1407.0, + 293.0, + 1407.0, + 331.0, + 1358.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 335.0, + 278.0, + 335.0, + 278.0, + 358.0, + 249.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 325.0, + 691.0, + 325.0, + 691.0, + 363.0, + 289.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 325.0, + 1252.0, + 325.0, + 1252.0, + 363.0, + 778.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 325.0, + 1407.0, + 325.0, + 1407.0, + 363.0, + 1332.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 367.0, + 278.0, + 367.0, + 278.0, + 392.0, + 249.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 359.0, + 296.0, + 359.0, + 296.0, + 396.0, + 292.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 359.0, + 1407.0, + 359.0, + 1407.0, + 396.0, + 387.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 252.0, + 403.0, + 277.0, + 403.0, + 277.0, + 424.0, + 252.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 395.0, + 1036.0, + 395.0, + 1036.0, + 428.0, + 290.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1156.0, + 374.0, + 1156.0, + 374.0, + 1195.0, + 241.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1156.0, + 1028.0, + 1156.0, + 1028.0, + 1195.0, + 948.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1156.0, + 1405.0, + 1156.0, + 1405.0, + 1195.0, + 1186.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1192.0, + 432.0, + 1192.0, + 432.0, + 1227.0, + 249.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1192.0, + 1404.0, + 1192.0, + 1404.0, + 1227.0, + 603.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 1224.0, + 612.0, + 1224.0, + 612.0, + 1258.0, + 250.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1224.0, + 919.0, + 1224.0, + 919.0, + 1258.0, + 634.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1224.0, + 1115.0, + 1224.0, + 1115.0, + 1258.0, + 939.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 515.0, + 373.0, + 515.0, + 373.0, + 557.0, + 239.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 515.0, + 825.0, + 515.0, + 825.0, + 557.0, + 483.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 515.0, + 1269.0, + 515.0, + 1269.0, + 557.0, + 1258.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 515.0, + 1410.0, + 515.0, + 1410.0, + 557.0, + 1376.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 250.0, + 558.0, + 280.0, + 558.0, + 280.0, + 585.0, + 250.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 549.0, + 296.0, + 549.0, + 296.0, + 590.0, + 287.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 549.0, + 528.0, + 549.0, + 528.0, + 590.0, + 336.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 549.0, + 578.0, + 549.0, + 578.0, + 590.0, + 564.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 549.0, + 871.0, + 549.0, + 871.0, + 590.0, + 685.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 549.0, + 1407.0, + 549.0, + 1407.0, + 590.0, + 911.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 600.0, + 280.0, + 600.0, + 280.0, + 630.0, + 247.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 585.0, + 296.0, + 585.0, + 296.0, + 638.0, + 285.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 585.0, + 874.0, + 585.0, + 874.0, + 638.0, + 707.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 585.0, + 1074.0, + 585.0, + 1074.0, + 638.0, + 910.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 585.0, + 1123.0, + 585.0, + 1123.0, + 638.0, + 1113.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 585.0, + 1411.0, + 585.0, + 1411.0, + 638.0, + 1219.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 634.0, + 281.0, + 634.0, + 281.0, + 664.0, + 247.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 628.0, + 841.0, + 628.0, + 841.0, + 668.0, + 289.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 788.0, + 281.0, + 788.0, + 281.0, + 813.0, + 249.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 775.0, + 447.0, + 775.0, + 447.0, + 821.0, + 287.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 775.0, + 933.0, + 775.0, + 933.0, + 821.0, + 494.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 775.0, + 1024.0, + 775.0, + 1024.0, + 821.0, + 967.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 775.0, + 1280.0, + 775.0, + 1280.0, + 821.0, + 1202.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 775.0, + 1410.0, + 775.0, + 1410.0, + 821.0, + 1399.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 814.0, + 716.0, + 814.0, + 716.0, + 849.0, + 242.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 197.0, + 855.0, + 197.0, + 855.0, + 241.0, + 249.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 197.0, + 1406.0, + 197.0, + 1406.0, + 241.0, + 944.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 232.0, + 396.0, + 232.0, + 396.0, + 274.0, + 241.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 232.0, + 519.0, + 232.0, + 519.0, + 274.0, + 455.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 232.0, + 883.0, + 232.0, + 883.0, + 274.0, + 806.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 232.0, + 983.0, + 232.0, + 983.0, + 274.0, + 968.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1821.0, + 860.0, + 1821.0, + 860.0, + 1884.0, + 288.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1821.0, + 1412.0, + 1821.0, + 1412.0, + 1884.0, + 982.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1870.0, + 667.0, + 1870.0, + 667.0, + 1903.0, + 296.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 929.0, + 280.0, + 929.0, + 280.0, + 953.0, + 249.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 920.0, + 373.0, + 920.0, + 373.0, + 958.0, + 290.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 920.0, + 1171.0, + 920.0, + 1171.0, + 958.0, + 497.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 920.0, + 1404.0, + 920.0, + 1404.0, + 958.0, + 1293.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 962.0, + 278.0, + 962.0, + 278.0, + 987.0, + 248.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 957.0, + 1407.0, + 957.0, + 1407.0, + 991.0, + 291.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 996.0, + 278.0, + 996.0, + 278.0, + 1022.0, + 248.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 989.0, + 918.0, + 989.0, + 918.0, + 1026.0, + 291.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 989.0, + 1405.0, + 989.0, + 1405.0, + 1026.0, + 1092.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 249.0, + 1030.0, + 278.0, + 1030.0, + 278.0, + 1055.0, + 249.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1023.0, + 1267.0, + 1023.0, + 1267.0, + 1061.0, + 291.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1910.0, + 535.0, + 1910.0, + 535.0, + 1973.0, + 356.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1910.0, + 709.0, + 1910.0, + 709.0, + 1973.0, + 657.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1910.0, + 1168.0, + 1910.0, + 1168.0, + 1973.0, + 1159.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1957.0, + 819.0, + 1957.0, + 819.0, + 2022.0, + 353.0, + 2022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1957.0, + 1274.0, + 1957.0, + 1274.0, + 2022.0, + 1267.0, + 2022.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 242, + 843, + 1403, + 843, + 1403, + 952, + 242, + 952 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 240, + 1791, + 1405, + 1791, + 1405, + 1895, + 240, + 1895 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 241, + 392, + 1409, + 392, + 1409, + 536, + 241, + 536 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 644, + 1903, + 1055, + 1903, + 1055, + 2003, + 644, + 2003 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 242, + 1532, + 1407, + 1532, + 1407, + 1680, + 242, + 1680 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 305, + 1028, + 1286, + 1028, + 1286, + 1210, + 305, + 1210 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 296, + 1303, + 1447, + 1303, + 1447, + 1429, + 296, + 1429 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 242, + 261, + 1402, + 261, + 1402, + 332, + 242, + 332 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 243, + 1461, + 1403, + 1461, + 1403, + 1504, + 243, + 1504 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 242, + 1257, + 1173, + 1257, + 1173, + 1293, + 242, + 1293 + ], + "score": 0.909 + }, + { + "category_id": 8, + "poly": [ + 410, + 1690, + 1287, + 1690, + 1287, + 1777, + 410, + 1777 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 237, + 543, + 1394, + 543, + 1394, + 578, + 237, + 578 + ], + "score": 0.889 + }, + { + "category_id": 8, + "poly": [ + 356, + 587, + 1278, + 587, + 1278, + 820, + 356, + 820 + ], + "score": 0.872 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1912, + 1401, + 1912, + 1401, + 1944, + 1351, + 1944 + ], + "score": 0.872 + }, + { + "category_id": 8, + "poly": [ + 372, + 342, + 1320, + 342, + 1320, + 386, + 372, + 386 + ], + "score": 0.862 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1716, + 1401, + 1716, + 1401, + 1747, + 1365, + 1747 + ], + "score": 0.848 + }, + { + "category_id": 9, + "poly": [ + 1365, + 348, + 1401, + 348, + 1401, + 379, + 1365, + 379 + ], + "score": 0.795 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2085, + 840, + 2085 + ], + "score": 0.782 + }, + { + "category_id": 1, + "poly": [ + 358, + 195, + 1182, + 195, + 1182, + 242, + 358, + 242 + ], + "score": 0.705 + }, + { + "category_id": 0, + "poly": [ + 295, + 981, + 952, + 981, + 952, + 1015, + 295, + 1015 + ], + "score": 0.697 + }, + { + "category_id": 1, + "poly": [ + 242, + 1230, + 276, + 1230, + 276, + 1253, + 242, + 1253 + ], + "score": 0.651 + }, + { + "category_id": 2, + "poly": [ + 243, + 209, + 275, + 209, + 275, + 232, + 243, + 232 + ], + "score": 0.526 + }, + { + "category_id": 1, + "poly": [ + 295, + 981, + 952, + 981, + 952, + 1015, + 295, + 1015 + ], + "score": 0.215 + }, + { + "category_id": 1, + "poly": [ + 356, + 587, + 1278, + 587, + 1278, + 820, + 356, + 820 + ], + "score": 0.125 + }, + { + "category_id": 14, + "poly": [ + 644, + 1901, + 1054, + 1901, + 1054, + 2003, + 644, + 2003 + ], + "score": 0.94, + "latex": "d _ { j , i } ^ { * } \\equiv \\sum _ { 1 \\leq h \\leq \\lambda _ { 1 , i } \\atop \\overline { { X } } _ { j } \\in U _ { h , i } } p _ { h , i } + \\sum _ { 1 \\leq h \\leq \\lambda _ { 2 , i } \\atop \\overline { { X } } _ { j } \\in V _ { h , i } } q _ { h , i } ," + }, + { + "category_id": 13, + "poly": [ + 590, + 429, + 718, + 429, + 718, + 464, + 590, + 464 + ], + "score": 0.93, + "latex": "U _ { h , i } \\cup V _ { h ^ { \\prime } , i }" + }, + { + "category_id": 14, + "poly": [ + 359, + 586, + 1278, + 586, + 1278, + 820, + 359, + 820 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\mathcal { X } = \\{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \\} , \\ Z _ { i } = \\{ X _ { 1 } , X _ { 2 } \\} , \\ \\lambda _ { i } = 5 , \\ \\lambda _ { 1 , i } = 2 , \\ \\lambda _ { 2 , i } = 1 } \\\\ & { ( W _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { i } } = ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \\} ) , } \\\\ & { ( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { i } } = ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} , \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} ) , } \\\\ & { ( W _ { h , i } \\cap ( \\mathcal { X } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { i } } = ( \\phi , \\phi , \\phi , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} ) , } \\\\ & { ( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } = ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} ) , \\ ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } = ( \\phi , \\{ X _ { 3 } , X _ { 4 } \\} ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 778, + 848, + 901, + 848, + 901, + 882, + 778, + 882 + ], + "score": 0.93, + "latex": "\\lambda _ { 1 , i } + \\lambda _ { 2 , i }" + }, + { + "category_id": 13, + "poly": [ + 584, + 880, + 744, + 880, + 744, + 919, + 584, + 919 + ], + "score": 0.93, + "latex": "\\mathcal { P } ( \\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i } )" + }, + { + "category_id": 13, + "poly": [ + 940, + 916, + 1015, + 916, + 1015, + 951, + 940, + 951 + ], + "score": 0.93, + "latex": "\\mathcal { O } ( \\lambda _ { i } ^ { 3 } )" + }, + { + "category_id": 13, + "poly": [ + 447, + 842, + 631, + 842, + 631, + 882, + 447, + 882 + ], + "score": 0.93, + "latex": "Z _ { i } \\subseteq \\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i }" + }, + { + "category_id": 14, + "poly": [ + 410, + 1686, + 1287, + 1686, + 1287, + 1779, + 410, + 1779 + ], + "score": 0.92, + "latex": "H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) \\equiv \\sum _ { i = 1 } ^ { n } ( H _ { \\mathrm { s c o r e } } ^ { * ( i ) } ( p . , i , q . , i ) + H _ { \\mathrm { o n e } } ^ { * ( i ) } ( p . , i , q . , i ) ) + H _ { \\mathrm { c y c l e } } ^ { * } ( p , q , r ) ," + }, + { + "category_id": 13, + "poly": [ + 868, + 430, + 983, + 430, + 983, + 464, + 868, + 464 + ], + "score": 0.92, + "latex": "U _ { h , i } , V _ { h ^ { \\prime } , i }" + }, + { + "category_id": 13, + "poly": [ + 671, + 1136, + 842, + 1136, + 842, + 1170, + 671, + 1170 + ], + "score": 0.92, + "latex": "Z _ { i } = Z \\cup \\{ X \\}" + }, + { + "category_id": 13, + "poly": [ + 652, + 1605, + 985, + 1605, + 985, + 1648, + 652, + 1648 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\sum _ { h = 0 } ^ { \\lambda _ { 1 , i } } p _ { h , i } = \\sum _ { h = 0 } ^ { \\lambda _ { 2 , i } } q _ { h , i } = 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 372, + 396, + 488, + 396, + 488, + 427, + 372, + 427 + ], + "score": 0.91, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 1093, + 918, + 1209, + 918, + 1209, + 949, + 1093, + 949 + ], + "score": 0.91, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 1066, + 1612, + 1189, + 1612, + 1189, + 1643, + 1066, + 1643 + ], + "score": 0.91, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 430, + 1027, + 550, + 1027, + 550, + 1067, + 430, + 1067 + ], + "score": 0.91, + "latex": "( W _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 1241, + 1540, + 1401, + 1540, + 1401, + 1575, + 1241, + 1575 + ], + "score": 0.9, + "latex": "p _ { h , i } , q _ { h ^ { \\prime } , i } \\in \\mathbb { B }" + }, + { + "category_id": 13, + "poly": [ + 1199, + 266, + 1318, + 266, + 1318, + 296, + 1199, + 296 + ], + "score": 0.9, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 410, + 1174, + 499, + 1174, + 499, + 1207, + 410, + 1207 + ], + "score": 0.9, + "latex": "Z \\neq Z ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1575, + 826, + 1575, + 826, + 1608, + 373, + 1608 + ], + "score": 0.9, + "latex": "0 \\leq h \\leq \\lambda _ { 1 , i } , 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 2 , i } , 1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 1122, + 502, + 1330, + 502, + 1330, + 537, + 1122, + 537 + ], + "score": 0.89, + "latex": "{ \\textstyle \\sum _ { i = 1 } ^ { n } } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } )" + }, + { + "category_id": 13, + "poly": [ + 822, + 1032, + 1104, + 1032, + 1104, + 1065, + 822, + 1065 + ], + "score": 0.89, + "latex": "\\lambda \\lambda _ { i } , Z ^ { \\prime } \\phi , Z \\phi" + }, + { + "category_id": 13, + "poly": [ + 373, + 261, + 1119, + 261, + 1119, + 301, + 373, + 301 + ], + "score": 0.89, + "latex": "Z _ { i } \\subseteq \\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i } , W _ { 0 , i } = U _ { 0 , i } = V _ { 0 , i } = \\phi , \\lambda _ { i } , \\lambda _ { 1 , i } , \\lambda _ { 2 , i } \\in \\mathbb { N } \\cup \\{ 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 371, + 1789, + 831, + 1789, + 831, + 1833, + 371, + 1833 + ], + "score": 0.89, + "latex": "\\pmb { p } = ( ( p _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } ) _ { i = 1 } ^ { n } , \\pmb { q } = ( ( q _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } ) _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 386, + 1070, + 455, + 1070, + 455, + 1099, + 386, + 1099 + ], + "score": 0.89, + "latex": "d = 1" + }, + { + "category_id": 14, + "poly": [ + 376, + 343, + 1322, + 343, + 1322, + 385, + 376, + 385 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\{ W _ { 0 , i } , \\cdots , W _ { \\lambda _ { i } , i } \\} \\subseteq \\{ U \\cup V \\mid ( U , V ) \\in \\{ U _ { 0 , i } , \\cdots , U _ { \\lambda _ { 1 , i } , i } \\} \\times \\{ V _ { 0 , i } , \\cdots , V _ { \\lambda _ { 2 , i } , i } \\} \\} , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 906, + 1136, + 1279, + 1136, + 1279, + 1171, + 906, + 1171 + ], + "score": 0.89, + "latex": "\\lambda \\lambda _ { 1 , i } + \\lambda _ { 2 , i } , Z ^ { \\prime } Z \\cup \\{ X \\} ." + }, + { + "category_id": 13, + "poly": [ + 563, + 1174, + 658, + 1174, + 658, + 1203, + 563, + 1203 + ], + "score": 0.89, + "latex": "Z Z ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 452, + 1140, + 627, + 1140, + 627, + 1171, + 452, + 1171 + ], + "score": 0.88, + "latex": "\\lambda _ { 1 , i } + \\lambda _ { 2 , i } < \\lambda" + }, + { + "category_id": 13, + "poly": [ + 1043, + 1835, + 1092, + 1835, + 1092, + 1863, + 1043, + 1863 + ], + "score": 0.88, + "latex": "\\mathbf { \\omega } _ { p , q }" + }, + { + "category_id": 13, + "poly": [ + 1137, + 884, + 1168, + 884, + 1168, + 915, + 1137, + 915 + ], + "score": 0.88, + "latex": "Z _ { i }" + }, + { + "category_id": 13, + "poly": [ + 407, + 1532, + 641, + 1532, + 641, + 1575, + 407, + 1575 + ], + "score": 0.87, + "latex": "( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } }" + }, + { + "category_id": 13, + "poly": [ + 487, + 1104, + 655, + 1104, + 655, + 1139, + 487, + 1139 + ], + "score": 0.85, + "latex": "\\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i } \\setminus Z" + }, + { + "category_id": 13, + "poly": [ + 487, + 1067, + 682, + 1067, + 682, + 1104, + 487, + 1104 + ], + "score": 0.85, + "latex": "| \\cup _ { h = 0 } ^ { \\lambda _ { i } } W _ { h , i } | - 1" + }, + { + "category_id": 14, + "poly": [ + 456, + 194, + 1181, + 194, + 1181, + 242, + 456, + 242 + ], + "score": 0.82, + "latex": "\\because ( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( W _ { h , i } \\cap ( \\mathcal { X } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } \\mathrm { ~ i n ~ } ( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } ," + }, + { + "category_id": 13, + "poly": [ + 1164, + 460, + 1399, + 460, + 1399, + 503, + 1164, + 503 + ], + "score": 0.82, + "latex": "( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } }" + }, + { + "category_id": 13, + "poly": [ + 427, + 1106, + 456, + 1106, + 456, + 1133, + 427, + 1133 + ], + "score": 0.8, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 804, + 1532, + 1028, + 1532, + 1028, + 1576, + 804, + 1576 + ], + "score": 0.69, + "latex": "( p _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( q _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } }" + }, + { + "category_id": 13, + "poly": [ + 462, + 195, + 1181, + 195, + 1181, + 241, + 462, + 241 + ], + "score": 0.62, + "latex": "\\begin{array} { r } { ( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( W _ { h , i } \\cap ( \\mathcal { X } \\setminus Z _ { i } ) ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } \\mathrm { ~ i n ~ } ( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 665, + 1033, + 690, + 1033, + 690, + 1060, + 665, + 1060 + ], + "score": 0.61, + "latex": "Z" + }, + { + "category_id": 14, + "poly": [ + 287, + 1299, + 1390, + 1299, + 1390, + 1433, + 287, + 1433 + ], + "score": 0.6, + "latex": "\\begin{array} { r l } & { Z _ { i } = \\phi , \\lambda _ { 1 , i } = 0 , \\lambda _ { 2 , i } = 5 : ( \\phi ) \\times ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 2 } , X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 1 } , X _ { 3 } , X _ { 3 } \\} , } \\\\ & { Z _ { i } = \\{ X _ { 1 } \\} , \\lambda _ { 1 , i } = 1 , \\lambda _ { 2 , i } = 3 : ( \\phi , \\{ X _ { 1 } \\} ) \\times ( \\phi , \\{ X _ { 2 } \\} , \\{ X _ { 3 } , X _ { 4 } \\} , \\{ X _ { 2 } , X _ { 3 } , X _ { 4 } \\} ) , } \\\\ & { Z _ { i } = \\{ X _ { 1 } , X _ { 2 } \\} , \\lambda _ { 1 , i } = 2 , \\lambda _ { 2 , i } = 1 : ( \\phi , \\{ X _ { 1 } \\} , \\{ X _ { 1 } , X _ { 2 } \\} ) \\times ( \\phi , \\{ X _ { 3 } , X _ { 4 } \\} ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 464, + 197, + 641, + 197, + 641, + 240, + 464, + 240 + ], + "score": 0.53, + "latex": "( W _ { h , i } \\cap Z _ { i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } }" + }, + { + "category_id": 14, + "poly": [ + 300, + 1348, + 859, + 1348, + 859, + 1383, + 300, + 1383 + ], + "score": 0.31, + "latex": "Z _ { i } = \\{ X _ { 1 } \\} , \\lambda _ { 1 , i } = 1 , \\lambda _ { 2 , i } = 3 : ( \\phi , \\{ X _ { 1 } \\} ) \\times ( \\phi , \\{ X _ { 1 } \\} )" + }, + { + "category_id": 15, + "poly": [ + 234.0, + 1454.0, + 1407.0, + 1454.0, + 1407.0, + 1512.0, + 234.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 863.0, + 2058.0, + 863.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 978.0, + 951.0, + 978.0, + 951.0, + 1021.0, + 296.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 210.0, + 278.0, + 210.0, + 278.0, + 235.0, + 240.0, + 235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 833.0, + 446.0, + 833.0, + 446.0, + 895.0, + 242.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 833.0, + 777.0, + 833.0, + 777.0, + 895.0, + 632.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 833.0, + 1412.0, + 833.0, + 1412.0, + 895.0, + 902.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 890.0, + 277.0, + 890.0, + 277.0, + 913.0, + 242.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 280.0, + 862.0, + 583.0, + 862.0, + 583.0, + 942.0, + 280.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 862.0, + 1136.0, + 862.0, + 1136.0, + 942.0, + 745.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 862.0, + 1415.0, + 862.0, + 1415.0, + 942.0, + 1169.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 916.0, + 939.0, + 916.0, + 939.0, + 952.0, + 293.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 916.0, + 1092.0, + 916.0, + 1092.0, + 952.0, + 1016.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 916.0, + 1222.0, + 916.0, + 1222.0, + 952.0, + 1210.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 243.0, + 1769.0, + 370.0, + 1769.0, + 370.0, + 1861.0, + 243.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 1769.0, + 1042.0, + 1769.0, + 1042.0, + 1861.0, + 832.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1769.0, + 1422.0, + 1769.0, + 1422.0, + 1861.0, + 1093.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 243.0, + 1826.0, + 1042.0, + 1826.0, + 1042.0, + 1869.0, + 243.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1826.0, + 1406.0, + 1826.0, + 1406.0, + 1869.0, + 1093.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1862.0, + 411.0, + 1862.0, + 411.0, + 1899.0, + 236.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 399.0, + 281.0, + 399.0, + 281.0, + 427.0, + 240.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 394.0, + 371.0, + 394.0, + 371.0, + 430.0, + 291.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 394.0, + 1405.0, + 394.0, + 1405.0, + 430.0, + 489.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 428.0, + 589.0, + 428.0, + 589.0, + 524.0, + 240.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 428.0, + 867.0, + 428.0, + 867.0, + 524.0, + 719.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 428.0, + 1163.0, + 428.0, + 1163.0, + 524.0, + 984.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 428.0, + 1423.0, + 428.0, + 1423.0, + 524.0, + 1400.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 498.0, + 1121.0, + 498.0, + 1121.0, + 542.0, + 291.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 498.0, + 1345.0, + 498.0, + 1345.0, + 542.0, + 1331.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 506.0, + 277.0, + 506.0, + 277.0, + 534.0, + 240.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1521.0, + 406.0, + 1521.0, + 406.0, + 1591.0, + 239.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1521.0, + 803.0, + 1521.0, + 803.0, + 1591.0, + 642.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1521.0, + 1240.0, + 1521.0, + 1240.0, + 1591.0, + 1029.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1521.0, + 1407.0, + 1521.0, + 1407.0, + 1591.0, + 1402.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1576.0, + 280.0, + 1576.0, + 280.0, + 1606.0, + 239.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1572.0, + 372.0, + 1572.0, + 372.0, + 1609.0, + 292.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1572.0, + 1407.0, + 1572.0, + 1407.0, + 1609.0, + 827.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1618.0, + 278.0, + 1618.0, + 278.0, + 1642.0, + 241.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 280.0, + 1590.0, + 651.0, + 1590.0, + 651.0, + 1664.0, + 280.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1590.0, + 1065.0, + 1590.0, + 1065.0, + 1664.0, + 986.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1590.0, + 1417.0, + 1590.0, + 1417.0, + 1664.0, + 1190.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1650.0, + 279.0, + 1650.0, + 279.0, + 1678.0, + 239.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1645.0, + 1408.0, + 1645.0, + 1408.0, + 1681.0, + 292.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.25, + 1532.0, + 522.25, + 1532.0, + 522.25, + 1559.0, + 469.25, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.25, + 1617.5, + 874.25, + 1617.5, + 874.25, + 1635.5, + 826.25, + 1635.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1017.0, + 429.0, + 1017.0, + 429.0, + 1080.0, + 300.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1017.0, + 664.0, + 1017.0, + 664.0, + 1080.0, + 551.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1017.0, + 821.0, + 1017.0, + 821.0, + 1080.0, + 691.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1017.0, + 1112.0, + 1017.0, + 1112.0, + 1080.0, + 1105.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1055.0, + 385.0, + 1055.0, + 385.0, + 1116.0, + 299.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1055.0, + 486.0, + 1055.0, + 486.0, + 1116.0, + 456.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1055.0, + 725.0, + 1055.0, + 725.0, + 1116.0, + 683.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1105.0, + 339.0, + 1105.0, + 339.0, + 1138.0, + 307.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1093.0, + 426.0, + 1093.0, + 426.0, + 1153.0, + 374.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1093.0, + 486.0, + 1093.0, + 486.0, + 1153.0, + 457.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1093.0, + 704.0, + 1093.0, + 704.0, + 1153.0, + 656.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1137.0, + 341.0, + 1137.0, + 341.0, + 1170.0, + 307.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1133.0, + 451.0, + 1133.0, + 451.0, + 1172.0, + 422.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1133.0, + 670.0, + 1133.0, + 670.0, + 1172.0, + 628.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1133.0, + 905.0, + 1133.0, + 905.0, + 1172.0, + 843.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1133.0, + 1285.0, + 1133.0, + 1285.0, + 1172.0, + 1280.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1174.0, + 339.0, + 1174.0, + 339.0, + 1208.0, + 307.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1172.0, + 409.0, + 1172.0, + 409.0, + 1206.0, + 381.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 1172.0, + 562.0, + 1172.0, + 562.0, + 1206.0, + 500.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1172.0, + 786.0, + 1172.0, + 786.0, + 1206.0, + 659.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1391.0, + 1305.0, + 1447.0, + 1305.0, + 1447.0, + 1345.0, + 1391.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 231.0, + 255.0, + 372.0, + 255.0, + 372.0, + 310.0, + 231.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 255.0, + 1198.0, + 255.0, + 1198.0, + 310.0, + 1120.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 255.0, + 1410.0, + 255.0, + 1410.0, + 310.0, + 1319.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 297.0, + 546.0, + 297.0, + 546.0, + 334.0, + 237.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1255.0, + 1178.0, + 1255.0, + 1178.0, + 1296.0, + 235.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 539.0, + 1396.0, + 539.0, + 1396.0, + 583.0, + 237.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 176.0, + 455.0, + 176.0, + 455.0, + 261.0, + 343.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 176.0, + 1198.0, + 176.0, + 1198.0, + 261.0, + 1182.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1230.0, + 278.0, + 1230.0, + 278.0, + 1256.0, + 240.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 978.0, + 951.0, + 978.0, + 951.0, + 1021.0, + 296.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 761.0, + 358.0, + 761.0, + 358.0, + 826.0, + 352.0, + 826.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 346, + 243, + 1352, + 243, + 1352, + 873, + 346, + 873 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 241, + 1408, + 1407, + 1408, + 1407, + 1543, + 241, + 1543 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 245, + 1143, + 1406, + 1143, + 1406, + 1246, + 245, + 1246 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 294, + 934, + 1407, + 934, + 1407, + 1090, + 294, + 1090 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 419, + 1304, + 1280, + 1304, + 1280, + 1402, + 419, + 1402 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 446, + 1638, + 1253, + 1638, + 1253, + 1716, + 446, + 1716 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 248, + 1556, + 1408, + 1556, + 1408, + 1632, + 248, + 1632 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 242, + 1725, + 1411, + 1725, + 1411, + 1801, + 242, + 1801 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 236, + 1940, + 1410, + 1940, + 1410, + 2010, + 236, + 2010 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 325, + 1855, + 1330, + 1855, + 1330, + 1933, + 325, + 1933 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 250, + 1258, + 1078, + 1258, + 1078, + 1293, + 250, + 1293 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 234, + 1812, + 1224, + 1812, + 1224, + 1848, + 234, + 1848 + ], + "score": 0.913 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1648, + 1400, + 1648, + 1400, + 1680, + 1351, + 1680 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1337, + 1400, + 1337, + 1400, + 1369, + 1351, + 1369 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1866, + 1400, + 1866, + 1400, + 1899, + 1352, + 1899 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 840, + 2060, + 859, + 2060, + 859, + 2086, + 840, + 2086 + ], + "score": 0.681 + }, + { + "category_id": 2, + "poly": [ + 840, + 2060, + 859, + 2060, + 859, + 2086, + 840, + 2086 + ], + "score": 0.159 + }, + { + "category_id": 14, + "poly": [ + 420, + 1302, + 1276, + 1302, + 1276, + 1402, + 420, + 1402 + ], + "score": 0.94, + "latex": "H _ { \\mathrm { s c o r r e } } ^ { * ( i ) } ( { \\pmb p } _ { . , i } , { \\pmb q } _ { . , i } ) \\equiv \\sum _ { h = 1 } ^ { \\lambda _ { 1 , i } } s _ { 1 , h , i } p _ { h , i } + \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } s _ { 2 , h , i } q _ { h , i } + \\sum _ { h = 1 } ^ { \\lambda _ { 1 , i } } \\sum _ { h ^ { \\prime } = 1 } ^ { \\lambda _ { 2 , i } } t _ { h , h ^ { \\prime } , i } p _ { h , i } q _ { h ^ { \\prime } , i } ," + }, + { + "category_id": 13, + "poly": [ + 1100, + 1727, + 1193, + 1727, + 1193, + 1761, + 1100, + 1761 + ], + "score": 0.93, + "latex": "\\xi _ { 1 , i } , \\xi _ { 2 , i }" + }, + { + "category_id": 14, + "poly": [ + 325, + 1854, + 1326, + 1854, + 1326, + 1933, + 325, + 1933 + ], + "score": 0.93, + "latex": "H _ { \\mathrm { c y c l e } } ^ { * } ( p , q , r ) \\equiv \\sum _ { 1 \\leq i < j < k \\leq n } \\delta _ { 1 } ^ { * } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) + \\sum _ { 1 \\leq i < j \\leq n } \\delta _ { 2 } ^ { * } ( d _ { i , j } ^ { * } r _ { i , j } + d _ { j , i } ^ { * } ( 1 - r _ { i , j } ) ) ," + }, + { + "category_id": 14, + "poly": [ + 447, + 1637, + 1249, + 1637, + 1249, + 1717, + 447, + 1717 + ], + "score": 0.92, + "latex": "H _ { \\mathrm { o n e } } ^ { * ( i ) } ( p . , . , q . , i ) \\equiv \\sum _ { 1 \\leq h < h ^ { \\prime } \\leq \\lambda _ { 1 , i } } \\xi _ { 1 , i } p _ { h , i } p _ { h ^ { \\prime } , i } + \\sum _ { 1 \\leq h < h ^ { \\prime } \\leq \\lambda _ { 2 , i } } \\xi _ { 2 , i } q _ { h , i } q _ { h ^ { \\prime } , i } ," + }, + { + "category_id": 13, + "poly": [ + 296, + 1758, + 624, + 1758, + 624, + 1803, + 296, + 1803 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\sum _ { h = 0 } ^ { \\lambda _ { 1 , i } } p _ { h , i } = \\sum _ { h = 0 } ^ { \\lambda _ { 2 , i } } q _ { h , i } = 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 851, + 1726, + 1057, + 1726, + 1057, + 1761, + 851, + 1761 + ], + "score": 0.92, + "latex": "0 < \\xi _ { 1 , i } , \\xi _ { 2 , i } \\in \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 883, + 1176, + 1166, + 1176, + 1166, + 1215, + 883, + 1215 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { n } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } ) + \\binom { n } { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 376, + 1146, + 526, + 1146, + 526, + 1178, + 376, + 1178 + ], + "score": 0.92, + "latex": "1 \\leq j , i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 951, + 1943, + 1018, + 1943, + 1018, + 1976, + 951, + 1976 + ], + "score": 0.91, + "latex": "\\delta _ { 1 } ^ { * } , \\delta _ { 2 } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 642, + 1942, + 815, + 1942, + 815, + 1976, + 642, + 1976 + ], + "score": 0.91, + "latex": "0 < \\delta _ { 1 } ^ { * } , \\delta _ { 2 } ^ { * } \\in \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 378, + 1410, + 515, + 1410, + 515, + 1441, + 378, + 1441 + ], + "score": 0.9, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 473, + 1216, + 570, + 1216, + 570, + 1247, + 473, + 1247 + ], + "score": 0.9, + "latex": "p _ { 0 , i } , q _ { 0 , i }" + }, + { + "category_id": 13, + "poly": [ + 375, + 1728, + 500, + 1728, + 500, + 1757, + 375, + 1757 + ], + "score": 0.89, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 1143, + 429, + 1187, + 429, + 1187, + 443, + 1143, + 443 + ], + "score": 0.88, + "latex": "2 \\to 4" + }, + { + "category_id": 13, + "poly": [ + 710, + 418, + 755, + 418, + 755, + 431, + 710, + 431 + ], + "score": 0.87, + "latex": "1 2" + }, + { + "category_id": 13, + "poly": [ + 1003, + 353, + 1047, + 353, + 1047, + 366, + 1003, + 366 + ], + "score": 0.87, + "latex": "2 \\to 3" + }, + { + "category_id": 13, + "poly": [ + 1282, + 533, + 1326, + 533, + 1326, + 546, + 1282, + 546 + ], + "score": 0.86, + "latex": "3 \\to 4" + }, + { + "category_id": 13, + "poly": [ + 296, + 1508, + 701, + 1508, + 701, + 1544, + 296, + 1544 + ], + "score": 0.84, + "latex": "\\log S ^ { ( i ) } ( V _ { h ^ { \\prime } , i } \\mid \\mathcal { D } ) - \\log S ^ { ( i ) } ( \\phi \\mid \\mathcal { D } )" + }, + { + "category_id": 13, + "poly": [ + 864, + 663, + 907, + 663, + 907, + 676, + 864, + 676 + ], + "score": 0.83, + "latex": "1 3" + }, + { + "category_id": 13, + "poly": [ + 297, + 999, + 762, + 999, + 762, + 1031, + 297, + 1031 + ], + "score": 0.83, + "latex": "\\{ 4 \\} , U _ { 1 , 2 } = \\{ 3 \\} , U _ { 1 , 3 } = \\{ 1 \\} , V _ { 1 , 3 } = \\{ 4 \\}" + }, + { + "category_id": 13, + "poly": [ + 1003, + 533, + 1047, + 533, + 1047, + 547, + 1003, + 547 + ], + "score": 0.82, + "latex": "1 4" + }, + { + "category_id": 13, + "poly": [ + 855, + 560, + 916, + 560, + 916, + 577, + 855, + 577 + ], + "score": 0.81, + "latex": "\\{ 4 \\} \\to 1" + }, + { + "category_id": 13, + "poly": [ + 549, + 572, + 610, + 572, + 610, + 589, + 549, + 589 + ], + "score": 0.8, + "latex": "\\left\\{ 2 \\right\\} \\to 1" + }, + { + "category_id": 13, + "poly": [ + 453, + 1588, + 687, + 1588, + 687, + 1630, + 453, + 1630 + ], + "score": 0.8, + "latex": "( U _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } , ( V _ { h , i } ) _ { h = 0 } ^ { \\lambda _ { 2 , i } }" + }, + { + "category_id": 13, + "poly": [ + 1023, + 937, + 1403, + 937, + 1403, + 968, + 1023, + 968 + ], + "score": 0.79, + "latex": "n = 4 , \\lambda _ { 1 , 1 } = 3 , \\lambda _ { 2 , 1 } = 1 , \\lambda _ { 1 , 2 } =" + }, + { + "category_id": 13, + "poly": [ + 701, + 624, + 763, + 624, + 763, + 642, + 701, + 642 + ], + "score": 0.78, + "latex": "\\{ 3 \\} \\to 1" + }, + { + "category_id": 13, + "poly": [ + 854, + 326, + 916, + 326, + 916, + 345, + 854, + 345 + ], + "score": 0.77, + "latex": "\\{ 3 \\} \\to 2" + }, + { + "category_id": 13, + "poly": [ + 310, + 966, + 1407, + 966, + 1407, + 1000, + 310, + 1000 + ], + "score": 0.69, + "latex": "\\bar { \\lambda } _ { 2 , 2 } = 0 , \\lambda _ { 1 , 3 } = \\bar { 1 } , \\lambda _ { 2 , 3 } = 1 , \\lambda _ { 1 , 4 } = 0 , \\lambda _ { 2 , 4 } = 0 , U _ { 1 , 1 } = \\{ 2 , 3 \\} , U _ { 2 , 1 } = \\{ 3 \\} , U _ { 3 , 1 } = \\{ 2 \\} , V _ { 1 , 1 } = \\{ 3 \\} , U _ { 2 , 2 } = \\{ 3 \\} , U _ { 3 , 2 } = \\{ 3 \\} , U _ { 2 , 2 } = \\{ 3 \\} ," + }, + { + "category_id": 13, + "poly": [ + 751, + 1440, + 1404, + 1440, + 1404, + 1482, + 751, + 1482 + ], + "score": 0.67, + "latex": "s _ { 1 , h , i } = \\_ \\log S ^ { ( i ) } ( U _ { h , i } \\mid \\mathcal { D } ) + \\log S ^ { ( i ) } ( \\phi \\mid \\mathcal { D } ) , s _ { 2 , h , i } =" + }, + { + "category_id": 13, + "poly": [ + 378, + 572, + 456, + 572, + 456, + 589, + 378, + 589 + ], + "score": 0.67, + "latex": "\\{ 2 , 3 \\} 1" + }, + { + "category_id": 13, + "poly": [ + 867, + 1002, + 916, + 1002, + 916, + 1029, + 867, + 1029 + ], + "score": 0.6, + "latex": "\\mathbf { \\mu } _ { p , q }" + }, + { + "category_id": 13, + "poly": [ + 310, + 1473, + 1415, + 1473, + 1415, + 1511, + 310, + 1511 + ], + "score": 0.45, + "latex": "\\begin{array} { r } { - \\log S ^ { ( i ) } ( V _ { h , i } \\mid \\mathcal D ) + \\log S ^ { ( i ) } ( \\phi \\mid \\mathcal D ) , t _ { h , h ^ { \\prime } , i } = - \\log S ^ { ( i ) } ( U _ { h , i } \\cup V _ { h ^ { \\prime } , i } \\mid \\mathcal D ) + \\log S ^ { ( i ) } ( U _ { h , i } \\mid \\mathcal D ) + } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1012, + 1003, + 1050, + 1003, + 1050, + 1025, + 1012, + 1025 + ], + "score": 0.26, + "latex": ": \\textbf { { r } }" + }, + { + "category_id": 13, + "poly": [ + 1032, + 1004, + 1050, + 1004, + 1050, + 1024, + 1032, + 1024 + ], + "score": 0.26, + "latex": "\\mathbfit { \\Delta } \\mathbf { r }" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 319.0, + 853.0, + 319.0, + 853.0, + 348.0, + 847.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 319.0, + 924.0, + 319.0, + 924.0, + 348.0, + 917.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 341.0, + 1002.0, + 341.0, + 1002.0, + 373.0, + 994.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 341.0, + 1055.0, + 341.0, + 1055.0, + 373.0, + 1048.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 406.0, + 709.0, + 406.0, + 709.0, + 438.0, + 703.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 406.0, + 764.0, + 406.0, + 764.0, + 438.0, + 756.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 418.0, + 1142.0, + 418.0, + 1142.0, + 448.0, + 1136.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 418.0, + 1196.0, + 418.0, + 1196.0, + 448.0, + 1188.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 526.0, + 1002.0, + 526.0, + 1002.0, + 550.0, + 999.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 526.0, + 1053.0, + 526.0, + 1053.0, + 550.0, + 1048.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 523.0, + 1281.0, + 523.0, + 1281.0, + 553.0, + 1275.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 523.0, + 1335.0, + 523.0, + 1335.0, + 553.0, + 1327.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 562.0, + 377.0, + 562.0, + 377.0, + 593.0, + 369.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 562.0, + 463.0, + 562.0, + 463.0, + 593.0, + 457.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 562.0, + 548.0, + 562.0, + 548.0, + 593.0, + 540.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 562.0, + 618.0, + 562.0, + 618.0, + 593.0, + 611.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 551.0, + 854.0, + 551.0, + 854.0, + 582.0, + 848.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 551.0, + 923.0, + 551.0, + 923.0, + 582.0, + 917.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 616.0, + 700.0, + 616.0, + 700.0, + 647.0, + 695.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 616.0, + 771.0, + 616.0, + 771.0, + 647.0, + 764.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 653.0, + 863.0, + 653.0, + 863.0, + 683.0, + 856.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 653.0, + 916.0, + 653.0, + 916.0, + 683.0, + 908.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 796.0, + 772.0, + 796.0, + 772.0, + 827.0, + 693.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 784.0, + 1204.0, + 784.0, + 1204.0, + 816.0, + 1125.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 929.0, + 1022.0, + 929.0, + 1022.0, + 975.0, + 291.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 929.0, + 1407.0, + 929.0, + 1407.0, + 975.0, + 1404.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 963.0, + 309.0, + 963.0, + 309.0, + 1004.0, + 292.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 993.0, + 866.0, + 993.0, + 866.0, + 1035.0, + 763.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 993.0, + 1011.0, + 993.0, + 1011.0, + 1035.0, + 917.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 993.0, + 1408.0, + 993.0, + 1408.0, + 1035.0, + 1051.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1025.0, + 1406.0, + 1025.0, + 1406.0, + 1063.0, + 295.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1056.0, + 613.0, + 1056.0, + 613.0, + 1093.0, + 295.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 861.0, + 2059.0, + 861.0, + 2092.0, + 839.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 861.0, + 2059.0, + 861.0, + 2092.0, + 839.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1405.0, + 377.0, + 1405.0, + 377.0, + 1446.0, + 238.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1405.0, + 1407.0, + 1405.0, + 1407.0, + 1446.0, + 516.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1447.0, + 279.0, + 1447.0, + 279.0, + 1474.0, + 240.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1434.0, + 750.0, + 1434.0, + 750.0, + 1483.0, + 286.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1405.0, + 1434.0, + 1408.0, + 1434.0, + 1408.0, + 1483.0, + 1405.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1479.0, + 278.0, + 1479.0, + 278.0, + 1506.0, + 240.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1469.0, + 309.0, + 1469.0, + 309.0, + 1515.0, + 285.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1512.0, + 281.0, + 1512.0, + 281.0, + 1539.0, + 242.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1505.0, + 295.0, + 1505.0, + 295.0, + 1544.0, + 287.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1505.0, + 713.0, + 1505.0, + 713.0, + 1544.0, + 702.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1142.0, + 375.0, + 1142.0, + 375.0, + 1181.0, + 236.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1142.0, + 1406.0, + 1142.0, + 1406.0, + 1181.0, + 527.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1184.0, + 280.0, + 1184.0, + 280.0, + 1208.0, + 242.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1174.0, + 882.0, + 1174.0, + 882.0, + 1217.0, + 287.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1174.0, + 1407.0, + 1174.0, + 1407.0, + 1217.0, + 1167.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1211.0, + 472.0, + 1211.0, + 472.0, + 1247.0, + 241.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1211.0, + 801.0, + 1211.0, + 801.0, + 1247.0, + 571.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 229.0, + 1554.0, + 452.0, + 1554.0, + 452.0, + 1647.0, + 229.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1554.0, + 1404.0, + 1554.0, + 1404.0, + 1647.0, + 688.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1722.0, + 295.0, + 1722.0, + 295.0, + 1821.0, + 240.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1722.0, + 850.0, + 1722.0, + 850.0, + 1821.0, + 625.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1722.0, + 1099.0, + 1722.0, + 1099.0, + 1821.0, + 1058.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1722.0, + 1408.0, + 1722.0, + 1408.0, + 1821.0, + 1194.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1768.5, + 514.0, + 1768.5, + 514.0, + 1790.5, + 466.0, + 1790.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1939.0, + 641.0, + 1939.0, + 641.0, + 1980.0, + 240.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1939.0, + 950.0, + 1939.0, + 950.0, + 1980.0, + 816.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1939.0, + 1407.0, + 1939.0, + 1407.0, + 1980.0, + 1019.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1972.0, + 543.0, + 1972.0, + 543.0, + 2014.0, + 235.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1255.0, + 1082.0, + 1255.0, + 1082.0, + 1299.0, + 239.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1810.0, + 1223.0, + 1810.0, + 1223.0, + 1851.0, + 235.0, + 1851.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 237, + 897, + 1408, + 897, + 1408, + 1500, + 237, + 1500 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 241, + 383, + 1405, + 383, + 1405, + 581, + 241, + 581 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 242, + 266, + 1407, + 266, + 1407, + 367, + 242, + 367 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 571, + 591, + 1125, + 591, + 1125, + 682, + 571, + 682 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 240, + 1854, + 1405, + 1854, + 1405, + 2016, + 240, + 2016 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 245, + 1749, + 1403, + 1749, + 1403, + 1851, + 245, + 1851 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 571, + 690, + 1124, + 690, + 1124, + 782, + 571, + 782 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 241, + 785, + 1404, + 785, + 1404, + 886, + 241, + 886 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 343, + 1506, + 1305, + 1506, + 1305, + 1744, + 343, + 1744 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1351, + 719, + 1401, + 719, + 1401, + 751, + 1351, + 751 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1523, + 1401, + 1523, + 1401, + 1554, + 1351, + 1554 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1351, + 619, + 1401, + 619, + 1401, + 651, + 1351, + 651 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 249, + 197, + 1029, + 197, + 1029, + 238, + 249, + 238 + ], + "score": 0.623 + }, + { + "category_id": 2, + "poly": [ + 839, + 2061, + 860, + 2061, + 860, + 2086, + 839, + 2086 + ], + "score": 0.584 + }, + { + "category_id": 2, + "poly": [ + 249, + 197, + 1029, + 197, + 1029, + 238, + 249, + 238 + ], + "score": 0.375 + }, + { + "category_id": 2, + "poly": [ + 839, + 2061, + 860, + 2061, + 860, + 2086, + 839, + 2086 + ], + "score": 0.318 + }, + { + "category_id": 8, + "poly": [ + 574, + 1506, + 1002, + 1506, + 1002, + 1571, + 574, + 1571 + ], + "score": 0.092 + }, + { + "category_id": 14, + "poly": [ + 343, + 1501, + 1303, + 1501, + 1303, + 1745, + 343, + 1745 + ], + "score": 0.94, + "latex": "\\begin{array} { c } { \\displaystyle \\operatorname* { m a x } _ { 1 \\leq i \\leq n } \\operatorname* { m a x } \\{ \\eta _ { 1 , i } , \\eta _ { 2 , i } \\} < \\delta _ { 1 } ^ { * } < \\displaystyle \\frac { \\delta _ { 2 } ^ { * } } { n - 2 } , } \\\\ { \\displaystyle \\operatorname* { m a x } _ { 1 \\leq j \\leq n } \\displaystyle \\operatorname* { m a x } _ { 1 \\leq i \\leq 1 _ { 1 , i } } \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 1 , i } } \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 2 , i } } ( - s _ { 1 , h , i } - t _ { h , h ^ { \\prime \\prime } , i } + s _ { 1 , h ^ { \\prime } , i } + t _ { h ^ { \\prime } , h ^ { \\prime \\prime } , i } ) , } \\\\ { \\displaystyle \\eta _ { 1 , i } \\equiv \\displaystyle \\operatorname* { m a x } _ { 1 \\leq j \\leq n } \\displaystyle \\frac { \\operatorname* { m a x } _ { i } } { X _ { j } \\in U _ { h , i } } \\operatorname* { m a x } _ { X _ { j } \\in U _ { h , i } } \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime \\prime } \\leq \\lambda _ { 1 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \\prime \\prime } , h , i } + s _ { 2 , h ^ { \\prime } , i } + t _ { h ^ { \\prime \\prime } , h ^ { \\prime \\prime } , i } ) , } \\\\ { \\displaystyle \\eta _ { 2 , i } \\equiv \\displaystyle \\operatorname* { m a x } _ { 1 \\leq j \\leq n } \\displaystyle \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 2 , i } } \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime } \\leq \\lambda _ { 2 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \\prime \\prime } , h , i } + s _ { 2 , h ^ { \\prime } , i } + t _ { h ^ { \\prime \\prime } , h ^ { \\prime } , i } ) , } \\\\ { \\displaystyle \\sum _ { k \\geq \\ell \\leq n } \\displaystyle \\sum _ { X _ { j } \\in V _ { h , i } } \\sum _ { j \\in V _ { h ^ { \\prime } } , i \\subset V _ { h , i } } ( \\operatorname* { m a x } _ { 0 \\leq h ^ { \\prime \\prime } \\leq \\lambda _ { 1 , i } } ( - s _ { 2 , h , i } - t _ { h ^ { \\prime \\prime } , h , i } + s _ { 2 , h ^ { \\prime } , i } + t _ { h ^ { \\prime \\prime } , h ^ { \\prime \\prime } , i } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 345, + 505, + 482, + 505, + 482, + 551, + 345, + 551 + ], + "score": 0.93, + "latex": "p _ { h , i } ^ { ( h ^ { * } , i ^ { * } ) } = 0" + }, + { + "category_id": 13, + "poly": [ + 1193, + 817, + 1325, + 817, + 1325, + 854, + 1193, + 854 + ], + "score": 0.93, + "latex": "\\mathcal { O } ( \\lambda _ { 1 , i } \\lambda _ { 2 , i } )" + }, + { + "category_id": 13, + "poly": [ + 1225, + 380, + 1365, + 380, + 1365, + 424, + 1225, + 424 + ], + "score": 0.92, + "latex": "\\scriptstyle \\sum _ { h = 1 } ^ { \\lambda _ { 1 , i ^ { * } } } p _ { h , i ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 1066, + 463, + 1396, + 463, + 1396, + 508, + 1066, + 508 + ], + "score": 0.92, + "latex": "\\pmb { p } ^ { ( h ^ { * } , i ^ { * } ) } = ( ( p _ { h , i } ^ { ( h ^ { * } , i ^ { * } ) } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } ) _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 376, + 999, + 706, + 999, + 706, + 1045, + 376, + 1045 + ], + "score": 0.92, + "latex": "r ^ { ( i ^ { * } , j ^ { * } ) } = ( r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } ) _ { 1 \\leq i < j \\leq n }" + }, + { + "category_id": 13, + "poly": [ + 813, + 1197, + 901, + 1197, + 901, + 1229, + 813, + 1229 + ], + "score": 0.92, + "latex": "( 1 , 1 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 872, + 900, + 1108, + 900, + 1108, + 936, + 872, + 936 + ], + "score": 0.91, + "latex": "( r _ { i ^ { * } , j ^ { * } } , d _ { j ^ { * } , i ^ { * } } ^ { * } , d _ { i ^ { * } , j ^ { * } } ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 401, + 1468, + 467, + 1468, + 467, + 1501, + 401, + 1501 + ], + "score": 0.91, + "latex": "\\delta _ { 1 } ^ { * } , \\delta _ { 2 } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 614, + 549, + 706, + 549, + 706, + 581, + 614, + 581 + ], + "score": 0.91, + "latex": "\\xi _ { 1 , i } , \\xi _ { 2 , i }" + }, + { + "category_id": 13, + "poly": [ + 1011, + 1041, + 1099, + 1041, + 1099, + 1073, + 1011, + 1073 + ], + "score": 0.91, + "latex": "( 1 , 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 1118, + 1261, + 1293, + 1261, + 1293, + 1295, + 1118, + 1295 + ], + "score": 0.91, + "latex": "H _ { \\mathrm { t r a n s } } ( r ) = 0" + }, + { + "category_id": 13, + "poly": [ + 471, + 966, + 559, + 966, + 559, + 999, + 471, + 999 + ], + "score": 0.91, + "latex": "( 0 , 1 , 0 )" + }, + { + "category_id": 13, + "poly": [ + 432, + 1292, + 767, + 1292, + 767, + 1328, + 432, + 1328 + ], + "score": 0.91, + "latex": "H _ { \\mathrm { t r a n s } } ( \\mathbf { \\bar { r } } ) > H _ { \\mathrm { t r a n s } } ( { r ^ { ( i ^ { * } , j ^ { * } ) } } )" + }, + { + "category_id": 13, + "poly": [ + 954, + 1295, + 1043, + 1295, + 1043, + 1328, + 954, + 1328 + ], + "score": 0.91, + "latex": "( 1 , 1 , 0 )" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1369, + 1301, + 1369, + 1301, + 1401, + 1213, + 1401 + ], + "score": 0.91, + "latex": "( 0 , 0 , 0 )" + }, + { + "category_id": 13, + "poly": [ + 719, + 420, + 1049, + 420, + 1049, + 465, + 719, + 465 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { p _ { h ^ { * } , i ^ { * } } = 1 , \\sum _ { h = 1 } ^ { \\lambda _ { 1 , i ^ { * } } } p _ { h , i ^ { * } } > 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 348, + 1397, + 849, + 1397, + 849, + 1435, + 348, + 1435 + ], + "score": 0.91, + "latex": "\\bar { H _ { \\mathrm { t o t a l } } ^ { * } } ( p , \\bar { q } , r ) - \\bar { H _ { \\mathrm { t o t a l } } ^ { * } } ( p , q , r ^ { ( i ^ { * } , j ^ { * } ) } ) \\geq \\delta _ { 1 } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1228, + 365, + 1228, + 365, + 1262, + 297, + 1262 + ], + "score": 0.91, + "latex": "\\delta _ { 1 } ^ { * } , \\delta _ { 2 } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 573, + 1895, + 806, + 1895, + 806, + 1929, + 573, + 1929 + ], + "score": 0.9, + "latex": "1 \\leq i ^ { * } < j ^ { * } \\leq n" + }, + { + "category_id": 13, + "poly": [ + 1087, + 1369, + 1176, + 1369, + 1176, + 1401, + 1087, + 1401 + ], + "score": 0.9, + "latex": "( 1 , 0 , 0 )" + }, + { + "category_id": 13, + "poly": [ + 340, + 1751, + 422, + 1751, + 422, + 1780, + 340, + 1780 + ], + "score": 0.9, + "latex": "n \\geq 3" + }, + { + "category_id": 13, + "poly": [ + 374, + 786, + 497, + 786, + 497, + 817, + 374, + 817 + ], + "score": 0.9, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 825, + 1370, + 913, + 1370, + 913, + 1401, + 825, + 1401 + ], + "score": 0.9, + "latex": "( 0 , 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 772, + 1000, + 992, + 1000, + 992, + 1043, + 772, + 1043 + ], + "score": 0.9, + "latex": "r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = 1 - r _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 974, + 1891, + 1312, + 1891, + 1312, + 1929, + 974, + 1929 + ], + "score": 0.9, + "latex": "H _ { \\mathrm { t r a n s } } ( r ) > H _ { \\mathrm { t r a n s } } ( r ^ { ( i ^ { * } , j ^ { * } ) } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1113, + 648, + 1113, + 648, + 1155, + 297, + 1155 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { s _ { 1 , h ^ { * * } , i ^ { * } } - \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * * } , h , i ^ { * } } q _ { h , i ^ { * } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1219, + 1111, + 1402, + 1111, + 1402, + 1157, + 1219, + 1157 + ], + "score": 0.9, + "latex": "p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = 0" + }, + { + "category_id": 13, + "poly": [ + 297, + 1813, + 659, + 1813, + 659, + 1852, + 297, + 1852 + ], + "score": 0.89, + "latex": "\\begin{array} { r l } { { \\mathcal { O } ( \\sum _ { i = 1 } ^ { n } n \\lambda _ { 1 , i } \\lambda _ { 2 , i } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } ) ) } \\quad } & { { } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1206, + 428, + 1406, + 428, + 1406, + 462, + 1206, + 462 + ], + "score": 0.89, + "latex": "H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) -" + }, + { + "category_id": 13, + "poly": [ + 1342, + 1401, + 1371, + 1401, + 1371, + 1434, + 1342, + 1434 + ], + "score": 0.89, + "latex": "{ { \\delta } _ { 1 } ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 737, + 1112, + 1158, + 1112, + 1158, + 1157, + 737, + 1157 + ], + "score": 0.88, + "latex": "\\pmb { p } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = ( ( p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } ) _ { h = 0 } ^ { \\lambda _ { 1 , i } } ) _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1969, + 616, + 1969, + 616, + 2014, + 297, + 2014 + ], + "score": 0.88, + "latex": "r ^ { ( i ^ { * } , j ^ { * } ) } = ( r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } ) _ { 1 \\leq i < j \\leq n }" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1928, + 1397, + 1928, + 1397, + 1970, + 1099, + 1970 + ], + "score": 0.88, + "latex": "r = ( r _ { i , j } ) _ { 1 \\leq i < j \\leq n } \\in \\mathbb { B } ^ { \\binom { n } { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 295, + 852, + 413, + 852, + 413, + 884, + 295, + 884 + ], + "score": 0.87, + "latex": "1 \\leq i \\leq n" + }, + { + "category_id": 13, + "poly": [ + 1205, + 1295, + 1406, + 1295, + 1406, + 1329, + 1205, + 1329 + ], + "score": 0.87, + "latex": "H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) \\sim" + }, + { + "category_id": 13, + "poly": [ + 297, + 464, + 983, + 464, + 983, + 504, + 297, + 504 + ], + "score": 0.86, + "latex": "\\begin{array} { r } { H _ { \\mathrm { t o t a l } } ^ { * } ( p ^ { ( h ^ { * } , i ^ { * } ) } , q , r ) \\geq \\xi _ { 1 , i ^ { * } } + s _ { 1 , h ^ { * } , i ^ { * } } + \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 447, + 1073, + 1404, + 1073, + 1404, + 1112, + 447, + 1112 + ], + "score": 0.86, + "latex": "\\begin{array} { r } { H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) - H _ { \\mathrm { t o t a l } } ^ { * } ( p ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } , q , r ) \\ge \\delta _ { 2 } ^ { * } + s _ { 1 , h ^ { * } , i ^ { * } } + \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } - \\frac { 1 - \\lambda _ { 1 } } { \\lambda _ { 2 } } | \\nabla p _ { h } ( q _ { h } ^ { * } , h ^ { * } , i ^ { * } ) | _ { \\mathbb { H } } ( q _ { h } ^ { * } , h ^ { * } , i ^ { * } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 728, + 966, + 1395, + 966, + 1395, + 1001, + 728, + 1001 + ], + "score": 0.86, + "latex": "H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ) - H _ { \\mathrm { t o t a l } } ^ { * } ( p , q , r ^ { ( i ^ { * } , j ^ { * } ) } ) \\geq \\delta _ { 2 } ^ { * } - ( n - 2 ) \\delta _ { 1 } ^ { * }" + }, + { + "category_id": 14, + "poly": [ + 569, + 584, + 1125, + 584, + 1125, + 785, + 569, + 785 + ], + "score": 0.85, + "latex": "\\begin{array} { r l r } & { } & { \\displaystyle \\operatorname* { m a x } _ { 0 \\le h \\le \\lambda _ { 1 , i } } \\bigl ( - s _ { 1 , h , i } - \\sum _ { h ^ { \\prime } = 1 } ^ { \\lambda _ { 2 , i } } \\operatorname* { m i n } \\bigl \\{ 0 , t _ { h , h ^ { \\prime } , i } \\bigr \\} \\bigr ) \\bigr ) < \\xi _ { 1 , i } , } \\\\ & { } & { \\displaystyle \\operatorname* { m a x } _ { 0 \\le h \\le \\lambda _ { 2 , i } } \\bigl ( - s _ { 2 , h , i } - \\sum _ { h ^ { \\prime } = 1 } ^ { \\lambda _ { 1 , i } } \\operatorname* { m i n } \\bigl \\{ 0 , t _ { h ^ { \\prime } , h , i } \\bigr \\} \\bigr ) \\bigr ) < \\xi _ { 2 , i } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 647, + 937, + 1396, + 937, + 1396, + 964, + 647, + 964 + ], + "score": 0.84, + "latex": "X _ { j ^ { * } } ~ \\in ~ U _ { h ^ { * } , i ^ { * } } , X _ { j ^ { * } } ~ \\notin ~ U _ { h ^ { * * } , i ^ { * } } ~ \\subset ~ U _ { h ^ { * } , i ^ { * } } , p _ { h ^ { * } , i ^ { * } } ~ = ~ 1 , p _ { h ^ { * * } , i ^ { * } } ~ = ~ 0" + }, + { + "category_id": 13, + "poly": [ + 627, + 1854, + 1196, + 1854, + 1196, + 1894, + 627, + 1894 + ], + "score": 0.84, + "latex": "\\begin{array} { r } { H _ { \\mathrm { t r a n s } } ( { \\pmb r } ) \\equiv \\sum _ { 1 \\leq i < j < k \\leq n } R ( r _ { i , j } , r _ { j , k } , r _ { i , k } ) > 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1308, + 1042, + 1395, + 1042, + 1395, + 1073, + 1308, + 1073 + ], + "score": 0.84, + "latex": "( 0 , 1 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 312, + 1328, + 1403, + 1328, + 1403, + 1369, + 312, + 1369 + ], + "score": 0.84, + "latex": "\\begin{array} { r } { T _ { \\mathrm { t o t a l } } ^ { * } \\big ( p ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } , q , r ^ { ( i ^ { * } , j ^ { * } ) } \\big ) \\geq \\delta _ { 1 } ^ { * } + s _ { 1 , h ^ { * } , i ^ { * } } + \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * } , h , i ^ { * } } q _ { h , i ^ { * } } - s _ { 1 , h ^ { * * } , i ^ { * } } - \\sum _ { h = 1 } ^ { \\lambda _ { 2 , i } } t _ { h ^ { * * } , h , i ^ { * } } q _ { h , i ^ { * } } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 511, + 506, + 869, + 506, + 869, + 550, + 511, + 550 + ], + "score": 0.84, + "latex": "( h , i ) = ( h ^ { * } , i ^ { * } ) , p _ { h , i } ^ { ( h ^ { * } , i ^ { * } ) } = p _ { h , i }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1934, + 858, + 1934, + 858, + 1969, + 297, + 1969 + ], + "score": 0.82, + "latex": "R ( r _ { 1 } , r _ { 2 } , r _ { 3 } ) = r _ { 1 } r _ { 2 } ( 1 - r _ { 3 } ) + ( 1 - r _ { 1 } ) ( 1 - r _ { 2 } ) r _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 326, + 1160, + 521, + 1160, + 521, + 1196, + 326, + 1196 + ], + "score": 0.81, + "latex": "( h , i ) \\ = \\ ( h ^ { * } , i ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 537, + 1154, + 728, + 1154, + 728, + 1200, + 537, + 1200 + ], + "score": 0.81, + "latex": "p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = 1" + }, + { + "category_id": 13, + "poly": [ + 1310, + 519, + 1331, + 519, + 1331, + 546, + 1310, + 546 + ], + "score": 0.81, + "latex": "\\pmb { p }" + }, + { + "category_id": 13, + "poly": [ + 1083, + 1439, + 1104, + 1439, + 1104, + 1466, + 1083, + 1466 + ], + "score": 0.8, + "latex": "\\pmb { p }" + }, + { + "category_id": 13, + "poly": [ + 1381, + 519, + 1401, + 519, + 1401, + 546, + 1381, + 546 + ], + "score": 0.78, + "latex": "\\pmb q" + }, + { + "category_id": 13, + "poly": [ + 1158, + 1439, + 1178, + 1439, + 1178, + 1466, + 1158, + 1466 + ], + "score": 0.78, + "latex": "\\pmb q" + }, + { + "category_id": 13, + "poly": [ + 935, + 1936, + 1088, + 1936, + 1088, + 1968, + 935, + 1968 + ], + "score": 0.76, + "latex": "r _ { 1 } , r _ { 2 } , r _ { 3 } \\in \\mathbb { B }" + }, + { + "category_id": 13, + "poly": [ + 763, + 1161, + 968, + 1161, + 968, + 1196, + 763, + 1196 + ], + "score": 0.73, + "latex": "( h , i ) \\ = \\ ( h ^ { * * } , i ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 1231, + 1000, + 1400, + 1000, + 1400, + 1042, + 1231, + 1042 + ], + "score": 0.73, + "latex": "r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = r _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 1097, + 1969, + 1255, + 1969, + 1255, + 2014, + 1097, + 2014 + ], + "score": 0.72, + "latex": "r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = r _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 985, + 1155, + 1205, + 1155, + 1205, + 1200, + 985, + 1200 + ], + "score": 0.67, + "latex": "p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } = p _ { h , i }" + }, + { + "category_id": 13, + "poly": [ + 754, + 1268, + 774, + 1268, + 774, + 1290, + 754, + 1290 + ], + "score": 0.67, + "latex": "\\mathbfit { \\Delta } \\mathbf { r }" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1005, + 1215, + 1005, + 1215, + 1040, + 1028, + 1040 + ], + "score": 0.65, + "latex": "( i , j ) = ( i ^ { * } , j ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 677, + 1970, + 1084, + 1970, + 1084, + 2014, + 677, + 2014 + ], + "score": 0.44, + "latex": "r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = 1 - r _ { i , j } \\ i f ( i , j ) = ( i ^ { * } , j ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1002, + 1400, + 1002, + 1400, + 1044, + 1028, + 1044 + ], + "score": 0.31, + "latex": "( i , j ) = ( i ^ { * } , j ^ { * } ) , r _ { i , j } ^ { ( i ^ { * } , j ^ { * } ) } = r _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 319, + 1156, + 719, + 1156, + 719, + 1200, + 319, + 1200 + ], + "score": 0.29, + "latex": "( h , i ) \\ = \\ ( h ^ { * } , i ^ { * } ) , p _ { h , i } ^ { ( h ^ { * } , h ^ { * * } , i ^ { * } ) } \\ = \\ 1" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 193.0, + 1030.0, + 193.0, + 1030.0, + 246.0, + 236.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2060.0, + 862.0, + 2060.0, + 862.0, + 2090.0, + 840.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 193.0, + 1030.0, + 193.0, + 1030.0, + 246.0, + 236.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2060.0, + 862.0, + 2060.0, + 862.0, + 2090.0, + 840.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 904.0, + 280.0, + 904.0, + 280.0, + 932.0, + 239.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 893.0, + 871.0, + 893.0, + 871.0, + 939.0, + 292.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 893.0, + 1406.0, + 893.0, + 1406.0, + 939.0, + 1109.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 936.0, + 280.0, + 936.0, + 280.0, + 966.0, + 239.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 928.0, + 646.0, + 928.0, + 646.0, + 971.0, + 290.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 928.0, + 1410.0, + 928.0, + 1410.0, + 971.0, + 1397.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 970.0, + 280.0, + 970.0, + 280.0, + 1040.0, + 240.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 958.0, + 375.0, + 958.0, + 375.0, + 1071.0, + 286.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 958.0, + 727.0, + 958.0, + 727.0, + 1071.0, + 707.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 958.0, + 1413.0, + 958.0, + 1413.0, + 1071.0, + 1401.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1043.0, + 280.0, + 1043.0, + 280.0, + 1072.0, + 239.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1032.0, + 1010.0, + 1032.0, + 1010.0, + 1078.0, + 290.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1032.0, + 1307.0, + 1032.0, + 1307.0, + 1078.0, + 1100.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1032.0, + 1410.0, + 1032.0, + 1410.0, + 1078.0, + 1396.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1082.0, + 278.0, + 1082.0, + 278.0, + 1111.0, + 239.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1067.0, + 446.0, + 1067.0, + 446.0, + 1120.0, + 288.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1122.0, + 278.0, + 1122.0, + 278.0, + 1152.0, + 240.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 280.0, + 1094.0, + 296.0, + 1094.0, + 296.0, + 1174.0, + 280.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1094.0, + 736.0, + 1094.0, + 736.0, + 1174.0, + 649.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1094.0, + 1218.0, + 1094.0, + 1218.0, + 1174.0, + 1159.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1094.0, + 1408.0, + 1094.0, + 1408.0, + 1174.0, + 1403.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1166.0, + 278.0, + 1166.0, + 278.0, + 1195.0, + 239.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1146.0, + 318.0, + 1146.0, + 318.0, + 1211.0, + 288.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1142.0, + 1412.0, + 1142.0, + 1412.0, + 1203.0, + 1206.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1200.0, + 278.0, + 1200.0, + 278.0, + 1230.0, + 239.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1193.0, + 812.0, + 1193.0, + 812.0, + 1230.0, + 294.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1193.0, + 1408.0, + 1193.0, + 1408.0, + 1230.0, + 902.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1232.0, + 278.0, + 1232.0, + 278.0, + 1262.0, + 239.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1226.0, + 296.0, + 1226.0, + 296.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1226.0, + 1406.0, + 1226.0, + 1406.0, + 1266.0, + 366.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1266.0, + 280.0, + 1266.0, + 280.0, + 1295.0, + 239.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1262.0, + 753.0, + 1262.0, + 753.0, + 1298.0, + 293.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 1262.0, + 1117.0, + 1262.0, + 1117.0, + 1298.0, + 775.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1262.0, + 1409.0, + 1262.0, + 1409.0, + 1298.0, + 1294.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1299.0, + 280.0, + 1299.0, + 280.0, + 1329.0, + 239.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1284.0, + 431.0, + 1284.0, + 431.0, + 1338.0, + 289.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1284.0, + 953.0, + 1284.0, + 953.0, + 1338.0, + 768.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 1284.0, + 1204.0, + 1284.0, + 1204.0, + 1338.0, + 1044.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1338.0, + 278.0, + 1338.0, + 278.0, + 1368.0, + 239.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1311.0, + 311.0, + 1311.0, + 311.0, + 1388.0, + 292.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1311.0, + 1421.0, + 1311.0, + 1421.0, + 1388.0, + 1404.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1370.0, + 280.0, + 1370.0, + 280.0, + 1400.0, + 239.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1364.0, + 824.0, + 1364.0, + 824.0, + 1404.0, + 293.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1364.0, + 1086.0, + 1364.0, + 1086.0, + 1404.0, + 914.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1364.0, + 1212.0, + 1364.0, + 1212.0, + 1404.0, + 1177.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1364.0, + 1408.0, + 1364.0, + 1408.0, + 1404.0, + 1302.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1404.0, + 280.0, + 1404.0, + 280.0, + 1434.0, + 239.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1394.0, + 347.0, + 1394.0, + 347.0, + 1440.0, + 289.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 1394.0, + 1341.0, + 1394.0, + 1341.0, + 1440.0, + 850.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1394.0, + 1409.0, + 1394.0, + 1409.0, + 1440.0, + 1372.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1438.0, + 280.0, + 1438.0, + 280.0, + 1466.0, + 240.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 1082.0, + 1432.0, + 1082.0, + 1469.0, + 294.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1432.0, + 1157.0, + 1432.0, + 1157.0, + 1469.0, + 1105.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1432.0, + 1405.0, + 1432.0, + 1405.0, + 1469.0, + 1179.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1471.0, + 280.0, + 1471.0, + 280.0, + 1500.0, + 240.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1466.0, + 400.0, + 1466.0, + 400.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1466.0, + 601.0, + 1466.0, + 601.0, + 1503.0, + 468.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 1006.5, + 1404.0, + 1006.5, + 1404.0, + 1041.5, + 1320.0, + 1041.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1087.0, + 1232.0, + 1087.0, + 1232.0, + 1118.0, + 1152.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1108.0, + 1141.0, + 1108.0, + 1141.0, + 1143.0, + 1054.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.75, + 1148.0, + 1041.75, + 1148.0, + 1041.75, + 1201.0, + 540.75, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1337.5, + 1242.0, + 1337.5, + 1242.0, + 1364.5, + 1182.0, + 1364.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 367.0, + 1224.0, + 367.0, + 1224.0, + 438.0, + 284.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 367.0, + 1416.0, + 367.0, + 1416.0, + 438.0, + 1366.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 392.0, + 281.0, + 392.0, + 281.0, + 421.0, + 238.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 416.0, + 296.0, + 416.0, + 296.0, + 525.0, + 238.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 416.0, + 1065.0, + 416.0, + 1065.0, + 525.0, + 1050.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1407.0, + 416.0, + 1413.0, + 416.0, + 1413.0, + 525.0, + 1407.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 517.0, + 279.0, + 517.0, + 279.0, + 546.0, + 238.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 498.0, + 1309.0, + 498.0, + 1309.0, + 557.0, + 870.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 498.0, + 1380.0, + 498.0, + 1380.0, + 557.0, + 1332.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 498.0, + 1409.0, + 498.0, + 1409.0, + 557.0, + 1402.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 551.0, + 279.0, + 551.0, + 279.0, + 579.0, + 237.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 542.0, + 613.0, + 542.0, + 613.0, + 583.0, + 291.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 542.0, + 844.0, + 542.0, + 844.0, + 583.0, + 707.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.25, + 472.0, + 839.25, + 472.0, + 839.25, + 493.5, + 779.25, + 493.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.25, + 499.0, + 762.25, + 499.0, + 762.25, + 554.0, + 349.25, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.25, + 508.5, + 397.25, + 508.5, + 397.25, + 554.0, + 292.25, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 273.0, + 278.0, + 273.0, + 278.0, + 300.0, + 241.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 268.0, + 1404.0, + 268.0, + 1404.0, + 303.0, + 292.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 309.0, + 278.0, + 309.0, + 278.0, + 332.0, + 242.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 301.0, + 1404.0, + 301.0, + 1404.0, + 337.0, + 291.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 340.0, + 280.0, + 340.0, + 280.0, + 368.0, + 241.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 335.0, + 911.0, + 335.0, + 911.0, + 368.0, + 289.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1857.0, + 280.0, + 1857.0, + 280.0, + 1887.0, + 240.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1848.0, + 626.0, + 1848.0, + 626.0, + 1898.0, + 288.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1848.0, + 1410.0, + 1848.0, + 1410.0, + 1898.0, + 1197.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1897.0, + 280.0, + 1897.0, + 280.0, + 1928.0, + 240.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1889.0, + 572.0, + 1889.0, + 572.0, + 1931.0, + 289.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1889.0, + 973.0, + 1889.0, + 973.0, + 1931.0, + 807.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1889.0, + 1408.0, + 1889.0, + 1408.0, + 1931.0, + 1313.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1936.0, + 281.0, + 1936.0, + 281.0, + 2007.0, + 239.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1926.0, + 296.0, + 1926.0, + 296.0, + 2040.0, + 285.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1926.0, + 1096.0, + 1926.0, + 1096.0, + 2040.0, + 1089.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1926.0, + 1411.0, + 1926.0, + 1411.0, + 2040.0, + 1398.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1951.5, + 482.0, + 1951.5, + 482.0, + 2023.5, + 284.0, + 2023.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1753.0, + 280.0, + 1753.0, + 280.0, + 1781.0, + 241.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1748.0, + 339.0, + 1748.0, + 339.0, + 1784.0, + 291.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1748.0, + 1408.0, + 1748.0, + 1408.0, + 1784.0, + 423.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1789.0, + 279.0, + 1789.0, + 279.0, + 1813.0, + 242.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1782.0, + 1406.0, + 1782.0, + 1406.0, + 1816.0, + 289.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1820.0, + 279.0, + 1820.0, + 279.0, + 1848.0, + 241.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1809.0, + 296.0, + 1809.0, + 296.0, + 1859.0, + 288.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1809.0, + 672.0, + 1809.0, + 672.0, + 1859.0, + 660.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 789.0, + 280.0, + 789.0, + 280.0, + 817.0, + 241.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 787.0, + 373.0, + 787.0, + 373.0, + 818.0, + 290.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 787.0, + 1405.0, + 787.0, + 1405.0, + 818.0, + 498.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 234.0, + 818.0, + 1192.0, + 818.0, + 1192.0, + 853.0, + 234.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 818.0, + 1405.0, + 818.0, + 1405.0, + 853.0, + 1326.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 852.0, + 294.0, + 852.0, + 294.0, + 888.0, + 237.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 852.0, + 426.0, + 852.0, + 426.0, + 888.0, + 414.0, + 888.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 303, + 249, + 1391, + 249, + 1391, + 540, + 303, + 540 + ], + "score": 0.981, + "html": "
Namenm∑=1 |IIil n∑=1 βi( -1)Ωi=1X n *
N=100N= 1000N = 10000
insurance273529843538834036
water3256610083165216735
alarm37446509182922729081
barley484841140051813101552
hailfinder5646626561446924277
hepar2706123145348376654782
" + }, + { + "category_id": 1, + "poly": [ + 295, + 1724, + 1405, + 1724, + 1405, + 1927, + 295, + 1927 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 241, + 626, + 1408, + 626, + 1408, + 864, + 241, + 864 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 277, + 971, + 1405, + 971, + 1405, + 1236, + 277, + 1236 + ], + "score": 0.961 + }, + { + "category_id": 2, + "poly": [ + 331, + 1946, + 740, + 1946, + 740, + 2008, + 331, + 2008 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 243, + 1665, + 923, + 1665, + 923, + 1700, + 243, + 1700 + ], + "score": 0.931 + }, + { + "category_id": 7, + "poly": [ + 323, + 543, + 731, + 543, + 731, + 573, + 323, + 573 + ], + "score": 0.92 + }, + { + "category_id": 6, + "poly": [ + 460, + 217, + 1234, + 217, + 1234, + 247, + 460, + 247 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 255, + 900, + 657, + 900, + 657, + 939, + 255, + 939 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 841, + 2060, + 859, + 2060, + 859, + 2085, + 841, + 2085 + ], + "score": 0.719 + }, + { + "category_id": 1, + "poly": [ + 307, + 1311, + 1309, + 1311, + 1309, + 1612, + 307, + 1612 + ], + "score": 0.687 + }, + { + "category_id": 6, + "poly": [ + 297, + 1269, + 927, + 1269, + 927, + 1301, + 297, + 1301 + ], + "score": 0.583 + }, + { + "category_id": 0, + "poly": [ + 297, + 1269, + 927, + 1269, + 927, + 1301, + 297, + 1301 + ], + "score": 0.277 + }, + { + "category_id": 2, + "poly": [ + 841, + 2060, + 859, + 2060, + 859, + 2085, + 841, + 2085 + ], + "score": 0.208 + }, + { + "category_id": 5, + "poly": [ + 307, + 1311, + 1309, + 1311, + 1309, + 1612, + 307, + 1612 + ], + "score": 0.17, + "html": "
1: Input: D,i,m Output: L Initialize: L ← {φ},L' ← {Φ},L" ←
2: 3:ford=1tomdo for W in L'do
4:forXinX\\{Xi}\\Wdo
5:if Si(W'|D)< Si(WU {X} |D) for all W' CWU{X},W' ∈ L then
6:L"←L"U{WU{X}}.
7:if L"≠Φ then L ← LUL",L'← L",L" ←Φelse break
8:forWinLdo
9:if there exist W' C W that satisfies Si(W|D) ≤ Si(W'|D) then L ← L\\{W}
" + }, + { + "category_id": 13, + "poly": [ + 455, + 1889, + 665, + 1889, + 665, + 1927, + 455, + 1927 + ], + "score": 0.96, + "latex": "{ \\textstyle \\sum _ { i = 1 } ^ { n } } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } )" + }, + { + "category_id": 13, + "poly": [ + 722, + 1791, + 821, + 1791, + 821, + 1825, + 722, + 1825 + ], + "score": 0.93, + "latex": "n ( n - 1 )" + }, + { + "category_id": 13, + "poly": [ + 1114, + 1070, + 1262, + 1070, + 1262, + 1106, + 1114, + 1106 + ], + "score": 0.93, + "latex": "\\mathcal { P } ( \\mathcal { X } \\backslash \\{ X _ { i } \\} )" + }, + { + "category_id": 13, + "poly": [ + 1186, + 1756, + 1286, + 1756, + 1286, + 1793, + 1186, + 1793 + ], + "score": 0.92, + "latex": "\\textstyle \\sum _ { i = 1 } ^ { n } \\lambda _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1026, + 1790, + 1192, + 1790, + 1192, + 1825, + 1026, + 1825 + ], + "score": 0.92, + "latex": "O ( n ( n - 1 ) ^ { \\frac { m } { 2 } } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 791, + 656, + 791, + 656, + 827, + 297, + 827 + ], + "score": 0.92, + "latex": "H _ { \\mathrm { t r a n s } } ( \\pmb { r } ) - H _ { \\mathrm { t r a n s } } ( \\pmb { r } ^ { ( 1 , 3 ) } ) = 3 ." + }, + { + "category_id": 13, + "poly": [ + 1003, + 627, + 1302, + 627, + 1302, + 662, + 1003, + 662 + ], + "score": 0.92, + "latex": "( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) = ( 1 , 1 , 0 )" + }, + { + "category_id": 13, + "poly": [ + 549, + 826, + 903, + 826, + 903, + 861, + 549, + 861 + ], + "score": 0.91, + "latex": "( i ^ { * } , j ^ { * } ) \\in \\{ ( 1 , 2 ) , ( 2 , 3 ) , ( 1 , 3 ) \\}" + }, + { + "category_id": 13, + "poly": [ + 987, + 793, + 1362, + 793, + 1362, + 827, + 987, + 827 + ], + "score": 0.91, + "latex": "H _ { \\mathrm { t r a n s } } ( \\pmb { r } ) - H _ { \\mathrm { t r a n s } } ( \\pmb { r } ^ { ( i ^ { * } , j ^ { * } ) } ) > 0" + }, + { + "category_id": 13, + "poly": [ + 410, + 1502, + 500, + 1502, + 500, + 1536, + 410, + 1536 + ], + "score": 0.9, + "latex": "{ \\mathcal { L } } ^ { \\prime \\prime } \\neq \\phi" + }, + { + "category_id": 13, + "poly": [ + 336, + 1825, + 415, + 1825, + 415, + 1854, + 336, + 1854 + ], + "score": 0.89, + "latex": "m > 2" + }, + { + "category_id": 13, + "poly": [ + 785, + 1568, + 1073, + 1568, + 1073, + 1604, + 785, + 1604 + ], + "score": 0.89, + "latex": "S _ { i } ( W \\mid \\mathcal { D } ) \\le S _ { i } ( W ^ { \\prime } \\mid \\mathcal { D } )" + }, + { + "category_id": 13, + "poly": [ + 942, + 1434, + 1235, + 1434, + 1235, + 1469, + 942, + 1469 + ], + "score": 0.89, + "latex": "W ^ { \\prime } \\subset W \\cup \\{ X \\} , W ^ { \\prime } \\in { \\mathcal { L } }" + }, + { + "category_id": 13, + "poly": [ + 785, + 1312, + 1127, + 1312, + 1127, + 1348, + 785, + 1348 + ], + "score": 0.89, + "latex": "{ \\mathcal { L } } \\gets \\{ \\phi \\} , { \\mathcal { L } } ^ { \\prime } \\gets \\{ \\phi \\} , { \\mathcal { L } } ^ { \\prime \\prime } \\gets \\phi" + }, + { + "category_id": 13, + "poly": [ + 1203, + 1039, + 1276, + 1039, + 1276, + 1067, + 1203, + 1067 + ], + "score": 0.89, + "latex": "\\alpha = 1" + }, + { + "category_id": 13, + "poly": [ + 531, + 1569, + 638, + 1569, + 638, + 1599, + 531, + 1599 + ], + "score": 0.89, + "latex": "W ^ { \\prime } \\subset W" + }, + { + "category_id": 13, + "poly": [ + 409, + 662, + 1407, + 662, + 1407, + 695, + 409, + 695 + ], + "score": 0.88, + "latex": "R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( 1 - r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) + R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( r _ { 1 , 2 } , 1 - r _ { 2 , 3 } , r _ { 1 , 3 } ) + R ( r _ { 1 , 2 } , 1 - r _ { 2 , 3 } , r _ { 1 , 3 } )" + }, + { + "category_id": 13, + "poly": [ + 562, + 1503, + 931, + 1503, + 931, + 1536, + 562, + 1536 + ], + "score": 0.87, + "latex": "\\mathcal { L } \\gets \\mathcal { L } \\cup \\mathcal { L } ^ { \\prime \\prime } , \\mathcal { L } ^ { \\prime } \\mathcal { L } ^ { \\prime \\prime } , \\mathcal { L } ^ { \\prime \\prime } \\phi" + }, + { + "category_id": 13, + "poly": [ + 393, + 763, + 457, + 763, + 457, + 789, + 393, + 789 + ], + "score": 0.87, + "latex": "3 < i" + }, + { + "category_id": 13, + "poly": [ + 297, + 696, + 817, + 696, + 817, + 728, + 297, + 728 + ], + "score": 0.86, + "latex": "R ( r _ { 1 , 2 } , r _ { 2 , 3 } , r _ { 1 , 3 } ) - R ( r _ { 1 , 2 } , r _ { 2 , 3 } , 1 - r _ { 1 , 3 } ) = 3 ." + }, + { + "category_id": 13, + "poly": [ + 529, + 1405, + 699, + 1405, + 699, + 1437, + 529, + 1437 + ], + "score": 0.86, + "latex": "\\chi \\setminus \\{ X _ { i } \\} \\setminus W" + }, + { + "category_id": 13, + "poly": [ + 386, + 1346, + 454, + 1346, + 454, + 1373, + 386, + 1373 + ], + "score": 0.85, + "latex": "d = 1" + }, + { + "category_id": 13, + "poly": [ + 493, + 1375, + 523, + 1375, + 523, + 1403, + 493, + 1403 + ], + "score": 0.84, + "latex": "\\mathcal { L } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 506, + 1467, + 793, + 1467, + 793, + 1499, + 506, + 1499 + ], + "score": 0.83, + "latex": "{ \\mathcal { L } } ^ { \\prime \\prime } \\gets { \\mathcal { L } } ^ { \\prime \\prime } \\cup \\{ W \\cup \\{ X \\} \\}" + }, + { + "category_id": 13, + "poly": [ + 1136, + 1568, + 1304, + 1568, + 1304, + 1604, + 1136, + 1604 + ], + "score": 0.82, + "latex": "{ \\mathcal { L } } \\gets { \\mathcal { L } } \\setminus \\{ W \\}" + }, + { + "category_id": 13, + "poly": [ + 1110, + 696, + 1401, + 696, + 1401, + 727, + 1110, + 727 + ], + "score": 0.82, + "latex": "R ( r _ { 1 , 2 } , r _ { 2 , i } , r _ { 1 , i } ) - R ( 1 -" + }, + { + "category_id": 13, + "poly": [ + 494, + 1436, + 864, + 1436, + 864, + 1466, + 494, + 1466 + ], + "score": 0.8, + "latex": "S _ { i } ( W ^ { \\prime } \\mid \\mathcal { D } ) < S _ { i } ( W \\cup \\{ X \\} \\mid \\mathcal { D } )" + }, + { + "category_id": 13, + "poly": [ + 469, + 1407, + 497, + 1407, + 497, + 1433, + 469, + 1433 + ], + "score": 0.75, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 732, + 761, + 1397, + 761, + 1397, + 791, + 732, + 791 + ], + "score": 0.67, + "latex": "H _ { \\mathrm { t r a n s } } ( \\pmb { r } ) - H _ { \\mathrm { t r a n s } } ( \\pmb { r } ^ { ( 1 , 2 ) } ) + H _ { \\mathrm { t r a n s } } ( \\pmb { r } ) - H _ { \\mathrm { t r a n s } } ( \\pmb { r } ^ { ( 2 , 3 ) } ) +" + }, + { + "category_id": 13, + "poly": [ + 451, + 1540, + 474, + 1540, + 474, + 1567, + 451, + 1567 + ], + "score": 0.62, + "latex": "\\mathcal { L }" + }, + { + "category_id": 13, + "poly": [ + 486, + 1350, + 514, + 1350, + 514, + 1372, + 486, + 1372 + ], + "score": 0.57, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 297, + 1005, + 404, + 1005, + 404, + 1036, + 297, + 1036 + ], + "score": 0.55, + "latex": "N = 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 428, + 1376, + 461, + 1376, + 461, + 1403, + 428, + 1403 + ], + "score": 0.52, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 427, + 1315, + 516, + 1315, + 516, + 1345, + 427, + 1345 + ], + "score": 0.47, + "latex": "\\mathcal { D } , i , m" + }, + { + "category_id": 13, + "poly": [ + 630, + 1315, + 653, + 1315, + 653, + 1343, + 630, + 1343 + ], + "score": 0.45, + "latex": "\\mathcal { L }" + }, + { + "category_id": 13, + "poly": [ + 386, + 1538, + 419, + 1538, + 419, + 1567, + 386, + 1567 + ], + "score": 0.44, + "latex": "W" + }, + { + "category_id": 13, + "poly": [ + 297, + 729, + 840, + 729, + 840, + 762, + 297, + 762 + ], + "score": 0.33, + "latex": "r _ { 1 , 2 } , r _ { 2 , i } , r _ { 1 , i } ) + R ( r _ { 2 , 3 } , r _ { 3 , i } , r _ { 2 , i } ) - R ( 1 - r _ { 2 , 3 } , r _ { 3 , i } )" + }, + { + "category_id": 13, + "poly": [ + 785, + 728, + 1343, + 728, + 1343, + 760, + 785, + 760 + ], + "score": 0.29, + "latex": "{ } _ { 3 } , r _ { 3 , i } , r _ { 2 , i } ) + R ( r _ { 1 , 3 } , r _ { 3 , i } , r _ { 1 , i } ) - R ( 1 - r _ { 1 , 3 } , r _ { 3 , i } , r _ { 1 } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1005, + 553, + 1005, + 553, + 1037, + 298, + 1037 + ], + "score": 0.27, + "latex": "N = 1 0 0 , 1 0 0 0 , 1 0 0 0 0" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1941.0, + 739.0, + 1941.0, + 739.0, + 1980.0, + 332.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1973.0, + 619.0, + 1973.0, + 619.0, + 2012.0, + 331.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1662.0, + 925.0, + 1662.0, + 925.0, + 1706.0, + 238.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 540.0, + 733.0, + 540.0, + 733.0, + 575.0, + 339.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 210.0, + 1240.0, + 210.0, + 1240.0, + 256.0, + 458.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 897.0, + 661.0, + 897.0, + 661.0, + 947.0, + 241.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 840.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1266.0, + 928.0, + 1266.0, + 928.0, + 1306.0, + 296.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1266.0, + 928.0, + 1266.0, + 928.0, + 1306.0, + 296.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 840.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1726.0, + 1403.0, + 1726.0, + 1403.0, + 1759.0, + 295.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1741.0, + 1185.0, + 1741.0, + 1185.0, + 1808.0, + 284.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1741.0, + 1416.0, + 1741.0, + 1416.0, + 1808.0, + 1287.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1792.0, + 721.0, + 1792.0, + 721.0, + 1825.0, + 292.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1792.0, + 1025.0, + 1792.0, + 1025.0, + 1825.0, + 822.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1792.0, + 1406.0, + 1792.0, + 1406.0, + 1825.0, + 1193.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1824.0, + 335.0, + 1824.0, + 335.0, + 1857.0, + 295.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1824.0, + 1405.0, + 1824.0, + 1405.0, + 1857.0, + 416.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1857.0, + 1405.0, + 1857.0, + 1405.0, + 1894.0, + 295.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1887.0, + 454.0, + 1887.0, + 454.0, + 1932.0, + 293.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1887.0, + 1407.0, + 1887.0, + 1407.0, + 1932.0, + 666.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 630.0, + 279.0, + 630.0, + 279.0, + 659.0, + 239.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 624.0, + 1002.0, + 624.0, + 1002.0, + 664.0, + 291.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 624.0, + 1408.0, + 624.0, + 1408.0, + 664.0, + 1303.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 663.0, + 278.0, + 663.0, + 278.0, + 694.0, + 239.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 657.0, + 408.0, + 657.0, + 408.0, + 700.0, + 290.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 696.0, + 279.0, + 696.0, + 279.0, + 727.0, + 239.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 691.0, + 1109.0, + 691.0, + 1109.0, + 732.0, + 818.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 691.0, + 1407.0, + 691.0, + 1407.0, + 732.0, + 1402.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 730.0, + 279.0, + 730.0, + 279.0, + 760.0, + 239.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 724.0, + 296.0, + 724.0, + 296.0, + 766.0, + 290.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 724.0, + 1408.0, + 724.0, + 1408.0, + 766.0, + 1344.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 763.0, + 281.0, + 763.0, + 281.0, + 794.0, + 239.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 754.0, + 392.0, + 754.0, + 392.0, + 796.0, + 289.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 754.0, + 731.0, + 754.0, + 731.0, + 796.0, + 458.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 754.0, + 1412.0, + 754.0, + 1412.0, + 796.0, + 1398.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 796.0, + 278.0, + 796.0, + 278.0, + 826.0, + 239.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 785.0, + 296.0, + 785.0, + 296.0, + 833.0, + 290.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 785.0, + 986.0, + 785.0, + 986.0, + 833.0, + 657.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 785.0, + 1412.0, + 785.0, + 1412.0, + 833.0, + 1363.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 831.0, + 281.0, + 831.0, + 281.0, + 860.0, + 240.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 824.0, + 548.0, + 824.0, + 548.0, + 865.0, + 290.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 824.0, + 916.0, + 824.0, + 916.0, + 865.0, + 904.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 826.0, + 1407.0, + 826.0, + 1407.0, + 860.0, + 1372.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 972.0, + 1405.0, + 972.0, + 1405.0, + 1007.0, + 295.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1006.0, + 1404.0, + 1006.0, + 1404.0, + 1038.0, + 554.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1038.0, + 1202.0, + 1038.0, + 1202.0, + 1073.0, + 295.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1038.0, + 1404.0, + 1038.0, + 1404.0, + 1073.0, + 1277.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1072.0, + 1113.0, + 1072.0, + 1113.0, + 1107.0, + 295.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1072.0, + 1404.0, + 1072.0, + 1404.0, + 1107.0, + 1263.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1106.0, + 1403.0, + 1106.0, + 1403.0, + 1138.0, + 296.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1136.0, + 1406.0, + 1136.0, + 1406.0, + 1175.0, + 294.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1171.0, + 1403.0, + 1171.0, + 1403.0, + 1205.0, + 295.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1204.0, + 567.0, + 1204.0, + 567.0, + 1236.0, + 293.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1311.0, + 426.0, + 1311.0, + 426.0, + 1350.0, + 308.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1311.0, + 629.0, + 1311.0, + 629.0, + 1350.0, + 517.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1311.0, + 784.0, + 1311.0, + 784.0, + 1350.0, + 654.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1311.0, + 1132.0, + 1311.0, + 1132.0, + 1350.0, + 1128.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1342.0, + 385.0, + 1342.0, + 385.0, + 1378.0, + 306.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1342.0, + 485.0, + 1342.0, + 485.0, + 1378.0, + 455.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1342.0, + 555.0, + 1342.0, + 555.0, + 1378.0, + 515.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1374.0, + 340.0, + 1374.0, + 340.0, + 1408.0, + 308.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1373.0, + 427.0, + 1373.0, + 427.0, + 1406.0, + 382.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1373.0, + 492.0, + 1373.0, + 492.0, + 1406.0, + 462.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1373.0, + 563.0, + 1373.0, + 563.0, + 1406.0, + 524.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1405.0, + 340.0, + 1405.0, + 340.0, + 1439.0, + 306.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1401.0, + 468.0, + 1401.0, + 468.0, + 1441.0, + 422.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1401.0, + 528.0, + 1401.0, + 528.0, + 1441.0, + 498.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1401.0, + 741.0, + 1401.0, + 741.0, + 1441.0, + 700.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1434.0, + 340.0, + 1434.0, + 340.0, + 1468.0, + 306.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1430.0, + 493.0, + 1430.0, + 493.0, + 1471.0, + 461.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1430.0, + 941.0, + 1430.0, + 941.0, + 1471.0, + 865.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1430.0, + 1299.0, + 1430.0, + 1299.0, + 1471.0, + 1236.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1466.0, + 340.0, + 1466.0, + 340.0, + 1499.0, + 308.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1460.0, + 807.0, + 1460.0, + 807.0, + 1500.0, + 794.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1503.0, + 340.0, + 1503.0, + 340.0, + 1537.0, + 308.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1499.0, + 409.0, + 1499.0, + 409.0, + 1539.0, + 379.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 1499.0, + 561.0, + 1499.0, + 561.0, + 1539.0, + 501.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1499.0, + 1062.0, + 1499.0, + 1062.0, + 1539.0, + 932.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1536.0, + 385.0, + 1536.0, + 385.0, + 1572.0, + 305.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1536.0, + 450.0, + 1536.0, + 450.0, + 1572.0, + 420.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1536.0, + 516.0, + 1536.0, + 516.0, + 1572.0, + 475.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1570.0, + 342.0, + 1570.0, + 342.0, + 1603.0, + 306.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1563.0, + 530.0, + 1563.0, + 530.0, + 1606.0, + 378.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1563.0, + 784.0, + 1563.0, + 784.0, + 1606.0, + 639.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1563.0, + 1135.0, + 1563.0, + 1135.0, + 1606.0, + 1074.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1563.0, + 1313.0, + 1563.0, + 1313.0, + 1606.0, + 1305.0, + 1606.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 322, + 248, + 1372, + 248, + 1372, + 540, + 322, + 540 + ], + "score": 0.982, + "html": "
Name£ λi/n(n-1)²Σ=1 （1,i+λ2,i）/∑=1 1
N=100N=1000N = 10000N = 100N= 1000N = 10000
insurance0.098730.246771.127420.613670.472850.32476
Water0.000970.001260.004290.726800.705880.44014
alarm0.038140.047380.189380.453320.355370.21617
barley0.001710.002920.014640.767170.755380.54149
hailfinder0.000850.004090.025250.827730.601780.33365
hepar20.000210.000030.000210.496940.633460.31284
" + }, + { + "category_id": 5, + "poly": [ + 299, + 658, + 1403, + 658, + 1403, + 950, + 299, + 950 + ], + "score": 0.981, + "html": "
Name∑i=1(1,i+ 入2,i）+(2） *∑=1(入1,i + λ2,i)(入1,i +λ2,i + 1) + (2)
N= 100N =1000N = 10000N= 100N= 1000N = 10000
insurance56676716613375902385482
water613648820143417205881
alarm1489147226283676127985169004
barley1247136219686758344624796
hailfinder1659195729672212708480578
hepar24777283639104499169164136939
" + }, + { + "category_id": 1, + "poly": [ + 296, + 1475, + 1405, + 1475, + 1405, + 1808, + 296, + 1808 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1294, + 1404, + 1294, + 1404, + 1460, + 297, + 1460 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 275, + 1909, + 1401, + 1909, + 1401, + 2008, + 275, + 2008 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 284, + 1046, + 1399, + 1046, + 1399, + 1112, + 284, + 1112 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 300, + 1212, + 1400, + 1212, + 1400, + 1280, + 300, + 1280 + ], + "score": 0.949 + }, + { + "category_id": 7, + "poly": [ + 349, + 543, + 800, + 543, + 800, + 573, + 349, + 573 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 294, + 1152, + 750, + 1152, + 750, + 1186, + 294, + 1186 + ], + "score": 0.913 + }, + { + "category_id": 7, + "poly": [ + 324, + 953, + 775, + 953, + 775, + 983, + 324, + 983 + ], + "score": 0.912 + }, + { + "category_id": 6, + "poly": [ + 405, + 217, + 1288, + 217, + 1288, + 247, + 405, + 247 + ], + "score": 0.911 + }, + { + "category_id": 6, + "poly": [ + 301, + 625, + 1396, + 625, + 1396, + 657, + 301, + 657 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 294, + 1848, + 600, + 1848, + 600, + 1880, + 294, + 1880 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 858, + 2061, + 858, + 2085, + 840, + 2085 + ], + "score": 0.8 + }, + { + "category_id": 2, + "poly": [ + 238, + 952, + 279, + 952, + 279, + 1978, + 238, + 1978 + ], + "score": 0.104 + }, + { + "category_id": 2, + "poly": [ + 241, + 1039, + 280, + 1039, + 280, + 1988, + 241, + 1988 + ], + "score": 0.1 + }, + { + "category_id": 13, + "poly": [ + 1093, + 1574, + 1172, + 1574, + 1172, + 1606, + 1093, + 1606 + ], + "score": 0.92, + "latex": "I ^ { 2 } + I" + }, + { + "category_id": 13, + "poly": [ + 853, + 1641, + 1326, + 1641, + 1326, + 1678, + 853, + 1678 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { n } ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } ) ( \\lambda _ { 1 , i } + \\lambda _ { 2 , i } + 1 ) + \\binom { n } { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 847, + 1942, + 982, + 1942, + 982, + 1971, + 847, + 1971 + ], + "score": 0.88, + "latex": "N = 1 0 0 0 0" + }, + { + "category_id": 13, + "poly": [ + 1181, + 1613, + 1230, + 1613, + 1230, + 1642, + 1181, + 1642 + ], + "score": 0.84, + "latex": "\\mathbf { \\Delta } _ { p , q }" + }, + { + "category_id": 13, + "poly": [ + 1215, + 1577, + 1234, + 1577, + 1234, + 1604, + 1215, + 1604 + ], + "score": 0.73, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 805, + 1615, + 826, + 1615, + 826, + 1637, + 805, + 1637 + ], + "score": 0.73, + "latex": "n" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 539.0, + 806.0, + 539.0, + 806.0, + 577.0, + 344.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1147.0, + 754.0, + 1147.0, + 754.0, + 1192.0, + 293.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 949.0, + 781.0, + 949.0, + 781.0, + 987.0, + 319.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 213.0, + 1294.0, + 213.0, + 1294.0, + 253.0, + 402.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 623.0, + 1394.0, + 623.0, + 1394.0, + 662.0, + 302.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1844.0, + 603.0, + 1844.0, + 603.0, + 1884.0, + 295.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2058.0, + 859.0, + 2058.0, + 859.0, + 2089.0, + 836.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 1053.0, + 277.0, + 1053.0, + 277.0, + 1084.0, + 241.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1085.0, + 278.0, + 1085.0, + 278.0, + 1123.0, + 238.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1158.0, + 277.0, + 1158.0, + 277.0, + 1196.0, + 238.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1218.0, + 278.0, + 1218.0, + 278.0, + 1254.0, + 238.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1253.0, + 278.0, + 1253.0, + 278.0, + 1291.0, + 238.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1298.0, + 278.0, + 1298.0, + 278.0, + 1334.0, + 238.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1334.0, + 278.0, + 1334.0, + 278.0, + 1369.0, + 238.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1365.0, + 278.0, + 1365.0, + 278.0, + 1402.0, + 238.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1399.0, + 278.0, + 1399.0, + 278.0, + 1436.0, + 238.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1432.0, + 281.0, + 1432.0, + 281.0, + 1473.0, + 236.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1480.0, + 278.0, + 1480.0, + 278.0, + 1517.0, + 238.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1514.0, + 279.0, + 1514.0, + 279.0, + 1549.0, + 238.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1549.0, + 277.0, + 1549.0, + 277.0, + 1581.0, + 239.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1582.0, + 278.0, + 1582.0, + 278.0, + 1614.0, + 239.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1615.0, + 278.0, + 1615.0, + 278.0, + 1647.0, + 239.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1647.0, + 279.0, + 1647.0, + 279.0, + 1682.0, + 238.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1682.0, + 278.0, + 1682.0, + 278.0, + 1714.0, + 239.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1715.0, + 278.0, + 1715.0, + 278.0, + 1746.0, + 239.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1748.0, + 279.0, + 1748.0, + 279.0, + 1784.0, + 238.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1779.0, + 279.0, + 1779.0, + 279.0, + 1816.0, + 238.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1852.0, + 278.0, + 1852.0, + 278.0, + 1887.0, + 236.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1915.0, + 277.0, + 1915.0, + 277.0, + 1947.0, + 239.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1948.0, + 276.0, + 1948.0, + 276.0, + 1981.0, + 239.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1052.0, + 279.0, + 1052.0, + 279.0, + 1085.0, + 238.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1083.0, + 280.0, + 1083.0, + 280.0, + 1121.0, + 237.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1157.0, + 278.0, + 1157.0, + 278.0, + 1193.0, + 238.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1218.0, + 278.0, + 1218.0, + 278.0, + 1251.0, + 238.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1251.0, + 279.0, + 1251.0, + 279.0, + 1292.0, + 237.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1300.0, + 279.0, + 1300.0, + 279.0, + 1334.0, + 238.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1332.0, + 279.0, + 1332.0, + 279.0, + 1367.0, + 238.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1366.0, + 279.0, + 1366.0, + 279.0, + 1400.0, + 238.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1400.0, + 278.0, + 1400.0, + 278.0, + 1435.0, + 238.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1432.0, + 280.0, + 1432.0, + 280.0, + 1471.0, + 237.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1479.0, + 279.0, + 1479.0, + 279.0, + 1518.0, + 237.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1515.0, + 279.0, + 1515.0, + 279.0, + 1548.0, + 238.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1548.0, + 278.0, + 1548.0, + 278.0, + 1582.0, + 238.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1581.0, + 279.0, + 1581.0, + 279.0, + 1614.0, + 238.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1614.0, + 279.0, + 1614.0, + 279.0, + 1649.0, + 238.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1647.0, + 280.0, + 1647.0, + 280.0, + 1681.0, + 238.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1681.0, + 280.0, + 1681.0, + 280.0, + 1715.0, + 238.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1715.0, + 280.0, + 1715.0, + 280.0, + 1747.0, + 238.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1747.0, + 280.0, + 1747.0, + 280.0, + 1780.0, + 238.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1779.0, + 281.0, + 1779.0, + 281.0, + 1816.0, + 236.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1852.0, + 280.0, + 1852.0, + 280.0, + 1885.0, + 237.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 1916.0, + 277.0, + 1916.0, + 277.0, + 1945.0, + 239.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1947.0, + 278.0, + 1947.0, + 278.0, + 1980.0, + 238.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1477.0, + 1403.0, + 1477.0, + 1403.0, + 1511.0, + 296.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1511.0, + 1405.0, + 1511.0, + 1405.0, + 1546.0, + 295.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1541.0, + 1405.0, + 1541.0, + 1405.0, + 1580.0, + 294.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1576.0, + 1092.0, + 1576.0, + 1092.0, + 1611.0, + 294.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1576.0, + 1214.0, + 1576.0, + 1214.0, + 1611.0, + 1173.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1576.0, + 1403.0, + 1576.0, + 1403.0, + 1611.0, + 1235.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1608.0, + 804.0, + 1608.0, + 804.0, + 1645.0, + 293.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1608.0, + 1180.0, + 1608.0, + 1180.0, + 1645.0, + 827.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1608.0, + 1405.0, + 1608.0, + 1405.0, + 1645.0, + 1231.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1640.0, + 852.0, + 1640.0, + 852.0, + 1682.0, + 293.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 1640.0, + 1406.0, + 1640.0, + 1406.0, + 1682.0, + 1327.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1678.0, + 1403.0, + 1678.0, + 1403.0, + 1709.0, + 296.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1709.0, + 1403.0, + 1709.0, + 1403.0, + 1744.0, + 294.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1744.0, + 1406.0, + 1744.0, + 1406.0, + 1778.0, + 295.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1776.0, + 648.0, + 1776.0, + 648.0, + 1811.0, + 295.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1295.0, + 1405.0, + 1295.0, + 1405.0, + 1328.0, + 295.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1363.0, + 294.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1362.0, + 1405.0, + 1362.0, + 1405.0, + 1395.0, + 295.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 1406.0, + 1392.0, + 1406.0, + 1429.0, + 294.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1428.0, + 732.0, + 1428.0, + 732.0, + 1464.0, + 294.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1944.0, + 295.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 846.0, + 1942.0, + 846.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1977.0, + 983.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1975.0, + 722.0, + 1975.0, + 722.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1045.0, + 1405.0, + 1045.0, + 1405.0, + 1084.0, + 293.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1080.0, + 1091.0, + 1080.0, + 1091.0, + 1115.0, + 293.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1209.0, + 1404.0, + 1209.0, + 1404.0, + 1249.0, + 294.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1243.0, + 584.0, + 1243.0, + 584.0, + 1286.0, + 294.0, + 1286.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 208, + 1368, + 208, + 1368, + 752, + 301, + 752 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 296, + 767, + 1406, + 767, + 1406, + 923, + 296, + 923 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 296, + 1479, + 1405, + 1479, + 1405, + 1745, + 296, + 1745 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 1760, + 1404, + 1760, + 1404, + 1959, + 298, + 1959 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 292, + 1017, + 1405, + 1017, + 1405, + 1284, + 292, + 1284 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 296, + 1299, + 1401, + 1299, + 1401, + 1365, + 296, + 1365 + ], + "score": 0.944 + }, + { + "category_id": 2, + "poly": [ + 328, + 1978, + 1220, + 1978, + 1220, + 2008, + 328, + 2008 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 293, + 1408, + 510, + 1408, + 510, + 1446, + 293, + 1446 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2084, + 840, + 2084 + ], + "score": 0.809 + }, + { + "category_id": 2, + "poly": [ + 239, + 1015, + 279, + 1015, + 279, + 1968, + 239, + 1968 + ], + "score": 0.099 + }, + { + "category_id": 13, + "poly": [ + 732, + 797, + 1176, + 797, + 1176, + 836, + 732, + 836 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { \\bar { n } } ( \\log S ^ { ( i ) } ( \\Pi _ { i } \\vert \\mathcal { \\bar { D } } ) - \\log S ^ { ( i ) } ( \\phi \\vert \\mathcal { D } ) ) } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 210.0, + 522.0, + 210.0, + 522.0, + 236.0, + 454.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 214.0, + 877.0, + 214.0, + 877.0, + 234.0, + 833.0, + 234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 210.0, + 1245.0, + 210.0, + 1245.0, + 235.0, + 1198.0, + 235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 248.0, + 345.0, + 248.0, + 345.0, + 265.0, + 319.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 248.0, + 712.0, + 248.0, + 712.0, + 265.0, + 685.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 248.0, + 1078.0, + 248.0, + 1078.0, + 265.0, + 1052.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 280.0, + 345.0, + 280.0, + 345.0, + 378.0, + 306.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 280.0, + 724.0, + 280.0, + 724.0, + 390.0, + 659.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 280.0, + 1078.0, + 280.0, + 1078.0, + 379.0, + 1038.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 345.0, + 343.0, + 345.0, + 343.0, + 364.0, + 318.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 344.0, + 713.0, + 344.0, + 713.0, + 365.0, + 684.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 344.0, + 1079.0, + 344.0, + 1079.0, + 365.0, + 1050.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 377.0, + 346.0, + 377.0, + 346.0, + 398.0, + 317.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 377.0, + 713.0, + 377.0, + 713.0, + 398.0, + 683.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 377.0, + 1080.0, + 377.0, + 1080.0, + 398.0, + 1050.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 411.0, + 345.0, + 411.0, + 345.0, + 429.0, + 317.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 411.0, + 712.0, + 411.0, + 712.0, + 429.0, + 684.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 411.0, + 1078.0, + 411.0, + 1078.0, + 429.0, + 1051.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 446.0, + 404.0, + 446.0, + 404.0, + 465.0, + 382.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 444.0, + 509.0, + 444.0, + 509.0, + 481.0, + 468.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 444.0, + 604.0, + 444.0, + 604.0, + 465.0, + 562.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 446.0, + 771.0, + 446.0, + 771.0, + 465.0, + 749.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 440.0, + 881.0, + 440.0, + 881.0, + 484.0, + 831.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 444.0, + 971.0, + 444.0, + 971.0, + 465.0, + 929.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 446.0, + 1138.0, + 446.0, + 1138.0, + 465.0, + 1117.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 444.0, + 1242.0, + 444.0, + 1242.0, + 478.0, + 1203.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 444.0, + 1337.0, + 444.0, + 1337.0, + 465.0, + 1296.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 485.0, + 515.0, + 485.0, + 515.0, + 517.0, + 463.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 488.0, + 890.0, + 488.0, + 890.0, + 512.0, + 821.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 488.0, + 1248.0, + 488.0, + 1248.0, + 514.0, + 1193.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 525.0, + 345.0, + 525.0, + 345.0, + 544.0, + 319.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 525.0, + 712.0, + 525.0, + 712.0, + 543.0, + 685.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 526.0, + 1078.0, + 526.0, + 1078.0, + 544.0, + 1052.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 558.0, + 343.0, + 558.0, + 343.0, + 658.0, + 303.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 558.0, + 725.0, + 558.0, + 725.0, + 668.0, + 659.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 557.0, + 1078.0, + 557.0, + 1078.0, + 658.0, + 1038.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 622.0, + 343.0, + 622.0, + 343.0, + 641.0, + 319.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 621.0, + 713.0, + 621.0, + 713.0, + 642.0, + 684.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 622.0, + 1078.0, + 622.0, + 1078.0, + 641.0, + 1052.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 655.0, + 343.0, + 655.0, + 343.0, + 674.0, + 318.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 654.0, + 713.0, + 654.0, + 713.0, + 675.0, + 683.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 655.0, + 1078.0, + 655.0, + 1078.0, + 673.0, + 1051.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 688.0, + 343.0, + 688.0, + 343.0, + 706.0, + 318.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 688.0, + 712.0, + 688.0, + 712.0, + 706.0, + 684.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 688.0, + 1078.0, + 688.0, + 1078.0, + 706.0, + 1051.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 723.0, + 404.0, + 723.0, + 404.0, + 742.0, + 382.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 723.0, + 503.0, + 723.0, + 503.0, + 740.0, + 473.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 724.0, + 601.0, + 724.0, + 601.0, + 740.0, + 563.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 723.0, + 771.0, + 723.0, + 771.0, + 742.0, + 748.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 722.0, + 872.0, + 722.0, + 872.0, + 743.0, + 839.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 722.0, + 971.0, + 722.0, + 971.0, + 742.0, + 931.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 723.0, + 1138.0, + 723.0, + 1138.0, + 740.0, + 1117.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 723.0, + 1237.0, + 723.0, + 1237.0, + 740.0, + 1208.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 724.0, + 1335.0, + 724.0, + 1335.0, + 740.0, + 1298.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 732.0, + 508.0, + 732.0, + 508.0, + 756.0, + 469.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 733.0, + 875.0, + 733.0, + 875.0, + 755.0, + 836.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 732.0, + 1242.0, + 732.0, + 1242.0, + 756.0, + 1203.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 768.0, + 1404.0, + 768.0, + 1404.0, + 801.0, + 295.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 789.0, + 731.0, + 789.0, + 731.0, + 844.0, + 289.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 789.0, + 1409.0, + 789.0, + 1409.0, + 844.0, + 1177.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 830.0, + 1404.0, + 830.0, + 1404.0, + 868.0, + 294.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 861.0, + 1404.0, + 861.0, + 1404.0, + 898.0, + 295.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 893.0, + 658.0, + 893.0, + 658.0, + 926.0, + 295.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1971.0, + 1224.0, + 1971.0, + 1224.0, + 2014.0, + 329.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1402.0, + 514.0, + 1402.0, + 514.0, + 1453.0, + 291.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 861.0, + 2060.0, + 861.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1023.0, + 278.0, + 1023.0, + 278.0, + 1057.0, + 238.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1056.0, + 278.0, + 1056.0, + 278.0, + 1091.0, + 238.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1090.0, + 278.0, + 1090.0, + 278.0, + 1124.0, + 238.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1124.0, + 278.0, + 1124.0, + 278.0, + 1158.0, + 238.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1157.0, + 278.0, + 1157.0, + 278.0, + 1191.0, + 238.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1190.0, + 278.0, + 1190.0, + 278.0, + 1224.0, + 238.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1223.0, + 278.0, + 1223.0, + 278.0, + 1257.0, + 238.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1256.0, + 279.0, + 1256.0, + 279.0, + 1295.0, + 237.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1304.0, + 278.0, + 1304.0, + 278.0, + 1340.0, + 238.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1336.0, + 281.0, + 1336.0, + 281.0, + 1374.0, + 236.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1417.0, + 280.0, + 1417.0, + 280.0, + 1452.0, + 237.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1483.0, + 280.0, + 1483.0, + 280.0, + 1521.0, + 236.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1515.0, + 280.0, + 1515.0, + 280.0, + 1554.0, + 236.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1550.0, + 279.0, + 1550.0, + 279.0, + 1585.0, + 238.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1585.0, + 278.0, + 1585.0, + 278.0, + 1619.0, + 238.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1618.0, + 279.0, + 1618.0, + 279.0, + 1652.0, + 238.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1651.0, + 278.0, + 1651.0, + 278.0, + 1685.0, + 238.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1685.0, + 278.0, + 1685.0, + 278.0, + 1719.0, + 238.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1719.0, + 277.0, + 1719.0, + 277.0, + 1755.0, + 238.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1766.0, + 279.0, + 1766.0, + 279.0, + 1800.0, + 237.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1799.0, + 279.0, + 1799.0, + 279.0, + 1834.0, + 237.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1832.0, + 279.0, + 1832.0, + 279.0, + 1866.0, + 238.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1866.0, + 279.0, + 1866.0, + 279.0, + 1900.0, + 238.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1898.0, + 279.0, + 1898.0, + 279.0, + 1932.0, + 238.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1932.0, + 279.0, + 1932.0, + 279.0, + 1965.0, + 238.0, + 1965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 1406.0, + 1477.0, + 1406.0, + 1519.0, + 293.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1512.0, + 1405.0, + 1512.0, + 1405.0, + 1550.0, + 293.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1545.0, + 1405.0, + 1545.0, + 1405.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1581.0, + 1403.0, + 1581.0, + 1403.0, + 1613.0, + 295.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1613.0, + 1405.0, + 1613.0, + 1405.0, + 1649.0, + 294.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1644.0, + 1406.0, + 1644.0, + 1406.0, + 1683.0, + 294.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1714.0, + 798.0, + 1714.0, + 798.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1762.0, + 1402.0, + 1762.0, + 1402.0, + 1795.0, + 294.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1795.0, + 1406.0, + 1795.0, + 1406.0, + 1829.0, + 294.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1825.0, + 1405.0, + 1825.0, + 1405.0, + 1864.0, + 293.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1859.0, + 1405.0, + 1859.0, + 1405.0, + 1898.0, + 293.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1892.0, + 1406.0, + 1892.0, + 1406.0, + 1930.0, + 293.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1928.0, + 495.0, + 1928.0, + 495.0, + 1961.0, + 294.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1017.0, + 1404.0, + 1017.0, + 1404.0, + 1054.0, + 294.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1051.0, + 1406.0, + 1051.0, + 1406.0, + 1087.0, + 295.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1082.0, + 1404.0, + 1082.0, + 1404.0, + 1124.0, + 295.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1121.0, + 1404.0, + 1121.0, + 1404.0, + 1153.0, + 295.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1153.0, + 1406.0, + 1153.0, + 1406.0, + 1186.0, + 294.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1185.0, + 1408.0, + 1185.0, + 1408.0, + 1219.0, + 294.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1218.0, + 1406.0, + 1218.0, + 1406.0, + 1254.0, + 295.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 447.0, + 1251.0, + 447.0, + 1290.0, + 295.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1298.0, + 1405.0, + 1298.0, + 1405.0, + 1335.0, + 295.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1333.0, + 771.0, + 1333.0, + 771.0, + 1366.0, + 296.0, + 1366.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 248, + 127, + 1412, + 127, + 1412, + 2009, + 248, + 2009 + ], + "score": 0.774 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 866, + 2061, + 866, + 2086, + 836, + 2086 + ], + "score": 0.762 + }, + { + "category_id": 0, + "poly": [ + 295, + 200, + 456, + 200, + 456, + 235, + 295, + 235 + ], + "score": 0.719 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 866, + 2061, + 866, + 2086, + 836, + 2086 + ], + "score": 0.244 + }, + { + "category_id": 2, + "poly": [ + 237, + 213, + 283, + 213, + 283, + 2017, + 237, + 2017 + ], + "score": 0.24 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 197.0, + 459.0, + 197.0, + 459.0, + 239.0, + 295.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 209.0, + 281.0, + 209.0, + 281.0, + 240.0, + 235.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 252.0, + 281.0, + 252.0, + 281.0, + 284.0, + 235.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 278.0, + 283.0, + 278.0, + 283.0, + 318.0, + 235.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 330.0, + 278.0, + 330.0, + 278.0, + 361.0, + 237.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 353.0, + 281.0, + 353.0, + 281.0, + 395.0, + 235.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 405.0, + 281.0, + 405.0, + 281.0, + 437.0, + 235.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 433.0, + 281.0, + 433.0, + 281.0, + 464.0, + 235.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 459.0, + 283.0, + 459.0, + 283.0, + 498.0, + 233.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 508.0, + 281.0, + 508.0, + 281.0, + 542.0, + 237.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 536.0, + 281.0, + 536.0, + 281.0, + 574.0, + 237.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 585.0, + 281.0, + 585.0, + 281.0, + 619.0, + 237.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 613.0, + 281.0, + 613.0, + 281.0, + 647.0, + 237.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 639.0, + 283.0, + 639.0, + 283.0, + 681.0, + 235.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 691.0, + 281.0, + 691.0, + 281.0, + 722.0, + 235.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 718.0, + 281.0, + 718.0, + 281.0, + 750.0, + 235.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 744.0, + 283.0, + 744.0, + 283.0, + 784.0, + 233.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 792.0, + 283.0, + 792.0, + 283.0, + 829.0, + 233.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 821.0, + 281.0, + 821.0, + 281.0, + 855.0, + 237.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 849.0, + 283.0, + 849.0, + 283.0, + 889.0, + 235.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 899.0, + 278.0, + 899.0, + 278.0, + 933.0, + 237.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 925.0, + 281.0, + 925.0, + 281.0, + 966.0, + 235.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 976.0, + 281.0, + 976.0, + 281.0, + 1008.0, + 235.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1002.0, + 281.0, + 1002.0, + 281.0, + 1046.0, + 235.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1052.0, + 278.0, + 1052.0, + 278.0, + 1085.0, + 237.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1079.0, + 283.0, + 1079.0, + 283.0, + 1119.0, + 235.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1129.0, + 281.0, + 1129.0, + 281.0, + 1163.0, + 237.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1155.0, + 281.0, + 1155.0, + 281.0, + 1188.0, + 237.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1182.0, + 283.0, + 1182.0, + 283.0, + 1228.0, + 235.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1232.0, + 283.0, + 1232.0, + 283.0, + 1274.0, + 235.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1282.0, + 278.0, + 1282.0, + 278.0, + 1315.0, + 237.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1311.0, + 281.0, + 1311.0, + 281.0, + 1343.0, + 237.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1337.0, + 283.0, + 1337.0, + 283.0, + 1379.0, + 235.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1389.0, + 281.0, + 1389.0, + 281.0, + 1420.0, + 235.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1413.0, + 281.0, + 1413.0, + 281.0, + 1458.0, + 233.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1462.0, + 281.0, + 1462.0, + 281.0, + 1496.0, + 237.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1490.0, + 283.0, + 1490.0, + 283.0, + 1530.0, + 233.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1539.0, + 281.0, + 1539.0, + 281.0, + 1573.0, + 237.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1567.0, + 283.0, + 1567.0, + 283.0, + 1609.0, + 235.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1615.0, + 283.0, + 1615.0, + 283.0, + 1652.0, + 233.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1645.0, + 281.0, + 1645.0, + 281.0, + 1682.0, + 237.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1694.0, + 281.0, + 1694.0, + 281.0, + 1726.0, + 237.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1720.0, + 283.0, + 1720.0, + 283.0, + 1760.0, + 233.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1771.0, + 281.0, + 1771.0, + 281.0, + 1803.0, + 235.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1797.0, + 281.0, + 1797.0, + 281.0, + 1839.0, + 235.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1847.0, + 281.0, + 1847.0, + 281.0, + 1881.0, + 237.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 235.0, + 1877.0, + 281.0, + 1877.0, + 281.0, + 1908.0, + 235.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1900.0, + 281.0, + 1900.0, + 281.0, + 1934.0, + 237.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 237.0, + 1932.0, + 281.0, + 1932.0, + 281.0, + 1966.0, + 237.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1978.0, + 283.0, + 1978.0, + 283.0, + 2015.0, + 233.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 196.0, + 462.0, + 196.0, + 462.0, + 240.0, + 242.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 250.0, + 1404.0, + 250.0, + 1404.0, + 283.0, + 238.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 279.0, + 281.0, + 279.0, + 281.0, + 312.0, + 240.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 277.0, + 506.0, + 277.0, + 506.0, + 310.0, + 323.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 329.0, + 285.0, + 329.0, + 285.0, + 360.0, + 240.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 327.0, + 1406.0, + 327.0, + 1406.0, + 360.0, + 292.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 356.0, + 281.0, + 356.0, + 281.0, + 387.0, + 240.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 354.0, + 803.0, + 354.0, + 803.0, + 387.0, + 321.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 405.0, + 283.0, + 405.0, + 283.0, + 434.0, + 240.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 403.0, + 1404.0, + 403.0, + 1404.0, + 436.0, + 292.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 432.0, + 281.0, + 432.0, + 281.0, + 465.0, + 240.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 432.0, + 1406.0, + 432.0, + 1406.0, + 465.0, + 323.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 461.0, + 281.0, + 461.0, + 281.0, + 494.0, + 240.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 459.0, + 1362.0, + 459.0, + 1362.0, + 492.0, + 323.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 511.0, + 285.0, + 511.0, + 285.0, + 539.0, + 242.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 508.0, + 1406.0, + 508.0, + 1406.0, + 542.0, + 292.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 535.0, + 281.0, + 535.0, + 281.0, + 568.0, + 240.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 533.0, + 1042.0, + 533.0, + 1042.0, + 568.0, + 318.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 585.0, + 285.0, + 585.0, + 285.0, + 616.0, + 240.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 583.0, + 1404.0, + 583.0, + 1404.0, + 616.0, + 292.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 614.0, + 281.0, + 614.0, + 281.0, + 647.0, + 240.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 610.0, + 1408.0, + 610.0, + 1408.0, + 649.0, + 319.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 641.0, + 279.0, + 641.0, + 279.0, + 674.0, + 240.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 639.0, + 391.0, + 639.0, + 391.0, + 674.0, + 321.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 690.0, + 283.0, + 690.0, + 283.0, + 721.0, + 240.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 684.0, + 1408.0, + 684.0, + 1408.0, + 725.0, + 287.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 715.0, + 281.0, + 715.0, + 281.0, + 748.0, + 240.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 711.0, + 1406.0, + 711.0, + 1406.0, + 752.0, + 319.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 746.0, + 279.0, + 746.0, + 279.0, + 779.0, + 240.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 742.0, + 786.0, + 742.0, + 786.0, + 779.0, + 316.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 795.0, + 285.0, + 795.0, + 285.0, + 826.0, + 240.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 791.0, + 1406.0, + 791.0, + 1406.0, + 831.0, + 289.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 822.0, + 281.0, + 822.0, + 281.0, + 855.0, + 240.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 822.0, + 1406.0, + 822.0, + 1406.0, + 855.0, + 321.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 851.0, + 281.0, + 851.0, + 281.0, + 882.0, + 240.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 847.0, + 393.0, + 847.0, + 393.0, + 882.0, + 323.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 899.0, + 285.0, + 899.0, + 285.0, + 930.0, + 240.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 892.0, + 1410.0, + 892.0, + 1410.0, + 934.0, + 287.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 928.0, + 281.0, + 928.0, + 281.0, + 961.0, + 240.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 926.0, + 1302.0, + 926.0, + 1302.0, + 959.0, + 323.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 975.0, + 281.0, + 975.0, + 281.0, + 1006.0, + 240.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 969.0, + 1408.0, + 969.0, + 1408.0, + 1010.0, + 287.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1004.0, + 281.0, + 1004.0, + 281.0, + 1037.0, + 240.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1002.0, + 755.0, + 1002.0, + 755.0, + 1035.0, + 321.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1051.0, + 281.0, + 1051.0, + 281.0, + 1082.0, + 240.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1045.0, + 1408.0, + 1045.0, + 1408.0, + 1087.0, + 292.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1080.0, + 281.0, + 1080.0, + 281.0, + 1113.0, + 240.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1078.0, + 768.0, + 1078.0, + 768.0, + 1111.0, + 321.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1130.0, + 285.0, + 1130.0, + 285.0, + 1161.0, + 240.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1128.0, + 1408.0, + 1128.0, + 1408.0, + 1161.0, + 292.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1155.0, + 281.0, + 1155.0, + 281.0, + 1188.0, + 240.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1153.0, + 1410.0, + 1153.0, + 1410.0, + 1192.0, + 321.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1184.0, + 281.0, + 1184.0, + 281.0, + 1217.0, + 240.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1186.0, + 468.0, + 1186.0, + 468.0, + 1213.0, + 325.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1231.0, + 1410.0, + 1231.0, + 1410.0, + 1264.0, + 242.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1281.0, + 1408.0, + 1281.0, + 1408.0, + 1314.0, + 240.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1310.0, + 281.0, + 1310.0, + 281.0, + 1343.0, + 240.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1310.0, + 1408.0, + 1310.0, + 1408.0, + 1343.0, + 323.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1338.0, + 281.0, + 1338.0, + 281.0, + 1369.0, + 240.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1336.0, + 664.0, + 1336.0, + 664.0, + 1369.0, + 321.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1388.0, + 283.0, + 1388.0, + 283.0, + 1419.0, + 240.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1382.0, + 1408.0, + 1382.0, + 1408.0, + 1423.0, + 290.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1415.0, + 279.0, + 1415.0, + 279.0, + 1446.0, + 240.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1411.0, + 780.0, + 1411.0, + 780.0, + 1448.0, + 316.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1464.0, + 285.0, + 1464.0, + 285.0, + 1493.0, + 240.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1462.0, + 1406.0, + 1462.0, + 1406.0, + 1495.0, + 292.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1491.0, + 281.0, + 1491.0, + 281.0, + 1524.0, + 240.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1491.0, + 1179.0, + 1491.0, + 1179.0, + 1524.0, + 321.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1532.0, + 1406.0, + 1532.0, + 1406.0, + 1576.0, + 240.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1568.0, + 279.0, + 1568.0, + 279.0, + 1601.0, + 240.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1566.0, + 1096.0, + 1566.0, + 1096.0, + 1603.0, + 319.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1617.0, + 285.0, + 1617.0, + 285.0, + 1648.0, + 240.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1615.0, + 1408.0, + 1615.0, + 1408.0, + 1648.0, + 292.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1644.0, + 279.0, + 1644.0, + 279.0, + 1675.0, + 240.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1642.0, + 489.0, + 1642.0, + 489.0, + 1675.0, + 323.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1694.0, + 283.0, + 1694.0, + 283.0, + 1722.0, + 240.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1692.0, + 1406.0, + 1692.0, + 1406.0, + 1725.0, + 292.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1720.0, + 281.0, + 1720.0, + 281.0, + 1753.0, + 240.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1720.0, + 840.0, + 1720.0, + 840.0, + 1753.0, + 323.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1770.0, + 283.0, + 1770.0, + 283.0, + 1801.0, + 240.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 1406.0, + 1768.0, + 1406.0, + 1801.0, + 294.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1799.0, + 279.0, + 1799.0, + 279.0, + 1832.0, + 240.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1797.0, + 1150.0, + 1797.0, + 1150.0, + 1830.0, + 323.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1848.0, + 285.0, + 1848.0, + 285.0, + 1877.0, + 242.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1846.0, + 1406.0, + 1846.0, + 1406.0, + 1879.0, + 292.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1875.0, + 281.0, + 1875.0, + 281.0, + 1906.0, + 240.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1869.0, + 1408.0, + 1869.0, + 1408.0, + 1910.0, + 319.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1902.0, + 281.0, + 1902.0, + 281.0, + 1935.0, + 240.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1898.0, + 1408.0, + 1898.0, + 1408.0, + 1937.0, + 321.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 240.0, + 1931.0, + 279.0, + 1931.0, + 279.0, + 1964.0, + 240.0, + 1964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1929.0, + 678.0, + 1929.0, + 678.0, + 1962.0, + 323.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 242.0, + 1978.0, + 1406.0, + 1978.0, + 1406.0, + 2012.0, + 242.0, + 2012.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 410, + 1176, + 1405, + 1176, + 1405, + 1415, + 410, + 1415 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 410, + 861, + 1403, + 861, + 1403, + 1018, + 410, + 1018 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 410, + 1729, + 1405, + 1729, + 1405, + 1909, + 410, + 1909 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 410, + 1464, + 1404, + 1464, + 1404, + 1683, + 410, + 1683 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 408, + 1065, + 1309, + 1065, + 1309, + 1129, + 408, + 1129 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 366, + 1138, + 644, + 1138, + 644, + 1166, + 366, + 1166 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 366, + 822, + 562, + 822, + 562, + 850, + 366, + 850 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 366, + 1027, + 797, + 1027, + 797, + 1056, + 366, + 1056 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 361, + 1425, + 1306, + 1425, + 1306, + 1455, + 361, + 1455 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 863, + 2061, + 863, + 2085, + 836, + 2085 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 388, + 1691, + 1126, + 1691, + 1126, + 1720, + 388, + 1720 + ], + "score": 0.816 + }, + { + "category_id": 0, + "poly": [ + 247, + 765, + 434, + 765, + 434, + 801, + 247, + 801 + ], + "score": 0.807 + }, + { + "category_id": 1, + "poly": [ + 254, + 202, + 1408, + 202, + 1408, + 712, + 254, + 712 + ], + "score": 0.76 + }, + { + "category_id": 2, + "poly": [ + 237, + 161, + 285, + 161, + 285, + 1970, + 237, + 1970 + ], + "score": 0.199 + }, + { + "category_id": 2, + "poly": [ + 238, + 224, + 279, + 224, + 279, + 720, + 238, + 720 + ], + "score": 0.174 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 868.0, + 2058.0, + 868.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 241.0, + 773.0, + 281.0, + 773.0, + 281.0, + 801.0, + 241.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 762.0, + 440.0, + 762.0, + 440.0, + 806.0, + 294.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 208.0, + 279.0, + 208.0, + 279.0, + 240.0, + 236.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 234.0, + 279.0, + 234.0, + 279.0, + 266.0, + 236.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 260.0, + 282.0, + 260.0, + 282.0, + 300.0, + 233.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 314.0, + 279.0, + 314.0, + 279.0, + 346.0, + 236.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 342.0, + 279.0, + 342.0, + 279.0, + 373.0, + 236.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 368.0, + 282.0, + 368.0, + 282.0, + 405.0, + 238.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 415.0, + 284.0, + 415.0, + 284.0, + 453.0, + 233.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 445.0, + 279.0, + 445.0, + 279.0, + 479.0, + 236.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 473.0, + 279.0, + 473.0, + 279.0, + 505.0, + 236.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 499.0, + 282.0, + 499.0, + 282.0, + 541.0, + 233.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 550.0, + 279.0, + 550.0, + 279.0, + 584.0, + 236.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 580.0, + 279.0, + 580.0, + 279.0, + 614.0, + 236.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 628.0, + 279.0, + 628.0, + 279.0, + 660.0, + 236.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 656.0, + 282.0, + 656.0, + 282.0, + 688.0, + 236.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 684.0, + 279.0, + 684.0, + 279.0, + 718.0, + 236.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 775.0, + 282.0, + 775.0, + 282.0, + 807.0, + 236.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 823.0, + 284.0, + 823.0, + 284.0, + 861.0, + 233.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 863.0, + 282.0, + 863.0, + 282.0, + 894.0, + 236.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 891.0, + 279.0, + 891.0, + 279.0, + 924.0, + 236.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 922.0, + 284.0, + 922.0, + 284.0, + 960.0, + 233.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 958.0, + 279.0, + 958.0, + 279.0, + 990.0, + 236.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 990.0, + 284.0, + 990.0, + 284.0, + 1028.0, + 233.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1028.0, + 284.0, + 1028.0, + 284.0, + 1065.0, + 233.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1067.0, + 279.0, + 1067.0, + 279.0, + 1101.0, + 236.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1099.0, + 282.0, + 1099.0, + 282.0, + 1139.0, + 233.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1137.0, + 282.0, + 1137.0, + 282.0, + 1177.0, + 233.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1179.0, + 282.0, + 1179.0, + 282.0, + 1211.0, + 236.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1203.0, + 282.0, + 1203.0, + 282.0, + 1242.0, + 233.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1239.0, + 279.0, + 1239.0, + 279.0, + 1272.0, + 238.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1266.0, + 279.0, + 1266.0, + 279.0, + 1300.0, + 238.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 1300.0, + 279.0, + 1300.0, + 279.0, + 1332.0, + 238.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1328.0, + 279.0, + 1328.0, + 279.0, + 1362.0, + 236.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1362.0, + 279.0, + 1362.0, + 279.0, + 1394.0, + 236.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1386.0, + 282.0, + 1386.0, + 282.0, + 1425.0, + 233.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1427.0, + 279.0, + 1427.0, + 279.0, + 1461.0, + 236.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1465.0, + 279.0, + 1465.0, + 279.0, + 1499.0, + 236.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1499.0, + 279.0, + 1499.0, + 279.0, + 1533.0, + 236.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1533.0, + 279.0, + 1533.0, + 279.0, + 1567.0, + 236.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1567.0, + 279.0, + 1567.0, + 279.0, + 1598.0, + 236.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1592.0, + 279.0, + 1592.0, + 279.0, + 1626.0, + 236.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1626.0, + 279.0, + 1626.0, + 279.0, + 1658.0, + 236.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 233.0, + 1652.0, + 282.0, + 1652.0, + 282.0, + 1692.0, + 233.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1694.0, + 279.0, + 1694.0, + 279.0, + 1728.0, + 236.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1732.0, + 282.0, + 1732.0, + 282.0, + 1763.0, + 236.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1760.0, + 279.0, + 1760.0, + 279.0, + 1791.0, + 236.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1791.0, + 279.0, + 1791.0, + 279.0, + 1823.0, + 236.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1821.0, + 279.0, + 1821.0, + 279.0, + 1853.0, + 236.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1853.0, + 282.0, + 1853.0, + 282.0, + 1885.0, + 236.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 1881.0, + 282.0, + 1881.0, + 282.0, + 1913.0, + 236.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 239.0, + 216.0, + 274.0, + 216.0, + 274.0, + 237.0, + 239.0, + 237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 236.0, + 279.0, + 236.0, + 279.0, + 266.0, + 236.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 264.0, + 279.0, + 264.0, + 279.0, + 296.0, + 238.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 313.0, + 279.0, + 313.0, + 279.0, + 344.0, + 238.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 341.0, + 278.0, + 341.0, + 278.0, + 372.0, + 236.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 368.0, + 281.0, + 368.0, + 281.0, + 402.0, + 236.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 418.0, + 279.0, + 418.0, + 279.0, + 448.0, + 236.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 447.0, + 279.0, + 447.0, + 279.0, + 477.0, + 238.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 474.0, + 279.0, + 474.0, + 279.0, + 503.0, + 238.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 501.0, + 281.0, + 501.0, + 281.0, + 536.0, + 236.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 551.0, + 279.0, + 551.0, + 279.0, + 582.0, + 238.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 579.0, + 281.0, + 579.0, + 281.0, + 614.0, + 236.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 238.0, + 629.0, + 279.0, + 629.0, + 279.0, + 659.0, + 238.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 236.0, + 656.0, + 279.0, + 656.0, + 279.0, + 688.0, + 236.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 234.0, + 684.0, + 278.0, + 684.0, + 278.0, + 716.0, + 234.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1173.0, + 1405.0, + 1173.0, + 1405.0, + 1208.0, + 407.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1205.0, + 1379.0, + 1205.0, + 1379.0, + 1234.0, + 450.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1234.0, + 1405.0, + 1234.0, + 1405.0, + 1270.0, + 410.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1265.0, + 655.0, + 1265.0, + 655.0, + 1294.0, + 451.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1298.0, + 1404.0, + 1298.0, + 1404.0, + 1331.0, + 410.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1325.0, + 789.0, + 1325.0, + 789.0, + 1358.0, + 451.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1358.0, + 1407.0, + 1358.0, + 1407.0, + 1390.0, + 410.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1386.0, + 1034.0, + 1386.0, + 1034.0, + 1416.0, + 451.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 858.0, + 1405.0, + 858.0, + 1405.0, + 894.0, + 408.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 887.0, + 684.0, + 887.0, + 684.0, + 920.0, + 450.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 921.0, + 1130.0, + 921.0, + 1130.0, + 953.0, + 408.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 954.0, + 1360.0, + 954.0, + 1360.0, + 987.0, + 409.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 987.0, + 1401.0, + 987.0, + 1401.0, + 1022.0, + 407.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1726.0, + 1407.0, + 1726.0, + 1407.0, + 1762.0, + 406.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1754.0, + 521.0, + 1754.0, + 521.0, + 1788.0, + 446.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1822.0, + 409.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1819.0, + 772.0, + 1819.0, + 772.0, + 1848.0, + 450.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1884.0, + 411.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1878.0, + 790.0, + 1878.0, + 790.0, + 1911.0, + 448.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1463.0, + 1395.0, + 1463.0, + 1395.0, + 1496.0, + 407.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1497.0, + 947.0, + 1497.0, + 947.0, + 1528.0, + 408.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1529.0, + 1330.0, + 1529.0, + 1330.0, + 1564.0, + 407.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1562.0, + 1407.0, + 1562.0, + 1407.0, + 1596.0, + 409.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1591.0, + 659.0, + 1591.0, + 659.0, + 1623.0, + 449.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1620.0, + 1408.0, + 1620.0, + 1408.0, + 1658.0, + 411.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1652.0, + 781.0, + 1652.0, + 781.0, + 1683.0, + 450.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1062.0, + 1316.0, + 1062.0, + 1316.0, + 1096.0, + 407.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1097.0, + 1245.0, + 1097.0, + 1245.0, + 1130.0, + 406.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1135.0, + 648.0, + 1135.0, + 648.0, + 1170.0, + 361.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 818.0, + 567.0, + 818.0, + 567.0, + 853.0, + 363.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1024.0, + 801.0, + 1024.0, + 801.0, + 1060.0, + 362.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1421.0, + 1309.0, + 1421.0, + 1309.0, + 1460.0, + 359.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1686.0, + 1124.0, + 1686.0, + 1124.0, + 1726.0, + 393.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 210.0, + 278.0, + 210.0, + 278.0, + 236.0, + 248.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 207.0, + 1404.0, + 207.0, + 1404.0, + 238.0, + 293.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 236.0, + 280.0, + 236.0, + 280.0, + 265.0, + 246.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 232.0, + 1406.0, + 232.0, + 1406.0, + 268.0, + 322.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 264.0, + 278.0, + 264.0, + 278.0, + 296.0, + 246.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 1052.0, + 261.0, + 1052.0, + 293.0, + 323.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 315.0, + 277.0, + 315.0, + 277.0, + 341.0, + 248.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 308.0, + 1406.0, + 308.0, + 1406.0, + 345.0, + 294.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 344.0, + 277.0, + 344.0, + 277.0, + 371.0, + 248.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 339.0, + 1406.0, + 339.0, + 1406.0, + 374.0, + 322.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 369.0, + 278.0, + 369.0, + 278.0, + 401.0, + 246.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 367.0, + 770.0, + 367.0, + 770.0, + 399.0, + 320.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 419.0, + 277.0, + 419.0, + 277.0, + 447.0, + 248.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 415.0, + 1408.0, + 415.0, + 1408.0, + 451.0, + 298.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 447.0, + 280.0, + 447.0, + 280.0, + 476.0, + 246.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 443.0, + 1407.0, + 443.0, + 1407.0, + 479.0, + 322.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 246.0, + 475.0, + 280.0, + 475.0, + 280.0, + 504.0, + 246.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 471.0, + 1406.0, + 471.0, + 1406.0, + 507.0, + 320.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 503.0, + 278.0, + 503.0, + 278.0, + 535.0, + 247.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 500.0, + 512.0, + 500.0, + 512.0, + 531.0, + 321.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 552.0, + 278.0, + 552.0, + 278.0, + 579.0, + 248.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 546.0, + 1404.0, + 546.0, + 1404.0, + 584.0, + 292.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 580.0, + 280.0, + 580.0, + 280.0, + 612.0, + 247.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 576.0, + 783.0, + 576.0, + 783.0, + 608.0, + 323.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 630.0, + 278.0, + 630.0, + 278.0, + 657.0, + 248.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 627.0, + 1406.0, + 627.0, + 1406.0, + 659.0, + 294.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 247.0, + 658.0, + 280.0, + 658.0, + 280.0, + 688.0, + 247.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 654.0, + 1408.0, + 654.0, + 1408.0, + 688.0, + 320.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 248.0, + 688.0, + 277.0, + 688.0, + 277.0, + 715.0, + 248.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 682.0, + 400.0, + 682.0, + 400.0, + 712.0, + 322.0, + 712.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/HyeSin4FPB/HyeSin4FPB_middle.json b/parse/train/HyeSin4FPB/HyeSin4FPB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..b3cd001456ac2809dab0c2a5c52df242a1fa1d73 --- /dev/null +++ b/parse/train/HyeSin4FPB/HyeSin4FPB_middle.json @@ -0,0 +1,70714 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 348, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 337, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 337, + 96 + ], + "score": 1.0, + "content": "LEARNING TO CONTROL PDES", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 349, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 349, + 117 + ], + "score": 1.0, + "content": "WITH DIFFERENTIABLE PHYSICS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 241, + 157 + ], + "lines": [ + { + "bbox": [ + 111, + 135, + 167, + 147 + ], + "spans": [ + { + "bbox": [ + 111, + 135, + 167, + 147 + ], + "score": 1.0, + "content": "Philipp Holl", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 241, + 157 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 241, + 157 + ], + "score": 1.0, + "content": "Technical University of Munich", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 266, + 135, + 332, + 157 + ], + "lines": [ + { + "bbox": [ + 264, + 135, + 333, + 146 + ], + "spans": [ + { + "bbox": [ + 264, + 135, + 333, + 146 + ], + "score": 1.0, + "content": "Vladlen Koltun", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 264, + 146, + 308, + 157 + ], + "spans": [ + { + "bbox": [ + 264, + 146, + 308, + 157 + ], + "score": 1.0, + "content": "Intel Labs", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 357, + 135, + 485, + 157 + ], + "lines": [ + { + "bbox": [ + 355, + 134, + 414, + 149 + ], + "spans": [ + { + "bbox": [ + 355, + 134, + 414, + 149 + ], + "score": 1.0, + "content": "Nils Thuerey", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 356, + 146, + 484, + 158 + ], + "spans": [ + { + "bbox": [ + 356, + 146, + 484, + 158 + ], + "score": 1.0, + "content": "Technical University of Munich", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 278, + 187, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 143, + 210, + 468, + 363 + ], + "lines": [ + { + "bbox": [ + 141, + 209, + 469, + 223 + ], + "spans": [ + { + "bbox": [ + 141, + 209, + 469, + 223 + ], + "score": 1.0, + "content": "Predicting outcomes and planning interactions with the physical world are long-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 221, + 469, + 234 + ], + "spans": [ + { + "bbox": [ + 141, + 221, + 469, + 234 + ], + "score": 1.0, + "content": "standing goals for machine learning. A variety of such tasks involves contin-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 232, + 469, + 245 + ], + "spans": [ + { + "bbox": [ + 141, + 232, + 469, + 245 + ], + "score": 1.0, + "content": "uous physical systems, which can be described by partial differential equations", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 243, + 469, + 255 + ], + "spans": [ + { + "bbox": [ + 142, + 243, + 469, + 255 + ], + "score": 1.0, + "content": "(PDEs) with many degrees of freedom. Existing methods that aim to control the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 253, + 470, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 470, + 267 + ], + "score": 1.0, + "content": "dynamics of such systems are typically limited to relatively short time frames or a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 264, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 264, + 470, + 277 + ], + "score": 1.0, + "content": "small number of interaction parameters. We present a novel hierarchical predictor-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 276, + 469, + 287 + ], + "spans": [ + { + "bbox": [ + 142, + 276, + 469, + 287 + ], + "score": 1.0, + "content": "corrector scheme which enables neural networks to learn to understand and control", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 287, + 470, + 299 + ], + "spans": [ + { + "bbox": [ + 142, + 287, + 470, + 299 + ], + "score": 1.0, + "content": "complex nonlinear physical systems over long time frames. We propose to split", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 297, + 470, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 470, + 311 + ], + "score": 1.0, + "content": "the problem into two distinct tasks: planning and control. To this end, we intro-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "score": 1.0, + "content": "duce a predictor network that plans optimal trajectories and a control network that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 320, + 469, + 331 + ], + "spans": [ + { + "bbox": [ + 142, + 320, + 469, + 331 + ], + "score": 1.0, + "content": "infers the corresponding control parameters. Both stages are trained end-to-end", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 330, + 469, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 469, + 343 + ], + "score": 1.0, + "content": "using a differentiable PDE solver. We demonstrate that our method successfully", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 342, + 469, + 353 + ], + "spans": [ + { + "bbox": [ + 142, + 342, + 469, + 353 + ], + "score": 1.0, + "content": "develops an understanding of complex physical systems and learns to control them", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 352, + 455, + 365 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 455, + 365 + ], + "score": 1.0, + "content": "for tasks involving PDEs such as the incompressible Navier-Stokes equations.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 109, + 383, + 206, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 208, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 208, + 397 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "Intelligent systems that operate in the physical world must be able to perceive, predict, and interact", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "with physical phenomena (Battaglia et al., 2013). In this work, we consider physical systems that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "can be characterized by partial differential equations (PDEs). PDEs constitute the most fundamen-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "tal description of evolving systems and are used to describe every physical theory, from quantum", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "mechanics and general relativity to turbulent flows (Courant & Hilbert, 1962; Smith, 1985). We", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "aim to endow artificial intelligent agents with the ability to direct the evolution of such systems via", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 473, + 190, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 190, + 483 + ], + "score": 1.0, + "content": "continuous controls.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "Such optimal control problems have typically been addressed via iterative optimization. Differ-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "entiable solvers and the adjoint method enable efficient optimization of high-dimensional sys-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "tems (Toussaint et al., 2018; de Avila Belbute-Peres et al., 2018; Schenck & Fox, 2018). However,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 523, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 506, + 535 + ], + "score": 1.0, + "content": "direct optimization through gradient descent (single shooting) at test time is resource-intensive and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "may be difficult to deploy in interactive settings. More advanced methods exist, such as multiple", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "shooting and collocation, but they commonly rely on modeling assumptions that limit their applica-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 555, + 439, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 439, + 568 + ], + "score": 1.0, + "content": "bility, and still require computationally intensive iterative optimization at test time.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "Iterative optimization methods are expensive because they have to start optimizing from scratch and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "typically require a large number of iterations to reach an optimum. In many real-world control prob-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "lems, however, agents have to repeatedly make decisions in specialized environments, and reaction", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "times are limited to a fraction of a second. This motivates the use of data-driven models such as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "deep neural networks, which combine short inference times with the capacity to build an internal", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 627, + 245, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 245, + 639 + ], + "score": 1.0, + "content": "representation of the environment.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "We present a novel deep learning approach that can learn to represent solution manifolds for a given", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "physical environment, and is orders of magnitude faster than iterative optimization techniques. The", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 667, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 677 + ], + "score": 1.0, + "content": "core of our method is a hierarchical predictor-corrector scheme that temporally divides the problem", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "into easier subproblems. This enables us to combine models specialized to different time scales in", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "order to control long sequences of complex high-dimensional systems. We train our models using a", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "differentiable PDE solver that can provide the agent with feedback of how interactions at any point", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "in time affect the outcome. Our models learn to represent manifolds containing a large number of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 473, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 473, + 733 + ], + "score": 1.0, + "content": "solutions, and can thereby avoid local minima that can trap classic optimization techniques.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 348, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 337, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 337, + 96 + ], + "score": 1.0, + "content": "LEARNING TO CONTROL PDES", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 97, + 349, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 349, + 117 + ], + "score": 1.0, + "content": "WITH DIFFERENTIABLE PHYSICS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 241, + 157 + ], + "lines": [ + { + "bbox": [ + 111, + 135, + 167, + 147 + ], + "spans": [ + { + "bbox": [ + 111, + 135, + 167, + 147 + ], + "score": 1.0, + "content": "Philipp Holl", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 241, + 157 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 241, + 157 + ], + "score": 1.0, + "content": "Technical University of Munich", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 111, + 135, + 241, + 157 + ] + }, + { + "type": "text", + "bbox": [ + 266, + 135, + 332, + 157 + ], + "lines": [ + { + "bbox": [ + 264, + 135, + 333, + 146 + ], + "spans": [ + { + "bbox": [ + 264, + 135, + 333, + 146 + ], + "score": 1.0, + "content": "Vladlen Koltun", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 264, + 146, + 308, + 157 + ], + "spans": [ + { + "bbox": [ + 264, + 146, + 308, + 157 + ], + "score": 1.0, + "content": "Intel Labs", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 264, + 135, + 333, + 157 + ] + }, + { + "type": "text", + "bbox": [ + 357, + 135, + 485, + 157 + ], + "lines": [ + { + "bbox": [ + 355, + 134, + 414, + 149 + ], + "spans": [ + { + "bbox": [ + 355, + 134, + 414, + 149 + ], + "score": 1.0, + "content": "Nils Thuerey", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 356, + 146, + 484, + 158 + ], + "spans": [ + { + "bbox": [ + 356, + 146, + 484, + 158 + ], + "score": 1.0, + "content": "Technical University of Munich", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 355, + 134, + 484, + 158 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 187, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 143, + 210, + 468, + 363 + ], + "lines": [ + { + "bbox": [ + 141, + 209, + 469, + 223 + ], + "spans": [ + { + "bbox": [ + 141, + 209, + 469, + 223 + ], + "score": 1.0, + "content": "Predicting outcomes and planning interactions with the physical world are long-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 221, + 469, + 234 + ], + "spans": [ + { + "bbox": [ + 141, + 221, + 469, + 234 + ], + "score": 1.0, + "content": "standing goals for machine learning. A variety of such tasks involves contin-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 232, + 469, + 245 + ], + "spans": [ + { + "bbox": [ + 141, + 232, + 469, + 245 + ], + "score": 1.0, + "content": "uous physical systems, which can be described by partial differential equations", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 243, + 469, + 255 + ], + "spans": [ + { + "bbox": [ + 142, + 243, + 469, + 255 + ], + "score": 1.0, + "content": "(PDEs) with many degrees of freedom. Existing methods that aim to control the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 253, + 470, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 470, + 267 + ], + "score": 1.0, + "content": "dynamics of such systems are typically limited to relatively short time frames or a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 264, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 264, + 470, + 277 + ], + "score": 1.0, + "content": "small number of interaction parameters. We present a novel hierarchical predictor-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 276, + 469, + 287 + ], + "spans": [ + { + "bbox": [ + 142, + 276, + 469, + 287 + ], + "score": 1.0, + "content": "corrector scheme which enables neural networks to learn to understand and control", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 287, + 470, + 299 + ], + "spans": [ + { + "bbox": [ + 142, + 287, + 470, + 299 + ], + "score": 1.0, + "content": "complex nonlinear physical systems over long time frames. We propose to split", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 297, + 470, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 470, + 311 + ], + "score": 1.0, + "content": "the problem into two distinct tasks: planning and control. To this end, we intro-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "score": 1.0, + "content": "duce a predictor network that plans optimal trajectories and a control network that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 320, + 469, + 331 + ], + "spans": [ + { + "bbox": [ + 142, + 320, + 469, + 331 + ], + "score": 1.0, + "content": "infers the corresponding control parameters. Both stages are trained end-to-end", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 330, + 469, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 469, + 343 + ], + "score": 1.0, + "content": "using a differentiable PDE solver. We demonstrate that our method successfully", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 342, + 469, + 353 + ], + "spans": [ + { + "bbox": [ + 142, + 342, + 469, + 353 + ], + "score": 1.0, + "content": "develops an understanding of complex physical systems and learns to control them", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 352, + 455, + 365 + ], + "spans": [ + { + "bbox": [ + 141, + 352, + 455, + 365 + ], + "score": 1.0, + "content": "for tasks involving PDEs such as the incompressible Navier-Stokes equations.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 15.5, + "bbox_fs": [ + 141, + 209, + 470, + 365 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 383, + 206, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 208, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 208, + 397 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "Intelligent systems that operate in the physical world must be able to perceive, predict, and interact", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "with physical phenomena (Battaglia et al., 2013). In this work, we consider physical systems that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "can be characterized by partial differential equations (PDEs). PDEs constitute the most fundamen-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "tal description of evolving systems and are used to describe every physical theory, from quantum", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "mechanics and general relativity to turbulent flows (Courant & Hilbert, 1962; Smith, 1985). We", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "aim to endow artificial intelligent agents with the ability to direct the evolution of such systems via", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 473, + 190, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 190, + 483 + ], + "score": 1.0, + "content": "continuous controls.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 407, + 505, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "Such optimal control problems have typically been addressed via iterative optimization. Differ-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "entiable solvers and the adjoint method enable efficient optimization of high-dimensional sys-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "tems (Toussaint et al., 2018; de Avila Belbute-Peres et al., 2018; Schenck & Fox, 2018). However,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 523, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 506, + 535 + ], + "score": 1.0, + "content": "direct optimization through gradient descent (single shooting) at test time is resource-intensive and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "may be difficult to deploy in interactive settings. More advanced methods exist, such as multiple", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "shooting and collocation, but they commonly rely on modeling assumptions that limit their applica-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 555, + 439, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 439, + 568 + ], + "score": 1.0, + "content": "bility, and still require computationally intensive iterative optimization at test time.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 489, + 506, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "Iterative optimization methods are expensive because they have to start optimizing from scratch and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "typically require a large number of iterations to reach an optimum. In many real-world control prob-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "lems, however, agents have to repeatedly make decisions in specialized environments, and reaction", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "times are limited to a fraction of a second. This motivates the use of data-driven models such as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "deep neural networks, which combine short inference times with the capacity to build an internal", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 627, + 245, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 245, + 639 + ], + "score": 1.0, + "content": "representation of the environment.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 572, + 505, + 639 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "We present a novel deep learning approach that can learn to represent solution manifolds for a given", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "physical environment, and is orders of magnitude faster than iterative optimization techniques. The", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 667, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 677 + ], + "score": 1.0, + "content": "core of our method is a hierarchical predictor-corrector scheme that temporally divides the problem", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "into easier subproblems. This enables us to combine models specialized to different time scales in", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "order to control long sequences of complex high-dimensional systems. We train our models using a", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "differentiable PDE solver that can provide the agent with feedback of how interactions at any point", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "in time affect the outcome. Our models learn to represent manifolds containing a large number of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 473, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 473, + 733 + ], + "score": 1.0, + "content": "solutions, and can thereby avoid local minima that can trap classic optimization techniques.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 643, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "We evaluate our method on a variety of control tasks in systems governed by advection-diffusion", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "PDEs such as the Navier-Stokes equations. We quantitatively evaluate the resulting sequences on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 504, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 504, + 117 + ], + "score": 1.0, + "content": "how well they approximate the target state and how much force was exerted on the physical system.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 491, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 491, + 128 + ], + "score": 1.0, + "content": "Our method yields stable control for significantly longer time spans than alternative approaches.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 107, + 142, + 200, + 155 + ], + "lines": [ + { + "bbox": [ + 104, + 141, + 201, + 158 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 201, + 158 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 505, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "Physical problems commonly involve nonlinear PDEs, often with many degrees of freedom. In this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 177, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 177, + 506, + 192 + ], + "score": 1.0, + "content": "context, several works have proposed methods for improving the solution of PDE problems (Long", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 189, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 201 + ], + "score": 1.0, + "content": "et al., 2018; Bar-Sinai et al., 2019; Hsieh et al., 2019) or used PDE formulations for unsupervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 200, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 212 + ], + "score": 1.0, + "content": "optimization (Raissi et al., 2018). Lagrangian fluid simulation has been tackled with regression", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "forests (Ladicky et al., 2015), graph neural networks (Mrowca et al., 2018; Li et al., 2019), and con-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 504, + 233 + ], + "score": 1.0, + "content": "tinuous convolutions (Ummenhofer et al., 2020). Data-driven turbulence models were trained with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "score": 1.0, + "content": "MLPs (Ling et al., 2016). Fully-convolutional networks were trained for pressure inference (Tomp-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "son et al., 2017) and advection components were used in adversarial settings (Xie et al., 2018).", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "Temporal updates in reduced spaces were learned via the Koopman operator (Morton et al., 2018).", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 265, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 506, + 278 + ], + "score": 1.0, + "content": "In a related area, deep networks have been used to predict chemical properties and the outcome of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 362, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 362, + 289 + ], + "score": 1.0, + "content": "chemical reactions (Gilmer et al., 2017; Bradshaw et al., 2019).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "Differentiable solvers have been shown to be useful in a variety of settings. Degrave et al. (2019)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "and de Avila Belbute-Peres et al. (2018) developed differentiable simulators for rigid body me-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "chanics. (See Popovic et al. (2000) for earlier work in computer graphics.) Toussaint et al. (2018)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "applied related techniques to manipulation planning. Specialized solvers were developed to infer", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 349 + ], + "score": 1.0, + "content": "protein structures (Ingraham et al., 2019), interact with liquids (Schenck & Fox, 2018), control soft", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "robots (Hu et al., 2019), and solve inverse problems that involve cloth (Liang et al., 2019). Like ours,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "these works typically leverage the automatic differentiation of deep learning pipelines (Griewank", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "& Walther, 2008; Maclaurin et al., 2015; Amos & Kolter, 2017; Mensch & Blondel, 2018; van", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Merrienboer et al. ¨ , 2018; Chen et al., 2018; Bradbury et al., 2018; Paszke et al., 2019; Tokui et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "2019). However, while the works above target Lagrangian solvers, i.e. reference frames moving", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "with the simulated material, we address grid-based solvers, which are particularly appropriate for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 414, + 231, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 231, + 427 + ], + "score": 1.0, + "content": "dense, volumetric phenomena.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 431, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 444 + ], + "score": 1.0, + "content": "The adjoint method (Lions, 1971; Pironneau, 1974; Jameson, 1988; Giles & Pierce, 2000; Bewley,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 441, + 504, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 504, + 454 + ], + "score": 1.0, + "content": "2001; McNamara et al., 2004) is used by most machine learning frameworks, where it is commonly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "known as reverse mode differentiation (Werbos, 2006; Chen et al., 2018). While a variety of spe-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "cialized adjoint solvers exist (Griewank et al., 1996; Fournier et al., 2012; Farrell et al., 2013), these", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "packages do not interface with production machine learning frameworks. A supporting contribu-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 321, + 498 + ], + "score": 1.0, + "content": "tion of our work is a differentiable PDE solver called", + "type": "text" + }, + { + "bbox": [ + 322, + 487, + 345, + 497 + ], + "score": 0.9, + "content": "\\Phi _ { \\mathrm { F l o w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "that integrates with TensorFlow (Abadi", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "et al., 2016) and PyTorch (Paszke et al., 2019). It is publicly available at https://github.com/tum-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 162, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 162, + 520 + ], + "score": 1.0, + "content": "pbs/PhiFlow.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 107, + 535, + 177, + 548 + ], + "lines": [ + { + "bbox": [ + 104, + 533, + 178, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 178, + 551 + ], + "score": 1.0, + "content": "3 PROBLEM", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 443, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 446, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 217, + 573 + ], + "score": 1.0, + "content": "Consider a physical system", + "type": "text" + }, + { + "bbox": [ + 218, + 560, + 248, + 572 + ], + "score": 0.93, + "content": "{ \\pmb u } ( { \\pmb x } , t )", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 559, + 446, + 573 + ], + "score": 1.0, + "content": "whose natural evolution is described by the PDE", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 576, + 377, + 604 + ], + "lines": [ + { + "bbox": [ + 233, + 576, + 377, + 604 + ], + "spans": [ + { + "bbox": [ + 233, + 576, + 377, + 604 + ], + "score": 0.95, + "content": "\\frac { \\partial \\pmb { u } } { \\partial t } = \\mathcal { P } \\left( \\pmb { u } , \\frac { \\partial \\pmb { u } } { \\partial \\pmb { x } } , \\frac { \\partial ^ { 2 } \\pmb { u } } { \\partial \\pmb { x } ^ { 2 } } , . . . , \\pmb { y } ( t ) \\right) ,", + "type": "interline_equation", + "image_path": "250455228ebaa194791dce9e4e390d57f3f9ba2e523968e7139740ecb2979fa3.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 576, + 377, + 590.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 233, + 590.0, + 377, + 604.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 134, + 620 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 608, + 144, + 618 + ], + "score": 0.81, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 607, + 347, + 620 + ], + "score": 1.0, + "content": "models the physical behavior of the system and", + "type": "text" + }, + { + "bbox": [ + 348, + 608, + 366, + 620 + ], + "score": 0.92, + "content": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { y } } } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "denotes external factors that can", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "score": 1.0, + "content": "influence the system. We now introduce an agent that can interact with the system by controlling", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 631, + 504, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 504, + 641 + ], + "score": 1.0, + "content": "certain parameters of the dynamics. This could be the rotation of a motor or fine-grained control", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 640, + 386, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 338, + 654 + ], + "score": 1.0, + "content": "over a field. We factor out this influence into a force term", + "type": "text" + }, + { + "bbox": [ + 338, + 641, + 347, + 650 + ], + "score": 0.82, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 640, + 386, + 654 + ], + "score": 1.0, + "content": ", yielding", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 656, + 382, + 684 + ], + "lines": [ + { + "bbox": [ + 229, + 656, + 382, + 684 + ], + "spans": [ + { + "bbox": [ + 229, + 656, + 382, + 684 + ], + "score": 0.95, + "content": "\\frac { \\partial \\pmb { u } } { \\partial t } = \\mathcal { P } \\left( \\pmb { u } , \\frac { \\partial \\pmb { u } } { \\partial \\pmb { x } } , \\frac { \\partial ^ { 2 } \\pmb { u } } { \\partial \\pmb { x } ^ { 2 } } , . . . \\right) + \\pmb { F } ( t ) .", + "type": "interline_equation", + "image_path": "f1c8f1ec6f4b56380ee05f8b6c422bac600cfdaadd9c09749d8a5d1409fd963a.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 656, + 382, + 670.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 229, + 670.0, + 382, + 684.0 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 350, + 699 + ], + "score": 1.0, + "content": "The agent can now be modelled as a function that computes", + "type": "text" + }, + { + "bbox": [ + 350, + 687, + 371, + 700 + ], + "score": 0.91, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 687, + 505, + 699 + ], + "score": 1.0, + "content": ". As solutions of nonlinear PDEs", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "were shown to yield low-dimensional manifolds (Foias et al., 1988; Titi, 1990), we target solution", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 161, + 722 + ], + "score": 1.0, + "content": "manifolds of", + "type": "text" + }, + { + "bbox": [ + 161, + 710, + 182, + 722 + ], + "score": 0.92, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 710, + 271, + 722 + ], + "score": 1.0, + "content": "for a given choice of", + "type": "text" + }, + { + "bbox": [ + 272, + 711, + 280, + 720 + ], + "score": 0.85, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "with suitable boundary conditions. This motivates our", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 721, + 298, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 298, + 733 + ], + "score": 1.0, + "content": "choice to employ deep networks for our agents.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "We evaluate our method on a variety of control tasks in systems governed by advection-diffusion", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "PDEs such as the Navier-Stokes equations. We quantitatively evaluate the resulting sequences on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 504, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 504, + 117 + ], + "score": 1.0, + "content": "how well they approximate the target state and how much force was exerted on the physical system.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 491, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 491, + 128 + ], + "score": 1.0, + "content": "Our method yields stable control for significantly longer time spans than alternative approaches.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 505, + 128 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 142, + 200, + 155 + ], + "lines": [ + { + "bbox": [ + 104, + 141, + 201, + 158 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 201, + 158 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 505, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "Physical problems commonly involve nonlinear PDEs, often with many degrees of freedom. In this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 177, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 177, + 506, + 192 + ], + "score": 1.0, + "content": "context, several works have proposed methods for improving the solution of PDE problems (Long", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 189, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 201 + ], + "score": 1.0, + "content": "et al., 2018; Bar-Sinai et al., 2019; Hsieh et al., 2019) or used PDE formulations for unsupervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 200, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 212 + ], + "score": 1.0, + "content": "optimization (Raissi et al., 2018). Lagrangian fluid simulation has been tackled with regression", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "forests (Ladicky et al., 2015), graph neural networks (Mrowca et al., 2018; Li et al., 2019), and con-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 504, + 233 + ], + "score": 1.0, + "content": "tinuous convolutions (Ummenhofer et al., 2020). Data-driven turbulence models were trained with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "score": 1.0, + "content": "MLPs (Ling et al., 2016). Fully-convolutional networks were trained for pressure inference (Tomp-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "son et al., 2017) and advection components were used in adversarial settings (Xie et al., 2018).", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "Temporal updates in reduced spaces were learned via the Koopman operator (Morton et al., 2018).", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 265, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 506, + 278 + ], + "score": 1.0, + "content": "In a related area, deep networks have been used to predict chemical properties and the outcome of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 362, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 362, + 289 + ], + "score": 1.0, + "content": "chemical reactions (Gilmer et al., 2017; Bradshaw et al., 2019).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 167, + 506, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "Differentiable solvers have been shown to be useful in a variety of settings. Degrave et al. (2019)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "and de Avila Belbute-Peres et al. (2018) developed differentiable simulators for rigid body me-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "chanics. (See Popovic et al. (2000) for earlier work in computer graphics.) Toussaint et al. (2018)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "applied related techniques to manipulation planning. Specialized solvers were developed to infer", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 349 + ], + "score": 1.0, + "content": "protein structures (Ingraham et al., 2019), interact with liquids (Schenck & Fox, 2018), control soft", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "robots (Hu et al., 2019), and solve inverse problems that involve cloth (Liang et al., 2019). Like ours,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "these works typically leverage the automatic differentiation of deep learning pipelines (Griewank", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "& Walther, 2008; Maclaurin et al., 2015; Amos & Kolter, 2017; Mensch & Blondel, 2018; van", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Merrienboer et al. ¨ , 2018; Chen et al., 2018; Bradbury et al., 2018; Paszke et al., 2019; Tokui et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "2019). However, while the works above target Lagrangian solvers, i.e. reference frames moving", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "with the simulated material, we address grid-based solvers, which are particularly appropriate for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 414, + 231, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 231, + 427 + ], + "score": 1.0, + "content": "dense, volumetric phenomena.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 293, + 506, + 427 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 431, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 444 + ], + "score": 1.0, + "content": "The adjoint method (Lions, 1971; Pironneau, 1974; Jameson, 1988; Giles & Pierce, 2000; Bewley,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 441, + 504, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 504, + 454 + ], + "score": 1.0, + "content": "2001; McNamara et al., 2004) is used by most machine learning frameworks, where it is commonly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "known as reverse mode differentiation (Werbos, 2006; Chen et al., 2018). While a variety of spe-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "cialized adjoint solvers exist (Griewank et al., 1996; Fournier et al., 2012; Farrell et al., 2013), these", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "packages do not interface with production machine learning frameworks. A supporting contribu-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 321, + 498 + ], + "score": 1.0, + "content": "tion of our work is a differentiable PDE solver called", + "type": "text" + }, + { + "bbox": [ + 322, + 487, + 345, + 497 + ], + "score": 0.9, + "content": "\\Phi _ { \\mathrm { F l o w } }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "that integrates with TensorFlow (Abadi", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "et al., 2016) and PyTorch (Paszke et al., 2019). It is publicly available at https://github.com/tum-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 162, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 162, + 520 + ], + "score": 1.0, + "content": "pbs/PhiFlow.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 104, + 429, + 505, + 520 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 535, + 177, + 548 + ], + "lines": [ + { + "bbox": [ + 104, + 533, + 178, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 178, + 551 + ], + "score": 1.0, + "content": "3 PROBLEM", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 443, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 446, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 217, + 573 + ], + "score": 1.0, + "content": "Consider a physical system", + "type": "text" + }, + { + "bbox": [ + 218, + 560, + 248, + 572 + ], + "score": 0.93, + "content": "{ \\pmb u } ( { \\pmb x } , t )", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 559, + 446, + 573 + ], + "score": 1.0, + "content": "whose natural evolution is described by the PDE", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 559, + 446, + 573 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 576, + 377, + 604 + ], + "lines": [ + { + "bbox": [ + 233, + 576, + 377, + 604 + ], + "spans": [ + { + "bbox": [ + 233, + 576, + 377, + 604 + ], + "score": 0.95, + "content": "\\frac { \\partial \\pmb { u } } { \\partial t } = \\mathcal { P } \\left( \\pmb { u } , \\frac { \\partial \\pmb { u } } { \\partial \\pmb { x } } , \\frac { \\partial ^ { 2 } \\pmb { u } } { \\partial \\pmb { x } ^ { 2 } } , . . . , \\pmb { y } ( t ) \\right) ,", + "type": "interline_equation", + "image_path": "250455228ebaa194791dce9e4e390d57f3f9ba2e523968e7139740ecb2979fa3.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 576, + 377, + 590.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 233, + 590.0, + 377, + 604.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 134, + 620 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 608, + 144, + 618 + ], + "score": 0.81, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 607, + 347, + 620 + ], + "score": 1.0, + "content": "models the physical behavior of the system and", + "type": "text" + }, + { + "bbox": [ + 348, + 608, + 366, + 620 + ], + "score": 0.92, + "content": "\\mathbf { \\nabla } _ { \\mathbf { \\boldsymbol { y } } } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "denotes external factors that can", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "score": 1.0, + "content": "influence the system. We now introduce an agent that can interact with the system by controlling", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 631, + 504, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 504, + 641 + ], + "score": 1.0, + "content": "certain parameters of the dynamics. This could be the rotation of a motor or fine-grained control", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 640, + 386, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 338, + 654 + ], + "score": 1.0, + "content": "over a field. We factor out this influence into a force term", + "type": "text" + }, + { + "bbox": [ + 338, + 641, + 347, + 650 + ], + "score": 0.82, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 640, + 386, + 654 + ], + "score": 1.0, + "content": ", yielding", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 607, + 505, + 654 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 656, + 382, + 684 + ], + "lines": [ + { + "bbox": [ + 229, + 656, + 382, + 684 + ], + "spans": [ + { + "bbox": [ + 229, + 656, + 382, + 684 + ], + "score": 0.95, + "content": "\\frac { \\partial \\pmb { u } } { \\partial t } = \\mathcal { P } \\left( \\pmb { u } , \\frac { \\partial \\pmb { u } } { \\partial \\pmb { x } } , \\frac { \\partial ^ { 2 } \\pmb { u } } { \\partial \\pmb { x } ^ { 2 } } , . . . \\right) + \\pmb { F } ( t ) .", + "type": "interline_equation", + "image_path": "f1c8f1ec6f4b56380ee05f8b6c422bac600cfdaadd9c09749d8a5d1409fd963a.jpg" + } + ] + } + ], + "index": 44.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 656, + 382, + 670.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 229, + 670.0, + 382, + 684.0 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 350, + 699 + ], + "score": 1.0, + "content": "The agent can now be modelled as a function that computes", + "type": "text" + }, + { + "bbox": [ + 350, + 687, + 371, + 700 + ], + "score": 0.91, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 687, + 505, + 699 + ], + "score": 1.0, + "content": ". As solutions of nonlinear PDEs", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "were shown to yield low-dimensional manifolds (Foias et al., 1988; Titi, 1990), we target solution", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 161, + 722 + ], + "score": 1.0, + "content": "manifolds of", + "type": "text" + }, + { + "bbox": [ + 161, + 710, + 182, + 722 + ], + "score": 0.92, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 710, + 271, + 722 + ], + "score": 1.0, + "content": "for a given choice of", + "type": "text" + }, + { + "bbox": [ + 272, + 711, + 280, + 720 + ], + "score": 0.85, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "with suitable boundary conditions. This motivates our", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 721, + 298, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 298, + 733 + ], + "score": 1.0, + "content": "choice to employ deep networks for our agents.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5, + "bbox_fs": [ + 106, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "In most real-world scenarios, it is not possible to observe the full state of a physical system. When", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "considering a cloud of smoke, for example, the smoke density may be observable while the velocity", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "field may not be seen directly. We model the imperfect information by defining the observable state", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 117, + 127 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 117, + 125, + 126 + ], + "score": 0.78, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 115, + 136, + 127 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 137, + 115, + 158, + 127 + ], + "score": 0.91, + "content": "\\pmb { o } ( \\pmb { u } )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 115, + 505, + 127 + ], + "score": 1.0, + "content": ". The observable state is problem dependent, and our agent is conditioned only on these", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 343, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 331, + 139 + ], + "score": 1.0, + "content": "observations, i.e. it does not have access to the full state", + "type": "text" + }, + { + "bbox": [ + 331, + 128, + 339, + 136 + ], + "score": 0.72, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 126, + 343, + 139 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 466, + 155 + ], + "score": 1.0, + "content": "Using the above notation, we define the control task as follows. An initial observable state", + "type": "text" + }, + { + "bbox": [ + 467, + 145, + 478, + 154 + ], + "score": 0.86, + "content": "\\scriptstyle { o _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 143, + 505, + 155 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 220, + 167 + ], + "score": 1.0, + "content": "PDE as well as a target state", + "type": "text" + }, + { + "bbox": [ + 220, + 155, + 231, + 164 + ], + "score": 0.87, + "content": "\\ b { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "are given (Figure 1a). We are interested in a reconstructed trajectory", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 126, + 177 + ], + "score": 0.91, + "content": "{ \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\Psi \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\Psi \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf { \\Psi \\mathbf } \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 165, + 239, + 177 + ], + "score": 1.0, + "content": "that matches these states at", + "type": "text" + }, + { + "bbox": [ + 239, + 166, + 249, + 177 + ], + "score": 0.87, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 165, + 267, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 267, + 166, + 276, + 176 + ], + "score": 0.86, + "content": "t _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 165, + 296, + 177 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 297, + 165, + 424, + 177 + ], + "score": 0.91, + "content": "\\pmb { o } _ { 0 } = \\pmb { o } ( \\pmb { u } ( t _ { 0 } ) ) , \\pmb { o } ^ { * } = \\pmb { o } ( \\pmb { u } ( t _ { * } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 165, + 505, + 177 + ], + "score": 1.0, + "content": ", and minimizes the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 385, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 324, + 189 + ], + "score": 1.0, + "content": "amount of force applied within the simulation domain", + "type": "text" + }, + { + "bbox": [ + 324, + 177, + 333, + 186 + ], + "score": 0.78, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 176, + 385, + 189 + ], + "score": 1.0, + "content": "(Figure 1b):", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 191, + 379, + 221 + ], + "lines": [ + { + "bbox": [ + 232, + 191, + 379, + 221 + ], + "spans": [ + { + "bbox": [ + 232, + 191, + 379, + 221 + ], + "score": 0.95, + "content": "L _ { F } [ { \\pmb u } ( t ) ] = \\int _ { t _ { 0 } } ^ { t _ { * } } \\int _ { \\mathcal { D } } | { \\pmb F } _ { \\pmb u } ( t ) | ^ { 2 } d x d t .", + "type": "interline_equation", + "image_path": "08ed5b49e32a508ea2abaca648ae486eb750d47db49b0116ed5307140487d26c.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 232, + 191, + 379, + 206.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 232, + 206.0, + 379, + 221.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 231, + 296, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 230, + 297, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 210, + 243 + ], + "score": 1.0, + "content": "Taking discrete time steps", + "type": "text" + }, + { + "bbox": [ + 211, + 231, + 223, + 241 + ], + "score": 0.84, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 230, + 297, + 243 + ], + "score": 1.0, + "content": ", the reconstructed", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 241, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 147, + 254 + ], + "score": 1.0, + "content": "trajectory", + "type": "text" + }, + { + "bbox": [ + 147, + 244, + 155, + 252 + ], + "score": 0.75, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 241, + 222, + 254 + ], + "score": 1.0, + "content": "is a sequence of", + "type": "text" + }, + { + "bbox": [ + 222, + 242, + 296, + 254 + ], + "score": 0.91, + "content": "n = ( t _ { * } - t _ { 0 } ) / \\Delta t", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 254, + 134, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 134, + 264 + ], + "score": 1.0, + "content": "states.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 296, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 297, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 297, + 282 + ], + "score": 1.0, + "content": "When an observable dimension cannot be con-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 280, + 297, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 297, + 293 + ], + "score": 1.0, + "content": "trolled directly, there may not exist any trajec-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 297, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 126, + 303 + ], + "score": 1.0, + "content": "tory", + "type": "text" + }, + { + "bbox": [ + 126, + 291, + 145, + 304 + ], + "score": 0.92, + "content": "{ \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi { } \\mathbf \\Psi { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { \\Psi } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\Psi \\mathbf { } \\mathbf \\mathbf \\mathbf { \\Psi \\mathbf } \\mathbf \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf { \\mathbf } \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 292, + 225, + 303 + ], + "score": 1.0, + "content": "that matches both", + "type": "text" + }, + { + "bbox": [ + 225, + 293, + 236, + 303 + ], + "score": 0.86, + "content": "\\scriptstyle \\mathbf { o } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 292, + 257, + 303 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 292, + 268, + 302 + ], + "score": 0.86, + "content": "\\boldsymbol { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 292, + 297, + 303 + ], + "score": 1.0, + "content": ". This", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 303, + 297, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 297, + 314 + ], + "score": 1.0, + "content": "can stem from either physical constraints or nu-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 313, + 297, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 297, + 325 + ], + "score": 1.0, + "content": "merical limitations. In these cases, we settle for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 297, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 189, + 338 + ], + "score": 1.0, + "content": "an approximation of", + "type": "text" + }, + { + "bbox": [ + 189, + 325, + 200, + 335 + ], + "score": 0.86, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 324, + 297, + 338 + ], + "score": 1.0, + "content": ". To measure the quality", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 335, + 297, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 297, + 348 + ], + "score": 1.0, + "content": "of the approximation of the target, we define an", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 346, + 297, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 174, + 358 + ], + "score": 1.0, + "content": "observation loss", + "type": "text" + }, + { + "bbox": [ + 174, + 347, + 187, + 358 + ], + "score": 0.89, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 346, + 297, + 358 + ], + "score": 1.0, + "content": ". The form of this loss can", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 357, + 298, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 298, + 369 + ], + "score": 1.0, + "content": "be chosen to fit the problem. We combine Eq. 3", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 367, + 297, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 297, + 380 + ], + "score": 1.0, + "content": "and the observation loss into the objective func-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 125, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 125, + 391 + ], + "score": 1.0, + "content": "tion", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 118, + 396, + 272, + 410 + ], + "lines": [ + { + "bbox": [ + 118, + 396, + 272, + 410 + ], + "spans": [ + { + "bbox": [ + 118, + 396, + 272, + 410 + ], + "score": 0.9, + "content": "\\begin{array} { r } { L [ { \\pmb u } ( t ) ] = \\alpha \\cdot L _ { \\pmb { F } } [ { \\pmb u } ( t ) ] + L _ { o } ^ { * } ( { \\pmb u } ( t _ { * } ) ) , } \\end{array}", + "type": "interline_equation", + "image_path": "74c8a451b52e116474c274bb9717655c98b10ffd0c7e498c6a7d182f95248e35.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 118, + 396, + 272, + 410 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "image", + "bbox": [ + 314, + 235, + 494, + 339 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 314, + 235, + 494, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 314, + 235, + 494, + 339 + ], + "spans": [ + { + "bbox": [ + 314, + 235, + 494, + 339 + ], + "score": 0.967, + "type": "image", + "image_path": "04f9b9adaa901eb5c66f102a4edbde131607e1474dca23bebf6fc936d4486367.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 314, + 235, + 494, + 248.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 314, + 248.0, + 494, + 261.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 314, + 261.0, + 494, + 274.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 314, + 274.0, + 494, + 287.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 314, + 287.0, + 494, + 300.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 314, + 300.0, + 494, + 313.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 314, + 313.0, + 494, + 326.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 314, + 326.0, + 494, + 339.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 303, + 351, + 505, + 406 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "Figure 1: Illustration of possible trajectories. The", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 303, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 303, + 362, + 506, + 374 + ], + "score": 1.0, + "content": "grey lines represent the unperturbed evolution of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "the physical system. The amount of applied force", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 303, + 383, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 303, + 383, + 505, + 397 + ], + "score": 1.0, + "content": "corresponds to how far the trajectory deviates", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 394, + 414, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 414, + 406 + ], + "score": 1.0, + "content": "from the natural evolution.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + } + ], + "index": 32.75 + }, + { + "type": "text", + "bbox": [ + 105, + 414, + 504, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 127, + 427 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 128, + 414, + 157, + 424 + ], + "score": 0.91, + "content": "\\alpha > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 412, + 506, + 427 + ], + "score": 1.0, + "content": ". We use square brackets to denote functionals, i.e. functions depending on fields or", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 425, + 233, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 233, + 437 + ], + "score": 1.0, + "content": "series rather than single values.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 107, + 452, + 208, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 209, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 209, + 466 + ], + "score": 1.0, + "content": "4 PRELIMINARIES", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 220, + 490 + ], + "score": 1.0, + "content": "Differentiable solvers. Let", + "type": "text" + }, + { + "bbox": [ + 220, + 477, + 250, + 489 + ], + "score": 0.93, + "content": "{ \\pmb u } ( { \\pmb x } , t )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "be described by a PDE as in Eq. 1. A regular solver can move", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 487, + 280, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 280, + 501 + ], + "score": 1.0, + "content": "the system forward in time via Euler steps:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 504, + 442, + 518 + ], + "lines": [ + { + "bbox": [ + 169, + 504, + 442, + 518 + ], + "spans": [ + { + "bbox": [ + 169, + 504, + 442, + 518 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\pmb { u } ( t _ { i + 1 } ) = \\mathrm { S o l v e r } [ \\pmb { u } ( t _ { i } ) , \\pmb { y } ( t _ { i } ) ] = \\pmb { u } ( t _ { i } ) + \\Delta t \\cdot \\mathcal { P } \\left( \\pmb { u } ( t _ { i } ) , . . . , \\pmb { y } ( t _ { i } ) \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "c7447aaeb2d69df2e41bf0c43d0175255694212e592d26c570785dda908f0fa4.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 169, + 504, + 442, + 518 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 342, + 536 + ], + "score": 1.0, + "content": "Each step moves the system forward by a time increment", + "type": "text" + }, + { + "bbox": [ + 342, + 523, + 355, + 532 + ], + "score": 0.81, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 522, + 505, + 536 + ], + "score": 1.0, + "content": ". Repeated execution produces a tra-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 137, + 547 + ], + "score": 1.0, + "content": "jectory", + "type": "text" + }, + { + "bbox": [ + 138, + 533, + 155, + 545 + ], + "score": 0.91, + "content": "u ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 532, + 505, + 547 + ], + "score": 1.0, + "content": "that approximates a solution to the PDE. This functionality for time advancement by", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "itself is not well-suited to solve optimization problems, since gradients can only be approximated", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "by finite differencing. For high-dimensional or continuous systems, this method becomes computa-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 567, + 496, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 496, + 578 + ], + "score": 1.0, + "content": "tionally expensive because a full trajectory needs to be computed for each optimizable parameter.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 504, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "Differentiable solvers resolve this issue by solving the adjoint problem (Pontryagin, 1962) via ana-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "lytic derivatives. The adjoint problem computes the same mathematical expressions while working", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "with lower-dimensional vectors. A differentiable solver can efficiently compute the derivatives with", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 235, + 629 + ], + "score": 1.0, + "content": "respect to any of its inputs, i.e.", + "type": "text" + }, + { + "bbox": [ + 236, + 616, + 306, + 628 + ], + "score": 0.93, + "content": "\\partial \\pmb { u } ( t _ { i + 1 } ) / \\partial \\pmb { u } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 616, + 326, + 629 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 326, + 616, + 396, + 628 + ], + "score": 0.93, + "content": "\\partial \\pmb { u } ( t _ { i + 1 } ) / \\partial \\pmb { y } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 616, + 506, + 629 + ], + "score": 1.0, + "content": ". This allows for gradient-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 626, + 467, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 467, + 640 + ], + "score": 1.0, + "content": "based optimization of inputs or control parameters over an arbitrary number of time steps.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52 + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 657 + ], + "score": 1.0, + "content": "Iterative trajectory optimization. Many techniques exist that try to find optimal trajectories by", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 236, + 667 + ], + "score": 1.0, + "content": "starting with an initial guess for", + "type": "text" + }, + { + "bbox": [ + 236, + 655, + 257, + 667 + ], + "score": 0.92, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "and slightly changing it until reaching an optimum. The sim-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "plest of these is known as single shooting. In one optimization step, it simulates the full dynamics,", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "then backpropagates the loss through the whole sequence to optimize the controls (Kraft, 1985;", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 251, + 701 + ], + "score": 1.0, + "content": "Leineweber et al., 2003). Replacing", + "type": "text" + }, + { + "bbox": [ + 252, + 689, + 272, + 700 + ], + "score": 0.9, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 687, + 329, + 701 + ], + "score": 1.0, + "content": "with an agent", + "type": "text" + }, + { + "bbox": [ + 329, + 688, + 375, + 700 + ], + "score": 0.93, + "content": "\\mathbf { } F ( t | o _ { t } , o ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 687, + 505, + 701 + ], + "score": 1.0, + "content": ", which can be parameterized by", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 104, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 379, + 712 + ], + "score": 1.0, + "content": "a deep network, yields a simple training method. For a sequence of", + "type": "text" + }, + { + "bbox": [ + 379, + 701, + 386, + 709 + ], + "score": 0.78, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 698, + 496, + 712 + ], + "score": 1.0, + "content": "frames, this setup contains", + "type": "text" + }, + { + "bbox": [ + 497, + 701, + 504, + 709 + ], + "score": 0.67, + "content": "n", + "type": "inline_equation" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "linked copies of the agent and is depicted in Figure 2. We refer to such an agent as a control force", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 106, + 721, + 176, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 176, + 732 + ], + "score": 1.0, + "content": "estimator (CFE).", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 58.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "In most real-world scenarios, it is not possible to observe the full state of a physical system. When", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "considering a cloud of smoke, for example, the smoke density may be observable while the velocity", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "field may not be seen directly. We model the imperfect information by defining the observable state", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 117, + 127 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 117, + 125, + 126 + ], + "score": 0.78, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 115, + 136, + 127 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 137, + 115, + 158, + 127 + ], + "score": 0.91, + "content": "\\pmb { o } ( \\pmb { u } )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 115, + 505, + 127 + ], + "score": 1.0, + "content": ". The observable state is problem dependent, and our agent is conditioned only on these", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 343, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 331, + 139 + ], + "score": 1.0, + "content": "observations, i.e. it does not have access to the full state", + "type": "text" + }, + { + "bbox": [ + 331, + 128, + 339, + 136 + ], + "score": 0.72, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 126, + 343, + 139 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 139 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 466, + 155 + ], + "score": 1.0, + "content": "Using the above notation, we define the control task as follows. An initial observable state", + "type": "text" + }, + { + "bbox": [ + 467, + 145, + 478, + 154 + ], + "score": 0.86, + "content": "\\scriptstyle { o _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 143, + 505, + 155 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 220, + 167 + ], + "score": 1.0, + "content": "PDE as well as a target state", + "type": "text" + }, + { + "bbox": [ + 220, + 155, + 231, + 164 + ], + "score": 0.87, + "content": "\\ b { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "are given (Figure 1a). We are interested in a reconstructed trajectory", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 126, + 177 + ], + "score": 0.91, + "content": "{ \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf \\Psi \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\Psi \\mathbf { } \\mathbf \\mathbf \\mathbf { } \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf { \\Psi \\mathbf } \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 165, + 239, + 177 + ], + "score": 1.0, + "content": "that matches these states at", + "type": "text" + }, + { + "bbox": [ + 239, + 166, + 249, + 177 + ], + "score": 0.87, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 165, + 267, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 267, + 166, + 276, + 176 + ], + "score": 0.86, + "content": "t _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 165, + 296, + 177 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 297, + 165, + 424, + 177 + ], + "score": 0.91, + "content": "\\pmb { o } _ { 0 } = \\pmb { o } ( \\pmb { u } ( t _ { 0 } ) ) , \\pmb { o } ^ { * } = \\pmb { o } ( \\pmb { u } ( t _ { * } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 165, + 505, + 177 + ], + "score": 1.0, + "content": ", and minimizes the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 385, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 324, + 189 + ], + "score": 1.0, + "content": "amount of force applied within the simulation domain", + "type": "text" + }, + { + "bbox": [ + 324, + 177, + 333, + 186 + ], + "score": 0.78, + "content": "\\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 176, + 385, + 189 + ], + "score": 1.0, + "content": "(Figure 1b):", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 143, + 505, + 189 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 191, + 379, + 221 + ], + "lines": [ + { + "bbox": [ + 232, + 191, + 379, + 221 + ], + "spans": [ + { + "bbox": [ + 232, + 191, + 379, + 221 + ], + "score": 0.95, + "content": "L _ { F } [ { \\pmb u } ( t ) ] = \\int _ { t _ { 0 } } ^ { t _ { * } } \\int _ { \\mathcal { D } } | { \\pmb F } _ { \\pmb u } ( t ) | ^ { 2 } d x d t .", + "type": "interline_equation", + "image_path": "08ed5b49e32a508ea2abaca648ae486eb750d47db49b0116ed5307140487d26c.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 232, + 191, + 379, + 206.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 232, + 206.0, + 379, + 221.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 231, + 296, + 264 + ], + "lines": [ + { + "bbox": [ + 106, + 230, + 297, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 210, + 243 + ], + "score": 1.0, + "content": "Taking discrete time steps", + "type": "text" + }, + { + "bbox": [ + 211, + 231, + 223, + 241 + ], + "score": 0.84, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 230, + 297, + 243 + ], + "score": 1.0, + "content": ", the reconstructed", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 241, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 147, + 254 + ], + "score": 1.0, + "content": "trajectory", + "type": "text" + }, + { + "bbox": [ + 147, + 244, + 155, + 252 + ], + "score": 0.75, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 241, + 222, + 254 + ], + "score": 1.0, + "content": "is a sequence of", + "type": "text" + }, + { + "bbox": [ + 222, + 242, + 296, + 254 + ], + "score": 0.91, + "content": "n = ( t _ { * } - t _ { 0 } ) / \\Delta t", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 254, + 134, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 134, + 264 + ], + "score": 1.0, + "content": "states.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 230, + 297, + 264 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 296, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 297, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 297, + 282 + ], + "score": 1.0, + "content": "When an observable dimension cannot be con-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 280, + 297, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 297, + 293 + ], + "score": 1.0, + "content": "trolled directly, there may not exist any trajec-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 297, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 126, + 303 + ], + "score": 1.0, + "content": "tory", + "type": "text" + }, + { + "bbox": [ + 126, + 291, + 145, + 304 + ], + "score": 0.92, + "content": "{ \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } { \\mathbf { } } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi { } \\mathbf \\Psi { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf { } \\mathbf \\mathbf { \\Psi } \\mathbf \\mathbf \\Psi \\mathbf { } \\mathbf \\mathbf \\Psi \\Psi \\mathbf { } \\mathbf \\mathbf \\mathbf { \\Psi \\mathbf } \\mathbf \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf { \\mathbf } \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 292, + 225, + 303 + ], + "score": 1.0, + "content": "that matches both", + "type": "text" + }, + { + "bbox": [ + 225, + 293, + 236, + 303 + ], + "score": 0.86, + "content": "\\scriptstyle \\mathbf { o } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 292, + 257, + 303 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 292, + 268, + 302 + ], + "score": 0.86, + "content": "\\boldsymbol { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 292, + 297, + 303 + ], + "score": 1.0, + "content": ". This", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 303, + 297, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 297, + 314 + ], + "score": 1.0, + "content": "can stem from either physical constraints or nu-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 313, + 297, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 297, + 325 + ], + "score": 1.0, + "content": "merical limitations. In these cases, we settle for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 297, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 189, + 338 + ], + "score": 1.0, + "content": "an approximation of", + "type": "text" + }, + { + "bbox": [ + 189, + 325, + 200, + 335 + ], + "score": 0.86, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 324, + 297, + 338 + ], + "score": 1.0, + "content": ". To measure the quality", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 335, + 297, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 297, + 348 + ], + "score": 1.0, + "content": "of the approximation of the target, we define an", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 346, + 297, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 174, + 358 + ], + "score": 1.0, + "content": "observation loss", + "type": "text" + }, + { + "bbox": [ + 174, + 347, + 187, + 358 + ], + "score": 0.89, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 346, + 297, + 358 + ], + "score": 1.0, + "content": ". The form of this loss can", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 357, + 298, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 298, + 369 + ], + "score": 1.0, + "content": "be chosen to fit the problem. We combine Eq. 3", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 367, + 297, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 297, + 380 + ], + "score": 1.0, + "content": "and the observation loss into the objective func-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 125, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 125, + 391 + ], + "score": 1.0, + "content": "tion", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 269, + 298, + 391 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 118, + 396, + 272, + 410 + ], + "lines": [ + { + "bbox": [ + 118, + 396, + 272, + 410 + ], + "spans": [ + { + "bbox": [ + 118, + 396, + 272, + 410 + ], + "score": 0.9, + "content": "\\begin{array} { r } { L [ { \\pmb u } ( t ) ] = \\alpha \\cdot L _ { \\pmb { F } } [ { \\pmb u } ( t ) ] + L _ { o } ^ { * } ( { \\pmb u } ( t _ { * } ) ) , } \\end{array}", + "type": "interline_equation", + "image_path": "74c8a451b52e116474c274bb9717655c98b10ffd0c7e498c6a7d182f95248e35.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 118, + 396, + 272, + 410 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "image", + "bbox": [ + 314, + 235, + 494, + 339 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 314, + 235, + 494, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 314, + 235, + 494, + 339 + ], + "spans": [ + { + "bbox": [ + 314, + 235, + 494, + 339 + ], + "score": 0.967, + "type": "image", + "image_path": "04f9b9adaa901eb5c66f102a4edbde131607e1474dca23bebf6fc936d4486367.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 314, + 235, + 494, + 248.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 314, + 248.0, + 494, + 261.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 314, + 261.0, + 494, + 274.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 314, + 274.0, + 494, + 287.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 314, + 287.0, + 494, + 300.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 314, + 300.0, + 494, + 313.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 314, + 313.0, + 494, + 326.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 314, + 326.0, + 494, + 339.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 303, + 351, + 505, + 406 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "Figure 1: Illustration of possible trajectories. The", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 303, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 303, + 362, + 506, + 374 + ], + "score": 1.0, + "content": "grey lines represent the unperturbed evolution of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "the physical system. The amount of applied force", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 303, + 383, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 303, + 383, + 505, + 397 + ], + "score": 1.0, + "content": "corresponds to how far the trajectory deviates", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 394, + 414, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 414, + 406 + ], + "score": 1.0, + "content": "from the natural evolution.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + } + ], + "index": 32.75 + }, + { + "type": "text", + "bbox": [ + 105, + 414, + 504, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 127, + 427 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 128, + 414, + 157, + 424 + ], + "score": 0.91, + "content": "\\alpha > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 412, + 506, + 427 + ], + "score": 1.0, + "content": ". We use square brackets to denote functionals, i.e. functions depending on fields or", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 425, + 233, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 233, + 437 + ], + "score": 1.0, + "content": "series rather than single values.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 412, + 506, + 437 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 452, + 208, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 209, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 209, + 466 + ], + "score": 1.0, + "content": "4 PRELIMINARIES", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 220, + 490 + ], + "score": 1.0, + "content": "Differentiable solvers. Let", + "type": "text" + }, + { + "bbox": [ + 220, + 477, + 250, + 489 + ], + "score": 0.93, + "content": "{ \\pmb u } ( { \\pmb x } , t )", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "be described by a PDE as in Eq. 1. A regular solver can move", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 487, + 280, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 280, + 501 + ], + "score": 1.0, + "content": "the system forward in time via Euler steps:", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 476, + 505, + 501 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 504, + 442, + 518 + ], + "lines": [ + { + "bbox": [ + 169, + 504, + 442, + 518 + ], + "spans": [ + { + "bbox": [ + 169, + 504, + 442, + 518 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\pmb { u } ( t _ { i + 1 } ) = \\mathrm { S o l v e r } [ \\pmb { u } ( t _ { i } ) , \\pmb { y } ( t _ { i } ) ] = \\pmb { u } ( t _ { i } ) + \\Delta t \\cdot \\mathcal { P } \\left( \\pmb { u } ( t _ { i } ) , . . . , \\pmb { y } ( t _ { i } ) \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "c7447aaeb2d69df2e41bf0c43d0175255694212e592d26c570785dda908f0fa4.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 169, + 504, + 442, + 518 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 342, + 536 + ], + "score": 1.0, + "content": "Each step moves the system forward by a time increment", + "type": "text" + }, + { + "bbox": [ + 342, + 523, + 355, + 532 + ], + "score": 0.81, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 522, + 505, + 536 + ], + "score": 1.0, + "content": ". Repeated execution produces a tra-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 137, + 547 + ], + "score": 1.0, + "content": "jectory", + "type": "text" + }, + { + "bbox": [ + 138, + 533, + 155, + 545 + ], + "score": 0.91, + "content": "u ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 532, + 505, + 547 + ], + "score": 1.0, + "content": "that approximates a solution to the PDE. This functionality for time advancement by", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "itself is not well-suited to solve optimization problems, since gradients can only be approximated", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "by finite differencing. For high-dimensional or continuous systems, this method becomes computa-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 567, + 496, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 496, + 578 + ], + "score": 1.0, + "content": "tionally expensive because a full trajectory needs to be computed for each optimizable parameter.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 522, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 504, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "Differentiable solvers resolve this issue by solving the adjoint problem (Pontryagin, 1962) via ana-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "lytic derivatives. The adjoint problem computes the same mathematical expressions while working", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "with lower-dimensional vectors. A differentiable solver can efficiently compute the derivatives with", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 235, + 629 + ], + "score": 1.0, + "content": "respect to any of its inputs, i.e.", + "type": "text" + }, + { + "bbox": [ + 236, + 616, + 306, + 628 + ], + "score": 0.93, + "content": "\\partial \\pmb { u } ( t _ { i + 1 } ) / \\partial \\pmb { u } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 616, + 326, + 629 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 326, + 616, + 396, + 628 + ], + "score": 0.93, + "content": "\\partial \\pmb { u } ( t _ { i + 1 } ) / \\partial \\pmb { y } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 616, + 506, + 629 + ], + "score": 1.0, + "content": ". This allows for gradient-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 626, + 467, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 467, + 640 + ], + "score": 1.0, + "content": "based optimization of inputs or control parameters over an arbitrary number of time steps.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52, + "bbox_fs": [ + 105, + 582, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 657 + ], + "score": 1.0, + "content": "Iterative trajectory optimization. Many techniques exist that try to find optimal trajectories by", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 236, + 667 + ], + "score": 1.0, + "content": "starting with an initial guess for", + "type": "text" + }, + { + "bbox": [ + 236, + 655, + 257, + 667 + ], + "score": 0.92, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "and slightly changing it until reaching an optimum. The sim-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "plest of these is known as single shooting. In one optimization step, it simulates the full dynamics,", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "then backpropagates the loss through the whole sequence to optimize the controls (Kraft, 1985;", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 251, + 701 + ], + "score": 1.0, + "content": "Leineweber et al., 2003). Replacing", + "type": "text" + }, + { + "bbox": [ + 252, + 689, + 272, + 700 + ], + "score": 0.9, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 687, + 329, + 701 + ], + "score": 1.0, + "content": "with an agent", + "type": "text" + }, + { + "bbox": [ + 329, + 688, + 375, + 700 + ], + "score": 0.93, + "content": "\\mathbf { } F ( t | o _ { t } , o ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 687, + 505, + 701 + ], + "score": 1.0, + "content": ", which can be parameterized by", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 104, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 379, + 712 + ], + "score": 1.0, + "content": "a deep network, yields a simple training method. For a sequence of", + "type": "text" + }, + { + "bbox": [ + 379, + 701, + 386, + 709 + ], + "score": 0.78, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 698, + 496, + 712 + ], + "score": 1.0, + "content": "frames, this setup contains", + "type": "text" + }, + { + "bbox": [ + 497, + 701, + 504, + 709 + ], + "score": 0.67, + "content": "n", + "type": "inline_equation" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "linked copies of the agent and is depicted in Figure 2. We refer to such an agent as a control force", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 106, + 721, + 176, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 176, + 732 + ], + "score": 1.0, + "content": "estimator (CFE).", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 58.5, + "bbox_fs": [ + 104, + 642, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "Optimizing such a chain of CFEs is both computationally expensive and causes gradients to pass", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "through a potentially long sequence of highly nonlinear simulation steps. When the reconstruction", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 107, + 106, + 114, + 114 + ], + "score": 0.7, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 105, + 429, + 117 + ], + "score": 1.0, + "content": "is close to an optimal trajectory, this is not a problem because the gradients", + "type": "text" + }, + { + "bbox": [ + 429, + 105, + 446, + 115 + ], + "score": 0.8, + "content": "\\Delta u", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "are small and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "the operations executed by the solver are differentiable by construction. The solver can therefore be", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "locally approximated by a first-order polynomial and the gradients can be safely backpropagated.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 146, + 152 + ], + "score": 1.0, + "content": "For large", + "type": "text" + }, + { + "bbox": [ + 147, + 138, + 163, + 147 + ], + "score": 0.7, + "content": "\\Delta u", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 136, + 506, + 152 + ], + "score": 1.0, + "content": ", e.g. at the beginning of an optimization, this approximation breaks down, causing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 163 + ], + "score": 1.0, + "content": "the gradients to become unstable while passing through the chain. This instability in the training", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "score": 1.0, + "content": "process can prevent single-shooting approaches from converging and deep networks from learning", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 282, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 282, + 184 + ], + "score": 1.0, + "content": "unless they are initialized near an optimum.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "Alternatives to single shooting exist, promising better and more efficient convergence. Multi-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "ple shooting (Bock & Plitt, 1984) splits the trajectory into segments with additional defect con-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "straints. Depending on the physical system, this method may have to be adjusted for specific prob-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "lems (Treuille et al., 2003). Collocation schemes (Hargraves & Paris, 1987) model trajectories with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 242 + ], + "score": 1.0, + "content": "splines. While this works well for particle trajectories, it is poorly suited for Eulerian solvers where", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "the evolution of individual points does not reflect the overall motion. Model reduction can be used", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "to reduce the dimensionality or nonlinearity of the problem, but generally requires domain-specific", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "knowledge. When applicable, these methods can converge faster or in a more stable manner than", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "single shooting. However, as we are focusing on a general optimization scheme in this work, we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 367, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 367, + 298 + ], + "score": 1.0, + "content": "will use single shooting and its variants as baseline comparisons.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "Supervised and differentiable physics losses. One of the key ingredients in training a machine", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 314, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 326 + ], + "score": 1.0, + "content": "learning model is the choice of loss function. For many tasks, supervised losses are used, i.e. losses", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 504, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 504, + 337 + ], + "score": 1.0, + "content": "that directly compare the output of the model for a specific input with the desired ground truth.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "While supervised losses can be employed for trajectory optimization, far better loss functions are", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "score": 1.0, + "content": "possible when a differentiable solver is available. We will refer to these as differentiable physics loss", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "functions. In this work, we employ a combination of supervised and differentiable physics losses,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 368, + 306, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 306, + 381 + ], + "score": 1.0, + "content": "as both come with advantages and disadvantages.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "One key limitation of supervised losses is that they can only measure the error of a single time step.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "Therefore, an agent cannot get any measure of how its output would influence future time steps.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "Another problem arises from the form of supervised training data which comprises input-output", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "pairs, which may be obtained directly from data generation or through iterative optimization. Since", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "optimal control problems are generally not unimodal, there can exist multiple possible outputs for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "one input. This ambiguity in the supervised training process will lead to suboptimal predictions as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 452, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 464 + ], + "score": 1.0, + "content": "the network will try to find a compromise between all possible outputs instead of picking one of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 461, + 132, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 132, + 474 + ], + "score": 1.0, + "content": "them.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 478, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "score": 1.0, + "content": "Differentiable physics losses solve these problems by allowing the agent to be directly optimized", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 489, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 504 + ], + "score": 1.0, + "content": "for the desired objective (Eq. 4). Unlike supervised losses, differentiable physics losses require a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "differentiable solver to backpropagate the gradients through the simulation. Multiple time steps can", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "be chained together, which is a key requirement since the objective (Eq. 4) explicitly depends on all", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 182, + 536 + ], + "score": 1.0, + "content": "time steps through", + "type": "text" + }, + { + "bbox": [ + 182, + 523, + 221, + 535 + ], + "score": 0.93, + "content": "L _ { F } [ { \\pmb u } ( t ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "(Eq. 3). As with iterative solvers, one optimization step for a sequence", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 117, + 546 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 536, + 125, + 544 + ], + "score": 0.72, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 534, + 248, + 546 + ], + "score": 1.0, + "content": "frames then invokes the agent", + "type": "text" + }, + { + "bbox": [ + 248, + 536, + 255, + 544 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "times before computing the loss, each invocation followed by", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "a solver step. The employed differentiable solver backpropagates the gradients through the whole", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "sequence, which gives the model feedback on (i) how its decisions change the future trajectory and", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5 + }, + { + "type": "image", + "bbox": [ + 149, + 584, + 461, + 686 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 149, + 584, + 461, + 686 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 149, + 584, + 461, + 686 + ], + "spans": [ + { + "bbox": [ + 149, + 584, + 461, + 686 + ], + "score": 0.971, + "type": "image", + "image_path": "fe0db1ae7bc69b9fe1a3415a31b4c96f379f1f645c3e225deb79e3c8a50de3a5.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 149, + 584, + 461, + 618.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 149, + 618.0, + 461, + 652.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 149, + 652.0, + 461, + 686.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 696, + 505, + 730 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "Figure 2: Single-shooting optimization with a control force estimator (CFE). (a) The forward pass", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 707, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 505, + 721 + ], + "score": 1.0, + "content": "simulates the full sequence. (b) Backpropagation computes the adjoint problem. (c) Weight updates", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 718, + 271, + 730 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 271, + 730 + ], + "score": 1.0, + "content": "are accumulated and applied to the CFE.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "Optimizing such a chain of CFEs is both computationally expensive and causes gradients to pass", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "through a potentially long sequence of highly nonlinear simulation steps. When the reconstruction", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 107, + 106, + 114, + 114 + ], + "score": 0.7, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 105, + 429, + 117 + ], + "score": 1.0, + "content": "is close to an optimal trajectory, this is not a problem because the gradients", + "type": "text" + }, + { + "bbox": [ + 429, + 105, + 446, + 115 + ], + "score": 0.8, + "content": "\\Delta u", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "are small and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "the operations executed by the solver are differentiable by construction. The solver can therefore be", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "locally approximated by a first-order polynomial and the gradients can be safely backpropagated.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 146, + 152 + ], + "score": 1.0, + "content": "For large", + "type": "text" + }, + { + "bbox": [ + 147, + 138, + 163, + 147 + ], + "score": 0.7, + "content": "\\Delta u", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 136, + 506, + 152 + ], + "score": 1.0, + "content": ", e.g. at the beginning of an optimization, this approximation breaks down, causing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 163 + ], + "score": 1.0, + "content": "the gradients to become unstable while passing through the chain. This instability in the training", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "score": 1.0, + "content": "process can prevent single-shooting approaches from converging and deep networks from learning", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 282, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 282, + 184 + ], + "score": 1.0, + "content": "unless they are initialized near an optimum.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 82, + 506, + 184 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "Alternatives to single shooting exist, promising better and more efficient convergence. Multi-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "ple shooting (Bock & Plitt, 1984) splits the trajectory into segments with additional defect con-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "straints. Depending on the physical system, this method may have to be adjusted for specific prob-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "lems (Treuille et al., 2003). Collocation schemes (Hargraves & Paris, 1987) model trajectories with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 231, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 242 + ], + "score": 1.0, + "content": "splines. While this works well for particle trajectories, it is poorly suited for Eulerian solvers where", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "the evolution of individual points does not reflect the overall motion. Model reduction can be used", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "to reduce the dimensionality or nonlinearity of the problem, but generally requires domain-specific", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "knowledge. When applicable, these methods can converge faster or in a more stable manner than", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "single shooting. However, as we are focusing on a general optimization scheme in this work, we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 367, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 367, + 298 + ], + "score": 1.0, + "content": "will use single shooting and its variants as baseline comparisons.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 187, + 505, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "Supervised and differentiable physics losses. One of the key ingredients in training a machine", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 314, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 326 + ], + "score": 1.0, + "content": "learning model is the choice of loss function. For many tasks, supervised losses are used, i.e. losses", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 504, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 504, + 337 + ], + "score": 1.0, + "content": "that directly compare the output of the model for a specific input with the desired ground truth.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "While supervised losses can be employed for trajectory optimization, far better loss functions are", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 359 + ], + "score": 1.0, + "content": "possible when a differentiable solver is available. We will refer to these as differentiable physics loss", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "functions. In this work, we employ a combination of supervised and differentiable physics losses,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 368, + 306, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 306, + 381 + ], + "score": 1.0, + "content": "as both come with advantages and disadvantages.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 303, + 506, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "One key limitation of supervised losses is that they can only measure the error of a single time step.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "Therefore, an agent cannot get any measure of how its output would influence future time steps.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "Another problem arises from the form of supervised training data which comprises input-output", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 431 + ], + "score": 1.0, + "content": "pairs, which may be obtained directly from data generation or through iterative optimization. Since", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "optimal control problems are generally not unimodal, there can exist multiple possible outputs for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "one input. This ambiguity in the supervised training process will lead to suboptimal predictions as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 452, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 464 + ], + "score": 1.0, + "content": "the network will try to find a compromise between all possible outputs instead of picking one of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 461, + 132, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 132, + 474 + ], + "score": 1.0, + "content": "them.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 385, + 506, + 474 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 478, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "score": 1.0, + "content": "Differentiable physics losses solve these problems by allowing the agent to be directly optimized", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 489, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 504 + ], + "score": 1.0, + "content": "for the desired objective (Eq. 4). Unlike supervised losses, differentiable physics losses require a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "differentiable solver to backpropagate the gradients through the simulation. Multiple time steps can", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "be chained together, which is a key requirement since the objective (Eq. 4) explicitly depends on all", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 182, + 536 + ], + "score": 1.0, + "content": "time steps through", + "type": "text" + }, + { + "bbox": [ + 182, + 523, + 221, + 535 + ], + "score": 0.93, + "content": "L _ { F } [ { \\pmb u } ( t ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "(Eq. 3). As with iterative solvers, one optimization step for a sequence", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 117, + 546 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 536, + 125, + 544 + ], + "score": 0.72, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 534, + 248, + 546 + ], + "score": 1.0, + "content": "frames then invokes the agent", + "type": "text" + }, + { + "bbox": [ + 248, + 536, + 255, + 544 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "times before computing the loss, each invocation followed by", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 557 + ], + "score": 1.0, + "content": "a solver step. The employed differentiable solver backpropagates the gradients through the whole", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "sequence, which gives the model feedback on (i) how its decisions change the future trajectory and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "(ii) how to handle states as input that were reached because of its previous decisions. Since no", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 502, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 502, + 107 + ], + "score": 1.0, + "content": "ground truth needs to be provided, multi-modal problems naturally converge towards one solution.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 478, + 506, + 569 + ] + }, + { + "type": "image", + "bbox": [ + 149, + 584, + 461, + 686 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 149, + 584, + 461, + 686 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 149, + 584, + 461, + 686 + ], + "spans": [ + { + "bbox": [ + 149, + 584, + 461, + 686 + ], + "score": 0.971, + "type": "image", + "image_path": "fe0db1ae7bc69b9fe1a3415a31b4c96f379f1f645c3e225deb79e3c8a50de3a5.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 149, + 584, + 461, + 618.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 149, + 618.0, + 461, + 652.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 149, + 652.0, + 461, + 686.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 696, + 505, + 730 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "Figure 2: Single-shooting optimization with a control force estimator (CFE). (a) The forward pass", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 707, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 505, + 721 + ], + "score": 1.0, + "content": "simulates the full sequence. (b) Backpropagation computes the adjoint problem. (c) Weight updates", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 718, + 271, + 730 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 271, + 730 + ], + "score": 1.0, + "content": "are accumulated and applied to the CFE.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "(ii) how to handle states as input that were reached because of its previous decisions. Since no", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 502, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 502, + 107 + ], + "score": 1.0, + "content": "ground truth needs to be provided, multi-modal problems naturally converge towards one solution.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 126, + 173, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 124, + 174, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 174, + 140 + ], + "score": 1.0, + "content": "5 METHOD", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 153, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "In order to optimally interact with a physical system, an agent has to (i) build an internal repre-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 164, + 504, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 292, + 176 + ], + "score": 1.0, + "content": "sentation of an optimal observable trajectory", + "type": "text" + }, + { + "bbox": [ + 293, + 164, + 325, + 177 + ], + "score": 0.93, + "content": "o ( { \\boldsymbol { u } } ( t ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 165, + 504, + 176 + ], + "score": 1.0, + "content": "and (ii) learn what actions to take to move", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 177, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 188 + ], + "score": 1.0, + "content": "the system along the desired trajectory. These two steps strongly resemble the predictor-corrector", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 249, + 199 + ], + "score": 1.0, + "content": "method (Press et al., 2007). Given", + "type": "text" + }, + { + "bbox": [ + 250, + 187, + 268, + 199 + ], + "score": 0.91, + "content": "o ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 187, + 432, + 199 + ], + "score": 1.0, + "content": ", a predictor-corrector method computes", + "type": "text" + }, + { + "bbox": [ + 432, + 186, + 475, + 199 + ], + "score": 0.92, + "content": "o ( t + \\Delta t )", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "in two", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 372, + 210 + ], + "score": 1.0, + "content": "steps. First, a prediction step approximates the next state, yielding", + "type": "text" + }, + { + "bbox": [ + 372, + 198, + 418, + 209 + ], + "score": 0.92, + "content": "o ^ { p } ( t + \\Delta t )", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 196, + 505, + 210 + ], + "score": 1.0, + "content": ". Then, the correction", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 207, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 127, + 222 + ], + "score": 1.0, + "content": "uses", + "type": "text" + }, + { + "bbox": [ + 127, + 209, + 174, + 220 + ], + "score": 0.93, + "content": "o ^ { p } ( t + \\Delta t )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 207, + 360, + 222 + ], + "score": 1.0, + "content": "to refine the initial approximation and obtain", + "type": "text" + }, + { + "bbox": [ + 361, + 209, + 403, + 221 + ], + "score": 0.91, + "content": "o ( t + \\Delta t )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 207, + 505, + 222 + ], + "score": 1.0, + "content": ". Each step can, to some", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 219, + 242, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 242, + 232 + ], + "score": 1.0, + "content": "degree, be learned independently.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 236, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "score": 1.0, + "content": "This motivates splitting the agent into two neural networks: an observation predictor (OP) network", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 227, + 261 + ], + "score": 1.0, + "content": "that infers intermediate states", + "type": "text" + }, + { + "bbox": [ + 227, + 248, + 253, + 259 + ], + "score": 0.84, + "content": "o ^ { p } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 246, + 257, + 261 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 258, + 247, + 335, + 259 + ], + "score": 0.68, + "content": "i \\in \\{ 1 , 2 , . . . n - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 246, + 506, + 261 + ], + "score": 1.0, + "content": ", planning out a trajectory, and a corrector", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 257, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 104, + 257, + 293, + 273 + ], + "score": 1.0, + "content": "network (CFE) that estimates the control force", + "type": "text" + }, + { + "bbox": [ + 293, + 258, + 367, + 271 + ], + "score": 0.91, + "content": "F ( t _ { i } | \\mathbf { \\sigma } ( \\dot { \\mathbf { u } _ { i } } ) , \\mathbf { \\sigma } _ { i + 1 } ^ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 257, + 506, + 273 + ], + "score": 1.0, + "content": "to follow that trajectory as closely", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "as possible. This splitting has the added benefit of exposing the planned trajectory, which would", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "otherwise be inaccessible. As we will demonstrate, it is crucial for the prediction stage to incorporate", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "score": 1.0, + "content": "knowledge about longer time spans. We address this by modelling the prediction as a temporally", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 302, + 428, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 428, + 315 + ], + "score": 1.0, + "content": "hierarchical process, recursively dividing the problem into smaller subproblems.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 300, + 331 + ], + "score": 1.0, + "content": "To achieve this, we let the OP not directly infer", + "type": "text" + }, + { + "bbox": [ + 300, + 318, + 380, + 331 + ], + "score": 0.92, + "content": "o ^ { p } ( t _ { i + 1 } | o ( u _ { i } ) , o ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "but instead model it to predict", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 321, + 343 + ], + "score": 1.0, + "content": "the optimal center point between two states at times", + "type": "text" + }, + { + "bbox": [ + 321, + 332, + 342, + 342 + ], + "score": 0.9, + "content": "t _ { i } , t _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 330, + 367, + 343 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 368, + 330, + 486, + 342 + ], + "score": 0.9, + "content": "i , j \\in \\{ 1 , 2 , . . . n - 1 \\} , j > i", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 330, + 505, + 343 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 339, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 107, + 341, + 201, + 353 + ], + "score": 0.9, + "content": "o ^ { p } ( ( \\bar { t _ { i } } + t _ { j } ) / 2 | o _ { i } , \\bar { o } _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 339, + 506, + 355 + ], + "score": 1.0, + "content": ". This function is much more general than predicting the state of the next", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "time step since two arbitrary states can be passed as arguments. Recursive OP evaluations can then", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 444, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 267, + 375 + ], + "score": 1.0, + "content": "partition the sequence until a prediction", + "type": "text" + }, + { + "bbox": [ + 267, + 363, + 293, + 375 + ], + "score": 0.93, + "content": "o ^ { p } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 363, + 371, + 375 + ], + "score": 1.0, + "content": "for every time step", + "type": "text" + }, + { + "bbox": [ + 371, + 364, + 380, + 374 + ], + "score": 0.84, + "content": "t _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 363, + 444, + 375 + ], + "score": 1.0, + "content": "has been made.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "score": 1.0, + "content": "This scheme naturally enables scaling to arbitrary time frames or arbitrary temporal resolutions,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "assuming that the OP can correctly anticipate the physical behavior. Since physical systems often", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "exhibit different behaviors on different time scales and the OP can be called with states separated by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "arbitrary time spans, we condition the OP on the time scale it is evaluated on by instantiating and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "training a unique version of the model for every time scale. This simplifies training and does not", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "significantly increase the model complexity as we use factors of two for the time scales, and hence", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 281, + 459 + ], + "score": 1.0, + "content": "the number of required models scales with", + "type": "text" + }, + { + "bbox": [ + 281, + 446, + 322, + 458 + ], + "score": 0.93, + "content": "{ \\mathcal { O } } ( \\log _ { 2 } n )", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 445, + 471, + 459 + ], + "score": 1.0, + "content": ". We will refer to one instance of an", + "type": "text" + }, + { + "bbox": [ + 471, + 446, + 491, + 457 + ], + "score": 0.9, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 456, + 474, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 398, + 469 + ], + "score": 1.0, + "content": "the time span between its input states, measured in the number of frames", + "type": "text" + }, + { + "bbox": [ + 399, + 456, + 470, + 469 + ], + "score": 0.92, + "content": "n = ( t _ { j } - t _ { i } ) / \\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 457, + 474, + 469 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 105, + 473, + 504, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 292, + 487 + ], + "score": 1.0, + "content": "Execution order. With the CFE and", + "type": "text" + }, + { + "bbox": [ + 293, + 474, + 312, + 485 + ], + "score": 0.88, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 471, + 506, + 487 + ], + "score": 1.0, + "content": "as building blocks, many algorithms for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 332, + 497 + ], + "score": 1.0, + "content": "solving the control problem, i.e. for computing", + "type": "text" + }, + { + "bbox": [ + 332, + 484, + 353, + 497 + ], + "score": 0.91, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 483, + 505, + 497 + ], + "score": 1.0, + "content": ", can be assembled and trained.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 277, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 277, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 277, + 507 + ], + "score": 1.0, + "content": "We compared a variety of algorithms and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 506, + 277, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 277, + 518 + ], + "score": 1.0, + "content": "found that a scheme we will refer to as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 518, + 277, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 277, + 529 + ], + "score": 1.0, + "content": "prediction refinement produces the best re-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 277, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 277, + 540 + ], + "score": 1.0, + "content": "sults. It is based on the following prin-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 539, + 278, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 278, + 551 + ], + "score": 1.0, + "content": "ciples: (i) always use the finest scale OP", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 551, + 277, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 277, + 561 + ], + "score": 1.0, + "content": "possible to make a prediction, (ii) execute", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 277, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 277, + 573 + ], + "score": 1.0, + "content": "the CFE followed by a solver step as soon", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 573, + 277, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 277, + 583 + ], + "score": 1.0, + "content": "as possible, (iii) refine predictions after the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 583, + 277, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 277, + 594 + ], + "score": 1.0, + "content": "solver has computed the next state. The al-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 594, + 277, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 277, + 606 + ], + "score": 1.0, + "content": "gorithm that realizes these goals is shown", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 605, + 277, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 248, + 617 + ], + "score": 1.0, + "content": "in Appendix B with an example for", + "type": "text" + }, + { + "bbox": [ + 248, + 605, + 273, + 615 + ], + "score": 0.89, + "content": "n = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 605, + 277, + 617 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 614, + 277, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 277, + 629 + ], + "score": 1.0, + "content": "To understand the algorithm and resulting", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 627, + 277, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 277, + 639 + ], + "score": 1.0, + "content": "execution orders, it is helpful to consider", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 639, + 204, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 204, + 650 + ], + "score": 1.0, + "content": "simpler algorithms first.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 38.5 + }, + { + "type": "image", + "bbox": [ + 284, + 500, + 505, + 624 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 284, + 500, + 505, + 624 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 284, + 500, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 284, + 500, + 505, + 624 + ], + "score": 0.939, + "type": "image", + "image_path": "7911148c7e6b1b70d655a2ba10e2f82c5d71c414546e6419403414d0e694ff15.jpg" + } + ] + } + ], + "index": 50, + "virtual_lines": [ + { + "bbox": [ + 284, + 500, + 505, + 513.7777777777778 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 284, + 513.7777777777778, + 505, + 527.5555555555557 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 284, + 527.5555555555557, + 505, + 541.3333333333335 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 284, + 541.3333333333335, + 505, + 555.1111111111113 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 284, + 555.1111111111113, + 505, + 568.8888888888891 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 284, + 568.8888888888891, + 505, + 582.666666666667 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 284, + 582.666666666667, + 505, + 596.4444444444448 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 284, + 596.4444444444448, + 505, + 610.2222222222226 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 284, + 610.2222222222226, + 505, + 624.0000000000005 + ], + "spans": [], + "index": 54 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 284, + 636, + 504, + 648 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 284, + 635, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 284, + 635, + 505, + 649 + ], + "score": 1.0, + "content": "Figure 3: Overview of the different execution schemes.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 55 + } + ], + "index": 52.5 + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 261, + 668 + ], + "score": 1.0, + "content": "The simplest combination of CFE and", + "type": "text" + }, + { + "bbox": [ + 261, + 655, + 280, + 666 + ], + "score": 0.89, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 653, + 505, + 668 + ], + "score": 1.0, + "content": "invocations that solves the full trajectory, shown in Fig-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "ure 3a, can be described as follows. Initially, all intermediate states are predicted hierarchically. The", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 253, + 690 + ], + "score": 1.0, + "content": "first prediction is the half-way point", + "type": "text" + }, + { + "bbox": [ + 253, + 677, + 320, + 689 + ], + "score": 0.92, + "content": "\\pmb { o } ^ { p } \\big ( t _ { n / 2 } | \\pmb { o } _ { 0 } , \\pmb { o } ^ { * } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 676, + 392, + 690 + ], + "score": 1.0, + "content": ", generated by the", + "type": "text" + }, + { + "bbox": [ + 393, + 677, + 412, + 688 + ], + "score": 0.9, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 676, + 506, + 690 + ], + "score": 1.0, + "content": ". Using that as input to", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 119, + 702 + ], + "score": 1.0, + "content": "an", + "type": "text" + }, + { + "bbox": [ + 119, + 688, + 146, + 700 + ], + "score": 0.92, + "content": "\\mathrm { O P } _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 687, + 265, + 702 + ], + "score": 1.0, + "content": "results in new predictions at", + "type": "text" + }, + { + "bbox": [ + 266, + 689, + 309, + 700 + ], + "score": 0.9, + "content": "t _ { n / 4 } , t _ { 3 n / 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 687, + 506, + 702 + ], + "score": 1.0, + "content": ". Continuing with this scheme, a prediction can", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 178, + 711 + ], + "score": 1.0, + "content": "be made for each", + "type": "text" + }, + { + "bbox": [ + 178, + 699, + 253, + 710 + ], + "score": 0.91, + "content": "t _ { i } , i \\in { 1 , . . . , n - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 699, + 505, + 711 + ], + "score": 1.0, + "content": ". Next, the actual trajectory is evaluated step by step. For each", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 125, + 722 + ], + "score": 1.0, + "content": "step", + "type": "text" + }, + { + "bbox": [ + 125, + 711, + 133, + 721 + ], + "score": 0.83, + "content": "t _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 710, + 285, + 722 + ], + "score": 1.0, + "content": ", the CFE computes the control force", + "type": "text" + }, + { + "bbox": [ + 285, + 710, + 309, + 722 + ], + "score": 0.92, + "content": "\\pmb { F } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 710, + 419, + 722 + ], + "score": 1.0, + "content": "conditioned on the state at", + "type": "text" + }, + { + "bbox": [ + 420, + 711, + 428, + 721 + ], + "score": 0.86, + "content": "t _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "and the prediction", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 107, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 142, + 733 + ], + "score": 0.93, + "content": "o ^ { p } ( t _ { i + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 720, + 171, + 734 + ], + "score": 1.0, + "content": ". Once", + "type": "text" + }, + { + "bbox": [ + 171, + 721, + 195, + 732 + ], + "score": 0.92, + "content": "\\pmb { F } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 720, + 448, + 734 + ], + "score": 1.0, + "content": "is known, the solver can step the simulation to the next state at", + "type": "text" + }, + { + "bbox": [ + 449, + 721, + 467, + 732 + ], + "score": 0.91, + "content": "t _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 720, + 505, + 734 + ], + "score": 1.0, + "content": ". This al-", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 59 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 107 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 126, + 173, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 124, + 174, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 174, + 140 + ], + "score": 1.0, + "content": "5 METHOD", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 153, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "In order to optimally interact with a physical system, an agent has to (i) build an internal repre-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 164, + 504, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 292, + 176 + ], + "score": 1.0, + "content": "sentation of an optimal observable trajectory", + "type": "text" + }, + { + "bbox": [ + 293, + 164, + 325, + 177 + ], + "score": 0.93, + "content": "o ( { \\boldsymbol { u } } ( t ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 165, + 504, + 176 + ], + "score": 1.0, + "content": "and (ii) learn what actions to take to move", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 177, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 188 + ], + "score": 1.0, + "content": "the system along the desired trajectory. These two steps strongly resemble the predictor-corrector", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 249, + 199 + ], + "score": 1.0, + "content": "method (Press et al., 2007). Given", + "type": "text" + }, + { + "bbox": [ + 250, + 187, + 268, + 199 + ], + "score": 0.91, + "content": "o ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 187, + 432, + 199 + ], + "score": 1.0, + "content": ", a predictor-corrector method computes", + "type": "text" + }, + { + "bbox": [ + 432, + 186, + 475, + 199 + ], + "score": 0.92, + "content": "o ( t + \\Delta t )", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "in two", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 372, + 210 + ], + "score": 1.0, + "content": "steps. First, a prediction step approximates the next state, yielding", + "type": "text" + }, + { + "bbox": [ + 372, + 198, + 418, + 209 + ], + "score": 0.92, + "content": "o ^ { p } ( t + \\Delta t )", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 196, + 505, + 210 + ], + "score": 1.0, + "content": ". Then, the correction", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 207, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 127, + 222 + ], + "score": 1.0, + "content": "uses", + "type": "text" + }, + { + "bbox": [ + 127, + 209, + 174, + 220 + ], + "score": 0.93, + "content": "o ^ { p } ( t + \\Delta t )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 207, + 360, + 222 + ], + "score": 1.0, + "content": "to refine the initial approximation and obtain", + "type": "text" + }, + { + "bbox": [ + 361, + 209, + 403, + 221 + ], + "score": 0.91, + "content": "o ( t + \\Delta t )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 207, + 505, + 222 + ], + "score": 1.0, + "content": ". Each step can, to some", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 219, + 242, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 242, + 232 + ], + "score": 1.0, + "content": "degree, be learned independently.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 154, + 505, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 236, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "score": 1.0, + "content": "This motivates splitting the agent into two neural networks: an observation predictor (OP) network", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 227, + 261 + ], + "score": 1.0, + "content": "that infers intermediate states", + "type": "text" + }, + { + "bbox": [ + 227, + 248, + 253, + 259 + ], + "score": 0.84, + "content": "o ^ { p } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 246, + 257, + 261 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 258, + 247, + 335, + 259 + ], + "score": 0.68, + "content": "i \\in \\{ 1 , 2 , . . . n - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 246, + 506, + 261 + ], + "score": 1.0, + "content": ", planning out a trajectory, and a corrector", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 257, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 104, + 257, + 293, + 273 + ], + "score": 1.0, + "content": "network (CFE) that estimates the control force", + "type": "text" + }, + { + "bbox": [ + 293, + 258, + 367, + 271 + ], + "score": 0.91, + "content": "F ( t _ { i } | \\mathbf { \\sigma } ( \\dot { \\mathbf { u } _ { i } } ) , \\mathbf { \\sigma } _ { i + 1 } ^ { p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 257, + 506, + 273 + ], + "score": 1.0, + "content": "to follow that trajectory as closely", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "as possible. This splitting has the added benefit of exposing the planned trajectory, which would", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "otherwise be inaccessible. As we will demonstrate, it is crucial for the prediction stage to incorporate", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 305 + ], + "score": 1.0, + "content": "knowledge about longer time spans. We address this by modelling the prediction as a temporally", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 302, + 428, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 428, + 315 + ], + "score": 1.0, + "content": "hierarchical process, recursively dividing the problem into smaller subproblems.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13, + "bbox_fs": [ + 104, + 235, + 506, + 315 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 300, + 331 + ], + "score": 1.0, + "content": "To achieve this, we let the OP not directly infer", + "type": "text" + }, + { + "bbox": [ + 300, + 318, + 380, + 331 + ], + "score": 0.92, + "content": "o ^ { p } ( t _ { i + 1 } | o ( u _ { i } ) , o ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "but instead model it to predict", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 321, + 343 + ], + "score": 1.0, + "content": "the optimal center point between two states at times", + "type": "text" + }, + { + "bbox": [ + 321, + 332, + 342, + 342 + ], + "score": 0.9, + "content": "t _ { i } , t _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 330, + 367, + 343 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 368, + 330, + 486, + 342 + ], + "score": 0.9, + "content": "i , j \\in \\{ 1 , 2 , . . . n - 1 \\} , j > i", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 330, + 505, + 343 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 339, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 107, + 341, + 201, + 353 + ], + "score": 0.9, + "content": "o ^ { p } ( ( \\bar { t _ { i } } + t _ { j } ) / 2 | o _ { i } , \\bar { o } _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 339, + 506, + 355 + ], + "score": 1.0, + "content": ". This function is much more general than predicting the state of the next", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "time step since two arbitrary states can be passed as arguments. Recursive OP evaluations can then", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 444, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 267, + 375 + ], + "score": 1.0, + "content": "partition the sequence until a prediction", + "type": "text" + }, + { + "bbox": [ + 267, + 363, + 293, + 375 + ], + "score": 0.93, + "content": "o ^ { p } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 363, + 371, + 375 + ], + "score": 1.0, + "content": "for every time step", + "type": "text" + }, + { + "bbox": [ + 371, + 364, + 380, + 374 + ], + "score": 0.84, + "content": "t _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 363, + 444, + 375 + ], + "score": 1.0, + "content": "has been made.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 318, + 506, + 375 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "score": 1.0, + "content": "This scheme naturally enables scaling to arbitrary time frames or arbitrary temporal resolutions,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "assuming that the OP can correctly anticipate the physical behavior. Since physical systems often", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "exhibit different behaviors on different time scales and the OP can be called with states separated by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "arbitrary time spans, we condition the OP on the time scale it is evaluated on by instantiating and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "training a unique version of the model for every time scale. This simplifies training and does not", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "significantly increase the model complexity as we use factors of two for the time scales, and hence", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 445, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 281, + 459 + ], + "score": 1.0, + "content": "the number of required models scales with", + "type": "text" + }, + { + "bbox": [ + 281, + 446, + 322, + 458 + ], + "score": 0.93, + "content": "{ \\mathcal { O } } ( \\log _ { 2 } n )", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 445, + 471, + 459 + ], + "score": 1.0, + "content": ". We will refer to one instance of an", + "type": "text" + }, + { + "bbox": [ + 471, + 446, + 491, + 457 + ], + "score": 0.9, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 445, + 505, + 459 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 456, + 474, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 398, + 469 + ], + "score": 1.0, + "content": "the time span between its input states, measured in the number of frames", + "type": "text" + }, + { + "bbox": [ + 399, + 456, + 470, + 469 + ], + "score": 0.92, + "content": "n = ( t _ { j } - t _ { i } ) / \\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 457, + 474, + 469 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 379, + 506, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 473, + 504, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 292, + 487 + ], + "score": 1.0, + "content": "Execution order. With the CFE and", + "type": "text" + }, + { + "bbox": [ + 293, + 474, + 312, + 485 + ], + "score": 0.88, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 471, + 506, + 487 + ], + "score": 1.0, + "content": "as building blocks, many algorithms for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 332, + 497 + ], + "score": 1.0, + "content": "solving the control problem, i.e. for computing", + "type": "text" + }, + { + "bbox": [ + 332, + 484, + 353, + 497 + ], + "score": 0.91, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 483, + 505, + 497 + ], + "score": 1.0, + "content": ", can be assembled and trained.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 471, + 506, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 277, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 277, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 277, + 507 + ], + "score": 1.0, + "content": "We compared a variety of algorithms and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 506, + 277, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 277, + 518 + ], + "score": 1.0, + "content": "found that a scheme we will refer to as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 518, + 277, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 277, + 529 + ], + "score": 1.0, + "content": "prediction refinement produces the best re-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 277, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 277, + 540 + ], + "score": 1.0, + "content": "sults. It is based on the following prin-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 539, + 278, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 278, + 551 + ], + "score": 1.0, + "content": "ciples: (i) always use the finest scale OP", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 551, + 277, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 277, + 561 + ], + "score": 1.0, + "content": "possible to make a prediction, (ii) execute", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 277, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 277, + 573 + ], + "score": 1.0, + "content": "the CFE followed by a solver step as soon", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 573, + 277, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 277, + 583 + ], + "score": 1.0, + "content": "as possible, (iii) refine predictions after the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 583, + 277, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 277, + 594 + ], + "score": 1.0, + "content": "solver has computed the next state. The al-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 594, + 277, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 277, + 606 + ], + "score": 1.0, + "content": "gorithm that realizes these goals is shown", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 605, + 277, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 248, + 617 + ], + "score": 1.0, + "content": "in Appendix B with an example for", + "type": "text" + }, + { + "bbox": [ + 248, + 605, + 273, + 615 + ], + "score": 0.89, + "content": "n = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 605, + 277, + 617 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 614, + 277, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 277, + 629 + ], + "score": 1.0, + "content": "To understand the algorithm and resulting", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 627, + 277, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 277, + 639 + ], + "score": 1.0, + "content": "execution orders, it is helpful to consider", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 639, + 204, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 204, + 650 + ], + "score": 1.0, + "content": "simpler algorithms first.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 495, + 278, + 650 + ] + }, + { + "type": "image", + "bbox": [ + 284, + 500, + 505, + 624 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 284, + 500, + 505, + 624 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 284, + 500, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 284, + 500, + 505, + 624 + ], + "score": 0.939, + "type": "image", + "image_path": "7911148c7e6b1b70d655a2ba10e2f82c5d71c414546e6419403414d0e694ff15.jpg" + } + ] + } + ], + "index": 50, + "virtual_lines": [ + { + "bbox": [ + 284, + 500, + 505, + 513.7777777777778 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 284, + 513.7777777777778, + 505, + 527.5555555555557 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 284, + 527.5555555555557, + 505, + 541.3333333333335 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 284, + 541.3333333333335, + 505, + 555.1111111111113 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 284, + 555.1111111111113, + 505, + 568.8888888888891 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 284, + 568.8888888888891, + 505, + 582.666666666667 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 284, + 582.666666666667, + 505, + 596.4444444444448 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 284, + 596.4444444444448, + 505, + 610.2222222222226 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 284, + 610.2222222222226, + 505, + 624.0000000000005 + ], + "spans": [], + "index": 54 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 284, + 636, + 504, + 648 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 284, + 635, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 284, + 635, + 505, + 649 + ], + "score": 1.0, + "content": "Figure 3: Overview of the different execution schemes.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 55 + } + ], + "index": 52.5 + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 261, + 668 + ], + "score": 1.0, + "content": "The simplest combination of CFE and", + "type": "text" + }, + { + "bbox": [ + 261, + 655, + 280, + 666 + ], + "score": 0.89, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 653, + 505, + 668 + ], + "score": 1.0, + "content": "invocations that solves the full trajectory, shown in Fig-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "ure 3a, can be described as follows. Initially, all intermediate states are predicted hierarchically. The", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 253, + 690 + ], + "score": 1.0, + "content": "first prediction is the half-way point", + "type": "text" + }, + { + "bbox": [ + 253, + 677, + 320, + 689 + ], + "score": 0.92, + "content": "\\pmb { o } ^ { p } \\big ( t _ { n / 2 } | \\pmb { o } _ { 0 } , \\pmb { o } ^ { * } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 676, + 392, + 690 + ], + "score": 1.0, + "content": ", generated by the", + "type": "text" + }, + { + "bbox": [ + 393, + 677, + 412, + 688 + ], + "score": 0.9, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 676, + 506, + 690 + ], + "score": 1.0, + "content": ". Using that as input to", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 119, + 702 + ], + "score": 1.0, + "content": "an", + "type": "text" + }, + { + "bbox": [ + 119, + 688, + 146, + 700 + ], + "score": 0.92, + "content": "\\mathrm { O P } _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 687, + 265, + 702 + ], + "score": 1.0, + "content": "results in new predictions at", + "type": "text" + }, + { + "bbox": [ + 266, + 689, + 309, + 700 + ], + "score": 0.9, + "content": "t _ { n / 4 } , t _ { 3 n / 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 687, + 506, + 702 + ], + "score": 1.0, + "content": ". Continuing with this scheme, a prediction can", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 178, + 711 + ], + "score": 1.0, + "content": "be made for each", + "type": "text" + }, + { + "bbox": [ + 178, + 699, + 253, + 710 + ], + "score": 0.91, + "content": "t _ { i } , i \\in { 1 , . . . , n - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 699, + 505, + 711 + ], + "score": 1.0, + "content": ". Next, the actual trajectory is evaluated step by step. For each", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 125, + 722 + ], + "score": 1.0, + "content": "step", + "type": "text" + }, + { + "bbox": [ + 125, + 711, + 133, + 721 + ], + "score": 0.83, + "content": "t _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 710, + 285, + 722 + ], + "score": 1.0, + "content": ", the CFE computes the control force", + "type": "text" + }, + { + "bbox": [ + 285, + 710, + 309, + 722 + ], + "score": 0.92, + "content": "\\pmb { F } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 710, + 419, + 722 + ], + "score": 1.0, + "content": "conditioned on the state at", + "type": "text" + }, + { + "bbox": [ + 420, + 711, + 428, + 721 + ], + "score": 0.86, + "content": "t _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "and the prediction", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 107, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 142, + 733 + ], + "score": 0.93, + "content": "o ^ { p } ( t _ { i + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 720, + 171, + 734 + ], + "score": 1.0, + "content": ". Once", + "type": "text" + }, + { + "bbox": [ + 171, + 721, + 195, + 732 + ], + "score": 0.92, + "content": "\\pmb { F } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 720, + 448, + 734 + ], + "score": 1.0, + "content": "is known, the solver can step the simulation to the next state at", + "type": "text" + }, + { + "bbox": [ + 449, + 721, + 467, + 732 + ], + "score": 0.91, + "content": "t _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 720, + 505, + 734 + ], + "score": 1.0, + "content": ". This al-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 238, + 95 + ], + "score": 1.0, + "content": "gorithm finds a trajectory in time", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 238, + 82, + 261, + 95 + ], + "score": 0.92, + "content": "{ \\mathcal { O } } ( n )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 261, + 82, + 284, + 95 + ], + "score": 1.0, + "content": "since", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 284, + 84, + 292, + 93 + ], + "score": 0.69, + "content": "n", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 293, + 82, + 349, + 95 + ], + "score": 1.0, + "content": "CFE calls and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 350, + 83, + 371, + 93 + ], + "score": 0.82, + "content": "n { - } 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 372, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "OP calls are required in total (see", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Appendix B). However, there are inherent problems with this algorithm. The physical constraints", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "of the PDE and potential approximation errors of the CFE can result in observations that are only", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "matched partially. This can result in the reconstructed trajectory exhibiting undesirable oscillations,", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "often visible as jittering. When subsequent predictions do not line up perfectly, large forces may be", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 500, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 500, + 150 + ], + "score": 1.0, + "content": "applied by the CFE or the reconstructed trajectory might stop following the predictions altogether.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 59, + "bbox_fs": [ + 105, + 653, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 238, + 95 + ], + "score": 1.0, + "content": "gorithm finds a trajectory in time", + "type": "text" + }, + { + "bbox": [ + 238, + 82, + 261, + 95 + ], + "score": 0.92, + "content": "{ \\mathcal { O } } ( n )", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 82, + 284, + 95 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 284, + 84, + 292, + 93 + ], + "score": 0.69, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 82, + 349, + 95 + ], + "score": 1.0, + "content": "CFE calls and", + "type": "text" + }, + { + "bbox": [ + 350, + 83, + 371, + 93 + ], + "score": 0.82, + "content": "n { - } 1", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "OP calls are required in total (see", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Appendix B). However, there are inherent problems with this algorithm. The physical constraints", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "of the PDE and potential approximation errors of the CFE can result in observations that are only", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "matched partially. This can result in the reconstructed trajectory exhibiting undesirable oscillations,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "often visible as jittering. When subsequent predictions do not line up perfectly, large forces may be", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 500, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 500, + 150 + ], + "score": 1.0, + "content": "applied by the CFE or the reconstructed trajectory might stop following the predictions altogether.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "This problem can be alleviated by changing the execution order of the two-stage algorithm described", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 504, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 504, + 177 + ], + "score": 1.0, + "content": "above. The resulting algorithm is shown in Figure 3b and will be referred to as staggered execution.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "In this setup, the simulation is advanced as soon as a prediction for the next observable state exists", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 188, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 310, + 199 + ], + "score": 1.0, + "content": "and OPs are only executed when their state at time", + "type": "text" + }, + { + "bbox": [ + 311, + 188, + 319, + 198 + ], + "score": 0.85, + "content": "t _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 188, + 504, + 199 + ], + "score": 1.0, + "content": "is available. This staggered execution scheme", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 197, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 506, + 211 + ], + "score": 1.0, + "content": "allows future predictions to take deviations from the predicted trajectory into account, preventing a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 385, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 243, + 222 + ], + "score": 1.0, + "content": "divergence of the actual evolution", + "type": "text" + }, + { + "bbox": [ + 244, + 209, + 276, + 221 + ], + "score": 0.93, + "content": "\\mathbf { \\nabla } _ { o ( u ( t ) ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 208, + 357, + 222 + ], + "score": 1.0, + "content": "from the prediction", + "type": "text" + }, + { + "bbox": [ + 357, + 209, + 380, + 221 + ], + "score": 0.92, + "content": "o ^ { p } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 208, + 385, + 222 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "While the staggered execution allows most predictions to correct for deviations from the predicted", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 149, + 249 + ], + "score": 1.0, + "content": "trajectory", + "type": "text" + }, + { + "bbox": [ + 149, + 237, + 160, + 247 + ], + "score": 0.84, + "content": "o ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 237, + 505, + 249 + ], + "score": 1.0, + "content": ", this scheme leaves several predictions unmodified. Most notably, the prediction", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 142, + 261 + ], + "score": 0.92, + "content": "o ^ { \\bar { p ( } _ { t _ { n / 2 } } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 247, + 505, + 261 + ], + "score": 1.0, + "content": ", which is inferred from just the initial state and the desired target, remains unchanged.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "This prediction must therefore be able to guide the reconstruction in the right direction without", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 268, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 104, + 268, + 348, + 284 + ], + "score": 1.0, + "content": "knowing about deviations in the system that occurred up to", + "type": "text" + }, + { + "bbox": [ + 349, + 271, + 377, + 282 + ], + "score": 0.92, + "content": "t _ { n / 2 - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 268, + 506, + 284 + ], + "score": 1.0, + "content": ". As a practical consequence, a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "score": 1.0, + "content": "network trained with this scheme typically learns to average over the deviations, resulting in blurred", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 292, + 236, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 236, + 304 + ], + "score": 1.0, + "content": "predictions (see Appendix D.2).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 317, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "Algorithm 1: Recursive algorithm computing the prediction refinement. The algorithm is called via", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 328, + 406, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 114, + 343 + ], + "score": 1.0, + "content": "R", + "type": "text" + }, + { + "bbox": [ + 114, + 329, + 223, + 341 + ], + "score": 0.31, + "content": "\\mathrm { e c o n s t r u c t } [ o _ { 0 } , o _ { * } , a b s e n t ]", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 328, + 367, + 343 + ], + "score": 1.0, + "content": "to reconstruct a full trajectory from", + "type": "text" + }, + { + "bbox": [ + 368, + 331, + 379, + 341 + ], + "score": 0.82, + "content": "o _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 328, + 390, + 343 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 390, + 331, + 402, + 340 + ], + "score": 0.83, + "content": "^ { o _ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 328, + 406, + 343 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "table", + "bbox": [ + 105, + 343, + 496, + 562 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 105, + 343, + 496, + 562 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 343, + 496, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 496, + 562 + ], + "score": 0.269, + "html": "
function Reconstruct[o(uo),On, O2n]; Input : Initial observation o(uo),observation On,optional observation O2n Output: Observation of the reconstructed state o(un)
if n=1 then F ← CFE[o(uo),01]
U1 ← Solver[uo,F]
return o(u1)
else
On/2 ← OP[o(uo),On]
o(un/2) ← Reconstruct[o(uo),On/2,On]
if O2n present then
03n/2 ← OP[0n,02n]
On ← OP[o(un/2),03n/2] else
1 03n/2 ← absent
end
o(un) ← Reconstruct[o(un/2),On,03n/2] return o(un)
", + "type": "table", + "image_path": "cbea6aa149ddeac0694e3da47a8c3253c757c7875cc6551629168fcf27af86c3.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 105, + 343, + 496, + 416.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 105, + 416.0, + 496, + 489.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 105, + 489.0, + 496, + 562.0 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "The prediction refinement scheme, listed in Algorithm 1 and illustrated in Figure 3c, solves this", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "problem by re-evaluating existing predictions whenever the simulation progesses in time. Not all", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "predictions need to be updated, though, and an update to a prediction at a finer time scale can", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "depend on a sequence of other predictions. The prediction refinement algorithm that achieves this in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "an optimal form is listed in Appendix B. While the resulting execution order is difficult to follow for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 631, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 237, + 646 + ], + "score": 1.0, + "content": "longer sequences with more than", + "type": "text" + }, + { + "bbox": [ + 238, + 632, + 264, + 642 + ], + "score": 0.9, + "content": "n = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 631, + 505, + 646 + ], + "score": 1.0, + "content": "frames, we give an overview of the algorithm by considering", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 196, + 656 + ], + "score": 1.0, + "content": "the prediction for time", + "type": "text" + }, + { + "bbox": [ + 196, + 644, + 215, + 655 + ], + "score": 0.9, + "content": "t _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 643, + 369, + 656 + ], + "score": 1.0, + "content": ". After the first center-frame prediction", + "type": "text" + }, + { + "bbox": [ + 370, + 643, + 406, + 655 + ], + "score": 0.92, + "content": "o ^ { p } \\big ( t _ { n / 2 } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 643, + 431, + 656 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 432, + 644, + 439, + 653 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "-frame sequence", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 154, + 667 + ], + "score": 1.0, + "content": "is made by", + "type": "text" + }, + { + "bbox": [ + 154, + 654, + 173, + 665 + ], + "score": 0.89, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 654, + 506, + 667 + ], + "score": 1.0, + "content": ", the prediction refinement algorithm calls itself recursively until all frames up to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 664, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 133, + 677 + ], + "score": 1.0, + "content": "frame", + "type": "text" + }, + { + "bbox": [ + 133, + 665, + 151, + 676 + ], + "score": 0.87, + "content": "n / 4", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 664, + 506, + 677 + ], + "score": 1.0, + "content": "are reconstructed from the CFE and the solver. The center prediction is then updated", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 674, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 104, + 674, + 130, + 691 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 131, + 677, + 158, + 689 + ], + "score": 0.91, + "content": "\\mathrm { O P } _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 674, + 476, + 691 + ], + "score": 1.0, + "content": "for the next smaller time scale compared to the previous prediction. The call of", + "type": "text" + }, + { + "bbox": [ + 477, + 676, + 504, + 689 + ], + "score": 0.91, + "content": "\\mathrm { O P } _ { n / 2 }", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 173, + 701 + ], + "score": 1.0, + "content": "also depends on", + "type": "text" + }, + { + "bbox": [ + 173, + 687, + 213, + 700 + ], + "score": 0.93, + "content": "o ^ { p } ( t _ { 3 n / 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 687, + 326, + 701 + ], + "score": 1.0, + "content": ", which was predicted using", + "type": "text" + }, + { + "bbox": [ + 327, + 687, + 354, + 701 + ], + "score": 0.9, + "content": "\\mathrm { O P } _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 687, + 506, + 701 + ], + "score": 1.0, + "content": ". After half of the remaining distance", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 369, + 713 + ], + "score": 1.0, + "content": "to the center is reconstructed by the solver, the center prediction at", + "type": "text" + }, + { + "bbox": [ + 369, + 700, + 387, + 712 + ], + "score": 0.89, + "content": "t _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 698, + 505, + 713 + ], + "score": 1.0, + "content": "is updated again, this time by", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 120, + 723 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 710, + 148, + 722 + ], + "score": 0.91, + "content": "\\mathrm { O P } _ { n / 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 709, + 505, + 723 + ], + "score": 1.0, + "content": ", including all prediction dependencies. Hence, the center prediction is continually refined", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "every time the temporal distance between the latest reconstruction and the prediction halves, until", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 30.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [], + "index": 2.5, + "bbox_fs": [ + 105, + 82, + 506, + 150 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "This problem can be alleviated by changing the execution order of the two-stage algorithm described", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 504, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 504, + 177 + ], + "score": 1.0, + "content": "above. The resulting algorithm is shown in Figure 3b and will be referred to as staggered execution.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "In this setup, the simulation is advanced as soon as a prediction for the next observable state exists", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 188, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 310, + 199 + ], + "score": 1.0, + "content": "and OPs are only executed when their state at time", + "type": "text" + }, + { + "bbox": [ + 311, + 188, + 319, + 198 + ], + "score": 0.85, + "content": "t _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 188, + 504, + 199 + ], + "score": 1.0, + "content": "is available. This staggered execution scheme", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 197, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 506, + 211 + ], + "score": 1.0, + "content": "allows future predictions to take deviations from the predicted trajectory into account, preventing a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 385, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 243, + 222 + ], + "score": 1.0, + "content": "divergence of the actual evolution", + "type": "text" + }, + { + "bbox": [ + 244, + 209, + 276, + 221 + ], + "score": 0.93, + "content": "\\mathbf { \\nabla } _ { o ( u ( t ) ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 208, + 357, + 222 + ], + "score": 1.0, + "content": "from the prediction", + "type": "text" + }, + { + "bbox": [ + 357, + 209, + 380, + 221 + ], + "score": 0.92, + "content": "o ^ { p } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 208, + 385, + 222 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 154, + 506, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 226, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "While the staggered execution allows most predictions to correct for deviations from the predicted", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 149, + 249 + ], + "score": 1.0, + "content": "trajectory", + "type": "text" + }, + { + "bbox": [ + 149, + 237, + 160, + 247 + ], + "score": 0.84, + "content": "o ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 237, + 505, + 249 + ], + "score": 1.0, + "content": ", this scheme leaves several predictions unmodified. Most notably, the prediction", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 142, + 261 + ], + "score": 0.92, + "content": "o ^ { \\bar { p ( } _ { t _ { n / 2 } } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 247, + 505, + 261 + ], + "score": 1.0, + "content": ", which is inferred from just the initial state and the desired target, remains unchanged.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 271 + ], + "score": 1.0, + "content": "This prediction must therefore be able to guide the reconstruction in the right direction without", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 268, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 104, + 268, + 348, + 284 + ], + "score": 1.0, + "content": "knowing about deviations in the system that occurred up to", + "type": "text" + }, + { + "bbox": [ + 349, + 271, + 377, + 282 + ], + "score": 0.92, + "content": "t _ { n / 2 - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 268, + 506, + 284 + ], + "score": 1.0, + "content": ". As a practical consequence, a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "score": 1.0, + "content": "network trained with this scheme typically learns to average over the deviations, resulting in blurred", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 292, + 236, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 236, + 304 + ], + "score": 1.0, + "content": "predictions (see Appendix D.2).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 226, + 506, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 317, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "Algorithm 1: Recursive algorithm computing the prediction refinement. The algorithm is called via", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 328, + 406, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 114, + 343 + ], + "score": 1.0, + "content": "R", + "type": "text" + }, + { + "bbox": [ + 114, + 329, + 223, + 341 + ], + "score": 0.31, + "content": "\\mathrm { e c o n s t r u c t } [ o _ { 0 } , o _ { * } , a b s e n t ]", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 328, + 367, + 343 + ], + "score": 1.0, + "content": "to reconstruct a full trajectory from", + "type": "text" + }, + { + "bbox": [ + 368, + 331, + 379, + 341 + ], + "score": 0.82, + "content": "o _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 328, + 390, + 343 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 390, + 331, + 402, + 340 + ], + "score": 0.83, + "content": "^ { o _ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 328, + 406, + 343 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 317, + 505, + 343 + ] + }, + { + "type": "table", + "bbox": [ + 105, + 343, + 496, + 562 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 105, + 343, + 496, + 562 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 343, + 496, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 496, + 562 + ], + "score": 0.269, + "html": "
function Reconstruct[o(uo),On, O2n]; Input : Initial observation o(uo),observation On,optional observation O2n Output: Observation of the reconstructed state o(un)
if n=1 then F ← CFE[o(uo),01]
U1 ← Solver[uo,F]
return o(u1)
else
On/2 ← OP[o(uo),On]
o(un/2) ← Reconstruct[o(uo),On/2,On]
if O2n present then
03n/2 ← OP[0n,02n]
On ← OP[o(un/2),03n/2] else
1 03n/2 ← absent
end
o(un) ← Reconstruct[o(un/2),On,03n/2] return o(un)
", + "type": "table", + "image_path": "cbea6aa149ddeac0694e3da47a8c3253c757c7875cc6551629168fcf27af86c3.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 105, + 343, + 496, + 416.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 105, + 416.0, + 496, + 489.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 105, + 489.0, + 496, + 562.0 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "The prediction refinement scheme, listed in Algorithm 1 and illustrated in Figure 3c, solves this", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "problem by re-evaluating existing predictions whenever the simulation progesses in time. Not all", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "predictions need to be updated, though, and an update to a prediction at a finer time scale can", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "depend on a sequence of other predictions. The prediction refinement algorithm that achieves this in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "an optimal form is listed in Appendix B. While the resulting execution order is difficult to follow for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 631, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 237, + 646 + ], + "score": 1.0, + "content": "longer sequences with more than", + "type": "text" + }, + { + "bbox": [ + 238, + 632, + 264, + 642 + ], + "score": 0.9, + "content": "n = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 631, + 505, + 646 + ], + "score": 1.0, + "content": "frames, we give an overview of the algorithm by considering", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 196, + 656 + ], + "score": 1.0, + "content": "the prediction for time", + "type": "text" + }, + { + "bbox": [ + 196, + 644, + 215, + 655 + ], + "score": 0.9, + "content": "t _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 643, + 369, + 656 + ], + "score": 1.0, + "content": ". After the first center-frame prediction", + "type": "text" + }, + { + "bbox": [ + 370, + 643, + 406, + 655 + ], + "score": 0.92, + "content": "o ^ { p } \\big ( t _ { n / 2 } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 643, + 431, + 656 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 432, + 644, + 439, + 653 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "-frame sequence", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 154, + 667 + ], + "score": 1.0, + "content": "is made by", + "type": "text" + }, + { + "bbox": [ + 154, + 654, + 173, + 665 + ], + "score": 0.89, + "content": "\\mathrm { O P } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 654, + 506, + 667 + ], + "score": 1.0, + "content": ", the prediction refinement algorithm calls itself recursively until all frames up to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 664, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 133, + 677 + ], + "score": 1.0, + "content": "frame", + "type": "text" + }, + { + "bbox": [ + 133, + 665, + 151, + 676 + ], + "score": 0.87, + "content": "n / 4", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 664, + 506, + 677 + ], + "score": 1.0, + "content": "are reconstructed from the CFE and the solver. The center prediction is then updated", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 674, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 104, + 674, + 130, + 691 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 131, + 677, + 158, + 689 + ], + "score": 0.91, + "content": "\\mathrm { O P } _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 674, + 476, + 691 + ], + "score": 1.0, + "content": "for the next smaller time scale compared to the previous prediction. The call of", + "type": "text" + }, + { + "bbox": [ + 477, + 676, + 504, + 689 + ], + "score": 0.91, + "content": "\\mathrm { O P } _ { n / 2 }", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 173, + 701 + ], + "score": 1.0, + "content": "also depends on", + "type": "text" + }, + { + "bbox": [ + 173, + 687, + 213, + 700 + ], + "score": 0.93, + "content": "o ^ { p } ( t _ { 3 n / 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 687, + 326, + 701 + ], + "score": 1.0, + "content": ", which was predicted using", + "type": "text" + }, + { + "bbox": [ + 327, + 687, + 354, + 701 + ], + "score": 0.9, + "content": "\\mathrm { O P } _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 687, + 506, + 701 + ], + "score": 1.0, + "content": ". After half of the remaining distance", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 369, + 713 + ], + "score": 1.0, + "content": "to the center is reconstructed by the solver, the center prediction at", + "type": "text" + }, + { + "bbox": [ + 369, + 700, + 387, + 712 + ], + "score": 0.89, + "content": "t _ { n / 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 698, + 505, + 713 + ], + "score": 1.0, + "content": "is updated again, this time by", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 120, + 723 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 710, + 148, + 722 + ], + "score": 0.91, + "content": "\\mathrm { O P } _ { n / 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 709, + 505, + 723 + ], + "score": 1.0, + "content": ", including all prediction dependencies. Hence, the center prediction is continually refined", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "every time the temporal distance between the latest reconstruction and the prediction halves, until", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 384, + 291 + ], + "score": 1.0, + "content": "the reconstruction reaches that frame. This way, all final predictions", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 385, + 279, + 411, + 291 + ], + "score": 0.93, + "content": "o ^ { p } ( t _ { i } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 411, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "are conditioned on the", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 290, + 497, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 249, + 303 + ], + "score": 1.0, + "content": "reconstruction of the previous state", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 249, + 290, + 281, + 302 + ], + "score": 0.93, + "content": "{ \\pmb u } ( t _ { i - 1 } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 282, + 290, + 497, + 303 + ], + "score": 1.0, + "content": "and can therefore account for all previous deviations.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 30.5, + "bbox_fs": [ + 104, + 577, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 80, + 504, + 205 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 80, + 504, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 80, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 504, + 205 + ], + "score": 0.967, + "type": "image", + "image_path": "961b474c3eb218cd35652159d49cc658bf68382844f8043e43a50b77366a6a4d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 80, + 504, + 121.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 121.66666666666666, + 504, + 163.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 163.33333333333331, + 504, + 204.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 223, + 505, + 268 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "Figure 4: Trajectories for an example control task using Burger’s equation. Initial and target states", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 234, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 247 + ], + "score": 1.0, + "content": "are plotted with thick dashed lines in red and blue, respectively. Inferred states are shown as solid", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "lines. (a) Natural evolution. (b) Reconstruction using a CFE chain. (c,d) Reconstructions using our", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 257, + 502, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 502, + 270 + ], + "score": 1.0, + "content": "hierarchical predictor-corrector scheme. (e) Ground-truth trajectory generated with constant force.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 504, + 302 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 384, + 291 + ], + "score": 1.0, + "content": "the reconstruction reaches that frame. This way, all final predictions", + "type": "text" + }, + { + "bbox": [ + 385, + 279, + 411, + 291 + ], + "score": 0.93, + "content": "o ^ { p } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "are conditioned on the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 290, + 497, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 249, + 303 + ], + "score": 1.0, + "content": "reconstruction of the previous state", + "type": "text" + }, + { + "bbox": [ + 249, + 290, + 281, + 302 + ], + "score": 0.93, + "content": "{ \\pmb u } ( t _ { i - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 290, + 497, + 303 + ], + "score": 1.0, + "content": "and can therefore account for all previous deviations.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 106, + 307, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 319 + ], + "score": 1.0, + "content": "The prediction refinement scheme requires the same number of force inferences but an increased", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 318, + 504, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 424, + 330 + ], + "score": 1.0, + "content": "number of OP evaluations compared to the simpler algorithms. With a total of", + "type": "text" + }, + { + "bbox": [ + 424, + 318, + 504, + 330 + ], + "score": 0.94, + "content": "3 n - 2 \\log _ { 2 } ( n ) - 3", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 365, + 342 + ], + "score": 1.0, + "content": "OP evaluations (see Appendix B), it is of the same complexity,", + "type": "text" + }, + { + "bbox": [ + 365, + 329, + 388, + 341 + ], + "score": 0.91, + "content": "{ \\mathcal { O } } ( n )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 329, + 505, + 342 + ], + "score": 1.0, + "content": ". In practice, this refinement", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "scheme incurs only a small overhead in terms of computation, which is outweighed by the significant", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 351, + 297, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 297, + 363 + ], + "score": 1.0, + "content": "gains in quality of the learned control function.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 378, + 172, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 174, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 174, + 393 + ], + "score": 1.0, + "content": "6 RESULTS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "We evaluate the capabilities of our method to learn to control physical PDEs in three different test", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "environments of increasing complexity. We first target a simple but nonlinear 1D equation, for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "which we present an ablation study to quantify accuracy. We then study two-dimensional problems:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "score": 1.0, + "content": "an incompressible fluid and a fluid with complex boundaries and indirect control. Full details are", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "given in Appendix D. Supplemental material containing additional sequences for all of the tests can", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 457, + 383, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 383, + 470 + ], + "score": 1.0, + "content": "be downloaded from https://ge.in.tum.de/publications/2020-iclr-holl.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "Burger’s equation. Burger’s equation is a nonlinear PDE that describes the time evolution of a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 485, + 367, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 155, + 498 + ], + "score": 1.0, + "content": "single field,", + "type": "text" + }, + { + "bbox": [ + 156, + 488, + 163, + 496 + ], + "score": 0.72, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 485, + 367, + 498 + ], + "score": 1.0, + "content": "(LeVeque, 1992). Using Eq. 1, it can be written as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 501, + 386, + 529 + ], + "lines": [ + { + "bbox": [ + 225, + 501, + 386, + 529 + ], + "spans": [ + { + "bbox": [ + 225, + 501, + 386, + 529 + ], + "score": 0.95, + "content": "\\mathcal { P } \\left( u , \\frac { \\partial u } { \\partial x } , \\frac { \\partial ^ { 2 } u } { \\partial x ^ { 2 } } \\right) = - u \\cdot \\frac { \\partial u } { \\partial x } + \\nu \\frac { \\partial ^ { 2 } u } { \\partial x ^ { 2 } } .", + "type": "interline_equation", + "image_path": "add3bab97283c6c8d71e236a730c91827c8748af50f0c75ab23157c9530ec379.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 225, + 501, + 386, + 515.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 225, + 515.0, + 386, + 529.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "Examples of the unperturbed evolution are shown in Figure 4a. We let the whole state be observable", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 543, + 456, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 191, + 558 + ], + "score": 1.0, + "content": "and controllable, i.e.", + "type": "text" + }, + { + "bbox": [ + 191, + 544, + 238, + 556 + ], + "score": 0.93, + "content": "o ( t ) = u ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 543, + 318, + 558 + ], + "score": 1.0, + "content": ", which implies that", + "type": "text" + }, + { + "bbox": [ + 318, + 545, + 330, + 555 + ], + "score": 0.87, + "content": "\\boldsymbol { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 543, + 456, + 558 + ], + "score": 1.0, + "content": "can always be reached exactly.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "The results of our ablation study with this equation are shown in Table 1. The table compares the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "resulting forces applied by differently trained models when reconstructing a ground-truth sequence", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 582, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 251, + 597 + ], + "score": 1.0, + "content": "(Figure 4e). The variant denoted by", + "type": "text" + }, + { + "bbox": [ + 252, + 583, + 272, + 594 + ], + "score": 0.27, + "content": "C F E", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 582, + 505, + 597 + ], + "score": 1.0, + "content": "chain uses a neural network to infer the force without any", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 592, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 505, + 609 + ], + "score": 1.0, + "content": "intermediate predictions. With a supervised loss, this method learns to approximate a single step", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "score": 1.0, + "content": "well. However, for longer sequences, results quickly deviate from an ideal trajectory and diverge", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "score": 1.0, + "content": "because the network never learned to account for errors made in previous steps (Figure 4b). Training", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 626, + 504, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 504, + 640 + ], + "score": 1.0, + "content": "the network with the objective loss (Eq. 4) using the differentiable solver greatly increases the quality", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 638, + 504, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 312, + 651 + ], + "score": 1.0, + "content": "of the reconstructions. On average, it applies only", + "type": "text" + }, + { + "bbox": [ + 312, + 638, + 332, + 649 + ], + "score": 0.88, + "content": "3 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 638, + 504, + 651 + ], + "score": 1.0, + "content": "of the force used by the supervised model", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 648, + 361, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 361, + 661 + ], + "score": 1.0, + "content": "as it learns to correct the temporal evolution of the PDE model.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "Next, we evaluate variants of our predictor-corrector approach, which hierarchically predicts in-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 321, + 690 + ], + "score": 1.0, + "content": "termediate states. Here, the CFE is implemented as", + "type": "text" + }, + { + "bbox": [ + 321, + 677, + 453, + 689 + ], + "score": 0.91, + "content": "F ( t _ { i } ) = ( o ^ { p } ( t _ { i + 1 } ) - u ( t _ { i } ) ) / \\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". Unlike the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "simple CFE chain above, training with the supervised loss and staggered execution produces stable", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "(albeit jittering) trajectories that successfully converge to the target state (Figure 4c). Surprisingly,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "this supervised method reaches almost the same accuracy as the differentiable CFE, despite not hav-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "score": 1.0, + "content": "ing access to physics-based gradients. However, employing the differentiable physics loss greatly", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 80, + 504, + 205 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 80, + 504, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 80, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 504, + 205 + ], + "score": 0.967, + "type": "image", + "image_path": "961b474c3eb218cd35652159d49cc658bf68382844f8043e43a50b77366a6a4d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 80, + 504, + 121.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 121.66666666666666, + 504, + 163.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 163.33333333333331, + 504, + 204.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 223, + 505, + 268 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "Figure 4: Trajectories for an example control task using Burger’s equation. Initial and target states", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 234, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 247 + ], + "score": 1.0, + "content": "are plotted with thick dashed lines in red and blue, respectively. Inferred states are shown as solid", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "lines. (a) Natural evolution. (b) Reconstruction using a CFE chain. (c,d) Reconstructions using our", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 257, + 502, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 502, + 270 + ], + "score": 1.0, + "content": "hierarchical predictor-corrector scheme. (e) Ground-truth trajectory generated with constant force.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 504, + 302 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 279, + 505, + 303 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 307, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 106, + 307, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 319 + ], + "score": 1.0, + "content": "The prediction refinement scheme requires the same number of force inferences but an increased", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 318, + 504, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 424, + 330 + ], + "score": 1.0, + "content": "number of OP evaluations compared to the simpler algorithms. With a total of", + "type": "text" + }, + { + "bbox": [ + 424, + 318, + 504, + 330 + ], + "score": 0.94, + "content": "3 n - 2 \\log _ { 2 } ( n ) - 3", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 365, + 342 + ], + "score": 1.0, + "content": "OP evaluations (see Appendix B), it is of the same complexity,", + "type": "text" + }, + { + "bbox": [ + 365, + 329, + 388, + 341 + ], + "score": 0.91, + "content": "{ \\mathcal { O } } ( n )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 329, + 505, + 342 + ], + "score": 1.0, + "content": ". In practice, this refinement", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "scheme incurs only a small overhead in terms of computation, which is outweighed by the significant", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 351, + 297, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 297, + 363 + ], + "score": 1.0, + "content": "gains in quality of the learned control function.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 307, + 505, + 363 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 378, + 172, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 174, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 174, + 393 + ], + "score": 1.0, + "content": "6 RESULTS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "We evaluate the capabilities of our method to learn to control physical PDEs in three different test", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "environments of increasing complexity. We first target a simple but nonlinear 1D equation, for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "which we present an ablation study to quantify accuracy. We then study two-dimensional problems:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 505, + 448 + ], + "score": 1.0, + "content": "an incompressible fluid and a fluid with complex boundaries and indirect control. Full details are", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "given in Appendix D. Supplemental material containing additional sequences for all of the tests can", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 457, + 383, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 383, + 470 + ], + "score": 1.0, + "content": "be downloaded from https://ge.in.tum.de/publications/2020-iclr-holl.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 403, + 506, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 504, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "Burger’s equation. Burger’s equation is a nonlinear PDE that describes the time evolution of a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 485, + 367, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 155, + 498 + ], + "score": 1.0, + "content": "single field,", + "type": "text" + }, + { + "bbox": [ + 156, + 488, + 163, + 496 + ], + "score": 0.72, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 485, + 367, + 498 + ], + "score": 1.0, + "content": "(LeVeque, 1992). Using Eq. 1, it can be written as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 474, + 506, + 498 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 501, + 386, + 529 + ], + "lines": [ + { + "bbox": [ + 225, + 501, + 386, + 529 + ], + "spans": [ + { + "bbox": [ + 225, + 501, + 386, + 529 + ], + "score": 0.95, + "content": "\\mathcal { P } \\left( u , \\frac { \\partial u } { \\partial x } , \\frac { \\partial ^ { 2 } u } { \\partial x ^ { 2 } } \\right) = - u \\cdot \\frac { \\partial u } { \\partial x } + \\nu \\frac { \\partial ^ { 2 } u } { \\partial x ^ { 2 } } .", + "type": "interline_equation", + "image_path": "add3bab97283c6c8d71e236a730c91827c8748af50f0c75ab23157c9530ec379.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 225, + 501, + 386, + 515.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 225, + 515.0, + 386, + 529.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "Examples of the unperturbed evolution are shown in Figure 4a. We let the whole state be observable", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 543, + 456, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 191, + 558 + ], + "score": 1.0, + "content": "and controllable, i.e.", + "type": "text" + }, + { + "bbox": [ + 191, + 544, + 238, + 556 + ], + "score": 0.93, + "content": "o ( t ) = u ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 543, + 318, + 558 + ], + "score": 1.0, + "content": ", which implies that", + "type": "text" + }, + { + "bbox": [ + 318, + 545, + 330, + 555 + ], + "score": 0.87, + "content": "\\boldsymbol { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 543, + 456, + 558 + ], + "score": 1.0, + "content": "can always be reached exactly.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 533, + 505, + 558 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "The results of our ablation study with this equation are shown in Table 1. The table compares the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "resulting forces applied by differently trained models when reconstructing a ground-truth sequence", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 582, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 251, + 597 + ], + "score": 1.0, + "content": "(Figure 4e). The variant denoted by", + "type": "text" + }, + { + "bbox": [ + 252, + 583, + 272, + 594 + ], + "score": 0.27, + "content": "C F E", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 582, + 505, + 597 + ], + "score": 1.0, + "content": "chain uses a neural network to infer the force without any", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 592, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 505, + 609 + ], + "score": 1.0, + "content": "intermediate predictions. With a supervised loss, this method learns to approximate a single step", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "score": 1.0, + "content": "well. However, for longer sequences, results quickly deviate from an ideal trajectory and diverge", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "score": 1.0, + "content": "because the network never learned to account for errors made in previous steps (Figure 4b). Training", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 626, + 504, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 504, + 640 + ], + "score": 1.0, + "content": "the network with the objective loss (Eq. 4) using the differentiable solver greatly increases the quality", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 638, + 504, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 312, + 651 + ], + "score": 1.0, + "content": "of the reconstructions. On average, it applies only", + "type": "text" + }, + { + "bbox": [ + 312, + 638, + 332, + 649 + ], + "score": 0.88, + "content": "3 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 638, + 504, + 651 + ], + "score": 1.0, + "content": "of the force used by the supervised model", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 648, + 361, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 361, + 661 + ], + "score": 1.0, + "content": "as it learns to correct the temporal evolution of the PDE model.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 561, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "Next, we evaluate variants of our predictor-corrector approach, which hierarchically predicts in-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 321, + 690 + ], + "score": 1.0, + "content": "termediate states. Here, the CFE is implemented as", + "type": "text" + }, + { + "bbox": [ + 321, + 677, + 453, + 689 + ], + "score": 0.91, + "content": "F ( t _ { i } ) = ( o ^ { p } ( t _ { i + 1 } ) - u ( t _ { i } ) ) / \\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". Unlike the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "simple CFE chain above, training with the supervised loss and staggered execution produces stable", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "(albeit jittering) trajectories that successfully converge to the target state (Figure 4c). Surprisingly,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "this supervised method reaches almost the same accuracy as the differentiable CFE, despite not hav-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "score": 1.0, + "content": "ing access to physics-based gradients. However, employing the differentiable physics loss greatly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "improves the reconstruction quality, producing solutions that are hard to distinguish from ideal tra-", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 392, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 404 + ], + "score": 1.0, + "content": "jectories (Figure 4d). The prediction refinement scheme further improves the accuracy, but the", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "differences to the staggered execution are relatively small as the predictions of the latter are already", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 415, + 164, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 164, + 426 + ], + "score": 1.0, + "content": "very accurate.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 666, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 102, + 462, + 198 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 109, + 89, + 494, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 87, + 496, + 102 + ], + "spans": [ + { + "bbox": [ + 111, + 87, + 496, + 102 + ], + "score": 1.0, + "content": "Table 1: Quantitative reconstruction evaluation using Burger’s equation, avg. for 100 examples.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 148, + 102, + 462, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 148, + 102, + 462, + 198 + ], + "spans": [ + { + "bbox": [ + 148, + 102, + 462, + 198 + ], + "score": 0.975, + "html": "
Execution schemeTraining lossForce S|F|dtInference time (ms)
CFE chainSupervised83.4± 2.00.024 ± 0.013
CFE chainDiff. Physics28.8±0.80.024 ± 0.013
StaggeredSupervised34.3 ± 1.11.15 ± 0.19
StaggeredDiff. Physics15.3 ± 0.71.15 ± 0.19
RefinedDiff. Physics14.2 ± 0.73.05 ± 0.37
Iterative optim. (60 iter.)Diff. Physics15.3 ± 1.652.7 ± 2.1
Iterative optim. (300 iter.)Diff. Physics10.2 ± 1.9264.0 ± 3.0
", + "type": "table", + "image_path": "cc4216be5f10841072302e21cddf247b32652eb3eac049a78b58769dbab8eb7c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 148, + 102, + 462, + 134.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 148, + 134.0, + 462, + 166.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 148, + 166.0, + 462, + 198.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "image", + "bbox": [ + 113, + 209, + 496, + 339 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 209, + 496, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 209, + 496, + 339 + ], + "spans": [ + { + "bbox": [ + 113, + 209, + 496, + 339 + ], + "score": 0.975, + "type": "image", + "image_path": "bb005f8da3bcc11e8c835add5b8a8bb7cdcdde53a3a133ce4bf604710068fdea.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 113, + 209, + 496, + 252.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 113, + 252.33333333333334, + 496, + 295.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 113, + 295.6666666666667, + 496, + 339.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 348, + 504, + 370 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "Figure 5: Example reconstructed trajectory from (a) the natural flow test set and (b) the shape test", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 359, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 187, + 371 + ], + "score": 1.0, + "content": "set. The target state", + "type": "text" + }, + { + "bbox": [ + 187, + 360, + 198, + 369 + ], + "score": 0.85, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 359, + 288, + 371 + ], + "score": 1.0, + "content": "is shown on the right.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "improves the reconstruction quality, producing solutions that are hard to distinguish from ideal tra-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 392, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 404 + ], + "score": 1.0, + "content": "jectories (Figure 4d). The prediction refinement scheme further improves the accuracy, but the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "differences to the staggered execution are relatively small as the predictions of the latter are already", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 415, + 164, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 164, + 426 + ], + "score": 1.0, + "content": "very accurate.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 444 + ], + "score": 1.0, + "content": "Table 1 also lists the results of classic shooting-based optimization applied to this problem. To match", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "the quality of the staggered execution scheme, the shooting method requires around 60 optimization", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "steps. These steps are significantly more expensive to compute, despite the fast convergence. After", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "around 300 iterations, the classic optimization reaches an optimal value of 10.2 and the loss stops", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "decreasing. Starting the iterative optimization with our method as an initial guess pushes the opti-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "mum slightly lower to 10.1. Thus, even this relatively simple problem shows the advantages of our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 496, + 180, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 180, + 509 + ], + "score": 1.0, + "content": "learned approach.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "Incompressible fluid flow. Next, we apply our algorithm to two-dimensional fluid dynamics", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "problems, which are challenging due to the complexities of the governing Navier-Stokes equa-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 535, + 387, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 282, + 547 + ], + "score": 1.0, + "content": "tions (Batchelor, 1967). For a velocity field", + "type": "text" + }, + { + "bbox": [ + 282, + 537, + 289, + 545 + ], + "score": 0.76, + "content": "\\pmb { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 535, + 387, + 547 + ], + "score": 1.0, + "content": ", these can be written as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 227, + 549, + 383, + 563 + ], + "lines": [ + { + "bbox": [ + 227, + 549, + 383, + 563 + ], + "spans": [ + { + "bbox": [ + 227, + 549, + 383, + 563 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\mathcal { P } ( \\pmb { v } , \\nabla \\pmb { v } ) = - \\pmb { v } \\cdot \\nabla \\pmb { v } + \\nu \\nabla ^ { 2 } \\pmb { v } - \\nabla p , } \\end{array}", + "type": "interline_equation", + "image_path": "4ecccd60623507053d4992a808e765bea32f85c5abe95cfcf0b22a316fcc6edc.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 227, + 549, + 383, + 563 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 504, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 226, + 579 + ], + "score": 1.0, + "content": "subject to the hard constraints", + "type": "text" + }, + { + "bbox": [ + 226, + 566, + 264, + 577 + ], + "score": 0.92, + "content": "\\nabla \\cdot \\pmb { v } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 565, + 281, + 579 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 281, + 566, + 322, + 578 + ], + "score": 0.93, + "content": "\\nabla \\times p = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 565, + 352, + 579 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 352, + 568, + 359, + 578 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 565, + 443, + 579 + ], + "score": 1.0, + "content": "denotes pressure and", + "type": "text" + }, + { + "bbox": [ + 444, + 568, + 450, + 576 + ], + "score": 0.78, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 565, + 505, + 579 + ], + "score": 1.0, + "content": "the viscosity.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 274, + 590 + ], + "score": 1.0, + "content": "In addition, we consider a passive density", + "type": "text" + }, + { + "bbox": [ + 275, + 579, + 281, + 588 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 577, + 398, + 590 + ], + "score": 1.0, + "content": "that moves with the fluid via", + "type": "text" + }, + { + "bbox": [ + 398, + 577, + 472, + 589 + ], + "score": 0.91, + "content": "\\partial \\rho / \\partial t = - \\pmb { v } \\cdot \\nabla \\rho ,", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 577, + 505, + 590 + ], + "score": 1.0, + "content": ". We set", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 588, + 428, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 114, + 598 + ], + "score": 0.67, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 588, + 183, + 600 + ], + "score": 1.0, + "content": "to be hidden and", + "type": "text" + }, + { + "bbox": [ + 183, + 590, + 190, + 600 + ], + "score": 0.82, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 588, + 417, + 600 + ], + "score": 1.0, + "content": "to be observable, and allow forces to be applied to all of", + "type": "text" + }, + { + "bbox": [ + 417, + 590, + 424, + 598 + ], + "score": 0.78, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 588, + 428, + 600 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 195, + 617 + ], + "score": 1.0, + "content": "We run our tests on a", + "type": "text" + }, + { + "bbox": [ + 195, + 604, + 216, + 615 + ], + "score": 0.87, + "content": "1 2 8 ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "grid, resulting in more than 16,000 effective continuous control param-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "eters. We train the OP and CFE networks for two different tasks: reconstruction of natural fluid", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "flows and controlled shape transitions. Example sequences are shown in Figure 5 and a quantita-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "tive evaluation, averaged over 100 examples, is given in Table 2. While all methods manage to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "approximate the target state well, there are considerable differences in the amount of force applied.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "The supervised technique exerts significantly more force than the methods based on the differen-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 671, + 504, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 504, + 683 + ], + "score": 1.0, + "content": "tiable solver, resulting in jittering reconstructions. The prediction refinement scheme produces the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 682, + 477, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 477, + 693 + ], + "score": 1.0, + "content": "smoothest transitions, converging to about half the loss of the staggered, non-refined variant.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "We compare our method to classic shooting algorithms for this incompressible flow problem. While", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "a direct shooting method fails to converge, a more advanced multi-scale shooting approach still re-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "quires 1500 iterations to obtain a level of accuracy that our model achieves almost instantly. In", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 102, + 462, + 198 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 109, + 89, + 494, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 87, + 496, + 102 + ], + "spans": [ + { + "bbox": [ + 111, + 87, + 496, + 102 + ], + "score": 1.0, + "content": "Table 1: Quantitative reconstruction evaluation using Burger’s equation, avg. for 100 examples.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 148, + 102, + 462, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 148, + 102, + 462, + 198 + ], + "spans": [ + { + "bbox": [ + 148, + 102, + 462, + 198 + ], + "score": 0.975, + "html": "
Execution schemeTraining lossForce S|F|dtInference time (ms)
CFE chainSupervised83.4± 2.00.024 ± 0.013
CFE chainDiff. Physics28.8±0.80.024 ± 0.013
StaggeredSupervised34.3 ± 1.11.15 ± 0.19
StaggeredDiff. Physics15.3 ± 0.71.15 ± 0.19
RefinedDiff. Physics14.2 ± 0.73.05 ± 0.37
Iterative optim. (60 iter.)Diff. Physics15.3 ± 1.652.7 ± 2.1
Iterative optim. (300 iter.)Diff. Physics10.2 ± 1.9264.0 ± 3.0
", + "type": "table", + "image_path": "cc4216be5f10841072302e21cddf247b32652eb3eac049a78b58769dbab8eb7c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 148, + 102, + 462, + 134.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 148, + 134.0, + 462, + 166.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 148, + 166.0, + 462, + 198.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "image", + "bbox": [ + 113, + 209, + 496, + 339 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 209, + 496, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 209, + 496, + 339 + ], + "spans": [ + { + "bbox": [ + 113, + 209, + 496, + 339 + ], + "score": 0.975, + "type": "image", + "image_path": "bb005f8da3bcc11e8c835add5b8a8bb7cdcdde53a3a133ce4bf604710068fdea.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 113, + 209, + 496, + 252.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 113, + 252.33333333333334, + 496, + 295.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 113, + 295.6666666666667, + 496, + 339.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 348, + 504, + 370 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "Figure 5: Example reconstructed trajectory from (a) the natural flow test set and (b) the shape test", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 359, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 187, + 371 + ], + "score": 1.0, + "content": "set. The target state", + "type": "text" + }, + { + "bbox": [ + 187, + 360, + 198, + 369 + ], + "score": 0.85, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 359, + 288, + 371 + ], + "score": 1.0, + "content": "is shown on the right.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 505, + 425 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 105, + 380, + 506, + 426 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 444 + ], + "score": 1.0, + "content": "Table 1 also lists the results of classic shooting-based optimization applied to this problem. To match", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "the quality of the staggered execution scheme, the shooting method requires around 60 optimization", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "steps. These steps are significantly more expensive to compute, despite the fast convergence. After", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "around 300 iterations, the classic optimization reaches an optimal value of 10.2 and the loss stops", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "decreasing. Starting the iterative optimization with our method as an initial guess pushes the opti-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "mum slightly lower to 10.1. Thus, even this relatively simple problem shows the advantages of our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 496, + 180, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 180, + 509 + ], + "score": 1.0, + "content": "learned approach.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 429, + 505, + 509 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "Incompressible fluid flow. Next, we apply our algorithm to two-dimensional fluid dynamics", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "problems, which are challenging due to the complexities of the governing Navier-Stokes equa-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 535, + 387, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 282, + 547 + ], + "score": 1.0, + "content": "tions (Batchelor, 1967). For a velocity field", + "type": "text" + }, + { + "bbox": [ + 282, + 537, + 289, + 545 + ], + "score": 0.76, + "content": "\\pmb { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 535, + 387, + 547 + ], + "score": 1.0, + "content": ", these can be written as", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 513, + 505, + 547 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 227, + 549, + 383, + 563 + ], + "lines": [ + { + "bbox": [ + 227, + 549, + 383, + 563 + ], + "spans": [ + { + "bbox": [ + 227, + 549, + 383, + 563 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\mathcal { P } ( \\pmb { v } , \\nabla \\pmb { v } ) = - \\pmb { v } \\cdot \\nabla \\pmb { v } + \\nu \\nabla ^ { 2 } \\pmb { v } - \\nabla p , } \\end{array}", + "type": "interline_equation", + "image_path": "4ecccd60623507053d4992a808e765bea32f85c5abe95cfcf0b22a316fcc6edc.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 227, + 549, + 383, + 563 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 504, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 226, + 579 + ], + "score": 1.0, + "content": "subject to the hard constraints", + "type": "text" + }, + { + "bbox": [ + 226, + 566, + 264, + 577 + ], + "score": 0.92, + "content": "\\nabla \\cdot \\pmb { v } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 565, + 281, + 579 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 281, + 566, + 322, + 578 + ], + "score": 0.93, + "content": "\\nabla \\times p = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 565, + 352, + 579 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 352, + 568, + 359, + 578 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 565, + 443, + 579 + ], + "score": 1.0, + "content": "denotes pressure and", + "type": "text" + }, + { + "bbox": [ + 444, + 568, + 450, + 576 + ], + "score": 0.78, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 565, + 505, + 579 + ], + "score": 1.0, + "content": "the viscosity.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 274, + 590 + ], + "score": 1.0, + "content": "In addition, we consider a passive density", + "type": "text" + }, + { + "bbox": [ + 275, + 579, + 281, + 588 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 577, + 398, + 590 + ], + "score": 1.0, + "content": "that moves with the fluid via", + "type": "text" + }, + { + "bbox": [ + 398, + 577, + 472, + 589 + ], + "score": 0.91, + "content": "\\partial \\rho / \\partial t = - \\pmb { v } \\cdot \\nabla \\rho ,", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 577, + 505, + 590 + ], + "score": 1.0, + "content": ". We set", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 588, + 428, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 114, + 598 + ], + "score": 0.67, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 588, + 183, + 600 + ], + "score": 1.0, + "content": "to be hidden and", + "type": "text" + }, + { + "bbox": [ + 183, + 590, + 190, + 600 + ], + "score": 0.82, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 588, + 417, + 600 + ], + "score": 1.0, + "content": "to be observable, and allow forces to be applied to all of", + "type": "text" + }, + { + "bbox": [ + 417, + 590, + 424, + 598 + ], + "score": 0.78, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 588, + 428, + 600 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 565, + 505, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 604, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 195, + 617 + ], + "score": 1.0, + "content": "We run our tests on a", + "type": "text" + }, + { + "bbox": [ + 195, + 604, + 216, + 615 + ], + "score": 0.87, + "content": "1 2 8 ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "grid, resulting in more than 16,000 effective continuous control param-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "eters. We train the OP and CFE networks for two different tasks: reconstruction of natural fluid", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "flows and controlled shape transitions. Example sequences are shown in Figure 5 and a quantita-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "tive evaluation, averaged over 100 examples, is given in Table 2. While all methods manage to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "approximate the target state well, there are considerable differences in the amount of force applied.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "The supervised technique exerts significantly more force than the methods based on the differen-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 671, + 504, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 504, + 683 + ], + "score": 1.0, + "content": "tiable solver, resulting in jittering reconstructions. The prediction refinement scheme produces the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 682, + 477, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 477, + 693 + ], + "score": 1.0, + "content": "smoothest transitions, converging to about half the loss of the staggered, non-refined variant.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 604, + 506, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "We compare our method to classic shooting algorithms for this incompressible flow problem. While", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "a direct shooting method fails to converge, a more advanced multi-scale shooting approach still re-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "quires 1500 iterations to obtain a level of accuracy that our model achieves almost instantly. In", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "addition, our model successfully learns a solution manifold, while iterative optimization techniques", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 311, + 504, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 504, + 321 + ], + "score": 1.0, + "content": "essentially start from scratch every time. This global view leads our model to more intuitive solu-", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "tions and decreases the likelihood of convergence to undesirable local minima. The solutions of our", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 332, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 344 + ], + "score": 1.0, + "content": "method can also be used as initial guesses for iterative solvers, as illustrated in Appendix D.4. We", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 342, + 504, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 477, + 356 + ], + "score": 1.0, + "content": "find that the iterative optimizer with an initial guess converges to solutions that require only", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 477, + 343, + 504, + 353 + ], + "score": 0.85, + "content": "5 7 . 4 \\%", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "of the force achieved by the iterative optimizer with default initialization. This illustrates how the", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "more global view of the learned solution manifold can improve the solutions of regular optimization", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 376, + 129, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 129, + 388 + ], + "score": 1.0, + "content": "runs.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 699, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 112, + 477, + 164 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 504, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "Table 2: A comparison of methods in terms of final cost for (a) the natural flow setup and (b) the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 480, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 480, + 113 + ], + "score": 1.0, + "content": "shape transitions. The initial distribution is sampled randomly and evolved to the target state.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 132, + 112, + 477, + 164 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 112, + 477, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 112, + 477, + 164 + ], + "score": 0.946, + "html": "
ExecutionLossa) Force LFa)Obs.Lb)Force LFb) Obs.L
StaggeredSupervised243±111.53 ± 0.23n/an/a
StaggeredDiff. Physics22.6 ± 1.10.64± 0.0889±60.331 ± 0.134
RefinedDiff. Physics11.7 ± 0.60.88 ± 0.1175±40.126 ± 0.010
", + "type": "table", + "image_path": "87d6d4a14d62fdb91a87f03d9e09c724dc5a53c53d16def5586b31272cde41ed.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 132, + 112, + 477, + 129.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 132, + 129.33333333333334, + 477, + 146.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 132, + 146.66666666666669, + 477, + 164.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "image", + "bbox": [ + 112, + 178, + 497, + 253 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 178, + 497, + 253 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 178, + 497, + 253 + ], + "spans": [ + { + "bbox": [ + 112, + 178, + 497, + 253 + ], + "score": 0.96, + "type": "image", + "image_path": "c5958a3994726defbe3e6cfbae4e40a4f536f27e96897e25c9de013f2d7e28b8.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 112, + 178, + 497, + 203.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 112, + 203.0, + 497, + 228.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 112, + 228.0, + 497, + 253.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 261, + 504, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 261, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 506, + 273 + ], + "score": 1.0, + "content": "Figure 6: Indirect control sequence. Obstacles are marked white, control regions light blue. The", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 270, + 492, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 492, + 286 + ], + "score": 1.0, + "content": "white arrows indicate the velocity field. The domain is enclosed in a solid box with an open top.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "addition, our model successfully learns a solution manifold, while iterative optimization techniques", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 311, + 504, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 504, + 321 + ], + "score": 1.0, + "content": "essentially start from scratch every time. This global view leads our model to more intuitive solu-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "tions and decreases the likelihood of convergence to undesirable local minima. The solutions of our", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 332, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 344 + ], + "score": 1.0, + "content": "method can also be used as initial guesses for iterative solvers, as illustrated in Appendix D.4. We", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 342, + 504, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 477, + 356 + ], + "score": 1.0, + "content": "find that the iterative optimizer with an initial guess converges to solutions that require only", + "type": "text" + }, + { + "bbox": [ + 477, + 343, + 504, + 353 + ], + "score": 0.85, + "content": "5 7 . 4 \\%", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "of the force achieved by the iterative optimizer with default initialization. This illustrates how the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "more global view of the learned solution manifold can improve the solutions of regular optimization", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 376, + 129, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 129, + 388 + ], + "score": 1.0, + "content": "runs.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 504, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "Splitting the task into prediction and correction ensures that intermediate predicted states are phys-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "ically plausible and allows us to generalize to new tasks. For example, we can infer transitions", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "score": 1.0, + "content": "involving multiple shapes, despite training only on individual shapes. This is demonstrated in Ap-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 425, + 156, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 425, + 156, + 437 + ], + "score": 1.0, + "content": "pendix D.2.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 455 + ], + "score": 1.0, + "content": "Incompressible fluid with indirect control. The next experiment increases the complexity of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "fluid control problem by adding obstacles to the simulated domain and limiting the area that can be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "controlled by the network. An example sequence in this setting is shown in Figure 6. As before,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 176, + 488 + ], + "score": 1.0, + "content": "only the density", + "type": "text" + }, + { + "bbox": [ + 177, + 477, + 184, + 487 + ], + "score": 0.78, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "is observable. Here, the goal is to move the smoke from its initial position", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "near the center into one of the three “buckets” at the top. Control forces can only be applied in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "the peripheral regions, which are outside the visible smoke distribution. Only by synchronizing the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 507, + 504, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 504, + 521 + ], + "score": 1.0, + "content": "5000 continuous control parameters can a directed velocity field be constructed in the central region.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "We first infer trajectories using a trained CFE network and predictions that move the smoke into the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 369, + 547 + ], + "score": 1.0, + "content": "desired bucket in a straight line. This baseline manages to transfer", + "type": "text" + }, + { + "bbox": [ + 369, + 536, + 419, + 547 + ], + "score": 0.91, + "content": "8 9 \\% \\pm 2 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "of the smoke into the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "target bucket. Next we enable the hierarchical predictions and train the OPs. This version manages", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 557, + 504, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 158, + 570 + ], + "score": 1.0, + "content": "to maneuver", + "type": "text" + }, + { + "bbox": [ + 159, + 558, + 229, + 569 + ], + "score": 0.92, + "content": "9 9 . 2 2 \\% \\pm 0 . 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 557, + 443, + 570 + ], + "score": 1.0, + "content": "of the smoke into the desired buckets while requiring", + "type": "text" + }, + { + "bbox": [ + 444, + 558, + 504, + 569 + ], + "score": 0.9, + "content": "1 9 . 1 \\% \\pm 1 . { \\bar { 0 } } \\%", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 568, + 149, + 581 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 149, + 581 + ], + "score": 1.0, + "content": "less force.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 586, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "For comparison, Table 3 also lists success rate and execution time for a direct optimization. Despite", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 251, + 610 + ], + "score": 1.0, + "content": "only obtaining a low success rate of", + "type": "text" + }, + { + "bbox": [ + 251, + 597, + 270, + 607 + ], + "score": 0.87, + "content": "82 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 596, + 505, + 610 + ], + "score": 1.0, + "content": ", the shooting method requires several orders of magnitude", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "longer than evaluating our trained model. Since all optimizations are independent of each other,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 619, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 630 + ], + "score": 1.0, + "content": "some find better solutions than others, reflected in the higher standard deviation. The increased", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "number of free parameters and complexity of the fluid dynamics to be controlled make this problem", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "intractable for the shooting method, while our model can leverage the learned representation to infer", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 651, + 374, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 374, + 664 + ], + "score": 1.0, + "content": "a solution very quickly. Further details are given in Appendix D.3.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 108, + 682, + 201, + 695 + ], + "lines": [ + { + "bbox": [ + 104, + 680, + 203, + 698 + ], + "spans": [ + { + "bbox": [ + 104, + 680, + 203, + 698 + ], + "score": 1.0, + "content": "7 CONCLUSIONS", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "We have demonstrated that deep learning models in conjunction with a differentiable physics solver", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "can successfully predict the behavior of complex physical systems and learn to control them. The in-", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 112, + 477, + 164 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 504, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "Table 2: A comparison of methods in terms of final cost for (a) the natural flow setup and (b) the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 480, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 480, + 113 + ], + "score": 1.0, + "content": "shape transitions. The initial distribution is sampled randomly and evolved to the target state.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 132, + 112, + 477, + 164 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 112, + 477, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 112, + 477, + 164 + ], + "score": 0.946, + "html": "
ExecutionLossa) Force LFa)Obs.Lb)Force LFb) Obs.L
StaggeredSupervised243±111.53 ± 0.23n/an/a
StaggeredDiff. Physics22.6 ± 1.10.64± 0.0889±60.331 ± 0.134
RefinedDiff. Physics11.7 ± 0.60.88 ± 0.1175±40.126 ± 0.010
", + "type": "table", + "image_path": "87d6d4a14d62fdb91a87f03d9e09c724dc5a53c53d16def5586b31272cde41ed.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 132, + 112, + 477, + 129.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 132, + 129.33333333333334, + 477, + 146.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 132, + 146.66666666666669, + 477, + 164.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "image", + "bbox": [ + 112, + 178, + 497, + 253 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 178, + 497, + 253 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 178, + 497, + 253 + ], + "spans": [ + { + "bbox": [ + 112, + 178, + 497, + 253 + ], + "score": 0.96, + "type": "image", + "image_path": "c5958a3994726defbe3e6cfbae4e40a4f536f27e96897e25c9de013f2d7e28b8.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 112, + 178, + 497, + 203.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 112, + 203.0, + 497, + 228.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 112, + 228.0, + 497, + 253.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 261, + 504, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 261, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 506, + 273 + ], + "score": 1.0, + "content": "Figure 6: Indirect control sequence. Obstacles are marked white, control regions light blue. The", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 270, + 492, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 492, + 286 + ], + "score": 1.0, + "content": "white arrows indicate the velocity field. The domain is enclosed in a solid box with an open top.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 387 + ], + "lines": [], + "index": 13.5, + "bbox_fs": [ + 105, + 298, + 505, + 388 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 504, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "Splitting the task into prediction and correction ensures that intermediate predicted states are phys-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "ically plausible and allows us to generalize to new tasks. For example, we can infer transitions", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "score": 1.0, + "content": "involving multiple shapes, despite training only on individual shapes. This is demonstrated in Ap-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 425, + 156, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 425, + 156, + 437 + ], + "score": 1.0, + "content": "pendix D.2.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 104, + 392, + 506, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 505, + 455 + ], + "score": 1.0, + "content": "Incompressible fluid with indirect control. The next experiment increases the complexity of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "fluid control problem by adding obstacles to the simulated domain and limiting the area that can be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "controlled by the network. An example sequence in this setting is shown in Figure 6. As before,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 176, + 488 + ], + "score": 1.0, + "content": "only the density", + "type": "text" + }, + { + "bbox": [ + 177, + 477, + 184, + 487 + ], + "score": 0.78, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "is observable. Here, the goal is to move the smoke from its initial position", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "near the center into one of the three “buckets” at the top. Control forces can only be applied in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "the peripheral regions, which are outside the visible smoke distribution. Only by synchronizing the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 507, + 504, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 504, + 521 + ], + "score": 1.0, + "content": "5000 continuous control parameters can a directed velocity field be constructed in the central region.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 441, + 505, + 521 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "We first infer trajectories using a trained CFE network and predictions that move the smoke into the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 369, + 547 + ], + "score": 1.0, + "content": "desired bucket in a straight line. This baseline manages to transfer", + "type": "text" + }, + { + "bbox": [ + 369, + 536, + 419, + 547 + ], + "score": 0.91, + "content": "8 9 \\% \\pm 2 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "of the smoke into the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "target bucket. Next we enable the hierarchical predictions and train the OPs. This version manages", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 557, + 504, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 158, + 570 + ], + "score": 1.0, + "content": "to maneuver", + "type": "text" + }, + { + "bbox": [ + 159, + 558, + 229, + 569 + ], + "score": 0.92, + "content": "9 9 . 2 2 \\% \\pm 0 . 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 557, + 443, + 570 + ], + "score": 1.0, + "content": "of the smoke into the desired buckets while requiring", + "type": "text" + }, + { + "bbox": [ + 444, + 558, + 504, + 569 + ], + "score": 0.9, + "content": "1 9 . 1 \\% \\pm 1 . { \\bar { 0 } } \\%", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 568, + 149, + 581 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 149, + 581 + ], + "score": 1.0, + "content": "less force.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 525, + 506, + 581 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 586, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "For comparison, Table 3 also lists success rate and execution time for a direct optimization. Despite", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 251, + 610 + ], + "score": 1.0, + "content": "only obtaining a low success rate of", + "type": "text" + }, + { + "bbox": [ + 251, + 597, + 270, + 607 + ], + "score": 0.87, + "content": "82 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 596, + 505, + 610 + ], + "score": 1.0, + "content": ", the shooting method requires several orders of magnitude", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "longer than evaluating our trained model. Since all optimizations are independent of each other,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 619, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 630 + ], + "score": 1.0, + "content": "some find better solutions than others, reflected in the higher standard deviation. The increased", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "number of free parameters and complexity of the fluid dynamics to be controlled make this problem", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "intractable for the shooting method, while our model can leverage the learned representation to infer", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 651, + 374, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 374, + 664 + ], + "score": 1.0, + "content": "a solution very quickly. Further details are given in Appendix D.3.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 586, + 505, + 664 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 682, + 201, + 695 + ], + "lines": [ + { + "bbox": [ + 104, + 680, + 203, + 698 + ], + "spans": [ + { + "bbox": [ + 104, + 680, + 203, + 698 + ], + "score": 1.0, + "content": "7 CONCLUSIONS", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "We have demonstrated that deep learning models in conjunction with a differentiable physics solver", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "can successfully predict the behavior of complex physical systems and learn to control them. The in-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "troduction of a hierarchical predictor-corrector architecture allowed the model to learn to reconstruct", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 200, + 443, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 443, + 215 + ], + "score": 1.0, + "content": "long sequences by treating the physical behavior on different time scales separately.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 112, + 468, + 170 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 503, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 88, + 504, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 88, + 504, + 101 + ], + "score": 1.0, + "content": "Table 3: Comparison of different methods on the task of moving a distribution of smoke into the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 312, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 312, + 113 + ], + "score": 1.0, + "content": "target region by applying forces outside the region.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 139, + 112, + 468, + 170 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 139, + 112, + 468, + 170 + ], + "spans": [ + { + "bbox": [ + 139, + 112, + 468, + 170 + ], + "score": 0.978, + "html": "
MethodOptimized quantityInside target (%)Inference time (ms)
Straight trajectoryCFE89.5 ± 2.631.46 ± 0.20
Staggered predictionsCFE,OPn99.22 ± 0.1567.40 ± 0.20
Iterative optim.Control velocity82.1 ± 7.3266.5 · 103
", + "type": "table", + "image_path": "3597e06625e12f3067e9bf38c78de557cf7d2df71a93bc7f564d311963f8fad9.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 139, + 112, + 468, + 131.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 139, + 131.33333333333334, + 468, + 150.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 139, + 150.66666666666669, + 468, + 170.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 106, + 190, + 503, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "troduction of a hierarchical predictor-corrector architecture allowed the model to learn to reconstruct", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 200, + 443, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 443, + 215 + ], + "score": 1.0, + "content": "long sequences by treating the physical behavior on different time scales separately.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 108, + 218, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "We have shown that using a differentiable solver greatly benefits the quality of solutions since the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "networks can learn how their decisions will affect the future. In our experiments, hierarchical infer-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 240, + 501, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 501, + 253 + ], + "score": 1.0, + "content": "ence schemes outperform traditional sequential agents because they can easily learn to plan ahead.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 257, + 505, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "To model realistic environments, we have introduced observations to our pipeline which restrict the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "information available to the learning agent. While the PDE solver still requires full state information", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 279, + 428, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 428, + 291 + ], + "score": 1.0, + "content": "to run the simulation, this restriction does not apply when the agent is deployed.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 309 + ], + "score": 1.0, + "content": "While we do not believe that learning approaches will replace iterative optimization, our method", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 307, + 504, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 504, + 319 + ], + "score": 1.0, + "content": "shows that it is possible to learn representations of solution manifolds for optimal control trajectories", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "using data-driven approaches. Fast inference is vital in time-critical applications and can also be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "used in conjunction with classical solvers to speed up convergence and ultimately produce better", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 339, + 148, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 148, + 352 + ], + "score": 1.0, + "content": "solutions.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 368, + 244, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 245, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 245, + 383 + ], + "score": 1.0, + "content": "8 ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 488, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 491, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 491, + 407 + ], + "score": 1.0, + "content": "This work was supported in part by the ERC Starting Grant realFlow (ERC-2015-StG-637014).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 107, + 421, + 176, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 176, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 176, + 435 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 505, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 115, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. TensorFlow: A system for large-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 462, + 504, + 476 + ], + "spans": [ + { + "bbox": [ + 115, + 462, + 504, + 476 + ], + "score": 1.0, + "content": "scale machine learning. In Symposium on Operating Systems Design and Implementation, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 502, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 482, + 504, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 504, + 495 + ], + "score": 1.0, + "content": "Brandon Amos and J Zico Kolter. OptNet: Differentiable optimization as a layer in neural networks.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 493, + 181, + 505 + ], + "spans": [ + { + "bbox": [ + 116, + 493, + 181, + 505 + ], + "score": 1.0, + "content": "In ICML, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 506, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "Yohai Bar-Sinai, Stephan Hoyer, Jason Hickey, and Michael P. Brenner. Learning data-driven dis-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 523, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 506, + 538 + ], + "score": 1.0, + "content": "cretizations for partial differential equations. Proceedings of the National Academy of Sciences,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 536, + 178, + 547 + ], + "spans": [ + { + "bbox": [ + 117, + 536, + 178, + 547 + ], + "score": 1.0, + "content": "116(31), 2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 482, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 482, + 569 + ], + "score": 1.0, + "content": "George K. Batchelor. An Introduction to Fluid Dynamics. Cambridge University Press, 1967.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 506, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum. Simulation as an engine of physical", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 586, + 469, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 586, + 469, + 599 + ], + "score": 1.0, + "content": "scene understanding. Proceedings of the National Academy of Sciences, 110(45), 2013.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 504, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 619 + ], + "score": 1.0, + "content": "Thomas R. Bewley. Flow control: new challenges for a new renaissance. Progress in Aerospace", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 617, + 196, + 629 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 196, + 629 + ], + "score": 1.0, + "content": "Sciences, 37, 2001.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 105, + 636, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "Hans Georg Bock and Karl-Josef Plitt. A multiple shooting algorithm for direct solution of optimal", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 648, + 356, + 660 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 356, + 660 + ], + "score": 1.0, + "content": "control problems. IFAC Proceedings Volumes, 17(2), 1984.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 678, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 504, + 691 + ], + "score": 1.0, + "content": "Maclaurin, and Skye Wanderman-Milne. JAX: Composable transformations of Python+NumPy", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 690, + 220, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 220, + 701 + ], + "score": 1.0, + "content": "programs. GitHub, 2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "score": 1.0, + "content": "John Bradshaw, Matt J Kusner, Brooks Paige, Marwin HS Segler, and Jose Miguel Hern ´ andez-´", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 720, + 370, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 370, + 732 + ], + "score": 1.0, + "content": "Lobato. A generative model for electron paths. In ICLR, 2019.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 112, + 468, + 170 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 503, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 88, + 504, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 88, + 504, + 101 + ], + "score": 1.0, + "content": "Table 3: Comparison of different methods on the task of moving a distribution of smoke into the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 312, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 312, + 113 + ], + "score": 1.0, + "content": "target region by applying forces outside the region.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 139, + 112, + 468, + 170 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 139, + 112, + 468, + 170 + ], + "spans": [ + { + "bbox": [ + 139, + 112, + 468, + 170 + ], + "score": 0.978, + "html": "
MethodOptimized quantityInside target (%)Inference time (ms)
Straight trajectoryCFE89.5 ± 2.631.46 ± 0.20
Staggered predictionsCFE,OPn99.22 ± 0.1567.40 ± 0.20
Iterative optim.Control velocity82.1 ± 7.3266.5 · 103
", + "type": "table", + "image_path": "3597e06625e12f3067e9bf38c78de557cf7d2df71a93bc7f564d311963f8fad9.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 139, + 112, + 468, + 131.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 139, + 131.33333333333334, + 468, + 150.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 139, + 150.66666666666669, + 468, + 170.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 106, + 190, + 503, + 213 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 190, + 505, + 215 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 218, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "We have shown that using a differentiable solver greatly benefits the quality of solutions since the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "networks can learn how their decisions will affect the future. In our experiments, hierarchical infer-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 240, + 501, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 501, + 253 + ], + "score": 1.0, + "content": "ence schemes outperform traditional sequential agents because they can easily learn to plan ahead.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 218, + 505, + 253 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 257, + 505, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "To model realistic environments, we have introduced observations to our pipeline which restrict the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "information available to the learning agent. While the PDE solver still requires full state information", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 279, + 428, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 428, + 291 + ], + "score": 1.0, + "content": "to run the simulation, this restriction does not apply when the agent is deployed.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 257, + 505, + 291 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 505, + 309 + ], + "score": 1.0, + "content": "While we do not believe that learning approaches will replace iterative optimization, our method", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 307, + 504, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 504, + 319 + ], + "score": 1.0, + "content": "shows that it is possible to learn representations of solution manifolds for optimal control trajectories", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "using data-driven approaches. Fast inference is vital in time-critical applications and can also be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "used in conjunction with classical solvers to speed up convergence and ultimately produce better", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 339, + 148, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 148, + 352 + ], + "score": 1.0, + "content": "solutions.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 295, + 505, + 352 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 368, + 244, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 245, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 245, + 383 + ], + "score": 1.0, + "content": "8 ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 488, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 491, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 491, + 407 + ], + "score": 1.0, + "content": "This work was supported in part by the ERC Starting Grant realFlow (ERC-2015-StG-637014).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 392, + 491, + 407 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 421, + 176, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 176, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 176, + 435 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 505, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 115, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. TensorFlow: A system for large-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 462, + 504, + 476 + ], + "spans": [ + { + "bbox": [ + 115, + 462, + 504, + 476 + ], + "score": 1.0, + "content": "scale machine learning. In Symposium on Operating Systems Design and Implementation, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 440, + 505, + 476 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 502, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 482, + 504, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 504, + 495 + ], + "score": 1.0, + "content": "Brandon Amos and J Zico Kolter. OptNet: Differentiable optimization as a layer in neural networks.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 493, + 181, + 505 + ], + "spans": [ + { + "bbox": [ + 116, + 493, + 181, + 505 + ], + "score": 1.0, + "content": "In ICML, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 482, + 504, + 505 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 506, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "Yohai Bar-Sinai, Stephan Hoyer, Jason Hickey, and Michael P. Brenner. Learning data-driven dis-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 523, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 506, + 538 + ], + "score": 1.0, + "content": "cretizations for partial differential equations. Proceedings of the National Academy of Sciences,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 117, + 536, + 178, + 547 + ], + "spans": [ + { + "bbox": [ + 117, + 536, + 178, + 547 + ], + "score": 1.0, + "content": "116(31), 2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 513, + 506, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 482, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 482, + 569 + ], + "score": 1.0, + "content": "George K. Batchelor. An Introduction to Fluid Dynamics. Cambridge University Press, 1967.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 554, + 482, + 569 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 506, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum. Simulation as an engine of physical", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 586, + 469, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 586, + 469, + 599 + ], + "score": 1.0, + "content": "scene understanding. Proceedings of the National Academy of Sciences, 110(45), 2013.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 575, + 505, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 504, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 619 + ], + "score": 1.0, + "content": "Thomas R. Bewley. Flow control: new challenges for a new renaissance. Progress in Aerospace", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 617, + 196, + 629 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 196, + 629 + ], + "score": 1.0, + "content": "Sciences, 37, 2001.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 604, + 505, + 629 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 636, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "Hans Georg Bock and Karl-Josef Plitt. A multiple shooting algorithm for direct solution of optimal", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 648, + 356, + 660 + ], + "spans": [ + { + "bbox": [ + 115, + 648, + 356, + 660 + ], + "score": 1.0, + "content": "control problems. IFAC Proceedings Volumes, 17(2), 1984.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 636, + 505, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 678, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 504, + 691 + ], + "score": 1.0, + "content": "Maclaurin, and Skye Wanderman-Milne. JAX: Composable transformations of Python+NumPy", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 690, + 220, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 220, + 701 + ], + "score": 1.0, + "content": "programs. GitHub, 2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 668, + 505, + 701 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 504, + 723 + ], + "score": 1.0, + "content": "John Bradshaw, Matt J Kusner, Brooks Paige, Marwin HS Segler, and Jose Miguel Hern ´ andez-´", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 720, + 370, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 370, + 732 + ], + "score": 1.0, + "content": "Lobato. A generative model for electron paths. In ICLR, 2019.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 708, + 504, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Tian Qi Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differ-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 435, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 435, + 106 + ], + "score": 1.0, + "content": "ential equations. In Advances in Neural Information Processing Systems, 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 504, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 504, + 125 + ], + "score": 1.0, + "content": "Alexandre Joel Chorin. The numerical solution of the Navier-Stokes equations for an incompressible", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 385, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 385, + 135 + ], + "score": 1.0, + "content": "fluid. Bulletin of the American Mathematical Society, 73(6), 1967.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "Richard Courant and David Hilbert. Methods of Mathematical Physics, volume II: Partial Differen-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 153, + 300, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 153, + 300, + 166 + ], + "score": 1.0, + "content": "tial Equations. Interscience Publishers, 1962.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 172, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Filipe de Avila Belbute-Peres, Kevin Smith, Kelsey Allen, Josh Tenenbaum, and J Zico Kolter.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 183, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 183, + 505, + 197 + ], + "score": 1.0, + "content": "End-to-end differentiable physics for learning and control. In Advances in Neural Information", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 224, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 224, + 207 + ], + "score": 1.0, + "content": "Processing Systems, 2018.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 213, + 504, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 228 + ], + "score": 1.0, + "content": "Jonas Degrave, Michiel Hermans, Joni Dambre, and Francis wyffels. A differentiable physics engine", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 225, + 387, + 237 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 387, + 237 + ], + "score": 1.0, + "content": "for deep learning in robotics. Frontiers in Neurorobotics, 13, 2019.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 505, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 505, + 256 + ], + "score": 1.0, + "content": "Patrick E Farrell, David A Ham, Simon W Funke, and Marie E Rognes. Automated derivation of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 254, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 506, + 269 + ], + "score": 1.0, + "content": "adjoint of high-level transient finite element programs. SIAM Journal on Scientific Computing,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 266, + 169, + 278 + ], + "spans": [ + { + "bbox": [ + 116, + 266, + 169, + 278 + ], + "score": 1.0, + "content": "35(4), 2013.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 105, + 284, + 504, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 299 + ], + "score": 1.0, + "content": "Ciprian Foias, George R Sell, and Roger Temam. Inertial manifolds for nonlinear evolutionary", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 296, + 349, + 308 + ], + "spans": [ + { + "bbox": [ + 116, + 296, + 349, + 308 + ], + "score": 1.0, + "content": "equations. Journal of Differential Equations, 73(2), 1988.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 315, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "David A Fournier, Hans J Skaug, Johnoel Ancheta, James Ianelli, Arni Magnusson, Mark N Maun-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 326, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 116, + 326, + 505, + 338 + ], + "score": 1.0, + "content": "der, Anders Nielsen, and John Sibert. Ad model builder: Using automatic differentiation for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 116, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "statistical inference of highly parameterized complex nonlinear models. Optimization Methods", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 348, + 225, + 360 + ], + "spans": [ + { + "bbox": [ + 115, + 348, + 225, + 360 + ], + "score": 1.0, + "content": "and Software, 27(2), 2012.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 504, + 380 + ], + "score": 1.0, + "content": "Michael B Giles and Niles A Pierce. An introduction to the adjoint approach to design. Flow,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 117, + 378, + 273, + 390 + ], + "spans": [ + { + "bbox": [ + 117, + 378, + 273, + 390 + ], + "score": 1.0, + "content": "Turbulence and Combustion, 65, 2000.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 503, + 420 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 409, + 345, + 420 + ], + "spans": [ + { + "bbox": [ + 116, + 409, + 345, + 420 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. In ICML, 2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 426, + 433, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 433, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 433, + 442 + ], + "score": 1.0, + "content": "Gene H Golub and Charles F Van Loan. Matrix Computations. JHU press, 2012.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 446, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 445, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 460 + ], + "score": 1.0, + "content": "Andreas Griewank and Andrea Walther. Evaluating Derivatives: Principles and Techniques of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 458, + 336, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 458, + 336, + 469 + ], + "score": 1.0, + "content": "Algorithmic Differentiation. SIAM, 2nd edition, 2008.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 108, + 476, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 489 + ], + "score": 1.0, + "content": "Andreas Griewank, David Juedes, and Jean Utke. Algorithm 755: ADOL-C: A package for the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 318, + 500 + ], + "score": 1.0, + "content": "automatic differentiation of algorithms written in", + "type": "text" + }, + { + "bbox": [ + 319, + 488, + 348, + 498 + ], + "score": 0.5, + "content": "\\mathrm { C / C } { + + }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 488, + 505, + 500 + ], + "score": 1.0, + "content": ". ACM Transactions on Mathematical", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 498, + 243, + 511 + ], + "spans": [ + { + "bbox": [ + 116, + 498, + 243, + 511 + ], + "score": 1.0, + "content": "Software (TOMS), 22(2), 1996.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 503, + 541 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 531 + ], + "score": 1.0, + "content": "Charles R Hargraves and Stephen W Paris. Direct trajectory optimization using nonlinear program-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 529, + 443, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 443, + 541 + ], + "score": 1.0, + "content": "ming and collocation. Journal of Guidance, Control, and Dynamics, 10(4), 1987.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 504, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "Francis Harlow and Eddie Welch. Numerical calculation of time-dependent viscous incompressible", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 559, + 365, + 571 + ], + "spans": [ + { + "bbox": [ + 117, + 559, + 365, + 571 + ], + "score": 1.0, + "content": "flow of fluid with free surface. Physics of Fluids, 8(12), 1965.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 108, + 577, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "Carsten Hartmann, Juan C Latorre, Wei Zhang, and Grigorios A Pavliotis. Optimal control of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 589, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 116, + 589, + 505, + 602 + ], + "score": 1.0, + "content": "multiscale systems using reduced-order models. Journal of Computational Dynamics, 1(2), 2014.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 608, + 504, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 504, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 504, + 623 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 619, + 212, + 630 + ], + "spans": [ + { + "bbox": [ + 116, + 619, + 212, + 630 + ], + "score": 1.0, + "content": "nition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 105, + 638, + 504, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 652 + ], + "score": 1.0, + "content": "Jun-Ting Hsieh, Shengjia Zhao, Stephan Eismann, Lucia Mirabella, and Stefano Ermon. Learning", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 649, + 379, + 662 + ], + "spans": [ + { + "bbox": [ + 115, + 649, + 379, + 662 + ], + "score": 1.0, + "content": "neural PDE solvers with convergence guarantees. In ICLR, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "Yuanming Hu, Jiancheng Liu, Andrew Spielberg, Joshua B Tenenbaum, William T Freeman, Jia-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 679, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 114, + 679, + 505, + 692 + ], + "score": 1.0, + "content": "jun Wu, Daniela Rus, and Wojciech Matusik. ChainQueen: A real-time differentiable physical", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 289, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 289, + 702 + ], + "score": 1.0, + "content": "simulator for soft robotics. In ICRA, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "John Ingraham, Adam Riesselman, Chris Sander, and Debora Marks. Learning protein structure", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 117, + 721, + 305, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 305, + 732 + ], + "score": 1.0, + "content": "with a differentiable simulator. In ICLR, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Tian Qi Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differ-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 435, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 435, + 106 + ], + "score": 1.0, + "content": "ential equations. In Advances in Neural Information Processing Systems, 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 504, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 504, + 125 + ], + "score": 1.0, + "content": "Alexandre Joel Chorin. The numerical solution of the Navier-Stokes equations for an incompressible", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 385, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 385, + 135 + ], + "score": 1.0, + "content": "fluid. Bulletin of the American Mathematical Society, 73(6), 1967.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 112, + 504, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "Richard Courant and David Hilbert. Methods of Mathematical Physics, volume II: Partial Differen-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 153, + 300, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 153, + 300, + 166 + ], + "score": 1.0, + "content": "tial Equations. Interscience Publishers, 1962.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 142, + 505, + 166 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 172, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Filipe de Avila Belbute-Peres, Kevin Smith, Kelsey Allen, Josh Tenenbaum, and J Zico Kolter.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 183, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 115, + 183, + 505, + 197 + ], + "score": 1.0, + "content": "End-to-end differentiable physics for learning and control. In Advances in Neural Information", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 224, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 224, + 207 + ], + "score": 1.0, + "content": "Processing Systems, 2018.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 172, + 505, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 213, + 504, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 228 + ], + "score": 1.0, + "content": "Jonas Degrave, Michiel Hermans, Joni Dambre, and Francis wyffels. A differentiable physics engine", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 225, + 387, + 237 + ], + "spans": [ + { + "bbox": [ + 115, + 225, + 387, + 237 + ], + "score": 1.0, + "content": "for deep learning in robotics. Frontiers in Neurorobotics, 13, 2019.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 212, + 505, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 505, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 244, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 505, + 256 + ], + "score": 1.0, + "content": "Patrick E Farrell, David A Ham, Simon W Funke, and Marie E Rognes. Automated derivation of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 254, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 506, + 269 + ], + "score": 1.0, + "content": "adjoint of high-level transient finite element programs. SIAM Journal on Scientific Computing,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 266, + 169, + 278 + ], + "spans": [ + { + "bbox": [ + 116, + 266, + 169, + 278 + ], + "score": 1.0, + "content": "35(4), 2013.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 244, + 506, + 278 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 284, + 504, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 299 + ], + "score": 1.0, + "content": "Ciprian Foias, George R Sell, and Roger Temam. Inertial manifolds for nonlinear evolutionary", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 296, + 349, + 308 + ], + "spans": [ + { + "bbox": [ + 116, + 296, + 349, + 308 + ], + "score": 1.0, + "content": "equations. Journal of Differential Equations, 73(2), 1988.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 283, + 505, + 308 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 315, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "David A Fournier, Hans J Skaug, Johnoel Ancheta, James Ianelli, Arni Magnusson, Mark N Maun-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 326, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 116, + 326, + 505, + 338 + ], + "score": 1.0, + "content": "der, Anders Nielsen, and John Sibert. Ad model builder: Using automatic differentiation for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 338, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 116, + 338, + 505, + 350 + ], + "score": 1.0, + "content": "statistical inference of highly parameterized complex nonlinear models. Optimization Methods", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 348, + 225, + 360 + ], + "spans": [ + { + "bbox": [ + 115, + 348, + 225, + 360 + ], + "score": 1.0, + "content": "and Software, 27(2), 2012.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 315, + 505, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 504, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 504, + 380 + ], + "score": 1.0, + "content": "Michael B Giles and Niles A Pierce. An introduction to the adjoint approach to design. Flow,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 117, + 378, + 273, + 390 + ], + "spans": [ + { + "bbox": [ + 117, + 378, + 273, + 390 + ], + "score": 1.0, + "content": "Turbulence and Combustion, 65, 2000.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 366, + 504, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 503, + 420 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 409, + 345, + 420 + ], + "spans": [ + { + "bbox": [ + 116, + 409, + 345, + 420 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. In ICML, 2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 397, + 505, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 426, + 433, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 433, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 433, + 442 + ], + "score": 1.0, + "content": "Gene H Golub and Charles F Van Loan. Matrix Computations. JHU press, 2012.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 425, + 433, + 442 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 446, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 445, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 460 + ], + "score": 1.0, + "content": "Andreas Griewank and Andrea Walther. Evaluating Derivatives: Principles and Techniques of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 458, + 336, + 469 + ], + "spans": [ + { + "bbox": [ + 115, + 458, + 336, + 469 + ], + "score": 1.0, + "content": "Algorithmic Differentiation. SIAM, 2nd edition, 2008.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 445, + 506, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 476, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 489 + ], + "score": 1.0, + "content": "Andreas Griewank, David Juedes, and Jean Utke. Algorithm 755: ADOL-C: A package for the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 318, + 500 + ], + "score": 1.0, + "content": "automatic differentiation of algorithms written in", + "type": "text" + }, + { + "bbox": [ + 319, + 488, + 348, + 498 + ], + "score": 0.5, + "content": "\\mathrm { C / C } { + + }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 488, + 505, + 500 + ], + "score": 1.0, + "content": ". ACM Transactions on Mathematical", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 498, + 243, + 511 + ], + "spans": [ + { + "bbox": [ + 116, + 498, + 243, + 511 + ], + "score": 1.0, + "content": "Software (TOMS), 22(2), 1996.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 475, + 505, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 503, + 541 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 531 + ], + "score": 1.0, + "content": "Charles R Hargraves and Stephen W Paris. Direct trajectory optimization using nonlinear program-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 529, + 443, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 443, + 541 + ], + "score": 1.0, + "content": "ming and collocation. Journal of Guidance, Control, and Dynamics, 10(4), 1987.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 517, + 505, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 504, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "Francis Harlow and Eddie Welch. Numerical calculation of time-dependent viscous incompressible", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 559, + 365, + 571 + ], + "spans": [ + { + "bbox": [ + 117, + 559, + 365, + 571 + ], + "score": 1.0, + "content": "flow of fluid with free surface. Physics of Fluids, 8(12), 1965.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 546, + 505, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 577, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "Carsten Hartmann, Juan C Latorre, Wei Zhang, and Grigorios A Pavliotis. Optimal control of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 589, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 116, + 589, + 505, + 602 + ], + "score": 1.0, + "content": "multiscale systems using reduced-order models. Journal of Computational Dynamics, 1(2), 2014.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 578, + 506, + 602 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 608, + 504, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 504, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 504, + 623 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 619, + 212, + 630 + ], + "spans": [ + { + "bbox": [ + 116, + 619, + 212, + 630 + ], + "score": 1.0, + "content": "nition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 605, + 504, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 638, + 504, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 652 + ], + "score": 1.0, + "content": "Jun-Ting Hsieh, Shengjia Zhao, Stephan Eismann, Lucia Mirabella, and Stefano Ermon. Learning", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 649, + 379, + 662 + ], + "spans": [ + { + "bbox": [ + 115, + 649, + 379, + 662 + ], + "score": 1.0, + "content": "neural PDE solvers with convergence guarantees. In ICLR, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 636, + 505, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 504, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "Yuanming Hu, Jiancheng Liu, Andrew Spielberg, Joshua B Tenenbaum, William T Freeman, Jia-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 679, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 114, + 679, + 505, + 692 + ], + "score": 1.0, + "content": "jun Wu, Daniela Rus, and Wojciech Matusik. ChainQueen: A real-time differentiable physical", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 289, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 289, + 702 + ], + "score": 1.0, + "content": "simulator for soft robotics. In ICRA, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 668, + 505, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "John Ingraham, Adam Riesselman, Chris Sander, and Debora Marks. Learning protein structure", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 117, + 721, + 305, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 305, + 732 + ], + "score": 1.0, + "content": "with a differentiable simulator. In ICLR, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 104, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Antony Jameson. Aerodynamic design via control theory. Journal of Scientific Computing, 3(3),", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 142, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 142, + 105 + ], + "score": 1.0, + "content": "1988.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 112, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 126 + ], + "score": 1.0, + "content": "Dieter Kraft. On converting optimal control problems into nonlinear programming problems. In", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 123, + 322, + 136 + ], + "spans": [ + { + "bbox": [ + 117, + 123, + 322, + 136 + ], + "score": 1.0, + "content": "Computational Mathematical Programming. 1985.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 109, + 142, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "Lubor Ladicky, SoHyeon Jeong, Barbara Solenthaler, Marc Pollefeys, and Markus Gross. Data-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "driven fluid simulations using regression forests. ACM Transactions on Graphics (TOG), 34(6),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 143, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 143, + 177 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 183, + 376, + 196 + ], + "lines": [ + { + "bbox": [ + 105, + 183, + 377, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 377, + 198 + ], + "score": 1.0, + "content": "Horace Lamb. Hydrodynamics. Cambridge University Press, 1932.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 203, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 504, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 504, + 216 + ], + "score": 1.0, + "content": "Daniel B Leineweber, Irene Bauer, Hans Georg Bock, and Johannes P Schloder. An efficient mul-¨", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 116, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "tiple shooting based reduced SQP strategy for large-scale dynamic process optimization. Part 1:", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 225, + 402, + 239 + ], + "spans": [ + { + "bbox": [ + 116, + 225, + 402, + 239 + ], + "score": 1.0, + "content": "Theoretical aspects. Computers & Chemical Engineering, 27(2), 2003.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 433, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 433, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 433, + 258 + ], + "score": 1.0, + "content": "Randall J. LeVeque. Numerical Methods for Conservation Laws. Springer, 1992.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 501, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 504, + 276 + ], + "score": 1.0, + "content": "Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 275, + 467, + 287 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 467, + 287 + ], + "score": 1.0, + "content": "dynamics for manipulating rigid bodies, deformable objects, and fluids. In ICLR, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 293, + 503, + 317 + ], + "lines": [ + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "Junbang Liang, Ming C. Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse prob-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 305, + 391, + 317 + ], + "spans": [ + { + "bbox": [ + 115, + 305, + 391, + 317 + ], + "score": 1.0, + "content": "lems. In Advances in Neural Information Processing Systems, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 504, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 338 + ], + "score": 1.0, + "content": "Julia Ling, Andrew Kurzawski, and Jeremy Templeton. Reynolds averaged turbulence modelling", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 335, + 499, + 348 + ], + "spans": [ + { + "bbox": [ + 115, + 335, + 499, + 348 + ], + "score": 1.0, + "content": "using deep neural networks with embedded invariance. Journal of Fluid Mechanics, 807, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "score": 1.0, + "content": "Jacques Louis Lions. Optimal Control of Systems Governed by Partial Differential Equations.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 365, + 210, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 210, + 378 + ], + "score": 1.0, + "content": "Springer-Verlag, 1971.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 504, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "Zichao Long, Yiping Lu, Xianzhong Ma, and Bin Dong. PDE-net: Learning PDEs from data. In", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 394, + 171, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 171, + 408 + ], + "score": 1.0, + "content": "ICML, 2018.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 415, + 504, + 438 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 430 + ], + "score": 1.0, + "content": "Dougal Maclaurin, David Duvenaud, and Ryan P. Adams. Autograd: Effortless gradients in Numpy.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 426, + 227, + 439 + ], + "spans": [ + { + "bbox": [ + 116, + 426, + 227, + 439 + ], + "score": 1.0, + "content": "In ICML Workshops, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 445, + 504, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "Antoine McNamara, Adrien Treuille, Zoran Popovic, and Jos Stam. Fluid control using the adjoint ´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 457, + 364, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 457, + 364, + 468 + ], + "score": 1.0, + "content": "method. ACM Transactions on Graphics (TOG), 23(3), 2004.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 504, + 498 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "Arthur Mensch and Mathieu Blondel. Differentiable dynamic programming for structured prediction", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 487, + 239, + 498 + ], + "spans": [ + { + "bbox": [ + 116, + 487, + 239, + 498 + ], + "score": 1.0, + "content": "and attention. In ICML, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 505, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "Jeremy Morton, Antony Jameson, Mykel J Kochenderfer, and Freddie Witherden. Deep dynamical", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 515, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 505, + 531 + ], + "score": 1.0, + "content": "modeling and control of unsteady fluid flows. In Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 528, + 178, + 540 + ], + "spans": [ + { + "bbox": [ + 115, + 528, + 178, + 540 + ], + "score": 1.0, + "content": "Systems, 2018.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 546, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "Damian Mrowca, Chengxu Zhuang, Elias Wang, Nick Haber, Fei-Fei Li, Josh Tenenbaum, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "Daniel L. Yamins. Flexible neural representation for physics prediction. In Advances in Neural", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 569, + 274, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 569, + 274, + 581 + ], + "score": 1.0, + "content": "Information Processing Systems, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 115, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "high-performance deep learning library. In Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 620, + 144, + 634 + ], + "spans": [ + { + "bbox": [ + 115, + 620, + 144, + 634 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 503, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 639, + 504, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 504, + 653 + ], + "score": 1.0, + "content": "Olivier Pironneau. On optimum design in fluid mechanics. Journal of Fluid Mechanics, 64(1), 1974.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 479, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 479, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 479, + 673 + ], + "score": 1.0, + "content": "Lev Semenovich Pontryagin. Mathematical Theory of Optimal Processes. John Wiley, 1962.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 502, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 503, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 503, + 691 + ], + "score": 1.0, + "content": "Jovan Popovic, Steven M. Seitz, Michael A. Erdmann, Zoran Popovic, and Andrew P. Witkin. In-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 690, + 403, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 403, + 702 + ], + "score": 1.0, + "content": "teractive manipulation of rigid body simulations. In SIGGRAPH, 2000.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 109, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "William H Press, Saul A Teukolsky, William T Vetterling, and Brian P Flannery. Numerical Recipes.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 720, + 307, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 307, + 732 + ], + "score": 1.0, + "content": "Cambridge University Press, 3rd edition, 2007.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Antony Jameson. Aerodynamic design via control theory. Journal of Scientific Computing, 3(3),", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 142, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 142, + 105 + ], + "score": 1.0, + "content": "1988.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 112, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 126 + ], + "score": 1.0, + "content": "Dieter Kraft. On converting optimal control problems into nonlinear programming problems. In", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 123, + 322, + 136 + ], + "spans": [ + { + "bbox": [ + 117, + 123, + 322, + 136 + ], + "score": 1.0, + "content": "Computational Mathematical Programming. 1985.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 506, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 142, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "Lubor Ladicky, SoHyeon Jeong, Barbara Solenthaler, Marc Pollefeys, and Markus Gross. Data-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 506, + 168 + ], + "score": 1.0, + "content": "driven fluid simulations using regression forests. ACM Transactions on Graphics (TOG), 34(6),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 143, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 143, + 177 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 142, + 506, + 177 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 183, + 376, + 196 + ], + "lines": [ + { + "bbox": [ + 105, + 183, + 377, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 377, + 198 + ], + "score": 1.0, + "content": "Horace Lamb. Hydrodynamics. Cambridge University Press, 1932.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 183, + 377, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 203, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 504, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 504, + 216 + ], + "score": 1.0, + "content": "Daniel B Leineweber, Irene Bauer, Hans Georg Bock, and Johannes P Schloder. An efficient mul-¨", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 116, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "tiple shooting based reduced SQP strategy for large-scale dynamic process optimization. Part 1:", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 225, + 402, + 239 + ], + "spans": [ + { + "bbox": [ + 116, + 225, + 402, + 239 + ], + "score": 1.0, + "content": "Theoretical aspects. Computers & Chemical Engineering, 27(2), 2003.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 203, + 505, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 433, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 433, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 433, + 258 + ], + "score": 1.0, + "content": "Randall J. LeVeque. Numerical Methods for Conservation Laws. Springer, 1992.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 243, + 433, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 501, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 504, + 276 + ], + "score": 1.0, + "content": "Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 275, + 467, + 287 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 467, + 287 + ], + "score": 1.0, + "content": "dynamics for manipulating rigid bodies, deformable objects, and fluids. In ICLR, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 263, + 504, + 287 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 293, + 503, + 317 + ], + "lines": [ + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "Junbang Liang, Ming C. Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse prob-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 305, + 391, + 317 + ], + "spans": [ + { + "bbox": [ + 115, + 305, + 391, + 317 + ], + "score": 1.0, + "content": "lems. In Advances in Neural Information Processing Systems, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 294, + 505, + 317 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 504, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 338 + ], + "score": 1.0, + "content": "Julia Ling, Andrew Kurzawski, and Jeremy Templeton. Reynolds averaged turbulence modelling", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 335, + 499, + 348 + ], + "spans": [ + { + "bbox": [ + 115, + 335, + 499, + 348 + ], + "score": 1.0, + "content": "using deep neural networks with embedded invariance. Journal of Fluid Mechanics, 807, 2016.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 323, + 505, + 348 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "score": 1.0, + "content": "Jacques Louis Lions. Optimal Control of Systems Governed by Partial Differential Equations.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 365, + 210, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 210, + 378 + ], + "score": 1.0, + "content": "Springer-Verlag, 1971.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 354, + 505, + 378 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 504, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "Zichao Long, Yiping Lu, Xianzhong Ma, and Bin Dong. PDE-net: Learning PDEs from data. In", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 394, + 171, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 171, + 408 + ], + "score": 1.0, + "content": "ICML, 2018.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 384, + 506, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 415, + 504, + 438 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 430 + ], + "score": 1.0, + "content": "Dougal Maclaurin, David Duvenaud, and Ryan P. Adams. Autograd: Effortless gradients in Numpy.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 426, + 227, + 439 + ], + "spans": [ + { + "bbox": [ + 116, + 426, + 227, + 439 + ], + "score": 1.0, + "content": "In ICML Workshops, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 412, + 505, + 439 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 445, + 504, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "Antoine McNamara, Adrien Treuille, Zoran Popovic, and Jos Stam. Fluid control using the adjoint ´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 457, + 364, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 457, + 364, + 468 + ], + "score": 1.0, + "content": "method. ACM Transactions on Graphics (TOG), 23(3), 2004.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 444, + 505, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 504, + 498 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "Arthur Mensch and Mathieu Blondel. Differentiable dynamic programming for structured prediction", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 487, + 239, + 498 + ], + "spans": [ + { + "bbox": [ + 116, + 487, + 239, + 498 + ], + "score": 1.0, + "content": "and attention. In ICML, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 476, + 505, + 498 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 505, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "Jeremy Morton, Antony Jameson, Mykel J Kochenderfer, and Freddie Witherden. Deep dynamical", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 515, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 505, + 531 + ], + "score": 1.0, + "content": "modeling and control of unsteady fluid flows. In Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 528, + 178, + 540 + ], + "spans": [ + { + "bbox": [ + 115, + 528, + 178, + 540 + ], + "score": 1.0, + "content": "Systems, 2018.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 506, + 505, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 546, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "Damian Mrowca, Chengxu Zhuang, Elias Wang, Nick Haber, Fei-Fei Li, Josh Tenenbaum, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "Daniel L. Yamins. Flexible neural representation for physics prediction. In Advances in Neural", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 569, + 274, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 569, + 274, + 581 + ], + "score": 1.0, + "content": "Information Processing Systems, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 547, + 505, + 581 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 115, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 115, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "high-performance deep learning library. In Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 620, + 144, + 634 + ], + "spans": [ + { + "bbox": [ + 115, + 620, + 144, + 634 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 588, + 506, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 503, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 639, + 504, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 504, + 653 + ], + "score": 1.0, + "content": "Olivier Pironneau. On optimum design in fluid mechanics. Journal of Fluid Mechanics, 64(1), 1974.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 639, + 504, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 479, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 479, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 479, + 673 + ], + "score": 1.0, + "content": "Lev Semenovich Pontryagin. Mathematical Theory of Optimal Processes. John Wiley, 1962.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 659, + 479, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 678, + 502, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 503, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 503, + 691 + ], + "score": 1.0, + "content": "Jovan Popovic, Steven M. Seitz, Michael A. Erdmann, Zoran Popovic, and Andrew P. Witkin. In-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 690, + 403, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 403, + 702 + ], + "score": 1.0, + "content": "teractive manipulation of rigid body simulations. In SIGGRAPH, 2000.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 680, + 503, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "William H Press, Saul A Teukolsky, William T Vetterling, and Brian P Flannery. Numerical Recipes.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 720, + 307, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 307, + 732 + ], + "score": 1.0, + "content": "Cambridge University Press, 3rd edition, 2007.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Maziar Raissi, Alireza Yazdani, and George Karniadakis. Hidden fluid mechanics: A Navier-Stokes", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "informed deep learning framework for assimilating flow visualization data. arXiv:1808.04327,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 103, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 143, + 117 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 136 + ], + "score": 1.0, + "content": "Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedi-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 505, + 146 + ], + "score": 1.0, + "content": "cal image segmentation. In Medical Image Computing and Computer-Assisted Intervention, 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 503, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 504, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 504, + 166 + ], + "score": 1.0, + "content": "Connor Schenck and Dieter Fox. SPNets: Differentiable fluid dynamics for deep neural networks.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 280, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 280, + 176 + ], + "score": 1.0, + "content": "In Conference on Robot Learning, 2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 503, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Gordon D Smith. Numerical Solution of Partial Differential Equations: Finite Difference Methods.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 194, + 243, + 206 + ], + "spans": [ + { + "bbox": [ + 116, + 194, + 243, + 206 + ], + "score": 1.0, + "content": "Oxford University Press, 1985.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 212, + 294, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 295, + 225 + ], + "score": 1.0, + "content": "Jos Stam. Stable Fluids. In SIGGRAPH, 1999.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 102, + 231, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "Edriss S Titi. On approximate inertial manifolds to the Navier-Stokes equations. Journal of Mathe-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 243, + 314, + 255 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 314, + 255 + ], + "score": 1.0, + "content": "matical Analysis and Applications, 149(2), 1990.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "Seiya Tokui, Ryosuke Okuta, Takuya Akiba, Yusuke Niitani, Toru Ogawa, Shunta Saito, Shuji", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 272, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 504, + 284 + ], + "score": 1.0, + "content": "Suzuki, Kota Uenishi, Brian Vogel, and Hiroyuki Yamazaki Vincent. Chainer: A deep learn-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 115, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "ing framework for accelerating the research cycle. In Conference on Knowledge Discovery and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 295, + 228, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 295, + 228, + 307 + ], + "score": 1.0, + "content": "Data Mining (KDD), 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "Jonathan Tompson, Kristofer Schlachter, Pablo Sprechmann, and Ken Perlin. Accelerating Eulerian", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 325, + 365, + 336 + ], + "spans": [ + { + "bbox": [ + 116, + 325, + 365, + 336 + ], + "score": 1.0, + "content": "fluid simulation with convolutional networks. In ICML, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 504, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "Marc Toussaint, Kelsey Allen, Kevin Smith, and Joshua B Tenenbaum. Differentiable physics and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 355, + 495, + 367 + ], + "spans": [ + { + "bbox": [ + 117, + 355, + 495, + 367 + ], + "score": 1.0, + "content": "stable modes for tool-use and manipulation planning. In Robotics: Science and Systems, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 372, + 504, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "Adrien Treuille, Antoine McNamara, Zoran Popovic, and Jos Stam. Keyframe control of smoke ´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 385, + 380, + 397 + ], + "spans": [ + { + "bbox": [ + 116, + 385, + 380, + 397 + ], + "score": 1.0, + "content": "simulations. ACM Transactions on Graphics (TOG), 22(3), 2003.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 504, + 426 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "Adrien Treuille, Andrew Lewis, and Zoran Popovic. Model reduction for real-time fluids. ´ ACM", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 414, + 304, + 426 + ], + "spans": [ + { + "bbox": [ + 117, + 414, + 304, + 426 + ], + "score": 1.0, + "content": "Transactions on Graphics (TOG), 25(3), 2006.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 504, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "Benjamin Ummenhofer, Lukas Prantl, Nils Thurey, and Vladlen Koltun. Lagrangian fluid simulation ¨", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 444, + 302, + 456 + ], + "spans": [ + { + "bbox": [ + 117, + 444, + 302, + 456 + ], + "score": 1.0, + "content": "with continuous convolutions. In ICLR, 2020.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "Bart van Merrienboer, Dan Moldovan, and Alexander Wiltschko. Tangent: Automatic differentiation ¨", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 473, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 115, + 473, + 505, + 487 + ], + "score": 1.0, + "content": "using source-code transformation for dynamically typed array programming. In Advances in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 484, + 304, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 484, + 304, + 497 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 503, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 517 + ], + "score": 1.0, + "content": "Hermann von Helmholtz. Uber integrale der hydrodynamischen gleichungen, welcher der wirbel- ¨", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 515, + 416, + 527 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 416, + 527 + ], + "score": 1.0, + "content": "bewegungen entsprechen. Journal fur Mathematik Bd. LV. Heft ¨ , 1:4, 1858.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 503, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "Paul J Werbos. Backwards differentiation in AD and neural nets: Past links and new opportunities.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 544, + 476, + 557 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 476, + 557 + ], + "score": 1.0, + "content": "In Automatic Differentiation: Applications, Theory, and Implementations. Springer, 2006.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 563, + 503, + 586 + ], + "lines": [ + { + "bbox": [ + 106, + 562, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 577 + ], + "score": 1.0, + "content": "You Xie, Erik Franz, Mengyu Chu, and Nils Thuerey. tempoGAN: A temporally coherent, volumet-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 575, + 490, + 588 + ], + "spans": [ + { + "bbox": [ + 115, + 575, + 490, + 588 + ], + "score": 1.0, + "content": "ric GAN for super-resolution fluid flow. ACM Transactions on Graphics (TOG), 37(4), 2018.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 104, + 593, + 500, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 593, + 499, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 499, + 606 + ], + "score": 1.0, + "content": "Kemin Zhou, John C. Doyle, and Keith Glover. Robust and Optimal Control. Prentice Hall, 1996.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Maziar Raissi, Alireza Yazdani, and George Karniadakis. Hidden fluid mechanics: A Navier-Stokes", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "informed deep learning framework for assimilating flow visualization data. arXiv:1808.04327,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 103, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 143, + 117 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 506, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 136 + ], + "score": 1.0, + "content": "Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedi-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 505, + 146 + ], + "score": 1.0, + "content": "cal image segmentation. In Medical Image Computing and Computer-Assisted Intervention, 2015.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 123, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 503, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 504, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 504, + 166 + ], + "score": 1.0, + "content": "Connor Schenck and Dieter Fox. SPNets: Differentiable fluid dynamics for deep neural networks.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 280, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 280, + 176 + ], + "score": 1.0, + "content": "In Conference on Robot Learning, 2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 152, + 504, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 503, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Gordon D Smith. Numerical Solution of Partial Differential Equations: Finite Difference Methods.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 194, + 243, + 206 + ], + "spans": [ + { + "bbox": [ + 116, + 194, + 243, + 206 + ], + "score": 1.0, + "content": "Oxford University Press, 1985.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 182, + 505, + 206 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 212, + 294, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 295, + 225 + ], + "score": 1.0, + "content": "Jos Stam. Stable Fluids. In SIGGRAPH, 1999.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 212, + 295, + 225 + ] + }, + { + "type": "text", + "bbox": [ + 102, + 231, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "Edriss S Titi. On approximate inertial manifolds to the Navier-Stokes equations. Journal of Mathe-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 243, + 314, + 255 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 314, + 255 + ], + "score": 1.0, + "content": "matical Analysis and Applications, 149(2), 1990.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 231, + 505, + 255 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "Seiya Tokui, Ryosuke Okuta, Takuya Akiba, Yusuke Niitani, Toru Ogawa, Shunta Saito, Shuji", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 272, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 115, + 272, + 504, + 284 + ], + "score": 1.0, + "content": "Suzuki, Kota Uenishi, Brian Vogel, and Hiroyuki Yamazaki Vincent. Chainer: A deep learn-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 115, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "ing framework for accelerating the research cycle. In Conference on Knowledge Discovery and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 295, + 228, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 295, + 228, + 307 + ], + "score": 1.0, + "content": "Data Mining (KDD), 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 261, + 506, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "Jonathan Tompson, Kristofer Schlachter, Pablo Sprechmann, and Ken Perlin. Accelerating Eulerian", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 325, + 365, + 336 + ], + "spans": [ + { + "bbox": [ + 116, + 325, + 365, + 336 + ], + "score": 1.0, + "content": "fluid simulation with convolutional networks. In ICML, 2017.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 313, + 505, + 336 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 504, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "Marc Toussaint, Kelsey Allen, Kevin Smith, and Joshua B Tenenbaum. Differentiable physics and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 355, + 495, + 367 + ], + "spans": [ + { + "bbox": [ + 117, + 355, + 495, + 367 + ], + "score": 1.0, + "content": "stable modes for tool-use and manipulation planning. In Robotics: Science and Systems, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 342, + 505, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 372, + 504, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "Adrien Treuille, Antoine McNamara, Zoran Popovic, and Jos Stam. Keyframe control of smoke ´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 385, + 380, + 397 + ], + "spans": [ + { + "bbox": [ + 116, + 385, + 380, + 397 + ], + "score": 1.0, + "content": "simulations. ACM Transactions on Graphics (TOG), 22(3), 2003.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 373, + 505, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 504, + 426 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "Adrien Treuille, Andrew Lewis, and Zoran Popovic. Model reduction for real-time fluids. ´ ACM", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 414, + 304, + 426 + ], + "spans": [ + { + "bbox": [ + 117, + 414, + 304, + 426 + ], + "score": 1.0, + "content": "Transactions on Graphics (TOG), 25(3), 2006.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 403, + 505, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 504, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "Benjamin Ummenhofer, Lukas Prantl, Nils Thurey, and Vladlen Koltun. Lagrangian fluid simulation ¨", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 444, + 302, + 456 + ], + "spans": [ + { + "bbox": [ + 117, + 444, + 302, + 456 + ], + "score": 1.0, + "content": "with continuous convolutions. In ICLR, 2020.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 433, + 505, + 456 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "Bart van Merrienboer, Dan Moldovan, and Alexander Wiltschko. Tangent: Automatic differentiation ¨", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 473, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 115, + 473, + 505, + 487 + ], + "score": 1.0, + "content": "using source-code transformation for dynamically typed array programming. In Advances in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 484, + 304, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 484, + 304, + 497 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 2018.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 462, + 505, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 503, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 517 + ], + "score": 1.0, + "content": "Hermann von Helmholtz. Uber integrale der hydrodynamischen gleichungen, welcher der wirbel- ¨", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 515, + 416, + 527 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 416, + 527 + ], + "score": 1.0, + "content": "bewegungen entsprechen. Journal fur Mathematik Bd. LV. Heft ¨ , 1:4, 1858.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 503, + 505, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 503, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "Paul J Werbos. Backwards differentiation in AD and neural nets: Past links and new opportunities.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 544, + 476, + 557 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 476, + 557 + ], + "score": 1.0, + "content": "In Automatic Differentiation: Applications, Theory, and Implementations. Springer, 2006.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 533, + 505, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 563, + 503, + 586 + ], + "lines": [ + { + "bbox": [ + 106, + 562, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 577 + ], + "score": 1.0, + "content": "You Xie, Erik Franz, Mengyu Chu, and Nils Thuerey. tempoGAN: A temporally coherent, volumet-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 575, + 490, + 588 + ], + "spans": [ + { + "bbox": [ + 115, + 575, + 490, + 588 + ], + "score": 1.0, + "content": "ric GAN for super-resolution fluid flow. ACM Transactions on Graphics (TOG), 37(4), 2018.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 562, + 505, + 588 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 593, + 500, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 593, + 499, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 499, + 606 + ], + "score": 1.0, + "content": "Kemin Zhou, John C. Doyle, and Keith Glover. Robust and Optimal Control. Prentice Hall, 1996.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 593, + 499, + 606 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 105, + 503, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "Our solver is publicly available at https://github.com/tum-pbs/PhiFlow, licensed as MIT. It is imple-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "mented via existing machine learning frameworks to benefit from the built-in automatic differentia-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 127, + 335, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 335, + 140 + ], + "score": 1.0, + "content": "tion and to enable tight integration with neural networks.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "For the experiments shown here we used the popular machine learning framework Tensor-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "Flow (Abadi et al., 2016). However, our solver is written in a framework-independent way and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 167, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 178 + ], + "score": 1.0, + "content": "also supports PyTorch (Paszke et al., 2019). Both frameworks allow for a low-level NumPy-like", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "implementation which is well suited for basic PDE building blocks. The following paragraphs out-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 188, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 201 + ], + "score": 1.0, + "content": "line how we implemented these building blocks and how they can be put together to solve the PDEs", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 200, + 188, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 188, + 211 + ], + "score": 1.0, + "content": "shown in Section 6.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 216, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 505, + 228 + ], + "score": 1.0, + "content": "Staggered grids. Many of the experiments presented in Section 6 use PDEs which track velocities.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "We adopt the marker-and-cell method (Harlow & Welch, 1965), storing densities in a regular grid", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 239, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 504, + 251 + ], + "score": 1.0, + "content": "and velocity in a staggered grid. Unlike regular grids, where all components are sampled at the cen-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "ters of grid cells, staggered grids sample vector fields in a staggered form. Each vector component", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 260, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 272 + ], + "score": 1.0, + "content": "is sampled in the center of the cell face perpendicular to that direction. This sampling allows for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "an exact formulation of the divergence of a staggered vector field, decreasing discretization errors", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "in many cases. On the other hand, it complicates operations that combine vector fields with regular", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 316, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 316, + 306 + ], + "score": 1.0, + "content": "fields such as transport or density-dependent forces.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 309, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 106, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 505, + 323 + ], + "score": 1.0, + "content": "We use staggered grids for the velocities in all of our experiments. The buoyancy operation, which", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 320, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 334 + ], + "score": 1.0, + "content": "applies an upward force proportional to the smoke density, interpolates the density to the staggered", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "grid. For the transport, also called advection, of regular or staggered fields, we interpolate the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "staggered field to grid cell centers or face centers, respectively. These interpolations are implemented", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "in TensorFlow using basic tensor operations, similar to the differential operators. We implemented", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 365, + 439, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 439, + 378 + ], + "score": 1.0, + "content": "all differential operators that act on vector fields to support staggered grids as well.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 105, + 381, + 503, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Differential operators. For the experiments outlined in this paper, we have implemented the fol-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 392, + 225, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 225, + 405 + ], + "score": 1.0, + "content": "lowing differential operators:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 132, + 412, + 486, + 468 + ], + "lines": [ + { + "bbox": [ + 132, + 412, + 377, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 361, + 424 + ], + "score": 1.0, + "content": "• Gradient of scalar fields in any number of dimensions,", + "type": "text" + }, + { + "bbox": [ + 361, + 413, + 377, + 423 + ], + "score": 0.79, + "content": "\\nabla x", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 426, + 485, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 461, + 439 + ], + "score": 1.0, + "content": "• Divergence of regular and staggered vector fields in any number of dimensions,", + "type": "text" + }, + { + "bbox": [ + 462, + 428, + 485, + 437 + ], + "score": 0.86, + "content": "\\nabla \\cdot \\pmb { x }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 133, + 441, + 353, + 453 + ], + "spans": [ + { + "bbox": [ + 133, + 441, + 324, + 453 + ], + "score": 1.0, + "content": "• Curl of staggered vector fields in 2D and 3D,", + "type": "text" + }, + { + "bbox": [ + 324, + 441, + 353, + 452 + ], + "score": 0.86, + "content": "\\nabla \\times \\mathbf x", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 131, + 455, + 378, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 358, + 467 + ], + "score": 1.0, + "content": "• Laplace of scalar fields in any number of dimensions,", + "type": "text" + }, + { + "bbox": [ + 358, + 455, + 378, + 466 + ], + "score": 0.87, + "content": "\\nabla ^ { 2 } x", + "type": "inline_equation" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "All differential operators are local operations, i.e. they only act on a small neighbourhood of grid", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 487, + 504, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 504, + 498 + ], + "score": 1.0, + "content": "points. In the context of machine learning, this suggests implementing them as convolution opera-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "tions with a fixed kernel. Indeed, all differential operators can be expressed this way and we have", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 398, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 398, + 520 + ], + "score": 1.0, + "content": "implemented some low-dimensional versions of them using this method.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "This method does, however, scale poorly with the dimensionality of the physical system as the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "convolutional kernels pick up a large number of zeros, thus wasting computations. Therefore, we", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 547, + 461, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 461, + 560 + ], + "score": 1.0, + "content": "express n-dimensional differential operators using basic mathematical tensor operations.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 501, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 502, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 502, + 577 + ], + "score": 1.0, + "content": "Consider the gradient computation in 1D, which results in a staggered grid. Each resulting value is", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "interline_equation", + "bbox": [ + 263, + 578, + 346, + 592 + ], + "lines": [ + { + "bbox": [ + 263, + 578, + 346, + 592 + ], + "spans": [ + { + "bbox": [ + 263, + 578, + 346, + 592 + ], + "score": 0.92, + "content": "( \\nabla x ) _ { i } = x _ { i } - x _ { i - 1 } ,", + "type": "interline_equation", + "image_path": "2ce2e36a34504a5bdd2cbc30eab55fa5c90128feea383bb029b31c0326678ce8.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 263, + 578, + 346, + 592 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "score": 1.0, + "content": "assuming the result is staggered at the lower faces of each grid cell. This operation can be imple-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 271, + 619 + ], + "score": 1.0, + "content": "mented as a 1D convolution with kernel", + "type": "text" + }, + { + "bbox": [ + 271, + 606, + 302, + 618 + ], + "score": 0.92, + "content": "( - 1 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "or as a vector operation which subtracts the array", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 616, + 249, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 249, + 629 + ], + "score": 1.0, + "content": "from itself, shifted by one element.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 504, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "In a low-dimensional setting, the convolution operation will be faster as it is highly optimized and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "can be executed on GPUs with one call. In higher dimensions, however, the vector-based version", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "is faster and more practical because it avoids unnecessary computations and can be coded in a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "dimension-independent fashion. Both convolutions and basic mathematical operations are supported", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "by all common machine learning frameworks, eliminating the need to implement custom gradient", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 149, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 149, + 701 + ], + "score": 1.0, + "content": "functions.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 106, + 705, + 457, + 717 + ], + "lines": [ + { + "bbox": [ + 105, + 703, + 458, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 458, + 720 + ], + "score": 1.0, + "content": "Advection. PDEs containing material derivatives can be solved using an advection step", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "interline_equation", + "bbox": [ + 267, + 720, + 343, + 733 + ], + "lines": [ + { + "bbox": [ + 267, + 720, + 343, + 733 + ], + "spans": [ + { + "bbox": [ + 267, + 720, + 343, + 733 + ], + "score": 0.91, + "content": "f \\gets \\mathrm { A d v e c t } [ f , { \\boldsymbol v } ]", + "type": "interline_equation", + "image_path": "c13fe980bfb2328b82580e6fc7b8bc7ead642e783397cf8755f72018c9b4cd42.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 267, + 720, + 343, + 733 + ], + "spans": [], + "index": 48 + } + ] + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 81, + 468, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 470, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 470, + 96 + ], + "score": 1.0, + "content": "A IMPLEMENTATION DETAILS OF THE DIFFERENTIABLE PDE SOLVER", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 105, + 503, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "Our solver is publicly available at https://github.com/tum-pbs/PhiFlow, licensed as MIT. It is imple-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "mented via existing machine learning frameworks to benefit from the built-in automatic differentia-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 127, + 335, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 335, + 140 + ], + "score": 1.0, + "content": "tion and to enable tight integration with neural networks.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 104, + 505, + 140 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 210 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "For the experiments shown here we used the popular machine learning framework Tensor-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "Flow (Abadi et al., 2016). However, our solver is written in a framework-independent way and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 167, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 505, + 178 + ], + "score": 1.0, + "content": "also supports PyTorch (Paszke et al., 2019). Both frameworks allow for a low-level NumPy-like", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "implementation which is well suited for basic PDE building blocks. The following paragraphs out-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 188, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 201 + ], + "score": 1.0, + "content": "line how we implemented these building blocks and how they can be put together to solve the PDEs", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 200, + 188, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 188, + 211 + ], + "score": 1.0, + "content": "shown in Section 6.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 145, + 505, + 211 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 216, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 505, + 228 + ], + "score": 1.0, + "content": "Staggered grids. Many of the experiments presented in Section 6 use PDEs which track velocities.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "We adopt the marker-and-cell method (Harlow & Welch, 1965), storing densities in a regular grid", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 239, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 504, + 251 + ], + "score": 1.0, + "content": "and velocity in a staggered grid. Unlike regular grids, where all components are sampled at the cen-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "ters of grid cells, staggered grids sample vector fields in a staggered form. Each vector component", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 260, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 272 + ], + "score": 1.0, + "content": "is sampled in the center of the cell face perpendicular to that direction. This sampling allows for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "an exact formulation of the divergence of a staggered vector field, decreasing discretization errors", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 294 + ], + "score": 1.0, + "content": "in many cases. On the other hand, it complicates operations that combine vector fields with regular", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 316, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 316, + 306 + ], + "score": 1.0, + "content": "fields such as transport or density-dependent forces.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 217, + 505, + 306 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 309, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 106, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 505, + 323 + ], + "score": 1.0, + "content": "We use staggered grids for the velocities in all of our experiments. The buoyancy operation, which", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 320, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 334 + ], + "score": 1.0, + "content": "applies an upward force proportional to the smoke density, interpolates the density to the staggered", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "grid. For the transport, also called advection, of regular or staggered fields, we interpolate the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "staggered field to grid cell centers or face centers, respectively. These interpolations are implemented", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "in TensorFlow using basic tensor operations, similar to the differential operators. We implemented", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 365, + 439, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 439, + 378 + ], + "score": 1.0, + "content": "all differential operators that act on vector fields to support staggered grids as well.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 310, + 506, + 378 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 381, + 503, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "Differential operators. For the experiments outlined in this paper, we have implemented the fol-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 392, + 225, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 225, + 405 + ], + "score": 1.0, + "content": "lowing differential operators:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 381, + 505, + 405 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 412, + 486, + 468 + ], + "lines": [ + { + "bbox": [ + 132, + 412, + 377, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 361, + 424 + ], + "score": 1.0, + "content": "• Gradient of scalar fields in any number of dimensions,", + "type": "text" + }, + { + "bbox": [ + 361, + 413, + 377, + 423 + ], + "score": 0.79, + "content": "\\nabla x", + "type": "inline_equation" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 132, + 426, + 485, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 461, + 439 + ], + "score": 1.0, + "content": "• Divergence of regular and staggered vector fields in any number of dimensions,", + "type": "text" + }, + { + "bbox": [ + 462, + 428, + 485, + 437 + ], + "score": 0.86, + "content": "\\nabla \\cdot \\pmb { x }", + "type": "inline_equation" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 133, + 441, + 353, + 453 + ], + "spans": [ + { + "bbox": [ + 133, + 441, + 324, + 453 + ], + "score": 1.0, + "content": "• Curl of staggered vector fields in 2D and 3D,", + "type": "text" + }, + { + "bbox": [ + 324, + 441, + 353, + 452 + ], + "score": 0.86, + "content": "\\nabla \\times \\mathbf x", + "type": "inline_equation" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 131, + 455, + 378, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 358, + 467 + ], + "score": 1.0, + "content": "• Laplace of scalar fields in any number of dimensions,", + "type": "text" + }, + { + "bbox": [ + 358, + 455, + 378, + 466 + ], + "score": 0.87, + "content": "\\nabla ^ { 2 } x", + "type": "inline_equation" + } + ], + "index": 28, + "is_list_start_line": true + } + ], + "index": 26.5, + "bbox_fs": [ + 131, + 412, + 485, + 467 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "All differential operators are local operations, i.e. they only act on a small neighbourhood of grid", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 487, + 504, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 504, + 498 + ], + "score": 1.0, + "content": "points. In the context of machine learning, this suggests implementing them as convolution opera-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "tions with a fixed kernel. Indeed, all differential operators can be expressed this way and we have", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 398, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 398, + 520 + ], + "score": 1.0, + "content": "implemented some low-dimensional versions of them using this method.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 475, + 506, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 505, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "This method does, however, scale poorly with the dimensionality of the physical system as the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "convolutional kernels pick up a large number of zeros, thus wasting computations. Therefore, we", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 547, + 461, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 461, + 560 + ], + "score": 1.0, + "content": "express n-dimensional differential operators using basic mathematical tensor operations.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 525, + 505, + 560 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 501, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 502, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 502, + 577 + ], + "score": 1.0, + "content": "Consider the gradient computation in 1D, which results in a staggered grid. Each resulting value is", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 563, + 502, + 577 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 263, + 578, + 346, + 592 + ], + "lines": [ + { + "bbox": [ + 263, + 578, + 346, + 592 + ], + "spans": [ + { + "bbox": [ + 263, + 578, + 346, + 592 + ], + "score": 0.92, + "content": "( \\nabla x ) _ { i } = x _ { i } - x _ { i - 1 } ,", + "type": "interline_equation", + "image_path": "2ce2e36a34504a5bdd2cbc30eab55fa5c90128feea383bb029b31c0326678ce8.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 263, + 578, + 346, + 592 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 628 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "score": 1.0, + "content": "assuming the result is staggered at the lower faces of each grid cell. This operation can be imple-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 271, + 619 + ], + "score": 1.0, + "content": "mented as a 1D convolution with kernel", + "type": "text" + }, + { + "bbox": [ + 271, + 606, + 302, + 618 + ], + "score": 0.92, + "content": "( - 1 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "or as a vector operation which subtracts the array", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 616, + 249, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 249, + 629 + ], + "score": 1.0, + "content": "from itself, shifted by one element.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 594, + 505, + 629 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 504, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "In a low-dimensional setting, the convolution operation will be faster as it is highly optimized and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "can be executed on GPUs with one call. In higher dimensions, however, the vector-based version", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "is faster and more practical because it avoids unnecessary computations and can be coded in a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "dimension-independent fashion. Both convolutions and basic mathematical operations are supported", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "by all common machine learning frameworks, eliminating the need to implement custom gradient", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 149, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 149, + 701 + ], + "score": 1.0, + "content": "functions.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 104, + 633, + 506, + 701 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 705, + 457, + 717 + ], + "lines": [ + { + "bbox": [ + 105, + 703, + 458, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 458, + 720 + ], + "score": 1.0, + "content": "Advection. PDEs containing material derivatives can be solved using an advection step", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 703, + 458, + 720 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 267, + 720, + 343, + 733 + ], + "lines": [ + { + "bbox": [ + 267, + 720, + 343, + 733 + ], + "spans": [ + { + "bbox": [ + 267, + 720, + 343, + 733 + ], + "score": 0.91, + "content": "f \\gets \\mathrm { A d v e c t } [ f , { \\boldsymbol v } ]", + "type": "interline_equation", + "image_path": "c13fe980bfb2328b82580e6fc7b8bc7ead642e783397cf8755f72018c9b4cd42.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 267, + 720, + 343, + 733 + ], + "spans": [], + "index": 48 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 243, + 95 + ], + "score": 1.0, + "content": "which moves each value of a field", + "type": "text" + }, + { + "bbox": [ + 243, + 83, + 250, + 94 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 82, + 418, + 95 + ], + "score": 1.0, + "content": "in the direction specified by a vector field", + "type": "text" + }, + { + "bbox": [ + 418, + 84, + 425, + 92 + ], + "score": 0.71, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 82, + 504, + 95 + ], + "score": 1.0, + "content": ". We implement the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "advection with semi-Lagrangian step (Stam, 1999) that looks back in time and supports regular and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 201, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 201, + 117 + ], + "score": 1.0, + "content": "staggered vector fields.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 121, + 504, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 317, + 134 + ], + "score": 1.0, + "content": "To determine the advected value of a grid cell or face", + "type": "text" + }, + { + "bbox": [ + 318, + 123, + 341, + 133 + ], + "score": 0.87, + "content": "x _ { \\mathrm { { t a r g e t } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 120, + 362, + 134 + ], + "score": 1.0, + "content": ", first", + "type": "text" + }, + { + "bbox": [ + 362, + 124, + 369, + 131 + ], + "score": 0.77, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "is interpolated to that point. Then", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 412, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 412, + 145 + ], + "score": 1.0, + "content": "the origin location is determined by following the vector backwards in time,", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 150, + 370, + 164 + ], + "lines": [ + { + "bbox": [ + 240, + 150, + 370, + 164 + ], + "spans": [ + { + "bbox": [ + 240, + 150, + 370, + 164 + ], + "score": 0.91, + "content": "x _ { \\mathrm { o r i g i n } } = x _ { \\mathrm { t a r g e t } } - \\Delta t \\cdot \\pmb { v } ( x _ { \\mathrm { t a r g e t } } ) .", + "type": "interline_equation", + "image_path": "b4a62133ea695497bd865af56fd101fb9fd35d1853b67f9426dd72d454498825.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 240, + 150, + 370, + 164 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 171, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "The final value is determined by linearly interpolating between the neighbouring grid cells around", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 107, + 183, + 130, + 194 + ], + "score": 0.83, + "content": "x _ { \\mathrm { o r i g i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 181, + 506, + 195 + ], + "score": 1.0, + "content": ". All of these operations are, again, implemented using basic mathematical operations. Hence,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 193, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 287, + 205 + ], + "score": 1.0, + "content": "gradients can be provided by the framework.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Poisson problems. Incompressible fluids, governed by the Navier-Stokes equations, are subject to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 187, + 233 + ], + "score": 1.0, + "content": "the hard constraints", + "type": "text" + }, + { + "bbox": [ + 187, + 221, + 227, + 231 + ], + "score": 0.91, + "content": "\\nabla \\cdot \\boldsymbol { v } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 220, + 245, + 233 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 245, + 221, + 289, + 232 + ], + "score": 0.92, + "content": "\\nabla \\times p = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 220, + 316, + 233 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 317, + 223, + 324, + 232 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "denotes the pressure. A numerical solver can", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 207, + 244 + ], + "score": 1.0, + "content": "achieve this by finding a", + "type": "text" + }, + { + "bbox": [ + 207, + 234, + 213, + 243 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "such that these constraints are satisfied. This step is often called Chorin", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 243, + 504, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 254 + ], + "score": 1.0, + "content": "Projection, or Helmholtz decomposition, and is closely related to the fundamental theorem of vector", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 382, + 267 + ], + "score": 1.0, + "content": "calculus (von Helmholtz, 1858; Chorin, 1967). On a grid, solving for", + "type": "text" + }, + { + "bbox": [ + 383, + 255, + 389, + 265 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 252, + 505, + 267 + ], + "score": 1.0, + "content": "is equal to solving a Poisson", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 208, + 277 + ], + "score": 1.0, + "content": "problem, i.e. a system of", + "type": "text" + }, + { + "bbox": [ + 208, + 265, + 218, + 275 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 265, + 288, + 277 + ], + "score": 1.0, + "content": "linear equations,", + "type": "text" + }, + { + "bbox": [ + 288, + 265, + 336, + 276 + ], + "score": 0.93, + "content": "A p = \\nabla \\cdot { \\boldsymbol { u } }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 265, + 364, + 277 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 365, + 265, + 375, + 275 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "is the total number of grid cells.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 275, + 429, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 125, + 288 + ], + "score": 1.0, + "content": "The", + "type": "text" + }, + { + "bbox": [ + 125, + 276, + 159, + 288 + ], + "score": 0.9, + "content": "( N \\cdot N )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 275, + 189, + 288 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 189, + 276, + 198, + 285 + ], + "score": 0.77, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 275, + 429, + 288 + ], + "score": 1.0, + "content": "is sparse and its entries are located at predictable indices.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 292, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "score": 1.0, + "content": "We numerically solve this Poisson problem with a conjugate gradient (CG) algorithm (Golub &", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 294, + 316 + ], + "score": 1.0, + "content": "Van Loan, 2012) that iteratively approximates", + "type": "text" + }, + { + "bbox": [ + 295, + 306, + 301, + 315 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 303, + 505, + 316 + ], + "score": 1.0, + "content": ". Since hundreds of CG steps typically need to be", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 315, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 326 + ], + "score": 1.0, + "content": "performed for each Poisson solve, it is unfeasible to unroll this chain of iterations, and store all", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 235, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 235, + 338 + ], + "score": 1.0, + "content": "intermediate results in memory.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 504, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 504, + 354 + ], + "score": 1.0, + "content": "To ensure that the automatic differentiation chain is not broken, we instead solve the adjoint problem.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 286, + 365 + ], + "score": 1.0, + "content": "For the pressure solve operation, the matrix", + "type": "text" + }, + { + "bbox": [ + 286, + 354, + 295, + 363 + ], + "score": 0.77, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "is symmetric and positive-definite. This causes the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "adjoint problem to have the same mathematical form (McNamara et al., 2004) as the original prob-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 374, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 388 + ], + "score": 1.0, + "content": "lem. Therefore we implement the gradient for the pressure solve by performing a pressure solve", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "on the gradient. We believe that this is a good example of leveraging the methodology of adjoint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 396, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 411 + ], + "score": 1.0, + "content": "method optimizations (Giles & Pierce, 2000; Treuille et al., 2006; Pontryagin, 1962) within deep", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "learning. With this formalism we arrive at a differentiable solver framework that closely integrates", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 417, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 417, + 505, + 434 + ], + "score": 1.0, + "content": "numerical methods for forward problems with support for inverse problems such as deep learning", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 430, + 201, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 201, + 442 + ], + "score": 1.0, + "content": "via the adjoint method.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 446, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "Solving Burger’s equation and the Navier-Stokes equations. Using the basic building blocks out-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "lined above, solving the PDEs becomes relatively simple. Burger’s equation involves an advection", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 468, + 326, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 326, + 482 + ], + "score": 1.0, + "content": "and a diffusion term which we evaluate independently.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 487, + 374, + 501 + ], + "lines": [ + { + "bbox": [ + 236, + 487, + 374, + 501 + ], + "spans": [ + { + "bbox": [ + 236, + 487, + 374, + 501 + ], + "score": 0.89, + "content": "\\mathrm { S o l v e r } [ u ] = \\mathrm { D i f f u s e } [ \\mathrm { A d v e c t } [ u , u ] ]", + "type": "interline_equation", + "image_path": "3cccb68531c623ccd6844782acc91707a18cc33aa2b8edd6653df36a295b6ea7.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 236, + 487, + 374, + 501 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 504, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 259, + 520 + ], + "score": 1.0, + "content": "where we explicitly compute Diffuse", + "type": "text" + }, + { + "bbox": [ + 259, + 506, + 332, + 519 + ], + "score": 0.82, + "content": "[ \\boldsymbol { u } ] = \\boldsymbol { u } + \\nu \\nabla ^ { 2 } \\boldsymbol { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 506, + 394, + 520 + ], + "score": 1.0, + "content": "with viscosity", + "type": "text" + }, + { + "bbox": [ + 395, + 510, + 401, + 517 + ], + "score": 0.75, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 506, + 505, + 520 + ], + "score": 1.0, + "content": ". The advection is semi-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 518, + 307, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 307, + 531 + ], + "score": 1.0, + "content": "Lagrangian with back-tracing as described above.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 429, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 429, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 429, + 549 + ], + "score": 1.0, + "content": "Solving the Navier-Stokes equations, typically comprises of the following steps:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 132, + 556, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 132, + 556, + 318, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 243, + 570 + ], + "score": 1.0, + "content": "• Transporting the density,", + "type": "text" + }, + { + "bbox": [ + 244, + 556, + 318, + 569 + ], + "score": 0.66, + "content": "\\rho \\gets \\mathrm { A d v e c t } [ \\rho , \\pmb { v } ]", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 133, + 573, + 323, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 573, + 247, + 586 + ], + "score": 1.0, + "content": "• Transporting the velocity,", + "type": "text" + }, + { + "bbox": [ + 247, + 573, + 323, + 585 + ], + "score": 0.84, + "content": "\\pmb { v } \\mathrm { A d v e c t } [ \\pmb { v } , \\pmb { v } ]", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 132, + 588, + 320, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 291, + 602 + ], + "score": 1.0, + "content": "• Applying diffusion if the viscosity is", + "type": "text" + }, + { + "bbox": [ + 291, + 590, + 316, + 600 + ], + "score": 0.89, + "content": "\\nu > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 588, + 320, + 602 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 132, + 605, + 418, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 249, + 618 + ], + "score": 1.0, + "content": "• Applying buoyancy force,", + "type": "text" + }, + { + "bbox": [ + 249, + 606, + 309, + 617 + ], + "score": 0.93, + "content": "{ \\pmb v } { \\pmb v } - { \\pmb \\beta } \\cdot { \\boldsymbol \\rho }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 605, + 408, + 618 + ], + "score": 1.0, + "content": "with buoyancy direction", + "type": "text" + }, + { + "bbox": [ + 409, + 606, + 418, + 617 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 39 + }, + { + "bbox": [ + 131, + 621, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 374, + 636 + ], + "score": 1.0, + "content": "• Enforcing incompressibility by solving for the pressure,", + "type": "text" + }, + { + "bbox": [ + 375, + 622, + 480, + 634 + ], + "score": 0.91, + "content": "p \\operatorname { S o l v e } [ A p = \\nabla \\cdot \\pmb { v } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 621, + 505, + 636 + ], + "score": 1.0, + "content": ", then", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 633, + 197, + 645 + ], + "spans": [ + { + "bbox": [ + 142, + 633, + 197, + 645 + ], + "score": 0.71, + "content": "\\pmb { v } \\pmb { v } - \\nabla p", + "type": "inline_equation" + } + ], + "index": 41 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 429, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 653, + 430, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 430, + 667 + ], + "score": 1.0, + "content": "These steps are executed in this order to advance the simulation forward in time.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "title", + "bbox": [ + 107, + 684, + 329, + 697 + ], + "lines": [ + { + "bbox": [ + 104, + 682, + 330, + 699 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 330, + 699 + ], + "score": 1.0, + "content": "B COMPLEXITY OF EXECUTION SCHEMES", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 391, + 722 + ], + "score": 1.0, + "content": "The staggered execution scheme recursively splits a sequence of length", + "type": "text" + }, + { + "bbox": [ + 391, + 712, + 398, + 720 + ], + "score": 0.76, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "into smaller sequences, as", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 248, + 733 + ], + "score": 1.0, + "content": "depicted in Fig. 3b and Fig. 7a for", + "type": "text" + }, + { + "bbox": [ + 248, + 721, + 274, + 730 + ], + "score": 0.9, + "content": "n = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 720, + 505, + 733 + ], + "score": 1.0, + "content": ". With each level of recursion depth, the sequence length", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 243, + 95 + ], + "score": 1.0, + "content": "which moves each value of a field", + "type": "text" + }, + { + "bbox": [ + 243, + 83, + 250, + 94 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 82, + 418, + 95 + ], + "score": 1.0, + "content": "in the direction specified by a vector field", + "type": "text" + }, + { + "bbox": [ + 418, + 84, + 425, + 92 + ], + "score": 0.71, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 82, + 504, + 95 + ], + "score": 1.0, + "content": ". We implement the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "advection with semi-Lagrangian step (Stam, 1999) that looks back in time and supports regular and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 201, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 201, + 117 + ], + "score": 1.0, + "content": "staggered vector fields.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 121, + 504, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 317, + 134 + ], + "score": 1.0, + "content": "To determine the advected value of a grid cell or face", + "type": "text" + }, + { + "bbox": [ + 318, + 123, + 341, + 133 + ], + "score": 0.87, + "content": "x _ { \\mathrm { { t a r g e t } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 120, + 362, + 134 + ], + "score": 1.0, + "content": ", first", + "type": "text" + }, + { + "bbox": [ + 362, + 124, + 369, + 131 + ], + "score": 0.77, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "is interpolated to that point. Then", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 412, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 412, + 145 + ], + "score": 1.0, + "content": "the origin location is determined by following the vector backwards in time,", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 120, + 505, + 145 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 150, + 370, + 164 + ], + "lines": [ + { + "bbox": [ + 240, + 150, + 370, + 164 + ], + "spans": [ + { + "bbox": [ + 240, + 150, + 370, + 164 + ], + "score": 0.91, + "content": "x _ { \\mathrm { o r i g i n } } = x _ { \\mathrm { t a r g e t } } - \\Delta t \\cdot \\pmb { v } ( x _ { \\mathrm { t a r g e t } } ) .", + "type": "interline_equation", + "image_path": "b4a62133ea695497bd865af56fd101fb9fd35d1853b67f9426dd72d454498825.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 240, + 150, + 370, + 164 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 171, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "The final value is determined by linearly interpolating between the neighbouring grid cells around", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 107, + 183, + 130, + 194 + ], + "score": 0.83, + "content": "x _ { \\mathrm { o r i g i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 181, + 506, + 195 + ], + "score": 1.0, + "content": ". All of these operations are, again, implemented using basic mathematical operations. Hence,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 193, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 287, + 205 + ], + "score": 1.0, + "content": "gradients can be provided by the framework.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 170, + 506, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Poisson problems. Incompressible fluids, governed by the Navier-Stokes equations, are subject to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 187, + 233 + ], + "score": 1.0, + "content": "the hard constraints", + "type": "text" + }, + { + "bbox": [ + 187, + 221, + 227, + 231 + ], + "score": 0.91, + "content": "\\nabla \\cdot \\boldsymbol { v } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 220, + 245, + 233 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 245, + 221, + 289, + 232 + ], + "score": 0.92, + "content": "\\nabla \\times p = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 220, + 316, + 233 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 317, + 223, + 324, + 232 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "denotes the pressure. A numerical solver can", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 207, + 244 + ], + "score": 1.0, + "content": "achieve this by finding a", + "type": "text" + }, + { + "bbox": [ + 207, + 234, + 213, + 243 + ], + "score": 0.8, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 231, + 505, + 244 + ], + "score": 1.0, + "content": "such that these constraints are satisfied. This step is often called Chorin", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 243, + 504, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 254 + ], + "score": 1.0, + "content": "Projection, or Helmholtz decomposition, and is closely related to the fundamental theorem of vector", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 382, + 267 + ], + "score": 1.0, + "content": "calculus (von Helmholtz, 1858; Chorin, 1967). On a grid, solving for", + "type": "text" + }, + { + "bbox": [ + 383, + 255, + 389, + 265 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 252, + 505, + 267 + ], + "score": 1.0, + "content": "is equal to solving a Poisson", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 208, + 277 + ], + "score": 1.0, + "content": "problem, i.e. a system of", + "type": "text" + }, + { + "bbox": [ + 208, + 265, + 218, + 275 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 265, + 288, + 277 + ], + "score": 1.0, + "content": "linear equations,", + "type": "text" + }, + { + "bbox": [ + 288, + 265, + 336, + 276 + ], + "score": 0.93, + "content": "A p = \\nabla \\cdot { \\boldsymbol { u } }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 265, + 364, + 277 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 365, + 265, + 375, + 275 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "is the total number of grid cells.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 275, + 429, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 125, + 288 + ], + "score": 1.0, + "content": "The", + "type": "text" + }, + { + "bbox": [ + 125, + 276, + 159, + 288 + ], + "score": 0.9, + "content": "( N \\cdot N )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 275, + 189, + 288 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 189, + 276, + 198, + 285 + ], + "score": 0.77, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 275, + 429, + 288 + ], + "score": 1.0, + "content": "is sparse and its entries are located at predictable indices.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 209, + 505, + 288 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 292, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 305 + ], + "score": 1.0, + "content": "We numerically solve this Poisson problem with a conjugate gradient (CG) algorithm (Golub &", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 294, + 316 + ], + "score": 1.0, + "content": "Van Loan, 2012) that iteratively approximates", + "type": "text" + }, + { + "bbox": [ + 295, + 306, + 301, + 315 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 303, + 505, + 316 + ], + "score": 1.0, + "content": ". Since hundreds of CG steps typically need to be", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 315, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 326 + ], + "score": 1.0, + "content": "performed for each Poisson solve, it is unfeasible to unroll this chain of iterations, and store all", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 235, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 235, + 338 + ], + "score": 1.0, + "content": "intermediate results in memory.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 291, + 505, + 338 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 504, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 504, + 354 + ], + "score": 1.0, + "content": "To ensure that the automatic differentiation chain is not broken, we instead solve the adjoint problem.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 286, + 365 + ], + "score": 1.0, + "content": "For the pressure solve operation, the matrix", + "type": "text" + }, + { + "bbox": [ + 286, + 354, + 295, + 363 + ], + "score": 0.77, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "is symmetric and positive-definite. This causes the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "adjoint problem to have the same mathematical form (McNamara et al., 2004) as the original prob-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 374, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 388 + ], + "score": 1.0, + "content": "lem. Therefore we implement the gradient for the pressure solve by performing a pressure solve", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "on the gradient. We believe that this is a good example of leveraging the methodology of adjoint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 396, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 411 + ], + "score": 1.0, + "content": "method optimizations (Giles & Pierce, 2000; Treuille et al., 2006; Pontryagin, 1962) within deep", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "learning. With this formalism we arrive at a differentiable solver framework that closely integrates", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 417, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 417, + 505, + 434 + ], + "score": 1.0, + "content": "numerical methods for forward problems with support for inverse problems such as deep learning", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 430, + 201, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 201, + 442 + ], + "score": 1.0, + "content": "via the adjoint method.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 104, + 343, + 506, + 442 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 446, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "Solving Burger’s equation and the Navier-Stokes equations. Using the basic building blocks out-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "lined above, solving the PDEs becomes relatively simple. Burger’s equation involves an advection", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 468, + 326, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 326, + 482 + ], + "score": 1.0, + "content": "and a diffusion term which we evaluate independently.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 446, + 505, + 482 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 487, + 374, + 501 + ], + "lines": [ + { + "bbox": [ + 236, + 487, + 374, + 501 + ], + "spans": [ + { + "bbox": [ + 236, + 487, + 374, + 501 + ], + "score": 0.89, + "content": "\\mathrm { S o l v e r } [ u ] = \\mathrm { D i f f u s e } [ \\mathrm { A d v e c t } [ u , u ] ]", + "type": "interline_equation", + "image_path": "3cccb68531c623ccd6844782acc91707a18cc33aa2b8edd6653df36a295b6ea7.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 236, + 487, + 374, + 501 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 504, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 259, + 520 + ], + "score": 1.0, + "content": "where we explicitly compute Diffuse", + "type": "text" + }, + { + "bbox": [ + 259, + 506, + 332, + 519 + ], + "score": 0.82, + "content": "[ \\boldsymbol { u } ] = \\boldsymbol { u } + \\nu \\nabla ^ { 2 } \\boldsymbol { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 506, + 394, + 520 + ], + "score": 1.0, + "content": "with viscosity", + "type": "text" + }, + { + "bbox": [ + 395, + 510, + 401, + 517 + ], + "score": 0.75, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 506, + 505, + 520 + ], + "score": 1.0, + "content": ". The advection is semi-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 518, + 307, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 307, + 531 + ], + "score": 1.0, + "content": "Lagrangian with back-tracing as described above.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 506, + 505, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 429, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 429, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 429, + 549 + ], + "score": 1.0, + "content": "Solving the Navier-Stokes equations, typically comprises of the following steps:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 534, + 429, + 549 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 556, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 132, + 556, + 318, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 243, + 570 + ], + "score": 1.0, + "content": "• Transporting the density,", + "type": "text" + }, + { + "bbox": [ + 244, + 556, + 318, + 569 + ], + "score": 0.66, + "content": "\\rho \\gets \\mathrm { A d v e c t } [ \\rho , \\pmb { v } ]", + "type": "inline_equation" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 133, + 573, + 323, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 573, + 247, + 586 + ], + "score": 1.0, + "content": "• Transporting the velocity,", + "type": "text" + }, + { + "bbox": [ + 247, + 573, + 323, + 585 + ], + "score": 0.84, + "content": "\\pmb { v } \\mathrm { A d v e c t } [ \\pmb { v } , \\pmb { v } ]", + "type": "inline_equation" + } + ], + "index": 37, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 588, + 320, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 291, + 602 + ], + "score": 1.0, + "content": "• Applying diffusion if the viscosity is", + "type": "text" + }, + { + "bbox": [ + 291, + 590, + 316, + 600 + ], + "score": 0.89, + "content": "\\nu > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 588, + 320, + 602 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 605, + 418, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 249, + 618 + ], + "score": 1.0, + "content": "• Applying buoyancy force,", + "type": "text" + }, + { + "bbox": [ + 249, + 606, + 309, + 617 + ], + "score": 0.93, + "content": "{ \\pmb v } { \\pmb v } - { \\pmb \\beta } \\cdot { \\boldsymbol \\rho }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 605, + 408, + 618 + ], + "score": 1.0, + "content": "with buoyancy direction", + "type": "text" + }, + { + "bbox": [ + 409, + 606, + 418, + 617 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 39, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 621, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 374, + 636 + ], + "score": 1.0, + "content": "• Enforcing incompressibility by solving for the pressure,", + "type": "text" + }, + { + "bbox": [ + 375, + 622, + 480, + 634 + ], + "score": 0.91, + "content": "p \\operatorname { S o l v e } [ A p = \\nabla \\cdot \\pmb { v } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 621, + 505, + 636 + ], + "score": 1.0, + "content": ", then", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 633, + 197, + 645 + ], + "spans": [ + { + "bbox": [ + 142, + 633, + 197, + 645 + ], + "score": 0.71, + "content": "\\pmb { v } \\pmb { v } - \\nabla p", + "type": "inline_equation" + } + ], + "index": 41, + "is_list_end_line": true + } + ], + "index": 38.5, + "bbox_fs": [ + 131, + 556, + 505, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 429, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 653, + 430, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 430, + 667 + ], + "score": 1.0, + "content": "These steps are executed in this order to advance the simulation forward in time.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 653, + 430, + 667 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 684, + 329, + 697 + ], + "lines": [ + { + "bbox": [ + 104, + 682, + 330, + 699 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 330, + 699 + ], + "score": 1.0, + "content": "B COMPLEXITY OF EXECUTION SCHEMES", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 391, + 722 + ], + "score": 1.0, + "content": "The staggered execution scheme recursively splits a sequence of length", + "type": "text" + }, + { + "bbox": [ + 391, + 712, + 398, + 720 + ], + "score": 0.76, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "into smaller sequences, as", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 248, + 733 + ], + "score": 1.0, + "content": "depicted in Fig. 3b and Fig. 7a for", + "type": "text" + }, + { + "bbox": [ + 248, + 721, + 274, + 730 + ], + "score": 0.9, + "content": "n = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 720, + 505, + 733 + ], + "score": 1.0, + "content": ". With each level of recursion depth, the sequence length", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 82, + 500, + 275 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 82, + 500, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 82, + 500, + 275 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 500, + 275 + ], + "score": 0.97, + "type": "image", + "image_path": "4d43b1ae1b1738b52abd4ab4b777f320c19cac4ba1a21a634a85b61ae3bf59d6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 82, + 500, + 146.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 146.33333333333331, + 500, + 210.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 210.66666666666663, + 500, + 274.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 285, + 505, + 320 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 182, + 299 + ], + "score": 1.0, + "content": "Figure 7: OP and", + "type": "text" + }, + { + "bbox": [ + 183, + 286, + 213, + 297 + ], + "score": 0.33, + "content": "\\mathrm { C F E } { + } \\mathrm { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "olver (Sol) executions for a sequence of length 8 performed by (a) the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 104, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "staggered execution scheme, (b) the prediction refinement scheme. The execution order is from top", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 307, + 150, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 150, + 320 + ], + "score": 1.0, + "content": "to bottom.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 337, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 351 + ], + "score": 1.0, + "content": "is cut in half and twice as many predictions need to be performed. The maximum depth depends on", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 348, + 408, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 187, + 362 + ], + "score": 1.0, + "content": "the sequence length", + "type": "text" + }, + { + "bbox": [ + 187, + 349, + 218, + 360 + ], + "score": 0.91, + "content": "t _ { n } - t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 348, + 294, + 362 + ], + "score": 1.0, + "content": "and the time steps", + "type": "text" + }, + { + "bbox": [ + 294, + 349, + 307, + 358 + ], + "score": 0.85, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 348, + 408, + 362 + ], + "score": 1.0, + "content": "performed by the solver,", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 364, + 365, + 392 + ], + "lines": [ + { + "bbox": [ + 246, + 364, + 365, + 392 + ], + "spans": [ + { + "bbox": [ + 246, + 364, + 365, + 392 + ], + "score": 0.95, + "content": "d _ { \\operatorname* { m a x } } = \\log _ { 2 } \\left( \\frac { t _ { n } - t _ { 0 } } { \\Delta t } \\right) - 1 .", + "type": "interline_equation", + "image_path": "a83bf99bf916b1388ac18139bc033ea34b6b551a0cab0d9300163bd113aeed38.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 364, + 365, + 378.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 246, + 378.0, + 365, + 392.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 445, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 446, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 446, + 409 + ], + "score": 1.0, + "content": "Therefore, the total number of predictions, equal to the number of OP evaluations, is", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 412, + 407, + 448 + ], + "lines": [ + { + "bbox": [ + 204, + 412, + 407, + 448 + ], + "spans": [ + { + "bbox": [ + 204, + 412, + 407, + 448 + ], + "score": 0.93, + "content": "N _ { \\mathrm { O P } } = 1 + 2 + 4 + \\ldots + n / 2 = \\sum _ { k = 0 } ^ { d _ { \\operatorname* { m a x } } } 2 ^ { k } = n - 1 .", + "type": "interline_equation", + "image_path": "295e2e44a7311e1e4ebd8e87c02bfab487e0eb72740851cc0b8bedb600b5afd8.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 412, + 407, + 430.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 204, + 430.0, + 407, + 448.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 506, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "The prediction refinement scheme performs more predictions, as can be seen in Fig. 7b.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "To understand the number of OP evaluations, we need to consider the recursive algorithm", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 478, + 504, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 159, + 493 + ], + "score": 1.0, + "content": "Reconstruct", + "type": "text" + }, + { + "bbox": [ + 159, + 479, + 212, + 491 + ], + "score": 0.86, + "content": "[ { \\pmb u } _ { 0 } , { \\pmb o } _ { n } , { \\pmb o } _ { 2 n } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 478, + 254, + 493 + ], + "score": 1.0, + "content": ", listed in", + "type": "text" + }, + { + "bbox": [ + 254, + 479, + 279, + 491 + ], + "score": 0.32, + "content": "\\mathrm { A l g ~ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 478, + 496, + 493 + ], + "score": 1.0, + "content": ", that reconstructs a sequence or partial sequence of", + "type": "text" + }, + { + "bbox": [ + 497, + 481, + 504, + 489 + ], + "score": 0.68, + "content": "n", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 313, + 503 + ], + "score": 1.0, + "content": "frames. For the first invocation, the last parameter", + "type": "text" + }, + { + "bbox": [ + 313, + 492, + 329, + 502 + ], + "score": 0.86, + "content": "{ \\pmb O } _ { 2 n }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "is absent, but for subsequences, that is not", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 381, + 514 + ], + "score": 1.0, + "content": "necessarily the case. Each invocation performs one OP evaluation if", + "type": "text" + }, + { + "bbox": [ + 381, + 502, + 397, + 513 + ], + "score": 0.89, + "content": "{ \\pmb O } _ { 2 n }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "is absent, otherwise three.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 512, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 506, + 524 + ], + "score": 1.0, + "content": "By counting the sequences for which this condition is fulfilled, we can compute the total number of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 524, + 211, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 211, + 534 + ], + "score": 1.0, + "content": "network evaluations to be", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 537, + 412, + 572 + ], + "lines": [ + { + "bbox": [ + 199, + 537, + 412, + 572 + ], + "spans": [ + { + "bbox": [ + 199, + 537, + 412, + 572 + ], + "score": 0.93, + "content": "N _ { \\mathrm { O P } } = 3 \\sum _ { k = 0 } ^ { d _ { \\operatorname* { m a x } } } 2 ^ { k } - 2 \\log _ { 2 } ( n ) = 3 n - 2 \\log _ { 2 } ( n ) - 3 .", + "type": "interline_equation", + "image_path": "6bb607c633491b0c8007edc5bdf841f16e00178d066dabecce17a23c01306de2.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 199, + 537, + 412, + 554.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 199, + 554.5, + 412, + 572.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 591, + 351, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 352, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 352, + 607 + ], + "score": 1.0, + "content": "C NETWORK ARCHITECTURES AND TRAINING", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 615, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "All neural networks used in this work are based on a modified U-net architecture (Ronneberger", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "et al., 2015). The U-net represents a typical multi-level convolutional network architecture with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "skip connections, which we modify by using residual blocks (He et al., 2016) instead of regular", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 648, + 457, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 457, + 662 + ], + "score": 1.0, + "content": "convolutions for each level. We slightly modify this basic layout for some experiments.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "The network used for predicting observations for the fluid example is detailed in Tab. 4. The input", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "to the network are two feature maps containing the current state and the target state. Zero-padding is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "applied to the input, so that all strided convolutions do not require padding. Next, five residual blocks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "are executed in order, each decreasing the resolution (1/2, 1/4, 1/8, 1/16, 1/32) while increasing the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "number of feature maps (4, 8, 16, 16, 16). Each block performs a convolution with kernel size 2", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "and stride 2, followed by two residual blocks with kernel size 3 and symmetric padding. Inside each", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 82, + 500, + 275 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 82, + 500, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 82, + 500, + 275 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 500, + 275 + ], + "score": 0.97, + "type": "image", + "image_path": "4d43b1ae1b1738b52abd4ab4b777f320c19cac4ba1a21a634a85b61ae3bf59d6.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 82, + 500, + 146.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 146.33333333333331, + 500, + 210.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 210.66666666666663, + 500, + 274.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 285, + 505, + 320 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 182, + 299 + ], + "score": 1.0, + "content": "Figure 7: OP and", + "type": "text" + }, + { + "bbox": [ + 183, + 286, + 213, + 297 + ], + "score": 0.33, + "content": "\\mathrm { C F E } { + } \\mathrm { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "olver (Sol) executions for a sequence of length 8 performed by (a) the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 104, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "staggered execution scheme, (b) the prediction refinement scheme. The execution order is from top", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 307, + 150, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 150, + 320 + ], + "score": 1.0, + "content": "to bottom.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 337, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 351 + ], + "score": 1.0, + "content": "is cut in half and twice as many predictions need to be performed. The maximum depth depends on", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 348, + 408, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 187, + 362 + ], + "score": 1.0, + "content": "the sequence length", + "type": "text" + }, + { + "bbox": [ + 187, + 349, + 218, + 360 + ], + "score": 0.91, + "content": "t _ { n } - t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 348, + 294, + 362 + ], + "score": 1.0, + "content": "and the time steps", + "type": "text" + }, + { + "bbox": [ + 294, + 349, + 307, + 358 + ], + "score": 0.85, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 348, + 408, + 362 + ], + "score": 1.0, + "content": "performed by the solver,", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 336, + 505, + 362 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 364, + 365, + 392 + ], + "lines": [ + { + "bbox": [ + 246, + 364, + 365, + 392 + ], + "spans": [ + { + "bbox": [ + 246, + 364, + 365, + 392 + ], + "score": 0.95, + "content": "d _ { \\operatorname* { m a x } } = \\log _ { 2 } \\left( \\frac { t _ { n } - t _ { 0 } } { \\Delta t } \\right) - 1 .", + "type": "interline_equation", + "image_path": "a83bf99bf916b1388ac18139bc033ea34b6b551a0cab0d9300163bd113aeed38.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 364, + 365, + 378.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 246, + 378.0, + 365, + 392.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 396, + 445, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 446, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 446, + 409 + ], + "score": 1.0, + "content": "Therefore, the total number of predictions, equal to the number of OP evaluations, is", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 395, + 446, + 409 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 412, + 407, + 448 + ], + "lines": [ + { + "bbox": [ + 204, + 412, + 407, + 448 + ], + "spans": [ + { + "bbox": [ + 204, + 412, + 407, + 448 + ], + "score": 0.93, + "content": "N _ { \\mathrm { O P } } = 1 + 2 + 4 + \\ldots + n / 2 = \\sum _ { k = 0 } ^ { d _ { \\operatorname* { m a x } } } 2 ^ { k } = n - 1 .", + "type": "interline_equation", + "image_path": "295e2e44a7311e1e4ebd8e87c02bfab487e0eb72740851cc0b8bedb600b5afd8.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 412, + 407, + 430.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 204, + 430.0, + 407, + 448.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 506, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "The prediction refinement scheme performs more predictions, as can be seen in Fig. 7b.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "To understand the number of OP evaluations, we need to consider the recursive algorithm", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 478, + 504, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 159, + 493 + ], + "score": 1.0, + "content": "Reconstruct", + "type": "text" + }, + { + "bbox": [ + 159, + 479, + 212, + 491 + ], + "score": 0.86, + "content": "[ { \\pmb u } _ { 0 } , { \\pmb o } _ { n } , { \\pmb o } _ { 2 n } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 478, + 254, + 493 + ], + "score": 1.0, + "content": ", listed in", + "type": "text" + }, + { + "bbox": [ + 254, + 479, + 279, + 491 + ], + "score": 0.32, + "content": "\\mathrm { A l g ~ 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 478, + 496, + 493 + ], + "score": 1.0, + "content": ", that reconstructs a sequence or partial sequence of", + "type": "text" + }, + { + "bbox": [ + 497, + 481, + 504, + 489 + ], + "score": 0.68, + "content": "n", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 490, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 313, + 503 + ], + "score": 1.0, + "content": "frames. For the first invocation, the last parameter", + "type": "text" + }, + { + "bbox": [ + 313, + 492, + 329, + 502 + ], + "score": 0.86, + "content": "{ \\pmb O } _ { 2 n }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 490, + 506, + 503 + ], + "score": 1.0, + "content": "is absent, but for subsequences, that is not", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 381, + 514 + ], + "score": 1.0, + "content": "necessarily the case. Each invocation performs one OP evaluation if", + "type": "text" + }, + { + "bbox": [ + 381, + 502, + 397, + 513 + ], + "score": 0.89, + "content": "{ \\pmb O } _ { 2 n }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "is absent, otherwise three.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 512, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 506, + 524 + ], + "score": 1.0, + "content": "By counting the sequences for which this condition is fulfilled, we can compute the total number of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 524, + 211, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 211, + 534 + ], + "score": 1.0, + "content": "network evaluations to be", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 457, + 506, + 534 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 537, + 412, + 572 + ], + "lines": [ + { + "bbox": [ + 199, + 537, + 412, + 572 + ], + "spans": [ + { + "bbox": [ + 199, + 537, + 412, + 572 + ], + "score": 0.93, + "content": "N _ { \\mathrm { O P } } = 3 \\sum _ { k = 0 } ^ { d _ { \\operatorname* { m a x } } } 2 ^ { k } - 2 \\log _ { 2 } ( n ) = 3 n - 2 \\log _ { 2 } ( n ) - 3 .", + "type": "interline_equation", + "image_path": "6bb607c633491b0c8007edc5bdf841f16e00178d066dabecce17a23c01306de2.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 199, + 537, + 412, + 554.5 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 199, + 554.5, + 412, + 572.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 591, + 351, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 352, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 352, + 607 + ], + "score": 1.0, + "content": "C NETWORK ARCHITECTURES AND TRAINING", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 615, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "All neural networks used in this work are based on a modified U-net architecture (Ronneberger", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "et al., 2015). The U-net represents a typical multi-level convolutional network architecture with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "skip connections, which we modify by using residual blocks (He et al., 2016) instead of regular", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 648, + 457, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 457, + 662 + ], + "score": 1.0, + "content": "convolutions for each level. We slightly modify this basic layout for some experiments.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 615, + 505, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "The network used for predicting observations for the fluid example is detailed in Tab. 4. The input", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "to the network are two feature maps containing the current state and the target state. Zero-padding is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "applied to the input, so that all strided convolutions do not require padding. Next, five residual blocks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "are executed in order, each decreasing the resolution (1/2, 1/4, 1/8, 1/16, 1/32) while increasing the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "number of feature maps (4, 8, 16, 16, 16). Each block performs a convolution with kernel size 2", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "and stride 2, followed by two residual blocks with kernel size 3 and symmetric padding. Inside each", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "score": 1.0, + "content": "block, the number of feature maps stays constant. Three more residual blocks are executed on the", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 372, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 385 + ], + "score": 1.0, + "content": "lowest resolution of the bowtie structure, after which the decoder part of the network commences,", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 384, + 265, + 395 + ], + "spans": [ + { + "bbox": [ + 107, + 384, + 265, + 395 + ], + "score": 1.0, + "content": "translating features into spatial content.", + "type": "text", + "cross_page": true + } + ], + "index": 21 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 109, + 455, + 330 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 89, + 502, + 109 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 87, + 504, + 103 + ], + "spans": [ + { + "bbox": [ + 105, + 87, + 504, + 103 + ], + "score": 1.0, + "content": "Table 4: Layers comprising the observation predictor network used in the direct fluid control exper-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 135, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 135, + 113 + ], + "score": 1.0, + "content": "iment.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 155, + 109, + 455, + 330 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 155, + 109, + 455, + 330 + ], + "spans": [ + { + "bbox": [ + 155, + 109, + 455, + 330 + ], + "score": 0.983, + "html": "
LayerResolutionFeature Maps
Input1282
Pad1592
Strided convolution + 2x Residual block794
Strided convolution+ 2x Residual block398
Strided convolution+ 2x Residual block1916
Strided convolution+ 2x Residual block916
Strided convolution+ 2x Residual block416
3x Residual block416
Upsample + Concatenate832
Convolution+ 2x Residual block816
Upsample+Concatenate1632
Convolution + 2x Residual block1616
Upsample+Concatenate3224
Convolution + 2x Residual block3216
Upsample+ Concatenate6420
Convolution + 2x Residual block6416
Upsample+ Concatenate12818
Convolution1281
", + "type": "table", + "image_path": "9dc969cd5b010da3b85cb8852932683ed869c9222a79153594be06b9ce48ca2a.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 155, + 109, + 455, + 122.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 155, + 122.0, + 455, + 135.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 155, + 135.0, + 455, + 148.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 155, + 148.0, + 455, + 161.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 155, + 161.0, + 455, + 174.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 155, + 174.0, + 455, + 187.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 155, + 187.0, + 455, + 200.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 155, + 200.0, + 455, + 213.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 155, + 213.0, + 455, + 226.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 155, + 226.0, + 455, + 239.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 155, + 239.0, + 455, + 252.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 155, + 252.0, + 455, + 265.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 155, + 265.0, + 455, + 278.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 155, + 278.0, + 455, + 291.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 155, + 291.0, + 455, + 304.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 155, + 304.0, + 455, + 317.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 155, + 317.0, + 455, + 330.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 5.25 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 504, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 374 + ], + "score": 1.0, + "content": "block, the number of feature maps stays constant. Three more residual blocks are executed on the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 372, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 385 + ], + "score": 1.0, + "content": "lowest resolution of the bowtie structure, after which the decoder part of the network commences,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 384, + 265, + 395 + ], + "spans": [ + { + "bbox": [ + 107, + 384, + 265, + 395 + ], + "score": 1.0, + "content": "translating features into spatial content.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 413 + ], + "score": 1.0, + "content": "The decoder works as follows: Starting with the lowest resolution, the feature maps are upsampled", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "with linear interpolation. The upsampled maps and the output of the previous block of same reso-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 423, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 434 + ], + "score": 1.0, + "content": "lution are then concatenated. Next, a convolution with 16 filters, a kernel size of 2 and symmetric", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "padding, followed by two more residual blocks, is executed. When the original resolution is reached,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 444, + 438, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 438, + 457 + ], + "score": 1.0, + "content": "only one feature map is produced instead of 16, forming the output of the network.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "Depending on the dimensionality of the problem, either 1D or 2D convolutions are used. The net-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "work used for the indirect control task is modified in the following ways: (i) It produces two output", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 262, + 496 + ], + "score": 1.0, + "content": "feature maps, representing the velocity", + "type": "text" + }, + { + "bbox": [ + 262, + 483, + 294, + 496 + ], + "score": 0.92, + "content": "( v _ { x } , v _ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 483, + 505, + 496 + ], + "score": 1.0, + "content": ". (ii) Four feature maps of the lowest resolution (4x4)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "are fed into a dense layer producing four output feature maps. These and the other feature maps are", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "concatenated before moving to the upsampling stage. This modification ensures that the receptive", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 515, + 272, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 272, + 528 + ], + "score": 1.0, + "content": "field of the network is the whole domain.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "All networks were implemented in TensorFlow (Abadi et al., 2016) and trained using the ADAM", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 542, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 558 + ], + "score": 1.0, + "content": "optimizer on an Nvidia GTX 1080 Ti. We use batch sizes ranging from 4 to 16. Supervised training", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "of all networks converges within a few minutes, for which we iteratively decrease the learning rate", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 563, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 128, + 579 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 565, + 150, + 576 + ], + "score": 0.9, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 563, + 162, + 579 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 163, + 565, + 184, + 576 + ], + "score": 0.9, + "content": "1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 563, + 506, + 579 + ], + "score": 1.0, + "content": ". We stop supervised training after a few epochs, comprising between 2000 and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 576, + 452, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 452, + 590 + ], + "score": 1.0, + "content": "10.000 iterations, as the networks usually converge within a fraction of the first epoch.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 457, + 606 + ], + "score": 1.0, + "content": "For training with the differentiable solver, we start with a decreased learning rate of", + "type": "text" + }, + { + "bbox": [ + 458, + 593, + 480, + 604 + ], + "score": 0.89, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 592, + 506, + 606 + ], + "score": 1.0, + "content": "since", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "the backpropagation through long chains is more challenging than training with a supervised loss.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "Optimization steps are also considerably more expensive since the whole chain needs to be executed,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "which includes a forward and backward simulation pass. For the fluid examples, an optimization", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "score": 1.0, + "content": "step takes 1-2 seconds to complete for the 2D fluid problems. We let the networks run about 100.000", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 649, + 405, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 405, + 660 + ], + "score": 1.0, + "content": "iterations, which takes between one and two days for the shown examples.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "title", + "bbox": [ + 107, + 689, + 447, + 700 + ], + "lines": [ + { + "bbox": [ + 104, + 687, + 450, + 704 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 450, + 704 + ], + "score": 1.0, + "content": "D DETAILED DESCRIPTION AND ANALYSIS OF THE EXPERIMENTS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 443, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 444, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 444, + 733 + ], + "score": 1.0, + "content": "In the following paragraphs, we give further details on the experiments of Section 6.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 109, + 455, + 330 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 89, + 502, + 109 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 87, + 504, + 103 + ], + "spans": [ + { + "bbox": [ + 105, + 87, + 504, + 103 + ], + "score": 1.0, + "content": "Table 4: Layers comprising the observation predictor network used in the direct fluid control exper-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 135, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 135, + 113 + ], + "score": 1.0, + "content": "iment.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 155, + 109, + 455, + 330 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 155, + 109, + 455, + 330 + ], + "spans": [ + { + "bbox": [ + 155, + 109, + 455, + 330 + ], + "score": 0.983, + "html": "
LayerResolutionFeature Maps
Input1282
Pad1592
Strided convolution + 2x Residual block794
Strided convolution+ 2x Residual block398
Strided convolution+ 2x Residual block1916
Strided convolution+ 2x Residual block916
Strided convolution+ 2x Residual block416
3x Residual block416
Upsample + Concatenate832
Convolution+ 2x Residual block816
Upsample+Concatenate1632
Convolution + 2x Residual block1616
Upsample+Concatenate3224
Convolution + 2x Residual block3216
Upsample+ Concatenate6420
Convolution + 2x Residual block6416
Upsample+ Concatenate12818
Convolution1281
", + "type": "table", + "image_path": "9dc969cd5b010da3b85cb8852932683ed869c9222a79153594be06b9ce48ca2a.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 155, + 109, + 455, + 122.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 155, + 122.0, + 455, + 135.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 155, + 135.0, + 455, + 148.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 155, + 148.0, + 455, + 161.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 155, + 161.0, + 455, + 174.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 155, + 174.0, + 455, + 187.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 155, + 187.0, + 455, + 200.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 155, + 200.0, + 455, + 213.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 155, + 213.0, + 455, + 226.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 155, + 226.0, + 455, + 239.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 155, + 239.0, + 455, + 252.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 155, + 252.0, + 455, + 265.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 155, + 265.0, + 455, + 278.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 155, + 278.0, + 455, + 291.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 155, + 291.0, + 455, + 304.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 155, + 304.0, + 455, + 317.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 155, + 317.0, + 455, + 330.0 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 5.25 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 504, + 394 + ], + "lines": [], + "index": 20, + "bbox_fs": [ + 105, + 360, + 506, + 395 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 413 + ], + "score": 1.0, + "content": "The decoder works as follows: Starting with the lowest resolution, the feature maps are upsampled", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "with linear interpolation. The upsampled maps and the output of the previous block of same reso-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 423, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 434 + ], + "score": 1.0, + "content": "lution are then concatenated. Next, a convolution with 16 filters, a kernel size of 2 and symmetric", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "padding, followed by two more residual blocks, is executed. When the original resolution is reached,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 444, + 438, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 438, + 457 + ], + "score": 1.0, + "content": "only one feature map is produced instead of 16, forming the output of the network.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 399, + 505, + 457 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "Depending on the dimensionality of the problem, either 1D or 2D convolutions are used. The net-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "work used for the indirect control task is modified in the following ways: (i) It produces two output", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 262, + 496 + ], + "score": 1.0, + "content": "feature maps, representing the velocity", + "type": "text" + }, + { + "bbox": [ + 262, + 483, + 294, + 496 + ], + "score": 0.92, + "content": "( v _ { x } , v _ { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 483, + 505, + 496 + ], + "score": 1.0, + "content": ". (ii) Four feature maps of the lowest resolution (4x4)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "are fed into a dense layer producing four output feature maps. These and the other feature maps are", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "concatenated before moving to the upsampling stage. This modification ensures that the receptive", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 515, + 272, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 272, + 528 + ], + "score": 1.0, + "content": "field of the network is the whole domain.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 461, + 505, + 528 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "All networks were implemented in TensorFlow (Abadi et al., 2016) and trained using the ADAM", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 542, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 558 + ], + "score": 1.0, + "content": "optimizer on an Nvidia GTX 1080 Ti. We use batch sizes ranging from 4 to 16. Supervised training", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "of all networks converges within a few minutes, for which we iteratively decrease the learning rate", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 563, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 128, + 579 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 565, + 150, + 576 + ], + "score": 0.9, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 563, + 162, + 579 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 163, + 565, + 184, + 576 + ], + "score": 0.9, + "content": "1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 563, + 506, + 579 + ], + "score": 1.0, + "content": ". We stop supervised training after a few epochs, comprising between 2000 and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 576, + 452, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 452, + 590 + ], + "score": 1.0, + "content": "10.000 iterations, as the networks usually converge within a fraction of the first epoch.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 104, + 533, + 506, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 457, + 606 + ], + "score": 1.0, + "content": "For training with the differentiable solver, we start with a decreased learning rate of", + "type": "text" + }, + { + "bbox": [ + 458, + 593, + 480, + 604 + ], + "score": 0.89, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 592, + 506, + 606 + ], + "score": 1.0, + "content": "since", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "the backpropagation through long chains is more challenging than training with a supervised loss.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "Optimization steps are also considerably more expensive since the whole chain needs to be executed,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "which includes a forward and backward simulation pass. For the fluid examples, an optimization", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 649 + ], + "score": 1.0, + "content": "step takes 1-2 seconds to complete for the 2D fluid problems. We let the networks run about 100.000", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 649, + 405, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 405, + 660 + ], + "score": 1.0, + "content": "iterations, which takes between one and two days for the shown examples.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 592, + 506, + 660 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 689, + 447, + 700 + ], + "lines": [ + { + "bbox": [ + 104, + 687, + 450, + 704 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 450, + 704 + ], + "score": 1.0, + "content": "D DETAILED DESCRIPTION AND ANALYSIS OF THE EXPERIMENTS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 443, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 444, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 444, + 733 + ], + "score": 1.0, + "content": "In the following paragraphs, we give further details on the experiments of Section 6.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 719, + 444, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 79, + 501, + 382 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 79, + 501, + 382 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 79, + 501, + 382 + ], + "spans": [ + { + "bbox": [ + 111, + 79, + 501, + 382 + ], + "score": 0.964, + "type": "image", + "image_path": "616bbed2522d0f40a9cfb0dce77c21be4b96022b4714e049ae484e46678b655f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 79, + 501, + 180.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 180.0, + 501, + 281.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 281.0, + 501, + 382.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 390, + 506, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "Figure 8: Trajectories for example control tasks using Burger’s equation. The initial state is plot-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "ted in red, the target state in blue. (a) Natural evolution, (b) Ground truth trajectory generated", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "with constant force, (c) CFE with supervised loss, (d) CFE with differentiable physics loss, (e) Re-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "construction with supervised loss, (f) Reconstruction with differentiable physics loss and staggered", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 434, + 460, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 460, + 448 + ], + "score": 1.0, + "content": "execution, (g) Reconstruction with differentiable physics loss and prediction refinement.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 469, + 225, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 469, + 227, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 227, + 482 + ], + "score": 1.0, + "content": "D.1 BURGER’S EQUATION", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 490, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 504 + ], + "score": 1.0, + "content": "For this experiment, we simulate Burger’s equation (Eq. 6) on a one-dimensional grid with 32 sam-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "ples over a course of 32 time steps. The typical behavior of Burger’s equation in 1D exhibits shock", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 510, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 187, + 527 + ], + "score": 1.0, + "content": "waves that move in", + "type": "text" + }, + { + "bbox": [ + 188, + 514, + 203, + 523 + ], + "score": 0.86, + "content": "+ x", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 510, + 215, + 527 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 216, + 514, + 231, + 523 + ], + "score": 0.85, + "content": "- x", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 510, + 285, + 527 + ], + "score": 1.0, + "content": "direction for", + "type": "text" + }, + { + "bbox": [ + 286, + 512, + 326, + 524 + ], + "score": 0.93, + "content": "u ( x ) > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 510, + 339, + 527 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 340, + 512, + 380, + 524 + ], + "score": 0.93, + "content": "u ( x ) \\bar { < } 0", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 510, + 505, + 527 + ], + "score": 1.0, + "content": ", respectively. When opposing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 523, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 537 + ], + "score": 1.0, + "content": "waves clash, they both weaken until only the stronger wave survives and keeps moving. Examples", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 534, + 225, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 225, + 546 + ], + "score": 1.0, + "content": "are shown in Figs. 4a and 8a.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 504, + 574 + ], + "lines": [ + { + "bbox": [ + 106, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 324, + 564 + ], + "score": 1.0, + "content": "All 32 samples are observable and controllable, i.e.", + "type": "text" + }, + { + "bbox": [ + 324, + 551, + 375, + 563 + ], + "score": 0.93, + "content": "o ( t ) = u ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 551, + 505, + 564 + ], + "score": 1.0, + "content": ". Thus, we can enforce that all", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 562, + 444, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 444, + 575 + ], + "score": 1.0, + "content": "trajectories reach the target state exactly by choosing the force for the last step to be", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 581, + 361, + 606 + ], + "lines": [ + { + "bbox": [ + 250, + 581, + 361, + 606 + ], + "spans": [ + { + "bbox": [ + 250, + 581, + 361, + 606 + ], + "score": 0.94, + "content": "F ( t _ { n - 1 } ) = \\frac { o ^ { * } - u ( t _ { n - 1 } ) } { \\Delta t } .", + "type": "interline_equation", + "image_path": "0e89ff775327e69887ac8382f0b93dedae13f2b33b5d23b1c0d2b56059f98118.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 250, + 581, + 361, + 606 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 504, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "To measure the quality of a solution, it is therefore sufficient to consider the applied force", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 622, + 367, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 159, + 640 + ], + "score": 0.92, + "content": "\\int _ { t _ { 0 } } ^ { t _ { * } } | F ( t ) | d t", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 622, + 367, + 641 + ], + "score": 1.0, + "content": "which is detailed for the tested methods in Table 1.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "Network training. Both for the CFE chains as well as for the observation prediction models, we use", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "the same network architecture, described in Appendix C. We train the networks on 3600 randomly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 293, + 678 + ], + "score": 1.0, + "content": "generated scenes with constant driving forces,", + "type": "text" + }, + { + "bbox": [ + 293, + 666, + 348, + 678 + ], + "score": 0.68, + "content": "\\mathbf { \\tilde { \\cal F } } ( t ) = \\mathrm { c o n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 666, + 505, + 678 + ], + "score": 1.0, + "content": ". The examples are initialized with two", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "Gaussian waves of random amplitude, size and position, set to clash in the center. In each time", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "step, a constant Gaussian force with the same randomized parameters is applied to the system to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 700, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 711 + ], + "score": 1.0, + "content": "steer it away from its natural evolution. Constant forces have a larger impact on the evolution than", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "temporally varying forces since the effects of temporally varying forces can partly cancel out over", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "score": 1.0, + "content": "time. The ground truth sequence can therefore be regarded as a near-perfect but not necessarily", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 79, + 501, + 382 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 79, + 501, + 382 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 79, + 501, + 382 + ], + "spans": [ + { + "bbox": [ + 111, + 79, + 501, + 382 + ], + "score": 0.964, + "type": "image", + "image_path": "616bbed2522d0f40a9cfb0dce77c21be4b96022b4714e049ae484e46678b655f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 79, + 501, + 180.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 180.0, + 501, + 281.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 281.0, + 501, + 382.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 390, + 506, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "Figure 8: Trajectories for example control tasks using Burger’s equation. The initial state is plot-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "ted in red, the target state in blue. (a) Natural evolution, (b) Ground truth trajectory generated", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 425 + ], + "score": 1.0, + "content": "with constant force, (c) CFE with supervised loss, (d) CFE with differentiable physics loss, (e) Re-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "construction with supervised loss, (f) Reconstruction with differentiable physics loss and staggered", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 434, + 460, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 460, + 448 + ], + "score": 1.0, + "content": "execution, (g) Reconstruction with differentiable physics loss and prediction refinement.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 107, + 469, + 225, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 469, + 227, + 482 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 227, + 482 + ], + "score": 1.0, + "content": "D.1 BURGER’S EQUATION", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 490, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 504 + ], + "score": 1.0, + "content": "For this experiment, we simulate Burger’s equation (Eq. 6) on a one-dimensional grid with 32 sam-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "ples over a course of 32 time steps. The typical behavior of Burger’s equation in 1D exhibits shock", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 510, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 187, + 527 + ], + "score": 1.0, + "content": "waves that move in", + "type": "text" + }, + { + "bbox": [ + 188, + 514, + 203, + 523 + ], + "score": 0.86, + "content": "+ x", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 510, + 215, + 527 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 216, + 514, + 231, + 523 + ], + "score": 0.85, + "content": "- x", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 510, + 285, + 527 + ], + "score": 1.0, + "content": "direction for", + "type": "text" + }, + { + "bbox": [ + 286, + 512, + 326, + 524 + ], + "score": 0.93, + "content": "u ( x ) > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 510, + 339, + 527 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 340, + 512, + 380, + 524 + ], + "score": 0.93, + "content": "u ( x ) \\bar { < } 0", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 510, + 505, + 527 + ], + "score": 1.0, + "content": ", respectively. When opposing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 523, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 537 + ], + "score": 1.0, + "content": "waves clash, they both weaken until only the stronger wave survives and keeps moving. Examples", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 534, + 225, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 225, + 546 + ], + "score": 1.0, + "content": "are shown in Figs. 4a and 8a.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 490, + 505, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 550, + 504, + 574 + ], + "lines": [ + { + "bbox": [ + 106, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 324, + 564 + ], + "score": 1.0, + "content": "All 32 samples are observable and controllable, i.e.", + "type": "text" + }, + { + "bbox": [ + 324, + 551, + 375, + 563 + ], + "score": 0.93, + "content": "o ( t ) = u ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 551, + 505, + 564 + ], + "score": 1.0, + "content": ". Thus, we can enforce that all", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 562, + 444, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 444, + 575 + ], + "score": 1.0, + "content": "trajectories reach the target state exactly by choosing the force for the last step to be", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 551, + 505, + 575 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 581, + 361, + 606 + ], + "lines": [ + { + "bbox": [ + 250, + 581, + 361, + 606 + ], + "spans": [ + { + "bbox": [ + 250, + 581, + 361, + 606 + ], + "score": 0.94, + "content": "F ( t _ { n - 1 } ) = \\frac { o ^ { * } - u ( t _ { n - 1 } ) } { \\Delta t } .", + "type": "interline_equation", + "image_path": "0e89ff775327e69887ac8382f0b93dedae13f2b33b5d23b1c0d2b56059f98118.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 250, + 581, + 361, + 606 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 504, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "To measure the quality of a solution, it is therefore sufficient to consider the applied force", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 622, + 367, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 159, + 640 + ], + "score": 0.92, + "content": "\\int _ { t _ { 0 } } ^ { t _ { * } } | F ( t ) | d t", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 622, + 367, + 641 + ], + "score": 1.0, + "content": "which is detailed for the tested methods in Table 1.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 612, + 505, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "Network training. Both for the CFE chains as well as for the observation prediction models, we use", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "the same network architecture, described in Appendix C. We train the networks on 3600 randomly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 293, + 678 + ], + "score": 1.0, + "content": "generated scenes with constant driving forces,", + "type": "text" + }, + { + "bbox": [ + 293, + 666, + 348, + 678 + ], + "score": 0.68, + "content": "\\mathbf { \\tilde { \\cal F } } ( t ) = \\mathrm { c o n s t }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 666, + 505, + 678 + ], + "score": 1.0, + "content": ". The examples are initialized with two", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "Gaussian waves of random amplitude, size and position, set to clash in the center. In each time", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "step, a constant Gaussian force with the same randomized parameters is applied to the system to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 700, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 711 + ], + "score": 1.0, + "content": "steer it away from its natural evolution. Constant forces have a larger impact on the evolution than", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "temporally varying forces since the effects of temporally varying forces can partly cancel out over", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "score": 1.0, + "content": "time. The ground truth sequence can therefore be regarded as a near-perfect but not necessarily", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "optimal trajectory. Figs. 4d and 8b display such examples. The same trajectories, without any forces", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 320, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 320, + 106 + ], + "score": 1.0, + "content": "applied, are shown in sub-figures (a) for comparison.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 643, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "optimal trajectory. Figs. 4d and 8b display such examples. The same trajectories, without any forces", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 320, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 320, + 106 + ], + "score": 1.0, + "content": "applied, are shown in sub-figures (a) for comparison.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 504, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "We pretrain all networks (OPs or CFE, depending on the method) with a supervised observation", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 119, + 130, + 135 + ], + "spans": [ + { + "bbox": [ + 104, + 119, + 130, + 135 + ], + "score": 1.0, + "content": "loss,", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 130, + 398, + 160 + ], + "lines": [ + { + "bbox": [ + 212, + 130, + 398, + 160 + ], + "spans": [ + { + "bbox": [ + 212, + 130, + 398, + 160 + ], + "score": 0.94, + "content": "L _ { o } ^ { \\mathrm { s u p } } = \\left| \\operatorname { O P } [ o ( t _ { i } ) , o ( t _ { j } ) ] - u ^ { \\mathrm { G T } } \\left( { \\frac { t _ { i } + t _ { j } } { 2 } } \\right) \\right| ^ { 2 } .", + "type": "interline_equation", + "image_path": "5dc56e4df962cd968a704591000f2069fb2689e0914a8db9c0abf9216f590f58.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 130, + 398, + 145.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 212, + 145.0, + 398, + 160.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 161, + 503, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 175 + ], + "score": 1.0, + "content": "The resulting trajectory after supervised training for the CFE chain is shown in Figure 4b and Fig-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 433, + 185 + ], + "score": 1.0, + "content": "ure 8c. For the observation prediction models, the trajectories are shown in Figure", + "type": "text" + }, + { + "bbox": [ + 434, + 173, + 445, + 182 + ], + "score": 0.7, + "content": "_ \\mathrm { 4 c }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "and Figure 8e.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 189, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 188, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 202 + ], + "score": 1.0, + "content": "After pretraining, we train all OP networks end-to-end with our objective loss function (see Eq. 4),", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 200, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 213 + ], + "score": 1.0, + "content": "making use of the differentiable solver. For this experiment, we choose the mean squared difference", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 210, + 240, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 240, + 223 + ], + "score": 1.0, + "content": "for the observation loss function:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 226, + 354, + 241 + ], + "lines": [ + { + "bbox": [ + 257, + 226, + 354, + 241 + ], + "spans": [ + { + "bbox": [ + 257, + 226, + 354, + 241 + ], + "score": 0.92, + "content": "L _ { o } ^ { * } = \\left| o ( u ( t _ { * } ) ) - o ^ { * } \\right| ^ { 2 } .", + "type": "interline_equation", + "image_path": "5092c43fc3ed720c42cf9a92a704cb863e851d4dfa995624a83cf936f30cf5a5.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 257, + 226, + 354, + 241 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 269 + ], + "lines": [ + { + "bbox": [ + 105, + 244, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 260 + ], + "score": 1.0, + "content": "We test both the staggered execution scheme and the prediction refinement scheme, shown in Fig-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 193, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 193, + 271 + ], + "score": 1.0, + "content": "ure 8f and Figure 8g.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 274, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "Results. Table 1 compares the resulting forces inferred by different methods. The results are aver-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 285, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 297 + ], + "score": 1.0, + "content": "aged over a set of 100 examples from the test set which is sampled from the same distribution as the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 296, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 325, + 308 + ], + "score": 1.0, + "content": "training set. The CFE chains both fail to converge to", + "type": "text" + }, + { + "bbox": [ + 325, + 297, + 335, + 306 + ], + "score": 0.87, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 296, + 505, + 308 + ], + "score": 1.0, + "content": ". While the differentiable physics version", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 197, + 320 + ], + "score": 1.0, + "content": "manages to produce a", + "type": "text" + }, + { + "bbox": [ + 197, + 308, + 219, + 319 + ], + "score": 0.89, + "content": "u _ { n - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 307, + 281, + 320 + ], + "score": 1.0, + "content": "that resembles", + "type": "text" + }, + { + "bbox": [ + 281, + 308, + 292, + 317 + ], + "score": 0.85, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 307, + 506, + 320 + ], + "score": 1.0, + "content": ", the supervised version completely deviates from an", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 383, + 331 + ], + "score": 1.0, + "content": "optimal trajectory. This shows that learning to infer the control force", + "type": "text" + }, + { + "bbox": [ + 384, + 318, + 407, + 330 + ], + "score": 0.93, + "content": "F ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 317, + 451, + 331 + ], + "score": 1.0, + "content": "only from", + "type": "text" + }, + { + "bbox": [ + 451, + 318, + 472, + 330 + ], + "score": 0.88, + "content": "u ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 317, + 476, + 331 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 476, + 318, + 487, + 329 + ], + "score": 0.67, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 111, + 339 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "is very difficult as the model needs to learn to anticipate the physical behavior over any length of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 339, + 131, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 131, + 353 + ], + "score": 1.0, + "content": "time.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "Compared to the CFE chains, the hierarchical models require much less force and learn to converge", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 140, + 380 + ], + "score": 1.0, + "content": "towards", + "type": "text" + }, + { + "bbox": [ + 141, + 369, + 151, + 378 + ], + "score": 0.84, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 368, + 505, + 380 + ], + "score": 1.0, + "content": ". Still, the supervised training applies much more force to the system than required, the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 378, + 345, + 392 + ], + "score": 1.0, + "content": "reasons for which become obvious when inspecting Figure", + "type": "text" + }, + { + "bbox": [ + 345, + 379, + 357, + 389 + ], + "score": 0.25, + "content": "^ \\mathrm { 4 b }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 378, + 505, + 392 + ], + "score": 1.0, + "content": "and Fig. 8e. While each state seems", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "score": 1.0, + "content": "close to the ground truth individually, the control oscillates undesirably, requiring counter-actions", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 399, + 160, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 160, + 413 + ], + "score": 1.0, + "content": "later in time.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "The methods using the differentiable solver significantly outperform their supervised counterparts", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "and exhibit an excellent performance that is very close the ground truth solutions in terms of required", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "forces. On many examples, they even reach the target state with less force than was applied by the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 450, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 465 + ], + "score": 1.0, + "content": "ground truth simulation. This would not be possible with the supervised loss alone, but by having", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "access to the gradient-based feedback from the differentiable solver, they can learn to find more", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 471, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 486 + ], + "score": 1.0, + "content": "efficient trajectories with respect to the objective loss. This allows the networks to learn applying", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 483, + 468, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 468, + 496 + ], + "score": 1.0, + "content": "forces in different locations that make the system approach the target state with less force.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 505, + 589 + ], + "lines": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "score": 1.0, + "content": "Figure 4e and Fig.8f,g show examples of this. The ground truth applies the same force in each step,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 297, + 524 + ], + "score": 1.0, + "content": "thereby continuously increasing the first sample", + "type": "text" + }, + { + "bbox": [ + 297, + 511, + 335, + 523 + ], + "score": 0.91, + "content": "u ( x = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 511, + 506, + 524 + ], + "score": 1.0, + "content": ", and the supervised method tries to imitate", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 353, + 535 + ], + "score": 1.0, + "content": "this behavior. The governing equation then slowly propagates", + "type": "text" + }, + { + "bbox": [ + 353, + 522, + 391, + 534 + ], + "score": 0.9, + "content": "u ( x = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 522, + 436, + 535 + ], + "score": 1.0, + "content": "in positive", + "type": "text" + }, + { + "bbox": [ + 436, + 524, + 443, + 532 + ], + "score": 0.74, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "direction since", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 172, + 545 + ], + "score": 0.9, + "content": "u ( x = 0 ) > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 533, + 505, + 546 + ], + "score": 1.0, + "content": ". The learning methods that use a differentiable solver make use of this fact by", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 218, + 558 + ], + "score": 1.0, + "content": "applying much more force", + "type": "text" + }, + { + "bbox": [ + 218, + 544, + 283, + 556 + ], + "score": 0.92, + "content": "F ( x = 0 ) > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "at this point than the ground truth, even overshooting", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "the target state. Later, when this value had time to propagate to the right, the model corrects this", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 268, + 578 + ], + "score": 1.0, + "content": "overshoot by applying a negative force", + "type": "text" + }, + { + "bbox": [ + 268, + 566, + 333, + 578 + ], + "score": 0.92, + "content": "F ( x = 0 ) \\ { \\overset { } { < } } \\ { \\overset { - } { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 567, + 505, + 578 + ], + "score": 1.0, + "content": ". Using this trick, these models reach the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 577, + 486, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 195, + 590 + ], + "score": 1.0, + "content": "target state with up to", + "type": "text" + }, + { + "bbox": [ + 195, + 577, + 215, + 587 + ], + "score": 0.86, + "content": "13 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 577, + 486, + 590 + ], + "score": 1.0, + "content": "less force than the ground truth on the sequence shown in Figure 4.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 504, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 504, + 607 + ], + "score": 1.0, + "content": "Figure 9 analyzes the variance of inferred forces. The supervised methods often fail to properly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "score": 1.0, + "content": "converge to the target state, resulting in large forces in the last step, visible as a second peak in the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "supervised CFE chain. The formulation of the loss (Eq. 3) suppresses force spikes. In the solutions", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "inferred by our method, the likelihood of large forces falls off multi-exponentially as a consequence.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "This means that large forces are exponentially rare, which is the expected behavior given the L2", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 200, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 200, + 662 + ], + "score": 1.0, + "content": "regularizer from Eq. 3.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "We also compare our results to a single-shooting baseline which is able to find near-optimal solutions", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "at the cost of higher computation times. The classic optimization uses the ADAM optimizer with a", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "learning rate of 0.01 and converges after around 300 iterations. To reach the quality of the staggered", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "prediction scheme, it requires only around 60 iterations. This quick convergence can be explained", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "by the relatively simple setup that is dominated by linear effects. Therefore, the gradients are stable,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 288, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 288, + 732 + ], + "score": 1.0, + "content": "even when propagated through many frames.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 81, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 504, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 124 + ], + "score": 1.0, + "content": "We pretrain all networks (OPs or CFE, depending on the method) with a supervised observation", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 119, + 130, + 135 + ], + "spans": [ + { + "bbox": [ + 104, + 119, + 130, + 135 + ], + "score": 1.0, + "content": "loss,", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 109, + 505, + 135 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 130, + 398, + 160 + ], + "lines": [ + { + "bbox": [ + 212, + 130, + 398, + 160 + ], + "spans": [ + { + "bbox": [ + 212, + 130, + 398, + 160 + ], + "score": 0.94, + "content": "L _ { o } ^ { \\mathrm { s u p } } = \\left| \\operatorname { O P } [ o ( t _ { i } ) , o ( t _ { j } ) ] - u ^ { \\mathrm { G T } } \\left( { \\frac { t _ { i } + t _ { j } } { 2 } } \\right) \\right| ^ { 2 } .", + "type": "interline_equation", + "image_path": "5dc56e4df962cd968a704591000f2069fb2689e0914a8db9c0abf9216f590f58.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 130, + 398, + 145.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 212, + 145.0, + 398, + 160.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 161, + 503, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 175 + ], + "score": 1.0, + "content": "The resulting trajectory after supervised training for the CFE chain is shown in Figure 4b and Fig-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 433, + 185 + ], + "score": 1.0, + "content": "ure 8c. For the observation prediction models, the trajectories are shown in Figure", + "type": "text" + }, + { + "bbox": [ + 434, + 173, + 445, + 182 + ], + "score": 0.7, + "content": "_ \\mathrm { 4 c }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "and Figure 8e.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 160, + 505, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 189, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 105, + 188, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 202 + ], + "score": 1.0, + "content": "After pretraining, we train all OP networks end-to-end with our objective loss function (see Eq. 4),", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 200, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 213 + ], + "score": 1.0, + "content": "making use of the differentiable solver. For this experiment, we choose the mean squared difference", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 210, + 240, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 240, + 223 + ], + "score": 1.0, + "content": "for the observation loss function:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 188, + 505, + 223 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 226, + 354, + 241 + ], + "lines": [ + { + "bbox": [ + 257, + 226, + 354, + 241 + ], + "spans": [ + { + "bbox": [ + 257, + 226, + 354, + 241 + ], + "score": 0.92, + "content": "L _ { o } ^ { * } = \\left| o ( u ( t _ { * } ) ) - o ^ { * } \\right| ^ { 2 } .", + "type": "interline_equation", + "image_path": "5092c43fc3ed720c42cf9a92a704cb863e851d4dfa995624a83cf936f30cf5a5.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 257, + 226, + 354, + 241 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 269 + ], + "lines": [ + { + "bbox": [ + 105, + 244, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 260 + ], + "score": 1.0, + "content": "We test both the staggered execution scheme and the prediction refinement scheme, shown in Fig-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 193, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 193, + 271 + ], + "score": 1.0, + "content": "ure 8f and Figure 8g.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 244, + 505, + 271 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 274, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "Results. Table 1 compares the resulting forces inferred by different methods. The results are aver-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 285, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 297 + ], + "score": 1.0, + "content": "aged over a set of 100 examples from the test set which is sampled from the same distribution as the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 296, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 325, + 308 + ], + "score": 1.0, + "content": "training set. The CFE chains both fail to converge to", + "type": "text" + }, + { + "bbox": [ + 325, + 297, + 335, + 306 + ], + "score": 0.87, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 296, + 505, + 308 + ], + "score": 1.0, + "content": ". While the differentiable physics version", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 197, + 320 + ], + "score": 1.0, + "content": "manages to produce a", + "type": "text" + }, + { + "bbox": [ + 197, + 308, + 219, + 319 + ], + "score": 0.89, + "content": "u _ { n - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 307, + 281, + 320 + ], + "score": 1.0, + "content": "that resembles", + "type": "text" + }, + { + "bbox": [ + 281, + 308, + 292, + 317 + ], + "score": 0.85, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 307, + 506, + 320 + ], + "score": 1.0, + "content": ", the supervised version completely deviates from an", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 383, + 331 + ], + "score": 1.0, + "content": "optimal trajectory. This shows that learning to infer the control force", + "type": "text" + }, + { + "bbox": [ + 384, + 318, + 407, + 330 + ], + "score": 0.93, + "content": "F ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 317, + 451, + 331 + ], + "score": 1.0, + "content": "only from", + "type": "text" + }, + { + "bbox": [ + 451, + 318, + 472, + 330 + ], + "score": 0.88, + "content": "u ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 317, + 476, + 331 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 476, + 318, + 487, + 329 + ], + "score": 0.67, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 329, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 111, + 339 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 329, + 506, + 342 + ], + "score": 1.0, + "content": "is very difficult as the model needs to learn to anticipate the physical behavior over any length of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 339, + 131, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 131, + 353 + ], + "score": 1.0, + "content": "time.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 275, + 506, + 353 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "Compared to the CFE chains, the hierarchical models require much less force and learn to converge", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 140, + 380 + ], + "score": 1.0, + "content": "towards", + "type": "text" + }, + { + "bbox": [ + 141, + 369, + 151, + 378 + ], + "score": 0.84, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 368, + 505, + 380 + ], + "score": 1.0, + "content": ". Still, the supervised training applies much more force to the system than required, the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 378, + 345, + 392 + ], + "score": 1.0, + "content": "reasons for which become obvious when inspecting Figure", + "type": "text" + }, + { + "bbox": [ + 345, + 379, + 357, + 389 + ], + "score": 0.25, + "content": "^ \\mathrm { 4 b }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 378, + 505, + 392 + ], + "score": 1.0, + "content": "and Fig. 8e. While each state seems", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "score": 1.0, + "content": "close to the ground truth individually, the control oscillates undesirably, requiring counter-actions", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 399, + 160, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 160, + 413 + ], + "score": 1.0, + "content": "later in time.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 104, + 356, + 505, + 413 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "The methods using the differentiable solver significantly outperform their supervised counterparts", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "and exhibit an excellent performance that is very close the ground truth solutions in terms of required", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "forces. On many examples, they even reach the target state with less force than was applied by the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 450, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 465 + ], + "score": 1.0, + "content": "ground truth simulation. This would not be possible with the supervised loss alone, but by having", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "access to the gradient-based feedback from the differentiable solver, they can learn to find more", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 471, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 486 + ], + "score": 1.0, + "content": "efficient trajectories with respect to the objective loss. This allows the networks to learn applying", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 483, + 468, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 468, + 496 + ], + "score": 1.0, + "content": "forces in different locations that make the system approach the target state with less force.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 417, + 506, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 505, + 589 + ], + "lines": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "score": 1.0, + "content": "Figure 4e and Fig.8f,g show examples of this. The ground truth applies the same force in each step,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 297, + 524 + ], + "score": 1.0, + "content": "thereby continuously increasing the first sample", + "type": "text" + }, + { + "bbox": [ + 297, + 511, + 335, + 523 + ], + "score": 0.91, + "content": "u ( x = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 511, + 506, + 524 + ], + "score": 1.0, + "content": ", and the supervised method tries to imitate", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 353, + 535 + ], + "score": 1.0, + "content": "this behavior. The governing equation then slowly propagates", + "type": "text" + }, + { + "bbox": [ + 353, + 522, + 391, + 534 + ], + "score": 0.9, + "content": "u ( x = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 522, + 436, + 535 + ], + "score": 1.0, + "content": "in positive", + "type": "text" + }, + { + "bbox": [ + 436, + 524, + 443, + 532 + ], + "score": 0.74, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "direction since", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 172, + 545 + ], + "score": 0.9, + "content": "u ( x = 0 ) > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 533, + 505, + 546 + ], + "score": 1.0, + "content": ". The learning methods that use a differentiable solver make use of this fact by", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 218, + 558 + ], + "score": 1.0, + "content": "applying much more force", + "type": "text" + }, + { + "bbox": [ + 218, + 544, + 283, + 556 + ], + "score": 0.92, + "content": "F ( x = 0 ) > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "at this point than the ground truth, even overshooting", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "the target state. Later, when this value had time to propagate to the right, the model corrects this", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 268, + 578 + ], + "score": 1.0, + "content": "overshoot by applying a negative force", + "type": "text" + }, + { + "bbox": [ + 268, + 566, + 333, + 578 + ], + "score": 0.92, + "content": "F ( x = 0 ) \\ { \\overset { } { < } } \\ { \\overset { - } { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 567, + 505, + 578 + ], + "score": 1.0, + "content": ". Using this trick, these models reach the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 577, + 486, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 195, + 590 + ], + "score": 1.0, + "content": "target state with up to", + "type": "text" + }, + { + "bbox": [ + 195, + 577, + 215, + 587 + ], + "score": 0.86, + "content": "13 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 577, + 486, + 590 + ], + "score": 1.0, + "content": "less force than the ground truth on the sequence shown in Figure 4.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 499, + 506, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 504, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 504, + 607 + ], + "score": 1.0, + "content": "Figure 9 analyzes the variance of inferred forces. The supervised methods often fail to properly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 504, + 617 + ], + "score": 1.0, + "content": "converge to the target state, resulting in large forces in the last step, visible as a second peak in the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "supervised CFE chain. The formulation of the loss (Eq. 3) suppresses force spikes. In the solutions", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "inferred by our method, the likelihood of large forces falls off multi-exponentially as a consequence.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "This means that large forces are exponentially rare, which is the expected behavior given the L2", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 649, + 200, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 200, + 662 + ], + "score": 1.0, + "content": "regularizer from Eq. 3.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 594, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "We also compare our results to a single-shooting baseline which is able to find near-optimal solutions", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "at the cost of higher computation times. The classic optimization uses the ADAM optimizer with a", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "learning rate of 0.01 and converges after around 300 iterations. To reach the quality of the staggered", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "prediction scheme, it requires only around 60 iterations. This quick convergence can be explained", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "by the relatively simple setup that is dominated by linear effects. Therefore, the gradients are stable,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 288, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 288, + 732 + ], + "score": 1.0, + "content": "even when propagated through many frames.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 665, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 104, + 436, + 289 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 156, + 104, + 436, + 289 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 104, + 436, + 289 + ], + "spans": [ + { + "bbox": [ + 156, + 104, + 436, + 289 + ], + "score": 0.973, + "type": "image", + "image_path": "9f693fd80541c30a3cfb1931b91806ef6230eca90d757d1d37a380d34e00312e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 156, + 104, + 436, + 165.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 156, + 165.66666666666666, + 436, + 227.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 156, + 227.33333333333331, + 436, + 289.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 304, + 503, + 327 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "Figure 9: Histogram comparing the frequency of force strengths of different methods, summed over", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 315, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 107, + 315, + 286, + 329 + ], + "score": 1.0, + "content": "100 examples from the Burger’s experiment.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 505, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "The computation times, shown in Tab. 1, were recorded on a single GTX 1080 Ti. We run 100", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 360, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 373 + ], + "score": 1.0, + "content": "examples in parallel to reduce the relative overhead caused by GPU instruction queuing. For the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "network-based methods, we average the inference time over 100 runs. We perform 10 runs for the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 383, + 198, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 198, + 394 + ], + "score": 1.0, + "content": "optimization methods.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 108, + 410, + 263, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 264, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 264, + 422 + ], + "score": 1.0, + "content": "D.2 INCOMPRESSIBLE FLUID FLOW", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 430, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "The incompressible Navier-Stokes equations model dynamics of fluids such as water or air, which", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "can develop highly complex and chaotic behavior. The phenomenon of turbulence is generally", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "seen as one of the few remaining fundamental and unsolved problems of classical physics. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "challenging nature of the equations indicates that typically a very significant computational effort", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "and a large number of degrees of freedom are required to numerically compute solutions. Here,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 366, + 498 + ], + "score": 1.0, + "content": "we target an incompressible two-dimensional gas with viscosity", + "type": "text" + }, + { + "bbox": [ + 366, + 488, + 373, + 496 + ], + "score": 0.67, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 486, + 506, + 498 + ], + "score": 1.0, + "content": ", described by the Navier-Stokes", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 234, + 510 + ], + "score": 1.0, + "content": "equations for the velocity field", + "type": "text" + }, + { + "bbox": [ + 235, + 498, + 242, + 507 + ], + "score": 0.71, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 497, + 506, + 510 + ], + "score": 1.0, + "content": ". We assume a constant fluid density throughout the simulation,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 508, + 462, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 135, + 521 + ], + "score": 1.0, + "content": "setting", + "type": "text" + }, + { + "bbox": [ + 136, + 509, + 205, + 520 + ], + "score": 0.88, + "content": "\\rho _ { f } = \\mathrm { c o n s t . } \\equiv 1 .", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 508, + 462, + 521 + ], + "score": 1.0, + "content": ". The gas velocity is controllable and, according to Eq. 1, we set", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 535, + 387, + 562 + ], + "lines": [ + { + "bbox": [ + 223, + 535, + 387, + 562 + ], + "spans": [ + { + "bbox": [ + 223, + 535, + 387, + 562 + ], + "score": 0.92, + "content": "\\mathscr { P } ( v , \\nabla v ) = - ( v \\cdot \\nabla ) v + \\nu \\nabla ^ { 2 } v - \\frac { \\nabla p } { \\rho _ { f } }", + "type": "interline_equation", + "image_path": "fe8dd6d4d8a855f7ec851c7d3fc4fdfd74ee0aa2a3aff58a8832236d8d5cdcde.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 535, + 387, + 548.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 223, + 548.5, + 387, + 562.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 231, + 589 + ], + "score": 1.0, + "content": "subject to the hard constraints", + "type": "text" + }, + { + "bbox": [ + 232, + 577, + 277, + 587 + ], + "score": 0.91, + "content": "\\nabla \\cdot \\pmb { v } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 577, + 296, + 589 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 297, + 577, + 345, + 588 + ], + "score": 0.93, + "content": "\\nabla \\times p = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 577, + 505, + 589 + ], + "score": 1.0, + "content": ". For our experiments, we target fluids", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 266, + 600 + ], + "score": 1.0, + "content": "with low viscosities, such as air, and set", + "type": "text" + }, + { + "bbox": [ + 266, + 588, + 291, + 598 + ], + "score": 0.9, + "content": "\\nu = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "in the equation above as the transport steps implicitly", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 598, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 613 + ], + "score": 1.0, + "content": "apply numerical diffusion that is on average higher than the targeted one. For fluids with a larger", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 350, + 623 + ], + "score": 1.0, + "content": "viscosity, the Poisson solver outlined above for computing", + "type": "text" + }, + { + "bbox": [ + 350, + 612, + 357, + 622 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "could be used to implicitly solve a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 620, + 263, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 252, + 635 + ], + "score": 1.0, + "content": "vector-valued diffusion equation for", + "type": "text" + }, + { + "bbox": [ + 253, + 623, + 259, + 631 + ], + "score": 0.76, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 620, + 263, + 635 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "However, incorporating a significant amount of viscosity would make the control problem easier to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "solve for most cases, as viscosity suppresses small scale structures in the motion. Hence, in order", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "score": 1.0, + "content": "to create a challenging environment for training our networks, we have but a minimal amount of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 671, + 235, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 235, + 683 + ], + "score": 1.0, + "content": "diffusion in the physical model.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 230, + 700 + ], + "score": 1.0, + "content": "In addition to the velocity field", + "type": "text" + }, + { + "bbox": [ + 230, + 690, + 237, + 698 + ], + "score": 0.73, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 687, + 404, + 700 + ], + "score": 1.0, + "content": ", we consider a smoke density distribution", + "type": "text" + }, + { + "bbox": [ + 404, + 690, + 411, + 699 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "which moves passively", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 231, + 711 + ], + "score": 1.0, + "content": "with the fluid. The evolution of", + "type": "text" + }, + { + "bbox": [ + 231, + 700, + 237, + 711 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 698, + 349, + 711 + ], + "score": 1.0, + "content": "is described by the equation", + "type": "text" + }, + { + "bbox": [ + 349, + 699, + 418, + 711 + ], + "score": 0.91, + "content": "\\partial \\rho / \\partial t = - v \\cdot \\nabla \\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 698, + 505, + 711 + ], + "score": 1.0, + "content": ". We treat the velocity", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 436, + 723 + ], + "score": 1.0, + "content": "field as hidden from observation, letting only the smoke density be observed, i.e.", + "type": "text" + }, + { + "bbox": [ + 436, + 709, + 484, + 722 + ], + "score": 0.93, + "content": "o ( t ) = \\rho ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 709, + 506, + 723 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 720, + 463, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 196, + 733 + ], + "score": 1.0, + "content": "stack the two fields as", + "type": "text" + }, + { + "bbox": [ + 196, + 721, + 239, + 732 + ], + "score": 0.93, + "content": "u = ( v , \\rho )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 720, + 463, + 733 + ], + "score": 1.0, + "content": "to write the system as one PDE, compatible with Eq. 1.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 104, + 436, + 289 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 156, + 104, + 436, + 289 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 156, + 104, + 436, + 289 + ], + "spans": [ + { + "bbox": [ + 156, + 104, + 436, + 289 + ], + "score": 0.973, + "type": "image", + "image_path": "9f693fd80541c30a3cfb1931b91806ef6230eca90d757d1d37a380d34e00312e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 156, + 104, + 436, + 165.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 156, + 165.66666666666666, + 436, + 227.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 156, + 227.33333333333331, + 436, + 289.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 304, + 503, + 327 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "Figure 9: Histogram comparing the frequency of force strengths of different methods, summed over", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 315, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 107, + 315, + 286, + 329 + ], + "score": 1.0, + "content": "100 examples from the Burger’s experiment.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 505, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "The computation times, shown in Tab. 1, were recorded on a single GTX 1080 Ti. We run 100", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 360, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 373 + ], + "score": 1.0, + "content": "examples in parallel to reduce the relative overhead caused by GPU instruction queuing. For the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 384 + ], + "score": 1.0, + "content": "network-based methods, we average the inference time over 100 runs. We perform 10 runs for the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 383, + 198, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 198, + 394 + ], + "score": 1.0, + "content": "optimization methods.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 349, + 506, + 394 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 410, + 263, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 264, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 264, + 422 + ], + "score": 1.0, + "content": "D.2 INCOMPRESSIBLE FLUID FLOW", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 430, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 431, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 505, + 443 + ], + "score": 1.0, + "content": "The incompressible Navier-Stokes equations model dynamics of fluids such as water or air, which", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "can develop highly complex and chaotic behavior. The phenomenon of turbulence is generally", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "seen as one of the few remaining fundamental and unsolved problems of classical physics. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "challenging nature of the equations indicates that typically a very significant computational effort", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "and a large number of degrees of freedom are required to numerically compute solutions. Here,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 366, + 498 + ], + "score": 1.0, + "content": "we target an incompressible two-dimensional gas with viscosity", + "type": "text" + }, + { + "bbox": [ + 366, + 488, + 373, + 496 + ], + "score": 0.67, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 486, + 506, + 498 + ], + "score": 1.0, + "content": ", described by the Navier-Stokes", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 234, + 510 + ], + "score": 1.0, + "content": "equations for the velocity field", + "type": "text" + }, + { + "bbox": [ + 235, + 498, + 242, + 507 + ], + "score": 0.71, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 497, + 506, + 510 + ], + "score": 1.0, + "content": ". We assume a constant fluid density throughout the simulation,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 508, + 462, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 135, + 521 + ], + "score": 1.0, + "content": "setting", + "type": "text" + }, + { + "bbox": [ + 136, + 509, + 205, + 520 + ], + "score": 0.88, + "content": "\\rho _ { f } = \\mathrm { c o n s t . } \\equiv 1 .", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 508, + 462, + 521 + ], + "score": 1.0, + "content": ". The gas velocity is controllable and, according to Eq. 1, we set", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 431, + 506, + 521 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 535, + 387, + 562 + ], + "lines": [ + { + "bbox": [ + 223, + 535, + 387, + 562 + ], + "spans": [ + { + "bbox": [ + 223, + 535, + 387, + 562 + ], + "score": 0.92, + "content": "\\mathscr { P } ( v , \\nabla v ) = - ( v \\cdot \\nabla ) v + \\nu \\nabla ^ { 2 } v - \\frac { \\nabla p } { \\rho _ { f } }", + "type": "interline_equation", + "image_path": "fe8dd6d4d8a855f7ec851c7d3fc4fdfd74ee0aa2a3aff58a8832236d8d5cdcde.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 535, + 387, + 548.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 223, + 548.5, + 387, + 562.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 231, + 589 + ], + "score": 1.0, + "content": "subject to the hard constraints", + "type": "text" + }, + { + "bbox": [ + 232, + 577, + 277, + 587 + ], + "score": 0.91, + "content": "\\nabla \\cdot \\pmb { v } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 577, + 296, + 589 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 297, + 577, + 345, + 588 + ], + "score": 0.93, + "content": "\\nabla \\times p = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 577, + 505, + 589 + ], + "score": 1.0, + "content": ". For our experiments, we target fluids", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 266, + 600 + ], + "score": 1.0, + "content": "with low viscosities, such as air, and set", + "type": "text" + }, + { + "bbox": [ + 266, + 588, + 291, + 598 + ], + "score": 0.9, + "content": "\\nu = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "in the equation above as the transport steps implicitly", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 598, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 613 + ], + "score": 1.0, + "content": "apply numerical diffusion that is on average higher than the targeted one. For fluids with a larger", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 350, + 623 + ], + "score": 1.0, + "content": "viscosity, the Poisson solver outlined above for computing", + "type": "text" + }, + { + "bbox": [ + 350, + 612, + 357, + 622 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "could be used to implicitly solve a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 620, + 263, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 252, + 635 + ], + "score": 1.0, + "content": "vector-valued diffusion equation for", + "type": "text" + }, + { + "bbox": [ + 253, + 623, + 259, + 631 + ], + "score": 0.76, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 620, + 263, + 635 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 577, + 506, + 635 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "However, incorporating a significant amount of viscosity would make the control problem easier to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "solve for most cases, as viscosity suppresses small scale structures in the motion. Hence, in order", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "score": 1.0, + "content": "to create a challenging environment for training our networks, we have but a minimal amount of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 671, + 235, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 235, + 683 + ], + "score": 1.0, + "content": "diffusion in the physical model.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 638, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 230, + 700 + ], + "score": 1.0, + "content": "In addition to the velocity field", + "type": "text" + }, + { + "bbox": [ + 230, + 690, + 237, + 698 + ], + "score": 0.73, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 687, + 404, + 700 + ], + "score": 1.0, + "content": ", we consider a smoke density distribution", + "type": "text" + }, + { + "bbox": [ + 404, + 690, + 411, + 699 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "which moves passively", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 231, + 711 + ], + "score": 1.0, + "content": "with the fluid. The evolution of", + "type": "text" + }, + { + "bbox": [ + 231, + 700, + 237, + 711 + ], + "score": 0.83, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 698, + 349, + 711 + ], + "score": 1.0, + "content": "is described by the equation", + "type": "text" + }, + { + "bbox": [ + 349, + 699, + 418, + 711 + ], + "score": 0.91, + "content": "\\partial \\rho / \\partial t = - v \\cdot \\nabla \\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 698, + 505, + 711 + ], + "score": 1.0, + "content": ". We treat the velocity", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 436, + 723 + ], + "score": 1.0, + "content": "field as hidden from observation, letting only the smoke density be observed, i.e.", + "type": "text" + }, + { + "bbox": [ + 436, + 709, + 484, + 722 + ], + "score": 0.93, + "content": "o ( t ) = \\rho ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 709, + 506, + 723 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 720, + 463, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 196, + 733 + ], + "score": 1.0, + "content": "stack the two fields as", + "type": "text" + }, + { + "bbox": [ + 196, + 721, + 239, + 732 + ], + "score": 0.93, + "content": "u = ( v , \\rho )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 720, + 463, + 733 + ], + "score": 1.0, + "content": "to write the system as one PDE, compatible with Eq. 1.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 114, + 145, + 298, + 158 + ], + "lines": [ + { + "bbox": [ + 112, + 144, + 299, + 160 + ], + "spans": [ + { + "bbox": [ + 112, + 144, + 299, + 160 + ], + "score": 1.0, + "content": "(a) Supervised, staggered execution", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 156, + 502, + 628 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 156, + 502, + 628 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 156, + 502, + 628 + ], + "spans": [ + { + "bbox": [ + 110, + 156, + 502, + 628 + ], + "score": 0.946, + "type": "image", + "image_path": "20c06b665ffa6d65e5d618435cf74ebc4ccf18e07da0c09d642a701d7a59f3df.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 110, + 156, + 502, + 313.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 313.33333333333337, + 502, + 470.66666666666674 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 110, + 470.66666666666674, + 502, + 628.0000000000001 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 637, + 504, + 671 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "Figure 10: Reconstruction of an example natural flow sequence from the test set. The predictions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 647, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 118, + 659 + ], + "score": 0.83, + "content": "\\pmb { o } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 647, + 272, + 661 + ], + "score": 1.0, + "content": "are plotted above the reconstructions", + "type": "text" + }, + { + "bbox": [ + 272, + 651, + 279, + 658 + ], + "score": 0.59, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 647, + 351, + 661 + ], + "score": 1.0, + "content": ". The target state", + "type": "text" + }, + { + "bbox": [ + 351, + 649, + 362, + 659 + ], + "score": 0.86, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 647, + 506, + 661 + ], + "score": 1.0, + "content": "is shown in the last column of the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 659, + 156, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 156, + 672 + ], + "score": 1.0, + "content": "predictions.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.5 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 114, + 145, + 298, + 158 + ], + "lines": [ + { + "bbox": [ + 112, + 144, + 299, + 160 + ], + "spans": [ + { + "bbox": [ + 112, + 144, + 299, + 160 + ], + "score": 1.0, + "content": "(a) Supervised, staggered execution", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 112, + 144, + 299, + 160 + ] + }, + { + "type": "image", + "bbox": [ + 110, + 156, + 502, + 628 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 156, + 502, + 628 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 156, + 502, + 628 + ], + "spans": [ + { + "bbox": [ + 110, + 156, + 502, + 628 + ], + "score": 0.946, + "type": "image", + "image_path": "20c06b665ffa6d65e5d618435cf74ebc4ccf18e07da0c09d642a701d7a59f3df.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 110, + 156, + 502, + 313.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 313.33333333333337, + 502, + 470.66666666666674 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 110, + 470.66666666666674, + 502, + 628.0000000000001 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 637, + 504, + 671 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "Figure 10: Reconstruction of an example natural flow sequence from the test set. The predictions", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 647, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 118, + 659 + ], + "score": 0.83, + "content": "\\pmb { o } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 647, + 272, + 661 + ], + "score": 1.0, + "content": "are plotted above the reconstructions", + "type": "text" + }, + { + "bbox": [ + 272, + 651, + 279, + 658 + ], + "score": 0.59, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 647, + 351, + 661 + ], + "score": 1.0, + "content": ". The target state", + "type": "text" + }, + { + "bbox": [ + 351, + 649, + 362, + 659 + ], + "score": 0.86, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 647, + 506, + 661 + ], + "score": 1.0, + "content": "is shown in the last column of the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 659, + 156, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 156, + 672 + ], + "score": 1.0, + "content": "predictions.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "For the OP and CFE networks, we use the 2D network architecture described in Appendix C. Instead", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "of directly generating the velocity update in the CFE network for this problem setup, we make use", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 457, + 117 + ], + "score": 1.0, + "content": "of stream functions (Lamb, 1932). Hence, the CFE network outputs a vector potential", + "type": "text" + }, + { + "bbox": [ + 457, + 105, + 466, + 114 + ], + "score": 0.82, + "content": "\\Phi", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "of which", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 139, + 128 + ], + "score": 1.0, + "content": "the curl", + "type": "text" + }, + { + "bbox": [ + 140, + 115, + 168, + 126 + ], + "score": 0.91, + "content": "\\nabla \\times \\Phi", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "is used as a velocity update. This setup numerically simplifies the incompressibility", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "condition of the Navier-Stokes equations but retains the same number of effective control parame-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 138, + 126, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 126, + 149 + ], + "score": 1.0, + "content": "ters.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "Datasets. We generate training and test datasets for two distinct tasks: flow reconstruction and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 317, + 178 + ], + "score": 1.0, + "content": "shape transition. Both datasets have a resolution of", + "type": "text" + }, + { + "bbox": [ + 317, + 165, + 361, + 176 + ], + "score": 0.9, + "content": "1 2 8 \\times 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "with the velocity fields being sam-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "pled in staggered form (see Appendix A). This results in over 16.000 effective continuous control", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 363, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 275, + 200 + ], + "score": 1.0, + "content": "parameters that make up the control force", + "type": "text" + }, + { + "bbox": [ + 275, + 187, + 299, + 199 + ], + "score": 0.92, + "content": "\\pmb { F } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 187, + 354, + 200 + ], + "score": 1.0, + "content": "for each step", + "type": "text" + }, + { + "bbox": [ + 354, + 188, + 358, + 197 + ], + "score": 0.58, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 187, + 363, + 200 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "score": 1.0, + "content": "The flow reconstruction dataset is comprised of ground-truth sequences where the initial states", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 107, + 215, + 139, + 227 + ], + "score": 0.92, + "content": "( \\rho _ { 0 } , v _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "are randomly sampled and then simulated for 64 time steps. The resulting smoke density is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 234, + 239 + ], + "score": 1.0, + "content": "then taken to be the target state,", + "type": "text" + }, + { + "bbox": [ + 234, + 226, + 317, + 238 + ], + "score": 0.92, + "content": "o ^ { * } \\equiv \\rho ^ { * } = \\rho ^ { \\mathrm { s i m } } ( t _ { 6 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 225, + 505, + 239 + ], + "score": 1.0, + "content": ". Since we use fully convolutional networks for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "both CFE and OPs, the open domain boundary must be handled carefully. If smoke was lost from", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "score": 1.0, + "content": "the simulation, because it crossed the outer boundary, a neural network would see the smoke simply", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "score": 1.0, + "content": "vanish unless it was explicitly given the domain size as input. To avoid these problems, we run the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 270, + 477, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 338, + 281 + ], + "score": 1.0, + "content": "simulation backwards in time and remove all smoke from", + "type": "text" + }, + { + "bbox": [ + 338, + 271, + 349, + 281 + ], + "score": 0.84, + "content": "\\rho _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 270, + 477, + 281 + ], + "score": 1.0, + "content": "that left the simulation domain.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 372, + 300 + ], + "score": 1.0, + "content": "For the shape transition dataset, we sample initial and target states", + "type": "text" + }, + { + "bbox": [ + 372, + 289, + 383, + 298 + ], + "score": 0.86, + "content": "\\rho _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 285, + 401, + 300 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 288, + 412, + 298 + ], + "score": 0.86, + "content": "\\rho _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 285, + 505, + 300 + ], + "score": 1.0, + "content": "by randomly choosing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "a shape from a library containing ten basic geometric shapes and placing it at a random location", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 460, + 321 + ], + "score": 1.0, + "content": "inside the domain. These can then be used for reconstructing sequences of any length", + "type": "text" + }, + { + "bbox": [ + 460, + 311, + 467, + 318 + ], + "score": 0.66, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 308, + 505, + 321 + ], + "score": 1.0, + "content": ". For the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 319, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 345, + 333 + ], + "score": 1.0, + "content": "results on shape transition presented in section 6, we choose", + "type": "text" + }, + { + "bbox": [ + 346, + 320, + 377, + 330 + ], + "score": 0.89, + "content": "n = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 319, + 505, + 333 + ], + "score": 1.0, + "content": "because all interesting behavior", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "can be seen within that time frame. Due to the linear interpolation used in the advection step (see", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 185, + 354 + ], + "score": 1.0, + "content": "Appendix A), both", + "type": "text" + }, + { + "bbox": [ + 185, + 344, + 192, + 353 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 342, + 210, + 354 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 211, + 344, + 218, + 352 + ], + "score": 0.71, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 342, + 506, + 354 + ], + "score": 1.0, + "content": "smear out over time. This numerical limitation makes it impossible to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 353, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 364 + ], + "score": 1.0, + "content": "match target states exactly in this task as the density will become blurry over time. While we could", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "score": 1.0, + "content": "generate ground-truth sequences using a classical optimizer, we refrain from doing so because (i)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "these trajectories are not guaranteed to be optimal and (ii) we want to see how well the model can", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 385, + 272, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 272, + 397 + ], + "score": 1.0, + "content": "learn from scratch, without initialization.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 433, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 435, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 435, + 417 + ], + "score": 1.0, + "content": "Training. We pretrain the CFE on the natural flow dataset with a supervised loss,", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 418, + 385, + 435 + ], + "lines": [ + { + "bbox": [ + 226, + 418, + 385, + 435 + ], + "spans": [ + { + "bbox": [ + 226, + 418, + 385, + 435 + ], + "score": 0.92, + "content": "L _ { \\mathrm { s u p } } ^ { \\mathrm { C F E } } ( \\boldsymbol { u } ( t ) ) = | \\boldsymbol { v } _ { \\boldsymbol { u } ( t ) } + \\boldsymbol { F } ( t ) - \\boldsymbol { v } ^ { * } ( t ) | ^ { 2 }", + "type": "interline_equation", + "image_path": "556cc85e0827c4efd9f7e58a1fdacc6fd61c0c6f533546b417944bd8979ea43a.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 226, + 418, + 385, + 435 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 504, + 483 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 133, + 452 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 440, + 156, + 452 + ], + "score": 0.92, + "content": "{ \\pmb v } ^ { * } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 439, + 504, + 452 + ], + "score": 1.0, + "content": "denotes the velocity from ground truth sequences. This supervised training alone con-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 431, + 463 + ], + "score": 1.0, + "content": "stitutes a good loss for the CFE as it only needs to consider single-step intervals", + "type": "text" + }, + { + "bbox": [ + 432, + 451, + 445, + 461 + ], + "score": 0.86, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "while the OPs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "score": 1.0, + "content": "handle longer sequences. Nevertheless, we found that using the differentiable solver with an obser-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 471, + 155, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 155, + 486 + ], + "score": 1.0, + "content": "vation loss,", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 481, + 410, + 496 + ], + "lines": [ + { + "bbox": [ + 200, + 481, + 410, + 496 + ], + "spans": [ + { + "bbox": [ + 200, + 481, + 410, + 496 + ], + "score": 0.89, + "content": "{ \\cal L } _ { o } ^ { \\mathrm { C F E } } = | B _ { r } ( { \\pmb o } ^ { * } ) - B _ { r } ( \\mathrm { S o l v e r } [ { \\pmb u } + \\mathrm { C F E } [ { \\pmb u } , { \\pmb o } ^ { * } ] ] ) | ^ { 2 } ,", + "type": "interline_equation", + "image_path": "9447117977b16fb73d06628ef25398c04d289e83a9ce126a0adef37170ed9717.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 200, + 481, + 410, + 496 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "further improves the accuracy of the inferred force without sacrificing the ground truth match. Here", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 509, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 107, + 509, + 133, + 522 + ], + "score": 0.92, + "content": "B _ { r } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 509, + 334, + 523 + ], + "score": 1.0, + "content": "denotes a blur function with a kernel of the form", + "type": "text" + }, + { + "bbox": [ + 335, + 509, + 360, + 524 + ], + "score": 0.93, + "content": "\\frac { 1 } { 1 + x / r }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 509, + 506, + 523 + ], + "score": 1.0, + "content": ". The blur helps make the gradients", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 521, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 537 + ], + "score": 1.0, + "content": "smoother and creates non-zero gradients in places where prediction and target do not overlap. During", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 263, + 545 + ], + "score": 1.0, + "content": "training, we start with a large radius of", + "type": "text" + }, + { + "bbox": [ + 263, + 533, + 308, + 543 + ], + "score": 0.89, + "content": "r = 1 6 \\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 534, + 323, + 545 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 323, + 534, + 336, + 544 + ], + "score": 0.91, + "content": "B _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 534, + 461, + 545 + ], + "score": 1.0, + "content": "and successively decrease it to", + "type": "text" + }, + { + "bbox": [ + 461, + 533, + 501, + 543 + ], + "score": 0.89, + "content": "r = 2 \\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 534, + 505, + 545 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 155, + 558 + ], + "score": 1.0, + "content": "We choose", + "type": "text" + }, + { + "bbox": [ + 155, + 546, + 163, + 554 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 543, + 206, + 558 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 206, + 545, + 221, + 555 + ], + "score": 0.89, + "content": "L _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 543, + 242, + 558 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 242, + 545, + 255, + 556 + ], + "score": 0.9, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "are of the same magnitude when the force loss spikes (see", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 554, + 144, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 144, + 568 + ], + "score": 1.0, + "content": "Fig. 15).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 571, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "After the CFE is trained, we successively train the OPs starting with the smallest time scale. For", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "the OPs, we train different models for natural flow reconstruction and shape transition, both based", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "on the same CFE model. We pre-train all OPs independently with a supervised observation loss", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "before jointly training them end-to-end with objective loss function (Eq. 4) and the differentiable", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "solver to find the optimal trajectory. We use the OPs trained with the staggered execution scheme", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "as initialization for the prediction refinement scheme. The complexity of solving the Navier-Stokes", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 636, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 506, + 653 + ], + "score": 1.0, + "content": "equations over many time steps in this example requires such a fully supervised initialization step.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "Without it, this setting is so non-linear that the learning process does not converge to a good so-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 658, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 658, + 505, + 674 + ], + "score": 1.0, + "content": "lution. Hence, it illustrates the importance of combining supervised and unsupervised (requiring", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 670, + 375, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 375, + 684 + ], + "score": 1.0, + "content": "differentiable physics) training for challenging learning objectives.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "A comparison of the different losses is shown in Fig. 10. The predictions, shown in the top rows", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "of each subfigure, illustrate the differences between the three methods. The supervised predictions,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "especially the long-term predictions (central images), are blurry because the network learns to aver-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "age over all ground truth sequences that match the given initial and target state. The differentiable", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51.5 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "22", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "For the OP and CFE networks, we use the 2D network architecture described in Appendix C. Instead", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "of directly generating the velocity update in the CFE network for this problem setup, we make use", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 457, + 117 + ], + "score": 1.0, + "content": "of stream functions (Lamb, 1932). Hence, the CFE network outputs a vector potential", + "type": "text" + }, + { + "bbox": [ + 457, + 105, + 466, + 114 + ], + "score": 0.82, + "content": "\\Phi", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 104, + 506, + 117 + ], + "score": 1.0, + "content": "of which", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 139, + 128 + ], + "score": 1.0, + "content": "the curl", + "type": "text" + }, + { + "bbox": [ + 140, + 115, + 168, + 126 + ], + "score": 0.91, + "content": "\\nabla \\times \\Phi", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "is used as a velocity update. This setup numerically simplifies the incompressibility", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "condition of the Navier-Stokes equations but retains the same number of effective control parame-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 138, + 126, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 126, + 149 + ], + "score": 1.0, + "content": "ters.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 81, + 506, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "Datasets. We generate training and test datasets for two distinct tasks: flow reconstruction and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 317, + 178 + ], + "score": 1.0, + "content": "shape transition. Both datasets have a resolution of", + "type": "text" + }, + { + "bbox": [ + 317, + 165, + 361, + 176 + ], + "score": 0.9, + "content": "1 2 8 \\times 1 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 164, + 506, + 178 + ], + "score": 1.0, + "content": "with the velocity fields being sam-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "pled in staggered form (see Appendix A). This results in over 16.000 effective continuous control", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 363, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 275, + 200 + ], + "score": 1.0, + "content": "parameters that make up the control force", + "type": "text" + }, + { + "bbox": [ + 275, + 187, + 299, + 199 + ], + "score": 0.92, + "content": "\\pmb { F } ( t _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 187, + 354, + 200 + ], + "score": 1.0, + "content": "for each step", + "type": "text" + }, + { + "bbox": [ + 354, + 188, + 358, + 197 + ], + "score": 0.58, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 187, + 363, + 200 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 154, + 506, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "score": 1.0, + "content": "The flow reconstruction dataset is comprised of ground-truth sequences where the initial states", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 107, + 215, + 139, + 227 + ], + "score": 0.92, + "content": "( \\rho _ { 0 } , v _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "are randomly sampled and then simulated for 64 time steps. The resulting smoke density is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 234, + 239 + ], + "score": 1.0, + "content": "then taken to be the target state,", + "type": "text" + }, + { + "bbox": [ + 234, + 226, + 317, + 238 + ], + "score": 0.92, + "content": "o ^ { * } \\equiv \\rho ^ { * } = \\rho ^ { \\mathrm { s i m } } ( t _ { 6 4 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 225, + 505, + 239 + ], + "score": 1.0, + "content": ". Since we use fully convolutional networks for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "both CFE and OPs, the open domain boundary must be handled carefully. If smoke was lost from", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "score": 1.0, + "content": "the simulation, because it crossed the outer boundary, a neural network would see the smoke simply", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "score": 1.0, + "content": "vanish unless it was explicitly given the domain size as input. To avoid these problems, we run the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 270, + 477, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 338, + 281 + ], + "score": 1.0, + "content": "simulation backwards in time and remove all smoke from", + "type": "text" + }, + { + "bbox": [ + 338, + 271, + 349, + 281 + ], + "score": 0.84, + "content": "\\rho _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 270, + 477, + 281 + ], + "score": 1.0, + "content": "that left the simulation domain.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 203, + 506, + 281 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 372, + 300 + ], + "score": 1.0, + "content": "For the shape transition dataset, we sample initial and target states", + "type": "text" + }, + { + "bbox": [ + 372, + 289, + 383, + 298 + ], + "score": 0.86, + "content": "\\rho _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 285, + 401, + 300 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 288, + 412, + 298 + ], + "score": 0.86, + "content": "\\rho _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 285, + 505, + 300 + ], + "score": 1.0, + "content": "by randomly choosing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "a shape from a library containing ten basic geometric shapes and placing it at a random location", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 460, + 321 + ], + "score": 1.0, + "content": "inside the domain. These can then be used for reconstructing sequences of any length", + "type": "text" + }, + { + "bbox": [ + 460, + 311, + 467, + 318 + ], + "score": 0.66, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 308, + 505, + 321 + ], + "score": 1.0, + "content": ". For the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 319, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 345, + 333 + ], + "score": 1.0, + "content": "results on shape transition presented in section 6, we choose", + "type": "text" + }, + { + "bbox": [ + 346, + 320, + 377, + 330 + ], + "score": 0.89, + "content": "n = 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 319, + 505, + 333 + ], + "score": 1.0, + "content": "because all interesting behavior", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "can be seen within that time frame. Due to the linear interpolation used in the advection step (see", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 342, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 185, + 354 + ], + "score": 1.0, + "content": "Appendix A), both", + "type": "text" + }, + { + "bbox": [ + 185, + 344, + 192, + 353 + ], + "score": 0.81, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 342, + 210, + 354 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 211, + 344, + 218, + 352 + ], + "score": 0.71, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 342, + 506, + 354 + ], + "score": 1.0, + "content": "smear out over time. This numerical limitation makes it impossible to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 353, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 364 + ], + "score": 1.0, + "content": "match target states exactly in this task as the density will become blurry over time. While we could", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "score": 1.0, + "content": "generate ground-truth sequences using a classical optimizer, we refrain from doing so because (i)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "these trajectories are not guaranteed to be optimal and (ii) we want to see how well the model can", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 385, + 272, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 272, + 397 + ], + "score": 1.0, + "content": "learn from scratch, without initialization.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 285, + 506, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 433, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 435, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 435, + 417 + ], + "score": 1.0, + "content": "Training. We pretrain the CFE on the natural flow dataset with a supervised loss,", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 400, + 435, + 417 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 418, + 385, + 435 + ], + "lines": [ + { + "bbox": [ + 226, + 418, + 385, + 435 + ], + "spans": [ + { + "bbox": [ + 226, + 418, + 385, + 435 + ], + "score": 0.92, + "content": "L _ { \\mathrm { s u p } } ^ { \\mathrm { C F E } } ( \\boldsymbol { u } ( t ) ) = | \\boldsymbol { v } _ { \\boldsymbol { u } ( t ) } + \\boldsymbol { F } ( t ) - \\boldsymbol { v } ^ { * } ( t ) | ^ { 2 }", + "type": "interline_equation", + "image_path": "556cc85e0827c4efd9f7e58a1fdacc6fd61c0c6f533546b417944bd8979ea43a.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 226, + 418, + 385, + 435 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 504, + 483 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 504, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 133, + 452 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 440, + 156, + 452 + ], + "score": 0.92, + "content": "{ \\pmb v } ^ { * } ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 439, + 504, + 452 + ], + "score": 1.0, + "content": "denotes the velocity from ground truth sequences. This supervised training alone con-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 431, + 463 + ], + "score": 1.0, + "content": "stitutes a good loss for the CFE as it only needs to consider single-step intervals", + "type": "text" + }, + { + "bbox": [ + 432, + 451, + 445, + 461 + ], + "score": 0.86, + "content": "\\Delta t", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "while the OPs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "score": 1.0, + "content": "handle longer sequences. Nevertheless, we found that using the differentiable solver with an obser-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 471, + 155, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 155, + 486 + ], + "score": 1.0, + "content": "vation loss,", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 439, + 506, + 486 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 481, + 410, + 496 + ], + "lines": [ + { + "bbox": [ + 200, + 481, + 410, + 496 + ], + "spans": [ + { + "bbox": [ + 200, + 481, + 410, + 496 + ], + "score": 0.89, + "content": "{ \\cal L } _ { o } ^ { \\mathrm { C F E } } = | B _ { r } ( { \\pmb o } ^ { * } ) - B _ { r } ( \\mathrm { S o l v e r } [ { \\pmb u } + \\mathrm { C F E } [ { \\pmb u } , { \\pmb o } ^ { * } ] ] ) | ^ { 2 } ,", + "type": "interline_equation", + "image_path": "9447117977b16fb73d06628ef25398c04d289e83a9ce126a0adef37170ed9717.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 200, + 481, + 410, + 496 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "further improves the accuracy of the inferred force without sacrificing the ground truth match. Here", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 509, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 107, + 509, + 133, + 522 + ], + "score": 0.92, + "content": "B _ { r } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 509, + 334, + 523 + ], + "score": 1.0, + "content": "denotes a blur function with a kernel of the form", + "type": "text" + }, + { + "bbox": [ + 335, + 509, + 360, + 524 + ], + "score": 0.93, + "content": "\\frac { 1 } { 1 + x / r }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 509, + 506, + 523 + ], + "score": 1.0, + "content": ". The blur helps make the gradients", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 521, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 537 + ], + "score": 1.0, + "content": "smoother and creates non-zero gradients in places where prediction and target do not overlap. During", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 263, + 545 + ], + "score": 1.0, + "content": "training, we start with a large radius of", + "type": "text" + }, + { + "bbox": [ + 263, + 533, + 308, + 543 + ], + "score": 0.89, + "content": "r = 1 6 \\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 534, + 323, + 545 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 323, + 534, + 336, + 544 + ], + "score": 0.91, + "content": "B _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 534, + 461, + 545 + ], + "score": 1.0, + "content": "and successively decrease it to", + "type": "text" + }, + { + "bbox": [ + 461, + 533, + 501, + 543 + ], + "score": 0.89, + "content": "r = 2 \\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 534, + 505, + 545 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 155, + 558 + ], + "score": 1.0, + "content": "We choose", + "type": "text" + }, + { + "bbox": [ + 155, + 546, + 163, + 554 + ], + "score": 0.77, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 543, + 206, + 558 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 206, + 545, + 221, + 555 + ], + "score": 0.89, + "content": "L _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 543, + 242, + 558 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 242, + 545, + 255, + 556 + ], + "score": 0.9, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "are of the same magnitude when the force loss spikes (see", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 554, + 144, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 144, + 568 + ], + "score": 1.0, + "content": "Fig. 15).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 498, + 506, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 571, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "After the CFE is trained, we successively train the OPs starting with the smallest time scale. For", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "the OPs, we train different models for natural flow reconstruction and shape transition, both based", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "on the same CFE model. We pre-train all OPs independently with a supervised observation loss", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "before jointly training them end-to-end with objective loss function (Eq. 4) and the differentiable", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "solver to find the optimal trajectory. We use the OPs trained with the staggered execution scheme", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "as initialization for the prediction refinement scheme. The complexity of solving the Navier-Stokes", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 636, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 506, + 653 + ], + "score": 1.0, + "content": "equations over many time steps in this example requires such a fully supervised initialization step.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "Without it, this setting is so non-linear that the learning process does not converge to a good so-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 658, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 658, + 505, + 674 + ], + "score": 1.0, + "content": "lution. Hence, it illustrates the importance of combining supervised and unsupervised (requiring", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 670, + 375, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 375, + 684 + ], + "score": 1.0, + "content": "differentiable physics) training for challenging learning objectives.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 44.5, + "bbox_fs": [ + 104, + 572, + 506, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "A comparison of the different losses is shown in Fig. 10. The predictions, shown in the top rows", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "of each subfigure, illustrate the differences between the three methods. The supervised predictions,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "especially the long-term predictions (central images), are blurry because the network learns to aver-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "age over all ground truth sequences that match the given initial and target state. The differentiable", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 114, + 84, + 233, + 96 + ], + "lines": [ + { + "bbox": [ + 113, + 82, + 234, + 100 + ], + "spans": [ + { + "bbox": [ + 113, + 82, + 234, + 100 + ], + "score": 1.0, + "content": "(a) Predicted trajectory", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 94, + 502, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 94, + 502, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 94, + 502, + 198 + ], + "spans": [ + { + "bbox": [ + 111, + 94, + 502, + 198 + ], + "score": 0.943, + "type": "image", + "image_path": "f6300ec4109fa9b0a44a256fe5fa33c143c0456be46eaba8451f6da96edfcc82.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 111, + 94, + 502, + 128.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 128.66666666666666, + 502, + 163.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 163.33333333333331, + 502, + 197.99999999999997 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 209, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Figure 11: Reconstruction of multiple shapes using prediction refinement. The CFE is trained on the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "natural flow dataset and OPs are trained on single-shape transitions. The predictions of all shapes", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 232, + 312, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 312, + 244 + ], + "score": 1.0, + "content": "are added and passed to the CFE as one prediction.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 279 + ], + "score": 1.0, + "content": "physics solver largely resolves this issue. The predictions are much sharper but the long-term pre-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 279, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 505, + 289 + ], + "score": 1.0, + "content": "dictions still do not account for short-term deviations. This can be seen in the central prediction of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 289, + 504, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 294, + 302 + ], + "score": 1.0, + "content": "Fig. 10b which shows hints of the target state", + "type": "text" + }, + { + "bbox": [ + 295, + 290, + 306, + 300 + ], + "score": 0.85, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 289, + 496, + 302 + ], + "score": 1.0, + "content": ", despite the fact that the actual reconstruction", + "type": "text" + }, + { + "bbox": [ + 496, + 291, + 504, + 299 + ], + "score": 0.51, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 299, + 504, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 496, + 313 + ], + "score": 1.0, + "content": "cannot reach that state at that time. The refined prediction, shown in subfigure (c), is closer to", + "type": "text" + }, + { + "bbox": [ + 496, + 303, + 504, + 310 + ], + "score": 0.57, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 311, + 339, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 339, + 324 + ], + "score": 1.0, + "content": "since it is conditioned on the previous reconstructed state.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 340 + ], + "score": 1.0, + "content": "In the training data, we let the network transform one shape into another at a random location. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "differentiable solver and the long-term intuition provided by our execution scheme make it possible", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 350, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 505, + 363 + ], + "score": 1.0, + "content": "to train networks that can infer accurate sequences of control forces. In most cases, the target shapes", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "are closely matched. As our networks infer sequences over time, we refer readers to the supplemental", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 372, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 383 + ], + "score": 1.0, + "content": "material (https://ge.in.tum.de/publications/2020-iclr-holl), which contains animations of additional", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 384, + 152, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 152, + 396 + ], + "score": 1.0, + "content": "sequences.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Generalization to multiple shapes. Splitting the reconstruction task into prediction and correction", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 411, + 504, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 417, + 423 + ], + "score": 1.0, + "content": "has the additional benefit of having full access to the intermediate predictions", + "type": "text" + }, + { + "bbox": [ + 418, + 411, + 429, + 421 + ], + "score": 0.85, + "content": "\\pmb { o } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 411, + 504, + 423 + ], + "score": 1.0, + "content": ". These model real", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "states of the system so classical processing or filter operations can be applied to them as well. We", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 298, + 446 + ], + "score": 1.0, + "content": "demonstrate this by generalizing our method to", + "type": "text" + }, + { + "bbox": [ + 298, + 433, + 327, + 443 + ], + "score": 0.9, + "content": "m > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "shapes that evolve within the same domain.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "Figure 11 shows an example of two weakly-interacting shape transitions. We implement this by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 323, + 468 + ], + "score": 1.0, + "content": "executing the OPs independently for each transition", + "type": "text" + }, + { + "bbox": [ + 323, + 455, + 392, + 467 + ], + "score": 0.94, + "content": "\\bar { k \\ } \\in \\ \\bar { \\{ 1 , 2 , . . . m \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "while inferring the control", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 464, + 504, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 130, + 480 + ], + "score": 1.0, + "content": "force", + "type": "text" + }, + { + "bbox": [ + 130, + 466, + 151, + 478 + ], + "score": 0.92, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 464, + 497, + 480 + ], + "score": 1.0, + "content": "on the joint system. This is achieved by adding the predictions of the smoke density", + "type": "text" + }, + { + "bbox": [ + 497, + 468, + 504, + 478 + ], + "score": 0.73, + "content": "\\rho", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 102, + 474, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 102, + 474, + 264, + 493 + ], + "score": 1.0, + "content": "before passing it to the CFE network,", + "type": "text" + }, + { + "bbox": [ + 265, + 476, + 330, + 489 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\tilde { \\pmb { o } ^ { p } } = \\sum _ { k = 1 } ^ { m ^ { * } } \\pmb { o } _ { k } ^ { p } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 474, + 505, + 493 + ], + "score": 1.0, + "content": ". The resulting force is then applied to all", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "sequences individually so that smoke from one transition does not end up in another target state.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "Using this scheme, we can define start and end positions for arbitrarily many shapes and let them", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 510, + 172, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 172, + 523 + ], + "score": 1.0, + "content": "evolve together.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "Evaluation of force strengths The average force strengths are detailed in Tab. 2 while Figure 12", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "gives a more detailed analysis of the force strengths. As expected from using a L2 regularizer on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "the force, large values are exponentially rare in the solutions inferred from our test set. None of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "the hierarchical execution schemes exhibit large outliers. The prediction refinement requires the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "least amount of force to match the target, slightly ahead of the staggered execution trained with the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "same loss. The supervised training produces trajectories with reduced continuity that result in larger", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 592, + 192, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 192, + 605 + ], + "score": 1.0, + "content": "forces being applied.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 349, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 351, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 351, + 634 + ], + "score": 1.0, + "content": "D.3 INCOMPRESSIBLE FLUID WITH INDIRECT CONTROL", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "As a fourth test environment, we target a case with increased complexity, where the network does", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 656, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 667 + ], + "score": 1.0, + "content": "not have the means anymore to directly control the full fluid volume. Instead, the network can", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "only apply forces in the peripheral regions, with a total of more than 5000 control parameters per", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "step. The obstacles prevent fluid from passing through them and the domain is enclosed with solid", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "boundaries from the left, right and bottom. This leads to additional hard constraints and interplays", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "between constraints in the physical model, and as such provides an interesting and challenging test", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "case for our method. The domain has three target regions (buckets) separated by walls at the top of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "the domain, into which a volume of smoke should be transported from any position in the center", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40.5 + } + ], + "page_idx": 22, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 114, + 84, + 233, + 96 + ], + "lines": [ + { + "bbox": [ + 113, + 82, + 234, + 100 + ], + "spans": [ + { + "bbox": [ + 113, + 82, + 234, + 100 + ], + "score": 1.0, + "content": "(a) Predicted trajectory", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 113, + 82, + 234, + 100 + ] + }, + { + "type": "image", + "bbox": [ + 111, + 94, + 502, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 94, + 502, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 94, + 502, + 198 + ], + "spans": [ + { + "bbox": [ + 111, + 94, + 502, + 198 + ], + "score": 0.943, + "type": "image", + "image_path": "f6300ec4109fa9b0a44a256fe5fa33c143c0456be46eaba8451f6da96edfcc82.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 111, + 94, + 502, + 128.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 128.66666666666666, + 502, + 163.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 163.33333333333331, + 502, + 197.99999999999997 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 209, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Figure 11: Reconstruction of multiple shapes using prediction refinement. The CFE is trained on the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "natural flow dataset and OPs are trained on single-shape transitions. The predictions of all shapes", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 232, + 312, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 312, + 244 + ], + "score": 1.0, + "content": "are added and passed to the CFE as one prediction.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 505, + 279 + ], + "score": 1.0, + "content": "physics solver largely resolves this issue. The predictions are much sharper but the long-term pre-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 279, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 505, + 289 + ], + "score": 1.0, + "content": "dictions still do not account for short-term deviations. This can be seen in the central prediction of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 289, + 504, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 294, + 302 + ], + "score": 1.0, + "content": "Fig. 10b which shows hints of the target state", + "type": "text" + }, + { + "bbox": [ + 295, + 290, + 306, + 300 + ], + "score": 0.85, + "content": "o ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 289, + 496, + 302 + ], + "score": 1.0, + "content": ", despite the fact that the actual reconstruction", + "type": "text" + }, + { + "bbox": [ + 496, + 291, + 504, + 299 + ], + "score": 0.51, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 299, + 504, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 496, + 313 + ], + "score": 1.0, + "content": "cannot reach that state at that time. The refined prediction, shown in subfigure (c), is closer to", + "type": "text" + }, + { + "bbox": [ + 496, + 303, + 504, + 310 + ], + "score": 0.57, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 311, + 339, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 339, + 324 + ], + "score": 1.0, + "content": "since it is conditioned on the previous reconstructed state.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 267, + 505, + 324 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 340 + ], + "score": 1.0, + "content": "In the training data, we let the network transform one shape into another at a random location. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "differentiable solver and the long-term intuition provided by our execution scheme make it possible", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 350, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 505, + 363 + ], + "score": 1.0, + "content": "to train networks that can infer accurate sequences of control forces. In most cases, the target shapes", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "are closely matched. As our networks infer sequences over time, we refer readers to the supplemental", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 372, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 383 + ], + "score": 1.0, + "content": "material (https://ge.in.tum.de/publications/2020-iclr-holl), which contains animations of additional", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 384, + 152, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 152, + 396 + ], + "score": 1.0, + "content": "sequences.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 329, + 505, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Generalization to multiple shapes. Splitting the reconstruction task into prediction and correction", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 411, + 504, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 417, + 423 + ], + "score": 1.0, + "content": "has the additional benefit of having full access to the intermediate predictions", + "type": "text" + }, + { + "bbox": [ + 418, + 411, + 429, + 421 + ], + "score": 0.85, + "content": "\\pmb { o } ^ { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 411, + 504, + 423 + ], + "score": 1.0, + "content": ". These model real", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "states of the system so classical processing or filter operations can be applied to them as well. We", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 298, + 446 + ], + "score": 1.0, + "content": "demonstrate this by generalizing our method to", + "type": "text" + }, + { + "bbox": [ + 298, + 433, + 327, + 443 + ], + "score": 0.9, + "content": "m > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "shapes that evolve within the same domain.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "Figure 11 shows an example of two weakly-interacting shape transitions. We implement this by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 323, + 468 + ], + "score": 1.0, + "content": "executing the OPs independently for each transition", + "type": "text" + }, + { + "bbox": [ + 323, + 455, + 392, + 467 + ], + "score": 0.94, + "content": "\\bar { k \\ } \\in \\ \\bar { \\{ 1 , 2 , . . . m \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "while inferring the control", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 464, + 504, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 130, + 480 + ], + "score": 1.0, + "content": "force", + "type": "text" + }, + { + "bbox": [ + 130, + 466, + 151, + 478 + ], + "score": 0.92, + "content": "\\mathbf { } F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 464, + 497, + 480 + ], + "score": 1.0, + "content": "on the joint system. This is achieved by adding the predictions of the smoke density", + "type": "text" + }, + { + "bbox": [ + 497, + 468, + 504, + 478 + ], + "score": 0.73, + "content": "\\rho", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 102, + 474, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 102, + 474, + 264, + 493 + ], + "score": 1.0, + "content": "before passing it to the CFE network,", + "type": "text" + }, + { + "bbox": [ + 265, + 476, + 330, + 489 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\tilde { \\pmb { o } ^ { p } } = \\sum _ { k = 1 } ^ { m ^ { * } } \\pmb { o } _ { k } ^ { p } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 474, + 505, + 493 + ], + "score": 1.0, + "content": ". The resulting force is then applied to all", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "sequences individually so that smoke from one transition does not end up in another target state.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "Using this scheme, we can define start and end positions for arbitrarily many shapes and let them", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 510, + 172, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 172, + 523 + ], + "score": 1.0, + "content": "evolve together.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23, + "bbox_fs": [ + 102, + 400, + 506, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "Evaluation of force strengths The average force strengths are detailed in Tab. 2 while Figure 12", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "gives a more detailed analysis of the force strengths. As expected from using a L2 regularizer on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "the force, large values are exponentially rare in the solutions inferred from our test set. None of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "the hierarchical execution schemes exhibit large outliers. The prediction refinement requires the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "least amount of force to match the target, slightly ahead of the staggered execution trained with the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "same loss. The supervised training produces trajectories with reduced continuity that result in larger", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 592, + 192, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 192, + 605 + ], + "score": 1.0, + "content": "forces being applied.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 525, + 506, + 605 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 349, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 351, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 351, + 634 + ], + "score": 1.0, + "content": "D.3 INCOMPRESSIBLE FLUID WITH INDIRECT CONTROL", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "As a fourth test environment, we target a case with increased complexity, where the network does", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 656, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 667 + ], + "score": 1.0, + "content": "not have the means anymore to directly control the full fluid volume. Instead, the network can", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "only apply forces in the peripheral regions, with a total of more than 5000 control parameters per", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "step. The obstacles prevent fluid from passing through them and the domain is enclosed with solid", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "boundaries from the left, right and bottom. This leads to additional hard constraints and interplays", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "between constraints in the physical model, and as such provides an interesting and challenging test", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "case for our method. The domain has three target regions (buckets) separated by walls at the top of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "the domain, into which a volume of smoke should be transported from any position in the center", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 256, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 505, + 270 + ], + "score": 1.0, + "content": "part. Both initial position and the target bucket are randomized for our training set of 3600 examples", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 267, + 389, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 389, + 280 + ], + "score": 1.0, + "content": "and test set of 100 examples. Each sequence consists of 16 time steps.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 644, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 134, + 434, + 319 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 155, + 134, + 434, + 319 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 155, + 134, + 434, + 319 + ], + "spans": [ + { + "bbox": [ + 155, + 134, + 434, + 319 + ], + "score": 0.925, + "type": "image", + "image_path": "2b9254a24433f718e82c560f601544338677f446c0e546092eb4791011aba195.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 155, + 134, + 434, + 195.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 155, + 195.66666666666666, + 434, + 257.3333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 155, + 257.3333333333333, + 434, + 319.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 334, + 504, + 358 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "Figure 12: Histogram comparing the frequency of force strengths applied in the direct fluid control", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 345, + 380, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 380, + 358 + ], + "score": 1.0, + "content": "experiment on the natural flow dataset, summed over 100 examples.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 108, + 423, + 502, + 677 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 423, + 502, + 677 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 423, + 502, + 677 + ], + "spans": [ + { + "bbox": [ + 108, + 423, + 502, + 677 + ], + "score": 0.973, + "type": "image", + "image_path": "7ce4b93ab1b93d35c2c8e72d6a62834b8590c4a106862d27585cf6b8af8c3f9f.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 423, + 502, + 507.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 507.6666666666667, + 502, + 592.3333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 592.3333333333334, + 502, + 677.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 688, + 499, + 700 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 687, + 500, + 702 + ], + "spans": [ + { + "bbox": [ + 110, + 687, + 500, + 702 + ], + "score": 1.0, + "content": "Figure 13: Two reconstructed trajectories from the test set of the indirect smoke control problem.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 7.0 + } + ], + "page_idx": 23, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 134, + 434, + 319 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 155, + 134, + 434, + 319 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 155, + 134, + 434, + 319 + ], + "spans": [ + { + "bbox": [ + 155, + 134, + 434, + 319 + ], + "score": 0.925, + "type": "image", + "image_path": "2b9254a24433f718e82c560f601544338677f446c0e546092eb4791011aba195.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 155, + 134, + 434, + 195.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 155, + 195.66666666666666, + 434, + 257.3333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 155, + 257.3333333333333, + 434, + 319.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 334, + 504, + 358 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "Figure 12: Histogram comparing the frequency of force strengths applied in the direct fluid control", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 345, + 380, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 380, + 358 + ], + "score": 1.0, + "content": "experiment on the natural flow dataset, summed over 100 examples.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 108, + 423, + 502, + 677 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 423, + 502, + 677 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 423, + 502, + 677 + ], + "spans": [ + { + "bbox": [ + 108, + 423, + 502, + 677 + ], + "score": 0.973, + "type": "image", + "image_path": "7ce4b93ab1b93d35c2c8e72d6a62834b8590c4a106862d27585cf6b8af8c3f9f.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 423, + 502, + 507.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 507.6666666666667, + 502, + 592.3333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 592.3333333333334, + 502, + 677.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 688, + 499, + 700 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 687, + 500, + 702 + ], + "spans": [ + { + "bbox": [ + 110, + 687, + 500, + 702 + ], + "score": 1.0, + "content": "Figure 13: Two reconstructed trajectories from the test set of the indirect smoke control problem.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 7.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 230, + 79, + 385, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 230, + 79, + 385, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 230, + 79, + 385, + 188 + ], + "spans": [ + { + "bbox": [ + 230, + 79, + 385, + 188 + ], + "score": 0.961, + "type": "image", + "image_path": "881c602cfd6572e4460b1d256d477216aa7d06cd96572e2a1acc2260694cbf71.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 79, + 385, + 92.625 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 230, + 92.625, + 385, + 106.25 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 230, + 106.25, + 385, + 119.875 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 230, + 119.875, + 385, + 133.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 230, + 133.5, + 385, + 147.125 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 230, + 147.125, + 385, + 160.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 230, + 160.75, + 385, + 174.375 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 230, + 174.375, + 385, + 188.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 200, + 505, + 233 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "Figure 14: Iterative optimization of the indirect incompressible fluid control problem. The graph", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 209, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 224 + ], + "score": 1.0, + "content": "shows the fraction of smoke that ends up in the correct bucket vs number of optimization steps,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 219, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 219, + 234 + ], + "score": 1.0, + "content": "averaged over 10 examples.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 107, + 257, + 503, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 256, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 505, + 270 + ], + "score": 1.0, + "content": "part. Both initial position and the target bucket are randomized for our training set of 3600 examples", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 267, + 389, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 389, + 280 + ], + "score": 1.0, + "content": "and test set of 100 examples. Each sequence consists of 16 time steps.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 285, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 297 + ], + "score": 1.0, + "content": "In this case the control is indirect since the smoke density lies outside the controlled area at all", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 296, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 505, + 307 + ], + "score": 1.0, + "content": "times. Only the incompressibility condition allows the network to influence the velocity outside", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 305, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 320 + ], + "score": 1.0, + "content": "the controlled area. This forces the model to consider the global context and synchronize a large", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "number of parameters to create a desired flow field. The requirement of complex synchronized force", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "fields makes generating reliable training data difficult, as manual or random sampling is unlikely", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "to produce a directed velocity field in the center. We therefore skip the pretraining process and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "directly train the CFE using the differentiable solver, while the OP networks are trained as before", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 360, + 171, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 126, + 374 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 362, + 167, + 372 + ], + "score": 0.88, + "content": "r = 2 \\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 360, + 171, + 374 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 378, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 391 + ], + "score": 1.0, + "content": "To evaluate how well the learning method performs, we measure how much of the smoke density", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "ends up inside the buckets and how much force was applied in total. For reference, we replace the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 401, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 412 + ], + "score": 1.0, + "content": "observation predictions with an algorithm that moves the smoke towards the bucket in a straight", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 411, + 504, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 452, + 424 + ], + "score": 1.0, + "content": "line. Averaged over 100 examples from the test set, the resulting model manages to put", + "type": "text" + }, + { + "bbox": [ + 453, + 411, + 504, + 422 + ], + "score": 0.91, + "content": "8 9 \\% \\pm 2 . \\bar { 6 } \\%", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "of the smoke into the target bucket. In contrast, the model trained with our full algorithm moves", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 433, + 491, + 446 + ], + "spans": [ + { + "bbox": [ + 107, + 433, + 177, + 444 + ], + "score": 0.92, + "content": "9 9 . 2 2 \\% \\pm 0 . 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 433, + 386, + 446 + ], + "score": 1.0, + "content": "of the smoke into the target buckets while requiring", + "type": "text" + }, + { + "bbox": [ + 386, + 433, + 447, + 444 + ], + "score": 0.91, + "content": "1 9 . 1 \\% \\pm 1 . 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 433, + 491, + 446 + ], + "score": 1.0, + "content": "less force.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "We also compare our method to an iterative optimization which directly optimizes the control ve-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "locities. We use the ADAM optimizer with a learning rate of 0.1. Despite the highly non-linear", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "score": 1.0, + "content": "setup, the gradients are stable enough to quickly let the smoke flow in the right direction. Fig. 14", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 483, + 504, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 504, + 495 + ], + "score": 1.0, + "content": "shows how the trajectories improve during optimization. After around 60 optimization steps, the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 494, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 507 + ], + "score": 1.0, + "content": "smoke distribution starts reaching the target bucket in some examples. Over the next 600 iterations,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 505, + 491, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 273, + 517 + ], + "score": 1.0, + "content": "it converges to a a configuration in which", + "type": "text" + }, + { + "bbox": [ + 273, + 505, + 317, + 516 + ], + "score": 0.89, + "content": "8 2 . 1 \\pm 7 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 505, + 491, + 517 + ], + "score": 1.0, + "content": "of the smoke ends up in the correct bucket.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 108, + 533, + 299, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 301, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 301, + 546 + ], + "score": 1.0, + "content": "D.4 COMPARISON TO SHOOTING METHODS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 555, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "We compare the sequences inferred by our trained models to classical shooting optimizations using", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 327, + 579 + ], + "score": 1.0, + "content": "our differentiable physics solver to directly optimize", + "type": "text" + }, + { + "bbox": [ + 327, + 566, + 348, + 578 + ], + "score": 0.94, + "content": "F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 566, + 447, + 579 + ], + "score": 1.0, + "content": "with the objective loss", + "type": "text" + }, + { + "bbox": [ + 447, + 567, + 456, + 576 + ], + "score": 0.59, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "(Eq. 4) for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "score": 1.0, + "content": "a single input. We make use of stream functions (Lamb, 1932), as in the second experiment, to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "ensure the incompressibility condition is fulfilled. For this comparison, the velocities of all steps are", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 598, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 275, + 612 + ], + "score": 1.0, + "content": "initialized with a normal distribution with", + "type": "text" + }, + { + "bbox": [ + 275, + 600, + 301, + 610 + ], + "score": 0.9, + "content": "\\mu = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 598, + 319, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 319, + 599, + 357, + 610 + ], + "score": 0.88, + "content": "\\sigma = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 598, + 506, + 612 + ], + "score": 1.0, + "content": "so that the initial trajectory does not", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 299, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 242, + 622 + ], + "score": 1.0, + "content": "significantly alter the initial state,", + "type": "text" + }, + { + "bbox": [ + 243, + 610, + 295, + 622 + ], + "score": 0.93, + "content": "u ( t ) \\approx u ( t _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 610, + 299, + 622 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "We first show how a simple single-shooting algorithm (Zhou et al., 1996) fares with our Navier-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "Stokes setup. When solving the resulting optimization problem using single-shooting, strong ar-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "tifacts in the reconstructions can be observed, as shown in Figure 17a. This undesirable behavior", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 660, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 463, + 673 + ], + "score": 1.0, + "content": "stems from the nonlinearity of the Navier-Stokes equations, which causes the gradients", + "type": "text" + }, + { + "bbox": [ + 464, + 660, + 504, + 671 + ], + "score": 0.89, + "content": "\\Delta \\boldsymbol { u } \\gg 0", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "to become noisy and unreliable when they are recurrently backpropagated through many time steps.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 682, + 463, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 463, + 694 + ], + "score": 1.0, + "content": "Unsurprisingly, the single-shooting optimizer converges to a undesirable local minimum.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "As single-shooting is well known to have problems with non-trivial problem settings, we employ a", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "multi-scale shooting (MS) method (Hartmann et al., 2014). This solver first computes the trajectory", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "on a coarsely discretized version of the problem before iteratively refining the discretization. For", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + } + ], + "page_idx": 24, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 230, + 79, + 385, + 188 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 230, + 79, + 385, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 230, + 79, + 385, + 188 + ], + "spans": [ + { + "bbox": [ + 230, + 79, + 385, + 188 + ], + "score": 0.961, + "type": "image", + "image_path": "881c602cfd6572e4460b1d256d477216aa7d06cd96572e2a1acc2260694cbf71.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 230, + 79, + 385, + 92.625 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 230, + 92.625, + 385, + 106.25 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 230, + 106.25, + 385, + 119.875 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 230, + 119.875, + 385, + 133.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 230, + 133.5, + 385, + 147.125 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 230, + 147.125, + 385, + 160.75 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 230, + 160.75, + 385, + 174.375 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 230, + 174.375, + 385, + 188.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 200, + 505, + 233 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "Figure 14: Iterative optimization of the indirect incompressible fluid control problem. The graph", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 209, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 224 + ], + "score": 1.0, + "content": "shows the fraction of smoke that ends up in the correct bucket vs number of optimization steps,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 219, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 219, + 234 + ], + "score": 1.0, + "content": "averaged over 10 examples.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 107, + 257, + 503, + 279 + ], + "lines": [], + "index": 11.5, + "bbox_fs": [ + 105, + 256, + 505, + 280 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 285, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 297 + ], + "score": 1.0, + "content": "In this case the control is indirect since the smoke density lies outside the controlled area at all", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 296, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 505, + 307 + ], + "score": 1.0, + "content": "times. Only the incompressibility condition allows the network to influence the velocity outside", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 305, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 320 + ], + "score": 1.0, + "content": "the controlled area. This forces the model to consider the global context and synchronize a large", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "number of parameters to create a desired flow field. The requirement of complex synchronized force", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "fields makes generating reliable training data difficult, as manual or random sampling is unlikely", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "to produce a directed velocity field in the center. We therefore skip the pretraining process and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "directly train the CFE using the differentiable solver, while the OP networks are trained as before", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 360, + 171, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 126, + 374 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 362, + 167, + 372 + ], + "score": 0.88, + "content": "r = 2 \\Delta x", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 360, + 171, + 374 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 285, + 506, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 378, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 391 + ], + "score": 1.0, + "content": "To evaluate how well the learning method performs, we measure how much of the smoke density", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "ends up inside the buckets and how much force was applied in total. For reference, we replace the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 401, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 412 + ], + "score": 1.0, + "content": "observation predictions with an algorithm that moves the smoke towards the bucket in a straight", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 411, + 504, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 452, + 424 + ], + "score": 1.0, + "content": "line. Averaged over 100 examples from the test set, the resulting model manages to put", + "type": "text" + }, + { + "bbox": [ + 453, + 411, + 504, + 422 + ], + "score": 0.91, + "content": "8 9 \\% \\pm 2 . \\bar { 6 } \\%", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "of the smoke into the target bucket. In contrast, the model trained with our full algorithm moves", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 433, + 491, + 446 + ], + "spans": [ + { + "bbox": [ + 107, + 433, + 177, + 444 + ], + "score": 0.92, + "content": "9 9 . 2 2 \\% \\pm 0 . 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 433, + 386, + 446 + ], + "score": 1.0, + "content": "of the smoke into the target buckets while requiring", + "type": "text" + }, + { + "bbox": [ + 386, + 433, + 447, + 444 + ], + "score": 0.91, + "content": "1 9 . 1 \\% \\pm 1 . 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 433, + 491, + 446 + ], + "score": 1.0, + "content": "less force.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 378, + 506, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "We also compare our method to an iterative optimization which directly optimizes the control ve-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "locities. We use the ADAM optimizer with a learning rate of 0.1. Despite the highly non-linear", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "score": 1.0, + "content": "setup, the gradients are stable enough to quickly let the smoke flow in the right direction. Fig. 14", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 483, + 504, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 504, + 495 + ], + "score": 1.0, + "content": "shows how the trajectories improve during optimization. After around 60 optimization steps, the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 494, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 507 + ], + "score": 1.0, + "content": "smoke distribution starts reaching the target bucket in some examples. Over the next 600 iterations,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 505, + 491, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 273, + 517 + ], + "score": 1.0, + "content": "it converges to a a configuration in which", + "type": "text" + }, + { + "bbox": [ + 273, + 505, + 317, + 516 + ], + "score": 0.89, + "content": "8 2 . 1 \\pm 7 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 505, + 491, + 517 + ], + "score": 1.0, + "content": "of the smoke ends up in the correct bucket.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 450, + 506, + 517 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 533, + 299, + 544 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 301, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 301, + 546 + ], + "score": 1.0, + "content": "D.4 COMPARISON TO SHOOTING METHODS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 555, + 505, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "We compare the sequences inferred by our trained models to classical shooting optimizations using", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 327, + 579 + ], + "score": 1.0, + "content": "our differentiable physics solver to directly optimize", + "type": "text" + }, + { + "bbox": [ + 327, + 566, + 348, + 578 + ], + "score": 0.94, + "content": "F ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 566, + 447, + 579 + ], + "score": 1.0, + "content": "with the objective loss", + "type": "text" + }, + { + "bbox": [ + 447, + 567, + 456, + 576 + ], + "score": 0.59, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "(Eq. 4) for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 590 + ], + "score": 1.0, + "content": "a single input. We make use of stream functions (Lamb, 1932), as in the second experiment, to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 506, + 601 + ], + "score": 1.0, + "content": "ensure the incompressibility condition is fulfilled. For this comparison, the velocities of all steps are", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 598, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 275, + 612 + ], + "score": 1.0, + "content": "initialized with a normal distribution with", + "type": "text" + }, + { + "bbox": [ + 275, + 600, + 301, + 610 + ], + "score": 0.9, + "content": "\\mu = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 598, + 319, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 319, + 599, + 357, + 610 + ], + "score": 0.88, + "content": "\\sigma = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 598, + 506, + 612 + ], + "score": 1.0, + "content": "so that the initial trajectory does not", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 299, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 242, + 622 + ], + "score": 1.0, + "content": "significantly alter the initial state,", + "type": "text" + }, + { + "bbox": [ + 243, + 610, + 295, + 622 + ], + "score": 0.93, + "content": "u ( t ) \\approx u ( t _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 610, + 299, + 622 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 554, + 506, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "We first show how a simple single-shooting algorithm (Zhou et al., 1996) fares with our Navier-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "Stokes setup. When solving the resulting optimization problem using single-shooting, strong ar-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "tifacts in the reconstructions can be observed, as shown in Figure 17a. This undesirable behavior", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 660, + 504, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 463, + 673 + ], + "score": 1.0, + "content": "stems from the nonlinearity of the Navier-Stokes equations, which causes the gradients", + "type": "text" + }, + { + "bbox": [ + 464, + 660, + 504, + 671 + ], + "score": 0.89, + "content": "\\Delta \\boldsymbol { u } \\gg 0", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "to become noisy and unreliable when they are recurrently backpropagated through many time steps.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 682, + 463, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 463, + 694 + ], + "score": 1.0, + "content": "Unsurprisingly, the single-shooting optimizer converges to a undesirable local minimum.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 626, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "As single-shooting is well known to have problems with non-trivial problem settings, we employ a", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "multi-scale shooting (MS) method (Hartmann et al., 2014). This solver first computes the trajectory", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "on a coarsely discretized version of the problem before iteratively refining the discretization. For", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "the first resolution, we use 1/16 of the original width and height which both reduces the number", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "of control parameters and reduces nonlinear effects from the physics model. By employing an", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "exponential learning rate decay, this multi-scale optimization converges reliably for all examples.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "We use the ADAM optimizer to compute the control variable updates from the gradients of the", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 439, + 250, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 250, + 452 + ], + "score": 1.0, + "content": "differentiable Navier-Stokes solver.", + "type": "text", + "cross_page": true + } + ], + "index": 13 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 81, + 457, + 189 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 149, + 81, + 457, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 149, + 81, + 457, + 189 + ], + "spans": [ + { + "bbox": [ + 149, + 81, + 457, + 189 + ], + "score": 0.941, + "type": "image", + "image_path": "a35f4c51deab2577e34560efeae92f303f1d17b9afb6e054f7dcf87dc12b4ccc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 149, + 81, + 457, + 117.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 149, + 117.0, + 457, + 153.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 149, + 153.0, + 457, + 189.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 201, + 502, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 199, + 502, + 215 + ], + "spans": [ + { + "bbox": [ + 108, + 199, + 502, + 215 + ], + "score": 1.0, + "content": "Figure 15: Mean convergence curves of the adjoint method optimization for 100 shape transitions.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 148, + 228, + 451, + 338 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 148, + 228, + 451, + 338 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 148, + 228, + 451, + 338 + ], + "spans": [ + { + "bbox": [ + 148, + 228, + 451, + 338 + ], + "score": 0.964, + "type": "image", + "image_path": "2ba7229c2a13450b6f2c8ddd6faf5efd14e1cb7f4bc750c7d76728352be6b08b.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 148, + 228, + 451, + 264.6666666666667 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 148, + 264.6666666666667, + 451, + 301.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 148, + 301.33333333333337, + 451, + 338.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 348, + 504, + 371 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 347, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 362 + ], + "score": 1.0, + "content": "Figure 16: Mean convergence curves of the adjoint method optimization for 100 shape transitions,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 359, + 386, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 386, + 372 + ], + "score": 1.0, + "content": "taking the reconstruction from the refinement scheme as initial guess.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "the first resolution, we use 1/16 of the original width and height which both reduces the number", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "of control parameters and reduces nonlinear effects from the physics model. By employing an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "exponential learning rate decay, this multi-scale optimization converges reliably for all examples.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "We use the ADAM optimizer to compute the control variable updates from the gradients of the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 439, + 250, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 250, + 452 + ], + "score": 1.0, + "content": "differentiable Navier-Stokes solver.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "An averaged set of representative convergence curves for this setup is shown in Figure 15. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 460, + 480 + ], + "score": 1.0, + "content": "objective loss (Eq. 4) is shown in its decomposed state as the sum of the observation loss", + "type": "text" + }, + { + "bbox": [ + 460, + 468, + 473, + 479 + ], + "score": 0.9, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 468, + 505, + 480 + ], + "score": 1.0, + "content": ", shown", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 477, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 237, + 492 + ], + "score": 1.0, + "content": "in Figure 15a, and the force loss", + "type": "text" + }, + { + "bbox": [ + 237, + 479, + 252, + 489 + ], + "score": 0.87, + "content": "L _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 477, + 506, + 492 + ], + "score": 1.0, + "content": ", shown in Figure 15b. Due to the initialization of all velocities", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 446, + 502 + ], + "score": 1.0, + "content": "with small values, the force loss starts out small. For the first 1000 iteration steps,", + "type": "text" + }, + { + "bbox": [ + 447, + 489, + 460, + 501 + ], + "score": 0.88, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "dominates", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 338, + 513 + ], + "score": 1.0, + "content": "which causes the system to move towards the target state", + "type": "text" + }, + { + "bbox": [ + 339, + 501, + 350, + 511 + ], + "score": 0.86, + "content": "\\ b { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 500, + 505, + 513 + ], + "score": 1.0, + "content": ". This trajectory is not ideal, however,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "as more force than necessary is applied. Once observation loss and force loss are of the same", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 523, + 374, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 374, + 535 + ], + "score": 1.0, + "content": "magnitude, the optimization refines the trajectory to use less force.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 539, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "We found that the trajectories predicted by our neural network based method correspond to perform-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "ing about 1500 steps with the MS optimization while requiring less tuning. Reconstructions of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "score": 1.0, + "content": "same example are compared in Figure 17. Performing the MS optimization up to this point took", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "131 seconds on a GTX 1080 Ti graphics card for a single 16-frame sequence while the network", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 581, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 506, + 596 + ], + "score": 1.0, + "content": "inference ran for 0.5 seconds. For longer sequences, this gap grows further because the network", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 213, + 606 + ], + "score": 1.0, + "content": "inference time scales with", + "type": "text" + }, + { + "bbox": [ + 213, + 594, + 235, + 605 + ], + "score": 0.9, + "content": "{ \\mathcal { O } } ( n )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 594, + 505, + 606 + ], + "score": 1.0, + "content": ". This could only be matched if the number of iterations for the MS", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 208, + 618 + ], + "score": 1.0, + "content": "optimization scaled with", + "type": "text" + }, + { + "bbox": [ + 209, + 605, + 230, + 617 + ], + "score": 0.89, + "content": "O ( 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 605, + 506, + 618 + ], + "score": 1.0, + "content": ", which is not the case for most problems. These tests indicate that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "our model has successfully internalized the behavior of a large class of physical behavior, and can", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "exert the right amount of force to reach the intended goal. The large number of iterations required", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 639, + 485, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 485, + 650 + ], + "score": 1.0, + "content": "for the single-case shooting optimization highlights the complexity of the individual solutions.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 666 + ], + "score": 1.0, + "content": "Interestingly, the network also benefits from the much more difficult task to learn a whole manifold", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "of solutions: comparing solutions with similar observation loss for the MS algorithm and our net-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "work, the former often finds solutions that are unintuitive and contain noticeable detours, e.g., not", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "taking a straight path for the density matching examples of Fig. 5. In such situations, our network", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "benefits from having to represent the solution manifold, instead of aiming for single task optimiza-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "tions. As the solutions are changing relatively smoothly, the complex task effectively regularizes the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "inference of new solutions and gives the network a more global view. Instead, the shooting optimiza-", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + } + ], + "page_idx": 25, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 81, + 457, + 189 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 149, + 81, + 457, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 149, + 81, + 457, + 189 + ], + "spans": [ + { + "bbox": [ + 149, + 81, + 457, + 189 + ], + "score": 0.941, + "type": "image", + "image_path": "a35f4c51deab2577e34560efeae92f303f1d17b9afb6e054f7dcf87dc12b4ccc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 149, + 81, + 457, + 117.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 149, + 117.0, + 457, + 153.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 149, + 153.0, + 457, + 189.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 201, + 502, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 199, + 502, + 215 + ], + "spans": [ + { + "bbox": [ + 108, + 199, + 502, + 215 + ], + "score": 1.0, + "content": "Figure 15: Mean convergence curves of the adjoint method optimization for 100 shape transitions.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 148, + 228, + 451, + 338 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 148, + 228, + 451, + 338 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 148, + 228, + 451, + 338 + ], + "spans": [ + { + "bbox": [ + 148, + 228, + 451, + 338 + ], + "score": 0.964, + "type": "image", + "image_path": "2ba7229c2a13450b6f2c8ddd6faf5efd14e1cb7f4bc750c7d76728352be6b08b.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 148, + 228, + 451, + 264.6666666666667 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 148, + 264.6666666666667, + 451, + 301.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 148, + 301.33333333333337, + 451, + 338.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 348, + 504, + 371 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 347, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 362 + ], + "score": 1.0, + "content": "Figure 16: Mean convergence curves of the adjoint method optimization for 100 shape transitions,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 359, + 386, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 386, + 372 + ], + "score": 1.0, + "content": "taking the reconstruction from the refinement scheme as initial guess.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 451 + ], + "lines": [], + "index": 11, + "bbox_fs": [ + 105, + 395, + 505, + 452 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "An averaged set of representative convergence curves for this setup is shown in Figure 15. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 460, + 480 + ], + "score": 1.0, + "content": "objective loss (Eq. 4) is shown in its decomposed state as the sum of the observation loss", + "type": "text" + }, + { + "bbox": [ + 460, + 468, + 473, + 479 + ], + "score": 0.9, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 468, + 505, + 480 + ], + "score": 1.0, + "content": ", shown", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 477, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 237, + 492 + ], + "score": 1.0, + "content": "in Figure 15a, and the force loss", + "type": "text" + }, + { + "bbox": [ + 237, + 479, + 252, + 489 + ], + "score": 0.87, + "content": "L _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 477, + 506, + 492 + ], + "score": 1.0, + "content": ", shown in Figure 15b. Due to the initialization of all velocities", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 446, + 502 + ], + "score": 1.0, + "content": "with small values, the force loss starts out small. For the first 1000 iteration steps,", + "type": "text" + }, + { + "bbox": [ + 447, + 489, + 460, + 501 + ], + "score": 0.88, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "dominates", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 338, + 513 + ], + "score": 1.0, + "content": "which causes the system to move towards the target state", + "type": "text" + }, + { + "bbox": [ + 339, + 501, + 350, + 511 + ], + "score": 0.86, + "content": "\\ b { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 500, + 505, + 513 + ], + "score": 1.0, + "content": ". This trajectory is not ideal, however,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "as more force than necessary is applied. Once observation loss and force loss are of the same", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 523, + 374, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 374, + 535 + ], + "score": 1.0, + "content": "magnitude, the optimization refines the trajectory to use less force.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 456, + 506, + 535 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 539, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "We found that the trajectories predicted by our neural network based method correspond to perform-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "ing about 1500 steps with the MS optimization while requiring less tuning. Reconstructions of the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 573 + ], + "score": 1.0, + "content": "same example are compared in Figure 17. Performing the MS optimization up to this point took", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "131 seconds on a GTX 1080 Ti graphics card for a single 16-frame sequence while the network", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 581, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 506, + 596 + ], + "score": 1.0, + "content": "inference ran for 0.5 seconds. For longer sequences, this gap grows further because the network", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 213, + 606 + ], + "score": 1.0, + "content": "inference time scales with", + "type": "text" + }, + { + "bbox": [ + 213, + 594, + 235, + 605 + ], + "score": 0.9, + "content": "{ \\mathcal { O } } ( n )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 594, + 505, + 606 + ], + "score": 1.0, + "content": ". This could only be matched if the number of iterations for the MS", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 208, + 618 + ], + "score": 1.0, + "content": "optimization scaled with", + "type": "text" + }, + { + "bbox": [ + 209, + 605, + 230, + 617 + ], + "score": 0.89, + "content": "O ( 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 605, + 506, + 618 + ], + "score": 1.0, + "content": ", which is not the case for most problems. These tests indicate that", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "our model has successfully internalized the behavior of a large class of physical behavior, and can", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "exert the right amount of force to reach the intended goal. The large number of iterations required", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 639, + 485, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 485, + 650 + ], + "score": 1.0, + "content": "for the single-case shooting optimization highlights the complexity of the individual solutions.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 538, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 666 + ], + "score": 1.0, + "content": "Interestingly, the network also benefits from the much more difficult task to learn a whole manifold", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "of solutions: comparing solutions with similar observation loss for the MS algorithm and our net-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "work, the former often finds solutions that are unintuitive and contain noticeable detours, e.g., not", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "taking a straight path for the density matching examples of Fig. 5. In such situations, our network", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "benefits from having to represent the solution manifold, instead of aiming for single task optimiza-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "tions. As the solutions are changing relatively smoothly, the complex task effectively regularizes the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "inference of new solutions and gives the network a more global view. Instead, the shooting optimiza-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "score": 1.0, + "content": "tions have to purely rely on local gradients for single-shooting or manually crafted multi-resolution", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 389, + 176, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 176, + 401 + ], + "score": 1.0, + "content": "schemes for MS.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 655, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 83, + 464, + 302 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 147, + 83, + 464, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 83, + 464, + 302 + ], + "spans": [ + { + "bbox": [ + 147, + 83, + 464, + 302 + ], + "score": 0.976, + "type": "image", + "image_path": "1f836a2a4cf6af34fbb92e278991e6ed0639e7ff3201c757ecc5042351a8d534.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 147, + 83, + 464, + 156.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 147, + 156.0, + 464, + 229.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 147, + 229.0, + 464, + 302.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 313, + 506, + 358 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "Figure 17: Example reconstruction of a shape transition. (a) Direct shooting optimization, 2300", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "iterations, (b) multi-scale shooting optimization, 1500 iterations, (c) output of our neural network", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "based method with prediction refinement. Our model infers the shown solution in a single pass, and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 347, + 255, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 255, + 360 + ], + "score": 1.0, + "content": "generalizes to a large class of inputs.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 378, + 504, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 391 + ], + "score": 1.0, + "content": "tions have to purely rely on local gradients for single-shooting or manually crafted multi-resolution", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 389, + 176, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 176, + 401 + ], + "score": 1.0, + "content": "schemes for MS.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "Our method can also be employed to support the MS optimization by initializing it with the velocities", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 362, + 429 + ], + "score": 1.0, + "content": "inferred by the networks. In this case, shown in Figure 16, both", + "type": "text" + }, + { + "bbox": [ + 362, + 418, + 376, + 429 + ], + "score": 0.89, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 417, + 393, + 429 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 394, + 418, + 409, + 429 + ], + "score": 0.89, + "content": "L _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "decrease right from the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "beginning, similar to the behavior in Figure 15 from iteration 1500 on. The reconstructed trajectory", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 438, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 453 + ], + "score": 1.0, + "content": "from the neural-network-based method is so close to the optimum that the multi-resolution approach", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 449, + 240, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 240, + 464 + ], + "score": 1.0, + "content": "described above is not necessary.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 475, + 229, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 230, + 488 + ], + "score": 1.0, + "content": "D.5 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "In Fig. 18, we provide a visual overview of a sub-set of the sequences that can be found in the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "supplemental materials. It contains 16 randomly selected reconstructions for each of the natural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 518, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 529 + ], + "score": 1.0, + "content": "flow, the shape transitions, and the indirect control examples. In addition, the supplemental mate-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "rial, available at https://ge.in.tum.de/publications/2020-iclr-holl, highlights the differences between", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 540, + 357, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 357, + 552 + ], + "score": 1.0, + "content": "unsupervised, staggered, and refined versions of our approach.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + } + ], + "page_idx": 26, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 83, + 464, + 302 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 147, + 83, + 464, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 83, + 464, + 302 + ], + "spans": [ + { + "bbox": [ + 147, + 83, + 464, + 302 + ], + "score": 0.976, + "type": "image", + "image_path": "1f836a2a4cf6af34fbb92e278991e6ed0639e7ff3201c757ecc5042351a8d534.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 147, + 83, + 464, + 156.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 147, + 156.0, + 464, + 229.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 147, + 229.0, + 464, + 302.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 313, + 506, + 358 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "Figure 17: Example reconstruction of a shape transition. (a) Direct shooting optimization, 2300", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "score": 1.0, + "content": "iterations, (b) multi-scale shooting optimization, 1500 iterations, (c) output of our neural network", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "based method with prediction refinement. Our model infers the shown solution in a single pass, and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 347, + 255, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 255, + 360 + ], + "score": 1.0, + "content": "generalizes to a large class of inputs.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 378, + 504, + 401 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 378, + 506, + 401 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "Our method can also be employed to support the MS optimization by initializing it with the velocities", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 362, + 429 + ], + "score": 1.0, + "content": "inferred by the networks. In this case, shown in Figure 16, both", + "type": "text" + }, + { + "bbox": [ + 362, + 418, + 376, + 429 + ], + "score": 0.89, + "content": "L _ { o } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 417, + 393, + 429 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 394, + 418, + 409, + 429 + ], + "score": 0.89, + "content": "L _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "decrease right from the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "beginning, similar to the behavior in Figure 15 from iteration 1500 on. The reconstructed trajectory", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 438, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 453 + ], + "score": 1.0, + "content": "from the neural-network-based method is so close to the optimum that the multi-resolution approach", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 449, + 240, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 240, + 464 + ], + "score": 1.0, + "content": "described above is not necessary.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 406, + 505, + 464 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 475, + 229, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 230, + 488 + ], + "score": 1.0, + "content": "D.5 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "In Fig. 18, we provide a visual overview of a sub-set of the sequences that can be found in the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "supplemental materials. It contains 16 randomly selected reconstructions for each of the natural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 518, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 529 + ], + "score": 1.0, + "content": "flow, the shape transitions, and the indirect control examples. In addition, the supplemental mate-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "rial, available at https://ge.in.tum.de/publications/2020-iclr-holl, highlights the differences between", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 540, + 357, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 357, + 552 + ], + "score": 1.0, + "content": "unsupervised, staggered, and refined versions of our approach.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 496, + 505, + 552 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 149, + 200, + 215, + 210 + ], + "lines": [ + { + "bbox": [ + 149, + 199, + 217, + 212 + ], + "spans": [ + { + "bbox": [ + 149, + 199, + 217, + 212 + ], + "score": 0.969, + "content": "(a) Natural flow", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 146, + 211, + 465, + 566 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 211, + 465, + 566 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 211, + 465, + 566 + ], + "spans": [ + { + "bbox": [ + 146, + 211, + 465, + 566 + ], + "score": 0.975, + "type": "image", + "image_path": "457ed3d46d9fd3c00a2c5bc78d1814abb331afcfbda61a70a5bb5cc1272ab318.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 146, + 211, + 465, + 329.3333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 146, + 329.3333333333333, + 465, + 447.66666666666663 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 146, + 447.66666666666663, + 465, + 566.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 578, + 505, + 612 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "Figure 18: Five additional sequences from the test sets of the natural flow and shape transition", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "setups. The first nine frames contain frames from our reconstruction. The far right image shows the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 600, + 135, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 135, + 614 + ], + "score": 1.0, + "content": "target.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.5 + } + ], + "page_idx": 27, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "28", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 149, + 200, + 215, + 210 + ], + "lines": [ + { + "bbox": [ + 149, + 199, + 217, + 212 + ], + "spans": [ + { + "bbox": [ + 149, + 199, + 217, + 212 + ], + "score": 0.969, + "content": "(a) Natural flow", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 146, + 211, + 465, + 566 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 211, + 465, + 566 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 211, + 465, + 566 + ], + "spans": [ + { + "bbox": [ + 146, + 211, + 465, + 566 + ], + "score": 0.975, + "type": "image", + "image_path": "457ed3d46d9fd3c00a2c5bc78d1814abb331afcfbda61a70a5bb5cc1272ab318.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 146, + 211, + 465, + 329.3333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 146, + 329.3333333333333, + 465, + 447.66666666666663 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 146, + 447.66666666666663, + 465, + 566.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 578, + 505, + 612 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "Figure 18: Five additional sequences from the test sets of the natural flow and shape transition", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "setups. The first nine frames contain frames from our reconstruction. The far right image shows the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 600, + 135, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 135, + 614 + ], + "score": 1.0, + "content": "target.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.5 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/KbV-UZRKb3g/KbV-UZRKb3g.md b/parse/train/KbV-UZRKb3g/KbV-UZRKb3g.md new file mode 100644 index 0000000000000000000000000000000000000000..3cc2dd7c72ace21543edd569f722d5aecc941aba --- /dev/null +++ b/parse/train/KbV-UZRKb3g/KbV-UZRKb3g.md @@ -0,0 +1,225 @@ +# Understanding and Improving Early Stopping for Learning with Noisy Labels + +Yingbin Bai1∗ Erkun Yang2∗ Bo Han3 Yanhua Yang2 Jiatong Li4 Yinian Mao4 Gang Niu5 Tongliang Liu1† + +1TML Lab, University of Sydney; 2Xidian University; 3Hong Kong Baptist University; 4Meituan-Dianping Group; 5RIKEN AIP + +# Abstract + +The memorization effect of deep neural network (DNN) plays a pivotal role in many state-of-the-art label-noise learning methods. To exploit this property, the early stopping trick, which stops the optimization at the early stage of training, is usually adopted. Current methods generally decide the early stopping point by considering a DNN as a whole. However, a DNN can be considered as a composition of a series of layers, and we find that the latter layers in a DNN are much more sensitive to label noise, while their former counterparts are quite robust. Therefore, selecting a stopping point for the whole network may make different DNN layers antagonistically affect each other, thus degrading the final performance. In this paper, we propose to separate a DNN into different parts and progressively train them to address this problem. Instead of the early stopping which trains a whole DNN all at once, we initially train former DNN layers by optimizing the DNN with a relatively large number of epochs. During training, we progressively train the latter DNN layers by using a smaller number of epochs with the preceding layers fixed to counteract the impact of noisy labels. We term the proposed method as progressive early stopping (PES). Despite its simplicity, compared with the traditional early stopping, PES can help to obtain more promising and stable results. Furthermore, by combining PES with existing approaches on noisy label training, we achieve state-of-the-art performance on image classification benchmarks. The code is made public at https://github.com/tmllab/PES. + +# 1 Introduction + +Deep networks have revolutionized a wide variety of tasks, such as image processing, speech recognition, and language modeling [7], However, this highly relies on the availability of large annotated data, which may not be feasible in practice. Instead, many large datasets with lower quality annotations are collected from online queries [5] or social-network tagging [18]. Such annotations inevitably contain mistakes or label noise. As deep networks have large model capacities, they can easily memorize and eventually overfit the noisy labels, leading to poor generalization performance [36]. Therefore, it is of great importance to develop a methodology that is robust to noisy annotations. + +Existing methods on learning with noisy labels (LNL) can be mainly categorized into two groups: model-based and model-free algorithms. Methods in the first category mainly model noisy labels with the noise transition matrix [24, 34, 33, 30]. With perfectly estimated noise transition matrix, models trained with corrected losses can approximate to the models trained with clean labels. However, current methods are usually fragile to estimate the noise transition matrix for heavy noisy data and are also hard to handle a large number of classes [9]. The second type explores the dynamic process of optimization policies, which relates to the memorization effect−deep neural networks tend to first memorize and fit majority (clean) patterns and then overfit minority (noisy) patterns [2]. Recently, based on this phenomenon, many methods [9, 26, 15, 16, 29] have been proposed and achieved promising performance. + +![](images/c7512ec26780dca9b15dc452cb391a4d5b9e7d1c3d1347ce8d597d9d4f66ce84.jpg) +Figure 1: We train a ResNet-18 model on CIFAR-10 with three types of noisy labels and evaluate the impact of noisy labels on the representations from the 9-th layer, the 17-th layer, and the final layer. The X-axis is the number of epochs for the first block of the network. The curves present the mean of five runs and the best performances are indicated with dotted vertical lines. + +To exploit the memorization effect, when the double descent phenomenon [3, 22, 11] cannot be guaranteed to occur, a core issue is to study when to stop the optimization of the network. While stopping the training for too few epochs can avoid overfitting to noisy labels, it can also make the network underfit to clean labels. Current methods [25, 23] usually adopt an early stopping strategy, which decides the stopping point by considering the network as a whole. However, since DNNs are usually optimized with stochastic gradient descent (SGD) with backpropagation, supervisory signals will gradually propagate through the whole network from latter layers (i.e., layers that are closer to output layers) to former layers (i.e., layers that are closer to input layers). Noting that the output layer is followed by the empirical risk in the optimization procedure. We hypothesize that noisy labels may have more severe impacts for the latter layers, which is different from current methods [9, 15] that usually stop the training of the whole network at once. + +To empirically verify the above hypothesis, we analyze the impact of noisy labels on representations from different layers with different training epochs. To quantitatively measure the impact of noisy labels from intermediate layers, we first train the whole network on noisy data with different training epochs and fix the parameters for the selected layer and its previous layers. We then reinitialize and optimize the rest layers with clean data, and the final classification performance is adopted to evaluate the impact of noisy labels. For the final layer, we directly report the overall classification performance. As illustrated in Figure 1, we can see that latter layers always achieve the best performance at relatively smaller epoch numbers and then exhibit stronger performance drops with additional training epochs, which verifies the hypothesis that noisy data may have more severe impacts for latter layers. With this understanding, we can infer that the early stopping, which optimizes the network all at once, may fail to fully exploit the memorization effect and induce sub-optimal performance. + +To address the above problem, we propose to optimize a DNN by considering it as a composition of several DNN parts and present a novel progressive early stopping (PES) method. Specifically, we initially train former DNN layers by optimizing them with a relatively large number of epochs. Then, to alleviate the impact of noisy labels for latter layers, we reinitialize and progressively train latter DNN layers by using smaller numbers of epochs with preceding DNN layers fixed. Since different layers are progressively trained with different early stopping epochs, we term the proposed method as progressive early stopping (PES). Despite its simplicity, compared with normal early stopping trick, PES can help to better exploit the memorization effect and obtain more promising and stable results. Moreover, since the model size and training epochs are gradually reduced during the optimization procedure, the training time of PES is only slightly greater than that of the normal early stopping. Finally, by combining PES with existing approaches on noisy label training tasks, we establish new state-of-the-art (SOTA) results on CIFAR-10 and CIFAR-100 with synthetic noise. We also achieve competitive results on one dataset with real-world noise: Clothing-1M [32]. + +The rest of the paper is organized as follows. In Section 2, we first introduce the proposed progressive early stopping and then present the details of the proposed algorithm by combining our method with existing approaches on noisy label training tasks. Section 3 shows the experimental results of our proposed method. Related works are briefly reviewed in Section 4. Finally, concluding remarks are given in Section 5. + +# 2 Proposed Method + +Let $D$ be the distribution of a pair of random variables $( X , Y ) \in \mathcal { X } \times \{ 1 , . . . K \}$ , where $\boldsymbol { X }$ indicates the variable of instances, $\mathbf { Y }$ is the variable of labels, $\mathcal { X }$ denotes the feature space, and $K$ is the number of classes. In many real-world problems, examples independently drawn from the distribution $D$ are unavailable. Before being observed, the clean labels are usually randomly corrupted into noisy labels. Let $\tilde { D }$ be the distribution of the noisy example $( X , { \tilde { Y } } )$ , where $\tilde { Y }$ indicates the variable of noisy labels. For label-noise learning, we can only access a sample set $\{ \pmb { x } _ { i } , \tilde { y } _ { i } ) \} _ { i = 1 } ^ { n }$ independently drawn from $\tilde { D }$ . The aim is to learn a robust classifier from the noisy sample set that can classify test instances accurately. + +In the following, we first elaborate on the proposed progressive early stopping (PES). Then, based on PES, we provide a learning algorithm that learns with confident examples and semi-supervised learning techniques. + +# 2.1 Progressive Early Stopping + +When trained with noisy labels, if clean labels are of majority within each noisy class, deep networks tend to first fit clean labels during an early learning stage before eventually memorizing the wrong labels, which can be explained by the memorization effect. Many current methods utilize this property to counteract the influence of noisy labels by stopping the optimization at an early learning phase. Specifically, a deep classifier can be obtained by optimizing the following objective function with a relatively small epoch number $T$ : + +$$ +\operatorname* { m i n } _ { \Theta } \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \mathcal { L } ( f ( x _ { i } ; \Theta ) , \tilde { y } _ { i } ) , +$$ + +where $f ( \cdot ; \Theta )$ is a deep classifier with model parameters $\Theta$ and $\mathcal { L }$ is the cross-entropy loss. When trained with noisy data, early learning regularization (ELR) [16] reveals that, for the most commonly used cross-entropy loss, the gradient is well correlated with the correct direction at the early learning phase. Therefore, with a properly defined small epoch number $T$ , the classifier can have higher accuracy than at initialization. While, if we continue to optimize the deep model after $T$ epochs, the classifier will be able to memorize more noise labels. Therefore, it is critical to select a proper epoch number $T$ to utilize the memorization effect and alleviate the influence of noisy labels. + +Current methods typically select the epoch number $T$ by considering the network as a whole. However, as Figure 1 makes clear, the impact of noisy labels on different DNN layers are different, which implies that the traditional early stopping trick, which optimizes the whole network all at once, may make different DNN layers to be antagonistically affected by each other, thus degrading the final model performance. To this end, we propose to separate a DNN into different parts and progressively train layers in different parts with different training epochs. Specifically, assume that the whole network $f ( \cdot ; \Theta )$ can be constituted with $L$ DNN parts + +$$ +\begin{array} { l } { z _ { 1 } = f _ { 1 } ( \pmb { x } ; \Theta _ { 1 } ) , } \\ { z _ { l } = f _ { l } ( z _ { l - 1 } ; \Theta _ { l } ) , \quad l = 2 , \ldots , L } \end{array} +$$ + +where $f _ { l } ( \cdot ; \Theta _ { l } )$ is the $l$ -th DNN part and $z _ { l }$ is the corresponding output. The output of the last part $z _ { L }$ is the prediction. The network $f ( \cdot ; \Theta )$ can also be represented as $f ( \cdot ; \Theta _ { 1 } , . . . \Theta _ { L } )$ . To counteract the impact of noisy labels, We initially optimize the parameter $\Theta _ { 1 }$ for the first part by training the whole network for $T _ { 1 }$ epochs with the following objective + +$$ +\operatorname* { m i n } _ { \Theta _ { 1 } \ldots \Theta _ { k } } \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \mathcal { L } ( f ( \pmb { x } _ { i } ; \Theta _ { 1 } , \ldots , \Theta _ { L } ) , \tilde { y } _ { i } ) . +$$ + +![](images/93c2dd94c2a0f5dd8e852ee49fbcbcd3facdbcf530179d0abf899ddba8541bf1.jpg) +Figure 2: Performance of the traditional early stopping trick and the proposed PES on CIFAR-10 with different types of label noise. The lines present the mean of five runs. + +Then, we keep the obtained parameter $\Theta _ { 1 } ^ { * }$ fixed, reinitialize and progressively learn the $l$ -th $( l =$ $2 , \ldots , L )$ DNN part with the parameters for preceding DNN parts fixed. The training procedure is conducted with $T _ { l }$ epochs by optimizing the following objective + +$$ +\operatorname* { m i n } _ { \Theta _ { l } \ldots \Theta _ { k } } \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \mathcal { L } ( f ( x _ { i } ; \Theta _ { 1 } ^ { * } , \ldots , \Theta _ { l - 1 } ^ { * } , \Theta _ { l } , \ldots , \Theta _ { L } ) , \tilde { y } _ { i } ) , \quad l = 2 \ldots L +$$ + +We gradually optimize the $( l + 1 )$ -th DNN part with the obtained parameter $\Theta _ { l } ^ { * }$ fixed, the optimization is continued until all the parameters have been optimized. As elaborated above, latter DNN parts are more sensitive to noisy labels than their former counterparts. Therefore, for the above initializing optimization in Eq. (3) and the following $L - 1$ steps of optimization in Eq. (4), we gradually reduce the training epochs (i.e. $T _ { 1 } \geq T _ { 2 } \geq \cdot \cdot \cdot \geq T _ { L } )$ to better exploit the memorization effect. After optimization, we can obtain the final network as $f ( \cdot , \Theta ) = \bar { f } ( \cdot ; \Theta _ { 1 } ^ { * } , \ldots , \Theta _ { L } ^ { * } )$ . Since this model is obtained by progressively exploiting the early stopping strategies for different DNN parts, we term the proposed method as progressive early stopping (PES). + +To explicitly verify the effectiveness of the proposed PES method, we conduct several pilot experiments, which compare the traditional early stopping and PES with label noise from different types and different levels. The results are illustrated in Figure 2, from which we can see that, compared with models trained with traditional early stopping, models trained with PES can achieve superior classification accuracy with smaller variations in all the cases. Current state-of-the-art methods [15] usually adopt models with the traditional early stopping as base models to distill confident examples and then utilize semi-supervised learning techniques by considering confident examples as labeled data and other noisy examples as unlabeled data to further improve the results. The final performance still heavily relies on the base model trained with noisy labels. By improving the performance of the base model, our method combined with semi-supervised learning techniques is able to establish new state-of-the-art results. In the following subsections, we will elaborate on how to utilize PES to distill confident examples and further combine it with semi-supervised learning techniques. + +# 2.2 Learning with Confident Examples + +Based on the deep network optimized with progressive early stopping, we can select confident examples to facilitate the model training. Here, confident examples refer to examples that have high probabilities with clean labels. In this paper, we treat examples whose predictions are consistent with given labels as confident examples. In addition, to make the results more robust, we generate two different augmentations for any given input and use the average prediction to decide its predicted label. Formally, we can obtain the confident example set $\mathcal { D } _ { l }$ as + +$$ +\begin{array} { c } { \mathcal { D } _ { l } = \{ ( \pmb { x } _ { i } , \tilde { y } _ { i } ) | \tilde { y } _ { i } = \hat { y } _ { i } , i = 1 , \ldots , n \} , } \\ { \hat { y } _ { i } = \underset { k \in \{ 1 , \ldots , K \} } { \arg \operatorname* { m a x } } \ \frac { 1 } { 2 } [ f ^ { k } ( \mathrm { A u g m e n t } ( \pmb { x } _ { i } ) ; \Theta ) + f ^ { k } ( \mathrm { A u g m e n t } ( \pmb { x } _ { i } ) ; \Theta ) ] , } \end{array} +$$ + +where Augment $( \cdot )$ indicates normal data augmentation operation including horizontal random flip and random crops, and $f ^ { k } ( x ; \Theta )$ is the predicted probability of $_ { \textbf { \em x } }$ belonging to class $k$ . Note that Augment $\left( \cdot \right)$ is a stochastic transformation, so the two terms in Eq (5) are not identical. The average + +Input: Neural network with trainable parameters $\boldsymbol { \Theta } = \{ \Theta _ { 1 } , \dots , \Theta _ { L } \}$ , Noisy training dataset +$\{ \bar { \pmb { x } _ { i } } , \tilde { y } _ { i } ) \} _ { i = 1 } ^ { n }$ , Number of training epochs for different part: $T _ { 1 } , \dots , T _ { L }$ , and training epochs $T _ { c }$ +for refining with confident examples. +for $i = 1 , \dots , T _ { 1 }$ do +Optimize network parameter $\Theta$ with Eq. (3); +for $l = 2$ , . . . , $L$ do Froze $\left\{ \Theta _ { 1 } , \ldots , \Theta _ { l - 1 } \right\}$ and re-initialize $\big \{ \Theta _ { l } , \dots , \Theta _ { L } \big \}$ ; for $i = 1 , \dots , T _ { l }$ do Optimize network parameter $\big \{ \Theta _ { l } , \dots , \Theta _ { L } \big \}$ with Eq. (4); +Unfroze $\Theta$ ; +for $i = 1 , \dots , T _ { c }$ do Extract confident example set $\mathcal { D } _ { l }$ and unlabeled set $\mathcal { D } _ { u }$ with classifier $f ( \cdot , \Theta )$ by Eq. (7); Training the classifier $f ( \cdot , \Theta )$ with MixMatch loss on $\mathcal { D } _ { l }$ and $\mathcal { D } _ { u }$ ; +Evaluate the obtained classifier $f ( \cdot , \Theta )$ . + +prediction of augmented examples provides a more stable prediction and is found empirically to improve performance. After obtaining the confident example set, one can easily train a classifier by considering confident examples as clean data. However, since the number of confident examples for different classes can vary greatly, directly training the model with the obtained confident example set may introduce a severe class imbalance problem. To this end, we adopt a weighted classification loss + +$$ +\mathcal { L } _ { c } = \sum _ { i = 1 } ^ { N } w _ { y _ { i } } \mathcal { L } _ { p } ( \tilde { y } _ { i } , f ( \pmb { x } _ { i } ; \Theta ) ) , +$$ + +where $w _ { i }$ is the corresponding class weight. Assuming that $\sigma _ { k } = | \{ ( \pmb { x } _ { i } , \tilde { y } _ { i } ) | \tilde { y } _ { i } = k , ( \pmb { x } _ { i } , \tilde { y } _ { i } ) \in \mathcal { D } _ { l } \} |$ denotes the cardinality of the confident example set belonging to the $k$ -th class. Then, we can set $w _ { i } = \sigma _ { i } / ( \sum _ { j = 1 } ^ { K } \sigma _ { j } )$ to indicate the corresponding class importance. + +# 2.3 Combining with Semi-Supervised Learning + +Training with only confident examples neglects the rest data and may suffer from insufficient training examples. To tackle this problem, we further resort to semi-supervised learning techniques by considering confident examples as labeled data and other noisy examples as unlabeled data. Specifically, the labeled data set and unlabeled data set can be obtained as + +$$ +\begin{array} { c } { { \displaystyle { \int } { \mathcal D } _ { l } = \{ ( \boldsymbol x _ { i } , \tilde { y } _ { i } ) | \tilde { y } _ { i } = \hat { y } _ { i } , i = 1 , . . . n \} } } \\ { { \displaystyle { \mathcal D } _ { u } = \{ \boldsymbol x _ { i } | \tilde { y } _ { i } \neq \hat { y } _ { i } , i = 1 , . . . n \} } } \\ { { \displaystyle { \hat { y } _ { i } = \arg \operatorname* { m a x } _ { k \in \{ 1 , . . . K \} } \frac { 1 } { 2 } [ f ^ { k } ( \mathrm { A u g m e n t } ( \boldsymbol x _ { i } ) ; \Theta ) + f ^ { k } ( \mathrm { A u g m e n t } ( \boldsymbol x _ { i } ) ; \Theta ) ] } , } } \end{array} +$$ + +where the labeled data set $\mathcal { D } _ { l }$ is the same as that in Eq (6), and $\mathcal { D } _ { u }$ is the rest unlabeled data set. Similar to [15], we adopt MixMatch [4] as the semi-supervised learning framework to train the final classification models. For more details about semi-supervised learning, we refer to [4]. The whole learning algorithm is summarized in Algorithm 1. + +# 3 Experiments + +# 3.1 Datasets and Implementation Details + +Datasets: We evaluate our method on two synthetic datasets, CIFAR-10 and CIFAR-100 [12] with different levels of symmetric, pairflip, and instance-dependent label noise (abbreviated as instance label noise) and a real-world dataset Clothing-1M [32]. Both CIFAR-10 and CIFAR-100 contain $5 0 \mathrm { k }$ training images and $1 0 \mathrm { k }$ test images of size $3 2 \times 3 2$ . Following previous works [9, 31, 16, 29], symmetric noise is generated by uniformly flipping labels for a percentage of the training dataset to all possible labels. Pairflip noise flips noisy labels into their adjacent class. And, instance noise is generated by image features. More details about the synthetic label noise are given in the supplementary material. For the flipping rate, it can include [9, 31] or ex-include [15, 16] true labels. We use the flipping rate including correct labels in Table 3 to compare with results in [15], and use without correct labels in the rest of the experiments. Clothing-1M [32] is a large-scale dataset with real-world noisy labels, whose images are clawed from the online shopping websites, and labels are generated based on surrounding texts. It contains 1 million training images, and $1 5 \mathrm { k }$ validation images, and 10k test images with clean labels. + +Table 1: Preliminary analysis of the performance and the quality of extracted confident examples on CIFAR-10. The mean and standard deviation are computed over five runs. + +
MetricsMethodsSym-20%Sym-50%Pair-45%Inst-20%Inst-40%
Test AccuracyEarly StoppingPES82.55±2.4685.87±1.5970.76±1.2475.87±1.3360.62±5.5962.40±2.3484.41±0.9086.58±0.4574.73±2.6577.07±1.18
Label PrecisionEarly StoppingPES98.81±0.1598.96±0.0994.65±0.1995.46±0.1472.53±5.2672.99±2.2798.70±0.4398.52±0.1990.77±1.8790.63±0.92
Label RecallEarly StoppingPES88.51±2.2692.67±1.4375.18±1.0081.03±1.8367.84±5.0671.06±2.2790.37±1.0193.24±0.6082.15±3.1785.91±0.68
+ +Baselines: Semi-supervised learning may strongly boost the performance, we separately compare our method with approaches with or without semi-supervised learning. For the comparison with baselines with semi-supervised learning, we combine our proposed method with MixMatch used in [15] as indicated in Subsection 2.3. (1) Approaches without semi-supervised learning: Co-teaching [9], Forward [24], Joint Optim [25], T-revision [31], DMI [34], and CDR [29]. (2) Methods with semi-supervised learning: M-correction [1], DivideMix [15], and $\mathrm { E L R + }$ [16]. We also adopt standard training with cross-entropy (CE) and MixUp [37] as baselines to show improvements. + +Network structure and optimization: Our method is implemented by PyTorch v1.6. Baseline methods are implemented based on public codes with hyper-parameters set according to the original papers. For DivideMix and $\mathrm { E L R + }$ , we evaluate the test accuracy with the first network. To better demonstrate the robustness of our algorithm, we keep the hyper-parameters fixed for different types of label noise. More technique details are given in the supplementary material. + +For experiments without semi-supervised learning, we follow [31], and use ResNet-18 [10] for CIFAR-10 and ResNet-34 for CIFAR-100. We split networks into three parts, the layers above block 4 as part 1, block 4 of ResNet as part 2, and the final layer as part 3. $T _ { 1 }$ is defined as 25 for CIFAR-10 and 30 for CIFAR-100, $T _ { 2 }$ as 7, and $T _ { 3 }$ as 5. The network is trained for 200 epochs and SGD with 0.9 momentum is used. The initial learning rate is set to 0.1 and decayed with a factor of 10 at the 100th and 150th epoch respectively, and a weight decay is set to $1 0 ^ { - 4 }$ . For $T _ { 2 }$ and $T _ { 3 }$ , we employ an Adam optimizer with a learning rate of $1 0 ^ { - 4 }$ . + +For experiments with semi-supervised learning, we follow the setting of [15] with PreAct Resnet-18. We set the final layer as part 2, the rest as part 1. $T _ { 1 }$ is defined as 20 for CIFAR-10 and 35 for CIFAR-100, and $T _ { 2 }$ as 5. The network is trained for 300 epochs. For optimization, we use a single cycle of cosine annealing [19], and the learning rate begins from $2 \times \mathrm { \bar { 1 0 ^ { - 2 } } }$ and ends at $2 \times 1 0 ^ { - 4 }$ , with a weight decay of $5 \times 1 0 ^ { - 4 }$ . An Adam optimizer is adopted with a learning rate of $1 0 ^ { - 4 }$ for $T _ { 2 }$ . For hyper-parameters from MixMatch, we set them according to the original paper [4]. + +For Clothing-1M [32], we follow the previous work [25], and employ a ResNet-50 [10] pre-trained on ImageNet [13]. We set the final layer as part 2, the rest as part 1. $T _ { 1 }$ and $T _ { 2 }$ are defined as 20 and 7 respectively. The network is trained with CE loss for 50 epochs and SGD is used with 0.9 momentum and a weight decay of $1 0 ^ { - 3 }$ . The learning rate is $5 \times 1 0 ^ { - 3 }$ and decayed by a factor of 10 at the 20th and 30th epoch respectively. We employ an Adam optimizer with a learning rate of $5 \times 1 0 ^ { - 6 }$ for $T _ { 2 }$ + +# 3.2 Preliminary Experiments + +In Figure 2, we can observe that with the PES trick, the performance of classifiers is generally improved compared with that the traditional early stopping trick. In this section, we further carefully analyze the quality of extracted labels by examining them from three aspects, i.e., test accuracy, label precision, and label recall. Here, label precision indicates the ratio of the number of extracted confident examples with correct labels in the total confident example set, and label recall represents the ratio of the number of confident examples with correct labels among the total correctly labeled examples. Specifically, we train a neural network on CIFAR-10 with different kinds and levels of label noise for 25 epochs respectively and report the performance for each case before and after the proposed PES is applied. + +Table 2: Comparison with state-of-the-art methods without semi-supervised learning on CIFAR-10 and CIFAR-100. The mean and standard deviation computed over five runs are presented. + +
DatasetMethodSymmetricPairflipInstance
20%50%45%20%40%
CIFAR10CE84.00±0.6675.51±1.2463.34±6.0385.10±0.6877.00±2.17
Co-teaching87.16±0.1172.80±0.4570.11±1.1686.54±0.1180.98±0.39
Forward85.63±0.5277.92±0.6660.15±1.9785.29±0.3874.72±3.24
Joint Optim89.70±0.1185.00±0.1782.63±1.3889.69±0.4282.62±0.57
T-revision89.63±0.1383.40±0.6577.06±6.4790.46±0.1385.37±3.36
DMI88.18±0.3678.28±0.4857.60±14.5689.14±0.3684.78±1.97
CDR89.72±0.3882.64±0.8973.67±0.5490.41±0.3483.07±1.33
Ours92.38±0.4087.45±0.3588.43±1.0892.69±0.4489.73±0.51
CIFAR100CE51.43±0.5837.69±3.4534.10±2.0452.19±1.4242.26±1.29
Co-teaching59.28±0.4741.37±0.0833.22±0.4857.24±0.6945.69±0.99
Forward57.75±0.3744.66±1.0127.88±0.8058.76±0.6644.50±0.72
Joint Optim64.55±0.3850.22±0.4142.61±0.6165.15±0.3155.57±0.41
T-revision65.40±1.0750.24±1.4541.10±1.9560.71±0.7351.54±0.91
DMI58.73±0.7044.25±1.1426.90±0.4558.05±0.2047.36±0.68
CDR66.52±0.2455.30±0.9643.87±1.3567.33±0.6755.94±0.56
Ours68.89±0.4558.90±2.7257.18±1.4470.49±0.7965.68±1.41
+ +Table 3: Comparison with state-of-the-art methods with semi-supervised learning on CIFAR-10 and CIFAR-100 with symmetric label noise from different levels. Results with \* are token from [15]. The mean and standard deviation are computed over three runs. + +
DatasetCIFAR-10CIFAR-100
Methods /NoiseSym-20%Sym-50%Sym-80%Sym-20%Sym-50%Sym-80%
CE86.5±0.680.6±0.263.7±0.857.9±0.447.3±0.222.3±1.2
MixUp93.2±0.388.2±0.373.3±0.369.5±0.257.1±0.634.1±0.6
M-correction*94.092.086.873.966.148.2
DivideMix*95.294.293.075.272.858.3
DivideMix95.6±0.194.6±0.192.9±0.375.3±0.172.7±0.656.4±0.3
ELR+94.9±0.293.6±0.190.4±0.275.5±0.271.0±0.250.4±0.8
Ours (Semi)95.9±0.195.1±0.293.1±0.277.4±0.374.3±0.661.6±0.6
+ +Results in Table 1 clearly show that, compared with the traditional early stopping, PES can help to obtain higher accuracies, precisions, and recalls for most cases. For instance-dependent label noise, PES can achieve higher recall values with comparable label precision values. Note that models with high recall values can help to collect more confident examples, which is critical for learning with confident examples and semi-supervised learning. Therefore, by enhancing the performance of the initial model, PES can help to improve the final classification performance in all cases, which is also verified by the experiments in Section 3.3. + +# 3.3 Classification Accuracy Evaluation + +Synthetic datasets. We first verify the effectiveness of our proposed method without semi-supervised learning techniques on two synthetic datasets: CIFAR-10 and CIFAR-100. For both of these two datasets, we leave $10 \%$ of data with noisy labels as noisy validation set. Results are presented in Table 2, which shows that our proposed method can consistently outperform all other baselines across various settings by a large margin. + +Table 4: Comparison with state-of-the-art methods with semi-supervised learning on CIFAR-10 and CIFAR-100 with instance-dependent and pairflip label noise from different levels. The mean and standard deviation are computed over three runs. + +
DatasetCIFAR-10CIFAR-100
Methods/NoiseInst-20%Inst-40%Pair-45%Inst-20%Inst-40%Pair-45%
CEMixUpDivideMixELR+87.5±0.578.9±0.774.9±1.756.8±0.448.2±0.538.5±0.6
93.3±0.295.5±0.187.6±0.594.5±0.282.4±1.085.6±1.767.1±0.155.0±0.170.9±0.144.2±0.548.2±1.0
75.2±0.270.9±0.1
94.9±0.194.3±0.286.1±1.275.8±0.174.3±0.365.3±1.3
Ours (Semi)95.9±0.195.3±0.194.5±0.377.6±0.376.1±0.473.6±1.7
+ +Table 5: Compassion with state-of-the-art methods on Clothing-1M. Results of baseline methods are taken from the original papers. ours represent the results obtained by PES with a single network and ours\* indicate the results obtained by PES with an ensemble model. + +
CEForwardJoint-OptimDMIT-revisionDivideMix*ELR+*OursOurs*
69.2169.8472.1672.4674.1874.7674.8174.6474.99
+ +![](images/61ae512b2fb46bee07bd5e6157d516e400cd564030c729a91354232a89157ebe.jpg) +Figure 3: Sensitivity analysis for different training iteration numbers: $T _ { 2 }$ and $T _ { 3 }$ + +Table 3 and Table 4 present the mean accuracy and standard deviation for our method and all baselines on CIFAR-10 and CIFAR-100, respectively. From the results, we can get that the proposed method can outperform all baselines in all cases. For pairflip label noise, the advantages of our proposed method become more apparent, and it significantly outperforms state-of-the-art methods by over $8 \%$ on both CIFAR-10 and CIFAR-100. These empirical results support our proposal that PES can improve the quality of selected confident examples, which helps improve performance and reduce the variance of the final classifier. + +Real-world dataset. We evaluate the performance of the proposed method on a real-world dataset with Clothing-1M [32] and select methods such as CE, Forward, Joint-Optim, DMI, and T-revision, which use a single network, and also methods such as DivideMix and $\mathrm { E L R + }$ , which adopt an ensemble model with two different networks, as baselines. We also report the results for the proposed PES with a single network as ours and the results for PES, which ensembles two networks, as ours\*. The overall results are reported in Table 5, from which we can observe that the proposed PES with a single network can outperform all baselines using a single network. And with an ensemble model, which contains two different networks, our method can outperform all the adopted baselines. These results clearly demonstrate that, by improving the performance of the initial classification network, our method is more flexible to handle such real-world noise problems. + +# 3.4 Sensitivity Analysis + +In this section, we investigate the hyper-parameter sensitivity for the training iteration number $T _ { 2 }$ and $T _ { 3 }$ , respectively. We firstly analyze the training epoch number for the second DNN part by varying $T _ { 2 }$ from the range of [0, 10]. The results are illustrated in Figure 3a, from which can find that, with the increasing of $T _ { 2 }$ , the performance of PES first increase and then decrease in all the cases except for $45 \%$ Pairflip noise on the CIFAR-10 dataset. While the model achieves the best performance with $T _ { 2 }$ as 7 for all types of noisy labels. Then we fix $T _ { 2 }$ as 7, and analyze the impact of the third DNN part by varying $T _ { 3 }$ from the range of [0, 10]. The results are shown in Figure 3b. Although the performance variance for different $T _ { 3 }$ is smaller than that for $T _ { 2 }$ , we can still observe that the best performance can be obtained when $T _ { 3 }$ is set as 5. More importantly, from these two figures, we can get that both $T _ { 2 }$ and $T _ { 3 }$ are robust to the different types of noisy labels. + +Table 6: Training time comparison for baselines on CIFAR-10 with $50 \%$ Symmetric label noise. + +
CECo-teachingCDRT-revisionELR+DivideMixOursOurs (Semi)
0.9h1.5h3.0h3.5h2.2h5.5h1.0h3.1h
+ +# 3.5 Training Time Comparison + +In this section, we compare the training time of our method and other state-of-the-art baselines. All the experiments are conducted on a server with a single Nvidia V100 GPU. The training times for all the methods are reported in Table 6, from which we can get that our algorithm with cross-entropy loss achieves the fastest speed across all baselines, only about 1 hour. Our method combining with MixMatch [4] is also fast, only a little more than half of the training time of DivideMix. The time of $\mathrm { E L R + }$ [16] shows superior, but $\mathrm { E L R + }$ trains the network with fewer epochs, with 200 epochs compared with ours for 300 epochs. + +# 4 Related work + +Learning with noisy data has been well studied [17, 6, 21, 27, 20]. Current works can be mainly categorized into two groups: model-based and model-free methods. In this section, we briefly review some closely related works. + +The first type models the relationship between clean labels and noisy labels by estimating the noise transition matrix and build a loss function to correct the loss [24, 30, 35, 28]. [24] first combines algorithms for estimating the noise rates and loss correction techniques together and introduces two alternative procedures for loss correction. It also proves that both of the two procedures enjoy formal robustness guarantees w.r.t. the clean data distribution. DMI [34] proposes an information-theoretic loss function, which utilizes Shannon’s mutual information and is robustness to different kinds of label noise. T-revision [31] estimates the noise transition matrix without anchor points by adding a fine-tuned slack variables. Although these methods have made certain progress, they are usually fragile to estimate the noise transition matrix for heavy noisy data and are also hard to handle a large number of classes. Therefore, in this paper, we mainly focus on the model-free methods. + +The second strand mainly counteracts noisy labels by exploiting the memorization effect that deep networks tend to first memorize and fit majority (clean) patterns and then overfit minority (noisy) patterns [2]. To exploit this property, Co-teaching [9] employs two networks with different initialization and uses small loss to select confident examples. M-correction [1] uses two Gaussian Mixture Models to identify confident examples, instead of using networks themselves. DivideMix [15] extends Co-teaching [9] and employs two Beta Mixture Model to select confident examples. MixMatch [4] is then adopted to leverage unconfident examples with a semi-supervised learning framework. All the above methods exploit the memorization effect by considering the adopted network as a whole. Recently, [14] shows that networks training with noisy labels can produce good representations, if the structure of networks suits the targeted tasks. Our method further explains that noisy labels have different impacts for different layers in a DNN. And latter layers will receive earlier and more severe impact than their former counterparts. Therefore, by considering a DNN as a composition of several layers and training different layers with different epochs, our method is able to better exploit the memorization effect and achieve superior performance. + +# 5 Conclusion + +In this work, we provide a progressive early stopping (PES) method to better exploit the memorization effect of deep neural networks (DNN) for noisy-label learning. We first find that the impact of noisy labels for former layers in a DNN is much less and later than that for latter DNN layers, and then build upon this insight to propose the PES method, which separates a DNN into different parts and progressively train each part to counteract the different impacts of noisy labels for different DNN layers. To show that PES can boost the performance of state-of-the-art methods, we conduct extensive experiments across multiple synthetic and real-world noisy datasets and demonstrate that the proposed PES can help to obtain substantial performance improvements compared to current state-of-the-art baselines. The main limitation of our method lies in that, by splitting a DNN into different parts, PES introduces several additional hyper-parameters that need to be tuned carefully. In the future, we will extend the work in the following aspects. First, we will study other mechanisms that distinguishing desired and undesired memorization rather than early stopping, e.g., the gradient ascent trick [8]. Second, we are interested in combining PES with interesting ideas from semi-supervised learning and unsupervised learning. + +# Acknowledgments and Disclosure of Funding + +YB was partially supported by Agriculture Consultant and Smart Management. BH was supported by the RGC Early Career Scheme No. 22200720, NSFC Young Scientists Fund No. 62006202 and HKBU CSD Departmental Incentive Grant. YY was partially supported by Key Research and Development Program of Shaanxi (ProgramNo. 2021ZDLGY01-03). GN was supported by JST AIP Acceleration Research Grant Number JPMJCR20U3, Japan. TL was partially supported by Australian Research Council Projects DE-190101473 and IC-190100031. + +References +[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O’Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In ICML, pages 312–321, 2019. +[2] Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S. Kanwal, Tegan Maharaj, Asja Fischer, Aaron C. Courville, Yoshua Bengio, and Simon Lacoste-Julien. A closer look at memorization in deep networks. In ICML, pages 233–242, 2017. +[3] Mikhail Belkin, Daniel Hsu, Siyuan Ma, and Soumik Mandal. Reconciling modern machine-learning practice and the classical bias–variance trade-off. Proceedings of the National Academy of Sciences, 116(32):15849–15854, 2019. +[4] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, pages 5050–5060, 2019. +[5] Youngchul Cha and Junghoo Cho. Social-network analysis using topic models. In SIGIR, pages 565–574, 2012. +[6] Jacob Goldberger and Ehud Ben-Reuven. Training deep neural-networks using a noise adaptation layer. In ICLR, 2017. +[7] Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning. The MIT Press, 2016. +[8] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor W. Tsang, and Masashi Sugiyama. SIGUA: forgetting may make learning with noisy labels more robust. In ICML, pages 4006–4016, 2020. +[9] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages 8527–8537, 2018. +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. +[11] Takashi Ishida, Ikko Yamane, Tomoya Sakai, Gang Niu, and Masashi Sugiyama. Do we need zero training loss after achieving zero training error? In ICML, pages 4604–4614, 2020. +[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. +[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, pages 1097–1105, 2012. +[14] Jingling Li, Mozhi Zhang, Keyulu Xu, John P. Dickerson, and Jimmy Ba. Noisy labels can induce good representations. arXiv preprint arXiv:2012.12896, 2020. +[15] Junnan Li, Richard Socher, and Steven C. H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR, 2020. +[16] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. In NeurIPS, pages 20331–20342, 2020. +[17] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on pattern analysis and machine intelligence, 38(3):447–461, 2016. +[18] Wei Liu, Yu-Gang Jiang, Jiebo Luo, and Shih-Fu Chang. Noise resistant graph ranking for improved web image search. In CVPR, pages 849–856, 2011. +[19] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In ICLR, 2017. +[20] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah M. Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. In ICML, pages 6543–6553, 2020. +[21] Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah M. Erfani, Shu-Tao Xia, Sudanthi N. R. Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In ICML, pages 3361–3370, 2018. +[22] Preetum Nakkiran, Gal Kaplun, Yamini Bansal, Tristan Yang, Boaz Barak, and Ilya Sutskever. Deep double descent: Where bigger models and more data hurt. In ICLR, 2020. +[23] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi-Phuong-Nhung Ngo, Thi Hoai Phuong Nguyen, Laura Beggel, and Thomas Brox. SELF: learning to filter noisy labels with self-ensembling. In ICLR, 2020. +[24] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, pages 2233–2241, 2017. +[25] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, pages 5552–5560, 2018. +[26] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, pages 8688–8696, 2018. +[27] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In ICCV, pages 322–330, 2019. +[28] Songhua Wu, Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Nannan Wang, Haifeng Liu, and Gang Niu. Class2simi: A noise reduction perspective on learning with noisy labels. In ICML, pages 11285–11295, 2021. +[29] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hindering the memorization of noisy labels. In ICLR, 2021. +[30] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, pages 7597–7610, 2020. +[31] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, pages 6835–6846, 2019. +[32] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In CVPR, pages 2691–2699, 2015. +[33] Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In NeurIPS, pages 6256–6268, 2020. +[34] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L_dmi: A novel information-theoretic loss function for training deep nets robust to label noise. In NeurIPS, pages 6222–6233, 2019. +[35] Yu Yao, Tongliang Liu, Bo Han, Mingming Gong, Jiankang Deng, Gang Niu, and Masashi Sugiyama. Dual T: reducing estimation error for transition matrix in label-noise learning. In NeurIPS, pages 7260–7271, 2020. +[36] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In ICLR, 2016. +[37] Hongyi Zhang, Moustapha Cissé, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. \ No newline at end of file diff --git a/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_content_list.json b/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8bf029bdef25840681a214c7cc57e6ae81898527 --- /dev/null +++ b/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_content_list.json @@ -0,0 +1,976 @@ +[ + { + "type": "text", + "text": "Understanding and Improving Early Stopping for Learning with Noisy Labels ", + "text_level": 1, + "bbox": [ + 199, + 122, + 800, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingbin Bai1∗ Erkun Yang2∗ Bo Han3 Yanhua Yang2 Jiatong Li4 Yinian Mao4 Gang Niu5 Tongliang Liu1† ", + "bbox": [ + 292, + 224, + 710, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1TML Lab, University of Sydney; 2Xidian University; 3Hong Kong Baptist University; 4Meituan-Dianping Group; 5RIKEN AIP ", + "bbox": [ + 214, + 260, + 784, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 325, + 535, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The memorization effect of deep neural network (DNN) plays a pivotal role in many state-of-the-art label-noise learning methods. To exploit this property, the early stopping trick, which stops the optimization at the early stage of training, is usually adopted. Current methods generally decide the early stopping point by considering a DNN as a whole. However, a DNN can be considered as a composition of a series of layers, and we find that the latter layers in a DNN are much more sensitive to label noise, while their former counterparts are quite robust. Therefore, selecting a stopping point for the whole network may make different DNN layers antagonistically affect each other, thus degrading the final performance. In this paper, we propose to separate a DNN into different parts and progressively train them to address this problem. Instead of the early stopping which trains a whole DNN all at once, we initially train former DNN layers by optimizing the DNN with a relatively large number of epochs. During training, we progressively train the latter DNN layers by using a smaller number of epochs with the preceding layers fixed to counteract the impact of noisy labels. We term the proposed method as progressive early stopping (PES). Despite its simplicity, compared with the traditional early stopping, PES can help to obtain more promising and stable results. Furthermore, by combining PES with existing approaches on noisy label training, we achieve state-of-the-art performance on image classification benchmarks. The code is made public at https://github.com/tmllab/PES. ", + "bbox": [ + 233, + 357, + 766, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction ", + "text_level": 1, + "bbox": [ + 174, + 659, + 310, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep networks have revolutionized a wide variety of tasks, such as image processing, speech recognition, and language modeling [7], However, this highly relies on the availability of large annotated data, which may not be feasible in practice. Instead, many large datasets with lower quality annotations are collected from online queries [5] or social-network tagging [18]. Such annotations inevitably contain mistakes or label noise. As deep networks have large model capacities, they can easily memorize and eventually overfit the noisy labels, leading to poor generalization performance [36]. Therefore, it is of great importance to develop a methodology that is robust to noisy annotations. ", + "bbox": [ + 174, + 690, + 825, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods on learning with noisy labels (LNL) can be mainly categorized into two groups: model-based and model-free algorithms. Methods in the first category mainly model noisy labels with the noise transition matrix [24, 34, 33, 30]. With perfectly estimated noise transition matrix, models trained with corrected losses can approximate to the models trained with clean labels. However, current methods are usually fragile to estimate the noise transition matrix for heavy noisy data and are also hard to handle a large number of classes [9]. The second type explores the dynamic process of optimization policies, which relates to the memorization effect−deep neural networks tend to first memorize and fit majority (clean) patterns and then overfit minority (noisy) patterns [2]. Recently, based on this phenomenon, many methods [9, 26, 15, 16, 29] have been proposed and achieved promising performance. ", + "bbox": [ + 176, + 808, + 825, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7512ec26780dca9b15dc452cb391a4d5b9e7d1c3d1347ce8d597d9d4f66ce84.jpg", + "image_caption": [ + "Figure 1: We train a ResNet-18 model on CIFAR-10 with three types of noisy labels and evaluate the impact of noisy labels on the representations from the 9-th layer, the 17-th layer, and the final layer. The X-axis is the number of epochs for the first block of the network. The curves present the mean of five runs and the best performances are indicated with dotted vertical lines. " + ], + "image_footnote": [], + "bbox": [ + 178, + 101, + 820, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 325, + 825, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To exploit the memorization effect, when the double descent phenomenon [3, 22, 11] cannot be guaranteed to occur, a core issue is to study when to stop the optimization of the network. While stopping the training for too few epochs can avoid overfitting to noisy labels, it can also make the network underfit to clean labels. Current methods [25, 23] usually adopt an early stopping strategy, which decides the stopping point by considering the network as a whole. However, since DNNs are usually optimized with stochastic gradient descent (SGD) with backpropagation, supervisory signals will gradually propagate through the whole network from latter layers (i.e., layers that are closer to output layers) to former layers (i.e., layers that are closer to input layers). Noting that the output layer is followed by the empirical risk in the optimization procedure. We hypothesize that noisy labels may have more severe impacts for the latter layers, which is different from current methods [9, 15] that usually stop the training of the whole network at once. ", + "bbox": [ + 174, + 414, + 825, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To empirically verify the above hypothesis, we analyze the impact of noisy labels on representations from different layers with different training epochs. To quantitatively measure the impact of noisy labels from intermediate layers, we first train the whole network on noisy data with different training epochs and fix the parameters for the selected layer and its previous layers. We then reinitialize and optimize the rest layers with clean data, and the final classification performance is adopted to evaluate the impact of noisy labels. For the final layer, we directly report the overall classification performance. As illustrated in Figure 1, we can see that latter layers always achieve the best performance at relatively smaller epoch numbers and then exhibit stronger performance drops with additional training epochs, which verifies the hypothesis that noisy data may have more severe impacts for latter layers. With this understanding, we can infer that the early stopping, which optimizes the network all at once, may fail to fully exploit the memorization effect and induce sub-optimal performance. ", + "bbox": [ + 174, + 573, + 825, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above problem, we propose to optimize a DNN by considering it as a composition of several DNN parts and present a novel progressive early stopping (PES) method. Specifically, we initially train former DNN layers by optimizing them with a relatively large number of epochs. Then, to alleviate the impact of noisy labels for latter layers, we reinitialize and progressively train latter DNN layers by using smaller numbers of epochs with preceding DNN layers fixed. Since different layers are progressively trained with different early stopping epochs, we term the proposed method as progressive early stopping (PES). Despite its simplicity, compared with normal early stopping trick, PES can help to better exploit the memorization effect and obtain more promising and stable results. Moreover, since the model size and training epochs are gradually reduced during the optimization procedure, the training time of PES is only slightly greater than that of the normal early stopping. Finally, by combining PES with existing approaches on noisy label training tasks, we establish new state-of-the-art (SOTA) results on CIFAR-10 and CIFAR-100 with synthetic noise. We also achieve competitive results on one dataset with real-world noise: Clothing-1M [32]. ", + "bbox": [ + 174, + 731, + 825, + 911 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rest of the paper is organized as follows. In Section 2, we first introduce the proposed progressive early stopping and then present the details of the proposed algorithm by combining our method with existing approaches on noisy label training tasks. Section 3 shows the experimental results of our proposed method. Related works are briefly reviewed in Section 4. Finally, concluding remarks are given in Section 5. ", + "bbox": [ + 174, + 90, + 825, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Proposed Method ", + "text_level": 1, + "bbox": [ + 174, + 180, + 352, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $D$ be the distribution of a pair of random variables $( X , Y ) \\in \\mathcal { X } \\times \\{ 1 , . . . K \\}$ , where $\\boldsymbol { X }$ indicates the variable of instances, $\\mathbf { Y }$ is the variable of labels, $\\mathcal { X }$ denotes the feature space, and $K$ is the number of classes. In many real-world problems, examples independently drawn from the distribution $D$ are unavailable. Before being observed, the clean labels are usually randomly corrupted into noisy labels. Let $\\tilde { D }$ be the distribution of the noisy example $( X , { \\tilde { Y } } )$ , where $\\tilde { Y }$ indicates the variable of noisy labels. For label-noise learning, we can only access a sample set $\\{ \\pmb { x } _ { i } , \\tilde { y } _ { i } ) \\} _ { i = 1 } ^ { n }$ independently drawn from $\\tilde { D }$ . The aim is to learn a robust classifier from the noisy sample set that can classify test instances accurately. ", + "bbox": [ + 173, + 212, + 826, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we first elaborate on the proposed progressive early stopping (PES). Then, based on PES, we provide a learning algorithm that learns with confident examples and semi-supervised learning techniques. ", + "bbox": [ + 174, + 334, + 825, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Progressive Early Stopping ", + "text_level": 1, + "bbox": [ + 176, + 393, + 401, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When trained with noisy labels, if clean labels are of majority within each noisy class, deep networks tend to first fit clean labels during an early learning stage before eventually memorizing the wrong labels, which can be explained by the memorization effect. Many current methods utilize this property to counteract the influence of noisy labels by stopping the optimization at an early learning phase. Specifically, a deep classifier can be obtained by optimizing the following objective function with a relatively small epoch number $T$ : ", + "bbox": [ + 173, + 419, + 826, + 502 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/c0c8c02abfce6194ad608c4820ea855fb6789ee70cf2039f7d84b2ec9d9be1ed.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\Theta } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( x _ { i } ; \\Theta ) , \\tilde { y } _ { i } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 405, + 511, + 589, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f ( \\cdot ; \\Theta )$ is a deep classifier with model parameters $\\Theta$ and $\\mathcal { L }$ is the cross-entropy loss. When trained with noisy data, early learning regularization (ELR) [16] reveals that, for the most commonly used cross-entropy loss, the gradient is well correlated with the correct direction at the early learning phase. Therefore, with a properly defined small epoch number $T$ , the classifier can have higher accuracy than at initialization. While, if we continue to optimize the deep model after $T$ epochs, the classifier will be able to memorize more noise labels. Therefore, it is critical to select a proper epoch number $T$ to utilize the memorization effect and alleviate the influence of noisy labels. ", + "bbox": [ + 173, + 559, + 825, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current methods typically select the epoch number $T$ by considering the network as a whole. However, as Figure 1 makes clear, the impact of noisy labels on different DNN layers are different, which implies that the traditional early stopping trick, which optimizes the whole network all at once, may make different DNN layers to be antagonistically affected by each other, thus degrading the final model performance. To this end, we propose to separate a DNN into different parts and progressively train layers in different parts with different training epochs. Specifically, assume that the whole network $f ( \\cdot ; \\Theta )$ can be constituted with $L$ DNN parts ", + "bbox": [ + 173, + 662, + 825, + 761 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/c49b4c2da11083b29e1461cfbacf31110bb24c4b4430f414be9d136f95e70e52.jpg", + "text": "$$\n\\begin{array} { l } { z _ { 1 } = f _ { 1 } ( \\pmb { x } ; \\Theta _ { 1 } ) , } \\\\ { z _ { l } = f _ { l } ( z _ { l - 1 } ; \\Theta _ { l } ) , \\quad l = 2 , \\ldots , L } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 383, + 766, + 612, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f _ { l } ( \\cdot ; \\Theta _ { l } )$ is the $l$ -th DNN part and $z _ { l }$ is the corresponding output. The output of the last part $z _ { L }$ is the prediction. The network $f ( \\cdot ; \\Theta )$ can also be represented as $f ( \\cdot ; \\Theta _ { 1 } , . . . \\Theta _ { L } )$ . To counteract the impact of noisy labels, We initially optimize the parameter $\\Theta _ { 1 }$ for the first part by training the whole network for $T _ { 1 }$ epochs with the following objective ", + "bbox": [ + 174, + 810, + 825, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/8f9f562c089aae98a23044dad475a993b51a64c95507fed67028e7a03c4884fc.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\Theta _ { 1 } \\ldots \\Theta _ { k } } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( \\pmb { x } _ { i } ; \\Theta _ { 1 } , \\ldots , \\Theta _ { L } ) , \\tilde { y } _ { i } ) .\n$$", + "text_format": "latex", + "bbox": [ + 364, + 875, + 633, + 916 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/93c2dd94c2a0f5dd8e852ee49fbcbcd3facdbcf530179d0abf899ddba8541bf1.jpg", + "image_caption": [ + "Figure 2: Performance of the traditional early stopping trick and the proposed PES on CIFAR-10 with different types of label noise. The lines present the mean of five runs. " + ], + "image_footnote": [], + "bbox": [ + 178, + 102, + 818, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we keep the obtained parameter $\\Theta _ { 1 } ^ { * }$ fixed, reinitialize and progressively learn the $l$ -th $( l =$ $2 , \\ldots , L )$ DNN part with the parameters for preceding DNN parts fixed. The training procedure is conducted with $T _ { l }$ epochs by optimizing the following objective ", + "bbox": [ + 174, + 287, + 825, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/6cf91797c88f1e24c4a6f13692f97ac06e4c0a1ac0851bf9e6baf25712dea2a2.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\Theta _ { l } \\ldots \\Theta _ { k } } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( x _ { i } ; \\Theta _ { 1 } ^ { * } , \\ldots , \\Theta _ { l - 1 } ^ { * } , \\Theta _ { l } , \\ldots , \\Theta _ { L } ) , \\tilde { y } _ { i } ) , \\quad l = 2 \\ldots L\n$$", + "text_format": "latex", + "bbox": [ + 272, + 335, + 725, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We gradually optimize the $( l + 1 )$ -th DNN part with the obtained parameter $\\Theta _ { l } ^ { * }$ fixed, the optimization is continued until all the parameters have been optimized. As elaborated above, latter DNN parts are more sensitive to noisy labels than their former counterparts. Therefore, for the above initializing optimization in Eq. (3) and the following $L - 1$ steps of optimization in Eq. (4), we gradually reduce the training epochs (i.e. $T _ { 1 } \\geq T _ { 2 } \\geq \\cdot \\cdot \\cdot \\geq T _ { L } )$ to better exploit the memorization effect. After optimization, we can obtain the final network as $f ( \\cdot , \\Theta ) = \\bar { f } ( \\cdot ; \\Theta _ { 1 } ^ { * } , \\ldots , \\Theta _ { L } ^ { * } )$ . Since this model is obtained by progressively exploiting the early stopping strategies for different DNN parts, we term the proposed method as progressive early stopping (PES). ", + "bbox": [ + 173, + 381, + 825, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explicitly verify the effectiveness of the proposed PES method, we conduct several pilot experiments, which compare the traditional early stopping and PES with label noise from different types and different levels. The results are illustrated in Figure 2, from which we can see that, compared with models trained with traditional early stopping, models trained with PES can achieve superior classification accuracy with smaller variations in all the cases. Current state-of-the-art methods [15] usually adopt models with the traditional early stopping as base models to distill confident examples and then utilize semi-supervised learning techniques by considering confident examples as labeled data and other noisy examples as unlabeled data to further improve the results. The final performance still heavily relies on the base model trained with noisy labels. By improving the performance of the base model, our method combined with semi-supervised learning techniques is able to establish new state-of-the-art results. In the following subsections, we will elaborate on how to utilize PES to distill confident examples and further combine it with semi-supervised learning techniques. ", + "bbox": [ + 173, + 497, + 825, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Learning with Confident Examples ", + "text_level": 1, + "bbox": [ + 174, + 680, + 457, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the deep network optimized with progressive early stopping, we can select confident examples to facilitate the model training. Here, confident examples refer to examples that have high probabilities with clean labels. In this paper, we treat examples whose predictions are consistent with given labels as confident examples. In addition, to make the results more robust, we generate two different augmentations for any given input and use the average prediction to decide its predicted label. Formally, we can obtain the confident example set $\\mathcal { D } _ { l }$ as ", + "bbox": [ + 174, + 704, + 825, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/f0ca8b2be795a56146eeba463a76de4df49127ec5a09e9825eff949f223573dc.jpg", + "text": "$$\n\\begin{array} { c } { \\mathcal { D } _ { l } = \\{ ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = \\hat { y } _ { i } , i = 1 , \\ldots , n \\} , } \\\\ { \\hat { y } _ { i } = \\underset { k \\in \\{ 1 , \\ldots , K \\} } { \\arg \\operatorname* { m a x } } \\ \\frac { 1 } { 2 } [ f ^ { k } ( \\mathrm { A u g m e n t } ( \\pmb { x } _ { i } ) ; \\Theta ) + f ^ { k } ( \\mathrm { A u g m e n t } ( \\pmb { x } _ { i } ) ; \\Theta ) ] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 281, + 803, + 715, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Augment $( \\cdot )$ indicates normal data augmentation operation including horizontal random flip and random crops, and $f ^ { k } ( x ; \\Theta )$ is the predicted probability of $_ { \\textbf { \\em x } }$ belonging to class $k$ . Note that Augment $\\left( \\cdot \\right)$ is a stochastic transformation, so the two terms in Eq (5) are not identical. The average ", + "bbox": [ + 176, + 868, + 825, + 912 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Neural network with trainable parameters $\\boldsymbol { \\Theta } = \\{ \\Theta _ { 1 } , \\dots , \\Theta _ { L } \\}$ , Noisy training dataset \n$\\{ \\bar { \\pmb { x } _ { i } } , \\tilde { y } _ { i } ) \\} _ { i = 1 } ^ { n }$ , Number of training epochs for different part: $T _ { 1 } , \\dots , T _ { L }$ , and training epochs $T _ { c }$ \nfor refining with confident examples. \nfor $i = 1 , \\dots , T _ { 1 }$ do \nOptimize network parameter $\\Theta$ with Eq. (3); \nfor $l = 2$ , . . . , $L$ do Froze $\\left\\{ \\Theta _ { 1 } , \\ldots , \\Theta _ { l - 1 } \\right\\}$ and re-initialize $\\big \\{ \\Theta _ { l } , \\dots , \\Theta _ { L } \\big \\}$ ; for $i = 1 , \\dots , T _ { l }$ do Optimize network parameter $\\big \\{ \\Theta _ { l } , \\dots , \\Theta _ { L } \\big \\}$ with Eq. (4); \nUnfroze $\\Theta$ ; \nfor $i = 1 , \\dots , T _ { c }$ do Extract confident example set $\\mathcal { D } _ { l }$ and unlabeled set $\\mathcal { D } _ { u }$ with classifier $f ( \\cdot , \\Theta )$ by Eq. (7); Training the classifier $f ( \\cdot , \\Theta )$ with MixMatch loss on $\\mathcal { D } _ { l }$ and $\\mathcal { D } _ { u }$ ; \nEvaluate the obtained classifier $f ( \\cdot , \\Theta )$ . ", + "bbox": [ + 171, + 112, + 792, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prediction of augmented examples provides a more stable prediction and is found empirically to improve performance. After obtaining the confident example set, one can easily train a classifier by considering confident examples as clean data. However, since the number of confident examples for different classes can vary greatly, directly training the model with the obtained confident example set may introduce a severe class imbalance problem. To this end, we adopt a weighted classification loss ", + "bbox": [ + 173, + 353, + 825, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/2d952161669f286508b4ff5d83652860d38fb213b9f942b25cfcf93e6118bade.jpg", + "text": "$$\n\\mathcal { L } _ { c } = \\sum _ { i = 1 } ^ { N } w _ { y _ { i } } \\mathcal { L } _ { p } ( \\tilde { y } _ { i } , f ( \\pmb { x } _ { i } ; \\Theta ) ) ,\n$$", + "text_format": "latex", + "bbox": [ + 393, + 439, + 602, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w _ { i }$ is the corresponding class weight. Assuming that $\\sigma _ { k } = | \\{ ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = k , ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) \\in \\mathcal { D } _ { l } \\} |$ denotes the cardinality of the confident example set belonging to the $k$ -th class. Then, we can set $w _ { i } = \\sigma _ { i } / ( \\sum _ { j = 1 } ^ { K } \\sigma _ { j } )$ to indicate the corresponding class importance. ", + "bbox": [ + 173, + 484, + 825, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Combining with Semi-Supervised Learning ", + "text_level": 1, + "bbox": [ + 173, + 546, + 516, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training with only confident examples neglects the rest data and may suffer from insufficient training examples. To tackle this problem, we further resort to semi-supervised learning techniques by considering confident examples as labeled data and other noisy examples as unlabeled data. Specifically, the labeled data set and unlabeled data set can be obtained as ", + "bbox": [ + 176, + 571, + 825, + 627 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a92713a47a810a07f4fbd6225dc073925113f0874b4ff035aa8d78e39c94acb1.jpg", + "text": "$$\n\\begin{array} { c } { { \\displaystyle { \\int } { \\mathcal D } _ { l } = \\{ ( \\boldsymbol x _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = \\hat { y } _ { i } , i = 1 , . . . n \\} } } \\\\ { { \\displaystyle { \\mathcal D } _ { u } = \\{ \\boldsymbol x _ { i } | \\tilde { y } _ { i } \\neq \\hat { y } _ { i } , i = 1 , . . . n \\} } } \\\\ { { \\displaystyle { \\hat { y } _ { i } = \\arg \\operatorname* { m a x } _ { k \\in \\{ 1 , . . . K \\} } \\frac { 1 } { 2 } [ f ^ { k } ( \\mathrm { A u g m e n t } ( \\boldsymbol x _ { i } ) ; \\Theta ) + f ^ { k } ( \\mathrm { A u g m e n t } ( \\boldsymbol x _ { i } ) ; \\Theta ) ] } , } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 282, + 633, + 714, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the labeled data set $\\mathcal { D } _ { l }$ is the same as that in Eq (6), and $\\mathcal { D } _ { u }$ is the rest unlabeled data set. Similar to [15], we adopt MixMatch [4] as the semi-supervised learning framework to train the final classification models. For more details about semi-supervised learning, we refer to [4]. The whole learning algorithm is summarized in Algorithm 1. ", + "bbox": [ + 178, + 710, + 820, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Experiments ", + "text_level": 1, + "bbox": [ + 174, + 785, + 312, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Datasets and Implementation Details ", + "text_level": 1, + "bbox": [ + 176, + 815, + 470, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: We evaluate our method on two synthetic datasets, CIFAR-10 and CIFAR-100 [12] with different levels of symmetric, pairflip, and instance-dependent label noise (abbreviated as instance label noise) and a real-world dataset Clothing-1M [32]. Both CIFAR-10 and CIFAR-100 contain $5 0 \\mathrm { k }$ training images and $1 0 \\mathrm { k }$ test images of size $3 2 \\times 3 2$ . Following previous works [9, 31, 16, 29], symmetric noise is generated by uniformly flipping labels for a percentage of the training dataset to all possible labels. Pairflip noise flips noisy labels into their adjacent class. And, instance noise is generated by image features. More details about the synthetic label noise are given in the supplementary material. For the flipping rate, it can include [9, 31] or ex-include [15, 16] true labels. We use the flipping rate including correct labels in Table 3 to compare with results in [15], and use without correct labels in the rest of the experiments. Clothing-1M [32] is a large-scale dataset with real-world noisy labels, whose images are clawed from the online shopping websites, and labels are generated based on surrounding texts. It contains 1 million training images, and $1 5 \\mathrm { k }$ validation images, and 10k test images with clean labels. ", + "bbox": [ + 174, + 842, + 825, + 912 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/00db2b4f97c9dec5546d30e722898d28a5de7b60aab6b4c0b55768812cd9dac7.jpg", + "table_caption": [ + "Table 1: Preliminary analysis of the performance and the quality of extracted confident examples on CIFAR-10. The mean and standard deviation are computed over five runs. " + ], + "table_footnote": [], + "table_body": "
MetricsMethodsSym-20%Sym-50%Pair-45%Inst-20%Inst-40%
Test AccuracyEarly StoppingPES82.55±2.4685.87±1.5970.76±1.2475.87±1.3360.62±5.5962.40±2.3484.41±0.9086.58±0.4574.73±2.6577.07±1.18
Label PrecisionEarly StoppingPES98.81±0.1598.96±0.0994.65±0.1995.46±0.1472.53±5.2672.99±2.2798.70±0.4398.52±0.1990.77±1.8790.63±0.92
Label RecallEarly StoppingPES88.51±2.2692.67±1.4375.18±1.0081.03±1.8367.84±5.0671.06±2.2790.37±1.0193.24±0.6082.15±3.1785.91±0.68
", + "bbox": [ + 176, + 131, + 823, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 262, + 825, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: Semi-supervised learning may strongly boost the performance, we separately compare our method with approaches with or without semi-supervised learning. For the comparison with baselines with semi-supervised learning, we combine our proposed method with MixMatch used in [15] as indicated in Subsection 2.3. (1) Approaches without semi-supervised learning: Co-teaching [9], Forward [24], Joint Optim [25], T-revision [31], DMI [34], and CDR [29]. (2) Methods with semi-supervised learning: M-correction [1], DivideMix [15], and $\\mathrm { E L R + }$ [16]. We also adopt standard training with cross-entropy (CE) and MixUp [37] as baselines to show improvements. ", + "bbox": [ + 173, + 380, + 825, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network structure and optimization: Our method is implemented by PyTorch v1.6. Baseline methods are implemented based on public codes with hyper-parameters set according to the original papers. For DivideMix and $\\mathrm { E L R + }$ , we evaluate the test accuracy with the first network. To better demonstrate the robustness of our algorithm, we keep the hyper-parameters fixed for different types of label noise. More technique details are given in the supplementary material. ", + "bbox": [ + 174, + 483, + 825, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For experiments without semi-supervised learning, we follow [31], and use ResNet-18 [10] for CIFAR-10 and ResNet-34 for CIFAR-100. We split networks into three parts, the layers above block 4 as part 1, block 4 of ResNet as part 2, and the final layer as part 3. $T _ { 1 }$ is defined as 25 for CIFAR-10 and 30 for CIFAR-100, $T _ { 2 }$ as 7, and $T _ { 3 }$ as 5. The network is trained for 200 epochs and SGD with 0.9 momentum is used. The initial learning rate is set to 0.1 and decayed with a factor of 10 at the 100th and 150th epoch respectively, and a weight decay is set to $1 0 ^ { - 4 }$ . For $T _ { 2 }$ and $T _ { 3 }$ , we employ an Adam optimizer with a learning rate of $1 0 ^ { - 4 }$ . ", + "bbox": [ + 173, + 559, + 825, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For experiments with semi-supervised learning, we follow the setting of [15] with PreAct Resnet-18. We set the final layer as part 2, the rest as part 1. $T _ { 1 }$ is defined as 20 for CIFAR-10 and 35 for CIFAR-100, and $T _ { 2 }$ as 5. The network is trained for 300 epochs. For optimization, we use a single cycle of cosine annealing [19], and the learning rate begins from $2 \\times \\mathrm { \\bar { 1 0 ^ { - 2 } } }$ and ends at $2 \\times 1 0 ^ { - 4 }$ , with a weight decay of $5 \\times 1 0 ^ { - 4 }$ . An Adam optimizer is adopted with a learning rate of $1 0 ^ { - 4 }$ for $T _ { 2 }$ . For hyper-parameters from MixMatch, we set them according to the original paper [4]. ", + "bbox": [ + 174, + 662, + 826, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For Clothing-1M [32], we follow the previous work [25], and employ a ResNet-50 [10] pre-trained on ImageNet [13]. We set the final layer as part 2, the rest as part 1. $T _ { 1 }$ and $T _ { 2 }$ are defined as 20 and 7 respectively. The network is trained with CE loss for 50 epochs and SGD is used with 0.9 momentum and a weight decay of $1 0 ^ { - 3 }$ . The learning rate is $5 \\times 1 0 ^ { - 3 }$ and decayed by a factor of 10 at the 20th and 30th epoch respectively. We employ an Adam optimizer with a learning rate of $5 \\times 1 0 ^ { - 6 }$ for $T _ { 2 }$ ", + "bbox": [ + 174, + 752, + 825, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Preliminary Experiments ", + "text_level": 1, + "bbox": [ + 174, + 842, + 390, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 2, we can observe that with the PES trick, the performance of classifiers is generally improved compared with that the traditional early stopping trick. In this section, we further carefully analyze the quality of extracted labels by examining them from three aspects, i.e., test accuracy, label precision, and label recall. Here, label precision indicates the ratio of the number of extracted confident examples with correct labels in the total confident example set, and label recall represents the ratio of the number of confident examples with correct labels among the total correctly labeled examples. Specifically, we train a neural network on CIFAR-10 with different kinds and levels of label noise for 25 epochs respectively and report the performance for each case before and after the proposed PES is applied. ", + "bbox": [ + 174, + 869, + 825, + 911 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/faf2277129f4013a50d7d87c0c93cc22f297e8287191fc976ac1e57bb5e493c0.jpg", + "table_caption": [ + "Table 2: Comparison with state-of-the-art methods without semi-supervised learning on CIFAR-10 and CIFAR-100. The mean and standard deviation computed over five runs are presented. " + ], + "table_footnote": [], + "table_body": "
DatasetMethodSymmetricPairflipInstance
20%50%45%20%40%
CIFAR10CE84.00±0.6675.51±1.2463.34±6.0385.10±0.6877.00±2.17
Co-teaching87.16±0.1172.80±0.4570.11±1.1686.54±0.1180.98±0.39
Forward85.63±0.5277.92±0.6660.15±1.9785.29±0.3874.72±3.24
Joint Optim89.70±0.1185.00±0.1782.63±1.3889.69±0.4282.62±0.57
T-revision89.63±0.1383.40±0.6577.06±6.4790.46±0.1385.37±3.36
DMI88.18±0.3678.28±0.4857.60±14.5689.14±0.3684.78±1.97
CDR89.72±0.3882.64±0.8973.67±0.5490.41±0.3483.07±1.33
Ours92.38±0.4087.45±0.3588.43±1.0892.69±0.4489.73±0.51
CIFAR100CE51.43±0.5837.69±3.4534.10±2.0452.19±1.4242.26±1.29
Co-teaching59.28±0.4741.37±0.0833.22±0.4857.24±0.6945.69±0.99
Forward57.75±0.3744.66±1.0127.88±0.8058.76±0.6644.50±0.72
Joint Optim64.55±0.3850.22±0.4142.61±0.6165.15±0.3155.57±0.41
T-revision65.40±1.0750.24±1.4541.10±1.9560.71±0.7351.54±0.91
DMI58.73±0.7044.25±1.1426.90±0.4558.05±0.2047.36±0.68
CDR66.52±0.2455.30±0.9643.87±1.3567.33±0.6755.94±0.56
Ours68.89±0.4558.90±2.7257.18±1.4470.49±0.7965.68±1.41
", + "bbox": [ + 189, + 131, + 803, + 400 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/db3492c5afea51549904ef40bb3ab4ac7a34d054f0b1e535e19df964e1db2048.jpg", + "table_caption": [ + "Table 3: Comparison with state-of-the-art methods with semi-supervised learning on CIFAR-10 and CIFAR-100 with symmetric label noise from different levels. Results with \\* are token from [15]. The mean and standard deviation are computed over three runs. " + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10CIFAR-100
Methods /NoiseSym-20%Sym-50%Sym-80%Sym-20%Sym-50%Sym-80%
CE86.5±0.680.6±0.263.7±0.857.9±0.447.3±0.222.3±1.2
MixUp93.2±0.388.2±0.373.3±0.369.5±0.257.1±0.634.1±0.6
M-correction*94.092.086.873.966.148.2
DivideMix*95.294.293.075.272.858.3
DivideMix95.6±0.194.6±0.192.9±0.375.3±0.172.7±0.656.4±0.3
ELR+94.9±0.293.6±0.190.4±0.275.5±0.271.0±0.250.4±0.8
Ours (Semi)95.9±0.195.1±0.293.1±0.277.4±0.374.3±0.661.6±0.6
", + "bbox": [ + 214, + 463, + 784, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 612, + 825, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results in Table 1 clearly show that, compared with the traditional early stopping, PES can help to obtain higher accuracies, precisions, and recalls for most cases. For instance-dependent label noise, PES can achieve higher recall values with comparable label precision values. Note that models with high recall values can help to collect more confident examples, which is critical for learning with confident examples and semi-supervised learning. Therefore, by enhancing the performance of the initial model, PES can help to improve the final classification performance in all cases, which is also verified by the experiments in Section 3.3. ", + "bbox": [ + 174, + 702, + 825, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Classification Accuracy Evaluation ", + "text_level": 1, + "bbox": [ + 176, + 815, + 457, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic datasets. We first verify the effectiveness of our proposed method without semi-supervised learning techniques on two synthetic datasets: CIFAR-10 and CIFAR-100. For both of these two datasets, we leave $10 \\%$ of data with noisy labels as noisy validation set. Results are presented in Table 2, which shows that our proposed method can consistently outperform all other baselines across various settings by a large margin. ", + "bbox": [ + 174, + 842, + 825, + 911 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/30f68d18682a41f23ee5c16e39c70d68c6be5ad1521db00b67346216a25d82d5.jpg", + "table_caption": [ + "Table 4: Comparison with state-of-the-art methods with semi-supervised learning on CIFAR-10 and CIFAR-100 with instance-dependent and pairflip label noise from different levels. The mean and standard deviation are computed over three runs. " + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10CIFAR-100
Methods/NoiseInst-20%Inst-40%Pair-45%Inst-20%Inst-40%Pair-45%
CEMixUpDivideMixELR+87.5±0.578.9±0.774.9±1.756.8±0.448.2±0.538.5±0.6
93.3±0.295.5±0.187.6±0.594.5±0.282.4±1.085.6±1.767.1±0.155.0±0.170.9±0.144.2±0.548.2±1.0
75.2±0.270.9±0.1
94.9±0.194.3±0.286.1±1.275.8±0.174.3±0.365.3±1.3
Ours (Semi)95.9±0.195.3±0.194.5±0.377.6±0.376.1±0.473.6±1.7
", + "bbox": [ + 222, + 145, + 772, + 252 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd26d18ed5ae3d06898df87d4bef68e8bf94d61f481d809a16bbd2362705f40d.jpg", + "table_caption": [ + "Table 5: Compassion with state-of-the-art methods on Clothing-1M. Results of baseline methods are taken from the original papers. ours represent the results obtained by PES with a single network and ours\\* indicate the results obtained by PES with an ensemble model. " + ], + "table_footnote": [], + "table_body": "
CEForwardJoint-OptimDMIT-revisionDivideMix*ELR+*OursOurs*
69.2169.8472.1672.4674.1874.7674.8174.6474.99
", + "bbox": [ + 207, + 324, + 790, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61ae512b2fb46bee07bd5e6157d516e400cd564030c729a91354232a89157ebe.jpg", + "image_caption": [ + "Figure 3: Sensitivity analysis for different training iteration numbers: $T _ { 2 }$ and $T _ { 3 }$ " + ], + "image_footnote": [], + "bbox": [ + 284, + 386, + 712, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 and Table 4 present the mean accuracy and standard deviation for our method and all baselines on CIFAR-10 and CIFAR-100, respectively. From the results, we can get that the proposed method can outperform all baselines in all cases. For pairflip label noise, the advantages of our proposed method become more apparent, and it significantly outperforms state-of-the-art methods by over $8 \\%$ on both CIFAR-10 and CIFAR-100. These empirical results support our proposal that PES can improve the quality of selected confident examples, which helps improve performance and reduce the variance of the final classifier. ", + "bbox": [ + 173, + 582, + 825, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-world dataset. We evaluate the performance of the proposed method on a real-world dataset with Clothing-1M [32] and select methods such as CE, Forward, Joint-Optim, DMI, and T-revision, which use a single network, and also methods such as DivideMix and $\\mathrm { E L R + }$ , which adopt an ensemble model with two different networks, as baselines. We also report the results for the proposed PES with a single network as ours and the results for PES, which ensembles two networks, as ours\\*. The overall results are reported in Table 5, from which we can observe that the proposed PES with a single network can outperform all baselines using a single network. And with an ensemble model, which contains two different networks, our method can outperform all the adopted baselines. These results clearly demonstrate that, by improving the performance of the initial classification network, our method is more flexible to handle such real-world noise problems. ", + "bbox": [ + 174, + 685, + 825, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Sensitivity Analysis ", + "text_level": 1, + "bbox": [ + 174, + 843, + 348, + 858 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we investigate the hyper-parameter sensitivity for the training iteration number $T _ { 2 }$ and $T _ { 3 }$ , respectively. We firstly analyze the training epoch number for the second DNN part by varying $T _ { 2 }$ from the range of [0, 10]. The results are illustrated in Figure 3a, from which can find that, with the increasing of $T _ { 2 }$ , the performance of PES first increase and then decrease in all the cases except for $45 \\%$ Pairflip noise on the CIFAR-10 dataset. While the model achieves the best performance with $T _ { 2 }$ as 7 for all types of noisy labels. Then we fix $T _ { 2 }$ as 7, and analyze the impact of the third DNN part by varying $T _ { 3 }$ from the range of [0, 10]. The results are shown in Figure 3b. Although the performance variance for different $T _ { 3 }$ is smaller than that for $T _ { 2 }$ , we can still observe that the best performance can be obtained when $T _ { 3 }$ is set as 5. More importantly, from these two figures, we can get that both $T _ { 2 }$ and $T _ { 3 }$ are robust to the different types of noisy labels. ", + "bbox": [ + 176, + 869, + 825, + 911 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1be43f84b04f81deb931ee482ba81d70209b96391a05c3c3cca2b69977f03a7a.jpg", + "table_caption": [ + "Table 6: Training time comparison for baselines on CIFAR-10 with $50 \\%$ Symmetric label noise. " + ], + "table_footnote": [], + "table_body": "
CECo-teachingCDRT-revisionELR+DivideMixOursOurs (Semi)
0.9h1.5h3.0h3.5h2.2h5.5h1.0h3.1h
", + "bbox": [ + 225, + 117, + 763, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 161, + 825, + 260 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training Time Comparison ", + "text_level": 1, + "bbox": [ + 174, + 275, + 401, + 290 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we compare the training time of our method and other state-of-the-art baselines. All the experiments are conducted on a server with a single Nvidia V100 GPU. The training times for all the methods are reported in Table 6, from which we can get that our algorithm with cross-entropy loss achieves the fastest speed across all baselines, only about 1 hour. Our method combining with MixMatch [4] is also fast, only a little more than half of the training time of DivideMix. The time of $\\mathrm { E L R + }$ [16] shows superior, but $\\mathrm { E L R + }$ trains the network with fewer epochs, with 200 epochs compared with ours for 300 epochs. ", + "bbox": [ + 174, + 300, + 825, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Related work ", + "text_level": 1, + "bbox": [ + 174, + 416, + 316, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Learning with noisy data has been well studied [17, 6, 21, 27, 20]. Current works can be mainly categorized into two groups: model-based and model-free methods. In this section, we briefly review some closely related works. ", + "bbox": [ + 174, + 446, + 825, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The first type models the relationship between clean labels and noisy labels by estimating the noise transition matrix and build a loss function to correct the loss [24, 30, 35, 28]. [24] first combines algorithms for estimating the noise rates and loss correction techniques together and introduces two alternative procedures for loss correction. It also proves that both of the two procedures enjoy formal robustness guarantees w.r.t. the clean data distribution. DMI [34] proposes an information-theoretic loss function, which utilizes Shannon’s mutual information and is robustness to different kinds of label noise. T-revision [31] estimates the noise transition matrix without anchor points by adding a fine-tuned slack variables. Although these methods have made certain progress, they are usually fragile to estimate the noise transition matrix for heavy noisy data and are also hard to handle a large number of classes. Therefore, in this paper, we mainly focus on the model-free methods. ", + "bbox": [ + 174, + 494, + 825, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The second strand mainly counteracts noisy labels by exploiting the memorization effect that deep networks tend to first memorize and fit majority (clean) patterns and then overfit minority (noisy) patterns [2]. To exploit this property, Co-teaching [9] employs two networks with different initialization and uses small loss to select confident examples. M-correction [1] uses two Gaussian Mixture Models to identify confident examples, instead of using networks themselves. DivideMix [15] extends Co-teaching [9] and employs two Beta Mixture Model to select confident examples. MixMatch [4] is then adopted to leverage unconfident examples with a semi-supervised learning framework. All the above methods exploit the memorization effect by considering the adopted network as a whole. Recently, [14] shows that networks training with noisy labels can produce good representations, if the structure of networks suits the targeted tasks. Our method further explains that noisy labels have different impacts for different layers in a DNN. And latter layers will receive earlier and more severe impact than their former counterparts. Therefore, by considering a DNN as a composition of several layers and training different layers with different epochs, our method is able to better exploit the memorization effect and achieve superior performance. ", + "bbox": [ + 173, + 640, + 826, + 833 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Conclusion ", + "text_level": 1, + "bbox": [ + 174, + 852, + 299, + 869 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we provide a progressive early stopping (PES) method to better exploit the memorization effect of deep neural networks (DNN) for noisy-label learning. We first find that the impact of noisy labels for former layers in a DNN is much less and later than that for latter DNN layers, and then build upon this insight to propose the PES method, which separates a DNN into different parts and progressively train each part to counteract the different impacts of noisy labels for different DNN layers. To show that PES can boost the performance of state-of-the-art methods, we conduct extensive experiments across multiple synthetic and real-world noisy datasets and demonstrate that the proposed PES can help to obtain substantial performance improvements compared to current state-of-the-art baselines. The main limitation of our method lies in that, by splitting a DNN into different parts, PES introduces several additional hyper-parameters that need to be tuned carefully. In the future, we will extend the work in the following aspects. First, we will study other mechanisms that distinguishing desired and undesired memorization rather than early stopping, e.g., the gradient ascent trick [8]. Second, we are interested in combining PES with interesting ideas from semi-supervised learning and unsupervised learning. ", + "bbox": [ + 173, + 883, + 821, + 911 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 90, + 825, + 257 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Acknowledgments and Disclosure of Funding ", + "text_level": 1, + "bbox": [ + 174, + 276, + 553, + 294 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "YB was partially supported by Agriculture Consultant and Smart Management. BH was supported by the RGC Early Career Scheme No. 22200720, NSFC Young Scientists Fund No. 62006202 and HKBU CSD Departmental Incentive Grant. YY was partially supported by Key Research and Development Program of Shaanxi (ProgramNo. 2021ZDLGY01-03). GN was supported by JST AIP Acceleration Research Grant Number JPMJCR20U3, Japan. TL was partially supported by Australian Research Council Projects DE-190101473 and IC-190100031. ", + "bbox": [ + 174, + 308, + 825, + 391 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "References \n[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O’Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In ICML, pages 312–321, 2019. \n[2] Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S. Kanwal, Tegan Maharaj, Asja Fischer, Aaron C. Courville, Yoshua Bengio, and Simon Lacoste-Julien. A closer look at memorization in deep networks. In ICML, pages 233–242, 2017. \n[3] Mikhail Belkin, Daniel Hsu, Siyuan Ma, and Soumik Mandal. Reconciling modern machine-learning practice and the classical bias–variance trade-off. Proceedings of the National Academy of Sciences, 116(32):15849–15854, 2019. \n[4] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, pages 5050–5060, 2019. \n[5] Youngchul Cha and Junghoo Cho. Social-network analysis using topic models. In SIGIR, pages 565–574, 2012. \n[6] Jacob Goldberger and Ehud Ben-Reuven. Training deep neural-networks using a noise adaptation layer. In ICLR, 2017. \n[7] Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning. The MIT Press, 2016. \n[8] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor W. Tsang, and Masashi Sugiyama. SIGUA: forgetting may make learning with noisy labels more robust. In ICML, pages 4006–4016, 2020. \n[9] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages 8527–8537, 2018. \n[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. \n[11] Takashi Ishida, Ikko Yamane, Tomoya Sakai, Gang Niu, and Masashi Sugiyama. Do we need zero training loss after achieving zero training error? In ICML, pages 4604–4614, 2020. \n[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. \n[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, pages 1097–1105, 2012. \n[14] Jingling Li, Mozhi Zhang, Keyulu Xu, John P. Dickerson, and Jimmy Ba. Noisy labels can induce good representations. arXiv preprint arXiv:2012.12896, 2020. \n[15] Junnan Li, Richard Socher, and Steven C. H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR, 2020. \n[16] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. In NeurIPS, pages 20331–20342, 2020. \n[17] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on pattern analysis and machine intelligence, 38(3):447–461, 2016. \n[18] Wei Liu, Yu-Gang Jiang, Jiebo Luo, and Shih-Fu Chang. Noise resistant graph ranking for improved web image search. In CVPR, pages 849–856, 2011. \n[19] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In ICLR, 2017. \n[20] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah M. Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. In ICML, pages 6543–6553, 2020. \n[21] Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah M. Erfani, Shu-Tao Xia, Sudanthi N. R. Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In ICML, pages 3361–3370, 2018. \n[22] Preetum Nakkiran, Gal Kaplun, Yamini Bansal, Tristan Yang, Boaz Barak, and Ilya Sutskever. Deep double descent: Where bigger models and more data hurt. In ICLR, 2020. \n[23] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi-Phuong-Nhung Ngo, Thi Hoai Phuong Nguyen, Laura Beggel, and Thomas Brox. SELF: learning to filter noisy labels with self-ensembling. In ICLR, 2020. \n[24] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, pages 2233–2241, 2017. \n[25] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, pages 5552–5560, 2018. \n[26] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, pages 8688–8696, 2018. \n[27] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In ICCV, pages 322–330, 2019. \n[28] Songhua Wu, Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Nannan Wang, Haifeng Liu, and Gang Niu. Class2simi: A noise reduction perspective on learning with noisy labels. In ICML, pages 11285–11295, 2021. \n[29] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hindering the memorization of noisy labels. In ICLR, 2021. \n[30] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, pages 7597–7610, 2020. \n[31] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, pages 6835–6846, 2019. \n[32] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In CVPR, pages 2691–2699, 2015. \n[33] Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In NeurIPS, pages 6256–6268, 2020. \n[34] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L_dmi: A novel information-theoretic loss function for training deep nets robust to label noise. In NeurIPS, pages 6222–6233, 2019. \n[35] Yu Yao, Tongliang Liu, Bo Han, Mingming Gong, Jiankang Deng, Gang Niu, and Masashi Sugiyama. Dual T: reducing estimation error for transition matrix in label-noise learning. In NeurIPS, pages 7260–7271, 2020. \n[36] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In ICLR, 2016. \n[37] Hongyi Zhang, Moustapha Cissé, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. ", + "bbox": [ + 171, + 88, + 828, + 917 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 89, + 826, + 670 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_middle.json b/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..67e31a886a7c47512b4e70a423f3dafb8ebc8484 --- /dev/null +++ b/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_middle.json @@ -0,0 +1,29811 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 122, + 97, + 490, + 137 + ], + "lines": [ + { + "bbox": [ + 120, + 95, + 492, + 119 + ], + "spans": [ + { + "bbox": [ + 120, + 95, + 492, + 119 + ], + "score": 1.0, + "content": "Understanding and Improving Early Stopping for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 118, + 410, + 138 + ], + "spans": [ + { + "bbox": [ + 201, + 118, + 410, + 138 + ], + "score": 1.0, + "content": "Learning with Noisy Labels", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 179, + 178, + 435, + 203 + ], + "lines": [ + { + "bbox": [ + 179, + 177, + 429, + 191 + ], + "spans": [ + { + "bbox": [ + 179, + 177, + 313, + 191 + ], + "score": 1.0, + "content": "Yingbin Bai1∗ Erkun Yang2∗", + "type": "text" + }, + { + "bbox": [ + 315, + 178, + 356, + 190 + ], + "score": 1.0, + "content": "Bo Han3", + "type": "text" + }, + { + "bbox": [ + 365, + 178, + 429, + 191 + ], + "score": 1.0, + "content": "Yanhua Yang2", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 190, + 437, + 203 + ], + "spans": [ + { + "bbox": [ + 179, + 190, + 233, + 203 + ], + "score": 1.0, + "content": "Jiatong Li4", + "type": "text" + }, + { + "bbox": [ + 240, + 190, + 302, + 203 + ], + "score": 1.0, + "content": "Yinian Mao4", + "type": "text" + }, + { + "bbox": [ + 307, + 190, + 357, + 203 + ], + "score": 1.0, + "content": "Gang Niu5", + "type": "text" + }, + { + "bbox": [ + 365, + 190, + 437, + 203 + ], + "score": 1.0, + "content": "Tongliang Liu1†", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 131, + 206, + 480, + 230 + ], + "lines": [ + { + "bbox": [ + 129, + 204, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 129, + 204, + 482, + 221 + ], + "score": 1.0, + "content": "1TML Lab, University of Sydney; 2Xidian University; 3Hong Kong Baptist University;", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 221, + 216, + 390, + 232 + ], + "spans": [ + { + "bbox": [ + 221, + 216, + 390, + 232 + ], + "score": 1.0, + "content": "4Meituan-Dianping Group; 5RIKEN AIP", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 283, + 258, + 328, + 271 + ], + "lines": [ + { + "bbox": [ + 282, + 258, + 330, + 272 + ], + "spans": [ + { + "bbox": [ + 282, + 258, + 330, + 272 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 283, + 469, + 501 + ], + "lines": [ + { + "bbox": [ + 141, + 282, + 469, + 296 + ], + "spans": [ + { + "bbox": [ + 141, + 282, + 469, + 296 + ], + "score": 1.0, + "content": "The memorization effect of deep neural network (DNN) plays a pivotal role in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 294, + 469, + 307 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 469, + 307 + ], + "score": 1.0, + "content": "many state-of-the-art label-noise learning methods. To exploit this property, the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 304, + 470, + 318 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 470, + 318 + ], + "score": 1.0, + "content": "early stopping trick, which stops the optimization at the early stage of training,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 316, + 470, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 316, + 470, + 329 + ], + "score": 1.0, + "content": "is usually adopted. Current methods generally decide the early stopping point", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 327, + 470, + 340 + ], + "spans": [ + { + "bbox": [ + 141, + 327, + 470, + 340 + ], + "score": 1.0, + "content": "by considering a DNN as a whole. However, a DNN can be considered as a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 338, + 470, + 351 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 470, + 351 + ], + "score": 1.0, + "content": "composition of a series of layers, and we find that the latter layers in a DNN are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 349, + 470, + 361 + ], + "spans": [ + { + "bbox": [ + 141, + 349, + 470, + 361 + ], + "score": 1.0, + "content": "much more sensitive to label noise, while their former counterparts are quite robust.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 359, + 470, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 359, + 470, + 372 + ], + "score": 1.0, + "content": "Therefore, selecting a stopping point for the whole network may make different", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 370, + 470, + 383 + ], + "spans": [ + { + "bbox": [ + 141, + 370, + 470, + 383 + ], + "score": 1.0, + "content": "DNN layers antagonistically affect each other, thus degrading the final performance.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 381, + 469, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 469, + 394 + ], + "score": 1.0, + "content": "In this paper, we propose to separate a DNN into different parts and progressively", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 392, + 470, + 406 + ], + "spans": [ + { + "bbox": [ + 141, + 392, + 470, + 406 + ], + "score": 1.0, + "content": "train them to address this problem. Instead of the early stopping which trains a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 402, + 469, + 416 + ], + "spans": [ + { + "bbox": [ + 141, + 402, + 469, + 416 + ], + "score": 1.0, + "content": "whole DNN all at once, we initially train former DNN layers by optimizing the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 414, + 469, + 427 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 469, + 427 + ], + "score": 1.0, + "content": "DNN with a relatively large number of epochs. During training, we progressively", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 425, + 469, + 438 + ], + "spans": [ + { + "bbox": [ + 141, + 425, + 469, + 438 + ], + "score": 1.0, + "content": "train the latter DNN layers by using a smaller number of epochs with the preceding", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 437, + 469, + 448 + ], + "spans": [ + { + "bbox": [ + 141, + 437, + 469, + 448 + ], + "score": 1.0, + "content": "layers fixed to counteract the impact of noisy labels. We term the proposed method", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 447, + 470, + 460 + ], + "spans": [ + { + "bbox": [ + 141, + 447, + 470, + 460 + ], + "score": 1.0, + "content": "as progressive early stopping (PES). Despite its simplicity, compared with the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 458, + 470, + 471 + ], + "spans": [ + { + "bbox": [ + 141, + 458, + 470, + 471 + ], + "score": 1.0, + "content": "traditional early stopping, PES can help to obtain more promising and stable results.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 469, + 470, + 482 + ], + "spans": [ + { + "bbox": [ + 141, + 469, + 470, + 482 + ], + "score": 1.0, + "content": "Furthermore, by combining PES with existing approaches on noisy label training,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 480, + 469, + 492 + ], + "spans": [ + { + "bbox": [ + 142, + 480, + 469, + 492 + ], + "score": 1.0, + "content": "we achieve state-of-the-art performance on image classification benchmarks. The", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 491, + 391, + 502 + ], + "spans": [ + { + "bbox": [ + 141, + 491, + 391, + 502 + ], + "score": 1.0, + "content": "code is made public at https://github.com/tmllab/PES.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 107, + 522, + 190, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 192, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 192, + 538 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "Deep networks have revolutionized a wide variety of tasks, such as image processing, speech", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "recognition, and language modeling [7], However, this highly relies on the availability of large", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "annotated data, which may not be feasible in practice. Instead, many large datasets with lower", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 580, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 505, + 593 + ], + "score": 1.0, + "content": "quality annotations are collected from online queries [5] or social-network tagging [18]. Such", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "annotations inevitably contain mistakes or label noise. As deep networks have large model capacities,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "they can easily memorize and eventually overfit the noisy labels, leading to poor generalization", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "performance [36]. Therefore, it is of great importance to develop a methodology that is robust to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 624, + 182, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 182, + 636 + ], + "score": 1.0, + "content": "noisy annotations.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 108, + 640, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 507, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 507, + 653 + ], + "score": 1.0, + "content": "Existing methods on learning with noisy labels (LNL) can be mainly categorized into two groups:", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "model-based and model-free algorithms. Methods in the first category mainly model noisy labels with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "the noise transition matrix [24, 34, 33, 30]. With perfectly estimated noise transition matrix, models", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 673, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 506, + 684 + ], + "score": 1.0, + "content": "trained with corrected losses can approximate to the models trained with clean labels. However,", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 692, + 358, + 713 + ], + "lines": [ + { + "bbox": [ + 117, + 689, + 177, + 703 + ], + "spans": [ + { + "bbox": [ + 117, + 689, + 177, + 703 + ], + "score": 1.0, + "content": "∗co-first author", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 700, + 358, + 714 + ], + "spans": [ + { + "bbox": [ + 118, + 700, + 295, + 714 + ], + "score": 1.0, + "content": "†Correspondence to Tongliang Liu (tongliang.liu", + "type": "text" + }, + { + "bbox": [ + 295, + 703, + 303, + 711 + ], + "score": 0.28, + "content": "@", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 700, + 358, + 714 + ], + "score": 1.0, + "content": "sydney.edu.au)", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 122, + 97, + 490, + 137 + ], + "lines": [ + { + "bbox": [ + 120, + 95, + 492, + 119 + ], + "spans": [ + { + "bbox": [ + 120, + 95, + 492, + 119 + ], + "score": 1.0, + "content": "Understanding and Improving Early Stopping for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 118, + 410, + 138 + ], + "spans": [ + { + "bbox": [ + 201, + 118, + 410, + 138 + ], + "score": 1.0, + "content": "Learning with Noisy Labels", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 179, + 178, + 435, + 203 + ], + "lines": [ + { + "bbox": [ + 179, + 177, + 429, + 191 + ], + "spans": [ + { + "bbox": [ + 179, + 177, + 313, + 191 + ], + "score": 1.0, + "content": "Yingbin Bai1∗ Erkun Yang2∗", + "type": "text" + }, + { + "bbox": [ + 315, + 178, + 356, + 190 + ], + "score": 1.0, + "content": "Bo Han3", + "type": "text" + }, + { + "bbox": [ + 365, + 178, + 429, + 191 + ], + "score": 1.0, + "content": "Yanhua Yang2", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 190, + 437, + 203 + ], + "spans": [ + { + "bbox": [ + 179, + 190, + 233, + 203 + ], + "score": 1.0, + "content": "Jiatong Li4", + "type": "text" + }, + { + "bbox": [ + 240, + 190, + 302, + 203 + ], + "score": 1.0, + "content": "Yinian Mao4", + "type": "text" + }, + { + "bbox": [ + 307, + 190, + 357, + 203 + ], + "score": 1.0, + "content": "Gang Niu5", + "type": "text" + }, + { + "bbox": [ + 365, + 190, + 437, + 203 + ], + "score": 1.0, + "content": "Tongliang Liu1†", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 179, + 177, + 437, + 203 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 206, + 480, + 230 + ], + "lines": [ + { + "bbox": [ + 129, + 204, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 129, + 204, + 482, + 221 + ], + "score": 1.0, + "content": "1TML Lab, University of Sydney; 2Xidian University; 3Hong Kong Baptist University;", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 221, + 216, + 390, + 232 + ], + "spans": [ + { + "bbox": [ + 221, + 216, + 390, + 232 + ], + "score": 1.0, + "content": "4Meituan-Dianping Group; 5RIKEN AIP", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 129, + 204, + 482, + 232 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 258, + 328, + 271 + ], + "lines": [ + { + "bbox": [ + 282, + 258, + 330, + 272 + ], + "spans": [ + { + "bbox": [ + 282, + 258, + 330, + 272 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 283, + 469, + 501 + ], + "lines": [ + { + "bbox": [ + 141, + 282, + 469, + 296 + ], + "spans": [ + { + "bbox": [ + 141, + 282, + 469, + 296 + ], + "score": 1.0, + "content": "The memorization effect of deep neural network (DNN) plays a pivotal role in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 294, + 469, + 307 + ], + "spans": [ + { + "bbox": [ + 141, + 294, + 469, + 307 + ], + "score": 1.0, + "content": "many state-of-the-art label-noise learning methods. To exploit this property, the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 304, + 470, + 318 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 470, + 318 + ], + "score": 1.0, + "content": "early stopping trick, which stops the optimization at the early stage of training,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 316, + 470, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 316, + 470, + 329 + ], + "score": 1.0, + "content": "is usually adopted. Current methods generally decide the early stopping point", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 327, + 470, + 340 + ], + "spans": [ + { + "bbox": [ + 141, + 327, + 470, + 340 + ], + "score": 1.0, + "content": "by considering a DNN as a whole. However, a DNN can be considered as a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 338, + 470, + 351 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 470, + 351 + ], + "score": 1.0, + "content": "composition of a series of layers, and we find that the latter layers in a DNN are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 349, + 470, + 361 + ], + "spans": [ + { + "bbox": [ + 141, + 349, + 470, + 361 + ], + "score": 1.0, + "content": "much more sensitive to label noise, while their former counterparts are quite robust.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 359, + 470, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 359, + 470, + 372 + ], + "score": 1.0, + "content": "Therefore, selecting a stopping point for the whole network may make different", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 370, + 470, + 383 + ], + "spans": [ + { + "bbox": [ + 141, + 370, + 470, + 383 + ], + "score": 1.0, + "content": "DNN layers antagonistically affect each other, thus degrading the final performance.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 381, + 469, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 469, + 394 + ], + "score": 1.0, + "content": "In this paper, we propose to separate a DNN into different parts and progressively", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 392, + 470, + 406 + ], + "spans": [ + { + "bbox": [ + 141, + 392, + 470, + 406 + ], + "score": 1.0, + "content": "train them to address this problem. Instead of the early stopping which trains a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 402, + 469, + 416 + ], + "spans": [ + { + "bbox": [ + 141, + 402, + 469, + 416 + ], + "score": 1.0, + "content": "whole DNN all at once, we initially train former DNN layers by optimizing the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 414, + 469, + 427 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 469, + 427 + ], + "score": 1.0, + "content": "DNN with a relatively large number of epochs. During training, we progressively", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 425, + 469, + 438 + ], + "spans": [ + { + "bbox": [ + 141, + 425, + 469, + 438 + ], + "score": 1.0, + "content": "train the latter DNN layers by using a smaller number of epochs with the preceding", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 437, + 469, + 448 + ], + "spans": [ + { + "bbox": [ + 141, + 437, + 469, + 448 + ], + "score": 1.0, + "content": "layers fixed to counteract the impact of noisy labels. We term the proposed method", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 447, + 470, + 460 + ], + "spans": [ + { + "bbox": [ + 141, + 447, + 470, + 460 + ], + "score": 1.0, + "content": "as progressive early stopping (PES). Despite its simplicity, compared with the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 458, + 470, + 471 + ], + "spans": [ + { + "bbox": [ + 141, + 458, + 470, + 471 + ], + "score": 1.0, + "content": "traditional early stopping, PES can help to obtain more promising and stable results.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 469, + 470, + 482 + ], + "spans": [ + { + "bbox": [ + 141, + 469, + 470, + 482 + ], + "score": 1.0, + "content": "Furthermore, by combining PES with existing approaches on noisy label training,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 480, + 469, + 492 + ], + "spans": [ + { + "bbox": [ + 142, + 480, + 469, + 492 + ], + "score": 1.0, + "content": "we achieve state-of-the-art performance on image classification benchmarks. The", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 491, + 391, + 502 + ], + "spans": [ + { + "bbox": [ + 141, + 491, + 391, + 502 + ], + "score": 1.0, + "content": "code is made public at https://github.com/tmllab/PES.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 16.5, + "bbox_fs": [ + 141, + 282, + 470, + 502 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 522, + 190, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 192, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 192, + 538 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "Deep networks have revolutionized a wide variety of tasks, such as image processing, speech", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "recognition, and language modeling [7], However, this highly relies on the availability of large", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "annotated data, which may not be feasible in practice. Instead, many large datasets with lower", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 580, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 505, + 593 + ], + "score": 1.0, + "content": "quality annotations are collected from online queries [5] or social-network tagging [18]. Such", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "annotations inevitably contain mistakes or label noise. As deep networks have large model capacities,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "they can easily memorize and eventually overfit the noisy labels, leading to poor generalization", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "performance [36]. Therefore, it is of great importance to develop a methodology that is robust to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 624, + 182, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 182, + 636 + ], + "score": 1.0, + "content": "noisy annotations.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 547, + 506, + 636 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 640, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 507, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 507, + 653 + ], + "score": 1.0, + "content": "Existing methods on learning with noisy labels (LNL) can be mainly categorized into two groups:", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "model-based and model-free algorithms. Methods in the first category mainly model noisy labels with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "the noise transition matrix [24, 34, 33, 30]. With perfectly estimated noise transition matrix, models", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 673, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 506, + 684 + ], + "score": 1.0, + "content": "trained with corrected losses can approximate to the models trained with clean labels. However,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "current methods are usually fragile to estimate the noise transition matrix for heavy noisy data and", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "are also hard to handle a large number of classes [9]. The second type explores the dynamic process", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "of optimization policies, which relates to the memorization effect−deep neural networks tend to first", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 291, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 303 + ], + "score": 1.0, + "content": "memorize and fit majority (clean) patterns and then overfit minority (noisy) patterns [2]. Recently,", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 302, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 314 + ], + "score": 1.0, + "content": "based on this phenomenon, many methods [9, 26, 15, 16, 29] have been proposed and achieved", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 313, + 205, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 205, + 326 + ], + "score": 1.0, + "content": "promising performance.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 639, + 507, + 684 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 80, + 502, + 192 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 80, + 502, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 80, + 502, + 192 + ], + "spans": [ + { + "bbox": [ + 109, + 80, + 502, + 192 + ], + "score": 0.965, + "type": "image", + "image_path": "c7512ec26780dca9b15dc452cb391a4d5b9e7d1c3d1347ce8d597d9d4f66ce84.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 80, + 502, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 117.33333333333334, + 502, + 154.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 154.66666666666669, + 502, + 192.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 198, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "Figure 1: We train a ResNet-18 model on CIFAR-10 with three types of noisy labels and evaluate the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 209, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "impact of noisy labels on the representations from the 9-th layer, the 17-th layer, and the final layer.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "The X-axis is the number of epochs for the first block of the network. The curves present the mean of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 232, + 407, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 407, + 243 + ], + "score": 1.0, + "content": "five runs and the best performances are indicated with dotted vertical lines.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 270 + ], + "score": 1.0, + "content": "current methods are usually fragile to estimate the noise transition matrix for heavy noisy data and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "are also hard to handle a large number of classes [9]. The second type explores the dynamic process", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "of optimization policies, which relates to the memorization effect−deep neural networks tend to first", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 291, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 506, + 303 + ], + "score": 1.0, + "content": "memorize and fit majority (clean) patterns and then overfit minority (noisy) patterns [2]. Recently,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 302, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 314 + ], + "score": 1.0, + "content": "based on this phenomenon, many methods [9, 26, 15, 16, 29] have been proposed and achieved", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 313, + 205, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 205, + 326 + ], + "score": 1.0, + "content": "promising performance.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 505, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "score": 1.0, + "content": "To exploit the memorization effect, when the double descent phenomenon [3, 22, 11] cannot be", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 340, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 352 + ], + "score": 1.0, + "content": "guaranteed to occur, a core issue is to study when to stop the optimization of the network. While", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "stopping the training for too few epochs can avoid overfitting to noisy labels, it can also make the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 360, + 507, + 376 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 507, + 376 + ], + "score": 1.0, + "content": "network underfit to clean labels. Current methods [25, 23] usually adopt an early stopping strategy,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 372, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 385 + ], + "score": 1.0, + "content": "which decides the stopping point by considering the network as a whole. However, since DNNs are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 384, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 396 + ], + "score": 1.0, + "content": "usually optimized with stochastic gradient descent (SGD) with backpropagation, supervisory signals", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "will gradually propagate through the whole network from latter layers (i.e., layers that are closer to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "output layers) to former layers (i.e., layers that are closer to input layers). Noting that the output layer", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 415, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 505, + 430 + ], + "score": 1.0, + "content": "is followed by the empirical risk in the optimization procedure. We hypothesize that noisy labels may", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 427, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 440 + ], + "score": 1.0, + "content": "have more severe impacts for the latter layers, which is different from current methods [9, 15] that", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 438, + 326, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 326, + 450 + ], + "score": 1.0, + "content": "usually stop the training of the whole network at once.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 504, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 504, + 467 + ], + "score": 1.0, + "content": "To empirically verify the above hypothesis, we analyze the impact of noisy labels on representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "from different layers with different training epochs. To quantitatively measure the impact of noisy", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 474, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 490 + ], + "score": 1.0, + "content": "labels from intermediate layers, we first train the whole network on noisy data with different training", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "epochs and fix the parameters for the selected layer and its previous layers. We then reinitialize and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "optimize the rest layers with clean data, and the final classification performance is adopted to evaluate", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 509, + 507, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 507, + 522 + ], + "score": 1.0, + "content": "the impact of noisy labels. For the final layer, we directly report the overall classification performance.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "As illustrated in Figure 1, we can see that latter layers always achieve the best performance at relatively", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 531, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 506, + 543 + ], + "score": 1.0, + "content": "smaller epoch numbers and then exhibit stronger performance drops with additional training epochs,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "which verifies the hypothesis that noisy data may have more severe impacts for latter layers. With", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 552, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 567 + ], + "score": 1.0, + "content": "this understanding, we can infer that the early stopping, which optimizes the network all at once, may", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 563, + 432, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 432, + 576 + ], + "score": 1.0, + "content": "fail to fully exploit the memorization effect and induce sub-optimal performance.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 506, + 593 + ], + "score": 1.0, + "content": "To address the above problem, we propose to optimize a DNN by considering it as a composition of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 590, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 605 + ], + "score": 1.0, + "content": "several DNN parts and present a novel progressive early stopping (PES) method. Specifically, we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 600, + 507, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 507, + 616 + ], + "score": 1.0, + "content": "initially train former DNN layers by optimizing them with a relatively large number of epochs. Then,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "to alleviate the impact of noisy labels for latter layers, we reinitialize and progressively train latter", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "DNN layers by using smaller numbers of epochs with preceding DNN layers fixed. Since different", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 635, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 506, + 648 + ], + "score": 1.0, + "content": "layers are progressively trained with different early stopping epochs, we term the proposed method as", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 645, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 659 + ], + "score": 1.0, + "content": "progressive early stopping (PES). Despite its simplicity, compared with normal early stopping trick,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 654, + 507, + 669 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 507, + 669 + ], + "score": 1.0, + "content": "PES can help to better exploit the memorization effect and obtain more promising and stable results.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "Moreover, since the model size and training epochs are gradually reduced during the optimization", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 676, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 104, + 676, + 507, + 693 + ], + "score": 1.0, + "content": "procedure, the training time of PES is only slightly greater than that of the normal early stopping.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "Finally, by combining PES with existing approaches on noisy label training tasks, we establish new", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "state-of-the-art (SOTA) results on CIFAR-10 and CIFAR-100 with synthetic noise. We also achieve", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 711, + 411, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 411, + 724 + ], + "score": 1.0, + "content": "competitive results on one dataset with real-world noise: Clothing-1M [32].", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 41 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 80, + 502, + 192 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 80, + 502, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 80, + 502, + 192 + ], + "spans": [ + { + "bbox": [ + 109, + 80, + 502, + 192 + ], + "score": 0.965, + "type": "image", + "image_path": "c7512ec26780dca9b15dc452cb391a4d5b9e7d1c3d1347ce8d597d9d4f66ce84.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 80, + 502, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 117.33333333333334, + 502, + 154.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 154.66666666666669, + 502, + 192.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 198, + 505, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "Figure 1: We train a ResNet-18 model on CIFAR-10 with three types of noisy labels and evaluate the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 209, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "impact of noisy labels on the representations from the 9-th layer, the 17-th layer, and the final layer.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "The X-axis is the number of epochs for the first block of the network. The curves present the mean of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 232, + 407, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 407, + 243 + ], + "score": 1.0, + "content": "five runs and the best performances are indicated with dotted vertical lines.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 505, + 324 + ], + "lines": [], + "index": 9.5, + "bbox_fs": [ + 105, + 258, + 506, + 326 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 328, + 505, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "score": 1.0, + "content": "To exploit the memorization effect, when the double descent phenomenon [3, 22, 11] cannot be", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 340, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 352 + ], + "score": 1.0, + "content": "guaranteed to occur, a core issue is to study when to stop the optimization of the network. While", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "stopping the training for too few epochs can avoid overfitting to noisy labels, it can also make the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 360, + 507, + 376 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 507, + 376 + ], + "score": 1.0, + "content": "network underfit to clean labels. Current methods [25, 23] usually adopt an early stopping strategy,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 372, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 385 + ], + "score": 1.0, + "content": "which decides the stopping point by considering the network as a whole. However, since DNNs are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 384, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 396 + ], + "score": 1.0, + "content": "usually optimized with stochastic gradient descent (SGD) with backpropagation, supervisory signals", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "will gradually propagate through the whole network from latter layers (i.e., layers that are closer to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "output layers) to former layers (i.e., layers that are closer to input layers). Noting that the output layer", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 415, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 104, + 415, + 505, + 430 + ], + "score": 1.0, + "content": "is followed by the empirical risk in the optimization procedure. We hypothesize that noisy labels may", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 427, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 440 + ], + "score": 1.0, + "content": "have more severe impacts for the latter layers, which is different from current methods [9, 15] that", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 438, + 326, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 326, + 450 + ], + "score": 1.0, + "content": "usually stop the training of the whole network at once.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 328, + 507, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 454, + 505, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 504, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 504, + 467 + ], + "score": 1.0, + "content": "To empirically verify the above hypothesis, we analyze the impact of noisy labels on representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "from different layers with different training epochs. To quantitatively measure the impact of noisy", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 474, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 490 + ], + "score": 1.0, + "content": "labels from intermediate layers, we first train the whole network on noisy data with different training", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "epochs and fix the parameters for the selected layer and its previous layers. We then reinitialize and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "optimize the rest layers with clean data, and the final classification performance is adopted to evaluate", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 509, + 507, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 507, + 522 + ], + "score": 1.0, + "content": "the impact of noisy labels. For the final layer, we directly report the overall classification performance.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "As illustrated in Figure 1, we can see that latter layers always achieve the best performance at relatively", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 531, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 506, + 543 + ], + "score": 1.0, + "content": "smaller epoch numbers and then exhibit stronger performance drops with additional training epochs,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "which verifies the hypothesis that noisy data may have more severe impacts for latter layers. With", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 552, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 567 + ], + "score": 1.0, + "content": "this understanding, we can infer that the early stopping, which optimizes the network all at once, may", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 563, + 432, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 432, + 576 + ], + "score": 1.0, + "content": "fail to fully exploit the memorization effect and induce sub-optimal performance.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 455, + 507, + 576 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 506, + 593 + ], + "score": 1.0, + "content": "To address the above problem, we propose to optimize a DNN by considering it as a composition of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 590, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 605 + ], + "score": 1.0, + "content": "several DNN parts and present a novel progressive early stopping (PES) method. Specifically, we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 600, + 507, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 507, + 616 + ], + "score": 1.0, + "content": "initially train former DNN layers by optimizing them with a relatively large number of epochs. Then,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "to alleviate the impact of noisy labels for latter layers, we reinitialize and progressively train latter", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "DNN layers by using smaller numbers of epochs with preceding DNN layers fixed. Since different", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 635, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 506, + 648 + ], + "score": 1.0, + "content": "layers are progressively trained with different early stopping epochs, we term the proposed method as", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 645, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 659 + ], + "score": 1.0, + "content": "progressive early stopping (PES). Despite its simplicity, compared with normal early stopping trick,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 654, + 507, + 669 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 507, + 669 + ], + "score": 1.0, + "content": "PES can help to better exploit the memorization effect and obtain more promising and stable results.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "Moreover, since the model size and training epochs are gradually reduced during the optimization", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 676, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 104, + 676, + 507, + 693 + ], + "score": 1.0, + "content": "procedure, the training time of PES is only slightly greater than that of the normal early stopping.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "Finally, by combining PES with existing approaches on noisy label training tasks, we establish new", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "state-of-the-art (SOTA) results on CIFAR-10 and CIFAR-100 with synthetic noise. We also achieve", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 711, + 411, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 411, + 724 + ], + "score": 1.0, + "content": "competitive results on one dataset with real-world noise: Clothing-1M [32].", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 41, + "bbox_fs": [ + 104, + 579, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "The rest of the paper is organized as follows. In Section 2, we first introduce the proposed progressive", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 97 + ], + "score": 1.0, + "content": "early stopping and then present the details of the proposed algorithm by combining our method with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "existing approaches on noisy label training tasks. Section 3 shows the experimental results of our", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "proposed method. Related works are briefly reviewed in Section 4. Finally, concluding remarks are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 117, + 183, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 183, + 127 + ], + "score": 1.0, + "content": "given in Section 5.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 107, + 143, + 216, + 157 + ], + "lines": [ + { + "bbox": [ + 104, + 142, + 217, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 217, + 160 + ], + "score": 1.0, + "content": "2 Proposed Method", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 168, + 506, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 122, + 181 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 170, + 132, + 179 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 169, + 322, + 181 + ], + "score": 1.0, + "content": "be the distribution of a pair of random variables", + "type": "text" + }, + { + "bbox": [ + 323, + 169, + 425, + 181 + ], + "score": 0.92, + "content": "( X , Y ) \\in \\mathcal { X } \\times \\{ 1 , . . . K \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 169, + 455, + 181 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 455, + 170, + 466, + 179 + ], + "score": 0.81, + "content": "\\boldsymbol { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 169, + 505, + 181 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 180, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 205, + 191 + ], + "score": 1.0, + "content": "the variable of instances,", + "type": "text" + }, + { + "bbox": [ + 205, + 180, + 215, + 190 + ], + "score": 0.75, + "content": "\\mathbf { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 180, + 310, + 191 + ], + "score": 1.0, + "content": "is the variable of labels,", + "type": "text" + }, + { + "bbox": [ + 310, + 181, + 320, + 190 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 180, + 438, + 191 + ], + "score": 1.0, + "content": "denotes the feature space, and", + "type": "text" + }, + { + "bbox": [ + 439, + 181, + 449, + 190 + ], + "score": 0.84, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 180, + 505, + 191 + ], + "score": 1.0, + "content": "is the number", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 191, + 504, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 494, + 203 + ], + "score": 1.0, + "content": "of classes. In many real-world problems, examples independently drawn from the distribution", + "type": "text" + }, + { + "bbox": [ + 494, + 191, + 504, + 201 + ], + "score": 0.81, + "content": "D", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "are unavailable. Before being observed, the clean labels are usually randomly corrupted into noisy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 152, + 227 + ], + "score": 1.0, + "content": "labels. Let", + "type": "text" + }, + { + "bbox": [ + 153, + 213, + 162, + 224 + ], + "score": 0.85, + "content": "\\tilde { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 213, + 329, + 227 + ], + "score": 1.0, + "content": "be the distribution of the noisy example", + "type": "text" + }, + { + "bbox": [ + 329, + 213, + 361, + 226 + ], + "score": 0.93, + "content": "( X , { \\tilde { Y } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 213, + 393, + 227 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 393, + 213, + 403, + 224 + ], + "score": 0.86, + "content": "\\tilde { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "indicates the variable of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 223, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 223, + 393, + 240 + ], + "score": 1.0, + "content": "noisy labels. For label-noise learning, we can only access a sample set", + "type": "text" + }, + { + "bbox": [ + 394, + 225, + 444, + 238 + ], + "score": 0.93, + "content": "\\{ \\pmb { x } _ { i } , \\tilde { y } _ { i } ) \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 223, + 506, + 240 + ], + "score": 1.0, + "content": "independently", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 155, + 251 + ], + "score": 1.0, + "content": "drawn from", + "type": "text" + }, + { + "bbox": [ + 156, + 237, + 165, + 248 + ], + "score": 0.85, + "content": "\\tilde { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 237, + 506, + 251 + ], + "score": 1.0, + "content": ". The aim is to learn a robust classifier from the noisy sample set that can classify test", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 191, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 191, + 262 + ], + "score": 1.0, + "content": "instances accurately.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 506, + 278 + ], + "score": 1.0, + "content": "In the following, we first elaborate on the proposed progressive early stopping (PES). Then, based", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "on PES, we provide a learning algorithm that learns with confident examples and semi-supervised", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 286, + 189, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 189, + 300 + ], + "score": 1.0, + "content": "learning techniques.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 312, + 246, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 248, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 248, + 328 + ], + "score": 1.0, + "content": "2.1 Progressive Early Stopping", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 332, + 506, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 333, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 505, + 344 + ], + "score": 1.0, + "content": "When trained with noisy labels, if clean labels are of majority within each noisy class, deep networks", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 343, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 357 + ], + "score": 1.0, + "content": "tend to first fit clean labels during an early learning stage before eventually memorizing the wrong", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 352, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 104, + 352, + 506, + 369 + ], + "score": 1.0, + "content": "labels, which can be explained by the memorization effect. Many current methods utilize this property", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 366, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 506, + 378 + ], + "score": 1.0, + "content": "to counteract the influence of noisy labels by stopping the optimization at an early learning phase.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "Specifically, a deep classifier can be obtained by optimizing the following objective function with a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 387, + 241, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 229, + 399 + ], + "score": 1.0, + "content": "relatively small epoch number", + "type": "text" + }, + { + "bbox": [ + 230, + 388, + 238, + 397 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 387, + 241, + 399 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 248, + 405, + 361, + 438 + ], + "lines": [ + { + "bbox": [ + 248, + 405, + 361, + 438 + ], + "spans": [ + { + "bbox": [ + 248, + 405, + 361, + 438 + ], + "score": 0.95, + "content": "\\operatorname* { m i n } _ { \\Theta } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( x _ { i } ; \\Theta ) , \\tilde { y } _ { i } ) ,", + "type": "interline_equation", + "image_path": "c0c8c02abfce6194ad608c4820ea855fb6789ee70cf2039f7d84b2ec9d9be1ed.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 248, + 405, + 361, + 421.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 248, + 421.5, + 361, + 438.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 443, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 133, + 456 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 443, + 163, + 456 + ], + "score": 0.92, + "content": "f ( \\cdot ; \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 443, + 337, + 456 + ], + "score": 1.0, + "content": "is a deep classifier with model parameters", + "type": "text" + }, + { + "bbox": [ + 338, + 444, + 347, + 453 + ], + "score": 0.83, + "content": "\\Theta", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 443, + 365, + 456 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 365, + 444, + 373, + 453 + ], + "score": 0.83, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "is the cross-entropy loss. When", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 454, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 504, + 466 + ], + "score": 1.0, + "content": "trained with noisy data, early learning regularization (ELR) [16] reveals that, for the most commonly", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "score": 1.0, + "content": "used cross-entropy loss, the gradient is well correlated with the correct direction at the early learning", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 369, + 489 + ], + "score": 1.0, + "content": "phase. Therefore, with a properly defined small epoch number", + "type": "text" + }, + { + "bbox": [ + 370, + 477, + 378, + 486 + ], + "score": 0.78, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 477, + 506, + 489 + ], + "score": 1.0, + "content": ", the classifier can have higher", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 448, + 500 + ], + "score": 1.0, + "content": "accuracy than at initialization. While, if we continue to optimize the deep model after", + "type": "text" + }, + { + "bbox": [ + 448, + 488, + 457, + 497 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "epochs, the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 497, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 511 + ], + "score": 1.0, + "content": "classifier will be able to memorize more noise labels. Therefore, it is critical to select a proper epoch", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 509, + 453, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 139, + 520 + ], + "score": 1.0, + "content": "number", + "type": "text" + }, + { + "bbox": [ + 140, + 509, + 148, + 519 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 509, + 453, + 520 + ], + "score": 1.0, + "content": "to utilize the memorization effect and alleviate the influence of noisy labels.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 304, + 538 + ], + "score": 1.0, + "content": "Current methods typically select the epoch number", + "type": "text" + }, + { + "bbox": [ + 305, + 526, + 313, + 535 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "by considering the network as a whole. However,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 536, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 548 + ], + "score": 1.0, + "content": "as Figure 1 makes clear, the impact of noisy labels on different DNN layers are different, which", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "implies that the traditional early stopping trick, which optimizes the whole network all at once, may", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "make different DNN layers to be antagonistically affected by each other, thus degrading the final", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "model performance. To this end, we propose to separate a DNN into different parts and progressively", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "train layers in different parts with different training epochs. Specifically, assume that the whole", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 591, + 322, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 141, + 603 + ], + "score": 1.0, + "content": "network", + "type": "text" + }, + { + "bbox": [ + 142, + 591, + 171, + 603 + ], + "score": 0.93, + "content": "f ( \\cdot ; \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 591, + 267, + 603 + ], + "score": 1.0, + "content": "can be constituted with", + "type": "text" + }, + { + "bbox": [ + 268, + 591, + 276, + 601 + ], + "score": 0.81, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 591, + 322, + 603 + ], + "score": 1.0, + "content": "DNN parts", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 607, + 375, + 638 + ], + "lines": [ + { + "bbox": [ + 235, + 607, + 375, + 638 + ], + "spans": [ + { + "bbox": [ + 235, + 607, + 375, + 638 + ], + "score": 0.92, + "content": "\\begin{array} { l } { z _ { 1 } = f _ { 1 } ( \\pmb { x } ; \\Theta _ { 1 } ) , } \\\\ { z _ { l } = f _ { l } ( z _ { l - 1 } ; \\Theta _ { l } ) , \\quad l = 2 , \\ldots , L } \\end{array}", + "type": "interline_equation", + "image_path": "c49b4c2da11083b29e1461cfbacf31110bb24c4b4430f414be9d136f95e70e52.jpg" + } + ] + } + ], + "index": 40.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 607, + 375, + 622.5 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 235, + 622.5, + 375, + 638.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 642, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 133, + 655 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 643, + 168, + 655 + ], + "score": 0.93, + "content": "f _ { l } ( \\cdot ; \\Theta _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 642, + 193, + 655 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 194, + 644, + 198, + 653 + ], + "score": 0.74, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 642, + 270, + 655 + ], + "score": 1.0, + "content": "-th DNN part and", + "type": "text" + }, + { + "bbox": [ + 270, + 645, + 280, + 654 + ], + "score": 0.85, + "content": "z _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 642, + 506, + 655 + ], + "score": 1.0, + "content": "is the corresponding output. The output of the last part", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 107, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 107, + 656, + 119, + 665 + ], + "score": 0.84, + "content": "z _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 653, + 245, + 667 + ], + "score": 1.0, + "content": "is the prediction. The network", + "type": "text" + }, + { + "bbox": [ + 245, + 654, + 274, + 666 + ], + "score": 0.92, + "content": "f ( \\cdot ; \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 653, + 383, + 667 + ], + "score": 1.0, + "content": "can also be represented as", + "type": "text" + }, + { + "bbox": [ + 383, + 654, + 443, + 666 + ], + "score": 0.91, + "content": "f ( \\cdot ; \\Theta _ { 1 } , . . . \\Theta _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 653, + 506, + 667 + ], + "score": 1.0, + "content": ". To counteract", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 362, + 677 + ], + "score": 1.0, + "content": "the impact of noisy labels, We initially optimize the parameter", + "type": "text" + }, + { + "bbox": [ + 363, + 665, + 375, + 676 + ], + "score": 0.88, + "content": "\\Theta _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "for the first part by training the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 676, + 339, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 182, + 687 + ], + "score": 1.0, + "content": "whole network for", + "type": "text" + }, + { + "bbox": [ + 182, + 676, + 194, + 687 + ], + "score": 0.88, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 676, + 339, + 687 + ], + "score": 1.0, + "content": "epochs with the following objective", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 693, + 388, + 726 + ], + "lines": [ + { + "bbox": [ + 223, + 693, + 388, + 726 + ], + "spans": [ + { + "bbox": [ + 223, + 693, + 388, + 726 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\Theta _ { 1 } \\ldots \\Theta _ { k } } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( \\pmb { x } _ { i } ; \\Theta _ { 1 } , \\ldots , \\Theta _ { L } ) , \\tilde { y } _ { i } ) .", + "type": "interline_equation", + "image_path": "8f9f562c089aae98a23044dad475a993b51a64c95507fed67028e7a03c4884fc.jpg" + } + ] + } + ], + "index": 46.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 693, + 388, + 709.5 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 223, + 709.5, + 388, + 726.0 + ], + "spans": [], + "index": 47 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "The rest of the paper is organized as follows. In Section 2, we first introduce the proposed progressive", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 97 + ], + "score": 1.0, + "content": "early stopping and then present the details of the proposed algorithm by combining our method with", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "existing approaches on noisy label training tasks. Section 3 shows the experimental results of our", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "proposed method. Related works are briefly reviewed in Section 4. Finally, concluding remarks are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 117, + 183, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 183, + 127 + ], + "score": 1.0, + "content": "given in Section 5.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 73, + 505, + 127 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 143, + 216, + 157 + ], + "lines": [ + { + "bbox": [ + 104, + 142, + 217, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 217, + 160 + ], + "score": 1.0, + "content": "2 Proposed Method", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 168, + 506, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 169, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 122, + 181 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 170, + 132, + 179 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 169, + 322, + 181 + ], + "score": 1.0, + "content": "be the distribution of a pair of random variables", + "type": "text" + }, + { + "bbox": [ + 323, + 169, + 425, + 181 + ], + "score": 0.92, + "content": "( X , Y ) \\in \\mathcal { X } \\times \\{ 1 , . . . K \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 169, + 455, + 181 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 455, + 170, + 466, + 179 + ], + "score": 0.81, + "content": "\\boldsymbol { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 169, + 505, + 181 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 180, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 205, + 191 + ], + "score": 1.0, + "content": "the variable of instances,", + "type": "text" + }, + { + "bbox": [ + 205, + 180, + 215, + 190 + ], + "score": 0.75, + "content": "\\mathbf { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 180, + 310, + 191 + ], + "score": 1.0, + "content": "is the variable of labels,", + "type": "text" + }, + { + "bbox": [ + 310, + 181, + 320, + 190 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 180, + 438, + 191 + ], + "score": 1.0, + "content": "denotes the feature space, and", + "type": "text" + }, + { + "bbox": [ + 439, + 181, + 449, + 190 + ], + "score": 0.84, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 180, + 505, + 191 + ], + "score": 1.0, + "content": "is the number", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 191, + 504, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 494, + 203 + ], + "score": 1.0, + "content": "of classes. In many real-world problems, examples independently drawn from the distribution", + "type": "text" + }, + { + "bbox": [ + 494, + 191, + 504, + 201 + ], + "score": 0.81, + "content": "D", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "are unavailable. Before being observed, the clean labels are usually randomly corrupted into noisy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 152, + 227 + ], + "score": 1.0, + "content": "labels. Let", + "type": "text" + }, + { + "bbox": [ + 153, + 213, + 162, + 224 + ], + "score": 0.85, + "content": "\\tilde { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 213, + 329, + 227 + ], + "score": 1.0, + "content": "be the distribution of the noisy example", + "type": "text" + }, + { + "bbox": [ + 329, + 213, + 361, + 226 + ], + "score": 0.93, + "content": "( X , { \\tilde { Y } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 213, + 393, + 227 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 393, + 213, + 403, + 224 + ], + "score": 0.86, + "content": "\\tilde { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "indicates the variable of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 223, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 223, + 393, + 240 + ], + "score": 1.0, + "content": "noisy labels. For label-noise learning, we can only access a sample set", + "type": "text" + }, + { + "bbox": [ + 394, + 225, + 444, + 238 + ], + "score": 0.93, + "content": "\\{ \\pmb { x } _ { i } , \\tilde { y } _ { i } ) \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 223, + 506, + 240 + ], + "score": 1.0, + "content": "independently", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 155, + 251 + ], + "score": 1.0, + "content": "drawn from", + "type": "text" + }, + { + "bbox": [ + 156, + 237, + 165, + 248 + ], + "score": 0.85, + "content": "\\tilde { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 237, + 506, + 251 + ], + "score": 1.0, + "content": ". The aim is to learn a robust classifier from the noisy sample set that can classify test", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 191, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 191, + 262 + ], + "score": 1.0, + "content": "instances accurately.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 169, + 506, + 262 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 265, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 506, + 278 + ], + "score": 1.0, + "content": "In the following, we first elaborate on the proposed progressive early stopping (PES). Then, based", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "on PES, we provide a learning algorithm that learns with confident examples and semi-supervised", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 286, + 189, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 189, + 300 + ], + "score": 1.0, + "content": "learning techniques.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 265, + 506, + 300 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 312, + 246, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 248, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 248, + 328 + ], + "score": 1.0, + "content": "2.1 Progressive Early Stopping", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 332, + 506, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 333, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 505, + 344 + ], + "score": 1.0, + "content": "When trained with noisy labels, if clean labels are of majority within each noisy class, deep networks", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 343, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 357 + ], + "score": 1.0, + "content": "tend to first fit clean labels during an early learning stage before eventually memorizing the wrong", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 352, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 104, + 352, + 506, + 369 + ], + "score": 1.0, + "content": "labels, which can be explained by the memorization effect. Many current methods utilize this property", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 366, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 506, + 378 + ], + "score": 1.0, + "content": "to counteract the influence of noisy labels by stopping the optimization at an early learning phase.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "Specifically, a deep classifier can be obtained by optimizing the following objective function with a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 387, + 241, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 229, + 399 + ], + "score": 1.0, + "content": "relatively small epoch number", + "type": "text" + }, + { + "bbox": [ + 230, + 388, + 238, + 397 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 387, + 241, + 399 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 333, + 506, + 399 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 248, + 405, + 361, + 438 + ], + "lines": [ + { + "bbox": [ + 248, + 405, + 361, + 438 + ], + "spans": [ + { + "bbox": [ + 248, + 405, + 361, + 438 + ], + "score": 0.95, + "content": "\\operatorname* { m i n } _ { \\Theta } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( x _ { i } ; \\Theta ) , \\tilde { y } _ { i } ) ,", + "type": "interline_equation", + "image_path": "c0c8c02abfce6194ad608c4820ea855fb6789ee70cf2039f7d84b2ec9d9be1ed.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 248, + 405, + 361, + 421.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 248, + 421.5, + 361, + 438.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 443, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 133, + 456 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 443, + 163, + 456 + ], + "score": 0.92, + "content": "f ( \\cdot ; \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 443, + 337, + 456 + ], + "score": 1.0, + "content": "is a deep classifier with model parameters", + "type": "text" + }, + { + "bbox": [ + 338, + 444, + 347, + 453 + ], + "score": 0.83, + "content": "\\Theta", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 443, + 365, + 456 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 365, + 444, + 373, + 453 + ], + "score": 0.83, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "is the cross-entropy loss. When", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 454, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 504, + 466 + ], + "score": 1.0, + "content": "trained with noisy data, early learning regularization (ELR) [16] reveals that, for the most commonly", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "score": 1.0, + "content": "used cross-entropy loss, the gradient is well correlated with the correct direction at the early learning", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 369, + 489 + ], + "score": 1.0, + "content": "phase. Therefore, with a properly defined small epoch number", + "type": "text" + }, + { + "bbox": [ + 370, + 477, + 378, + 486 + ], + "score": 0.78, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 477, + 506, + 489 + ], + "score": 1.0, + "content": ", the classifier can have higher", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 448, + 500 + ], + "score": 1.0, + "content": "accuracy than at initialization. While, if we continue to optimize the deep model after", + "type": "text" + }, + { + "bbox": [ + 448, + 488, + 457, + 497 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "epochs, the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 497, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 511 + ], + "score": 1.0, + "content": "classifier will be able to memorize more noise labels. Therefore, it is critical to select a proper epoch", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 509, + 453, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 139, + 520 + ], + "score": 1.0, + "content": "number", + "type": "text" + }, + { + "bbox": [ + 140, + 509, + 148, + 519 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 509, + 453, + 520 + ], + "score": 1.0, + "content": "to utilize the memorization effect and alleviate the influence of noisy labels.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 443, + 506, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 304, + 538 + ], + "score": 1.0, + "content": "Current methods typically select the epoch number", + "type": "text" + }, + { + "bbox": [ + 305, + 526, + 313, + 535 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "by considering the network as a whole. However,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 536, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 548 + ], + "score": 1.0, + "content": "as Figure 1 makes clear, the impact of noisy labels on different DNN layers are different, which", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "implies that the traditional early stopping trick, which optimizes the whole network all at once, may", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "make different DNN layers to be antagonistically affected by each other, thus degrading the final", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "model performance. To this end, we propose to separate a DNN into different parts and progressively", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "train layers in different parts with different training epochs. Specifically, assume that the whole", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 591, + 322, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 141, + 603 + ], + "score": 1.0, + "content": "network", + "type": "text" + }, + { + "bbox": [ + 142, + 591, + 171, + 603 + ], + "score": 0.93, + "content": "f ( \\cdot ; \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 591, + 267, + 603 + ], + "score": 1.0, + "content": "can be constituted with", + "type": "text" + }, + { + "bbox": [ + 268, + 591, + 276, + 601 + ], + "score": 0.81, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 591, + 322, + 603 + ], + "score": 1.0, + "content": "DNN parts", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 525, + 506, + 603 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 607, + 375, + 638 + ], + "lines": [ + { + "bbox": [ + 235, + 607, + 375, + 638 + ], + "spans": [ + { + "bbox": [ + 235, + 607, + 375, + 638 + ], + "score": 0.92, + "content": "\\begin{array} { l } { z _ { 1 } = f _ { 1 } ( \\pmb { x } ; \\Theta _ { 1 } ) , } \\\\ { z _ { l } = f _ { l } ( z _ { l - 1 } ; \\Theta _ { l } ) , \\quad l = 2 , \\ldots , L } \\end{array}", + "type": "interline_equation", + "image_path": "c49b4c2da11083b29e1461cfbacf31110bb24c4b4430f414be9d136f95e70e52.jpg" + } + ] + } + ], + "index": 40.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 607, + 375, + 622.5 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 235, + 622.5, + 375, + 638.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 642, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 133, + 655 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 643, + 168, + 655 + ], + "score": 0.93, + "content": "f _ { l } ( \\cdot ; \\Theta _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 642, + 193, + 655 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 194, + 644, + 198, + 653 + ], + "score": 0.74, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 642, + 270, + 655 + ], + "score": 1.0, + "content": "-th DNN part and", + "type": "text" + }, + { + "bbox": [ + 270, + 645, + 280, + 654 + ], + "score": 0.85, + "content": "z _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 642, + 506, + 655 + ], + "score": 1.0, + "content": "is the corresponding output. The output of the last part", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 107, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 107, + 656, + 119, + 665 + ], + "score": 0.84, + "content": "z _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 653, + 245, + 667 + ], + "score": 1.0, + "content": "is the prediction. The network", + "type": "text" + }, + { + "bbox": [ + 245, + 654, + 274, + 666 + ], + "score": 0.92, + "content": "f ( \\cdot ; \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 653, + 383, + 667 + ], + "score": 1.0, + "content": "can also be represented as", + "type": "text" + }, + { + "bbox": [ + 383, + 654, + 443, + 666 + ], + "score": 0.91, + "content": "f ( \\cdot ; \\Theta _ { 1 } , . . . \\Theta _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 653, + 506, + 667 + ], + "score": 1.0, + "content": ". To counteract", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 362, + 677 + ], + "score": 1.0, + "content": "the impact of noisy labels, We initially optimize the parameter", + "type": "text" + }, + { + "bbox": [ + 363, + 665, + 375, + 676 + ], + "score": 0.88, + "content": "\\Theta _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "for the first part by training the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 676, + 339, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 182, + 687 + ], + "score": 1.0, + "content": "whole network for", + "type": "text" + }, + { + "bbox": [ + 182, + 676, + 194, + 687 + ], + "score": 0.88, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 676, + 339, + 687 + ], + "score": 1.0, + "content": "epochs with the following objective", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 642, + 506, + 687 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 693, + 388, + 726 + ], + "lines": [ + { + "bbox": [ + 223, + 693, + 388, + 726 + ], + "spans": [ + { + "bbox": [ + 223, + 693, + 388, + 726 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { \\Theta _ { 1 } \\ldots \\Theta _ { k } } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( \\pmb { x } _ { i } ; \\Theta _ { 1 } , \\ldots , \\Theta _ { L } ) , \\tilde { y } _ { i } ) .", + "type": "interline_equation", + "image_path": "8f9f562c089aae98a23044dad475a993b51a64c95507fed67028e7a03c4884fc.jpg" + } + ] + } + ], + "index": 46.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 693, + 388, + 709.5 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 223, + 709.5, + 388, + 726.0 + ], + "spans": [], + "index": 47 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 81, + 501, + 190 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 81, + 501, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 81, + 501, + 190 + ], + "spans": [ + { + "bbox": [ + 109, + 81, + 501, + 190 + ], + "score": 0.965, + "type": "image", + "image_path": "93c2dd94c2a0f5dd8e852ee49fbcbcd3facdbcf530179d0abf899ddba8541bf1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 81, + 501, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 117.33333333333334, + 501, + 153.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 153.66666666666669, + 501, + 190.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 196, + 506, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "Figure 2: Performance of the traditional early stopping trick and the proposed PES on CIFAR-10", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 207, + 405, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 405, + 220 + ], + "score": 1.0, + "content": "with different types of label noise. The lines present the mean of five runs.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 266, + 242 + ], + "score": 1.0, + "content": "Then, we keep the obtained parameter", + "type": "text" + }, + { + "bbox": [ + 266, + 230, + 280, + 241 + ], + "score": 0.89, + "content": "\\Theta _ { 1 } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 228, + 467, + 242 + ], + "score": 1.0, + "content": "fixed, reinitialize and progressively learn the", + "type": "text" + }, + { + "bbox": [ + 467, + 230, + 471, + 239 + ], + "score": 0.68, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 228, + 487, + 242 + ], + "score": 1.0, + "content": "-th", + "type": "text" + }, + { + "bbox": [ + 487, + 230, + 505, + 240 + ], + "score": 0.79, + "content": "( l =", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 145, + 251 + ], + "score": 0.88, + "content": "2 , \\ldots , L )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "DNN part with the parameters for preceding DNN parts fixed. The training procedure is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 251, + 364, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 170, + 263 + ], + "score": 1.0, + "content": "conducted with", + "type": "text" + }, + { + "bbox": [ + 170, + 251, + 180, + 262 + ], + "score": 0.85, + "content": "T _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 251, + 364, + 263 + ], + "score": 1.0, + "content": "epochs by optimizing the following objective", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 266, + 444, + 298 + ], + "lines": [ + { + "bbox": [ + 167, + 266, + 444, + 298 + ], + "spans": [ + { + "bbox": [ + 167, + 266, + 444, + 298 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { \\Theta _ { l } \\ldots \\Theta _ { k } } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( x _ { i } ; \\Theta _ { 1 } ^ { * } , \\ldots , \\Theta _ { l - 1 } ^ { * } , \\Theta _ { l } , \\ldots , \\Theta _ { L } ) , \\tilde { y } _ { i } ) , \\quad l = 2 \\ldots L", + "type": "interline_equation", + "image_path": "6cf91797c88f1e24c4a6f13692f97ac06e4c0a1ac0851bf9e6baf25712dea2a2.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 167, + 266, + 444, + 276.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 167, + 276.6666666666667, + 444, + 287.33333333333337 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 167, + 287.33333333333337, + 444, + 298.00000000000006 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 211, + 315 + ], + "score": 1.0, + "content": "We gradually optimize the", + "type": "text" + }, + { + "bbox": [ + 211, + 302, + 237, + 314 + ], + "score": 0.91, + "content": "( l + 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 302, + 400, + 315 + ], + "score": 1.0, + "content": "-th DNN part with the obtained parameter", + "type": "text" + }, + { + "bbox": [ + 401, + 303, + 414, + 315 + ], + "score": 0.89, + "content": "\\Theta _ { l } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "fixed, the optimization", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "is continued until all the parameters have been optimized. As elaborated above, latter DNN parts are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 322, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 104, + 322, + 506, + 338 + ], + "score": 1.0, + "content": "more sensitive to noisy labels than their former counterparts. Therefore, for the above initializing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 270, + 348 + ], + "score": 1.0, + "content": "optimization in Eq. (3) and the following", + "type": "text" + }, + { + "bbox": [ + 270, + 335, + 295, + 346 + ], + "score": 0.86, + "content": "L - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "steps of optimization in Eq. (4), we gradually reduce", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 209, + 359 + ], + "score": 1.0, + "content": "the training epochs (i.e.", + "type": "text" + }, + { + "bbox": [ + 209, + 346, + 304, + 357 + ], + "score": 0.89, + "content": "T _ { 1 } \\geq T _ { 2 } \\geq \\cdot \\cdot \\cdot \\geq T _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 346, + 506, + 359 + ], + "score": 1.0, + "content": "to better exploit the memorization effect. After", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 306, + 370 + ], + "score": 1.0, + "content": "optimization, we can obtain the final network as", + "type": "text" + }, + { + "bbox": [ + 307, + 357, + 419, + 369 + ], + "score": 0.91, + "content": "f ( \\cdot , \\Theta ) = \\bar { f } ( \\cdot ; \\Theta _ { 1 } ^ { * } , \\ldots , \\Theta _ { L } ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 357, + 506, + 370 + ], + "score": 1.0, + "content": ". Since this model is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "obtained by progressively exploiting the early stopping strategies for different DNN parts, we term", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 378, + 339, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 339, + 392 + ], + "score": 1.0, + "content": "the proposed method as progressive early stopping (PES).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "To explicitly verify the effectiveness of the proposed PES method, we conduct several pilot experi-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "ments, which compare the traditional early stopping and PES with label noise from different types", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 416, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 506, + 430 + ], + "score": 1.0, + "content": "and different levels. The results are illustrated in Figure 2, from which we can see that, compared", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 440 + ], + "score": 1.0, + "content": "with models trained with traditional early stopping, models trained with PES can achieve superior", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "classification accuracy with smaller variations in all the cases. Current state-of-the-art methods [15]", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "usually adopt models with the traditional early stopping as base models to distill confident examples", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "and then utilize semi-supervised learning techniques by considering confident examples as labeled", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "data and other noisy examples as unlabeled data to further improve the results. The final performance", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "still heavily relies on the base model trained with noisy labels. By improving the performance of the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "base model, our method combined with semi-supervised learning techniques is able to establish new", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "state-of-the-art results. In the following subsections, we will elaborate on how to utilize PES to distill", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 515, + 447, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 447, + 528 + ], + "score": 1.0, + "content": "confident examples and further combine it with semi-supervised learning techniques.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 539, + 280, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 281, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 281, + 553 + ], + "score": 1.0, + "content": "2.2 Learning with Confident Examples", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 573 + ], + "score": 1.0, + "content": "Based on the deep network optimized with progressive early stopping, we can select confident", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 570, + 504, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 504, + 582 + ], + "score": 1.0, + "content": "examples to facilitate the model training. Here, confident examples refer to examples that have high", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "probabilities with clean labels. In this paper, we treat examples whose predictions are consistent with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 591, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 506, + 605 + ], + "score": 1.0, + "content": "given labels as confident examples. In addition, to make the results more robust, we generate two", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 601, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 616 + ], + "score": 1.0, + "content": "different augmentations for any given input and use the average prediction to decide its predicted", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 614, + 358, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 334, + 626 + ], + "score": 1.0, + "content": "label. Formally, we can obtain the confident example set", + "type": "text" + }, + { + "bbox": [ + 334, + 614, + 345, + 625 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 614, + 358, + 626 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 636, + 438, + 681 + ], + "lines": [ + { + "bbox": [ + 172, + 636, + 438, + 681 + ], + "spans": [ + { + "bbox": [ + 172, + 636, + 438, + 681 + ], + "score": 0.91, + "content": "\\begin{array} { c } { \\mathcal { D } _ { l } = \\{ ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = \\hat { y } _ { i } , i = 1 , \\ldots , n \\} , } \\\\ { \\hat { y } _ { i } = \\underset { k \\in \\{ 1 , \\ldots , K \\} } { \\arg \\operatorname* { m a x } } \\ \\frac { 1 } { 2 } [ f ^ { k } ( \\mathrm { A u g m e n t } ( \\pmb { x } _ { i } ) ; \\Theta ) + f ^ { k } ( \\mathrm { A u g m e n t } ( \\pmb { x } _ { i } ) ; \\Theta ) ] , } \\end{array}", + "type": "interline_equation", + "image_path": "f0ca8b2be795a56146eeba463a76de4df49127ec5a09e9825eff949f223573dc.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 172, + 636, + 438, + 651.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 172, + 651.0, + 438, + 666.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 172, + 666.0, + 438, + 681.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 688, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 171, + 701 + ], + "score": 1.0, + "content": "where Augment", + "type": "text" + }, + { + "bbox": [ + 171, + 688, + 182, + 700 + ], + "score": 0.68, + "content": "( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "indicates normal data augmentation operation including horizontal random flip", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 203, + 713 + ], + "score": 1.0, + "content": "and random crops, and", + "type": "text" + }, + { + "bbox": [ + 203, + 699, + 240, + 712 + ], + "score": 0.93, + "content": "f ^ { k } ( x ; \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 699, + 366, + 713 + ], + "score": 1.0, + "content": "is the predicted probability of", + "type": "text" + }, + { + "bbox": [ + 367, + 702, + 374, + 710 + ], + "score": 0.76, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 699, + 452, + 713 + ], + "score": 1.0, + "content": "belonging to class", + "type": "text" + }, + { + "bbox": [ + 453, + 700, + 459, + 710 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 699, + 506, + 713 + ], + "score": 1.0, + "content": ". Note that", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 144, + 723 + ], + "score": 1.0, + "content": "Augment", + "type": "text" + }, + { + "bbox": [ + 144, + 711, + 155, + 723 + ], + "score": 0.73, + "content": "\\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "is a stochastic transformation, so the two terms in Eq (5) are not identical. The average", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "4", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 81, + 501, + 190 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 81, + 501, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 81, + 501, + 190 + ], + "spans": [ + { + "bbox": [ + 109, + 81, + 501, + 190 + ], + "score": 0.965, + "type": "image", + "image_path": "93c2dd94c2a0f5dd8e852ee49fbcbcd3facdbcf530179d0abf899ddba8541bf1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 81, + 501, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 117.33333333333334, + 501, + 153.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 153.66666666666669, + 501, + 190.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 196, + 506, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "Figure 2: Performance of the traditional early stopping trick and the proposed PES on CIFAR-10", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 207, + 405, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 405, + 220 + ], + "score": 1.0, + "content": "with different types of label noise. The lines present the mean of five runs.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 266, + 242 + ], + "score": 1.0, + "content": "Then, we keep the obtained parameter", + "type": "text" + }, + { + "bbox": [ + 266, + 230, + 280, + 241 + ], + "score": 0.89, + "content": "\\Theta _ { 1 } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 228, + 467, + 242 + ], + "score": 1.0, + "content": "fixed, reinitialize and progressively learn the", + "type": "text" + }, + { + "bbox": [ + 467, + 230, + 471, + 239 + ], + "score": 0.68, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 228, + 487, + 242 + ], + "score": 1.0, + "content": "-th", + "type": "text" + }, + { + "bbox": [ + 487, + 230, + 505, + 240 + ], + "score": 0.79, + "content": "( l =", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 145, + 251 + ], + "score": 0.88, + "content": "2 , \\ldots , L )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "DNN part with the parameters for preceding DNN parts fixed. The training procedure is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 251, + 364, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 170, + 263 + ], + "score": 1.0, + "content": "conducted with", + "type": "text" + }, + { + "bbox": [ + 170, + 251, + 180, + 262 + ], + "score": 0.85, + "content": "T _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 251, + 364, + 263 + ], + "score": 1.0, + "content": "epochs by optimizing the following objective", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 228, + 505, + 263 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 266, + 444, + 298 + ], + "lines": [ + { + "bbox": [ + 167, + 266, + 444, + 298 + ], + "spans": [ + { + "bbox": [ + 167, + 266, + 444, + 298 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { \\Theta _ { l } \\ldots \\Theta _ { k } } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( x _ { i } ; \\Theta _ { 1 } ^ { * } , \\ldots , \\Theta _ { l - 1 } ^ { * } , \\Theta _ { l } , \\ldots , \\Theta _ { L } ) , \\tilde { y } _ { i } ) , \\quad l = 2 \\ldots L", + "type": "interline_equation", + "image_path": "6cf91797c88f1e24c4a6f13692f97ac06e4c0a1ac0851bf9e6baf25712dea2a2.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 167, + 266, + 444, + 276.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 167, + 276.6666666666667, + 444, + 287.33333333333337 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 167, + 287.33333333333337, + 444, + 298.00000000000006 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 211, + 315 + ], + "score": 1.0, + "content": "We gradually optimize the", + "type": "text" + }, + { + "bbox": [ + 211, + 302, + 237, + 314 + ], + "score": 0.91, + "content": "( l + 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 302, + 400, + 315 + ], + "score": 1.0, + "content": "-th DNN part with the obtained parameter", + "type": "text" + }, + { + "bbox": [ + 401, + 303, + 414, + 315 + ], + "score": 0.89, + "content": "\\Theta _ { l } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "fixed, the optimization", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "is continued until all the parameters have been optimized. As elaborated above, latter DNN parts are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 322, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 104, + 322, + 506, + 338 + ], + "score": 1.0, + "content": "more sensitive to noisy labels than their former counterparts. Therefore, for the above initializing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 270, + 348 + ], + "score": 1.0, + "content": "optimization in Eq. (3) and the following", + "type": "text" + }, + { + "bbox": [ + 270, + 335, + 295, + 346 + ], + "score": 0.86, + "content": "L - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "steps of optimization in Eq. (4), we gradually reduce", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 209, + 359 + ], + "score": 1.0, + "content": "the training epochs (i.e.", + "type": "text" + }, + { + "bbox": [ + 209, + 346, + 304, + 357 + ], + "score": 0.89, + "content": "T _ { 1 } \\geq T _ { 2 } \\geq \\cdot \\cdot \\cdot \\geq T _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 346, + 506, + 359 + ], + "score": 1.0, + "content": "to better exploit the memorization effect. After", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 306, + 370 + ], + "score": 1.0, + "content": "optimization, we can obtain the final network as", + "type": "text" + }, + { + "bbox": [ + 307, + 357, + 419, + 369 + ], + "score": 0.91, + "content": "f ( \\cdot , \\Theta ) = \\bar { f } ( \\cdot ; \\Theta _ { 1 } ^ { * } , \\ldots , \\Theta _ { L } ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 357, + 506, + 370 + ], + "score": 1.0, + "content": ". Since this model is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "obtained by progressively exploiting the early stopping strategies for different DNN parts, we term", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 378, + 339, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 339, + 392 + ], + "score": 1.0, + "content": "the proposed method as progressive early stopping (PES).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 302, + 506, + 392 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "To explicitly verify the effectiveness of the proposed PES method, we conduct several pilot experi-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "ments, which compare the traditional early stopping and PES with label noise from different types", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 416, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 506, + 430 + ], + "score": 1.0, + "content": "and different levels. The results are illustrated in Figure 2, from which we can see that, compared", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 428, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 506, + 440 + ], + "score": 1.0, + "content": "with models trained with traditional early stopping, models trained with PES can achieve superior", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "classification accuracy with smaller variations in all the cases. Current state-of-the-art methods [15]", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "usually adopt models with the traditional early stopping as base models to distill confident examples", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "and then utilize semi-supervised learning techniques by considering confident examples as labeled", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "data and other noisy examples as unlabeled data to further improve the results. The final performance", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "still heavily relies on the base model trained with noisy labels. By improving the performance of the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "base model, our method combined with semi-supervised learning techniques is able to establish new", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "state-of-the-art results. In the following subsections, we will elaborate on how to utilize PES to distill", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 515, + 447, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 447, + 528 + ], + "score": 1.0, + "content": "confident examples and further combine it with semi-supervised learning techniques.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 394, + 506, + 528 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 539, + 280, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 281, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 281, + 553 + ], + "score": 1.0, + "content": "2.2 Learning with Confident Examples", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 573 + ], + "score": 1.0, + "content": "Based on the deep network optimized with progressive early stopping, we can select confident", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 570, + 504, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 570, + 504, + 582 + ], + "score": 1.0, + "content": "examples to facilitate the model training. Here, confident examples refer to examples that have high", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "score": 1.0, + "content": "probabilities with clean labels. In this paper, we treat examples whose predictions are consistent with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 591, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 506, + 605 + ], + "score": 1.0, + "content": "given labels as confident examples. In addition, to make the results more robust, we generate two", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 601, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 616 + ], + "score": 1.0, + "content": "different augmentations for any given input and use the average prediction to decide its predicted", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 614, + 358, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 334, + 626 + ], + "score": 1.0, + "content": "label. Formally, we can obtain the confident example set", + "type": "text" + }, + { + "bbox": [ + 334, + 614, + 345, + 625 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 614, + 358, + 626 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 558, + 506, + 626 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 636, + 438, + 681 + ], + "lines": [ + { + "bbox": [ + 172, + 636, + 438, + 681 + ], + "spans": [ + { + "bbox": [ + 172, + 636, + 438, + 681 + ], + "score": 0.91, + "content": "\\begin{array} { c } { \\mathcal { D } _ { l } = \\{ ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = \\hat { y } _ { i } , i = 1 , \\ldots , n \\} , } \\\\ { \\hat { y } _ { i } = \\underset { k \\in \\{ 1 , \\ldots , K \\} } { \\arg \\operatorname* { m a x } } \\ \\frac { 1 } { 2 } [ f ^ { k } ( \\mathrm { A u g m e n t } ( \\pmb { x } _ { i } ) ; \\Theta ) + f ^ { k } ( \\mathrm { A u g m e n t } ( \\pmb { x } _ { i } ) ; \\Theta ) ] , } \\end{array}", + "type": "interline_equation", + "image_path": "f0ca8b2be795a56146eeba463a76de4df49127ec5a09e9825eff949f223573dc.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 172, + 636, + 438, + 651.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 172, + 651.0, + 438, + 666.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 172, + 666.0, + 438, + 681.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 688, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 171, + 701 + ], + "score": 1.0, + "content": "where Augment", + "type": "text" + }, + { + "bbox": [ + 171, + 688, + 182, + 700 + ], + "score": 0.68, + "content": "( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "indicates normal data augmentation operation including horizontal random flip", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 203, + 713 + ], + "score": 1.0, + "content": "and random crops, and", + "type": "text" + }, + { + "bbox": [ + 203, + 699, + 240, + 712 + ], + "score": 0.93, + "content": "f ^ { k } ( x ; \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 699, + 366, + 713 + ], + "score": 1.0, + "content": "is the predicted probability of", + "type": "text" + }, + { + "bbox": [ + 367, + 702, + 374, + 710 + ], + "score": 0.76, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 699, + 452, + 713 + ], + "score": 1.0, + "content": "belonging to class", + "type": "text" + }, + { + "bbox": [ + 453, + 700, + 459, + 710 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 699, + 506, + 713 + ], + "score": 1.0, + "content": ". Note that", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 144, + 723 + ], + "score": 1.0, + "content": "Augment", + "type": "text" + }, + { + "bbox": [ + 144, + 711, + 155, + 723 + ], + "score": 0.73, + "content": "\\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "is a stochastic transformation, so the two terms in Eq (5) are not identical. The average", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 687, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 89, + 485, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 89, + 480, + 102 + ], + "spans": [ + { + "bbox": [ + 106, + 89, + 305, + 102 + ], + "score": 1.0, + "content": "Input: Neural network with trainable parameters", + "type": "text" + }, + { + "bbox": [ + 305, + 89, + 385, + 102 + ], + "score": 0.9, + "content": "\\boldsymbol { \\Theta } = \\{ \\Theta _ { 1 } , \\dots , \\Theta _ { L } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 89, + 480, + 102 + ], + "score": 1.0, + "content": ", Noisy training dataset", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 99, + 484, + 114 + ], + "spans": [ + { + "bbox": [ + 107, + 100, + 157, + 113 + ], + "score": 0.92, + "content": "\\{ \\bar { \\pmb { x } _ { i } } , \\tilde { y } _ { i } ) \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 99, + 342, + 114 + ], + "score": 1.0, + "content": ", Number of training epochs for different part:", + "type": "text" + }, + { + "bbox": [ + 343, + 102, + 388, + 112 + ], + "score": 0.91, + "content": "T _ { 1 } , \\dots , T _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 99, + 472, + 114 + ], + "score": 1.0, + "content": ", and training epochs", + "type": "text" + }, + { + "bbox": [ + 473, + 101, + 484, + 112 + ], + "score": 0.79, + "content": "T _ { c }", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 111, + 256, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 256, + 124 + ], + "score": 1.0, + "content": "for refining with confident examples.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 122, + 192, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 121, + 135 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 123, + 177, + 134 + ], + "score": 0.91, + "content": "i = 1 , \\dots , T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 122, + 192, + 135 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 303, + 146 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 239, + 146 + ], + "score": 1.0, + "content": "Optimize network parameter", + "type": "text" + }, + { + "bbox": [ + 239, + 133, + 248, + 142 + ], + "score": 0.77, + "content": "\\Theta", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 131, + 303, + 146 + ], + "score": 1.0, + "content": "with Eq. (3);", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 146, + 185, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 121, + 157 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 146, + 142, + 156 + ], + "score": 0.79, + "content": "l = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 146, + 162, + 157 + ], + "score": 1.0, + "content": ", . . . ,", + "type": "text" + }, + { + "bbox": [ + 162, + 147, + 170, + 156 + ], + "score": 0.57, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 146, + 185, + 157 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 155, + 342, + 171 + ], + "spans": [ + { + "bbox": [ + 119, + 155, + 147, + 171 + ], + "score": 1.0, + "content": "Froze", + "type": "text" + }, + { + "bbox": [ + 148, + 157, + 213, + 169 + ], + "score": 0.91, + "content": "\\left\\{ \\Theta _ { 1 } , \\ldots , \\Theta _ { l - 1 } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 155, + 280, + 171 + ], + "score": 1.0, + "content": "and re-initialize", + "type": "text" + }, + { + "bbox": [ + 281, + 156, + 338, + 169 + ], + "score": 0.92, + "content": "\\big \\{ \\Theta _ { l } , \\dots , \\Theta _ { L } \\big \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 155, + 342, + 171 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 167, + 205, + 180 + ], + "spans": [ + { + "bbox": [ + 121, + 167, + 137, + 180 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 137, + 169, + 191, + 179 + ], + "score": 0.84, + "content": "i = 1 , \\dots , T _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 167, + 205, + 180 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 122, + 178, + 367, + 192 + ], + "spans": [ + { + "bbox": [ + 122, + 178, + 254, + 192 + ], + "score": 1.0, + "content": "Optimize network parameter", + "type": "text" + }, + { + "bbox": [ + 255, + 178, + 312, + 191 + ], + "score": 0.93, + "content": "\\big \\{ \\Theta _ { l } , \\dots , \\Theta _ { L } \\big \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 178, + 367, + 192 + ], + "score": 1.0, + "content": "with Eq. (4);", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 195, + 155, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 141, + 207 + ], + "score": 1.0, + "content": "Unfroze", + "type": "text" + }, + { + "bbox": [ + 142, + 196, + 150, + 206 + ], + "score": 0.7, + "content": "\\Theta", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 195, + 155, + 207 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 206, + 191, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 121, + 218 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 207, + 176, + 218 + ], + "score": 0.9, + "content": "i = 1 , \\dots , T _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 206, + 191, + 218 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 119, + 215, + 478, + 231 + ], + "spans": [ + { + "bbox": [ + 119, + 215, + 242, + 231 + ], + "score": 1.0, + "content": "Extract confident example set", + "type": "text" + }, + { + "bbox": [ + 242, + 218, + 254, + 229 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 215, + 327, + 231 + ], + "score": 1.0, + "content": "and unlabeled set", + "type": "text" + }, + { + "bbox": [ + 327, + 218, + 341, + 229 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 215, + 401, + 231 + ], + "score": 1.0, + "content": "with classifier", + "type": "text" + }, + { + "bbox": [ + 401, + 217, + 430, + 230 + ], + "score": 0.93, + "content": "f ( \\cdot , \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 215, + 478, + 231 + ], + "score": 1.0, + "content": "by Eq. (7);", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 119, + 227, + 386, + 241 + ], + "spans": [ + { + "bbox": [ + 119, + 227, + 210, + 241 + ], + "score": 1.0, + "content": "Training the classifier", + "type": "text" + }, + { + "bbox": [ + 211, + 229, + 240, + 241 + ], + "score": 0.93, + "content": "f ( \\cdot , \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 227, + 336, + 241 + ], + "score": 1.0, + "content": "with MixMatch loss on", + "type": "text" + }, + { + "bbox": [ + 337, + 229, + 349, + 240 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 227, + 367, + 241 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 229, + 381, + 240 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 227, + 386, + 241 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 267, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 233, + 255 + ], + "score": 1.0, + "content": "Evaluate the obtained classifier", + "type": "text" + }, + { + "bbox": [ + 233, + 242, + 262, + 254 + ], + "score": 0.91, + "content": "f ( \\cdot , \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 240, + 267, + 255 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 280, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "prediction of augmented examples provides a more stable prediction and is found empirically to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "improve performance. After obtaining the confident example set, one can easily train a classifier by", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "considering confident examples as clean data. However, since the number of confident examples for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 311, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 327 + ], + "score": 1.0, + "content": "different classes can vary greatly, directly training the model with the obtained confident example set", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "may introduce a severe class imbalance problem. To this end, we adopt a weighted classification loss", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 348, + 369, + 382 + ], + "lines": [ + { + "bbox": [ + 241, + 348, + 369, + 382 + ], + "spans": [ + { + "bbox": [ + 241, + 348, + 369, + 382 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { c } = \\sum _ { i = 1 } ^ { N } w _ { y _ { i } } \\mathcal { L } _ { p } ( \\tilde { y } _ { i } , f ( \\pmb { x } _ { i } ; \\Theta ) ) ,", + "type": "interline_equation", + "image_path": "2d952161669f286508b4ff5d83652860d38fb213b9f942b25cfcf93e6118bade.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 348, + 369, + 365.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 241, + 365.0, + 369, + 382.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 504, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 133, + 397 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 387, + 145, + 396 + ], + "score": 0.85, + "content": "w _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 384, + 345, + 397 + ], + "score": 1.0, + "content": "is the corresponding class weight. Assuming that", + "type": "text" + }, + { + "bbox": [ + 346, + 384, + 504, + 397 + ], + "score": 0.91, + "content": "\\sigma _ { k } = | \\{ ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = k , ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) \\in \\mathcal { D } _ { l } \\} |", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 388, + 408 + ], + "score": 1.0, + "content": "denotes the cardinality of the confident example set belonging to the", + "type": "text" + }, + { + "bbox": [ + 388, + 397, + 395, + 405 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "-th class. Then, we can set", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 406, + 382, + 423 + ], + "spans": [ + { + "bbox": [ + 107, + 406, + 189, + 422 + ], + "score": 0.92, + "content": "w _ { i } = \\sigma _ { i } / ( \\sum _ { j = 1 } ^ { K } \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 407, + 382, + 423 + ], + "score": 1.0, + "content": "to indicate the corresponding class importance.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 106, + 433, + 316, + 446 + ], + "lines": [ + { + "bbox": [ + 104, + 430, + 317, + 450 + ], + "spans": [ + { + "bbox": [ + 104, + 430, + 317, + 450 + ], + "score": 1.0, + "content": "2.3 Combining with Semi-Supervised Learning", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 453, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "Training with only confident examples neglects the rest data and may suffer from insufficient", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "training examples. To tackle this problem, we further resort to semi-supervised learning techniques", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 507, + 488 + ], + "score": 1.0, + "content": "by considering confident examples as labeled data and other noisy examples as unlabeled data.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 404, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 404, + 499 + ], + "score": 1.0, + "content": "Specifically, the labeled data set and unlabeled data set can be obtained as", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 502, + 437, + 560 + ], + "lines": [ + { + "bbox": [ + 173, + 502, + 437, + 560 + ], + "spans": [ + { + "bbox": [ + 173, + 502, + 437, + 560 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\displaystyle { \\int } { \\mathcal D } _ { l } = \\{ ( \\boldsymbol x _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = \\hat { y } _ { i } , i = 1 , . . . n \\} } } \\\\ { { \\displaystyle { \\mathcal D } _ { u } = \\{ \\boldsymbol x _ { i } | \\tilde { y } _ { i } \\neq \\hat { y } _ { i } , i = 1 , . . . n \\} } } \\\\ { { \\displaystyle { \\hat { y } _ { i } = \\arg \\operatorname* { m a x } _ { k \\in \\{ 1 , . . . K \\} } \\frac { 1 } { 2 } [ f ^ { k } ( \\mathrm { A u g m e n t } ( \\boldsymbol x _ { i } ) ; \\Theta ) + f ^ { k } ( \\mathrm { A u g m e n t } ( \\boldsymbol x _ { i } ) ; \\Theta ) ] } , } } \\end{array}", + "type": "interline_equation", + "image_path": "a92713a47a810a07f4fbd6225dc073925113f0874b4ff035aa8d78e39c94acb1.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 173, + 502, + 437, + 521.3333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 173, + 521.3333333333334, + 437, + 540.6666666666667 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 173, + 540.6666666666667, + 437, + 560.0000000000001 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 563, + 502, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 216, + 575 + ], + "score": 1.0, + "content": "where the labeled data set", + "type": "text" + }, + { + "bbox": [ + 216, + 563, + 228, + 574 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 563, + 369, + 575 + ], + "score": 1.0, + "content": "is the same as that in Eq (6), and", + "type": "text" + }, + { + "bbox": [ + 369, + 564, + 383, + 574 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "is the rest unlabeled data set.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 574, + 504, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 574, + 504, + 586 + ], + "score": 1.0, + "content": "Similar to [15], we adopt MixMatch [4] as the semi-supervised learning framework to train the final", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "classification models. For more details about semi-supervised learning, we refer to [4]. The whole", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 596, + 308, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 308, + 608 + ], + "score": 1.0, + "content": "learning algorithm is summarized in Algorithm 1.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 191, + 636 + ], + "lines": [ + { + "bbox": [ + 104, + 621, + 193, + 639 + ], + "spans": [ + { + "bbox": [ + 104, + 621, + 193, + 639 + ], + "score": 1.0, + "content": "3 Experiments", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 288, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 288, + 660 + ], + "score": 1.0, + "content": "3.1 Datasets and Implementation Details", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 504, + 678 + ], + "score": 1.0, + "content": "Datasets: We evaluate our method on two synthetic datasets, CIFAR-10 and CIFAR-100 [12] with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "different levels of symmetric, pairflip, and instance-dependent label noise (abbreviated as instance", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "label noise) and a real-world dataset Clothing-1M [32]. Both CIFAR-10 and CIFAR-100 contain", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 123, + 711 + ], + "score": 0.46, + "content": "5 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 699, + 206, + 713 + ], + "score": 1.0, + "content": "training images and", + "type": "text" + }, + { + "bbox": [ + 206, + 700, + 223, + 711 + ], + "score": 0.26, + "content": "1 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 699, + 299, + 713 + ], + "score": 1.0, + "content": "test images of size", + "type": "text" + }, + { + "bbox": [ + 300, + 700, + 333, + 711 + ], + "score": 0.89, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 699, + 506, + 713 + ], + "score": 1.0, + "content": ". Following previous works [9, 31, 16, 29],", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 710, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 507, + 725 + ], + "score": 1.0, + "content": "symmetric noise is generated by uniformly flipping labels for a percentage of the training dataset", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 89, + 485, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 89, + 480, + 102 + ], + "spans": [ + { + "bbox": [ + 106, + 89, + 305, + 102 + ], + "score": 1.0, + "content": "Input: Neural network with trainable parameters", + "type": "text" + }, + { + "bbox": [ + 305, + 89, + 385, + 102 + ], + "score": 0.9, + "content": "\\boldsymbol { \\Theta } = \\{ \\Theta _ { 1 } , \\dots , \\Theta _ { L } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 89, + 480, + 102 + ], + "score": 1.0, + "content": ", Noisy training dataset", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 107, + 99, + 484, + 114 + ], + "spans": [ + { + "bbox": [ + 107, + 100, + 157, + 113 + ], + "score": 0.92, + "content": "\\{ \\bar { \\pmb { x } _ { i } } , \\tilde { y } _ { i } ) \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 99, + 342, + 114 + ], + "score": 1.0, + "content": ", Number of training epochs for different part:", + "type": "text" + }, + { + "bbox": [ + 343, + 102, + 388, + 112 + ], + "score": 0.91, + "content": "T _ { 1 } , \\dots , T _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 99, + 472, + 114 + ], + "score": 1.0, + "content": ", and training epochs", + "type": "text" + }, + { + "bbox": [ + 473, + 101, + 484, + 112 + ], + "score": 0.79, + "content": "T _ { c }", + "type": "inline_equation" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 111, + 256, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 256, + 124 + ], + "score": 1.0, + "content": "for refining with confident examples.", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 122, + 192, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 121, + 135 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 123, + 177, + 134 + ], + "score": 0.91, + "content": "i = 1 , \\dots , T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 122, + 192, + 135 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 131, + 303, + 146 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 239, + 146 + ], + "score": 1.0, + "content": "Optimize network parameter", + "type": "text" + }, + { + "bbox": [ + 239, + 133, + 248, + 142 + ], + "score": 0.77, + "content": "\\Theta", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 131, + 303, + 146 + ], + "score": 1.0, + "content": "with Eq. (3);", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 146, + 185, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 121, + 157 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 146, + 142, + 156 + ], + "score": 0.79, + "content": "l = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 146, + 162, + 157 + ], + "score": 1.0, + "content": ", . . . ,", + "type": "text" + }, + { + "bbox": [ + 162, + 147, + 170, + 156 + ], + "score": 0.57, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 146, + 185, + 157 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 119, + 155, + 342, + 171 + ], + "spans": [ + { + "bbox": [ + 119, + 155, + 147, + 171 + ], + "score": 1.0, + "content": "Froze", + "type": "text" + }, + { + "bbox": [ + 148, + 157, + 213, + 169 + ], + "score": 0.91, + "content": "\\left\\{ \\Theta _ { 1 } , \\ldots , \\Theta _ { l - 1 } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 155, + 280, + 171 + ], + "score": 1.0, + "content": "and re-initialize", + "type": "text" + }, + { + "bbox": [ + 281, + 156, + 338, + 169 + ], + "score": 0.92, + "content": "\\big \\{ \\Theta _ { l } , \\dots , \\Theta _ { L } \\big \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 155, + 342, + 171 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 121, + 167, + 205, + 180 + ], + "spans": [ + { + "bbox": [ + 121, + 167, + 137, + 180 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 137, + 169, + 191, + 179 + ], + "score": 0.84, + "content": "i = 1 , \\dots , T _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 167, + 205, + 180 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 122, + 178, + 367, + 192 + ], + "spans": [ + { + "bbox": [ + 122, + 178, + 254, + 192 + ], + "score": 1.0, + "content": "Optimize network parameter", + "type": "text" + }, + { + "bbox": [ + 255, + 178, + 312, + 191 + ], + "score": 0.93, + "content": "\\big \\{ \\Theta _ { l } , \\dots , \\Theta _ { L } \\big \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 178, + 367, + 192 + ], + "score": 1.0, + "content": "with Eq. (4);", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 195, + 155, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 141, + 207 + ], + "score": 1.0, + "content": "Unfroze", + "type": "text" + }, + { + "bbox": [ + 142, + 196, + 150, + 206 + ], + "score": 0.7, + "content": "\\Theta", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 195, + 155, + 207 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 206, + 191, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 121, + 218 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 207, + 176, + 218 + ], + "score": 0.9, + "content": "i = 1 , \\dots , T _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 206, + 191, + 218 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 119, + 215, + 478, + 231 + ], + "spans": [ + { + "bbox": [ + 119, + 215, + 242, + 231 + ], + "score": 1.0, + "content": "Extract confident example set", + "type": "text" + }, + { + "bbox": [ + 242, + 218, + 254, + 229 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 215, + 327, + 231 + ], + "score": 1.0, + "content": "and unlabeled set", + "type": "text" + }, + { + "bbox": [ + 327, + 218, + 341, + 229 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 215, + 401, + 231 + ], + "score": 1.0, + "content": "with classifier", + "type": "text" + }, + { + "bbox": [ + 401, + 217, + 430, + 230 + ], + "score": 0.93, + "content": "f ( \\cdot , \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 215, + 478, + 231 + ], + "score": 1.0, + "content": "by Eq. (7);", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 119, + 227, + 386, + 241 + ], + "spans": [ + { + "bbox": [ + 119, + 227, + 210, + 241 + ], + "score": 1.0, + "content": "Training the classifier", + "type": "text" + }, + { + "bbox": [ + 211, + 229, + 240, + 241 + ], + "score": 0.93, + "content": "f ( \\cdot , \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 227, + 336, + 241 + ], + "score": 1.0, + "content": "with MixMatch loss on", + "type": "text" + }, + { + "bbox": [ + 337, + 229, + 349, + 240 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 227, + 367, + 241 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 229, + 381, + 240 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 227, + 386, + 241 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 240, + 267, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 233, + 255 + ], + "score": 1.0, + "content": "Evaluate the obtained classifier", + "type": "text" + }, + { + "bbox": [ + 233, + 242, + 262, + 254 + ], + "score": 0.91, + "content": "f ( \\cdot , \\Theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 240, + 267, + 255 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 89, + 484, + 255 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 280, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 293 + ], + "score": 1.0, + "content": "prediction of augmented examples provides a more stable prediction and is found empirically to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "improve performance. After obtaining the confident example set, one can easily train a classifier by", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "considering confident examples as clean data. However, since the number of confident examples for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 311, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 327 + ], + "score": 1.0, + "content": "different classes can vary greatly, directly training the model with the obtained confident example set", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "may introduce a severe class imbalance problem. To this end, we adopt a weighted classification loss", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 280, + 506, + 336 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 348, + 369, + 382 + ], + "lines": [ + { + "bbox": [ + 241, + 348, + 369, + 382 + ], + "spans": [ + { + "bbox": [ + 241, + 348, + 369, + 382 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { c } = \\sum _ { i = 1 } ^ { N } w _ { y _ { i } } \\mathcal { L } _ { p } ( \\tilde { y } _ { i } , f ( \\pmb { x } _ { i } ; \\Theta ) ) ,", + "type": "interline_equation", + "image_path": "2d952161669f286508b4ff5d83652860d38fb213b9f942b25cfcf93e6118bade.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 348, + 369, + 365.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 241, + 365.0, + 369, + 382.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 504, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 133, + 397 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 387, + 145, + 396 + ], + "score": 0.85, + "content": "w _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 384, + 345, + 397 + ], + "score": 1.0, + "content": "is the corresponding class weight. Assuming that", + "type": "text" + }, + { + "bbox": [ + 346, + 384, + 504, + 397 + ], + "score": 0.91, + "content": "\\sigma _ { k } = | \\{ ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = k , ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) \\in \\mathcal { D } _ { l } \\} |", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 388, + 408 + ], + "score": 1.0, + "content": "denotes the cardinality of the confident example set belonging to the", + "type": "text" + }, + { + "bbox": [ + 388, + 397, + 395, + 405 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "-th class. Then, we can set", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 406, + 382, + 423 + ], + "spans": [ + { + "bbox": [ + 107, + 406, + 189, + 422 + ], + "score": 0.92, + "content": "w _ { i } = \\sigma _ { i } / ( \\sum _ { j = 1 } ^ { K } \\sigma _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 407, + 382, + 423 + ], + "score": 1.0, + "content": "to indicate the corresponding class importance.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 384, + 506, + 423 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 433, + 316, + 446 + ], + "lines": [ + { + "bbox": [ + 104, + 430, + 317, + 450 + ], + "spans": [ + { + "bbox": [ + 104, + 430, + 317, + 450 + ], + "score": 1.0, + "content": "2.3 Combining with Semi-Supervised Learning", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 453, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "Training with only confident examples neglects the rest data and may suffer from insufficient", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "training examples. To tackle this problem, we further resort to semi-supervised learning techniques", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 507, + 488 + ], + "score": 1.0, + "content": "by considering confident examples as labeled data and other noisy examples as unlabeled data.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 404, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 404, + 499 + ], + "score": 1.0, + "content": "Specifically, the labeled data set and unlabeled data set can be obtained as", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 454, + 507, + 499 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 502, + 437, + 560 + ], + "lines": [ + { + "bbox": [ + 173, + 502, + 437, + 560 + ], + "spans": [ + { + "bbox": [ + 173, + 502, + 437, + 560 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\displaystyle { \\int } { \\mathcal D } _ { l } = \\{ ( \\boldsymbol x _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = \\hat { y } _ { i } , i = 1 , . . . n \\} } } \\\\ { { \\displaystyle { \\mathcal D } _ { u } = \\{ \\boldsymbol x _ { i } | \\tilde { y } _ { i } \\neq \\hat { y } _ { i } , i = 1 , . . . n \\} } } \\\\ { { \\displaystyle { \\hat { y } _ { i } = \\arg \\operatorname* { m a x } _ { k \\in \\{ 1 , . . . K \\} } \\frac { 1 } { 2 } [ f ^ { k } ( \\mathrm { A u g m e n t } ( \\boldsymbol x _ { i } ) ; \\Theta ) + f ^ { k } ( \\mathrm { A u g m e n t } ( \\boldsymbol x _ { i } ) ; \\Theta ) ] } , } } \\end{array}", + "type": "interline_equation", + "image_path": "a92713a47a810a07f4fbd6225dc073925113f0874b4ff035aa8d78e39c94acb1.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 173, + 502, + 437, + 521.3333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 173, + 521.3333333333334, + 437, + 540.6666666666667 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 173, + 540.6666666666667, + 437, + 560.0000000000001 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 563, + 502, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 216, + 575 + ], + "score": 1.0, + "content": "where the labeled data set", + "type": "text" + }, + { + "bbox": [ + 216, + 563, + 228, + 574 + ], + "score": 0.88, + "content": "\\mathcal { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 563, + 369, + 575 + ], + "score": 1.0, + "content": "is the same as that in Eq (6), and", + "type": "text" + }, + { + "bbox": [ + 369, + 564, + 383, + 574 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "is the rest unlabeled data set.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 574, + 504, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 574, + 504, + 586 + ], + "score": 1.0, + "content": "Similar to [15], we adopt MixMatch [4] as the semi-supervised learning framework to train the final", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "classification models. For more details about semi-supervised learning, we refer to [4]. The whole", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 596, + 308, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 308, + 608 + ], + "score": 1.0, + "content": "learning algorithm is summarized in Algorithm 1.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 563, + 505, + 608 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 191, + 636 + ], + "lines": [ + { + "bbox": [ + 104, + 621, + 193, + 639 + ], + "spans": [ + { + "bbox": [ + 104, + 621, + 193, + 639 + ], + "score": 1.0, + "content": "3 Experiments", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 288, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 288, + 660 + ], + "score": 1.0, + "content": "3.1 Datasets and Implementation Details", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 504, + 678 + ], + "score": 1.0, + "content": "Datasets: We evaluate our method on two synthetic datasets, CIFAR-10 and CIFAR-100 [12] with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "different levels of symmetric, pairflip, and instance-dependent label noise (abbreviated as instance", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "label noise) and a real-world dataset Clothing-1M [32]. Both CIFAR-10 and CIFAR-100 contain", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 123, + 711 + ], + "score": 0.46, + "content": "5 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 699, + 206, + 713 + ], + "score": 1.0, + "content": "training images and", + "type": "text" + }, + { + "bbox": [ + 206, + 700, + 223, + 711 + ], + "score": 0.26, + "content": "1 0 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 699, + 299, + 713 + ], + "score": 1.0, + "content": "test images of size", + "type": "text" + }, + { + "bbox": [ + 300, + 700, + 333, + 711 + ], + "score": 0.89, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 699, + 506, + 713 + ], + "score": 1.0, + "content": ". Following previous works [9, 31, 16, 29],", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 710, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 507, + 725 + ], + "score": 1.0, + "content": "symmetric noise is generated by uniformly flipping labels for a percentage of the training dataset", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "to all possible labels. Pairflip noise flips noisy labels into their adjacent class. And, instance", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 220, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 231 + ], + "score": 1.0, + "content": "noise is generated by image features. More details about the synthetic label noise are given in the", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 231, + 506, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 242 + ], + "score": 1.0, + "content": "supplementary material. For the flipping rate, it can include [9, 31] or ex-include [15, 16] true labels.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "We use the flipping rate including correct labels in Table 3 to compare with results in [15], and use", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 253, + 504, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 504, + 264 + ], + "score": 1.0, + "content": "without correct labels in the rest of the experiments. Clothing-1M [32] is a large-scale dataset with", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "real-world noisy labels, whose images are clawed from the online shopping websites, and labels", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 445, + 287 + ], + "score": 1.0, + "content": "are generated based on surrounding texts. It contains 1 million training images, and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 445, + 274, + 462, + 285 + ], + "score": 0.46, + "content": "1 5 \\mathrm { k }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 462, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "validation", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 286, + 293, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 293, + 297 + ], + "score": 1.0, + "content": "images, and 10k test images with clean labels.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 667, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 104, + 504, + 190 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 77, + 504, + 100 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 90 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 90 + ], + "score": 1.0, + "content": "Table 1: Preliminary analysis of the performance and the quality of extracted confident examples on", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 87, + 404, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 87, + 404, + 101 + ], + "score": 1.0, + "content": "CIFAR-10. The mean and standard deviation are computed over five runs.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 104, + 504, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 104, + 504, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 104, + 504, + 190 + ], + "score": 0.984, + "html": "
MetricsMethodsSym-20%Sym-50%Pair-45%Inst-20%Inst-40%
Test AccuracyEarly StoppingPES82.55±2.4685.87±1.5970.76±1.2475.87±1.3360.62±5.5962.40±2.3484.41±0.9086.58±0.4574.73±2.6577.07±1.18
Label PrecisionEarly StoppingPES98.81±0.1598.96±0.0994.65±0.1995.46±0.1472.53±5.2672.99±2.2798.70±0.4398.52±0.1990.77±1.8790.63±0.92
Label RecallEarly StoppingPES88.51±2.2692.67±1.4375.18±1.0081.03±1.8367.84±5.0671.06±2.2790.37±1.0193.24±0.6082.15±3.1785.91±0.68
", + "type": "table", + "image_path": "00db2b4f97c9dec5546d30e722898d28a5de7b60aab6b4c0b55768812cd9dac7.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 108, + 104, + 504, + 132.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 132.66666666666666, + 504, + 161.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 161.33333333333331, + 504, + 189.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 106, + 208, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "to all possible labels. Pairflip noise flips noisy labels into their adjacent class. And, instance", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 220, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 231 + ], + "score": 1.0, + "content": "noise is generated by image features. More details about the synthetic label noise are given in the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 231, + 506, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 242 + ], + "score": 1.0, + "content": "supplementary material. For the flipping rate, it can include [9, 31] or ex-include [15, 16] true labels.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "We use the flipping rate including correct labels in Table 3 to compare with results in [15], and use", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 253, + 504, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 504, + 264 + ], + "score": 1.0, + "content": "without correct labels in the rest of the experiments. Clothing-1M [32] is a large-scale dataset with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "real-world noisy labels, whose images are clawed from the online shopping websites, and labels", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 445, + 287 + ], + "score": 1.0, + "content": "are generated based on surrounding texts. It contains 1 million training images, and", + "type": "text" + }, + { + "bbox": [ + 445, + 274, + 462, + 285 + ], + "score": 0.46, + "content": "1 5 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "validation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 286, + 293, + 297 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 293, + 297 + ], + "score": 1.0, + "content": "images, and 10k test images with clean labels.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "score": 1.0, + "content": "Baselines: Semi-supervised learning may strongly boost the performance, we separately compare", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "our method with approaches with or without semi-supervised learning. For the comparison with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "baselines with semi-supervised learning, we combine our proposed method with MixMatch used in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 333, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 348 + ], + "score": 1.0, + "content": "[15] as indicated in Subsection 2.3. (1) Approaches without semi-supervised learning: Co-teaching", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "[9], Forward [24], Joint Optim [25], T-revision [31], DMI [34], and CDR [29]. (2) Methods with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 363, + 369 + ], + "score": 1.0, + "content": "semi-supervised learning: M-correction [1], DivideMix [15], and", + "type": "text" + }, + { + "bbox": [ + 363, + 357, + 389, + 367 + ], + "score": 0.5, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "[16]. We also adopt standard", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 367, + 451, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 451, + 380 + ], + "score": 1.0, + "content": "training with cross-entropy (CE) and MixUp [37] as baselines to show improvements.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 505, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "Network structure and optimization: Our method is implemented by PyTorch v1.6. Baseline", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "methods are implemented based on public codes with hyper-parameters set according to the original", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 220, + 418 + ], + "score": 1.0, + "content": "papers. For DivideMix and", + "type": "text" + }, + { + "bbox": [ + 221, + 405, + 247, + 416 + ], + "score": 0.73, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 404, + 506, + 418 + ], + "score": 1.0, + "content": ", we evaluate the test accuracy with the first network. To better", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "demonstrate the robustness of our algorithm, we keep the hyper-parameters fixed for different types", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 427, + 423, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 423, + 440 + ], + "score": 1.0, + "content": "of label noise. More technique details are given in the supplementary material.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 443, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "For experiments without semi-supervised learning, we follow [31], and use ResNet-18 [10] for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 454, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 506, + 466 + ], + "score": 1.0, + "content": "CIFAR-10 and ResNet-34 for CIFAR-100. We split networks into three parts, the layers above block", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 372, + 477 + ], + "score": 1.0, + "content": "4 as part 1, block 4 of ResNet as part 2, and the final layer as part 3.", + "type": "text" + }, + { + "bbox": [ + 372, + 465, + 384, + 476 + ], + "score": 0.86, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "is defined as 25 for CIFAR-10", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 203, + 488 + ], + "score": 1.0, + "content": "and 30 for CIFAR-100,", + "type": "text" + }, + { + "bbox": [ + 203, + 476, + 215, + 487 + ], + "score": 0.86, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 475, + 254, + 488 + ], + "score": 1.0, + "content": "as 7, and", + "type": "text" + }, + { + "bbox": [ + 254, + 476, + 266, + 487 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "as 5. The network is trained for 200 epochs and SGD with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "score": 1.0, + "content": "0.9 momentum is used. The initial learning rate is set to 0.1 and decayed with a factor of 10 at the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 496, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 104, + 496, + 360, + 511 + ], + "score": 1.0, + "content": "100th and 150th epoch respectively, and a weight decay is set to", + "type": "text" + }, + { + "bbox": [ + 361, + 497, + 382, + 508 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 496, + 402, + 511 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 403, + 498, + 414, + 509 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 496, + 432, + 511 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 498, + 443, + 509 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 496, + 506, + 511 + ], + "score": 1.0, + "content": ", we employ an", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 507, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 264, + 521 + ], + "score": 1.0, + "content": "Adam optimizer with a learning rate of", + "type": "text" + }, + { + "bbox": [ + 264, + 508, + 286, + 519 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 507, + 290, + 521 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 506, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 507, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 507, + 538 + ], + "score": 1.0, + "content": "For experiments with semi-supervised learning, we follow the setting of [15] with PreAct Resnet-18.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 316, + 549 + ], + "score": 1.0, + "content": "We set the final layer as part 2, the rest as part 1.", + "type": "text" + }, + { + "bbox": [ + 317, + 536, + 328, + 547 + ], + "score": 0.85, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "is defined as 20 for CIFAR-10 and 35 for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 176, + 559 + ], + "score": 1.0, + "content": "CIFAR-100, and", + "type": "text" + }, + { + "bbox": [ + 176, + 547, + 187, + 558 + ], + "score": 0.89, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "as 5. The network is trained for 300 epochs. For optimization, we use a single", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 373, + 570 + ], + "score": 1.0, + "content": "cycle of cosine annealing [19], and the learning rate begins from", + "type": "text" + }, + { + "bbox": [ + 374, + 558, + 413, + 569 + ], + "score": 0.91, + "content": "2 \\times \\mathrm { \\bar { 1 0 ^ { - 2 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 557, + 463, + 570 + ], + "score": 1.0, + "content": "and ends at", + "type": "text" + }, + { + "bbox": [ + 463, + 558, + 502, + 568 + ], + "score": 0.91, + "content": "2 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 557, + 506, + 570 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 508, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 198, + 582 + ], + "score": 1.0, + "content": "with a weight decay of", + "type": "text" + }, + { + "bbox": [ + 198, + 569, + 237, + 579 + ], + "score": 0.92, + "content": "5 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 568, + 455, + 582 + ], + "score": 1.0, + "content": ". An Adam optimizer is adopted with a learning rate of", + "type": "text" + }, + { + "bbox": [ + 455, + 569, + 477, + 579 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 568, + 491, + 582 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 492, + 569, + 503, + 580 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 568, + 508, + 582 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 579, + 455, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 455, + 592 + ], + "score": 1.0, + "content": "For hyper-parameters from MixMatch, we set them according to the original paper [4].", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "For Clothing-1M [32], we follow the previous work [25], and employ a ResNet-50 [10] pre-trained", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 372, + 619 + ], + "score": 1.0, + "content": "on ImageNet [13]. We set the final layer as part 2, the rest as part 1.", + "type": "text" + }, + { + "bbox": [ + 372, + 608, + 384, + 618 + ], + "score": 0.87, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 607, + 401, + 619 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 608, + 413, + 618 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 607, + 506, + 619 + ], + "score": 1.0, + "content": "are defined as 20 and 7", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "respectively. The network is trained with CE loss for 50 epochs and SGD is used with 0.9 momentum", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 196, + 642 + ], + "score": 1.0, + "content": "and a weight decay of", + "type": "text" + }, + { + "bbox": [ + 196, + 628, + 217, + 639 + ], + "score": 0.91, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 627, + 301, + 642 + ], + "score": 1.0, + "content": ". The learning rate is", + "type": "text" + }, + { + "bbox": [ + 301, + 628, + 340, + 640 + ], + "score": 0.92, + "content": "5 \\times 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 627, + 506, + 642 + ], + "score": 1.0, + "content": "and decayed by a factor of 10 at the 20th", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 639, + 503, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 437, + 653 + ], + "score": 1.0, + "content": "and 30th epoch respectively. We employ an Adam optimizer with a learning rate of", + "type": "text" + }, + { + "bbox": [ + 437, + 639, + 476, + 650 + ], + "score": 0.92, + "content": "5 \\times 1 0 ^ { - 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 639, + 491, + 653 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 492, + 640, + 503, + 651 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "title", + "bbox": [ + 107, + 667, + 239, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 240, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 240, + 682 + ], + "score": 1.0, + "content": "3.2 Preliminary Experiments", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "In Figure 2, we can observe that with the PES trick, the performance of classifiers is generally", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "score": 1.0, + "content": "improved compared with that the traditional early stopping trick. In this section, we further carefully", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 710, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 725 + ], + "score": 1.0, + "content": "analyze the quality of extracted labels by examining them from three aspects, i.e., test accuracy,", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 104, + 504, + 190 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 77, + 504, + 100 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 90 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 90 + ], + "score": 1.0, + "content": "Table 1: Preliminary analysis of the performance and the quality of extracted confident examples on", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 87, + 404, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 87, + 404, + 101 + ], + "score": 1.0, + "content": "CIFAR-10. The mean and standard deviation are computed over five runs.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 104, + 504, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 104, + 504, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 104, + 504, + 190 + ], + "score": 0.984, + "html": "
MetricsMethodsSym-20%Sym-50%Pair-45%Inst-20%Inst-40%
Test AccuracyEarly StoppingPES82.55±2.4685.87±1.5970.76±1.2475.87±1.3360.62±5.5962.40±2.3484.41±0.9086.58±0.4574.73±2.6577.07±1.18
Label PrecisionEarly StoppingPES98.81±0.1598.96±0.0994.65±0.1995.46±0.1472.53±5.2672.99±2.2798.70±0.4398.52±0.1990.77±1.8790.63±0.92
Label RecallEarly StoppingPES88.51±2.2692.67±1.4375.18±1.0081.03±1.8367.84±5.0671.06±2.2790.37±1.0193.24±0.6082.15±3.1785.91±0.68
", + "type": "table", + "image_path": "00db2b4f97c9dec5546d30e722898d28a5de7b60aab6b4c0b55768812cd9dac7.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 108, + 104, + 504, + 132.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 132.66666666666666, + 504, + 161.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 161.33333333333331, + 504, + 189.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 106, + 208, + 505, + 297 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 105, + 209, + 506, + 297 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 315 + ], + "score": 1.0, + "content": "Baselines: Semi-supervised learning may strongly boost the performance, we separately compare", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "our method with approaches with or without semi-supervised learning. For the comparison with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "baselines with semi-supervised learning, we combine our proposed method with MixMatch used in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 333, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 348 + ], + "score": 1.0, + "content": "[15] as indicated in Subsection 2.3. (1) Approaches without semi-supervised learning: Co-teaching", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "[9], Forward [24], Joint Optim [25], T-revision [31], DMI [34], and CDR [29]. (2) Methods with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 363, + 369 + ], + "score": 1.0, + "content": "semi-supervised learning: M-correction [1], DivideMix [15], and", + "type": "text" + }, + { + "bbox": [ + 363, + 357, + 389, + 367 + ], + "score": 0.5, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "[16]. We also adopt standard", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 367, + 451, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 451, + 380 + ], + "score": 1.0, + "content": "training with cross-entropy (CE) and MixUp [37] as baselines to show improvements.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 300, + 506, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 505, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "Network structure and optimization: Our method is implemented by PyTorch v1.6. Baseline", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "methods are implemented based on public codes with hyper-parameters set according to the original", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 220, + 418 + ], + "score": 1.0, + "content": "papers. For DivideMix and", + "type": "text" + }, + { + "bbox": [ + 221, + 405, + 247, + 416 + ], + "score": 0.73, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 404, + 506, + 418 + ], + "score": 1.0, + "content": ", we evaluate the test accuracy with the first network. To better", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "demonstrate the robustness of our algorithm, we keep the hyper-parameters fixed for different types", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 427, + 423, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 423, + 440 + ], + "score": 1.0, + "content": "of label noise. More technique details are given in the supplementary material.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 384, + 506, + 440 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 443, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "For experiments without semi-supervised learning, we follow [31], and use ResNet-18 [10] for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 454, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 506, + 466 + ], + "score": 1.0, + "content": "CIFAR-10 and ResNet-34 for CIFAR-100. We split networks into three parts, the layers above block", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 372, + 477 + ], + "score": 1.0, + "content": "4 as part 1, block 4 of ResNet as part 2, and the final layer as part 3.", + "type": "text" + }, + { + "bbox": [ + 372, + 465, + 384, + 476 + ], + "score": 0.86, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "is defined as 25 for CIFAR-10", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 203, + 488 + ], + "score": 1.0, + "content": "and 30 for CIFAR-100,", + "type": "text" + }, + { + "bbox": [ + 203, + 476, + 215, + 487 + ], + "score": 0.86, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 475, + 254, + 488 + ], + "score": 1.0, + "content": "as 7, and", + "type": "text" + }, + { + "bbox": [ + 254, + 476, + 266, + 487 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "as 5. The network is trained for 200 epochs and SGD with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "score": 1.0, + "content": "0.9 momentum is used. The initial learning rate is set to 0.1 and decayed with a factor of 10 at the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 496, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 104, + 496, + 360, + 511 + ], + "score": 1.0, + "content": "100th and 150th epoch respectively, and a weight decay is set to", + "type": "text" + }, + { + "bbox": [ + 361, + 497, + 382, + 508 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 496, + 402, + 511 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 403, + 498, + 414, + 509 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 496, + 432, + 511 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 498, + 443, + 509 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 496, + 506, + 511 + ], + "score": 1.0, + "content": ", we employ an", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 507, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 264, + 521 + ], + "score": 1.0, + "content": "Adam optimizer with a learning rate of", + "type": "text" + }, + { + "bbox": [ + 264, + 508, + 286, + 519 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 507, + 290, + 521 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 443, + 506, + 521 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 506, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 507, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 507, + 538 + ], + "score": 1.0, + "content": "For experiments with semi-supervised learning, we follow the setting of [15] with PreAct Resnet-18.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 316, + 549 + ], + "score": 1.0, + "content": "We set the final layer as part 2, the rest as part 1.", + "type": "text" + }, + { + "bbox": [ + 317, + 536, + 328, + 547 + ], + "score": 0.85, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 536, + 506, + 549 + ], + "score": 1.0, + "content": "is defined as 20 for CIFAR-10 and 35 for", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 176, + 559 + ], + "score": 1.0, + "content": "CIFAR-100, and", + "type": "text" + }, + { + "bbox": [ + 176, + 547, + 187, + 558 + ], + "score": 0.89, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "as 5. The network is trained for 300 epochs. For optimization, we use a single", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 373, + 570 + ], + "score": 1.0, + "content": "cycle of cosine annealing [19], and the learning rate begins from", + "type": "text" + }, + { + "bbox": [ + 374, + 558, + 413, + 569 + ], + "score": 0.91, + "content": "2 \\times \\mathrm { \\bar { 1 0 ^ { - 2 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 557, + 463, + 570 + ], + "score": 1.0, + "content": "and ends at", + "type": "text" + }, + { + "bbox": [ + 463, + 558, + 502, + 568 + ], + "score": 0.91, + "content": "2 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 557, + 506, + 570 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 508, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 198, + 582 + ], + "score": 1.0, + "content": "with a weight decay of", + "type": "text" + }, + { + "bbox": [ + 198, + 569, + 237, + 579 + ], + "score": 0.92, + "content": "5 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 568, + 455, + 582 + ], + "score": 1.0, + "content": ". An Adam optimizer is adopted with a learning rate of", + "type": "text" + }, + { + "bbox": [ + 455, + 569, + 477, + 579 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 568, + 491, + 582 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 492, + 569, + 503, + 580 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 568, + 508, + 582 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 579, + 455, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 455, + 592 + ], + "score": 1.0, + "content": "For hyper-parameters from MixMatch, we set them according to the original paper [4].", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 525, + 508, + 592 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "For Clothing-1M [32], we follow the previous work [25], and employ a ResNet-50 [10] pre-trained", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 372, + 619 + ], + "score": 1.0, + "content": "on ImageNet [13]. We set the final layer as part 2, the rest as part 1.", + "type": "text" + }, + { + "bbox": [ + 372, + 608, + 384, + 618 + ], + "score": 0.87, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 607, + 401, + 619 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 608, + 413, + 618 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 607, + 506, + 619 + ], + "score": 1.0, + "content": "are defined as 20 and 7", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "respectively. The network is trained with CE loss for 50 epochs and SGD is used with 0.9 momentum", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 196, + 642 + ], + "score": 1.0, + "content": "and a weight decay of", + "type": "text" + }, + { + "bbox": [ + 196, + 628, + 217, + 639 + ], + "score": 0.91, + "content": "1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 627, + 301, + 642 + ], + "score": 1.0, + "content": ". The learning rate is", + "type": "text" + }, + { + "bbox": [ + 301, + 628, + 340, + 640 + ], + "score": 0.92, + "content": "5 \\times 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 627, + 506, + 642 + ], + "score": 1.0, + "content": "and decayed by a factor of 10 at the 20th", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 639, + 503, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 437, + 653 + ], + "score": 1.0, + "content": "and 30th epoch respectively. We employ an Adam optimizer with a learning rate of", + "type": "text" + }, + { + "bbox": [ + 437, + 639, + 476, + 650 + ], + "score": 0.92, + "content": "5 \\times 1 0 ^ { - 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 639, + 491, + 653 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 492, + 640, + 503, + 651 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 596, + 506, + 653 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 667, + 239, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 240, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 240, + 682 + ], + "score": 1.0, + "content": "3.2 Preliminary Experiments", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "In Figure 2, we can observe that with the PES trick, the performance of classifiers is generally", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "score": 1.0, + "content": "improved compared with that the traditional early stopping trick. In this section, we further carefully", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 710, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 725 + ], + "score": 1.0, + "content": "analyze the quality of extracted labels by examining them from three aspects, i.e., test accuracy,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "score": 1.0, + "content": "label precision, and label recall. Here, label precision indicates the ratio of the number of extracted", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "confident examples with correct labels in the total confident example set, and label recall represents", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "the ratio of the number of confident examples with correct labels among the total correctly labeled", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 518, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 530 + ], + "score": 1.0, + "content": "examples. Specifically, we train a neural network on CIFAR-10 with different kinds and levels of", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "label noise for 25 epochs respectively and report the performance for each case before and after the", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 540, + 209, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 209, + 552 + ], + "score": 1.0, + "content": "proposed PES is applied.", + "type": "text", + "cross_page": true + } + ], + "index": 16 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 688, + 506, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 104, + 492, + 317 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 77, + 505, + 100 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 78, + 505, + 89 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 505, + 89 + ], + "score": 1.0, + "content": "Table 2: Comparison with state-of-the-art methods without semi-supervised learning on CIFAR-10", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 87, + 466, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 87, + 466, + 101 + ], + "score": 1.0, + "content": "and CIFAR-100. The mean and standard deviation computed over five runs are presented.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 116, + 104, + 492, + 317 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 104, + 492, + 317 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 492, + 317 + ], + "score": 0.986, + "html": "
DatasetMethodSymmetricPairflipInstance
20%50%45%20%40%
CIFAR10CE84.00±0.6675.51±1.2463.34±6.0385.10±0.6877.00±2.17
Co-teaching87.16±0.1172.80±0.4570.11±1.1686.54±0.1180.98±0.39
Forward85.63±0.5277.92±0.6660.15±1.9785.29±0.3874.72±3.24
Joint Optim89.70±0.1185.00±0.1782.63±1.3889.69±0.4282.62±0.57
T-revision89.63±0.1383.40±0.6577.06±6.4790.46±0.1385.37±3.36
DMI88.18±0.3678.28±0.4857.60±14.5689.14±0.3684.78±1.97
CDR89.72±0.3882.64±0.8973.67±0.5490.41±0.3483.07±1.33
Ours92.38±0.4087.45±0.3588.43±1.0892.69±0.4489.73±0.51
CIFAR100CE51.43±0.5837.69±3.4534.10±2.0452.19±1.4242.26±1.29
Co-teaching59.28±0.4741.37±0.0833.22±0.4857.24±0.6945.69±0.99
Forward57.75±0.3744.66±1.0127.88±0.8058.76±0.6644.50±0.72
Joint Optim64.55±0.3850.22±0.4142.61±0.6165.15±0.3155.57±0.41
T-revision65.40±1.0750.24±1.4541.10±1.9560.71±0.7351.54±0.91
DMI58.73±0.7044.25±1.1426.90±0.4558.05±0.2047.36±0.68
CDR66.52±0.2455.30±0.9643.87±1.3567.33±0.6755.94±0.56
Ours68.89±0.4558.90±2.7257.18±1.4470.49±0.7965.68±1.41
", + "type": "table", + "image_path": "faf2277129f4013a50d7d87c0c93cc22f297e8287191fc976ac1e57bb5e493c0.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 116, + 104, + 492, + 175.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 116, + 175.0, + 492, + 246.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 116, + 246.0, + 492, + 317.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 131, + 367, + 480, + 475 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 329, + 505, + 363 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "Table 3: Comparison with state-of-the-art methods with semi-supervised learning on CIFAR-10 and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "CIFAR-100 with symmetric label noise from different levels. Results with * are token from [15]. The", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 351, + 344, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 344, + 364 + ], + "score": 1.0, + "content": "mean and standard deviation are computed over three runs.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 131, + 367, + 480, + 475 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 131, + 367, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 367, + 480, + 475 + ], + "score": 0.984, + "html": "
DatasetCIFAR-10CIFAR-100
Methods /NoiseSym-20%Sym-50%Sym-80%Sym-20%Sym-50%Sym-80%
CE86.5±0.680.6±0.263.7±0.857.9±0.447.3±0.222.3±1.2
MixUp93.2±0.388.2±0.373.3±0.369.5±0.257.1±0.634.1±0.6
M-correction*94.092.086.873.966.148.2
DivideMix*95.294.293.075.272.858.3
DivideMix95.6±0.194.6±0.192.9±0.375.3±0.172.7±0.656.4±0.3
ELR+94.9±0.293.6±0.190.4±0.275.5±0.271.0±0.250.4±0.8
Ours (Semi)95.9±0.195.1±0.293.1±0.277.4±0.374.3±0.661.6±0.6
", + "type": "table", + "image_path": "db3492c5afea51549904ef40bb3ab4ac7a34d054f0b1e535e19df964e1db2048.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 131, + 367, + 480, + 403.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 131, + 403.0, + 480, + 439.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 131, + 439.0, + 480, + 475.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 485, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "score": 1.0, + "content": "label precision, and label recall. Here, label precision indicates the ratio of the number of extracted", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "confident examples with correct labels in the total confident example set, and label recall represents", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "the ratio of the number of confident examples with correct labels among the total correctly labeled", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 518, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 530 + ], + "score": 1.0, + "content": "examples. Specifically, we train a neural network on CIFAR-10 with different kinds and levels of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "label noise for 25 epochs respectively and report the performance for each case before and after the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 540, + 209, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 209, + 552 + ], + "score": 1.0, + "content": "proposed PES is applied.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 569 + ], + "score": 1.0, + "content": "Results in Table 1 clearly show that, compared with the traditional early stopping, PES can help to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 567, + 507, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 507, + 580 + ], + "score": 1.0, + "content": "obtain higher accuracies, precisions, and recalls for most cases. For instance-dependent label noise,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "PES can achieve higher recall values with comparable label precision values. Note that models with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "high recall values can help to collect more confident examples, which is critical for learning with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "confident examples and semi-supervised learning. Therefore, by enhancing the performance of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "initial model, PES can help to improve the final classification performance in all cases, which is also", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 623, + 277, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 277, + 634 + ], + "score": 1.0, + "content": "verified by the experiments in Section 3.3.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 280, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 280, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 280, + 660 + ], + "score": 1.0, + "content": "3.3 Classification Accuracy Evaluation", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "Synthetic datasets. We first verify the effectiveness of our proposed method without semi-supervised", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "learning techniques on two synthetic datasets: CIFAR-10 and CIFAR-100. For both of these two", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 183, + 701 + ], + "score": 1.0, + "content": "datasets, we leave", + "type": "text" + }, + { + "bbox": [ + 183, + 689, + 203, + 699 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "of data with noisy labels as noisy validation set. Results are presented in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "Table 2, which shows that our proposed method can consistently outperform all other baselines across", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 712, + 244, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 244, + 723 + ], + "score": 1.0, + "content": "various settings by a large margin.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 104, + 492, + 317 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 77, + 505, + 100 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 78, + 505, + 89 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 505, + 89 + ], + "score": 1.0, + "content": "Table 2: Comparison with state-of-the-art methods without semi-supervised learning on CIFAR-10", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 87, + 466, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 87, + 466, + 101 + ], + "score": 1.0, + "content": "and CIFAR-100. The mean and standard deviation computed over five runs are presented.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 116, + 104, + 492, + 317 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 104, + 492, + 317 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 492, + 317 + ], + "score": 0.986, + "html": "
DatasetMethodSymmetricPairflipInstance
20%50%45%20%40%
CIFAR10CE84.00±0.6675.51±1.2463.34±6.0385.10±0.6877.00±2.17
Co-teaching87.16±0.1172.80±0.4570.11±1.1686.54±0.1180.98±0.39
Forward85.63±0.5277.92±0.6660.15±1.9785.29±0.3874.72±3.24
Joint Optim89.70±0.1185.00±0.1782.63±1.3889.69±0.4282.62±0.57
T-revision89.63±0.1383.40±0.6577.06±6.4790.46±0.1385.37±3.36
DMI88.18±0.3678.28±0.4857.60±14.5689.14±0.3684.78±1.97
CDR89.72±0.3882.64±0.8973.67±0.5490.41±0.3483.07±1.33
Ours92.38±0.4087.45±0.3588.43±1.0892.69±0.4489.73±0.51
CIFAR100CE51.43±0.5837.69±3.4534.10±2.0452.19±1.4242.26±1.29
Co-teaching59.28±0.4741.37±0.0833.22±0.4857.24±0.6945.69±0.99
Forward57.75±0.3744.66±1.0127.88±0.8058.76±0.6644.50±0.72
Joint Optim64.55±0.3850.22±0.4142.61±0.6165.15±0.3155.57±0.41
T-revision65.40±1.0750.24±1.4541.10±1.9560.71±0.7351.54±0.91
DMI58.73±0.7044.25±1.1426.90±0.4558.05±0.2047.36±0.68
CDR66.52±0.2455.30±0.9643.87±1.3567.33±0.6755.94±0.56
Ours68.89±0.4558.90±2.7257.18±1.4470.49±0.7965.68±1.41
", + "type": "table", + "image_path": "faf2277129f4013a50d7d87c0c93cc22f297e8287191fc976ac1e57bb5e493c0.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 116, + 104, + 492, + 175.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 116, + 175.0, + 492, + 246.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 116, + 246.0, + 492, + 317.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 131, + 367, + 480, + 475 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 329, + 505, + 363 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "Table 3: Comparison with state-of-the-art methods with semi-supervised learning on CIFAR-10 and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "CIFAR-100 with symmetric label noise from different levels. Results with * are token from [15]. The", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 351, + 344, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 344, + 364 + ], + "score": 1.0, + "content": "mean and standard deviation are computed over three runs.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 131, + 367, + 480, + 475 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 131, + 367, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 367, + 480, + 475 + ], + "score": 0.984, + "html": "
DatasetCIFAR-10CIFAR-100
Methods /NoiseSym-20%Sym-50%Sym-80%Sym-20%Sym-50%Sym-80%
CE86.5±0.680.6±0.263.7±0.857.9±0.447.3±0.222.3±1.2
MixUp93.2±0.388.2±0.373.3±0.369.5±0.257.1±0.634.1±0.6
M-correction*94.092.086.873.966.148.2
DivideMix*95.294.293.075.272.858.3
DivideMix95.6±0.194.6±0.192.9±0.375.3±0.172.7±0.656.4±0.3
ELR+94.9±0.293.6±0.190.4±0.275.5±0.271.0±0.250.4±0.8
Ours (Semi)95.9±0.195.1±0.293.1±0.277.4±0.374.3±0.661.6±0.6
", + "type": "table", + "image_path": "db3492c5afea51549904ef40bb3ab4ac7a34d054f0b1e535e19df964e1db2048.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 131, + 367, + 480, + 403.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 131, + 403.0, + 480, + 439.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 131, + 439.0, + 480, + 475.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 485, + 505, + 551 + ], + "lines": [], + "index": 13.5, + "bbox_fs": [ + 104, + 486, + 506, + 552 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 556, + 505, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 569 + ], + "score": 1.0, + "content": "Results in Table 1 clearly show that, compared with the traditional early stopping, PES can help to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 567, + 507, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 507, + 580 + ], + "score": 1.0, + "content": "obtain higher accuracies, precisions, and recalls for most cases. For instance-dependent label noise,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "PES can achieve higher recall values with comparable label precision values. Note that models with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "high recall values can help to collect more confident examples, which is critical for learning with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "confident examples and semi-supervised learning. Therefore, by enhancing the performance of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "initial model, PES can help to improve the final classification performance in all cases, which is also", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 623, + 277, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 277, + 634 + ], + "score": 1.0, + "content": "verified by the experiments in Section 3.3.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 557, + 507, + 634 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 280, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 280, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 280, + 660 + ], + "score": 1.0, + "content": "3.3 Classification Accuracy Evaluation", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "Synthetic datasets. We first verify the effectiveness of our proposed method without semi-supervised", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "learning techniques on two synthetic datasets: CIFAR-10 and CIFAR-100. For both of these two", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 183, + 701 + ], + "score": 1.0, + "content": "datasets, we leave", + "type": "text" + }, + { + "bbox": [ + 183, + 689, + 203, + 699 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "of data with noisy labels as noisy validation set. Results are presented in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "Table 2, which shows that our proposed method can consistently outperform all other baselines across", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 712, + 244, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 244, + 723 + ], + "score": 1.0, + "content": "various settings by a large margin.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 667, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 115, + 473, + 200 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 77, + 505, + 110 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 90 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 90 + ], + "score": 1.0, + "content": "Table 4: Comparison with state-of-the-art methods with semi-supervised learning on CIFAR-10 and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 88, + 505, + 100 + ], + "spans": [ + { + "bbox": [ + 106, + 88, + 505, + 100 + ], + "score": 1.0, + "content": "CIFAR-100 with instance-dependent and pairflip label noise from different levels. The mean and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 100, + 304, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 304, + 111 + ], + "score": 1.0, + "content": "standard deviation are computed over three runs.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 136, + 115, + 473, + 200 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 115, + 473, + 200 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 473, + 200 + ], + "score": 0.98, + "html": "
DatasetCIFAR-10CIFAR-100
Methods/NoiseInst-20%Inst-40%Pair-45%Inst-20%Inst-40%Pair-45%
CEMixUpDivideMixELR+87.5±0.578.9±0.774.9±1.756.8±0.448.2±0.538.5±0.6
93.3±0.295.5±0.187.6±0.594.5±0.282.4±1.085.6±1.767.1±0.155.0±0.170.9±0.144.2±0.548.2±1.0
75.2±0.270.9±0.1
94.9±0.194.3±0.286.1±1.275.8±0.174.3±0.365.3±1.3
Ours (Semi)95.9±0.195.3±0.194.5±0.377.6±0.376.1±0.473.6±1.7
", + "type": "table", + "image_path": "30f68d18682a41f23ee5c16e39c70d68c6be5ad1521db00b67346216a25d82d5.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 136, + 115, + 473, + 143.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 136, + 143.33333333333334, + 473, + 171.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 136, + 171.66666666666669, + 473, + 200.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 127, + 257, + 484, + 283 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 218, + 505, + 252 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "Table 5: Compassion with state-of-the-art methods on Clothing-1M. Results of baseline methods are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "taken from the original papers. ours represent the results obtained by PES with a single network and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 240, + 380, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 380, + 252 + ], + "score": 1.0, + "content": "ours* indicate the results obtained by PES with an ensemble model.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "table_body", + "bbox": [ + 127, + 257, + 484, + 283 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 127, + 257, + 484, + 283 + ], + "spans": [ + { + "bbox": [ + 127, + 257, + 484, + 283 + ], + "score": 0.96, + "html": "
CEForwardJoint-OptimDMIT-revisionDivideMix*ELR+*OursOurs*
69.2169.8472.1672.4674.1874.7674.8174.6474.99
", + "type": "table", + "image_path": "fd26d18ed5ae3d06898df87d4bef68e8bf94d61f481d809a16bbd2362705f40d.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 127, + 257, + 484, + 265.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 127, + 265.6666666666667, + 484, + 274.33333333333337 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 127, + 274.33333333333337, + 484, + 283.00000000000006 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 8.5 + }, + { + "type": "image", + "bbox": [ + 174, + 306, + 436, + 430 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 174, + 306, + 436, + 430 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 306, + 436, + 430 + ], + "spans": [ + { + "bbox": [ + 174, + 306, + 436, + 430 + ], + "score": 0.97, + "type": "image", + "image_path": "61ae512b2fb46bee07bd5e6157d516e400cd564030c729a91354232a89157ebe.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 174, + 306, + 436, + 347.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 174, + 347.3333333333333, + 436, + 388.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 174, + 388.66666666666663, + 436, + 429.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 142, + 437, + 466, + 449 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 436, + 464, + 451 + ], + "spans": [ + { + "bbox": [ + 143, + 436, + 422, + 451 + ], + "score": 1.0, + "content": "Figure 3: Sensitivity analysis for different training iteration numbers:", + "type": "text" + }, + { + "bbox": [ + 423, + 438, + 434, + 448 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 436, + 452, + 451 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 452, + 438, + 464, + 448 + ], + "score": 0.87, + "content": "T _ { 3 }", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 14.0 + }, + { + "type": "text", + "bbox": [ + 106, + 461, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "Table 3 and Table 4 present the mean accuracy and standard deviation for our method and all baselines", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "score": 1.0, + "content": "on CIFAR-10 and CIFAR-100, respectively. From the results, we can get that the proposed method", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "can outperform all baselines in all cases. For pairflip label noise, the advantages of our proposed", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "method become more apparent, and it significantly outperforms state-of-the-art methods by over", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 121, + 516 + ], + "score": 0.83, + "content": "8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "on both CIFAR-10 and CIFAR-100. These empirical results support our proposal that PES can", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "improve the quality of selected confident examples, which helps improve performance and reduce the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 528, + 227, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 227, + 539 + ], + "score": 1.0, + "content": "variance of the final classifier.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 506, + 556 + ], + "score": 1.0, + "content": "Real-world dataset. We evaluate the performance of the proposed method on a real-world dataset", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 555, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 506, + 567 + ], + "score": 1.0, + "content": "with Clothing-1M [32] and select methods such as CE, Forward, Joint-Optim, DMI, and T-revision,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 376, + 578 + ], + "score": 1.0, + "content": "which use a single network, and also methods such as DivideMix and", + "type": "text" + }, + { + "bbox": [ + 376, + 566, + 401, + 576 + ], + "score": 0.71, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 565, + 505, + 578 + ], + "score": 1.0, + "content": ", which adopt an ensemble", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "model with two different networks, as baselines. We also report the results for the proposed PES", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 588, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 506, + 599 + ], + "score": 1.0, + "content": "with a single network as ours and the results for PES, which ensembles two networks, as ours*. The", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 599, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 506, + 610 + ], + "score": 1.0, + "content": "overall results are reported in Table 5, from which we can observe that the proposed PES with a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "single network can outperform all baselines using a single network. And with an ensemble model,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 621, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 632 + ], + "score": 1.0, + "content": "which contains two different networks, our method can outperform all the adopted baselines. These", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 631, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 644 + ], + "score": 1.0, + "content": "results clearly demonstrate that, by improving the performance of the initial classification network,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 642, + 388, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 388, + 654 + ], + "score": 1.0, + "content": "our method is more flexible to handle such real-world noise problems.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 107, + 668, + 213, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 214, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 214, + 682 + ], + "score": 1.0, + "content": "3.4 Sensitivity Analysis", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 475, + 702 + ], + "score": 1.0, + "content": "In this section, we investigate the hyper-parameter sensitivity for the training iteration number", + "type": "text" + }, + { + "bbox": [ + 476, + 689, + 487, + 700 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 107, + 701, + 118, + 711 + ], + "score": 0.85, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 699, + 505, + 713 + ], + "score": 1.0, + "content": ", respectively. We firstly analyze the training epoch number for the second DNN part by varying", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 712, + 118, + 722 + ], + "score": 0.86, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "from the range of [0, 10]. The results are illustrated in Figure 3a, from which can find that, with", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 115, + 473, + 200 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 77, + 505, + 110 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 90 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 90 + ], + "score": 1.0, + "content": "Table 4: Comparison with state-of-the-art methods with semi-supervised learning on CIFAR-10 and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 88, + 505, + 100 + ], + "spans": [ + { + "bbox": [ + 106, + 88, + 505, + 100 + ], + "score": 1.0, + "content": "CIFAR-100 with instance-dependent and pairflip label noise from different levels. The mean and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 100, + 304, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 304, + 111 + ], + "score": 1.0, + "content": "standard deviation are computed over three runs.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 136, + 115, + 473, + 200 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 115, + 473, + 200 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 473, + 200 + ], + "score": 0.98, + "html": "
DatasetCIFAR-10CIFAR-100
Methods/NoiseInst-20%Inst-40%Pair-45%Inst-20%Inst-40%Pair-45%
CEMixUpDivideMixELR+87.5±0.578.9±0.774.9±1.756.8±0.448.2±0.538.5±0.6
93.3±0.295.5±0.187.6±0.594.5±0.282.4±1.085.6±1.767.1±0.155.0±0.170.9±0.144.2±0.548.2±1.0
75.2±0.270.9±0.1
94.9±0.194.3±0.286.1±1.275.8±0.174.3±0.365.3±1.3
Ours (Semi)95.9±0.195.3±0.194.5±0.377.6±0.376.1±0.473.6±1.7
", + "type": "table", + "image_path": "30f68d18682a41f23ee5c16e39c70d68c6be5ad1521db00b67346216a25d82d5.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 136, + 115, + 473, + 143.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 136, + 143.33333333333334, + 473, + 171.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 136, + 171.66666666666669, + 473, + 200.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 127, + 257, + 484, + 283 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 218, + 505, + 252 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "Table 5: Compassion with state-of-the-art methods on Clothing-1M. Results of baseline methods are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "taken from the original papers. ours represent the results obtained by PES with a single network and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 240, + 380, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 380, + 252 + ], + "score": 1.0, + "content": "ours* indicate the results obtained by PES with an ensemble model.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "table_body", + "bbox": [ + 127, + 257, + 484, + 283 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 127, + 257, + 484, + 283 + ], + "spans": [ + { + "bbox": [ + 127, + 257, + 484, + 283 + ], + "score": 0.96, + "html": "
CEForwardJoint-OptimDMIT-revisionDivideMix*ELR+*OursOurs*
69.2169.8472.1672.4674.1874.7674.8174.6474.99
", + "type": "table", + "image_path": "fd26d18ed5ae3d06898df87d4bef68e8bf94d61f481d809a16bbd2362705f40d.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 127, + 257, + 484, + 265.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 127, + 265.6666666666667, + 484, + 274.33333333333337 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 127, + 274.33333333333337, + 484, + 283.00000000000006 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 8.5 + }, + { + "type": "image", + "bbox": [ + 174, + 306, + 436, + 430 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 174, + 306, + 436, + 430 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 306, + 436, + 430 + ], + "spans": [ + { + "bbox": [ + 174, + 306, + 436, + 430 + ], + "score": 0.97, + "type": "image", + "image_path": "61ae512b2fb46bee07bd5e6157d516e400cd564030c729a91354232a89157ebe.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 174, + 306, + 436, + 347.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 174, + 347.3333333333333, + 436, + 388.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 174, + 388.66666666666663, + 436, + 429.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 142, + 437, + 466, + 449 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 436, + 464, + 451 + ], + "spans": [ + { + "bbox": [ + 143, + 436, + 422, + 451 + ], + "score": 1.0, + "content": "Figure 3: Sensitivity analysis for different training iteration numbers:", + "type": "text" + }, + { + "bbox": [ + 423, + 438, + 434, + 448 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 436, + 452, + 451 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 452, + 438, + 464, + 448 + ], + "score": 0.87, + "content": "T _ { 3 }", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 14.0 + }, + { + "type": "text", + "bbox": [ + 106, + 461, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "Table 3 and Table 4 present the mean accuracy and standard deviation for our method and all baselines", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "score": 1.0, + "content": "on CIFAR-10 and CIFAR-100, respectively. From the results, we can get that the proposed method", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "can outperform all baselines in all cases. For pairflip label noise, the advantages of our proposed", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "method become more apparent, and it significantly outperforms state-of-the-art methods by over", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 121, + 516 + ], + "score": 0.83, + "content": "8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "on both CIFAR-10 and CIFAR-100. These empirical results support our proposal that PES can", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "improve the quality of selected confident examples, which helps improve performance and reduce the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 528, + 227, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 227, + 539 + ], + "score": 1.0, + "content": "variance of the final classifier.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 462, + 506, + 539 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 506, + 556 + ], + "score": 1.0, + "content": "Real-world dataset. We evaluate the performance of the proposed method on a real-world dataset", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 555, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 506, + 567 + ], + "score": 1.0, + "content": "with Clothing-1M [32] and select methods such as CE, Forward, Joint-Optim, DMI, and T-revision,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 376, + 578 + ], + "score": 1.0, + "content": "which use a single network, and also methods such as DivideMix and", + "type": "text" + }, + { + "bbox": [ + 376, + 566, + 401, + 576 + ], + "score": 0.71, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 565, + 505, + 578 + ], + "score": 1.0, + "content": ", which adopt an ensemble", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "model with two different networks, as baselines. We also report the results for the proposed PES", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 588, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 506, + 599 + ], + "score": 1.0, + "content": "with a single network as ours and the results for PES, which ensembles two networks, as ours*. The", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 599, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 506, + 610 + ], + "score": 1.0, + "content": "overall results are reported in Table 5, from which we can observe that the proposed PES with a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "single network can outperform all baselines using a single network. And with an ensemble model,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 621, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 505, + 632 + ], + "score": 1.0, + "content": "which contains two different networks, our method can outperform all the adopted baselines. These", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 631, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 644 + ], + "score": 1.0, + "content": "results clearly demonstrate that, by improving the performance of the initial classification network,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 642, + 388, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 388, + 654 + ], + "score": 1.0, + "content": "our method is more flexible to handle such real-world noise problems.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 543, + 506, + 654 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 668, + 213, + 680 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 214, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 214, + 682 + ], + "score": 1.0, + "content": "3.4 Sensitivity Analysis", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 475, + 702 + ], + "score": 1.0, + "content": "In this section, we investigate the hyper-parameter sensitivity for the training iteration number", + "type": "text" + }, + { + "bbox": [ + 476, + 689, + 487, + 700 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 107, + 701, + 118, + 711 + ], + "score": 0.85, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 699, + 505, + 713 + ], + "score": 1.0, + "content": ", respectively. We firstly analyze the training epoch number for the second DNN part by varying", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 712, + 118, + 722 + ], + "score": 0.86, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "from the range of [0, 10]. The results are illustrated in Figure 3a, from which can find that, with", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 175, + 142 + ], + "score": 1.0, + "content": "the increasing of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 176, + 129, + 187, + 140 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 187, + 128, + 505, + 142 + ], + "score": 1.0, + "content": ", the performance of PES first increase and then decrease in all the cases except", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 121, + 152 + ], + "score": 1.0, + "content": "for", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 121, + 140, + 141, + 150 + ], + "score": 0.85, + "content": "45 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 141, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "Pairflip noise on the CIFAR-10 dataset. While the model achieves the best performance", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 127, + 163 + ], + "score": 1.0, + "content": "with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 127, + 151, + 138, + 162 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 138, + 150, + 326, + 163 + ], + "score": 1.0, + "content": "as 7 for all types of noisy labels. Then we fix", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 326, + 151, + 338, + 162 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 338, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "as 7, and analyze the impact of the third", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 192, + 175 + ], + "score": 1.0, + "content": "DNN part by varying", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 193, + 162, + 205, + 172 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 205, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "from the range of [0, 10]. The results are shown in Figure 3b. Although the", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 248, + 185 + ], + "score": 1.0, + "content": "performance variance for different", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 249, + 173, + 260, + 183 + ], + "score": 0.87, + "content": "T _ { 3 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 261, + 172, + 355, + 185 + ], + "score": 1.0, + "content": "is smaller than that for", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 355, + 173, + 367, + 183 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 367, + 172, + 506, + 185 + ], + "score": 1.0, + "content": ", we can still observe that the best", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 248, + 196 + ], + "score": 1.0, + "content": "performance can be obtained when", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 248, + 184, + 260, + 194 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 260, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "is set as 5. More importantly, from these two figures, we can", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 194, + 391, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 158, + 207 + ], + "score": 1.0, + "content": "get that both", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 159, + 195, + 170, + 205 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 171, + 194, + 188, + 207 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 189, + 195, + 200, + 205 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 200, + 194, + 391, + 207 + ], + "score": 1.0, + "content": "are robust to the different types of noisy labels.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 688, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 93, + 467, + 120 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 109, + 77, + 497, + 89 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 76, + 499, + 90 + ], + "spans": [ + { + "bbox": [ + 112, + 76, + 383, + 90 + ], + "score": 1.0, + "content": "Table 6: Training time comparison for baselines on CIFAR-10 with", + "type": "text" + }, + { + "bbox": [ + 383, + 78, + 403, + 88 + ], + "score": 0.83, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 76, + 499, + 90 + ], + "score": 1.0, + "content": "Symmetric label noise.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 138, + 93, + 467, + 120 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 93, + 467, + 120 + ], + "spans": [ + { + "bbox": [ + 138, + 93, + 467, + 120 + ], + "score": 0.96, + "html": "
CECo-teachingCDRT-revisionELR+DivideMixOursOurs (Semi)
0.9h1.5h3.0h3.5h2.2h5.5h1.0h3.1h
", + "type": "table", + "image_path": "1be43f84b04f81deb931ee482ba81d70209b96391a05c3c3cca2b69977f03a7a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 138, + 93, + 467, + 102.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 138, + 102.0, + 467, + 111.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 138, + 111.0, + 467, + 120.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 128, + 505, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 128, + 505, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 175, + 142 + ], + "score": 1.0, + "content": "the increasing of", + "type": "text" + }, + { + "bbox": [ + 176, + 129, + 187, + 140 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 128, + 505, + 142 + ], + "score": 1.0, + "content": ", the performance of PES first increase and then decrease in all the cases except", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 121, + 152 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 140, + 141, + 150 + ], + "score": 0.85, + "content": "45 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "Pairflip noise on the CIFAR-10 dataset. While the model achieves the best performance", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 127, + 163 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 151, + 138, + 162 + ], + "score": 0.87, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 150, + 326, + 163 + ], + "score": 1.0, + "content": "as 7 for all types of noisy labels. Then we fix", + "type": "text" + }, + { + "bbox": [ + 326, + 151, + 338, + 162 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "as 7, and analyze the impact of the third", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 192, + 175 + ], + "score": 1.0, + "content": "DNN part by varying", + "type": "text" + }, + { + "bbox": [ + 193, + 162, + 205, + 172 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "from the range of [0, 10]. The results are shown in Figure 3b. Although the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 248, + 185 + ], + "score": 1.0, + "content": "performance variance for different", + "type": "text" + }, + { + "bbox": [ + 249, + 173, + 260, + 183 + ], + "score": 0.87, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 172, + 355, + 185 + ], + "score": 1.0, + "content": "is smaller than that for", + "type": "text" + }, + { + "bbox": [ + 355, + 173, + 367, + 183 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 172, + 506, + 185 + ], + "score": 1.0, + "content": ", we can still observe that the best", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 248, + 196 + ], + "score": 1.0, + "content": "performance can be obtained when", + "type": "text" + }, + { + "bbox": [ + 248, + 184, + 260, + 194 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "is set as 5. More importantly, from these two figures, we can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 194, + 391, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 158, + 207 + ], + "score": 1.0, + "content": "get that both", + "type": "text" + }, + { + "bbox": [ + 159, + 195, + 170, + 205 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 194, + 188, + 207 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 189, + 195, + 200, + 205 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 194, + 391, + 207 + ], + "score": 1.0, + "content": "are robust to the different types of noisy labels.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 218, + 246, + 230 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 248, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 248, + 232 + ], + "score": 1.0, + "content": "3.5 Training Time Comparison", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 238, + 505, + 315 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "score": 1.0, + "content": "In this section, we compare the training time of our method and other state-of-the-art baselines. All", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "the experiments are conducted on a server with a single Nvidia V100 GPU. The training times for all", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 274 + ], + "score": 1.0, + "content": "the methods are reported in Table 6, from which we can get that our algorithm with cross-entropy", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "loss achieves the fastest speed across all baselines, only about 1 hour. Our method combining with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "MixMatch [4] is also fast, only a little more than half of the training time of DivideMix. The time", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 118, + 306 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 293, + 145, + 303 + ], + "score": 0.65, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 292, + 248, + 306 + ], + "score": 1.0, + "content": "[16] shows superior, but", + "type": "text" + }, + { + "bbox": [ + 248, + 293, + 275, + 303 + ], + "score": 0.62, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "trains the network with fewer epochs, with 200 epochs", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 304, + 252, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 252, + 316 + ], + "score": 1.0, + "content": "compared with ours for 300 epochs.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 107, + 330, + 194, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 196, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 196, + 345 + ], + "score": 1.0, + "content": "4 Related work", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "Learning with noisy data has been well studied [17, 6, 21, 27, 20]. Current works can be mainly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "categorized into two groups: model-based and model-free methods. In this section, we briefly review", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 376, + 220, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 220, + 388 + ], + "score": 1.0, + "content": "some closely related works.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "score": 1.0, + "content": "The first type models the relationship between clean labels and noisy labels by estimating the noise", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "transition matrix and build a loss function to correct the loss [24, 30, 35, 28]. [24] first combines", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "algorithms for estimating the noise rates and loss correction techniques together and introduces two", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "alternative procedures for loss correction. It also proves that both of the two procedures enjoy formal", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "robustness guarantees w.r.t. the clean data distribution. DMI [34] proposes an information-theoretic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "score": 1.0, + "content": "loss function, which utilizes Shannon’s mutual information and is robustness to different kinds of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "label noise. T-revision [31] estimates the noise transition matrix without anchor points by adding", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "a fine-tuned slack variables. Although these methods have made certain progress, they are usually", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "score": 1.0, + "content": "fragile to estimate the noise transition matrix for heavy noisy data and are also hard to handle a large", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 491, + 462, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 462, + 503 + ], + "score": 1.0, + "content": "number of classes. Therefore, in this paper, we mainly focus on the model-free methods.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 507, + 506, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "score": 1.0, + "content": "The second strand mainly counteracts noisy labels by exploiting the memorization effect that deep", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 517, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 104, + 517, + 506, + 532 + ], + "score": 1.0, + "content": "networks tend to first memorize and fit majority (clean) patterns and then overfit minority (noisy)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 528, + 507, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 507, + 542 + ], + "score": 1.0, + "content": "patterns [2]. To exploit this property, Co-teaching [9] employs two networks with different initializa-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 541, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 552 + ], + "score": 1.0, + "content": "tion and uses small loss to select confident examples. M-correction [1] uses two Gaussian Mixture", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "Models to identify confident examples, instead of using networks themselves. DivideMix [15] extends", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 575 + ], + "score": 1.0, + "content": "Co-teaching [9] and employs two Beta Mixture Model to select confident examples. MixMatch [4]", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "is then adopted to leverage unconfident examples with a semi-supervised learning framework. All", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "the above methods exploit the memorization effect by considering the adopted network as a whole.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "Recently, [14] shows that networks training with noisy labels can produce good representations, if", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "the structure of networks suits the targeted tasks. Our method further explains that noisy labels have", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "different impacts for different layers in a DNN. And latter layers will receive earlier and more severe", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "impact than their former counterparts. Therefore, by considering a DNN as a composition of several", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "layers and training different layers with different epochs, our method is able to better exploit the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 649, + 329, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 329, + 662 + ], + "score": 1.0, + "content": "memorization effect and achieve superior performance.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 107, + 675, + 183, + 689 + ], + "lines": [ + { + "bbox": [ + 104, + 673, + 185, + 691 + ], + "spans": [ + { + "bbox": [ + 104, + 673, + 185, + 691 + ], + "score": 1.0, + "content": "5 Conclusion", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "In this work, we provide a progressive early stopping (PES) method to better exploit the memorization", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 710, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 504, + 723 + ], + "score": 1.0, + "content": "effect of deep neural networks (DNN) for noisy-label learning. We first find that the impact of noisy", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 93, + 467, + 120 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 109, + 77, + 497, + 89 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 76, + 499, + 90 + ], + "spans": [ + { + "bbox": [ + 112, + 76, + 383, + 90 + ], + "score": 1.0, + "content": "Table 6: Training time comparison for baselines on CIFAR-10 with", + "type": "text" + }, + { + "bbox": [ + 383, + 78, + 403, + 88 + ], + "score": 0.83, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 76, + 499, + 90 + ], + "score": 1.0, + "content": "Symmetric label noise.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 138, + 93, + 467, + 120 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 93, + 467, + 120 + ], + "spans": [ + { + "bbox": [ + 138, + 93, + 467, + 120 + ], + "score": 0.96, + "html": "
CECo-teachingCDRT-revisionELR+DivideMixOursOurs (Semi)
0.9h1.5h3.0h3.5h2.2h5.5h1.0h3.1h
", + "type": "table", + "image_path": "1be43f84b04f81deb931ee482ba81d70209b96391a05c3c3cca2b69977f03a7a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 138, + 93, + 467, + 102.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 138, + 102.0, + 467, + 111.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 138, + 111.0, + 467, + 120.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 128, + 505, + 206 + ], + "lines": [], + "index": 7, + "bbox_fs": [ + 105, + 128, + 506, + 207 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 218, + 246, + 230 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 248, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 248, + 232 + ], + "score": 1.0, + "content": "3.5 Training Time Comparison", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 238, + 505, + 315 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 251 + ], + "score": 1.0, + "content": "In this section, we compare the training time of our method and other state-of-the-art baselines. All", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "the experiments are conducted on a server with a single Nvidia V100 GPU. The training times for all", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 274 + ], + "score": 1.0, + "content": "the methods are reported in Table 6, from which we can get that our algorithm with cross-entropy", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "loss achieves the fastest speed across all baselines, only about 1 hour. Our method combining with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "MixMatch [4] is also fast, only a little more than half of the training time of DivideMix. The time", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 292, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 118, + 306 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 293, + 145, + 303 + ], + "score": 0.65, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 292, + 248, + 306 + ], + "score": 1.0, + "content": "[16] shows superior, but", + "type": "text" + }, + { + "bbox": [ + 248, + 293, + 275, + 303 + ], + "score": 0.62, + "content": "\\mathrm { E L R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 292, + 505, + 306 + ], + "score": 1.0, + "content": "trains the network with fewer epochs, with 200 epochs", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 304, + 252, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 252, + 316 + ], + "score": 1.0, + "content": "compared with ours for 300 epochs.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 237, + 506, + 316 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 330, + 194, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 196, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 196, + 345 + ], + "score": 1.0, + "content": "4 Related work", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "Learning with noisy data has been well studied [17, 6, 21, 27, 20]. Current works can be mainly", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "categorized into two groups: model-based and model-free methods. In this section, we briefly review", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 376, + 220, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 220, + 388 + ], + "score": 1.0, + "content": "some closely related works.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 353, + 505, + 388 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "score": 1.0, + "content": "The first type models the relationship between clean labels and noisy labels by estimating the noise", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "transition matrix and build a loss function to correct the loss [24, 30, 35, 28]. [24] first combines", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "algorithms for estimating the noise rates and loss correction techniques together and introduces two", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "alternative procedures for loss correction. It also proves that both of the two procedures enjoy formal", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 505, + 449 + ], + "score": 1.0, + "content": "robustness guarantees w.r.t. the clean data distribution. DMI [34] proposes an information-theoretic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "score": 1.0, + "content": "loss function, which utilizes Shannon’s mutual information and is robustness to different kinds of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "label noise. T-revision [31] estimates the noise transition matrix without anchor points by adding", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "a fine-tuned slack variables. Although these methods have made certain progress, they are usually", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "score": 1.0, + "content": "fragile to estimate the noise transition matrix for heavy noisy data and are also hard to handle a large", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 491, + 462, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 462, + 503 + ], + "score": 1.0, + "content": "number of classes. Therefore, in this paper, we mainly focus on the model-free methods.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 392, + 506, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 507, + 506, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "score": 1.0, + "content": "The second strand mainly counteracts noisy labels by exploiting the memorization effect that deep", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 517, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 104, + 517, + 506, + 532 + ], + "score": 1.0, + "content": "networks tend to first memorize and fit majority (clean) patterns and then overfit minority (noisy)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 528, + 507, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 507, + 542 + ], + "score": 1.0, + "content": "patterns [2]. To exploit this property, Co-teaching [9] employs two networks with different initializa-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 541, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 552 + ], + "score": 1.0, + "content": "tion and uses small loss to select confident examples. M-correction [1] uses two Gaussian Mixture", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "Models to identify confident examples, instead of using networks themselves. DivideMix [15] extends", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 575 + ], + "score": 1.0, + "content": "Co-teaching [9] and employs two Beta Mixture Model to select confident examples. MixMatch [4]", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "is then adopted to leverage unconfident examples with a semi-supervised learning framework. All", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "the above methods exploit the memorization effect by considering the adopted network as a whole.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "Recently, [14] shows that networks training with noisy labels can produce good representations, if", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "the structure of networks suits the targeted tasks. Our method further explains that noisy labels have", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "different impacts for different layers in a DNN. And latter layers will receive earlier and more severe", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "impact than their former counterparts. Therefore, by considering a DNN as a composition of several", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "layers and training different layers with different epochs, our method is able to better exploit the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 649, + 329, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 329, + 662 + ], + "score": 1.0, + "content": "memorization effect and achieve superior performance.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 506, + 507, + 662 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 675, + 183, + 689 + ], + "lines": [ + { + "bbox": [ + 104, + 673, + 185, + 691 + ], + "spans": [ + { + "bbox": [ + 104, + 673, + 185, + 691 + ], + "score": 1.0, + "content": "5 Conclusion", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "In this work, we provide a progressive early stopping (PES) method to better exploit the memorization", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 710, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 504, + 723 + ], + "score": 1.0, + "content": "effect of deep neural networks (DNN) for noisy-label learning. We first find that the impact of noisy", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "labels for former layers in a DNN is much less and later than that for latter DNN layers, and then", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "build upon this insight to propose the PES method, which separates a DNN into different parts and", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "progressively train each part to counteract the different impacts of noisy labels for different DNN", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "layers. To show that PES can boost the performance of state-of-the-art methods, we conduct extensive", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 117, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 128 + ], + "score": 1.0, + "content": "experiments across multiple synthetic and real-world noisy datasets and demonstrate that the proposed", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "PES can help to obtain substantial performance improvements compared to current state-of-the-art", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "baselines. The main limitation of our method lies in that, by splitting a DNN into different parts, PES", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "introduces several additional hyper-parameters that need to be tuned carefully. In the future, we will", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 158, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 158, + 506, + 173 + ], + "score": 1.0, + "content": "extend the work in the following aspects. First, we will study other mechanisms that distinguishing", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 169, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 184 + ], + "score": 1.0, + "content": "desired and undesired memorization rather than early stopping, e.g., the gradient ascent trick [8].", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 180, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 196 + ], + "score": 1.0, + "content": "Second, we are interested in combining PES with interesting ideas from semi-supervised learning", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 192, + 216, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 216, + 206 + ], + "score": 1.0, + "content": "and unsupervised learning.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 699, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "labels for former layers in a DNN is much less and later than that for latter DNN layers, and then", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "build upon this insight to propose the PES method, which separates a DNN into different parts and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "progressively train each part to counteract the different impacts of noisy labels for different DNN", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "layers. To show that PES can boost the performance of state-of-the-art methods, we conduct extensive", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 117, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 128 + ], + "score": 1.0, + "content": "experiments across multiple synthetic and real-world noisy datasets and demonstrate that the proposed", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "PES can help to obtain substantial performance improvements compared to current state-of-the-art", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "baselines. The main limitation of our method lies in that, by splitting a DNN into different parts, PES", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "introduces several additional hyper-parameters that need to be tuned carefully. In the future, we will", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 158, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 158, + 506, + 173 + ], + "score": 1.0, + "content": "extend the work in the following aspects. First, we will study other mechanisms that distinguishing", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 169, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 506, + 184 + ], + "score": 1.0, + "content": "desired and undesired memorization rather than early stopping, e.g., the gradient ascent trick [8].", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 180, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 506, + 196 + ], + "score": 1.0, + "content": "Second, we are interested in combining PES with interesting ideas from semi-supervised learning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 192, + 216, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 216, + 206 + ], + "score": 1.0, + "content": "and unsupervised learning.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 107, + 219, + 339, + 233 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 341, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 341, + 237 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 244, + 505, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "YB was partially supported by Agriculture Consultant and Smart Management. BH was supported", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "by the RGC Early Career Scheme No. 22200720, NSFC Young Scientists Fund No. 62006202", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "score": 1.0, + "content": "and HKBU CSD Departmental Incentive Grant. YY was partially supported by Key Research and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "Development Program of Shaanxi (ProgramNo. 2021ZDLGY01-03). GN was supported by JST", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 286, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 104, + 286, + 505, + 302 + ], + "score": 1.0, + "content": "AIP Acceleration Research Grant Number JPMJCR20U3, Japan. TL was partially supported by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 299, + 402, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 402, + 312 + ], + "score": 1.0, + "content": "Australian Research Council Projects DE-190101473 and IC-190100031.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 742, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 204 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 104, + 73, + 506, + 206 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 219, + 339, + 233 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 341, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 341, + 237 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 244, + 505, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "YB was partially supported by Agriculture Consultant and Smart Management. BH was supported", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "by the RGC Early Career Scheme No. 22200720, NSFC Young Scientists Fund No. 62006202", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "score": 1.0, + "content": "and HKBU CSD Departmental Incentive Grant. YY was partially supported by Key Research and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "Development Program of Shaanxi (ProgramNo. 2021ZDLGY01-03). GN was supported by JST", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 286, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 104, + 286, + 505, + 302 + ], + "score": 1.0, + "content": "AIP Acceleration Research Grant Number JPMJCR20U3, Japan. TL was partially supported by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 299, + 402, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 402, + 312 + ], + "score": 1.0, + "content": "Australian Research Council Projects DE-190101473 and IC-190100031.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 104, + 243, + 506, + 312 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 70, + 507, + 727 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 109, + 88, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 109, + 88, + 506, + 102 + ], + "score": 1.0, + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O’Connor, and Kevin McGuinness. Unsupervised label", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 99, + 374, + 113 + ], + "spans": [ + { + "bbox": [ + 125, + 99, + 374, + 113 + ], + "score": 1.0, + "content": "noise modeling and loss correction. In ICML, pages 312–321, 2019.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 117, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 109, + 117, + 507, + 131 + ], + "score": 1.0, + "content": "[2] Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 128, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 126, + 128, + 506, + 139 + ], + "score": 1.0, + "content": "Kanwal, Tegan Maharaj, Asja Fischer, Aaron C. Courville, Yoshua Bengio, and Simon Lacoste-Julien. A", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 136, + 414, + 149 + ], + "spans": [ + { + "bbox": [ + 124, + 136, + 414, + 149 + ], + "score": 1.0, + "content": "closer look at memorization in deep networks. In ICML, pages 233–242, 2017.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 154, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 109, + 154, + 506, + 168 + ], + "score": 1.0, + "content": "[3] Mikhail Belkin, Daniel Hsu, Siyuan Ma, and Soumik Mandal. Reconciling modern machine-learning", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 164, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 125, + 164, + 507, + 178 + ], + "score": 1.0, + "content": "practice and the classical bias–variance trade-off. Proceedings of the National Academy of Sciences,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 174, + 234, + 186 + ], + "spans": [ + { + "bbox": [ + 126, + 174, + 234, + 186 + ], + "score": 1.0, + "content": "116(32):15849–15854, 2019.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 109, + 192, + 507, + 205 + ], + "spans": [ + { + "bbox": [ + 109, + 192, + 507, + 205 + ], + "score": 1.0, + "content": "[4] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 202, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 125, + 202, + 481, + 216 + ], + "score": 1.0, + "content": "Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, pages 5050–5060, 2019.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 109, + 220, + 507, + 234 + ], + "spans": [ + { + "bbox": [ + 109, + 220, + 507, + 234 + ], + "score": 1.0, + "content": "[5] Youngchul Cha and Junghoo Cho. Social-network analysis using topic models. In SIGIR, pages 565–574,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 229, + 150, + 242 + ], + "spans": [ + { + "bbox": [ + 126, + 229, + 150, + 242 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 108, + 247, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 108, + 247, + 506, + 261 + ], + "score": 1.0, + "content": "[6] Jacob Goldberger and Ehud Ben-Reuven. Training deep neural-networks using a noise adaptation layer. In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 257, + 175, + 271 + ], + "spans": [ + { + "bbox": [ + 125, + 257, + 175, + 271 + ], + "score": 1.0, + "content": "ICLR, 2017.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 109, + 275, + 507, + 289 + ], + "spans": [ + { + "bbox": [ + 109, + 275, + 507, + 289 + ], + "score": 1.0, + "content": "[7] Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning. The MIT Press,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 126, + 286, + 150, + 298 + ], + "spans": [ + { + "bbox": [ + 126, + 286, + 150, + 298 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 109, + 302, + 507, + 316 + ], + "spans": [ + { + "bbox": [ + 109, + 302, + 507, + 316 + ], + "score": 1.0, + "content": "[8] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor W. Tsang, and Masashi Sugiyama. SIGUA:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 313, + 473, + 327 + ], + "spans": [ + { + "bbox": [ + 126, + 313, + 473, + 327 + ], + "score": 1.0, + "content": "forgetting may make learning with noisy labels more robust. In ICML, pages 4006–4016, 2020.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 109, + 330, + 507, + 344 + ], + "spans": [ + { + "bbox": [ + 109, + 330, + 507, + 344 + ], + "score": 1.0, + "content": "[9] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 124, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 124, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 352, + 195, + 363 + ], + "spans": [ + { + "bbox": [ + 126, + 352, + 195, + 363 + ], + "score": 1.0, + "content": "8527–8537, 2018.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 368, + 507, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 507, + 382 + ], + "score": 1.0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 125, + 378, + 246, + 392 + ], + "spans": [ + { + "bbox": [ + 125, + 378, + 246, + 392 + ], + "score": 1.0, + "content": "In CVPR, pages 770–778, 2016.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "score": 1.0, + "content": "[11] Takashi Ishida, Ikko Yamane, Tomoya Sakai, Gang Niu, and Masashi Sugiyama. Do we need zero training", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 125, + 406, + 397, + 420 + ], + "spans": [ + { + "bbox": [ + 125, + 406, + 397, + 420 + ], + "score": 1.0, + "content": "loss after achieving zero training error? In ICML, pages 4604–4614, 2020.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 123, + 435, + 176, + 446 + ], + "spans": [ + { + "bbox": [ + 123, + 435, + 176, + 446 + ], + "score": 1.0, + "content": "report, 2009.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "score": 1.0, + "content": "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 463, + 325, + 474 + ], + "spans": [ + { + "bbox": [ + 126, + 463, + 325, + 474 + ], + "score": 1.0, + "content": "neural networks. In NeurIPS, pages 1097–1105, 2012.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "[14] Jingling Li, Mozhi Zhang, Keyulu Xu, John P. Dickerson, and Jimmy Ba. Noisy labels can induce good", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 124, + 489, + 333, + 502 + ], + "spans": [ + { + "bbox": [ + 124, + 489, + 333, + 502 + ], + "score": 1.0, + "content": "representations. arXiv preprint arXiv:2012.12896, 2020.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "[15] Junnan Li, Richard Socher, and Steven C. H. Hoi. Dividemix: Learning with noisy labels as semi-supervised", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 125, + 516, + 218, + 530 + ], + "spans": [ + { + "bbox": [ + 125, + 516, + 218, + 530 + ], + "score": 1.0, + "content": "learning. In ICLR, 2020.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 536, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 547 + ], + "score": 1.0, + "content": "[16] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regular-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 546, + 439, + 557 + ], + "spans": [ + { + "bbox": [ + 126, + 546, + 439, + 557 + ], + "score": 1.0, + "content": "ization prevents memorization of noisy labels. In NeurIPS, pages 20331–20342, 2020.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "score": 1.0, + "content": "[17] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 125, + 573, + 420, + 586 + ], + "spans": [ + { + "bbox": [ + 125, + 573, + 420, + 586 + ], + "score": 1.0, + "content": "Transactions on pattern analysis and machine intelligence, 38(3):447–461, 2016.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "[18] Wei Liu, Yu-Gang Jiang, Jiebo Luo, and Shih-Fu Chang. Noise resistant graph ranking for improved web", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 601, + 298, + 613 + ], + "spans": [ + { + "bbox": [ + 126, + 601, + 298, + 613 + ], + "score": 1.0, + "content": "image search. In CVPR, pages 849–856, 2011.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 619, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 630 + ], + "score": 1.0, + "content": "[19] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In ICLR, 2017.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 637, + 507, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 507, + 648 + ], + "score": 1.0, + "content": "[20] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah M. Erfani, and James Bailey. Normal-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 647, + 451, + 658 + ], + "spans": [ + { + "bbox": [ + 126, + 647, + 451, + 658 + ], + "score": 1.0, + "content": "ized loss functions for deep learning with noisy labels. In ICML, pages 6543–6553, 2020.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 664, + 507, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 507, + 676 + ], + "score": 1.0, + "content": "[21] Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah M. Erfani, Shu-Tao Xia, Sudanthi N. R.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 125, + 672, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 125, + 672, + 506, + 687 + ], + "score": 1.0, + "content": "Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In ICML, pages", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 685, + 194, + 695 + ], + "spans": [ + { + "bbox": [ + 126, + 685, + 194, + 695 + ], + "score": 1.0, + "content": "3361–3370, 2018.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 700, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 506, + 716 + ], + "score": 1.0, + "content": "[22] Preetum Nakkiran, Gal Kaplun, Yamini Bansal, Tristan Yang, Boaz Barak, and Ilya Sutskever. Deep", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 125, + 711, + 394, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 711, + 394, + 725 + ], + "score": 1.0, + "content": "double descent: Where bigger models and more data hurt. In ICLR, 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 23.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 70, + 507, + 727 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 88, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 109, + 88, + 506, + 102 + ], + "score": 1.0, + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O’Connor, and Kevin McGuinness. Unsupervised label", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 99, + 374, + 113 + ], + "spans": [ + { + "bbox": [ + 125, + 99, + 374, + 113 + ], + "score": 1.0, + "content": "noise modeling and loss correction. In ICML, pages 312–321, 2019.", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 117, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 109, + 117, + 507, + 131 + ], + "score": 1.0, + "content": "[2] Devansh Arpit, Stanislaw Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S.", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 128, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 126, + 128, + 506, + 139 + ], + "score": 1.0, + "content": "Kanwal, Tegan Maharaj, Asja Fischer, Aaron C. Courville, Yoshua Bengio, and Simon Lacoste-Julien. A", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 136, + 414, + 149 + ], + "spans": [ + { + "bbox": [ + 124, + 136, + 414, + 149 + ], + "score": 1.0, + "content": "closer look at memorization in deep networks. In ICML, pages 233–242, 2017.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 154, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 109, + 154, + 506, + 168 + ], + "score": 1.0, + "content": "[3] Mikhail Belkin, Daniel Hsu, Siyuan Ma, and Soumik Mandal. Reconciling modern machine-learning", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 164, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 125, + 164, + 507, + 178 + ], + "score": 1.0, + "content": "practice and the classical bias–variance trade-off. Proceedings of the National Academy of Sciences,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 174, + 234, + 186 + ], + "spans": [ + { + "bbox": [ + 126, + 174, + 234, + 186 + ], + "score": 1.0, + "content": "116(32):15849–15854, 2019.", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 192, + 507, + 205 + ], + "spans": [ + { + "bbox": [ + 109, + 192, + 507, + 205 + ], + "score": 1.0, + "content": "[4] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel.", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 202, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 125, + 202, + 481, + 216 + ], + "score": 1.0, + "content": "Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, pages 5050–5060, 2019.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 220, + 507, + 234 + ], + "spans": [ + { + "bbox": [ + 109, + 220, + 507, + 234 + ], + "score": 1.0, + "content": "[5] Youngchul Cha and Junghoo Cho. Social-network analysis using topic models. In SIGIR, pages 565–574,", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 229, + 150, + 242 + ], + "spans": [ + { + "bbox": [ + 126, + 229, + 150, + 242 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 247, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 108, + 247, + 506, + 261 + ], + "score": 1.0, + "content": "[6] Jacob Goldberger and Ehud Ben-Reuven. Training deep neural-networks using a noise adaptation layer. In", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 257, + 175, + 271 + ], + "spans": [ + { + "bbox": [ + 125, + 257, + 175, + 271 + ], + "score": 1.0, + "content": "ICLR, 2017.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 275, + 507, + 289 + ], + "spans": [ + { + "bbox": [ + 109, + 275, + 507, + 289 + ], + "score": 1.0, + "content": "[7] Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning. The MIT Press,", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 286, + 150, + 298 + ], + "spans": [ + { + "bbox": [ + 126, + 286, + 150, + 298 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 302, + 507, + 316 + ], + "spans": [ + { + "bbox": [ + 109, + 302, + 507, + 316 + ], + "score": 1.0, + "content": "[8] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor W. Tsang, and Masashi Sugiyama. SIGUA:", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 313, + 473, + 327 + ], + "spans": [ + { + "bbox": [ + 126, + 313, + 473, + 327 + ], + "score": 1.0, + "content": "forgetting may make learning with noisy labels more robust. In ICML, pages 4006–4016, 2020.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 330, + 507, + 344 + ], + "spans": [ + { + "bbox": [ + 109, + 330, + 507, + 344 + ], + "score": 1.0, + "content": "[9] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama.", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 124, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 352, + 195, + 363 + ], + "spans": [ + { + "bbox": [ + 126, + 352, + 195, + 363 + ], + "score": 1.0, + "content": "8527–8537, 2018.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 368, + 507, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 507, + 382 + ], + "score": 1.0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 378, + 246, + 392 + ], + "spans": [ + { + "bbox": [ + 125, + 378, + 246, + 392 + ], + "score": 1.0, + "content": "In CVPR, pages 770–778, 2016.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 410 + ], + "score": 1.0, + "content": "[11] Takashi Ishida, Ikko Yamane, Tomoya Sakai, Gang Niu, and Masashi Sugiyama. Do we need zero training", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 406, + 397, + 420 + ], + "spans": [ + { + "bbox": [ + 125, + 406, + 397, + 420 + ], + "score": 1.0, + "content": "loss after achieving zero training error? In ICML, pages 4604–4614, 2020.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 123, + 435, + 176, + 446 + ], + "spans": [ + { + "bbox": [ + 123, + 435, + 176, + 446 + ], + "score": 1.0, + "content": "report, 2009.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "score": 1.0, + "content": "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 463, + 325, + 474 + ], + "spans": [ + { + "bbox": [ + 126, + 463, + 325, + 474 + ], + "score": 1.0, + "content": "neural networks. In NeurIPS, pages 1097–1105, 2012.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "[14] Jingling Li, Mozhi Zhang, Keyulu Xu, John P. Dickerson, and Jimmy Ba. Noisy labels can induce good", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 489, + 333, + 502 + ], + "spans": [ + { + "bbox": [ + 124, + 489, + 333, + 502 + ], + "score": 1.0, + "content": "representations. arXiv preprint arXiv:2012.12896, 2020.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "[15] Junnan Li, Richard Socher, and Steven C. H. Hoi. Dividemix: Learning with noisy labels as semi-supervised", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 516, + 218, + 530 + ], + "spans": [ + { + "bbox": [ + 125, + 516, + 218, + 530 + ], + "score": 1.0, + "content": "learning. In ICLR, 2020.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 536, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 506, + 547 + ], + "score": 1.0, + "content": "[16] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regular-", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 546, + 439, + 557 + ], + "spans": [ + { + "bbox": [ + 126, + 546, + 439, + 557 + ], + "score": 1.0, + "content": "ization prevents memorization of noisy labels. In NeurIPS, pages 20331–20342, 2020.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "score": 1.0, + "content": "[17] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 573, + 420, + 586 + ], + "spans": [ + { + "bbox": [ + 125, + 573, + 420, + 586 + ], + "score": 1.0, + "content": "Transactions on pattern analysis and machine intelligence, 38(3):447–461, 2016.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "[18] Wei Liu, Yu-Gang Jiang, Jiebo Luo, and Shih-Fu Chang. Noise resistant graph ranking for improved web", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 601, + 298, + 613 + ], + "spans": [ + { + "bbox": [ + 126, + 601, + 298, + 613 + ], + "score": 1.0, + "content": "image search. In CVPR, pages 849–856, 2011.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 619, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 630 + ], + "score": 1.0, + "content": "[19] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In ICLR, 2017.", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 637, + 507, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 507, + 648 + ], + "score": 1.0, + "content": "[20] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah M. Erfani, and James Bailey. Normal-", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 647, + 451, + 658 + ], + "spans": [ + { + "bbox": [ + 126, + 647, + 451, + 658 + ], + "score": 1.0, + "content": "ized loss functions for deep learning with noisy labels. In ICML, pages 6543–6553, 2020.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 664, + 507, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 507, + 676 + ], + "score": 1.0, + "content": "[21] Xingjun Ma, Yisen Wang, Michael E. Houle, Shuo Zhou, Sarah M. Erfani, Shu-Tao Xia, Sudanthi N. R.", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 672, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 125, + 672, + 506, + 687 + ], + "score": 1.0, + "content": "Wijewickrema, and James Bailey. Dimensionality-driven learning with noisy labels. In ICML, pages", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 685, + 194, + 695 + ], + "spans": [ + { + "bbox": [ + 126, + 685, + 194, + 695 + ], + "score": 1.0, + "content": "3361–3370, 2018.", + "type": "text" + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 700, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 506, + 716 + ], + "score": 1.0, + "content": "[22] Preetum Nakkiran, Gal Kaplun, Yamini Bansal, Tristan Yang, Boaz Barak, and Ilya Sutskever. Deep", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 711, + 394, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 711, + 394, + 725 + ], + "score": 1.0, + "content": "double descent: Where bigger models and more data hurt. In ICLR, 2020.", + "type": "text" + } + ], + "index": 47, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 74, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 106, + 74, + 506, + 87 + ], + "score": 1.0, + "content": "[23] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi-Phuong-Nhung Ngo, Thi Hoai Phuong Nguyen,", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 124, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "Laura Beggel, and Thomas Brox. SELF: learning to filter noisy labels with self-ensembling. In ICLR,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 93, + 150, + 106 + ], + "spans": [ + { + "bbox": [ + 125, + 93, + 150, + 106 + ], + "score": 1.0, + "content": "2020.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 109, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 104, + 109, + 506, + 127 + ], + "score": 1.0, + "content": "[24] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 122, + 492, + 134 + ], + "spans": [ + { + "bbox": [ + 126, + 122, + 492, + 134 + ], + "score": 1.0, + "content": "neural networks robust to label noise: A loss correction approach. In CVPR, pages 2233–2241, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "[25] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 149, + 352, + 162 + ], + "spans": [ + { + "bbox": [ + 126, + 149, + 352, + 162 + ], + "score": 1.0, + "content": "learning with noisy labels. In CVPR, pages 5552–5560, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 166, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 181 + ], + "score": 1.0, + "content": "[26] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 177, + 384, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 177, + 384, + 190 + ], + "score": 1.0, + "content": "learning with open-set noisy labels. In CVPR, pages 8688–8696, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 194, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 505, + 209 + ], + "score": 1.0, + "content": "[27] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 205, + 378, + 218 + ], + "spans": [ + { + "bbox": [ + 126, + 205, + 378, + 218 + ], + "score": 1.0, + "content": "for robust learning with noisy labels. In ICCV, pages 322–330, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 222, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 236 + ], + "score": 1.0, + "content": "[28] Songhua Wu, Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Nannan Wang, Haifeng Liu, and", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 231, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 124, + 231, + 506, + 247 + ], + "score": 1.0, + "content": "Gang Niu. Class2simi: A noise reduction perspective on learning with noisy labels. In ICML, pages", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 242, + 202, + 255 + ], + "spans": [ + { + "bbox": [ + 125, + 242, + 202, + 255 + ], + "score": 1.0, + "content": "11285–11295, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 260, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 275 + ], + "score": 1.0, + "content": "[29] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 271, + 400, + 284 + ], + "spans": [ + { + "bbox": [ + 126, + 271, + 400, + 284 + ], + "score": 1.0, + "content": "early-learning: Hindering the memorization of noisy labels. In ICLR, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "[30] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 297, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 124, + 297, + 506, + 312 + ], + "score": 1.0, + "content": "Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 124, + 308, + 252, + 322 + ], + "spans": [ + { + "bbox": [ + 124, + 308, + 252, + 322 + ], + "score": 1.0, + "content": "NeurIPS, pages 7597–7610, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "[31] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 336, + 474, + 349 + ], + "spans": [ + { + "bbox": [ + 126, + 336, + 474, + 349 + ], + "score": 1.0, + "content": "anchor points really indispensable in label-noise learning? In NeurIPS, pages 6835–6846, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "[32] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 365, + 360, + 378 + ], + "spans": [ + { + "bbox": [ + 126, + 365, + 360, + 378 + ], + "score": 1.0, + "content": "data for image classification. In CVPR, pages 2691–2699, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "[33] Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 392, + 340, + 406 + ], + "spans": [ + { + "bbox": [ + 124, + 392, + 340, + 406 + ], + "score": 1.0, + "content": "consistency training. In NeurIPS, pages 6256–6268, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 409, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 104, + 409, + 506, + 424 + ], + "score": 1.0, + "content": "[34] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L_dmi: A novel information-theoretic loss function", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 420, + 418, + 433 + ], + "spans": [ + { + "bbox": [ + 126, + 420, + 418, + 433 + ], + "score": 1.0, + "content": "for training deep nets robust to label noise. In NeurIPS, pages 6222–6233, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 438, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 452 + ], + "score": 1.0, + "content": "[35] Yu Yao, Tongliang Liu, Bo Han, Mingming Gong, Jiankang Deng, Gang Niu, and Masashi Sugiyama. Dual", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 446, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 124, + 446, + 507, + 462 + ], + "score": 1.0, + "content": "T: reducing estimation error for transition matrix in label-noise learning. In NeurIPS, pages 7260–7271,", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 458, + 150, + 470 + ], + "spans": [ + { + "bbox": [ + 126, + 458, + 150, + 470 + ], + "score": 1.0, + "content": "2020.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 474, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 104, + 474, + 505, + 491 + ], + "score": 1.0, + "content": "[36] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 486, + 342, + 499 + ], + "spans": [ + { + "bbox": [ + 126, + 486, + 342, + 499 + ], + "score": 1.0, + "content": "learning requires rethinking generalization. In ICLR, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 502, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 502, + 506, + 518 + ], + "score": 1.0, + "content": "[37] Hongyi Zhang, Moustapha Cissé, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 515, + 236, + 526 + ], + "spans": [ + { + "bbox": [ + 126, + 515, + 236, + 526 + ], + "score": 1.0, + "content": "minimization. In ICLR, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 70, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 71, + 506, + 531 + ], + "lines": [ + { + "bbox": [ + 106, + 74, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 106, + 74, + 506, + 87 + ], + "score": 1.0, + "content": "[23] Duc Tam Nguyen, Chaithanya Kumar Mummadi, Thi-Phuong-Nhung Ngo, Thi Hoai Phuong Nguyen,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 124, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 124, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "Laura Beggel, and Thomas Brox. SELF: learning to filter noisy labels with self-ensembling. In ICLR,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 93, + 150, + 106 + ], + "spans": [ + { + "bbox": [ + 125, + 93, + 150, + 106 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 109, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 104, + 109, + 506, + 127 + ], + "score": 1.0, + "content": "[24] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 122, + 492, + 134 + ], + "spans": [ + { + "bbox": [ + 126, + 122, + 492, + 134 + ], + "score": 1.0, + "content": "neural networks robust to label noise: A loss correction approach. In CVPR, pages 2233–2241, 2017.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "[25] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 149, + 352, + 162 + ], + "spans": [ + { + "bbox": [ + 126, + 149, + 352, + 162 + ], + "score": 1.0, + "content": "learning with noisy labels. In CVPR, pages 5552–5560, 2018.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 166, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 181 + ], + "score": 1.0, + "content": "[26] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 177, + 384, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 177, + 384, + 190 + ], + "score": 1.0, + "content": "learning with open-set noisy labels. In CVPR, pages 8688–8696, 2018.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 194, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 505, + 209 + ], + "score": 1.0, + "content": "[27] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 205, + 378, + 218 + ], + "spans": [ + { + "bbox": [ + 126, + 205, + 378, + 218 + ], + "score": 1.0, + "content": "for robust learning with noisy labels. In ICCV, pages 322–330, 2019.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 222, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 236 + ], + "score": 1.0, + "content": "[28] Songhua Wu, Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Nannan Wang, Haifeng Liu, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 124, + 231, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 124, + 231, + 506, + 247 + ], + "score": 1.0, + "content": "Gang Niu. Class2simi: A noise reduction perspective on learning with noisy labels. In ICML, pages", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 242, + 202, + 255 + ], + "spans": [ + { + "bbox": [ + 125, + 242, + 202, + 255 + ], + "score": 1.0, + "content": "11285–11295, 2021.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 260, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 275 + ], + "score": 1.0, + "content": "[29] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 271, + 400, + 284 + ], + "spans": [ + { + "bbox": [ + 126, + 271, + 400, + 284 + ], + "score": 1.0, + "content": "early-learning: Hindering the memorization of noisy labels. In ICLR, 2021.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "[30] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 124, + 297, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 124, + 297, + 506, + 312 + ], + "score": 1.0, + "content": "Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 124, + 308, + 252, + 322 + ], + "spans": [ + { + "bbox": [ + 124, + 308, + 252, + 322 + ], + "score": 1.0, + "content": "NeurIPS, pages 7597–7610, 2020.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "[31] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 336, + 474, + 349 + ], + "spans": [ + { + "bbox": [ + 126, + 336, + 474, + 349 + ], + "score": 1.0, + "content": "anchor points really indispensable in label-noise learning? In NeurIPS, pages 6835–6846, 2019.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "[32] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 365, + 360, + 378 + ], + "spans": [ + { + "bbox": [ + 126, + 365, + 360, + 378 + ], + "score": 1.0, + "content": "data for image classification. In CVPR, pages 2691–2699, 2015.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "[33] Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 124, + 392, + 340, + 406 + ], + "spans": [ + { + "bbox": [ + 124, + 392, + 340, + 406 + ], + "score": 1.0, + "content": "consistency training. In NeurIPS, pages 6256–6268, 2020.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 409, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 104, + 409, + 506, + 424 + ], + "score": 1.0, + "content": "[34] Yilun Xu, Peng Cao, Yuqing Kong, and Yizhou Wang. L_dmi: A novel information-theoretic loss function", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 420, + 418, + 433 + ], + "spans": [ + { + "bbox": [ + 126, + 420, + 418, + 433 + ], + "score": 1.0, + "content": "for training deep nets robust to label noise. In NeurIPS, pages 6222–6233, 2019.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 438, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 452 + ], + "score": 1.0, + "content": "[35] Yu Yao, Tongliang Liu, Bo Han, Mingming Gong, Jiankang Deng, Gang Niu, and Masashi Sugiyama. Dual", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 124, + 446, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 124, + 446, + 507, + 462 + ], + "score": 1.0, + "content": "T: reducing estimation error for transition matrix in label-noise learning. In NeurIPS, pages 7260–7271,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 458, + 150, + 470 + ], + "spans": [ + { + "bbox": [ + 126, + 458, + 150, + 470 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 474, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 104, + 474, + 505, + 491 + ], + "score": 1.0, + "content": "[36] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 486, + 342, + 499 + ], + "spans": [ + { + "bbox": [ + 126, + 486, + 342, + 499 + ], + "score": 1.0, + "content": "learning requires rethinking generalization. In ICLR, 2016.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 502, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 502, + 506, + 518 + ], + "score": 1.0, + "content": "[37] Hongyi Zhang, Moustapha Cissé, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 515, + 236, + 526 + ], + "spans": [ + { + "bbox": [ + 126, + 515, + 236, + 526 + ], + "score": 1.0, + "content": "minimization. In ICLR, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 16.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 71, + 506, + 531 + ], + "lines": [], + "index": 16.5, + "bbox_fs": [ + 104, + 74, + 507, + 526 + ], + "lines_deleted": true + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_model.json b/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_model.json new file mode 100644 index 0000000000000000000000000000000000000000..af923ba2e0317792168ad3a8b8e5ad5f335e5e3c --- /dev/null +++ b/parse/train/KbV-UZRKb3g/KbV-UZRKb3g_model.json @@ -0,0 +1,15086 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 788, + 1303, + 788, + 1303, + 1394, + 398, + 1394 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1521, + 1404, + 1521, + 1404, + 1765, + 298, + 1765 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 1778, + 1405, + 1778, + 1405, + 1901, + 300, + 1901 + ], + "score": 0.969 + }, + { + "category_id": 0, + "poly": [ + 341, + 271, + 1362, + 271, + 1362, + 382, + 341, + 382 + ], + "score": 0.963 + }, + { + "category_id": 2, + "poly": [ + 330, + 1924, + 996, + 1924, + 996, + 1981, + 330, + 1981 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 499, + 495, + 1209, + 495, + 1209, + 565, + 499, + 565 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 365, + 573, + 1334, + 573, + 1334, + 639, + 365, + 639 + ], + "score": 0.921 + }, + { + "category_id": 0, + "poly": [ + 299, + 1450, + 530, + 1450, + 530, + 1489, + 299, + 1489 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 298, + 2033, + 1070, + 2033, + 1070, + 2062, + 298, + 2062 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 788, + 717, + 912, + 717, + 912, + 754, + 788, + 754 + ], + "score": 0.886 + }, + { + "category_id": 13, + "poly": [ + 821, + 1954, + 842, + 1954, + 842, + 1977, + 821, + 1977 + ], + "score": 0.28, + "latex": "@" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 266.0, + 1367.0, + 266.0, + 1367.0, + 333.0, + 334.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 328.0, + 1141.0, + 328.0, + 1141.0, + 385.0, + 559.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1916.0, + 492.0, + 1916.0, + 492.0, + 1954.0, + 327.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1946.0, + 820.0, + 1946.0, + 820.0, + 1986.0, + 328.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1946.0, + 996.0, + 1946.0, + 996.0, + 1986.0, + 843.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1447.0, + 536.0, + 1447.0, + 536.0, + 1495.0, + 292.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2028.0, + 1074.0, + 2028.0, + 1074.0, + 2066.0, + 294.0, + 2066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 717.0, + 917.0, + 717.0, + 917.0, + 756.0, + 784.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 786.0, + 1305.0, + 786.0, + 1305.0, + 824.0, + 394.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 819.0, + 1304.0, + 819.0, + 1304.0, + 855.0, + 394.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 847.0, + 1307.0, + 847.0, + 1307.0, + 886.0, + 393.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 880.0, + 1307.0, + 880.0, + 1307.0, + 916.0, + 393.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 909.0, + 1308.0, + 909.0, + 1308.0, + 945.0, + 393.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 940.0, + 1307.0, + 940.0, + 1307.0, + 975.0, + 394.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 971.0, + 1307.0, + 971.0, + 1307.0, + 1004.0, + 394.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 999.0, + 1307.0, + 999.0, + 1307.0, + 1036.0, + 393.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1030.0, + 1308.0, + 1030.0, + 1308.0, + 1065.0, + 394.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1059.0, + 1305.0, + 1059.0, + 1305.0, + 1096.0, + 393.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1090.0, + 1308.0, + 1090.0, + 1308.0, + 1128.0, + 392.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1119.0, + 1305.0, + 1119.0, + 1305.0, + 1157.0, + 392.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1150.0, + 1304.0, + 1150.0, + 1304.0, + 1188.0, + 393.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1181.0, + 1305.0, + 1181.0, + 1305.0, + 1219.0, + 392.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1214.0, + 1305.0, + 1214.0, + 1305.0, + 1246.0, + 394.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1243.0, + 1306.0, + 1243.0, + 1306.0, + 1278.0, + 392.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1274.0, + 1308.0, + 1274.0, + 1308.0, + 1310.0, + 394.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1303.0, + 1307.0, + 1303.0, + 1307.0, + 1339.0, + 394.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1335.0, + 1305.0, + 1335.0, + 1305.0, + 1367.0, + 395.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1364.0, + 1087.0, + 1364.0, + 1087.0, + 1397.0, + 394.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1520.0, + 1405.0, + 1520.0, + 1405.0, + 1556.0, + 293.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1550.0, + 1405.0, + 1550.0, + 1405.0, + 1588.0, + 292.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1582.0, + 1405.0, + 1582.0, + 1405.0, + 1616.0, + 296.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1612.0, + 1405.0, + 1612.0, + 1405.0, + 1648.0, + 293.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1641.0, + 1408.0, + 1641.0, + 1408.0, + 1679.0, + 292.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1670.0, + 1405.0, + 1670.0, + 1405.0, + 1709.0, + 292.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1705.0, + 1405.0, + 1705.0, + 1405.0, + 1739.0, + 294.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1736.0, + 506.0, + 1736.0, + 506.0, + 1767.0, + 296.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1776.0, + 1409.0, + 1776.0, + 1409.0, + 1815.0, + 294.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1810.0, + 1405.0, + 1810.0, + 1405.0, + 1843.0, + 295.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1841.0, + 1405.0, + 1841.0, + 1405.0, + 1873.0, + 295.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1870.0, + 1407.0, + 1870.0, + 1407.0, + 1902.0, + 295.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 494.0, + 870.0, + 494.0, + 870.0, + 532.0, + 498.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 495.0, + 991.0, + 495.0, + 991.0, + 530.0, + 875.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 495.0, + 1193.0, + 495.0, + 1193.0, + 533.0, + 1014.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 530.0, + 648.0, + 530.0, + 648.0, + 565.0, + 499.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 530.0, + 839.0, + 530.0, + 839.0, + 564.0, + 668.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 528.0, + 994.0, + 528.0, + 994.0, + 566.0, + 854.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 530.0, + 1216.0, + 530.0, + 1216.0, + 566.0, + 1015.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 569.0, + 1340.0, + 569.0, + 1340.0, + 615.0, + 361.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 602.0, + 1085.0, + 602.0, + 1085.0, + 646.0, + 615.0, + 646.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1611, + 1405, + 1611, + 1405, + 2007, + 298, + 2007 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 913, + 1404, + 913, + 1404, + 1248, + 298, + 1248 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1263, + 1405, + 1263, + 1405, + 1598, + 298, + 1598 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 717, + 1404, + 717, + 1404, + 900, + 298, + 900 + ], + "score": 0.977 + }, + { + "category_id": 4, + "poly": [ + 296, + 551, + 1405, + 551, + 1405, + 675, + 296, + 675 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 303, + 223, + 1395, + 223, + 1395, + 535, + 303, + 535 + ], + "score": 0.965 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2085, + 841, + 2085 + ], + "score": 0.724 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.105 + }, + { + "category_id": 13, + "poly": [ + 535, + 505, + 586, + 505, + 586, + 533, + 535, + 533 + ], + "score": 0.28, + "latex": "50 \\%" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 552.0, + 1403.0, + 552.0, + 1403.0, + 585.0, + 295.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 583.0, + 1408.0, + 583.0, + 1408.0, + 615.0, + 295.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 612.0, + 1406.0, + 612.0, + 1406.0, + 646.0, + 293.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 645.0, + 1133.0, + 645.0, + 1133.0, + 677.0, + 295.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 228.0, + 344.0, + 228.0, + 344.0, + 244.0, + 324.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 226.0, + 707.0, + 226.0, + 707.0, + 245.0, + 685.0, + 245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 228.0, + 1070.0, + 228.0, + 1070.0, + 244.0, + 1049.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 276.0, + 344.0, + 276.0, + 344.0, + 297.0, + 323.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 294.0, + 541.0, + 294.0, + 541.0, + 303.0, + 522.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 300.0, + 326.0, + 300.0, + 326.0, + 399.0, + 305.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 299.0, + 706.0, + 299.0, + 706.0, + 402.0, + 668.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 299.0, + 1055.0, + 299.0, + 1055.0, + 400.0, + 1032.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 325.0, + 344.0, + 325.0, + 344.0, + 348.0, + 324.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 317.0, + 1069.0, + 317.0, + 1069.0, + 339.0, + 1046.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 352.0, + 707.0, + 352.0, + 707.0, + 374.0, + 686.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 378.0, + 344.0, + 378.0, + 344.0, + 397.0, + 323.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 384.0, + 706.0, + 384.0, + 706.0, + 405.0, + 685.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 388.0, + 814.0, + 388.0, + 814.0, + 405.0, + 779.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 371.0, + 1070.0, + 371.0, + 1070.0, + 391.0, + 1049.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 389.0, + 1234.0, + 389.0, + 1234.0, + 406.0, + 1202.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 403.0, + 443.0, + 403.0, + 443.0, + 428.0, + 378.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 411.0, + 707.0, + 411.0, + 707.0, + 431.0, + 685.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 402.0, + 807.0, + 402.0, + 807.0, + 427.0, + 739.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 410.0, + 1070.0, + 410.0, + 1070.0, + 429.0, + 1048.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 405.0, + 1169.0, + 405.0, + 1169.0, + 425.0, + 1106.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 431.0, + 342.0, + 431.0, + 342.0, + 446.0, + 324.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 419.0, + 449.0, + 419.0, + 449.0, + 460.0, + 378.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 433.0, + 707.0, + 433.0, + 707.0, + 454.0, + 685.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 421.0, + 812.0, + 421.0, + 812.0, + 460.0, + 742.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 441.0, + 1071.0, + 441.0, + 1071.0, + 460.0, + 1048.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 421.0, + 1177.0, + 421.0, + 1177.0, + 459.0, + 1105.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 452.0, + 709.0, + 452.0, + 709.0, + 472.0, + 685.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 461.0, + 414.0, + 461.0, + 414.0, + 482.0, + 391.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 463.0, + 475.0, + 463.0, + 475.0, + 478.0, + 457.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 461.0, + 540.0, + 461.0, + 540.0, + 481.0, + 517.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 461.0, + 602.0, + 461.0, + 602.0, + 481.0, + 580.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 461.0, + 669.0, + 461.0, + 669.0, + 479.0, + 642.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 461.0, + 777.0, + 461.0, + 777.0, + 482.0, + 755.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 461.0, + 840.0, + 461.0, + 840.0, + 481.0, + 818.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 461.0, + 902.0, + 461.0, + 902.0, + 481.0, + 880.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 461.0, + 966.0, + 461.0, + 966.0, + 481.0, + 942.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 461.0, + 1032.0, + 461.0, + 1032.0, + 481.0, + 1003.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 461.0, + 1141.0, + 461.0, + 1141.0, + 482.0, + 1117.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 461.0, + 1203.0, + 461.0, + 1203.0, + 481.0, + 1182.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 461.0, + 1265.0, + 461.0, + 1265.0, + 481.0, + 1241.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 461.0, + 1328.0, + 461.0, + 1328.0, + 481.0, + 1306.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 461.0, + 1395.0, + 461.0, + 1395.0, + 481.0, + 1366.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 474.0, + 526.0, + 474.0, + 526.0, + 496.0, + 472.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 473.0, + 890.0, + 473.0, + 890.0, + 496.0, + 832.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 474.0, + 1253.0, + 474.0, + 1253.0, + 496.0, + 1198.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 505.0, + 534.0, + 505.0, + 534.0, + 535.0, + 383.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 503.0, + 936.0, + 503.0, + 936.0, + 537.0, + 762.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 505.0, + 1300.0, + 505.0, + 1300.0, + 535.0, + 1122.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 372.5, + 636.0, + 372.5, + 636.0, + 418.5, + 520.0, + 418.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 318.5, + 652.0, + 318.5, + 652.0, + 327.5, + 634.0, + 327.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2094.0, + 839.0, + 2094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2093.0, + 838.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1611.0, + 1406.0, + 1611.0, + 1406.0, + 1648.0, + 295.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1640.0, + 1407.0, + 1640.0, + 1407.0, + 1681.0, + 292.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1667.0, + 1409.0, + 1667.0, + 1409.0, + 1712.0, + 292.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1702.0, + 1407.0, + 1702.0, + 1407.0, + 1738.0, + 295.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1734.0, + 1407.0, + 1734.0, + 1407.0, + 1770.0, + 295.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1764.0, + 1406.0, + 1764.0, + 1406.0, + 1800.0, + 295.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1792.0, + 1408.0, + 1792.0, + 1408.0, + 1832.0, + 292.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1819.0, + 1410.0, + 1819.0, + 1410.0, + 1861.0, + 291.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1853.0, + 1406.0, + 1853.0, + 1406.0, + 1890.0, + 293.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1879.0, + 1410.0, + 1879.0, + 1410.0, + 1927.0, + 291.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1915.0, + 1405.0, + 1915.0, + 1405.0, + 1951.0, + 295.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1403.0, + 1943.0, + 1403.0, + 1979.0, + 295.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1976.0, + 1142.0, + 1976.0, + 1142.0, + 2012.0, + 295.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 912.0, + 1405.0, + 912.0, + 1405.0, + 950.0, + 293.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 946.0, + 1407.0, + 946.0, + 1407.0, + 980.0, + 292.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 975.0, + 1404.0, + 975.0, + 1404.0, + 1010.0, + 293.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1001.0, + 1409.0, + 1001.0, + 1409.0, + 1045.0, + 291.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1036.0, + 1406.0, + 1036.0, + 1406.0, + 1071.0, + 294.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1067.0, + 1406.0, + 1067.0, + 1406.0, + 1102.0, + 294.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1095.0, + 1408.0, + 1095.0, + 1408.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1125.0, + 1406.0, + 1125.0, + 1406.0, + 1163.0, + 293.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1153.0, + 1405.0, + 1153.0, + 1405.0, + 1195.0, + 291.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1188.0, + 1406.0, + 1188.0, + 1406.0, + 1223.0, + 293.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1218.0, + 906.0, + 1218.0, + 906.0, + 1252.0, + 296.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1264.0, + 1402.0, + 1264.0, + 1402.0, + 1298.0, + 295.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1292.0, + 1405.0, + 1292.0, + 1405.0, + 1330.0, + 293.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1318.0, + 1406.0, + 1318.0, + 1406.0, + 1363.0, + 292.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1354.0, + 1406.0, + 1354.0, + 1406.0, + 1390.0, + 292.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1384.0, + 1406.0, + 1384.0, + 1406.0, + 1421.0, + 292.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1416.0, + 1410.0, + 1416.0, + 1410.0, + 1450.0, + 295.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1445.0, + 1405.0, + 1445.0, + 1405.0, + 1480.0, + 295.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1476.0, + 1407.0, + 1476.0, + 1407.0, + 1511.0, + 295.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1506.0, + 1405.0, + 1506.0, + 1405.0, + 1540.0, + 295.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1534.0, + 1406.0, + 1534.0, + 1406.0, + 1575.0, + 292.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1566.0, + 1202.0, + 1566.0, + 1202.0, + 1602.0, + 293.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 719.0, + 1404.0, + 719.0, + 1404.0, + 750.0, + 296.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 748.0, + 1405.0, + 748.0, + 1405.0, + 783.0, + 294.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 778.0, + 1406.0, + 778.0, + 1406.0, + 814.0, + 294.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 809.0, + 1406.0, + 809.0, + 1406.0, + 844.0, + 293.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 839.0, + 1405.0, + 839.0, + 1405.0, + 874.0, + 294.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 571.0, + 872.0, + 571.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1459, + 1404, + 1459, + 1404, + 1675, + 297, + 1675 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1231, + 1404, + 1231, + 1404, + 1446, + 297, + 1446 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 468, + 1406, + 468, + 1406, + 726, + 297, + 726 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 924, + 1406, + 924, + 1406, + 1108, + 297, + 1108 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 202, + 1405, + 202, + 1405, + 356, + 298, + 356 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1786, + 1403, + 1786, + 1403, + 1910, + 298, + 1910 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 737, + 1405, + 737, + 1405, + 831, + 299, + 831 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 622, + 1931, + 1075, + 1931, + 1075, + 2014, + 622, + 2014 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 655, + 1692, + 1043, + 1692, + 1043, + 1770, + 655, + 1770 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 693, + 1130, + 1005, + 1130, + 1005, + 1214, + 693, + 1214 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 300, + 868, + 686, + 868, + 686, + 902, + 300, + 902 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 299, + 399, + 602, + 399, + 602, + 438, + 299, + 438 + ], + "score": 0.923 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1716, + 1400, + 1716, + 1400, + 1745, + 1366, + 1745 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1155, + 1400, + 1155, + 1400, + 1185, + 1366, + 1185 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1957, + 1400, + 1957, + 1400, + 1986, + 1366, + 1986 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2085, + 841, + 2085 + ], + "score": 0.694 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.366 + }, + { + "category_id": 14, + "poly": [ + 691, + 1126, + 1004, + 1126, + 1004, + 1217, + 691, + 1217 + ], + "score": 0.95, + "latex": "\\operatorname* { m i n } _ { \\Theta } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( x _ { i } ; \\Theta ) , \\tilde { y } _ { i } ) ," + }, + { + "category_id": 14, + "poly": [ + 621, + 1926, + 1078, + 1926, + 1078, + 2017, + 621, + 2017 + ], + "score": 0.94, + "latex": "\\operatorname* { m i n } _ { \\Theta _ { 1 } \\ldots \\Theta _ { k } } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( \\pmb { x } _ { i } ; \\Theta _ { 1 } , \\ldots , \\Theta _ { L } ) , \\tilde { y } _ { i } ) ." + }, + { + "category_id": 13, + "poly": [ + 1095, + 626, + 1235, + 626, + 1235, + 662, + 1095, + 662 + ], + "score": 0.93, + "latex": "\\{ \\pmb { x } _ { i } , \\tilde { y } _ { i } ) \\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 915, + 592, + 1004, + 592, + 1004, + 630, + 915, + 630 + ], + "score": 0.93, + "latex": "( X , { \\tilde { Y } } )" + }, + { + "category_id": 13, + "poly": [ + 395, + 1642, + 477, + 1642, + 477, + 1676, + 395, + 1676 + ], + "score": 0.93, + "latex": "f ( \\cdot ; \\Theta )" + }, + { + "category_id": 13, + "poly": [ + 373, + 1787, + 468, + 1787, + 468, + 1821, + 373, + 1821 + ], + "score": 0.93, + "latex": "f _ { l } ( \\cdot ; \\Theta _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 898, + 471, + 1181, + 471, + 1181, + 505, + 898, + 505 + ], + "score": 0.92, + "latex": "( X , Y ) \\in \\mathcal { X } \\times \\{ 1 , . . . K \\}" + }, + { + "category_id": 13, + "poly": [ + 682, + 1818, + 762, + 1818, + 762, + 1851, + 682, + 1851 + ], + "score": 0.92, + "latex": "f ( \\cdot ; \\Theta )" + }, + { + "category_id": 14, + "poly": [ + 655, + 1688, + 1044, + 1688, + 1044, + 1773, + 655, + 1773 + ], + "score": 0.92, + "latex": "\\begin{array} { l } { z _ { 1 } = f _ { 1 } ( \\pmb { x } ; \\Theta _ { 1 } ) , } \\\\ { z _ { l } = f _ { l } ( z _ { l - 1 } ; \\Theta _ { l } ) , \\quad l = 2 , \\ldots , L } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 373, + 1233, + 455, + 1233, + 455, + 1267, + 373, + 1267 + ], + "score": 0.92, + "latex": "f ( \\cdot ; \\Theta )" + }, + { + "category_id": 13, + "poly": [ + 1065, + 1818, + 1232, + 1818, + 1232, + 1851, + 1065, + 1851 + ], + "score": 0.91, + "latex": "f ( \\cdot ; \\Theta _ { 1 } , . . . \\Theta _ { L } )" + }, + { + "category_id": 13, + "poly": [ + 508, + 1880, + 539, + 1880, + 539, + 1909, + 508, + 1909 + ], + "score": 0.88, + "latex": "T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1009, + 1849, + 1044, + 1849, + 1044, + 1878, + 1009, + 1878 + ], + "score": 0.88, + "latex": "\\Theta _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1094, + 592, + 1122, + 592, + 1122, + 624, + 1094, + 624 + ], + "score": 0.86, + "latex": "\\tilde { Y }" + }, + { + "category_id": 13, + "poly": [ + 425, + 592, + 452, + 592, + 452, + 624, + 425, + 624 + ], + "score": 0.85, + "latex": "\\tilde { D }" + }, + { + "category_id": 13, + "poly": [ + 752, + 1792, + 780, + 1792, + 780, + 1818, + 752, + 1818 + ], + "score": 0.85, + "latex": "z _ { l }" + }, + { + "category_id": 13, + "poly": [ + 434, + 659, + 460, + 659, + 460, + 690, + 434, + 690 + ], + "score": 0.85, + "latex": "\\tilde { D }" + }, + { + "category_id": 13, + "poly": [ + 1220, + 503, + 1249, + 503, + 1249, + 529, + 1220, + 529 + ], + "score": 0.84, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 298, + 1823, + 332, + 1823, + 332, + 1848, + 298, + 1848 + ], + "score": 0.84, + "latex": "z _ { L }" + }, + { + "category_id": 13, + "poly": [ + 939, + 1234, + 964, + 1234, + 964, + 1261, + 939, + 1261 + ], + "score": 0.83, + "latex": "\\Theta" + }, + { + "category_id": 13, + "poly": [ + 1016, + 1235, + 1038, + 1235, + 1038, + 1261, + 1016, + 1261 + ], + "score": 0.83, + "latex": "\\mathcal { L }" + }, + { + "category_id": 13, + "poly": [ + 639, + 1078, + 662, + 1078, + 662, + 1104, + 639, + 1104 + ], + "score": 0.83, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 340, + 473, + 367, + 473, + 367, + 499, + 340, + 499 + ], + "score": 0.82, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 745, + 1644, + 767, + 1644, + 767, + 1670, + 745, + 1670 + ], + "score": 0.81, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 1266, + 473, + 1296, + 473, + 1296, + 499, + 1266, + 499 + ], + "score": 0.81, + "latex": "\\boldsymbol { X }" + }, + { + "category_id": 13, + "poly": [ + 1374, + 532, + 1401, + 532, + 1401, + 559, + 1374, + 559 + ], + "score": 0.81, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 389, + 1416, + 413, + 1416, + 413, + 1442, + 389, + 1442 + ], + "score": 0.81, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 863, + 503, + 890, + 503, + 890, + 529, + 863, + 529 + ], + "score": 0.81, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 848, + 1462, + 872, + 1462, + 872, + 1488, + 848, + 1488 + ], + "score": 0.8, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1247, + 1356, + 1270, + 1356, + 1270, + 1382, + 1247, + 1382 + ], + "score": 0.79, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1326, + 1051, + 1326, + 1051, + 1352, + 1028, + 1352 + ], + "score": 0.78, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 571, + 502, + 599, + 502, + 599, + 529, + 571, + 529 + ], + "score": 0.75, + "latex": "\\mathbf { Y }" + }, + { + "category_id": 13, + "poly": [ + 539, + 1789, + 552, + 1789, + 552, + 1815, + 539, + 1815 + ], + "score": 0.74, + "latex": "l" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 861.0, + 691.0, + 861.0, + 691.0, + 912.0, + 292.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 397.0, + 605.0, + 397.0, + 605.0, + 446.0, + 291.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2058.0, + 860.0, + 2058.0, + 860.0, + 2089.0, + 839.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1461.0, + 847.0, + 1461.0, + 847.0, + 1495.0, + 296.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1461.0, + 1407.0, + 1461.0, + 1407.0, + 1495.0, + 873.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1489.0, + 1404.0, + 1489.0, + 1404.0, + 1524.0, + 295.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1519.0, + 1406.0, + 1519.0, + 1406.0, + 1558.0, + 292.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1552.0, + 1405.0, + 1552.0, + 1405.0, + 1587.0, + 295.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 1404.0, + 1581.0, + 1404.0, + 1618.0, + 294.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1610.0, + 1406.0, + 1610.0, + 1406.0, + 1647.0, + 292.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1642.0, + 394.0, + 1642.0, + 394.0, + 1677.0, + 295.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1642.0, + 744.0, + 1642.0, + 744.0, + 1677.0, + 478.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1642.0, + 897.0, + 1642.0, + 897.0, + 1677.0, + 768.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1231.0, + 372.0, + 1231.0, + 372.0, + 1268.0, + 294.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1231.0, + 938.0, + 1231.0, + 938.0, + 1268.0, + 456.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1231.0, + 1015.0, + 1231.0, + 1015.0, + 1268.0, + 965.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1231.0, + 1405.0, + 1231.0, + 1405.0, + 1268.0, + 1039.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1262.0, + 1402.0, + 1262.0, + 1402.0, + 1297.0, + 295.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1290.0, + 1406.0, + 1290.0, + 1406.0, + 1331.0, + 292.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1325.0, + 1027.0, + 1325.0, + 1027.0, + 1359.0, + 295.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1325.0, + 1406.0, + 1325.0, + 1406.0, + 1359.0, + 1052.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1354.0, + 1246.0, + 1354.0, + 1246.0, + 1389.0, + 295.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1354.0, + 1404.0, + 1354.0, + 1404.0, + 1389.0, + 1271.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1383.0, + 1405.0, + 1383.0, + 1405.0, + 1420.0, + 294.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1416.0, + 388.0, + 1416.0, + 388.0, + 1447.0, + 295.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1416.0, + 1260.0, + 1416.0, + 1260.0, + 1447.0, + 414.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 470.0, + 339.0, + 470.0, + 339.0, + 505.0, + 294.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 470.0, + 897.0, + 470.0, + 897.0, + 505.0, + 368.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 470.0, + 1265.0, + 470.0, + 1265.0, + 505.0, + 1182.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 470.0, + 1404.0, + 470.0, + 1404.0, + 505.0, + 1297.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 501.0, + 570.0, + 501.0, + 570.0, + 532.0, + 296.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 501.0, + 862.0, + 501.0, + 862.0, + 532.0, + 600.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 501.0, + 1219.0, + 501.0, + 1219.0, + 532.0, + 891.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 501.0, + 1404.0, + 501.0, + 1404.0, + 532.0, + 1250.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 531.0, + 1373.0, + 531.0, + 1373.0, + 566.0, + 295.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 560.0, + 1404.0, + 560.0, + 1404.0, + 597.0, + 292.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 592.0, + 424.0, + 592.0, + 424.0, + 631.0, + 294.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 592.0, + 914.0, + 592.0, + 914.0, + 631.0, + 453.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 592.0, + 1093.0, + 592.0, + 1093.0, + 631.0, + 1005.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 592.0, + 1406.0, + 592.0, + 1406.0, + 631.0, + 1123.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 622.0, + 1094.0, + 622.0, + 1094.0, + 668.0, + 290.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 622.0, + 1408.0, + 622.0, + 1408.0, + 668.0, + 1236.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 659.0, + 433.0, + 659.0, + 433.0, + 698.0, + 294.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 659.0, + 1406.0, + 659.0, + 1406.0, + 698.0, + 461.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 693.0, + 532.0, + 693.0, + 532.0, + 728.0, + 294.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 926.0, + 1403.0, + 926.0, + 1403.0, + 958.0, + 296.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 953.0, + 1406.0, + 953.0, + 1406.0, + 993.0, + 292.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 980.0, + 1406.0, + 980.0, + 1406.0, + 1025.0, + 291.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1019.0, + 1407.0, + 1019.0, + 1407.0, + 1050.0, + 295.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1046.0, + 1408.0, + 1046.0, + 1408.0, + 1082.0, + 295.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1077.0, + 638.0, + 1077.0, + 638.0, + 1109.0, + 295.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1077.0, + 671.0, + 1077.0, + 671.0, + 1109.0, + 663.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 204.0, + 1403.0, + 204.0, + 1403.0, + 237.0, + 296.0, + 237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 233.0, + 1405.0, + 233.0, + 1405.0, + 270.0, + 295.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 299.0, + 293.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 292.0, + 1405.0, + 292.0, + 1405.0, + 333.0, + 293.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 327.0, + 510.0, + 327.0, + 510.0, + 355.0, + 293.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1786.0, + 372.0, + 1786.0, + 372.0, + 1822.0, + 294.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1786.0, + 538.0, + 1786.0, + 538.0, + 1822.0, + 469.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1786.0, + 751.0, + 1786.0, + 751.0, + 1822.0, + 553.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1786.0, + 1407.0, + 1786.0, + 1407.0, + 1822.0, + 781.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1816.0, + 297.0, + 1816.0, + 297.0, + 1853.0, + 293.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1816.0, + 681.0, + 1816.0, + 681.0, + 1853.0, + 333.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1816.0, + 1064.0, + 1816.0, + 1064.0, + 1853.0, + 763.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1816.0, + 1407.0, + 1816.0, + 1407.0, + 1853.0, + 1233.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1847.0, + 1008.0, + 1847.0, + 1008.0, + 1882.0, + 294.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1847.0, + 1403.0, + 1847.0, + 1403.0, + 1882.0, + 1045.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1878.0, + 507.0, + 1878.0, + 507.0, + 1911.0, + 296.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1878.0, + 943.0, + 1878.0, + 943.0, + 1911.0, + 540.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 737.0, + 1406.0, + 737.0, + 1406.0, + 774.0, + 293.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 768.0, + 1405.0, + 768.0, + 1405.0, + 802.0, + 294.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 797.0, + 527.0, + 797.0, + 527.0, + 835.0, + 295.0, + 835.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1097, + 1405, + 1097, + 1405, + 1465, + 297, + 1465 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 296, + 839, + 1405, + 839, + 1405, + 1087, + 296, + 1087 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1552, + 1405, + 1552, + 1405, + 1739, + 298, + 1739 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 302, + 1912, + 1403, + 1912, + 1403, + 2009, + 302, + 2009 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 635, + 1405, + 635, + 1405, + 730, + 298, + 730 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 477, + 1771, + 1218, + 1771, + 1218, + 1890, + 477, + 1890 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 303, + 226, + 1392, + 226, + 1392, + 529, + 303, + 529 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 293, + 545, + 1407, + 545, + 1407, + 609, + 293, + 609 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 465, + 744, + 1233, + 744, + 1233, + 828, + 465, + 828 + ], + "score": 0.945 + }, + { + "category_id": 0, + "poly": [ + 299, + 1498, + 780, + 1498, + 780, + 1532, + 299, + 1532 + ], + "score": 0.929 + }, + { + "category_id": 9, + "poly": [ + 1366, + 771, + 1400, + 771, + 1400, + 800, + 1366, + 800 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1818, + 1400, + 1818, + 1400, + 1847, + 1366, + 1847 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.78 + }, + { + "category_id": 13, + "poly": [ + 565, + 1943, + 669, + 1943, + 669, + 1979, + 565, + 1979 + ], + "score": 0.93, + "latex": "f ^ { k } ( x ; \\Theta )" + }, + { + "category_id": 14, + "poly": [ + 464, + 740, + 1234, + 740, + 1234, + 830, + 464, + 830 + ], + "score": 0.93, + "latex": "\\operatorname* { m i n } _ { \\Theta _ { l } \\ldots \\Theta _ { k } } \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\mathcal { L } ( f ( x _ { i } ; \\Theta _ { 1 } ^ { * } , \\ldots , \\Theta _ { l - 1 } ^ { * } , \\Theta _ { l } , \\ldots , \\Theta _ { L } ) , \\tilde { y } _ { i } ) , \\quad l = 2 \\ldots L" + }, + { + "category_id": 14, + "poly": [ + 480, + 1768, + 1218, + 1768, + 1218, + 1894, + 480, + 1894 + ], + "score": 0.91, + "latex": "\\begin{array} { c } { \\mathcal { D } _ { l } = \\{ ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = \\hat { y } _ { i } , i = 1 , \\ldots , n \\} , } \\\\ { \\hat { y } _ { i } = \\underset { k \\in \\{ 1 , \\ldots , K \\} } { \\arg \\operatorname* { m a x } } \\ \\frac { 1 } { 2 } [ f ^ { k } ( \\mathrm { A u g m e n t } ( \\pmb { x } _ { i } ) ; \\Theta ) + f ^ { k } ( \\mathrm { A u g m e n t } ( \\pmb { x } _ { i } ) ; \\Theta ) ] , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 588, + 841, + 660, + 841, + 660, + 874, + 588, + 874 + ], + "score": 0.91, + "latex": "( l + 1 )" + }, + { + "category_id": 13, + "poly": [ + 853, + 992, + 1166, + 992, + 1166, + 1027, + 853, + 1027 + ], + "score": 0.91, + "latex": "f ( \\cdot , \\Theta ) = \\bar { f } ( \\cdot ; \\Theta _ { 1 } ^ { * } , \\ldots , \\Theta _ { L } ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 1114, + 843, + 1150, + 843, + 1150, + 875, + 1114, + 875 + ], + "score": 0.89, + "latex": "\\Theta _ { l } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 741, + 639, + 778, + 639, + 778, + 670, + 741, + 670 + ], + "score": 0.89, + "latex": "\\Theta _ { 1 } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 583, + 963, + 845, + 963, + 845, + 994, + 583, + 994 + ], + "score": 0.89, + "latex": "T _ { 1 } \\geq T _ { 2 } \\geq \\cdot \\cdot \\cdot \\geq T _ { L } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 668, + 403, + 668, + 403, + 699, + 297, + 699 + ], + "score": 0.88, + "latex": "2 , \\ldots , L )" + }, + { + "category_id": 13, + "poly": [ + 929, + 1708, + 961, + 1708, + 961, + 1737, + 929, + 1737 + ], + "score": 0.88, + "latex": "\\mathcal { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 751, + 933, + 821, + 933, + 821, + 962, + 751, + 962 + ], + "score": 0.86, + "latex": "L - 1" + }, + { + "category_id": 13, + "poly": [ + 474, + 699, + 502, + 699, + 502, + 728, + 474, + 728 + ], + "score": 0.85, + "latex": "T _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1259, + 1947, + 1277, + 1947, + 1277, + 1973, + 1259, + 1973 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1354, + 639, + 1405, + 639, + 1405, + 668, + 1354, + 668 + ], + "score": 0.79, + "latex": "( l =" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1951, + 1040, + 1951, + 1040, + 1973, + 1020, + 1973 + ], + "score": 0.76, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 13, + "poly": [ + 401, + 1976, + 431, + 1976, + 431, + 2009, + 401, + 2009 + ], + "score": 0.73, + "latex": "\\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 1299, + 639, + 1311, + 639, + 1311, + 665, + 1299, + 665 + ], + "score": 0.68, + "latex": "l" + }, + { + "category_id": 13, + "poly": [ + 477, + 1912, + 507, + 1912, + 507, + 1946, + 477, + 1946 + ], + "score": 0.68, + "latex": "( \\cdot )" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 241.0, + 347.0, + 241.0, + 347.0, + 262.0, + 322.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 241.0, + 711.0, + 241.0, + 711.0, + 262.0, + 685.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 241.0, + 1073.0, + 241.0, + 1073.0, + 262.0, + 1048.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 290.0, + 346.0, + 290.0, + 346.0, + 395.0, + 304.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 290.0, + 710.0, + 290.0, + 710.0, + 397.0, + 667.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 290.0, + 1073.0, + 290.0, + 1073.0, + 398.0, + 1030.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 406.0, + 347.0, + 406.0, + 347.0, + 425.0, + 324.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 411.0, + 647.0, + 411.0, + 647.0, + 436.0, + 557.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 405.0, + 710.0, + 405.0, + 710.0, + 425.0, + 685.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 411.0, + 1010.0, + 411.0, + 1010.0, + 436.0, + 920.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 405.0, + 1073.0, + 405.0, + 1073.0, + 425.0, + 1048.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 411.0, + 1373.0, + 411.0, + 1373.0, + 436.0, + 1283.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 428.0, + 589.0, + 428.0, + 589.0, + 451.0, + 556.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 428.0, + 952.0, + 428.0, + 952.0, + 451.0, + 921.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 430.0, + 1313.0, + 430.0, + 1313.0, + 449.0, + 1283.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 454.0, + 409.0, + 454.0, + 409.0, + 473.0, + 384.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 454.0, + 473.0, + 454.0, + 473.0, + 473.0, + 447.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 454.0, + 537.0, + 454.0, + 537.0, + 474.0, + 511.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 452.0, + 601.0, + 452.0, + 601.0, + 476.0, + 573.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 452.0, + 669.0, + 452.0, + 669.0, + 474.0, + 633.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 454.0, + 772.0, + 454.0, + 772.0, + 473.0, + 747.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 454.0, + 836.0, + 454.0, + 836.0, + 473.0, + 810.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 452.0, + 901.0, + 452.0, + 901.0, + 476.0, + 872.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 452.0, + 965.0, + 452.0, + 965.0, + 476.0, + 935.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 452.0, + 1032.0, + 452.0, + 1032.0, + 474.0, + 996.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 454.0, + 1135.0, + 454.0, + 1135.0, + 473.0, + 1110.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 454.0, + 1199.0, + 454.0, + 1199.0, + 473.0, + 1174.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 452.0, + 1264.0, + 452.0, + 1264.0, + 476.0, + 1235.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 452.0, + 1328.0, + 452.0, + 1328.0, + 476.0, + 1298.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 452.0, + 1396.0, + 452.0, + 1396.0, + 474.0, + 1359.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 469.0, + 529.0, + 469.0, + 529.0, + 491.0, + 470.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 469.0, + 892.0, + 469.0, + 892.0, + 491.0, + 833.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 467.0, + 1256.0, + 467.0, + 1256.0, + 492.0, + 1194.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 497.0, + 591.0, + 497.0, + 591.0, + 528.0, + 383.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 496.0, + 935.0, + 496.0, + 935.0, + 530.0, + 762.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 497.0, + 1302.0, + 497.0, + 1302.0, + 527.0, + 1121.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 348.0, + 591.0, + 348.0, + 591.0, + 358.0, + 585.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 361.0, + 615.0, + 361.0, + 615.0, + 371.0, + 604.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 544.0, + 1404.0, + 544.0, + 1404.0, + 580.0, + 295.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 576.0, + 1127.0, + 576.0, + 1127.0, + 612.0, + 295.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1495.0, + 782.0, + 1495.0, + 782.0, + 1538.0, + 294.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1097.0, + 1408.0, + 1097.0, + 1408.0, + 1135.0, + 294.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1129.0, + 1405.0, + 1129.0, + 1405.0, + 1165.0, + 292.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1158.0, + 1407.0, + 1158.0, + 1407.0, + 1196.0, + 294.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1190.0, + 1406.0, + 1190.0, + 1406.0, + 1224.0, + 294.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1221.0, + 1405.0, + 1221.0, + 1405.0, + 1255.0, + 295.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 1405.0, + 1251.0, + 1405.0, + 1285.0, + 295.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1280.0, + 1407.0, + 1280.0, + 1407.0, + 1318.0, + 294.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1311.0, + 1405.0, + 1311.0, + 1405.0, + 1346.0, + 295.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 1405.0, + 1339.0, + 1405.0, + 1376.0, + 294.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1370.0, + 1406.0, + 1370.0, + 1406.0, + 1408.0, + 292.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1400.0, + 1403.0, + 1400.0, + 1403.0, + 1435.0, + 294.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1431.0, + 1244.0, + 1431.0, + 1244.0, + 1469.0, + 294.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 840.0, + 587.0, + 840.0, + 587.0, + 876.0, + 294.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 840.0, + 1113.0, + 840.0, + 1113.0, + 876.0, + 661.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 840.0, + 1403.0, + 840.0, + 1403.0, + 876.0, + 1151.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 1405.0, + 872.0, + 1405.0, + 906.0, + 294.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 897.0, + 1407.0, + 897.0, + 1407.0, + 940.0, + 290.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 932.0, + 750.0, + 932.0, + 750.0, + 968.0, + 293.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 932.0, + 1406.0, + 932.0, + 1406.0, + 968.0, + 822.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 962.0, + 582.0, + 962.0, + 582.0, + 999.0, + 294.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 962.0, + 1407.0, + 962.0, + 1407.0, + 999.0, + 846.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 992.0, + 852.0, + 992.0, + 852.0, + 1029.0, + 294.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 992.0, + 1406.0, + 992.0, + 1406.0, + 1029.0, + 1167.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1021.0, + 1407.0, + 1021.0, + 1407.0, + 1061.0, + 291.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1052.0, + 942.0, + 1052.0, + 942.0, + 1090.0, + 294.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1552.0, + 1407.0, + 1552.0, + 1407.0, + 1592.0, + 293.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1585.0, + 1402.0, + 1585.0, + 1402.0, + 1617.0, + 296.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1615.0, + 1405.0, + 1615.0, + 1405.0, + 1651.0, + 293.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1644.0, + 1407.0, + 1644.0, + 1407.0, + 1681.0, + 291.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1672.0, + 1406.0, + 1672.0, + 1406.0, + 1713.0, + 293.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1706.0, + 928.0, + 1706.0, + 928.0, + 1740.0, + 295.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1706.0, + 996.0, + 1706.0, + 996.0, + 1740.0, + 962.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 476.0, + 1909.0, + 476.0, + 1949.0, + 295.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1909.0, + 1404.0, + 1909.0, + 1404.0, + 1949.0, + 508.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1943.0, + 564.0, + 1943.0, + 564.0, + 1981.0, + 296.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1943.0, + 1019.0, + 1943.0, + 1019.0, + 1981.0, + 670.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1943.0, + 1258.0, + 1943.0, + 1258.0, + 1981.0, + 1041.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1943.0, + 1407.0, + 1943.0, + 1407.0, + 1981.0, + 1278.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1974.0, + 400.0, + 1974.0, + 400.0, + 2011.0, + 297.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2011.0, + 432.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 634.0, + 740.0, + 634.0, + 740.0, + 673.0, + 292.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 634.0, + 1298.0, + 634.0, + 1298.0, + 673.0, + 779.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 634.0, + 1353.0, + 634.0, + 1353.0, + 673.0, + 1312.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 668.0, + 1405.0, + 668.0, + 1405.0, + 702.0, + 404.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 698.0, + 473.0, + 698.0, + 473.0, + 732.0, + 295.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 698.0, + 1012.0, + 698.0, + 1012.0, + 732.0, + 503.0, + 732.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1853, + 1404, + 1853, + 1404, + 2009, + 298, + 2009 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 303, + 1564, + 1397, + 1564, + 1397, + 1689, + 303, + 1689 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1068, + 1404, + 1068, + 1404, + 1170, + 297, + 1170 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1261, + 1403, + 1261, + 1403, + 1383, + 300, + 1383 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 778, + 1404, + 778, + 1404, + 937, + 297, + 937 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 481, + 1398, + 1217, + 1398, + 1217, + 1552, + 481, + 1552 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 671, + 971, + 1028, + 971, + 1028, + 1060, + 671, + 1060 + ], + "score": 0.959 + }, + { + "category_id": 0, + "poly": [ + 297, + 1204, + 878, + 1204, + 878, + 1239, + 297, + 1239 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 300, + 1797, + 801, + 1797, + 801, + 1831, + 300, + 1831 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 298, + 1729, + 533, + 1729, + 533, + 1767, + 298, + 1767 + ], + "score": 0.923 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1001, + 1400, + 1001, + 1400, + 1031, + 1366, + 1031 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1460, + 1400, + 1460, + 1400, + 1490, + 1366, + 1490 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.721 + }, + { + "category_id": 1, + "poly": [ + 294, + 248, + 1349, + 248, + 1349, + 713, + 294, + 713 + ], + "score": 0.532 + }, + { + "category_id": 6, + "poly": [ + 300, + 210, + 1124, + 210, + 1124, + 243, + 300, + 243 + ], + "score": 0.516 + }, + { + "category_id": 5, + "poly": [ + 294, + 248, + 1349, + 248, + 1349, + 713, + 294, + 713 + ], + "score": 0.096, + "html": "
for refining with confident examples.Input: Neural network with trainable parameters O = {O1,..,OL}, Noisy training dataset {xi, yi)}"=1, Number of training epochs for different part: T1,...,TL, and training epochs Tc
fori=1,...,Tdo
L Optimize network parameter O with Eq. (3);
forl= 2,...,L doFroze {Θ1,...,Θt-1} and re-initialize {Θt,...,ΘL};
for i=1,...,T do
L Optimize network parameter {Ot,..., ΘL} with Eq. (4);
Unfroze ;
for i=1,...,Tc do
Extract confident example set Dl and unlabeled set Du with classifier f(·, O) by Eq. (7);
Training the classifier f(·,O) with MixMatch loss on Dt and Du;
Evaluate the obtained classifier f(-, O).
" + }, + { + "category_id": 14, + "poly": [ + 671, + 967, + 1026, + 967, + 1026, + 1063, + 671, + 1063 + ], + "score": 0.95, + "latex": "\\mathcal { L } _ { c } = \\sum _ { i = 1 } ^ { N } w _ { y _ { i } } \\mathcal { L } _ { p } ( \\tilde { y } _ { i } , f ( \\pmb { x } _ { i } ; \\Theta ) ) ," + }, + { + "category_id": 13, + "poly": [ + 709, + 497, + 868, + 497, + 868, + 531, + 709, + 531 + ], + "score": 0.93, + "latex": "\\big \\{ \\Theta _ { l } , \\dots , \\Theta _ { L } \\big \\}" + }, + { + "category_id": 13, + "poly": [ + 1115, + 605, + 1197, + 605, + 1197, + 639, + 1115, + 639 + ], + "score": 0.93, + "latex": "f ( \\cdot , \\Theta )" + }, + { + "category_id": 13, + "poly": [ + 587, + 637, + 668, + 637, + 668, + 670, + 587, + 670 + ], + "score": 0.93, + "latex": "f ( \\cdot , \\Theta )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1129, + 527, + 1129, + 527, + 1174, + 298, + 1174 + ], + "score": 0.92, + "latex": "w _ { i } = \\sigma _ { i } / ( \\sum _ { j = 1 } ^ { K } \\sigma _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 781, + 436, + 940, + 436, + 940, + 472, + 781, + 472 + ], + "score": 0.92, + "latex": "\\big \\{ \\Theta _ { l } , \\dots , \\Theta _ { L } \\big \\}" + }, + { + "category_id": 13, + "poly": [ + 298, + 280, + 438, + 280, + 438, + 315, + 298, + 315 + ], + "score": 0.92, + "latex": "\\{ \\bar { \\pmb { x } _ { i } } , \\tilde { y } _ { i } ) \\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 962, + 1069, + 1402, + 1069, + 1402, + 1103, + 962, + 1103 + ], + "score": 0.91, + "latex": "\\sigma _ { k } = | \\{ ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = k , ( \\pmb { x } _ { i } , \\tilde { y } _ { i } ) \\in \\mathcal { D } _ { l } \\} |" + }, + { + "category_id": 13, + "poly": [ + 649, + 674, + 730, + 674, + 730, + 708, + 649, + 708 + ], + "score": 0.91, + "latex": "f ( \\cdot , \\Theta )" + }, + { + "category_id": 13, + "poly": [ + 339, + 343, + 492, + 343, + 492, + 373, + 339, + 373 + ], + "score": 0.91, + "latex": "i = 1 , \\dots , T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 412, + 437, + 594, + 437, + 594, + 471, + 412, + 471 + ], + "score": 0.91, + "latex": "\\left\\{ \\Theta _ { 1 } , \\ldots , \\Theta _ { l - 1 } \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 953, + 284, + 1079, + 284, + 1079, + 313, + 953, + 313 + ], + "score": 0.91, + "latex": "T _ { 1 } , \\dots , T _ { L }" + }, + { + "category_id": 14, + "poly": [ + 483, + 1396, + 1215, + 1396, + 1215, + 1556, + 483, + 1556 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { \\displaystyle { \\int } { \\mathcal D } _ { l } = \\{ ( \\boldsymbol x _ { i } , \\tilde { y } _ { i } ) | \\tilde { y } _ { i } = \\hat { y } _ { i } , i = 1 , . . . n \\} } } \\\\ { { \\displaystyle { \\mathcal D } _ { u } = \\{ \\boldsymbol x _ { i } | \\tilde { y } _ { i } \\neq \\hat { y } _ { i } , i = 1 , . . . n \\} } } \\\\ { { \\displaystyle { \\hat { y } _ { i } = \\arg \\operatorname* { m a x } _ { k \\in \\{ 1 , . . . K \\} } \\frac { 1 } { 2 } [ f ^ { k } ( \\mathrm { A u g m e n t } ( \\boldsymbol x _ { i } ) ; \\Theta ) + f ^ { k } ( \\mathrm { A u g m e n t } ( \\boldsymbol x _ { i } ) ; \\Theta ) ] } , } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 339, + 577, + 491, + 577, + 491, + 608, + 339, + 608 + ], + "score": 0.9, + "latex": "i = 1 , \\dots , T _ { c }" + }, + { + "category_id": 13, + "poly": [ + 849, + 249, + 1070, + 249, + 1070, + 284, + 849, + 284 + ], + "score": 0.9, + "latex": "\\boldsymbol { \\Theta } = \\{ \\Theta _ { 1 } , \\dots , \\Theta _ { L } \\}" + }, + { + "category_id": 13, + "poly": [ + 1021, + 638, + 1059, + 638, + 1059, + 668, + 1021, + 668 + ], + "score": 0.9, + "latex": "\\mathcal { D } _ { u }" + }, + { + "category_id": 13, + "poly": [ + 910, + 607, + 949, + 607, + 949, + 637, + 910, + 637 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { u }" + }, + { + "category_id": 13, + "poly": [ + 1026, + 1567, + 1066, + 1567, + 1066, + 1596, + 1026, + 1596 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { u }" + }, + { + "category_id": 13, + "poly": [ + 834, + 1946, + 926, + 1946, + 926, + 1975, + 834, + 1975 + ], + "score": 0.89, + "latex": "3 2 \\times 3 2" + }, + { + "category_id": 13, + "poly": [ + 937, + 638, + 970, + 638, + 970, + 667, + 937, + 667 + ], + "score": 0.88, + "latex": "\\mathcal { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 674, + 607, + 708, + 607, + 708, + 637, + 674, + 637 + ], + "score": 0.88, + "latex": "\\mathcal { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 601, + 1566, + 635, + 1566, + 635, + 1596, + 601, + 1596 + ], + "score": 0.88, + "latex": "\\mathcal { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1075, + 403, + 1075, + 403, + 1100, + 371, + 1100 + ], + "score": 0.85, + "latex": "w _ { i }" + }, + { + "category_id": 13, + "poly": [ + 382, + 471, + 532, + 471, + 532, + 499, + 382, + 499 + ], + "score": 0.84, + "latex": "i = 1 , \\dots , T _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1080, + 1103, + 1098, + 1103, + 1098, + 1127, + 1080, + 1127 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1314, + 281, + 1345, + 281, + 1345, + 312, + 1314, + 312 + ], + "score": 0.79, + "latex": "T _ { c }" + }, + { + "category_id": 13, + "poly": [ + 338, + 407, + 396, + 407, + 396, + 436, + 338, + 436 + ], + "score": 0.79, + "latex": "l = 2" + }, + { + "category_id": 13, + "poly": [ + 665, + 371, + 690, + 371, + 690, + 397, + 665, + 397 + ], + "score": 0.77, + "latex": "\\Theta" + }, + { + "category_id": 13, + "poly": [ + 395, + 547, + 419, + 547, + 419, + 574, + 395, + 574 + ], + "score": 0.7, + "latex": "\\Theta" + }, + { + "category_id": 13, + "poly": [ + 452, + 409, + 474, + 409, + 474, + 435, + 452, + 435 + ], + "score": 0.57, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 297, + 1945, + 344, + 1945, + 344, + 1976, + 297, + 1976 + ], + "score": 0.46, + "latex": "5 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 574, + 1946, + 620, + 1946, + 620, + 1975, + 574, + 1975 + ], + "score": 0.26, + "latex": "1 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1197.0, + 882.0, + 1197.0, + 882.0, + 1251.0, + 290.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1797.0, + 802.0, + 1797.0, + 802.0, + 1834.0, + 294.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1726.0, + 537.0, + 1726.0, + 537.0, + 1775.0, + 290.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2092.0, + 838.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 200.0, + 1126.0, + 200.0, + 1126.0, + 255.0, + 293.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1855.0, + 1400.0, + 1855.0, + 1400.0, + 1884.0, + 297.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1882.0, + 1406.0, + 1882.0, + 1406.0, + 1920.0, + 293.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1915.0, + 1405.0, + 1915.0, + 1405.0, + 1948.0, + 296.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 1944.0, + 573.0, + 1944.0, + 573.0, + 1982.0, + 345.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1944.0, + 833.0, + 1944.0, + 833.0, + 1982.0, + 621.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1944.0, + 1408.0, + 1944.0, + 1408.0, + 1982.0, + 927.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1974.0, + 1409.0, + 1974.0, + 1409.0, + 2015.0, + 293.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1566.0, + 600.0, + 1566.0, + 600.0, + 1598.0, + 297.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1566.0, + 1025.0, + 1566.0, + 1025.0, + 1598.0, + 636.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1566.0, + 1403.0, + 1566.0, + 1403.0, + 1598.0, + 1067.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1595.0, + 1402.0, + 1595.0, + 1402.0, + 1628.0, + 298.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1624.0, + 1403.0, + 1624.0, + 1403.0, + 1660.0, + 297.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1657.0, + 856.0, + 1657.0, + 856.0, + 1690.0, + 296.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1068.0, + 370.0, + 1068.0, + 370.0, + 1104.0, + 296.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1068.0, + 961.0, + 1068.0, + 961.0, + 1104.0, + 404.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1099.0, + 1079.0, + 1099.0, + 1079.0, + 1134.0, + 296.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 1099.0, + 1406.0, + 1099.0, + 1406.0, + 1134.0, + 1099.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1132.0, + 297.0, + 1132.0, + 297.0, + 1177.0, + 294.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1132.0, + 1062.0, + 1132.0, + 1062.0, + 1177.0, + 528.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1263.0, + 1405.0, + 1263.0, + 1405.0, + 1295.0, + 295.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1292.0, + 1405.0, + 1292.0, + 1405.0, + 1327.0, + 294.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1322.0, + 1409.0, + 1322.0, + 1409.0, + 1358.0, + 294.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1352.0, + 1123.0, + 1352.0, + 1123.0, + 1387.0, + 294.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 778.0, + 1406.0, + 778.0, + 1406.0, + 816.0, + 295.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 812.0, + 1404.0, + 812.0, + 1404.0, + 845.0, + 297.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 841.0, + 1405.0, + 841.0, + 1405.0, + 875.0, + 296.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 866.0, + 1407.0, + 866.0, + 1407.0, + 910.0, + 292.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 902.0, + 1405.0, + 902.0, + 1405.0, + 936.0, + 296.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 249.0, + 848.0, + 249.0, + 848.0, + 285.0, + 296.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 249.0, + 1335.0, + 249.0, + 1335.0, + 285.0, + 1071.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 276.0, + 952.0, + 276.0, + 952.0, + 319.0, + 439.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 276.0, + 1313.0, + 276.0, + 1313.0, + 319.0, + 1080.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 276.0, + 1351.0, + 276.0, + 1351.0, + 319.0, + 1346.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 310.0, + 712.0, + 310.0, + 712.0, + 345.0, + 293.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 340.0, + 338.0, + 340.0, + 338.0, + 375.0, + 295.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 340.0, + 534.0, + 340.0, + 534.0, + 375.0, + 493.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 366.0, + 664.0, + 366.0, + 664.0, + 406.0, + 297.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 366.0, + 842.0, + 366.0, + 842.0, + 406.0, + 691.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 408.0, + 337.0, + 408.0, + 337.0, + 438.0, + 296.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 408.0, + 451.0, + 408.0, + 451.0, + 438.0, + 397.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 408.0, + 514.0, + 408.0, + 514.0, + 438.0, + 475.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 432.0, + 411.0, + 432.0, + 411.0, + 475.0, + 333.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 432.0, + 780.0, + 432.0, + 780.0, + 475.0, + 595.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 432.0, + 951.0, + 432.0, + 951.0, + 475.0, + 941.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 466.0, + 381.0, + 466.0, + 381.0, + 501.0, + 337.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 466.0, + 572.0, + 466.0, + 572.0, + 501.0, + 533.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 495.0, + 708.0, + 495.0, + 708.0, + 535.0, + 339.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 495.0, + 1020.0, + 495.0, + 1020.0, + 535.0, + 869.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 543.0, + 394.0, + 543.0, + 394.0, + 577.0, + 295.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 543.0, + 433.0, + 543.0, + 433.0, + 577.0, + 420.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 573.0, + 338.0, + 573.0, + 338.0, + 608.0, + 295.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 573.0, + 532.0, + 573.0, + 532.0, + 608.0, + 492.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 598.0, + 673.0, + 598.0, + 673.0, + 643.0, + 333.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 598.0, + 909.0, + 598.0, + 909.0, + 643.0, + 709.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 598.0, + 1114.0, + 598.0, + 1114.0, + 643.0, + 950.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 598.0, + 1328.0, + 598.0, + 1328.0, + 643.0, + 1198.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 632.0, + 586.0, + 632.0, + 586.0, + 672.0, + 331.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 632.0, + 936.0, + 632.0, + 936.0, + 672.0, + 669.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 632.0, + 1020.0, + 632.0, + 1020.0, + 672.0, + 971.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 632.0, + 1073.0, + 632.0, + 1073.0, + 672.0, + 1060.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 669.0, + 648.0, + 669.0, + 648.0, + 710.0, + 294.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 669.0, + 742.0, + 669.0, + 742.0, + 710.0, + 731.0, + 710.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 301, + 290, + 1400, + 290, + 1400, + 528, + 301, + 528 + ], + "score": 0.984, + "html": "
MetricsMethodsSym-20%Sym-50%Pair-45%Inst-20%Inst-40%
Test AccuracyEarly StoppingPES82.55±2.4685.87±1.5970.76±1.2475.87±1.3360.62±5.5962.40±2.3484.41±0.9086.58±0.4574.73±2.6577.07±1.18
Label PrecisionEarly StoppingPES98.81±0.1598.96±0.0994.65±0.1995.46±0.1472.53±5.2672.99±2.2798.70±0.4398.52±0.1990.77±1.8790.63±0.92
Label RecallEarly StoppingPES88.51±2.2692.67±1.4375.18±1.0081.03±1.8367.84±5.0671.06±2.2790.37±1.0193.24±0.6082.15±3.1785.91±0.68
" + }, + { + "category_id": 1, + "poly": [ + 297, + 838, + 1404, + 838, + 1404, + 1053, + 297, + 1053 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1460, + 1406, + 1460, + 1406, + 1643, + 298, + 1643 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1231, + 1404, + 1231, + 1404, + 1446, + 297, + 1446 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 579, + 1404, + 579, + 1404, + 825, + 297, + 825 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1656, + 1405, + 1656, + 1405, + 1811, + 298, + 1811 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1066, + 1403, + 1066, + 1403, + 1219, + 298, + 1219 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1914, + 1403, + 1914, + 1403, + 2007, + 299, + 2007 + ], + "score": 0.972 + }, + { + "category_id": 6, + "poly": [ + 296, + 215, + 1402, + 215, + 1402, + 278, + 296, + 278 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 299, + 1854, + 664, + 1854, + 664, + 1888, + 299, + 1888 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 858, + 2062, + 858, + 2085, + 840, + 2085 + ], + "score": 0.785 + }, + { + "category_id": 13, + "poly": [ + 838, + 1747, + 945, + 1747, + 945, + 1778, + 838, + 1778 + ], + "score": 0.92, + "latex": "5 \\times 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1216, + 1777, + 1323, + 1777, + 1323, + 1808, + 1216, + 1808 + ], + "score": 0.92, + "latex": "5 \\times 1 0 ^ { - 6 }" + }, + { + "category_id": 13, + "poly": [ + 552, + 1581, + 659, + 1581, + 659, + 1611, + 552, + 1611 + ], + "score": 0.92, + "latex": "5 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 1039, + 1550, + 1148, + 1550, + 1148, + 1581, + 1039, + 1581 + ], + "score": 0.91, + "latex": "2 \\times \\mathrm { \\bar { 1 0 ^ { - 2 } } }" + }, + { + "category_id": 13, + "poly": [ + 1288, + 1550, + 1396, + 1550, + 1396, + 1580, + 1288, + 1580 + ], + "score": 0.91, + "latex": "2 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 546, + 1747, + 605, + 1747, + 605, + 1777, + 546, + 1777 + ], + "score": 0.91, + "latex": "1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1265, + 1581, + 1325, + 1581, + 1325, + 1610, + 1265, + 1610 + ], + "score": 0.9, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 1003, + 1383, + 1063, + 1383, + 1063, + 1413, + 1003, + 1413 + ], + "score": 0.9, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 736, + 1413, + 795, + 1413, + 795, + 1443, + 736, + 1443 + ], + "score": 0.9, + "latex": "1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 490, + 1522, + 522, + 1522, + 522, + 1551, + 490, + 1551 + ], + "score": 0.89, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1120, + 1386, + 1151, + 1386, + 1151, + 1415, + 1120, + 1415 + ], + "score": 0.88, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 1386, + 1233, + 1386, + 1233, + 1415, + 1202, + 1415 + ], + "score": 0.88, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 707, + 1324, + 739, + 1324, + 739, + 1354, + 707, + 1354 + ], + "score": 0.88, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1689, + 1148, + 1689, + 1148, + 1718, + 1115, + 1718 + ], + "score": 0.87, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1367, + 1779, + 1398, + 1779, + 1398, + 1809, + 1367, + 1809 + ], + "score": 0.87, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1367, + 1583, + 1398, + 1583, + 1398, + 1612, + 1367, + 1612 + ], + "score": 0.87, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1689, + 1067, + 1689, + 1067, + 1718, + 1035, + 1718 + ], + "score": 0.87, + "latex": "T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1035, + 1294, + 1068, + 1294, + 1068, + 1324, + 1035, + 1324 + ], + "score": 0.86, + "latex": "T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 565, + 1324, + 598, + 1324, + 598, + 1354, + 565, + 1354 + ], + "score": 0.86, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 881, + 1491, + 913, + 1491, + 913, + 1521, + 881, + 1521 + ], + "score": 0.85, + "latex": "T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 614, + 1127, + 687, + 1127, + 687, + 1156, + 614, + 1156 + ], + "score": 0.73, + "latex": "\\mathrm { E L R + }" + }, + { + "category_id": 13, + "poly": [ + 678, + 460, + 794, + 460, + 794, + 486, + 678, + 486 + ], + "score": 0.51, + "latex": "8 8 . 5 1 { \\pm } 2 . 2 6 " + }, + { + "category_id": 13, + "poly": [ + 1011, + 992, + 1083, + 992, + 1083, + 1020, + 1011, + 1020 + ], + "score": 0.5, + "latex": "\\mathrm { E L R + }" + }, + { + "category_id": 13, + "poly": [ + 824, + 460, + 941, + 460, + 941, + 486, + 824, + 486 + ], + "score": 0.48, + "latex": "7 5 . 1 8 { \\pm } 1 . 0 0" + }, + { + "category_id": 13, + "poly": [ + 1118, + 460, + 1232, + 460, + 1232, + 486, + 1118, + 486 + ], + "score": 0.47, + "latex": "9 0 . 3 7 { \\pm } 1 . 0 1 " + }, + { + "category_id": 13, + "poly": [ + 678, + 395, + 793, + 395, + 793, + 421, + 678, + 421 + ], + "score": 0.47, + "latex": "9 8 . 8 1 { \\pm } 0 . 1 5 " + }, + { + "category_id": 13, + "poly": [ + 1238, + 763, + 1284, + 763, + 1284, + 792, + 1238, + 792 + ], + "score": 0.46, + "latex": "1 5 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 824, + 395, + 940, + 395, + 940, + 421, + 824, + 421 + ], + "score": 0.44, + "latex": "9 4 . 6 5 { \\pm } 0 . 1 9" + }, + { + "category_id": 13, + "poly": [ + 971, + 492, + 1087, + 492, + 1087, + 519, + 971, + 519 + ], + "score": 0.42, + "latex": "{ \\bf 7 1 . 0 6 } \\pm 2 . 2 7" + }, + { + "category_id": 13, + "poly": [ + 970, + 427, + 1087, + 427, + 1087, + 453, + 970, + 453 + ], + "score": 0.35, + "latex": "7 2 . 9 9 { \\scriptstyle \\pm 2 . 2 7 }" + }, + { + "category_id": 13, + "poly": [ + 677, + 492, + 794, + 492, + 794, + 519, + 677, + 519 + ], + "score": 0.3, + "latex": "\\mathbf { 9 } 2 . 6 7 { \\pm } 1 . 4 3 " + }, + { + "category_id": 13, + "poly": [ + 970, + 460, + 1087, + 460, + 1087, + 486, + 970, + 486 + ], + "score": 0.28, + "latex": "6 7 . 8 4 \\pm 5 . 0 6" + }, + { + "category_id": 13, + "poly": [ + 1118, + 396, + 1233, + 396, + 1233, + 420, + 1118, + 420 + ], + "score": 0.26, + "latex": "\\mathbf { 9 8 . 7 0 { \\scriptstyle \\pm 0 . 4 3 } }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 215.0, + 1404.0, + 215.0, + 1404.0, + 252.0, + 293.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 243.0, + 1124.0, + 243.0, + 1124.0, + 282.0, + 295.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 667.0, + 1849.0, + 667.0, + 1895.0, + 292.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2060.0, + 861.0, + 2060.0, + 861.0, + 2091.0, + 840.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 836.0, + 1405.0, + 836.0, + 1405.0, + 875.0, + 292.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 870.0, + 1404.0, + 870.0, + 1404.0, + 905.0, + 295.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 898.0, + 1405.0, + 898.0, + 1405.0, + 936.0, + 294.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 927.0, + 1405.0, + 927.0, + 1405.0, + 968.0, + 292.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 959.0, + 1406.0, + 959.0, + 1406.0, + 995.0, + 293.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 991.0, + 1010.0, + 991.0, + 1010.0, + 1025.0, + 295.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 991.0, + 1405.0, + 991.0, + 1405.0, + 1025.0, + 1084.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1021.0, + 1255.0, + 1021.0, + 1255.0, + 1056.0, + 295.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1459.0, + 1409.0, + 1459.0, + 1409.0, + 1495.0, + 295.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1490.0, + 880.0, + 1490.0, + 880.0, + 1525.0, + 295.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1490.0, + 1407.0, + 1490.0, + 1407.0, + 1525.0, + 914.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1517.0, + 489.0, + 1517.0, + 489.0, + 1554.0, + 293.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1517.0, + 1406.0, + 1517.0, + 1406.0, + 1554.0, + 523.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1549.0, + 1038.0, + 1549.0, + 1038.0, + 1584.0, + 292.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1549.0, + 1287.0, + 1549.0, + 1287.0, + 1584.0, + 1149.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1549.0, + 1407.0, + 1549.0, + 1407.0, + 1584.0, + 1397.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1578.0, + 551.0, + 1578.0, + 551.0, + 1617.0, + 293.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1578.0, + 1264.0, + 1578.0, + 1264.0, + 1617.0, + 660.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1578.0, + 1366.0, + 1578.0, + 1366.0, + 1617.0, + 1326.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1578.0, + 1412.0, + 1578.0, + 1412.0, + 1617.0, + 1399.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1611.0, + 1265.0, + 1611.0, + 1265.0, + 1646.0, + 295.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1231.0, + 1405.0, + 1231.0, + 1405.0, + 1266.0, + 295.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1262.0, + 1406.0, + 1262.0, + 1406.0, + 1297.0, + 295.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1293.0, + 1034.0, + 1293.0, + 1034.0, + 1327.0, + 292.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1293.0, + 1406.0, + 1293.0, + 1406.0, + 1327.0, + 1069.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1321.0, + 564.0, + 1321.0, + 564.0, + 1357.0, + 292.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1321.0, + 706.0, + 1321.0, + 706.0, + 1357.0, + 599.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1321.0, + 1404.0, + 1321.0, + 1404.0, + 1357.0, + 740.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1352.0, + 1404.0, + 1352.0, + 1404.0, + 1388.0, + 292.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1379.0, + 1002.0, + 1379.0, + 1002.0, + 1421.0, + 291.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1379.0, + 1119.0, + 1379.0, + 1119.0, + 1421.0, + 1064.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1379.0, + 1201.0, + 1379.0, + 1201.0, + 1421.0, + 1152.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1379.0, + 1406.0, + 1379.0, + 1406.0, + 1421.0, + 1234.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1411.0, + 735.0, + 1411.0, + 735.0, + 1448.0, + 292.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1411.0, + 806.0, + 1411.0, + 806.0, + 1448.0, + 796.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 581.0, + 1405.0, + 581.0, + 1405.0, + 616.0, + 294.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 612.0, + 1405.0, + 612.0, + 1405.0, + 644.0, + 294.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 643.0, + 1407.0, + 643.0, + 1407.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 670.0, + 1406.0, + 670.0, + 1406.0, + 708.0, + 292.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 704.0, + 1402.0, + 704.0, + 1402.0, + 734.0, + 296.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 733.0, + 1406.0, + 733.0, + 1406.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 762.0, + 1237.0, + 762.0, + 1237.0, + 799.0, + 294.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 762.0, + 1406.0, + 762.0, + 1406.0, + 799.0, + 1285.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 795.0, + 814.0, + 795.0, + 814.0, + 826.0, + 295.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1658.0, + 1405.0, + 1658.0, + 1405.0, + 1691.0, + 296.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1687.0, + 1034.0, + 1687.0, + 1034.0, + 1720.0, + 295.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1687.0, + 1114.0, + 1687.0, + 1114.0, + 1720.0, + 1068.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1687.0, + 1406.0, + 1687.0, + 1406.0, + 1720.0, + 1149.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1715.0, + 1408.0, + 1715.0, + 1408.0, + 1753.0, + 291.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1744.0, + 545.0, + 1744.0, + 545.0, + 1784.0, + 292.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1744.0, + 837.0, + 1744.0, + 837.0, + 1784.0, + 606.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 1744.0, + 1407.0, + 1744.0, + 1407.0, + 1784.0, + 946.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1775.0, + 1215.0, + 1775.0, + 1215.0, + 1816.0, + 293.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 1775.0, + 1366.0, + 1775.0, + 1366.0, + 1816.0, + 1324.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1775.0, + 1410.0, + 1775.0, + 1410.0, + 1816.0, + 1399.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1067.0, + 1404.0, + 1067.0, + 1404.0, + 1100.0, + 296.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1096.0, + 1405.0, + 1096.0, + 1405.0, + 1133.0, + 294.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1124.0, + 613.0, + 1124.0, + 613.0, + 1162.0, + 292.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1124.0, + 1408.0, + 1124.0, + 1408.0, + 1162.0, + 688.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1157.0, + 1405.0, + 1157.0, + 1405.0, + 1194.0, + 294.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1187.0, + 1177.0, + 1187.0, + 1177.0, + 1224.0, + 294.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1950.0, + 293.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1945.0, + 1402.0, + 1945.0, + 1402.0, + 1979.0, + 295.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1974.0, + 1408.0, + 1974.0, + 1408.0, + 2015.0, + 293.0, + 2015.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 324, + 289, + 1368, + 289, + 1368, + 883, + 324, + 883 + ], + "score": 0.986, + "html": "
DatasetMethodSymmetricPairflipInstance
20%50%45%20%40%
CIFAR10CE84.00±0.6675.51±1.2463.34±6.0385.10±0.6877.00±2.17
Co-teaching87.16±0.1172.80±0.4570.11±1.1686.54±0.1180.98±0.39
Forward85.63±0.5277.92±0.6660.15±1.9785.29±0.3874.72±3.24
Joint Optim89.70±0.1185.00±0.1782.63±1.3889.69±0.4282.62±0.57
T-revision89.63±0.1383.40±0.6577.06±6.4790.46±0.1385.37±3.36
DMI88.18±0.3678.28±0.4857.60±14.5689.14±0.3684.78±1.97
CDR89.72±0.3882.64±0.8973.67±0.5490.41±0.3483.07±1.33
Ours92.38±0.4087.45±0.3588.43±1.0892.69±0.4489.73±0.51
CIFAR100CE51.43±0.5837.69±3.4534.10±2.0452.19±1.4242.26±1.29
Co-teaching59.28±0.4741.37±0.0833.22±0.4857.24±0.6945.69±0.99
Forward57.75±0.3744.66±1.0127.88±0.8058.76±0.6644.50±0.72
Joint Optim64.55±0.3850.22±0.4142.61±0.6165.15±0.3155.57±0.41
T-revision65.40±1.0750.24±1.4541.10±1.9560.71±0.7351.54±0.91
DMI58.73±0.7044.25±1.1426.90±0.4558.05±0.2047.36±0.68
CDR66.52±0.2455.30±0.9643.87±1.3567.33±0.6755.94±0.56
Ours68.89±0.4558.90±2.7257.18±1.4470.49±0.7965.68±1.41
" + }, + { + "category_id": 5, + "poly": [ + 366, + 1021, + 1335, + 1021, + 1335, + 1320, + 366, + 1320 + ], + "score": 0.984, + "html": "
DatasetCIFAR-10CIFAR-100
Methods /NoiseSym-20%Sym-50%Sym-80%Sym-20%Sym-50%Sym-80%
CE86.5±0.680.6±0.263.7±0.857.9±0.447.3±0.222.3±1.2
MixUp93.2±0.388.2±0.373.3±0.369.5±0.257.1±0.634.1±0.6
M-correction*94.092.086.873.966.148.2
DivideMix*95.294.293.075.272.858.3
DivideMix95.6±0.194.6±0.192.9±0.375.3±0.172.7±0.656.4±0.3
ELR+94.9±0.293.6±0.190.4±0.275.5±0.271.0±0.250.4±0.8
Ours (Semi)95.9±0.195.1±0.293.1±0.277.4±0.374.3±0.661.6±0.6
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1546, + 1404, + 1546, + 1404, + 1761, + 298, + 1761 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1854, + 1403, + 1854, + 1403, + 2007, + 299, + 2007 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1349, + 1405, + 1349, + 1405, + 1533, + 297, + 1533 + ], + "score": 0.971 + }, + { + "category_id": 6, + "poly": [ + 297, + 216, + 1403, + 216, + 1403, + 279, + 297, + 279 + ], + "score": 0.931 + }, + { + "category_id": 0, + "poly": [ + 300, + 1797, + 779, + 1797, + 779, + 1831, + 300, + 1831 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.713 + }, + { + "category_id": 6, + "poly": [ + 302, + 915, + 1403, + 915, + 1403, + 1009, + 302, + 1009 + ], + "score": 0.662 + }, + { + "category_id": 6, + "poly": [ + 302, + 916, + 1404, + 916, + 1404, + 1009, + 302, + 1009 + ], + "score": 0.319 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2084, + 841, + 2084 + ], + "score": 0.096 + }, + { + "category_id": 13, + "poly": [ + 510, + 1915, + 564, + 1915, + 564, + 1944, + 510, + 1944 + ], + "score": 0.86, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 962, + 1253, + 1055, + 1253, + 1055, + 1279, + 962, + 1279 + ], + "score": 0.58, + "latex": "7 5 . 5 { \\pm } 0 . 2 " + }, + { + "category_id": 13, + "poly": [ + 833, + 1253, + 926, + 1253, + 926, + 1279, + 833, + 1279 + ], + "score": 0.54, + "latex": "9 0 . 4 \\pm 0 . 2 " + }, + { + "category_id": 13, + "poly": [ + 962, + 1220, + 1055, + 1220, + 1055, + 1247, + 962, + 1247 + ], + "score": 0.53, + "latex": "7 5 . 3 { \\pm } 0 . 1 " + }, + { + "category_id": 13, + "poly": [ + 833, + 1220, + 926, + 1220, + 926, + 1247, + 833, + 1247 + ], + "score": 0.52, + "latex": "9 2 . 9 { \\pm } 0 . 3 " + }, + { + "category_id": 13, + "poly": [ + 705, + 1253, + 799, + 1253, + 799, + 1279, + 705, + 1279 + ], + "score": 0.5, + "latex": "9 3 . 6 { \\pm } 0 . 1 " + }, + { + "category_id": 13, + "poly": [ + 704, + 1221, + 798, + 1221, + 798, + 1247, + 704, + 1247 + ], + "score": 0.49, + "latex": "9 4 . 6 { \\pm } 0 . 1 " + }, + { + "category_id": 13, + "poly": [ + 577, + 1220, + 669, + 1220, + 669, + 1247, + 577, + 1247 + ], + "score": 0.48, + "latex": "9 5 . 6 { \\pm } 0 . 1 " + }, + { + "category_id": 13, + "poly": [ + 1090, + 1220, + 1184, + 1220, + 1184, + 1247, + 1090, + 1247 + ], + "score": 0.47, + "latex": "7 2 . 7 { \\pm } 0 . 6 " + }, + { + "category_id": 13, + "poly": [ + 632, + 719, + 748, + 719, + 748, + 746, + 632, + 746 + ], + "score": 0.4, + "latex": "6 4 . 5 5 { \\pm } 0 . 3 8 " + }, + { + "category_id": 13, + "poly": [ + 577, + 1253, + 670, + 1253, + 670, + 1279, + 577, + 1279 + ], + "score": 0.34, + "latex": "9 4 . 9 { \\pm } 0 . 2 " + }, + { + "category_id": 13, + "poly": [ + 631, + 655, + 748, + 655, + 748, + 682, + 631, + 682 + ], + "score": 0.32, + "latex": "5 9 . 2 8 { \\scriptstyle \\pm 0 . 4 7 }" + }, + { + "category_id": 13, + "poly": [ + 1219, + 1253, + 1313, + 1253, + 1313, + 1279, + 1219, + 1279 + ], + "score": 0.28, + "latex": "5 0 . 4 { \\pm } 0 . 8 " + }, + { + "category_id": 13, + "poly": [ + 1219, + 1220, + 1312, + 1220, + 1312, + 1247, + 1219, + 1247 + ], + "score": 0.28, + "latex": "5 6 . 4 { \\pm } 0 . 3 " + }, + { + "category_id": 13, + "poly": [ + 779, + 847, + 895, + 847, + 895, + 874, + 779, + 874 + ], + "score": 0.28, + "latex": "{ \\pm 8 . 9 0 { \\pm } 2 . 7 2 }" + }, + { + "category_id": 13, + "poly": [ + 779, + 752, + 895, + 752, + 895, + 778, + 779, + 778 + ], + "score": 0.27, + "latex": "5 0 . 2 4 \\pm 1 . 4 5" + }, + { + "category_id": 13, + "poly": [ + 1091, + 1253, + 1184, + 1253, + 1184, + 1279, + 1091, + 1279 + ], + "score": 0.27, + "latex": "7 1 . 0 { \\pm } 0 . 2 " + }, + { + "category_id": 13, + "poly": [ + 705, + 1286, + 798, + 1286, + 798, + 1312, + 705, + 1312 + ], + "score": 0.26, + "latex": "{ \\bf 9 5 . 1 } { \\bf \\pm 0 . 2 }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 217.0, + 1404.0, + 217.0, + 1404.0, + 249.0, + 296.0, + 249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 244.0, + 1295.0, + 244.0, + 1295.0, + 281.0, + 295.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1794.0, + 779.0, + 1794.0, + 779.0, + 1836.0, + 292.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 916.0, + 1405.0, + 916.0, + 1405.0, + 950.0, + 296.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 945.0, + 1404.0, + 945.0, + 1404.0, + 980.0, + 296.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 975.0, + 958.0, + 975.0, + 958.0, + 1013.0, + 295.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 917.0, + 1405.0, + 917.0, + 1405.0, + 950.0, + 297.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 947.0, + 1404.0, + 947.0, + 1404.0, + 981.0, + 297.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 976.0, + 957.0, + 976.0, + 957.0, + 1013.0, + 296.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2092.0, + 840.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1548.0, + 1405.0, + 1548.0, + 1405.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1576.0, + 1409.0, + 1576.0, + 1409.0, + 1612.0, + 293.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1608.0, + 1404.0, + 1608.0, + 1404.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1639.0, + 1405.0, + 1639.0, + 1405.0, + 1673.0, + 294.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1668.0, + 1404.0, + 1668.0, + 1404.0, + 1703.0, + 294.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1699.0, + 1406.0, + 1699.0, + 1406.0, + 1733.0, + 294.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1731.0, + 772.0, + 1731.0, + 772.0, + 1762.0, + 294.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1855.0, + 1404.0, + 1855.0, + 1404.0, + 1888.0, + 297.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1920.0, + 294.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1915.0, + 509.0, + 1915.0, + 509.0, + 1948.0, + 295.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1915.0, + 1404.0, + 1915.0, + 1404.0, + 1948.0, + 565.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1407.0, + 1942.0, + 1407.0, + 1982.0, + 293.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1978.0, + 680.0, + 1978.0, + 680.0, + 2011.0, + 297.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1351.0, + 1403.0, + 1351.0, + 1403.0, + 1383.0, + 296.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1380.0, + 1406.0, + 1380.0, + 1406.0, + 1416.0, + 295.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1411.0, + 1407.0, + 1411.0, + 1407.0, + 1446.0, + 295.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1441.0, + 1407.0, + 1441.0, + 1407.0, + 1474.0, + 292.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1468.0, + 1407.0, + 1468.0, + 1407.0, + 1508.0, + 291.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1502.0, + 581.0, + 1502.0, + 581.0, + 1535.0, + 292.0, + 1535.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1511, + 1405, + 1511, + 1405, + 1816, + 298, + 1816 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1283, + 1404, + 1283, + 1404, + 1497, + 297, + 1497 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 379, + 321, + 1316, + 321, + 1316, + 557, + 379, + 557 + ], + "score": 0.98, + "html": "
DatasetCIFAR-10CIFAR-100
Methods/NoiseInst-20%Inst-40%Pair-45%Inst-20%Inst-40%Pair-45%
CEMixUpDivideMixELR+87.5±0.578.9±0.774.9±1.756.8±0.448.2±0.538.5±0.6
93.3±0.295.5±0.187.6±0.594.5±0.282.4±1.085.6±1.767.1±0.155.0±0.170.9±0.144.2±0.548.2±1.0
75.2±0.270.9±0.1
94.9±0.194.3±0.286.1±1.275.8±0.174.3±0.365.3±1.3
Ours (Semi)95.9±0.195.3±0.194.5±0.377.6±0.376.1±0.473.6±1.7
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1915, + 1403, + 1915, + 1403, + 2008, + 300, + 2008 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 485, + 852, + 1212, + 852, + 1212, + 1197, + 485, + 1197 + ], + "score": 0.97 + }, + { + "category_id": 5, + "poly": [ + 353, + 714, + 1346, + 714, + 1346, + 787, + 353, + 787 + ], + "score": 0.96, + "html": "
CEForwardJoint-OptimDMIT-revisionDivideMix*ELR+*OursOurs*
69.2169.8472.1672.4674.1874.7674.8174.6474.99
" + }, + { + "category_id": 4, + "poly": [ + 397, + 1214, + 1296, + 1214, + 1296, + 1249, + 397, + 1249 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 299, + 1857, + 593, + 1857, + 593, + 1889, + 299, + 1889 + ], + "score": 0.921 + }, + { + "category_id": 6, + "poly": [ + 298, + 216, + 1404, + 216, + 1404, + 308, + 298, + 308 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2085, + 841, + 2085 + ], + "score": 0.799 + }, + { + "category_id": 6, + "poly": [ + 299, + 608, + 1404, + 608, + 1404, + 700, + 299, + 700 + ], + "score": 0.759 + }, + { + "category_id": 1, + "poly": [ + 299, + 608, + 1404, + 608, + 1404, + 700, + 299, + 700 + ], + "score": 0.226 + }, + { + "category_id": 13, + "poly": [ + 1323, + 1916, + 1355, + 1916, + 1355, + 1946, + 1323, + 1946 + ], + "score": 0.88, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1175, + 1217, + 1207, + 1217, + 1207, + 1246, + 1175, + 1246 + ], + "score": 0.87, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1258, + 1217, + 1289, + 1217, + 1289, + 1246, + 1258, + 1246 + ], + "score": 0.87, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1979, + 328, + 1979, + 328, + 2006, + 298, + 2006 + ], + "score": 0.86, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1948, + 328, + 1948, + 328, + 1976, + 298, + 1976 + ], + "score": 0.85, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1406, + 338, + 1406, + 338, + 1435, + 297, + 1435 + ], + "score": 0.83, + "latex": "8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1046, + 1573, + 1116, + 1573, + 1116, + 1602, + 1046, + 1602 + ], + "score": 0.71, + "latex": "\\mathrm { E L R + }" + }, + { + "category_id": 13, + "poly": [ + 710, + 489, + 803, + 489, + 803, + 514, + 710, + 514 + ], + "score": 0.49, + "latex": "9 4 . 3 { \\pm } 0 . 2 " + }, + { + "category_id": 13, + "poly": [ + 834, + 489, + 926, + 489, + 926, + 515, + 834, + 515 + ], + "score": 0.46, + "latex": "8 6 . 1 { \\pm } 1 . 2 " + }, + { + "category_id": 13, + "poly": [ + 710, + 393, + 802, + 393, + 802, + 419, + 710, + 419 + ], + "score": 0.45, + "latex": "7 8 . 9 2 0 . 7" + }, + { + "category_id": 13, + "poly": [ + 833, + 394, + 925, + 394, + 925, + 419, + 833, + 419 + ], + "score": 0.4, + "latex": "7 4 . 9 2 1 . 7" + }, + { + "category_id": 13, + "poly": [ + 957, + 489, + 1050, + 489, + 1050, + 515, + 957, + 515 + ], + "score": 0.34, + "latex": "7 5 . 8 { \\pm } 0 . 1 " + }, + { + "category_id": 13, + "poly": [ + 710, + 457, + 803, + 457, + 803, + 483, + 710, + 483 + ], + "score": 0.33, + "latex": "9 4 . 5 { \\pm } 0 . 2 " + }, + { + "category_id": 13, + "poly": [ + 833, + 425, + 927, + 425, + 927, + 451, + 833, + 451 + ], + "score": 0.32, + "latex": "8 2 . 4 \\pm 1 . 0" + }, + { + "category_id": 13, + "poly": [ + 833, + 457, + 926, + 457, + 926, + 483, + 833, + 483 + ], + "score": 0.31, + "latex": "8 5 . 6 \\pm 1 . 7" + }, + { + "category_id": 13, + "poly": [ + 755, + 361, + 801, + 361, + 801, + 386, + 755, + 386 + ], + "score": 0.26, + "latex": "40 \\%" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 875.0, + 527.0, + 875.0, + 527.0, + 894.0, + 503.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 874.0, + 892.0, + 874.0, + 892.0, + 895.0, + 864.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 934.0, + 527.0, + 934.0, + 527.0, + 1040.0, + 486.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 946.0, + 841.0, + 946.0, + 841.0, + 971.0, + 686.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 935.0, + 889.0, + 935.0, + 889.0, + 1042.0, + 843.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 946.0, + 1202.0, + 946.0, + 1202.0, + 971.0, + 1049.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 963.0, + 824.0, + 963.0, + 824.0, + 991.0, + 686.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 975.0, + 1041.0, + 975.0, + 1041.0, + 982.0, + 1029.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 963.0, + 1188.0, + 963.0, + 1188.0, + 991.0, + 1049.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 992.0, + 527.0, + 992.0, + 527.0, + 1014.0, + 506.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 982.0, + 818.0, + 982.0, + 818.0, + 1007.0, + 685.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 993.0, + 890.0, + 993.0, + 890.0, + 1013.0, + 867.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 983.0, + 1182.0, + 983.0, + 1182.0, + 1009.0, + 1049.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1011.0, + 677.0, + 1011.0, + 677.0, + 1022.0, + 659.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1003.0, + 833.0, + 1003.0, + 833.0, + 1027.0, + 686.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1011.0, + 1042.0, + 1011.0, + 1042.0, + 1022.0, + 1027.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1003.0, + 1194.0, + 1003.0, + 1194.0, + 1027.0, + 1049.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1052.0, + 527.0, + 1052.0, + 527.0, + 1072.0, + 504.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1066.0, + 693.0, + 1066.0, + 693.0, + 1078.0, + 678.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1063.0, + 777.0, + 1063.0, + 777.0, + 1077.0, + 748.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1068.0, + 836.0, + 1068.0, + 836.0, + 1079.0, + 823.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1052.0, + 890.0, + 1052.0, + 890.0, + 1072.0, + 865.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1061.0, + 1005.0, + 1061.0, + 1005.0, + 1077.0, + 902.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1062.0, + 1071.0, + 1062.0, + 1071.0, + 1076.0, + 1038.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1060.0, + 1144.0, + 1060.0, + 1144.0, + 1076.0, + 1090.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1079.0, + 620.0, + 1079.0, + 620.0, + 1087.0, + 608.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1111.0, + 528.0, + 1111.0, + 528.0, + 1130.0, + 505.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1109.0, + 892.0, + 1109.0, + 892.0, + 1132.0, + 866.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 1121.0, + 549.0, + 1121.0, + 549.0, + 1140.0, + 533.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1121.0, + 622.0, + 1121.0, + 622.0, + 1141.0, + 606.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1119.0, + 697.0, + 1119.0, + 697.0, + 1142.0, + 677.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1121.0, + 766.0, + 1121.0, + 766.0, + 1139.0, + 750.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1119.0, + 846.0, + 1119.0, + 846.0, + 1141.0, + 816.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1121.0, + 913.0, + 1121.0, + 913.0, + 1140.0, + 896.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1120.0, + 987.0, + 1120.0, + 987.0, + 1142.0, + 968.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1119.0, + 1060.0, + 1119.0, + 1060.0, + 1142.0, + 1039.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1123.0, + 1128.0, + 1123.0, + 1128.0, + 1138.0, + 1116.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1119.0, + 1207.0, + 1119.0, + 1207.0, + 1142.0, + 1179.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1135.0, + 715.0, + 1135.0, + 715.0, + 1162.0, + 656.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1135.0, + 1079.0, + 1135.0, + 1079.0, + 1162.0, + 1018.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1161.0, + 707.0, + 1161.0, + 707.0, + 1205.0, + 628.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1161.0, + 1072.0, + 1161.0, + 1072.0, + 1205.0, + 990.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1213.0, + 1174.0, + 1213.0, + 1174.0, + 1253.0, + 398.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1213.0, + 1257.0, + 1213.0, + 1257.0, + 1253.0, + 1208.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1213.0, + 1300.0, + 1213.0, + 1300.0, + 1253.0, + 1290.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1853.0, + 597.0, + 1853.0, + 597.0, + 1896.0, + 293.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 214.0, + 1404.0, + 214.0, + 1404.0, + 250.0, + 293.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 245.0, + 1404.0, + 245.0, + 1404.0, + 280.0, + 295.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 278.0, + 845.0, + 278.0, + 845.0, + 311.0, + 296.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2089.0, + 839.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 606.0, + 1404.0, + 606.0, + 1404.0, + 642.0, + 293.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 637.0, + 1403.0, + 637.0, + 1403.0, + 671.0, + 295.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 668.0, + 1056.0, + 668.0, + 1056.0, + 702.0, + 295.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1510.0, + 1406.0, + 1510.0, + 1406.0, + 1546.0, + 295.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1542.0, + 1408.0, + 1542.0, + 1408.0, + 1577.0, + 295.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1571.0, + 1045.0, + 1571.0, + 1045.0, + 1607.0, + 295.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1571.0, + 1405.0, + 1571.0, + 1405.0, + 1607.0, + 1117.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1601.0, + 1403.0, + 1601.0, + 1403.0, + 1637.0, + 295.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1634.0, + 1406.0, + 1634.0, + 1406.0, + 1666.0, + 296.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1665.0, + 1406.0, + 1665.0, + 1406.0, + 1696.0, + 296.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1693.0, + 1408.0, + 1693.0, + 1408.0, + 1728.0, + 295.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1726.0, + 1405.0, + 1726.0, + 1405.0, + 1757.0, + 296.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1753.0, + 1408.0, + 1753.0, + 1408.0, + 1789.0, + 293.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1786.0, + 1079.0, + 1786.0, + 1079.0, + 1818.0, + 296.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1285.0, + 1405.0, + 1285.0, + 1405.0, + 1319.0, + 294.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1313.0, + 1406.0, + 1313.0, + 1406.0, + 1351.0, + 292.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1346.0, + 1404.0, + 1346.0, + 1404.0, + 1380.0, + 295.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1377.0, + 1405.0, + 1377.0, + 1405.0, + 1411.0, + 294.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1403.0, + 1404.0, + 1403.0, + 1404.0, + 1440.0, + 339.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1438.0, + 1404.0, + 1438.0, + 1404.0, + 1472.0, + 295.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1469.0, + 633.0, + 1469.0, + 633.0, + 1499.0, + 296.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 1322.0, + 1913.0, + 1322.0, + 1951.0, + 294.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1913.0, + 1407.0, + 1913.0, + 1407.0, + 1951.0, + 1356.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 297.0, + 1942.0, + 297.0, + 1983.0, + 293.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1983.0, + 329.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1976.0, + 1403.0, + 1976.0, + 1403.0, + 2010.0, + 329.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 606.0, + 1404.0, + 606.0, + 1404.0, + 642.0, + 293.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 637.0, + 1403.0, + 637.0, + 1403.0, + 671.0, + 295.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 668.0, + 1056.0, + 668.0, + 1056.0, + 702.0, + 295.0, + 702.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1410, + 1406, + 1410, + 1406, + 1836, + 297, + 1836 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1091, + 1404, + 1091, + 1404, + 1396, + 298, + 1396 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 662, + 1404, + 662, + 1404, + 876, + 298, + 876 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 984, + 1403, + 984, + 1403, + 1077, + 299, + 1077 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 357, + 1404, + 357, + 1404, + 573, + 297, + 573 + ], + "score": 0.974 + }, + { + "category_id": 5, + "poly": [ + 386, + 261, + 1298, + 261, + 1298, + 334, + 386, + 334 + ], + "score": 0.96, + "html": "
CECo-teachingCDRT-revisionELR+DivideMixOursOurs (Semi)
0.9h1.5h3.0h3.5h2.2h5.5h1.0h3.1h
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1945, + 1398, + 1945, + 1398, + 2007, + 297, + 2007 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 299, + 606, + 686, + 606, + 686, + 640, + 299, + 640 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 917, + 541, + 917, + 541, + 954, + 299, + 954 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 298, + 1877, + 509, + 1877, + 509, + 1915, + 298, + 1915 + ], + "score": 0.904 + }, + { + "category_id": 6, + "poly": [ + 303, + 215, + 1381, + 215, + 1381, + 248, + 303, + 248 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 858, + 2061, + 858, + 2084, + 840, + 2084 + ], + "score": 0.8 + }, + { + "category_id": 13, + "poly": [ + 525, + 542, + 557, + 542, + 557, + 571, + 525, + 571 + ], + "score": 0.88, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 908, + 420, + 940, + 420, + 940, + 450, + 908, + 450 + ], + "score": 0.88, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 442, + 542, + 474, + 542, + 474, + 571, + 442, + 571 + ], + "score": 0.88, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 691, + 512, + 723, + 512, + 723, + 541, + 691, + 541 + ], + "score": 0.88, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 988, + 481, + 1020, + 481, + 1020, + 511, + 988, + 511 + ], + "score": 0.88, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 537, + 451, + 570, + 451, + 570, + 480, + 537, + 480 + ], + "score": 0.88, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 489, + 360, + 520, + 360, + 520, + 390, + 489, + 390 + ], + "score": 0.88, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 692, + 482, + 724, + 482, + 724, + 511, + 692, + 511 + ], + "score": 0.87, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 354, + 421, + 385, + 421, + 385, + 450, + 354, + 450 + ], + "score": 0.87, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 338, + 390, + 393, + 390, + 393, + 418, + 338, + 418 + ], + "score": 0.85, + "latex": "45 \\%" + }, + { + "category_id": 13, + "poly": [ + 1066, + 217, + 1121, + 217, + 1121, + 246, + 1066, + 246 + ], + "score": 0.83, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 329, + 815, + 404, + 815, + 404, + 844, + 329, + 844 + ], + "score": 0.65, + "latex": "\\mathrm { E L R + }" + }, + { + "category_id": 13, + "poly": [ + 691, + 815, + 766, + 815, + 766, + 843, + 691, + 843 + ], + "score": 0.62, + "latex": "\\mathrm { E L R + }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 603.0, + 690.0, + 603.0, + 690.0, + 646.0, + 293.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 914.0, + 545.0, + 914.0, + 545.0, + 959.0, + 293.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1871.0, + 514.0, + 1871.0, + 514.0, + 1922.0, + 289.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 213.0, + 1065.0, + 213.0, + 1065.0, + 252.0, + 312.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 213.0, + 1387.0, + 213.0, + 1387.0, + 252.0, + 1122.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 860.0, + 2060.0, + 860.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1408.0, + 1404.0, + 1408.0, + 1404.0, + 1446.0, + 294.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1437.0, + 1407.0, + 1437.0, + 1407.0, + 1479.0, + 291.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1469.0, + 1411.0, + 1469.0, + 1411.0, + 1507.0, + 291.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1503.0, + 1404.0, + 1503.0, + 1404.0, + 1534.0, + 295.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1529.0, + 1407.0, + 1529.0, + 1407.0, + 1569.0, + 292.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1560.0, + 1406.0, + 1560.0, + 1406.0, + 1598.0, + 294.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1593.0, + 1406.0, + 1593.0, + 1406.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1622.0, + 1407.0, + 1622.0, + 1407.0, + 1657.0, + 295.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1652.0, + 1407.0, + 1652.0, + 1407.0, + 1687.0, + 295.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 1406.0, + 1681.0, + 1406.0, + 1719.0, + 294.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1713.0, + 1404.0, + 1713.0, + 1404.0, + 1747.0, + 295.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 1406.0, + 1743.0, + 1406.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1406.0, + 1773.0, + 1406.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1804.0, + 915.0, + 1804.0, + 915.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1090.0, + 1406.0, + 1090.0, + 1406.0, + 1127.0, + 293.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1121.0, + 1406.0, + 1121.0, + 1406.0, + 1157.0, + 293.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1151.0, + 1406.0, + 1151.0, + 1406.0, + 1188.0, + 293.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1183.0, + 1406.0, + 1183.0, + 1406.0, + 1218.0, + 294.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1213.0, + 1405.0, + 1213.0, + 1405.0, + 1249.0, + 293.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1242.0, + 1406.0, + 1242.0, + 1406.0, + 1278.0, + 294.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1270.0, + 1404.0, + 1270.0, + 1404.0, + 1309.0, + 292.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1302.0, + 1404.0, + 1302.0, + 1404.0, + 1340.0, + 292.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1332.0, + 1405.0, + 1332.0, + 1405.0, + 1370.0, + 293.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1364.0, + 1284.0, + 1364.0, + 1284.0, + 1398.0, + 293.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 660.0, + 1406.0, + 660.0, + 1406.0, + 699.0, + 292.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 694.0, + 1405.0, + 694.0, + 1405.0, + 729.0, + 294.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 720.0, + 1404.0, + 720.0, + 1404.0, + 762.0, + 293.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 753.0, + 1405.0, + 753.0, + 1405.0, + 789.0, + 293.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1404.0, + 782.0, + 1404.0, + 818.0, + 293.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 813.0, + 328.0, + 813.0, + 328.0, + 851.0, + 293.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 813.0, + 690.0, + 813.0, + 690.0, + 851.0, + 405.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 813.0, + 1405.0, + 813.0, + 1405.0, + 851.0, + 767.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 845.0, + 700.0, + 845.0, + 700.0, + 879.0, + 293.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 982.0, + 1405.0, + 982.0, + 1405.0, + 1021.0, + 293.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1017.0, + 1405.0, + 1017.0, + 1405.0, + 1050.0, + 294.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1047.0, + 613.0, + 1047.0, + 613.0, + 1079.0, + 294.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 356.0, + 488.0, + 356.0, + 488.0, + 396.0, + 292.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 356.0, + 1405.0, + 356.0, + 1405.0, + 396.0, + 521.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 390.0, + 337.0, + 390.0, + 337.0, + 424.0, + 295.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 390.0, + 1405.0, + 390.0, + 1405.0, + 424.0, + 394.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 419.0, + 353.0, + 419.0, + 353.0, + 454.0, + 295.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 419.0, + 907.0, + 419.0, + 907.0, + 454.0, + 386.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 419.0, + 1405.0, + 419.0, + 1405.0, + 454.0, + 941.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 449.0, + 536.0, + 449.0, + 536.0, + 487.0, + 294.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 449.0, + 1406.0, + 449.0, + 1406.0, + 487.0, + 571.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 480.0, + 691.0, + 480.0, + 691.0, + 514.0, + 293.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 480.0, + 987.0, + 480.0, + 987.0, + 514.0, + 725.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 480.0, + 1406.0, + 480.0, + 1406.0, + 514.0, + 1021.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 509.0, + 690.0, + 509.0, + 690.0, + 546.0, + 292.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 509.0, + 1405.0, + 509.0, + 1405.0, + 546.0, + 724.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 539.0, + 441.0, + 539.0, + 441.0, + 576.0, + 292.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 539.0, + 524.0, + 539.0, + 524.0, + 576.0, + 475.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 539.0, + 1088.0, + 539.0, + 1088.0, + 576.0, + 558.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1943.0, + 1403.0, + 1943.0, + 1403.0, + 1980.0, + 293.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1402.0, + 1973.0, + 1402.0, + 2011.0, + 295.0, + 2011.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 202, + 1405, + 202, + 1405, + 568, + 297, + 568 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 679, + 1404, + 679, + 1404, + 862, + 298, + 862 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 299, + 609, + 944, + 609, + 944, + 649, + 299, + 649 + ], + "score": 0.93 + }, + { + "category_id": 2, + "poly": [ + 837, + 2062, + 865, + 2062, + 865, + 2085, + 837, + 2085 + ], + "score": 0.842 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 605.0, + 949.0, + 605.0, + 949.0, + 659.0, + 293.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 203.0, + 1405.0, + 203.0, + 1405.0, + 237.0, + 295.0, + 237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 233.0, + 1405.0, + 233.0, + 1405.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 265.0, + 1405.0, + 265.0, + 1405.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 295.0, + 1405.0, + 295.0, + 1405.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 326.0, + 1405.0, + 326.0, + 1405.0, + 358.0, + 292.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 389.0, + 294.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 385.0, + 1405.0, + 385.0, + 1405.0, + 419.0, + 295.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 415.0, + 1405.0, + 415.0, + 1405.0, + 449.0, + 295.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 441.0, + 1406.0, + 441.0, + 1406.0, + 483.0, + 291.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 472.0, + 1408.0, + 472.0, + 1408.0, + 512.0, + 292.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 501.0, + 1406.0, + 501.0, + 1406.0, + 545.0, + 294.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 534.0, + 601.0, + 534.0, + 601.0, + 573.0, + 295.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 676.0, + 1406.0, + 676.0, + 1406.0, + 716.0, + 292.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 709.0, + 1405.0, + 709.0, + 1405.0, + 744.0, + 293.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 739.0, + 1406.0, + 739.0, + 1406.0, + 778.0, + 293.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 769.0, + 1408.0, + 769.0, + 1408.0, + 808.0, + 292.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 796.0, + 1405.0, + 796.0, + 1405.0, + 840.0, + 291.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 831.0, + 1119.0, + 831.0, + 1119.0, + 867.0, + 294.0, + 867.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 294, + 195, + 1409, + 195, + 1409, + 2021, + 294, + 2021 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 299, + 200, + 455, + 200, + 455, + 234, + 299, + 234 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 863, + 2061, + 863, + 2086, + 836, + 2086 + ], + "score": 0.781 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 863, + 2061, + 863, + 2086, + 836, + 2086 + ], + "score": 0.198 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 196.0, + 460.0, + 196.0, + 460.0, + 240.0, + 295.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 868.0, + 2058.0, + 868.0, + 2098.0, + 832.0, + 2098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 868.0, + 2058.0, + 868.0, + 2098.0, + 832.0, + 2098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 197.0, + 462.0, + 197.0, + 462.0, + 242.0, + 295.0, + 242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 247.0, + 1407.0, + 247.0, + 1407.0, + 285.0, + 304.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 277.0, + 1041.0, + 277.0, + 1041.0, + 316.0, + 348.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 326.0, + 1409.0, + 326.0, + 1409.0, + 364.0, + 304.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 356.0, + 1407.0, + 356.0, + 1407.0, + 388.0, + 352.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 380.0, + 1151.0, + 380.0, + 1151.0, + 416.0, + 346.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 430.0, + 1407.0, + 430.0, + 1407.0, + 468.0, + 304.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 458.0, + 1409.0, + 458.0, + 1409.0, + 496.0, + 348.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 486.0, + 650.0, + 486.0, + 650.0, + 518.0, + 350.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 534.0, + 1409.0, + 534.0, + 1409.0, + 572.0, + 304.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 562.0, + 1337.0, + 562.0, + 1337.0, + 600.0, + 348.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 612.0, + 1409.0, + 612.0, + 1409.0, + 651.0, + 304.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 687.0, + 1407.0, + 687.0, + 1407.0, + 727.0, + 302.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 715.0, + 488.0, + 715.0, + 488.0, + 753.0, + 348.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 765.0, + 1409.0, + 765.0, + 1409.0, + 803.0, + 304.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 795.0, + 418.0, + 795.0, + 418.0, + 829.0, + 350.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 841.0, + 1409.0, + 841.0, + 1409.0, + 879.0, + 304.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 871.0, + 1315.0, + 871.0, + 1315.0, + 909.0, + 350.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 919.0, + 1409.0, + 919.0, + 1409.0, + 958.0, + 304.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 945.0, + 1407.0, + 945.0, + 1407.0, + 986.0, + 346.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 978.0, + 542.0, + 978.0, + 542.0, + 1010.0, + 350.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1024.0, + 1411.0, + 1024.0, + 1411.0, + 1062.0, + 292.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1052.0, + 684.0, + 1052.0, + 684.0, + 1090.0, + 348.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1100.0, + 1407.0, + 1100.0, + 1407.0, + 1140.0, + 292.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1130.0, + 1103.0, + 1130.0, + 1103.0, + 1168.0, + 348.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 1407.0, + 1178.0, + 1407.0, + 1216.0, + 294.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1209.0, + 491.0, + 1209.0, + 491.0, + 1240.0, + 343.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1254.0, + 1407.0, + 1254.0, + 1407.0, + 1293.0, + 292.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1287.0, + 903.0, + 1287.0, + 903.0, + 1319.0, + 350.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1335.0, + 1405.0, + 1335.0, + 1405.0, + 1367.0, + 296.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1361.0, + 927.0, + 1361.0, + 927.0, + 1397.0, + 346.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1409.0, + 1407.0, + 1409.0, + 1407.0, + 1447.0, + 292.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1435.0, + 608.0, + 1435.0, + 608.0, + 1473.0, + 348.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1489.0, + 1407.0, + 1489.0, + 1407.0, + 1521.0, + 294.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1517.0, + 1221.0, + 1517.0, + 1221.0, + 1549.0, + 350.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1563.0, + 1407.0, + 1563.0, + 1407.0, + 1602.0, + 294.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1592.0, + 1169.0, + 1592.0, + 1169.0, + 1630.0, + 348.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1640.0, + 1407.0, + 1640.0, + 1407.0, + 1678.0, + 292.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1672.0, + 828.0, + 1672.0, + 828.0, + 1704.0, + 350.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1720.0, + 1407.0, + 1720.0, + 1407.0, + 1752.0, + 296.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1770.0, + 1409.0, + 1770.0, + 1409.0, + 1802.0, + 296.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1798.0, + 1255.0, + 1798.0, + 1255.0, + 1830.0, + 350.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1846.0, + 1409.0, + 1846.0, + 1409.0, + 1878.0, + 296.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1868.0, + 1407.0, + 1868.0, + 1407.0, + 1911.0, + 348.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1904.0, + 540.0, + 1904.0, + 540.0, + 1931.0, + 352.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1945.0, + 1407.0, + 1945.0, + 1407.0, + 1991.0, + 290.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1977.0, + 1097.0, + 1977.0, + 1097.0, + 2015.0, + 348.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 637.5, + 417.0, + 637.5, + 417.0, + 674.0, + 351.0, + 674.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 293, + 199, + 1408, + 199, + 1408, + 1476, + 293, + 1476 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.591 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.514 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 206.0, + 1407.0, + 206.0, + 1407.0, + 242.0, + 296.0, + 242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 231.0, + 1408.0, + 231.0, + 1408.0, + 270.0, + 347.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 259.0, + 417.0, + 259.0, + 417.0, + 295.0, + 348.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 304.0, + 1407.0, + 304.0, + 1407.0, + 353.0, + 290.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 341.0, + 1369.0, + 341.0, + 1369.0, + 373.0, + 351.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 389.0, + 1405.0, + 389.0, + 1405.0, + 424.0, + 294.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 416.0, + 979.0, + 416.0, + 979.0, + 452.0, + 350.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 463.0, + 1407.0, + 463.0, + 1407.0, + 503.0, + 294.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 493.0, + 1068.0, + 493.0, + 1068.0, + 529.0, + 350.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 539.0, + 1405.0, + 539.0, + 1405.0, + 582.0, + 291.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 571.0, + 1051.0, + 571.0, + 1051.0, + 607.0, + 350.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 618.0, + 1405.0, + 618.0, + 1405.0, + 658.0, + 293.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 644.0, + 1407.0, + 644.0, + 1407.0, + 688.0, + 345.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 674.0, + 563.0, + 674.0, + 563.0, + 710.0, + 348.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 724.0, + 1405.0, + 724.0, + 1405.0, + 764.0, + 294.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 754.0, + 1113.0, + 754.0, + 1113.0, + 790.0, + 350.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 803.0, + 1405.0, + 803.0, + 1405.0, + 839.0, + 294.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 827.0, + 1408.0, + 827.0, + 1408.0, + 869.0, + 347.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 856.0, + 700.0, + 856.0, + 700.0, + 895.0, + 347.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 906.0, + 1407.0, + 906.0, + 1407.0, + 947.0, + 293.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 936.0, + 1319.0, + 936.0, + 1319.0, + 972.0, + 350.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 987.0, + 1405.0, + 987.0, + 1405.0, + 1023.0, + 294.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1014.0, + 1001.0, + 1014.0, + 1001.0, + 1050.0, + 350.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1063.0, + 1405.0, + 1063.0, + 1405.0, + 1099.0, + 294.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1090.0, + 945.0, + 1090.0, + 945.0, + 1129.0, + 347.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1137.0, + 1407.0, + 1137.0, + 1407.0, + 1179.0, + 291.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1169.0, + 1163.0, + 1169.0, + 1163.0, + 1205.0, + 350.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1218.0, + 1407.0, + 1218.0, + 1407.0, + 1258.0, + 293.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 1241.0, + 1410.0, + 1241.0, + 1410.0, + 1285.0, + 345.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1274.0, + 417.0, + 1274.0, + 417.0, + 1308.0, + 350.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1319.0, + 1405.0, + 1319.0, + 1405.0, + 1364.0, + 291.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1352.0, + 950.0, + 1352.0, + 950.0, + 1388.0, + 350.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1397.0, + 1407.0, + 1397.0, + 1407.0, + 1440.0, + 291.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1431.0, + 656.0, + 1431.0, + 656.0, + 1463.0, + 351.0, + 1463.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/LkFG3lB13U5/LkFG3lB13U5.md b/parse/train/LkFG3lB13U5/LkFG3lB13U5.md new file mode 100644 index 0000000000000000000000000000000000000000..b70889051d99c325169c7a0f7b46a12f4a4fe4b8 --- /dev/null +++ b/parse/train/LkFG3lB13U5/LkFG3lB13U5.md @@ -0,0 +1,914 @@ +# ADAPTIVE FEDERATED OPTIMIZATION + +Sashank J. Reddi∗, Zachary Charles∗, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konecný, Sanjiv Kumar, H. Brendan McMahan ˇ + +Google Research +{sashank, zachcharles, manzilzaheer, zachgarrett, krush, +konkey, sanjivk, mcmahan}@google.com + +# ABSTRACT + +Federated learning is a distributed machine learning paradigm in which a large number of clients coordinate with a central server to learn a model without sharing their own training data. Standard federated optimization methods such as Federated Averaging (FEDAVG) are often difficult to tune and exhibit unfavorable convergence behavior. In non-federated settings, adaptive optimization methods have had notable success in combating such issues. In this work, we propose federated versions of adaptive optimizers, including ADAGRAD, ADAM, and YOGI, and analyze their convergence in the presence of heterogeneous data for general nonconvex settings. Our results highlight the interplay between client heterogeneity and communication efficiency. We also perform extensive experiments on these methods and show that the use of adaptive optimizers can significantly improve the performance of federated learning. + +# 1 INTRODUCTION + +Federated learning (FL) is a machine learning paradigm in which multiple clients cooperate to learn a model under the orchestration of a central server (McMahan et al., 2017). In FL, raw client data is never shared with the server or other clients. This distinguishes FL from traditional distributed optimization, and requires contending with heterogeneous data. FL has two primary settings, crosssilo (eg. FL between large institutions) and cross-device (eg. FL across edge devices) (Kairouz et al., 2019, Table 1). In cross-silo FL, most clients participate in every round and can maintain state between rounds. In the more challenging cross-device FL, our primary focus, only a small fraction of clients participate in each round, and clients cannot maintain state across rounds. For a more in-depth discussion of FL and the challenges involved, we defer to Kairouz et al. (2019) and Li et al. (2019a). + +Standard optimization methods, such as distributed SGD, are often unsuitable in FL and can incur high communication costs. To remedy this, many federated optimization methods use local client updates, in which clients update their models multiple times before communicating with the server. This can greatly reduce the amount of communication required to train a model. One such method is FEDAVG (McMahan et al., 2017), in which clients perform multiple epochs of SGD on their local datasets. The clients communicate their models to the server, which averages them to form a new global model. While FEDAVG has seen great success, recent works have highlighted its convergence issues in some settings (Karimireddy et al., 2019; Hsu et al., 2019). This is due to a variety of factors including (1) client drift (Karimireddy et al., 2019), where local client models move away from globally optimal models, and (2) a lack of adaptivity. FEDAVG is similar in spirit to SGD, and may be unsuitable for settings with heavy-tail stochastic gradient noise distributions, which often arise when training language models (Zhang et al., 2019a). Such settings benefit from adaptive learning rates, which incorporate knowledge of past iterations to perform more informed optimization. + +In this paper, we focus on the second issue and present a simple framework for incorporating adaptivity in FL. In particular, we propose a general optimization framework in which (1) clients perform multiple epochs of training using a client optimizer to minimize loss on their local data and (2) server updates its global model by applying a gradient-based server optimizer to the average of the clients’ model updates. We show that FEDAVG is the special case where SGD is used as both client and server optimizer and server learning rate is 1. This framework can also seamlessly incorporate adaptivity by using adaptive optimizers as client or server optimizers. Building upon this, we develop novel adaptive optimization techniques for FL by using per-coordinate methods as server optimizers. By focusing on adaptive server optimization, we enable use of adaptive learning rates without increase in client storage or communication costs, and ensure compatibility with cross-device FL. + +Main contributions In light of the above, we highlight the main contributions of the paper. + +• We study a general framework for federated optimization using server and client optimizers. This framework generalizes many existing federated optimization methods, including FEDAVG. • We use this framework to design novel, cross-device compatible, adaptive federated optimization methods, and provide convergence analysis in general nonconvex settings. To the best of our knowledge, these are the first methods for FL using adaptive server optimization. We show an important interplay between the number of local steps and the heterogeneity among clients. • We introduce comprehensive and reproducible empirical benchmarks for comparing federated optimization methods. These benchmarks consist of seven diverse and representative FL tasks involving both image and text data, with varying amounts of heterogeneity and numbers of clients. • We demonstrate strong empirical performance of our adaptive optimizers throughout, improving upon commonly used baselines. Our results show that our methods can be easier to tune, and highlight their utility in cross-device settings. + +Related work FEDAVG was first introduced by McMahan et al. (2017), who showed it can dramatically reduce communication costs. Many variants have since been proposed to tackle issues such as convergence and client drift. Examples include adding a regularization term in the client objectives towards the broadcast model (Li et al., 2018), and server momentum (Hsu et al., 2019). When clients are homogeneous, FEDAVG reduces to local SGD (Zinkevich et al., 2010), which has been analyzed by many works (Stich, 2019; Yu et al., 2019; Wang & Joshi, 2018; Stich & Karimireddy, 2019; Basu et al., 2019). In order to analyze FEDAVG in heterogeneous settings, many works derive convergence rates depending on the amount of heterogeneity (Li et al., 2018; Wang et al., 2019; Khaled et al., 2019; Li et al., 2019b). Typically, the convergence rate of FEDAVG gets worse with client heterogeneity. By using control variates to reduce client drift, the SCAFFOLD method (Karimireddy et al., 2019) achieves convergence rates that are independent of the amount of heterogeneity. While effective in cross-silo FL, the method is incompatible with cross-device FL as it requires clients to maintain state across rounds. For more detailed comparisons, we defer to Kairouz et al. (2019). + +Adaptive methods have been the subject of significant theoretical and empirical study, in both convex (McMahan & Streeter, 2010b; Duchi et al., 2011; Kingma & Ba, 2015) and non-convex settings (Li & Orabona, 2018; Ward et al., 2018; Wu et al., 2019). Reddi et al. (2019); Zaheer et al. (2018) study convergence failures of ADAM in certain non-convex settings, and develop an adaptive optimizer, YOGI, designed to improve convergence. While most work on adaptive methods focuses on non-FL settings, Xie et al. (2019) propose ADAALTER, a method for FL using adaptive client optimization. Conceptually, our approach is also related to the LOOKAHEAD optimizer (Zhang et al., 2019b), which was designed for non-FL settings. Similar to ADAALTER, an adaptive FL variant of LOOKAHEAD entails adaptive client optimization (see Appendix B.3 for more details). We note that both ADAALTER and LOOKAHEAD are, in fact, special cases of our framework (see Algorithm 1) and the primary novelty of our work comes in focusing on adaptive server optimization. This allows us to avoid aggregating optimizer states across clients, making our methods require at most half as much communication and client memory usage per round (see Appendix B.3 for details). + +Notation For $a , b \in \mathbb { R } ^ { d }$ , we let ${ \sqrt { a } } , a ^ { 2 }$ and $a / b$ denote the element-wise square root, square, and division of the vectors. For $\theta _ { i } \in \mathbb { R } ^ { d }$ , we use both $\theta _ { i , j }$ and $[ \theta _ { i } ] _ { j }$ to denote its $j ^ { \mathrm { t h } }$ coordinate. + +# 2 FEDERATED LEARNING AND FEDAVG + +In federated learning, we solve an optimization problem of the form: + +$$ +\operatorname* { m i n } _ { x \in \mathbb { R } ^ { d } } f ( x ) = \frac { 1 } { m } \sum _ { i = 1 } ^ { m } F _ { i } ( x ) , +$$ + +where $F _ { i } ( x ) = \mathbb { E } _ { z \sim \mathcal { D } _ { i } } [ f _ { i } ( x , z ) ]$ , is the loss function of the $i ^ { \mathrm { { t h } } }$ client, $z \in { \mathcal { Z } }$ , and $\mathcal { D } _ { i }$ is the data distribution for the $i ^ { \mathrm { { t h } } }$ client. For $i \neq j$ , $\mathcal { D } _ { i }$ and $\mathcal { D } _ { j }$ may be very different. The functions $F _ { i }$ (and therefore $f$ ) may be nonconvex. For each $i$ and $x$ , we assume access to an unbiased stochastic gradient $g _ { i } ( x )$ of the client’s true gradient $\nabla F _ { i } ( x )$ . In addition, we make the following assumptions. + +Assumption 1 (Lipschitz Gradient). The function $F _ { i }$ is $L$ -smooth for all $i \in [ m ]$ i.e., $\parallel \nabla F _ { i } ( x ) -$ $\nabla F _ { i } ( y ) \| \leq L \| x - y \|$ , for all $x , y \in \mathbb { R } ^ { d }$ . + +Assumption 2 (Bounded Variance). The function $F _ { i }$ have $\sigma _ { l }$ -bounded (local) variance i.e., $\mathbb { E } [ \| \nabla [ \widehat { f } _ { i } ( x , z ) ] _ { j } - [ \nabla F _ { i } ( x ) ] _ { j } \| ^ { 2 } ] = \sigma _ { l , j } ^ { 2 }$ for all $x \in \mathbb { R } ^ { d }$ , $j \in [ d ]$ and $i \in [ m ]$ . Furthermore, we assume the (global) variance is bounded, $\begin{array} { r } { \left( 1 / m \right) \sum _ { i = 1 } ^ { m } \| \nabla [ F _ { i } ( x ) ] _ { j } - [ \nabla f ( x ) ] _ { j } \| ^ { 2 } \leq \sigma _ { g , j } ^ { 2 } } \end{array}$ for all $\boldsymbol { x } \in \mathbb { R } ^ { d }$ and $j \in [ d ]$ . + +Assumption 3 (Bounded Gradients). The function $f _ { i } ( x , z )$ have $G$ -bounded gradients i.e., for any $i \in [ m ]$ , $\boldsymbol { x } \in \mathbb { R } ^ { d }$ and $z \in { \mathcal { Z } }$ we have $| [ \nabla f _ { i } ( x , z ) ] _ { j } | \leq G$ for all $j \in [ d ]$ . + +With a slight abuse of notation, we use $\sigma _ { l } ^ { 2 }$ and $\sigma _ { g } ^ { 2 }$ to denote $\textstyle \sum _ { j = 1 } ^ { d } \sigma _ { l , j } ^ { 2 }$ and $\textstyle \sum _ { j = 1 } ^ { d } \sigma _ { g , j } ^ { 2 }$ . Assumptions 1 and 3 are fairly standard in nonconvex optimization literature (Reddi et al., 2016; Ward et al., 2018; Zaheer et al., 2018). We make no further assumptions regarding the similarity of clients datasets. Assumption 2 is a form of bounded variance, but between the client objective functions and the overall objective function. This assumption has been used throughout various works on federated optimization (Li et al., 2018; Wang et al., 2019). Intuitively, the parameter $\sigma _ { g }$ quantifies similarity of client objective functions. Note $\sigma _ { g } = 0$ corresponds to the i.i.d. setting. + +A common approach to solving (1) in federated settings is FEDAVG (McMahan et al., 2017). At each round of FEDAVG, a subset of clients are selected (typically randomly) and the server broadcasts its global model to each client. In parallel, the clients run SGD on their own loss function, and send the resulting model to the server. The server then updates its global model as the average of these local models. See Algorithm 3 in the appendix for more details. + +Suppose that at round $t$ , the server has model $x _ { t }$ and samples a set $s$ of clients. Let $\ v x _ { i } ^ { t }$ denote the model of each client $i \in S$ after local training. We rewrite FEDAVG’s update as + +$$ +x _ { t + 1 } = { \frac { 1 } { | S | } } \sum _ { i \in S } x _ { i } ^ { t } = x _ { t } - { \frac { 1 } { | S | } } \sum _ { i \in S } \left( x _ { t } - x _ { i } ^ { t } \right) . +$$ + +Let $\Delta _ { i } ^ { t } : = x _ { i } ^ { t } - x _ { t }$ and $\begin{array} { r } { \Delta _ { t } : = \left( { 1 } / { | { S } | } \right) \sum _ { i \in \mathcal { S } } \Delta _ { i } ^ { t } } \end{array}$ . Then the server update in FEDAVG is equivalent to applying SGD to the “pseudo-gradient” $\bar { - } \Delta _ { t }$ with learning rate $\eta = 1$ . This formulation makes it clear that other choices of $\eta$ are possible. One could also utilize optimizers other than SGD on the clients, or use an alternative update rule on the server. This family of algorithms, which we refer to collectively as FEDOPT, is formalized in Algorithm 1. + +# Algorithm 1 FEDOPT + +
1: Input:xo,CLIENTOPT,SERVEROPT
2: fort=O,.,T-1do
3: Samplea subset Sofclients
4: xt,0=xt
5: for each client i ∈ S in parallel do
6: for k=0,.,K-1 do
Compute an unbiased estimate gt,k of VFi(xt,k) 7:
x,+1=CLEO(,,,t) 8:
9: △=x,K-xt
10: △t=∑ies△t 11: Xt+1 = SERVEROPT(xt,-△t,n,t)
+ +In Algorithm 1, CLIENTOPT and SERVEROPT are gradient-based optimizers with learning rates $\eta _ { l }$ and $\eta$ respectively. Intuitively, CLIENTOPT aims to minimize (1) based on each client’s local data while SERVEROPT optimizes from a global perspective. FEDOPT naturally allows the use of adaptive optimizers (eg. ADAM, YOGI, etc.), as well as techniques such as server-side momentum (leading to FEDAVGM, proposed by Hsu et al. (2019)). In its most general form, FEDOPT uses a CLIENTOPT whose updates can depend on globally aggregated statistics (e.g. server updates in the previous iterations). We also allow $\eta$ and $\eta _ { l }$ to depend on the round $t$ in order to encompass learning rate schedules. While we focus on specific adaptive optimizers in this work, we can in principle use any adaptive optimizer (e.g. AMSGRAD (Reddi et al., 2019), ADABOUND (Luo et al., 2019)). + +While FEDOPT has intuitive benefits over FEDAVG, it also raises a fundamental question: Can the negative of the average model difference $\Delta _ { t }$ be used as a pseudo-gradient in general server optimizer updates? In this paper, we provide an affirmative answer to this question by establishing a theoretical basis for FEDOPT. We will show that the use of the term SERVEROPT is justified, as we can guarantee convergence across a wide variety of server optimizers, including ADAGRAD, ADAM, and YOGI, thus developing principled adaptive optimizers for FL based on our framework. + +# 3 ADAPTIVE FEDERATED OPTIMIZATION + +In this section, we specialize FEDOPT to settings where SERVEROPT is an adaptive optimization method (one of ADAGRAD, YOGI or ADAM) and CLIENTOPT is SGD. By using adaptive methods (which generally require maintaining state) on the server and SGD on the clients, we ensure our methods have the same communication cost as FEDAVG and work in cross-device settings. + +Algorithm 2 provides pseudo-code for our methods. An alternate version using batched data and example-based weighting (as opposed to uniform weighting) of clients is given in Algorithm 5. The parameter $\tau$ controls the algorithms’ degree of adaptivity, with smaller values of $\tau$ representing higher degrees of adaptivity. Note that the server updates of our methods are invariant to fixed multiplicative changes to the client learning rate $\eta _ { l }$ for appropriately chosen $\tau$ , though as we shall see shortly, we will require $\eta _ { l }$ to be sufficiently small in our analysis. + +# Algorithm 2 FEDADAGRAD FEDYOGI , and FEDADAM + +1: Initialization: $x _ { 0 } , v _ { - 1 } \ge \tau ^ { 2 }$ , decay parameters $\beta _ { 1 } , \beta _ { 2 } \in [ 0 , 1 )$ +2: for $t = 0 , \cdots , T - 1$ do +3: Sample subset $s$ of clients +4: $x _ { i , 0 } ^ { t } \dot { = } x _ { t }$ +5: for each client $i \in S$ in parallel do +6: for $k = 0 , \cdots , K - \bar { 1 }$ do +7: Compute an unbiased estimate $g _ { i , k } ^ { t }$ of $\nabla F _ { i } ( x _ { i , k } ^ { t } )$ +8: $\begin{array} { c } { x _ { i , k + 1 } ^ { t } = x _ { i , k } ^ { t } - \eta _ { l } g _ { i , k } ^ { t } } \\ { \Delta _ { i } ^ { t } = x _ { i , K } ^ { t } - x _ { t } } \\ { \Delta _ { t } = \beta _ { 1 } \Delta _ { t - 1 } + \left( 1 - \beta _ { 1 } \right) \left( \frac { 1 } { \left| S \right| } \sum _ { i \in S } \Delta _ { i } ^ { t } \right) } \end{array}$ +9: +10: +11: $v _ { t } = v _ { t - 1 } + \Delta _ { t } ^ { 2 }$ (FEDADAGRAD) +12: $\begin{array} { r l } & { v _ { t } = v _ { t - 1 } - ( 1 - \beta _ { 2 } ) \Delta _ { t } ^ { 2 } \mathrm { s i g n } ( v _ { t - 1 } - \Delta _ { t } ^ { 2 } ) \ ( \mathrm { F E D Y O G I } ) } \\ & { v _ { t } = \beta _ { 2 } v _ { t - 1 } + ( 1 - \beta _ { 2 } ) \Delta _ { t } ^ { 2 } \left( \mathrm { F E D A D A M } \right) } \\ & { x _ { t + 1 } = x _ { t } + \eta \frac { \Delta _ { t } } { \sqrt { v _ { t } } + \tau } } \end{array}$ +13: +14: + +We provide convergence analyses of these methods in general nonconvex settings, assuming full participation, i.e. $\bar { \mathcal { S } } = [ m ]$ . For expository purposes, we assume $\beta _ { 1 } = 0$ , though our analysis can be directly extended to $\beta _ { 1 } > 0$ . Our analysis can also be extended to partial participation (i.e. $| S | < m$ , see Appendix A.2.1 for details). Furthermore, non-uniform weighted averaging typically used in FEDAVG (McMahan et al., 2017) can also be incorporated into our analysis fairly easily. + +Theorem 1. Let Assumptions $I$ to 3 hold, and let $L , G , \sigma _ { l } , \sigma _ { g }$ be as defined therein. Let $\sigma ^ { 2 } = $ $\sigma _ { l } ^ { 2 } + 6 K \sigma _ { g } ^ { 2 }$ . Consider the following conditions for $\eta _ { l }$ : + +(Condition I) + +(Condition II) + +$$ +\begin{array} { r l } & { \eta _ { l } \le \displaystyle \frac { 1 } { K } \operatorname* { m i n } \left\{ \frac { 1 } { 1 6 L } , \frac { 1 } { T ^ { 1 / 6 } } \left[ \frac { \tau } { 1 2 0 L ^ { 2 } G } \right] ^ { 1 / 3 } \right\} , } \\ & { \eta _ { l } \le \displaystyle \frac { 1 } { 3 K } \operatorname* { m i n } \left\{ \frac { 1 } { T ^ { 1 / 1 0 } } \left[ \frac { \tau ^ { 3 } } { L ^ { 2 } G ^ { 3 } } \right] ^ { 1 / 5 } , \frac { 1 } { T ^ { 1 / 8 } } \left[ \frac { \tau ^ { 2 } } { L ^ { 3 } G \eta } \right] ^ { 1 / 4 } \right\} . } \end{array} +$$ + +Then the iterates of Algorithm 2 for FEDADAGRAD satisfy + +$$ +\begin{array} { r l r l } & { u n d e r ~ C o n d i t i o n ~ I o n l y , } & & { \underset { 0 \leq t \leq T - 1 } { \operatorname* { m i n } } \mathbb { E } \| \nabla f ( x _ { t } ) \| ^ { 2 } \leq \odot \left( \left[ \frac { G } { \sqrt { T } } + \frac { \tau } { \eta _ { t } K T } \right] ( \Psi + \Psi _ { v a r } ) \right) , } \\ & { u n d e r ~ b o t h ~ C o n d i t i o n ~ I \ll { I I } , } & & { \underset { 0 \leq t \leq T - 1 } { \operatorname* { m i n } } \mathbb { E } \| \nabla f ( x _ { t } ) \| ^ { 2 } \leq \odot \left( \left[ \frac { G } { \sqrt { T } } + \frac { \tau } { \eta _ { t } K T } \right] \left( \Psi + \widetilde { \Psi } _ { \mathrm { v a r } } \right) \right) . } \end{array} +$$ + +Here, we define + +$$ +\begin{array} { c } { { \Psi = \displaystyle \frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \eta } + \frac { 5 \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \tau } \sigma ^ { 2 } , } } \\ { { \Psi _ { \mathrm { v a r } } = \displaystyle \frac { d ( \eta _ { l } K G ^ { 2 } + \tau \eta L ) } { \tau } \Bigg [ 1 + \log { \displaystyle \frac { \tau ^ { 2 } + \eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \tau ^ { 2 } } } \Bigg ] , } } \\ { { \widetilde { \Psi } _ { \mathrm { v a r } } = \displaystyle \frac { 2 \eta _ { l } K G ^ { 2 } + \tau \eta L } { \tau ^ { 2 } } \left[ \displaystyle \frac { 2 \eta _ { l } ^ { 2 } K T } { m } \sigma _ { l } ^ { 2 } + 1 0 \eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \sigma ^ { 2 } \right] . } } \end{array} +$$ + +All proofs are relegated to Appendix A due to space constraints. When $\eta _ { l }$ satisfies the condition in the second part the above result, we obtain a convergence rate depending on $\operatorname* { m i n } \{ \Psi _ { \mathrm { v a r } } , \widetilde { \Psi } _ { \mathrm { v a r } } \}$ . To obtain an explicit dependence on $T$ and $K$ , we simplify the above result for a specific choice of $\eta , \eta _ { l }$ and $\tau$ . + +Corollary 1. Suppose √ $\eta _ { l }$ is such that the conditions in Theorem $I$ are satisfied and $\eta _ { l } = \Theta \big ( { 1 } / { ( K L \sqrt { T } } \big )$ . Also suppose $\eta = \Theta ( \sqrt { K m } )$ and $\tau = G / _ { L }$ . Then, for sufficiently large $T$ , the iterates of Algorithm 2 for FEDADAGRAD satisfy + +$$ +\operatorname* { m i n } _ { 0 \leq t \leq T - 1 } \mathbb { E } \| \nabla f ( x _ { t } ) \| ^ { 2 } = \mathcal { O } \Bigg ( \frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \sqrt { m K T } } + \frac { 2 \sigma _ { l } ^ { 2 } L } { G ^ { 2 } \sqrt { m K T } } + \frac { \sigma ^ { 2 } } { G K T } + \frac { \sigma ^ { 2 } L \sqrt { m } } { G ^ { 2 } \sqrt { K } T ^ { 3 / 2 } } \Bigg ) . +$$ + +We defer a detailed discussion about our analysis and its implication to the end of the section. + +Analysis of FEDADAM Next, we provide the convergence analysis of FEDADAM. The proof of FEDYOGI is very similar and hence, we omit the details of FEDYOGI’s analysis. + +Theorem 2. Let Assumptions 1 to 3 hold, and $L , G , \sigma _ { l } , \sigma _ { g }$ be as defined therein. Let $\sigma ^ { 2 } = \sigma _ { l } ^ { 2 } + 6 K \sigma _ { g } ^ { 2 }$ Suppose the client learning rate satisfies $\eta _ { l } \leq 1 / 1 6 L K$ and + +$$ +\eta _ { l } \leq \frac { 1 } { 6 K } \operatorname* { m i n } \left\{ \left[ \frac { \tau } { G L } \right] ^ { 1 / 2 } , \left[ \frac { \tau ^ { 2 } } { G L ^ { 3 } \eta } \right] ^ { 1 / 4 } , \left[ \frac { \tau } { G L ^ { 2 } } \right] ^ { 1 / 3 } \right\} . +$$ + +Then the iterates of Algorithm 2 for FEDADAM satisfy + +$$ +\operatorname* { m i n } _ { 0 \leq t \leq T - 1 } \mathbb { E } \| \nabla f ( x _ { t } ) \| ^ { 2 } = \odot \left( \frac { \sqrt { \beta _ { 2 } } \eta _ { l } K G + \tau } { \eta _ { l } K T } \left( \Psi + \Psi _ { \mathrm { v a r } } \right) \right) , +$$ + +where + +$$ +\begin{array} { c } { { \Psi = \displaystyle \frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \eta } + \frac { 5 \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \tau } \sigma ^ { 2 } , } } \\ { { \Psi _ { \nu a r } = \left( G + \displaystyle \frac { \eta L } { 2 } \right) \left[ \displaystyle \frac { 4 \eta _ { l } ^ { 2 } K T } { m \tau ^ { 2 } } \sigma _ { l } ^ { 2 } + \displaystyle \frac { 2 0 \eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \tau ^ { 2 } } \sigma ^ { 2 } \right] . } } \end{array} +$$ + +Similar to the FEDADAGRAD case, we restate the above result for a specific choice of $\eta _ { l } , \eta$ and $\tau$ in order to highlight the dependence of $K$ and $T$ . + +Corollary 2. Suppose $\eta _ { l }$ is chosen such that the conditions in Theorem 2 are satisfied and that√ $\eta _ { l } = \Theta \bigl ( 1 / ( K L \sqrt { T } ) \bigr )$ . Also, suppose $\eta = \Theta ( \sqrt { K m } )$ and $\tau = G / _ { L }$ . Then, for sufficiently large $T$ , the iterates of Algorithm 2 for FEDADAM satisfy + +$$ +\operatorname* { m i n } _ { 0 \leq t \leq T - 1 } \mathbb { E } \| \nabla f ( x _ { t } ) \| ^ { 2 } = \mathcal { O } \Bigg ( \frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \sqrt { m K T } } + \frac { 2 \sigma _ { l } ^ { 2 } L } { G ^ { 2 } \sqrt { m K T } } + \frac { \sigma ^ { 2 } } { G K T } + \frac { \sigma ^ { 2 } L \sqrt { m } } { G ^ { 2 } \sqrt { K } T ^ { 3 / 2 } } \Bigg ) . +$$ + +Remark 1. The server learning rate $\eta = 1$ typically used in FEDAVG does not necessarily minimize the upper bound in Theorems 1 & 2. The effect of $\cdot _ { g }$ , a measure of client heterogeneity, on convergence can be reduced by choosing sufficiently $\eta _ { l }$ and a reasonably large $\eta$ (e.g. see Corollary 1). Thus, the effect of client heterogeneity can be reduced by carefully choosing client and server learning rates, but not removed entirely. Our empirical analysis (eg. Figure 1) supports this conclusion. + +Discussion. We briefly discuss our theoretical analysis and its implications in the FL setting. The convergence rates for FEDADAGRAD and FEDADAM are similar, so our discussion applies to all the adaptive federated optimization algorithms (including FEDYOGI) proposed in the paper. + +(i) Comparison of convergence rates. When $T$ is sufficiently large compared to $K$ , $\mathcal { O } ( { } ^ { 1 } / { \sqrt { m K T } } )$ is the dominant term in Corollary $1 \ \& \ 2$ . Thus, we effectively obtain a convergence rate of $\mathcal { O } ( { } ^ { 1 } / \sqrt { m K T } )$ , which matches the best known rate for the general non-convex setting of our interest (e.g. see (Karimireddy et al., 2019)). We also note that in the $i . i . d$ setting considered in (Wang & Joshi, 2018), which corresponds to $\sigma _ { g } = 0$ , we match their convergence rates. Similar to the centralized setting, it is possible to obtain convergence rates with better dependence on constants for federated adaptive methods, compared to FEDAVG, by incorporating non-uniform bounds on gradients across coordinates (Zaheer et al., 2018). + +(ii) Learning rates $\pmb { \& }$ their decay. The client learning rate of $^ 1 / \sqrt { T }$ in our analysis requires knowledge of the number of rounds $T$ a priori; however, it is easy to generalize our analysis to the case where $\eta _ { l }$ is decayed at a rate of $^ { 1 } / \sqrt { t }$ . Observe that one must decay $\eta _ { l }$ , not the server learning rate $\eta$ , to obtain convergence. This is because the client drift introduced by the local updates does not vanish as $T \to \infty$ when $\eta _ { l }$ is constant. As we show in Appendix E.6, learning rate decay can improve empirical performance. Also, note the inverse relationship between $\eta _ { l }$ and $\eta$ in Corollary 1 & 2, which we observe in our empirical analysis (see Appendix E.4). + +(iii) Communication efficiency $\pmb { \& }$ local steps. The total communication cost of the algorithms depends on the number of communication rounds $T$ . From Corollary 1 & 2, it is clear that a larger $K$ leads to fewer rounds of communication as long as $K = \mathcal { O } ( T \bar { \sigma } _ { l } ^ { 2 } / \sigma _ { g } ^ { 2 } )$ . Thus, the number of local iterations can be large when either the ratio $\sigma _ { l } ^ { 2 } / \sigma _ { g } ^ { 2 }$ or $T$ is large. In the $i . i . d$ setting where $\sigma _ { g } = 0$ , unsurprisingly, $K$ can be very large. + +(iv) Client heterogeneity. While careful selection of client and server learning rates can reduce the effect of client heterogeneity (see Remark 1), it does not completely remove it. In highly heterogeneous settings, it may be necessary to use mechanisms such as control variates (Karimireddy et al., 2019). However, our empirical analysis suggest that for moderate, naturally arising heterogeneity, adaptive optimizers are quite effective, especially in cross-device settings (see Figure 1). Furthermore, our algorithms can be directly combined with such mechanisms. + +As mentioned earlier, for the sake of simplicity, our analysis assumes full-participation $\mathcal { S } = [ m ] )$ . Our analysis can be directly generalized to limited participation at the cost of an additional variance term in our rates that depends on $| { \cal S } | / m$ , the fraction of clients sampled (see Section A.2.1 for details). + +4 EXPERIMENTAL EVALUATION: DATASETS, TASKS, AND METHODS + +We evaluate our algorithms on what we believe is the most extensive and representative suite of federated datasets and modeling tasks to date. We wish to understand how server adaptivity can help improve convergence, especially in cross-device settings. To accomplish this, we conduct simulations on seven diverse and representative learning tasks across five datasets. Notably, three of the five have a naturally-arising client partitioning, highly representative of real-world FL problems. + +Datasets, models, and tasks We use five datasets: CIFAR-10, CIFAR-100 (Krizhevsky & Hinton, 2009), EMNIST (Cohen et al., 2017), Shakespeare (McMahan et al., 2017), and Stack Overflow (Authors, 2019). The first three are image datasets, the last two are text datasets. For CIFAR-10 and CIFAR-100, we train ResNet-18 (replacing batch norm with group norm (Hsieh et al., 2019)). For EMNIST, we train a CNN for character recognition (EMNIST CR) and a bottleneck autoencoder (EMNIST AE). For Shakespeare, we train an RNN for next-character-prediction. For Stack Overflow, we perform tag prediction using logistic regression on bag-of-words vectors (SO LR) and train an RNN to do next-word-prediction (SO NWP). For full details of the datasets, see Appendix C. + +Implementation We implement all algorithms in TensorFlow Federated (Ingerman & Ostrowski, 2019). Clients are sampled uniformly at random, without replacement in a given round, but with replacement across rounds. Our implementation has two important characteristics. First, instead of doing $K$ training steps per client, we do $E$ epochs of training over each client’s dataset. Second, to account for varying numbers of gradient steps per client, we weight the average of the client outputs $\Delta _ { i } ^ { t }$ by each client’s number of training samples. This follows the approach of (McMahan et al., 2017), and can often outperform uniform weighting (Zaheer et al., 2018). For full descriptions of the algorithms used, see Appendix B. + +Optimizers and hyperparameters We compare FEDADAGRAD, FEDADAM, and FEDYOGI (with adaptivity $\tau$ ) to FEDOPT where CLIENTOPT and SERVEROPT are SGD with learning rates $\eta _ { l }$ and $\eta$ . For the server, we use a momentum parameter of 0 (FEDAVG), and 0.9 (FEDAVGM). We fix the client batch size on a per-task level (see Appendix D.3). For FEDADAM and FEDYOGI, we fix a momentum parameter $\beta _ { 1 } = 0 . 9$ and a second moment parameter $\beta _ { 2 } = 0 . 9 9$ . We also compare to SCAFFOLD (see Appendix B.2 for implementation details). For SO NWP, we sample 50 clients per round, while for all other tasks we sample 10. We use $E = 1$ local epochs throughout. + +We select $\eta , ~ \eta$ , and $\tau$ by grid-search tuning. While this is often done using validation data in centralized settings, such data is often inaccessible in FL, especially cross-device FL. Therefore, we tune by selecting the parameters that minimize the average training loss over the last 100 rounds of training. We run 1500 rounds of training on the EMNIST CR, Shakespeare, and Stack Overflow tasks, 3000 rounds for EMNIST AE, and 4000 rounds for the CIFAR tasks. For more details and a record of the best hyperparameters, see Appendix D. + +Validation metrics For all tasks, we measure the performance on a validation set throughout training. For Stack Overflow tasks, the validation set contains 10,000 randomly sampled test examples (due to the size of the test dataset, see Table 2). For all other tasks, we use the entire test set. Since all algorithms exchange equal-sized objects between server and clients, we use the number of communication rounds as a proxy for wall-clock training time. + +# 5 EXPERIMENTAL EVALUATION: RESULTS + +# 5.1 COMPARISONS BETWEEN METHODS + +We compare the convergence of our adaptive methods to non-adaptive methods: FEDAVG, FEDAVGM and SCAFFOLD. Plots of validation performances for each task/optimizer are in Figure 1, and Table 1 summarizes the last-100-round validation performance. Due to space constraints, results for EMNIST CR are in Appendix E.1, and full test set results for Stack Overflow are in Appendix E.2. + +Sparse-gradient tasks Text data often produces long-tailed feature distributions, leading to approximately-sparse gradients which adaptive optimizers can capitalize on (Zhang et al., 2019a). Both Stack Overflow tasks exhibit such behavior, though they are otherwise dramatically different— in feature representation (bag-of-words vs. variable-length token sequence), model architecture (GLM vs deep network), and optimization landscape (convex vs nonconvex). In both tasks, words that do not appear in a client’s dataset produce nearzero client updates. Thus, the accumulators $v _ { t , j }$ in Algorithm 2 remain small for parameters tied to rare words, allowing large + +
FED..ADAGRADADAMYoGIAvGMAvG
CIFAR-1072.177.478.077.472.8
CIFAR-10047.952.552.452.444.7
EMNIST CR85.185.685.585.284.9
SHAKESPEARE57.557.057.257.356.9
SO NWP23.825.225.223.819.5
SO LR67.165.865.936.930.0
EMNIST AE4.201.010.981.656.47
+ +Table 1: Average validation performance over the last 100 rounds: $\%$ accuracy for rows 1–5; Recall $\textcircled { a } 5$ $( \times 1 0 0 )$ for Stack Overflow LR; and MSE $( \times 1 0 0 0 )$ for EMNIST AE. Performance within $0 . 5 \%$ of the best result for each task are shown in bold. + +updates to be made when they do occur. This intuition is born out in Figure 1, where adaptive optimizers dramatically outperform non-adaptive ones. For the non-convex NWP task, momentum is also critical, whereas it slightly hinders performance for the convex LR task. + +Dense-gradient tasks CIFAR-10/100, EMNIST AE/CR, and Shakespeare lack a sparse-gradient structure. Shakespeare is relatively easy—most optimizers perform well after enough rounds once suitably tuned, though FEDADAGRAD converges faster. For CIFAR-10/100 and EMNIST AE, adaptivity and momentum offer substantial improvements over FEDAVG. Moreover, FEDYOGI and FEDADAM have faster initial convergence than FEDAVGM on these tasks. Notably, FEDADAM and FEDYOGI perform comparably to or better than non-adaptive optimizers throughout, and close to or better than + +![](images/291f261557be507186a2a9304e87c9225666c4791011ea1091541e22f3f95278.jpg) +Figure 1: Validation accuracy of adaptive and non-adaptive methods, as well as SCAFFOLD, using constant learning rates $\eta$ and $\eta _ { l }$ tuned to achieve the best training performance over the last 100 communication rounds; see Appendix D.2 for grids. + +FEDADAGRAD throughout. As we discuss below, FEDADAM and FEDYOGI actually enable easier learning rate tuning than FEDAVGM in many tasks. + +Comparison to SCAFFOLD On all tasks, SCAFFOLD performs comparably to or worse than FEDAVG and our adaptive methods. On Stack Overflow, SCAFFOLD and FEDAVG are nearly identical. This is because the number of clients (342,477) makes it unlikely we sample any client more than once. Intuitively, SCAFFOLD does not have a chance to use its client control variates. In other tasks, SCAFFOLD performs worse than other methods. We present two possible explanations: First, we only sample a small fraction of clients at each round, so most users are sampled infrequently. Intuitively, the client control variates can become stale, and may consequently degrade the performance. Second, SCAFFOLD is similar to variance reduction methods such as SVRG (Johnson & Zhang, 2013). While theoretically performant, such methods often perform worse than SGD in practice (Defazio & Bottou, 2018). As shown by Defazio et al. (2014), variance reduction often only accelerates convergence when close to a critical point. In cross-device settings (where the number of communication rounds are limited), SCAFFOLD may actually reduce empirical performance. + +# 5.2 EASE OF TUNING + +![](images/a1ecef7595988e2f35b8b5c2ee3db0d5eb2f0bad2ba1b6103616e6c3ce2fd289.jpg) +Figure 2: Validation accuracy (averaged over the last 100 rounds) of FEDADAM, FEDYOGI, and FEDAVGM for various client/server learning rates combination on the SO NWP task. For FEDADAM and FEDYOGI, we set $\tau = 1 0 ^ { - 3 }$ . + +Obtaining optimal performance involves tuning $\eta _ { l } , \eta$ , and for the adaptive methods, $\tau$ . To quantify how easy it is to tune various methods, we plot their validation performance as a function of $\eta _ { l }$ and $\eta$ Figure 2 gives results for FEDADAM, FEDYOGI, and FEDAVGM on Stack Overflow NWP. Plots for all other optimizers and tasks are in Appendix E.3. For FEDAVGM, there are only a few good values of $\eta _ { l }$ for each $\eta$ , while for FEDADAM and FEDYOGI, there are many good values of $\eta _ { l }$ for a range of $\eta$ . Thus, FEDADAM and FEDYOGI are arguably easier to tune in this setting. Similar results hold for other tasks and optimizers (Figures 5 to 11). + +![](images/cd2738f345eafd8cd9372344ebd4c03605653f82aa8c8583ca069896d16d9d6c.jpg) +Figure 3: Validation performance of FEDADAGRAD, FEDADAM, and FEDYOGI for varying $\tau$ on various tasks. The learning rates $\eta$ and $\eta _ { l }$ are tuned for each $\tau$ to achieve the best training performance on the last 100 communication rounds. + +This leads to a natural question: Is the reduction in the need to tune $\eta _ { l }$ and $\eta$ offset by the need to tune the adaptivity $\tau ?$ In fact, while we tune $\tau$ in Figure 1, our results are relatively robust to $\tau$ . To demonstrate, we plot the best validation performance for various $\tau$ in Figure 3. For nearly all tasks and optimizers, $\dot { \tau } = 1 0 ^ { - 3 }$ works almost as well all other values. This aligns with work by Zaheer et al. (2018), who show that moderately large $\tau$ yield better performance for centralized adaptive optimizers. FEDADAM and FEDYOGI see only small differences in performance among $\tau$ on all tasks except Stack Overflow LR (for which FEDADAGRAD is the best optimizer, and is robust to $\tau$ ). + +# 5.3 OTHER FINDINGS + +We present additional empirical analyses in Appendix E. These include EMNIST CR results (Appendix E.1), Stack Overflow results on the full test dataset (Appendix E.2), client/server learning rate heat maps for all optimizers and tasks (Appendix E.3), an analysis of the relationship between $\eta$ and $\eta _ { l }$ (Appendix E.4), and experiments with learning rate decay (Appendix E.6). + +# 6 CONCLUSION + +In this paper, we demonstrated that adaptive optimizers can be powerful tools in improving the convergence of FL. By using a simple client/server optimizer framework, we can incorporate adaptivity into FL in a principled, intuitive, and theoretically-justified manner. We also developed comprehensive benchmarks for comparing federated optimization algorithms. To encourage reproducibility and breadth of comparison, we have attempted to describe our experiments as rigorously as possible, and have created an open-source framework with all models, datasets, and code. We believe our work raises many important questions about how best to perform federated optimization. Example directions for future research include understanding how the use of adaptivity affects differential privacy and fairness. + +# REFERENCES + +The TensorFlow Federated Authors. TensorFlow Federated Stack Overflow dataset, 2019. URL https://www.tensorflow.org/federated/api_docs/python/tff/ simulation/datasets/stackoverflow/load_data. + +Debraj Basu, Deepesh Data, Can Karakus, and Suhas Diggavi. Qsparse-local-SGD: Distributed SGD with quantization, sparsification and local computations. In Advances in Neural Information Processing Systems, pp. 14668–14679, 2019. + +Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex Ingerman, Vladimir Ivanov, Chloé Kiddon, Jakub Konecný, Stefano Mazzocchi, Brendan McMahan, Timon ˇ Van Overveldt, David Petrou, Daniel Ramage, and Jason Roselander. Towards federated learning at scale: System design. In A. Talwalkar, V. Smith, and M. Zaharia (eds.), Proceedings of Machine Learning and Systems, volume 1, pp. 374–388. Proceedings of MLSys, 2019. URL https://proceedings.mlsys.org/paper/2019/file/ bd686fd640be98efaae0091fa301e613-Paper.pdf. + +Sebastian Caldas, Peter Wu, Tian Li, Jakub Konecný, H Brendan McMahan, Virginia Smith, and ˇ Ameet Talwalkar. LEAF: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. + +Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. EMNIST: Extending MNIST to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pp. 2921–2926. IEEE, 2017. + +Aaron Defazio and Léon Bottou. On the ineffectiveness of variance reduced optimization for deep learning. arXiv preprint arXiv:1812.04529, 2018. + +Aaron Defazio, Francis Bach, and Simon Lacoste-Julien. SAGA: A fast incremental gradient method with support for non-strongly convex composite objectives. In NIPS, pp. 1646–1654, 2014. + +John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011. + +Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training ImageNet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. + +Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip B Gibbons. The non-IID data quagmire of decentralized machine learning. arXiv preprint arXiv:1910.00189, 2019. + +Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. + +Alex Ingerman and Krzys Ostrowski. Introducing TensorFlow Federated, 2019. URL https://medium.com/tensorflow/ introducing-tensorflow-federated-a4147aa20041. + +Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive variance reduction. In Advances in Neural Information Processing Systems, pp. 315–323, 2013. + +Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Keith Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. arXiv preprint arXiv:1912.04977, 2019. + +Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for on-device federated learning. arXiv preprint arXiv:1910.06378, 2019. + +Ahmed Khaled, Konstantin Mishchenko, and Peter Richtárik. First analysis of local GD on heterogeneous data. arXiv preprint arXiv:1909.04715, 2019. + +Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. + +Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009. + +Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. arXiv preprint arXiv:1812.06127, 2018. + +Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. arXiv preprint arXiv:1908.07873, 2019a. + +Wei Li and Andrew McCallum. Pachinko allocation: DAG-structured mixture models of topic correlations. In Proceedings of the 23rd international conference on Machine learning, pp. 577– 584, 2006. + +Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. arXiv preprint arXiv:1907.02189, 2019b. + +Xiaoyu Li and Francesco Orabona. On the convergence of stochastic gradient descent with adaptive stepsizes. arXiv preprint arXiv:1805.08114, 2018. + +Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic bound of learning rate. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. URL https://openreview. net/forum?id=Bkg3g2R9FX. + +Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April 2017, Fort Lauderdale, FL, USA, pp. 1273–1282, 2017. URL http://proceedings.mlr. press/v54/mcmahan17a.html. + +H. Brendan McMahan and Matthew Streeter. Adaptive bound optimization for online convex optimization. In COLT, 2010a. + +H. Brendan McMahan and Matthew J. Streeter. Adaptive bound optimization for online convex optimization. In COLT The 23rd Conference on Learning Theory, 2010b. + +Sashank J Reddi, Ahmed Hefny, Suvrit Sra, Barnabás Póczós, and Alex Smola. Stochastic variance reduction for nonconvex optimization. arXiv:1603.06160, 2016. + +Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of ADAM and beyond. arXiv preprint arXiv:1904.09237, 2019. + +Sebastian U. Stich. Local SGD converges fast and communicates little. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id= S1g2JnRcFX. + +Sebastian U Stich and Sai Praneeth Karimireddy. The error-feedback framework: Better rates for SGD with delayed gradients and compressed communication. arXiv preprint arXiv:1909.05350, 2019. + +Jianyu Wang and Gauri Joshi. Cooperative SGD: A unified framework for the design and analysis of communication-efficient SGD algorithms. arXiv preprint arXiv:1808.07576, 2018. + +Shiqiang Wang, Tiffany Tuor, Theodoros Salonidis, Kin K Leung, Christian Makaya, Ting He, and Kevin Chan. Adaptive federated learning in resource constrained edge computing systems. IEEE Journal on Selected Areas in Communications, 37(6):1205–1221, 2019. + +Rachel Ward, Xiaoxia Wu, and Leon Bottou. Adagrad stepsizes: Sharp convergence over nonconvex landscapes, from any initialization. arXiv preprint arXiv:1806.01811, 2018. + +Xiaoxia Wu, Simon S Du, and Rachel Ward. Global convergence of adaptive gradient methods for an over-parameterized neural network. arXiv preprint arXiv:1902.07111, 2019. + +Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 3–19, 2018. + +Cong Xie, Oluwasanmi Koyejo, Indranil Gupta, and Haibin Lin. Local AdaAlter: Communicationefficient stochastic gradient descent with adaptive learning rates. arXiv preprint arXiv:1911.09030, 2019. + +Hao Yu, Sen Yang, and Shenghuo Zhu. Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pp. 5693–5700, 2019. + +Manzil Zaheer, Sashank Reddi, Devendra Sachan, Satyen Kale, and Sanjiv Kumar. Adaptive methods for nonconvex optimization. In Advances in Neural Information Processing Systems, pp. 9815– 9825, 2018. + +Jingzhao Zhang, Sai Praneeth Karimireddy, Andreas Veit, Seungyeon Kim, Sashank J. Reddi, Sanjiv Kumar, and Suvrit Sra. Why ADAM beats SGD for attention models. arXiv preprint arxiv:1912.03194, 2019a. + +Michael R. Zhang, James Lucas, Jimmy Ba, and Geoffrey E. Hinton. Lookahead optimizer: k steps forward, 1 step back. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, 8-14 December 2019, Vancouver, BC, Canada, pp. 9593–9604, 2019b. + +Martin Zinkevich, Markus Weimer, Lihong Li, and Alex J Smola. Parallelized stochastic gradient descent. In Advances in neural information processing systems, pp. 2595–2603, 2010. + +# A PROOF OF RESULTS + +# A.1 MAIN CHALLENGES + +We first recap some of the central challenges to our analysis. Theoretical analyses of optimization methods for federated learning are much different than analyses for centralized settings. The key factors complicating the analysis are: + +1. Clients performing multiple local updates. +2. Data heterogeneity. +3. Understanding the communication complexity. + +As a result of (1), the updates from the clients to the server are not gradients, or even unbiased estimates of gradients, they are pseudo-gradients (see Section 2). These pseudo-gradients are challenging to analyze as they can have both high bias (their expectation is not the gradient of the empirical loss function) and high variance (due to compounding variance across client updates) and are therefore challenging to bound. This is exacerbated by (2), which we quantify by the parameter $\sigma _ { g }$ in Section 2. Things are further complicated by (3), as we must obtain a good trade-off between the number of client updates taken per round ( $K$ in Algorithms 1 and 2) and the number of communication rounds $T$ . Such trade-offs do not exist in centralized optimization. + +# A.2 PROOF OF THEOREM 1 + +Proof of Theorem 1. Recall that the server update of FEDADAGRAD is the following + +$$ +x _ { t + 1 , i } = x _ { t , i } + \eta \frac { \Delta _ { t , i } } { \sqrt { v _ { t , i } } + \tau } , +$$ + +for all $i \in [ d ]$ . Since the function $f$ is $L$ -smooth, we have the following: + +$$ +\begin{array} { l } { f ( x _ { t + 1 } ) \leq f ( x _ { t } ) + \langle \nabla f ( x _ { t } ) , x _ { t + 1 } - x _ { t } \rangle + \displaystyle \frac { L } { 2 } \| x _ { t + 1 } - x _ { t } \| ^ { 2 } } \\ { \displaystyle = f ( x _ { t } ) + \eta \left. \nabla f ( x _ { t } ) , \frac { \Delta _ { t } } { \sqrt { v _ { t } } + \tau } \right. + \displaystyle \frac { \eta ^ { 2 } L } { 2 } \sum _ { i = 1 } ^ { d } \frac { \Delta _ { t , i } ^ { 2 } } { ( \sqrt { v _ { t , i } } + \tau ) ^ { 2 } } } \end{array} +$$ + +The second step follows simply from FEDADAGRAD’s update. We take the expectation of $f ( x _ { t + 1 } )$ (over randomness at time step $t$ ) in the above inequality: + +$$ +\begin{array} { l } { \displaystyle \ddot { c } _ { i } [ f ( x _ { i + 1 } ) ] \leq f ( x _ { i } ) + \eta \left. \nabla f ( x _ { i } ) , \mathbb { E } _ { k } \left[ \frac { \Delta _ { i } } { \sqrt { \Psi _ { t - 1 } } + \tau } \right] \right. + \frac { \eta ^ { 2 } L } { 2 } \displaystyle \sum _ { i = 1 } ^ { d } \mathbb { E } _ { k } \left[ \frac { \Delta _ { i } ^ { 2 } \rho } { \left( \sqrt { \nu _ { t - 1 } } + \tau \right) ^ { 2 } } \right] } \\ { = f ( x _ { i } ) + \eta \left. \nabla f ( x _ { i } ) , \mathbb { E } _ { k } \left[ \frac { \Delta _ { i } } { \sqrt { \nu _ { t + 1 } } + \tau } - \frac { \Delta _ { i } } { \sqrt { \nu _ { t - 1 } } + \tau } + \frac { \Delta _ { i } } { \sqrt { \nu _ { t - 1 } } + \tau } \right] \right. } \\ { \displaystyle \qquad + \frac { \eta ^ { 2 } L } { 2 } \displaystyle \sum _ { i = 1 } ^ { d } \mathbb { E } _ { k } \left[ \frac { \Delta _ { i , j } ^ { 2 } } { \left( \sqrt { \nu _ { t , j } } + \tau \right) ^ { 2 } } \right] } \\ { = f ( x _ { i } ) + \eta \underbrace { \left. \nabla f ( x _ { i } ) , \mathbb { E } _ { k } \left[ \frac { \Delta _ { i } } { \sqrt { \nu _ { t - 1 } } + \tau } \right] \right. } _ { \displaystyle \mathrm { ~ ( ~ 1 ~ ) ~ } } + \eta \left. \nabla f ( x _ { i } ) , \mathbb { E } _ { k } \left[ \frac { \Delta _ { i } } { \sqrt { \nu _ { t } } + \tau } - \frac { \Delta _ { i } } { \sqrt { \nu _ { t - 1 } } + \tau } \right] \right. } \\ { \displaystyle \qquad + \frac { \eta ^ { 2 } L } { 2 } \displaystyle \sum _ { i = 1 } ^ { d } \mathbb { E } _ { k } \left[ \frac { \Delta _ { i , j } ^ { 2 } } { \left( \sqrt { \nu _ { t - 1 } } + \tau \right) ^ { 2 } } \right] } \end{array} +$$ + +We will first bound $T _ { 2 }$ in the following manner: + +$$ +\begin{array} { r l } & { T _ { 2 } = \bigg \langle \nabla f ( { x } _ { t } ) , \mathbb { E } _ { t } \left[ \frac { \Delta _ { t } } { \sqrt { v _ { t } } + \tau } - \frac { \Delta _ { t } } { \sqrt { v _ { t - 1 } } + \tau } \right] \bigg \rangle } \\ & { \quad = \mathbb { E } _ { t } \displaystyle \sum _ { j = 1 } ^ { d } [ \nabla f ( { x } _ { t } ) ] _ { j } \times \left[ \frac { \Delta _ { t , j } } { \sqrt { v _ { t , j } } + \tau } - \frac { \Delta _ { t , j } } { \sqrt { v _ { t - 1 , j } } + \tau } \right] } \\ & { \quad = \mathbb { E } _ { t } \displaystyle \sum _ { j = 1 } ^ { d } [ \nabla f ( { x } _ { t } ) ] _ { j } \times \Delta _ { t , j } \times \left[ \frac { \sqrt { v _ { t - 1 , j } } - \sqrt { v _ { t , j } } } { ( \sqrt { v _ { t , j } } + \tau ) ( \sqrt { v _ { t - 1 , j } } + \tau ) } \right] } \end{array} +$$ + +and recalling $v _ { t } = v _ { t - 1 } + \Delta _ { t } ^ { 2 }$ so $- \Delta _ { t , j } ^ { 2 } = ( \sqrt { v _ { t - 1 , j } } - \sqrt { v _ { t , j } } ) ( \sqrt { v _ { t - 1 , j } } + \sqrt { v _ { t , j } } ) )$ we have, + +$$ +\begin{array} { r l r } { { = \mathbb { E } _ { t } \sum _ { j = 1 } ^ { d } [ \nabla f ( x _ { t } ) ] _ { j } \times \Delta _ { t , j } \times [ \overline { { ( \sqrt { v _ { t , j } } + \tau ) ( \sqrt { v _ { t - 1 , j } } + \tau ) ( \sqrt { v _ { t - 1 , j } } + \sqrt { v _ { t , j } } ) } } ] } } \\ & { \leq \mathbb { E } _ { t } \sum _ { j = 1 } ^ { d } \vert \nabla f ( x _ { t } ) \vert _ { j } \vert \times \vert \Delta _ { t , j } \vert \times [ \overline { { ( \sqrt { v _ { t , j } } + \tau ) ( \sqrt { v _ { t - 1 , j } } + \tau ) ( \sqrt { v _ { t - 1 , j } } + \sqrt { v _ { t , j } } ) } } ] } \\ & { \leq \mathbb { E } _ { t } \sum _ { j = 1 } ^ { d } \vert \nabla f ( x _ { t } ) \vert _ { j } \vert \times \vert \Delta _ { t , j } \vert \times [ \overline { { ( v _ { t , j } + \tau ) ( \sqrt { v _ { t - 1 , j } } + \tau ) } } ] } & { \mathrm { s i n c e } v _ { t - 1 , j } \geq \tau } \\ & { \leq \mathbb { E } _ { t } \sum _ { j = 1 } ^ { d } \vert \nabla f ( x _ { t } ) \vert _ { j } \vert \times \vert \Delta _ { t , j } \vert \times [ \overline { { ( v _ { t , j } + \tau ^ { 2 } ) ( \sqrt { v _ { t - 1 , j } } + \tau ) } } ] } & { \mathrm { s i n c e } v _ { t - 1 , j } \geq \tau } \end{array} +$$ + +Here $v _ { t - 1 , j } \geq \tau$ since $v _ { - 1 } \geq \tau$ (see the initialization of Algorithm 2) and $v _ { t , j }$ is increasing in $t$ . The above bound can be further upper bounded in the following manner: + +$$ +\begin{array} { r l } { T _ { 2 } \le \mathbb { E } _ { t } \displaystyle \sum _ { j = 1 } ^ { d } \eta _ { l } K G ^ { 2 } \left[ \displaystyle \frac { \Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) ( \sqrt { v _ { t - 1 , j } } + \tau ) } \right] } & { \mathrm { ~ s i n c e ~ } [ \nabla f ( x _ { t } ) ] _ { i } \le G \mathrm { ~ a n d ~ } \Delta _ { t , i } \le \eta _ { l } K G } \\ { \le \mathbb { E } _ { t } \displaystyle \sum _ { j = 1 } ^ { d } \displaystyle \frac { \eta _ { l } K G ^ { 2 } } { \tau } \left[ \displaystyle \frac { \Delta _ { t , j } ^ { 2 } } { \sum _ { l = 0 } ^ { t } \Delta _ { l , j } ^ { 2 } + \tau ^ { 2 } } \right] } & { \mathrm { ~ s i n c e ~ } \sqrt { v _ { t - 1 , j } } \ge 0 . } \end{array} +$$ + +Bounding $T _ { 1 }$ We now turn our attention to bounding the term $T _ { 1 }$ , which we need to be sufficiently negative. We observe the following: + +$$ +\begin{array} { r l } & { T _ { 1 } = \bigg \langle \nabla f ( x _ { t } ) , \mathbb { E } _ { t } \left[ \frac { \Delta _ { t } } { \sqrt { v _ { t - 1 } } + \tau } \right] \bigg \rangle } \\ & { \quad = \bigg \langle \frac { \nabla f ( x _ { t } ) } { \sqrt { v _ { t - 1 } } + \tau } , \mathbb { E } _ { t } \left[ \Delta _ { t } - \eta _ { l } K \nabla f ( x _ { t } ) + \eta _ { l } K \nabla f ( x _ { t } ) \right] \bigg \rangle } \\ & { \quad = - \eta _ { l } K \displaystyle \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } + \underbrace { \left. \frac { \nabla f ( x _ { t } ) } { \sqrt { v _ { t - 1 } } + \tau } , \mathbb { E } _ { t } \left[ \Delta _ { t } + \eta _ { l } K \nabla f ( x _ { t } ) \right] \right. } _ { T _ { 2 } } . } \end{array} +$$ + +In order to bound $T _ { 1 }$ , we use the following upper bound on $T _ { 3 }$ (which captures the difference between the actual update $\Delta _ { t }$ and an appropriate scaling of $- \nabla f ( x _ { t } ) )$ : + +$$ +\begin{array} { r l } & { T _ { 3 } = \left. \frac { \nabla f ( \boldsymbol { x } _ { t } ) } { \sqrt { v _ { t - 1 } } + \tau } , \mathbb { E } _ { t } \left[ \Delta _ { t } + \eta _ { l } K \nabla f ( \boldsymbol { x } _ { t } ) \right] \right. } \\ & { \quad = \left. \frac { \nabla f ( \boldsymbol { x } _ { t } ) } { \sqrt { v _ { t - 1 } } + \tau } , \mathbb { E } _ { t } \left[ - \displaystyle \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \eta _ { l } g _ { i , k } ^ { t } + \eta _ { l } K \nabla f ( \boldsymbol { x } _ { t } ) \right] \right. } \\ & { \quad = \left. \frac { \nabla f ( \boldsymbol { x } _ { t } ) } { \sqrt { v _ { t - 1 } } + \tau } , \mathbb { E } _ { t } \left[ - \displaystyle \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \eta _ { l } \nabla F _ { i } ( x _ { i , k } ^ { t } ) + \eta _ { l } K \nabla f ( \boldsymbol { x } _ { t } ) \right] \right. . } \end{array} +$$ + +Here we used the fact that $\begin{array} { r } { \nabla f ( x _ { t } ) = \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \nabla F _ { i } ( x _ { t } ) } \end{array}$ and $g _ { i , k } ^ { t }$ is an unbiased estimator of the gradient at $\boldsymbol { x } _ { i , k } ^ { t }$ , we further bound $T _ { 3 }$ as follows using a simple application of the fact that $a b \leq$ $( a ^ { 2 } + b ^ { 2 } ) / 2$ . : + +$$ +\begin{array} { r l } { r _ { 3 } \leq \displaystyle \frac { \eta _ { l } K } { 2 } \sum _ { j = 1 } ^ { d } \frac { \displaystyle \sum _ { j } ^ { d } } { \sqrt { v _ { t - 1 , j } } + \tau } + \frac { \eta _ { l } } { 2 K } \mathbb { E } _ { t } \left[ \left\| \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \frac { \nabla F _ { i } ( x _ { i , k } ^ { t } ) } { \sqrt { v _ { t - 1 } } + \tau } - \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \frac { \nabla F _ { i } ( x _ { t } ) } { \sqrt { v _ { t - 1 } } + \tau } \right\| ^ { 2 } \right] } & { } \\ { \leq \displaystyle \frac { \eta _ { l } K } { 2 } \sum _ { j = 1 } ^ { d } \frac { \displaystyle \left[ \nabla f ( x _ { t } ) \right] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } + \frac { \eta _ { l } } { 2 m } \mathbb { E } _ { t } \left[ \displaystyle \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \left\| \frac { \nabla F _ { i } ( x _ { i , k } ^ { t } ) - \nabla F _ { i } ( x _ { t } ) } { \sqrt { v _ { t - 1 } } + \tau } \right\| ^ { 2 } \right] } & { } \\ { \leq \displaystyle \frac { \eta _ { l } K } { 2 } \sum _ { j = 1 } ^ { d } \frac { \displaystyle \left[ \nabla f ( x _ { t } ) \right] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } + \frac { \eta _ { l } L ^ { 2 } } { 2 m \tau } \mathbb { E } _ { t } \left[ \displaystyle \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \| x _ { i , k } ^ { t } - x _ { t } \| ^ { 2 } \right] } & { \mathrm { u s i n g ~ A s s u m p t i o n ~ 1 ~ a n d ~ } v _ { t - 1 } \geq 0 } \end{array} +$$ + +The second inequality follows from Lemma 6. The last inequality follows from $L$ -Lipschitz nature of the gradient (Assumption 1). We now prove a lemma that bounds the “drift” of the $\dot { \boldsymbol { x } } _ { i , k } ^ { t }$ from $x _ { t }$ : + +Lemma 3. For any step-size satisfying $\begin{array} { r } { \eta _ { l } \le \frac { 1 } { 8 L K } } \end{array}$ , we can bound the drift for any $k \in \{ 0 , \cdots , K - 1 \}$ as + +$$ +\frac { 1 } { m } \sum _ { i = 1 } ^ { m } \mathbb { E } \| x _ { i , k } ^ { t } - x _ { t } \| ^ { 2 } \leq 5 K \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 2 K \sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \eta _ { l } ^ { 2 } \mathbb { E } [ \| \nabla f ( x _ { t } ) ) ) \| ^ { 2 } \big ] . +$$ + +Proof. The result trivially holds for $k = 1$ since $x _ { i , 0 } ^ { t } = x _ { t }$ for all $i \in [ m ]$ . We now turn our attention to the case where $k \geq 1$ . To prove the above result, we observe that for any client $i \in [ m ]$ and $k \in [ K ]$ , + +$$ +\begin{array} { r l } { { \tilde { \mathbf { \mu } } _ { i , k , - } ^ { t } - x _ { t } \| ^ { 2 } = \mathbb { E } \big \| x _ { i , k - 1 } ^ { t } - x _ { t } - \eta _ { l } g _ { i , k - 1 } ^ { t } \big \| ^ { 2 } } } \\ & { \leq \mathbb { E } \big \| x _ { i , k - 1 } ^ { t } - x _ { t } - \eta _ { l } ( g _ { i , k - 1 } ^ { t } - \nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) + \nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \nabla F _ { i } ( x _ { t } ) + \nabla F _ { i } ( x _ { t } ) - \nabla f ( x _ { t } ) + \nabla F _ { i } ( x _ { t } ) - \nabla F ( x _ { t } ) \big \| ^ { 2 } } \\ & { \leq ( 1 + \frac { 1 } { 2 K - 1 } ) \mathbb { E } \big \| x _ { i , k - 1 } ^ { t } - x _ { t } \big \| ^ { 2 } + \mathbb { E } \big \| \eta _ { l } ( g _ { i , k - 1 } ^ { t } - \nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) ) \big \| ^ { 2 } } \\ & { \qquad + 6 K \mathbb { E } [ \big \| \eta _ { l } ( \nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \nabla F _ { i } ( x _ { t } ) ) \big \| ^ { 2 } ] + 6 K \mathbb { E } [ \big \| \eta _ { l } ( \nabla F _ { i } ( x _ { t } ) - \nabla f ( x _ { t } ) ) \big \| ^ { 2 } \big ] + 6 K \mathbb { E } [ \big \| \eta _ { l } } \end{array} +$$ + +The first inequality uses the fact that $g _ { k - 1 , i } ^ { t }$ is an unbiased estimator of $\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } )$ and Lemma 7. The above quantity can be further bounded by the following: + +$$ +\begin{array} { r l r } { { \mathbb { E } \Vert x _ { i , k } ^ { t } - x _ { t } \Vert ^ { 2 } \leq ( 1 + \frac { 1 } { 2 K - 1 } ) \mathbb { E } \Vert x _ { i , k - 1 } ^ { t } - x _ { t } \Vert ^ { 2 } + \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } \sigma _ { l , j } ^ { 2 } + 6 K \eta _ { l } ^ { 2 } \mathbb { E } \Vert L ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \Vert ^ { 2 } } } \\ & { } & { + 6 K \mathbb { E } [ \Vert \eta _ { l } ( \nabla F _ { i } ( x _ { t } ) - \nabla f ( x _ { t } ) ) \Vert ^ { 2 } ] + 6 K \mathbb { E } [ \Vert \eta _ { l } \nabla f ( x _ { t } ) ) ) \Vert ^ { 2 } ] } \\ & { } & { = ( 1 + \frac { 1 } { 2 K - 1 } + 6 K \eta _ { l } ^ { 2 } L ^ { 2 } ) \mathbb { E } \Vert ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \Vert ^ { 2 } + \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } \sigma _ { l , j } ^ { 2 } } \\ & { } & { + 6 K \mathbb { E } [ \Vert \eta _ { l } ( \nabla F _ { i } ( x _ { t } ) - \nabla f ( x _ { t } ) ) \Vert ^ { 2 } ] + 6 K \eta _ { l } ^ { 2 } \mathbb { E } [ \Vert \nabla f ( x _ { t } ) ) ) \Vert ^ { 2 } ] } \end{array} +$$ + +Here, the first inequality follows from Assumption 1 and 2. Averaging over the clients $i$ , we obtain the following: + +$$ +\begin{array} { r l r } { { \frac { 1 } { n } \sum _ { i = 1 } ^ { m } \mathbb { E } \| x _ { i , k } ^ { t } - x _ { t } \| ^ { 2 } \leq ( 1 + \frac { 1 } { 2 K - 1 } + 6 K \eta _ { l } ^ { 2 } L ^ { 2 } ) \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \mathbb { E } \| x _ { i , k - 1 } ^ { t } - x _ { t } \| ^ { 2 } + \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } \sigma _ { l , j } ^ { 2 } } } \\ & { } & { + \frac { 6 K } { m } \sum _ { i = 1 } ^ { m } \mathbb { E } [ \| \eta _ { l } ( \nabla F _ { i } ( x _ { t } ) - \nabla f ( x _ { t } ) ) \| ^ { 2 } ] + 6 K \eta _ { l } ^ { 2 } \mathbb { E } [ \| \nabla f ( x _ { t } ) ) ) } \\ & { } & { \leq ( 1 + \frac { 1 } { 2 K - 1 } + 6 K \eta _ { l } ^ { 2 } L ^ { 2 } ) \frac { 1 } { m } \displaystyle \sum _ { i = 1 } ^ { m } \mathbb { E } \| x _ { i , k - 1 } ^ { t } - x _ { t } \| ^ { 2 } + \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K o ( \eta _ { l } ^ { 2 } ) ) } \\ & { } & { + 6 K \eta _ { l } ^ { 2 } \mathbb { E } [ \| \nabla f ( x _ { t } ) ) \| ^ { 2 } ] } \end{array} +$$ + +From the above, we get the following inequality: + +$$ +\begin{array} { r l r } { { \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \mathbb { E } \| x _ { i , k } ^ { t } - x _ { t } \| ^ { 2 } \leq ( 1 + \frac { 1 } { K - 1 } ) \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \mathbb { E } \| x _ { i , k - 1 } ^ { t } - x _ { t } \| ^ { 2 } + \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } } \\ & { } & { \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad + 6 K \eta _ { l } ^ { 2 } \mathbb { E } [ \| \nabla f ( x _ { t } ) ) ) \| ^ { 2 } \Big ] } \end{array} +$$ + +Unrolling the recursion, we obtain the following: + +$$ +\begin{array} { r l } & { \displaystyle \frac { 1 } { n } \sum _ { i = 1 } ^ { m } \mathbb { E } \| x _ { i , k } ^ { t } - x _ { t } \| ^ { 2 } \leq \sum _ { p = 0 } ^ { k - 1 } \left( 1 + \frac { 1 } { K - 1 } \right) ^ { p } \left[ \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) + 6 K \eta _ { l } ^ { 2 } \mathbb { E } [ \| \nabla f ( x _ { t } ) ) ) \| ^ { 2 } \right] } \\ & { \leq \left( K - 1 \right) \times \left[ \left( 1 + \frac { 1 } { K - 1 } \right) ^ { K } - 1 \right] \times \left[ \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) + 6 K \eta _ { l } ^ { 2 } \mathbb { E } [ \| \nabla f ( x _ { t } ) ) ) \| ^ { 2 } \right] } \\ & { \leq \left[ 5 K \eta _ { l } ^ { 2 } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \eta _ { l } ^ { 2 } \mathbb { E } [ \| \nabla f ( x _ { t } ) ) ) \| ^ { 2 } \right] , } \end{array} +$$ + +concluding the proof of Lemma 3. The last inequality uses the fact that $( 1 + \frac { 1 } { K - 1 } ) ^ { K } \leq 5$ for $K > 1$ . + +Using the above lemma in Equation 6 and Condition I, we can bound $T _ { 3 }$ in the following manner: + +$$ +\begin{array} { r l } & { T _ { 3 } \leq \displaystyle \frac { \eta _ { l } K } { 2 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } + \frac { \eta _ { l } L ^ { 2 } } { 2 m \tau } \mathbb { E } _ { t } \left[ \displaystyle \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \sum _ { j = 1 } ^ { d } ( [ x _ { i , k } ^ { t } ] _ { j } - [ x _ { t } ] _ { j } ) ^ { 2 } \right] } \\ & { \quad \leq \displaystyle \frac { \eta _ { l } K } { 2 } \displaystyle \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } + \frac { \eta _ { l } K L ^ { 2 } } { 2 \tau } \left[ 5 K \eta _ { l } ^ { 2 } \mathbb { E } \displaystyle \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \eta _ { l } ^ { 2 } \mathbb { E } [ \| \nabla f ( x _ { t } ) ) ) \| ^ { 2 } \right] } \\ & { \quad \leq \displaystyle \frac { 3 \eta _ { l } K } { 4 } \displaystyle \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } + \frac { 5 \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \tau } \mathbb { E } \displaystyle \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } \end{array} +$$ + +Here we used the fact that $\sqrt { v _ { t - 1 , j } } \leq \eta _ { l } K G \sqrt { T }$ and Condition I in Theorem 1. Using the above bound in Equation 5, we get + +$$ +T _ { 1 } \leq - \frac { \eta _ { l } K } { 4 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } + \frac { 5 \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \tau } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) +$$ + +Putting the pieces together Substituting in Equation (3), bounds $T _ { 1 }$ in Equation (8) and bound $T _ { 2 }$ in Equation (4), we obtain + +$$ +\begin{array} { r l } & { \mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \le f ( x _ { t } ) + \eta \times \left[ - \frac { \eta _ { l } K } { 4 } \displaystyle \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } + \frac { 5 \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \tau } \mathbb { E } \displaystyle \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) \right] } \\ & { \quad \quad \quad \quad \quad + \eta \times \mathbb { E } \displaystyle \sum _ { j = 1 } ^ { d } \frac { \eta _ { l } K G ^ { 2 } } { \tau } \left[ \frac { \Delta _ { t , j } ^ { 2 } } { \displaystyle \sum _ { l = 0 } ^ { t } \Delta _ { l , j } ^ { 2 } + \tau ^ { 2 } } \right] + \frac { \eta ^ { 2 } } { 2 } \displaystyle \sum _ { j = 1 } ^ { d } L \mathbb { E } \left[ \frac { \Delta _ { t , j } ^ { 2 } } { \displaystyle \sum _ { l = 0 } ^ { t } \Delta _ { l , j } ^ { 2 } + \tau ^ { 2 } } \right] . } \end{array} +$$ + +Rearranging the above inequality and summing it from $t = 0$ to $T - 1$ , we get + +$$ +\begin{array} { r } { \displaystyle \sum _ { t = 0 } ^ { T - 1 } \frac { \eta _ { l } K } { 4 } \sum _ { j = 1 } ^ { d } \mathbb { E } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } \leq \frac { f ( x _ { 0 } ) - \mathbb { E } [ f ( x _ { T } ) ] } { \eta } + \frac { 5 \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \tau } \displaystyle \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } \\ { + \displaystyle \sum _ { t = 0 } ^ { T - 1 } \mathbb { E } \sum _ { j = 1 } ^ { d } \left( \frac { \eta _ { l } K G ^ { 2 } } { \tau } + \frac { \eta L } { 2 } \right) \times \left[ \frac { \Delta _ { t , j } ^ { 2 } } { \sum _ { l = 0 } ^ { t } \Delta _ { l , j } ^ { 2 } + \tau ^ { 2 } } \right] } \end{array} +$$ + +The first inequality uses simple telescoping sum. For completing the proof, we need the following result. + +Lemma 4. The following upper bound holds for Algorithm 2 (FEDADAGRAD): + +$$ +\begin{array} { r l } & { \underset { t = 0 } { \overset { T \mathrm { - } 1 } { \sum } } \underset { j = 1 } { \overset { d } { \sum } } \frac { \Delta _ { t , j } ^ { 2 } } { \sum _ { l = 0 } ^ { t } \Delta _ { l , j } ^ { 2 } + \tau ^ { 2 } } \leq \Bigg [ \operatorname* { m i n } \left\{ d + \underset { j = 1 } { \overset { d } { \sum } } \log \left( 1 + \frac { \eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \tau ^ { 2 } } \right) \right. } \\ & { \qquad + \left. \frac { 4 \eta _ { l } ^ { 2 } K T } { m \tau ^ { 2 } } \underset { j = 1 } { \overset { d } { \sum } } \sigma _ { l , j } ^ { 2 } + 2 0 \eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \mathbb { E } \underset { j = 1 } { \overset { d } { \sum } } \frac { ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } { \tau ^ { 2 } } + \frac { 4 0 \eta _ { l } ^ { 4 } K ^ { 2 } L ^ { 2 } } { \tau ^ { 2 } } \underset { t = 0 } { \overset { T \mathrm { - } 1 } { \sum } } \mathbb { E } \left[ \left\| \nabla f ( x _ { t } ) \right\| \right. \right. } \end{array} +$$ + +Proof. We bound the desired quantity in the following manner: + +$$ +\begin{array} { r } { \stackrel { T \to 1 } { \sum } \displaystyle \sum _ { j = 0 } ^ { d } \frac { \Delta _ { t , j } ^ { 2 } } { \sum _ { l = 0 } ^ { t } \Delta _ { l , j } ^ { 2 } + \tau ^ { 2 } } \leq d + \mathbb { E } \displaystyle \sum _ { j = 1 } ^ { d } \log \left( 1 + \frac { \sum _ { l = 0 } ^ { T - 1 } \Delta _ { l , j } ^ { 2 } } { \tau ^ { 2 } } \right) \leq d + \displaystyle \sum _ { j = 1 } ^ { d } \log \left( 1 + \frac { \eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \tau ^ { 2 } } \right) . } \end{array} +$$ + +An alternate way of the bounding this quantity is as follows: + +$$ +\begin{array} { r l } & { \mathbb { E } \displaystyle \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \frac { \Delta _ { t , j } ^ { 2 } } { \sum _ { l = 0 } ^ { t } \Delta _ { t , j } ^ { 2 } + \tau ^ { 2 } } \leq \mathbb { E } \displaystyle \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \frac { \Delta _ { t , j } ^ { 2 } } { \tau ^ { 2 } } } \\ & { \leq \mathbb { E } \displaystyle \sum _ { t = 0 } ^ { T - 1 } \left\| \frac { \Delta _ { t } + \eta _ { l } K \nabla f ( x _ { t } ) - \eta _ { l } K \nabla f ( x _ { t } ) } { \tau } \right\| ^ { 2 } } \\ & { \leq 2 \mathbb { E } \displaystyle \sum _ { t = 0 } ^ { T - 1 } \left[ \left\| \frac { \Delta _ { t } + \eta _ { l } K \nabla f ( x _ { t } ) } { \tau } \right\| ^ { 2 } + \eta _ { l } ^ { 2 } K ^ { 2 } \left\| \frac { \nabla f ( x _ { t } ) } { \tau } \right\| ^ { 2 } \right] . } \end{array} +$$ + +The first quantity in the above bound can be further bounded as follows: + +$$ +\begin{array} { r l } & { \mathbb { E } _ { \frac { d } { d } } ^ { \lambda - 1 } \Bigg [ \Bigg | \frac { 1 } { \tau } \Bigg ( - \frac { 1 } { m } \sum _ { i = 1 } ^ { \infty } - \lambda _ { i } ^ { \lambda } \sum _ { j = 1 } ^ { \infty } - \lambda _ { i } ^ { \lambda } \nabla F _ { i } ( \xi _ { i } , \xi _ { i } ) \Bigg ) \Bigg | ^ { 2 } } \\ & { - 2 \mathbb { E } _ { \frac { d } { d } } ^ { \lambda } \Bigg [ \Bigg | \frac { 1 } { \tau } \Bigg ( \frac { 1 } { m } \sum _ { i = 1 } ^ { \infty } \sum _ { j = 1 } ^ { \infty } \lambda _ { i } ^ { \lambda } \sum _ { j = 1 } ^ { \lambda } - \lambda _ { i } ^ { \lambda } \nabla F _ { i } ( \xi _ { i } , \xi _ { i } ) + 2 \mathbb { E } _ { \frac { d } { d } } ^ { \lambda } \lambda _ { i } ^ { \lambda } \xi _ { i } ^ { \lambda } \xi _ { i } ^ { \lambda } \xi _ { j } - 3 \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \Bigg ) - } \\ & - 2 \mathbb { E } _ { \frac { d } { d } } ^ { \lambda } \mathbb { E } ^ { \lambda - 1 } \Bigg [ \Bigg | \frac { 1 } { \tau } \Bigg ( \frac { 1 } { m } \sum _ { i = 1 } ^ { \infty } \lambda _ { i } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb { E } ^ { \lambda } \mathbb E \end{array} +$$ + +Here, the first inequality follows from simple application of the fact that $a b \le ( a ^ { 2 } + b ^ { 2 } ) / 2$ . The result follows. □ + +Substituting the above bound in Equation (10), we obtain: + +$$ +\begin{array} { r l } & { \frac { \displaystyle { n K } } { \displaystyle { 4 } } \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \mathbb { E } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \pi _ { t - 1 , j } } + \tau } } \\ & { \leq \frac { f \left( x _ { 0 } \right) - \mathbb { E } \left[ f ( x _ { T } ) \right] } { \eta } + \frac { 5 \eta _ { t } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \tau } \mathbb { E } \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } \\ & { \qquad + \displaystyle { \sum _ { j = 1 } ^ { d } \left( \frac { \eta _ { t } K G ^ { 2 } } { \tau } + \frac { \eta L } { 2 } \right) \times \left[ \operatorname* { m i n } \left\{ d + d \log \left( 1 + \frac { \eta _ { t } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \tau ^ { 2 } } \right) + \right. \right.} } \\ & { \quad \quad \left. \frac { 4 \eta _ { t } ^ { 2 } K T } { m \tau ^ { 2 } } \sum _ { j = 1 } ^ { d } \sigma _ { l , j } ^ { 2 } + 2 0 \eta _ { t } ^ { 4 } K ^ { 3 } L ^ { 2 } T \sum _ { j = 1 } ^ { d } \frac { ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } { \tau ^ { 2 } } + \frac { 4 0 \eta _ { t } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \tau ^ { 2 } } \sum _ { t = 0 } ^ { T - 1 } \mathbb { E } \left[ \left\| \nabla f ( x _ { t } ) \right\| ^ { 2 } \right] \right\} } \end{array} +$$ + +We observe the following: + +$$ +\sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla \mathbb { E } f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { v _ { t - 1 , j } } + \tau } \geq \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \mathbb { E } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \eta _ { t } K G \sqrt { T } + \tau } \geq \frac { T } { \eta _ { t } K G \sqrt { T } + \tau } \operatorname* { m i n } _ { 0 \leq t \leq T } \mathbb { E } \| \nabla f ( x _ { t } ) \| ^ { 2 } . +$$ + +The second part of Theorem 1 follows from using the above inequality in Equation (12). Note that the first part of Theorem 1 is obtain from the part of Lemma 4. □ + +# A.2.1 LIMITED PARTICIPATION + +For limited participation, the main changed in the proof is in Equation (11). The rest of the proof is similar so we mainly focus on Equation (11) here. Let $s$ be the sampled set at the $t ^ { \mathrm { { t h } } }$ iteration such that $| S | = s$ . In partial participation, we assume that the set $s$ is sampled uniformly from all subsets of $[ m ]$ with size $s$ . In this case, for the first term in Equation (11), we have + +$$ +\begin{array} { r l } { \sum _ { j = 1 } ^ { N } \Bigg [ \Bigg | \int _ { - 1 } ^ { 1 } \Bigg ( \int _ { 0 } ^ { 1 } \sum _ { j = 0 } ^ { 1 } \sum _ { i = 0 } ^ { N } \sum _ { j = 1 } ^ { N } \Bigg ) \Bigg | ^ { 2 } \Bigg ] ^ { 2 } } & { \mathrm { d } \Gamma } \\ { = } & { \mathcal { D } _ { j } ^ { \mathrm { R } } \Bigg [ \Bigg | \int _ { - 1 } ^ { 1 } \Bigg ( \int _ { 0 } ^ { 1 } \sum _ { j = 0 } ^ { 1 } \sum _ { i = 0 } ^ { N } \sum _ { j = 1 } ^ { N } \delta ( \sum _ { i = 0 } ^ { N } \epsilon _ { j i } ) + \mathrm { d } \Omega _ { i } ^ { \mathrm { R } } \bigg ( \epsilon _ { j i } \Big ) - \mathrm { d } \Omega _ { j } ^ { \mathrm { R } } \bigg ( \epsilon _ { j i } \Big ) - \mathrm { d } \Omega _ { i } ^ { \mathrm { R } } \bigg ( \epsilon _ { j i } \Big ) - \mathrm { d } \Omega _ { j } ^ { \mathrm { R } } \bigg ) \Bigg | ^ { 2 } } \\ & { \leq \epsilon _ { j } ^ { \mathrm { R } } \Bigg ( \int _ { - 1 } ^ { 1 } \epsilon _ { j i } ^ { \mathrm { R } } \bigg ) ^ { 2 } \Bigg [ \Bigg | \int _ { 0 } ^ { 1 } \sum _ { j = 1 } ^ { N } \delta ( \sum _ { j = 1 } ^ { N } \epsilon _ { j i } ) - \mathrm { d } \Omega _ { i } ^ { \mathrm { R } } \bigg ( \epsilon _ { j i } ^ { \mathrm { R } } \bigg ) - \mathrm { d } \Omega _ { j } ^ { \mathrm { R } } \Bigg | ^ { 2 } \Bigg ] ^ { 2 } + \Bigg | \int _ { - 1 } ^ { 1 } \mathrm { d } \Omega _ { i } ^ { \mathrm { R } } \bigg ( \epsilon _ { j i } ^ { \mathrm { R } } \bigg ( \epsilon _ { j i } ^ { \mathrm { R } } \bigg ) - \mathrm { v } _ { \mathrm { R } } ^ { \mathrm { R } } \bigg ( \epsilon _ { j i } ^ { \mathrm { R } } \bigg ) \Bigg | ^ { 2 } } \\ & \qquad + \mathrm { d } \Omega _ { j } ^ { \mathrm { R } } \Bigg | ^ { 2 } \Bigg ] + \Bigg | \int _ { 0 } ^ { 1 } \int _ { - 1 } ^ { 1 } \mathrm { d } \Omega _ { i } ^ { \mathrm { R } } \Bigg ( \int _ { 0 } ^ { 1 } \sum _ { j = 1 } ^ { N } \mathrm { d } \ \end{array} +$$ + +Note that the expectation here also includes $s$ . The first inequality is obtained from the fact that $( a + b + c ) ^ { 2 } \leq 3 { \dot { ( } } a ^ { 2 } + b ^ { 2 } + c ^ { 2 } )$ . The second inequality is obtained from the fact that set $s$ is uniformly + +sampled from all subsets of $[ m ]$ with size $s$ . The third and fourth inequalities are similar to the one used in proof of Theorem 1. Substituting the above bound in Equation (10) gives the desired convergence rate. + +# A.3 PROOF OF THEOREM 2 + +Proof of Theorem 2. The proof strategy is similar to that of FEDADAGRAD except that we need to handle the exponential moving average in FEDADAM. We note that the update of FEDADAM is the following + +$$ +x _ { t + 1 } = x _ { t } + \eta \frac { \Delta _ { t } } { \sqrt { v _ { t } } + \tau } , +$$ + +for all $i \in [ d ]$ . Using the $L$ -smooth nature of function $f$ and the above update rule, we have the following: + +$$ +f ( \boldsymbol { x } _ { t + 1 } ) \leq f ( \boldsymbol { x } _ { t } ) + \eta \left. \nabla f ( \boldsymbol { x } _ { t } ) , \frac { \Delta _ { t } } { \sqrt { v _ { t } } + \tau } \right. + \frac { \eta ^ { 2 } L } { 2 } \sum _ { i = 1 } ^ { d } \frac { \Delta _ { t , i } ^ { 2 } } { ( \sqrt { v _ { t , i } } + \tau ) ^ { 2 } } +$$ + +The second step follows simply from FEDADAM’s update. We take the expectation of $f ( x _ { t + 1 } )$ (over randomness at time step $t$ ) and rewrite the above inequality as: + +$$ +\begin{array} { r } { \bar { c } _ { t } [ f ( x _ { t + 1 } ) ] \leq f ( x _ { t } ) + \eta \left. \nabla f ( x _ { t } ) , \mathbb { E } _ { t } \left[ \frac { \Delta _ { t } } { \sqrt { v _ { t } } + \tau } - \frac { \Delta _ { t } } { \sqrt { \beta _ { 2 } v _ { t - 1 } } + \tau } + \frac { \Delta _ { t } } { \sqrt { \beta _ { 2 } v _ { t - 1 } } + \tau } \right] \right. + \frac { \eta ^ { 2 } L } { 2 } \displaystyle \sum _ { j = 1 } ^ { d } } \\ { = f ( x _ { t } ) + \eta \underbrace { \left. \nabla f ( x _ { t } ) , \mathbb { E } _ { t } \left[ \frac { \Delta _ { t } } { \sqrt { \beta _ { 2 } v _ { t - 1 } } + \tau } \right] \right. } _ { R _ { 1 } } + \eta \underbrace { \left. \nabla f ( x _ { t } ) , \mathbb { E } _ { t } \left[ \frac { \Delta _ { t } } { \sqrt { v _ { t } } + \tau } - \frac { \Delta _ { t } } { \sqrt { \beta _ { 2 } v _ { t - 1 } } } \right. \right. } _ { R _ { 2 } } } \\ { \left. \left. + \frac { \eta ^ { 2 } L } { 2 } \displaystyle \sum _ { j = 1 } ^ { d } \mathbb { E } _ { t } \left[ \frac { \Delta _ { t , j } ^ { 2 } } { ( \sqrt { v _ { t , j } } + \tau ) ^ { 2 } } \right] \right. \right. } \end{array} +$$ + +Bounding $R _ { 2 }$ . We observe the following about $R _ { 2 }$ + +$$ +\begin{array} { r l } & { H _ { 2 } = \mathbf { E } _ { \frac { 1 } { 2 } - 1 } ^ { + } \nabla f ( x ) \Big | _ { \mathcal { S } } \times \left[ \frac { \Delta _ { t , x } } { \sqrt { \nabla f _ { \mathbf { S } } \phi _ { t + 1 } } } - \frac { \Delta _ { t , x } } { \sqrt { \nabla f _ { \mathbf { S } } \phi _ { t - 1 , \Delta } } + \tau } \right] } \\ & { \quad = \mathbf { E } _ { \frac { 1 } { 2 } - 1 } ^ { + } \nabla f ( x ) \Big | _ { \mathcal { S } } \Delta _ { t , x } \simeq \Delta _ { t , \Phi } \times \left[ \frac { \sqrt { \omega _ { 2 } \phi _ { 1 } } \mathbf { e } _ { 1 } - \sqrt { \nabla f _ { \mathbf { S } } \mathbf { e } _ { 1 } } } { \left( \sqrt { \omega _ { 2 } } \omega _ { 1 } \right) \tau \left( \sqrt { \hat { \omega } _ { 2 } \omega _ { 1 } } - \Delta _ { t , 1 } \right) \tau } \right] } \\ & { \quad \quad - \mathbf { E } _ { \frac { 1 } { 2 } - 1 } ^ { + } \nabla f ( x ) \Big | _ { \mathcal { S } } \times \Delta _ { t , \Phi } \times \left[ \frac { - ( 1 - \omega _ { 2 } ) \Delta _ { t , x } ^ { 2 } } { \left( \sqrt { \omega _ { 2 } } \omega _ { 1 } + \gamma \right)} \left( \sqrt { \hat { \omega } _ { 2 } \omega _ { 1 } } - \sqrt { \nabla f _ { \mathbf { S } } \mathbf { e } _ { 1 } - \sqrt { \nabla f _ { \mathbf { S } } \mathbf { e } _ { 1 } } } \right) + \sqrt { \mathbf { e } _ { 1 } \omega _ { 2 } } \right] } \\ & { \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \Delta _ { t } ^ { 2 } } \\ & { \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \Delta _ { t } ^ { 2 } } \\ & { \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \Delta _ { t } ^ { 2 } } \\ & \quad \quad \quad \quad \quad \quad \end{array} +$$ + +Bounding $R _ { 1 }$ . The term $R _ { 1 }$ can be bounded as follows: + +$$ +\begin{array} { r l } & { R _ { 1 } = \Bigg \langle \nabla f ( x _ { t } ) , \mathbb { E } _ { t } \left[ \frac { \Delta _ { t } } { \sqrt { \beta _ { 2 } v _ { t - 1 } } + \tau } \right] \Bigg \rangle } \\ & { \quad = \Bigg \langle \frac { \nabla f ( x _ { t } ) } { \sqrt { \beta _ { 2 } v _ { t - 1 } } + \tau } , \mathbb { E } _ { t } \left[ \Delta _ { t } - \eta _ { t } K \nabla f ( x _ { t } ) + \eta _ { t } K \nabla f ( x _ { t } ) \right] \Bigg \rangle } \\ & { \quad = - \eta _ { t } K \displaystyle \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } v _ { t - 1 , j } } + \tau } + \underbrace { \Bigg \langle \frac { \nabla f ( x _ { t } ) } { \sqrt { \beta _ { 2 } v _ { t - 1 } } + \tau } , \mathbb { E } _ { t } \left[ \Delta _ { t } + \eta _ { t } K \nabla f ( x _ { t } ) \right] \Bigg \rangle } _ { R _ { 3 } } . } \end{array} +$$ + +Bounding $R _ { 3 }$ . The term $R _ { 3 }$ can be bounded in exactly the same way as term $T _ { 3 }$ in proof of Theorem 1: + +$$ +R _ { 3 } \leq \frac { \eta _ { l } K } { 2 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } v _ { t - 1 , j } } + \tau } + \frac { \eta _ { l } L ^ { 2 } } { 2 m \tau } \mathbb { E } _ { t } \left[ \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \| x _ { i , k } ^ { t } - x _ { t } \| ^ { 2 } \right] +$$ + +Substituting the above inequality in Equation (15), we get + +$$ +R _ { 1 } \leq - \frac { \eta _ { l } K } { 2 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } v _ { t - 1 , j } } + \tau } + \frac { \eta _ { l } L ^ { 2 } } { 2 m \tau } \mathbb { E } _ { t } \left[ \sum _ { i = 1 } ^ { m } \sum _ { k = 0 } ^ { K - 1 } \| x _ { i , k } ^ { t } - x _ { t } \| ^ { 2 } \right] +$$ + +Here we used the fact that $\sqrt { v _ { t - 1 , j } } \leq \eta _ { l } K G$ and conditions in Theorem 2. Using Lemma 3, we obtain the following bound on $R _ { 1 }$ : + +$$ +R _ { 1 } \leq - \frac { \eta _ { l } K } { 4 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } v _ { t - 1 , j } } + \tau } + \frac { 5 \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \tau } \mathbb { E } _ { t } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) +$$ + +Putting pieces together. Substituting bounds $R _ { 1 }$ and $R _ { 2 }$ in Equation (14), we have + +$$ +\begin{array} { r l } { { \mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \leq f ( x _ { t } ) - \frac { \eta \eta _ { l } K } { 4 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } } v _ { t - 1 , j } } + \tau + \frac { 5 \eta \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \tau } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } } \\ & { + ( \frac { \eta \sqrt { 1 - \beta _ { 2 } } G } { \tau } ) \sum _ { j = 1 } ^ { d } \mathbb { E } _ { t } [ \frac { \Delta _ { t , j } ^ { 2 } } { \sqrt { v _ { t , j } } + \tau ) } ] + ( \frac { \eta ^ { 2 } L } { 2 } ) \sum _ { j = 1 } ^ { d } \mathbb { E } _ { t } [ \frac { \Delta _ { t , j } ^ { 2 } } { v _ { t , j } + \tau ^ { 2 } ) } ] } \end{array} +$$ + +Summing over $t = 0$ to $T - 1$ and using telescoping sum, we have + +$$ +\mathbb { E } [ f ( x _ { T } ) ] \leq f ( x _ { 0 } ) - \frac { \eta \eta _ { l } K } { 4 } \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \mathbb { E } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } v _ { t - 1 , j } } + \tau } + \frac { 5 \eta \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \tau } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) +$$ + +To bound this term further, we need the following result. + +Lemma 5. The following upper bound holds for Algorithm 2 (FEDADAM): + +$$ +\sum _ { = 0 } ^ { { { \nu - 1 } } } \sum _ { j = 1 } ^ { d } \mathbb { E } \left[ \frac { \Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } \right] \leq \frac { 4 \eta _ { l } ^ { 2 } K T } { m \tau ^ { 2 } } \sum _ { j = 1 } ^ { d } \sigma _ { l , j } ^ { 2 } + \frac { 2 0 \eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \tau ^ { 2 } } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) + \frac { 4 0 \eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \tau ^ { 2 } } \sum _ { t = 1 } ^ { T - 1 } \frac { \Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } \sum _ { l = 1 } ^ { d } \frac { \Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } \sum _ { \ell = 1 } ^ { d } \frac { \Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } \sum _ { j = 1 } ^ { d } \frac { \Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } \sum _ { l = 1 } ^ { d } \frac { \Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } \sum _ { j = 1 } ^ { d } \frac { \Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } \sum _ { l = 1 } ^ { d } \frac { \Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } \sum _ { l = 1 } ^ { d } \frac { \Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } +$$ + +Proof. + +$$ +\mathbb { E } \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \frac { \Delta _ { t , j } ^ { 2 } } { \left( 1 - \beta _ { 2 } \right) \sum _ { l = 0 } ^ { t } \beta _ { 2 } ^ { t - l } \Delta _ { t , j } ^ { 2 } + \tau ^ { 2 } } \le \mathbb { E } \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \frac { \Delta _ { t , j } ^ { 2 } } { \tau ^ { 2 } } +$$ + +The rest of the proof follows along the lines of proof of Lemma 4. Using the same argument, we get + +$$ +\begin{array} { r l r } { { \sum _ { j = 1 } ^ { d } \mathbb { E } [ \frac { \Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \tau ^ { 2 } ) } ] \leq \frac { 4 \eta _ { l } ^ { 2 } K T } { m \tau ^ { 2 } } \sum _ { j = 1 } ^ { d } \sigma _ { l , j } ^ { 2 } + \frac { 2 0 \eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \tau ^ { 2 } } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } } \\ & { } & { + \frac { 4 0 \eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \tau ^ { 2 } } \sum _ { t = 0 } ^ { T - 1 } \mathbb { E } [ \nabla f ( x _ { t } ) ^ { 2 } ] , } \end{array} +$$ + +which is the desired result. + +Substituting the bound obtained from above lemma in Equation (17) and using a similar argument for bounding + +$$ +\left( \frac { \eta \sqrt { 1 - \beta _ { 2 } } G } { \tau } \right) \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \mathbb { E } \left[ \frac { \Delta _ { t , j } ^ { 2 } } { \sqrt { v _ { t , j } } + \tau ) } \right] , +$$ + +we obtain + +$$ +\begin{array} { c } { { \displaystyle z _ { t } [ f ( x _ { T } ) ] \leq f ( x _ { 0 } ) - \frac { \eta \eta _ { l } K } { 8 } \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } v _ { t - 1 , j } } + \tau } + \frac { 5 \eta \eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \tau } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) } } \\ { { + \left( \eta \sqrt { 1 - \beta _ { 2 } } G + \frac { \eta ^ { 2 } L } { 2 } \right) \times \left[ \frac { 4 \eta _ { l } ^ { 2 } K T } { m \tau ^ { 2 } } \sum _ { j = 1 } ^ { d } \sigma _ { l , j } ^ { 2 } + \frac { 2 0 \eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } T } { \tau ^ { 2 } } \mathbb { E } \sum _ { j = 1 } ^ { d } ( \sigma _ { l , j } ^ { 2 } + 6 K \sigma _ { g , j } ^ { 2 } ) \right] \sigma _ { g , l } ^ { 2 } } } \end{array} +$$ + +The above inequality is obtained due to the fact that: + +$$ +\left( \sqrt { 1 - \beta _ { 2 } } G + \frac { \eta L } { 2 } \right) \frac { 4 0 \eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \tau ^ { 2 } } \leq \frac { \eta _ { l } K } { 1 6 } \left( \frac { 1 } { \sqrt { \beta _ { 2 } } \eta _ { l } K G } + \frac { 1 } { \tau } \right) . +$$ + +The above condition follows from the condition on $\eta _ { l }$ in Theorem 2. We also observe the following: + +$$ +\sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } v _ { t - 1 , j } } + \tau } \geq \sum _ { t = 0 } ^ { T - 1 } \sum _ { j = 1 } ^ { d } \frac { [ \nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \sqrt { \beta _ { 2 } } \eta _ { l } K G + \tau } \geq \frac { T } { \sqrt { \beta _ { 2 } } \eta _ { l } K G + \tau } \operatorname* { m i n } _ { \eta \leq t \leq T } \Vert \nabla f ( x _ { t } ) \Vert ^ { 2 } . +$$ + +Substituting this bound in the above equation yields the desired result. + +# A.4 AUXILIARY LEMMATTA + +Lemma 6. For random variables $z _ { 1 } , \ldots , z _ { r }$ , we have + +$$ +\begin{array} { r } { \mathbb { E } \left[ \| z _ { 1 } + . . . + z _ { r } \| ^ { 2 } \right] \leq r \mathbb { E } \left[ \| z _ { 1 } \| ^ { 2 } + . . . + \| z _ { r } \| ^ { 2 } \right] . } \end{array} +$$ + +Lemma 7. For independent, mean $O$ random variables $z _ { 1 } , \ldots , z _ { r }$ , we have + +$$ +\begin{array} { r } { \mathbb { E } \left[ \| z _ { 1 } + . . . + z _ { r } \| ^ { 2 } \right] = \mathbb { E } \left[ \| z _ { 1 } \| ^ { 2 } + . . . + \| z _ { r } \| ^ { 2 } \right] . } \end{array} +$$ + +# B FEDERATED ALGORITHMS: IMPLEMENTATIONS AND PRACTICAL CONSIDERATIONS + +# B.1 FEDAVG AND FEDOPT + +In Algorithm 3, we give a simplified version of the FEDAVG algorithm by McMahan et al. (2017), that applies to the setup given in Section 2. We write $\mathbf { S } \mathbf { G } \mathbf { D } _ { K } ( x _ { t } , \eta _ { l } , f _ { i } )$ to denote $K$ steps of SGD using gradients $\nabla f _ { i } ( x , \bar { z } )$ for $z \sim D _ { i }$ with (local) learning rate $\eta _ { l }$ , starting from $x _ { t }$ . As noted in Section 2, Algorithm 3 is the special case of Algorithm 1 where CLIENTOPT is SGD, and SERVEROPT is SGD with learning rate 1. + +While Algorithms 1, 2, and 3 are useful for understanding relations between federated optimization methods, we are also interested in practical versions of these algorithms. In particular, Algorithms 1, 2, and 3 are stated in terms of a kind of ‘gradient oracle’, where we compute unbiased estimates + +# Algorithm 3 Simplified FEDAVG + +
Input: xo fort=0,.,T-1do
Samplea subsetSofclients
xt=xt
for each client i ∈ Sinparallel do
xt = SGDk(xt, nt, fi) for i∈ S (in parallel)
xt+1=∑iesxt
+ +of the client’s gradient. In practical scenarios, we often only have access to finite data samples, the number of which may vary between clients. + +Instead, we assume that in (1), each client distribution $\mathcal { D } _ { i }$ is the uniform distribution over some finite set $D _ { i }$ of size $n _ { i }$ . The $n _ { i }$ may vary significantly between clients, requiring extra care when implementing federated optimization methods. We assume the set $D _ { i }$ is partitioned into a collection of batches $B _ { i }$ , each of size $B$ . For $b \in B _ { i }$ , we let $f _ { i } ( x ; b )$ denote the average loss on this batch at $x$ with corresponding gradient $\nabla f _ { i } ( x ; b )$ . Thus, if $b$ is sampled uniformly at random from $B _ { i }$ , $\nabla f _ { i } ( x ; b )$ is an unbiased estimate of $\nabla F _ { i } ( x )$ . + +When training, instead of uniformly using $K$ gradient steps, as in Algorithm 1, we will instead perform $E$ epochs of training over each client’s dataset. Additionally, we will take a weighted average of the client updates, where we weight according to the number of examples $n _ { i }$ in each client’s dataset. This leads to a batched data version of FEDOPT in Algorithm 4, and a batched data version of FEDADAGRAD, FEDADAM, and FEDYOGI given in Algorithm 5. + +# Algorithm 4 FEDOPT - Batched data + +Input: $x _ { 0 }$ , CLIENTOPT, SERVEROPT +for $t = 0 , \cdots , T - 1$ do Sample a subset $s$ of clients $x _ { i } ^ { t } = x _ { t }$ for each client $i \in S$ in parallel do for $e = 1 , \ldots , E$ do for $b \in B _ { i }$ do $\begin{array} { c } { \mathsf { \Gamma } _ { i } \gets \mathsf { \Gamma } _ { i } \gets } \\ { g _ { i } ^ { t } = \nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\ { x _ { i } ^ { t } = \mathrm { C } _ { \mathrm { L I E N T O P T } } ( x _ { i } ^ { t } , g _ { i } ^ { t } , \eta _ { l } , t ) } \\ { \Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\ { n = \sum _ { i \in \mathcal { S } } n _ { i } , \Delta _ { t } = \sum _ { i \in \mathcal { S } } \frac { n _ { i } } { n } \Delta _ { i } ^ { t } } \\ { x _ { t + 1 } = \mathrm { S E R V E R O P T } ( x _ { t } , - \Delta _ { t } , \eta , t ) } \end{array}$ + +In Section 5, we use Algorithm 4 when implementing FEDAVG and FEDAVGM. In particular, FEDAVG and FEDAVGM correspond to Algorithm 4 where CLIENTOPT and SERVEROPT are SGD. FEDAVG uses vanilla SGD on the server, while FEDAVGM uses SGD with a momentum parameter of 0.9. In both methods, we tune both client learning rate $\eta _ { l }$ and server learning rate $\eta$ . This means that the version of FEDAVG used in all experiments is strictly more general than that in (McMahan et al., 2017), which corresponds to FEDOPT where CLIENTOPT and SERVEROPT are SGD, and SERVEROPT has a learning rate of 1. + +We use Algorithm 5 for all implementations FEDADAGRAD, FEDADAM, and FEDYOGI in Section 5. For FEDADAGRAD, we set $\beta _ { 1 } = \beta _ { 2 } = 0$ (as typical versions of ADAGRAD do not use momentum). For FEDADAM and FEDYOGI we set $\beta _ { 1 } = { \bar { 0 . 9 } } , \beta _ { 2 } = 0 . 9 9$ . While these parameters are generally good choices (Zaheer et al., 2018), we emphasize that better results may be obtainable by tuning these parameters. + +# B.2 SCAFFOLD + +As discussed in Section 5, we compare all five optimizers above to SCAFFOLD (Karimireddy et al., 2019) on our various tasks. There are a few important notes about the validity of this comparison. + +# Algorithm 5 FEDADAGRAD FEDYOGI , and FEDADAM - Batched data + +Input: $x _ { 0 } , v _ { - 1 } \ge \tau ^ { 2 }$ , optional $\beta _ { 1 } , \beta _ { 2 } \in [ 0 , 1 )$ for FEDYOGI and FEDADAM +for $t = 0 , \cdots , T - 1$ do Sample a subset $s$ of clients $x _ { i } ^ { t } = x _ { t }$ for each client $i \in S$ in parallel do for $e = 1 , \ldots , E$ do for $b \in B _ { i }$ do $\begin{array} { r l } & { ~ x _ { i } ^ { t } = \check { x } _ { i } ^ { t } - \eta _ { l } \nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\ & { ~ \Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\ & { n = \sum _ { i \in S } n _ { i } , ~ \Delta _ { t } = \sum _ { i \in S } \frac { n _ { i } } { n } \Delta _ { i } ^ { t } } \\ & { \Delta _ { t } = \beta _ { 1 } \Delta _ { t - 1 } + ( 1 - \beta _ { 1 } ) \Delta _ { t } } \end{array}$ $v _ { t } = v _ { t - 1 } + \Delta _ { t } ^ { 2 }$ (FEDADAGRAD) $v _ { t } = v _ { t - 1 } - ( 1 - \beta _ { 2 } ) \Delta _ { t } ^ { 2 } \mathrm { s i g n } ( v _ { t - 1 } - \Delta _ { t } ^ { 2 } ) \left( \mathbf { F E D Y O G I } \right)$ vt = β2vt−1 + (1 − β2)∆2t (FEDADAM) xt+1 = xt + η √∆tvt+τ + +1. In cross-device settings, this is not a fair comparison. In particular, SCAFFOLD does not work in settings where clients cannot maintain state across rounds, as may be the case for federated learning systems on edge devices, such as cell phones. + +2. SCAFFOLD has two variants described by Karimireddy et al. (2019). In Option I, the control variate of a client is updated using a full gradient computation. This effectively requires performing an extra pass over each client’s dataset, as compared to Algorithm 1. In order to normalize the amount of client work, we instead use Option II, in which the clients’ control variates are updated using the difference between the server model and the client’s learned model. This requires the same amount of client work as FEDAVG and Algorithm 2. + +For practical reasons, we implement a version of SCAFFOLD mirroring Algorithm 4, in which we perform $E$ epochs over the client’s dataset, and perform weighted averaging of client models. For posterity, we give the full pseudo-code of the version of SCAFFOLD used in our experiments in Algorithm 6. This is a simple adaptiation of Option II of the SCAFFOLD algorithm in (Karimireddy et al., 2019) to the same setting as Algorithm 4. In particular, we let $n _ { i }$ denote the number of examples in client $i$ ’s local dataset. + +# Algorithm 6 SCAFFOLD SCAFFOLD, Option II - Batched data + +Input: $x _ { 0 } , c , \eta _ { l } , \eta$ +for $t = 0 , \cdots , T - 1 \cdot$ do Sample a subset $s$ of clients $x _ { i } ^ { t } = x _ { t }$ for each client $i \in S$ in parallel do for $e = 1 , \ldots , E$ do for $b \in B _ { i }$ do $\begin{array} { r l } & { \quad g _ { i } ^ { t } = \breve { \nabla } f _ { i } ( x _ { i } ^ { t } ; b ) } \\ & { \quad \quad x _ { i } ^ { t } = x _ { i } ^ { t } - \eta _ { l } ( g _ { i } ^ { t } - c _ { i } + c ) } \\ & { \quad c _ { i } ^ { + } = c _ { i } - c + ( E | \mathcal { B } _ { i } | \eta _ { l } ) ^ { - 1 } ( x _ { i } ^ { t } - x _ { i } ) } \\ & { \quad \Delta x _ { i } = x _ { i } ^ { t } - x _ { t } , \Delta c _ { i } = c _ { i } ^ { + } - c _ { i } } \\ & { \quad c _ { i } = c _ { i } ^ { + } } \\ & { \quad n = \sum _ { i \in \mathcal { S } } n _ { i } , \Delta x = \sum _ { i \in \mathcal { S } } \frac { n _ { i } } { n } \Delta x _ { i } , \Delta c = \sum _ { i \in \mathcal { S } } \frac { n _ { i } } { n } \Delta c _ { i } } \\ & { \quad x _ { t + 1 } = x _ { t } + \eta \Delta x , c = c + \frac { | \mathcal { S } | } { N } \Delta c } \end{array}$ + +In this algorithm, $c _ { i }$ is the control variate of client $i$ , and $c$ is the running average of these control variates. In practice, we must initialize the control variates $c _ { i }$ in some way when sampling a client $i$ for the first time. In our implementation, we set $c _ { i } = c$ the first time we sample a client $i$ . This has the advantage of exactly recovering FEDAVG when each client is sampled at most once. To initialize $c$ , we use the all zeros vector. + +We compare this version of SCAFFOLD to FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG on our tasks, tuning the learning rates in the same way (using the same grids as in Appendix D.2). In particular, $\eta _ { l } , \eta$ are tuned to obtain the best training performance over the last 100 communication rounds. We use the same federated hyperparameters for SCAFFOLD as discussed in Section 4. Namely, we set $E = 1$ , and sample 10 clients per round for all tasks except Stack Overflow NWP, where we sample 50. The results are given in Figure 1 in Section 5. + +# B.3 LOOKAHEAD, ADAALTER, AND CLIENT ADAPTIVITY + +The LOOKAHEAD optimizer (Zhang et al., 2019b) is primarily designed for non-FL settings. LOOKAHEAD uses a generic optimizer in the inner loop and updates its parameters using a “outer” learning rate. Thus, unlike FEDOPT, LOOKAHEAD uses a single generic optimizer and is thus conceptually different. In fact, LOOKAHEAD can be seen as a special case of FEDOPT in non-FL settings which uses a generic optimizer CLIENTOPT as a client optimizer, and SGD as the server optimizer. While there are multiple ways LOOKAHEAD could be generalized to a federated setting, one straightforward version would simply use an adaptive method as the CLIENTOPT. On the other hand, ADAALTER (Xie et al., 2019) is designed specifically for distributed settings. In ADAALTER, clients use a local optimizer similar to ADAGRAD (McMahan & Streeter, 2010a; Duchi et al., 2011) to perform multiple epochs of training on their local datasets. Both LOOKAHEAD and ADAALTER use client adaptivity, which is fundamentally different from the adaptive server optimizers proposed in Algorithm 2. + +To illustrate the differences, consider the client-to-server communication in ADAALTER. This requires communicating both the model weights and the client accumulators (used to perform the adaptive optimization, analogous to $v _ { t }$ in Algorithm 2). In the case of ADAALTER, the client accumulator is the same size as the model’s trainable weights. Thus, the client-to-server communication doubles for this method, relative to FEDAVG. In ADAALTER, the server averages the client accumulators and broadcasts the average to the next set of clients, who use this to initialize their adaptive optimizers. This means that the server-to-client communication also doubles relative to FEDAVG. The same would occur for any adaptive client optimizer in the distributed version of LOOKAHEAD described above. For similar reasons, we see that client adaptive methods also increase the amount of memory needed on the client (as they must store the current accumulator). By contrast, our adaptive server methods (Algorithm 2) do not require extra communication or client memory relative to FEDAVG. Thus, we see that server-side adaptive optimization benefits from lower per-round communication and client memory requirements, which are of paramount importance for FL applications (Bonawitz et al., 2019). + +# C DATASET & MODELS + +Here we provide detailed description of the datasets and models used in the paper. We use federated versions of vision datasets EMNIST (Cohen et al., 2017), CIFAR-10 (Krizhevsky & Hinton, 2009), and CIFAR-100 (Krizhevsky & Hinton, 2009), and language modeling datasets Shakespeare (McMahan et al., 2017) and StackOverflow (Authors, 2019). Statistics for the training datasets can be found in Table 2. We give descriptions of the datasets, models, and tasks below. Statistics on the number of clients and examples in both the training and test splits of the datasets are given in Table 2. + +Table 2: Dataset statistics. + +
DATASETTRAIN CLIENTSTRAIN EXAMPLESTEST CLIENTSTEST EXAMPLES
CIFAR-1050050,00010010,000
CIFAR-10050050,00010010,000
EMNIST-623,400671,5853,40077,483
SHAKESPEARE71516,0687152,356
STACKOVERFLOW342,477135,818,730204,08816,586,035
+ +# C.1 CIFAR-10/CIFAR-100 + +We create a federated version of CIFAR-10 by randomly partitioning the training data among 500 clients, with each client receiving 100 examples. We use the same approach as Hsu et al. (2019), where we apply latent Dirichlet allocation (LDA) over the labels of CIFAR-10 to create a federated dataset. Each client has an associated multinomial distribution over labels from which its examples are drawn. The multinomial is drawn from a symmetric Dirichlet distribution with parameter 0.1. + +For CIFAR-100, we perform a similar partitioning of 100 examples to 500 clients, but using a more sophisticated approach. We use a two step LDA process over the coarse and fine labels. We randomly partition the data to reflect the "coarse" and "fine" label structure of CIFAR-100 by using the Pachinko Allocation Method (PAM) (Li & McCallum, 2006). This creates more realistic client datasets, whose label distributions better resemble practical heterogeneous client datasets. We have made publicly available the specific federated version of CIFAR-100 we used for all experiments, though we avoid giving a link in this work in order to avoid de-anonymization. For complete details on how the dataset was created, see Appendix F. + +We train a modified ResNet-18 on both datasets, where the batch normalization layers are replaced by group normalization layers (Wu & He, 2018). We use two groups in each group normalization layer. As shown by Hsieh et al. (2019), group normalization can lead to significant gains in accuracy over batch normalization in federated settings. + +Preprocessing CIFAR-10 and CIFAR-100 consist of images with 3 channels of $3 2 \times 3 2$ pixels each. Each pixel is represented by an unsigned int8. As is standard with CIFAR datasets, we perform preprocessing on both training and test images. For training images, we perform a random crop to shape (24, 24, 3), followed by a random horizontal flip. For testing images, we centrally crop the image to $( 2 4 , 2 4 , 3 )$ . For both training and testing images, we then normalize the pixel values according to their mean and standard deviation. Namely, given an image $x$ , we compute $( x - \mu ) / \sigma$ where $\mu$ is the average of the pixel values in $x$ , and $\sigma$ is the standard deviation. + +# C.2 EMNIST + +EMNIST consists of images of digits and upper and lower case English characters, with 62 total classes. The federated version of EMNIST (Caldas et al., 2018) partitions the digits by their author. The dataset has natural heterogeneity stemming from the writing style of each person. We perform two distinct tasks on EMNIST, autoencoder training (EMNIST AE) and character recognition (EMNIST CR). For EMNIST AE, we train the “MNIST” autoencoder (Zaheer et al., 2018). This is a densely connected autoencoder with layers of size $( 2 8 \times 2 8 ) - 1 0 0 0 - 5 0 0 - 2 5 0 - 3 0$ and a symmetric decoder. A full description of the model is in Table 3. For EMNIST CR, we use a convolutional network. The network has two convolutional layers (with $3 \times 3$ kernels), max pooling, and dropout, followed by a 128 unit dense layer. A full description of the model is in Table 4. + +Table 3: EMNIST autoencoder model architecture. We use a sigmoid activation at all dense layers. + +
LayerOutput Shape#of Trainable Parameters
Input7840
Dense1000785000
Dense500500500
Dense250125250
Dense307530
Dense2507750
Dense500125500
Dense1000501000
Dense784784784
+ +Table 4: EMNIST character recognition model architecture. + +
LayerOutput Shape#of Trainable ParametersActivationHyperparameters
Input(28,28,1)0ReLUkernel size = 3; strides=(1,1)
Conv2d(26,26,32)320
Conv2d(24,24,64)18496
MaxPool2d(12,12,64)0
Dropout(12,12,64)0p= 0.25
Flatten92160
Dense1281179776
Dropout1280
Dense627998softmaxp= 0.5
+ +# C.3 SHAKESPEARE + +Shakespeare is a language modeling dataset built from the collective works of William Shakespeare. In this dataset, each client corresponds to a speaking role with at least two lines. The dataset consists of 715 clients. Each client’s lines are partitioned into training and test sets. Here, the task is to do next character prediction. We use an RNN that first takes a series of characters as input and embeds each of them into a learned 8-dimensional space. The embedded characters are then passed through 2 LSTM layers, each with 256 nodes, followed by a densely connected softmax output layer. We split the lines of each speaking role into into sequences of 80 characters, padding if necessary. We use a vocabulary size of 90; 86 for the characters contained in the Shakespeare dataset, and 4 extra characters for padding, out-of-vocabulary, beginning of line and end of line tokens. We train our model to take a sequence of 80 characters, and predict a sequence of 80 characters formed by shifting the input sequence by one (so that its last character is the new character we are actually trying to predict). Therefore, our output dimension is $8 0 \times 9 0$ . A full description of the model is in Table 5. + +Table 5: Shakespeare model architecture. + +
LayerOutput Shape#of Trainable Parameters
Input800
Embedding(80,8)720
LSTM(80,256)271360
LSTM(80,256)525312
Dense(80,90)23130
+ +# C.4 STACK OVERFLOW + +Stack Overflow is a language modeling dataset consisting of question and answers from the question and answer site, Stack Overflow. The questions and answers also have associated metadata, including tags. The dataset contains 342,477 unique users which we use as clients. We perform two tasks on this dataset: tag prediction via logistic regression (Stack Overflow LR, SO LR for short), and next-word prediction (Stack Overflow NWP, SO NWP for short). For both tasks, we restrict to the $1 0 , 0 0 0 { \mathrm { m o s t } }$ frequently used words. For Stack Overflow LR, we restrict to the 500 most frequent tags and adopt a one-versus-rest classification strategy, where each question/answer is represented as a bag-of-words vector (normalized to have sum 1). + +For Stack Overflow NWP, we restrict each client to the first 1000 sentences in their dataset (if they contain this many, otherwise we use the full dataset). We also perform padding and truncation to ensure that sentences have 20 words. We then represent the sentence as a sequence of indices corresponding to the 10,000 frequently used words, as well as indices representing padding, out-ofvocabulary words, beginning of sentence, and end of sentence. We perform next-word-prediction on these sequences using an RNN that embeds each word in a sentence into a learned 96-dimensional space. It then feeds the embedded words into a single LSTM layer of hidden dimension 670, followed by a densely connected softmax output layer. A full description of the model is in Table 6. The metric used in the main body is the top-1 accuracy over the proper 10,000-word vocabulary; it does not include padding, out-of-vocab, or beginning or end of sentence tokens when computing the accuracy.klkjlkj + +Table 6: Stack Overflow next word prediction model architecture. + +
LayerOutput Shape#of Trainable Parameters
Input200
Embedding(20,96)960384
LSTM(20,670)2055560
Dense(20,96)64416
Dense(20,10004)970388
+ +# D EXPERIMENT HYPERPARAMETERS + +# D.1 HYPERPARAMETER TUNING + +Throughout our experiments, we compare the performance of different instantiations of Algorithm 1 that use different server optimizers. We use SGD, SGD with momentum (denoted SGDM), ADAGRAD, ADAM, and YOGI. For the client optimizer, we use mini-batch SGD throughout. For all tasks, we tune the client learning rate $\eta _ { l }$ and server learning rate $\eta$ by using a large grid search. Full descriptions of the per-task server and client learning rate grids are given in Appendix D.2. + +We use the version of FEDADAGRAD, FEDADAM, and FEDYOGI in Algorithm 5. We let $\beta _ { 1 } = 0$ for FEDADAGRAD, and we let $\beta _ { 1 } = 0 . 9 , \beta _ { 2 } = 0 . 9 9$ for FEDADAM, and FEDYOGI. For FEDAVG and FEDAVGM, we use Algorithm 4, where CLIENTOPT, SERVEROPT are SGD. For FEDAVGM, the server SGD optimizer uses a momentum parameter of 0.9. For FEDADAGRAD, FEDADAM, and FEDYOGI, we tune the parameter $\tau$ in Algorithm 5. + +When tuning parameters, we select the best hyperparameters $( \eta _ { l } , \eta , \tau )$ based on the average training loss over the last 100 communication rounds. Note that at each round, we only see a fraction of the total users (10 for each task except Stack Overflow NWP, which uses 50). Thus, the training loss at a given round is a noisy estimate of the population-level training loss, which is why we averave over a window of communication rounds. + +# D.2 HYPERPARAMETER GRIDS + +Below, we give the client learning rate $\textcircled { \eta _ { l } }$ in Algorithm 1) and server learning rate ( $\dot { \eta }$ in Algorithm 1) grids used for each task. These grids were chosen based on an initial evaluation over the grids + +$$ +\eta _ { l } \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \dotsc , 1 0 ^ { 0 . 5 } \} +$$ + +$$ +\eta \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \dots , 1 0 ^ { 1 } \} +$$ + +These grids were then refined for Stack Overflow LR and EMNIST AE in an attempt to ensure that the best client/server learning rate combinations for each optimizer was contained in the interior of the learning rate grids. We generally found that these two tasks required searching larger learning rates than the other two tasks. The final grids were as follows: + +CIFAR-10: + +$$ +\begin{array} { r } { \eta _ { l } \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dots , 1 0 ^ { 0 . 5 } \} } \\ { \eta \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dots , 1 0 ^ { 1 } \} } \end{array} +$$ + +CIFAR-100: + +$$ +\begin{array} { r } { \eta _ { l } \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dotsc , 1 0 ^ { 0 . 5 } \} } \\ { \eta \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dotsc , 1 0 ^ { 1 } \} } \end{array} +$$ + +EMNIST AE: + +$$ +\begin{array} { r } { \eta _ { l } \in \{ 1 0 ^ { - 1 . 5 } , 1 0 ^ { - 1 } , \dots , 1 0 ^ { 2 } \} } \\ { \eta \in \{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \dots , 1 0 ^ { 1 } \} } \end{array} +$$ + +EMNIST CR: + +$$ +\begin{array} { r } { \eta _ { l } \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dotsc , 1 0 ^ { 0 . 5 } \} } \\ { \eta \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dotsc , 1 0 ^ { 1 } \} } \end{array} +$$ + +Shakespeare: + +$$ +\begin{array} { r } { \eta _ { l } \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dotsc , 1 0 ^ { 0 . 5 } \} } \\ { \eta \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dotsc , 1 0 ^ { 1 } \} } \end{array} +$$ + +# StackOverflow LR: + +$$ +\begin{array} { c } { { \eta _ { l } \in \{ 1 0 ^ { - 1 } , 1 0 ^ { - 0 . 5 } , \dots , 1 0 ^ { 3 } \} } } \\ { { \eta \in \{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \dots , 1 0 ^ { 1 . 5 } \} } } \end{array} +$$ + +# StackOverflow NWP: + +$$ +\begin{array} { r } { \eta _ { l } \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dotsc , 1 0 ^ { 0 . 5 } \} } \\ { \eta \in \{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \dotsc , 1 0 ^ { 1 } \} } \end{array} +$$ + +For all tasks, we tune $\tau$ over the grid: + +$$ +\tau \in \{ 1 0 ^ { - 5 } , \ldots , 1 0 ^ { - 1 } \} . +$$ + +# D.3 PER-TASK BATCH SIZES + +Given the large number of hyperparameters to tune, and to avoid conflating variables, we fix the batch size at a per-task level. When comparing centralized training to federated training in Section 5, we use the same batch size in both federated and centralized training. A full summary of the batch sizes is given in Table 7. + +Table 7: Client batch sizes used for each task. + +
TASKBATCH SIZE
CIFAR-1020
CIFAR-10020
EMNISTAE20
EMNISTCR20
SHAKESPEARE4
STACKOVERFLOWLR100
STACKOVERFLOW NWP16
+ +# D.4 BEST PERFORMING HYPERPARAMETERS + +In this section, we present, for each optimizer, the best client and server learning rates and values of $\tau$ found for the tasks discussed in Section 5. Specifically, these are the hyperparameters used in Figure Figure 1 and table Table 1. The validation metrics in Table 1 are obtained using the learning rates in Table 8 and the values of $\tau$ in Table 9. As discussed in Section 4, we choose $\eta , \eta _ { l }$ and $\tau$ to be the parameters that minimizer the average training loss over the last 100 communication rounds. + +Table 8: The base-10 logarithm of the client $( \eta _ { l } )$ and server $( \eta )$ learning rate combinations that achieve the accuracies from Table 1. See Appendix D.2 for a full description of the grids. + +
FEDADAGRADFEDADAMFEDYOGIFEDAVGMFEDAVG
nnmnmnmnmn
CIFAR-10-3/2-1-3/2-2-3/2-2-3/2-12-120
CIFAR-100-1-1-3/20-3/20-3/20-11
EMNISTAE3/2-3/21-3/21-3/2/010
EMNIST CR-3/2-1-3/2-5/2-3/2-5/2-3/2-1/2-10
SHAKESPEARE0-1/20-20-20-1/200
STACKOVERFLOWLR212-1/22-1/22020
STACKOVERFLOW NWP -1/2-3/2-1/2-2-1/2-2-1/20-1/20
+ +Table 9: The base-10 logarithm of the parameter $\tau$ (as defined in Algorithm 2) that achieve the validation metrics in Table 1. + +
FEDADAGRAD FEDADAM FEDYOGI
CIFAR-10-2-3-3
CIFAR-100-2-1-1
EMNISTAE-3-3-3
EMNIST CR-2-4-4
SHAKESPEARE-1-3-3
STACKOVERFLOWLR-2-5-5
STACKOVERFLOWNWP-4-5-5
+ +![](images/2cfd7a23d061067471c578b5c9cb5c0fbe27e8391814637353f1209c09dbb3a9.jpg) +Figure 4: Validation accuracy on EMNIST CR using constant learning rates $\eta , \eta _ { l }$ , and $\tau$ tuned to achieve the best training performance on the last 100 communication rounds; see Appendix D for hyperparameter grids. + +Table 10: Test set performance for Stack Overflow tasks after training: Accuracy for NWP and Recall $\textcircled { a } 5$ $( \times 1 0 0 )$ for LR. Performance within within $0 . 5 \%$ of the best result are shown in bold. + +
FED...ADAGRADADAMYOGIAvGMAvG
STACK OVERFLOW NWP24.425.725.724.520.5
STACK OVERFLOW LR66.865.266.546.540.6
+ +# E ADDITIONAL EXPERIMENTAL RESULTS + +# E.1 RESULTS ON EMNIST CR + +We plot the validation accuracy of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, FEDAVG, and SCAFFOLD on EMNIST CR. As in Figure 1, we tune the learning rates $\eta _ { l } , \eta$ and adaptivity $\tau$ by selecting the parameters obtaining the smallest training loss, averaged over the last 100 training rounds. The results are given in Figure 4. + +We see that all methods are roughly comparably throughout the duration of training. This reflects the fact that the dataset is quite simple, and most clients have all classes in their local datasets, reducing any heterogeneity among classes. Note that SCAFFOLD performs slightly worse than FEDAVG and all other methods here. As discussed in Section 5, this may be due to the presence of stale client control variates, and the communication-limited regime of our experiments. + +# E.2 STACK OVERFLOW TEST SET PERFORMANCE + +As discussed in Section 5, in order to compute a measure of performance for the Stack Overflow tasks as training progresses, we use a subsampled version of the test dataset, due to its prohibitively large number of clients and examples. In particular, at each round of training, we sample 10,000 random test samples, and use this as a measure of performance over time. However, once training is completed, we also evaluate on the full test dataset. For the Stack Overflow experiments described in Section 5, the final test accuracy is given in Table 10. + +# E.3 LEARNING RATE ROBUSTNESS + +In this section, we showcase what combinations of client learning rate $\eta _ { l }$ and server learning rate $\eta$ performed well for each optimizer and task. As in Figure 2, we plot, for each optimizer, task, and pair $( \eta _ { l } , \eta )$ , the validation set performance (averaged over the last 100 rounds). As in Section 5, we fix $\tau = 1 0 ^ { - 3 }$ throughout. The results, for the CIFAR-10, CIFAR-100, EMNIST AE, EMNIST CR, + +Shakespeare, Stack Overflow LR, and Stack Overflow NWP tasks are given in Figures 5, 6, 7, 8, 9, 10, and 11, respectively. + +While the general trends depend on the optimizer and task, we see that in many cases, the adaptive methods have rectangular regions that perform well. This implies a kind of robustness to fixing one of $\eta , \eta _ { l }$ , and varying the other. On the other hand, FEDAVGM and FEDAVG often have triangular regions that perform well, suggesting that $\eta$ and $\eta _ { l }$ should be tuned simultaneously. + +![](images/3c228a3a0a506f7fab5984fad48909c5f19d018f5100b7319cb0ea2a5cd5bbad.jpg) +Figure 5: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the CIFAR-10 task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\bar { \tau } = 1 0 ^ { - 3 }$ . + +![](images/988976289e57389c260e2ce5ff2bcfe18be6722e20e33a23e74e0c18d86179b5.jpg) +Figure 6: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the CIFAR-100 task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\tau = 1 0 ^ { - 3 }$ . + +![](images/a23451f198d579719bd5c12522f418c42bc52fa09bd6fee577a2a6cd9144cba7.jpg) +Figure 7: Validation MSE (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the EMNIST AE task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\tau = 1 0 ^ { - 3 }$ . + +![](images/c472049d458c0bdb90d3a9865546ea6d2a6905247c0626ad243509dc4925cb64.jpg) +Figure 8: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the EMNIST CR task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\tau = 1 0 ^ { - 3 }$ . + +![](images/110df1d101eebec91adcafb1b92bcfbd8eb49008a8ef880e5dd00d979aa15b32.jpg) +Figure 9: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Shakespeare task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\tau = 1 0 ^ { - 3 }$ . + +![](images/36a8c5f6ace9eb2d7ece4dab87f90356b80a7333fb4a7383edd182ee807213cb.jpg) +Figure 10: Validation recall $\textcircled { \alpha } 5$ (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Stack Overflow LR task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\tau = 1 0 ^ { - 3 }$ . + +![](images/4238061faf9efe9b6eef6f047f57581529cb3114358e9247fc74a3a2af45e1e3.jpg) +Figure 11: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Stack Overflow NWP task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\tau = 1 0 ^ { - 3 }$ . + +In order to better understand the results in Appendix E.3, we plot the relation between optimal choices of client and server learning rates. For each optimizer, task, and client learning rate $\eta _ { l }$ , we find the best corresponding server learning rate $\eta$ among the grids listed in Appendix D.2. Throughout, we fix $\tau = 1 0 ^ { - 3 }$ for the adaptive methods. We omit any points for which the final validation loss is within $10 \%$ of the worst-recorded validation loss over all hyperparameters. Essentially, we omit client learning rates that did not lead to any training of the model. The results are given in Figure 12. + +![](images/50c44eab15de90c5967e9a2de54c1c02f46cb58fea87d1a6bfb848613b707602.jpg) +Figure 12: The best server learning rate in our hyperparameter tuning grids for each client learning rate, optimizers, and task. We select the server learning rates based on the average validation performance over the last 100 communication rounds. For FEDADAGRAD, FEDADAM, and FEDYOGI, we fix $\tau = 1 0 ^ { - 3 }$ . We omit all client learning rates for which all server learning rates did not change the initial validation loss by more than $10 \%$ . + +In virtually all tasks, we see a clear inverse relationship between client learning rate $\eta _ { l }$ and server learning rate $\eta$ for FEDAVG and FEDAVGM. As discussed in Section 5, this supports the observation that for the non-adaptive methods, $\eta _ { l }$ and $\eta$ must in some sense be tuned simultaneously. On the other hand, for adaptive optimizers on most tasks we see much more stability in the best server learning rate $\eta$ as the client learning rate $\eta _ { l }$ varies. This supports our observation in Section 5 that for adaptive methods, tuning the client learning rate is more important. + +Notably, we see a clear exception to this on the Stack Overflow LR task, where there is a definitive inverse relationship between learning rates among all optimizers. The EMNIST AE task also displays somewhat different behavior. While there are still noticeable inverse relationships between learning rates for FEDAVG and FEDAVGM, the range of good client learning rates is relatively small. We emphasize that this task is fundamentally different than the remaining tasks. As noted by Zaheer et al. (2018), the primary obstacle in training bottleneck autoencoders is escaping saddle points, not in converging to critical points. Thus, we expect there to be qualitative differences between EMNIST AE and other tasks, even EMNIST CR (which uses the same dataset). + +# E.5 ROBUSTNESS OF THE ADAPTIVITY PARAMETER + +As discussed in Section 5, we plot, for each adaptive optimizer and task, the validation accuracy as a function of the adaptivity parameter $\tau$ . In particular, for each value of $\tau$ (which we vary over $\{ 1 0 ^ { - 5 } , \dotsc , 1 0 ^ { - 1 } \}$ , see Appendix D), we plot the best possible last-100-rounds validation set performance. Specifically, we plot the average validation performance over the last 100 rounds using the best a posteriori values of client and server learning rates. The results are given in Figure 13. + +![](images/88a3306ab9d22edfbff475be8101d1e8b35ad59803132c108b2b9ff96ab1bfdf.jpg) +Figure 13: Validation performance of FEDADAGRAD, FEDADAM, and FEDYOGI for varying $\tau$ on various tasks. The learning rates $\eta$ and $\eta _ { l }$ are tuned for each $\tau$ to achieve the best training performance on the last 100 communication rounds. + +# E.6 IMPROVING PERFORMANCE WITH LEARNING RATE DECAY + +Despite the success of adaptive methods, it is natural to ask if there is still more to be gained. To test this, we trained the EMNIST CR model in a centralized fashion on a shuffled version of the dataset. We trained for 100 epochs and used tuned learning rates for each (centralized) optimizer, achieving an accuracy of $8 8 \%$ (see Table 11, CENTRALIZED row), significantly above the best EMNIST CR results from Table 1. The theoretical results in Section 3 point to a partial explanation, as they only hold when the client learning rate is small or is decayed over time. + +To validate this, we ran the same hyperparameter grids on the federated EMNIST CR task, but using a client learning rate schedule. We use a “staircase” exponential decay schedule (EXPDECAY) where the client learning rate $\eta _ { l }$ is decreased by a factor of 0.1 every 500 rounds. This is analogous in some sense to standard staircase learning rate schedules in centralized optimization (Goyal et al., 2017). Table 11 gives the results. EXPDECAY improves the accuracy of all optimizers, and allows most to get close to the best centralized accuracy. While we do not close the gap with centralized optimization entirely, we suspect that further tuning of the amount and frequency of decay may lead to even better accuracies. However, this may also require performing significantly more communication rounds, as the theoretical results in Section 3 are primarily asymptotic. In communication-limited settings, the added benefit of learning rate decay seems to be modest. + +Table 11: (Top) Test accuracy $( \% )$ of a model trained centrally with various optimizers. (Bottom) Average test accuracy $( \% )$ over the last 100 rounds of various federated optimizers on the EMNIST CR task, using constant learning rates or the EXPDECAY schedule for $\eta _ { l }$ . Accuracies (for the federated tasks) within $\mathrm { { \bar { 0 } . 5 \% } }$ of the best result are shown in bold. + +
ADAGRAD CENTRALIZED 88.0ADAM 87.9YOGI 88.0SGDM 87.7SGD 87.7
FED... ADAGRADADAMYOGIAvGMAVG
CONSTANT ηt85.1 85.685.585.284.9
EXPDECAY85.3 86.286.285.885.2
+ +# F CREATING A FEDERATED CIFAR-100 + +Overview We use the Pachinko Allocation Method (PAM) (Li & McCallum, 2006) to create a federated CIFAR-100. PAM is a topic modeling method in which the correlations between individual words in a vocabulary are represented by a rooted directed acyclic graph (DAG) whose leaves are the vocabulary words. The interior nodes are topics with Dirichlet distributions over their child nodes. To generate a document, we sample multinomial distributions from each interior node’s Dirichlet distribution. To sample a word from the document, we begin at the root, and draw a child node its multinomial distribution, and continue doing this until we reach a leaf node. + +To partition CIFAR-100 across clients, we use the label structure of CIFAR-100. Each image in the dataset has a fine label (often referred to as its label) which is a member of a coarse label. For example, the fine label “seal” is a member of the coarse label “aquatic mammals”. There are 20 coarse labels in CIFAR-100, each with 5 fine labels. We represent this structure as a DAG $G$ , with a root whose children are the coarse labels. Each coarse label is an interior node whose child nodes are its fine labels. The root node has a symmetric Dirichlet distribution with parameter $\alpha$ over the coarse labels, and each coarse label has a symmetric Dirichlet distribution with parameter $\beta$ . + +We associate each client to a document. That is, we draw a multinomial distribution from the Dirichlet prior at the root $\left( \operatorname { D i r } ( \alpha ) \right)$ and a multinomial distribution from the Dirichlet prior at each coarse label $\bar { ( } \mathrm { D i r } ( \beta ) )$ . To create the client dataset, we draw leaf nodes from this DAG using Pachinko allocation, randomly sample an example with the given fine label, and assign it to the client’s dataset. We do this 100 times for each of 500 distinct training clients. + +While more complex than LDA, this approach creates more realistic heterogeneity among client datasets by creating correlations between label frequencies for fine labels within the same coarse label set. Intuitively, if a client’s dataset has many images of dolphins, they are likely to also have pictures of whales. By using a small $\alpha$ at the root, client datasets become more likely to focus on a few coarse labels. By using a larger $\beta$ for the coarse-to-fine label distributions, clients are more likely to have multiple fine labels from the same coarse label. + +One important note: Once we sample a fine label, we randomly select an element with that label without replacement. This ensures no two clients have overlapping examples. In more detail, suppose we have sample a fine label $y$ with coarse label $c$ for client $m$ , and there is only one remaining such example $( x , { \bar { c } } , y )$ . We assign $( x , c , y )$ to client $m$ ’s dataset, and remove the leaf node $y$ from the DAG $G$ . We also remove $y$ from the multinomial distribution $\theta _ { c }$ that client $m$ has associated to coarse label $c$ , which we refer to as renormalization with respect to $y$ (Algorithm 8). If $i$ has no remaining children after pruning node $j$ , we also remove node $i$ from $G$ and re-normalize the root multinomial $\theta _ { r }$ with respect to $c$ . For all subsequent clients, we draw multinomials from this pruned $G$ according to symmetric Dirichlet distributions with the same parameters as before, but with one fewer category. + +Notation and method Let $\mathcal { C }$ denote the set of coarse labels and $\mathcal { V }$ the set of fine labels, and let $s$ denote the CIFAR-100 dataset. This consists of examples $( x , c , y )$ where $x$ is an image vector, $c \in { \mathcal { C } }$ is a coarse label set, and $y \in \mathcal { V }$ is a fine label with $y \in c$ . For $c \in \mathcal { C } , y \in \mathcal { D }$ , we let $S _ { c }$ and $\mathcal { S } _ { y }$ denote the set of examples in $s$ with coarse label $c$ and fine label $y$ . For $v \in G$ we let $| G [ v ] |$ denote the set of children of $v$ in $G$ . For $\gamma \in \mathbb { R }$ , let $\operatorname { D i r } ( \gamma , k )$ denote the symmetric Dirichlet distribution with $k$ categories. + +# Algorithm 7 Creating a federated CIFAR-100 + +Input: $N , M \in \mathbb { Z } _ { > 0 } , \alpha , \beta \in \mathbb { R } _ { \ge 0 }$ +for $m = 1 , \cdots , M$ do Sample $\theta _ { r } \sim \operatorname { D i r } ( \alpha , | G [ r ] | )$ for $c \in { \mathcal { C } } \cap G [ r ]$ do Sample $\theta _ { c } \sim \operatorname { D i r } ( \beta , | G [ c ] | )$ $D _ { m } = \varnothing$ for $n = 1 , \cdots N$ do Sample c ∼ Multinomial $\left( \theta _ { r } \right)$ Sample y ∼ Multinomial $\left( \theta _ { c } \right)$ Select $( x , c , y ) \in S$ uniformly at random $D _ { m } = D _ { m } \cup \{ ( x , c , y ) \}$ $\mathcal { S } = \mathcal { S } \backslash \{ ( x , c , \overset { \cdot } { y } ) \}$ if $\begin{array} { r } { { \cal S } _ { y } = \emptyset } \end{array}$ then G = G\{y} θc = RENORMALIZE(θc, y) if Sc = ∅ then G = G\{c} θr = RENORMALIZE(θr, c) + +
Algorithm8RENORMALIZE
Initialization: θ= (p1,...,Pk),i∈[K]
a=∑k≠iPk
fork ∈[K],k≠ido
Pk=pk/a
Return0' = (pi,···P'-1,P'+1,···,PK)
+ +Let $M$ denote the number of clients, $N$ the number of examples per client, and $D _ { m }$ the dataset for client $m \in \{ 1 , \cdots , M \}$ . A full description of our method is given in Algorithm 7. For our experiments, we use $N = 1 0 0$ , $M = 5 0 0$ , $\alpha = 0 . 1 , \beta = 1 0$ . In Figure 14, we plot the distribution of unique labels among the 500 training clients. Each client has only a fraction of the overall labels in the distribution. Moreover, there is variance in the number of unique labels, with most clients having between 20 and 30, and some having over 40. Some client datasets have very few unique labels. While this is primarily an artifact of performing without replacement sampling, this helps increase the heterogeneity of the dataset in a way that can reflect practical concerns, as in many settings, clients may only have a few types of labels in their dataset. + +![](images/759e0a67cf8e490c2d7eb60abe974fcab397e55f7a8e0758467c8c94ae7999c1.jpg) +Figure 14: The distribution of the number of unique labels among training client datasets in our federated CIFAR-100 dataset. \ No newline at end of file diff --git a/parse/train/LkFG3lB13U5/LkFG3lB13U5_content_list.json b/parse/train/LkFG3lB13U5/LkFG3lB13U5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2b113158beecf84d38d2f158d647f8148d508974 --- /dev/null +++ b/parse/train/LkFG3lB13U5/LkFG3lB13U5_content_list.json @@ -0,0 +1,4380 @@ +[ + { + "type": "text", + "text": "ADAPTIVE FEDERATED OPTIMIZATION ", + "text_level": 1, + "bbox": [ + 176, + 99, + 638, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sashank J. Reddi∗, Zachary Charles∗, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Konecný, Sanjiv Kumar, H. Brendan McMahan ˇ ", + "bbox": [ + 186, + 143, + 772, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research \n{sashank, zachcharles, manzilzaheer, zachgarrett, krush, \nkonkey, sanjivk, mcmahan}@google.com ", + "bbox": [ + 186, + 174, + 732, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 250, + 544, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning is a distributed machine learning paradigm in which a large number of clients coordinate with a central server to learn a model without sharing their own training data. Standard federated optimization methods such as Federated Averaging (FEDAVG) are often difficult to tune and exhibit unfavorable convergence behavior. In non-federated settings, adaptive optimization methods have had notable success in combating such issues. In this work, we propose federated versions of adaptive optimizers, including ADAGRAD, ADAM, and YOGI, and analyze their convergence in the presence of heterogeneous data for general nonconvex settings. Our results highlight the interplay between client heterogeneity and communication efficiency. We also perform extensive experiments on these methods and show that the use of adaptive optimizers can significantly improve the performance of federated learning. ", + "bbox": [ + 233, + 281, + 766, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 473, + 336, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning (FL) is a machine learning paradigm in which multiple clients cooperate to learn a model under the orchestration of a central server (McMahan et al., 2017). In FL, raw client data is never shared with the server or other clients. This distinguishes FL from traditional distributed optimization, and requires contending with heterogeneous data. FL has two primary settings, crosssilo (eg. FL between large institutions) and cross-device (eg. FL across edge devices) (Kairouz et al., 2019, Table 1). In cross-silo FL, most clients participate in every round and can maintain state between rounds. In the more challenging cross-device FL, our primary focus, only a small fraction of clients participate in each round, and clients cannot maintain state across rounds. For a more in-depth discussion of FL and the challenges involved, we defer to Kairouz et al. (2019) and Li et al. (2019a). ", + "bbox": [ + 174, + 505, + 825, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Standard optimization methods, such as distributed SGD, are often unsuitable in FL and can incur high communication costs. To remedy this, many federated optimization methods use local client updates, in which clients update their models multiple times before communicating with the server. This can greatly reduce the amount of communication required to train a model. One such method is FEDAVG (McMahan et al., 2017), in which clients perform multiple epochs of SGD on their local datasets. The clients communicate their models to the server, which averages them to form a new global model. While FEDAVG has seen great success, recent works have highlighted its convergence issues in some settings (Karimireddy et al., 2019; Hsu et al., 2019). This is due to a variety of factors including (1) client drift (Karimireddy et al., 2019), where local client models move away from globally optimal models, and (2) a lack of adaptivity. FEDAVG is similar in spirit to SGD, and may be unsuitable for settings with heavy-tail stochastic gradient noise distributions, which often arise when training language models (Zhang et al., 2019a). Such settings benefit from adaptive learning rates, which incorporate knowledge of past iterations to perform more informed optimization. ", + "bbox": [ + 174, + 633, + 825, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we focus on the second issue and present a simple framework for incorporating adaptivity in FL. In particular, we propose a general optimization framework in which (1) clients perform multiple epochs of training using a client optimizer to minimize loss on their local data and (2) server updates its global model by applying a gradient-based server optimizer to the average of the clients’ model updates. We show that FEDAVG is the special case where SGD is used as both client and server optimizer and server learning rate is 1. This framework can also seamlessly incorporate adaptivity by using adaptive optimizers as client or server optimizers. Building upon this, we develop novel adaptive optimization techniques for FL by using per-coordinate methods as server optimizers. By focusing on adaptive server optimization, we enable use of adaptive learning rates without increase in client storage or communication costs, and ensure compatibility with cross-device FL. ", + "bbox": [ + 174, + 818, + 823, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 159 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main contributions In light of the above, we highlight the main contributions of the paper. ", + "bbox": [ + 178, + 165, + 779, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We study a general framework for federated optimization using server and client optimizers. This framework generalizes many existing federated optimization methods, including FEDAVG. • We use this framework to design novel, cross-device compatible, adaptive federated optimization methods, and provide convergence analysis in general nonconvex settings. To the best of our knowledge, these are the first methods for FL using adaptive server optimization. We show an important interplay between the number of local steps and the heterogeneity among clients. • We introduce comprehensive and reproducible empirical benchmarks for comparing federated optimization methods. These benchmarks consist of seven diverse and representative FL tasks involving both image and text data, with varying amounts of heterogeneity and numbers of clients. • We demonstrate strong empirical performance of our adaptive optimizers throughout, improving upon commonly used baselines. Our results show that our methods can be easier to tune, and highlight their utility in cross-device settings. ", + "bbox": [ + 178, + 190, + 826, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Related work FEDAVG was first introduced by McMahan et al. (2017), who showed it can dramatically reduce communication costs. Many variants have since been proposed to tackle issues such as convergence and client drift. Examples include adding a regularization term in the client objectives towards the broadcast model (Li et al., 2018), and server momentum (Hsu et al., 2019). When clients are homogeneous, FEDAVG reduces to local SGD (Zinkevich et al., 2010), which has been analyzed by many works (Stich, 2019; Yu et al., 2019; Wang & Joshi, 2018; Stich & Karimireddy, 2019; Basu et al., 2019). In order to analyze FEDAVG in heterogeneous settings, many works derive convergence rates depending on the amount of heterogeneity (Li et al., 2018; Wang et al., 2019; Khaled et al., 2019; Li et al., 2019b). Typically, the convergence rate of FEDAVG gets worse with client heterogeneity. By using control variates to reduce client drift, the SCAFFOLD method (Karimireddy et al., 2019) achieves convergence rates that are independent of the amount of heterogeneity. While effective in cross-silo FL, the method is incompatible with cross-device FL as it requires clients to maintain state across rounds. For more detailed comparisons, we defer to Kairouz et al. (2019). ", + "bbox": [ + 173, + 380, + 825, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adaptive methods have been the subject of significant theoretical and empirical study, in both convex (McMahan & Streeter, 2010b; Duchi et al., 2011; Kingma & Ba, 2015) and non-convex settings (Li & Orabona, 2018; Ward et al., 2018; Wu et al., 2019). Reddi et al. (2019); Zaheer et al. (2018) study convergence failures of ADAM in certain non-convex settings, and develop an adaptive optimizer, YOGI, designed to improve convergence. While most work on adaptive methods focuses on non-FL settings, Xie et al. (2019) propose ADAALTER, a method for FL using adaptive client optimization. Conceptually, our approach is also related to the LOOKAHEAD optimizer (Zhang et al., 2019b), which was designed for non-FL settings. Similar to ADAALTER, an adaptive FL variant of LOOKAHEAD entails adaptive client optimization (see Appendix B.3 for more details). We note that both ADAALTER and LOOKAHEAD are, in fact, special cases of our framework (see Algorithm 1) and the primary novelty of our work comes in focusing on adaptive server optimization. This allows us to avoid aggregating optimizer states across clients, making our methods require at most half as much communication and client memory usage per round (see Appendix B.3 for details). ", + "bbox": [ + 173, + 564, + 825, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notation For $a , b \\in \\mathbb { R } ^ { d }$ , we let ${ \\sqrt { a } } , a ^ { 2 }$ and $a / b$ denote the element-wise square root, square, and division of the vectors. For $\\theta _ { i } \\in \\mathbb { R } ^ { d }$ , we use both $\\theta _ { i , j }$ and $[ \\theta _ { i } ] _ { j }$ to denote its $j ^ { \\mathrm { t h } }$ coordinate. ", + "bbox": [ + 173, + 747, + 825, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 FEDERATED LEARNING AND FEDAVG ", + "text_level": 1, + "bbox": [ + 174, + 796, + 519, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In federated learning, we solve an optimization problem of the form: ", + "bbox": [ + 173, + 827, + 624, + 842 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/1e92002b69629c7c98895c3a52f42672bac76a3ac172efccb91bb0ec97aaca12.jpg", + "text": "$$\n\\operatorname* { m i n } _ { x \\in \\mathbb { R } ^ { d } } f ( x ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } F _ { i } ( x ) ,\n$$", + "text_format": "latex", + "bbox": [ + 408, + 847, + 588, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $F _ { i } ( x ) = \\mathbb { E } _ { z \\sim \\mathcal { D } _ { i } } [ f _ { i } ( x , z ) ]$ , is the loss function of the $i ^ { \\mathrm { { t h } } }$ client, $z \\in { \\mathcal { Z } }$ , and $\\mathcal { D } _ { i }$ is the data distribution for the $i ^ { \\mathrm { { t h } } }$ client. For $i \\neq j$ , $\\mathcal { D } _ { i }$ and $\\mathcal { D } _ { j }$ may be very different. The functions $F _ { i }$ (and therefore $f$ ) may be nonconvex. For each $i$ and $x$ , we assume access to an unbiased stochastic gradient $g _ { i } ( x )$ of the client’s true gradient $\\nabla F _ { i } ( x )$ . In addition, we make the following assumptions. ", + "bbox": [ + 174, + 893, + 825, + 925 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assumption 1 (Lipschitz Gradient). The function $F _ { i }$ is $L$ -smooth for all $i \\in [ m ]$ i.e., $\\parallel \\nabla F _ { i } ( x ) -$ $\\nabla F _ { i } ( y ) \\| \\leq L \\| x - y \\|$ , for all $x , y \\in \\mathbb { R } ^ { d }$ . ", + "bbox": [ + 173, + 136, + 823, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assumption 2 (Bounded Variance). The function $F _ { i }$ have $\\sigma _ { l }$ -bounded (local) variance i.e., $\\mathbb { E } [ \\| \\nabla [ \\widehat { f } _ { i } ( x , z ) ] _ { j } - [ \\nabla F _ { i } ( x ) ] _ { j } \\| ^ { 2 } ] = \\sigma _ { l , j } ^ { 2 }$ for all $x \\in \\mathbb { R } ^ { d }$ , $j \\in [ d ]$ and $i \\in [ m ]$ . Furthermore, we assume the (global) variance is bounded, $\\begin{array} { r } { \\left( 1 / m \\right) \\sum _ { i = 1 } ^ { m } \\| \\nabla [ F _ { i } ( x ) ] _ { j } - [ \\nabla f ( x ) ] _ { j } \\| ^ { 2 } \\leq \\sigma _ { g , j } ^ { 2 } } \\end{array}$ for all $\\boldsymbol { x } \\in \\mathbb { R } ^ { d }$ and $j \\in [ d ]$ . ", + "bbox": [ + 173, + 170, + 825, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assumption 3 (Bounded Gradients). The function $f _ { i } ( x , z )$ have $G$ -bounded gradients i.e., for any $i \\in [ m ]$ , $\\boldsymbol { x } \\in \\mathbb { R } ^ { d }$ and $z \\in { \\mathcal { Z } }$ we have $| [ \\nabla f _ { i } ( x , z ) ] _ { j } | \\leq G$ for all $j \\in [ d ]$ . ", + "bbox": [ + 174, + 236, + 823, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With a slight abuse of notation, we use $\\sigma _ { l } ^ { 2 }$ and $\\sigma _ { g } ^ { 2 }$ to denote $\\textstyle \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 }$ and $\\textstyle \\sum _ { j = 1 } ^ { d } \\sigma _ { g , j } ^ { 2 }$ . Assumptions 1 and 3 are fairly standard in nonconvex optimization literature (Reddi et al., 2016; Ward et al., 2018; Zaheer et al., 2018). We make no further assumptions regarding the similarity of clients datasets. Assumption 2 is a form of bounded variance, but between the client objective functions and the overall objective function. This assumption has been used throughout various works on federated optimization (Li et al., 2018; Wang et al., 2019). Intuitively, the parameter $\\sigma _ { g }$ quantifies similarity of client objective functions. Note $\\sigma _ { g } = 0$ corresponds to the i.i.d. setting. ", + "bbox": [ + 173, + 280, + 825, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A common approach to solving (1) in federated settings is FEDAVG (McMahan et al., 2017). At each round of FEDAVG, a subset of clients are selected (typically randomly) and the server broadcasts its global model to each client. In parallel, the clients run SGD on their own loss function, and send the resulting model to the server. The server then updates its global model as the average of these local models. See Algorithm 3 in the appendix for more details. ", + "bbox": [ + 173, + 386, + 825, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose that at round $t$ , the server has model $x _ { t }$ and samples a set $s$ of clients. Let $\\ v x _ { i } ^ { t }$ denote the model of each client $i \\in S$ after local training. We rewrite FEDAVG’s update as ", + "bbox": [ + 174, + 460, + 821, + 489 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/43e92881083d9a5561a6396b1be0718d263b1f1718f9e7b7d644b230f97efb05.jpg", + "text": "$$\nx _ { t + 1 } = { \\frac { 1 } { | S | } } \\sum _ { i \\in S } x _ { i } ^ { t } = x _ { t } - { \\frac { 1 } { | S | } } \\sum _ { i \\in S } \\left( x _ { t } - x _ { i } ^ { t } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 344, + 496, + 651, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\Delta _ { i } ^ { t } : = x _ { i } ^ { t } - x _ { t }$ and $\\begin{array} { r } { \\Delta _ { t } : = \\left( { 1 } / { | { S } | } \\right) \\sum _ { i \\in \\mathcal { S } } \\Delta _ { i } ^ { t } } \\end{array}$ . Then the server update in FEDAVG is equivalent to applying SGD to the “pseudo-gradient” $\\bar { - } \\Delta _ { t }$ with learning rate $\\eta = 1$ . This formulation makes it clear that other choices of $\\eta$ are possible. One could also utilize optimizers other than SGD on the clients, or use an alternative update rule on the server. This family of algorithms, which we refer to collectively as FEDOPT, is formalized in Algorithm 1. ", + "bbox": [ + 173, + 541, + 825, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 FEDOPT ", + "text_level": 1, + "bbox": [ + 173, + 627, + 323, + 642 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/2592fd4b1598c7bfe3873a636a0f2dff92e8d381323eec82ae75910138d9078b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
1: Input:xo,CLIENTOPT,SERVEROPT
2: fort=O,.,T-1do
3: Samplea subset Sofclients
4: xt,0=xt
5: for each client i ∈ S in parallel do
6: for k=0,.,K-1 do
Compute an unbiased estimate gt,k of VFi(xt,k) 7:
x,+1=CLEO(,,,t) 8:
9: △=x,K-xt
10: △t=∑ies△t 11: Xt+1 = SERVEROPT(xt,-△t,n,t)
", + "bbox": [ + 178, + 645, + 598, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Algorithm 1, CLIENTOPT and SERVEROPT are gradient-based optimizers with learning rates $\\eta _ { l }$ and $\\eta$ respectively. Intuitively, CLIENTOPT aims to minimize (1) based on each client’s local data while SERVEROPT optimizes from a global perspective. FEDOPT naturally allows the use of adaptive optimizers (eg. ADAM, YOGI, etc.), as well as techniques such as server-side momentum (leading to FEDAVGM, proposed by Hsu et al. (2019)). In its most general form, FEDOPT uses a CLIENTOPT whose updates can depend on globally aggregated statistics (e.g. server updates in the previous iterations). We also allow $\\eta$ and $\\eta _ { l }$ to depend on the round $t$ in order to encompass learning rate schedules. While we focus on specific adaptive optimizers in this work, we can in principle use any adaptive optimizer (e.g. AMSGRAD (Reddi et al., 2019), ADABOUND (Luo et al., 2019)). ", + "bbox": [ + 173, + 840, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 145 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While FEDOPT has intuitive benefits over FEDAVG, it also raises a fundamental question: Can the negative of the average model difference $\\Delta _ { t }$ be used as a pseudo-gradient in general server optimizer updates? In this paper, we provide an affirmative answer to this question by establishing a theoretical basis for FEDOPT. We will show that the use of the term SERVEROPT is justified, as we can guarantee convergence across a wide variety of server optimizers, including ADAGRAD, ADAM, and YOGI, thus developing principled adaptive optimizers for FL based on our framework. ", + "bbox": [ + 174, + 151, + 825, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 ADAPTIVE FEDERATED OPTIMIZATION ", + "text_level": 1, + "bbox": [ + 176, + 253, + 529, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we specialize FEDOPT to settings where SERVEROPT is an adaptive optimization method (one of ADAGRAD, YOGI or ADAM) and CLIENTOPT is SGD. By using adaptive methods (which generally require maintaining state) on the server and SGD on the clients, we ensure our methods have the same communication cost as FEDAVG and work in cross-device settings. ", + "bbox": [ + 174, + 285, + 825, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 2 provides pseudo-code for our methods. An alternate version using batched data and example-based weighting (as opposed to uniform weighting) of clients is given in Algorithm 5. The parameter $\\tau$ controls the algorithms’ degree of adaptivity, with smaller values of $\\tau$ representing higher degrees of adaptivity. Note that the server updates of our methods are invariant to fixed multiplicative changes to the client learning rate $\\eta _ { l }$ for appropriately chosen $\\tau$ , though as we shall see shortly, we will require $\\eta _ { l }$ to be sufficiently small in our analysis. ", + "bbox": [ + 173, + 347, + 825, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 2 FEDADAGRAD FEDYOGI , and FEDADAM ", + "text_level": 1, + "bbox": [ + 173, + 446, + 575, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: Initialization: $x _ { 0 } , v _ { - 1 } \\ge \\tau ^ { 2 }$ , decay parameters $\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )$ \n2: for $t = 0 , \\cdots , T - 1$ do \n3: Sample subset $s$ of clients \n4: $x _ { i , 0 } ^ { t } \\dot { = } x _ { t }$ \n5: for each client $i \\in S$ in parallel do \n6: for $k = 0 , \\cdots , K - \\bar { 1 }$ do \n7: Compute an unbiased estimate $g _ { i , k } ^ { t }$ of $\\nabla F _ { i } ( x _ { i , k } ^ { t } )$ \n8: $\\begin{array} { c } { x _ { i , k + 1 } ^ { t } = x _ { i , k } ^ { t } - \\eta _ { l } g _ { i , k } ^ { t } } \\\\ { \\Delta _ { i } ^ { t } = x _ { i , K } ^ { t } - x _ { t } } \\\\ { \\Delta _ { t } = \\beta _ { 1 } \\Delta _ { t - 1 } + \\left( 1 - \\beta _ { 1 } \\right) \\left( \\frac { 1 } { \\left| S \\right| } \\sum _ { i \\in S } \\Delta _ { i } ^ { t } \\right) } \\end{array}$ \n9: \n10: \n11: $v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }$ (FEDADAGRAD) \n12: $\\begin{array} { r l } & { v _ { t } = v _ { t - 1 } - ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\mathrm { s i g n } ( v _ { t - 1 } - \\Delta _ { t } ^ { 2 } ) \\ ( \\mathrm { F E D Y O G I } ) } \\\\ & { v _ { t } = \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\left( \\mathrm { F E D A D A M } \\right) } \\\\ & { x _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } } \\end{array}$ \n13: \n14: ", + "bbox": [ + 176, + 467, + 602, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide convergence analyses of these methods in general nonconvex settings, assuming full participation, i.e. $\\bar { \\mathcal { S } } = [ m ]$ . For expository purposes, we assume $\\beta _ { 1 } = 0$ , though our analysis can be directly extended to $\\beta _ { 1 } > 0$ . Our analysis can also be extended to partial participation (i.e. $| S | < m$ , see Appendix A.2.1 for details). Furthermore, non-uniform weighted averaging typically used in FEDAVG (McMahan et al., 2017) can also be incorporated into our analysis fairly easily. ", + "bbox": [ + 173, + 733, + 826, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Let Assumptions $I$ to 3 hold, and let $L , G , \\sigma _ { l } , \\sigma _ { g }$ be as defined therein. Let $\\sigma ^ { 2 } = $ $\\sigma _ { l } ^ { 2 } + 6 K \\sigma _ { g } ^ { 2 }$ . Consider the following conditions for $\\eta _ { l }$ : ", + "bbox": [ + 173, + 805, + 825, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(Condition I) ", + "bbox": [ + 173, + 854, + 263, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(Condition II) ", + "bbox": [ + 173, + 893, + 267, + 909 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/1a3347da564634f07515e78770ddadeb285d2b3ed5f780010df4a7f6a8388a15.jpg", + "text": "$$\n\\begin{array} { r l } & { \\eta _ { l } \\le \\displaystyle \\frac { 1 } { K } \\operatorname* { m i n } \\left\\{ \\frac { 1 } { 1 6 L } , \\frac { 1 } { T ^ { 1 / 6 } } \\left[ \\frac { \\tau } { 1 2 0 L ^ { 2 } G } \\right] ^ { 1 / 3 } \\right\\} , } \\\\ & { \\eta _ { l } \\le \\displaystyle \\frac { 1 } { 3 K } \\operatorname* { m i n } \\left\\{ \\frac { 1 } { T ^ { 1 / 1 0 } } \\left[ \\frac { \\tau ^ { 3 } } { L ^ { 2 } G ^ { 3 } } \\right] ^ { 1 / 5 } , \\frac { 1 } { T ^ { 1 / 8 } } \\left[ \\frac { \\tau ^ { 2 } } { L ^ { 3 } G \\eta } \\right] ^ { 1 / 4 } \\right\\} . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 349, + 843, + 738, + 922 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then the iterates of Algorithm 2 for FEDADAGRAD satisfy ", + "bbox": [ + 173, + 103, + 558, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/6d227a0445973c9ea45cc137c9b833b013e2b27e0a67359d5cea6897678e26c4.jpg", + "text": "$$\n\\begin{array} { r l r l } & { u n d e r ~ C o n d i t i o n ~ I o n l y , } & & { \\underset { 0 \\leq t \\leq T - 1 } { \\operatorname* { m i n } } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } \\leq \\odot \\left( \\left[ \\frac { G } { \\sqrt { T } } + \\frac { \\tau } { \\eta _ { t } K T } \\right] ( \\Psi + \\Psi _ { v a r } ) \\right) , } \\\\ & { u n d e r ~ b o t h ~ C o n d i t i o n ~ I \\ll { I I } , } & & { \\underset { 0 \\leq t \\leq T - 1 } { \\operatorname* { m i n } } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } \\leq \\odot \\left( \\left[ \\frac { G } { \\sqrt { T } } + \\frac { \\tau } { \\eta _ { t } K T } \\right] \\left( \\Psi + \\widetilde { \\Psi } _ { \\mathrm { v a r } } \\right) \\right) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 166, + 123, + 803, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we define ", + "bbox": [ + 173, + 199, + 277, + 214 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0a2a2d5da3167cf10c370390393db317fa6834be07374f6ba61f51fd8bdceaa2.jpg", + "text": "$$\n\\begin{array} { c } { { \\Psi = \\displaystyle \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\sigma ^ { 2 } , } } \\\\ { { \\Psi _ { \\mathrm { v a r } } = \\displaystyle \\frac { d ( \\eta _ { l } K G ^ { 2 } + \\tau \\eta L ) } { \\tau } \\Bigg [ 1 + \\log { \\displaystyle \\frac { \\tau ^ { 2 } + \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } } \\Bigg ] , } } \\\\ { { \\widetilde { \\Psi } _ { \\mathrm { v a r } } = \\displaystyle \\frac { 2 \\eta _ { l } K G ^ { 2 } + \\tau \\eta L } { \\tau ^ { 2 } } \\left[ \\displaystyle \\frac { 2 \\eta _ { l } ^ { 2 } K T } { m } \\sigma _ { l } ^ { 2 } + 1 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\sigma ^ { 2 } \\right] . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 305, + 212, + 692, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All proofs are relegated to Appendix A due to space constraints. When $\\eta _ { l }$ satisfies the condition in the second part the above result, we obtain a convergence rate depending on $\\operatorname* { m i n } \\{ \\Psi _ { \\mathrm { v a r } } , \\widetilde { \\Psi } _ { \\mathrm { v a r } } \\}$ . To obtain an explicit dependence on $T$ and $K$ , we simplify the above result for a specific choice of $\\eta , \\eta _ { l }$ and $\\tau$ . ", + "bbox": [ + 174, + 335, + 825, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Corollary 1. Suppose √ $\\eta _ { l }$ is such that the conditions in Theorem $I$ are satisfied and $\\eta _ { l } = \\Theta \\big ( { 1 } / { ( K L \\sqrt { T } } \\big )$ . Also suppose $\\eta = \\Theta ( \\sqrt { K m } )$ and $\\tau = G / _ { L }$ . Then, for sufficiently large $T$ , the iterates of Algorithm 2 for FEDADAGRAD satisfy ", + "bbox": [ + 173, + 383, + 825, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/43b9fd992563d3b2e2e52d3a830ad9c43e4930752416313d7759c7ded13d3f34.jpg", + "text": "$$\n\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\mathcal { O } \\Bigg ( \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\sqrt { m K T } } + \\frac { 2 \\sigma _ { l } ^ { 2 } L } { G ^ { 2 } \\sqrt { m K T } } + \\frac { \\sigma ^ { 2 } } { G K T } + \\frac { \\sigma ^ { 2 } L \\sqrt { m } } { G ^ { 2 } \\sqrt { K } T ^ { 3 / 2 } } \\Bigg ) .\n$$", + "text_format": "latex", + "bbox": [ + 205, + 434, + 789, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We defer a detailed discussion about our analysis and its implication to the end of the section. ", + "bbox": [ + 169, + 498, + 785, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis of FEDADAM Next, we provide the convergence analysis of FEDADAM. The proof of FEDYOGI is very similar and hence, we omit the details of FEDYOGI’s analysis. ", + "bbox": [ + 173, + 520, + 823, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. Let Assumptions 1 to 3 hold, and $L , G , \\sigma _ { l } , \\sigma _ { g }$ be as defined therein. Let $\\sigma ^ { 2 } = \\sigma _ { l } ^ { 2 } + 6 K \\sigma _ { g } ^ { 2 }$ Suppose the client learning rate satisfies $\\eta _ { l } \\leq 1 / 1 6 L K$ and ", + "bbox": [ + 173, + 551, + 825, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/e8928e419e9469dc75d92d0746178a7f01edcb0c938a07b652ba1145cfa29abf.jpg", + "text": "$$\n\\eta _ { l } \\leq \\frac { 1 } { 6 K } \\operatorname* { m i n } \\left\\{ \\left[ \\frac { \\tau } { G L } \\right] ^ { 1 / 2 } , \\left[ \\frac { \\tau ^ { 2 } } { G L ^ { 3 } \\eta } \\right] ^ { 1 / 4 } , \\left[ \\frac { \\tau } { G L ^ { 2 } } \\right] ^ { 1 / 3 } \\right\\} .\n$$", + "text_format": "latex", + "bbox": [ + 313, + 589, + 681, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the iterates of Algorithm 2 for FEDADAM satisfy ", + "bbox": [ + 176, + 636, + 531, + 651 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/4b92b739de8bef239c7482a5ea7c381ade04f5280fe2322fdfff1491c1c14c1c.jpg", + "text": "$$\n\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\odot \\left( \\frac { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } { \\eta _ { l } K T } \\left( \\Psi + \\Psi _ { \\mathrm { v a r } } \\right) \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 299, + 657, + 699, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 173, + 699, + 217, + 713 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/134d7b1d2fa61708e74810c75283b81fe424571d1047e4237a6f800cd422b0a5.jpg", + "text": "$$\n\\begin{array} { c } { { \\Psi = \\displaystyle \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\sigma ^ { 2 } , } } \\\\ { { \\Psi _ { \\nu a r } = \\left( G + \\displaystyle \\frac { \\eta L } { 2 } \\right) \\left[ \\displaystyle \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sigma _ { l } ^ { 2 } + \\displaystyle \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\sigma ^ { 2 } \\right] . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 316, + 708, + 681, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to the FEDADAGRAD case, we restate the above result for a specific choice of $\\eta _ { l } , \\eta$ and $\\tau$ in order to highlight the dependence of $K$ and $T$ . ", + "bbox": [ + 173, + 787, + 825, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Corollary 2. Suppose $\\eta _ { l }$ is chosen such that the conditions in Theorem 2 are satisfied and that√ $\\eta _ { l } = \\Theta \\bigl ( 1 / ( K L \\sqrt { T } ) \\bigr )$ . Also, suppose $\\eta = \\Theta ( \\sqrt { K m } )$ and $\\tau = G / _ { L }$ . Then, for sufficiently large $T$ , the iterates of Algorithm 2 for FEDADAM satisfy ", + "bbox": [ + 174, + 820, + 825, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/8506f0697321f1583c8f1fe7772c08b405c66458be413e11447b3a78200c04f4.jpg", + "text": "$$\n\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\mathcal { O } \\Bigg ( \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\sqrt { m K T } } + \\frac { 2 \\sigma _ { l } ^ { 2 } L } { G ^ { 2 } \\sqrt { m K T } } + \\frac { \\sigma ^ { 2 } } { G K T } + \\frac { \\sigma ^ { 2 } L \\sqrt { m } } { G ^ { 2 } \\sqrt { K } T ^ { 3 / 2 } } \\Bigg ) .\n$$", + "text_format": "latex", + "bbox": [ + 207, + 869, + 789, + 912 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1. The server learning rate $\\eta = 1$ typically used in FEDAVG does not necessarily minimize the upper bound in Theorems 1 & 2. The effect of $\\cdot _ { g }$ , a measure of client heterogeneity, on convergence can be reduced by choosing sufficiently $\\eta _ { l }$ and a reasonably large $\\eta$ (e.g. see Corollary 1). Thus, the effect of client heterogeneity can be reduced by carefully choosing client and server learning rates, but not removed entirely. Our empirical analysis (eg. Figure 1) supports this conclusion. ", + "bbox": [ + 174, + 103, + 825, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. We briefly discuss our theoretical analysis and its implications in the FL setting. The convergence rates for FEDADAGRAD and FEDADAM are similar, so our discussion applies to all the adaptive federated optimization algorithms (including FEDYOGI) proposed in the paper. ", + "bbox": [ + 174, + 183, + 823, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(i) Comparison of convergence rates. When $T$ is sufficiently large compared to $K$ , $\\mathcal { O } ( { } ^ { 1 } / { \\sqrt { m K T } } )$ is the dominant term in Corollary $1 \\ \\& \\ 2$ . Thus, we effectively obtain a convergence rate of $\\mathcal { O } ( { } ^ { 1 } / \\sqrt { m K T } )$ , which matches the best known rate for the general non-convex setting of our interest (e.g. see (Karimireddy et al., 2019)). We also note that in the $i . i . d$ setting considered in (Wang & Joshi, 2018), which corresponds to $\\sigma _ { g } = 0$ , we match their convergence rates. Similar to the centralized setting, it is possible to obtain convergence rates with better dependence on constants for federated adaptive methods, compared to FEDAVG, by incorporating non-uniform bounds on gradients across coordinates (Zaheer et al., 2018). ", + "bbox": [ + 176, + 231, + 825, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(ii) Learning rates $\\pmb { \\& }$ their decay. The client learning rate of $^ 1 / \\sqrt { T }$ in our analysis requires knowledge of the number of rounds $T$ a priori; however, it is easy to generalize our analysis to the case where $\\eta _ { l }$ is decayed at a rate of $^ { 1 } / \\sqrt { t }$ . Observe that one must decay $\\eta _ { l }$ , not the server learning rate $\\eta$ , to obtain convergence. This is because the client drift introduced by the local updates does not vanish as $T \\to \\infty$ when $\\eta _ { l }$ is constant. As we show in Appendix E.6, learning rate decay can improve empirical performance. Also, note the inverse relationship between $\\eta _ { l }$ and $\\eta$ in Corollary 1 & 2, which we observe in our empirical analysis (see Appendix E.4). ", + "bbox": [ + 173, + 342, + 825, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(iii) Communication efficiency $\\pmb { \\& }$ local steps. The total communication cost of the algorithms depends on the number of communication rounds $T$ . From Corollary 1 & 2, it is clear that a larger $K$ leads to fewer rounds of communication as long as $K = \\mathcal { O } ( T \\bar { \\sigma } _ { l } ^ { 2 } / \\sigma _ { g } ^ { 2 } )$ . Thus, the number of local iterations can be large when either the ratio $\\sigma _ { l } ^ { 2 } / \\sigma _ { g } ^ { 2 }$ or $T$ is large. In the $i . i . d$ setting where $\\sigma _ { g } = 0$ , unsurprisingly, $K$ can be very large. ", + "bbox": [ + 169, + 439, + 825, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(iv) Client heterogeneity. While careful selection of client and server learning rates can reduce the effect of client heterogeneity (see Remark 1), it does not completely remove it. In highly heterogeneous settings, it may be necessary to use mechanisms such as control variates (Karimireddy et al., 2019). However, our empirical analysis suggest that for moderate, naturally arising heterogeneity, adaptive optimizers are quite effective, especially in cross-device settings (see Figure 1). Furthermore, our algorithms can be directly combined with such mechanisms. ", + "bbox": [ + 171, + 511, + 825, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As mentioned earlier, for the sake of simplicity, our analysis assumes full-participation $\\mathcal { S } = [ m ] )$ . Our analysis can be directly generalized to limited participation at the cost of an additional variance term in our rates that depends on $| { \\cal S } | / m$ , the fraction of clients sampled (see Section A.2.1 for details). ", + "bbox": [ + 174, + 599, + 825, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 EXPERIMENTAL EVALUATION: DATASETS, TASKS, AND METHODS ", + "bbox": [ + 171, + 660, + 753, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our algorithms on what we believe is the most extensive and representative suite of federated datasets and modeling tasks to date. We wish to understand how server adaptivity can help improve convergence, especially in cross-device settings. To accomplish this, we conduct simulations on seven diverse and representative learning tasks across five datasets. Notably, three of the five have a naturally-arising client partitioning, highly representative of real-world FL problems. ", + "bbox": [ + 174, + 690, + 825, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets, models, and tasks We use five datasets: CIFAR-10, CIFAR-100 (Krizhevsky & Hinton, 2009), EMNIST (Cohen et al., 2017), Shakespeare (McMahan et al., 2017), and Stack Overflow (Authors, 2019). The first three are image datasets, the last two are text datasets. For CIFAR-10 and CIFAR-100, we train ResNet-18 (replacing batch norm with group norm (Hsieh et al., 2019)). For EMNIST, we train a CNN for character recognition (EMNIST CR) and a bottleneck autoencoder (EMNIST AE). For Shakespeare, we train an RNN for next-character-prediction. For Stack Overflow, we perform tag prediction using logistic regression on bag-of-words vectors (SO LR) and train an RNN to do next-word-prediction (SO NWP). For full details of the datasets, see Appendix C. ", + "bbox": [ + 173, + 766, + 825, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation We implement all algorithms in TensorFlow Federated (Ingerman & Ostrowski, 2019). Clients are sampled uniformly at random, without replacement in a given round, but with replacement across rounds. Our implementation has two important characteristics. First, instead of doing $K$ training steps per client, we do $E$ epochs of training over each client’s dataset. Second, to account for varying numbers of gradient steps per client, we weight the average of the client outputs $\\Delta _ { i } ^ { t }$ by each client’s number of training samples. This follows the approach of (McMahan et al., 2017), and can often outperform uniform weighting (Zaheer et al., 2018). For full descriptions of the algorithms used, see Appendix B. ", + "bbox": [ + 176, + 882, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 171 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimizers and hyperparameters We compare FEDADAGRAD, FEDADAM, and FEDYOGI (with adaptivity $\\tau$ ) to FEDOPT where CLIENTOPT and SERVEROPT are SGD with learning rates $\\eta _ { l }$ and $\\eta$ . For the server, we use a momentum parameter of 0 (FEDAVG), and 0.9 (FEDAVGM). We fix the client batch size on a per-task level (see Appendix D.3). For FEDADAM and FEDYOGI, we fix a momentum parameter $\\beta _ { 1 } = 0 . 9$ and a second moment parameter $\\beta _ { 2 } = 0 . 9 9$ . We also compare to SCAFFOLD (see Appendix B.2 for implementation details). For SO NWP, we sample 50 clients per round, while for all other tasks we sample 10. We use $E = 1$ local epochs throughout. ", + "bbox": [ + 173, + 179, + 825, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We select $\\eta , ~ \\eta$ , and $\\tau$ by grid-search tuning. While this is often done using validation data in centralized settings, such data is often inaccessible in FL, especially cross-device FL. Therefore, we tune by selecting the parameters that minimize the average training loss over the last 100 rounds of training. We run 1500 rounds of training on the EMNIST CR, Shakespeare, and Stack Overflow tasks, 3000 rounds for EMNIST AE, and 4000 rounds for the CIFAR tasks. For more details and a record of the best hyperparameters, see Appendix D. ", + "bbox": [ + 174, + 281, + 825, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Validation metrics For all tasks, we measure the performance on a validation set throughout training. For Stack Overflow tasks, the validation set contains 10,000 randomly sampled test examples (due to the size of the test dataset, see Table 2). For all other tasks, we use the entire test set. Since all algorithms exchange equal-sized objects between server and clients, we use the number of communication rounds as a proxy for wall-clock training time. ", + "bbox": [ + 174, + 371, + 825, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 EXPERIMENTAL EVALUATION: RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 458, + 539, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 COMPARISONS BETWEEN METHODS ", + "text_level": 1, + "bbox": [ + 178, + 489, + 460, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the convergence of our adaptive methods to non-adaptive methods: FEDAVG, FEDAVGM and SCAFFOLD. Plots of validation performances for each task/optimizer are in Figure 1, and Table 1 summarizes the last-100-round validation performance. Due to space constraints, results for EMNIST CR are in Appendix E.1, and full test set results for Stack Overflow are in Appendix E.2. ", + "bbox": [ + 174, + 515, + 825, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sparse-gradient tasks Text data often produces long-tailed feature distributions, leading to approximately-sparse gradients which adaptive optimizers can capitalize on (Zhang et al., 2019a). Both Stack Overflow tasks exhibit such behavior, though they are otherwise dramatically different— in feature representation (bag-of-words vs. variable-length token sequence), model architecture (GLM vs deep network), and optimization landscape (convex vs nonconvex). In both tasks, words that do not appear in a client’s dataset produce nearzero client updates. Thus, the accumulators $v _ { t , j }$ in Algorithm 2 remain small for parameters tied to rare words, allowing large ", + "bbox": [ + 174, + 577, + 452, + 792 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/241593e968cb0ffd4a91a36315da5518b624bd18588c254e50d8dbf78f761dc6.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Average validation performance over the last 100 rounds: $\\%$ accuracy for rows 1–5; Recall $\\textcircled { a } 5$ $( \\times 1 0 0 )$ for Stack Overflow LR; and MSE $( \\times 1 0 0 0 )$ for EMNIST AE. Performance within $0 . 5 \\%$ of the best result for each task are shown in bold. " + ], + "table_body": "
FED..ADAGRADADAMYoGIAvGMAvG
CIFAR-1072.177.478.077.472.8
CIFAR-10047.952.552.452.444.7
EMNIST CR85.185.685.585.284.9
SHAKESPEARE57.557.057.257.356.9
SO NWP23.825.225.223.819.5
SO LR67.165.865.936.930.0
EMNIST AE4.201.010.981.656.47
", + "bbox": [ + 470, + 579, + 815, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "updates to be made when they do occur. This intuition is born out in Figure 1, where adaptive optimizers dramatically outperform non-adaptive ones. For the non-convex NWP task, momentum is also critical, whereas it slightly hinders performance for the convex LR task. ", + "bbox": [ + 174, + 794, + 821, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dense-gradient tasks CIFAR-10/100, EMNIST AE/CR, and Shakespeare lack a sparse-gradient structure. Shakespeare is relatively easy—most optimizers perform well after enough rounds once suitably tuned, though FEDADAGRAD converges faster. For CIFAR-10/100 and EMNIST AE, adaptivity and momentum offer substantial improvements over FEDAVG. Moreover, FEDYOGI and FEDADAM have faster initial convergence than FEDAVGM on these tasks. Notably, FEDADAM and FEDYOGI perform comparably to or better than non-adaptive optimizers throughout, and close to or better than ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/291f261557be507186a2a9304e87c9225666c4791011ea1091541e22f3f95278.jpg", + "image_caption": [ + "Figure 1: Validation accuracy of adaptive and non-adaptive methods, as well as SCAFFOLD, using constant learning rates $\\eta$ and $\\eta _ { l }$ tuned to achieve the best training performance over the last 100 communication rounds; see Appendix D.2 for grids. " + ], + "image_footnote": [], + "bbox": [ + 204, + 87, + 792, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FEDADAGRAD throughout. As we discuss below, FEDADAM and FEDYOGI actually enable easier learning rate tuning than FEDAVGM in many tasks. ", + "bbox": [ + 174, + 392, + 823, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to SCAFFOLD On all tasks, SCAFFOLD performs comparably to or worse than FEDAVG and our adaptive methods. On Stack Overflow, SCAFFOLD and FEDAVG are nearly identical. This is because the number of clients (342,477) makes it unlikely we sample any client more than once. Intuitively, SCAFFOLD does not have a chance to use its client control variates. In other tasks, SCAFFOLD performs worse than other methods. We present two possible explanations: First, we only sample a small fraction of clients at each round, so most users are sampled infrequently. Intuitively, the client control variates can become stale, and may consequently degrade the performance. Second, SCAFFOLD is similar to variance reduction methods such as SVRG (Johnson & Zhang, 2013). While theoretically performant, such methods often perform worse than SGD in practice (Defazio & Bottou, 2018). As shown by Defazio et al. (2014), variance reduction often only accelerates convergence when close to a critical point. In cross-device settings (where the number of communication rounds are limited), SCAFFOLD may actually reduce empirical performance. ", + "bbox": [ + 173, + 428, + 826, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 EASE OF TUNING ", + "text_level": 1, + "bbox": [ + 174, + 611, + 333, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1ecef7595988e2f35b8b5c2ee3db0d5eb2f0bad2ba1b6103616e6c3ce2fd289.jpg", + "image_caption": [ + "Figure 2: Validation accuracy (averaged over the last 100 rounds) of FEDADAM, FEDYOGI, and FEDAVGM for various client/server learning rates combination on the SO NWP task. For FEDADAM and FEDYOGI, we set $\\tau = 1 0 ^ { - 3 }$ . " + ], + "image_footnote": [], + "bbox": [ + 197, + 642, + 794, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Obtaining optimal performance involves tuning $\\eta _ { l } , \\eta$ , and for the adaptive methods, $\\tau$ . To quantify how easy it is to tune various methods, we plot their validation performance as a function of $\\eta _ { l }$ and $\\eta$ Figure 2 gives results for FEDADAM, FEDYOGI, and FEDAVGM on Stack Overflow NWP. Plots for all other optimizers and tasks are in Appendix E.3. For FEDAVGM, there are only a few good values of $\\eta _ { l }$ for each $\\eta$ , while for FEDADAM and FEDYOGI, there are many good values of $\\eta _ { l }$ for a range of $\\eta$ . Thus, FEDADAM and FEDYOGI are arguably easier to tune in this setting. Similar results hold for other tasks and optimizers (Figures 5 to 11). ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd2738f345eafd8cd9372344ebd4c03605653f82aa8c8583ca069896d16d9d6c.jpg", + "image_caption": [ + "Figure 3: Validation performance of FEDADAGRAD, FEDADAM, and FEDYOGI for varying $\\tau$ on various tasks. The learning rates $\\eta$ and $\\eta _ { l }$ are tuned for each $\\tau$ to achieve the best training performance on the last 100 communication rounds. " + ], + "image_footnote": [], + "bbox": [ + 204, + 104, + 792, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 411, + 823, + 439 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This leads to a natural question: Is the reduction in the need to tune $\\eta _ { l }$ and $\\eta$ offset by the need to tune the adaptivity $\\tau ?$ In fact, while we tune $\\tau$ in Figure 1, our results are relatively robust to $\\tau$ . To demonstrate, we plot the best validation performance for various $\\tau$ in Figure 3. For nearly all tasks and optimizers, $\\dot { \\tau } = 1 0 ^ { - 3 }$ works almost as well all other values. This aligns with work by Zaheer et al. (2018), who show that moderately large $\\tau$ yield better performance for centralized adaptive optimizers. FEDADAM and FEDYOGI see only small differences in performance among $\\tau$ on all tasks except Stack Overflow LR (for which FEDADAGRAD is the best optimizer, and is robust to $\\tau$ ). ", + "bbox": [ + 174, + 445, + 825, + 541 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3 OTHER FINDINGS ", + "text_level": 1, + "bbox": [ + 176, + 564, + 338, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present additional empirical analyses in Appendix E. These include EMNIST CR results (Appendix E.1), Stack Overflow results on the full test dataset (Appendix E.2), client/server learning rate heat maps for all optimizers and tasks (Appendix E.3), an analysis of the relationship between $\\eta$ and $\\eta _ { l }$ (Appendix E.4), and experiments with learning rate decay (Appendix E.6). ", + "bbox": [ + 174, + 592, + 825, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 672, + 318, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we demonstrated that adaptive optimizers can be powerful tools in improving the convergence of FL. By using a simple client/server optimizer framework, we can incorporate adaptivity into FL in a principled, intuitive, and theoretically-justified manner. We also developed comprehensive benchmarks for comparing federated optimization algorithms. To encourage reproducibility and breadth of comparison, we have attempted to describe our experiments as rigorously as possible, and have created an open-source framework with all models, datasets, and code. We believe our work raises many important questions about how best to perform federated optimization. Example directions for future research include understanding how the use of adaptivity affects differential privacy and fairness. ", + "bbox": [ + 174, + 707, + 825, + 830 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 856, + 285, + 871 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The TensorFlow Federated Authors. TensorFlow Federated Stack Overflow dataset, 2019. URL https://www.tensorflow.org/federated/api_docs/python/tff/ simulation/datasets/stackoverflow/load_data. ", + "bbox": [ + 176, + 882, + 825, + 922 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Debraj Basu, Deepesh Data, Can Karakus, and Suhas Diggavi. Qsparse-local-SGD: Distributed SGD with quantization, sparsification and local computations. In Advances in Neural Information Processing Systems, pp. 14668–14679, 2019. ", + "bbox": [ + 178, + 103, + 823, + 145 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex Ingerman, Vladimir Ivanov, Chloé Kiddon, Jakub Konecný, Stefano Mazzocchi, Brendan McMahan, Timon ˇ Van Overveldt, David Petrou, Daniel Ramage, and Jason Roselander. Towards federated learning at scale: System design. In A. Talwalkar, V. Smith, and M. Zaharia (eds.), Proceedings of Machine Learning and Systems, volume 1, pp. 374–388. Proceedings of MLSys, 2019. URL https://proceedings.mlsys.org/paper/2019/file/ bd686fd640be98efaae0091fa301e613-Paper.pdf. ", + "bbox": [ + 174, + 156, + 825, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sebastian Caldas, Peter Wu, Tian Li, Jakub Konecný, H Brendan McMahan, Virginia Smith, and ˇ Ameet Talwalkar. LEAF: A benchmark for federated settings. arXiv preprint arXiv:1812.01097, 2018. ", + "bbox": [ + 174, + 262, + 826, + 303 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. EMNIST: Extending MNIST to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pp. 2921–2926. IEEE, 2017. ", + "bbox": [ + 173, + 314, + 823, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aaron Defazio and Léon Bottou. On the ineffectiveness of variance reduced optimization for deep learning. arXiv preprint arXiv:1812.04529, 2018. ", + "bbox": [ + 171, + 366, + 823, + 395 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aaron Defazio, Francis Bach, and Simon Lacoste-Julien. SAGA: A fast incremental gradient method with support for non-strongly convex composite objectives. In NIPS, pp. 1646–1654, 2014. ", + "bbox": [ + 173, + 404, + 823, + 433 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011. ", + "bbox": [ + 171, + 443, + 823, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training ImageNet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. ", + "bbox": [ + 176, + 481, + 823, + 523 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip B Gibbons. The non-IID data quagmire of decentralized machine learning. arXiv preprint arXiv:1910.00189, 2019. ", + "bbox": [ + 176, + 534, + 823, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019. ", + "bbox": [ + 174, + 571, + 821, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Ingerman and Krzys Ostrowski. Introducing TensorFlow Federated, 2019. URL https://medium.com/tensorflow/ introducing-tensorflow-federated-a4147aa20041. ", + "bbox": [ + 174, + 609, + 826, + 652 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive variance reduction. In Advances in Neural Information Processing Systems, pp. 315–323, 2013. ", + "bbox": [ + 174, + 662, + 823, + 691 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Keith Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. arXiv preprint arXiv:1912.04977, 2019. ", + "bbox": [ + 176, + 700, + 823, + 743 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for on-device federated learning. arXiv preprint arXiv:1910.06378, 2019. ", + "bbox": [ + 176, + 752, + 825, + 795 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ahmed Khaled, Konstantin Mishchenko, and Peter Richtárik. First analysis of local GD on heterogeneous data. arXiv preprint arXiv:1909.04715, 2019. ", + "bbox": [ + 169, + 804, + 825, + 833 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. ", + "bbox": [ + 174, + 843, + 825, + 886 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009. ", + "bbox": [ + 176, + 896, + 823, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. arXiv preprint arXiv:1812.06127, 2018. ", + "bbox": [ + 173, + 103, + 826, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. arXiv preprint arXiv:1908.07873, 2019a. ", + "bbox": [ + 173, + 140, + 825, + 169 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Wei Li and Andrew McCallum. Pachinko allocation: DAG-structured mixture models of topic correlations. In Proceedings of the 23rd international conference on Machine learning, pp. 577– 584, 2006. ", + "bbox": [ + 174, + 176, + 825, + 218 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of FedAvg on non-IID data. arXiv preprint arXiv:1907.02189, 2019b. ", + "bbox": [ + 173, + 227, + 825, + 255 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xiaoyu Li and Francesco Orabona. On the convergence of stochastic gradient descent with adaptive stepsizes. arXiv preprint arXiv:1805.08114, 2018. ", + "bbox": [ + 174, + 262, + 825, + 291 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic bound of learning rate. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. URL https://openreview. net/forum?id=Bkg3g2R9FX. ", + "bbox": [ + 173, + 299, + 826, + 356 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April 2017, Fort Lauderdale, FL, USA, pp. 1273–1282, 2017. URL http://proceedings.mlr. press/v54/mcmahan17a.html. ", + "bbox": [ + 173, + 363, + 826, + 433 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "H. Brendan McMahan and Matthew Streeter. Adaptive bound optimization for online convex optimization. In COLT, 2010a. ", + "bbox": [ + 171, + 440, + 823, + 469 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "H. Brendan McMahan and Matthew J. Streeter. Adaptive bound optimization for online convex optimization. In COLT The 23rd Conference on Learning Theory, 2010b. ", + "bbox": [ + 173, + 476, + 823, + 506 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sashank J Reddi, Ahmed Hefny, Suvrit Sra, Barnabás Póczós, and Alex Smola. Stochastic variance reduction for nonconvex optimization. arXiv:1603.06160, 2016. ", + "bbox": [ + 174, + 512, + 823, + 541 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of ADAM and beyond. arXiv preprint arXiv:1904.09237, 2019. ", + "bbox": [ + 173, + 549, + 823, + 578 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sebastian U. Stich. Local SGD converges fast and communicates little. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id= S1g2JnRcFX. ", + "bbox": [ + 174, + 585, + 826, + 627 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sebastian U Stich and Sai Praneeth Karimireddy. The error-feedback framework: Better rates for SGD with delayed gradients and compressed communication. arXiv preprint arXiv:1909.05350, 2019. ", + "bbox": [ + 174, + 636, + 826, + 678 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jianyu Wang and Gauri Joshi. Cooperative SGD: A unified framework for the design and analysis of communication-efficient SGD algorithms. arXiv preprint arXiv:1808.07576, 2018. ", + "bbox": [ + 171, + 685, + 825, + 715 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Shiqiang Wang, Tiffany Tuor, Theodoros Salonidis, Kin K Leung, Christian Makaya, Ting He, and Kevin Chan. Adaptive federated learning in resource constrained edge computing systems. IEEE Journal on Selected Areas in Communications, 37(6):1205–1221, 2019. ", + "bbox": [ + 173, + 722, + 826, + 765 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rachel Ward, Xiaoxia Wu, and Leon Bottou. Adagrad stepsizes: Sharp convergence over nonconvex landscapes, from any initialization. arXiv preprint arXiv:1806.01811, 2018. ", + "bbox": [ + 173, + 772, + 825, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xiaoxia Wu, Simon S Du, and Rachel Ward. Global convergence of adaptive gradient methods for an over-parameterized neural network. arXiv preprint arXiv:1902.07111, 2019. ", + "bbox": [ + 173, + 808, + 821, + 838 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 3–19, 2018. ", + "bbox": [ + 176, + 844, + 821, + 875 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cong Xie, Oluwasanmi Koyejo, Indranil Gupta, and Haibin Lin. Local AdaAlter: Communicationefficient stochastic gradient descent with adaptive learning rates. arXiv preprint arXiv:1911.09030, 2019. ", + "bbox": [ + 174, + 882, + 825, + 922 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hao Yu, Sen Yang, and Shenghuo Zhu. Parallel restarted SGD with faster convergence and less communication: Demystifying why model averaging works for deep learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pp. 5693–5700, 2019. ", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Manzil Zaheer, Sashank Reddi, Devendra Sachan, Satyen Kale, and Sanjiv Kumar. Adaptive methods for nonconvex optimization. In Advances in Neural Information Processing Systems, pp. 9815– 9825, 2018. ", + "bbox": [ + 173, + 154, + 825, + 195 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jingzhao Zhang, Sai Praneeth Karimireddy, Andreas Veit, Seungyeon Kim, Sashank J. Reddi, Sanjiv Kumar, and Suvrit Sra. Why ADAM beats SGD for attention models. arXiv preprint arxiv:1912.03194, 2019a. ", + "bbox": [ + 171, + 204, + 825, + 246 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Michael R. Zhang, James Lucas, Jimmy Ba, and Geoffrey E. Hinton. Lookahead optimizer: k steps forward, 1 step back. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, 8-14 December 2019, Vancouver, BC, Canada, pp. 9593–9604, 2019b. ", + "bbox": [ + 173, + 255, + 825, + 324 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Martin Zinkevich, Markus Weimer, Lihong Li, and Alex J Smola. Parallelized stochastic gradient descent. In Advances in neural information processing systems, pp. 2595–2603, 2010. ", + "bbox": [ + 173, + 333, + 823, + 362 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A PROOF OF RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 102, + 372, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1 MAIN CHALLENGES ", + "text_level": 1, + "bbox": [ + 176, + 137, + 359, + 152 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We first recap some of the central challenges to our analysis. Theoretical analyses of optimization methods for federated learning are much different than analyses for centralized settings. The key factors complicating the analysis are: ", + "bbox": [ + 173, + 166, + 825, + 208 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "1. Clients performing multiple local updates. \n2. Data heterogeneity. \n3. Understanding the communication complexity. ", + "bbox": [ + 210, + 222, + 539, + 289 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As a result of (1), the updates from the clients to the server are not gradients, or even unbiased estimates of gradients, they are pseudo-gradients (see Section 2). These pseudo-gradients are challenging to analyze as they can have both high bias (their expectation is not the gradient of the empirical loss function) and high variance (due to compounding variance across client updates) and are therefore challenging to bound. This is exacerbated by (2), which we quantify by the parameter $\\sigma _ { g }$ in Section 2. Things are further complicated by (3), as we must obtain a good trade-off between the number of client updates taken per round ( $K$ in Algorithms 1 and 2) and the number of communication rounds $T$ . Such trade-offs do not exist in centralized optimization. ", + "bbox": [ + 173, + 303, + 826, + 412 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.2 PROOF OF THEOREM 1 ", + "text_level": 1, + "bbox": [ + 176, + 436, + 375, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Proof of Theorem 1. Recall that the server update of FEDADAGRAD is the following ", + "bbox": [ + 176, + 464, + 728, + 481 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/70e8f36e5a1da551ed96f15bc036a6dbdec58dfee69595e72a57d119957bca0d.jpg", + "text": "$$\nx _ { t + 1 , i } = x _ { t , i } + \\eta \\frac { \\Delta _ { t , i } } { \\sqrt { v _ { t , i } } + \\tau } ,\n$$", + "text_format": "latex", + "bbox": [ + 401, + 492, + 593, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "for all $i \\in [ d ]$ . Since the function $f$ is $L$ -smooth, we have the following: ", + "bbox": [ + 173, + 541, + 643, + 556 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/a02d4e87b347ccb793d5a7bc761291d399804b6a07c0c04243b749afdacd2e0a.jpg", + "text": "$$\n\\begin{array} { l } { f ( x _ { t + 1 } ) \\leq f ( x _ { t } ) + \\langle \\nabla f ( x _ { t } ) , x _ { t + 1 } - x _ { t } \\rangle + \\displaystyle \\frac { L } { 2 } \\| x _ { t + 1 } - x _ { t } \\| ^ { 2 } } \\\\ { \\displaystyle = f ( x _ { t } ) + \\eta \\left. \\nabla f ( x _ { t } ) , \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } \\right. + \\displaystyle \\frac { \\eta ^ { 2 } L } { 2 } \\sum _ { i = 1 } ^ { d } \\frac { \\Delta _ { t , i } ^ { 2 } } { ( \\sqrt { v _ { t , i } } + \\tau ) ^ { 2 } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 267, + 569, + 728, + 646 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The second step follows simply from FEDADAGRAD’s update. We take the expectation of $f ( x _ { t + 1 } )$ (over randomness at time step $t$ ) in the above inequality: ", + "bbox": [ + 174, + 657, + 828, + 686 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/4b80d6a909d656a015e6c315e7b4af7d453e3b5da9042aa0d788a35d2cbed587.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\ddot { c } _ { i } [ f ( x _ { i + 1 } ) ] \\leq f ( x _ { i } ) + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\Psi _ { t - 1 } } + \\tau } \\right] \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } ^ { 2 } \\rho } { \\left( \\sqrt { \\nu _ { t - 1 } } + \\tau \\right) ^ { 2 } } \\right] } \\\\ { = f ( x _ { i } ) + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t + 1 } } + \\tau } - \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } + \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } \\\\ { \\displaystyle \\qquad + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i , j } ^ { 2 } } { \\left( \\sqrt { \\nu _ { t , j } } + \\tau \\right) ^ { 2 } } \\right] } \\\\ { = f ( x _ { i } ) + \\eta \\underbrace { \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } _ { \\displaystyle \\mathrm { ~ ( ~ 1 ~ ) ~ } } + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t } } + \\tau } - \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } \\\\ { \\displaystyle \\qquad + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i , j } ^ { 2 } } { \\left( \\sqrt { \\nu _ { t - 1 } } + \\tau \\right) ^ { 2 } } \\right] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 699, + 839, + 928 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We will first bound $T _ { 2 }$ in the following manner: ", + "bbox": [ + 171, + 103, + 486, + 118 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/df03e20292f518fafce8e9aa33fed901c8bdbcb91da1627722123ed477863934.jpg", + "text": "$$\n\\begin{array} { r l } & { T _ { 2 } = \\bigg \\langle \\nabla f ( { x } _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { v _ { t - 1 } } + \\tau } \\right] \\bigg \\rangle } \\\\ & { \\quad = \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } [ \\nabla f ( { x } _ { t } ) ] _ { j } \\times \\left[ \\frac { \\Delta _ { t , j } } { \\sqrt { v _ { t , j } } + \\tau } - \\frac { \\Delta _ { t , j } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\right] } \\\\ & { \\quad = \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } [ \\nabla f ( { x } _ { t } ) ] _ { j } \\times \\Delta _ { t , j } \\times \\left[ \\frac { \\sqrt { v _ { t - 1 , j } } - \\sqrt { v _ { t , j } } } { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } \\right] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 171, + 126, + 598, + 256 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "and recalling $v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }$ so $- \\Delta _ { t , j } ^ { 2 } = ( \\sqrt { v _ { t - 1 , j } } - \\sqrt { v _ { t , j } } ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) )$ we have, ", + "bbox": [ + 169, + 262, + 772, + 284 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/89a45736bfecac0a9a4c9b691d21a2ec279cc835dfa861191c16c256200c233e.jpg", + "text": "$$\n\\begin{array} { r l r } { { = \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } [ \\nabla f ( x _ { t } ) ] _ { j } \\times \\Delta _ { t , j } \\times [ \\overline { { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) } } ] } } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) } } ] } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( v _ { t , j } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } } ] } & { \\mathrm { s i n c e } v _ { t - 1 , j } \\geq \\tau } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( v _ { t , j } + \\tau ^ { 2 } ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } } ] } & { \\mathrm { s i n c e } v _ { t - 1 , j } \\geq \\tau } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 191, + 291, + 838, + 429 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Here $v _ { t - 1 , j } \\geq \\tau$ since $v _ { - 1 } \\geq \\tau$ (see the initialization of Algorithm 2) and $v _ { t , j }$ is increasing in $t$ . The above bound can be further upper bounded in the following manner: ", + "bbox": [ + 173, + 436, + 825, + 465 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/78157b390d5ece8a6d5a4fcc282b59da5c64013fc66f3ef02080752845ef50e4.jpg", + "text": "$$\n\\begin{array} { r l } { T _ { 2 } \\le \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } \\eta _ { l } K G ^ { 2 } \\left[ \\displaystyle \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } \\right] } & { \\mathrm { ~ s i n c e ~ } [ \\nabla f ( x _ { t } ) ] _ { i } \\le G \\mathrm { ~ a n d ~ } \\Delta _ { t , i } \\le \\eta _ { l } K G } \\\\ { \\le \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } \\displaystyle \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } \\left[ \\displaystyle \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] } & { \\mathrm { ~ s i n c e ~ } \\sqrt { v _ { t - 1 , j } } \\ge 0 . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 191, + 473, + 807, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Bounding $T _ { 1 }$ We now turn our attention to bounding the term $T _ { 1 }$ , which we need to be sufficiently negative. We observe the following: ", + "bbox": [ + 171, + 580, + 823, + 611 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/e50b63fd171470cab38be028f883c04445fb08d3cd1f7574b91e6d4330590510.jpg", + "text": "$$\n\\begin{array} { r l } & { T _ { 1 } = \\bigg \\langle \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t - 1 } } + \\tau } \\right] \\bigg \\rangle } \\\\ & { \\quad = \\bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } - \\eta _ { l } K \\nabla f ( x _ { t } ) + \\eta _ { l } K \\nabla f ( x _ { t } ) \\right] \\bigg \\rangle } \\\\ & { \\quad = - \\eta _ { l } K \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\underbrace { \\left. \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) \\right] \\right. } _ { T _ { 2 } } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 256, + 618, + 740, + 747 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In order to bound $T _ { 1 }$ , we use the following upper bound on $T _ { 3 }$ (which captures the difference between the actual update $\\Delta _ { t }$ and an appropriate scaling of $- \\nabla f ( x _ { t } ) )$ : ", + "bbox": [ + 171, + 758, + 825, + 789 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/681eaac06a8b37de6351e2cb8ab0982ee32cb67d32339873a7c300da74320fbf.jpg", + "text": "$$\n\\begin{array} { r l } & { T _ { 3 } = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. } \\\\ & { \\quad = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ - \\displaystyle \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\eta _ { l } g _ { i , k } ^ { t } + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. } \\\\ & { \\quad = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ - \\displaystyle \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\eta _ { l } \\nabla F _ { i } ( x _ { i , k } ^ { t } ) + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 171, + 797, + 643, + 925 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Here we used the fact that $\\begin{array} { r } { \\nabla f ( x _ { t } ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\nabla F _ { i } ( x _ { t } ) } \\end{array}$ and $g _ { i , k } ^ { t }$ is an unbiased estimator of the gradient at $\\boldsymbol { x } _ { i , k } ^ { t }$ , we further bound $T _ { 3 }$ as follows using a simple application of the fact that $a b \\leq$ $( a ^ { 2 } + b ^ { 2 } ) / 2$ . : ", + "bbox": [ + 173, + 99, + 826, + 150 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/135fb6d27abb4f64ac497fc2e6dfe5212caeb8bde829228a6b2a68b4ae53ff54.jpg", + "text": "$$\n\\begin{array} { r l } { r _ { 3 } \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\sum _ { j } ^ { d } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } } { 2 K } \\mathbb { E } _ { t } \\left[ \\left\\| \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\frac { \\nabla F _ { i } ( x _ { i , k } ^ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } - \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\frac { \\nabla F _ { i } ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } \\right\\| ^ { 2 } \\right] } & { } \\\\ { \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\left[ \\nabla f ( x _ { t } ) \\right] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } } { 2 m } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\left\\| \\frac { \\nabla F _ { i } ( x _ { i , k } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } \\right\\| ^ { 2 } \\right] } & { } \\\\ { \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\left[ \\nabla f ( x _ { t } ) \\right] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right] } & { \\mathrm { u s i n g ~ A s s u m p t i o n ~ 1 ~ a n d ~ } v _ { t - 1 } \\geq 0 } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 150, + 838, + 297 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The second inequality follows from Lemma 6. The last inequality follows from $L$ -Lipschitz nature of the gradient (Assumption 1). We now prove a lemma that bounds the “drift” of the $\\dot { \\boldsymbol { x } } _ { i , k } ^ { t }$ from $x _ { t }$ : ", + "bbox": [ + 174, + 310, + 826, + 340 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Lemma 3. For any step-size satisfying $\\begin{array} { r } { \\eta _ { l } \\le \\frac { 1 } { 8 L K } } \\end{array}$ , we can bound the drift for any $k \\in \\{ 0 , \\cdots , K - 1 \\}$ as ", + "bbox": [ + 171, + 347, + 825, + 376 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/877d85de065b125fcbdd573ab793a40671501e8b9fe2d9027ba7faad2e0b960c.jpg", + "text": "$$\n\\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 2 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\big ] .\n$$", + "text_format": "latex", + "bbox": [ + 228, + 375, + 767, + 421 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Proof. The result trivially holds for $k = 1$ since $x _ { i , 0 } ^ { t } = x _ { t }$ for all $i \\in [ m ]$ . We now turn our attention to the case where $k \\geq 1$ . To prove the above result, we observe that for any client $i \\in [ m ]$ and $k \\in [ K ]$ , ", + "bbox": [ + 173, + 434, + 826, + 479 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/b92da40aa6cb42737b33859fef28eadcd6c3cc547d30b73fe54c6f61fb504c6f.jpg", + "text": "$$\n\\begin{array} { r l } { { \\tilde { \\mathbf { \\mu } } _ { i , k , - } ^ { t } - x _ { t } \\| ^ { 2 } = \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } - \\eta _ { l } g _ { i , k - 1 } ^ { t } \\big \\| ^ { 2 } } } \\\\ & { \\leq \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } - \\eta _ { l } ( g _ { i , k - 1 } ^ { t } - \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) + \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) + \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) + \\nabla F _ { i } ( x _ { t } ) - \\nabla F ( x _ { t } ) \\big \\| ^ { 2 } } \\\\ & { \\leq ( 1 + \\frac { 1 } { 2 K - 1 } ) \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\big \\| ^ { 2 } + \\mathbb { E } \\big \\| \\eta _ { l } ( g _ { i , k - 1 } ^ { t } - \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) ) \\big \\| ^ { 2 } } \\\\ & { \\qquad + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) ) \\big \\| ^ { 2 } ] + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\big \\| ^ { 2 } \\big ] + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 478, + 839, + 583 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The first inequality uses the fact that $g _ { k - 1 , i } ^ { t }$ is an unbiased estimator of $\\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } )$ and Lemma 7. The above quantity can be further bounded by the following: ", + "bbox": [ + 174, + 582, + 861, + 612 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/62ce5a6b5255a01075f279a3e565de1e807443577d0e01b14c00e33e37b7137c.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\mathbb { E } \\Vert x _ { i , k } ^ { t } - x _ { t } \\Vert ^ { 2 } \\leq ( 1 + \\frac { 1 } { 2 K - 1 } ) \\mathbb { E } \\Vert x _ { i , k - 1 } ^ { t } - x _ { t } \\Vert ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\Vert L ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \\Vert ^ { 2 } } } \\\\ & { } & { + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\Vert ^ { 2 } ] + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } \\nabla f ( x _ { t } ) ) ) \\Vert ^ { 2 } ] } \\\\ & { } & { = ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\mathbb { E } \\Vert ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \\Vert ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } } \\\\ & { } & { + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\Vert ^ { 2 } ] + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\Vert \\nabla f ( x _ { t } ) ) ) \\Vert ^ { 2 } ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 613, + 821, + 746 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Here, the first inequality follows from Assumption 1 and 2. Averaging over the clients $i$ , we obtain the following: ", + "bbox": [ + 173, + 744, + 823, + 772 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/9907b46276cfcd32c409d8f54bf5fb2a33c7b852993595a81028be61508d4d18.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\frac { 1 } { n } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } } } \\\\ & { } & { + \\frac { 6 K } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } [ \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\| ^ { 2 } ] + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) } \\\\ & { } & { \\leq ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\frac { 1 } { m } \\displaystyle \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K o ( \\eta _ { l } ^ { 2 } ) ) } \\\\ & { } & { + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) \\| ^ { 2 } ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 772, + 839, + 929 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "From the above, we get the following inequality: ", + "bbox": [ + 174, + 103, + 493, + 118 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/06509a0c7208e1d1f63971a170637487f5b95ad5edfbd77444befa383de925f4.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq ( 1 + \\frac { 1 } { K - 1 } ) \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { } & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\Big ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 194, + 122, + 803, + 189 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Unrolling the recursion, we obtain the following: ", + "bbox": [ + 173, + 189, + 496, + 204 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/b0cd2701990e2815bcc71e7a5558696a35f4fa2fbc4e531360177bd889510cfa.jpg", + "text": "$$\n\\begin{array} { r l } & { \\displaystyle \\frac { 1 } { n } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq \\sum _ { p = 0 } ^ { k - 1 } \\left( 1 + \\frac { 1 } { K - 1 } \\right) ^ { p } \\left[ \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\leq \\left( K - 1 \\right) \\times \\left[ \\left( 1 + \\frac { 1 } { K - 1 } \\right) ^ { K } - 1 \\right] \\times \\left[ \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\leq \\left[ 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 207, + 826, + 358 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "concluding the proof of Lemma 3. The last inequality uses the fact that $( 1 + \\frac { 1 } { K - 1 } ) ^ { K } \\leq 5$ for $K > 1$ . ", + "bbox": [ + 174, + 361, + 825, + 392 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Using the above lemma in Equation 6 and Condition I, we can bound $T _ { 3 }$ in the following manner: ", + "bbox": [ + 171, + 406, + 812, + 422 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/5e9af182310dca586e22d5d9dba22a9cad9a2456f50adb4f15c609ecd0f4e85a.jpg", + "text": "$$\n\\begin{array} { r l } & { T _ { 3 } \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\sum _ { j = 1 } ^ { d } ( [ x _ { i , k } ^ { t } ] _ { j } - [ x _ { t } ] _ { j } ) ^ { 2 } \\right] } \\\\ & { \\quad \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } K L ^ { 2 } } { 2 \\tau } \\left[ 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\quad \\leq \\displaystyle \\frac { 3 \\eta _ { l } K } { 4 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 169, + 425, + 816, + 571 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Here we used the fact that $\\sqrt { v _ { t - 1 , j } } \\leq \\eta _ { l } K G \\sqrt { T }$ and Condition I in Theorem 1. Using the above bound in Equation 5, we get ", + "bbox": [ + 174, + 575, + 823, + 606 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/b6b61e73493c395d11f2595904023099798dc8f8e6f87c85393db4be641d33b0.jpg", + "text": "$$\nT _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )\n$$", + "text_format": "latex", + "bbox": [ + 281, + 608, + 717, + 654 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Putting the pieces together Substituting in Equation (3), bounds $T _ { 1 }$ in Equation (8) and bound $T _ { 2 }$ in Equation (4), we obtain ", + "bbox": [ + 173, + 664, + 825, + 694 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/01dbf84a6063916b2196bacb1b99de374543c8715cdbd8b7b93761144da7dae5.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \\le f ( x _ { t } ) + \\eta \\times \\left[ - \\frac { \\eta _ { l } K } { 4 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad + \\eta \\times \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\displaystyle \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] + \\frac { \\eta ^ { 2 } } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } L \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\displaystyle \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 194, + 695, + 802, + 792 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Rearranging the above inequality and summing it from $t = 0$ to $T - 1$ , we get ", + "bbox": [ + 173, + 794, + 687, + 810 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/a89d28cfd4b662cd940c97da3b5e2ab805c88b8444fe6b6dac9c9f6c5041205e.jpg", + "text": "$$\n\\begin{array} { r } { \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\leq \\frac { f ( x _ { 0 } ) - \\mathbb { E } [ f ( x _ { T } ) ] } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\\\ { + \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\left( \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } + \\frac { \\eta L } { 2 } \\right) \\times \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 212, + 830, + 787, + 922 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The first inequality uses simple telescoping sum. For completing the proof, we need the following result. ", + "bbox": [ + 171, + 102, + 825, + 133 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Lemma 4. The following upper bound holds for Algorithm 2 (FEDADAGRAD): ", + "bbox": [ + 171, + 141, + 696, + 159 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/4e1147b2e7475f6e8e8fce3208937bbce8cda5b7b739aab827184761c2c3094c.jpg", + "text": "$$\n\\begin{array} { r l } & { \\underset { t = 0 } { \\overset { T \\mathrm { - } 1 } { \\sum } } \\underset { j = 1 } { \\overset { d } { \\sum } } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\leq \\Bigg [ \\operatorname* { m i n } \\left\\{ d + \\underset { j = 1 } { \\overset { d } { \\sum } } \\log \\left( 1 + \\frac { \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) \\right. } \\\\ & { \\qquad + \\left. \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\underset { j = 1 } { \\overset { d } { \\sum } } \\sigma _ { l , j } ^ { 2 } + 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\mathbb { E } \\underset { j = 1 } { \\overset { d } { \\sum } } \\frac { ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } { \\tau ^ { 2 } } + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 2 } L ^ { 2 } } { \\tau ^ { 2 } } \\underset { t = 0 } { \\overset { T \\mathrm { - } 1 } { \\sum } } \\mathbb { E } \\left[ \\left\\| \\nabla f ( x _ { t } ) \\right\\| \\right. \\right. } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 165, + 839, + 258 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Proof. We bound the desired quantity in the following manner: ", + "bbox": [ + 173, + 290, + 589, + 305 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/f830567f03181e138bb5c7ca6e8d097a5604bdc2e90fba0c4ce861105aa65d08.jpg", + "text": "$$\n\\begin{array} { r } { \\stackrel { T \\to 1 } { \\sum } \\displaystyle \\sum _ { j = 0 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\leq d + \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } \\log \\left( 1 + \\frac { \\sum _ { l = 0 } ^ { T - 1 } \\Delta _ { l , j } ^ { 2 } } { \\tau ^ { 2 } } \\right) \\leq d + \\displaystyle \\sum _ { j = 1 } ^ { d } \\log \\left( 1 + \\frac { \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 313, + 834, + 358 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "An alternate way of the bounding this quantity is as follows: ", + "bbox": [ + 173, + 364, + 570, + 381 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/f221e4673ceef98f84efd7e8812ee80530c22604df75c2c5ec4970ef2546c0d4.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { t , j } ^ { 2 } + \\tau ^ { 2 } } \\leq \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\tau ^ { 2 } } } \\\\ & { \\leq \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\left\\| \\frac { \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) - \\eta _ { l } K \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } } \\\\ & { \\leq 2 \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\left[ \\left\\| \\frac { \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } + \\eta _ { l } ^ { 2 } K ^ { 2 } \\left\\| \\frac { \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 302, + 387, + 694, + 523 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The first quantity in the above bound can be further bounded as follows: ", + "bbox": [ + 173, + 529, + 645, + 544 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/3bf3db4fce63d62f2c4ad22cd269923cf76161218d94f6b6c6a36975222e5d97.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda - 1 } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( - \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } - \\lambda _ { i } ^ { \\lambda } \\sum _ { j = 1 } ^ { \\infty } - \\lambda _ { i } ^ { \\lambda } \\nabla F _ { i } ( \\xi _ { i } , \\xi _ { i } ) \\Bigg ) \\Bigg | ^ { 2 } } \\\\ & { - 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } \\sum _ { j = 1 } ^ { \\infty } \\lambda _ { i } ^ { \\lambda } \\sum _ { j = 1 } ^ { \\lambda } - \\lambda _ { i } ^ { \\lambda } \\nabla F _ { i } ( \\xi _ { i } , \\xi _ { i } ) + 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\lambda _ { i } ^ { \\lambda } \\xi _ { i } ^ { \\lambda } \\xi _ { i } ^ { \\lambda } \\xi _ { j } - 3 \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\Bigg ) - } \\\\ & - 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\mathbb { E } ^ { \\lambda - 1 } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } \\lambda _ { i } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb E \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 551, + 839, + 892 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Here, the first inequality follows from simple application of the fact that $a b \\le ( a ^ { 2 } + b ^ { 2 } ) / 2$ . The result follows. □ ", + "bbox": [ + 174, + 895, + 826, + 924 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Substituting the above bound in Equation (10), we obtain: ", + "bbox": [ + 173, + 103, + 553, + 118 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/b9883d2eb2dc53244494a52fbe203284635df98d004ce0d80f6c1d8722c8ec7d.jpg", + "text": "$$\n\\begin{array} { r l } & { \\frac { \\displaystyle { n K } } { \\displaystyle { 4 } } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\pi _ { t - 1 , j } } + \\tau } } \\\\ & { \\leq \\frac { f \\left( x _ { 0 } \\right) - \\mathbb { E } \\left[ f ( x _ { T } ) \\right] } { \\eta } + \\frac { 5 \\eta _ { t } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\\\ & { \\qquad + \\displaystyle { \\sum _ { j = 1 } ^ { d } \\left( \\frac { \\eta _ { t } K G ^ { 2 } } { \\tau } + \\frac { \\eta L } { 2 } \\right) \\times \\left[ \\operatorname* { m i n } \\left\\{ d + d \\log \\left( 1 + \\frac { \\eta _ { t } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) + \\right. \\right.} } \\\\ & { \\quad \\quad \\left. \\frac { 4 \\eta _ { t } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + 2 0 \\eta _ { t } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\sum _ { j = 1 } ^ { d } \\frac { ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } { \\tau ^ { 2 } } + \\frac { 4 0 \\eta _ { t } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } \\left[ \\left\\| \\nabla f ( x _ { t } ) \\right\\| ^ { 2 } \\right] \\right\\} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 122, + 839, + 306 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "We observe the following: ", + "bbox": [ + 174, + 330, + 348, + 344 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/32ab3f3d2fdf07a6c34f4f486222c9477f6fadd36c17f09624fb4b6f3a41c7b6.jpg", + "text": "$$\n\\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla \\mathbb { E } f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\geq \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\eta _ { t } K G \\sqrt { T } + \\tau } \\geq \\frac { T } { \\eta _ { t } K G \\sqrt { T } + \\tau } \\operatorname* { m i n } _ { 0 \\leq t \\leq T } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 200, + 348, + 794, + 393 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "The second part of Theorem 1 follows from using the above inequality in Equation (12). Note that the first part of Theorem 1 is obtain from the part of Lemma 4. □ ", + "bbox": [ + 173, + 396, + 825, + 425 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "A.2.1 LIMITED PARTICIPATION ", + "text_level": 1, + "bbox": [ + 176, + 439, + 403, + 454 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "For limited participation, the main changed in the proof is in Equation (11). The rest of the proof is similar so we mainly focus on Equation (11) here. Let $s$ be the sampled set at the $t ^ { \\mathrm { { t h } } }$ iteration such that $| S | = s$ . In partial participation, we assume that the set $s$ is sampled uniformly from all subsets of $[ m ]$ with size $s$ . In this case, for the first term in Equation (11), we have ", + "bbox": [ + 173, + 463, + 825, + 518 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/b799b7729cf8d3e983e7798f43c9e68895804b5af4937a7a71e8edf4df8dbd51.jpg", + "text": "$$\n\\begin{array} { r l } { \\sum _ { j = 1 } ^ { N } \\Bigg [ \\Bigg | \\int _ { - 1 } ^ { 1 } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 0 } ^ { 1 } \\sum _ { i = 0 } ^ { N } \\sum _ { j = 1 } ^ { N } \\Bigg ) \\Bigg | ^ { 2 } \\Bigg ] ^ { 2 } } & { \\mathrm { d } \\Gamma } \\\\ { = } & { \\mathcal { D } _ { j } ^ { \\mathrm { R } } \\Bigg [ \\Bigg | \\int _ { - 1 } ^ { 1 } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 0 } ^ { 1 } \\sum _ { i = 0 } ^ { N } \\sum _ { j = 1 } ^ { N } \\delta ( \\sum _ { i = 0 } ^ { N } \\epsilon _ { j i } ) + \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\bigg ) \\Bigg | ^ { 2 } } \\\\ & { \\leq \\epsilon _ { j } ^ { \\mathrm { R } } \\Bigg ( \\int _ { - 1 } ^ { 1 } \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) ^ { 2 } \\Bigg [ \\Bigg | \\int _ { 0 } ^ { 1 } \\sum _ { j = 1 } ^ { N } \\delta ( \\sum _ { j = 1 } ^ { N } \\epsilon _ { j i } ) - \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\Bigg | ^ { 2 } \\Bigg ] ^ { 2 } + \\Bigg | \\int _ { - 1 } ^ { 1 } \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) - \\mathrm { v } _ { \\mathrm { R } } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) \\Bigg | ^ { 2 } } \\\\ & \\qquad + \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\Bigg | ^ { 2 } \\Bigg ] + \\Bigg | \\int _ { 0 } ^ { 1 } \\int _ { - 1 } ^ { 1 } \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 1 } ^ { N } \\mathrm { d } \\ \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 523, + 839, + 882 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Note that the expectation here also includes $s$ . The first inequality is obtained from the fact that $( a + b + c ) ^ { 2 } \\leq 3 { \\dot { ( } } a ^ { 2 } + b ^ { 2 } + c ^ { 2 } )$ . The second inequality is obtained from the fact that set $s$ is uniformly ", + "bbox": [ + 171, + 895, + 823, + 924 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "sampled from all subsets of $[ m ]$ with size $s$ . The third and fourth inequalities are similar to the one used in proof of Theorem 1. Substituting the above bound in Equation (10) gives the desired convergence rate. ", + "bbox": [ + 174, + 102, + 825, + 146 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "A.3 PROOF OF THEOREM 2 ", + "text_level": 1, + "bbox": [ + 176, + 169, + 377, + 184 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Proof of Theorem 2. The proof strategy is similar to that of FEDADAGRAD except that we need to handle the exponential moving average in FEDADAM. We note that the update of FEDADAM is the following ", + "bbox": [ + 174, + 198, + 825, + 241 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/4dfc7f601b27403d8437017dac13006da4bff9de10fc9f0683eb9bea340e8c9d.jpg", + "text": "$$\nx _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } ,\n$$", + "text_format": "latex", + "bbox": [ + 416, + 247, + 581, + 284 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "for all $i \\in [ d ]$ . Using the $L$ -smooth nature of function $f$ and the above update rule, we have the following: ", + "bbox": [ + 173, + 295, + 825, + 325 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/2e104f1e0635f188755318d4665c1d9be0fa34468dd053e57fbec0d4ba2f6d96.jpg", + "text": "$$\nf ( \\boldsymbol { x } _ { t + 1 } ) \\leq f ( \\boldsymbol { x } _ { t } ) + \\eta \\left. \\nabla f ( \\boldsymbol { x } _ { t } ) , \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\sum _ { i = 1 } ^ { d } \\frac { \\Delta _ { t , i } ^ { 2 } } { ( \\sqrt { v _ { t , i } } + \\tau ) ^ { 2 } }\n$$", + "text_format": "latex", + "bbox": [ + 267, + 338, + 728, + 382 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "The second step follows simply from FEDADAM’s update. We take the expectation of $f ( x _ { t + 1 } )$ (over randomness at time step $t$ ) and rewrite the above inequality as: ", + "bbox": [ + 174, + 395, + 823, + 424 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/24d642061b4389eae21ff08c5e8289d958fc3329cfb685492876b7902596d324.jpg", + "text": "$$\n\\begin{array} { r } { \\bar { c } _ { t } [ f ( x _ { t + 1 } ) ] \\leq f ( x _ { t } ) + \\eta \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } + \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } } \\\\ { = f ( x _ { t } ) + \\eta \\underbrace { \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\right. } _ { R _ { 1 } } + \\eta \\underbrace { \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } } \\right. \\right. } _ { R _ { 2 } } } \\\\ { \\left. \\left. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( \\sqrt { v _ { t , j } } + \\tau ) ^ { 2 } } \\right] \\right. \\right. } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 435, + 839, + 592 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Bounding $R _ { 2 }$ . We observe the following about $R _ { 2 }$ ", + "bbox": [ + 173, + 625, + 524, + 640 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/14caa1b90349bf8677c43b103ed85506a24aeb712af8d7eb4efb7b30522c60f0.jpg", + "text": "$$\n\\begin{array} { r l } & { H _ { 2 } = \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\times \\left[ \\frac { \\Delta _ { t , x } } { \\sqrt { \\nabla f _ { \\mathbf { S } } \\phi _ { t + 1 } } } - \\frac { \\Delta _ { t , x } } { \\sqrt { \\nabla f _ { \\mathbf { S } } \\phi _ { t - 1 , \\Delta } } + \\tau } \\right] } \\\\ & { \\quad = \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\Delta _ { t , x } \\simeq \\Delta _ { t , \\Phi } \\times \\left[ \\frac { \\sqrt { \\omega _ { 2 } \\phi _ { 1 } } \\mathbf { e } _ { 1 } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } } } { \\left( \\sqrt { \\omega _ { 2 } } \\omega _ { 1 } \\right) \\tau \\left( \\sqrt { \\hat { \\omega } _ { 2 } \\omega _ { 1 } } - \\Delta _ { t , 1 } \\right) \\tau } \\right] } \\\\ & { \\quad \\quad - \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\times \\Delta _ { t , \\Phi } \\times \\left[ \\frac { - ( 1 - \\omega _ { 2 } ) \\Delta _ { t , x } ^ { 2 } } { \\left( \\sqrt { \\omega _ { 2 } } \\omega _ { 1 } + \\gamma \\right)} \\left( \\sqrt { \\hat { \\omega } _ { 2 } \\omega _ { 1 } } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } } } \\right) + \\sqrt { \\mathbf { e } _ { 1 } \\omega _ { 2 } } \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & \\quad \\quad \\quad \\quad \\quad \\quad \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 163, + 651, + 816, + 934 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Bounding $R _ { 1 }$ . The term $R _ { 1 }$ can be bounded as follows: ", + "bbox": [ + 173, + 102, + 553, + 118 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/79285d7c20a293db19d90a840732db61a34f5a16ef6a2e0efbaaa1f9d4430e1a.jpg", + "text": "$$\n\\begin{array} { r l } & { R _ { 1 } = \\Bigg \\langle \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\Bigg \\rangle } \\\\ & { \\quad = \\Bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } - \\eta _ { t } K \\nabla f ( x _ { t } ) + \\eta _ { t } K \\nabla f ( x _ { t } ) \\right] \\Bigg \\rangle } \\\\ & { \\quad = - \\eta _ { t } K \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\underbrace { \\Bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { t } K \\nabla f ( x _ { t } ) \\right] \\Bigg \\rangle } _ { R _ { 3 } } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 235, + 119, + 763, + 268 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Bounding $R _ { 3 }$ . The term $R _ { 3 }$ can be bounded in exactly the same way as term $T _ { 3 }$ in proof of Theorem 1: ", + "bbox": [ + 171, + 277, + 825, + 308 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/1529630d222e38076b1968b366a7864e920b2807055a709cea39b00619c49e24.jpg", + "text": "$$\nR _ { 3 } \\leq \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right]\n$$", + "text_format": "latex", + "bbox": [ + 274, + 306, + 723, + 352 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Substituting the above inequality in Equation (15), we get ", + "bbox": [ + 174, + 354, + 555, + 369 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/2cb5efae5bcb75510b7bf1a2c4b6baed75f041a6e9b59b6f0e34b56452914193.jpg", + "text": "$$\nR _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right]\n$$", + "text_format": "latex", + "bbox": [ + 267, + 372, + 728, + 416 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Here we used the fact that $\\sqrt { v _ { t - 1 , j } } \\leq \\eta _ { l } K G$ and conditions in Theorem 2. Using Lemma 3, we obtain the following bound on $R _ { 1 }$ : ", + "bbox": [ + 176, + 420, + 823, + 449 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/fb802f805c675d6605c42ec4824c5cae647ce1d82396839cbfac9945fa900004.jpg", + "text": "$$\nR _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )\n$$", + "text_format": "latex", + "bbox": [ + 266, + 450, + 732, + 497 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Putting pieces together. Substituting bounds $R _ { 1 }$ and $R _ { 2 }$ in Equation (14), we have ", + "bbox": [ + 174, + 506, + 732, + 522 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/581c12f407c6907c3db8057492fd35368a978c92b9872459ee4463a4f20962ef.jpg", + "text": "$$\n\\begin{array} { r l } { { \\mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \\leq f ( x _ { t } ) - \\frac { \\eta \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } } v _ { t - 1 , j } } + \\tau + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { + ( \\frac { \\eta \\sqrt { 1 - \\beta _ { 2 } } G } { \\tau } ) \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sqrt { v _ { t , j } } + \\tau ) } ] + ( \\frac { \\eta ^ { 2 } L } { 2 } ) \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { v _ { t , j } + \\tau ^ { 2 } ) } ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 525, + 816, + 617 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Summing over $t = 0$ to $T - 1$ and using telescoping sum, we have ", + "bbox": [ + 173, + 617, + 611, + 632 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/2bb1c75401afd25f833a08734903cd0b4c8abefd23fb829ec05ba8702971b25a.jpg", + "text": "$$\n\\mathbb { E } [ f ( x _ { T } ) ] \\leq f ( x _ { 0 } ) - \\frac { \\eta \\eta _ { l } K } { 4 } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )\n$$", + "text_format": "latex", + "bbox": [ + 181, + 635, + 823, + 728 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "To bound this term further, we need the following result. ", + "bbox": [ + 171, + 743, + 547, + 758 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Lemma 5. The following upper bound holds for Algorithm 2 (FEDADAM): ", + "bbox": [ + 169, + 765, + 669, + 780 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/e2983ee88e0f3318e98488d18def796751a244c1b770753a2331705158205c09.jpg", + "text": "$$\n\\sum _ { = 0 } ^ { { { \\nu - 1 } } } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\right] \\leq \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 1 } ^ { T - 1 } \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { \\ell = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) }\n$$", + "text_format": "latex", + "bbox": [ + 181, + 782, + 839, + 828 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Proof. ", + "bbox": [ + 173, + 840, + 217, + 854 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/23d9eee56aad3aa19eeedd3480a4fa77878237aa9a04b8b242f952d518d4db2f.jpg", + "text": "$$\n\\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { l = 0 } ^ { t } \\beta _ { 2 } ^ { t - l } \\Delta _ { t , j } ^ { 2 } + \\tau ^ { 2 } } \\le \\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\tau ^ { 2 } }\n$$", + "text_format": "latex", + "bbox": [ + 299, + 858, + 696, + 904 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "The rest of the proof follows along the lines of proof of Lemma 4. Using the same argument, we get ", + "bbox": [ + 169, + 102, + 825, + 119 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/8116e20a8628c1291d8690901c627f03535e92acf917808d74434203d7b5dc89.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\sum _ { j = 1 } ^ { d } \\mathbb { E } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } ] \\leq \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { } & { + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } [ \\nabla f ( x _ { t } ) ^ { 2 } ] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 236, + 125, + 761, + 215 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "which is the desired result. ", + "bbox": [ + 174, + 219, + 349, + 234 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Substituting the bound obtained from above lemma in Equation (17) and using a similar argument for bounding ", + "bbox": [ + 174, + 248, + 828, + 276 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/db2a2e7a966fb340bee0a14d463b9efdb27af424086a3216d4baffdefcab5eee.jpg", + "text": "$$\n\\left( \\frac { \\eta \\sqrt { 1 - \\beta _ { 2 } } G } { \\tau } \\right) \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sqrt { v _ { t , j } } + \\tau ) } \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 356, + 275, + 640, + 320 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "we obtain ", + "bbox": [ + 173, + 321, + 240, + 337 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/4a9bc09902b93927839786af52a1d3af237b4586b75d29799085b5989a6fd191.jpg", + "text": "$$\n\\begin{array} { c } { { \\displaystyle z _ { t } [ f ( x _ { T } ) ] \\leq f ( x _ { 0 } ) - \\frac { \\eta \\eta _ { l } K } { 8 } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ { { + \\left( \\eta \\sqrt { 1 - \\beta _ { 2 } } G + \\frac { \\eta ^ { 2 } L } { 2 } \\right) \\times \\left[ \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) \\right] \\sigma _ { g , l } ^ { 2 } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 340, + 839, + 438 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "The above inequality is obtained due to the fact that: ", + "bbox": [ + 174, + 441, + 516, + 457 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/a4a0191d5000373a0b05493f25fbea3ac0f36d43b0d7010ecb4316f3e3f6dc37.jpg", + "text": "$$\n\\left( \\sqrt { 1 - \\beta _ { 2 } } G + \\frac { \\eta L } { 2 } \\right) \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\leq \\frac { \\eta _ { l } K } { 1 6 } \\left( \\frac { 1 } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G } + \\frac { 1 } { \\tau } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 289, + 462, + 709, + 498 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "The above condition follows from the condition on $\\eta _ { l }$ in Theorem 2. We also observe the following: ", + "bbox": [ + 171, + 502, + 826, + 518 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/a68d4e1375af9101dde5e6d6ff208aecccefacf479dd4df606fdc277f5489324.jpg", + "text": "$$\n\\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } \\geq \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } \\geq \\frac { T } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } \\operatorname* { m i n } _ { \\eta \\leq t \\leq T } \\Vert \\nabla f ( x _ { t } ) \\Vert ^ { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 197, + 523, + 799, + 569 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Substituting this bound in the above equation yields the desired result. ", + "bbox": [ + 173, + 574, + 633, + 590 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "A.4 AUXILIARY LEMMATTA ", + "text_level": 1, + "bbox": [ + 174, + 606, + 383, + 621 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Lemma 6. For random variables $z _ { 1 } , \\ldots , z _ { r }$ , we have ", + "bbox": [ + 171, + 631, + 531, + 647 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/a99b2ee87abfad1cdb2125a638328910b46443d8a5b875a8ffa4f54104940353.jpg", + "text": "$$\n\\begin{array} { r } { \\mathbb { E } \\left[ \\| z _ { 1 } + . . . + z _ { r } \\| ^ { 2 } \\right] \\leq r \\mathbb { E } \\left[ \\| z _ { 1 } \\| ^ { 2 } + . . . + \\| z _ { r } \\| ^ { 2 } \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 331, + 651, + 665, + 671 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Lemma 7. For independent, mean $O$ random variables $z _ { 1 } , \\ldots , z _ { r }$ , we have ", + "bbox": [ + 174, + 676, + 666, + 691 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/12cabfb3dc55a2bd1c78426ee8b6886b880e7f05856a147bd15b538b6c95e27e.jpg", + "text": "$$\n\\begin{array} { r } { \\mathbb { E } \\left[ \\| z _ { 1 } + . . . + z _ { r } \\| ^ { 2 } \\right] = \\mathbb { E } \\left[ \\| z _ { 1 } \\| ^ { 2 } + . . . + \\| z _ { r } \\| ^ { 2 } \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 334, + 695, + 660, + 715 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "B FEDERATED ALGORITHMS: IMPLEMENTATIONS AND PRACTICAL CONSIDERATIONS ", + "text_level": 1, + "bbox": [ + 176, + 732, + 745, + 766 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "B.1 FEDAVG AND FEDOPT ", + "text_level": 1, + "bbox": [ + 174, + 780, + 374, + 796 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "In Algorithm 3, we give a simplified version of the FEDAVG algorithm by McMahan et al. (2017), that applies to the setup given in Section 2. We write $\\mathbf { S } \\mathbf { G } \\mathbf { D } _ { K } ( x _ { t } , \\eta _ { l } , f _ { i } )$ to denote $K$ steps of SGD using gradients $\\nabla f _ { i } ( x , \\bar { z } )$ for $z \\sim D _ { i }$ with (local) learning rate $\\eta _ { l }$ , starting from $x _ { t }$ . As noted in Section 2, Algorithm 3 is the special case of Algorithm 1 where CLIENTOPT is SGD, and SERVEROPT is SGD with learning rate 1. ", + "bbox": [ + 174, + 806, + 825, + 876 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "While Algorithms 1, 2, and 3 are useful for understanding relations between federated optimization methods, we are also interested in practical versions of these algorithms. In particular, Algorithms 1, 2, and 3 are stated in terms of a kind of ‘gradient oracle’, where we compute unbiased estimates ", + "bbox": [ + 174, + 882, + 825, + 924 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Algorithm 3 Simplified FEDAVG ", + "text_level": 1, + "bbox": [ + 174, + 103, + 397, + 118 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/aefb27f699f699b25c2789d831eb4f00c476aa4f78358921e8bac4a2bc3ca6e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input: xo fort=0,.,T-1do
Samplea subsetSofclients
xt=xt
for each client i ∈ Sinparallel do
xt = SGDk(xt, nt, fi) for i∈ S (in parallel)
xt+1=∑iesxt
", + "bbox": [ + 183, + 121, + 534, + 228 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "of the client’s gradient. In practical scenarios, we often only have access to finite data samples, the number of which may vary between clients. ", + "bbox": [ + 173, + 257, + 821, + 285 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Instead, we assume that in (1), each client distribution $\\mathcal { D } _ { i }$ is the uniform distribution over some finite set $D _ { i }$ of size $n _ { i }$ . The $n _ { i }$ may vary significantly between clients, requiring extra care when implementing federated optimization methods. We assume the set $D _ { i }$ is partitioned into a collection of batches $B _ { i }$ , each of size $B$ . For $b \\in B _ { i }$ , we let $f _ { i } ( x ; b )$ denote the average loss on this batch at $x$ with corresponding gradient $\\nabla f _ { i } ( x ; b )$ . Thus, if $b$ is sampled uniformly at random from $B _ { i }$ , $\\nabla f _ { i } ( x ; b )$ is an unbiased estimate of $\\nabla F _ { i } ( x )$ . ", + "bbox": [ + 173, + 291, + 825, + 375 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "When training, instead of uniformly using $K$ gradient steps, as in Algorithm 1, we will instead perform $E$ epochs of training over each client’s dataset. Additionally, we will take a weighted average of the client updates, where we weight according to the number of examples $n _ { i }$ in each client’s dataset. This leads to a batched data version of FEDOPT in Algorithm 4, and a batched data version of FEDADAGRAD, FEDADAM, and FEDYOGI given in Algorithm 5. ", + "bbox": [ + 173, + 380, + 825, + 449 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Algorithm 4 FEDOPT - Batched data ", + "text_level": 1, + "bbox": [ + 174, + 467, + 419, + 482 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Input: $x _ { 0 }$ , CLIENTOPT, SERVEROPT \nfor $t = 0 , \\cdots , T - 1$ do Sample a subset $s$ of clients $x _ { i } ^ { t } = x _ { t }$ for each client $i \\in S$ in parallel do for $e = 1 , \\ldots , E$ do for $b \\in B _ { i }$ do $\\begin{array} { c } { \\mathsf { \\Gamma } _ { i } \\gets \\mathsf { \\Gamma } _ { i } \\gets } \\\\ { g _ { i } ^ { t } = \\nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ { x _ { i } ^ { t } = \\mathrm { C } _ { \\mathrm { L I E N T O P T } } ( x _ { i } ^ { t } , g _ { i } ^ { t } , \\eta _ { l } , t ) } \\\\ { \\Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\\\ { n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , \\Delta _ { t } = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\\\ { x _ { t + 1 } = \\mathrm { S E R V E R O P T } ( x _ { t } , - \\Delta _ { t } , \\eta , t ) } \\end{array}$ ", + "bbox": [ + 181, + 484, + 493, + 660 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "In Section 5, we use Algorithm 4 when implementing FEDAVG and FEDAVGM. In particular, FEDAVG and FEDAVGM correspond to Algorithm 4 where CLIENTOPT and SERVEROPT are SGD. FEDAVG uses vanilla SGD on the server, while FEDAVGM uses SGD with a momentum parameter of 0.9. In both methods, we tune both client learning rate $\\eta _ { l }$ and server learning rate $\\eta$ . This means that the version of FEDAVG used in all experiments is strictly more general than that in (McMahan et al., 2017), which corresponds to FEDOPT where CLIENTOPT and SERVEROPT are SGD, and SERVEROPT has a learning rate of 1. ", + "bbox": [ + 173, + 678, + 825, + 773 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "We use Algorithm 5 for all implementations FEDADAGRAD, FEDADAM, and FEDYOGI in Section 5. For FEDADAGRAD, we set $\\beta _ { 1 } = \\beta _ { 2 } = 0$ (as typical versions of ADAGRAD do not use momentum). For FEDADAM and FEDYOGI we set $\\beta _ { 1 } = { \\bar { 0 . 9 } } , \\beta _ { 2 } = 0 . 9 9$ . While these parameters are generally good choices (Zaheer et al., 2018), we emphasize that better results may be obtainable by tuning these parameters. ", + "bbox": [ + 174, + 780, + 825, + 849 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "B.2 SCAFFOLD ", + "text_level": 1, + "bbox": [ + 174, + 868, + 307, + 883 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "As discussed in Section 5, we compare all five optimizers above to SCAFFOLD (Karimireddy et al., 2019) on our various tasks. There are a few important notes about the validity of this comparison. ", + "bbox": [ + 174, + 895, + 825, + 924 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Algorithm 5 FEDADAGRAD FEDYOGI , and FEDADAM - Batched data ", + "text_level": 1, + "bbox": [ + 176, + 104, + 676, + 122 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Input: $x _ { 0 } , v _ { - 1 } \\ge \\tau ^ { 2 }$ , optional $\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )$ for FEDYOGI and FEDADAM \nfor $t = 0 , \\cdots , T - 1$ do Sample a subset $s$ of clients $x _ { i } ^ { t } = x _ { t }$ for each client $i \\in S$ in parallel do for $e = 1 , \\ldots , E$ do for $b \\in B _ { i }$ do $\\begin{array} { r l } & { ~ x _ { i } ^ { t } = \\check { x } _ { i } ^ { t } - \\eta _ { l } \\nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ & { ~ \\Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\\\ & { n = \\sum _ { i \\in S } n _ { i } , ~ \\Delta _ { t } = \\sum _ { i \\in S } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\\\ & { \\Delta _ { t } = \\beta _ { 1 } \\Delta _ { t - 1 } + ( 1 - \\beta _ { 1 } ) \\Delta _ { t } } \\end{array}$ $v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }$ (FEDADAGRAD) $v _ { t } = v _ { t - 1 } - ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\mathrm { s i g n } ( v _ { t - 1 } - \\Delta _ { t } ^ { 2 } ) \\left( \\mathbf { F E D Y O G I } \\right)$ vt = β2vt−1 + (1 − β2)∆2t (FEDADAM) xt+1 = xt + η √∆tvt+τ ", + "bbox": [ + 178, + 127, + 687, + 373 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "1. In cross-device settings, this is not a fair comparison. In particular, SCAFFOLD does not work in settings where clients cannot maintain state across rounds, as may be the case for federated learning systems on edge devices, such as cell phones. ", + "bbox": [ + 212, + 400, + 825, + 441 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "2. SCAFFOLD has two variants described by Karimireddy et al. (2019). In Option I, the control variate of a client is updated using a full gradient computation. This effectively requires performing an extra pass over each client’s dataset, as compared to Algorithm 1. In order to normalize the amount of client work, we instead use Option II, in which the clients’ control variates are updated using the difference between the server model and the client’s learned model. This requires the same amount of client work as FEDAVG and Algorithm 2. ", + "bbox": [ + 212, + 445, + 825, + 527 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "For practical reasons, we implement a version of SCAFFOLD mirroring Algorithm 4, in which we perform $E$ epochs over the client’s dataset, and perform weighted averaging of client models. For posterity, we give the full pseudo-code of the version of SCAFFOLD used in our experiments in Algorithm 6. This is a simple adaptiation of Option II of the SCAFFOLD algorithm in (Karimireddy et al., 2019) to the same setting as Algorithm 4. In particular, we let $n _ { i }$ denote the number of examples in client $i$ ’s local dataset. ", + "bbox": [ + 173, + 539, + 825, + 621 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Algorithm 6 SCAFFOLD SCAFFOLD, Option II - Batched data ", + "text_level": 1, + "bbox": [ + 173, + 633, + 521, + 650 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Input: $x _ { 0 } , c , \\eta _ { l } , \\eta$ \nfor $t = 0 , \\cdots , T - 1 \\cdot$ do Sample a subset $s$ of clients $x _ { i } ^ { t } = x _ { t }$ for each client $i \\in S$ in parallel do for $e = 1 , \\ldots , E$ do for $b \\in B _ { i }$ do $\\begin{array} { r l } & { \\quad g _ { i } ^ { t } = \\breve { \\nabla } f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ & { \\quad \\quad x _ { i } ^ { t } = x _ { i } ^ { t } - \\eta _ { l } ( g _ { i } ^ { t } - c _ { i } + c ) } \\\\ & { \\quad c _ { i } ^ { + } = c _ { i } - c + ( E | \\mathcal { B } _ { i } | \\eta _ { l } ) ^ { - 1 } ( x _ { i } ^ { t } - x _ { i } ) } \\\\ & { \\quad \\Delta x _ { i } = x _ { i } ^ { t } - x _ { t } , \\Delta c _ { i } = c _ { i } ^ { + } - c _ { i } } \\\\ & { \\quad c _ { i } = c _ { i } ^ { + } } \\\\ & { \\quad n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , \\Delta x = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta x _ { i } , \\Delta c = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta c _ { i } } \\\\ & { \\quad x _ { t + 1 } = x _ { t } + \\eta \\Delta x , c = c + \\frac { | \\mathcal { S } | } { N } \\Delta c } \\end{array}$ ", + "bbox": [ + 176, + 655, + 601, + 863 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "In this algorithm, $c _ { i }$ is the control variate of client $i$ , and $c$ is the running average of these control variates. In practice, we must initialize the control variates $c _ { i }$ in some way when sampling a client $i$ for the first time. In our implementation, we set $c _ { i } = c$ the first time we sample a client $i$ . This has the advantage of exactly recovering FEDAVG when each client is sampled at most once. To initialize $c$ , we use the all zeros vector. ", + "bbox": [ + 174, + 882, + 825, + 924 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "We compare this version of SCAFFOLD to FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG on our tasks, tuning the learning rates in the same way (using the same grids as in Appendix D.2). In particular, $\\eta _ { l } , \\eta$ are tuned to obtain the best training performance over the last 100 communication rounds. We use the same federated hyperparameters for SCAFFOLD as discussed in Section 4. Namely, we set $E = 1$ , and sample 10 clients per round for all tasks except Stack Overflow NWP, where we sample 50. The results are given in Figure 1 in Section 5. ", + "bbox": [ + 174, + 137, + 825, + 220 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "B.3 LOOKAHEAD, ADAALTER, AND CLIENT ADAPTIVITY ", + "text_level": 1, + "bbox": [ + 178, + 242, + 578, + 256 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "The LOOKAHEAD optimizer (Zhang et al., 2019b) is primarily designed for non-FL settings. LOOKAHEAD uses a generic optimizer in the inner loop and updates its parameters using a “outer” learning rate. Thus, unlike FEDOPT, LOOKAHEAD uses a single generic optimizer and is thus conceptually different. In fact, LOOKAHEAD can be seen as a special case of FEDOPT in non-FL settings which uses a generic optimizer CLIENTOPT as a client optimizer, and SGD as the server optimizer. While there are multiple ways LOOKAHEAD could be generalized to a federated setting, one straightforward version would simply use an adaptive method as the CLIENTOPT. On the other hand, ADAALTER (Xie et al., 2019) is designed specifically for distributed settings. In ADAALTER, clients use a local optimizer similar to ADAGRAD (McMahan & Streeter, 2010a; Duchi et al., 2011) to perform multiple epochs of training on their local datasets. Both LOOKAHEAD and ADAALTER use client adaptivity, which is fundamentally different from the adaptive server optimizers proposed in Algorithm 2. ", + "bbox": [ + 174, + 268, + 825, + 419 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "To illustrate the differences, consider the client-to-server communication in ADAALTER. This requires communicating both the model weights and the client accumulators (used to perform the adaptive optimization, analogous to $v _ { t }$ in Algorithm 2). In the case of ADAALTER, the client accumulator is the same size as the model’s trainable weights. Thus, the client-to-server communication doubles for this method, relative to FEDAVG. In ADAALTER, the server averages the client accumulators and broadcasts the average to the next set of clients, who use this to initialize their adaptive optimizers. This means that the server-to-client communication also doubles relative to FEDAVG. The same would occur for any adaptive client optimizer in the distributed version of LOOKAHEAD described above. For similar reasons, we see that client adaptive methods also increase the amount of memory needed on the client (as they must store the current accumulator). By contrast, our adaptive server methods (Algorithm 2) do not require extra communication or client memory relative to FEDAVG. Thus, we see that server-side adaptive optimization benefits from lower per-round communication and client memory requirements, which are of paramount importance for FL applications (Bonawitz et al., 2019). ", + "bbox": [ + 174, + 425, + 825, + 616 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "C DATASET & MODELS ", + "text_level": 1, + "bbox": [ + 176, + 640, + 385, + 656 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "Here we provide detailed description of the datasets and models used in the paper. We use federated versions of vision datasets EMNIST (Cohen et al., 2017), CIFAR-10 (Krizhevsky & Hinton, 2009), and CIFAR-100 (Krizhevsky & Hinton, 2009), and language modeling datasets Shakespeare (McMahan et al., 2017) and StackOverflow (Authors, 2019). Statistics for the training datasets can be found in Table 2. We give descriptions of the datasets, models, and tasks below. Statistics on the number of clients and examples in both the training and test splits of the datasets are given in Table 2. ", + "bbox": [ + 174, + 672, + 825, + 756 + ], + "page_idx": 23 + }, + { + "type": "table", + "img_path": "images/64c387ac10e2d1ad5cbf52d2f118f0817fd2d7cb047ee8a592d862fcb7e0acf0.jpg", + "table_caption": [ + "Table 2: Dataset statistics. " + ], + "table_footnote": [], + "table_body": "
DATASETTRAIN CLIENTSTRAIN EXAMPLESTEST CLIENTSTEST EXAMPLES
CIFAR-1050050,00010010,000
CIFAR-10050050,00010010,000
EMNIST-623,400671,5853,40077,483
SHAKESPEARE71516,0687152,356
STACKOVERFLOW342,477135,818,730204,08816,586,035
", + "bbox": [ + 192, + 813, + 802, + 904 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "C.1 CIFAR-10/CIFAR-100 ", + "text_level": 1, + "bbox": [ + 174, + 103, + 382, + 117 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "We create a federated version of CIFAR-10 by randomly partitioning the training data among 500 clients, with each client receiving 100 examples. We use the same approach as Hsu et al. (2019), where we apply latent Dirichlet allocation (LDA) over the labels of CIFAR-10 to create a federated dataset. Each client has an associated multinomial distribution over labels from which its examples are drawn. The multinomial is drawn from a symmetric Dirichlet distribution with parameter 0.1. ", + "bbox": [ + 174, + 133, + 825, + 202 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "For CIFAR-100, we perform a similar partitioning of 100 examples to 500 clients, but using a more sophisticated approach. We use a two step LDA process over the coarse and fine labels. We randomly partition the data to reflect the \"coarse\" and \"fine\" label structure of CIFAR-100 by using the Pachinko Allocation Method (PAM) (Li & McCallum, 2006). This creates more realistic client datasets, whose label distributions better resemble practical heterogeneous client datasets. We have made publicly available the specific federated version of CIFAR-100 we used for all experiments, though we avoid giving a link in this work in order to avoid de-anonymization. For complete details on how the dataset was created, see Appendix F. ", + "bbox": [ + 174, + 209, + 823, + 318 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "We train a modified ResNet-18 on both datasets, where the batch normalization layers are replaced by group normalization layers (Wu & He, 2018). We use two groups in each group normalization layer. As shown by Hsieh et al. (2019), group normalization can lead to significant gains in accuracy over batch normalization in federated settings. ", + "bbox": [ + 174, + 325, + 825, + 380 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "Preprocessing CIFAR-10 and CIFAR-100 consist of images with 3 channels of $3 2 \\times 3 2$ pixels each. Each pixel is represented by an unsigned int8. As is standard with CIFAR datasets, we perform preprocessing on both training and test images. For training images, we perform a random crop to shape (24, 24, 3), followed by a random horizontal flip. For testing images, we centrally crop the image to $( 2 4 , 2 4 , 3 )$ . For both training and testing images, we then normalize the pixel values according to their mean and standard deviation. Namely, given an image $x$ , we compute $( x - \\mu ) / \\sigma$ where $\\mu$ is the average of the pixel values in $x$ , and $\\sigma$ is the standard deviation. ", + "bbox": [ + 174, + 405, + 825, + 502 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "C.2 EMNIST ", + "text_level": 1, + "bbox": [ + 174, + 529, + 282, + 544 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "EMNIST consists of images of digits and upper and lower case English characters, with 62 total classes. The federated version of EMNIST (Caldas et al., 2018) partitions the digits by their author. The dataset has natural heterogeneity stemming from the writing style of each person. We perform two distinct tasks on EMNIST, autoencoder training (EMNIST AE) and character recognition (EMNIST CR). For EMNIST AE, we train the “MNIST” autoencoder (Zaheer et al., 2018). This is a densely connected autoencoder with layers of size $( 2 8 \\times 2 8 ) - 1 0 0 0 - 5 0 0 - 2 5 0 - 3 0$ and a symmetric decoder. A full description of the model is in Table 3. For EMNIST CR, we use a convolutional network. The network has two convolutional layers (with $3 \\times 3$ kernels), max pooling, and dropout, followed by a 128 unit dense layer. A full description of the model is in Table 4. ", + "bbox": [ + 174, + 559, + 825, + 683 + ], + "page_idx": 24 + }, + { + "type": "table", + "img_path": "images/1cac799a631a563c437bba36e917a0f9a929e4099680e3f7cbd5c5c792362666.jpg", + "table_caption": [ + "Table 3: EMNIST autoencoder model architecture. We use a sigmoid activation at all dense layers. " + ], + "table_footnote": [], + "table_body": "
LayerOutput Shape#of Trainable Parameters
Input7840
Dense1000785000
Dense500500500
Dense250125250
Dense307530
Dense2507750
Dense500125500
Dense1000501000
Dense784784784
", + "bbox": [ + 326, + 744, + 669, + 898 + ], + "page_idx": 24 + }, + { + "type": "table", + "img_path": "images/49a2be33aa30ee88c8550157d98584ccc7fbc15519e726d42098154e6e21d648.jpg", + "table_caption": [ + "Table 4: EMNIST character recognition model architecture. " + ], + "table_footnote": [], + "table_body": "
LayerOutput Shape#of Trainable ParametersActivationHyperparameters
Input(28,28,1)0ReLUkernel size = 3; strides=(1,1)
Conv2d(26,26,32)320
Conv2d(24,24,64)18496
MaxPool2d(12,12,64)0
Dropout(12,12,64)0p= 0.25
Flatten92160
Dense1281179776
Dropout1280
Dense627998softmaxp= 0.5
", + "bbox": [ + 189, + 138, + 808, + 280 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "C.3 SHAKESPEARE ", + "text_level": 1, + "bbox": [ + 174, + 313, + 320, + 327 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "Shakespeare is a language modeling dataset built from the collective works of William Shakespeare. In this dataset, each client corresponds to a speaking role with at least two lines. The dataset consists of 715 clients. Each client’s lines are partitioned into training and test sets. Here, the task is to do next character prediction. We use an RNN that first takes a series of characters as input and embeds each of them into a learned 8-dimensional space. The embedded characters are then passed through 2 LSTM layers, each with 256 nodes, followed by a densely connected softmax output layer. We split the lines of each speaking role into into sequences of 80 characters, padding if necessary. We use a vocabulary size of 90; 86 for the characters contained in the Shakespeare dataset, and 4 extra characters for padding, out-of-vocabulary, beginning of line and end of line tokens. We train our model to take a sequence of 80 characters, and predict a sequence of 80 characters formed by shifting the input sequence by one (so that its last character is the new character we are actually trying to predict). Therefore, our output dimension is $8 0 \\times 9 0$ . A full description of the model is in Table 5. ", + "bbox": [ + 173, + 338, + 825, + 502 + ], + "page_idx": 25 + }, + { + "type": "table", + "img_path": "images/4606797e00b0dcdcdf28eda69d71d35b7b1aaa87bd586e020980b2381bb56585.jpg", + "table_caption": [ + "Table 5: Shakespeare model architecture. " + ], + "table_footnote": [], + "table_body": "
LayerOutput Shape#of Trainable Parameters
Input800
Embedding(80,8)720
LSTM(80,256)271360
LSTM(80,256)525312
Dense(80,90)23130
", + "bbox": [ + 310, + 555, + 686, + 655 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "C.4 STACK OVERFLOW ", + "text_level": 1, + "bbox": [ + 174, + 685, + 348, + 700 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "Stack Overflow is a language modeling dataset consisting of question and answers from the question and answer site, Stack Overflow. The questions and answers also have associated metadata, including tags. The dataset contains 342,477 unique users which we use as clients. We perform two tasks on this dataset: tag prediction via logistic regression (Stack Overflow LR, SO LR for short), and next-word prediction (Stack Overflow NWP, SO NWP for short). For both tasks, we restrict to the $1 0 , 0 0 0 { \\mathrm { m o s t } }$ frequently used words. For Stack Overflow LR, we restrict to the 500 most frequent tags and adopt a one-versus-rest classification strategy, where each question/answer is represented as a bag-of-words vector (normalized to have sum 1). ", + "bbox": [ + 173, + 712, + 825, + 821 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "For Stack Overflow NWP, we restrict each client to the first 1000 sentences in their dataset (if they contain this many, otherwise we use the full dataset). We also perform padding and truncation to ensure that sentences have 20 words. We then represent the sentence as a sequence of indices corresponding to the 10,000 frequently used words, as well as indices representing padding, out-ofvocabulary words, beginning of sentence, and end of sentence. We perform next-word-prediction on these sequences using an RNN that embeds each word in a sentence into a learned 96-dimensional space. It then feeds the embedded words into a single LSTM layer of hidden dimension 670, followed by a densely connected softmax output layer. A full description of the model is in Table 6. The metric used in the main body is the top-1 accuracy over the proper 10,000-word vocabulary; it does not include padding, out-of-vocab, or beginning or end of sentence tokens when computing the accuracy.klkjlkj ", + "bbox": [ + 174, + 827, + 825, + 924 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 26 + }, + { + "type": "table", + "img_path": "images/07674b4a3346fba520b0df21158e36a0bd0b6bf6b8925b1eccbbe2b0ed6b2119.jpg", + "table_caption": [ + "Table 6: Stack Overflow next word prediction model architecture. " + ], + "table_footnote": [], + "table_body": "
LayerOutput Shape#of Trainable Parameters
Input200
Embedding(20,96)960384
LSTM(20,670)2055560
Dense(20,96)64416
Dense(20,10004)970388
", + "bbox": [ + 310, + 213, + 686, + 311 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "D EXPERIMENT HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 173, + 345, + 496, + 362 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "D.1 HYPERPARAMETER TUNING ", + "text_level": 1, + "bbox": [ + 176, + 377, + 411, + 392 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "Throughout our experiments, we compare the performance of different instantiations of Algorithm 1 that use different server optimizers. We use SGD, SGD with momentum (denoted SGDM), ADAGRAD, ADAM, and YOGI. For the client optimizer, we use mini-batch SGD throughout. For all tasks, we tune the client learning rate $\\eta _ { l }$ and server learning rate $\\eta$ by using a large grid search. Full descriptions of the per-task server and client learning rate grids are given in Appendix D.2. ", + "bbox": [ + 173, + 402, + 826, + 472 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "We use the version of FEDADAGRAD, FEDADAM, and FEDYOGI in Algorithm 5. We let $\\beta _ { 1 } = 0$ for FEDADAGRAD, and we let $\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9$ for FEDADAM, and FEDYOGI. For FEDAVG and FEDAVGM, we use Algorithm 4, where CLIENTOPT, SERVEROPT are SGD. For FEDAVGM, the server SGD optimizer uses a momentum parameter of 0.9. For FEDADAGRAD, FEDADAM, and FEDYOGI, we tune the parameter $\\tau$ in Algorithm 5. ", + "bbox": [ + 174, + 478, + 825, + 547 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "When tuning parameters, we select the best hyperparameters $( \\eta _ { l } , \\eta , \\tau )$ based on the average training loss over the last 100 communication rounds. Note that at each round, we only see a fraction of the total users (10 for each task except Stack Overflow NWP, which uses 50). Thus, the training loss at a given round is a noisy estimate of the population-level training loss, which is why we averave over a window of communication rounds. ", + "bbox": [ + 173, + 553, + 825, + 622 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "D.2 HYPERPARAMETER GRIDS ", + "text_level": 1, + "bbox": [ + 176, + 640, + 400, + 655 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "Below, we give the client learning rate $\\textcircled { \\eta _ { l } }$ in Algorithm 1) and server learning rate ( $\\dot { \\eta }$ in Algorithm 1) grids used for each task. These grids were chosen based on an initial evaluation over the grids ", + "bbox": [ + 173, + 665, + 823, + 694 + ], + "page_idx": 26 + }, + { + "type": "equation", + "img_path": "images/091adbb736908132cf026b29b99d714c6b88f7f7a7ae64332e741b84825d7273.jpg", + "text": "$$\n\\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \\dotsc , 1 0 ^ { 0 . 5 } \\}\n$$", + "text_format": "latex", + "bbox": [ + 374, + 700, + 624, + 719 + ], + "page_idx": 26 + }, + { + "type": "equation", + "img_path": "images/959fa51de3117b8b618ad47658706dbb38adce7165a265e04d2de3df575cffdc.jpg", + "text": "$$\n\\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \\dots , 1 0 ^ { 1 } \\}\n$$", + "text_format": "latex", + "bbox": [ + 380, + 726, + 617, + 744 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "These grids were then refined for Stack Overflow LR and EMNIST AE in an attempt to ensure that the best client/server learning rate combinations for each optimizer was contained in the interior of the learning rate grids. We generally found that these two tasks required searching larger learning rates than the other two tasks. The final grids were as follows: ", + "bbox": [ + 174, + 748, + 825, + 804 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "CIFAR-10: ", + "bbox": [ + 174, + 809, + 253, + 824 + ], + "page_idx": 26 + }, + { + "type": "equation", + "img_path": "images/9e6d7fab16dfcd5cacf80dce1a9993e9a5e85693b8b629186a033f737e0833ce.jpg", + "text": "$$\n\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 1 } \\} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 393, + 820, + 602, + 864 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "CIFAR-100: ", + "bbox": [ + 174, + 872, + 261, + 887 + ], + "page_idx": 26 + }, + { + "type": "equation", + "img_path": "images/8dc8f8156718af7f955b03bf75e7f6511ce5db902a5c743fe27b2b8a6b770d78.jpg", + "text": "$$\n\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 393, + 883, + 602, + 926 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "EMNIST AE: ", + "bbox": [ + 173, + 102, + 271, + 118 + ], + "page_idx": 27 + }, + { + "type": "equation", + "img_path": "images/b565073787858f2a311a5ab65a60e1e3bf5a43a684fe4c409e2b4a2a22d64f72.jpg", + "text": "$$\n\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 1 . 5 } , 1 0 ^ { - 1 } , \\dots , 1 0 ^ { 2 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \\dots , 1 0 ^ { 1 } \\} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 398, + 113, + 598, + 155 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "EMNIST CR: ", + "bbox": [ + 173, + 162, + 272, + 178 + ], + "page_idx": 27 + }, + { + "type": "equation", + "img_path": "images/5829fa2b6c99ae68d686975dd4e00cba1ae1f7acf303ca8632112c0596ed9480.jpg", + "text": "$$\n\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 395, + 174, + 602, + 215 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "Shakespeare: ", + "bbox": [ + 173, + 223, + 267, + 238 + ], + "page_idx": 27 + }, + { + "type": "equation", + "img_path": "images/1925b4444836ff6a20b1d96a8b03bc1e2991f89d4e3a73cc9ae5823d393261c2.jpg", + "text": "$$\n\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 395, + 234, + 602, + 275 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "StackOverflow LR: ", + "text_level": 1, + "bbox": [ + 173, + 284, + 310, + 297 + ], + "page_idx": 27 + }, + { + "type": "equation", + "img_path": "images/aa9cd39d8e20a456a48e55a688be30e9d5d8267907f22e9a98cf561f51a602ca.jpg", + "text": "$$\n\\begin{array} { c } { { \\eta _ { l } \\in \\{ 1 0 ^ { - 1 } , 1 0 ^ { - 0 . 5 } , \\dots , 1 0 ^ { 3 } \\} } } \\\\ { { \\eta \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \\dots , 1 0 ^ { 1 . 5 } \\} } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 397, + 294, + 599, + 335 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "StackOverflow NWP: ", + "text_level": 1, + "bbox": [ + 174, + 343, + 325, + 358 + ], + "page_idx": 27 + }, + { + "type": "equation", + "img_path": "images/7014d8516c84e3d145a877f3a8396b4672e15c2311e6cb4ec613ffe7b3479892.jpg", + "text": "$$\n\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 393, + 353, + 602, + 396 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "For all tasks, we tune $\\tau$ over the grid: ", + "bbox": [ + 174, + 404, + 419, + 417 + ], + "page_idx": 27 + }, + { + "type": "equation", + "img_path": "images/99916ba71eb9cc55ca05ba10dacb5c7da17683c6eea6a8fa7ac8d86b868fbfc5.jpg", + "text": "$$\n\\tau \\in \\{ 1 0 ^ { - 5 } , \\ldots , 1 0 ^ { - 1 } \\} .\n$$", + "text_format": "latex", + "bbox": [ + 419, + 424, + 576, + 444 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "D.3 PER-TASK BATCH SIZES ", + "text_level": 1, + "bbox": [ + 174, + 458, + 382, + 473 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "Given the large number of hyperparameters to tune, and to avoid conflating variables, we fix the batch size at a per-task level. When comparing centralized training to federated training in Section 5, we use the same batch size in both federated and centralized training. A full summary of the batch sizes is given in Table 7. ", + "bbox": [ + 174, + 484, + 825, + 539 + ], + "page_idx": 27 + }, + { + "type": "table", + "img_path": "images/6c7462e1397fb086dda2f8cdd601dcf42bfb28989d5b6963c3977419b97a70e3.jpg", + "table_caption": [ + "Table 7: Client batch sizes used for each task. " + ], + "table_footnote": [], + "table_body": "
TASKBATCH SIZE
CIFAR-1020
CIFAR-10020
EMNISTAE20
EMNISTCR20
SHAKESPEARE4
STACKOVERFLOWLR100
STACKOVERFLOW NWP16
", + "bbox": [ + 359, + 592, + 637, + 718 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "D.4 BEST PERFORMING HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 176, + 739, + 495, + 753 + ], + "page_idx": 27 + }, + { + "type": "text", + "text": "In this section, we present, for each optimizer, the best client and server learning rates and values of $\\tau$ found for the tasks discussed in Section 5. Specifically, these are the hyperparameters used in Figure Figure 1 and table Table 1. The validation metrics in Table 1 are obtained using the learning rates in Table 8 and the values of $\\tau$ in Table 9. As discussed in Section 4, we choose $\\eta , \\eta _ { l }$ and $\\tau$ to be the parameters that minimizer the average training loss over the last 100 communication rounds. ", + "bbox": [ + 173, + 765, + 825, + 834 + ], + "page_idx": 27 + }, + { + "type": "table", + "img_path": "images/18f4d3def369eb284a8c94b42035a6239166b092d7473e6f6b782d8c4f16463f.jpg", + "table_caption": [ + "Table 8: The base-10 logarithm of the client $( \\eta _ { l } )$ and server $( \\eta )$ learning rate combinations that achieve the accuracies from Table 1. See Appendix D.2 for a full description of the grids. " + ], + "table_footnote": [], + "table_body": "
FEDADAGRADFEDADAMFEDYOGIFEDAVGMFEDAVG
nnmnmnmnmn
CIFAR-10-3/2-1-3/2-2-3/2-2-3/2-12-120
CIFAR-100-1-1-3/20-3/20-3/20-11
EMNISTAE3/2-3/21-3/21-3/2/010
EMNIST CR-3/2-1-3/2-5/2-3/2-5/2-3/2-1/2-10
SHAKESPEARE0-1/20-20-20-1/200
STACKOVERFLOWLR212-1/22-1/22020
STACKOVERFLOW NWP -1/2-3/2-1/2-2-1/2-2-1/20-1/20
", + "bbox": [ + 192, + 261, + 805, + 400 + ], + "page_idx": 28 + }, + { + "type": "table", + "img_path": "images/b46f78a535bf0ba9996c36dce6049238dc97473b3dafec87f773b3016281aa05.jpg", + "table_caption": [ + "Table 9: The base-10 logarithm of the parameter $\\tau$ (as defined in Algorithm 2) that achieve the validation metrics in Table 1. " + ], + "table_footnote": [], + "table_body": "
FEDADAGRAD FEDADAM FEDYOGI
CIFAR-10-2-3-3
CIFAR-100-2-1-1
EMNISTAE-3-3-3
EMNIST CR-2-4-4
SHAKESPEARE-1-3-3
STACKOVERFLOWLR-2-5-5
STACKOVERFLOWNWP-4-5-5
", + "bbox": [ + 269, + 683, + 727, + 808 + ], + "page_idx": 28 + }, + { + "type": "image", + "img_path": "images/2cfd7a23d061067471c578b5c9cb5c0fbe27e8391814637353f1209c09dbb3a9.jpg", + "image_caption": [ + "Figure 4: Validation accuracy on EMNIST CR using constant learning rates $\\eta , \\eta _ { l }$ , and $\\tau$ tuned to achieve the best training performance on the last 100 communication rounds; see Appendix D for hyperparameter grids. " + ], + "image_footnote": [], + "bbox": [ + 349, + 112, + 643, + 277 + ], + "page_idx": 29 + }, + { + "type": "table", + "img_path": "images/22c05c52675ef94844cfd19ea818df09b50ac07abae2ea6a55b2da64c1871581.jpg", + "table_caption": [ + "Table 10: Test set performance for Stack Overflow tasks after training: Accuracy for NWP and Recall $\\textcircled { a } 5$ $( \\times 1 0 0 )$ for LR. Performance within within $0 . 5 \\%$ of the best result are shown in bold. " + ], + "table_footnote": [], + "table_body": "
FED...ADAGRADADAMYOGIAvGMAvG
STACK OVERFLOW NWP24.425.725.724.520.5
STACK OVERFLOW LR66.865.266.546.540.6
", + "bbox": [ + 276, + 420, + 720, + 484 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "E ADDITIONAL EXPERIMENTAL RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 513, + 535, + 530 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "E.1 RESULTS ON EMNIST CR ", + "text_level": 1, + "bbox": [ + 176, + 544, + 400, + 559 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "We plot the validation accuracy of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, FEDAVG, and SCAFFOLD on EMNIST CR. As in Figure 1, we tune the learning rates $\\eta _ { l } , \\eta$ and adaptivity $\\tau$ by selecting the parameters obtaining the smallest training loss, averaged over the last 100 training rounds. The results are given in Figure 4. ", + "bbox": [ + 174, + 570, + 825, + 626 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "We see that all methods are roughly comparably throughout the duration of training. This reflects the fact that the dataset is quite simple, and most clients have all classes in their local datasets, reducing any heterogeneity among classes. Note that SCAFFOLD performs slightly worse than FEDAVG and all other methods here. As discussed in Section 5, this may be due to the presence of stale client control variates, and the communication-limited regime of our experiments. ", + "bbox": [ + 174, + 632, + 825, + 702 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "E.2 STACK OVERFLOW TEST SET PERFORMANCE ", + "text_level": 1, + "bbox": [ + 174, + 718, + 526, + 732 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "As discussed in Section 5, in order to compute a measure of performance for the Stack Overflow tasks as training progresses, we use a subsampled version of the test dataset, due to its prohibitively large number of clients and examples. In particular, at each round of training, we sample 10,000 random test samples, and use this as a measure of performance over time. However, once training is completed, we also evaluate on the full test dataset. For the Stack Overflow experiments described in Section 5, the final test accuracy is given in Table 10. ", + "bbox": [ + 174, + 743, + 825, + 827 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "E.3 LEARNING RATE ROBUSTNESS ", + "text_level": 1, + "bbox": [ + 176, + 843, + 426, + 857 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "In this section, we showcase what combinations of client learning rate $\\eta _ { l }$ and server learning rate $\\eta$ performed well for each optimizer and task. As in Figure 2, we plot, for each optimizer, task, and pair $( \\eta _ { l } , \\eta )$ , the validation set performance (averaged over the last 100 rounds). As in Section 5, we fix $\\tau = 1 0 ^ { - 3 }$ throughout. The results, for the CIFAR-10, CIFAR-100, EMNIST AE, EMNIST CR, ", + "bbox": [ + 176, + 868, + 825, + 922 + ], + "page_idx": 29 + }, + { + "type": "text", + "text": "Shakespeare, Stack Overflow LR, and Stack Overflow NWP tasks are given in Figures 5, 6, 7, 8, 9, 10, and 11, respectively. ", + "bbox": [ + 173, + 103, + 825, + 132 + ], + "page_idx": 30 + }, + { + "type": "text", + "text": "While the general trends depend on the optimizer and task, we see that in many cases, the adaptive methods have rectangular regions that perform well. This implies a kind of robustness to fixing one of $\\eta , \\eta _ { l }$ , and varying the other. On the other hand, FEDAVGM and FEDAVG often have triangular regions that perform well, suggesting that $\\eta$ and $\\eta _ { l }$ should be tuned simultaneously. ", + "bbox": [ + 173, + 138, + 825, + 194 + ], + "page_idx": 30 + }, + { + "type": "image", + "img_path": "images/3c228a3a0a506f7fab5984fad48909c5f19d018f5100b7319cb0ea2a5cd5bbad.jpg", + "image_caption": [ + "Figure 5: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the CIFAR-10 task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\\bar { \\tau } = 1 0 ^ { - 3 }$ . " + ], + "image_footnote": [], + "bbox": [ + 204, + 207, + 797, + 479 + ], + "page_idx": 30 + }, + { + "type": "image", + "img_path": "images/988976289e57389c260e2ce5ff2bcfe18be6722e20e33a23e74e0c18d86179b5.jpg", + "image_caption": [ + "Figure 6: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the CIFAR-100 task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\\tau = 1 0 ^ { - 3 }$ . " + ], + "image_footnote": [], + "bbox": [ + 202, + 132, + 797, + 407 + ], + "page_idx": 31 + }, + { + "type": "image", + "img_path": "images/a23451f198d579719bd5c12522f418c42bc52fa09bd6fee577a2a6cd9144cba7.jpg", + "image_caption": [ + "Figure 7: Validation MSE (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the EMNIST AE task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\\tau = 1 0 ^ { - 3 }$ . " + ], + "image_footnote": [], + "bbox": [ + 200, + 559, + 797, + 811 + ], + "page_idx": 31 + }, + { + "type": "image", + "img_path": "images/c472049d458c0bdb90d3a9865546ea6d2a6905247c0626ad243509dc4925cb64.jpg", + "image_caption": [ + "Figure 8: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the EMNIST CR task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\\tau = 1 0 ^ { - 3 }$ . " + ], + "image_footnote": [], + "bbox": [ + 204, + 132, + 797, + 407 + ], + "page_idx": 32 + }, + { + "type": "image", + "img_path": "images/110df1d101eebec91adcafb1b92bcfbd8eb49008a8ef880e5dd00d979aa15b32.jpg", + "image_caption": [ + "Figure 9: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Shakespeare task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\\tau = 1 0 ^ { - 3 }$ . " + ], + "image_footnote": [], + "bbox": [ + 202, + 547, + 797, + 820 + ], + "page_idx": 32 + }, + { + "type": "image", + "img_path": "images/36a8c5f6ace9eb2d7ece4dab87f90356b80a7333fb4a7383edd182ee807213cb.jpg", + "image_caption": [ + "Figure 10: Validation recall $\\textcircled { \\alpha } 5$ (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Stack Overflow LR task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\\tau = 1 0 ^ { - 3 }$ . " + ], + "image_footnote": [], + "bbox": [ + 202, + 145, + 799, + 392 + ], + "page_idx": 33 + }, + { + "type": "image", + "img_path": "images/4238061faf9efe9b6eef6f047f57581529cb3114358e9247fc74a3a2af45e1e3.jpg", + "image_caption": [ + "Figure 11: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Stack Overflow NWP task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set $\\tau = 1 0 ^ { - 3 }$ . " + ], + "image_footnote": [], + "bbox": [ + 202, + 545, + 797, + 818 + ], + "page_idx": 33 + }, + { + "type": "text", + "text": "In order to better understand the results in Appendix E.3, we plot the relation between optimal choices of client and server learning rates. For each optimizer, task, and client learning rate $\\eta _ { l }$ , we find the best corresponding server learning rate $\\eta$ among the grids listed in Appendix D.2. Throughout, we fix $\\tau = 1 0 ^ { - 3 }$ for the adaptive methods. We omit any points for which the final validation loss is within $10 \\%$ of the worst-recorded validation loss over all hyperparameters. Essentially, we omit client learning rates that did not lead to any training of the model. The results are given in Figure 12. ", + "bbox": [ + 174, + 133, + 825, + 217 + ], + "page_idx": 34 + }, + { + "type": "image", + "img_path": "images/50c44eab15de90c5967e9a2de54c1c02f46cb58fea87d1a6bfb848613b707602.jpg", + "image_caption": [ + "Figure 12: The best server learning rate in our hyperparameter tuning grids for each client learning rate, optimizers, and task. We select the server learning rates based on the average validation performance over the last 100 communication rounds. For FEDADAGRAD, FEDADAM, and FEDYOGI, we fix $\\tau = 1 0 ^ { - 3 }$ . We omit all client learning rates for which all server learning rates did not change the initial validation loss by more than $10 \\%$ . " + ], + "image_footnote": [], + "bbox": [ + 196, + 247, + 800, + 601 + ], + "page_idx": 34 + }, + { + "type": "text", + "text": "In virtually all tasks, we see a clear inverse relationship between client learning rate $\\eta _ { l }$ and server learning rate $\\eta$ for FEDAVG and FEDAVGM. As discussed in Section 5, this supports the observation that for the non-adaptive methods, $\\eta _ { l }$ and $\\eta$ must in some sense be tuned simultaneously. On the other hand, for adaptive optimizers on most tasks we see much more stability in the best server learning rate $\\eta$ as the client learning rate $\\eta _ { l }$ varies. This supports our observation in Section 5 that for adaptive methods, tuning the client learning rate is more important. ", + "bbox": [ + 173, + 724, + 825, + 808 + ], + "page_idx": 34 + }, + { + "type": "text", + "text": "Notably, we see a clear exception to this on the Stack Overflow LR task, where there is a definitive inverse relationship between learning rates among all optimizers. The EMNIST AE task also displays somewhat different behavior. While there are still noticeable inverse relationships between learning rates for FEDAVG and FEDAVGM, the range of good client learning rates is relatively small. We emphasize that this task is fundamentally different than the remaining tasks. As noted by Zaheer et al. (2018), the primary obstacle in training bottleneck autoencoders is escaping saddle points, not in converging to critical points. Thus, we expect there to be qualitative differences between EMNIST AE and other tasks, even EMNIST CR (which uses the same dataset). ", + "bbox": [ + 173, + 814, + 825, + 924 + ], + "page_idx": 34 + }, + { + "type": "text", + "text": "E.5 ROBUSTNESS OF THE ADAPTIVITY PARAMETER ", + "text_level": 1, + "bbox": [ + 178, + 104, + 540, + 117 + ], + "page_idx": 35 + }, + { + "type": "text", + "text": "As discussed in Section 5, we plot, for each adaptive optimizer and task, the validation accuracy as a function of the adaptivity parameter $\\tau$ . In particular, for each value of $\\tau$ (which we vary over $\\{ 1 0 ^ { - 5 } , \\dotsc , 1 0 ^ { - 1 } \\}$ , see Appendix D), we plot the best possible last-100-rounds validation set performance. Specifically, we plot the average validation performance over the last 100 rounds using the best a posteriori values of client and server learning rates. The results are given in Figure 13. ", + "bbox": [ + 174, + 131, + 825, + 200 + ], + "page_idx": 35 + }, + { + "type": "image", + "img_path": "images/88a3306ab9d22edfbff475be8101d1e8b35ad59803132c108b2b9ff96ab1bfdf.jpg", + "image_caption": [ + "Figure 13: Validation performance of FEDADAGRAD, FEDADAM, and FEDYOGI for varying $\\tau$ on various tasks. The learning rates $\\eta$ and $\\eta _ { l }$ are tuned for each $\\tau$ to achieve the best training performance on the last 100 communication rounds. " + ], + "image_footnote": [], + "bbox": [ + 205, + 224, + 792, + 566 + ], + "page_idx": 35 + }, + { + "type": "text", + "text": "E.6 IMPROVING PERFORMANCE WITH LEARNING RATE DECAY ", + "text_level": 1, + "bbox": [ + 174, + 671, + 616, + 684 + ], + "page_idx": 35 + }, + { + "type": "text", + "text": "Despite the success of adaptive methods, it is natural to ask if there is still more to be gained. To test this, we trained the EMNIST CR model in a centralized fashion on a shuffled version of the dataset. We trained for 100 epochs and used tuned learning rates for each (centralized) optimizer, achieving an accuracy of $8 8 \\%$ (see Table 11, CENTRALIZED row), significantly above the best EMNIST CR results from Table 1. The theoretical results in Section 3 point to a partial explanation, as they only hold when the client learning rate is small or is decayed over time. ", + "bbox": [ + 173, + 698, + 825, + 780 + ], + "page_idx": 35 + }, + { + "type": "text", + "text": "To validate this, we ran the same hyperparameter grids on the federated EMNIST CR task, but using a client learning rate schedule. We use a “staircase” exponential decay schedule (EXPDECAY) where the client learning rate $\\eta _ { l }$ is decreased by a factor of 0.1 every 500 rounds. This is analogous in some sense to standard staircase learning rate schedules in centralized optimization (Goyal et al., 2017). Table 11 gives the results. EXPDECAY improves the accuracy of all optimizers, and allows most to get close to the best centralized accuracy. While we do not close the gap with centralized optimization entirely, we suspect that further tuning of the amount and frequency of decay may lead to even better accuracies. However, this may also require performing significantly more communication rounds, as the theoretical results in Section 3 are primarily asymptotic. In communication-limited settings, the added benefit of learning rate decay seems to be modest. ", + "bbox": [ + 173, + 787, + 825, + 924 + ], + "page_idx": 35 + }, + { + "type": "table", + "img_path": "images/9ab3ed2946f8c758203144c35238f338ebc6565635feb02aa760ffd0caeae7ca.jpg", + "table_caption": [ + "Table 11: (Top) Test accuracy $( \\% )$ of a model trained centrally with various optimizers. (Bottom) Average test accuracy $( \\% )$ over the last 100 rounds of various federated optimizers on the EMNIST CR task, using constant learning rates or the EXPDECAY schedule for $\\eta _ { l }$ . Accuracies (for the federated tasks) within $\\mathrm { { \\bar { 0 } . 5 \\% } }$ of the best result are shown in bold. " + ], + "table_footnote": [], + "table_body": "
ADAGRAD CENTRALIZED 88.0ADAM 87.9YOGI 88.0SGDM 87.7SGD 87.7
FED... ADAGRADADAMYOGIAvGMAVG
CONSTANT ηt85.1 85.685.585.284.9
EXPDECAY85.3 86.286.285.885.2
", + "bbox": [ + 313, + 183, + 684, + 273 + ], + "page_idx": 36 + }, + { + "type": "text", + "text": "F CREATING A FEDERATED CIFAR-100 ", + "text_level": 1, + "bbox": [ + 173, + 303, + 524, + 320 + ], + "page_idx": 36 + }, + { + "type": "text", + "text": "Overview We use the Pachinko Allocation Method (PAM) (Li & McCallum, 2006) to create a federated CIFAR-100. PAM is a topic modeling method in which the correlations between individual words in a vocabulary are represented by a rooted directed acyclic graph (DAG) whose leaves are the vocabulary words. The interior nodes are topics with Dirichlet distributions over their child nodes. To generate a document, we sample multinomial distributions from each interior node’s Dirichlet distribution. To sample a word from the document, we begin at the root, and draw a child node its multinomial distribution, and continue doing this until we reach a leaf node. ", + "bbox": [ + 173, + 334, + 825, + 430 + ], + "page_idx": 36 + }, + { + "type": "text", + "text": "To partition CIFAR-100 across clients, we use the label structure of CIFAR-100. Each image in the dataset has a fine label (often referred to as its label) which is a member of a coarse label. For example, the fine label “seal” is a member of the coarse label “aquatic mammals”. There are 20 coarse labels in CIFAR-100, each with 5 fine labels. We represent this structure as a DAG $G$ , with a root whose children are the coarse labels. Each coarse label is an interior node whose child nodes are its fine labels. The root node has a symmetric Dirichlet distribution with parameter $\\alpha$ over the coarse labels, and each coarse label has a symmetric Dirichlet distribution with parameter $\\beta$ . ", + "bbox": [ + 174, + 436, + 825, + 532 + ], + "page_idx": 36 + }, + { + "type": "text", + "text": "We associate each client to a document. That is, we draw a multinomial distribution from the Dirichlet prior at the root $\\left( \\operatorname { D i r } ( \\alpha ) \\right)$ and a multinomial distribution from the Dirichlet prior at each coarse label $\\bar { ( } \\mathrm { D i r } ( \\beta ) )$ . To create the client dataset, we draw leaf nodes from this DAG using Pachinko allocation, randomly sample an example with the given fine label, and assign it to the client’s dataset. We do this 100 times for each of 500 distinct training clients. ", + "bbox": [ + 174, + 539, + 825, + 608 + ], + "page_idx": 36 + }, + { + "type": "text", + "text": "While more complex than LDA, this approach creates more realistic heterogeneity among client datasets by creating correlations between label frequencies for fine labels within the same coarse label set. Intuitively, if a client’s dataset has many images of dolphins, they are likely to also have pictures of whales. By using a small $\\alpha$ at the root, client datasets become more likely to focus on a few coarse labels. By using a larger $\\beta$ for the coarse-to-fine label distributions, clients are more likely to have multiple fine labels from the same coarse label. ", + "bbox": [ + 174, + 614, + 825, + 696 + ], + "page_idx": 36 + }, + { + "type": "text", + "text": "One important note: Once we sample a fine label, we randomly select an element with that label without replacement. This ensures no two clients have overlapping examples. In more detail, suppose we have sample a fine label $y$ with coarse label $c$ for client $m$ , and there is only one remaining such example $( x , { \\bar { c } } , y )$ . We assign $( x , c , y )$ to client $m$ ’s dataset, and remove the leaf node $y$ from the DAG $G$ . We also remove $y$ from the multinomial distribution $\\theta _ { c }$ that client $m$ has associated to coarse label $c$ , which we refer to as renormalization with respect to $y$ (Algorithm 8). If $i$ has no remaining children after pruning node $j$ , we also remove node $i$ from $G$ and re-normalize the root multinomial $\\theta _ { r }$ with respect to $c$ . For all subsequent clients, we draw multinomials from this pruned $G$ according to symmetric Dirichlet distributions with the same parameters as before, but with one fewer category. ", + "bbox": [ + 173, + 703, + 825, + 827 + ], + "page_idx": 36 + }, + { + "type": "text", + "text": "Notation and method Let $\\mathcal { C }$ denote the set of coarse labels and $\\mathcal { V }$ the set of fine labels, and let $s$ denote the CIFAR-100 dataset. This consists of examples $( x , c , y )$ where $x$ is an image vector, $c \\in { \\mathcal { C } }$ is a coarse label set, and $y \\in \\mathcal { V }$ is a fine label with $y \\in c$ . For $c \\in \\mathcal { C } , y \\in \\mathcal { D }$ , we let $S _ { c }$ and $\\mathcal { S } _ { y }$ denote the set of examples in $s$ with coarse label $c$ and fine label $y$ . For $v \\in G$ we let $| G [ v ] |$ denote the set of children of $v$ in $G$ . For $\\gamma \\in \\mathbb { R }$ , let $\\operatorname { D i r } ( \\gamma , k )$ denote the symmetric Dirichlet distribution with $k$ categories. ", + "bbox": [ + 173, + 840, + 825, + 924 + ], + "page_idx": 36 + }, + { + "type": "text", + "text": "Algorithm 7 Creating a federated CIFAR-100 ", + "text_level": 1, + "bbox": [ + 174, + 103, + 475, + 117 + ], + "page_idx": 37 + }, + { + "type": "text", + "text": "Input: $N , M \\in \\mathbb { Z } _ { > 0 } , \\alpha , \\beta \\in \\mathbb { R } _ { \\ge 0 }$ \nfor $m = 1 , \\cdots , M$ do Sample $\\theta _ { r } \\sim \\operatorname { D i r } ( \\alpha , | G [ r ] | )$ for $c \\in { \\mathcal { C } } \\cap G [ r ]$ do Sample $\\theta _ { c } \\sim \\operatorname { D i r } ( \\beta , | G [ c ] | )$ $D _ { m } = \\varnothing$ for $n = 1 , \\cdots N$ do Sample c ∼ Multinomial $\\left( \\theta _ { r } \\right)$ Sample y ∼ Multinomial $\\left( \\theta _ { c } \\right)$ Select $( x , c , y ) \\in S$ uniformly at random $D _ { m } = D _ { m } \\cup \\{ ( x , c , y ) \\}$ $\\mathcal { S } = \\mathcal { S } \\backslash \\{ ( x , c , \\overset { \\cdot } { y } ) \\}$ if $\\begin{array} { r } { { \\cal S } _ { y } = \\emptyset } \\end{array}$ then G = G\\{y} θc = RENORMALIZE(θc, y) if Sc = ∅ then G = G\\{c} θr = RENORMALIZE(θr, c) ", + "bbox": [ + 174, + 119, + 508, + 372 + ], + "page_idx": 37 + }, + { + "type": "table", + "img_path": "images/efaa025501c7065b26a782907225a109bc4617af98f087301f0971bf59bbce79.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Algorithm8RENORMALIZE
Initialization: θ= (p1,...,Pk),i∈[K]
a=∑k≠iPk
fork ∈[K],k≠ido
Pk=pk/a
Return0' = (pi,···P'-1,P'+1,···,PK)
", + "bbox": [ + 173, + 387, + 825, + 488 + ], + "page_idx": 37 + }, + { + "type": "text", + "text": "Let $M$ denote the number of clients, $N$ the number of examples per client, and $D _ { m }$ the dataset for client $m \\in \\{ 1 , \\cdots , M \\}$ . A full description of our method is given in Algorithm 7. For our experiments, we use $N = 1 0 0$ , $M = 5 0 0$ , $\\alpha = 0 . 1 , \\beta = 1 0$ . In Figure 14, we plot the distribution of unique labels among the 500 training clients. Each client has only a fraction of the overall labels in the distribution. Moreover, there is variance in the number of unique labels, with most clients having between 20 and 30, and some having over 40. Some client datasets have very few unique labels. While this is primarily an artifact of performing without replacement sampling, this helps increase the heterogeneity of the dataset in a way that can reflect practical concerns, as in many settings, clients may only have a few types of labels in their dataset. ", + "bbox": [ + 173, + 515, + 826, + 638 + ], + "page_idx": 37 + }, + { + "type": "image", + "img_path": "images/759e0a67cf8e490c2d7eb60abe974fcab397e55f7a8e0758467c8c94ae7999c1.jpg", + "image_caption": [ + "Figure 14: The distribution of the number of unique labels among training client datasets in our federated CIFAR-100 dataset. " + ], + "image_footnote": [], + "bbox": [ + 339, + 695, + 629, + 876 + ], + "page_idx": 37 + } +] \ No newline at end of file diff --git a/parse/train/LkFG3lB13U5/LkFG3lB13U5_middle.json b/parse/train/LkFG3lB13U5/LkFG3lB13U5_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..45c5fcb0bc741f45358f5a911ab53c2b1e58052d --- /dev/null +++ b/parse/train/LkFG3lB13U5/LkFG3lB13U5_middle.json @@ -0,0 +1,90059 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 391, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 394, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 394, + 99 + ], + "score": 1.0, + "content": "ADAPTIVE FEDERATED OPTIMIZATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 114, + 114, + 473, + 137 + ], + "lines": [ + { + "bbox": [ + 111, + 114, + 474, + 128 + ], + "spans": [ + { + "bbox": [ + 111, + 114, + 474, + 128 + ], + "score": 1.0, + "content": "Sashank J. Reddi∗, Zachary Charles∗, Manzil Zaheer, Zachary Garrett, Keith Rush,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 126, + 352, + 138 + ], + "spans": [ + { + "bbox": [ + 112, + 126, + 352, + 138 + ], + "score": 1.0, + "content": "Jakub Konecný, Sanjiv Kumar, H. Brendan McMahan ˇ", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 114, + 138, + 448, + 169 + ], + "lines": [ + { + "bbox": [ + 111, + 136, + 183, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 136, + 183, + 149 + ], + "score": 1.0, + "content": "Google Research", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 147, + 448, + 160 + ], + "spans": [ + { + "bbox": [ + 113, + 147, + 448, + 160 + ], + "score": 1.0, + "content": "{sashank, zachcharles, manzilzaheer, zachgarrett, krush,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 158, + 331, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 331, + 171 + ], + "score": 1.0, + "content": "konkey, sanjivk, mcmahan}@google.com", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 198, + 333, + 210 + ], + "lines": [ + { + "bbox": [ + 276, + 198, + 335, + 211 + ], + "spans": [ + { + "bbox": [ + 276, + 198, + 335, + 211 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 223, + 469, + 353 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "score": 1.0, + "content": "Federated learning is a distributed machine learning paradigm in which a large", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 234, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 247 + ], + "score": 1.0, + "content": "number of clients coordinate with a central server to learn a model without sharing", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 245, + 470, + 256 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 256 + ], + "score": 1.0, + "content": "their own training data. Standard federated optimization methods such as Fed-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 256, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 142, + 256, + 469, + 267 + ], + "score": 1.0, + "content": "erated Averaging (FEDAVG) are often difficult to tune and exhibit unfavorable", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "convergence behavior. In non-federated settings, adaptive optimization methods", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 278, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 289 + ], + "score": 1.0, + "content": "have had notable success in combating such issues. In this work, we propose fed-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "score": 1.0, + "content": "erated versions of adaptive optimizers, including ADAGRAD, ADAM, and YOGI,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "score": 1.0, + "content": "and analyze their convergence in the presence of heterogeneous data for general", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 309, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 309, + 469, + 322 + ], + "score": 1.0, + "content": "nonconvex settings. Our results highlight the interplay between client heterogeneity", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 320, + 469, + 332 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 469, + 332 + ], + "score": 1.0, + "content": "and communication efficiency. We also perform extensive experiments on these", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 331, + 469, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 331, + 469, + 343 + ], + "score": 1.0, + "content": "methods and show that the use of adaptive optimizers can significantly improve the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 341, + 283, + 355 + ], + "spans": [ + { + "bbox": [ + 141, + 341, + 283, + 355 + ], + "score": 1.0, + "content": "performance of federated learning.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 108, + 375, + 206, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 374, + 208, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 208, + 390 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "Federated learning (FL) is a machine learning paradigm in which multiple clients cooperate to learn", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "a model under the orchestration of a central server (McMahan et al., 2017). In FL, raw client data", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "is never shared with the server or other clients. This distinguishes FL from traditional distributed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 432, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 444 + ], + "score": 1.0, + "content": "optimization, and requires contending with heterogeneous data. FL has two primary settings, cross-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "score": 1.0, + "content": "silo (eg. FL between large institutions) and cross-device (eg. FL across edge devices) (Kairouz et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 467 + ], + "score": 1.0, + "content": "2019, Table 1). In cross-silo FL, most clients participate in every round and can maintain state", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "between rounds. In the more challenging cross-device FL, our primary focus, only a small fraction of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "clients participate in each round, and clients cannot maintain state across rounds. For a more in-depth", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "score": 1.0, + "content": "discussion of FL and the challenges involved, we defer to Kairouz et al. (2019) and Li et al. (2019a).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 514 + ], + "score": 1.0, + "content": "Standard optimization methods, such as distributed SGD, are often unsuitable in FL and can incur", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 514, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 505, + 525 + ], + "score": 1.0, + "content": "high communication costs. To remedy this, many federated optimization methods use local client", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 524, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 507, + 536 + ], + "score": 1.0, + "content": "updates, in which clients update their models multiple times before communicating with the server.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "This can greatly reduce the amount of communication required to train a model. One such method is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "score": 1.0, + "content": "FEDAVG (McMahan et al., 2017), in which clients perform multiple epochs of SGD on their local", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "datasets. The clients communicate their models to the server, which averages them to form a new", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 566, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 581 + ], + "score": 1.0, + "content": "global model. While FEDAVG has seen great success, recent works have highlighted its convergence", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "issues in some settings (Karimireddy et al., 2019; Hsu et al., 2019). This is due to a variety of factors", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "including (1) client drift (Karimireddy et al., 2019), where local client models move away from", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 598, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 613 + ], + "score": 1.0, + "content": "globally optimal models, and (2) a lack of adaptivity. FEDAVG is similar in spirit to SGD, and may", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "be unsuitable for settings with heavy-tail stochastic gradient noise distributions, which often arise", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 619, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 635 + ], + "score": 1.0, + "content": "when training language models (Zhang et al., 2019a). Such settings benefit from adaptive learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 631, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 481, + 644 + ], + "score": 1.0, + "content": "rates, which incorporate knowledge of past iterations to perform more informed optimization.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 662 + ], + "score": 1.0, + "content": "In this paper, we focus on the second issue and present a simple framework for incorporating", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "adaptivity in FL. In particular, we propose a general optimization framework in which (1) clients", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "perform multiple epochs of training using a client optimizer to minimize loss on their local data and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "(2) server updates its global model by applying a gradient-based server optimizer to the average of the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 691, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 505, + 703 + ], + "score": 1.0, + "content": "clients’ model updates. We show that FEDAVG is the special case where SGD is used as both client", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 701, + 506, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 506, + 715 + ], + "score": 1.0, + "content": "and server optimizer and server learning rate is 1. This framework can also seamlessly incorporate", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 269, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 721, + 271, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 721, + 271, + 733 + ], + "score": 1.0, + "content": "∗Authors contributed equally to this work", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 79, + 391, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 394, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 394, + 99 + ], + "score": 1.0, + "content": "ADAPTIVE FEDERATED OPTIMIZATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 114, + 114, + 473, + 137 + ], + "lines": [ + { + "bbox": [ + 111, + 114, + 474, + 128 + ], + "spans": [ + { + "bbox": [ + 111, + 114, + 474, + 128 + ], + "score": 1.0, + "content": "Sashank J. Reddi∗, Zachary Charles∗, Manzil Zaheer, Zachary Garrett, Keith Rush,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 126, + 352, + 138 + ], + "spans": [ + { + "bbox": [ + 112, + 126, + 352, + 138 + ], + "score": 1.0, + "content": "Jakub Konecný, Sanjiv Kumar, H. Brendan McMahan ˇ", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 111, + 114, + 474, + 138 + ] + }, + { + "type": "list", + "bbox": [ + 114, + 138, + 448, + 169 + ], + "lines": [ + { + "bbox": [ + 111, + 136, + 183, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 136, + 183, + 149 + ], + "score": 1.0, + "content": "Google Research", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 147, + 448, + 160 + ], + "spans": [ + { + "bbox": [ + 113, + 147, + 448, + 160 + ], + "score": 1.0, + "content": "{sashank, zachcharles, manzilzaheer, zachgarrett, krush,", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 158, + 331, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 158, + 331, + 171 + ], + "score": 1.0, + "content": "konkey, sanjivk, mcmahan}@google.com", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + } + ], + "index": 4, + "bbox_fs": [ + 111, + 136, + 448, + 171 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 198, + 333, + 210 + ], + "lines": [ + { + "bbox": [ + 276, + 198, + 335, + 211 + ], + "spans": [ + { + "bbox": [ + 276, + 198, + 335, + 211 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 223, + 469, + 353 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "score": 1.0, + "content": "Federated learning is a distributed machine learning paradigm in which a large", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 234, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 247 + ], + "score": 1.0, + "content": "number of clients coordinate with a central server to learn a model without sharing", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 245, + 470, + 256 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 256 + ], + "score": 1.0, + "content": "their own training data. Standard federated optimization methods such as Fed-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 256, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 142, + 256, + 469, + 267 + ], + "score": 1.0, + "content": "erated Averaging (FEDAVG) are often difficult to tune and exhibit unfavorable", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "convergence behavior. In non-federated settings, adaptive optimization methods", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 278, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 289 + ], + "score": 1.0, + "content": "have had notable success in combating such issues. In this work, we propose fed-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 300 + ], + "score": 1.0, + "content": "erated versions of adaptive optimizers, including ADAGRAD, ADAM, and YOGI,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 470, + 311 + ], + "score": 1.0, + "content": "and analyze their convergence in the presence of heterogeneous data for general", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 309, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 309, + 469, + 322 + ], + "score": 1.0, + "content": "nonconvex settings. Our results highlight the interplay between client heterogeneity", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 320, + 469, + 332 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 469, + 332 + ], + "score": 1.0, + "content": "and communication efficiency. We also perform extensive experiments on these", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 331, + 469, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 331, + 469, + 343 + ], + "score": 1.0, + "content": "methods and show that the use of adaptive optimizers can significantly improve the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 341, + 283, + 355 + ], + "spans": [ + { + "bbox": [ + 141, + 341, + 283, + 355 + ], + "score": 1.0, + "content": "performance of federated learning.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5, + "bbox_fs": [ + 141, + 222, + 470, + 355 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 375, + 206, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 374, + 208, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 208, + 390 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "Federated learning (FL) is a machine learning paradigm in which multiple clients cooperate to learn", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "a model under the orchestration of a central server (McMahan et al., 2017). In FL, raw client data", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "is never shared with the server or other clients. This distinguishes FL from traditional distributed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 432, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 444 + ], + "score": 1.0, + "content": "optimization, and requires contending with heterogeneous data. FL has two primary settings, cross-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "score": 1.0, + "content": "silo (eg. FL between large institutions) and cross-device (eg. FL across edge devices) (Kairouz et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 467 + ], + "score": 1.0, + "content": "2019, Table 1). In cross-silo FL, most clients participate in every round and can maintain state", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "between rounds. In the more challenging cross-device FL, our primary focus, only a small fraction of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "clients participate in each round, and clients cannot maintain state across rounds. For a more in-depth", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 497 + ], + "score": 1.0, + "content": "discussion of FL and the challenges involved, we defer to Kairouz et al. (2019) and Li et al. (2019a).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 399, + 506, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 514 + ], + "score": 1.0, + "content": "Standard optimization methods, such as distributed SGD, are often unsuitable in FL and can incur", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 514, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 505, + 525 + ], + "score": 1.0, + "content": "high communication costs. To remedy this, many federated optimization methods use local client", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 524, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 507, + 536 + ], + "score": 1.0, + "content": "updates, in which clients update their models multiple times before communicating with the server.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "This can greatly reduce the amount of communication required to train a model. One such method is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "score": 1.0, + "content": "FEDAVG (McMahan et al., 2017), in which clients perform multiple epochs of SGD on their local", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "datasets. The clients communicate their models to the server, which averages them to form a new", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 566, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 581 + ], + "score": 1.0, + "content": "global model. While FEDAVG has seen great success, recent works have highlighted its convergence", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "issues in some settings (Karimireddy et al., 2019; Hsu et al., 2019). This is due to a variety of factors", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "including (1) client drift (Karimireddy et al., 2019), where local client models move away from", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 598, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 613 + ], + "score": 1.0, + "content": "globally optimal models, and (2) a lack of adaptivity. FEDAVG is similar in spirit to SGD, and may", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "be unsuitable for settings with heavy-tail stochastic gradient noise distributions, which often arise", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 619, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 635 + ], + "score": 1.0, + "content": "when training language models (Zhang et al., 2019a). Such settings benefit from adaptive learning", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 631, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 481, + 644 + ], + "score": 1.0, + "content": "rates, which incorporate knowledge of past iterations to perform more informed optimization.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 503, + 507, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 662 + ], + "score": 1.0, + "content": "In this paper, we focus on the second issue and present a simple framework for incorporating", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "adaptivity in FL. In particular, we propose a general optimization framework in which (1) clients", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "perform multiple epochs of training using a client optimizer to minimize loss on their local data and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "(2) server updates its global model by applying a gradient-based server optimizer to the average of the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 691, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 505, + 703 + ], + "score": 1.0, + "content": "clients’ model updates. We show that FEDAVG is the special case where SGD is used as both client", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 701, + 506, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 506, + 715 + ], + "score": 1.0, + "content": "and server optimizer and server learning rate is 1. This framework can also seamlessly incorporate", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "adaptivity by using adaptive optimizers as client or server optimizers. Building upon this, we develop", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "novel adaptive optimization techniques for FL by using per-coordinate methods as server optimizers.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "By focusing on adaptive server optimization, we enable use of adaptive learning rates without increase", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 461, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 461, + 127 + ], + "score": 1.0, + "content": "in client storage or communication costs, and ensure compatibility with cross-device FL.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 646, + 506, + 715 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "adaptivity by using adaptive optimizers as client or server optimizers. Building upon this, we develop", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "novel adaptive optimization techniques for FL by using per-coordinate methods as server optimizers.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "By focusing on adaptive server optimization, we enable use of adaptive learning rates without increase", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 461, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 461, + 127 + ], + "score": 1.0, + "content": "in client storage or communication costs, and ensure compatibility with cross-device FL.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 109, + 131, + 477, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 479, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 479, + 145 + ], + "score": 1.0, + "content": "Main contributions In light of the above, we highlight the main contributions of the paper.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 109, + 151, + 506, + 292 + ], + "lines": [ + { + "bbox": [ + 109, + 150, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 109, + 150, + 506, + 164 + ], + "score": 1.0, + "content": "• We study a general framework for federated optimization using server and client optimizers. This", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 162, + 483, + 175 + ], + "spans": [ + { + "bbox": [ + 118, + 162, + 483, + 175 + ], + "score": 1.0, + "content": "framework generalizes many existing federated optimization methods, including FEDAVG.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 176, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 110, + 176, + 505, + 190 + ], + "score": 1.0, + "content": "• We use this framework to design novel, cross-device compatible, adaptive federated optimization", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 188, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 118, + 188, + 506, + 201 + ], + "score": 1.0, + "content": "methods, and provide convergence analysis in general nonconvex settings. To the best of our", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 118, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "knowledge, these are the first methods for FL using adaptive server optimization. We show an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 118, + 210, + 485, + 221 + ], + "spans": [ + { + "bbox": [ + 118, + 210, + 485, + 221 + ], + "score": 1.0, + "content": "important interplay between the number of local steps and the heterogeneity among clients.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 113, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "• We introduce comprehensive and reproducible empirical benchmarks for comparing federated", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 118, + 234, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 118, + 234, + 506, + 247 + ], + "score": 1.0, + "content": "optimization methods. These benchmarks consist of seven diverse and representative FL tasks", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 118, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 118, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "involving both image and text data, with varying amounts of heterogeneity and numbers of clients.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 109, + 258, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 109, + 258, + 506, + 273 + ], + "score": 1.0, + "content": "• We demonstrate strong empirical performance of our adaptive optimizers throughout, improving", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 118, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 118, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "upon commonly used baselines. Our results show that our methods can be easier to tune, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 118, + 280, + 302, + 294 + ], + "spans": [ + { + "bbox": [ + 118, + 280, + 302, + 294 + ], + "score": 1.0, + "content": "highlight their utility in cross-device settings.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 506, + 312 + ], + "score": 1.0, + "content": "Related work FEDAVG was first introduced by McMahan et al. (2017), who showed it can dramati-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 312, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 323 + ], + "score": 1.0, + "content": "cally reduce communication costs. Many variants have since been proposed to tackle issues such as", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "convergence and client drift. Examples include adding a regularization term in the client objectives", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "towards the broadcast model (Li et al., 2018), and server momentum (Hsu et al., 2019). When clients", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 344, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 355 + ], + "score": 1.0, + "content": "are homogeneous, FEDAVG reduces to local SGD (Zinkevich et al., 2010), which has been analyzed", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "by many works (Stich, 2019; Yu et al., 2019; Wang & Joshi, 2018; Stich & Karimireddy, 2019; Basu", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "et al., 2019). In order to analyze FEDAVG in heterogeneous settings, many works derive convergence", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "rates depending on the amount of heterogeneity (Li et al., 2018; Wang et al., 2019; Khaled et al., 2019;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "score": 1.0, + "content": "Li et al., 2019b). Typically, the convergence rate of FEDAVG gets worse with client heterogeneity.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "By using control variates to reduce client drift, the SCAFFOLD method (Karimireddy et al., 2019)", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "achieves convergence rates that are independent of the amount of heterogeneity. While effective in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "cross-silo FL, the method is incompatible with cross-device FL as it requires clients to maintain state", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 430, + 429, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 429, + 442 + ], + "score": 1.0, + "content": "across rounds. For more detailed comparisons, we defer to Kairouz et al. (2019).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 447, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "score": 1.0, + "content": "Adaptive methods have been the subject of significant theoretical and empirical study, in both", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "convex (McMahan & Streeter, 2010b; Duchi et al., 2011; Kingma & Ba, 2015) and non-convex", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "settings (Li & Orabona, 2018; Ward et al., 2018; Wu et al., 2019). Reddi et al. (2019); Zaheer et al.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "(2018) study convergence failures of ADAM in certain non-convex settings, and develop an adaptive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "optimizer, YOGI, designed to improve convergence. While most work on adaptive methods focuses", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "on non-FL settings, Xie et al. (2019) propose ADAALTER, a method for FL using adaptive client", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "optimization. Conceptually, our approach is also related to the LOOKAHEAD optimizer (Zhang et al.,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "2019b), which was designed for non-FL settings. Similar to ADAALTER, an adaptive FL variant of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "LOOKAHEAD entails adaptive client optimization (see Appendix B.3 for more details). We note that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 542, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 506, + 555 + ], + "score": 1.0, + "content": "both ADAALTER and LOOKAHEAD are, in fact, special cases of our framework (see Algorithm 1)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 506, + 567 + ], + "score": 1.0, + "content": "and the primary novelty of our work comes in focusing on adaptive server optimization. This allows", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 565, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 577 + ], + "score": 1.0, + "content": "us to avoid aggregating optimizer states across clients, making our methods require at most half as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 576, + 464, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 464, + 588 + ], + "score": 1.0, + "content": "much communication and client memory usage per round (see Appendix B.3 for details).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 165, + 605 + ], + "score": 1.0, + "content": "Notation For", + "type": "text" + }, + { + "bbox": [ + 166, + 591, + 206, + 604 + ], + "score": 0.9, + "content": "a , b \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 591, + 237, + 605 + ], + "score": 1.0, + "content": ", we let", + "type": "text" + }, + { + "bbox": [ + 238, + 591, + 266, + 604 + ], + "score": 0.91, + "content": "{ \\sqrt { a } } , a ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 591, + 285, + 605 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 286, + 592, + 302, + 604 + ], + "score": 0.89, + "content": "a / b", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 591, + 505, + 605 + ], + "score": 1.0, + "content": "denote the element-wise square root, square, and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 603, + 468, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 216, + 617 + ], + "score": 1.0, + "content": "division of the vectors. For", + "type": "text" + }, + { + "bbox": [ + 216, + 604, + 249, + 615 + ], + "score": 0.92, + "content": "\\theta _ { i } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 603, + 303, + 617 + ], + "score": 1.0, + "content": ", we use both", + "type": "text" + }, + { + "bbox": [ + 303, + 604, + 318, + 617 + ], + "score": 0.91, + "content": "\\theta _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 603, + 336, + 617 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 337, + 604, + 355, + 617 + ], + "score": 0.92, + "content": "[ \\theta _ { i } ] _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 603, + 408, + 617 + ], + "score": 1.0, + "content": "to denote its", + "type": "text" + }, + { + "bbox": [ + 408, + 604, + 420, + 616 + ], + "score": 0.88, + "content": "j ^ { \\mathrm { t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 603, + 468, + 617 + ], + "score": 1.0, + "content": "coordinate.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "title", + "bbox": [ + 107, + 631, + 318, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 319, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 319, + 646 + ], + "score": 1.0, + "content": "2 FEDERATED LEARNING AND FEDAVG", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 382, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 382, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 382, + 668 + ], + "score": 1.0, + "content": "In federated learning, we solve an optimization problem of the form:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 671, + 360, + 704 + ], + "lines": [ + { + "bbox": [ + 250, + 671, + 360, + 704 + ], + "spans": [ + { + "bbox": [ + 250, + 671, + 360, + 704 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { x \\in \\mathbb { R } ^ { d } } f ( x ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } F _ { i } ( x ) ,", + "type": "interline_equation", + "image_path": "1e92002b69629c7c98895c3a52f42672bac76a3ac172efccb91bb0ec97aaca12.jpg" + } + ] + } + ], + "index": 47.5, + "virtual_lines": [ + { + "bbox": [ + 250, + 671, + 360, + 687.5 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 250, + 687.5, + 360, + 704.0 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 708, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 134, + 721 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 709, + 239, + 721 + ], + "score": 0.92, + "content": "F _ { i } ( x ) = \\mathbb { E } _ { z \\sim \\mathcal { D } _ { i } } [ f _ { i } ( x , z ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 707, + 352, + 721 + ], + "score": 1.0, + "content": ", is the loss function of the", + "type": "text" + }, + { + "bbox": [ + 352, + 708, + 363, + 719 + ], + "score": 0.87, + "content": "i ^ { \\mathrm { { t h } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 707, + 393, + 721 + ], + "score": 1.0, + "content": "client,", + "type": "text" + }, + { + "bbox": [ + 393, + 709, + 423, + 720 + ], + "score": 0.9, + "content": "z \\in { \\mathcal { Z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 707, + 445, + 721 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 446, + 709, + 458, + 720 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 707, + 506, + 721 + ], + "score": 1.0, + "content": "is the data", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 718, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 186, + 734 + ], + "score": 1.0, + "content": "distribution for the", + "type": "text" + }, + { + "bbox": [ + 186, + 721, + 197, + 730 + ], + "score": 0.84, + "content": "i ^ { \\mathrm { { t h } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 718, + 243, + 734 + ], + "score": 1.0, + "content": "client. For", + "type": "text" + }, + { + "bbox": [ + 244, + 721, + 267, + 732 + ], + "score": 0.83, + "content": "i \\neq j", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 718, + 272, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 272, + 721, + 284, + 732 + ], + "score": 0.72, + "content": "\\mathcal { D } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 718, + 303, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 304, + 721, + 316, + 733 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 718, + 471, + 734 + ], + "score": 1.0, + "content": "may be very different. The functions", + "type": "text" + }, + { + "bbox": [ + 472, + 721, + 483, + 732 + ], + "score": 0.88, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 718, + 506, + 734 + ], + "score": 1.0, + "content": "(and", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 126 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 127 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 109, + 131, + 477, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 479, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 479, + 145 + ], + "score": 1.0, + "content": "Main contributions In light of the above, we highlight the main contributions of the paper.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 129, + 479, + 145 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 151, + 506, + 292 + ], + "lines": [ + { + "bbox": [ + 109, + 150, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 109, + 150, + 506, + 164 + ], + "score": 1.0, + "content": "• We study a general framework for federated optimization using server and client optimizers. This", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 162, + 483, + 175 + ], + "spans": [ + { + "bbox": [ + 118, + 162, + 483, + 175 + ], + "score": 1.0, + "content": "framework generalizes many existing federated optimization methods, including FEDAVG.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 176, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 110, + 176, + 505, + 190 + ], + "score": 1.0, + "content": "• We use this framework to design novel, cross-device compatible, adaptive federated optimization", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 188, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 118, + 188, + 506, + 201 + ], + "score": 1.0, + "content": "methods, and provide convergence analysis in general nonconvex settings. To the best of our", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 118, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "knowledge, these are the first methods for FL using adaptive server optimization. We show an", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 118, + 210, + 485, + 221 + ], + "spans": [ + { + "bbox": [ + 118, + 210, + 485, + 221 + ], + "score": 1.0, + "content": "important interplay between the number of local steps and the heterogeneity among clients.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 113, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "• We introduce comprehensive and reproducible empirical benchmarks for comparing federated", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 118, + 234, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 118, + 234, + 506, + 247 + ], + "score": 1.0, + "content": "optimization methods. These benchmarks consist of seven diverse and representative FL tasks", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 118, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 118, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "involving both image and text data, with varying amounts of heterogeneity and numbers of clients.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 109, + 258, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 109, + 258, + 506, + 273 + ], + "score": 1.0, + "content": "• We demonstrate strong empirical performance of our adaptive optimizers throughout, improving", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 118, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 118, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "upon commonly used baselines. Our results show that our methods can be easier to tune, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 118, + 280, + 302, + 294 + ], + "spans": [ + { + "bbox": [ + 118, + 280, + 302, + 294 + ], + "score": 1.0, + "content": "highlight their utility in cross-device settings.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5, + "bbox_fs": [ + 109, + 150, + 506, + 294 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 506, + 312 + ], + "score": 1.0, + "content": "Related work FEDAVG was first introduced by McMahan et al. (2017), who showed it can dramati-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 312, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 323 + ], + "score": 1.0, + "content": "cally reduce communication costs. Many variants have since been proposed to tackle issues such as", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "convergence and client drift. Examples include adding a regularization term in the client objectives", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "towards the broadcast model (Li et al., 2018), and server momentum (Hsu et al., 2019). When clients", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 344, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 355 + ], + "score": 1.0, + "content": "are homogeneous, FEDAVG reduces to local SGD (Zinkevich et al., 2010), which has been analyzed", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "by many works (Stich, 2019; Yu et al., 2019; Wang & Joshi, 2018; Stich & Karimireddy, 2019; Basu", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "et al., 2019). In order to analyze FEDAVG in heterogeneous settings, many works derive convergence", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "rates depending on the amount of heterogeneity (Li et al., 2018; Wang et al., 2019; Khaled et al., 2019;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 400 + ], + "score": 1.0, + "content": "Li et al., 2019b). Typically, the convergence rate of FEDAVG gets worse with client heterogeneity.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "By using control variates to reduce client drift, the SCAFFOLD method (Karimireddy et al., 2019)", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "achieves convergence rates that are independent of the amount of heterogeneity. While effective in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "cross-silo FL, the method is incompatible with cross-device FL as it requires clients to maintain state", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 430, + 429, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 429, + 442 + ], + "score": 1.0, + "content": "across rounds. For more detailed comparisons, we defer to Kairouz et al. (2019).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 301, + 506, + 442 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 447, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 457 + ], + "score": 1.0, + "content": "Adaptive methods have been the subject of significant theoretical and empirical study, in both", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "convex (McMahan & Streeter, 2010b; Duchi et al., 2011; Kingma & Ba, 2015) and non-convex", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "settings (Li & Orabona, 2018; Ward et al., 2018; Wu et al., 2019). Reddi et al. (2019); Zaheer et al.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "(2018) study convergence failures of ADAM in certain non-convex settings, and develop an adaptive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "optimizer, YOGI, designed to improve convergence. While most work on adaptive methods focuses", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "on non-FL settings, Xie et al. (2019) propose ADAALTER, a method for FL using adaptive client", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "optimization. Conceptually, our approach is also related to the LOOKAHEAD optimizer (Zhang et al.,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "2019b), which was designed for non-FL settings. Similar to ADAALTER, an adaptive FL variant of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "LOOKAHEAD entails adaptive client optimization (see Appendix B.3 for more details). We note that", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 542, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 506, + 555 + ], + "score": 1.0, + "content": "both ADAALTER and LOOKAHEAD are, in fact, special cases of our framework (see Algorithm 1)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 506, + 567 + ], + "score": 1.0, + "content": "and the primary novelty of our work comes in focusing on adaptive server optimization. This allows", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 565, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 577 + ], + "score": 1.0, + "content": "us to avoid aggregating optimizer states across clients, making our methods require at most half as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 576, + 464, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 464, + 588 + ], + "score": 1.0, + "content": "much communication and client memory usage per round (see Appendix B.3 for details).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 446, + 506, + 588 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 165, + 605 + ], + "score": 1.0, + "content": "Notation For", + "type": "text" + }, + { + "bbox": [ + 166, + 591, + 206, + 604 + ], + "score": 0.9, + "content": "a , b \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 591, + 237, + 605 + ], + "score": 1.0, + "content": ", we let", + "type": "text" + }, + { + "bbox": [ + 238, + 591, + 266, + 604 + ], + "score": 0.91, + "content": "{ \\sqrt { a } } , a ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 591, + 285, + 605 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 286, + 592, + 302, + 604 + ], + "score": 0.89, + "content": "a / b", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 591, + 505, + 605 + ], + "score": 1.0, + "content": "denote the element-wise square root, square, and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 603, + 468, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 216, + 617 + ], + "score": 1.0, + "content": "division of the vectors. For", + "type": "text" + }, + { + "bbox": [ + 216, + 604, + 249, + 615 + ], + "score": 0.92, + "content": "\\theta _ { i } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 603, + 303, + 617 + ], + "score": 1.0, + "content": ", we use both", + "type": "text" + }, + { + "bbox": [ + 303, + 604, + 318, + 617 + ], + "score": 0.91, + "content": "\\theta _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 603, + 336, + 617 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 337, + 604, + 355, + 617 + ], + "score": 0.92, + "content": "[ \\theta _ { i } ] _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 603, + 408, + 617 + ], + "score": 1.0, + "content": "to denote its", + "type": "text" + }, + { + "bbox": [ + 408, + 604, + 420, + 616 + ], + "score": 0.88, + "content": "j ^ { \\mathrm { t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 603, + 468, + 617 + ], + "score": 1.0, + "content": "coordinate.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 591, + 505, + 617 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 631, + 318, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 319, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 319, + 646 + ], + "score": 1.0, + "content": "2 FEDERATED LEARNING AND FEDAVG", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 382, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 382, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 382, + 668 + ], + "score": 1.0, + "content": "In federated learning, we solve an optimization problem of the form:", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 655, + 382, + 668 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 671, + 360, + 704 + ], + "lines": [ + { + "bbox": [ + 250, + 671, + 360, + 704 + ], + "spans": [ + { + "bbox": [ + 250, + 671, + 360, + 704 + ], + "score": 0.94, + "content": "\\operatorname* { m i n } _ { x \\in \\mathbb { R } ^ { d } } f ( x ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } F _ { i } ( x ) ,", + "type": "interline_equation", + "image_path": "1e92002b69629c7c98895c3a52f42672bac76a3ac172efccb91bb0ec97aaca12.jpg" + } + ] + } + ], + "index": 47.5, + "virtual_lines": [ + { + "bbox": [ + 250, + 671, + 360, + 687.5 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 250, + 687.5, + 360, + 704.0 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 708, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 134, + 721 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 709, + 239, + 721 + ], + "score": 0.92, + "content": "F _ { i } ( x ) = \\mathbb { E } _ { z \\sim \\mathcal { D } _ { i } } [ f _ { i } ( x , z ) ]", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 707, + 352, + 721 + ], + "score": 1.0, + "content": ", is the loss function of the", + "type": "text" + }, + { + "bbox": [ + 352, + 708, + 363, + 719 + ], + "score": 0.87, + "content": "i ^ { \\mathrm { { t h } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 707, + 393, + 721 + ], + "score": 1.0, + "content": "client,", + "type": "text" + }, + { + "bbox": [ + 393, + 709, + 423, + 720 + ], + "score": 0.9, + "content": "z \\in { \\mathcal { Z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 707, + 445, + 721 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 446, + 709, + 458, + 720 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 707, + 506, + 721 + ], + "score": 1.0, + "content": "is the data", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 718, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 718, + 186, + 734 + ], + "score": 1.0, + "content": "distribution for the", + "type": "text" + }, + { + "bbox": [ + 186, + 721, + 197, + 730 + ], + "score": 0.84, + "content": "i ^ { \\mathrm { { t h } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 718, + 243, + 734 + ], + "score": 1.0, + "content": "client. For", + "type": "text" + }, + { + "bbox": [ + 244, + 721, + 267, + 732 + ], + "score": 0.83, + "content": "i \\neq j", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 718, + 272, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 272, + 721, + 284, + 732 + ], + "score": 0.72, + "content": "\\mathcal { D } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 718, + 303, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 304, + 721, + 316, + 733 + ], + "score": 0.9, + "content": "\\mathcal { D } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 718, + 471, + 734 + ], + "score": 1.0, + "content": "may be very different. The functions", + "type": "text" + }, + { + "bbox": [ + 472, + 721, + 483, + 732 + ], + "score": 0.88, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 718, + 506, + 734 + ], + "score": 1.0, + "content": "(and", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 144, + 95 + ], + "score": 1.0, + "content": "therefore", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 144, + 83, + 151, + 94 + ], + "score": 0.81, + "content": "f", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 151, + 82, + 267, + 95 + ], + "score": 1.0, + "content": ") may be nonconvex. For each", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 268, + 84, + 272, + 92 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 273, + 82, + 289, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 289, + 85, + 296, + 92 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 297, + 82, + 505, + 95 + ], + "score": 1.0, + "content": ", we assume access to an unbiased stochastic gradient", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 92, + 474, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 129, + 106 + ], + "score": 0.91, + "content": "g _ { i } ( x )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 129, + 92, + 240, + 107 + ], + "score": 1.0, + "content": "of the client’s true gradient", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 240, + 93, + 273, + 105 + ], + "score": 0.93, + "content": "\\nabla F _ { i } ( x )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 273, + 92, + 474, + 107 + ], + "score": 1.0, + "content": ". In addition, we make the following assumptions.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 707, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 144, + 95 + ], + "score": 1.0, + "content": "therefore", + "type": "text" + }, + { + "bbox": [ + 144, + 83, + 151, + 94 + ], + "score": 0.81, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 82, + 267, + 95 + ], + "score": 1.0, + "content": ") may be nonconvex. For each", + "type": "text" + }, + { + "bbox": [ + 268, + 84, + 272, + 92 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 82, + 289, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 289, + 85, + 296, + 92 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 82, + 505, + 95 + ], + "score": 1.0, + "content": ", we assume access to an unbiased stochastic gradient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 92, + 474, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 129, + 106 + ], + "score": 0.91, + "content": "g _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 92, + 240, + 107 + ], + "score": 1.0, + "content": "of the client’s true gradient", + "type": "text" + }, + { + "bbox": [ + 240, + 93, + 273, + 105 + ], + "score": 0.93, + "content": "\\nabla F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 92, + 474, + 107 + ], + "score": 1.0, + "content": ". In addition, we make the following assumptions.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 108, + 504, + 132 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 312, + 122 + ], + "score": 1.0, + "content": "Assumption 1 (Lipschitz Gradient). The function", + "type": "text" + }, + { + "bbox": [ + 312, + 109, + 323, + 119 + ], + "score": 0.88, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 107, + 334, + 122 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 334, + 109, + 342, + 118 + ], + "score": 0.77, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 107, + 404, + 122 + ], + "score": 1.0, + "content": "-smooth for all", + "type": "text" + }, + { + "bbox": [ + 405, + 108, + 437, + 120 + ], + "score": 0.93, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 107, + 456, + 122 + ], + "score": 1.0, + "content": "i.e.,", + "type": "text" + }, + { + "bbox": [ + 456, + 108, + 505, + 121 + ], + "score": 0.91, + "content": "\\parallel \\nabla F _ { i } ( x ) -", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 119, + 272, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 198, + 133 + ], + "score": 0.91, + "content": "\\nabla F _ { i } ( y ) \\| \\leq L \\| x - y \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 119, + 228, + 132 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 228, + 119, + 268, + 132 + ], + "score": 0.92, + "content": "x , y \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 119, + 272, + 132 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 135, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 507, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 325, + 148 + ], + "score": 1.0, + "content": "Assumption 2 (Bounded Variance). The function", + "type": "text" + }, + { + "bbox": [ + 325, + 136, + 336, + 146 + ], + "score": 0.84, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 133, + 363, + 148 + ], + "score": 1.0, + "content": "have", + "type": "text" + }, + { + "bbox": [ + 364, + 136, + 374, + 146 + ], + "score": 0.84, + "content": "\\sigma _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 133, + 507, + 148 + ], + "score": 1.0, + "content": "-bounded (local) variance i.e.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 145, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 270, + 159 + ], + "score": 0.9, + "content": "\\mathbb { E } [ \\| \\nabla [ \\widehat { f } _ { i } ( x , z ) ] _ { j } - [ \\nabla F _ { i } ( x ) ] _ { j } \\| ^ { 2 } ] = \\sigma _ { l , j } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 145, + 301, + 159 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 302, + 146, + 335, + 157 + ], + "score": 0.85, + "content": "x \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 145, + 340, + 159 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 340, + 146, + 372, + 158 + ], + "score": 0.86, + "content": "j \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 145, + 393, + 159 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 394, + 146, + 428, + 158 + ], + "score": 0.91, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 145, + 506, + 159 + ], + "score": 1.0, + "content": ". Furthermore, we", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 156, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 104, + 156, + 283, + 174 + ], + "score": 1.0, + "content": "assume the (global) variance is bounded,", + "type": "text" + }, + { + "bbox": [ + 283, + 158, + 474, + 173 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\left( 1 / m \\right) \\sum _ { i = 1 } ^ { m } \\| \\nabla [ F _ { i } ( x ) ] _ { j } - [ \\nabla f ( x ) ] _ { j } \\| ^ { 2 } \\leq \\sigma _ { g , j } ^ { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 156, + 506, + 174 + ], + "score": 1.0, + "content": "for all", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 170, + 189, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 137, + 183 + ], + "score": 0.9, + "content": "\\boldsymbol { x } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 170, + 156, + 186 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 172, + 185, + 185 + ], + "score": 0.91, + "content": "j \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 170, + 189, + 186 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 504, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 506, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 312, + 202 + ], + "score": 1.0, + "content": "Assumption 3 (Bounded Gradients). The function", + "type": "text" + }, + { + "bbox": [ + 313, + 187, + 345, + 199 + ], + "score": 0.93, + "content": "f _ { i } ( x , z )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 185, + 367, + 202 + ], + "score": 1.0, + "content": "have", + "type": "text" + }, + { + "bbox": [ + 368, + 188, + 377, + 198 + ], + "score": 0.74, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 185, + 506, + 202 + ], + "score": 1.0, + "content": "-bounded gradients i.e., for any", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 198, + 391, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 137, + 212 + ], + "score": 0.88, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 198, + 141, + 213 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 141, + 199, + 172, + 210 + ], + "score": 0.89, + "content": "\\boldsymbol { x } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 198, + 191, + 213 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 191, + 200, + 218, + 210 + ], + "score": 0.89, + "content": "z \\in { \\mathcal { Z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 198, + 254, + 213 + ], + "score": 1.0, + "content": "we have", + "type": "text" + }, + { + "bbox": [ + 254, + 199, + 330, + 212 + ], + "score": 0.91, + "content": "| [ \\nabla f _ { i } ( x , z ) ] _ { j } | \\leq G", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 198, + 359, + 213 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 359, + 199, + 387, + 212 + ], + "score": 0.91, + "content": "j \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 198, + 391, + 213 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 101, + 218, + 509, + 241 + ], + "spans": [ + { + "bbox": [ + 101, + 218, + 258, + 241 + ], + "score": 1.0, + "content": "With a slight abuse of notation, we use", + "type": "text" + }, + { + "bbox": [ + 258, + 223, + 270, + 236 + ], + "score": 0.9, + "content": "\\sigma _ { l } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 218, + 287, + 241 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 287, + 223, + 299, + 237 + ], + "score": 0.9, + "content": "\\sigma _ { g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 218, + 338, + 241 + ], + "score": 1.0, + "content": "to denote", + "type": "text" + }, + { + "bbox": [ + 338, + 221, + 380, + 237 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 218, + 398, + 241 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 221, + 442, + 237 + ], + "score": 0.91, + "content": "\\textstyle \\sum _ { j = 1 } ^ { d } \\sigma _ { g , j } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 218, + 509, + 241 + ], + "score": 1.0, + "content": ". Assumptions 1", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 235, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 247 + ], + "score": 1.0, + "content": "and 3 are fairly standard in nonconvex optimization literature (Reddi et al., 2016; Ward et al., 2018;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "Zaheer et al., 2018). We make no further assumptions regarding the similarity of clients datasets.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 258, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 269 + ], + "score": 1.0, + "content": "Assumption 2 is a form of bounded variance, but between the client objective functions and the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 268, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 279 + ], + "score": 1.0, + "content": "overall objective function. This assumption has been used throughout various works on federated", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 400, + 291 + ], + "score": 1.0, + "content": "optimization (Li et al., 2018; Wang et al., 2019). Intuitively, the parameter", + "type": "text" + }, + { + "bbox": [ + 401, + 280, + 412, + 291 + ], + "score": 0.86, + "content": "\\sigma _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 279, + 506, + 291 + ], + "score": 1.0, + "content": "quantifies similarity of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 288, + 394, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 233, + 303 + ], + "score": 1.0, + "content": "client objective functions. Note", + "type": "text" + }, + { + "bbox": [ + 234, + 290, + 263, + 302 + ], + "score": 0.91, + "content": "\\sigma _ { g } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 288, + 394, + 303 + ], + "score": 1.0, + "content": "corresponds to the i.i.d. setting.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "A common approach to solving (1) in federated settings is FEDAVG (McMahan et al., 2017). At each", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "score": 1.0, + "content": "round of FEDAVG, a subset of clients are selected (typically randomly) and the server broadcasts its", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "global model to each client. In parallel, the clients run SGD on their own loss function, and send the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "resulting model to the server. The server then updates its global model as the average of these local", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 340, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 340, + 362 + ], + "score": 1.0, + "content": "models. See Algorithm 3 in the appendix for more details.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 503, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 198, + 378 + ], + "score": 1.0, + "content": "Suppose that at round", + "type": "text" + }, + { + "bbox": [ + 198, + 367, + 203, + 376 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 365, + 293, + 378 + ], + "score": 1.0, + "content": ", the server has model", + "type": "text" + }, + { + "bbox": [ + 294, + 367, + 304, + 377 + ], + "score": 0.85, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 365, + 379, + 378 + ], + "score": 1.0, + "content": "and samples a set", + "type": "text" + }, + { + "bbox": [ + 380, + 366, + 388, + 376 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 365, + 448, + 378 + ], + "score": 1.0, + "content": "of clients. Let", + "type": "text" + }, + { + "bbox": [ + 448, + 366, + 459, + 378 + ], + "score": 0.89, + "content": "\\ v x _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "denote the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 376, + 425, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 190, + 389 + ], + "score": 1.0, + "content": "model of each client", + "type": "text" + }, + { + "bbox": [ + 190, + 377, + 214, + 387 + ], + "score": 0.9, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 376, + 425, + 389 + ], + "score": 1.0, + "content": "after local training. We rewrite FEDAVG’s update as", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 393, + 399, + 424 + ], + "lines": [ + { + "bbox": [ + 211, + 393, + 399, + 424 + ], + "spans": [ + { + "bbox": [ + 211, + 393, + 399, + 424 + ], + "score": 0.93, + "content": "x _ { t + 1 } = { \\frac { 1 } { | S | } } \\sum _ { i \\in S } x _ { i } ^ { t } = x _ { t } - { \\frac { 1 } { | S | } } \\sum _ { i \\in S } \\left( x _ { t } - x _ { i } ^ { t } \\right) .", + "type": "interline_equation", + "image_path": "43e92881083d9a5561a6396b1be0718d263b1f1718f9e7b7d644b230f97efb05.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 211, + 393, + 399, + 408.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 211, + 408.5, + 399, + 424.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 505, + 485 + ], + "lines": [ + { + "bbox": [ + 104, + 427, + 507, + 446 + ], + "spans": [ + { + "bbox": [ + 104, + 427, + 122, + 446 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 430, + 182, + 442 + ], + "score": 0.92, + "content": "\\Delta _ { i } ^ { t } : = x _ { i } ^ { t } - x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 427, + 200, + 446 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 200, + 429, + 294, + 443 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\Delta _ { t } : = \\left( { 1 } / { | { S } | } \\right) \\sum _ { i \\in \\mathcal { S } } \\Delta _ { i } ^ { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 427, + 507, + 446 + ], + "score": 1.0, + "content": ". Then the server update in FEDAVG is equivalent to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 441, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 269, + 454 + ], + "score": 1.0, + "content": "applying SGD to the “pseudo-gradient”", + "type": "text" + }, + { + "bbox": [ + 270, + 442, + 290, + 452 + ], + "score": 0.87, + "content": "\\bar { - } \\Delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 441, + 366, + 454 + ], + "score": 1.0, + "content": "with learning rate", + "type": "text" + }, + { + "bbox": [ + 367, + 442, + 393, + 453 + ], + "score": 0.91, + "content": "\\eta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 441, + 506, + 454 + ], + "score": 1.0, + "content": ". This formulation makes it", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 451, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 214, + 465 + ], + "score": 1.0, + "content": "clear that other choices of", + "type": "text" + }, + { + "bbox": [ + 214, + 454, + 221, + 464 + ], + "score": 0.81, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 451, + 506, + 465 + ], + "score": 1.0, + "content": "are possible. One could also utilize optimizers other than SGD on the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "clients, or use an alternative update rule on the server. This family of algorithms, which we refer to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 473, + 325, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 325, + 485 + ], + "score": 1.0, + "content": "collectively as FEDOPT, is formalized in Algorithm 1.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "title", + "bbox": [ + 106, + 497, + 198, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 199, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 199, + 511 + ], + "score": 1.0, + "content": "Algorithm 1 FEDOPT", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 109, + 511, + 366, + 648 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 109, + 511, + 366, + 648 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 511, + 366, + 648 + ], + "spans": [ + { + "bbox": [ + 109, + 511, + 366, + 648 + ], + "score": 0.9, + "html": "
1: Input:xo,CLIENTOPT,SERVEROPT
2: fort=O,.,T-1do
3: Samplea subset Sofclients
4: xt,0=xt
5: for each client i ∈ S in parallel do
6: for k=0,.,K-1 do
Compute an unbiased estimate gt,k of VFi(xt,k) 7:
x,+1=CLEO(,,,t) 8:
9: △=x,K-xt
10: △t=∑ies△t 11: Xt+1 = SERVEROPT(xt,-△t,n,t)
", + "type": "table", + "image_path": "2592fd4b1598c7bfe3873a636a0f2dff92e8d381323eec82ae75910138d9078b.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 109, + 511, + 366, + 556.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 109, + 556.6666666666666, + 366, + 602.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 109, + 602.3333333333333, + 366, + 647.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "In Algorithm 1, CLIENTOPT and SERVEROPT are gradient-based optimizers with learning rates", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 115, + 689 + ], + "score": 0.81, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 677, + 135, + 690 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 136, + 680, + 142, + 689 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "respectively. Intuitively, CLIENTOPT aims to minimize (1) based on each client’s local", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "data while SERVEROPT optimizes from a global perspective. FEDOPT naturally allows the use of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "adaptive optimizers (eg. ADAM, YOGI, etc.), as well as techniques such as server-side momentum", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 711, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 722 + ], + "score": 1.0, + "content": "(leading to FEDAVGM, proposed by Hsu et al. (2019)). In its most general form, FEDOPT uses a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 721, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 734 + ], + "score": 1.0, + "content": "CLIENTOPT whose updates can depend on globally aggregated statistics (e.g. server updates in the", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 107 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 108, + 504, + 132 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 312, + 122 + ], + "score": 1.0, + "content": "Assumption 1 (Lipschitz Gradient). The function", + "type": "text" + }, + { + "bbox": [ + 312, + 109, + 323, + 119 + ], + "score": 0.88, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 107, + 334, + 122 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 334, + 109, + 342, + 118 + ], + "score": 0.77, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 107, + 404, + 122 + ], + "score": 1.0, + "content": "-smooth for all", + "type": "text" + }, + { + "bbox": [ + 405, + 108, + 437, + 120 + ], + "score": 0.93, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 107, + 456, + 122 + ], + "score": 1.0, + "content": "i.e.,", + "type": "text" + }, + { + "bbox": [ + 456, + 108, + 505, + 121 + ], + "score": 0.91, + "content": "\\parallel \\nabla F _ { i } ( x ) -", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 119, + 272, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 198, + 133 + ], + "score": 0.91, + "content": "\\nabla F _ { i } ( y ) \\| \\leq L \\| x - y \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 119, + 228, + 132 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 228, + 119, + 268, + 132 + ], + "score": 0.92, + "content": "x , y \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 119, + 272, + 132 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 107, + 505, + 133 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 135, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 507, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 325, + 148 + ], + "score": 1.0, + "content": "Assumption 2 (Bounded Variance). The function", + "type": "text" + }, + { + "bbox": [ + 325, + 136, + 336, + 146 + ], + "score": 0.84, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 133, + 363, + 148 + ], + "score": 1.0, + "content": "have", + "type": "text" + }, + { + "bbox": [ + 364, + 136, + 374, + 146 + ], + "score": 0.84, + "content": "\\sigma _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 133, + 507, + 148 + ], + "score": 1.0, + "content": "-bounded (local) variance i.e.,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 145, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 270, + 159 + ], + "score": 0.9, + "content": "\\mathbb { E } [ \\| \\nabla [ \\widehat { f } _ { i } ( x , z ) ] _ { j } - [ \\nabla F _ { i } ( x ) ] _ { j } \\| ^ { 2 } ] = \\sigma _ { l , j } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 145, + 301, + 159 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 302, + 146, + 335, + 157 + ], + "score": 0.85, + "content": "x \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 145, + 340, + 159 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 340, + 146, + 372, + 158 + ], + "score": 0.86, + "content": "j \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 145, + 393, + 159 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 394, + 146, + 428, + 158 + ], + "score": 0.91, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 145, + 506, + 159 + ], + "score": 1.0, + "content": ". Furthermore, we", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 156, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 104, + 156, + 283, + 174 + ], + "score": 1.0, + "content": "assume the (global) variance is bounded,", + "type": "text" + }, + { + "bbox": [ + 283, + 158, + 474, + 173 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\left( 1 / m \\right) \\sum _ { i = 1 } ^ { m } \\| \\nabla [ F _ { i } ( x ) ] _ { j } - [ \\nabla f ( x ) ] _ { j } \\| ^ { 2 } \\leq \\sigma _ { g , j } ^ { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 156, + 506, + 174 + ], + "score": 1.0, + "content": "for all", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 170, + 189, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 137, + 183 + ], + "score": 0.9, + "content": "\\boldsymbol { x } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 170, + 156, + 186 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 172, + 185, + 185 + ], + "score": 0.91, + "content": "j \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 170, + 189, + 186 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 133, + 507, + 186 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 504, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 185, + 506, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 312, + 202 + ], + "score": 1.0, + "content": "Assumption 3 (Bounded Gradients). The function", + "type": "text" + }, + { + "bbox": [ + 313, + 187, + 345, + 199 + ], + "score": 0.93, + "content": "f _ { i } ( x , z )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 185, + 367, + 202 + ], + "score": 1.0, + "content": "have", + "type": "text" + }, + { + "bbox": [ + 368, + 188, + 377, + 198 + ], + "score": 0.74, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 185, + 506, + 202 + ], + "score": 1.0, + "content": "-bounded gradients i.e., for any", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 198, + 391, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 137, + 212 + ], + "score": 0.88, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 198, + 141, + 213 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 141, + 199, + 172, + 210 + ], + "score": 0.89, + "content": "\\boldsymbol { x } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 198, + 191, + 213 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 191, + 200, + 218, + 210 + ], + "score": 0.89, + "content": "z \\in { \\mathcal { Z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 198, + 254, + 213 + ], + "score": 1.0, + "content": "we have", + "type": "text" + }, + { + "bbox": [ + 254, + 199, + 330, + 212 + ], + "score": 0.91, + "content": "| [ \\nabla f _ { i } ( x , z ) ] _ { j } | \\leq G", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 198, + 359, + 213 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 359, + 199, + 387, + 212 + ], + "score": 0.91, + "content": "j \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 198, + 391, + 213 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 185, + 506, + 213 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 101, + 218, + 509, + 241 + ], + "spans": [ + { + "bbox": [ + 101, + 218, + 258, + 241 + ], + "score": 1.0, + "content": "With a slight abuse of notation, we use", + "type": "text" + }, + { + "bbox": [ + 258, + 223, + 270, + 236 + ], + "score": 0.9, + "content": "\\sigma _ { l } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 218, + 287, + 241 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 287, + 223, + 299, + 237 + ], + "score": 0.9, + "content": "\\sigma _ { g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 218, + 338, + 241 + ], + "score": 1.0, + "content": "to denote", + "type": "text" + }, + { + "bbox": [ + 338, + 221, + 380, + 237 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 218, + 398, + 241 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 221, + 442, + 237 + ], + "score": 0.91, + "content": "\\textstyle \\sum _ { j = 1 } ^ { d } \\sigma _ { g , j } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 218, + 509, + 241 + ], + "score": 1.0, + "content": ". Assumptions 1", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 235, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 247 + ], + "score": 1.0, + "content": "and 3 are fairly standard in nonconvex optimization literature (Reddi et al., 2016; Ward et al., 2018;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "Zaheer et al., 2018). We make no further assumptions regarding the similarity of clients datasets.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 258, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 269 + ], + "score": 1.0, + "content": "Assumption 2 is a form of bounded variance, but between the client objective functions and the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 268, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 279 + ], + "score": 1.0, + "content": "overall objective function. This assumption has been used throughout various works on federated", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 400, + 291 + ], + "score": 1.0, + "content": "optimization (Li et al., 2018; Wang et al., 2019). Intuitively, the parameter", + "type": "text" + }, + { + "bbox": [ + 401, + 280, + 412, + 291 + ], + "score": 0.86, + "content": "\\sigma _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 279, + 506, + 291 + ], + "score": 1.0, + "content": "quantifies similarity of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 288, + 394, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 233, + 303 + ], + "score": 1.0, + "content": "client objective functions. Note", + "type": "text" + }, + { + "bbox": [ + 234, + 290, + 263, + 302 + ], + "score": 0.91, + "content": "\\sigma _ { g } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 288, + 394, + 303 + ], + "score": 1.0, + "content": "corresponds to the i.i.d. setting.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13, + "bbox_fs": [ + 101, + 218, + 509, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "A common approach to solving (1) in federated settings is FEDAVG (McMahan et al., 2017). At each", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 329 + ], + "score": 1.0, + "content": "round of FEDAVG, a subset of clients are selected (typically randomly) and the server broadcasts its", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "global model to each client. In parallel, the clients run SGD on their own loss function, and send the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "resulting model to the server. The server then updates its global model as the average of these local", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 340, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 340, + 362 + ], + "score": 1.0, + "content": "models. See Algorithm 3 in the appendix for more details.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 306, + 505, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 503, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 198, + 378 + ], + "score": 1.0, + "content": "Suppose that at round", + "type": "text" + }, + { + "bbox": [ + 198, + 367, + 203, + 376 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 365, + 293, + 378 + ], + "score": 1.0, + "content": ", the server has model", + "type": "text" + }, + { + "bbox": [ + 294, + 367, + 304, + 377 + ], + "score": 0.85, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 365, + 379, + 378 + ], + "score": 1.0, + "content": "and samples a set", + "type": "text" + }, + { + "bbox": [ + 380, + 366, + 388, + 376 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 365, + 448, + 378 + ], + "score": 1.0, + "content": "of clients. Let", + "type": "text" + }, + { + "bbox": [ + 448, + 366, + 459, + 378 + ], + "score": 0.89, + "content": "\\ v x _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "denote the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 376, + 425, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 190, + 389 + ], + "score": 1.0, + "content": "model of each client", + "type": "text" + }, + { + "bbox": [ + 190, + 377, + 214, + 387 + ], + "score": 0.9, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 376, + 425, + 389 + ], + "score": 1.0, + "content": "after local training. We rewrite FEDAVG’s update as", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 365, + 505, + 389 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 393, + 399, + 424 + ], + "lines": [ + { + "bbox": [ + 211, + 393, + 399, + 424 + ], + "spans": [ + { + "bbox": [ + 211, + 393, + 399, + 424 + ], + "score": 0.93, + "content": "x _ { t + 1 } = { \\frac { 1 } { | S | } } \\sum _ { i \\in S } x _ { i } ^ { t } = x _ { t } - { \\frac { 1 } { | S | } } \\sum _ { i \\in S } \\left( x _ { t } - x _ { i } ^ { t } \\right) .", + "type": "interline_equation", + "image_path": "43e92881083d9a5561a6396b1be0718d263b1f1718f9e7b7d644b230f97efb05.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 211, + 393, + 399, + 408.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 211, + 408.5, + 399, + 424.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 505, + 485 + ], + "lines": [ + { + "bbox": [ + 104, + 427, + 507, + 446 + ], + "spans": [ + { + "bbox": [ + 104, + 427, + 122, + 446 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 430, + 182, + 442 + ], + "score": 0.92, + "content": "\\Delta _ { i } ^ { t } : = x _ { i } ^ { t } - x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 427, + 200, + 446 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 200, + 429, + 294, + 443 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\Delta _ { t } : = \\left( { 1 } / { | { S } | } \\right) \\sum _ { i \\in \\mathcal { S } } \\Delta _ { i } ^ { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 427, + 507, + 446 + ], + "score": 1.0, + "content": ". Then the server update in FEDAVG is equivalent to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 441, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 269, + 454 + ], + "score": 1.0, + "content": "applying SGD to the “pseudo-gradient”", + "type": "text" + }, + { + "bbox": [ + 270, + 442, + 290, + 452 + ], + "score": 0.87, + "content": "\\bar { - } \\Delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 441, + 366, + 454 + ], + "score": 1.0, + "content": "with learning rate", + "type": "text" + }, + { + "bbox": [ + 367, + 442, + 393, + 453 + ], + "score": 0.91, + "content": "\\eta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 441, + 506, + 454 + ], + "score": 1.0, + "content": ". This formulation makes it", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 451, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 214, + 465 + ], + "score": 1.0, + "content": "clear that other choices of", + "type": "text" + }, + { + "bbox": [ + 214, + 454, + 221, + 464 + ], + "score": 0.81, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 451, + 506, + 465 + ], + "score": 1.0, + "content": "are possible. One could also utilize optimizers other than SGD on the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "clients, or use an alternative update rule on the server. This family of algorithms, which we refer to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 473, + 325, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 325, + 485 + ], + "score": 1.0, + "content": "collectively as FEDOPT, is formalized in Algorithm 1.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 427, + 507, + 485 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 497, + 198, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 199, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 199, + 511 + ], + "score": 1.0, + "content": "Algorithm 1 FEDOPT", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 109, + 511, + 366, + 648 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 109, + 511, + 366, + 648 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 511, + 366, + 648 + ], + "spans": [ + { + "bbox": [ + 109, + 511, + 366, + 648 + ], + "score": 0.9, + "html": "
1: Input:xo,CLIENTOPT,SERVEROPT
2: fort=O,.,T-1do
3: Samplea subset Sofclients
4: xt,0=xt
5: for each client i ∈ S in parallel do
6: for k=0,.,K-1 do
Compute an unbiased estimate gt,k of VFi(xt,k) 7:
x,+1=CLEO(,,,t) 8:
9: △=x,K-xt
10: △t=∑ies△t 11: Xt+1 = SERVEROPT(xt,-△t,n,t)
", + "type": "table", + "image_path": "2592fd4b1598c7bfe3873a636a0f2dff92e8d381323eec82ae75910138d9078b.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 109, + 511, + 366, + 556.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 109, + 556.6666666666666, + 366, + 602.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 109, + 602.3333333333333, + 366, + 647.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "In Algorithm 1, CLIENTOPT and SERVEROPT are gradient-based optimizers with learning rates", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 115, + 689 + ], + "score": 0.81, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 677, + 135, + 690 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 136, + 680, + 142, + 689 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "respectively. Intuitively, CLIENTOPT aims to minimize (1) based on each client’s local", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "data while SERVEROPT optimizes from a global perspective. FEDOPT naturally allows the use of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "adaptive optimizers (eg. ADAM, YOGI, etc.), as well as techniques such as server-side momentum", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 711, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 722 + ], + "score": 1.0, + "content": "(leading to FEDAVGM, proposed by Hsu et al. (2019)). In its most general form, FEDOPT uses a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 721, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 734 + ], + "score": 1.0, + "content": "CLIENTOPT whose updates can depend on globally aggregated statistics (e.g. server updates in the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 245, + 96 + ], + "score": 1.0, + "content": "previous iterations). We also allow", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 246, + 84, + 253, + 94 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 253, + 81, + 270, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 270, + 84, + 280, + 94 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 280, + 81, + 374, + 96 + ], + "score": 1.0, + "content": "to depend on the round", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 375, + 84, + 380, + 92 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 380, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "in order to encompass learning", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "rate schedules. While we focus on specific adaptive optimizers in this work, we can in principle use", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 486, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 486, + 117 + ], + "score": 1.0, + "content": "any adaptive optimizer (e.g. AMSGRAD (Reddi et al., 2019), ADABOUND (Luo et al., 2019)).", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 666, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 245, + 96 + ], + "score": 1.0, + "content": "previous iterations). We also allow", + "type": "text" + }, + { + "bbox": [ + 246, + 84, + 253, + 94 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 81, + 270, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 270, + 84, + 280, + 94 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 81, + 374, + 96 + ], + "score": 1.0, + "content": "to depend on the round", + "type": "text" + }, + { + "bbox": [ + 375, + 84, + 380, + 92 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "in order to encompass learning", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "rate schedules. While we focus on specific adaptive optimizers in this work, we can in principle use", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 486, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 486, + 117 + ], + "score": 1.0, + "content": "any adaptive optimizer (e.g. AMSGRAD (Reddi et al., 2019), ADABOUND (Luo et al., 2019)).", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 120, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "While FEDOPT has intuitive benefits over FEDAVG, it also raises a fundamental question: Can the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 266, + 144 + ], + "score": 1.0, + "content": "negative of the average model difference", + "type": "text" + }, + { + "bbox": [ + 267, + 132, + 279, + 142 + ], + "score": 0.87, + "content": "\\Delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "be used as a pseudo-gradient in general server optimizer", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "updates? In this paper, we provide an affirmative answer to this question by establishing a theoretical", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 166 + ], + "score": 1.0, + "content": "basis for FEDOPT. We will show that the use of the term SERVEROPT is justified, as we can guarantee", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 506, + 176 + ], + "score": 1.0, + "content": "convergence across a wide variety of server optimizers, including ADAGRAD, ADAM, and YOGI,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 425, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 425, + 186 + ], + "score": 1.0, + "content": "thus developing principled adaptive optimizers for FL based on our framework.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 108, + 201, + 324, + 214 + ], + "lines": [ + { + "bbox": [ + 104, + 200, + 326, + 216 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 326, + 216 + ], + "score": 1.0, + "content": "3 ADAPTIVE FEDERATED OPTIMIZATION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "In this section, we specialize FEDOPT to settings where SERVEROPT is an adaptive optimization", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "score": 1.0, + "content": "method (one of ADAGRAD, YOGI or ADAM) and CLIENTOPT is SGD. By using adaptive methods", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 260 + ], + "score": 1.0, + "content": "(which generally require maintaining state) on the server and SGD on the clients, we ensure our", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 471, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 471, + 272 + ], + "score": 1.0, + "content": "methods have the same communication cost as FEDAVG and work in cross-device settings.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "Algorithm 2 provides pseudo-code for our methods. An alternate version using batched data and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "example-based weighting (as opposed to uniform weighting) of clients is given in Algorithm 5. The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 147, + 309 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 148, + 299, + 155, + 307 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 297, + 419, + 309 + ], + "score": 1.0, + "content": "controls the algorithms’ degree of adaptivity, with smaller values of", + "type": "text" + }, + { + "bbox": [ + 420, + 299, + 426, + 307 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "representing higher", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 308, + 504, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 504, + 320 + ], + "score": 1.0, + "content": "degrees of adaptivity. Note that the server updates of our methods are invariant to fixed multiplicative", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 244, + 331 + ], + "score": 1.0, + "content": "changes to the client learning rate", + "type": "text" + }, + { + "bbox": [ + 244, + 320, + 253, + 330 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 318, + 355, + 331 + ], + "score": 1.0, + "content": "for appropriately chosen", + "type": "text" + }, + { + "bbox": [ + 356, + 320, + 362, + 328 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 318, + 505, + 331 + ], + "score": 1.0, + "content": ", though as we shall see shortly, we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 329, + 322, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 154, + 342 + ], + "score": 1.0, + "content": "will require", + "type": "text" + }, + { + "bbox": [ + 155, + 331, + 164, + 341 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 329, + 322, + 342 + ], + "score": 1.0, + "content": "to be sufficiently small in our analysis.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 106, + 354, + 352, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 353, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 228, + 369 + ], + "score": 1.0, + "content": "Algorithm 2 FEDADAGRAD", + "type": "text" + }, + { + "bbox": [ + 232, + 352, + 353, + 370 + ], + "score": 1.0, + "content": "FEDYOGI , and FEDADAM", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 370, + 369, + 569 + ], + "lines": [ + { + "bbox": [ + 110, + 369, + 369, + 385 + ], + "spans": [ + { + "bbox": [ + 110, + 369, + 181, + 385 + ], + "score": 1.0, + "content": "1: Initialization:", + "type": "text" + }, + { + "bbox": [ + 181, + 371, + 234, + 383 + ], + "score": 0.88, + "content": "x _ { 0 } , v _ { - 1 } \\ge \\tau ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 369, + 309, + 385 + ], + "score": 1.0, + "content": ", decay parameters", + "type": "text" + }, + { + "bbox": [ + 310, + 371, + 369, + 384 + ], + "score": 0.91, + "content": "\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 109, + 381, + 225, + 394 + ], + "spans": [ + { + "bbox": [ + 109, + 381, + 138, + 393 + ], + "score": 1.0, + "content": "2: for", + "type": "text" + }, + { + "bbox": [ + 138, + 383, + 209, + 394 + ], + "score": 0.88, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 381, + 225, + 393 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 110, + 393, + 247, + 405 + ], + "spans": [ + { + "bbox": [ + 110, + 393, + 122, + 405 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 137, + 393, + 198, + 405 + ], + "score": 1.0, + "content": "Sample subset", + "type": "text" + }, + { + "bbox": [ + 198, + 394, + 206, + 403 + ], + "score": 0.8, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 393, + 247, + 405 + ], + "score": 1.0, + "content": "of clients", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 110, + 401, + 182, + 421 + ], + "spans": [ + { + "bbox": [ + 110, + 403, + 122, + 416 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 138, + 403, + 178, + 417 + ], + "score": 0.5, + "content": "x _ { i , 0 } ^ { t } \\dot { = } x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 401, + 182, + 421 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 110, + 415, + 282, + 427 + ], + "spans": [ + { + "bbox": [ + 110, + 415, + 121, + 427 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 137, + 415, + 199, + 427 + ], + "score": 1.0, + "content": "for each client", + "type": "text" + }, + { + "bbox": [ + 199, + 416, + 222, + 426 + ], + "score": 0.88, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 415, + 282, + 427 + ], + "score": 1.0, + "content": "in parallel do", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 110, + 426, + 258, + 438 + ], + "spans": [ + { + "bbox": [ + 110, + 426, + 121, + 438 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 153, + 426, + 168, + 437 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 168, + 426, + 243, + 438 + ], + "score": 0.89, + "content": "k = 0 , \\cdots , K - \\bar { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 426, + 258, + 437 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 110, + 434, + 363, + 452 + ], + "spans": [ + { + "bbox": [ + 110, + 437, + 122, + 450 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 166, + 434, + 292, + 452 + ], + "score": 1.0, + "content": "Compute an unbiased estimate", + "type": "text" + }, + { + "bbox": [ + 293, + 437, + 309, + 451 + ], + "score": 0.91, + "content": "g _ { i , k } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 434, + 321, + 452 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 321, + 437, + 363, + 451 + ], + "score": 0.91, + "content": "\\nabla F _ { i } ( x _ { i , k } ^ { t } )", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 110, + 449, + 311, + 498 + ], + "spans": [ + { + "bbox": [ + 110, + 450, + 122, + 463 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 137, + 449, + 311, + 498 + ], + "score": 0.84, + "content": "\\begin{array} { c } { x _ { i , k + 1 } ^ { t } = x _ { i , k } ^ { t } - \\eta _ { l } g _ { i , k } ^ { t } } \\\\ { \\Delta _ { i } ^ { t } = x _ { i , K } ^ { t } - x _ { t } } \\\\ { \\Delta _ { t } = \\beta _ { 1 } \\Delta _ { t - 1 } + \\left( 1 - \\beta _ { 1 } \\right) \\left( \\frac { 1 } { \\left| S \\right| } \\sum _ { i \\in S } \\Delta _ { i } ^ { t } \\right) } \\end{array}", + "type": "inline_equation", + "image_path": "64a4df90b9ef875531f06b07407ff8cc2911e941ad716e76b11347461ef989cc.jpg" + } + ], + "index": 29 + }, + { + "bbox": [ + 109, + 464, + 122, + 479 + ], + "spans": [ + { + "bbox": [ + 109, + 464, + 122, + 479 + ], + "score": 1.0, + "content": "9:", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 481, + 122, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 122, + 496 + ], + "score": 1.0, + "content": "10:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 499, + 281, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 122, + 514 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 140, + 500, + 208, + 513 + ], + "score": 0.71, + "content": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 499, + 281, + 513 + ], + "score": 1.0, + "content": "(FEDADAGRAD)", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 515, + 367, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 122, + 532 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 138, + 515, + 367, + 568 + ], + "score": 0.7, + "content": "\\begin{array} { r l } & { v _ { t } = v _ { t - 1 } - ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\mathrm { s i g n } ( v _ { t - 1 } - \\Delta _ { t } ^ { 2 } ) \\ ( \\mathrm { F E D Y O G I } ) } \\\\ & { v _ { t } = \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\left( \\mathrm { F E D A D A M } \\right) } \\\\ & { x _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } } \\end{array}", + "type": "inline_equation", + "image_path": "32ee17883ba5478b7ac54f615779eedc5713aae3bffd1c1a8bbdb8e1021bffff.jpg" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 534, + 122, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 122, + 549 + ], + "score": 1.0, + "content": "13:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 550, + 122, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 122, + 564 + ], + "score": 1.0, + "content": "14:", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 506, + 636 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "We provide convergence analyses of these methods in general nonconvex settings, assuming full", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 177, + 605 + ], + "score": 1.0, + "content": "participation, i.e.", + "type": "text" + }, + { + "bbox": [ + 178, + 592, + 213, + 604 + ], + "score": 0.91, + "content": "\\bar { \\mathcal { S } } = [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 592, + 363, + 605 + ], + "score": 1.0, + "content": ". For expository purposes, we assume", + "type": "text" + }, + { + "bbox": [ + 363, + 592, + 393, + 603 + ], + "score": 0.92, + "content": "\\beta _ { 1 } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 592, + 506, + 605 + ], + "score": 1.0, + "content": ", though our analysis can be", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 600, + 507, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 187, + 617 + ], + "score": 1.0, + "content": "directly extended to", + "type": "text" + }, + { + "bbox": [ + 187, + 604, + 217, + 614 + ], + "score": 0.89, + "content": "\\beta _ { 1 } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 600, + 467, + 617 + ], + "score": 1.0, + "content": ". Our analysis can also be extended to partial participation (i.e.", + "type": "text" + }, + { + "bbox": [ + 467, + 603, + 502, + 615 + ], + "score": 0.92, + "content": "| S | < m", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 600, + 507, + 617 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "see Appendix A.2.1 for details). Furthermore, non-uniform weighted averaging typically used in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 623, + 461, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 461, + 637 + ], + "score": 1.0, + "content": "FEDAVG (McMahan et al., 2017) can also be incorporated into our analysis fairly easily.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 638, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 104, + 636, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 233, + 652 + ], + "score": 1.0, + "content": "Theorem 1. Let Assumptions", + "type": "text" + }, + { + "bbox": [ + 233, + 639, + 240, + 649 + ], + "score": 0.28, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 636, + 318, + 652 + ], + "score": 1.0, + "content": "to 3 hold, and let", + "type": "text" + }, + { + "bbox": [ + 319, + 639, + 367, + 651 + ], + "score": 0.93, + "content": "L , G , \\sigma _ { l } , \\sigma _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 636, + 479, + 652 + ], + "score": 1.0, + "content": "be as defined therein. Let", + "type": "text" + }, + { + "bbox": [ + 480, + 638, + 505, + 650 + ], + "score": 0.85, + "content": "\\sigma ^ { 2 } = ", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 649, + 325, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 155, + 664 + ], + "score": 0.92, + "content": "\\sigma _ { l } ^ { 2 } + 6 K \\sigma _ { g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 649, + 311, + 665 + ], + "score": 1.0, + "content": ". Consider the following conditions for", + "type": "text" + }, + { + "bbox": [ + 311, + 653, + 320, + 663 + ], + "score": 0.81, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 649, + 325, + 665 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 677, + 161, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 162, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 162, + 691 + ], + "score": 1.0, + "content": "(Condition I)", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 164, + 720 + ], + "lines": [ + { + "bbox": [ + 106, + 707, + 165, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 707, + 165, + 722 + ], + "score": 1.0, + "content": "(Condition II)", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 668, + 452, + 731 + ], + "lines": [ + { + "bbox": [ + 214, + 668, + 452, + 731 + ], + "spans": [ + { + "bbox": [ + 214, + 668, + 452, + 731 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\eta _ { l } \\le \\displaystyle \\frac { 1 } { K } \\operatorname* { m i n } \\left\\{ \\frac { 1 } { 1 6 L } , \\frac { 1 } { T ^ { 1 / 6 } } \\left[ \\frac { \\tau } { 1 2 0 L ^ { 2 } G } \\right] ^ { 1 / 3 } \\right\\} , } \\\\ & { \\eta _ { l } \\le \\displaystyle \\frac { 1 } { 3 K } \\operatorname* { m i n } \\left\\{ \\frac { 1 } { T ^ { 1 / 1 0 } } \\left[ \\frac { \\tau ^ { 3 } } { L ^ { 2 } G ^ { 3 } } \\right] ^ { 1 / 5 } , \\frac { 1 } { T ^ { 1 / 8 } } \\left[ \\frac { \\tau ^ { 2 } } { L ^ { 3 } G \\eta } \\right] ^ { 1 / 4 } \\right\\} . } \\end{array}", + "type": "interline_equation", + "image_path": "1a3347da564634f07515e78770ddadeb285d2b3ed5f780010df4a7f6a8388a15.jpg" + } + ] + } + ], + "index": 45.0, + "virtual_lines": [ + { + "bbox": [ + 214, + 668, + 452, + 683.75 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 214, + 683.75, + 452, + 699.5 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 214, + 699.5, + 452, + 715.25 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 214, + 715.25, + 452, + 731.0 + ], + "spans": [], + "index": 47 + } + ] + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 115 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 104, + 81, + 506, + 117 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 120, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "While FEDOPT has intuitive benefits over FEDAVG, it also raises a fundamental question: Can the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 266, + 144 + ], + "score": 1.0, + "content": "negative of the average model difference", + "type": "text" + }, + { + "bbox": [ + 267, + 132, + 279, + 142 + ], + "score": 0.87, + "content": "\\Delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "be used as a pseudo-gradient in general server optimizer", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "updates? In this paper, we provide an affirmative answer to this question by establishing a theoretical", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 166 + ], + "score": 1.0, + "content": "basis for FEDOPT. We will show that the use of the term SERVEROPT is justified, as we can guarantee", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 506, + 176 + ], + "score": 1.0, + "content": "convergence across a wide variety of server optimizers, including ADAGRAD, ADAM, and YOGI,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 425, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 425, + 186 + ], + "score": 1.0, + "content": "thus developing principled adaptive optimizers for FL based on our framework.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 120, + 506, + 186 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 201, + 324, + 214 + ], + "lines": [ + { + "bbox": [ + 104, + 200, + 326, + 216 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 326, + 216 + ], + "score": 1.0, + "content": "3 ADAPTIVE FEDERATED OPTIMIZATION", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "In this section, we specialize FEDOPT to settings where SERVEROPT is an adaptive optimization", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "score": 1.0, + "content": "method (one of ADAGRAD, YOGI or ADAM) and CLIENTOPT is SGD. By using adaptive methods", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 260 + ], + "score": 1.0, + "content": "(which generally require maintaining state) on the server and SGD on the clients, we ensure our", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 471, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 471, + 272 + ], + "score": 1.0, + "content": "methods have the same communication cost as FEDAVG and work in cross-device settings.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 227, + 506, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "Algorithm 2 provides pseudo-code for our methods. An alternate version using batched data and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "example-based weighting (as opposed to uniform weighting) of clients is given in Algorithm 5. The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 147, + 309 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 148, + 299, + 155, + 307 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 297, + 419, + 309 + ], + "score": 1.0, + "content": "controls the algorithms’ degree of adaptivity, with smaller values of", + "type": "text" + }, + { + "bbox": [ + 420, + 299, + 426, + 307 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "representing higher", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 308, + 504, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 504, + 320 + ], + "score": 1.0, + "content": "degrees of adaptivity. Note that the server updates of our methods are invariant to fixed multiplicative", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 244, + 331 + ], + "score": 1.0, + "content": "changes to the client learning rate", + "type": "text" + }, + { + "bbox": [ + 244, + 320, + 253, + 330 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 318, + 355, + 331 + ], + "score": 1.0, + "content": "for appropriately chosen", + "type": "text" + }, + { + "bbox": [ + 356, + 320, + 362, + 328 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 318, + 505, + 331 + ], + "score": 1.0, + "content": ", though as we shall see shortly, we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 329, + 322, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 154, + 342 + ], + "score": 1.0, + "content": "will require", + "type": "text" + }, + { + "bbox": [ + 155, + 331, + 164, + 341 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 329, + 322, + 342 + ], + "score": 1.0, + "content": "to be sufficiently small in our analysis.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 275, + 506, + 342 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 354, + 352, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 353, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 228, + 369 + ], + "score": 1.0, + "content": "Algorithm 2 FEDADAGRAD", + "type": "text" + }, + { + "bbox": [ + 232, + 352, + 353, + 370 + ], + "score": 1.0, + "content": "FEDYOGI , and FEDADAM", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "index", + "bbox": [ + 108, + 370, + 369, + 569 + ], + "lines": [ + { + "bbox": [ + 110, + 369, + 369, + 385 + ], + "spans": [ + { + "bbox": [ + 110, + 369, + 181, + 385 + ], + "score": 1.0, + "content": "1: Initialization:", + "type": "text" + }, + { + "bbox": [ + 181, + 371, + 234, + 383 + ], + "score": 0.88, + "content": "x _ { 0 } , v _ { - 1 } \\ge \\tau ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 369, + 309, + 385 + ], + "score": 1.0, + "content": ", decay parameters", + "type": "text" + }, + { + "bbox": [ + 310, + 371, + 369, + 384 + ], + "score": 0.91, + "content": "\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )", + "type": "inline_equation" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 381, + 225, + 394 + ], + "spans": [ + { + "bbox": [ + 109, + 381, + 138, + 393 + ], + "score": 1.0, + "content": "2: for", + "type": "text" + }, + { + "bbox": [ + 138, + 383, + 209, + 394 + ], + "score": 0.88, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 381, + 225, + 393 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 393, + 247, + 405 + ], + "spans": [ + { + "bbox": [ + 110, + 393, + 122, + 405 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 137, + 393, + 198, + 405 + ], + "score": 1.0, + "content": "Sample subset", + "type": "text" + }, + { + "bbox": [ + 198, + 394, + 206, + 403 + ], + "score": 0.8, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 393, + 247, + 405 + ], + "score": 1.0, + "content": "of clients", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 401, + 182, + 421 + ], + "spans": [ + { + "bbox": [ + 110, + 403, + 122, + 416 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 138, + 403, + 178, + 417 + ], + "score": 0.5, + "content": "x _ { i , 0 } ^ { t } \\dot { = } x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 401, + 182, + 421 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 415, + 282, + 427 + ], + "spans": [ + { + "bbox": [ + 110, + 415, + 121, + 427 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 137, + 415, + 199, + 427 + ], + "score": 1.0, + "content": "for each client", + "type": "text" + }, + { + "bbox": [ + 199, + 416, + 222, + 426 + ], + "score": 0.88, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 415, + 282, + 427 + ], + "score": 1.0, + "content": "in parallel do", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 426, + 258, + 438 + ], + "spans": [ + { + "bbox": [ + 110, + 426, + 121, + 438 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 153, + 426, + 168, + 437 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 168, + 426, + 243, + 438 + ], + "score": 0.89, + "content": "k = 0 , \\cdots , K - \\bar { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 426, + 258, + 437 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 434, + 363, + 452 + ], + "spans": [ + { + "bbox": [ + 110, + 437, + 122, + 450 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 166, + 434, + 292, + 452 + ], + "score": 1.0, + "content": "Compute an unbiased estimate", + "type": "text" + }, + { + "bbox": [ + 293, + 437, + 309, + 451 + ], + "score": 0.91, + "content": "g _ { i , k } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 434, + 321, + 452 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 321, + 437, + 363, + 451 + ], + "score": 0.91, + "content": "\\nabla F _ { i } ( x _ { i , k } ^ { t } )", + "type": "inline_equation" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 449, + 311, + 498 + ], + "spans": [ + { + "bbox": [ + 110, + 450, + 122, + 463 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 137, + 449, + 311, + 498 + ], + "score": 0.84, + "content": "\\begin{array} { c } { x _ { i , k + 1 } ^ { t } = x _ { i , k } ^ { t } - \\eta _ { l } g _ { i , k } ^ { t } } \\\\ { \\Delta _ { i } ^ { t } = x _ { i , K } ^ { t } - x _ { t } } \\\\ { \\Delta _ { t } = \\beta _ { 1 } \\Delta _ { t - 1 } + \\left( 1 - \\beta _ { 1 } \\right) \\left( \\frac { 1 } { \\left| S \\right| } \\sum _ { i \\in S } \\Delta _ { i } ^ { t } \\right) } \\end{array}", + "type": "inline_equation", + "image_path": "64a4df90b9ef875531f06b07407ff8cc2911e941ad716e76b11347461ef989cc.jpg" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 464, + 122, + 479 + ], + "spans": [ + { + "bbox": [ + 109, + 464, + 122, + 479 + ], + "score": 1.0, + "content": "9:", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 481, + 122, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 481, + 122, + 496 + ], + "score": 1.0, + "content": "10:", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 499, + 281, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 122, + 514 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 140, + 500, + 208, + 513 + ], + "score": 0.71, + "content": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 499, + 281, + 513 + ], + "score": 1.0, + "content": "(FEDADAGRAD)", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 515, + 367, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 122, + 532 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 138, + 515, + 367, + 568 + ], + "score": 0.7, + "content": "\\begin{array} { r l } & { v _ { t } = v _ { t - 1 } - ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\mathrm { s i g n } ( v _ { t - 1 } - \\Delta _ { t } ^ { 2 } ) \\ ( \\mathrm { F E D Y O G I } ) } \\\\ & { v _ { t } = \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\left( \\mathrm { F E D A D A M } \\right) } \\\\ & { x _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } } \\end{array}", + "type": "inline_equation", + "image_path": "32ee17883ba5478b7ac54f615779eedc5713aae3bffd1c1a8bbdb8e1021bffff.jpg" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 534, + 122, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 122, + 549 + ], + "score": 1.0, + "content": "13:", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 550, + 122, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 122, + 564 + ], + "score": 1.0, + "content": "14:", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 369, + 369, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 506, + 636 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "We provide convergence analyses of these methods in general nonconvex settings, assuming full", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 177, + 605 + ], + "score": 1.0, + "content": "participation, i.e.", + "type": "text" + }, + { + "bbox": [ + 178, + 592, + 213, + 604 + ], + "score": 0.91, + "content": "\\bar { \\mathcal { S } } = [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 592, + 363, + 605 + ], + "score": 1.0, + "content": ". For expository purposes, we assume", + "type": "text" + }, + { + "bbox": [ + 363, + 592, + 393, + 603 + ], + "score": 0.92, + "content": "\\beta _ { 1 } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 592, + 506, + 605 + ], + "score": 1.0, + "content": ", though our analysis can be", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 600, + 507, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 187, + 617 + ], + "score": 1.0, + "content": "directly extended to", + "type": "text" + }, + { + "bbox": [ + 187, + 604, + 217, + 614 + ], + "score": 0.89, + "content": "\\beta _ { 1 } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 600, + 467, + 617 + ], + "score": 1.0, + "content": ". Our analysis can also be extended to partial participation (i.e.", + "type": "text" + }, + { + "bbox": [ + 467, + 603, + 502, + 615 + ], + "score": 0.92, + "content": "| S | < m", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 600, + 507, + 617 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "see Appendix A.2.1 for details). Furthermore, non-uniform weighted averaging typically used in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 623, + 461, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 461, + 637 + ], + "score": 1.0, + "content": "FEDAVG (McMahan et al., 2017) can also be incorporated into our analysis fairly easily.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 581, + 507, + 637 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 638, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 104, + 636, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 233, + 652 + ], + "score": 1.0, + "content": "Theorem 1. Let Assumptions", + "type": "text" + }, + { + "bbox": [ + 233, + 639, + 240, + 649 + ], + "score": 0.28, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 636, + 318, + 652 + ], + "score": 1.0, + "content": "to 3 hold, and let", + "type": "text" + }, + { + "bbox": [ + 319, + 639, + 367, + 651 + ], + "score": 0.93, + "content": "L , G , \\sigma _ { l } , \\sigma _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 636, + 479, + 652 + ], + "score": 1.0, + "content": "be as defined therein. Let", + "type": "text" + }, + { + "bbox": [ + 480, + 638, + 505, + 650 + ], + "score": 0.85, + "content": "\\sigma ^ { 2 } = ", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 649, + 325, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 155, + 664 + ], + "score": 0.92, + "content": "\\sigma _ { l } ^ { 2 } + 6 K \\sigma _ { g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 649, + 311, + 665 + ], + "score": 1.0, + "content": ". Consider the following conditions for", + "type": "text" + }, + { + "bbox": [ + 311, + 653, + 320, + 663 + ], + "score": 0.81, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 649, + 325, + 665 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 104, + 636, + 505, + 665 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 677, + 161, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 162, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 162, + 691 + ], + "score": 1.0, + "content": "(Condition I)", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 676, + 162, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 164, + 720 + ], + "lines": [ + { + "bbox": [ + 106, + 707, + 165, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 707, + 165, + 722 + ], + "score": 1.0, + "content": "(Condition II)", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 707, + 165, + 722 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 668, + 452, + 731 + ], + "lines": [ + { + "bbox": [ + 214, + 668, + 452, + 731 + ], + "spans": [ + { + "bbox": [ + 214, + 668, + 452, + 731 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\eta _ { l } \\le \\displaystyle \\frac { 1 } { K } \\operatorname* { m i n } \\left\\{ \\frac { 1 } { 1 6 L } , \\frac { 1 } { T ^ { 1 / 6 } } \\left[ \\frac { \\tau } { 1 2 0 L ^ { 2 } G } \\right] ^ { 1 / 3 } \\right\\} , } \\\\ & { \\eta _ { l } \\le \\displaystyle \\frac { 1 } { 3 K } \\operatorname* { m i n } \\left\\{ \\frac { 1 } { T ^ { 1 / 1 0 } } \\left[ \\frac { \\tau ^ { 3 } } { L ^ { 2 } G ^ { 3 } } \\right] ^ { 1 / 5 } , \\frac { 1 } { T ^ { 1 / 8 } } \\left[ \\frac { \\tau ^ { 2 } } { L ^ { 3 } G \\eta } \\right] ^ { 1 / 4 } \\right\\} . } \\end{array}", + "type": "interline_equation", + "image_path": "1a3347da564634f07515e78770ddadeb285d2b3ed5f780010df4a7f6a8388a15.jpg" + } + ] + } + ], + "index": 45.0, + "virtual_lines": [ + { + "bbox": [ + 214, + 668, + 452, + 683.75 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 214, + 683.75, + 452, + 699.5 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 214, + 699.5, + 452, + 715.25 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 214, + 715.25, + 452, + 731.0 + ], + "spans": [], + "index": 47 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 342, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 342, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 342, + 96 + ], + "score": 1.0, + "content": "Then the iterates of Algorithm 2 for FEDADAGRAD satisfy", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 102, + 98, + 492, + 157 + ], + "lines": [ + { + "bbox": [ + 102, + 98, + 492, + 157 + ], + "spans": [ + { + "bbox": [ + 102, + 98, + 492, + 157 + ], + "score": 0.93, + "content": "\\begin{array} { r l r l } & { u n d e r ~ C o n d i t i o n ~ I o n l y , } & & { \\underset { 0 \\leq t \\leq T - 1 } { \\operatorname* { m i n } } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } \\leq \\odot \\left( \\left[ \\frac { G } { \\sqrt { T } } + \\frac { \\tau } { \\eta _ { t } K T } \\right] ( \\Psi + \\Psi _ { v a r } ) \\right) , } \\\\ & { u n d e r ~ b o t h ~ C o n d i t i o n ~ I \\ll { I I } , } & & { \\underset { 0 \\leq t \\leq T - 1 } { \\operatorname* { m i n } } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } \\leq \\odot \\left( \\left[ \\frac { G } { \\sqrt { T } } + \\frac { \\tau } { \\eta _ { t } K T } \\right] \\left( \\Psi + \\widetilde { \\Psi } _ { \\mathrm { v a r } } \\right) \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "6d227a0445973c9ea45cc137c9b833b013e2b27e0a67359d5cea6897678e26c4.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 102, + 98, + 492, + 117.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 102, + 117.66666666666667, + 492, + 137.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 102, + 137.33333333333334, + 492, + 157.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 158, + 170, + 170 + ], + "lines": [ + { + "bbox": [ + 105, + 157, + 170, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 170, + 172 + ], + "score": 1.0, + "content": "Here, we define", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 168, + 424, + 260 + ], + "lines": [ + { + "bbox": [ + 187, + 168, + 424, + 260 + ], + "spans": [ + { + "bbox": [ + 187, + 168, + 424, + 260 + ], + "score": 0.93, + "content": "\\begin{array} { c } { { \\Psi = \\displaystyle \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\sigma ^ { 2 } , } } \\\\ { { \\Psi _ { \\mathrm { v a r } } = \\displaystyle \\frac { d ( \\eta _ { l } K G ^ { 2 } + \\tau \\eta L ) } { \\tau } \\Bigg [ 1 + \\log { \\displaystyle \\frac { \\tau ^ { 2 } + \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } } \\Bigg ] , } } \\\\ { { \\widetilde { \\Psi } _ { \\mathrm { v a r } } = \\displaystyle \\frac { 2 \\eta _ { l } K G ^ { 2 } + \\tau \\eta L } { \\tau ^ { 2 } } \\left[ \\displaystyle \\frac { 2 \\eta _ { l } ^ { 2 } K T } { m } \\sigma _ { l } ^ { 2 } + 1 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\sigma ^ { 2 } \\right] . } } \\end{array}", + "type": "interline_equation", + "image_path": "0a2a2d5da3167cf10c370390393db317fa6834be07374f6ba61f51fd8bdceaa2.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 168, + 424, + 183.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 187, + 183.33333333333334, + 424, + 198.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 187, + 198.66666666666669, + 424, + 214.00000000000003 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 187, + 214.00000000000003, + 424, + 229.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 229.33333333333337, + 424, + 244.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 244.6666666666667, + 424, + 260.00000000000006 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 383, + 279 + ], + "score": 1.0, + "content": "All proofs are relegated to Appendix A due to space constraints. When", + "type": "text" + }, + { + "bbox": [ + 384, + 268, + 393, + 278 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "satisfies the condition in the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 277, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 392, + 293 + ], + "score": 1.0, + "content": "second part the above result, we obtain a convergence rate depending on", + "type": "text" + }, + { + "bbox": [ + 392, + 277, + 462, + 291 + ], + "score": 0.91, + "content": "\\operatorname* { m i n } \\{ \\Psi _ { \\mathrm { v a r } } , \\widetilde { \\Psi } _ { \\mathrm { v a r } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 279, + 505, + 293 + ], + "score": 1.0, + "content": ". To obtain", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 291, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 212, + 302 + ], + "score": 1.0, + "content": "an explicit dependence on", + "type": "text" + }, + { + "bbox": [ + 212, + 291, + 220, + 300 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 291, + 238, + 302 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 239, + 291, + 249, + 300 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 291, + 459, + 302 + ], + "score": 1.0, + "content": ", we simplify the above result for a specific choice of", + "type": "text" + }, + { + "bbox": [ + 460, + 292, + 478, + 302 + ], + "score": 0.87, + "content": "\\eta , \\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 291, + 496, + 302 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 496, + 293, + 503, + 300 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 291, + 506, + 302 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 304, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 197, + 318 + ], + "score": 1.0, + "content": "Corollary 1. Suppose √", + "type": "text" + }, + { + "bbox": [ + 197, + 307, + 205, + 316 + ], + "score": 0.77, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 304, + 357, + 318 + ], + "score": 1.0, + "content": "is such that the conditions in Theorem", + "type": "text" + }, + { + "bbox": [ + 357, + 306, + 362, + 315 + ], + "score": 0.34, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 304, + 430, + 318 + ], + "score": 1.0, + "content": "are satisfied and", + "type": "text" + }, + { + "bbox": [ + 430, + 304, + 503, + 317 + ], + "score": 0.92, + "content": "\\eta _ { l } = \\Theta \\big ( { 1 } / { ( K L \\sqrt { T } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 304, + 506, + 318 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 316, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 160, + 330 + ], + "score": 1.0, + "content": "Also suppose", + "type": "text" + }, + { + "bbox": [ + 160, + 316, + 221, + 329 + ], + "score": 0.92, + "content": "\\eta = \\Theta ( \\sqrt { K m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 316, + 240, + 330 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 241, + 317, + 275, + 329 + ], + "score": 0.91, + "content": "\\tau = G / _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 316, + 386, + 330 + ], + "score": 1.0, + "content": ". Then, for sufficiently large", + "type": "text" + }, + { + "bbox": [ + 386, + 318, + 394, + 327 + ], + "score": 0.71, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 316, + 506, + 330 + ], + "score": 1.0, + "content": ", the iterates of Algorithm 2", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 327, + 213, + 341 + ], + "spans": [ + { + "bbox": [ + 104, + 327, + 213, + 341 + ], + "score": 1.0, + "content": "for FEDADAGRAD satisfy", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 126, + 344, + 483, + 378 + ], + "lines": [ + { + "bbox": [ + 126, + 344, + 483, + 378 + ], + "spans": [ + { + "bbox": [ + 126, + 344, + 483, + 378 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\mathcal { O } \\Bigg ( \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\sqrt { m K T } } + \\frac { 2 \\sigma _ { l } ^ { 2 } L } { G ^ { 2 } \\sqrt { m K T } } + \\frac { \\sigma ^ { 2 } } { G K T } + \\frac { \\sigma ^ { 2 } L \\sqrt { m } } { G ^ { 2 } \\sqrt { K } T ^ { 3 / 2 } } \\Bigg ) .", + "type": "interline_equation", + "image_path": "43b9fd992563d3b2e2e52d3a830ad9c43e4930752416313d7759c7ded13d3f34.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 126, + 344, + 483, + 355.3333333333333 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 126, + 355.3333333333333, + 483, + 366.66666666666663 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 126, + 366.66666666666663, + 483, + 377.99999999999994 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 395, + 481, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 396, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 482, + 408 + ], + "score": 1.0, + "content": "We defer a detailed discussion about our analysis and its implication to the end of the section.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 412, + 504, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "Analysis of FEDADAM Next, we provide the convergence analysis of FEDADAM. The proof of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 429, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 429, + 436 + ], + "score": 1.0, + "content": "FEDYOGI is very similar and hence, we omit the details of FEDYOGI’s analysis.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 437, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 104, + 434, + 503, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 286, + 455 + ], + "score": 1.0, + "content": "Theorem 2. Let Assumptions 1 to 3 hold, and", + "type": "text" + }, + { + "bbox": [ + 287, + 438, + 335, + 451 + ], + "score": 0.9, + "content": "L , G , \\sigma _ { l } , \\sigma _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 434, + 434, + 455 + ], + "score": 1.0, + "content": "be as defined therein. Let", + "type": "text" + }, + { + "bbox": [ + 434, + 437, + 503, + 451 + ], + "score": 0.92, + "content": "\\sigma ^ { 2 } = \\sigma _ { l } ^ { 2 } + 6 K \\sigma _ { g } ^ { 2 }", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 450, + 340, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 270, + 463 + ], + "score": 1.0, + "content": "Suppose the client learning rate satisfies", + "type": "text" + }, + { + "bbox": [ + 270, + 451, + 320, + 462 + ], + "score": 0.9, + "content": "\\eta _ { l } \\leq 1 / 1 6 L K", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 450, + 340, + 463 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 467, + 417, + 500 + ], + "lines": [ + { + "bbox": [ + 192, + 467, + 417, + 500 + ], + "spans": [ + { + "bbox": [ + 192, + 467, + 417, + 500 + ], + "score": 0.93, + "content": "\\eta _ { l } \\leq \\frac { 1 } { 6 K } \\operatorname* { m i n } \\left\\{ \\left[ \\frac { \\tau } { G L } \\right] ^ { 1 / 2 } , \\left[ \\frac { \\tau ^ { 2 } } { G L ^ { 3 } \\eta } \\right] ^ { 1 / 4 } , \\left[ \\frac { \\tau } { G L ^ { 2 } } \\right] ^ { 1 / 3 } \\right\\} .", + "type": "interline_equation", + "image_path": "e8928e419e9469dc75d92d0746178a7f01edcb0c938a07b652ba1145cfa29abf.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 192, + 467, + 417, + 483.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 192, + 483.5, + 417, + 500.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 504, + 325, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 503, + 325, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 325, + 518 + ], + "score": 1.0, + "content": "Then the iterates of Algorithm 2 for FEDADAM satisfy", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 521, + 428, + 549 + ], + "lines": [ + { + "bbox": [ + 183, + 521, + 428, + 549 + ], + "spans": [ + { + "bbox": [ + 183, + 521, + 428, + 549 + ], + "score": 0.89, + "content": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\odot \\left( \\frac { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } { \\eta _ { l } K T } \\left( \\Psi + \\Psi _ { \\mathrm { v a r } } \\right) \\right) ,", + "type": "interline_equation", + "image_path": "4b92b739de8bef239c7482a5ea7c381ade04f5280fe2322fdfff1491c1c14c1c.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 183, + 521, + 428, + 530.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 183, + 530.3333333333334, + 428, + 539.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 183, + 539.6666666666667, + 428, + 549.0000000000001 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 133, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 135, + 567 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 561, + 417, + 619 + ], + "lines": [ + { + "bbox": [ + 194, + 561, + 417, + 619 + ], + "spans": [ + { + "bbox": [ + 194, + 561, + 417, + 619 + ], + "score": 0.75, + "content": "\\begin{array} { c } { { \\Psi = \\displaystyle \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\sigma ^ { 2 } , } } \\\\ { { \\Psi _ { \\nu a r } = \\left( G + \\displaystyle \\frac { \\eta L } { 2 } \\right) \\left[ \\displaystyle \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sigma _ { l } ^ { 2 } + \\displaystyle \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\sigma ^ { 2 } \\right] . } } \\end{array}", + "type": "interline_equation", + "image_path": "134d7b1d2fa61708e74810c75283b81fe424571d1047e4237a6f800cd422b0a5.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 194, + 561, + 417, + 575.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 194, + 575.5, + 417, + 590.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 194, + 590.0, + 417, + 604.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 194, + 604.5, + 417, + 619.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 624, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 449, + 637 + ], + "score": 1.0, + "content": "Similar to the FEDADAGRAD case, we restate the above result for a specific choice of", + "type": "text" + }, + { + "bbox": [ + 450, + 627, + 469, + 637 + ], + "score": 0.78, + "content": "\\eta _ { l } , \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 624, + 487, + 637 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 487, + 627, + 493, + 635 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 636, + 294, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 254, + 648 + ], + "score": 1.0, + "content": "order to highlight the dependence of", + "type": "text" + }, + { + "bbox": [ + 254, + 636, + 264, + 646 + ], + "score": 0.85, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 636, + 282, + 648 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 282, + 636, + 290, + 646 + ], + "score": 0.84, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 636, + 294, + 648 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 201, + 663 + ], + "score": 1.0, + "content": "Corollary 2. Suppose", + "type": "text" + }, + { + "bbox": [ + 201, + 652, + 210, + 662 + ], + "score": 0.75, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 650, + 506, + 663 + ], + "score": 1.0, + "content": "is chosen such that the conditions in Theorem 2 are satisfied and that√", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 661, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 107, + 662, + 183, + 675 + ], + "score": 0.91, + "content": "\\eta _ { l } = \\Theta \\bigl ( 1 / ( K L \\sqrt { T } ) \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 662, + 246, + 676 + ], + "score": 1.0, + "content": ". Also, suppose", + "type": "text" + }, + { + "bbox": [ + 247, + 661, + 308, + 675 + ], + "score": 0.92, + "content": "\\eta = \\Theta ( \\sqrt { K m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 662, + 328, + 676 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 662, + 363, + 675 + ], + "score": 0.9, + "content": "\\tau = G / _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 662, + 478, + 676 + ], + "score": 1.0, + "content": ". Then, for sufficiently large", + "type": "text" + }, + { + "bbox": [ + 478, + 663, + 486, + 672 + ], + "score": 0.73, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 662, + 505, + 676 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 673, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 289, + 686 + ], + "score": 1.0, + "content": "iterates of Algorithm 2 for FEDADAM satisfy", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "interline_equation", + "bbox": [ + 127, + 689, + 483, + 723 + ], + "lines": [ + { + "bbox": [ + 127, + 689, + 483, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 689, + 483, + 723 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\mathcal { O } \\Bigg ( \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\sqrt { m K T } } + \\frac { 2 \\sigma _ { l } ^ { 2 } L } { G ^ { 2 } \\sqrt { m K T } } + \\frac { \\sigma ^ { 2 } } { G K T } + \\frac { \\sigma ^ { 2 } L \\sqrt { m } } { G ^ { 2 } \\sqrt { K } T ^ { 3 / 2 } } \\Bigg ) .", + "type": "interline_equation", + "image_path": "8506f0697321f1583c8f1fe7772c08b405c66458be413e11447b3a78200c04f4.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 127, + 689, + 483, + 700.3333333333334 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 127, + 700.3333333333334, + 483, + 711.6666666666667 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 127, + 711.6666666666667, + 483, + 723.0000000000001 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 342, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 342, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 342, + 96 + ], + "score": 1.0, + "content": "Then the iterates of Algorithm 2 for FEDADAGRAD satisfy", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 342, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 102, + 98, + 492, + 157 + ], + "lines": [ + { + "bbox": [ + 102, + 98, + 492, + 157 + ], + "spans": [ + { + "bbox": [ + 102, + 98, + 492, + 157 + ], + "score": 0.93, + "content": "\\begin{array} { r l r l } & { u n d e r ~ C o n d i t i o n ~ I o n l y , } & & { \\underset { 0 \\leq t \\leq T - 1 } { \\operatorname* { m i n } } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } \\leq \\odot \\left( \\left[ \\frac { G } { \\sqrt { T } } + \\frac { \\tau } { \\eta _ { t } K T } \\right] ( \\Psi + \\Psi _ { v a r } ) \\right) , } \\\\ & { u n d e r ~ b o t h ~ C o n d i t i o n ~ I \\ll { I I } , } & & { \\underset { 0 \\leq t \\leq T - 1 } { \\operatorname* { m i n } } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } \\leq \\odot \\left( \\left[ \\frac { G } { \\sqrt { T } } + \\frac { \\tau } { \\eta _ { t } K T } \\right] \\left( \\Psi + \\widetilde { \\Psi } _ { \\mathrm { v a r } } \\right) \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "6d227a0445973c9ea45cc137c9b833b013e2b27e0a67359d5cea6897678e26c4.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 102, + 98, + 492, + 117.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 102, + 117.66666666666667, + 492, + 137.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 102, + 137.33333333333334, + 492, + 157.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 158, + 170, + 170 + ], + "lines": [ + { + "bbox": [ + 105, + 157, + 170, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 170, + 172 + ], + "score": 1.0, + "content": "Here, we define", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 157, + 170, + 172 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 168, + 424, + 260 + ], + "lines": [ + { + "bbox": [ + 187, + 168, + 424, + 260 + ], + "spans": [ + { + "bbox": [ + 187, + 168, + 424, + 260 + ], + "score": 0.93, + "content": "\\begin{array} { c } { { \\Psi = \\displaystyle \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\sigma ^ { 2 } , } } \\\\ { { \\Psi _ { \\mathrm { v a r } } = \\displaystyle \\frac { d ( \\eta _ { l } K G ^ { 2 } + \\tau \\eta L ) } { \\tau } \\Bigg [ 1 + \\log { \\displaystyle \\frac { \\tau ^ { 2 } + \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } } \\Bigg ] , } } \\\\ { { \\widetilde { \\Psi } _ { \\mathrm { v a r } } = \\displaystyle \\frac { 2 \\eta _ { l } K G ^ { 2 } + \\tau \\eta L } { \\tau ^ { 2 } } \\left[ \\displaystyle \\frac { 2 \\eta _ { l } ^ { 2 } K T } { m } \\sigma _ { l } ^ { 2 } + 1 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\sigma ^ { 2 } \\right] . } } \\end{array}", + "type": "interline_equation", + "image_path": "0a2a2d5da3167cf10c370390393db317fa6834be07374f6ba61f51fd8bdceaa2.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 168, + 424, + 183.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 187, + 183.33333333333334, + 424, + 198.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 187, + 198.66666666666669, + 424, + 214.00000000000003 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 187, + 214.00000000000003, + 424, + 229.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 187, + 229.33333333333337, + 424, + 244.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 187, + 244.6666666666667, + 424, + 260.00000000000006 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 383, + 279 + ], + "score": 1.0, + "content": "All proofs are relegated to Appendix A due to space constraints. When", + "type": "text" + }, + { + "bbox": [ + 384, + 268, + 393, + 278 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "satisfies the condition in the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 277, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 392, + 293 + ], + "score": 1.0, + "content": "second part the above result, we obtain a convergence rate depending on", + "type": "text" + }, + { + "bbox": [ + 392, + 277, + 462, + 291 + ], + "score": 0.91, + "content": "\\operatorname* { m i n } \\{ \\Psi _ { \\mathrm { v a r } } , \\widetilde { \\Psi } _ { \\mathrm { v a r } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 279, + 505, + 293 + ], + "score": 1.0, + "content": ". To obtain", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 291, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 212, + 302 + ], + "score": 1.0, + "content": "an explicit dependence on", + "type": "text" + }, + { + "bbox": [ + 212, + 291, + 220, + 300 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 291, + 238, + 302 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 239, + 291, + 249, + 300 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 291, + 459, + 302 + ], + "score": 1.0, + "content": ", we simplify the above result for a specific choice of", + "type": "text" + }, + { + "bbox": [ + 460, + 292, + 478, + 302 + ], + "score": 0.87, + "content": "\\eta , \\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 291, + 496, + 302 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 496, + 293, + 503, + 300 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 291, + 506, + 302 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 266, + 506, + 302 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 304, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 197, + 318 + ], + "score": 1.0, + "content": "Corollary 1. Suppose √", + "type": "text" + }, + { + "bbox": [ + 197, + 307, + 205, + 316 + ], + "score": 0.77, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 304, + 357, + 318 + ], + "score": 1.0, + "content": "is such that the conditions in Theorem", + "type": "text" + }, + { + "bbox": [ + 357, + 306, + 362, + 315 + ], + "score": 0.34, + "content": "I", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 304, + 430, + 318 + ], + "score": 1.0, + "content": "are satisfied and", + "type": "text" + }, + { + "bbox": [ + 430, + 304, + 503, + 317 + ], + "score": 0.92, + "content": "\\eta _ { l } = \\Theta \\big ( { 1 } / { ( K L \\sqrt { T } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 304, + 506, + 318 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 316, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 160, + 330 + ], + "score": 1.0, + "content": "Also suppose", + "type": "text" + }, + { + "bbox": [ + 160, + 316, + 221, + 329 + ], + "score": 0.92, + "content": "\\eta = \\Theta ( \\sqrt { K m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 316, + 240, + 330 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 241, + 317, + 275, + 329 + ], + "score": 0.91, + "content": "\\tau = G / _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 316, + 386, + 330 + ], + "score": 1.0, + "content": ". Then, for sufficiently large", + "type": "text" + }, + { + "bbox": [ + 386, + 318, + 394, + 327 + ], + "score": 0.71, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 316, + 506, + 330 + ], + "score": 1.0, + "content": ", the iterates of Algorithm 2", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 327, + 213, + 341 + ], + "spans": [ + { + "bbox": [ + 104, + 327, + 213, + 341 + ], + "score": 1.0, + "content": "for FEDADAGRAD satisfy", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 304, + 506, + 341 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 126, + 344, + 483, + 378 + ], + "lines": [ + { + "bbox": [ + 126, + 344, + 483, + 378 + ], + "spans": [ + { + "bbox": [ + 126, + 344, + 483, + 378 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\mathcal { O } \\Bigg ( \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\sqrt { m K T } } + \\frac { 2 \\sigma _ { l } ^ { 2 } L } { G ^ { 2 } \\sqrt { m K T } } + \\frac { \\sigma ^ { 2 } } { G K T } + \\frac { \\sigma ^ { 2 } L \\sqrt { m } } { G ^ { 2 } \\sqrt { K } T ^ { 3 / 2 } } \\Bigg ) .", + "type": "interline_equation", + "image_path": "43b9fd992563d3b2e2e52d3a830ad9c43e4930752416313d7759c7ded13d3f34.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 126, + 344, + 483, + 355.3333333333333 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 126, + 355.3333333333333, + 483, + 366.66666666666663 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 126, + 366.66666666666663, + 483, + 377.99999999999994 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 395, + 481, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 396, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 482, + 408 + ], + "score": 1.0, + "content": "We defer a detailed discussion about our analysis and its implication to the end of the section.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 396, + 482, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 412, + 504, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "Analysis of FEDADAM Next, we provide the convergence analysis of FEDADAM. The proof of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 429, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 429, + 436 + ], + "score": 1.0, + "content": "FEDYOGI is very similar and hence, we omit the details of FEDYOGI’s analysis.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 412, + 505, + 436 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 437, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 104, + 434, + 503, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 286, + 455 + ], + "score": 1.0, + "content": "Theorem 2. Let Assumptions 1 to 3 hold, and", + "type": "text" + }, + { + "bbox": [ + 287, + 438, + 335, + 451 + ], + "score": 0.9, + "content": "L , G , \\sigma _ { l } , \\sigma _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 434, + 434, + 455 + ], + "score": 1.0, + "content": "be as defined therein. Let", + "type": "text" + }, + { + "bbox": [ + 434, + 437, + 503, + 451 + ], + "score": 0.92, + "content": "\\sigma ^ { 2 } = \\sigma _ { l } ^ { 2 } + 6 K \\sigma _ { g } ^ { 2 }", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 450, + 340, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 270, + 463 + ], + "score": 1.0, + "content": "Suppose the client learning rate satisfies", + "type": "text" + }, + { + "bbox": [ + 270, + 451, + 320, + 462 + ], + "score": 0.9, + "content": "\\eta _ { l } \\leq 1 / 1 6 L K", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 450, + 340, + 463 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 434, + 503, + 463 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 467, + 417, + 500 + ], + "lines": [ + { + "bbox": [ + 192, + 467, + 417, + 500 + ], + "spans": [ + { + "bbox": [ + 192, + 467, + 417, + 500 + ], + "score": 0.93, + "content": "\\eta _ { l } \\leq \\frac { 1 } { 6 K } \\operatorname* { m i n } \\left\\{ \\left[ \\frac { \\tau } { G L } \\right] ^ { 1 / 2 } , \\left[ \\frac { \\tau ^ { 2 } } { G L ^ { 3 } \\eta } \\right] ^ { 1 / 4 } , \\left[ \\frac { \\tau } { G L ^ { 2 } } \\right] ^ { 1 / 3 } \\right\\} .", + "type": "interline_equation", + "image_path": "e8928e419e9469dc75d92d0746178a7f01edcb0c938a07b652ba1145cfa29abf.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 192, + 467, + 417, + 483.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 192, + 483.5, + 417, + 500.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 504, + 325, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 503, + 325, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 325, + 518 + ], + "score": 1.0, + "content": "Then the iterates of Algorithm 2 for FEDADAM satisfy", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 503, + 325, + 518 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 521, + 428, + 549 + ], + "lines": [ + { + "bbox": [ + 183, + 521, + 428, + 549 + ], + "spans": [ + { + "bbox": [ + 183, + 521, + 428, + 549 + ], + "score": 0.89, + "content": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\odot \\left( \\frac { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } { \\eta _ { l } K T } \\left( \\Psi + \\Psi _ { \\mathrm { v a r } } \\right) \\right) ,", + "type": "interline_equation", + "image_path": "4b92b739de8bef239c7482a5ea7c381ade04f5280fe2322fdfff1491c1c14c1c.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 183, + 521, + 428, + 530.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 183, + 530.3333333333334, + 428, + 539.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 183, + 539.6666666666667, + 428, + 549.0000000000001 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 554, + 133, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 135, + 567 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 552, + 135, + 567 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 194, + 561, + 417, + 619 + ], + "lines": [ + { + "bbox": [ + 194, + 561, + 417, + 619 + ], + "spans": [ + { + "bbox": [ + 194, + 561, + 417, + 619 + ], + "score": 0.75, + "content": "\\begin{array} { c } { { \\Psi = \\displaystyle \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\sigma ^ { 2 } , } } \\\\ { { \\Psi _ { \\nu a r } = \\left( G + \\displaystyle \\frac { \\eta L } { 2 } \\right) \\left[ \\displaystyle \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sigma _ { l } ^ { 2 } + \\displaystyle \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\sigma ^ { 2 } \\right] . } } \\end{array}", + "type": "interline_equation", + "image_path": "134d7b1d2fa61708e74810c75283b81fe424571d1047e4237a6f800cd422b0a5.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 194, + 561, + 417, + 575.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 194, + 575.5, + 417, + 590.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 194, + 590.0, + 417, + 604.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 194, + 604.5, + 417, + 619.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 624, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 449, + 637 + ], + "score": 1.0, + "content": "Similar to the FEDADAGRAD case, we restate the above result for a specific choice of", + "type": "text" + }, + { + "bbox": [ + 450, + 627, + 469, + 637 + ], + "score": 0.78, + "content": "\\eta _ { l } , \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 624, + 487, + 637 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 487, + 627, + 493, + 635 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 636, + 294, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 254, + 648 + ], + "score": 1.0, + "content": "order to highlight the dependence of", + "type": "text" + }, + { + "bbox": [ + 254, + 636, + 264, + 646 + ], + "score": 0.85, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 636, + 282, + 648 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 282, + 636, + 290, + 646 + ], + "score": 0.84, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 636, + 294, + 648 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 624, + 505, + 648 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 106, + 650, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 201, + 663 + ], + "score": 1.0, + "content": "Corollary 2. Suppose", + "type": "text" + }, + { + "bbox": [ + 201, + 652, + 210, + 662 + ], + "score": 0.75, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 650, + 506, + 663 + ], + "score": 1.0, + "content": "is chosen such that the conditions in Theorem 2 are satisfied and that√", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 661, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 107, + 662, + 183, + 675 + ], + "score": 0.91, + "content": "\\eta _ { l } = \\Theta \\bigl ( 1 / ( K L \\sqrt { T } ) \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 662, + 246, + 676 + ], + "score": 1.0, + "content": ". Also, suppose", + "type": "text" + }, + { + "bbox": [ + 247, + 661, + 308, + 675 + ], + "score": 0.92, + "content": "\\eta = \\Theta ( \\sqrt { K m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 662, + 328, + 676 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 662, + 363, + 675 + ], + "score": 0.9, + "content": "\\tau = G / _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 662, + 478, + 676 + ], + "score": 1.0, + "content": ". Then, for sufficiently large", + "type": "text" + }, + { + "bbox": [ + 478, + 663, + 486, + 672 + ], + "score": 0.73, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 662, + 505, + 676 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 673, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 289, + 686 + ], + "score": 1.0, + "content": "iterates of Algorithm 2 for FEDADAM satisfy", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 650, + 506, + 686 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 127, + 689, + 483, + 723 + ], + "lines": [ + { + "bbox": [ + 127, + 689, + 483, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 689, + 483, + 723 + ], + "score": 0.93, + "content": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\mathcal { O } \\Bigg ( \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\sqrt { m K T } } + \\frac { 2 \\sigma _ { l } ^ { 2 } L } { G ^ { 2 } \\sqrt { m K T } } + \\frac { \\sigma ^ { 2 } } { G K T } + \\frac { \\sigma ^ { 2 } L \\sqrt { m } } { G ^ { 2 } \\sqrt { K } T ^ { 3 / 2 } } \\Bigg ) .", + "type": "interline_equation", + "image_path": "8506f0697321f1583c8f1fe7772c08b405c66458be413e11447b3a78200c04f4.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 127, + 689, + 483, + 700.3333333333334 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 127, + 700.3333333333334, + 483, + 711.6666666666667 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 127, + 711.6666666666667, + 483, + 723.0000000000001 + ], + "spans": [], + "index": 43 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 254, + 96 + ], + "score": 1.0, + "content": "Remark 1. The server learning rate", + "type": "text" + }, + { + "bbox": [ + 254, + 83, + 279, + 94 + ], + "score": 0.91, + "content": "\\eta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "typically used in FEDAVG does not necessarily minimize", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 296, + 108 + ], + "score": 1.0, + "content": "the upper bound in Theorems 1 & 2. The effect of", + "type": "text" + }, + { + "bbox": [ + 296, + 95, + 308, + 106 + ], + "score": 0.84, + "content": "\\cdot _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 92, + 505, + 108 + ], + "score": 1.0, + "content": ", a measure of client heterogeneity, on convergence", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 264, + 117 + ], + "score": 1.0, + "content": "can be reduced by choosing sufficiently", + "type": "text" + }, + { + "bbox": [ + 264, + 106, + 273, + 116 + ], + "score": 0.65, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 104, + 369, + 117 + ], + "score": 1.0, + "content": "and a reasonably large", + "type": "text" + }, + { + "bbox": [ + 369, + 106, + 376, + 116 + ], + "score": 0.28, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "(e.g. see Corollary 1). Thus, the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "effect of client heterogeneity can be reduced by carefully choosing client and server learning rates,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 460, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 460, + 138 + ], + "score": 1.0, + "content": "but not removed entirely. Our empirical analysis (eg. Figure 1) supports this conclusion.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 504, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 157 + ], + "score": 1.0, + "content": "Discussion. We briefly discuss our theoretical analysis and its implications in the FL setting. The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 156, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 505, + 168 + ], + "score": 1.0, + "content": "convergence rates for FEDADAGRAD and FEDADAM are similar, so our discussion applies to all the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 460, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 460, + 180 + ], + "score": 1.0, + "content": "adaptive federated optimization algorithms (including FEDYOGI) proposed in the paper.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 108, + 183, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 107, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 107, + 182, + 296, + 196 + ], + "score": 1.0, + "content": "(i) Comparison of convergence rates. When", + "type": "text" + }, + { + "bbox": [ + 297, + 183, + 306, + 193 + ], + "score": 0.78, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 182, + 440, + 196 + ], + "score": 1.0, + "content": "is sufficiently large compared to", + "type": "text" + }, + { + "bbox": [ + 441, + 183, + 451, + 193 + ], + "score": 0.61, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 182, + 455, + 196 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 456, + 183, + 505, + 195 + ], + "score": 0.9, + "content": "\\mathcal { O } ( { } ^ { 1 } / { \\sqrt { m K T } } )", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 120, + 192, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 120, + 192, + 266, + 207 + ], + "score": 1.0, + "content": "is the dominant term in Corollary", + "type": "text" + }, + { + "bbox": [ + 266, + 194, + 292, + 204 + ], + "score": 0.28, + "content": "1 \\ \\& \\ 2", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 192, + 506, + 207 + ], + "score": 1.0, + "content": ". Thus, we effectively obtain a convergence rate of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 123, + 204, + 172, + 217 + ], + "score": 0.91, + "content": "\\mathcal { O } ( { } ^ { 1 } / \\sqrt { m K T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 204, + 506, + 216 + ], + "score": 1.0, + "content": ", which matches the best known rate for the general non-convex setting of our interest", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 121, + 214, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 121, + 214, + 369, + 229 + ], + "score": 1.0, + "content": "(e.g. see (Karimireddy et al., 2019)). We also note that in the", + "type": "text" + }, + { + "bbox": [ + 370, + 216, + 389, + 226 + ], + "score": 0.28, + "content": "i . i . d", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 214, + 506, + 229 + ], + "score": 1.0, + "content": "setting considered in (Wang", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 121, + 224, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 121, + 224, + 275, + 239 + ], + "score": 1.0, + "content": "& Joshi, 2018), which corresponds to", + "type": "text" + }, + { + "bbox": [ + 275, + 227, + 305, + 238 + ], + "score": 0.92, + "content": "\\sigma _ { g } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 224, + 506, + 239 + ], + "score": 1.0, + "content": ", we match their convergence rates. Similar to the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 121, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 121, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "centralized setting, it is possible to obtain convergence rates with better dependence on constants", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 122, + 248, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 122, + 248, + 505, + 259 + ], + "score": 1.0, + "content": "for federated adaptive methods, compared to FEDAVG, by incorporating non-uniform bounds on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 121, + 258, + 323, + 271 + ], + "spans": [ + { + "bbox": [ + 121, + 258, + 323, + 271 + ], + "score": 1.0, + "content": "gradients across coordinates (Zaheer et al., 2018).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 187, + 283 + ], + "score": 1.0, + "content": "(ii) Learning rates", + "type": "text" + }, + { + "bbox": [ + 188, + 271, + 197, + 281 + ], + "score": 0.35, + "content": "\\pmb { \\& }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 271, + 357, + 283 + ], + "score": 1.0, + "content": "their decay. The client learning rate of", + "type": "text" + }, + { + "bbox": [ + 357, + 271, + 379, + 282 + ], + "score": 0.92, + "content": "^ 1 / \\sqrt { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 271, + 506, + 283 + ], + "score": 1.0, + "content": "in our analysis requires knowl-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 122, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 122, + 281, + 240, + 294 + ], + "score": 1.0, + "content": "edge of the number of rounds", + "type": "text" + }, + { + "bbox": [ + 241, + 282, + 249, + 291 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "a priori; however, it is easy to generalize our analysis to the case", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 121, + 291, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 121, + 291, + 149, + 306 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 150, + 294, + 159, + 304 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 291, + 253, + 306 + ], + "score": 1.0, + "content": "is decayed at a rate of", + "type": "text" + }, + { + "bbox": [ + 254, + 292, + 272, + 304 + ], + "score": 0.9, + "content": "^ { 1 } / \\sqrt { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 291, + 397, + 306 + ], + "score": 1.0, + "content": ". Observe that one must decay", + "type": "text" + }, + { + "bbox": [ + 397, + 294, + 406, + 304 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 291, + 505, + 306 + ], + "score": 1.0, + "content": ", not the server learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 121, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 121, + 303, + 140, + 315 + ], + "score": 1.0, + "content": "rate", + "type": "text" + }, + { + "bbox": [ + 141, + 305, + 147, + 315 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 303, + 505, + 315 + ], + "score": 1.0, + "content": ", to obtain convergence. This is because the client drift introduced by the local updates", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 121, + 312, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 121, + 312, + 201, + 327 + ], + "score": 1.0, + "content": "does not vanish as", + "type": "text" + }, + { + "bbox": [ + 201, + 314, + 237, + 324 + ], + "score": 0.91, + "content": "T \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 312, + 263, + 327 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 263, + 315, + 273, + 325 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 312, + 506, + 327 + ], + "score": 1.0, + "content": "is constant. As we show in Appendix E.6, learning rate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 121, + 323, + 504, + 338 + ], + "spans": [ + { + "bbox": [ + 121, + 323, + 469, + 338 + ], + "score": 1.0, + "content": "decay can improve empirical performance. Also, note the inverse relationship between", + "type": "text" + }, + { + "bbox": [ + 470, + 326, + 479, + 336 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 323, + 497, + 338 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 497, + 326, + 504, + 336 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 121, + 335, + 458, + 347 + ], + "spans": [ + { + "bbox": [ + 121, + 335, + 458, + 347 + ], + "score": 1.0, + "content": "in Corollary 1 & 2, which we observe in our empirical analysis (see Appendix E.4).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 104, + 348, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 103, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 103, + 347, + 238, + 360 + ], + "score": 1.0, + "content": "(iii) Communication efficiency", + "type": "text" + }, + { + "bbox": [ + 239, + 347, + 248, + 358 + ], + "score": 0.33, + "content": "\\pmb { \\& }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "local steps. The total communication cost of the algorithms", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 121, + 357, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 121, + 357, + 317, + 371 + ], + "score": 1.0, + "content": "depends on the number of communication rounds", + "type": "text" + }, + { + "bbox": [ + 317, + 358, + 326, + 368 + ], + "score": 0.74, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 357, + 506, + 371 + ], + "score": 1.0, + "content": ". From Corollary 1 & 2, it is clear that a larger", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 123, + 368, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 123, + 369, + 133, + 379 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 368, + 343, + 381 + ], + "score": 1.0, + "content": "leads to fewer rounds of communication as long as", + "type": "text" + }, + { + "bbox": [ + 343, + 368, + 416, + 382 + ], + "score": 0.92, + "content": "K = \\mathcal { O } ( T \\bar { \\sigma } _ { l } ^ { 2 } / \\sigma _ { g } ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 368, + 506, + 381 + ], + "score": 1.0, + "content": ". Thus, the number of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 122, + 381, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 122, + 381, + 320, + 394 + ], + "score": 1.0, + "content": "local iterations can be large when either the ratio", + "type": "text" + }, + { + "bbox": [ + 320, + 381, + 347, + 395 + ], + "score": 0.94, + "content": "\\sigma _ { l } ^ { 2 } / \\sigma _ { g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 381, + 359, + 394 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 360, + 382, + 368, + 392 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 381, + 429, + 394 + ], + "score": 1.0, + "content": "is large. In the", + "type": "text" + }, + { + "bbox": [ + 429, + 382, + 448, + 392 + ], + "score": 0.37, + "content": "i . i . d", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "setting where", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 123, + 393, + 302, + 406 + ], + "spans": [ + { + "bbox": [ + 123, + 393, + 152, + 405 + ], + "score": 0.91, + "content": "\\sigma _ { g } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 393, + 217, + 406 + ], + "score": 1.0, + "content": ", unsurprisingly,", + "type": "text" + }, + { + "bbox": [ + 218, + 393, + 228, + 403 + ], + "score": 0.75, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 393, + 302, + 406 + ], + "score": 1.0, + "content": "can be very large.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 105, + 405, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 102, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 102, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "(iv) Client heterogeneity. While careful selection of client and server learning rates can reduce", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 122, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 122, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "the effect of client heterogeneity (see Remark 1), it does not completely remove it. In highly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 122, + 426, + 507, + 439 + ], + "spans": [ + { + "bbox": [ + 122, + 426, + 507, + 439 + ], + "score": 1.0, + "content": "heterogeneous settings, it may be necessary to use mechanisms such as control variates (Karim-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 121, + 435, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 121, + 435, + 505, + 451 + ], + "score": 1.0, + "content": "ireddy et al., 2019). However, our empirical analysis suggest that for moderate, naturally arising", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 122, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 122, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "heterogeneity, adaptive optimizers are quite effective, especially in cross-device settings (see", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 122, + 459, + 478, + 470 + ], + "spans": [ + { + "bbox": [ + 122, + 459, + 478, + 470 + ], + "score": 1.0, + "content": "Figure 1). Furthermore, our algorithms can be directly combined with such mechanisms.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 463, + 488 + ], + "score": 1.0, + "content": "As mentioned earlier, for the sake of simplicity, our analysis assumes full-participation", + "type": "text" + }, + { + "bbox": [ + 463, + 475, + 502, + 487 + ], + "score": 0.86, + "content": "\\mathcal { S } = [ m ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 474, + 507, + 488 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 487, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 506, + 498 + ], + "score": 1.0, + "content": "Our analysis can be directly generalized to limited participation at the cost of an additional variance", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 497, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 236, + 509 + ], + "score": 1.0, + "content": "term in our rates that depends on", + "type": "text" + }, + { + "bbox": [ + 236, + 497, + 263, + 509 + ], + "score": 0.92, + "content": "| { \\cal S } | / m", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 497, + 506, + 509 + ], + "score": 1.0, + "content": ", the fraction of clients sampled (see Section A.2.1 for details).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 105, + 523, + 461, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 463, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 463, + 537 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL EVALUATION: DATASETS, TASKS, AND METHODS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "We evaluate our algorithms on what we believe is the most extensive and representative suite of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 557, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 571 + ], + "score": 1.0, + "content": "federated datasets and modeling tasks to date. We wish to understand how server adaptivity can help", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "improve convergence, especially in cross-device settings. To accomplish this, we conduct simulations", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "on seven diverse and representative learning tasks across five datasets. Notably, three of the five have", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 590, + 454, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 454, + 603 + ], + "score": 1.0, + "content": "a naturally-arising client partitioning, highly representative of real-world FL problems.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 607, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 507, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 507, + 619 + ], + "score": 1.0, + "content": "Datasets, models, and tasks We use five datasets: CIFAR-10, CIFAR-100 (Krizhevsky & Hinton,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "2009), EMNIST (Cohen et al., 2017), Shakespeare (McMahan et al., 2017), and Stack Overflow (Au-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "score": 1.0, + "content": "thors, 2019). The first three are image datasets, the last two are text datasets. For CIFAR-10 and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 638, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 652 + ], + "score": 1.0, + "content": "CIFAR-100, we train ResNet-18 (replacing batch norm with group norm (Hsieh et al., 2019)). For", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "EMNIST, we train a CNN for character recognition (EMNIST CR) and a bottleneck autoencoder", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 660, + 507, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 507, + 674 + ], + "score": 1.0, + "content": "(EMNIST AE). For Shakespeare, we train an RNN for next-character-prediction. For Stack Overflow,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "we perform tag prediction using logistic regression on bag-of-words vectors (SO LR) and train an", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 681, + 479, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 479, + 694 + ], + "score": 1.0, + "content": "RNN to do next-word-prediction (SO NWP). For full details of the datasets, see Appendix C.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Implementation We implement all algorithms in TensorFlow Federated (Ingerman & Ostrowski,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "2019). Clients are sampled uniformly at random, without replacement in a given round, but with", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "replacement across rounds. Our implementation has two important characteristics. First, instead of", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 52 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 137 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 254, + 96 + ], + "score": 1.0, + "content": "Remark 1. The server learning rate", + "type": "text" + }, + { + "bbox": [ + 254, + 83, + 279, + 94 + ], + "score": 0.91, + "content": "\\eta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "typically used in FEDAVG does not necessarily minimize", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 296, + 108 + ], + "score": 1.0, + "content": "the upper bound in Theorems 1 & 2. The effect of", + "type": "text" + }, + { + "bbox": [ + 296, + 95, + 308, + 106 + ], + "score": 0.84, + "content": "\\cdot _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 92, + 505, + 108 + ], + "score": 1.0, + "content": ", a measure of client heterogeneity, on convergence", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 264, + 117 + ], + "score": 1.0, + "content": "can be reduced by choosing sufficiently", + "type": "text" + }, + { + "bbox": [ + 264, + 106, + 273, + 116 + ], + "score": 0.65, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 104, + 369, + 117 + ], + "score": 1.0, + "content": "and a reasonably large", + "type": "text" + }, + { + "bbox": [ + 369, + 106, + 376, + 116 + ], + "score": 0.28, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "(e.g. see Corollary 1). Thus, the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "effect of client heterogeneity can be reduced by carefully choosing client and server learning rates,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 460, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 460, + 138 + ], + "score": 1.0, + "content": "but not removed entirely. Our empirical analysis (eg. Figure 1) supports this conclusion.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 81, + 506, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 504, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 157 + ], + "score": 1.0, + "content": "Discussion. We briefly discuss our theoretical analysis and its implications in the FL setting. The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 156, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 505, + 168 + ], + "score": 1.0, + "content": "convergence rates for FEDADAGRAD and FEDADAM are similar, so our discussion applies to all the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 460, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 460, + 180 + ], + "score": 1.0, + "content": "adaptive federated optimization algorithms (including FEDYOGI) proposed in the paper.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 145, + 505, + 180 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 183, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 107, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 107, + 182, + 296, + 196 + ], + "score": 1.0, + "content": "(i) Comparison of convergence rates. When", + "type": "text" + }, + { + "bbox": [ + 297, + 183, + 306, + 193 + ], + "score": 0.78, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 182, + 440, + 196 + ], + "score": 1.0, + "content": "is sufficiently large compared to", + "type": "text" + }, + { + "bbox": [ + 441, + 183, + 451, + 193 + ], + "score": 0.61, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 182, + 455, + 196 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 456, + 183, + 505, + 195 + ], + "score": 0.9, + "content": "\\mathcal { O } ( { } ^ { 1 } / { \\sqrt { m K T } } )", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 120, + 192, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 120, + 192, + 266, + 207 + ], + "score": 1.0, + "content": "is the dominant term in Corollary", + "type": "text" + }, + { + "bbox": [ + 266, + 194, + 292, + 204 + ], + "score": 0.28, + "content": "1 \\ \\& \\ 2", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 192, + 506, + 207 + ], + "score": 1.0, + "content": ". Thus, we effectively obtain a convergence rate of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 123, + 204, + 172, + 217 + ], + "score": 0.91, + "content": "\\mathcal { O } ( { } ^ { 1 } / \\sqrt { m K T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 204, + 506, + 216 + ], + "score": 1.0, + "content": ", which matches the best known rate for the general non-convex setting of our interest", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 121, + 214, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 121, + 214, + 369, + 229 + ], + "score": 1.0, + "content": "(e.g. see (Karimireddy et al., 2019)). We also note that in the", + "type": "text" + }, + { + "bbox": [ + 370, + 216, + 389, + 226 + ], + "score": 0.28, + "content": "i . i . d", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 214, + 506, + 229 + ], + "score": 1.0, + "content": "setting considered in (Wang", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 121, + 224, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 121, + 224, + 275, + 239 + ], + "score": 1.0, + "content": "& Joshi, 2018), which corresponds to", + "type": "text" + }, + { + "bbox": [ + 275, + 227, + 305, + 238 + ], + "score": 0.92, + "content": "\\sigma _ { g } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 224, + 506, + 239 + ], + "score": 1.0, + "content": ", we match their convergence rates. Similar to the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 121, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 121, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "centralized setting, it is possible to obtain convergence rates with better dependence on constants", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 122, + 248, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 122, + 248, + 505, + 259 + ], + "score": 1.0, + "content": "for federated adaptive methods, compared to FEDAVG, by incorporating non-uniform bounds on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 121, + 258, + 323, + 271 + ], + "spans": [ + { + "bbox": [ + 121, + 258, + 323, + 271 + ], + "score": 1.0, + "content": "gradients across coordinates (Zaheer et al., 2018).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 107, + 182, + 506, + 271 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 346 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 187, + 283 + ], + "score": 1.0, + "content": "(ii) Learning rates", + "type": "text" + }, + { + "bbox": [ + 188, + 271, + 197, + 281 + ], + "score": 0.35, + "content": "\\pmb { \\& }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 271, + 357, + 283 + ], + "score": 1.0, + "content": "their decay. The client learning rate of", + "type": "text" + }, + { + "bbox": [ + 357, + 271, + 379, + 282 + ], + "score": 0.92, + "content": "^ 1 / \\sqrt { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 271, + 506, + 283 + ], + "score": 1.0, + "content": "in our analysis requires knowl-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 122, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 122, + 281, + 240, + 294 + ], + "score": 1.0, + "content": "edge of the number of rounds", + "type": "text" + }, + { + "bbox": [ + 241, + 282, + 249, + 291 + ], + "score": 0.8, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "a priori; however, it is easy to generalize our analysis to the case", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 121, + 291, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 121, + 291, + 149, + 306 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 150, + 294, + 159, + 304 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 291, + 253, + 306 + ], + "score": 1.0, + "content": "is decayed at a rate of", + "type": "text" + }, + { + "bbox": [ + 254, + 292, + 272, + 304 + ], + "score": 0.9, + "content": "^ { 1 } / \\sqrt { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 291, + 397, + 306 + ], + "score": 1.0, + "content": ". Observe that one must decay", + "type": "text" + }, + { + "bbox": [ + 397, + 294, + 406, + 304 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 291, + 505, + 306 + ], + "score": 1.0, + "content": ", not the server learning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 121, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 121, + 303, + 140, + 315 + ], + "score": 1.0, + "content": "rate", + "type": "text" + }, + { + "bbox": [ + 141, + 305, + 147, + 315 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 303, + 505, + 315 + ], + "score": 1.0, + "content": ", to obtain convergence. This is because the client drift introduced by the local updates", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 121, + 312, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 121, + 312, + 201, + 327 + ], + "score": 1.0, + "content": "does not vanish as", + "type": "text" + }, + { + "bbox": [ + 201, + 314, + 237, + 324 + ], + "score": 0.91, + "content": "T \\to \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 312, + 263, + 327 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 263, + 315, + 273, + 325 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 312, + 506, + 327 + ], + "score": 1.0, + "content": "is constant. As we show in Appendix E.6, learning rate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 121, + 323, + 504, + 338 + ], + "spans": [ + { + "bbox": [ + 121, + 323, + 469, + 338 + ], + "score": 1.0, + "content": "decay can improve empirical performance. Also, note the inverse relationship between", + "type": "text" + }, + { + "bbox": [ + 470, + 326, + 479, + 336 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 323, + 497, + 338 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 497, + 326, + 504, + 336 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 121, + 335, + 458, + 347 + ], + "spans": [ + { + "bbox": [ + 121, + 335, + 458, + 347 + ], + "score": 1.0, + "content": "in Corollary 1 & 2, which we observe in our empirical analysis (see Appendix E.4).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 271, + 506, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 348, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 103, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 103, + 347, + 238, + 360 + ], + "score": 1.0, + "content": "(iii) Communication efficiency", + "type": "text" + }, + { + "bbox": [ + 239, + 347, + 248, + 358 + ], + "score": 0.33, + "content": "\\pmb { \\& }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "local steps. The total communication cost of the algorithms", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 121, + 357, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 121, + 357, + 317, + 371 + ], + "score": 1.0, + "content": "depends on the number of communication rounds", + "type": "text" + }, + { + "bbox": [ + 317, + 358, + 326, + 368 + ], + "score": 0.74, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 357, + 506, + 371 + ], + "score": 1.0, + "content": ". From Corollary 1 & 2, it is clear that a larger", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 123, + 368, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 123, + 369, + 133, + 379 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 368, + 343, + 381 + ], + "score": 1.0, + "content": "leads to fewer rounds of communication as long as", + "type": "text" + }, + { + "bbox": [ + 343, + 368, + 416, + 382 + ], + "score": 0.92, + "content": "K = \\mathcal { O } ( T \\bar { \\sigma } _ { l } ^ { 2 } / \\sigma _ { g } ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 368, + 506, + 381 + ], + "score": 1.0, + "content": ". Thus, the number of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 122, + 381, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 122, + 381, + 320, + 394 + ], + "score": 1.0, + "content": "local iterations can be large when either the ratio", + "type": "text" + }, + { + "bbox": [ + 320, + 381, + 347, + 395 + ], + "score": 0.94, + "content": "\\sigma _ { l } ^ { 2 } / \\sigma _ { g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 381, + 359, + 394 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 360, + 382, + 368, + 392 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 381, + 429, + 394 + ], + "score": 1.0, + "content": "is large. In the", + "type": "text" + }, + { + "bbox": [ + 429, + 382, + 448, + 392 + ], + "score": 0.37, + "content": "i . i . d", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "setting where", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 123, + 393, + 302, + 406 + ], + "spans": [ + { + "bbox": [ + 123, + 393, + 152, + 405 + ], + "score": 0.91, + "content": "\\sigma _ { g } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 393, + 217, + 406 + ], + "score": 1.0, + "content": ", unsurprisingly,", + "type": "text" + }, + { + "bbox": [ + 218, + 393, + 228, + 403 + ], + "score": 0.75, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 393, + 302, + 406 + ], + "score": 1.0, + "content": "can be very large.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 103, + 347, + 506, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 405, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 102, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 102, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "(iv) Client heterogeneity. While careful selection of client and server learning rates can reduce", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 122, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 122, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "the effect of client heterogeneity (see Remark 1), it does not completely remove it. In highly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 122, + 426, + 507, + 439 + ], + "spans": [ + { + "bbox": [ + 122, + 426, + 507, + 439 + ], + "score": 1.0, + "content": "heterogeneous settings, it may be necessary to use mechanisms such as control variates (Karim-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 121, + 435, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 121, + 435, + 505, + 451 + ], + "score": 1.0, + "content": "ireddy et al., 2019). However, our empirical analysis suggest that for moderate, naturally arising", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 122, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 122, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "heterogeneity, adaptive optimizers are quite effective, especially in cross-device settings (see", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 122, + 459, + 478, + 470 + ], + "spans": [ + { + "bbox": [ + 122, + 459, + 478, + 470 + ], + "score": 1.0, + "content": "Figure 1). Furthermore, our algorithms can be directly combined with such mechanisms.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 102, + 405, + 507, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 463, + 488 + ], + "score": 1.0, + "content": "As mentioned earlier, for the sake of simplicity, our analysis assumes full-participation", + "type": "text" + }, + { + "bbox": [ + 463, + 475, + 502, + 487 + ], + "score": 0.86, + "content": "\\mathcal { S } = [ m ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 474, + 507, + 488 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 487, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 506, + 498 + ], + "score": 1.0, + "content": "Our analysis can be directly generalized to limited participation at the cost of an additional variance", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 497, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 236, + 509 + ], + "score": 1.0, + "content": "term in our rates that depends on", + "type": "text" + }, + { + "bbox": [ + 236, + 497, + 263, + 509 + ], + "score": 0.92, + "content": "| { \\cal S } | / m", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 497, + 506, + 509 + ], + "score": 1.0, + "content": ", the fraction of clients sampled (see Section A.2.1 for details).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 474, + 507, + 509 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 523, + 461, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 463, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 463, + 537 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL EVALUATION: DATASETS, TASKS, AND METHODS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 522, + 463, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "We evaluate our algorithms on what we believe is the most extensive and representative suite of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 557, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 505, + 571 + ], + "score": 1.0, + "content": "federated datasets and modeling tasks to date. We wish to understand how server adaptivity can help", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "improve convergence, especially in cross-device settings. To accomplish this, we conduct simulations", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "on seven diverse and representative learning tasks across five datasets. Notably, three of the five have", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 590, + 454, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 454, + 603 + ], + "score": 1.0, + "content": "a naturally-arising client partitioning, highly representative of real-world FL problems.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 548, + 506, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 607, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 507, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 507, + 619 + ], + "score": 1.0, + "content": "Datasets, models, and tasks We use five datasets: CIFAR-10, CIFAR-100 (Krizhevsky & Hinton,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "2009), EMNIST (Cohen et al., 2017), Shakespeare (McMahan et al., 2017), and Stack Overflow (Au-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "score": 1.0, + "content": "thors, 2019). The first three are image datasets, the last two are text datasets. For CIFAR-10 and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 638, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 652 + ], + "score": 1.0, + "content": "CIFAR-100, we train ResNet-18 (replacing batch norm with group norm (Hsieh et al., 2019)). For", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "EMNIST, we train a CNN for character recognition (EMNIST CR) and a bottleneck autoencoder", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 660, + 507, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 507, + 674 + ], + "score": 1.0, + "content": "(EMNIST AE). For Shakespeare, we train an RNN for next-character-prediction. For Stack Overflow,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "we perform tag prediction using logistic regression on bag-of-words vectors (SO LR) and train an", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 681, + 479, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 479, + 694 + ], + "score": 1.0, + "content": "RNN to do next-word-prediction (SO NWP). For full details of the datasets, see Appendix C.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 606, + 507, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Implementation We implement all algorithms in TensorFlow Federated (Ingerman & Ostrowski,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "2019). Clients are sampled uniformly at random, without replacement in a given round, but with", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 721, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 732 + ], + "score": 1.0, + "content": "replacement across rounds. Our implementation has two important characteristics. First, instead of", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 82, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 132, + 94 + ], + "score": 1.0, + "content": "doing", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 133, + 83, + 143, + 93 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 144, + 82, + 274, + 94 + ], + "score": 1.0, + "content": "training steps per client, we do", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 275, + 83, + 284, + 93 + ], + "score": 0.62, + "content": "E", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 285, + 82, + 506, + 94 + ], + "score": 1.0, + "content": "epochs of training over each client’s dataset. Second,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 105 + ], + "score": 1.0, + "content": "to account for varying numbers of gradient steps per client, we weight the average of the client", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 139, + 116 + ], + "score": 1.0, + "content": "outputs", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 140, + 104, + 152, + 116 + ], + "score": 0.88, + "content": "\\Delta _ { i } ^ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 153, + 104, + 505, + 116 + ], + "score": 1.0, + "content": "by each client’s number of training samples. This follows the approach of (McMahan", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "et al., 2017), and can often outperform uniform weighting (Zaheer et al., 2018). For full descriptions", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 269, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 269, + 137 + ], + "score": 1.0, + "content": "of the algorithms used, see Appendix B.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 52, + "bbox_fs": [ + 106, + 699, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 132, + 94 + ], + "score": 1.0, + "content": "doing", + "type": "text" + }, + { + "bbox": [ + 133, + 83, + 143, + 93 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 82, + 274, + 94 + ], + "score": 1.0, + "content": "training steps per client, we do", + "type": "text" + }, + { + "bbox": [ + 275, + 83, + 284, + 93 + ], + "score": 0.62, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 82, + 506, + 94 + ], + "score": 1.0, + "content": "epochs of training over each client’s dataset. Second,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 105 + ], + "score": 1.0, + "content": "to account for varying numbers of gradient steps per client, we weight the average of the client", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 139, + 116 + ], + "score": 1.0, + "content": "outputs", + "type": "text" + }, + { + "bbox": [ + 140, + 104, + 152, + 116 + ], + "score": 0.88, + "content": "\\Delta _ { i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 104, + 505, + 116 + ], + "score": 1.0, + "content": "by each client’s number of training samples. This follows the approach of (McMahan", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "et al., 2017), and can often outperform uniform weighting (Zaheer et al., 2018). For full descriptions", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 269, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 269, + 137 + ], + "score": 1.0, + "content": "of the algorithms used, see Appendix B.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 505, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "Optimizers and hyperparameters We compare FEDADAGRAD, FEDADAM, and FEDYOGI (with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 149, + 166 + ], + "score": 1.0, + "content": "adaptivity", + "type": "text" + }, + { + "bbox": [ + 149, + 155, + 156, + 163 + ], + "score": 0.65, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 152, + 477, + 166 + ], + "score": 1.0, + "content": ") to FEDOPT where CLIENTOPT and SERVEROPT are SGD with learning rates", + "type": "text" + }, + { + "bbox": [ + 477, + 155, + 487, + 164 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 152, + 506, + 166 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 164, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 167, + 113, + 175 + ], + "score": 0.73, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 164, + 505, + 176 + ], + "score": 1.0, + "content": ". For the server, we use a momentum parameter of 0 (FEDAVG), and 0.9 (FEDAVGM). We fix the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 174, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 506, + 187 + ], + "score": 1.0, + "content": "client batch size on a per-task level (see Appendix D.3). For FEDADAM and FEDYOGI, we fix a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 199, + 198 + ], + "score": 1.0, + "content": "momentum parameter", + "type": "text" + }, + { + "bbox": [ + 199, + 185, + 237, + 196 + ], + "score": 0.92, + "content": "\\beta _ { 1 } = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 185, + 374, + 198 + ], + "score": 1.0, + "content": "and a second moment parameter", + "type": "text" + }, + { + "bbox": [ + 374, + 185, + 417, + 196 + ], + "score": 0.91, + "content": "\\beta _ { 2 } = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 185, + 506, + 198 + ], + "score": 1.0, + "content": ". We also compare to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "SCAFFOLD (see Appendix B.2 for implementation details). For SO NWP, we sample 50 clients per", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 207, + 451, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 322, + 219 + ], + "score": 1.0, + "content": "round, while for all other tasks we sample 10. We use", + "type": "text" + }, + { + "bbox": [ + 322, + 207, + 350, + 217 + ], + "score": 0.9, + "content": "E = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 207, + 451, + 219 + ], + "score": 1.0, + "content": "local epochs throughout.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 223, + 505, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 149, + 235 + ], + "score": 1.0, + "content": "We select", + "type": "text" + }, + { + "bbox": [ + 150, + 225, + 171, + 235 + ], + "score": 0.63, + "content": "\\eta , ~ \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 222, + 194, + 235 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 194, + 226, + 201, + 233 + ], + "score": 0.71, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "by grid-search tuning. While this is often done using validation data in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "centralized settings, such data is often inaccessible in FL, especially cross-device FL. Therefore, we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "tune by selecting the parameters that minimize the average training loss over the last 100 rounds", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 256, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 505, + 267 + ], + "score": 1.0, + "content": "of training. We run 1500 rounds of training on the EMNIST CR, Shakespeare, and Stack Overflow", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "tasks, 3000 rounds for EMNIST AE, and 4000 rounds for the CIFAR tasks. For more details and a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 277, + 319, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 319, + 290 + ], + "score": 1.0, + "content": "record of the best hyperparameters, see Appendix D.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 507, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 507, + 307 + ], + "score": 1.0, + "content": "Validation metrics For all tasks, we measure the performance on a validation set throughout training.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "For Stack Overflow tasks, the validation set contains 10,000 randomly sampled test examples (due", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "to the size of the test dataset, see Table 2). For all other tasks, we use the entire test set. Since", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "score": 1.0, + "content": "all algorithms exchange equal-sized objects between server and clients, we use the number of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 336, + 358, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 358, + 349 + ], + "score": 1.0, + "content": "communication rounds as a proxy for wall-clock training time.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 363, + 330, + 376 + ], + "lines": [ + { + "bbox": [ + 104, + 362, + 332, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 332, + 378 + ], + "score": 1.0, + "content": "5 EXPERIMENTAL EVALUATION: RESULTS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 109, + 388, + 282, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 284, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 284, + 399 + ], + "score": 1.0, + "content": "5.1 COMPARISONS BETWEEN METHODS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "We compare the convergence of our adaptive methods to non-adaptive methods: FEDAVG, FEDAVGM", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "score": 1.0, + "content": "and SCAFFOLD. Plots of validation performances for each task/optimizer are in Figure 1, and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "Table 1 summarizes the last-100-round validation performance. Due to space constraints, results for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 440, + 502, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 502, + 452 + ], + "score": 1.0, + "content": "EMNIST CR are in Appendix E.1, and full test set results for Stack Overflow are in Appendix E.2.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 457, + 277, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 277, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 277, + 468 + ], + "score": 1.0, + "content": "Sparse-gradient tasks Text data often", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 467, + 279, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 279, + 480 + ], + "score": 1.0, + "content": "produces long-tailed feature distributions,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 478, + 277, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 277, + 491 + ], + "score": 1.0, + "content": "leading to approximately-sparse gradients", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 489, + 276, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 276, + 500 + ], + "score": 1.0, + "content": "which adaptive optimizers can capitalize", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 500, + 279, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 279, + 511 + ], + "score": 1.0, + "content": "on (Zhang et al., 2019a). Both Stack Over-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 510, + 277, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 277, + 523 + ], + "score": 1.0, + "content": "flow tasks exhibit such behavior, though", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 522, + 278, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 278, + 533 + ], + "score": 1.0, + "content": "they are otherwise dramatically different—", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 532, + 279, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 279, + 544 + ], + "score": 1.0, + "content": "in feature representation (bag-of-words vs.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 543, + 279, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 279, + 554 + ], + "score": 1.0, + "content": "variable-length token sequence), model ar-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 552, + 278, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 278, + 565 + ], + "score": 1.0, + "content": "chitecture (GLM vs deep network), and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 565, + 279, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 279, + 576 + ], + "score": 1.0, + "content": "optimization landscape (convex vs non-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 575, + 278, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 278, + 586 + ], + "score": 1.0, + "content": "convex). In both tasks, words that do not", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 586, + 279, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 279, + 598 + ], + "score": 1.0, + "content": "appear in a client’s dataset produce near-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 597, + 277, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 277, + 608 + ], + "score": 1.0, + "content": "zero client updates. Thus, the accumulators", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 107, + 607, + 279, + 619 + ], + "spans": [ + { + "bbox": [ + 107, + 609, + 122, + 619 + ], + "score": 0.88, + "content": "v _ { t , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 607, + 279, + 619 + ], + "score": 1.0, + "content": "in Algorithm 2 remain small for pa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 618, + 277, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 277, + 630 + ], + "score": 1.0, + "content": "rameters tied to rare words, allowing large", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 36.5 + }, + { + "type": "table", + "bbox": [ + 288, + 459, + 499, + 561 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 288, + 459, + 499, + 561 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 288, + 459, + 499, + 561 + ], + "spans": [ + { + "bbox": [ + 288, + 459, + 499, + 561 + ], + "score": 0.978, + "html": "
FED..ADAGRADADAMYoGIAvGMAvG
CIFAR-1072.177.478.077.472.8
CIFAR-10047.952.552.452.444.7
EMNIST CR85.185.685.585.284.9
SHAKESPEARE57.557.057.257.356.9
SO NWP23.825.225.223.819.5
SO LR67.165.865.936.930.0
EMNIST AE4.201.010.981.656.47
", + "type": "table", + "image_path": "241593e968cb0ffd4a91a36315da5518b624bd18588c254e50d8dbf78f761dc6.jpg" + } + ] + } + ], + "index": 48.5, + "virtual_lines": [ + { + "bbox": [ + 288, + 459, + 499, + 471.75 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 288, + 471.75, + 499, + 484.5 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 288, + 484.5, + 499, + 497.25 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 288, + 497.25, + 499, + 510.0 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 288, + 510.0, + 499, + 522.75 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 288, + 522.75, + 499, + 535.5 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 288, + 535.5, + 499, + 548.25 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 288, + 548.25, + 499, + 561.0 + ], + "spans": [], + "index": 52 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 284, + 565, + 505, + 619 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 284, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 284, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "Table 1: Average validation performance over the last", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 284, + 575, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 284, + 575, + 331, + 588 + ], + "score": 1.0, + "content": "100 rounds:", + "type": "text" + }, + { + "bbox": [ + 332, + 576, + 342, + 587 + ], + "score": 0.82, + "content": "\\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 575, + 458, + 588 + ], + "score": 1.0, + "content": "accuracy for rows 1–5; Recall", + "type": "text" + }, + { + "bbox": [ + 459, + 576, + 473, + 586 + ], + "score": 0.64, + "content": "\\textcircled { a } 5", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 576, + 504, + 587 + ], + "score": 0.79, + "content": "( \\times 1 0 0 )", + "type": "inline_equation" + } + ], + "index": 54 + }, + { + "bbox": [ + 283, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 283, + 586, + 416, + 599 + ], + "score": 1.0, + "content": "for Stack Overflow LR; and MSE", + "type": "text" + }, + { + "bbox": [ + 416, + 587, + 452, + 598 + ], + "score": 0.89, + "content": "( \\times 1 0 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "for EMNIST", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 284, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 284, + 597, + 381, + 609 + ], + "score": 1.0, + "content": "AE. Performance within", + "type": "text" + }, + { + "bbox": [ + 381, + 598, + 404, + 608 + ], + "score": 0.88, + "content": "0 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "of the best result for each", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 285, + 609, + 379, + 619 + ], + "spans": [ + { + "bbox": [ + 285, + 609, + 379, + 619 + ], + "score": 1.0, + "content": "task are shown in bold.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 55 + } + ], + "index": 51.75 + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 503, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "updates to be made when they do occur. This intuition is born out in Figure 1, where adaptive", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "optimizers dramatically outperform non-adaptive ones. For the non-convex NWP task, momentum is", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 650, + 414, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 414, + 662 + ], + "score": 1.0, + "content": "also critical, whereas it slightly hinders performance for the convex LR task.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 59 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "Dense-gradient tasks CIFAR-10/100, EMNIST AE/CR, and Shakespeare lack a sparse-gradient", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "structure. Shakespeare is relatively easy—most optimizers perform well after enough rounds once suit-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "ably tuned, though FEDADAGRAD converges faster. For CIFAR-10/100 and EMNIST AE, adaptivity", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "and momentum offer substantial improvements over FEDAVG. Moreover, FEDYOGI and FEDADAM", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "have faster initial convergence than FEDAVGM on these tasks. Notably, FEDADAM and FEDYOGI", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "score": 1.0, + "content": "perform comparably to or better than non-adaptive optimizers throughout, and close to or better than", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 63.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 7 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 136 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 137 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 505, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "Optimizers and hyperparameters We compare FEDADAGRAD, FEDADAM, and FEDYOGI (with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 149, + 166 + ], + "score": 1.0, + "content": "adaptivity", + "type": "text" + }, + { + "bbox": [ + 149, + 155, + 156, + 163 + ], + "score": 0.65, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 152, + 477, + 166 + ], + "score": 1.0, + "content": ") to FEDOPT where CLIENTOPT and SERVEROPT are SGD with learning rates", + "type": "text" + }, + { + "bbox": [ + 477, + 155, + 487, + 164 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 152, + 506, + 166 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 164, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 167, + 113, + 175 + ], + "score": 0.73, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 164, + 505, + 176 + ], + "score": 1.0, + "content": ". For the server, we use a momentum parameter of 0 (FEDAVG), and 0.9 (FEDAVGM). We fix the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 174, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 506, + 187 + ], + "score": 1.0, + "content": "client batch size on a per-task level (see Appendix D.3). For FEDADAM and FEDYOGI, we fix a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 185, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 199, + 198 + ], + "score": 1.0, + "content": "momentum parameter", + "type": "text" + }, + { + "bbox": [ + 199, + 185, + 237, + 196 + ], + "score": 0.92, + "content": "\\beta _ { 1 } = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 185, + 374, + 198 + ], + "score": 1.0, + "content": "and a second moment parameter", + "type": "text" + }, + { + "bbox": [ + 374, + 185, + 417, + 196 + ], + "score": 0.91, + "content": "\\beta _ { 2 } = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 185, + 506, + 198 + ], + "score": 1.0, + "content": ". We also compare to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "SCAFFOLD (see Appendix B.2 for implementation details). For SO NWP, we sample 50 clients per", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 207, + 451, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 322, + 219 + ], + "score": 1.0, + "content": "round, while for all other tasks we sample 10. We use", + "type": "text" + }, + { + "bbox": [ + 322, + 207, + 350, + 217 + ], + "score": 0.9, + "content": "E = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 207, + 451, + 219 + ], + "score": 1.0, + "content": "local epochs throughout.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 142, + 506, + 219 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 223, + 505, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 149, + 235 + ], + "score": 1.0, + "content": "We select", + "type": "text" + }, + { + "bbox": [ + 150, + 225, + 171, + 235 + ], + "score": 0.63, + "content": "\\eta , ~ \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 222, + 194, + 235 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 194, + 226, + 201, + 233 + ], + "score": 0.71, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "by grid-search tuning. While this is often done using validation data in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "centralized settings, such data is often inaccessible in FL, especially cross-device FL. Therefore, we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "tune by selecting the parameters that minimize the average training loss over the last 100 rounds", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 256, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 505, + 267 + ], + "score": 1.0, + "content": "of training. We run 1500 rounds of training on the EMNIST CR, Shakespeare, and Stack Overflow", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "tasks, 3000 rounds for EMNIST AE, and 4000 rounds for the CIFAR tasks. For more details and a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 277, + 319, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 319, + 290 + ], + "score": 1.0, + "content": "record of the best hyperparameters, see Appendix D.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 222, + 505, + 290 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 293, + 507, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 507, + 307 + ], + "score": 1.0, + "content": "Validation metrics For all tasks, we measure the performance on a validation set throughout training.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "For Stack Overflow tasks, the validation set contains 10,000 randomly sampled test examples (due", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "to the size of the test dataset, see Table 2). For all other tasks, we use the entire test set. Since", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "score": 1.0, + "content": "all algorithms exchange equal-sized objects between server and clients, we use the number of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 336, + 358, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 358, + 349 + ], + "score": 1.0, + "content": "communication rounds as a proxy for wall-clock training time.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 293, + 507, + 349 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 363, + 330, + 376 + ], + "lines": [ + { + "bbox": [ + 104, + 362, + 332, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 332, + 378 + ], + "score": 1.0, + "content": "5 EXPERIMENTAL EVALUATION: RESULTS", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 109, + 388, + 282, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 284, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 284, + 399 + ], + "score": 1.0, + "content": "5.1 COMPARISONS BETWEEN METHODS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "We compare the convergence of our adaptive methods to non-adaptive methods: FEDAVG, FEDAVGM", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 431 + ], + "score": 1.0, + "content": "and SCAFFOLD. Plots of validation performances for each task/optimizer are in Figure 1, and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "Table 1 summarizes the last-100-round validation performance. Due to space constraints, results for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 440, + 502, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 502, + 452 + ], + "score": 1.0, + "content": "EMNIST CR are in Appendix E.1, and full test set results for Stack Overflow are in Appendix E.2.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 407, + 506, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 457, + 277, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 277, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 277, + 468 + ], + "score": 1.0, + "content": "Sparse-gradient tasks Text data often", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 467, + 279, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 279, + 480 + ], + "score": 1.0, + "content": "produces long-tailed feature distributions,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 478, + 277, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 277, + 491 + ], + "score": 1.0, + "content": "leading to approximately-sparse gradients", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 489, + 276, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 276, + 500 + ], + "score": 1.0, + "content": "which adaptive optimizers can capitalize", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 500, + 279, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 279, + 511 + ], + "score": 1.0, + "content": "on (Zhang et al., 2019a). Both Stack Over-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 510, + 277, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 277, + 523 + ], + "score": 1.0, + "content": "flow tasks exhibit such behavior, though", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 522, + 278, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 278, + 533 + ], + "score": 1.0, + "content": "they are otherwise dramatically different—", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 532, + 279, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 279, + 544 + ], + "score": 1.0, + "content": "in feature representation (bag-of-words vs.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 543, + 279, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 279, + 554 + ], + "score": 1.0, + "content": "variable-length token sequence), model ar-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 552, + 278, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 278, + 565 + ], + "score": 1.0, + "content": "chitecture (GLM vs deep network), and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 565, + 279, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 279, + 576 + ], + "score": 1.0, + "content": "optimization landscape (convex vs non-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 575, + 278, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 278, + 586 + ], + "score": 1.0, + "content": "convex). In both tasks, words that do not", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 586, + 279, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 279, + 598 + ], + "score": 1.0, + "content": "appear in a client’s dataset produce near-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 597, + 277, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 277, + 608 + ], + "score": 1.0, + "content": "zero client updates. Thus, the accumulators", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 107, + 607, + 279, + 619 + ], + "spans": [ + { + "bbox": [ + 107, + 609, + 122, + 619 + ], + "score": 0.88, + "content": "v _ { t , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 607, + 279, + 619 + ], + "score": 1.0, + "content": "in Algorithm 2 remain small for pa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 618, + 277, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 277, + 630 + ], + "score": 1.0, + "content": "rameters tied to rare words, allowing large", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 457, + 279, + 630 + ] + }, + { + "type": "table", + "bbox": [ + 288, + 459, + 499, + 561 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 288, + 459, + 499, + 561 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 288, + 459, + 499, + 561 + ], + "spans": [ + { + "bbox": [ + 288, + 459, + 499, + 561 + ], + "score": 0.978, + "html": "
FED..ADAGRADADAMYoGIAvGMAvG
CIFAR-1072.177.478.077.472.8
CIFAR-10047.952.552.452.444.7
EMNIST CR85.185.685.585.284.9
SHAKESPEARE57.557.057.257.356.9
SO NWP23.825.225.223.819.5
SO LR67.165.865.936.930.0
EMNIST AE4.201.010.981.656.47
", + "type": "table", + "image_path": "241593e968cb0ffd4a91a36315da5518b624bd18588c254e50d8dbf78f761dc6.jpg" + } + ] + } + ], + "index": 48.5, + "virtual_lines": [ + { + "bbox": [ + 288, + 459, + 499, + 471.75 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 288, + 471.75, + 499, + 484.5 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 288, + 484.5, + 499, + 497.25 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 288, + 497.25, + 499, + 510.0 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 288, + 510.0, + 499, + 522.75 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 288, + 522.75, + 499, + 535.5 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 288, + 535.5, + 499, + 548.25 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 288, + 548.25, + 499, + 561.0 + ], + "spans": [], + "index": 52 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 284, + 565, + 505, + 619 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 284, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 284, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "Table 1: Average validation performance over the last", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 284, + 575, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 284, + 575, + 331, + 588 + ], + "score": 1.0, + "content": "100 rounds:", + "type": "text" + }, + { + "bbox": [ + 332, + 576, + 342, + 587 + ], + "score": 0.82, + "content": "\\%", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 575, + 458, + 588 + ], + "score": 1.0, + "content": "accuracy for rows 1–5; Recall", + "type": "text" + }, + { + "bbox": [ + 459, + 576, + 473, + 586 + ], + "score": 0.64, + "content": "\\textcircled { a } 5", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 576, + 504, + 587 + ], + "score": 0.79, + "content": "( \\times 1 0 0 )", + "type": "inline_equation" + } + ], + "index": 54 + }, + { + "bbox": [ + 283, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 283, + 586, + 416, + 599 + ], + "score": 1.0, + "content": "for Stack Overflow LR; and MSE", + "type": "text" + }, + { + "bbox": [ + 416, + 587, + 452, + 598 + ], + "score": 0.89, + "content": "( \\times 1 0 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "for EMNIST", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 284, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 284, + 597, + 381, + 609 + ], + "score": 1.0, + "content": "AE. Performance within", + "type": "text" + }, + { + "bbox": [ + 381, + 598, + 404, + 608 + ], + "score": 0.88, + "content": "0 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "of the best result for each", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 285, + 609, + 379, + 619 + ], + "spans": [ + { + "bbox": [ + 285, + 609, + 379, + 619 + ], + "score": 1.0, + "content": "task are shown in bold.", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 55 + } + ], + "index": 51.75 + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 503, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "updates to be made when they do occur. This intuition is born out in Figure 1, where adaptive", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "optimizers dramatically outperform non-adaptive ones. For the non-convex NWP task, momentum is", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 650, + 414, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 414, + 662 + ], + "score": 1.0, + "content": "also critical, whereas it slightly hinders performance for the convex LR task.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 59, + "bbox_fs": [ + 105, + 628, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "Dense-gradient tasks CIFAR-10/100, EMNIST AE/CR, and Shakespeare lack a sparse-gradient", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "structure. Shakespeare is relatively easy—most optimizers perform well after enough rounds once suit-", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "ably tuned, though FEDADAGRAD converges faster. For CIFAR-10/100 and EMNIST AE, adaptivity", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "and momentum offer substantial improvements over FEDAVG. Moreover, FEDYOGI and FEDADAM", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "have faster initial convergence than FEDAVGM on these tasks. Notably, FEDADAM and FEDYOGI", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "score": 1.0, + "content": "perform comparably to or better than non-adaptive optimizers throughout, and close to or better than", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 63.5, + "bbox_fs": [ + 105, + 666, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 69, + 485, + 238 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 69, + 485, + 238 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 69, + 485, + 238 + ], + "spans": [ + { + "bbox": [ + 125, + 69, + 485, + 238 + ], + "score": 0.976, + "type": "image", + "image_path": "291f261557be507186a2a9304e87c9225666c4791011ea1091541e22f3f95278.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 69, + 485, + 125.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 125.33333333333334, + 485, + 181.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 181.66666666666669, + 485, + 238.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 256, + 505, + 289 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "score": 1.0, + "content": "Figure 1: Validation accuracy of adaptive and non-adaptive methods, as well as SCAFFOLD, using", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 201, + 279 + ], + "score": 1.0, + "content": "constant learning rates", + "type": "text" + }, + { + "bbox": [ + 202, + 269, + 209, + 279 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 267, + 228, + 279 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 269, + 237, + 279 + ], + "score": 0.78, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 267, + 505, + 279 + ], + "score": 1.0, + "content": "tuned to achieve the best training performance over the last 100", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 278, + 316, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 316, + 290 + ], + "score": 1.0, + "content": "communication rounds; see Appendix D.2 for grids.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 311, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "score": 1.0, + "content": "FEDADAGRAD throughout. As we discuss below, FEDADAM and FEDYOGI actually enable easier", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 323, + 314, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 314, + 334 + ], + "score": 1.0, + "content": "learning rate tuning than FEDAVGM in many tasks.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 339, + 506, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 505, + 351 + ], + "score": 1.0, + "content": "Comparison to SCAFFOLD On all tasks, SCAFFOLD performs comparably to or worse than", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "FEDAVG and our adaptive methods. On Stack Overflow, SCAFFOLD and FEDAVG are nearly", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 359, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 373 + ], + "score": 1.0, + "content": "identical. This is because the number of clients (342,477) makes it unlikely we sample any client", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "more than once. Intuitively, SCAFFOLD does not have a chance to use its client control variates. In", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "score": 1.0, + "content": "other tasks, SCAFFOLD performs worse than other methods. We present two possible explanations:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "score": 1.0, + "content": "First, we only sample a small fraction of clients at each round, so most users are sampled infrequently.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "Intuitively, the client control variates can become stale, and may consequently degrade the perfor-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 415, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 426 + ], + "score": 1.0, + "content": "mance. Second, SCAFFOLD is similar to variance reduction methods such as SVRG (Johnson", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "& Zhang, 2013). While theoretically performant, such methods often perform worse than SGD in", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "practice (Defazio & Bottou, 2018). As shown by Defazio et al. (2014), variance reduction often only", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 447, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 458 + ], + "score": 1.0, + "content": "accelerates convergence when close to a critical point. In cross-device settings (where the number of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 456, + 487, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 487, + 470 + ], + "score": 1.0, + "content": "communication rounds are limited), SCAFFOLD may actually reduce empirical performance.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 107, + 484, + 204, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 206, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 206, + 497 + ], + "score": 1.0, + "content": "5.2 EASE OF TUNING", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 121, + 509, + 486, + 615 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 121, + 509, + 486, + 615 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 121, + 509, + 486, + 615 + ], + "spans": [ + { + "bbox": [ + 121, + 509, + 486, + 615 + ], + "score": 0.961, + "type": "image", + "image_path": "a1ecef7595988e2f35b8b5c2ee3db0d5eb2f0bad2ba1b6103616e6c3ce2fd289.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 121, + 509, + 486, + 544.3333333333334 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 121, + 544.3333333333334, + 486, + 579.6666666666667 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 121, + 579.6666666666667, + 486, + 615.0000000000001 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 630, + 505, + 663 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "Figure 2: Validation accuracy (averaged over the last 100 rounds) of FEDADAM, FEDYOGI, and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "FEDAVGM for various client/server learning rates combination on the SO NWP task. For FEDADAM", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 651, + 241, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 196, + 663 + ], + "score": 1.0, + "content": "and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 196, + 651, + 236, + 662 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 651, + 241, + 663 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 300, + 690 + ], + "score": 1.0, + "content": "Obtaining optimal performance involves tuning", + "type": "text" + }, + { + "bbox": [ + 301, + 680, + 320, + 689 + ], + "score": 0.86, + "content": "\\eta _ { l } , \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 677, + 446, + 690 + ], + "score": 1.0, + "content": ", and for the adaptive methods,", + "type": "text" + }, + { + "bbox": [ + 446, + 680, + 452, + 688 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". To quantify", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 687, + 503, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 469, + 701 + ], + "score": 1.0, + "content": "how easy it is to tune various methods, we plot their validation performance as a function of", + "type": "text" + }, + { + "bbox": [ + 470, + 690, + 479, + 700 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 687, + 496, + 701 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 497, + 690, + 503, + 700 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "Figure 2 gives results for FEDADAM, FEDYOGI, and FEDAVGM on Stack Overflow NWP. Plots for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "all other optimizers and tasks are in Appendix E.3. For FEDAVGM, there are only a few good values", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 117, + 733 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 722, + 126, + 732 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 720, + 162, + 733 + ], + "score": 1.0, + "content": "for each", + "type": "text" + }, + { + "bbox": [ + 162, + 722, + 168, + 732 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 720, + 439, + 733 + ], + "score": 1.0, + "content": ", while for FEDADAM and FEDYOGI, there are many good values of", + "type": "text" + }, + { + "bbox": [ + 439, + 722, + 448, + 732 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "for a range of", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 69, + 485, + 238 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 69, + 485, + 238 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 69, + 485, + 238 + ], + "spans": [ + { + "bbox": [ + 125, + 69, + 485, + 238 + ], + "score": 0.976, + "type": "image", + "image_path": "291f261557be507186a2a9304e87c9225666c4791011ea1091541e22f3f95278.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 69, + 485, + 125.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 125.33333333333334, + 485, + 181.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 181.66666666666669, + 485, + 238.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 256, + 505, + 289 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 270 + ], + "score": 1.0, + "content": "Figure 1: Validation accuracy of adaptive and non-adaptive methods, as well as SCAFFOLD, using", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 201, + 279 + ], + "score": 1.0, + "content": "constant learning rates", + "type": "text" + }, + { + "bbox": [ + 202, + 269, + 209, + 279 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 267, + 228, + 279 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 269, + 237, + 279 + ], + "score": 0.78, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 267, + 505, + 279 + ], + "score": 1.0, + "content": "tuned to achieve the best training performance over the last 100", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 278, + 316, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 316, + 290 + ], + "score": 1.0, + "content": "communication rounds; see Appendix D.2 for grids.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 311, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "score": 1.0, + "content": "FEDADAGRAD throughout. As we discuss below, FEDADAM and FEDYOGI actually enable easier", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 323, + 314, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 314, + 334 + ], + "score": 1.0, + "content": "learning rate tuning than FEDAVGM in many tasks.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 310, + 505, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 339, + 506, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 505, + 351 + ], + "score": 1.0, + "content": "Comparison to SCAFFOLD On all tasks, SCAFFOLD performs comparably to or worse than", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "FEDAVG and our adaptive methods. On Stack Overflow, SCAFFOLD and FEDAVG are nearly", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 359, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 373 + ], + "score": 1.0, + "content": "identical. This is because the number of clients (342,477) makes it unlikely we sample any client", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "more than once. Intuitively, SCAFFOLD does not have a chance to use its client control variates. In", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "score": 1.0, + "content": "other tasks, SCAFFOLD performs worse than other methods. We present two possible explanations:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "score": 1.0, + "content": "First, we only sample a small fraction of clients at each round, so most users are sampled infrequently.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "score": 1.0, + "content": "Intuitively, the client control variates can become stale, and may consequently degrade the perfor-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 415, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 426 + ], + "score": 1.0, + "content": "mance. Second, SCAFFOLD is similar to variance reduction methods such as SVRG (Johnson", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "& Zhang, 2013). While theoretically performant, such methods often perform worse than SGD in", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 104, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "practice (Defazio & Bottou, 2018). As shown by Defazio et al. (2014), variance reduction often only", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 447, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 458 + ], + "score": 1.0, + "content": "accelerates convergence when close to a critical point. In cross-device settings (where the number of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 456, + 487, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 487, + 470 + ], + "score": 1.0, + "content": "communication rounds are limited), SCAFFOLD may actually reduce empirical performance.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 339, + 506, + 470 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 484, + 204, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 206, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 206, + 497 + ], + "score": 1.0, + "content": "5.2 EASE OF TUNING", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 121, + 509, + 486, + 615 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 121, + 509, + 486, + 615 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 121, + 509, + 486, + 615 + ], + "spans": [ + { + "bbox": [ + 121, + 509, + 486, + 615 + ], + "score": 0.961, + "type": "image", + "image_path": "a1ecef7595988e2f35b8b5c2ee3db0d5eb2f0bad2ba1b6103616e6c3ce2fd289.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 121, + 509, + 486, + 544.3333333333334 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 121, + 544.3333333333334, + 486, + 579.6666666666667 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 121, + 579.6666666666667, + 486, + 615.0000000000001 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 630, + 505, + 663 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 642 + ], + "score": 1.0, + "content": "Figure 2: Validation accuracy (averaged over the last 100 rounds) of FEDADAM, FEDYOGI, and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "FEDAVGM for various client/server learning rates combination on the SO NWP task. For FEDADAM", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 651, + 241, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 196, + 663 + ], + "score": 1.0, + "content": "and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 196, + 651, + 236, + 662 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 651, + 241, + 663 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 300, + 690 + ], + "score": 1.0, + "content": "Obtaining optimal performance involves tuning", + "type": "text" + }, + { + "bbox": [ + 301, + 680, + 320, + 689 + ], + "score": 0.86, + "content": "\\eta _ { l } , \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 677, + 446, + 690 + ], + "score": 1.0, + "content": ", and for the adaptive methods,", + "type": "text" + }, + { + "bbox": [ + 446, + 680, + 452, + 688 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". To quantify", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 687, + 503, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 469, + 701 + ], + "score": 1.0, + "content": "how easy it is to tune various methods, we plot their validation performance as a function of", + "type": "text" + }, + { + "bbox": [ + 470, + 690, + 479, + 700 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 687, + 496, + 701 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 497, + 690, + 503, + 700 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "Figure 2 gives results for FEDADAM, FEDYOGI, and FEDAVGM on Stack Overflow NWP. Plots for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "all other optimizers and tasks are in Appendix E.3. For FEDAVGM, there are only a few good values", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 117, + 733 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 722, + 126, + 732 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 720, + 162, + 733 + ], + "score": 1.0, + "content": "for each", + "type": "text" + }, + { + "bbox": [ + 162, + 722, + 168, + 732 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 720, + 439, + 733 + ], + "score": 1.0, + "content": ", while for FEDADAM and FEDYOGI, there are many good values of", + "type": "text" + }, + { + "bbox": [ + 439, + 722, + 448, + 732 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "for a range of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 113, + 338 + ], + "score": 0.74, + "content": "\\eta", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 114, + 325, + 506, + 339 + ], + "score": 1.0, + "content": ". Thus, FEDADAM and FEDYOGI are arguably easier to tune in this setting. Similar results hold for", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 337, + 285, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 285, + 349 + ], + "score": 1.0, + "content": "other tasks and optimizers (Figures 5 to 11).", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 677, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 83, + 485, + 248 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 83, + 485, + 248 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 83, + 485, + 248 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 485, + 248 + ], + "score": 0.971, + "type": "image", + "image_path": "cd2738f345eafd8cd9372344ebd4c03605653f82aa8c8583ca069896d16d9d6c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 83, + 485, + 138.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 138.0, + 485, + 193.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 193.0, + 485, + 248.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 266, + 505, + 299 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 265, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 484, + 279 + ], + "score": 1.0, + "content": "Figure 3: Validation performance of FEDADAGRAD, FEDADAM, and FEDYOGI for varying", + "type": "text" + }, + { + "bbox": [ + 484, + 268, + 491, + 276 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 265, + 505, + 279 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 233, + 289 + ], + "score": 1.0, + "content": "various tasks. The learning rates", + "type": "text" + }, + { + "bbox": [ + 234, + 279, + 240, + 289 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 277, + 257, + 289 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 279, + 266, + 289 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 277, + 339, + 289 + ], + "score": 1.0, + "content": "are tuned for each", + "type": "text" + }, + { + "bbox": [ + 339, + 279, + 346, + 287 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "to achieve the best training performance", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 288, + 264, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 264, + 299 + ], + "score": 1.0, + "content": "on the last 100 communication rounds.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 326, + 504, + 348 + ], + "lines": [ + { + "bbox": [ + 106, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 113, + 338 + ], + "score": 0.74, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 325, + 506, + 339 + ], + "score": 1.0, + "content": ". Thus, FEDADAM and FEDYOGI are arguably easier to tune in this setting. Similar results hold for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 337, + 285, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 285, + 349 + ], + "score": 1.0, + "content": "other tasks and optimizers (Figures 5 to 11).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 353, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 383, + 366 + ], + "score": 1.0, + "content": "This leads to a natural question: Is the reduction in the need to tune", + "type": "text" + }, + { + "bbox": [ + 383, + 356, + 393, + 365 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 353, + 411, + 366 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 411, + 356, + 418, + 365 + ], + "score": 0.81, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "offset by the need to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 183, + 376 + ], + "score": 1.0, + "content": "tune the adaptivity", + "type": "text" + }, + { + "bbox": [ + 184, + 365, + 195, + 375 + ], + "score": 0.79, + "content": "\\tau ?", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 364, + 286, + 376 + ], + "score": 1.0, + "content": "In fact, while we tune", + "type": "text" + }, + { + "bbox": [ + 286, + 366, + 293, + 375 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 364, + 481, + 376 + ], + "score": 1.0, + "content": "in Figure 1, our results are relatively robust to", + "type": "text" + }, + { + "bbox": [ + 482, + 366, + 488, + 374 + ], + "score": 0.69, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 364, + 505, + 376 + ], + "score": 1.0, + "content": ". To", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 376, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 368, + 387 + ], + "score": 1.0, + "content": "demonstrate, we plot the best validation performance for various", + "type": "text" + }, + { + "bbox": [ + 369, + 377, + 376, + 385 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 376, + 505, + 387 + ], + "score": 1.0, + "content": "in Figure 3. For nearly all tasks", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 172, + 398 + ], + "score": 1.0, + "content": "and optimizers,", + "type": "text" + }, + { + "bbox": [ + 172, + 385, + 213, + 396 + ], + "score": 0.92, + "content": "\\dot { \\tau } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "works almost as well all other values. This aligns with work by Zaheer", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 295, + 409 + ], + "score": 1.0, + "content": "et al. (2018), who show that moderately large", + "type": "text" + }, + { + "bbox": [ + 295, + 399, + 302, + 407 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "yield better performance for centralized adaptive", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 408, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 451, + 419 + ], + "score": 1.0, + "content": "optimizers. FEDADAM and FEDYOGI see only small differences in performance among", + "type": "text" + }, + { + "bbox": [ + 451, + 409, + 458, + 417 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 408, + 504, + 419 + ], + "score": 1.0, + "content": "on all tasks", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 418, + 484, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 471, + 430 + ], + "score": 1.0, + "content": "except Stack Overflow LR (for which FEDADAGRAD is the best optimizer, and is robust to", + "type": "text" + }, + { + "bbox": [ + 471, + 420, + 478, + 428 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 418, + 484, + 430 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 447, + 207, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 208, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 208, + 460 + ], + "score": 1.0, + "content": "5.3 OTHER FINDINGS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 469, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 507, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 507, + 482 + ], + "score": 1.0, + "content": "We present additional empirical analyses in Appendix E. These include EMNIST CR results (Ap-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 480, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 492 + ], + "score": 1.0, + "content": "pendix E.1), Stack Overflow results on the full test dataset (Appendix E.2), client/server learning rate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 491, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 480, + 503 + ], + "score": 1.0, + "content": "heat maps for all optimizers and tasks (Appendix E.3), an analysis of the relationship between", + "type": "text" + }, + { + "bbox": [ + 480, + 493, + 487, + 502 + ], + "score": 0.77, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 491, + 505, + 503 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 502, + 417, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 115, + 513 + ], + "score": 0.8, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 502, + 417, + 514 + ], + "score": 1.0, + "content": "(Appendix E.4), and experiments with learning rate decay (Appendix E.6).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 533, + 195, + 545 + ], + "lines": [ + { + "bbox": [ + 104, + 531, + 197, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 197, + 549 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "score": 1.0, + "content": "In this paper, we demonstrated that adaptive optimizers can be powerful tools in improving the conver-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 570, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 506, + 585 + ], + "score": 1.0, + "content": "gence of FL. By using a simple client/server optimizer framework, we can incorporate adaptivity into", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "FL in a principled, intuitive, and theoretically-justified manner. We also developed comprehensive", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "benchmarks for comparing federated optimization algorithms. To encourage reproducibility and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 603, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 506, + 616 + ], + "score": 1.0, + "content": "breadth of comparison, we have attempted to describe our experiments as rigorously as possible,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "and have created an open-source framework with all models, datasets, and code. We believe our", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "work raises many important questions about how best to perform federated optimization. Example", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "directions for future research include understanding how the use of adaptivity affects differential", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 646, + 191, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 191, + 658 + ], + "score": 1.0, + "content": "privacy and fairness.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 108, + 678, + 175, + 690 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 176, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 176, + 691 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 274, + 710 + ], + "score": 1.0, + "content": "The TensorFlow Federated Authors.", + "type": "text" + }, + { + "bbox": [ + 290, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "TensorFlow Federated Stack Overflow dataset,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "2019. URL https://www.tensorflow.org/federated/api_docs/python/tff/", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 720, + 378, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 378, + 732 + ], + "score": 1.0, + "content": "simulation/datasets/stackoverflow/load_data.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 83, + 485, + 248 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 83, + 485, + 248 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 83, + 485, + 248 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 485, + 248 + ], + "score": 0.971, + "type": "image", + "image_path": "cd2738f345eafd8cd9372344ebd4c03605653f82aa8c8583ca069896d16d9d6c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 83, + 485, + 138.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 138.0, + 485, + 193.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 193.0, + 485, + 248.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 266, + 505, + 299 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 265, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 484, + 279 + ], + "score": 1.0, + "content": "Figure 3: Validation performance of FEDADAGRAD, FEDADAM, and FEDYOGI for varying", + "type": "text" + }, + { + "bbox": [ + 484, + 268, + 491, + 276 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 265, + 505, + 279 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 233, + 289 + ], + "score": 1.0, + "content": "various tasks. The learning rates", + "type": "text" + }, + { + "bbox": [ + 234, + 279, + 240, + 289 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 277, + 257, + 289 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 279, + 266, + 289 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 277, + 339, + 289 + ], + "score": 1.0, + "content": "are tuned for each", + "type": "text" + }, + { + "bbox": [ + 339, + 279, + 346, + 287 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "to achieve the best training performance", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 288, + 264, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 264, + 299 + ], + "score": 1.0, + "content": "on the last 100 communication rounds.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 326, + 504, + 348 + ], + "lines": [], + "index": 6.5, + "bbox_fs": [ + 106, + 325, + 506, + 349 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 353, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 383, + 366 + ], + "score": 1.0, + "content": "This leads to a natural question: Is the reduction in the need to tune", + "type": "text" + }, + { + "bbox": [ + 383, + 356, + 393, + 365 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 353, + 411, + 366 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 411, + 356, + 418, + 365 + ], + "score": 0.81, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "offset by the need to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 183, + 376 + ], + "score": 1.0, + "content": "tune the adaptivity", + "type": "text" + }, + { + "bbox": [ + 184, + 365, + 195, + 375 + ], + "score": 0.79, + "content": "\\tau ?", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 364, + 286, + 376 + ], + "score": 1.0, + "content": "In fact, while we tune", + "type": "text" + }, + { + "bbox": [ + 286, + 366, + 293, + 375 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 364, + 481, + 376 + ], + "score": 1.0, + "content": "in Figure 1, our results are relatively robust to", + "type": "text" + }, + { + "bbox": [ + 482, + 366, + 488, + 374 + ], + "score": 0.69, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 364, + 505, + 376 + ], + "score": 1.0, + "content": ". To", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 376, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 368, + 387 + ], + "score": 1.0, + "content": "demonstrate, we plot the best validation performance for various", + "type": "text" + }, + { + "bbox": [ + 369, + 377, + 376, + 385 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 376, + 505, + 387 + ], + "score": 1.0, + "content": "in Figure 3. For nearly all tasks", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 172, + 398 + ], + "score": 1.0, + "content": "and optimizers,", + "type": "text" + }, + { + "bbox": [ + 172, + 385, + 213, + 396 + ], + "score": 0.92, + "content": "\\dot { \\tau } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "works almost as well all other values. This aligns with work by Zaheer", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 295, + 409 + ], + "score": 1.0, + "content": "et al. (2018), who show that moderately large", + "type": "text" + }, + { + "bbox": [ + 295, + 399, + 302, + 407 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "yield better performance for centralized adaptive", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 408, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 451, + 419 + ], + "score": 1.0, + "content": "optimizers. FEDADAM and FEDYOGI see only small differences in performance among", + "type": "text" + }, + { + "bbox": [ + 451, + 409, + 458, + 417 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 408, + 504, + 419 + ], + "score": 1.0, + "content": "on all tasks", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 418, + 484, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 471, + 430 + ], + "score": 1.0, + "content": "except Stack Overflow LR (for which FEDADAGRAD is the best optimizer, and is robust to", + "type": "text" + }, + { + "bbox": [ + 471, + 420, + 478, + 428 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 418, + 484, + 430 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 353, + 506, + 430 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 447, + 207, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 208, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 208, + 460 + ], + "score": 1.0, + "content": "5.3 OTHER FINDINGS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 469, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 507, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 507, + 482 + ], + "score": 1.0, + "content": "We present additional empirical analyses in Appendix E. These include EMNIST CR results (Ap-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 480, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 492 + ], + "score": 1.0, + "content": "pendix E.1), Stack Overflow results on the full test dataset (Appendix E.2), client/server learning rate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 491, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 480, + 503 + ], + "score": 1.0, + "content": "heat maps for all optimizers and tasks (Appendix E.3), an analysis of the relationship between", + "type": "text" + }, + { + "bbox": [ + 480, + 493, + 487, + 502 + ], + "score": 0.77, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 491, + 505, + 503 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 502, + 417, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 115, + 513 + ], + "score": 0.8, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 502, + 417, + 514 + ], + "score": 1.0, + "content": "(Appendix E.4), and experiments with learning rate decay (Appendix E.6).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 468, + 507, + 514 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 533, + 195, + 545 + ], + "lines": [ + { + "bbox": [ + 104, + 531, + 197, + 549 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 197, + 549 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 573 + ], + "score": 1.0, + "content": "In this paper, we demonstrated that adaptive optimizers can be powerful tools in improving the conver-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 570, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 506, + 585 + ], + "score": 1.0, + "content": "gence of FL. By using a simple client/server optimizer framework, we can incorporate adaptivity into", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "FL in a principled, intuitive, and theoretically-justified manner. We also developed comprehensive", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "benchmarks for comparing federated optimization algorithms. To encourage reproducibility and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 603, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 506, + 616 + ], + "score": 1.0, + "content": "breadth of comparison, we have attempted to describe our experiments as rigorously as possible,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "and have created an open-source framework with all models, datasets, and code. We believe our", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "work raises many important questions about how best to perform federated optimization. Example", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "directions for future research include understanding how the use of adaptivity affects differential", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 646, + 191, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 191, + 658 + ], + "score": 1.0, + "content": "privacy and fairness.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 560, + 506, + 658 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 678, + 175, + 690 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 176, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 176, + 691 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 274, + 710 + ], + "score": 1.0, + "content": "The TensorFlow Federated Authors.", + "type": "text" + }, + { + "bbox": [ + 290, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "TensorFlow Federated Stack Overflow dataset,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "2019. URL https://www.tensorflow.org/federated/api_docs/python/tff/", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 720, + 378, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 378, + 732 + ], + "score": 1.0, + "content": "simulation/datasets/stackoverflow/load_data.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Debraj Basu, Deepesh Data, Can Karakus, and Suhas Diggavi. Qsparse-local-SGD: Distributed", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "SGD with quantization, sparsification and local computations. In Advances in Neural Information", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 299, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 299, + 117 + ], + "score": 1.0, + "content": "Processing Systems, pp. 14668–14679, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 124, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex Ingerman, Vladimir", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "Ivanov, Chloé Kiddon, Jakub Konecný, Stefano Mazzocchi, Brendan McMahan, Timon ˇ", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 506, + 157 + ], + "score": 1.0, + "content": "Van Overveldt, David Petrou, Daniel Ramage, and Jason Roselander. Towards feder-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 506, + 167 + ], + "score": 1.0, + "content": "ated learning at scale: System design. In A. Talwalkar, V. Smith, and M. Zaharia", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 115, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "(eds.), Proceedings of Machine Learning and Systems, volume 1, pp. 374–388. Proceed-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 177, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 177, + 506, + 191 + ], + "score": 1.0, + "content": "ings of MLSys, 2019. URL https://proceedings.mlsys.org/paper/2019/file/", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 187, + 373, + 201 + ], + "spans": [ + { + "bbox": [ + 115, + 187, + 373, + 201 + ], + "score": 1.0, + "content": "bd686fd640be98efaae0091fa301e613-Paper.pdf.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 506, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "Sebastian Caldas, Peter Wu, Tian Li, Jakub Konecný, H Brendan McMahan, Virginia Smith, and ˇ", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 115, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "Ameet Talwalkar. LEAF: A benchmark for federated settings. arXiv preprint arXiv:1812.01097,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 229, + 142, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 229, + 142, + 241 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 249, + 504, + 282 + ], + "lines": [ + { + "bbox": [ + 107, + 249, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 107, + 249, + 506, + 261 + ], + "score": 1.0, + "content": "Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. EMNIST: Extending MNIST", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 258, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 115, + 258, + 506, + 273 + ], + "score": 1.0, + "content": "to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pp.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 271, + 218, + 282 + ], + "spans": [ + { + "bbox": [ + 116, + 271, + 218, + 282 + ], + "score": 1.0, + "content": "2921–2926. IEEE, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 105, + 290, + 504, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 304 + ], + "score": 1.0, + "content": "Aaron Defazio and Léon Bottou. On the ineffectiveness of variance reduced optimization for deep", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 301, + 317, + 313 + ], + "spans": [ + { + "bbox": [ + 115, + 301, + 317, + 313 + ], + "score": 1.0, + "content": "learning. arXiv preprint arXiv:1812.04529, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 320, + 504, + 343 + ], + "lines": [ + { + "bbox": [ + 107, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 107, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "Aaron Defazio, Francis Bach, and Simon Lacoste-Julien. SAGA: A fast incremental gradient method", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 331, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 116, + 331, + 482, + 344 + ], + "score": 1.0, + "content": "with support for non-strongly convex composite objectives. In NIPS, pp. 1646–1654, 2014.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 105, + 351, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 362, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 482, + 374 + ], + "score": 1.0, + "content": "stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 381, + 504, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 403, + 362, + 415 + ], + "spans": [ + { + "bbox": [ + 116, + 403, + 362, + 415 + ], + "score": 1.0, + "content": "ImageNet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 423, + 504, + 445 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip B Gibbons. The non-IID data quagmire of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 433, + 409, + 446 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 409, + 446 + ], + "score": 1.0, + "content": "decentralized machine learning. arXiv preprint arXiv:1910.00189, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 503, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 464, + 465, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 464, + 465, + 477 + ], + "score": 1.0, + "content": "distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 506, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 507, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 257, + 497 + ], + "score": 1.0, + "content": "Alex Ingerman and Krzys", + "type": "text" + }, + { + "bbox": [ + 264, + 483, + 315, + 495 + ], + "score": 1.0, + "content": "Ostrowski.", + "type": "text" + }, + { + "bbox": [ + 357, + 482, + 507, + 496 + ], + "score": 1.0, + "content": "Introducing TensorFlow Fed-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 493, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 147, + 507 + ], + "score": 1.0, + "content": "erated,", + "type": "text" + }, + { + "bbox": [ + 169, + 493, + 198, + 507 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + }, + { + "bbox": [ + 274, + 493, + 301, + 507 + ], + "score": 1.0, + "content": "URL", + "type": "text" + }, + { + "bbox": [ + 323, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "https://medium.com/tensorflow/", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 505, + 392, + 517 + ], + "spans": [ + { + "bbox": [ + 116, + 505, + 392, + 517 + ], + "score": 1.0, + "content": "introducing-tensorflow-federated-a4147aa20041.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 504, + 548 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive variance", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 535, + 465, + 549 + ], + "spans": [ + { + "bbox": [ + 115, + 535, + 465, + 549 + ], + "score": 1.0, + "content": "reduction. In Advances in Neural Information Processing Systems, pp. 315–323, 2013.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 108, + 555, + 504, + 589 + ], + "lines": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "Bhagoji, Keith Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 577, + 446, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 446, + 590 + ], + "score": 1.0, + "content": "and open problems in federated learning. arXiv preprint arXiv:1912.04977, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 116, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for on-device federated", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 619, + 317, + 631 + ], + "spans": [ + { + "bbox": [ + 116, + 619, + 317, + 631 + ], + "score": 1.0, + "content": "learning. arXiv preprint arXiv:1910.06378, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 104, + 637, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 507, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 507, + 652 + ], + "score": 1.0, + "content": "Ahmed Khaled, Konstantin Mishchenko, and Peter Richtárik. First analysis of local GD on heteroge-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 649, + 327, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 649, + 327, + 661 + ], + "score": 1.0, + "content": "neous data. arXiv preprint arXiv:1909.04715, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 679, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 679, + 506, + 693 + ], + "score": 1.0, + "content": "Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 117, + 690, + 268, + 702 + ], + "spans": [ + { + "bbox": [ + 117, + 690, + 268, + 702 + ], + "score": 1.0, + "content": "Conference Track Proceedings, 2015.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 108, + 710, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 248, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 248, + 732 + ], + "score": 1.0, + "content": "Technical report, Citeseer, 2009.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Debraj Basu, Deepesh Data, Can Karakus, and Suhas Diggavi. Qsparse-local-SGD: Distributed", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "SGD with quantization, sparsification and local computations. In Advances in Neural Information", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 299, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 299, + 117 + ], + "score": 1.0, + "content": "Processing Systems, pp. 14668–14679, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 506, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 124, + 505, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex Ingerman, Vladimir", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "Ivanov, Chloé Kiddon, Jakub Konecný, Stefano Mazzocchi, Brendan McMahan, Timon ˇ", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 506, + 157 + ], + "score": 1.0, + "content": "Van Overveldt, David Petrou, Daniel Ramage, and Jason Roselander. Towards feder-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 506, + 167 + ], + "score": 1.0, + "content": "ated learning at scale: System design. In A. Talwalkar, V. Smith, and M. Zaharia", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 115, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "(eds.), Proceedings of Machine Learning and Systems, volume 1, pp. 374–388. Proceed-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 177, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 177, + 506, + 191 + ], + "score": 1.0, + "content": "ings of MLSys, 2019. URL https://proceedings.mlsys.org/paper/2019/file/", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 187, + 373, + 201 + ], + "spans": [ + { + "bbox": [ + 115, + 187, + 373, + 201 + ], + "score": 1.0, + "content": "bd686fd640be98efaae0091fa301e613-Paper.pdf.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 122, + 506, + 201 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 506, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "Sebastian Caldas, Peter Wu, Tian Li, Jakub Konecný, H Brendan McMahan, Virginia Smith, and ˇ", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 115, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "Ameet Talwalkar. LEAF: A benchmark for federated settings. arXiv preprint arXiv:1812.01097,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 229, + 142, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 229, + 142, + 241 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 207, + 506, + 241 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 249, + 504, + 282 + ], + "lines": [ + { + "bbox": [ + 107, + 249, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 107, + 249, + 506, + 261 + ], + "score": 1.0, + "content": "Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. EMNIST: Extending MNIST", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 258, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 115, + 258, + 506, + 273 + ], + "score": 1.0, + "content": "to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pp.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 271, + 218, + 282 + ], + "spans": [ + { + "bbox": [ + 116, + 271, + 218, + 282 + ], + "score": 1.0, + "content": "2921–2926. IEEE, 2017.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 107, + 249, + 506, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 290, + 504, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 304 + ], + "score": 1.0, + "content": "Aaron Defazio and Léon Bottou. On the ineffectiveness of variance reduced optimization for deep", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 301, + 317, + 313 + ], + "spans": [ + { + "bbox": [ + 115, + 301, + 317, + 313 + ], + "score": 1.0, + "content": "learning. arXiv preprint arXiv:1812.04529, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 288, + 505, + 313 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 320, + 504, + 343 + ], + "lines": [ + { + "bbox": [ + 107, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 107, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "Aaron Defazio, Francis Bach, and Simon Lacoste-Julien. SAGA: A fast incremental gradient method", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 331, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 116, + 331, + 482, + 344 + ], + "score": 1.0, + "content": "with support for non-strongly convex composite objectives. In NIPS, pp. 1646–1654, 2014.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 107, + 320, + 505, + 344 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 351, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 505, + 363 + ], + "score": 1.0, + "content": "John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 362, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 482, + 374 + ], + "score": 1.0, + "content": "stochastic optimization. Journal of Machine Learning Research, 12(Jul):2121–2159, 2011.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 351, + 505, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 381, + 504, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 391, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 391, + 506, + 406 + ], + "score": 1.0, + "content": "Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 403, + 362, + 415 + ], + "spans": [ + { + "bbox": [ + 116, + 403, + 362, + 415 + ], + "score": 1.0, + "content": "ImageNet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 380, + 506, + 415 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 423, + 504, + 445 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip B Gibbons. The non-IID data quagmire of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 433, + 409, + 446 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 409, + 446 + ], + "score": 1.0, + "content": "decentralized machine learning. arXiv preprint arXiv:1910.00189, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 423, + 505, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 503, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 464, + 465, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 464, + 465, + 477 + ], + "score": 1.0, + "content": "distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 453, + 505, + 477 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 506, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 507, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 257, + 497 + ], + "score": 1.0, + "content": "Alex Ingerman and Krzys", + "type": "text" + }, + { + "bbox": [ + 264, + 483, + 315, + 495 + ], + "score": 1.0, + "content": "Ostrowski.", + "type": "text" + }, + { + "bbox": [ + 357, + 482, + 507, + 496 + ], + "score": 1.0, + "content": "Introducing TensorFlow Fed-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 493, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 147, + 507 + ], + "score": 1.0, + "content": "erated,", + "type": "text" + }, + { + "bbox": [ + 169, + 493, + 198, + 507 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + }, + { + "bbox": [ + 274, + 493, + 301, + 507 + ], + "score": 1.0, + "content": "URL", + "type": "text" + }, + { + "bbox": [ + 323, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "https://medium.com/tensorflow/", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 505, + 392, + 517 + ], + "spans": [ + { + "bbox": [ + 116, + 505, + 392, + 517 + ], + "score": 1.0, + "content": "introducing-tensorflow-federated-a4147aa20041.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 482, + 507, + 517 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 504, + 548 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive variance", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 535, + 465, + 549 + ], + "spans": [ + { + "bbox": [ + 115, + 535, + 465, + 549 + ], + "score": 1.0, + "content": "reduction. In Advances in Neural Information Processing Systems, pp. 315–323, 2013.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 525, + 505, + 549 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 555, + 504, + 589 + ], + "lines": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "Bhagoji, Keith Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 577, + 446, + 590 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 446, + 590 + ], + "score": 1.0, + "content": "and open problems in federated learning. arXiv preprint arXiv:1912.04977, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 554, + 505, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 116, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "Ananda Theertha Suresh. SCAFFOLD: Stochastic controlled averaging for on-device federated", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 619, + 317, + 631 + ], + "spans": [ + { + "bbox": [ + 116, + 619, + 317, + 631 + ], + "score": 1.0, + "content": "learning. arXiv preprint arXiv:1910.06378, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 597, + 505, + 631 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 637, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 507, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 507, + 652 + ], + "score": 1.0, + "content": "Ahmed Khaled, Konstantin Mishchenko, and Peter Richtárik. First analysis of local GD on heteroge-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 649, + 327, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 649, + 327, + 661 + ], + "score": 1.0, + "content": "neous data. arXiv preprint arXiv:1909.04715, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 636, + 507, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 679, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 679, + 506, + 693 + ], + "score": 1.0, + "content": "Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 117, + 690, + 268, + 702 + ], + "spans": [ + { + "bbox": [ + 117, + 690, + 268, + 702 + ], + "score": 1.0, + "content": "Conference Track Proceedings, 2015.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 106, + 668, + 506, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 710, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 720, + 248, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 248, + 732 + ], + "score": 1.0, + "content": "Technical report, Citeseer, 2009.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 708, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 92, + 486, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 92, + 486, + 106 + ], + "score": 1.0, + "content": "Federated optimization in heterogeneous networks. arXiv preprint arXiv:1812.06127, 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 505, + 134 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 507, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 507, + 126 + ], + "score": 1.0, + "content": "Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 122, + 410, + 134 + ], + "spans": [ + { + "bbox": [ + 116, + 122, + 410, + 134 + ], + "score": 1.0, + "content": "methods, and future directions. arXiv preprint arXiv:1908.07873, 2019a.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 140, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 139, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 505, + 153 + ], + "score": 1.0, + "content": "Wei Li and Andrew McCallum. Pachinko allocation: DAG-structured mixture models of topic", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 150, + 507, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 150, + 507, + 163 + ], + "score": 1.0, + "content": "correlations. In Proceedings of the 23rd international conference on Machine learning, pp. 577–", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 161, + 162, + 173 + ], + "spans": [ + { + "bbox": [ + 115, + 161, + 162, + 173 + ], + "score": 1.0, + "content": "584, 2006.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 505, + 202 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "score": 1.0, + "content": "Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 190, + 387, + 203 + ], + "spans": [ + { + "bbox": [ + 116, + 190, + 387, + 203 + ], + "score": 1.0, + "content": "FedAvg on non-IID data. arXiv preprint arXiv:1907.02189, 2019b.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "Xiaoyu Li and Francesco Orabona. On the convergence of stochastic gradient descent with adaptive", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 220, + 320, + 231 + ], + "spans": [ + { + "bbox": [ + 115, + 220, + 320, + 231 + ], + "score": 1.0, + "content": "stepsizes. arXiv preprint arXiv:1805.08114, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 506, + 282 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "score": 1.0, + "content": "Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 248, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 248, + 506, + 261 + ], + "score": 1.0, + "content": "bound of learning rate. In 7th International Conference on Learning Representations, ICLR 2019,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 258, + 508, + 272 + ], + "spans": [ + { + "bbox": [ + 114, + 258, + 508, + 272 + ], + "score": 1.0, + "content": "New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. URL https://openreview.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 270, + 259, + 282 + ], + "spans": [ + { + "bbox": [ + 116, + 270, + 259, + 282 + ], + "score": 1.0, + "content": "net/forum?id=Bkg3g2R9FX.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 506, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 507, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 507, + 300 + ], + "score": 1.0, + "content": "Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 117, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "Communication-efficient learning of deep networks from decentralized data. In Proceedings of the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 309, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 116, + 309, + 506, + 321 + ], + "score": 1.0, + "content": "20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 318, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 114, + 318, + 507, + 333 + ], + "score": 1.0, + "content": "2017, Fort Lauderdale, FL, USA, pp. 1273–1282, 2017. URL http://proceedings.mlr.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 331, + 271, + 343 + ], + "spans": [ + { + "bbox": [ + 115, + 331, + 271, + 343 + ], + "score": 1.0, + "content": "press/v54/mcmahan17a.html.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 105, + 349, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 362 + ], + "score": 1.0, + "content": "H. Brendan McMahan and Matthew Streeter. Adaptive bound optimization for online convex", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 360, + 242, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 360, + 242, + 372 + ], + "score": 1.0, + "content": "optimization. In COLT, 2010a.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 377, + 504, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 377, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 391 + ], + "score": 1.0, + "content": "H. Brendan McMahan and Matthew J. Streeter. Adaptive bound optimization for online convex", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 389, + 412, + 401 + ], + "spans": [ + { + "bbox": [ + 116, + 389, + 412, + 401 + ], + "score": 1.0, + "content": "optimization. In COLT The 23rd Conference on Learning Theory, 2010b.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 504, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "Sashank J Reddi, Ahmed Hefny, Suvrit Sra, Barnabás Póczós, and Alex Smola. Stochastic variance", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 417, + 375, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 375, + 430 + ], + "score": 1.0, + "content": "reduction for nonconvex optimization. arXiv:1603.06160, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 504, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of ADAM and beyond. arXiv", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 447, + 254, + 459 + ], + "spans": [ + { + "bbox": [ + 115, + 447, + 254, + 459 + ], + "score": 1.0, + "content": "preprint arXiv:1904.09237, 2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 506, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 507, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 507, + 477 + ], + "score": 1.0, + "content": "Sebastian U. Stich. Local SGD converges fast and communicates little. In International Confer-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 506, + 488 + ], + "score": 1.0, + "content": "ence on Learning Representations, 2019. URL https://openreview.net/forum?id=", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 487, + 181, + 498 + ], + "spans": [ + { + "bbox": [ + 116, + 487, + 181, + 498 + ], + "score": 1.0, + "content": "S1g2JnRcFX.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 506, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "Sebastian U Stich and Sai Praneeth Karimireddy. The error-feedback framework: Better rates for", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 115, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "SGD with delayed gradients and compressed communication. arXiv preprint arXiv:1909.05350,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 524, + 143, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 524, + 143, + 538 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 105, + 543, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 544, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 506, + 557 + ], + "score": 1.0, + "content": "Jianyu Wang and Gauri Joshi. Cooperative SGD: A unified framework for the design and analysis of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 554, + 450, + 567 + ], + "spans": [ + { + "bbox": [ + 115, + 554, + 450, + 567 + ], + "score": 1.0, + "content": "communication-efficient SGD algorithms. arXiv preprint arXiv:1808.07576, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 506, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "Shiqiang Wang, Tiffany Tuor, Theodoros Salonidis, Kin K Leung, Christian Makaya, Ting He, and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "Kevin Chan. Adaptive federated learning in resource constrained edge computing systems. IEEE", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 595, + 404, + 606 + ], + "spans": [ + { + "bbox": [ + 116, + 595, + 404, + 606 + ], + "score": 1.0, + "content": "Journal on Selected Areas in Communications, 37(6):1205–1221, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 612, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "Rachel Ward, Xiaoxia Wu, and Leon Bottou. Adagrad stepsizes: Sharp convergence over nonconvex", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 623, + 423, + 636 + ], + "spans": [ + { + "bbox": [ + 116, + 623, + 423, + 636 + ], + "score": 1.0, + "content": "landscapes, from any initialization. arXiv preprint arXiv:1806.01811, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 503, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Xiaoxia Wu, Simon S Du, and Rachel Ward. Global convergence of adaptive gradient methods for an", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 652, + 424, + 664 + ], + "spans": [ + { + "bbox": [ + 115, + 652, + 424, + 664 + ], + "score": 1.0, + "content": "over-parameterized neural network. arXiv preprint arXiv:1902.07111, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 108, + 669, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 684 + ], + "score": 1.0, + "content": "Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 680, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 288, + 693 + ], + "score": 1.0, + "content": "Computer Vision (ECCV), pp. 3–19, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Cong Xie, Oluwasanmi Koyejo, Indranil Gupta, and Haibin Lin. Local AdaAlter: Communication-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "efficient stochastic gradient descent with adaptive learning rates. arXiv preprint arXiv:1911.09030,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 114, + 718, + 143, + 733 + ], + "spans": [ + { + "bbox": [ + 114, + 718, + 143, + 733 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 92, + 486, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 92, + 486, + 106 + ], + "score": 1.0, + "content": "Federated optimization in heterogeneous networks. arXiv preprint arXiv:1812.06127, 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 81, + 506, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 505, + 134 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 507, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 507, + 126 + ], + "score": 1.0, + "content": "Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 122, + 410, + 134 + ], + "spans": [ + { + "bbox": [ + 116, + 122, + 410, + 134 + ], + "score": 1.0, + "content": "methods, and future directions. arXiv preprint arXiv:1908.07873, 2019a.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 109, + 507, + 134 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 140, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 139, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 505, + 153 + ], + "score": 1.0, + "content": "Wei Li and Andrew McCallum. Pachinko allocation: DAG-structured mixture models of topic", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 150, + 507, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 150, + 507, + 163 + ], + "score": 1.0, + "content": "correlations. In Proceedings of the 23rd international conference on Machine learning, pp. 577–", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 161, + 162, + 173 + ], + "spans": [ + { + "bbox": [ + 115, + 161, + 162, + 173 + ], + "score": 1.0, + "content": "584, 2006.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 139, + 507, + 173 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 505, + 202 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "score": 1.0, + "content": "Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 190, + 387, + 203 + ], + "spans": [ + { + "bbox": [ + 116, + 190, + 387, + 203 + ], + "score": 1.0, + "content": "FedAvg on non-IID data. arXiv preprint arXiv:1907.02189, 2019b.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 179, + 505, + 203 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "Xiaoyu Li and Francesco Orabona. On the convergence of stochastic gradient descent with adaptive", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 220, + 320, + 231 + ], + "spans": [ + { + "bbox": [ + 115, + 220, + 320, + 231 + ], + "score": 1.0, + "content": "stepsizes. arXiv preprint arXiv:1805.08114, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 208, + 505, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 506, + 282 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 250 + ], + "score": 1.0, + "content": "Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 248, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 248, + 506, + 261 + ], + "score": 1.0, + "content": "bound of learning rate. In 7th International Conference on Learning Representations, ICLR 2019,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 258, + 508, + 272 + ], + "spans": [ + { + "bbox": [ + 114, + 258, + 508, + 272 + ], + "score": 1.0, + "content": "New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. URL https://openreview.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 270, + 259, + 282 + ], + "spans": [ + { + "bbox": [ + 116, + 270, + 259, + 282 + ], + "score": 1.0, + "content": "net/forum?id=Bkg3g2R9FX.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 237, + 508, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 288, + 506, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 507, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 507, + 300 + ], + "score": 1.0, + "content": "Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 117, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "Communication-efficient learning of deep networks from decentralized data. In Proceedings of the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 309, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 116, + 309, + 506, + 321 + ], + "score": 1.0, + "content": "20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 318, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 114, + 318, + 507, + 333 + ], + "score": 1.0, + "content": "2017, Fort Lauderdale, FL, USA, pp. 1273–1282, 2017. URL http://proceedings.mlr.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 331, + 271, + 343 + ], + "spans": [ + { + "bbox": [ + 115, + 331, + 271, + 343 + ], + "score": 1.0, + "content": "press/v54/mcmahan17a.html.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 286, + 507, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 349, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 362 + ], + "score": 1.0, + "content": "H. Brendan McMahan and Matthew Streeter. Adaptive bound optimization for online convex", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 360, + 242, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 360, + 242, + 372 + ], + "score": 1.0, + "content": "optimization. In COLT, 2010a.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 347, + 505, + 372 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 377, + 504, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 377, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 391 + ], + "score": 1.0, + "content": "H. Brendan McMahan and Matthew J. Streeter. Adaptive bound optimization for online convex", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 389, + 412, + 401 + ], + "spans": [ + { + "bbox": [ + 116, + 389, + 412, + 401 + ], + "score": 1.0, + "content": "optimization. In COLT The 23rd Conference on Learning Theory, 2010b.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 377, + 505, + 401 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 504, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "Sashank J Reddi, Ahmed Hefny, Suvrit Sra, Barnabás Póczós, and Alex Smola. Stochastic variance", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 417, + 375, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 375, + 430 + ], + "score": 1.0, + "content": "reduction for nonconvex optimization. arXiv:1603.06160, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 406, + 505, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 504, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 448 + ], + "score": 1.0, + "content": "Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of ADAM and beyond. arXiv", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 447, + 254, + 459 + ], + "spans": [ + { + "bbox": [ + 115, + 447, + 254, + 459 + ], + "score": 1.0, + "content": "preprint arXiv:1904.09237, 2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 434, + 505, + 459 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 506, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 507, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 507, + 477 + ], + "score": 1.0, + "content": "Sebastian U. Stich. Local SGD converges fast and communicates little. In International Confer-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 506, + 488 + ], + "score": 1.0, + "content": "ence on Learning Representations, 2019. URL https://openreview.net/forum?id=", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 487, + 181, + 498 + ], + "spans": [ + { + "bbox": [ + 116, + 487, + 181, + 498 + ], + "score": 1.0, + "content": "S1g2JnRcFX.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 464, + 507, + 498 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 504, + 506, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "Sebastian U Stich and Sai Praneeth Karimireddy. The error-feedback framework: Better rates for", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 115, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "SGD with delayed gradients and compressed communication. arXiv preprint arXiv:1909.05350,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 524, + 143, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 524, + 143, + 538 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 504, + 506, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 543, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 544, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 506, + 557 + ], + "score": 1.0, + "content": "Jianyu Wang and Gauri Joshi. Cooperative SGD: A unified framework for the design and analysis of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 554, + 450, + 567 + ], + "spans": [ + { + "bbox": [ + 115, + 554, + 450, + 567 + ], + "score": 1.0, + "content": "communication-efficient SGD algorithms. arXiv preprint arXiv:1808.07576, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 544, + 506, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 506, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 506, + 586 + ], + "score": 1.0, + "content": "Shiqiang Wang, Tiffany Tuor, Theodoros Salonidis, Kin K Leung, Christian Makaya, Ting He, and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "Kevin Chan. Adaptive federated learning in resource constrained edge computing systems. IEEE", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 595, + 404, + 606 + ], + "spans": [ + { + "bbox": [ + 116, + 595, + 404, + 606 + ], + "score": 1.0, + "content": "Journal on Selected Areas in Communications, 37(6):1205–1221, 2019.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 572, + 506, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 612, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "Rachel Ward, Xiaoxia Wu, and Leon Bottou. Adagrad stepsizes: Sharp convergence over nonconvex", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 623, + 423, + 636 + ], + "spans": [ + { + "bbox": [ + 116, + 623, + 423, + 636 + ], + "score": 1.0, + "content": "landscapes, from any initialization. arXiv preprint arXiv:1806.01811, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 612, + 506, + 636 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 503, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 654 + ], + "score": 1.0, + "content": "Xiaoxia Wu, Simon S Du, and Rachel Ward. Global convergence of adaptive gradient methods for an", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 652, + 424, + 664 + ], + "spans": [ + { + "bbox": [ + 115, + 652, + 424, + 664 + ], + "score": 1.0, + "content": "over-parameterized neural network. arXiv preprint arXiv:1902.07111, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 640, + 505, + 664 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 669, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 684 + ], + "score": 1.0, + "content": "Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 680, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 288, + 693 + ], + "score": 1.0, + "content": "Computer Vision (ECCV), pp. 3–19, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 668, + 505, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Cong Xie, Oluwasanmi Koyejo, Indranil Gupta, and Haibin Lin. Local AdaAlter: Communication-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "efficient stochastic gradient descent with adaptive learning rates. arXiv preprint arXiv:1911.09030,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 114, + 718, + 143, + 733 + ], + "spans": [ + { + "bbox": [ + 114, + 718, + 143, + 733 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 699, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Hao Yu, Sen Yang, and Shenghuo Zhu. Parallel restarted SGD with faster convergence and less", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "communication: Demystifying why model averaging works for deep learning. In Proceedings of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 442, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 442, + 117 + ], + "score": 1.0, + "content": "the AAAI Conference on Artificial Intelligence, volume 33, pp. 5693–5700, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 505, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Manzil Zaheer, Sashank Reddi, Devendra Sachan, Satyen Kale, and Sanjiv Kumar. Adaptive methods", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 506, + 146 + ], + "score": 1.0, + "content": "for nonconvex optimization. In Advances in Neural Information Processing Systems, pp. 9815–", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 167, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 167, + 156 + ], + "score": 1.0, + "content": "9825, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 162, + 505, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "score": 1.0, + "content": "Jingzhao Zhang, Sai Praneeth Karimireddy, Andreas Veit, Seungyeon Kim, Sashank J. Reddi,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 505, + 185 + ], + "score": 1.0, + "content": "Sanjiv Kumar, and Suvrit Sra. Why ADAM beats SGD for attention models. arXiv preprint", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 184, + 222, + 196 + ], + "spans": [ + { + "bbox": [ + 115, + 184, + 222, + 196 + ], + "score": 1.0, + "content": "arxiv:1912.03194, 2019a.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 202, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "Michael R. Zhang, James Lucas, Jimmy Ba, and Geoffrey E. Hinton. Lookahead optimizer: k", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 213, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 506, + 228 + ], + "score": 1.0, + "content": "steps forward, 1 step back. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 223, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 114, + 223, + 506, + 239 + ], + "score": 1.0, + "content": "d’Alché-Buc, Emily B. Fox, and Roman Garnett (eds.), Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 245, + 400, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 400, + 259 + ], + "score": 1.0, + "content": "8-14 December 2019, Vancouver, BC, Canada, pp. 9593–9604, 2019b.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 504, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "score": 1.0, + "content": "Martin Zinkevich, Markus Weimer, Lihong Li, and Alex J Smola. Parallelized stochastic gradient", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 275, + 461, + 288 + ], + "spans": [ + { + "bbox": [ + 116, + 275, + 461, + 288 + ], + "score": 1.0, + "content": "descent. In Advances in neural information processing systems, pp. 2595–2603, 2010.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Hao Yu, Sen Yang, and Shenghuo Zhu. Parallel restarted SGD with faster convergence and less", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "communication: Demystifying why model averaging works for deep learning. In Proceedings of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 442, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 442, + 117 + ], + "score": 1.0, + "content": "the AAAI Conference on Artificial Intelligence, volume 33, pp. 5693–5700, 2019.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 506, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 505, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Manzil Zaheer, Sashank Reddi, Devendra Sachan, Satyen Kale, and Sanjiv Kumar. Adaptive methods", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 506, + 146 + ], + "score": 1.0, + "content": "for nonconvex optimization. In Advances in Neural Information Processing Systems, pp. 9815–", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 167, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 167, + 156 + ], + "score": 1.0, + "content": "9825, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 122, + 506, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 162, + 505, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "score": 1.0, + "content": "Jingzhao Zhang, Sai Praneeth Karimireddy, Andreas Veit, Seungyeon Kim, Sashank J. Reddi,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 505, + 185 + ], + "score": 1.0, + "content": "Sanjiv Kumar, and Suvrit Sra. Why ADAM beats SGD for attention models. arXiv preprint", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 184, + 222, + 196 + ], + "spans": [ + { + "bbox": [ + 115, + 184, + 222, + 196 + ], + "score": 1.0, + "content": "arxiv:1912.03194, 2019a.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 162, + 506, + 196 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 202, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "Michael R. Zhang, James Lucas, Jimmy Ba, and Geoffrey E. Hinton. Lookahead optimizer: k", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 213, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 506, + 228 + ], + "score": 1.0, + "content": "steps forward, 1 step back. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 223, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 114, + 223, + 506, + 239 + ], + "score": 1.0, + "content": "d’Alché-Buc, Emily B. Fox, and Roman Garnett (eds.), Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 245, + 400, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 400, + 259 + ], + "score": 1.0, + "content": "8-14 December 2019, Vancouver, BC, Canada, pp. 9593–9604, 2019b.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 203, + 506, + 259 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 504, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "score": 1.0, + "content": "Martin Zinkevich, Markus Weimer, Lihong Li, and Alex J Smola. Parallelized stochastic gradient", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 275, + 461, + 288 + ], + "spans": [ + { + "bbox": [ + 116, + 275, + 461, + 288 + ], + "score": 1.0, + "content": "descent. In Advances in neural information processing systems, pp. 2595–2603, 2010.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 264, + 505, + 288 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 228, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 229, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 229, + 96 + ], + "score": 1.0, + "content": "A PROOF OF RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 109, + 220, + 121 + ], + "lines": [ + { + "bbox": [ + 106, + 108, + 221, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 221, + 122 + ], + "score": 1.0, + "content": "A.1 MAIN CHALLENGES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "We first recap some of the central challenges to our analysis. Theoretical analyses of optimization", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "methods for federated learning are much different than analyses for centralized settings. The key", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 154, + 257, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 257, + 167 + ], + "score": 1.0, + "content": "factors complicating the analysis are:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 129, + 176, + 330, + 229 + ], + "lines": [ + { + "bbox": [ + 129, + 175, + 313, + 191 + ], + "spans": [ + { + "bbox": [ + 129, + 175, + 313, + 191 + ], + "score": 1.0, + "content": "1. Clients performing multiple local updates.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 128, + 195, + 223, + 211 + ], + "spans": [ + { + "bbox": [ + 128, + 195, + 223, + 211 + ], + "score": 1.0, + "content": "2. Data heterogeneity.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 216, + 331, + 231 + ], + "spans": [ + { + "bbox": [ + 129, + 216, + 331, + 231 + ], + "score": 1.0, + "content": "3. Understanding the communication complexity.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 506, + 327 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "As a result of (1), the updates from the clients to the server are not gradients, or even unbiased estimates", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 250, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 266 + ], + "score": 1.0, + "content": "of gradients, they are pseudo-gradients (see Section 2). These pseudo-gradients are challenging to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "analyze as they can have both high bias (their expectation is not the gradient of the empirical loss", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "score": 1.0, + "content": "function) and high variance (due to compounding variance across client updates) and are therefore", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 283, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 442, + 296 + ], + "score": 1.0, + "content": "challenging to bound. This is exacerbated by (2), which we quantify by the parameter", + "type": "text" + }, + { + "bbox": [ + 443, + 285, + 454, + 296 + ], + "score": 0.88, + "content": "\\sigma _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 283, + 506, + 296 + ], + "score": 1.0, + "content": "in Section 2.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "Things are further complicated by (3), as we must obtain a good trade-off between the number of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 232, + 318 + ], + "score": 1.0, + "content": "client updates taken per round (", + "type": "text" + }, + { + "bbox": [ + 232, + 306, + 242, + 315 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 304, + 505, + 318 + ], + "score": 1.0, + "content": "in Algorithms 1 and 2) and the number of communication rounds", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 315, + 342, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 115, + 326 + ], + "score": 0.63, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 315, + 342, + 328 + ], + "score": 1.0, + "content": ". Such trade-offs do not exist in centralized optimization.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 346, + 230, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 231, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 231, + 359 + ], + "score": 1.0, + "content": "A.2 PROOF OF THEOREM 1", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 108, + 368, + 446, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 448, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 448, + 383 + ], + "score": 1.0, + "content": "Proof of Theorem 1. Recall that the server update of FEDADAGRAD is the following", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 390, + 363, + 419 + ], + "lines": [ + { + "bbox": [ + 246, + 390, + 363, + 419 + ], + "spans": [ + { + "bbox": [ + 246, + 390, + 363, + 419 + ], + "score": 0.94, + "content": "x _ { t + 1 , i } = x _ { t , i } + \\eta \\frac { \\Delta _ { t , i } } { \\sqrt { v _ { t , i } } + \\tau } ,", + "type": "interline_equation", + "image_path": "70e8f36e5a1da551ed96f15bc036a6dbdec58dfee69595e72a57d119957bca0d.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 390, + 363, + 404.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 246, + 404.5, + 363, + 419.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 394, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 395, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 133, + 444 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 429, + 160, + 442 + ], + "score": 0.92, + "content": "i \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 427, + 239, + 444 + ], + "score": 1.0, + "content": ". Since the function", + "type": "text" + }, + { + "bbox": [ + 240, + 430, + 246, + 441 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 427, + 257, + 444 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 257, + 430, + 265, + 439 + ], + "score": 0.83, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 427, + 395, + 444 + ], + "score": 1.0, + "content": "-smooth, we have the following:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 451, + 446, + 512 + ], + "lines": [ + { + "bbox": [ + 164, + 451, + 446, + 512 + ], + "spans": [ + { + "bbox": [ + 164, + 451, + 446, + 512 + ], + "score": 0.95, + "content": "\\begin{array} { l } { f ( x _ { t + 1 } ) \\leq f ( x _ { t } ) + \\langle \\nabla f ( x _ { t } ) , x _ { t + 1 } - x _ { t } \\rangle + \\displaystyle \\frac { L } { 2 } \\| x _ { t + 1 } - x _ { t } \\| ^ { 2 } } \\\\ { \\displaystyle = f ( x _ { t } ) + \\eta \\left. \\nabla f ( x _ { t } ) , \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } \\right. + \\displaystyle \\frac { \\eta ^ { 2 } L } { 2 } \\sum _ { i = 1 } ^ { d } \\frac { \\Delta _ { t , i } ^ { 2 } } { ( \\sqrt { v _ { t , i } } + \\tau ) ^ { 2 } } } \\end{array}", + "type": "interline_equation", + "image_path": "a02d4e87b347ccb793d5a7bc761291d399804b6a07c0c04243b749afdacd2e0a.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 164, + 451, + 446, + 471.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 164, + 471.3333333333333, + 446, + 491.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 164, + 491.66666666666663, + 446, + 511.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 507, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 471, + 535 + ], + "score": 1.0, + "content": "The second step follows simply from FEDADAGRAD’s update. We take the expectation of", + "type": "text" + }, + { + "bbox": [ + 471, + 521, + 505, + 533 + ], + "score": 0.91, + "content": "f ( x _ { t + 1 } )", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 532, + 333, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 227, + 545 + ], + "score": 1.0, + "content": "(over randomness at time step", + "type": "text" + }, + { + "bbox": [ + 228, + 533, + 233, + 542 + ], + "score": 0.64, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 532, + 333, + 545 + ], + "score": 1.0, + "content": ") in the above inequality:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 554, + 514, + 735 + ], + "lines": [ + { + "bbox": [ + 111, + 554, + 514, + 735 + ], + "spans": [ + { + "bbox": [ + 111, + 554, + 514, + 735 + ], + "score": 0.88, + "content": "\\begin{array} { l } { \\displaystyle \\ddot { c } _ { i } [ f ( x _ { i + 1 } ) ] \\leq f ( x _ { i } ) + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\Psi _ { t - 1 } } + \\tau } \\right] \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } ^ { 2 } \\rho } { \\left( \\sqrt { \\nu _ { t - 1 } } + \\tau \\right) ^ { 2 } } \\right] } \\\\ { = f ( x _ { i } ) + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t + 1 } } + \\tau } - \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } + \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } \\\\ { \\displaystyle \\qquad + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i , j } ^ { 2 } } { \\left( \\sqrt { \\nu _ { t , j } } + \\tau \\right) ^ { 2 } } \\right] } \\\\ { = f ( x _ { i } ) + \\eta \\underbrace { \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } _ { \\displaystyle \\mathrm { ~ ( ~ 1 ~ ) ~ } } + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t } } + \\tau } - \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } \\\\ { \\displaystyle \\qquad + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i , j } ^ { 2 } } { \\left( \\sqrt { \\nu _ { t - 1 } } + \\tau \\right) ^ { 2 } } \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "4b80d6a909d656a015e6c315e7b4af7d453e3b5da9042aa0d788a35d2cbed587.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 111, + 554, + 514, + 614.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 111, + 614.3333333333334, + 514, + 674.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 111, + 674.6666666666667, + 514, + 735.0000000000001 + ], + "spans": [], + "index": 28 + } + ] + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 228, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 229, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 229, + 96 + ], + "score": 1.0, + "content": "A PROOF OF RESULTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 109, + 220, + 121 + ], + "lines": [ + { + "bbox": [ + 106, + 108, + 221, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 221, + 122 + ], + "score": 1.0, + "content": "A.1 MAIN CHALLENGES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "We first recap some of the central challenges to our analysis. Theoretical analyses of optimization", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "methods for federated learning are much different than analyses for centralized settings. The key", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 154, + 257, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 257, + 167 + ], + "score": 1.0, + "content": "factors complicating the analysis are:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 131, + 505, + 167 + ] + }, + { + "type": "index", + "bbox": [ + 129, + 176, + 330, + 229 + ], + "lines": [ + { + "bbox": [ + 129, + 175, + 313, + 191 + ], + "spans": [ + { + "bbox": [ + 129, + 175, + 313, + 191 + ], + "score": 1.0, + "content": "1. Clients performing multiple local updates.", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 195, + 223, + 211 + ], + "spans": [ + { + "bbox": [ + 128, + 195, + 223, + 211 + ], + "score": 1.0, + "content": "2. Data heterogeneity.", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 129, + 216, + 331, + 231 + ], + "spans": [ + { + "bbox": [ + 129, + 216, + 331, + 231 + ], + "score": 1.0, + "content": "3. Understanding the communication complexity.", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + } + ], + "index": 6, + "bbox_fs": [ + 128, + 175, + 331, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 506, + 327 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "As a result of (1), the updates from the clients to the server are not gradients, or even unbiased estimates", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 250, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 266 + ], + "score": 1.0, + "content": "of gradients, they are pseudo-gradients (see Section 2). These pseudo-gradients are challenging to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 505, + 275 + ], + "score": 1.0, + "content": "analyze as they can have both high bias (their expectation is not the gradient of the empirical loss", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 285 + ], + "score": 1.0, + "content": "function) and high variance (due to compounding variance across client updates) and are therefore", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 283, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 442, + 296 + ], + "score": 1.0, + "content": "challenging to bound. This is exacerbated by (2), which we quantify by the parameter", + "type": "text" + }, + { + "bbox": [ + 443, + 285, + 454, + 296 + ], + "score": 0.88, + "content": "\\sigma _ { g }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 283, + 506, + 296 + ], + "score": 1.0, + "content": "in Section 2.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "Things are further complicated by (3), as we must obtain a good trade-off between the number of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 304, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 232, + 318 + ], + "score": 1.0, + "content": "client updates taken per round (", + "type": "text" + }, + { + "bbox": [ + 232, + 306, + 242, + 315 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 304, + 505, + 318 + ], + "score": 1.0, + "content": "in Algorithms 1 and 2) and the number of communication rounds", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 315, + 342, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 115, + 326 + ], + "score": 0.63, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 315, + 342, + 328 + ], + "score": 1.0, + "content": ". Such trade-offs do not exist in centralized optimization.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 240, + 506, + 328 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 346, + 230, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 231, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 231, + 359 + ], + "score": 1.0, + "content": "A.2 PROOF OF THEOREM 1", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 108, + 368, + 446, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 448, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 448, + 383 + ], + "score": 1.0, + "content": "Proof of Theorem 1. Recall that the server update of FEDADAGRAD is the following", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 367, + 448, + 383 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 390, + 363, + 419 + ], + "lines": [ + { + "bbox": [ + 246, + 390, + 363, + 419 + ], + "spans": [ + { + "bbox": [ + 246, + 390, + 363, + 419 + ], + "score": 0.94, + "content": "x _ { t + 1 , i } = x _ { t , i } + \\eta \\frac { \\Delta _ { t , i } } { \\sqrt { v _ { t , i } } + \\tau } ,", + "type": "interline_equation", + "image_path": "70e8f36e5a1da551ed96f15bc036a6dbdec58dfee69595e72a57d119957bca0d.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 390, + 363, + 404.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 246, + 404.5, + 363, + 419.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 429, + 394, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 395, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 133, + 444 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 429, + 160, + 442 + ], + "score": 0.92, + "content": "i \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 427, + 239, + 444 + ], + "score": 1.0, + "content": ". Since the function", + "type": "text" + }, + { + "bbox": [ + 240, + 430, + 246, + 441 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 427, + 257, + 444 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 257, + 430, + 265, + 439 + ], + "score": 0.83, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 427, + 395, + 444 + ], + "score": 1.0, + "content": "-smooth, we have the following:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 427, + 395, + 444 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 451, + 446, + 512 + ], + "lines": [ + { + "bbox": [ + 164, + 451, + 446, + 512 + ], + "spans": [ + { + "bbox": [ + 164, + 451, + 446, + 512 + ], + "score": 0.95, + "content": "\\begin{array} { l } { f ( x _ { t + 1 } ) \\leq f ( x _ { t } ) + \\langle \\nabla f ( x _ { t } ) , x _ { t + 1 } - x _ { t } \\rangle + \\displaystyle \\frac { L } { 2 } \\| x _ { t + 1 } - x _ { t } \\| ^ { 2 } } \\\\ { \\displaystyle = f ( x _ { t } ) + \\eta \\left. \\nabla f ( x _ { t } ) , \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } \\right. + \\displaystyle \\frac { \\eta ^ { 2 } L } { 2 } \\sum _ { i = 1 } ^ { d } \\frac { \\Delta _ { t , i } ^ { 2 } } { ( \\sqrt { v _ { t , i } } + \\tau ) ^ { 2 } } } \\end{array}", + "type": "interline_equation", + "image_path": "a02d4e87b347ccb793d5a7bc761291d399804b6a07c0c04243b749afdacd2e0a.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 164, + 451, + 446, + 471.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 164, + 471.3333333333333, + 446, + 491.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 164, + 491.66666666666663, + 446, + 511.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 507, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 471, + 535 + ], + "score": 1.0, + "content": "The second step follows simply from FEDADAGRAD’s update. We take the expectation of", + "type": "text" + }, + { + "bbox": [ + 471, + 521, + 505, + 533 + ], + "score": 0.91, + "content": "f ( x _ { t + 1 } )", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 532, + 333, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 227, + 545 + ], + "score": 1.0, + "content": "(over randomness at time step", + "type": "text" + }, + { + "bbox": [ + 228, + 533, + 233, + 542 + ], + "score": 0.64, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 532, + 333, + 545 + ], + "score": 1.0, + "content": ") in the above inequality:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 519, + 505, + 545 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 554, + 514, + 735 + ], + "lines": [ + { + "bbox": [ + 111, + 554, + 514, + 735 + ], + "spans": [ + { + "bbox": [ + 111, + 554, + 514, + 735 + ], + "score": 0.88, + "content": "\\begin{array} { l } { \\displaystyle \\ddot { c } _ { i } [ f ( x _ { i + 1 } ) ] \\leq f ( x _ { i } ) + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\Psi _ { t - 1 } } + \\tau } \\right] \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } ^ { 2 } \\rho } { \\left( \\sqrt { \\nu _ { t - 1 } } + \\tau \\right) ^ { 2 } } \\right] } \\\\ { = f ( x _ { i } ) + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t + 1 } } + \\tau } - \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } + \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } \\\\ { \\displaystyle \\qquad + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i , j } ^ { 2 } } { \\left( \\sqrt { \\nu _ { t , j } } + \\tau \\right) ^ { 2 } } \\right] } \\\\ { = f ( x _ { i } ) + \\eta \\underbrace { \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } _ { \\displaystyle \\mathrm { ~ ( ~ 1 ~ ) ~ } } + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t } } + \\tau } - \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } \\\\ { \\displaystyle \\qquad + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i , j } ^ { 2 } } { \\left( \\sqrt { \\nu _ { t - 1 } } + \\tau \\right) ^ { 2 } } \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "4b80d6a909d656a015e6c315e7b4af7d453e3b5da9042aa0d788a35d2cbed587.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 111, + 554, + 514, + 614.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 111, + 614.3333333333334, + 514, + 674.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 111, + 674.6666666666667, + 514, + 735.0000000000001 + ], + "spans": [], + "index": 28 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 298, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 299, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 185, + 96 + ], + "score": 1.0, + "content": "We will first bound", + "type": "text" + }, + { + "bbox": [ + 185, + 83, + 196, + 93 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 80, + 299, + 96 + ], + "score": 1.0, + "content": "in the following manner:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 105, + 100, + 366, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 366, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 366, + 203 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { T _ { 2 } = \\bigg \\langle \\nabla f ( { x } _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { v _ { t - 1 } } + \\tau } \\right] \\bigg \\rangle } \\\\ & { \\quad = \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } [ \\nabla f ( { x } _ { t } ) ] _ { j } \\times \\left[ \\frac { \\Delta _ { t , j } } { \\sqrt { v _ { t , j } } + \\tau } - \\frac { \\Delta _ { t , j } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\right] } \\\\ & { \\quad = \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } [ \\nabla f ( { x } _ { t } ) ] _ { j } \\times \\Delta _ { t , j } \\times \\left[ \\frac { \\sqrt { v _ { t - 1 , j } } - \\sqrt { v _ { t , j } } } { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "df03e20292f518fafce8e9aa33fed901c8bdbcb91da1627722123ed477863934.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 105, + 100, + 366, + 134.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 134.33333333333334, + 366, + 168.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 105, + 168.66666666666669, + 366, + 203.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 208, + 473, + 225 + ], + "lines": [ + { + "bbox": [ + 104, + 208, + 476, + 226 + ], + "spans": [ + { + "bbox": [ + 104, + 208, + 160, + 226 + ], + "score": 1.0, + "content": "and recalling", + "type": "text" + }, + { + "bbox": [ + 161, + 210, + 227, + 223 + ], + "score": 0.89, + "content": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 208, + 238, + 226 + ], + "score": 1.0, + "content": "so", + "type": "text" + }, + { + "bbox": [ + 238, + 211, + 434, + 224 + ], + "score": 0.89, + "content": "- \\Delta _ { t , j } ^ { 2 } = ( \\sqrt { v _ { t - 1 , j } } - \\sqrt { v _ { t , j } } ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 208, + 476, + 226 + ], + "score": 1.0, + "content": "we have,", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 231, + 513, + 340 + ], + "lines": [ + { + "bbox": [ + 117, + 231, + 513, + 340 + ], + "spans": [ + { + "bbox": [ + 117, + 231, + 513, + 340 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } { { = \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } [ \\nabla f ( x _ { t } ) ] _ { j } \\times \\Delta _ { t , j } \\times [ \\overline { { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) } } ] } } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) } } ] } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( v _ { t , j } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } } ] } & { \\mathrm { s i n c e } v _ { t - 1 , j } \\geq \\tau } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( v _ { t , j } + \\tau ^ { 2 } ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } } ] } & { \\mathrm { s i n c e } v _ { t - 1 , j } \\geq \\tau } \\end{array}", + "type": "interline_equation", + "image_path": "89a45736bfecac0a9a4c9b691d21a2ec279cc835dfa861191c16c256200c233e.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 117, + 231, + 513, + 267.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 117, + 267.3333333333333, + 513, + 303.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 117, + 303.66666666666663, + 513, + 339.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 505, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 128, + 360 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 128, + 347, + 173, + 359 + ], + "score": 0.92, + "content": "v _ { t - 1 , j } \\geq \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 345, + 196, + 360 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 196, + 347, + 232, + 358 + ], + "score": 0.91, + "content": "v _ { - 1 } \\geq \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 345, + 399, + 360 + ], + "score": 1.0, + "content": "(see the initialization of Algorithm 2) and", + "type": "text" + }, + { + "bbox": [ + 400, + 348, + 415, + 359 + ], + "score": 0.89, + "content": "v _ { t , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 345, + 478, + 360 + ], + "score": 1.0, + "content": "is increasing in", + "type": "text" + }, + { + "bbox": [ + 479, + 348, + 483, + 356 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 345, + 505, + 360 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 356, + 381, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 381, + 370 + ], + "score": 1.0, + "content": "above bound can be further upper bounded in the following manner:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 375, + 494, + 449 + ], + "lines": [ + { + "bbox": [ + 117, + 375, + 494, + 449 + ], + "spans": [ + { + "bbox": [ + 117, + 375, + 494, + 449 + ], + "score": 0.93, + "content": "\\begin{array} { r l } { T _ { 2 } \\le \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } \\eta _ { l } K G ^ { 2 } \\left[ \\displaystyle \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } \\right] } & { \\mathrm { ~ s i n c e ~ } [ \\nabla f ( x _ { t } ) ] _ { i } \\le G \\mathrm { ~ a n d ~ } \\Delta _ { t , i } \\le \\eta _ { l } K G } \\\\ { \\le \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } \\displaystyle \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } \\left[ \\displaystyle \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] } & { \\mathrm { ~ s i n c e ~ } \\sqrt { v _ { t - 1 , j } } \\ge 0 . } \\end{array}", + "type": "interline_equation", + "image_path": "78157b390d5ece8a6d5a4fcc282b59da5c64013fc66f3ef02080752845ef50e4.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 117, + 375, + 494, + 399.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 117, + 399.6666666666667, + 494, + 424.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 117, + 424.33333333333337, + 494, + 449.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 460, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 150, + 475 + ], + "score": 1.0, + "content": "Bounding", + "type": "text" + }, + { + "bbox": [ + 151, + 462, + 162, + 472 + ], + "score": 0.86, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 460, + 360, + 475 + ], + "score": 1.0, + "content": "We now turn our attention to bounding the term", + "type": "text" + }, + { + "bbox": [ + 361, + 462, + 372, + 472 + ], + "score": 0.87, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 460, + 505, + 475 + ], + "score": 1.0, + "content": ", which we need to be sufficiently", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 471, + 252, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 252, + 485 + ], + "score": 1.0, + "content": "negative. We observe the following:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 157, + 490, + 453, + 592 + ], + "lines": [ + { + "bbox": [ + 157, + 490, + 453, + 592 + ], + "spans": [ + { + "bbox": [ + 157, + 490, + 453, + 592 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { T _ { 1 } = \\bigg \\langle \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t - 1 } } + \\tau } \\right] \\bigg \\rangle } \\\\ & { \\quad = \\bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } - \\eta _ { l } K \\nabla f ( x _ { t } ) + \\eta _ { l } K \\nabla f ( x _ { t } ) \\right] \\bigg \\rangle } \\\\ & { \\quad = - \\eta _ { l } K \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\underbrace { \\left. \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) \\right] \\right. } _ { T _ { 2 } } . } \\end{array}", + "type": "interline_equation", + "image_path": "e50b63fd171470cab38be028f883c04445fb08d3cd1f7574b91e6d4330590510.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 157, + 490, + 453, + 524.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 157, + 524.0, + 453, + 558.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 157, + 558.0, + 453, + 592.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 601, + 505, + 625 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 176, + 615 + ], + "score": 1.0, + "content": "In order to bound", + "type": "text" + }, + { + "bbox": [ + 176, + 603, + 187, + 613 + ], + "score": 0.87, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 601, + 338, + 615 + ], + "score": 1.0, + "content": ", we use the following upper bound on", + "type": "text" + }, + { + "bbox": [ + 338, + 603, + 349, + 613 + ], + "score": 0.87, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "(which captures the difference between", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 612, + 356, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 177, + 625 + ], + "score": 1.0, + "content": "the actual update", + "type": "text" + }, + { + "bbox": [ + 177, + 614, + 189, + 624 + ], + "score": 0.88, + "content": "\\Delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 612, + 309, + 625 + ], + "score": 1.0, + "content": "and an appropriate scaling of", + "type": "text" + }, + { + "bbox": [ + 309, + 613, + 352, + 625 + ], + "score": 0.91, + "content": "- \\nabla f ( x _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 612, + 356, + 625 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "interline_equation", + "bbox": [ + 105, + 632, + 394, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 394, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 394, + 733 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { T _ { 3 } = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. } \\\\ & { \\quad = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ - \\displaystyle \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\eta _ { l } g _ { i , k } ^ { t } + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. } \\\\ & { \\quad = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ - \\displaystyle \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\eta _ { l } \\nabla F _ { i } ( x _ { i , k } ^ { t } ) + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. . } \\end{array}", + "type": "interline_equation", + "image_path": "681eaac06a8b37de6351e2cb8ab0982ee32cb67d32339873a7c300da74320fbf.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 105, + 632, + 394, + 665.6666666666666 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 105, + 665.6666666666666, + 394, + 699.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 105, + 699.3333333333333, + 394, + 732.9999999999999 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 298, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 299, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 185, + 96 + ], + "score": 1.0, + "content": "We will first bound", + "type": "text" + }, + { + "bbox": [ + 185, + 83, + 196, + 93 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 80, + 299, + 96 + ], + "score": 1.0, + "content": "in the following manner:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 80, + 299, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 105, + 100, + 366, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 366, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 366, + 203 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { T _ { 2 } = \\bigg \\langle \\nabla f ( { x } _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { v _ { t - 1 } } + \\tau } \\right] \\bigg \\rangle } \\\\ & { \\quad = \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } [ \\nabla f ( { x } _ { t } ) ] _ { j } \\times \\left[ \\frac { \\Delta _ { t , j } } { \\sqrt { v _ { t , j } } + \\tau } - \\frac { \\Delta _ { t , j } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\right] } \\\\ & { \\quad = \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } [ \\nabla f ( { x } _ { t } ) ] _ { j } \\times \\Delta _ { t , j } \\times \\left[ \\frac { \\sqrt { v _ { t - 1 , j } } - \\sqrt { v _ { t , j } } } { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "df03e20292f518fafce8e9aa33fed901c8bdbcb91da1627722123ed477863934.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 105, + 100, + 366, + 134.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 134.33333333333334, + 366, + 168.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 105, + 168.66666666666669, + 366, + 203.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 208, + 473, + 225 + ], + "lines": [ + { + "bbox": [ + 104, + 208, + 476, + 226 + ], + "spans": [ + { + "bbox": [ + 104, + 208, + 160, + 226 + ], + "score": 1.0, + "content": "and recalling", + "type": "text" + }, + { + "bbox": [ + 161, + 210, + 227, + 223 + ], + "score": 0.89, + "content": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 208, + 238, + 226 + ], + "score": 1.0, + "content": "so", + "type": "text" + }, + { + "bbox": [ + 238, + 211, + 434, + 224 + ], + "score": 0.89, + "content": "- \\Delta _ { t , j } ^ { 2 } = ( \\sqrt { v _ { t - 1 , j } } - \\sqrt { v _ { t , j } } ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 208, + 476, + 226 + ], + "score": 1.0, + "content": "we have,", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 104, + 208, + 476, + 226 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 231, + 513, + 340 + ], + "lines": [ + { + "bbox": [ + 117, + 231, + 513, + 340 + ], + "spans": [ + { + "bbox": [ + 117, + 231, + 513, + 340 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } { { = \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } [ \\nabla f ( x _ { t } ) ] _ { j } \\times \\Delta _ { t , j } \\times [ \\overline { { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) } } ] } } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) } } ] } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( v _ { t , j } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } } ] } & { \\mathrm { s i n c e } v _ { t - 1 , j } \\geq \\tau } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( v _ { t , j } + \\tau ^ { 2 } ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } } ] } & { \\mathrm { s i n c e } v _ { t - 1 , j } \\geq \\tau } \\end{array}", + "type": "interline_equation", + "image_path": "89a45736bfecac0a9a4c9b691d21a2ec279cc835dfa861191c16c256200c233e.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 117, + 231, + 513, + 267.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 117, + 267.3333333333333, + 513, + 303.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 117, + 303.66666666666663, + 513, + 339.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 346, + 505, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 128, + 360 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 128, + 347, + 173, + 359 + ], + "score": 0.92, + "content": "v _ { t - 1 , j } \\geq \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 345, + 196, + 360 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 196, + 347, + 232, + 358 + ], + "score": 0.91, + "content": "v _ { - 1 } \\geq \\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 345, + 399, + 360 + ], + "score": 1.0, + "content": "(see the initialization of Algorithm 2) and", + "type": "text" + }, + { + "bbox": [ + 400, + 348, + 415, + 359 + ], + "score": 0.89, + "content": "v _ { t , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 345, + 478, + 360 + ], + "score": 1.0, + "content": "is increasing in", + "type": "text" + }, + { + "bbox": [ + 479, + 348, + 483, + 356 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 345, + 505, + 360 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 356, + 381, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 381, + 370 + ], + "score": 1.0, + "content": "above bound can be further upper bounded in the following manner:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 345, + 505, + 370 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 375, + 494, + 449 + ], + "lines": [ + { + "bbox": [ + 117, + 375, + 494, + 449 + ], + "spans": [ + { + "bbox": [ + 117, + 375, + 494, + 449 + ], + "score": 0.93, + "content": "\\begin{array} { r l } { T _ { 2 } \\le \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } \\eta _ { l } K G ^ { 2 } \\left[ \\displaystyle \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } \\right] } & { \\mathrm { ~ s i n c e ~ } [ \\nabla f ( x _ { t } ) ] _ { i } \\le G \\mathrm { ~ a n d ~ } \\Delta _ { t , i } \\le \\eta _ { l } K G } \\\\ { \\le \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } \\displaystyle \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } \\left[ \\displaystyle \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] } & { \\mathrm { ~ s i n c e ~ } \\sqrt { v _ { t - 1 , j } } \\ge 0 . } \\end{array}", + "type": "interline_equation", + "image_path": "78157b390d5ece8a6d5a4fcc282b59da5c64013fc66f3ef02080752845ef50e4.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 117, + 375, + 494, + 399.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 117, + 399.6666666666667, + 494, + 424.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 117, + 424.33333333333337, + 494, + 449.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 460, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 150, + 475 + ], + "score": 1.0, + "content": "Bounding", + "type": "text" + }, + { + "bbox": [ + 151, + 462, + 162, + 472 + ], + "score": 0.86, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 460, + 360, + 475 + ], + "score": 1.0, + "content": "We now turn our attention to bounding the term", + "type": "text" + }, + { + "bbox": [ + 361, + 462, + 372, + 472 + ], + "score": 0.87, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 460, + 505, + 475 + ], + "score": 1.0, + "content": ", which we need to be sufficiently", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 471, + 252, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 252, + 485 + ], + "score": 1.0, + "content": "negative. We observe the following:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 460, + 505, + 485 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 157, + 490, + 453, + 592 + ], + "lines": [ + { + "bbox": [ + 157, + 490, + 453, + 592 + ], + "spans": [ + { + "bbox": [ + 157, + 490, + 453, + 592 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { T _ { 1 } = \\bigg \\langle \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t - 1 } } + \\tau } \\right] \\bigg \\rangle } \\\\ & { \\quad = \\bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } - \\eta _ { l } K \\nabla f ( x _ { t } ) + \\eta _ { l } K \\nabla f ( x _ { t } ) \\right] \\bigg \\rangle } \\\\ & { \\quad = - \\eta _ { l } K \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\underbrace { \\left. \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) \\right] \\right. } _ { T _ { 2 } } . } \\end{array}", + "type": "interline_equation", + "image_path": "e50b63fd171470cab38be028f883c04445fb08d3cd1f7574b91e6d4330590510.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 157, + 490, + 453, + 524.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 157, + 524.0, + 453, + 558.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 157, + 558.0, + 453, + 592.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 601, + 505, + 625 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 176, + 615 + ], + "score": 1.0, + "content": "In order to bound", + "type": "text" + }, + { + "bbox": [ + 176, + 603, + 187, + 613 + ], + "score": 0.87, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 601, + 338, + 615 + ], + "score": 1.0, + "content": ", we use the following upper bound on", + "type": "text" + }, + { + "bbox": [ + 338, + 603, + 349, + 613 + ], + "score": 0.87, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "(which captures the difference between", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 612, + 356, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 177, + 625 + ], + "score": 1.0, + "content": "the actual update", + "type": "text" + }, + { + "bbox": [ + 177, + 614, + 189, + 624 + ], + "score": 0.88, + "content": "\\Delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 612, + 309, + 625 + ], + "score": 1.0, + "content": "and an appropriate scaling of", + "type": "text" + }, + { + "bbox": [ + 309, + 613, + 352, + 625 + ], + "score": 0.91, + "content": "- \\nabla f ( x _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 612, + 356, + 625 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 601, + 505, + 625 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 105, + 632, + 394, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 394, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 394, + 733 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { T _ { 3 } = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. } \\\\ & { \\quad = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ - \\displaystyle \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\eta _ { l } g _ { i , k } ^ { t } + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. } \\\\ & { \\quad = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ - \\displaystyle \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\eta _ { l } \\nabla F _ { i } ( x _ { i , k } ^ { t } ) + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. . } \\end{array}", + "type": "interline_equation", + "image_path": "681eaac06a8b37de6351e2cb8ab0982ee32cb67d32339873a7c300da74320fbf.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 105, + 632, + 394, + 665.6666666666666 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 105, + 665.6666666666666, + 394, + 699.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 105, + 699.3333333333333, + 394, + 732.9999999999999 + ], + "spans": [], + "index": 22 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 79, + 506, + 119 + ], + "lines": [ + { + "bbox": [ + 103, + 75, + 509, + 99 + ], + "spans": [ + { + "bbox": [ + 103, + 75, + 218, + 99 + ], + "score": 1.0, + "content": "Here we used the fact that", + "type": "text" + }, + { + "bbox": [ + 218, + 80, + 338, + 94 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\nabla f ( x _ { t } ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\nabla F _ { i } ( x _ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 75, + 357, + 99 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 358, + 81, + 374, + 95 + ], + "score": 0.92, + "content": "g _ { i , k } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 75, + 509, + 99 + ], + "score": 1.0, + "content": "is an unbiased estimator of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 103, + 90, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 103, + 90, + 153, + 110 + ], + "score": 1.0, + "content": "gradient at", + "type": "text" + }, + { + "bbox": [ + 153, + 93, + 170, + 108 + ], + "score": 0.91, + "content": "\\boldsymbol { x } _ { i , k } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 90, + 248, + 110 + ], + "score": 1.0, + "content": ", we further bound", + "type": "text" + }, + { + "bbox": [ + 249, + 95, + 260, + 105 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 90, + 481, + 110 + ], + "score": 1.0, + "content": "as follows using a simple application of the fact that", + "type": "text" + }, + { + "bbox": [ + 482, + 94, + 505, + 105 + ], + "score": 0.8, + "content": "a b \\leq", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 164, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 155, + 119 + ], + "score": 0.87, + "content": "( a ^ { 2 } + b ^ { 2 } ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 105, + 164, + 120 + ], + "score": 1.0, + "content": ". :", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 119, + 513, + 236 + ], + "lines": [ + { + "bbox": [ + 111, + 119, + 513, + 236 + ], + "spans": [ + { + "bbox": [ + 111, + 119, + 513, + 236 + ], + "score": 0.92, + "content": "\\begin{array} { r l } { r _ { 3 } \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\sum _ { j } ^ { d } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } } { 2 K } \\mathbb { E } _ { t } \\left[ \\left\\| \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\frac { \\nabla F _ { i } ( x _ { i , k } ^ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } - \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\frac { \\nabla F _ { i } ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } \\right\\| ^ { 2 } \\right] } & { } \\\\ { \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\left[ \\nabla f ( x _ { t } ) \\right] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } } { 2 m } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\left\\| \\frac { \\nabla F _ { i } ( x _ { i , k } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } \\right\\| ^ { 2 } \\right] } & { } \\\\ { \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\left[ \\nabla f ( x _ { t } ) \\right] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right] } & { \\mathrm { u s i n g ~ A s s u m p t i o n ~ 1 ~ a n d ~ } v _ { t - 1 } \\geq 0 } \\end{array}", + "type": "interline_equation", + "image_path": "135fb6d27abb4f64ac497fc2e6dfe5212caeb8bde829228a6b2a68b4ae53ff54.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 111, + 119, + 513, + 158.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 158.0, + 513, + 197.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 111, + 197.0, + 513, + 236.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 506, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 246, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 419, + 259 + ], + "score": 1.0, + "content": "The second inequality follows from Lemma 6. The last inequality follows from", + "type": "text" + }, + { + "bbox": [ + 419, + 248, + 426, + 257 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 246, + 506, + 259 + ], + "score": 1.0, + "content": "-Lipschitz nature of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 256, + 493, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 437, + 272 + ], + "score": 1.0, + "content": "the gradient (Assumption 1). We now prove a lemma that bounds the “drift” of the", + "type": "text" + }, + { + "bbox": [ + 437, + 257, + 454, + 271 + ], + "score": 0.92, + "content": "\\dot { \\boldsymbol { x } } _ { i , k } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 256, + 477, + 272 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 477, + 259, + 488, + 269 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 256, + 493, + 272 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 105, + 275, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 104, + 274, + 504, + 294 + ], + "spans": [ + { + "bbox": [ + 104, + 274, + 260, + 294 + ], + "score": 1.0, + "content": "Lemma 3. For any step-size satisfying", + "type": "text" + }, + { + "bbox": [ + 260, + 276, + 301, + 290 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\eta _ { l } \\le \\frac { 1 } { 8 L K } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 274, + 423, + 294 + ], + "score": 1.0, + "content": ", we can bound the drift for any", + "type": "text" + }, + { + "bbox": [ + 423, + 277, + 504, + 289 + ], + "score": 0.9, + "content": "k \\in \\{ 0 , \\cdots , K - 1 \\}", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 289, + 119, + 300 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 119, + 300 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 297, + 470, + 334 + ], + "lines": [ + { + "bbox": [ + 140, + 297, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 140, + 297, + 470, + 334 + ], + "score": 0.93, + "content": "\\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 2 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\big ] .", + "type": "interline_equation", + "image_path": "877d85de065b125fcbdd573ab793a40671501e8b9fe2d9027ba7faad2e0b960c.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 140, + 297, + 470, + 309.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 140, + 309.3333333333333, + 470, + 321.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 140, + 321.66666666666663, + 470, + 333.99999999999994 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 344, + 506, + 380 + ], + "lines": [ + { + "bbox": [ + 104, + 342, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 342, + 249, + 361 + ], + "score": 1.0, + "content": "Proof. The result trivially holds for", + "type": "text" + }, + { + "bbox": [ + 249, + 345, + 275, + 355 + ], + "score": 0.9, + "content": "k = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 342, + 298, + 361 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 298, + 345, + 337, + 358 + ], + "score": 0.95, + "content": "x _ { i , 0 } ^ { t } = x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 342, + 365, + 361 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 365, + 345, + 396, + 357 + ], + "score": 0.93, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 342, + 507, + 361 + ], + "score": 1.0, + "content": ". We now turn our attention", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 182, + 370 + ], + "score": 1.0, + "content": "to the case where", + "type": "text" + }, + { + "bbox": [ + 182, + 357, + 210, + 369 + ], + "score": 0.9, + "content": "k \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 357, + 452, + 370 + ], + "score": 1.0, + "content": ". To prove the above result, we observe that for any client", + "type": "text" + }, + { + "bbox": [ + 452, + 357, + 486, + 370 + ], + "score": 0.92, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 366, + 146, + 382 + ], + "spans": [ + { + "bbox": [ + 107, + 368, + 140, + 380 + ], + "score": 0.91, + "content": "k \\in [ K ]", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 366, + 146, + 382 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 379, + 514, + 462 + ], + "lines": [ + { + "bbox": [ + 111, + 379, + 514, + 462 + ], + "spans": [ + { + "bbox": [ + 111, + 379, + 514, + 462 + ], + "score": 0.9, + "content": "\\begin{array} { r l } { { \\tilde { \\mathbf { \\mu } } _ { i , k , - } ^ { t } - x _ { t } \\| ^ { 2 } = \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } - \\eta _ { l } g _ { i , k - 1 } ^ { t } \\big \\| ^ { 2 } } } \\\\ & { \\leq \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } - \\eta _ { l } ( g _ { i , k - 1 } ^ { t } - \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) + \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) + \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) + \\nabla F _ { i } ( x _ { t } ) - \\nabla F ( x _ { t } ) \\big \\| ^ { 2 } } \\\\ & { \\leq ( 1 + \\frac { 1 } { 2 K - 1 } ) \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\big \\| ^ { 2 } + \\mathbb { E } \\big \\| \\eta _ { l } ( g _ { i , k - 1 } ^ { t } - \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) ) \\big \\| ^ { 2 } } \\\\ & { \\qquad + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) ) \\big \\| ^ { 2 } ] + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\big \\| ^ { 2 } \\big ] + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } } \\end{array}", + "type": "interline_equation", + "image_path": "b92da40aa6cb42737b33859fef28eadcd6c3cc547d30b73fe54c6f61fb504c6f.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 111, + 379, + 514, + 406.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 111, + 406.6666666666667, + 514, + 434.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 111, + 434.33333333333337, + 514, + 462.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 527, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 507, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 254, + 476 + ], + "score": 1.0, + "content": "The first inequality uses the fact that", + "type": "text" + }, + { + "bbox": [ + 255, + 462, + 281, + 475 + ], + "score": 0.93, + "content": "g _ { k - 1 , i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 460, + 393, + 476 + ], + "score": 1.0, + "content": "is an unbiased estimator of", + "type": "text" + }, + { + "bbox": [ + 393, + 462, + 445, + 476 + ], + "score": 0.93, + "content": "\\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 460, + 507, + 476 + ], + "score": 1.0, + "content": "and Lemma 7.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 473, + 350, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 350, + 487 + ], + "score": 1.0, + "content": "The above quantity can be further bounded by the following:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 486, + 503, + 591 + ], + "lines": [ + { + "bbox": [ + 111, + 486, + 503, + 591 + ], + "spans": [ + { + "bbox": [ + 111, + 486, + 503, + 591 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { { \\mathbb { E } \\Vert x _ { i , k } ^ { t } - x _ { t } \\Vert ^ { 2 } \\leq ( 1 + \\frac { 1 } { 2 K - 1 } ) \\mathbb { E } \\Vert x _ { i , k - 1 } ^ { t } - x _ { t } \\Vert ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\Vert L ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \\Vert ^ { 2 } } } \\\\ & { } & { + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\Vert ^ { 2 } ] + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } \\nabla f ( x _ { t } ) ) ) \\Vert ^ { 2 } ] } \\\\ & { } & { = ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\mathbb { E } \\Vert ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \\Vert ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } } \\\\ & { } & { + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\Vert ^ { 2 } ] + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\Vert \\nabla f ( x _ { t } ) ) ) \\Vert ^ { 2 } ] } \\end{array}", + "type": "interline_equation", + "image_path": "62ce5a6b5255a01075f279a3e565de1e807443577d0e01b14c00e33e37b7137c.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 111, + 486, + 503, + 521.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 111, + 521.0, + 503, + 556.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 111, + 556.0, + 503, + 591.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 504, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 455, + 603 + ], + "score": 1.0, + "content": "Here, the first inequality follows from Assumption 1 and 2. Averaging over the clients", + "type": "text" + }, + { + "bbox": [ + 455, + 591, + 459, + 600 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 588, + 505, + 603 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 599, + 165, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 165, + 614 + ], + "score": 1.0, + "content": "the following:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 612, + 514, + 736 + ], + "lines": [ + { + "bbox": [ + 111, + 612, + 514, + 736 + ], + "spans": [ + { + "bbox": [ + 111, + 612, + 514, + 736 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } { { \\frac { 1 } { n } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } } } \\\\ & { } & { + \\frac { 6 K } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } [ \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\| ^ { 2 } ] + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) } \\\\ & { } & { \\leq ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\frac { 1 } { m } \\displaystyle \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K o ( \\eta _ { l } ^ { 2 } ) ) } \\\\ & { } & { + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) \\| ^ { 2 } ] } \\end{array}", + "type": "interline_equation", + "image_path": "9907b46276cfcd32c409d8f54bf5fb2a33c7b852993595a81028be61508d4d18.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 111, + 612, + 514, + 653.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 111, + 653.3333333333334, + 514, + 694.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 111, + 694.6666666666667, + 514, + 736.0000000000001 + ], + "spans": [], + "index": 28 + } + ] + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 79, + 506, + 119 + ], + "lines": [ + { + "bbox": [ + 103, + 75, + 509, + 99 + ], + "spans": [ + { + "bbox": [ + 103, + 75, + 218, + 99 + ], + "score": 1.0, + "content": "Here we used the fact that", + "type": "text" + }, + { + "bbox": [ + 218, + 80, + 338, + 94 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\nabla f ( x _ { t } ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\nabla F _ { i } ( x _ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 75, + 357, + 99 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 358, + 81, + 374, + 95 + ], + "score": 0.92, + "content": "g _ { i , k } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 75, + 509, + 99 + ], + "score": 1.0, + "content": "is an unbiased estimator of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 103, + 90, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 103, + 90, + 153, + 110 + ], + "score": 1.0, + "content": "gradient at", + "type": "text" + }, + { + "bbox": [ + 153, + 93, + 170, + 108 + ], + "score": 0.91, + "content": "\\boldsymbol { x } _ { i , k } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 90, + 248, + 110 + ], + "score": 1.0, + "content": ", we further bound", + "type": "text" + }, + { + "bbox": [ + 249, + 95, + 260, + 105 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 90, + 481, + 110 + ], + "score": 1.0, + "content": "as follows using a simple application of the fact that", + "type": "text" + }, + { + "bbox": [ + 482, + 94, + 505, + 105 + ], + "score": 0.8, + "content": "a b \\leq", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 164, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 155, + 119 + ], + "score": 0.87, + "content": "( a ^ { 2 } + b ^ { 2 } ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 105, + 164, + 120 + ], + "score": 1.0, + "content": ". :", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 103, + 75, + 509, + 120 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 119, + 513, + 236 + ], + "lines": [ + { + "bbox": [ + 111, + 119, + 513, + 236 + ], + "spans": [ + { + "bbox": [ + 111, + 119, + 513, + 236 + ], + "score": 0.92, + "content": "\\begin{array} { r l } { r _ { 3 } \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\sum _ { j } ^ { d } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } } { 2 K } \\mathbb { E } _ { t } \\left[ \\left\\| \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\frac { \\nabla F _ { i } ( x _ { i , k } ^ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } - \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\frac { \\nabla F _ { i } ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } \\right\\| ^ { 2 } \\right] } & { } \\\\ { \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\left[ \\nabla f ( x _ { t } ) \\right] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } } { 2 m } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\left\\| \\frac { \\nabla F _ { i } ( x _ { i , k } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } \\right\\| ^ { 2 } \\right] } & { } \\\\ { \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\left[ \\nabla f ( x _ { t } ) \\right] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right] } & { \\mathrm { u s i n g ~ A s s u m p t i o n ~ 1 ~ a n d ~ } v _ { t - 1 } \\geq 0 } \\end{array}", + "type": "interline_equation", + "image_path": "135fb6d27abb4f64ac497fc2e6dfe5212caeb8bde829228a6b2a68b4ae53ff54.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 111, + 119, + 513, + 158.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 158.0, + 513, + 197.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 111, + 197.0, + 513, + 236.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 506, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 246, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 419, + 259 + ], + "score": 1.0, + "content": "The second inequality follows from Lemma 6. The last inequality follows from", + "type": "text" + }, + { + "bbox": [ + 419, + 248, + 426, + 257 + ], + "score": 0.79, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 246, + 506, + 259 + ], + "score": 1.0, + "content": "-Lipschitz nature of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 256, + 493, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 437, + 272 + ], + "score": 1.0, + "content": "the gradient (Assumption 1). We now prove a lemma that bounds the “drift” of the", + "type": "text" + }, + { + "bbox": [ + 437, + 257, + 454, + 271 + ], + "score": 0.92, + "content": "\\dot { \\boldsymbol { x } } _ { i , k } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 256, + 477, + 272 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 477, + 259, + 488, + 269 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 256, + 493, + 272 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 246, + 506, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 275, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 104, + 274, + 504, + 294 + ], + "spans": [ + { + "bbox": [ + 104, + 274, + 260, + 294 + ], + "score": 1.0, + "content": "Lemma 3. For any step-size satisfying", + "type": "text" + }, + { + "bbox": [ + 260, + 276, + 301, + 290 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\eta _ { l } \\le \\frac { 1 } { 8 L K } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 274, + 423, + 294 + ], + "score": 1.0, + "content": ", we can bound the drift for any", + "type": "text" + }, + { + "bbox": [ + 423, + 277, + 504, + 289 + ], + "score": 0.9, + "content": "k \\in \\{ 0 , \\cdots , K - 1 \\}", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 289, + 119, + 300 + ], + "spans": [ + { + "bbox": [ + 104, + 289, + 119, + 300 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 274, + 504, + 300 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 297, + 470, + 334 + ], + "lines": [ + { + "bbox": [ + 140, + 297, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 140, + 297, + 470, + 334 + ], + "score": 0.93, + "content": "\\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 2 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\big ] .", + "type": "interline_equation", + "image_path": "877d85de065b125fcbdd573ab793a40671501e8b9fe2d9027ba7faad2e0b960c.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 140, + 297, + 470, + 309.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 140, + 309.3333333333333, + 470, + 321.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 140, + 321.66666666666663, + 470, + 333.99999999999994 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 344, + 506, + 380 + ], + "lines": [ + { + "bbox": [ + 104, + 342, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 342, + 249, + 361 + ], + "score": 1.0, + "content": "Proof. The result trivially holds for", + "type": "text" + }, + { + "bbox": [ + 249, + 345, + 275, + 355 + ], + "score": 0.9, + "content": "k = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 342, + 298, + 361 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 298, + 345, + 337, + 358 + ], + "score": 0.95, + "content": "x _ { i , 0 } ^ { t } = x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 342, + 365, + 361 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 365, + 345, + 396, + 357 + ], + "score": 0.93, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 342, + 507, + 361 + ], + "score": 1.0, + "content": ". We now turn our attention", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 357, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 182, + 370 + ], + "score": 1.0, + "content": "to the case where", + "type": "text" + }, + { + "bbox": [ + 182, + 357, + 210, + 369 + ], + "score": 0.9, + "content": "k \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 357, + 452, + 370 + ], + "score": 1.0, + "content": ". To prove the above result, we observe that for any client", + "type": "text" + }, + { + "bbox": [ + 452, + 357, + 486, + 370 + ], + "score": 0.92, + "content": "i \\in [ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 357, + 505, + 370 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 366, + 146, + 382 + ], + "spans": [ + { + "bbox": [ + 107, + 368, + 140, + 380 + ], + "score": 0.91, + "content": "k \\in [ K ]", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 366, + 146, + 382 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 342, + 507, + 382 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 379, + 514, + 462 + ], + "lines": [ + { + "bbox": [ + 111, + 379, + 514, + 462 + ], + "spans": [ + { + "bbox": [ + 111, + 379, + 514, + 462 + ], + "score": 0.9, + "content": "\\begin{array} { r l } { { \\tilde { \\mathbf { \\mu } } _ { i , k , - } ^ { t } - x _ { t } \\| ^ { 2 } = \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } - \\eta _ { l } g _ { i , k - 1 } ^ { t } \\big \\| ^ { 2 } } } \\\\ & { \\leq \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } - \\eta _ { l } ( g _ { i , k - 1 } ^ { t } - \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) + \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) + \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) + \\nabla F _ { i } ( x _ { t } ) - \\nabla F ( x _ { t } ) \\big \\| ^ { 2 } } \\\\ & { \\leq ( 1 + \\frac { 1 } { 2 K - 1 } ) \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\big \\| ^ { 2 } + \\mathbb { E } \\big \\| \\eta _ { l } ( g _ { i , k - 1 } ^ { t } - \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) ) \\big \\| ^ { 2 } } \\\\ & { \\qquad + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) ) \\big \\| ^ { 2 } ] + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\big \\| ^ { 2 } \\big ] + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } } \\end{array}", + "type": "interline_equation", + "image_path": "b92da40aa6cb42737b33859fef28eadcd6c3cc547d30b73fe54c6f61fb504c6f.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 111, + 379, + 514, + 406.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 111, + 406.6666666666667, + 514, + 434.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 111, + 434.33333333333337, + 514, + 462.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 527, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 507, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 254, + 476 + ], + "score": 1.0, + "content": "The first inequality uses the fact that", + "type": "text" + }, + { + "bbox": [ + 255, + 462, + 281, + 475 + ], + "score": 0.93, + "content": "g _ { k - 1 , i } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 460, + 393, + 476 + ], + "score": 1.0, + "content": "is an unbiased estimator of", + "type": "text" + }, + { + "bbox": [ + 393, + 462, + 445, + 476 + ], + "score": 0.93, + "content": "\\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 460, + 507, + 476 + ], + "score": 1.0, + "content": "and Lemma 7.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 473, + 350, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 350, + 487 + ], + "score": 1.0, + "content": "The above quantity can be further bounded by the following:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 460, + 507, + 487 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 486, + 503, + 591 + ], + "lines": [ + { + "bbox": [ + 111, + 486, + 503, + 591 + ], + "spans": [ + { + "bbox": [ + 111, + 486, + 503, + 591 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { { \\mathbb { E } \\Vert x _ { i , k } ^ { t } - x _ { t } \\Vert ^ { 2 } \\leq ( 1 + \\frac { 1 } { 2 K - 1 } ) \\mathbb { E } \\Vert x _ { i , k - 1 } ^ { t } - x _ { t } \\Vert ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\Vert L ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \\Vert ^ { 2 } } } \\\\ & { } & { + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\Vert ^ { 2 } ] + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } \\nabla f ( x _ { t } ) ) ) \\Vert ^ { 2 } ] } \\\\ & { } & { = ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\mathbb { E } \\Vert ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \\Vert ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } } \\\\ & { } & { + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\Vert ^ { 2 } ] + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\Vert \\nabla f ( x _ { t } ) ) ) \\Vert ^ { 2 } ] } \\end{array}", + "type": "interline_equation", + "image_path": "62ce5a6b5255a01075f279a3e565de1e807443577d0e01b14c00e33e37b7137c.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 111, + 486, + 503, + 521.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 111, + 521.0, + 503, + 556.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 111, + 556.0, + 503, + 591.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 504, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 455, + 603 + ], + "score": 1.0, + "content": "Here, the first inequality follows from Assumption 1 and 2. Averaging over the clients", + "type": "text" + }, + { + "bbox": [ + 455, + 591, + 459, + 600 + ], + "score": 0.77, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 588, + 505, + 603 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 599, + 165, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 165, + 614 + ], + "score": 1.0, + "content": "the following:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 588, + 505, + 614 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 612, + 514, + 736 + ], + "lines": [ + { + "bbox": [ + 111, + 612, + 514, + 736 + ], + "spans": [ + { + "bbox": [ + 111, + 612, + 514, + 736 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } { { \\frac { 1 } { n } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } } } \\\\ & { } & { + \\frac { 6 K } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } [ \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\| ^ { 2 } ] + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) } \\\\ & { } & { \\leq ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\frac { 1 } { m } \\displaystyle \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K o ( \\eta _ { l } ^ { 2 } ) ) } \\\\ & { } & { + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) \\| ^ { 2 } ] } \\end{array}", + "type": "interline_equation", + "image_path": "9907b46276cfcd32c409d8f54bf5fb2a33c7b852993595a81028be61508d4d18.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 111, + 612, + 514, + 653.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 111, + 653.3333333333334, + 514, + 694.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 111, + 694.6666666666667, + 514, + 736.0000000000001 + ], + "spans": [], + "index": 28 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 302, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 303, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 303, + 97 + ], + "score": 1.0, + "content": "From the above, we get the following inequality:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 97, + 492, + 150 + ], + "lines": [ + { + "bbox": [ + 119, + 97, + 492, + 150 + ], + "spans": [ + { + "bbox": [ + 119, + 97, + 492, + 150 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { { \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq ( 1 + \\frac { 1 } { K - 1 } ) \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { } & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\Big ] } \\end{array}", + "type": "interline_equation", + "image_path": "06509a0c7208e1d1f63971a170637487f5b95ad5edfbd77444befa383de925f4.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 119, + 97, + 492, + 114.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 114.66666666666667, + 492, + 132.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 119, + 132.33333333333334, + 492, + 150.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 304, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 148, + 305, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 305, + 165 + ], + "score": 1.0, + "content": "Unrolling the recursion, we obtain the following:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 164, + 506, + 284 + ], + "lines": [ + { + "bbox": [ + 111, + 164, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 111, + 164, + 506, + 284 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\displaystyle \\frac { 1 } { n } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq \\sum _ { p = 0 } ^ { k - 1 } \\left( 1 + \\frac { 1 } { K - 1 } \\right) ^ { p } \\left[ \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\leq \\left( K - 1 \\right) \\times \\left[ \\left( 1 + \\frac { 1 } { K - 1 } \\right) ^ { K } - 1 \\right] \\times \\left[ \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\leq \\left[ 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] , } \\end{array}", + "type": "interline_equation", + "image_path": "b0cd2701990e2815bcc71e7a5558696a35f4fa2fbc4e531360177bd889510cfa.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 111, + 164, + 506, + 204.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 111, + 204.0, + 506, + 244.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 244.0, + 506, + 284.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 505, + 311 + ], + "lines": [ + { + "bbox": [ + 104, + 285, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 411, + 302 + ], + "score": 1.0, + "content": "concluding the proof of Lemma 3. The last inequality uses the fact that", + "type": "text" + }, + { + "bbox": [ + 412, + 286, + 489, + 301 + ], + "score": 0.93, + "content": "( 1 + \\frac { 1 } { K - 1 } ) ^ { K } \\leq 5", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 285, + 507, + 302 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 297, + 502, + 312 + ], + "spans": [ + { + "bbox": [ + 107, + 299, + 135, + 310 + ], + "score": 0.86, + "content": "K > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 297, + 141, + 312 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 497, + 303, + 502, + 307 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 105, + 322, + 497, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 498, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 384, + 336 + ], + "score": 1.0, + "content": "Using the above lemma in Equation 6 and Condition I, we can bound", + "type": "text" + }, + { + "bbox": [ + 385, + 324, + 396, + 334 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 322, + 498, + 336 + ], + "score": 1.0, + "content": "in the following manner:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 104, + 337, + 500, + 453 + ], + "lines": [ + { + "bbox": [ + 104, + 337, + 500, + 453 + ], + "spans": [ + { + "bbox": [ + 104, + 337, + 500, + 453 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { T _ { 3 } \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\sum _ { j = 1 } ^ { d } ( [ x _ { i , k } ^ { t } ] _ { j } - [ x _ { t } ] _ { j } ) ^ { 2 } \\right] } \\\\ & { \\quad \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } K L ^ { 2 } } { 2 \\tau } \\left[ 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\quad \\leq \\displaystyle \\frac { 3 \\eta _ { l } K } { 4 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "5e9af182310dca586e22d5d9dba22a9cad9a2456f50adb4f15c609ecd0f4e85a.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 104, + 337, + 500, + 375.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 104, + 375.6666666666667, + 500, + 414.33333333333337 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 104, + 414.33333333333337, + 500, + 453.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 504, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 217, + 471 + ], + "score": 1.0, + "content": "Here we used the fact that", + "type": "text" + }, + { + "bbox": [ + 217, + 456, + 306, + 470 + ], + "score": 0.92, + "content": "\\sqrt { v _ { t - 1 , j } } \\leq \\eta _ { l } K G \\sqrt { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 455, + 506, + 471 + ], + "score": 1.0, + "content": "and Condition I in Theorem 1. Using the above", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 466, + 223, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 223, + 482 + ], + "score": 1.0, + "content": "bound in Equation 5, we get", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 482, + 439, + 518 + ], + "lines": [ + { + "bbox": [ + 172, + 482, + 439, + 518 + ], + "spans": [ + { + "bbox": [ + 172, + 482, + 439, + 518 + ], + "score": 0.94, + "content": "T _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )", + "type": "interline_equation", + "image_path": "b6b61e73493c395d11f2595904023099798dc8f8e6f87c85393db4be641d33b0.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 172, + 482, + 439, + 494.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 172, + 494.0, + 439, + 506.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 172, + 506.0, + 439, + 518.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 526, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 504, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 374, + 541 + ], + "score": 1.0, + "content": "Putting the pieces together Substituting in Equation (3), bounds", + "type": "text" + }, + { + "bbox": [ + 374, + 527, + 386, + 538 + ], + "score": 0.88, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 524, + 492, + 541 + ], + "score": 1.0, + "content": "in Equation (8) and bound", + "type": "text" + }, + { + "bbox": [ + 493, + 529, + 504, + 538 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 537, + 214, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 214, + 550 + ], + "score": 1.0, + "content": "in Equation (4), we obtain", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 551, + 491, + 628 + ], + "lines": [ + { + "bbox": [ + 119, + 551, + 491, + 628 + ], + "spans": [ + { + "bbox": [ + 119, + 551, + 491, + 628 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \\le f ( x _ { t } ) + \\eta \\times \\left[ - \\frac { \\eta _ { l } K } { 4 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad + \\eta \\times \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\displaystyle \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] + \\frac { \\eta ^ { 2 } } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } L \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\displaystyle \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "01dbf84a6063916b2196bacb1b99de374543c8715cdbd8b7b93761144da7dae5.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 119, + 551, + 491, + 576.6666666666666 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 119, + 576.6666666666666, + 491, + 602.3333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 119, + 602.3333333333333, + 491, + 627.9999999999999 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 629, + 421, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 629, + 421, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 327, + 643 + ], + "score": 1.0, + "content": "Rearranging the above inequality and summing it from", + "type": "text" + }, + { + "bbox": [ + 327, + 630, + 351, + 640 + ], + "score": 0.88, + "content": "t = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 629, + 362, + 643 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 362, + 630, + 388, + 640 + ], + "score": 0.88, + "content": "T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 629, + 421, + 643 + ], + "score": 1.0, + "content": ", we get", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 658, + 482, + 731 + ], + "lines": [ + { + "bbox": [ + 130, + 658, + 482, + 731 + ], + "spans": [ + { + "bbox": [ + 130, + 658, + 482, + 731 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\leq \\frac { f ( x _ { 0 } ) - \\mathbb { E } [ f ( x _ { T } ) ] } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\\\ { + \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\left( \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } + \\frac { \\eta L } { 2 } \\right) \\times \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "a89d28cfd4b662cd940c97da3b5e2ab805c88b8444fe6b6dac9c9f6c5041205e.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 130, + 658, + 482, + 682.3333333333334 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 130, + 682.3333333333334, + 482, + 706.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 130, + 706.6666666666667, + 482, + 731.0000000000001 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 302, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 303, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 303, + 97 + ], + "score": 1.0, + "content": "From the above, we get the following inequality:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 80, + 303, + 97 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 97, + 492, + 150 + ], + "lines": [ + { + "bbox": [ + 119, + 97, + 492, + 150 + ], + "spans": [ + { + "bbox": [ + 119, + 97, + 492, + 150 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { { \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq ( 1 + \\frac { 1 } { K - 1 } ) \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { } & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\Big ] } \\end{array}", + "type": "interline_equation", + "image_path": "06509a0c7208e1d1f63971a170637487f5b95ad5edfbd77444befa383de925f4.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 119, + 97, + 492, + 114.66666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 114.66666666666667, + 492, + 132.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 119, + 132.33333333333334, + 492, + 150.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 304, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 148, + 305, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 305, + 165 + ], + "score": 1.0, + "content": "Unrolling the recursion, we obtain the following:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 148, + 305, + 165 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 164, + 506, + 284 + ], + "lines": [ + { + "bbox": [ + 111, + 164, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 111, + 164, + 506, + 284 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\displaystyle \\frac { 1 } { n } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq \\sum _ { p = 0 } ^ { k - 1 } \\left( 1 + \\frac { 1 } { K - 1 } \\right) ^ { p } \\left[ \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\leq \\left( K - 1 \\right) \\times \\left[ \\left( 1 + \\frac { 1 } { K - 1 } \\right) ^ { K } - 1 \\right] \\times \\left[ \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\leq \\left[ 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] , } \\end{array}", + "type": "interline_equation", + "image_path": "b0cd2701990e2815bcc71e7a5558696a35f4fa2fbc4e531360177bd889510cfa.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 111, + 164, + 506, + 204.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 111, + 204.0, + 506, + 244.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 244.0, + 506, + 284.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 505, + 311 + ], + "lines": [ + { + "bbox": [ + 104, + 285, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 411, + 302 + ], + "score": 1.0, + "content": "concluding the proof of Lemma 3. The last inequality uses the fact that", + "type": "text" + }, + { + "bbox": [ + 412, + 286, + 489, + 301 + ], + "score": 0.93, + "content": "( 1 + \\frac { 1 } { K - 1 } ) ^ { K } \\leq 5", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 285, + 507, + 302 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 297, + 502, + 312 + ], + "spans": [ + { + "bbox": [ + 107, + 299, + 135, + 310 + ], + "score": 0.86, + "content": "K > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 297, + 141, + 312 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 497, + 303, + 502, + 307 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 285, + 507, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 322, + 497, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 498, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 384, + 336 + ], + "score": 1.0, + "content": "Using the above lemma in Equation 6 and Condition I, we can bound", + "type": "text" + }, + { + "bbox": [ + 385, + 324, + 396, + 334 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 322, + 498, + 336 + ], + "score": 1.0, + "content": "in the following manner:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 322, + 498, + 336 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 104, + 337, + 500, + 453 + ], + "lines": [ + { + "bbox": [ + 104, + 337, + 500, + 453 + ], + "spans": [ + { + "bbox": [ + 104, + 337, + 500, + 453 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { T _ { 3 } \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\sum _ { j = 1 } ^ { d } ( [ x _ { i , k } ^ { t } ] _ { j } - [ x _ { t } ] _ { j } ) ^ { 2 } \\right] } \\\\ & { \\quad \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } K L ^ { 2 } } { 2 \\tau } \\left[ 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\quad \\leq \\displaystyle \\frac { 3 \\eta _ { l } K } { 4 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "5e9af182310dca586e22d5d9dba22a9cad9a2456f50adb4f15c609ecd0f4e85a.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 104, + 337, + 500, + 375.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 104, + 375.6666666666667, + 500, + 414.33333333333337 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 104, + 414.33333333333337, + 500, + 453.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 504, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 217, + 471 + ], + "score": 1.0, + "content": "Here we used the fact that", + "type": "text" + }, + { + "bbox": [ + 217, + 456, + 306, + 470 + ], + "score": 0.92, + "content": "\\sqrt { v _ { t - 1 , j } } \\leq \\eta _ { l } K G \\sqrt { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 455, + 506, + 471 + ], + "score": 1.0, + "content": "and Condition I in Theorem 1. Using the above", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 466, + 223, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 223, + 482 + ], + "score": 1.0, + "content": "bound in Equation 5, we get", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 455, + 506, + 482 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 482, + 439, + 518 + ], + "lines": [ + { + "bbox": [ + 172, + 482, + 439, + 518 + ], + "spans": [ + { + "bbox": [ + 172, + 482, + 439, + 518 + ], + "score": 0.94, + "content": "T _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )", + "type": "interline_equation", + "image_path": "b6b61e73493c395d11f2595904023099798dc8f8e6f87c85393db4be641d33b0.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 172, + 482, + 439, + 494.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 172, + 494.0, + 439, + 506.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 172, + 506.0, + 439, + 518.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 526, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 504, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 374, + 541 + ], + "score": 1.0, + "content": "Putting the pieces together Substituting in Equation (3), bounds", + "type": "text" + }, + { + "bbox": [ + 374, + 527, + 386, + 538 + ], + "score": 0.88, + "content": "T _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 524, + 492, + 541 + ], + "score": 1.0, + "content": "in Equation (8) and bound", + "type": "text" + }, + { + "bbox": [ + 493, + 529, + 504, + 538 + ], + "score": 0.88, + "content": "T _ { 2 }", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 537, + 214, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 214, + 550 + ], + "score": 1.0, + "content": "in Equation (4), we obtain", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 524, + 504, + 550 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 551, + 491, + 628 + ], + "lines": [ + { + "bbox": [ + 119, + 551, + 491, + 628 + ], + "spans": [ + { + "bbox": [ + 119, + 551, + 491, + 628 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \\le f ( x _ { t } ) + \\eta \\times \\left[ - \\frac { \\eta _ { l } K } { 4 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad + \\eta \\times \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\displaystyle \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] + \\frac { \\eta ^ { 2 } } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } L \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\displaystyle \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "01dbf84a6063916b2196bacb1b99de374543c8715cdbd8b7b93761144da7dae5.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 119, + 551, + 491, + 576.6666666666666 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 119, + 576.6666666666666, + 491, + 602.3333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 119, + 602.3333333333333, + 491, + 627.9999999999999 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 629, + 421, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 629, + 421, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 327, + 643 + ], + "score": 1.0, + "content": "Rearranging the above inequality and summing it from", + "type": "text" + }, + { + "bbox": [ + 327, + 630, + 351, + 640 + ], + "score": 0.88, + "content": "t = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 629, + 362, + 643 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 362, + 630, + 388, + 640 + ], + "score": 0.88, + "content": "T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 629, + 421, + 643 + ], + "score": 1.0, + "content": ", we get", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 629, + 421, + 643 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 130, + 658, + 482, + 731 + ], + "lines": [ + { + "bbox": [ + 130, + 658, + 482, + 731 + ], + "spans": [ + { + "bbox": [ + 130, + 658, + 482, + 731 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\leq \\frac { f ( x _ { 0 } ) - \\mathbb { E } [ f ( x _ { T } ) ] } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\\\ { + \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\left( \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } + \\frac { \\eta L } { 2 } \\right) \\times \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] } \\end{array}", + "type": "interline_equation", + "image_path": "a89d28cfd4b662cd940c97da3b5e2ab805c88b8444fe6b6dac9c9f6c5041205e.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 130, + 658, + 482, + 682.3333333333334 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 130, + 682.3333333333334, + 482, + 706.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 130, + 706.6666666666667, + 482, + 731.0000000000001 + ], + "spans": [], + "index": 27 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 505, + 97 + ], + "score": 1.0, + "content": "The first inequality uses simple telescoping sum. For completing the proof, we need the following", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 135, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 135, + 106 + ], + "score": 1.0, + "content": "result.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 426, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 428, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 428, + 128 + ], + "score": 1.0, + "content": "Lemma 4. The following upper bound holds for Algorithm 2 (FEDADAGRAD):", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 131, + 514, + 205 + ], + "lines": [ + { + "bbox": [ + 111, + 131, + 514, + 205 + ], + "spans": [ + { + "bbox": [ + 111, + 131, + 514, + 205 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\underset { t = 0 } { \\overset { T \\mathrm { - } 1 } { \\sum } } \\underset { j = 1 } { \\overset { d } { \\sum } } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\leq \\Bigg [ \\operatorname* { m i n } \\left\\{ d + \\underset { j = 1 } { \\overset { d } { \\sum } } \\log \\left( 1 + \\frac { \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) \\right. } \\\\ & { \\qquad + \\left. \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\underset { j = 1 } { \\overset { d } { \\sum } } \\sigma _ { l , j } ^ { 2 } + 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\mathbb { E } \\underset { j = 1 } { \\overset { d } { \\sum } } \\frac { ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } { \\tau ^ { 2 } } + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 2 } L ^ { 2 } } { \\tau ^ { 2 } } \\underset { t = 0 } { \\overset { T \\mathrm { - } 1 } { \\sum } } \\mathbb { E } \\left[ \\left\\| \\nabla f ( x _ { t } ) \\right\\| \\right. \\right. } \\end{array}", + "type": "interline_equation", + "image_path": "4e1147b2e7475f6e8e8fce3208937bbce8cda5b7b739aab827184761c2c3094c.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 111, + 131, + 514, + 155.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 155.66666666666666, + 514, + 180.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 111, + 180.33333333333331, + 514, + 204.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 361, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 229, + 362, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 362, + 244 + ], + "score": 1.0, + "content": "Proof. We bound the desired quantity in the following manner:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 248, + 511, + 284 + ], + "lines": [ + { + "bbox": [ + 111, + 248, + 511, + 284 + ], + "spans": [ + { + "bbox": [ + 111, + 248, + 511, + 284 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\stackrel { T \\to 1 } { \\sum } \\displaystyle \\sum _ { j = 0 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\leq d + \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } \\log \\left( 1 + \\frac { \\sum _ { l = 0 } ^ { T - 1 } \\Delta _ { l , j } ^ { 2 } } { \\tau ^ { 2 } } \\right) \\leq d + \\displaystyle \\sum _ { j = 1 } ^ { d } \\log \\left( 1 + \\frac { \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "f830567f03181e138bb5c7ca6e8d097a5604bdc2e90fba0c4ce861105aa65d08.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 111, + 248, + 511, + 260.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 260.0, + 511, + 272.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 272.0, + 511, + 284.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 349, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 289, + 349, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 349, + 303 + ], + "score": 1.0, + "content": "An alternate way of the bounding this quantity is as follows:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 307, + 425, + 415 + ], + "lines": [ + { + "bbox": [ + 185, + 307, + 425, + 415 + ], + "spans": [ + { + "bbox": [ + 185, + 307, + 425, + 415 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { t , j } ^ { 2 } + \\tau ^ { 2 } } \\leq \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\tau ^ { 2 } } } \\\\ & { \\leq \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\left\\| \\frac { \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) - \\eta _ { l } K \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } } \\\\ & { \\leq 2 \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\left[ \\left\\| \\frac { \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } + \\eta _ { l } ^ { 2 } K ^ { 2 } \\left\\| \\frac { \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "f221e4673ceef98f84efd7e8812ee80530c22604df75c2c5ec4970ef2546c0d4.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 185, + 307, + 425, + 322.42857142857144 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 185, + 322.42857142857144, + 425, + 337.8571428571429 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 185, + 337.8571428571429, + 425, + 353.28571428571433 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 185, + 353.28571428571433, + 425, + 368.7142857142858 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 185, + 368.7142857142858, + 425, + 384.1428571428572 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 185, + 384.1428571428572, + 425, + 399.57142857142867 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 185, + 399.57142857142867, + 425, + 415.0000000000001 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 395, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 396, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 396, + 432 + ], + "score": 1.0, + "content": "The first quantity in the above bound can be further bounded as follows:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 437, + 514, + 707 + ], + "lines": [ + { + "bbox": [ + 111, + 437, + 514, + 707 + ], + "spans": [ + { + "bbox": [ + 111, + 437, + 514, + 707 + ], + "score": 0.96, + "content": "\\begin{array} { r l } & { \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda - 1 } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( - \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } - \\lambda _ { i } ^ { \\lambda } \\sum _ { j = 1 } ^ { \\infty } - \\lambda _ { i } ^ { \\lambda } \\nabla F _ { i } ( \\xi _ { i } , \\xi _ { i } ) \\Bigg ) \\Bigg | ^ { 2 } } \\\\ & { - 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } \\sum _ { j = 1 } ^ { \\infty } \\lambda _ { i } ^ { \\lambda } \\sum _ { j = 1 } ^ { \\lambda } - \\lambda _ { i } ^ { \\lambda } \\nabla F _ { i } ( \\xi _ { i } , \\xi _ { i } ) + 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\lambda _ { i } ^ { \\lambda } \\xi _ { i } ^ { \\lambda } \\xi _ { i } ^ { \\lambda } \\xi _ { j } - 3 \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\Bigg ) - } \\\\ & - 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\mathbb { E } ^ { \\lambda - 1 } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } \\lambda _ { i } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb E \\end{array}", + "type": "interline_equation", + "image_path": "3bf3db4fce63d62f2c4ad22cd269923cf76161218d94f6b6c6a36975222e5d97.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 111, + 437, + 514, + 527.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 111, + 527.0, + 514, + 617.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 111, + 617.0, + 514, + 707.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 389, + 722 + ], + "score": 1.0, + "content": "Here, the first inequality follows from simple application of the fact that", + "type": "text" + }, + { + "bbox": [ + 389, + 709, + 460, + 722 + ], + "score": 0.93, + "content": "a b \\le ( a ^ { 2 } + b ^ { 2 } ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ". The result", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 719, + 504, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 143, + 733 + ], + "score": 1.0, + "content": "follows.", + "type": "text" + }, + { + "bbox": [ + 495, + 722, + 504, + 730 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 505, + 97 + ], + "score": 1.0, + "content": "The first inequality uses simple telescoping sum. For completing the proof, we need the following", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 135, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 135, + 106 + ], + "score": 1.0, + "content": "result.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 104, + 80, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 426, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 428, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 428, + 128 + ], + "score": 1.0, + "content": "Lemma 4. The following upper bound holds for Algorithm 2 (FEDADAGRAD):", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 111, + 428, + 128 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 131, + 514, + 205 + ], + "lines": [ + { + "bbox": [ + 111, + 131, + 514, + 205 + ], + "spans": [ + { + "bbox": [ + 111, + 131, + 514, + 205 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { \\underset { t = 0 } { \\overset { T \\mathrm { - } 1 } { \\sum } } \\underset { j = 1 } { \\overset { d } { \\sum } } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\leq \\Bigg [ \\operatorname* { m i n } \\left\\{ d + \\underset { j = 1 } { \\overset { d } { \\sum } } \\log \\left( 1 + \\frac { \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) \\right. } \\\\ & { \\qquad + \\left. \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\underset { j = 1 } { \\overset { d } { \\sum } } \\sigma _ { l , j } ^ { 2 } + 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\mathbb { E } \\underset { j = 1 } { \\overset { d } { \\sum } } \\frac { ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } { \\tau ^ { 2 } } + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 2 } L ^ { 2 } } { \\tau ^ { 2 } } \\underset { t = 0 } { \\overset { T \\mathrm { - } 1 } { \\sum } } \\mathbb { E } \\left[ \\left\\| \\nabla f ( x _ { t } ) \\right\\| \\right. \\right. } \\end{array}", + "type": "interline_equation", + "image_path": "4e1147b2e7475f6e8e8fce3208937bbce8cda5b7b739aab827184761c2c3094c.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 111, + 131, + 514, + 155.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 155.66666666666666, + 514, + 180.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 111, + 180.33333333333331, + 514, + 204.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 361, + 242 + ], + "lines": [ + { + "bbox": [ + 105, + 229, + 362, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 362, + 244 + ], + "score": 1.0, + "content": "Proof. We bound the desired quantity in the following manner:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 229, + 362, + 244 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 248, + 511, + 284 + ], + "lines": [ + { + "bbox": [ + 111, + 248, + 511, + 284 + ], + "spans": [ + { + "bbox": [ + 111, + 248, + 511, + 284 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\stackrel { T \\to 1 } { \\sum } \\displaystyle \\sum _ { j = 0 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\leq d + \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } \\log \\left( 1 + \\frac { \\sum _ { l = 0 } ^ { T - 1 } \\Delta _ { l , j } ^ { 2 } } { \\tau ^ { 2 } } \\right) \\leq d + \\displaystyle \\sum _ { j = 1 } ^ { d } \\log \\left( 1 + \\frac { \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "f830567f03181e138bb5c7ca6e8d097a5604bdc2e90fba0c4ce861105aa65d08.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 111, + 248, + 511, + 260.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 111, + 260.0, + 511, + 272.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 272.0, + 511, + 284.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 349, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 289, + 349, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 349, + 303 + ], + "score": 1.0, + "content": "An alternate way of the bounding this quantity is as follows:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 289, + 349, + 303 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 307, + 425, + 415 + ], + "lines": [ + { + "bbox": [ + 185, + 307, + 425, + 415 + ], + "spans": [ + { + "bbox": [ + 185, + 307, + 425, + 415 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { t , j } ^ { 2 } + \\tau ^ { 2 } } \\leq \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\tau ^ { 2 } } } \\\\ & { \\leq \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\left\\| \\frac { \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) - \\eta _ { l } K \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } } \\\\ & { \\leq 2 \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\left[ \\left\\| \\frac { \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } + \\eta _ { l } ^ { 2 } K ^ { 2 } \\left\\| \\frac { \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "f221e4673ceef98f84efd7e8812ee80530c22604df75c2c5ec4970ef2546c0d4.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 185, + 307, + 425, + 322.42857142857144 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 185, + 322.42857142857144, + 425, + 337.8571428571429 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 185, + 337.8571428571429, + 425, + 353.28571428571433 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 185, + 353.28571428571433, + 425, + 368.7142857142858 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 185, + 368.7142857142858, + 425, + 384.1428571428572 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 185, + 384.1428571428572, + 425, + 399.57142857142867 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 185, + 399.57142857142867, + 425, + 415.0000000000001 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 395, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 396, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 396, + 432 + ], + "score": 1.0, + "content": "The first quantity in the above bound can be further bounded as follows:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 418, + 396, + 432 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 437, + 514, + 707 + ], + "lines": [ + { + "bbox": [ + 111, + 437, + 514, + 707 + ], + "spans": [ + { + "bbox": [ + 111, + 437, + 514, + 707 + ], + "score": 0.96, + "content": "\\begin{array} { r l } & { \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda - 1 } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( - \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } - \\lambda _ { i } ^ { \\lambda } \\sum _ { j = 1 } ^ { \\infty } - \\lambda _ { i } ^ { \\lambda } \\nabla F _ { i } ( \\xi _ { i } , \\xi _ { i } ) \\Bigg ) \\Bigg | ^ { 2 } } \\\\ & { - 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } \\sum _ { j = 1 } ^ { \\infty } \\lambda _ { i } ^ { \\lambda } \\sum _ { j = 1 } ^ { \\lambda } - \\lambda _ { i } ^ { \\lambda } \\nabla F _ { i } ( \\xi _ { i } , \\xi _ { i } ) + 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\lambda _ { i } ^ { \\lambda } \\xi _ { i } ^ { \\lambda } \\xi _ { i } ^ { \\lambda } \\xi _ { j } - 3 \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\Bigg ) - } \\\\ & - 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\mathbb { E } ^ { \\lambda - 1 } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } \\lambda _ { i } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb E \\end{array}", + "type": "interline_equation", + "image_path": "3bf3db4fce63d62f2c4ad22cd269923cf76161218d94f6b6c6a36975222e5d97.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 111, + 437, + 514, + 527.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 111, + 527.0, + 514, + 617.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 111, + 617.0, + 514, + 707.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 389, + 722 + ], + "score": 1.0, + "content": "Here, the first inequality follows from simple application of the fact that", + "type": "text" + }, + { + "bbox": [ + 389, + 709, + 460, + 722 + ], + "score": 0.93, + "content": "a b \\le ( a ^ { 2 } + b ^ { 2 } ) / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ". The result", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 719, + 504, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 143, + 733 + ], + "score": 1.0, + "content": "follows.", + "type": "text" + }, + { + "bbox": [ + 495, + 722, + 504, + 730 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 339, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 339, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 339, + 95 + ], + "score": 1.0, + "content": "Substituting the above bound in Equation (10), we obtain:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 97, + 514, + 243 + ], + "lines": [ + { + "bbox": [ + 111, + 97, + 514, + 243 + ], + "spans": [ + { + "bbox": [ + 111, + 97, + 514, + 243 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\frac { \\displaystyle { n K } } { \\displaystyle { 4 } } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\pi _ { t - 1 , j } } + \\tau } } \\\\ & { \\leq \\frac { f \\left( x _ { 0 } \\right) - \\mathbb { E } \\left[ f ( x _ { T } ) \\right] } { \\eta } + \\frac { 5 \\eta _ { t } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\\\ & { \\qquad + \\displaystyle { \\sum _ { j = 1 } ^ { d } \\left( \\frac { \\eta _ { t } K G ^ { 2 } } { \\tau } + \\frac { \\eta L } { 2 } \\right) \\times \\left[ \\operatorname* { m i n } \\left\\{ d + d \\log \\left( 1 + \\frac { \\eta _ { t } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) + \\right. \\right.} } \\\\ & { \\quad \\quad \\left. \\frac { 4 \\eta _ { t } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + 2 0 \\eta _ { t } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\sum _ { j = 1 } ^ { d } \\frac { ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } { \\tau ^ { 2 } } + \\frac { 4 0 \\eta _ { t } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } \\left[ \\left\\| \\nabla f ( x _ { t } ) \\right\\| ^ { 2 } \\right] \\right\\} } \\end{array}", + "type": "interline_equation", + "image_path": "b9883d2eb2dc53244494a52fbe203284635df98d004ce0d80f6c1d8722c8ec7d.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 111, + 97, + 514, + 145.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 145.66666666666666, + 514, + 194.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 194.33333333333331, + 514, + 242.99999999999997 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 262, + 213, + 273 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 213, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 213, + 276 + ], + "score": 1.0, + "content": "We observe the following:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 123, + 276, + 486, + 312 + ], + "lines": [ + { + "bbox": [ + 123, + 276, + 486, + 312 + ], + "spans": [ + { + "bbox": [ + 123, + 276, + 486, + 312 + ], + "score": 0.94, + "content": "\\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla \\mathbb { E } f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\geq \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\eta _ { t } K G \\sqrt { T } + \\tau } \\geq \\frac { T } { \\eta _ { t } K G \\sqrt { T } + \\tau } \\operatorname* { m i n } _ { 0 \\leq t \\leq T } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } .", + "type": "interline_equation", + "image_path": "32ab3f3d2fdf07a6c34f4f486222c9477f6fadd36c17f09624fb4b6f3a41c7b6.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 123, + 276, + 486, + 288.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 123, + 288.0, + 486, + 300.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 123, + 300.0, + 486, + 312.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "The second part of Theorem 1 follows from using the above inequality in Equation (12). Note that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 324, + 504, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 358, + 337 + ], + "score": 1.0, + "content": "the first part of Theorem 1 is obtain from the part of Lemma 4.", + "type": "text" + }, + { + "bbox": [ + 496, + 327, + 504, + 335 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 348, + 247, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 248, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 248, + 361 + ], + "score": 1.0, + "content": "A.2.1 LIMITED PARTICIPATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 367, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "For limited participation, the main changed in the proof is in Equation (11). The rest of the proof is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 325, + 390 + ], + "score": 1.0, + "content": "similar so we mainly focus on Equation (11) here. Let", + "type": "text" + }, + { + "bbox": [ + 326, + 379, + 334, + 388 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 378, + 436, + 390 + ], + "score": 1.0, + "content": "be the sampled set at the", + "type": "text" + }, + { + "bbox": [ + 436, + 378, + 447, + 388 + ], + "score": 0.87, + "content": "t ^ { \\mathrm { { t h } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 378, + 506, + 390 + ], + "score": 1.0, + "content": "iteration such", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 124, + 401 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 389, + 155, + 401 + ], + "score": 0.89, + "content": "| S | = s", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 388, + 343, + 401 + ], + "score": 1.0, + "content": ". In partial participation, we assume that the set", + "type": "text" + }, + { + "bbox": [ + 344, + 389, + 352, + 399 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "is sampled uniformly from all subsets", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 400, + 405, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 118, + 411 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 401, + 132, + 412 + ], + "score": 0.88, + "content": "[ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 400, + 172, + 411 + ], + "score": 1.0, + "content": "with size", + "type": "text" + }, + { + "bbox": [ + 173, + 402, + 178, + 410 + ], + "score": 0.64, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 400, + 405, + 411 + ], + "score": 1.0, + "content": ". In this case, for the first term in Equation (11), we have", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 415, + 514, + 699 + ], + "lines": [ + { + "bbox": [ + 111, + 415, + 514, + 699 + ], + "spans": [ + { + "bbox": [ + 111, + 415, + 514, + 699 + ], + "score": 0.96, + "content": "\\begin{array} { r l } { \\sum _ { j = 1 } ^ { N } \\Bigg [ \\Bigg | \\int _ { - 1 } ^ { 1 } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 0 } ^ { 1 } \\sum _ { i = 0 } ^ { N } \\sum _ { j = 1 } ^ { N } \\Bigg ) \\Bigg | ^ { 2 } \\Bigg ] ^ { 2 } } & { \\mathrm { d } \\Gamma } \\\\ { = } & { \\mathcal { D } _ { j } ^ { \\mathrm { R } } \\Bigg [ \\Bigg | \\int _ { - 1 } ^ { 1 } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 0 } ^ { 1 } \\sum _ { i = 0 } ^ { N } \\sum _ { j = 1 } ^ { N } \\delta ( \\sum _ { i = 0 } ^ { N } \\epsilon _ { j i } ) + \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\bigg ) \\Bigg | ^ { 2 } } \\\\ & { \\leq \\epsilon _ { j } ^ { \\mathrm { R } } \\Bigg ( \\int _ { - 1 } ^ { 1 } \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) ^ { 2 } \\Bigg [ \\Bigg | \\int _ { 0 } ^ { 1 } \\sum _ { j = 1 } ^ { N } \\delta ( \\sum _ { j = 1 } ^ { N } \\epsilon _ { j i } ) - \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\Bigg | ^ { 2 } \\Bigg ] ^ { 2 } + \\Bigg | \\int _ { - 1 } ^ { 1 } \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) - \\mathrm { v } _ { \\mathrm { R } } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) \\Bigg | ^ { 2 } } \\\\ & \\qquad + \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\Bigg | ^ { 2 } \\Bigg ] + \\Bigg | \\int _ { 0 } ^ { 1 } \\int _ { - 1 } ^ { 1 } \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 1 } ^ { N } \\mathrm { d } \\ \\end{array}", + "type": "interline_equation", + "image_path": "b799b7729cf8d3e983e7798f43c9e68895804b5af4937a7a71e8edf4df8dbd51.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 111, + 415, + 514, + 509.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 111, + 509.6666666666667, + 514, + 604.3333333333334 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 111, + 604.3333333333334, + 514, + 699.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 289, + 722 + ], + "score": 1.0, + "content": "Note that the expectation here also includes", + "type": "text" + }, + { + "bbox": [ + 290, + 711, + 297, + 720 + ], + "score": 0.78, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ". The first inequality is obtained from the fact that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 228, + 733 + ], + "score": 0.91, + "content": "( a + b + c ) ^ { 2 } \\leq 3 { \\dot { ( } } a ^ { 2 } + b ^ { 2 } + c ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 720, + 446, + 733 + ], + "score": 1.0, + "content": ". The second inequality is obtained from the fact that set", + "type": "text" + }, + { + "bbox": [ + 446, + 721, + 454, + 730 + ], + "score": 0.83, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "is uniformly", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 339, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 339, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 339, + 95 + ], + "score": 1.0, + "content": "Substituting the above bound in Equation (10), we obtain:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 82, + 339, + 95 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 97, + 514, + 243 + ], + "lines": [ + { + "bbox": [ + 111, + 97, + 514, + 243 + ], + "spans": [ + { + "bbox": [ + 111, + 97, + 514, + 243 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\frac { \\displaystyle { n K } } { \\displaystyle { 4 } } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\pi _ { t - 1 , j } } + \\tau } } \\\\ & { \\leq \\frac { f \\left( x _ { 0 } \\right) - \\mathbb { E } \\left[ f ( x _ { T } ) \\right] } { \\eta } + \\frac { 5 \\eta _ { t } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\\\ & { \\qquad + \\displaystyle { \\sum _ { j = 1 } ^ { d } \\left( \\frac { \\eta _ { t } K G ^ { 2 } } { \\tau } + \\frac { \\eta L } { 2 } \\right) \\times \\left[ \\operatorname* { m i n } \\left\\{ d + d \\log \\left( 1 + \\frac { \\eta _ { t } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) + \\right. \\right.} } \\\\ & { \\quad \\quad \\left. \\frac { 4 \\eta _ { t } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + 2 0 \\eta _ { t } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\sum _ { j = 1 } ^ { d } \\frac { ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } { \\tau ^ { 2 } } + \\frac { 4 0 \\eta _ { t } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } \\left[ \\left\\| \\nabla f ( x _ { t } ) \\right\\| ^ { 2 } \\right] \\right\\} } \\end{array}", + "type": "interline_equation", + "image_path": "b9883d2eb2dc53244494a52fbe203284635df98d004ce0d80f6c1d8722c8ec7d.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 111, + 97, + 514, + 145.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 145.66666666666666, + 514, + 194.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 194.33333333333331, + 514, + 242.99999999999997 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 262, + 213, + 273 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 213, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 213, + 276 + ], + "score": 1.0, + "content": "We observe the following:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 259, + 213, + 276 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 123, + 276, + 486, + 312 + ], + "lines": [ + { + "bbox": [ + 123, + 276, + 486, + 312 + ], + "spans": [ + { + "bbox": [ + 123, + 276, + 486, + 312 + ], + "score": 0.94, + "content": "\\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla \\mathbb { E } f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\geq \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\eta _ { t } K G \\sqrt { T } + \\tau } \\geq \\frac { T } { \\eta _ { t } K G \\sqrt { T } + \\tau } \\operatorname* { m i n } _ { 0 \\leq t \\leq T } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } .", + "type": "interline_equation", + "image_path": "32ab3f3d2fdf07a6c34f4f486222c9477f6fadd36c17f09624fb4b6f3a41c7b6.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 123, + 276, + 486, + 288.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 123, + 288.0, + 486, + 300.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 123, + 300.0, + 486, + 312.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "The second part of Theorem 1 follows from using the above inequality in Equation (12). Note that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 324, + 504, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 358, + 337 + ], + "score": 1.0, + "content": "the first part of Theorem 1 is obtain from the part of Lemma 4.", + "type": "text" + }, + { + "bbox": [ + 496, + 327, + 504, + 335 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 313, + 506, + 337 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 348, + 247, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 248, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 248, + 361 + ], + "score": 1.0, + "content": "A.2.1 LIMITED PARTICIPATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 367, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "For limited participation, the main changed in the proof is in Equation (11). The rest of the proof is", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 325, + 390 + ], + "score": 1.0, + "content": "similar so we mainly focus on Equation (11) here. Let", + "type": "text" + }, + { + "bbox": [ + 326, + 379, + 334, + 388 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 378, + 436, + 390 + ], + "score": 1.0, + "content": "be the sampled set at the", + "type": "text" + }, + { + "bbox": [ + 436, + 378, + 447, + 388 + ], + "score": 0.87, + "content": "t ^ { \\mathrm { { t h } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 378, + 506, + 390 + ], + "score": 1.0, + "content": "iteration such", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 124, + 401 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 124, + 389, + 155, + 401 + ], + "score": 0.89, + "content": "| S | = s", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 388, + 343, + 401 + ], + "score": 1.0, + "content": ". In partial participation, we assume that the set", + "type": "text" + }, + { + "bbox": [ + 344, + 389, + 352, + 399 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "is sampled uniformly from all subsets", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 400, + 405, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 118, + 411 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 401, + 132, + 412 + ], + "score": 0.88, + "content": "[ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 400, + 172, + 411 + ], + "score": 1.0, + "content": "with size", + "type": "text" + }, + { + "bbox": [ + 173, + 402, + 178, + 410 + ], + "score": 0.64, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 400, + 405, + 411 + ], + "score": 1.0, + "content": ". In this case, for the first term in Equation (11), we have", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 367, + 506, + 412 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 415, + 514, + 699 + ], + "lines": [ + { + "bbox": [ + 111, + 415, + 514, + 699 + ], + "spans": [ + { + "bbox": [ + 111, + 415, + 514, + 699 + ], + "score": 0.96, + "content": "\\begin{array} { r l } { \\sum _ { j = 1 } ^ { N } \\Bigg [ \\Bigg | \\int _ { - 1 } ^ { 1 } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 0 } ^ { 1 } \\sum _ { i = 0 } ^ { N } \\sum _ { j = 1 } ^ { N } \\Bigg ) \\Bigg | ^ { 2 } \\Bigg ] ^ { 2 } } & { \\mathrm { d } \\Gamma } \\\\ { = } & { \\mathcal { D } _ { j } ^ { \\mathrm { R } } \\Bigg [ \\Bigg | \\int _ { - 1 } ^ { 1 } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 0 } ^ { 1 } \\sum _ { i = 0 } ^ { N } \\sum _ { j = 1 } ^ { N } \\delta ( \\sum _ { i = 0 } ^ { N } \\epsilon _ { j i } ) + \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\bigg ) \\Bigg | ^ { 2 } } \\\\ & { \\leq \\epsilon _ { j } ^ { \\mathrm { R } } \\Bigg ( \\int _ { - 1 } ^ { 1 } \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) ^ { 2 } \\Bigg [ \\Bigg | \\int _ { 0 } ^ { 1 } \\sum _ { j = 1 } ^ { N } \\delta ( \\sum _ { j = 1 } ^ { N } \\epsilon _ { j i } ) - \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\Bigg | ^ { 2 } \\Bigg ] ^ { 2 } + \\Bigg | \\int _ { - 1 } ^ { 1 } \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) - \\mathrm { v } _ { \\mathrm { R } } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) \\Bigg | ^ { 2 } } \\\\ & \\qquad + \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\Bigg | ^ { 2 } \\Bigg ] + \\Bigg | \\int _ { 0 } ^ { 1 } \\int _ { - 1 } ^ { 1 } \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 1 } ^ { N } \\mathrm { d } \\ \\end{array}", + "type": "interline_equation", + "image_path": "b799b7729cf8d3e983e7798f43c9e68895804b5af4937a7a71e8edf4df8dbd51.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 111, + 415, + 514, + 509.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 111, + 509.6666666666667, + 514, + 604.3333333333334 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 111, + 604.3333333333334, + 514, + 699.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 289, + 722 + ], + "score": 1.0, + "content": "Note that the expectation here also includes", + "type": "text" + }, + { + "bbox": [ + 290, + 711, + 297, + 720 + ], + "score": 0.78, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 709, + 505, + 722 + ], + "score": 1.0, + "content": ". The first inequality is obtained from the fact that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 228, + 733 + ], + "score": 0.91, + "content": "( a + b + c ) ^ { 2 } \\leq 3 { \\dot { ( } } a ^ { 2 } + b ^ { 2 } + c ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 720, + 446, + 733 + ], + "score": 1.0, + "content": ". The second inequality is obtained from the fact that set", + "type": "text" + }, + { + "bbox": [ + 446, + 721, + 454, + 730 + ], + "score": 0.83, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "is uniformly", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 225, + 95 + ], + "score": 1.0, + "content": "sampled from all subsets of", + "type": "text" + }, + { + "bbox": [ + 225, + 82, + 240, + 94 + ], + "score": 0.88, + "content": "[ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 81, + 282, + 95 + ], + "score": 1.0, + "content": "with size", + "type": "text" + }, + { + "bbox": [ + 283, + 84, + 289, + 92 + ], + "score": 0.6, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 81, + 505, + 95 + ], + "score": 1.0, + "content": ". The third and fourth inequalities are similar to the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "one used in proof of Theorem 1. Substituting the above bound in Equation (10) gives the desired", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 179, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 179, + 116 + ], + "score": 1.0, + "content": "convergence rate.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 134, + 231, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 134, + 232, + 147 + ], + "spans": [ + { + "bbox": [ + 106, + 134, + 232, + 147 + ], + "score": 1.0, + "content": "A.3 PROOF OF THEOREM 2", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 157, + 505, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 157, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 505, + 171 + ], + "score": 1.0, + "content": "Proof of Theorem 2. The proof strategy is similar to that of FEDADAGRAD except that we need to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 167, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 181 + ], + "score": 1.0, + "content": "handle the exponential moving average in FEDADAM. We note that the update of FEDADAM is the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 178, + 149, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 149, + 193 + ], + "score": 1.0, + "content": "following", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 255, + 196, + 356, + 225 + ], + "lines": [ + { + "bbox": [ + 255, + 196, + 356, + 225 + ], + "spans": [ + { + "bbox": [ + 255, + 196, + 356, + 225 + ], + "score": 0.94, + "content": "x _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } ,", + "type": "interline_equation", + "image_path": "4dfc7f601b27403d8437017dac13006da4bff9de10fc9f0683eb9bea340e8c9d.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 255, + 196, + 356, + 210.5 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 255, + 210.5, + 356, + 225.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 135, + 248 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 135, + 235, + 164, + 247 + ], + "score": 0.94, + "content": "i \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 234, + 212, + 248 + ], + "score": 1.0, + "content": ". Using the", + "type": "text" + }, + { + "bbox": [ + 212, + 236, + 220, + 245 + ], + "score": 0.81, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 234, + 332, + 248 + ], + "score": 1.0, + "content": "-smooth nature of function", + "type": "text" + }, + { + "bbox": [ + 332, + 236, + 339, + 247 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "and the above update rule, we have the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 245, + 150, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 150, + 259 + ], + "score": 1.0, + "content": "following:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 268, + 446, + 303 + ], + "lines": [ + { + "bbox": [ + 164, + 268, + 446, + 303 + ], + "spans": [ + { + "bbox": [ + 164, + 268, + 446, + 303 + ], + "score": 0.93, + "content": "f ( \\boldsymbol { x } _ { t + 1 } ) \\leq f ( \\boldsymbol { x } _ { t } ) + \\eta \\left. \\nabla f ( \\boldsymbol { x } _ { t } ) , \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\sum _ { i = 1 } ^ { d } \\frac { \\Delta _ { t , i } ^ { 2 } } { ( \\sqrt { v _ { t , i } } + \\tau ) ^ { 2 } }", + "type": "interline_equation", + "image_path": "2e104f1e0635f188755318d4665c1d9be0fa34468dd053e57fbec0d4ba2f6d96.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 164, + 268, + 446, + 279.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 164, + 279.6666666666667, + 446, + 291.33333333333337 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 164, + 291.33333333333337, + 446, + 303.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 447, + 328 + ], + "score": 1.0, + "content": "The second step follows simply from FEDADAM’s update. We take the expectation of", + "type": "text" + }, + { + "bbox": [ + 447, + 313, + 480, + 326 + ], + "score": 0.93, + "content": "f ( x _ { t + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 311, + 506, + 328 + ], + "score": 1.0, + "content": "(over", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 324, + 358, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 205, + 337 + ], + "score": 1.0, + "content": "randomness at time step", + "type": "text" + }, + { + "bbox": [ + 205, + 326, + 210, + 334 + ], + "score": 0.62, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 324, + 358, + 337 + ], + "score": 1.0, + "content": ") and rewrite the above inequality as:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 345, + 514, + 469 + ], + "lines": [ + { + "bbox": [ + 111, + 345, + 514, + 469 + ], + "spans": [ + { + "bbox": [ + 111, + 345, + 514, + 469 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\bar { c } _ { t } [ f ( x _ { t + 1 } ) ] \\leq f ( x _ { t } ) + \\eta \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } + \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } } \\\\ { = f ( x _ { t } ) + \\eta \\underbrace { \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\right. } _ { R _ { 1 } } + \\eta \\underbrace { \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } } \\right. \\right. } _ { R _ { 2 } } } \\\\ { \\left. \\left. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( \\sqrt { v _ { t , j } } + \\tau ) ^ { 2 } } \\right] \\right. \\right. } \\end{array}", + "type": "interline_equation", + "image_path": "24d642061b4389eae21ff08c5e8289d958fc3329cfb685492876b7902596d324.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 111, + 345, + 514, + 386.3333333333333 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 111, + 386.3333333333333, + 514, + 427.66666666666663 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 111, + 427.66666666666663, + 514, + 468.99999999999994 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 321, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 317, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 150, + 510 + ], + "score": 1.0, + "content": "Bounding", + "type": "text" + }, + { + "bbox": [ + 151, + 496, + 163, + 506 + ], + "score": 0.86, + "content": "R _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 492, + 304, + 510 + ], + "score": 1.0, + "content": ". We observe the following about", + "type": "text" + }, + { + "bbox": [ + 304, + 496, + 317, + 506 + ], + "score": 0.89, + "content": "R _ { 2 }", + "type": "inline_equation" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 100, + 516, + 500, + 740 + ], + "lines": [ + { + "bbox": [ + 100, + 516, + 500, + 740 + ], + "spans": [ + { + "bbox": [ + 100, + 516, + 500, + 740 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { H _ { 2 } = \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\times \\left[ \\frac { \\Delta _ { t , x } } { \\sqrt { \\nabla f _ { \\mathbf { S } } \\phi _ { t + 1 } } } - \\frac { \\Delta _ { t , x } } { \\sqrt { \\nabla f _ { \\mathbf { S } } \\phi _ { t - 1 , \\Delta } } + \\tau } \\right] } \\\\ & { \\quad = \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\Delta _ { t , x } \\simeq \\Delta _ { t , \\Phi } \\times \\left[ \\frac { \\sqrt { \\omega _ { 2 } \\phi _ { 1 } } \\mathbf { e } _ { 1 } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } } } { \\left( \\sqrt { \\omega _ { 2 } } \\omega _ { 1 } \\right) \\tau \\left( \\sqrt { \\hat { \\omega } _ { 2 } \\omega _ { 1 } } - \\Delta _ { t , 1 } \\right) \\tau } \\right] } \\\\ & { \\quad \\quad - \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\times \\Delta _ { t , \\Phi } \\times \\left[ \\frac { - ( 1 - \\omega _ { 2 } ) \\Delta _ { t , x } ^ { 2 } } { \\left( \\sqrt { \\omega _ { 2 } } \\omega _ { 1 } + \\gamma \\right)} \\left( \\sqrt { \\hat { \\omega } _ { 2 } \\omega _ { 1 } } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } } } \\right) + \\sqrt { \\mathbf { e } _ { 1 } \\omega _ { 2 } } \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & \\quad \\quad \\quad \\quad \\quad \\quad \\end{array}", + "type": "interline_equation", + "image_path": "14caa1b90349bf8677c43b103ed85506a24aeb712af8d7eb4efb7b30522c60f0.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 100, + 516, + 500, + 590.6666666666666 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 100, + 590.6666666666666, + 500, + 665.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 100, + 665.3333333333333, + 500, + 739.9999999999999 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 81, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 225, + 95 + ], + "score": 1.0, + "content": "sampled from all subsets of", + "type": "text" + }, + { + "bbox": [ + 225, + 82, + 240, + 94 + ], + "score": 0.88, + "content": "[ m ]", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 81, + 282, + 95 + ], + "score": 1.0, + "content": "with size", + "type": "text" + }, + { + "bbox": [ + 283, + 84, + 289, + 92 + ], + "score": 0.6, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 81, + 505, + 95 + ], + "score": 1.0, + "content": ". The third and fourth inequalities are similar to the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "one used in proof of Theorem 1. Substituting the above bound in Equation (10) gives the desired", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 179, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 179, + 116 + ], + "score": 1.0, + "content": "convergence rate.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 116 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 134, + 231, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 134, + 232, + 147 + ], + "spans": [ + { + "bbox": [ + 106, + 134, + 232, + 147 + ], + "score": 1.0, + "content": "A.3 PROOF OF THEOREM 2", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 157, + 505, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 157, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 505, + 171 + ], + "score": 1.0, + "content": "Proof of Theorem 2. The proof strategy is similar to that of FEDADAGRAD except that we need to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 167, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 181 + ], + "score": 1.0, + "content": "handle the exponential moving average in FEDADAM. We note that the update of FEDADAM is the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 178, + 149, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 149, + 193 + ], + "score": 1.0, + "content": "following", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 157, + 505, + 193 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 255, + 196, + 356, + 225 + ], + "lines": [ + { + "bbox": [ + 255, + 196, + 356, + 225 + ], + "spans": [ + { + "bbox": [ + 255, + 196, + 356, + 225 + ], + "score": 0.94, + "content": "x _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } ,", + "type": "interline_equation", + "image_path": "4dfc7f601b27403d8437017dac13006da4bff9de10fc9f0683eb9bea340e8c9d.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 255, + 196, + 356, + 210.5 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 255, + 210.5, + 356, + 225.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 135, + 248 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 135, + 235, + 164, + 247 + ], + "score": 0.94, + "content": "i \\in [ d ]", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 234, + 212, + 248 + ], + "score": 1.0, + "content": ". Using the", + "type": "text" + }, + { + "bbox": [ + 212, + 236, + 220, + 245 + ], + "score": 0.81, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 234, + 332, + 248 + ], + "score": 1.0, + "content": "-smooth nature of function", + "type": "text" + }, + { + "bbox": [ + 332, + 236, + 339, + 247 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 234, + 505, + 248 + ], + "score": 1.0, + "content": "and the above update rule, we have the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 245, + 150, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 150, + 259 + ], + "score": 1.0, + "content": "following:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 234, + 505, + 259 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 268, + 446, + 303 + ], + "lines": [ + { + "bbox": [ + 164, + 268, + 446, + 303 + ], + "spans": [ + { + "bbox": [ + 164, + 268, + 446, + 303 + ], + "score": 0.93, + "content": "f ( \\boldsymbol { x } _ { t + 1 } ) \\leq f ( \\boldsymbol { x } _ { t } ) + \\eta \\left. \\nabla f ( \\boldsymbol { x } _ { t } ) , \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\sum _ { i = 1 } ^ { d } \\frac { \\Delta _ { t , i } ^ { 2 } } { ( \\sqrt { v _ { t , i } } + \\tau ) ^ { 2 } }", + "type": "interline_equation", + "image_path": "2e104f1e0635f188755318d4665c1d9be0fa34468dd053e57fbec0d4ba2f6d96.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 164, + 268, + 446, + 279.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 164, + 279.6666666666667, + 446, + 291.33333333333337 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 164, + 291.33333333333337, + 446, + 303.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 504, + 336 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 447, + 328 + ], + "score": 1.0, + "content": "The second step follows simply from FEDADAM’s update. We take the expectation of", + "type": "text" + }, + { + "bbox": [ + 447, + 313, + 480, + 326 + ], + "score": 0.93, + "content": "f ( x _ { t + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 311, + 506, + 328 + ], + "score": 1.0, + "content": "(over", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 324, + 358, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 205, + 337 + ], + "score": 1.0, + "content": "randomness at time step", + "type": "text" + }, + { + "bbox": [ + 205, + 326, + 210, + 334 + ], + "score": 0.62, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 324, + 358, + 337 + ], + "score": 1.0, + "content": ") and rewrite the above inequality as:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 311, + 506, + 337 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 345, + 514, + 469 + ], + "lines": [ + { + "bbox": [ + 111, + 345, + 514, + 469 + ], + "spans": [ + { + "bbox": [ + 111, + 345, + 514, + 469 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\bar { c } _ { t } [ f ( x _ { t + 1 } ) ] \\leq f ( x _ { t } ) + \\eta \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } + \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } } \\\\ { = f ( x _ { t } ) + \\eta \\underbrace { \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\right. } _ { R _ { 1 } } + \\eta \\underbrace { \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } } \\right. \\right. } _ { R _ { 2 } } } \\\\ { \\left. \\left. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( \\sqrt { v _ { t , j } } + \\tau ) ^ { 2 } } \\right] \\right. \\right. } \\end{array}", + "type": "interline_equation", + "image_path": "24d642061b4389eae21ff08c5e8289d958fc3329cfb685492876b7902596d324.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 111, + 345, + 514, + 386.3333333333333 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 111, + 386.3333333333333, + 514, + 427.66666666666663 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 111, + 427.66666666666663, + 514, + 468.99999999999994 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 321, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 317, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 150, + 510 + ], + "score": 1.0, + "content": "Bounding", + "type": "text" + }, + { + "bbox": [ + 151, + 496, + 163, + 506 + ], + "score": 0.86, + "content": "R _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 492, + 304, + 510 + ], + "score": 1.0, + "content": ". We observe the following about", + "type": "text" + }, + { + "bbox": [ + 304, + 496, + 317, + 506 + ], + "score": 0.89, + "content": "R _ { 2 }", + "type": "inline_equation" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 492, + 317, + 510 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 100, + 516, + 500, + 740 + ], + "lines": [ + { + "bbox": [ + 100, + 516, + 500, + 740 + ], + "spans": [ + { + "bbox": [ + 100, + 516, + 500, + 740 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { H _ { 2 } = \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\times \\left[ \\frac { \\Delta _ { t , x } } { \\sqrt { \\nabla f _ { \\mathbf { S } } \\phi _ { t + 1 } } } - \\frac { \\Delta _ { t , x } } { \\sqrt { \\nabla f _ { \\mathbf { S } } \\phi _ { t - 1 , \\Delta } } + \\tau } \\right] } \\\\ & { \\quad = \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\Delta _ { t , x } \\simeq \\Delta _ { t , \\Phi } \\times \\left[ \\frac { \\sqrt { \\omega _ { 2 } \\phi _ { 1 } } \\mathbf { e } _ { 1 } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } } } { \\left( \\sqrt { \\omega _ { 2 } } \\omega _ { 1 } \\right) \\tau \\left( \\sqrt { \\hat { \\omega } _ { 2 } \\omega _ { 1 } } - \\Delta _ { t , 1 } \\right) \\tau } \\right] } \\\\ & { \\quad \\quad - \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\times \\Delta _ { t , \\Phi } \\times \\left[ \\frac { - ( 1 - \\omega _ { 2 } ) \\Delta _ { t , x } ^ { 2 } } { \\left( \\sqrt { \\omega _ { 2 } } \\omega _ { 1 } + \\gamma \\right)} \\left( \\sqrt { \\hat { \\omega } _ { 2 } \\omega _ { 1 } } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } } } \\right) + \\sqrt { \\mathbf { e } _ { 1 } \\omega _ { 2 } } \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & \\quad \\quad \\quad \\quad \\quad \\quad \\end{array}", + "type": "interline_equation", + "image_path": "14caa1b90349bf8677c43b103ed85506a24aeb712af8d7eb4efb7b30522c60f0.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 100, + 516, + 500, + 590.6666666666666 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 100, + 590.6666666666666, + 500, + 665.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 100, + 665.3333333333333, + 500, + 739.9999999999999 + ], + "spans": [], + "index": 22 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 339, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 340, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 150, + 96 + ], + "score": 1.0, + "content": "Bounding", + "type": "text" + }, + { + "bbox": [ + 151, + 83, + 163, + 93 + ], + "score": 0.87, + "content": "R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 81, + 214, + 96 + ], + "score": 1.0, + "content": ". The term", + "type": "text" + }, + { + "bbox": [ + 214, + 83, + 227, + 93 + ], + "score": 0.88, + "content": "R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 81, + 340, + 96 + ], + "score": 1.0, + "content": "can be bounded as follows:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 95, + 467, + 213 + ], + "lines": [ + { + "bbox": [ + 144, + 95, + 467, + 213 + ], + "spans": [ + { + "bbox": [ + 144, + 95, + 467, + 213 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { R _ { 1 } = \\Bigg \\langle \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\Bigg \\rangle } \\\\ & { \\quad = \\Bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } - \\eta _ { t } K \\nabla f ( x _ { t } ) + \\eta _ { t } K \\nabla f ( x _ { t } ) \\right] \\Bigg \\rangle } \\\\ & { \\quad = - \\eta _ { t } K \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\underbrace { \\Bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { t } K \\nabla f ( x _ { t } ) \\right] \\Bigg \\rangle } _ { R _ { 3 } } . } \\end{array}", + "type": "interline_equation", + "image_path": "79285d7c20a293db19d90a840732db61a34f5a16ef6a2e0efbaaa1f9d4430e1a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 144, + 95, + 467, + 134.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 144, + 134.33333333333334, + 467, + 173.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 144, + 173.66666666666669, + 467, + 213.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 220, + 505, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 151, + 234 + ], + "score": 1.0, + "content": "Bounding", + "type": "text" + }, + { + "bbox": [ + 152, + 222, + 164, + 232 + ], + "score": 0.87, + "content": "R _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 219, + 218, + 234 + ], + "score": 1.0, + "content": ". The term", + "type": "text" + }, + { + "bbox": [ + 218, + 222, + 231, + 232 + ], + "score": 0.88, + "content": "R _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 219, + 443, + 234 + ], + "score": 1.0, + "content": "can be bounded in exactly the same way as term", + "type": "text" + }, + { + "bbox": [ + 443, + 222, + 454, + 232 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 219, + 506, + 234 + ], + "score": 1.0, + "content": "in proof of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 231, + 156, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 156, + 244 + ], + "score": 1.0, + "content": "Theorem 1:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 243, + 443, + 279 + ], + "lines": [ + { + "bbox": [ + 168, + 243, + 443, + 279 + ], + "spans": [ + { + "bbox": [ + 168, + 243, + 443, + 279 + ], + "score": 0.93, + "content": "R _ { 3 } \\leq \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "1529630d222e38076b1968b366a7864e920b2807055a709cea39b00619c49e24.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 168, + 243, + 443, + 255.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 168, + 255.0, + 443, + 267.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 168, + 267.0, + 443, + 279.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 281, + 340, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 340, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 340, + 295 + ], + "score": 1.0, + "content": "Substituting the above inequality in Equation (15), we get", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 295, + 446, + 330 + ], + "lines": [ + { + "bbox": [ + 164, + 295, + 446, + 330 + ], + "spans": [ + { + "bbox": [ + 164, + 295, + 446, + 330 + ], + "score": 0.93, + "content": "R _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "2cb5efae5bcb75510b7bf1a2c4b6baed75f041a6e9b59b6f0e34b56452914193.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 164, + 295, + 446, + 306.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 164, + 306.6666666666667, + 446, + 318.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 164, + 318.33333333333337, + 446, + 330.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 333, + 504, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 218, + 346 + ], + "score": 1.0, + "content": "Here we used the fact that", + "type": "text" + }, + { + "bbox": [ + 218, + 334, + 292, + 345 + ], + "score": 0.9, + "content": "\\sqrt { v _ { t - 1 , j } } \\leq \\eta _ { l } K G", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 332, + 505, + 346 + ], + "score": 1.0, + "content": "and conditions in Theorem 2. Using Lemma 3, we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 342, + 248, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 229, + 357 + ], + "score": 1.0, + "content": "obtain the following bound on", + "type": "text" + }, + { + "bbox": [ + 230, + 345, + 242, + 355 + ], + "score": 0.85, + "content": "R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 342, + 248, + 357 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 163, + 357, + 448, + 394 + ], + "lines": [ + { + "bbox": [ + 163, + 357, + 448, + 394 + ], + "spans": [ + { + "bbox": [ + 163, + 357, + 448, + 394 + ], + "score": 0.93, + "content": "R _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )", + "type": "interline_equation", + "image_path": "fb802f805c675d6605c42ec4824c5cae647ce1d82396839cbfac9945fa900004.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 163, + 357, + 448, + 369.3333333333333 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 163, + 369.3333333333333, + 448, + 381.66666666666663 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 163, + 381.66666666666663, + 448, + 393.99999999999994 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 448, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 450, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 297, + 415 + ], + "score": 1.0, + "content": "Putting pieces together. Substituting bounds", + "type": "text" + }, + { + "bbox": [ + 298, + 402, + 310, + 412 + ], + "score": 0.89, + "content": "R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 400, + 329, + 415 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 329, + 402, + 342, + 412 + ], + "score": 0.89, + "content": "R _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 400, + 450, + 415 + ], + "score": 1.0, + "content": "in Equation (14), we have", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 416, + 500, + 489 + ], + "lines": [ + { + "bbox": [ + 111, + 416, + 500, + 489 + ], + "spans": [ + { + "bbox": [ + 111, + 416, + 500, + 489 + ], + "score": 0.94, + "content": "\\begin{array} { r l } { { \\mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \\leq f ( x _ { t } ) - \\frac { \\eta \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } } v _ { t - 1 , j } } + \\tau + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { + ( \\frac { \\eta \\sqrt { 1 - \\beta _ { 2 } } G } { \\tau } ) \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sqrt { v _ { t , j } } + \\tau ) } ] + ( \\frac { \\eta ^ { 2 } L } { 2 } ) \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { v _ { t , j } + \\tau ^ { 2 } ) } ] } \\end{array}", + "type": "interline_equation", + "image_path": "581c12f407c6907c3db8057492fd35368a978c92b9872459ee4463a4f20962ef.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 111, + 416, + 500, + 440.3333333333333 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 111, + 440.3333333333333, + 500, + 464.66666666666663 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 111, + 464.66666666666663, + 500, + 488.99999999999994 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 374, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 375, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 167, + 504 + ], + "score": 1.0, + "content": "Summing over", + "type": "text" + }, + { + "bbox": [ + 168, + 490, + 191, + 500 + ], + "score": 0.9, + "content": "t = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 488, + 202, + 504 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 202, + 490, + 228, + 500 + ], + "score": 0.88, + "content": "T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 488, + 375, + 504 + ], + "score": 1.0, + "content": "and using telescoping sum, we have", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 503, + 504, + 577 + ], + "lines": [ + { + "bbox": [ + 111, + 503, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 111, + 503, + 504, + 577 + ], + "score": 0.94, + "content": "\\mathbb { E } [ f ( x _ { T } ) ] \\leq f ( x _ { 0 } ) - \\frac { \\eta \\eta _ { l } K } { 4 } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )", + "type": "interline_equation", + "image_path": "2bb1c75401afd25f833a08734903cd0b4c8abefd23fb829ec05ba8702971b25a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 111, + 503, + 504, + 527.6666666666666 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 111, + 527.6666666666666, + 504, + 552.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 111, + 552.3333333333333, + 504, + 576.9999999999999 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 589, + 335, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 334, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 334, + 603 + ], + "score": 1.0, + "content": "To bound this term further, we need the following result.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 104, + 606, + 410, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 410, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 410, + 620 + ], + "score": 1.0, + "content": "Lemma 5. The following upper bound holds for Algorithm 2 (FEDADAM):", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 620, + 514, + 656 + ], + "lines": [ + { + "bbox": [ + 111, + 620, + 514, + 656 + ], + "spans": [ + { + "bbox": [ + 111, + 620, + 514, + 656 + ], + "score": 0.87, + "content": "\\sum _ { = 0 } ^ { { { \\nu - 1 } } } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\right] \\leq \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 1 } ^ { T - 1 } \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { \\ell = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) }", + "type": "interline_equation", + "image_path": "e2983ee88e0f3318e98488d18def796751a244c1b770753a2331705158205c09.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 111, + 620, + 514, + 632.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 111, + 632.0, + 514, + 644.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 111, + 644.0, + 514, + 656.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 133, + 677 + ], + "lines": [ + { + "bbox": [ + 104, + 664, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 135, + 681 + ], + "score": 1.0, + "content": "Proof.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 680, + 426, + 716 + ], + "lines": [ + { + "bbox": [ + 183, + 680, + 426, + 716 + ], + "spans": [ + { + "bbox": [ + 183, + 680, + 426, + 716 + ], + "score": 0.93, + "content": "\\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { l = 0 } ^ { t } \\beta _ { 2 } ^ { t - l } \\Delta _ { t , j } ^ { 2 } + \\tau ^ { 2 } } \\le \\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\tau ^ { 2 } }", + "type": "interline_equation", + "image_path": "23d9eee56aad3aa19eeedd3480a4fa77878237aa9a04b8b242f952d518d4db2f.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 183, + 680, + 426, + 698.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 183, + 698.0, + 426, + 716.0 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 312, + 761 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "20", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 339, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 340, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 150, + 96 + ], + "score": 1.0, + "content": "Bounding", + "type": "text" + }, + { + "bbox": [ + 151, + 83, + 163, + 93 + ], + "score": 0.87, + "content": "R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 81, + 214, + 96 + ], + "score": 1.0, + "content": ". The term", + "type": "text" + }, + { + "bbox": [ + 214, + 83, + 227, + 93 + ], + "score": 0.88, + "content": "R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 81, + 340, + 96 + ], + "score": 1.0, + "content": "can be bounded as follows:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 340, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 95, + 467, + 213 + ], + "lines": [ + { + "bbox": [ + 144, + 95, + 467, + 213 + ], + "spans": [ + { + "bbox": [ + 144, + 95, + 467, + 213 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { R _ { 1 } = \\Bigg \\langle \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\Bigg \\rangle } \\\\ & { \\quad = \\Bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } - \\eta _ { t } K \\nabla f ( x _ { t } ) + \\eta _ { t } K \\nabla f ( x _ { t } ) \\right] \\Bigg \\rangle } \\\\ & { \\quad = - \\eta _ { t } K \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\underbrace { \\Bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { t } K \\nabla f ( x _ { t } ) \\right] \\Bigg \\rangle } _ { R _ { 3 } } . } \\end{array}", + "type": "interline_equation", + "image_path": "79285d7c20a293db19d90a840732db61a34f5a16ef6a2e0efbaaa1f9d4430e1a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 144, + 95, + 467, + 134.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 144, + 134.33333333333334, + 467, + 173.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 144, + 173.66666666666669, + 467, + 213.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 220, + 505, + 244 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 151, + 234 + ], + "score": 1.0, + "content": "Bounding", + "type": "text" + }, + { + "bbox": [ + 152, + 222, + 164, + 232 + ], + "score": 0.87, + "content": "R _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 219, + 218, + 234 + ], + "score": 1.0, + "content": ". The term", + "type": "text" + }, + { + "bbox": [ + 218, + 222, + 231, + 232 + ], + "score": 0.88, + "content": "R _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 219, + 443, + 234 + ], + "score": 1.0, + "content": "can be bounded in exactly the same way as term", + "type": "text" + }, + { + "bbox": [ + 443, + 222, + 454, + 232 + ], + "score": 0.88, + "content": "T _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 219, + 506, + 234 + ], + "score": 1.0, + "content": "in proof of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 231, + 156, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 156, + 244 + ], + "score": 1.0, + "content": "Theorem 1:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 219, + 506, + 244 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 243, + 443, + 279 + ], + "lines": [ + { + "bbox": [ + 168, + 243, + 443, + 279 + ], + "spans": [ + { + "bbox": [ + 168, + 243, + 443, + 279 + ], + "score": 0.93, + "content": "R _ { 3 } \\leq \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "1529630d222e38076b1968b366a7864e920b2807055a709cea39b00619c49e24.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 168, + 243, + 443, + 255.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 168, + 255.0, + 443, + 267.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 168, + 267.0, + 443, + 279.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 281, + 340, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 340, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 340, + 295 + ], + "score": 1.0, + "content": "Substituting the above inequality in Equation (15), we get", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 279, + 340, + 295 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 295, + 446, + 330 + ], + "lines": [ + { + "bbox": [ + 164, + 295, + 446, + 330 + ], + "spans": [ + { + "bbox": [ + 164, + 295, + 446, + 330 + ], + "score": 0.93, + "content": "R _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "2cb5efae5bcb75510b7bf1a2c4b6baed75f041a6e9b59b6f0e34b56452914193.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 164, + 295, + 446, + 306.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 164, + 306.6666666666667, + 446, + 318.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 164, + 318.33333333333337, + 446, + 330.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 333, + 504, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 332, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 218, + 346 + ], + "score": 1.0, + "content": "Here we used the fact that", + "type": "text" + }, + { + "bbox": [ + 218, + 334, + 292, + 345 + ], + "score": 0.9, + "content": "\\sqrt { v _ { t - 1 , j } } \\leq \\eta _ { l } K G", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 332, + 505, + 346 + ], + "score": 1.0, + "content": "and conditions in Theorem 2. Using Lemma 3, we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 342, + 248, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 229, + 357 + ], + "score": 1.0, + "content": "obtain the following bound on", + "type": "text" + }, + { + "bbox": [ + 230, + 345, + 242, + 355 + ], + "score": 0.85, + "content": "R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 342, + 248, + 357 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 332, + 505, + 357 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 163, + 357, + 448, + 394 + ], + "lines": [ + { + "bbox": [ + 163, + 357, + 448, + 394 + ], + "spans": [ + { + "bbox": [ + 163, + 357, + 448, + 394 + ], + "score": 0.93, + "content": "R _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )", + "type": "interline_equation", + "image_path": "fb802f805c675d6605c42ec4824c5cae647ce1d82396839cbfac9945fa900004.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 163, + 357, + 448, + 369.3333333333333 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 163, + 369.3333333333333, + 448, + 381.66666666666663 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 163, + 381.66666666666663, + 448, + 393.99999999999994 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 448, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 450, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 297, + 415 + ], + "score": 1.0, + "content": "Putting pieces together. Substituting bounds", + "type": "text" + }, + { + "bbox": [ + 298, + 402, + 310, + 412 + ], + "score": 0.89, + "content": "R _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 400, + 329, + 415 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 329, + 402, + 342, + 412 + ], + "score": 0.89, + "content": "R _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 400, + 450, + 415 + ], + "score": 1.0, + "content": "in Equation (14), we have", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 400, + 450, + 415 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 416, + 500, + 489 + ], + "lines": [ + { + "bbox": [ + 111, + 416, + 500, + 489 + ], + "spans": [ + { + "bbox": [ + 111, + 416, + 500, + 489 + ], + "score": 0.94, + "content": "\\begin{array} { r l } { { \\mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \\leq f ( x _ { t } ) - \\frac { \\eta \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } } v _ { t - 1 , j } } + \\tau + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { + ( \\frac { \\eta \\sqrt { 1 - \\beta _ { 2 } } G } { \\tau } ) \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sqrt { v _ { t , j } } + \\tau ) } ] + ( \\frac { \\eta ^ { 2 } L } { 2 } ) \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { v _ { t , j } + \\tau ^ { 2 } ) } ] } \\end{array}", + "type": "interline_equation", + "image_path": "581c12f407c6907c3db8057492fd35368a978c92b9872459ee4463a4f20962ef.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 111, + 416, + 500, + 440.3333333333333 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 111, + 440.3333333333333, + 500, + 464.66666666666663 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 111, + 464.66666666666663, + 500, + 488.99999999999994 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 374, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 375, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 167, + 504 + ], + "score": 1.0, + "content": "Summing over", + "type": "text" + }, + { + "bbox": [ + 168, + 490, + 191, + 500 + ], + "score": 0.9, + "content": "t = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 488, + 202, + 504 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 202, + 490, + 228, + 500 + ], + "score": 0.88, + "content": "T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 488, + 375, + 504 + ], + "score": 1.0, + "content": "and using telescoping sum, we have", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 488, + 375, + 504 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 503, + 504, + 577 + ], + "lines": [ + { + "bbox": [ + 111, + 503, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 111, + 503, + 504, + 577 + ], + "score": 0.94, + "content": "\\mathbb { E } [ f ( x _ { T } ) ] \\leq f ( x _ { 0 } ) - \\frac { \\eta \\eta _ { l } K } { 4 } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )", + "type": "interline_equation", + "image_path": "2bb1c75401afd25f833a08734903cd0b4c8abefd23fb829ec05ba8702971b25a.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 111, + 503, + 504, + 527.6666666666666 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 111, + 527.6666666666666, + 504, + 552.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 111, + 552.3333333333333, + 504, + 576.9999999999999 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 589, + 335, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 334, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 334, + 603 + ], + "score": 1.0, + "content": "To bound this term further, we need the following result.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 587, + 334, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 606, + 410, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 410, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 410, + 620 + ], + "score": 1.0, + "content": "Lemma 5. The following upper bound holds for Algorithm 2 (FEDADAM):", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 605, + 410, + 620 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 620, + 514, + 656 + ], + "lines": [ + { + "bbox": [ + 111, + 620, + 514, + 656 + ], + "spans": [ + { + "bbox": [ + 111, + 620, + 514, + 656 + ], + "score": 0.87, + "content": "\\sum _ { = 0 } ^ { { { \\nu - 1 } } } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\right] \\leq \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 1 } ^ { T - 1 } \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { \\ell = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) }", + "type": "interline_equation", + "image_path": "e2983ee88e0f3318e98488d18def796751a244c1b770753a2331705158205c09.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 111, + 620, + 514, + 632.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 111, + 632.0, + 514, + 644.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 111, + 644.0, + 514, + 656.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 133, + 677 + ], + "lines": [ + { + "bbox": [ + 104, + 664, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 135, + 681 + ], + "score": 1.0, + "content": "Proof.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 664, + 135, + 681 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 680, + 426, + 716 + ], + "lines": [ + { + "bbox": [ + 183, + 680, + 426, + 716 + ], + "spans": [ + { + "bbox": [ + 183, + 680, + 426, + 716 + ], + "score": 0.93, + "content": "\\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { l = 0 } ^ { t } \\beta _ { 2 } ^ { t - l } \\Delta _ { t , j } ^ { 2 } + \\tau ^ { 2 } } \\le \\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\tau ^ { 2 } }", + "type": "interline_equation", + "image_path": "23d9eee56aad3aa19eeedd3480a4fa77878237aa9a04b8b242f952d518d4db2f.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 183, + 680, + 426, + 698.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 183, + 698.0, + 426, + 716.0 + ], + "spans": [], + "index": 33 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 506, + 97 + ], + "score": 1.0, + "content": "The rest of the proof follows along the lines of proof of Lemma 4. Using the same argument, we get", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 145, + 99, + 466, + 171 + ], + "lines": [ + { + "bbox": [ + 145, + 99, + 466, + 171 + ], + "spans": [ + { + "bbox": [ + 145, + 99, + 466, + 171 + ], + "score": 0.95, + "content": "\\begin{array} { r l r } { { \\sum _ { j = 1 } ^ { d } \\mathbb { E } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } ] \\leq \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { } & { + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } [ \\nabla f ( x _ { t } ) ^ { 2 } ] , } \\end{array}", + "type": "interline_equation", + "image_path": "8116e20a8628c1291d8690901c627f03535e92acf917808d74434203d7b5dc89.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 145, + 99, + 466, + 123.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 145, + 123.0, + 466, + 147.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 145, + 147.0, + 466, + 171.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 214, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 214, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 214, + 186 + ], + "score": 1.0, + "content": "which is the desired result.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 197, + 507, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "Substituting the bound obtained from above lemma in Equation (17) and using a similar argument for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 207, + 148, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 148, + 222 + ], + "score": 1.0, + "content": "bounding", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 218, + 392, + 254 + ], + "lines": [ + { + "bbox": [ + 218, + 218, + 392, + 254 + ], + "spans": [ + { + "bbox": [ + 218, + 218, + 392, + 254 + ], + "score": 0.94, + "content": "\\left( \\frac { \\eta \\sqrt { 1 - \\beta _ { 2 } } G } { \\tau } \\right) \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sqrt { v _ { t , j } } + \\tau ) } \\right] ,", + "type": "interline_equation", + "image_path": "db2a2e7a966fb340bee0a14d463b9efdb27af424086a3216d4baffdefcab5eee.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 218, + 218, + 392, + 236.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 218, + 236.0, + 392, + 254.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 255, + 147, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 148, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 148, + 267 + ], + "score": 1.0, + "content": "we obtain", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 270, + 514, + 347 + ], + "lines": [ + { + "bbox": [ + 111, + 270, + 514, + 347 + ], + "spans": [ + { + "bbox": [ + 111, + 270, + 514, + 347 + ], + "score": 0.92, + "content": "\\begin{array} { c } { { \\displaystyle z _ { t } [ f ( x _ { T } ) ] \\leq f ( x _ { 0 } ) - \\frac { \\eta \\eta _ { l } K } { 8 } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ { { + \\left( \\eta \\sqrt { 1 - \\beta _ { 2 } } G + \\frac { \\eta ^ { 2 } L } { 2 } \\right) \\times \\left[ \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) \\right] \\sigma _ { g , l } ^ { 2 } } } \\end{array}", + "type": "interline_equation", + "image_path": "4a9bc09902b93927839786af52a1d3af237b4586b75d29799085b5989a6fd191.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 111, + 270, + 514, + 295.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 111, + 295.6666666666667, + 514, + 321.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 111, + 321.33333333333337, + 514, + 347.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 350, + 316, + 362 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 318, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 318, + 363 + ], + "score": 1.0, + "content": "The above inequality is obtained due to the fact that:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 366, + 434, + 395 + ], + "lines": [ + { + "bbox": [ + 177, + 366, + 434, + 395 + ], + "spans": [ + { + "bbox": [ + 177, + 366, + 434, + 395 + ], + "score": 0.93, + "content": "\\left( \\sqrt { 1 - \\beta _ { 2 } } G + \\frac { \\eta L } { 2 } \\right) \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\leq \\frac { \\eta _ { l } K } { 1 6 } \\left( \\frac { 1 } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G } + \\frac { 1 } { \\tau } \\right) .", + "type": "interline_equation", + "image_path": "a4a0191d5000373a0b05493f25fbea3ac0f36d43b0d7010ecb4316f3e3f6dc37.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 177, + 366, + 434, + 375.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 177, + 375.6666666666667, + 434, + 385.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 177, + 385.33333333333337, + 434, + 395.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 398, + 506, + 411 + ], + "lines": [ + { + "bbox": [ + 104, + 396, + 507, + 413 + ], + "spans": [ + { + "bbox": [ + 104, + 396, + 311, + 413 + ], + "score": 1.0, + "content": "The above condition follows from the condition on", + "type": "text" + }, + { + "bbox": [ + 312, + 401, + 321, + 411 + ], + "score": 0.85, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 396, + 507, + 413 + ], + "score": 1.0, + "content": "in Theorem 2. We also observe the following:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 121, + 415, + 489, + 451 + ], + "lines": [ + { + "bbox": [ + 121, + 415, + 489, + 451 + ], + "spans": [ + { + "bbox": [ + 121, + 415, + 489, + 451 + ], + "score": 0.93, + "content": "\\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } \\geq \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } \\geq \\frac { T } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } \\operatorname* { m i n } _ { \\eta \\leq t \\leq T } \\Vert \\nabla f ( x _ { t } ) \\Vert ^ { 2 } .", + "type": "interline_equation", + "image_path": "a68d4e1375af9101dde5e6d6ff208aecccefacf479dd4df606fdc277f5489324.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 121, + 415, + 489, + 427.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 121, + 427.0, + 489, + 439.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 121, + 439.0, + 489, + 451.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 388, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 388, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 388, + 469 + ], + "score": 1.0, + "content": "Substituting this bound in the above equation yields the desired result.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 480, + 235, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 480, + 235, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 235, + 492 + ], + "score": 1.0, + "content": "A.4 AUXILIARY LEMMATTA", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 500, + 325, + 513 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 324, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 243, + 514 + ], + "score": 1.0, + "content": "Lemma 6. For random variables", + "type": "text" + }, + { + "bbox": [ + 243, + 503, + 284, + 513 + ], + "score": 0.88, + "content": "z _ { 1 } , \\ldots , z _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 499, + 324, + 514 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 516, + 407, + 532 + ], + "lines": [ + { + "bbox": [ + 203, + 516, + 407, + 532 + ], + "spans": [ + { + "bbox": [ + 203, + 516, + 407, + 532 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathbb { E } \\left[ \\| z _ { 1 } + . . . + z _ { r } \\| ^ { 2 } \\right] \\leq r \\mathbb { E } \\left[ \\| z _ { 1 } \\| ^ { 2 } + . . . + \\| z _ { r } \\| ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "a99b2ee87abfad1cdb2125a638328910b46443d8a5b875a8ffa4f54104940353.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 203, + 516, + 407, + 532 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 408, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 410, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 248, + 550 + ], + "score": 1.0, + "content": "Lemma 7. For independent, mean", + "type": "text" + }, + { + "bbox": [ + 248, + 537, + 255, + 546 + ], + "score": 0.65, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 534, + 328, + 550 + ], + "score": 1.0, + "content": "random variables", + "type": "text" + }, + { + "bbox": [ + 329, + 538, + 370, + 548 + ], + "score": 0.88, + "content": "z _ { 1 } , \\ldots , z _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 534, + 410, + 550 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 551, + 404, + 567 + ], + "lines": [ + { + "bbox": [ + 205, + 551, + 404, + 567 + ], + "spans": [ + { + "bbox": [ + 205, + 551, + 404, + 567 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\mathbb { E } \\left[ \\| z _ { 1 } + . . . + z _ { r } \\| ^ { 2 } \\right] = \\mathbb { E } \\left[ \\| z _ { 1 } \\| ^ { 2 } + . . . + \\| z _ { r } \\| ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "12cabfb3dc55a2bd1c78426ee8b6886b880e7f05856a147bd15b538b6c95e27e.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 205, + 551, + 404, + 567 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 580, + 456, + 607 + ], + "lines": [ + { + "bbox": [ + 104, + 579, + 458, + 595 + ], + "spans": [ + { + "bbox": [ + 104, + 579, + 458, + 595 + ], + "score": 1.0, + "content": "B FEDERATED ALGORITHMS: IMPLEMENTATIONS AND PRACTICAL", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 595, + 221, + 608 + ], + "spans": [ + { + "bbox": [ + 127, + 595, + 221, + 608 + ], + "score": 1.0, + "content": "CONSIDERATIONS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 107, + 618, + 229, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 231, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 231, + 632 + ], + "score": 1.0, + "content": "B.1 FEDAVG AND FEDOPT", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 639, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "In Algorithm 3, we give a simplified version of the FEDAVG algorithm by McMahan et al. (2017), that", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 304, + 663 + ], + "score": 1.0, + "content": "applies to the setup given in Section 2. We write", + "type": "text" + }, + { + "bbox": [ + 304, + 650, + 373, + 662 + ], + "score": 0.93, + "content": "\\mathbf { S } \\mathbf { G } \\mathbf { D } _ { K } ( x _ { t } , \\eta _ { l } , f _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 649, + 414, + 663 + ], + "score": 1.0, + "content": "to denote", + "type": "text" + }, + { + "bbox": [ + 414, + 651, + 425, + 660 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 649, + 505, + 663 + ], + "score": 1.0, + "content": "steps of SGD using", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 660, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 145, + 674 + ], + "score": 1.0, + "content": "gradients", + "type": "text" + }, + { + "bbox": [ + 146, + 661, + 186, + 673 + ], + "score": 0.92, + "content": "\\nabla f _ { i } ( x , \\bar { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 660, + 201, + 674 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 201, + 661, + 232, + 672 + ], + "score": 0.91, + "content": "z \\sim D _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 660, + 334, + 674 + ], + "score": 1.0, + "content": "with (local) learning rate", + "type": "text" + }, + { + "bbox": [ + 334, + 663, + 343, + 672 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 660, + 401, + 674 + ], + "score": 1.0, + "content": ", starting from", + "type": "text" + }, + { + "bbox": [ + 402, + 663, + 412, + 672 + ], + "score": 0.86, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 660, + 506, + 674 + ], + "score": 1.0, + "content": ". As noted in Section 2,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 672, + 504, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 504, + 684 + ], + "score": 1.0, + "content": "Algorithm 3 is the special case of Algorithm 1 where CLIENTOPT is SGD, and SERVEROPT is SGD", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 682, + 189, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 189, + 694 + ], + "score": 1.0, + "content": "with learning rate 1.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "While Algorithms 1, 2, and 3 are useful for understanding relations between federated optimization", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "methods, we are also interested in practical versions of these algorithms. In particular, Algorithms", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "1, 2, and 3 are stated in terms of a kind of ‘gradient oracle’, where we compute unbiased estimates", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 174, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 495, + 176, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 495, + 176, + 505, + 186 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 456, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 496, + 458, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 496, + 458, + 504, + 466 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 506, + 97 + ], + "score": 1.0, + "content": "The rest of the proof follows along the lines of proof of Lemma 4. Using the same argument, we get", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 79, + 506, + 97 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 145, + 99, + 466, + 171 + ], + "lines": [ + { + "bbox": [ + 145, + 99, + 466, + 171 + ], + "spans": [ + { + "bbox": [ + 145, + 99, + 466, + 171 + ], + "score": 0.95, + "content": "\\begin{array} { r l r } { { \\sum _ { j = 1 } ^ { d } \\mathbb { E } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } ] \\leq \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { } & { + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } [ \\nabla f ( x _ { t } ) ^ { 2 } ] , } \\end{array}", + "type": "interline_equation", + "image_path": "8116e20a8628c1291d8690901c627f03535e92acf917808d74434203d7b5dc89.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 145, + 99, + 466, + 123.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 145, + 123.0, + 466, + 147.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 145, + 147.0, + 466, + 171.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 214, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 214, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 214, + 186 + ], + "score": 1.0, + "content": "which is the desired result.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 174, + 214, + 186 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 197, + 507, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 210 + ], + "score": 1.0, + "content": "Substituting the bound obtained from above lemma in Equation (17) and using a similar argument for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 207, + 148, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 148, + 222 + ], + "score": 1.0, + "content": "bounding", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 196, + 505, + 222 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 218, + 392, + 254 + ], + "lines": [ + { + "bbox": [ + 218, + 218, + 392, + 254 + ], + "spans": [ + { + "bbox": [ + 218, + 218, + 392, + 254 + ], + "score": 0.94, + "content": "\\left( \\frac { \\eta \\sqrt { 1 - \\beta _ { 2 } } G } { \\tau } \\right) \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sqrt { v _ { t , j } } + \\tau ) } \\right] ,", + "type": "interline_equation", + "image_path": "db2a2e7a966fb340bee0a14d463b9efdb27af424086a3216d4baffdefcab5eee.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 218, + 218, + 392, + 236.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 218, + 236.0, + 392, + 254.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 255, + 147, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 148, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 148, + 267 + ], + "score": 1.0, + "content": "we obtain", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 255, + 148, + 267 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 270, + 514, + 347 + ], + "lines": [ + { + "bbox": [ + 111, + 270, + 514, + 347 + ], + "spans": [ + { + "bbox": [ + 111, + 270, + 514, + 347 + ], + "score": 0.92, + "content": "\\begin{array} { c } { { \\displaystyle z _ { t } [ f ( x _ { T } ) ] \\leq f ( x _ { 0 } ) - \\frac { \\eta \\eta _ { l } K } { 8 } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ { { + \\left( \\eta \\sqrt { 1 - \\beta _ { 2 } } G + \\frac { \\eta ^ { 2 } L } { 2 } \\right) \\times \\left[ \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) \\right] \\sigma _ { g , l } ^ { 2 } } } \\end{array}", + "type": "interline_equation", + "image_path": "4a9bc09902b93927839786af52a1d3af237b4586b75d29799085b5989a6fd191.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 111, + 270, + 514, + 295.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 111, + 295.6666666666667, + 514, + 321.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 111, + 321.33333333333337, + 514, + 347.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 350, + 316, + 362 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 318, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 318, + 363 + ], + "score": 1.0, + "content": "The above inequality is obtained due to the fact that:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 349, + 318, + 363 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 366, + 434, + 395 + ], + "lines": [ + { + "bbox": [ + 177, + 366, + 434, + 395 + ], + "spans": [ + { + "bbox": [ + 177, + 366, + 434, + 395 + ], + "score": 0.93, + "content": "\\left( \\sqrt { 1 - \\beta _ { 2 } } G + \\frac { \\eta L } { 2 } \\right) \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\leq \\frac { \\eta _ { l } K } { 1 6 } \\left( \\frac { 1 } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G } + \\frac { 1 } { \\tau } \\right) .", + "type": "interline_equation", + "image_path": "a4a0191d5000373a0b05493f25fbea3ac0f36d43b0d7010ecb4316f3e3f6dc37.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 177, + 366, + 434, + 375.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 177, + 375.6666666666667, + 434, + 385.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 177, + 385.33333333333337, + 434, + 395.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 398, + 506, + 411 + ], + "lines": [ + { + "bbox": [ + 104, + 396, + 507, + 413 + ], + "spans": [ + { + "bbox": [ + 104, + 396, + 311, + 413 + ], + "score": 1.0, + "content": "The above condition follows from the condition on", + "type": "text" + }, + { + "bbox": [ + 312, + 401, + 321, + 411 + ], + "score": 0.85, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 396, + 507, + 413 + ], + "score": 1.0, + "content": "in Theorem 2. We also observe the following:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 396, + 507, + 413 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 121, + 415, + 489, + 451 + ], + "lines": [ + { + "bbox": [ + 121, + 415, + 489, + 451 + ], + "spans": [ + { + "bbox": [ + 121, + 415, + 489, + 451 + ], + "score": 0.93, + "content": "\\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } \\geq \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } \\geq \\frac { T } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } \\operatorname* { m i n } _ { \\eta \\leq t \\leq T } \\Vert \\nabla f ( x _ { t } ) \\Vert ^ { 2 } .", + "type": "interline_equation", + "image_path": "a68d4e1375af9101dde5e6d6ff208aecccefacf479dd4df606fdc277f5489324.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 121, + 415, + 489, + 427.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 121, + 427.0, + 489, + 439.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 121, + 439.0, + 489, + 451.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 388, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 388, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 388, + 469 + ], + "score": 1.0, + "content": "Substituting this bound in the above equation yields the desired result.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 455, + 388, + 469 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 480, + 235, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 480, + 235, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 235, + 492 + ], + "score": 1.0, + "content": "A.4 AUXILIARY LEMMATTA", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 500, + 325, + 513 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 324, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 243, + 514 + ], + "score": 1.0, + "content": "Lemma 6. For random variables", + "type": "text" + }, + { + "bbox": [ + 243, + 503, + 284, + 513 + ], + "score": 0.88, + "content": "z _ { 1 } , \\ldots , z _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 499, + 324, + 514 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 499, + 324, + 514 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 516, + 407, + 532 + ], + "lines": [ + { + "bbox": [ + 203, + 516, + 407, + 532 + ], + "spans": [ + { + "bbox": [ + 203, + 516, + 407, + 532 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathbb { E } \\left[ \\| z _ { 1 } + . . . + z _ { r } \\| ^ { 2 } \\right] \\leq r \\mathbb { E } \\left[ \\| z _ { 1 } \\| ^ { 2 } + . . . + \\| z _ { r } \\| ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "a99b2ee87abfad1cdb2125a638328910b46443d8a5b875a8ffa4f54104940353.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 203, + 516, + 407, + 532 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 408, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 410, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 248, + 550 + ], + "score": 1.0, + "content": "Lemma 7. For independent, mean", + "type": "text" + }, + { + "bbox": [ + 248, + 537, + 255, + 546 + ], + "score": 0.65, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 534, + 328, + 550 + ], + "score": 1.0, + "content": "random variables", + "type": "text" + }, + { + "bbox": [ + 329, + 538, + 370, + 548 + ], + "score": 0.88, + "content": "z _ { 1 } , \\ldots , z _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 534, + 410, + 550 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 534, + 410, + 550 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 551, + 404, + 567 + ], + "lines": [ + { + "bbox": [ + 205, + 551, + 404, + 567 + ], + "spans": [ + { + "bbox": [ + 205, + 551, + 404, + 567 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\mathbb { E } \\left[ \\| z _ { 1 } + . . . + z _ { r } \\| ^ { 2 } \\right] = \\mathbb { E } \\left[ \\| z _ { 1 } \\| ^ { 2 } + . . . + \\| z _ { r } \\| ^ { 2 } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "12cabfb3dc55a2bd1c78426ee8b6886b880e7f05856a147bd15b538b6c95e27e.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 205, + 551, + 404, + 567 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 580, + 456, + 607 + ], + "lines": [ + { + "bbox": [ + 104, + 579, + 458, + 595 + ], + "spans": [ + { + "bbox": [ + 104, + 579, + 458, + 595 + ], + "score": 1.0, + "content": "B FEDERATED ALGORITHMS: IMPLEMENTATIONS AND PRACTICAL", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 595, + 221, + 608 + ], + "spans": [ + { + "bbox": [ + 127, + 595, + 221, + 608 + ], + "score": 1.0, + "content": "CONSIDERATIONS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 107, + 618, + 229, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 231, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 231, + 632 + ], + "score": 1.0, + "content": "B.1 FEDAVG AND FEDOPT", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 639, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "In Algorithm 3, we give a simplified version of the FEDAVG algorithm by McMahan et al. (2017), that", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 304, + 663 + ], + "score": 1.0, + "content": "applies to the setup given in Section 2. We write", + "type": "text" + }, + { + "bbox": [ + 304, + 650, + 373, + 662 + ], + "score": 0.93, + "content": "\\mathbf { S } \\mathbf { G } \\mathbf { D } _ { K } ( x _ { t } , \\eta _ { l } , f _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 649, + 414, + 663 + ], + "score": 1.0, + "content": "to denote", + "type": "text" + }, + { + "bbox": [ + 414, + 651, + 425, + 660 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 649, + 505, + 663 + ], + "score": 1.0, + "content": "steps of SGD using", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 660, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 145, + 674 + ], + "score": 1.0, + "content": "gradients", + "type": "text" + }, + { + "bbox": [ + 146, + 661, + 186, + 673 + ], + "score": 0.92, + "content": "\\nabla f _ { i } ( x , \\bar { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 660, + 201, + 674 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 201, + 661, + 232, + 672 + ], + "score": 0.91, + "content": "z \\sim D _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 660, + 334, + 674 + ], + "score": 1.0, + "content": "with (local) learning rate", + "type": "text" + }, + { + "bbox": [ + 334, + 663, + 343, + 672 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 660, + 401, + 674 + ], + "score": 1.0, + "content": ", starting from", + "type": "text" + }, + { + "bbox": [ + 402, + 663, + 412, + 672 + ], + "score": 0.86, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 660, + 506, + 674 + ], + "score": 1.0, + "content": ". As noted in Section 2,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 672, + 504, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 672, + 504, + 684 + ], + "score": 1.0, + "content": "Algorithm 3 is the special case of Algorithm 1 where CLIENTOPT is SGD, and SERVEROPT is SGD", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 682, + 189, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 189, + 694 + ], + "score": 1.0, + "content": "with learning rate 1.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 104, + 639, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "While Algorithms 1, 2, and 3 are useful for understanding relations between federated optimization", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "methods, we are also interested in practical versions of these algorithms. In particular, Algorithms", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "1, 2, and 3 are stated in terms of a kind of ‘gradient oracle’, where we compute unbiased estimates", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 243, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 243, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 243, + 97 + ], + "score": 1.0, + "content": "Algorithm 3 Simplified FEDAVG", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 112, + 96, + 327, + 181 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 112, + 96, + 327, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 96, + 327, + 181 + ], + "spans": [ + { + "bbox": [ + 112, + 96, + 327, + 181 + ], + "score": 0.554, + "html": "
Input: xo fort=0,.,T-1do
Samplea subsetSofclients
xt=xt
for each client i ∈ Sinparallel do
xt = SGDk(xt, nt, fi) for i∈ S (in parallel)
xt+1=∑iesxt
", + "type": "table", + "image_path": "aefb27f699f699b25c2789d831eb4f00c476aa4f78358921e8bac4a2bc3ca6e7.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 112, + 96, + 327, + 110.16666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 110.16666666666667, + 327, + 124.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 112, + 124.33333333333334, + 327, + 138.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 112, + 138.5, + 327, + 152.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 112, + 152.66666666666666, + 327, + 166.83333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 112, + 166.83333333333331, + 327, + 180.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 503, + 226 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "of the client’s gradient. In practical scenarios, we often only have access to finite data samples, the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 214, + 284, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 284, + 227 + ], + "score": 1.0, + "content": "number of which may vary between clients.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 231, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 334, + 244 + ], + "score": 1.0, + "content": "Instead, we assume that in (1), each client distribution", + "type": "text" + }, + { + "bbox": [ + 334, + 232, + 347, + 243 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "is the uniform distribution over some", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 241, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 144, + 254 + ], + "score": 1.0, + "content": "finite set", + "type": "text" + }, + { + "bbox": [ + 145, + 243, + 157, + 253 + ], + "score": 0.89, + "content": "D _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 241, + 189, + 254 + ], + "score": 1.0, + "content": "of size", + "type": "text" + }, + { + "bbox": [ + 189, + 244, + 199, + 253 + ], + "score": 0.84, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 241, + 222, + 254 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 222, + 244, + 233, + 253 + ], + "score": 0.86, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 241, + 505, + 254 + ], + "score": 1.0, + "content": "may vary significantly between clients, requiring extra care when", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 370, + 265 + ], + "score": 1.0, + "content": "implementing federated optimization methods. We assume the set", + "type": "text" + }, + { + "bbox": [ + 370, + 253, + 383, + 264 + ], + "score": 0.89, + "content": "D _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "is partitioned into a collection", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 263, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 150, + 275 + ], + "score": 1.0, + "content": "of batches", + "type": "text" + }, + { + "bbox": [ + 150, + 264, + 161, + 275 + ], + "score": 0.88, + "content": "B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 264, + 216, + 275 + ], + "score": 1.0, + "content": ", each of size", + "type": "text" + }, + { + "bbox": [ + 216, + 264, + 225, + 273 + ], + "score": 0.78, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 264, + 245, + 275 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 245, + 264, + 272, + 275 + ], + "score": 0.91, + "content": "b \\in B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 264, + 303, + 275 + ], + "score": 1.0, + "content": ", we let", + "type": "text" + }, + { + "bbox": [ + 304, + 263, + 335, + 276 + ], + "score": 0.93, + "content": "f _ { i } ( x ; b )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 264, + 497, + 275 + ], + "score": 1.0, + "content": "denote the average loss on this batch at", + "type": "text" + }, + { + "bbox": [ + 497, + 266, + 504, + 273 + ], + "score": 0.73, + "content": "x", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 219, + 287 + ], + "score": 1.0, + "content": "with corresponding gradient", + "type": "text" + }, + { + "bbox": [ + 219, + 275, + 258, + 286 + ], + "score": 0.91, + "content": "\\nabla f _ { i } ( x ; b )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 275, + 295, + 287 + ], + "score": 1.0, + "content": ". Thus, if", + "type": "text" + }, + { + "bbox": [ + 295, + 275, + 300, + 284 + ], + "score": 0.8, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 275, + 450, + 287 + ], + "score": 1.0, + "content": "is sampled uniformly at random from", + "type": "text" + }, + { + "bbox": [ + 450, + 275, + 461, + 285 + ], + "score": 0.83, + "content": "B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 275, + 465, + 287 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 465, + 274, + 505, + 286 + ], + "score": 0.92, + "content": "\\nabla f _ { i } ( x ; b )", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 284, + 249, + 298 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 212, + 298 + ], + "score": 1.0, + "content": "is an unbiased estimate of", + "type": "text" + }, + { + "bbox": [ + 212, + 286, + 244, + 297 + ], + "score": 0.92, + "content": "\\nabla F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 284, + 249, + 298 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 284, + 315 + ], + "score": 1.0, + "content": "When training, instead of uniformly using", + "type": "text" + }, + { + "bbox": [ + 285, + 302, + 295, + 312 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 301, + 505, + 315 + ], + "score": 1.0, + "content": "gradient steps, as in Algorithm 1, we will instead", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 140, + 326 + ], + "score": 1.0, + "content": "perform", + "type": "text" + }, + { + "bbox": [ + 140, + 313, + 150, + 322 + ], + "score": 0.8, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 312, + 505, + 326 + ], + "score": 1.0, + "content": "epochs of training over each client’s dataset. Additionally, we will take a weighted average", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 427, + 336 + ], + "score": 1.0, + "content": "of the client updates, where we weight according to the number of examples", + "type": "text" + }, + { + "bbox": [ + 428, + 325, + 438, + 334 + ], + "score": 0.85, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "in each client’s", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "dataset. This leads to a batched data version of FEDOPT in Algorithm 4, and a batched data version", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 344, + 380, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 380, + 356 + ], + "score": 1.0, + "content": "of FEDADAGRAD, FEDADAM, and FEDYOGI given in Algorithm 5.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 107, + 370, + 257, + 382 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 259, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 259, + 381 + ], + "score": 1.0, + "content": "Algorithm 4 FEDOPT - Batched data", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 111, + 384, + 302, + 523 + ], + "lines": [ + { + "bbox": [ + 116, + 383, + 267, + 396 + ], + "spans": [ + { + "bbox": [ + 116, + 383, + 142, + 396 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 143, + 386, + 154, + 395 + ], + "score": 0.79, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 383, + 267, + 396 + ], + "score": 1.0, + "content": ", CLIENTOPT, SERVEROPT", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 394, + 218, + 406 + ], + "spans": [ + { + "bbox": [ + 117, + 394, + 131, + 406 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 132, + 395, + 202, + 406 + ], + "score": 0.87, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 394, + 218, + 406 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 404, + 246, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 198, + 416 + ], + "score": 1.0, + "content": "Sample a subset", + "type": "text" + }, + { + "bbox": [ + 198, + 406, + 206, + 415 + ], + "score": 0.77, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 404, + 246, + 416 + ], + "score": 1.0, + "content": "of clients", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 415, + 164, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 164, + 428 + ], + "score": 0.65, + "content": "x _ { i } ^ { t } = x _ { t }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 426, + 275, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 191, + 438 + ], + "score": 1.0, + "content": "for each client", + "type": "text" + }, + { + "bbox": [ + 192, + 427, + 216, + 437 + ], + "score": 0.87, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 426, + 275, + 438 + ], + "score": 1.0, + "content": "in parallel do", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 145, + 437, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 145, + 437, + 161, + 449 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 161, + 438, + 216, + 448 + ], + "score": 0.86, + "content": "e = 1 , \\ldots , E", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 437, + 230, + 449 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 160, + 447, + 219, + 460 + ], + "spans": [ + { + "bbox": [ + 160, + 447, + 176, + 460 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 176, + 449, + 204, + 459 + ], + "score": 0.76, + "content": "b \\in B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 447, + 219, + 460 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 456, + 302, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 302, + 523 + ], + "score": 0.84, + "content": "\\begin{array} { c } { \\mathsf { \\Gamma } _ { i } \\gets \\mathsf { \\Gamma } _ { i } \\gets } \\\\ { g _ { i } ^ { t } = \\nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ { x _ { i } ^ { t } = \\mathrm { C } _ { \\mathrm { L I E N T O P T } } ( x _ { i } ^ { t } , g _ { i } ^ { t } , \\eta _ { l } , t ) } \\\\ { \\Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\\\ { n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , \\Delta _ { t } = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\\\ { x _ { t + 1 } = \\mathrm { S E R V E R O P T } ( x _ { t } , - \\Delta _ { t } , \\eta , t ) } \\end{array}", + "type": "inline_equation", + "image_path": "fed4b01c62352d990507536924135eaf8af1c0a0d9a93e28aa25f1270d58775a.jpg" + } + ], + "index": 28 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 537, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "In Section 5, we use Algorithm 4 when implementing FEDAVG and FEDAVGM. In particular,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "FEDAVG and FEDAVGM correspond to Algorithm 4 where CLIENTOPT and SERVEROPT are SGD.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "FEDAVG uses vanilla SGD on the server, while FEDAVGM uses SGD with a momentum parameter", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 338, + 583 + ], + "score": 1.0, + "content": "of 0.9. In both methods, we tune both client learning rate", + "type": "text" + }, + { + "bbox": [ + 338, + 572, + 348, + 581 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 569, + 446, + 583 + ], + "score": 1.0, + "content": "and server learning rate", + "type": "text" + }, + { + "bbox": [ + 446, + 572, + 453, + 581 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 569, + 505, + 583 + ], + "score": 1.0, + "content": ". This means", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "that the version of FEDAVG used in all experiments is strictly more general than that in (McMahan", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "et al., 2017), which corresponds to FEDOPT where CLIENTOPT and SERVEROPT are SGD, and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 602, + 257, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 257, + 614 + ], + "score": 1.0, + "content": "SERVEROPT has a learning rate of 1.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 618, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "score": 1.0, + "content": "We use Algorithm 5 for all implementations FEDADAGRAD, FEDADAM, and FEDYOGI in Section 5.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 629, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 217, + 642 + ], + "score": 1.0, + "content": "For FEDADAGRAD, we set", + "type": "text" + }, + { + "bbox": [ + 217, + 630, + 271, + 641 + ], + "score": 0.92, + "content": "\\beta _ { 1 } = \\beta _ { 2 } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 629, + 506, + 642 + ], + "score": 1.0, + "content": "(as typical versions of ADAGRAD do not use momentum).", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 262, + 653 + ], + "score": 1.0, + "content": "For FEDADAM and FEDYOGI we set", + "type": "text" + }, + { + "bbox": [ + 262, + 640, + 347, + 652 + ], + "score": 0.84, + "content": "\\beta _ { 1 } = { \\bar { 0 . 9 } } , \\beta _ { 2 } = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 639, + 505, + 653 + ], + "score": 1.0, + "content": ". While these parameters are generally", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 649, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 665 + ], + "score": 1.0, + "content": "good choices (Zaheer et al., 2018), we emphasize that better results may be obtainable by tuning", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 661, + 178, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 178, + 675 + ], + "score": 1.0, + "content": "these parameters.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 688, + 188, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 190, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 190, + 702 + ], + "score": 1.0, + "content": "B.2 SCAFFOLD", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "As discussed in Section 5, we compare all five optimizers above to SCAFFOLD (Karimireddy et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 719, + 497, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 497, + 733 + ], + "score": 1.0, + "content": "2019) on our various tasks. There are a few important notes about the validity of this comparison.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "22", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 243, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 243, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 243, + 97 + ], + "score": 1.0, + "content": "Algorithm 3 Simplified FEDAVG", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 112, + 96, + 327, + 181 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 112, + 96, + 327, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 96, + 327, + 181 + ], + "spans": [ + { + "bbox": [ + 112, + 96, + 327, + 181 + ], + "score": 0.554, + "html": "
Input: xo fort=0,.,T-1do
Samplea subsetSofclients
xt=xt
for each client i ∈ Sinparallel do
xt = SGDk(xt, nt, fi) for i∈ S (in parallel)
xt+1=∑iesxt
", + "type": "table", + "image_path": "aefb27f699f699b25c2789d831eb4f00c476aa4f78358921e8bac4a2bc3ca6e7.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 112, + 96, + 327, + 110.16666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 110.16666666666667, + 327, + 124.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 112, + 124.33333333333334, + 327, + 138.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 112, + 138.5, + 327, + 152.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 112, + 152.66666666666666, + 327, + 166.83333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 112, + 166.83333333333331, + 327, + 180.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 503, + 226 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "of the client’s gradient. In practical scenarios, we often only have access to finite data samples, the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 214, + 284, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 284, + 227 + ], + "score": 1.0, + "content": "number of which may vary between clients.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 203, + 505, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 231, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 334, + 244 + ], + "score": 1.0, + "content": "Instead, we assume that in (1), each client distribution", + "type": "text" + }, + { + "bbox": [ + 334, + 232, + 347, + 243 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "is the uniform distribution over some", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 241, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 144, + 254 + ], + "score": 1.0, + "content": "finite set", + "type": "text" + }, + { + "bbox": [ + 145, + 243, + 157, + 253 + ], + "score": 0.89, + "content": "D _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 241, + 189, + 254 + ], + "score": 1.0, + "content": "of size", + "type": "text" + }, + { + "bbox": [ + 189, + 244, + 199, + 253 + ], + "score": 0.84, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 241, + 222, + 254 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 222, + 244, + 233, + 253 + ], + "score": 0.86, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 241, + 505, + 254 + ], + "score": 1.0, + "content": "may vary significantly between clients, requiring extra care when", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 370, + 265 + ], + "score": 1.0, + "content": "implementing federated optimization methods. We assume the set", + "type": "text" + }, + { + "bbox": [ + 370, + 253, + 383, + 264 + ], + "score": 0.89, + "content": "D _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "is partitioned into a collection", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 263, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 150, + 275 + ], + "score": 1.0, + "content": "of batches", + "type": "text" + }, + { + "bbox": [ + 150, + 264, + 161, + 275 + ], + "score": 0.88, + "content": "B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 264, + 216, + 275 + ], + "score": 1.0, + "content": ", each of size", + "type": "text" + }, + { + "bbox": [ + 216, + 264, + 225, + 273 + ], + "score": 0.78, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 264, + 245, + 275 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 245, + 264, + 272, + 275 + ], + "score": 0.91, + "content": "b \\in B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 264, + 303, + 275 + ], + "score": 1.0, + "content": ", we let", + "type": "text" + }, + { + "bbox": [ + 304, + 263, + 335, + 276 + ], + "score": 0.93, + "content": "f _ { i } ( x ; b )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 264, + 497, + 275 + ], + "score": 1.0, + "content": "denote the average loss on this batch at", + "type": "text" + }, + { + "bbox": [ + 497, + 266, + 504, + 273 + ], + "score": 0.73, + "content": "x", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 274, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 219, + 287 + ], + "score": 1.0, + "content": "with corresponding gradient", + "type": "text" + }, + { + "bbox": [ + 219, + 275, + 258, + 286 + ], + "score": 0.91, + "content": "\\nabla f _ { i } ( x ; b )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 275, + 295, + 287 + ], + "score": 1.0, + "content": ". Thus, if", + "type": "text" + }, + { + "bbox": [ + 295, + 275, + 300, + 284 + ], + "score": 0.8, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 275, + 450, + 287 + ], + "score": 1.0, + "content": "is sampled uniformly at random from", + "type": "text" + }, + { + "bbox": [ + 450, + 275, + 461, + 285 + ], + "score": 0.83, + "content": "B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 275, + 465, + 287 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 465, + 274, + 505, + 286 + ], + "score": 0.92, + "content": "\\nabla f _ { i } ( x ; b )", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 284, + 249, + 298 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 212, + 298 + ], + "score": 1.0, + "content": "is an unbiased estimate of", + "type": "text" + }, + { + "bbox": [ + 212, + 286, + 244, + 297 + ], + "score": 0.92, + "content": "\\nabla F _ { i } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 284, + 249, + 298 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 231, + 506, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 301, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 301, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 284, + 315 + ], + "score": 1.0, + "content": "When training, instead of uniformly using", + "type": "text" + }, + { + "bbox": [ + 285, + 302, + 295, + 312 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 301, + 505, + 315 + ], + "score": 1.0, + "content": "gradient steps, as in Algorithm 1, we will instead", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 140, + 326 + ], + "score": 1.0, + "content": "perform", + "type": "text" + }, + { + "bbox": [ + 140, + 313, + 150, + 322 + ], + "score": 0.8, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 312, + 505, + 326 + ], + "score": 1.0, + "content": "epochs of training over each client’s dataset. Additionally, we will take a weighted average", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 427, + 336 + ], + "score": 1.0, + "content": "of the client updates, where we weight according to the number of examples", + "type": "text" + }, + { + "bbox": [ + 428, + 325, + 438, + 334 + ], + "score": 0.85, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "in each client’s", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "dataset. This leads to a batched data version of FEDOPT in Algorithm 4, and a batched data version", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 344, + 380, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 380, + 356 + ], + "score": 1.0, + "content": "of FEDADAGRAD, FEDADAM, and FEDYOGI given in Algorithm 5.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 301, + 506, + 356 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 370, + 257, + 382 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 259, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 259, + 381 + ], + "score": 1.0, + "content": "Algorithm 4 FEDOPT - Batched data", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "list", + "bbox": [ + 111, + 384, + 302, + 523 + ], + "lines": [ + { + "bbox": [ + 116, + 383, + 267, + 396 + ], + "spans": [ + { + "bbox": [ + 116, + 383, + 142, + 396 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 143, + 386, + 154, + 395 + ], + "score": 0.79, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 383, + 267, + 396 + ], + "score": 1.0, + "content": ", CLIENTOPT, SERVEROPT", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 117, + 394, + 218, + 406 + ], + "spans": [ + { + "bbox": [ + 117, + 394, + 131, + 406 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 132, + 395, + 202, + 406 + ], + "score": 0.87, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 394, + 218, + 406 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 130, + 404, + 246, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 198, + 416 + ], + "score": 1.0, + "content": "Sample a subset", + "type": "text" + }, + { + "bbox": [ + 198, + 406, + 206, + 415 + ], + "score": 0.77, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 404, + 246, + 416 + ], + "score": 1.0, + "content": "of clients", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 415, + 164, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 164, + 428 + ], + "score": 0.65, + "content": "x _ { i } ^ { t } = x _ { t }", + "type": "inline_equation" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 426, + 275, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 191, + 438 + ], + "score": 1.0, + "content": "for each client", + "type": "text" + }, + { + "bbox": [ + 192, + 427, + 216, + 437 + ], + "score": 0.87, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 426, + 275, + 438 + ], + "score": 1.0, + "content": "in parallel do", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 437, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 145, + 437, + 161, + 449 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 161, + 438, + 216, + 448 + ], + "score": 0.86, + "content": "e = 1 , \\ldots , E", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 437, + 230, + 449 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 160, + 447, + 219, + 460 + ], + "spans": [ + { + "bbox": [ + 160, + 447, + 176, + 460 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 176, + 449, + 204, + 459 + ], + "score": 0.76, + "content": "b \\in B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 447, + 219, + 460 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 130, + 456, + 302, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 302, + 523 + ], + "score": 0.84, + "content": "\\begin{array} { c } { \\mathsf { \\Gamma } _ { i } \\gets \\mathsf { \\Gamma } _ { i } \\gets } \\\\ { g _ { i } ^ { t } = \\nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ { x _ { i } ^ { t } = \\mathrm { C } _ { \\mathrm { L I E N T O P T } } ( x _ { i } ^ { t } , g _ { i } ^ { t } , \\eta _ { l } , t ) } \\\\ { \\Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\\\ { n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , \\Delta _ { t } = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\\\ { x _ { t + 1 } = \\mathrm { S E R V E R O P T } ( x _ { t } , - \\Delta _ { t } , \\eta , t ) } \\end{array}", + "type": "inline_equation", + "image_path": "fed4b01c62352d990507536924135eaf8af1c0a0d9a93e28aa25f1270d58775a.jpg" + } + ], + "index": 28 + } + ], + "index": 24.5, + "bbox_fs": [ + 116, + 383, + 302, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 537, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "In Section 5, we use Algorithm 4 when implementing FEDAVG and FEDAVGM. In particular,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 560 + ], + "score": 1.0, + "content": "FEDAVG and FEDAVGM correspond to Algorithm 4 where CLIENTOPT and SERVEROPT are SGD.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "FEDAVG uses vanilla SGD on the server, while FEDAVGM uses SGD with a momentum parameter", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 338, + 583 + ], + "score": 1.0, + "content": "of 0.9. In both methods, we tune both client learning rate", + "type": "text" + }, + { + "bbox": [ + 338, + 572, + 348, + 581 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 569, + 446, + 583 + ], + "score": 1.0, + "content": "and server learning rate", + "type": "text" + }, + { + "bbox": [ + 446, + 572, + 453, + 581 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 569, + 505, + 583 + ], + "score": 1.0, + "content": ". This means", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "that the version of FEDAVG used in all experiments is strictly more general than that in (McMahan", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "et al., 2017), which corresponds to FEDOPT where CLIENTOPT and SERVEROPT are SGD, and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 602, + 257, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 257, + 614 + ], + "score": 1.0, + "content": "SERVEROPT has a learning rate of 1.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 537, + 506, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 618, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "score": 1.0, + "content": "We use Algorithm 5 for all implementations FEDADAGRAD, FEDADAM, and FEDYOGI in Section 5.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 629, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 217, + 642 + ], + "score": 1.0, + "content": "For FEDADAGRAD, we set", + "type": "text" + }, + { + "bbox": [ + 217, + 630, + 271, + 641 + ], + "score": 0.92, + "content": "\\beta _ { 1 } = \\beta _ { 2 } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 629, + 506, + 642 + ], + "score": 1.0, + "content": "(as typical versions of ADAGRAD do not use momentum).", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 262, + 653 + ], + "score": 1.0, + "content": "For FEDADAM and FEDYOGI we set", + "type": "text" + }, + { + "bbox": [ + 262, + 640, + 347, + 652 + ], + "score": 0.84, + "content": "\\beta _ { 1 } = { \\bar { 0 . 9 } } , \\beta _ { 2 } = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 639, + 505, + 653 + ], + "score": 1.0, + "content": ". While these parameters are generally", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 649, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 665 + ], + "score": 1.0, + "content": "good choices (Zaheer et al., 2018), we emphasize that better results may be obtainable by tuning", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 661, + 178, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 178, + 675 + ], + "score": 1.0, + "content": "these parameters.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 619, + 506, + 675 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 688, + 188, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 190, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 190, + 702 + ], + "score": 1.0, + "content": "B.2 SCAFFOLD", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "As discussed in Section 5, we compare all five optimizers above to SCAFFOLD (Karimireddy et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 719, + 497, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 497, + 733 + ], + "score": 1.0, + "content": "2019) on our various tasks. There are a few important notes about the validity of this comparison.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 414, + 97 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 415, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 228, + 98 + ], + "score": 1.0, + "content": "Algorithm 5 FEDADAGRAD", + "type": "text" + }, + { + "bbox": [ + 233, + 82, + 415, + 99 + ], + "score": 1.0, + "content": "FEDYOGI , and FEDADAM - Batched data", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 109, + 101, + 421, + 296 + ], + "lines": [ + { + "bbox": [ + 115, + 99, + 418, + 113 + ], + "spans": [ + { + "bbox": [ + 115, + 99, + 142, + 113 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 143, + 100, + 197, + 113 + ], + "score": 0.86, + "content": "x _ { 0 } , v _ { - 1 } \\ge \\tau ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 99, + 236, + 113 + ], + "score": 1.0, + "content": ", optional", + "type": "text" + }, + { + "bbox": [ + 236, + 101, + 295, + 113 + ], + "score": 0.92, + "content": "\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 99, + 418, + 113 + ], + "score": 1.0, + "content": "for FEDYOGI and FEDADAM", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 218, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 131, + 123 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 132, + 112, + 203, + 123 + ], + "score": 0.85, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 111, + 218, + 123 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 122, + 246, + 134 + ], + "spans": [ + { + "bbox": [ + 131, + 122, + 198, + 134 + ], + "score": 1.0, + "content": "Sample a subset", + "type": "text" + }, + { + "bbox": [ + 198, + 123, + 206, + 132 + ], + "score": 0.8, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 122, + 246, + 134 + ], + "score": 1.0, + "content": "of clients", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 133, + 164, + 145 + ], + "spans": [ + { + "bbox": [ + 131, + 133, + 164, + 145 + ], + "score": 0.7, + "content": "x _ { i } ^ { t } = x _ { t }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 144, + 275, + 155 + ], + "spans": [ + { + "bbox": [ + 131, + 144, + 191, + 155 + ], + "score": 1.0, + "content": "for each client", + "type": "text" + }, + { + "bbox": [ + 192, + 145, + 216, + 154 + ], + "score": 0.88, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 144, + 275, + 155 + ], + "score": 1.0, + "content": "in parallel do", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 154, + 231, + 167 + ], + "spans": [ + { + "bbox": [ + 145, + 154, + 161, + 167 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 161, + 155, + 216, + 166 + ], + "score": 0.87, + "content": "e = 1 , \\ldots , E", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 154, + 231, + 167 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 164, + 219, + 177 + ], + "spans": [ + { + "bbox": [ + 160, + 164, + 176, + 177 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 176, + 166, + 203, + 176 + ], + "score": 0.81, + "content": "b \\in B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 164, + 219, + 177 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 175, + 273, + 226 + ], + "spans": [ + { + "bbox": [ + 129, + 175, + 273, + 226 + ], + "score": 0.4, + "content": "\\begin{array} { r l } & { ~ x _ { i } ^ { t } = \\check { x } _ { i } ^ { t } - \\eta _ { l } \\nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ & { ~ \\Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\\\ & { n = \\sum _ { i \\in S } n _ { i } , ~ \\Delta _ { t } = \\sum _ { i \\in S } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\\\ & { \\Delta _ { t } = \\beta _ { 1 } \\Delta _ { t - 1 } + ( 1 - \\beta _ { 1 } ) \\Delta _ { t } } \\end{array}", + "type": "inline_equation", + "image_path": "c6e4c3eff5d7e07aa14233cc539c77625b318473644ddb924ddc117114efe8ea.jpg" + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 228, + 275, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 229, + 200, + 241 + ], + "score": 0.7, + "content": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 228, + 275, + 243 + ], + "score": 1.0, + "content": "(FEDADAGRAD)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 246, + 358, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 358, + 259 + ], + "score": 0.36, + "content": "v _ { t } = v _ { t - 1 } - ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\mathrm { s i g n } ( v _ { t - 1 } - \\Delta _ { t } ^ { 2 } ) \\left( \\mathbf { F E D Y O G I } \\right)", + "type": "inline_equation", + "image_path": "9059cb1294b942c2760c1e3df43c3ffce86132b0f17ee464a8c3b398fd644370.jpg" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 262, + 303, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 303, + 279 + ], + "score": 1.0, + "content": "vt = β2vt−1 + (1 − β2)∆2t (FEDADAM)", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 278, + 221, + 297 + ], + "spans": [ + { + "bbox": [ + 129, + 278, + 221, + 297 + ], + "score": 1.0, + "content": "xt+1 = xt + η √∆tvt+τ", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 130, + 317, + 505, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 505, + 329 + ], + "score": 1.0, + "content": "1. In cross-device settings, this is not a fair comparison. In particular, SCAFFOLD does not", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 142, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "work in settings where clients cannot maintain state across rounds, as may be the case for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 338, + 402, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 402, + 352 + ], + "score": 1.0, + "content": "federated learning systems on edge devices, such as cell phones.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 130, + 353, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 129, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 129, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "2. SCAFFOLD has two variants described by Karimireddy et al. (2019). In Option I, the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 505, + 376 + ], + "score": 1.0, + "content": "control variate of a client is updated using a full gradient computation. This effectively", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 141, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "requires performing an extra pass over each client’s dataset, as compared to Algorithm 1. In", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 141, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "order to normalize the amount of client work, we instead use Option II, in which the clients’", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 141, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "control variates are updated using the difference between the server model and the client’s", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 142, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "learned model. This requires the same amount of client work as FEDAVG and Algorithm 2.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 427, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 440 + ], + "score": 1.0, + "content": "For practical reasons, we implement a version of SCAFFOLD mirroring Algorithm 4, in which we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 141, + 451 + ], + "score": 1.0, + "content": "perform", + "type": "text" + }, + { + "bbox": [ + 142, + 439, + 151, + 449 + ], + "score": 0.76, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "epochs over the client’s dataset, and perform weighted averaging of client models. For", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "posterity, we give the full pseudo-code of the version of SCAFFOLD used in our experiments in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "Algorithm 6. This is a simple adaptiation of Option II of the SCAFFOLD algorithm in (Karimireddy", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 369, + 482 + ], + "score": 1.0, + "content": "et al., 2019) to the same setting as Algorithm 4. In particular, we let", + "type": "text" + }, + { + "bbox": [ + 370, + 472, + 380, + 482 + ], + "score": 0.84, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "denote the number of examples", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 480, + 209, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 141, + 493 + ], + "score": 1.0, + "content": "in client", + "type": "text" + }, + { + "bbox": [ + 141, + 482, + 146, + 491 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 480, + 209, + 493 + ], + "score": 1.0, + "content": "’s local dataset.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 106, + 502, + 319, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 321, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 321, + 518 + ], + "score": 1.0, + "content": "Algorithm 6 SCAFFOLD SCAFFOLD, Option II - Batched data", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 519, + 368, + 684 + ], + "lines": [ + { + "bbox": [ + 116, + 516, + 188, + 529 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 142, + 529 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 143, + 519, + 188, + 528 + ], + "score": 0.25, + "content": "x _ { 0 } , c , \\eta _ { l } , \\eta", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 527, + 217, + 538 + ], + "spans": [ + { + "bbox": [ + 116, + 527, + 131, + 538 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 132, + 528, + 207, + 538 + ], + "score": 0.82, + "content": "t = 0 , \\cdots , T - 1 \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 527, + 217, + 538 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 537, + 246, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 198, + 549 + ], + "score": 1.0, + "content": "Sample a subset", + "type": "text" + }, + { + "bbox": [ + 198, + 538, + 206, + 548 + ], + "score": 0.79, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 537, + 246, + 549 + ], + "score": 1.0, + "content": "of clients", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 131, + 548, + 164, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 164, + 561 + ], + "score": 0.63, + "content": "x _ { i } ^ { t } = x _ { t }", + "type": "inline_equation", + "image_path": "1f8a06a3b0ecfb1dd1923fd0df3fe63fa73bba2617e728e661329e4f6fde228b.jpg" + } + ], + "index": 32 + }, + { + "bbox": [ + 131, + 559, + 275, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 191, + 571 + ], + "score": 1.0, + "content": "for each client", + "type": "text" + }, + { + "bbox": [ + 192, + 560, + 216, + 570 + ], + "score": 0.87, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 559, + 275, + 571 + ], + "score": 1.0, + "content": "in parallel do", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 146, + 570, + 231, + 582 + ], + "spans": [ + { + "bbox": [ + 146, + 570, + 161, + 582 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 161, + 571, + 216, + 581 + ], + "score": 0.87, + "content": "e = 1 , \\ldots , E", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 570, + 231, + 582 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 160, + 580, + 218, + 592 + ], + "spans": [ + { + "bbox": [ + 160, + 580, + 176, + 592 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 176, + 582, + 204, + 592 + ], + "score": 0.83, + "content": "b \\in B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 580, + 218, + 592 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 129, + 591, + 368, + 684 + ], + "spans": [ + { + "bbox": [ + 129, + 591, + 368, + 684 + ], + "score": 0.82, + "content": "\\begin{array} { r l } & { \\quad g _ { i } ^ { t } = \\breve { \\nabla } f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ & { \\quad \\quad x _ { i } ^ { t } = x _ { i } ^ { t } - \\eta _ { l } ( g _ { i } ^ { t } - c _ { i } + c ) } \\\\ & { \\quad c _ { i } ^ { + } = c _ { i } - c + ( E | \\mathcal { B } _ { i } | \\eta _ { l } ) ^ { - 1 } ( x _ { i } ^ { t } - x _ { i } ) } \\\\ & { \\quad \\Delta x _ { i } = x _ { i } ^ { t } - x _ { t } , \\Delta c _ { i } = c _ { i } ^ { + } - c _ { i } } \\\\ & { \\quad c _ { i } = c _ { i } ^ { + } } \\\\ & { \\quad n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , \\Delta x = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta x _ { i } , \\Delta c = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta c _ { i } } \\\\ & { \\quad x _ { t + 1 } = x _ { t } + \\eta \\Delta x , c = c + \\frac { | \\mathcal { S } | } { N } \\Delta c } \\end{array}", + "type": "inline_equation", + "image_path": "69ab3e69634393f08b554fc4cb522461331aa6d923889c045ef63bb15a03cd46.jpg" + } + ], + "index": 36 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 179, + 711 + ], + "score": 1.0, + "content": "In this algorithm,", + "type": "text" + }, + { + "bbox": [ + 180, + 701, + 189, + 710 + ], + "score": 0.82, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 699, + 313, + 711 + ], + "score": 1.0, + "content": "is the control variate of client", + "type": "text" + }, + { + "bbox": [ + 314, + 700, + 318, + 709 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 699, + 339, + 711 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 340, + 702, + 345, + 709 + ], + "score": 0.76, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "is the running average of these control", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 711, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 341, + 722 + ], + "score": 1.0, + "content": "variates. In practice, we must initialize the control variates", + "type": "text" + }, + { + "bbox": [ + 342, + 712, + 350, + 721 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 711, + 498, + 722 + ], + "score": 1.0, + "content": "in some way when sampling a client", + "type": "text" + }, + { + "bbox": [ + 499, + 711, + 504, + 720 + ], + "score": 0.6, + "content": "i", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 301, + 733 + ], + "score": 1.0, + "content": "for the first time. In our implementation, we set", + "type": "text" + }, + { + "bbox": [ + 301, + 722, + 328, + 732 + ], + "score": 0.9, + "content": "c _ { i } = c", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 720, + 460, + 733 + ], + "score": 1.0, + "content": "the first time we sample a client", + "type": "text" + }, + { + "bbox": [ + 460, + 721, + 465, + 730 + ], + "score": 0.61, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 720, + 505, + 733 + ], + "score": 1.0, + "content": ". This has", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 22, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 414, + 97 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 415, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 228, + 98 + ], + "score": 1.0, + "content": "Algorithm 5 FEDADAGRAD", + "type": "text" + }, + { + "bbox": [ + 233, + 82, + 415, + 99 + ], + "score": 1.0, + "content": "FEDYOGI , and FEDADAM - Batched data", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "list", + "bbox": [ + 109, + 101, + 421, + 296 + ], + "lines": [ + { + "bbox": [ + 115, + 99, + 418, + 113 + ], + "spans": [ + { + "bbox": [ + 115, + 99, + 142, + 113 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 143, + 100, + 197, + 113 + ], + "score": 0.86, + "content": "x _ { 0 } , v _ { - 1 } \\ge \\tau ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 99, + 236, + 113 + ], + "score": 1.0, + "content": ", optional", + "type": "text" + }, + { + "bbox": [ + 236, + 101, + 295, + 113 + ], + "score": 0.92, + "content": "\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 99, + 418, + 113 + ], + "score": 1.0, + "content": "for FEDYOGI and FEDADAM", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 111, + 218, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 131, + 123 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 132, + 112, + 203, + 123 + ], + "score": 0.85, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 111, + 218, + 123 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 122, + 246, + 134 + ], + "spans": [ + { + "bbox": [ + 131, + 122, + 198, + 134 + ], + "score": 1.0, + "content": "Sample a subset", + "type": "text" + }, + { + "bbox": [ + 198, + 123, + 206, + 132 + ], + "score": 0.8, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 122, + 246, + 134 + ], + "score": 1.0, + "content": "of clients", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 133, + 164, + 145 + ], + "spans": [ + { + "bbox": [ + 131, + 133, + 164, + 145 + ], + "score": 0.7, + "content": "x _ { i } ^ { t } = x _ { t }", + "type": "inline_equation" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 144, + 275, + 155 + ], + "spans": [ + { + "bbox": [ + 131, + 144, + 191, + 155 + ], + "score": 1.0, + "content": "for each client", + "type": "text" + }, + { + "bbox": [ + 192, + 145, + 216, + 154 + ], + "score": 0.88, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 144, + 275, + 155 + ], + "score": 1.0, + "content": "in parallel do", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 154, + 231, + 167 + ], + "spans": [ + { + "bbox": [ + 145, + 154, + 161, + 167 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 161, + 155, + 216, + 166 + ], + "score": 0.87, + "content": "e = 1 , \\ldots , E", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 154, + 231, + 167 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 160, + 164, + 219, + 177 + ], + "spans": [ + { + "bbox": [ + 160, + 164, + 176, + 177 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 176, + 166, + 203, + 176 + ], + "score": 0.81, + "content": "b \\in B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 164, + 219, + 177 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 129, + 175, + 273, + 226 + ], + "spans": [ + { + "bbox": [ + 129, + 175, + 273, + 226 + ], + "score": 0.4, + "content": "\\begin{array} { r l } & { ~ x _ { i } ^ { t } = \\check { x } _ { i } ^ { t } - \\eta _ { l } \\nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ & { ~ \\Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\\\ & { n = \\sum _ { i \\in S } n _ { i } , ~ \\Delta _ { t } = \\sum _ { i \\in S } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\\\ & { \\Delta _ { t } = \\beta _ { 1 } \\Delta _ { t - 1 } + ( 1 - \\beta _ { 1 } ) \\Delta _ { t } } \\end{array}", + "type": "inline_equation", + "image_path": "c6e4c3eff5d7e07aa14233cc539c77625b318473644ddb924ddc117114efe8ea.jpg" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 134, + 228, + 275, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 229, + 200, + 241 + ], + "score": 0.7, + "content": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 228, + 275, + 243 + ], + "score": 1.0, + "content": "(FEDADAGRAD)", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 133, + 246, + 358, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 358, + 259 + ], + "score": 0.36, + "content": "v _ { t } = v _ { t - 1 } - ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\mathrm { s i g n } ( v _ { t - 1 } - \\Delta _ { t } ^ { 2 } ) \\left( \\mathbf { F E D Y O G I } \\right)", + "type": "inline_equation", + "image_path": "9059cb1294b942c2760c1e3df43c3ffce86132b0f17ee464a8c3b398fd644370.jpg" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 262, + 303, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 303, + 279 + ], + "score": 1.0, + "content": "vt = β2vt−1 + (1 − β2)∆2t (FEDADAM)", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 129, + 278, + 221, + 297 + ], + "spans": [ + { + "bbox": [ + 129, + 278, + 221, + 297 + ], + "score": 1.0, + "content": "xt+1 = xt + η √∆tvt+τ", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + } + ], + "index": 6.5, + "bbox_fs": [ + 115, + 99, + 418, + 297 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 317, + 505, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 317, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 505, + 329 + ], + "score": 1.0, + "content": "1. In cross-device settings, this is not a fair comparison. In particular, SCAFFOLD does not", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 142, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "work in settings where clients cannot maintain state across rounds, as may be the case for", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 338, + 402, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 402, + 352 + ], + "score": 1.0, + "content": "federated learning systems on edge devices, such as cell phones.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 130, + 317, + 505, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 353, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 129, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 129, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "2. SCAFFOLD has two variants described by Karimireddy et al. (2019). In Option I, the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 505, + 376 + ], + "score": 1.0, + "content": "control variate of a client is updated using a full gradient computation. This effectively", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 141, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "requires performing an extra pass over each client’s dataset, as compared to Algorithm 1. In", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 141, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "order to normalize the amount of client work, we instead use Option II, in which the clients’", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 396, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 141, + 396, + 506, + 408 + ], + "score": 1.0, + "content": "control variates are updated using the difference between the server model and the client’s", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 142, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "learned model. This requires the same amount of client work as FEDAVG and Algorithm 2.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 129, + 353, + 506, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 427, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 440 + ], + "score": 1.0, + "content": "For practical reasons, we implement a version of SCAFFOLD mirroring Algorithm 4, in which we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 141, + 451 + ], + "score": 1.0, + "content": "perform", + "type": "text" + }, + { + "bbox": [ + 142, + 439, + 151, + 449 + ], + "score": 0.76, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "epochs over the client’s dataset, and perform weighted averaging of client models. For", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "posterity, we give the full pseudo-code of the version of SCAFFOLD used in our experiments in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "Algorithm 6. This is a simple adaptiation of Option II of the SCAFFOLD algorithm in (Karimireddy", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 369, + 482 + ], + "score": 1.0, + "content": "et al., 2019) to the same setting as Algorithm 4. In particular, we let", + "type": "text" + }, + { + "bbox": [ + 370, + 472, + 380, + 482 + ], + "score": 0.84, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "denote the number of examples", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 480, + 209, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 141, + 493 + ], + "score": 1.0, + "content": "in client", + "type": "text" + }, + { + "bbox": [ + 141, + 482, + 146, + 491 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 480, + 209, + 493 + ], + "score": 1.0, + "content": "’s local dataset.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 104, + 427, + 506, + 493 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 502, + 319, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 321, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 321, + 518 + ], + "score": 1.0, + "content": "Algorithm 6 SCAFFOLD SCAFFOLD, Option II - Batched data", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "list", + "bbox": [ + 108, + 519, + 368, + 684 + ], + "lines": [ + { + "bbox": [ + 116, + 516, + 188, + 529 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 142, + 529 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 143, + 519, + 188, + 528 + ], + "score": 0.25, + "content": "x _ { 0 } , c , \\eta _ { l } , \\eta", + "type": "inline_equation" + } + ], + "index": 29, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 116, + 527, + 217, + 538 + ], + "spans": [ + { + "bbox": [ + 116, + 527, + 131, + 538 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 132, + 528, + 207, + 538 + ], + "score": 0.82, + "content": "t = 0 , \\cdots , T - 1 \\cdot", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 527, + 217, + 538 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 130, + 537, + 246, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 198, + 549 + ], + "score": 1.0, + "content": "Sample a subset", + "type": "text" + }, + { + "bbox": [ + 198, + 538, + 206, + 548 + ], + "score": 0.79, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 537, + 246, + 549 + ], + "score": 1.0, + "content": "of clients", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 548, + 164, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 164, + 561 + ], + "score": 0.63, + "content": "x _ { i } ^ { t } = x _ { t }", + "type": "inline_equation", + "image_path": "1f8a06a3b0ecfb1dd1923fd0df3fe63fa73bba2617e728e661329e4f6fde228b.jpg" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 559, + 275, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 191, + 571 + ], + "score": 1.0, + "content": "for each client", + "type": "text" + }, + { + "bbox": [ + 192, + 560, + 216, + 570 + ], + "score": 0.87, + "content": "i \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 559, + 275, + 571 + ], + "score": 1.0, + "content": "in parallel do", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 570, + 231, + 582 + ], + "spans": [ + { + "bbox": [ + 146, + 570, + 161, + 582 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 161, + 571, + 216, + 581 + ], + "score": 0.87, + "content": "e = 1 , \\ldots , E", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 570, + 231, + 582 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 160, + 580, + 218, + 592 + ], + "spans": [ + { + "bbox": [ + 160, + 580, + 176, + 592 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 176, + 582, + 204, + 592 + ], + "score": 0.83, + "content": "b \\in B _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 580, + 218, + 592 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 129, + 591, + 368, + 684 + ], + "spans": [ + { + "bbox": [ + 129, + 591, + 368, + 684 + ], + "score": 0.82, + "content": "\\begin{array} { r l } & { \\quad g _ { i } ^ { t } = \\breve { \\nabla } f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ & { \\quad \\quad x _ { i } ^ { t } = x _ { i } ^ { t } - \\eta _ { l } ( g _ { i } ^ { t } - c _ { i } + c ) } \\\\ & { \\quad c _ { i } ^ { + } = c _ { i } - c + ( E | \\mathcal { B } _ { i } | \\eta _ { l } ) ^ { - 1 } ( x _ { i } ^ { t } - x _ { i } ) } \\\\ & { \\quad \\Delta x _ { i } = x _ { i } ^ { t } - x _ { t } , \\Delta c _ { i } = c _ { i } ^ { + } - c _ { i } } \\\\ & { \\quad c _ { i } = c _ { i } ^ { + } } \\\\ & { \\quad n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , \\Delta x = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta x _ { i } , \\Delta c = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta c _ { i } } \\\\ & { \\quad x _ { t + 1 } = x _ { t } + \\eta \\Delta x , c = c + \\frac { | \\mathcal { S } | } { N } \\Delta c } \\end{array}", + "type": "inline_equation", + "image_path": "69ab3e69634393f08b554fc4cb522461331aa6d923889c045ef63bb15a03cd46.jpg" + } + ], + "index": 36 + } + ], + "index": 32.5, + "bbox_fs": [ + 116, + 516, + 368, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 179, + 711 + ], + "score": 1.0, + "content": "In this algorithm,", + "type": "text" + }, + { + "bbox": [ + 180, + 701, + 189, + 710 + ], + "score": 0.82, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 699, + 313, + 711 + ], + "score": 1.0, + "content": "is the control variate of client", + "type": "text" + }, + { + "bbox": [ + 314, + 700, + 318, + 709 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 699, + 339, + 711 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 340, + 702, + 345, + 709 + ], + "score": 0.76, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "is the running average of these control", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 711, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 341, + 722 + ], + "score": 1.0, + "content": "variates. In practice, we must initialize the control variates", + "type": "text" + }, + { + "bbox": [ + 342, + 712, + 350, + 721 + ], + "score": 0.85, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 711, + 498, + 722 + ], + "score": 1.0, + "content": "in some way when sampling a client", + "type": "text" + }, + { + "bbox": [ + 499, + 711, + 504, + 720 + ], + "score": 0.6, + "content": "i", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 301, + 733 + ], + "score": 1.0, + "content": "for the first time. In our implementation, we set", + "type": "text" + }, + { + "bbox": [ + 301, + 722, + 328, + 732 + ], + "score": 0.9, + "content": "c _ { i } = c", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 720, + 460, + 733 + ], + "score": 1.0, + "content": "the first time we sample a client", + "type": "text" + }, + { + "bbox": [ + 460, + 721, + 465, + 730 + ], + "score": 0.61, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 720, + 505, + 733 + ], + "score": 1.0, + "content": ". This has", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "the advantage of exactly recovering FEDAVG when each client is sampled at most once. To initialize", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 93, + 225, + 106 + ], + "spans": [ + { + "bbox": [ + 107, + 96, + 112, + 104 + ], + "score": 0.34, + "content": "c", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 113, + 93, + 225, + 106 + ], + "score": 1.0, + "content": ", we use the all zeros vector.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 699, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "the advantage of exactly recovering FEDAVG when each client is sampled at most once. To initialize", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 93, + 225, + 106 + ], + "spans": [ + { + "bbox": [ + 107, + 96, + 112, + 104 + ], + "score": 0.34, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 93, + 225, + 106 + ], + "score": 1.0, + "content": ", we use the all zeros vector.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 109, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 121 + ], + "score": 1.0, + "content": "We compare this version of SCAFFOLD to FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "and FEDAVG on our tasks, tuning the learning rates in the same way (using the same grids as in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 223, + 144 + ], + "score": 1.0, + "content": "Appendix D.2). In particular,", + "type": "text" + }, + { + "bbox": [ + 223, + 133, + 242, + 143 + ], + "score": 0.89, + "content": "\\eta _ { l } , \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "are tuned to obtain the best training performance over the last 100", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "communication rounds. We use the same federated hyperparameters for SCAFFOLD as discussed in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 209, + 165 + ], + "score": 1.0, + "content": "Section 4. Namely, we set", + "type": "text" + }, + { + "bbox": [ + 210, + 154, + 237, + 163 + ], + "score": 0.9, + "content": "E = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 153, + 506, + 165 + ], + "score": 1.0, + "content": ", and sample 10 clients per round for all tasks except Stack Overflow", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 403, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 403, + 176 + ], + "score": 1.0, + "content": "NWP, where we sample 50. The results are given in Figure 1 in Section 5.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 109, + 192, + 354, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 356, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 356, + 204 + ], + "score": 1.0, + "content": "B.3 LOOKAHEAD, ADAALTER, AND CLIENT ADAPTIVITY", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 213, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 506, + 225 + ], + "score": 1.0, + "content": "The LOOKAHEAD optimizer (Zhang et al., 2019b) is primarily designed for non-FL settings. LOOKA-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "HEAD uses a generic optimizer in the inner loop and updates its parameters using a “outer” learning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 234, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 247 + ], + "score": 1.0, + "content": "rate. Thus, unlike FEDOPT, LOOKAHEAD uses a single generic optimizer and is thus conceptually dif-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "ferent. In fact, LOOKAHEAD can be seen as a special case of FEDOPT in non-FL settings which uses a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "generic optimizer CLIENTOPT as a client optimizer, and SGD as the server optimizer. While there are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 267, + 504, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 504, + 278 + ], + "score": 1.0, + "content": "multiple ways LOOKAHEAD could be generalized to a federated setting, one straightforward version", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 278, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 505, + 289 + ], + "score": 1.0, + "content": "would simply use an adaptive method as the CLIENTOPT. On the other hand, ADAALTER (Xie et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "2019) is designed specifically for distributed settings. In ADAALTER, clients use a local optimizer", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "similar to ADAGRAD (McMahan & Streeter, 2010a; Duchi et al., 2011) to perform multiple epochs", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "of training on their local datasets. Both LOOKAHEAD and ADAALTER use client adaptivity, which is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 321, + 450, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 450, + 332 + ], + "score": 1.0, + "content": "fundamentally different from the adaptive server optimizers proposed in Algorithm 2.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 337, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 350 + ], + "score": 1.0, + "content": "To illustrate the differences, consider the client-to-server communication in ADAALTER. This requires", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "communicating both the model weights and the client accumulators (used to perform the adaptive", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 215, + 371 + ], + "score": 1.0, + "content": "optimization, analogous to", + "type": "text" + }, + { + "bbox": [ + 216, + 360, + 225, + 370 + ], + "score": 0.85, + "content": "v _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "in Algorithm 2). In the case of ADAALTER, the client accumulator is", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "the same size as the model’s trainable weights. Thus, the client-to-server communication doubles for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "this method, relative to FEDAVG. In ADAALTER, the server averages the client accumulators and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 507, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 507, + 404 + ], + "score": 1.0, + "content": "broadcasts the average to the next set of clients, who use this to initialize their adaptive optimizers.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 402, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 506, + 414 + ], + "score": 1.0, + "content": "This means that the server-to-client communication also doubles relative to FEDAVG. The same", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 413, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 424 + ], + "score": 1.0, + "content": "would occur for any adaptive client optimizer in the distributed version of LOOKAHEAD described", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "above. For similar reasons, we see that client adaptive methods also increase the amount of memory", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "needed on the client (as they must store the current accumulator). By contrast, our adaptive server", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "methods (Algorithm 2) do not require extra communication or client memory relative to FEDAVG.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 455, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 504, + 466 + ], + "score": 1.0, + "content": "Thus, we see that server-side adaptive optimization benefits from lower per-round communication", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "and client memory requirements, which are of paramount importance for FL applications (Bonawitz", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 477, + 159, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 159, + 489 + ], + "score": 1.0, + "content": "et al., 2019).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 108, + 507, + 236, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 237, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 237, + 522 + ], + "score": 1.0, + "content": "C DATASET & MODELS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "Here we provide detailed description of the datasets and models used in the paper. We use federated", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "versions of vision datasets EMNIST (Cohen et al., 2017), CIFAR-10 (Krizhevsky & Hinton, 2009),", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "and CIFAR-100 (Krizhevsky & Hinton, 2009), and language modeling datasets Shakespeare (McMa-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "han et al., 2017) and StackOverflow (Authors, 2019). Statistics for the training datasets can be found", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 577, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 588 + ], + "score": 1.0, + "content": "in Table 2. We give descriptions of the datasets, models, and tasks below. Statistics on the number of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 587, + 470, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 470, + 600 + ], + "score": 1.0, + "content": "clients and examples in both the training and test splits of the datasets are given in Table 2.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + }, + { + "type": "table", + "bbox": [ + 118, + 644, + 491, + 716 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 253, + 621, + 357, + 633 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 252, + 620, + 359, + 633 + ], + "spans": [ + { + "bbox": [ + 252, + 620, + 359, + 633 + ], + "score": 1.0, + "content": "Table 2: Dataset statistics.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "table_body", + "bbox": [ + 118, + 644, + 491, + 716 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 644, + 491, + 716 + ], + "spans": [ + { + "bbox": [ + 118, + 644, + 491, + 716 + ], + "score": 0.983, + "html": "
DATASETTRAIN CLIENTSTRAIN EXAMPLESTEST CLIENTSTEST EXAMPLES
CIFAR-1050050,00010010,000
CIFAR-10050050,00010010,000
EMNIST-623,400671,5853,40077,483
SHAKESPEARE71516,0687152,356
STACKOVERFLOW342,477135,818,730204,08816,586,035
", + "type": "table", + "image_path": "64c387ac10e2d1ad5cbf52d2f118f0817fd2d7cb047ee8a592d862fcb7e0acf0.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 118, + 644, + 491, + 668.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 118, + 668.0, + 491, + 692.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 118, + 692.0, + 491, + 716.0 + ], + "spans": [], + "index": 44 + } + ] + } + ], + "index": 42.0 + } + ], + "page_idx": 23, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 109, + 505, + 175 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 506, + 121 + ], + "score": 1.0, + "content": "We compare this version of SCAFFOLD to FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 133 + ], + "score": 1.0, + "content": "and FEDAVG on our tasks, tuning the learning rates in the same way (using the same grids as in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 223, + 144 + ], + "score": 1.0, + "content": "Appendix D.2). In particular,", + "type": "text" + }, + { + "bbox": [ + 223, + 133, + 242, + 143 + ], + "score": 0.89, + "content": "\\eta _ { l } , \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "are tuned to obtain the best training performance over the last 100", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 154 + ], + "score": 1.0, + "content": "communication rounds. We use the same federated hyperparameters for SCAFFOLD as discussed in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 209, + 165 + ], + "score": 1.0, + "content": "Section 4. Namely, we set", + "type": "text" + }, + { + "bbox": [ + 210, + 154, + 237, + 163 + ], + "score": 0.9, + "content": "E = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 153, + 506, + 165 + ], + "score": 1.0, + "content": ", and sample 10 clients per round for all tasks except Stack Overflow", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 403, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 403, + 176 + ], + "score": 1.0, + "content": "NWP, where we sample 50. The results are given in Figure 1 in Section 5.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 110, + 506, + 176 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 192, + 354, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 356, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 356, + 204 + ], + "score": 1.0, + "content": "B.3 LOOKAHEAD, ADAALTER, AND CLIENT ADAPTIVITY", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 213, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 506, + 225 + ], + "score": 1.0, + "content": "The LOOKAHEAD optimizer (Zhang et al., 2019b) is primarily designed for non-FL settings. LOOKA-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "HEAD uses a generic optimizer in the inner loop and updates its parameters using a “outer” learning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 234, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 247 + ], + "score": 1.0, + "content": "rate. Thus, unlike FEDOPT, LOOKAHEAD uses a single generic optimizer and is thus conceptually dif-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "ferent. In fact, LOOKAHEAD can be seen as a special case of FEDOPT in non-FL settings which uses a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "generic optimizer CLIENTOPT as a client optimizer, and SGD as the server optimizer. While there are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 267, + 504, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 504, + 278 + ], + "score": 1.0, + "content": "multiple ways LOOKAHEAD could be generalized to a federated setting, one straightforward version", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 278, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 505, + 289 + ], + "score": 1.0, + "content": "would simply use an adaptive method as the CLIENTOPT. On the other hand, ADAALTER (Xie et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "2019) is designed specifically for distributed settings. In ADAALTER, clients use a local optimizer", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "similar to ADAGRAD (McMahan & Streeter, 2010a; Duchi et al., 2011) to perform multiple epochs", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "of training on their local datasets. Both LOOKAHEAD and ADAALTER use client adaptivity, which is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 321, + 450, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 450, + 332 + ], + "score": 1.0, + "content": "fundamentally different from the adaptive server optimizers proposed in Algorithm 2.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 213, + 506, + 332 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 337, + 505, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 350 + ], + "score": 1.0, + "content": "To illustrate the differences, consider the client-to-server communication in ADAALTER. This requires", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "communicating both the model weights and the client accumulators (used to perform the adaptive", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 215, + 371 + ], + "score": 1.0, + "content": "optimization, analogous to", + "type": "text" + }, + { + "bbox": [ + 216, + 360, + 225, + 370 + ], + "score": 0.85, + "content": "v _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "in Algorithm 2). In the case of ADAALTER, the client accumulator is", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "the same size as the model’s trainable weights. Thus, the client-to-server communication doubles for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "this method, relative to FEDAVG. In ADAALTER, the server averages the client accumulators and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 390, + 507, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 507, + 404 + ], + "score": 1.0, + "content": "broadcasts the average to the next set of clients, who use this to initialize their adaptive optimizers.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 402, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 506, + 414 + ], + "score": 1.0, + "content": "This means that the server-to-client communication also doubles relative to FEDAVG. The same", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 413, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 424 + ], + "score": 1.0, + "content": "would occur for any adaptive client optimizer in the distributed version of LOOKAHEAD described", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "above. For similar reasons, we see that client adaptive methods also increase the amount of memory", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "needed on the client (as they must store the current accumulator). By contrast, our adaptive server", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "methods (Algorithm 2) do not require extra communication or client memory relative to FEDAVG.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 455, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 504, + 466 + ], + "score": 1.0, + "content": "Thus, we see that server-side adaptive optimization benefits from lower per-round communication", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "and client memory requirements, which are of paramount importance for FL applications (Bonawitz", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 477, + 159, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 159, + 489 + ], + "score": 1.0, + "content": "et al., 2019).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 336, + 507, + 489 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 507, + 236, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 237, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 237, + 522 + ], + "score": 1.0, + "content": "C DATASET & MODELS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "Here we provide detailed description of the datasets and models used in the paper. We use federated", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "versions of vision datasets EMNIST (Cohen et al., 2017), CIFAR-10 (Krizhevsky & Hinton, 2009),", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "and CIFAR-100 (Krizhevsky & Hinton, 2009), and language modeling datasets Shakespeare (McMa-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "han et al., 2017) and StackOverflow (Authors, 2019). Statistics for the training datasets can be found", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 577, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 588 + ], + "score": 1.0, + "content": "in Table 2. We give descriptions of the datasets, models, and tasks below. Statistics on the number of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 587, + 470, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 470, + 600 + ], + "score": 1.0, + "content": "clients and examples in both the training and test splits of the datasets are given in Table 2.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 533, + 506, + 600 + ] + }, + { + "type": "table", + "bbox": [ + 118, + 644, + 491, + 716 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 253, + 621, + 357, + 633 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 252, + 620, + 359, + 633 + ], + "spans": [ + { + "bbox": [ + 252, + 620, + 359, + 633 + ], + "score": 1.0, + "content": "Table 2: Dataset statistics.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "table_body", + "bbox": [ + 118, + 644, + 491, + 716 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 644, + 491, + 716 + ], + "spans": [ + { + "bbox": [ + 118, + 644, + 491, + 716 + ], + "score": 0.983, + "html": "
DATASETTRAIN CLIENTSTRAIN EXAMPLESTEST CLIENTSTEST EXAMPLES
CIFAR-1050050,00010010,000
CIFAR-10050050,00010010,000
EMNIST-623,400671,5853,40077,483
SHAKESPEARE71516,0687152,356
STACKOVERFLOW342,477135,818,730204,08816,586,035
", + "type": "table", + "image_path": "64c387ac10e2d1ad5cbf52d2f118f0817fd2d7cb047ee8a592d862fcb7e0acf0.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 118, + 644, + 491, + 668.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 118, + 668.0, + 491, + 692.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 118, + 692.0, + 491, + 716.0 + ], + "spans": [], + "index": 44 + } + ] + } + ], + "index": 42.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 234, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 235, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 235, + 96 + ], + "score": 1.0, + "content": "C.1 CIFAR-10/CIFAR-100", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "We create a federated version of CIFAR-10 by randomly partitioning the training data among 500", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "clients, with each client receiving 100 examples. We use the same approach as Hsu et al. (2019),", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "where we apply latent Dirichlet allocation (LDA) over the labels of CIFAR-10 to create a federated", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "dataset. Each client has an associated multinomial distribution over labels from which its examples", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 496, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 496, + 161 + ], + "score": 1.0, + "content": "are drawn. The multinomial is drawn from a symmetric Dirichlet distribution with parameter 0.1.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "For CIFAR-100, we perform a similar partitioning of 100 examples to 500 clients, but using a more", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "sophisticated approach. We use a two step LDA process over the coarse and fine labels. We randomly", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "partition the data to reflect the \"coarse\" and \"fine\" label structure of CIFAR-100 by using the Pachinko", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "Allocation Method (PAM) (Li & McCallum, 2006). This creates more realistic client datasets, whose", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "score": 1.0, + "content": "label distributions better resemble practical heterogeneous client datasets. We have made publicly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "available the specific federated version of CIFAR-100 we used for all experiments, though we avoid", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "giving a link in this work in order to avoid de-anonymization. For complete details on how the dataset", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 241, + 225, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 225, + 254 + ], + "score": 1.0, + "content": "was created, see Appendix F.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 505, + 301 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "We train a modified ResNet-18 on both datasets, where the batch normalization layers are replaced", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 268, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 281 + ], + "score": 1.0, + "content": "by group normalization layers (Wu & He, 2018). We use two groups in each group normalization", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "layer. As shown by Hsieh et al. (2019), group normalization can lead to significant gains in accuracy", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 289, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 294, + 303 + ], + "score": 1.0, + "content": "over batch normalization in federated settings.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 321, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 443, + 334 + ], + "score": 1.0, + "content": "Preprocessing CIFAR-10 and CIFAR-100 consist of images with 3 channels of", + "type": "text" + }, + { + "bbox": [ + 443, + 322, + 477, + 333 + ], + "score": 0.88, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "pixels", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "score": 1.0, + "content": "each. Each pixel is represented by an unsigned int8. As is standard with CIFAR datasets, we perform", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "preprocessing on both training and test images. For training images, we perform a random crop", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 368 + ], + "score": 1.0, + "content": "to shape (24, 24, 3), followed by a random horizontal flip. For testing images, we centrally crop", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 160, + 378 + ], + "score": 1.0, + "content": "the image to", + "type": "text" + }, + { + "bbox": [ + 161, + 366, + 202, + 377 + ], + "score": 0.29, + "content": "( 2 4 , 2 4 , 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 365, + 505, + 378 + ], + "score": 1.0, + "content": ". For both training and testing images, we then normalize the pixel values", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 375, + 504, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 398, + 389 + ], + "score": 1.0, + "content": "according to their mean and standard deviation. Namely, given an image", + "type": "text" + }, + { + "bbox": [ + 398, + 378, + 405, + 386 + ], + "score": 0.67, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 375, + 460, + 389 + ], + "score": 1.0, + "content": ", we compute", + "type": "text" + }, + { + "bbox": [ + 461, + 375, + 504, + 388 + ], + "score": 0.93, + "content": "( x - \\mu ) / \\sigma", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 386, + 422, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 133, + 398 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 388, + 140, + 398 + ], + "score": 0.81, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 386, + 284, + 398 + ], + "score": 1.0, + "content": "is the average of the pixel values in", + "type": "text" + }, + { + "bbox": [ + 284, + 388, + 291, + 396 + ], + "score": 0.73, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 386, + 311, + 398 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 312, + 388, + 319, + 396 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 386, + 422, + 398 + ], + "score": 1.0, + "content": "is the standard deviation.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 419, + 173, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 174, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 174, + 433 + ], + "score": 1.0, + "content": "C.2 EMNIST", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "score": 1.0, + "content": "EMNIST consists of images of digits and upper and lower case English characters, with 62 total", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "score": 1.0, + "content": "classes. The federated version of EMNIST (Caldas et al., 2018) partitions the digits by their author.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "score": 1.0, + "content": "The dataset has natural heterogeneity stemming from the writing style of each person. We perform two", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "distinct tasks on EMNIST, autoencoder training (EMNIST AE) and character recognition (EMNIST", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "score": 1.0, + "content": "CR). For EMNIST AE, we train the “MNIST” autoencoder (Zaheer et al., 2018). This is a densely", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 281, + 509 + ], + "score": 1.0, + "content": "connected autoencoder with layers of size", + "type": "text" + }, + { + "bbox": [ + 281, + 497, + 433, + 509 + ], + "score": 0.88, + "content": "( 2 8 \\times 2 8 ) - 1 0 0 0 - 5 0 0 - 2 5 0 - 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "and a symmetric", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "decoder. A full description of the model is in Table 3. For EMNIST CR, we use a convolutional", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 336, + 532 + ], + "score": 1.0, + "content": "network. The network has two convolutional layers (with", + "type": "text" + }, + { + "bbox": [ + 337, + 519, + 360, + 529 + ], + "score": 0.9, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "kernels), max pooling, and dropout,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 529, + 428, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 428, + 542 + ], + "score": 1.0, + "content": "followed by a 128 unit dense layer. A full description of the model is in Table 4.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 200, + 590, + 410, + 712 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 567, + 502, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 565, + 503, + 581 + ], + "spans": [ + { + "bbox": [ + 107, + 565, + 503, + 581 + ], + "score": 1.0, + "content": "Table 3: EMNIST autoencoder model architecture. We use a sigmoid activation at all dense layers.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "table_body", + "bbox": [ + 200, + 590, + 410, + 712 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 590, + 410, + 712 + ], + "spans": [ + { + "bbox": [ + 200, + 590, + 410, + 712 + ], + "score": 0.978, + "html": "
LayerOutput Shape#of Trainable Parameters
Input7840
Dense1000785000
Dense500500500
Dense250125250
Dense307530
Dense2507750
Dense500125500
Dense1000501000
Dense784784784
", + "type": "table", + "image_path": "1cac799a631a563c437bba36e917a0f9a929e4099680e3f7cbd5c5c792362666.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 200, + 590, + 410, + 603.5555555555555 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 200, + 603.5555555555555, + 410, + 617.1111111111111 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 200, + 617.1111111111111, + 410, + 630.6666666666666 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 200, + 630.6666666666666, + 410, + 644.2222222222222 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 200, + 644.2222222222222, + 410, + 657.7777777777777 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 200, + 657.7777777777777, + 410, + 671.3333333333333 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 200, + 671.3333333333333, + 410, + 684.8888888888888 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 200, + 684.8888888888888, + 410, + 698.4444444444443 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 200, + 698.4444444444443, + 410, + 711.9999999999999 + ], + "spans": [], + "index": 44 + } + ] + } + ], + "index": 37.5 + } + ], + "page_idx": 24, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 234, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 235, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 235, + 96 + ], + "score": 1.0, + "content": "C.1 CIFAR-10/CIFAR-100", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "We create a federated version of CIFAR-10 by randomly partitioning the training data among 500", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "clients, with each client receiving 100 examples. We use the same approach as Hsu et al. (2019),", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "where we apply latent Dirichlet allocation (LDA) over the labels of CIFAR-10 to create a federated", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "dataset. Each client has an associated multinomial distribution over labels from which its examples", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 496, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 496, + 161 + ], + "score": 1.0, + "content": "are drawn. The multinomial is drawn from a symmetric Dirichlet distribution with parameter 0.1.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 106, + 505, + 161 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "For CIFAR-100, we perform a similar partitioning of 100 examples to 500 clients, but using a more", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "sophisticated approach. We use a two step LDA process over the coarse and fine labels. We randomly", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "partition the data to reflect the \"coarse\" and \"fine\" label structure of CIFAR-100 by using the Pachinko", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 210 + ], + "score": 1.0, + "content": "Allocation Method (PAM) (Li & McCallum, 2006). This creates more realistic client datasets, whose", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 222 + ], + "score": 1.0, + "content": "label distributions better resemble practical heterogeneous client datasets. We have made publicly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "available the specific federated version of CIFAR-100 we used for all experiments, though we avoid", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "giving a link in this work in order to avoid de-anonymization. For complete details on how the dataset", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 241, + 225, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 225, + 254 + ], + "score": 1.0, + "content": "was created, see Appendix F.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 166, + 506, + 254 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 505, + 301 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "We train a modified ResNet-18 on both datasets, where the batch normalization layers are replaced", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 268, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 281 + ], + "score": 1.0, + "content": "by group normalization layers (Wu & He, 2018). We use two groups in each group normalization", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "layer. As shown by Hsieh et al. (2019), group normalization can lead to significant gains in accuracy", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 289, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 294, + 303 + ], + "score": 1.0, + "content": "over batch normalization in federated settings.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 257, + 506, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 321, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 443, + 334 + ], + "score": 1.0, + "content": "Preprocessing CIFAR-10 and CIFAR-100 consist of images with 3 channels of", + "type": "text" + }, + { + "bbox": [ + 443, + 322, + 477, + 333 + ], + "score": 0.88, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "pixels", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "score": 1.0, + "content": "each. Each pixel is represented by an unsigned int8. As is standard with CIFAR datasets, we perform", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "preprocessing on both training and test images. For training images, we perform a random crop", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 368 + ], + "score": 1.0, + "content": "to shape (24, 24, 3), followed by a random horizontal flip. For testing images, we centrally crop", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 160, + 378 + ], + "score": 1.0, + "content": "the image to", + "type": "text" + }, + { + "bbox": [ + 161, + 366, + 202, + 377 + ], + "score": 0.29, + "content": "( 2 4 , 2 4 , 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 365, + 505, + 378 + ], + "score": 1.0, + "content": ". For both training and testing images, we then normalize the pixel values", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 375, + 504, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 398, + 389 + ], + "score": 1.0, + "content": "according to their mean and standard deviation. Namely, given an image", + "type": "text" + }, + { + "bbox": [ + 398, + 378, + 405, + 386 + ], + "score": 0.67, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 375, + 460, + 389 + ], + "score": 1.0, + "content": ", we compute", + "type": "text" + }, + { + "bbox": [ + 461, + 375, + 504, + 388 + ], + "score": 0.93, + "content": "( x - \\mu ) / \\sigma", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 386, + 422, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 133, + 398 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 388, + 140, + 398 + ], + "score": 0.81, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 386, + 284, + 398 + ], + "score": 1.0, + "content": "is the average of the pixel values in", + "type": "text" + }, + { + "bbox": [ + 284, + 388, + 291, + 396 + ], + "score": 0.73, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 386, + 311, + 398 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 312, + 388, + 319, + 396 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 386, + 422, + 398 + ], + "score": 1.0, + "content": "is the standard deviation.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 322, + 506, + 398 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 419, + 173, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 174, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 174, + 433 + ], + "score": 1.0, + "content": "C.2 EMNIST", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "score": 1.0, + "content": "EMNIST consists of images of digits and upper and lower case English characters, with 62 total", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "score": 1.0, + "content": "classes. The federated version of EMNIST (Caldas et al., 2018) partitions the digits by their author.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "score": 1.0, + "content": "The dataset has natural heterogeneity stemming from the writing style of each person. We perform two", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 487 + ], + "score": 1.0, + "content": "distinct tasks on EMNIST, autoencoder training (EMNIST AE) and character recognition (EMNIST", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "score": 1.0, + "content": "CR). For EMNIST AE, we train the “MNIST” autoencoder (Zaheer et al., 2018). This is a densely", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 281, + 509 + ], + "score": 1.0, + "content": "connected autoencoder with layers of size", + "type": "text" + }, + { + "bbox": [ + 281, + 497, + 433, + 509 + ], + "score": 0.88, + "content": "( 2 8 \\times 2 8 ) - 1 0 0 0 - 5 0 0 - 2 5 0 - 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "and a symmetric", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "decoder. A full description of the model is in Table 3. For EMNIST CR, we use a convolutional", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 336, + 532 + ], + "score": 1.0, + "content": "network. The network has two convolutional layers (with", + "type": "text" + }, + { + "bbox": [ + 337, + 519, + 360, + 529 + ], + "score": 0.9, + "content": "3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "kernels), max pooling, and dropout,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 529, + 428, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 428, + 542 + ], + "score": 1.0, + "content": "followed by a 128 unit dense layer. A full description of the model is in Table 4.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 443, + 506, + 542 + ] + }, + { + "type": "table", + "bbox": [ + 200, + 590, + 410, + 712 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 567, + 502, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 565, + 503, + 581 + ], + "spans": [ + { + "bbox": [ + 107, + 565, + 503, + 581 + ], + "score": 1.0, + "content": "Table 3: EMNIST autoencoder model architecture. We use a sigmoid activation at all dense layers.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "table_body", + "bbox": [ + 200, + 590, + 410, + 712 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 200, + 590, + 410, + 712 + ], + "spans": [ + { + "bbox": [ + 200, + 590, + 410, + 712 + ], + "score": 0.978, + "html": "
LayerOutput Shape#of Trainable Parameters
Input7840
Dense1000785000
Dense500500500
Dense250125250
Dense307530
Dense2507750
Dense500125500
Dense1000501000
Dense784784784
", + "type": "table", + "image_path": "1cac799a631a563c437bba36e917a0f9a929e4099680e3f7cbd5c5c792362666.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 200, + 590, + 410, + 603.5555555555555 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 200, + 603.5555555555555, + 410, + 617.1111111111111 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 200, + 617.1111111111111, + 410, + 630.6666666666666 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 200, + 630.6666666666666, + 410, + 644.2222222222222 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 200, + 644.2222222222222, + 410, + 657.7777777777777 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 200, + 657.7777777777777, + 410, + 671.3333333333333 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 200, + 671.3333333333333, + 410, + 684.8888888888888 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 200, + 684.8888888888888, + 410, + 698.4444444444443 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 200, + 698.4444444444443, + 410, + 711.9999999999999 + ], + "spans": [], + "index": 44 + } + ] + } + ], + "index": 37.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 110, + 495, + 222 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 185, + 89, + 425, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 185, + 89, + 425, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 89, + 425, + 102 + ], + "score": 1.0, + "content": "Table 4: EMNIST character recognition model architecture.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 116, + 110, + 495, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 110, + 495, + 222 + ], + "spans": [ + { + "bbox": [ + 116, + 110, + 495, + 222 + ], + "score": 0.982, + "html": "
LayerOutput Shape#of Trainable ParametersActivationHyperparameters
Input(28,28,1)0ReLUkernel size = 3; strides=(1,1)
Conv2d(26,26,32)320
Conv2d(24,24,64)18496
MaxPool2d(12,12,64)0
Dropout(12,12,64)0p= 0.25
Flatten92160
Dense1281179776
Dropout1280
Dense627998softmaxp= 0.5
", + "type": "table", + "image_path": "49a2be33aa30ee88c8550157d98584ccc7fbc15519e726d42098154e6e21d648.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 116, + 110, + 495, + 147.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 147.33333333333334, + 495, + 184.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 116, + 184.66666666666669, + 495, + 222.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "title", + "bbox": [ + 107, + 248, + 196, + 259 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 198, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 198, + 261 + ], + "score": 1.0, + "content": "C.3 SHAKESPEARE", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 506, + 281 + ], + "score": 1.0, + "content": "Shakespeare is a language modeling dataset built from the collective works of William Shakespeare.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 280, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 291 + ], + "score": 1.0, + "content": "In this dataset, each client corresponds to a speaking role with at least two lines. The dataset consists", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 289, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 303 + ], + "score": 1.0, + "content": "of 715 clients. Each client’s lines are partitioned into training and test sets. Here, the task is to do", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "next character prediction. We use an RNN that first takes a series of characters as input and embeds", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "each of them into a learned 8-dimensional space. The embedded characters are then passed through", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "2 LSTM layers, each with 256 nodes, followed by a densely connected softmax output layer. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "score": 1.0, + "content": "split the lines of each speaking role into into sequences of 80 characters, padding if necessary. We", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "use a vocabulary size of 90; 86 for the characters contained in the Shakespeare dataset, and 4 extra", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "characters for padding, out-of-vocabulary, beginning of line and end of line tokens. We train our", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 364, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 104, + 364, + 506, + 379 + ], + "score": 1.0, + "content": "model to take a sequence of 80 characters, and predict a sequence of 80 characters formed by shifting", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "the input sequence by one (so that its last character is the new character we are actually trying to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 387, + 501, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 283, + 399 + ], + "score": 1.0, + "content": "predict). Therefore, our output dimension is", + "type": "text" + }, + { + "bbox": [ + 284, + 387, + 317, + 397 + ], + "score": 0.89, + "content": "8 0 \\times 9 0", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 387, + 501, + 399 + ], + "score": 1.0, + "content": ". A full description of the model is in Table 5.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5 + }, + { + "type": "table", + "bbox": [ + 190, + 440, + 420, + 519 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 223, + 418, + 388, + 430 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 222, + 418, + 389, + 430 + ], + "spans": [ + { + "bbox": [ + 222, + 418, + 389, + 430 + ], + "score": 1.0, + "content": "Table 5: Shakespeare model architecture.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "table_body", + "bbox": [ + 190, + 440, + 420, + 519 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 190, + 440, + 420, + 519 + ], + "spans": [ + { + "bbox": [ + 190, + 440, + 420, + 519 + ], + "score": 0.98, + "html": "
LayerOutput Shape#of Trainable Parameters
Input800
Embedding(80,8)720
LSTM(80,256)271360
LSTM(80,256)525312
Dense(80,90)23130
", + "type": "table", + "image_path": "4606797e00b0dcdcdf28eda69d71d35b7b1aaa87bd586e020980b2381bb56585.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 440, + 420, + 453.1666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 190, + 453.1666666666667, + 420, + 466.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 190, + 466.33333333333337, + 420, + 479.50000000000006 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 190, + 479.50000000000006, + 420, + 492.66666666666674 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 190, + 492.66666666666674, + 420, + 505.8333333333334 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 190, + 505.8333333333334, + 420, + 519.0000000000001 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 18.75 + }, + { + "type": "title", + "bbox": [ + 107, + 543, + 213, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 214, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 214, + 556 + ], + "score": 1.0, + "content": "C.4 STACK OVERFLOW", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 504, + 576 + ], + "score": 1.0, + "content": "Stack Overflow is a language modeling dataset consisting of question and answers from the question", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "score": 1.0, + "content": "and answer site, Stack Overflow. The questions and answers also have associated metadata, including", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "tags. The dataset contains 342,477 unique users which we use as clients. We perform two tasks on this", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "dataset: tag prediction via logistic regression (Stack Overflow LR, SO LR for short), and next-word", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 607, + 504, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 454, + 620 + ], + "score": 1.0, + "content": "prediction (Stack Overflow NWP, SO NWP for short). For both tasks, we restrict to the", + "type": "text" + }, + { + "bbox": [ + 454, + 607, + 504, + 618 + ], + "score": 0.34, + "content": "1 0 , 0 0 0 { \\mathrm { m o s t } }", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "frequently used words. For Stack Overflow LR, we restrict to the 500 most frequent tags and adopt a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "one-versus-rest classification strategy, where each question/answer is represented as a bag-of-words", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 639, + 247, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 247, + 652 + ], + "score": 1.0, + "content": "vector (normalized to have sum 1).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "For Stack Overflow NWP, we restrict each client to the first 1000 sentences in their dataset (if", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "they contain this many, otherwise we use the full dataset). We also perform padding and truncation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "to ensure that sentences have 20 words. We then represent the sentence as a sequence of indices", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 689, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 700 + ], + "score": 1.0, + "content": "corresponding to the 10,000 frequently used words, as well as indices representing padding, out-of-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "vocabulary words, beginning of sentence, and end of sentence. We perform next-word-prediction on", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "these sequences using an RNN that embeds each word in a sentence into a learned 96-dimensional", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "score": 1.0, + "content": "space. It then feeds the embedded words into a single LSTM layer of hidden dimension 670, followed", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + } + ], + "page_idx": 25, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "26", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 110, + 495, + 222 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 185, + 89, + 425, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 185, + 89, + 425, + 102 + ], + "spans": [ + { + "bbox": [ + 185, + 89, + 425, + 102 + ], + "score": 1.0, + "content": "Table 4: EMNIST character recognition model architecture.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 116, + 110, + 495, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 110, + 495, + 222 + ], + "spans": [ + { + "bbox": [ + 116, + 110, + 495, + 222 + ], + "score": 0.982, + "html": "
LayerOutput Shape#of Trainable ParametersActivationHyperparameters
Input(28,28,1)0ReLUkernel size = 3; strides=(1,1)
Conv2d(26,26,32)320
Conv2d(24,24,64)18496
MaxPool2d(12,12,64)0
Dropout(12,12,64)0p= 0.25
Flatten92160
Dense1281179776
Dropout1280
Dense627998softmaxp= 0.5
", + "type": "table", + "image_path": "49a2be33aa30ee88c8550157d98584ccc7fbc15519e726d42098154e6e21d648.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 116, + 110, + 495, + 147.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 147.33333333333334, + 495, + 184.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 116, + 184.66666666666669, + 495, + 222.00000000000003 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "title", + "bbox": [ + 107, + 248, + 196, + 259 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 198, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 198, + 261 + ], + "score": 1.0, + "content": "C.3 SHAKESPEARE", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 506, + 281 + ], + "score": 1.0, + "content": "Shakespeare is a language modeling dataset built from the collective works of William Shakespeare.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 280, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 291 + ], + "score": 1.0, + "content": "In this dataset, each client corresponds to a speaking role with at least two lines. The dataset consists", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 289, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 303 + ], + "score": 1.0, + "content": "of 715 clients. Each client’s lines are partitioned into training and test sets. Here, the task is to do", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "next character prediction. We use an RNN that first takes a series of characters as input and embeds", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "each of them into a learned 8-dimensional space. The embedded characters are then passed through", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "2 LSTM layers, each with 256 nodes, followed by a densely connected softmax output layer. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 346 + ], + "score": 1.0, + "content": "split the lines of each speaking role into into sequences of 80 characters, padding if necessary. We", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "use a vocabulary size of 90; 86 for the characters contained in the Shakespeare dataset, and 4 extra", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "characters for padding, out-of-vocabulary, beginning of line and end of line tokens. We train our", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 364, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 104, + 364, + 506, + 379 + ], + "score": 1.0, + "content": "model to take a sequence of 80 characters, and predict a sequence of 80 characters formed by shifting", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "the input sequence by one (so that its last character is the new character we are actually trying to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 387, + 501, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 283, + 399 + ], + "score": 1.0, + "content": "predict). Therefore, our output dimension is", + "type": "text" + }, + { + "bbox": [ + 284, + 387, + 317, + 397 + ], + "score": 0.89, + "content": "8 0 \\times 9 0", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 387, + 501, + 399 + ], + "score": 1.0, + "content": ". A full description of the model is in Table 5.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 269, + 506, + 399 + ] + }, + { + "type": "table", + "bbox": [ + 190, + 440, + 420, + 519 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 223, + 418, + 388, + 430 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 222, + 418, + 389, + 430 + ], + "spans": [ + { + "bbox": [ + 222, + 418, + 389, + 430 + ], + "score": 1.0, + "content": "Table 5: Shakespeare model architecture.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "table_body", + "bbox": [ + 190, + 440, + 420, + 519 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 190, + 440, + 420, + 519 + ], + "spans": [ + { + "bbox": [ + 190, + 440, + 420, + 519 + ], + "score": 0.98, + "html": "
LayerOutput Shape#of Trainable Parameters
Input800
Embedding(80,8)720
LSTM(80,256)271360
LSTM(80,256)525312
Dense(80,90)23130
", + "type": "table", + "image_path": "4606797e00b0dcdcdf28eda69d71d35b7b1aaa87bd586e020980b2381bb56585.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 440, + 420, + 453.1666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 190, + 453.1666666666667, + 420, + 466.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 190, + 466.33333333333337, + 420, + 479.50000000000006 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 190, + 479.50000000000006, + 420, + 492.66666666666674 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 190, + 492.66666666666674, + 420, + 505.8333333333334 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 190, + 505.8333333333334, + 420, + 519.0000000000001 + ], + "spans": [], + "index": 23 + } + ] + } + ], + "index": 18.75 + }, + { + "type": "title", + "bbox": [ + 107, + 543, + 213, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 214, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 214, + 556 + ], + "score": 1.0, + "content": "C.4 STACK OVERFLOW", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 504, + 576 + ], + "score": 1.0, + "content": "Stack Overflow is a language modeling dataset consisting of question and answers from the question", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 588 + ], + "score": 1.0, + "content": "and answer site, Stack Overflow. The questions and answers also have associated metadata, including", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "tags. The dataset contains 342,477 unique users which we use as clients. We perform two tasks on this", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "dataset: tag prediction via logistic regression (Stack Overflow LR, SO LR for short), and next-word", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 607, + 504, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 454, + 620 + ], + "score": 1.0, + "content": "prediction (Stack Overflow NWP, SO NWP for short). For both tasks, we restrict to the", + "type": "text" + }, + { + "bbox": [ + 454, + 607, + 504, + 618 + ], + "score": 0.34, + "content": "1 0 , 0 0 0 { \\mathrm { m o s t } }", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "frequently used words. For Stack Overflow LR, we restrict to the 500 most frequent tags and adopt a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "one-versus-rest classification strategy, where each question/answer is represented as a bag-of-words", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 639, + 247, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 247, + 652 + ], + "score": 1.0, + "content": "vector (normalized to have sum 1).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 564, + 506, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "For Stack Overflow NWP, we restrict each client to the first 1000 sentences in their dataset (if", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "they contain this many, otherwise we use the full dataset). We also perform padding and truncation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "to ensure that sentences have 20 words. We then represent the sentence as a sequence of indices", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 689, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 700 + ], + "score": 1.0, + "content": "corresponding to the 10,000 frequently used words, as well as indices representing padding, out-of-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "vocabulary words, beginning of sentence, and end of sentence. We perform next-word-prediction on", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "these sequences using an RNN that embeds each word in a sentence into a learned 96-dimensional", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 732 + ], + "score": 1.0, + "content": "space. It then feeds the embedded words into a single LSTM layer of hidden dimension 670, followed", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "by a densely connected softmax output layer. A full description of the model is in Table 6. The", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "metric used in the main body is the top-1 accuracy over the proper 10,000-word vocabulary; it does", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "not include padding, out-of-vocab, or beginning or end of sentence tokens when computing the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 173, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 173, + 128 + ], + "score": 1.0, + "content": "accuracy.klkjlkj", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 656, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "by a densely connected softmax output layer. A full description of the model is in Table 6. The", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "metric used in the main body is the top-1 accuracy over the proper 10,000-word vocabulary; it does", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "not include padding, out-of-vocab, or beginning or end of sentence tokens when computing the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 173, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 173, + 128 + ], + "score": 1.0, + "content": "accuracy.klkjlkj", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table", + "bbox": [ + 190, + 169, + 420, + 247 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 173, + 145, + 436, + 158 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 173, + 144, + 437, + 159 + ], + "spans": [ + { + "bbox": [ + 173, + 144, + 437, + 159 + ], + "score": 1.0, + "content": "Table 6: Stack Overflow next word prediction model architecture.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 190, + 169, + 420, + 247 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 190, + 169, + 420, + 247 + ], + "spans": [ + { + "bbox": [ + 190, + 169, + 420, + 247 + ], + "score": 0.979, + "html": "
LayerOutput Shape#of Trainable Parameters
Input200
Embedding(20,96)960384
LSTM(20,670)2055560
Dense(20,96)64416
Dense(20,10004)970388
", + "type": "table", + "image_path": "07674b4a3346fba520b0df21158e36a0bd0b6bf6b8925b1eccbbe2b0ed6b2119.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 169, + 420, + 182.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 190, + 182.0, + 420, + 195.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 190, + 195.0, + 420, + 208.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 190, + 208.0, + 420, + 221.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 190, + 221.0, + 420, + 234.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 190, + 234.0, + 420, + 247.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 5.75 + }, + { + "type": "title", + "bbox": [ + 106, + 274, + 304, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 273, + 306, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 306, + 288 + ], + "score": 1.0, + "content": "D EXPERIMENT HYPERPARAMETERS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 299, + 252, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 253, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 253, + 311 + ], + "score": 1.0, + "content": "D.1 HYPERPARAMETER TUNING", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 319, + 506, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "Throughout our experiments, we compare the performance of different instantiations of Algorithm 1", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "that use different server optimizers. We use SGD, SGD with momentum (denoted SGDM), ADAGRAD,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "ADAM, and YOGI. For the client optimizer, we use mini-batch SGD throughout. For all tasks, we tune", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 197, + 365 + ], + "score": 1.0, + "content": "the client learning rate", + "type": "text" + }, + { + "bbox": [ + 198, + 354, + 207, + 363 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 351, + 303, + 365 + ], + "score": 1.0, + "content": "and server learning rate", + "type": "text" + }, + { + "bbox": [ + 304, + 354, + 311, + 363 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 351, + 506, + 365 + ], + "score": 1.0, + "content": "by using a large grid search. Full descriptions of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 362, + 409, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 409, + 375 + ], + "score": 1.0, + "content": "the per-task server and client learning rate grids are given in Appendix D.2.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 379, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 473, + 392 + ], + "score": 1.0, + "content": "We use the version of FEDADAGRAD, FEDADAM, and FEDYOGI in Algorithm 5. We let", + "type": "text" + }, + { + "bbox": [ + 473, + 380, + 504, + 391 + ], + "score": 0.89, + "content": "\\beta _ { 1 } = 0", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 233, + 402 + ], + "score": 1.0, + "content": "for FEDADAGRAD, and we let", + "type": "text" + }, + { + "bbox": [ + 234, + 390, + 318, + 402 + ], + "score": 0.78, + "content": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "for FEDADAM, and FEDYOGI. For FEDAVG", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "score": 1.0, + "content": "and FEDAVGM, we use Algorithm 4, where CLIENTOPT, SERVEROPT are SGD. For FEDAVGM,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "the server SGD optimizer uses a momentum parameter of 0.9. For FEDADAGRAD, FEDADAM, and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 422, + 315, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 242, + 435 + ], + "score": 1.0, + "content": "FEDYOGI, we tune the parameter", + "type": "text" + }, + { + "bbox": [ + 243, + 424, + 249, + 432 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 422, + 315, + 435 + ], + "score": 1.0, + "content": "in Algorithm 5.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 438, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 349, + 452 + ], + "score": 1.0, + "content": "When tuning parameters, we select the best hyperparameters", + "type": "text" + }, + { + "bbox": [ + 349, + 439, + 385, + 451 + ], + "score": 0.93, + "content": "( \\eta _ { l } , \\eta , \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 438, + 505, + 452 + ], + "score": 1.0, + "content": "based on the average training", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "loss over the last 100 communication rounds. Note that at each round, we only see a fraction of the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "total users (10 for each task except Stack Overflow NWP, which uses 50). Thus, the training loss at a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 506, + 484 + ], + "score": 1.0, + "content": "given round is a noisy estimate of the population-level training loss, which is why we averave over a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 483, + 246, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 246, + 493 + ], + "score": 1.0, + "content": "window of communication rounds.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 108, + 507, + 245, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 246, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 246, + 519 + ], + "score": 1.0, + "content": "D.2 HYPERPARAMETER GRIDS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 262, + 540 + ], + "score": 1.0, + "content": "Below, we give the client learning rate", + "type": "text" + }, + { + "bbox": [ + 262, + 529, + 271, + 540 + ], + "score": 0.77, + "content": "\\textcircled { \\eta _ { l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 527, + 434, + 540 + ], + "score": 1.0, + "content": "in Algorithm 1) and server learning rate (", + "type": "text" + }, + { + "bbox": [ + 434, + 530, + 441, + 540 + ], + "score": 0.74, + "content": "\\dot { \\eta }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "in Algorithm 1)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 538, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 482, + 552 + ], + "score": 1.0, + "content": "grids used for each task. These grids were chosen based on an initial evaluation over the grids", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 555, + 382, + 570 + ], + "lines": [ + { + "bbox": [ + 229, + 555, + 382, + 570 + ], + "spans": [ + { + "bbox": [ + 229, + 555, + 382, + 570 + ], + "score": 0.84, + "content": "\\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \\dotsc , 1 0 ^ { 0 . 5 } \\}", + "type": "interline_equation", + "image_path": "091adbb736908132cf026b29b99d714c6b88f7f7a7ae64332e741b84825d7273.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 229, + 555, + 382, + 570 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 575, + 378, + 590 + ], + "lines": [ + { + "bbox": [ + 233, + 575, + 378, + 590 + ], + "spans": [ + { + "bbox": [ + 233, + 575, + 378, + 590 + ], + "score": 0.84, + "content": "\\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \\dots , 1 0 ^ { 1 } \\}", + "type": "interline_equation", + "image_path": "959fa51de3117b8b618ad47658706dbb38adce7165a265e04d2de3df575cffdc.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 233, + 575, + 378, + 590 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "These grids were then refined for Stack Overflow LR and EMNIST AE in an attempt to ensure that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "score": 1.0, + "content": "the best client/server learning rate combinations for each optimizer was contained in the interior of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "score": 1.0, + "content": "the learning rate grids. We generally found that these two tasks required searching larger learning", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 625, + 355, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 355, + 638 + ], + "score": 1.0, + "content": "rates than the other two tasks. The final grids were as follows:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 641, + 155, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 156, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 156, + 654 + ], + "score": 1.0, + "content": "CIFAR-10:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 650, + 369, + 685 + ], + "lines": [ + { + "bbox": [ + 241, + 650, + 369, + 685 + ], + "spans": [ + { + "bbox": [ + 241, + 650, + 369, + 685 + ], + "score": 0.58, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "9e6d7fab16dfcd5cacf80dce1a9993e9a5e85693b8b629186a033f737e0833ce.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 650, + 369, + 667.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 241, + 667.5, + 369, + 685.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 160, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 691, + 161, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 161, + 706 + ], + "score": 1.0, + "content": "CIFAR-100:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 700, + 369, + 734 + ], + "lines": [ + { + "bbox": [ + 241, + 700, + 369, + 734 + ], + "spans": [ + { + "bbox": [ + 241, + 700, + 369, + 734 + ], + "score": 0.72, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "8dc8f8156718af7f955b03bf75e7f6511ce5db902a5c743fe27b2b8a6b770d78.jpg" + } + ] + } + ], + "index": 41.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 700, + 369, + 717.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 241, + 717.0, + 369, + 734.0 + ], + "spans": [], + "index": 42 + } + ] + } + ], + "page_idx": 26, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "27", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 505, + 128 + ], + "lines_deleted": true + }, + { + "type": "table", + "bbox": [ + 190, + 169, + 420, + 247 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 173, + 145, + 436, + 158 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 173, + 144, + 437, + 159 + ], + "spans": [ + { + "bbox": [ + 173, + 144, + 437, + 159 + ], + "score": 1.0, + "content": "Table 6: Stack Overflow next word prediction model architecture.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 190, + 169, + 420, + 247 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 190, + 169, + 420, + 247 + ], + "spans": [ + { + "bbox": [ + 190, + 169, + 420, + 247 + ], + "score": 0.979, + "html": "
LayerOutput Shape#of Trainable Parameters
Input200
Embedding(20,96)960384
LSTM(20,670)2055560
Dense(20,96)64416
Dense(20,10004)970388
", + "type": "table", + "image_path": "07674b4a3346fba520b0df21158e36a0bd0b6bf6b8925b1eccbbe2b0ed6b2119.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 169, + 420, + 182.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 190, + 182.0, + 420, + 195.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 190, + 195.0, + 420, + 208.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 190, + 208.0, + 420, + 221.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 190, + 221.0, + 420, + 234.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 190, + 234.0, + 420, + 247.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 5.75 + }, + { + "type": "title", + "bbox": [ + 106, + 274, + 304, + 287 + ], + "lines": [ + { + "bbox": [ + 106, + 273, + 306, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 306, + 288 + ], + "score": 1.0, + "content": "D EXPERIMENT HYPERPARAMETERS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 299, + 252, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 253, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 253, + 311 + ], + "score": 1.0, + "content": "D.1 HYPERPARAMETER TUNING", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 319, + 506, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "Throughout our experiments, we compare the performance of different instantiations of Algorithm 1", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "that use different server optimizers. We use SGD, SGD with momentum (denoted SGDM), ADAGRAD,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "ADAM, and YOGI. For the client optimizer, we use mini-batch SGD throughout. For all tasks, we tune", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 351, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 197, + 365 + ], + "score": 1.0, + "content": "the client learning rate", + "type": "text" + }, + { + "bbox": [ + 198, + 354, + 207, + 363 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 351, + 303, + 365 + ], + "score": 1.0, + "content": "and server learning rate", + "type": "text" + }, + { + "bbox": [ + 304, + 354, + 311, + 363 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 351, + 506, + 365 + ], + "score": 1.0, + "content": "by using a large grid search. Full descriptions of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 362, + 409, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 409, + 375 + ], + "score": 1.0, + "content": "the per-task server and client learning rate grids are given in Appendix D.2.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 320, + 506, + 375 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 379, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 380, + 504, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 473, + 392 + ], + "score": 1.0, + "content": "We use the version of FEDADAGRAD, FEDADAM, and FEDYOGI in Algorithm 5. We let", + "type": "text" + }, + { + "bbox": [ + 473, + 380, + 504, + 391 + ], + "score": 0.89, + "content": "\\beta _ { 1 } = 0", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 233, + 402 + ], + "score": 1.0, + "content": "for FEDADAGRAD, and we let", + "type": "text" + }, + { + "bbox": [ + 234, + 390, + 318, + 402 + ], + "score": 0.78, + "content": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "for FEDADAM, and FEDYOGI. For FEDAVG", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "score": 1.0, + "content": "and FEDAVGM, we use Algorithm 4, where CLIENTOPT, SERVEROPT are SGD. For FEDAVGM,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "the server SGD optimizer uses a momentum parameter of 0.9. For FEDADAGRAD, FEDADAM, and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 422, + 315, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 242, + 435 + ], + "score": 1.0, + "content": "FEDYOGI, we tune the parameter", + "type": "text" + }, + { + "bbox": [ + 243, + 424, + 249, + 432 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 422, + 315, + 435 + ], + "score": 1.0, + "content": "in Algorithm 5.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 380, + 506, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 438, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 349, + 452 + ], + "score": 1.0, + "content": "When tuning parameters, we select the best hyperparameters", + "type": "text" + }, + { + "bbox": [ + 349, + 439, + 385, + 451 + ], + "score": 0.93, + "content": "( \\eta _ { l } , \\eta , \\tau )", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 438, + 505, + 452 + ], + "score": 1.0, + "content": "based on the average training", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "loss over the last 100 communication rounds. Note that at each round, we only see a fraction of the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "total users (10 for each task except Stack Overflow NWP, which uses 50). Thus, the training loss at a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 506, + 484 + ], + "score": 1.0, + "content": "given round is a noisy estimate of the population-level training loss, which is why we averave over a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 483, + 246, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 246, + 493 + ], + "score": 1.0, + "content": "window of communication rounds.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 438, + 506, + 493 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 507, + 245, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 507, + 246, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 246, + 519 + ], + "score": 1.0, + "content": "D.2 HYPERPARAMETER GRIDS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 262, + 540 + ], + "score": 1.0, + "content": "Below, we give the client learning rate", + "type": "text" + }, + { + "bbox": [ + 262, + 529, + 271, + 540 + ], + "score": 0.77, + "content": "\\textcircled { \\eta _ { l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 527, + 434, + 540 + ], + "score": 1.0, + "content": "in Algorithm 1) and server learning rate (", + "type": "text" + }, + { + "bbox": [ + 434, + 530, + 441, + 540 + ], + "score": 0.74, + "content": "\\dot { \\eta }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "in Algorithm 1)", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 538, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 482, + 552 + ], + "score": 1.0, + "content": "grids used for each task. These grids were chosen based on an initial evaluation over the grids", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 527, + 506, + 552 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 555, + 382, + 570 + ], + "lines": [ + { + "bbox": [ + 229, + 555, + 382, + 570 + ], + "spans": [ + { + "bbox": [ + 229, + 555, + 382, + 570 + ], + "score": 0.84, + "content": "\\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \\dotsc , 1 0 ^ { 0 . 5 } \\}", + "type": "interline_equation", + "image_path": "091adbb736908132cf026b29b99d714c6b88f7f7a7ae64332e741b84825d7273.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 229, + 555, + 382, + 570 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 575, + 378, + 590 + ], + "lines": [ + { + "bbox": [ + 233, + 575, + 378, + 590 + ], + "spans": [ + { + "bbox": [ + 233, + 575, + 378, + 590 + ], + "score": 0.84, + "content": "\\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \\dots , 1 0 ^ { 1 } \\}", + "type": "interline_equation", + "image_path": "959fa51de3117b8b618ad47658706dbb38adce7165a265e04d2de3df575cffdc.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 233, + 575, + 378, + 590 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "These grids were then refined for Stack Overflow LR and EMNIST AE in an attempt to ensure that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "score": 1.0, + "content": "the best client/server learning rate combinations for each optimizer was contained in the interior of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 628 + ], + "score": 1.0, + "content": "the learning rate grids. We generally found that these two tasks required searching larger learning", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 625, + 355, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 355, + 638 + ], + "score": 1.0, + "content": "rates than the other two tasks. The final grids were as follows:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 592, + 506, + 638 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 641, + 155, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 156, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 156, + 654 + ], + "score": 1.0, + "content": "CIFAR-10:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 642, + 156, + 654 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 650, + 369, + 685 + ], + "lines": [ + { + "bbox": [ + 241, + 650, + 369, + 685 + ], + "spans": [ + { + "bbox": [ + 241, + 650, + 369, + 685 + ], + "score": 0.58, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "9e6d7fab16dfcd5cacf80dce1a9993e9a5e85693b8b629186a033f737e0833ce.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 650, + 369, + 667.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 241, + 667.5, + 369, + 685.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 160, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 691, + 161, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 161, + 706 + ], + "score": 1.0, + "content": "CIFAR-100:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40, + "bbox_fs": [ + 106, + 691, + 161, + 706 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 700, + 369, + 734 + ], + "lines": [ + { + "bbox": [ + 241, + 700, + 369, + 734 + ], + "spans": [ + { + "bbox": [ + 241, + 700, + 369, + 734 + ], + "score": 0.72, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "8dc8f8156718af7f955b03bf75e7f6511ce5db902a5c743fe27b2b8a6b770d78.jpg" + } + ] + } + ], + "index": 41.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 700, + 369, + 717.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 241, + 717.0, + 369, + 734.0 + ], + "spans": [], + "index": 42 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 166, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 167, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 167, + 96 + ], + "score": 1.0, + "content": "EMNIST AE:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 90, + 366, + 123 + ], + "lines": [ + { + "bbox": [ + 244, + 90, + 366, + 123 + ], + "spans": [ + { + "bbox": [ + 244, + 90, + 366, + 123 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 1 . 5 } , 1 0 ^ { - 1 } , \\dots , 1 0 ^ { 2 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \\dots , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "b565073787858f2a311a5ab65a60e1e3bf5a43a684fe4c409e2b4a2a22d64f72.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 244, + 90, + 366, + 106.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 244, + 106.5, + 366, + 123.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 167, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 130, + 168, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 168, + 144 + ], + "score": 1.0, + "content": "EMNIST CR:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 138, + 369, + 171 + ], + "lines": [ + { + "bbox": [ + 242, + 138, + 369, + 171 + ], + "spans": [ + { + "bbox": [ + 242, + 138, + 369, + 171 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "5829fa2b6c99ae68d686975dd4e00cba1ae1f7acf303ca8632112c0596ed9480.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 242, + 138, + 369, + 154.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 242, + 154.5, + 369, + 171.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 177, + 164, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 176, + 165, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 165, + 191 + ], + "score": 1.0, + "content": "Shakespeare:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 186, + 369, + 218 + ], + "lines": [ + { + "bbox": [ + 242, + 186, + 369, + 218 + ], + "spans": [ + { + "bbox": [ + 242, + 186, + 369, + 218 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "1925b4444836ff6a20b1d96a8b03bc1e2991f89d4e3a73cc9ae5823d393261c2.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 242, + 186, + 369, + 202.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 242, + 202.0, + 369, + 218.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 225, + 190, + 236 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 191, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 191, + 238 + ], + "score": 1.0, + "content": "StackOverflow LR:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 233, + 367, + 266 + ], + "lines": [ + { + "bbox": [ + 243, + 233, + 367, + 266 + ], + "spans": [ + { + "bbox": [ + 243, + 233, + 367, + 266 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\eta _ { l } \\in \\{ 1 0 ^ { - 1 } , 1 0 ^ { - 0 . 5 } , \\dots , 1 0 ^ { 3 } \\} } } \\\\ { { \\eta \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \\dots , 1 0 ^ { 1 . 5 } \\} } } \\end{array}", + "type": "interline_equation", + "image_path": "aa9cd39d8e20a456a48e55a688be30e9d5d8267907f22e9a98cf561f51a602ca.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 233, + 367, + 249.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 243, + 249.5, + 367, + 266.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 272, + 199, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 271, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 200, + 285 + ], + "score": 1.0, + "content": "StackOverflow NWP:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 280, + 369, + 314 + ], + "lines": [ + { + "bbox": [ + 241, + 280, + 369, + 314 + ], + "spans": [ + { + "bbox": [ + 241, + 280, + 369, + 314 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "7014d8516c84e3d145a877f3a8396b4672e15c2311e6cb4ec613ffe7b3479892.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 280, + 369, + 297.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 241, + 297.0, + 369, + 314.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 257, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 258, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 194, + 333 + ], + "score": 1.0, + "content": "For all tasks, we tune", + "type": "text" + }, + { + "bbox": [ + 194, + 323, + 201, + 330 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 318, + 258, + 333 + ], + "score": 1.0, + "content": "over the grid:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 336, + 353, + 352 + ], + "lines": [ + { + "bbox": [ + 257, + 336, + 353, + 352 + ], + "spans": [ + { + "bbox": [ + 257, + 336, + 353, + 352 + ], + "score": 0.9, + "content": "\\tau \\in \\{ 1 0 ^ { - 5 } , \\ldots , 1 0 ^ { - 1 } \\} .", + "type": "interline_equation", + "image_path": "99916ba71eb9cc55ca05ba10dacb5c7da17683c6eea6a8fa7ac8d86b868fbfc5.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 257, + 336, + 353, + 352 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 363, + 234, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 235, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 235, + 375 + ], + "score": 1.0, + "content": "D.3 PER-TASK BATCH SIZES", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 427 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 506, + 398 + ], + "score": 1.0, + "content": "Given the large number of hyperparameters to tune, and to avoid conflating variables, we fix the batch", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 393, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 408 + ], + "score": 1.0, + "content": "size at a per-task level. When comparing centralized training to federated training in Section 5, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 404, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 418 + ], + "score": 1.0, + "content": "use the same batch size in both federated and centralized training. A full summary of the batch sizes", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 416, + 185, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 185, + 428 + ], + "score": 1.0, + "content": "is given in Table 7.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "table", + "bbox": [ + 220, + 469, + 390, + 569 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 214, + 447, + 396, + 459 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 213, + 446, + 397, + 459 + ], + "spans": [ + { + "bbox": [ + 213, + 446, + 397, + 459 + ], + "score": 1.0, + "content": "Table 7: Client batch sizes used for each task.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "table_body", + "bbox": [ + 220, + 469, + 390, + 569 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 220, + 469, + 390, + 569 + ], + "spans": [ + { + "bbox": [ + 220, + 469, + 390, + 569 + ], + "score": 0.98, + "html": "
TASKBATCH SIZE
CIFAR-1020
CIFAR-10020
EMNISTAE20
EMNISTCR20
SHAKESPEARE4
STACKOVERFLOWLR100
STACKOVERFLOW NWP16
", + "type": "table", + "image_path": "6c7462e1397fb086dda2f8cdd601dcf42bfb28989d5b6963c3977419b97a70e3.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 220, + 469, + 390, + 483.2857142857143 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 220, + 483.2857142857143, + 390, + 497.57142857142856 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 220, + 497.57142857142856, + 390, + 511.85714285714283 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 220, + 511.85714285714283, + 390, + 526.1428571428571 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 220, + 526.1428571428571, + 390, + 540.4285714285714 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 220, + 540.4285714285714, + 390, + 554.7142857142858 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 220, + 554.7142857142858, + 390, + 569.0000000000001 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 24.0 + }, + { + "type": "title", + "bbox": [ + 108, + 586, + 303, + 597 + ], + "lines": [ + { + "bbox": [ + 106, + 586, + 304, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 304, + 598 + ], + "score": 1.0, + "content": "D.4 BEST PERFORMING HYPERPARAMETERS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "In this section, we present, for each optimizer, the best client and server learning rates and values", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 117, + 629 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 619, + 124, + 627 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "found for the tasks discussed in Section 5. Specifically, these are the hyperparameters used in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 626, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 642 + ], + "score": 1.0, + "content": "Figure Figure 1 and table Table 1. The validation metrics in Table 1 are obtained using the learning", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 237, + 651 + ], + "score": 1.0, + "content": "rates in Table 8 and the values of", + "type": "text" + }, + { + "bbox": [ + 238, + 641, + 245, + 649 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 638, + 438, + 651 + ], + "score": 1.0, + "content": "in Table 9. As discussed in Section 4, we choose", + "type": "text" + }, + { + "bbox": [ + 438, + 640, + 457, + 650 + ], + "score": 0.86, + "content": "\\eta , \\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 638, + 475, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 475, + 641, + 482, + 649 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "to be", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 650, + 492, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 492, + 662 + ], + "score": 1.0, + "content": "the parameters that minimizer the average training loss over the last 100 communication rounds.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + } + ], + "page_idx": 27, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "28", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 166, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 167, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 167, + 96 + ], + "score": 1.0, + "content": "EMNIST AE:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 83, + 167, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 90, + 366, + 123 + ], + "lines": [ + { + "bbox": [ + 244, + 90, + 366, + 123 + ], + "spans": [ + { + "bbox": [ + 244, + 90, + 366, + 123 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 1 . 5 } , 1 0 ^ { - 1 } , \\dots , 1 0 ^ { 2 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \\dots , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "b565073787858f2a311a5ab65a60e1e3bf5a43a684fe4c409e2b4a2a22d64f72.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 244, + 90, + 366, + 106.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 244, + 106.5, + 366, + 123.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 167, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 130, + 168, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 168, + 144 + ], + "score": 1.0, + "content": "EMNIST CR:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 130, + 168, + 144 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 138, + 369, + 171 + ], + "lines": [ + { + "bbox": [ + 242, + 138, + 369, + 171 + ], + "spans": [ + { + "bbox": [ + 242, + 138, + 369, + 171 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "5829fa2b6c99ae68d686975dd4e00cba1ae1f7acf303ca8632112c0596ed9480.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 242, + 138, + 369, + 154.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 242, + 154.5, + 369, + 171.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 177, + 164, + 189 + ], + "lines": [ + { + "bbox": [ + 105, + 176, + 165, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 165, + 191 + ], + "score": 1.0, + "content": "Shakespeare:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 176, + 165, + 191 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 186, + 369, + 218 + ], + "lines": [ + { + "bbox": [ + 242, + 186, + 369, + 218 + ], + "spans": [ + { + "bbox": [ + 242, + 186, + 369, + 218 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "1925b4444836ff6a20b1d96a8b03bc1e2991f89d4e3a73cc9ae5823d393261c2.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 242, + 186, + 369, + 202.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 242, + 202.0, + 369, + 218.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 225, + 190, + 236 + ], + "lines": [ + { + "bbox": [ + 105, + 223, + 191, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 191, + 238 + ], + "score": 1.0, + "content": "StackOverflow LR:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 233, + 367, + 266 + ], + "lines": [ + { + "bbox": [ + 243, + 233, + 367, + 266 + ], + "spans": [ + { + "bbox": [ + 243, + 233, + 367, + 266 + ], + "score": 0.9, + "content": "\\begin{array} { c } { { \\eta _ { l } \\in \\{ 1 0 ^ { - 1 } , 1 0 ^ { - 0 . 5 } , \\dots , 1 0 ^ { 3 } \\} } } \\\\ { { \\eta \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \\dots , 1 0 ^ { 1 . 5 } \\} } } \\end{array}", + "type": "interline_equation", + "image_path": "aa9cd39d8e20a456a48e55a688be30e9d5d8267907f22e9a98cf561f51a602ca.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 233, + 367, + 249.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 243, + 249.5, + 367, + 266.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 272, + 199, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 271, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 200, + 285 + ], + "score": 1.0, + "content": "StackOverflow NWP:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 280, + 369, + 314 + ], + "lines": [ + { + "bbox": [ + 241, + 280, + 369, + 314 + ], + "spans": [ + { + "bbox": [ + 241, + 280, + 369, + 314 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}", + "type": "interline_equation", + "image_path": "7014d8516c84e3d145a877f3a8396b4672e15c2311e6cb4ec613ffe7b3479892.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 280, + 369, + 297.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 241, + 297.0, + 369, + 314.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 257, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 258, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 194, + 333 + ], + "score": 1.0, + "content": "For all tasks, we tune", + "type": "text" + }, + { + "bbox": [ + 194, + 323, + 201, + 330 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 318, + 258, + 333 + ], + "score": 1.0, + "content": "over the grid:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 318, + 258, + 333 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 336, + 353, + 352 + ], + "lines": [ + { + "bbox": [ + 257, + 336, + 353, + 352 + ], + "spans": [ + { + "bbox": [ + 257, + 336, + 353, + 352 + ], + "score": 0.9, + "content": "\\tau \\in \\{ 1 0 ^ { - 5 } , \\ldots , 1 0 ^ { - 1 } \\} .", + "type": "interline_equation", + "image_path": "99916ba71eb9cc55ca05ba10dacb5c7da17683c6eea6a8fa7ac8d86b868fbfc5.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 257, + 336, + 353, + 352 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 363, + 234, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 235, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 235, + 375 + ], + "score": 1.0, + "content": "D.3 PER-TASK BATCH SIZES", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 427 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 506, + 398 + ], + "score": 1.0, + "content": "Given the large number of hyperparameters to tune, and to avoid conflating variables, we fix the batch", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 393, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 408 + ], + "score": 1.0, + "content": "size at a per-task level. When comparing centralized training to federated training in Section 5, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 404, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 418 + ], + "score": 1.0, + "content": "use the same batch size in both federated and centralized training. A full summary of the batch sizes", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 416, + 185, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 185, + 428 + ], + "score": 1.0, + "content": "is given in Table 7.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 383, + 506, + 428 + ] + }, + { + "type": "table", + "bbox": [ + 220, + 469, + 390, + 569 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 214, + 447, + 396, + 459 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 213, + 446, + 397, + 459 + ], + "spans": [ + { + "bbox": [ + 213, + 446, + 397, + 459 + ], + "score": 1.0, + "content": "Table 7: Client batch sizes used for each task.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "table_body", + "bbox": [ + 220, + 469, + 390, + 569 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 220, + 469, + 390, + 569 + ], + "spans": [ + { + "bbox": [ + 220, + 469, + 390, + 569 + ], + "score": 0.98, + "html": "
TASKBATCH SIZE
CIFAR-1020
CIFAR-10020
EMNISTAE20
EMNISTCR20
SHAKESPEARE4
STACKOVERFLOWLR100
STACKOVERFLOW NWP16
", + "type": "table", + "image_path": "6c7462e1397fb086dda2f8cdd601dcf42bfb28989d5b6963c3977419b97a70e3.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 220, + 469, + 390, + 483.2857142857143 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 220, + 483.2857142857143, + 390, + 497.57142857142856 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 220, + 497.57142857142856, + 390, + 511.85714285714283 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 220, + 511.85714285714283, + 390, + 526.1428571428571 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 220, + 526.1428571428571, + 390, + 540.4285714285714 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 220, + 540.4285714285714, + 390, + 554.7142857142858 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 220, + 554.7142857142858, + 390, + 569.0000000000001 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 24.0 + }, + { + "type": "title", + "bbox": [ + 108, + 586, + 303, + 597 + ], + "lines": [ + { + "bbox": [ + 106, + 586, + 304, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 304, + 598 + ], + "score": 1.0, + "content": "D.4 BEST PERFORMING HYPERPARAMETERS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "In this section, we present, for each optimizer, the best client and server learning rates and values", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 117, + 629 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 619, + 124, + 627 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "found for the tasks discussed in Section 5. Specifically, these are the hyperparameters used in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 626, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 642 + ], + "score": 1.0, + "content": "Figure Figure 1 and table Table 1. The validation metrics in Table 1 are obtained using the learning", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 237, + 651 + ], + "score": 1.0, + "content": "rates in Table 8 and the values of", + "type": "text" + }, + { + "bbox": [ + 238, + 641, + 245, + 649 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 638, + 438, + 651 + ], + "score": 1.0, + "content": "in Table 9. As discussed in Section 4, we choose", + "type": "text" + }, + { + "bbox": [ + 438, + 640, + 457, + 650 + ], + "score": 0.86, + "content": "\\eta , \\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 638, + 475, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 475, + 641, + 482, + 649 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "to be", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 650, + 492, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 492, + 662 + ], + "score": 1.0, + "content": "the parameters that minimizer the average training loss over the last 100 communication rounds.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 606, + 506, + 662 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 207, + 493, + 317 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 173, + 504, + 196 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 278, + 186 + ], + "score": 1.0, + "content": "Table 8: The base-10 logarithm of the client", + "type": "text" + }, + { + "bbox": [ + 279, + 174, + 293, + 185 + ], + "score": 0.76, + "content": "( \\eta _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 173, + 337, + 186 + ], + "score": 1.0, + "content": "and server", + "type": "text" + }, + { + "bbox": [ + 337, + 174, + 349, + 185 + ], + "score": 0.58, + "content": "( \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "learning rate combinations that achieve", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 185, + 431, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 431, + 198 + ], + "score": 1.0, + "content": "the accuracies from Table 1. See Appendix D.2 for a full description of the grids.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 118, + 207, + 493, + 317 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 207, + 493, + 317 + ], + "spans": [ + { + "bbox": [ + 118, + 207, + 493, + 317 + ], + "score": 0.981, + "html": "
FEDADAGRADFEDADAMFEDYOGIFEDAVGMFEDAVG
nnmnmnmnmn
CIFAR-10-3/2-1-3/2-2-3/2-2-3/2-12-120
CIFAR-100-1-1-3/20-3/20-3/20-11
EMNISTAE3/2-3/21-3/21-3/2/010
EMNIST CR-3/2-1-3/2-5/2-3/2-5/2-3/2-1/2-10
SHAKESPEARE0-1/20-20-20-1/200
STACKOVERFLOWLR212-1/22-1/22020
STACKOVERFLOW NWP -1/2-3/2-1/2-2-1/2-2-1/20-1/20
", + "type": "table", + "image_path": "18f4d3def369eb284a8c94b42035a6239166b092d7473e6f6b782d8c4f16463f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 118, + 207, + 493, + 243.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 118, + 243.66666666666666, + 493, + 280.3333333333333 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 118, + 280.3333333333333, + 493, + 317.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 165, + 541, + 445, + 640 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 507, + 505, + 529 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 311, + 520 + ], + "score": 1.0, + "content": "Table 9: The base-10 logarithm of the parameter", + "type": "text" + }, + { + "bbox": [ + 311, + 509, + 318, + 517 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 506, + 505, + 520 + ], + "score": 1.0, + "content": "(as defined in Algorithm 2) that achieve the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 517, + 224, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 224, + 529 + ], + "score": 1.0, + "content": "validation metrics in Table 1.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "table_body", + "bbox": [ + 165, + 541, + 445, + 640 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 165, + 541, + 445, + 640 + ], + "spans": [ + { + "bbox": [ + 165, + 541, + 445, + 640 + ], + "score": 0.977, + "html": "
FEDADAGRAD FEDADAM FEDYOGI
CIFAR-10-2-3-3
CIFAR-100-2-1-1
EMNISTAE-3-3-3
EMNIST CR-2-4-4
SHAKESPEARE-1-3-3
STACKOVERFLOWLR-2-5-5
STACKOVERFLOWNWP-4-5-5
", + "type": "table", + "image_path": "b46f78a535bf0ba9996c36dce6049238dc97473b3dafec87f773b3016281aa05.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 165, + 541, + 445, + 574.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 165, + 574.0, + 445, + 607.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 165, + 607.0, + 445, + 640.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.75 + } + ], + "page_idx": 28, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 207, + 493, + 317 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 173, + 504, + 196 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 278, + 186 + ], + "score": 1.0, + "content": "Table 8: The base-10 logarithm of the client", + "type": "text" + }, + { + "bbox": [ + 279, + 174, + 293, + 185 + ], + "score": 0.76, + "content": "( \\eta _ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 173, + 337, + 186 + ], + "score": 1.0, + "content": "and server", + "type": "text" + }, + { + "bbox": [ + 337, + 174, + 349, + 185 + ], + "score": 0.58, + "content": "( \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "learning rate combinations that achieve", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 185, + 431, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 431, + 198 + ], + "score": 1.0, + "content": "the accuracies from Table 1. See Appendix D.2 for a full description of the grids.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 118, + 207, + 493, + 317 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 207, + 493, + 317 + ], + "spans": [ + { + "bbox": [ + 118, + 207, + 493, + 317 + ], + "score": 0.981, + "html": "
FEDADAGRADFEDADAMFEDYOGIFEDAVGMFEDAVG
nnmnmnmnmn
CIFAR-10-3/2-1-3/2-2-3/2-2-3/2-12-120
CIFAR-100-1-1-3/20-3/20-3/20-11
EMNISTAE3/2-3/21-3/21-3/2/010
EMNIST CR-3/2-1-3/2-5/2-3/2-5/2-3/2-1/2-10
SHAKESPEARE0-1/20-20-20-1/200
STACKOVERFLOWLR212-1/22-1/22020
STACKOVERFLOW NWP -1/2-3/2-1/2-2-1/2-2-1/20-1/20
", + "type": "table", + "image_path": "18f4d3def369eb284a8c94b42035a6239166b092d7473e6f6b782d8c4f16463f.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 118, + 207, + 493, + 243.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 118, + 243.66666666666666, + 493, + 280.3333333333333 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 118, + 280.3333333333333, + 493, + 317.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 165, + 541, + 445, + 640 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 507, + 505, + 529 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 506, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 311, + 520 + ], + "score": 1.0, + "content": "Table 9: The base-10 logarithm of the parameter", + "type": "text" + }, + { + "bbox": [ + 311, + 509, + 318, + 517 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 506, + 505, + 520 + ], + "score": 1.0, + "content": "(as defined in Algorithm 2) that achieve the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 517, + 224, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 224, + 529 + ], + "score": 1.0, + "content": "validation metrics in Table 1.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "table_body", + "bbox": [ + 165, + 541, + 445, + 640 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 165, + 541, + 445, + 640 + ], + "spans": [ + { + "bbox": [ + 165, + 541, + 445, + 640 + ], + "score": 0.977, + "html": "
FEDADAGRAD FEDADAM FEDYOGI
CIFAR-10-2-3-3
CIFAR-100-2-1-1
EMNISTAE-3-3-3
EMNIST CR-2-4-4
SHAKESPEARE-1-3-3
STACKOVERFLOWLR-2-5-5
STACKOVERFLOWNWP-4-5-5
", + "type": "table", + "image_path": "b46f78a535bf0ba9996c36dce6049238dc97473b3dafec87f773b3016281aa05.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 165, + 541, + 445, + 574.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 165, + 574.0, + 445, + 607.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 165, + 607.0, + 445, + 640.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 214, + 89, + 394, + 220 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 214, + 89, + 394, + 220 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 214, + 89, + 394, + 220 + ], + "spans": [ + { + "bbox": [ + 214, + 89, + 394, + 220 + ], + "score": 0.969, + "type": "image", + "image_path": "2cfd7a23d061067471c578b5c9cb5c0fbe27e8391814637353f1209c09dbb3a9.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 89, + 394, + 102.1 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 214, + 102.1, + 394, + 115.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 214, + 115.19999999999999, + 394, + 128.29999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 214, + 128.29999999999998, + 394, + 141.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 214, + 141.39999999999998, + 394, + 154.49999999999997 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 214, + 154.49999999999997, + 394, + 167.59999999999997 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 214, + 167.59999999999997, + 394, + 180.69999999999996 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 214, + 180.69999999999996, + 394, + 193.79999999999995 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 214, + 193.79999999999995, + 394, + 206.89999999999995 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 214, + 206.89999999999995, + 394, + 219.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 241, + 505, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 419, + 255 + ], + "score": 1.0, + "content": "Figure 4: Validation accuracy on EMNIST CR using constant learning rates", + "type": "text" + }, + { + "bbox": [ + 419, + 244, + 439, + 253 + ], + "score": 0.39, + "content": "\\eta , \\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 241, + 460, + 255 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 461, + 244, + 468, + 252 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "tuned to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "achieve the best training performance on the last 100 communication rounds; see Appendix D for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 263, + 196, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 196, + 276 + ], + "score": 1.0, + "content": "hyperparameter grids.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + } + ], + "index": 7.75 + }, + { + "type": "table", + "bbox": [ + 169, + 333, + 441, + 384 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 299, + 505, + 321 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "Table 10: Test set performance for Stack Overflow tasks after training: Accuracy for NWP and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 309, + 490, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 133, + 321 + ], + "score": 1.0, + "content": "Recall", + "type": "text" + }, + { + "bbox": [ + 133, + 310, + 148, + 320 + ], + "score": 0.55, + "content": "\\textcircled { a } 5", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 310, + 180, + 321 + ], + "score": 0.76, + "content": "( \\times 1 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 309, + 322, + 321 + ], + "score": 1.0, + "content": "for LR. Performance within within", + "type": "text" + }, + { + "bbox": [ + 323, + 309, + 345, + 320 + ], + "score": 0.88, + "content": "0 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 309, + 490, + 321 + ], + "score": 1.0, + "content": "of the best result are shown in bold.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "table_body", + "bbox": [ + 169, + 333, + 441, + 384 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 333, + 441, + 384 + ], + "spans": [ + { + "bbox": [ + 169, + 333, + 441, + 384 + ], + "score": 0.975, + "html": "
FED...ADAGRADADAMYOGIAvGMAvG
STACK OVERFLOW NWP24.425.725.724.520.5
STACK OVERFLOW LR66.865.266.546.540.6
", + "type": "table", + "image_path": "22c05c52675ef94844cfd19ea818df09b50ac07abae2ea6a55b2da64c1871581.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 169, + 333, + 441, + 350.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 169, + 350.0, + 441, + 367.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 169, + 367.0, + 441, + 384.0 + ], + "spans": [], + "index": 17 + } + ] + } + ], + "index": 14.75 + }, + { + "type": "title", + "bbox": [ + 107, + 407, + 328, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 329, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 329, + 422 + ], + "score": 1.0, + "content": "E ADDITIONAL EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 431, + 245, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 246, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 246, + 444 + ], + "score": 1.0, + "content": "E.1 RESULTS ON EMNIST CR", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 507, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 507, + 466 + ], + "score": 1.0, + "content": "We plot the validation accuracy of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, FEDAVG,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 462, + 504, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 417, + 477 + ], + "score": 1.0, + "content": "and SCAFFOLD on EMNIST CR. As in Figure 1, we tune the learning rates", + "type": "text" + }, + { + "bbox": [ + 418, + 465, + 437, + 475 + ], + "score": 0.87, + "content": "\\eta _ { l } , \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 462, + 497, + 477 + ], + "score": 1.0, + "content": "and adaptivity", + "type": "text" + }, + { + "bbox": [ + 497, + 465, + 504, + 473 + ], + "score": 0.69, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 472, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 488 + ], + "score": 1.0, + "content": "by selecting the parameters obtaining the smallest training loss, averaged over the last 100 training", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 485, + 272, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 272, + 497 + ], + "score": 1.0, + "content": "rounds. The results are given in Figure 4.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 501, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "We see that all methods are roughly comparably throughout the duration of training. This reflects the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "fact that the dataset is quite simple, and most clients have all classes in their local datasets, reducing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "any heterogeneity among classes. Note that SCAFFOLD performs slightly worse than FEDAVG", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "and all other methods here. As discussed in Section 5, this may be due to the presence of stale client", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 545, + 410, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 410, + 556 + ], + "score": 1.0, + "content": "control variates, and the communication-limited regime of our experiments.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 107, + 569, + 322, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 322, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 322, + 581 + ], + "score": 1.0, + "content": "E.2 STACK OVERFLOW TEST SET PERFORMANCE", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 589, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "As discussed in Section 5, in order to compute a measure of performance for the Stack Overflow", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "tasks as training progresses, we use a subsampled version of the test dataset, due to its prohibitively", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "large number of clients and examples. In particular, at each round of training, we sample 10,000", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "random test samples, and use this as a measure of performance over time. However, once training is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "completed, we also evaluate on the full test dataset. For the Stack Overflow experiments described in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 642, + 321, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 321, + 655 + ], + "score": 1.0, + "content": "Section 5, the final test accuracy is given in Table 10.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 108, + 668, + 261, + 679 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 262, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 262, + 680 + ], + "score": 1.0, + "content": "E.3 LEARNING RATE ROBUSTNESS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 688, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 504, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 390, + 702 + ], + "score": 1.0, + "content": "In this section, we showcase what combinations of client learning rate", + "type": "text" + }, + { + "bbox": [ + 390, + 691, + 399, + 700 + ], + "score": 0.8, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 686, + 497, + 702 + ], + "score": 1.0, + "content": "and server learning rate", + "type": "text" + }, + { + "bbox": [ + 497, + 691, + 504, + 700 + ], + "score": 0.7, + "content": "\\eta", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "performed well for each optimizer and task. As in Figure 2, we plot, for each optimizer, task, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 709, + 125, + 723 + ], + "score": 1.0, + "content": "pair", + "type": "text" + }, + { + "bbox": [ + 125, + 710, + 151, + 722 + ], + "score": 0.89, + "content": "( \\eta _ { l } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 709, + 506, + 723 + ], + "score": 1.0, + "content": ", the validation set performance (averaged over the last 100 rounds). As in Section 5, we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 719, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 119, + 733 + ], + "score": 1.0, + "content": "fix", + "type": "text" + }, + { + "bbox": [ + 120, + 721, + 160, + 731 + ], + "score": 0.89, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 719, + 506, + 733 + ], + "score": 1.0, + "content": "throughout. The results, for the CIFAR-10, CIFAR-100, EMNIST AE, EMNIST CR,", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + } + ], + "page_idx": 29, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "30", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 214, + 89, + 394, + 220 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 214, + 89, + 394, + 220 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 214, + 89, + 394, + 220 + ], + "spans": [ + { + "bbox": [ + 214, + 89, + 394, + 220 + ], + "score": 0.969, + "type": "image", + "image_path": "2cfd7a23d061067471c578b5c9cb5c0fbe27e8391814637353f1209c09dbb3a9.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 89, + 394, + 102.1 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 214, + 102.1, + 394, + 115.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 214, + 115.19999999999999, + 394, + 128.29999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 214, + 128.29999999999998, + 394, + 141.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 214, + 141.39999999999998, + 394, + 154.49999999999997 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 214, + 154.49999999999997, + 394, + 167.59999999999997 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 214, + 167.59999999999997, + 394, + 180.69999999999996 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 214, + 180.69999999999996, + 394, + 193.79999999999995 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 214, + 193.79999999999995, + 394, + 206.89999999999995 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 214, + 206.89999999999995, + 394, + 219.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 241, + 505, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 419, + 255 + ], + "score": 1.0, + "content": "Figure 4: Validation accuracy on EMNIST CR using constant learning rates", + "type": "text" + }, + { + "bbox": [ + 419, + 244, + 439, + 253 + ], + "score": 0.39, + "content": "\\eta , \\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 241, + 460, + 255 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 461, + 244, + 468, + 252 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "tuned to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 264 + ], + "score": 1.0, + "content": "achieve the best training performance on the last 100 communication rounds; see Appendix D for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 263, + 196, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 196, + 276 + ], + "score": 1.0, + "content": "hyperparameter grids.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + } + ], + "index": 7.75 + }, + { + "type": "table", + "bbox": [ + 169, + 333, + 441, + 384 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 299, + 505, + 321 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 311 + ], + "score": 1.0, + "content": "Table 10: Test set performance for Stack Overflow tasks after training: Accuracy for NWP and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 309, + 490, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 133, + 321 + ], + "score": 1.0, + "content": "Recall", + "type": "text" + }, + { + "bbox": [ + 133, + 310, + 148, + 320 + ], + "score": 0.55, + "content": "\\textcircled { a } 5", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 310, + 180, + 321 + ], + "score": 0.76, + "content": "( \\times 1 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 309, + 322, + 321 + ], + "score": 1.0, + "content": "for LR. Performance within within", + "type": "text" + }, + { + "bbox": [ + 323, + 309, + 345, + 320 + ], + "score": 0.88, + "content": "0 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 309, + 490, + 321 + ], + "score": 1.0, + "content": "of the best result are shown in bold.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "table_body", + "bbox": [ + 169, + 333, + 441, + 384 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 333, + 441, + 384 + ], + "spans": [ + { + "bbox": [ + 169, + 333, + 441, + 384 + ], + "score": 0.975, + "html": "
FED...ADAGRADADAMYOGIAvGMAvG
STACK OVERFLOW NWP24.425.725.724.520.5
STACK OVERFLOW LR66.865.266.546.540.6
", + "type": "table", + "image_path": "22c05c52675ef94844cfd19ea818df09b50ac07abae2ea6a55b2da64c1871581.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 169, + 333, + 441, + 350.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 169, + 350.0, + 441, + 367.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 169, + 367.0, + 441, + 384.0 + ], + "spans": [], + "index": 17 + } + ] + } + ], + "index": 14.75 + }, + { + "type": "title", + "bbox": [ + 107, + 407, + 328, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 329, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 329, + 422 + ], + "score": 1.0, + "content": "E ADDITIONAL EXPERIMENTAL RESULTS", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 431, + 245, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 246, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 246, + 444 + ], + "score": 1.0, + "content": "E.1 RESULTS ON EMNIST CR", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 507, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 507, + 466 + ], + "score": 1.0, + "content": "We plot the validation accuracy of FEDADAGRAD, FEDADAM, FEDYOGI, FEDAVGM, FEDAVG,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 462, + 504, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 417, + 477 + ], + "score": 1.0, + "content": "and SCAFFOLD on EMNIST CR. As in Figure 1, we tune the learning rates", + "type": "text" + }, + { + "bbox": [ + 418, + 465, + 437, + 475 + ], + "score": 0.87, + "content": "\\eta _ { l } , \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 462, + 497, + 477 + ], + "score": 1.0, + "content": "and adaptivity", + "type": "text" + }, + { + "bbox": [ + 497, + 465, + 504, + 473 + ], + "score": 0.69, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 472, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 488 + ], + "score": 1.0, + "content": "by selecting the parameters obtaining the smallest training loss, averaged over the last 100 training", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 485, + 272, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 272, + 497 + ], + "score": 1.0, + "content": "rounds. The results are given in Figure 4.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 451, + 507, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 501, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "We see that all methods are roughly comparably throughout the duration of training. This reflects the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "fact that the dataset is quite simple, and most clients have all classes in their local datasets, reducing", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "any heterogeneity among classes. Note that SCAFFOLD performs slightly worse than FEDAVG", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "and all other methods here. As discussed in Section 5, this may be due to the presence of stale client", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 545, + 410, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 410, + 556 + ], + "score": 1.0, + "content": "control variates, and the communication-limited regime of our experiments.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 502, + 506, + 556 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 569, + 322, + 580 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 322, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 322, + 581 + ], + "score": 1.0, + "content": "E.2 STACK OVERFLOW TEST SET PERFORMANCE", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 589, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "As discussed in Section 5, in order to compute a measure of performance for the Stack Overflow", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "tasks as training progresses, we use a subsampled version of the test dataset, due to its prohibitively", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "large number of clients and examples. In particular, at each round of training, we sample 10,000", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "random test samples, and use this as a measure of performance over time. However, once training is", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "completed, we also evaluate on the full test dataset. For the Stack Overflow experiments described in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 642, + 321, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 321, + 655 + ], + "score": 1.0, + "content": "Section 5, the final test accuracy is given in Table 10.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 589, + 506, + 655 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 668, + 261, + 679 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 262, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 262, + 680 + ], + "score": 1.0, + "content": "E.3 LEARNING RATE ROBUSTNESS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 688, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 504, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 390, + 702 + ], + "score": 1.0, + "content": "In this section, we showcase what combinations of client learning rate", + "type": "text" + }, + { + "bbox": [ + 390, + 691, + 399, + 700 + ], + "score": 0.8, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 686, + 497, + 702 + ], + "score": 1.0, + "content": "and server learning rate", + "type": "text" + }, + { + "bbox": [ + 497, + 691, + 504, + 700 + ], + "score": 0.7, + "content": "\\eta", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "performed well for each optimizer and task. As in Figure 2, we plot, for each optimizer, task, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 709, + 125, + 723 + ], + "score": 1.0, + "content": "pair", + "type": "text" + }, + { + "bbox": [ + 125, + 710, + 151, + 722 + ], + "score": 0.89, + "content": "( \\eta _ { l } , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 709, + 506, + 723 + ], + "score": 1.0, + "content": ", the validation set performance (averaged over the last 100 rounds). As in Section 5, we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 719, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 119, + 733 + ], + "score": 1.0, + "content": "fix", + "type": "text" + }, + { + "bbox": [ + 120, + 721, + 160, + 731 + ], + "score": 0.89, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 719, + 506, + 733 + ], + "score": 1.0, + "content": "throughout. The results, for the CIFAR-10, CIFAR-100, EMNIST AE, EMNIST CR,", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5, + "bbox_fs": [ + 104, + 686, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 507, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 507, + 96 + ], + "score": 1.0, + "content": "Shakespeare, Stack Overflow LR, and Stack Overflow NWP tasks are given in Figures 5, 6, 7, 8, 9,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 205, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 205, + 106 + ], + "score": 1.0, + "content": "10, and 11, respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 122 + ], + "score": 1.0, + "content": "While the general trends depend on the optimizer and task, we see that in many cases, the adaptive", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "score": 1.0, + "content": "methods have rectangular regions that perform well. This implies a kind of robustness to fixing one", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 118, + 144 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 133, + 136, + 143 + ], + "score": 0.87, + "content": "\\eta , \\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 131, + 506, + 144 + ], + "score": 1.0, + "content": ", and varying the other. On the other hand, FEDAVGM and FEDAVG often have triangular", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 441, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 275, + 155 + ], + "score": 1.0, + "content": "regions that perform well, suggesting that", + "type": "text" + }, + { + "bbox": [ + 275, + 144, + 282, + 154 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 143, + 299, + 155 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 144, + 308, + 154 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 143, + 441, + 155 + ], + "score": 1.0, + "content": "should be tuned simultaneously.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 125, + 164, + 488, + 380 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 164, + 488, + 380 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 164, + 488, + 380 + ], + "spans": [ + { + "bbox": [ + 125, + 164, + 488, + 380 + ], + "score": 0.973, + "type": "image", + "image_path": "3c228a3a0a506f7fab5984fad48909c5f19d018f5100b7319cb0ea2a5cd5bbad.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 125, + 164, + 488, + 236.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 125, + 236.0, + 488, + 308.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 125, + 308.0, + 488, + 380.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 396, + 506, + 429 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 395, + 507, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 507, + 409 + ], + "score": 1.0, + "content": "Figure 5: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 417, + 439, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 393, + 429 + ], + "score": 1.0, + "content": "CIFAR-10 task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 394, + 417, + 434, + 428 + ], + "score": 0.91, + "content": "\\bar { \\tau } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 417, + 439, + 429 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ], + "page_idx": 30, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 507, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 507, + 96 + ], + "score": 1.0, + "content": "Shakespeare, Stack Overflow LR, and Stack Overflow NWP tasks are given in Figures 5, 6, 7, 8, 9,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 205, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 205, + 106 + ], + "score": 1.0, + "content": "10, and 11, respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 507, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 122 + ], + "score": 1.0, + "content": "While the general trends depend on the optimizer and task, we see that in many cases, the adaptive", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "score": 1.0, + "content": "methods have rectangular regions that perform well. This implies a kind of robustness to fixing one", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 118, + 144 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 118, + 133, + 136, + 143 + ], + "score": 0.87, + "content": "\\eta , \\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 131, + 506, + 144 + ], + "score": 1.0, + "content": ", and varying the other. On the other hand, FEDAVGM and FEDAVG often have triangular", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 441, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 275, + 155 + ], + "score": 1.0, + "content": "regions that perform well, suggesting that", + "type": "text" + }, + { + "bbox": [ + 275, + 144, + 282, + 154 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 143, + 299, + 155 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 144, + 308, + 154 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 143, + 441, + 155 + ], + "score": 1.0, + "content": "should be tuned simultaneously.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 110, + 506, + 155 + ] + }, + { + "type": "image", + "bbox": [ + 125, + 164, + 488, + 380 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 164, + 488, + 380 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 164, + 488, + 380 + ], + "spans": [ + { + "bbox": [ + 125, + 164, + 488, + 380 + ], + "score": 0.973, + "type": "image", + "image_path": "3c228a3a0a506f7fab5984fad48909c5f19d018f5100b7319cb0ea2a5cd5bbad.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 125, + 164, + 488, + 236.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 125, + 236.0, + 488, + 308.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 125, + 308.0, + 488, + 380.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 396, + 506, + 429 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 395, + 507, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 507, + 409 + ], + "score": 1.0, + "content": "Figure 5: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 417, + 439, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 393, + 429 + ], + "score": 1.0, + "content": "CIFAR-10 task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 394, + 417, + 434, + 428 + ], + "score": 0.91, + "content": "\\bar { \\tau } = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 417, + 439, + 429 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 105, + 488, + 323 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 105, + 488, + 323 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 105, + 488, + 323 + ], + "spans": [ + { + "bbox": [ + 124, + 105, + 488, + 323 + ], + "score": 0.97, + "type": "image", + "image_path": "988976289e57389c260e2ce5ff2bcfe18be6722e20e33a23e74e0c18d86179b5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 105, + 488, + 177.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 177.66666666666669, + 488, + 250.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 250.33333333333337, + 488, + 323.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 338, + 506, + 372 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "Figure 6: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 358, + 444, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 398, + 372 + ], + "score": 1.0, + "content": "CIFAR-100 task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 399, + 360, + 439, + 370 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 358, + 444, + 372 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 123, + 443, + 488, + 643 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 443, + 488, + 643 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 123, + 443, + 488, + 643 + ], + "spans": [ + { + "bbox": [ + 123, + 443, + 488, + 643 + ], + "score": 0.956, + "type": "image", + "image_path": "a23451f198d579719bd5c12522f418c42bc52fa09bd6fee577a2a6cd9144cba7.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 123, + 443, + 488, + 509.6666666666667 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 123, + 509.6666666666667, + 488, + 576.3333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 123, + 576.3333333333334, + 488, + 643.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 666, + 506, + 700 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "Figure 7: Validation MSE (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 686, + 450, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 405, + 700 + ], + "score": 1.0, + "content": "EMNIST AE task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 405, + 687, + 445, + 698 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 686, + 450, + 700 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ], + "page_idx": 31, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 105, + 488, + 323 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 105, + 488, + 323 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 105, + 488, + 323 + ], + "spans": [ + { + "bbox": [ + 124, + 105, + 488, + 323 + ], + "score": 0.97, + "type": "image", + "image_path": "988976289e57389c260e2ce5ff2bcfe18be6722e20e33a23e74e0c18d86179b5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 105, + 488, + 177.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 177.66666666666669, + 488, + 250.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 250.33333333333337, + 488, + 323.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 338, + 506, + 372 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "Figure 6: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 358, + 444, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 398, + 372 + ], + "score": 1.0, + "content": "CIFAR-100 task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 399, + 360, + 439, + 370 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 358, + 444, + 372 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 123, + 443, + 488, + 643 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 443, + 488, + 643 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 123, + 443, + 488, + 643 + ], + "spans": [ + { + "bbox": [ + 123, + 443, + 488, + 643 + ], + "score": 0.956, + "type": "image", + "image_path": "a23451f198d579719bd5c12522f418c42bc52fa09bd6fee577a2a6cd9144cba7.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 123, + 443, + 488, + 509.6666666666667 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 123, + 509.6666666666667, + 488, + 576.3333333333334 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 123, + 576.3333333333334, + 488, + 643.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 666, + 506, + 700 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "Figure 7: Validation MSE (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 686, + 450, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 405, + 700 + ], + "score": 1.0, + "content": "EMNIST AE task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 405, + 687, + 445, + 698 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 686, + 450, + 700 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 105, + 488, + 323 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 105, + 488, + 323 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 105, + 488, + 323 + ], + "spans": [ + { + "bbox": [ + 125, + 105, + 488, + 323 + ], + "score": 0.972, + "type": "image", + "image_path": "c472049d458c0bdb90d3a9865546ea6d2a6905247c0626ad243509dc4925cb64.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 105, + 488, + 177.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 177.66666666666669, + 488, + 250.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 250.33333333333337, + 488, + 323.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 338, + 507, + 371 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "Figure 8: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 358, + 449, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 358, + 405, + 372 + ], + "score": 1.0, + "content": "EMNIST CR task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 405, + 360, + 445, + 370 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 358, + 449, + 372 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 124, + 434, + 488, + 650 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 434, + 488, + 650 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 434, + 488, + 650 + ], + "spans": [ + { + "bbox": [ + 124, + 434, + 488, + 650 + ], + "score": 0.971, + "type": "image", + "image_path": "110df1d101eebec91adcafb1b92bcfbd8eb49008a8ef880e5dd00d979aa15b32.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 124, + 434, + 488, + 506.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 124, + 506.0, + 488, + 578.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 124, + 578.0, + 488, + 650.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 666, + 505, + 700 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 665, + 507, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 507, + 679 + ], + "score": 1.0, + "content": "Figure 9: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 686, + 446, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 401, + 700 + ], + "score": 1.0, + "content": "Shakespeare task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 402, + 687, + 442, + 698 + ], + "score": 0.92, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 686, + 446, + 700 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ], + "page_idx": 32, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "33", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 105, + 488, + 323 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 105, + 488, + 323 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 105, + 488, + 323 + ], + "spans": [ + { + "bbox": [ + 125, + 105, + 488, + 323 + ], + "score": 0.972, + "type": "image", + "image_path": "c472049d458c0bdb90d3a9865546ea6d2a6905247c0626ad243509dc4925cb64.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 125, + 105, + 488, + 177.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 125, + 177.66666666666669, + 488, + 250.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 125, + 250.33333333333337, + 488, + 323.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 338, + 507, + 371 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "Figure 8: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 358, + 449, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 358, + 405, + 372 + ], + "score": 1.0, + "content": "EMNIST CR task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 405, + 360, + 445, + 370 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 358, + 449, + 372 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 124, + 434, + 488, + 650 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 434, + 488, + 650 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 434, + 488, + 650 + ], + "spans": [ + { + "bbox": [ + 124, + 434, + 488, + 650 + ], + "score": 0.971, + "type": "image", + "image_path": "110df1d101eebec91adcafb1b92bcfbd8eb49008a8ef880e5dd00d979aa15b32.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 124, + 434, + 488, + 506.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 124, + 506.0, + 488, + 578.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 124, + 578.0, + 488, + 650.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 666, + 505, + 700 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 665, + 507, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 507, + 679 + ], + "score": 1.0, + "content": "Figure 9: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 686, + 446, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 401, + 700 + ], + "score": 1.0, + "content": "Shakespeare task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 402, + 687, + 442, + 698 + ], + "score": 0.92, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 686, + 446, + 700 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 115, + 489, + 311 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 115, + 489, + 311 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 115, + 489, + 311 + ], + "spans": [ + { + "bbox": [ + 124, + 115, + 489, + 311 + ], + "score": 0.97, + "type": "image", + "image_path": "36a8c5f6ace9eb2d7ece4dab87f90356b80a7333fb4a7383edd182ee807213cb.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 115, + 489, + 180.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 180.33333333333331, + 489, + 245.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 245.66666666666663, + 489, + 310.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 330, + 506, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 330, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 218, + 342 + ], + "score": 1.0, + "content": "Figure 10: Validation recall", + "type": "text" + }, + { + "bbox": [ + 218, + 331, + 233, + 340 + ], + "score": 0.62, + "content": "\\textcircled { \\alpha } 5", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 330, + 506, + 342 + ], + "score": 1.0, + "content": "(averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Stack", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 350, + 449, + 364 + ], + "spans": [ + { + "bbox": [ + 104, + 350, + 404, + 364 + ], + "score": 1.0, + "content": "Overflow LR task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 404, + 351, + 445, + 362 + ], + "score": 0.92, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 350, + 449, + 364 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 124, + 432, + 488, + 648 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 432, + 488, + 648 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 432, + 488, + 648 + ], + "spans": [ + { + "bbox": [ + 124, + 432, + 488, + 648 + ], + "score": 0.966, + "type": "image", + "image_path": "4238061faf9efe9b6eef6f047f57581529cb3114358e9247fc74a3a2af45e1e3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 124, + 432, + 488, + 504.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 124, + 504.0, + 488, + 576.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 124, + 576.0, + 488, + 648.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 663, + 506, + 697 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 663, + 507, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 507, + 677 + ], + "score": 1.0, + "content": "Figure 11: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 673, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 506, + 687 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Stack", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 684, + 459, + 698 + ], + "spans": [ + { + "bbox": [ + 104, + 684, + 413, + 698 + ], + "score": 1.0, + "content": "Overflow NWP task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 414, + 685, + 454, + 695 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 684, + 459, + 698 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ], + "page_idx": 33, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 115, + 489, + 311 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 115, + 489, + 311 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 115, + 489, + 311 + ], + "spans": [ + { + "bbox": [ + 124, + 115, + 489, + 311 + ], + "score": 0.97, + "type": "image", + "image_path": "36a8c5f6ace9eb2d7ece4dab87f90356b80a7333fb4a7383edd182ee807213cb.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 115, + 489, + 180.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 180.33333333333331, + 489, + 245.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 245.66666666666663, + 489, + 310.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 330, + 506, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 330, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 218, + 342 + ], + "score": 1.0, + "content": "Figure 10: Validation recall", + "type": "text" + }, + { + "bbox": [ + 218, + 331, + 233, + 340 + ], + "score": 0.62, + "content": "\\textcircled { \\alpha } 5", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 330, + 506, + 342 + ], + "score": 1.0, + "content": "(averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Stack", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 350, + 449, + 364 + ], + "spans": [ + { + "bbox": [ + 104, + 350, + 404, + 364 + ], + "score": 1.0, + "content": "Overflow LR task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 404, + 351, + 445, + 362 + ], + "score": 0.92, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 350, + 449, + 364 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 124, + 432, + 488, + 648 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 432, + 488, + 648 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 432, + 488, + 648 + ], + "spans": [ + { + "bbox": [ + 124, + 432, + 488, + 648 + ], + "score": 0.966, + "type": "image", + "image_path": "4238061faf9efe9b6eef6f047f57581529cb3114358e9247fc74a3a2af45e1e3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 124, + 432, + 488, + 504.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 124, + 504.0, + 488, + 576.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 124, + 576.0, + 488, + 648.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 663, + 506, + 697 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 663, + 507, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 507, + 677 + ], + "score": 1.0, + "content": "Figure 11: Validation accuracy (averaged over the last 100 rounds) of FEDADAGRAD, FEDADAM,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 673, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 506, + 687 + ], + "score": 1.0, + "content": "FEDYOGI, FEDAVGM, and FEDAVG for various client/server learning rates combination on the Stack", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 684, + 459, + 698 + ], + "spans": [ + { + "bbox": [ + 104, + 684, + 413, + 698 + ], + "score": 1.0, + "content": "Overflow NWP task. For FEDADAGRAD, FEDADAM, and FEDYOGI, we set", + "type": "text" + }, + { + "bbox": [ + 414, + 685, + 454, + 695 + ], + "score": 0.91, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 684, + 459, + 698 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "In order to better understand the results in Appendix E.3, we plot the relation between optimal choices", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 117, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 444, + 130 + ], + "score": 1.0, + "content": "of client and server learning rates. For each optimizer, task, and client learning rate", + "type": "text" + }, + { + "bbox": [ + 444, + 120, + 453, + 129 + ], + "score": 0.8, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 117, + 505, + 130 + ], + "score": 1.0, + "content": ", we find the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 128, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 265, + 141 + ], + "score": 1.0, + "content": "best corresponding server learning rate", + "type": "text" + }, + { + "bbox": [ + 266, + 131, + 273, + 140 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 128, + 506, + 141 + ], + "score": 1.0, + "content": "among the grids listed in Appendix D.2. Throughout, we", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 120, + 153 + ], + "score": 1.0, + "content": "fix", + "type": "text" + }, + { + "bbox": [ + 120, + 139, + 162, + 149 + ], + "score": 0.92, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 137, + 506, + 153 + ], + "score": 1.0, + "content": "for the adaptive methods. We omit any points for which the final validation loss is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 133, + 162 + ], + "score": 1.0, + "content": "within", + "type": "text" + }, + { + "bbox": [ + 134, + 150, + 153, + 160 + ], + "score": 0.84, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "of the worst-recorded validation loss over all hyperparameters. Essentially, we omit client", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 160, + 484, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 484, + 173 + ], + "score": 1.0, + "content": "learning rates that did not lead to any training of the model. The results are given in Figure 12.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 120, + 196, + 490, + 476 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 196, + 490, + 476 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 196, + 490, + 476 + ], + "spans": [ + { + "bbox": [ + 120, + 196, + 490, + 476 + ], + "score": 0.977, + "type": "image", + "image_path": "50c44eab15de90c5967e9a2de54c1c02f46cb58fea87d1a6bfb848613b707602.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 120, + 196, + 490, + 289.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 120, + 289.3333333333333, + 490, + 382.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 120, + 382.66666666666663, + 490, + 475.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 495, + 506, + 549 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "Figure 12: The best server learning rate in our hyperparameter tuning grids for each client learning rate,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "optimizers, and task. We select the server learning rates based on the average validation performance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 516, + 504, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 504, + 528 + ], + "score": 1.0, + "content": "over the last 100 communication rounds. For FEDADAGRAD, FEDADAM, and FEDYOGI, we fix", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 148, + 537 + ], + "score": 0.92, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 525, + 506, + 540 + ], + "score": 1.0, + "content": ". We omit all client learning rates for which all server learning rates did not change the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 538, + 270, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 248, + 550 + ], + "score": 1.0, + "content": "initial validation loss by more than", + "type": "text" + }, + { + "bbox": [ + 248, + 538, + 267, + 548 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 538, + 270, + 550 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 450, + 588 + ], + "score": 1.0, + "content": "In virtually all tasks, we see a clear inverse relationship between client learning rate", + "type": "text" + }, + { + "bbox": [ + 450, + 577, + 459, + 587 + ], + "score": 0.81, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 574, + 506, + 588 + ], + "score": 1.0, + "content": "and server", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 586, + 504, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 158, + 597 + ], + "score": 1.0, + "content": "learning rate", + "type": "text" + }, + { + "bbox": [ + 158, + 588, + 165, + 597 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 586, + 504, + 597 + ], + "score": 1.0, + "content": "for FEDAVG and FEDAVGM. As discussed in Section 5, this supports the observation", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 242, + 609 + ], + "score": 1.0, + "content": "that for the non-adaptive methods,", + "type": "text" + }, + { + "bbox": [ + 243, + 598, + 252, + 608 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 596, + 269, + 609 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 269, + 599, + 276, + 608 + ], + "score": 0.82, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "must in some sense be tuned simultaneously. On the other", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 605, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 621 + ], + "score": 1.0, + "content": "hand, for adaptive optimizers on most tasks we see much more stability in the best server learning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 123, + 631 + ], + "score": 1.0, + "content": "rate", + "type": "text" + }, + { + "bbox": [ + 124, + 620, + 130, + 629 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 618, + 232, + 631 + ], + "score": 1.0, + "content": "as the client learning rate", + "type": "text" + }, + { + "bbox": [ + 232, + 620, + 241, + 629 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "varies. This supports our observation in Section 5 that for adaptive", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 630, + 340, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 340, + 641 + ], + "score": 1.0, + "content": "methods, tuning the client learning rate is more important.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 645, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 646, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 657 + ], + "score": 1.0, + "content": "Notably, we see a clear exception to this on the Stack Overflow LR task, where there is a definitive", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "score": 1.0, + "content": "inverse relationship between learning rates among all optimizers. The EMNIST AE task also displays", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "somewhat different behavior. While there are still noticeable inverse relationships between learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "rates for FEDAVG and FEDAVGM, the range of good client learning rates is relatively small. We", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "emphasize that this task is fundamentally different than the remaining tasks. As noted by Zaheer et al.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "(2018), the primary obstacle in training bottleneck autoencoders is escaping saddle points, not in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "converging to critical points. Thus, we expect there to be qualitative differences between EMNIST", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 721, + 385, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 385, + 732 + ], + "score": 1.0, + "content": "AE and other tasks, even EMNIST CR (which uses the same dataset).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + } + ], + "page_idx": 34, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "35", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 112, + 83, + 416, + 93 + ], + "lines": [ + { + "bbox": [ + 109, + 82, + 419, + 94 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 419, + 94 + ], + "score": 1.0, + "content": "E.4 ON THE RELATION BETWEEN CLIENT AND SERVER LEARNING RATE", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "In order to better understand the results in Appendix E.3, we plot the relation between optimal choices", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 117, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 444, + 130 + ], + "score": 1.0, + "content": "of client and server learning rates. For each optimizer, task, and client learning rate", + "type": "text" + }, + { + "bbox": [ + 444, + 120, + 453, + 129 + ], + "score": 0.8, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 117, + 505, + 130 + ], + "score": 1.0, + "content": ", we find the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 128, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 265, + 141 + ], + "score": 1.0, + "content": "best corresponding server learning rate", + "type": "text" + }, + { + "bbox": [ + 266, + 131, + 273, + 140 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 128, + 506, + 141 + ], + "score": 1.0, + "content": "among the grids listed in Appendix D.2. Throughout, we", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 120, + 153 + ], + "score": 1.0, + "content": "fix", + "type": "text" + }, + { + "bbox": [ + 120, + 139, + 162, + 149 + ], + "score": 0.92, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 137, + 506, + 153 + ], + "score": 1.0, + "content": "for the adaptive methods. We omit any points for which the final validation loss is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 150, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 133, + 162 + ], + "score": 1.0, + "content": "within", + "type": "text" + }, + { + "bbox": [ + 134, + 150, + 153, + 160 + ], + "score": 0.84, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 150, + 505, + 162 + ], + "score": 1.0, + "content": "of the worst-recorded validation loss over all hyperparameters. Essentially, we omit client", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 160, + 484, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 484, + 173 + ], + "score": 1.0, + "content": "learning rates that did not lead to any training of the model. The results are given in Figure 12.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 106, + 506, + 173 + ] + }, + { + "type": "image", + "bbox": [ + 120, + 196, + 490, + 476 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 196, + 490, + 476 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 196, + 490, + 476 + ], + "spans": [ + { + "bbox": [ + 120, + 196, + 490, + 476 + ], + "score": 0.977, + "type": "image", + "image_path": "50c44eab15de90c5967e9a2de54c1c02f46cb58fea87d1a6bfb848613b707602.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 120, + 196, + 490, + 289.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 120, + 289.3333333333333, + 490, + 382.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 120, + 382.66666666666663, + 490, + 475.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 495, + 506, + 549 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "Figure 12: The best server learning rate in our hyperparameter tuning grids for each client learning rate,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "optimizers, and task. We select the server learning rates based on the average validation performance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 516, + 504, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 504, + 528 + ], + "score": 1.0, + "content": "over the last 100 communication rounds. For FEDADAGRAD, FEDADAM, and FEDYOGI, we fix", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 148, + 537 + ], + "score": 0.92, + "content": "\\tau = 1 0 ^ { - 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 525, + 506, + 540 + ], + "score": 1.0, + "content": ". We omit all client learning rates for which all server learning rates did not change the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 538, + 270, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 248, + 550 + ], + "score": 1.0, + "content": "initial validation loss by more than", + "type": "text" + }, + { + "bbox": [ + 248, + 538, + 267, + 548 + ], + "score": 0.86, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 538, + 270, + 550 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 450, + 588 + ], + "score": 1.0, + "content": "In virtually all tasks, we see a clear inverse relationship between client learning rate", + "type": "text" + }, + { + "bbox": [ + 450, + 577, + 459, + 587 + ], + "score": 0.81, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 574, + 506, + 588 + ], + "score": 1.0, + "content": "and server", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 586, + 504, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 158, + 597 + ], + "score": 1.0, + "content": "learning rate", + "type": "text" + }, + { + "bbox": [ + 158, + 588, + 165, + 597 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 586, + 504, + 597 + ], + "score": 1.0, + "content": "for FEDAVG and FEDAVGM. As discussed in Section 5, this supports the observation", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 242, + 609 + ], + "score": 1.0, + "content": "that for the non-adaptive methods,", + "type": "text" + }, + { + "bbox": [ + 243, + 598, + 252, + 608 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 596, + 269, + 609 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 269, + 599, + 276, + 608 + ], + "score": 0.82, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "must in some sense be tuned simultaneously. On the other", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 605, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 621 + ], + "score": 1.0, + "content": "hand, for adaptive optimizers on most tasks we see much more stability in the best server learning", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 123, + 631 + ], + "score": 1.0, + "content": "rate", + "type": "text" + }, + { + "bbox": [ + 124, + 620, + 130, + 629 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 618, + 232, + 631 + ], + "score": 1.0, + "content": "as the client learning rate", + "type": "text" + }, + { + "bbox": [ + 232, + 620, + 241, + 629 + ], + "score": 0.82, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "varies. This supports our observation in Section 5 that for adaptive", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 630, + 340, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 340, + 641 + ], + "score": 1.0, + "content": "methods, tuning the client learning rate is more important.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 574, + 506, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 645, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 646, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 657 + ], + "score": 1.0, + "content": "Notably, we see a clear exception to this on the Stack Overflow LR task, where there is a definitive", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 669 + ], + "score": 1.0, + "content": "inverse relationship between learning rates among all optimizers. The EMNIST AE task also displays", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "somewhat different behavior. While there are still noticeable inverse relationships between learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "rates for FEDAVG and FEDAVGM, the range of good client learning rates is relatively small. We", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "emphasize that this task is fundamentally different than the remaining tasks. As noted by Zaheer et al.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "(2018), the primary obstacle in training bottleneck autoencoders is escaping saddle points, not in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 711, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 721 + ], + "score": 1.0, + "content": "converging to critical points. Thus, we expect there to be qualitative differences between EMNIST", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 721, + 385, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 385, + 732 + ], + "score": 1.0, + "content": "AE and other tasks, even EMNIST CR (which uses the same dataset).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 646, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 83, + 331, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 333, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 333, + 95 + ], + "score": 1.0, + "content": "E.5 ROBUSTNESS OF THE ADAPTIVITY PARAMETER", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 104, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "As discussed in Section 5, we plot, for each adaptive optimizer and task, the validation accuracy", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 282, + 128 + ], + "score": 1.0, + "content": "as a function of the adaptivity parameter", + "type": "text" + }, + { + "bbox": [ + 282, + 117, + 289, + 125 + ], + "score": 0.63, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 114, + 427, + 128 + ], + "score": 1.0, + "content": ". In particular, for each value of", + "type": "text" + }, + { + "bbox": [ + 428, + 117, + 435, + 125 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "(which we vary", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 124, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 124, + 127, + 140 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 127, + 126, + 201, + 138 + ], + "score": 0.93, + "content": "\\{ 1 0 ^ { - 5 } , \\dotsc , 1 0 ^ { - 1 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 124, + 507, + 140 + ], + "score": 1.0, + "content": ", see Appendix D), we plot the best possible last-100-rounds validation set", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "score": 1.0, + "content": "performance. Specifically, we plot the average validation performance over the last 100 rounds using", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 492, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 492, + 161 + ], + "score": 1.0, + "content": "the best a posteriori values of client and server learning rates. The results are given in Figure 13.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 126, + 178, + 485, + 449 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 178, + 485, + 449 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 178, + 485, + 449 + ], + "spans": [ + { + "bbox": [ + 126, + 178, + 485, + 449 + ], + "score": 0.978, + "type": "image", + "image_path": "88a3306ab9d22edfbff475be8101d1e8b35ad59803132c108b2b9ff96ab1bfdf.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 126, + 178, + 485, + 268.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 126, + 268.3333333333333, + 485, + 358.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 126, + 358.66666666666663, + 485, + 448.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 468, + 504, + 501 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 484, + 480 + ], + "score": 1.0, + "content": "Figure 13: Validation performance of FEDADAGRAD, FEDADAM, and FEDYOGI for varying", + "type": "text" + }, + { + "bbox": [ + 484, + 470, + 492, + 478 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 233, + 491 + ], + "score": 1.0, + "content": "various tasks. The learning rates", + "type": "text" + }, + { + "bbox": [ + 233, + 480, + 240, + 490 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 479, + 257, + 491 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 480, + 266, + 490 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 479, + 339, + 491 + ], + "score": 1.0, + "content": "are tuned for each", + "type": "text" + }, + { + "bbox": [ + 339, + 481, + 346, + 488 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "to achieve the best training performance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 490, + 264, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 264, + 501 + ], + "score": 1.0, + "content": "on the last 100 communication rounds.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 107, + 532, + 377, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 379, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 379, + 544 + ], + "score": 1.0, + "content": "E.6 IMPROVING PERFORMANCE WITH LEARNING RATE DECAY", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 553, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 506, + 565 + ], + "score": 1.0, + "content": "Despite the success of adaptive methods, it is natural to ask if there is still more to be gained. To test", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "this, we trained the EMNIST CR model in a centralized fashion on a shuffled version of the dataset.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 573, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 589 + ], + "score": 1.0, + "content": "We trained for 100 epochs and used tuned learning rates for each (centralized) optimizer, achieving", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 168, + 598 + ], + "score": 1.0, + "content": "an accuracy of", + "type": "text" + }, + { + "bbox": [ + 168, + 586, + 188, + 596 + ], + "score": 0.87, + "content": "8 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "(see Table 11, CENTRALIZED row), significantly above the best EMNIST CR", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "results from Table 1. The theoretical results in Section 3 point to a partial explanation, as they only", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 606, + 374, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 374, + 620 + ], + "score": 1.0, + "content": "hold when the client learning rate is small or is decayed over time.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 624, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 622, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 622, + 506, + 638 + ], + "score": 1.0, + "content": "To validate this, we ran the same hyperparameter grids on the federated EMNIST CR task, but using", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "a client learning rate schedule. We use a “staircase” exponential decay schedule (EXPDECAY) where", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 196, + 658 + ], + "score": 1.0, + "content": "the client learning rate", + "type": "text" + }, + { + "bbox": [ + 197, + 647, + 206, + 657 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "is decreased by a factor of 0.1 every 500 rounds. This is analogous in some", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 656, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 656, + 506, + 668 + ], + "score": 1.0, + "content": "sense to standard staircase learning rate schedules in centralized optimization (Goyal et al., 2017).", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "Table 11 gives the results. EXPDECAY improves the accuracy of all optimizers, and allows most to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "get close to the best centralized accuracy. While we do not close the gap with centralized optimization", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 689, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 700 + ], + "score": 1.0, + "content": "entirely, we suspect that further tuning of the amount and frequency of decay may lead to even better", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "accuracies. However, this may also require performing significantly more communication rounds, as", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "the theoretical results in Section 3 are primarily asymptotic. In communication-limited settings, the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 720, + 334, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 334, + 733 + ], + "score": 1.0, + "content": "added benefit of learning rate decay seems to be modest.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + } + ], + "page_idx": 35, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 83, + 331, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 333, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 333, + 95 + ], + "score": 1.0, + "content": "E.5 ROBUSTNESS OF THE ADAPTIVITY PARAMETER", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 104, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "As discussed in Section 5, we plot, for each adaptive optimizer and task, the validation accuracy", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 282, + 128 + ], + "score": 1.0, + "content": "as a function of the adaptivity parameter", + "type": "text" + }, + { + "bbox": [ + 282, + 117, + 289, + 125 + ], + "score": 0.63, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 114, + 427, + 128 + ], + "score": 1.0, + "content": ". In particular, for each value of", + "type": "text" + }, + { + "bbox": [ + 428, + 117, + 435, + 125 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "(which we vary", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 124, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 124, + 127, + 140 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 127, + 126, + 201, + 138 + ], + "score": 0.93, + "content": "\\{ 1 0 ^ { - 5 } , \\dotsc , 1 0 ^ { - 1 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 124, + 507, + 140 + ], + "score": 1.0, + "content": ", see Appendix D), we plot the best possible last-100-rounds validation set", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "score": 1.0, + "content": "performance. Specifically, we plot the average validation performance over the last 100 rounds using", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 147, + 492, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 492, + 161 + ], + "score": 1.0, + "content": "the best a posteriori values of client and server learning rates. The results are given in Figure 13.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 104, + 507, + 161 + ] + }, + { + "type": "image", + "bbox": [ + 126, + 178, + 485, + 449 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 178, + 485, + 449 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 178, + 485, + 449 + ], + "spans": [ + { + "bbox": [ + 126, + 178, + 485, + 449 + ], + "score": 0.978, + "type": "image", + "image_path": "88a3306ab9d22edfbff475be8101d1e8b35ad59803132c108b2b9ff96ab1bfdf.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 126, + 178, + 485, + 268.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 126, + 268.3333333333333, + 485, + 358.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 126, + 358.66666666666663, + 485, + 448.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 468, + 504, + 501 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 484, + 480 + ], + "score": 1.0, + "content": "Figure 13: Validation performance of FEDADAGRAD, FEDADAM, and FEDYOGI for varying", + "type": "text" + }, + { + "bbox": [ + 484, + 470, + 492, + 478 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 233, + 491 + ], + "score": 1.0, + "content": "various tasks. The learning rates", + "type": "text" + }, + { + "bbox": [ + 233, + 480, + 240, + 490 + ], + "score": 0.79, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 479, + 257, + 491 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 480, + 266, + 490 + ], + "score": 0.84, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 479, + 339, + 491 + ], + "score": 1.0, + "content": "are tuned for each", + "type": "text" + }, + { + "bbox": [ + 339, + 481, + 346, + 488 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "to achieve the best training performance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 490, + 264, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 264, + 501 + ], + "score": 1.0, + "content": "on the last 100 communication rounds.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 107, + 532, + 377, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 379, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 379, + 544 + ], + "score": 1.0, + "content": "E.6 IMPROVING PERFORMANCE WITH LEARNING RATE DECAY", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 553, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 506, + 565 + ], + "score": 1.0, + "content": "Despite the success of adaptive methods, it is natural to ask if there is still more to be gained. To test", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "this, we trained the EMNIST CR model in a centralized fashion on a shuffled version of the dataset.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 573, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 589 + ], + "score": 1.0, + "content": "We trained for 100 epochs and used tuned learning rates for each (centralized) optimizer, achieving", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 168, + 598 + ], + "score": 1.0, + "content": "an accuracy of", + "type": "text" + }, + { + "bbox": [ + 168, + 586, + 188, + 596 + ], + "score": 0.87, + "content": "8 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "(see Table 11, CENTRALIZED row), significantly above the best EMNIST CR", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "results from Table 1. The theoretical results in Section 3 point to a partial explanation, as they only", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 606, + 374, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 374, + 620 + ], + "score": 1.0, + "content": "hold when the client learning rate is small or is decayed over time.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 553, + 506, + 620 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 624, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 622, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 622, + 506, + 638 + ], + "score": 1.0, + "content": "To validate this, we ran the same hyperparameter grids on the federated EMNIST CR task, but using", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "a client learning rate schedule. We use a “staircase” exponential decay schedule (EXPDECAY) where", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 196, + 658 + ], + "score": 1.0, + "content": "the client learning rate", + "type": "text" + }, + { + "bbox": [ + 197, + 647, + 206, + 657 + ], + "score": 0.83, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "is decreased by a factor of 0.1 every 500 rounds. This is analogous in some", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 656, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 656, + 506, + 668 + ], + "score": 1.0, + "content": "sense to standard staircase learning rate schedules in centralized optimization (Goyal et al., 2017).", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "Table 11 gives the results. EXPDECAY improves the accuracy of all optimizers, and allows most to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "get close to the best centralized accuracy. While we do not close the gap with centralized optimization", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 689, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 506, + 700 + ], + "score": 1.0, + "content": "entirely, we suspect that further tuning of the amount and frequency of decay may lead to even better", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "accuracies. However, this may also require performing significantly more communication rounds, as", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "the theoretical results in Section 3 are primarily asymptotic. In communication-limited settings, the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 720, + 334, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 334, + 733 + ], + "score": 1.0, + "content": "added benefit of learning rate decay seems to be modest.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 622, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 145, + 419, + 217 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 506, + 133 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 89, + 506, + 103 + ], + "spans": [ + { + "bbox": [ + 104, + 89, + 232, + 103 + ], + "score": 1.0, + "content": "Table 11: (Top) Test accuracy", + "type": "text" + }, + { + "bbox": [ + 232, + 90, + 248, + 101 + ], + "score": 0.82, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 89, + 506, + 103 + ], + "score": 1.0, + "content": "of a model trained centrally with various optimizers. (Bottom)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 196, + 113 + ], + "score": 1.0, + "content": "Average test accuracy", + "type": "text" + }, + { + "bbox": [ + 196, + 101, + 212, + 112 + ], + "score": 0.82, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 101, + 506, + 113 + ], + "score": 1.0, + "content": "over the last 100 rounds of various federated optimizers on the EMNIST", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 377, + 124 + ], + "score": 1.0, + "content": "CR task, using constant learning rates or the EXPDECAY schedule for", + "type": "text" + }, + { + "bbox": [ + 378, + 113, + 387, + 123 + ], + "score": 0.81, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 110, + 506, + 124 + ], + "score": 1.0, + "content": ". Accuracies (for the federated", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 328, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 160, + 133 + ], + "score": 1.0, + "content": "tasks) within", + "type": "text" + }, + { + "bbox": [ + 160, + 122, + 183, + 133 + ], + "score": 0.87, + "content": "\\mathrm { { \\bar { 0 } . 5 \\% } }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 122, + 328, + 133 + ], + "score": 1.0, + "content": "of the best result are shown in bold.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 192, + 145, + 419, + 217 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 192, + 145, + 419, + 217 + ], + "spans": [ + { + "bbox": [ + 192, + 145, + 419, + 217 + ], + "score": 0.979, + "html": "
ADAGRAD CENTRALIZED 88.0ADAM 87.9YOGI 88.0SGDM 87.7SGD 87.7
FED... ADAGRADADAMYOGIAvGMAVG
CONSTANT ηt85.1 85.685.585.284.9
EXPDECAY85.3 86.286.285.885.2
", + "type": "table", + "image_path": "9ab3ed2946f8c758203144c35238f338ebc6565635feb02aa760ffd0caeae7ca.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 192, + 145, + 419, + 159.4 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 192, + 159.4, + 419, + 173.8 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 192, + 173.8, + 419, + 188.20000000000002 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 192, + 188.20000000000002, + 419, + 202.60000000000002 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 192, + 202.60000000000002, + 419, + 217.00000000000003 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 3.75 + }, + { + "type": "title", + "bbox": [ + 106, + 240, + 321, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 321, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 321, + 255 + ], + "score": 1.0, + "content": "F CREATING A FEDERATED CIFAR-100", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 505, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 151, + 277 + ], + "score": 1.0, + "content": "Overview", + "type": "text" + }, + { + "bbox": [ + 152, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "We use the Pachinko Allocation Method (PAM) (Li & McCallum, 2006) to create a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "federated CIFAR-100. PAM is a topic modeling method in which the correlations between individual", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "words in a vocabulary are represented by a rooted directed acyclic graph (DAG) whose leaves are the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "vocabulary words. The interior nodes are topics with Dirichlet distributions over their child nodes.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 308, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 320 + ], + "score": 1.0, + "content": "To generate a document, we sample multinomial distributions from each interior node’s Dirichlet", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "distribution. To sample a word from the document, we begin at the root, and draw a child node its", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 411, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 411, + 342 + ], + "score": 1.0, + "content": "multinomial distribution, and continue doing this until we reach a leaf node.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 346, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "To partition CIFAR-100 across clients, we use the label structure of CIFAR-100. Each image in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "the dataset has a fine label (often referred to as its label) which is a member of a coarse label. For", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "score": 1.0, + "content": "example, the fine label “seal” is a member of the coarse label “aquatic mammals”. There are 20", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 465, + 391 + ], + "score": 1.0, + "content": "coarse labels in CIFAR-100, each with 5 fine labels. We represent this structure as a DAG", + "type": "text" + }, + { + "bbox": [ + 465, + 379, + 474, + 389 + ], + "score": 0.78, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 379, + 506, + 391 + ], + "score": 1.0, + "content": ", with a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "root whose children are the coarse labels. Each coarse label is an interior node whose child nodes are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 434, + 412 + ], + "score": 1.0, + "content": "its fine labels. The root node has a symmetric Dirichlet distribution with parameter", + "type": "text" + }, + { + "bbox": [ + 434, + 402, + 442, + 410 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "over the coarse", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 410, + 450, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 438, + 423 + ], + "score": 1.0, + "content": "labels, and each coarse label has a symmetric Dirichlet distribution with parameter", + "type": "text" + }, + { + "bbox": [ + 438, + 411, + 445, + 423 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 410, + 450, + 423 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "We associate each client to a document. That is, we draw a multinomial distribution from the Dirichlet", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 171, + 450 + ], + "score": 1.0, + "content": "prior at the root", + "type": "text" + }, + { + "bbox": [ + 172, + 438, + 206, + 450 + ], + "score": 0.88, + "content": "\\left( \\operatorname { D i r } ( \\alpha ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "and a multinomial distribution from the Dirichlet prior at each coarse label", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 448, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 141, + 461 + ], + "score": 0.84, + "content": "\\bar { ( } \\mathrm { D i r } ( \\beta ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 448, + 507, + 462 + ], + "score": 1.0, + "content": ". To create the client dataset, we draw leaf nodes from this DAG using Pachinko allocation,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "randomly sample an example with the given fine label, and assign it to the client’s dataset. We do this", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 470, + 307, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 307, + 483 + ], + "score": 1.0, + "content": "100 times for each of 500 distinct training clients.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "While more complex than LDA, this approach creates more realistic heterogeneity among client", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "datasets by creating correlations between label frequencies for fine labels within the same coarse", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 509, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 521 + ], + "score": 1.0, + "content": "label set. Intuitively, if a client’s dataset has many images of dolphins, they are likely to also have", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 254, + 532 + ], + "score": 1.0, + "content": "pictures of whales. By using a small", + "type": "text" + }, + { + "bbox": [ + 254, + 521, + 262, + 529 + ], + "score": 0.79, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 519, + 506, + 532 + ], + "score": 1.0, + "content": "at the root, client datasets become more likely to focus on a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 248, + 543 + ], + "score": 1.0, + "content": "few coarse labels. By using a larger", + "type": "text" + }, + { + "bbox": [ + 248, + 531, + 255, + 542 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "for the coarse-to-fine label distributions, clients are more likely", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 541, + 328, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 328, + 552 + ], + "score": 1.0, + "content": "to have multiple fine labels from the same coarse label.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 557, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "score": 1.0, + "content": "One important note: Once we sample a fine label, we randomly select an element with that label", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "without replacement. This ensures no two clients have overlapping examples. In more detail, suppose", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 218, + 591 + ], + "score": 1.0, + "content": "we have sample a fine label", + "type": "text" + }, + { + "bbox": [ + 218, + 581, + 225, + 590 + ], + "score": 0.77, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 579, + 296, + 591 + ], + "score": 1.0, + "content": "with coarse label", + "type": "text" + }, + { + "bbox": [ + 296, + 581, + 302, + 589 + ], + "score": 0.74, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 579, + 342, + 591 + ], + "score": 1.0, + "content": "for client", + "type": "text" + }, + { + "bbox": [ + 342, + 581, + 352, + 590 + ], + "score": 0.74, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 579, + 505, + 591 + ], + "score": 1.0, + "content": ", and there is only one remaining such", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 142, + 603 + ], + "score": 1.0, + "content": "example", + "type": "text" + }, + { + "bbox": [ + 142, + 590, + 174, + 601 + ], + "score": 0.92, + "content": "( x , { \\bar { c } } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 590, + 221, + 603 + ], + "score": 1.0, + "content": ". We assign", + "type": "text" + }, + { + "bbox": [ + 221, + 590, + 253, + 601 + ], + "score": 0.92, + "content": "( x , c , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 590, + 289, + 603 + ], + "score": 1.0, + "content": "to client", + "type": "text" + }, + { + "bbox": [ + 289, + 591, + 299, + 600 + ], + "score": 0.76, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 590, + 438, + 603 + ], + "score": 1.0, + "content": "’s dataset, and remove the leaf node", + "type": "text" + }, + { + "bbox": [ + 439, + 592, + 446, + 601 + ], + "score": 0.78, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "from the DAG", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 599, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 107, + 601, + 116, + 611 + ], + "score": 0.68, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 599, + 188, + 614 + ], + "score": 1.0, + "content": ". We also remove", + "type": "text" + }, + { + "bbox": [ + 189, + 603, + 195, + 612 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 599, + 337, + 614 + ], + "score": 1.0, + "content": "from the multinomial distribution", + "type": "text" + }, + { + "bbox": [ + 337, + 601, + 347, + 612 + ], + "score": 0.88, + "content": "\\theta _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 599, + 392, + 614 + ], + "score": 1.0, + "content": "that client", + "type": "text" + }, + { + "bbox": [ + 393, + 602, + 403, + 611 + ], + "score": 0.71, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 599, + 506, + 614 + ], + "score": 1.0, + "content": "has associated to coarse", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 128, + 624 + ], + "score": 1.0, + "content": "label", + "type": "text" + }, + { + "bbox": [ + 128, + 613, + 134, + 622 + ], + "score": 0.62, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 610, + 349, + 624 + ], + "score": 1.0, + "content": ", which we refer to as renormalization with respect to", + "type": "text" + }, + { + "bbox": [ + 350, + 613, + 357, + 623 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 610, + 427, + 624 + ], + "score": 1.0, + "content": "(Algorithm 8). If", + "type": "text" + }, + { + "bbox": [ + 428, + 612, + 433, + 621 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "has no remaining", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 217, + 635 + ], + "score": 1.0, + "content": "children after pruning node", + "type": "text" + }, + { + "bbox": [ + 217, + 623, + 223, + 633 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 621, + 313, + 635 + ], + "score": 1.0, + "content": ", we also remove node", + "type": "text" + }, + { + "bbox": [ + 313, + 623, + 318, + 632 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 621, + 340, + 635 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 341, + 623, + 350, + 632 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "and re-normalize the root multinomial", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 117, + 644 + ], + "score": 0.86, + "content": "\\theta _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 632, + 178, + 646 + ], + "score": 1.0, + "content": "with respect to", + "type": "text" + }, + { + "bbox": [ + 179, + 635, + 185, + 643 + ], + "score": 0.65, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 632, + 453, + 646 + ], + "score": 1.0, + "content": ". For all subsequent clients, we draw multinomials from this pruned", + "type": "text" + }, + { + "bbox": [ + 453, + 633, + 462, + 643 + ], + "score": 0.79, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "according", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 643, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 507, + 658 + ], + "score": 1.0, + "content": "to symmetric Dirichlet distributions with the same parameters as before, but with one fewer category.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 227, + 678 + ], + "score": 1.0, + "content": "Notation and method Let", + "type": "text" + }, + { + "bbox": [ + 227, + 667, + 234, + 677 + ], + "score": 0.78, + "content": "\\mathcal { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 666, + 372, + 678 + ], + "score": 1.0, + "content": "denote the set of coarse labels and", + "type": "text" + }, + { + "bbox": [ + 373, + 667, + 382, + 677 + ], + "score": 0.84, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 666, + 496, + 678 + ], + "score": 1.0, + "content": "the set of fine labels, and let", + "type": "text" + }, + { + "bbox": [ + 496, + 667, + 504, + 677 + ], + "score": 0.8, + "content": "s", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 677, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 335, + 690 + ], + "score": 1.0, + "content": "denote the CIFAR-100 dataset. This consists of examples", + "type": "text" + }, + { + "bbox": [ + 335, + 678, + 367, + 689 + ], + "score": 0.92, + "content": "( x , c , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 677, + 395, + 690 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 396, + 680, + 403, + 687 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 677, + 480, + 690 + ], + "score": 1.0, + "content": "is an image vector,", + "type": "text" + }, + { + "bbox": [ + 480, + 678, + 504, + 688 + ], + "score": 0.88, + "content": "c \\in { \\mathcal { C } }", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 204, + 701 + ], + "score": 1.0, + "content": "is a coarse label set, and", + "type": "text" + }, + { + "bbox": [ + 205, + 689, + 231, + 700 + ], + "score": 0.91, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 688, + 307, + 701 + ], + "score": 1.0, + "content": "is a fine label with", + "type": "text" + }, + { + "bbox": [ + 308, + 690, + 331, + 700 + ], + "score": 0.89, + "content": "y \\in c", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 688, + 350, + 701 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 351, + 689, + 403, + 700 + ], + "score": 0.8, + "content": "c \\in \\mathcal { C } , y \\in \\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 688, + 433, + 701 + ], + "score": 1.0, + "content": ", we let", + "type": "text" + }, + { + "bbox": [ + 433, + 689, + 444, + 699 + ], + "score": 0.89, + "content": "S _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 688, + 462, + 701 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 463, + 689, + 475, + 700 + ], + "score": 0.89, + "content": "\\mathcal { S } _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 688, + 505, + 701 + ], + "score": 1.0, + "content": "denote", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 196, + 712 + ], + "score": 1.0, + "content": "the set of examples in", + "type": "text" + }, + { + "bbox": [ + 196, + 700, + 204, + 709 + ], + "score": 0.83, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 699, + 276, + 712 + ], + "score": 1.0, + "content": "with coarse label", + "type": "text" + }, + { + "bbox": [ + 276, + 702, + 282, + 709 + ], + "score": 0.77, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 699, + 339, + 712 + ], + "score": 1.0, + "content": "and fine label", + "type": "text" + }, + { + "bbox": [ + 340, + 701, + 346, + 711 + ], + "score": 0.76, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 699, + 366, + 712 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 366, + 700, + 393, + 709 + ], + "score": 0.91, + "content": "v \\in G", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 699, + 421, + 712 + ], + "score": 1.0, + "content": "we let", + "type": "text" + }, + { + "bbox": [ + 421, + 699, + 446, + 711 + ], + "score": 0.91, + "content": "| G [ v ] |", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "denote the set", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 165, + 723 + ], + "score": 1.0, + "content": "of children of", + "type": "text" + }, + { + "bbox": [ + 165, + 712, + 172, + 720 + ], + "score": 0.75, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 709, + 183, + 723 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 184, + 711, + 192, + 720 + ], + "score": 0.76, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 709, + 213, + 723 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 213, + 710, + 241, + 722 + ], + "score": 0.91, + "content": "\\gamma \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 709, + 257, + 723 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 258, + 710, + 295, + 722 + ], + "score": 0.92, + "content": "\\operatorname { D i r } ( \\gamma , k )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 709, + 497, + 723 + ], + "score": 1.0, + "content": "denote the symmetric Dirichlet distribution with", + "type": "text" + }, + { + "bbox": [ + 497, + 711, + 504, + 720 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 721, + 152, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 152, + 734 + ], + "score": 1.0, + "content": "categories.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5 + } + ], + "page_idx": 36, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 145, + 419, + 217 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 506, + 133 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 89, + 506, + 103 + ], + "spans": [ + { + "bbox": [ + 104, + 89, + 232, + 103 + ], + "score": 1.0, + "content": "Table 11: (Top) Test accuracy", + "type": "text" + }, + { + "bbox": [ + 232, + 90, + 248, + 101 + ], + "score": 0.82, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 89, + 506, + 103 + ], + "score": 1.0, + "content": "of a model trained centrally with various optimizers. (Bottom)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 196, + 113 + ], + "score": 1.0, + "content": "Average test accuracy", + "type": "text" + }, + { + "bbox": [ + 196, + 101, + 212, + 112 + ], + "score": 0.82, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 101, + 506, + 113 + ], + "score": 1.0, + "content": "over the last 100 rounds of various federated optimizers on the EMNIST", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 377, + 124 + ], + "score": 1.0, + "content": "CR task, using constant learning rates or the EXPDECAY schedule for", + "type": "text" + }, + { + "bbox": [ + 378, + 113, + 387, + 123 + ], + "score": 0.81, + "content": "\\eta _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 110, + 506, + 124 + ], + "score": 1.0, + "content": ". Accuracies (for the federated", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 328, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 160, + 133 + ], + "score": 1.0, + "content": "tasks) within", + "type": "text" + }, + { + "bbox": [ + 160, + 122, + 183, + 133 + ], + "score": 0.87, + "content": "\\mathrm { { \\bar { 0 } . 5 \\% } }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 122, + 328, + 133 + ], + "score": 1.0, + "content": "of the best result are shown in bold.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 192, + 145, + 419, + 217 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 192, + 145, + 419, + 217 + ], + "spans": [ + { + "bbox": [ + 192, + 145, + 419, + 217 + ], + "score": 0.979, + "html": "
ADAGRAD CENTRALIZED 88.0ADAM 87.9YOGI 88.0SGDM 87.7SGD 87.7
FED... ADAGRADADAMYOGIAvGMAVG
CONSTANT ηt85.1 85.685.585.284.9
EXPDECAY85.3 86.286.285.885.2
", + "type": "table", + "image_path": "9ab3ed2946f8c758203144c35238f338ebc6565635feb02aa760ffd0caeae7ca.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 192, + 145, + 419, + 159.4 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 192, + 159.4, + 419, + 173.8 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 192, + 173.8, + 419, + 188.20000000000002 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 192, + 188.20000000000002, + 419, + 202.60000000000002 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 192, + 202.60000000000002, + 419, + 217.00000000000003 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 3.75 + }, + { + "type": "title", + "bbox": [ + 106, + 240, + 321, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 321, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 321, + 255 + ], + "score": 1.0, + "content": "F CREATING A FEDERATED CIFAR-100", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 505, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 151, + 277 + ], + "score": 1.0, + "content": "Overview", + "type": "text" + }, + { + "bbox": [ + 152, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "We use the Pachinko Allocation Method (PAM) (Li & McCallum, 2006) to create a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "federated CIFAR-100. PAM is a topic modeling method in which the correlations between individual", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "words in a vocabulary are represented by a rooted directed acyclic graph (DAG) whose leaves are the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "vocabulary words. The interior nodes are topics with Dirichlet distributions over their child nodes.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 308, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 320 + ], + "score": 1.0, + "content": "To generate a document, we sample multinomial distributions from each interior node’s Dirichlet", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 331 + ], + "score": 1.0, + "content": "distribution. To sample a word from the document, we begin at the root, and draw a child node its", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 411, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 411, + 342 + ], + "score": 1.0, + "content": "multinomial distribution, and continue doing this until we reach a leaf node.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 264, + 506, + 342 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 346, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "To partition CIFAR-100 across clients, we use the label structure of CIFAR-100. Each image in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "the dataset has a fine label (often referred to as its label) which is a member of a coarse label. For", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "score": 1.0, + "content": "example, the fine label “seal” is a member of the coarse label “aquatic mammals”. There are 20", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 465, + 391 + ], + "score": 1.0, + "content": "coarse labels in CIFAR-100, each with 5 fine labels. We represent this structure as a DAG", + "type": "text" + }, + { + "bbox": [ + 465, + 379, + 474, + 389 + ], + "score": 0.78, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 379, + 506, + 391 + ], + "score": 1.0, + "content": ", with a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "root whose children are the coarse labels. Each coarse label is an interior node whose child nodes are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 434, + 412 + ], + "score": 1.0, + "content": "its fine labels. The root node has a symmetric Dirichlet distribution with parameter", + "type": "text" + }, + { + "bbox": [ + 434, + 402, + 442, + 410 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "over the coarse", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 410, + 450, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 438, + 423 + ], + "score": 1.0, + "content": "labels, and each coarse label has a symmetric Dirichlet distribution with parameter", + "type": "text" + }, + { + "bbox": [ + 438, + 411, + 445, + 423 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 410, + 450, + 423 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 346, + 506, + 423 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "We associate each client to a document. That is, we draw a multinomial distribution from the Dirichlet", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 171, + 450 + ], + "score": 1.0, + "content": "prior at the root", + "type": "text" + }, + { + "bbox": [ + 172, + 438, + 206, + 450 + ], + "score": 0.88, + "content": "\\left( \\operatorname { D i r } ( \\alpha ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "and a multinomial distribution from the Dirichlet prior at each coarse label", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 448, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 141, + 461 + ], + "score": 0.84, + "content": "\\bar { ( } \\mathrm { D i r } ( \\beta ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 448, + 507, + 462 + ], + "score": 1.0, + "content": ". To create the client dataset, we draw leaf nodes from this DAG using Pachinko allocation,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 472 + ], + "score": 1.0, + "content": "randomly sample an example with the given fine label, and assign it to the client’s dataset. We do this", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 470, + 307, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 307, + 483 + ], + "score": 1.0, + "content": "100 times for each of 500 distinct training clients.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 427, + 507, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "While more complex than LDA, this approach creates more realistic heterogeneity among client", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "datasets by creating correlations between label frequencies for fine labels within the same coarse", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 509, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 521 + ], + "score": 1.0, + "content": "label set. Intuitively, if a client’s dataset has many images of dolphins, they are likely to also have", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 254, + 532 + ], + "score": 1.0, + "content": "pictures of whales. By using a small", + "type": "text" + }, + { + "bbox": [ + 254, + 521, + 262, + 529 + ], + "score": 0.79, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 519, + 506, + 532 + ], + "score": 1.0, + "content": "at the root, client datasets become more likely to focus on a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 248, + 543 + ], + "score": 1.0, + "content": "few coarse labels. By using a larger", + "type": "text" + }, + { + "bbox": [ + 248, + 531, + 255, + 542 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "for the coarse-to-fine label distributions, clients are more likely", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 541, + 328, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 328, + 552 + ], + "score": 1.0, + "content": "to have multiple fine labels from the same coarse label.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 487, + 506, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 557, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "score": 1.0, + "content": "One important note: Once we sample a fine label, we randomly select an element with that label", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "without replacement. This ensures no two clients have overlapping examples. In more detail, suppose", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 218, + 591 + ], + "score": 1.0, + "content": "we have sample a fine label", + "type": "text" + }, + { + "bbox": [ + 218, + 581, + 225, + 590 + ], + "score": 0.77, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 579, + 296, + 591 + ], + "score": 1.0, + "content": "with coarse label", + "type": "text" + }, + { + "bbox": [ + 296, + 581, + 302, + 589 + ], + "score": 0.74, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 579, + 342, + 591 + ], + "score": 1.0, + "content": "for client", + "type": "text" + }, + { + "bbox": [ + 342, + 581, + 352, + 590 + ], + "score": 0.74, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 579, + 505, + 591 + ], + "score": 1.0, + "content": ", and there is only one remaining such", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 142, + 603 + ], + "score": 1.0, + "content": "example", + "type": "text" + }, + { + "bbox": [ + 142, + 590, + 174, + 601 + ], + "score": 0.92, + "content": "( x , { \\bar { c } } , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 590, + 221, + 603 + ], + "score": 1.0, + "content": ". We assign", + "type": "text" + }, + { + "bbox": [ + 221, + 590, + 253, + 601 + ], + "score": 0.92, + "content": "( x , c , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 590, + 289, + 603 + ], + "score": 1.0, + "content": "to client", + "type": "text" + }, + { + "bbox": [ + 289, + 591, + 299, + 600 + ], + "score": 0.76, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 590, + 438, + 603 + ], + "score": 1.0, + "content": "’s dataset, and remove the leaf node", + "type": "text" + }, + { + "bbox": [ + 439, + 592, + 446, + 601 + ], + "score": 0.78, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "from the DAG", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 599, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 107, + 601, + 116, + 611 + ], + "score": 0.68, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 599, + 188, + 614 + ], + "score": 1.0, + "content": ". We also remove", + "type": "text" + }, + { + "bbox": [ + 189, + 603, + 195, + 612 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 599, + 337, + 614 + ], + "score": 1.0, + "content": "from the multinomial distribution", + "type": "text" + }, + { + "bbox": [ + 337, + 601, + 347, + 612 + ], + "score": 0.88, + "content": "\\theta _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 599, + 392, + 614 + ], + "score": 1.0, + "content": "that client", + "type": "text" + }, + { + "bbox": [ + 393, + 602, + 403, + 611 + ], + "score": 0.71, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 599, + 506, + 614 + ], + "score": 1.0, + "content": "has associated to coarse", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 128, + 624 + ], + "score": 1.0, + "content": "label", + "type": "text" + }, + { + "bbox": [ + 128, + 613, + 134, + 622 + ], + "score": 0.62, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 610, + 349, + 624 + ], + "score": 1.0, + "content": ", which we refer to as renormalization with respect to", + "type": "text" + }, + { + "bbox": [ + 350, + 613, + 357, + 623 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 610, + 427, + 624 + ], + "score": 1.0, + "content": "(Algorithm 8). If", + "type": "text" + }, + { + "bbox": [ + 428, + 612, + 433, + 621 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "has no remaining", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 217, + 635 + ], + "score": 1.0, + "content": "children after pruning node", + "type": "text" + }, + { + "bbox": [ + 217, + 623, + 223, + 633 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 621, + 313, + 635 + ], + "score": 1.0, + "content": ", we also remove node", + "type": "text" + }, + { + "bbox": [ + 313, + 623, + 318, + 632 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 621, + 340, + 635 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 341, + 623, + 350, + 632 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "and re-normalize the root multinomial", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 117, + 644 + ], + "score": 0.86, + "content": "\\theta _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 632, + 178, + 646 + ], + "score": 1.0, + "content": "with respect to", + "type": "text" + }, + { + "bbox": [ + 179, + 635, + 185, + 643 + ], + "score": 0.65, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 632, + 453, + 646 + ], + "score": 1.0, + "content": ". For all subsequent clients, we draw multinomials from this pruned", + "type": "text" + }, + { + "bbox": [ + 453, + 633, + 462, + 643 + ], + "score": 0.79, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "according", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 643, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 507, + 658 + ], + "score": 1.0, + "content": "to symmetric Dirichlet distributions with the same parameters as before, but with one fewer category.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39, + "bbox_fs": [ + 104, + 558, + 507, + 658 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 504, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 227, + 678 + ], + "score": 1.0, + "content": "Notation and method Let", + "type": "text" + }, + { + "bbox": [ + 227, + 667, + 234, + 677 + ], + "score": 0.78, + "content": "\\mathcal { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 666, + 372, + 678 + ], + "score": 1.0, + "content": "denote the set of coarse labels and", + "type": "text" + }, + { + "bbox": [ + 373, + 667, + 382, + 677 + ], + "score": 0.84, + "content": "\\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 666, + 496, + 678 + ], + "score": 1.0, + "content": "the set of fine labels, and let", + "type": "text" + }, + { + "bbox": [ + 496, + 667, + 504, + 677 + ], + "score": 0.8, + "content": "s", + "type": "inline_equation" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 677, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 335, + 690 + ], + "score": 1.0, + "content": "denote the CIFAR-100 dataset. This consists of examples", + "type": "text" + }, + { + "bbox": [ + 335, + 678, + 367, + 689 + ], + "score": 0.92, + "content": "( x , c , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 677, + 395, + 690 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 396, + 680, + 403, + 687 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 677, + 480, + 690 + ], + "score": 1.0, + "content": "is an image vector,", + "type": "text" + }, + { + "bbox": [ + 480, + 678, + 504, + 688 + ], + "score": 0.88, + "content": "c \\in { \\mathcal { C } }", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 688, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 204, + 701 + ], + "score": 1.0, + "content": "is a coarse label set, and", + "type": "text" + }, + { + "bbox": [ + 205, + 689, + 231, + 700 + ], + "score": 0.91, + "content": "y \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 688, + 307, + 701 + ], + "score": 1.0, + "content": "is a fine label with", + "type": "text" + }, + { + "bbox": [ + 308, + 690, + 331, + 700 + ], + "score": 0.89, + "content": "y \\in c", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 688, + 350, + 701 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 351, + 689, + 403, + 700 + ], + "score": 0.8, + "content": "c \\in \\mathcal { C } , y \\in \\mathcal { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 688, + 433, + 701 + ], + "score": 1.0, + "content": ", we let", + "type": "text" + }, + { + "bbox": [ + 433, + 689, + 444, + 699 + ], + "score": 0.89, + "content": "S _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 688, + 462, + 701 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 463, + 689, + 475, + 700 + ], + "score": 0.89, + "content": "\\mathcal { S } _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 688, + 505, + 701 + ], + "score": 1.0, + "content": "denote", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 196, + 712 + ], + "score": 1.0, + "content": "the set of examples in", + "type": "text" + }, + { + "bbox": [ + 196, + 700, + 204, + 709 + ], + "score": 0.83, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 699, + 276, + 712 + ], + "score": 1.0, + "content": "with coarse label", + "type": "text" + }, + { + "bbox": [ + 276, + 702, + 282, + 709 + ], + "score": 0.77, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 699, + 339, + 712 + ], + "score": 1.0, + "content": "and fine label", + "type": "text" + }, + { + "bbox": [ + 340, + 701, + 346, + 711 + ], + "score": 0.76, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 699, + 366, + 712 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 366, + 700, + 393, + 709 + ], + "score": 0.91, + "content": "v \\in G", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 699, + 421, + 712 + ], + "score": 1.0, + "content": "we let", + "type": "text" + }, + { + "bbox": [ + 421, + 699, + 446, + 711 + ], + "score": 0.91, + "content": "| G [ v ] |", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "denote the set", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 504, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 165, + 723 + ], + "score": 1.0, + "content": "of children of", + "type": "text" + }, + { + "bbox": [ + 165, + 712, + 172, + 720 + ], + "score": 0.75, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 709, + 183, + 723 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 184, + 711, + 192, + 720 + ], + "score": 0.76, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 709, + 213, + 723 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 213, + 710, + 241, + 722 + ], + "score": 0.91, + "content": "\\gamma \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 709, + 257, + 723 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 258, + 710, + 295, + 722 + ], + "score": 0.92, + "content": "\\operatorname { D i r } ( \\gamma , k )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 709, + 497, + 723 + ], + "score": 1.0, + "content": "denote the symmetric Dirichlet distribution with", + "type": "text" + }, + { + "bbox": [ + 497, + 711, + 504, + 720 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 721, + 152, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 152, + 734 + ], + "score": 1.0, + "content": "categories.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 666, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 291, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 294, + 96 + ], + "score": 1.0, + "content": "Algorithm 7 Creating a federated CIFAR-100", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 95, + 311, + 295 + ], + "lines": [ + { + "bbox": [ + 115, + 93, + 249, + 110 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 141, + 110 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 142, + 96, + 249, + 107 + ], + "score": 0.5, + "content": "N , M \\in \\mathbb { Z } _ { > 0 } , \\alpha , \\beta \\in \\mathbb { R } _ { \\ge 0 }", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 106, + 209, + 118 + ], + "spans": [ + { + "bbox": [ + 116, + 106, + 131, + 118 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 132, + 107, + 196, + 117 + ], + "score": 0.52, + "content": "m = 1 , \\cdots , M", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 106, + 209, + 118 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 116, + 243, + 131 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 163, + 131 + ], + "score": 1.0, + "content": "Sample", + "type": "text" + }, + { + "bbox": [ + 164, + 117, + 243, + 129 + ], + "score": 0.76, + "content": "\\theta _ { r } \\sim \\operatorname { D i r } ( \\alpha , | G [ r ] | )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 127, + 215, + 141 + ], + "spans": [ + { + "bbox": [ + 130, + 127, + 147, + 141 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 147, + 128, + 199, + 139 + ], + "score": 0.78, + "content": "c \\in { \\mathcal { C } } \\cap G [ r ]", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 127, + 215, + 141 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 144, + 136, + 257, + 153 + ], + "spans": [ + { + "bbox": [ + 144, + 136, + 178, + 153 + ], + "score": 1.0, + "content": "Sample", + "type": "text" + }, + { + "bbox": [ + 179, + 139, + 257, + 151 + ], + "score": 0.81, + "content": "\\theta _ { c } \\sim \\operatorname { D i r } ( \\beta , | G [ c ] | )", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 151, + 168, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 151, + 168, + 163 + ], + "score": 0.57, + "content": "D _ { m } = \\varnothing", + "type": "inline_equation", + "image_path": "88b407760e00cd79137d458acf36234cae322ad3e963e502b577cfcd93cad360.jpg" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 162, + 214, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 146, + 175 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 146, + 164, + 199, + 174 + ], + "score": 0.83, + "content": "n = 1 , \\cdots N", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 162, + 214, + 175 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 145, + 172, + 264, + 186 + ], + "spans": [ + { + "bbox": [ + 145, + 172, + 247, + 186 + ], + "score": 1.0, + "content": "Sample c ∼ Multinomial", + "type": "text" + }, + { + "bbox": [ + 248, + 174, + 264, + 185 + ], + "score": 0.49, + "content": "\\left( \\theta _ { r } \\right)", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 183, + 264, + 196 + ], + "spans": [ + { + "bbox": [ + 145, + 183, + 249, + 196 + ], + "score": 1.0, + "content": "Sample y ∼ Multinomial", + "type": "text" + }, + { + "bbox": [ + 249, + 185, + 264, + 196 + ], + "score": 0.62, + "content": "\\left( \\theta _ { c } \\right)", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 194, + 311, + 208 + ], + "spans": [ + { + "bbox": [ + 145, + 194, + 174, + 208 + ], + "score": 1.0, + "content": "Select", + "type": "text" + }, + { + "bbox": [ + 174, + 195, + 225, + 206 + ], + "score": 0.82, + "content": "( x , c , y ) \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 194, + 311, + 208 + ], + "score": 1.0, + "content": "uniformly at random", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 147, + 206, + 246, + 217 + ], + "spans": [ + { + "bbox": [ + 147, + 206, + 246, + 217 + ], + "score": 0.59, + "content": "D _ { m } = D _ { m } \\cup \\{ ( x , c , y ) \\}", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 146, + 217, + 222, + 228 + ], + "spans": [ + { + "bbox": [ + 146, + 217, + 222, + 228 + ], + "score": 0.35, + "content": "\\mathcal { S } = \\mathcal { S } \\backslash \\{ ( x , c , \\overset { \\cdot } { y } ) \\}", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 145, + 226, + 210, + 240 + ], + "spans": [ + { + "bbox": [ + 145, + 226, + 155, + 240 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 155, + 228, + 185, + 239 + ], + "score": 0.68, + "content": "\\begin{array} { r } { { \\cal S } _ { y } = \\emptyset } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 226, + 210, + 240 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 158, + 235, + 214, + 252 + ], + "spans": [ + { + "bbox": [ + 158, + 235, + 214, + 252 + ], + "score": 1.0, + "content": "G = G\\{y}", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 159, + 245, + 276, + 262 + ], + "spans": [ + { + "bbox": [ + 159, + 245, + 276, + 262 + ], + "score": 1.0, + "content": "θc = RENORMALIZE(θc, y)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 160, + 258, + 223, + 271 + ], + "spans": [ + { + "bbox": [ + 160, + 258, + 223, + 271 + ], + "score": 1.0, + "content": "if Sc = ∅ then", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 174, + 268, + 227, + 284 + ], + "spans": [ + { + "bbox": [ + 174, + 268, + 227, + 284 + ], + "score": 1.0, + "content": "G = G\\{c}", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 174, + 279, + 291, + 294 + ], + "spans": [ + { + "bbox": [ + 174, + 279, + 291, + 294 + ], + "score": 1.0, + "content": "θr = RENORMALIZE(θr, c)", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 9.5 + }, + { + "type": "table", + "bbox": [ + 106, + 307, + 505, + 387 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 307, + 505, + 387 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 307, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 387 + ], + "score": 0.49, + "html": "
Algorithm8RENORMALIZE
Initialization: θ= (p1,...,Pk),i∈[K]
a=∑k≠iPk
fork ∈[K],k≠ido
Pk=pk/a
Return0' = (pi,···P'-1,P'+1,···,PK)
", + "type": "table", + "image_path": "efaa025501c7065b26a782907225a109bc4617af98f087301f0971bf59bbce79.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 106, + 307, + 505, + 333.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 106, + 333.6666666666667, + 505, + 360.33333333333337 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 106, + 360.33333333333337, + 505, + 387.00000000000006 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 408, + 506, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 123, + 421 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 409, + 135, + 419 + ], + "score": 0.72, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 407, + 261, + 421 + ], + "score": 1.0, + "content": "denote the number of clients,", + "type": "text" + }, + { + "bbox": [ + 262, + 409, + 272, + 419 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 407, + 439, + 421 + ], + "score": 1.0, + "content": "the number of examples per client, and", + "type": "text" + }, + { + "bbox": [ + 439, + 409, + 457, + 420 + ], + "score": 0.9, + "content": "D _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "the dataset", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 147, + 432 + ], + "score": 1.0, + "content": "for client", + "type": "text" + }, + { + "bbox": [ + 148, + 419, + 223, + 431 + ], + "score": 0.91, + "content": "m \\in \\{ 1 , \\cdots , M \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 418, + 506, + 432 + ], + "score": 1.0, + "content": ". A full description of our method is given in Algorithm 7. For our", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 189, + 443 + ], + "score": 1.0, + "content": "experiments, we use", + "type": "text" + }, + { + "bbox": [ + 190, + 431, + 228, + 441 + ], + "score": 0.8, + "content": "N = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 430, + 231, + 443 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 231, + 430, + 271, + 441 + ], + "score": 0.76, + "content": "M = 5 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 430, + 275, + 443 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 275, + 430, + 342, + 442 + ], + "score": 0.68, + "content": "\\alpha = 0 . 1 , \\beta = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 430, + 506, + 443 + ], + "score": 1.0, + "content": ". In Figure 14, we plot the distribution of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "unique labels among the 500 training clients. Each client has only a fraction of the overall labels in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "the distribution. Moreover, there is variance in the number of unique labels, with most clients having", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "score": 1.0, + "content": "between 20 and 30, and some having over 40. Some client datasets have very few unique labels.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "While this is primarily an artifact of performing without replacement sampling, this helps increase the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "heterogeneity of the dataset in a way that can reflect practical concerns, as in many settings, clients", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 495, + 315, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 315, + 507 + ], + "score": 1.0, + "content": "may only have a few types of labels in their dataset.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 208, + 551, + 385, + 694 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 208, + 551, + 385, + 694 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 208, + 551, + 385, + 694 + ], + "spans": [ + { + "bbox": [ + 208, + 551, + 385, + 694 + ], + "score": 0.965, + "type": "image", + "image_path": "759e0a67cf8e490c2d7eb60abe974fcab397e55f7a8e0758467c8c94ae7999c1.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 551, + 385, + 565.3 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 208, + 565.3, + 385, + 579.5999999999999 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 208, + 579.5999999999999, + 385, + 593.8999999999999 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 208, + 593.8999999999999, + 385, + 608.1999999999998 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 208, + 608.1999999999998, + 385, + 622.4999999999998 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 208, + 622.4999999999998, + 385, + 636.7999999999997 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 208, + 636.7999999999997, + 385, + 651.0999999999997 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 208, + 651.0999999999997, + 385, + 665.3999999999996 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 208, + 665.3999999999996, + 385, + 679.6999999999996 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 208, + 679.6999999999996, + 385, + 693.9999999999995 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 704, + 504, + 726 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 703, + 506, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 506, + 718 + ], + "score": 1.0, + "content": "Figure 14: The distribution of the number of unique labels among training client datasets in our", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 715, + 228, + 727 + ], + "spans": [ + { + "bbox": [ + 106, + 715, + 228, + 727 + ], + "score": 1.0, + "content": "federated CIFAR-100 dataset.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "index": 38.5 + } + ], + "page_idx": 37, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "38", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 291, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 294, + 96 + ], + "score": 1.0, + "content": "Algorithm 7 Creating a federated CIFAR-100", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "list", + "bbox": [ + 107, + 95, + 311, + 295 + ], + "lines": [ + { + "bbox": [ + 115, + 93, + 249, + 110 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 141, + 110 + ], + "score": 1.0, + "content": "Input:", + "type": "text" + }, + { + "bbox": [ + 142, + 96, + 249, + 107 + ], + "score": 0.5, + "content": "N , M \\in \\mathbb { Z } _ { > 0 } , \\alpha , \\beta \\in \\mathbb { R } _ { \\ge 0 }", + "type": "inline_equation" + } + ], + "index": 1, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 116, + 106, + 209, + 118 + ], + "spans": [ + { + "bbox": [ + 116, + 106, + 131, + 118 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 132, + 107, + 196, + 117 + ], + "score": 0.52, + "content": "m = 1 , \\cdots , M", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 106, + 209, + 118 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 130, + 116, + 243, + 131 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 163, + 131 + ], + "score": 1.0, + "content": "Sample", + "type": "text" + }, + { + "bbox": [ + 164, + 117, + 243, + 129 + ], + "score": 0.76, + "content": "\\theta _ { r } \\sim \\operatorname { D i r } ( \\alpha , | G [ r ] | )", + "type": "inline_equation" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 130, + 127, + 215, + 141 + ], + "spans": [ + { + "bbox": [ + 130, + 127, + 147, + 141 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 147, + 128, + 199, + 139 + ], + "score": 0.78, + "content": "c \\in { \\mathcal { C } } \\cap G [ r ]", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 127, + 215, + 141 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 144, + 136, + 257, + 153 + ], + "spans": [ + { + "bbox": [ + 144, + 136, + 178, + 153 + ], + "score": 1.0, + "content": "Sample", + "type": "text" + }, + { + "bbox": [ + 179, + 139, + 257, + 151 + ], + "score": 0.81, + "content": "\\theta _ { c } \\sim \\operatorname { D i r } ( \\beta , | G [ c ] | )", + "type": "inline_equation" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 151, + 168, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 151, + 168, + 163 + ], + "score": 0.57, + "content": "D _ { m } = \\varnothing", + "type": "inline_equation", + "image_path": "88b407760e00cd79137d458acf36234cae322ad3e963e502b577cfcd93cad360.jpg" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 130, + 162, + 214, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 146, + 175 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 146, + 164, + 199, + 174 + ], + "score": 0.83, + "content": "n = 1 , \\cdots N", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 162, + 214, + 175 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 172, + 264, + 186 + ], + "spans": [ + { + "bbox": [ + 145, + 172, + 247, + 186 + ], + "score": 1.0, + "content": "Sample c ∼ Multinomial", + "type": "text" + }, + { + "bbox": [ + 248, + 174, + 264, + 185 + ], + "score": 0.49, + "content": "\\left( \\theta _ { r } \\right)", + "type": "inline_equation" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 183, + 264, + 196 + ], + "spans": [ + { + "bbox": [ + 145, + 183, + 249, + 196 + ], + "score": 1.0, + "content": "Sample y ∼ Multinomial", + "type": "text" + }, + { + "bbox": [ + 249, + 185, + 264, + 196 + ], + "score": 0.62, + "content": "\\left( \\theta _ { c } \\right)", + "type": "inline_equation" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 194, + 311, + 208 + ], + "spans": [ + { + "bbox": [ + 145, + 194, + 174, + 208 + ], + "score": 1.0, + "content": "Select", + "type": "text" + }, + { + "bbox": [ + 174, + 195, + 225, + 206 + ], + "score": 0.82, + "content": "( x , c , y ) \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 194, + 311, + 208 + ], + "score": 1.0, + "content": "uniformly at random", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 147, + 206, + 246, + 217 + ], + "spans": [ + { + "bbox": [ + 147, + 206, + 246, + 217 + ], + "score": 0.59, + "content": "D _ { m } = D _ { m } \\cup \\{ ( x , c , y ) \\}", + "type": "inline_equation" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 217, + 222, + 228 + ], + "spans": [ + { + "bbox": [ + 146, + 217, + 222, + 228 + ], + "score": 0.35, + "content": "\\mathcal { S } = \\mathcal { S } \\backslash \\{ ( x , c , \\overset { \\cdot } { y } ) \\}", + "type": "inline_equation" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 226, + 210, + 240 + ], + "spans": [ + { + "bbox": [ + 145, + 226, + 155, + 240 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 155, + 228, + 185, + 239 + ], + "score": 0.68, + "content": "\\begin{array} { r } { { \\cal S } _ { y } = \\emptyset } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 226, + 210, + 240 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 158, + 235, + 214, + 252 + ], + "spans": [ + { + "bbox": [ + 158, + 235, + 214, + 252 + ], + "score": 1.0, + "content": "G = G\\{y}", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 159, + 245, + 276, + 262 + ], + "spans": [ + { + "bbox": [ + 159, + 245, + 276, + 262 + ], + "score": 1.0, + "content": "θc = RENORMALIZE(θc, y)", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 160, + 258, + 223, + 271 + ], + "spans": [ + { + "bbox": [ + 160, + 258, + 223, + 271 + ], + "score": 1.0, + "content": "if Sc = ∅ then", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 174, + 268, + 227, + 284 + ], + "spans": [ + { + "bbox": [ + 174, + 268, + 227, + 284 + ], + "score": 1.0, + "content": "G = G\\{c}", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 174, + 279, + 291, + 294 + ], + "spans": [ + { + "bbox": [ + 174, + 279, + 291, + 294 + ], + "score": 1.0, + "content": "θr = RENORMALIZE(θr, c)", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + } + ], + "index": 9.5, + "bbox_fs": [ + 115, + 93, + 311, + 294 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 307, + 505, + 387 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 307, + 505, + 387 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 307, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 387 + ], + "score": 0.49, + "html": "
Algorithm8RENORMALIZE
Initialization: θ= (p1,...,Pk),i∈[K]
a=∑k≠iPk
fork ∈[K],k≠ido
Pk=pk/a
Return0' = (pi,···P'-1,P'+1,···,PK)
", + "type": "table", + "image_path": "efaa025501c7065b26a782907225a109bc4617af98f087301f0971bf59bbce79.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 106, + 307, + 505, + 333.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 106, + 333.6666666666667, + 505, + 360.33333333333337 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 106, + 360.33333333333337, + 505, + 387.00000000000006 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 408, + 506, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 123, + 421 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 409, + 135, + 419 + ], + "score": 0.72, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 407, + 261, + 421 + ], + "score": 1.0, + "content": "denote the number of clients,", + "type": "text" + }, + { + "bbox": [ + 262, + 409, + 272, + 419 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 407, + 439, + 421 + ], + "score": 1.0, + "content": "the number of examples per client, and", + "type": "text" + }, + { + "bbox": [ + 439, + 409, + 457, + 420 + ], + "score": 0.9, + "content": "D _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "the dataset", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 147, + 432 + ], + "score": 1.0, + "content": "for client", + "type": "text" + }, + { + "bbox": [ + 148, + 419, + 223, + 431 + ], + "score": 0.91, + "content": "m \\in \\{ 1 , \\cdots , M \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 418, + 506, + 432 + ], + "score": 1.0, + "content": ". A full description of our method is given in Algorithm 7. For our", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 189, + 443 + ], + "score": 1.0, + "content": "experiments, we use", + "type": "text" + }, + { + "bbox": [ + 190, + 431, + 228, + 441 + ], + "score": 0.8, + "content": "N = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 430, + 231, + 443 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 231, + 430, + 271, + 441 + ], + "score": 0.76, + "content": "M = 5 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 430, + 275, + 443 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 275, + 430, + 342, + 442 + ], + "score": 0.68, + "content": "\\alpha = 0 . 1 , \\beta = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 430, + 506, + 443 + ], + "score": 1.0, + "content": ". In Figure 14, we plot the distribution of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "unique labels among the 500 training clients. Each client has only a fraction of the overall labels in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "the distribution. Moreover, there is variance in the number of unique labels, with most clients having", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 506, + 475 + ], + "score": 1.0, + "content": "between 20 and 30, and some having over 40. Some client datasets have very few unique labels.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "While this is primarily an artifact of performing without replacement sampling, this helps increase the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "heterogeneity of the dataset in a way that can reflect practical concerns, as in many settings, clients", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 495, + 315, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 315, + 507 + ], + "score": 1.0, + "content": "may only have a few types of labels in their dataset.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 407, + 506, + 507 + ] + }, + { + "type": "image", + "bbox": [ + 208, + 551, + 385, + 694 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 208, + 551, + 385, + 694 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 208, + 551, + 385, + 694 + ], + "spans": [ + { + "bbox": [ + 208, + 551, + 385, + 694 + ], + "score": 0.965, + "type": "image", + "image_path": "759e0a67cf8e490c2d7eb60abe974fcab397e55f7a8e0758467c8c94ae7999c1.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 551, + 385, + 565.3 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 208, + 565.3, + 385, + 579.5999999999999 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 208, + 579.5999999999999, + 385, + 593.8999999999999 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 208, + 593.8999999999999, + 385, + 608.1999999999998 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 208, + 608.1999999999998, + 385, + 622.4999999999998 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 208, + 622.4999999999998, + 385, + 636.7999999999997 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 208, + 636.7999999999997, + 385, + 651.0999999999997 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 208, + 651.0999999999997, + 385, + 665.3999999999996 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 208, + 665.3999999999996, + 385, + 679.6999999999996 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 208, + 679.6999999999996, + 385, + 693.9999999999995 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 704, + 504, + 726 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 703, + 506, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 506, + 718 + ], + "score": 1.0, + "content": "Figure 14: The distribution of the number of unique labels among training client datasets in our", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 715, + 228, + 727 + ], + "spans": [ + { + "bbox": [ + 106, + 715, + 228, + 727 + ], + "score": 1.0, + "content": "federated CIFAR-100 dataset.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "index": 38.5 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/LkFG3lB13U5/LkFG3lB13U5_model.json b/parse/train/LkFG3lB13U5/LkFG3lB13U5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b016eb202d2fa83ed65a1497c90e2563a8884a1c --- /dev/null +++ b/parse/train/LkFG3lB13U5/LkFG3lB13U5_model.json @@ -0,0 +1,70416 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1397, + 1404, + 1397, + 1404, + 1787, + 298, + 1787 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 398, + 622, + 1303, + 622, + 1303, + 981, + 398, + 981 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1801, + 1402, + 1801, + 1402, + 1982, + 298, + 1982 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 1112, + 1404, + 1112, + 1404, + 1381, + 298, + 1381 + ], + "score": 0.965 + }, + { + "category_id": 0, + "poly": [ + 300, + 220, + 1088, + 220, + 1088, + 269, + 300, + 269 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 333, + 2006, + 749, + 2006, + 749, + 2033, + 333, + 2033 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 302, + 1042, + 573, + 1042, + 573, + 1077, + 302, + 1077 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 773, + 552, + 926, + 552, + 926, + 585, + 773, + 585 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 104, + 300, + 104 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.736 + }, + { + "category_id": 1, + "poly": [ + 317, + 384, + 1247, + 384, + 1247, + 472, + 317, + 472 + ], + "score": 0.531 + }, + { + "category_id": 1, + "poly": [ + 317, + 319, + 1314, + 319, + 1314, + 382, + 317, + 382 + ], + "score": 0.249 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 217.0, + 1095.0, + 217.0, + 1095.0, + 275.0, + 293.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 2003.0, + 753.0, + 2003.0, + 753.0, + 2037.0, + 331.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 579.0, + 1039.0, + 579.0, + 1086.0, + 294.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 551.0, + 932.0, + 551.0, + 932.0, + 588.0, + 769.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1398.0, + 1405.0, + 1398.0, + 1405.0, + 1430.0, + 297.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1429.0, + 1405.0, + 1429.0, + 1405.0, + 1461.0, + 296.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1457.0, + 1409.0, + 1457.0, + 1409.0, + 1491.0, + 293.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1485.0, + 1405.0, + 1485.0, + 1405.0, + 1519.0, + 295.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1513.0, + 1406.0, + 1513.0, + 1406.0, + 1551.0, + 293.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1543.0, + 1405.0, + 1543.0, + 1405.0, + 1582.0, + 293.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 1406.0, + 1574.0, + 1406.0, + 1614.0, + 293.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1603.0, + 1406.0, + 1603.0, + 1406.0, + 1643.0, + 293.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1634.0, + 1404.0, + 1634.0, + 1404.0, + 1670.0, + 294.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1662.0, + 1405.0, + 1662.0, + 1405.0, + 1703.0, + 292.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1695.0, + 1405.0, + 1695.0, + 1405.0, + 1731.0, + 294.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1721.0, + 1406.0, + 1721.0, + 1406.0, + 1764.0, + 292.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1755.0, + 1338.0, + 1755.0, + 1338.0, + 1791.0, + 294.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 619.0, + 1305.0, + 619.0, + 1305.0, + 657.0, + 393.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 650.0, + 1306.0, + 650.0, + 1306.0, + 688.0, + 393.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 682.0, + 1307.0, + 682.0, + 1307.0, + 713.0, + 394.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 712.0, + 1305.0, + 712.0, + 1305.0, + 744.0, + 395.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 741.0, + 1306.0, + 741.0, + 1306.0, + 776.0, + 392.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 773.0, + 1308.0, + 773.0, + 1308.0, + 805.0, + 394.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 803.0, + 1308.0, + 803.0, + 1308.0, + 834.0, + 394.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 831.0, + 1307.0, + 831.0, + 1307.0, + 866.0, + 394.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 860.0, + 1305.0, + 860.0, + 1305.0, + 897.0, + 392.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 891.0, + 1304.0, + 891.0, + 1304.0, + 924.0, + 394.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 920.0, + 1304.0, + 920.0, + 1304.0, + 954.0, + 394.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 949.0, + 788.0, + 949.0, + 788.0, + 987.0, + 392.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1797.0, + 1406.0, + 1797.0, + 1406.0, + 1839.0, + 293.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1830.0, + 1407.0, + 1830.0, + 1407.0, + 1869.0, + 293.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1860.0, + 1407.0, + 1860.0, + 1407.0, + 1896.0, + 292.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1890.0, + 1407.0, + 1890.0, + 1407.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1922.0, + 1404.0, + 1922.0, + 1404.0, + 1953.0, + 296.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1949.0, + 1406.0, + 1949.0, + 1406.0, + 1987.0, + 293.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1109.0, + 1405.0, + 1109.0, + 1405.0, + 1145.0, + 294.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1139.0, + 1406.0, + 1139.0, + 1406.0, + 1175.0, + 293.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1170.0, + 1405.0, + 1170.0, + 1405.0, + 1204.0, + 292.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1200.0, + 1408.0, + 1200.0, + 1408.0, + 1236.0, + 294.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1229.0, + 1406.0, + 1229.0, + 1406.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1258.0, + 1406.0, + 1258.0, + 1406.0, + 1298.0, + 293.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1290.0, + 1408.0, + 1290.0, + 1408.0, + 1326.0, + 294.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1321.0, + 1404.0, + 1321.0, + 1404.0, + 1353.0, + 296.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1350.0, + 1405.0, + 1350.0, + 1405.0, + 1382.0, + 296.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 379.0, + 511.0, + 379.0, + 511.0, + 414.0, + 311.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 409.0, + 1245.0, + 409.0, + 1245.0, + 446.0, + 314.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 440.0, + 920.0, + 440.0, + 920.0, + 476.0, + 312.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 317.0, + 1317.0, + 317.0, + 1317.0, + 358.0, + 310.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 351.0, + 978.0, + 351.0, + 978.0, + 386.0, + 313.0, + 386.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 837, + 1405, + 837, + 1405, + 1226, + 297, + 1226 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1242, + 1405, + 1242, + 1405, + 1632, + 297, + 1632 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1404, + 229, + 1404, + 352, + 299, + 352 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 697, + 1869, + 1002, + 1869, + 1002, + 1955, + 697, + 1955 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 299, + 1968, + 1404, + 1968, + 1404, + 2037, + 299, + 2037 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 297, + 1645, + 1403, + 1645, + 1403, + 1713, + 297, + 1713 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 296, + 1821, + 1063, + 1821, + 1063, + 1855, + 296, + 1855 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 299, + 1753, + 886, + 1753, + 886, + 1790, + 299, + 1790 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 303, + 365, + 1326, + 365, + 1326, + 398, + 303, + 398 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1895, + 1400, + 1895, + 1400, + 1925, + 1366, + 1925 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 304, + 420, + 1407, + 420, + 1407, + 813, + 304, + 813 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.718 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.244 + }, + { + "category_id": 14, + "poly": [ + 696, + 1866, + 1001, + 1866, + 1001, + 1956, + 696, + 1956 + ], + "score": 0.94, + "latex": "\\operatorname* { m i n } _ { x \\in \\mathbb { R } ^ { d } } f ( x ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } F _ { i } ( x ) ," + }, + { + "category_id": 13, + "poly": [ + 375, + 1970, + 664, + 1970, + 664, + 2003, + 375, + 2003 + ], + "score": 0.92, + "latex": "F _ { i } ( x ) = \\mathbb { E } _ { z \\sim \\mathcal { D } _ { i } } [ f _ { i } ( x , z ) ]" + }, + { + "category_id": 13, + "poly": [ + 937, + 1679, + 988, + 1679, + 988, + 1714, + 937, + 1714 + ], + "score": 0.92, + "latex": "[ \\theta _ { i } ] _ { j }" + }, + { + "category_id": 13, + "poly": [ + 602, + 1679, + 693, + 1679, + 693, + 1710, + 602, + 1710 + ], + "score": 0.92, + "latex": "\\theta _ { i } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 844, + 1680, + 885, + 1680, + 885, + 1714, + 844, + 1714 + ], + "score": 0.91, + "latex": "\\theta _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 662, + 1644, + 741, + 1644, + 741, + 1678, + 662, + 1678 + ], + "score": 0.91, + "latex": "{ \\sqrt { a } } , a ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1094, + 1971, + 1176, + 1971, + 1176, + 2000, + 1094, + 2000 + ], + "score": 0.9, + "latex": "z \\in { \\mathcal { Z } }" + }, + { + "category_id": 13, + "poly": [ + 462, + 1643, + 573, + 1643, + 573, + 1678, + 462, + 1678 + ], + "score": 0.9, + "latex": "a , b \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 845, + 2004, + 880, + 2004, + 880, + 2038, + 845, + 2038 + ], + "score": 0.9, + "latex": "\\mathcal { D } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 795, + 1646, + 839, + 1646, + 839, + 1679, + 795, + 1679 + ], + "score": 0.89, + "latex": "a / b" + }, + { + "category_id": 13, + "poly": [ + 1239, + 1972, + 1273, + 1972, + 1273, + 2001, + 1239, + 2001 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1135, + 1678, + 1169, + 1678, + 1169, + 1712, + 1135, + 1712 + ], + "score": 0.88, + "latex": "j ^ { \\mathrm { t h } }" + }, + { + "category_id": 13, + "poly": [ + 1312, + 2003, + 1342, + 2003, + 1342, + 2034, + 1312, + 2034 + ], + "score": 0.88, + "latex": "F _ { i }" + }, + { + "category_id": 13, + "poly": [ + 980, + 1967, + 1011, + 1967, + 1011, + 1998, + 980, + 1998 + ], + "score": 0.87, + "latex": "i ^ { \\mathrm { { t h } } }" + }, + { + "category_id": 13, + "poly": [ + 519, + 2003, + 549, + 2003, + 549, + 2030, + 519, + 2030 + ], + "score": 0.84, + "latex": "i ^ { \\mathrm { { t h } } }" + }, + { + "category_id": 13, + "poly": [ + 678, + 2003, + 744, + 2003, + 744, + 2035, + 678, + 2035 + ], + "score": 0.83, + "latex": "i \\neq j" + }, + { + "category_id": 13, + "poly": [ + 757, + 2003, + 790, + 2003, + 790, + 2034, + 757, + 2034 + ], + "score": 0.72, + "latex": "\\mathcal { D } _ { i }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1751.0, + 888.0, + 1751.0, + 888.0, + 1795.0, + 292.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 837.0, + 1407.0, + 837.0, + 1407.0, + 868.0, + 296.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 867.0, + 1403.0, + 867.0, + 1403.0, + 899.0, + 295.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 897.0, + 1405.0, + 897.0, + 1405.0, + 931.0, + 294.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 926.0, + 1405.0, + 926.0, + 1405.0, + 960.0, + 294.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 956.0, + 1403.0, + 956.0, + 1403.0, + 988.0, + 296.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 983.0, + 1406.0, + 983.0, + 1406.0, + 1021.0, + 292.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1015.0, + 1405.0, + 1015.0, + 1405.0, + 1049.0, + 294.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 1406.0, + 1044.0, + 1406.0, + 1080.0, + 294.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1073.0, + 1408.0, + 1073.0, + 1408.0, + 1112.0, + 292.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1104.0, + 1405.0, + 1104.0, + 1405.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1135.0, + 1405.0, + 1135.0, + 1405.0, + 1169.0, + 294.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1164.0, + 1406.0, + 1164.0, + 1406.0, + 1200.0, + 295.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1195.0, + 1192.0, + 1195.0, + 1192.0, + 1229.0, + 294.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1240.0, + 1403.0, + 1240.0, + 1403.0, + 1272.0, + 296.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1270.0, + 1407.0, + 1270.0, + 1407.0, + 1306.0, + 294.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 1406.0, + 1298.0, + 1406.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1330.0, + 1405.0, + 1330.0, + 1405.0, + 1366.0, + 295.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1360.0, + 1405.0, + 1360.0, + 1405.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1426.0, + 295.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1420.0, + 1405.0, + 1420.0, + 1405.0, + 1455.0, + 292.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1447.0, + 1406.0, + 1447.0, + 1406.0, + 1484.0, + 294.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1407.0, + 1479.0, + 1407.0, + 1515.0, + 295.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1507.0, + 1406.0, + 1507.0, + 1406.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1540.0, + 1406.0, + 1540.0, + 1406.0, + 1576.0, + 295.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1570.0, + 1406.0, + 1570.0, + 1406.0, + 1604.0, + 294.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1602.0, + 1290.0, + 1602.0, + 1290.0, + 1634.0, + 296.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 259.0, + 1408.0, + 259.0, + 1408.0, + 295.0, + 293.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1283.0, + 321.0, + 1283.0, + 353.0, + 294.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1964.0, + 374.0, + 1964.0, + 374.0, + 2005.0, + 293.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1964.0, + 979.0, + 1964.0, + 979.0, + 2005.0, + 665.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1964.0, + 1093.0, + 1964.0, + 1093.0, + 2005.0, + 1012.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1964.0, + 1238.0, + 1964.0, + 1238.0, + 2005.0, + 1177.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1964.0, + 1406.0, + 1964.0, + 1406.0, + 2005.0, + 1274.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1996.0, + 518.0, + 1996.0, + 518.0, + 2041.0, + 293.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1996.0, + 677.0, + 1996.0, + 677.0, + 2041.0, + 550.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1996.0, + 756.0, + 1996.0, + 756.0, + 2041.0, + 745.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1996.0, + 844.0, + 1996.0, + 844.0, + 2041.0, + 791.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1996.0, + 1311.0, + 1996.0, + 1311.0, + 2041.0, + 881.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1996.0, + 1406.0, + 1996.0, + 1406.0, + 2041.0, + 1343.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1642.0, + 461.0, + 1642.0, + 461.0, + 1682.0, + 295.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1642.0, + 661.0, + 1642.0, + 661.0, + 1682.0, + 574.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1642.0, + 794.0, + 1642.0, + 794.0, + 1682.0, + 742.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1642.0, + 1405.0, + 1642.0, + 1405.0, + 1682.0, + 840.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1675.0, + 601.0, + 1675.0, + 601.0, + 1716.0, + 293.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1675.0, + 843.0, + 1675.0, + 843.0, + 1716.0, + 694.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1675.0, + 936.0, + 1675.0, + 936.0, + 1716.0, + 886.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 1675.0, + 1134.0, + 1675.0, + 1134.0, + 1716.0, + 989.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1675.0, + 1302.0, + 1675.0, + 1302.0, + 1716.0, + 1170.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1821.0, + 1063.0, + 1821.0, + 1063.0, + 1858.0, + 295.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 360.0, + 1332.0, + 360.0, + 1332.0, + 405.0, + 293.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 419.0, + 1406.0, + 419.0, + 1406.0, + 457.0, + 304.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 452.0, + 1343.0, + 452.0, + 1343.0, + 488.0, + 328.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 489.0, + 1404.0, + 489.0, + 1404.0, + 528.0, + 307.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 523.0, + 1406.0, + 523.0, + 1406.0, + 559.0, + 330.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 552.0, + 1404.0, + 552.0, + 1404.0, + 587.0, + 330.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 584.0, + 1349.0, + 584.0, + 1349.0, + 616.0, + 330.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 620.0, + 1406.0, + 620.0, + 1406.0, + 658.0, + 315.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 651.0, + 1406.0, + 651.0, + 1406.0, + 687.0, + 329.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 682.0, + 1408.0, + 682.0, + 1408.0, + 718.0, + 328.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 718.0, + 1406.0, + 718.0, + 1406.0, + 760.0, + 305.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 751.0, + 1406.0, + 751.0, + 1406.0, + 786.0, + 329.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 779.0, + 841.0, + 779.0, + 841.0, + 819.0, + 329.0, + 819.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 617, + 1405, + 617, + 1405, + 839, + 296, + 839 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1193, + 1405, + 1193, + 1405, + 1348, + 297, + 1348 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 850, + 1404, + 850, + 1404, + 1003, + 297, + 1003 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 375, + 1405, + 375, + 1405, + 514, + 297, + 514 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 296, + 1851, + 1405, + 1851, + 1405, + 2036, + 296, + 2036 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 587, + 1094, + 1108, + 1094, + 1108, + 1177, + 587, + 1177 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 299, + 1016, + 1399, + 1016, + 1399, + 1080, + 299, + 1080 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1401, + 229, + 1401, + 293, + 294, + 293 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 298, + 521, + 1402, + 521, + 1402, + 592, + 298, + 592 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 297, + 300, + 1400, + 300, + 1400, + 368, + 297, + 368 + ], + "score": 0.931 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.913 + }, + { + "category_id": 5, + "poly": [ + 305, + 1421, + 1017, + 1421, + 1017, + 1802, + 305, + 1802 + ], + "score": 0.9, + "html": "
1: Input:xo,CLIENTOPT,SERVEROPT
2: fort=O,.,T-1do
3: Samplea subset Sofclients
4: xt,0=xt
5: for each client i ∈ S in parallel do
6: for k=0,.,K-1 do
Compute an unbiased estimate gt,k of VFi(xt,k) 7:
x,+1=CLEO(,,,t) 8:
9: △=x,K-xt
10: △t=∑ies△t 11: Xt+1 = SERVEROPT(xt,-△t,n,t)
" + }, + { + "category_id": 0, + "poly": [ + 297, + 1383, + 550, + 1383, + 550, + 1416, + 297, + 1416 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.65 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.417 + }, + { + "category_id": 14, + "poly": [ + 588, + 1093, + 1111, + 1093, + 1111, + 1178, + 588, + 1178 + ], + "score": 0.93, + "latex": "x _ { t + 1 } = { \\frac { 1 } { | S | } } \\sum _ { i \\in S } x _ { i } ^ { t } = x _ { t } - { \\frac { 1 } { | S | } } \\sum _ { i \\in S } \\left( x _ { t } - x _ { i } ^ { t } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 669, + 260, + 759, + 260, + 759, + 294, + 669, + 294 + ], + "score": 0.93, + "latex": "\\nabla F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 870, + 522, + 959, + 522, + 959, + 555, + 870, + 555 + ], + "score": 0.93, + "latex": "f _ { i } ( x , z )" + }, + { + "category_id": 13, + "poly": [ + 940, + 616, + 1057, + 616, + 1057, + 661, + 940, + 661 + ], + "score": 0.93, + "latex": "\\textstyle \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1125, + 301, + 1214, + 301, + 1214, + 336, + 1125, + 336 + ], + "score": 0.93, + "latex": "i \\in [ m ]" + }, + { + "category_id": 13, + "poly": [ + 342, + 1195, + 507, + 1195, + 507, + 1230, + 342, + 1230 + ], + "score": 0.92, + "latex": "\\Delta _ { i } ^ { t } : = x _ { i } ^ { t } - x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 635, + 333, + 746, + 333, + 746, + 368, + 635, + 368 + ], + "score": 0.92, + "latex": "x , y \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 999, + 555, + 1077, + 555, + 1077, + 589, + 999, + 589 + ], + "score": 0.91, + "latex": "j \\in [ d ]" + }, + { + "category_id": 13, + "poly": [ + 708, + 555, + 919, + 555, + 919, + 590, + 708, + 590 + ], + "score": 0.91, + "latex": "| [ \\nabla f _ { i } ( x , z ) ] _ { j } | \\leq G" + }, + { + "category_id": 13, + "poly": [ + 650, + 807, + 733, + 807, + 733, + 840, + 650, + 840 + ], + "score": 0.91, + "latex": "\\sigma _ { g } = 0" + }, + { + "category_id": 13, + "poly": [ + 436, + 479, + 514, + 479, + 514, + 514, + 436, + 514 + ], + "score": 0.91, + "latex": "j \\in [ d ]" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1228, + 1092, + 1228, + 1092, + 1259, + 1020, + 1259 + ], + "score": 0.91, + "latex": "\\eta = 1" + }, + { + "category_id": 13, + "poly": [ + 1107, + 616, + 1228, + 616, + 1228, + 660, + 1107, + 660 + ], + "score": 0.91, + "latex": "\\textstyle \\sum _ { j = 1 } ^ { d } \\sigma _ { g , j } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 334, + 550, + 334, + 550, + 370, + 297, + 370 + ], + "score": 0.91, + "latex": "\\nabla F _ { i } ( y ) \\| \\leq L \\| x - y \\|" + }, + { + "category_id": 13, + "poly": [ + 1095, + 407, + 1191, + 407, + 1191, + 440, + 1095, + 440 + ], + "score": 0.91, + "latex": "i \\in [ m ]" + }, + { + "category_id": 13, + "poly": [ + 297, + 260, + 360, + 260, + 360, + 295, + 297, + 295 + ], + "score": 0.91, + "latex": "g _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1268, + 300, + 1404, + 300, + 1404, + 337, + 1268, + 337 + ], + "score": 0.91, + "latex": "\\parallel \\nabla F _ { i } ( x ) -" + }, + { + "category_id": 13, + "poly": [ + 558, + 1193, + 819, + 1193, + 819, + 1232, + 558, + 1232 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\Delta _ { t } : = \\left( { 1 } / { | { S } | } \\right) \\sum _ { i \\in \\mathcal { S } } \\Delta _ { i } ^ { t } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 530, + 1049, + 595, + 1049, + 595, + 1076, + 530, + 1076 + ], + "score": 0.9, + "latex": "i \\in S" + }, + { + "category_id": 13, + "poly": [ + 297, + 477, + 383, + 477, + 383, + 510, + 297, + 510 + ], + "score": 0.9, + "latex": "\\boldsymbol { x } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 718, + 621, + 750, + 621, + 750, + 657, + 718, + 657 + ], + "score": 0.9, + "latex": "\\sigma _ { l } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 799, + 620, + 831, + 620, + 831, + 660, + 799, + 660 + ], + "score": 0.9, + "latex": "\\sigma _ { g } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 296, + 405, + 750, + 405, + 750, + 444, + 296, + 444 + ], + "score": 0.9, + "latex": "\\mathbb { E } [ \\| \\nabla [ \\widehat { f } _ { i } ( x , z ) ] _ { j } - [ \\nabla F _ { i } ( x ) ] _ { j } \\| ^ { 2 } ] = \\sigma _ { l , j } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 894, + 1601, + 1011, + 1601, + 1011, + 1641, + 894, + 1641 + ], + "score": 0.89, + "latex": "\\nabla F _ { i } ( x _ { i , k } ^ { t } )" + }, + { + "category_id": 13, + "poly": [ + 469, + 1573, + 678, + 1573, + 678, + 1604, + 469, + 1604 + ], + "score": 0.89, + "latex": "k = 0 , \\cdots , K - 1" + }, + { + "category_id": 13, + "poly": [ + 394, + 554, + 480, + 554, + 480, + 585, + 394, + 585 + ], + "score": 0.89, + "latex": "\\boldsymbol { x } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1247, + 1017, + 1277, + 1017, + 1277, + 1051, + 1247, + 1051 + ], + "score": 0.89, + "latex": "\\ v x _ { i } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 532, + 556, + 606, + 556, + 606, + 585, + 532, + 585 + ], + "score": 0.89, + "latex": "z \\in { \\mathcal { Z } }" + }, + { + "category_id": 13, + "poly": [ + 386, + 1452, + 583, + 1452, + 583, + 1482, + 386, + 1482 + ], + "score": 0.88, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 788, + 441, + 1317, + 441, + 1317, + 481, + 788, + 481 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\left( 1 / m \\right) \\sum _ { i = 1 } ^ { m } \\| \\nabla [ F _ { i } ( x ) ] _ { j } - [ \\nabla f ( x ) ] _ { j } \\| ^ { 2 } \\leq \\sigma _ { g , j } ^ { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 555, + 383, + 555, + 383, + 589, + 297, + 589 + ], + "score": 0.88, + "latex": "i \\in [ m ]" + }, + { + "category_id": 13, + "poly": [ + 553, + 1544, + 619, + 1544, + 619, + 1571, + 553, + 1571 + ], + "score": 0.88, + "latex": "i \\in S" + }, + { + "category_id": 13, + "poly": [ + 869, + 303, + 898, + 303, + 898, + 333, + 869, + 333 + ], + "score": 0.88, + "latex": "F _ { i }" + }, + { + "category_id": 13, + "poly": [ + 750, + 1230, + 808, + 1230, + 808, + 1257, + 750, + 1257 + ], + "score": 0.87, + "latex": "\\bar { - } \\Delta _ { t }" + }, + { + "category_id": 13, + "poly": [ + 815, + 1602, + 859, + 1602, + 859, + 1640, + 815, + 1640 + ], + "score": 0.87, + "latex": "g _ { i , k } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 947, + 407, + 1036, + 407, + 1036, + 440, + 947, + 440 + ], + "score": 0.86, + "latex": "j \\in [ d ]" + }, + { + "category_id": 13, + "poly": [ + 1114, + 780, + 1146, + 780, + 1146, + 810, + 1114, + 810 + ], + "score": 0.86, + "latex": "\\sigma _ { g }" + }, + { + "category_id": 13, + "poly": [ + 839, + 407, + 933, + 407, + 933, + 437, + 839, + 437 + ], + "score": 0.85, + "latex": "x \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 817, + 1022, + 846, + 1022, + 846, + 1048, + 817, + 1048 + ], + "score": 0.85, + "latex": "x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 904, + 378, + 934, + 378, + 934, + 407, + 904, + 407 + ], + "score": 0.84, + "latex": "F _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1012, + 380, + 1039, + 380, + 1039, + 407, + 1012, + 407 + ], + "score": 0.84, + "latex": "\\sigma _ { l }" + }, + { + "category_id": 13, + "poly": [ + 596, + 1262, + 615, + 1262, + 615, + 1289, + 596, + 1289 + ], + "score": 0.81, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1056, + 1019, + 1079, + 1019, + 1079, + 1046, + 1056, + 1046 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 297, + 1890, + 322, + 1890, + 322, + 1916, + 297, + 1916 + ], + "score": 0.81, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 401, + 232, + 421, + 232, + 421, + 262, + 401, + 262 + ], + "score": 0.81, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 378, + 1889, + 397, + 1889, + 397, + 1916, + 378, + 1916 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 745, + 234, + 758, + 234, + 758, + 258, + 745, + 258 + ], + "score": 0.78, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 571, + 1482, + 594, + 1482, + 594, + 1508, + 571, + 1508 + ], + "score": 0.78, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 930, + 303, + 952, + 303, + 952, + 330, + 930, + 330 + ], + "score": 0.77, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 417, + 1426, + 448, + 1426, + 448, + 1451, + 417, + 1451 + ], + "score": 0.77, + "latex": "x _ { 0 }" + }, + { + "category_id": 14, + "poly": [ + 379, + 1635, + 901, + 1635, + 901, + 1797, + 379, + 1797 + ], + "score": 0.77, + "latex": "\\begin{array} { r l } & { \\qquad x _ { i , k + 1 } ^ { t } = \\mathrm { C L I E N T O P T } ( x _ { i , k } ^ { t } , g _ { i , k } ^ { t } , \\eta _ { l } , t ) } \\\\ & { \\qquad \\Delta _ { i } ^ { t } = x _ { i , K } ^ { t } - x _ { t } } \\\\ & { \\qquad \\Delta _ { t } = \\frac { 1 } { | \\mathcal { S } | } \\sum _ { i \\in \\mathcal { S } } \\Delta _ { i } ^ { t } } \\\\ & { x _ { t + 1 } = \\mathrm { S E R V E R O P T } ( x _ { t } , - \\Delta _ { t } , \\eta , t ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 551, + 1021, + 565, + 1021, + 565, + 1045, + 551, + 1045 + ], + "score": 0.76, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 805, + 237, + 824, + 237, + 824, + 258, + 805, + 258 + ], + "score": 0.75, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1023, + 523, + 1048, + 523, + 1048, + 551, + 1023, + 551 + ], + "score": 0.74, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 385, + 1508, + 495, + 1508, + 495, + 1547, + 385, + 1547 + ], + "score": 0.57, + "latex": "{ x _ { i , 0 } ^ { t } } \\dot { = } { x _ { t } }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1381.0, + 553.0, + 1381.0, + 553.0, + 1420.0, + 294.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 281.0, + 606.0, + 717.0, + 606.0, + 717.0, + 671.0, + 281.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 606.0, + 798.0, + 606.0, + 798.0, + 671.0, + 751.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 606.0, + 939.0, + 606.0, + 939.0, + 671.0, + 832.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 606.0, + 1106.0, + 606.0, + 1106.0, + 671.0, + 1058.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 606.0, + 1416.0, + 606.0, + 1416.0, + 671.0, + 1229.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 654.0, + 1407.0, + 654.0, + 1407.0, + 688.0, + 293.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 683.0, + 1408.0, + 683.0, + 1408.0, + 719.0, + 294.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 717.0, + 1405.0, + 717.0, + 1405.0, + 748.0, + 296.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 746.0, + 1403.0, + 746.0, + 1403.0, + 777.0, + 295.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 775.0, + 1113.0, + 775.0, + 1113.0, + 810.0, + 295.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 775.0, + 1406.0, + 775.0, + 1406.0, + 810.0, + 1147.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 802.0, + 649.0, + 802.0, + 649.0, + 843.0, + 295.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 802.0, + 1096.0, + 802.0, + 1096.0, + 843.0, + 734.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1188.0, + 341.0, + 1188.0, + 341.0, + 1240.0, + 290.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1188.0, + 557.0, + 1188.0, + 557.0, + 1240.0, + 508.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1188.0, + 1410.0, + 1188.0, + 1410.0, + 1240.0, + 820.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1225.0, + 749.0, + 1225.0, + 749.0, + 1263.0, + 295.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1225.0, + 1019.0, + 1225.0, + 1019.0, + 1263.0, + 809.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1225.0, + 1407.0, + 1225.0, + 1407.0, + 1263.0, + 1093.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1255.0, + 595.0, + 1255.0, + 595.0, + 1292.0, + 295.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1255.0, + 1406.0, + 1255.0, + 1406.0, + 1292.0, + 616.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1284.0, + 1405.0, + 1284.0, + 1405.0, + 1320.0, + 295.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1314.0, + 903.0, + 1314.0, + 903.0, + 1348.0, + 296.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 850.0, + 1405.0, + 850.0, + 1405.0, + 887.0, + 294.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 882.0, + 1404.0, + 882.0, + 1404.0, + 915.0, + 295.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 911.0, + 1405.0, + 911.0, + 1405.0, + 945.0, + 293.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 940.0, + 1405.0, + 940.0, + 1405.0, + 977.0, + 294.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 969.0, + 947.0, + 969.0, + 947.0, + 1006.0, + 294.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 371.0, + 903.0, + 371.0, + 903.0, + 412.0, + 294.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 371.0, + 1011.0, + 371.0, + 1011.0, + 412.0, + 935.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 371.0, + 1410.0, + 371.0, + 1410.0, + 412.0, + 1040.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 405.0, + 838.0, + 405.0, + 838.0, + 443.0, + 751.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 405.0, + 946.0, + 405.0, + 946.0, + 443.0, + 934.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 405.0, + 1094.0, + 405.0, + 1094.0, + 443.0, + 1037.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 405.0, + 1406.0, + 405.0, + 1406.0, + 443.0, + 1192.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 435.0, + 787.0, + 435.0, + 787.0, + 486.0, + 290.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 435.0, + 1408.0, + 435.0, + 1408.0, + 486.0, + 1318.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 473.0, + 296.0, + 473.0, + 296.0, + 517.0, + 291.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 473.0, + 435.0, + 473.0, + 435.0, + 517.0, + 384.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 473.0, + 525.0, + 473.0, + 525.0, + 517.0, + 515.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1889.0, + 293.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1882.0, + 296.0, + 1882.0, + 296.0, + 1918.0, + 291.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1882.0, + 377.0, + 1882.0, + 377.0, + 1918.0, + 323.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1882.0, + 1406.0, + 1882.0, + 1406.0, + 1918.0, + 398.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1949.0, + 294.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1979.0, + 295.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1975.0, + 1406.0, + 1975.0, + 1406.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2039.0, + 295.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1016.0, + 550.0, + 1016.0, + 550.0, + 1052.0, + 296.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 1016.0, + 816.0, + 1016.0, + 816.0, + 1052.0, + 566.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1016.0, + 1055.0, + 1016.0, + 1055.0, + 1052.0, + 847.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1016.0, + 1246.0, + 1016.0, + 1246.0, + 1052.0, + 1080.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1016.0, + 1403.0, + 1016.0, + 1403.0, + 1052.0, + 1278.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1045.0, + 529.0, + 1045.0, + 529.0, + 1083.0, + 294.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1045.0, + 1183.0, + 1045.0, + 1183.0, + 1083.0, + 596.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 400.0, + 229.0, + 400.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 229.0, + 744.0, + 229.0, + 744.0, + 265.0, + 422.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 229.0, + 804.0, + 229.0, + 804.0, + 265.0, + 759.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 825.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 257.0, + 296.0, + 257.0, + 296.0, + 298.0, + 293.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 257.0, + 668.0, + 257.0, + 668.0, + 298.0, + 361.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 257.0, + 1318.0, + 257.0, + 1318.0, + 298.0, + 760.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 515.0, + 869.0, + 515.0, + 869.0, + 562.0, + 292.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 515.0, + 1022.0, + 515.0, + 1022.0, + 562.0, + 960.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 515.0, + 1406.0, + 515.0, + 1406.0, + 562.0, + 1049.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 550.0, + 296.0, + 550.0, + 296.0, + 594.0, + 291.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 550.0, + 393.0, + 550.0, + 393.0, + 594.0, + 384.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 550.0, + 531.0, + 550.0, + 531.0, + 594.0, + 481.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 550.0, + 707.0, + 550.0, + 707.0, + 594.0, + 607.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 550.0, + 998.0, + 550.0, + 998.0, + 594.0, + 920.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 550.0, + 1088.0, + 550.0, + 1088.0, + 594.0, + 1078.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 299.0, + 868.0, + 299.0, + 868.0, + 339.0, + 295.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 299.0, + 929.0, + 299.0, + 929.0, + 339.0, + 899.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 299.0, + 1124.0, + 299.0, + 1124.0, + 339.0, + 953.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 299.0, + 1267.0, + 299.0, + 1267.0, + 339.0, + 1215.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 331.0, + 296.0, + 331.0, + 296.0, + 369.0, + 293.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 331.0, + 634.0, + 331.0, + 634.0, + 369.0, + 551.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 331.0, + 757.0, + 331.0, + 757.0, + 369.0, + 747.0, + 369.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 629, + 1403, + 629, + 1403, + 750, + 298, + 750 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 765, + 1404, + 765, + 1404, + 947, + 297, + 947 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 335, + 1405, + 335, + 1405, + 518, + 298, + 518 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1614, + 1406, + 1614, + 1406, + 1768, + 297, + 1768 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1401, + 229, + 1401, + 322, + 298, + 322 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 296, + 1774, + 1403, + 1774, + 1403, + 1843, + 296, + 1843 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 300, + 561, + 902, + 561, + 902, + 597, + 300, + 597 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 104, + 300, + 104 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 296, + 985, + 978, + 985, + 978, + 1024, + 296, + 1024 + ], + "score": 0.89 + }, + { + "category_id": 8, + "poly": [ + 596, + 1939, + 1253, + 1939, + 1253, + 2030, + 596, + 2030 + ], + "score": 0.85 + }, + { + "category_id": 8, + "poly": [ + 597, + 1860, + 1099, + 1860, + 1099, + 1934, + 597, + 1934 + ], + "score": 0.799 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.781 + }, + { + "category_id": 1, + "poly": [ + 302, + 1030, + 1026, + 1030, + 1026, + 1581, + 302, + 1581 + ], + "score": 0.642 + }, + { + "category_id": 1, + "poly": [ + 297, + 1967, + 456, + 1967, + 456, + 2002, + 297, + 2002 + ], + "score": 0.585 + }, + { + "category_id": 1, + "poly": [ + 297, + 1881, + 448, + 1881, + 448, + 1915, + 297, + 1915 + ], + "score": 0.577 + }, + { + "category_id": 9, + "poly": [ + 297, + 1881, + 448, + 1881, + 448, + 1915, + 297, + 1915 + ], + "score": 0.222 + }, + { + "category_id": 9, + "poly": [ + 297, + 1967, + 456, + 1967, + 456, + 2002, + 297, + 2002 + ], + "score": 0.115 + }, + { + "category_id": 14, + "poly": [ + 595, + 1858, + 1258, + 1858, + 1258, + 2033, + 595, + 2033 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { \\eta _ { l } \\le \\displaystyle \\frac { 1 } { K } \\operatorname* { m i n } \\left\\{ \\frac { 1 } { 1 6 L } , \\frac { 1 } { T ^ { 1 / 6 } } \\left[ \\frac { \\tau } { 1 2 0 L ^ { 2 } G } \\right] ^ { 1 / 3 } \\right\\} , } \\\\ & { \\eta _ { l } \\le \\displaystyle \\frac { 1 } { 3 K } \\operatorname* { m i n } \\left\\{ \\frac { 1 } { T ^ { 1 / 1 0 } } \\left[ \\frac { \\tau ^ { 3 } } { L ^ { 2 } G ^ { 3 } } \\right] ^ { 1 / 5 } , \\frac { 1 } { T ^ { 1 / 8 } } \\left[ \\frac { \\tau ^ { 2 } } { L ^ { 3 } G \\eta } \\right] ^ { 1 / 4 } \\right\\} . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 887, + 1775, + 1021, + 1775, + 1021, + 1811, + 887, + 1811 + ], + "score": 0.93, + "latex": "L , G , \\sigma _ { l } , \\sigma _ { g }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1807, + 431, + 1807, + 431, + 1847, + 297, + 1847 + ], + "score": 0.92, + "latex": "\\sigma _ { l } ^ { 2 } + 6 K \\sigma _ { g } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1299, + 1675, + 1397, + 1675, + 1397, + 1709, + 1299, + 1709 + ], + "score": 0.92, + "latex": "| S | < m" + }, + { + "category_id": 13, + "poly": [ + 1011, + 1646, + 1093, + 1646, + 1093, + 1677, + 1011, + 1677 + ], + "score": 0.92, + "latex": "\\beta _ { 1 } = 0" + }, + { + "category_id": 13, + "poly": [ + 495, + 1645, + 593, + 1645, + 593, + 1678, + 495, + 1678 + ], + "score": 0.91, + "latex": "\\bar { \\mathcal { S } } = [ m ]" + }, + { + "category_id": 13, + "poly": [ + 894, + 1214, + 1011, + 1214, + 1011, + 1254, + 894, + 1254 + ], + "score": 0.91, + "latex": "\\nabla F _ { i } ( x _ { i , k } ^ { t } )" + }, + { + "category_id": 13, + "poly": [ + 862, + 1033, + 1026, + 1033, + 1026, + 1067, + 862, + 1067 + ], + "score": 0.91, + "latex": "\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 814, + 1214, + 859, + 1214, + 859, + 1254, + 814, + 1254 + ], + "score": 0.91, + "latex": "g _ { i , k } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 469, + 1186, + 677, + 1186, + 677, + 1217, + 469, + 1217 + ], + "score": 0.89, + "latex": "k = 0 , \\cdots , K - \\bar { 1 }" + }, + { + "category_id": 13, + "poly": [ + 521, + 1678, + 603, + 1678, + 603, + 1707, + 521, + 1707 + ], + "score": 0.89, + "latex": "\\beta _ { 1 } > 0" + }, + { + "category_id": 13, + "poly": [ + 554, + 1158, + 619, + 1158, + 619, + 1185, + 554, + 1185 + ], + "score": 0.88, + "latex": "i \\in S" + }, + { + "category_id": 13, + "poly": [ + 504, + 1032, + 651, + 1032, + 651, + 1065, + 504, + 1065 + ], + "score": 0.88, + "latex": "x _ { 0 } , v _ { - 1 } \\ge \\tau ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 386, + 1065, + 583, + 1065, + 583, + 1095, + 386, + 1095 + ], + "score": 0.88, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 742, + 367, + 777, + 367, + 777, + 397, + 742, + 397 + ], + "score": 0.87, + "latex": "\\Delta _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1334, + 1773, + 1404, + 1773, + 1404, + 1808, + 1334, + 1808 + ], + "score": 0.85, + "latex": "\\sigma ^ { 2 } = " + }, + { + "category_id": 14, + "poly": [ + 382, + 1248, + 866, + 1248, + 866, + 1386, + 382, + 1386 + ], + "score": 0.84, + "latex": "\\begin{array} { c } { x _ { i , k + 1 } ^ { t } = x _ { i , k } ^ { t } - \\eta _ { l } g _ { i , k } ^ { t } } \\\\ { \\Delta _ { i } ^ { t } = x _ { i , K } ^ { t } - x _ { t } } \\\\ { \\Delta _ { t } = \\beta _ { 1 } \\Delta _ { t - 1 } + \\left( 1 - \\beta _ { 1 } \\right) \\left( \\frac { 1 } { \\left| S \\right| } \\sum _ { i \\in S } \\Delta _ { i } ^ { t } \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 680, + 891, + 705, + 891, + 705, + 917, + 680, + 917 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 752, + 235, + 778, + 235, + 778, + 263, + 752, + 263 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 431, + 921, + 457, + 921, + 457, + 948, + 431, + 948 + ], + "score": 0.82, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 865, + 1814, + 889, + 1814, + 889, + 1842, + 865, + 1842 + ], + "score": 0.81, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 552, + 1095, + 574, + 1095, + 574, + 1121, + 552, + 1121 + ], + "score": 0.8, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 684, + 236, + 703, + 236, + 703, + 263, + 684, + 263 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1167, + 832, + 1186, + 832, + 1186, + 854, + 1167, + 854 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 412, + 833, + 431, + 833, + 431, + 854, + 412, + 854 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 989, + 891, + 1008, + 891, + 1008, + 913, + 989, + 913 + ], + "score": 0.72, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1042, + 234, + 1056, + 234, + 1056, + 258, + 1042, + 258 + ], + "score": 0.72, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 391, + 1390, + 578, + 1390, + 578, + 1425, + 391, + 1425 + ], + "score": 0.71, + "latex": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }" + }, + { + "category_id": 14, + "poly": [ + 384, + 1432, + 1021, + 1432, + 1021, + 1578, + 384, + 1578 + ], + "score": 0.7, + "latex": "\\begin{array} { r l } & { v _ { t } = v _ { t - 1 } - ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\mathrm { s i g n } ( v _ { t - 1 } - \\Delta _ { t } ^ { 2 } ) \\ ( \\mathrm { F E D Y O G I } ) } \\\\ & { v _ { t } = \\beta _ { 2 } v _ { t - 1 } + ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\left( \\mathrm { F E D A D A M } \\right) } \\\\ & { x _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 385, + 1122, + 495, + 1122, + 495, + 1160, + 385, + 1160 + ], + "score": 0.5, + "latex": "x _ { i , 0 } ^ { t } \\dot { = } x _ { t }" + }, + { + "category_id": 14, + "poly": [ + 385, + 1122, + 495, + 1122, + 495, + 1160, + 385, + 1160 + ], + "score": 0.47, + "latex": "x _ { i , 0 } ^ { t } \\dot { = } x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 649, + 1777, + 667, + 1777, + 667, + 1804, + 649, + 1804 + ], + "score": 0.28, + "latex": "I" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 557.0, + 906.0, + 557.0, + 906.0, + 602.0, + 291.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 984.0, + 636.0, + 984.0, + 636.0, + 1026.0, + 293.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 980.0, + 981.0, + 980.0, + 981.0, + 1029.0, + 646.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 837.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 632.0, + 1403.0, + 632.0, + 1403.0, + 664.0, + 294.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 660.0, + 1407.0, + 660.0, + 1407.0, + 695.0, + 293.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 692.0, + 1407.0, + 692.0, + 1407.0, + 724.0, + 294.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 716.0, + 1309.0, + 716.0, + 1309.0, + 756.0, + 292.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 765.0, + 1405.0, + 765.0, + 1405.0, + 801.0, + 295.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 797.0, + 1405.0, + 797.0, + 1405.0, + 832.0, + 295.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 827.0, + 411.0, + 827.0, + 411.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 827.0, + 1166.0, + 827.0, + 1166.0, + 861.0, + 432.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 827.0, + 1406.0, + 827.0, + 1406.0, + 861.0, + 1187.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 857.0, + 1402.0, + 857.0, + 1402.0, + 889.0, + 296.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 886.0, + 679.0, + 886.0, + 679.0, + 921.0, + 295.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 886.0, + 988.0, + 886.0, + 988.0, + 921.0, + 706.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 886.0, + 1405.0, + 886.0, + 1405.0, + 921.0, + 1009.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 916.0, + 430.0, + 916.0, + 430.0, + 952.0, + 295.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 916.0, + 896.0, + 916.0, + 896.0, + 952.0, + 458.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 336.0, + 1405.0, + 336.0, + 1405.0, + 371.0, + 295.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 367.0, + 741.0, + 367.0, + 741.0, + 402.0, + 295.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 367.0, + 1405.0, + 367.0, + 1405.0, + 402.0, + 778.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 396.0, + 1405.0, + 396.0, + 1405.0, + 430.0, + 293.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 424.0, + 1405.0, + 424.0, + 1405.0, + 462.0, + 293.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 458.0, + 1407.0, + 458.0, + 1407.0, + 490.0, + 295.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 485.0, + 1182.0, + 485.0, + 1182.0, + 519.0, + 294.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1616.0, + 1404.0, + 1616.0, + 1404.0, + 1649.0, + 296.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1645.0, + 494.0, + 1645.0, + 494.0, + 1682.0, + 294.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1645.0, + 1010.0, + 1645.0, + 1010.0, + 1682.0, + 594.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1645.0, + 1406.0, + 1645.0, + 1406.0, + 1682.0, + 1094.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1668.0, + 520.0, + 1668.0, + 520.0, + 1716.0, + 292.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1668.0, + 1298.0, + 1668.0, + 1298.0, + 1716.0, + 604.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1668.0, + 1411.0, + 1668.0, + 1411.0, + 1716.0, + 1398.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1704.0, + 1406.0, + 1704.0, + 1406.0, + 1741.0, + 295.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1732.0, + 1281.0, + 1732.0, + 1281.0, + 1772.0, + 294.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 226.0, + 683.0, + 226.0, + 683.0, + 269.0, + 291.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 226.0, + 751.0, + 226.0, + 751.0, + 269.0, + 704.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 226.0, + 1041.0, + 226.0, + 1041.0, + 269.0, + 779.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 226.0, + 1407.0, + 226.0, + 1407.0, + 269.0, + 1057.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1352.0, + 290.0, + 1352.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1769.0, + 648.0, + 1769.0, + 648.0, + 1812.0, + 291.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1769.0, + 886.0, + 1769.0, + 886.0, + 1812.0, + 668.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 1769.0, + 1333.0, + 1769.0, + 1333.0, + 1812.0, + 1022.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1804.0, + 864.0, + 1804.0, + 864.0, + 1849.0, + 432.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1804.0, + 905.0, + 1804.0, + 905.0, + 1849.0, + 890.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1026.0, + 503.0, + 1026.0, + 503.0, + 1071.0, + 307.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1026.0, + 861.0, + 1026.0, + 861.0, + 1071.0, + 652.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1026.0, + 1030.0, + 1026.0, + 1030.0, + 1071.0, + 1027.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1060.0, + 385.0, + 1060.0, + 385.0, + 1094.0, + 305.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1060.0, + 626.0, + 1060.0, + 626.0, + 1094.0, + 584.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1092.0, + 339.0, + 1092.0, + 339.0, + 1126.0, + 307.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1093.0, + 551.0, + 1093.0, + 551.0, + 1125.0, + 383.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1093.0, + 687.0, + 1093.0, + 687.0, + 1125.0, + 575.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1114.0, + 384.0, + 1114.0, + 384.0, + 1170.0, + 381.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1114.0, + 507.0, + 1114.0, + 507.0, + 1170.0, + 496.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1121.0, + 339.0, + 1121.0, + 339.0, + 1157.0, + 307.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1155.0, + 337.0, + 1155.0, + 337.0, + 1188.0, + 308.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1155.0, + 553.0, + 1155.0, + 553.0, + 1188.0, + 383.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1155.0, + 785.0, + 1155.0, + 785.0, + 1188.0, + 620.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1185.0, + 338.0, + 1185.0, + 338.0, + 1217.0, + 308.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 1184.0, + 468.0, + 1184.0, + 468.0, + 1216.0, + 425.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1184.0, + 717.0, + 1184.0, + 717.0, + 1216.0, + 678.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1214.0, + 339.0, + 1214.0, + 339.0, + 1251.0, + 307.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1206.0, + 813.0, + 1206.0, + 813.0, + 1256.0, + 463.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1206.0, + 893.0, + 1206.0, + 893.0, + 1256.0, + 860.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1206.0, + 1015.0, + 1206.0, + 1015.0, + 1256.0, + 1012.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1252.0, + 339.0, + 1252.0, + 339.0, + 1288.0, + 307.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1233.0, + 733.0, + 1233.0, + 733.0, + 1302.0, + 456.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 1291.0, + 340.0, + 1291.0, + 340.0, + 1332.0, + 304.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1337.0, + 341.0, + 1337.0, + 341.0, + 1378.0, + 296.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1325.0, + 381.0, + 1325.0, + 381.0, + 1386.0, + 378.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1389.0, + 341.0, + 1389.0, + 341.0, + 1430.0, + 296.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 1387.0, + 782.0, + 1387.0, + 782.0, + 1427.0, + 579.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1437.0, + 340.0, + 1437.0, + 340.0, + 1478.0, + 296.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1486.0, + 341.0, + 1486.0, + 341.0, + 1525.0, + 295.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1528.0, + 341.0, + 1528.0, + 341.0, + 1567.0, + 295.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1525.5, + 633.0, + 1525.5, + 633.0, + 1577.5, + 378.0, + 1577.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1964.0, + 460.0, + 1964.0, + 460.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1879.0, + 451.0, + 1879.0, + 451.0, + 1920.0, + 295.0, + 1920.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1806, + 1405, + 1806, + 1405, + 1905, + 299, + 1905 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 543, + 1558, + 1153, + 1558, + 1153, + 1719, + 543, + 1719 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 297, + 846, + 1405, + 846, + 1405, + 946, + 297, + 946 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 299, + 740, + 1403, + 740, + 1403, + 840, + 299, + 840 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 523, + 470, + 1173, + 470, + 1173, + 722, + 523, + 722 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 534, + 1298, + 1158, + 1298, + 1158, + 1391, + 534, + 1391 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 295, + 1736, + 1404, + 1736, + 1404, + 1800, + 295, + 1800 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 295, + 1216, + 1403, + 1216, + 1403, + 1286, + 295, + 1286 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 297, + 1146, + 1401, + 1146, + 1401, + 1210, + 297, + 1210 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 509, + 1449, + 1191, + 1449, + 1191, + 1527, + 509, + 1527 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 353, + 960, + 1343, + 960, + 1343, + 1050, + 353, + 1050 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 291, + 1099, + 1338, + 1099, + 1338, + 1134, + 291, + 1134 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 300, + 1402, + 904, + 1402, + 904, + 1436, + 300, + 1436 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 296, + 1540, + 370, + 1540, + 370, + 1571, + 296, + 1571 + ], + "score": 0.92 + }, + { + "category_id": 8, + "poly": [ + 355, + 1919, + 1345, + 1919, + 1345, + 2008, + 355, + 2008 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 295, + 228, + 952, + 228, + 952, + 264, + 295, + 264 + ], + "score": 0.872 + }, + { + "category_id": 1, + "poly": [ + 297, + 441, + 473, + 441, + 473, + 473, + 297, + 473 + ], + "score": 0.819 + }, + { + "category_id": 2, + "poly": [ + 840, + 2087, + 859, + 2087, + 859, + 2113, + 840, + 2113 + ], + "score": 0.768 + }, + { + "category_id": 1, + "poly": [ + 288, + 276, + 1356, + 276, + 1356, + 431, + 288, + 431 + ], + "score": 0.453 + }, + { + "category_id": 8, + "poly": [ + 295, + 276, + 1356, + 276, + 1356, + 431, + 295, + 431 + ], + "score": 0.143 + }, + { + "category_id": 8, + "poly": [ + 307, + 276, + 1352, + 276, + 1352, + 350, + 307, + 350 + ], + "score": 0.091 + }, + { + "category_id": 14, + "poly": [ + 520, + 467, + 1180, + 467, + 1180, + 724, + 520, + 724 + ], + "score": 0.93, + "latex": "\\begin{array} { c } { { \\Psi = \\displaystyle \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\sigma ^ { 2 } , } } \\\\ { { \\Psi _ { \\mathrm { v a r } } = \\displaystyle \\frac { d ( \\eta _ { l } K G ^ { 2 } + \\tau \\eta L ) } { \\tau } \\Bigg [ 1 + \\log { \\displaystyle \\frac { \\tau ^ { 2 } + \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } } \\Bigg ] , } } \\\\ { { \\widetilde { \\Psi } _ { \\mathrm { v a r } } = \\displaystyle \\frac { 2 \\eta _ { l } K G ^ { 2 } + \\tau \\eta L } { \\tau ^ { 2 } } \\left[ \\displaystyle \\frac { 2 \\eta _ { l } ^ { 2 } K T } { m } \\sigma _ { l } ^ { 2 } + 1 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\sigma ^ { 2 } \\right] . } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 285, + 274, + 1367, + 274, + 1367, + 437, + 285, + 437 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r l } & { u n d e r ~ C o n d i t i o n ~ I o n l y , } & & { \\underset { 0 \\leq t \\leq T - 1 } { \\operatorname* { m i n } } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } \\leq \\odot \\left( \\left[ \\frac { G } { \\sqrt { T } } + \\frac { \\tau } { \\eta _ { t } K T } \\right] ( \\Psi + \\Psi _ { v a r } ) \\right) , } \\\\ & { u n d e r ~ b o t h ~ C o n d i t i o n ~ I \\ll { I I } , } & & { \\underset { 0 \\leq t \\leq T - 1 } { \\operatorname* { m i n } } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } \\leq \\odot \\left( \\left[ \\frac { G } { \\sqrt { T } } + \\frac { \\tau } { \\eta _ { t } K T } \\right] \\left( \\Psi + \\widetilde { \\Psi } _ { \\mathrm { v a r } } \\right) \\right) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 353, + 1915, + 1344, + 1915, + 1344, + 2009, + 353, + 2009 + ], + "score": 0.93, + "latex": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\mathcal { O } \\Bigg ( \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\sqrt { m K T } } + \\frac { 2 \\sigma _ { l } ^ { 2 } L } { G ^ { 2 } \\sqrt { m K T } } + \\frac { \\sigma ^ { 2 } } { G K T } + \\frac { \\sigma ^ { 2 } L \\sqrt { m } } { G ^ { 2 } \\sqrt { K } T ^ { 3 / 2 } } \\Bigg ) ." + }, + { + "category_id": 14, + "poly": [ + 352, + 956, + 1344, + 956, + 1344, + 1051, + 352, + 1051 + ], + "score": 0.93, + "latex": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\mathcal { O } \\Bigg ( \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\sqrt { m K T } } + \\frac { 2 \\sigma _ { l } ^ { 2 } L } { G ^ { 2 } \\sqrt { m K T } } + \\frac { \\sigma ^ { 2 } } { G K T } + \\frac { \\sigma ^ { 2 } L \\sqrt { m } } { G ^ { 2 } \\sqrt { K } T ^ { 3 / 2 } } \\Bigg ) ." + }, + { + "category_id": 14, + "poly": [ + 536, + 1298, + 1160, + 1298, + 1160, + 1391, + 536, + 1391 + ], + "score": 0.93, + "latex": "\\eta _ { l } \\leq \\frac { 1 } { 6 K } \\operatorname* { m i n } \\left\\{ \\left[ \\frac { \\tau } { G L } \\right] ^ { 1 / 2 } , \\left[ \\frac { \\tau ^ { 2 } } { G L ^ { 3 } \\eta } \\right] ^ { 1 / 4 } , \\left[ \\frac { \\tau } { G L ^ { 2 } } \\right] ^ { 1 / 3 } \\right\\} ." + }, + { + "category_id": 13, + "poly": [ + 687, + 1838, + 857, + 1838, + 857, + 1875, + 687, + 1875 + ], + "score": 0.92, + "latex": "\\eta = \\Theta ( \\sqrt { K m } )" + }, + { + "category_id": 13, + "poly": [ + 447, + 880, + 616, + 880, + 616, + 916, + 447, + 916 + ], + "score": 0.92, + "latex": "\\eta = \\Theta ( \\sqrt { K m } )" + }, + { + "category_id": 13, + "poly": [ + 1207, + 1214, + 1399, + 1214, + 1399, + 1254, + 1207, + 1254 + ], + "score": 0.92, + "latex": "\\sigma ^ { 2 } = \\sigma _ { l } ^ { 2 } + 6 K \\sigma _ { g } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 846, + 1398, + 846, + 1398, + 881, + 1196, + 881 + ], + "score": 0.92, + "latex": "\\eta _ { l } = \\Theta \\big ( { 1 } / { ( K L \\sqrt { T } } \\big )" + }, + { + "category_id": 13, + "poly": [ + 670, + 881, + 765, + 881, + 765, + 916, + 670, + 916 + ], + "score": 0.91, + "latex": "\\tau = G / _ { L }" + }, + { + "category_id": 13, + "poly": [ + 1090, + 771, + 1285, + 771, + 1285, + 810, + 1090, + 810 + ], + "score": 0.91, + "latex": "\\operatorname* { m i n } \\{ \\Psi _ { \\mathrm { v a r } } , \\widetilde { \\Psi } _ { \\mathrm { v a r } } \\}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1841, + 509, + 1841, + 509, + 1875, + 298, + 1875 + ], + "score": 0.91, + "latex": "\\eta _ { l } = \\Theta \\bigl ( 1 / ( K L \\sqrt { T } ) \\bigr )" + }, + { + "category_id": 13, + "poly": [ + 798, + 1219, + 931, + 1219, + 931, + 1253, + 798, + 1253 + ], + "score": 0.9, + "latex": "L , G , \\sigma _ { l } , \\sigma _ { g }" + }, + { + "category_id": 13, + "poly": [ + 913, + 1840, + 1010, + 1840, + 1010, + 1875, + 913, + 1875 + ], + "score": 0.9, + "latex": "\\tau = G / _ { L }" + }, + { + "category_id": 13, + "poly": [ + 752, + 1253, + 891, + 1253, + 891, + 1286, + 752, + 1286 + ], + "score": 0.9, + "latex": "\\eta _ { l } \\leq 1 / 1 6 L K" + }, + { + "category_id": 14, + "poly": [ + 510, + 1449, + 1190, + 1449, + 1190, + 1527, + 510, + 1527 + ], + "score": 0.89, + "latex": "\\operatorname* { m i n } _ { 0 \\leq t \\leq T - 1 } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } = \\odot \\left( \\frac { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } { \\eta _ { l } K T } \\left( \\Psi + \\Psi _ { \\mathrm { v a r } } \\right) \\right) ," + }, + { + "category_id": 13, + "poly": [ + 1278, + 813, + 1329, + 813, + 1329, + 840, + 1278, + 840 + ], + "score": 0.87, + "latex": "\\eta , \\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 707, + 1769, + 735, + 1769, + 735, + 1795, + 707, + 1795 + ], + "score": 0.85, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 786, + 1769, + 808, + 1769, + 808, + 1795, + 786, + 1795 + ], + "score": 0.84, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 590, + 809, + 613, + 809, + 613, + 835, + 590, + 835 + ], + "score": 0.83, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1067, + 747, + 1092, + 747, + 1092, + 774, + 1067, + 774 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 664, + 809, + 692, + 809, + 692, + 835, + 664, + 835 + ], + "score": 0.83, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1250, + 1743, + 1304, + 1743, + 1304, + 1770, + 1250, + 1770 + ], + "score": 0.78, + "latex": "\\eta _ { l } , \\eta" + }, + { + "category_id": 13, + "poly": [ + 1354, + 1744, + 1372, + 1744, + 1372, + 1765, + 1354, + 1765 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 549, + 853, + 572, + 853, + 572, + 879, + 549, + 879 + ], + "score": 0.77, + "latex": "\\eta _ { l }" + }, + { + "category_id": 14, + "poly": [ + 540, + 1561, + 1159, + 1561, + 1159, + 1721, + 540, + 1721 + ], + "score": 0.75, + "latex": "\\begin{array} { c } { { \\Psi = \\displaystyle \\frac { f ( x _ { 0 } ) - f ( x ^ { * } ) } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\sigma ^ { 2 } , } } \\\\ { { \\Psi _ { \\nu a r } = \\left( G + \\displaystyle \\frac { \\eta L } { 2 } \\right) \\left[ \\displaystyle \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sigma _ { l } ^ { 2 } + \\displaystyle \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\sigma ^ { 2 } \\right] . } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1380, + 814, + 1398, + 814, + 1398, + 835, + 1380, + 835 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 560, + 1813, + 586, + 1813, + 586, + 1840, + 560, + 1840 + ], + "score": 0.75, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1329, + 1843, + 1351, + 1843, + 1351, + 1869, + 1329, + 1869 + ], + "score": 0.73, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1074, + 884, + 1097, + 884, + 1097, + 910, + 1074, + 910 + ], + "score": 0.71, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 993, + 850, + 1008, + 850, + 1008, + 875, + 993, + 875 + ], + "score": 0.34, + "latex": "I" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2119.0, + 839.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1806.0, + 559.0, + 1806.0, + 559.0, + 1844.0, + 297.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1806.0, + 1406.0, + 1806.0, + 1406.0, + 1844.0, + 587.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1839.0, + 297.0, + 1839.0, + 297.0, + 1878.0, + 294.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1839.0, + 686.0, + 1839.0, + 686.0, + 1878.0, + 510.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1839.0, + 912.0, + 1839.0, + 912.0, + 1878.0, + 858.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1839.0, + 1328.0, + 1839.0, + 1328.0, + 1878.0, + 1011.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1839.0, + 1405.0, + 1839.0, + 1405.0, + 1878.0, + 1352.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1870.0, + 803.0, + 1870.0, + 803.0, + 1908.0, + 296.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 845.0, + 548.0, + 845.0, + 548.0, + 884.0, + 296.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 845.0, + 992.0, + 845.0, + 992.0, + 884.0, + 573.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 845.0, + 1195.0, + 845.0, + 1195.0, + 884.0, + 1009.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 845.0, + 1408.0, + 845.0, + 1408.0, + 884.0, + 1399.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 880.0, + 446.0, + 880.0, + 446.0, + 919.0, + 291.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 880.0, + 669.0, + 880.0, + 669.0, + 919.0, + 617.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 880.0, + 1073.0, + 880.0, + 1073.0, + 919.0, + 766.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 880.0, + 1406.0, + 880.0, + 1406.0, + 919.0, + 1098.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 910.0, + 592.0, + 910.0, + 592.0, + 949.0, + 291.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 741.0, + 1066.0, + 741.0, + 1066.0, + 776.0, + 295.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 741.0, + 1403.0, + 741.0, + 1403.0, + 776.0, + 1093.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 775.0, + 1089.0, + 775.0, + 1089.0, + 814.0, + 294.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 775.0, + 1404.0, + 775.0, + 1404.0, + 814.0, + 1286.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 809.0, + 589.0, + 809.0, + 589.0, + 841.0, + 294.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 809.0, + 663.0, + 809.0, + 663.0, + 841.0, + 614.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 809.0, + 1277.0, + 809.0, + 1277.0, + 841.0, + 693.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 809.0, + 1379.0, + 809.0, + 1379.0, + 841.0, + 1330.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 809.0, + 1408.0, + 809.0, + 1408.0, + 841.0, + 1399.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1736.0, + 1249.0, + 1736.0, + 1249.0, + 1772.0, + 295.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1736.0, + 1353.0, + 1736.0, + 1353.0, + 1772.0, + 1305.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1736.0, + 1404.0, + 1736.0, + 1404.0, + 1772.0, + 1373.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1767.0, + 706.0, + 1767.0, + 706.0, + 1801.0, + 294.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1767.0, + 785.0, + 1767.0, + 785.0, + 1801.0, + 736.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1767.0, + 818.0, + 1767.0, + 818.0, + 1801.0, + 809.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1207.0, + 797.0, + 1207.0, + 797.0, + 1266.0, + 289.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1207.0, + 1206.0, + 1207.0, + 1206.0, + 1266.0, + 932.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 1207.0, + 1411.0, + 1207.0, + 1411.0, + 1266.0, + 1400.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 751.0, + 1251.0, + 751.0, + 1288.0, + 295.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1251.0, + 946.0, + 1251.0, + 946.0, + 1288.0, + 892.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1145.0, + 1405.0, + 1145.0, + 1405.0, + 1181.0, + 296.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1175.0, + 1193.0, + 1175.0, + 1193.0, + 1213.0, + 293.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1100.0, + 1339.0, + 1100.0, + 1339.0, + 1135.0, + 295.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1399.0, + 905.0, + 1399.0, + 905.0, + 1441.0, + 293.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1534.0, + 376.0, + 1534.0, + 376.0, + 1575.0, + 293.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 951.0, + 225.0, + 951.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 438.0, + 474.0, + 438.0, + 474.0, + 478.0, + 294.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 330.0, + 1378.0, + 330.0, + 1378.0, + 455.0, + 1368.0, + 455.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1687, + 1404, + 1687, + 1404, + 1928, + 297, + 1928 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1522, + 1403, + 1522, + 1403, + 1673, + 298, + 1673 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 300, + 1943, + 1403, + 1943, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1403, + 229, + 1403, + 381, + 299, + 381 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 299, + 403, + 1402, + 403, + 1402, + 495, + 299, + 495 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 298, + 1321, + 1405, + 1321, + 1405, + 1413, + 298, + 1413 + ], + "score": 0.959 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 300, + 509, + 1404, + 509, + 1404, + 749, + 300, + 749 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 292, + 1126, + 1404, + 1126, + 1404, + 1307, + 292, + 1307 + ], + "score": 0.781 + }, + { + "category_id": 1, + "poly": [ + 296, + 753, + 1404, + 753, + 1404, + 963, + 296, + 963 + ], + "score": 0.777 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.768 + }, + { + "category_id": 1, + "poly": [ + 292, + 1454, + 1283, + 1454, + 1283, + 1490, + 292, + 1490 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 290, + 967, + 1403, + 967, + 1403, + 1123, + 290, + 1123 + ], + "score": 0.712 + }, + { + "category_id": 0, + "poly": [ + 292, + 1454, + 1283, + 1454, + 1283, + 1490, + 292, + 1490 + ], + "score": 0.216 + }, + { + "category_id": 13, + "poly": [ + 891, + 1059, + 966, + 1059, + 966, + 1099, + 891, + 1099 + ], + "score": 0.94, + "latex": "\\sigma _ { l } ^ { 2 } / \\sigma _ { g } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 657, + 1381, + 731, + 1381, + 731, + 1415, + 657, + 1415 + ], + "score": 0.92, + "latex": "| { \\cal S } | / m" + }, + { + "category_id": 13, + "poly": [ + 955, + 1024, + 1157, + 1024, + 1157, + 1063, + 955, + 1063 + ], + "score": 0.92, + "latex": "K = \\mathcal { O } ( T \\bar { \\sigma } _ { l } ^ { 2 } / \\sigma _ { g } ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 994, + 753, + 1053, + 753, + 1053, + 786, + 994, + 786 + ], + "score": 0.92, + "latex": "^ 1 / \\sqrt { T }" + }, + { + "category_id": 13, + "poly": [ + 766, + 631, + 848, + 631, + 848, + 663, + 766, + 663 + ], + "score": 0.92, + "latex": "\\sigma _ { g } = 0" + }, + { + "category_id": 13, + "poly": [ + 342, + 1093, + 424, + 1093, + 424, + 1126, + 342, + 1126 + ], + "score": 0.91, + "latex": "\\sigma _ { g } = 0" + }, + { + "category_id": 13, + "poly": [ + 560, + 873, + 660, + 873, + 660, + 901, + 560, + 901 + ], + "score": 0.91, + "latex": "T \\to \\infty" + }, + { + "category_id": 13, + "poly": [ + 708, + 232, + 777, + 232, + 777, + 262, + 708, + 262 + ], + "score": 0.91, + "latex": "\\eta = 1" + }, + { + "category_id": 13, + "poly": [ + 343, + 569, + 479, + 569, + 479, + 603, + 343, + 603 + ], + "score": 0.91, + "latex": "\\mathcal { O } ( { } ^ { 1 } / \\sqrt { m K T } )" + }, + { + "category_id": 13, + "poly": [ + 1267, + 509, + 1403, + 509, + 1403, + 543, + 1267, + 543 + ], + "score": 0.9, + "latex": "\\mathcal { O } ( { } ^ { 1 } / { \\sqrt { m K T } } )" + }, + { + "category_id": 13, + "poly": [ + 706, + 813, + 756, + 813, + 756, + 845, + 706, + 845 + ], + "score": 0.9, + "latex": "^ { 1 } / \\sqrt { t }" + }, + { + "category_id": 13, + "poly": [ + 1288, + 1321, + 1395, + 1321, + 1395, + 1355, + 1288, + 1355 + ], + "score": 0.86, + "latex": "\\mathcal { S } = [ m ] )" + }, + { + "category_id": 13, + "poly": [ + 824, + 264, + 856, + 264, + 856, + 295, + 824, + 295 + ], + "score": 0.84, + "latex": "\\cdot _ { g }" + }, + { + "category_id": 13, + "poly": [ + 417, + 817, + 443, + 817, + 443, + 845, + 417, + 845 + ], + "score": 0.84, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 733, + 877, + 760, + 877, + 760, + 905, + 733, + 905 + ], + "score": 0.82, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1306, + 907, + 1332, + 907, + 1332, + 934, + 1306, + 934 + ], + "score": 0.82, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1104, + 818, + 1129, + 818, + 1129, + 845, + 1104, + 845 + ], + "score": 0.82, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 670, + 784, + 694, + 784, + 694, + 810, + 670, + 810 + ], + "score": 0.8, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1000, + 1062, + 1023, + 1062, + 1023, + 1089, + 1000, + 1089 + ], + "score": 0.79, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 392, + 848, + 410, + 848, + 410, + 875, + 392, + 875 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 825, + 511, + 850, + 511, + 850, + 537, + 825, + 537 + ], + "score": 0.78, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 342, + 1026, + 371, + 1026, + 371, + 1053, + 342, + 1053 + ], + "score": 0.77, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1383, + 907, + 1401, + 907, + 1401, + 934, + 1383, + 934 + ], + "score": 0.76, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 606, + 1093, + 636, + 1093, + 636, + 1120, + 606, + 1120 + ], + "score": 0.75, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 883, + 997, + 906, + 997, + 906, + 1023, + 883, + 1023 + ], + "score": 0.74, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 736, + 296, + 761, + 296, + 761, + 323, + 736, + 323 + ], + "score": 0.65, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1225, + 510, + 1253, + 510, + 1253, + 538, + 1225, + 538 + ], + "score": 0.61, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1193, + 1062, + 1246, + 1062, + 1246, + 1090, + 1193, + 1090 + ], + "score": 0.37, + "latex": "i . i . d" + }, + { + "category_id": 13, + "poly": [ + 523, + 753, + 549, + 753, + 549, + 781, + 523, + 781 + ], + "score": 0.35, + "latex": "\\pmb { \\& }" + }, + { + "category_id": 13, + "poly": [ + 664, + 966, + 691, + 966, + 691, + 995, + 664, + 995 + ], + "score": 0.33, + "latex": "\\pmb { \\& }" + }, + { + "category_id": 13, + "poly": [ + 1026, + 297, + 1045, + 297, + 1045, + 323, + 1026, + 323 + ], + "score": 0.28, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 741, + 540, + 813, + 540, + 813, + 569, + 741, + 569 + ], + "score": 0.28, + "latex": "1 \\ \\& \\ 2" + }, + { + "category_id": 13, + "poly": [ + 1028, + 600, + 1081, + 600, + 1081, + 628, + 1028, + 628 + ], + "score": 0.28, + "latex": "i . i . d" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1452.0, + 1288.0, + 1452.0, + 1288.0, + 1493.0, + 292.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1684.0, + 1409.0, + 1684.0, + 1409.0, + 1721.0, + 295.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1714.0, + 1407.0, + 1714.0, + 1407.0, + 1752.0, + 295.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1745.0, + 1406.0, + 1745.0, + 1406.0, + 1782.0, + 294.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1406.0, + 1773.0, + 1406.0, + 1813.0, + 294.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1805.0, + 1405.0, + 1805.0, + 1405.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1834.0, + 1409.0, + 1834.0, + 1409.0, + 1873.0, + 292.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1865.0, + 1406.0, + 1865.0, + 1406.0, + 1901.0, + 292.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1894.0, + 1333.0, + 1894.0, + 1333.0, + 1930.0, + 295.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1523.0, + 1405.0, + 1523.0, + 1405.0, + 1556.0, + 297.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1549.0, + 1405.0, + 1549.0, + 1405.0, + 1588.0, + 293.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1581.0, + 1407.0, + 1581.0, + 1407.0, + 1617.0, + 293.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1611.0, + 1407.0, + 1611.0, + 1407.0, + 1646.0, + 293.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1639.0, + 1263.0, + 1639.0, + 1263.0, + 1676.0, + 293.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1975.0, + 296.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1407.0, + 2003.0, + 1407.0, + 2035.0, + 295.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 707.0, + 226.0, + 707.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 268.0, + 778.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 256.0, + 823.0, + 256.0, + 823.0, + 300.0, + 294.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 256.0, + 1405.0, + 256.0, + 1405.0, + 300.0, + 857.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 735.0, + 291.0, + 735.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 291.0, + 1025.0, + 291.0, + 1025.0, + 327.0, + 762.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 327.0, + 1046.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1408.0, + 319.0, + 1408.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 348.0, + 1280.0, + 348.0, + 1280.0, + 386.0, + 293.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 403.0, + 1404.0, + 403.0, + 1404.0, + 437.0, + 297.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 435.0, + 1404.0, + 435.0, + 1404.0, + 468.0, + 294.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 461.0, + 1280.0, + 461.0, + 1280.0, + 502.0, + 293.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1318.0, + 1287.0, + 1318.0, + 1287.0, + 1357.0, + 292.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1318.0, + 1409.0, + 1318.0, + 1409.0, + 1357.0, + 1396.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1354.0, + 1406.0, + 1354.0, + 1406.0, + 1384.0, + 296.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1381.0, + 656.0, + 1381.0, + 656.0, + 1415.0, + 293.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1381.0, + 1408.0, + 1381.0, + 1408.0, + 1415.0, + 732.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 508.0, + 824.0, + 508.0, + 824.0, + 545.0, + 299.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 508.0, + 1224.0, + 508.0, + 1224.0, + 545.0, + 851.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 508.0, + 1266.0, + 508.0, + 1266.0, + 545.0, + 1254.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 536.0, + 740.0, + 536.0, + 740.0, + 575.0, + 335.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 536.0, + 1406.0, + 536.0, + 1406.0, + 575.0, + 814.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 569.0, + 1406.0, + 569.0, + 1406.0, + 602.0, + 480.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 596.0, + 1027.0, + 596.0, + 1027.0, + 637.0, + 337.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 596.0, + 1408.0, + 596.0, + 1408.0, + 637.0, + 1082.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 624.0, + 765.0, + 624.0, + 765.0, + 666.0, + 337.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 624.0, + 1406.0, + 624.0, + 1406.0, + 666.0, + 849.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 657.0, + 1405.0, + 657.0, + 1405.0, + 694.0, + 338.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 689.0, + 1404.0, + 689.0, + 1404.0, + 722.0, + 340.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 717.0, + 898.0, + 717.0, + 898.0, + 754.0, + 338.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1125.0, + 1405.0, + 1125.0, + 1405.0, + 1160.0, + 286.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1155.0, + 1403.0, + 1155.0, + 1403.0, + 1190.0, + 339.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1185.0, + 1409.0, + 1185.0, + 1409.0, + 1220.0, + 339.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 1211.0, + 1405.0, + 1211.0, + 1405.0, + 1253.0, + 337.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1245.0, + 1405.0, + 1245.0, + 1405.0, + 1280.0, + 339.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1277.0, + 1328.0, + 1277.0, + 1328.0, + 1308.0, + 339.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 753.0, + 522.0, + 753.0, + 522.0, + 787.0, + 293.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 753.0, + 993.0, + 753.0, + 993.0, + 787.0, + 550.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 753.0, + 1406.0, + 753.0, + 1406.0, + 787.0, + 1054.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 782.0, + 669.0, + 782.0, + 669.0, + 819.0, + 339.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 782.0, + 1407.0, + 782.0, + 1407.0, + 819.0, + 695.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 810.0, + 416.0, + 810.0, + 416.0, + 851.0, + 337.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 810.0, + 705.0, + 810.0, + 705.0, + 851.0, + 444.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 810.0, + 1103.0, + 810.0, + 1103.0, + 851.0, + 757.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 810.0, + 1405.0, + 810.0, + 1405.0, + 851.0, + 1130.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 843.0, + 391.0, + 843.0, + 391.0, + 877.0, + 337.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 843.0, + 1405.0, + 843.0, + 1405.0, + 877.0, + 411.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 869.0, + 559.0, + 869.0, + 559.0, + 909.0, + 337.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 869.0, + 732.0, + 869.0, + 732.0, + 909.0, + 661.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 869.0, + 1406.0, + 869.0, + 1406.0, + 909.0, + 761.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 899.0, + 1305.0, + 899.0, + 1305.0, + 939.0, + 337.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 899.0, + 1382.0, + 899.0, + 1382.0, + 939.0, + 1333.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 899.0, + 1406.0, + 899.0, + 1406.0, + 939.0, + 1402.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 932.0, + 1274.0, + 932.0, + 1274.0, + 966.0, + 338.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1452.0, + 1288.0, + 1452.0, + 1288.0, + 1493.0, + 292.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 966.0, + 663.0, + 966.0, + 663.0, + 1000.0, + 288.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 966.0, + 1404.0, + 966.0, + 1404.0, + 1000.0, + 692.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 993.0, + 882.0, + 993.0, + 882.0, + 1032.0, + 337.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 993.0, + 1406.0, + 993.0, + 1406.0, + 1032.0, + 907.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1024.0, + 954.0, + 1024.0, + 954.0, + 1061.0, + 372.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1024.0, + 1406.0, + 1024.0, + 1406.0, + 1061.0, + 1158.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1060.0, + 890.0, + 1060.0, + 890.0, + 1097.0, + 339.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1060.0, + 999.0, + 1060.0, + 999.0, + 1097.0, + 967.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1060.0, + 1192.0, + 1060.0, + 1192.0, + 1097.0, + 1024.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1060.0, + 1405.0, + 1060.0, + 1405.0, + 1097.0, + 1247.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 1092.0, + 341.0, + 1092.0, + 341.0, + 1129.0, + 337.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 1092.0, + 605.0, + 1092.0, + 605.0, + 1129.0, + 425.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1092.0, + 841.0, + 1092.0, + 841.0, + 1129.0, + 637.0, + 1129.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 802, + 1277, + 1388, + 1277, + 1388, + 1559, + 802, + 1559 + ], + "score": 0.978, + "html": "
FED..ADAGRADADAMYoGIAvGMAvG
CIFAR-1072.177.478.077.472.8
CIFAR-10047.952.552.452.444.7
EMNIST CR85.185.685.585.284.9
SHAKESPEARE57.557.057.257.356.9
SO NWP23.825.225.223.819.5
SO LR67.165.865.936.930.0
EMNIST AE4.201.010.981.656.47
" + }, + { + "category_id": 1, + "poly": [ + 297, + 396, + 1404, + 396, + 1404, + 607, + 297, + 607 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 621, + 1403, + 621, + 1403, + 802, + 298, + 802 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 818, + 1405, + 818, + 1405, + 968, + 299, + 968 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1272, + 772, + 1272, + 772, + 1747, + 298, + 1747 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1852, + 1404, + 1852, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1404, + 229, + 1404, + 380, + 298, + 380 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1134, + 1403, + 1134, + 1403, + 1254, + 299, + 1254 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 299, + 1749, + 1398, + 1749, + 1398, + 1839, + 299, + 1839 + ], + "score": 0.96 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 303, + 1078, + 786, + 1078, + 786, + 1108, + 303, + 1108 + ], + "score": 0.875 + }, + { + "category_id": 0, + "poly": [ + 300, + 1011, + 918, + 1011, + 918, + 1046, + 300, + 1046 + ], + "score": 0.791 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2111, + 842, + 2111 + ], + "score": 0.73 + }, + { + "category_id": 7, + "poly": [ + 791, + 1571, + 1403, + 1571, + 1403, + 1721, + 791, + 1721 + ], + "score": 0.72 + }, + { + "category_id": 6, + "poly": [ + 791, + 1571, + 1403, + 1571, + 1403, + 1721, + 791, + 1721 + ], + "score": 0.4 + }, + { + "category_id": 1, + "poly": [ + 300, + 1011, + 918, + 1011, + 918, + 1046, + 300, + 1046 + ], + "score": 0.098 + }, + { + "category_id": 13, + "poly": [ + 554, + 516, + 661, + 516, + 661, + 547, + 554, + 547 + ], + "score": 0.92, + "latex": "\\beta _ { 1 } = 0 . 9" + }, + { + "category_id": 13, + "poly": [ + 1041, + 516, + 1161, + 516, + 1161, + 547, + 1041, + 547 + ], + "score": 0.91, + "latex": "\\beta _ { 2 } = 0 . 9 9" + }, + { + "category_id": 13, + "poly": [ + 897, + 577, + 973, + 577, + 973, + 604, + 897, + 604 + ], + "score": 0.9, + "latex": "E = 1" + }, + { + "category_id": 13, + "poly": [ + 1158, + 1632, + 1256, + 1632, + 1256, + 1663, + 1158, + 1663 + ], + "score": 0.89, + "latex": "( \\times 1 0 0 0 )" + }, + { + "category_id": 13, + "poly": [ + 389, + 290, + 424, + 290, + 424, + 323, + 389, + 323 + ], + "score": 0.88, + "latex": "\\Delta _ { i } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1693, + 340, + 1693, + 340, + 1722, + 298, + 1722 + ], + "score": 0.88, + "latex": "v _ { t , j }" + }, + { + "category_id": 13, + "poly": [ + 1061, + 1662, + 1123, + 1662, + 1123, + 1691, + 1061, + 1691 + ], + "score": 0.88, + "latex": "0 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1327, + 432, + 1353, + 432, + 1353, + 458, + 1327, + 458 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 923, + 1601, + 951, + 1601, + 951, + 1631, + 923, + 1631 + ], + "score": 0.82, + "latex": "\\%" + }, + { + "category_id": 13, + "poly": [ + 1322, + 1602, + 1402, + 1602, + 1402, + 1633, + 1322, + 1633 + ], + "score": 0.79, + "latex": "( \\times 1 0 0 )" + }, + { + "category_id": 13, + "poly": [ + 370, + 232, + 399, + 232, + 399, + 259, + 370, + 259 + ], + "score": 0.76, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 298, + 465, + 316, + 465, + 316, + 488, + 298, + 488 + ], + "score": 0.73, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 540, + 628, + 560, + 628, + 560, + 649, + 540, + 649 + ], + "score": 0.71, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 416, + 432, + 435, + 432, + 435, + 455, + 416, + 455 + ], + "score": 0.65, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1275, + 1602, + 1315, + 1602, + 1315, + 1630, + 1275, + 1630 + ], + "score": 0.64, + "latex": "\\textcircled { a } 5" + }, + { + "category_id": 13, + "poly": [ + 417, + 627, + 476, + 627, + 476, + 654, + 417, + 654 + ], + "score": 0.63, + "latex": "\\eta , ~ \\eta" + }, + { + "category_id": 13, + "poly": [ + 764, + 232, + 791, + 232, + 791, + 259, + 764, + 259 + ], + "score": 0.62, + "latex": "E" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1077.0, + 791.0, + 1077.0, + 791.0, + 1111.0, + 296.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1007.0, + 923.0, + 1007.0, + 923.0, + 1052.0, + 291.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1569.0, + 1405.0, + 1569.0, + 1405.0, + 1604.0, + 789.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1599.0, + 922.0, + 1599.0, + 922.0, + 1634.0, + 789.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1599.0, + 1274.0, + 1599.0, + 1274.0, + 1634.0, + 952.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1599.0, + 1321.0, + 1599.0, + 1321.0, + 1634.0, + 1316.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1628.0, + 1157.0, + 1628.0, + 1157.0, + 1664.0, + 788.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1628.0, + 1405.0, + 1628.0, + 1405.0, + 1664.0, + 1257.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1661.0, + 1060.0, + 1661.0, + 1060.0, + 1694.0, + 789.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1661.0, + 1405.0, + 1661.0, + 1405.0, + 1694.0, + 1124.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1692.0, + 1054.0, + 1692.0, + 1054.0, + 1720.0, + 792.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1569.0, + 1405.0, + 1569.0, + 1405.0, + 1604.0, + 789.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1599.0, + 922.0, + 1599.0, + 922.0, + 1634.0, + 789.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1599.0, + 1274.0, + 1599.0, + 1274.0, + 1634.0, + 952.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1599.0, + 1321.0, + 1599.0, + 1321.0, + 1634.0, + 1316.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1628.0, + 1157.0, + 1628.0, + 1157.0, + 1664.0, + 788.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1628.0, + 1405.0, + 1628.0, + 1405.0, + 1664.0, + 1257.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1661.0, + 1060.0, + 1661.0, + 1060.0, + 1694.0, + 789.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1661.0, + 1405.0, + 1661.0, + 1405.0, + 1694.0, + 1124.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1692.0, + 1054.0, + 1692.0, + 1054.0, + 1720.0, + 792.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 396.0, + 1404.0, + 396.0, + 1404.0, + 430.0, + 296.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 424.0, + 415.0, + 424.0, + 415.0, + 463.0, + 292.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 424.0, + 1326.0, + 424.0, + 1326.0, + 463.0, + 436.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 424.0, + 1406.0, + 424.0, + 1406.0, + 463.0, + 1354.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 458.0, + 297.0, + 458.0, + 297.0, + 489.0, + 293.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 458.0, + 1405.0, + 458.0, + 1405.0, + 489.0, + 317.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 486.0, + 1407.0, + 486.0, + 1407.0, + 520.0, + 295.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 516.0, + 553.0, + 516.0, + 553.0, + 550.0, + 294.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 516.0, + 1040.0, + 516.0, + 1040.0, + 550.0, + 662.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 516.0, + 1406.0, + 516.0, + 1406.0, + 550.0, + 1162.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 544.0, + 1405.0, + 544.0, + 1405.0, + 580.0, + 295.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 576.0, + 896.0, + 576.0, + 896.0, + 610.0, + 294.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 576.0, + 1255.0, + 576.0, + 1255.0, + 610.0, + 974.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 619.0, + 416.0, + 619.0, + 416.0, + 655.0, + 293.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 619.0, + 539.0, + 619.0, + 539.0, + 655.0, + 477.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 619.0, + 1405.0, + 619.0, + 1405.0, + 655.0, + 561.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 651.0, + 1405.0, + 651.0, + 1405.0, + 685.0, + 293.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 681.0, + 1405.0, + 681.0, + 1405.0, + 717.0, + 294.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 712.0, + 1405.0, + 712.0, + 1405.0, + 744.0, + 293.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 740.0, + 1405.0, + 740.0, + 1405.0, + 774.0, + 294.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 770.0, + 887.0, + 770.0, + 887.0, + 806.0, + 294.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 814.0, + 1409.0, + 814.0, + 1409.0, + 854.0, + 294.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 846.0, + 1404.0, + 846.0, + 1404.0, + 881.0, + 294.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 877.0, + 1405.0, + 877.0, + 1405.0, + 910.0, + 295.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 905.0, + 1407.0, + 905.0, + 1407.0, + 940.0, + 293.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 936.0, + 996.0, + 936.0, + 996.0, + 972.0, + 294.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1270.0, + 772.0, + 1270.0, + 772.0, + 1302.0, + 295.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1299.0, + 776.0, + 1299.0, + 776.0, + 1335.0, + 293.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1330.0, + 772.0, + 1330.0, + 772.0, + 1364.0, + 295.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1360.0, + 769.0, + 1360.0, + 769.0, + 1391.0, + 296.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1390.0, + 775.0, + 1390.0, + 775.0, + 1421.0, + 296.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1418.0, + 772.0, + 1418.0, + 772.0, + 1454.0, + 294.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1451.0, + 774.0, + 1451.0, + 774.0, + 1482.0, + 296.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 776.0, + 1479.0, + 776.0, + 1512.0, + 295.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1510.0, + 775.0, + 1510.0, + 775.0, + 1540.0, + 295.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1536.0, + 774.0, + 1536.0, + 774.0, + 1572.0, + 294.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1570.0, + 775.0, + 1570.0, + 775.0, + 1601.0, + 297.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1599.0, + 773.0, + 1599.0, + 773.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1629.0, + 775.0, + 1629.0, + 775.0, + 1662.0, + 294.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1660.0, + 771.0, + 1660.0, + 771.0, + 1691.0, + 296.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1687.0, + 297.0, + 1687.0, + 297.0, + 1722.0, + 293.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1687.0, + 776.0, + 1687.0, + 776.0, + 1722.0, + 341.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1717.0, + 770.0, + 1717.0, + 770.0, + 1751.0, + 295.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1888.0, + 293.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1884.0, + 1408.0, + 1884.0, + 1408.0, + 1919.0, + 294.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1951.0, + 292.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1946.0, + 1405.0, + 1946.0, + 1405.0, + 1977.0, + 296.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 369.0, + 230.0, + 369.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 230.0, + 763.0, + 230.0, + 763.0, + 263.0, + 400.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 230.0, + 1406.0, + 230.0, + 1406.0, + 263.0, + 792.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 259.0, + 1406.0, + 259.0, + 1406.0, + 294.0, + 292.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 388.0, + 290.0, + 388.0, + 323.0, + 294.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 323.0, + 425.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 355.0, + 294.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 348.0, + 748.0, + 348.0, + 748.0, + 383.0, + 293.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1132.0, + 1405.0, + 1132.0, + 1405.0, + 1169.0, + 294.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1161.0, + 1407.0, + 1161.0, + 1407.0, + 1199.0, + 294.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1228.0, + 295.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1223.0, + 1397.0, + 1223.0, + 1397.0, + 1257.0, + 295.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1747.0, + 1404.0, + 1747.0, + 1404.0, + 1782.0, + 294.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1777.0, + 1406.0, + 1777.0, + 1406.0, + 1811.0, + 295.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1808.0, + 1151.0, + 1808.0, + 1151.0, + 1841.0, + 295.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1007.0, + 923.0, + 1007.0, + 923.0, + 1052.0, + 291.0, + 1052.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 942, + 1406, + 942, + 1406, + 1303, + 296, + 1303 + ], + "score": 0.982 + }, + { + "category_id": 3, + "poly": [ + 348, + 193, + 1348, + 193, + 1348, + 662, + 348, + 662 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1883, + 1404, + 1883, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 337, + 1416, + 1352, + 1416, + 1352, + 1710, + 337, + 1710 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 297, + 713, + 1405, + 713, + 1405, + 804, + 297, + 804 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 298, + 1752, + 1404, + 1752, + 1404, + 1842, + 298, + 1842 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 299, + 865, + 1402, + 865, + 1402, + 928, + 299, + 928 + ], + "score": 0.947 + }, + { + "category_id": 0, + "poly": [ + 299, + 1346, + 569, + 1346, + 569, + 1375, + 299, + 1375 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.791 + }, + { + "category_id": 13, + "poly": [ + 546, + 1811, + 658, + 1811, + 658, + 1840, + 546, + 1840 + ], + "score": 0.91, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 837, + 1890, + 890, + 1890, + 890, + 1916, + 837, + 1916 + ], + "score": 0.86, + "latex": "\\eta _ { l } , \\eta" + }, + { + "category_id": 13, + "poly": [ + 1221, + 2007, + 1245, + 2007, + 1245, + 2035, + 1221, + 2035 + ], + "score": 0.84, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1306, + 1918, + 1331, + 1918, + 1331, + 1945, + 1306, + 1945 + ], + "score": 0.84, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 327, + 2007, + 352, + 2007, + 352, + 2035, + 327, + 2035 + ], + "score": 0.84, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1381, + 1919, + 1398, + 1919, + 1398, + 1945, + 1381, + 1945 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 635, + 749, + 661, + 749, + 661, + 775, + 635, + 775 + ], + "score": 0.78, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 562, + 749, + 582, + 749, + 582, + 775, + 562, + 775 + ], + "score": 0.78, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 451, + 2008, + 469, + 2008, + 469, + 2035, + 451, + 2035 + ], + "score": 0.76, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1240, + 1890, + 1258, + 1890, + 1258, + 1912, + 1240, + 1912 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 207.0, + 403.0, + 207.0, + 403.0, + 232.0, + 368.0, + 232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 200.0, + 555.0, + 200.0, + 555.0, + 221.0, + 482.0, + 221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 209.0, + 746.0, + 209.0, + 746.0, + 230.0, + 714.0, + 230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 200.0, + 902.0, + 200.0, + 902.0, + 221.0, + 823.0, + 221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 207.0, + 1099.0, + 207.0, + 1099.0, + 231.0, + 1057.0, + 231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 200.0, + 1252.0, + 200.0, + 1252.0, + 221.0, + 1169.0, + 221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 250.0, + 401.0, + 250.0, + 401.0, + 271.0, + 370.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 248.0, + 760.0, + 248.0, + 760.0, + 353.0, + 689.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 227.0, + 1066.0, + 227.0, + 1066.0, + 376.0, + 1041.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 264.0, + 377.0, + 264.0, + 377.0, + 338.0, + 356.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 264.0, + 1097.0, + 264.0, + 1097.0, + 284.0, + 1058.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 289.0, + 403.0, + 289.0, + 403.0, + 314.0, + 369.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 317.0, + 746.0, + 317.0, + 746.0, + 341.0, + 709.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 317.0, + 1098.0, + 317.0, + 1098.0, + 341.0, + 1057.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 331.0, + 402.0, + 331.0, + 402.0, + 356.0, + 369.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 374.0, + 402.0, + 374.0, + 402.0, + 395.0, + 370.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 374.0, + 745.0, + 374.0, + 745.0, + 394.0, + 714.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 386.0, + 407.0, + 386.0, + 407.0, + 401.0, + 393.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 383.0, + 600.0, + 383.0, + 600.0, + 404.0, + 439.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 384.0, + 658.0, + 384.0, + 658.0, + 403.0, + 613.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 386.0, + 752.0, + 386.0, + 752.0, + 402.0, + 737.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 383.0, + 825.0, + 383.0, + 825.0, + 402.0, + 781.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 384.0, + 943.0, + 384.0, + 943.0, + 403.0, + 839.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 383.0, + 1003.0, + 383.0, + 1003.0, + 405.0, + 956.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 383.0, + 1196.0, + 383.0, + 1196.0, + 405.0, + 1149.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 382.0, + 1272.0, + 382.0, + 1272.0, + 405.0, + 1225.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 383.0, + 1346.0, + 383.0, + 1346.0, + 405.0, + 1302.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 399.0, + 588.0, + 399.0, + 588.0, + 420.0, + 449.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 401.0, + 930.0, + 401.0, + 930.0, + 419.0, + 794.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 399.0, + 1280.0, + 399.0, + 1280.0, + 420.0, + 1142.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 444.0, + 397.0, + 444.0, + 397.0, + 472.0, + 362.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 437.0, + 562.0, + 437.0, + 562.0, + 463.0, + 462.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 444.0, + 741.0, + 444.0, + 741.0, + 472.0, + 705.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 439.0, + 922.0, + 439.0, + 922.0, + 460.0, + 790.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 448.0, + 1083.0, + 448.0, + 1083.0, + 469.0, + 1050.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 438.0, + 1272.0, + 438.0, + 1272.0, + 459.0, + 1124.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 489.0, + 409.0, + 489.0, + 409.0, + 593.0, + 334.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 511.0, + 624.0, + 511.0, + 624.0, + 534.0, + 543.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 490.0, + 738.0, + 490.0, + 738.0, + 580.0, + 688.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 492.0, + 1094.0, + 492.0, + 1094.0, + 592.0, + 1022.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 530.0, + 609.0, + 530.0, + 609.0, + 549.0, + 526.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 528.0, + 740.0, + 528.0, + 740.0, + 553.0, + 705.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 543.0, + 602.0, + 543.0, + 602.0, + 568.0, + 542.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 550.0, + 544.0, + 550.0, + 544.0, + 561.0, + 524.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 564.0, + 606.0, + 564.0, + 606.0, + 602.0, + 543.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 570.0, + 738.0, + 570.0, + 738.0, + 595.0, + 708.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 598.0, + 614.0, + 598.0, + 614.0, + 615.0, + 524.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 622.0, + 651.0, + 622.0, + 651.0, + 643.0, + 415.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 621.0, + 835.0, + 621.0, + 835.0, + 644.0, + 796.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 621.0, + 918.0, + 621.0, + 918.0, + 644.0, + 872.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 621.0, + 996.0, + 621.0, + 996.0, + 644.0, + 950.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 613.0, + 1089.0, + 613.0, + 1089.0, + 641.0, + 1051.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 621.0, + 1179.0, + 621.0, + 1179.0, + 644.0, + 1140.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 622.0, + 1262.0, + 622.0, + 1262.0, + 644.0, + 1216.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 621.0, + 1340.0, + 621.0, + 1340.0, + 644.0, + 1294.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 639.0, + 581.0, + 639.0, + 581.0, + 660.0, + 444.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 634.0, + 925.0, + 634.0, + 925.0, + 661.0, + 786.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 636.0, + 1269.0, + 636.0, + 1269.0, + 661.0, + 1130.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 369.5, + 1106.0, + 369.5, + 1106.0, + 401.5, + 1059.0, + 401.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.25, + 611.5, + 400.25, + 611.5, + 400.25, + 639.5, + 366.25, + 639.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.75, + 610.0, + 743.75, + 610.0, + 743.75, + 640.5, + 710.75, + 640.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1423.0, + 590.0, + 1423.0, + 590.0, + 1444.0, + 388.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1421.0, + 931.0, + 1421.0, + 931.0, + 1447.0, + 735.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1422.0, + 1279.0, + 1422.0, + 1279.0, + 1447.0, + 1077.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1432.0, + 646.0, + 1432.0, + 646.0, + 1452.0, + 618.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1432.0, + 990.0, + 1432.0, + 990.0, + 1452.0, + 962.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1432.0, + 1333.0, + 1432.0, + 1333.0, + 1452.0, + 1306.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1443.0, + 437.0, + 1443.0, + 437.0, + 1466.0, + 358.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1445.0, + 464.0, + 1445.0, + 464.0, + 1462.0, + 440.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1445.0, + 489.0, + 1445.0, + 489.0, + 1462.0, + 466.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1445.0, + 515.0, + 1445.0, + 515.0, + 1462.0, + 491.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1445.0, + 566.0, + 1445.0, + 566.0, + 1462.0, + 517.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1445.0, + 591.0, + 1445.0, + 591.0, + 1461.0, + 568.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1443.0, + 781.0, + 1443.0, + 781.0, + 1466.0, + 702.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1445.0, + 807.0, + 1445.0, + 807.0, + 1462.0, + 783.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1445.0, + 833.0, + 1445.0, + 833.0, + 1462.0, + 810.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 1445.0, + 858.0, + 1445.0, + 858.0, + 1462.0, + 834.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1445.0, + 884.0, + 1445.0, + 884.0, + 1462.0, + 861.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1445.0, + 910.0, + 1445.0, + 910.0, + 1461.0, + 885.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1445.0, + 936.0, + 1445.0, + 936.0, + 1462.0, + 911.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1444.0, + 1151.0, + 1444.0, + 1151.0, + 1464.0, + 1047.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1445.0, + 1177.0, + 1445.0, + 1177.0, + 1461.0, + 1152.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1445.0, + 1202.0, + 1445.0, + 1202.0, + 1461.0, + 1178.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1445.0, + 1228.0, + 1445.0, + 1228.0, + 1461.0, + 1205.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1445.0, + 1279.0, + 1445.0, + 1279.0, + 1462.0, + 1230.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1465.0, + 411.0, + 1465.0, + 411.0, + 1518.0, + 342.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1471.0, + 438.0, + 1471.0, + 438.0, + 1488.0, + 414.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1471.0, + 464.0, + 1471.0, + 464.0, + 1488.0, + 439.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1471.0, + 489.0, + 1471.0, + 489.0, + 1487.0, + 466.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 1471.0, + 515.0, + 1471.0, + 515.0, + 1488.0, + 490.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1471.0, + 567.0, + 1471.0, + 567.0, + 1487.0, + 517.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1471.0, + 591.0, + 1471.0, + 591.0, + 1488.0, + 568.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1469.0, + 646.0, + 1469.0, + 646.0, + 1491.0, + 619.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1465.0, + 755.0, + 1465.0, + 755.0, + 1518.0, + 685.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1471.0, + 781.0, + 1471.0, + 781.0, + 1487.0, + 757.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1471.0, + 807.0, + 1471.0, + 807.0, + 1488.0, + 783.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1471.0, + 833.0, + 1471.0, + 833.0, + 1487.0, + 810.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1470.0, + 858.0, + 1470.0, + 858.0, + 1488.0, + 835.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1471.0, + 910.0, + 1471.0, + 910.0, + 1487.0, + 861.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1471.0, + 936.0, + 1471.0, + 936.0, + 1488.0, + 912.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1469.0, + 990.0, + 1469.0, + 990.0, + 1491.0, + 963.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 1465.0, + 1203.0, + 1465.0, + 1203.0, + 1516.0, + 1028.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1471.0, + 1229.0, + 1471.0, + 1229.0, + 1488.0, + 1205.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1471.0, + 1254.0, + 1471.0, + 1254.0, + 1487.0, + 1230.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1471.0, + 1280.0, + 1471.0, + 1280.0, + 1488.0, + 1256.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1470.0, + 1333.0, + 1470.0, + 1333.0, + 1491.0, + 1307.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1496.0, + 411.0, + 1496.0, + 411.0, + 1516.0, + 359.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1497.0, + 437.0, + 1497.0, + 437.0, + 1513.0, + 414.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1497.0, + 462.0, + 1497.0, + 462.0, + 1514.0, + 439.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1497.0, + 489.0, + 1497.0, + 489.0, + 1513.0, + 465.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1497.0, + 515.0, + 1497.0, + 515.0, + 1514.0, + 491.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1497.0, + 567.0, + 1497.0, + 567.0, + 1514.0, + 517.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1497.0, + 591.0, + 1497.0, + 591.0, + 1513.0, + 568.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1495.0, + 782.0, + 1495.0, + 782.0, + 1517.0, + 702.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1497.0, + 807.0, + 1497.0, + 807.0, + 1513.0, + 783.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1497.0, + 833.0, + 1497.0, + 833.0, + 1513.0, + 810.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 1497.0, + 858.0, + 1497.0, + 858.0, + 1513.0, + 834.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1497.0, + 911.0, + 1497.0, + 911.0, + 1513.0, + 861.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1497.0, + 936.0, + 1497.0, + 936.0, + 1514.0, + 912.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1495.0, + 1254.0, + 1495.0, + 1254.0, + 1516.0, + 1047.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1497.0, + 1280.0, + 1497.0, + 1280.0, + 1514.0, + 1256.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1506.0, + 364.0, + 1506.0, + 364.0, + 1650.0, + 338.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1508.0, + 646.0, + 1508.0, + 646.0, + 1529.0, + 619.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1506.0, + 710.0, + 1506.0, + 710.0, + 1653.0, + 681.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1508.0, + 990.0, + 1508.0, + 990.0, + 1529.0, + 963.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1508.0, + 1051.0, + 1508.0, + 1051.0, + 1549.0, + 1030.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1508.0, + 1333.0, + 1508.0, + 1333.0, + 1529.0, + 1308.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1522.0, + 411.0, + 1522.0, + 411.0, + 1542.0, + 356.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1523.0, + 437.0, + 1523.0, + 437.0, + 1538.0, + 412.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1523.0, + 464.0, + 1523.0, + 464.0, + 1539.0, + 440.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1523.0, + 489.0, + 1523.0, + 489.0, + 1538.0, + 466.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1523.0, + 515.0, + 1523.0, + 515.0, + 1538.0, + 491.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1523.0, + 540.0, + 1523.0, + 540.0, + 1539.0, + 517.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1523.0, + 567.0, + 1523.0, + 567.0, + 1538.0, + 543.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1523.0, + 592.0, + 1523.0, + 592.0, + 1538.0, + 568.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1521.0, + 660.0, + 1521.0, + 660.0, + 1592.0, + 638.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1522.0, + 755.0, + 1522.0, + 755.0, + 1542.0, + 699.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1523.0, + 782.0, + 1523.0, + 782.0, + 1538.0, + 757.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1523.0, + 807.0, + 1523.0, + 807.0, + 1539.0, + 784.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 1523.0, + 857.0, + 1523.0, + 857.0, + 1539.0, + 808.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1523.0, + 885.0, + 1523.0, + 885.0, + 1539.0, + 862.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1523.0, + 911.0, + 1523.0, + 911.0, + 1538.0, + 886.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1523.0, + 936.0, + 1523.0, + 936.0, + 1538.0, + 912.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1522.0, + 1004.0, + 1522.0, + 1004.0, + 1592.0, + 982.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1522.0, + 1253.0, + 1522.0, + 1253.0, + 1542.0, + 1043.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1523.0, + 1280.0, + 1523.0, + 1280.0, + 1539.0, + 1256.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1521.0, + 1349.0, + 1521.0, + 1349.0, + 1592.0, + 1326.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1546.0, + 516.0, + 1546.0, + 516.0, + 1567.0, + 357.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1548.0, + 541.0, + 1548.0, + 541.0, + 1565.0, + 517.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1548.0, + 567.0, + 1548.0, + 567.0, + 1565.0, + 543.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1548.0, + 592.0, + 1548.0, + 592.0, + 1565.0, + 568.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1546.0, + 646.0, + 1546.0, + 646.0, + 1569.0, + 620.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1547.0, + 860.0, + 1547.0, + 860.0, + 1567.0, + 700.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1547.0, + 885.0, + 1547.0, + 885.0, + 1565.0, + 861.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1548.0, + 911.0, + 1548.0, + 911.0, + 1565.0, + 886.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1548.0, + 936.0, + 1548.0, + 936.0, + 1565.0, + 912.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1546.0, + 990.0, + 1546.0, + 990.0, + 1569.0, + 964.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 1539.0, + 1254.0, + 1539.0, + 1254.0, + 1607.0, + 1028.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1548.0, + 1280.0, + 1548.0, + 1280.0, + 1565.0, + 1256.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1546.0, + 1333.0, + 1546.0, + 1333.0, + 1569.0, + 1308.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1571.0, + 567.0, + 1571.0, + 567.0, + 1595.0, + 354.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1574.0, + 592.0, + 1574.0, + 592.0, + 1591.0, + 568.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1571.0, + 911.0, + 1571.0, + 911.0, + 1595.0, + 698.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1574.0, + 936.0, + 1574.0, + 936.0, + 1591.0, + 912.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1571.0, + 1126.0, + 1571.0, + 1126.0, + 1597.0, + 1041.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1574.0, + 1151.0, + 1574.0, + 1151.0, + 1591.0, + 1127.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1573.0, + 1254.0, + 1573.0, + 1254.0, + 1592.0, + 1152.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1574.0, + 1279.0, + 1574.0, + 1279.0, + 1591.0, + 1256.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1599.0, + 567.0, + 1599.0, + 567.0, + 1620.0, + 356.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1600.0, + 592.0, + 1600.0, + 592.0, + 1617.0, + 568.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1588.0, + 645.0, + 1588.0, + 645.0, + 1606.0, + 622.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1599.0, + 911.0, + 1599.0, + 911.0, + 1620.0, + 699.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1600.0, + 936.0, + 1600.0, + 936.0, + 1617.0, + 912.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1585.0, + 990.0, + 1585.0, + 990.0, + 1607.0, + 964.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1595.0, + 1151.0, + 1595.0, + 1151.0, + 1651.0, + 1030.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1599.0, + 1177.0, + 1599.0, + 1177.0, + 1617.0, + 1152.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1600.0, + 1280.0, + 1600.0, + 1280.0, + 1617.0, + 1178.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1588.0, + 1332.0, + 1588.0, + 1332.0, + 1606.0, + 1309.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1623.0, + 567.0, + 1623.0, + 567.0, + 1645.0, + 356.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1626.0, + 591.0, + 1626.0, + 591.0, + 1641.0, + 568.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1627.0, + 635.0, + 1627.0, + 635.0, + 1644.0, + 622.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1623.0, + 911.0, + 1623.0, + 911.0, + 1645.0, + 699.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1626.0, + 936.0, + 1626.0, + 936.0, + 1641.0, + 912.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1627.0, + 979.0, + 1627.0, + 979.0, + 1644.0, + 965.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1623.0, + 1177.0, + 1623.0, + 1177.0, + 1646.0, + 1042.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1626.0, + 1229.0, + 1626.0, + 1229.0, + 1642.0, + 1178.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1626.0, + 1280.0, + 1626.0, + 1280.0, + 1642.0, + 1230.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 1627.0, + 1323.0, + 1627.0, + 1323.0, + 1644.0, + 1310.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1650.0, + 592.0, + 1650.0, + 592.0, + 1670.0, + 356.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1650.0, + 885.0, + 1650.0, + 885.0, + 1670.0, + 699.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1651.0, + 936.0, + 1651.0, + 936.0, + 1668.0, + 886.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1651.0, + 1100.0, + 1651.0, + 1100.0, + 1670.0, + 1042.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 1651.0, + 1126.0, + 1651.0, + 1126.0, + 1668.0, + 1101.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1651.0, + 1150.0, + 1651.0, + 1150.0, + 1668.0, + 1127.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1651.0, + 1177.0, + 1651.0, + 1177.0, + 1668.0, + 1152.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1651.0, + 1202.0, + 1651.0, + 1202.0, + 1668.0, + 1178.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1651.0, + 1228.0, + 1651.0, + 1228.0, + 1668.0, + 1205.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1651.0, + 1254.0, + 1651.0, + 1254.0, + 1668.0, + 1229.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1651.0, + 1280.0, + 1651.0, + 1280.0, + 1668.0, + 1256.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1670.0, + 595.0, + 1670.0, + 595.0, + 1692.0, + 383.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1664.0, + 637.0, + 1664.0, + 637.0, + 1682.0, + 612.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1670.0, + 939.0, + 1670.0, + 939.0, + 1692.0, + 727.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1665.0, + 980.0, + 1665.0, + 980.0, + 1682.0, + 956.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1669.0, + 1281.0, + 1669.0, + 1281.0, + 1691.0, + 1070.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 1664.0, + 1324.0, + 1664.0, + 1324.0, + 1682.0, + 1300.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 1684.0, + 579.0, + 1684.0, + 579.0, + 1708.0, + 400.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1684.0, + 924.0, + 1684.0, + 924.0, + 1707.0, + 744.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1684.0, + 1266.0, + 1684.0, + 1266.0, + 1708.0, + 1087.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 711.0, + 1406.0, + 711.0, + 1406.0, + 750.0, + 294.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 744.0, + 561.0, + 744.0, + 561.0, + 776.0, + 294.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 744.0, + 634.0, + 744.0, + 634.0, + 776.0, + 583.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 744.0, + 1403.0, + 744.0, + 1403.0, + 776.0, + 662.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 773.0, + 879.0, + 773.0, + 879.0, + 806.0, + 295.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1752.0, + 1405.0, + 1752.0, + 1405.0, + 1785.0, + 294.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1780.0, + 1405.0, + 1780.0, + 1405.0, + 1816.0, + 295.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1811.0, + 545.0, + 1811.0, + 545.0, + 1842.0, + 292.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1811.0, + 670.0, + 1811.0, + 670.0, + 1842.0, + 659.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1340.0, + 574.0, + 1340.0, + 574.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 942.0, + 1403.0, + 942.0, + 1403.0, + 975.0, + 296.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 969.0, + 1404.0, + 969.0, + 1404.0, + 1007.0, + 293.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 999.0, + 1406.0, + 999.0, + 1406.0, + 1037.0, + 293.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1031.0, + 1404.0, + 1031.0, + 1404.0, + 1067.0, + 294.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1060.0, + 1408.0, + 1060.0, + 1408.0, + 1099.0, + 294.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1091.0, + 1407.0, + 1091.0, + 1407.0, + 1126.0, + 294.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1121.0, + 1408.0, + 1121.0, + 1408.0, + 1157.0, + 294.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1153.0, + 1404.0, + 1153.0, + 1404.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1177.0, + 1406.0, + 1177.0, + 1406.0, + 1217.0, + 293.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1209.0, + 1407.0, + 1209.0, + 1407.0, + 1248.0, + 291.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1242.0, + 1404.0, + 1242.0, + 1404.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1269.0, + 1354.0, + 1269.0, + 1354.0, + 1307.0, + 293.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1883.0, + 836.0, + 1883.0, + 836.0, + 1919.0, + 294.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1883.0, + 1239.0, + 1883.0, + 1239.0, + 1919.0, + 891.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1919.0, + 1259.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1305.0, + 1910.0, + 1305.0, + 1949.0, + 292.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 1910.0, + 1380.0, + 1910.0, + 1380.0, + 1949.0, + 1332.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1910.0, + 1409.0, + 1910.0, + 1409.0, + 1949.0, + 1399.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 326.0, + 2002.0, + 326.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 2002.0, + 450.0, + 2002.0, + 450.0, + 2038.0, + 353.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 2002.0, + 1220.0, + 2002.0, + 1220.0, + 2038.0, + 470.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2038.0, + 1246.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 863.0, + 1404.0, + 863.0, + 1404.0, + 902.0, + 294.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 898.0, + 873.0, + 898.0, + 873.0, + 930.0, + 297.0, + 930.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1557, + 1405, + 1557, + 1405, + 1828, + 298, + 1828 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 983, + 1404, + 983, + 1404, + 1194, + 298, + 1194 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1304, + 1405, + 1304, + 1405, + 1426, + 299, + 1426 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 348, + 232, + 1348, + 232, + 1348, + 689, + 348, + 689 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 301, + 1944, + 1403, + 1944, + 1403, + 2033, + 301, + 2033 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 297, + 739, + 1404, + 739, + 1404, + 831, + 297, + 831 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 299, + 907, + 1400, + 907, + 1400, + 969, + 299, + 969 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 300, + 1885, + 487, + 1885, + 487, + 1918, + 300, + 1918 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 300, + 1243, + 575, + 1243, + 575, + 1274, + 300, + 1274 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 300, + 1481, + 543, + 1481, + 543, + 1516, + 300, + 1516 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.751 + }, + { + "category_id": 13, + "poly": [ + 479, + 1072, + 592, + 1072, + 592, + 1102, + 479, + 1102 + ], + "score": 0.92, + "latex": "\\dot { \\tau } = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 716, + 775, + 741, + 775, + 741, + 803, + 716, + 803 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1066, + 989, + 1092, + 989, + 1092, + 1016, + 1066, + 1016 + ], + "score": 0.82, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1144, + 989, + 1163, + 989, + 1163, + 1015, + 1144, + 1015 + ], + "score": 0.81, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 297, + 1398, + 322, + 1398, + 322, + 1427, + 297, + 1427 + ], + "score": 0.8, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 650, + 776, + 669, + 776, + 669, + 803, + 650, + 803 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 512, + 1015, + 544, + 1015, + 544, + 1042, + 512, + 1042 + ], + "score": 0.79, + "latex": "\\tau ?" + }, + { + "category_id": 13, + "poly": [ + 1254, + 1138, + 1273, + 1138, + 1273, + 1160, + 1254, + 1160 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1025, + 1048, + 1045, + 1048, + 1045, + 1071, + 1025, + 1071 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1336, + 1370, + 1355, + 1370, + 1355, + 1396, + 1336, + 1396 + ], + "score": 0.77, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 796, + 1018, + 816, + 1018, + 816, + 1042, + 796, + 1042 + ], + "score": 0.76, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 821, + 1109, + 841, + 1109, + 841, + 1131, + 821, + 1131 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 297, + 913, + 316, + 913, + 316, + 939, + 297, + 939 + ], + "score": 0.74, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 943, + 776, + 963, + 776, + 963, + 798, + 943, + 798 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1346, + 747, + 1366, + 747, + 1366, + 768, + 1346, + 768 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1310, + 1169, + 1328, + 1169, + 1328, + 1190, + 1310, + 1190 + ], + "score": 0.72, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1339, + 1019, + 1358, + 1019, + 1358, + 1041, + 1339, + 1041 + ], + "score": 0.69, + "latex": "\\tau" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 247.0, + 413.0, + 247.0, + 413.0, + 270.0, + 368.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 238.0, + 566.0, + 238.0, + 566.0, + 260.0, + 493.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 246.0, + 747.0, + 246.0, + 747.0, + 271.0, + 712.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 240.0, + 909.0, + 240.0, + 909.0, + 259.0, + 830.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 247.0, + 1108.0, + 247.0, + 1108.0, + 270.0, + 1057.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 238.0, + 1264.0, + 238.0, + 1264.0, + 260.0, + 1180.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 281.0, + 479.0, + 281.0, + 479.0, + 289.0, + 469.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 264.0, + 1066.0, + 264.0, + 1066.0, + 311.0, + 1042.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 285.0, + 414.0, + 285.0, + 414.0, + 313.0, + 367.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 288.0, + 768.0, + 288.0, + 768.0, + 391.0, + 685.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 288.0, + 1107.0, + 288.0, + 1107.0, + 311.0, + 1054.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 301.0, + 377.0, + 301.0, + 377.0, + 377.0, + 354.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 318.0, + 424.0, + 318.0, + 424.0, + 326.0, + 415.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 302.0, + 1066.0, + 302.0, + 1066.0, + 415.0, + 1039.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 324.0, + 413.0, + 324.0, + 413.0, + 353.0, + 369.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 327.0, + 1108.0, + 327.0, + 1108.0, + 351.0, + 1057.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 361.0, + 763.0, + 361.0, + 763.0, + 373.0, + 748.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 369.0, + 411.0, + 369.0, + 411.0, + 393.0, + 369.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 369.0, + 1107.0, + 369.0, + 1107.0, + 393.0, + 1057.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 376.0, + 1124.0, + 376.0, + 1124.0, + 390.0, + 1109.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 407.0, + 441.0, + 407.0, + 441.0, + 439.0, + 369.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 419.0, + 551.0, + 419.0, + 551.0, + 439.0, + 506.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 420.0, + 647.0, + 420.0, + 647.0, + 441.0, + 618.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 411.0, + 745.0, + 411.0, + 745.0, + 433.0, + 714.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 419.0, + 890.0, + 419.0, + 890.0, + 440.0, + 845.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 419.0, + 945.0, + 419.0, + 945.0, + 440.0, + 901.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 420.0, + 996.0, + 420.0, + 996.0, + 441.0, + 961.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 409.0, + 1107.0, + 409.0, + 1107.0, + 433.0, + 1057.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 405.0, + 1191.0, + 405.0, + 1191.0, + 440.0, + 1145.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 409.0, + 1243.0, + 409.0, + 1243.0, + 439.0, + 1198.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 419.0, + 1295.0, + 419.0, + 1295.0, + 440.0, + 1250.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 420.0, + 1338.0, + 420.0, + 1338.0, + 441.0, + 1306.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 431.0, + 584.0, + 431.0, + 584.0, + 461.0, + 475.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 432.0, + 924.0, + 432.0, + 924.0, + 461.0, + 814.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 432.0, + 1276.0, + 432.0, + 1276.0, + 461.0, + 1167.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 471.0, + 408.0, + 471.0, + 408.0, + 499.0, + 360.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 464.0, + 572.0, + 464.0, + 572.0, + 490.0, + 473.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 473.0, + 749.0, + 473.0, + 749.0, + 498.0, + 705.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 465.0, + 933.0, + 465.0, + 933.0, + 487.0, + 800.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 473.0, + 1094.0, + 473.0, + 1094.0, + 498.0, + 1050.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 465.0, + 1285.0, + 465.0, + 1285.0, + 487.0, + 1136.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 506.0, + 423.0, + 506.0, + 423.0, + 622.0, + 333.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 519.0, + 529.0, + 519.0, + 529.0, + 544.0, + 512.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 520.0, + 589.0, + 520.0, + 589.0, + 549.0, + 566.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 513.0, + 749.0, + 513.0, + 749.0, + 607.0, + 688.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 510.0, + 766.0, + 510.0, + 766.0, + 525.0, + 751.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 513.0, + 1110.0, + 513.0, + 1110.0, + 605.0, + 1036.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 524.0, + 1163.0, + 524.0, + 1163.0, + 537.0, + 1148.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 522.0, + 1222.0, + 522.0, + 1222.0, + 548.0, + 1200.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 527.0, + 1275.0, + 527.0, + 1275.0, + 545.0, + 1255.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 535.0, + 1322.0, + 535.0, + 1322.0, + 544.0, + 1313.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 554.0, + 749.0, + 554.0, + 749.0, + 579.0, + 705.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 553.0, + 1092.0, + 553.0, + 1092.0, + 579.0, + 1050.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 560.0, + 1325.0, + 560.0, + 1325.0, + 570.0, + 1313.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 567.0, + 406.0, + 567.0, + 406.0, + 597.0, + 360.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 602.0, + 436.0, + 602.0, + 436.0, + 639.0, + 358.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 587.0, + 520.0, + 587.0, + 520.0, + 627.0, + 438.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 595.0, + 748.0, + 595.0, + 748.0, + 619.0, + 706.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 595.0, + 1092.0, + 595.0, + 1092.0, + 619.0, + 1050.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 614.0, + 497.0, + 614.0, + 497.0, + 646.0, + 438.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 646.0, + 489.0, + 646.0, + 489.0, + 666.0, + 444.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 646.0, + 544.0, + 646.0, + 544.0, + 666.0, + 499.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 645.0, + 601.0, + 645.0, + 601.0, + 668.0, + 554.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 646.0, + 648.0, + 646.0, + 648.0, + 666.0, + 607.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 646.0, + 834.0, + 646.0, + 834.0, + 667.0, + 788.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 646.0, + 889.0, + 646.0, + 889.0, + 667.0, + 842.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 645.0, + 943.0, + 645.0, + 943.0, + 667.0, + 893.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 646.0, + 987.0, + 646.0, + 987.0, + 667.0, + 953.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 646.0, + 1178.0, + 646.0, + 1178.0, + 665.0, + 1131.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 646.0, + 1234.0, + 646.0, + 1234.0, + 667.0, + 1186.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 645.0, + 1288.0, + 645.0, + 1288.0, + 667.0, + 1237.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 646.0, + 1330.0, + 646.0, + 1330.0, + 667.0, + 1297.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 661.0, + 575.0, + 661.0, + 575.0, + 687.0, + 469.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 661.0, + 919.0, + 661.0, + 919.0, + 687.0, + 814.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 661.0, + 1263.0, + 661.0, + 1263.0, + 687.0, + 1157.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 418.5, + 498.0, + 418.5, + 498.0, + 441.5, + 451.0, + 441.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.75, + 418.5, + 606.75, + 418.5, + 606.75, + 441.5, + 559.75, + 441.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.75, + 417.5, + 778.75, + 417.5, + 778.75, + 441.5, + 732.75, + 441.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.25, + 417.0, + 836.25, + 417.0, + 836.25, + 441.5, + 787.25, + 441.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 418.5, + 1139.0, + 418.5, + 1139.0, + 441.5, + 1093.0, + 441.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 632.5, + 434.0, + 632.5, + 434.0, + 671.5, + 360.0, + 671.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 634.0, + 780.0, + 634.0, + 780.0, + 670.0, + 705.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.75, + 634.0, + 1122.75, + 634.0, + 1122.75, + 669.5, + 1048.75, + 669.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 737.0, + 1345.0, + 737.0, + 1345.0, + 777.0, + 292.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 737.0, + 1405.0, + 737.0, + 1405.0, + 777.0, + 1367.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 771.0, + 649.0, + 771.0, + 649.0, + 804.0, + 295.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 771.0, + 715.0, + 771.0, + 715.0, + 804.0, + 670.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 771.0, + 942.0, + 771.0, + 942.0, + 804.0, + 742.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 771.0, + 1404.0, + 771.0, + 1404.0, + 804.0, + 964.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 801.0, + 734.0, + 801.0, + 734.0, + 831.0, + 296.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1887.0, + 489.0, + 1887.0, + 489.0, + 1920.0, + 297.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1238.0, + 579.0, + 1238.0, + 579.0, + 1279.0, + 293.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1475.0, + 549.0, + 1475.0, + 549.0, + 1525.0, + 291.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1557.0, + 1408.0, + 1557.0, + 1408.0, + 1593.0, + 293.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1586.0, + 1407.0, + 1586.0, + 1407.0, + 1627.0, + 291.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1615.0, + 1405.0, + 1615.0, + 1405.0, + 1654.0, + 292.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1645.0, + 1406.0, + 1645.0, + 1406.0, + 1683.0, + 293.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1677.0, + 1407.0, + 1677.0, + 1407.0, + 1713.0, + 295.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1705.0, + 1403.0, + 1705.0, + 1403.0, + 1741.0, + 295.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1735.0, + 1403.0, + 1735.0, + 1403.0, + 1771.0, + 295.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1766.0, + 1405.0, + 1766.0, + 1405.0, + 1802.0, + 295.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1797.0, + 532.0, + 1797.0, + 532.0, + 1830.0, + 293.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 983.0, + 1065.0, + 983.0, + 1065.0, + 1017.0, + 294.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 983.0, + 1143.0, + 983.0, + 1143.0, + 1017.0, + 1093.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 983.0, + 1404.0, + 983.0, + 1404.0, + 1017.0, + 1164.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1013.0, + 511.0, + 1013.0, + 511.0, + 1047.0, + 294.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1013.0, + 795.0, + 1013.0, + 795.0, + 1047.0, + 545.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1013.0, + 1338.0, + 1013.0, + 1338.0, + 1047.0, + 817.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1013.0, + 1404.0, + 1013.0, + 1404.0, + 1047.0, + 1359.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1045.0, + 1024.0, + 1045.0, + 1024.0, + 1075.0, + 296.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1045.0, + 1404.0, + 1045.0, + 1404.0, + 1075.0, + 1046.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1071.0, + 478.0, + 1071.0, + 478.0, + 1108.0, + 293.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1071.0, + 1406.0, + 1071.0, + 1406.0, + 1108.0, + 593.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1103.0, + 820.0, + 1103.0, + 820.0, + 1137.0, + 293.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1103.0, + 1404.0, + 1103.0, + 1404.0, + 1137.0, + 842.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1135.0, + 1253.0, + 1135.0, + 1253.0, + 1165.0, + 296.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1135.0, + 1402.0, + 1135.0, + 1402.0, + 1165.0, + 1274.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1163.0, + 1309.0, + 1163.0, + 1309.0, + 1197.0, + 294.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 1163.0, + 1346.0, + 1163.0, + 1346.0, + 1197.0, + 1329.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1301.0, + 1409.0, + 1301.0, + 1409.0, + 1340.0, + 293.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1334.0, + 1406.0, + 1334.0, + 1406.0, + 1368.0, + 293.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1366.0, + 1335.0, + 1366.0, + 1335.0, + 1399.0, + 295.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1366.0, + 1405.0, + 1366.0, + 1405.0, + 1399.0, + 1356.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1395.0, + 1161.0, + 1395.0, + 1161.0, + 1430.0, + 323.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1941.0, + 762.0, + 1941.0, + 762.0, + 1974.0, + 297.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1940.0, + 1407.0, + 1940.0, + 1407.0, + 1976.0, + 807.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1970.0, + 1404.0, + 1970.0, + 1404.0, + 2007.0, + 320.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2001.0, + 1052.0, + 2001.0, + 1052.0, + 2036.0, + 324.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 905.0, + 296.0, + 905.0, + 296.0, + 942.0, + 293.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 905.0, + 1407.0, + 905.0, + 1407.0, + 942.0, + 317.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 937.0, + 793.0, + 937.0, + 793.0, + 972.0, + 295.0, + 972.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 345, + 1405, + 345, + 1405, + 556, + 298, + 556 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 303, + 229, + 1401, + 229, + 1401, + 321, + 303, + 321 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 105, + 300, + 105 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 298, + 1260, + 1399, + 1260, + 1399, + 1322, + 298, + 1322 + ], + "score": 0.863 + }, + { + "category_id": 1, + "poly": [ + 299, + 1460, + 1401, + 1460, + 1401, + 1523, + 299, + 1523 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 298, + 1344, + 1406, + 1344, + 1406, + 1437, + 298, + 1437 + ], + "score": 0.849 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 296, + 692, + 1402, + 692, + 1402, + 785, + 296, + 785 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 300, + 1175, + 1401, + 1175, + 1401, + 1238, + 300, + 1238 + ], + "score": 0.815 + }, + { + "category_id": 1, + "poly": [ + 293, + 806, + 1402, + 806, + 1402, + 870, + 293, + 870 + ], + "score": 0.812 + }, + { + "category_id": 1, + "poly": [ + 301, + 1658, + 1403, + 1658, + 1403, + 1752, + 301, + 1752 + ], + "score": 0.811 + }, + { + "category_id": 1, + "poly": [ + 301, + 1544, + 1400, + 1544, + 1400, + 1637, + 301, + 1637 + ], + "score": 0.777 + }, + { + "category_id": 1, + "poly": [ + 291, + 1772, + 1404, + 1772, + 1404, + 1836, + 291, + 1836 + ], + "score": 0.773 + }, + { + "category_id": 1, + "poly": [ + 298, + 578, + 1407, + 578, + 1407, + 669, + 298, + 669 + ], + "score": 0.771 + }, + { + "category_id": 1, + "poly": [ + 296, + 891, + 1402, + 891, + 1402, + 955, + 296, + 955 + ], + "score": 0.768 + }, + { + "category_id": 1, + "poly": [ + 301, + 1060, + 1401, + 1060, + 1401, + 1153, + 301, + 1153 + ], + "score": 0.761 + }, + { + "category_id": 1, + "poly": [ + 294, + 976, + 1402, + 976, + 1402, + 1040, + 294, + 1040 + ], + "score": 0.728 + }, + { + "category_id": 1, + "poly": [ + 300, + 1973, + 1401, + 1973, + 1401, + 2034, + 300, + 2034 + ], + "score": 0.723 + }, + { + "category_id": 1, + "poly": [ + 299, + 1857, + 1403, + 1857, + 1403, + 1951, + 299, + 1951 + ], + "score": 0.667 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 294.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 341.0, + 1406.0, + 341.0, + 1406.0, + 380.0, + 293.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 373.0, + 1403.0, + 373.0, + 1403.0, + 407.0, + 322.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 403.0, + 1408.0, + 403.0, + 1408.0, + 437.0, + 322.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 434.0, + 1406.0, + 434.0, + 1406.0, + 466.0, + 320.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 463.0, + 1406.0, + 463.0, + 1406.0, + 497.0, + 322.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 493.0, + 1406.0, + 493.0, + 1406.0, + 531.0, + 321.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 522.0, + 1038.0, + 522.0, + 1038.0, + 559.0, + 320.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 259.0, + 1406.0, + 259.0, + 1406.0, + 297.0, + 323.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 289.0, + 832.0, + 289.0, + 832.0, + 327.0, + 322.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1261.0, + 1405.0, + 1261.0, + 1405.0, + 1296.0, + 296.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1290.0, + 1293.0, + 1290.0, + 1293.0, + 1325.0, + 324.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1460.0, + 1405.0, + 1460.0, + 1405.0, + 1495.0, + 295.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1488.0, + 1294.0, + 1488.0, + 1294.0, + 1526.0, + 320.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1340.0, + 714.0, + 1340.0, + 714.0, + 1381.0, + 292.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1342.0, + 875.0, + 1342.0, + 875.0, + 1376.0, + 736.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1341.0, + 1411.0, + 1341.0, + 1411.0, + 1378.0, + 993.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1372.0, + 410.0, + 1372.0, + 410.0, + 1409.0, + 320.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1371.0, + 551.0, + 1371.0, + 551.0, + 1409.0, + 471.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1371.0, + 838.0, + 1371.0, + 838.0, + 1409.0, + 763.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1375.0, + 1404.0, + 1375.0, + 1404.0, + 1409.0, + 899.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1404.0, + 1089.0, + 1404.0, + 1089.0, + 1438.0, + 323.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 693.0, + 1406.0, + 693.0, + 1406.0, + 726.0, + 298.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 719.0, + 1408.0, + 719.0, + 1408.0, + 761.0, + 320.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 754.0, + 608.0, + 754.0, + 608.0, + 784.0, + 324.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1176.0, + 1405.0, + 1176.0, + 1405.0, + 1212.0, + 296.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1205.0, + 1138.0, + 1205.0, + 1138.0, + 1240.0, + 321.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 801.0, + 1405.0, + 801.0, + 1405.0, + 846.0, + 293.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 837.0, + 883.0, + 837.0, + 883.0, + 871.0, + 321.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1659.0, + 1405.0, + 1659.0, + 1405.0, + 1693.0, + 296.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1689.0, + 1404.0, + 1689.0, + 1404.0, + 1723.0, + 324.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1720.0, + 882.0, + 1720.0, + 882.0, + 1753.0, + 324.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1541.0, + 1405.0, + 1541.0, + 1405.0, + 1580.0, + 292.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1574.0, + 1405.0, + 1574.0, + 1405.0, + 1611.0, + 322.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1604.0, + 1240.0, + 1604.0, + 1240.0, + 1640.0, + 321.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1767.0, + 1409.0, + 1767.0, + 1409.0, + 1812.0, + 294.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1804.0, + 911.0, + 1804.0, + 911.0, + 1837.0, + 322.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 577.0, + 1405.0, + 577.0, + 1405.0, + 612.0, + 295.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 607.0, + 1408.0, + 607.0, + 1408.0, + 643.0, + 320.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 637.0, + 397.0, + 637.0, + 397.0, + 670.0, + 321.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 891.0, + 1404.0, + 891.0, + 1404.0, + 927.0, + 298.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 922.0, + 1341.0, + 922.0, + 1341.0, + 956.0, + 323.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1057.0, + 1407.0, + 1057.0, + 1407.0, + 1097.0, + 294.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1088.0, + 1406.0, + 1088.0, + 1406.0, + 1128.0, + 321.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1121.0, + 1008.0, + 1121.0, + 1008.0, + 1155.0, + 323.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 975.0, + 1403.0, + 975.0, + 1403.0, + 1011.0, + 296.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1008.0, + 1341.0, + 1008.0, + 1341.0, + 1041.0, + 322.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1967.0, + 1406.0, + 1967.0, + 1406.0, + 2011.0, + 294.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2000.0, + 690.0, + 2000.0, + 690.0, + 2035.0, + 323.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1858.0, + 1405.0, + 1858.0, + 1405.0, + 1892.0, + 297.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1887.0, + 1408.0, + 1887.0, + 1408.0, + 1925.0, + 324.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1919.0, + 747.0, + 1919.0, + 747.0, + 1952.0, + 326.0, + 1952.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 862, + 2088, + 862, + 2113, + 835, + 2113 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 296, + 1210, + 1402, + 1210, + 1402, + 1273, + 296, + 1273 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 293, + 1510, + 1404, + 1510, + 1404, + 1575, + 293, + 1575 + ], + "score": 0.785 + }, + { + "category_id": 1, + "poly": [ + 298, + 1291, + 1406, + 1291, + 1406, + 1383, + 298, + 1383 + ], + "score": 0.767 + }, + { + "category_id": 1, + "poly": [ + 293, + 970, + 1402, + 970, + 1402, + 1034, + 293, + 1034 + ], + "score": 0.762 + }, + { + "category_id": 1, + "poly": [ + 298, + 1401, + 1407, + 1401, + 1407, + 1493, + 298, + 1493 + ], + "score": 0.756 + }, + { + "category_id": 1, + "poly": [ + 299, + 580, + 1403, + 580, + 1403, + 644, + 299, + 644 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 296, + 1049, + 1401, + 1049, + 1401, + 1114, + 296, + 1114 + ], + "score": 0.747 + }, + { + "category_id": 1, + "poly": [ + 298, + 1129, + 1400, + 1129, + 1400, + 1194, + 298, + 1194 + ], + "score": 0.74 + }, + { + "category_id": 1, + "poly": [ + 297, + 1590, + 1407, + 1590, + 1407, + 1684, + 297, + 1684 + ], + "score": 0.726 + }, + { + "category_id": 1, + "poly": [ + 295, + 500, + 1403, + 500, + 1403, + 563, + 295, + 563 + ], + "score": 0.707 + }, + { + "category_id": 1, + "poly": [ + 297, + 660, + 1408, + 660, + 1408, + 784, + 297, + 784 + ], + "score": 0.702 + }, + { + "category_id": 1, + "poly": [ + 297, + 1780, + 1399, + 1780, + 1399, + 1845, + 297, + 1845 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 297, + 800, + 1408, + 800, + 1408, + 953, + 297, + 953 + ], + "score": 0.685 + }, + { + "category_id": 1, + "poly": [ + 302, + 1861, + 1399, + 1861, + 1399, + 1926, + 302, + 1926 + ], + "score": 0.668 + }, + { + "category_id": 1, + "poly": [ + 296, + 1701, + 1403, + 1701, + 1403, + 1765, + 296, + 1765 + ], + "score": 0.66 + }, + { + "category_id": 1, + "poly": [ + 299, + 1943, + 1405, + 1943, + 1405, + 2033, + 299, + 2033 + ], + "score": 0.626 + }, + { + "category_id": 1, + "poly": [ + 296, + 228, + 1406, + 228, + 1406, + 294, + 296, + 294 + ], + "score": 0.615 + }, + { + "category_id": 1, + "poly": [ + 298, + 390, + 1404, + 390, + 1404, + 481, + 298, + 481 + ], + "score": 0.608 + }, + { + "category_id": 1, + "poly": [ + 295, + 309, + 1404, + 309, + 1404, + 374, + 295, + 374 + ], + "score": 0.583 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1208.0, + 1404.0, + 1208.0, + 1404.0, + 1247.0, + 295.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1242.0, + 707.0, + 1242.0, + 707.0, + 1275.0, + 320.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1512.0, + 1406.0, + 1512.0, + 1406.0, + 1548.0, + 295.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1541.0, + 1251.0, + 1541.0, + 1251.0, + 1576.0, + 321.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1289.0, + 1409.0, + 1289.0, + 1409.0, + 1326.0, + 295.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1319.0, + 1406.0, + 1319.0, + 1406.0, + 1356.0, + 321.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1353.0, + 503.0, + 1353.0, + 503.0, + 1384.0, + 324.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 966.0, + 1405.0, + 966.0, + 1405.0, + 1008.0, + 292.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1001.0, + 674.0, + 1001.0, + 674.0, + 1035.0, + 324.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1401.0, + 1405.0, + 1401.0, + 1405.0, + 1435.0, + 296.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1429.0, + 1408.0, + 1429.0, + 1408.0, + 1467.0, + 320.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1458.0, + 398.0, + 1458.0, + 398.0, + 1495.0, + 320.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 579.0, + 1405.0, + 579.0, + 1405.0, + 618.0, + 295.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 612.0, + 891.0, + 612.0, + 891.0, + 644.0, + 322.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1048.0, + 1405.0, + 1048.0, + 1405.0, + 1087.0, + 294.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1082.0, + 1145.0, + 1082.0, + 1145.0, + 1115.0, + 323.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1129.0, + 1404.0, + 1129.0, + 1404.0, + 1167.0, + 294.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1160.0, + 1043.0, + 1160.0, + 1043.0, + 1195.0, + 323.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1591.0, + 1407.0, + 1591.0, + 1407.0, + 1629.0, + 296.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1659.0, + 321.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1654.0, + 1124.0, + 1654.0, + 1124.0, + 1684.0, + 323.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 498.0, + 1405.0, + 498.0, + 1405.0, + 537.0, + 294.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 529.0, + 1075.0, + 529.0, + 1075.0, + 565.0, + 323.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 659.0, + 1406.0, + 659.0, + 1406.0, + 695.0, + 294.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 689.0, + 1408.0, + 689.0, + 1408.0, + 725.0, + 320.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 718.0, + 1413.0, + 718.0, + 1413.0, + 757.0, + 319.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 751.0, + 720.0, + 751.0, + 720.0, + 784.0, + 323.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1779.0, + 1404.0, + 1779.0, + 1404.0, + 1818.0, + 295.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1813.0, + 1178.0, + 1813.0, + 1178.0, + 1845.0, + 322.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 797.0, + 1410.0, + 797.0, + 1410.0, + 836.0, + 294.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 832.0, + 1405.0, + 832.0, + 1405.0, + 865.0, + 325.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 861.0, + 1406.0, + 861.0, + 1406.0, + 894.0, + 323.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 886.0, + 1411.0, + 886.0, + 1411.0, + 927.0, + 319.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 921.0, + 753.0, + 921.0, + 753.0, + 953.0, + 321.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1856.0, + 1404.0, + 1856.0, + 1404.0, + 1901.0, + 296.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1891.0, + 801.0, + 1891.0, + 801.0, + 1926.0, + 322.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1700.0, + 1407.0, + 1700.0, + 1407.0, + 1738.0, + 294.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1732.0, + 1175.0, + 1732.0, + 1175.0, + 1768.0, + 324.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1408.0, + 1942.0, + 1408.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1970.0, + 1407.0, + 1970.0, + 1407.0, + 2008.0, + 323.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1997.0, + 399.0, + 1997.0, + 399.0, + 2037.0, + 319.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 1408.0, + 227.0, + 1408.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 258.0, + 1350.0, + 258.0, + 1350.0, + 295.0, + 323.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 388.0, + 1404.0, + 388.0, + 1404.0, + 425.0, + 295.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 419.0, + 1409.0, + 419.0, + 1409.0, + 455.0, + 320.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 449.0, + 452.0, + 449.0, + 452.0, + 483.0, + 321.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 304.0, + 1409.0, + 304.0, + 1409.0, + 351.0, + 293.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 341.0, + 1139.0, + 341.0, + 1139.0, + 374.0, + 324.0, + 374.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2112, + 835, + 2112 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1402, + 229, + 1402, + 323, + 299, + 323 + ], + "score": 0.761 + }, + { + "category_id": 1, + "poly": [ + 293, + 452, + 1403, + 452, + 1403, + 544, + 293, + 544 + ], + "score": 0.734 + }, + { + "category_id": 1, + "poly": [ + 296, + 341, + 1403, + 341, + 1403, + 433, + 296, + 433 + ], + "score": 0.731 + }, + { + "category_id": 1, + "poly": [ + 297, + 735, + 1401, + 735, + 1401, + 799, + 297, + 799 + ], + "score": 0.719 + }, + { + "category_id": 1, + "poly": [ + 296, + 563, + 1405, + 563, + 1405, + 716, + 296, + 716 + ], + "score": 0.382 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 229.0, + 1407.0, + 229.0, + 1407.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 260.0, + 1407.0, + 260.0, + 1407.0, + 298.0, + 323.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 291.0, + 1229.0, + 291.0, + 1229.0, + 325.0, + 324.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 451.0, + 1408.0, + 451.0, + 1408.0, + 488.0, + 295.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 482.0, + 1404.0, + 482.0, + 1404.0, + 516.0, + 322.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 512.0, + 619.0, + 512.0, + 619.0, + 546.0, + 321.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 341.0, + 1404.0, + 341.0, + 1404.0, + 375.0, + 295.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 373.0, + 1406.0, + 373.0, + 1406.0, + 406.0, + 323.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 398.0, + 465.0, + 398.0, + 465.0, + 435.0, + 322.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 771.0, + 293.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 764.0, + 1283.0, + 764.0, + 1283.0, + 800.0, + 324.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 564.0, + 1406.0, + 564.0, + 1406.0, + 601.0, + 295.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 592.0, + 1406.0, + 592.0, + 1406.0, + 634.0, + 320.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 620.0, + 1406.0, + 620.0, + 1406.0, + 665.0, + 319.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 653.0, + 1407.0, + 653.0, + 1407.0, + 690.0, + 322.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 683.0, + 1112.0, + 683.0, + 1112.0, + 720.0, + 322.0, + 720.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 668, + 1407, + 668, + 1407, + 911, + 297, + 911 + ], + "score": 0.979 + }, + { + "category_id": 8, + "poly": [ + 459, + 1257, + 1239, + 1257, + 1239, + 1418, + 459, + 1418 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 297, + 1542, + 1440, + 1542, + 1440, + 2039, + 297, + 2039 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 297, + 367, + 1404, + 367, + 1404, + 460, + 297, + 460 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 688, + 1088, + 1011, + 1088, + 1011, + 1165, + 688, + 1165 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 299, + 1448, + 1409, + 1448, + 1409, + 1512, + 299, + 1512 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 300, + 304, + 613, + 304, + 613, + 337, + 300, + 337 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 813, + 74, + 813, + 105, + 299, + 105 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 300, + 1024, + 1240, + 1024, + 1240, + 1059, + 300, + 1059 + ], + "score": 0.892 + }, + { + "category_id": 1, + "poly": [ + 296, + 1192, + 1096, + 1192, + 1096, + 1227, + 296, + 1227 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1357, + 1401, + 1357, + 1401, + 1388, + 1366, + 1388 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2113, + 835, + 2113 + ], + "score": 0.857 + }, + { + "category_id": 0, + "poly": [ + 301, + 962, + 639, + 962, + 639, + 994, + 301, + 994 + ], + "score": 0.845 + }, + { + "category_id": 0, + "poly": [ + 302, + 226, + 634, + 226, + 634, + 262, + 302, + 262 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 361, + 491, + 919, + 491, + 919, + 637, + 361, + 637 + ], + "score": 0.719 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1974, + 1402, + 1974, + 1402, + 2004, + 1365, + 2004 + ], + "score": 0.194 + }, + { + "category_id": 14, + "poly": [ + 457, + 1255, + 1240, + 1255, + 1240, + 1423, + 457, + 1423 + ], + "score": 0.95, + "latex": "\\begin{array} { l } { f ( x _ { t + 1 } ) \\leq f ( x _ { t } ) + \\langle \\nabla f ( x _ { t } ) , x _ { t + 1 } - x _ { t } \\rangle + \\displaystyle \\frac { L } { 2 } \\| x _ { t + 1 } - x _ { t } \\| ^ { 2 } } \\\\ { \\displaystyle = f ( x _ { t } ) + \\eta \\left. \\nabla f ( x _ { t } ) , \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } \\right. + \\displaystyle \\frac { \\eta ^ { 2 } L } { 2 } \\sum _ { i = 1 } ^ { d } \\frac { \\Delta _ { t , i } ^ { 2 } } { ( \\sqrt { v _ { t , i } } + \\tau ) ^ { 2 } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 686, + 1086, + 1011, + 1086, + 1011, + 1165, + 686, + 1165 + ], + "score": 0.94, + "latex": "x _ { t + 1 , i } = x _ { t , i } + \\eta \\frac { \\Delta _ { t , i } } { \\sqrt { v _ { t , i } } + \\tau } ," + }, + { + "category_id": 13, + "poly": [ + 372, + 1194, + 447, + 1194, + 447, + 1228, + 372, + 1228 + ], + "score": 0.92, + "latex": "i \\in [ d ]" + }, + { + "category_id": 13, + "poly": [ + 1310, + 1449, + 1404, + 1449, + 1404, + 1483, + 1310, + 1483 + ], + "score": 0.91, + "latex": "f ( x _ { t + 1 } )" + }, + { + "category_id": 14, + "poly": [ + 310, + 1539, + 1428, + 1539, + 1428, + 2044, + 310, + 2044 + ], + "score": 0.88, + "latex": "\\begin{array} { l } { \\displaystyle \\ddot { c } _ { i } [ f ( x _ { i + 1 } ) ] \\leq f ( x _ { i } ) + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\Psi _ { t - 1 } } + \\tau } \\right] \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } ^ { 2 } \\rho } { \\left( \\sqrt { \\nu _ { t - 1 } } + \\tau \\right) ^ { 2 } } \\right] } \\\\ { = f ( x _ { i } ) + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t + 1 } } + \\tau } - \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } + \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } \\\\ { \\displaystyle \\qquad + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i , j } ^ { 2 } } { \\left( \\sqrt { \\nu _ { t , j } } + \\tau \\right) ^ { 2 } } \\right] } \\\\ { = f ( x _ { i } ) + \\eta \\underbrace { \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } _ { \\displaystyle \\mathrm { ~ ( ~ 1 ~ ) ~ } } + \\eta \\left. \\nabla f ( x _ { i } ) , \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t } } + \\tau } - \\frac { \\Delta _ { i } } { \\sqrt { \\nu _ { t - 1 } } + \\tau } \\right] \\right. } \\\\ { \\displaystyle \\qquad + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { i = 1 } ^ { d } \\mathbb { E } _ { k } \\left[ \\frac { \\Delta _ { i , j } ^ { 2 } } { \\left( \\sqrt { \\nu _ { t - 1 } } + \\tau \\right) ^ { 2 } } \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1231, + 792, + 1263, + 792, + 1263, + 823, + 1231, + 823 + ], + "score": 0.88, + "latex": "\\sigma _ { g }" + }, + { + "category_id": 13, + "poly": [ + 667, + 1196, + 686, + 1196, + 686, + 1226, + 667, + 1226 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 716, + 1196, + 737, + 1196, + 737, + 1222, + 716, + 1222 + ], + "score": 0.83, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 646, + 850, + 674, + 850, + 674, + 876, + 646, + 876 + ], + "score": 0.76, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 634, + 1483, + 648, + 1483, + 648, + 1507, + 634, + 1507 + ], + "score": 0.64, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 297, + 880, + 320, + 880, + 320, + 906, + 297, + 906 + ], + "score": 0.63, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1188, + 1917, + 1213, + 1917, + 1213, + 1935, + 1188, + 1935 + ], + "score": 0.47, + "latex": "T _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 302.0, + 615.0, + 302.0, + 615.0, + 340.0, + 295.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 962.0, + 643.0, + 962.0, + 643.0, + 998.0, + 296.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 222.0, + 637.0, + 222.0, + 637.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 669.0, + 1407.0, + 669.0, + 1407.0, + 706.0, + 294.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 696.0, + 1405.0, + 696.0, + 1405.0, + 739.0, + 292.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 764.0, + 295.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 759.0, + 1405.0, + 759.0, + 1405.0, + 792.0, + 295.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 787.0, + 1230.0, + 787.0, + 1230.0, + 824.0, + 294.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 787.0, + 1408.0, + 787.0, + 1408.0, + 824.0, + 1264.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 816.0, + 1407.0, + 816.0, + 1407.0, + 852.0, + 292.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 847.0, + 645.0, + 847.0, + 645.0, + 885.0, + 294.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 847.0, + 1405.0, + 847.0, + 1405.0, + 885.0, + 675.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 877.0, + 952.0, + 877.0, + 952.0, + 913.0, + 321.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 366.0, + 1404.0, + 366.0, + 1404.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 395.0, + 1405.0, + 395.0, + 1405.0, + 436.0, + 292.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 428.0, + 716.0, + 428.0, + 716.0, + 464.0, + 294.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1443.0, + 1309.0, + 1443.0, + 1309.0, + 1488.0, + 293.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1480.0, + 633.0, + 1480.0, + 633.0, + 1516.0, + 296.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1480.0, + 927.0, + 1480.0, + 927.0, + 1516.0, + 649.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1020.0, + 1245.0, + 1020.0, + 1245.0, + 1066.0, + 292.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1188.0, + 371.0, + 1188.0, + 371.0, + 1234.0, + 293.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1188.0, + 666.0, + 1188.0, + 666.0, + 1234.0, + 448.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1188.0, + 715.0, + 1188.0, + 715.0, + 1234.0, + 687.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1188.0, + 1099.0, + 1188.0, + 1099.0, + 1234.0, + 738.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 488.0, + 872.0, + 488.0, + 872.0, + 532.0, + 359.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 544.0, + 620.0, + 544.0, + 620.0, + 588.0, + 358.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 600.0, + 921.0, + 600.0, + 921.0, + 643.0, + 359.0, + 643.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 438, + 1362, + 1259, + 1362, + 1259, + 1652, + 438, + 1652 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 294, + 1754, + 1093, + 1754, + 1093, + 2034, + 294, + 2034 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 329, + 1048, + 1373, + 1048, + 1373, + 1247, + 329, + 1247 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 296, + 279, + 1019, + 279, + 1019, + 562, + 296, + 562 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 297, + 962, + 1403, + 962, + 1403, + 1027, + 297, + 1027 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 329, + 643, + 1257, + 643, + 1257, + 943, + 329, + 943 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 294, + 1280, + 1402, + 1280, + 1402, + 1346, + 294, + 1346 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 289, + 580, + 1316, + 580, + 1316, + 626, + 289, + 626 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 294, + 228, + 829, + 228, + 829, + 263, + 294, + 263 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 814, + 74, + 814, + 106, + 298, + 106 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 294, + 1672, + 1403, + 1672, + 1403, + 1737, + 294, + 1737 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.86 + }, + { + "category_id": 9, + "poly": [ + 1364, + 1558, + 1402, + 1558, + 1402, + 1589, + 1364, + 1589 + ], + "score": 0.848 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1181, + 1401, + 1181, + 1401, + 1211, + 1365, + 1211 + ], + "score": 0.838 + }, + { + "category_id": 8, + "poly": [ + 324, + 643, + 1477, + 643, + 1477, + 943, + 324, + 943 + ], + "score": 0.477 + }, + { + "category_id": 14, + "poly": [ + 437, + 1362, + 1260, + 1362, + 1260, + 1646, + 437, + 1646 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { T _ { 1 } = \\bigg \\langle \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t - 1 } } + \\tau } \\right] \\bigg \\rangle } \\\\ & { \\quad = \\bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } - \\eta _ { l } K \\nabla f ( x _ { t } ) + \\eta _ { l } K \\nabla f ( x _ { t } ) \\right] \\bigg \\rangle } \\\\ & { \\quad = - \\eta _ { l } K \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\underbrace { \\left. \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) \\right] \\right. } _ { T _ { 2 } } . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 294, + 1756, + 1096, + 1756, + 1096, + 2037, + 294, + 2037 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { T _ { 3 } = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. } \\\\ & { \\quad = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ - \\displaystyle \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\eta _ { l } g _ { i , k } ^ { t } + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. } \\\\ & { \\quad = \\left. \\frac { \\nabla f ( \\boldsymbol { x } _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ - \\displaystyle \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\eta _ { l } \\nabla F _ { i } ( x _ { i , k } ^ { t } ) + \\eta _ { l } K \\nabla f ( \\boldsymbol { x } _ { t } ) \\right] \\right. . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 293, + 279, + 1019, + 279, + 1019, + 565, + 293, + 565 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { T _ { 2 } = \\bigg \\langle \\nabla f ( { x } _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { v _ { t - 1 } } + \\tau } \\right] \\bigg \\rangle } \\\\ & { \\quad = \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } [ \\nabla f ( { x } _ { t } ) ] _ { j } \\times \\left[ \\frac { \\Delta _ { t , j } } { \\sqrt { v _ { t , j } } + \\tau } - \\frac { \\Delta _ { t , j } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\right] } \\\\ & { \\quad = \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } [ \\nabla f ( { x } _ { t } ) ] _ { j } \\times \\Delta _ { t , j } \\times \\left[ \\frac { \\sqrt { v _ { t - 1 , j } } - \\sqrt { v _ { t , j } } } { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } \\right] } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 327, + 1043, + 1374, + 1043, + 1374, + 1249, + 327, + 1249 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } { T _ { 2 } \\le \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } \\eta _ { l } K G ^ { 2 } \\left[ \\displaystyle \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } \\right] } & { \\mathrm { ~ s i n c e ~ } [ \\nabla f ( x _ { t } ) ] _ { i } \\le G \\mathrm { ~ a n d ~ } \\Delta _ { t , i } \\le \\eta _ { l } K G } \\\\ { \\le \\mathbb { E } _ { t } \\displaystyle \\sum _ { j = 1 } ^ { d } \\displaystyle \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } \\left[ \\displaystyle \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] } & { \\mathrm { ~ s i n c e ~ } \\sqrt { v _ { t - 1 , j } } \\ge 0 . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 357, + 966, + 481, + 966, + 481, + 998, + 357, + 998 + ], + "score": 0.92, + "latex": "v _ { t - 1 , j } \\geq \\tau" + }, + { + "category_id": 14, + "poly": [ + 327, + 644, + 1425, + 644, + 1425, + 946, + 327, + 946 + ], + "score": 0.92, + "latex": "\\begin{array} { r l r } { { = \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } [ \\nabla f ( x _ { t } ) ] _ { j } \\times \\Delta _ { t , j } \\times [ \\overline { { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) } } ] } } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( \\sqrt { v _ { t , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) } } ] } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( v _ { t , j } + \\tau ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } } ] } & { \\mathrm { s i n c e } v _ { t - 1 , j } \\geq \\tau } \\\\ & { \\leq \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } \\vert \\nabla f ( x _ { t } ) \\vert _ { j } \\vert \\times \\vert \\Delta _ { t , j } \\vert \\times [ \\overline { { ( v _ { t , j } + \\tau ^ { 2 } ) ( \\sqrt { v _ { t - 1 , j } } + \\tau ) } } ] } & { \\mathrm { s i n c e } v _ { t - 1 , j } \\geq \\tau } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 860, + 1704, + 978, + 1704, + 978, + 1737, + 860, + 1737 + ], + "score": 0.91, + "latex": "- \\nabla f ( x _ { t } ) )" + }, + { + "category_id": 13, + "poly": [ + 547, + 966, + 646, + 966, + 646, + 996, + 547, + 996 + ], + "score": 0.91, + "latex": "v _ { - 1 } \\geq \\tau" + }, + { + "category_id": 13, + "poly": [ + 663, + 587, + 1207, + 587, + 1207, + 624, + 663, + 624 + ], + "score": 0.89, + "latex": "- \\Delta _ { t , j } ^ { 2 } = ( \\sqrt { v _ { t - 1 , j } } - \\sqrt { v _ { t , j } } ) ( \\sqrt { v _ { t - 1 , j } } + \\sqrt { v _ { t , j } } ) )" + }, + { + "category_id": 13, + "poly": [ + 1112, + 968, + 1155, + 968, + 1155, + 999, + 1112, + 999 + ], + "score": 0.89, + "latex": "v _ { t , j }" + }, + { + "category_id": 13, + "poly": [ + 448, + 585, + 631, + 585, + 631, + 620, + 448, + 620 + ], + "score": 0.89, + "latex": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 515, + 232, + 546, + 232, + 546, + 261, + 515, + 261 + ], + "score": 0.88, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 493, + 1707, + 527, + 1707, + 527, + 1734, + 493, + 1734 + ], + "score": 0.88, + "latex": "\\Delta _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1003, + 1284, + 1034, + 1284, + 1034, + 1313, + 1003, + 1313 + ], + "score": 0.87, + "latex": "T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 490, + 1676, + 521, + 1676, + 521, + 1704, + 490, + 1704 + ], + "score": 0.87, + "latex": "T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 941, + 1675, + 972, + 1675, + 972, + 1704, + 941, + 1704 + ], + "score": 0.87, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 420, + 1284, + 452, + 1284, + 452, + 1313, + 420, + 1313 + ], + "score": 0.86, + "latex": "T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1209, + 1078, + 1367, + 1078, + 1367, + 1113, + 1209, + 1113 + ], + "score": 0.78, + "latex": "\\Delta _ { t , i } \\leq \\eta _ { l } K G" + }, + { + "category_id": 13, + "poly": [ + 1331, + 967, + 1344, + 967, + 1344, + 991, + 1331, + 991 + ], + "score": 0.71, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1328, + 874, + 1468, + 874, + 1468, + 913, + 1328, + 913 + ], + "score": 0.53, + "latex": "v _ { t - 1 , j } \\geq \\tau ^ { 2 } ." + }, + { + "category_id": 13, + "poly": [ + 986, + 1077, + 1159, + 1077, + 1159, + 1113, + 986, + 1113 + ], + "score": 0.47, + "latex": "[ \\nabla f ( x _ { t } ) ] _ { i } \\leq G" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 961.0, + 356.0, + 961.0, + 356.0, + 1001.0, + 293.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 961.0, + 546.0, + 961.0, + 546.0, + 1001.0, + 482.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 961.0, + 1111.0, + 961.0, + 1111.0, + 1001.0, + 647.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 961.0, + 1330.0, + 961.0, + 1330.0, + 1001.0, + 1156.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 961.0, + 1404.0, + 961.0, + 1404.0, + 1001.0, + 1345.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1060.0, + 991.0, + 1060.0, + 1029.0, + 294.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1278.0, + 419.0, + 1278.0, + 419.0, + 1320.0, + 293.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1278.0, + 1002.0, + 1278.0, + 1002.0, + 1320.0, + 453.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1278.0, + 1405.0, + 1278.0, + 1405.0, + 1320.0, + 1035.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1310.0, + 701.0, + 1310.0, + 701.0, + 1349.0, + 294.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 578.0, + 447.0, + 578.0, + 447.0, + 629.0, + 291.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 578.0, + 662.0, + 578.0, + 662.0, + 629.0, + 632.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 578.0, + 1323.0, + 578.0, + 1323.0, + 629.0, + 1208.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 224.0, + 514.0, + 224.0, + 514.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 224.0, + 832.0, + 224.0, + 832.0, + 269.0, + 547.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1671.0, + 489.0, + 1671.0, + 489.0, + 1711.0, + 294.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1671.0, + 940.0, + 1671.0, + 940.0, + 1711.0, + 522.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 1671.0, + 1405.0, + 1671.0, + 1405.0, + 1711.0, + 973.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1702.0, + 492.0, + 1702.0, + 492.0, + 1738.0, + 296.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1702.0, + 859.0, + 1702.0, + 859.0, + 1738.0, + 528.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1702.0, + 991.0, + 1702.0, + 991.0, + 1738.0, + 979.0, + 1738.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 221, + 1408, + 221, + 1408, + 331, + 295, + 331 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 296, + 957, + 1406, + 957, + 1406, + 1058, + 296, + 1058 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 306, + 1357, + 1398, + 1357, + 1398, + 1638, + 306, + 1638 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 294, + 1709, + 1480, + 1709, + 1480, + 2043, + 294, + 2043 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 294, + 336, + 1444, + 336, + 1444, + 653, + 294, + 653 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 389, + 834, + 1308, + 834, + 1308, + 925, + 389, + 925 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 297, + 1061, + 1558, + 1061, + 1558, + 1280, + 297, + 1280 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 814, + 74, + 814, + 106, + 298, + 106 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 298, + 686, + 1406, + 686, + 1406, + 750, + 298, + 750 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 297, + 1639, + 1401, + 1639, + 1401, + 1702, + 297, + 1702 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2113, + 835, + 2113 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 292, + 766, + 1403, + 766, + 1403, + 830, + 292, + 830 + ], + "score": 0.847 + }, + { + "category_id": 9, + "poly": [ + 1365, + 652, + 1402, + 652, + 1402, + 681, + 1365, + 681 + ], + "score": 0.836 + }, + { + "category_id": 9, + "poly": [ + 1364, + 862, + 1401, + 862, + 1401, + 894, + 1364, + 894 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 299, + 1283, + 1464, + 1283, + 1464, + 1349, + 299, + 1349 + ], + "score": 0.678 + }, + { + "category_id": 13, + "poly": [ + 830, + 959, + 938, + 959, + 938, + 997, + 830, + 997 + ], + "score": 0.95, + "latex": "x _ { i , 0 } ^ { t } = x _ { t }" + }, + { + "category_id": 14, + "poly": [ + 311, + 1352, + 1399, + 1352, + 1399, + 1643, + 311, + 1643 + ], + "score": 0.94, + "latex": "\\begin{array} { r l r } { { \\mathbb { E } \\Vert x _ { i , k } ^ { t } - x _ { t } \\Vert ^ { 2 } \\leq ( 1 + \\frac { 1 } { 2 K - 1 } ) \\mathbb { E } \\Vert x _ { i , k - 1 } ^ { t } - x _ { t } \\Vert ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\Vert L ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \\Vert ^ { 2 } } } \\\\ & { } & { + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\Vert ^ { 2 } ] + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } \\nabla f ( x _ { t } ) ) ) \\Vert ^ { 2 } ] } \\\\ & { } & { = ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\mathbb { E } \\Vert ( x _ { i , k - 1 } ^ { t } - x _ { t } ) \\Vert ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } } \\\\ & { } & { + 6 K \\mathbb { E } [ \\Vert \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\Vert ^ { 2 } ] + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\Vert \\nabla f ( x _ { t } ) ) ) \\Vert ^ { 2 } ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1015, + 959, + 1101, + 959, + 1101, + 993, + 1015, + 993 + ], + "score": 0.93, + "latex": "i \\in [ m ]" + }, + { + "category_id": 13, + "poly": [ + 1093, + 1284, + 1237, + 1284, + 1237, + 1323, + 1093, + 1323 + ], + "score": 0.93, + "latex": "\\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } )" + }, + { + "category_id": 13, + "poly": [ + 724, + 768, + 837, + 768, + 837, + 808, + 724, + 808 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\eta _ { l } \\le \\frac { 1 } { 8 L K } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 390, + 827, + 1308, + 827, + 1308, + 928, + 390, + 928 + ], + "score": 0.93, + "latex": "\\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 2 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\big ] ." + }, + { + "category_id": 13, + "poly": [ + 709, + 1284, + 781, + 1284, + 781, + 1322, + 709, + 1322 + ], + "score": 0.93, + "latex": "g _ { k - 1 , i } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 995, + 225, + 1040, + 225, + 1040, + 265, + 995, + 265 + ], + "score": 0.92, + "latex": "g _ { i , k } ^ { t }" + }, + { + "category_id": 14, + "poly": [ + 310, + 332, + 1427, + 332, + 1427, + 658, + 310, + 658 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } { r _ { 3 } \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\sum _ { j } ^ { d } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } } { 2 K } \\mathbb { E } _ { t } \\left[ \\left\\| \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\frac { \\nabla F _ { i } ( x _ { i , k } ^ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } - \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\frac { \\nabla F _ { i } ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } \\right\\| ^ { 2 } \\right] } & { } \\\\ { \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\left[ \\nabla f ( x _ { t } ) \\right] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } } { 2 m } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\left\\| \\frac { \\nabla F _ { i } ( x _ { i , k } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) } { \\sqrt { v _ { t - 1 } } + \\tau } \\right\\| ^ { 2 } \\right] } & { } \\\\ { \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { \\displaystyle \\left[ \\nabla f ( x _ { t } ) \\right] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right] } & { \\mathrm { u s i n g ~ A s s u m p t i o n ~ 1 ~ a n d ~ } v _ { t - 1 } \\geq 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1216, + 716, + 1263, + 716, + 1263, + 755, + 1216, + 755 + ], + "score": 0.92, + "latex": "\\dot { \\boldsymbol { x } } _ { i , k } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 608, + 223, + 941, + 223, + 941, + 262, + 608, + 262 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\nabla f ( x _ { t } ) = \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\nabla F _ { i } ( x _ { t } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1257, + 993, + 1351, + 993, + 1351, + 1028, + 1257, + 1028 + ], + "score": 0.92, + "latex": "i \\in [ m ]" + }, + { + "category_id": 14, + "poly": [ + 310, + 1702, + 1428, + 1702, + 1428, + 2047, + 310, + 2047 + ], + "score": 0.92, + "latex": "\\begin{array} { r l r } { { \\frac { 1 } { n } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } } } \\\\ & { } & { + \\frac { 6 K } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } [ \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\| ^ { 2 } ] + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) } \\\\ & { } & { \\leq ( 1 + \\frac { 1 } { 2 K - 1 } + 6 K \\eta _ { l } ^ { 2 } L ^ { 2 } ) \\frac { 1 } { m } \\displaystyle \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K o ( \\eta _ { l } ^ { 2 } ) ) } \\\\ & { } & { + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) \\| ^ { 2 } ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1349, + 589, + 1454, + 589, + 1454, + 620, + 1349, + 620 + ], + "score": 0.91, + "latex": "v _ { t - 1 } \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 427, + 261, + 473, + 261, + 473, + 300, + 427, + 300 + ], + "score": 0.91, + "latex": "\\boldsymbol { x } _ { i , k } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1024, + 390, + 1024, + 390, + 1058, + 298, + 1058 + ], + "score": 0.91, + "latex": "k \\in [ K ]" + }, + { + "category_id": 13, + "poly": [ + 1177, + 770, + 1401, + 770, + 1401, + 805, + 1177, + 805 + ], + "score": 0.9, + "latex": "k \\in \\{ 0 , \\cdots , K - 1 \\}" + }, + { + "category_id": 14, + "poly": [ + 310, + 1053, + 1428, + 1053, + 1428, + 1284, + 310, + 1284 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } { { \\tilde { \\mathbf { \\mu } } _ { i , k , - } ^ { t } - x _ { t } \\| ^ { 2 } = \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } - \\eta _ { l } g _ { i , k - 1 } ^ { t } \\big \\| ^ { 2 } } } \\\\ & { \\leq \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } - \\eta _ { l } ( g _ { i , k - 1 } ^ { t } - \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) + \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) + \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) + \\nabla F _ { i } ( x _ { t } ) - \\nabla F ( x _ { t } ) \\big \\| ^ { 2 } } \\\\ & { \\leq ( 1 + \\frac { 1 } { 2 K - 1 } ) \\mathbb { E } \\big \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\big \\| ^ { 2 } + \\mathbb { E } \\big \\| \\eta _ { l } ( g _ { i , k - 1 } ^ { t } - \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) ) \\big \\| ^ { 2 } } \\\\ & { \\qquad + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { i , k - 1 } ^ { t } ) - \\nabla F _ { i } ( x _ { t } ) ) \\big \\| ^ { 2 } ] + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } ( \\nabla F _ { i } ( x _ { t } ) - \\nabla f ( x _ { t } ) ) \\big \\| ^ { 2 } \\big ] + 6 K \\mathbb { E } [ \\big \\| \\eta _ { l } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 507, + 994, + 584, + 994, + 584, + 1025, + 507, + 1025 + ], + "score": 0.9, + "latex": "k \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 694, + 960, + 764, + 960, + 764, + 988, + 694, + 988 + ], + "score": 0.9, + "latex": "k = 1" + }, + { + "category_id": 13, + "poly": [ + 692, + 264, + 724, + 264, + 724, + 294, + 692, + 294 + ], + "score": 0.88, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 295, + 297, + 433, + 297, + 433, + 333, + 295, + 333 + ], + "score": 0.87, + "latex": "( a ^ { 2 } + b ^ { 2 } ) / 2" + }, + { + "category_id": 13, + "poly": [ + 1327, + 721, + 1356, + 721, + 1356, + 748, + 1327, + 748 + ], + "score": 0.84, + "latex": "x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1339, + 262, + 1405, + 262, + 1405, + 294, + 1339, + 294 + ], + "score": 0.8, + "latex": "a b \\leq" + }, + { + "category_id": 13, + "poly": [ + 1165, + 690, + 1186, + 690, + 1186, + 715, + 1165, + 715 + ], + "score": 0.79, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 1265, + 1643, + 1277, + 1643, + 1277, + 1667, + 1265, + 1667 + ], + "score": 0.77, + "latex": "i" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 210.0, + 607.0, + 210.0, + 607.0, + 276.0, + 287.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 210.0, + 994.0, + 210.0, + 994.0, + 276.0, + 942.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 210.0, + 1415.0, + 210.0, + 1415.0, + 276.0, + 1041.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 250.0, + 426.0, + 250.0, + 426.0, + 307.0, + 288.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 250.0, + 691.0, + 250.0, + 691.0, + 307.0, + 474.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 250.0, + 1338.0, + 250.0, + 1338.0, + 307.0, + 725.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1406.0, + 250.0, + 1411.0, + 250.0, + 1411.0, + 307.0, + 1406.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 294.0, + 294.0, + 294.0, + 294.0, + 336.0, + 290.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 294.0, + 457.0, + 294.0, + 457.0, + 336.0, + 434.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 952.0, + 693.0, + 952.0, + 693.0, + 1004.0, + 290.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 952.0, + 829.0, + 952.0, + 829.0, + 1004.0, + 765.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 952.0, + 1014.0, + 952.0, + 1014.0, + 1004.0, + 939.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 952.0, + 1409.0, + 952.0, + 1409.0, + 1004.0, + 1102.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 994.0, + 506.0, + 994.0, + 506.0, + 1029.0, + 294.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 994.0, + 1256.0, + 994.0, + 1256.0, + 1029.0, + 585.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 994.0, + 1404.0, + 994.0, + 1404.0, + 1029.0, + 1352.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1017.0, + 297.0, + 1017.0, + 297.0, + 1063.0, + 291.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1017.0, + 407.0, + 1017.0, + 407.0, + 1063.0, + 391.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 686.0, + 1164.0, + 686.0, + 1164.0, + 722.0, + 296.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 686.0, + 1406.0, + 686.0, + 1406.0, + 722.0, + 1187.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 713.0, + 1215.0, + 713.0, + 1215.0, + 756.0, + 293.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 713.0, + 1326.0, + 713.0, + 1326.0, + 756.0, + 1264.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1357.0, + 713.0, + 1370.0, + 713.0, + 1370.0, + 756.0, + 1357.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1635.0, + 1264.0, + 1635.0, + 1264.0, + 1676.0, + 293.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1635.0, + 1405.0, + 1635.0, + 1405.0, + 1676.0, + 1278.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 461.0, + 1665.0, + 461.0, + 1706.0, + 294.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 762.0, + 723.0, + 762.0, + 723.0, + 818.0, + 289.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 762.0, + 1176.0, + 762.0, + 1176.0, + 818.0, + 838.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 762.0, + 1406.0, + 762.0, + 1406.0, + 818.0, + 1402.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 803.0, + 333.0, + 803.0, + 333.0, + 836.0, + 291.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1280.0, + 708.0, + 1280.0, + 708.0, + 1324.0, + 292.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1280.0, + 1092.0, + 1280.0, + 1092.0, + 1324.0, + 782.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1280.0, + 1411.0, + 1280.0, + 1411.0, + 1324.0, + 1238.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1315.0, + 974.0, + 1315.0, + 974.0, + 1354.0, + 293.0, + 1354.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 366, + 1831, + 1338, + 1831, + 1338, + 2030, + 366, + 2030 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 330, + 276, + 1369, + 276, + 1369, + 413, + 330, + 413 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 333, + 1536, + 1359, + 1536, + 1359, + 1743, + 333, + 1743 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 297, + 460, + 1404, + 460, + 1404, + 789, + 297, + 789 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 296, + 940, + 1400, + 940, + 1400, + 1257, + 296, + 1257 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 477, + 1345, + 1223, + 1345, + 1223, + 1439, + 477, + 1439 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 1462, + 1403, + 1462, + 1403, + 1528, + 296, + 1528 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 298, + 1269, + 1402, + 1269, + 1402, + 1336, + 298, + 1336 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 298, + 796, + 1404, + 796, + 1404, + 865, + 298, + 865 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 840, + 229, + 840, + 263, + 298, + 263 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 296, + 1749, + 1171, + 1749, + 1171, + 1785, + 296, + 1785 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 292, + 896, + 1383, + 896, + 1383, + 931, + 292, + 931 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 297, + 418, + 847, + 418, + 847, + 452, + 297, + 452 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.918 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1375, + 1401, + 1375, + 1401, + 1406, + 1365, + 1406 + ], + "score": 0.871 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1964, + 1402, + 1964, + 1402, + 1995, + 1350, + 1995 + ], + "score": 0.824 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2114, + 835, + 2114 + ], + "score": 0.795 + }, + { + "category_id": 9, + "poly": [ + 1364, + 1677, + 1402, + 1677, + 1402, + 1708, + 1364, + 1708 + ], + "score": 0.731 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2114, + 835, + 2114 + ], + "score": 0.151 + }, + { + "category_id": 14, + "poly": [ + 478, + 1341, + 1221, + 1341, + 1221, + 1441, + 478, + 1441 + ], + "score": 0.94, + "latex": "T _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )" + }, + { + "category_id": 14, + "poly": [ + 332, + 1532, + 1366, + 1532, + 1366, + 1746, + 332, + 1746 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { \\mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \\le f ( x _ { t } ) + \\eta \\times \\left[ - \\frac { \\eta _ { l } K } { 4 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad + \\eta \\times \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\displaystyle \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] + \\frac { \\eta ^ { 2 } } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } L \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\displaystyle \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 362, + 1828, + 1340, + 1828, + 1340, + 2033, + 362, + 2033 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\leq \\frac { f ( x _ { 0 } ) - \\mathbb { E } [ f ( x _ { T } ) ] } { \\eta } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\\\ { + \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } \\left( \\frac { \\eta _ { l } K G ^ { 2 } } { \\tau } + \\frac { \\eta L } { 2 } \\right) \\times \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1145, + 796, + 1359, + 796, + 1359, + 837, + 1145, + 837 + ], + "score": 0.93, + "latex": "( 1 + \\frac { 1 } { K - 1 } ) ^ { K } \\leq 5" + }, + { + "category_id": 14, + "poly": [ + 331, + 271, + 1369, + 271, + 1369, + 417, + 331, + 417 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } { { \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq ( 1 + \\frac { 1 } { K - 1 } ) \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k - 1 } ^ { t } - x _ { t } \\| ^ { 2 } + \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { } & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\Big ] } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 310, + 458, + 1407, + 458, + 1407, + 791, + 310, + 791 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\displaystyle \\frac { 1 } { n } \\sum _ { i = 1 } ^ { m } \\mathbb { E } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\leq \\sum _ { p = 0 } ^ { k - 1 } \\left( 1 + \\frac { 1 } { K - 1 } \\right) ^ { p } \\left[ \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\leq \\left( K - 1 \\right) \\times \\left[ \\left( 1 + \\frac { 1 } { K - 1 } \\right) ^ { K } - 1 \\right] \\times \\left[ \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 6 K \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\leq \\left[ 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 605, + 1268, + 852, + 1268, + 852, + 1308, + 605, + 1308 + ], + "score": 0.92, + "latex": "\\sqrt { v _ { t - 1 , j } } \\leq \\eta _ { l } K G \\sqrt { T }" + }, + { + "category_id": 14, + "poly": [ + 291, + 937, + 1390, + 937, + 1390, + 1261, + 291, + 1261 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { T _ { 3 } \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\displaystyle \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\sum _ { j = 1 } ^ { d } ( [ x _ { i , k } ^ { t } ] _ { j } - [ x _ { t } ] _ { j } ) ^ { 2 } \\right] } \\\\ & { \\quad \\leq \\displaystyle \\frac { \\eta _ { l } K } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } K L ^ { 2 } } { 2 \\tau } \\left[ 5 K \\eta _ { l } ^ { 2 } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + 3 0 K ^ { 2 } \\eta _ { l } ^ { 2 } \\mathbb { E } [ \\| \\nabla f ( x _ { t } ) ) ) \\| ^ { 2 } \\right] } \\\\ & { \\quad \\leq \\displaystyle \\frac { 3 \\eta _ { l } K } { 4 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1041, + 1466, + 1073, + 1466, + 1073, + 1495, + 1041, + 1495 + ], + "score": 0.88, + "latex": "T _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1070, + 900, + 1101, + 900, + 1101, + 929, + 1070, + 929 + ], + "score": 0.88, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 911, + 1752, + 976, + 1752, + 976, + 1779, + 911, + 1779 + ], + "score": 0.88, + "latex": "t = 0" + }, + { + "category_id": 13, + "poly": [ + 1007, + 1751, + 1078, + 1751, + 1078, + 1780, + 1007, + 1780 + ], + "score": 0.88, + "latex": "T - 1" + }, + { + "category_id": 13, + "poly": [ + 1370, + 1470, + 1400, + 1470, + 1400, + 1495, + 1370, + 1495 + ], + "score": 0.88, + "latex": "T _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 833, + 376, + 833, + 376, + 863, + 298, + 863 + ], + "score": 0.86, + "latex": "K > 1" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2123.0, + 830.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1457.0, + 1040.0, + 1457.0, + 1040.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1457.0, + 1369.0, + 1457.0, + 1369.0, + 1503.0, + 1074.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1457.0, + 1404.0, + 1457.0, + 1404.0, + 1503.0, + 1401.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1494.0, + 596.0, + 1494.0, + 596.0, + 1530.0, + 294.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1266.0, + 604.0, + 1266.0, + 604.0, + 1310.0, + 292.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1266.0, + 1407.0, + 1266.0, + 1407.0, + 1310.0, + 853.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 621.0, + 1297.0, + 621.0, + 1339.0, + 293.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 792.0, + 1144.0, + 792.0, + 1144.0, + 841.0, + 291.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 792.0, + 1410.0, + 792.0, + 1410.0, + 841.0, + 1360.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 826.0, + 297.0, + 826.0, + 297.0, + 869.0, + 292.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 826.0, + 392.0, + 826.0, + 392.0, + 869.0, + 377.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1381.0, + 844.0, + 1397.0, + 844.0, + 1397.0, + 854.0, + 1381.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 223.0, + 843.0, + 223.0, + 843.0, + 271.0, + 293.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1748.0, + 910.0, + 1748.0, + 910.0, + 1788.0, + 296.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 1748.0, + 1006.0, + 1748.0, + 1006.0, + 1788.0, + 977.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1748.0, + 1170.0, + 1748.0, + 1170.0, + 1788.0, + 1079.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 895.0, + 1069.0, + 895.0, + 1069.0, + 935.0, + 297.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 895.0, + 1385.0, + 895.0, + 1385.0, + 935.0, + 1102.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 412.0, + 848.0, + 412.0, + 848.0, + 459.0, + 295.0, + 459.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 516, + 856, + 1179, + 856, + 1179, + 1151, + 516, + 1151 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 296, + 366, + 1500, + 366, + 1500, + 567, + 296, + 567 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 295, + 1216, + 1628, + 1216, + 1628, + 1951, + 295, + 1951 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 294, + 227, + 1404, + 227, + 1404, + 295, + 294, + 295 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1406, + 1971, + 1406, + 2035, + 299, + 2035 + ], + "score": 0.932 + }, + { + "category_id": 8, + "poly": [ + 296, + 691, + 1425, + 691, + 1425, + 789, + 296, + 789 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 295, + 639, + 1004, + 639, + 1004, + 674, + 295, + 674 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 295, + 805, + 970, + 805, + 970, + 839, + 295, + 839 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 295, + 1165, + 1099, + 1165, + 1099, + 1199, + 295, + 1199 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 294, + 313, + 1184, + 313, + 1184, + 350, + 294, + 350 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2113, + 835, + 2113 + ], + "score": 0.864 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1086, + 1402, + 1086, + 1402, + 1119, + 1350, + 1119 + ], + "score": 0.772 + }, + { + "category_id": 14, + "poly": [ + 310, + 1214, + 1428, + 1214, + 1428, + 1965, + 310, + 1965 + ], + "score": 0.96, + "latex": "\\begin{array} { r l } & { \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda - 1 } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( - \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } - \\lambda _ { i } ^ { \\lambda } \\sum _ { j = 1 } ^ { \\infty } - \\lambda _ { i } ^ { \\lambda } \\nabla F _ { i } ( \\xi _ { i } , \\xi _ { i } ) \\Bigg ) \\Bigg | ^ { 2 } } \\\\ & { - 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } \\sum _ { j = 1 } ^ { \\infty } \\lambda _ { i } ^ { \\lambda } \\sum _ { j = 1 } ^ { \\lambda } - \\lambda _ { i } ^ { \\lambda } \\nabla F _ { i } ( \\xi _ { i } , \\xi _ { i } ) + 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\lambda _ { i } ^ { \\lambda } \\xi _ { i } ^ { \\lambda } \\xi _ { i } ^ { \\lambda } \\xi _ { j } - 3 \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\Bigg ) - } \\\\ & - 2 \\mathbb { E } _ { \\frac { d } { d } } ^ { \\lambda } \\mathbb { E } ^ { \\lambda - 1 } \\Bigg [ \\Bigg | \\frac { 1 } { \\tau } \\Bigg ( \\frac { 1 } { m } \\sum _ { i = 1 } ^ { \\infty } \\lambda _ { i } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb { E } ^ { \\lambda } \\mathbb E \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 516, + 855, + 1183, + 855, + 1183, + 1153, + 516, + 1153 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { t , j } ^ { 2 } + \\tau ^ { 2 } } \\leq \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\tau ^ { 2 } } } \\\\ & { \\leq \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\left\\| \\frac { \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) - \\eta _ { l } K \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } } \\\\ & { \\leq 2 \\mathbb { E } \\displaystyle \\sum _ { t = 0 } ^ { T - 1 } \\left[ \\left\\| \\frac { \\Delta _ { t } + \\eta _ { l } K \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } + \\eta _ { l } ^ { 2 } K ^ { 2 } \\left\\| \\frac { \\nabla f ( x _ { t } ) } { \\tau } \\right\\| ^ { 2 } \\right] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1082, + 1971, + 1279, + 1971, + 1279, + 2007, + 1082, + 2007 + ], + "score": 0.93, + "latex": "a b \\le ( a ^ { 2 } + b ^ { 2 } ) / 2" + }, + { + "category_id": 14, + "poly": [ + 310, + 689, + 1422, + 689, + 1422, + 791, + 310, + 791 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\stackrel { T \\to 1 } { \\sum } \\displaystyle \\sum _ { j = 0 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\leq d + \\mathbb { E } \\displaystyle \\sum _ { j = 1 } ^ { d } \\log \\left( 1 + \\frac { \\sum _ { l = 0 } ^ { T - 1 } \\Delta _ { l , j } ^ { 2 } } { \\tau ^ { 2 } } \\right) \\leq d + \\displaystyle \\sum _ { j = 1 } ^ { d } \\log \\left( 1 + \\frac { \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 310, + 365, + 1428, + 365, + 1428, + 571, + 310, + 571 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { \\underset { t = 0 } { \\overset { T \\mathrm { - } 1 } { \\sum } } \\underset { j = 1 } { \\overset { d } { \\sum } } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sum _ { l = 0 } ^ { t } \\Delta _ { l , j } ^ { 2 } + \\tau ^ { 2 } } \\leq \\Bigg [ \\operatorname* { m i n } \\left\\{ d + \\underset { j = 1 } { \\overset { d } { \\sum } } \\log \\left( 1 + \\frac { \\eta _ { l } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) \\right. } \\\\ & { \\qquad + \\left. \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\underset { j = 1 } { \\overset { d } { \\sum } } \\sigma _ { l , j } ^ { 2 } + 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\mathbb { E } \\underset { j = 1 } { \\overset { d } { \\sum } } \\frac { ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } { \\tau ^ { 2 } } + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 2 } L ^ { 2 } } { \\tau ^ { 2 } } \\underset { t = 0 } { \\overset { T \\mathrm { - } 1 } { \\sum } } \\mathbb { E } \\left[ \\left\\| \\nabla f ( x _ { t } ) \\right\\| \\right. \\right. } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 224.0, + 1405.0, + 224.0, + 1405.0, + 270.0, + 291.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 257.0, + 375.0, + 257.0, + 375.0, + 296.0, + 292.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1081.0, + 1971.0, + 1081.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 1280.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1998.0, + 398.0, + 1998.0, + 398.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 2008.0, + 1401.0, + 2008.0, + 1401.0, + 2029.0, + 1377.0, + 2029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 637.0, + 1007.0, + 637.0, + 1007.0, + 680.0, + 294.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 803.0, + 972.0, + 803.0, + 972.0, + 844.0, + 294.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1163.0, + 1102.0, + 1163.0, + 1102.0, + 1202.0, + 293.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 311.0, + 1189.0, + 311.0, + 1189.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1020, + 1405, + 1020, + 1405, + 1143, + 296, + 1143 + ], + "score": 0.977 + }, + { + "category_id": 8, + "poly": [ + 298, + 277, + 1447, + 277, + 1447, + 672, + 298, + 672 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 297, + 1155, + 1612, + 1155, + 1612, + 1952, + 297, + 1952 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 297, + 874, + 1404, + 874, + 1404, + 937, + 297, + 937 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 294, + 1971, + 1402, + 1971, + 1402, + 2036, + 294, + 2036 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 298, + 728, + 592, + 728, + 592, + 760, + 298, + 760 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 943, + 228, + 943, + 263, + 297, + 263 + ], + "score": 0.921 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 814, + 74, + 814, + 106, + 298, + 106 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 300, + 969, + 688, + 969, + 688, + 1001, + 300, + 1001 + ], + "score": 0.911 + }, + { + "category_id": 8, + "poly": [ + 346, + 771, + 1352, + 771, + 1352, + 864, + 346, + 864 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2113, + 835, + 2113 + ], + "score": 0.858 + }, + { + "category_id": 9, + "poly": [ + 1351, + 670, + 1401, + 670, + 1401, + 700, + 1351, + 700 + ], + "score": 0.847 + }, + { + "category_id": 14, + "poly": [ + 310, + 1155, + 1428, + 1155, + 1428, + 1943, + 310, + 1943 + ], + "score": 0.96, + "latex": "\\begin{array} { r l } { \\sum _ { j = 1 } ^ { N } \\Bigg [ \\Bigg | \\int _ { - 1 } ^ { 1 } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 0 } ^ { 1 } \\sum _ { i = 0 } ^ { N } \\sum _ { j = 1 } ^ { N } \\Bigg ) \\Bigg | ^ { 2 } \\Bigg ] ^ { 2 } } & { \\mathrm { d } \\Gamma } \\\\ { = } & { \\mathcal { D } _ { j } ^ { \\mathrm { R } } \\Bigg [ \\Bigg | \\int _ { - 1 } ^ { 1 } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 0 } ^ { 1 } \\sum _ { i = 0 } ^ { N } \\sum _ { j = 1 } ^ { N } \\delta ( \\sum _ { i = 0 } ^ { N } \\epsilon _ { j i } ) + \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } \\Big ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\bigg ) \\Bigg | ^ { 2 } } \\\\ & { \\leq \\epsilon _ { j } ^ { \\mathrm { R } } \\Bigg ( \\int _ { - 1 } ^ { 1 } \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) ^ { 2 } \\Bigg [ \\Bigg | \\int _ { 0 } ^ { 1 } \\sum _ { j = 1 } ^ { N } \\delta ( \\sum _ { j = 1 } ^ { N } \\epsilon _ { j i } ) - \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) - \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\Bigg | ^ { 2 } \\Bigg ] ^ { 2 } + \\Bigg | \\int _ { - 1 } ^ { 1 } \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) - \\mathrm { v } _ { \\mathrm { R } } ^ { \\mathrm { R } } \\bigg ( \\epsilon _ { j i } ^ { \\mathrm { R } } \\bigg ) \\Bigg | ^ { 2 } } \\\\ & \\qquad + \\mathrm { d } \\Omega _ { j } ^ { \\mathrm { R } } \\Bigg | ^ { 2 } \\Bigg ] + \\Bigg | \\int _ { 0 } ^ { 1 } \\int _ { - 1 } ^ { 1 } \\mathrm { d } \\Omega _ { i } ^ { \\mathrm { R } } \\Bigg ( \\int _ { 0 } ^ { 1 } \\sum _ { j = 1 } ^ { N } \\mathrm { d } \\ \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 342, + 767, + 1350, + 767, + 1350, + 867, + 342, + 867 + ], + "score": 0.94, + "latex": "\\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla \\mathbb { E } f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { v _ { t - 1 , j } } + \\tau } \\geq \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\eta _ { t } K G \\sqrt { T } + \\tau } \\geq \\frac { T } { \\eta _ { t } K G \\sqrt { T } + \\tau } \\operatorname* { m i n } _ { 0 \\leq t \\leq T } \\mathbb { E } \\| \\nabla f ( x _ { t } ) \\| ^ { 2 } ." + }, + { + "category_id": 14, + "poly": [ + 310, + 271, + 1428, + 271, + 1428, + 677, + 310, + 677 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\frac { \\displaystyle { n K } } { \\displaystyle { 4 } } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\pi _ { t - 1 , j } } + \\tau } } \\\\ & { \\leq \\frac { f \\left( x _ { 0 } \\right) - \\mathbb { E } \\left[ f ( x _ { T } ) \\right] } { \\eta } + \\frac { 5 \\eta _ { t } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } \\\\ & { \\qquad + \\displaystyle { \\sum _ { j = 1 } ^ { d } \\left( \\frac { \\eta _ { t } K G ^ { 2 } } { \\tau } + \\frac { \\eta L } { 2 } \\right) \\times \\left[ \\operatorname* { m i n } \\left\\{ d + d \\log \\left( 1 + \\frac { \\eta _ { t } ^ { 2 } K ^ { 2 } G ^ { 2 } T } { \\tau ^ { 2 } } \\right) + \\right. \\right.} } \\\\ & { \\quad \\quad \\left. \\frac { 4 \\eta _ { t } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + 2 0 \\eta _ { t } ^ { 4 } K ^ { 3 } L ^ { 2 } T \\sum _ { j = 1 } ^ { d } \\frac { ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } { \\tau ^ { 2 } } + \\frac { 4 0 \\eta _ { t } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } \\left[ \\left\\| \\nabla f ( x _ { t } ) \\right\\| ^ { 2 } \\right] \\right\\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 295, + 2001, + 634, + 2001, + 634, + 2037, + 295, + 2037 + ], + "score": 0.91, + "latex": "( a + b + c ) ^ { 2 } \\leq 3 { \\dot { ( } } a ^ { 2 } + b ^ { 2 } + c ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 347, + 1083, + 432, + 1083, + 432, + 1114, + 347, + 1114 + ], + "score": 0.89, + "latex": "| S | = s" + }, + { + "category_id": 13, + "poly": [ + 329, + 1115, + 369, + 1115, + 369, + 1145, + 329, + 1145 + ], + "score": 0.88, + "latex": "[ m ]" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1051, + 1243, + 1051, + 1243, + 1079, + 1213, + 1079 + ], + "score": 0.87, + "latex": "t ^ { \\mathrm { { t h } } }" + }, + { + "category_id": 13, + "poly": [ + 1241, + 2005, + 1262, + 2005, + 1262, + 2030, + 1241, + 2030 + ], + "score": 0.83, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 956, + 1083, + 979, + 1083, + 979, + 1109, + 956, + 1109 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 906, + 1053, + 929, + 1053, + 929, + 1079, + 906, + 1079 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 806, + 1976, + 827, + 1976, + 827, + 2001, + 806, + 2001 + ], + "score": 0.78, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 481, + 1118, + 496, + 1118, + 496, + 1139, + 481, + 1139 + ], + "score": 0.64, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 967.0, + 691.0, + 967.0, + 691.0, + 1004.0, + 297.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2122.0, + 831.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1021.0, + 1405.0, + 1021.0, + 1405.0, + 1054.0, + 295.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1051.0, + 905.0, + 1051.0, + 905.0, + 1084.0, + 293.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1051.0, + 1212.0, + 1051.0, + 1212.0, + 1084.0, + 930.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1051.0, + 1406.0, + 1051.0, + 1406.0, + 1084.0, + 1244.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1079.0, + 346.0, + 1079.0, + 346.0, + 1116.0, + 293.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1079.0, + 955.0, + 1079.0, + 955.0, + 1116.0, + 433.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1079.0, + 1406.0, + 1079.0, + 1406.0, + 1116.0, + 980.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1112.0, + 328.0, + 1112.0, + 328.0, + 1144.0, + 296.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1112.0, + 480.0, + 1112.0, + 480.0, + 1144.0, + 370.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1112.0, + 1126.0, + 1112.0, + 1126.0, + 1144.0, + 497.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 871.0, + 1406.0, + 871.0, + 1406.0, + 909.0, + 294.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 902.0, + 996.0, + 902.0, + 996.0, + 937.0, + 295.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 909.0, + 1401.0, + 909.0, + 1401.0, + 931.0, + 1379.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 805.0, + 1970.0, + 805.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2007.0, + 828.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 2000.0, + 1240.0, + 2000.0, + 1240.0, + 2038.0, + 635.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2038.0, + 1263.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 722.0, + 594.0, + 722.0, + 594.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 943.0, + 229.0, + 943.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 437, + 1404, + 437, + 1404, + 532, + 298, + 532 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 227, + 1405, + 227, + 1405, + 323, + 298, + 323 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 294, + 962, + 1649, + 962, + 1649, + 1301, + 294, + 1301 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 301, + 1440, + 1414, + 1440, + 1414, + 2043, + 301, + 2043 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 709, + 551, + 989, + 551, + 989, + 625, + 709, + 625 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 458, + 750, + 1240, + 750, + 1240, + 840, + 458, + 840 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 299, + 871, + 1402, + 871, + 1402, + 936, + 299, + 936 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 296, + 652, + 1404, + 652, + 1404, + 718, + 296, + 718 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 296, + 1375, + 893, + 1375, + 893, + 1410, + 296, + 1410 + ], + "score": 0.932 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 814, + 74, + 814, + 106, + 299, + 106 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 301, + 374, + 643, + 374, + 643, + 407, + 301, + 407 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1351, + 778, + 1400, + 778, + 1400, + 810, + 1351, + 810 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 865, + 2087, + 865, + 2113, + 835, + 2113 + ], + "score": 0.857 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1297, + 1402, + 1297, + 1402, + 1327, + 1350, + 1327 + ], + "score": 0.341 + }, + { + "category_id": 14, + "poly": [ + 280, + 1435, + 1391, + 1435, + 1391, + 2057, + 280, + 2057 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { H _ { 2 } = \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\times \\left[ \\frac { \\Delta _ { t , x } } { \\sqrt { \\nabla f _ { \\mathbf { S } } \\phi _ { t + 1 } } } - \\frac { \\Delta _ { t , x } } { \\sqrt { \\nabla f _ { \\mathbf { S } } \\phi _ { t - 1 , \\Delta } } + \\tau } \\right] } \\\\ & { \\quad = \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\Delta _ { t , x } \\simeq \\Delta _ { t , \\Phi } \\times \\left[ \\frac { \\sqrt { \\omega _ { 2 } \\phi _ { 1 } } \\mathbf { e } _ { 1 } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } } } { \\left( \\sqrt { \\omega _ { 2 } } \\omega _ { 1 } \\right) \\tau \\left( \\sqrt { \\hat { \\omega } _ { 2 } \\omega _ { 1 } } - \\Delta _ { t , 1 } \\right) \\tau } \\right] } \\\\ & { \\quad \\quad - \\mathbf { E } _ { \\frac { 1 } { 2 } - 1 } ^ { + } \\nabla f ( x ) \\Big | _ { \\mathcal { S } } \\times \\Delta _ { t , \\Phi } \\times \\left[ \\frac { - ( 1 - \\omega _ { 2 } ) \\Delta _ { t , x } ^ { 2 } } { \\left( \\sqrt { \\omega _ { 2 } } \\omega _ { 1 } + \\gamma \\right)} \\left( \\sqrt { \\hat { \\omega } _ { 2 } \\omega _ { 1 } } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } - \\sqrt { \\nabla f _ { \\mathbf { S } } \\mathbf { e } _ { 1 } } } \\right) + \\sqrt { \\mathbf { e } _ { 1 } \\omega _ { 2 } } \\right] } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & { \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\Delta _ { t } ^ { 2 } } \\\\ & \\quad \\quad \\quad \\quad \\quad \\quad \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 709, + 547, + 990, + 547, + 990, + 625, + 709, + 625 + ], + "score": 0.94, + "latex": "x _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } ," + }, + { + "category_id": 13, + "poly": [ + 376, + 654, + 458, + 654, + 458, + 688, + 376, + 688 + ], + "score": 0.94, + "latex": "i \\in [ d ]" + }, + { + "category_id": 14, + "poly": [ + 458, + 746, + 1241, + 746, + 1241, + 843, + 458, + 843 + ], + "score": 0.93, + "latex": "f ( \\boldsymbol { x } _ { t + 1 } ) \\leq f ( \\boldsymbol { x } _ { t } ) + \\eta \\left. \\nabla f ( \\boldsymbol { x } _ { t } ) , \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\sum _ { i = 1 } ^ { d } \\frac { \\Delta _ { t , i } ^ { 2 } } { ( \\sqrt { v _ { t , i } } + \\tau ) ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 1243, + 872, + 1336, + 872, + 1336, + 906, + 1243, + 906 + ], + "score": 0.93, + "latex": "f ( x _ { t + 1 } )" + }, + { + "category_id": 14, + "poly": [ + 310, + 961, + 1428, + 961, + 1428, + 1305, + 310, + 1305 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\bar { c } _ { t } [ f ( x _ { t + 1 } ) ] \\leq f ( x _ { t } ) + \\eta \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } + \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\right. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } } \\\\ { = f ( x _ { t } ) + \\eta \\underbrace { \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\right. } _ { R _ { 1 } } + \\eta \\underbrace { \\left. \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } - \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } } \\right. \\right. } _ { R _ { 2 } } } \\\\ { \\left. \\left. + \\frac { \\eta ^ { 2 } L } { 2 } \\displaystyle \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( \\sqrt { v _ { t , j } } + \\tau ) ^ { 2 } } \\right] \\right. \\right. } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 847, + 1379, + 882, + 1379, + 882, + 1408, + 847, + 1408 + ], + "score": 0.89, + "latex": "R _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 627, + 230, + 669, + 230, + 669, + 263, + 627, + 263 + ], + "score": 0.88, + "latex": "[ m ]" + }, + { + "category_id": 13, + "poly": [ + 420, + 1379, + 455, + 1379, + 455, + 1408, + 420, + 1408 + ], + "score": 0.86, + "latex": "R _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 924, + 656, + 944, + 656, + 944, + 687, + 924, + 687 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 590, + 656, + 612, + 656, + 612, + 682, + 590, + 682 + ], + "score": 0.81, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 571, + 906, + 585, + 906, + 585, + 930, + 571, + 930 + ], + "score": 0.62, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 787, + 236, + 804, + 236, + 804, + 258, + 787, + 258 + ], + "score": 0.6, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 374.0, + 645.0, + 374.0, + 645.0, + 410.0, + 296.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 437.0, + 1405.0, + 437.0, + 1405.0, + 475.0, + 294.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 466.0, + 1405.0, + 466.0, + 1405.0, + 505.0, + 292.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 495.0, + 414.0, + 495.0, + 414.0, + 537.0, + 293.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 626.0, + 227.0, + 626.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 227.0, + 786.0, + 227.0, + 786.0, + 265.0, + 670.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 265.0, + 805.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 259.0, + 1403.0, + 259.0, + 1403.0, + 293.0, + 295.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 295.0, + 498.0, + 295.0, + 498.0, + 324.0, + 294.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 865.0, + 1242.0, + 865.0, + 1242.0, + 913.0, + 293.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 865.0, + 1406.0, + 865.0, + 1406.0, + 913.0, + 1337.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 902.0, + 570.0, + 902.0, + 570.0, + 938.0, + 295.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 902.0, + 996.0, + 902.0, + 996.0, + 938.0, + 586.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 650.0, + 375.0, + 650.0, + 375.0, + 690.0, + 294.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 650.0, + 589.0, + 650.0, + 589.0, + 690.0, + 459.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 650.0, + 923.0, + 650.0, + 923.0, + 690.0, + 613.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 650.0, + 1404.0, + 650.0, + 1404.0, + 690.0, + 945.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.25, + 682.0, + 419.25, + 682.0, + 419.25, + 721.0, + 293.25, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1369.0, + 419.0, + 1369.0, + 419.0, + 1419.0, + 293.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1369.0, + 846.0, + 1369.0, + 846.0, + 1419.0, + 456.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1369.0, + 898.0, + 1369.0, + 898.0, + 1419.0, + 883.0, + 1419.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 309, + 1159, + 1386, + 1159, + 1386, + 1354, + 309, + 1354 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 509, + 1893, + 1186, + 1893, + 1186, + 1990, + 509, + 1990 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 458, + 824, + 1236, + 824, + 1236, + 918, + 458, + 918 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 404, + 268, + 1295, + 268, + 1295, + 591, + 404, + 591 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 455, + 998, + 1246, + 998, + 1246, + 1093, + 455, + 1093 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 301, + 1403, + 1400, + 1403, + 1400, + 1602, + 301, + 1602 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 300, + 925, + 1401, + 925, + 1401, + 989, + 300, + 989 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 294, + 613, + 1403, + 613, + 1403, + 678, + 294, + 678 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 469, + 682, + 1227, + 682, + 1227, + 775, + 469, + 775 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 296, + 227, + 942, + 227, + 942, + 263, + 296, + 263 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 299, + 1115, + 1247, + 1115, + 1247, + 1150, + 299, + 1150 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 298, + 782, + 945, + 782, + 945, + 816, + 298, + 816 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 814, + 73, + 814, + 106, + 297, + 106 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 292, + 1637, + 932, + 1637, + 932, + 1672, + 292, + 1672 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 295, + 1361, + 1041, + 1361, + 1041, + 1394, + 295, + 1394 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 297, + 1851, + 371, + 1851, + 371, + 1883, + 297, + 1883 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1028, + 1401, + 1028, + 1401, + 1059, + 1350, + 1059 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1350, + 490, + 1402, + 490, + 1402, + 522, + 1350, + 522 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 291, + 1684, + 1140, + 1684, + 1140, + 1719, + 291, + 1719 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 867, + 2087, + 867, + 2114, + 834, + 2114 + ], + "score": 0.86 + }, + { + "category_id": 8, + "poly": [ + 296, + 1726, + 1625, + 1726, + 1625, + 1821, + 296, + 1821 + ], + "score": 0.72 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1598, + 1402, + 1598, + 1402, + 1628, + 1350, + 1628 + ], + "score": 0.662 + }, + { + "category_id": 8, + "poly": [ + 297, + 1726, + 1625, + 1726, + 1625, + 1821, + 297, + 1821 + ], + "score": 0.63 + }, + { + "category_id": 14, + "poly": [ + 401, + 266, + 1298, + 266, + 1298, + 592, + 401, + 592 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { R _ { 1 } = \\Bigg \\langle \\nabla f ( x _ { t } ) , \\mathbb { E } _ { t } \\left[ \\frac { \\Delta _ { t } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } \\right] \\Bigg \\rangle } \\\\ & { \\quad = \\Bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } - \\eta _ { t } K \\nabla f ( x _ { t } ) + \\eta _ { t } K \\nabla f ( x _ { t } ) \\right] \\Bigg \\rangle } \\\\ & { \\quad = - \\eta _ { t } K \\displaystyle \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\underbrace { \\Bigg \\langle \\frac { \\nabla f ( x _ { t } ) } { \\sqrt { \\beta _ { 2 } v _ { t - 1 } } + \\tau } , \\mathbb { E } _ { t } \\left[ \\Delta _ { t } + \\eta _ { t } K \\nabla f ( x _ { t } ) \\right] \\Bigg \\rangle } _ { R _ { 3 } } . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 310, + 1399, + 1400, + 1399, + 1400, + 1604, + 310, + 1604 + ], + "score": 0.94, + "latex": "\\mathbb { E } [ f ( x _ { T } ) ] \\leq f ( x _ { 0 } ) - \\frac { \\eta \\eta _ { l } K } { 4 } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )" + }, + { + "category_id": 14, + "poly": [ + 311, + 1156, + 1390, + 1156, + 1390, + 1359, + 311, + 1359 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } { { \\mathbb { E } _ { t } [ f ( x _ { t + 1 } ) ] \\leq f ( x _ { t } ) - \\frac { \\eta \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } } v _ { t - 1 , j } } + \\tau + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { + ( \\frac { \\eta \\sqrt { 1 - \\beta _ { 2 } } G } { \\tau } ) \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sqrt { v _ { t , j } } + \\tau ) } ] + ( \\frac { \\eta ^ { 2 } L } { 2 } ) \\sum _ { j = 1 } ^ { d } \\mathbb { E } _ { t } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { v _ { t , j } + \\tau ^ { 2 } ) } ] } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 510, + 1889, + 1185, + 1889, + 1185, + 1990, + 510, + 1990 + ], + "score": 0.93, + "latex": "\\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\left( 1 - \\beta _ { 2 } \\right) \\sum _ { l = 0 } ^ { t } \\beta _ { 2 } ^ { t - l } \\Delta _ { t , j } ^ { 2 } + \\tau ^ { 2 } } \\le \\mathbb { E } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { \\tau ^ { 2 } }" + }, + { + "category_id": 14, + "poly": [ + 468, + 676, + 1231, + 676, + 1231, + 777, + 468, + 777 + ], + "score": 0.93, + "latex": "R _ { 3 } \\leq \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right]" + }, + { + "category_id": 14, + "poly": [ + 454, + 993, + 1245, + 993, + 1245, + 1095, + 454, + 1095 + ], + "score": 0.93, + "latex": "R _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 4 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } } { 2 \\tau } \\mathbb { E } _ { t } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } )" + }, + { + "category_id": 14, + "poly": [ + 458, + 821, + 1240, + 821, + 1240, + 919, + 458, + 919 + ], + "score": 0.93, + "latex": "R _ { 1 } \\leq - \\frac { \\eta _ { l } K } { 2 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { \\eta _ { l } L ^ { 2 } } { 2 m \\tau } \\mathbb { E } _ { t } \\left[ \\sum _ { i = 1 } ^ { m } \\sum _ { k = 0 } ^ { K - 1 } \\| x _ { i , k } ^ { t } - x _ { t } \\| ^ { 2 } \\right]" + }, + { + "category_id": 13, + "poly": [ + 608, + 928, + 813, + 928, + 813, + 961, + 608, + 961 + ], + "score": 0.9, + "latex": "\\sqrt { v _ { t - 1 , j } } \\leq \\eta _ { l } K G" + }, + { + "category_id": 13, + "poly": [ + 467, + 1363, + 531, + 1363, + 531, + 1389, + 467, + 1389 + ], + "score": 0.9, + "latex": "t = 0" + }, + { + "category_id": 13, + "poly": [ + 828, + 1118, + 863, + 1118, + 863, + 1147, + 828, + 1147 + ], + "score": 0.89, + "latex": "R _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 915, + 1118, + 951, + 1118, + 951, + 1147, + 915, + 1147 + ], + "score": 0.89, + "latex": "R _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 563, + 1363, + 634, + 1363, + 634, + 1390, + 563, + 1390 + ], + "score": 0.88, + "latex": "T - 1" + }, + { + "category_id": 13, + "poly": [ + 607, + 617, + 643, + 617, + 643, + 646, + 607, + 646 + ], + "score": 0.88, + "latex": "R _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 596, + 232, + 631, + 232, + 631, + 261, + 596, + 261 + ], + "score": 0.88, + "latex": "R _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1232, + 617, + 1263, + 617, + 1263, + 647, + 1232, + 647 + ], + "score": 0.88, + "latex": "T _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 423, + 617, + 458, + 617, + 458, + 646, + 423, + 646 + ], + "score": 0.87, + "latex": "R _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 420, + 232, + 455, + 232, + 455, + 261, + 420, + 261 + ], + "score": 0.87, + "latex": "R _ { 1 }" + }, + { + "category_id": 14, + "poly": [ + 310, + 1723, + 1428, + 1723, + 1428, + 1823, + 310, + 1823 + ], + "score": 0.87, + "latex": "\\sum _ { = 0 } ^ { { { \\nu - 1 } } } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\right] \\leq \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 1 } ^ { T - 1 } \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { \\ell = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { j = 1 } ^ { d } \\frac { \\Delta _ { j , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } \\sum _ { l = 1 } ^ { d } \\frac { \\Delta _ { l , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) }" + }, + { + "category_id": 13, + "poly": [ + 639, + 960, + 674, + 960, + 674, + 987, + 639, + 987 + ], + "score": 0.85, + "latex": "R _ { 1 }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2122.0, + 830.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 924.0, + 607.0, + 924.0, + 607.0, + 963.0, + 295.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 924.0, + 1405.0, + 924.0, + 1405.0, + 963.0, + 814.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 952.0, + 638.0, + 952.0, + 638.0, + 993.0, + 294.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 952.0, + 690.0, + 952.0, + 690.0, + 993.0, + 675.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 609.0, + 422.0, + 609.0, + 422.0, + 651.0, + 294.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 609.0, + 606.0, + 609.0, + 606.0, + 651.0, + 459.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 609.0, + 1231.0, + 609.0, + 1231.0, + 651.0, + 644.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 609.0, + 1406.0, + 609.0, + 1406.0, + 651.0, + 1264.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 644.0, + 434.0, + 644.0, + 434.0, + 678.0, + 294.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 419.0, + 226.0, + 419.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 226.0, + 595.0, + 226.0, + 595.0, + 267.0, + 456.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 226.0, + 945.0, + 226.0, + 945.0, + 267.0, + 632.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1112.0, + 827.0, + 1112.0, + 827.0, + 1155.0, + 293.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1112.0, + 914.0, + 1112.0, + 914.0, + 1155.0, + 864.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1112.0, + 1250.0, + 1112.0, + 1250.0, + 1155.0, + 952.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 777.0, + 945.0, + 777.0, + 945.0, + 822.0, + 294.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1633.0, + 928.0, + 1633.0, + 928.0, + 1676.0, + 293.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1357.0, + 466.0, + 1357.0, + 466.0, + 1400.0, + 295.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1357.0, + 562.0, + 1357.0, + 562.0, + 1400.0, + 532.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1357.0, + 1044.0, + 1357.0, + 1044.0, + 1400.0, + 635.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1846.0, + 376.0, + 1846.0, + 376.0, + 1893.0, + 291.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1683.0, + 1141.0, + 1683.0, + 1141.0, + 1723.0, + 295.0, + 1723.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1775, + 1405, + 1775, + 1405, + 1929, + 298, + 1929 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1403, + 1942, + 1403, + 2036, + 299, + 2036 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 406, + 278, + 1297, + 278, + 1297, + 472, + 406, + 472 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 295, + 753, + 1507, + 753, + 1507, + 960, + 295, + 960 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 608, + 611, + 1092, + 611, + 1092, + 706, + 608, + 706 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 493, + 1019, + 1204, + 1019, + 1204, + 1098, + 493, + 1098 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 340, + 1158, + 1354, + 1158, + 1354, + 1251, + 340, + 1251 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 299, + 548, + 1409, + 548, + 1409, + 611, + 299, + 611 + ], + "score": 0.939 + }, + { + "category_id": 8, + "poly": [ + 566, + 1435, + 1131, + 1435, + 1131, + 1481, + 566, + 1481 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 297, + 1266, + 1079, + 1266, + 1079, + 1300, + 297, + 1300 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 292, + 1108, + 1406, + 1108, + 1406, + 1143, + 292, + 1143 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 299, + 1719, + 638, + 1719, + 638, + 1753, + 299, + 1753 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 298, + 484, + 596, + 484, + 596, + 517, + 298, + 517 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 296, + 711, + 410, + 711, + 410, + 742, + 296, + 742 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 299, + 973, + 879, + 973, + 879, + 1007, + 299, + 1007 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 291, + 227, + 1403, + 227, + 1403, + 265, + 291, + 265 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 300, + 1613, + 1267, + 1613, + 1267, + 1688, + 300, + 1688 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 298, + 1489, + 1135, + 1489, + 1135, + 1523, + 298, + 1523 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 299, + 1335, + 653, + 1335, + 653, + 1368, + 299, + 1368 + ], + "score": 0.896 + }, + { + "category_id": 8, + "poly": [ + 571, + 1531, + 1123, + 1531, + 1123, + 1579, + 571, + 1579 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 1374, + 485, + 1403, + 485, + 1403, + 515, + 1374, + 515 + ], + "score": 0.816 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1268, + 1404, + 1268, + 1404, + 1297, + 1374, + 1297 + ], + "score": 0.797 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 293, + 1391, + 905, + 1391, + 905, + 1425, + 293, + 1425 + ], + "score": 0.68 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.137 + }, + { + "category_id": 14, + "poly": [ + 404, + 275, + 1297, + 275, + 1297, + 475, + 404, + 475 + ], + "score": 0.95, + "latex": "\\begin{array} { r l r } { { \\sum _ { j = 1 } ^ { d } \\mathbb { E } [ \\frac { \\Delta _ { t , j } ^ { 2 } } { ( v _ { t , j } + \\tau ^ { 2 } ) } ] \\leq \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 3 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ & { } & { + \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\sum _ { t = 0 } ^ { T - 1 } \\mathbb { E } [ \\nabla f ( x _ { t } ) ^ { 2 } ] , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 607, + 606, + 1091, + 606, + 1091, + 706, + 607, + 706 + ], + "score": 0.94, + "latex": "\\left( \\frac { \\eta \\sqrt { 1 - \\beta _ { 2 } } G } { \\tau } \\right) \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\mathbb { E } \\left[ \\frac { \\Delta _ { t , j } ^ { 2 } } { \\sqrt { v _ { t , j } } + \\tau ) } \\right] ," + }, + { + "category_id": 14, + "poly": [ + 337, + 1153, + 1360, + 1153, + 1360, + 1255, + 337, + 1255 + ], + "score": 0.93, + "latex": "\\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } \\geq \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } \\geq \\frac { T } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G + \\tau } \\operatorname* { m i n } _ { \\eta \\leq t \\leq T } \\Vert \\nabla f ( x _ { t } ) \\Vert ^ { 2 } ." + }, + { + "category_id": 14, + "poly": [ + 493, + 1018, + 1208, + 1018, + 1208, + 1099, + 493, + 1099 + ], + "score": 0.93, + "latex": "\\left( \\sqrt { 1 - \\beta _ { 2 } } G + \\frac { \\eta L } { 2 } \\right) \\frac { 4 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } } { \\tau ^ { 2 } } \\leq \\frac { \\eta _ { l } K } { 1 6 } \\left( \\frac { 1 } { \\sqrt { \\beta _ { 2 } } \\eta _ { l } K G } + \\frac { 1 } { \\tau } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 846, + 1808, + 1037, + 1808, + 1037, + 1840, + 846, + 1840 + ], + "score": 0.93, + "latex": "\\mathbf { S } \\mathbf { G } \\mathbf { D } _ { K } ( x _ { t } , \\eta _ { l } , f _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 406, + 1837, + 517, + 1837, + 517, + 1870, + 406, + 1870 + ], + "score": 0.92, + "latex": "\\nabla f _ { i } ( x , \\bar { z } )" + }, + { + "category_id": 14, + "poly": [ + 310, + 751, + 1428, + 751, + 1428, + 965, + 310, + 965 + ], + "score": 0.92, + "latex": "\\begin{array} { c } { { \\displaystyle z _ { t } [ f ( x _ { T } ) ] \\leq f ( x _ { 0 } ) - \\frac { \\eta \\eta _ { l } K } { 8 } \\sum _ { t = 0 } ^ { T - 1 } \\sum _ { j = 1 } ^ { d } \\frac { [ \\nabla f ( x _ { t } ) ] _ { j } ^ { 2 } } { \\sqrt { \\beta _ { 2 } v _ { t - 1 , j } } + \\tau } + \\frac { 5 \\eta \\eta _ { l } ^ { 3 } K ^ { 2 } L ^ { 2 } T } { 2 \\tau } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) } } \\\\ { { + \\left( \\eta \\sqrt { 1 - \\beta _ { 2 } } G + \\frac { \\eta ^ { 2 } L } { 2 } \\right) \\times \\left[ \\frac { 4 \\eta _ { l } ^ { 2 } K T } { m \\tau ^ { 2 } } \\sum _ { j = 1 } ^ { d } \\sigma _ { l , j } ^ { 2 } + \\frac { 2 0 \\eta _ { l } ^ { 4 } K ^ { 4 } L ^ { 2 } T } { \\tau ^ { 2 } } \\mathbb { E } \\sum _ { j = 1 } ^ { d } ( \\sigma _ { l , j } ^ { 2 } + 6 K \\sigma _ { g , j } ^ { 2 } ) \\right] \\sigma _ { g , l } ^ { 2 } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 561, + 1838, + 645, + 1838, + 645, + 1868, + 561, + 1868 + ], + "score": 0.91, + "latex": "z \\sim D _ { i }" + }, + { + "category_id": 14, + "poly": [ + 564, + 1435, + 1133, + 1435, + 1133, + 1480, + 564, + 1480 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\mathbb { E } \\left[ \\| z _ { 1 } + . . . + z _ { r } \\| ^ { 2 } \\right] \\leq r \\mathbb { E } \\left[ \\| z _ { 1 } \\| ^ { 2 } + . . . + \\| z _ { r } \\| ^ { 2 } \\right] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 914, + 1496, + 1028, + 1496, + 1028, + 1523, + 914, + 1523 + ], + "score": 0.88, + "latex": "z _ { 1 } , \\ldots , z _ { r }" + }, + { + "category_id": 14, + "poly": [ + 572, + 1533, + 1123, + 1533, + 1123, + 1577, + 572, + 1577 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\mathbb { E } \\left[ \\| z _ { 1 } + . . . + z _ { r } \\| ^ { 2 } \\right] = \\mathbb { E } \\left[ \\| z _ { 1 } \\| ^ { 2 } + . . . + \\| z _ { r } \\| ^ { 2 } \\right] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 677, + 1398, + 791, + 1398, + 791, + 1425, + 677, + 1425 + ], + "score": 0.88, + "latex": "z _ { 1 } , \\ldots , z _ { r }" + }, + { + "category_id": 13, + "poly": [ + 1117, + 1842, + 1145, + 1842, + 1145, + 1868, + 1117, + 1868 + ], + "score": 0.86, + "latex": "x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 867, + 1115, + 892, + 1115, + 892, + 1142, + 867, + 1142 + ], + "score": 0.85, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 930, + 1844, + 954, + 1844, + 954, + 1869, + 930, + 1869 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1152, + 1809, + 1181, + 1809, + 1181, + 1835, + 1152, + 1835 + ], + "score": 0.82, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 691, + 1493, + 709, + 1493, + 709, + 1517, + 691, + 1517 + ], + "score": 0.65, + "latex": "O" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1717.0, + 642.0, + 1717.0, + 642.0, + 1757.0, + 294.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1611.0, + 1274.0, + 1611.0, + 1274.0, + 1655.0, + 291.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1654.0, + 615.0, + 1654.0, + 615.0, + 1689.0, + 353.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1336.0, + 653.0, + 1336.0, + 653.0, + 1369.0, + 297.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 489.0, + 1403.0, + 489.0, + 1403.0, + 517.0, + 1377.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1380.0, + 1273.0, + 1402.0, + 1273.0, + 1402.0, + 1296.0, + 1380.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1777.0, + 1406.0, + 1777.0, + 1406.0, + 1810.0, + 296.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1804.0, + 845.0, + 1804.0, + 845.0, + 1844.0, + 293.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1804.0, + 1151.0, + 1804.0, + 1151.0, + 1844.0, + 1038.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1804.0, + 1405.0, + 1804.0, + 1405.0, + 1844.0, + 1182.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1835.0, + 405.0, + 1835.0, + 405.0, + 1873.0, + 291.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1835.0, + 560.0, + 1835.0, + 560.0, + 1873.0, + 518.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1835.0, + 929.0, + 1835.0, + 929.0, + 1873.0, + 646.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1835.0, + 1116.0, + 1835.0, + 1116.0, + 1873.0, + 955.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1835.0, + 1408.0, + 1835.0, + 1408.0, + 1873.0, + 1146.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1867.0, + 1402.0, + 1867.0, + 1402.0, + 1900.0, + 296.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1896.0, + 526.0, + 1896.0, + 526.0, + 1929.0, + 296.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1404.0, + 1940.0, + 1404.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 546.0, + 1405.0, + 546.0, + 1405.0, + 585.0, + 294.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 575.0, + 412.0, + 575.0, + 412.0, + 618.0, + 293.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1264.0, + 1080.0, + 1264.0, + 1080.0, + 1304.0, + 295.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1100.0, + 866.0, + 1100.0, + 866.0, + 1149.0, + 291.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1100.0, + 1409.0, + 1100.0, + 1409.0, + 1149.0, + 893.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 485.0, + 597.0, + 485.0, + 597.0, + 517.0, + 297.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 709.0, + 412.0, + 709.0, + 412.0, + 744.0, + 294.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 971.0, + 884.0, + 971.0, + 884.0, + 1010.0, + 295.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 222.0, + 1406.0, + 222.0, + 1406.0, + 272.0, + 293.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1484.0, + 690.0, + 1484.0, + 690.0, + 1529.0, + 294.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1484.0, + 913.0, + 1484.0, + 913.0, + 1529.0, + 710.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1484.0, + 1139.0, + 1484.0, + 1139.0, + 1529.0, + 1029.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1388.0, + 676.0, + 1388.0, + 676.0, + 1430.0, + 295.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1388.0, + 900.0, + 1388.0, + 900.0, + 1430.0, + 792.0, + 1430.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 643, + 1405, + 643, + 1405, + 826, + 297, + 826 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1493, + 1405, + 1493, + 1405, + 1705, + 297, + 1705 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1719, + 1405, + 1719, + 1405, + 1871, + 298, + 1871 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 838, + 1404, + 838, + 1404, + 990, + 297, + 990 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 295, + 567, + 1399, + 567, + 1399, + 628, + 295, + 628 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1404, + 1972, + 1404, + 2034, + 299, + 2034 + ], + "score": 0.946 + }, + { + "category_id": 0, + "poly": [ + 299, + 1913, + 524, + 1913, + 524, + 1945, + 299, + 1945 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.857 + }, + { + "category_id": 1, + "poly": [ + 310, + 1069, + 839, + 1069, + 839, + 1455, + 310, + 1455 + ], + "score": 0.845 + }, + { + "category_id": 0, + "poly": [ + 299, + 1030, + 716, + 1030, + 716, + 1062, + 299, + 1062 + ], + "score": 0.833 + }, + { + "category_id": 0, + "poly": [ + 298, + 228, + 675, + 228, + 675, + 262, + 298, + 262 + ], + "score": 0.771 + }, + { + "category_id": 5, + "poly": [ + 313, + 269, + 911, + 269, + 911, + 504, + 313, + 504 + ], + "score": 0.554, + "html": "
Input: xo fort=0,.,T-1do
Samplea subsetSofclients
xt=xt
for each client i ∈ Sinparallel do
xt = SGDk(xt, nt, fi) for i∈ S (in parallel)
xt+1=∑iesxt
" + }, + { + "category_id": 1, + "poly": [ + 313, + 269, + 911, + 269, + 911, + 504, + 313, + 504 + ], + "score": 0.394 + }, + { + "category_id": 13, + "poly": [ + 845, + 733, + 931, + 733, + 931, + 767, + 845, + 767 + ], + "score": 0.93, + "latex": "f _ { i } ( x ; b )" + }, + { + "category_id": 13, + "poly": [ + 591, + 796, + 680, + 796, + 680, + 827, + 591, + 827 + ], + "score": 0.92, + "latex": "\\nabla F _ { i } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1294, + 763, + 1403, + 763, + 1403, + 797, + 1294, + 797 + ], + "score": 0.92, + "latex": "\\nabla f _ { i } ( x ; b )" + }, + { + "category_id": 13, + "poly": [ + 605, + 1751, + 753, + 1751, + 753, + 1782, + 605, + 1782 + ], + "score": 0.92, + "latex": "\\beta _ { 1 } = \\beta _ { 2 } = 0" + }, + { + "category_id": 13, + "poly": [ + 610, + 764, + 719, + 764, + 719, + 796, + 610, + 796 + ], + "score": 0.91, + "latex": "\\nabla f _ { i } ( x ; b )" + }, + { + "category_id": 13, + "poly": [ + 682, + 734, + 758, + 734, + 758, + 764, + 682, + 764 + ], + "score": 0.91, + "latex": "b \\in B _ { i }" + }, + { + "category_id": 13, + "poly": [ + 930, + 646, + 964, + 646, + 964, + 675, + 930, + 675 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 403, + 675, + 438, + 675, + 438, + 705, + 403, + 705 + ], + "score": 0.89, + "latex": "D _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1029, + 704, + 1064, + 704, + 1064, + 734, + 1029, + 734 + ], + "score": 0.89, + "latex": "D _ { i }" + }, + { + "category_id": 13, + "poly": [ + 419, + 734, + 449, + 734, + 449, + 764, + 419, + 764 + ], + "score": 0.88, + "latex": "B _ { i }" + }, + { + "category_id": 13, + "poly": [ + 534, + 1188, + 600, + 1188, + 600, + 1215, + 534, + 1215 + ], + "score": 0.87, + "latex": "i \\in S" + }, + { + "category_id": 13, + "poly": [ + 534, + 388, + 599, + 388, + 599, + 415, + 534, + 415 + ], + "score": 0.87, + "latex": "i \\in S" + }, + { + "category_id": 13, + "poly": [ + 367, + 1099, + 563, + 1099, + 563, + 1129, + 367, + 1129 + ], + "score": 0.87, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 449, + 1218, + 600, + 1218, + 600, + 1247, + 449, + 1247 + ], + "score": 0.86, + "latex": "e = 1 , \\ldots , E" + }, + { + "category_id": 13, + "poly": [ + 619, + 678, + 648, + 678, + 648, + 705, + 619, + 705 + ], + "score": 0.86, + "latex": "n _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1189, + 904, + 1219, + 904, + 1219, + 930, + 1189, + 930 + ], + "score": 0.85, + "latex": "n _ { i }" + }, + { + "category_id": 13, + "poly": [ + 706, + 417, + 772, + 417, + 772, + 445, + 706, + 445 + ], + "score": 0.85, + "latex": "i \\in S" + }, + { + "category_id": 13, + "poly": [ + 729, + 1780, + 966, + 1780, + 966, + 1812, + 729, + 1812 + ], + "score": 0.84, + "latex": "\\beta _ { 1 } = { \\bar { 0 . 9 } } , \\beta _ { 2 } = 0 . 9 9" + }, + { + "category_id": 14, + "poly": [ + 362, + 1269, + 841, + 1269, + 841, + 1453, + 362, + 1453 + ], + "score": 0.84, + "latex": "\\begin{array} { c } { \\mathsf { \\Gamma } _ { i } \\gets \\mathsf { \\Gamma } _ { i } \\gets } \\\\ { g _ { i } ^ { t } = \\nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ { x _ { i } ^ { t } = \\mathrm { C } _ { \\mathrm { L I E N T O P T } } ( x _ { i } ^ { t } , g _ { i } ^ { t } , \\eta _ { l } , t ) } \\\\ { \\Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\\\ { n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , \\Delta _ { t } = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\\\ { x _ { t + 1 } = \\mathrm { S E R V E R O P T } ( x _ { t } , - \\Delta _ { t } , \\eta , t ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 526, + 678, + 555, + 678, + 555, + 705, + 526, + 705 + ], + "score": 0.84, + "latex": "n _ { i }" + }, + { + "category_id": 13, + "poly": [ + 367, + 298, + 564, + 298, + 564, + 328, + 367, + 328 + ], + "score": 0.84, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 941, + 1590, + 967, + 1590, + 967, + 1616, + 941, + 1616 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1252, + 764, + 1282, + 764, + 1282, + 794, + 1252, + 794 + ], + "score": 0.83, + "latex": "B _ { i }" + }, + { + "category_id": 13, + "poly": [ + 391, + 871, + 417, + 871, + 417, + 897, + 391, + 897 + ], + "score": 0.8, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 409, + 416, + 662, + 416, + 662, + 449, + 409, + 449 + ], + "score": 0.8, + "latex": "x _ { i } ^ { t } = \\mathsf { S G D } _ { K } ( x _ { t } , \\eta _ { l } , \\bar { f } _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 821, + 765, + 836, + 765, + 836, + 791, + 821, + 791 + ], + "score": 0.8, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 398, + 1073, + 429, + 1073, + 429, + 1098, + 398, + 1098 + ], + "score": 0.79, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 792, + 840, + 821, + 840, + 821, + 868, + 792, + 868 + ], + "score": 0.79, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1240, + 1590, + 1259, + 1590, + 1259, + 1616, + 1240, + 1616 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 601, + 735, + 626, + 735, + 626, + 761, + 601, + 761 + ], + "score": 0.78, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 552, + 1128, + 574, + 1128, + 574, + 1154, + 552, + 1154 + ], + "score": 0.77, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 491, + 1248, + 567, + 1248, + 567, + 1276, + 491, + 1276 + ], + "score": 0.76, + "latex": "b \\in B _ { i }" + }, + { + "category_id": 13, + "poly": [ + 552, + 328, + 575, + 328, + 575, + 354, + 552, + 354 + ], + "score": 0.76, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1382, + 739, + 1401, + 739, + 1401, + 761, + 1382, + 761 + ], + "score": 0.73, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 398, + 274, + 430, + 274, + 430, + 297, + 398, + 297 + ], + "score": 0.73, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 366, + 1155, + 457, + 1155, + 457, + 1190, + 366, + 1190 + ], + "score": 0.65, + "latex": "x _ { i } ^ { t } = x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 366, + 356, + 458, + 356, + 458, + 389, + 366, + 389 + ], + "score": 0.64, + "latex": "x _ { i } ^ { t } = x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 365, + 449, + 600, + 449, + 600, + 495, + 365, + 495 + ], + "score": 0.35, + "latex": "\\begin{array} { r } { x _ { t + 1 } = \\frac { 1 } { | \\mathcal { S } | } \\sum _ { i \\in \\mathcal { S } } x _ { i } ^ { t } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 729, + 1781, + 834, + 1781, + 834, + 1811, + 729, + 1811 + ], + "score": 0.25, + "latex": "\\beta _ { 1 } = 0 . 9" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1909.0, + 528.0, + 1909.0, + 528.0, + 1952.0, + 292.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1030.0, + 720.0, + 1030.0, + 720.0, + 1061.0, + 296.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 223.0, + 677.0, + 223.0, + 677.0, + 271.0, + 292.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 642.0, + 929.0, + 642.0, + 929.0, + 679.0, + 294.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 642.0, + 1406.0, + 642.0, + 1406.0, + 679.0, + 965.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 671.0, + 402.0, + 671.0, + 402.0, + 708.0, + 294.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 671.0, + 525.0, + 671.0, + 525.0, + 708.0, + 439.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 671.0, + 618.0, + 671.0, + 618.0, + 708.0, + 556.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 671.0, + 1405.0, + 671.0, + 1405.0, + 708.0, + 649.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 703.0, + 1028.0, + 703.0, + 1028.0, + 738.0, + 295.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 703.0, + 1405.0, + 703.0, + 1405.0, + 738.0, + 1065.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 735.0, + 418.0, + 735.0, + 418.0, + 766.0, + 296.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 735.0, + 600.0, + 735.0, + 600.0, + 766.0, + 450.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 735.0, + 681.0, + 735.0, + 681.0, + 766.0, + 627.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 735.0, + 844.0, + 735.0, + 844.0, + 766.0, + 759.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 735.0, + 1381.0, + 735.0, + 1381.0, + 766.0, + 932.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 735.0, + 1405.0, + 735.0, + 1405.0, + 766.0, + 1402.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 764.0, + 609.0, + 764.0, + 609.0, + 799.0, + 294.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 764.0, + 820.0, + 764.0, + 820.0, + 799.0, + 720.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 764.0, + 1251.0, + 764.0, + 1251.0, + 799.0, + 837.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 764.0, + 1293.0, + 764.0, + 1293.0, + 799.0, + 1283.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 789.0, + 590.0, + 789.0, + 590.0, + 829.0, + 291.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 789.0, + 693.0, + 789.0, + 693.0, + 829.0, + 681.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1494.0, + 1408.0, + 1494.0, + 1408.0, + 1528.0, + 294.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1522.0, + 1408.0, + 1522.0, + 1408.0, + 1558.0, + 294.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1553.0, + 1407.0, + 1553.0, + 1407.0, + 1589.0, + 294.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1583.0, + 940.0, + 1583.0, + 940.0, + 1620.0, + 294.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1583.0, + 1239.0, + 1583.0, + 1239.0, + 1620.0, + 968.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1583.0, + 1405.0, + 1583.0, + 1405.0, + 1620.0, + 1260.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1614.0, + 1405.0, + 1614.0, + 1405.0, + 1648.0, + 295.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1642.0, + 1405.0, + 1642.0, + 1405.0, + 1677.0, + 292.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1673.0, + 716.0, + 1673.0, + 716.0, + 1707.0, + 295.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1720.0, + 1408.0, + 1720.0, + 1408.0, + 1753.0, + 296.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1749.0, + 604.0, + 1749.0, + 604.0, + 1786.0, + 293.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1749.0, + 1408.0, + 1749.0, + 1408.0, + 1786.0, + 754.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1775.0, + 728.0, + 1775.0, + 728.0, + 1815.0, + 292.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1775.0, + 1405.0, + 1775.0, + 1405.0, + 1815.0, + 967.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1805.0, + 1405.0, + 1805.0, + 1405.0, + 1849.0, + 292.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1838.0, + 496.0, + 1838.0, + 496.0, + 1875.0, + 295.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 838.0, + 791.0, + 838.0, + 791.0, + 875.0, + 295.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 838.0, + 1405.0, + 838.0, + 1405.0, + 875.0, + 822.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 868.0, + 390.0, + 868.0, + 390.0, + 906.0, + 292.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 868.0, + 1404.0, + 868.0, + 1404.0, + 906.0, + 418.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 898.0, + 1188.0, + 898.0, + 1188.0, + 935.0, + 294.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 898.0, + 1406.0, + 898.0, + 1406.0, + 935.0, + 1220.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 930.0, + 1404.0, + 930.0, + 1404.0, + 963.0, + 295.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 957.0, + 1057.0, + 957.0, + 1057.0, + 990.0, + 295.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 566.0, + 1403.0, + 566.0, + 1403.0, + 601.0, + 295.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 597.0, + 789.0, + 597.0, + 789.0, + 633.0, + 295.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 1382.0, + 1999.0, + 1382.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1066.0, + 397.0, + 1066.0, + 397.0, + 1100.0, + 323.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 1066.0, + 743.0, + 1066.0, + 743.0, + 1100.0, + 430.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1097.0, + 366.0, + 1097.0, + 366.0, + 1128.0, + 325.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1097.0, + 606.0, + 1097.0, + 606.0, + 1128.0, + 564.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1124.0, + 551.0, + 1124.0, + 551.0, + 1158.0, + 362.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1124.0, + 685.0, + 1124.0, + 685.0, + 1158.0, + 575.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1157.0, + 463.0, + 1157.0, + 463.0, + 1190.0, + 458.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1186.0, + 533.0, + 1186.0, + 533.0, + 1217.0, + 364.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1186.0, + 765.0, + 1186.0, + 765.0, + 1217.0, + 601.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1215.0, + 448.0, + 1215.0, + 448.0, + 1248.0, + 404.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1215.0, + 640.0, + 1215.0, + 640.0, + 1248.0, + 601.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1244.0, + 490.0, + 1244.0, + 490.0, + 1279.0, + 446.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1244.0, + 609.0, + 1244.0, + 609.0, + 1279.0, + 568.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 264.0, + 397.0, + 264.0, + 397.0, + 302.0, + 322.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 264.0, + 435.0, + 264.0, + 435.0, + 302.0, + 431.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 294.0, + 366.0, + 294.0, + 366.0, + 328.0, + 322.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 294.0, + 604.0, + 294.0, + 604.0, + 328.0, + 565.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 325.0, + 551.0, + 325.0, + 551.0, + 358.0, + 363.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 325.0, + 687.0, + 325.0, + 687.0, + 358.0, + 576.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 352.0, + 365.0, + 352.0, + 365.0, + 394.0, + 362.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 352.0, + 465.0, + 352.0, + 465.0, + 394.0, + 459.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 384.0, + 533.0, + 384.0, + 533.0, + 418.0, + 364.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 384.0, + 767.0, + 384.0, + 767.0, + 418.0, + 600.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 411.0, + 408.0, + 411.0, + 408.0, + 453.0, + 404.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 411.0, + 705.0, + 411.0, + 705.0, + 453.0, + 663.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 411.0, + 911.0, + 411.0, + 911.0, + 453.0, + 773.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 445.0, + 364.0, + 445.0, + 364.0, + 498.0, + 357.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 445.0, + 607.0, + 445.0, + 607.0, + 498.0, + 601.0, + 498.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1188, + 1404, + 1188, + 1404, + 1369, + 296, + 1369 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1943, + 1404, + 1943, + 1404, + 2035, + 298, + 2035 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 362, + 881, + 1403, + 881, + 1403, + 973, + 362, + 973 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 363, + 982, + 1404, + 982, + 1404, + 1163, + 363, + 1163 + ], + "score": 0.933 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 297, + 1397, + 887, + 1397, + 887, + 1432, + 297, + 1432 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 302, + 1443, + 1023, + 1443, + 1023, + 1902, + 302, + 1902 + ], + "score": 0.817 + }, + { + "category_id": 0, + "poly": [ + 302, + 233, + 1150, + 233, + 1150, + 272, + 302, + 272 + ], + "score": 0.575 + }, + { + "category_id": 1, + "poly": [ + 304, + 282, + 1171, + 282, + 1171, + 823, + 304, + 823 + ], + "score": 0.384 + }, + { + "category_id": 5, + "poly": [ + 304, + 282, + 1171, + 282, + 1171, + 823, + 304, + 823 + ], + "score": 0.163, + "html": "
Input: xo,U-1 ≥ τ²,optional β1, β2 ∈ [0,1) for FEDYOGI and FEDADAM fort=0,,T-1do SampleasubsetSofclients foreachclienti ∈Sinparallel do fore=1,...,Edo for b ∈Bi do x=xt-nVfi(xt;b）
△=x-xt n=∑iesni,△t=∑ies² n △t=β△t-1+(1-β1)△t
Ut = Ut-1 + △² (FEDADAGRAD)
Ut =Ut-1-(1-β2)△²sign(Ut-1-△²) (FEDY0G1)
Ut = β2Ut-1 + (1 - β2)△² (FEDADAM) xt+1=xt+nt+T △t
" + }, + { + "category_id": 6, + "poly": [ + 302, + 233, + 1150, + 233, + 1150, + 272, + 302, + 272 + ], + "score": 0.091 + }, + { + "category_id": 13, + "poly": [ + 658, + 281, + 821, + 281, + 821, + 315, + 658, + 315 + ], + "score": 0.92, + "latex": "\\beta _ { 1 } , \\beta _ { 2 } \\in [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 838, + 2007, + 912, + 2007, + 912, + 2034, + 838, + 2034 + ], + "score": 0.9, + "latex": "c _ { i } = c" + }, + { + "category_id": 13, + "poly": [ + 534, + 403, + 600, + 403, + 600, + 430, + 534, + 430 + ], + "score": 0.88, + "latex": "i \\in S" + }, + { + "category_id": 13, + "poly": [ + 534, + 1557, + 600, + 1557, + 600, + 1584, + 534, + 1584 + ], + "score": 0.87, + "latex": "i \\in S" + }, + { + "category_id": 13, + "poly": [ + 449, + 433, + 600, + 433, + 600, + 462, + 449, + 462 + ], + "score": 0.87, + "latex": "e = 1 , \\ldots , E" + }, + { + "category_id": 13, + "poly": [ + 449, + 1587, + 600, + 1587, + 600, + 1616, + 449, + 1616 + ], + "score": 0.87, + "latex": "e = 1 , \\ldots , E" + }, + { + "category_id": 13, + "poly": [ + 398, + 280, + 549, + 280, + 549, + 314, + 398, + 314 + ], + "score": 0.86, + "latex": "x _ { 0 } , v _ { - 1 } \\ge \\tau ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 950, + 1978, + 974, + 1978, + 974, + 2004, + 950, + 2004 + ], + "score": 0.85, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 367, + 313, + 564, + 313, + 564, + 343, + 367, + 343 + ], + "score": 0.85, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1313, + 1057, + 1313, + 1057, + 1339, + 1028, + 1339 + ], + "score": 0.84, + "latex": "n _ { i }" + }, + { + "category_id": 13, + "poly": [ + 491, + 1617, + 567, + 1617, + 567, + 1645, + 491, + 1645 + ], + "score": 0.83, + "latex": "b \\in B _ { i }" + }, + { + "category_id": 14, + "poly": [ + 361, + 1642, + 1024, + 1642, + 1024, + 1900, + 361, + 1900 + ], + "score": 0.82, + "latex": "\\begin{array} { r l } & { \\quad g _ { i } ^ { t } = \\breve { \\nabla } f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ & { \\quad \\quad x _ { i } ^ { t } = x _ { i } ^ { t } - \\eta _ { l } ( g _ { i } ^ { t } - c _ { i } + c ) } \\\\ & { \\quad c _ { i } ^ { + } = c _ { i } - c + ( E | \\mathcal { B } _ { i } | \\eta _ { l } ) ^ { - 1 } ( x _ { i } ^ { t } - x _ { i } ) } \\\\ & { \\quad \\Delta x _ { i } = x _ { i } ^ { t } - x _ { t } , \\Delta c _ { i } = c _ { i } ^ { + } - c _ { i } } \\\\ & { \\quad c _ { i } = c _ { i } ^ { + } } \\\\ & { \\quad n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , \\Delta x = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta x _ { i } , \\Delta c = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta c _ { i } } \\\\ & { \\quad x _ { t + 1 } = x _ { t } + \\eta \\Delta x , c = c + \\frac { | \\mathcal { S } | } { N } \\Delta c } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 500, + 1948, + 525, + 1948, + 525, + 1974, + 500, + 1974 + ], + "score": 0.82, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 367, + 1468, + 575, + 1468, + 575, + 1497, + 367, + 1497 + ], + "score": 0.82, + "latex": "t = 0 , \\cdots , T - 1 \\cdot" + }, + { + "category_id": 13, + "poly": [ + 491, + 463, + 566, + 463, + 566, + 491, + 491, + 491 + ], + "score": 0.81, + "latex": "b \\in B _ { i }" + }, + { + "category_id": 13, + "poly": [ + 552, + 343, + 574, + 343, + 574, + 369, + 552, + 369 + ], + "score": 0.8, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 552, + 1497, + 574, + 1497, + 574, + 1523, + 552, + 1523 + ], + "score": 0.79, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 945, + 1950, + 961, + 1950, + 961, + 1971, + 945, + 1971 + ], + "score": 0.76, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 395, + 1221, + 421, + 1221, + 421, + 1248, + 395, + 1248 + ], + "score": 0.76, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 394, + 1341, + 407, + 1341, + 407, + 1366, + 394, + 1366 + ], + "score": 0.73, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 873, + 1947, + 886, + 1947, + 886, + 1971, + 873, + 1971 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 373, + 637, + 558, + 637, + 558, + 672, + 373, + 672 + ], + "score": 0.7, + "latex": "v _ { t } = v _ { t - 1 } + \\Delta _ { t } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 366, + 371, + 458, + 371, + 458, + 404, + 366, + 404 + ], + "score": 0.7, + "latex": "x _ { i } ^ { t } = x _ { t }" + }, + { + "category_id": 14, + "poly": [ + 366, + 1524, + 458, + 1524, + 458, + 1559, + 366, + 1559 + ], + "score": 0.63, + "latex": "x _ { i } ^ { t } = x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1279, + 2005, + 1292, + 2005, + 1292, + 2030, + 1279, + 2030 + ], + "score": 0.61, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1387, + 1976, + 1402, + 1976, + 1402, + 2001, + 1387, + 2001 + ], + "score": 0.6, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 398, + 1444, + 429, + 1444, + 429, + 1467, + 398, + 1467 + ], + "score": 0.44, + "latex": "x _ { 0 }" + }, + { + "category_id": 14, + "poly": [ + 361, + 487, + 760, + 487, + 760, + 630, + 361, + 630 + ], + "score": 0.4, + "latex": "\\begin{array} { r l } & { ~ x _ { i } ^ { t } = \\check { x } _ { i } ^ { t } - \\eta _ { l } \\nabla f _ { i } ( x _ { i } ^ { t } ; b ) } \\\\ & { ~ \\Delta _ { i } ^ { t } = x _ { i } ^ { t } - x _ { t } } \\\\ & { n = \\sum _ { i \\in S } n _ { i } , ~ \\Delta _ { t } = \\sum _ { i \\in S } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\\\ & { \\Delta _ { t } = \\beta _ { 1 } \\Delta _ { t - 1 } + ( 1 - \\beta _ { 1 } ) \\Delta _ { t } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 366, + 776, + 609, + 776, + 609, + 821, + 366, + 821 + ], + "score": 0.37, + "latex": "\\begin{array} { r } { x _ { t + 1 } = x _ { t } + \\eta \\frac { \\Delta _ { t } } { \\sqrt { v _ { t } } + \\tau } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 370, + 684, + 995, + 684, + 995, + 722, + 370, + 722 + ], + "score": 0.36, + "latex": "v _ { t } = v _ { t - 1 } - ( 1 - \\beta _ { 2 } ) \\Delta _ { t } ^ { 2 } \\mathrm { s i g n } ( v _ { t - 1 } - \\Delta _ { t } ^ { 2 } ) \\left( \\mathbf { F E D Y O G I } \\right)" + }, + { + "category_id": 14, + "poly": [ + 366, + 563, + 755, + 563, + 755, + 600, + 366, + 600 + ], + "score": 0.3, + "latex": "\\begin{array} { r } { n = \\sum _ { i \\in \\mathcal { S } } n _ { i } , ~ \\Delta _ { t } = \\sum _ { i \\in \\mathcal { S } } \\frac { n _ { i } } { n } \\Delta _ { i } ^ { t } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 398, + 1443, + 524, + 1443, + 524, + 1468, + 398, + 1468 + ], + "score": 0.25, + "latex": "x _ { 0 } , c , \\eta _ { l } , \\eta" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1392.0, + 893.0, + 1392.0, + 893.0, + 1441.0, + 293.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2124.0, + 828.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 232.0, + 636.0, + 232.0, + 636.0, + 274.0, + 294.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 229.0, + 1154.0, + 229.0, + 1154.0, + 276.0, + 648.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 232.0, + 636.0, + 232.0, + 636.0, + 274.0, + 294.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 229.0, + 1154.0, + 229.0, + 1154.0, + 276.0, + 648.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1188.0, + 1405.0, + 1188.0, + 1405.0, + 1223.0, + 295.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1218.0, + 394.0, + 1218.0, + 394.0, + 1255.0, + 292.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1218.0, + 1406.0, + 1218.0, + 1406.0, + 1255.0, + 422.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1248.0, + 1406.0, + 1248.0, + 1406.0, + 1285.0, + 291.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1278.0, + 1404.0, + 1278.0, + 1404.0, + 1314.0, + 294.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1306.0, + 1027.0, + 1306.0, + 1027.0, + 1341.0, + 294.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1306.0, + 1404.0, + 1306.0, + 1404.0, + 1341.0, + 1058.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1336.0, + 393.0, + 1336.0, + 393.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1336.0, + 581.0, + 1336.0, + 581.0, + 1371.0, + 408.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 499.0, + 1943.0, + 499.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1943.0, + 872.0, + 1943.0, + 872.0, + 1977.0, + 526.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1943.0, + 944.0, + 1943.0, + 944.0, + 1977.0, + 887.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1977.0, + 962.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1976.0, + 949.0, + 1976.0, + 949.0, + 2006.0, + 296.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1976.0, + 1386.0, + 1976.0, + 1386.0, + 2006.0, + 975.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 837.0, + 2001.0, + 837.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 2001.0, + 1278.0, + 2001.0, + 1278.0, + 2037.0, + 913.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 2001.0, + 1404.0, + 2001.0, + 1404.0, + 2037.0, + 1293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 881.0, + 1404.0, + 881.0, + 1404.0, + 915.0, + 362.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 913.0, + 1404.0, + 913.0, + 1404.0, + 946.0, + 395.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 939.0, + 1117.0, + 939.0, + 1117.0, + 979.0, + 393.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 981.0, + 1405.0, + 981.0, + 1405.0, + 1016.0, + 359.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1012.0, + 1405.0, + 1012.0, + 1405.0, + 1047.0, + 394.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1044.0, + 1405.0, + 1044.0, + 1405.0, + 1075.0, + 393.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1071.0, + 1406.0, + 1071.0, + 1406.0, + 1105.0, + 394.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1101.0, + 1406.0, + 1101.0, + 1406.0, + 1135.0, + 394.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1130.0, + 1408.0, + 1130.0, + 1408.0, + 1165.0, + 395.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1435.0, + 397.0, + 1435.0, + 397.0, + 1472.0, + 324.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1464.0, + 366.0, + 1464.0, + 366.0, + 1497.0, + 324.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1464.0, + 604.0, + 1464.0, + 604.0, + 1497.0, + 576.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1494.0, + 551.0, + 1494.0, + 551.0, + 1527.0, + 363.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1494.0, + 686.0, + 1494.0, + 686.0, + 1527.0, + 575.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1521.0, + 465.0, + 1521.0, + 465.0, + 1563.0, + 459.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1554.0, + 533.0, + 1554.0, + 533.0, + 1587.0, + 364.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1554.0, + 765.0, + 1554.0, + 765.0, + 1587.0, + 601.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1584.0, + 448.0, + 1584.0, + 448.0, + 1617.0, + 406.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1584.0, + 643.0, + 1584.0, + 643.0, + 1617.0, + 601.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1613.0, + 490.0, + 1613.0, + 490.0, + 1646.0, + 446.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1613.0, + 608.0, + 1613.0, + 608.0, + 1646.0, + 568.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1808.0, + 950.0, + 1808.0, + 950.0, + 1857.5, + 690.0, + 1857.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 276.0, + 397.0, + 276.0, + 397.0, + 316.0, + 320.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 276.0, + 657.0, + 276.0, + 657.0, + 316.0, + 550.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 276.0, + 1163.0, + 276.0, + 1163.0, + 316.0, + 822.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 310.0, + 366.0, + 310.0, + 366.0, + 342.0, + 322.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 310.0, + 607.0, + 310.0, + 607.0, + 342.0, + 565.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 341.0, + 551.0, + 341.0, + 551.0, + 373.0, + 365.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 341.0, + 684.0, + 341.0, + 684.0, + 373.0, + 575.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 368.0, + 465.0, + 368.0, + 465.0, + 410.0, + 459.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 401.0, + 533.0, + 401.0, + 533.0, + 432.0, + 364.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 401.0, + 766.0, + 401.0, + 766.0, + 432.0, + 601.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 428.0, + 448.0, + 428.0, + 448.0, + 465.0, + 403.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 428.0, + 643.0, + 428.0, + 643.0, + 465.0, + 601.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 458.0, + 490.0, + 458.0, + 490.0, + 494.0, + 447.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 458.0, + 610.0, + 458.0, + 610.0, + 494.0, + 567.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 550.0, + 360.0, + 550.0, + 360.0, + 636.0, + 357.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 634.0, + 372.0, + 634.0, + 372.0, + 677.0, + 366.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 634.0, + 766.0, + 634.0, + 766.0, + 677.0, + 559.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 678.0, + 369.0, + 678.0, + 369.0, + 732.0, + 365.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 678.0, + 1001.0, + 678.0, + 1001.0, + 732.0, + 996.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 730.0, + 842.0, + 730.0, + 842.0, + 777.0, + 367.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 774.0, + 614.0, + 774.0, + 614.0, + 825.0, + 359.0, + 825.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 22, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 592, + 1405, + 592, + 1405, + 923, + 298, + 923 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 938, + 1405, + 938, + 1405, + 1357, + 298, + 1357 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 330, + 1790, + 1366, + 1790, + 1366, + 1991, + 330, + 1991 + ], + "score": 0.983, + "html": "
DATASETTRAIN CLIENTSTRAIN EXAMPLESTEST CLIENTSTEST EXAMPLES
CIFAR-1050050,00010010,000
CIFAR-10050050,00010010,000
EMNIST-623,400671,5853,40077,483
SHAKESPEARE71516,0687152,356
STACKOVERFLOW342,477135,818,730204,08816,586,035
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1483, + 1405, + 1483, + 1405, + 1665, + 298, + 1665 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 305, + 1404, + 305, + 1404, + 488, + 298, + 488 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1401, + 229, + 1401, + 292, + 297, + 292 + ], + "score": 0.947 + }, + { + "category_id": 6, + "poly": [ + 705, + 1727, + 994, + 1727, + 994, + 1759, + 705, + 1759 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 300, + 1409, + 657, + 1409, + 657, + 1445, + 300, + 1445 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.872 + }, + { + "category_id": 0, + "poly": [ + 304, + 534, + 984, + 534, + 984, + 564, + 304, + 564 + ], + "score": 0.589 + }, + { + "category_id": 1, + "poly": [ + 304, + 534, + 984, + 534, + 984, + 564, + 304, + 564 + ], + "score": 0.273 + }, + { + "category_id": 13, + "poly": [ + 584, + 428, + 659, + 428, + 659, + 455, + 584, + 455 + ], + "score": 0.9, + "latex": "E = 1" + }, + { + "category_id": 13, + "poly": [ + 621, + 372, + 674, + 372, + 674, + 399, + 621, + 399 + ], + "score": 0.89, + "latex": "\\eta _ { l } , \\eta" + }, + { + "category_id": 13, + "poly": [ + 600, + 1002, + 627, + 1002, + 627, + 1029, + 600, + 1029 + ], + "score": 0.85, + "latex": "v _ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 267, + 313, + 267, + 313, + 289, + 298, + 289 + ], + "score": 0.34, + "latex": "c" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1724.0, + 998.0, + 1724.0, + 998.0, + 1761.0, + 702.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1405.0, + 661.0, + 1405.0, + 661.0, + 1452.0, + 294.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 533.0, + 990.0, + 533.0, + 990.0, + 567.0, + 297.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 592.0, + 1408.0, + 592.0, + 1408.0, + 626.0, + 295.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 620.0, + 1406.0, + 620.0, + 1406.0, + 661.0, + 292.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 651.0, + 1407.0, + 651.0, + 1407.0, + 688.0, + 293.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 682.0, + 1407.0, + 682.0, + 1407.0, + 717.0, + 293.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 711.0, + 1407.0, + 711.0, + 1407.0, + 748.0, + 292.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 743.0, + 1402.0, + 743.0, + 1402.0, + 773.0, + 296.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 773.0, + 1403.0, + 773.0, + 1403.0, + 804.0, + 295.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 802.0, + 1405.0, + 802.0, + 1405.0, + 836.0, + 295.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 831.0, + 1405.0, + 831.0, + 1405.0, + 865.0, + 293.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 862.0, + 1405.0, + 862.0, + 1405.0, + 895.0, + 293.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 893.0, + 1251.0, + 893.0, + 1251.0, + 924.0, + 296.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 935.0, + 1405.0, + 935.0, + 1405.0, + 973.0, + 293.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 968.0, + 1405.0, + 968.0, + 1405.0, + 1003.0, + 293.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 998.0, + 599.0, + 998.0, + 599.0, + 1031.0, + 293.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 998.0, + 1406.0, + 998.0, + 1406.0, + 1031.0, + 628.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1060.0, + 296.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1055.0, + 1406.0, + 1055.0, + 1406.0, + 1091.0, + 293.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1085.0, + 1409.0, + 1085.0, + 1409.0, + 1123.0, + 293.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1119.0, + 1406.0, + 1119.0, + 1406.0, + 1150.0, + 296.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 1405.0, + 1148.0, + 1405.0, + 1179.0, + 295.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1175.0, + 1405.0, + 1175.0, + 1405.0, + 1212.0, + 293.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1205.0, + 1406.0, + 1205.0, + 1406.0, + 1241.0, + 295.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1236.0, + 1408.0, + 1236.0, + 1408.0, + 1271.0, + 295.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1266.0, + 1402.0, + 1266.0, + 1402.0, + 1297.0, + 296.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1296.0, + 1406.0, + 1296.0, + 1406.0, + 1331.0, + 295.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1325.0, + 443.0, + 1325.0, + 443.0, + 1361.0, + 293.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1482.0, + 1405.0, + 1482.0, + 1405.0, + 1517.0, + 295.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1512.0, + 1405.0, + 1512.0, + 1405.0, + 1548.0, + 295.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1541.0, + 1408.0, + 1541.0, + 1408.0, + 1578.0, + 293.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1571.0, + 1405.0, + 1571.0, + 1405.0, + 1608.0, + 293.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1604.0, + 1405.0, + 1604.0, + 1405.0, + 1636.0, + 296.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1631.0, + 1307.0, + 1631.0, + 1307.0, + 1667.0, + 293.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 307.0, + 1406.0, + 307.0, + 1406.0, + 338.0, + 296.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 335.0, + 1405.0, + 335.0, + 1405.0, + 371.0, + 295.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 366.0, + 620.0, + 366.0, + 620.0, + 402.0, + 294.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 366.0, + 1406.0, + 366.0, + 1406.0, + 402.0, + 675.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1405.0, + 397.0, + 1405.0, + 429.0, + 293.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 426.0, + 583.0, + 426.0, + 583.0, + 461.0, + 294.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 426.0, + 1406.0, + 426.0, + 1406.0, + 461.0, + 660.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 456.0, + 1122.0, + 456.0, + 1122.0, + 491.0, + 293.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 261.0, + 627.0, + 261.0, + 627.0, + 295.0, + 314.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 533.0, + 990.0, + 533.0, + 990.0, + 567.0, + 297.0, + 567.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 23, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 462, + 1402, + 462, + 1402, + 702, + 299, + 702 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 894, + 1403, + 894, + 1403, + 1106, + 298, + 1106 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1232, + 1405, + 1232, + 1405, + 1504, + 298, + 1504 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 558, + 1641, + 1140, + 1641, + 1140, + 1978, + 558, + 1978 + ], + "score": 0.978, + "html": "
LayerOutput Shape#of Trainable Parameters
Input7840
Dense1000785000
Dense500500500
Dense250125250
Dense307530
Dense2507750
Dense500125500
Dense1000501000
Dense784784784
" + }, + { + "category_id": 1, + "poly": [ + 299, + 295, + 1403, + 295, + 1403, + 446, + 299, + 446 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 717, + 1403, + 717, + 1403, + 838, + 298, + 838 + ], + "score": 0.974 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 299, + 1166, + 482, + 1166, + 482, + 1198, + 299, + 1198 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.872 + }, + { + "category_id": 0, + "poly": [ + 299, + 230, + 652, + 230, + 652, + 260, + 299, + 260 + ], + "score": 0.787 + }, + { + "category_id": 6, + "poly": [ + 296, + 1577, + 1395, + 1577, + 1395, + 1612, + 296, + 1612 + ], + "score": 0.711 + }, + { + "category_id": 13, + "poly": [ + 1281, + 1044, + 1401, + 1044, + 1401, + 1078, + 1281, + 1078 + ], + "score": 0.93, + "latex": "( x - \\mu ) / \\sigma" + }, + { + "category_id": 13, + "poly": [ + 937, + 1444, + 1002, + 1444, + 1002, + 1472, + 937, + 1472 + ], + "score": 0.9, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 783, + 1383, + 1205, + 1383, + 1205, + 1414, + 783, + 1414 + ], + "score": 0.88, + "latex": "( 2 8 \\times 2 8 ) - 1 0 0 0 - 5 0 0 - 2 5 0 - 3 0" + }, + { + "category_id": 13, + "poly": [ + 1232, + 896, + 1327, + 896, + 1327, + 925, + 1232, + 925 + ], + "score": 0.88, + "latex": "3 2 \\times 3 2" + }, + { + "category_id": 13, + "poly": [ + 371, + 1079, + 391, + 1079, + 391, + 1106, + 371, + 1106 + ], + "score": 0.81, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 867, + 1079, + 888, + 1079, + 888, + 1102, + 867, + 1102 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 790, + 1080, + 809, + 1080, + 809, + 1102, + 790, + 1102 + ], + "score": 0.73, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1107, + 1051, + 1126, + 1051, + 1126, + 1073, + 1107, + 1073 + ], + "score": 0.67, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 448, + 1017, + 563, + 1017, + 563, + 1049, + 448, + 1049 + ], + "score": 0.29, + "latex": "( 2 4 , 2 4 , 3 )" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1163.0, + 486.0, + 1163.0, + 486.0, + 1204.0, + 295.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2083.0, + 871.0, + 2083.0, + 871.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 654.0, + 229.0, + 654.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1572.0, + 1399.0, + 1572.0, + 1399.0, + 1616.0, + 299.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 462.0, + 1406.0, + 462.0, + 1406.0, + 496.0, + 295.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 492.0, + 1406.0, + 492.0, + 1406.0, + 528.0, + 293.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 520.0, + 1406.0, + 520.0, + 1406.0, + 557.0, + 292.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 551.0, + 1406.0, + 551.0, + 1406.0, + 585.0, + 294.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 578.0, + 1406.0, + 578.0, + 1406.0, + 618.0, + 293.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 610.0, + 1404.0, + 610.0, + 1404.0, + 646.0, + 293.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 639.0, + 1408.0, + 639.0, + 1408.0, + 676.0, + 293.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 672.0, + 627.0, + 672.0, + 627.0, + 706.0, + 294.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 895.0, + 1231.0, + 895.0, + 1231.0, + 929.0, + 296.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 895.0, + 1405.0, + 895.0, + 1405.0, + 929.0, + 1328.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 924.0, + 1407.0, + 924.0, + 1407.0, + 962.0, + 293.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 954.0, + 1408.0, + 954.0, + 1408.0, + 995.0, + 292.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 983.0, + 1407.0, + 983.0, + 1407.0, + 1024.0, + 292.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1016.0, + 447.0, + 1016.0, + 447.0, + 1050.0, + 294.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1016.0, + 1405.0, + 1016.0, + 1405.0, + 1050.0, + 564.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1044.0, + 1106.0, + 1044.0, + 1106.0, + 1082.0, + 293.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1044.0, + 1280.0, + 1044.0, + 1280.0, + 1082.0, + 1127.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1074.0, + 370.0, + 1074.0, + 370.0, + 1108.0, + 296.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1074.0, + 789.0, + 1074.0, + 789.0, + 1108.0, + 392.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1074.0, + 866.0, + 1074.0, + 866.0, + 1108.0, + 810.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1074.0, + 1174.0, + 1074.0, + 1174.0, + 1108.0, + 889.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1232.0, + 1406.0, + 1232.0, + 1406.0, + 1269.0, + 293.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1261.0, + 1408.0, + 1261.0, + 1408.0, + 1298.0, + 293.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1289.0, + 1407.0, + 1289.0, + 1407.0, + 1328.0, + 293.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1323.0, + 1405.0, + 1323.0, + 1405.0, + 1355.0, + 296.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1349.0, + 1405.0, + 1349.0, + 1405.0, + 1387.0, + 293.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1381.0, + 782.0, + 1381.0, + 782.0, + 1416.0, + 293.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1381.0, + 1405.0, + 1381.0, + 1405.0, + 1416.0, + 1206.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1411.0, + 1406.0, + 1411.0, + 1406.0, + 1447.0, + 295.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1440.0, + 936.0, + 1440.0, + 936.0, + 1478.0, + 292.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1440.0, + 1408.0, + 1440.0, + 1408.0, + 1478.0, + 1003.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1470.0, + 1190.0, + 1470.0, + 1190.0, + 1506.0, + 295.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 295.0, + 1405.0, + 295.0, + 1405.0, + 331.0, + 295.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 325.0, + 1405.0, + 325.0, + 1405.0, + 361.0, + 294.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 390.0, + 293.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 384.0, + 1404.0, + 384.0, + 1404.0, + 419.0, + 294.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 414.0, + 1379.0, + 414.0, + 1379.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 715.0, + 1405.0, + 715.0, + 1405.0, + 750.0, + 293.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 746.0, + 1407.0, + 746.0, + 1407.0, + 782.0, + 293.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 775.0, + 1404.0, + 775.0, + 1404.0, + 816.0, + 292.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 804.0, + 817.0, + 804.0, + 817.0, + 843.0, + 292.0, + 843.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 24, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 746, + 1405, + 746, + 1405, + 1107, + 297, + 1107 + ], + "score": 0.985 + }, + { + "category_id": 5, + "poly": [ + 324, + 308, + 1375, + 308, + 1375, + 619, + 324, + 619 + ], + "score": 0.982, + "html": "
LayerOutput Shape#of Trainable ParametersActivationHyperparameters
Input(28,28,1)0ReLUkernel size = 3; strides=(1,1)
Conv2d(26,26,32)320
Conv2d(24,24,64)18496
MaxPool2d(12,12,64)0
Dropout(12,12,64)0p= 0.25
Flatten92160
Dense1281179776
Dropout1280
Dense627998softmaxp= 0.5
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1567, + 1404, + 1567, + 1404, + 1809, + 297, + 1809 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1822, + 1404, + 1822, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 529, + 1224, + 1169, + 1224, + 1169, + 1443, + 529, + 1443 + ], + "score": 0.98, + "html": "
LayerOutput Shape#of Trainable Parameters
Input800
Embedding(80,8)720
LSTM(80,256)271360
LSTM(80,256)525312
Dense(80,90)23130
" + }, + { + "category_id": 0, + "poly": [ + 299, + 1510, + 592, + 1510, + 592, + 1543, + 299, + 1543 + ], + "score": 0.922 + }, + { + "category_id": 6, + "poly": [ + 515, + 249, + 1181, + 249, + 1181, + 282, + 515, + 282 + ], + "score": 0.915 + }, + { + "category_id": 6, + "poly": [ + 621, + 1162, + 1078, + 1162, + 1078, + 1195, + 621, + 1195 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 298, + 690, + 547, + 690, + 547, + 721, + 298, + 721 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.856 + }, + { + "category_id": 13, + "poly": [ + 789, + 1076, + 882, + 1076, + 882, + 1105, + 789, + 1105 + ], + "score": 0.89, + "latex": "8 0 \\times 9 0" + }, + { + "category_id": 13, + "poly": [ + 1297, + 418, + 1370, + 418, + 1370, + 446, + 1297, + 446 + ], + "score": 0.76, + "latex": "\\scriptstyle ( = ( 1 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 1312, + 390, + 1370, + 390, + 1370, + 418, + 1312, + 418 + ], + "score": 0.75, + "latex": "\\left( 1 , 1 \\right)" + }, + { + "category_id": 13, + "poly": [ + 1184, + 419, + 1222, + 419, + 1222, + 442, + 1184, + 442 + ], + "score": 0.57, + "latex": "= 3" + }, + { + "category_id": 13, + "poly": [ + 1184, + 392, + 1221, + 392, + 1221, + 415, + 1184, + 415 + ], + "score": 0.49, + "latex": "= 3" + }, + { + "category_id": 13, + "poly": [ + 1224, + 444, + 1308, + 444, + 1308, + 472, + 1224, + 472 + ], + "score": 0.37, + "latex": "\\ B = ( 2 , 2 )" + }, + { + "category_id": 13, + "poly": [ + 1263, + 1688, + 1402, + 1688, + 1402, + 1718, + 1263, + 1718 + ], + "score": 0.34, + "latex": "1 0 , 0 0 0 { \\mathrm { m o s t } }" + }, + { + "category_id": 13, + "poly": [ + 1173, + 472, + 1272, + 472, + 1272, + 502, + 1173, + 502 + ], + "score": 0.3, + "latex": "p = 0 . 2 5" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1509.0, + 596.0, + 1509.0, + 596.0, + 1545.0, + 296.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 248.0, + 1183.0, + 248.0, + 1183.0, + 284.0, + 516.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1162.0, + 1081.0, + 1162.0, + 1081.0, + 1197.0, + 618.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 687.0, + 551.0, + 687.0, + 551.0, + 726.0, + 295.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 748.0, + 1408.0, + 748.0, + 1408.0, + 781.0, + 296.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 779.0, + 1406.0, + 779.0, + 1406.0, + 811.0, + 295.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 803.0, + 1405.0, + 803.0, + 1405.0, + 843.0, + 292.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 836.0, + 1406.0, + 836.0, + 1406.0, + 870.0, + 292.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 868.0, + 1403.0, + 868.0, + 1403.0, + 900.0, + 295.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 894.0, + 1403.0, + 894.0, + 1403.0, + 930.0, + 294.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 924.0, + 1406.0, + 924.0, + 1406.0, + 963.0, + 292.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 958.0, + 1406.0, + 958.0, + 1406.0, + 991.0, + 295.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 985.0, + 1406.0, + 985.0, + 1406.0, + 1022.0, + 295.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1012.0, + 1406.0, + 1012.0, + 1406.0, + 1055.0, + 291.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 1405.0, + 1044.0, + 1405.0, + 1082.0, + 294.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1076.0, + 788.0, + 1076.0, + 788.0, + 1110.0, + 292.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1076.0, + 1392.0, + 1076.0, + 1392.0, + 1110.0, + 883.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1568.0, + 1402.0, + 1568.0, + 1402.0, + 1601.0, + 296.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1596.0, + 1405.0, + 1596.0, + 1405.0, + 1636.0, + 294.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1629.0, + 1405.0, + 1629.0, + 1405.0, + 1663.0, + 295.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1659.0, + 1405.0, + 1659.0, + 1405.0, + 1692.0, + 295.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1687.0, + 1262.0, + 1687.0, + 1262.0, + 1723.0, + 292.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1687.0, + 1407.0, + 1687.0, + 1407.0, + 1723.0, + 1403.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1715.0, + 1406.0, + 1715.0, + 1406.0, + 1752.0, + 294.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1747.0, + 1405.0, + 1747.0, + 1405.0, + 1784.0, + 294.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1777.0, + 688.0, + 1777.0, + 688.0, + 1812.0, + 294.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1823.0, + 1405.0, + 1823.0, + 1405.0, + 1857.0, + 294.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1852.0, + 1404.0, + 1852.0, + 1404.0, + 1887.0, + 294.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1918.0, + 293.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1915.0, + 1407.0, + 1915.0, + 1407.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 25, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 529, + 470, + 1169, + 470, + 1169, + 688, + 529, + 688 + ], + "score": 0.979, + "html": "
LayerOutput Shape#of Trainable Parameters
Input200
Embedding(20,96)960384
LSTM(20,670)2055560
Dense(20,96)64416
Dense(20,10004)970388
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1053, + 1405, + 1053, + 1405, + 1206, + 298, + 1206 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 887, + 1406, + 887, + 1406, + 1040, + 297, + 1040 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1219, + 1405, + 1219, + 1405, + 1372, + 297, + 1372 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1648, + 1404, + 1648, + 1404, + 1770, + 298, + 1770 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1403, + 228, + 1403, + 353, + 298, + 353 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 296, + 1466, + 1402, + 1466, + 1402, + 1530, + 296, + 1530 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 670, + 1808, + 1025, + 1808, + 1025, + 1901, + 670, + 1901 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 671, + 1948, + 1024, + 1948, + 1024, + 2039, + 671, + 2039 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 639, + 1543, + 1058, + 1543, + 1058, + 1642, + 639, + 1642 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 300, + 1410, + 682, + 1410, + 682, + 1442, + 300, + 1442 + ], + "score": 0.917 + }, + { + "category_id": 6, + "poly": [ + 482, + 405, + 1213, + 405, + 1213, + 440, + 482, + 440 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 298, + 1921, + 445, + 1921, + 445, + 1955, + 298, + 1955 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 300, + 832, + 701, + 832, + 701, + 864, + 300, + 864 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 864, + 2087, + 864, + 2112, + 834, + 2112 + ], + "score": 0.874 + }, + { + "category_id": 0, + "poly": [ + 296, + 762, + 847, + 762, + 847, + 798, + 296, + 798 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 298, + 1783, + 431, + 1783, + 431, + 1815, + 298, + 1815 + ], + "score": 0.836 + }, + { + "category_id": 0, + "poly": [ + 297, + 762, + 847, + 762, + 847, + 798, + 297, + 798 + ], + "score": 0.142 + }, + { + "category_id": 13, + "poly": [ + 972, + 1221, + 1072, + 1221, + 1072, + 1254, + 972, + 1254 + ], + "score": 0.93, + "latex": "( \\eta _ { l } , \\eta , \\tau )" + }, + { + "category_id": 13, + "poly": [ + 1316, + 1056, + 1402, + 1056, + 1402, + 1087, + 1316, + 1087 + ], + "score": 0.89, + "latex": "\\beta _ { 1 } = 0" + }, + { + "category_id": 14, + "poly": [ + 649, + 1598, + 1050, + 1598, + 1050, + 1641, + 649, + 1641 + ], + "score": 0.84, + "latex": "\\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \\dots , 1 0 ^ { 1 } \\}" + }, + { + "category_id": 14, + "poly": [ + 637, + 1543, + 1063, + 1543, + 1063, + 1586, + 637, + 1586 + ], + "score": 0.84, + "latex": "\\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , 1 0 ^ { - 2 } , \\dotsc , 1 0 ^ { 0 . 5 } \\}" + }, + { + "category_id": 13, + "poly": [ + 550, + 984, + 576, + 984, + 576, + 1010, + 550, + 1010 + ], + "score": 0.84, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 845, + 984, + 864, + 984, + 864, + 1011, + 845, + 1011 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 650, + 1086, + 886, + 1086, + 886, + 1117, + 650, + 1117 + ], + "score": 0.78, + "latex": "\\beta _ { 1 } = 0 . 9 , \\beta _ { 2 } = 0 . 9 9" + }, + { + "category_id": 13, + "poly": [ + 730, + 1472, + 755, + 1472, + 755, + 1500, + 730, + 1500 + ], + "score": 0.77, + "latex": "\\textcircled { \\eta _ { l } }" + }, + { + "category_id": 13, + "poly": [ + 675, + 1180, + 694, + 1180, + 694, + 1202, + 675, + 1202 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1207, + 1473, + 1226, + 1473, + 1226, + 1500, + 1207, + 1500 + ], + "score": 0.74, + "latex": "\\dot { \\eta }" + }, + { + "category_id": 14, + "poly": [ + 672, + 1947, + 1027, + 1947, + 1027, + 2040, + 672, + 2040 + ], + "score": 0.72, + "latex": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 671, + 1806, + 1027, + 1806, + 1027, + 1903, + 671, + 1903 + ], + "score": 0.58, + "latex": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 1 } \\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 651, + 1086, + 756, + 1086, + 756, + 1117, + 651, + 1117 + ], + "score": 0.37, + "latex": "\\beta _ { 1 } = 0 . 9" + }, + { + "category_id": 14, + "poly": [ + 671, + 1808, + 1027, + 1808, + 1027, + 1850, + 671, + 1850 + ], + "score": 0.29, + "latex": "\\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 0 . 5 } \\}" + }, + { + "category_id": 14, + "poly": [ + 676, + 1859, + 1021, + 1859, + 1021, + 1900, + 676, + 1900 + ], + "score": 0.29, + "latex": "\\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dots , 1 0 ^ { 1 } \\}" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1411.0, + 684.0, + 1411.0, + 684.0, + 1442.0, + 296.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 402.0, + 1216.0, + 402.0, + 1216.0, + 442.0, + 482.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 833.0, + 704.0, + 833.0, + 704.0, + 864.0, + 297.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 761.0, + 850.0, + 761.0, + 850.0, + 801.0, + 295.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 761.0, + 850.0, + 761.0, + 850.0, + 801.0, + 295.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1056.0, + 1315.0, + 1056.0, + 1315.0, + 1089.0, + 296.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1056.0, + 1406.0, + 1056.0, + 1406.0, + 1089.0, + 1403.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1086.0, + 649.0, + 1086.0, + 649.0, + 1119.0, + 295.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1086.0, + 1405.0, + 1086.0, + 1405.0, + 1119.0, + 887.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1115.0, + 1407.0, + 1115.0, + 1407.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1143.0, + 1407.0, + 1143.0, + 1407.0, + 1178.0, + 293.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1173.0, + 674.0, + 1173.0, + 674.0, + 1209.0, + 293.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1173.0, + 875.0, + 1173.0, + 875.0, + 1209.0, + 695.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 890.0, + 1404.0, + 890.0, + 1404.0, + 923.0, + 296.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 919.0, + 1407.0, + 919.0, + 1407.0, + 953.0, + 294.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 946.0, + 1407.0, + 946.0, + 1407.0, + 985.0, + 294.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 975.0, + 549.0, + 975.0, + 549.0, + 1015.0, + 294.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 975.0, + 844.0, + 975.0, + 844.0, + 1015.0, + 577.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 975.0, + 1407.0, + 975.0, + 1407.0, + 1015.0, + 865.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1007.0, + 1137.0, + 1007.0, + 1137.0, + 1043.0, + 294.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1218.0, + 971.0, + 1218.0, + 971.0, + 1257.0, + 294.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1218.0, + 1405.0, + 1218.0, + 1405.0, + 1257.0, + 1073.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 1403.0, + 1251.0, + 1403.0, + 1284.0, + 295.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1278.0, + 1407.0, + 1278.0, + 1407.0, + 1317.0, + 292.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1310.0, + 1407.0, + 1310.0, + 1407.0, + 1347.0, + 295.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1343.0, + 686.0, + 1343.0, + 686.0, + 1372.0, + 296.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1645.0, + 1406.0, + 1645.0, + 1406.0, + 1682.0, + 293.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1406.0, + 1679.0, + 1406.0, + 1711.0, + 294.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1707.0, + 1405.0, + 1707.0, + 1405.0, + 1745.0, + 292.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1738.0, + 988.0, + 1738.0, + 988.0, + 1773.0, + 293.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 264.0, + 292.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1404.0, + 290.0, + 1404.0, + 327.0, + 293.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 320.0, + 481.0, + 320.0, + 481.0, + 357.0, + 293.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1466.0, + 729.0, + 1466.0, + 729.0, + 1502.0, + 295.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1466.0, + 1206.0, + 1466.0, + 1206.0, + 1502.0, + 756.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 1466.0, + 1406.0, + 1466.0, + 1406.0, + 1502.0, + 1227.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1495.0, + 1341.0, + 1495.0, + 1341.0, + 1534.0, + 292.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1922.0, + 448.0, + 1922.0, + 448.0, + 1963.0, + 296.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1785.0, + 434.0, + 1785.0, + 434.0, + 1819.0, + 297.0, + 1819.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 26, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1067, + 1405, + 1067, + 1405, + 1188, + 298, + 1188 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 612, + 1305, + 1086, + 1305, + 1086, + 1583, + 612, + 1583 + ], + "score": 0.98, + "html": "
TASKBATCH SIZE
CIFAR-1020
CIFAR-10020
EMNISTAE20
EMNISTCR20
SHAKESPEARE4
STACKOVERFLOWLR100
STACKOVERFLOW NWP16
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1684, + 1405, + 1684, + 1405, + 1838, + 297, + 1838 + ], + "score": 0.979 + }, + { + "category_id": 8, + "poly": [ + 672, + 781, + 1023, + 781, + 1023, + 871, + 672, + 871 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 676, + 649, + 1021, + 649, + 1021, + 738, + 676, + 738 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 671, + 517, + 1024, + 517, + 1024, + 607, + 671, + 607 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 671, + 385, + 1024, + 385, + 1024, + 474, + 671, + 474 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 679, + 253, + 1017, + 253, + 1017, + 341, + 679, + 341 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 716, + 937, + 981, + 937, + 981, + 979, + 716, + 979 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 298, + 890, + 715, + 890, + 715, + 922, + 298, + 922 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 299, + 1011, + 652, + 1011, + 652, + 1042, + 299, + 1042 + ], + "score": 0.917 + }, + { + "category_id": 6, + "poly": [ + 596, + 1242, + 1101, + 1242, + 1101, + 1277, + 596, + 1277 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 297, + 494, + 457, + 494, + 457, + 527, + 297, + 527 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 813, + 74, + 813, + 105, + 299, + 105 + ], + "score": 0.904 + }, + { + "category_id": 1, + "poly": [ + 296, + 359, + 464, + 359, + 464, + 394, + 296, + 394 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 300, + 1629, + 844, + 1629, + 844, + 1660, + 300, + 1660 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.862 + }, + { + "category_id": 0, + "poly": [ + 298, + 757, + 555, + 757, + 555, + 790, + 298, + 790 + ], + "score": 0.809 + }, + { + "category_id": 1, + "poly": [ + 297, + 227, + 463, + 227, + 463, + 262, + 297, + 262 + ], + "score": 0.779 + }, + { + "category_id": 0, + "poly": [ + 297, + 625, + 529, + 625, + 529, + 658, + 297, + 658 + ], + "score": 0.613 + }, + { + "category_id": 1, + "poly": [ + 297, + 625, + 529, + 625, + 529, + 658, + 297, + 658 + ], + "score": 0.361 + }, + { + "category_id": 0, + "poly": [ + 297, + 227, + 463, + 227, + 463, + 262, + 297, + 262 + ], + "score": 0.146 + }, + { + "category_id": 1, + "poly": [ + 298, + 757, + 555, + 757, + 555, + 790, + 298, + 790 + ], + "score": 0.131 + }, + { + "category_id": 14, + "poly": [ + 671, + 780, + 1025, + 780, + 1025, + 873, + 671, + 873 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 715, + 936, + 983, + 936, + 983, + 978, + 715, + 978 + ], + "score": 0.9, + "latex": "\\tau \\in \\{ 1 0 ^ { - 5 } , \\ldots , 1 0 ^ { - 1 } \\} ." + }, + { + "category_id": 14, + "poly": [ + 676, + 649, + 1021, + 649, + 1021, + 740, + 676, + 740 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { { \\eta _ { l } \\in \\{ 1 0 ^ { - 1 } , 1 0 ^ { - 0 . 5 } , \\dots , 1 0 ^ { 3 } \\} } } \\\\ { { \\eta \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \\dots , 1 0 ^ { 1 . 5 } \\} } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 673, + 386, + 1025, + 386, + 1025, + 476, + 673, + 476 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 673, + 517, + 1025, + 517, + 1025, + 608, + 673, + 608 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 0 . 5 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 3 } , 1 0 ^ { - 2 . 5 } , \\dotsc , 1 0 ^ { 1 } \\} } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 680, + 252, + 1018, + 252, + 1018, + 342, + 680, + 342 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\eta _ { l } \\in \\{ 1 0 ^ { - 1 . 5 } , 1 0 ^ { - 1 } , \\dots , 1 0 ^ { 2 } \\} } \\\\ { \\eta \\in \\{ 1 0 ^ { - 2 } , 1 0 ^ { - 1 . 5 } , \\dots , 1 0 ^ { 1 } \\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1219, + 1780, + 1271, + 1780, + 1271, + 1808, + 1219, + 1808 + ], + "score": 0.86, + "latex": "\\eta , \\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 540, + 898, + 559, + 898, + 559, + 918, + 540, + 918 + ], + "score": 0.78, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1321, + 1781, + 1341, + 1781, + 1341, + 1803, + 1321, + 1803 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 662, + 1781, + 681, + 1781, + 681, + 1803, + 662, + 1803 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 328, + 1720, + 347, + 1720, + 347, + 1743, + 328, + 1743 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1011.0, + 653.0, + 1011.0, + 653.0, + 1044.0, + 296.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1240.0, + 1104.0, + 1240.0, + 1104.0, + 1277.0, + 593.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1630.0, + 847.0, + 1630.0, + 847.0, + 1662.0, + 296.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 755.0, + 557.0, + 755.0, + 557.0, + 794.0, + 295.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 621.0, + 533.0, + 621.0, + 533.0, + 662.0, + 294.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 466.0, + 231.0, + 466.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1064.0, + 1406.0, + 1064.0, + 1406.0, + 1107.0, + 295.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1093.0, + 1405.0, + 1093.0, + 1405.0, + 1134.0, + 293.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1123.0, + 1403.0, + 1123.0, + 1403.0, + 1163.0, + 293.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1157.0, + 514.0, + 1157.0, + 514.0, + 1190.0, + 293.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1686.0, + 1403.0, + 1686.0, + 1403.0, + 1719.0, + 296.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1713.0, + 327.0, + 1713.0, + 327.0, + 1749.0, + 295.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1749.0, + 348.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1741.0, + 1406.0, + 1741.0, + 1406.0, + 1785.0, + 292.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1773.0, + 661.0, + 1773.0, + 661.0, + 1811.0, + 292.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1773.0, + 1218.0, + 1773.0, + 1218.0, + 1811.0, + 682.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1773.0, + 1320.0, + 1773.0, + 1320.0, + 1811.0, + 1272.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1811.0, + 1342.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1806.0, + 1368.0, + 1806.0, + 1368.0, + 1839.0, + 296.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 885.0, + 539.0, + 885.0, + 539.0, + 927.0, + 294.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 885.0, + 717.0, + 885.0, + 717.0, + 927.0, + 560.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 491.0, + 461.0, + 491.0, + 461.0, + 533.0, + 294.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 363.0, + 468.0, + 363.0, + 468.0, + 400.0, + 295.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 466.0, + 231.0, + 466.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 621.0, + 533.0, + 621.0, + 533.0, + 662.0, + 294.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 755.0, + 557.0, + 755.0, + 557.0, + 794.0, + 295.0, + 794.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 27, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 328, + 577, + 1371, + 577, + 1371, + 883, + 328, + 883 + ], + "score": 0.981, + "html": "
FEDADAGRADFEDADAMFEDYOGIFEDAVGMFEDAVG
nnmnmnmnmn
CIFAR-10-3/2-1-3/2-2-3/2-2-3/2-12-120
CIFAR-100-1-1-3/20-3/20-3/20-11
EMNISTAE3/2-3/21-3/21-3/2/010
EMNIST CR-3/2-1-3/2-5/2-3/2-5/2-3/2-1/2-10
SHAKESPEARE0-1/20-20-20-1/200
STACKOVERFLOWLR212-1/22-1/22020
STACKOVERFLOW NWP -1/2-3/2-1/2-2-1/2-2-1/20-1/20
" + }, + { + "category_id": 5, + "poly": [ + 461, + 1503, + 1238, + 1503, + 1238, + 1780, + 461, + 1780 + ], + "score": 0.977, + "html": "
FEDADAGRAD FEDADAM FEDYOGI
CIFAR-10-2-3-3
CIFAR-100-2-1-1
EMNISTAE-3-3-3
EMNIST CR-2-4-4
SHAKESPEARE-1-3-3
STACKOVERFLOWLR-2-5-5
STACKOVERFLOWNWP-4-5-5
" + }, + { + "category_id": 6, + "poly": [ + 296, + 483, + 1400, + 483, + 1400, + 547, + 296, + 547 + ], + "score": 0.912 + }, + { + "category_id": 6, + "poly": [ + 297, + 1410, + 1403, + 1410, + 1403, + 1472, + 297, + 1472 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.85 + }, + { + "category_id": 13, + "poly": [ + 775, + 486, + 816, + 486, + 816, + 516, + 775, + 516 + ], + "score": 0.76, + "latex": "( \\eta _ { l } )" + }, + { + "category_id": 13, + "poly": [ + 866, + 1416, + 886, + 1416, + 886, + 1438, + 866, + 1438 + ], + "score": 0.72, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 938, + 486, + 972, + 486, + 972, + 516, + 938, + 516 + ], + "score": 0.58, + "latex": "( \\eta )" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 483.0, + 774.0, + 483.0, + 774.0, + 519.0, + 294.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 483.0, + 937.0, + 483.0, + 937.0, + 519.0, + 817.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 483.0, + 1404.0, + 483.0, + 1404.0, + 519.0, + 973.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 514.0, + 1199.0, + 514.0, + 1199.0, + 550.0, + 295.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1407.0, + 865.0, + 1407.0, + 865.0, + 1445.0, + 292.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1407.0, + 1403.0, + 1407.0, + 1403.0, + 1445.0, + 887.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1438.0, + 624.0, + 1438.0, + 624.0, + 1470.0, + 296.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 830.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 28, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1638, + 1404, + 1638, + 1404, + 1820, + 298, + 1820 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1394, + 1404, + 1394, + 1404, + 1545, + 298, + 1545 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 472, + 925, + 1227, + 925, + 1227, + 1067, + 472, + 1067 + ], + "score": 0.975, + "html": "
FED...ADAGRADADAMYOGIAvGMAvG
STACK OVERFLOW NWP24.425.725.724.520.5
STACK OVERFLOW LR66.865.266.546.540.6
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1258, + 1405, + 1258, + 1405, + 1378, + 298, + 1378 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1913, + 1404, + 1913, + 1404, + 2033, + 300, + 2033 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 595, + 248, + 1097, + 248, + 1097, + 613, + 595, + 613 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 297, + 670, + 1403, + 670, + 1403, + 764, + 297, + 764 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 298, + 1133, + 912, + 1133, + 912, + 1168, + 298, + 1168 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 300, + 1199, + 683, + 1199, + 683, + 1232, + 300, + 1232 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 813, + 76, + 813, + 104, + 299, + 104 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 302, + 1856, + 726, + 1856, + 726, + 1887, + 302, + 1887 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 298, + 1583, + 895, + 1583, + 895, + 1613, + 298, + 1613 + ], + "score": 0.877 + }, + { + "category_id": 6, + "poly": [ + 293, + 831, + 1405, + 831, + 1405, + 894, + 293, + 894 + ], + "score": 0.856 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 13, + "poly": [ + 334, + 2004, + 446, + 2004, + 446, + 2031, + 334, + 2031 + ], + "score": 0.89, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 349, + 1974, + 420, + 1974, + 420, + 2006, + 349, + 2006 + ], + "score": 0.89, + "latex": "( \\eta _ { l } , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 898, + 861, + 960, + 861, + 960, + 891, + 898, + 891 + ], + "score": 0.88, + "latex": "0 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1162, + 1294, + 1216, + 1294, + 1216, + 1320, + 1162, + 1320 + ], + "score": 0.87, + "latex": "\\eta _ { l } , \\eta" + }, + { + "category_id": 13, + "poly": [ + 1085, + 1920, + 1110, + 1920, + 1110, + 1945, + 1085, + 1945 + ], + "score": 0.8, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1281, + 678, + 1301, + 678, + 1301, + 701, + 1281, + 701 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 419, + 862, + 500, + 862, + 500, + 894, + 419, + 894 + ], + "score": 0.76, + "latex": "( \\times 1 0 0 )" + }, + { + "category_id": 13, + "poly": [ + 1382, + 1920, + 1401, + 1920, + 1401, + 1946, + 1382, + 1946 + ], + "score": 0.7, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1382, + 1294, + 1401, + 1294, + 1401, + 1316, + 1382, + 1316 + ], + "score": 0.69, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 371, + 862, + 412, + 862, + 412, + 890, + 371, + 890 + ], + "score": 0.55, + "latex": "\\textcircled { a } 5" + }, + { + "category_id": 13, + "poly": [ + 1195, + 679, + 1221, + 679, + 1221, + 705, + 1195, + 705 + ], + "score": 0.42, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1166, + 678, + 1221, + 678, + 1221, + 705, + 1166, + 705 + ], + "score": 0.39, + "latex": "\\eta , \\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1167, + 679, + 1185, + 679, + 1185, + 705, + 1167, + 705 + ], + "score": 0.28, + "latex": "\\eta" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 259.0, + 674.0, + 259.0, + 674.0, + 299.0, + 619.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 250.0, + 935.0, + 250.0, + 935.0, + 279.0, + 797.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 354.0, + 633.0, + 354.0, + 633.0, + 477.0, + 594.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 398.0, + 674.0, + 398.0, + 674.0, + 437.0, + 619.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 553.0, + 830.0, + 553.0, + 830.0, + 586.0, + 772.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 554.0, + 969.0, + 554.0, + 969.0, + 587.0, + 897.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 553.0, + 1098.0, + 553.0, + 1098.0, + 586.0, + 1026.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 581.0, + 980.0, + 581.0, + 980.0, + 612.0, + 755.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.75, + 532.5, + 680.75, + 532.5, + 680.75, + 582.5, + 626.75, + 582.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 671.0, + 1165.0, + 671.0, + 1165.0, + 709.0, + 295.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 671.0, + 1280.0, + 671.0, + 1280.0, + 709.0, + 1222.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 671.0, + 1405.0, + 671.0, + 1405.0, + 709.0, + 1302.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 701.0, + 1404.0, + 701.0, + 1404.0, + 735.0, + 295.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 733.0, + 546.0, + 733.0, + 546.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1129.0, + 914.0, + 1129.0, + 914.0, + 1174.0, + 293.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1199.0, + 686.0, + 1199.0, + 686.0, + 1235.0, + 294.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1855.0, + 730.0, + 1855.0, + 730.0, + 1891.0, + 295.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1581.0, + 895.0, + 1581.0, + 895.0, + 1616.0, + 296.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 833.0, + 1403.0, + 833.0, + 1403.0, + 865.0, + 296.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 861.0, + 370.0, + 861.0, + 370.0, + 894.0, + 293.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 861.0, + 418.0, + 861.0, + 418.0, + 894.0, + 413.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 861.0, + 897.0, + 861.0, + 897.0, + 894.0, + 501.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 861.0, + 1363.0, + 861.0, + 1363.0, + 894.0, + 961.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2120.0, + 830.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1637.0, + 1405.0, + 1637.0, + 1405.0, + 1672.0, + 294.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1669.0, + 1406.0, + 1669.0, + 1406.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1698.0, + 1406.0, + 1698.0, + 1406.0, + 1732.0, + 292.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1726.0, + 1406.0, + 1726.0, + 1406.0, + 1764.0, + 292.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1758.0, + 1405.0, + 1758.0, + 1405.0, + 1793.0, + 294.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1786.0, + 892.0, + 1786.0, + 892.0, + 1822.0, + 293.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1395.0, + 1404.0, + 1395.0, + 1404.0, + 1428.0, + 296.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1460.0, + 293.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1488.0, + 292.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1483.0, + 1406.0, + 1483.0, + 1406.0, + 1518.0, + 293.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1514.0, + 1139.0, + 1514.0, + 1139.0, + 1547.0, + 296.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1255.0, + 1410.0, + 1255.0, + 1410.0, + 1295.0, + 293.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1285.0, + 1161.0, + 1285.0, + 1161.0, + 1326.0, + 293.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1285.0, + 1381.0, + 1285.0, + 1381.0, + 1326.0, + 1217.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1285.0, + 1405.0, + 1285.0, + 1405.0, + 1326.0, + 1402.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1312.0, + 1406.0, + 1312.0, + 1406.0, + 1358.0, + 292.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1348.0, + 758.0, + 1348.0, + 758.0, + 1383.0, + 295.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1908.0, + 1084.0, + 1908.0, + 1084.0, + 1952.0, + 293.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1908.0, + 1381.0, + 1908.0, + 1381.0, + 1952.0, + 1111.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1952.0, + 1402.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1970.0, + 348.0, + 1970.0, + 348.0, + 2010.0, + 291.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1970.0, + 1408.0, + 1970.0, + 1408.0, + 2010.0, + 421.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 333.0, + 1999.0, + 333.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1999.0, + 1408.0, + 1999.0, + 1408.0, + 2038.0, + 447.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 29, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 349, + 458, + 1358, + 458, + 1358, + 1058, + 349, + 1058 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 306, + 1405, + 306, + 1405, + 429, + 297, + 429 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1404, + 229, + 1404, + 292, + 296, + 292 + ], + "score": 0.929 + }, + { + "category_id": 4, + "poly": [ + 295, + 1100, + 1408, + 1100, + 1408, + 1193, + 295, + 1193 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2112, + 835, + 2112 + ], + "score": 0.83 + }, + { + "category_id": 13, + "poly": [ + 1095, + 1160, + 1207, + 1160, + 1207, + 1190, + 1095, + 1190 + ], + "score": 0.91, + "latex": "\\bar { \\tau } = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 329, + 371, + 380, + 371, + 380, + 399, + 329, + 399 + ], + "score": 0.87, + "latex": "\\eta , \\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 833, + 401, + 858, + 401, + 858, + 428, + 833, + 428 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 765, + 402, + 784, + 402, + 784, + 429, + 765, + 429 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 464.0, + 568.0, + 464.0, + 568.0, + 483.0, + 427.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 464.0, + 903.0, + 464.0, + 903.0, + 483.0, + 778.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 461.0, + 1243.0, + 461.0, + 1243.0, + 485.0, + 1125.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 472.0, + 651.0, + 472.0, + 651.0, + 491.0, + 627.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 472.0, + 996.0, + 472.0, + 996.0, + 491.0, + 972.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 472.0, + 1340.0, + 472.0, + 1340.0, + 491.0, + 1316.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 484.0, + 602.0, + 484.0, + 602.0, + 504.0, + 366.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 485.0, + 790.0, + 485.0, + 790.0, + 502.0, + 711.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 486.0, + 812.0, + 486.0, + 812.0, + 500.0, + 792.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 486.0, + 837.0, + 486.0, + 837.0, + 500.0, + 818.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 485.0, + 865.0, + 485.0, + 865.0, + 501.0, + 842.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 486.0, + 889.0, + 486.0, + 889.0, + 500.0, + 870.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 485.0, + 942.0, + 485.0, + 942.0, + 501.0, + 892.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 484.0, + 1210.0, + 484.0, + 1210.0, + 504.0, + 1053.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 486.0, + 1232.0, + 486.0, + 1232.0, + 500.0, + 1213.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 486.0, + 1259.0, + 486.0, + 1259.0, + 500.0, + 1239.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 484.0, + 1286.0, + 484.0, + 1286.0, + 501.0, + 1262.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 505.0, + 601.0, + 505.0, + 601.0, + 690.0, + 344.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 499.0, + 652.0, + 499.0, + 652.0, + 521.0, + 626.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 505.0, + 762.0, + 505.0, + 762.0, + 649.0, + 688.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 513.0, + 789.0, + 513.0, + 789.0, + 525.0, + 765.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 513.0, + 813.0, + 513.0, + 813.0, + 525.0, + 791.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 509.0, + 946.0, + 509.0, + 946.0, + 528.0, + 817.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 499.0, + 997.0, + 499.0, + 997.0, + 521.0, + 971.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 506.0, + 1184.0, + 506.0, + 1184.0, + 555.0, + 1035.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 510.0, + 1259.0, + 510.0, + 1259.0, + 527.0, + 1185.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 510.0, + 1288.0, + 510.0, + 1288.0, + 528.0, + 1260.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 499.0, + 1341.0, + 499.0, + 1341.0, + 521.0, + 1314.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 533.0, + 602.0, + 533.0, + 602.0, + 555.0, + 366.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 530.0, + 651.0, + 530.0, + 651.0, + 548.0, + 627.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 532.0, + 765.0, + 532.0, + 765.0, + 557.0, + 711.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 538.0, + 786.0, + 538.0, + 786.0, + 552.0, + 767.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 538.0, + 813.0, + 538.0, + 813.0, + 551.0, + 792.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 536.0, + 946.0, + 536.0, + 946.0, + 554.0, + 815.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 530.0, + 996.0, + 530.0, + 996.0, + 548.0, + 971.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 533.0, + 1134.0, + 533.0, + 1134.0, + 554.0, + 1055.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 538.0, + 1156.0, + 538.0, + 1156.0, + 552.0, + 1137.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 537.0, + 1288.0, + 537.0, + 1288.0, + 553.0, + 1160.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 529.0, + 1341.0, + 529.0, + 1341.0, + 550.0, + 1313.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 561.0, + 601.0, + 561.0, + 601.0, + 582.0, + 364.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 560.0, + 945.0, + 560.0, + 945.0, + 582.0, + 707.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 558.0, + 998.0, + 558.0, + 998.0, + 580.0, + 970.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 544.0, + 1288.0, + 544.0, + 1288.0, + 589.0, + 1034.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 561.0, + 1355.0, + 561.0, + 1355.0, + 632.0, + 1332.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 586.0, + 601.0, + 586.0, + 601.0, + 607.0, + 364.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 588.0, + 651.0, + 588.0, + 651.0, + 606.0, + 627.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 585.0, + 945.0, + 585.0, + 945.0, + 607.0, + 707.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 588.0, + 996.0, + 588.0, + 996.0, + 606.0, + 971.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 577.0, + 1157.0, + 577.0, + 1157.0, + 690.0, + 1034.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 588.0, + 1184.0, + 588.0, + 1184.0, + 605.0, + 1158.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 588.0, + 1288.0, + 588.0, + 1288.0, + 605.0, + 1185.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 588.0, + 1339.0, + 588.0, + 1339.0, + 606.0, + 1316.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 612.0, + 472.0, + 612.0, + 472.0, + 634.0, + 363.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 614.0, + 602.0, + 614.0, + 602.0, + 630.0, + 498.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 612.0, + 815.0, + 612.0, + 815.0, + 634.0, + 707.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 612.0, + 945.0, + 612.0, + 945.0, + 631.0, + 840.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 612.0, + 1289.0, + 612.0, + 1289.0, + 634.0, + 1051.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 616.0, + 1340.0, + 616.0, + 1340.0, + 635.0, + 1316.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 638.0, + 602.0, + 638.0, + 602.0, + 660.0, + 363.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 644.0, + 652.0, + 644.0, + 652.0, + 666.0, + 627.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 636.0, + 945.0, + 636.0, + 945.0, + 689.0, + 689.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 644.0, + 997.0, + 644.0, + 997.0, + 666.0, + 970.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 637.0, + 1289.0, + 637.0, + 1289.0, + 661.0, + 1048.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 644.0, + 1340.0, + 644.0, + 1340.0, + 666.0, + 1314.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 664.0, + 599.0, + 664.0, + 599.0, + 685.0, + 363.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 661.0, + 945.0, + 661.0, + 945.0, + 687.0, + 704.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 661.0, + 1183.0, + 661.0, + 1183.0, + 687.0, + 1049.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 665.0, + 1289.0, + 665.0, + 1289.0, + 682.0, + 1184.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 675.0, + 651.0, + 675.0, + 651.0, + 694.0, + 628.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 673.0, + 997.0, + 673.0, + 997.0, + 695.0, + 970.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 675.0, + 1340.0, + 675.0, + 1340.0, + 694.0, + 1316.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 690.0, + 522.0, + 690.0, + 522.0, + 711.0, + 364.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 690.0, + 598.0, + 690.0, + 598.0, + 708.0, + 549.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 687.0, + 946.0, + 687.0, + 946.0, + 712.0, + 707.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 690.0, + 1288.0, + 690.0, + 1288.0, + 711.0, + 1050.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 708.0, + 601.0, + 708.0, + 601.0, + 730.0, + 389.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 704.0, + 642.0, + 704.0, + 642.0, + 720.0, + 628.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 712.0, + 944.0, + 712.0, + 944.0, + 730.0, + 736.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 706.0, + 984.0, + 706.0, + 984.0, + 718.0, + 974.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 711.0, + 1287.0, + 711.0, + 1287.0, + 729.0, + 1079.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 704.0, + 1329.0, + 704.0, + 1329.0, + 720.0, + 1317.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 723.0, + 586.0, + 723.0, + 586.0, + 747.0, + 407.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 723.0, + 930.0, + 723.0, + 930.0, + 747.0, + 751.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 723.0, + 1273.0, + 723.0, + 1273.0, + 748.0, + 1094.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 774.0, + 723.0, + 774.0, + 723.0, + 793.0, + 601.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 782.0, + 817.0, + 782.0, + 817.0, + 799.0, + 792.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 770.0, + 1063.0, + 770.0, + 1063.0, + 795.0, + 947.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 782.0, + 1161.0, + 782.0, + 1161.0, + 801.0, + 1137.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 790.0, + 611.0, + 790.0, + 611.0, + 814.0, + 531.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 795.0, + 763.0, + 795.0, + 763.0, + 812.0, + 612.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 794.0, + 1111.0, + 794.0, + 1111.0, + 814.0, + 875.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 811.0, + 817.0, + 811.0, + 817.0, + 829.0, + 792.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 811.0, + 1161.0, + 811.0, + 1161.0, + 829.0, + 1138.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 813.0, + 538.0, + 813.0, + 538.0, + 1001.0, + 509.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 814.0, + 882.0, + 814.0, + 882.0, + 902.0, + 856.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 819.0, + 560.0, + 819.0, + 560.0, + 841.0, + 530.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 820.0, + 636.0, + 820.0, + 636.0, + 836.0, + 609.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 820.0, + 662.0, + 820.0, + 662.0, + 837.0, + 636.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 820.0, + 689.0, + 820.0, + 689.0, + 837.0, + 662.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 820.0, + 765.0, + 820.0, + 765.0, + 836.0, + 738.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 820.0, + 903.0, + 820.0, + 903.0, + 841.0, + 874.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 820.0, + 930.0, + 820.0, + 930.0, + 836.0, + 902.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 820.0, + 955.0, + 820.0, + 955.0, + 836.0, + 927.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 820.0, + 981.0, + 820.0, + 981.0, + 836.0, + 954.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 820.0, + 1034.0, + 820.0, + 1034.0, + 839.0, + 977.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 820.0, + 1057.0, + 820.0, + 1057.0, + 836.0, + 1031.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 820.0, + 1111.0, + 820.0, + 1111.0, + 839.0, + 1056.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 821.0, + 585.0, + 821.0, + 585.0, + 837.0, + 558.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 821.0, + 713.0, + 821.0, + 713.0, + 837.0, + 688.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 822.0, + 610.0, + 822.0, + 610.0, + 835.0, + 584.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 822.0, + 739.0, + 822.0, + 739.0, + 835.0, + 714.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 844.0, + 687.0, + 844.0, + 687.0, + 866.0, + 531.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 847.0, + 714.0, + 847.0, + 714.0, + 863.0, + 688.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 847.0, + 738.0, + 847.0, + 738.0, + 863.0, + 715.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 847.0, + 765.0, + 847.0, + 765.0, + 863.0, + 740.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 840.0, + 817.0, + 840.0, + 817.0, + 858.0, + 792.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 844.0, + 1110.0, + 844.0, + 1110.0, + 866.0, + 875.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 839.0, + 1162.0, + 839.0, + 1162.0, + 860.0, + 1135.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 871.0, + 766.0, + 871.0, + 766.0, + 892.0, + 527.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 870.0, + 833.0, + 870.0, + 833.0, + 941.0, + 793.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 871.0, + 1110.0, + 871.0, + 1110.0, + 892.0, + 871.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 866.0, + 1177.0, + 866.0, + 1177.0, + 941.0, + 1137.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 896.0, + 766.0, + 896.0, + 766.0, + 918.0, + 530.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 898.0, + 817.0, + 898.0, + 817.0, + 916.0, + 793.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 888.0, + 1110.0, + 888.0, + 1110.0, + 1000.0, + 855.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 898.0, + 1160.0, + 898.0, + 1160.0, + 917.0, + 1137.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 921.0, + 766.0, + 921.0, + 766.0, + 943.0, + 528.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 927.0, + 817.0, + 927.0, + 817.0, + 945.0, + 793.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 920.0, + 1110.0, + 920.0, + 1110.0, + 945.0, + 870.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 927.0, + 1161.0, + 927.0, + 1161.0, + 945.0, + 1138.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 948.0, + 766.0, + 948.0, + 766.0, + 970.0, + 529.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 956.0, + 817.0, + 956.0, + 817.0, + 974.0, + 793.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 947.0, + 1110.0, + 947.0, + 1110.0, + 971.0, + 870.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 956.0, + 1161.0, + 956.0, + 1161.0, + 974.0, + 1138.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 973.0, + 766.0, + 973.0, + 766.0, + 995.0, + 528.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 973.0, + 1110.0, + 973.0, + 1110.0, + 995.0, + 872.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 985.0, + 817.0, + 985.0, + 817.0, + 1003.0, + 793.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 985.0, + 1160.0, + 985.0, + 1160.0, + 1003.0, + 1138.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1000.0, + 766.0, + 1000.0, + 766.0, + 1020.0, + 528.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1000.0, + 927.0, + 1000.0, + 927.0, + 1019.0, + 871.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1001.0, + 1056.0, + 1001.0, + 1056.0, + 1018.0, + 930.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1001.0, + 1082.0, + 1001.0, + 1082.0, + 1017.0, + 1059.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1001.0, + 1106.0, + 1001.0, + 1106.0, + 1018.0, + 1083.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1017.0, + 806.0, + 1017.0, + 806.0, + 1029.0, + 796.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1016.0, + 1150.0, + 1016.0, + 1150.0, + 1031.0, + 1137.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1022.0, + 765.0, + 1022.0, + 765.0, + 1040.0, + 558.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1022.0, + 1109.0, + 1022.0, + 1109.0, + 1040.0, + 902.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1033.0, + 750.0, + 1033.0, + 750.0, + 1057.0, + 572.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1033.0, + 1094.0, + 1033.0, + 1094.0, + 1057.0, + 915.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 555.5, + 662.0, + 555.5, + 662.0, + 649.0, + 643.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.25, + 556.0, + 1006.25, + 556.0, + 1006.25, + 648.0, + 987.25, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.25, + 559.0, + 649.25, + 559.0, + 649.25, + 577.0, + 629.25, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 559.0, + 1338.0, + 559.0, + 1338.0, + 577.0, + 1315.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 612.0, + 495.0, + 612.0, + 495.0, + 629.0, + 470.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.25, + 613.0, + 841.25, + 613.0, + 841.25, + 630.0, + 812.25, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 690.0, + 549.0, + 690.0, + 549.0, + 706.0, + 520.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1098.0, + 1409.0, + 1098.0, + 1409.0, + 1137.0, + 292.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1404.0, + 1130.0, + 1404.0, + 1165.0, + 294.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1159.0, + 1094.0, + 1159.0, + 1094.0, + 1194.0, + 292.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1159.0, + 1220.0, + 1159.0, + 1220.0, + 1194.0, + 1208.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 306.0, + 1403.0, + 306.0, + 1403.0, + 341.0, + 294.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 338.0, + 1405.0, + 338.0, + 1405.0, + 371.0, + 295.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 366.0, + 328.0, + 366.0, + 328.0, + 402.0, + 295.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 366.0, + 1406.0, + 366.0, + 1406.0, + 402.0, + 381.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 398.0, + 764.0, + 398.0, + 764.0, + 433.0, + 292.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 398.0, + 832.0, + 398.0, + 832.0, + 433.0, + 785.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 398.0, + 1225.0, + 398.0, + 1225.0, + 433.0, + 859.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1409.0, + 225.0, + 1409.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 257.0, + 570.0, + 257.0, + 570.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 30, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 347, + 293, + 1358, + 293, + 1358, + 899, + 347, + 899 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 343, + 1231, + 1358, + 1231, + 1358, + 1787, + 343, + 1787 + ], + "score": 0.956 + }, + { + "category_id": 4, + "poly": [ + 297, + 1851, + 1407, + 1851, + 1407, + 1945, + 297, + 1945 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 295, + 941, + 1408, + 941, + 1408, + 1034, + 295, + 1034 + ], + "score": 0.945 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2113, + 835, + 2113 + ], + "score": 0.854 + }, + { + "category_id": 13, + "poly": [ + 1126, + 1911, + 1238, + 1911, + 1238, + 1941, + 1126, + 1941 + ], + "score": 0.91, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 1000, + 1221, + 1000, + 1221, + 1030, + 1109, + 1030 + ], + "score": 0.91, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 919, + 614, + 1155, + 614, + 1155, + 885, + 919, + 885 + ], + "score": 0.43, + "latex": "{ \\begin{array} { r l r } & { { \\mathrm { C F ~ F A R - 1 0 0 ~ F e ~ d e d s y } } } \\\\ { 1 } & { 1 } & { 2 0 1 \\ 1 \\ 1 0 \\ 1 \\ 1 0 \\ 1 \\ 1 0 \\ 1 \\ 0 } \\\\ { 2 } & { 2 6 5 } & { 3 1 \\ 1 \\ 2 \\ 1 0 \\ 1 \\ 1 5 \\ 1 0 \\ 1 \\ 1 5 \\ 1 0 } \\\\ { 4 } & { 1 2 0 1 } & { 5 1 \\ 1 \\ 1 \\ 1 0 \\ 1 \\ 1 \\ 1 0 \\ 1 \\ 1 \\ 1 0 } \\\\ { 4 } & { 1 5 0 1 } & { 1 5 7 \\ 2 4 5 \\ 2 \\ 2 4 4 1 \\ \\ 1 2 \\ 1 0 \\ 1 \\ 1 \\ 1 \\ 0 } \\\\ { 1 } & { 1 5 0 1 } & { 1 5 7 \\ 2 4 4 4 1 \\ \\ 1 0 \\ 1 \\ 3 2 \\ 1 2 \\ 1 0 \\ 1 \\ 1 \\ 0 } \\\\ { 1 } & { 2 6 3 } & { 2 7 7 \\ 1 \\ 1 \\ 0 1 \\ 5 \\ 2 1 \\ 0 1 \\ \\ 1 0 \\ 1 \\ \\ 1 \\ 3 2 \\ 1 \\ 0 } \\\\ { 4 } & { 2 8 2 \\ \\ 3 7 \\ \\ \\ 1 1 \\ 0 9 \\ \\ 6 5 \\ 5 \\ 4 \\ 9 \\ \\ 1 4 \\ \\ 1 \\ } & { 1 2 0 2 \\ \\ 0 1 \\ \\ 1 0 \\ 1 \\ 0 \\ 1 \\ 0 } \\\\ { 5 1 } & { 1 5 3 \\ 1 \\ \\ 3 1 \\ \\ 3 \\ 4 \\ 3 \\ 2 \\ \\ 1 \\ 3 \\ 2 \\ 2 \\ 1 \\ 1 \\ 3 \\ 0 } \\\\ { 4 } & { 1 0 \\ \\ 1 1 \\ \\ 1 \\ 2 \\ 3 \\ 3 1 \\ \\ 3 \\ 3 \\ 2 \\ \\ 3 \\ 1 \\ 2 \\ 1 \\ \\ 1 0 \\ \\ 1 \\ \\ 0 } \\\\ { 1 } & { 1 0 \\ 0 \\ 0 \\ \\ 3 \\ \\ 1 \\ 3 \\ 2 \\ 1 \\ 1 \\ \\ 1 0 \\ 1 \\ \\ 1 \\ 0 \\ 1 \\ \\ 1 \\ 0 \\ 1 \\ \\ 1 \\ 0 } \\\\ { 9 . 2 5 2 0 \\ 1 \\ \\ 3 \\ 1 \\ \\ 3 \\ 2 1 \\ 0 1 \\ \\ 2 1 \\ \\ 1 0 \\ 1 \\ \\ 1 0 \\ 1 \\ \\ 1 \\ 0 1 \\ \\ 1 \\ 0 1 \\ \\ 1 \\ 0 1 \\ } \\end{array} }" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 302.0, + 575.0, + 302.0, + 575.0, + 324.0, + 420.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 313.0, + 651.0, + 313.0, + 651.0, + 331.0, + 628.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 302.0, + 910.0, + 302.0, + 910.0, + 324.0, + 771.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 310.0, + 996.0, + 310.0, + 996.0, + 332.0, + 971.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 301.0, + 1250.0, + 301.0, + 1250.0, + 327.0, + 1119.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 313.0, + 1340.0, + 313.0, + 1340.0, + 331.0, + 1315.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 324.0, + 420.0, + 324.0, + 420.0, + 344.0, + 366.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 327.0, + 443.0, + 327.0, + 443.0, + 341.0, + 423.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 325.0, + 471.0, + 325.0, + 471.0, + 342.0, + 447.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 325.0, + 496.0, + 325.0, + 496.0, + 342.0, + 472.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 325.0, + 522.0, + 325.0, + 522.0, + 342.0, + 498.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 325.0, + 547.0, + 325.0, + 547.0, + 342.0, + 524.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 325.0, + 572.0, + 325.0, + 572.0, + 342.0, + 549.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 325.0, + 599.0, + 325.0, + 599.0, + 342.0, + 575.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 324.0, + 763.0, + 324.0, + 763.0, + 344.0, + 710.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 327.0, + 787.0, + 327.0, + 787.0, + 339.0, + 766.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 327.0, + 813.0, + 327.0, + 813.0, + 339.0, + 792.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 325.0, + 840.0, + 325.0, + 840.0, + 342.0, + 817.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 327.0, + 864.0, + 327.0, + 864.0, + 341.0, + 844.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 327.0, + 890.0, + 327.0, + 890.0, + 341.0, + 870.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 327.0, + 916.0, + 327.0, + 916.0, + 339.0, + 894.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 325.0, + 943.0, + 325.0, + 943.0, + 342.0, + 920.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 325.0, + 1133.0, + 325.0, + 1133.0, + 342.0, + 1109.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 325.0, + 1158.0, + 325.0, + 1158.0, + 342.0, + 1134.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 325.0, + 1183.0, + 325.0, + 1183.0, + 342.0, + 1160.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 325.0, + 1209.0, + 325.0, + 1209.0, + 342.0, + 1186.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 325.0, + 1234.0, + 325.0, + 1234.0, + 342.0, + 1211.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 327.0, + 1260.0, + 327.0, + 1260.0, + 339.0, + 1238.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 324.0, + 1286.0, + 324.0, + 1286.0, + 342.0, + 1263.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 343.0, + 420.0, + 343.0, + 420.0, + 530.0, + 345.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 351.0, + 445.0, + 351.0, + 445.0, + 367.0, + 421.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 351.0, + 471.0, + 351.0, + 471.0, + 367.0, + 446.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 351.0, + 496.0, + 351.0, + 496.0, + 367.0, + 472.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 351.0, + 522.0, + 351.0, + 522.0, + 367.0, + 498.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 351.0, + 547.0, + 351.0, + 547.0, + 367.0, + 524.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 351.0, + 574.0, + 351.0, + 574.0, + 367.0, + 549.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 351.0, + 599.0, + 351.0, + 599.0, + 367.0, + 575.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 350.0, + 651.0, + 350.0, + 651.0, + 368.0, + 628.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 343.0, + 763.0, + 343.0, + 763.0, + 531.0, + 688.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 351.0, + 789.0, + 351.0, + 789.0, + 367.0, + 765.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 351.0, + 814.0, + 351.0, + 814.0, + 367.0, + 790.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 351.0, + 840.0, + 351.0, + 840.0, + 367.0, + 817.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 352.0, + 864.0, + 352.0, + 864.0, + 366.0, + 844.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 352.0, + 890.0, + 352.0, + 890.0, + 366.0, + 870.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 351.0, + 917.0, + 351.0, + 917.0, + 367.0, + 893.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 351.0, + 943.0, + 351.0, + 943.0, + 367.0, + 918.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 349.0, + 997.0, + 349.0, + 997.0, + 371.0, + 971.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 344.0, + 1108.0, + 344.0, + 1108.0, + 397.0, + 1036.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 351.0, + 1133.0, + 351.0, + 1133.0, + 367.0, + 1109.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 351.0, + 1158.0, + 351.0, + 1158.0, + 367.0, + 1134.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 351.0, + 1183.0, + 351.0, + 1183.0, + 367.0, + 1160.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 351.0, + 1209.0, + 351.0, + 1209.0, + 367.0, + 1186.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 351.0, + 1234.0, + 351.0, + 1234.0, + 367.0, + 1211.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 351.0, + 1261.0, + 351.0, + 1261.0, + 367.0, + 1237.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 351.0, + 1286.0, + 351.0, + 1286.0, + 367.0, + 1263.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 349.0, + 1341.0, + 349.0, + 1341.0, + 370.0, + 1313.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 377.0, + 471.0, + 377.0, + 471.0, + 394.0, + 368.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 377.0, + 496.0, + 377.0, + 496.0, + 393.0, + 472.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 377.0, + 522.0, + 377.0, + 522.0, + 393.0, + 498.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 378.0, + 546.0, + 378.0, + 546.0, + 390.0, + 524.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 377.0, + 574.0, + 377.0, + 574.0, + 393.0, + 549.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 377.0, + 598.0, + 377.0, + 598.0, + 393.0, + 575.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 374.0, + 790.0, + 374.0, + 790.0, + 396.0, + 709.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 377.0, + 814.0, + 377.0, + 814.0, + 393.0, + 791.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 378.0, + 839.0, + 378.0, + 839.0, + 390.0, + 818.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 378.0, + 864.0, + 378.0, + 864.0, + 392.0, + 844.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 378.0, + 890.0, + 378.0, + 890.0, + 392.0, + 870.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 377.0, + 917.0, + 377.0, + 917.0, + 393.0, + 893.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 377.0, + 943.0, + 377.0, + 943.0, + 393.0, + 920.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 375.0, + 1134.0, + 375.0, + 1134.0, + 396.0, + 1053.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 377.0, + 1158.0, + 377.0, + 1158.0, + 393.0, + 1135.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 378.0, + 1182.0, + 378.0, + 1182.0, + 390.0, + 1161.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 378.0, + 1208.0, + 378.0, + 1208.0, + 390.0, + 1187.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 378.0, + 1233.0, + 378.0, + 1233.0, + 392.0, + 1214.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 377.0, + 1261.0, + 377.0, + 1261.0, + 393.0, + 1238.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 377.0, + 1286.0, + 377.0, + 1286.0, + 393.0, + 1263.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 389.0, + 652.0, + 389.0, + 652.0, + 408.0, + 628.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 389.0, + 996.0, + 389.0, + 996.0, + 408.0, + 972.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 387.0, + 1055.0, + 387.0, + 1055.0, + 428.0, + 1035.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 389.0, + 1340.0, + 389.0, + 1340.0, + 408.0, + 1315.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 401.0, + 599.0, + 401.0, + 599.0, + 447.0, + 362.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 401.0, + 668.0, + 401.0, + 668.0, + 471.0, + 628.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 400.0, + 892.0, + 400.0, + 892.0, + 450.0, + 706.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 402.0, + 917.0, + 402.0, + 917.0, + 445.0, + 893.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 402.0, + 943.0, + 402.0, + 943.0, + 445.0, + 918.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 401.0, + 1011.0, + 401.0, + 1011.0, + 471.0, + 972.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 401.0, + 1236.0, + 401.0, + 1236.0, + 530.0, + 1033.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 402.0, + 1261.0, + 402.0, + 1261.0, + 445.0, + 1237.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 402.0, + 1286.0, + 402.0, + 1286.0, + 445.0, + 1263.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 402.0, + 1355.0, + 402.0, + 1355.0, + 471.0, + 1314.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 452.0, + 601.0, + 452.0, + 601.0, + 474.0, + 363.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 452.0, + 944.0, + 452.0, + 944.0, + 474.0, + 707.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 452.0, + 1289.0, + 452.0, + 1289.0, + 474.0, + 1050.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 466.0, + 652.0, + 466.0, + 652.0, + 485.0, + 628.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 466.0, + 995.0, + 466.0, + 995.0, + 485.0, + 973.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 466.0, + 1340.0, + 466.0, + 1340.0, + 485.0, + 1315.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 476.0, + 600.0, + 476.0, + 600.0, + 500.0, + 362.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 478.0, + 945.0, + 478.0, + 945.0, + 500.0, + 706.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 476.0, + 1262.0, + 476.0, + 1262.0, + 501.0, + 1048.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 480.0, + 1286.0, + 480.0, + 1286.0, + 496.0, + 1263.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 503.0, + 522.0, + 503.0, + 522.0, + 525.0, + 363.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 504.0, + 599.0, + 504.0, + 599.0, + 522.0, + 523.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 505.0, + 651.0, + 505.0, + 651.0, + 524.0, + 628.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 503.0, + 945.0, + 503.0, + 945.0, + 525.0, + 706.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 505.0, + 996.0, + 505.0, + 996.0, + 524.0, + 973.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 502.0, + 1133.0, + 502.0, + 1133.0, + 526.0, + 1048.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 501.0, + 1288.0, + 501.0, + 1288.0, + 522.0, + 1158.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 505.0, + 1340.0, + 505.0, + 1340.0, + 524.0, + 1315.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 530.0, + 572.0, + 530.0, + 572.0, + 550.0, + 363.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 531.0, + 599.0, + 531.0, + 599.0, + 547.0, + 575.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 530.0, + 918.0, + 530.0, + 918.0, + 551.0, + 706.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 531.0, + 943.0, + 531.0, + 943.0, + 547.0, + 920.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 530.0, + 1233.0, + 530.0, + 1233.0, + 551.0, + 1050.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 531.0, + 1262.0, + 531.0, + 1262.0, + 547.0, + 1234.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 531.0, + 1286.0, + 531.0, + 1286.0, + 548.0, + 1264.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 552.0, + 600.0, + 552.0, + 600.0, + 570.0, + 392.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 546.0, + 642.0, + 546.0, + 642.0, + 560.0, + 627.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 552.0, + 944.0, + 552.0, + 944.0, + 570.0, + 736.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 547.0, + 984.0, + 547.0, + 984.0, + 559.0, + 975.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 551.0, + 1288.0, + 551.0, + 1288.0, + 569.0, + 1078.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 547.0, + 1328.0, + 547.0, + 1328.0, + 558.0, + 1318.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 563.0, + 586.0, + 563.0, + 586.0, + 588.0, + 407.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 563.0, + 929.0, + 563.0, + 929.0, + 587.0, + 752.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 563.0, + 1272.0, + 563.0, + 1272.0, + 587.0, + 1096.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 611.0, + 729.0, + 611.0, + 729.0, + 636.0, + 592.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 623.0, + 817.0, + 623.0, + 817.0, + 640.0, + 791.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 623.0, + 1160.0, + 623.0, + 1160.0, + 640.0, + 1156.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 634.0, + 636.0, + 634.0, + 636.0, + 653.0, + 533.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 634.0, + 662.0, + 634.0, + 662.0, + 652.0, + 638.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 637.0, + 686.0, + 637.0, + 686.0, + 649.0, + 664.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 637.0, + 711.0, + 637.0, + 711.0, + 649.0, + 689.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 637.0, + 737.0, + 637.0, + 737.0, + 649.0, + 714.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 635.0, + 765.0, + 635.0, + 765.0, + 652.0, + 740.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 633.0, + 918.0, + 633.0, + 918.0, + 654.0, + 876.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 654.0, + 558.0, + 654.0, + 558.0, + 707.0, + 513.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 660.0, + 714.0, + 660.0, + 714.0, + 677.0, + 559.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 661.0, + 738.0, + 661.0, + 738.0, + 677.0, + 715.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 661.0, + 763.0, + 661.0, + 763.0, + 677.0, + 740.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 659.0, + 818.0, + 659.0, + 818.0, + 681.0, + 790.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 655.0, + 918.0, + 655.0, + 918.0, + 739.0, + 856.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 661.0, + 1160.0, + 661.0, + 1160.0, + 680.0, + 1156.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 684.0, + 714.0, + 684.0, + 714.0, + 706.0, + 531.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 687.0, + 739.0, + 687.0, + 739.0, + 703.0, + 715.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 687.0, + 763.0, + 687.0, + 763.0, + 703.0, + 740.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 684.0, + 918.0, + 684.0, + 918.0, + 706.0, + 874.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 695.0, + 535.0, + 695.0, + 535.0, + 738.0, + 514.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 699.0, + 817.0, + 699.0, + 817.0, + 717.0, + 791.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 699.0, + 1160.0, + 699.0, + 1160.0, + 718.0, + 1156.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 711.0, + 740.0, + 711.0, + 740.0, + 732.0, + 527.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 712.0, + 765.0, + 712.0, + 765.0, + 730.0, + 741.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 712.0, + 832.0, + 712.0, + 832.0, + 781.0, + 809.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 710.0, + 918.0, + 710.0, + 918.0, + 732.0, + 871.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 712.0, + 1177.0, + 712.0, + 1177.0, + 780.0, + 1156.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 727.0, + 765.0, + 727.0, + 765.0, + 841.0, + 511.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 738.0, + 817.0, + 738.0, + 817.0, + 756.0, + 792.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 727.0, + 918.0, + 727.0, + 918.0, + 841.0, + 855.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 737.0, + 1161.0, + 737.0, + 1161.0, + 759.0, + 1156.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 762.0, + 583.0, + 762.0, + 583.0, + 783.0, + 527.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 763.0, + 608.0, + 763.0, + 608.0, + 781.0, + 586.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 763.0, + 763.0, + 763.0, + 763.0, + 781.0, + 609.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 762.0, + 918.0, + 762.0, + 918.0, + 783.0, + 872.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 775.0, + 818.0, + 775.0, + 818.0, + 797.0, + 791.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 776.0, + 1160.0, + 776.0, + 1160.0, + 795.0, + 1156.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 786.0, + 582.0, + 786.0, + 582.0, + 811.0, + 526.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 790.0, + 609.0, + 790.0, + 609.0, + 806.0, + 586.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 789.0, + 765.0, + 789.0, + 765.0, + 806.0, + 611.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 789.0, + 918.0, + 789.0, + 918.0, + 808.0, + 872.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 812.0, + 584.0, + 812.0, + 584.0, + 836.0, + 527.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 815.0, + 609.0, + 815.0, + 609.0, + 832.0, + 585.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 815.0, + 634.0, + 815.0, + 634.0, + 832.0, + 611.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 815.0, + 686.0, + 815.0, + 686.0, + 832.0, + 663.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 815.0, + 738.0, + 815.0, + 738.0, + 832.0, + 689.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 815.0, + 765.0, + 815.0, + 765.0, + 832.0, + 740.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 813.0, + 818.0, + 813.0, + 818.0, + 835.0, + 791.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 812.0, + 918.0, + 812.0, + 918.0, + 836.0, + 871.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 813.0, + 1161.0, + 813.0, + 1161.0, + 835.0, + 1156.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 840.0, + 584.0, + 840.0, + 584.0, + 860.0, + 527.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 841.0, + 609.0, + 841.0, + 609.0, + 857.0, + 586.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 841.0, + 635.0, + 841.0, + 635.0, + 857.0, + 612.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 840.0, + 660.0, + 840.0, + 660.0, + 857.0, + 638.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 841.0, + 686.0, + 841.0, + 686.0, + 857.0, + 663.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 841.0, + 712.0, + 841.0, + 712.0, + 857.0, + 689.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 841.0, + 738.0, + 841.0, + 738.0, + 857.0, + 715.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 841.0, + 763.0, + 841.0, + 763.0, + 858.0, + 740.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 840.0, + 918.0, + 840.0, + 918.0, + 860.0, + 871.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 858.0, + 767.0, + 858.0, + 767.0, + 882.0, + 555.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 854.0, + 809.0, + 854.0, + 809.0, + 872.0, + 788.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 858.0, + 918.0, + 858.0, + 918.0, + 880.0, + 899.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 873.0, + 751.0, + 873.0, + 751.0, + 897.0, + 571.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 873.0, + 1096.0, + 873.0, + 1096.0, + 897.0, + 915.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 323.5, + 1109.0, + 323.5, + 1109.0, + 342.0, + 1053.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 504.0, + 1160.0, + 504.0, + 1160.0, + 520.5, + 1131.0, + 520.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.75, + 814.0, + 659.75, + 814.0, + 659.75, + 832.0, + 637.75, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 854.0, + 1152.0, + 854.0, + 1152.0, + 870.5, + 1135.0, + 870.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1241.0, + 574.0, + 1241.0, + 574.0, + 1263.0, + 416.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1243.0, + 906.0, + 1243.0, + 906.0, + 1261.0, + 768.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1243.0, + 1247.0, + 1243.0, + 1247.0, + 1262.0, + 1116.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1256.0, + 417.0, + 1256.0, + 417.0, + 1442.0, + 343.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1264.0, + 442.0, + 1264.0, + 442.0, + 1280.0, + 418.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1263.0, + 468.0, + 1263.0, + 468.0, + 1279.0, + 445.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1261.0, + 494.0, + 1261.0, + 494.0, + 1281.0, + 470.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1263.0, + 518.0, + 1263.0, + 518.0, + 1280.0, + 495.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1264.0, + 543.0, + 1264.0, + 543.0, + 1280.0, + 519.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1263.0, + 569.0, + 1263.0, + 569.0, + 1279.0, + 546.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1265.0, + 593.0, + 1265.0, + 593.0, + 1279.0, + 573.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1256.0, + 667.0, + 1256.0, + 667.0, + 1438.0, + 622.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1256.0, + 761.0, + 1256.0, + 761.0, + 1442.0, + 687.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1263.0, + 787.0, + 1263.0, + 787.0, + 1280.0, + 763.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1263.0, + 812.0, + 1263.0, + 812.0, + 1280.0, + 789.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1263.0, + 837.0, + 1263.0, + 837.0, + 1280.0, + 813.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1263.0, + 887.0, + 1263.0, + 887.0, + 1280.0, + 839.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1264.0, + 911.0, + 1264.0, + 911.0, + 1278.0, + 891.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1263.0, + 938.0, + 1263.0, + 938.0, + 1280.0, + 914.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1256.0, + 1011.0, + 1256.0, + 1011.0, + 1438.0, + 965.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1257.0, + 1105.0, + 1257.0, + 1105.0, + 1341.0, + 1034.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1263.0, + 1130.0, + 1263.0, + 1130.0, + 1279.0, + 1107.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1264.0, + 1156.0, + 1264.0, + 1156.0, + 1280.0, + 1133.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1264.0, + 1180.0, + 1264.0, + 1180.0, + 1280.0, + 1157.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1263.0, + 1230.0, + 1263.0, + 1230.0, + 1280.0, + 1183.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1263.0, + 1257.0, + 1263.0, + 1257.0, + 1280.0, + 1234.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1263.0, + 1281.0, + 1263.0, + 1281.0, + 1280.0, + 1258.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1256.0, + 1356.0, + 1256.0, + 1356.0, + 1440.0, + 1309.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1287.0, + 417.0, + 1287.0, + 417.0, + 1309.0, + 365.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1290.0, + 442.0, + 1290.0, + 442.0, + 1303.0, + 420.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1288.0, + 468.0, + 1288.0, + 468.0, + 1305.0, + 445.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1288.0, + 493.0, + 1288.0, + 493.0, + 1304.0, + 470.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1290.0, + 517.0, + 1290.0, + 517.0, + 1303.0, + 497.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1289.0, + 542.0, + 1289.0, + 542.0, + 1303.0, + 522.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1290.0, + 567.0, + 1290.0, + 567.0, + 1304.0, + 547.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1289.0, + 594.0, + 1289.0, + 594.0, + 1305.0, + 571.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1288.0, + 648.0, + 1288.0, + 648.0, + 1307.0, + 625.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1286.0, + 761.0, + 1286.0, + 761.0, + 1307.0, + 709.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1290.0, + 785.0, + 1290.0, + 785.0, + 1303.0, + 763.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1288.0, + 812.0, + 1288.0, + 812.0, + 1305.0, + 789.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1288.0, + 837.0, + 1288.0, + 837.0, + 1304.0, + 813.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1289.0, + 861.0, + 1289.0, + 861.0, + 1303.0, + 841.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1288.0, + 887.0, + 1288.0, + 887.0, + 1304.0, + 863.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1288.0, + 913.0, + 1288.0, + 913.0, + 1304.0, + 890.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1289.0, + 938.0, + 1289.0, + 938.0, + 1305.0, + 914.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1285.0, + 996.0, + 1285.0, + 996.0, + 1311.0, + 965.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1286.0, + 1105.0, + 1286.0, + 1105.0, + 1309.0, + 1053.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1290.0, + 1129.0, + 1290.0, + 1129.0, + 1303.0, + 1108.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1288.0, + 1156.0, + 1288.0, + 1156.0, + 1304.0, + 1133.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1288.0, + 1180.0, + 1288.0, + 1180.0, + 1304.0, + 1157.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1289.0, + 1205.0, + 1289.0, + 1205.0, + 1303.0, + 1185.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1289.0, + 1229.0, + 1289.0, + 1229.0, + 1303.0, + 1209.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1290.0, + 1256.0, + 1290.0, + 1256.0, + 1303.0, + 1235.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1290.0, + 1280.0, + 1290.0, + 1280.0, + 1304.0, + 1260.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1286.0, + 1338.0, + 1286.0, + 1338.0, + 1309.0, + 1313.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1312.0, + 417.0, + 1312.0, + 417.0, + 1332.0, + 366.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1314.0, + 443.0, + 1314.0, + 443.0, + 1330.0, + 418.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1314.0, + 468.0, + 1314.0, + 468.0, + 1330.0, + 445.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1314.0, + 493.0, + 1314.0, + 493.0, + 1330.0, + 470.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1314.0, + 518.0, + 1314.0, + 518.0, + 1330.0, + 495.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1315.0, + 542.0, + 1315.0, + 542.0, + 1329.0, + 522.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1314.0, + 569.0, + 1314.0, + 569.0, + 1330.0, + 546.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1314.0, + 594.0, + 1314.0, + 594.0, + 1330.0, + 571.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1314.0, + 652.0, + 1314.0, + 652.0, + 1337.0, + 623.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1312.0, + 762.0, + 1312.0, + 762.0, + 1332.0, + 709.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1314.0, + 787.0, + 1314.0, + 787.0, + 1330.0, + 763.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1314.0, + 812.0, + 1314.0, + 812.0, + 1330.0, + 789.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1314.0, + 837.0, + 1314.0, + 837.0, + 1330.0, + 813.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1313.0, + 863.0, + 1313.0, + 863.0, + 1330.0, + 839.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1315.0, + 885.0, + 1315.0, + 885.0, + 1329.0, + 866.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1314.0, + 938.0, + 1314.0, + 938.0, + 1330.0, + 914.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1314.0, + 996.0, + 1314.0, + 996.0, + 1337.0, + 967.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1312.0, + 1106.0, + 1312.0, + 1106.0, + 1332.0, + 1053.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1314.0, + 1130.0, + 1314.0, + 1130.0, + 1330.0, + 1107.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1314.0, + 1156.0, + 1314.0, + 1156.0, + 1330.0, + 1133.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1315.0, + 1179.0, + 1315.0, + 1179.0, + 1329.0, + 1159.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1314.0, + 1207.0, + 1314.0, + 1207.0, + 1330.0, + 1183.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1315.0, + 1229.0, + 1315.0, + 1229.0, + 1329.0, + 1209.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1315.0, + 1255.0, + 1315.0, + 1255.0, + 1329.0, + 1235.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1315.0, + 1280.0, + 1315.0, + 1280.0, + 1329.0, + 1260.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1311.0, + 1341.0, + 1311.0, + 1341.0, + 1338.0, + 1309.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1338.0, + 417.0, + 1338.0, + 417.0, + 1359.0, + 364.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1339.0, + 443.0, + 1339.0, + 443.0, + 1355.0, + 418.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1338.0, + 467.0, + 1338.0, + 467.0, + 1355.0, + 445.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1338.0, + 493.0, + 1338.0, + 493.0, + 1355.0, + 470.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1338.0, + 517.0, + 1338.0, + 517.0, + 1355.0, + 495.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1338.0, + 544.0, + 1338.0, + 544.0, + 1355.0, + 521.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1340.0, + 567.0, + 1340.0, + 567.0, + 1354.0, + 547.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1339.0, + 595.0, + 1339.0, + 595.0, + 1355.0, + 572.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1342.0, + 652.0, + 1342.0, + 652.0, + 1364.0, + 623.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1338.0, + 761.0, + 1338.0, + 761.0, + 1359.0, + 708.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1339.0, + 787.0, + 1339.0, + 787.0, + 1355.0, + 762.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1339.0, + 812.0, + 1339.0, + 812.0, + 1355.0, + 789.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1338.0, + 837.0, + 1338.0, + 837.0, + 1355.0, + 813.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1338.0, + 862.0, + 1338.0, + 862.0, + 1355.0, + 839.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1338.0, + 887.0, + 1338.0, + 887.0, + 1355.0, + 864.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1338.0, + 913.0, + 1338.0, + 913.0, + 1355.0, + 890.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1339.0, + 938.0, + 1339.0, + 938.0, + 1355.0, + 914.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1340.0, + 997.0, + 1340.0, + 997.0, + 1365.0, + 964.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1327.0, + 1105.0, + 1327.0, + 1105.0, + 1442.0, + 1033.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1338.0, + 1130.0, + 1338.0, + 1130.0, + 1355.0, + 1106.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1339.0, + 1156.0, + 1339.0, + 1156.0, + 1355.0, + 1133.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1339.0, + 1180.0, + 1339.0, + 1180.0, + 1355.0, + 1157.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1339.0, + 1207.0, + 1339.0, + 1207.0, + 1355.0, + 1183.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1338.0, + 1231.0, + 1338.0, + 1231.0, + 1355.0, + 1208.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1339.0, + 1257.0, + 1339.0, + 1257.0, + 1355.0, + 1234.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1339.0, + 1283.0, + 1339.0, + 1283.0, + 1356.0, + 1258.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1339.0, + 1341.0, + 1339.0, + 1341.0, + 1367.0, + 1308.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1363.0, + 417.0, + 1363.0, + 417.0, + 1384.0, + 363.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1365.0, + 441.0, + 1365.0, + 441.0, + 1379.0, + 421.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1364.0, + 493.0, + 1364.0, + 493.0, + 1380.0, + 445.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1364.0, + 518.0, + 1364.0, + 518.0, + 1380.0, + 495.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1365.0, + 542.0, + 1365.0, + 542.0, + 1379.0, + 522.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1364.0, + 569.0, + 1364.0, + 569.0, + 1380.0, + 546.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1364.0, + 594.0, + 1364.0, + 594.0, + 1380.0, + 571.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1370.0, + 652.0, + 1370.0, + 652.0, + 1392.0, + 623.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1362.0, + 761.0, + 1362.0, + 761.0, + 1384.0, + 706.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1364.0, + 787.0, + 1364.0, + 787.0, + 1380.0, + 763.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1365.0, + 810.0, + 1365.0, + 810.0, + 1379.0, + 790.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1365.0, + 834.0, + 1365.0, + 834.0, + 1379.0, + 816.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1365.0, + 861.0, + 1365.0, + 861.0, + 1379.0, + 841.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1364.0, + 888.0, + 1364.0, + 888.0, + 1380.0, + 864.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1364.0, + 913.0, + 1364.0, + 913.0, + 1380.0, + 890.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1364.0, + 938.0, + 1364.0, + 938.0, + 1380.0, + 914.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1370.0, + 996.0, + 1370.0, + 996.0, + 1392.0, + 967.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1363.0, + 1105.0, + 1363.0, + 1105.0, + 1384.0, + 1049.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1365.0, + 1128.0, + 1365.0, + 1128.0, + 1379.0, + 1108.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1365.0, + 1154.0, + 1365.0, + 1154.0, + 1379.0, + 1134.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1365.0, + 1178.0, + 1365.0, + 1178.0, + 1379.0, + 1159.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1364.0, + 1231.0, + 1364.0, + 1231.0, + 1380.0, + 1208.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1364.0, + 1257.0, + 1364.0, + 1257.0, + 1380.0, + 1234.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1364.0, + 1281.0, + 1364.0, + 1281.0, + 1380.0, + 1258.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 1370.0, + 1339.0, + 1370.0, + 1339.0, + 1392.0, + 1310.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1388.0, + 417.0, + 1388.0, + 417.0, + 1410.0, + 362.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1389.0, + 442.0, + 1389.0, + 442.0, + 1407.0, + 420.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1389.0, + 468.0, + 1389.0, + 468.0, + 1407.0, + 445.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1389.0, + 493.0, + 1389.0, + 493.0, + 1407.0, + 470.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1389.0, + 518.0, + 1389.0, + 518.0, + 1407.0, + 495.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1389.0, + 543.0, + 1389.0, + 543.0, + 1407.0, + 521.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1391.0, + 569.0, + 1391.0, + 569.0, + 1407.0, + 546.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1389.0, + 595.0, + 1389.0, + 595.0, + 1407.0, + 571.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1397.0, + 652.0, + 1397.0, + 652.0, + 1419.0, + 623.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1388.0, + 761.0, + 1388.0, + 761.0, + 1410.0, + 705.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1389.0, + 787.0, + 1389.0, + 787.0, + 1407.0, + 763.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1389.0, + 812.0, + 1389.0, + 812.0, + 1407.0, + 789.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1389.0, + 837.0, + 1389.0, + 837.0, + 1407.0, + 813.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1391.0, + 862.0, + 1391.0, + 862.0, + 1407.0, + 839.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1389.0, + 887.0, + 1389.0, + 887.0, + 1407.0, + 864.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1389.0, + 913.0, + 1389.0, + 913.0, + 1407.0, + 890.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1389.0, + 938.0, + 1389.0, + 938.0, + 1407.0, + 914.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1399.0, + 995.0, + 1399.0, + 995.0, + 1419.0, + 967.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1388.0, + 1105.0, + 1388.0, + 1105.0, + 1410.0, + 1049.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1391.0, + 1130.0, + 1391.0, + 1130.0, + 1407.0, + 1107.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1392.0, + 1155.0, + 1392.0, + 1155.0, + 1404.0, + 1134.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1391.0, + 1182.0, + 1391.0, + 1182.0, + 1407.0, + 1157.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1391.0, + 1206.0, + 1391.0, + 1206.0, + 1407.0, + 1183.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1389.0, + 1230.0, + 1389.0, + 1230.0, + 1407.0, + 1208.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1389.0, + 1257.0, + 1389.0, + 1257.0, + 1407.0, + 1234.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1389.0, + 1283.0, + 1389.0, + 1283.0, + 1407.0, + 1258.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1395.0, + 1339.0, + 1395.0, + 1339.0, + 1421.0, + 1309.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1413.0, + 395.0, + 1413.0, + 395.0, + 1434.0, + 363.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1417.0, + 415.0, + 1417.0, + 415.0, + 1429.0, + 396.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1416.0, + 441.0, + 1416.0, + 441.0, + 1429.0, + 422.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1416.0, + 467.0, + 1416.0, + 467.0, + 1429.0, + 447.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1414.0, + 493.0, + 1414.0, + 493.0, + 1432.0, + 470.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1414.0, + 518.0, + 1414.0, + 518.0, + 1432.0, + 495.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1416.0, + 544.0, + 1416.0, + 544.0, + 1432.0, + 519.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1417.0, + 567.0, + 1417.0, + 567.0, + 1430.0, + 547.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1416.0, + 593.0, + 1416.0, + 593.0, + 1429.0, + 573.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1412.0, + 739.0, + 1412.0, + 739.0, + 1434.0, + 706.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1417.0, + 760.0, + 1417.0, + 760.0, + 1429.0, + 740.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1416.0, + 784.0, + 1416.0, + 784.0, + 1429.0, + 766.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1416.0, + 810.0, + 1416.0, + 810.0, + 1429.0, + 790.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1414.0, + 837.0, + 1414.0, + 837.0, + 1432.0, + 814.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1414.0, + 862.0, + 1414.0, + 862.0, + 1432.0, + 840.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1417.0, + 885.0, + 1417.0, + 885.0, + 1430.0, + 866.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1414.0, + 912.0, + 1414.0, + 912.0, + 1432.0, + 891.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1414.0, + 939.0, + 1414.0, + 939.0, + 1432.0, + 916.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1413.0, + 1083.0, + 1413.0, + 1083.0, + 1434.0, + 1050.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1416.0, + 1104.0, + 1416.0, + 1104.0, + 1429.0, + 1084.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1416.0, + 1128.0, + 1416.0, + 1128.0, + 1429.0, + 1110.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1416.0, + 1155.0, + 1416.0, + 1155.0, + 1429.0, + 1135.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1414.0, + 1180.0, + 1414.0, + 1180.0, + 1432.0, + 1158.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1414.0, + 1206.0, + 1414.0, + 1206.0, + 1432.0, + 1184.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1416.0, + 1230.0, + 1416.0, + 1230.0, + 1432.0, + 1207.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1414.0, + 1257.0, + 1414.0, + 1257.0, + 1432.0, + 1234.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 1414.0, + 1283.0, + 1414.0, + 1283.0, + 1432.0, + 1259.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1426.0, + 652.0, + 1426.0, + 652.0, + 1446.0, + 622.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1426.0, + 996.0, + 1426.0, + 996.0, + 1446.0, + 965.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1426.0, + 1339.0, + 1426.0, + 1339.0, + 1446.0, + 1309.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1434.0, + 597.0, + 1434.0, + 597.0, + 1455.0, + 391.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1434.0, + 941.0, + 1434.0, + 941.0, + 1455.0, + 734.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1434.0, + 1285.0, + 1434.0, + 1285.0, + 1455.0, + 1078.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1447.0, + 583.0, + 1447.0, + 583.0, + 1470.0, + 403.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1449.0, + 926.0, + 1449.0, + 926.0, + 1469.0, + 750.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 1449.0, + 1270.0, + 1449.0, + 1270.0, + 1469.0, + 1092.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1551.0, + 727.0, + 1551.0, + 727.0, + 1572.0, + 590.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1551.0, + 1068.0, + 1551.0, + 1068.0, + 1573.0, + 938.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1566.0, + 536.0, + 1566.0, + 536.0, + 1752.0, + 510.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 1566.0, + 833.0, + 1566.0, + 833.0, + 1748.0, + 808.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1566.0, + 881.0, + 1566.0, + 881.0, + 1752.0, + 854.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1569.0, + 1176.0, + 1569.0, + 1176.0, + 1748.0, + 1154.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1572.0, + 560.0, + 1572.0, + 560.0, + 1593.0, + 530.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1572.0, + 817.0, + 1572.0, + 817.0, + 1592.0, + 788.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1572.0, + 904.0, + 1572.0, + 904.0, + 1593.0, + 874.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1572.0, + 1161.0, + 1572.0, + 1161.0, + 1592.0, + 1132.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1573.0, + 581.0, + 1573.0, + 581.0, + 1590.0, + 560.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1573.0, + 608.0, + 1573.0, + 608.0, + 1590.0, + 586.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1573.0, + 658.0, + 1573.0, + 658.0, + 1590.0, + 636.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1573.0, + 683.0, + 1573.0, + 683.0, + 1590.0, + 660.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1573.0, + 709.0, + 1573.0, + 709.0, + 1589.0, + 686.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1573.0, + 733.0, + 1573.0, + 733.0, + 1589.0, + 710.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1573.0, + 761.0, + 1573.0, + 761.0, + 1590.0, + 736.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1573.0, + 926.0, + 1573.0, + 926.0, + 1590.0, + 904.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1573.0, + 1001.0, + 1573.0, + 1001.0, + 1590.0, + 978.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1573.0, + 1027.0, + 1573.0, + 1027.0, + 1590.0, + 1004.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1573.0, + 1053.0, + 1573.0, + 1053.0, + 1590.0, + 1029.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1573.0, + 1104.0, + 1573.0, + 1104.0, + 1590.0, + 1079.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1574.0, + 632.0, + 1574.0, + 632.0, + 1590.0, + 609.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1574.0, + 977.0, + 1574.0, + 977.0, + 1590.0, + 954.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1574.0, + 1076.0, + 1574.0, + 1076.0, + 1588.0, + 1056.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1575.0, + 949.0, + 1575.0, + 949.0, + 1589.0, + 929.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1596.0, + 559.0, + 1596.0, + 559.0, + 1618.0, + 531.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1598.0, + 581.0, + 1598.0, + 581.0, + 1615.0, + 560.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1598.0, + 607.0, + 1598.0, + 607.0, + 1615.0, + 585.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1598.0, + 633.0, + 1598.0, + 633.0, + 1614.0, + 609.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1598.0, + 658.0, + 1598.0, + 658.0, + 1615.0, + 636.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1600.0, + 682.0, + 1600.0, + 682.0, + 1613.0, + 661.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1598.0, + 709.0, + 1598.0, + 709.0, + 1614.0, + 686.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1600.0, + 732.0, + 1600.0, + 732.0, + 1613.0, + 711.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1599.0, + 760.0, + 1599.0, + 760.0, + 1615.0, + 736.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1597.0, + 813.0, + 1597.0, + 813.0, + 1618.0, + 789.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1593.0, + 904.0, + 1593.0, + 904.0, + 1619.0, + 873.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1600.0, + 924.0, + 1600.0, + 924.0, + 1613.0, + 906.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1600.0, + 949.0, + 1600.0, + 949.0, + 1614.0, + 931.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1600.0, + 975.0, + 1600.0, + 975.0, + 1614.0, + 955.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1600.0, + 1000.0, + 1600.0, + 1000.0, + 1614.0, + 981.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1598.0, + 1027.0, + 1598.0, + 1027.0, + 1615.0, + 1004.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1598.0, + 1053.0, + 1598.0, + 1053.0, + 1615.0, + 1029.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1598.0, + 1078.0, + 1598.0, + 1078.0, + 1615.0, + 1055.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1599.0, + 1104.0, + 1599.0, + 1104.0, + 1615.0, + 1079.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1593.0, + 1162.0, + 1593.0, + 1162.0, + 1621.0, + 1130.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1620.0, + 561.0, + 1620.0, + 561.0, + 1644.0, + 532.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1625.0, + 581.0, + 1625.0, + 581.0, + 1639.0, + 562.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1625.0, + 607.0, + 1625.0, + 607.0, + 1639.0, + 587.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1625.0, + 631.0, + 1625.0, + 631.0, + 1639.0, + 611.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1624.0, + 659.0, + 1624.0, + 659.0, + 1640.0, + 636.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1624.0, + 683.0, + 1624.0, + 683.0, + 1640.0, + 660.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 1625.0, + 708.0, + 1625.0, + 708.0, + 1639.0, + 689.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1624.0, + 734.0, + 1624.0, + 734.0, + 1640.0, + 711.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1624.0, + 760.0, + 1624.0, + 760.0, + 1640.0, + 737.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1624.0, + 817.0, + 1624.0, + 817.0, + 1646.0, + 788.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1616.0, + 905.0, + 1616.0, + 905.0, + 1645.0, + 874.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1625.0, + 924.0, + 1625.0, + 924.0, + 1639.0, + 906.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1625.0, + 949.0, + 1625.0, + 949.0, + 1639.0, + 931.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1625.0, + 975.0, + 1625.0, + 975.0, + 1639.0, + 956.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1625.0, + 1000.0, + 1625.0, + 1000.0, + 1639.0, + 982.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1624.0, + 1027.0, + 1624.0, + 1027.0, + 1640.0, + 1004.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1624.0, + 1053.0, + 1624.0, + 1053.0, + 1640.0, + 1029.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1624.0, + 1078.0, + 1624.0, + 1078.0, + 1640.0, + 1055.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1624.0, + 1104.0, + 1624.0, + 1104.0, + 1640.0, + 1080.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1621.0, + 1162.0, + 1621.0, + 1162.0, + 1648.0, + 1130.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1648.0, + 580.0, + 1648.0, + 580.0, + 1668.0, + 528.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1650.0, + 605.0, + 1650.0, + 605.0, + 1664.0, + 586.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1649.0, + 632.0, + 1649.0, + 632.0, + 1666.0, + 610.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1649.0, + 659.0, + 1649.0, + 659.0, + 1665.0, + 636.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1649.0, + 683.0, + 1649.0, + 683.0, + 1665.0, + 660.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1649.0, + 709.0, + 1649.0, + 709.0, + 1665.0, + 686.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1650.0, + 732.0, + 1650.0, + 732.0, + 1664.0, + 712.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1649.0, + 760.0, + 1649.0, + 760.0, + 1665.0, + 737.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1651.0, + 816.0, + 1651.0, + 816.0, + 1674.0, + 788.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1646.0, + 924.0, + 1646.0, + 924.0, + 1671.0, + 870.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1650.0, + 949.0, + 1650.0, + 949.0, + 1664.0, + 931.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1650.0, + 975.0, + 1650.0, + 975.0, + 1664.0, + 955.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1650.0, + 1000.0, + 1650.0, + 1000.0, + 1664.0, + 982.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1648.0, + 1053.0, + 1648.0, + 1053.0, + 1665.0, + 1031.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1648.0, + 1077.0, + 1648.0, + 1077.0, + 1666.0, + 1055.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1649.0, + 1103.0, + 1649.0, + 1103.0, + 1666.0, + 1080.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1648.0, + 1162.0, + 1648.0, + 1162.0, + 1675.0, + 1130.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1672.0, + 580.0, + 1672.0, + 580.0, + 1693.0, + 528.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1675.0, + 605.0, + 1675.0, + 605.0, + 1689.0, + 586.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1675.0, + 631.0, + 1675.0, + 631.0, + 1689.0, + 612.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1675.0, + 657.0, + 1675.0, + 657.0, + 1689.0, + 637.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1673.0, + 683.0, + 1673.0, + 683.0, + 1690.0, + 660.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1674.0, + 709.0, + 1674.0, + 709.0, + 1690.0, + 686.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1674.0, + 733.0, + 1674.0, + 733.0, + 1691.0, + 710.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1675.0, + 758.0, + 1675.0, + 758.0, + 1689.0, + 738.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1682.0, + 815.0, + 1682.0, + 815.0, + 1698.0, + 786.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1672.0, + 904.0, + 1672.0, + 904.0, + 1693.0, + 871.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1675.0, + 924.0, + 1675.0, + 924.0, + 1689.0, + 905.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1675.0, + 949.0, + 1675.0, + 949.0, + 1689.0, + 929.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1675.0, + 975.0, + 1675.0, + 975.0, + 1689.0, + 956.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1675.0, + 1000.0, + 1675.0, + 1000.0, + 1689.0, + 982.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1675.0, + 1025.0, + 1675.0, + 1025.0, + 1689.0, + 1006.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1675.0, + 1050.0, + 1675.0, + 1050.0, + 1689.0, + 1032.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 1675.0, + 1076.0, + 1675.0, + 1076.0, + 1689.0, + 1057.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1675.0, + 1101.0, + 1675.0, + 1101.0, + 1689.0, + 1082.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1679.0, + 1161.0, + 1679.0, + 1161.0, + 1701.0, + 1132.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1698.0, + 559.0, + 1698.0, + 559.0, + 1720.0, + 528.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1699.0, + 581.0, + 1699.0, + 581.0, + 1716.0, + 560.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1701.0, + 605.0, + 1701.0, + 605.0, + 1715.0, + 587.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1699.0, + 632.0, + 1699.0, + 632.0, + 1716.0, + 610.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1701.0, + 657.0, + 1701.0, + 657.0, + 1714.0, + 637.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1699.0, + 682.0, + 1699.0, + 682.0, + 1716.0, + 660.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1699.0, + 734.0, + 1699.0, + 734.0, + 1716.0, + 686.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1699.0, + 758.0, + 1699.0, + 758.0, + 1716.0, + 737.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1698.0, + 903.0, + 1698.0, + 903.0, + 1720.0, + 871.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1699.0, + 925.0, + 1699.0, + 925.0, + 1716.0, + 904.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1699.0, + 950.0, + 1699.0, + 950.0, + 1716.0, + 929.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1699.0, + 976.0, + 1699.0, + 976.0, + 1716.0, + 954.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1701.0, + 999.0, + 1701.0, + 999.0, + 1715.0, + 982.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1699.0, + 1026.0, + 1699.0, + 1026.0, + 1716.0, + 1004.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1699.0, + 1051.0, + 1699.0, + 1051.0, + 1716.0, + 1031.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1699.0, + 1077.0, + 1699.0, + 1077.0, + 1716.0, + 1055.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1699.0, + 1103.0, + 1699.0, + 1103.0, + 1716.0, + 1080.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1708.0, + 816.0, + 1708.0, + 816.0, + 1729.0, + 788.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1708.0, + 1161.0, + 1708.0, + 1161.0, + 1729.0, + 1132.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1722.0, + 560.0, + 1722.0, + 560.0, + 1744.0, + 528.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1725.0, + 580.0, + 1725.0, + 580.0, + 1739.0, + 562.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1727.0, + 605.0, + 1727.0, + 605.0, + 1739.0, + 587.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1725.0, + 630.0, + 1725.0, + 630.0, + 1739.0, + 612.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1727.0, + 657.0, + 1727.0, + 657.0, + 1739.0, + 638.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1727.0, + 681.0, + 1727.0, + 681.0, + 1739.0, + 662.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1725.0, + 706.0, + 1725.0, + 706.0, + 1739.0, + 688.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1725.0, + 732.0, + 1725.0, + 732.0, + 1739.0, + 712.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1727.0, + 758.0, + 1727.0, + 758.0, + 1739.0, + 738.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1725.0, + 925.0, + 1725.0, + 925.0, + 1739.0, + 906.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1727.0, + 949.0, + 1727.0, + 949.0, + 1739.0, + 931.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1725.0, + 974.0, + 1725.0, + 974.0, + 1739.0, + 956.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1727.0, + 1000.0, + 1727.0, + 1000.0, + 1739.0, + 982.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1724.0, + 1026.0, + 1724.0, + 1026.0, + 1741.0, + 1005.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1725.0, + 1050.0, + 1725.0, + 1050.0, + 1739.0, + 1032.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1727.0, + 1076.0, + 1727.0, + 1076.0, + 1739.0, + 1056.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1724.0, + 1104.0, + 1724.0, + 1104.0, + 1741.0, + 1080.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1736.0, + 817.0, + 1736.0, + 817.0, + 1756.0, + 787.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1736.0, + 1161.0, + 1736.0, + 1161.0, + 1756.0, + 1132.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1744.0, + 762.0, + 1744.0, + 762.0, + 1765.0, + 555.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1744.0, + 1105.0, + 1744.0, + 1105.0, + 1765.0, + 898.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1757.0, + 748.0, + 1757.0, + 748.0, + 1780.0, + 568.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1757.0, + 1092.0, + 1757.0, + 1092.0, + 1780.0, + 912.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.75, + 1313.0, + 911.75, + 1313.0, + 911.75, + 1330.0, + 890.75, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1365.5, + 1205.0, + 1365.5, + 1205.0, + 1378.0, + 1183.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.25, + 1648.0, + 1027.25, + 1648.0, + 1027.25, + 1665.0, + 1001.25, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1720.0, + 900.0, + 1720.0, + 900.0, + 1747.0, + 874.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 1408.0, + 1849.0, + 1408.0, + 1888.0, + 292.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1908.0, + 1125.0, + 1908.0, + 1125.0, + 1946.0, + 291.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1908.0, + 1250.0, + 1908.0, + 1250.0, + 1946.0, + 1239.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 938.0, + 1407.0, + 938.0, + 1407.0, + 978.0, + 292.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 969.0, + 1405.0, + 969.0, + 1405.0, + 1006.0, + 294.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 997.0, + 1108.0, + 997.0, + 1108.0, + 1036.0, + 292.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 997.0, + 1235.0, + 997.0, + 1235.0, + 1036.0, + 1222.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 830.0, + 2125.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 31, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 348, + 294, + 1357, + 294, + 1357, + 898, + 348, + 898 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 347, + 1208, + 1357, + 1208, + 1357, + 1808, + 347, + 1808 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 296, + 1850, + 1404, + 1850, + 1404, + 1945, + 296, + 1945 + ], + "score": 0.953 + }, + { + "category_id": 4, + "poly": [ + 295, + 941, + 1409, + 941, + 1409, + 1033, + 295, + 1033 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.856 + }, + { + "category_id": 13, + "poly": [ + 1117, + 1910, + 1229, + 1910, + 1229, + 1940, + 1117, + 1940 + ], + "score": 0.92, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1126, + 1000, + 1238, + 1000, + 1238, + 1029, + 1126, + 1029 + ], + "score": 0.91, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 301.0, + 576.0, + 301.0, + 576.0, + 324.0, + 417.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 312.0, + 651.0, + 312.0, + 651.0, + 330.0, + 627.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 301.0, + 913.0, + 301.0, + 913.0, + 324.0, + 769.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 312.0, + 996.0, + 312.0, + 996.0, + 330.0, + 971.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 301.0, + 1251.0, + 301.0, + 1251.0, + 325.0, + 1117.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 312.0, + 1339.0, + 312.0, + 1339.0, + 330.0, + 1316.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 323.0, + 399.0, + 323.0, + 399.0, + 344.0, + 366.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 326.0, + 442.0, + 326.0, + 442.0, + 339.0, + 422.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 326.0, + 469.0, + 326.0, + 469.0, + 339.0, + 448.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 324.0, + 496.0, + 324.0, + 496.0, + 341.0, + 472.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 325.0, + 522.0, + 325.0, + 522.0, + 341.0, + 498.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 325.0, + 548.0, + 325.0, + 548.0, + 341.0, + 523.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 324.0, + 573.0, + 324.0, + 573.0, + 340.0, + 550.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 322.0, + 599.0, + 322.0, + 599.0, + 343.0, + 576.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 323.0, + 762.0, + 323.0, + 762.0, + 344.0, + 710.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 324.0, + 814.0, + 324.0, + 814.0, + 340.0, + 791.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 325.0, + 837.0, + 325.0, + 837.0, + 339.0, + 818.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 324.0, + 865.0, + 324.0, + 865.0, + 340.0, + 842.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 326.0, + 891.0, + 326.0, + 891.0, + 339.0, + 870.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 324.0, + 916.0, + 324.0, + 916.0, + 341.0, + 893.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 324.0, + 943.0, + 324.0, + 943.0, + 341.0, + 920.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 324.0, + 1107.0, + 324.0, + 1107.0, + 344.0, + 1055.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 325.0, + 1130.0, + 325.0, + 1130.0, + 339.0, + 1110.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 326.0, + 1156.0, + 326.0, + 1156.0, + 339.0, + 1136.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 325.0, + 1182.0, + 325.0, + 1182.0, + 339.0, + 1162.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 326.0, + 1207.0, + 326.0, + 1207.0, + 339.0, + 1188.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 326.0, + 1234.0, + 326.0, + 1234.0, + 339.0, + 1213.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 325.0, + 1259.0, + 325.0, + 1259.0, + 339.0, + 1239.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 324.0, + 1286.0, + 324.0, + 1286.0, + 341.0, + 1261.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 345.0, + 418.0, + 345.0, + 418.0, + 397.0, + 347.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 352.0, + 442.0, + 352.0, + 442.0, + 366.0, + 423.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 352.0, + 468.0, + 352.0, + 468.0, + 366.0, + 448.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 349.0, + 496.0, + 349.0, + 496.0, + 367.0, + 474.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 349.0, + 521.0, + 349.0, + 521.0, + 367.0, + 498.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 349.0, + 546.0, + 349.0, + 546.0, + 367.0, + 524.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 349.0, + 573.0, + 349.0, + 573.0, + 367.0, + 550.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 352.0, + 596.0, + 352.0, + 596.0, + 366.0, + 576.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 337.0, + 651.0, + 337.0, + 651.0, + 355.0, + 628.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 341.0, + 762.0, + 341.0, + 762.0, + 431.0, + 691.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 349.0, + 788.0, + 349.0, + 788.0, + 367.0, + 765.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 351.0, + 814.0, + 351.0, + 814.0, + 367.0, + 791.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 352.0, + 837.0, + 352.0, + 837.0, + 366.0, + 818.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 349.0, + 864.0, + 349.0, + 864.0, + 367.0, + 842.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 349.0, + 892.0, + 349.0, + 892.0, + 367.0, + 868.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 349.0, + 916.0, + 349.0, + 916.0, + 367.0, + 894.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 351.0, + 943.0, + 351.0, + 943.0, + 367.0, + 920.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 335.0, + 996.0, + 335.0, + 996.0, + 357.0, + 970.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 343.0, + 1105.0, + 343.0, + 1105.0, + 396.0, + 1035.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 352.0, + 1130.0, + 352.0, + 1130.0, + 366.0, + 1110.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 351.0, + 1157.0, + 351.0, + 1157.0, + 367.0, + 1134.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 352.0, + 1182.0, + 352.0, + 1182.0, + 366.0, + 1162.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 352.0, + 1207.0, + 352.0, + 1207.0, + 366.0, + 1189.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 351.0, + 1235.0, + 351.0, + 1235.0, + 367.0, + 1212.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 351.0, + 1261.0, + 351.0, + 1261.0, + 367.0, + 1237.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 349.0, + 1286.0, + 349.0, + 1286.0, + 367.0, + 1263.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 337.0, + 1340.0, + 337.0, + 1340.0, + 355.0, + 1316.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 363.0, + 651.0, + 363.0, + 651.0, + 382.0, + 628.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 363.0, + 996.0, + 363.0, + 996.0, + 382.0, + 973.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 363.0, + 1340.0, + 363.0, + 1340.0, + 382.0, + 1316.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 376.0, + 418.0, + 376.0, + 418.0, + 393.0, + 367.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 377.0, + 442.0, + 377.0, + 442.0, + 391.0, + 423.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 376.0, + 470.0, + 376.0, + 470.0, + 392.0, + 447.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 376.0, + 496.0, + 376.0, + 496.0, + 392.0, + 472.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 377.0, + 520.0, + 377.0, + 520.0, + 391.0, + 500.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 377.0, + 545.0, + 377.0, + 545.0, + 391.0, + 526.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 377.0, + 571.0, + 377.0, + 571.0, + 391.0, + 551.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 376.0, + 598.0, + 376.0, + 598.0, + 392.0, + 575.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 374.0, + 761.0, + 374.0, + 761.0, + 395.0, + 710.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 376.0, + 788.0, + 376.0, + 788.0, + 392.0, + 765.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 376.0, + 813.0, + 376.0, + 813.0, + 392.0, + 790.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 377.0, + 837.0, + 377.0, + 837.0, + 391.0, + 819.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 377.0, + 863.0, + 377.0, + 863.0, + 391.0, + 843.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 377.0, + 889.0, + 377.0, + 889.0, + 391.0, + 870.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 376.0, + 917.0, + 376.0, + 917.0, + 392.0, + 894.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 376.0, + 943.0, + 376.0, + 943.0, + 392.0, + 920.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 375.0, + 1104.0, + 375.0, + 1104.0, + 395.0, + 1054.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 376.0, + 1132.0, + 376.0, + 1132.0, + 392.0, + 1109.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 376.0, + 1157.0, + 376.0, + 1157.0, + 392.0, + 1134.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 377.0, + 1181.0, + 377.0, + 1181.0, + 391.0, + 1162.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 377.0, + 1207.0, + 377.0, + 1207.0, + 391.0, + 1189.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 377.0, + 1233.0, + 377.0, + 1233.0, + 391.0, + 1213.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 376.0, + 1261.0, + 376.0, + 1261.0, + 392.0, + 1238.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 376.0, + 1286.0, + 376.0, + 1286.0, + 392.0, + 1263.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 384.0, + 371.0, + 384.0, + 371.0, + 530.0, + 344.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 389.0, + 651.0, + 389.0, + 651.0, + 406.0, + 628.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 389.0, + 995.0, + 389.0, + 995.0, + 407.0, + 971.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 388.0, + 1055.0, + 388.0, + 1055.0, + 427.0, + 1035.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 389.0, + 1340.0, + 389.0, + 1340.0, + 407.0, + 1316.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 400.0, + 419.0, + 400.0, + 419.0, + 421.0, + 363.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 403.0, + 442.0, + 403.0, + 442.0, + 417.0, + 423.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 401.0, + 470.0, + 401.0, + 470.0, + 418.0, + 447.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 401.0, + 496.0, + 401.0, + 496.0, + 418.0, + 472.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 403.0, + 520.0, + 403.0, + 520.0, + 417.0, + 500.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 401.0, + 548.0, + 401.0, + 548.0, + 418.0, + 524.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 401.0, + 573.0, + 401.0, + 573.0, + 418.0, + 550.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 403.0, + 596.0, + 403.0, + 596.0, + 417.0, + 576.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 401.0, + 668.0, + 401.0, + 668.0, + 470.0, + 646.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 398.0, + 762.0, + 398.0, + 762.0, + 422.0, + 705.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 401.0, + 787.0, + 401.0, + 787.0, + 418.0, + 764.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 401.0, + 816.0, + 401.0, + 816.0, + 419.0, + 788.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 401.0, + 840.0, + 401.0, + 840.0, + 418.0, + 817.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 401.0, + 865.0, + 401.0, + 865.0, + 418.0, + 842.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 403.0, + 889.0, + 403.0, + 889.0, + 417.0, + 870.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 403.0, + 915.0, + 403.0, + 915.0, + 417.0, + 895.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 401.0, + 943.0, + 401.0, + 943.0, + 418.0, + 920.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 400.0, + 1011.0, + 400.0, + 1011.0, + 471.0, + 988.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 402.0, + 1104.0, + 402.0, + 1104.0, + 420.0, + 1048.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 403.0, + 1130.0, + 403.0, + 1130.0, + 417.0, + 1110.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 401.0, + 1159.0, + 401.0, + 1159.0, + 419.0, + 1133.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 403.0, + 1182.0, + 403.0, + 1182.0, + 417.0, + 1162.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 403.0, + 1207.0, + 403.0, + 1207.0, + 417.0, + 1188.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 401.0, + 1235.0, + 401.0, + 1235.0, + 418.0, + 1212.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 401.0, + 1261.0, + 401.0, + 1261.0, + 418.0, + 1238.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 403.0, + 1283.0, + 403.0, + 1283.0, + 417.0, + 1264.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 400.0, + 1355.0, + 400.0, + 1355.0, + 471.0, + 1331.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 426.0, + 445.0, + 426.0, + 445.0, + 447.0, + 364.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 427.0, + 497.0, + 427.0, + 497.0, + 443.0, + 446.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 427.0, + 522.0, + 427.0, + 522.0, + 443.0, + 499.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 429.0, + 546.0, + 429.0, + 546.0, + 442.0, + 524.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 429.0, + 572.0, + 429.0, + 572.0, + 442.0, + 551.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 428.0, + 598.0, + 428.0, + 598.0, + 444.0, + 575.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 413.0, + 653.0, + 413.0, + 653.0, + 435.0, + 627.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 417.0, + 813.0, + 417.0, + 813.0, + 529.0, + 691.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 427.0, + 839.0, + 427.0, + 839.0, + 444.0, + 817.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 427.0, + 866.0, + 427.0, + 866.0, + 443.0, + 841.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 427.0, + 891.0, + 427.0, + 891.0, + 444.0, + 869.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 429.0, + 915.0, + 429.0, + 915.0, + 443.0, + 895.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 427.0, + 941.0, + 427.0, + 941.0, + 444.0, + 920.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 412.0, + 996.0, + 412.0, + 996.0, + 435.0, + 971.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 417.0, + 1132.0, + 417.0, + 1132.0, + 529.0, + 1035.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 429.0, + 1184.0, + 429.0, + 1184.0, + 442.0, + 1134.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 429.0, + 1209.0, + 429.0, + 1209.0, + 442.0, + 1185.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 427.0, + 1235.0, + 427.0, + 1235.0, + 444.0, + 1213.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 427.0, + 1286.0, + 427.0, + 1286.0, + 444.0, + 1238.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 413.0, + 1340.0, + 413.0, + 1340.0, + 435.0, + 1315.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 439.0, + 653.0, + 439.0, + 653.0, + 461.0, + 627.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 440.0, + 996.0, + 440.0, + 996.0, + 458.0, + 971.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 439.0, + 1341.0, + 439.0, + 1341.0, + 461.0, + 1313.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 451.0, + 549.0, + 451.0, + 549.0, + 473.0, + 363.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 454.0, + 573.0, + 454.0, + 573.0, + 470.0, + 550.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 454.0, + 598.0, + 454.0, + 598.0, + 470.0, + 574.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 451.0, + 842.0, + 451.0, + 842.0, + 473.0, + 707.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 455.0, + 864.0, + 455.0, + 864.0, + 469.0, + 844.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 452.0, + 893.0, + 452.0, + 893.0, + 470.0, + 866.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 454.0, + 917.0, + 454.0, + 917.0, + 470.0, + 894.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 454.0, + 943.0, + 454.0, + 943.0, + 470.0, + 918.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 451.0, + 1208.0, + 451.0, + 1208.0, + 472.0, + 1050.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 452.0, + 1236.0, + 452.0, + 1236.0, + 470.0, + 1211.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 454.0, + 1286.0, + 454.0, + 1286.0, + 470.0, + 1238.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 466.0, + 651.0, + 466.0, + 651.0, + 485.0, + 628.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 466.0, + 996.0, + 466.0, + 996.0, + 485.0, + 973.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 466.0, + 1340.0, + 466.0, + 1340.0, + 485.0, + 1316.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 478.0, + 522.0, + 478.0, + 522.0, + 499.0, + 362.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 478.0, + 573.0, + 478.0, + 573.0, + 495.0, + 523.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 479.0, + 598.0, + 479.0, + 598.0, + 495.0, + 575.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 477.0, + 894.0, + 477.0, + 894.0, + 499.0, + 706.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 480.0, + 915.0, + 480.0, + 915.0, + 494.0, + 895.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 479.0, + 943.0, + 479.0, + 943.0, + 495.0, + 920.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 478.0, + 1208.0, + 478.0, + 1208.0, + 499.0, + 1049.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 478.0, + 1236.0, + 478.0, + 1236.0, + 495.0, + 1209.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 479.0, + 1285.0, + 479.0, + 1285.0, + 496.0, + 1237.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 492.0, + 651.0, + 492.0, + 651.0, + 510.0, + 628.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 489.0, + 997.0, + 489.0, + 997.0, + 511.0, + 971.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 491.0, + 1340.0, + 491.0, + 1340.0, + 513.0, + 1315.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 500.0, + 598.0, + 500.0, + 598.0, + 526.0, + 363.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 502.0, + 866.0, + 502.0, + 866.0, + 524.0, + 707.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 505.0, + 892.0, + 505.0, + 892.0, + 522.0, + 868.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 505.0, + 917.0, + 505.0, + 917.0, + 522.0, + 893.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 505.0, + 943.0, + 505.0, + 943.0, + 521.0, + 918.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 502.0, + 1286.0, + 502.0, + 1286.0, + 524.0, + 1050.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 517.0, + 651.0, + 517.0, + 651.0, + 536.0, + 628.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 517.0, + 996.0, + 517.0, + 996.0, + 536.0, + 973.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 516.0, + 1340.0, + 516.0, + 1340.0, + 538.0, + 1315.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 529.0, + 548.0, + 529.0, + 548.0, + 550.0, + 363.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 530.0, + 573.0, + 530.0, + 573.0, + 546.0, + 550.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 531.0, + 598.0, + 531.0, + 598.0, + 547.0, + 575.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 529.0, + 764.0, + 529.0, + 764.0, + 550.0, + 706.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 529.0, + 893.0, + 529.0, + 893.0, + 547.0, + 765.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 530.0, + 917.0, + 530.0, + 917.0, + 547.0, + 894.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 530.0, + 943.0, + 530.0, + 943.0, + 547.0, + 918.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 529.0, + 1236.0, + 529.0, + 1236.0, + 550.0, + 1050.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 530.0, + 1286.0, + 530.0, + 1286.0, + 547.0, + 1238.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 551.0, + 601.0, + 551.0, + 601.0, + 569.0, + 392.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 547.0, + 640.0, + 547.0, + 640.0, + 558.0, + 630.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 546.0, + 945.0, + 546.0, + 945.0, + 572.0, + 731.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 547.0, + 984.0, + 547.0, + 984.0, + 558.0, + 975.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 550.0, + 1289.0, + 550.0, + 1289.0, + 569.0, + 1078.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 546.0, + 1328.0, + 546.0, + 1328.0, + 558.0, + 1318.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 562.0, + 586.0, + 562.0, + 586.0, + 587.0, + 407.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 562.0, + 930.0, + 562.0, + 930.0, + 587.0, + 750.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 562.0, + 1273.0, + 562.0, + 1273.0, + 587.0, + 1094.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 612.0, + 732.0, + 612.0, + 732.0, + 634.0, + 591.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 623.0, + 817.0, + 623.0, + 817.0, + 640.0, + 791.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 610.0, + 1071.0, + 610.0, + 1071.0, + 636.0, + 940.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 623.0, + 1160.0, + 623.0, + 1160.0, + 640.0, + 1137.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 633.0, + 611.0, + 633.0, + 611.0, + 654.0, + 533.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 637.0, + 634.0, + 637.0, + 634.0, + 649.0, + 613.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 634.0, + 661.0, + 634.0, + 661.0, + 650.0, + 638.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 637.0, + 685.0, + 637.0, + 685.0, + 649.0, + 664.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 634.0, + 712.0, + 634.0, + 712.0, + 650.0, + 688.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 637.0, + 737.0, + 637.0, + 737.0, + 649.0, + 715.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 634.0, + 764.0, + 634.0, + 764.0, + 650.0, + 740.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 633.0, + 980.0, + 633.0, + 980.0, + 654.0, + 876.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 634.0, + 1005.0, + 634.0, + 1005.0, + 650.0, + 982.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 637.0, + 1029.0, + 637.0, + 1029.0, + 649.0, + 1007.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 637.0, + 1055.0, + 637.0, + 1055.0, + 649.0, + 1033.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 634.0, + 1081.0, + 634.0, + 1081.0, + 650.0, + 1058.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 635.0, + 1105.0, + 635.0, + 1105.0, + 649.0, + 1086.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 653.0, + 662.0, + 653.0, + 662.0, + 708.0, + 513.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 661.0, + 686.0, + 661.0, + 686.0, + 677.0, + 663.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 662.0, + 710.0, + 662.0, + 710.0, + 676.0, + 691.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 662.0, + 737.0, + 662.0, + 737.0, + 675.0, + 716.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 660.0, + 764.0, + 660.0, + 764.0, + 677.0, + 740.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 647.0, + 817.0, + 647.0, + 817.0, + 665.0, + 792.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 654.0, + 1032.0, + 654.0, + 1032.0, + 740.0, + 856.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 660.0, + 1056.0, + 660.0, + 1056.0, + 677.0, + 1033.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 660.0, + 1081.0, + 660.0, + 1081.0, + 677.0, + 1058.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 660.0, + 1107.0, + 660.0, + 1107.0, + 677.0, + 1084.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 646.0, + 1162.0, + 646.0, + 1162.0, + 668.0, + 1136.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 672.0, + 817.0, + 672.0, + 817.0, + 691.0, + 792.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 671.0, + 1162.0, + 671.0, + 1162.0, + 693.0, + 1136.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 682.0, + 662.0, + 682.0, + 662.0, + 707.0, + 532.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 686.0, + 686.0, + 686.0, + 686.0, + 703.0, + 663.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 687.0, + 710.0, + 687.0, + 710.0, + 701.0, + 691.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 686.0, + 738.0, + 686.0, + 738.0, + 703.0, + 715.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 686.0, + 764.0, + 686.0, + 764.0, + 703.0, + 739.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 681.0, + 1056.0, + 681.0, + 1056.0, + 707.0, + 875.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 686.0, + 1081.0, + 686.0, + 1081.0, + 703.0, + 1058.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 685.0, + 1107.0, + 685.0, + 1107.0, + 703.0, + 1084.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 695.0, + 534.0, + 695.0, + 534.0, + 739.0, + 513.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 699.0, + 817.0, + 699.0, + 817.0, + 716.0, + 791.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 698.0, + 1160.0, + 698.0, + 1160.0, + 716.0, + 1137.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 711.0, + 687.0, + 711.0, + 687.0, + 731.0, + 528.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 713.0, + 710.0, + 713.0, + 710.0, + 727.0, + 691.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 712.0, + 738.0, + 712.0, + 738.0, + 728.0, + 715.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 712.0, + 764.0, + 712.0, + 764.0, + 728.0, + 740.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 710.0, + 832.0, + 710.0, + 832.0, + 780.0, + 810.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 709.0, + 1056.0, + 709.0, + 1056.0, + 731.0, + 871.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 712.0, + 1082.0, + 712.0, + 1082.0, + 728.0, + 1058.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 713.0, + 1105.0, + 713.0, + 1105.0, + 727.0, + 1086.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 711.0, + 1175.0, + 711.0, + 1175.0, + 779.0, + 1156.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 728.0, + 712.0, + 728.0, + 712.0, + 839.0, + 512.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 738.0, + 736.0, + 738.0, + 736.0, + 752.0, + 716.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 737.0, + 764.0, + 737.0, + 764.0, + 755.0, + 739.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 725.0, + 817.0, + 725.0, + 817.0, + 743.0, + 792.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 728.0, + 1032.0, + 728.0, + 1032.0, + 839.0, + 856.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 737.0, + 1056.0, + 737.0, + 1056.0, + 753.0, + 1033.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 737.0, + 1081.0, + 737.0, + 1081.0, + 753.0, + 1058.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 737.0, + 1108.0, + 737.0, + 1108.0, + 753.0, + 1085.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 722.0, + 1161.0, + 722.0, + 1161.0, + 745.0, + 1136.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 749.0, + 819.0, + 749.0, + 819.0, + 770.0, + 791.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 750.0, + 1161.0, + 750.0, + 1161.0, + 769.0, + 1137.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 762.0, + 688.0, + 762.0, + 688.0, + 782.0, + 528.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 763.0, + 712.0, + 763.0, + 712.0, + 780.0, + 690.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 763.0, + 738.0, + 763.0, + 738.0, + 779.0, + 715.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 763.0, + 764.0, + 763.0, + 764.0, + 780.0, + 739.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 762.0, + 925.0, + 762.0, + 925.0, + 782.0, + 872.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 763.0, + 1056.0, + 763.0, + 1056.0, + 780.0, + 929.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 765.0, + 1081.0, + 765.0, + 1081.0, + 778.0, + 1060.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 763.0, + 1107.0, + 763.0, + 1107.0, + 780.0, + 1082.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 774.0, + 818.0, + 774.0, + 818.0, + 796.0, + 791.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 777.0, + 1161.0, + 777.0, + 1161.0, + 795.0, + 1137.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 786.0, + 714.0, + 786.0, + 714.0, + 810.0, + 527.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 789.0, + 738.0, + 789.0, + 738.0, + 806.0, + 715.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 789.0, + 762.0, + 789.0, + 762.0, + 806.0, + 739.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 788.0, + 926.0, + 788.0, + 926.0, + 809.0, + 872.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 791.0, + 951.0, + 791.0, + 951.0, + 804.0, + 931.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 788.0, + 1058.0, + 788.0, + 1058.0, + 806.0, + 955.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 789.0, + 1107.0, + 789.0, + 1107.0, + 806.0, + 1059.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 800.0, + 818.0, + 800.0, + 818.0, + 822.0, + 791.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 800.0, + 1162.0, + 800.0, + 1162.0, + 822.0, + 1136.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 813.0, + 583.0, + 813.0, + 583.0, + 835.0, + 528.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 814.0, + 688.0, + 814.0, + 688.0, + 831.0, + 586.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 815.0, + 713.0, + 815.0, + 713.0, + 831.0, + 690.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 815.0, + 738.0, + 815.0, + 738.0, + 831.0, + 715.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 815.0, + 764.0, + 815.0, + 764.0, + 831.0, + 739.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 813.0, + 926.0, + 813.0, + 926.0, + 835.0, + 872.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 815.0, + 953.0, + 815.0, + 953.0, + 831.0, + 930.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 815.0, + 980.0, + 815.0, + 980.0, + 831.0, + 955.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 815.0, + 1005.0, + 815.0, + 1005.0, + 831.0, + 981.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 815.0, + 1030.0, + 815.0, + 1030.0, + 831.0, + 1007.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 815.0, + 1056.0, + 815.0, + 1056.0, + 831.0, + 1033.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 815.0, + 1107.0, + 815.0, + 1107.0, + 831.0, + 1058.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 828.0, + 817.0, + 828.0, + 817.0, + 846.0, + 792.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 828.0, + 1161.0, + 828.0, + 1161.0, + 846.0, + 1137.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 839.0, + 582.0, + 839.0, + 582.0, + 860.0, + 528.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 840.0, + 609.0, + 840.0, + 609.0, + 857.0, + 586.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 840.0, + 688.0, + 840.0, + 688.0, + 857.0, + 612.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 841.0, + 710.0, + 841.0, + 710.0, + 855.0, + 691.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 840.0, + 738.0, + 840.0, + 738.0, + 857.0, + 714.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 838.0, + 764.0, + 838.0, + 764.0, + 858.0, + 740.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 839.0, + 928.0, + 839.0, + 928.0, + 859.0, + 871.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 841.0, + 951.0, + 841.0, + 951.0, + 855.0, + 931.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 840.0, + 980.0, + 840.0, + 980.0, + 857.0, + 956.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 840.0, + 1005.0, + 840.0, + 1005.0, + 857.0, + 982.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 840.0, + 1030.0, + 840.0, + 1030.0, + 857.0, + 1007.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 840.0, + 1056.0, + 840.0, + 1056.0, + 857.0, + 1032.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 843.0, + 1081.0, + 843.0, + 1081.0, + 855.0, + 1059.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 843.0, + 1107.0, + 843.0, + 1107.0, + 855.0, + 1085.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 860.0, + 766.0, + 860.0, + 766.0, + 879.0, + 558.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 854.0, + 807.0, + 854.0, + 807.0, + 870.0, + 792.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 861.0, + 1109.0, + 861.0, + 1109.0, + 880.0, + 902.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 855.0, + 1152.0, + 855.0, + 1152.0, + 869.0, + 1137.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 873.0, + 750.0, + 873.0, + 750.0, + 897.0, + 572.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 873.0, + 1095.0, + 873.0, + 1095.0, + 897.0, + 915.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.25, + 324.0, + 416.25, + 324.0, + 416.25, + 340.0, + 394.25, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.25, + 324.0, + 786.25, + 324.0, + 786.25, + 340.0, + 765.25, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1213.0, + 581.0, + 1213.0, + 581.0, + 1232.0, + 413.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1212.0, + 917.0, + 1212.0, + 917.0, + 1233.0, + 764.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1208.0, + 1257.0, + 1208.0, + 1257.0, + 1235.0, + 1110.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1222.0, + 995.0, + 1222.0, + 995.0, + 1240.0, + 972.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1222.0, + 1339.0, + 1222.0, + 1339.0, + 1240.0, + 1316.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1233.0, + 417.0, + 1233.0, + 417.0, + 1254.0, + 366.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1236.0, + 445.0, + 1236.0, + 445.0, + 1249.0, + 421.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1235.0, + 470.0, + 1235.0, + 470.0, + 1251.0, + 447.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 1234.0, + 496.0, + 1234.0, + 496.0, + 1251.0, + 473.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1234.0, + 520.0, + 1234.0, + 520.0, + 1251.0, + 498.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1234.0, + 572.0, + 1234.0, + 572.0, + 1250.0, + 522.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1234.0, + 601.0, + 1234.0, + 601.0, + 1250.0, + 573.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1234.0, + 763.0, + 1234.0, + 763.0, + 1254.0, + 710.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1235.0, + 787.0, + 1235.0, + 787.0, + 1251.0, + 764.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1234.0, + 814.0, + 1234.0, + 814.0, + 1251.0, + 791.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1235.0, + 839.0, + 1235.0, + 839.0, + 1251.0, + 816.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1234.0, + 891.0, + 1234.0, + 891.0, + 1251.0, + 840.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1236.0, + 916.0, + 1236.0, + 916.0, + 1249.0, + 894.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1235.0, + 942.0, + 1235.0, + 942.0, + 1251.0, + 919.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1234.0, + 1108.0, + 1234.0, + 1108.0, + 1254.0, + 1054.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1236.0, + 1131.0, + 1236.0, + 1131.0, + 1249.0, + 1110.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1235.0, + 1157.0, + 1235.0, + 1157.0, + 1251.0, + 1134.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1236.0, + 1182.0, + 1236.0, + 1182.0, + 1250.0, + 1162.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1236.0, + 1208.0, + 1236.0, + 1208.0, + 1249.0, + 1186.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1234.0, + 1234.0, + 1234.0, + 1234.0, + 1251.0, + 1210.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1234.0, + 1261.0, + 1234.0, + 1261.0, + 1251.0, + 1237.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1234.0, + 1286.0, + 1234.0, + 1286.0, + 1251.0, + 1262.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1250.0, + 418.0, + 1250.0, + 418.0, + 1441.0, + 343.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1260.0, + 496.0, + 1260.0, + 496.0, + 1277.0, + 422.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1260.0, + 521.0, + 1260.0, + 521.0, + 1277.0, + 498.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1260.0, + 547.0, + 1260.0, + 547.0, + 1277.0, + 522.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1260.0, + 601.0, + 1260.0, + 601.0, + 1277.0, + 549.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1260.0, + 651.0, + 1260.0, + 651.0, + 1279.0, + 628.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1252.0, + 866.0, + 1252.0, + 866.0, + 1340.0, + 691.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1260.0, + 917.0, + 1260.0, + 917.0, + 1277.0, + 867.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1260.0, + 942.0, + 1260.0, + 942.0, + 1277.0, + 919.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1258.0, + 997.0, + 1258.0, + 997.0, + 1280.0, + 970.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1253.0, + 1106.0, + 1253.0, + 1106.0, + 1339.0, + 1035.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1260.0, + 1158.0, + 1260.0, + 1158.0, + 1277.0, + 1108.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1260.0, + 1183.0, + 1260.0, + 1183.0, + 1277.0, + 1160.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1260.0, + 1209.0, + 1260.0, + 1209.0, + 1277.0, + 1185.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1260.0, + 1236.0, + 1260.0, + 1236.0, + 1277.0, + 1210.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1260.0, + 1286.0, + 1260.0, + 1286.0, + 1277.0, + 1237.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1260.0, + 1340.0, + 1260.0, + 1340.0, + 1279.0, + 1316.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1285.0, + 444.0, + 1285.0, + 444.0, + 1305.0, + 366.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1286.0, + 469.0, + 1286.0, + 469.0, + 1303.0, + 446.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1286.0, + 521.0, + 1286.0, + 521.0, + 1303.0, + 470.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1286.0, + 599.0, + 1286.0, + 599.0, + 1303.0, + 522.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1281.0, + 839.0, + 1281.0, + 839.0, + 1308.0, + 707.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1286.0, + 917.0, + 1286.0, + 917.0, + 1302.0, + 842.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1286.0, + 943.0, + 1286.0, + 943.0, + 1302.0, + 919.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1283.0, + 1109.0, + 1283.0, + 1109.0, + 1307.0, + 1053.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1286.0, + 1157.0, + 1286.0, + 1157.0, + 1302.0, + 1134.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1287.0, + 1182.0, + 1287.0, + 1182.0, + 1301.0, + 1162.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1286.0, + 1236.0, + 1286.0, + 1236.0, + 1302.0, + 1184.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1286.0, + 1261.0, + 1286.0, + 1261.0, + 1302.0, + 1237.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1286.0, + 1287.0, + 1286.0, + 1287.0, + 1303.0, + 1262.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1298.0, + 652.0, + 1298.0, + 652.0, + 1317.0, + 628.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1298.0, + 995.0, + 1298.0, + 995.0, + 1317.0, + 972.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1298.0, + 1339.0, + 1298.0, + 1339.0, + 1317.0, + 1316.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1309.0, + 599.0, + 1309.0, + 599.0, + 1331.0, + 363.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1311.0, + 667.0, + 1311.0, + 667.0, + 1381.0, + 644.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1310.0, + 788.0, + 1310.0, + 788.0, + 1331.0, + 706.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1311.0, + 815.0, + 1311.0, + 815.0, + 1328.0, + 790.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1311.0, + 839.0, + 1311.0, + 839.0, + 1328.0, + 816.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1311.0, + 943.0, + 1311.0, + 943.0, + 1328.0, + 840.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 1311.0, + 1012.0, + 1311.0, + 1012.0, + 1382.0, + 989.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1308.0, + 1106.0, + 1308.0, + 1106.0, + 1333.0, + 1050.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1311.0, + 1132.0, + 1311.0, + 1132.0, + 1328.0, + 1108.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1311.0, + 1157.0, + 1311.0, + 1157.0, + 1328.0, + 1134.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1311.0, + 1262.0, + 1311.0, + 1262.0, + 1328.0, + 1160.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 1312.0, + 1287.0, + 1312.0, + 1287.0, + 1328.0, + 1264.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 1311.0, + 1355.0, + 1311.0, + 1355.0, + 1382.0, + 1332.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1335.0, + 600.0, + 1335.0, + 600.0, + 1357.0, + 363.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1338.0, + 652.0, + 1338.0, + 652.0, + 1356.0, + 628.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 1327.0, + 942.0, + 1327.0, + 942.0, + 1440.0, + 689.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1338.0, + 995.0, + 1338.0, + 995.0, + 1356.0, + 972.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1326.0, + 1185.0, + 1326.0, + 1185.0, + 1440.0, + 1034.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1338.0, + 1209.0, + 1338.0, + 1209.0, + 1354.0, + 1186.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1336.0, + 1286.0, + 1336.0, + 1286.0, + 1354.0, + 1210.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1338.0, + 1339.0, + 1338.0, + 1339.0, + 1356.0, + 1316.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1359.0, + 600.0, + 1359.0, + 600.0, + 1385.0, + 362.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1359.0, + 943.0, + 1359.0, + 943.0, + 1385.0, + 705.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1362.0, + 1288.0, + 1362.0, + 1288.0, + 1384.0, + 1051.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1374.0, + 653.0, + 1374.0, + 653.0, + 1396.0, + 627.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1376.0, + 995.0, + 1376.0, + 995.0, + 1394.0, + 972.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1374.0, + 1340.0, + 1374.0, + 1340.0, + 1396.0, + 1315.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1388.0, + 598.0, + 1388.0, + 598.0, + 1409.0, + 363.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1388.0, + 866.0, + 1388.0, + 866.0, + 1410.0, + 706.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1389.0, + 944.0, + 1389.0, + 944.0, + 1406.0, + 867.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1388.0, + 1288.0, + 1388.0, + 1288.0, + 1409.0, + 1050.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1408.0, + 599.0, + 1408.0, + 599.0, + 1437.0, + 361.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1413.0, + 652.0, + 1413.0, + 652.0, + 1434.0, + 625.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1410.0, + 918.0, + 1410.0, + 918.0, + 1437.0, + 705.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1415.0, + 943.0, + 1415.0, + 943.0, + 1432.0, + 919.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1415.0, + 995.0, + 1415.0, + 995.0, + 1433.0, + 972.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1411.0, + 1286.0, + 1411.0, + 1286.0, + 1437.0, + 1049.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1414.0, + 1341.0, + 1414.0, + 1341.0, + 1435.0, + 1313.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1439.0, + 599.0, + 1439.0, + 599.0, + 1460.0, + 363.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1439.0, + 942.0, + 1439.0, + 942.0, + 1460.0, + 706.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1437.0, + 1264.0, + 1437.0, + 1264.0, + 1460.0, + 1050.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1442.0, + 1284.0, + 1442.0, + 1284.0, + 1456.0, + 1265.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1460.0, + 601.0, + 1460.0, + 601.0, + 1482.0, + 391.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1454.0, + 642.0, + 1454.0, + 642.0, + 1469.0, + 627.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1460.0, + 946.0, + 1460.0, + 946.0, + 1482.0, + 735.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1456.0, + 984.0, + 1456.0, + 984.0, + 1467.0, + 975.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1460.0, + 1289.0, + 1460.0, + 1289.0, + 1482.0, + 1079.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 1454.0, + 1330.0, + 1454.0, + 1330.0, + 1470.0, + 1317.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1473.0, + 587.0, + 1473.0, + 587.0, + 1497.0, + 407.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1473.0, + 931.0, + 1473.0, + 931.0, + 1497.0, + 751.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1473.0, + 1273.0, + 1473.0, + 1273.0, + 1497.0, + 1094.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1521.0, + 739.0, + 1521.0, + 739.0, + 1544.0, + 585.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1530.0, + 817.0, + 1530.0, + 817.0, + 1552.0, + 790.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1521.0, + 1078.0, + 1521.0, + 1078.0, + 1545.0, + 934.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1532.0, + 1161.0, + 1532.0, + 1161.0, + 1549.0, + 1136.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1543.0, + 610.0, + 1543.0, + 610.0, + 1564.0, + 531.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1545.0, + 635.0, + 1545.0, + 635.0, + 1561.0, + 611.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1544.0, + 660.0, + 1544.0, + 660.0, + 1561.0, + 638.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1544.0, + 685.0, + 1544.0, + 685.0, + 1561.0, + 664.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 1545.0, + 713.0, + 1545.0, + 713.0, + 1561.0, + 689.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1545.0, + 739.0, + 1545.0, + 739.0, + 1561.0, + 714.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1546.0, + 762.0, + 1546.0, + 762.0, + 1559.0, + 741.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1543.0, + 1005.0, + 1543.0, + 1005.0, + 1564.0, + 875.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1544.0, + 1082.0, + 1544.0, + 1082.0, + 1561.0, + 1007.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1545.0, + 1108.0, + 1545.0, + 1108.0, + 1561.0, + 1083.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1564.0, + 763.0, + 1564.0, + 763.0, + 1651.0, + 510.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1570.0, + 816.0, + 1570.0, + 816.0, + 1589.0, + 793.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1563.0, + 1110.0, + 1563.0, + 1110.0, + 1706.0, + 854.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1570.0, + 1161.0, + 1570.0, + 1161.0, + 1589.0, + 1136.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1594.0, + 765.0, + 1594.0, + 765.0, + 1615.0, + 532.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1594.0, + 1109.0, + 1594.0, + 1109.0, + 1615.0, + 875.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1609.0, + 817.0, + 1609.0, + 817.0, + 1627.0, + 792.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1608.0, + 1161.0, + 1608.0, + 1161.0, + 1627.0, + 1138.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1620.0, + 661.0, + 1620.0, + 661.0, + 1642.0, + 527.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1622.0, + 765.0, + 1622.0, + 765.0, + 1638.0, + 662.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1620.0, + 834.0, + 1620.0, + 834.0, + 1690.0, + 807.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1620.0, + 1111.0, + 1620.0, + 1111.0, + 1640.0, + 872.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1620.0, + 1177.0, + 1620.0, + 1177.0, + 1690.0, + 1154.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1637.0, + 765.0, + 1637.0, + 765.0, + 1749.0, + 511.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1647.0, + 816.0, + 1647.0, + 816.0, + 1666.0, + 793.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1645.0, + 1110.0, + 1645.0, + 1110.0, + 1668.0, + 872.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1647.0, + 1161.0, + 1647.0, + 1161.0, + 1666.0, + 1136.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1670.0, + 766.0, + 1670.0, + 766.0, + 1695.0, + 526.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1670.0, + 927.0, + 1670.0, + 927.0, + 1695.0, + 870.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1671.0, + 1111.0, + 1671.0, + 1111.0, + 1691.0, + 929.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1686.0, + 817.0, + 1686.0, + 817.0, + 1705.0, + 793.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1686.0, + 1161.0, + 1686.0, + 1161.0, + 1705.0, + 1136.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1697.0, + 765.0, + 1697.0, + 765.0, + 1720.0, + 526.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1695.0, + 1110.0, + 1695.0, + 1110.0, + 1748.0, + 858.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1721.0, + 609.0, + 1721.0, + 609.0, + 1747.0, + 526.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1724.0, + 739.0, + 1724.0, + 739.0, + 1742.0, + 610.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1724.0, + 764.0, + 1724.0, + 764.0, + 1742.0, + 741.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1723.0, + 818.0, + 1723.0, + 818.0, + 1745.0, + 792.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1723.0, + 927.0, + 1723.0, + 927.0, + 1745.0, + 872.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1724.0, + 954.0, + 1724.0, + 954.0, + 1742.0, + 929.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1724.0, + 979.0, + 1724.0, + 979.0, + 1742.0, + 956.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1724.0, + 1005.0, + 1724.0, + 1005.0, + 1742.0, + 980.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1724.0, + 1030.0, + 1724.0, + 1030.0, + 1742.0, + 1007.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1724.0, + 1083.0, + 1724.0, + 1083.0, + 1742.0, + 1031.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1724.0, + 1108.0, + 1724.0, + 1108.0, + 1742.0, + 1084.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1723.0, + 1162.0, + 1723.0, + 1162.0, + 1745.0, + 1135.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1750.0, + 585.0, + 1750.0, + 585.0, + 1769.0, + 527.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1751.0, + 609.0, + 1751.0, + 609.0, + 1767.0, + 586.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1751.0, + 635.0, + 1751.0, + 635.0, + 1767.0, + 611.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1750.0, + 660.0, + 1750.0, + 660.0, + 1767.0, + 638.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1750.0, + 765.0, + 1750.0, + 765.0, + 1767.0, + 661.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1747.0, + 954.0, + 1747.0, + 954.0, + 1770.0, + 870.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1751.0, + 979.0, + 1751.0, + 979.0, + 1767.0, + 956.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1751.0, + 1005.0, + 1751.0, + 1005.0, + 1767.0, + 980.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1751.0, + 1030.0, + 1751.0, + 1030.0, + 1767.0, + 1007.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1749.0, + 1109.0, + 1749.0, + 1109.0, + 1768.0, + 1032.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1768.0, + 766.0, + 1768.0, + 766.0, + 1790.0, + 557.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1762.0, + 808.0, + 1762.0, + 808.0, + 1782.0, + 791.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1769.0, + 1110.0, + 1769.0, + 1110.0, + 1791.0, + 901.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1766.0, + 1151.0, + 1766.0, + 1151.0, + 1780.0, + 1136.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1783.0, + 749.0, + 1783.0, + 749.0, + 1806.0, + 571.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1783.0, + 1094.0, + 1783.0, + 1094.0, + 1807.0, + 916.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.25, + 1221.0, + 653.25, + 1221.0, + 653.25, + 1240.0, + 626.25, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.25, + 1285.0, + 1130.25, + 1285.0, + 1130.25, + 1302.0, + 1108.25, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1848.0, + 1409.0, + 1848.0, + 1409.0, + 1888.0, + 293.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1879.0, + 1404.0, + 1879.0, + 1404.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1907.0, + 1116.0, + 1907.0, + 1116.0, + 1945.0, + 291.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1907.0, + 1241.0, + 1907.0, + 1241.0, + 1945.0, + 1230.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 938.0, + 1408.0, + 938.0, + 1408.0, + 977.0, + 293.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 970.0, + 1405.0, + 970.0, + 1405.0, + 1005.0, + 294.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 997.0, + 1125.0, + 997.0, + 1125.0, + 1034.0, + 290.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 997.0, + 1249.0, + 997.0, + 1249.0, + 1034.0, + 1239.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 32, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 345, + 322, + 1359, + 322, + 1359, + 865, + 345, + 865 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 346, + 1201, + 1358, + 1201, + 1358, + 1802, + 346, + 1802 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 296, + 917, + 1406, + 917, + 1406, + 1011, + 296, + 1011 + ], + "score": 0.95 + }, + { + "category_id": 4, + "poly": [ + 296, + 1844, + 1406, + 1844, + 1406, + 1937, + 296, + 1937 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.854 + }, + { + "category_id": 13, + "poly": [ + 1124, + 977, + 1237, + 977, + 1237, + 1007, + 1124, + 1007 + ], + "score": 0.92, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 1150, + 1903, + 1263, + 1903, + 1263, + 1933, + 1150, + 1933 + ], + "score": 0.91, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 608, + 920, + 648, + 920, + 648, + 947, + 608, + 947 + ], + "score": 0.62, + "latex": "\\textcircled { \\alpha } 5" + }, + { + "category_id": 13, + "poly": [ + 876, + 1519, + 1158, + 1519, + 1158, + 1793, + 876, + 1793 + ], + "score": 0.5, + "latex": "\\begin{array} { r } { \\mathrm { S t a c c ~ O n e r t e r ~ W o t e ~ P M W P , ~ F e d e N G } } \\\\ { \\mathrm { 1 . 0 ~ S ~ s ~ a ~ 8 0 ~ 0 . 3 ~ 1 . 2 ~ 1 . ~ 1 . ~ 0 ~ s ~ 0 . } } \\\\ { \\mathrm { 2 . 5 . ~ 5 . 7 ~ 5 . } } \\\\ { \\mathrm { 0 . 0 . 5 - 1 . 7 ~ 5 . } } \\\\ { \\mathrm { 3 . 0 . 1 ~ 5 . ~ 6 . ~ 7 . ~ 2 ~ 4 . 1 ~ 2 . ~ 1 . ~ 0 . } } \\\\ { \\mathrm { 0 . 0 . 1 ~ 8 . ~ 7 . ~ 2 ~ 1 . ~ 1 . ~ 0 . } } \\\\ { \\mathrm { 0 . 5 . ~ 6 . 0 ~ 1 . 2 ~ 5 . ~ 5 . ~ 7 . ~ 3 1 . ~ 1 . ~ 0 . ~ 9 . } } \\\\ { \\mathrm { 1 . 0 . ~ 6 . 0 ~ 0 . ~ 0 . 9 ~ 5 . ~ 7 . ~ 0 3 . ~ 1 . ~ 1 . ~ 1 . ~ 3 } } \\\\ { \\mathrm { 1 . 5 . ~ 5 0 ~ ~ 0 . 0 ~ 0 . ~ 0 . 3 ~ 1 . ~ 6 . ~ 5 . ~ 7 . ~ 0 . ~ 0 } } \\\\ { \\mathrm { 2 . 0 . ~ 6 . 0 ~ 0 . ~ 0 . 0 ~ 0 . 0 ~ 0 . 3 ~ 2 . ~ 0 . 7 ~ 0 . ~ 0 . } } \\\\ { \\mathrm { 2 . 5 . ~ 5 0 ~ 0 . ~ 0 . 0 ~ 0 . ~ 0 . 0 ~ 0 . 0 ~ 1 . 2 ~ 1 . ~ 0 . } } \\\\ { \\mathrm { 2 . 5 . ~ 5 0 ~ 0 . ~ 0 . 0 ~ 0 . ~ 0 . 0 ~ 0 . 0 ~ 0 . 1 . ~ 1 . ~ 0 . } } \\\\ { \\mathrm { 3 . 0 . 0 ~ 0 . ~ 0 . 0 ~ 0 . ~ 0 . 0 ~ 0 . ~ 0 . 0 ~ 1 . ~ 1 . ~ 0 . } } \\\\ { \\mathrm { 3 . 0 . 2 ~ 5 . ~ 2 . ~ 0 . ~ 1 . ~ 5 . ~ 1 . ~ 0 . ~ 0 . 0 ~ 0 . 5 } } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 326.0, + 593.0, + 326.0, + 593.0, + 345.0, + 401.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 326.0, + 929.0, + 326.0, + 929.0, + 345.0, + 752.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 326.0, + 1266.0, + 326.0, + 1266.0, + 345.0, + 1100.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 335.0, + 652.0, + 335.0, + 652.0, + 352.0, + 629.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 335.0, + 996.0, + 335.0, + 996.0, + 352.0, + 973.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 335.0, + 1339.0, + 335.0, + 1339.0, + 352.0, + 1317.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 345.0, + 604.0, + 345.0, + 604.0, + 526.0, + 344.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 361.0, + 652.0, + 361.0, + 652.0, + 380.0, + 629.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 344.0, + 947.0, + 344.0, + 947.0, + 526.0, + 689.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 361.0, + 996.0, + 361.0, + 996.0, + 380.0, + 974.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 344.0, + 1292.0, + 344.0, + 1292.0, + 526.0, + 1032.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 361.0, + 1341.0, + 361.0, + 1341.0, + 380.0, + 1316.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 391.0, + 461.0, + 391.0, + 461.0, + 413.0, + 363.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 395.0, + 484.0, + 395.0, + 484.0, + 408.0, + 462.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 395.0, + 507.0, + 395.0, + 507.0, + 408.0, + 485.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 393.0, + 604.0, + 393.0, + 604.0, + 410.0, + 509.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 382.0, + 669.0, + 382.0, + 669.0, + 450.0, + 629.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 391.0, + 946.0, + 391.0, + 946.0, + 414.0, + 707.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 381.0, + 1013.0, + 381.0, + 1013.0, + 492.0, + 974.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 391.0, + 1289.0, + 391.0, + 1289.0, + 413.0, + 1052.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 381.0, + 1358.0, + 381.0, + 1358.0, + 492.0, + 1317.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 415.0, + 601.0, + 415.0, + 601.0, + 435.0, + 365.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 415.0, + 652.0, + 415.0, + 652.0, + 433.0, + 628.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 415.0, + 948.0, + 415.0, + 948.0, + 436.0, + 707.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 415.0, + 996.0, + 415.0, + 996.0, + 433.0, + 973.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 415.0, + 1291.0, + 415.0, + 1291.0, + 436.0, + 1052.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 415.0, + 1339.0, + 415.0, + 1339.0, + 433.0, + 1316.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 439.0, + 601.0, + 439.0, + 601.0, + 460.0, + 361.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 439.0, + 669.0, + 439.0, + 669.0, + 492.0, + 629.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 439.0, + 946.0, + 439.0, + 946.0, + 460.0, + 705.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 442.0, + 997.0, + 442.0, + 997.0, + 460.0, + 974.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 439.0, + 1266.0, + 439.0, + 1266.0, + 460.0, + 1048.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 442.0, + 1289.0, + 442.0, + 1289.0, + 455.0, + 1269.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 442.0, + 1341.0, + 442.0, + 1341.0, + 460.0, + 1317.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 462.0, + 604.0, + 462.0, + 604.0, + 482.0, + 361.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 468.0, + 652.0, + 468.0, + 652.0, + 487.0, + 629.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 459.0, + 923.0, + 459.0, + 923.0, + 484.0, + 705.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 463.0, + 947.0, + 463.0, + 947.0, + 480.0, + 924.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 468.0, + 998.0, + 468.0, + 998.0, + 487.0, + 974.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 461.0, + 1267.0, + 461.0, + 1267.0, + 482.0, + 1048.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 465.0, + 1288.0, + 465.0, + 1288.0, + 479.0, + 1269.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 468.0, + 1341.0, + 468.0, + 1341.0, + 487.0, + 1317.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 485.0, + 604.0, + 485.0, + 604.0, + 507.0, + 361.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 494.0, + 654.0, + 494.0, + 654.0, + 517.0, + 628.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 485.0, + 922.0, + 485.0, + 922.0, + 507.0, + 705.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 488.0, + 945.0, + 488.0, + 945.0, + 502.0, + 925.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 495.0, + 996.0, + 495.0, + 996.0, + 514.0, + 973.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 485.0, + 1222.0, + 485.0, + 1222.0, + 507.0, + 1048.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 488.0, + 1265.0, + 488.0, + 1265.0, + 501.0, + 1243.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 488.0, + 1287.0, + 488.0, + 1287.0, + 502.0, + 1269.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 495.0, + 1339.0, + 495.0, + 1339.0, + 514.0, + 1317.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 506.0, + 605.0, + 506.0, + 605.0, + 532.0, + 359.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 508.0, + 924.0, + 508.0, + 924.0, + 530.0, + 705.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 512.0, + 945.0, + 512.0, + 945.0, + 526.0, + 925.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 508.0, + 1266.0, + 508.0, + 1266.0, + 530.0, + 1048.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 512.0, + 1288.0, + 512.0, + 1288.0, + 526.0, + 1270.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 524.0, + 643.0, + 524.0, + 643.0, + 539.0, + 630.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 524.0, + 988.0, + 524.0, + 988.0, + 539.0, + 974.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 524.0, + 1331.0, + 524.0, + 1331.0, + 539.0, + 1318.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 530.0, + 605.0, + 530.0, + 605.0, + 549.0, + 388.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 530.0, + 948.0, + 530.0, + 948.0, + 549.0, + 732.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 530.0, + 1291.0, + 530.0, + 1291.0, + 549.0, + 1076.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 543.0, + 578.0, + 543.0, + 578.0, + 565.0, + 411.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 541.0, + 924.0, + 541.0, + 924.0, + 565.0, + 755.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 541.0, + 1267.0, + 541.0, + 1267.0, + 565.0, + 1098.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 621.0, + 749.0, + 621.0, + 749.0, + 640.0, + 575.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 618.0, + 1089.0, + 618.0, + 1089.0, + 643.0, + 921.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 629.0, + 817.0, + 629.0, + 817.0, + 648.0, + 794.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 629.0, + 1162.0, + 629.0, + 1162.0, + 648.0, + 1139.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 638.0, + 770.0, + 638.0, + 770.0, + 657.0, + 623.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 640.0, + 558.0, + 640.0, + 558.0, + 661.0, + 529.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 640.0, + 902.0, + 640.0, + 902.0, + 661.0, + 873.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 640.0, + 1111.0, + 640.0, + 1111.0, + 657.0, + 1088.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 641.0, + 533.0, + 641.0, + 533.0, + 820.0, + 507.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 641.0, + 629.0, + 641.0, + 629.0, + 656.0, + 577.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 641.0, + 1042.0, + 641.0, + 1042.0, + 657.0, + 1014.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 641.0, + 1064.0, + 641.0, + 1064.0, + 657.0, + 1041.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 642.0, + 580.0, + 642.0, + 580.0, + 655.0, + 554.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 642.0, + 879.0, + 642.0, + 879.0, + 822.0, + 853.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 642.0, + 925.0, + 642.0, + 925.0, + 655.0, + 898.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 642.0, + 947.0, + 642.0, + 947.0, + 655.0, + 923.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 642.0, + 973.0, + 642.0, + 973.0, + 655.0, + 946.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 642.0, + 995.0, + 642.0, + 995.0, + 655.0, + 969.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 642.0, + 1018.0, + 642.0, + 1018.0, + 655.0, + 992.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 642.0, + 1089.0, + 642.0, + 1089.0, + 655.0, + 1063.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 663.0, + 771.0, + 663.0, + 771.0, + 685.0, + 529.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 656.0, + 817.0, + 656.0, + 817.0, + 674.0, + 795.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 662.0, + 1113.0, + 662.0, + 1113.0, + 685.0, + 873.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 656.0, + 1162.0, + 656.0, + 1162.0, + 674.0, + 1139.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 686.0, + 770.0, + 686.0, + 770.0, + 707.0, + 528.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 675.0, + 835.0, + 675.0, + 835.0, + 787.0, + 795.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 686.0, + 1113.0, + 686.0, + 1113.0, + 707.0, + 872.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 675.0, + 1178.0, + 675.0, + 1178.0, + 786.0, + 1139.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 708.0, + 768.0, + 708.0, + 768.0, + 729.0, + 529.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 709.0, + 818.0, + 709.0, + 818.0, + 728.0, + 795.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 708.0, + 1113.0, + 708.0, + 1113.0, + 732.0, + 873.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 709.0, + 1162.0, + 709.0, + 1162.0, + 728.0, + 1139.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 734.0, + 770.0, + 734.0, + 770.0, + 754.0, + 526.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 737.0, + 818.0, + 737.0, + 818.0, + 754.0, + 795.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 733.0, + 1111.0, + 733.0, + 1111.0, + 754.0, + 869.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 737.0, + 1162.0, + 737.0, + 1162.0, + 754.0, + 1139.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 757.0, + 770.0, + 757.0, + 770.0, + 777.0, + 526.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 763.0, + 820.0, + 763.0, + 820.0, + 781.0, + 795.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 757.0, + 1111.0, + 757.0, + 1111.0, + 777.0, + 869.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 763.0, + 1163.0, + 763.0, + 1163.0, + 780.0, + 1139.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 779.0, + 767.0, + 779.0, + 767.0, + 802.0, + 527.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 779.0, + 1111.0, + 779.0, + 1111.0, + 802.0, + 871.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 790.0, + 818.0, + 790.0, + 818.0, + 809.0, + 794.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 790.0, + 1161.0, + 790.0, + 1161.0, + 809.0, + 1139.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 805.0, + 767.0, + 805.0, + 767.0, + 823.0, + 528.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 805.0, + 1086.0, + 805.0, + 1086.0, + 823.0, + 872.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 805.0, + 1111.0, + 805.0, + 1111.0, + 822.0, + 1088.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 818.0, + 809.0, + 818.0, + 809.0, + 832.0, + 789.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 818.0, + 1153.0, + 818.0, + 1153.0, + 832.0, + 1139.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 824.0, + 768.0, + 824.0, + 768.0, + 843.0, + 554.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 824.0, + 1112.0, + 824.0, + 1112.0, + 843.0, + 897.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 835.0, + 745.0, + 835.0, + 745.0, + 860.0, + 576.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 837.0, + 1089.0, + 837.0, + 1089.0, + 858.0, + 921.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 486.0, + 1243.0, + 486.0, + 1243.0, + 502.5, + 1218.0, + 502.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1205.0, + 605.0, + 1205.0, + 605.0, + 1228.0, + 388.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1207.0, + 940.0, + 1207.0, + 940.0, + 1226.0, + 742.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 1205.0, + 1281.0, + 1205.0, + 1281.0, + 1228.0, + 1088.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1216.0, + 652.0, + 1216.0, + 652.0, + 1234.0, + 628.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1216.0, + 996.0, + 1216.0, + 996.0, + 1234.0, + 971.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1216.0, + 1340.0, + 1216.0, + 1340.0, + 1234.0, + 1315.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1227.0, + 419.0, + 1227.0, + 419.0, + 1248.0, + 366.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1228.0, + 444.0, + 1228.0, + 444.0, + 1244.0, + 421.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1228.0, + 470.0, + 1228.0, + 470.0, + 1244.0, + 447.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1228.0, + 495.0, + 1228.0, + 495.0, + 1244.0, + 472.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1228.0, + 521.0, + 1228.0, + 521.0, + 1244.0, + 498.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1228.0, + 573.0, + 1228.0, + 573.0, + 1244.0, + 524.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1228.0, + 598.0, + 1228.0, + 598.0, + 1244.0, + 575.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1227.0, + 762.0, + 1227.0, + 762.0, + 1248.0, + 710.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1229.0, + 786.0, + 1229.0, + 786.0, + 1243.0, + 766.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1229.0, + 813.0, + 1229.0, + 813.0, + 1243.0, + 793.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1228.0, + 839.0, + 1228.0, + 839.0, + 1244.0, + 816.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1228.0, + 865.0, + 1228.0, + 865.0, + 1244.0, + 842.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1228.0, + 891.0, + 1228.0, + 891.0, + 1244.0, + 868.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1228.0, + 917.0, + 1228.0, + 917.0, + 1244.0, + 893.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1228.0, + 942.0, + 1228.0, + 942.0, + 1244.0, + 919.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1227.0, + 1106.0, + 1227.0, + 1106.0, + 1248.0, + 1054.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1229.0, + 1130.0, + 1229.0, + 1130.0, + 1243.0, + 1110.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1228.0, + 1158.0, + 1228.0, + 1158.0, + 1244.0, + 1135.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1228.0, + 1183.0, + 1228.0, + 1183.0, + 1244.0, + 1160.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1228.0, + 1209.0, + 1228.0, + 1209.0, + 1244.0, + 1186.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 1229.0, + 1233.0, + 1229.0, + 1233.0, + 1243.0, + 1213.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1228.0, + 1261.0, + 1228.0, + 1261.0, + 1244.0, + 1238.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1228.0, + 1286.0, + 1228.0, + 1286.0, + 1244.0, + 1262.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1245.0, + 419.0, + 1245.0, + 419.0, + 1435.0, + 344.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1254.0, + 444.0, + 1254.0, + 444.0, + 1271.0, + 420.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1254.0, + 470.0, + 1254.0, + 470.0, + 1271.0, + 447.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1254.0, + 495.0, + 1254.0, + 495.0, + 1271.0, + 471.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1252.0, + 522.0, + 1252.0, + 522.0, + 1271.0, + 498.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1254.0, + 574.0, + 1254.0, + 574.0, + 1271.0, + 523.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1254.0, + 598.0, + 1254.0, + 598.0, + 1271.0, + 575.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1252.0, + 652.0, + 1252.0, + 652.0, + 1274.0, + 626.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1244.0, + 762.0, + 1244.0, + 762.0, + 1436.0, + 688.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1254.0, + 788.0, + 1254.0, + 788.0, + 1271.0, + 764.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1254.0, + 814.0, + 1254.0, + 814.0, + 1271.0, + 791.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1254.0, + 840.0, + 1254.0, + 840.0, + 1271.0, + 816.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1254.0, + 865.0, + 1254.0, + 865.0, + 1271.0, + 842.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1254.0, + 891.0, + 1254.0, + 891.0, + 1271.0, + 868.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1254.0, + 917.0, + 1254.0, + 917.0, + 1271.0, + 893.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1254.0, + 942.0, + 1254.0, + 942.0, + 1271.0, + 919.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1251.0, + 996.0, + 1251.0, + 996.0, + 1274.0, + 970.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1245.0, + 1107.0, + 1245.0, + 1107.0, + 1436.0, + 1032.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1254.0, + 1132.0, + 1254.0, + 1132.0, + 1271.0, + 1108.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1254.0, + 1158.0, + 1254.0, + 1158.0, + 1271.0, + 1133.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1254.0, + 1183.0, + 1254.0, + 1183.0, + 1271.0, + 1160.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1254.0, + 1209.0, + 1254.0, + 1209.0, + 1271.0, + 1186.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1254.0, + 1261.0, + 1254.0, + 1261.0, + 1271.0, + 1211.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1254.0, + 1286.0, + 1254.0, + 1286.0, + 1271.0, + 1262.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1251.0, + 1341.0, + 1251.0, + 1341.0, + 1274.0, + 1313.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1278.0, + 419.0, + 1278.0, + 419.0, + 1299.0, + 366.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1278.0, + 549.0, + 1278.0, + 549.0, + 1297.0, + 420.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1279.0, + 574.0, + 1279.0, + 574.0, + 1296.0, + 550.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1280.0, + 599.0, + 1280.0, + 599.0, + 1296.0, + 575.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1278.0, + 762.0, + 1278.0, + 762.0, + 1299.0, + 710.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 1279.0, + 788.0, + 1279.0, + 788.0, + 1296.0, + 764.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1281.0, + 811.0, + 1281.0, + 811.0, + 1295.0, + 792.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1279.0, + 839.0, + 1279.0, + 839.0, + 1296.0, + 817.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1280.0, + 865.0, + 1280.0, + 865.0, + 1296.0, + 842.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1279.0, + 891.0, + 1279.0, + 891.0, + 1296.0, + 868.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1280.0, + 917.0, + 1280.0, + 917.0, + 1296.0, + 893.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1279.0, + 942.0, + 1279.0, + 942.0, + 1296.0, + 919.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1278.0, + 1106.0, + 1278.0, + 1106.0, + 1299.0, + 1054.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1280.0, + 1132.0, + 1280.0, + 1132.0, + 1296.0, + 1109.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1279.0, + 1157.0, + 1279.0, + 1157.0, + 1296.0, + 1133.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1280.0, + 1183.0, + 1280.0, + 1183.0, + 1296.0, + 1160.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1279.0, + 1209.0, + 1279.0, + 1209.0, + 1296.0, + 1186.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1280.0, + 1235.0, + 1280.0, + 1235.0, + 1296.0, + 1211.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1280.0, + 1286.0, + 1280.0, + 1286.0, + 1296.0, + 1237.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1293.0, + 652.0, + 1293.0, + 652.0, + 1311.0, + 627.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1293.0, + 996.0, + 1293.0, + 996.0, + 1311.0, + 971.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1291.0, + 1341.0, + 1291.0, + 1341.0, + 1312.0, + 1314.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1304.0, + 574.0, + 1304.0, + 574.0, + 1350.0, + 362.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1305.0, + 599.0, + 1305.0, + 599.0, + 1348.0, + 575.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1307.0, + 667.0, + 1307.0, + 667.0, + 1375.0, + 626.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1304.0, + 762.0, + 1304.0, + 762.0, + 1350.0, + 706.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1305.0, + 867.0, + 1305.0, + 867.0, + 1348.0, + 763.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1305.0, + 893.0, + 1305.0, + 893.0, + 1348.0, + 868.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1305.0, + 918.0, + 1305.0, + 918.0, + 1348.0, + 894.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1305.0, + 942.0, + 1305.0, + 942.0, + 1348.0, + 919.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1305.0, + 1011.0, + 1305.0, + 1011.0, + 1375.0, + 971.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1304.0, + 1211.0, + 1304.0, + 1211.0, + 1350.0, + 1050.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1305.0, + 1235.0, + 1305.0, + 1235.0, + 1348.0, + 1212.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1305.0, + 1286.0, + 1305.0, + 1286.0, + 1348.0, + 1237.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1304.0, + 1355.0, + 1304.0, + 1355.0, + 1375.0, + 1314.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1355.0, + 499.0, + 1355.0, + 499.0, + 1377.0, + 362.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1355.0, + 573.0, + 1355.0, + 573.0, + 1375.0, + 518.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1357.0, + 598.0, + 1357.0, + 598.0, + 1374.0, + 575.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1355.0, + 918.0, + 1355.0, + 918.0, + 1377.0, + 707.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1357.0, + 942.0, + 1357.0, + 942.0, + 1375.0, + 919.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1356.0, + 1237.0, + 1356.0, + 1237.0, + 1377.0, + 1050.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1357.0, + 1286.0, + 1357.0, + 1286.0, + 1374.0, + 1238.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1370.0, + 652.0, + 1370.0, + 652.0, + 1389.0, + 627.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1368.0, + 997.0, + 1368.0, + 997.0, + 1390.0, + 970.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1368.0, + 1341.0, + 1368.0, + 1341.0, + 1390.0, + 1314.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1380.0, + 574.0, + 1380.0, + 574.0, + 1405.0, + 361.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1383.0, + 599.0, + 1383.0, + 599.0, + 1399.0, + 575.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1379.0, + 917.0, + 1379.0, + 917.0, + 1405.0, + 705.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1383.0, + 942.0, + 1383.0, + 942.0, + 1400.0, + 919.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1380.0, + 1286.0, + 1380.0, + 1286.0, + 1405.0, + 1048.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1405.0, + 419.0, + 1405.0, + 419.0, + 1430.0, + 361.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1407.0, + 573.0, + 1407.0, + 573.0, + 1427.0, + 420.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1408.0, + 598.0, + 1408.0, + 598.0, + 1425.0, + 575.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1408.0, + 644.0, + 1408.0, + 644.0, + 1429.0, + 626.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1403.0, + 942.0, + 1403.0, + 942.0, + 1430.0, + 705.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1408.0, + 986.0, + 1408.0, + 986.0, + 1428.0, + 972.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1405.0, + 1286.0, + 1405.0, + 1286.0, + 1430.0, + 1049.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 1408.0, + 1332.0, + 1408.0, + 1332.0, + 1428.0, + 1314.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1433.0, + 444.0, + 1433.0, + 444.0, + 1453.0, + 363.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1433.0, + 522.0, + 1433.0, + 522.0, + 1451.0, + 445.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1435.0, + 573.0, + 1435.0, + 573.0, + 1451.0, + 523.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1435.0, + 598.0, + 1435.0, + 598.0, + 1451.0, + 575.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1432.0, + 893.0, + 1432.0, + 893.0, + 1453.0, + 706.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1435.0, + 942.0, + 1435.0, + 942.0, + 1451.0, + 894.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1431.0, + 1286.0, + 1431.0, + 1286.0, + 1454.0, + 1050.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1450.0, + 603.0, + 1450.0, + 603.0, + 1476.0, + 389.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1448.0, + 642.0, + 1448.0, + 642.0, + 1462.0, + 626.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1450.0, + 947.0, + 1450.0, + 947.0, + 1475.0, + 734.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1450.0, + 984.0, + 1450.0, + 984.0, + 1461.0, + 975.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1451.0, + 1291.0, + 1451.0, + 1291.0, + 1476.0, + 1077.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1450.0, + 1328.0, + 1450.0, + 1328.0, + 1461.0, + 1318.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1467.0, + 587.0, + 1467.0, + 587.0, + 1491.0, + 406.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1467.0, + 930.0, + 1467.0, + 930.0, + 1490.0, + 751.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1467.0, + 1274.0, + 1467.0, + 1274.0, + 1491.0, + 1094.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1518.0, + 760.0, + 1518.0, + 760.0, + 1536.0, + 564.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1525.0, + 817.0, + 1525.0, + 817.0, + 1545.0, + 789.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1537.0, + 586.0, + 1537.0, + 586.0, + 1558.0, + 531.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1540.0, + 608.0, + 1540.0, + 608.0, + 1553.0, + 588.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1538.0, + 635.0, + 1538.0, + 635.0, + 1555.0, + 611.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1538.0, + 661.0, + 1538.0, + 661.0, + 1555.0, + 637.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1538.0, + 686.0, + 1538.0, + 686.0, + 1555.0, + 663.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1540.0, + 711.0, + 1540.0, + 711.0, + 1553.0, + 691.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1538.0, + 738.0, + 1538.0, + 738.0, + 1555.0, + 714.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1538.0, + 763.0, + 1538.0, + 763.0, + 1555.0, + 740.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1558.0, + 688.0, + 1558.0, + 688.0, + 1644.0, + 511.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1564.0, + 738.0, + 1564.0, + 738.0, + 1581.0, + 690.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1564.0, + 764.0, + 1564.0, + 764.0, + 1581.0, + 740.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1561.0, + 817.0, + 1561.0, + 817.0, + 1583.0, + 792.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1558.0, + 875.0, + 1558.0, + 875.0, + 1645.0, + 855.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1588.0, + 714.0, + 1588.0, + 714.0, + 1610.0, + 531.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1589.0, + 764.0, + 1589.0, + 764.0, + 1606.0, + 715.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1601.0, + 818.0, + 1601.0, + 818.0, + 1623.0, + 792.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1601.0, + 1162.0, + 1601.0, + 1162.0, + 1623.0, + 1159.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1614.0, + 583.0, + 1614.0, + 583.0, + 1635.0, + 528.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1613.0, + 738.0, + 1613.0, + 738.0, + 1633.0, + 586.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1614.0, + 764.0, + 1614.0, + 764.0, + 1633.0, + 740.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1615.0, + 833.0, + 1615.0, + 833.0, + 1684.0, + 809.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1612.0, + 875.0, + 1612.0, + 875.0, + 1636.0, + 871.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1614.0, + 1177.0, + 1614.0, + 1177.0, + 1684.0, + 1159.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1631.0, + 738.0, + 1631.0, + 738.0, + 1700.0, + 512.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1641.0, + 764.0, + 1641.0, + 764.0, + 1658.0, + 740.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1639.0, + 817.0, + 1639.0, + 817.0, + 1661.0, + 791.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1630.0, + 875.0, + 1630.0, + 875.0, + 1699.0, + 856.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1666.0, + 583.0, + 1666.0, + 583.0, + 1687.0, + 528.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1666.0, + 609.0, + 1666.0, + 609.0, + 1684.0, + 586.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1666.0, + 635.0, + 1666.0, + 635.0, + 1684.0, + 610.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1665.0, + 764.0, + 1665.0, + 764.0, + 1685.0, + 638.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1664.0, + 875.0, + 1664.0, + 875.0, + 1688.0, + 871.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1680.0, + 817.0, + 1680.0, + 817.0, + 1699.0, + 793.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 1689.0, + 634.0, + 1689.0, + 634.0, + 1741.0, + 513.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1692.0, + 661.0, + 1692.0, + 661.0, + 1710.0, + 637.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1691.0, + 764.0, + 1691.0, + 764.0, + 1711.0, + 662.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1689.0, + 875.0, + 1689.0, + 875.0, + 1742.0, + 855.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1716.0, + 583.0, + 1716.0, + 583.0, + 1740.0, + 527.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1718.0, + 609.0, + 1718.0, + 609.0, + 1736.0, + 586.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1718.0, + 635.0, + 1718.0, + 635.0, + 1734.0, + 611.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1718.0, + 660.0, + 1718.0, + 660.0, + 1736.0, + 637.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1718.0, + 688.0, + 1718.0, + 688.0, + 1736.0, + 662.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 1718.0, + 713.0, + 1718.0, + 713.0, + 1736.0, + 689.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1718.0, + 738.0, + 1718.0, + 738.0, + 1736.0, + 715.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1718.0, + 764.0, + 1718.0, + 764.0, + 1736.0, + 740.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1718.0, + 808.0, + 1718.0, + 808.0, + 1738.0, + 794.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1715.0, + 875.0, + 1715.0, + 875.0, + 1739.0, + 871.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1744.0, + 584.0, + 1744.0, + 584.0, + 1763.0, + 528.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1745.0, + 609.0, + 1745.0, + 609.0, + 1761.0, + 586.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1746.0, + 633.0, + 1746.0, + 633.0, + 1760.0, + 613.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1744.0, + 661.0, + 1744.0, + 661.0, + 1761.0, + 638.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1744.0, + 686.0, + 1744.0, + 686.0, + 1761.0, + 663.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 1744.0, + 713.0, + 1744.0, + 713.0, + 1761.0, + 689.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1744.0, + 738.0, + 1744.0, + 738.0, + 1761.0, + 715.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1745.0, + 764.0, + 1745.0, + 764.0, + 1761.0, + 740.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1744.0, + 875.0, + 1744.0, + 875.0, + 1763.0, + 872.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1762.0, + 766.0, + 1762.0, + 766.0, + 1784.0, + 557.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1756.0, + 808.0, + 1756.0, + 808.0, + 1776.0, + 788.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1778.0, + 750.0, + 1778.0, + 750.0, + 1800.0, + 572.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1778.0, + 1095.0, + 1778.0, + 1095.0, + 1800.0, + 916.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1330.0, + 1161.0, + 1330.0, + 1161.0, + 1347.0, + 1131.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1356.0, + 525.0, + 1356.0, + 525.0, + 1372.5, + 493.0, + 1372.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 918.0, + 607.0, + 918.0, + 607.0, + 952.0, + 295.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 918.0, + 1406.0, + 918.0, + 1406.0, + 952.0, + 649.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 948.0, + 1407.0, + 948.0, + 1407.0, + 982.0, + 295.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 974.0, + 1123.0, + 974.0, + 1123.0, + 1012.0, + 291.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 974.0, + 1249.0, + 974.0, + 1249.0, + 1012.0, + 1238.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1843.0, + 1409.0, + 1843.0, + 1409.0, + 1881.0, + 294.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1872.0, + 1407.0, + 1872.0, + 1407.0, + 1909.0, + 293.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1900.0, + 1149.0, + 1900.0, + 1149.0, + 1939.0, + 291.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 1900.0, + 1276.0, + 1900.0, + 1276.0, + 1939.0, + 1264.0, + 1939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 829.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 33, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1793, + 1404, + 1793, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1597, + 1404, + 1597, + 1404, + 1779, + 297, + 1779 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 297, + 1405, + 297, + 1405, + 479, + 298, + 479 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 336, + 545, + 1363, + 545, + 1363, + 1324, + 336, + 1324 + ], + "score": 0.977 + }, + { + "category_id": 4, + "poly": [ + 295, + 1376, + 1406, + 1376, + 1406, + 1526, + 295, + 1526 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2112, + 835, + 2112 + ], + "score": 0.857 + }, + { + "category_id": 2, + "poly": [ + 312, + 233, + 1158, + 233, + 1158, + 260, + 312, + 260 + ], + "score": 0.577 + }, + { + "category_id": 13, + "poly": [ + 335, + 387, + 452, + 387, + 452, + 416, + 335, + 416 + ], + "score": 0.92, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1464, + 412, + 1464, + 412, + 1494, + 296, + 1494 + ], + "score": 0.92, + "latex": "\\tau = 1 0 ^ { - 3 }" + }, + { + "category_id": 13, + "poly": [ + 690, + 1496, + 742, + 1496, + 742, + 1524, + 690, + 1524 + ], + "score": 0.86, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 373, + 418, + 425, + 418, + 425, + 447, + 373, + 447 + ], + "score": 0.84, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 675, + 1663, + 700, + 1663, + 700, + 1691, + 675, + 1691 + ], + "score": 0.82, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 749, + 1664, + 769, + 1664, + 769, + 1691, + 749, + 1691 + ], + "score": 0.82, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 646, + 1723, + 672, + 1723, + 672, + 1749, + 646, + 1749 + ], + "score": 0.82, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1251, + 1604, + 1277, + 1604, + 1277, + 1631, + 1251, + 1631 + ], + "score": 0.81, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1236, + 334, + 1261, + 334, + 1261, + 360, + 1236, + 360 + ], + "score": 0.8, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 345, + 1723, + 363, + 1723, + 363, + 1749, + 345, + 1749 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 441, + 1634, + 460, + 1634, + 460, + 1661, + 441, + 1661 + ], + "score": 0.78, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 739, + 364, + 759, + 364, + 759, + 391, + 739, + 391 + ], + "score": 0.76, + "latex": "\\eta" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 549.0, + 558.0, + 549.0, + 558.0, + 571.0, + 483.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 549.0, + 918.0, + 549.0, + 918.0, + 571.0, + 834.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 551.0, + 1273.0, + 551.0, + 1273.0, + 570.0, + 1189.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 566.0, + 394.0, + 566.0, + 394.0, + 736.0, + 334.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 565.0, + 751.0, + 565.0, + 751.0, + 737.0, + 690.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 570.0, + 770.0, + 570.0, + 770.0, + 585.0, + 754.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 593.0, + 867.0, + 593.0, + 867.0, + 607.0, + 852.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 571.0, + 1107.0, + 571.0, + 1107.0, + 739.0, + 1044.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 570.0, + 1256.0, + 570.0, + 1256.0, + 585.0, + 1239.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 607.0, + 398.0, + 607.0, + 398.0, + 634.0, + 360.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 608.0, + 750.0, + 608.0, + 750.0, + 631.0, + 716.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 636.0, + 772.0, + 636.0, + 772.0, + 651.0, + 756.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 632.0, + 1125.0, + 632.0, + 1125.0, + 647.0, + 1110.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 634.0, + 1318.0, + 634.0, + 1318.0, + 644.0, + 1306.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 653.0, + 395.0, + 653.0, + 395.0, + 676.0, + 351.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 662.0, + 1090.0, + 662.0, + 1090.0, + 679.0, + 1066.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 697.0, + 750.0, + 697.0, + 750.0, + 719.0, + 704.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 721.0, + 1104.0, + 721.0, + 1104.0, + 742.0, + 1064.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 725.0, + 1222.0, + 725.0, + 1222.0, + 740.0, + 1208.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 737.0, + 430.0, + 737.0, + 430.0, + 757.0, + 384.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 737.0, + 504.0, + 737.0, + 504.0, + 759.0, + 463.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 737.0, + 579.0, + 737.0, + 579.0, + 759.0, + 538.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 736.0, + 652.0, + 736.0, + 652.0, + 759.0, + 618.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 736.0, + 786.0, + 736.0, + 786.0, + 760.0, + 741.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 737.0, + 913.0, + 737.0, + 913.0, + 759.0, + 872.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 737.0, + 973.0, + 737.0, + 973.0, + 759.0, + 937.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 735.0, + 1233.0, + 735.0, + 1233.0, + 761.0, + 1197.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 737.0, + 1296.0, + 737.0, + 1296.0, + 759.0, + 1260.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 737.0, + 1359.0, + 737.0, + 1359.0, + 759.0, + 1327.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 750.0, + 598.0, + 750.0, + 598.0, + 779.0, + 444.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 754.0, + 954.0, + 754.0, + 954.0, + 777.0, + 799.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 754.0, + 1309.0, + 754.0, + 1309.0, + 777.0, + 1154.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 820.0, + 564.0, + 820.0, + 564.0, + 843.0, + 476.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 821.0, + 927.0, + 821.0, + 927.0, + 846.0, + 824.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 823.0, + 1294.0, + 823.0, + 1294.0, + 843.0, + 1158.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 838.0, + 399.0, + 838.0, + 399.0, + 1011.0, + 334.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 838.0, + 750.0, + 838.0, + 750.0, + 1011.0, + 690.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 844.0, + 1097.0, + 844.0, + 1097.0, + 1010.0, + 1044.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 926.0, + 395.0, + 926.0, + 395.0, + 948.0, + 351.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 930.0, + 419.0, + 930.0, + 419.0, + 948.0, + 399.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 927.0, + 738.0, + 927.0, + 738.0, + 949.0, + 710.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 954.0, + 1097.0, + 954.0, + 1097.0, + 974.0, + 1061.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 970.0, + 394.0, + 970.0, + 394.0, + 991.0, + 351.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 971.0, + 749.0, + 971.0, + 749.0, + 991.0, + 704.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1009.0, + 430.0, + 1009.0, + 430.0, + 1032.0, + 386.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 1009.0, + 521.0, + 1009.0, + 521.0, + 1034.0, + 476.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1009.0, + 611.0, + 1009.0, + 611.0, + 1032.0, + 567.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1009.0, + 786.0, + 1009.0, + 786.0, + 1032.0, + 741.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 1009.0, + 851.0, + 1009.0, + 851.0, + 1034.0, + 805.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1010.0, + 912.0, + 1010.0, + 912.0, + 1031.0, + 871.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1010.0, + 974.0, + 1010.0, + 974.0, + 1032.0, + 940.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1010.0, + 1129.0, + 1010.0, + 1129.0, + 1031.0, + 1090.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1009.0, + 1194.0, + 1009.0, + 1194.0, + 1032.0, + 1156.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1007.0, + 1262.0, + 1007.0, + 1262.0, + 1034.0, + 1225.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1010.0, + 1327.0, + 1010.0, + 1327.0, + 1032.0, + 1293.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 1026.0, + 598.0, + 1026.0, + 598.0, + 1049.0, + 444.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1026.0, + 954.0, + 1026.0, + 954.0, + 1049.0, + 800.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1025.0, + 1305.0, + 1025.0, + 1305.0, + 1050.0, + 1149.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1093.0, + 946.0, + 1093.0, + 946.0, + 1115.0, + 791.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1110.0, + 743.0, + 1110.0, + 743.0, + 1133.0, + 711.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1116.0, + 712.0, + 1116.0, + 712.0, + 1283.0, + 683.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1118.0, + 761.0, + 1118.0, + 761.0, + 1129.0, + 750.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1153.0, + 928.0, + 1153.0, + 928.0, + 1173.0, + 843.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1161.0, + 743.0, + 1161.0, + 743.0, + 1183.0, + 707.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1171.0, + 909.0, + 1171.0, + 909.0, + 1191.0, + 845.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1166.0, + 990.0, + 1166.0, + 990.0, + 1182.0, + 974.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1185.0, + 903.0, + 1185.0, + 903.0, + 1212.0, + 846.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1206.0, + 908.0, + 1206.0, + 908.0, + 1226.0, + 845.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1214.0, + 739.0, + 1214.0, + 739.0, + 1237.0, + 703.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1220.0, + 763.0, + 1220.0, + 763.0, + 1228.0, + 753.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1221.0, + 899.0, + 1221.0, + 899.0, + 1244.0, + 844.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1264.0, + 744.0, + 1264.0, + 744.0, + 1286.0, + 703.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1282.0, + 777.0, + 1282.0, + 777.0, + 1302.0, + 732.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1282.0, + 867.0, + 1282.0, + 867.0, + 1304.0, + 826.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1282.0, + 958.0, + 1282.0, + 958.0, + 1304.0, + 916.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1298.0, + 948.0, + 1298.0, + 948.0, + 1324.0, + 791.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 652.0, + 752.0, + 652.0, + 752.0, + 677.0, + 706.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.75, + 695.5, + 398.75, + 695.5, + 398.75, + 720.5, + 349.75, + 720.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 735.5, + 852.0, + 735.5, + 852.0, + 759.5, + 803.0, + 759.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.75, + 735.5, + 1171.75, + 735.5, + 1171.75, + 759.5, + 1125.75, + 759.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 881.0, + 753.0, + 881.0, + 753.0, + 903.0, + 714.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.25, + 928.5, + 749.25, + 928.5, + 749.25, + 940.0, + 729.25, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1375.0, + 1407.0, + 1375.0, + 1407.0, + 1411.0, + 294.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1405.0, + 1406.0, + 1405.0, + 1406.0, + 1442.0, + 294.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1435.0, + 1402.0, + 1435.0, + 1402.0, + 1468.0, + 294.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1461.0, + 295.0, + 1461.0, + 295.0, + 1502.0, + 292.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1461.0, + 1407.0, + 1461.0, + 1407.0, + 1502.0, + 413.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1495.0, + 689.0, + 1495.0, + 689.0, + 1528.0, + 295.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1495.0, + 752.0, + 1495.0, + 752.0, + 1528.0, + 743.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 230.0, + 1165.0, + 230.0, + 1165.0, + 263.0, + 304.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1796.0, + 1404.0, + 1796.0, + 1404.0, + 1826.0, + 296.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1820.0, + 1405.0, + 1820.0, + 1405.0, + 1860.0, + 292.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1891.0, + 291.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1882.0, + 1406.0, + 1882.0, + 1406.0, + 1917.0, + 291.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1407.0, + 1912.0, + 1407.0, + 1949.0, + 294.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1977.0, + 296.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1975.0, + 1405.0, + 1975.0, + 1405.0, + 2005.0, + 292.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1072.0, + 2003.0, + 1072.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1596.0, + 1250.0, + 1596.0, + 1250.0, + 1634.0, + 292.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1596.0, + 1406.0, + 1596.0, + 1406.0, + 1634.0, + 1278.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1629.0, + 440.0, + 1629.0, + 440.0, + 1660.0, + 295.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1629.0, + 1402.0, + 1629.0, + 1402.0, + 1660.0, + 461.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1657.0, + 674.0, + 1657.0, + 674.0, + 1693.0, + 292.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1657.0, + 748.0, + 1657.0, + 748.0, + 1693.0, + 701.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1693.0, + 770.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1683.0, + 1405.0, + 1683.0, + 1405.0, + 1726.0, + 292.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1718.0, + 344.0, + 1718.0, + 344.0, + 1753.0, + 294.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1718.0, + 645.0, + 1718.0, + 645.0, + 1753.0, + 364.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1718.0, + 1405.0, + 1718.0, + 1405.0, + 1753.0, + 673.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1750.0, + 947.0, + 1750.0, + 947.0, + 1781.0, + 295.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 296.0, + 1405.0, + 296.0, + 1405.0, + 333.0, + 293.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 326.0, + 1235.0, + 326.0, + 1235.0, + 363.0, + 293.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 326.0, + 1405.0, + 326.0, + 1405.0, + 363.0, + 1262.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 357.0, + 738.0, + 357.0, + 738.0, + 394.0, + 293.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 357.0, + 1406.0, + 357.0, + 1406.0, + 394.0, + 760.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 383.0, + 334.0, + 383.0, + 334.0, + 426.0, + 292.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 383.0, + 1407.0, + 383.0, + 1407.0, + 426.0, + 453.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 417.0, + 372.0, + 417.0, + 372.0, + 451.0, + 295.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 417.0, + 1405.0, + 417.0, + 1405.0, + 451.0, + 426.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 447.0, + 1347.0, + 447.0, + 1347.0, + 482.0, + 295.0, + 482.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 34, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1734, + 1405, + 1734, + 1405, + 2034, + 297, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1537, + 1405, + 1537, + 1405, + 1718, + 297, + 1718 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 351, + 495, + 1348, + 495, + 1348, + 1248, + 351, + 1248 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 290, + 1404, + 290, + 1404, + 444, + 298, + 444 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 298, + 1300, + 1400, + 1300, + 1400, + 1392, + 298, + 1392 + ], + "score": 0.945 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 835, + 2089, + 865, + 2089, + 865, + 2112, + 835, + 2112 + ], + "score": 0.856 + }, + { + "category_id": 0, + "poly": [ + 299, + 1479, + 1048, + 1479, + 1048, + 1508, + 299, + 1508 + ], + "score": 0.749 + }, + { + "category_id": 0, + "poly": [ + 304, + 232, + 920, + 232, + 920, + 260, + 304, + 260 + ], + "score": 0.562 + }, + { + "category_id": 1, + "poly": [ + 299, + 1479, + 1048, + 1479, + 1048, + 1508, + 299, + 1508 + ], + "score": 0.168 + }, + { + "category_id": 13, + "poly": [ + 355, + 350, + 561, + 350, + 561, + 385, + 355, + 385 + ], + "score": 0.93, + "latex": "\\{ 1 0 ^ { - 5 } , \\dotsc , 1 0 ^ { - 1 } \\}" + }, + { + "category_id": 13, + "poly": [ + 468, + 1629, + 523, + 1629, + 523, + 1658, + 468, + 1658 + ], + "score": 0.87, + "latex": "8 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 716, + 1335, + 741, + 1335, + 741, + 1363, + 716, + 1363 + ], + "score": 0.84, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 548, + 1799, + 574, + 1799, + 574, + 1826, + 548, + 1826 + ], + "score": 0.83, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 649, + 1336, + 669, + 1336, + 669, + 1363, + 649, + 1363 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1346, + 1307, + 1367, + 1307, + 1367, + 1328, + 1346, + 1328 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 943, + 1337, + 963, + 1337, + 963, + 1358, + 943, + 1358 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1189, + 326, + 1210, + 326, + 1210, + 349, + 1189, + 349 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 785, + 327, + 804, + 327, + 804, + 349, + 785, + 349 + ], + "score": 0.63, + "latex": "\\tau" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 505.0, + 412.0, + 505.0, + 412.0, + 530.0, + 368.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 497.0, + 566.0, + 497.0, + 566.0, + 519.0, + 493.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 506.0, + 747.0, + 506.0, + 747.0, + 528.0, + 714.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 497.0, + 909.0, + 497.0, + 909.0, + 519.0, + 828.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 505.0, + 1108.0, + 505.0, + 1108.0, + 528.0, + 1056.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 497.0, + 1264.0, + 497.0, + 1264.0, + 519.0, + 1180.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 538.0, + 480.0, + 538.0, + 480.0, + 547.0, + 468.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 524.0, + 1065.0, + 524.0, + 1065.0, + 569.0, + 1041.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 537.0, + 1124.0, + 537.0, + 1124.0, + 550.0, + 1110.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 546.0, + 412.0, + 546.0, + 412.0, + 570.0, + 368.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 546.0, + 765.0, + 546.0, + 765.0, + 651.0, + 684.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 549.0, + 820.0, + 549.0, + 820.0, + 562.0, + 805.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 546.0, + 1108.0, + 546.0, + 1108.0, + 570.0, + 1055.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 561.0, + 413.0, + 561.0, + 413.0, + 637.0, + 355.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 559.0, + 1108.0, + 559.0, + 1108.0, + 634.0, + 1041.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 627.0, + 412.0, + 627.0, + 412.0, + 651.0, + 368.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 620.0, + 762.0, + 620.0, + 762.0, + 630.0, + 750.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 622.0, + 1108.0, + 622.0, + 1108.0, + 674.0, + 1043.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 636.0, + 1121.0, + 636.0, + 1121.0, + 647.0, + 1111.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 664.0, + 1283.0, + 664.0, + 1283.0, + 678.0, + 1267.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 673.0, + 612.0, + 673.0, + 612.0, + 702.0, + 450.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 677.0, + 834.0, + 677.0, + 834.0, + 697.0, + 787.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 677.0, + 893.0, + 677.0, + 893.0, + 697.0, + 844.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 677.0, + 949.0, + 677.0, + 949.0, + 698.0, + 900.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 677.0, + 998.0, + 677.0, + 998.0, + 698.0, + 956.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 677.0, + 1191.0, + 677.0, + 1191.0, + 697.0, + 1145.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 677.0, + 1244.0, + 677.0, + 1244.0, + 698.0, + 1197.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 677.0, + 1296.0, + 677.0, + 1296.0, + 698.0, + 1250.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 679.0, + 1344.0, + 679.0, + 1344.0, + 699.0, + 1305.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 693.0, + 583.0, + 693.0, + 583.0, + 719.0, + 476.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 693.0, + 923.0, + 693.0, + 923.0, + 719.0, + 816.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 693.0, + 1275.0, + 693.0, + 1275.0, + 719.0, + 1168.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 767.0, + 414.0, + 767.0, + 414.0, + 795.0, + 367.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 761.0, + 572.0, + 761.0, + 572.0, + 783.0, + 487.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 770.0, + 756.0, + 770.0, + 756.0, + 794.0, + 712.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 760.0, + 923.0, + 760.0, + 923.0, + 786.0, + 824.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 770.0, + 1100.0, + 770.0, + 1100.0, + 794.0, + 1055.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 761.0, + 1284.0, + 761.0, + 1284.0, + 783.0, + 1150.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 805.0, + 430.0, + 805.0, + 430.0, + 901.0, + 354.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 809.0, + 484.0, + 809.0, + 484.0, + 830.0, + 464.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 810.0, + 538.0, + 810.0, + 538.0, + 830.0, + 520.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 816.0, + 592.0, + 816.0, + 592.0, + 829.0, + 576.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 810.0, + 647.0, + 810.0, + 647.0, + 830.0, + 629.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 802.0, + 775.0, + 802.0, + 775.0, + 918.0, + 684.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 817.0, + 883.0, + 817.0, + 883.0, + 840.0, + 865.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 810.0, + 1100.0, + 810.0, + 1100.0, + 904.0, + 1038.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 865.0, + 756.0, + 865.0, + 756.0, + 894.0, + 710.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 891.0, + 413.0, + 891.0, + 413.0, + 918.0, + 368.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 899.0, + 756.0, + 899.0, + 756.0, + 926.0, + 711.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 883.0, + 871.0, + 883.0, + 871.0, + 923.0, + 788.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 891.0, + 1099.0, + 891.0, + 1099.0, + 916.0, + 1057.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 912.0, + 849.0, + 912.0, + 849.0, + 941.0, + 785.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 917.0, + 792.0, + 917.0, + 792.0, + 933.0, + 764.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 942.0, + 498.0, + 942.0, + 498.0, + 963.0, + 450.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 942.0, + 607.0, + 942.0, + 607.0, + 963.0, + 559.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 942.0, + 654.0, + 942.0, + 654.0, + 963.0, + 618.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 942.0, + 840.0, + 942.0, + 840.0, + 963.0, + 795.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 942.0, + 896.0, + 942.0, + 896.0, + 963.0, + 849.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 942.0, + 997.0, + 942.0, + 997.0, + 963.0, + 960.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 942.0, + 1184.0, + 942.0, + 1184.0, + 963.0, + 1138.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 942.0, + 1239.0, + 942.0, + 1239.0, + 963.0, + 1193.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 942.0, + 1294.0, + 942.0, + 1294.0, + 963.0, + 1248.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 945.0, + 1333.0, + 945.0, + 1333.0, + 962.0, + 1307.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 955.0, + 583.0, + 955.0, + 583.0, + 984.0, + 475.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 955.0, + 927.0, + 955.0, + 927.0, + 984.0, + 820.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 956.0, + 1272.0, + 956.0, + 1272.0, + 983.0, + 1164.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1034.0, + 750.0, + 1034.0, + 750.0, + 1058.0, + 706.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1025.0, + 942.0, + 1025.0, + 942.0, + 1048.0, + 790.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1072.0, + 749.0, + 1072.0, + 749.0, + 1101.0, + 704.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1075.0, + 770.0, + 1075.0, + 770.0, + 1107.0, + 748.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1084.0, + 875.0, + 1084.0, + 875.0, + 1107.0, + 857.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1087.0, + 717.0, + 1087.0, + 717.0, + 1166.0, + 691.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1113.0, + 750.0, + 1113.0, + 750.0, + 1142.0, + 704.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1121.0, + 981.0, + 1121.0, + 981.0, + 1130.0, + 971.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1155.0, + 748.0, + 1155.0, + 748.0, + 1180.0, + 706.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1193.0, + 748.0, + 1193.0, + 748.0, + 1222.0, + 706.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1206.0, + 835.0, + 1206.0, + 835.0, + 1226.0, + 787.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1206.0, + 944.0, + 1206.0, + 944.0, + 1227.0, + 896.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1206.0, + 988.0, + 1206.0, + 988.0, + 1227.0, + 953.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 1219.0, + 919.0, + 1219.0, + 919.0, + 1248.0, + 812.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.75, + 666.0, + 445.75, + 666.0, + 445.75, + 702.0, + 367.75, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.75, + 667.0, + 779.75, + 667.0, + 779.75, + 703.0, + 710.75, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.75, + 663.5, + 1139.75, + 663.5, + 1139.75, + 703.0, + 1055.75, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 675.5, + 656.0, + 675.5, + 656.0, + 699.5, + 612.0, + 699.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.25, + 928.0, + 787.25, + 928.0, + 787.25, + 969.0, + 710.25, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.25, + 929.0, + 445.25, + 929.0, + 445.25, + 968.5, + 367.25, + 968.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.25, + 928.0, + 1130.25, + 928.0, + 1130.25, + 968.0, + 1055.25, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 939.5, + 554.0, + 939.5, + 554.0, + 963.5, + 505.0, + 963.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 939.5, + 952.0, + 939.5, + 952.0, + 963.5, + 902.0, + 963.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1078.0, + 818.0, + 1078.0, + 818.0, + 1104.0, + 806.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1203.5, + 781.0, + 1203.5, + 781.0, + 1227.5, + 733.0, + 1227.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.75, + 1203.5, + 892.75, + 1203.5, + 892.75, + 1227.5, + 841.75, + 1227.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 1345.0, + 1297.0, + 1345.0, + 1336.0, + 293.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1297.0, + 1405.0, + 1297.0, + 1405.0, + 1336.0, + 1368.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1332.0, + 648.0, + 1332.0, + 648.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1332.0, + 715.0, + 1332.0, + 715.0, + 1365.0, + 670.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1332.0, + 942.0, + 1332.0, + 942.0, + 1365.0, + 742.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1332.0, + 1404.0, + 1332.0, + 1404.0, + 1365.0, + 964.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1362.0, + 734.0, + 1362.0, + 734.0, + 1392.0, + 296.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1475.0, + 1053.0, + 1475.0, + 1053.0, + 1512.0, + 294.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 926.0, + 230.0, + 926.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1729.0, + 1406.0, + 1729.0, + 1406.0, + 1773.0, + 291.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1764.0, + 1405.0, + 1764.0, + 1405.0, + 1799.0, + 292.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1793.0, + 547.0, + 1793.0, + 547.0, + 1829.0, + 292.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1793.0, + 1406.0, + 1793.0, + 1406.0, + 1829.0, + 575.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1824.0, + 1407.0, + 1824.0, + 1407.0, + 1856.0, + 291.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1851.0, + 1407.0, + 1851.0, + 1407.0, + 1889.0, + 292.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1884.0, + 1406.0, + 1884.0, + 1406.0, + 1919.0, + 295.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1915.0, + 1406.0, + 1915.0, + 1406.0, + 1946.0, + 296.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 929.0, + 2000.0, + 929.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1537.0, + 1406.0, + 1537.0, + 1406.0, + 1572.0, + 294.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1567.0, + 1408.0, + 1567.0, + 1408.0, + 1602.0, + 294.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1592.0, + 1407.0, + 1592.0, + 1407.0, + 1638.0, + 292.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1628.0, + 467.0, + 1628.0, + 467.0, + 1663.0, + 294.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1628.0, + 1405.0, + 1628.0, + 1405.0, + 1663.0, + 524.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1658.0, + 1405.0, + 1658.0, + 1405.0, + 1694.0, + 292.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1685.0, + 1040.0, + 1685.0, + 1040.0, + 1723.0, + 292.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 290.0, + 1404.0, + 290.0, + 1404.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 319.0, + 784.0, + 319.0, + 784.0, + 358.0, + 292.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 319.0, + 1188.0, + 319.0, + 1188.0, + 358.0, + 805.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 358.0, + 1211.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 345.0, + 354.0, + 345.0, + 354.0, + 389.0, + 289.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 345.0, + 1410.0, + 345.0, + 1410.0, + 389.0, + 562.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 378.0, + 1406.0, + 378.0, + 1406.0, + 421.0, + 292.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 411.0, + 1367.0, + 411.0, + 1367.0, + 448.0, + 294.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1475.0, + 1053.0, + 1475.0, + 1053.0, + 1512.0, + 294.0, + 1512.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 35, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 963, + 1404, + 963, + 1404, + 1174, + 298, + 1174 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 737, + 1405, + 737, + 1405, + 949, + 297, + 949 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1354, + 1403, + 1354, + 1403, + 1534, + 298, + 1534 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 534, + 404, + 1165, + 404, + 1165, + 605, + 534, + 605 + ], + "score": 0.979, + "html": "
ADAGRAD CENTRALIZED 88.0ADAM 87.9YOGI 88.0SGDM 87.7SGD 87.7
FED... ADAGRADADAMYOGIAvGMAVG
CONSTANT ηt85.1 85.685.585.284.9
EXPDECAY85.3 86.286.285.885.2
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1188, + 1403, + 1188, + 1403, + 1340, + 298, + 1340 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1852, + 1404, + 1852, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 6, + "poly": [ + 297, + 249, + 1406, + 249, + 1406, + 372, + 297, + 372 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 297, + 1549, + 1404, + 1549, + 1404, + 1822, + 297, + 1822 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 297, + 667, + 892, + 667, + 892, + 706, + 297, + 706 + ], + "score": 0.932 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2112, + 835, + 2112 + ], + "score": 0.852 + }, + { + "category_id": 13, + "poly": [ + 616, + 1641, + 705, + 1641, + 705, + 1672, + 616, + 1672 + ], + "score": 0.92, + "latex": "( x , c , y )" + }, + { + "category_id": 13, + "poly": [ + 397, + 1640, + 486, + 1640, + 486, + 1672, + 397, + 1672 + ], + "score": 0.92, + "latex": "( x , { \\bar { c } } , y )" + }, + { + "category_id": 13, + "poly": [ + 933, + 1884, + 1022, + 1884, + 1022, + 1916, + 933, + 1916 + ], + "score": 0.92, + "latex": "( x , c , y )" + }, + { + "category_id": 13, + "poly": [ + 717, + 1974, + 822, + 1974, + 822, + 2007, + 717, + 2007 + ], + "score": 0.92, + "latex": "\\operatorname { D i r } ( \\gamma , k )" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1946, + 1092, + 1946, + 1092, + 1972, + 1019, + 1972 + ], + "score": 0.91, + "latex": "v \\in G" + }, + { + "category_id": 13, + "poly": [ + 570, + 1915, + 643, + 1915, + 643, + 1945, + 570, + 1945 + ], + "score": 0.91, + "latex": "y \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 1171, + 1944, + 1240, + 1944, + 1240, + 1977, + 1171, + 1977 + ], + "score": 0.91, + "latex": "| G [ v ] |" + }, + { + "category_id": 13, + "poly": [ + 594, + 1974, + 670, + 1974, + 670, + 2006, + 594, + 2006 + ], + "score": 0.91, + "latex": "\\gamma \\in \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1287, + 1915, + 1320, + 1915, + 1320, + 1947, + 1287, + 1947 + ], + "score": 0.89, + "latex": "\\mathcal { S } _ { y }" + }, + { + "category_id": 13, + "poly": [ + 1205, + 1915, + 1236, + 1915, + 1236, + 1944, + 1205, + 1944 + ], + "score": 0.89, + "latex": "S _ { c }" + }, + { + "category_id": 13, + "poly": [ + 856, + 1917, + 920, + 1917, + 920, + 1945, + 856, + 1945 + ], + "score": 0.89, + "latex": "y \\in c" + }, + { + "category_id": 13, + "poly": [ + 938, + 1670, + 966, + 1670, + 966, + 1700, + 938, + 1700 + ], + "score": 0.88, + "latex": "\\theta _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1336, + 1884, + 1400, + 1884, + 1400, + 1912, + 1336, + 1912 + ], + "score": 0.88, + "latex": "c \\in { \\mathcal { C } }" + }, + { + "category_id": 13, + "poly": [ + 478, + 1218, + 574, + 1218, + 574, + 1252, + 478, + 1252 + ], + "score": 0.88, + "latex": "\\left( \\operatorname { D i r } ( \\alpha ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 446, + 340, + 509, + 340, + 509, + 370, + 446, + 370 + ], + "score": 0.87, + "latex": "\\mathrm { { \\bar { 0 } . 5 \\% } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1760, + 325, + 1760, + 325, + 1790, + 297, + 1790 + ], + "score": 0.86, + "latex": "\\theta _ { r }" + }, + { + "category_id": 13, + "poly": [ + 690, + 1476, + 711, + 1476, + 711, + 1506, + 690, + 1506 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 296, + 1248, + 393, + 1248, + 393, + 1282, + 296, + 1282 + ], + "score": 0.84, + "latex": "\\bar { ( } \\mathrm { D i r } ( \\beta ) )" + }, + { + "category_id": 13, + "poly": [ + 1037, + 1855, + 1062, + 1855, + 1062, + 1883, + 1037, + 1883 + ], + "score": 0.84, + "latex": "\\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 1218, + 1144, + 1238, + 1144, + 1238, + 1175, + 1218, + 1175 + ], + "score": 0.83, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 547, + 1945, + 569, + 1945, + 569, + 1971, + 547, + 1971 + ], + "score": 0.83, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 546, + 282, + 590, + 282, + 590, + 313, + 546, + 313 + ], + "score": 0.82, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 647, + 251, + 690, + 251, + 690, + 282, + 647, + 282 + ], + "score": 0.82, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 1382, + 1975, + 1401, + 1975, + 1401, + 2001, + 1382, + 2001 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 525, + 1675, + 544, + 1675, + 544, + 1701, + 525, + 1701 + ], + "score": 0.81, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1050, + 315, + 1075, + 315, + 1075, + 343, + 1050, + 343 + ], + "score": 0.81, + "latex": "\\eta _ { l }" + }, + { + "category_id": 13, + "poly": [ + 948, + 1732, + 973, + 1732, + 973, + 1757, + 948, + 1757 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1379, + 1855, + 1401, + 1855, + 1401, + 1881, + 1379, + 1881 + ], + "score": 0.8, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 975, + 1915, + 1121, + 1915, + 1121, + 1945, + 975, + 1945 + ], + "score": 0.8, + "latex": "c \\in \\mathcal { C } , y \\in \\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 1260, + 1761, + 1286, + 1761, + 1286, + 1788, + 1260, + 1788 + ], + "score": 0.79, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 973, + 1704, + 992, + 1704, + 992, + 1731, + 973, + 1731 + ], + "score": 0.79, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 708, + 1449, + 730, + 1449, + 730, + 1472, + 708, + 1472 + ], + "score": 0.79, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1220, + 1645, + 1239, + 1645, + 1239, + 1671, + 1220, + 1671 + ], + "score": 0.78, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 633, + 1855, + 652, + 1855, + 652, + 1881, + 633, + 1881 + ], + "score": 0.78, + "latex": "\\mathcal { C }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1889, + 1120, + 1889, + 1120, + 1911, + 1100, + 1911 + ], + "score": 0.78, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1294, + 1055, + 1318, + 1055, + 1318, + 1082, + 1294, + 1082 + ], + "score": 0.78, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 769, + 1950, + 785, + 1950, + 785, + 1971, + 769, + 1971 + ], + "score": 0.77, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 605, + 1732, + 621, + 1732, + 621, + 1761, + 605, + 1761 + ], + "score": 0.77, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 607, + 1615, + 627, + 1615, + 627, + 1641, + 607, + 1641 + ], + "score": 0.77, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 512, + 1975, + 536, + 1975, + 536, + 2001, + 512, + 2001 + ], + "score": 0.76, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1118, + 1230, + 1118, + 1230, + 1140, + 1208, + 1140 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 945, + 1949, + 963, + 1949, + 963, + 1975, + 945, + 1975 + ], + "score": 0.76, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 804, + 1644, + 831, + 1644, + 831, + 1667, + 804, + 1667 + ], + "score": 0.76, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 460, + 1979, + 478, + 1979, + 478, + 2001, + 460, + 2001 + ], + "score": 0.75, + "latex": "v" + }, + { + "category_id": 13, + "poly": [ + 1189, + 1702, + 1203, + 1702, + 1203, + 1727, + 1189, + 1727 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 951, + 1616, + 978, + 1616, + 978, + 1639, + 951, + 1639 + ], + "score": 0.74, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 824, + 1616, + 840, + 1616, + 840, + 1638, + 824, + 1638 + ], + "score": 0.74, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 871, + 1732, + 884, + 1732, + 884, + 1757, + 871, + 1757 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1092, + 1674, + 1120, + 1674, + 1120, + 1698, + 1092, + 1698 + ], + "score": 0.71, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 298, + 1670, + 323, + 1670, + 323, + 1698, + 298, + 1698 + ], + "score": 0.68, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 498, + 1765, + 514, + 1765, + 514, + 1787, + 498, + 1787 + ], + "score": 0.65, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 358, + 1705, + 374, + 1705, + 374, + 1728, + 358, + 1728 + ], + "score": 0.62, + "latex": "c" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 249.0, + 646.0, + 249.0, + 646.0, + 287.0, + 291.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 249.0, + 1407.0, + 249.0, + 1407.0, + 287.0, + 691.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 283.0, + 545.0, + 283.0, + 545.0, + 315.0, + 295.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 283.0, + 1406.0, + 283.0, + 1406.0, + 315.0, + 591.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 308.0, + 1049.0, + 308.0, + 1049.0, + 345.0, + 292.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 308.0, + 1406.0, + 308.0, + 1406.0, + 345.0, + 1076.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 341.0, + 445.0, + 341.0, + 445.0, + 372.0, + 294.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 341.0, + 912.0, + 341.0, + 912.0, + 372.0, + 510.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 668.0, + 892.0, + 668.0, + 892.0, + 709.0, + 293.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 963.0, + 1405.0, + 963.0, + 1405.0, + 997.0, + 294.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 994.0, + 1404.0, + 994.0, + 1404.0, + 1025.0, + 296.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1022.0, + 1405.0, + 1022.0, + 1405.0, + 1056.0, + 293.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1053.0, + 1293.0, + 1053.0, + 1293.0, + 1087.0, + 294.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 1053.0, + 1406.0, + 1053.0, + 1406.0, + 1087.0, + 1319.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1081.0, + 1405.0, + 1081.0, + 1405.0, + 1115.0, + 293.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1111.0, + 1207.0, + 1111.0, + 1207.0, + 1146.0, + 295.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1111.0, + 1405.0, + 1111.0, + 1405.0, + 1146.0, + 1231.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1140.0, + 1217.0, + 1140.0, + 1217.0, + 1177.0, + 293.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1140.0, + 1252.0, + 1140.0, + 1252.0, + 1177.0, + 1239.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 736.0, + 422.0, + 736.0, + 422.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 736.0, + 1407.0, + 736.0, + 1407.0, + 772.0, + 423.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 766.0, + 1406.0, + 766.0, + 1406.0, + 800.0, + 294.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 797.0, + 1405.0, + 797.0, + 1405.0, + 831.0, + 294.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 827.0, + 1407.0, + 827.0, + 1407.0, + 860.0, + 292.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 856.0, + 1406.0, + 856.0, + 1406.0, + 891.0, + 294.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 1405.0, + 887.0, + 1405.0, + 921.0, + 295.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 917.0, + 1144.0, + 917.0, + 1144.0, + 951.0, + 295.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1354.0, + 1405.0, + 1354.0, + 1405.0, + 1389.0, + 294.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1384.0, + 1405.0, + 1384.0, + 1405.0, + 1419.0, + 294.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1414.0, + 1407.0, + 1414.0, + 1407.0, + 1449.0, + 293.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1444.0, + 707.0, + 1444.0, + 707.0, + 1479.0, + 293.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1444.0, + 1408.0, + 1444.0, + 1408.0, + 1479.0, + 731.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1474.0, + 689.0, + 1474.0, + 689.0, + 1509.0, + 294.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1474.0, + 1407.0, + 1474.0, + 1407.0, + 1509.0, + 712.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1504.0, + 913.0, + 1504.0, + 913.0, + 1535.0, + 294.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1188.0, + 1405.0, + 1188.0, + 1405.0, + 1221.0, + 297.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1217.0, + 477.0, + 1217.0, + 477.0, + 1252.0, + 293.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1217.0, + 1405.0, + 1217.0, + 1405.0, + 1252.0, + 575.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1246.0, + 1409.0, + 1246.0, + 1409.0, + 1286.0, + 394.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1280.0, + 1405.0, + 1280.0, + 1405.0, + 1313.0, + 294.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1307.0, + 853.0, + 1307.0, + 853.0, + 1343.0, + 294.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 632.0, + 1851.0, + 632.0, + 1886.0, + 295.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1851.0, + 1036.0, + 1851.0, + 1036.0, + 1886.0, + 653.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1851.0, + 1378.0, + 1851.0, + 1378.0, + 1886.0, + 1063.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1851.0, + 1406.0, + 1851.0, + 1406.0, + 1886.0, + 1402.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 932.0, + 1881.0, + 932.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1881.0, + 1099.0, + 1881.0, + 1099.0, + 1917.0, + 1023.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 1881.0, + 1335.0, + 1881.0, + 1335.0, + 1917.0, + 1121.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 1401.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 569.0, + 1913.0, + 569.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1913.0, + 855.0, + 1913.0, + 855.0, + 1948.0, + 644.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1913.0, + 974.0, + 1913.0, + 974.0, + 1948.0, + 921.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1913.0, + 1204.0, + 1913.0, + 1204.0, + 1948.0, + 1122.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1913.0, + 1286.0, + 1913.0, + 1286.0, + 1948.0, + 1237.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1948.0, + 1321.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 546.0, + 1943.0, + 546.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1943.0, + 768.0, + 1943.0, + 768.0, + 1978.0, + 570.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1943.0, + 944.0, + 1943.0, + 944.0, + 1978.0, + 786.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1943.0, + 1018.0, + 1943.0, + 1018.0, + 1978.0, + 964.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1943.0, + 1170.0, + 1943.0, + 1170.0, + 1978.0, + 1093.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1978.0, + 1241.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 459.0, + 1971.0, + 459.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 1971.0, + 511.0, + 1971.0, + 511.0, + 2010.0, + 479.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1971.0, + 593.0, + 1971.0, + 593.0, + 2010.0, + 537.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1971.0, + 716.0, + 1971.0, + 716.0, + 2010.0, + 671.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1971.0, + 1381.0, + 1971.0, + 1381.0, + 2010.0, + 823.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1971.0, + 1409.0, + 1971.0, + 1409.0, + 2010.0, + 1402.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 424.0, + 2003.0, + 424.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1550.0, + 1404.0, + 1550.0, + 1404.0, + 1582.0, + 297.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1578.0, + 1406.0, + 1578.0, + 1406.0, + 1617.0, + 294.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1610.0, + 606.0, + 1610.0, + 606.0, + 1643.0, + 295.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1610.0, + 823.0, + 1610.0, + 823.0, + 1643.0, + 628.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1610.0, + 950.0, + 1610.0, + 950.0, + 1643.0, + 841.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1610.0, + 1404.0, + 1610.0, + 1404.0, + 1643.0, + 979.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1639.0, + 396.0, + 1639.0, + 396.0, + 1675.0, + 294.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1639.0, + 615.0, + 1639.0, + 615.0, + 1675.0, + 487.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1639.0, + 803.0, + 1639.0, + 803.0, + 1675.0, + 706.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 1639.0, + 1219.0, + 1639.0, + 1219.0, + 1675.0, + 832.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1639.0, + 1405.0, + 1639.0, + 1405.0, + 1675.0, + 1240.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 297.0, + 1665.0, + 297.0, + 1708.0, + 294.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1665.0, + 524.0, + 1665.0, + 524.0, + 1708.0, + 324.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1665.0, + 937.0, + 1665.0, + 937.0, + 1708.0, + 545.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1665.0, + 1091.0, + 1665.0, + 1091.0, + 1708.0, + 967.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 1665.0, + 1406.0, + 1665.0, + 1406.0, + 1708.0, + 1121.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1697.0, + 357.0, + 1697.0, + 357.0, + 1736.0, + 292.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1697.0, + 972.0, + 1697.0, + 972.0, + 1736.0, + 375.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1697.0, + 1188.0, + 1697.0, + 1188.0, + 1736.0, + 993.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1697.0, + 1405.0, + 1697.0, + 1405.0, + 1736.0, + 1204.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1727.0, + 604.0, + 1727.0, + 604.0, + 1765.0, + 292.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1727.0, + 870.0, + 1727.0, + 870.0, + 1765.0, + 622.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1727.0, + 947.0, + 1727.0, + 947.0, + 1765.0, + 885.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1727.0, + 1406.0, + 1727.0, + 1406.0, + 1765.0, + 974.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1757.0, + 497.0, + 1757.0, + 497.0, + 1796.0, + 326.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1757.0, + 1259.0, + 1757.0, + 1259.0, + 1796.0, + 515.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1757.0, + 1405.0, + 1757.0, + 1405.0, + 1796.0, + 1287.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1787.0, + 1410.0, + 1787.0, + 1410.0, + 1828.0, + 291.0, + 1828.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 36, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1134, + 1406, + 1134, + 1406, + 1408, + 296, + 1408 + ], + "score": 0.981 + }, + { + "category_id": 3, + "poly": [ + 578, + 1531, + 1071, + 1531, + 1071, + 1928, + 578, + 1928 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 296, + 1958, + 1402, + 1958, + 1402, + 2019, + 296, + 2019 + ], + "score": 0.943 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2113, + 835, + 2113 + ], + "score": 0.867 + }, + { + "category_id": 0, + "poly": [ + 299, + 229, + 811, + 229, + 811, + 261, + 299, + 261 + ], + "score": 0.628 + }, + { + "category_id": 5, + "poly": [ + 295, + 855, + 1405, + 855, + 1405, + 1077, + 295, + 1077 + ], + "score": 0.49, + "html": "
Algorithm8RENORMALIZE
Initialization: θ= (p1,...,Pk),i∈[K]
a=∑k≠iPk
fork ∈[K],k≠ido
Pk=pk/a
Return0' = (pi,···P'-1,P'+1,···,PK)
" + }, + { + "category_id": 1, + "poly": [ + 299, + 265, + 866, + 265, + 866, + 822, + 299, + 822 + ], + "score": 0.322 + }, + { + "category_id": 8, + "poly": [ + 299, + 265, + 866, + 265, + 866, + 822, + 299, + 822 + ], + "score": 0.161 + }, + { + "category_id": 6, + "poly": [ + 299, + 229, + 811, + 229, + 811, + 261, + 299, + 261 + ], + "score": 0.123 + }, + { + "category_id": 1, + "poly": [ + 307, + 898, + 784, + 898, + 784, + 1075, + 307, + 1075 + ], + "score": 0.101 + }, + { + "category_id": 13, + "poly": [ + 412, + 1166, + 620, + 1166, + 620, + 1199, + 412, + 1199 + ], + "score": 0.91, + "latex": "m \\in \\{ 1 , \\cdots , M \\}" + }, + { + "category_id": 13, + "poly": [ + 1222, + 1137, + 1270, + 1137, + 1270, + 1167, + 1222, + 1167 + ], + "score": 0.9, + "latex": "D _ { m }" + }, + { + "category_id": 13, + "poly": [ + 408, + 456, + 553, + 456, + 553, + 485, + 408, + 485 + ], + "score": 0.83, + "latex": "n = 1 , \\cdots N" + }, + { + "category_id": 13, + "poly": [ + 485, + 544, + 627, + 544, + 627, + 574, + 485, + 574 + ], + "score": 0.82, + "latex": "( x , c , y ) \\in S" + }, + { + "category_id": 13, + "poly": [ + 498, + 387, + 714, + 387, + 714, + 420, + 498, + 420 + ], + "score": 0.81, + "latex": "\\theta _ { c } \\sim \\operatorname { D i r } ( \\beta , | G [ c ] | )" + }, + { + "category_id": 13, + "poly": [ + 528, + 1199, + 634, + 1199, + 634, + 1225, + 528, + 1225 + ], + "score": 0.8, + "latex": "N = 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 410, + 358, + 553, + 358, + 553, + 388, + 410, + 388 + ], + "score": 0.78, + "latex": "c \\in { \\mathcal { C } } \\cap G [ r ]" + }, + { + "category_id": 13, + "poly": [ + 728, + 1137, + 757, + 1137, + 757, + 1165, + 728, + 1165 + ], + "score": 0.78, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 644, + 1197, + 755, + 1197, + 755, + 1226, + 644, + 1226 + ], + "score": 0.76, + "latex": "M = 5 0 0" + }, + { + "category_id": 13, + "poly": [ + 456, + 327, + 675, + 327, + 675, + 359, + 456, + 359 + ], + "score": 0.76, + "latex": "\\theta _ { r } \\sim \\operatorname { D i r } ( \\alpha , | G [ r ] | )" + }, + { + "category_id": 13, + "poly": [ + 408, + 1031, + 783, + 1031, + 783, + 1068, + 408, + 1068 + ], + "score": 0.75, + "latex": "\\theta ^ { \\prime } = ( p _ { 1 } ^ { \\prime } , \\cdot \\cdot \\cdot p _ { i - 1 } ^ { \\prime } , p _ { i + 1 } ^ { \\prime } , \\cdot \\cdot \\cdot , p _ { K } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 344, + 1137, + 377, + 1137, + 377, + 1165, + 344, + 1165 + ], + "score": 0.72, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 766, + 1197, + 952, + 1197, + 952, + 1228, + 766, + 1228 + ], + "score": 0.68, + "latex": "\\alpha = 0 . 1 , \\beta = 1 0" + }, + { + "category_id": 13, + "poly": [ + 432, + 634, + 516, + 634, + 516, + 665, + 432, + 665 + ], + "score": 0.68, + "latex": "\\begin{array} { r } { { \\cal S } _ { y } = \\emptyset } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 482, + 899, + 776, + 899, + 776, + 932, + 482, + 932 + ], + "score": 0.63, + "latex": "\\theta = ( p _ { 1 } , \\ldots , p _ { K } ) , i \\in [ K ]" + }, + { + "category_id": 13, + "poly": [ + 693, + 515, + 736, + 515, + 736, + 545, + 693, + 545 + ], + "score": 0.62, + "latex": "\\left( \\theta _ { c } \\right)" + }, + { + "category_id": 13, + "poly": [ + 409, + 573, + 684, + 573, + 684, + 603, + 409, + 603 + ], + "score": 0.59, + "latex": "D _ { m } = D _ { m } \\cup \\{ ( x , c , y ) \\}" + }, + { + "category_id": 14, + "poly": [ + 366, + 422, + 467, + 422, + 467, + 454, + 366, + 454 + ], + "score": 0.57, + "latex": "D _ { m } = \\varnothing" + }, + { + "category_id": 13, + "poly": [ + 367, + 299, + 545, + 299, + 545, + 327, + 367, + 327 + ], + "score": 0.52, + "latex": "m = 1 , \\cdots , M" + }, + { + "category_id": 13, + "poly": [ + 395, + 267, + 694, + 267, + 694, + 299, + 395, + 299 + ], + "score": 0.5, + "latex": "N , M \\in \\mathbb { Z } _ { > 0 } , \\alpha , \\beta \\in \\mathbb { R } _ { \\ge 0 }" + }, + { + "category_id": 13, + "poly": [ + 689, + 484, + 734, + 484, + 734, + 514, + 689, + 514 + ], + "score": 0.49, + "latex": "\\left( \\theta _ { r } \\right)" + }, + { + "category_id": 13, + "poly": [ + 365, + 422, + 467, + 422, + 467, + 454, + 365, + 454 + ], + "score": 0.41, + "latex": "D _ { m } = \\varnothing" + }, + { + "category_id": 13, + "poly": [ + 408, + 603, + 617, + 603, + 617, + 635, + 408, + 635 + ], + "score": 0.35, + "latex": "\\mathcal { S } = \\mathcal { S } \\backslash \\{ ( x , c , \\overset { \\cdot } { y } ) \\}" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1531.0, + 1060.0, + 1531.0, + 1060.0, + 1564.0, + 648.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1549.0, + 643.0, + 1549.0, + 643.0, + 1583.0, + 603.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1593.0, + 641.0, + 1593.0, + 641.0, + 1631.0, + 604.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1630.0, + 645.0, + 1630.0, + 645.0, + 1807.0, + 601.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1648.0, + 615.0, + 1648.0, + 615.0, + 1788.0, + 577.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1813.0, + 642.0, + 1813.0, + 642.0, + 1849.0, + 604.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1862.0, + 637.0, + 1862.0, + 637.0, + 1885.0, + 619.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1875.0, + 735.0, + 1875.0, + 735.0, + 1901.0, + 699.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1873.0, + 833.0, + 1873.0, + 833.0, + 1902.0, + 796.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1873.0, + 930.0, + 1873.0, + 930.0, + 1902.0, + 893.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1875.0, + 1025.0, + 1875.0, + 1025.0, + 1901.0, + 991.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1897.0, + 999.0, + 1897.0, + 999.0, + 1928.0, + 707.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1953.0, + 1407.0, + 1953.0, + 1407.0, + 1996.0, + 293.0, + 1996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1987.0, + 635.0, + 1987.0, + 635.0, + 2021.0, + 295.0, + 2021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 817.0, + 225.0, + 817.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 817.0, + 225.0, + 817.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1133.0, + 343.0, + 1133.0, + 343.0, + 1172.0, + 293.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1133.0, + 727.0, + 1133.0, + 727.0, + 1172.0, + 378.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1133.0, + 1221.0, + 1133.0, + 1221.0, + 1172.0, + 758.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1133.0, + 1406.0, + 1133.0, + 1406.0, + 1172.0, + 1271.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1163.0, + 411.0, + 1163.0, + 411.0, + 1202.0, + 293.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1163.0, + 1407.0, + 1163.0, + 1407.0, + 1202.0, + 621.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1196.0, + 527.0, + 1196.0, + 527.0, + 1231.0, + 293.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1196.0, + 643.0, + 1196.0, + 643.0, + 1231.0, + 635.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1196.0, + 765.0, + 1196.0, + 765.0, + 1231.0, + 756.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1196.0, + 1407.0, + 1196.0, + 1407.0, + 1231.0, + 953.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1224.0, + 1406.0, + 1224.0, + 1406.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1251.0, + 1406.0, + 1251.0, + 1406.0, + 1293.0, + 293.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1283.0, + 1408.0, + 1283.0, + 1408.0, + 1320.0, + 293.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1313.0, + 1404.0, + 1313.0, + 1404.0, + 1352.0, + 293.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1345.0, + 1404.0, + 1345.0, + 1404.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1376.0, + 875.0, + 1376.0, + 875.0, + 1409.0, + 295.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 261.0, + 394.0, + 261.0, + 394.0, + 306.0, + 322.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 295.0, + 366.0, + 295.0, + 366.0, + 329.0, + 323.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 295.0, + 583.0, + 295.0, + 583.0, + 329.0, + 546.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 324.0, + 455.0, + 324.0, + 455.0, + 364.0, + 362.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 324.0, + 680.0, + 324.0, + 680.0, + 364.0, + 676.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 353.0, + 409.0, + 353.0, + 409.0, + 393.0, + 362.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 353.0, + 599.0, + 353.0, + 599.0, + 393.0, + 554.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 379.0, + 497.0, + 379.0, + 497.0, + 426.0, + 401.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 379.0, + 720.0, + 379.0, + 720.0, + 426.0, + 715.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 421.0, + 473.0, + 421.0, + 473.0, + 456.0, + 468.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 450.0, + 407.0, + 450.0, + 407.0, + 487.0, + 363.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 450.0, + 596.0, + 450.0, + 596.0, + 487.0, + 554.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 478.0, + 688.0, + 478.0, + 688.0, + 517.0, + 404.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 478.0, + 738.0, + 478.0, + 738.0, + 517.0, + 735.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 510.0, + 692.0, + 510.0, + 692.0, + 547.0, + 405.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 539.0, + 484.0, + 539.0, + 484.0, + 578.0, + 405.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 539.0, + 866.0, + 539.0, + 866.0, + 578.0, + 628.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 567.0, + 408.0, + 567.0, + 408.0, + 611.0, + 404.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 567.0, + 689.0, + 567.0, + 689.0, + 611.0, + 685.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 595.0, + 407.0, + 595.0, + 407.0, + 642.0, + 403.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 595.0, + 622.0, + 595.0, + 622.0, + 642.0, + 618.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 629.0, + 431.0, + 629.0, + 431.0, + 667.0, + 403.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 629.0, + 584.0, + 629.0, + 584.0, + 667.0, + 517.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 653.0, + 596.0, + 653.0, + 596.0, + 700.0, + 439.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 683.0, + 767.0, + 683.0, + 767.0, + 728.0, + 443.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 718.0, + 621.0, + 718.0, + 621.0, + 755.0, + 445.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 747.0, + 632.0, + 747.0, + 632.0, + 789.0, + 484.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 775.0, + 809.0, + 775.0, + 809.0, + 818.0, + 485.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 893.0, + 481.0, + 893.0, + 481.0, + 935.0, + 321.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 925.0, + 487.0, + 925.0, + 487.0, + 970.0, + 321.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 959.0, + 574.0, + 959.0, + 574.0, + 998.0, + 321.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 992.0, + 497.0, + 992.0, + 497.0, + 1029.0, + 361.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1027.0, + 407.0, + 1027.0, + 407.0, + 1069.0, + 320.0, + 1069.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 37, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/SJeS16EKPr/SJeS16EKPr.md b/parse/train/SJeS16EKPr/SJeS16EKPr.md new file mode 100644 index 0000000000000000000000000000000000000000..1338051f39952ca027b07ffec6187f748d60451c --- /dev/null +++ b/parse/train/SJeS16EKPr/SJeS16EKPr.md @@ -0,0 +1,668 @@ +# LEARNING RELEVANT FEATURES FOR STATISTICAL INFERENCE + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Given two views of data, we consider the problem of finding the features of one view which can be most faithfully inferred from the other. We find that these are also the most correlated variables in the sense of deep canonical correlation analysis (DCCA). Moreover, we show that these variables can be used to construct a non-parametric representation of the implied joint probability distribution. This representation can be used to compute the expectations of functions over one view of data conditioned on the other, such as Bayesian estimators and their standard deviations. We test the approach using inference on occluded MNIST images, and show that our representation contains multiple modes. Surprisingly, when applied to supervised learning (one dataset consists of labels), this approach automatically provides regularization and faster convergence compared to the cross-entropy objective. We also explore using this approach to discover salient independent variables of a single dataset. + +# 1 INTRODUCTION + +Given samples $( x _ { 1 } , y _ { 1 } ) , \dotsc , ( x _ { n } , y _ { n } )$ from an unknown joint probability distribution $p ( x , y )$ , we want to construct a useful representation of the conditional probabilities $p ( x | y )$ and $p ( y | x )$ , so that we that we can infer one view from the other on new data. + +For instance, $x$ and $y$ could be past and future histories of dynamical data, visual and auditory inputs, actions and their effects, etc. + +Following the approach introduced in Beny & Osborne (2013; 2015b) in the context of quantum ´ information theory, we look at the problem as follows: the conditional distributions $p ( y | x )$ can be thought of as representing a noisy communication channel (stochastic map). This channel is a linear map between spaces of typically ludicrously large dimensions (the spaces of all probability distributions over $x$ or $y$ ). We want a pair of small subspaces which best represent the channel. Specifically, we look for those vectors representing probability distributions over $x$ which lose least distinguishability under the channel, where the distinguishability is measured by the $\chi ^ { 2 }$ divergence. + +We show in Section 3 that this is equivalent to performing a certain singular value decomposition of the channel (seen as an operator in Hilbert space) and keep only the components with the largest singular values. Moreover, the full singular value decomposition is equivalent to the decomposition in terms of canonical variables introduced in Lancaster (1958), namely, + +$$ +p ( \boldsymbol { y } | \boldsymbol { x } ) = p ( \boldsymbol { y } ) \sum _ { i = 1 } ^ { D } \eta _ { i } u _ { i } ( \boldsymbol { x } ) v _ { i } ( \boldsymbol { y } ) , +$$ + +where $u _ { i }$ and $u _ { j }$ are real non-linear functions such that $\mathbb { E } ( u _ { i } u _ { j } ) = \delta _ { i j }$ , $\mathbb { E } ( v _ { i } v _ { j } ) = \delta _ { i j }$ , and $0 ~ <$ $\eta _ { D } \leq \cdot \cdot \cdot \leq \eta _ { 1 } \leq \eta _ { 0 } = 1$ are the singular values. + +The practical advantage of this representation for inference (or prediction) is that it reduces the evaluation of conditional expectations to that of empirical averages over the (unconditional) marginal $p ( y )$ . + +As observed in Michaeli et al. (2016), the span of the first $k$ canonical variables $u _ { i } , \ v _ { j }$ is what is learned by the deep canonical correlation analysis (DCCA) (Andrew et al., 2013). Indeed, these + +variables are those which maximize the correlations $\mathbb { E } ( u _ { i } v _ { i } )$ subject to the same constraints as above. +(This reduces to CCA (Hotelling, 1936) when $u _ { i }$ , $v _ { j }$ are linear maps). + +In this work, besides establishing this new information-theoretical interpretation of canonical variables and DCCA, we experiment with using this representation for performing inference on new data. Moreover, we propose a strategy for extracting disentangled variables from the canonical variables, inspired by analytical solutions. + +# 2 RELATED WORK + +This general problem (of building an effective representation of the conditional probability distributions implied by joint samples) covers many existing approaches in different contexts. For instance, if the variables $y$ has few possible states, then it reduces to a classification problem, usually solved by minimizing the crossentropy between a predicted distribution and the one-hot encoding of the classes. + +When $y$ has a large number of states, or is fundamentally continuous, existing approaches usually do not model the whole conditional distribution, but either provide the average (regression), or approximately sample from it. + +The main class of methods which allows for sampling from the conditional distributions are variational: a deterministic neural networks produces the parameters of analytical classes of probabilities. +This includes variational autoencoders Kingma & Welling (2013) (e.g., Iten et al. (2018); Wang et al. +(2016)), and approaches based on the minimum description length principle such as Gregor et al. +(2013). + +Alternatively, it may also be possible to use adversarial training Goodfellow et al. (2014), by using a conditional Mirza & Osindero (2014) version of an energy-based GAN Zhao et al. (2016). + +By contrast, our approach doesn’t require the training of a generative model. Instead, conditional expectations are constructed as linear combinations of unconditional empirical averages over the training data. + +Previous approaches to equipping CCA or DCCA with an information-theoretic interpretation have explored different directions. For instance, in Wang et al. (2016), the authors generalize a probabilistic interpretation for CCA in terms of gaussian distributions, which leads to a variational approach. In Painsky et al. (2018), additional constraints on the mutual information between the data and the learned variables are added to the optimizations. + +A previous attempt at designing a numerical solution for our singular value problem can be found in Beny (2018b). In that work, the relevant variables were represented through a PCA kernel produced ´ by Monte Carlo sampling, but wasn’t practical. + +# 3 THEORY + +We formalize the problem by assuming that our data was sampled from an unknown joint distribution $p ( x , y )$ over two random variables $X$ and $Y$ . + +Let $V _ { X }$ and $V _ { Y }$ denote the linear spaces spanned by all probability distributions over $X$ and $Y$ respectively. Here we assume that $X$ and $Y$ take finitely many values for simplicity, but this formalism can be straightforwardly extended to infinite-dimensional vector spaces. + +We will need inner products on $V _ { X }$ and $V _ { Y }$ , to make them into real Hilbert spaces. We use the Fisher information metrics evaluated at the points $p ( x )$ and $p ( y )$ respectively (marginals of $p ( x , y ) \mathrm { . }$ ), that is, + +$$ +\langle \mu , \mu ^ { \prime } \rangle _ { X } : = \sum _ { x } { \frac { \mu ( x ) \mu ^ { \prime } ( x ) } { p ( x ) } } \quad { \mathrm { a n d } } \quad \langle \nu , \nu ^ { \prime } \rangle _ { Y } : = \sum _ { y } { \frac { \nu ( y ) \nu ^ { \prime } ( y ) } { p ( y ) } } +$$ + +for any vectors $\mu , \mu ^ { \prime } \in V _ { X }$ and $\nu , \nu ^ { \prime } \in V _ { Y }$ . + +Below we also call the marginals $p _ { X }$ and $p _ { Y }$ respectively when omitting their arguments $( p _ { X } ( x ) \equiv$ $p ( x )$ and $p _ { Y } ( y ) \equiv p ( y ) \rangle$ . + +These inner products allow us to define the $\chi ^ { 2 }$ divergence: + +$$ +\chi ^ { 2 } ( q , p _ { X } ) = \langle q - p _ { X } , q - p _ { X } \rangle _ { X } , +$$ + +which a measures of statistical distinguishability between $q$ and $p _ { X }$ . Specifically, it quantifies how easy it is to reject the null hypothesis that the state is $p _ { X }$ when it is actually $q$ , based on the empirical distribution obtained from independent samples. It is also the lowest order approximation of the Kullback-Leibler divergence. + +The joint distribution $p ( x , y )$ yields conditional distributions $p ( y | x )$ and $p ( x | y )$ . These can be understood as the components (or kernels) of stochastic maps $\mathcal { N } : V _ { X } V _ { Y }$ and $\mathcal { N ^ { * } } : V _ { X } V _ { Y }$ respectively. Explicitely, if $\mu \in V _ { X }$ and $\nu \in V _ { Y }$ , then the images $\mathcal { N } ( \mu ) \in V _ { Y }$ and $\mathcal { N } ^ { * } ( \nu ) \in V _ { X }$ are defined by + +$$ +{ \mathcal { N } } ( \mu ) ( y ) = \sum _ { x } p ( y | x ) \mu ( x ) \quad { \mathrm { a n d } } \quad { \mathcal { N } } ^ { * } ( \nu ) ( x ) = \sum _ { y } p ( x | y ) \nu ( y ) . +$$ + +These stochastic maps $\mathcal { N }$ and $\mathcal { N } ^ { * }$ perform inference of one variable given some (possibly imperfect) knowledge about the other, with priors given by the marginals $p ( x )$ or $p ( y )$ of $p ( x , y )$ depending on the direction of the inference. Importantly, $\mathcal { N } ^ { * }$ is the transpose of $\mathcal { N }$ in terms of the inner products defined above: + +$$ +\langle \nu , \mathcal { N } ( \mu ) \rangle _ { Y } = \langle \mathcal { N } ^ { * } ( \nu ) , \mu \rangle _ { X } . +$$ + +We now have the tools to address the problem mentioned in the introduction. The distinguishability between $q \in V _ { X }$ and $p _ { X }$ after the action of the channel $\mathcal { N }$ is $\chi ^ { 2 } ( \mathcal { N } ( q ) , p _ { Y } )$ since $p _ { Y } = \mathcal { N } ( p _ { X } )$ . Hence we want to find the distributions $p$ which maximize the relevance Beny & Osborne (2013). ´ + +$$ +\eta ( q ) = \frac { \chi ^ { 2 } ( \mathcal { N } ( q ) , p _ { Y } ) } { \chi ^ { 2 } ( q , p _ { X } ) } = \frac { \langle \mathcal { N } ( \mu ) , \mathcal { N } ( \mu ) \rangle _ { Y } } { \langle \mu , \mu \rangle _ { X } } , +$$ + +where $\mu = q - p _ { X }$ . The inner-product formulation makes it clear that this amounts to finding the eigenvector with largest eigenvalue for the symmetric map $\mathcal { N } ^ { * } \mathcal { N }$ , which is also the singular vector with largest singular value for $\mathcal { N }$ . On can then go on to find the eigenvector with next largest eigenvalue and so on, which are automatically orthogonal. + +In practice, the inner products are more tractable to compute if we express elements $\mu \in V _ { X }$ and $\nu \in V _ { Y }$ in terms of variables $f$ and $g$ as $\mu = p _ { X } f$ and $\nu = p _ { Y } g$ , or + +$$ +\mu ( x ) = p ( x ) f ( x ) \quad { \mathrm { a n d } } \quad \nu ( y ) = p ( y ) g ( y ) +$$ + +for all $x , y$ . Indeed, this yields simply + +$$ +\langle \mu , \mu ^ { \prime } \rangle _ { X } = \operatorname { \mathbb { E } } ( f f ^ { \prime } ) \quad { \mathrm { a n d } } \quad \langle \nu , \nu ^ { \prime } \rangle _ { Y } = \operatorname { \mathbb { E } } ( g g ^ { \prime } ) . +$$ + +We are now in measure to make the connection with DCCA Andrew et al. (2013). Indeed, the aims of DCCA is to maximize the correlations $\operatorname { c o r r } ( f , g ) = \mathbb { E } ( f g )$ over function $f ( x )$ and $g ( y )$ such that $\mathbb { E } ( f ^ { 2 } ) = \mathbb { E } ( g ^ { 2 } ) = 1$ . But, using $\mu = p _ { X } f$ and $\nu = p _ { Y } g$ , we have + +$$ +\mathbb { E } ( f g ) = \langle \nu , \mathcal { N } ( \mu ) \rangle _ { Y } , +$$ + +which is maximized by the left- and right- singular vectors $\mu$ and $\nu$ of $\mathcal { N }$ with largest singular value. + +Given all the singular vectors $\mu _ { i } = p _ { X } f _ { i }$ and $\nu _ { i } = p _ { Y } g _ { i }$ with singular values $\eta _ { i }$ , we obtain the representation + +$$ +\mathcal { N } ( \mu ) = \sum _ { i } \eta _ { i } \nu _ { i } \langle \mu _ { i } , \mu \rangle _ { X } , +$$ + +which, using a more standard notation and the Kronecker delta $\delta _ { x }$ , yields Eq. 1: + +$$ +p ( y | x ) = { \mathcal { N } } ( \delta _ { x } ) ( y ) = { \frac { 1 } { p ( x ) } } \sum _ { i } \eta _ { i } \nu _ { i } ( y ) \mu _ { i } ( x ) = p ( y ) \sum _ { i } \eta _ { i } v _ { i } ( y ) u _ { i } ( x ) , +$$ + +where $\mu _ { i } = p _ { X } u _ { i }$ and $\nu _ { i } = p _ { Y } v _ { i }$ . + +For the purpose of the optimization and inference, we do not need to full diagonal decomposition, but just functions $f _ { i } = \mu _ { i } / p _ { X }$ and $g _ { j } = \nu _ { j } / p _ { Y } , i , j = 1 , \ldots , k _ { 0 }$ , which have the same span as the canonical variables $u _ { i }$ and $v _ { j }$ respectively for $i , j = 1 , \ldots , k _ { 0 }$ (assuming that $\eta _ { 1 } , \ldots , \eta _ { k _ { 0 } }$ are the largest singular vector). Below we refer to $f _ { i }$ and $g _ { j }$ as the $k _ { 0 }$ most relevant variables. + +Because these functions may not be orthogonal, we need the covariance matrices + +$$ +K _ { i j } = \langle \mu _ { i } , \mu _ { j } \rangle _ { X } = \mathbb { E } ( f _ { i } f _ { j } ) , \quad L _ { i j } = \langle \nu _ { i } , \nu _ { j } \rangle _ { X } = \mathbb { E } ( g _ { i } g _ { j } ) , \quad A _ { i j } = \langle \nu _ { i } , N ( \mu _ { j } ) \rangle _ { Y } = \mathbb { E } ( g _ { i } f _ { j } ) . +$$ + +If $N _ { i j }$ denote the components of $\mathcal { N }$ in the sense that $\begin{array} { r } { \mathcal { N } ( p _ { X } f _ { j } ) = \sum _ { i } N _ { i j } p _ { Y } g _ { i } } \end{array}$ , then, using our inner products to isolate $N _ { i j }$ , we obtain $N = L ^ { - 1 } A$ . Similarly, the components of $\mathcal { N } ^ { * }$ are $N _ { i j } ^ { * } =$ $K ^ { - 1 } A ^ { \top }$ . This implies that the sum of the square of the singular values of $\mathcal { N }$ restricted to the spans of the vectors $p _ { X } f _ { i }$ and $p _ { Y } g _ { j }$ for all $i , j$ , which is what we want to maximize, is just given by + +$$ +\sum _ { i = 1 } ^ { k _ { 0 } } \eta _ { i } ^ { 2 } = \mathrm { T r } \left( N ^ { \ast } N \right) = \mathrm { T r } ( K ^ { - 1 } A ^ { \top } L ^ { - 1 } A ) . +$$ + +This is the DCCA objective. Below we use the objective function $C = k _ { 0 } - \mathrm { T r } \left( N ^ { \ast } N \right)$ , for the cosmetic reason that its optimal value is zero. + +Moreover, the corresponding truncated representation of the conditional distribution is + +$$ +p ( \boldsymbol { y } | \boldsymbol { x } ) = \mathcal { N } ( \delta _ { \boldsymbol { x } } ) ( \boldsymbol { y } ) \simeq p ( \boldsymbol { y } ) \sum _ { i , j = 1 } ^ { k _ { 0 } } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } g _ { i } ( \boldsymbol { y } ) f _ { j } ( \boldsymbol { x } ) , +$$ + +where we used the fact that the components of $\delta _ { x }$ are $\begin{array} { r } { \delta _ { j } = \sum _ { i } K _ { j i } ^ { - 1 } f _ { i } ( x ) } \end{array}$ . + +Of course, This approach can produce a faithful representation of the correlations only if $\mathcal { N }$ is actually close to being of rank $k _ { 0 }$ (see Appendix $\mathbf { B }$ for a more precise statement). If we interpret the relevant subspace as a space of probability over latent variable, this means that our latent variables have at most $k _ { 0 }$ discrete states. + +However, even if the rank $k _ { 0 }$ corner of $\mathcal { N }$ is a not a good approximation, this strategy allows us to nevertheless do the correct inference on certain random variables, namely those which are in the span of the canonical variables! + +Indeed, the exact conditional expectation of $g _ { k }$ is (assuming $D$ is the actual rank of $\mathcal { N }$ ), + +$$ +\begin{array} { l } { { \displaystyle \sum _ { y } g _ { k } ( y ) p ( y | x ) = \sum _ { i , j = 1 } ^ { D } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } \mathbb { E } ( g _ { k } g _ { i } ) f _ { j } ( x ) } } \\ { { \displaystyle \qquad = \sum _ { j = 1 } ^ { D } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) = \sum _ { j = 1 } ^ { k _ { 0 } } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) } , } \end{array} +$$ + +where the last truncation is exact if $k \leq k _ { 0 }$ due to the assumption that the basis $f _ { i }$ and $g _ { j }$ have the same span as the $k _ { 0 }$ largest right and left singular vectors of $\mathcal { N }$ respectively. + +For instance, if $p ( x , y )$ is Gaussian, the canonical variables can be computed analytically, as in Lancaster (1958) or Beny (2018a) in the multivariate case. Solutions for other distributions were ´ also computed in Eagleson (1964). + +Notably, for any two dimensional Gaussian, the space of $k$ most relevant variables is simply spanned by the moments $f _ { n } ( x ) = x ^ { n }$ and $g _ { n } ( y ) = y ^ { n }$ for $n = 0 , \ldots , k - 1$ . Hence, in this case the first $k$ moments can be inferred exactly using only the $k + 1$ most relevant variables (See Appendix C). + +# 4 ALGORITHM + +Let us explicit the algorithm resulting from the above analysis. + +We assume that we are given independent samples $( x _ { 1 } , y _ { 1 } ) , ( x _ { 1 } , y _ { 2 } ) , . . .$ from the otherwise unknown joint distribution $p ( x , y )$ . + +We first perform DCCA (Andrew et al., 2013). That is, we need two independent deterministic feedforward neural networks. The first maps $x$ to a set of $k _ { 0 }$ real-valued variables $f _ { 1 } ( x ) , \ldots , f _ { k _ { 0 } } ( x )$ . The second maps $y$ to a different set of $k _ { 0 }$ variables $g _ { 1 } ( y ) , \ldots , g _ { k _ { 0 } } ( y )$ . + +The parameters of the neural networks are to be set to minimize the objective function + +$$ +C = k _ { 0 } - \operatorname { T r } { \left( K ^ { - 1 } A ^ { \top } L ^ { - 1 } A \right) } , +$$ + +where the matrices $K , L , A$ can be approximated over a mini-batch $( x _ { n } , y _ { n } )$ , $n = 1 , \ldots , N$ via + +$$ +K _ { i j } = { \frac { 1 } { N } } \sum _ { n = 1 } ^ { N } f _ { i } ( x _ { n } ) f _ { j } ( x _ { n } ) , \quad L _ { i j } = { \frac { 1 } { N } } \sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) g _ { j } ( y _ { n } ) , \quad A _ { i j } = { \frac { 1 } { N } } \sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) f _ { j } ( x _ { n } ) . +$$ + +We found that, provided the batch size is sufficiently large compared to $k _ { 0 }$ (about 10 times in our experience), this can be minimized using ADAM or direct gradient descent. However, to guarantee stability when using large $k _ { 0 }$ , we needed to explicit the gradient of the objective function in order to force the use of the Moore-Penrose pseudo-inverses for $K ^ { - 1 }$ and $L ^ { - 1 }$ in both the forward and backward passes, in addition to using 64 bits floats in these computations. + +Once the relevant variables have been learned, we still need to use the training data in a second step. Indeed, suppose that we wish to use our model to infer the value of some function $\Theta ( x )$ , i.e., to compute its approximate expectation value in terms of the conditional distribution $x \mapsto p ( x | y )$ . Then we need to store, for each variable $j = 1 , \ldots , k _ { 0 }$ , the quantities + +$$ +\Theta _ { j } = \frac { 1 } { N _ { \mathrm { f u l l } } } \sum _ { n = 1 } ^ { N _ { \mathrm { f u l l } } } \Theta ( x _ { n } ) f _ { j } ( x _ { n } ) , +$$ + +where the average is to be taken on the full training batch (of size $N _ { \mathrm { f u l l } } ,$ . The same can be done exchanging $x$ with $y$ and $f _ { j }$ with $g _ { j }$ for the reverse inference. + +For instance, if a data point $x$ is composed of real components $x ^ { a }$ —such as pixel color components for an image—and we are interested in the estimator which minimize the expected $l ^ { 2 }$ distance to the predicted values of these components, then we need the expectation values of the components $\Theta ( x ) = x ^ { a }$ for all $a$ , and possibly higher moments to gain more knowledge about the shape of the posterior distribution, such as the second moments $\Theta ^ { \prime } ( \bar { x } ) = x ^ { 2 }$ , etc. + +Inference can then be performed with new data using + +$$ +\overline { { { \Theta } } } = \sum _ { x } p ( x | y ) \Theta ( x ) \approx \sum _ { i , j = 1 } ^ { k _ { 0 } } ( K ^ { - 1 } A ^ { \top } L ^ { - 1 } ) _ { j i } \Theta _ { j } g _ { i } ( y ) . +$$ + +Moreover, the accuracy of thof the relevant variables, i.e., t depend on, for which $k _ { 0 }$ $\Theta$ taken in the span. $\begin{array} { r } { \Theta ( x ) = \sum _ { i = 1 } ^ { k } c _ { i } f _ { i } ( x ) } \end{array}$ $\begin{array} { r } { \Theta _ { j } = \sum _ { i = 1 } ^ { k _ { 0 } } c _ { i } K _ { i j } } \end{array}$ + +The reverse inference formulas are obtained simply by the exchanges $K L$ , $A A ^ { \top }$ , and $g _ { i } f _ { i }$ . + +# 5 EXPERIMENTS + +In all our experiments, we used the ADAM optimizer with learning rate 0.001. We used the Flux package (Innes, 2018) for Julia, as well as Tensorflow. + +As usual the data is divided into a training set and a testing set. No aspect of the testing set is used during training. The loss function refers to Eq. (7). In order to monitor overfitting, we compute a “test loss” and a “training loss”. The test loss is computed from the trained variables using only the test data, and accordingly, the training loss is computed purely using the training data. + +Moreover, when performing inference on test data using Eq. (10), we use the covariances $A , L , K$ and expectations $\Theta _ { j }$ (Equ. (9)) built from the training data only. + +# 5.1 INFERENCE ON OCCLUDED MNIST + +In this experiment, we use the left and right halves of the MNIST digit images as correlated variables $X$ and $Y$ . The goal is to obtain the expected left halves given the right halves, or vice versa. + +The training set was augmented by random small rotations and displacements to make the task more ambiguous, as we want to explore the uncertainty in the prediction. + +The relevant variables were represented by two convolutional neural networks of identical architecture. They are composed of four convolutional layers and one fully connected layer, an architecture that performs well for supervised learning on this dataset. For ease of implementation, these CNN have the whole image as input, but with either half zeroed (same value as black pixels). Half-width CNNs with proper padding at the cut perform similarly. + +After training, we used the training dataset to also compute the expected pixel gray value as well as their covariance for each relevant variable using Eq. (9). + +These were used into Eq. (10) to compute the mean pixel gray values and their covariances over the conditional probability of $X$ given $Y$ on test data. This mean is the Bayesian estimator for the $l ^ { 2 }$ distance between half images, i.e., it should minimize the expected distance $d _ { l ^ { 2 } }$ over the conditional distribution, where $\begin{array} { r } { d _ { l ^ { 2 } } ^ { 2 } ( x , \overline { { y } } ) = \sum _ { i } ( x _ { i } - y _ { i } ) ^ { 2 } } \end{array}$ , where $x _ { i } \in [ 0 , 1 ]$ is the value of the $\mathrm { i } ^ { t h }$ pixel. (This is equivalent to minimizing the mean square error). + +The results on a randomly selected subset of test digits is shown in Fig. 1. For each example, we also computed the images obtained by adding plus or minus one standard deviation along the direction of greatest variance in the space of relevant variables. This reveals the main ambiguities (such as between 8 and 3 or 7 and 9 which share a similar right half). + +The graph of the singular values shows that the rank cutoff of 200 is too low to capture all of the relevant variables (the sudden drop at the end is not robust to an increase in the cutoff), but the results are reasonable nevertheless. This shows that our representation of the conditional distributions contains valuable information besides the simple mean. + +# 5.2 SUPERVISED LEARNING + +In the context of a supervised classification task, one of the dataset (the labels) is of sufficiently low dimensionality that we can use a complete basis over its probability space as our relevant variables, such as the standard one-hot encoding of labels. This serves as a good first sanity test for our approach. Surprisingly, we find that it converges faster than standard approaches, and without the need for regularization. + +Let the variable $Y$ stands for the labels, with values in $\{ 1 , \ldots , k \}$ . The probability space consists of vectors with $k$ real components. The canonical basis corresponds to the one-hot encoding $g _ { i } ( j ) =$ $\delta _ { i j }$ (Kronecker delta). All we need is a neural network to encode $k$ variables $f _ { 1 } , \ldots , f _ { k }$ on $X$ . After learning the most relevant variables $f _ { i }$ , we apply the reverse of Eq. (10) for function $\Theta ( y ) = y$ , and use the maximum component of expected value $\overline { y }$ to infer the labels from the data. + +Let us refer to this procedure as DCCI (Deep Canonical Correlations based Inference). + +We tested this approach on the MNIST and CIFAR10 datasets, and compared the results to the standard cross-entropy objective (Fig. 2). + +We plotted the accuracy as function of the epoch rather than clock time which would depend on many factors. But the time per epoch is roughly the same for each approaches in the above experiments. Indeed, the training time is dominated by the forward and backward evaluations of the neural networks which are identical. (However, the time it takes to evaluate our objective can become significant for much larger number of labels $k$ , since it involves the inversion of matrices of dimension $k$ . This is in addition to the fact that a greater dimension would require also larger batches.) + +We found that, without regularization, simply changing the objective from cross-entropy to DCCI provided a large improvement both of convergence speed and final accuracy for both models. + +![](images/201f3ce00c47dbafcb9ef33dbcddfe321ec4b0f6d68f48039079f27a17681d75.jpg) +Figure 1: Left mosaic: the left halves of the MNIST digits in a random sample from the test set are inferred from the right half, with cutoff $k _ { 0 } = 2 0 0$ . Three images are shown for each digit. Within each triplet, the middle image represents the mean over pixel intensity of the inferred condition distribution, while left and right images corresponds to a plus and minus one standard deviation from the mean in the direction of largest covariance (in the space of half-images). A particularly interesting example is highlighted. Top-right: loss per epoch for $k _ { 0 } ~ = ~ 2 0 0$ . Bottom-right: the singular values for different values of the cutoff $k _ { 0 }$ , after 150 epochs in each case. + +![](images/f1a73e5112ee1b5d1d1b809be05b2c39b82dfe510aa9cbf963a2647ac4f9eb88.jpg) +Figure 2: Loss and inaccuracy (error rate) on test sets for two classification tasks. The models were trained either using the cross-entropy (CE) or our approach (DCCI), with or without regularization layers. On the MNIST dataset, we used an “all CNN” network, and for the CIFAR10 dataset we used a short VGG variation with 10 convolutions and 3 fully connected layers. In the regularized form, post-activation Batchnorm layers were placed after each convolutional layers on the VGG network. What is shown is the mean over 10 independent runs for MNIST and 5 runs for CIFAR10. The shaded area spans the standard deviation. ADAM with default parameters was used in all cases. No data augmentation was used except for horizontal flips for CIFAR10 (resulting in epochs of 100,000 images). + +![](images/03b9882c59b4695eeffc013e7377800c0cf409c81087a146cf7fe4fc6b062c90.jpg) +Figure 3: Top-left: First 20 relevant variables (on $X$ ) determined by DCCA for a system where $X$ consists of two coordinates uniformly sampled over a circle and a surrounding ring, and $Y$ consists of the same points but shifted by a small normally distributed vector. The variables are arranged from left-to-right and top-to-bottom in order of decreasing relevance. Top-right: the same variables multiplied by the marginal $p _ { X }$ . Bottom row: introducing a gap in the ring allows for a monotonous function of the angle to serve as second most relevant variable (instead of the sine/cosine couple). Hence the angle is automatically “disentangled” from the other variables. (Mid-gray represents the value 0). + +On MNIST, DCCI alone also outperformed cross-entropy with dropout. (Dropout did not yield any improvement in conjunction with DCCI). However, adding batch-normalization layers on the CIFAR example, erased any distinction between DCCI and cross-entropy. + +# 5.3 STRUCTURE OF THE RELEVANT VARIABLES + +We mentioned in Section 3 that if $p ( x , y )$ is a two-dimensional Gaussian distribution with zero mean, then the $n$ most relevant variables of $X$ are the first $n$ powers of $X$ itself, independently of the covariance matrix. This implies that the canonical variables are the Hermite polynomials in $X$ (which results from applying the Gram-Schmidt procedure to the basis $\{ 1 , x , x ^ { 2 } , \bar { . . . } \bar \} \}$ . + +A similar property holds for multivariate Gaussians, namely, the less relevant singular values are polynomials in the more relevant ones. If this is true more generally, it should be possible to further compress and organize the latent space extracted with DCCA by finding a minimal set of generators, which ought to also be in the span of the most relevant variables. + +We applied DCCA to a synthetic dataset to explore this idea, shown in Fig. 3. In this case, we actually performed a final SVD to obtain the unique uncorrelated canonical variables, and ordered them by decreasing relevance (their respective singular values). + +Here, $X$ consists of two real numbers, distributed uniformly within a ring and a disk. The variable $Y$ is obtained by adding a random Gaussian shift to $X$ with a small standard deviation. The more relevant variables ought to be those which are more robust to such small random displacement. This formalizes the idea that we are interested in extracting “large-scale” variables Beny (2018b). ´ + +We would expect the relevant independent variables to be: the binary variable indicating whether the point is in the disk or the ring and the angle around the ring, followed by the radial component in the ring, and finally the Cartesian coordinates inside the disk. This is precisely what we see in Fig. 3. + +Indeed—if we put aside for now the fact that the angle itself is not directly represented—besides the constant function, the two most relevant variables are the sine and cosine of the angle, followed by the binary variable separating the disk from the ring. + +But these variables ought to span the space of probabilities over the relevant variables, not just the variables themselves. Hence the next six variables are sines and cosines of smaller wavelength, which can encode probability distributions which are increasingly more precisely localized, down to a precision (wavelength) comparable with the diameter of the inner disk. Accordingly, the next two most relevant variables are the Cartesian coordinates inside the disk. This is followed by additional moments of the angle, down to a wavelength equal to the ring’s thickness, at which point we see the radius in the ring appear. + +![](images/139ed066324617d38572ae37b20db57c2963cbc5f2e6fcfe36fcf651198597fd.jpg) +Figure 4: Left: Best mean squared error for images reconstructed from the $k$ most relevant variables, as a function of $k$ (the latent dimension). This logarithmic plot shows that improvements stop once the dimension reaches 19 (where the two lines cross). Right: images produced by the generator from latent variables sampled according to the best Gaussian fit in latent space, for feature subspaces of dimensions 2, 8 and 19. + +As mentioned, we see that the angle itself is not represented, likely because it is discontinuous. However, as shown also in Fig. 3, creating a gap in the ring allows for the angle to emerge as most relevant variable. This suggests that this approach may be able to automatically learn intrinsic coordinates of the latent variable manifold. + +# 5.4 INDEPENDENT VARIABLES AND GENERATIVE MODEL + +If we postulate that the independent (or disentangled) relevant latent variables can be found in the linear span of the relevant variables, we can attempt to extract them by optimizing a neural network composed of two parts. Firstly, a linear layer maps the relevant variables to a small number of outputs (equal to the latent dimension). The purpose of this linear layer is to find the independent variables. These latent variables are then processed by an arbitrarily complex generative network to produce a possible value of the variable $X$ . As objective function, we may us an appropriate measure of similarity between the output and the data element from which the variables were obtained. + +We tested this idea as follows. We took $X$ to consist of the MNIST digits, and produced $Y$ by randomly permuting neighboring pixels in the image, until the mean displacement per pixel is of order 1. In addition, we added independent Gaussian noise to the pixel values. (Hence the noise map $\mathcal { N }$ simulates the coarse-graining channel introduced in Beny & Osborne (2015a)). ´ + +As in the previous experiment, we do so to implement our intuition that the more relevant variables ought to be the ones which are of larger scale, or more robust to local perturbations. + +The relevant variables of the clean images were produced by the same convolutional neural network as in Section 5.2, while the variables of the coarse-grained images were extracted by a network of the same geometry, but with half the number of filters and neurons. + +We extracted the 1000 most relevant out of 1200 learned variables in this way. (The least relevant variables in this system happen to be highly dependent on the total number of variables and hence cannot be trusted to be correct). As a second step, we trained a linear layer coupled to a network composed of 5 fully-connected layers of 800 hidden neurons each. We refer to the number of output neurons in the first linear layer as the latent dimension. + +As input, this network received the variables extracted from MNIST images using the above convolutional neural net (after it was fully trained using DCCA), and was trained to minimize the mean square error between its output and the original MNIST digit. + +The resulting best mean square errors are shown in Fig. 4, as function of the latent dimension. Here we see a distinct change of polynomial scaling law at dimension 19. Increasing the dimension further provides no improvement. This behaviour is compatible with our hypothesis that the extra variables are just functions of those first twenty variables (functions which are effectively re-implemented by the generative network). + +Images generated by sampling from a Gaussian approximation of the latent distribution for different latent dimensions are shown in Fig. 4. Below dimension 20, most generated image can be recognized as a specific digit. + +# 6 OUTLOOK + +We studied the classical (non-quantum) form of the theory introduced in Beny & Osborne (2013), ´ and found that the relevant observables of that theory are just the most correlated canonical variables in the sense of DCCA Andrew et al. (2013), and can be learned effectively using standard machine learning methods. + +This point of views on DCCA provided us with several new insights. The first is that the learned relevant variables provide a useful representation of a joint probability distribution. We showed that performing inference using this representation can outperform crossentropy in predicting classes. Our experiments on halves of MNIST also show that the conditional distribution we obtain can effectively represent the uncertainty in the prediction of high-dimensional data. + +A second insight relates to the interpretation of the canonical variables as spanning directions in the space of probability distributions. As suggested by the gaussian solutions and our experiment on synthetic data, we postulate that the canonical variables are functions of a small number of independent generators contained in their span. This hypothesis is supported by our experiment on MNIST, but further work is required to find a way to cleanly extract these variables. + +We have yet to explore the potential applications of one of the salient aspect of this approach to inference, the fact that the canonical variables learned using DCCA are also those which can be most reliably predicted, irrespective of the value of the cutoff. To see why this is potentially significant, we observe that a central feature of scientific exploration is that we are not so concerned with making predictions about some given variables, as much as we are with discovering variables which can be predicted. + +Another important feature of this approach is the fact that the resulting model allows for the direct evaluation of the expectation values in the posterior distribution without sampling. In particular this allows for the evaluation of credible intervals. Hence it should be especially suited to scientific applications where the ability to quantify uncertainty is essential. + +Finally, the relationship that we established with theory of quantum origin points towards a potential quantum generalization of DCCA that would apply to quantum data, or classical measurements of quantum systems. + +# ACKNOWLEDGMENTS + +We would like to thank Joel B ¨ eny and Raban Iten for helpful suggestions. We are also in- ´ debted to an anonymous ICLR2020 referee for pointing out the connection between our approach and DCCA. This work was supported by the National Research Foundation of Korea (NRF2018R1D1A1A02048436). + +# REFERENCES + +Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analysis. In International conference on machine learning, pp. 1247–1255, 2013. + +C. Beny and T. J. Osborne. Information geometric approach to the renormalisation group. ´ Phys. Rev. A, 92:022330, 2015a. doi: 10.1103/PhysRevA.92.022330. (arXiv:1206.7004). +Cedric B ´ eny. Coarse-grained distinguishability of field interactions. ´ Quantum, 2:67, 2018a. (arXiv:1509.03249). +Cedric B ´ eny. Inferring relevant features: from qft to pca. ´ International Journal of Quantum Information, 16:1840012, 2018b. (arXiv:1802.05756). +Cedric B ´ eny and Tobias J Osborne. Renormalisation as an inference problem. ´ (arXiv:1310.3188), 2013. +Cedric B ´ eny and Tobias J Osborne. The renormalisation group via statistical inference. ´ New J. Phys., 17:083005, 2015b. doi: 10.1088/1367-2630/17/8/083005. (arXiv:1402.4949). +GK Eagleson. Polynomial expansions of bivariate distributions. The Annals of Mathematical Statistics, 35(3):1208–1215, 1964. +Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211–218, 1936. +Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014. +Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. arXiv preprint arXiv:1310.8499, 2013. +Harold Hotelling. Relations between two sets of variates. Biometrika, 28(3/4):321–377, 1936. +Mike Innes. Flux: Elegant machine learning with julia. Journal of Open Source Software, 2018. doi: 10.21105/joss.00602. +Raban Iten, Tony Metger, Henrik Wilming, L´ıdia Del Rio, and Renato Renner. Discovering physical concepts with neural networks. (arXiv:1807.10300), 2018. +Diederik P Kingma and Max Welling. Auto-encoding variational bayes. (arXiv:1312.6114), 2013. +HO Lancaster. The structure of bivariate distributions. The Annals of Mathematical Statistics, 29 (3):719–736, 1958. +Tomer Michaeli, Weiran Wang, and Karen Livescu. Nonparametric canonical correlation analysis. In International Conference on Machine Learning, pp. 1967–1976, 2016. +Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. +M. Ohya and D. Petz. Quantum entropy and its use. Springer Verlag, 2004. +Amichai Painsky, Meir Feder, and Naftali Tishby. An information-theoretic framework for nonlinear canonical correlation analysis. arXiv preprint arXiv:1810.13259, 2018. +Weiran Wang, Xinchen Yan, Honglak Lee, and Karen Livescu. Deep variational canonical correlation analysis. arXiv preprint arXiv:1610.03454, 2016. +Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126, 2016. + +# A EXTRA INFORMATION ABOUT THE ALGORITHM + +# A.1 ALTERNATIVE INTERPRETATION OF THE OBJECTIVE + +If we write $F _ { i j } : = f _ { j } ( x _ { i } )$ and $G _ { i j } : = g _ { j } ( y _ { i } )$ for the value of our variables on the dataset, then $\begin{array} { r } { K = \frac { 1 } { N } F ^ { \top } F } \end{array}$ , $\begin{array} { r } { L = \frac { 1 } { N } G ^ { \top } G } \end{array}$ and $\begin{array} { r } { A = \frac { 1 } { N } G ^ { \top } F } \end{array}$ . The DCCA objective can then be written as + +$$ +\operatorname { T r } \left( K ^ { - 1 } A ^ { \top } L ^ { - 1 } A \right) = \operatorname { T r } \left( P Q \right) +$$ + +where $P = F ( F ^ { \top } F ) ^ { - 1 } F ^ { \top }$ and $Q = G ( G ^ { \top } G ) ^ { - 1 } G ^ { \top }$ are the projectors on the ranges of $F$ and $G$ respectively. Hence, we are maximizing the overlap between those ranges (which represent possible linear combinations of datapoints, respectively determined from variables of one or the other correlated views.) + +# A.2 HEURISTIC + +Batch size—In our experiments, we observed that the batch size during training needs to be an order of magnitude larger than the number of variables (rank cutoff). When the batch size was too small, learning seemed to converge normally in terms of training and test loss, but resulted in variables which yield dramatically different losses when evaluated on larger batches, and yield spurious predictions. + +Constant variables—The loss function $C$ takes value between 0 and $k _ { 0 } - 1$ because the constant variable always has relevance 1. The constant variable could be enforced a priori rather than learned, which, due to the objective, automatically forces the learned variables to have zero expectation values (be orthogonal to the constant variable). This might have advantages in certain circumstances, but in our experiments we found that this sometime hindered convergence. + +Invertibility issues—The covariance matrices $K$ and $L$ can be ill-conditioned, potentially causing the gradient to “explode” because of the inverses $K ^ { - 1 }$ et $L ^ { - 1 }$ involved in the loss function. This can be avoided either by using the Moore-Penrose pseudo-inverse, or by replacing $K ^ { - 1 }$ by $( K + \epsilon { \bf 1 } ) ^ { - 1 }$ in the loss for some small positive number $\epsilon$ , and likewise for $L ^ { - 1 }$ . + +Symmetries in the loss function—The loss $C$ only depends on the span of the variables $f _ { i }$ and $g _ { j }$ , hence it has a very large group of symmetries. In particular, it is invariant under a change of the norm of each variable independently from each other. Because of that, it is preferable not to have a linear last layer. Using a hyperbolic tangent as last nonlinearity worked in our experiments. + +Regularization—In all our tests, dropout had no beneficial effect. In fact, our objective seems to already provide a form of regularization, as shown in Section 5.2. + +# B THEORY IN MORE DETAILS + +We consider two correlated random variables $X$ and $Y$ with a joint probability distribution $p ( x , y )$ . We assume that we are able to numerically evaluate expectations with respect to this distribution, for instance because we can sample from it. We want to use this ability in order to compute expectations with respect to the conditional distributions $p _ { X | Y } ( x | y ) \ : = \ : \dot { p } ( x , y ) / p _ { X } ( x )$ and $p _ { Y | X } ( y | x ) ~ = ~ p ( x , y ) / p _ { Y } ( y )$ , where $\begin{array} { r } { p _ { X } ( x ) \ = \ \sum _ { y } p ( x , y ) } \end{array}$ and $\begin{array} { r } { p _ { Y } ( y ) \ = \ \sum _ { x } p ( x , y ) } \end{array}$ are the marginals of $p$ . Below we sometime remove the subscripts $X , X | Y$ or $Y | X$ if there is no ambiguity. + +For instance, suppose we generated samples of $y$ given $x$ , through explicit knowledge of $p _ { Y \mid X }$ . Then the evaluation of expectations with respect to $p _ { X | Y }$ is the subject of Bayesian inference. However, this is generally done in a context where the variable $X$ has low dimensionality and parameterizes a hand-crafted model. Our approach, however, is free of such a model and the variable $X$ can be of very high dimensionality. + +# B.1 INNER PRODUCT ON PROBABILITY VECTORS + +In order to define our strategy, we need to equip the spaces of probability distributions for $X$ and $Y$ with an inner product structure. Let us focus on $X$ , and assume that it takes discrete values to avoid unnecessary technicalities. The set of probability vectors is a convex subset of the real linear space $V _ { X } = \mathbb { R } ^ { n }$ . Let us equip this space with the product + +$$ +\langle \mu , \mu ^ { \prime } \rangle _ { X } : = \sum _ { x } { \frac { \mu ( x ) \mu ^ { \prime } ( x ) } { p _ { X } ( x ) } } +$$ + +for any $\mu , \mu ^ { \prime } \in V _ { X }$ . We also write $\| \mu \| _ { X } ^ { 2 } = \langle \mu , \mu \rangle _ { X }$ . Importantly, this depends explicitly on the fixed probability vector $p _ { X } ( x )$ , which we took to be the marginal of $p ( x , y )$ . If $p _ { X }$ has full support, this makes $V _ { X }$ into a real inner product space. The same can be done for the variable $Y$ , yielding the inner product $\langle \nu , \nu ^ { \prime } \rangle _ { Y }$ for $\nu , \nu ^ { \prime } \in V _ { Y }$ . + +Had we interpreted $\mu$ and $\mu ^ { \prime }$ as tangent vectors to $V _ { X }$ , considered as a manifold, this would be the Fisher information (Riemannian) metric, as in Beny & Osborne (2015b). But this quantity is also ´ meaningful for finite vectors: the induced norm distance between $p _ { X }$ and any probability vector $q$ is the $\chi ^ { 2 }$ -divergence: + +$$ +\chi ^ { 2 } ( q , p _ { X } ) = \langle q - p _ { X } , q - p _ { X } \rangle _ { X } . +$$ + +The set of conditional probability distributions $p _ { Y \mid X }$ form a stochastic map, i.e., a linear map $\mathcal { N }$ $V _ { X } \to V _ { Y }$ , $\mu \mapsto { \mathcal { N } } ( \mu )$ , where + +$$ +\mathcal { N } ( \mu ) ( y ) = \sum _ { x } p _ { Y | X } ( y | x ) \mu ( x ) +$$ + +for any $\mu \in V _ { X }$ . + +It is straightforward to check that the stochastic map $\mathcal { N } ^ { * }$ defined by + +$$ +\mathcal { N ^ { * } } ( \nu ) ( x ) = \sum _ { x } p _ { X | Y } ( x | y ) \nu ( x ) +$$ + +is the transpose $\mathcal { N } ^ { * }$ of $\mathcal { N }$ with respect to the inner products we defined (Ohya & Petz, 2004), i.e., for all $\nu \in V _ { Y }$ and $\mu \in V _ { X }$ , + +$$ +\langle \nu , \mathcal { N } ( \mu ) \rangle _ { Y } = \langle \mathcal { N } ^ { * } ( \nu ) , \mu \rangle _ { X } . +$$ + +Also, we observe that $\mathcal { N } ( p _ { X } ) = p _ { Y }$ and $\mathcal { N } ^ { * } ( p _ { Y } ) = p _ { X }$ . + +# B.2 EIGEN-RELEVANCE DECOMPOSITION + +We can use the inner products on $V _ { X }$ and $V _ { Y }$ to define a singular value decomposition of the stochastic map $\mathcal { N }$ . That is, there is an orthonormal family $u _ { 1 } , \ldots , u _ { k }$ of $V _ { X }$ and an orthonormal family $v _ { 1 } , \ldots , v _ { k }$ of $V _ { Y }$ , such that + +$$ +\begin{array} { r } { \mathcal { N } ( u _ { j } ) = \eta _ { j } v _ { j } , } \end{array} +$$ + +for $j = 1 , \dotsc , k$ . For each $j , \eta _ { j }$ is a singular value of $\mathcal { N }$ , whose square we call the relevance of the vector $v _ { j }$ . Moreover $\eta _ { j } \in [ 0 , 1 ]$ since the $\chi ^ { 2 }$ divergence is contractive under any stochastic map. Given that ${ \ddot { \mathcal { N } } } ^ { * }$ is the transpose of $\mathcal { N }$ : + +$$ +\begin{array} { r } { \mathcal { N } ^ { * } ( v _ { j } ) = \eta _ { j } u _ { j } . } \end{array} +$$ + +Equivalently, $u _ { j }$ is an eigenvector of $\mathcal { N } ^ { \ast } \circ \mathcal { N }$ and $v _ { j }$ is an eigenvectors of $\mathcal { N } \circ \mathcal { N } ^ { \ast }$ , both with eigenvalue $\eta _ { j } ^ { 2 }$ . + +Because $\mathcal { N }$ maps $p _ { X }$ to $p _ { Y }$ , we always have the dual eigenvectors $u _ { 0 } ~ = ~ p _ { X }$ and $v _ { 0 } = p _ { Y }$ with eigenvalue 1. + +# B.3 LOW-RANK APPROXIMATION + +Typically, the dimension $k$ of the space of probabilities is more than astronomically large. For instance, if the values of $X$ consists of small 256 gray level images of $2 8 \times 2 8$ pixels, then $k =$ + +$2 5 6 ^ { 2 8 ^ { 2 } } \simeq 1 0 ^ { 1 8 8 8 }$ . However, in many case, only very few of these dimensions may be relevant for the purpose of inferring other variables. + +The core of our approach is to approximate $\mathcal { N }$ and $\mathcal { N } ^ { * }$ by restricting them to the span of the first $k _ { 0 }$ eigenvectors $u _ { j }$ and $v _ { j }$ with largest singular values $\eta _ { j }$ . That is, if we order the singular values $\eta _ { j }$ , $j = 1 , \dots , k$ in decreasing order, we propose to use the approximations + +$$ +\begin{array} { l } { { \displaystyle \mathcal { N } _ { 0 } ( \mu ) = \sum _ { j \leq k _ { 0 } } \eta _ { j } \langle u _ { j } , \mu \rangle _ { X } v _ { j } } } \\ { { \displaystyle \mathcal { N } _ { 0 } ^ { * } ( \nu ) = \sum _ { j \leq k _ { 0 } } \eta _ { j } \langle v _ { j } , \nu \rangle _ { Y } u _ { j } } } \end{array} +$$ + +to $\mathcal { N }$ and $\mathcal { N } ^ { * }$ respectively, for some $k _ { 0 }$ typically much smaller than $k$ , and any $\mu \in V _ { X }$ , $\nu \in V _ { Y }$ . + +We denote the components of $\mathcal { N } _ { 0 }$ and $\mathcal { N } _ { 0 } ^ { \ast }$ by $q ( y | x )$ and $q ( x | y )$ , e.g., + +$$ +{ \mathcal { N } } _ { 0 } ( \mu ) ( y ) = \sum _ { x } q ( y | x ) \mu ( x ) . +$$ + +Since $\mathcal { N } _ { 0 }$ and $\mathcal { N } _ { 0 } ^ { \ast }$ are adjoint, we can define $q ( x , y ) = q ( x | y ) p _ { Y } ( y ) = q ( y | x ) p _ { X } ( x )$ . Although the marginals of $\dot { \boldsymbol { q } } ( \boldsymbol { x } , \boldsymbol { y } )$ are the probability distributions $p _ { X }$ and $p _ { Y }$ , the numbers $q ( x , y )$ are not necessarily positive. + +The quality of this approximation for a given $k _ { 0 }$ does not directly depend on the dimensionality of $X$ and $Y$ , but only on the amount of correlations between the two variables. Our aim is to use a $k _ { 0 }$ small enough that the components of $\mathcal { N } _ { 0 }$ and $\mathcal { N } _ { 0 } ^ { \ast }$ can be computed explicitly. + +Theorem 1. $\mathcal { N } _ { 0 }$ is the map of rank $k _ { 0 }$ which minimizes the average distance + +$$ +\sum _ { x } p ( x ) \| \mathcal { N } _ { 0 } ( \delta _ { x } ) - \mathcal { N } ( \delta _ { x } ) \| _ { Y } ^ { 2 } = \sum _ { x y } \frac { ( q ( x , y ) - p ( x , y ) ) ^ { 2 } } { p ( x ) p ( y ) } . +$$ + +Proof. The low rank approximation $\mathcal { N } _ { 0 }$ minimizes the distance $\| \mathcal { N } _ { 0 } - \mathcal { N } \| _ { \mathrm { F } }$ where + +$$ +\| \mathcal { M } \| _ { F } ^ { 2 } = \mathrm { T r } \left( \mathcal { M } ^ { * } \mathcal { M } \right) +$$ + +is the Hilbert-Schmidt (or Frobenius) norm (Eckart & Young, 1936). This follows from the fact that this is also the $l ^ { 2 }$ -norm of the vector of singular values of $\mathcal { M }$ . Let us find the explicit form of the trace. Each possible value $x$ of the variable $X$ is associated with a probability distribution $\delta _ { x } ( y ) = 1$ when $x = y$ and zero otherwise. These distributions form an orthogonal basis of $V _ { X }$ , and have norms $\langle \delta _ { x } , \delta _ { x } \rangle = 1 / p _ { X } ( x )$ . Therefore, + +$$ +\begin{array} { l } { \displaystyle \operatorname { T r } \left( \mathcal { M ^ { * } M } \right) = \sum _ { x } p _ { X } ( x ) \langle \delta _ { x } , \mathcal { M ^ { * } M } ( \delta _ { x } ) \rangle _ { Y } } \\ { \displaystyle \quad = \sum _ { x } p _ { X } ( x ) \| \mathcal { M } ( \delta _ { x } ) \| _ { Y } ^ { 2 } } \end{array} +$$ + +# B.4 RELEVANT VARIABLES + +We express the elements $\mu \in \ V _ { X }$ and $\nu \in \ V _ { Y }$ in terms of the marginals $p _ { X }$ and $p _ { Y }$ as simple products: + +$$ +\mu ( x ) = p _ { X } ( x ) f ( x ) \quad { \mathrm { a n d } } \quad \nu ( y ) = p _ { Y } ( y ) g ( y ) +$$ + +for all $x , y$ , where $f$ and $g$ are real functions of $x$ and $y$ . + +The inner products then simply become correlations among variables. Using also $\mu ^ { \prime } = p _ { X } f ^ { \prime }$ and $\nu ^ { \prime } = p _ { Y } g ^ { \prime }$ , we obtain + +$$ +\begin{array} { l } { { \langle \mu , \mu ^ { \prime } \rangle _ { X } = \displaystyle \sum _ { x } p _ { X } ( x ) f ( x ) f ^ { \prime } ( x ) = \overline { { { f f ^ { \prime } } } } , } } \\ { { \langle \nu , \nu ^ { \prime } \rangle _ { Y } = \displaystyle \sum _ { y } p _ { Y } ( y ) g ( y ) g ^ { \prime } ( y ) = \overline { { { g g ^ { \prime } } } } . } } \end{array} +$$ + +These are simple expectation values with respect to $p$ , which we assumed is the type of quantity we can evaluate for arbitrary functions $f , f ^ { \prime } , g , \bar { g ^ { \prime } }$ . + +Since $\mathcal { N } ^ { * } \mathcal { N }$ is self-adjoint in terms of this inner product, its eigenvectors $u _ { i }$ are orthogonal, and hence the corresponding variables $a _ { i }$ defined by ${ u } _ { i } \bar { ( x ) } = p _ { X } ( x ) \bar { a _ { i } } ( x )$ are uncorrelated. Indeed, + +$$ +\overline { { a _ { i } a _ { j } } } = \langle u _ { i } , u _ { j } \rangle _ { X } = 0 , +$$ + +for all $i , j$ . Moreover, accounting for the eigenvector $u _ { 0 } = p _ { X }$ (corresponding to the constant feature $a _ { 0 } ( x ) = 1$ for all $x$ ), + +$$ +\overline { { a } } _ { i } = 0 +$$ + +for all $i \neq 0$ . Hence we trivially have + +$$ +\overline { { a _ { i } a _ { j } } } = \overline { { a } } _ { i } \overline { { a } } _ { j } +$$ + +for all $i , j \neq 0$ . + +Likewise for the eigenvectors of $\mathcal { N N } ^ { * }$ . If $v _ { i } ( y ) = p _ { Y } ( y ) b _ { i } ( y )$ : + +$$ +\overline { { b _ { i } b _ { j } } } = \langle v _ { i } , v _ { j } \rangle _ { Y } = 0 = \overline { { b } } _ { i } \overline { { b } } _ { j } . +$$ + +for all $i , j \neq 0$ . + +Importantly, this does not mean that the variables $u _ { 1 } , u _ { 2 } , \ldots$ nor $v _ { 1 } , v _ { 2 } , \ldots$ are “disentangled”, i.e., they are not statistically independent. These variables represent components in the space of probability vectors, rather than the “sample” space. They should be understood as spanning a subspace of the space of functions over the relevant independent variables. We discuss this in more detail in Section 5.3. + +# B.5 CORNERS OF $\mathcal { N }$ AND LOSS FUNCTION + +The final piece of puzzle we need, is the ability to express the components (corners) of $\mathcal { N }$ and $\mathcal { N } ^ { * }$ in the span of possible non-orthogonal families of variables. + +Let us therefore consider two arbitrary families $f _ { 1 } , \ldots , f _ { k _ { 0 } }$ and $g _ { 1 } , \ldots , g _ { k _ { 0 } }$ of variables, which respectively represent the vectors $p _ { X } f _ { j } \in V _ { X }$ and $p _ { Y } g _ { j } \in V _ { Y }$ . + +Firstly, we need matrices representing the components of the inner products on $V _ { X }$ and $V _ { Y }$ . Those are the symmetric matrices + +$$ +\begin{array} { r } { K _ { i j } = \langle p _ { X } f _ { i } , p _ { X } f _ { j } \rangle _ { X } = \overline { { f _ { i } f _ { j } } } , } \\ { L _ { i j } = \langle p _ { Y } g _ { i } , p _ { Y } g _ { j } \rangle _ { Y } = \overline { { g _ { i } g _ { j } } } . } \end{array} +$$ + +The components $N _ { i j }$ of $\mathcal { N }$ are defined by + +$$ +\mathcal { N } ( p _ { X } f _ { j } ) = \sum _ { i } N _ { i j } p _ { Y } g _ { i } . +$$ + +Taking the inner product with $p _ { Y } g _ { k }$ , we obtain + +$$ +\langle p _ { Y } g _ { k } , \mathcal { N } ( p _ { X } f _ { j } ) \rangle = \sum _ { i } N _ { i j } L _ { k i } . +$$ + +The left-hand side can be computed using Equ. 13. It is the matrix + +$$ +\begin{array} { l } { { \displaystyle { \cal A } _ { k j } = \langle p _ { Y } g _ { k } , { \cal N } ( p _ { X } f _ { j } ) \rangle } } \\ { { \displaystyle ~ = \sum _ { x , y } \frac { p _ { Y } ( y ) g _ { k } ( y ) p _ { Y | X } ( y | x ) p _ { X } ( x ) f _ { j } ( x ) } { p _ { Y } ( y ) } } } \\ { { \displaystyle ~ = \sum _ { x , y } p ( x , y ) g _ { k } ( y ) f _ { j } ( x ) = \overline { { g _ { k } f _ { j } } } . } } \end{array} +$$ + +Therefore, in matrix notation, Equ. (34) is $A = L N$ , or + +$$ +N = L ^ { - 1 } A . +$$ + +The components $N _ { i j } ^ { * }$ of $\mathcal { N } ^ { * }$ are obtained by just swapping $X$ and $Y$ , yielding + +$$ +N ^ { * } = K ^ { - 1 } A ^ { \top } . +$$ + +Hence the singular values of the corner of $\mathcal { N }$ defined by the variables $f _ { j }$ and $g _ { j }$ are just the squareroot of the eigenvalues of the matrix $N ^ { * } N = K ^ { - 1 } A ^ { \top } L ^ { - 1 } A$ . In order to find the variables $f _ { j }$ and $g _ { j }$ with the same span as the first $k _ { 0 }$ eigenvectors $u _ { j } , v _ { j }$ , we just need to maximize all the eigenvalues of $N ^ { * } N$ . A simple way to do this is to use (minus) the trace of $N ^ { * } N$ as loss function, since it is the sum of the square of the singular values. We call $\mathrm { T r } \left( N ^ { * } N \right)$ the relevance of the subspaces defines by the variables $f _ { j }$ ad $g _ { i }$ for all $i , j$ . This yields the loss/cost function: + +$$ +C = k _ { 0 } - \operatorname { T r } \left( N ^ { \ast } N \right) = k _ { 0 } - \operatorname { T r } \left( K ^ { - 1 } A ^ { \top } L ^ { - 1 } A \right) . +$$ + +Once optimal variables have been found, one can obtain the components of the eigenvectors in the span of $f _ { 1 } , \ldots , f _ { k _ { 0 } }$ through standard numerical diagonalization of $N ^ { * } N$ . + +# B.6 INFERENCE + +The variables minimizing $C$ can be used to infer one variable from the other. For instance, given $y$ , the inferred probability distribution over $x$ is given by $p _ { X | Y } ( x | y ) = \mathcal { N } ^ { * } ( \delta _ { y } ) ( x )$ , where $\delta _ { y } ( \bar { y } ^ { \prime } )$ is 1 when $y = y ^ { \prime }$ and zero otherwise. In order to compute this, we first need the components of the distribution $\delta _ { y }$ in terms of the family $p _ { Y } g _ { 1 } , \ldots , p _ { Y } g _ { k _ { 0 } }$ , i.e., the real numbers $( \delta _ { y } ) _ { j }$ such that + +$$ +\delta _ { y } ( y ^ { \prime } ) = p _ { Y } ( y ^ { \prime } ) \sum _ { i = 1 } ^ { k _ { 0 } } ( \delta _ { y } ) _ { i } g _ { i } ( y ^ { \prime } ) + r ( y ^ { \prime } ) , +$$ + +where $\langle r , p _ { Y } \delta _ { i } \rangle _ { Y } = 0$ for all $i$ . Taking the inner product with $p _ { Y } g _ { j }$ , we obtain + +$$ +\langle p _ { Y } g _ { j } , \delta _ { y } \rangle _ { Y } = \sum _ { i = 1 } ^ { k _ { 0 } } ( \delta _ { y } ) _ { i } L _ { j i } , +$$ + +where the left hand side is also just + +$$ +\langle p _ { Y } g _ { j } , \delta _ { y } \rangle _ { Y } = g _ { j } ( y ) . +$$ + +Therefore the components of $\delta _ { y }$ are explicitly + +$$ +( \delta _ { y } ) _ { i } = \sum _ { j } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) . +$$ + +It follows that + +$$ +\begin{array} { l } { { \displaystyle p _ { X | Y } ( x | y ) = \mathcal { N } ^ { * } ( \delta _ { y } ) ( x ) \approx \mathcal { N } _ { 0 } ^ { * } ( \delta _ { y } ) ( x ) } } \\ { { \displaystyle ~ = \sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) f _ { k } ( x ) . } } \end{array} +$$ + +Then, for instance, the expected inferred value of $X$ is + +$$ +\overline { { { x } } } = \sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) \sum _ { x } p _ { X } ( x ) x f _ { k } ( x ) . +$$ + +For the inference of $Y$ from $x$ , we have + +$$ +p _ { Y | X } ( y | x ) \approx \sum _ { i j k } N _ { k i } ( K ^ { - 1 } ) _ { i j } f _ { j } ( x ) g _ { k } ( y ) . +$$ + +# C ANALYTICAL EXAMPLE + +When $p ( x , y )$ is any multivariate Gaussian distribution, everything can be computed analytically. Let us consider here the one-dimensional case. We use $p ( x ) \ \propto \ \exp { \left( - x ^ { 2 } / 2 \tau ^ { 2 } \right) }$ , and the conditional $\underline { { { p } } } ( y | x ) ~ \propto ~ \exp \left( - ( y - x ) ^ { 2 } / 2 \sigma ^ { 2 } \right)$ . That is, $y$ is equal to $x$ but with some added Gaussian noise. This gives + +$$ +p _ { X | Y } ( x | y ) \propto \exp \left( - \frac { ( x - \gamma y ) ^ { 2 } } { 2 \tau ^ { 2 } ( 1 - \gamma ) } \right) , \quad \mathrm { w h e r e } \quad \gamma = \frac { \tau ^ { 2 } } { \sigma ^ { 2 } + \tau ^ { 2 } } . +$$ + +It was show in Lancaster (1958), that the most relevant subspace of dimension $k _ { 0 }$ on the variable $X$ is simply spanned by the variables + +$$ +f _ { n } ( x ) = x ^ { n } , +$$ + +$n = 0 , \ldots , k _ { 0 } - 1$ . Similarly for $Y$ ; + +$$ +g _ { n } ( y ) = y ^ { n } . +$$ + +This independence of the relevant variables on the detailed parameters of $p$ is a general property of Gaussian joint distributions. + +This means, for instance, that the most relevant feature $( n = 1$ ) for predicting the value of $X$ given $Y = y$ is simply $Y$ itself. The higher order variables have to do with inferring extra aspects of the probability distribution over $X$ . + +A set of orthogonal variables can be obtain from the Gram-Schmidt procedure, which, if done from small to large $n$ much necessarily yield the eigenvectors $u _ { n }$ and $v _ { n }$ . For illustration purpose, let us work with the non-orthogonal vectors $f _ { n }$ and $g _ { n }$ , keeping only the first $k _ { 0 } = 3$ vectors. + +The three matrices (correlators) we need can be easily computed: + +$$ +\begin{array} { l c c } { { K = \left( \begin{array} { c c c } { { 1 } } & { { 0 } } & { { \tau ^ { 2 } } } \\ { { 0 } } & { { \tau ^ { 2 } } } & { { 0 } } \\ { { \tau ^ { 2 } } } & { { 0 } } & { { 3 \tau ^ { 4 } } } \end{array} \right) } } & { { L = \left( \begin{array} { c c c } { { 1 } } & { { 0 } } & { { \tau ^ { 2 } + \sigma ^ { 2 } } } \\ { { 0 } } & { { \tau ^ { 2 } + \sigma ^ { 2 } } } & { { 0 } } \\ { { \tau ^ { 2 } + \sigma ^ { 2 } } } & { { 0 } } & { { 3 ( \tau ^ { 2 } + \sigma ^ { 2 } ) ^ { 2 } } } \end{array} \right) } } \\ { { A = \left( \begin{array} { c c c } { { 1 } } & { { 0 } } & { { \tau ^ { 2 } } } \\ { { 0 } } & { { \tau ^ { 2 } } } & { { 0 } } \\ { { \tau ^ { 2 } + \sigma ^ { 2 } } } & { { 0 } } & { { \tau ^ { 2 } ( \sigma ^ { 2 } + 3 \tau ^ { 2 } ) } } \end{array} \right) . } } \end{array} +$$ + +We obtain + +$$ +M = K ^ { - 1 } A ^ { \top } L ^ { - 1 } A = \left( \begin{array} { c c c } { { 1 } } & { { 0 } } & { { \tau ^ { 2 } ( 1 - \gamma ^ { 2 } ) } } \\ { { 0 } } & { { \gamma } } & { { 0 } } \\ { { 0 } } & { { 0 } } & { { \gamma ^ { 2 } } } \end{array} \right) . +$$ + +The eigenvalues of $M$ can be read on the diagonal, and the corresponding eigenvectors are $( 1 , 0 , 0 )$ , $( 0 , 1 , \bar { 0 } )$ and $( - \tau ^ { 2 } , 0 , 1 )$ , which means that the eigenfunctions are in order $\bar { u } _ { 0 } ( x ) = 1$ , $u _ { 1 } ( x ) = x$ and u2(x) = x2 − τ 2. + +Because we are working with continuous variables, the true rank of $\mathcal { N }$ is infinite, even for any finite cutoff on the singular values. Nevertheless, it is instructive to see how the approximate inference fares for rank $k _ { 0 } = 3$ . Given the value $y$ for $Y$ , the inferred distribution over $X$ is + +$$ +\mathcal { N } _ { 0 } ^ { * } ( \delta _ { y } ) ( x ) = p _ { X } ( x ) p _ { Y } ( y ) \sum _ { j , k = 0 } ^ { 2 } ( K ^ { - 1 } A ^ { \top } L ^ { - 1 } ) _ { k j } y ^ { j } x ^ { k } . +$$ + +The approximately inferred first and second moments of $X$ is given by integrating the above times $x$ (resp. $x ^ { 2 }$ ) over $x$ . We obtain + +$$ +{ \overline { { x } } } = \gamma y \quad { \mathrm { a n d } } \quad { \overline { { x ^ { 2 } } } } = \gamma ^ { 2 } y ^ { 2 } + ( 1 - \gamma ) \tau ^ { 2 } , +$$ + +which are actually exact: they are equal to the first two moments of $X$ over $p _ { X | Y }$ as given in Eq. (46). + +In fact, it is easy to see that this would be true for the first $k _ { 0 } - 1$ moments had we kept the $k _ { 0 }$ most relevant variables. \ No newline at end of file diff --git a/parse/train/SJeS16EKPr/SJeS16EKPr_content_list.json b/parse/train/SJeS16EKPr/SJeS16EKPr_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6182c118f4c07aef24f47d51ecbbcf461ad6287b --- /dev/null +++ b/parse/train/SJeS16EKPr/SJeS16EKPr_content_list.json @@ -0,0 +1,3054 @@ +[ + { + "type": "text", + "text": "LEARNING RELEVANT FEATURES FOR STATISTICAL INFERENCE ", + "text_level": 1, + "bbox": [ + 174, + 101, + 821, + 145 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 170, + 400, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 234, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given two views of data, we consider the problem of finding the features of one view which can be most faithfully inferred from the other. We find that these are also the most correlated variables in the sense of deep canonical correlation analysis (DCCA). Moreover, we show that these variables can be used to construct a non-parametric representation of the implied joint probability distribution. This representation can be used to compute the expectations of functions over one view of data conditioned on the other, such as Bayesian estimators and their standard deviations. We test the approach using inference on occluded MNIST images, and show that our representation contains multiple modes. Surprisingly, when applied to supervised learning (one dataset consists of labels), this approach automatically provides regularization and faster convergence compared to the cross-entropy objective. We also explore using this approach to discover salient independent variables of a single dataset. ", + "bbox": [ + 233, + 267, + 764, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 476, + 336, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given samples $( x _ { 1 } , y _ { 1 } ) , \\dotsc , ( x _ { n } , y _ { n } )$ from an unknown joint probability distribution $p ( x , y )$ , we want to construct a useful representation of the conditional probabilities $p ( x | y )$ and $p ( y | x )$ , so that we that we can infer one view from the other on new data. ", + "bbox": [ + 176, + 507, + 825, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For instance, $x$ and $y$ could be past and future histories of dynamical data, visual and auditory inputs, actions and their effects, etc. ", + "bbox": [ + 174, + 556, + 820, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following the approach introduced in Beny & Osborne (2013; 2015b) in the context of quantum ´ information theory, we look at the problem as follows: the conditional distributions $p ( y | x )$ can be thought of as representing a noisy communication channel (stochastic map). This channel is a linear map between spaces of typically ludicrously large dimensions (the spaces of all probability distributions over $x$ or $y$ ). We want a pair of small subspaces which best represent the channel. Specifically, we look for those vectors representing probability distributions over $x$ which lose least distinguishability under the channel, where the distinguishability is measured by the $\\chi ^ { 2 }$ divergence. ", + "bbox": [ + 173, + 592, + 825, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We show in Section 3 that this is equivalent to performing a certain singular value decomposition of the channel (seen as an operator in Hilbert space) and keep only the components with the largest singular values. Moreover, the full singular value decomposition is equivalent to the decomposition in terms of canonical variables introduced in Lancaster (1958), namely, ", + "bbox": [ + 173, + 695, + 825, + 752 + ], + "page_idx": 0 + }, + { + "type": "equation", + "img_path": "images/cd0226efb4f314de95eef92b4975cd11169f564dba4f7e7b395e743af0ef5d3d.jpg", + "text": "$$\np ( \\boldsymbol { y } | \\boldsymbol { x } ) = p ( \\boldsymbol { y } ) \\sum _ { i = 1 } ^ { D } \\eta _ { i } u _ { i } ( \\boldsymbol { x } ) v _ { i } ( \\boldsymbol { y } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 390, + 760, + 606, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $u _ { i }$ and $u _ { j }$ are real non-linear functions such that $\\mathbb { E } ( u _ { i } u _ { j } ) = \\delta _ { i j }$ , $\\mathbb { E } ( v _ { i } v _ { j } ) = \\delta _ { i j }$ , and $0 ~ <$ $\\eta _ { D } \\leq \\cdot \\cdot \\cdot \\leq \\eta _ { 1 } \\leq \\eta _ { 0 } = 1$ are the singular values. ", + "bbox": [ + 173, + 810, + 823, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The practical advantage of this representation for inference (or prediction) is that it reduces the evaluation of conditional expectations to that of empirical averages over the (unconditional) marginal $p ( y )$ . ", + "bbox": [ + 173, + 845, + 825, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As observed in Michaeli et al. (2016), the span of the first $k$ canonical variables $u _ { i } , \\ v _ { j }$ is what is learned by the deep canonical correlation analysis (DCCA) (Andrew et al., 2013). Indeed, these ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "variables are those which maximize the correlations $\\mathbb { E } ( u _ { i } v _ { i } )$ subject to the same constraints as above. \n(This reduces to CCA (Hotelling, 1936) when $u _ { i }$ , $v _ { j }$ are linear maps). ", + "bbox": [ + 171, + 103, + 821, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, besides establishing this new information-theoretical interpretation of canonical variables and DCCA, we experiment with using this representation for performing inference on new data. Moreover, we propose a strategy for extracting disentangled variables from the canonical variables, inspired by analytical solutions. ", + "bbox": [ + 174, + 138, + 823, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 215, + 341, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This general problem (of building an effective representation of the conditional probability distributions implied by joint samples) covers many existing approaches in different contexts. For instance, if the variables $y$ has few possible states, then it reduces to a classification problem, usually solved by minimizing the crossentropy between a predicted distribution and the one-hot encoding of the classes. ", + "bbox": [ + 174, + 247, + 825, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When $y$ has a large number of states, or is fundamentally continuous, existing approaches usually do not model the whole conditional distribution, but either provide the average (regression), or approximately sample from it. ", + "bbox": [ + 174, + 324, + 825, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main class of methods which allows for sampling from the conditional distributions are variational: a deterministic neural networks produces the parameters of analytical classes of probabilities. \nThis includes variational autoencoders Kingma & Welling (2013) (e.g., Iten et al. (2018); Wang et al. \n(2016)), and approaches based on the minimum description length principle such as Gregor et al. \n(2013). ", + "bbox": [ + 174, + 373, + 825, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternatively, it may also be possible to use adversarial training Goodfellow et al. (2014), by using a conditional Mirza & Osindero (2014) version of an energy-based GAN Zhao et al. (2016). ", + "bbox": [ + 174, + 450, + 823, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By contrast, our approach doesn’t require the training of a generative model. Instead, conditional expectations are constructed as linear combinations of unconditional empirical averages over the training data. ", + "bbox": [ + 174, + 484, + 825, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous approaches to equipping CCA or DCCA with an information-theoretic interpretation have explored different directions. For instance, in Wang et al. (2016), the authors generalize a probabilistic interpretation for CCA in terms of gaussian distributions, which leads to a variational approach. In Painsky et al. (2018), additional constraints on the mutual information between the data and the learned variables are added to the optimizations. ", + "bbox": [ + 174, + 534, + 825, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A previous attempt at designing a numerical solution for our singular value problem can be found in Beny (2018b). In that work, the relevant variables were represented through a PCA kernel produced ´ by Monte Carlo sampling, but wasn’t practical. ", + "bbox": [ + 174, + 611, + 825, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 THEORY ", + "text_level": 1, + "bbox": [ + 176, + 674, + 276, + 690 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formalize the problem by assuming that our data was sampled from an unknown joint distribution $p ( x , y )$ over two random variables $X$ and $Y$ . ", + "bbox": [ + 173, + 707, + 823, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $V _ { X }$ and $V _ { Y }$ denote the linear spaces spanned by all probability distributions over $X$ and $Y$ respectively. Here we assume that $X$ and $Y$ take finitely many values for simplicity, but this formalism can be straightforwardly extended to infinite-dimensional vector spaces. ", + "bbox": [ + 174, + 741, + 823, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We will need inner products on $V _ { X }$ and $V _ { Y }$ , to make them into real Hilbert spaces. We use the Fisher information metrics evaluated at the points $p ( x )$ and $p ( y )$ respectively (marginals of $p ( x , y ) \\mathrm { . }$ ), that is, ", + "bbox": [ + 173, + 790, + 825, + 830 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/8d970395cca7301e1fc810391ed22d29995147be6897298796ca3b5d2b948d9a.jpg", + "text": "$$\n\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } : = \\sum _ { x } { \\frac { \\mu ( x ) \\mu ^ { \\prime } ( x ) } { p ( x ) } } \\quad { \\mathrm { a n d } } \\quad \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } : = \\sum _ { y } { \\frac { \\nu ( y ) \\nu ^ { \\prime } ( y ) } { p ( y ) } }\n$$", + "text_format": "latex", + "bbox": [ + 290, + 829, + 707, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for any vectors $\\mu , \\mu ^ { \\prime } \\in V _ { X }$ and $\\nu , \\nu ^ { \\prime } \\in V _ { Y }$ . ", + "bbox": [ + 174, + 873, + 455, + 890 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Below we also call the marginals $p _ { X }$ and $p _ { Y }$ respectively when omitting their arguments $( p _ { X } ( x ) \\equiv$ $p ( x )$ and $p _ { Y } ( y ) \\equiv p ( y ) \\rangle$ . ", + "bbox": [ + 173, + 895, + 825, + 925 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These inner products allow us to define the $\\chi ^ { 2 }$ divergence: ", + "bbox": [ + 173, + 102, + 558, + 118 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/94f8cac2c398495c9f82111acc830c8d5618725abd8916dddbcaf9f6f22579a4.jpg", + "text": "$$\n\\chi ^ { 2 } ( q , p _ { X } ) = \\langle q - p _ { X } , q - p _ { X } \\rangle _ { X } ,\n$$", + "text_format": "latex", + "bbox": [ + 385, + 126, + 611, + 145 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which a measures of statistical distinguishability between $q$ and $p _ { X }$ . Specifically, it quantifies how easy it is to reject the null hypothesis that the state is $p _ { X }$ when it is actually $q$ , based on the empirical distribution obtained from independent samples. It is also the lowest order approximation of the Kullback-Leibler divergence. ", + "bbox": [ + 173, + 152, + 825, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The joint distribution $p ( x , y )$ yields conditional distributions $p ( y | x )$ and $p ( x | y )$ . These can be understood as the components (or kernels) of stochastic maps $\\mathcal { N } : V _ { X } V _ { Y }$ and $\\mathcal { N ^ { * } } : V _ { X } V _ { Y }$ respectively. Explicitely, if $\\mu \\in V _ { X }$ and $\\nu \\in V _ { Y }$ , then the images $\\mathcal { N } ( \\mu ) \\in V _ { Y }$ and $\\mathcal { N } ^ { * } ( \\nu ) \\in V _ { X }$ are defined by ", + "bbox": [ + 173, + 214, + 825, + 271 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/83efb853a41eb697ad26f5434c59976d25c58e676b4343b10adc019e7bfde6eb.jpg", + "text": "$$\n{ \\mathcal { N } } ( \\mu ) ( y ) = \\sum _ { x } p ( y | x ) \\mu ( x ) \\quad { \\mathrm { a n d } } \\quad { \\mathcal { N } } ^ { * } ( \\nu ) ( x ) = \\sum _ { y } p ( x | y ) \\nu ( y ) .\n$$", + "text_format": "latex", + "bbox": [ + 279, + 272, + 717, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These stochastic maps $\\mathcal { N }$ and $\\mathcal { N } ^ { * }$ perform inference of one variable given some (possibly imperfect) knowledge about the other, with priors given by the marginals $p ( x )$ or $p ( y )$ of $p ( x , y )$ depending on the direction of the inference. Importantly, $\\mathcal { N } ^ { * }$ is the transpose of $\\mathcal { N }$ in terms of the inner products defined above: ", + "bbox": [ + 173, + 311, + 825, + 367 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/677d78296d305d22115c0b00ccfa8cc209f4016332046b6803faa126393f9014.jpg", + "text": "$$\n\\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } = \\langle \\mathcal { N } ^ { * } ( \\nu ) , \\mu \\rangle _ { X } .\n$$", + "text_format": "latex", + "bbox": [ + 401, + 368, + 596, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now have the tools to address the problem mentioned in the introduction. The distinguishability between $q \\in V _ { X }$ and $p _ { X }$ after the action of the channel $\\mathcal { N }$ is $\\chi ^ { 2 } ( \\mathcal { N } ( q ) , p _ { Y } )$ since $p _ { Y } = \\mathcal { N } ( p _ { X } )$ . Hence we want to find the distributions $p$ which maximize the relevance Beny & Osborne (2013). ´ ", + "bbox": [ + 173, + 398, + 825, + 443 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/8efbc9ac41de1e15ecde109b8c4a2a5349e45c8fbbd6588ce8bc655870bab5e1.jpg", + "text": "$$\n\\eta ( q ) = \\frac { \\chi ^ { 2 } ( \\mathcal { N } ( q ) , p _ { Y } ) } { \\chi ^ { 2 } ( q , p _ { X } ) } = \\frac { \\langle \\mathcal { N } ( \\mu ) , \\mathcal { N } ( \\mu ) \\rangle _ { Y } } { \\langle \\mu , \\mu \\rangle _ { X } } ,\n$$", + "text_format": "latex", + "bbox": [ + 354, + 449, + 643, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu = q - p _ { X }$ . The inner-product formulation makes it clear that this amounts to finding the eigenvector with largest eigenvalue for the symmetric map $\\mathcal { N } ^ { * } \\mathcal { N }$ , which is also the singular vector with largest singular value for $\\mathcal { N }$ . On can then go on to find the eigenvector with next largest eigenvalue and so on, which are automatically orthogonal. ", + "bbox": [ + 173, + 492, + 825, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, the inner products are more tractable to compute if we express elements $\\mu \\in V _ { X }$ and $\\nu \\in V _ { Y }$ in terms of variables $f$ and $g$ as $\\mu = p _ { X } f$ and $\\nu = p _ { Y } g$ , or ", + "bbox": [ + 173, + 555, + 823, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/6cd28b0f04c36b2ba6bd2535b122fe953ab7bf7a2a9122ea177314f4b21519e4.jpg", + "text": "$$\n\\mu ( x ) = p ( x ) f ( x ) \\quad { \\mathrm { a n d } } \\quad \\nu ( y ) = p ( y ) g ( y )\n$$", + "text_format": "latex", + "bbox": [ + 354, + 593, + 643, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for all $x , y$ . Indeed, this yields simply ", + "bbox": [ + 173, + 618, + 421, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/687c2c575cdafbc77f87dfaa38adba10014f26a4561574c1643cda0cc9e5e626.jpg", + "text": "$$\n\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } = \\operatorname { \\mathbb { E } } ( f f ^ { \\prime } ) \\quad { \\mathrm { a n d } } \\quad \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } = \\operatorname { \\mathbb { E } } ( g g ^ { \\prime } ) .\n$$", + "text_format": "latex", + "bbox": [ + 343, + 641, + 653, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are now in measure to make the connection with DCCA Andrew et al. (2013). Indeed, the aims of DCCA is to maximize the correlations $\\operatorname { c o r r } ( f , g ) = \\mathbb { E } ( f g )$ over function $f ( x )$ and $g ( y )$ such that $\\mathbb { E } ( f ^ { 2 } ) = \\mathbb { E } ( g ^ { 2 } ) = 1$ . But, using $\\mu = p _ { X } f$ and $\\nu = p _ { Y } g$ , we have ", + "bbox": [ + 173, + 675, + 826, + 718 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/0a9ed56485d8397567b19155defc4b1b6b44cf20e2c3c5b2dbe6898eb1ca495e.jpg", + "text": "$$\n\\mathbb { E } ( f g ) = \\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } ,\n$$", + "text_format": "latex", + "bbox": [ + 423, + 727, + 571, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is maximized by the left- and right- singular vectors $\\mu$ and $\\nu$ of $\\mathcal { N }$ with largest singular value. ", + "bbox": [ + 173, + 751, + 821, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given all the singular vectors $\\mu _ { i } = p _ { X } f _ { i }$ and $\\nu _ { i } = p _ { Y } g _ { i }$ with singular values $\\eta _ { i }$ , we obtain the representation ", + "bbox": [ + 173, + 773, + 823, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/0899cfdd5bcded9b0b8fce6ad2c29faa2b4a24e5a10c10304414279e734be06b.jpg", + "text": "$$\n\\mathcal { N } ( \\mu ) = \\sum _ { i } \\eta _ { i } \\nu _ { i } \\langle \\mu _ { i } , \\mu \\rangle _ { X } ,\n$$", + "text_format": "latex", + "bbox": [ + 408, + 803, + 588, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which, using a more standard notation and the Kronecker delta $\\delta _ { x }$ , yields Eq. 1: ", + "bbox": [ + 173, + 840, + 694, + 857 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/cac76d1ff3ed640c21c85d26037938c5c2ae2e784b2c708e314ec2840f41a26a.jpg", + "text": "$$\np ( y | x ) = { \\mathcal { N } } ( \\delta _ { x } ) ( y ) = { \\frac { 1 } { p ( x ) } } \\sum _ { i } \\eta _ { i } \\nu _ { i } ( y ) \\mu _ { i } ( x ) = p ( y ) \\sum _ { i } \\eta _ { i } v _ { i } ( y ) u _ { i } ( x ) ,\n$$", + "text_format": "latex", + "bbox": [ + 261, + 863, + 735, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu _ { i } = p _ { X } u _ { i }$ and $\\nu _ { i } = p _ { Y } v _ { i }$ . ", + "bbox": [ + 173, + 910, + 397, + 925 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the purpose of the optimization and inference, we do not need to full diagonal decomposition, but just functions $f _ { i } = \\mu _ { i } / p _ { X }$ and $g _ { j } = \\nu _ { j } / p _ { Y } , i , j = 1 , \\ldots , k _ { 0 }$ , which have the same span as the canonical variables $u _ { i }$ and $v _ { j }$ respectively for $i , j = 1 , \\ldots , k _ { 0 }$ (assuming that $\\eta _ { 1 } , \\ldots , \\eta _ { k _ { 0 } }$ are the largest singular vector). Below we refer to $f _ { i }$ and $g _ { j }$ as the $k _ { 0 }$ most relevant variables. ", + "bbox": [ + 173, + 128, + 825, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because these functions may not be orthogonal, we need the covariance matrices ", + "bbox": [ + 176, + 193, + 705, + 208 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/454c207fea8af6fb7e54a62760ae9f11fdb9b33fecf0925c06e090123d1dd7e3.jpg", + "text": "$$\nK _ { i j } = \\langle \\mu _ { i } , \\mu _ { j } \\rangle _ { X } = \\mathbb { E } ( f _ { i } f _ { j } ) , \\quad L _ { i j } = \\langle \\nu _ { i } , \\nu _ { j } \\rangle _ { X } = \\mathbb { E } ( g _ { i } g _ { j } ) , \\quad A _ { i j } = \\langle \\nu _ { i } , N ( \\mu _ { j } ) \\rangle _ { Y } = \\mathbb { E } ( g _ { i } f _ { j } ) .\n$$", + "text_format": "latex", + "bbox": [ + 184, + 214, + 810, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If $N _ { i j }$ denote the components of $\\mathcal { N }$ in the sense that $\\begin{array} { r } { \\mathcal { N } ( p _ { X } f _ { j } ) = \\sum _ { i } N _ { i j } p _ { Y } g _ { i } } \\end{array}$ , then, using our inner products to isolate $N _ { i j }$ , we obtain $N = L ^ { - 1 } A$ . Similarly, the components of $\\mathcal { N } ^ { * }$ are $N _ { i j } ^ { * } =$ $K ^ { - 1 } A ^ { \\top }$ . This implies that the sum of the square of the singular values of $\\mathcal { N }$ restricted to the spans of the vectors $p _ { X } f _ { i }$ and $p _ { Y } g _ { j }$ for all $i , j$ , which is what we want to maximize, is just given by ", + "bbox": [ + 173, + 244, + 825, + 308 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/f1587c308117d7252ef0335d24a7a140dc31be7bcbc658867615280d59a8460e.jpg", + "text": "$$\n\\sum _ { i = 1 } ^ { k _ { 0 } } \\eta _ { i } ^ { 2 } = \\mathrm { T r } \\left( N ^ { \\ast } N \\right) = \\mathrm { T r } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A ) .\n$$", + "text_format": "latex", + "bbox": [ + 349, + 315, + 643, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This is the DCCA objective. Below we use the objective function $C = k _ { 0 } - \\mathrm { T r } \\left( N ^ { \\ast } N \\right)$ , for the cosmetic reason that its optimal value is zero. ", + "bbox": [ + 171, + 364, + 823, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, the corresponding truncated representation of the conditional distribution is ", + "bbox": [ + 169, + 401, + 743, + 416 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/d0d1286f608b3021f23eb2026fa6e81a31d907cd4c6e4cc4877260801cf01603.jpg", + "text": "$$\np ( \\boldsymbol { y } | \\boldsymbol { x } ) = \\mathcal { N } ( \\delta _ { \\boldsymbol { x } } ) ( \\boldsymbol { y } ) \\simeq p ( \\boldsymbol { y } ) \\sum _ { i , j = 1 } ^ { k _ { 0 } } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } g _ { i } ( \\boldsymbol { y } ) f _ { j } ( \\boldsymbol { x } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 302, + 422, + 694, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we used the fact that the components of $\\delta _ { x }$ are $\\begin{array} { r } { \\delta _ { j } = \\sum _ { i } K _ { j i } ^ { - 1 } f _ { i } ( x ) } \\end{array}$ . ", + "bbox": [ + 174, + 476, + 658, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Of course, This approach can produce a faithful representation of the correlations only if $\\mathcal { N }$ is actually close to being of rank $k _ { 0 }$ (see Appendix $\\mathbf { B }$ for a more precise statement). If we interpret the relevant subspace as a space of probability over latent variable, this means that our latent variables have at most $k _ { 0 }$ discrete states. ", + "bbox": [ + 173, + 498, + 825, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, even if the rank $k _ { 0 }$ corner of $\\mathcal { N }$ is a not a good approximation, this strategy allows us to nevertheless do the correct inference on certain random variables, namely those which are in the span of the canonical variables! ", + "bbox": [ + 173, + 560, + 825, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Indeed, the exact conditional expectation of $g _ { k }$ is (assuming $D$ is the actual rank of $\\mathcal { N }$ ), ", + "bbox": [ + 173, + 609, + 745, + 626 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/ec316fca8c448843caf2ac556e4420de40df72f1ce473ed26529bbebf75318f8.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle \\sum _ { y } g _ { k } ( y ) p ( y | x ) = \\sum _ { i , j = 1 } ^ { D } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } \\mathbb { E } ( g _ { k } g _ { i } ) f _ { j } ( x ) } } \\\\ { { \\displaystyle \\qquad = \\sum _ { j = 1 } ^ { D } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) = \\sum _ { j = 1 } ^ { k _ { 0 } } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 285, + 632, + 710, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the last truncation is exact if $k \\leq k _ { 0 }$ due to the assumption that the basis $f _ { i }$ and $g _ { j }$ have the same span as the $k _ { 0 }$ largest right and left singular vectors of $\\mathcal { N }$ respectively. ", + "bbox": [ + 171, + 729, + 823, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For instance, if $p ( x , y )$ is Gaussian, the canonical variables can be computed analytically, as in Lancaster (1958) or Beny (2018a) in the multivariate case. Solutions for other distributions were ´ also computed in Eagleson (1964). ", + "bbox": [ + 174, + 765, + 823, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, for any two dimensional Gaussian, the space of $k$ most relevant variables is simply spanned by the moments $f _ { n } ( x ) = x ^ { n }$ and $g _ { n } ( y ) = y ^ { n }$ for $n = 0 , \\ldots , k - 1$ . Hence, in this case the first $k$ moments can be inferred exactly using only the $k + 1$ most relevant variables (See Appendix C). ", + "bbox": [ + 174, + 814, + 825, + 857 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 ALGORITHM ", + "text_level": 1, + "bbox": [ + 174, + 877, + 312, + 893 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us explicit the algorithm resulting from the above analysis. ", + "bbox": [ + 173, + 909, + 583, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assume that we are given independent samples $( x _ { 1 } , y _ { 1 } ) , ( x _ { 1 } , y _ { 2 } ) , . . .$ from the otherwise unknown joint distribution $p ( x , y )$ . ", + "bbox": [ + 171, + 103, + 823, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first perform DCCA (Andrew et al., 2013). That is, we need two independent deterministic feedforward neural networks. The first maps $x$ to a set of $k _ { 0 }$ real-valued variables $f _ { 1 } ( x ) , \\ldots , f _ { k _ { 0 } } ( x )$ . The second maps $y$ to a different set of $k _ { 0 }$ variables $g _ { 1 } ( y ) , \\ldots , g _ { k _ { 0 } } ( y )$ . ", + "bbox": [ + 174, + 138, + 823, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The parameters of the neural networks are to be set to minimize the objective function ", + "bbox": [ + 171, + 188, + 740, + 202 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/18fde4f521fc48a4fa7b3af4096adc86fcb6939f0c44a4dcc3bba5d1292cffc4.jpg", + "text": "$$\nC = k _ { 0 } - \\operatorname { T r } { \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) } ,\n$$", + "text_format": "latex", + "bbox": [ + 395, + 207, + 601, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the matrices $K , L , A$ can be approximated over a mini-batch $( x _ { n } , y _ { n } )$ , $n = 1 , \\ldots , N$ via ", + "bbox": [ + 176, + 231, + 797, + 246 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/add296e57eaf2458c135e33a4e76c91636e008183e45eab3b6dc43aae8bb61df.jpg", + "text": "$$\nK _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } f _ { i } ( x _ { n } ) f _ { j } ( x _ { n } ) , \\quad L _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) g _ { j } ( y _ { n } ) , \\quad A _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) f _ { j } ( x _ { n } ) .\n$$", + "text_format": "latex", + "bbox": [ + 186, + 250, + 789, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We found that, provided the batch size is sufficiently large compared to $k _ { 0 }$ (about 10 times in our experience), this can be minimized using ADAM or direct gradient descent. However, to guarantee stability when using large $k _ { 0 }$ , we needed to explicit the gradient of the objective function in order to force the use of the Moore-Penrose pseudo-inverses for $K ^ { - 1 }$ and $L ^ { - 1 }$ in both the forward and backward passes, in addition to using 64 bits floats in these computations. ", + "bbox": [ + 173, + 304, + 825, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the relevant variables have been learned, we still need to use the training data in a second step. Indeed, suppose that we wish to use our model to infer the value of some function $\\Theta ( x )$ , i.e., to compute its approximate expectation value in terms of the conditional distribution $x \\mapsto p ( x | y )$ . Then we need to store, for each variable $j = 1 , \\ldots , k _ { 0 }$ , the quantities ", + "bbox": [ + 174, + 381, + 825, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/5f2e2f9c14a534f7c9a30a260312b27d1541ae4a51197d6d82177cf1c9d784b2.jpg", + "text": "$$\n\\Theta _ { j } = \\frac { 1 } { N _ { \\mathrm { f u l l } } } \\sum _ { n = 1 } ^ { N _ { \\mathrm { f u l l } } } \\Theta ( x _ { n } ) f _ { j } ( x _ { n } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 393, + 443, + 602, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the average is to be taken on the full training batch (of size $N _ { \\mathrm { f u l l } } ,$ . The same can be done exchanging $x$ with $y$ and $f _ { j }$ with $g _ { j }$ for the reverse inference. ", + "bbox": [ + 171, + 489, + 823, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For instance, if a data point $x$ is composed of real components $x ^ { a }$ —such as pixel color components for an image—and we are interested in the estimator which minimize the expected $l ^ { 2 }$ distance to the predicted values of these components, then we need the expectation values of the components $\\Theta ( x ) = x ^ { a }$ for all $a$ , and possibly higher moments to gain more knowledge about the shape of the posterior distribution, such as the second moments $\\Theta ^ { \\prime } ( \\bar { x } ) = x ^ { 2 }$ , etc. ", + "bbox": [ + 173, + 525, + 825, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference can then be performed with new data using ", + "bbox": [ + 174, + 602, + 524, + 616 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/e466bd2d915cdb5122b5c1773ab354787b45130d24c02c188710baa5586101e8.jpg", + "text": "$$\n\\overline { { { \\Theta } } } = \\sum _ { x } p ( x | y ) \\Theta ( x ) \\approx \\sum _ { i , j = 1 } ^ { k _ { 0 } } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } ) _ { j i } \\Theta _ { j } g _ { i } ( y ) .\n$$", + "text_format": "latex", + "bbox": [ + 310, + 621, + 686, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, the accuracy of thof the relevant variables, i.e., t depend on, for which $k _ { 0 }$ $\\Theta$ taken in the span. $\\begin{array} { r } { \\Theta ( x ) = \\sum _ { i = 1 } ^ { k } c _ { i } f _ { i } ( x ) } \\end{array}$ $\\begin{array} { r } { \\Theta _ { j } = \\sum _ { i = 1 } ^ { k _ { 0 } } c _ { i } K _ { i j } } \\end{array}$ ", + "bbox": [ + 174, + 678, + 823, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reverse inference formulas are obtained simply by the exchanges $K L$ , $A A ^ { \\top }$ , and $g _ { i } f _ { i }$ . ", + "bbox": [ + 173, + 717, + 825, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 765, + 328, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In all our experiments, we used the ADAM optimizer with learning rate 0.001. We used the Flux package (Innes, 2018) for Julia, as well as Tensorflow. ", + "bbox": [ + 173, + 796, + 823, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As usual the data is divided into a training set and a testing set. No aspect of the testing set is used during training. The loss function refers to Eq. (7). In order to monitor overfitting, we compute a “test loss” and a “training loss”. The test loss is computed from the trained variables using only the test data, and accordingly, the training loss is computed purely using the training data. ", + "bbox": [ + 174, + 832, + 825, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, when performing inference on test data using Eq. (10), we use the covariances $A , L , K$ and expectations $\\Theta _ { j }$ (Equ. (9)) built from the training data only. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1 INFERENCE ON OCCLUDED MNIST ", + "text_level": 1, + "bbox": [ + 176, + 103, + 457, + 117 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this experiment, we use the left and right halves of the MNIST digit images as correlated variables $X$ and $Y$ . The goal is to obtain the expected left halves given the right halves, or vice versa. ", + "bbox": [ + 173, + 138, + 821, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training set was augmented by random small rotations and displacements to make the task more ambiguous, as we want to explore the uncertainty in the prediction. ", + "bbox": [ + 176, + 174, + 821, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The relevant variables were represented by two convolutional neural networks of identical architecture. They are composed of four convolutional layers and one fully connected layer, an architecture that performs well for supervised learning on this dataset. For ease of implementation, these CNN have the whole image as input, but with either half zeroed (same value as black pixels). Half-width CNNs with proper padding at the cut perform similarly. ", + "bbox": [ + 174, + 210, + 825, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After training, we used the training dataset to also compute the expected pixel gray value as well as their covariance for each relevant variable using Eq. (9). ", + "bbox": [ + 176, + 286, + 821, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These were used into Eq. (10) to compute the mean pixel gray values and their covariances over the conditional probability of $X$ given $Y$ on test data. This mean is the Bayesian estimator for the $l ^ { 2 }$ distance between half images, i.e., it should minimize the expected distance $d _ { l ^ { 2 } }$ over the conditional distribution, where $\\begin{array} { r } { d _ { l ^ { 2 } } ^ { 2 } ( x , \\overline { { y } } ) = \\sum _ { i } ( x _ { i } - y _ { i } ) ^ { 2 } } \\end{array}$ , where $x _ { i } \\in [ 0 , 1 ]$ is the value of the $\\mathrm { i } ^ { t h }$ pixel. (This is equivalent to minimizing the mean square error). ", + "bbox": [ + 174, + 321, + 825, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results on a randomly selected subset of test digits is shown in Fig. 1. For each example, we also computed the images obtained by adding plus or minus one standard deviation along the direction of greatest variance in the space of relevant variables. This reveals the main ambiguities (such as between 8 and 3 or 7 and 9 which share a similar right half). ", + "bbox": [ + 174, + 398, + 825, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The graph of the singular values shows that the rank cutoff of 200 is too low to capture all of the relevant variables (the sudden drop at the end is not robust to an increase in the cutoff), but the results are reasonable nevertheless. This shows that our representation of the conditional distributions contains valuable information besides the simple mean. ", + "bbox": [ + 174, + 462, + 825, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2 SUPERVISED LEARNING ", + "text_level": 1, + "bbox": [ + 176, + 559, + 379, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the context of a supervised classification task, one of the dataset (the labels) is of sufficiently low dimensionality that we can use a complete basis over its probability space as our relevant variables, such as the standard one-hot encoding of labels. This serves as a good first sanity test for our approach. Surprisingly, we find that it converges faster than standard approaches, and without the need for regularization. ", + "bbox": [ + 174, + 594, + 825, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let the variable $Y$ stands for the labels, with values in $\\{ 1 , \\ldots , k \\}$ . The probability space consists of vectors with $k$ real components. The canonical basis corresponds to the one-hot encoding $g _ { i } ( j ) =$ $\\delta _ { i j }$ (Kronecker delta). All we need is a neural network to encode $k$ variables $f _ { 1 } , \\ldots , f _ { k }$ on $X$ . After learning the most relevant variables $f _ { i }$ , we apply the reverse of Eq. (10) for function $\\Theta ( y ) = y$ , and use the maximum component of expected value $\\overline { y }$ to infer the labels from the data. ", + "bbox": [ + 174, + 671, + 825, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let us refer to this procedure as DCCI (Deep Canonical Correlations based Inference). ", + "bbox": [ + 176, + 747, + 740, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We tested this approach on the MNIST and CIFAR10 datasets, and compared the results to the standard cross-entropy objective (Fig. 2). ", + "bbox": [ + 176, + 768, + 820, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We plotted the accuracy as function of the epoch rather than clock time which would depend on many factors. But the time per epoch is roughly the same for each approaches in the above experiments. Indeed, the training time is dominated by the forward and backward evaluations of the neural networks which are identical. (However, the time it takes to evaluate our objective can become significant for much larger number of labels $k$ , since it involves the inversion of matrices of dimension $k$ . This is in addition to the fact that a greater dimension would require also larger batches.) ", + "bbox": [ + 174, + 804, + 823, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We found that, without regularization, simply changing the objective from cross-entropy to DCCI provided a large improvement both of convergence speed and final accuracy for both models. ", + "bbox": [ + 174, + 895, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/201f3ce00c47dbafcb9ef33dbcddfe321ec4b0f6d68f48039079f27a17681d75.jpg", + "image_caption": [ + "Figure 1: Left mosaic: the left halves of the MNIST digits in a random sample from the test set are inferred from the right half, with cutoff $k _ { 0 } = 2 0 0$ . Three images are shown for each digit. Within each triplet, the middle image represents the mean over pixel intensity of the inferred condition distribution, while left and right images corresponds to a plus and minus one standard deviation from the mean in the direction of largest covariance (in the space of half-images). A particularly interesting example is highlighted. Top-right: loss per epoch for $k _ { 0 } ~ = ~ 2 0 0$ . Bottom-right: the singular values for different values of the cutoff $k _ { 0 }$ , after 150 epochs in each case. " + ], + "image_footnote": [], + "bbox": [ + 183, + 116, + 831, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1a73e5112ee1b5d1d1b809be05b2c39b82dfe510aa9cbf963a2647ac4f9eb88.jpg", + "image_caption": [ + "Figure 2: Loss and inaccuracy (error rate) on test sets for two classification tasks. The models were trained either using the cross-entropy (CE) or our approach (DCCI), with or without regularization layers. On the MNIST dataset, we used an “all CNN” network, and for the CIFAR10 dataset we used a short VGG variation with 10 convolutions and 3 fully connected layers. In the regularized form, post-activation Batchnorm layers were placed after each convolutional layers on the VGG network. What is shown is the mean over 10 independent runs for MNIST and 5 runs for CIFAR10. The shaded area spans the standard deviation. ADAM with default parameters was used in all cases. No data augmentation was used except for horizontal flips for CIFAR10 (resulting in epochs of 100,000 images). " + ], + "image_footnote": [], + "bbox": [ + 200, + 570, + 818, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03b9882c59b4695eeffc013e7377800c0cf409c81087a146cf7fe4fc6b062c90.jpg", + "image_caption": [ + "Figure 3: Top-left: First 20 relevant variables (on $X$ ) determined by DCCA for a system where $X$ consists of two coordinates uniformly sampled over a circle and a surrounding ring, and $Y$ consists of the same points but shifted by a small normally distributed vector. The variables are arranged from left-to-right and top-to-bottom in order of decreasing relevance. Top-right: the same variables multiplied by the marginal $p _ { X }$ . Bottom row: introducing a gap in the ring allows for a monotonous function of the angle to serve as second most relevant variable (instead of the sine/cosine couple). Hence the angle is automatically “disentangled” from the other variables. (Mid-gray represents the value 0). " + ], + "image_footnote": [], + "bbox": [ + 184, + 101, + 826, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On MNIST, DCCI alone also outperformed cross-entropy with dropout. (Dropout did not yield any improvement in conjunction with DCCI). However, adding batch-normalization layers on the CIFAR example, erased any distinction between DCCI and cross-entropy. ", + "bbox": [ + 174, + 460, + 825, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 STRUCTURE OF THE RELEVANT VARIABLES ", + "text_level": 1, + "bbox": [ + 174, + 520, + 514, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We mentioned in Section 3 that if $p ( x , y )$ is a two-dimensional Gaussian distribution with zero mean, then the $n$ most relevant variables of $X$ are the first $n$ powers of $X$ itself, independently of the covariance matrix. This implies that the canonical variables are the Hermite polynomials in $X$ (which results from applying the Gram-Schmidt procedure to the basis $\\{ 1 , x , x ^ { 2 } , \\bar { . . . } \\bar \\} \\}$ . ", + "bbox": [ + 174, + 545, + 825, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A similar property holds for multivariate Gaussians, namely, the less relevant singular values are polynomials in the more relevant ones. If this is true more generally, it should be possible to further compress and organize the latent space extracted with DCCA by finding a minimal set of generators, which ought to also be in the span of the most relevant variables. ", + "bbox": [ + 174, + 608, + 825, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We applied DCCA to a synthetic dataset to explore this idea, shown in Fig. 3. In this case, we actually performed a final SVD to obtain the unique uncorrelated canonical variables, and ordered them by decreasing relevance (their respective singular values). ", + "bbox": [ + 174, + 671, + 825, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $X$ consists of two real numbers, distributed uniformly within a ring and a disk. The variable $Y$ is obtained by adding a random Gaussian shift to $X$ with a small standard deviation. The more relevant variables ought to be those which are more robust to such small random displacement. This formalizes the idea that we are interested in extracting “large-scale” variables Beny (2018b). ´ ", + "bbox": [ + 174, + 720, + 825, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We would expect the relevant independent variables to be: the binary variable indicating whether the point is in the disk or the ring and the angle around the ring, followed by the radial component in the ring, and finally the Cartesian coordinates inside the disk. This is precisely what we see in Fig. 3. ", + "bbox": [ + 174, + 784, + 823, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Indeed—if we put aside for now the fact that the angle itself is not directly represented—besides the constant function, the two most relevant variables are the sine and cosine of the angle, followed by the binary variable separating the disk from the ring. ", + "bbox": [ + 174, + 847, + 823, + 888 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "But these variables ought to span the space of probabilities over the relevant variables, not just the variables themselves. Hence the next six variables are sines and cosines of smaller wavelength, which can encode probability distributions which are increasingly more precisely localized, down to a precision (wavelength) comparable with the diameter of the inner disk. Accordingly, the next two most relevant variables are the Cartesian coordinates inside the disk. This is followed by additional moments of the angle, down to a wavelength equal to the ring’s thickness, at which point we see the radius in the ring appear. ", + "bbox": [ + 174, + 895, + 821, + 922 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/139ed066324617d38572ae37b20db57c2963cbc5f2e6fcfe36fcf651198597fd.jpg", + "image_caption": [ + "Figure 4: Left: Best mean squared error for images reconstructed from the $k$ most relevant variables, as a function of $k$ (the latent dimension). This logarithmic plot shows that improvements stop once the dimension reaches 19 (where the two lines cross). Right: images produced by the generator from latent variables sampled according to the best Gaussian fit in latent space, for feature subspaces of dimensions 2, 8 and 19. " + ], + "image_footnote": [], + "bbox": [ + 189, + 98, + 810, + 284 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 404, + 825, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As mentioned, we see that the angle itself is not represented, likely because it is discontinuous. However, as shown also in Fig. 3, creating a gap in the ring allows for the angle to emerge as most relevant variable. This suggests that this approach may be able to automatically learn intrinsic coordinates of the latent variable manifold. ", + "bbox": [ + 174, + 481, + 825, + 536 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.4 INDEPENDENT VARIABLES AND GENERATIVE MODEL ", + "text_level": 1, + "bbox": [ + 173, + 563, + 580, + 577 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "If we postulate that the independent (or disentangled) relevant latent variables can be found in the linear span of the relevant variables, we can attempt to extract them by optimizing a neural network composed of two parts. Firstly, a linear layer maps the relevant variables to a small number of outputs (equal to the latent dimension). The purpose of this linear layer is to find the independent variables. These latent variables are then processed by an arbitrarily complex generative network to produce a possible value of the variable $X$ . As objective function, we may us an appropriate measure of similarity between the output and the data element from which the variables were obtained. ", + "bbox": [ + 174, + 602, + 825, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We tested this idea as follows. We took $X$ to consist of the MNIST digits, and produced $Y$ by randomly permuting neighboring pixels in the image, until the mean displacement per pixel is of order 1. In addition, we added independent Gaussian noise to the pixel values. (Hence the noise map $\\mathcal { N }$ simulates the coarse-graining channel introduced in Beny & Osborne (2015a)). ´ ", + "bbox": [ + 176, + 707, + 825, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As in the previous experiment, we do so to implement our intuition that the more relevant variables ought to be the ones which are of larger scale, or more robust to local perturbations. ", + "bbox": [ + 176, + 770, + 821, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The relevant variables of the clean images were produced by the same convolutional neural network as in Section 5.2, while the variables of the coarse-grained images were extracted by a network of the same geometry, but with half the number of filters and neurons. ", + "bbox": [ + 176, + 805, + 821, + 847 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We extracted the 1000 most relevant out of 1200 learned variables in this way. (The least relevant variables in this system happen to be highly dependent on the total number of variables and hence cannot be trusted to be correct). As a second step, we trained a linear layer coupled to a network composed of 5 fully-connected layers of 800 hidden neurons each. We refer to the number of output neurons in the first linear layer as the latent dimension. ", + "bbox": [ + 176, + 854, + 823, + 922 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As input, this network received the variables extracted from MNIST images using the above convolutional neural net (after it was fully trained using DCCA), and was trained to minimize the mean square error between its output and the original MNIST digit. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The resulting best mean square errors are shown in Fig. 4, as function of the latent dimension. Here we see a distinct change of polynomial scaling law at dimension 19. Increasing the dimension further provides no improvement. This behaviour is compatible with our hypothesis that the extra variables are just functions of those first twenty variables (functions which are effectively re-implemented by the generative network). ", + "bbox": [ + 174, + 152, + 823, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Images generated by sampling from a Gaussian approximation of the latent distribution for different latent dimensions are shown in Fig. 4. Below dimension 20, most generated image can be recognized as a specific digit. ", + "bbox": [ + 176, + 229, + 823, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 OUTLOOK ", + "text_level": 1, + "bbox": [ + 174, + 292, + 290, + 310 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We studied the classical (non-quantum) form of the theory introduced in Beny & Osborne (2013), ´ and found that the relevant observables of that theory are just the most correlated canonical variables in the sense of DCCA Andrew et al. (2013), and can be learned effectively using standard machine learning methods. ", + "bbox": [ + 174, + 325, + 823, + 382 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This point of views on DCCA provided us with several new insights. The first is that the learned relevant variables provide a useful representation of a joint probability distribution. We showed that performing inference using this representation can outperform crossentropy in predicting classes. Our experiments on halves of MNIST also show that the conditional distribution we obtain can effectively represent the uncertainty in the prediction of high-dimensional data. ", + "bbox": [ + 174, + 388, + 823, + 459 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A second insight relates to the interpretation of the canonical variables as spanning directions in the space of probability distributions. As suggested by the gaussian solutions and our experiment on synthetic data, we postulate that the canonical variables are functions of a small number of independent generators contained in their span. This hypothesis is supported by our experiment on MNIST, but further work is required to find a way to cleanly extract these variables. ", + "bbox": [ + 174, + 465, + 825, + 535 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We have yet to explore the potential applications of one of the salient aspect of this approach to inference, the fact that the canonical variables learned using DCCA are also those which can be most reliably predicted, irrespective of the value of the cutoff. To see why this is potentially significant, we observe that a central feature of scientific exploration is that we are not so concerned with making predictions about some given variables, as much as we are with discovering variables which can be predicted. ", + "bbox": [ + 174, + 542, + 825, + 626 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Another important feature of this approach is the fact that the resulting model allows for the direct evaluation of the expectation values in the posterior distribution without sampling. In particular this allows for the evaluation of credible intervals. Hence it should be especially suited to scientific applications where the ability to quantify uncertainty is essential. ", + "bbox": [ + 176, + 633, + 823, + 689 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, the relationship that we established with theory of quantum origin points towards a potential quantum generalization of DCCA that would apply to quantum data, or classical measurements of quantum systems. ", + "bbox": [ + 176, + 695, + 825, + 738 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 761, + 356, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We would like to thank Joel B ¨ eny and Raban Iten for helpful suggestions. We are also in- ´ debted to an anonymous ICLR2020 referee for pointing out the connection between our approach and DCCA. This work was supported by the National Research Foundation of Korea (NRF2018R1D1A1A02048436). ", + "bbox": [ + 174, + 792, + 825, + 848 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 871, + 285, + 886 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analysis. In International conference on machine learning, pp. 1247–1255, 2013. ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C. Beny and T. J. Osborne. Information geometric approach to the renormalisation group. ´ Phys. Rev. A, 92:022330, 2015a. doi: 10.1103/PhysRevA.92.022330. (arXiv:1206.7004). \nCedric B ´ eny. Coarse-grained distinguishability of field interactions. ´ Quantum, 2:67, 2018a. (arXiv:1509.03249). \nCedric B ´ eny. Inferring relevant features: from qft to pca. ´ International Journal of Quantum Information, 16:1840012, 2018b. (arXiv:1802.05756). \nCedric B ´ eny and Tobias J Osborne. Renormalisation as an inference problem. ´ (arXiv:1310.3188), 2013. \nCedric B ´ eny and Tobias J Osborne. The renormalisation group via statistical inference. ´ New J. Phys., 17:083005, 2015b. doi: 10.1088/1367-2630/17/8/083005. (arXiv:1402.4949). \nGK Eagleson. Polynomial expansions of bivariate distributions. The Annals of Mathematical Statistics, 35(3):1208–1215, 1964. \nCarl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211–218, 1936. \nIan Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014. \nKarol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. arXiv preprint arXiv:1310.8499, 2013. \nHarold Hotelling. Relations between two sets of variates. Biometrika, 28(3/4):321–377, 1936. \nMike Innes. Flux: Elegant machine learning with julia. Journal of Open Source Software, 2018. doi: 10.21105/joss.00602. \nRaban Iten, Tony Metger, Henrik Wilming, L´ıdia Del Rio, and Renato Renner. Discovering physical concepts with neural networks. (arXiv:1807.10300), 2018. \nDiederik P Kingma and Max Welling. Auto-encoding variational bayes. (arXiv:1312.6114), 2013. \nHO Lancaster. The structure of bivariate distributions. The Annals of Mathematical Statistics, 29 (3):719–736, 1958. \nTomer Michaeli, Weiran Wang, and Karen Livescu. Nonparametric canonical correlation analysis. In International Conference on Machine Learning, pp. 1967–1976, 2016. \nMehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. \nM. Ohya and D. Petz. Quantum entropy and its use. Springer Verlag, 2004. \nAmichai Painsky, Meir Feder, and Naftali Tishby. An information-theoretic framework for nonlinear canonical correlation analysis. arXiv preprint arXiv:1810.13259, 2018. \nWeiran Wang, Xinchen Yan, Honglak Lee, and Karen Livescu. Deep variational canonical correlation analysis. arXiv preprint arXiv:1610.03454, 2016. \nJunbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126, 2016. ", + "bbox": [ + 171, + 71, + 826, + 926 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A EXTRA INFORMATION ABOUT THE ALGORITHM ", + "text_level": 1, + "bbox": [ + 174, + 102, + 602, + 118 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.1 ALTERNATIVE INTERPRETATION OF THE OBJECTIVE ", + "text_level": 1, + "bbox": [ + 174, + 133, + 576, + 148 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "If we write $F _ { i j } : = f _ { j } ( x _ { i } )$ and $G _ { i j } : = g _ { j } ( y _ { i } )$ for the value of our variables on the dataset, then $\\begin{array} { r } { K = \\frac { 1 } { N } F ^ { \\top } F } \\end{array}$ , $\\begin{array} { r } { L = \\frac { 1 } { N } G ^ { \\top } G } \\end{array}$ and $\\begin{array} { r } { A = \\frac { 1 } { N } G ^ { \\top } F } \\end{array}$ . The DCCA objective can then be written as ", + "bbox": [ + 173, + 160, + 825, + 193 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/bb4208eb4b2d98539b5ca87ea715a86d6484f5c067aed48305d0f58fcc0ba1c9.jpg", + "text": "$$\n\\operatorname { T r } \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) = \\operatorname { T r } \\left( P Q \\right)\n$$", + "text_format": "latex", + "bbox": [ + 392, + 199, + 604, + 219 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $P = F ( F ^ { \\top } F ) ^ { - 1 } F ^ { \\top }$ and $Q = G ( G ^ { \\top } G ) ^ { - 1 } G ^ { \\top }$ are the projectors on the ranges of $F$ and $G$ respectively. Hence, we are maximizing the overlap between those ranges (which represent possible linear combinations of datapoints, respectively determined from variables of one or the other correlated views.) ", + "bbox": [ + 174, + 226, + 825, + 284 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.2 HEURISTIC ", + "text_level": 1, + "bbox": [ + 176, + 303, + 297, + 318 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Batch size—In our experiments, we observed that the batch size during training needs to be an order of magnitude larger than the number of variables (rank cutoff). When the batch size was too small, learning seemed to converge normally in terms of training and test loss, but resulted in variables which yield dramatically different losses when evaluated on larger batches, and yield spurious predictions. ", + "bbox": [ + 174, + 340, + 825, + 410 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Constant variables—The loss function $C$ takes value between 0 and $k _ { 0 } - 1$ because the constant variable always has relevance 1. The constant variable could be enforced a priori rather than learned, which, due to the objective, automatically forces the learned variables to have zero expectation values (be orthogonal to the constant variable). This might have advantages in certain circumstances, but in our experiments we found that this sometime hindered convergence. ", + "bbox": [ + 174, + 428, + 825, + 497 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Invertibility issues—The covariance matrices $K$ and $L$ can be ill-conditioned, potentially causing the gradient to “explode” because of the inverses $K ^ { - 1 }$ et $L ^ { - 1 }$ involved in the loss function. This can be avoided either by using the Moore-Penrose pseudo-inverse, or by replacing $K ^ { - 1 }$ by $( K + \\epsilon { \\bf 1 } ) ^ { - 1 }$ in the loss for some small positive number $\\epsilon$ , and likewise for $L ^ { - 1 }$ . ", + "bbox": [ + 174, + 515, + 825, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Symmetries in the loss function—The loss $C$ only depends on the span of the variables $f _ { i }$ and $g _ { j }$ , hence it has a very large group of symmetries. In particular, it is invariant under a change of the norm of each variable independently from each other. Because of that, it is preferable not to have a linear last layer. Using a hyperbolic tangent as last nonlinearity worked in our experiments. ", + "bbox": [ + 173, + 589, + 825, + 645 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Regularization—In all our tests, dropout had no beneficial effect. In fact, our objective seems to already provide a form of regularization, as shown in Section 5.2. ", + "bbox": [ + 174, + 662, + 823, + 690 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B THEORY IN MORE DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 713, + 431, + 729 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We consider two correlated random variables $X$ and $Y$ with a joint probability distribution $p ( x , y )$ . We assume that we are able to numerically evaluate expectations with respect to this distribution, for instance because we can sample from it. We want to use this ability in order to compute expectations with respect to the conditional distributions $p _ { X | Y } ( x | y ) \\ : = \\ : \\dot { p } ( x , y ) / p _ { X } ( x )$ and $p _ { Y | X } ( y | x ) ~ = ~ p ( x , y ) / p _ { Y } ( y )$ , where $\\begin{array} { r } { p _ { X } ( x ) \\ = \\ \\sum _ { y } p ( x , y ) } \\end{array}$ and $\\begin{array} { r } { p _ { Y } ( y ) \\ = \\ \\sum _ { x } p ( x , y ) } \\end{array}$ are the marginals of $p$ . Below we sometime remove the subscripts $X , X | Y$ or $Y | X$ if there is no ambiguity. ", + "bbox": [ + 173, + 744, + 825, + 848 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For instance, suppose we generated samples of $y$ given $x$ , through explicit knowledge of $p _ { Y \\mid X }$ . Then the evaluation of expectations with respect to $p _ { X | Y }$ is the subject of Bayesian inference. However, this is generally done in a context where the variable $X$ has low dimensionality and parameterizes a hand-crafted model. Our approach, however, is free of such a model and the variable $X$ can be of very high dimensionality. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B.1 INNER PRODUCT ON PROBABILITY VECTORS ", + "text_level": 1, + "bbox": [ + 178, + 103, + 522, + 117 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In order to define our strategy, we need to equip the spaces of probability distributions for $X$ and $Y$ with an inner product structure. Let us focus on $X$ , and assume that it takes discrete values to avoid unnecessary technicalities. The set of probability vectors is a convex subset of the real linear space $V _ { X } = \\mathbb { R } ^ { n }$ . Let us equip this space with the product ", + "bbox": [ + 173, + 128, + 825, + 185 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/1a12ff616c966e54cf1c282da79d92f2674b89326cb9bf7716455de6922806bb.jpg", + "text": "$$\n\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } : = \\sum _ { x } { \\frac { \\mu ( x ) \\mu ^ { \\prime } ( x ) } { p _ { X } ( x ) } }\n$$", + "text_format": "latex", + "bbox": [ + 406, + 193, + 591, + 231 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "for any $\\mu , \\mu ^ { \\prime } \\in V _ { X }$ . We also write $\\| \\mu \\| _ { X } ^ { 2 } = \\langle \\mu , \\mu \\rangle _ { X }$ . Importantly, this depends explicitly on the fixed probability vector $p _ { X } ( x )$ , which we took to be the marginal of $p ( x , y )$ . If $p _ { X }$ has full support, this makes $V _ { X }$ into a real inner product space. The same can be done for the variable $Y$ , yielding the inner product $\\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y }$ for $\\nu , \\nu ^ { \\prime } \\in V _ { Y }$ . ", + "bbox": [ + 173, + 238, + 825, + 295 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Had we interpreted $\\mu$ and $\\mu ^ { \\prime }$ as tangent vectors to $V _ { X }$ , considered as a manifold, this would be the Fisher information (Riemannian) metric, as in Beny & Osborne (2015b). But this quantity is also ´ meaningful for finite vectors: the induced norm distance between $p _ { X }$ and any probability vector $q$ is the $\\chi ^ { 2 }$ -divergence: ", + "bbox": [ + 173, + 301, + 825, + 357 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/11d45a6962684bf4a6b418780675e28df62328e758433c27c52a68b89d784428.jpg", + "text": "$$\n\\chi ^ { 2 } ( q , p _ { X } ) = \\langle q - p _ { X } , q - p _ { X } \\rangle _ { X } .\n$$", + "text_format": "latex", + "bbox": [ + 385, + 354, + 611, + 375 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The set of conditional probability distributions $p _ { Y \\mid X }$ form a stochastic map, i.e., a linear map $\\mathcal { N }$ $V _ { X } \\to V _ { Y }$ , $\\mu \\mapsto { \\mathcal { N } } ( \\mu )$ , where ", + "bbox": [ + 173, + 385, + 820, + 415 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/9f6aa5c7920c5d7ca7df6891ee1fa8eac0f726354cb2a190d7d57b440eed4689.jpg", + "text": "$$\n\\mathcal { N } ( \\mu ) ( y ) = \\sum _ { x } p _ { Y | X } ( y | x ) \\mu ( x )\n$$", + "text_format": "latex", + "bbox": [ + 392, + 421, + 606, + 455 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "for any $\\mu \\in V _ { X }$ . ", + "bbox": [ + 173, + 462, + 282, + 477 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is straightforward to check that the stochastic map $\\mathcal { N } ^ { * }$ defined by ", + "bbox": [ + 173, + 483, + 617, + 497 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/b4557f7afbfd59cd5d8971ce557d748cd9505101ca155f61defc18c9f1347d21.jpg", + "text": "$$\n\\mathcal { N ^ { * } } ( \\nu ) ( x ) = \\sum _ { x } p _ { X | Y } ( x | y ) \\nu ( x )\n$$", + "text_format": "latex", + "bbox": [ + 387, + 503, + 609, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "is the transpose $\\mathcal { N } ^ { * }$ of $\\mathcal { N }$ with respect to the inner products we defined (Ohya & Petz, 2004), i.e., for all $\\nu \\in V _ { Y }$ and $\\mu \\in V _ { X }$ , ", + "bbox": [ + 174, + 544, + 825, + 571 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/59de3bedea85948fa8b40b185bb2572dc7663c5393d2e0340dbd6abdad9dd49e.jpg", + "text": "$$\n\\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } = \\langle \\mathcal { N } ^ { * } ( \\nu ) , \\mu \\rangle _ { X } .\n$$", + "text_format": "latex", + "bbox": [ + 400, + 571, + 596, + 589 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Also, we observe that $\\mathcal { N } ( p _ { X } ) = p _ { Y }$ and $\\mathcal { N } ^ { * } ( p _ { Y } ) = p _ { X }$ . ", + "bbox": [ + 173, + 592, + 545, + 608 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.2 EIGEN-RELEVANCE DECOMPOSITION", + "text_level": 1, + "bbox": [ + 176, + 623, + 472, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We can use the inner products on $V _ { X }$ and $V _ { Y }$ to define a singular value decomposition of the stochastic map $\\mathcal { N }$ . That is, there is an orthonormal family $u _ { 1 } , \\ldots , u _ { k }$ of $V _ { X }$ and an orthonormal family $v _ { 1 } , \\ldots , v _ { k }$ of $V _ { Y }$ , such that ", + "bbox": [ + 174, + 650, + 825, + 691 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/8b1490f43f66bb1116b078ef858053ca3c89a396338ce45339fb6e4f50da5ac8.jpg", + "text": "$$\n\\begin{array} { r } { \\mathcal { N } ( u _ { j } ) = \\eta _ { j } v _ { j } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 446, + 690, + 550, + 708 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "for $j = 1 , \\dotsc , k$ . For each $j , \\eta _ { j }$ is a singular value of $\\mathcal { N }$ , whose square we call the relevance of the vector $v _ { j }$ . Moreover $\\eta _ { j } \\in [ 0 , 1 ]$ since the $\\chi ^ { 2 }$ divergence is contractive under any stochastic map. Given that ${ \\ddot { \\mathcal { N } } } ^ { * }$ is the transpose of $\\mathcal { N }$ : ", + "bbox": [ + 173, + 712, + 825, + 756 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/7bdb6bc1c44d9193a0074f170fe1194569b52d796fb680d65a6fafc14cfb139d.jpg", + "text": "$$\n\\begin{array} { r } { \\mathcal { N } ^ { * } ( v _ { j } ) = \\eta _ { j } u _ { j } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 442, + 762, + 555, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Equivalently, $u _ { j }$ is an eigenvector of $\\mathcal { N } ^ { \\ast } \\circ \\mathcal { N }$ and $v _ { j }$ is an eigenvectors of $\\mathcal { N } \\circ \\mathcal { N } ^ { \\ast }$ , both with eigenvalue $\\eta _ { j } ^ { 2 }$ . ", + "bbox": [ + 174, + 786, + 825, + 818 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Because $\\mathcal { N }$ maps $p _ { X }$ to $p _ { Y }$ , we always have the dual eigenvectors $u _ { 0 } ~ = ~ p _ { X }$ and $v _ { 0 } = p _ { Y }$ with eigenvalue 1. ", + "bbox": [ + 173, + 823, + 825, + 852 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.3 LOW-RANK APPROXIMATION ", + "text_level": 1, + "bbox": [ + 176, + 869, + 416, + 883 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Typically, the dimension $k$ of the space of probabilities is more than astronomically large. For instance, if the values of $X$ consists of small 256 gray level images of $2 8 \\times 2 8$ pixels, then $k =$ ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$2 5 6 ^ { 2 8 ^ { 2 } } \\simeq 1 0 ^ { 1 8 8 8 }$ . However, in many case, only very few of these dimensions may be relevant for the purpose of inferring other variables. ", + "bbox": [ + 171, + 102, + 823, + 132 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The core of our approach is to approximate $\\mathcal { N }$ and $\\mathcal { N } ^ { * }$ by restricting them to the span of the first $k _ { 0 }$ eigenvectors $u _ { j }$ and $v _ { j }$ with largest singular values $\\eta _ { j }$ . That is, if we order the singular values $\\eta _ { j }$ , $j = 1 , \\dots , k$ in decreasing order, we propose to use the approximations ", + "bbox": [ + 173, + 138, + 825, + 181 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/f3a5327242bb5b1d075b4f3e55813df2cb07b711bc8ac260eb13179d35ab16b0.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle \\mathcal { N } _ { 0 } ( \\mu ) = \\sum _ { j \\leq k _ { 0 } } \\eta _ { j } \\langle u _ { j } , \\mu \\rangle _ { X } v _ { j } } } \\\\ { { \\displaystyle \\mathcal { N } _ { 0 } ^ { * } ( \\nu ) = \\sum _ { j \\leq k _ { 0 } } \\eta _ { j } \\langle v _ { j } , \\nu \\rangle _ { Y } u _ { j } } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 403, + 184, + 594, + 257 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "to $\\mathcal { N }$ and $\\mathcal { N } ^ { * }$ respectively, for some $k _ { 0 }$ typically much smaller than $k$ , and any $\\mu \\in V _ { X }$ , $\\nu \\in V _ { Y }$ . ", + "bbox": [ + 171, + 276, + 805, + 292 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We denote the components of $\\mathcal { N } _ { 0 }$ and $\\mathcal { N } _ { 0 } ^ { \\ast }$ by $q ( y | x )$ and $q ( x | y )$ , e.g., ", + "bbox": [ + 174, + 296, + 629, + 314 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/a6e2b9880b1b37ead6fd6a4ffe84590a7907071edbd961c9b9d70841b48801b8.jpg", + "text": "$$\n{ \\mathcal { N } } _ { 0 } ( \\mu ) ( y ) = \\sum _ { x } q ( y | x ) \\mu ( x ) .\n$$", + "text_format": "latex", + "bbox": [ + 400, + 316, + 598, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Since $\\mathcal { N } _ { 0 }$ and $\\mathcal { N } _ { 0 } ^ { \\ast }$ are adjoint, we can define $q ( x , y ) = q ( x | y ) p _ { Y } ( y ) = q ( y | x ) p _ { X } ( x )$ . Although the marginals of $\\dot { \\boldsymbol { q } } ( \\boldsymbol { x } , \\boldsymbol { y } )$ are the probability distributions $p _ { X }$ and $p _ { Y }$ , the numbers $q ( x , y )$ are not necessarily positive. ", + "bbox": [ + 176, + 352, + 823, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The quality of this approximation for a given $k _ { 0 }$ does not directly depend on the dimensionality of $X$ and $Y$ , but only on the amount of correlations between the two variables. Our aim is to use a $k _ { 0 }$ small enough that the components of $\\mathcal { N } _ { 0 }$ and $\\mathcal { N } _ { 0 } ^ { \\ast }$ can be computed explicitly. ", + "bbox": [ + 176, + 401, + 825, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Theorem 1. $\\mathcal { N } _ { 0 }$ is the map of rank $k _ { 0 }$ which minimizes the average distance ", + "bbox": [ + 174, + 446, + 676, + 460 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/df8d7f7d7a094fb6a4ec97675e9fa6183150df7e30f9c93a72ddb88e92010910.jpg", + "text": "$$\n\\sum _ { x } p ( x ) \\| \\mathcal { N } _ { 0 } ( \\delta _ { x } ) - \\mathcal { N } ( \\delta _ { x } ) \\| _ { Y } ^ { 2 } = \\sum _ { x y } \\frac { ( q ( x , y ) - p ( x , y ) ) ^ { 2 } } { p ( x ) p ( y ) } .\n$$", + "text_format": "latex", + "bbox": [ + 302, + 463, + 694, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proof. The low rank approximation $\\mathcal { N } _ { 0 }$ minimizes the distance $\\| \\mathcal { N } _ { 0 } - \\mathcal { N } \\| _ { \\mathrm { F } }$ where ", + "bbox": [ + 173, + 517, + 718, + 534 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/253fa3db9ec0fe002e085ab83d9ded0c2df946910a937c7ba475785f06910eb1.jpg", + "text": "$$\n\\| \\mathcal { M } \\| _ { F } ^ { 2 } = \\mathrm { T r } \\left( \\mathcal { M } ^ { * } \\mathcal { M } \\right)\n$$", + "text_format": "latex", + "bbox": [ + 423, + 536, + 575, + 554 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "is the Hilbert-Schmidt (or Frobenius) norm (Eckart & Young, 1936). This follows from the fact that this is also the $l ^ { 2 }$ -norm of the vector of singular values of $\\mathcal { M }$ . Let us find the explicit form of the trace. Each possible value $x$ of the variable $X$ is associated with a probability distribution $\\delta _ { x } ( y ) = 1$ when $x = y$ and zero otherwise. These distributions form an orthogonal basis of $V _ { X }$ , and have norms $\\langle \\delta _ { x } , \\delta _ { x } \\rangle = 1 / p _ { X } ( x )$ . Therefore, ", + "bbox": [ + 173, + 556, + 825, + 627 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/6617f39c0c7d039718aac0cf6cf3a2830a2819f9724871e62ef353ed978bdd14.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\operatorname { T r } \\left( \\mathcal { M ^ { * } M } \\right) = \\sum _ { x } p _ { X } ( x ) \\langle \\delta _ { x } , \\mathcal { M ^ { * } M } ( \\delta _ { x } ) \\rangle _ { Y } } \\\\ { \\displaystyle \\quad = \\sum _ { x } p _ { X } ( x ) \\| \\mathcal { M } ( \\delta _ { x } ) \\| _ { Y } ^ { 2 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 349, + 631, + 647, + 698 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B.4 RELEVANT VARIABLES ", + "text_level": 1, + "bbox": [ + 174, + 731, + 377, + 744 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We express the elements $\\mu \\in \\ V _ { X }$ and $\\nu \\in \\ V _ { Y }$ in terms of the marginals $p _ { X }$ and $p _ { Y }$ as simple products: ", + "bbox": [ + 169, + 756, + 825, + 784 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/f0603455ab679d7dc3d9eef5e9809c2db30e86572bece69d545acc50922839b5.jpg", + "text": "$$\n\\mu ( x ) = p _ { X } ( x ) f ( x ) \\quad { \\mathrm { a n d } } \\quad \\nu ( y ) = p _ { Y } ( y ) g ( y )\n$$", + "text_format": "latex", + "bbox": [ + 343, + 782, + 655, + 800 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "for all $x , y$ , where $f$ and $g$ are real functions of $x$ and $y$ . ", + "bbox": [ + 173, + 801, + 539, + 815 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The inner products then simply become correlations among variables. Using also $\\mu ^ { \\prime } = p _ { X } f ^ { \\prime }$ and $\\nu ^ { \\prime } = p _ { Y } g ^ { \\prime }$ , we obtain ", + "bbox": [ + 173, + 821, + 825, + 851 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/f7cda9f3540d0af30f97eeb8acb938d9f6f74bc7df4d8f25db6a203aa534caed.jpg", + "text": "$$\n\\begin{array} { l } { { \\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } = \\displaystyle \\sum _ { x } p _ { X } ( x ) f ( x ) f ^ { \\prime } ( x ) = \\overline { { { f f ^ { \\prime } } } } , } } \\\\ { { \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } = \\displaystyle \\sum _ { y } p _ { Y } ( y ) g ( y ) g ^ { \\prime } ( y ) = \\overline { { { g g ^ { \\prime } } } } . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 364, + 853, + 632, + 922 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "These are simple expectation values with respect to $p$ , which we assumed is the type of quantity we can evaluate for arbitrary functions $f , f ^ { \\prime } , g , \\bar { g ^ { \\prime } }$ . ", + "bbox": [ + 173, + 103, + 823, + 133 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Since $\\mathcal { N } ^ { * } \\mathcal { N }$ is self-adjoint in terms of this inner product, its eigenvectors $u _ { i }$ are orthogonal, and hence the corresponding variables $a _ { i }$ defined by ${ u } _ { i } \\bar { ( x ) } = p _ { X } ( x ) \\bar { a _ { i } } ( x )$ are uncorrelated. Indeed, ", + "bbox": [ + 171, + 138, + 823, + 167 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/e9e69f6a4cec03a846afcb9709a37a4da5c1127809ce22499507efcdd41c32bb.jpg", + "text": "$$\n\\overline { { a _ { i } a _ { j } } } = \\langle u _ { i } , u _ { j } \\rangle _ { X } = 0 ,\n$$", + "text_format": "latex", + "bbox": [ + 421, + 174, + 575, + 193 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "for all $i , j$ . Moreover, accounting for the eigenvector $u _ { 0 } = p _ { X }$ (corresponding to the constant feature $a _ { 0 } ( x ) = 1$ for all $x$ ), ", + "bbox": [ + 173, + 199, + 826, + 227 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/16dffdd84bef144f0869194d1bd7081036059f9c15ef42b3ce9bcebd8ccfc950.jpg", + "text": "$$\n\\overline { { a } } _ { i } = 0\n$$", + "text_format": "latex", + "bbox": [ + 473, + 227, + 524, + 241 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "for all $i \\neq 0$ . Hence we trivially have ", + "bbox": [ + 173, + 246, + 419, + 261 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/1314457dd98ce6123c16959fd0d98f53893544881218a7696aa40d6a71ec0f7f.jpg", + "text": "$$\n\\overline { { a _ { i } a _ { j } } } = \\overline { { a } } _ { i } \\overline { { a } } _ { j }\n$$", + "text_format": "latex", + "bbox": [ + 455, + 261, + 540, + 276 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "for all $i , j \\neq 0$ . ", + "bbox": [ + 173, + 279, + 274, + 295 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Likewise for the eigenvectors of $\\mathcal { N N } ^ { * }$ . If $v _ { i } ( y ) = p _ { Y } ( y ) b _ { i } ( y )$ : ", + "bbox": [ + 173, + 299, + 589, + 316 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/64b88123c6d74a35b1c5dc9c829e55ebd9514f524ae9e97af52afe2fde456f6f.jpg", + "text": "$$\n\\overline { { b _ { i } b _ { j } } } = \\langle v _ { i } , v _ { j } \\rangle _ { Y } = 0 = \\overline { { b } } _ { i } \\overline { { b } } _ { j } .\n$$", + "text_format": "latex", + "bbox": [ + 401, + 323, + 596, + 342 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "for all $i , j \\neq 0$ . ", + "bbox": [ + 173, + 348, + 274, + 363 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Importantly, this does not mean that the variables $u _ { 1 } , u _ { 2 } , \\ldots$ nor $v _ { 1 } , v _ { 2 } , \\ldots$ are “disentangled”, i.e., they are not statistically independent. These variables represent components in the space of probability vectors, rather than the “sample” space. They should be understood as spanning a subspace of the space of functions over the relevant independent variables. We discuss this in more detail in Section 5.3. ", + "bbox": [ + 174, + 369, + 825, + 439 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B.5 CORNERS OF $\\mathcal { N }$ AND LOSS FUNCTION ", + "text_level": 1, + "bbox": [ + 176, + 457, + 477, + 472 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The final piece of puzzle we need, is the ability to express the components (corners) of $\\mathcal { N }$ and $\\mathcal { N } ^ { * }$ in the span of possible non-orthogonal families of variables. ", + "bbox": [ + 174, + 483, + 820, + 512 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Let us therefore consider two arbitrary families $f _ { 1 } , \\ldots , f _ { k _ { 0 } }$ and $g _ { 1 } , \\ldots , g _ { k _ { 0 } }$ of variables, which respectively represent the vectors $p _ { X } f _ { j } \\in V _ { X }$ and $p _ { Y } g _ { j } \\in V _ { Y }$ . ", + "bbox": [ + 176, + 518, + 820, + 549 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Firstly, we need matrices representing the components of the inner products on $V _ { X }$ and $V _ { Y }$ . Those are the symmetric matrices ", + "bbox": [ + 171, + 553, + 823, + 582 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/e3b0b84063132afb696e79be3654f58285ebf5ac5071aecd466e0c04ac8ad0dd.jpg", + "text": "$$\n\\begin{array} { r } { K _ { i j } = \\langle p _ { X } f _ { i } , p _ { X } f _ { j } \\rangle _ { X } = \\overline { { f _ { i } f _ { j } } } , } \\\\ { L _ { i j } = \\langle p _ { Y } g _ { i } , p _ { Y } g _ { j } \\rangle _ { Y } = \\overline { { g _ { i } g _ { j } } } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 393, + 588, + 601, + 628 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The components $N _ { i j }$ of $\\mathcal { N }$ are defined by ", + "bbox": [ + 176, + 640, + 447, + 655 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/2b7b8a822b8382b09ad5f7ef0f60b06484b822d4bcf9a16b5df140c5972bf464.jpg", + "text": "$$\n\\mathcal { N } ( p _ { X } f _ { j } ) = \\sum _ { i } N _ { i j } p _ { Y } g _ { i } .\n$$", + "text_format": "latex", + "bbox": [ + 408, + 661, + 588, + 695 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Taking the inner product with $p _ { Y } g _ { k }$ , we obtain ", + "bbox": [ + 176, + 702, + 480, + 717 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/b2186206361d77232b516411f4d298cafe0f44cb95c38a9d96ca5dc21515641c.jpg", + "text": "$$\n\\langle p _ { Y } g _ { k } , \\mathcal { N } ( p _ { X } f _ { j } ) \\rangle = \\sum _ { i } N _ { i j } L _ { k i } .\n$$", + "text_format": "latex", + "bbox": [ + 387, + 723, + 611, + 757 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The left-hand side can be computed using Equ. 13. It is the matrix ", + "bbox": [ + 173, + 763, + 607, + 779 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/9deaf3f2b4a503f27c9b2f2fc081d2425e38207c6b07dc4513908e2dfe0753c2.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle { \\cal A } _ { k j } = \\langle p _ { Y } g _ { k } , { \\cal N } ( p _ { X } f _ { j } ) \\rangle } } \\\\ { { \\displaystyle ~ = \\sum _ { x , y } \\frac { p _ { Y } ( y ) g _ { k } ( y ) p _ { Y | X } ( y | x ) p _ { X } ( x ) f _ { j } ( x ) } { p _ { Y } ( y ) } } } \\\\ { { \\displaystyle ~ = \\sum _ { x , y } p ( x , y ) g _ { k } ( y ) f _ { j } ( x ) = \\overline { { g _ { k } f _ { j } } } . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 343, + 784, + 653, + 881 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Therefore, in matrix notation, Equ. (34) is $A = L N$ , or ", + "bbox": [ + 173, + 886, + 537, + 901 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/ae48dea95919ca0dca6569335f749fd8f58bd0eaa7faf859df0a382270205e86.jpg", + "text": "$$\nN = L ^ { - 1 } A .\n$$", + "text_format": "latex", + "bbox": [ + 455, + 907, + 540, + 922 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The components $N _ { i j } ^ { * }$ of $\\mathcal { N } ^ { * }$ are obtained by just swapping $X$ and $Y$ , yielding ", + "bbox": [ + 173, + 102, + 681, + 119 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/8607cc6b778302cefea952b8aa87f1d995b39f936a8c08a3448aa169ed9bbd5d.jpg", + "text": "$$\nN ^ { * } = K ^ { - 1 } A ^ { \\top } .\n$$", + "text_format": "latex", + "bbox": [ + 444, + 130, + 553, + 147 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Hence the singular values of the corner of $\\mathcal { N }$ defined by the variables $f _ { j }$ and $g _ { j }$ are just the squareroot of the eigenvalues of the matrix $N ^ { * } N = K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A$ . In order to find the variables $f _ { j }$ and $g _ { j }$ with the same span as the first $k _ { 0 }$ eigenvectors $u _ { j } , v _ { j }$ , we just need to maximize all the eigenvalues of $N ^ { * } N$ . A simple way to do this is to use (minus) the trace of $N ^ { * } N$ as loss function, since it is the sum of the square of the singular values. We call $\\mathrm { T r } \\left( N ^ { * } N \\right)$ the relevance of the subspaces defines by the variables $f _ { j }$ ad $g _ { i }$ for all $i , j$ . This yields the loss/cost function: ", + "bbox": [ + 173, + 165, + 825, + 253 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/b4e2488c74a3f265625ac742d2edbb99eb5b739f7810f5d57a05baee1221ccf8.jpg", + "text": "$$\nC = k _ { 0 } - \\operatorname { T r } \\left( N ^ { \\ast } N \\right) = k _ { 0 } - \\operatorname { T r } \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 331, + 262, + 666, + 282 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Once optimal variables have been found, one can obtain the components of the eigenvectors in the span of $f _ { 1 } , \\ldots , f _ { k _ { 0 } }$ through standard numerical diagonalization of $N ^ { * } N$ . ", + "bbox": [ + 174, + 299, + 825, + 329 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "B.6 INFERENCE ", + "text_level": 1, + "bbox": [ + 174, + 348, + 299, + 363 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The variables minimizing $C$ can be used to infer one variable from the other. For instance, given $y$ , the inferred probability distribution over $x$ is given by $p _ { X | Y } ( x | y ) = \\mathcal { N } ^ { * } ( \\delta _ { y } ) ( x )$ , where $\\delta _ { y } ( \\bar { y } ^ { \\prime } )$ is 1 when $y = y ^ { \\prime }$ and zero otherwise. In order to compute this, we first need the components of the distribution $\\delta _ { y }$ in terms of the family $p _ { Y } g _ { 1 } , \\ldots , p _ { Y } g _ { k _ { 0 } }$ , i.e., the real numbers $( \\delta _ { y } ) _ { j }$ such that ", + "bbox": [ + 173, + 376, + 825, + 435 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/f872f27d4090ce21e2aad0a936530e9e6025e37a3b62a65bec1564c77eff912d.jpg", + "text": "$$\n\\delta _ { y } ( y ^ { \\prime } ) = p _ { Y } ( y ^ { \\prime } ) \\sum _ { i = 1 } ^ { k _ { 0 } } ( \\delta _ { y } ) _ { i } g _ { i } ( y ^ { \\prime } ) + r ( y ^ { \\prime } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 364, + 444, + 632, + 488 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "where $\\langle r , p _ { Y } \\delta _ { i } \\rangle _ { Y } = 0$ for all $i$ . Taking the inner product with $p _ { Y } g _ { j }$ , we obtain ", + "bbox": [ + 173, + 497, + 686, + 513 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/975c3a3c2adecafe908e86c2a9af738588c69578219d8b9bdf6fafdfe8f06ccb.jpg", + "text": "$$\n\\langle p _ { Y } g _ { j } , \\delta _ { y } \\rangle _ { Y } = \\sum _ { i = 1 } ^ { k _ { 0 } } ( \\delta _ { y } ) _ { i } L _ { j i } ,\n$$", + "text_format": "latex", + "bbox": [ + 403, + 523, + 593, + 568 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "where the left hand side is also just ", + "bbox": [ + 174, + 577, + 405, + 590 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/1ed107ef27991b943bccefb6b6bbba58532efbd7417b6e102276a318fbafa637.jpg", + "text": "$$\n\\langle p _ { Y } g _ { j } , \\delta _ { y } \\rangle _ { Y } = g _ { j } ( y ) .\n$$", + "text_format": "latex", + "bbox": [ + 424, + 602, + 573, + 619 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Therefore the components of $\\delta _ { y }$ are explicitly ", + "bbox": [ + 174, + 628, + 473, + 643 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/cfc6c6eba7b5b121c7de8a15d7862c781c1896c64d26b9e202ec11c8c4ad5386.jpg", + "text": "$$\n( \\delta _ { y } ) _ { i } = \\sum _ { j } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) .\n$$", + "text_format": "latex", + "bbox": [ + 411, + 654, + 588, + 689 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "It follows that ", + "bbox": [ + 173, + 705, + 267, + 719 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/1ea01b53b365947e50f1df1e28eada3c55875a450cc944384f38c51dcd92f6c4.jpg", + "text": "$$\n\\begin{array} { l } { { \\displaystyle p _ { X | Y } ( x | y ) = \\mathcal { N } ^ { * } ( \\delta _ { y } ) ( x ) \\approx \\mathcal { N } _ { 0 } ^ { * } ( \\delta _ { y } ) ( x ) } } \\\\ { { \\displaystyle ~ = \\sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) f _ { k } ( x ) . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 357, + 726, + 638, + 784 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Then, for instance, the expected inferred value of $X$ is ", + "bbox": [ + 173, + 799, + 531, + 814 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/ede1edc95774bb1a48d73bb19789c9232db331bf80900a374ee8cf177b94faca.jpg", + "text": "$$\n\\overline { { { x } } } = \\sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) \\sum _ { x } p _ { X } ( x ) x f _ { k } ( x ) .\n$$", + "text_format": "latex", + "bbox": [ + 349, + 823, + 648, + 859 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "For the inference of $Y$ from $x$ , we have ", + "bbox": [ + 173, + 869, + 434, + 883 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/a6d8ff69ee7efce2d0c01dcd31e4bc7d29a0fe27c24506780c53243d50eb0f5a.jpg", + "text": "$$\np _ { Y | X } ( y | x ) \\approx \\sum _ { i j k } N _ { k i } ( K ^ { - 1 } ) _ { i j } f _ { j } ( x ) g _ { k } ( y ) .\n$$", + "text_format": "latex", + "bbox": [ + 356, + 892, + 640, + 928 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C ANALYTICAL EXAMPLE ", + "text_level": 1, + "bbox": [ + 176, + 102, + 406, + 118 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "When $p ( x , y )$ is any multivariate Gaussian distribution, everything can be computed analytically. Let us consider here the one-dimensional case. We use $p ( x ) \\ \\propto \\ \\exp { \\left( - x ^ { 2 } / 2 \\tau ^ { 2 } \\right) }$ , and the conditional $\\underline { { { p } } } ( y | x ) ~ \\propto ~ \\exp \\left( - ( y - x ) ^ { 2 } / 2 \\sigma ^ { 2 } \\right)$ . That is, $y$ is equal to $x$ but with some added Gaussian noise. This gives ", + "bbox": [ + 173, + 132, + 825, + 193 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/10899a4858ab6715bf0ec506b7e356d22489e9c4630de49317621359c2d5ac92.jpg", + "text": "$$\np _ { X | Y } ( x | y ) \\propto \\exp \\left( - \\frac { ( x - \\gamma y ) ^ { 2 } } { 2 \\tau ^ { 2 } ( 1 - \\gamma ) } \\right) , \\quad \\mathrm { w h e r e } \\quad \\gamma = \\frac { \\tau ^ { 2 } } { \\sigma ^ { 2 } + \\tau ^ { 2 } } .\n$$", + "text_format": "latex", + "bbox": [ + 287, + 190, + 710, + 227 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "It was show in Lancaster (1958), that the most relevant subspace of dimension $k _ { 0 }$ on the variable $X$ is simply spanned by the variables ", + "bbox": [ + 171, + 236, + 821, + 263 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/73b73808bbc9159c5a966035aa4e8cbaac31db30f53dc1a63c210274fc805b87.jpg", + "text": "$$\nf _ { n } ( x ) = x ^ { n } ,\n$$", + "text_format": "latex", + "bbox": [ + 454, + 262, + 542, + 280 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "$n = 0 , \\ldots , k _ { 0 } - 1$ . Similarly for $Y$ ; ", + "bbox": [ + 173, + 281, + 411, + 297 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/8b45b2b5538fea1629bcb0b9df88672de2464ff4e3b331fa2b06daea2fd0cc03.jpg", + "text": "$$\ng _ { n } ( y ) = y ^ { n } .\n$$", + "text_format": "latex", + "bbox": [ + 455, + 295, + 542, + 313 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "This independence of the relevant variables on the detailed parameters of $p$ is a general property of Gaussian joint distributions. ", + "bbox": [ + 173, + 315, + 825, + 343 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "This means, for instance, that the most relevant feature $( n = 1$ ) for predicting the value of $X$ given $Y = y$ is simply $Y$ itself. The higher order variables have to do with inferring extra aspects of the probability distribution over $X$ . ", + "bbox": [ + 174, + 351, + 825, + 392 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "A set of orthogonal variables can be obtain from the Gram-Schmidt procedure, which, if done from small to large $n$ much necessarily yield the eigenvectors $u _ { n }$ and $v _ { n }$ . For illustration purpose, let us work with the non-orthogonal vectors $f _ { n }$ and $g _ { n }$ , keeping only the first $k _ { 0 } = 3$ vectors. ", + "bbox": [ + 174, + 398, + 825, + 443 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The three matrices (correlators) we need can be easily computed: ", + "bbox": [ + 173, + 448, + 601, + 463 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/21630e4089ab2ac003986e1aea6a1512c08f2f3a4e83cf67d0632043d2fe4ed8.jpg", + "text": "$$\n\\begin{array} { l c c } { { K = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } } } & { { 0 } } & { { 3 \\tau ^ { 4 } } } \\end{array} \\right) } } & { { L = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { 3 ( \\tau ^ { 2 } + \\sigma ^ { 2 } ) ^ { 2 } } } \\end{array} \\right) } } \\\\ { { A = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { \\tau ^ { 2 } ( \\sigma ^ { 2 } + 3 \\tau ^ { 2 } ) } } \\end{array} \\right) . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 196, + 469, + 774, + 571 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We obtain ", + "bbox": [ + 173, + 575, + 243, + 590 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/f535fc90744ea97d7a26b415efbbb17733ce6c77e7858e4a81860e170aef0734.jpg", + "text": "$$\nM = K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } ( 1 - \\gamma ^ { 2 } ) } } \\\\ { { 0 } } & { { \\gamma } } & { { 0 } } \\\\ { { 0 } } & { { 0 } } & { { \\gamma ^ { 2 } } } \\end{array} \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 338, + 585, + 660, + 636 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The eigenvalues of $M$ can be read on the diagonal, and the corresponding eigenvectors are $( 1 , 0 , 0 )$ , $( 0 , 1 , \\bar { 0 } )$ and $( - \\tau ^ { 2 } , 0 , 1 )$ , which means that the eigenfunctions are in order $\\bar { u } _ { 0 } ( x ) = 1$ , $u _ { 1 } ( x ) = x$ and u2(x) = x2 − τ 2. ", + "bbox": [ + 173, + 638, + 825, + 680 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Because we are working with continuous variables, the true rank of $\\mathcal { N }$ is infinite, even for any finite cutoff on the singular values. Nevertheless, it is instructive to see how the approximate inference fares for rank $k _ { 0 } = 3$ . Given the value $y$ for $Y$ , the inferred distribution over $X$ is ", + "bbox": [ + 174, + 686, + 825, + 729 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/8d97ff77c3cad717d7dd4af9483f1195bb46a057f2b481d6ef2627afbc141356.jpg", + "text": "$$\n\\mathcal { N } _ { 0 } ^ { * } ( \\delta _ { y } ) ( x ) = p _ { X } ( x ) p _ { Y } ( y ) \\sum _ { j , k = 0 } ^ { 2 } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } ) _ { k j } y ^ { j } x ^ { k } .\n$$", + "text_format": "latex", + "bbox": [ + 315, + 736, + 681, + 781 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The approximately inferred first and second moments of $X$ is given by integrating the above times $x$ (resp. $x ^ { 2 }$ ) over $x$ . We obtain ", + "bbox": [ + 173, + 786, + 825, + 815 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/82ea07d3179fe2084324edd3855c2cce89f193120aa41b2a4e445fa3416ed17f.jpg", + "text": "$$\n{ \\overline { { x } } } = \\gamma y \\quad { \\mathrm { a n d } } \\quad { \\overline { { x ^ { 2 } } } } = \\gamma ^ { 2 } y ^ { 2 } + ( 1 - \\gamma ) \\tau ^ { 2 } ,\n$$", + "text_format": "latex", + "bbox": [ + 361, + 821, + 633, + 840 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "which are actually exact: they are equal to the first two moments of $X$ over $p _ { X | Y }$ as given in Eq. (46). ", + "bbox": [ + 171, + 847, + 823, + 863 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "In fact, it is easy to see that this would be true for the first $k _ { 0 } - 1$ moments had we kept the $k _ { 0 }$ most relevant variables. ", + "bbox": [ + 174, + 868, + 825, + 898 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/parse/train/SJeS16EKPr/SJeS16EKPr_middle.json b/parse/train/SJeS16EKPr/SJeS16EKPr_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..738091b64aa37a64b9156979a65bd56c4dc5d426 --- /dev/null +++ b/parse/train/SJeS16EKPr/SJeS16EKPr_middle.json @@ -0,0 +1,60007 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 503, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 504, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 504, + 97 + ], + "score": 1.0, + "content": "LEARNING RELEVANT FEATURES FOR STATISTICAL IN-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 101, + 176, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 176, + 118 + ], + "score": 1.0, + "content": "FERENCE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 245, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 354 + ], + "lines": [ + { + "bbox": [ + 142, + 212, + 469, + 224 + ], + "spans": [ + { + "bbox": [ + 142, + 212, + 469, + 224 + ], + "score": 1.0, + "content": "Given two views of data, we consider the problem of finding the features of one", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "score": 1.0, + "content": "view which can be most faithfully inferred from the other. We find that these are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "score": 1.0, + "content": "also the most correlated variables in the sense of deep canonical correlation anal-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 257 + ], + "score": 1.0, + "content": "ysis (DCCA). Moreover, we show that these variables can be used to construct", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 267 + ], + "score": 1.0, + "content": "a non-parametric representation of the implied joint probability distribution. This", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "score": 1.0, + "content": "representation can be used to compute the expectations of functions over one view", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "score": 1.0, + "content": "of data conditioned on the other, such as Bayesian estimators and their standard", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 301 + ], + "score": 1.0, + "content": "deviations. We test the approach using inference on occluded MNIST images, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 469, + 311 + ], + "score": 1.0, + "content": "show that our representation contains multiple modes. Surprisingly, when applied", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "to supervised learning (one dataset consists of labels), this approach automatically", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "provides regularization and faster convergence compared to the cross-entropy ob-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 332, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 470, + 345 + ], + "score": 1.0, + "content": "jective. We also explore using this approach to discover salient independent vari-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 242, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 242, + 356 + ], + "score": 1.0, + "content": "ables of a single dataset.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 377, + 206, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 402, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 170, + 416 + ], + "score": 1.0, + "content": "Given samples", + "type": "text" + }, + { + "bbox": [ + 171, + 402, + 258, + 415 + ], + "score": 0.93, + "content": "( x _ { 1 } , y _ { 1 } ) , \\dotsc , ( x _ { n } , y _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 401, + 456, + 416 + ], + "score": 1.0, + "content": "from an unknown joint probability distribution", + "type": "text" + }, + { + "bbox": [ + 456, + 402, + 485, + 414 + ], + "score": 0.91, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 401, + 506, + 416 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 397, + 426 + ], + "score": 1.0, + "content": "want to construct a useful representation of the conditional probabilities", + "type": "text" + }, + { + "bbox": [ + 398, + 414, + 425, + 426 + ], + "score": 0.93, + "content": "p ( x | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 413, + 444, + 426 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 444, + 414, + 471, + 426 + ], + "score": 0.92, + "content": "p ( y | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 413, + 506, + 426 + ], + "score": 1.0, + "content": ", so that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 424, + 340, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 340, + 437 + ], + "score": 1.0, + "content": "we that we can infer one view from the other on new data.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 502, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 159, + 455 + ], + "score": 1.0, + "content": "For instance,", + "type": "text" + }, + { + "bbox": [ + 159, + 444, + 167, + 451 + ], + "score": 0.73, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 440, + 183, + 455 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 184, + 443, + 190, + 453 + ], + "score": 0.78, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 440, + 504, + 455 + ], + "score": 1.0, + "content": "could be past and future histories of dynamical data, visual and auditory inputs,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 453, + 223, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 223, + 464 + ], + "score": 1.0, + "content": "actions and their effects, etc.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "Following the approach introduced in Beny & Osborne (2013; 2015b) in the context of quantum ´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 458, + 493 + ], + "score": 1.0, + "content": "information theory, we look at the problem as follows: the conditional distributions", + "type": "text" + }, + { + "bbox": [ + 458, + 480, + 486, + 492 + ], + "score": 0.92, + "content": "p ( y | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "can", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "be thought of as representing a noisy communication channel (stochastic map). This channel is a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 502, + 504, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 504, + 515 + ], + "score": 1.0, + "content": "linear map between spaces of typically ludicrously large dimensions (the spaces of all probability", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 512, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 180, + 527 + ], + "score": 1.0, + "content": "distributions over", + "type": "text" + }, + { + "bbox": [ + 181, + 516, + 188, + 523 + ], + "score": 0.68, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 512, + 201, + 527 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 201, + 515, + 209, + 525 + ], + "score": 0.72, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 512, + 505, + 527 + ], + "score": 1.0, + "content": "). We want a pair of small subspaces which best represent the channel.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 430, + 538 + ], + "score": 1.0, + "content": "Specifically, we look for those vectors representing probability distributions over", + "type": "text" + }, + { + "bbox": [ + 430, + 526, + 437, + 534 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "which lose least", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 533, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 443, + 549 + ], + "score": 1.0, + "content": "distinguishability under the channel, where the distinguishability is measured by the", + "type": "text" + }, + { + "bbox": [ + 444, + 534, + 456, + 547 + ], + "score": 0.9, + "content": "\\chi ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 533, + 505, + 549 + ], + "score": 1.0, + "content": "divergence.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "We show in Section 3 that this is equivalent to performing a certain singular value decomposition", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "score": 1.0, + "content": "of the channel (seen as an operator in Hilbert space) and keep only the components with the largest", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "singular values. Moreover, the full singular value decomposition is equivalent to the decomposition", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 584, + 393, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 393, + 598 + ], + "score": 1.0, + "content": "in terms of canonical variables introduced in Lancaster (1958), namely,", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 602, + 371, + 637 + ], + "lines": [ + { + "bbox": [ + 239, + 602, + 371, + 637 + ], + "spans": [ + { + "bbox": [ + 239, + 602, + 371, + 637 + ], + "score": 0.95, + "content": "p ( \\boldsymbol { y } | \\boldsymbol { x } ) = p ( \\boldsymbol { y } ) \\sum _ { i = 1 } ^ { D } \\eta _ { i } u _ { i } ( \\boldsymbol { x } ) v _ { i } ( \\boldsymbol { y } ) ,", + "type": "interline_equation", + "image_path": "cd0226efb4f314de95eef92b4975cd11169f564dba4f7e7b395e743af0ef5d3d.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 602, + 371, + 619.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 239, + 619.5, + 371, + 637.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 642, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 134, + 656 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 645, + 144, + 654 + ], + "score": 0.83, + "content": "u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 642, + 164, + 656 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 645, + 175, + 655 + ], + "score": 0.84, + "content": "u _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 642, + 335, + 656 + ], + "score": 1.0, + "content": "are real non-linear functions such that", + "type": "text" + }, + { + "bbox": [ + 336, + 643, + 397, + 655 + ], + "score": 0.89, + "content": "\\mathbb { E } ( u _ { i } u _ { j } ) = \\delta _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 642, + 402, + 656 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 403, + 643, + 463, + 655 + ], + "score": 0.91, + "content": "\\mathbb { E } ( v _ { i } v _ { j } ) = \\delta _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 642, + 486, + 656 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 486, + 643, + 505, + 654 + ], + "score": 0.83, + "content": "0 ~ <", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 654, + 304, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 208, + 666 + ], + "score": 0.9, + "content": "\\eta _ { D } \\leq \\cdot \\cdot \\cdot \\leq \\eta _ { 1 } \\leq \\eta _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 654, + 304, + 666 + ], + "score": 1.0, + "content": "are the singular values.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 670, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "The practical advantage of this representation for inference (or prediction) is that it reduces the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "evaluation of conditional expectations to that of empirical averages over the (unconditional) marginal", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 691, + 131, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 126, + 705 + ], + "score": 0.9, + "content": "p ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 691, + 131, + 707 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 354, + 722 + ], + "score": 1.0, + "content": "As observed in Michaeli et al. (2016), the span of the first", + "type": "text" + }, + { + "bbox": [ + 354, + 710, + 361, + 720 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 709, + 444, + 722 + ], + "score": 1.0, + "content": "canonical variables", + "type": "text" + }, + { + "bbox": [ + 444, + 711, + 470, + 722 + ], + "score": 0.29, + "content": "u _ { i } , \\ v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "is what", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "is learned by the deep canonical correlation analysis (DCCA) (Andrew et al., 2013). Indeed, these", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 503, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 504, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 504, + 97 + ], + "score": 1.0, + "content": "LEARNING RELEVANT FEATURES FOR STATISTICAL IN-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 101, + 176, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 176, + 118 + ], + "score": 1.0, + "content": "FERENCE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 245, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 111, + 135, + 245, + 158 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 354 + ], + "lines": [ + { + "bbox": [ + 142, + 212, + 469, + 224 + ], + "spans": [ + { + "bbox": [ + 142, + 212, + 469, + 224 + ], + "score": 1.0, + "content": "Given two views of data, we consider the problem of finding the features of one", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "score": 1.0, + "content": "view which can be most faithfully inferred from the other. We find that these are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 246 + ], + "score": 1.0, + "content": "also the most correlated variables in the sense of deep canonical correlation anal-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 257 + ], + "score": 1.0, + "content": "ysis (DCCA). Moreover, we show that these variables can be used to construct", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 469, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 267 + ], + "score": 1.0, + "content": "a non-parametric representation of the implied joint probability distribution. This", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "score": 1.0, + "content": "representation can be used to compute the expectations of functions over one view", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "score": 1.0, + "content": "of data conditioned on the other, such as Bayesian estimators and their standard", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 301 + ], + "score": 1.0, + "content": "deviations. We test the approach using inference on occluded MNIST images, and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 469, + 311 + ], + "score": 1.0, + "content": "show that our representation contains multiple modes. Surprisingly, when applied", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "to supervised learning (one dataset consists of labels), this approach automatically", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "provides regularization and faster convergence compared to the cross-entropy ob-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 332, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 470, + 345 + ], + "score": 1.0, + "content": "jective. We also explore using this approach to discover salient independent vari-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 343, + 242, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 242, + 356 + ], + "score": 1.0, + "content": "ables of a single dataset.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11, + "bbox_fs": [ + 141, + 212, + 470, + 356 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 377, + 206, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 208, + 392 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 402, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 170, + 416 + ], + "score": 1.0, + "content": "Given samples", + "type": "text" + }, + { + "bbox": [ + 171, + 402, + 258, + 415 + ], + "score": 0.93, + "content": "( x _ { 1 } , y _ { 1 } ) , \\dotsc , ( x _ { n } , y _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 401, + 456, + 416 + ], + "score": 1.0, + "content": "from an unknown joint probability distribution", + "type": "text" + }, + { + "bbox": [ + 456, + 402, + 485, + 414 + ], + "score": 0.91, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 401, + 506, + 416 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 397, + 426 + ], + "score": 1.0, + "content": "want to construct a useful representation of the conditional probabilities", + "type": "text" + }, + { + "bbox": [ + 398, + 414, + 425, + 426 + ], + "score": 0.93, + "content": "p ( x | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 413, + 444, + 426 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 444, + 414, + 471, + 426 + ], + "score": 0.92, + "content": "p ( y | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 413, + 506, + 426 + ], + "score": 1.0, + "content": ", so that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 424, + 340, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 340, + 437 + ], + "score": 1.0, + "content": "we that we can infer one view from the other on new data.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 401, + 506, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 502, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 159, + 455 + ], + "score": 1.0, + "content": "For instance,", + "type": "text" + }, + { + "bbox": [ + 159, + 444, + 167, + 451 + ], + "score": 0.73, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 440, + 183, + 455 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 184, + 443, + 190, + 453 + ], + "score": 0.78, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 440, + 504, + 455 + ], + "score": 1.0, + "content": "could be past and future histories of dynamical data, visual and auditory inputs,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 453, + 223, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 223, + 464 + ], + "score": 1.0, + "content": "actions and their effects, etc.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 440, + 504, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "Following the approach introduced in Beny & Osborne (2013; 2015b) in the context of quantum ´", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 458, + 493 + ], + "score": 1.0, + "content": "information theory, we look at the problem as follows: the conditional distributions", + "type": "text" + }, + { + "bbox": [ + 458, + 480, + 486, + 492 + ], + "score": 0.92, + "content": "p ( y | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "can", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "be thought of as representing a noisy communication channel (stochastic map). This channel is a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 502, + 504, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 504, + 515 + ], + "score": 1.0, + "content": "linear map between spaces of typically ludicrously large dimensions (the spaces of all probability", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 512, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 180, + 527 + ], + "score": 1.0, + "content": "distributions over", + "type": "text" + }, + { + "bbox": [ + 181, + 516, + 188, + 523 + ], + "score": 0.68, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 512, + 201, + 527 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 201, + 515, + 209, + 525 + ], + "score": 0.72, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 512, + 505, + 527 + ], + "score": 1.0, + "content": "). We want a pair of small subspaces which best represent the channel.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 430, + 538 + ], + "score": 1.0, + "content": "Specifically, we look for those vectors representing probability distributions over", + "type": "text" + }, + { + "bbox": [ + 430, + 526, + 437, + 534 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "which lose least", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 533, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 443, + 549 + ], + "score": 1.0, + "content": "distinguishability under the channel, where the distinguishability is measured by the", + "type": "text" + }, + { + "bbox": [ + 444, + 534, + 456, + 547 + ], + "score": 0.9, + "content": "\\chi ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 533, + 505, + 549 + ], + "score": 1.0, + "content": "divergence.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 469, + 506, + 549 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "We show in Section 3 that this is equivalent to performing a certain singular value decomposition", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "score": 1.0, + "content": "of the channel (seen as an operator in Hilbert space) and keep only the components with the largest", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "singular values. Moreover, the full singular value decomposition is equivalent to the decomposition", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 584, + 393, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 393, + 598 + ], + "score": 1.0, + "content": "in terms of canonical variables introduced in Lancaster (1958), namely,", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 551, + 506, + 598 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 602, + 371, + 637 + ], + "lines": [ + { + "bbox": [ + 239, + 602, + 371, + 637 + ], + "spans": [ + { + "bbox": [ + 239, + 602, + 371, + 637 + ], + "score": 0.95, + "content": "p ( \\boldsymbol { y } | \\boldsymbol { x } ) = p ( \\boldsymbol { y } ) \\sum _ { i = 1 } ^ { D } \\eta _ { i } u _ { i } ( \\boldsymbol { x } ) v _ { i } ( \\boldsymbol { y } ) ,", + "type": "interline_equation", + "image_path": "cd0226efb4f314de95eef92b4975cd11169f564dba4f7e7b395e743af0ef5d3d.jpg" + } + ] + } + ], + "index": 35.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 602, + 371, + 619.5 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 239, + 619.5, + 371, + 637.0 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 642, + 504, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 134, + 656 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 645, + 144, + 654 + ], + "score": 0.83, + "content": "u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 642, + 164, + 656 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 645, + 175, + 655 + ], + "score": 0.84, + "content": "u _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 642, + 335, + 656 + ], + "score": 1.0, + "content": "are real non-linear functions such that", + "type": "text" + }, + { + "bbox": [ + 336, + 643, + 397, + 655 + ], + "score": 0.89, + "content": "\\mathbb { E } ( u _ { i } u _ { j } ) = \\delta _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 642, + 402, + 656 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 403, + 643, + 463, + 655 + ], + "score": 0.91, + "content": "\\mathbb { E } ( v _ { i } v _ { j } ) = \\delta _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 642, + 486, + 656 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 486, + 643, + 505, + 654 + ], + "score": 0.83, + "content": "0 ~ <", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 654, + 304, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 208, + 666 + ], + "score": 0.9, + "content": "\\eta _ { D } \\leq \\cdot \\cdot \\cdot \\leq \\eta _ { 1 } \\leq \\eta _ { 0 } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 654, + 304, + 666 + ], + "score": 1.0, + "content": "are the singular values.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 642, + 505, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 670, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "The practical advantage of this representation for inference (or prediction) is that it reduces the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "evaluation of conditional expectations to that of empirical averages over the (unconditional) marginal", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 691, + 131, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 126, + 705 + ], + "score": 0.9, + "content": "p ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 691, + 131, + 707 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 671, + 505, + 707 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 354, + 722 + ], + "score": 1.0, + "content": "As observed in Michaeli et al. (2016), the span of the first", + "type": "text" + }, + { + "bbox": [ + 354, + 710, + 361, + 720 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 709, + 444, + 722 + ], + "score": 1.0, + "content": "canonical variables", + "type": "text" + }, + { + "bbox": [ + 444, + 711, + 470, + 722 + ], + "score": 0.29, + "content": "u _ { i } , \\ v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "is what", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "is learned by the deep canonical correlation analysis (DCCA) (Andrew et al., 2013). Indeed, these", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 311, + 95 + ], + "score": 1.0, + "content": "variables are those which maximize the correlations", + "type": "text" + }, + { + "bbox": [ + 312, + 82, + 344, + 95 + ], + "score": 0.93, + "content": "\\mathbb { E } ( u _ { i } v _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "subject to the same constraints as above.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 92, + 386, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 291, + 107 + ], + "score": 1.0, + "content": "(This reduces to CCA (Hotelling, 1936) when", + "type": "text" + }, + { + "bbox": [ + 292, + 96, + 302, + 105 + ], + "score": 0.74, + "content": "u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 92, + 306, + 107 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 306, + 96, + 316, + 106 + ], + "score": 0.8, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 92, + 386, + 107 + ], + "score": 1.0, + "content": "are linear maps).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "In this work, besides establishing this new information-theoretical interpretation of canonical vari-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 133 + ], + "score": 1.0, + "content": "ables and DCCA, we experiment with using this representation for performing inference on new", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "data. Moreover, we propose a strategy for extracting disentangled variables from the canonical", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 276, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 276, + 155 + ], + "score": 1.0, + "content": "variables, inspired by analytical solutions.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 108, + 171, + 209, + 184 + ], + "lines": [ + { + "bbox": [ + 104, + 169, + 210, + 186 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 210, + 186 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 196, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "This general problem (of building an effective representation of the conditional probability distribu-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 207, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 221 + ], + "score": 1.0, + "content": "tions implied by joint samples) covers many existing approaches in different contexts. For instance,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 168, + 231 + ], + "score": 1.0, + "content": "if the variables", + "type": "text" + }, + { + "bbox": [ + 169, + 220, + 176, + 230 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "has few possible states, then it reduces to a classification problem, usually solved", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "score": 1.0, + "content": "by minimizing the crossentropy between a predicted distribution and the one-hot encoding of the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 140, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 140, + 253 + ], + "score": 1.0, + "content": "classes.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 257, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 133, + 270 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 133, + 259, + 140, + 269 + ], + "score": 0.74, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "has a large number of states, or is fundamentally continuous, existing approaches usually", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 267, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 505, + 281 + ], + "score": 1.0, + "content": "do not model the whole conditional distribution, but either provide the average (regression), or", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 280, + 229, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 229, + 291 + ], + "score": 1.0, + "content": "approximately sample from it.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 296, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "The main class of methods which allows for sampling from the conditional distributions are varia-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "tional: a deterministic neural networks produces the parameters of analytical classes of probabilities.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "This includes variational autoencoders Kingma & Welling (2013) (e.g., Iten et al. (2018); Wang et al.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "score": 1.0, + "content": "(2016)), and approaches based on the minimum description length principle such as Gregor et al.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 339, + 139, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 139, + 353 + ], + "score": 1.0, + "content": "(2013).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 371 + ], + "score": 1.0, + "content": "Alternatively, it may also be possible to use adversarial training Goodfellow et al. (2014), by using", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 368, + 476, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 476, + 380 + ], + "score": 1.0, + "content": "a conditional Mirza & Osindero (2014) version of an energy-based GAN Zhao et al. (2016).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "By contrast, our approach doesn’t require the training of a generative model. Instead, conditional", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "expectations are constructed as linear combinations of unconditional empirical averages over the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 406, + 163, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 163, + 420 + ], + "score": 1.0, + "content": "training data.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "Previous approaches to equipping CCA or DCCA with an information-theoretic interpretation have", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "explored different directions. For instance, in Wang et al. (2016), the authors generalize a probabilis-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 446, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 504, + 457 + ], + "score": 1.0, + "content": "tic interpretation for CCA in terms of gaussian distributions, which leads to a variational approach.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "In Painsky et al. (2018), additional constraints on the mutual information between the data and the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 468, + 301, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 301, + 480 + ], + "score": 1.0, + "content": "learned variables are added to the optimizations.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "A previous attempt at designing a numerical solution for our singular value problem can be found in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "Beny (2018b). In that work, the relevant variables were represented through a PCA kernel produced ´", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 506, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 297, + 519 + ], + "score": 1.0, + "content": "by Monte Carlo sampling, but wasn’t practical.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 108, + 534, + 169, + 547 + ], + "lines": [ + { + "bbox": [ + 104, + 531, + 172, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 172, + 550 + ], + "score": 1.0, + "content": "3 THEORY", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "We formalize the problem by assuming that our data was sampled from an unknown joint distribution", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 571, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 107, + 571, + 136, + 583 + ], + "score": 0.92, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 571, + 245, + 582 + ], + "score": 1.0, + "content": "over two random variables", + "type": "text" + }, + { + "bbox": [ + 246, + 571, + 256, + 581 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 571, + 273, + 582 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 274, + 571, + 283, + 581 + ], + "score": 0.81, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 571, + 287, + 582 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 122, + 600 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 588, + 137, + 599 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 587, + 156, + 600 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 588, + 170, + 599 + ], + "score": 0.88, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 587, + 451, + 600 + ], + "score": 1.0, + "content": "denote the linear spaces spanned by all probability distributions over", + "type": "text" + }, + { + "bbox": [ + 452, + 588, + 462, + 597 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 587, + 480, + 600 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 480, + 588, + 490, + 597 + ], + "score": 0.81, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "re-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 235, + 612 + ], + "score": 1.0, + "content": "spectively. Here we assume that", + "type": "text" + }, + { + "bbox": [ + 235, + 599, + 245, + 609 + ], + "score": 0.85, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 598, + 262, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 263, + 599, + 272, + 609 + ], + "score": 0.79, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "take finitely many values for simplicity, but this formalism", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 609, + 397, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 397, + 623 + ], + "score": 1.0, + "content": "can be straightforwardly extended to infinite-dimensional vector spaces.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 231, + 639 + ], + "score": 1.0, + "content": "We will need inner products on", + "type": "text" + }, + { + "bbox": [ + 231, + 627, + 245, + 637 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 625, + 263, + 639 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 263, + 627, + 277, + 637 + ], + "score": 0.89, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 625, + 505, + 639 + ], + "score": 1.0, + "content": ", to make them into real Hilbert spaces. We use the Fisher", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 282, + 651 + ], + "score": 1.0, + "content": "information metrics evaluated at the points", + "type": "text" + }, + { + "bbox": [ + 283, + 637, + 302, + 650 + ], + "score": 0.92, + "content": "p ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 637, + 321, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 322, + 637, + 341, + 650 + ], + "score": 0.91, + "content": "p ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 637, + 450, + 651 + ], + "score": 1.0, + "content": "respectively (marginals of", + "type": "text" + }, + { + "bbox": [ + 451, + 637, + 480, + 650 + ], + "score": 0.91, + "content": "p ( x , y ) \\mathrm { . }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "), that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 647, + 120, + 662 + ], + "spans": [ + { + "bbox": [ + 104, + 647, + 120, + 662 + ], + "score": 1.0, + "content": "is,", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 178, + 657, + 433, + 689 + ], + "lines": [ + { + "bbox": [ + 178, + 657, + 433, + 689 + ], + "spans": [ + { + "bbox": [ + 178, + 657, + 433, + 689 + ], + "score": 0.92, + "content": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } : = \\sum _ { x } { \\frac { \\mu ( x ) \\mu ^ { \\prime } ( x ) } { p ( x ) } } \\quad { \\mathrm { a n d } } \\quad \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } : = \\sum _ { y } { \\frac { \\nu ( y ) \\nu ^ { \\prime } ( y ) } { p ( y ) } }", + "type": "interline_equation", + "image_path": "8d970395cca7301e1fc810391ed22d29995147be6897298796ca3b5d2b948d9a.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 178, + 657, + 433, + 667.6666666666666 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 178, + 667.6666666666666, + 433, + 678.3333333333333 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 178, + 678.3333333333333, + 433, + 688.9999999999999 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 692, + 279, + 705 + ], + "lines": [ + { + "bbox": [ + 106, + 692, + 280, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 168, + 706 + ], + "score": 1.0, + "content": "for any vectors", + "type": "text" + }, + { + "bbox": [ + 169, + 693, + 214, + 704 + ], + "score": 0.92, + "content": "\\mu , \\mu ^ { \\prime } \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 692, + 233, + 706 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 233, + 692, + 276, + 704 + ], + "score": 0.91, + "content": "\\nu , \\nu ^ { \\prime } \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 692, + 280, + 706 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 240, + 723 + ], + "score": 1.0, + "content": "Below we also call the marginals", + "type": "text" + }, + { + "bbox": [ + 241, + 712, + 254, + 721 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 707, + 272, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 273, + 712, + 286, + 721 + ], + "score": 0.85, + "content": "p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 707, + 463, + 723 + ], + "score": 1.0, + "content": "respectively when omitting their arguments", + "type": "text" + }, + { + "bbox": [ + 464, + 709, + 506, + 722 + ], + "score": 0.88, + "content": "( p _ { X } ( x ) \\equiv", + "type": "inline_equation" + } + ], + "index": 46 + }, + { + "bbox": [ + 107, + 720, + 209, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 126, + 732 + ], + "score": 0.91, + "content": "p ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 720, + 144, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 144, + 720, + 203, + 733 + ], + "score": 0.93, + "content": "p _ { Y } ( y ) \\equiv p ( y ) \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 720, + 209, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 311, + 95 + ], + "score": 1.0, + "content": "variables are those which maximize the correlations", + "type": "text" + }, + { + "bbox": [ + 312, + 82, + 344, + 95 + ], + "score": 0.93, + "content": "\\mathbb { E } ( u _ { i } v _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "subject to the same constraints as above.", + "type": "text" + } + ], + "index": 0, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 92, + 386, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 291, + 107 + ], + "score": 1.0, + "content": "(This reduces to CCA (Hotelling, 1936) when", + "type": "text" + }, + { + "bbox": [ + 292, + 96, + 302, + 105 + ], + "score": 0.74, + "content": "u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 92, + 306, + 107 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 306, + 96, + 316, + 106 + ], + "score": 0.8, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 92, + 386, + 107 + ], + "score": 1.0, + "content": "are linear maps).", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 107 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "In this work, besides establishing this new information-theoretical interpretation of canonical vari-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 133 + ], + "score": 1.0, + "content": "ables and DCCA, we experiment with using this representation for performing inference on new", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "data. Moreover, we propose a strategy for extracting disentangled variables from the canonical", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 276, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 276, + 155 + ], + "score": 1.0, + "content": "variables, inspired by analytical solutions.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 111, + 506, + 155 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 171, + 209, + 184 + ], + "lines": [ + { + "bbox": [ + 104, + 169, + 210, + 186 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 210, + 186 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 196, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "This general problem (of building an effective representation of the conditional probability distribu-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 207, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 221 + ], + "score": 1.0, + "content": "tions implied by joint samples) covers many existing approaches in different contexts. For instance,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 168, + 231 + ], + "score": 1.0, + "content": "if the variables", + "type": "text" + }, + { + "bbox": [ + 169, + 220, + 176, + 230 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "has few possible states, then it reduces to a classification problem, usually solved", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 243 + ], + "score": 1.0, + "content": "by minimizing the crossentropy between a predicted distribution and the one-hot encoding of the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 140, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 140, + 253 + ], + "score": 1.0, + "content": "classes.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 196, + 505, + 253 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 257, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 133, + 270 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 133, + 259, + 140, + 269 + ], + "score": 0.74, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "has a large number of states, or is fundamentally continuous, existing approaches usually", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 267, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 505, + 281 + ], + "score": 1.0, + "content": "do not model the whole conditional distribution, but either provide the average (regression), or", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 280, + 229, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 229, + 291 + ], + "score": 1.0, + "content": "approximately sample from it.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 257, + 505, + 291 + ] + }, + { + "type": "list", + "bbox": [ + 107, + 296, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "The main class of methods which allows for sampling from the conditional distributions are varia-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "tional: a deterministic neural networks produces the parameters of analytical classes of probabilities.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 331 + ], + "score": 1.0, + "content": "This includes variational autoencoders Kingma & Welling (2013) (e.g., Iten et al. (2018); Wang et al.", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "score": 1.0, + "content": "(2016)), and approaches based on the minimum description length principle such as Gregor et al.", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 339, + 139, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 139, + 353 + ], + "score": 1.0, + "content": "(2013).", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 17, + "bbox_fs": [ + 105, + 295, + 505, + 353 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 504, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 371 + ], + "score": 1.0, + "content": "Alternatively, it may also be possible to use adversarial training Goodfellow et al. (2014), by using", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 368, + 476, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 476, + 380 + ], + "score": 1.0, + "content": "a conditional Mirza & Osindero (2014) version of an energy-based GAN Zhao et al. (2016).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 356, + 505, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "By contrast, our approach doesn’t require the training of a generative model. Instead, conditional", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "expectations are constructed as linear combinations of unconditional empirical averages over the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 406, + 163, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 163, + 420 + ], + "score": 1.0, + "content": "training data.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 384, + 505, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "Previous approaches to equipping CCA or DCCA with an information-theoretic interpretation have", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "explored different directions. For instance, in Wang et al. (2016), the authors generalize a probabilis-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 446, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 504, + 457 + ], + "score": 1.0, + "content": "tic interpretation for CCA in terms of gaussian distributions, which leads to a variational approach.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "In Painsky et al. (2018), additional constraints on the mutual information between the data and the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 468, + 301, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 301, + 480 + ], + "score": 1.0, + "content": "learned variables are added to the optimizations.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 424, + 505, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "A previous attempt at designing a numerical solution for our singular value problem can be found in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "Beny (2018b). In that work, the relevant variables were represented through a PCA kernel produced ´", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 506, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 297, + 519 + ], + "score": 1.0, + "content": "by Monte Carlo sampling, but wasn’t practical.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 484, + 505, + 519 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 534, + 169, + 547 + ], + "lines": [ + { + "bbox": [ + 104, + 531, + 172, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 531, + 172, + 550 + ], + "score": 1.0, + "content": "3 THEORY", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "We formalize the problem by assuming that our data was sampled from an unknown joint distribution", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 571, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 107, + 571, + 136, + 583 + ], + "score": 0.92, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 571, + 245, + 582 + ], + "score": 1.0, + "content": "over two random variables", + "type": "text" + }, + { + "bbox": [ + 246, + 571, + 256, + 581 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 571, + 273, + 582 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 274, + 571, + 283, + 581 + ], + "score": 0.81, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 571, + 287, + 582 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 559, + 505, + 583 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 122, + 600 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 588, + 137, + 599 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 587, + 156, + 600 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 588, + 170, + 599 + ], + "score": 0.88, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 587, + 451, + 600 + ], + "score": 1.0, + "content": "denote the linear spaces spanned by all probability distributions over", + "type": "text" + }, + { + "bbox": [ + 452, + 588, + 462, + 597 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 587, + 480, + 600 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 480, + 588, + 490, + 597 + ], + "score": 0.81, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "re-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 235, + 612 + ], + "score": 1.0, + "content": "spectively. Here we assume that", + "type": "text" + }, + { + "bbox": [ + 235, + 599, + 245, + 609 + ], + "score": 0.85, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 598, + 262, + 612 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 263, + 599, + 272, + 609 + ], + "score": 0.79, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "take finitely many values for simplicity, but this formalism", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 609, + 397, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 397, + 623 + ], + "score": 1.0, + "content": "can be straightforwardly extended to infinite-dimensional vector spaces.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 587, + 505, + 623 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 231, + 639 + ], + "score": 1.0, + "content": "We will need inner products on", + "type": "text" + }, + { + "bbox": [ + 231, + 627, + 245, + 637 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 625, + 263, + 639 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 263, + 627, + 277, + 637 + ], + "score": 0.89, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 625, + 505, + 639 + ], + "score": 1.0, + "content": ", to make them into real Hilbert spaces. We use the Fisher", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 282, + 651 + ], + "score": 1.0, + "content": "information metrics evaluated at the points", + "type": "text" + }, + { + "bbox": [ + 283, + 637, + 302, + 650 + ], + "score": 0.92, + "content": "p ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 637, + 321, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 322, + 637, + 341, + 650 + ], + "score": 0.91, + "content": "p ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 637, + 450, + 651 + ], + "score": 1.0, + "content": "respectively (marginals of", + "type": "text" + }, + { + "bbox": [ + 451, + 637, + 480, + 650 + ], + "score": 0.91, + "content": "p ( x , y ) \\mathrm { . }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "), that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 647, + 120, + 662 + ], + "spans": [ + { + "bbox": [ + 104, + 647, + 120, + 662 + ], + "score": 1.0, + "content": "is,", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 625, + 506, + 662 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 178, + 657, + 433, + 689 + ], + "lines": [ + { + "bbox": [ + 178, + 657, + 433, + 689 + ], + "spans": [ + { + "bbox": [ + 178, + 657, + 433, + 689 + ], + "score": 0.92, + "content": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } : = \\sum _ { x } { \\frac { \\mu ( x ) \\mu ^ { \\prime } ( x ) } { p ( x ) } } \\quad { \\mathrm { a n d } } \\quad \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } : = \\sum _ { y } { \\frac { \\nu ( y ) \\nu ^ { \\prime } ( y ) } { p ( y ) } }", + "type": "interline_equation", + "image_path": "8d970395cca7301e1fc810391ed22d29995147be6897298796ca3b5d2b948d9a.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 178, + 657, + 433, + 667.6666666666666 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 178, + 667.6666666666666, + 433, + 678.3333333333333 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 178, + 678.3333333333333, + 433, + 688.9999999999999 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 692, + 279, + 705 + ], + "lines": [ + { + "bbox": [ + 106, + 692, + 280, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 168, + 706 + ], + "score": 1.0, + "content": "for any vectors", + "type": "text" + }, + { + "bbox": [ + 169, + 693, + 214, + 704 + ], + "score": 0.92, + "content": "\\mu , \\mu ^ { \\prime } \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 692, + 233, + 706 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 233, + 692, + 276, + 704 + ], + "score": 0.91, + "content": "\\nu , \\nu ^ { \\prime } \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 692, + 280, + 706 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 692, + 280, + 706 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 240, + 723 + ], + "score": 1.0, + "content": "Below we also call the marginals", + "type": "text" + }, + { + "bbox": [ + 241, + 712, + 254, + 721 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 707, + 272, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 273, + 712, + 286, + 721 + ], + "score": 0.85, + "content": "p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 707, + 463, + 723 + ], + "score": 1.0, + "content": "respectively when omitting their arguments", + "type": "text" + }, + { + "bbox": [ + 464, + 709, + 506, + 722 + ], + "score": 0.88, + "content": "( p _ { X } ( x ) \\equiv", + "type": "inline_equation" + } + ], + "index": 46 + }, + { + "bbox": [ + 107, + 720, + 209, + 734 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 126, + 732 + ], + "score": 0.91, + "content": "p ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 720, + 144, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 144, + 720, + 203, + 733 + ], + "score": 0.93, + "content": "p _ { Y } ( y ) \\equiv p ( y ) \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 720, + 209, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 707, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 342, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 342, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 280, + 96 + ], + "score": 1.0, + "content": "These inner products allow us to define the", + "type": "text" + }, + { + "bbox": [ + 280, + 82, + 292, + 94 + ], + "score": 0.9, + "content": "\\chi ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 81, + 342, + 96 + ], + "score": 1.0, + "content": "divergence:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 100, + 374, + 115 + ], + "lines": [ + { + "bbox": [ + 236, + 100, + 374, + 115 + ], + "spans": [ + { + "bbox": [ + 236, + 100, + 374, + 115 + ], + "score": 0.92, + "content": "\\chi ^ { 2 } ( q , p _ { X } ) = \\langle q - p _ { X } , q - p _ { X } \\rangle _ { X } ,", + "type": "interline_equation", + "image_path": "94f8cac2c398495c9f82111acc830c8d5618725abd8916dddbcaf9f6f22579a4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 236, + 100, + 374, + 115 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 121, + 505, + 166 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 340, + 135 + ], + "score": 1.0, + "content": "which a measures of statistical distinguishability between", + "type": "text" + }, + { + "bbox": [ + 340, + 124, + 347, + 133 + ], + "score": 0.81, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 120, + 365, + 135 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 365, + 123, + 378, + 133 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 120, + 506, + 135 + ], + "score": 1.0, + "content": ". Specifically, it quantifies how", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 315, + 145 + ], + "score": 1.0, + "content": "easy it is to reject the null hypothesis that the state is", + "type": "text" + }, + { + "bbox": [ + 315, + 134, + 329, + 144 + ], + "score": 0.88, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 132, + 403, + 145 + ], + "score": 1.0, + "content": "when it is actually", + "type": "text" + }, + { + "bbox": [ + 404, + 134, + 410, + 144 + ], + "score": 0.79, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 132, + 505, + 145 + ], + "score": 1.0, + "content": ", based on the empirical", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "distribution obtained from independent samples. It is also the lowest order approximation of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 153, + 226, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 226, + 167 + ], + "score": 1.0, + "content": "Kullback-Leibler divergence.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 170, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 195, + 184 + ], + "score": 1.0, + "content": "The joint distribution", + "type": "text" + }, + { + "bbox": [ + 195, + 171, + 224, + 183 + ], + "score": 0.93, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 170, + 353, + 184 + ], + "score": 1.0, + "content": "yields conditional distributions", + "type": "text" + }, + { + "bbox": [ + 353, + 171, + 380, + 183 + ], + "score": 0.92, + "content": "p ( y | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 170, + 399, + 184 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 400, + 171, + 427, + 183 + ], + "score": 0.91, + "content": "p ( x | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 170, + 506, + 184 + ], + "score": 1.0, + "content": ". These can be un-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 181, + 503, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 349, + 194 + ], + "score": 1.0, + "content": "derstood as the components (or kernels) of stochastic maps", + "type": "text" + }, + { + "bbox": [ + 349, + 183, + 415, + 193 + ], + "score": 0.89, + "content": "\\mathcal { N } : V _ { X } V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 181, + 434, + 194 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 434, + 182, + 503, + 193 + ], + "score": 0.9, + "content": "\\mathcal { N ^ { * } } : V _ { X } V _ { Y }", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 216, + 206 + ], + "score": 1.0, + "content": "respectively. Explicitely, if", + "type": "text" + }, + { + "bbox": [ + 216, + 193, + 248, + 205 + ], + "score": 0.92, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 192, + 267, + 206 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 267, + 193, + 298, + 204 + ], + "score": 0.91, + "content": "\\nu \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 192, + 367, + 206 + ], + "score": 1.0, + "content": ", then the images", + "type": "text" + }, + { + "bbox": [ + 367, + 193, + 416, + 205 + ], + "score": 0.91, + "content": "\\mathcal { N } ( \\mu ) \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 192, + 434, + 206 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 435, + 193, + 489, + 205 + ], + "score": 0.91, + "content": "\\mathcal { N } ^ { * } ( \\nu ) \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 202, + 152, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 152, + 218 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 171, + 216, + 439, + 243 + ], + "lines": [ + { + "bbox": [ + 171, + 216, + 439, + 243 + ], + "spans": [ + { + "bbox": [ + 171, + 216, + 439, + 243 + ], + "score": 0.91, + "content": "{ \\mathcal { N } } ( \\mu ) ( y ) = \\sum _ { x } p ( y | x ) \\mu ( x ) \\quad { \\mathrm { a n d } } \\quad { \\mathcal { N } } ^ { * } ( \\nu ) ( x ) = \\sum _ { y } p ( x | y ) \\nu ( y ) .", + "type": "interline_equation", + "image_path": "83efb853a41eb697ad26f5434c59976d25c58e676b4343b10adc019e7bfde6eb.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 171, + 216, + 439, + 243 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 247, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 197, + 260 + ], + "score": 1.0, + "content": "These stochastic maps", + "type": "text" + }, + { + "bbox": [ + 197, + 248, + 208, + 258 + ], + "score": 0.82, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 247, + 225, + 260 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 225, + 248, + 240, + 258 + ], + "score": 0.9, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 247, + 505, + 260 + ], + "score": 1.0, + "content": "perform inference of one variable given some (possibly imperfect)", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 258, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 355, + 272 + ], + "score": 1.0, + "content": "knowledge about the other, with priors given by the marginals", + "type": "text" + }, + { + "bbox": [ + 355, + 259, + 375, + 271 + ], + "score": 0.92, + "content": "p ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 258, + 387, + 272 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 387, + 259, + 406, + 271 + ], + "score": 0.91, + "content": "p ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 258, + 418, + 272 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 418, + 259, + 447, + 271 + ], + "score": 0.92, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 258, + 505, + 272 + ], + "score": 1.0, + "content": "depending on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 279, + 282 + ], + "score": 1.0, + "content": "the direction of the inference. Importantly,", + "type": "text" + }, + { + "bbox": [ + 280, + 270, + 294, + 280 + ], + "score": 0.89, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 268, + 372, + 282 + ], + "score": 1.0, + "content": "is the transpose of", + "type": "text" + }, + { + "bbox": [ + 372, + 271, + 383, + 279 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "in terms of the inner products", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 280, + 167, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 167, + 292 + ], + "score": 1.0, + "content": "defined above:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 292, + 365, + 307 + ], + "lines": [ + { + "bbox": [ + 246, + 292, + 365, + 307 + ], + "spans": [ + { + "bbox": [ + 246, + 292, + 365, + 307 + ], + "score": 0.91, + "content": "\\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } = \\langle \\mathcal { N } ^ { * } ( \\nu ) , \\mu \\rangle _ { X } .", + "type": "interline_equation", + "image_path": "677d78296d305d22115c0b00ccfa8cc209f4016332046b6803faa126393f9014.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 246, + 292, + 365, + 307 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 330 + ], + "score": 1.0, + "content": "We now have the tools to address the problem mentioned in the introduction. The distinguishability", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 326, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 326, + 142, + 342 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 143, + 328, + 176, + 339 + ], + "score": 0.91, + "content": "q \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 326, + 196, + 342 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 196, + 330, + 210, + 339 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 326, + 336, + 342 + ], + "score": 1.0, + "content": "after the action of the channel", + "type": "text" + }, + { + "bbox": [ + 336, + 328, + 348, + 338 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 326, + 359, + 342 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 359, + 327, + 417, + 340 + ], + "score": 0.93, + "content": "\\chi ^ { 2 } ( \\mathcal { N } ( q ) , p _ { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 326, + 442, + 342 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 443, + 328, + 501, + 340 + ], + "score": 0.92, + "content": "p _ { Y } = \\mathcal { N } ( p _ { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 326, + 506, + 342 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 338, + 499, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 265, + 351 + ], + "score": 1.0, + "content": "Hence we want to find the distributions", + "type": "text" + }, + { + "bbox": [ + 266, + 341, + 272, + 351 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 338, + 499, + 351 + ], + "score": 1.0, + "content": "which maximize the relevance Beny & Osborne (2013). ´", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 356, + 394, + 385 + ], + "lines": [ + { + "bbox": [ + 217, + 356, + 394, + 385 + ], + "spans": [ + { + "bbox": [ + 217, + 356, + 394, + 385 + ], + "score": 0.94, + "content": "\\eta ( q ) = \\frac { \\chi ^ { 2 } ( \\mathcal { N } ( q ) , p _ { Y } ) } { \\chi ^ { 2 } ( q , p _ { X } ) } = \\frac { \\langle \\mathcal { N } ( \\mu ) , \\mathcal { N } ( \\mu ) \\rangle _ { Y } } { \\langle \\mu , \\mu \\rangle _ { X } } ,", + "type": "interline_equation", + "image_path": "8efbc9ac41de1e15ecde109b8c4a2a5349e45c8fbbd6588ce8bc655870bab5e1.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 356, + 394, + 370.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 217, + 370.5, + 394, + 385.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 390, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 134, + 405 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 392, + 191, + 403 + ], + "score": 0.9, + "content": "\\mu = q - p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 390, + 506, + 405 + ], + "score": 1.0, + "content": ". The inner-product formulation makes it clear that this amounts to finding", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 366, + 415 + ], + "score": 1.0, + "content": "the eigenvector with largest eigenvalue for the symmetric map", + "type": "text" + }, + { + "bbox": [ + 367, + 402, + 392, + 412 + ], + "score": 0.92, + "content": "\\mathcal { N } ^ { * } \\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 402, + 506, + 415 + ], + "score": 1.0, + "content": ", which is also the singular", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 255, + 426 + ], + "score": 1.0, + "content": "vector with largest singular value for", + "type": "text" + }, + { + "bbox": [ + 255, + 413, + 266, + 423 + ], + "score": 0.81, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 413, + 506, + 426 + ], + "score": 1.0, + "content": ". On can then go on to find the eigenvector with next largest", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 424, + 341, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 341, + 437 + ], + "score": 1.0, + "content": "eigenvalue and so on, which are automatically orthogonal.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 504, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 451, + 453 + ], + "score": 1.0, + "content": "In practice, the inner products are more tractable to compute if we express elements", + "type": "text" + }, + { + "bbox": [ + 451, + 441, + 486, + 452 + ], + "score": 0.92, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 451, + 376, + 465 + ], + "spans": [ + { + "bbox": [ + 107, + 452, + 138, + 463 + ], + "score": 0.91, + "content": "\\nu \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 451, + 223, + 465 + ], + "score": 1.0, + "content": "in terms of variables", + "type": "text" + }, + { + "bbox": [ + 223, + 452, + 231, + 464 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 451, + 248, + 465 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 249, + 453, + 255, + 464 + ], + "score": 0.79, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 451, + 267, + 465 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 267, + 452, + 306, + 464 + ], + "score": 0.91, + "content": "\\mu = p _ { X } f", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 451, + 324, + 465 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 324, + 453, + 361, + 464 + ], + "score": 0.9, + "content": "\\nu = p _ { Y } g", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 451, + 376, + 465 + ], + "score": 1.0, + "content": ", or", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 470, + 394, + 484 + ], + "lines": [ + { + "bbox": [ + 217, + 470, + 394, + 484 + ], + "spans": [ + { + "bbox": [ + 217, + 470, + 394, + 484 + ], + "score": 0.9, + "content": "\\mu ( x ) = p ( x ) f ( x ) \\quad { \\mathrm { a n d } } \\quad \\nu ( y ) = p ( y ) g ( y )", + "type": "interline_equation", + "image_path": "6cd28b0f04c36b2ba6bd2535b122fe953ab7bf7a2a9122ea177314f4b21519e4.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 217, + 470, + 394, + 484 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 490, + 258, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 259, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 133, + 505 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 493, + 150, + 502 + ], + "score": 0.86, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 488, + 259, + 505 + ], + "score": 1.0, + "content": ". Indeed, this yields simply", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 508, + 400, + 523 + ], + "lines": [ + { + "bbox": [ + 210, + 508, + 400, + 523 + ], + "spans": [ + { + "bbox": [ + 210, + 508, + 400, + 523 + ], + "score": 0.91, + "content": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } = \\operatorname { \\mathbb { E } } ( f f ^ { \\prime } ) \\quad { \\mathrm { a n d } } \\quad \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } = \\operatorname { \\mathbb { E } } ( g g ^ { \\prime } ) .", + "type": "interline_equation", + "image_path": "687c2c575cdafbc77f87dfaa38adba10014f26a4561574c1643cda0cc9e5e626.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 210, + 508, + 400, + 523 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 506, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "score": 1.0, + "content": "We are now in measure to make the connection with DCCA Andrew et al. (2013). Indeed, the aims", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 272, + 559 + ], + "score": 1.0, + "content": "of DCCA is to maximize the correlations", + "type": "text" + }, + { + "bbox": [ + 273, + 546, + 351, + 558 + ], + "score": 0.92, + "content": "\\operatorname { c o r r } ( f , g ) = \\mathbb { E } ( f g )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 546, + 408, + 559 + ], + "score": 1.0, + "content": "over function", + "type": "text" + }, + { + "bbox": [ + 408, + 546, + 428, + 558 + ], + "score": 0.92, + "content": "f ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 546, + 446, + 559 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 447, + 546, + 465, + 558 + ], + "score": 0.91, + "content": "g ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 556, + 370, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 188, + 569 + ], + "score": 0.93, + "content": "\\mathbb { E } ( f ^ { 2 } ) = \\mathbb { E } ( g ^ { 2 } ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 556, + 235, + 570 + ], + "score": 1.0, + "content": ". But, using", + "type": "text" + }, + { + "bbox": [ + 236, + 558, + 275, + 569 + ], + "score": 0.93, + "content": "\\mu = p _ { X } f", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 556, + 293, + 570 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 293, + 559, + 330, + 569 + ], + "score": 0.9, + "content": "\\nu = p _ { Y } g", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 556, + 370, + 570 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "interline_equation", + "bbox": [ + 259, + 576, + 350, + 590 + ], + "lines": [ + { + "bbox": [ + 259, + 576, + 350, + 590 + ], + "spans": [ + { + "bbox": [ + 259, + 576, + 350, + 590 + ], + "score": 0.89, + "content": "\\mathbb { E } ( f g ) = \\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } ,", + "type": "interline_equation", + "image_path": "0a9ed56485d8397567b19155defc4b1b6b44cf20e2c3c5b2dbe6898eb1ca495e.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 259, + 576, + 350, + 590 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 503, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 340, + 609 + ], + "score": 1.0, + "content": "which is maximized by the left- and right- singular vectors", + "type": "text" + }, + { + "bbox": [ + 340, + 599, + 348, + 608 + ], + "score": 0.83, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 595, + 365, + 609 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 366, + 598, + 372, + 606 + ], + "score": 0.77, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 595, + 384, + 609 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 384, + 596, + 395, + 606 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "with largest singular value.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 613, + 504, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 230, + 627 + ], + "score": 1.0, + "content": "Given all the singular vectors", + "type": "text" + }, + { + "bbox": [ + 231, + 614, + 279, + 625 + ], + "score": 0.93, + "content": "\\mu _ { i } = p _ { X } f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 612, + 298, + 627 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 615, + 345, + 625 + ], + "score": 0.89, + "content": "\\nu _ { i } = p _ { Y } g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 612, + 432, + 627 + ], + "score": 1.0, + "content": "with singular values", + "type": "text" + }, + { + "bbox": [ + 433, + 616, + 442, + 625 + ], + "score": 0.84, + "content": "\\eta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 612, + 506, + 627 + ], + "score": 1.0, + "content": ", we obtain the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 624, + 166, + 636 + ], + "spans": [ + { + "bbox": [ + 104, + 624, + 166, + 636 + ], + "score": 1.0, + "content": "representation", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 636, + 360, + 662 + ], + "lines": [ + { + "bbox": [ + 250, + 636, + 360, + 662 + ], + "spans": [ + { + "bbox": [ + 250, + 636, + 360, + 662 + ], + "score": 0.94, + "content": "\\mathcal { N } ( \\mu ) = \\sum _ { i } \\eta _ { i } \\nu _ { i } \\langle \\mu _ { i } , \\mu \\rangle _ { X } ,", + "type": "interline_equation", + "image_path": "0899cfdd5bcded9b0b8fce6ad2c29faa2b4a24e5a10c10304414279e734be06b.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 250, + 636, + 360, + 662 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 425, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 426, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 359, + 680 + ], + "score": 1.0, + "content": "which, using a more standard notation and the Kronecker delta", + "type": "text" + }, + { + "bbox": [ + 359, + 667, + 370, + 678 + ], + "score": 0.88, + "content": "\\delta _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 665, + 426, + 680 + ], + "score": 1.0, + "content": ", yields Eq. 1:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 160, + 684, + 450, + 714 + ], + "lines": [ + { + "bbox": [ + 160, + 684, + 450, + 714 + ], + "spans": [ + { + "bbox": [ + 160, + 684, + 450, + 714 + ], + "score": 0.94, + "content": "p ( y | x ) = { \\mathcal { N } } ( \\delta _ { x } ) ( y ) = { \\frac { 1 } { p ( x ) } } \\sum _ { i } \\eta _ { i } \\nu _ { i } ( y ) \\mu _ { i } ( x ) = p ( y ) \\sum _ { i } \\eta _ { i } v _ { i } ( y ) u _ { i } ( x ) ,", + "type": "interline_equation", + "image_path": "cac76d1ff3ed640c21c85d26037938c5c2ae2e784b2c708e314ec2840f41a26a.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 160, + 684, + 450, + 694.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 160, + 694.0, + 450, + 704.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 160, + 704.0, + 450, + 714.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 721, + 243, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 719, + 244, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 133, + 735 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 722, + 178, + 732 + ], + "score": 0.9, + "content": "\\mu _ { i } = p _ { X } u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 719, + 197, + 735 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 197, + 722, + 239, + 732 + ], + "score": 0.89, + "content": "\\nu _ { i } = p _ { Y } v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 719, + 244, + 735 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 342, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 342, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 280, + 96 + ], + "score": 1.0, + "content": "These inner products allow us to define the", + "type": "text" + }, + { + "bbox": [ + 280, + 82, + 292, + 94 + ], + "score": 0.9, + "content": "\\chi ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 81, + 342, + 96 + ], + "score": 1.0, + "content": "divergence:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 342, + 96 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 100, + 374, + 115 + ], + "lines": [ + { + "bbox": [ + 236, + 100, + 374, + 115 + ], + "spans": [ + { + "bbox": [ + 236, + 100, + 374, + 115 + ], + "score": 0.92, + "content": "\\chi ^ { 2 } ( q , p _ { X } ) = \\langle q - p _ { X } , q - p _ { X } \\rangle _ { X } ,", + "type": "interline_equation", + "image_path": "94f8cac2c398495c9f82111acc830c8d5618725abd8916dddbcaf9f6f22579a4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 236, + 100, + 374, + 115 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 121, + 505, + 166 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 340, + 135 + ], + "score": 1.0, + "content": "which a measures of statistical distinguishability between", + "type": "text" + }, + { + "bbox": [ + 340, + 124, + 347, + 133 + ], + "score": 0.81, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 120, + 365, + 135 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 365, + 123, + 378, + 133 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 120, + 506, + 135 + ], + "score": 1.0, + "content": ". Specifically, it quantifies how", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 315, + 145 + ], + "score": 1.0, + "content": "easy it is to reject the null hypothesis that the state is", + "type": "text" + }, + { + "bbox": [ + 315, + 134, + 329, + 144 + ], + "score": 0.88, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 132, + 403, + 145 + ], + "score": 1.0, + "content": "when it is actually", + "type": "text" + }, + { + "bbox": [ + 404, + 134, + 410, + 144 + ], + "score": 0.79, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 132, + 505, + 145 + ], + "score": 1.0, + "content": ", based on the empirical", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 156 + ], + "score": 1.0, + "content": "distribution obtained from independent samples. It is also the lowest order approximation of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 153, + 226, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 226, + 167 + ], + "score": 1.0, + "content": "Kullback-Leibler divergence.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 120, + 506, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 170, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 195, + 184 + ], + "score": 1.0, + "content": "The joint distribution", + "type": "text" + }, + { + "bbox": [ + 195, + 171, + 224, + 183 + ], + "score": 0.93, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 170, + 353, + 184 + ], + "score": 1.0, + "content": "yields conditional distributions", + "type": "text" + }, + { + "bbox": [ + 353, + 171, + 380, + 183 + ], + "score": 0.92, + "content": "p ( y | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 170, + 399, + 184 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 400, + 171, + 427, + 183 + ], + "score": 0.91, + "content": "p ( x | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 170, + 506, + 184 + ], + "score": 1.0, + "content": ". These can be un-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 181, + 503, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 349, + 194 + ], + "score": 1.0, + "content": "derstood as the components (or kernels) of stochastic maps", + "type": "text" + }, + { + "bbox": [ + 349, + 183, + 415, + 193 + ], + "score": 0.89, + "content": "\\mathcal { N } : V _ { X } V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 181, + 434, + 194 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 434, + 182, + 503, + 193 + ], + "score": 0.9, + "content": "\\mathcal { N ^ { * } } : V _ { X } V _ { Y }", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 216, + 206 + ], + "score": 1.0, + "content": "respectively. Explicitely, if", + "type": "text" + }, + { + "bbox": [ + 216, + 193, + 248, + 205 + ], + "score": 0.92, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 192, + 267, + 206 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 267, + 193, + 298, + 204 + ], + "score": 0.91, + "content": "\\nu \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 192, + 367, + 206 + ], + "score": 1.0, + "content": ", then the images", + "type": "text" + }, + { + "bbox": [ + 367, + 193, + 416, + 205 + ], + "score": 0.91, + "content": "\\mathcal { N } ( \\mu ) \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 192, + 434, + 206 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 435, + 193, + 489, + 205 + ], + "score": 0.91, + "content": "\\mathcal { N } ^ { * } ( \\nu ) \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 202, + 152, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 152, + 218 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 170, + 506, + 218 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 171, + 216, + 439, + 243 + ], + "lines": [ + { + "bbox": [ + 171, + 216, + 439, + 243 + ], + "spans": [ + { + "bbox": [ + 171, + 216, + 439, + 243 + ], + "score": 0.91, + "content": "{ \\mathcal { N } } ( \\mu ) ( y ) = \\sum _ { x } p ( y | x ) \\mu ( x ) \\quad { \\mathrm { a n d } } \\quad { \\mathcal { N } } ^ { * } ( \\nu ) ( x ) = \\sum _ { y } p ( x | y ) \\nu ( y ) .", + "type": "interline_equation", + "image_path": "83efb853a41eb697ad26f5434c59976d25c58e676b4343b10adc019e7bfde6eb.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 171, + 216, + 439, + 243 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 247, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 197, + 260 + ], + "score": 1.0, + "content": "These stochastic maps", + "type": "text" + }, + { + "bbox": [ + 197, + 248, + 208, + 258 + ], + "score": 0.82, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 247, + 225, + 260 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 225, + 248, + 240, + 258 + ], + "score": 0.9, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 247, + 505, + 260 + ], + "score": 1.0, + "content": "perform inference of one variable given some (possibly imperfect)", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 258, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 355, + 272 + ], + "score": 1.0, + "content": "knowledge about the other, with priors given by the marginals", + "type": "text" + }, + { + "bbox": [ + 355, + 259, + 375, + 271 + ], + "score": 0.92, + "content": "p ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 258, + 387, + 272 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 387, + 259, + 406, + 271 + ], + "score": 0.91, + "content": "p ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 258, + 418, + 272 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 418, + 259, + 447, + 271 + ], + "score": 0.92, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 258, + 505, + 272 + ], + "score": 1.0, + "content": "depending on", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 279, + 282 + ], + "score": 1.0, + "content": "the direction of the inference. Importantly,", + "type": "text" + }, + { + "bbox": [ + 280, + 270, + 294, + 280 + ], + "score": 0.89, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 268, + 372, + 282 + ], + "score": 1.0, + "content": "is the transpose of", + "type": "text" + }, + { + "bbox": [ + 372, + 271, + 383, + 279 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "in terms of the inner products", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 280, + 167, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 167, + 292 + ], + "score": 1.0, + "content": "defined above:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 247, + 506, + 292 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 292, + 365, + 307 + ], + "lines": [ + { + "bbox": [ + 246, + 292, + 365, + 307 + ], + "spans": [ + { + "bbox": [ + 246, + 292, + 365, + 307 + ], + "score": 0.91, + "content": "\\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } = \\langle \\mathcal { N } ^ { * } ( \\nu ) , \\mu \\rangle _ { X } .", + "type": "interline_equation", + "image_path": "677d78296d305d22115c0b00ccfa8cc209f4016332046b6803faa126393f9014.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 246, + 292, + 365, + 307 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 330 + ], + "score": 1.0, + "content": "We now have the tools to address the problem mentioned in the introduction. The distinguishability", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 326, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 326, + 142, + 342 + ], + "score": 1.0, + "content": "between", + "type": "text" + }, + { + "bbox": [ + 143, + 328, + 176, + 339 + ], + "score": 0.91, + "content": "q \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 326, + 196, + 342 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 196, + 330, + 210, + 339 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 326, + 336, + 342 + ], + "score": 1.0, + "content": "after the action of the channel", + "type": "text" + }, + { + "bbox": [ + 336, + 328, + 348, + 338 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 326, + 359, + 342 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 359, + 327, + 417, + 340 + ], + "score": 0.93, + "content": "\\chi ^ { 2 } ( \\mathcal { N } ( q ) , p _ { Y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 326, + 442, + 342 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 443, + 328, + 501, + 340 + ], + "score": 0.92, + "content": "p _ { Y } = \\mathcal { N } ( p _ { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 326, + 506, + 342 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 338, + 499, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 265, + 351 + ], + "score": 1.0, + "content": "Hence we want to find the distributions", + "type": "text" + }, + { + "bbox": [ + 266, + 341, + 272, + 351 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 338, + 499, + 351 + ], + "score": 1.0, + "content": "which maximize the relevance Beny & Osborne (2013). ´", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 315, + 506, + 351 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 356, + 394, + 385 + ], + "lines": [ + { + "bbox": [ + 217, + 356, + 394, + 385 + ], + "spans": [ + { + "bbox": [ + 217, + 356, + 394, + 385 + ], + "score": 0.94, + "content": "\\eta ( q ) = \\frac { \\chi ^ { 2 } ( \\mathcal { N } ( q ) , p _ { Y } ) } { \\chi ^ { 2 } ( q , p _ { X } ) } = \\frac { \\langle \\mathcal { N } ( \\mu ) , \\mathcal { N } ( \\mu ) \\rangle _ { Y } } { \\langle \\mu , \\mu \\rangle _ { X } } ,", + "type": "interline_equation", + "image_path": "8efbc9ac41de1e15ecde109b8c4a2a5349e45c8fbbd6588ce8bc655870bab5e1.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 356, + 394, + 370.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 217, + 370.5, + 394, + 385.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 390, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 134, + 405 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 392, + 191, + 403 + ], + "score": 0.9, + "content": "\\mu = q - p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 390, + 506, + 405 + ], + "score": 1.0, + "content": ". The inner-product formulation makes it clear that this amounts to finding", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 366, + 415 + ], + "score": 1.0, + "content": "the eigenvector with largest eigenvalue for the symmetric map", + "type": "text" + }, + { + "bbox": [ + 367, + 402, + 392, + 412 + ], + "score": 0.92, + "content": "\\mathcal { N } ^ { * } \\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 402, + 506, + 415 + ], + "score": 1.0, + "content": ", which is also the singular", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 255, + 426 + ], + "score": 1.0, + "content": "vector with largest singular value for", + "type": "text" + }, + { + "bbox": [ + 255, + 413, + 266, + 423 + ], + "score": 0.81, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 413, + 506, + 426 + ], + "score": 1.0, + "content": ". On can then go on to find the eigenvector with next largest", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 424, + 341, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 341, + 437 + ], + "score": 1.0, + "content": "eigenvalue and so on, which are automatically orthogonal.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 390, + 506, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 504, + 464 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 451, + 453 + ], + "score": 1.0, + "content": "In practice, the inner products are more tractable to compute if we express elements", + "type": "text" + }, + { + "bbox": [ + 451, + 441, + 486, + 452 + ], + "score": 0.92, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 451, + 376, + 465 + ], + "spans": [ + { + "bbox": [ + 107, + 452, + 138, + 463 + ], + "score": 0.91, + "content": "\\nu \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 451, + 223, + 465 + ], + "score": 1.0, + "content": "in terms of variables", + "type": "text" + }, + { + "bbox": [ + 223, + 452, + 231, + 464 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 451, + 248, + 465 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 249, + 453, + 255, + 464 + ], + "score": 0.79, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 451, + 267, + 465 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 267, + 452, + 306, + 464 + ], + "score": 0.91, + "content": "\\mu = p _ { X } f", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 451, + 324, + 465 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 324, + 453, + 361, + 464 + ], + "score": 0.9, + "content": "\\nu = p _ { Y } g", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 451, + 376, + 465 + ], + "score": 1.0, + "content": ", or", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 441, + 505, + 465 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 470, + 394, + 484 + ], + "lines": [ + { + "bbox": [ + 217, + 470, + 394, + 484 + ], + "spans": [ + { + "bbox": [ + 217, + 470, + 394, + 484 + ], + "score": 0.9, + "content": "\\mu ( x ) = p ( x ) f ( x ) \\quad { \\mathrm { a n d } } \\quad \\nu ( y ) = p ( y ) g ( y )", + "type": "interline_equation", + "image_path": "6cd28b0f04c36b2ba6bd2535b122fe953ab7bf7a2a9122ea177314f4b21519e4.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 217, + 470, + 394, + 484 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 490, + 258, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 259, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 133, + 505 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 493, + 150, + 502 + ], + "score": 0.86, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 488, + 259, + 505 + ], + "score": 1.0, + "content": ". Indeed, this yields simply", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 488, + 259, + 505 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 508, + 400, + 523 + ], + "lines": [ + { + "bbox": [ + 210, + 508, + 400, + 523 + ], + "spans": [ + { + "bbox": [ + 210, + 508, + 400, + 523 + ], + "score": 0.91, + "content": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } = \\operatorname { \\mathbb { E } } ( f f ^ { \\prime } ) \\quad { \\mathrm { a n d } } \\quad \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } = \\operatorname { \\mathbb { E } } ( g g ^ { \\prime } ) .", + "type": "interline_equation", + "image_path": "687c2c575cdafbc77f87dfaa38adba10014f26a4561574c1643cda0cc9e5e626.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 210, + 508, + 400, + 523 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 506, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "score": 1.0, + "content": "We are now in measure to make the connection with DCCA Andrew et al. (2013). Indeed, the aims", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 272, + 559 + ], + "score": 1.0, + "content": "of DCCA is to maximize the correlations", + "type": "text" + }, + { + "bbox": [ + 273, + 546, + 351, + 558 + ], + "score": 0.92, + "content": "\\operatorname { c o r r } ( f , g ) = \\mathbb { E } ( f g )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 546, + 408, + 559 + ], + "score": 1.0, + "content": "over function", + "type": "text" + }, + { + "bbox": [ + 408, + 546, + 428, + 558 + ], + "score": 0.92, + "content": "f ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 546, + 446, + 559 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 447, + 546, + 465, + 558 + ], + "score": 0.91, + "content": "g ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 556, + 370, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 188, + 569 + ], + "score": 0.93, + "content": "\\mathbb { E } ( f ^ { 2 } ) = \\mathbb { E } ( g ^ { 2 } ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 556, + 235, + 570 + ], + "score": 1.0, + "content": ". But, using", + "type": "text" + }, + { + "bbox": [ + 236, + 558, + 275, + 569 + ], + "score": 0.93, + "content": "\\mu = p _ { X } f", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 556, + 293, + 570 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 293, + 559, + 330, + 569 + ], + "score": 0.9, + "content": "\\nu = p _ { Y } g", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 556, + 370, + 570 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 535, + 506, + 570 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 259, + 576, + 350, + 590 + ], + "lines": [ + { + "bbox": [ + 259, + 576, + 350, + 590 + ], + "spans": [ + { + "bbox": [ + 259, + 576, + 350, + 590 + ], + "score": 0.89, + "content": "\\mathbb { E } ( f g ) = \\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } ,", + "type": "interline_equation", + "image_path": "0a9ed56485d8397567b19155defc4b1b6b44cf20e2c3c5b2dbe6898eb1ca495e.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 259, + 576, + 350, + 590 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 503, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 340, + 609 + ], + "score": 1.0, + "content": "which is maximized by the left- and right- singular vectors", + "type": "text" + }, + { + "bbox": [ + 340, + 599, + 348, + 608 + ], + "score": 0.83, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 595, + 365, + 609 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 366, + 598, + 372, + 606 + ], + "score": 0.77, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 595, + 384, + 609 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 384, + 596, + 395, + 606 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "with largest singular value.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 595, + 505, + 609 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 613, + 504, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 612, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 230, + 627 + ], + "score": 1.0, + "content": "Given all the singular vectors", + "type": "text" + }, + { + "bbox": [ + 231, + 614, + 279, + 625 + ], + "score": 0.93, + "content": "\\mu _ { i } = p _ { X } f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 612, + 298, + 627 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 615, + 345, + 625 + ], + "score": 0.89, + "content": "\\nu _ { i } = p _ { Y } g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 612, + 432, + 627 + ], + "score": 1.0, + "content": "with singular values", + "type": "text" + }, + { + "bbox": [ + 433, + 616, + 442, + 625 + ], + "score": 0.84, + "content": "\\eta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 612, + 506, + 627 + ], + "score": 1.0, + "content": ", we obtain the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 624, + 166, + 636 + ], + "spans": [ + { + "bbox": [ + 104, + 624, + 166, + 636 + ], + "score": 1.0, + "content": "representation", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 612, + 506, + 636 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 636, + 360, + 662 + ], + "lines": [ + { + "bbox": [ + 250, + 636, + 360, + 662 + ], + "spans": [ + { + "bbox": [ + 250, + 636, + 360, + 662 + ], + "score": 0.94, + "content": "\\mathcal { N } ( \\mu ) = \\sum _ { i } \\eta _ { i } \\nu _ { i } \\langle \\mu _ { i } , \\mu \\rangle _ { X } ,", + "type": "interline_equation", + "image_path": "0899cfdd5bcded9b0b8fce6ad2c29faa2b4a24e5a10c10304414279e734be06b.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 250, + 636, + 360, + 662 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 425, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 426, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 359, + 680 + ], + "score": 1.0, + "content": "which, using a more standard notation and the Kronecker delta", + "type": "text" + }, + { + "bbox": [ + 359, + 667, + 370, + 678 + ], + "score": 0.88, + "content": "\\delta _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 665, + 426, + 680 + ], + "score": 1.0, + "content": ", yields Eq. 1:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 665, + 426, + 680 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 160, + 684, + 450, + 714 + ], + "lines": [ + { + "bbox": [ + 160, + 684, + 450, + 714 + ], + "spans": [ + { + "bbox": [ + 160, + 684, + 450, + 714 + ], + "score": 0.94, + "content": "p ( y | x ) = { \\mathcal { N } } ( \\delta _ { x } ) ( y ) = { \\frac { 1 } { p ( x ) } } \\sum _ { i } \\eta _ { i } \\nu _ { i } ( y ) \\mu _ { i } ( x ) = p ( y ) \\sum _ { i } \\eta _ { i } v _ { i } ( y ) u _ { i } ( x ) ,", + "type": "interline_equation", + "image_path": "cac76d1ff3ed640c21c85d26037938c5c2ae2e784b2c708e314ec2840f41a26a.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 160, + 684, + 450, + 694.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 160, + 694.0, + 450, + 704.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 160, + 704.0, + 450, + 714.0 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 721, + 243, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 719, + 244, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 133, + 735 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 722, + 178, + 732 + ], + "score": 0.9, + "content": "\\mu _ { i } = p _ { X } u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 719, + 197, + 735 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 197, + 722, + 239, + 732 + ], + "score": 0.89, + "content": "\\nu _ { i } = p _ { Y } v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 719, + 244, + 735 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 719, + 244, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 102, + 505, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "For the purpose of the optimization and inference, we do not need to full diagonal decomposition,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 179, + 128 + ], + "score": 1.0, + "content": "but just functions", + "type": "text" + }, + { + "bbox": [ + 179, + 114, + 229, + 126 + ], + "score": 0.93, + "content": "f _ { i } = \\mu _ { i } / p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 113, + 248, + 128 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 248, + 114, + 368, + 126 + ], + "score": 0.75, + "content": "g _ { j } = \\nu _ { j } / p _ { Y } , i , j = 1 , \\ldots , k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 113, + 506, + 128 + ], + "score": 1.0, + "content": ", which have the same span as the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 187, + 138 + ], + "score": 1.0, + "content": "canonical variables", + "type": "text" + }, + { + "bbox": [ + 187, + 127, + 197, + 136 + ], + "score": 0.85, + "content": "u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 125, + 217, + 138 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 127, + 227, + 137 + ], + "score": 0.87, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 125, + 295, + 138 + ], + "score": 1.0, + "content": "respectively for", + "type": "text" + }, + { + "bbox": [ + 295, + 126, + 362, + 136 + ], + "score": 0.9, + "content": "i , j = 1 , \\ldots , k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 125, + 426, + 138 + ], + "score": 1.0, + "content": "(assuming that", + "type": "text" + }, + { + "bbox": [ + 426, + 127, + 473, + 137 + ], + "score": 0.91, + "content": "\\eta _ { 1 } , \\ldots , \\eta _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "are the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 136, + 453, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 277, + 149 + ], + "score": 1.0, + "content": "largest singular vector). Below we refer to", + "type": "text" + }, + { + "bbox": [ + 278, + 137, + 287, + 148 + ], + "score": 0.88, + "content": "f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 136, + 305, + 149 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 305, + 138, + 315, + 149 + ], + "score": 0.86, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 136, + 342, + 149 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 342, + 137, + 353, + 147 + ], + "score": 0.86, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 136, + 453, + 149 + ], + "score": 1.0, + "content": "most relevant variables.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 108, + 153, + 432, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 432, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 432, + 166 + ], + "score": 1.0, + "content": "Because these functions may not be orthogonal, we need the covariance matrices", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 170, + 496, + 185 + ], + "lines": [ + { + "bbox": [ + 113, + 170, + 496, + 185 + ], + "spans": [ + { + "bbox": [ + 113, + 170, + 496, + 185 + ], + "score": 0.9, + "content": "K _ { i j } = \\langle \\mu _ { i } , \\mu _ { j } \\rangle _ { X } = \\mathbb { E } ( f _ { i } f _ { j } ) , \\quad L _ { i j } = \\langle \\nu _ { i } , \\nu _ { j } \\rangle _ { X } = \\mathbb { E } ( g _ { i } g _ { j } ) , \\quad A _ { i j } = \\langle \\nu _ { i } , N ( \\mu _ { j } ) \\rangle _ { Y } = \\mathbb { E } ( g _ { i } f _ { j } ) .", + "type": "interline_equation", + "image_path": "454c207fea8af6fb7e54a62760ae9f11fdb9b33fecf0925c06e090123d1dd7e3.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 113, + 170, + 496, + 185 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 505, + 244 + ], + "lines": [ + { + "bbox": [ + 104, + 194, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 117, + 209 + ], + "score": 1.0, + "content": "If", + "type": "text" + }, + { + "bbox": [ + 117, + 196, + 133, + 208 + ], + "score": 0.91, + "content": "N _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 194, + 244, + 209 + ], + "score": 1.0, + "content": "denote the components of", + "type": "text" + }, + { + "bbox": [ + 244, + 196, + 255, + 205 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 194, + 327, + 209 + ], + "score": 1.0, + "content": "in the sense that", + "type": "text" + }, + { + "bbox": [ + 328, + 195, + 435, + 208 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathcal { N } ( p _ { X } f _ { j } ) = \\sum _ { i } N _ { i j } p _ { Y } g _ { i } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 194, + 506, + 209 + ], + "score": 1.0, + "content": ", then, using our", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 206, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 104, + 206, + 206, + 222 + ], + "score": 1.0, + "content": "inner products to isolate", + "type": "text" + }, + { + "bbox": [ + 207, + 208, + 222, + 220 + ], + "score": 0.91, + "content": "N _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 206, + 269, + 222 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + }, + { + "bbox": [ + 270, + 207, + 319, + 218 + ], + "score": 0.91, + "content": "N = L ^ { - 1 } A", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 206, + 444, + 222 + ], + "score": 1.0, + "content": ". Similarly, the components of", + "type": "text" + }, + { + "bbox": [ + 444, + 208, + 459, + 218 + ], + "score": 0.87, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 206, + 476, + 222 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 477, + 208, + 505, + 221 + ], + "score": 0.91, + "content": "N _ { i j } ^ { * } =", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 141, + 231 + ], + "score": 0.91, + "content": "K ^ { - 1 } A ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 220, + 403, + 234 + ], + "score": 1.0, + "content": ". This implies that the sum of the square of the singular values of", + "type": "text" + }, + { + "bbox": [ + 403, + 221, + 415, + 231 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "restricted to the spans", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 231, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 163, + 246 + ], + "score": 1.0, + "content": "of the vectors", + "type": "text" + }, + { + "bbox": [ + 163, + 233, + 185, + 244 + ], + "score": 0.92, + "content": "p _ { X } f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 231, + 203, + 246 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 203, + 234, + 225, + 245 + ], + "score": 0.9, + "content": "p _ { Y } g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 231, + 253, + 246 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 253, + 233, + 267, + 244 + ], + "score": 0.87, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 231, + 482, + 246 + ], + "score": 1.0, + "content": ", which is what we want to maximize, is just given by", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 250, + 394, + 285 + ], + "lines": [ + { + "bbox": [ + 214, + 250, + 394, + 285 + ], + "spans": [ + { + "bbox": [ + 214, + 250, + 394, + 285 + ], + "score": 0.93, + "content": "\\sum _ { i = 1 } ^ { k _ { 0 } } \\eta _ { i } ^ { 2 } = \\mathrm { T r } \\left( N ^ { \\ast } N \\right) = \\mathrm { T r } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A ) .", + "type": "interline_equation", + "image_path": "f1587c308117d7252ef0335d24a7a140dc31be7bcbc658867615280d59a8460e.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 250, + 394, + 267.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 214, + 267.5, + 394, + 285.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 289, + 504, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 380, + 303 + ], + "score": 1.0, + "content": "This is the DCCA objective. Below we use the objective function", + "type": "text" + }, + { + "bbox": [ + 380, + 290, + 470, + 302 + ], + "score": 0.91, + "content": "C = k _ { 0 } - \\mathrm { T r } \\left( N ^ { \\ast } N \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 290, + 505, + 303 + ], + "score": 1.0, + "content": ", for the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 301, + 291, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 291, + 314 + ], + "score": 1.0, + "content": "cosmetic reason that its optimal value is zero.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 104, + 318, + 455, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 454, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 454, + 331 + ], + "score": 1.0, + "content": "Moreover, the corresponding truncated representation of the conditional distribution is", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 335, + 425, + 371 + ], + "lines": [ + { + "bbox": [ + 185, + 335, + 425, + 371 + ], + "spans": [ + { + "bbox": [ + 185, + 335, + 425, + 371 + ], + "score": 0.93, + "content": "p ( \\boldsymbol { y } | \\boldsymbol { x } ) = \\mathcal { N } ( \\delta _ { \\boldsymbol { x } } ) ( \\boldsymbol { y } ) \\simeq p ( \\boldsymbol { y } ) \\sum _ { i , j = 1 } ^ { k _ { 0 } } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } g _ { i } ( \\boldsymbol { y } ) f _ { j } ( \\boldsymbol { x } ) ,", + "type": "interline_equation", + "image_path": "d0d1286f608b3021f23eb2026fa6e81a31d907cd4c6e4cc4877260801cf01603.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 335, + 425, + 353.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 185, + 353.0, + 425, + 371.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 377, + 403, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 404, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 294, + 392 + ], + "score": 1.0, + "content": "where we used the fact that the components of", + "type": "text" + }, + { + "bbox": [ + 294, + 379, + 304, + 389 + ], + "score": 0.88, + "content": "\\delta _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 376, + 320, + 392 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 321, + 377, + 401, + 392 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\delta _ { j } = \\sum _ { i } K _ { j i } ^ { - 1 } f _ { i } ( x ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 376, + 404, + 392 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 482, + 408 + ], + "score": 1.0, + "content": "Of course, This approach can produce a faithful representation of the correlations only if", + "type": "text" + }, + { + "bbox": [ + 483, + 396, + 493, + 405 + ], + "score": 0.81, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 228, + 419 + ], + "score": 1.0, + "content": "actually close to being of rank", + "type": "text" + }, + { + "bbox": [ + 229, + 407, + 239, + 417 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 407, + 300, + 419 + ], + "score": 1.0, + "content": "(see Appendix", + "type": "text" + }, + { + "bbox": [ + 301, + 407, + 309, + 416 + ], + "score": 0.4, + "content": "\\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "for a more precise statement). If we interpret the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "relevant subspace as a space of probability over latent variable, this means that our latent variables", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 429, + 232, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 159, + 440 + ], + "score": 1.0, + "content": "have at most", + "type": "text" + }, + { + "bbox": [ + 159, + 429, + 170, + 439 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 429, + 232, + 440 + ], + "score": 1.0, + "content": "discrete states.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 444, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 216, + 458 + ], + "score": 1.0, + "content": "However, even if the rank", + "type": "text" + }, + { + "bbox": [ + 216, + 446, + 227, + 456 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 443, + 269, + 458 + ], + "score": 1.0, + "content": "corner of", + "type": "text" + }, + { + "bbox": [ + 270, + 445, + 281, + 455 + ], + "score": 0.82, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 443, + 505, + 458 + ], + "score": 1.0, + "content": "is a not a good approximation, this strategy allows us", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "to nevertheless do the correct inference on certain random variables, namely those which are in the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 467, + 235, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 235, + 479 + ], + "score": 1.0, + "content": "span of the canonical variables!", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 456, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 458, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 283, + 497 + ], + "score": 1.0, + "content": "Indeed, the exact conditional expectation of", + "type": "text" + }, + { + "bbox": [ + 283, + 486, + 294, + 496 + ], + "score": 0.86, + "content": "g _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 483, + 348, + 497 + ], + "score": 1.0, + "content": "is (assuming", + "type": "text" + }, + { + "bbox": [ + 348, + 485, + 357, + 494 + ], + "score": 0.85, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 483, + 440, + 497 + ], + "score": 1.0, + "content": "is the actual rank of", + "type": "text" + }, + { + "bbox": [ + 440, + 484, + 450, + 494 + ], + "score": 0.8, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 483, + 458, + 497 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 175, + 501, + 435, + 574 + ], + "lines": [ + { + "bbox": [ + 175, + 501, + 435, + 574 + ], + "spans": [ + { + "bbox": [ + 175, + 501, + 435, + 574 + ], + "score": 0.95, + "content": "\\begin{array} { l } { { \\displaystyle \\sum _ { y } g _ { k } ( y ) p ( y | x ) = \\sum _ { i , j = 1 } ^ { D } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } \\mathbb { E } ( g _ { k } g _ { i } ) f _ { j } ( x ) } } \\\\ { { \\displaystyle \\qquad = \\sum _ { j = 1 } ^ { D } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) = \\sum _ { j = 1 } ^ { k _ { 0 } } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "ec316fca8c448843caf2ac556e4420de40df72f1ce473ed26529bbebf75318f8.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 175, + 501, + 435, + 525.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 175, + 525.3333333333334, + 435, + 549.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 175, + 549.6666666666667, + 435, + 574.0000000000001 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 578, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 250, + 591 + ], + "score": 1.0, + "content": "where the last truncation is exact if", + "type": "text" + }, + { + "bbox": [ + 251, + 578, + 281, + 590 + ], + "score": 0.93, + "content": "k \\leq k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 578, + 429, + 591 + ], + "score": 1.0, + "content": "due to the assumption that the basis", + "type": "text" + }, + { + "bbox": [ + 429, + 579, + 438, + 590 + ], + "score": 0.89, + "content": "f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 578, + 457, + 591 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 457, + 580, + 468, + 591 + ], + "score": 0.86, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "have the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 589, + 411, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 175, + 603 + ], + "score": 1.0, + "content": "same span as the", + "type": "text" + }, + { + "bbox": [ + 176, + 590, + 186, + 601 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 589, + 347, + 603 + ], + "score": 1.0, + "content": "largest right and left singular vectors of", + "type": "text" + }, + { + "bbox": [ + 348, + 590, + 358, + 599 + ], + "score": 0.82, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 589, + 411, + 603 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 606, + 504, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 173, + 619 + ], + "score": 1.0, + "content": "For instance, if", + "type": "text" + }, + { + "bbox": [ + 173, + 606, + 202, + 618 + ], + "score": 0.93, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "is Gaussian, the canonical variables can be computed analytically, as in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "Lancaster (1958) or Beny (2018a) in the multivariate case. Solutions for other distributions were ´", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 629, + 248, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 248, + 641 + ], + "score": 1.0, + "content": "also computed in Eagleson (1964).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 331, + 658 + ], + "score": 1.0, + "content": "Notably, for any two dimensional Gaussian, the space of", + "type": "text" + }, + { + "bbox": [ + 331, + 646, + 338, + 655 + ], + "score": 0.84, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "most relevant variables is simply spanned", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 656, + 504, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 174, + 669 + ], + "score": 1.0, + "content": "by the moments", + "type": "text" + }, + { + "bbox": [ + 174, + 656, + 224, + 668 + ], + "score": 0.93, + "content": "f _ { n } ( x ) = x ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 657, + 243, + 669 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 243, + 656, + 292, + 668 + ], + "score": 0.92, + "content": "g _ { n } ( y ) = y ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 657, + 308, + 669 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 308, + 657, + 380, + 668 + ], + "score": 0.91, + "content": "n = 0 , \\ldots , k - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 657, + 497, + 669 + ], + "score": 1.0, + "content": ". Hence, in this case the first", + "type": "text" + }, + { + "bbox": [ + 497, + 657, + 504, + 666 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 667, + 494, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 298, + 680 + ], + "score": 1.0, + "content": "moments can be inferred exactly using only the", + "type": "text" + }, + { + "bbox": [ + 298, + 668, + 322, + 678 + ], + "score": 0.89, + "content": "k + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 667, + 494, + 680 + ], + "score": 1.0, + "content": "most relevant variables (See Appendix C).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 107, + 695, + 191, + 708 + ], + "lines": [ + { + "bbox": [ + 105, + 694, + 192, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 192, + 710 + ], + "score": 1.0, + "content": "4 ALGORITHM", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 357, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 359, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 359, + 734 + ], + "score": 1.0, + "content": "Let us explicit the algorithm resulting from the above analysis.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + }, + { + "type": "discarded", + "bbox": [ + 105, + 83, + 345, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 348, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 348, + 95 + ], + "score": 1.0, + "content": "3.1 NON-DIAGONAL FORM AND RELEVANT VARIABLES", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 102, + 505, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "For the purpose of the optimization and inference, we do not need to full diagonal decomposition,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 179, + 128 + ], + "score": 1.0, + "content": "but just functions", + "type": "text" + }, + { + "bbox": [ + 179, + 114, + 229, + 126 + ], + "score": 0.93, + "content": "f _ { i } = \\mu _ { i } / p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 113, + 248, + 128 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 248, + 114, + 368, + 126 + ], + "score": 0.75, + "content": "g _ { j } = \\nu _ { j } / p _ { Y } , i , j = 1 , \\ldots , k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 113, + 506, + 128 + ], + "score": 1.0, + "content": ", which have the same span as the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 125, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 187, + 138 + ], + "score": 1.0, + "content": "canonical variables", + "type": "text" + }, + { + "bbox": [ + 187, + 127, + 197, + 136 + ], + "score": 0.85, + "content": "u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 125, + 217, + 138 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 127, + 227, + 137 + ], + "score": 0.87, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 125, + 295, + 138 + ], + "score": 1.0, + "content": "respectively for", + "type": "text" + }, + { + "bbox": [ + 295, + 126, + 362, + 136 + ], + "score": 0.9, + "content": "i , j = 1 , \\ldots , k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 125, + 426, + 138 + ], + "score": 1.0, + "content": "(assuming that", + "type": "text" + }, + { + "bbox": [ + 426, + 127, + 473, + 137 + ], + "score": 0.91, + "content": "\\eta _ { 1 } , \\ldots , \\eta _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 125, + 506, + 138 + ], + "score": 1.0, + "content": "are the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 136, + 453, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 277, + 149 + ], + "score": 1.0, + "content": "largest singular vector). Below we refer to", + "type": "text" + }, + { + "bbox": [ + 278, + 137, + 287, + 148 + ], + "score": 0.88, + "content": "f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 136, + 305, + 149 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 305, + 138, + 315, + 149 + ], + "score": 0.86, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 136, + 342, + 149 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 342, + 137, + 353, + 147 + ], + "score": 0.86, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 136, + 453, + 149 + ], + "score": 1.0, + "content": "most relevant variables.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 102, + 506, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 153, + 432, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 432, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 432, + 166 + ], + "score": 1.0, + "content": "Because these functions may not be orthogonal, we need the covariance matrices", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 151, + 432, + 166 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 170, + 496, + 185 + ], + "lines": [ + { + "bbox": [ + 113, + 170, + 496, + 185 + ], + "spans": [ + { + "bbox": [ + 113, + 170, + 496, + 185 + ], + "score": 0.9, + "content": "K _ { i j } = \\langle \\mu _ { i } , \\mu _ { j } \\rangle _ { X } = \\mathbb { E } ( f _ { i } f _ { j } ) , \\quad L _ { i j } = \\langle \\nu _ { i } , \\nu _ { j } \\rangle _ { X } = \\mathbb { E } ( g _ { i } g _ { j } ) , \\quad A _ { i j } = \\langle \\nu _ { i } , N ( \\mu _ { j } ) \\rangle _ { Y } = \\mathbb { E } ( g _ { i } f _ { j } ) .", + "type": "interline_equation", + "image_path": "454c207fea8af6fb7e54a62760ae9f11fdb9b33fecf0925c06e090123d1dd7e3.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 113, + 170, + 496, + 185 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 505, + 244 + ], + "lines": [ + { + "bbox": [ + 104, + 194, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 117, + 209 + ], + "score": 1.0, + "content": "If", + "type": "text" + }, + { + "bbox": [ + 117, + 196, + 133, + 208 + ], + "score": 0.91, + "content": "N _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 194, + 244, + 209 + ], + "score": 1.0, + "content": "denote the components of", + "type": "text" + }, + { + "bbox": [ + 244, + 196, + 255, + 205 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 194, + 327, + 209 + ], + "score": 1.0, + "content": "in the sense that", + "type": "text" + }, + { + "bbox": [ + 328, + 195, + 435, + 208 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathcal { N } ( p _ { X } f _ { j } ) = \\sum _ { i } N _ { i j } p _ { Y } g _ { i } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 194, + 506, + 209 + ], + "score": 1.0, + "content": ", then, using our", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 206, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 104, + 206, + 206, + 222 + ], + "score": 1.0, + "content": "inner products to isolate", + "type": "text" + }, + { + "bbox": [ + 207, + 208, + 222, + 220 + ], + "score": 0.91, + "content": "N _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 206, + 269, + 222 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + }, + { + "bbox": [ + 270, + 207, + 319, + 218 + ], + "score": 0.91, + "content": "N = L ^ { - 1 } A", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 206, + 444, + 222 + ], + "score": 1.0, + "content": ". Similarly, the components of", + "type": "text" + }, + { + "bbox": [ + 444, + 208, + 459, + 218 + ], + "score": 0.87, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 206, + 476, + 222 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 477, + 208, + 505, + 221 + ], + "score": 0.91, + "content": "N _ { i j } ^ { * } =", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 141, + 231 + ], + "score": 0.91, + "content": "K ^ { - 1 } A ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 220, + 403, + 234 + ], + "score": 1.0, + "content": ". This implies that the sum of the square of the singular values of", + "type": "text" + }, + { + "bbox": [ + 403, + 221, + 415, + 231 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "restricted to the spans", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 231, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 163, + 246 + ], + "score": 1.0, + "content": "of the vectors", + "type": "text" + }, + { + "bbox": [ + 163, + 233, + 185, + 244 + ], + "score": 0.92, + "content": "p _ { X } f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 231, + 203, + 246 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 203, + 234, + 225, + 245 + ], + "score": 0.9, + "content": "p _ { Y } g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 231, + 253, + 246 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 253, + 233, + 267, + 244 + ], + "score": 0.87, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 231, + 482, + 246 + ], + "score": 1.0, + "content": ", which is what we want to maximize, is just given by", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 104, + 194, + 506, + 246 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 250, + 394, + 285 + ], + "lines": [ + { + "bbox": [ + 214, + 250, + 394, + 285 + ], + "spans": [ + { + "bbox": [ + 214, + 250, + 394, + 285 + ], + "score": 0.93, + "content": "\\sum _ { i = 1 } ^ { k _ { 0 } } \\eta _ { i } ^ { 2 } = \\mathrm { T r } \\left( N ^ { \\ast } N \\right) = \\mathrm { T r } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A ) .", + "type": "interline_equation", + "image_path": "f1587c308117d7252ef0335d24a7a140dc31be7bcbc658867615280d59a8460e.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 250, + 394, + 267.5 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 214, + 267.5, + 394, + 285.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 289, + 504, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 380, + 303 + ], + "score": 1.0, + "content": "This is the DCCA objective. Below we use the objective function", + "type": "text" + }, + { + "bbox": [ + 380, + 290, + 470, + 302 + ], + "score": 0.91, + "content": "C = k _ { 0 } - \\mathrm { T r } \\left( N ^ { \\ast } N \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 290, + 505, + 303 + ], + "score": 1.0, + "content": ", for the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 301, + 291, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 291, + 314 + ], + "score": 1.0, + "content": "cosmetic reason that its optimal value is zero.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 290, + 505, + 314 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 318, + 455, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 454, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 454, + 331 + ], + "score": 1.0, + "content": "Moreover, the corresponding truncated representation of the conditional distribution is", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 317, + 454, + 331 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 335, + 425, + 371 + ], + "lines": [ + { + "bbox": [ + 185, + 335, + 425, + 371 + ], + "spans": [ + { + "bbox": [ + 185, + 335, + 425, + 371 + ], + "score": 0.93, + "content": "p ( \\boldsymbol { y } | \\boldsymbol { x } ) = \\mathcal { N } ( \\delta _ { \\boldsymbol { x } } ) ( \\boldsymbol { y } ) \\simeq p ( \\boldsymbol { y } ) \\sum _ { i , j = 1 } ^ { k _ { 0 } } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } g _ { i } ( \\boldsymbol { y } ) f _ { j } ( \\boldsymbol { x } ) ,", + "type": "interline_equation", + "image_path": "d0d1286f608b3021f23eb2026fa6e81a31d907cd4c6e4cc4877260801cf01603.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 335, + 425, + 353.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 185, + 353.0, + 425, + 371.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 377, + 403, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 404, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 294, + 392 + ], + "score": 1.0, + "content": "where we used the fact that the components of", + "type": "text" + }, + { + "bbox": [ + 294, + 379, + 304, + 389 + ], + "score": 0.88, + "content": "\\delta _ { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 376, + 320, + 392 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 321, + 377, + 401, + 392 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\delta _ { j } = \\sum _ { i } K _ { j i } ^ { - 1 } f _ { i } ( x ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 376, + 404, + 392 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 376, + 404, + 392 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 482, + 408 + ], + "score": 1.0, + "content": "Of course, This approach can produce a faithful representation of the correlations only if", + "type": "text" + }, + { + "bbox": [ + 483, + 396, + 493, + 405 + ], + "score": 0.81, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 228, + 419 + ], + "score": 1.0, + "content": "actually close to being of rank", + "type": "text" + }, + { + "bbox": [ + 229, + 407, + 239, + 417 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 407, + 300, + 419 + ], + "score": 1.0, + "content": "(see Appendix", + "type": "text" + }, + { + "bbox": [ + 301, + 407, + 309, + 416 + ], + "score": 0.4, + "content": "\\mathbf { B }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "for a more precise statement). If we interpret the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "relevant subspace as a space of probability over latent variable, this means that our latent variables", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 429, + 232, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 159, + 440 + ], + "score": 1.0, + "content": "have at most", + "type": "text" + }, + { + "bbox": [ + 159, + 429, + 170, + 439 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 429, + 232, + 440 + ], + "score": 1.0, + "content": "discrete states.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 395, + 506, + 440 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 444, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 216, + 458 + ], + "score": 1.0, + "content": "However, even if the rank", + "type": "text" + }, + { + "bbox": [ + 216, + 446, + 227, + 456 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 443, + 269, + 458 + ], + "score": 1.0, + "content": "corner of", + "type": "text" + }, + { + "bbox": [ + 270, + 445, + 281, + 455 + ], + "score": 0.82, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 443, + 505, + 458 + ], + "score": 1.0, + "content": "is a not a good approximation, this strategy allows us", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "to nevertheless do the correct inference on certain random variables, namely those which are in the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 467, + 235, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 235, + 479 + ], + "score": 1.0, + "content": "span of the canonical variables!", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 443, + 505, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 456, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 458, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 283, + 497 + ], + "score": 1.0, + "content": "Indeed, the exact conditional expectation of", + "type": "text" + }, + { + "bbox": [ + 283, + 486, + 294, + 496 + ], + "score": 0.86, + "content": "g _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 483, + 348, + 497 + ], + "score": 1.0, + "content": "is (assuming", + "type": "text" + }, + { + "bbox": [ + 348, + 485, + 357, + 494 + ], + "score": 0.85, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 483, + 440, + 497 + ], + "score": 1.0, + "content": "is the actual rank of", + "type": "text" + }, + { + "bbox": [ + 440, + 484, + 450, + 494 + ], + "score": 0.8, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 483, + 458, + 497 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 483, + 458, + 497 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 175, + 501, + 435, + 574 + ], + "lines": [ + { + "bbox": [ + 175, + 501, + 435, + 574 + ], + "spans": [ + { + "bbox": [ + 175, + 501, + 435, + 574 + ], + "score": 0.95, + "content": "\\begin{array} { l } { { \\displaystyle \\sum _ { y } g _ { k } ( y ) p ( y | x ) = \\sum _ { i , j = 1 } ^ { D } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } \\mathbb { E } ( g _ { k } g _ { i } ) f _ { j } ( x ) } } \\\\ { { \\displaystyle \\qquad = \\sum _ { j = 1 } ^ { D } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) = \\sum _ { j = 1 } ^ { k _ { 0 } } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) } , } \\end{array}", + "type": "interline_equation", + "image_path": "ec316fca8c448843caf2ac556e4420de40df72f1ce473ed26529bbebf75318f8.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 175, + 501, + 435, + 525.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 175, + 525.3333333333334, + 435, + 549.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 175, + 549.6666666666667, + 435, + 574.0000000000001 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 578, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 250, + 591 + ], + "score": 1.0, + "content": "where the last truncation is exact if", + "type": "text" + }, + { + "bbox": [ + 251, + 578, + 281, + 590 + ], + "score": 0.93, + "content": "k \\leq k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 578, + 429, + 591 + ], + "score": 1.0, + "content": "due to the assumption that the basis", + "type": "text" + }, + { + "bbox": [ + 429, + 579, + 438, + 590 + ], + "score": 0.89, + "content": "f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 578, + 457, + 591 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 457, + 580, + 468, + 591 + ], + "score": 0.86, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "have the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 589, + 411, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 175, + 603 + ], + "score": 1.0, + "content": "same span as the", + "type": "text" + }, + { + "bbox": [ + 176, + 590, + 186, + 601 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 589, + 347, + 603 + ], + "score": 1.0, + "content": "largest right and left singular vectors of", + "type": "text" + }, + { + "bbox": [ + 348, + 590, + 358, + 599 + ], + "score": 0.82, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 589, + 411, + 603 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 578, + 505, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 606, + 504, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 173, + 619 + ], + "score": 1.0, + "content": "For instance, if", + "type": "text" + }, + { + "bbox": [ + 173, + 606, + 202, + 618 + ], + "score": 0.93, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "is Gaussian, the canonical variables can be computed analytically, as in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "Lancaster (1958) or Beny (2018a) in the multivariate case. Solutions for other distributions were ´", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 629, + 248, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 248, + 641 + ], + "score": 1.0, + "content": "also computed in Eagleson (1964).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 606, + 506, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 679 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 331, + 658 + ], + "score": 1.0, + "content": "Notably, for any two dimensional Gaussian, the space of", + "type": "text" + }, + { + "bbox": [ + 331, + 646, + 338, + 655 + ], + "score": 0.84, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "most relevant variables is simply spanned", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 656, + 504, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 174, + 669 + ], + "score": 1.0, + "content": "by the moments", + "type": "text" + }, + { + "bbox": [ + 174, + 656, + 224, + 668 + ], + "score": 0.93, + "content": "f _ { n } ( x ) = x ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 657, + 243, + 669 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 243, + 656, + 292, + 668 + ], + "score": 0.92, + "content": "g _ { n } ( y ) = y ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 657, + 308, + 669 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 308, + 657, + 380, + 668 + ], + "score": 0.91, + "content": "n = 0 , \\ldots , k - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 657, + 497, + 669 + ], + "score": 1.0, + "content": ". Hence, in this case the first", + "type": "text" + }, + { + "bbox": [ + 497, + 657, + 504, + 666 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 667, + 494, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 298, + 680 + ], + "score": 1.0, + "content": "moments can be inferred exactly using only the", + "type": "text" + }, + { + "bbox": [ + 298, + 668, + 322, + 678 + ], + "score": 0.89, + "content": "k + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 667, + 494, + 680 + ], + "score": 1.0, + "content": "most relevant variables (See Appendix C).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 645, + 505, + 680 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 695, + 191, + 708 + ], + "lines": [ + { + "bbox": [ + 105, + 694, + 192, + 710 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 192, + 710 + ], + "score": 1.0, + "content": "4 ALGORITHM", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 357, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 359, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 359, + 734 + ], + "score": 1.0, + "content": "Let us explicit the algorithm resulting from the above analysis.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 719, + 359, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 319, + 96 + ], + "score": 1.0, + "content": "We assume that we are given independent samples", + "type": "text" + }, + { + "bbox": [ + 320, + 82, + 403, + 95 + ], + "score": 0.92, + "content": "( x _ { 1 } , y _ { 1 } ) , ( x _ { 1 } , y _ { 2 } ) , . . .", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "from the otherwise un-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 91, + 238, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 205, + 107 + ], + "score": 1.0, + "content": "known joint distribution", + "type": "text" + }, + { + "bbox": [ + 205, + 93, + 234, + 106 + ], + "score": 0.93, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 91, + 238, + 107 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 145 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "score": 1.0, + "content": "We first perform DCCA (Andrew et al., 2013). That is, we need two independent deterministic feed-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 275, + 135 + ], + "score": 1.0, + "content": "forward neural networks. The first maps", + "type": "text" + }, + { + "bbox": [ + 275, + 123, + 282, + 131 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 120, + 328, + 135 + ], + "score": 1.0, + "content": "to a set of", + "type": "text" + }, + { + "bbox": [ + 329, + 122, + 339, + 132 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 120, + 428, + 135 + ], + "score": 1.0, + "content": "real-valued variables", + "type": "text" + }, + { + "bbox": [ + 429, + 121, + 501, + 133 + ], + "score": 0.92, + "content": "f _ { 1 } ( x ) , \\ldots , f _ { k _ { 0 } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 120, + 505, + 135 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 389, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 178, + 145 + ], + "score": 1.0, + "content": "The second maps", + "type": "text" + }, + { + "bbox": [ + 178, + 134, + 185, + 144 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 132, + 264, + 145 + ], + "score": 1.0, + "content": "to a different set of", + "type": "text" + }, + { + "bbox": [ + 264, + 133, + 275, + 144 + ], + "score": 0.89, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 132, + 314, + 145 + ], + "score": 1.0, + "content": "variables", + "type": "text" + }, + { + "bbox": [ + 315, + 132, + 386, + 145 + ], + "score": 0.93, + "content": "g _ { 1 } ( y ) , \\ldots , g _ { k _ { 0 } } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 132, + 389, + 145 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 105, + 149, + 453, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 452, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 452, + 162 + ], + "score": 1.0, + "content": "The parameters of the neural networks are to be set to minimize the objective function", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 164, + 368, + 179 + ], + "lines": [ + { + "bbox": [ + 242, + 164, + 368, + 179 + ], + "spans": [ + { + "bbox": [ + 242, + 164, + 368, + 179 + ], + "score": 0.93, + "content": "C = k _ { 0 } - \\operatorname { T r } { \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) } ,", + "type": "interline_equation", + "image_path": "18fde4f521fc48a4fa7b3af4096adc86fcb6939f0c44a4dcc3bba5d1292cffc4.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 242, + 164, + 368, + 179 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 183, + 488, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 488, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 184, + 196 + ], + "score": 1.0, + "content": "where the matrices", + "type": "text" + }, + { + "bbox": [ + 184, + 183, + 217, + 195 + ], + "score": 0.91, + "content": "K , L , A", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 182, + 377, + 196 + ], + "score": 1.0, + "content": "can be approximated over a mini-batch", + "type": "text" + }, + { + "bbox": [ + 377, + 183, + 411, + 195 + ], + "score": 0.89, + "content": "( x _ { n } , y _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 182, + 415, + 196 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 416, + 183, + 471, + 195 + ], + "score": 0.82, + "content": "n = 1 , \\ldots , N", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 182, + 488, + 196 + ], + "score": 1.0, + "content": "via", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 114, + 198, + 483, + 233 + ], + "lines": [ + { + "bbox": [ + 114, + 198, + 483, + 233 + ], + "spans": [ + { + "bbox": [ + 114, + 198, + 483, + 233 + ], + "score": 0.94, + "content": "K _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } f _ { i } ( x _ { n } ) f _ { j } ( x _ { n } ) , \\quad L _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) g _ { j } ( y _ { n } ) , \\quad A _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) f _ { j } ( x _ { n } ) .", + "type": "interline_equation", + "image_path": "add296e57eaf2458c135e33a4e76c91636e008183e45eab3b6dc43aae8bb61df.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 114, + 198, + 483, + 209.66666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 114, + 209.66666666666666, + 483, + 221.33333333333331 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 114, + 221.33333333333331, + 483, + 232.99999999999997 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 241, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 399, + 255 + ], + "score": 1.0, + "content": "We found that, provided the batch size is sufficiently large compared to", + "type": "text" + }, + { + "bbox": [ + 399, + 243, + 410, + 253 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "(about 10 times in our", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "experience), this can be minimized using ADAM or direct gradient descent. However, to guarantee", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 213, + 276 + ], + "score": 1.0, + "content": "stability when using large", + "type": "text" + }, + { + "bbox": [ + 213, + 264, + 224, + 275 + ], + "score": 0.86, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 263, + 506, + 276 + ], + "score": 1.0, + "content": ", we needed to explicit the gradient of the objective function in order", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 345, + 287 + ], + "score": 1.0, + "content": "to force the use of the Moore-Penrose pseudo-inverses for", + "type": "text" + }, + { + "bbox": [ + 345, + 275, + 366, + 285 + ], + "score": 0.9, + "content": "K ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 274, + 385, + 287 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 386, + 275, + 404, + 285 + ], + "score": 0.9, + "content": "L ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "in both the forward and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 285, + 403, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 403, + 299 + ], + "score": 1.0, + "content": "backward passes, in addition to using 64 bits floats in these computations.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "Once the relevant variables have been learned, we still need to use the training data in a second", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 461, + 327 + ], + "score": 1.0, + "content": "step. Indeed, suppose that we wish to use our model to infer the value of some function", + "type": "text" + }, + { + "bbox": [ + 462, + 314, + 483, + 326 + ], + "score": 0.9, + "content": "\\Theta ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 313, + 505, + 327 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 451, + 338 + ], + "score": 1.0, + "content": "to compute its approximate expectation value in terms of the conditional distribution", + "type": "text" + }, + { + "bbox": [ + 451, + 325, + 501, + 337 + ], + "score": 0.93, + "content": "x \\mapsto p ( x | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 325, + 505, + 338 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 335, + 384, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 269, + 348 + ], + "score": 1.0, + "content": "Then we need to store, for each variable", + "type": "text" + }, + { + "bbox": [ + 269, + 336, + 324, + 348 + ], + "score": 0.92, + "content": "j = 1 , \\ldots , k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 335, + 384, + 348 + ], + "score": 1.0, + "content": ", the quantities", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 351, + 369, + 385 + ], + "lines": [ + { + "bbox": [ + 241, + 351, + 369, + 385 + ], + "spans": [ + { + "bbox": [ + 241, + 351, + 369, + 385 + ], + "score": 0.95, + "content": "\\Theta _ { j } = \\frac { 1 } { N _ { \\mathrm { f u l l } } } \\sum _ { n = 1 } ^ { N _ { \\mathrm { f u l l } } } \\Theta ( x _ { n } ) f _ { j } ( x _ { n } ) ,", + "type": "interline_equation", + "image_path": "5f2e2f9c14a534f7c9a30a260312b27d1541ae4a51197d6d82177cf1c9d784b2.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 351, + 369, + 368.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 241, + 368.0, + 369, + 385.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 388, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 380, + 401 + ], + "score": 1.0, + "content": "where the average is to be taken on the full training batch (of size", + "type": "text" + }, + { + "bbox": [ + 380, + 389, + 402, + 400 + ], + "score": 0.89, + "content": "N _ { \\mathrm { f u l l } } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 388, + 505, + 401 + ], + "score": 1.0, + "content": ". The same can be done", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 400, + 352, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 154, + 412 + ], + "score": 1.0, + "content": "exchanging", + "type": "text" + }, + { + "bbox": [ + 155, + 402, + 162, + 410 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 400, + 183, + 412 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 183, + 401, + 190, + 411 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 400, + 208, + 412 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 208, + 400, + 218, + 412 + ], + "score": 0.88, + "content": "f _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 400, + 239, + 412 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 239, + 401, + 250, + 412 + ], + "score": 0.86, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 400, + 352, + 412 + ], + "score": 1.0, + "content": "for the reverse inference.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 218, + 429 + ], + "score": 1.0, + "content": "For instance, if a data point", + "type": "text" + }, + { + "bbox": [ + 218, + 419, + 225, + 426 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 416, + 357, + 429 + ], + "score": 1.0, + "content": "is composed of real components", + "type": "text" + }, + { + "bbox": [ + 357, + 417, + 369, + 426 + ], + "score": 0.86, + "content": "x ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "—such as pixel color components", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 426, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 448, + 440 + ], + "score": 1.0, + "content": "for an image—and we are interested in the estimator which minimize the expected", + "type": "text" + }, + { + "bbox": [ + 448, + 427, + 457, + 437 + ], + "score": 0.85, + "content": "l ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 426, + 505, + 440 + ], + "score": 1.0, + "content": "distance to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "the predicted values of these components, then we need the expectation values of the components", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 107, + 449, + 154, + 461 + ], + "score": 0.92, + "content": "\\Theta ( x ) = x ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 448, + 182, + 462 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 183, + 452, + 189, + 459 + ], + "score": 0.72, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 448, + 505, + 462 + ], + "score": 1.0, + "content": ", and possibly higher moments to gain more knowledge about the shape of the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 380, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 311, + 473 + ], + "score": 1.0, + "content": "posterior distribution, such as the second moments", + "type": "text" + }, + { + "bbox": [ + 311, + 460, + 359, + 472 + ], + "score": 0.93, + "content": "\\Theta ^ { \\prime } ( \\bar { x } ) = x ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 460, + 380, + 473 + ], + "score": 1.0, + "content": ", etc.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 321, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 321, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 321, + 491 + ], + "score": 1.0, + "content": "Inference can then be performed with new data using", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 190, + 492, + 420, + 528 + ], + "lines": [ + { + "bbox": [ + 190, + 492, + 420, + 528 + ], + "spans": [ + { + "bbox": [ + 190, + 492, + 420, + 528 + ], + "score": 0.94, + "content": "\\overline { { { \\Theta } } } = \\sum _ { x } p ( x | y ) \\Theta ( x ) \\approx \\sum _ { i , j = 1 } ^ { k _ { 0 } } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } ) _ { j i } \\Theta _ { j } g _ { i } ( y ) .", + "type": "interline_equation", + "image_path": "e466bd2d915cdb5122b5c1773ab354787b45130d24c02c188710baa5586101e8.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 492, + 420, + 510.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 190, + 510.0, + 420, + 528.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 537, + 504, + 564 + ], + "lines": [ + { + "bbox": [ + 102, + 537, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 102, + 537, + 224, + 569 + ], + "score": 1.0, + "content": "Moreover, the accuracy of thof the relevant variables, i.e.,", + "type": "text" + }, + { + "bbox": [ + 315, + 537, + 359, + 569 + ], + "score": 1.0, + "content": "t depend on, for which", + "type": "text" + }, + { + "bbox": [ + 396, + 538, + 407, + 549 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 538, + 426, + 547 + ], + "score": 0.84, + "content": "\\Theta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 537, + 505, + 569 + ], + "score": 1.0, + "content": "taken in the span.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 225, + 549, + 434, + 563 + ], + "spans": [ + { + "bbox": [ + 225, + 549, + 315, + 563 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\Theta ( x ) = \\sum _ { i = 1 } ^ { k } c _ { i } f _ { i } ( x ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 549, + 434, + 563 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\Theta _ { j } = \\sum _ { i = 1 } ^ { k _ { 0 } } c _ { i } K _ { i j } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 568, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 397, + 582 + ], + "score": 1.0, + "content": "The reverse inference formulas are obtained simply by the exchanges", + "type": "text" + }, + { + "bbox": [ + 397, + 569, + 434, + 579 + ], + "score": 0.68, + "content": "K L", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 567, + 439, + 582 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 440, + 568, + 482, + 579 + ], + "score": 0.82, + "content": "A A ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 567, + 506, + 582 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 579, + 144, + 594 + ], + "spans": [ + { + "bbox": [ + 107, + 580, + 139, + 592 + ], + "score": 0.92, + "content": "g _ { i } f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 579, + 144, + 594 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 107, + 606, + 201, + 620 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 201, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 201, + 622 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 504, + 654 + ], + "lines": [ + { + "bbox": [ + 106, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "In all our experiments, we used the ADAM optimizer with learning rate 0.001. We used the Flux", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 325, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 325, + 655 + ], + "score": 1.0, + "content": "package (Innes, 2018) for Julia, as well as Tensorflow.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "As usual the data is divided into a training set and a testing set. No aspect of the testing set is used", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "during training. The loss function refers to Eq. (7). In order to monitor overfitting, we compute a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "score": 1.0, + "content": "“test loss” and a “training loss”. The test loss is computed from the trained variables using only the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 693, + 452, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 452, + 705 + ], + "score": 1.0, + "content": "test data, and accordingly, the training loss is computed purely using the training data.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 470, + 722 + ], + "score": 1.0, + "content": "Moreover, when performing inference on test data using Eq. (10), we use the covariances", + "type": "text" + }, + { + "bbox": [ + 470, + 709, + 504, + 721 + ], + "score": 0.89, + "content": "A , L , K", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 720, + 363, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 175, + 733 + ], + "score": 1.0, + "content": "and expectations", + "type": "text" + }, + { + "bbox": [ + 176, + 721, + 189, + 733 + ], + "score": 0.89, + "content": "\\Theta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 720, + 363, + 733 + ], + "score": 1.0, + "content": "(Equ. (9)) built from the training data only.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 319, + 96 + ], + "score": 1.0, + "content": "We assume that we are given independent samples", + "type": "text" + }, + { + "bbox": [ + 320, + 82, + 403, + 95 + ], + "score": 0.92, + "content": "( x _ { 1 } , y _ { 1 } ) , ( x _ { 1 } , y _ { 2 } ) , . . .", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "from the otherwise un-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 91, + 238, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 205, + 107 + ], + "score": 1.0, + "content": "known joint distribution", + "type": "text" + }, + { + "bbox": [ + 205, + 93, + 234, + 106 + ], + "score": 0.93, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 91, + 238, + 107 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 107 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 145 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 504, + 122 + ], + "score": 1.0, + "content": "We first perform DCCA (Andrew et al., 2013). That is, we need two independent deterministic feed-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 120, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 275, + 135 + ], + "score": 1.0, + "content": "forward neural networks. The first maps", + "type": "text" + }, + { + "bbox": [ + 275, + 123, + 282, + 131 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 120, + 328, + 135 + ], + "score": 1.0, + "content": "to a set of", + "type": "text" + }, + { + "bbox": [ + 329, + 122, + 339, + 132 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 120, + 428, + 135 + ], + "score": 1.0, + "content": "real-valued variables", + "type": "text" + }, + { + "bbox": [ + 429, + 121, + 501, + 133 + ], + "score": 0.92, + "content": "f _ { 1 } ( x ) , \\ldots , f _ { k _ { 0 } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 120, + 505, + 135 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 389, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 178, + 145 + ], + "score": 1.0, + "content": "The second maps", + "type": "text" + }, + { + "bbox": [ + 178, + 134, + 185, + 144 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 132, + 264, + 145 + ], + "score": 1.0, + "content": "to a different set of", + "type": "text" + }, + { + "bbox": [ + 264, + 133, + 275, + 144 + ], + "score": 0.89, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 132, + 314, + 145 + ], + "score": 1.0, + "content": "variables", + "type": "text" + }, + { + "bbox": [ + 315, + 132, + 386, + 145 + ], + "score": 0.93, + "content": "g _ { 1 } ( y ) , \\ldots , g _ { k _ { 0 } } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 132, + 389, + 145 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 110, + 505, + 145 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 149, + 453, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 452, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 452, + 162 + ], + "score": 1.0, + "content": "The parameters of the neural networks are to be set to minimize the objective function", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 148, + 452, + 162 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 164, + 368, + 179 + ], + "lines": [ + { + "bbox": [ + 242, + 164, + 368, + 179 + ], + "spans": [ + { + "bbox": [ + 242, + 164, + 368, + 179 + ], + "score": 0.93, + "content": "C = k _ { 0 } - \\operatorname { T r } { \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) } ,", + "type": "interline_equation", + "image_path": "18fde4f521fc48a4fa7b3af4096adc86fcb6939f0c44a4dcc3bba5d1292cffc4.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 242, + 164, + 368, + 179 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 183, + 488, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 488, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 184, + 196 + ], + "score": 1.0, + "content": "where the matrices", + "type": "text" + }, + { + "bbox": [ + 184, + 183, + 217, + 195 + ], + "score": 0.91, + "content": "K , L , A", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 182, + 377, + 196 + ], + "score": 1.0, + "content": "can be approximated over a mini-batch", + "type": "text" + }, + { + "bbox": [ + 377, + 183, + 411, + 195 + ], + "score": 0.89, + "content": "( x _ { n } , y _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 182, + 415, + 196 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 416, + 183, + 471, + 195 + ], + "score": 0.82, + "content": "n = 1 , \\ldots , N", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 182, + 488, + 196 + ], + "score": 1.0, + "content": "via", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 182, + 488, + 196 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 114, + 198, + 483, + 233 + ], + "lines": [ + { + "bbox": [ + 114, + 198, + 483, + 233 + ], + "spans": [ + { + "bbox": [ + 114, + 198, + 483, + 233 + ], + "score": 0.94, + "content": "K _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } f _ { i } ( x _ { n } ) f _ { j } ( x _ { n } ) , \\quad L _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) g _ { j } ( y _ { n } ) , \\quad A _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) f _ { j } ( x _ { n } ) .", + "type": "interline_equation", + "image_path": "add296e57eaf2458c135e33a4e76c91636e008183e45eab3b6dc43aae8bb61df.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 114, + 198, + 483, + 209.66666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 114, + 209.66666666666666, + 483, + 221.33333333333331 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 114, + 221.33333333333331, + 483, + 232.99999999999997 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 241, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 106, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 399, + 255 + ], + "score": 1.0, + "content": "We found that, provided the batch size is sufficiently large compared to", + "type": "text" + }, + { + "bbox": [ + 399, + 243, + 410, + 253 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "(about 10 times in our", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "experience), this can be minimized using ADAM or direct gradient descent. However, to guarantee", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 213, + 276 + ], + "score": 1.0, + "content": "stability when using large", + "type": "text" + }, + { + "bbox": [ + 213, + 264, + 224, + 275 + ], + "score": 0.86, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 263, + 506, + 276 + ], + "score": 1.0, + "content": ", we needed to explicit the gradient of the objective function in order", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 345, + 287 + ], + "score": 1.0, + "content": "to force the use of the Moore-Penrose pseudo-inverses for", + "type": "text" + }, + { + "bbox": [ + 345, + 275, + 366, + 285 + ], + "score": 0.9, + "content": "K ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 274, + 385, + 287 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 386, + 275, + 404, + 285 + ], + "score": 0.9, + "content": "L ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "in both the forward and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 285, + 403, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 403, + 299 + ], + "score": 1.0, + "content": "backward passes, in addition to using 64 bits floats in these computations.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 241, + 506, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "Once the relevant variables have been learned, we still need to use the training data in a second", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 461, + 327 + ], + "score": 1.0, + "content": "step. Indeed, suppose that we wish to use our model to infer the value of some function", + "type": "text" + }, + { + "bbox": [ + 462, + 314, + 483, + 326 + ], + "score": 0.9, + "content": "\\Theta ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 313, + 505, + 327 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 451, + 338 + ], + "score": 1.0, + "content": "to compute its approximate expectation value in terms of the conditional distribution", + "type": "text" + }, + { + "bbox": [ + 451, + 325, + 501, + 337 + ], + "score": 0.93, + "content": "x \\mapsto p ( x | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 325, + 505, + 338 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 335, + 384, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 269, + 348 + ], + "score": 1.0, + "content": "Then we need to store, for each variable", + "type": "text" + }, + { + "bbox": [ + 269, + 336, + 324, + 348 + ], + "score": 0.92, + "content": "j = 1 , \\ldots , k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 335, + 384, + 348 + ], + "score": 1.0, + "content": ", the quantities", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 302, + 505, + 348 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 351, + 369, + 385 + ], + "lines": [ + { + "bbox": [ + 241, + 351, + 369, + 385 + ], + "spans": [ + { + "bbox": [ + 241, + 351, + 369, + 385 + ], + "score": 0.95, + "content": "\\Theta _ { j } = \\frac { 1 } { N _ { \\mathrm { f u l l } } } \\sum _ { n = 1 } ^ { N _ { \\mathrm { f u l l } } } \\Theta ( x _ { n } ) f _ { j } ( x _ { n } ) ,", + "type": "interline_equation", + "image_path": "5f2e2f9c14a534f7c9a30a260312b27d1541ae4a51197d6d82177cf1c9d784b2.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 351, + 369, + 368.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 241, + 368.0, + 369, + 385.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 388, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 380, + 401 + ], + "score": 1.0, + "content": "where the average is to be taken on the full training batch (of size", + "type": "text" + }, + { + "bbox": [ + 380, + 389, + 402, + 400 + ], + "score": 0.89, + "content": "N _ { \\mathrm { f u l l } } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 388, + 505, + 401 + ], + "score": 1.0, + "content": ". The same can be done", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 400, + 352, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 154, + 412 + ], + "score": 1.0, + "content": "exchanging", + "type": "text" + }, + { + "bbox": [ + 155, + 402, + 162, + 410 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 400, + 183, + 412 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 183, + 401, + 190, + 411 + ], + "score": 0.81, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 400, + 208, + 412 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 208, + 400, + 218, + 412 + ], + "score": 0.88, + "content": "f _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 400, + 239, + 412 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 239, + 401, + 250, + 412 + ], + "score": 0.86, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 400, + 352, + 412 + ], + "score": 1.0, + "content": "for the reverse inference.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 388, + 505, + 412 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 218, + 429 + ], + "score": 1.0, + "content": "For instance, if a data point", + "type": "text" + }, + { + "bbox": [ + 218, + 419, + 225, + 426 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 416, + 357, + 429 + ], + "score": 1.0, + "content": "is composed of real components", + "type": "text" + }, + { + "bbox": [ + 357, + 417, + 369, + 426 + ], + "score": 0.86, + "content": "x ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "—such as pixel color components", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 426, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 448, + 440 + ], + "score": 1.0, + "content": "for an image—and we are interested in the estimator which minimize the expected", + "type": "text" + }, + { + "bbox": [ + 448, + 427, + 457, + 437 + ], + "score": 0.85, + "content": "l ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 426, + 505, + 440 + ], + "score": 1.0, + "content": "distance to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "the predicted values of these components, then we need the expectation values of the components", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 107, + 449, + 154, + 461 + ], + "score": 0.92, + "content": "\\Theta ( x ) = x ^ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 448, + 182, + 462 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 183, + 452, + 189, + 459 + ], + "score": 0.72, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 448, + 505, + 462 + ], + "score": 1.0, + "content": ", and possibly higher moments to gain more knowledge about the shape of the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 380, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 311, + 473 + ], + "score": 1.0, + "content": "posterior distribution, such as the second moments", + "type": "text" + }, + { + "bbox": [ + 311, + 460, + 359, + 472 + ], + "score": 0.93, + "content": "\\Theta ^ { \\prime } ( \\bar { x } ) = x ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 460, + 380, + 473 + ], + "score": 1.0, + "content": ", etc.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 416, + 505, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 321, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 321, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 321, + 491 + ], + "score": 1.0, + "content": "Inference can then be performed with new data using", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 475, + 321, + 491 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 190, + 492, + 420, + 528 + ], + "lines": [ + { + "bbox": [ + 190, + 492, + 420, + 528 + ], + "spans": [ + { + "bbox": [ + 190, + 492, + 420, + 528 + ], + "score": 0.94, + "content": "\\overline { { { \\Theta } } } = \\sum _ { x } p ( x | y ) \\Theta ( x ) \\approx \\sum _ { i , j = 1 } ^ { k _ { 0 } } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } ) _ { j i } \\Theta _ { j } g _ { i } ( y ) .", + "type": "interline_equation", + "image_path": "e466bd2d915cdb5122b5c1773ab354787b45130d24c02c188710baa5586101e8.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 492, + 420, + 510.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 190, + 510.0, + 420, + 528.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 537, + 504, + 564 + ], + "lines": [ + { + "bbox": [ + 102, + 537, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 102, + 537, + 224, + 569 + ], + "score": 1.0, + "content": "Moreover, the accuracy of thof the relevant variables, i.e.,", + "type": "text" + }, + { + "bbox": [ + 315, + 537, + 359, + 569 + ], + "score": 1.0, + "content": "t depend on, for which", + "type": "text" + }, + { + "bbox": [ + 396, + 538, + 407, + 549 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 538, + 426, + 547 + ], + "score": 0.84, + "content": "\\Theta", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 537, + 505, + 569 + ], + "score": 1.0, + "content": "taken in the span.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 225, + 549, + 434, + 563 + ], + "spans": [ + { + "bbox": [ + 225, + 549, + 315, + 563 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\Theta ( x ) = \\sum _ { i = 1 } ^ { k } c _ { i } f _ { i } ( x ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 549, + 434, + 563 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\Theta _ { j } = \\sum _ { i = 1 } ^ { k _ { 0 } } c _ { i } K _ { i j } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 102, + 537, + 505, + 569 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 568, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 397, + 582 + ], + "score": 1.0, + "content": "The reverse inference formulas are obtained simply by the exchanges", + "type": "text" + }, + { + "bbox": [ + 397, + 569, + 434, + 579 + ], + "score": 0.68, + "content": "K L", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 567, + 439, + 582 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 440, + 568, + 482, + 579 + ], + "score": 0.82, + "content": "A A ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 567, + 506, + 582 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 579, + 144, + 594 + ], + "spans": [ + { + "bbox": [ + 107, + 580, + 139, + 592 + ], + "score": 0.92, + "content": "g _ { i } f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 579, + 144, + 594 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 567, + 506, + 594 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 606, + 201, + 620 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 201, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 201, + 622 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 504, + 654 + ], + "lines": [ + { + "bbox": [ + 106, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "In all our experiments, we used the ADAM optimizer with learning rate 0.001. We used the Flux", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 325, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 325, + 655 + ], + "score": 1.0, + "content": "package (Innes, 2018) for Julia, as well as Tensorflow.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 632, + 505, + 655 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "As usual the data is divided into a training set and a testing set. No aspect of the testing set is used", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "during training. The loss function refers to Eq. (7). In order to monitor overfitting, we compute a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 505, + 694 + ], + "score": 1.0, + "content": "“test loss” and a “training loss”. The test loss is computed from the trained variables using only the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 693, + 452, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 452, + 705 + ], + "score": 1.0, + "content": "test data, and accordingly, the training loss is computed purely using the training data.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 659, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 470, + 722 + ], + "score": 1.0, + "content": "Moreover, when performing inference on test data using Eq. (10), we use the covariances", + "type": "text" + }, + { + "bbox": [ + 470, + 709, + 504, + 721 + ], + "score": 0.89, + "content": "A , L , K", + "type": "inline_equation" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 720, + 363, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 175, + 733 + ], + "score": 1.0, + "content": "and expectations", + "type": "text" + }, + { + "bbox": [ + 176, + 721, + 189, + 733 + ], + "score": 0.89, + "content": "\\Theta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 720, + 363, + 733 + ], + "score": 1.0, + "content": "(Equ. (9)) built from the training data only.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 709, + 504, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 280, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 282, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 282, + 95 + ], + "score": 1.0, + "content": "5.1 INFERENCE ON OCCLUDED MNIST", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 503, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "In this experiment, we use the left and right halves of the MNIST digit images as correlated variables", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 120, + 475, + 136 + ], + "spans": [ + { + "bbox": [ + 107, + 122, + 117, + 132 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 120, + 135, + 136 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 135, + 122, + 144, + 132 + ], + "score": 0.45, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 120, + 475, + 136 + ], + "score": 1.0, + "content": ". The goal is to obtain the expected left halves given the right halves, or vice versa.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 108, + 138, + 503, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "score": 1.0, + "content": "The training set was augmented by random small rotations and displacements to make the task more", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 150, + 376, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 376, + 162 + ], + "score": 1.0, + "content": "ambiguous, as we want to explore the uncertainty in the prediction.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 505, + 180 + ], + "score": 1.0, + "content": "The relevant variables were represented by two convolutional neural networks of identical architec-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "ture. They are composed of four convolutional layers and one fully connected layer, an architecture", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 189, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 505, + 201 + ], + "score": 1.0, + "content": "that performs well for supervised learning on this dataset. For ease of implementation, these CNN", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 198, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 212 + ], + "score": 1.0, + "content": "have the whole image as input, but with either half zeroed (same value as black pixels). Half-width", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 209, + 330, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 330, + 223 + ], + "score": 1.0, + "content": "CNNs with proper padding at the cut perform similarly.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 108, + 227, + 503, + 250 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "After training, we used the training dataset to also compute the expected pixel gray value as well as", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 237, + 332, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 332, + 251 + ], + "score": 1.0, + "content": "their covariance for each relevant variable using Eq. (9).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 255, + 505, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "score": 1.0, + "content": "These were used into Eq. (10) to compute the mean pixel gray values and their covariances over the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 265, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 213, + 277 + ], + "score": 1.0, + "content": "conditional probability of", + "type": "text" + }, + { + "bbox": [ + 213, + 267, + 223, + 276 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 265, + 249, + 277 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 250, + 266, + 259, + 276 + ], + "score": 0.77, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 265, + 495, + 277 + ], + "score": 1.0, + "content": "on test data. This mean is the Bayesian estimator for the", + "type": "text" + }, + { + "bbox": [ + 495, + 266, + 504, + 276 + ], + "score": 0.81, + "content": "l ^ { 2 }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 408, + 289 + ], + "score": 1.0, + "content": "distance between half images, i.e., it should minimize the expected distance", + "type": "text" + }, + { + "bbox": [ + 409, + 277, + 422, + 289 + ], + "score": 0.88, + "content": "d _ { l ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "over the conditional", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 287, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 185, + 302 + ], + "score": 1.0, + "content": "distribution, where", + "type": "text" + }, + { + "bbox": [ + 185, + 288, + 291, + 301 + ], + "score": 0.92, + "content": "\\begin{array} { r } { d _ { l ^ { 2 } } ^ { 2 } ( x , \\overline { { y } } ) = \\sum _ { i } ( x _ { i } - y _ { i } ) ^ { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 287, + 322, + 302 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 322, + 288, + 366, + 300 + ], + "score": 0.92, + "content": "x _ { i } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 287, + 441, + 302 + ], + "score": 1.0, + "content": "is the value of the", + "type": "text" + }, + { + "bbox": [ + 441, + 288, + 454, + 298 + ], + "score": 0.85, + "content": "\\mathrm { i } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 287, + 506, + 302 + ], + "score": 1.0, + "content": "pixel. (This", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 299, + 311, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 311, + 311 + ], + "score": 1.0, + "content": "is equivalent to minimizing the mean square error).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 316, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 505, + 328 + ], + "score": 1.0, + "content": "The results on a randomly selected subset of test digits is shown in Fig. 1. For each example, we also", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "computed the images obtained by adding plus or minus one standard deviation along the direction", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "of greatest variance in the space of relevant variables. This reveals the main ambiguities (such as", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 348, + 349, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 349, + 362 + ], + "score": 1.0, + "content": "between 8 and 3 or 7 and 9 which share a similar right half).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "The graph of the singular values shows that the rank cutoff of 200 is too low to capture all of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "relevant variables (the sudden drop at the end is not robust to an increase in the cutoff), but the re-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "sults are reasonable nevertheless. This shows that our representation of the conditional distributions", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 398, + 330, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 330, + 411 + ], + "score": 1.0, + "content": "contains valuable information besides the simple mean.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 108, + 443, + 232, + 454 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 234, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 234, + 456 + ], + "score": 1.0, + "content": "5.2 SUPERVISED LEARNING", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "In the context of a supervised classification task, one of the dataset (the labels) is of sufficiently low", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "dimensionality that we can use a complete basis over its probability space as our relevant variables,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "such as the standard one-hot encoding of labels. This serves as a good first sanity test for our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "approach. Surprisingly, we find that it converges faster than standard approaches, and without the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 514, + 201, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 201, + 529 + ], + "score": 1.0, + "content": "need for regularization.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 171, + 545 + ], + "score": 1.0, + "content": "Let the variable", + "type": "text" + }, + { + "bbox": [ + 171, + 532, + 180, + 542 + ], + "score": 0.75, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 531, + 324, + 545 + ], + "score": 1.0, + "content": "stands for the labels, with values in", + "type": "text" + }, + { + "bbox": [ + 324, + 532, + 367, + 544 + ], + "score": 0.94, + "content": "\\{ 1 , \\ldots , k \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 531, + 506, + 545 + ], + "score": 1.0, + "content": ". The probability space consists of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 542, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 158, + 557 + ], + "score": 1.0, + "content": "vectors with", + "type": "text" + }, + { + "bbox": [ + 159, + 544, + 165, + 553 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 542, + 471, + 557 + ], + "score": 1.0, + "content": "real components. The canonical basis corresponds to the one-hot encoding", + "type": "text" + }, + { + "bbox": [ + 471, + 543, + 505, + 555 + ], + "score": 0.91, + "content": "g _ { i } ( j ) =", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 107, + 554, + 119, + 567 + ], + "score": 0.88, + "content": "\\delta _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 554, + 365, + 566 + ], + "score": 1.0, + "content": "(Kronecker delta). All we need is a neural network to encode", + "type": "text" + }, + { + "bbox": [ + 365, + 555, + 372, + 564 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 554, + 411, + 566 + ], + "score": 1.0, + "content": "variables", + "type": "text" + }, + { + "bbox": [ + 411, + 554, + 453, + 565 + ], + "score": 0.91, + "content": "f _ { 1 } , \\ldots , f _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 554, + 467, + 566 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 467, + 555, + 477, + 564 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 554, + 506, + 566 + ], + "score": 1.0, + "content": ". After", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 250, + 578 + ], + "score": 1.0, + "content": "learning the most relevant variables", + "type": "text" + }, + { + "bbox": [ + 251, + 565, + 259, + 577 + ], + "score": 0.87, + "content": "f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 564, + 444, + 578 + ], + "score": 1.0, + "content": ", we apply the reverse of Eq. (10) for function", + "type": "text" + }, + { + "bbox": [ + 444, + 565, + 484, + 577 + ], + "score": 0.92, + "content": "\\Theta ( y ) = y", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 564, + 506, + 578 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 576, + 437, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 299, + 588 + ], + "score": 1.0, + "content": "use the maximum component of expected value", + "type": "text" + }, + { + "bbox": [ + 299, + 577, + 306, + 588 + ], + "score": 0.82, + "content": "\\overline { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 576, + 437, + 588 + ], + "score": 1.0, + "content": "to infer the labels from the data.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 592, + 453, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 454, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 454, + 606 + ], + "score": 1.0, + "content": "Let us refer to this procedure as DCCI (Deep Canonical Correlations based Inference).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 609, + 502, + 632 + ], + "lines": [ + { + "bbox": [ + 107, + 609, + 504, + 622 + ], + "spans": [ + { + "bbox": [ + 107, + 609, + 504, + 622 + ], + "score": 1.0, + "content": "We tested this approach on the MNIST and CIFAR10 datasets, and compared the results to the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 621, + 272, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 272, + 634 + ], + "score": 1.0, + "content": "standard cross-entropy objective (Fig. 2).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "We plotted the accuracy as function of the epoch rather than clock time which would depend on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "many factors. But the time per epoch is roughly the same for each approaches in the above experi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "ments. Indeed, the training time is dominated by the forward and backward evaluations of the neural", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 669, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 505, + 685 + ], + "score": 1.0, + "content": "networks which are identical. (However, the time it takes to evaluate our objective can become sig-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 273, + 694 + ], + "score": 1.0, + "content": "nificant for much larger number of labels", + "type": "text" + }, + { + "bbox": [ + 273, + 683, + 279, + 691 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 681, + 506, + 694 + ], + "score": 1.0, + "content": ", since it involves the inversion of matrices of dimension", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 692, + 473, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 113, + 703 + ], + "score": 0.72, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 692, + 473, + 705 + ], + "score": 1.0, + "content": ". This is in addition to the fact that a greater dimension would require also larger batches.)", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "We found that, without regularization, simply changing the objective from cross-entropy to DCCI", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 480, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 480, + 733 + ], + "score": 1.0, + "content": "provided a large improvement both of convergence speed and final accuracy for both models.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 280, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 282, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 282, + 95 + ], + "score": 1.0, + "content": "5.1 INFERENCE ON OCCLUDED MNIST", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 503, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "In this experiment, we use the left and right halves of the MNIST digit images as correlated variables", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 120, + 475, + 136 + ], + "spans": [ + { + "bbox": [ + 107, + 122, + 117, + 132 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 120, + 135, + 136 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 135, + 122, + 144, + 132 + ], + "score": 0.45, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 120, + 475, + 136 + ], + "score": 1.0, + "content": ". The goal is to obtain the expected left halves given the right halves, or vice versa.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 111, + 505, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 138, + 503, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 504, + 151 + ], + "score": 1.0, + "content": "The training set was augmented by random small rotations and displacements to make the task more", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 150, + 376, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 376, + 162 + ], + "score": 1.0, + "content": "ambiguous, as we want to explore the uncertainty in the prediction.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 138, + 504, + 162 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 505, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 505, + 180 + ], + "score": 1.0, + "content": "The relevant variables were represented by two convolutional neural networks of identical architec-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "ture. They are composed of four convolutional layers and one fully connected layer, an architecture", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 189, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 505, + 201 + ], + "score": 1.0, + "content": "that performs well for supervised learning on this dataset. For ease of implementation, these CNN", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 198, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 212 + ], + "score": 1.0, + "content": "have the whole image as input, but with either half zeroed (same value as black pixels). Half-width", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 209, + 330, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 330, + 223 + ], + "score": 1.0, + "content": "CNNs with proper padding at the cut perform similarly.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 166, + 505, + 223 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 227, + 503, + 250 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "After training, we used the training dataset to also compute the expected pixel gray value as well as", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 237, + 332, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 332, + 251 + ], + "score": 1.0, + "content": "their covariance for each relevant variable using Eq. (9).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 226, + 505, + 251 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 255, + 505, + 311 + ], + "lines": [ + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "score": 1.0, + "content": "These were used into Eq. (10) to compute the mean pixel gray values and their covariances over the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 265, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 213, + 277 + ], + "score": 1.0, + "content": "conditional probability of", + "type": "text" + }, + { + "bbox": [ + 213, + 267, + 223, + 276 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 265, + 249, + 277 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 250, + 266, + 259, + 276 + ], + "score": 0.77, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 265, + 495, + 277 + ], + "score": 1.0, + "content": "on test data. This mean is the Bayesian estimator for the", + "type": "text" + }, + { + "bbox": [ + 495, + 266, + 504, + 276 + ], + "score": 0.81, + "content": "l ^ { 2 }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 277, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 408, + 289 + ], + "score": 1.0, + "content": "distance between half images, i.e., it should minimize the expected distance", + "type": "text" + }, + { + "bbox": [ + 409, + 277, + 422, + 289 + ], + "score": 0.88, + "content": "d _ { l ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 277, + 505, + 289 + ], + "score": 1.0, + "content": "over the conditional", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 287, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 185, + 302 + ], + "score": 1.0, + "content": "distribution, where", + "type": "text" + }, + { + "bbox": [ + 185, + 288, + 291, + 301 + ], + "score": 0.92, + "content": "\\begin{array} { r } { d _ { l ^ { 2 } } ^ { 2 } ( x , \\overline { { y } } ) = \\sum _ { i } ( x _ { i } - y _ { i } ) ^ { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 287, + 322, + 302 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 322, + 288, + 366, + 300 + ], + "score": 0.92, + "content": "x _ { i } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 287, + 441, + 302 + ], + "score": 1.0, + "content": "is the value of the", + "type": "text" + }, + { + "bbox": [ + 441, + 288, + 454, + 298 + ], + "score": 0.85, + "content": "\\mathrm { i } ^ { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 287, + 506, + 302 + ], + "score": 1.0, + "content": "pixel. (This", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 299, + 311, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 311, + 311 + ], + "score": 1.0, + "content": "is equivalent to minimizing the mean square error).", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 255, + 506, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 316, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 505, + 328 + ], + "score": 1.0, + "content": "The results on a randomly selected subset of test digits is shown in Fig. 1. For each example, we also", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "computed the images obtained by adding plus or minus one standard deviation along the direction", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "of greatest variance in the space of relevant variables. This reveals the main ambiguities (such as", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 348, + 349, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 349, + 362 + ], + "score": 1.0, + "content": "between 8 and 3 or 7 and 9 which share a similar right half).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 316, + 506, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "The graph of the singular values shows that the rank cutoff of 200 is too low to capture all of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "relevant variables (the sudden drop at the end is not robust to an increase in the cutoff), but the re-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 505, + 400 + ], + "score": 1.0, + "content": "sults are reasonable nevertheless. This shows that our representation of the conditional distributions", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 398, + 330, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 330, + 411 + ], + "score": 1.0, + "content": "contains valuable information besides the simple mean.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 366, + 505, + 411 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 443, + 232, + 454 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 234, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 234, + 456 + ], + "score": 1.0, + "content": "5.2 SUPERVISED LEARNING", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "In the context of a supervised classification task, one of the dataset (the labels) is of sufficiently low", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "dimensionality that we can use a complete basis over its probability space as our relevant variables,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "such as the standard one-hot encoding of labels. This serves as a good first sanity test for our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "approach. Surprisingly, we find that it converges faster than standard approaches, and without the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 514, + 201, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 201, + 529 + ], + "score": 1.0, + "content": "need for regularization.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 471, + 506, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 171, + 545 + ], + "score": 1.0, + "content": "Let the variable", + "type": "text" + }, + { + "bbox": [ + 171, + 532, + 180, + 542 + ], + "score": 0.75, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 531, + 324, + 545 + ], + "score": 1.0, + "content": "stands for the labels, with values in", + "type": "text" + }, + { + "bbox": [ + 324, + 532, + 367, + 544 + ], + "score": 0.94, + "content": "\\{ 1 , \\ldots , k \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 531, + 506, + 545 + ], + "score": 1.0, + "content": ". The probability space consists of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 542, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 158, + 557 + ], + "score": 1.0, + "content": "vectors with", + "type": "text" + }, + { + "bbox": [ + 159, + 544, + 165, + 553 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 542, + 471, + 557 + ], + "score": 1.0, + "content": "real components. The canonical basis corresponds to the one-hot encoding", + "type": "text" + }, + { + "bbox": [ + 471, + 543, + 505, + 555 + ], + "score": 0.91, + "content": "g _ { i } ( j ) =", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 554, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 107, + 554, + 119, + 567 + ], + "score": 0.88, + "content": "\\delta _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 554, + 365, + 566 + ], + "score": 1.0, + "content": "(Kronecker delta). All we need is a neural network to encode", + "type": "text" + }, + { + "bbox": [ + 365, + 555, + 372, + 564 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 554, + 411, + 566 + ], + "score": 1.0, + "content": "variables", + "type": "text" + }, + { + "bbox": [ + 411, + 554, + 453, + 565 + ], + "score": 0.91, + "content": "f _ { 1 } , \\ldots , f _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 554, + 467, + 566 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 467, + 555, + 477, + 564 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 554, + 506, + 566 + ], + "score": 1.0, + "content": ". After", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 250, + 578 + ], + "score": 1.0, + "content": "learning the most relevant variables", + "type": "text" + }, + { + "bbox": [ + 251, + 565, + 259, + 577 + ], + "score": 0.87, + "content": "f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 564, + 444, + 578 + ], + "score": 1.0, + "content": ", we apply the reverse of Eq. (10) for function", + "type": "text" + }, + { + "bbox": [ + 444, + 565, + 484, + 577 + ], + "score": 0.92, + "content": "\\Theta ( y ) = y", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 564, + 506, + 578 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 576, + 437, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 299, + 588 + ], + "score": 1.0, + "content": "use the maximum component of expected value", + "type": "text" + }, + { + "bbox": [ + 299, + 577, + 306, + 588 + ], + "score": 0.82, + "content": "\\overline { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 576, + 437, + 588 + ], + "score": 1.0, + "content": "to infer the labels from the data.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 531, + 506, + 588 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 592, + 453, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 454, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 454, + 606 + ], + "score": 1.0, + "content": "Let us refer to this procedure as DCCI (Deep Canonical Correlations based Inference).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 592, + 454, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 609, + 502, + 632 + ], + "lines": [ + { + "bbox": [ + 107, + 609, + 504, + 622 + ], + "spans": [ + { + "bbox": [ + 107, + 609, + 504, + 622 + ], + "score": 1.0, + "content": "We tested this approach on the MNIST and CIFAR10 datasets, and compared the results to the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 621, + 272, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 272, + 634 + ], + "score": 1.0, + "content": "standard cross-entropy objective (Fig. 2).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 609, + 504, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "We plotted the accuracy as function of the epoch rather than clock time which would depend on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "many factors. But the time per epoch is roughly the same for each approaches in the above experi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "ments. Indeed, the training time is dominated by the forward and backward evaluations of the neural", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 669, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 505, + 685 + ], + "score": 1.0, + "content": "networks which are identical. (However, the time it takes to evaluate our objective can become sig-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 681, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 273, + 694 + ], + "score": 1.0, + "content": "nificant for much larger number of labels", + "type": "text" + }, + { + "bbox": [ + 273, + 683, + 279, + 691 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 681, + 506, + 694 + ], + "score": 1.0, + "content": ", since it involves the inversion of matrices of dimension", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 692, + 473, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 113, + 703 + ], + "score": 0.72, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 692, + 473, + 705 + ], + "score": 1.0, + "content": ". This is in addition to the fact that a greater dimension would require also larger batches.)", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 104, + 638, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "We found that, without regularization, simply changing the objective from cross-entropy to DCCI", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 480, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 480, + 733 + ], + "score": 1.0, + "content": "provided a large improvement both of convergence speed and final accuracy for both models.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 92, + 509, + 331 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 92, + 509, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 92, + 509, + 331 + ], + "spans": [ + { + "bbox": [ + 112, + 92, + 509, + 331 + ], + "score": 0.89, + "type": "image", + "image_path": "201f3ce00c47dbafcb9ef33dbcddfe321ec4b0f6d68f48039079f27a17681d75.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 92, + 509, + 171.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 171.66666666666669, + 509, + 251.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 251.33333333333337, + 509, + 331.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 345, + 506, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "Figure 1: Left mosaic: the left halves of the MNIST digits in a random sample from the test set are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 268, + 369 + ], + "score": 1.0, + "content": "inferred from the right half, with cutoff", + "type": "text" + }, + { + "bbox": [ + 268, + 357, + 308, + 368 + ], + "score": 0.92, + "content": "k _ { 0 } = 2 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 358, + 505, + 369 + ], + "score": 1.0, + "content": ". Three images are shown for each digit. Within", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "each triplet, the middle image represents the mean over pixel intensity of the inferred condition", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "distribution, while left and right images corresponds to a plus and minus one standard deviation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "score": 1.0, + "content": "from the mean in the direction of largest covariance (in the space of half-images). A particularly", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 378, + 414 + ], + "score": 1.0, + "content": "interesting example is highlighted. Top-right: loss per epoch for", + "type": "text" + }, + { + "bbox": [ + 379, + 401, + 422, + 412 + ], + "score": 0.91, + "content": "k _ { 0 } ~ = ~ 2 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 401, + 505, + 414 + ], + "score": 1.0, + "content": ". Bottom-right: the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 411, + 435, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 298, + 425 + ], + "score": 1.0, + "content": "singular values for different values of the cutoff", + "type": "text" + }, + { + "bbox": [ + 299, + 412, + 309, + 423 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 411, + 435, + 425 + ], + "score": 1.0, + "content": ", after 150 epochs in each case.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 123, + 452, + 501, + 605 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 452, + 501, + 605 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 123, + 452, + 501, + 605 + ], + "spans": [ + { + "bbox": [ + 123, + 452, + 501, + 605 + ], + "score": 0.962, + "type": "image", + "image_path": "f1a73e5112ee1b5d1d1b809be05b2c39b82dfe510aa9cbf963a2647ac4f9eb88.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 123, + 452, + 501, + 503.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 123, + 503.0, + 501, + 554.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 123, + 554.0, + 501, + 605.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 619, + 506, + 719 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "Figure 2: Loss and inaccuracy (error rate) on test sets for two classification tasks. The models were", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "trained either using the cross-entropy (CE) or our approach (DCCI), with or without regularization", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "layers. On the MNIST dataset, we used an “all CNN” network, and for the CIFAR10 dataset we used", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 651, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 505, + 666 + ], + "score": 1.0, + "content": "a short VGG variation with 10 convolutions and 3 fully connected layers. In the regularized form,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "score": 1.0, + "content": "post-activation Batchnorm layers were placed after each convolutional layers on the VGG network.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 675, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 505, + 687 + ], + "score": 1.0, + "content": "What is shown is the mean over 10 independent runs for MNIST and 5 runs for CIFAR10. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 685, + 505, + 697 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 505, + 697 + ], + "score": 1.0, + "content": "shaded area spans the standard deviation. ADAM with default parameters was used in all cases. No", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 696, + 506, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 506, + 709 + ], + "score": 1.0, + "content": "data augmentation was used except for horizontal flips for CIFAR10 (resulting in epochs of 100,000", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 706, + 144, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 144, + 722 + ], + "score": 1.0, + "content": "images).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + } + ], + "index": 14.0 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 92, + 509, + 331 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 92, + 509, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 92, + 509, + 331 + ], + "spans": [ + { + "bbox": [ + 112, + 92, + 509, + 331 + ], + "score": 0.89, + "type": "image", + "image_path": "201f3ce00c47dbafcb9ef33dbcddfe321ec4b0f6d68f48039079f27a17681d75.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 92, + 509, + 171.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 171.66666666666669, + 509, + 251.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 251.33333333333337, + 509, + 331.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 345, + 506, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "Figure 1: Left mosaic: the left halves of the MNIST digits in a random sample from the test set are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 268, + 369 + ], + "score": 1.0, + "content": "inferred from the right half, with cutoff", + "type": "text" + }, + { + "bbox": [ + 268, + 357, + 308, + 368 + ], + "score": 0.92, + "content": "k _ { 0 } = 2 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 358, + 505, + 369 + ], + "score": 1.0, + "content": ". Three images are shown for each digit. Within", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "each triplet, the middle image represents the mean over pixel intensity of the inferred condition", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "distribution, while left and right images corresponds to a plus and minus one standard deviation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "score": 1.0, + "content": "from the mean in the direction of largest covariance (in the space of half-images). A particularly", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 378, + 414 + ], + "score": 1.0, + "content": "interesting example is highlighted. Top-right: loss per epoch for", + "type": "text" + }, + { + "bbox": [ + 379, + 401, + 422, + 412 + ], + "score": 0.91, + "content": "k _ { 0 } ~ = ~ 2 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 401, + 505, + 414 + ], + "score": 1.0, + "content": ". Bottom-right: the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 411, + 435, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 298, + 425 + ], + "score": 1.0, + "content": "singular values for different values of the cutoff", + "type": "text" + }, + { + "bbox": [ + 299, + 412, + 309, + 423 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 411, + 435, + 425 + ], + "score": 1.0, + "content": ", after 150 epochs in each case.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "image", + "bbox": [ + 123, + 452, + 501, + 605 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 452, + 501, + 605 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 123, + 452, + 501, + 605 + ], + "spans": [ + { + "bbox": [ + 123, + 452, + 501, + 605 + ], + "score": 0.962, + "type": "image", + "image_path": "f1a73e5112ee1b5d1d1b809be05b2c39b82dfe510aa9cbf963a2647ac4f9eb88.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 123, + 452, + 501, + 503.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 123, + 503.0, + 501, + 554.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 123, + 554.0, + 501, + 605.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 619, + 506, + 719 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "Figure 2: Loss and inaccuracy (error rate) on test sets for two classification tasks. The models were", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "trained either using the cross-entropy (CE) or our approach (DCCI), with or without regularization", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "layers. On the MNIST dataset, we used an “all CNN” network, and for the CIFAR10 dataset we used", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 651, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 505, + 666 + ], + "score": 1.0, + "content": "a short VGG variation with 10 convolutions and 3 fully connected layers. In the regularized form,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "score": 1.0, + "content": "post-activation Batchnorm layers were placed after each convolutional layers on the VGG network.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 675, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 505, + 687 + ], + "score": 1.0, + "content": "What is shown is the mean over 10 independent runs for MNIST and 5 runs for CIFAR10. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 685, + 505, + 697 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 505, + 697 + ], + "score": 1.0, + "content": "shaded area spans the standard deviation. ADAM with default parameters was used in all cases. No", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 696, + 506, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 506, + 709 + ], + "score": 1.0, + "content": "data augmentation was used except for horizontal flips for CIFAR10 (resulting in epochs of 100,000", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 706, + 144, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 144, + 722 + ], + "score": 1.0, + "content": "images).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + } + ], + "index": 14.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 80, + 506, + 234 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 80, + 506, + 234 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 80, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 113, + 80, + 506, + 234 + ], + "score": 0.976, + "type": "image", + "image_path": "03b9882c59b4695eeffc013e7377800c0cf409c81087a146cf7fe4fc6b062c90.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 80, + 506, + 131.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 131.33333333333334, + 506, + 182.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 182.66666666666669, + 506, + 234.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 256, + 505, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 257, + 504, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 307, + 268 + ], + "score": 1.0, + "content": "Figure 3: Top-left: First 20 relevant variables (on", + "type": "text" + }, + { + "bbox": [ + 308, + 257, + 318, + 267 + ], + "score": 0.67, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 257, + 493, + 268 + ], + "score": 1.0, + "content": ") determined by DCCA for a system where", + "type": "text" + }, + { + "bbox": [ + 493, + 257, + 504, + 267 + ], + "score": 0.74, + "content": "X", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 460, + 280 + ], + "score": 1.0, + "content": "consists of two coordinates uniformly sampled over a circle and a surrounding ring, and", + "type": "text" + }, + { + "bbox": [ + 460, + 268, + 470, + 278 + ], + "score": 0.66, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "consists", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "of the same points but shifted by a small normally distributed vector. The variables are arranged", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "from left-to-right and top-to-bottom in order of decreasing relevance. Top-right: the same variables", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 215, + 313 + ], + "score": 1.0, + "content": "multiplied by the marginal", + "type": "text" + }, + { + "bbox": [ + 215, + 302, + 229, + 312 + ], + "score": 0.82, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 301, + 505, + 313 + ], + "score": 1.0, + "content": ". Bottom row: introducing a gap in the ring allows for a monotonous", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 104, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "function of the angle to serve as second most relevant variable (instead of the sine/cosine couple).", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "Hence the angle is automatically “disentangled” from the other variables. (Mid-gray represents the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 332, + 144, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 144, + 346 + ], + "score": 1.0, + "content": "value 0).", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "On MNIST, DCCI alone also outperformed cross-entropy with dropout. (Dropout did not yield", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "any improvement in conjunction with DCCI). However, adding batch-normalization layers on the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 386, + 402, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 402, + 401 + ], + "score": 1.0, + "content": "CIFAR example, erased any distinction between DCCI and cross-entropy.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 107, + 412, + 315, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 316, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 316, + 425 + ], + "score": 1.0, + "content": "5.3 STRUCTURE OF THE RELEVANT VARIABLES", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 432, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 252, + 445 + ], + "score": 1.0, + "content": "We mentioned in Section 3 that if", + "type": "text" + }, + { + "bbox": [ + 252, + 433, + 281, + 444 + ], + "score": 0.92, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 432, + 506, + 445 + ], + "score": 1.0, + "content": "is a two-dimensional Gaussian distribution with zero", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 444, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 169, + 455 + ], + "score": 1.0, + "content": "mean, then the", + "type": "text" + }, + { + "bbox": [ + 169, + 446, + 176, + 453 + ], + "score": 0.66, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 444, + 285, + 455 + ], + "score": 1.0, + "content": "most relevant variables of", + "type": "text" + }, + { + "bbox": [ + 285, + 444, + 295, + 453 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 444, + 345, + 455 + ], + "score": 1.0, + "content": "are the first", + "type": "text" + }, + { + "bbox": [ + 345, + 446, + 353, + 453 + ], + "score": 0.71, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 444, + 397, + 455 + ], + "score": 1.0, + "content": "powers of", + "type": "text" + }, + { + "bbox": [ + 397, + 444, + 408, + 453 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 444, + 506, + 455 + ], + "score": 1.0, + "content": "itself, independently of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 454, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 493, + 466 + ], + "score": 1.0, + "content": "the covariance matrix. This implies that the canonical variables are the Hermite polynomials in", + "type": "text" + }, + { + "bbox": [ + 494, + 455, + 504, + 464 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 465, + 456, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 390, + 478 + ], + "score": 1.0, + "content": "(which results from applying the Gram-Schmidt procedure to the basis", + "type": "text" + }, + { + "bbox": [ + 391, + 465, + 451, + 478 + ], + "score": 0.92, + "content": "\\{ 1 , x , x ^ { 2 } , \\bar { . . . } \\bar \\} \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 465, + 456, + 478 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "A similar property holds for multivariate Gaussians, namely, the less relevant singular values are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 494, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 505 + ], + "score": 1.0, + "content": "polynomials in the more relevant ones. If this is true more generally, it should be possible to further", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "compress and organize the latent space extracted with DCCA by finding a minimal set of generators,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 515, + 366, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 366, + 527 + ], + "score": 1.0, + "content": "which ought to also be in the span of the most relevant variables.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "score": 1.0, + "content": "We applied DCCA to a synthetic dataset to explore this idea, shown in Fig. 3. In this case, we", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "actually performed a final SVD to obtain the unique uncorrelated canonical variables, and ordered", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 555, + 361, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 361, + 567 + ], + "score": 1.0, + "content": "them by decreasing relevance (their respective singular values).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 571, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 131, + 583 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 131, + 571, + 142, + 581 + ], + "score": 0.77, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "consists of two real numbers, distributed uniformly within a ring and a disk. The variable", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 116, + 592 + ], + "score": 0.77, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 581, + 318, + 595 + ], + "score": 1.0, + "content": "is obtained by adding a random Gaussian shift to", + "type": "text" + }, + { + "bbox": [ + 319, + 583, + 329, + 592 + ], + "score": 0.79, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "with a small standard deviation. The more", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "relevant variables ought to be those which are more robust to such small random displacement. This", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 604, + 477, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 477, + 616 + ], + "score": 1.0, + "content": "formalizes the idea that we are interested in extracting “large-scale” variables Beny (2018b). ´", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 504, + 665 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "score": 1.0, + "content": "We would expect the relevant independent variables to be: the binary variable indicating whether", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "the point is in the disk or the ring and the angle around the ring, followed by the radial component", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "in the ring, and finally the Cartesian coordinates inside the disk. This is precisely what we see in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 653, + 137, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 653, + 137, + 667 + ], + "score": 1.0, + "content": "Fig. 3.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 671, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "Indeed—if we put aside for now the fact that the angle itself is not directly represented—besides the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "score": 1.0, + "content": "constant function, the two most relevant variables are the sine and cosine of the angle, followed by", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 691, + 318, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 318, + 707 + ], + "score": 1.0, + "content": "the binary variable separating the disk from the ring.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "But these variables ought to span the space of probabilities over the relevant variables, not just the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "variables themselves. Hence the next six variables are sines and cosines of smaller wavelength,", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 80, + 506, + 234 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 80, + 506, + 234 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 80, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 113, + 80, + 506, + 234 + ], + "score": 0.976, + "type": "image", + "image_path": "03b9882c59b4695eeffc013e7377800c0cf409c81087a146cf7fe4fc6b062c90.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 113, + 80, + 506, + 131.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 113, + 131.33333333333334, + 506, + 182.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 113, + 182.66666666666669, + 506, + 234.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 256, + 505, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 257, + 504, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 307, + 268 + ], + "score": 1.0, + "content": "Figure 3: Top-left: First 20 relevant variables (on", + "type": "text" + }, + { + "bbox": [ + 308, + 257, + 318, + 267 + ], + "score": 0.67, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 257, + 493, + 268 + ], + "score": 1.0, + "content": ") determined by DCCA for a system where", + "type": "text" + }, + { + "bbox": [ + 493, + 257, + 504, + 267 + ], + "score": 0.74, + "content": "X", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 460, + 280 + ], + "score": 1.0, + "content": "consists of two coordinates uniformly sampled over a circle and a surrounding ring, and", + "type": "text" + }, + { + "bbox": [ + 460, + 268, + 470, + 278 + ], + "score": 0.66, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "consists", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "of the same points but shifted by a small normally distributed vector. The variables are arranged", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 302 + ], + "score": 1.0, + "content": "from left-to-right and top-to-bottom in order of decreasing relevance. Top-right: the same variables", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 215, + 313 + ], + "score": 1.0, + "content": "multiplied by the marginal", + "type": "text" + }, + { + "bbox": [ + 215, + 302, + 229, + 312 + ], + "score": 0.82, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 301, + 505, + 313 + ], + "score": 1.0, + "content": ". Bottom row: introducing a gap in the ring allows for a monotonous", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 104, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "function of the angle to serve as second most relevant variable (instead of the sine/cosine couple).", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "Hence the angle is automatically “disentangled” from the other variables. (Mid-gray represents the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 332, + 144, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 144, + 346 + ], + "score": 1.0, + "content": "value 0).", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + } + ], + "index": 3.75 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "On MNIST, DCCI alone also outperformed cross-entropy with dropout. (Dropout did not yield", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "any improvement in conjunction with DCCI). However, adding batch-normalization layers on the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 386, + 402, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 402, + 401 + ], + "score": 1.0, + "content": "CIFAR example, erased any distinction between DCCI and cross-entropy.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 364, + 505, + 401 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 412, + 315, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 316, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 316, + 425 + ], + "score": 1.0, + "content": "5.3 STRUCTURE OF THE RELEVANT VARIABLES", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 432, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 252, + 445 + ], + "score": 1.0, + "content": "We mentioned in Section 3 that if", + "type": "text" + }, + { + "bbox": [ + 252, + 433, + 281, + 444 + ], + "score": 0.92, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 432, + 506, + 445 + ], + "score": 1.0, + "content": "is a two-dimensional Gaussian distribution with zero", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 444, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 169, + 455 + ], + "score": 1.0, + "content": "mean, then the", + "type": "text" + }, + { + "bbox": [ + 169, + 446, + 176, + 453 + ], + "score": 0.66, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 444, + 285, + 455 + ], + "score": 1.0, + "content": "most relevant variables of", + "type": "text" + }, + { + "bbox": [ + 285, + 444, + 295, + 453 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 444, + 345, + 455 + ], + "score": 1.0, + "content": "are the first", + "type": "text" + }, + { + "bbox": [ + 345, + 446, + 353, + 453 + ], + "score": 0.71, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 444, + 397, + 455 + ], + "score": 1.0, + "content": "powers of", + "type": "text" + }, + { + "bbox": [ + 397, + 444, + 408, + 453 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 444, + 506, + 455 + ], + "score": 1.0, + "content": "itself, independently of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 454, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 493, + 466 + ], + "score": 1.0, + "content": "the covariance matrix. This implies that the canonical variables are the Hermite polynomials in", + "type": "text" + }, + { + "bbox": [ + 494, + 455, + 504, + 464 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 465, + 456, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 390, + 478 + ], + "score": 1.0, + "content": "(which results from applying the Gram-Schmidt procedure to the basis", + "type": "text" + }, + { + "bbox": [ + 391, + 465, + 451, + 478 + ], + "score": 0.92, + "content": "\\{ 1 , x , x ^ { 2 } , \\bar { . . . } \\bar \\} \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 465, + 456, + 478 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 432, + 506, + 478 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "A similar property holds for multivariate Gaussians, namely, the less relevant singular values are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 494, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 506, + 505 + ], + "score": 1.0, + "content": "polynomials in the more relevant ones. If this is true more generally, it should be possible to further", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "compress and organize the latent space extracted with DCCA by finding a minimal set of generators,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 515, + 366, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 366, + 527 + ], + "score": 1.0, + "content": "which ought to also be in the span of the most relevant variables.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 482, + 506, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "score": 1.0, + "content": "We applied DCCA to a synthetic dataset to explore this idea, shown in Fig. 3. In this case, we", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 555 + ], + "score": 1.0, + "content": "actually performed a final SVD to obtain the unique uncorrelated canonical variables, and ordered", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 555, + 361, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 361, + 567 + ], + "score": 1.0, + "content": "them by decreasing relevance (their respective singular values).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 531, + 505, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 571, + 505, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 131, + 583 + ], + "score": 1.0, + "content": "Here,", + "type": "text" + }, + { + "bbox": [ + 131, + 571, + 142, + 581 + ], + "score": 0.77, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "consists of two real numbers, distributed uniformly within a ring and a disk. The variable", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 116, + 592 + ], + "score": 0.77, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 581, + 318, + 595 + ], + "score": 1.0, + "content": "is obtained by adding a random Gaussian shift to", + "type": "text" + }, + { + "bbox": [ + 319, + 583, + 329, + 592 + ], + "score": 0.79, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "with a small standard deviation. The more", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "relevant variables ought to be those which are more robust to such small random displacement. This", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 604, + 477, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 477, + 616 + ], + "score": 1.0, + "content": "formalizes the idea that we are interested in extracting “large-scale” variables Beny (2018b). ´", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 571, + 506, + 616 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 504, + 665 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 633 + ], + "score": 1.0, + "content": "We would expect the relevant independent variables to be: the binary variable indicating whether", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "the point is in the disk or the ring and the angle around the ring, followed by the radial component", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "in the ring, and finally the Cartesian coordinates inside the disk. This is precisely what we see in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 653, + 137, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 653, + 137, + 667 + ], + "score": 1.0, + "content": "Fig. 3.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 104, + 621, + 506, + 667 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 671, + 504, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "Indeed—if we put aside for now the fact that the angle itself is not directly represented—besides the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "score": 1.0, + "content": "constant function, the two most relevant variables are the sine and cosine of the angle, followed by", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 691, + 318, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 318, + 707 + ], + "score": 1.0, + "content": "the binary variable separating the disk from the ring.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 670, + 505, + 707 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "But these variables ought to span the space of probabilities over the relevant variables, not just the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "variables themselves. Hence the next six variables are sines and cosines of smaller wavelength,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "which can encode probability distributions which are increasingly more precisely localized, down to", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "a precision (wavelength) comparable with the diameter of the inner disk. Accordingly, the next two", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "most relevant variables are the Cartesian coordinates inside the disk. This is followed by additional", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "moments of the angle, down to a wavelength equal to the ring’s thickness, at which point we see the", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 364, + 208, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 208, + 377 + ], + "score": 1.0, + "content": "radius in the ring appear.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 78, + 496, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 78, + 496, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 78, + 496, + 225 + ], + "spans": [ + { + "bbox": [ + 116, + 78, + 496, + 225 + ], + "score": 0.964, + "type": "image", + "image_path": "139ed066324617d38572ae37b20db57c2963cbc5f2e6fcfe36fcf651198597fd.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 78, + 496, + 127.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 127.0, + 496, + 176.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 176.0, + 496, + 225.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 239, + 505, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 401, + 252 + ], + "score": 1.0, + "content": "Figure 4: Left: Best mean squared error for images reconstructed from the", + "type": "text" + }, + { + "bbox": [ + 401, + 240, + 408, + 249 + ], + "score": 0.75, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "most relevant variables,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 249, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 171, + 263 + ], + "score": 1.0, + "content": "as a function of", + "type": "text" + }, + { + "bbox": [ + 171, + 251, + 178, + 260 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 249, + 505, + 263 + ], + "score": 1.0, + "content": "(the latent dimension). This logarithmic plot shows that improvements stop once", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "score": 1.0, + "content": "the dimension reaches 19 (where the two lines cross). Right: images produced by the generator from", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "latent variables sampled according to the best Gaussian fit in latent space, for feature subspaces of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 282, + 204, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 204, + 296 + ], + "score": 1.0, + "content": "dimensions 2, 8 and 19.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 505, + 376 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "which can encode probability distributions which are increasingly more precisely localized, down to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "a precision (wavelength) comparable with the diameter of the inner disk. Accordingly, the next two", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "most relevant variables are the Cartesian coordinates inside the disk. This is followed by additional", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "moments of the angle, down to a wavelength equal to the ring’s thickness, at which point we see the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 364, + 208, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 208, + 377 + ], + "score": 1.0, + "content": "radius in the ring appear.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 381, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 504, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 504, + 393 + ], + "score": 1.0, + "content": "As mentioned, we see that the angle itself is not represented, likely because it is discontinuous.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 392, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 404 + ], + "score": 1.0, + "content": "However, as shown also in Fig. 3, creating a gap in the ring allows for the angle to emerge as", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "most relevant variable. This suggests that this approach may be able to automatically learn intrinsic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 415, + 279, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 279, + 425 + ], + "score": 1.0, + "content": "coordinates of the latent variable manifold.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 106, + 446, + 355, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 356, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 356, + 457 + ], + "score": 1.0, + "content": "5.4 INDEPENDENT VARIABLES AND GENERATIVE MODEL", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "If we postulate that the independent (or disentangled) relevant latent variables can be found in the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "linear span of the relevant variables, we can attempt to extract them by optimizing a neural network", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 499, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 511 + ], + "score": 1.0, + "content": "composed of two parts. Firstly, a linear layer maps the relevant variables to a small number of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "outputs (equal to the latent dimension). The purpose of this linear layer is to find the independent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "variables. These latent variables are then processed by an arbitrarily complex generative network to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 263, + 545 + ], + "score": 1.0, + "content": "produce a possible value of the variable", + "type": "text" + }, + { + "bbox": [ + 263, + 533, + 273, + 542 + ], + "score": 0.74, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 532, + 506, + 545 + ], + "score": 1.0, + "content": ". As objective function, we may us an appropriate measure", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 544, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 483, + 555 + ], + "score": 1.0, + "content": "of similarity between the output and the data element from which the variables were obtained.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 108, + 560, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 276, + 573 + ], + "score": 1.0, + "content": "We tested this idea as follows. We took", + "type": "text" + }, + { + "bbox": [ + 276, + 561, + 287, + 570 + ], + "score": 0.8, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 559, + 480, + 573 + ], + "score": 1.0, + "content": "to consist of the MNIST digits, and produced", + "type": "text" + }, + { + "bbox": [ + 481, + 560, + 491, + 570 + ], + "score": 0.56, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "score": 1.0, + "content": "randomly permuting neighboring pixels in the image, until the mean displacement per pixel is of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "order 1. In addition, we added independent Gaussian noise to the pixel values. (Hence the noise", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 593, + 455, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 126, + 606 + ], + "score": 1.0, + "content": "map", + "type": "text" + }, + { + "bbox": [ + 126, + 593, + 137, + 603 + ], + "score": 0.74, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 593, + 455, + 606 + ], + "score": 1.0, + "content": "simulates the coarse-graining channel introduced in Beny & Osborne (2015a)). ´", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 108, + 610, + 503, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "As in the previous experiment, we do so to implement our intuition that the more relevant variables", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 621, + 443, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 443, + 634 + ], + "score": 1.0, + "content": "ought to be the ones which are of larger scale, or more robust to local perturbations.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 108, + 638, + 503, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "The relevant variables of the clean images were produced by the same convolutional neural network", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "as in Section 5.2, while the variables of the coarse-grained images were extracted by a network of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 660, + 377, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 377, + 672 + ], + "score": 1.0, + "content": "the same geometry, but with half the number of filters and neurons.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 677, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 107, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 107, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "We extracted the 1000 most relevant out of 1200 learned variables in this way. (The least relevant", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 107, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "variables in this system happen to be highly dependent on the total number of variables and hence", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "cannot be trusted to be correct). As a second step, we trained a linear layer coupled to a network", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "composed of 5 fully-connected layers of 800 hidden neurons each. We refer to the number of output", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 721, + 327, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 327, + 733 + ], + "score": 1.0, + "content": "neurons in the first linear layer as the latent dimension.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 78, + 496, + 225 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 78, + 496, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 78, + 496, + 225 + ], + "spans": [ + { + "bbox": [ + 116, + 78, + 496, + 225 + ], + "score": 0.964, + "type": "image", + "image_path": "139ed066324617d38572ae37b20db57c2963cbc5f2e6fcfe36fcf651198597fd.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 78, + 496, + 127.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 127.0, + 496, + 176.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 176.0, + 496, + 225.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 239, + 505, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 401, + 252 + ], + "score": 1.0, + "content": "Figure 4: Left: Best mean squared error for images reconstructed from the", + "type": "text" + }, + { + "bbox": [ + 401, + 240, + 408, + 249 + ], + "score": 0.75, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "most relevant variables,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 249, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 171, + 263 + ], + "score": 1.0, + "content": "as a function of", + "type": "text" + }, + { + "bbox": [ + 171, + 251, + 178, + 260 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 249, + 505, + 263 + ], + "score": 1.0, + "content": "(the latent dimension). This logarithmic plot shows that improvements stop once", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "score": 1.0, + "content": "the dimension reaches 19 (where the two lines cross). Right: images produced by the generator from", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 285 + ], + "score": 1.0, + "content": "latent variables sampled according to the best Gaussian fit in latent space, for feature subspaces of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 282, + 204, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 204, + 296 + ], + "score": 1.0, + "content": "dimensions 2, 8 and 19.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 505, + 376 + ], + "lines": [], + "index": 10, + "bbox_fs": [ + 105, + 320, + 505, + 377 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 381, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 504, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 504, + 393 + ], + "score": 1.0, + "content": "As mentioned, we see that the angle itself is not represented, likely because it is discontinuous.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 392, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 404 + ], + "score": 1.0, + "content": "However, as shown also in Fig. 3, creating a gap in the ring allows for the angle to emerge as", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "most relevant variable. This suggests that this approach may be able to automatically learn intrinsic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 415, + 279, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 279, + 425 + ], + "score": 1.0, + "content": "coordinates of the latent variable manifold.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 381, + 505, + 425 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 446, + 355, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 356, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 356, + 457 + ], + "score": 1.0, + "content": "5.4 INDEPENDENT VARIABLES AND GENERATIVE MODEL", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "If we postulate that the independent (or disentangled) relevant latent variables can be found in the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "linear span of the relevant variables, we can attempt to extract them by optimizing a neural network", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 499, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 511 + ], + "score": 1.0, + "content": "composed of two parts. Firstly, a linear layer maps the relevant variables to a small number of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "outputs (equal to the latent dimension). The purpose of this linear layer is to find the independent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 534 + ], + "score": 1.0, + "content": "variables. These latent variables are then processed by an arbitrarily complex generative network to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 263, + 545 + ], + "score": 1.0, + "content": "produce a possible value of the variable", + "type": "text" + }, + { + "bbox": [ + 263, + 533, + 273, + 542 + ], + "score": 0.74, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 532, + 506, + 545 + ], + "score": 1.0, + "content": ". As objective function, we may us an appropriate measure", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 544, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 483, + 555 + ], + "score": 1.0, + "content": "of similarity between the output and the data element from which the variables were obtained.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 476, + 506, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 560, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 276, + 573 + ], + "score": 1.0, + "content": "We tested this idea as follows. We took", + "type": "text" + }, + { + "bbox": [ + 276, + 561, + 287, + 570 + ], + "score": 0.8, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 559, + 480, + 573 + ], + "score": 1.0, + "content": "to consist of the MNIST digits, and produced", + "type": "text" + }, + { + "bbox": [ + 481, + 560, + 491, + 570 + ], + "score": 0.56, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "score": 1.0, + "content": "randomly permuting neighboring pixels in the image, until the mean displacement per pixel is of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "order 1. In addition, we added independent Gaussian noise to the pixel values. (Hence the noise", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 593, + 455, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 126, + 606 + ], + "score": 1.0, + "content": "map", + "type": "text" + }, + { + "bbox": [ + 126, + 593, + 137, + 603 + ], + "score": 0.74, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 593, + 455, + 606 + ], + "score": 1.0, + "content": "simulates the coarse-graining channel introduced in Beny & Osborne (2015a)). ´", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 559, + 506, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 610, + 503, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 622 + ], + "score": 1.0, + "content": "As in the previous experiment, we do so to implement our intuition that the more relevant variables", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 621, + 443, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 443, + 634 + ], + "score": 1.0, + "content": "ought to be the ones which are of larger scale, or more robust to local perturbations.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 610, + 505, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 638, + 503, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "The relevant variables of the clean images were produced by the same convolutional neural network", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "as in Section 5.2, while the variables of the coarse-grained images were extracted by a network of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 660, + 377, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 377, + 672 + ], + "score": 1.0, + "content": "the same geometry, but with half the number of filters and neurons.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 637, + 505, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 677, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 107, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 107, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "We extracted the 1000 most relevant out of 1200 learned variables in this way. (The least relevant", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 107, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "variables in this system happen to be highly dependent on the total number of variables and hence", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "cannot be trusted to be correct). As a second step, we trained a linear layer coupled to a network", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "composed of 5 fully-connected layers of 800 hidden neurons each. We refer to the number of output", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 721, + 327, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 327, + 733 + ], + "score": 1.0, + "content": "neurons in the first linear layer as the latent dimension.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 677, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "As input, this network received the variables extracted from MNIST images using the above convo-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "lutional neural net (after it was fully trained using DCCA), and was trained to minimize the mean", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 353, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 353, + 117 + ], + "score": 1.0, + "content": "square error between its output and the original MNIST digit.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 107, + 121, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 107, + 121, + 506, + 133 + ], + "score": 1.0, + "content": "The resulting best mean square errors are shown in Fig. 4, as function of the latent dimension. Here", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "we see a distinct change of polynomial scaling law at dimension 19. Increasing the dimension further", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "provides no improvement. This behaviour is compatible with our hypothesis that the extra variables", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "score": 1.0, + "content": "are just functions of those first twenty variables (functions which are effectively re-implemented by", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 166, + 205, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 205, + 178 + ], + "score": 1.0, + "content": "the generative network).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 182, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Images generated by sampling from a Gaussian approximation of the latent distribution for different", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "latent dimensions are shown in Fig. 4. Below dimension 20, most generated image can be recognized", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 180, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 180, + 217 + ], + "score": 1.0, + "content": "as a specific digit.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 107, + 232, + 178, + 246 + ], + "lines": [ + { + "bbox": [ + 104, + 231, + 181, + 249 + ], + "spans": [ + { + "bbox": [ + 104, + 231, + 181, + 249 + ], + "score": 1.0, + "content": "6 OUTLOOK", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "We studied the classical (non-quantum) form of the theory introduced in Beny & Osborne (2013), ´", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "and found that the relevant observables of that theory are just the most correlated canonical variables", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "in the sense of DCCA Andrew et al. (2013), and can be learned effectively using standard machine", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 292, + 180, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 180, + 304 + ], + "score": 1.0, + "content": "learning methods.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 504, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "This point of views on DCCA provided us with several new insights. The first is that the learned", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "relevant variables provide a useful representation of a joint probability distribution. We showed that", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "performing inference using this representation can outperform crossentropy in predicting classes.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 341, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 354 + ], + "score": 1.0, + "content": "Our experiments on halves of MNIST also show that the conditional distribution we obtain can", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 424, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 424, + 366 + ], + "score": 1.0, + "content": "effectively represent the uncertainty in the prediction of high-dimensional data.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "A second insight relates to the interpretation of the canonical variables as spanning directions in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "the space of probability distributions. As suggested by the gaussian solutions and our experiment", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "on synthetic data, we postulate that the canonical variables are functions of a small number of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "independent generators contained in their span. This hypothesis is supported by our experiment on", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 412, + 443, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 443, + 426 + ], + "score": 1.0, + "content": "MNIST, but further work is required to find a way to cleanly extract these variables.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "We have yet to explore the potential applications of one of the salient aspect of this approach to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "inference, the fact that the canonical variables learned using DCCA are also those which can be most", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "reliably predicted, irrespective of the value of the cutoff. To see why this is potentially significant,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "we observe that a central feature of scientific exploration is that we are not so concerned with making", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "predictions about some given variables, as much as we are with discovering variables which can be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 485, + 148, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 148, + 496 + ], + "score": 1.0, + "content": "predicted.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 108, + 502, + 504, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "Another important feature of this approach is the fact that the resulting model allows for the direct", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "evaluation of the expectation values in the posterior distribution without sampling. In particular", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "this allows for the evaluation of credible intervals. Hence it should be especially suited to scientific", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 535, + 369, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 369, + 547 + ], + "score": 1.0, + "content": "applications where the ability to quantify uncertainty is essential.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 108, + 551, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "score": 1.0, + "content": "Finally, the relationship that we established with theory of quantum origin points towards a potential", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "score": 1.0, + "content": "quantum generalization of DCCA that would apply to quantum data, or classical measurements of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 575, + 181, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 181, + 586 + ], + "score": 1.0, + "content": "quantum systems.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 108, + 603, + 218, + 615 + ], + "lines": [ + { + "bbox": [ + 107, + 603, + 219, + 617 + ], + "spans": [ + { + "bbox": [ + 107, + 603, + 219, + 617 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 628, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "We would like to thank Joel B ¨ eny and Raban Iten for helpful suggestions. We are also in- ´", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "debted to an anonymous ICLR2020 referee for pointing out the connection between our approach", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "and DCCA. This work was supported by the National Research Foundation of Korea (NRF-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 660, + 218, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 218, + 673 + ], + "score": 1.0, + "content": "2018R1D1A1A02048436).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "title", + "bbox": [ + 108, + 690, + 175, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 176, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 176, + 703 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analysis.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 719, + 406, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 406, + 733 + ], + "score": 1.0, + "content": "In International conference on machine learning, pp. 1247–1255, 2013.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "As input, this network received the variables extracted from MNIST images using the above convo-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "lutional neural net (after it was fully trained using DCCA), and was trained to minimize the mean", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 353, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 353, + 117 + ], + "score": 1.0, + "content": "square error between its output and the original MNIST digit.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 107, + 121, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 107, + 121, + 506, + 133 + ], + "score": 1.0, + "content": "The resulting best mean square errors are shown in Fig. 4, as function of the latent dimension. Here", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "we see a distinct change of polynomial scaling law at dimension 19. Increasing the dimension further", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "provides no improvement. This behaviour is compatible with our hypothesis that the extra variables", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 168 + ], + "score": 1.0, + "content": "are just functions of those first twenty variables (functions which are effectively re-implemented by", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 166, + 205, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 205, + 178 + ], + "score": 1.0, + "content": "the generative network).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 121, + 506, + 178 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 182, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Images generated by sampling from a Gaussian approximation of the latent distribution for different", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "latent dimensions are shown in Fig. 4. Below dimension 20, most generated image can be recognized", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 180, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 180, + 217 + ], + "score": 1.0, + "content": "as a specific digit.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 182, + 505, + 217 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 232, + 178, + 246 + ], + "lines": [ + { + "bbox": [ + 104, + 231, + 181, + 249 + ], + "spans": [ + { + "bbox": [ + 104, + 231, + 181, + 249 + ], + "score": 1.0, + "content": "6 OUTLOOK", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "We studied the classical (non-quantum) form of the theory introduced in Beny & Osborne (2013), ´", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "and found that the relevant observables of that theory are just the most correlated canonical variables", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "in the sense of DCCA Andrew et al. (2013), and can be learned effectively using standard machine", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 292, + 180, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 180, + 304 + ], + "score": 1.0, + "content": "learning methods.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 258, + 506, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 504, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "This point of views on DCCA provided us with several new insights. The first is that the learned", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "relevant variables provide a useful representation of a joint probability distribution. We showed that", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "performing inference using this representation can outperform crossentropy in predicting classes.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 341, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 354 + ], + "score": 1.0, + "content": "Our experiments on halves of MNIST also show that the conditional distribution we obtain can", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 424, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 424, + 366 + ], + "score": 1.0, + "content": "effectively represent the uncertainty in the prediction of high-dimensional data.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 309, + 506, + 366 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 381 + ], + "score": 1.0, + "content": "A second insight relates to the interpretation of the canonical variables as spanning directions in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "the space of probability distributions. As suggested by the gaussian solutions and our experiment", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "on synthetic data, we postulate that the canonical variables are functions of a small number of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "independent generators contained in their span. This hypothesis is supported by our experiment on", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 412, + 443, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 443, + 426 + ], + "score": 1.0, + "content": "MNIST, but further work is required to find a way to cleanly extract these variables.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 370, + 505, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 442 + ], + "score": 1.0, + "content": "We have yet to explore the potential applications of one of the salient aspect of this approach to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "inference, the fact that the canonical variables learned using DCCA are also those which can be most", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "reliably predicted, irrespective of the value of the cutoff. To see why this is potentially significant,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "we observe that a central feature of scientific exploration is that we are not so concerned with making", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "predictions about some given variables, as much as we are with discovering variables which can be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 485, + 148, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 148, + 496 + ], + "score": 1.0, + "content": "predicted.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 429, + 505, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 502, + 504, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "Another important feature of this approach is the fact that the resulting model allows for the direct", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "evaluation of the expectation values in the posterior distribution without sampling. In particular", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "this allows for the evaluation of credible intervals. Hence it should be especially suited to scientific", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 535, + 369, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 369, + 547 + ], + "score": 1.0, + "content": "applications where the ability to quantify uncertainty is essential.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 502, + 506, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 551, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 564 + ], + "score": 1.0, + "content": "Finally, the relationship that we established with theory of quantum origin points towards a potential", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "score": 1.0, + "content": "quantum generalization of DCCA that would apply to quantum data, or classical measurements of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 575, + 181, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 181, + 586 + ], + "score": 1.0, + "content": "quantum systems.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 551, + 506, + 586 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 603, + 218, + 615 + ], + "lines": [ + { + "bbox": [ + 107, + 603, + 219, + 617 + ], + "spans": [ + { + "bbox": [ + 107, + 603, + 219, + 617 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 628, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "We would like to thank Joel B ¨ eny and Raban Iten for helpful suggestions. We are also in- ´", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "debted to an anonymous ICLR2020 referee for pointing out the connection between our approach", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "and DCCA. This work was supported by the National Research Foundation of Korea (NRF-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 660, + 218, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 218, + 673 + ], + "score": 1.0, + "content": "2018R1D1A1A02048436).", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 628, + 506, + 673 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 690, + 175, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 176, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 176, + 703 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. Deep canonical correlation analysis.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 719, + 406, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 406, + 733 + ], + "score": 1.0, + "content": "In International conference on machine learning, pp. 1247–1255, 2013.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 57, + 506, + 734 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 505, + 97 + ], + "score": 1.0, + "content": "C. Beny and T. J. Osborne. Information geometric approach to the renormalisation group. ´ Phys.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 450, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 450, + 107 + ], + "score": 1.0, + "content": "Rev. A, 92:022330, 2015a. doi: 10.1103/PhysRevA.92.022330. (arXiv:1206.7004).", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 130 + ], + "score": 1.0, + "content": "Cedric B ´ eny. Coarse-grained distinguishability of field interactions. ´ Quantum, 2:67, 2018a.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 127, + 202, + 141 + ], + "spans": [ + { + "bbox": [ + 115, + 127, + 202, + 141 + ], + "score": 1.0, + "content": "(arXiv:1509.03249).", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 164 + ], + "score": 1.0, + "content": "Cedric B ´ eny. Inferring relevant features: from qft to pca. ´ International Journal of Quantum Infor-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 160, + 318, + 174 + ], + "spans": [ + { + "bbox": [ + 114, + 160, + 318, + 174 + ], + "score": 1.0, + "content": "mation, 16:1840012, 2018b. (arXiv:1802.05756).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 183, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 198 + ], + "score": 1.0, + "content": "Cedric B ´ eny and Tobias J Osborne. Renormalisation as an inference problem. ´ (arXiv:1310.3188),", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 194, + 144, + 209 + ], + "spans": [ + { + "bbox": [ + 114, + 194, + 144, + 209 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 218, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 505, + 232 + ], + "score": 1.0, + "content": "Cedric B ´ eny and Tobias J Osborne. The renormalisation group via statistical inference. ´ New J.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 227, + 457, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 227, + 457, + 241 + ], + "score": 1.0, + "content": "Phys., 17:083005, 2015b. doi: 10.1088/1367-2630/17/8/083005. (arXiv:1402.4949).", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 250, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 265 + ], + "score": 1.0, + "content": "GK Eagleson. Polynomial expansions of bivariate distributions. The Annals of Mathematical Statis-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 262, + 235, + 276 + ], + "spans": [ + { + "bbox": [ + 114, + 262, + 235, + 276 + ], + "score": 1.0, + "content": "tics, 35(3):1208–1215, 1964.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 285, + 504, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 504, + 299 + ], + "score": 1.0, + "content": "Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychome-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 297, + 224, + 310 + ], + "spans": [ + { + "bbox": [ + 114, + 297, + 224, + 310 + ], + "score": 1.0, + "content": "trika, 1(3):211–218, 1936.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 319, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 333 + ], + "score": 1.0, + "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 115, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural infor-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 340, + 318, + 356 + ], + "spans": [ + { + "bbox": [ + 113, + 340, + 318, + 356 + ], + "score": 1.0, + "content": "mation processing systems, pp. 2672–2680, 2014.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 362, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 505, + 379 + ], + "score": 1.0, + "content": "Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregres-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 375, + 334, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 334, + 389 + ], + "score": 1.0, + "content": "sive networks. arXiv preprint arXiv:1310.8499, 2013.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 397, + 485, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 485, + 411 + ], + "score": 1.0, + "content": "Harold Hotelling. Relations between two sets of variates. Biometrika, 28(3/4):321–377, 1936.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 418, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 104, + 418, + 506, + 435 + ], + "score": 1.0, + "content": "Mike Innes. Flux: Elegant machine learning with julia. Journal of Open Source Software, 2018.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 432, + 223, + 446 + ], + "spans": [ + { + "bbox": [ + 115, + 432, + 223, + 446 + ], + "score": 1.0, + "content": "doi: 10.21105/joss.00602.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "score": 1.0, + "content": "Raban Iten, Tony Metger, Henrik Wilming, L´ıdia Del Rio, and Renato Renner. Discovering physical", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 466, + 353, + 479 + ], + "spans": [ + { + "bbox": [ + 114, + 466, + 353, + 479 + ], + "score": 1.0, + "content": "concepts with neural networks. (arXiv:1807.10300), 2018.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 488, + 502, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 502, + 502 + ], + "score": 1.0, + "content": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. (arXiv:1312.6114), 2013.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "HO Lancaster. The structure of bivariate distributions. The Annals of Mathematical Statistics, 29", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 522, + 196, + 534 + ], + "spans": [ + { + "bbox": [ + 114, + 522, + 196, + 534 + ], + "score": 1.0, + "content": "(3):719–736, 1958.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 543, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 559 + ], + "score": 1.0, + "content": "Tomer Michaeli, Weiran Wang, and Karen Livescu. Nonparametric canonical correlation analysis.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 556, + 411, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 556, + 411, + 570 + ], + "score": 1.0, + "content": "In International Conference on Machine Learning, pp. 1967–1976, 2016.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 578, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 594 + ], + "score": 1.0, + "content": "Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 590, + 215, + 603 + ], + "spans": [ + { + "bbox": [ + 114, + 590, + 215, + 603 + ], + "score": 1.0, + "content": "arXiv:1411.1784, 2014.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 613, + 410, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 410, + 627 + ], + "score": 1.0, + "content": "M. Ohya and D. Petz. Quantum entropy and its use. Springer Verlag, 2004.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 635, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 649 + ], + "score": 1.0, + "content": "Amichai Painsky, Meir Feder, and Naftali Tishby. An information-theoretic framework for non-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 113, + 645, + 429, + 661 + ], + "spans": [ + { + "bbox": [ + 113, + 645, + 429, + 661 + ], + "score": 1.0, + "content": "linear canonical correlation analysis. arXiv preprint arXiv:1810.13259, 2018.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "Weiran Wang, Xinchen Yan, Honglak Lee, and Karen Livescu. Deep variational canonical correla-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 680, + 335, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 335, + 694 + ], + "score": 1.0, + "content": "tion analysis. arXiv preprint arXiv:1610.03454, 2016.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 703, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 505, + 717 + ], + "score": 1.0, + "content": "Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 714, + 279, + 728 + ], + "spans": [ + { + "bbox": [ + 115, + 714, + 279, + 728 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1609.03126, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 18.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 57, + 506, + 734 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 505, + 97 + ], + "score": 1.0, + "content": "C. Beny and T. J. Osborne. Information geometric approach to the renormalisation group. ´ Phys.", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 93, + 450, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 450, + 107 + ], + "score": 1.0, + "content": "Rev. A, 92:022330, 2015a. doi: 10.1103/PhysRevA.92.022330. (arXiv:1206.7004).", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 116, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 130 + ], + "score": 1.0, + "content": "Cedric B ´ eny. Coarse-grained distinguishability of field interactions. ´ Quantum, 2:67, 2018a.", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 127, + 202, + 141 + ], + "spans": [ + { + "bbox": [ + 115, + 127, + 202, + 141 + ], + "score": 1.0, + "content": "(arXiv:1509.03249).", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 150, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 164 + ], + "score": 1.0, + "content": "Cedric B ´ eny. Inferring relevant features: from qft to pca. ´ International Journal of Quantum Infor-", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 160, + 318, + 174 + ], + "spans": [ + { + "bbox": [ + 114, + 160, + 318, + 174 + ], + "score": 1.0, + "content": "mation, 16:1840012, 2018b. (arXiv:1802.05756).", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 183, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 198 + ], + "score": 1.0, + "content": "Cedric B ´ eny and Tobias J Osborne. Renormalisation as an inference problem. ´ (arXiv:1310.3188),", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 194, + 144, + 209 + ], + "spans": [ + { + "bbox": [ + 114, + 194, + 144, + 209 + ], + "score": 1.0, + "content": "2013.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 218, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 505, + 232 + ], + "score": 1.0, + "content": "Cedric B ´ eny and Tobias J Osborne. The renormalisation group via statistical inference. ´ New J.", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 227, + 457, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 227, + 457, + 241 + ], + "score": 1.0, + "content": "Phys., 17:083005, 2015b. doi: 10.1088/1367-2630/17/8/083005. (arXiv:1402.4949).", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 250, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 265 + ], + "score": 1.0, + "content": "GK Eagleson. Polynomial expansions of bivariate distributions. The Annals of Mathematical Statis-", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 262, + 235, + 276 + ], + "spans": [ + { + "bbox": [ + 114, + 262, + 235, + 276 + ], + "score": 1.0, + "content": "tics, 35(3):1208–1215, 1964.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 285, + 504, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 504, + 299 + ], + "score": 1.0, + "content": "Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychome-", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 297, + 224, + 310 + ], + "spans": [ + { + "bbox": [ + 114, + 297, + 224, + 310 + ], + "score": 1.0, + "content": "trika, 1(3):211–218, 1936.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 319, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 333 + ], + "score": 1.0, + "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 115, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural infor-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 340, + 318, + 356 + ], + "spans": [ + { + "bbox": [ + 113, + 340, + 318, + 356 + ], + "score": 1.0, + "content": "mation processing systems, pp. 2672–2680, 2014.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 362, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 505, + 379 + ], + "score": 1.0, + "content": "Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregres-", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 375, + 334, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 334, + 389 + ], + "score": 1.0, + "content": "sive networks. arXiv preprint arXiv:1310.8499, 2013.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 397, + 485, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 485, + 411 + ], + "score": 1.0, + "content": "Harold Hotelling. Relations between two sets of variates. Biometrika, 28(3/4):321–377, 1936.", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 418, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 104, + 418, + 506, + 435 + ], + "score": 1.0, + "content": "Mike Innes. Flux: Elegant machine learning with julia. Journal of Open Source Software, 2018.", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 432, + 223, + 446 + ], + "spans": [ + { + "bbox": [ + 115, + 432, + 223, + 446 + ], + "score": 1.0, + "content": "doi: 10.21105/joss.00602.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "score": 1.0, + "content": "Raban Iten, Tony Metger, Henrik Wilming, L´ıdia Del Rio, and Renato Renner. Discovering physical", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 466, + 353, + 479 + ], + "spans": [ + { + "bbox": [ + 114, + 466, + 353, + 479 + ], + "score": 1.0, + "content": "concepts with neural networks. (arXiv:1807.10300), 2018.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 488, + 502, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 502, + 502 + ], + "score": 1.0, + "content": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. (arXiv:1312.6114), 2013.", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "HO Lancaster. The structure of bivariate distributions. The Annals of Mathematical Statistics, 29", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 522, + 196, + 534 + ], + "spans": [ + { + "bbox": [ + 114, + 522, + 196, + 534 + ], + "score": 1.0, + "content": "(3):719–736, 1958.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 543, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 559 + ], + "score": 1.0, + "content": "Tomer Michaeli, Weiran Wang, and Karen Livescu. Nonparametric canonical correlation analysis.", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 556, + 411, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 556, + 411, + 570 + ], + "score": 1.0, + "content": "In International Conference on Machine Learning, pp. 1967–1976, 2016.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 578, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 594 + ], + "score": 1.0, + "content": "Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 590, + 215, + 603 + ], + "spans": [ + { + "bbox": [ + 114, + 590, + 215, + 603 + ], + "score": 1.0, + "content": "arXiv:1411.1784, 2014.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 613, + 410, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 410, + 627 + ], + "score": 1.0, + "content": "M. Ohya and D. Petz. Quantum entropy and its use. Springer Verlag, 2004.", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 635, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 649 + ], + "score": 1.0, + "content": "Amichai Painsky, Meir Feder, and Naftali Tishby. An information-theoretic framework for non-", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 645, + 429, + 661 + ], + "spans": [ + { + "bbox": [ + 113, + 645, + 429, + 661 + ], + "score": 1.0, + "content": "linear canonical correlation analysis. arXiv preprint arXiv:1810.13259, 2018.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "Weiran Wang, Xinchen Yan, Honglak Lee, and Karen Livescu. Deep variational canonical correla-", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 680, + 335, + 694 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 335, + 694 + ], + "score": 1.0, + "content": "tion analysis. arXiv preprint arXiv:1610.03454, 2016.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 703, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 505, + 717 + ], + "score": 1.0, + "content": "Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network.", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 714, + 279, + 728 + ], + "spans": [ + { + "bbox": [ + 115, + 714, + 279, + 728 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1609.03126, 2016.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 81, + 506, + 728 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 369, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 370, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 370, + 96 + ], + "score": 1.0, + "content": "A EXTRA INFORMATION ABOUT THE ALGORITHM", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 106, + 353, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 353, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 353, + 119 + ], + "score": 1.0, + "content": "A.1 ALTERNATIVE INTERPRETATION OF THE OBJECTIVE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 127, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 155, + 141 + ], + "score": 1.0, + "content": "If we write", + "type": "text" + }, + { + "bbox": [ + 156, + 128, + 216, + 140 + ], + "score": 0.92, + "content": "F _ { i j } : = f _ { j } ( x _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 126, + 236, + 141 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 128, + 297, + 140 + ], + "score": 0.91, + "content": "G _ { i j } : = g _ { j } ( y _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 126, + 506, + 141 + ], + "score": 1.0, + "content": "for the value of our variables on the dataset, then", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 138, + 473, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 162, + 153 + ], + "score": 0.87, + "content": "\\begin{array} { r } { K = \\frac { 1 } { N } F ^ { \\top } F } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 138, + 165, + 154 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 166, + 140, + 219, + 153 + ], + "score": 0.88, + "content": "\\begin{array} { r } { L = \\frac { 1 } { N } G ^ { \\top } G } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 138, + 237, + 154 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 140, + 290, + 153 + ], + "score": 0.91, + "content": "\\begin{array} { r } { A = \\frac { 1 } { N } G ^ { \\top } F } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 138, + 473, + 154 + ], + "score": 1.0, + "content": ". The DCCA objective can then be written as", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 158, + 370, + 174 + ], + "lines": [ + { + "bbox": [ + 240, + 158, + 370, + 174 + ], + "spans": [ + { + "bbox": [ + 240, + 158, + 370, + 174 + ], + "score": 0.91, + "content": "\\operatorname { T r } \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) = \\operatorname { T r } \\left( P Q \\right)", + "type": "interline_equation", + "image_path": "bb4208eb4b2d98539b5ca87ea715a86d6484f5c067aed48305d0f58fcc0ba1c9.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 240, + 158, + 370, + 174 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 179, + 505, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 134, + 195 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 180, + 222, + 194 + ], + "score": 0.93, + "content": "P = F ( F ^ { \\top } F ) ^ { - 1 } F ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 179, + 242, + 195 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 243, + 180, + 331, + 194 + ], + "score": 0.93, + "content": "Q = G ( G ^ { \\top } G ) ^ { - 1 } G ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 179, + 477, + 195 + ], + "score": 1.0, + "content": "are the projectors on the ranges of", + "type": "text" + }, + { + "bbox": [ + 477, + 182, + 486, + 191 + ], + "score": 0.84, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 179, + 506, + 195 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 107, + 193, + 116, + 203 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "respectively. Hence, we are maximizing the overlap between those ranges (which represent pos-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "sible linear combinations of datapoints, respectively determined from variables of one or the other", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 213, + 181, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 181, + 227 + ], + "score": 1.0, + "content": "correlated views.)", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 108, + 240, + 182, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 183, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 183, + 254 + ], + "score": 1.0, + "content": "A.2 HEURISTIC", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "Batch size—In our experiments, we observed that the batch size during training needs to be an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "order of magnitude larger than the number of variables (rank cutoff). When the batch size was", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "too small, learning seemed to converge normally in terms of training and test loss, but resulted", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "in variables which yield dramatically different losses when evaluated on larger batches, and yield", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 314, + 192, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 192, + 327 + ], + "score": 1.0, + "content": "spurious predictions.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 270, + 352 + ], + "score": 1.0, + "content": "Constant variables—The loss function", + "type": "text" + }, + { + "bbox": [ + 270, + 340, + 279, + 349 + ], + "score": 0.8, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 340, + 389, + 352 + ], + "score": 1.0, + "content": "takes value between 0 and", + "type": "text" + }, + { + "bbox": [ + 389, + 340, + 418, + 351 + ], + "score": 0.9, + "content": "k _ { 0 } - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "because the constant", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 351, + 504, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 504, + 362 + ], + "score": 1.0, + "content": "variable always has relevance 1. The constant variable could be enforced a priori rather than learned,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 361, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 375 + ], + "score": 1.0, + "content": "which, due to the objective, automatically forces the learned variables to have zero expectation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 371, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 386 + ], + "score": 1.0, + "content": "values (be orthogonal to the constant variable). This might have advantages in certain circumstances,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 383, + 406, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 406, + 397 + ], + "score": 1.0, + "content": "but in our experiments we found that this sometime hindered convergence.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 295, + 422 + ], + "score": 1.0, + "content": "Invertibility issues—The covariance matrices", + "type": "text" + }, + { + "bbox": [ + 295, + 410, + 306, + 419 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 407, + 324, + 422 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 324, + 410, + 332, + 419 + ], + "score": 0.8, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 407, + 506, + 422 + ], + "score": 1.0, + "content": "can be ill-conditioned, potentially causing", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 301, + 432 + ], + "score": 1.0, + "content": "the gradient to “explode” because of the inverses", + "type": "text" + }, + { + "bbox": [ + 301, + 419, + 322, + 430 + ], + "score": 0.9, + "content": "K ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 419, + 332, + 432 + ], + "score": 1.0, + "content": "et", + "type": "text" + }, + { + "bbox": [ + 332, + 419, + 351, + 430 + ], + "score": 0.9, + "content": "L ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "involved in the loss function. This can", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 430, + 504, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 419, + 443 + ], + "score": 1.0, + "content": "be avoided either by using the Moore-Penrose pseudo-inverse, or by replacing", + "type": "text" + }, + { + "bbox": [ + 419, + 430, + 440, + 441 + ], + "score": 0.89, + "content": "K ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 430, + 454, + 443 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 454, + 430, + 504, + 443 + ], + "score": 0.91, + "content": "( K + \\epsilon { \\bf 1 } ) ^ { - 1 }", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 441, + 375, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 277, + 453 + ], + "score": 1.0, + "content": "in the loss for some small positive number", + "type": "text" + }, + { + "bbox": [ + 278, + 444, + 283, + 452 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 442, + 353, + 453 + ], + "score": 1.0, + "content": ", and likewise for", + "type": "text" + }, + { + "bbox": [ + 353, + 441, + 371, + 452 + ], + "score": 0.91, + "content": "L ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 442, + 375, + 453 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 467, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 104, + 465, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 285, + 482 + ], + "score": 1.0, + "content": "Symmetries in the loss function—The loss", + "type": "text" + }, + { + "bbox": [ + 286, + 468, + 295, + 477 + ], + "score": 0.8, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 465, + 463, + 482 + ], + "score": 1.0, + "content": "only depends on the span of the variables", + "type": "text" + }, + { + "bbox": [ + 464, + 468, + 473, + 479 + ], + "score": 0.88, + "content": "f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 465, + 491, + 482 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 491, + 469, + 501, + 479 + ], + "score": 0.84, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 465, + 506, + 482 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 478, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 506, + 490 + ], + "score": 1.0, + "content": "hence it has a very large group of symmetries. In particular, it is invariant under a change of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 488, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 507, + 502 + ], + "score": 1.0, + "content": "norm of each variable independently from each other. Because of that, it is preferable not to have a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 501, + 472, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 472, + 512 + ], + "score": 1.0, + "content": "linear last layer. Using a hyperbolic tangent as last nonlinearity worked in our experiments.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 504, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "Regularization—In all our tests, dropout had no beneficial effect. In fact, our objective seems to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 536, + 370, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 370, + 549 + ], + "score": 1.0, + "content": "already provide a form of regularization, as shown in Section 5.2.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "title", + "bbox": [ + 108, + 565, + 264, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 266, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 266, + 580 + ], + "score": 1.0, + "content": "B THEORY IN MORE DETAILS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 290, + 604 + ], + "score": 1.0, + "content": "We consider two correlated random variables", + "type": "text" + }, + { + "bbox": [ + 290, + 592, + 300, + 601 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 589, + 318, + 604 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 318, + 591, + 328, + 601 + ], + "score": 0.78, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 589, + 471, + 604 + ], + "score": 1.0, + "content": "with a joint probability distribution", + "type": "text" + }, + { + "bbox": [ + 472, + 591, + 501, + 603 + ], + "score": 0.92, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 589, + 505, + 604 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "We assume that we are able to numerically evaluate expectations with respect to this distribu-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 613, + 504, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 504, + 626 + ], + "score": 1.0, + "content": "tion, for instance because we can sample from it. We want to use this ability in order to com-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 623, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 104, + 623, + 364, + 637 + ], + "score": 1.0, + "content": "pute expectations with respect to the conditional distributions", + "type": "text" + }, + { + "bbox": [ + 364, + 624, + 486, + 636 + ], + "score": 0.91, + "content": "p _ { X | Y } ( x | y ) \\ : = \\ : \\dot { p } ( x , y ) / p _ { X } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 623, + 505, + 637 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 635, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 229, + 649 + ], + "score": 0.91, + "content": "p _ { Y | X } ( y | x ) ~ = ~ p ( x , y ) / p _ { Y } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 636, + 264, + 650 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 264, + 635, + 357, + 650 + ], + "score": 0.91, + "content": "\\begin{array} { r } { p _ { X } ( x ) \\ = \\ \\sum _ { y } p ( x , y ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 636, + 379, + 650 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 379, + 637, + 470, + 649 + ], + "score": 0.91, + "content": "\\begin{array} { r } { p _ { Y } ( y ) \\ = \\ \\sum _ { x } p ( x , y ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "are the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 648, + 504, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 160, + 661 + ], + "score": 1.0, + "content": "marginals of", + "type": "text" + }, + { + "bbox": [ + 160, + 651, + 167, + 660 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 648, + 351, + 661 + ], + "score": 1.0, + "content": ". Below we sometime remove the subscripts", + "type": "text" + }, + { + "bbox": [ + 351, + 649, + 387, + 661 + ], + "score": 0.82, + "content": "X , X | Y", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 648, + 400, + 661 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 401, + 649, + 421, + 661 + ], + "score": 0.91, + "content": "Y | X", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 648, + 504, + 661 + ], + "score": 1.0, + "content": "if there is no ambi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 659, + 133, + 675 + ], + "spans": [ + { + "bbox": [ + 104, + 659, + 133, + 675 + ], + "score": 1.0, + "content": "guity.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 293, + 690 + ], + "score": 1.0, + "content": "For instance, suppose we generated samples of", + "type": "text" + }, + { + "bbox": [ + 293, + 679, + 300, + 689 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 677, + 324, + 690 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 324, + 679, + 331, + 687 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 677, + 455, + 690 + ], + "score": 1.0, + "content": ", through explicit knowledge of", + "type": "text" + }, + { + "bbox": [ + 455, + 678, + 478, + 689 + ], + "score": 0.86, + "content": "p _ { Y \\mid X }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 290, + 702 + ], + "score": 1.0, + "content": "the evaluation of expectations with respect to", + "type": "text" + }, + { + "bbox": [ + 291, + 690, + 313, + 700 + ], + "score": 0.9, + "content": "p _ { X | Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "is the subject of Bayesian inference. However,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 322, + 712 + ], + "score": 1.0, + "content": "this is generally done in a context where the variable", + "type": "text" + }, + { + "bbox": [ + 322, + 699, + 333, + 709 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "has low dimensionality and parameterizes", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 454, + 722 + ], + "score": 1.0, + "content": "a hand-crafted model. Our approach, however, is free of such a model and the variable", + "type": "text" + }, + { + "bbox": [ + 454, + 710, + 465, + 720 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "can be of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 719, + 210, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 210, + 734 + ], + "score": 1.0, + "content": "very high dimensionality.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 369, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 370, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 370, + 96 + ], + "score": 1.0, + "content": "A EXTRA INFORMATION ABOUT THE ALGORITHM", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 106, + 353, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 353, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 353, + 119 + ], + "score": 1.0, + "content": "A.1 ALTERNATIVE INTERPRETATION OF THE OBJECTIVE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 127, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 126, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 155, + 141 + ], + "score": 1.0, + "content": "If we write", + "type": "text" + }, + { + "bbox": [ + 156, + 128, + 216, + 140 + ], + "score": 0.92, + "content": "F _ { i j } : = f _ { j } ( x _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 126, + 236, + 141 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 128, + 297, + 140 + ], + "score": 0.91, + "content": "G _ { i j } : = g _ { j } ( y _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 126, + 506, + 141 + ], + "score": 1.0, + "content": "for the value of our variables on the dataset, then", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 138, + 473, + 154 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 162, + 153 + ], + "score": 0.87, + "content": "\\begin{array} { r } { K = \\frac { 1 } { N } F ^ { \\top } F } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 138, + 165, + 154 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 166, + 140, + 219, + 153 + ], + "score": 0.88, + "content": "\\begin{array} { r } { L = \\frac { 1 } { N } G ^ { \\top } G } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 138, + 237, + 154 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 140, + 290, + 153 + ], + "score": 0.91, + "content": "\\begin{array} { r } { A = \\frac { 1 } { N } G ^ { \\top } F } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 138, + 473, + 154 + ], + "score": 1.0, + "content": ". The DCCA objective can then be written as", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 126, + 506, + 154 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 158, + 370, + 174 + ], + "lines": [ + { + "bbox": [ + 240, + 158, + 370, + 174 + ], + "spans": [ + { + "bbox": [ + 240, + 158, + 370, + 174 + ], + "score": 0.91, + "content": "\\operatorname { T r } \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) = \\operatorname { T r } \\left( P Q \\right)", + "type": "interline_equation", + "image_path": "bb4208eb4b2d98539b5ca87ea715a86d6484f5c067aed48305d0f58fcc0ba1c9.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 240, + 158, + 370, + 174 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 179, + 505, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 134, + 195 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 180, + 222, + 194 + ], + "score": 0.93, + "content": "P = F ( F ^ { \\top } F ) ^ { - 1 } F ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 179, + 242, + 195 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 243, + 180, + 331, + 194 + ], + "score": 0.93, + "content": "Q = G ( G ^ { \\top } G ) ^ { - 1 } G ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 179, + 477, + 195 + ], + "score": 1.0, + "content": "are the projectors on the ranges of", + "type": "text" + }, + { + "bbox": [ + 477, + 182, + 486, + 191 + ], + "score": 0.84, + "content": "F", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 179, + 506, + 195 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 107, + 193, + 116, + 203 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "respectively. Hence, we are maximizing the overlap between those ranges (which represent pos-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "sible linear combinations of datapoints, respectively determined from variables of one or the other", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 213, + 181, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 181, + 227 + ], + "score": 1.0, + "content": "correlated views.)", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 179, + 506, + 227 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 240, + 182, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 183, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 183, + 254 + ], + "score": 1.0, + "content": "A.2 HEURISTIC", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "Batch size—In our experiments, we observed that the batch size during training needs to be an", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "order of magnitude larger than the number of variables (rank cutoff). When the batch size was", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "too small, learning seemed to converge normally in terms of training and test loss, but resulted", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "in variables which yield dramatically different losses when evaluated on larger batches, and yield", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 314, + 192, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 192, + 327 + ], + "score": 1.0, + "content": "spurious predictions.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 270, + 505, + 327 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 270, + 352 + ], + "score": 1.0, + "content": "Constant variables—The loss function", + "type": "text" + }, + { + "bbox": [ + 270, + 340, + 279, + 349 + ], + "score": 0.8, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 340, + 389, + 352 + ], + "score": 1.0, + "content": "takes value between 0 and", + "type": "text" + }, + { + "bbox": [ + 389, + 340, + 418, + 351 + ], + "score": 0.9, + "content": "k _ { 0 } - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "because the constant", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 351, + 504, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 504, + 362 + ], + "score": 1.0, + "content": "variable always has relevance 1. The constant variable could be enforced a priori rather than learned,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 361, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 375 + ], + "score": 1.0, + "content": "which, due to the objective, automatically forces the learned variables to have zero expectation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 371, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 386 + ], + "score": 1.0, + "content": "values (be orthogonal to the constant variable). This might have advantages in certain circumstances,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 383, + 406, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 406, + 397 + ], + "score": 1.0, + "content": "but in our experiments we found that this sometime hindered convergence.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 340, + 506, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 295, + 422 + ], + "score": 1.0, + "content": "Invertibility issues—The covariance matrices", + "type": "text" + }, + { + "bbox": [ + 295, + 410, + 306, + 419 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 407, + 324, + 422 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 324, + 410, + 332, + 419 + ], + "score": 0.8, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 407, + 506, + 422 + ], + "score": 1.0, + "content": "can be ill-conditioned, potentially causing", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 301, + 432 + ], + "score": 1.0, + "content": "the gradient to “explode” because of the inverses", + "type": "text" + }, + { + "bbox": [ + 301, + 419, + 322, + 430 + ], + "score": 0.9, + "content": "K ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 419, + 332, + 432 + ], + "score": 1.0, + "content": "et", + "type": "text" + }, + { + "bbox": [ + 332, + 419, + 351, + 430 + ], + "score": 0.9, + "content": "L ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "involved in the loss function. This can", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 430, + 504, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 419, + 443 + ], + "score": 1.0, + "content": "be avoided either by using the Moore-Penrose pseudo-inverse, or by replacing", + "type": "text" + }, + { + "bbox": [ + 419, + 430, + 440, + 441 + ], + "score": 0.89, + "content": "K ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 430, + 454, + 443 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 454, + 430, + 504, + 443 + ], + "score": 0.91, + "content": "( K + \\epsilon { \\bf 1 } ) ^ { - 1 }", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 441, + 375, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 277, + 453 + ], + "score": 1.0, + "content": "in the loss for some small positive number", + "type": "text" + }, + { + "bbox": [ + 278, + 444, + 283, + 452 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 442, + 353, + 453 + ], + "score": 1.0, + "content": ", and likewise for", + "type": "text" + }, + { + "bbox": [ + 353, + 441, + 371, + 452 + ], + "score": 0.91, + "content": "L ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 442, + 375, + 453 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 407, + 506, + 453 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 467, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 104, + 465, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 285, + 482 + ], + "score": 1.0, + "content": "Symmetries in the loss function—The loss", + "type": "text" + }, + { + "bbox": [ + 286, + 468, + 295, + 477 + ], + "score": 0.8, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 465, + 463, + 482 + ], + "score": 1.0, + "content": "only depends on the span of the variables", + "type": "text" + }, + { + "bbox": [ + 464, + 468, + 473, + 479 + ], + "score": 0.88, + "content": "f _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 465, + 491, + 482 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 491, + 469, + 501, + 479 + ], + "score": 0.84, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 465, + 506, + 482 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 478, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 506, + 490 + ], + "score": 1.0, + "content": "hence it has a very large group of symmetries. In particular, it is invariant under a change of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 488, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 507, + 502 + ], + "score": 1.0, + "content": "norm of each variable independently from each other. Because of that, it is preferable not to have a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 501, + 472, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 472, + 512 + ], + "score": 1.0, + "content": "linear last layer. Using a hyperbolic tangent as last nonlinearity worked in our experiments.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 465, + 507, + 512 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 504, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 505, + 538 + ], + "score": 1.0, + "content": "Regularization—In all our tests, dropout had no beneficial effect. In fact, our objective seems to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 536, + 370, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 370, + 549 + ], + "score": 1.0, + "content": "already provide a form of regularization, as shown in Section 5.2.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 524, + 505, + 549 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 565, + 264, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 266, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 266, + 580 + ], + "score": 1.0, + "content": "B THEORY IN MORE DETAILS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 290, + 604 + ], + "score": 1.0, + "content": "We consider two correlated random variables", + "type": "text" + }, + { + "bbox": [ + 290, + 592, + 300, + 601 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 589, + 318, + 604 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 318, + 591, + 328, + 601 + ], + "score": 0.78, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 589, + 471, + 604 + ], + "score": 1.0, + "content": "with a joint probability distribution", + "type": "text" + }, + { + "bbox": [ + 472, + 591, + 501, + 603 + ], + "score": 0.92, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 589, + 505, + 604 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "We assume that we are able to numerically evaluate expectations with respect to this distribu-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 613, + 504, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 504, + 626 + ], + "score": 1.0, + "content": "tion, for instance because we can sample from it. We want to use this ability in order to com-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 623, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 104, + 623, + 364, + 637 + ], + "score": 1.0, + "content": "pute expectations with respect to the conditional distributions", + "type": "text" + }, + { + "bbox": [ + 364, + 624, + 486, + 636 + ], + "score": 0.91, + "content": "p _ { X | Y } ( x | y ) \\ : = \\ : \\dot { p } ( x , y ) / p _ { X } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 623, + 505, + 637 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 635, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 229, + 649 + ], + "score": 0.91, + "content": "p _ { Y | X } ( y | x ) ~ = ~ p ( x , y ) / p _ { Y } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 636, + 264, + 650 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 264, + 635, + 357, + 650 + ], + "score": 0.91, + "content": "\\begin{array} { r } { p _ { X } ( x ) \\ = \\ \\sum _ { y } p ( x , y ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 636, + 379, + 650 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 379, + 637, + 470, + 649 + ], + "score": 0.91, + "content": "\\begin{array} { r } { p _ { Y } ( y ) \\ = \\ \\sum _ { x } p ( x , y ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "are the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 648, + 504, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 160, + 661 + ], + "score": 1.0, + "content": "marginals of", + "type": "text" + }, + { + "bbox": [ + 160, + 651, + 167, + 660 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 648, + 351, + 661 + ], + "score": 1.0, + "content": ". Below we sometime remove the subscripts", + "type": "text" + }, + { + "bbox": [ + 351, + 649, + 387, + 661 + ], + "score": 0.82, + "content": "X , X | Y", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 648, + 400, + 661 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 401, + 649, + 421, + 661 + ], + "score": 0.91, + "content": "Y | X", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 648, + 504, + 661 + ], + "score": 1.0, + "content": "if there is no ambi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 659, + 133, + 675 + ], + "spans": [ + { + "bbox": [ + 104, + 659, + 133, + 675 + ], + "score": 1.0, + "content": "guity.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 104, + 589, + 505, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 293, + 690 + ], + "score": 1.0, + "content": "For instance, suppose we generated samples of", + "type": "text" + }, + { + "bbox": [ + 293, + 679, + 300, + 689 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 677, + 324, + 690 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 324, + 679, + 331, + 687 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 677, + 455, + 690 + ], + "score": 1.0, + "content": ", through explicit knowledge of", + "type": "text" + }, + { + "bbox": [ + 455, + 678, + 478, + 689 + ], + "score": 0.86, + "content": "p _ { Y \\mid X }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 290, + 702 + ], + "score": 1.0, + "content": "the evaluation of expectations with respect to", + "type": "text" + }, + { + "bbox": [ + 291, + 690, + 313, + 700 + ], + "score": 0.9, + "content": "p _ { X | Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "is the subject of Bayesian inference. However,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 322, + 712 + ], + "score": 1.0, + "content": "this is generally done in a context where the variable", + "type": "text" + }, + { + "bbox": [ + 322, + 699, + 333, + 709 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "has low dimensionality and parameterizes", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 454, + 722 + ], + "score": 1.0, + "content": "a hand-crafted model. Our approach, however, is free of such a model and the variable", + "type": "text" + }, + { + "bbox": [ + 454, + 710, + 465, + 720 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "can be of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 719, + 210, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 210, + 734 + ], + "score": 1.0, + "content": "very high dimensionality.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 677, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 82, + 320, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "score": 1.0, + "content": "B.1 INNER PRODUCT ON PROBABILITY VECTORS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 505, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 466, + 115 + ], + "score": 1.0, + "content": "In order to define our strategy, we need to equip the spaces of probability distributions for", + "type": "text" + }, + { + "bbox": [ + 466, + 104, + 477, + 113 + ], + "score": 0.8, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 102, + 494, + 115 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 495, + 103, + 504, + 113 + ], + "score": 0.75, + "content": "Y", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 115, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 298, + 126 + ], + "score": 1.0, + "content": "with an inner product structure. Let us focus on", + "type": "text" + }, + { + "bbox": [ + 299, + 115, + 308, + 124 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 115, + 505, + 126 + ], + "score": 1.0, + "content": ", and assume that it takes discrete values to avoid", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "score": 1.0, + "content": "unnecessary technicalities. The set of probability vectors is a convex subset of the real linear space", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 136, + 313, + 148 + ], + "spans": [ + { + "bbox": [ + 107, + 136, + 147, + 147 + ], + "score": 0.91, + "content": "V _ { X } = \\mathbb { R } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 136, + 313, + 148 + ], + "score": 1.0, + "content": ". Let us equip this space with the product", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 153, + 362, + 183 + ], + "lines": [ + { + "bbox": [ + 249, + 153, + 362, + 183 + ], + "spans": [ + { + "bbox": [ + 249, + 153, + 362, + 183 + ], + "score": 0.95, + "content": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } : = \\sum _ { x } { \\frac { \\mu ( x ) \\mu ^ { \\prime } ( x ) } { p _ { X } ( x ) } }", + "type": "interline_equation", + "image_path": "1a12ff616c966e54cf1c282da79d92f2674b89326cb9bf7716455de6922806bb.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 249, + 153, + 362, + 168.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 249, + 168.0, + 362, + 183.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 189, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 188, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 138, + 203 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 139, + 190, + 187, + 201 + ], + "score": 0.92, + "content": "\\mu , \\mu ^ { \\prime } \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 188, + 253, + 203 + ], + "score": 1.0, + "content": ". We also write", + "type": "text" + }, + { + "bbox": [ + 253, + 189, + 324, + 202 + ], + "score": 0.93, + "content": "\\| \\mu \\| _ { X } ^ { 2 } = \\langle \\mu , \\mu \\rangle _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 188, + 505, + 203 + ], + "score": 1.0, + "content": ". Importantly, this depends explicitly on the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 202, + 213 + ], + "score": 1.0, + "content": "fixed probability vector", + "type": "text" + }, + { + "bbox": [ + 203, + 201, + 229, + 212 + ], + "score": 0.92, + "content": "p _ { X } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 199, + 380, + 213 + ], + "score": 1.0, + "content": ", which we took to be the marginal of", + "type": "text" + }, + { + "bbox": [ + 380, + 201, + 409, + 213 + ], + "score": 0.94, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 199, + 423, + 213 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 423, + 201, + 437, + 212 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "has full support,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 210, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 152, + 225 + ], + "score": 1.0, + "content": "this makes", + "type": "text" + }, + { + "bbox": [ + 152, + 212, + 167, + 223 + ], + "score": 0.9, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 210, + 456, + 225 + ], + "score": 1.0, + "content": "into a real inner product space. The same can be done for the variable", + "type": "text" + }, + { + "bbox": [ + 456, + 212, + 465, + 221 + ], + "score": 0.78, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 210, + 505, + 225 + ], + "score": 1.0, + "content": ", yielding", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 274, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 177, + 235 + ], + "score": 1.0, + "content": "the inner product", + "type": "text" + }, + { + "bbox": [ + 177, + 222, + 210, + 235 + ], + "score": 0.93, + "content": "\\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 223, + 226, + 235 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 226, + 222, + 270, + 234 + ], + "score": 0.92, + "content": "\\nu , \\nu ^ { \\prime } \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 223, + 274, + 235 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 186, + 251 + ], + "score": 1.0, + "content": "Had we interpreted", + "type": "text" + }, + { + "bbox": [ + 186, + 241, + 194, + 251 + ], + "score": 0.8, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 240, + 212, + 251 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 212, + 239, + 222, + 251 + ], + "score": 0.88, + "content": "\\mu ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 240, + 309, + 251 + ], + "score": 1.0, + "content": "as tangent vectors to", + "type": "text" + }, + { + "bbox": [ + 309, + 240, + 323, + 250 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 240, + 504, + 251 + ], + "score": 1.0, + "content": ", considered as a manifold, this would be the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "Fisher information (Riemannian) metric, as in Beny & Osborne (2015b). But this quantity is also ´", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 261, + 367, + 275 + ], + "score": 1.0, + "content": "meaningful for finite vectors: the induced norm distance between", + "type": "text" + }, + { + "bbox": [ + 367, + 263, + 381, + 273 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 261, + 489, + 275 + ], + "score": 1.0, + "content": "and any probability vector", + "type": "text" + }, + { + "bbox": [ + 489, + 264, + 495, + 273 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 184, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 121, + 285 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 272, + 133, + 284 + ], + "score": 0.89, + "content": "\\chi ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 271, + 184, + 285 + ], + "score": 1.0, + "content": "-divergence:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 281, + 374, + 297 + ], + "lines": [ + { + "bbox": [ + 236, + 281, + 374, + 297 + ], + "spans": [ + { + "bbox": [ + 236, + 281, + 374, + 297 + ], + "score": 0.91, + "content": "\\chi ^ { 2 } ( q , p _ { X } ) = \\langle q - p _ { X } , q - p _ { X } \\rangle _ { X } .", + "type": "interline_equation", + "image_path": "11d45a6962684bf4a6b418780675e28df62328e758433c27c52a68b89d784428.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 236, + 281, + 374, + 297 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 502, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 498, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 297, + 319 + ], + "score": 1.0, + "content": "The set of conditional probability distributions", + "type": "text" + }, + { + "bbox": [ + 297, + 308, + 319, + 318 + ], + "score": 0.91, + "content": "p _ { Y \\mid X }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 303, + 487, + 319 + ], + "score": 1.0, + "content": "form a stochastic map, i.e., a linear map", + "type": "text" + }, + { + "bbox": [ + 487, + 306, + 498, + 315 + ], + "score": 0.45, + "content": "\\mathcal { N }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 317, + 230, + 330 + ], + "spans": [ + { + "bbox": [ + 107, + 317, + 149, + 329 + ], + "score": 0.86, + "content": "V _ { X } \\to V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 317, + 153, + 330 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 153, + 317, + 199, + 329 + ], + "score": 0.86, + "content": "\\mu \\mapsto { \\mathcal { N } } ( \\mu )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 317, + 230, + 330 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 334, + 371, + 361 + ], + "lines": [ + { + "bbox": [ + 240, + 334, + 371, + 361 + ], + "spans": [ + { + "bbox": [ + 240, + 334, + 371, + 361 + ], + "score": 0.94, + "content": "\\mathcal { N } ( \\mu ) ( y ) = \\sum _ { x } p _ { Y | X } ( y | x ) \\mu ( x )", + "type": "interline_equation", + "image_path": "9f6aa5c7920c5d7ca7df6891ee1fa8eac0f726354cb2a190d7d57b440eed4689.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 240, + 334, + 371, + 347.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 240, + 347.5, + 371, + 361.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 173, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 174, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 137, + 379 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 137, + 367, + 170, + 378 + ], + "score": 0.91, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 365, + 174, + 379 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 378, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 379, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 317, + 398 + ], + "score": 1.0, + "content": "It is straightforward to check that the stochastic map", + "type": "text" + }, + { + "bbox": [ + 318, + 383, + 333, + 393 + ], + "score": 0.86, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 379, + 379, + 398 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 399, + 373, + 426 + ], + "lines": [ + { + "bbox": [ + 237, + 399, + 373, + 426 + ], + "spans": [ + { + "bbox": [ + 237, + 399, + 373, + 426 + ], + "score": 0.94, + "content": "\\mathcal { N ^ { * } } ( \\nu ) ( x ) = \\sum _ { x } p _ { X | Y } ( x | y ) \\nu ( x )", + "type": "interline_equation", + "image_path": "b4557f7afbfd59cd5d8971ce557d748cd9505101ca155f61defc18c9f1347d21.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 237, + 399, + 373, + 412.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 237, + 412.5, + 373, + 426.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 172, + 445 + ], + "score": 1.0, + "content": "is the transpose", + "type": "text" + }, + { + "bbox": [ + 173, + 432, + 188, + 441 + ], + "score": 0.86, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 429, + 201, + 445 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 201, + 432, + 212, + 441 + ], + "score": 0.81, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 429, + 506, + 445 + ], + "score": 1.0, + "content": "with respect to the inner products we defined (Ohya & Petz, 2004), i.e.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 442, + 220, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 133, + 455 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 443, + 164, + 453 + ], + "score": 0.91, + "content": "\\nu \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 442, + 183, + 455 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 183, + 443, + 215, + 454 + ], + "score": 0.91, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 442, + 220, + 455 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 453, + 365, + 467 + ], + "lines": [ + { + "bbox": [ + 245, + 453, + 365, + 467 + ], + "spans": [ + { + "bbox": [ + 245, + 453, + 365, + 467 + ], + "score": 0.93, + "content": "\\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } = \\langle \\mathcal { N } ^ { * } ( \\nu ) , \\mu \\rangle _ { X } .", + "type": "interline_equation", + "image_path": "59de3bedea85948fa8b40b185bb2572dc7663c5393d2e0340dbd6abdad9dd49e.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 245, + 453, + 365, + 467 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 334, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 335, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 195, + 484 + ], + "score": 1.0, + "content": "Also, we observe that", + "type": "text" + }, + { + "bbox": [ + 195, + 469, + 251, + 482 + ], + "score": 0.93, + "content": "\\mathcal { N } ( p _ { X } ) = p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 466, + 270, + 484 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 270, + 469, + 330, + 482 + ], + "score": 0.91, + "content": "\\mathcal { N } ^ { * } ( p _ { Y } ) = p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 466, + 335, + 484 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 494, + 289, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 291, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 291, + 507 + ], + "score": 1.0, + "content": "B.2 EIGEN-RELEVANCE DECOMPOSITION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 237, + 528 + ], + "score": 1.0, + "content": "We can use the inner products on", + "type": "text" + }, + { + "bbox": [ + 238, + 516, + 252, + 527 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 515, + 270, + 528 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 270, + 516, + 283, + 527 + ], + "score": 0.9, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "to define a singular value decomposition of the stochas-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 140, + 540 + ], + "score": 1.0, + "content": "tic map", + "type": "text" + }, + { + "bbox": [ + 140, + 527, + 151, + 536 + ], + "score": 0.8, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 525, + 319, + 540 + ], + "score": 1.0, + "content": ". That is, there is an orthonormal family", + "type": "text" + }, + { + "bbox": [ + 319, + 528, + 363, + 538 + ], + "score": 0.9, + "content": "u _ { 1 } , \\ldots , u _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 525, + 377, + 540 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 377, + 527, + 392, + 537 + ], + "score": 0.9, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 525, + 505, + 540 + ], + "score": 1.0, + "content": "and an orthonormal family", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 538, + 217, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 149, + 549 + ], + "score": 0.9, + "content": "v _ { 1 } , \\ldots , v _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 538, + 161, + 549 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 161, + 538, + 174, + 549 + ], + "score": 0.89, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 538, + 217, + 549 + ], + "score": 1.0, + "content": ", such that", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 273, + 547, + 337, + 561 + ], + "lines": [ + { + "bbox": [ + 273, + 547, + 337, + 561 + ], + "spans": [ + { + "bbox": [ + 273, + 547, + 337, + 561 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathcal { N } ( u _ { j } ) = \\eta _ { j } v _ { j } , } \\end{array}", + "type": "interline_equation", + "image_path": "8b1490f43f66bb1116b078ef858053ca3c89a396338ce45339fb6e4f50da5ac8.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 273, + 547, + 337, + 561 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 121, + 577 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 565, + 176, + 576 + ], + "score": 0.94, + "content": "j = 1 , \\dotsc , k", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 563, + 220, + 577 + ], + "score": 1.0, + "content": ". For each", + "type": "text" + }, + { + "bbox": [ + 220, + 565, + 241, + 576 + ], + "score": 0.46, + "content": "j , \\eta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 563, + 331, + 577 + ], + "score": 1.0, + "content": "is a singular value of", + "type": "text" + }, + { + "bbox": [ + 332, + 564, + 343, + 574 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 563, + 506, + 577 + ], + "score": 1.0, + "content": ", whose square we call the relevance of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 148, + 589 + ], + "score": 1.0, + "content": "the vector", + "type": "text" + }, + { + "bbox": [ + 148, + 578, + 158, + 588 + ], + "score": 0.83, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 576, + 204, + 589 + ], + "score": 1.0, + "content": ". Moreover", + "type": "text" + }, + { + "bbox": [ + 204, + 577, + 246, + 588 + ], + "score": 0.91, + "content": "\\eta _ { j } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 576, + 285, + 589 + ], + "score": 1.0, + "content": "since the", + "type": "text" + }, + { + "bbox": [ + 285, + 576, + 297, + 588 + ], + "score": 0.9, + "content": "\\chi ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "divergence is contractive under any stochastic map.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 586, + 257, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 150, + 600 + ], + "score": 1.0, + "content": "Given that", + "type": "text" + }, + { + "bbox": [ + 150, + 587, + 165, + 597 + ], + "score": 0.85, + "content": "{ \\ddot { \\mathcal { N } } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 586, + 242, + 600 + ], + "score": 1.0, + "content": "is the transpose of", + "type": "text" + }, + { + "bbox": [ + 242, + 588, + 253, + 597 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 586, + 257, + 600 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "interline_equation", + "bbox": [ + 271, + 604, + 340, + 619 + ], + "lines": [ + { + "bbox": [ + 271, + 604, + 340, + 619 + ], + "spans": [ + { + "bbox": [ + 271, + 604, + 340, + 619 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathcal { N } ^ { * } ( v _ { j } ) = \\eta _ { j } u _ { j } . } \\end{array}", + "type": "interline_equation", + "image_path": "7bdb6bc1c44d9193a0074f170fe1194569b52d796fb680d65a6fafc14cfb139d.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 271, + 604, + 340, + 619 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 162, + 636 + ], + "score": 1.0, + "content": "Equivalently,", + "type": "text" + }, + { + "bbox": [ + 163, + 626, + 174, + 636 + ], + "score": 0.83, + "content": "u _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 621, + 262, + 636 + ], + "score": 1.0, + "content": "is an eigenvector of", + "type": "text" + }, + { + "bbox": [ + 262, + 623, + 299, + 634 + ], + "score": 0.91, + "content": "\\mathcal { N } ^ { \\ast } \\circ \\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 621, + 319, + 636 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 319, + 626, + 330, + 636 + ], + "score": 0.85, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 621, + 421, + 636 + ], + "score": 1.0, + "content": "is an eigenvectors of", + "type": "text" + }, + { + "bbox": [ + 421, + 623, + 457, + 634 + ], + "score": 0.9, + "content": "\\mathcal { N } \\circ \\mathcal { N } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 621, + 506, + 636 + ], + "score": 1.0, + "content": ", both with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 167, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 151, + 651 + ], + "score": 1.0, + "content": "eigenvalue", + "type": "text" + }, + { + "bbox": [ + 151, + 635, + 162, + 649 + ], + "score": 0.89, + "content": "\\eta _ { j } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 634, + 167, + 651 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 104, + 650, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 650, + 143, + 667 + ], + "score": 1.0, + "content": "Because", + "type": "text" + }, + { + "bbox": [ + 143, + 653, + 154, + 663 + ], + "score": 0.79, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 650, + 180, + 667 + ], + "score": 1.0, + "content": "maps", + "type": "text" + }, + { + "bbox": [ + 181, + 655, + 194, + 664 + ], + "score": 0.84, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 650, + 207, + 667 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 208, + 655, + 221, + 664 + ], + "score": 0.85, + "content": "p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 650, + 383, + 667 + ], + "score": 1.0, + "content": ", we always have the dual eigenvectors", + "type": "text" + }, + { + "bbox": [ + 383, + 655, + 423, + 664 + ], + "score": 0.89, + "content": "u _ { 0 } ~ = ~ p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 650, + 443, + 667 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 443, + 655, + 483, + 664 + ], + "score": 0.89, + "content": "v _ { 0 } = p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 650, + 506, + 667 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 663, + 163, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 163, + 676 + ], + "score": 1.0, + "content": "eigenvalue 1.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 108, + 689, + 255, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 256, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 256, + 701 + ], + "score": 1.0, + "content": "B.3 LOW-RANK APPROXIMATION", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 211, + 722 + ], + "score": 1.0, + "content": "Typically, the dimension", + "type": "text" + }, + { + "bbox": [ + 211, + 710, + 218, + 720 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "of the space of probabilities is more than astronomically large. For", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 210, + 733 + ], + "score": 1.0, + "content": "instance, if the values of", + "type": "text" + }, + { + "bbox": [ + 210, + 721, + 221, + 730 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 720, + 398, + 733 + ], + "score": 1.0, + "content": "consists of small 256 gray level images of", + "type": "text" + }, + { + "bbox": [ + 398, + 721, + 433, + 731 + ], + "score": 0.89, + "content": "2 8 \\times 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 720, + 485, + 733 + ], + "score": 1.0, + "content": "pixels, then", + "type": "text" + }, + { + "bbox": [ + 485, + 721, + 505, + 731 + ], + "score": 0.85, + "content": "k =", + "type": "inline_equation" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 109, + 82, + 320, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 322, + 95 + ], + "score": 1.0, + "content": "B.1 INNER PRODUCT ON PROBABILITY VECTORS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 505, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 466, + 115 + ], + "score": 1.0, + "content": "In order to define our strategy, we need to equip the spaces of probability distributions for", + "type": "text" + }, + { + "bbox": [ + 466, + 104, + 477, + 113 + ], + "score": 0.8, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 102, + 494, + 115 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 495, + 103, + 504, + 113 + ], + "score": 0.75, + "content": "Y", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 115, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 298, + 126 + ], + "score": 1.0, + "content": "with an inner product structure. Let us focus on", + "type": "text" + }, + { + "bbox": [ + 299, + 115, + 308, + 124 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 115, + 505, + 126 + ], + "score": 1.0, + "content": ", and assume that it takes discrete values to avoid", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "score": 1.0, + "content": "unnecessary technicalities. The set of probability vectors is a convex subset of the real linear space", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 136, + 313, + 148 + ], + "spans": [ + { + "bbox": [ + 107, + 136, + 147, + 147 + ], + "score": 0.91, + "content": "V _ { X } = \\mathbb { R } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 136, + 313, + 148 + ], + "score": 1.0, + "content": ". Let us equip this space with the product", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 102, + 505, + 148 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 153, + 362, + 183 + ], + "lines": [ + { + "bbox": [ + 249, + 153, + 362, + 183 + ], + "spans": [ + { + "bbox": [ + 249, + 153, + 362, + 183 + ], + "score": 0.95, + "content": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } : = \\sum _ { x } { \\frac { \\mu ( x ) \\mu ^ { \\prime } ( x ) } { p _ { X } ( x ) } }", + "type": "interline_equation", + "image_path": "1a12ff616c966e54cf1c282da79d92f2674b89326cb9bf7716455de6922806bb.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 249, + 153, + 362, + 168.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 249, + 168.0, + 362, + 183.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 189, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 188, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 138, + 203 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 139, + 190, + 187, + 201 + ], + "score": 0.92, + "content": "\\mu , \\mu ^ { \\prime } \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 188, + 253, + 203 + ], + "score": 1.0, + "content": ". We also write", + "type": "text" + }, + { + "bbox": [ + 253, + 189, + 324, + 202 + ], + "score": 0.93, + "content": "\\| \\mu \\| _ { X } ^ { 2 } = \\langle \\mu , \\mu \\rangle _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 188, + 505, + 203 + ], + "score": 1.0, + "content": ". Importantly, this depends explicitly on the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 202, + 213 + ], + "score": 1.0, + "content": "fixed probability vector", + "type": "text" + }, + { + "bbox": [ + 203, + 201, + 229, + 212 + ], + "score": 0.92, + "content": "p _ { X } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 199, + 380, + 213 + ], + "score": 1.0, + "content": ", which we took to be the marginal of", + "type": "text" + }, + { + "bbox": [ + 380, + 201, + 409, + 213 + ], + "score": 0.94, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 199, + 423, + 213 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 423, + 201, + 437, + 212 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "has full support,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 210, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 152, + 225 + ], + "score": 1.0, + "content": "this makes", + "type": "text" + }, + { + "bbox": [ + 152, + 212, + 167, + 223 + ], + "score": 0.9, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 210, + 456, + 225 + ], + "score": 1.0, + "content": "into a real inner product space. The same can be done for the variable", + "type": "text" + }, + { + "bbox": [ + 456, + 212, + 465, + 221 + ], + "score": 0.78, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 210, + 505, + 225 + ], + "score": 1.0, + "content": ", yielding", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 274, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 177, + 235 + ], + "score": 1.0, + "content": "the inner product", + "type": "text" + }, + { + "bbox": [ + 177, + 222, + 210, + 235 + ], + "score": 0.93, + "content": "\\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 223, + 226, + 235 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 226, + 222, + 270, + 234 + ], + "score": 0.92, + "content": "\\nu , \\nu ^ { \\prime } \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 223, + 274, + 235 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 188, + 505, + 235 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 239, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 186, + 251 + ], + "score": 1.0, + "content": "Had we interpreted", + "type": "text" + }, + { + "bbox": [ + 186, + 241, + 194, + 251 + ], + "score": 0.8, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 240, + 212, + 251 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 212, + 239, + 222, + 251 + ], + "score": 0.88, + "content": "\\mu ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 240, + 309, + 251 + ], + "score": 1.0, + "content": "as tangent vectors to", + "type": "text" + }, + { + "bbox": [ + 309, + 240, + 323, + 250 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 240, + 504, + 251 + ], + "score": 1.0, + "content": ", considered as a manifold, this would be the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "Fisher information (Riemannian) metric, as in Beny & Osborne (2015b). But this quantity is also ´", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 261, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 261, + 367, + 275 + ], + "score": 1.0, + "content": "meaningful for finite vectors: the induced norm distance between", + "type": "text" + }, + { + "bbox": [ + 367, + 263, + 381, + 273 + ], + "score": 0.86, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 261, + 489, + 275 + ], + "score": 1.0, + "content": "and any probability vector", + "type": "text" + }, + { + "bbox": [ + 489, + 264, + 495, + 273 + ], + "score": 0.8, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 261, + 505, + 275 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 184, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 121, + 285 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 272, + 133, + 284 + ], + "score": 0.89, + "content": "\\chi ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 271, + 184, + 285 + ], + "score": 1.0, + "content": "-divergence:", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 104, + 239, + 505, + 285 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 281, + 374, + 297 + ], + "lines": [ + { + "bbox": [ + 236, + 281, + 374, + 297 + ], + "spans": [ + { + "bbox": [ + 236, + 281, + 374, + 297 + ], + "score": 0.91, + "content": "\\chi ^ { 2 } ( q , p _ { X } ) = \\langle q - p _ { X } , q - p _ { X } \\rangle _ { X } .", + "type": "interline_equation", + "image_path": "11d45a6962684bf4a6b418780675e28df62328e758433c27c52a68b89d784428.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 236, + 281, + 374, + 297 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 502, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 498, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 297, + 319 + ], + "score": 1.0, + "content": "The set of conditional probability distributions", + "type": "text" + }, + { + "bbox": [ + 297, + 308, + 319, + 318 + ], + "score": 0.91, + "content": "p _ { Y \\mid X }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 303, + 487, + 319 + ], + "score": 1.0, + "content": "form a stochastic map, i.e., a linear map", + "type": "text" + }, + { + "bbox": [ + 487, + 306, + 498, + 315 + ], + "score": 0.45, + "content": "\\mathcal { N }", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 317, + 230, + 330 + ], + "spans": [ + { + "bbox": [ + 107, + 317, + 149, + 329 + ], + "score": 0.86, + "content": "V _ { X } \\to V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 317, + 153, + 330 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 153, + 317, + 199, + 329 + ], + "score": 0.86, + "content": "\\mu \\mapsto { \\mathcal { N } } ( \\mu )", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 317, + 230, + 330 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 303, + 498, + 330 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 334, + 371, + 361 + ], + "lines": [ + { + "bbox": [ + 240, + 334, + 371, + 361 + ], + "spans": [ + { + "bbox": [ + 240, + 334, + 371, + 361 + ], + "score": 0.94, + "content": "\\mathcal { N } ( \\mu ) ( y ) = \\sum _ { x } p _ { Y | X } ( y | x ) \\mu ( x )", + "type": "interline_equation", + "image_path": "9f6aa5c7920c5d7ca7df6891ee1fa8eac0f726354cb2a190d7d57b440eed4689.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 240, + 334, + 371, + 347.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 240, + 347.5, + 371, + 361.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 173, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 174, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 137, + 379 + ], + "score": 1.0, + "content": "for any", + "type": "text" + }, + { + "bbox": [ + 137, + 367, + 170, + 378 + ], + "score": 0.91, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 365, + 174, + 379 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 365, + 174, + 379 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 378, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 379, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 317, + 398 + ], + "score": 1.0, + "content": "It is straightforward to check that the stochastic map", + "type": "text" + }, + { + "bbox": [ + 318, + 383, + 333, + 393 + ], + "score": 0.86, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 379, + 379, + 398 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 379, + 379, + 398 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 399, + 373, + 426 + ], + "lines": [ + { + "bbox": [ + 237, + 399, + 373, + 426 + ], + "spans": [ + { + "bbox": [ + 237, + 399, + 373, + 426 + ], + "score": 0.94, + "content": "\\mathcal { N ^ { * } } ( \\nu ) ( x ) = \\sum _ { x } p _ { X | Y } ( x | y ) \\nu ( x )", + "type": "interline_equation", + "image_path": "b4557f7afbfd59cd5d8971ce557d748cd9505101ca155f61defc18c9f1347d21.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 237, + 399, + 373, + 412.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 237, + 412.5, + 373, + 426.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 172, + 445 + ], + "score": 1.0, + "content": "is the transpose", + "type": "text" + }, + { + "bbox": [ + 173, + 432, + 188, + 441 + ], + "score": 0.86, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 429, + 201, + 445 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 201, + 432, + 212, + 441 + ], + "score": 0.81, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 429, + 506, + 445 + ], + "score": 1.0, + "content": "with respect to the inner products we defined (Ohya & Petz, 2004), i.e.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 442, + 220, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 133, + 455 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 443, + 164, + 453 + ], + "score": 0.91, + "content": "\\nu \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 442, + 183, + 455 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 183, + 443, + 215, + 454 + ], + "score": 0.91, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 442, + 220, + 455 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 429, + 506, + 455 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 453, + 365, + 467 + ], + "lines": [ + { + "bbox": [ + 245, + 453, + 365, + 467 + ], + "spans": [ + { + "bbox": [ + 245, + 453, + 365, + 467 + ], + "score": 0.93, + "content": "\\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } = \\langle \\mathcal { N } ^ { * } ( \\nu ) , \\mu \\rangle _ { X } .", + "type": "interline_equation", + "image_path": "59de3bedea85948fa8b40b185bb2572dc7663c5393d2e0340dbd6abdad9dd49e.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 245, + 453, + 365, + 467 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 334, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 335, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 195, + 484 + ], + "score": 1.0, + "content": "Also, we observe that", + "type": "text" + }, + { + "bbox": [ + 195, + 469, + 251, + 482 + ], + "score": 0.93, + "content": "\\mathcal { N } ( p _ { X } ) = p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 466, + 270, + 484 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 270, + 469, + 330, + 482 + ], + "score": 0.91, + "content": "\\mathcal { N } ^ { * } ( p _ { Y } ) = p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 466, + 335, + 484 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 466, + 335, + 484 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 494, + 289, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 291, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 291, + 507 + ], + "score": 1.0, + "content": "B.2 EIGEN-RELEVANCE DECOMPOSITION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 237, + 528 + ], + "score": 1.0, + "content": "We can use the inner products on", + "type": "text" + }, + { + "bbox": [ + 238, + 516, + 252, + 527 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 515, + 270, + 528 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 270, + 516, + 283, + 527 + ], + "score": 0.9, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "to define a singular value decomposition of the stochas-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 525, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 140, + 540 + ], + "score": 1.0, + "content": "tic map", + "type": "text" + }, + { + "bbox": [ + 140, + 527, + 151, + 536 + ], + "score": 0.8, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 525, + 319, + 540 + ], + "score": 1.0, + "content": ". That is, there is an orthonormal family", + "type": "text" + }, + { + "bbox": [ + 319, + 528, + 363, + 538 + ], + "score": 0.9, + "content": "u _ { 1 } , \\ldots , u _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 525, + 377, + 540 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 377, + 527, + 392, + 537 + ], + "score": 0.9, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 525, + 505, + 540 + ], + "score": 1.0, + "content": "and an orthonormal family", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 538, + 217, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 149, + 549 + ], + "score": 0.9, + "content": "v _ { 1 } , \\ldots , v _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 538, + 161, + 549 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 161, + 538, + 174, + 549 + ], + "score": 0.89, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 538, + 217, + 549 + ], + "score": 1.0, + "content": ", such that", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 515, + 505, + 549 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 273, + 547, + 337, + 561 + ], + "lines": [ + { + "bbox": [ + 273, + 547, + 337, + 561 + ], + "spans": [ + { + "bbox": [ + 273, + 547, + 337, + 561 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathcal { N } ( u _ { j } ) = \\eta _ { j } v _ { j } , } \\end{array}", + "type": "interline_equation", + "image_path": "8b1490f43f66bb1116b078ef858053ca3c89a396338ce45339fb6e4f50da5ac8.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 273, + 547, + 337, + 561 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 121, + 577 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 122, + 565, + 176, + 576 + ], + "score": 0.94, + "content": "j = 1 , \\dotsc , k", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 563, + 220, + 577 + ], + "score": 1.0, + "content": ". For each", + "type": "text" + }, + { + "bbox": [ + 220, + 565, + 241, + 576 + ], + "score": 0.46, + "content": "j , \\eta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 563, + 331, + 577 + ], + "score": 1.0, + "content": "is a singular value of", + "type": "text" + }, + { + "bbox": [ + 332, + 564, + 343, + 574 + ], + "score": 0.83, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 563, + 506, + 577 + ], + "score": 1.0, + "content": ", whose square we call the relevance of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 148, + 589 + ], + "score": 1.0, + "content": "the vector", + "type": "text" + }, + { + "bbox": [ + 148, + 578, + 158, + 588 + ], + "score": 0.83, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 576, + 204, + 589 + ], + "score": 1.0, + "content": ". Moreover", + "type": "text" + }, + { + "bbox": [ + 204, + 577, + 246, + 588 + ], + "score": 0.91, + "content": "\\eta _ { j } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 576, + 285, + 589 + ], + "score": 1.0, + "content": "since the", + "type": "text" + }, + { + "bbox": [ + 285, + 576, + 297, + 588 + ], + "score": 0.9, + "content": "\\chi ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "divergence is contractive under any stochastic map.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 586, + 257, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 150, + 600 + ], + "score": 1.0, + "content": "Given that", + "type": "text" + }, + { + "bbox": [ + 150, + 587, + 165, + 597 + ], + "score": 0.85, + "content": "{ \\ddot { \\mathcal { N } } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 586, + 242, + 600 + ], + "score": 1.0, + "content": "is the transpose of", + "type": "text" + }, + { + "bbox": [ + 242, + 588, + 253, + 597 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 586, + 257, + 600 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 563, + 506, + 600 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 271, + 604, + 340, + 619 + ], + "lines": [ + { + "bbox": [ + 271, + 604, + 340, + 619 + ], + "spans": [ + { + "bbox": [ + 271, + 604, + 340, + 619 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathcal { N } ^ { * } ( v _ { j } ) = \\eta _ { j } u _ { j } . } \\end{array}", + "type": "interline_equation", + "image_path": "7bdb6bc1c44d9193a0074f170fe1194569b52d796fb680d65a6fafc14cfb139d.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 271, + 604, + 340, + 619 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 162, + 636 + ], + "score": 1.0, + "content": "Equivalently,", + "type": "text" + }, + { + "bbox": [ + 163, + 626, + 174, + 636 + ], + "score": 0.83, + "content": "u _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 621, + 262, + 636 + ], + "score": 1.0, + "content": "is an eigenvector of", + "type": "text" + }, + { + "bbox": [ + 262, + 623, + 299, + 634 + ], + "score": 0.91, + "content": "\\mathcal { N } ^ { \\ast } \\circ \\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 621, + 319, + 636 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 319, + 626, + 330, + 636 + ], + "score": 0.85, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 621, + 421, + 636 + ], + "score": 1.0, + "content": "is an eigenvectors of", + "type": "text" + }, + { + "bbox": [ + 421, + 623, + 457, + 634 + ], + "score": 0.9, + "content": "\\mathcal { N } \\circ \\mathcal { N } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 621, + 506, + 636 + ], + "score": 1.0, + "content": ", both with", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 167, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 151, + 651 + ], + "score": 1.0, + "content": "eigenvalue", + "type": "text" + }, + { + "bbox": [ + 151, + 635, + 162, + 649 + ], + "score": 0.89, + "content": "\\eta _ { j } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 634, + 167, + 651 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 621, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 104, + 650, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 104, + 650, + 143, + 667 + ], + "score": 1.0, + "content": "Because", + "type": "text" + }, + { + "bbox": [ + 143, + 653, + 154, + 663 + ], + "score": 0.79, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 650, + 180, + 667 + ], + "score": 1.0, + "content": "maps", + "type": "text" + }, + { + "bbox": [ + 181, + 655, + 194, + 664 + ], + "score": 0.84, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 650, + 207, + 667 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 208, + 655, + 221, + 664 + ], + "score": 0.85, + "content": "p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 650, + 383, + 667 + ], + "score": 1.0, + "content": ", we always have the dual eigenvectors", + "type": "text" + }, + { + "bbox": [ + 383, + 655, + 423, + 664 + ], + "score": 0.89, + "content": "u _ { 0 } ~ = ~ p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 650, + 443, + 667 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 443, + 655, + 483, + 664 + ], + "score": 0.89, + "content": "v _ { 0 } = p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 650, + 506, + 667 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 663, + 163, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 163, + 676 + ], + "score": 1.0, + "content": "eigenvalue 1.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 650, + 506, + 676 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 689, + 255, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 256, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 256, + 701 + ], + "score": 1.0, + "content": "B.3 LOW-RANK APPROXIMATION", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 211, + 722 + ], + "score": 1.0, + "content": "Typically, the dimension", + "type": "text" + }, + { + "bbox": [ + 211, + 710, + 218, + 720 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "of the space of probabilities is more than astronomically large. For", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 210, + 733 + ], + "score": 1.0, + "content": "instance, if the values of", + "type": "text" + }, + { + "bbox": [ + 210, + 721, + 221, + 730 + ], + "score": 0.83, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 720, + 398, + 733 + ], + "score": 1.0, + "content": "consists of small 256 gray level images of", + "type": "text" + }, + { + "bbox": [ + 398, + 721, + 433, + 731 + ], + "score": 0.89, + "content": "2 8 \\times 2 8", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 720, + 485, + 733 + ], + "score": 1.0, + "content": "pixels, then", + "type": "text" + }, + { + "bbox": [ + 485, + 721, + 505, + 731 + ], + "score": 0.85, + "content": "k =", + "type": "inline_equation" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 710, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 176, + 93 + ], + "score": 0.91, + "content": "2 5 6 ^ { 2 8 ^ { 2 } } \\simeq 1 0 ^ { 1 8 8 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 77, + 507, + 97 + ], + "score": 1.0, + "content": ". However, in many case, only very few of these dimensions may be relevant for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 267, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 267, + 106 + ], + "score": 1.0, + "content": "the purpose of inferring other variables.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 144 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 504, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 281, + 123 + ], + "score": 1.0, + "content": "The core of our approach is to approximate", + "type": "text" + }, + { + "bbox": [ + 281, + 111, + 292, + 120 + ], + "score": 0.8, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 110, + 310, + 123 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 310, + 110, + 325, + 120 + ], + "score": 0.87, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 110, + 493, + 123 + ], + "score": 1.0, + "content": "by restricting them to the span of the first", + "type": "text" + }, + { + "bbox": [ + 493, + 111, + 504, + 122 + ], + "score": 0.87, + "content": "k _ { 0 }", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 119, + 507, + 136 + ], + "spans": [ + { + "bbox": [ + 104, + 119, + 159, + 136 + ], + "score": 1.0, + "content": "eigenvectors", + "type": "text" + }, + { + "bbox": [ + 159, + 123, + 171, + 133 + ], + "score": 0.85, + "content": "u _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 119, + 189, + 136 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 190, + 123, + 200, + 134 + ], + "score": 0.86, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 119, + 315, + 136 + ], + "score": 1.0, + "content": "with largest singular values", + "type": "text" + }, + { + "bbox": [ + 315, + 123, + 325, + 134 + ], + "score": 0.86, + "content": "\\eta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 119, + 491, + 136 + ], + "score": 1.0, + "content": ". That is, if we order the singular values", + "type": "text" + }, + { + "bbox": [ + 491, + 124, + 501, + 134 + ], + "score": 0.84, + "content": "\\eta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 119, + 507, + 136 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 132, + 394, + 145 + ], + "spans": [ + { + "bbox": [ + 107, + 133, + 158, + 144 + ], + "score": 0.92, + "content": "j = 1 , \\dots , k", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 132, + 394, + 145 + ], + "score": 1.0, + "content": "in decreasing order, we propose to use the approximations", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 146, + 364, + 204 + ], + "lines": [ + { + "bbox": [ + 247, + 146, + 364, + 204 + ], + "spans": [ + { + "bbox": [ + 247, + 146, + 364, + 204 + ], + "score": 0.93, + "content": "\\begin{array} { l } { { \\displaystyle \\mathcal { N } _ { 0 } ( \\mu ) = \\sum _ { j \\leq k _ { 0 } } \\eta _ { j } \\langle u _ { j } , \\mu \\rangle _ { X } v _ { j } } } \\\\ { { \\displaystyle \\mathcal { N } _ { 0 } ^ { * } ( \\nu ) = \\sum _ { j \\leq k _ { 0 } } \\eta _ { j } \\langle v _ { j } , \\nu \\rangle _ { Y } u _ { j } } } \\end{array}", + "type": "interline_equation", + "image_path": "f3a5327242bb5b1d075b4f3e55813df2cb07b711bc8ac260eb13179d35ab16b0.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 146, + 364, + 175.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 247, + 175.0, + 364, + 204.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 219, + 493, + 232 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 494, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 117, + 234 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 220, + 128, + 230 + ], + "score": 0.82, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 217, + 146, + 234 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 146, + 220, + 161, + 230 + ], + "score": 0.87, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 217, + 252, + 234 + ], + "score": 1.0, + "content": "respectively, for some", + "type": "text" + }, + { + "bbox": [ + 252, + 220, + 263, + 231 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 217, + 378, + 234 + ], + "score": 1.0, + "content": "typically much smaller than", + "type": "text" + }, + { + "bbox": [ + 378, + 220, + 384, + 230 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 217, + 421, + 234 + ], + "score": 1.0, + "content": ", and any", + "type": "text" + }, + { + "bbox": [ + 422, + 220, + 454, + 231 + ], + "score": 0.86, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 217, + 458, + 234 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 459, + 220, + 490, + 231 + ], + "score": 0.81, + "content": "\\nu \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 217, + 494, + 234 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 385, + 249 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 385, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 227, + 251 + ], + "score": 1.0, + "content": "We denote the components of", + "type": "text" + }, + { + "bbox": [ + 227, + 236, + 241, + 248 + ], + "score": 0.89, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 234, + 259, + 251 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 259, + 236, + 274, + 249 + ], + "score": 0.9, + "content": "\\mathcal { N } _ { 0 } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 234, + 288, + 251 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 289, + 236, + 316, + 249 + ], + "score": 0.92, + "content": "q ( y | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 234, + 334, + 251 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 334, + 236, + 361, + 249 + ], + "score": 0.91, + "content": "q ( x | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 234, + 385, + 251 + ], + "score": 1.0, + "content": ", e.g.,", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 251, + 366, + 277 + ], + "lines": [ + { + "bbox": [ + 245, + 251, + 366, + 277 + ], + "spans": [ + { + "bbox": [ + 245, + 251, + 366, + 277 + ], + "score": 0.94, + "content": "{ \\mathcal { N } } _ { 0 } ( \\mu ) ( y ) = \\sum _ { x } q ( y | x ) \\mu ( x ) .", + "type": "interline_equation", + "image_path": "a6e2b9880b1b37ead6fd6a4ffe84590a7907071edbd961c9b9d70841b48801b8.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 245, + 251, + 366, + 277 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 279, + 504, + 312 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 132, + 293 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 132, + 280, + 146, + 290 + ], + "score": 0.88, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 278, + 165, + 293 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 166, + 279, + 181, + 291 + ], + "score": 0.9, + "content": "\\mathcal { N } _ { 0 } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 278, + 292, + 293 + ], + "score": 1.0, + "content": "are adjoint, we can define", + "type": "text" + }, + { + "bbox": [ + 292, + 279, + 458, + 291 + ], + "score": 0.93, + "content": "q ( x , y ) = q ( x | y ) p _ { Y } ( y ) = q ( y | x ) p _ { X } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 278, + 505, + 293 + ], + "score": 1.0, + "content": ". Although", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 289, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 176, + 303 + ], + "score": 1.0, + "content": "the marginals of", + "type": "text" + }, + { + "bbox": [ + 176, + 291, + 205, + 302 + ], + "score": 0.93, + "content": "\\dot { \\boldsymbol { q } } ( \\boldsymbol { x } , \\boldsymbol { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 289, + 338, + 303 + ], + "score": 1.0, + "content": "are the probability distributions", + "type": "text" + }, + { + "bbox": [ + 338, + 293, + 352, + 302 + ], + "score": 0.87, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 289, + 372, + 303 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 372, + 293, + 385, + 302 + ], + "score": 0.86, + "content": "p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 289, + 443, + 303 + ], + "score": 1.0, + "content": ", the numbers", + "type": "text" + }, + { + "bbox": [ + 443, + 291, + 472, + 302 + ], + "score": 0.92, + "content": "q ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 289, + 506, + 303 + ], + "score": 1.0, + "content": "are not", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 302, + 189, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 189, + 313 + ], + "score": 1.0, + "content": "necessarily positive.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 108, + 318, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 290, + 331 + ], + "score": 1.0, + "content": "The quality of this approximation for a given", + "type": "text" + }, + { + "bbox": [ + 290, + 319, + 302, + 329 + ], + "score": 0.87, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "does not directly depend on the dimensionality of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 329, + 504, + 342 + ], + "spans": [ + { + "bbox": [ + 107, + 330, + 117, + 339 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 329, + 135, + 342 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 135, + 330, + 144, + 339 + ], + "score": 0.75, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 329, + 493, + 342 + ], + "score": 1.0, + "content": ", but only on the amount of correlations between the two variables. Our aim is to use a", + "type": "text" + }, + { + "bbox": [ + 493, + 330, + 504, + 340 + ], + "score": 0.87, + "content": "k _ { 0 }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 340, + 417, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 256, + 354 + ], + "score": 1.0, + "content": "small enough that the components of", + "type": "text" + }, + { + "bbox": [ + 256, + 340, + 270, + 352 + ], + "score": 0.9, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 340, + 288, + 354 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 288, + 340, + 303, + 352 + ], + "score": 0.91, + "content": "\\mathcal { N } _ { 0 } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 340, + 417, + 354 + ], + "score": 1.0, + "content": "can be computed explicitly.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 414, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 415, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 159, + 367 + ], + "score": 1.0, + "content": "Theorem 1.", + "type": "text" + }, + { + "bbox": [ + 160, + 354, + 174, + 365 + ], + "score": 0.86, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 352, + 249, + 367 + ], + "score": 1.0, + "content": "is the map of rank", + "type": "text" + }, + { + "bbox": [ + 250, + 354, + 260, + 365 + ], + "score": 0.87, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 352, + 415, + 367 + ], + "score": 1.0, + "content": "which minimizes the average distance", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 367, + 425, + 399 + ], + "lines": [ + { + "bbox": [ + 185, + 367, + 425, + 399 + ], + "spans": [ + { + "bbox": [ + 185, + 367, + 425, + 399 + ], + "score": 0.93, + "content": "\\sum _ { x } p ( x ) \\| \\mathcal { N } _ { 0 } ( \\delta _ { x } ) - \\mathcal { N } ( \\delta _ { x } ) \\| _ { Y } ^ { 2 } = \\sum _ { x y } \\frac { ( q ( x , y ) - p ( x , y ) ) ^ { 2 } } { p ( x ) p ( y ) } .", + "type": "interline_equation", + "image_path": "df8d7f7d7a094fb6a4ec97675e9fa6183150df7e30f9c93a72ddb88e92010910.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 367, + 425, + 383.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 185, + 383.0, + 425, + 399.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 440, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 441, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 252, + 424 + ], + "score": 1.0, + "content": "Proof. The low rank approximation", + "type": "text" + }, + { + "bbox": [ + 252, + 411, + 266, + 422 + ], + "score": 0.9, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 410, + 361, + 424 + ], + "score": 1.0, + "content": "minimizes the distance", + "type": "text" + }, + { + "bbox": [ + 362, + 411, + 412, + 423 + ], + "score": 0.93, + "content": "\\| \\mathcal { N } _ { 0 } - \\mathcal { N } \\| _ { \\mathrm { F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 410, + 441, + 424 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 259, + 425, + 352, + 439 + ], + "lines": [ + { + "bbox": [ + 259, + 425, + 352, + 439 + ], + "spans": [ + { + "bbox": [ + 259, + 425, + 352, + 439 + ], + "score": 0.93, + "content": "\\| \\mathcal { M } \\| _ { F } ^ { 2 } = \\mathrm { T r } \\left( \\mathcal { M } ^ { * } \\mathcal { M } \\right)", + "type": "interline_equation", + "image_path": "253fa3db9ec0fe002e085ab83d9ded0c2df946910a937c7ba475785f06910eb1.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 259, + 425, + 352, + 439 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "is the Hilbert-Schmidt (or Frobenius) norm (Eckart & Young, 1936). This follows from the fact", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 187, + 465 + ], + "score": 1.0, + "content": "that this is also the", + "type": "text" + }, + { + "bbox": [ + 188, + 452, + 196, + 462 + ], + "score": 0.85, + "content": "l ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 451, + 366, + 465 + ], + "score": 1.0, + "content": "-norm of the vector of singular values of", + "type": "text" + }, + { + "bbox": [ + 366, + 453, + 379, + 462 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 451, + 505, + 465 + ], + "score": 1.0, + "content": ". Let us find the explicit form", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 462, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 244, + 477 + ], + "score": 1.0, + "content": "of the trace. Each possible value", + "type": "text" + }, + { + "bbox": [ + 244, + 466, + 251, + 474 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 462, + 315, + 477 + ], + "score": 1.0, + "content": "of the variable", + "type": "text" + }, + { + "bbox": [ + 315, + 464, + 325, + 474 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 462, + 505, + 477 + ], + "score": 1.0, + "content": "is associated with a probability distribution", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 149, + 487 + ], + "score": 0.92, + "content": "\\delta _ { x } ( y ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 475, + 173, + 487 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 173, + 476, + 199, + 486 + ], + "score": 0.88, + "content": "x = y", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 475, + 470, + 487 + ], + "score": 1.0, + "content": "and zero otherwise. These distributions form an orthogonal basis of", + "type": "text" + }, + { + "bbox": [ + 470, + 475, + 484, + 486 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 475, + 505, + 487 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 485, + 285, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 155, + 498 + ], + "score": 1.0, + "content": "have norms", + "type": "text" + }, + { + "bbox": [ + 155, + 486, + 237, + 498 + ], + "score": 0.93, + "content": "\\langle \\delta _ { x } , \\delta _ { x } \\rangle = 1 / p _ { X } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 485, + 285, + 498 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 500, + 396, + 553 + ], + "lines": [ + { + "bbox": [ + 214, + 500, + 396, + 553 + ], + "spans": [ + { + "bbox": [ + 214, + 500, + 396, + 553 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle \\operatorname { T r } \\left( \\mathcal { M ^ { * } M } \\right) = \\sum _ { x } p _ { X } ( x ) \\langle \\delta _ { x } , \\mathcal { M ^ { * } M } ( \\delta _ { x } ) \\rangle _ { Y } } \\\\ { \\displaystyle \\quad = \\sum _ { x } p _ { X } ( x ) \\| \\mathcal { M } ( \\delta _ { x } ) \\| _ { Y } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "6617f39c0c7d039718aac0cf6cf3a2830a2819f9724871e62ef353ed978bdd14.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 214, + 500, + 396, + 517.6666666666666 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 214, + 517.6666666666666, + 396, + 535.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 214, + 535.3333333333333, + 396, + 552.9999999999999 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 579, + 231, + 590 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 231, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 231, + 590 + ], + "score": 1.0, + "content": "B.4 RELEVANT VARIABLES", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 104, + 599, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 211, + 613 + ], + "score": 1.0, + "content": "We express the elements", + "type": "text" + }, + { + "bbox": [ + 211, + 600, + 248, + 611 + ], + "score": 0.92, + "content": "\\mu \\in \\ V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 598, + 268, + 613 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 268, + 600, + 304, + 610 + ], + "score": 0.91, + "content": "\\nu \\in \\ V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 598, + 414, + 613 + ], + "score": 1.0, + "content": "in terms of the marginals", + "type": "text" + }, + { + "bbox": [ + 414, + 602, + 428, + 611 + ], + "score": 0.87, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 598, + 448, + 613 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 449, + 602, + 461, + 611 + ], + "score": 0.86, + "content": "p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 598, + 505, + 613 + ], + "score": 1.0, + "content": "as simple", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 610, + 147, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 147, + 623 + ], + "score": 1.0, + "content": "products:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 620, + 401, + 634 + ], + "lines": [ + { + "bbox": [ + 210, + 620, + 401, + 634 + ], + "spans": [ + { + "bbox": [ + 210, + 620, + 401, + 634 + ], + "score": 0.88, + "content": "\\mu ( x ) = p _ { X } ( x ) f ( x ) \\quad { \\mathrm { a n d } } \\quad \\nu ( y ) = p _ { Y } ( y ) g ( y )", + "type": "interline_equation", + "image_path": "f0603455ab679d7dc3d9eef5e9809c2db30e86572bece69d545acc50922839b5.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 210, + 620, + 401, + 634 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 330, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 331, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 133, + 649 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 637, + 150, + 646 + ], + "score": 0.86, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 633, + 180, + 649 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 180, + 635, + 187, + 646 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 633, + 205, + 649 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 205, + 637, + 212, + 646 + ], + "score": 0.82, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 633, + 295, + 649 + ], + "score": 1.0, + "content": "are real functions of", + "type": "text" + }, + { + "bbox": [ + 296, + 637, + 303, + 645 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 633, + 320, + 649 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 321, + 637, + 327, + 646 + ], + "score": 0.78, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 633, + 331, + 649 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 651, + 505, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 439, + 665 + ], + "score": 1.0, + "content": "The inner products then simply become correlations among variables. Using also", + "type": "text" + }, + { + "bbox": [ + 440, + 651, + 486, + 663 + ], + "score": 0.93, + "content": "\\mu ^ { \\prime } = p _ { X } f ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 650, + 506, + 665 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 661, + 195, + 675 + ], + "spans": [ + { + "bbox": [ + 107, + 662, + 149, + 674 + ], + "score": 0.92, + "content": "\\nu ^ { \\prime } = p _ { Y } g ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 661, + 195, + 675 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 676, + 387, + 731 + ], + "lines": [ + { + "bbox": [ + 223, + 676, + 387, + 731 + ], + "spans": [ + { + "bbox": [ + 223, + 676, + 387, + 731 + ], + "score": 0.9, + "content": "\\begin{array} { l } { { \\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } = \\displaystyle \\sum _ { x } p _ { X } ( x ) f ( x ) f ^ { \\prime } ( x ) = \\overline { { { f f ^ { \\prime } } } } , } } \\\\ { { \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } = \\displaystyle \\sum _ { y } p _ { Y } ( y ) g ( y ) g ^ { \\prime } ( y ) = \\overline { { { g g ^ { \\prime } } } } . } } \\end{array}", + "type": "interline_equation", + "image_path": "f7cda9f3540d0af30f97eeb8acb938d9f6f74bc7df4d8f25db6a203aa534caed.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 223, + 676, + 387, + 694.3333333333334 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 223, + 694.3333333333334, + 387, + 712.6666666666667 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 223, + 712.6666666666667, + 387, + 731.0000000000001 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 555, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 496, + 556, + 504, + 565 + ], + "spans": [ + { + "bbox": [ + 496, + 556, + 504, + 565 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 77, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 176, + 93 + ], + "score": 0.91, + "content": "2 5 6 ^ { 2 8 ^ { 2 } } \\simeq 1 0 ^ { 1 8 8 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 77, + 507, + 97 + ], + "score": 1.0, + "content": ". However, in many case, only very few of these dimensions may be relevant for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 267, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 267, + 106 + ], + "score": 1.0, + "content": "the purpose of inferring other variables.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 77, + 507, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 144 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 504, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 281, + 123 + ], + "score": 1.0, + "content": "The core of our approach is to approximate", + "type": "text" + }, + { + "bbox": [ + 281, + 111, + 292, + 120 + ], + "score": 0.8, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 110, + 310, + 123 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 310, + 110, + 325, + 120 + ], + "score": 0.87, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 110, + 493, + 123 + ], + "score": 1.0, + "content": "by restricting them to the span of the first", + "type": "text" + }, + { + "bbox": [ + 493, + 111, + 504, + 122 + ], + "score": 0.87, + "content": "k _ { 0 }", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 119, + 507, + 136 + ], + "spans": [ + { + "bbox": [ + 104, + 119, + 159, + 136 + ], + "score": 1.0, + "content": "eigenvectors", + "type": "text" + }, + { + "bbox": [ + 159, + 123, + 171, + 133 + ], + "score": 0.85, + "content": "u _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 119, + 189, + 136 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 190, + 123, + 200, + 134 + ], + "score": 0.86, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 119, + 315, + 136 + ], + "score": 1.0, + "content": "with largest singular values", + "type": "text" + }, + { + "bbox": [ + 315, + 123, + 325, + 134 + ], + "score": 0.86, + "content": "\\eta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 119, + 491, + 136 + ], + "score": 1.0, + "content": ". That is, if we order the singular values", + "type": "text" + }, + { + "bbox": [ + 491, + 124, + 501, + 134 + ], + "score": 0.84, + "content": "\\eta _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 119, + 507, + 136 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 132, + 394, + 145 + ], + "spans": [ + { + "bbox": [ + 107, + 133, + 158, + 144 + ], + "score": 0.92, + "content": "j = 1 , \\dots , k", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 132, + 394, + 145 + ], + "score": 1.0, + "content": "in decreasing order, we propose to use the approximations", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 110, + 507, + 145 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 146, + 364, + 204 + ], + "lines": [ + { + "bbox": [ + 247, + 146, + 364, + 204 + ], + "spans": [ + { + "bbox": [ + 247, + 146, + 364, + 204 + ], + "score": 0.93, + "content": "\\begin{array} { l } { { \\displaystyle \\mathcal { N } _ { 0 } ( \\mu ) = \\sum _ { j \\leq k _ { 0 } } \\eta _ { j } \\langle u _ { j } , \\mu \\rangle _ { X } v _ { j } } } \\\\ { { \\displaystyle \\mathcal { N } _ { 0 } ^ { * } ( \\nu ) = \\sum _ { j \\leq k _ { 0 } } \\eta _ { j } \\langle v _ { j } , \\nu \\rangle _ { Y } u _ { j } } } \\end{array}", + "type": "interline_equation", + "image_path": "f3a5327242bb5b1d075b4f3e55813df2cb07b711bc8ac260eb13179d35ab16b0.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 146, + 364, + 175.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 247, + 175.0, + 364, + 204.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 219, + 493, + 232 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 494, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 117, + 234 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 220, + 128, + 230 + ], + "score": 0.82, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 217, + 146, + 234 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 146, + 220, + 161, + 230 + ], + "score": 0.87, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 217, + 252, + 234 + ], + "score": 1.0, + "content": "respectively, for some", + "type": "text" + }, + { + "bbox": [ + 252, + 220, + 263, + 231 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 217, + 378, + 234 + ], + "score": 1.0, + "content": "typically much smaller than", + "type": "text" + }, + { + "bbox": [ + 378, + 220, + 384, + 230 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 217, + 421, + 234 + ], + "score": 1.0, + "content": ", and any", + "type": "text" + }, + { + "bbox": [ + 422, + 220, + 454, + 231 + ], + "score": 0.86, + "content": "\\mu \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 217, + 458, + 234 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 459, + 220, + 490, + 231 + ], + "score": 0.81, + "content": "\\nu \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 217, + 494, + 234 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 217, + 494, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 385, + 249 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 385, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 227, + 251 + ], + "score": 1.0, + "content": "We denote the components of", + "type": "text" + }, + { + "bbox": [ + 227, + 236, + 241, + 248 + ], + "score": 0.89, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 234, + 259, + 251 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 259, + 236, + 274, + 249 + ], + "score": 0.9, + "content": "\\mathcal { N } _ { 0 } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 234, + 288, + 251 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 289, + 236, + 316, + 249 + ], + "score": 0.92, + "content": "q ( y | x )", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 234, + 334, + 251 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 334, + 236, + 361, + 249 + ], + "score": 0.91, + "content": "q ( x | y )", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 234, + 385, + 251 + ], + "score": 1.0, + "content": ", e.g.,", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 234, + 385, + 251 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 245, + 251, + 366, + 277 + ], + "lines": [ + { + "bbox": [ + 245, + 251, + 366, + 277 + ], + "spans": [ + { + "bbox": [ + 245, + 251, + 366, + 277 + ], + "score": 0.94, + "content": "{ \\mathcal { N } } _ { 0 } ( \\mu ) ( y ) = \\sum _ { x } q ( y | x ) \\mu ( x ) .", + "type": "interline_equation", + "image_path": "a6e2b9880b1b37ead6fd6a4ffe84590a7907071edbd961c9b9d70841b48801b8.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 245, + 251, + 366, + 277 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 279, + 504, + 312 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 132, + 293 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 132, + 280, + 146, + 290 + ], + "score": 0.88, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 278, + 165, + 293 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 166, + 279, + 181, + 291 + ], + "score": 0.9, + "content": "\\mathcal { N } _ { 0 } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 278, + 292, + 293 + ], + "score": 1.0, + "content": "are adjoint, we can define", + "type": "text" + }, + { + "bbox": [ + 292, + 279, + 458, + 291 + ], + "score": 0.93, + "content": "q ( x , y ) = q ( x | y ) p _ { Y } ( y ) = q ( y | x ) p _ { X } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 278, + 505, + 293 + ], + "score": 1.0, + "content": ". Although", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 289, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 176, + 303 + ], + "score": 1.0, + "content": "the marginals of", + "type": "text" + }, + { + "bbox": [ + 176, + 291, + 205, + 302 + ], + "score": 0.93, + "content": "\\dot { \\boldsymbol { q } } ( \\boldsymbol { x } , \\boldsymbol { y } )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 289, + 338, + 303 + ], + "score": 1.0, + "content": "are the probability distributions", + "type": "text" + }, + { + "bbox": [ + 338, + 293, + 352, + 302 + ], + "score": 0.87, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 289, + 372, + 303 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 372, + 293, + 385, + 302 + ], + "score": 0.86, + "content": "p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 289, + 443, + 303 + ], + "score": 1.0, + "content": ", the numbers", + "type": "text" + }, + { + "bbox": [ + 443, + 291, + 472, + 302 + ], + "score": 0.92, + "content": "q ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 289, + 506, + 303 + ], + "score": 1.0, + "content": "are not", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 302, + 189, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 189, + 313 + ], + "score": 1.0, + "content": "necessarily positive.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 278, + 506, + 313 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 318, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 290, + 331 + ], + "score": 1.0, + "content": "The quality of this approximation for a given", + "type": "text" + }, + { + "bbox": [ + 290, + 319, + 302, + 329 + ], + "score": 0.87, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "does not directly depend on the dimensionality of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 329, + 504, + 342 + ], + "spans": [ + { + "bbox": [ + 107, + 330, + 117, + 339 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 329, + 135, + 342 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 135, + 330, + 144, + 339 + ], + "score": 0.75, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 329, + 493, + 342 + ], + "score": 1.0, + "content": ", but only on the amount of correlations between the two variables. Our aim is to use a", + "type": "text" + }, + { + "bbox": [ + 493, + 330, + 504, + 340 + ], + "score": 0.87, + "content": "k _ { 0 }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 340, + 417, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 256, + 354 + ], + "score": 1.0, + "content": "small enough that the components of", + "type": "text" + }, + { + "bbox": [ + 256, + 340, + 270, + 352 + ], + "score": 0.9, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 340, + 288, + 354 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 288, + 340, + 303, + 352 + ], + "score": 0.91, + "content": "\\mathcal { N } _ { 0 } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 340, + 417, + 354 + ], + "score": 1.0, + "content": "can be computed explicitly.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 318, + 506, + 354 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 414, + 365 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 415, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 159, + 367 + ], + "score": 1.0, + "content": "Theorem 1.", + "type": "text" + }, + { + "bbox": [ + 160, + 354, + 174, + 365 + ], + "score": 0.86, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 352, + 249, + 367 + ], + "score": 1.0, + "content": "is the map of rank", + "type": "text" + }, + { + "bbox": [ + 250, + 354, + 260, + 365 + ], + "score": 0.87, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 352, + 415, + 367 + ], + "score": 1.0, + "content": "which minimizes the average distance", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 352, + 415, + 367 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 367, + 425, + 399 + ], + "lines": [ + { + "bbox": [ + 185, + 367, + 425, + 399 + ], + "spans": [ + { + "bbox": [ + 185, + 367, + 425, + 399 + ], + "score": 0.93, + "content": "\\sum _ { x } p ( x ) \\| \\mathcal { N } _ { 0 } ( \\delta _ { x } ) - \\mathcal { N } ( \\delta _ { x } ) \\| _ { Y } ^ { 2 } = \\sum _ { x y } \\frac { ( q ( x , y ) - p ( x , y ) ) ^ { 2 } } { p ( x ) p ( y ) } .", + "type": "interline_equation", + "image_path": "df8d7f7d7a094fb6a4ec97675e9fa6183150df7e30f9c93a72ddb88e92010910.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 185, + 367, + 425, + 383.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 185, + 383.0, + 425, + 399.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 440, + 423 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 441, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 252, + 424 + ], + "score": 1.0, + "content": "Proof. The low rank approximation", + "type": "text" + }, + { + "bbox": [ + 252, + 411, + 266, + 422 + ], + "score": 0.9, + "content": "\\mathcal { N } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 410, + 361, + 424 + ], + "score": 1.0, + "content": "minimizes the distance", + "type": "text" + }, + { + "bbox": [ + 362, + 411, + 412, + 423 + ], + "score": 0.93, + "content": "\\| \\mathcal { N } _ { 0 } - \\mathcal { N } \\| _ { \\mathrm { F } }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 410, + 441, + 424 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 410, + 441, + 424 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 259, + 425, + 352, + 439 + ], + "lines": [ + { + "bbox": [ + 259, + 425, + 352, + 439 + ], + "spans": [ + { + "bbox": [ + 259, + 425, + 352, + 439 + ], + "score": 0.93, + "content": "\\| \\mathcal { M } \\| _ { F } ^ { 2 } = \\mathrm { T r } \\left( \\mathcal { M } ^ { * } \\mathcal { M } \\right)", + "type": "interline_equation", + "image_path": "253fa3db9ec0fe002e085ab83d9ded0c2df946910a937c7ba475785f06910eb1.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 259, + 425, + 352, + 439 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "is the Hilbert-Schmidt (or Frobenius) norm (Eckart & Young, 1936). This follows from the fact", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 187, + 465 + ], + "score": 1.0, + "content": "that this is also the", + "type": "text" + }, + { + "bbox": [ + 188, + 452, + 196, + 462 + ], + "score": 0.85, + "content": "l ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 451, + 366, + 465 + ], + "score": 1.0, + "content": "-norm of the vector of singular values of", + "type": "text" + }, + { + "bbox": [ + 366, + 453, + 379, + 462 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 451, + 505, + 465 + ], + "score": 1.0, + "content": ". Let us find the explicit form", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 462, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 244, + 477 + ], + "score": 1.0, + "content": "of the trace. Each possible value", + "type": "text" + }, + { + "bbox": [ + 244, + 466, + 251, + 474 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 462, + 315, + 477 + ], + "score": 1.0, + "content": "of the variable", + "type": "text" + }, + { + "bbox": [ + 315, + 464, + 325, + 474 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 462, + 505, + 477 + ], + "score": 1.0, + "content": "is associated with a probability distribution", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 149, + 487 + ], + "score": 0.92, + "content": "\\delta _ { x } ( y ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 475, + 173, + 487 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 173, + 476, + 199, + 486 + ], + "score": 0.88, + "content": "x = y", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 475, + 470, + 487 + ], + "score": 1.0, + "content": "and zero otherwise. These distributions form an orthogonal basis of", + "type": "text" + }, + { + "bbox": [ + 470, + 475, + 484, + 486 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 475, + 505, + 487 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 485, + 285, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 155, + 498 + ], + "score": 1.0, + "content": "have norms", + "type": "text" + }, + { + "bbox": [ + 155, + 486, + 237, + 498 + ], + "score": 0.93, + "content": "\\langle \\delta _ { x } , \\delta _ { x } \\rangle = 1 / p _ { X } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 485, + 285, + 498 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 442, + 505, + 498 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 500, + 396, + 553 + ], + "lines": [ + { + "bbox": [ + 214, + 500, + 396, + 553 + ], + "spans": [ + { + "bbox": [ + 214, + 500, + 396, + 553 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle \\operatorname { T r } \\left( \\mathcal { M ^ { * } M } \\right) = \\sum _ { x } p _ { X } ( x ) \\langle \\delta _ { x } , \\mathcal { M ^ { * } M } ( \\delta _ { x } ) \\rangle _ { Y } } \\\\ { \\displaystyle \\quad = \\sum _ { x } p _ { X } ( x ) \\| \\mathcal { M } ( \\delta _ { x } ) \\| _ { Y } ^ { 2 } } \\end{array}", + "type": "interline_equation", + "image_path": "6617f39c0c7d039718aac0cf6cf3a2830a2819f9724871e62ef353ed978bdd14.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 214, + 500, + 396, + 517.6666666666666 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 214, + 517.6666666666666, + 396, + 535.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 214, + 535.3333333333333, + 396, + 552.9999999999999 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 579, + 231, + 590 + ], + "lines": [ + { + "bbox": [ + 106, + 579, + 231, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 231, + 590 + ], + "score": 1.0, + "content": "B.4 RELEVANT VARIABLES", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 104, + 599, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 598, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 211, + 613 + ], + "score": 1.0, + "content": "We express the elements", + "type": "text" + }, + { + "bbox": [ + 211, + 600, + 248, + 611 + ], + "score": 0.92, + "content": "\\mu \\in \\ V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 598, + 268, + 613 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 268, + 600, + 304, + 610 + ], + "score": 0.91, + "content": "\\nu \\in \\ V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 598, + 414, + 613 + ], + "score": 1.0, + "content": "in terms of the marginals", + "type": "text" + }, + { + "bbox": [ + 414, + 602, + 428, + 611 + ], + "score": 0.87, + "content": "p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 598, + 448, + 613 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 449, + 602, + 461, + 611 + ], + "score": 0.86, + "content": "p _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 598, + 505, + 613 + ], + "score": 1.0, + "content": "as simple", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 610, + 147, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 147, + 623 + ], + "score": 1.0, + "content": "products:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 598, + 505, + 623 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 620, + 401, + 634 + ], + "lines": [ + { + "bbox": [ + 210, + 620, + 401, + 634 + ], + "spans": [ + { + "bbox": [ + 210, + 620, + 401, + 634 + ], + "score": 0.88, + "content": "\\mu ( x ) = p _ { X } ( x ) f ( x ) \\quad { \\mathrm { a n d } } \\quad \\nu ( y ) = p _ { Y } ( y ) g ( y )", + "type": "interline_equation", + "image_path": "f0603455ab679d7dc3d9eef5e9809c2db30e86572bece69d545acc50922839b5.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 210, + 620, + 401, + 634 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 330, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 331, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 133, + 649 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 637, + 150, + 646 + ], + "score": 0.86, + "content": "x , y", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 633, + 180, + 649 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 180, + 635, + 187, + 646 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 633, + 205, + 649 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 205, + 637, + 212, + 646 + ], + "score": 0.82, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 633, + 295, + 649 + ], + "score": 1.0, + "content": "are real functions of", + "type": "text" + }, + { + "bbox": [ + 296, + 637, + 303, + 645 + ], + "score": 0.77, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 633, + 320, + 649 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 321, + 637, + 327, + 646 + ], + "score": 0.78, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 633, + 331, + 649 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 633, + 331, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 651, + 505, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 439, + 665 + ], + "score": 1.0, + "content": "The inner products then simply become correlations among variables. Using also", + "type": "text" + }, + { + "bbox": [ + 440, + 651, + 486, + 663 + ], + "score": 0.93, + "content": "\\mu ^ { \\prime } = p _ { X } f ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 650, + 506, + 665 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 661, + 195, + 675 + ], + "spans": [ + { + "bbox": [ + 107, + 662, + 149, + 674 + ], + "score": 0.92, + "content": "\\nu ^ { \\prime } = p _ { Y } g ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 661, + 195, + 675 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 650, + 506, + 675 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 676, + 387, + 731 + ], + "lines": [ + { + "bbox": [ + 223, + 676, + 387, + 731 + ], + "spans": [ + { + "bbox": [ + 223, + 676, + 387, + 731 + ], + "score": 0.9, + "content": "\\begin{array} { l } { { \\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } = \\displaystyle \\sum _ { x } p _ { X } ( x ) f ( x ) f ^ { \\prime } ( x ) = \\overline { { { f f ^ { \\prime } } } } , } } \\\\ { { \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } = \\displaystyle \\sum _ { y } p _ { Y } ( y ) g ( y ) g ^ { \\prime } ( y ) = \\overline { { { g g ^ { \\prime } } } } . } } \\end{array}", + "type": "interline_equation", + "image_path": "f7cda9f3540d0af30f97eeb8acb938d9f6f74bc7df4d8f25db6a203aa534caed.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 223, + 676, + 387, + 694.3333333333334 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 223, + 694.3333333333334, + 387, + 712.6666666666667 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 223, + 712.6666666666667, + 387, + 731.0000000000001 + ], + "spans": [], + "index": 38 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 313, + 96 + ], + "score": 1.0, + "content": "These are simple expectation values with respect to", + "type": "text" + }, + { + "bbox": [ + 314, + 85, + 320, + 94 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ", which we assumed is the type of quantity we", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 248, + 106 + ], + "score": 1.0, + "content": "can evaluate for arbitrary functions", + "type": "text" + }, + { + "bbox": [ + 249, + 93, + 290, + 106 + ], + "score": 0.93, + "content": "f , f ^ { \\prime } , g , \\bar { g ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 93, + 294, + 106 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 110, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 132, + 123 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 132, + 110, + 158, + 121 + ], + "score": 0.85, + "content": "\\mathcal { N } ^ { * } \\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 110, + 411, + 123 + ], + "score": 1.0, + "content": "is self-adjoint in terms of this inner product, its eigenvectors", + "type": "text" + }, + { + "bbox": [ + 411, + 113, + 421, + 122 + ], + "score": 0.87, + "content": "u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "are orthogonal, and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 488, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 244, + 133 + ], + "score": 1.0, + "content": "hence the corresponding variables", + "type": "text" + }, + { + "bbox": [ + 245, + 123, + 254, + 132 + ], + "score": 0.85, + "content": "a _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 122, + 300, + 133 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + }, + { + "bbox": [ + 300, + 121, + 385, + 134 + ], + "score": 0.93, + "content": "{ u } _ { i } \\bar { ( x ) } = p _ { X } ( x ) \\bar { a _ { i } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 122, + 488, + 133 + ], + "score": 1.0, + "content": "are uncorrelated. Indeed,", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 138, + 352, + 153 + ], + "lines": [ + { + "bbox": [ + 258, + 138, + 352, + 153 + ], + "spans": [ + { + "bbox": [ + 258, + 138, + 352, + 153 + ], + "score": 0.92, + "content": "\\overline { { a _ { i } a _ { j } } } = \\langle u _ { i } , u _ { j } \\rangle _ { X } = 0 ,", + "type": "interline_equation", + "image_path": "e9e69f6a4cec03a846afcb9709a37a4da5c1127809ce22499507efcdd41c32bb.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 258, + 138, + 352, + 153 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 158, + 506, + 180 + ], + "lines": [ + { + "bbox": [ + 105, + 156, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 132, + 171 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 132, + 159, + 146, + 169 + ], + "score": 0.82, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 156, + 315, + 171 + ], + "score": 1.0, + "content": ". Moreover, accounting for the eigenvector", + "type": "text" + }, + { + "bbox": [ + 315, + 160, + 352, + 169 + ], + "score": 0.91, + "content": "u _ { 0 } = p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 156, + 506, + 171 + ], + "score": 1.0, + "content": "(corresponding to the constant feature", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 168, + 192, + 181 + ], + "spans": [ + { + "bbox": [ + 107, + 169, + 149, + 181 + ], + "score": 0.92, + "content": "a _ { 0 } ( x ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 168, + 177, + 181 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 177, + 170, + 184, + 178 + ], + "score": 0.73, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 168, + 192, + 181 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 290, + 180, + 321, + 191 + ], + "lines": [ + { + "bbox": [ + 290, + 180, + 321, + 191 + ], + "spans": [ + { + "bbox": [ + 290, + 180, + 321, + 191 + ], + "score": 0.89, + "content": "\\overline { { a } } _ { i } = 0", + "type": "interline_equation", + "image_path": "16dffdd84bef144f0869194d1bd7081036059f9c15ef42b3ce9bcebd8ccfc950.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 290, + 180, + 321, + 191 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 195, + 257, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 258, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 133, + 208 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 195, + 156, + 207 + ], + "score": 0.9, + "content": "i \\neq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 194, + 258, + 208 + ], + "score": 1.0, + "content": ". Hence we trivially have", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 279, + 207, + 331, + 219 + ], + "lines": [ + { + "bbox": [ + 279, + 207, + 331, + 219 + ], + "spans": [ + { + "bbox": [ + 279, + 207, + 331, + 219 + ], + "score": 0.91, + "content": "\\overline { { a _ { i } a _ { j } } } = \\overline { { a } } _ { i } \\overline { { a } } _ { j }", + "type": "interline_equation", + "image_path": "1314457dd98ce6123c16959fd0d98f53893544881218a7696aa40d6a71ec0f7f.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 279, + 207, + 331, + 219 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 221, + 168, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 169, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 133, + 234 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 222, + 165, + 234 + ], + "score": 0.9, + "content": "i , j \\neq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 221, + 169, + 234 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 361, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 361, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 237, + 252 + ], + "score": 1.0, + "content": "Likewise for the eigenvectors of", + "type": "text" + }, + { + "bbox": [ + 237, + 239, + 262, + 249 + ], + "score": 0.88, + "content": "\\mathcal { N N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 236, + 276, + 252 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 276, + 238, + 357, + 251 + ], + "score": 0.9, + "content": "v _ { i } ( y ) = p _ { Y } ( y ) b _ { i } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 236, + 361, + 252 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 256, + 365, + 271 + ], + "lines": [ + { + "bbox": [ + 246, + 256, + 365, + 271 + ], + "spans": [ + { + "bbox": [ + 246, + 256, + 365, + 271 + ], + "score": 0.91, + "content": "\\overline { { b _ { i } b _ { j } } } = \\langle v _ { i } , v _ { j } \\rangle _ { Y } = 0 = \\overline { { b } } _ { i } \\overline { { b } } _ { j } .", + "type": "interline_equation", + "image_path": "64b88123c6d74a35b1c5dc9c829e55ebd9514f524ae9e97af52afe2fde456f6f.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 246, + 256, + 365, + 271 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 168, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 169, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 133, + 289 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 277, + 165, + 288 + ], + "score": 0.9, + "content": "i , j \\neq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 276, + 169, + 289 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 293, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 292, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 305, + 307 + ], + "score": 1.0, + "content": "Importantly, this does not mean that the variables", + "type": "text" + }, + { + "bbox": [ + 305, + 295, + 347, + 305 + ], + "score": 0.86, + "content": "u _ { 1 } , u _ { 2 } , \\ldots", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 292, + 366, + 307 + ], + "score": 1.0, + "content": "nor", + "type": "text" + }, + { + "bbox": [ + 366, + 296, + 406, + 305 + ], + "score": 0.87, + "content": "v _ { 1 } , v _ { 2 } , \\ldots", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 292, + 506, + 307 + ], + "score": 1.0, + "content": "are “disentangled”, i.e.,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 305, + 504, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 504, + 317 + ], + "score": 1.0, + "content": "they are not statistically independent. These variables represent components in the space of proba-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 314, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 329 + ], + "score": 1.0, + "content": "bility vectors, rather than the “sample” space. They should be understood as spanning a subspace", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 339 + ], + "score": 1.0, + "content": "of the space of functions over the relevant independent variables. We discuss this in more detail in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 337, + 156, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 156, + 348 + ], + "score": 1.0, + "content": "Section 5.3.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 362, + 292, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 190, + 376 + ], + "score": 1.0, + "content": "B.5 CORNERS OF", + "type": "text" + }, + { + "bbox": [ + 190, + 363, + 201, + 372 + ], + "score": 0.69, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 360, + 295, + 376 + ], + "score": 1.0, + "content": "AND LOSS FUNCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 502, + 406 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 459, + 396 + ], + "score": 1.0, + "content": "The final piece of puzzle we need, is the ability to express the components (corners) of", + "type": "text" + }, + { + "bbox": [ + 459, + 383, + 470, + 393 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 382, + 488, + 396 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 488, + 383, + 504, + 393 + ], + "score": 0.84, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 348, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 348, + 407 + ], + "score": 1.0, + "content": "in the span of possible non-orthogonal families of variables.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 411, + 502, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 504, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 307, + 425 + ], + "score": 1.0, + "content": "Let us therefore consider two arbitrary families", + "type": "text" + }, + { + "bbox": [ + 307, + 411, + 353, + 423 + ], + "score": 0.89, + "content": "f _ { 1 } , \\ldots , f _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 409, + 374, + 425 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 375, + 413, + 420, + 423 + ], + "score": 0.87, + "content": "g _ { 1 } , \\ldots , g _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 409, + 504, + 425 + ], + "score": 1.0, + "content": "of variables, which", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 421, + 359, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 241, + 435 + ], + "score": 1.0, + "content": "respectively represent the vectors", + "type": "text" + }, + { + "bbox": [ + 241, + 422, + 290, + 434 + ], + "score": 0.93, + "content": "p _ { X } f _ { j } \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 421, + 308, + 435 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 308, + 423, + 355, + 434 + ], + "score": 0.9, + "content": "p _ { Y } g _ { j } \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 421, + 359, + 435 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 105, + 438, + 504, + 461 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 426, + 451 + ], + "score": 1.0, + "content": "Firstly, we need matrices representing the components of the inner products on", + "type": "text" + }, + { + "bbox": [ + 426, + 439, + 440, + 450 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 438, + 459, + 451 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 459, + 439, + 473, + 450 + ], + "score": 0.88, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 438, + 505, + 451 + ], + "score": 1.0, + "content": ". Those", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 450, + 217, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 217, + 461 + ], + "score": 1.0, + "content": "are the symmetric matrices", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 466, + 368, + 498 + ], + "lines": [ + { + "bbox": [ + 241, + 466, + 368, + 498 + ], + "spans": [ + { + "bbox": [ + 241, + 466, + 368, + 498 + ], + "score": 0.91, + "content": "\\begin{array} { r } { K _ { i j } = \\langle p _ { X } f _ { i } , p _ { X } f _ { j } \\rangle _ { X } = \\overline { { f _ { i } f _ { j } } } , } \\\\ { L _ { i j } = \\langle p _ { Y } g _ { i } , p _ { Y } g _ { j } \\rangle _ { Y } = \\overline { { g _ { i } g _ { j } } } . } \\end{array}", + "type": "interline_equation", + "image_path": "e3b0b84063132afb696e79be3654f58285ebf5ac5071aecd466e0c04ac8ad0dd.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 466, + 368, + 482.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 241, + 482.0, + 368, + 498.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 507, + 274, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 275, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 175, + 523 + ], + "score": 1.0, + "content": "The components", + "type": "text" + }, + { + "bbox": [ + 175, + 508, + 191, + 520 + ], + "score": 0.91, + "content": "N _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 504, + 203, + 523 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 204, + 508, + 214, + 517 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 504, + 275, + 523 + ], + "score": 1.0, + "content": "are defined by", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 524, + 360, + 551 + ], + "lines": [ + { + "bbox": [ + 250, + 524, + 360, + 551 + ], + "spans": [ + { + "bbox": [ + 250, + 524, + 360, + 551 + ], + "score": 0.94, + "content": "\\mathcal { N } ( p _ { X } f _ { j } ) = \\sum _ { i } N _ { i j } p _ { Y } g _ { i } .", + "type": "interline_equation", + "image_path": "2b7b8a822b8382b09ad5f7ef0f60b06484b822d4bcf9a16b5df140c5972bf464.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 250, + 524, + 360, + 551 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 556, + 294, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 227, + 570 + ], + "score": 1.0, + "content": "Taking the inner product with", + "type": "text" + }, + { + "bbox": [ + 227, + 559, + 250, + 568 + ], + "score": 0.87, + "content": "p _ { Y } g _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 555, + 295, + 570 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 573, + 374, + 600 + ], + "lines": [ + { + "bbox": [ + 237, + 573, + 374, + 600 + ], + "spans": [ + { + "bbox": [ + 237, + 573, + 374, + 600 + ], + "score": 0.94, + "content": "\\langle p _ { Y } g _ { k } , \\mathcal { N } ( p _ { X } f _ { j } ) \\rangle = \\sum _ { i } N _ { i j } L _ { k i } .", + "type": "interline_equation", + "image_path": "b2186206361d77232b516411f4d298cafe0f44cb95c38a9d96ca5dc21515641c.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 237, + 573, + 374, + 600 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 372, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 373, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 373, + 618 + ], + "score": 1.0, + "content": "The left-hand side can be computed using Equ. 13. It is the matrix", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 621, + 400, + 698 + ], + "lines": [ + { + "bbox": [ + 210, + 621, + 400, + 698 + ], + "spans": [ + { + "bbox": [ + 210, + 621, + 400, + 698 + ], + "score": 0.94, + "content": "\\begin{array} { l } { { \\displaystyle { \\cal A } _ { k j } = \\langle p _ { Y } g _ { k } , { \\cal N } ( p _ { X } f _ { j } ) \\rangle } } \\\\ { { \\displaystyle ~ = \\sum _ { x , y } \\frac { p _ { Y } ( y ) g _ { k } ( y ) p _ { Y | X } ( y | x ) p _ { X } ( x ) f _ { j } ( x ) } { p _ { Y } ( y ) } } } \\\\ { { \\displaystyle ~ = \\sum _ { x , y } p ( x , y ) g _ { k } ( y ) f _ { j } ( x ) = \\overline { { g _ { k } f _ { j } } } . } } \\end{array}", + "type": "interline_equation", + "image_path": "9deaf3f2b4a503f27c9b2f2fc081d2425e38207c6b07dc4513908e2dfe0753c2.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 210, + 621, + 400, + 636.4 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 210, + 636.4, + 400, + 651.8 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 210, + 651.8, + 400, + 667.1999999999999 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 210, + 667.1999999999999, + 400, + 682.5999999999999 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 210, + 682.5999999999999, + 400, + 697.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 702, + 329, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 702, + 329, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 276, + 715 + ], + "score": 1.0, + "content": "Therefore, in matrix notation, Equ. (34) is", + "type": "text" + }, + { + "bbox": [ + 277, + 703, + 314, + 712 + ], + "score": 0.9, + "content": "A = L N", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 702, + 329, + 715 + ], + "score": 1.0, + "content": ", or", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 279, + 719, + 331, + 731 + ], + "lines": [ + { + "bbox": [ + 279, + 719, + 331, + 731 + ], + "spans": [ + { + "bbox": [ + 279, + 719, + 331, + 731 + ], + "score": 0.9, + "content": "N = L ^ { - 1 } A .", + "type": "interline_equation", + "image_path": "ae48dea95919ca0dca6569335f749fd8f58bd0eaa7faf859df0a382270205e86.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 279, + 719, + 331, + 731 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 313, + 96 + ], + "score": 1.0, + "content": "These are simple expectation values with respect to", + "type": "text" + }, + { + "bbox": [ + 314, + 85, + 320, + 94 + ], + "score": 0.78, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ", which we assumed is the type of quantity we", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 248, + 106 + ], + "score": 1.0, + "content": "can evaluate for arbitrary functions", + "type": "text" + }, + { + "bbox": [ + 249, + 93, + 290, + 106 + ], + "score": 0.93, + "content": "f , f ^ { \\prime } , g , \\bar { g ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 93, + 294, + 106 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 110, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 132, + 123 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 132, + 110, + 158, + 121 + ], + "score": 0.85, + "content": "\\mathcal { N } ^ { * } \\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 110, + 411, + 123 + ], + "score": 1.0, + "content": "is self-adjoint in terms of this inner product, its eigenvectors", + "type": "text" + }, + { + "bbox": [ + 411, + 113, + 421, + 122 + ], + "score": 0.87, + "content": "u _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "are orthogonal, and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 488, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 244, + 133 + ], + "score": 1.0, + "content": "hence the corresponding variables", + "type": "text" + }, + { + "bbox": [ + 245, + 123, + 254, + 132 + ], + "score": 0.85, + "content": "a _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 122, + 300, + 133 + ], + "score": 1.0, + "content": "defined by", + "type": "text" + }, + { + "bbox": [ + 300, + 121, + 385, + 134 + ], + "score": 0.93, + "content": "{ u } _ { i } \\bar { ( x ) } = p _ { X } ( x ) \\bar { a _ { i } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 122, + 488, + 133 + ], + "score": 1.0, + "content": "are uncorrelated. Indeed,", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 110, + 505, + 134 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 258, + 138, + 352, + 153 + ], + "lines": [ + { + "bbox": [ + 258, + 138, + 352, + 153 + ], + "spans": [ + { + "bbox": [ + 258, + 138, + 352, + 153 + ], + "score": 0.92, + "content": "\\overline { { a _ { i } a _ { j } } } = \\langle u _ { i } , u _ { j } \\rangle _ { X } = 0 ,", + "type": "interline_equation", + "image_path": "e9e69f6a4cec03a846afcb9709a37a4da5c1127809ce22499507efcdd41c32bb.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 258, + 138, + 352, + 153 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 158, + 506, + 180 + ], + "lines": [ + { + "bbox": [ + 105, + 156, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 132, + 171 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 132, + 159, + 146, + 169 + ], + "score": 0.82, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 156, + 315, + 171 + ], + "score": 1.0, + "content": ". Moreover, accounting for the eigenvector", + "type": "text" + }, + { + "bbox": [ + 315, + 160, + 352, + 169 + ], + "score": 0.91, + "content": "u _ { 0 } = p _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 156, + 506, + 171 + ], + "score": 1.0, + "content": "(corresponding to the constant feature", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 168, + 192, + 181 + ], + "spans": [ + { + "bbox": [ + 107, + 169, + 149, + 181 + ], + "score": 0.92, + "content": "a _ { 0 } ( x ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 168, + 177, + 181 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 177, + 170, + 184, + 178 + ], + "score": 0.73, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 168, + 192, + 181 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 156, + 506, + 181 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 290, + 180, + 321, + 191 + ], + "lines": [ + { + "bbox": [ + 290, + 180, + 321, + 191 + ], + "spans": [ + { + "bbox": [ + 290, + 180, + 321, + 191 + ], + "score": 0.89, + "content": "\\overline { { a } } _ { i } = 0", + "type": "interline_equation", + "image_path": "16dffdd84bef144f0869194d1bd7081036059f9c15ef42b3ce9bcebd8ccfc950.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 290, + 180, + 321, + 191 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 195, + 257, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 258, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 133, + 208 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 195, + 156, + 207 + ], + "score": 0.9, + "content": "i \\neq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 194, + 258, + 208 + ], + "score": 1.0, + "content": ". Hence we trivially have", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 194, + 258, + 208 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 279, + 207, + 331, + 219 + ], + "lines": [ + { + "bbox": [ + 279, + 207, + 331, + 219 + ], + "spans": [ + { + "bbox": [ + 279, + 207, + 331, + 219 + ], + "score": 0.91, + "content": "\\overline { { a _ { i } a _ { j } } } = \\overline { { a } } _ { i } \\overline { { a } } _ { j }", + "type": "interline_equation", + "image_path": "1314457dd98ce6123c16959fd0d98f53893544881218a7696aa40d6a71ec0f7f.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 279, + 207, + 331, + 219 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 221, + 168, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 169, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 133, + 234 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 222, + 165, + 234 + ], + "score": 0.9, + "content": "i , j \\neq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 221, + 169, + 234 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 221, + 169, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 361, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 361, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 237, + 252 + ], + "score": 1.0, + "content": "Likewise for the eigenvectors of", + "type": "text" + }, + { + "bbox": [ + 237, + 239, + 262, + 249 + ], + "score": 0.88, + "content": "\\mathcal { N N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 236, + 276, + 252 + ], + "score": 1.0, + "content": ". If", + "type": "text" + }, + { + "bbox": [ + 276, + 238, + 357, + 251 + ], + "score": 0.9, + "content": "v _ { i } ( y ) = p _ { Y } ( y ) b _ { i } ( y )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 236, + 361, + 252 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 236, + 361, + 252 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 256, + 365, + 271 + ], + "lines": [ + { + "bbox": [ + 246, + 256, + 365, + 271 + ], + "spans": [ + { + "bbox": [ + 246, + 256, + 365, + 271 + ], + "score": 0.91, + "content": "\\overline { { b _ { i } b _ { j } } } = \\langle v _ { i } , v _ { j } \\rangle _ { Y } = 0 = \\overline { { b } } _ { i } \\overline { { b } } _ { j } .", + "type": "interline_equation", + "image_path": "64b88123c6d74a35b1c5dc9c829e55ebd9514f524ae9e97af52afe2fde456f6f.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 246, + 256, + 365, + 271 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 168, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 169, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 133, + 289 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 133, + 277, + 165, + 288 + ], + "score": 0.9, + "content": "i , j \\neq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 276, + 169, + 289 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 276, + 169, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 293, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 292, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 305, + 307 + ], + "score": 1.0, + "content": "Importantly, this does not mean that the variables", + "type": "text" + }, + { + "bbox": [ + 305, + 295, + 347, + 305 + ], + "score": 0.86, + "content": "u _ { 1 } , u _ { 2 } , \\ldots", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 292, + 366, + 307 + ], + "score": 1.0, + "content": "nor", + "type": "text" + }, + { + "bbox": [ + 366, + 296, + 406, + 305 + ], + "score": 0.87, + "content": "v _ { 1 } , v _ { 2 } , \\ldots", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 292, + 506, + 307 + ], + "score": 1.0, + "content": "are “disentangled”, i.e.,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 305, + 504, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 504, + 317 + ], + "score": 1.0, + "content": "they are not statistically independent. These variables represent components in the space of proba-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 314, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 506, + 329 + ], + "score": 1.0, + "content": "bility vectors, rather than the “sample” space. They should be understood as spanning a subspace", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 339 + ], + "score": 1.0, + "content": "of the space of functions over the relevant independent variables. We discuss this in more detail in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 337, + 156, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 156, + 348 + ], + "score": 1.0, + "content": "Section 5.3.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 292, + 506, + 348 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 362, + 292, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 190, + 376 + ], + "score": 1.0, + "content": "B.5 CORNERS OF", + "type": "text" + }, + { + "bbox": [ + 190, + 363, + 201, + 372 + ], + "score": 0.69, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 360, + 295, + 376 + ], + "score": 1.0, + "content": "AND LOSS FUNCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 502, + 406 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 504, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 459, + 396 + ], + "score": 1.0, + "content": "The final piece of puzzle we need, is the ability to express the components (corners) of", + "type": "text" + }, + { + "bbox": [ + 459, + 383, + 470, + 393 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 382, + 488, + 396 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 488, + 383, + 504, + 393 + ], + "score": 0.84, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 348, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 348, + 407 + ], + "score": 1.0, + "content": "in the span of possible non-orthogonal families of variables.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 382, + 504, + 407 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 411, + 502, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 504, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 307, + 425 + ], + "score": 1.0, + "content": "Let us therefore consider two arbitrary families", + "type": "text" + }, + { + "bbox": [ + 307, + 411, + 353, + 423 + ], + "score": 0.89, + "content": "f _ { 1 } , \\ldots , f _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 409, + 374, + 425 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 375, + 413, + 420, + 423 + ], + "score": 0.87, + "content": "g _ { 1 } , \\ldots , g _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 409, + 504, + 425 + ], + "score": 1.0, + "content": "of variables, which", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 421, + 359, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 241, + 435 + ], + "score": 1.0, + "content": "respectively represent the vectors", + "type": "text" + }, + { + "bbox": [ + 241, + 422, + 290, + 434 + ], + "score": 0.93, + "content": "p _ { X } f _ { j } \\in V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 421, + 308, + 435 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 308, + 423, + 355, + 434 + ], + "score": 0.9, + "content": "p _ { Y } g _ { j } \\in V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 421, + 359, + 435 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 409, + 504, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 438, + 504, + 461 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 426, + 451 + ], + "score": 1.0, + "content": "Firstly, we need matrices representing the components of the inner products on", + "type": "text" + }, + { + "bbox": [ + 426, + 439, + 440, + 450 + ], + "score": 0.89, + "content": "V _ { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 438, + 459, + 451 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 459, + 439, + 473, + 450 + ], + "score": 0.88, + "content": "V _ { Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 438, + 505, + 451 + ], + "score": 1.0, + "content": ". Those", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 450, + 217, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 217, + 461 + ], + "score": 1.0, + "content": "are the symmetric matrices", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 438, + 505, + 461 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 466, + 368, + 498 + ], + "lines": [ + { + "bbox": [ + 241, + 466, + 368, + 498 + ], + "spans": [ + { + "bbox": [ + 241, + 466, + 368, + 498 + ], + "score": 0.91, + "content": "\\begin{array} { r } { K _ { i j } = \\langle p _ { X } f _ { i } , p _ { X } f _ { j } \\rangle _ { X } = \\overline { { f _ { i } f _ { j } } } , } \\\\ { L _ { i j } = \\langle p _ { Y } g _ { i } , p _ { Y } g _ { j } \\rangle _ { Y } = \\overline { { g _ { i } g _ { j } } } . } \\end{array}", + "type": "interline_equation", + "image_path": "e3b0b84063132afb696e79be3654f58285ebf5ac5071aecd466e0c04ac8ad0dd.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 241, + 466, + 368, + 482.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 241, + 482.0, + 368, + 498.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 507, + 274, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 275, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 175, + 523 + ], + "score": 1.0, + "content": "The components", + "type": "text" + }, + { + "bbox": [ + 175, + 508, + 191, + 520 + ], + "score": 0.91, + "content": "N _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 504, + 203, + 523 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 204, + 508, + 214, + 517 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 504, + 275, + 523 + ], + "score": 1.0, + "content": "are defined by", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 504, + 275, + 523 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 524, + 360, + 551 + ], + "lines": [ + { + "bbox": [ + 250, + 524, + 360, + 551 + ], + "spans": [ + { + "bbox": [ + 250, + 524, + 360, + 551 + ], + "score": 0.94, + "content": "\\mathcal { N } ( p _ { X } f _ { j } ) = \\sum _ { i } N _ { i j } p _ { Y } g _ { i } .", + "type": "interline_equation", + "image_path": "2b7b8a822b8382b09ad5f7ef0f60b06484b822d4bcf9a16b5df140c5972bf464.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 250, + 524, + 360, + 551 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 556, + 294, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 227, + 570 + ], + "score": 1.0, + "content": "Taking the inner product with", + "type": "text" + }, + { + "bbox": [ + 227, + 559, + 250, + 568 + ], + "score": 0.87, + "content": "p _ { Y } g _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 555, + 295, + 570 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 555, + 295, + 570 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 573, + 374, + 600 + ], + "lines": [ + { + "bbox": [ + 237, + 573, + 374, + 600 + ], + "spans": [ + { + "bbox": [ + 237, + 573, + 374, + 600 + ], + "score": 0.94, + "content": "\\langle p _ { Y } g _ { k } , \\mathcal { N } ( p _ { X } f _ { j } ) \\rangle = \\sum _ { i } N _ { i j } L _ { k i } .", + "type": "interline_equation", + "image_path": "b2186206361d77232b516411f4d298cafe0f44cb95c38a9d96ca5dc21515641c.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 237, + 573, + 374, + 600 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 372, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 373, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 373, + 618 + ], + "score": 1.0, + "content": "The left-hand side can be computed using Equ. 13. It is the matrix", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 605, + 373, + 618 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 621, + 400, + 698 + ], + "lines": [ + { + "bbox": [ + 210, + 621, + 400, + 698 + ], + "spans": [ + { + "bbox": [ + 210, + 621, + 400, + 698 + ], + "score": 0.94, + "content": "\\begin{array} { l } { { \\displaystyle { \\cal A } _ { k j } = \\langle p _ { Y } g _ { k } , { \\cal N } ( p _ { X } f _ { j } ) \\rangle } } \\\\ { { \\displaystyle ~ = \\sum _ { x , y } \\frac { p _ { Y } ( y ) g _ { k } ( y ) p _ { Y | X } ( y | x ) p _ { X } ( x ) f _ { j } ( x ) } { p _ { Y } ( y ) } } } \\\\ { { \\displaystyle ~ = \\sum _ { x , y } p ( x , y ) g _ { k } ( y ) f _ { j } ( x ) = \\overline { { g _ { k } f _ { j } } } . } } \\end{array}", + "type": "interline_equation", + "image_path": "9deaf3f2b4a503f27c9b2f2fc081d2425e38207c6b07dc4513908e2dfe0753c2.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 210, + 621, + 400, + 636.4 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 210, + 636.4, + 400, + 651.8 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 210, + 651.8, + 400, + 667.1999999999999 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 210, + 667.1999999999999, + 400, + 682.5999999999999 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 210, + 682.5999999999999, + 400, + 697.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 702, + 329, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 702, + 329, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 276, + 715 + ], + "score": 1.0, + "content": "Therefore, in matrix notation, Equ. (34) is", + "type": "text" + }, + { + "bbox": [ + 277, + 703, + 314, + 712 + ], + "score": 0.9, + "content": "A = L N", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 702, + 329, + 715 + ], + "score": 1.0, + "content": ", or", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 702, + 329, + 715 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 279, + 719, + 331, + 731 + ], + "lines": [ + { + "bbox": [ + 279, + 719, + 331, + 731 + ], + "spans": [ + { + "bbox": [ + 279, + 719, + 331, + 731 + ], + "score": 0.9, + "content": "N = L ^ { - 1 } A .", + "type": "interline_equation", + "image_path": "ae48dea95919ca0dca6569335f749fd8f58bd0eaa7faf859df0a382270205e86.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 279, + 719, + 331, + 731 + ], + "spans": [], + "index": 39 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 417, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 419, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 175, + 98 + ], + "score": 1.0, + "content": "The components", + "type": "text" + }, + { + "bbox": [ + 175, + 83, + 191, + 96 + ], + "score": 0.91, + "content": "N _ { i j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 80, + 203, + 98 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 204, + 83, + 219, + 92 + ], + "score": 0.87, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 80, + 341, + 98 + ], + "score": 1.0, + "content": "are obtained by just swapping", + "type": "text" + }, + { + "bbox": [ + 342, + 83, + 352, + 92 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 80, + 370, + 98 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 370, + 83, + 379, + 92 + ], + "score": 0.79, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 80, + 419, + 98 + ], + "score": 1.0, + "content": ", yielding", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 272, + 103, + 339, + 117 + ], + "lines": [ + { + "bbox": [ + 272, + 103, + 339, + 117 + ], + "spans": [ + { + "bbox": [ + 272, + 103, + 339, + 117 + ], + "score": 0.91, + "content": "N ^ { * } = K ^ { - 1 } A ^ { \\top } .", + "type": "interline_equation", + "image_path": "8607cc6b778302cefea952b8aa87f1d995b39f936a8c08a3448aa169ed9bbd5d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 272, + 103, + 339, + 117 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 131, + 505, + 201 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 276, + 145 + ], + "score": 1.0, + "content": "Hence the singular values of the corner of", + "type": "text" + }, + { + "bbox": [ + 277, + 133, + 288, + 142 + ], + "score": 0.81, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 131, + 386, + 145 + ], + "score": 1.0, + "content": "defined by the variables", + "type": "text" + }, + { + "bbox": [ + 387, + 133, + 397, + 145 + ], + "score": 0.89, + "content": "f _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 131, + 415, + 145 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 415, + 134, + 425, + 145 + ], + "score": 0.86, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "are just the square-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 141, + 504, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 251, + 160 + ], + "score": 1.0, + "content": "root of the eigenvalues of the matrix", + "type": "text" + }, + { + "bbox": [ + 252, + 144, + 348, + 155 + ], + "score": 0.91, + "content": "N ^ { * } N = K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 141, + 465, + 160 + ], + "score": 1.0, + "content": ". In order to find the variables", + "type": "text" + }, + { + "bbox": [ + 466, + 145, + 476, + 157 + ], + "score": 0.88, + "content": "f _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 141, + 493, + 160 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 493, + 146, + 504, + 157 + ], + "score": 0.84, + "content": "g _ { j }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 155, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 229, + 170 + ], + "score": 1.0, + "content": "with the same span as the first", + "type": "text" + }, + { + "bbox": [ + 230, + 156, + 241, + 167 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 155, + 294, + 170 + ], + "score": 1.0, + "content": "eigenvectors", + "type": "text" + }, + { + "bbox": [ + 294, + 158, + 320, + 168 + ], + "score": 0.7, + "content": "u _ { j } , v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 155, + 506, + 170 + ], + "score": 1.0, + "content": ", we just need to maximize all the eigenvalues", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 165, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 117, + 180 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 167, + 141, + 177 + ], + "score": 0.89, + "content": "N ^ { * } N", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 165, + 358, + 180 + ], + "score": 1.0, + "content": ". A simple way to do this is to use (minus) the trace of", + "type": "text" + }, + { + "bbox": [ + 358, + 167, + 382, + 177 + ], + "score": 0.9, + "content": "N ^ { * } N", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 165, + 505, + 180 + ], + "score": 1.0, + "content": "as loss function, since it is the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 305, + 191 + ], + "score": 1.0, + "content": "sum of the square of the singular values. We call", + "type": "text" + }, + { + "bbox": [ + 305, + 178, + 349, + 190 + ], + "score": 0.92, + "content": "\\mathrm { T r } \\left( N ^ { * } N \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 177, + 505, + 191 + ], + "score": 1.0, + "content": "the relevance of the subspaces defines", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 189, + 387, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 172, + 201 + ], + "score": 1.0, + "content": "by the variables", + "type": "text" + }, + { + "bbox": [ + 172, + 189, + 182, + 201 + ], + "score": 0.89, + "content": "f _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 189, + 195, + 201 + ], + "score": 1.0, + "content": "ad", + "type": "text" + }, + { + "bbox": [ + 195, + 190, + 204, + 200 + ], + "score": 0.84, + "content": "g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 189, + 232, + 201 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 233, + 189, + 246, + 200 + ], + "score": 0.87, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 189, + 387, + 201 + ], + "score": 1.0, + "content": ". This yields the loss/cost function:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 208, + 408, + 224 + ], + "lines": [ + { + "bbox": [ + 203, + 208, + 408, + 224 + ], + "spans": [ + { + "bbox": [ + 203, + 208, + 408, + 224 + ], + "score": 0.9, + "content": "C = k _ { 0 } - \\operatorname { T r } \\left( N ^ { \\ast } N \\right) = k _ { 0 } - \\operatorname { T r } \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) .", + "type": "interline_equation", + "image_path": "b4e2488c74a3f265625ac742d2edbb99eb5b739f7810f5d57a05baee1221ccf8.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 203, + 208, + 408, + 224 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 237, + 505, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "Once optimal variables have been found, one can obtain the components of the eigenvectors in the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 249, + 400, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 138, + 261 + ], + "score": 1.0, + "content": "span of", + "type": "text" + }, + { + "bbox": [ + 138, + 249, + 184, + 261 + ], + "score": 0.93, + "content": "f _ { 1 } , \\ldots , f _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 249, + 372, + 261 + ], + "score": 1.0, + "content": "through standard numerical diagonalization of", + "type": "text" + }, + { + "bbox": [ + 372, + 249, + 396, + 259 + ], + "score": 0.9, + "content": "N ^ { * } N", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 249, + 400, + 261 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 107, + 276, + 183, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 184, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 184, + 289 + ], + "score": 1.0, + "content": "B.6 INFERENCE", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 213, + 311 + ], + "score": 1.0, + "content": "The variables minimizing", + "type": "text" + }, + { + "bbox": [ + 213, + 299, + 222, + 309 + ], + "score": 0.8, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "can be used to infer one variable from the other. For instance, given", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 107, + 313, + 113, + 321 + ], + "score": 0.71, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 309, + 280, + 322 + ], + "score": 1.0, + "content": ", the inferred probability distribution over", + "type": "text" + }, + { + "bbox": [ + 281, + 311, + 288, + 319 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 309, + 335, + 322 + ], + "score": 1.0, + "content": "is given by", + "type": "text" + }, + { + "bbox": [ + 335, + 309, + 437, + 322 + ], + "score": 0.92, + "content": "p _ { X | Y } ( x | y ) = \\mathcal { N } ^ { * } ( \\delta _ { y } ) ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 309, + 468, + 322 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 468, + 309, + 494, + 322 + ], + "score": 0.92, + "content": "\\delta _ { y } ( \\bar { y } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 139, + 333 + ], + "score": 1.0, + "content": "1 when", + "type": "text" + }, + { + "bbox": [ + 139, + 321, + 169, + 333 + ], + "score": 0.91, + "content": "y = y ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "and zero otherwise. In order to compute this, we first need the components of the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 332, + 478, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 155, + 345 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 155, + 333, + 165, + 344 + ], + "score": 0.88, + "content": "\\delta _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 332, + 255, + 345 + ], + "score": 1.0, + "content": "in terms of the family", + "type": "text" + }, + { + "bbox": [ + 255, + 334, + 325, + 344 + ], + "score": 0.9, + "content": "p _ { Y } g _ { 1 } , \\ldots , p _ { Y } g _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 332, + 415, + 345 + ], + "score": 1.0, + "content": ", i.e., the real numbers", + "type": "text" + }, + { + "bbox": [ + 415, + 332, + 437, + 345 + ], + "score": 0.92, + "content": "( \\delta _ { y } ) _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 332, + 478, + 345 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 352, + 387, + 387 + ], + "lines": [ + { + "bbox": [ + 223, + 352, + 387, + 387 + ], + "spans": [ + { + "bbox": [ + 223, + 352, + 387, + 387 + ], + "score": 0.95, + "content": "\\delta _ { y } ( y ^ { \\prime } ) = p _ { Y } ( y ^ { \\prime } ) \\sum _ { i = 1 } ^ { k _ { 0 } } ( \\delta _ { y } ) _ { i } g _ { i } ( y ^ { \\prime } ) + r ( y ^ { \\prime } ) ,", + "type": "interline_equation", + "image_path": "f872f27d4090ce21e2aad0a936530e9e6025e37a3b62a65bec1564c77eff912d.jpg" + } + ] + } + ], + "index": 16.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 352, + 387, + 369.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 223, + 369.5, + 387, + 387.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 420, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 421, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 133, + 408 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 395, + 196, + 407 + ], + "score": 0.93, + "content": "\\langle r , p _ { Y } \\delta _ { i } \\rangle _ { Y } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 393, + 223, + 408 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 224, + 396, + 228, + 405 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 393, + 353, + 408 + ], + "score": 1.0, + "content": ". Taking the inner product with", + "type": "text" + }, + { + "bbox": [ + 353, + 397, + 375, + 407 + ], + "score": 0.88, + "content": "p _ { Y } g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 393, + 421, + 408 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 415, + 363, + 450 + ], + "lines": [ + { + "bbox": [ + 247, + 415, + 363, + 450 + ], + "spans": [ + { + "bbox": [ + 247, + 415, + 363, + 450 + ], + "score": 0.94, + "content": "\\langle p _ { Y } g _ { j } , \\delta _ { y } \\rangle _ { Y } = \\sum _ { i = 1 } ^ { k _ { 0 } } ( \\delta _ { y } ) _ { i } L _ { j i } ,", + "type": "interline_equation", + "image_path": "975c3a3c2adecafe908e86c2a9af738588c69578219d8b9bdf6fafdfe8f06ccb.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 415, + 363, + 432.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 247, + 432.5, + 363, + 450.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 457, + 248, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 249, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 249, + 469 + ], + "score": 1.0, + "content": "where the left hand side is also just", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 260, + 477, + 351, + 491 + ], + "lines": [ + { + "bbox": [ + 260, + 477, + 351, + 491 + ], + "spans": [ + { + "bbox": [ + 260, + 477, + 351, + 491 + ], + "score": 0.91, + "content": "\\langle p _ { Y } g _ { j } , \\delta _ { y } \\rangle _ { Y } = g _ { j } ( y ) .", + "type": "interline_equation", + "image_path": "1ed107ef27991b943bccefb6b6bbba58532efbd7417b6e102276a318fbafa637.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 260, + 477, + 351, + 491 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 290, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 291, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 224, + 514 + ], + "score": 1.0, + "content": "Therefore the components of", + "type": "text" + }, + { + "bbox": [ + 225, + 499, + 235, + 510 + ], + "score": 0.89, + "content": "\\delta _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 495, + 291, + 514 + ], + "score": 1.0, + "content": "are explicitly", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 518, + 360, + 546 + ], + "lines": [ + { + "bbox": [ + 252, + 518, + 360, + 546 + ], + "spans": [ + { + "bbox": [ + 252, + 518, + 360, + 546 + ], + "score": 0.94, + "content": "( \\delta _ { y } ) _ { i } = \\sum _ { j } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) .", + "type": "interline_equation", + "image_path": "cfc6c6eba7b5b121c7de8a15d7862c781c1896c64d26b9e202ec11c8c4ad5386.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 252, + 518, + 360, + 546 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 164, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 165, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 165, + 572 + ], + "score": 1.0, + "content": "It follows that", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 575, + 391, + 621 + ], + "lines": [ + { + "bbox": [ + 219, + 575, + 391, + 621 + ], + "spans": [ + { + "bbox": [ + 219, + 575, + 391, + 621 + ], + "score": 0.93, + "content": "\\begin{array} { l } { { \\displaystyle p _ { X | Y } ( x | y ) = \\mathcal { N } ^ { * } ( \\delta _ { y } ) ( x ) \\approx \\mathcal { N } _ { 0 } ^ { * } ( \\delta _ { y } ) ( x ) } } \\\\ { { \\displaystyle ~ = \\sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) f _ { k } ( x ) . } } \\end{array}", + "type": "interline_equation", + "image_path": "1ea01b53b365947e50f1df1e28eada3c55875a450cc944384f38c51dcd92f6c4.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 219, + 575, + 391, + 590.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 219, + 590.3333333333334, + 391, + 605.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 219, + 605.6666666666667, + 391, + 621.0000000000001 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 325, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 326, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 305, + 646 + ], + "score": 1.0, + "content": "Then, for instance, the expected inferred value of", + "type": "text" + }, + { + "bbox": [ + 305, + 634, + 315, + 644 + ], + "score": 0.86, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 633, + 326, + 646 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 652, + 397, + 681 + ], + "lines": [ + { + "bbox": [ + 214, + 652, + 397, + 681 + ], + "spans": [ + { + "bbox": [ + 214, + 652, + 397, + 681 + ], + "score": 0.93, + "content": "\\overline { { { x } } } = \\sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) \\sum _ { x } p _ { X } ( x ) x f _ { k } ( x ) .", + "type": "interline_equation", + "image_path": "ede1edc95774bb1a48d73bb19789c9232db331bf80900a374ee8cf177b94faca.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 652, + 397, + 666.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 214, + 666.5, + 397, + 681.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 689, + 266, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 266, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 188, + 702 + ], + "score": 1.0, + "content": "For the inference of", + "type": "text" + }, + { + "bbox": [ + 188, + 689, + 197, + 698 + ], + "score": 0.85, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 687, + 220, + 702 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 221, + 692, + 227, + 698 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 687, + 266, + 702 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 707, + 392, + 735 + ], + "lines": [ + { + "bbox": [ + 218, + 707, + 392, + 735 + ], + "spans": [ + { + "bbox": [ + 218, + 707, + 392, + 735 + ], + "score": 0.94, + "content": "p _ { Y | X } ( y | x ) \\approx \\sum _ { i j k } N _ { k i } ( K ^ { - 1 } ) _ { i j } f _ { j } ( x ) g _ { k } ( y ) .", + "type": "interline_equation", + "image_path": "a6d8ff69ee7efce2d0c01dcd31e4bc7d29a0fe27c24506780c53243d50eb0f5a.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 218, + 707, + 392, + 735 + ], + "spans": [], + "index": 33 + } + ] + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 417, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 419, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 175, + 98 + ], + "score": 1.0, + "content": "The components", + "type": "text" + }, + { + "bbox": [ + 175, + 83, + 191, + 96 + ], + "score": 0.91, + "content": "N _ { i j } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 80, + 203, + 98 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 204, + 83, + 219, + 92 + ], + "score": 0.87, + "content": "\\mathcal { N } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 80, + 341, + 98 + ], + "score": 1.0, + "content": "are obtained by just swapping", + "type": "text" + }, + { + "bbox": [ + 342, + 83, + 352, + 92 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 80, + 370, + 98 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 370, + 83, + 379, + 92 + ], + "score": 0.79, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 80, + 419, + 98 + ], + "score": 1.0, + "content": ", yielding", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 80, + 419, + 98 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 272, + 103, + 339, + 117 + ], + "lines": [ + { + "bbox": [ + 272, + 103, + 339, + 117 + ], + "spans": [ + { + "bbox": [ + 272, + 103, + 339, + 117 + ], + "score": 0.91, + "content": "N ^ { * } = K ^ { - 1 } A ^ { \\top } .", + "type": "interline_equation", + "image_path": "8607cc6b778302cefea952b8aa87f1d995b39f936a8c08a3448aa169ed9bbd5d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 272, + 103, + 339, + 117 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 131, + 505, + 201 + ], + "lines": [ + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 276, + 145 + ], + "score": 1.0, + "content": "Hence the singular values of the corner of", + "type": "text" + }, + { + "bbox": [ + 277, + 133, + 288, + 142 + ], + "score": 0.81, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 131, + 386, + 145 + ], + "score": 1.0, + "content": "defined by the variables", + "type": "text" + }, + { + "bbox": [ + 387, + 133, + 397, + 145 + ], + "score": 0.89, + "content": "f _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 131, + 415, + 145 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 415, + 134, + 425, + 145 + ], + "score": 0.86, + "content": "g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "are just the square-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 141, + 504, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 251, + 160 + ], + "score": 1.0, + "content": "root of the eigenvalues of the matrix", + "type": "text" + }, + { + "bbox": [ + 252, + 144, + 348, + 155 + ], + "score": 0.91, + "content": "N ^ { * } N = K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 141, + 465, + 160 + ], + "score": 1.0, + "content": ". In order to find the variables", + "type": "text" + }, + { + "bbox": [ + 466, + 145, + 476, + 157 + ], + "score": 0.88, + "content": "f _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 141, + 493, + 160 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 493, + 146, + 504, + 157 + ], + "score": 0.84, + "content": "g _ { j }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 155, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 229, + 170 + ], + "score": 1.0, + "content": "with the same span as the first", + "type": "text" + }, + { + "bbox": [ + 230, + 156, + 241, + 167 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 155, + 294, + 170 + ], + "score": 1.0, + "content": "eigenvectors", + "type": "text" + }, + { + "bbox": [ + 294, + 158, + 320, + 168 + ], + "score": 0.7, + "content": "u _ { j } , v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 155, + 506, + 170 + ], + "score": 1.0, + "content": ", we just need to maximize all the eigenvalues", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 165, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 117, + 180 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 167, + 141, + 177 + ], + "score": 0.89, + "content": "N ^ { * } N", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 165, + 358, + 180 + ], + "score": 1.0, + "content": ". A simple way to do this is to use (minus) the trace of", + "type": "text" + }, + { + "bbox": [ + 358, + 167, + 382, + 177 + ], + "score": 0.9, + "content": "N ^ { * } N", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 165, + 505, + 180 + ], + "score": 1.0, + "content": "as loss function, since it is the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 305, + 191 + ], + "score": 1.0, + "content": "sum of the square of the singular values. We call", + "type": "text" + }, + { + "bbox": [ + 305, + 178, + 349, + 190 + ], + "score": 0.92, + "content": "\\mathrm { T r } \\left( N ^ { * } N \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 177, + 505, + 191 + ], + "score": 1.0, + "content": "the relevance of the subspaces defines", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 189, + 387, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 172, + 201 + ], + "score": 1.0, + "content": "by the variables", + "type": "text" + }, + { + "bbox": [ + 172, + 189, + 182, + 201 + ], + "score": 0.89, + "content": "f _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 189, + 195, + 201 + ], + "score": 1.0, + "content": "ad", + "type": "text" + }, + { + "bbox": [ + 195, + 190, + 204, + 200 + ], + "score": 0.84, + "content": "g _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 189, + 232, + 201 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 233, + 189, + 246, + 200 + ], + "score": 0.87, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 189, + 387, + 201 + ], + "score": 1.0, + "content": ". This yields the loss/cost function:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5, + "bbox_fs": [ + 104, + 131, + 506, + 201 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 208, + 408, + 224 + ], + "lines": [ + { + "bbox": [ + 203, + 208, + 408, + 224 + ], + "spans": [ + { + "bbox": [ + 203, + 208, + 408, + 224 + ], + "score": 0.9, + "content": "C = k _ { 0 } - \\operatorname { T r } \\left( N ^ { \\ast } N \\right) = k _ { 0 } - \\operatorname { T r } \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) .", + "type": "interline_equation", + "image_path": "b4e2488c74a3f265625ac742d2edbb99eb5b739f7810f5d57a05baee1221ccf8.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 203, + 208, + 408, + 224 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 237, + 505, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "Once optimal variables have been found, one can obtain the components of the eigenvectors in the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 249, + 400, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 138, + 261 + ], + "score": 1.0, + "content": "span of", + "type": "text" + }, + { + "bbox": [ + 138, + 249, + 184, + 261 + ], + "score": 0.93, + "content": "f _ { 1 } , \\ldots , f _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 249, + 372, + 261 + ], + "score": 1.0, + "content": "through standard numerical diagonalization of", + "type": "text" + }, + { + "bbox": [ + 372, + 249, + 396, + 259 + ], + "score": 0.9, + "content": "N ^ { * } N", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 249, + 400, + 261 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 237, + 505, + 261 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 276, + 183, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 184, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 184, + 289 + ], + "score": 1.0, + "content": "B.6 INFERENCE", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 213, + 311 + ], + "score": 1.0, + "content": "The variables minimizing", + "type": "text" + }, + { + "bbox": [ + 213, + 299, + 222, + 309 + ], + "score": 0.8, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 297, + 505, + 311 + ], + "score": 1.0, + "content": "can be used to infer one variable from the other. For instance, given", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 107, + 313, + 113, + 321 + ], + "score": 0.71, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 309, + 280, + 322 + ], + "score": 1.0, + "content": ", the inferred probability distribution over", + "type": "text" + }, + { + "bbox": [ + 281, + 311, + 288, + 319 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 309, + 335, + 322 + ], + "score": 1.0, + "content": "is given by", + "type": "text" + }, + { + "bbox": [ + 335, + 309, + 437, + 322 + ], + "score": 0.92, + "content": "p _ { X | Y } ( x | y ) = \\mathcal { N } ^ { * } ( \\delta _ { y } ) ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 309, + 468, + 322 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 468, + 309, + 494, + 322 + ], + "score": 0.92, + "content": "\\delta _ { y } ( \\bar { y } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 139, + 333 + ], + "score": 1.0, + "content": "1 when", + "type": "text" + }, + { + "bbox": [ + 139, + 321, + 169, + 333 + ], + "score": 0.91, + "content": "y = y ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "and zero otherwise. In order to compute this, we first need the components of the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 332, + 478, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 155, + 345 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 155, + 333, + 165, + 344 + ], + "score": 0.88, + "content": "\\delta _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 332, + 255, + 345 + ], + "score": 1.0, + "content": "in terms of the family", + "type": "text" + }, + { + "bbox": [ + 255, + 334, + 325, + 344 + ], + "score": 0.9, + "content": "p _ { Y } g _ { 1 } , \\ldots , p _ { Y } g _ { k _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 332, + 415, + 345 + ], + "score": 1.0, + "content": ", i.e., the real numbers", + "type": "text" + }, + { + "bbox": [ + 415, + 332, + 437, + 345 + ], + "score": 0.92, + "content": "( \\delta _ { y } ) _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 332, + 478, + 345 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 297, + 506, + 345 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 352, + 387, + 387 + ], + "lines": [ + { + "bbox": [ + 223, + 352, + 387, + 387 + ], + "spans": [ + { + "bbox": [ + 223, + 352, + 387, + 387 + ], + "score": 0.95, + "content": "\\delta _ { y } ( y ^ { \\prime } ) = p _ { Y } ( y ^ { \\prime } ) \\sum _ { i = 1 } ^ { k _ { 0 } } ( \\delta _ { y } ) _ { i } g _ { i } ( y ^ { \\prime } ) + r ( y ^ { \\prime } ) ,", + "type": "interline_equation", + "image_path": "f872f27d4090ce21e2aad0a936530e9e6025e37a3b62a65bec1564c77eff912d.jpg" + } + ] + } + ], + "index": 16.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 352, + 387, + 369.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 223, + 369.5, + 387, + 387.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 394, + 420, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 421, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 133, + 408 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 395, + 196, + 407 + ], + "score": 0.93, + "content": "\\langle r , p _ { Y } \\delta _ { i } \\rangle _ { Y } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 393, + 223, + 408 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 224, + 396, + 228, + 405 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 393, + 353, + 408 + ], + "score": 1.0, + "content": ". Taking the inner product with", + "type": "text" + }, + { + "bbox": [ + 353, + 397, + 375, + 407 + ], + "score": 0.88, + "content": "p _ { Y } g _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 393, + 421, + 408 + ], + "score": 1.0, + "content": ", we obtain", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 393, + 421, + 408 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 415, + 363, + 450 + ], + "lines": [ + { + "bbox": [ + 247, + 415, + 363, + 450 + ], + "spans": [ + { + "bbox": [ + 247, + 415, + 363, + 450 + ], + "score": 0.94, + "content": "\\langle p _ { Y } g _ { j } , \\delta _ { y } \\rangle _ { Y } = \\sum _ { i = 1 } ^ { k _ { 0 } } ( \\delta _ { y } ) _ { i } L _ { j i } ,", + "type": "interline_equation", + "image_path": "975c3a3c2adecafe908e86c2a9af738588c69578219d8b9bdf6fafdfe8f06ccb.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 415, + 363, + 432.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 247, + 432.5, + 363, + 450.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 457, + 248, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 249, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 249, + 469 + ], + "score": 1.0, + "content": "where the left hand side is also just", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 455, + 249, + 469 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 260, + 477, + 351, + 491 + ], + "lines": [ + { + "bbox": [ + 260, + 477, + 351, + 491 + ], + "spans": [ + { + "bbox": [ + 260, + 477, + 351, + 491 + ], + "score": 0.91, + "content": "\\langle p _ { Y } g _ { j } , \\delta _ { y } \\rangle _ { Y } = g _ { j } ( y ) .", + "type": "interline_equation", + "image_path": "1ed107ef27991b943bccefb6b6bbba58532efbd7417b6e102276a318fbafa637.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 260, + 477, + 351, + 491 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 290, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 291, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 224, + 514 + ], + "score": 1.0, + "content": "Therefore the components of", + "type": "text" + }, + { + "bbox": [ + 225, + 499, + 235, + 510 + ], + "score": 0.89, + "content": "\\delta _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 495, + 291, + 514 + ], + "score": 1.0, + "content": "are explicitly", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 495, + 291, + 514 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 518, + 360, + 546 + ], + "lines": [ + { + "bbox": [ + 252, + 518, + 360, + 546 + ], + "spans": [ + { + "bbox": [ + 252, + 518, + 360, + 546 + ], + "score": 0.94, + "content": "( \\delta _ { y } ) _ { i } = \\sum _ { j } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) .", + "type": "interline_equation", + "image_path": "cfc6c6eba7b5b121c7de8a15d7862c781c1896c64d26b9e202ec11c8c4ad5386.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 252, + 518, + 360, + 546 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 559, + 164, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 165, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 165, + 572 + ], + "score": 1.0, + "content": "It follows that", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 558, + 165, + 572 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 575, + 391, + 621 + ], + "lines": [ + { + "bbox": [ + 219, + 575, + 391, + 621 + ], + "spans": [ + { + "bbox": [ + 219, + 575, + 391, + 621 + ], + "score": 0.93, + "content": "\\begin{array} { l } { { \\displaystyle p _ { X | Y } ( x | y ) = \\mathcal { N } ^ { * } ( \\delta _ { y } ) ( x ) \\approx \\mathcal { N } _ { 0 } ^ { * } ( \\delta _ { y } ) ( x ) } } \\\\ { { \\displaystyle ~ = \\sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) f _ { k } ( x ) . } } \\end{array}", + "type": "interline_equation", + "image_path": "1ea01b53b365947e50f1df1e28eada3c55875a450cc944384f38c51dcd92f6c4.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 219, + 575, + 391, + 590.3333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 219, + 590.3333333333334, + 391, + 605.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 219, + 605.6666666666667, + 391, + 621.0000000000001 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 325, + 645 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 326, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 305, + 646 + ], + "score": 1.0, + "content": "Then, for instance, the expected inferred value of", + "type": "text" + }, + { + "bbox": [ + 305, + 634, + 315, + 644 + ], + "score": 0.86, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 633, + 326, + 646 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 633, + 326, + 646 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 214, + 652, + 397, + 681 + ], + "lines": [ + { + "bbox": [ + 214, + 652, + 397, + 681 + ], + "spans": [ + { + "bbox": [ + 214, + 652, + 397, + 681 + ], + "score": 0.93, + "content": "\\overline { { { x } } } = \\sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) \\sum _ { x } p _ { X } ( x ) x f _ { k } ( x ) .", + "type": "interline_equation", + "image_path": "ede1edc95774bb1a48d73bb19789c9232db331bf80900a374ee8cf177b94faca.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 214, + 652, + 397, + 666.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 214, + 666.5, + 397, + 681.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 689, + 266, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 266, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 188, + 702 + ], + "score": 1.0, + "content": "For the inference of", + "type": "text" + }, + { + "bbox": [ + 188, + 689, + 197, + 698 + ], + "score": 0.85, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 687, + 220, + 702 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 221, + 692, + 227, + 698 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 687, + 266, + 702 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 687, + 266, + 702 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 707, + 392, + 735 + ], + "lines": [ + { + "bbox": [ + 218, + 707, + 392, + 735 + ], + "spans": [ + { + "bbox": [ + 218, + 707, + 392, + 735 + ], + "score": 0.94, + "content": "p _ { Y | X } ( y | x ) \\approx \\sum _ { i j k } N _ { k i } ( K ^ { - 1 } ) _ { i j } f _ { j } ( x ) g _ { k } ( y ) .", + "type": "interline_equation", + "image_path": "a6d8ff69ee7efce2d0c01dcd31e4bc7d29a0fe27c24506780c53243d50eb0f5a.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 218, + 707, + 392, + 735 + ], + "spans": [], + "index": 33 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 249, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 250, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 250, + 96 + ], + "score": 1.0, + "content": "C ANALYTICAL EXAMPLE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 105, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 132, + 119 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 132, + 106, + 161, + 118 + ], + "score": 0.94, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 106, + 506, + 119 + ], + "score": 1.0, + "content": "is any multivariate Gaussian distribution, everything can be computed analytically. Let", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 315, + 131 + ], + "score": 1.0, + "content": "us consider here the one-dimensional case. We use", + "type": "text" + }, + { + "bbox": [ + 316, + 117, + 421, + 131 + ], + "score": 0.92, + "content": "p ( x ) \\ \\propto \\ \\exp { \\left( - x ^ { 2 } / 2 \\tau ^ { 2 } \\right) }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 117, + 505, + 131 + ], + "score": 1.0, + "content": ", and the conditional", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 129, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 247, + 144 + ], + "score": 0.9, + "content": "\\underline { { { p } } } ( y | x ) ~ \\propto ~ \\exp \\left( - ( y - x ) ^ { 2 } / 2 \\sigma ^ { 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 129, + 288, + 145 + ], + "score": 1.0, + "content": ". That is,", + "type": "text" + }, + { + "bbox": [ + 288, + 133, + 295, + 143 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 129, + 342, + 145 + ], + "score": 1.0, + "content": "is equal to", + "type": "text" + }, + { + "bbox": [ + 342, + 133, + 349, + 141 + ], + "score": 0.81, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 129, + 506, + 145 + ], + "score": 1.0, + "content": "but with some added Gaussian noise.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 140, + 152, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 152, + 155 + ], + "score": 1.0, + "content": "This gives", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 176, + 151, + 435, + 180 + ], + "lines": [ + { + "bbox": [ + 176, + 151, + 435, + 180 + ], + "spans": [ + { + "bbox": [ + 176, + 151, + 435, + 180 + ], + "score": 0.93, + "content": "p _ { X | Y } ( x | y ) \\propto \\exp \\left( - \\frac { ( x - \\gamma y ) ^ { 2 } } { 2 \\tau ^ { 2 } ( 1 - \\gamma ) } \\right) , \\quad \\mathrm { w h e r e } \\quad \\gamma = \\frac { \\tau ^ { 2 } } { \\sigma ^ { 2 } + \\tau ^ { 2 } } .", + "type": "interline_equation", + "image_path": "10899a4858ab6715bf0ec506b7e356d22489e9c4630de49317621359c2d5ac92.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 176, + 151, + 435, + 160.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 176, + 160.66666666666666, + 435, + 170.33333333333331 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 176, + 170.33333333333331, + 435, + 179.99999999999997 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 187, + 503, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 420, + 199 + ], + "score": 1.0, + "content": "It was show in Lancaster (1958), that the most relevant subspace of dimension", + "type": "text" + }, + { + "bbox": [ + 420, + 188, + 431, + 199 + ], + "score": 0.89, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 187, + 493, + 199 + ], + "score": 1.0, + "content": "on the variable", + "type": "text" + }, + { + "bbox": [ + 494, + 188, + 504, + 197 + ], + "score": 0.8, + "content": "X", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 199, + 245, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 245, + 210 + ], + "score": 1.0, + "content": "is simply spanned by the variables", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "interline_equation", + "bbox": [ + 278, + 208, + 332, + 222 + ], + "lines": [ + { + "bbox": [ + 278, + 208, + 332, + 222 + ], + "spans": [ + { + "bbox": [ + 278, + 208, + 332, + 222 + ], + "score": 0.89, + "content": "f _ { n } ( x ) = x ^ { n } ,", + "type": "interline_equation", + "image_path": "73b73808bbc9159c5a966035aa4e8cbaac31db30f53dc1a63c210274fc805b87.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 278, + 208, + 332, + 222 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 252, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 253, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 181, + 236 + ], + "score": 0.9, + "content": "n = 0 , \\ldots , k _ { 0 } - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 223, + 239, + 237 + ], + "score": 1.0, + "content": ". Similarly for", + "type": "text" + }, + { + "bbox": [ + 239, + 225, + 248, + 234 + ], + "score": 0.34, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 223, + 253, + 237 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 279, + 234, + 332, + 248 + ], + "lines": [ + { + "bbox": [ + 279, + 234, + 332, + 248 + ], + "spans": [ + { + "bbox": [ + 279, + 234, + 332, + 248 + ], + "score": 0.88, + "content": "g _ { n } ( y ) = y ^ { n } .", + "type": "interline_equation", + "image_path": "8b45b2b5538fea1629bcb0b9df88672de2464ff4e3b331fa2b06daea2fd0cc03.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 279, + 234, + 332, + 248 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 401, + 263 + ], + "score": 1.0, + "content": "This independence of the relevant variables on the detailed parameters of", + "type": "text" + }, + { + "bbox": [ + 402, + 253, + 408, + 262 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "is a general property of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 221, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 221, + 274 + ], + "score": 1.0, + "content": "Gaussian joint distributions.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 311 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 331, + 291 + ], + "score": 1.0, + "content": "This means, for instance, that the most relevant feature", + "type": "text" + }, + { + "bbox": [ + 331, + 279, + 358, + 289 + ], + "score": 0.86, + "content": "( n = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 276, + 469, + 291 + ], + "score": 1.0, + "content": ") for predicting the value of", + "type": "text" + }, + { + "bbox": [ + 469, + 279, + 479, + 288 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 276, + 505, + 291 + ], + "score": 1.0, + "content": "given", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 136, + 300 + ], + "score": 0.9, + "content": "Y = y", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 289, + 176, + 302 + ], + "score": 1.0, + "content": "is simply", + "type": "text" + }, + { + "bbox": [ + 176, + 289, + 186, + 299 + ], + "score": 0.79, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "itself. The higher order variables have to do with inferring extra aspects of the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 300, + 235, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 221, + 312 + ], + "score": 1.0, + "content": "probability distribution over", + "type": "text" + }, + { + "bbox": [ + 221, + 300, + 231, + 310 + ], + "score": 0.85, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 300, + 235, + 312 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 316, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 504, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 504, + 329 + ], + "score": 1.0, + "content": "A set of orthogonal variables can be obtain from the Gram-Schmidt procedure, which, if done from", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 163, + 340 + ], + "score": 1.0, + "content": "small to large", + "type": "text" + }, + { + "bbox": [ + 164, + 330, + 171, + 338 + ], + "score": 0.72, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 328, + 334, + 340 + ], + "score": 1.0, + "content": "much necessarily yield the eigenvectors", + "type": "text" + }, + { + "bbox": [ + 334, + 330, + 347, + 339 + ], + "score": 0.87, + "content": "u _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 328, + 365, + 340 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 365, + 330, + 376, + 339 + ], + "score": 0.86, + "content": "v _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 328, + 505, + 340 + ], + "score": 1.0, + "content": ". For illustration purpose, let us", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 339, + 455, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 259, + 352 + ], + "score": 1.0, + "content": "work with the non-orthogonal vectors", + "type": "text" + }, + { + "bbox": [ + 259, + 339, + 271, + 351 + ], + "score": 0.89, + "content": "f _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 339, + 289, + 352 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 289, + 340, + 300, + 351 + ], + "score": 0.86, + "content": "g _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 339, + 390, + 352 + ], + "score": 1.0, + "content": ", keeping only the first", + "type": "text" + }, + { + "bbox": [ + 391, + 339, + 420, + 350 + ], + "score": 0.91, + "content": "k _ { 0 } = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 339, + 455, + 352 + ], + "score": 1.0, + "content": "vectors.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 355, + 368, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 369, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 369, + 370 + ], + "score": 1.0, + "content": "The three matrices (correlators) we need can be easily computed:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 120, + 372, + 474, + 453 + ], + "lines": [ + { + "bbox": [ + 120, + 372, + 474, + 453 + ], + "spans": [ + { + "bbox": [ + 120, + 372, + 474, + 453 + ], + "score": 0.95, + "content": "\\begin{array} { l c c } { { K = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } } } & { { 0 } } & { { 3 \\tau ^ { 4 } } } \\end{array} \\right) } } & { { L = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { 3 ( \\tau ^ { 2 } + \\sigma ^ { 2 } ) ^ { 2 } } } \\end{array} \\right) } } \\\\ { { A = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { \\tau ^ { 2 } ( \\sigma ^ { 2 } + 3 \\tau ^ { 2 } ) } } \\end{array} \\right) . } } \\end{array}", + "type": "interline_equation", + "image_path": "21630e4089ab2ac003986e1aea6a1512c08f2f3a4e83cf67d0632043d2fe4ed8.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 120, + 372, + 474, + 399.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 120, + 399.0, + 474, + 426.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 120, + 426.0, + 474, + 453.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 149, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 150, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 150, + 469 + ], + "score": 1.0, + "content": "We obtain", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 464, + 404, + 504 + ], + "lines": [ + { + "bbox": [ + 207, + 464, + 404, + 504 + ], + "spans": [ + { + "bbox": [ + 207, + 464, + 404, + 504 + ], + "score": 0.95, + "content": "M = K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } ( 1 - \\gamma ^ { 2 } ) } } \\\\ { { 0 } } & { { \\gamma } } & { { 0 } } \\\\ { { 0 } } & { { 0 } } & { { \\gamma ^ { 2 } } } \\end{array} \\right) .", + "type": "interline_equation", + "image_path": "f535fc90744ea97d7a26b415efbbb17733ce6c77e7858e4a81860e170aef0734.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 464, + 404, + 484.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 207, + 484.0, + 404, + 504.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 104, + 503, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 104, + 503, + 184, + 519 + ], + "score": 1.0, + "content": "The eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 184, + 506, + 196, + 515 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 503, + 469, + 519 + ], + "score": 1.0, + "content": "can be read on the diagonal, and the corresponding eigenvectors are", + "type": "text" + }, + { + "bbox": [ + 469, + 506, + 501, + 518 + ], + "score": 0.9, + "content": "( 1 , 0 , 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 503, + 505, + 519 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 515, + 504, + 530 + ], + "spans": [ + { + "bbox": [ + 107, + 516, + 139, + 528 + ], + "score": 0.9, + "content": "( 0 , 1 , \\bar { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 515, + 159, + 530 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 159, + 516, + 203, + 528 + ], + "score": 0.92, + "content": "( - \\tau ^ { 2 } , 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 515, + 408, + 530 + ], + "score": 1.0, + "content": ", which means that the eigenfunctions are in order", + "type": "text" + }, + { + "bbox": [ + 409, + 516, + 453, + 529 + ], + "score": 0.91, + "content": "\\bar { u } _ { 0 } ( x ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 515, + 458, + 530 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 459, + 518, + 504, + 529 + ], + "score": 0.89, + "content": "u _ { 1 } ( x ) = x", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 525, + 199, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 525, + 199, + 542 + ], + "score": 1.0, + "content": "and u2(x) = x2 − τ 2.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 376, + 558 + ], + "score": 1.0, + "content": "Because we are working with continuous variables, the true rank of", + "type": "text" + }, + { + "bbox": [ + 376, + 545, + 387, + 555 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "is infinite, even for any finite", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "cutoff on the singular values. Nevertheless, it is instructive to see how the approximate inference", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 567, + 434, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 162, + 579 + ], + "score": 1.0, + "content": "fares for rank", + "type": "text" + }, + { + "bbox": [ + 163, + 567, + 192, + 578 + ], + "score": 0.92, + "content": "k _ { 0 } = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 567, + 261, + 579 + ], + "score": 1.0, + "content": ". Given the value", + "type": "text" + }, + { + "bbox": [ + 261, + 569, + 267, + 578 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 567, + 283, + 579 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 283, + 567, + 292, + 577 + ], + "score": 0.82, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 567, + 412, + 579 + ], + "score": 1.0, + "content": ", the inferred distribution over", + "type": "text" + }, + { + "bbox": [ + 413, + 567, + 423, + 576 + ], + "score": 0.85, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 567, + 434, + 579 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 583, + 417, + 619 + ], + "lines": [ + { + "bbox": [ + 193, + 583, + 417, + 619 + ], + "spans": [ + { + "bbox": [ + 193, + 583, + 417, + 619 + ], + "score": 0.94, + "content": "\\mathcal { N } _ { 0 } ^ { * } ( \\delta _ { y } ) ( x ) = p _ { X } ( x ) p _ { Y } ( y ) \\sum _ { j , k = 0 } ^ { 2 } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } ) _ { k j } y ^ { j } x ^ { k } .", + "type": "interline_equation", + "image_path": "8d97ff77c3cad717d7dd4af9483f1195bb46a057f2b481d6ef2627afbc141356.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 583, + 417, + 601.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 193, + 601.0, + 417, + 619.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 623, + 505, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 335, + 636 + ], + "score": 1.0, + "content": "The approximately inferred first and second moments of", + "type": "text" + }, + { + "bbox": [ + 336, + 624, + 346, + 633 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 622, + 504, + 636 + ], + "score": 1.0, + "content": "is given by integrating the above times", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 634, + 231, + 646 + ], + "spans": [ + { + "bbox": [ + 107, + 638, + 113, + 644 + ], + "score": 0.67, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 634, + 140, + 646 + ], + "score": 1.0, + "content": "(resp.", + "type": "text" + }, + { + "bbox": [ + 140, + 634, + 151, + 644 + ], + "score": 0.81, + "content": "x ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 634, + 176, + 646 + ], + "score": 1.0, + "content": ") over", + "type": "text" + }, + { + "bbox": [ + 176, + 636, + 183, + 644 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 634, + 231, + 646 + ], + "score": 1.0, + "content": ". We obtain", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 651, + 388, + 666 + ], + "lines": [ + { + "bbox": [ + 221, + 651, + 388, + 666 + ], + "spans": [ + { + "bbox": [ + 221, + 651, + 388, + 666 + ], + "score": 0.92, + "content": "{ \\overline { { x } } } = \\gamma y \\quad { \\mathrm { a n d } } \\quad { \\overline { { x ^ { 2 } } } } = \\gamma ^ { 2 } y ^ { 2 } + ( 1 - \\gamma ) \\tau ^ { 2 } ,", + "type": "interline_equation", + "image_path": "82ea07d3179fe2084324edd3855c2cce89f193120aa41b2a4e445fa3416ed17f.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 221, + 651, + 388, + 666 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 671, + 504, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 370, + 685 + ], + "score": 1.0, + "content": "which are actually exact: they are equal to the first two moments of", + "type": "text" + }, + { + "bbox": [ + 371, + 673, + 380, + 682 + ], + "score": 0.85, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 670, + 401, + 685 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 401, + 674, + 423, + 685 + ], + "score": 0.89, + "content": "p _ { X | Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 670, + 505, + 685 + ], + "score": 1.0, + "content": "as given in Eq. (46).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 505, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 338, + 701 + ], + "score": 1.0, + "content": "In fact, it is easy to see that this would be true for the first", + "type": "text" + }, + { + "bbox": [ + 338, + 689, + 366, + 700 + ], + "score": 0.91, + "content": "k _ { 0 } - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 687, + 471, + 701 + ], + "score": 1.0, + "content": "moments had we kept the", + "type": "text" + }, + { + "bbox": [ + 471, + 689, + 482, + 700 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "most", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 700, + 182, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 182, + 711 + ], + "score": 1.0, + "content": "relevant variables.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 309, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 249, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 250, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 250, + 96 + ], + "score": 1.0, + "content": "C ANALYTICAL EXAMPLE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 105, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 132, + 119 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 132, + 106, + 161, + 118 + ], + "score": 0.94, + "content": "p ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 106, + 506, + 119 + ], + "score": 1.0, + "content": "is any multivariate Gaussian distribution, everything can be computed analytically. Let", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 315, + 131 + ], + "score": 1.0, + "content": "us consider here the one-dimensional case. We use", + "type": "text" + }, + { + "bbox": [ + 316, + 117, + 421, + 131 + ], + "score": 0.92, + "content": "p ( x ) \\ \\propto \\ \\exp { \\left( - x ^ { 2 } / 2 \\tau ^ { 2 } \\right) }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 117, + 505, + 131 + ], + "score": 1.0, + "content": ", and the conditional", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 129, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 247, + 144 + ], + "score": 0.9, + "content": "\\underline { { { p } } } ( y | x ) ~ \\propto ~ \\exp \\left( - ( y - x ) ^ { 2 } / 2 \\sigma ^ { 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 129, + 288, + 145 + ], + "score": 1.0, + "content": ". That is,", + "type": "text" + }, + { + "bbox": [ + 288, + 133, + 295, + 143 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 129, + 342, + 145 + ], + "score": 1.0, + "content": "is equal to", + "type": "text" + }, + { + "bbox": [ + 342, + 133, + 349, + 141 + ], + "score": 0.81, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 129, + 506, + 145 + ], + "score": 1.0, + "content": "but with some added Gaussian noise.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 140, + 152, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 152, + 155 + ], + "score": 1.0, + "content": "This gives", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 106, + 506, + 155 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 176, + 151, + 435, + 180 + ], + "lines": [ + { + "bbox": [ + 176, + 151, + 435, + 180 + ], + "spans": [ + { + "bbox": [ + 176, + 151, + 435, + 180 + ], + "score": 0.93, + "content": "p _ { X | Y } ( x | y ) \\propto \\exp \\left( - \\frac { ( x - \\gamma y ) ^ { 2 } } { 2 \\tau ^ { 2 } ( 1 - \\gamma ) } \\right) , \\quad \\mathrm { w h e r e } \\quad \\gamma = \\frac { \\tau ^ { 2 } } { \\sigma ^ { 2 } + \\tau ^ { 2 } } .", + "type": "interline_equation", + "image_path": "10899a4858ab6715bf0ec506b7e356d22489e9c4630de49317621359c2d5ac92.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 176, + 151, + 435, + 160.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 176, + 160.66666666666666, + 435, + 170.33333333333331 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 176, + 170.33333333333331, + 435, + 179.99999999999997 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 187, + 503, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 420, + 199 + ], + "score": 1.0, + "content": "It was show in Lancaster (1958), that the most relevant subspace of dimension", + "type": "text" + }, + { + "bbox": [ + 420, + 188, + 431, + 199 + ], + "score": 0.89, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 187, + 493, + 199 + ], + "score": 1.0, + "content": "on the variable", + "type": "text" + }, + { + "bbox": [ + 494, + 188, + 504, + 197 + ], + "score": 0.8, + "content": "X", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 199, + 245, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 245, + 210 + ], + "score": 1.0, + "content": "is simply spanned by the variables", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 187, + 504, + 210 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 278, + 208, + 332, + 222 + ], + "lines": [ + { + "bbox": [ + 278, + 208, + 332, + 222 + ], + "spans": [ + { + "bbox": [ + 278, + 208, + 332, + 222 + ], + "score": 0.89, + "content": "f _ { n } ( x ) = x ^ { n } ,", + "type": "interline_equation", + "image_path": "73b73808bbc9159c5a966035aa4e8cbaac31db30f53dc1a63c210274fc805b87.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 278, + 208, + 332, + 222 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 252, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 253, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 181, + 236 + ], + "score": 0.9, + "content": "n = 0 , \\ldots , k _ { 0 } - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 223, + 239, + 237 + ], + "score": 1.0, + "content": ". Similarly for", + "type": "text" + }, + { + "bbox": [ + 239, + 225, + 248, + 234 + ], + "score": 0.34, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 223, + 253, + 237 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 223, + 253, + 237 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 279, + 234, + 332, + 248 + ], + "lines": [ + { + "bbox": [ + 279, + 234, + 332, + 248 + ], + "spans": [ + { + "bbox": [ + 279, + 234, + 332, + 248 + ], + "score": 0.88, + "content": "g _ { n } ( y ) = y ^ { n } .", + "type": "interline_equation", + "image_path": "8b45b2b5538fea1629bcb0b9df88672de2464ff4e3b331fa2b06daea2fd0cc03.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 279, + 234, + 332, + 248 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 401, + 263 + ], + "score": 1.0, + "content": "This independence of the relevant variables on the detailed parameters of", + "type": "text" + }, + { + "bbox": [ + 402, + 253, + 408, + 262 + ], + "score": 0.81, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "is a general property of", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 221, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 221, + 274 + ], + "score": 1.0, + "content": "Gaussian joint distributions.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 249, + 506, + 274 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 311 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 331, + 291 + ], + "score": 1.0, + "content": "This means, for instance, that the most relevant feature", + "type": "text" + }, + { + "bbox": [ + 331, + 279, + 358, + 289 + ], + "score": 0.86, + "content": "( n = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 276, + 469, + 291 + ], + "score": 1.0, + "content": ") for predicting the value of", + "type": "text" + }, + { + "bbox": [ + 469, + 279, + 479, + 288 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 276, + 505, + 291 + ], + "score": 1.0, + "content": "given", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 136, + 300 + ], + "score": 0.9, + "content": "Y = y", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 289, + 176, + 302 + ], + "score": 1.0, + "content": "is simply", + "type": "text" + }, + { + "bbox": [ + 176, + 289, + 186, + 299 + ], + "score": 0.79, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "itself. The higher order variables have to do with inferring extra aspects of the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 300, + 235, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 221, + 312 + ], + "score": 1.0, + "content": "probability distribution over", + "type": "text" + }, + { + "bbox": [ + 221, + 300, + 231, + 310 + ], + "score": 0.85, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 300, + 235, + 312 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 276, + 505, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 316, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 504, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 504, + 329 + ], + "score": 1.0, + "content": "A set of orthogonal variables can be obtain from the Gram-Schmidt procedure, which, if done from", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 163, + 340 + ], + "score": 1.0, + "content": "small to large", + "type": "text" + }, + { + "bbox": [ + 164, + 330, + 171, + 338 + ], + "score": 0.72, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 328, + 334, + 340 + ], + "score": 1.0, + "content": "much necessarily yield the eigenvectors", + "type": "text" + }, + { + "bbox": [ + 334, + 330, + 347, + 339 + ], + "score": 0.87, + "content": "u _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 328, + 365, + 340 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 365, + 330, + 376, + 339 + ], + "score": 0.86, + "content": "v _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 328, + 505, + 340 + ], + "score": 1.0, + "content": ". For illustration purpose, let us", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 339, + 455, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 259, + 352 + ], + "score": 1.0, + "content": "work with the non-orthogonal vectors", + "type": "text" + }, + { + "bbox": [ + 259, + 339, + 271, + 351 + ], + "score": 0.89, + "content": "f _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 339, + 289, + 352 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 289, + 340, + 300, + 351 + ], + "score": 0.86, + "content": "g _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 339, + 390, + 352 + ], + "score": 1.0, + "content": ", keeping only the first", + "type": "text" + }, + { + "bbox": [ + 391, + 339, + 420, + 350 + ], + "score": 0.91, + "content": "k _ { 0 } = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 339, + 455, + 352 + ], + "score": 1.0, + "content": "vectors.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 317, + 505, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 355, + 368, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 369, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 369, + 370 + ], + "score": 1.0, + "content": "The three matrices (correlators) we need can be easily computed:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 353, + 369, + 370 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 120, + 372, + 474, + 453 + ], + "lines": [ + { + "bbox": [ + 120, + 372, + 474, + 453 + ], + "spans": [ + { + "bbox": [ + 120, + 372, + 474, + 453 + ], + "score": 0.95, + "content": "\\begin{array} { l c c } { { K = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } } } & { { 0 } } & { { 3 \\tau ^ { 4 } } } \\end{array} \\right) } } & { { L = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { 3 ( \\tau ^ { 2 } + \\sigma ^ { 2 } ) ^ { 2 } } } \\end{array} \\right) } } \\\\ { { A = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { \\tau ^ { 2 } ( \\sigma ^ { 2 } + 3 \\tau ^ { 2 } ) } } \\end{array} \\right) . } } \\end{array}", + "type": "interline_equation", + "image_path": "21630e4089ab2ac003986e1aea6a1512c08f2f3a4e83cf67d0632043d2fe4ed8.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 120, + 372, + 474, + 399.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 120, + 399.0, + 474, + 426.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 120, + 426.0, + 474, + 453.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 149, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 150, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 150, + 469 + ], + "score": 1.0, + "content": "We obtain", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 455, + 150, + 469 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 464, + 404, + 504 + ], + "lines": [ + { + "bbox": [ + 207, + 464, + 404, + 504 + ], + "spans": [ + { + "bbox": [ + 207, + 464, + 404, + 504 + ], + "score": 0.95, + "content": "M = K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } ( 1 - \\gamma ^ { 2 } ) } } \\\\ { { 0 } } & { { \\gamma } } & { { 0 } } \\\\ { { 0 } } & { { 0 } } & { { \\gamma ^ { 2 } } } \\end{array} \\right) .", + "type": "interline_equation", + "image_path": "f535fc90744ea97d7a26b415efbbb17733ce6c77e7858e4a81860e170aef0734.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 464, + 404, + 484.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 207, + 484.0, + 404, + 504.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 104, + 503, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 104, + 503, + 184, + 519 + ], + "score": 1.0, + "content": "The eigenvalues of", + "type": "text" + }, + { + "bbox": [ + 184, + 506, + 196, + 515 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 503, + 469, + 519 + ], + "score": 1.0, + "content": "can be read on the diagonal, and the corresponding eigenvectors are", + "type": "text" + }, + { + "bbox": [ + 469, + 506, + 501, + 518 + ], + "score": 0.9, + "content": "( 1 , 0 , 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 503, + 505, + 519 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 515, + 504, + 530 + ], + "spans": [ + { + "bbox": [ + 107, + 516, + 139, + 528 + ], + "score": 0.9, + "content": "( 0 , 1 , \\bar { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 515, + 159, + 530 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 159, + 516, + 203, + 528 + ], + "score": 0.92, + "content": "( - \\tau ^ { 2 } , 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 515, + 408, + 530 + ], + "score": 1.0, + "content": ", which means that the eigenfunctions are in order", + "type": "text" + }, + { + "bbox": [ + 409, + 516, + 453, + 529 + ], + "score": 0.91, + "content": "\\bar { u } _ { 0 } ( x ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 515, + 458, + 530 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 459, + 518, + 504, + 529 + ], + "score": 0.89, + "content": "u _ { 1 } ( x ) = x", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 525, + 199, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 525, + 199, + 542 + ], + "score": 1.0, + "content": "and u2(x) = x2 − τ 2.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 104, + 503, + 505, + 542 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 376, + 558 + ], + "score": 1.0, + "content": "Because we are working with continuous variables, the true rank of", + "type": "text" + }, + { + "bbox": [ + 376, + 545, + 387, + 555 + ], + "score": 0.84, + "content": "\\mathcal { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "is infinite, even for any finite", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "cutoff on the singular values. Nevertheless, it is instructive to see how the approximate inference", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 567, + 434, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 162, + 579 + ], + "score": 1.0, + "content": "fares for rank", + "type": "text" + }, + { + "bbox": [ + 163, + 567, + 192, + 578 + ], + "score": 0.92, + "content": "k _ { 0 } = 3", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 567, + 261, + 579 + ], + "score": 1.0, + "content": ". Given the value", + "type": "text" + }, + { + "bbox": [ + 261, + 569, + 267, + 578 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 567, + 283, + 579 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 283, + 567, + 292, + 577 + ], + "score": 0.82, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 567, + 412, + 579 + ], + "score": 1.0, + "content": ", the inferred distribution over", + "type": "text" + }, + { + "bbox": [ + 413, + 567, + 423, + 576 + ], + "score": 0.85, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 567, + 434, + 579 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 544, + 505, + 579 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 583, + 417, + 619 + ], + "lines": [ + { + "bbox": [ + 193, + 583, + 417, + 619 + ], + "spans": [ + { + "bbox": [ + 193, + 583, + 417, + 619 + ], + "score": 0.94, + "content": "\\mathcal { N } _ { 0 } ^ { * } ( \\delta _ { y } ) ( x ) = p _ { X } ( x ) p _ { Y } ( y ) \\sum _ { j , k = 0 } ^ { 2 } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } ) _ { k j } y ^ { j } x ^ { k } .", + "type": "interline_equation", + "image_path": "8d97ff77c3cad717d7dd4af9483f1195bb46a057f2b481d6ef2627afbc141356.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 583, + 417, + 601.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 193, + 601.0, + 417, + 619.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 623, + 505, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 335, + 636 + ], + "score": 1.0, + "content": "The approximately inferred first and second moments of", + "type": "text" + }, + { + "bbox": [ + 336, + 624, + 346, + 633 + ], + "score": 0.84, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 622, + 504, + 636 + ], + "score": 1.0, + "content": "is given by integrating the above times", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 634, + 231, + 646 + ], + "spans": [ + { + "bbox": [ + 107, + 638, + 113, + 644 + ], + "score": 0.67, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 634, + 140, + 646 + ], + "score": 1.0, + "content": "(resp.", + "type": "text" + }, + { + "bbox": [ + 140, + 634, + 151, + 644 + ], + "score": 0.81, + "content": "x ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 634, + 176, + 646 + ], + "score": 1.0, + "content": ") over", + "type": "text" + }, + { + "bbox": [ + 176, + 636, + 183, + 644 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 634, + 231, + 646 + ], + "score": 1.0, + "content": ". We obtain", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 622, + 504, + 646 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 651, + 388, + 666 + ], + "lines": [ + { + "bbox": [ + 221, + 651, + 388, + 666 + ], + "spans": [ + { + "bbox": [ + 221, + 651, + 388, + 666 + ], + "score": 0.92, + "content": "{ \\overline { { x } } } = \\gamma y \\quad { \\mathrm { a n d } } \\quad { \\overline { { x ^ { 2 } } } } = \\gamma ^ { 2 } y ^ { 2 } + ( 1 - \\gamma ) \\tau ^ { 2 } ,", + "type": "interline_equation", + "image_path": "82ea07d3179fe2084324edd3855c2cce89f193120aa41b2a4e445fa3416ed17f.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 221, + 651, + 388, + 666 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 671, + 504, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 370, + 685 + ], + "score": 1.0, + "content": "which are actually exact: they are equal to the first two moments of", + "type": "text" + }, + { + "bbox": [ + 371, + 673, + 380, + 682 + ], + "score": 0.85, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 670, + 401, + 685 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 401, + 674, + 423, + 685 + ], + "score": 0.89, + "content": "p _ { X | Y }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 670, + 505, + 685 + ], + "score": 1.0, + "content": "as given in Eq. (46).", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 670, + 505, + 685 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 505, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 338, + 701 + ], + "score": 1.0, + "content": "In fact, it is easy to see that this would be true for the first", + "type": "text" + }, + { + "bbox": [ + 338, + 689, + 366, + 700 + ], + "score": 0.91, + "content": "k _ { 0 } - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 687, + 471, + 701 + ], + "score": 1.0, + "content": "moments had we kept the", + "type": "text" + }, + { + "bbox": [ + 471, + 689, + 482, + 700 + ], + "score": 0.88, + "content": "k _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "most", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 700, + 182, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 182, + 711 + ], + "score": 1.0, + "content": "relevant variables.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 687, + 505, + 711 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/SJeS16EKPr/SJeS16EKPr_model.json b/parse/train/SJeS16EKPr/SJeS16EKPr_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d1be299c5515cb917b7711a997247718487843c --- /dev/null +++ b/parse/train/SJeS16EKPr/SJeS16EKPr_model.json @@ -0,0 +1,31986 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 589, + 1302, + 589, + 1302, + 986, + 398, + 986 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1304, + 1404, + 1304, + 1404, + 1520, + 297, + 1520 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1533, + 1403, + 1533, + 1403, + 1658, + 297, + 1658 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 1118, + 1404, + 1118, + 1404, + 1211, + 300, + 1211 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1863, + 1403, + 1863, + 1403, + 1958, + 297, + 1958 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 662, + 1678, + 1033, + 1678, + 1033, + 1769, + 662, + 1769 + ], + "score": 0.959 + }, + { + "category_id": 0, + "poly": [ + 299, + 223, + 1398, + 223, + 1398, + 321, + 299, + 321 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 296, + 1786, + 1401, + 1786, + 1401, + 1850, + 296, + 1850 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 296, + 1971, + 1400, + 1971, + 1400, + 2035, + 296, + 2035 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 299, + 1226, + 1396, + 1226, + 1396, + 1290, + 299, + 1290 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 302, + 1048, + 574, + 1048, + 574, + 1084, + 302, + 1084 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1707, + 1399, + 1707, + 1399, + 1737, + 1366, + 1737 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 773, + 519, + 927, + 519, + 927, + 553, + 773, + 553 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 313, + 376, + 681, + 376, + 681, + 437, + 313, + 437 + ], + "score": 0.806 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.695 + }, + { + "category_id": 14, + "poly": [ + 665, + 1674, + 1032, + 1674, + 1032, + 1770, + 665, + 1770 + ], + "score": 0.95, + "latex": "p ( \\boldsymbol { y } | \\boldsymbol { x } ) = p ( \\boldsymbol { y } ) \\sum _ { i = 1 } ^ { D } \\eta _ { i } u _ { i } ( \\boldsymbol { x } ) v _ { i } ( \\boldsymbol { y } ) ," + }, + { + "category_id": 13, + "poly": [ + 1106, + 1150, + 1183, + 1150, + 1183, + 1184, + 1106, + 1184 + ], + "score": 0.93, + "latex": "p ( x | y )" + }, + { + "category_id": 13, + "poly": [ + 475, + 1119, + 719, + 1119, + 719, + 1153, + 475, + 1153 + ], + "score": 0.93, + "latex": "( x _ { 1 } , y _ { 1 } ) , \\dotsc , ( x _ { n } , y _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1152, + 1311, + 1152, + 1311, + 1184, + 1235, + 1184 + ], + "score": 0.92, + "latex": "p ( y | x )" + }, + { + "category_id": 13, + "poly": [ + 1274, + 1335, + 1351, + 1335, + 1351, + 1369, + 1274, + 1369 + ], + "score": 0.92, + "latex": "p ( y | x )" + }, + { + "category_id": 13, + "poly": [ + 1269, + 1119, + 1349, + 1119, + 1349, + 1152, + 1269, + 1152 + ], + "score": 0.91, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 1120, + 1787, + 1287, + 1787, + 1287, + 1822, + 1120, + 1822 + ], + "score": 0.91, + "latex": "\\mathbb { E } ( v _ { i } v _ { j } ) = \\delta _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 1234, + 1486, + 1267, + 1486, + 1267, + 1520, + 1234, + 1520 + ], + "score": 0.9, + "latex": "\\chi ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1820, + 580, + 1820, + 580, + 1851, + 297, + 1851 + ], + "score": 0.9, + "latex": "\\eta _ { D } \\leq \\cdot \\cdot \\cdot \\leq \\eta _ { 1 } \\leq \\eta _ { 0 } = 1" + }, + { + "category_id": 13, + "poly": [ + 297, + 1925, + 350, + 1925, + 350, + 1960, + 297, + 1960 + ], + "score": 0.9, + "latex": "p ( y )" + }, + { + "category_id": 13, + "poly": [ + 934, + 1787, + 1105, + 1787, + 1105, + 1822, + 934, + 1822 + ], + "score": 0.89, + "latex": "\\mathbb { E } ( u _ { i } u _ { j } ) = \\delta _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 457, + 1794, + 488, + 1794, + 488, + 1821, + 457, + 1821 + ], + "score": 0.84, + "latex": "u _ { j }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1794, + 402, + 1794, + 402, + 1818, + 374, + 1818 + ], + "score": 0.83, + "latex": "u _ { i }" + }, + { + "category_id": 13, + "poly": [ + 985, + 1974, + 1004, + 1974, + 1004, + 2000, + 985, + 2000 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1351, + 1788, + 1405, + 1788, + 1405, + 1819, + 1351, + 1819 + ], + "score": 0.83, + "latex": "0 ~ <" + }, + { + "category_id": 13, + "poly": [ + 512, + 1233, + 530, + 1233, + 530, + 1259, + 512, + 1259 + ], + "score": 0.78, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1197, + 1463, + 1216, + 1463, + 1216, + 1485, + 1197, + 1485 + ], + "score": 0.75, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 444, + 1234, + 464, + 1234, + 464, + 1255, + 444, + 1255 + ], + "score": 0.73, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 561, + 1432, + 581, + 1432, + 581, + 1459, + 561, + 1459 + ], + "score": 0.72, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 503, + 1434, + 523, + 1434, + 523, + 1455, + 503, + 1455 + ], + "score": 0.68, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1279, + 1979, + 1308, + 1979, + 1308, + 2006, + 1279, + 2006 + ], + "score": 0.64, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1979, + 1263, + 1979, + 1263, + 2003, + 1236, + 2003 + ], + "score": 0.61, + "latex": "u _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1977, + 1308, + 1977, + 1308, + 2006, + 1235, + 2006 + ], + "score": 0.29, + "latex": "u _ { i } , \\ v _ { j }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 222.0, + 1402.0, + 222.0, + 1402.0, + 272.0, + 294.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 282.0, + 490.0, + 282.0, + 490.0, + 329.0, + 293.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 579.0, + 1044.0, + 579.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 518.0, + 932.0, + 518.0, + 932.0, + 556.0, + 770.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 589.0, + 1305.0, + 589.0, + 1305.0, + 623.0, + 395.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 621.0, + 1305.0, + 621.0, + 1305.0, + 653.0, + 396.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 650.0, + 1306.0, + 650.0, + 1306.0, + 685.0, + 394.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 682.0, + 1307.0, + 682.0, + 1307.0, + 715.0, + 392.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 714.0, + 1305.0, + 714.0, + 1305.0, + 744.0, + 392.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 744.0, + 1304.0, + 744.0, + 1304.0, + 776.0, + 394.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 770.0, + 1307.0, + 770.0, + 1307.0, + 807.0, + 393.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 803.0, + 1307.0, + 803.0, + 1307.0, + 838.0, + 393.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 834.0, + 1305.0, + 834.0, + 1305.0, + 866.0, + 394.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 865.0, + 1303.0, + 865.0, + 1303.0, + 897.0, + 394.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 895.0, + 1304.0, + 895.0, + 1304.0, + 930.0, + 393.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 924.0, + 1306.0, + 924.0, + 1306.0, + 959.0, + 393.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 954.0, + 673.0, + 954.0, + 673.0, + 989.0, + 394.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1303.0, + 1404.0, + 1303.0, + 1404.0, + 1342.0, + 294.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1337.0, + 1273.0, + 1337.0, + 1273.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1337.0, + 1405.0, + 1337.0, + 1405.0, + 1371.0, + 1352.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1366.0, + 1406.0, + 1366.0, + 1406.0, + 1401.0, + 294.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1397.0, + 1401.0, + 1397.0, + 1401.0, + 1432.0, + 294.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1424.0, + 502.0, + 1424.0, + 502.0, + 1464.0, + 292.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1424.0, + 560.0, + 1424.0, + 560.0, + 1464.0, + 524.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1424.0, + 1405.0, + 1424.0, + 1405.0, + 1464.0, + 582.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1456.0, + 1196.0, + 1456.0, + 1196.0, + 1495.0, + 294.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1456.0, + 1406.0, + 1456.0, + 1406.0, + 1495.0, + 1217.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1482.0, + 1233.0, + 1482.0, + 1233.0, + 1527.0, + 292.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 1482.0, + 1405.0, + 1482.0, + 1405.0, + 1527.0, + 1268.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1531.0, + 1405.0, + 1531.0, + 1405.0, + 1571.0, + 292.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1563.0, + 1407.0, + 1563.0, + 1407.0, + 1600.0, + 292.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1595.0, + 1405.0, + 1595.0, + 1405.0, + 1632.0, + 293.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1624.0, + 1094.0, + 1624.0, + 1094.0, + 1662.0, + 292.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1116.0, + 474.0, + 1116.0, + 474.0, + 1157.0, + 294.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1116.0, + 1268.0, + 1116.0, + 1268.0, + 1157.0, + 720.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 1116.0, + 1406.0, + 1116.0, + 1406.0, + 1157.0, + 1350.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1149.0, + 1105.0, + 1149.0, + 1105.0, + 1186.0, + 295.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1149.0, + 1234.0, + 1149.0, + 1234.0, + 1186.0, + 1184.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 1149.0, + 1406.0, + 1149.0, + 1406.0, + 1186.0, + 1312.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1180.0, + 947.0, + 1180.0, + 947.0, + 1214.0, + 296.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1864.0, + 1403.0, + 1864.0, + 1403.0, + 1898.0, + 296.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1893.0, + 1403.0, + 1893.0, + 1403.0, + 1928.0, + 293.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1921.0, + 296.0, + 1921.0, + 296.0, + 1964.0, + 288.0, + 1964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1921.0, + 366.0, + 1921.0, + 366.0, + 1964.0, + 351.0, + 1964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1785.0, + 373.0, + 1785.0, + 373.0, + 1824.0, + 295.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1785.0, + 456.0, + 1785.0, + 456.0, + 1824.0, + 403.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1785.0, + 933.0, + 1785.0, + 933.0, + 1824.0, + 489.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1785.0, + 1119.0, + 1785.0, + 1119.0, + 1824.0, + 1106.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1785.0, + 1350.0, + 1785.0, + 1350.0, + 1824.0, + 1288.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1818.0, + 846.0, + 1818.0, + 846.0, + 1851.0, + 581.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 984.0, + 1971.0, + 984.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1971.0, + 1234.0, + 1971.0, + 1234.0, + 2007.0, + 1005.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 1309.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1223.0, + 443.0, + 1223.0, + 443.0, + 1264.0, + 293.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1223.0, + 511.0, + 1223.0, + 511.0, + 1264.0, + 465.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1223.0, + 1401.0, + 1223.0, + 1401.0, + 1264.0, + 531.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1259.0, + 621.0, + 1259.0, + 621.0, + 1291.0, + 296.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 377.0, + 560.0, + 377.0, + 560.0, + 410.0, + 315.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 406.0, + 682.0, + 406.0, + 682.0, + 439.0, + 311.0, + 439.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 823, + 1403, + 823, + 1403, + 978, + 298, + 978 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 307, + 1402, + 307, + 1402, + 430, + 298, + 430 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 546, + 1403, + 546, + 1403, + 700, + 298, + 700 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1177, + 1403, + 1177, + 1403, + 1332, + 298, + 1332 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 716, + 1404, + 716, + 1404, + 809, + 298, + 809 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1346, + 1403, + 1346, + 1403, + 1440, + 299, + 1440 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 1069, + 1404, + 1069, + 1404, + 1162, + 298, + 1162 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 299, + 1632, + 1400, + 1632, + 1400, + 1726, + 299, + 1726 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 297, + 1740, + 1403, + 1740, + 1403, + 1829, + 297, + 1829 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 296, + 1556, + 1401, + 1556, + 1401, + 1617, + 296, + 1617 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 292, + 229, + 1399, + 229, + 1399, + 294, + 292, + 294 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 1971, + 1403, + 1971, + 1403, + 2037, + 296, + 2037 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 496, + 1830, + 1202, + 1830, + 1202, + 1912, + 496, + 1912 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 298, + 993, + 1400, + 993, + 1400, + 1055, + 298, + 1055 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 1924, + 777, + 1924, + 777, + 1959, + 298, + 1959 + ], + "score": 0.929 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 300, + 476, + 582, + 476, + 582, + 512, + 300, + 512 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 300, + 1484, + 471, + 1484, + 471, + 1520, + 300, + 1520 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.745 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1852, + 1400, + 1852, + 1400, + 1881, + 1366, + 1881 + ], + "score": 0.631 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1852, + 1400, + 1852, + 1400, + 1881, + 1366, + 1881 + ], + "score": 0.329 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.131 + }, + { + "category_id": 13, + "poly": [ + 867, + 230, + 957, + 230, + 957, + 264, + 867, + 264 + ], + "score": 0.93, + "latex": "\\mathbb { E } ( u _ { i } v _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 402, + 2002, + 566, + 2002, + 566, + 2037, + 402, + 2037 + ], + "score": 0.93, + "latex": "p _ { Y } ( y ) \\equiv p ( y ) \\rangle" + }, + { + "category_id": 14, + "poly": [ + 496, + 1827, + 1204, + 1827, + 1204, + 1915, + 496, + 1915 + ], + "score": 0.92, + "latex": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } : = \\sum _ { x } { \\frac { \\mu ( x ) \\mu ^ { \\prime } ( x ) } { p ( x ) } } \\quad { \\mathrm { a n d } } \\quad \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } : = \\sum _ { y } { \\frac { \\nu ( y ) \\nu ^ { \\prime } ( y ) } { p ( y ) } }" + }, + { + "category_id": 13, + "poly": [ + 470, + 1925, + 596, + 1925, + 596, + 1958, + 470, + 1958 + ], + "score": 0.92, + "latex": "\\mu , \\mu ^ { \\prime } \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1587, + 378, + 1587, + 378, + 1620, + 298, + 1620 + ], + "score": 0.92, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 787, + 1772, + 841, + 1772, + 841, + 1806, + 787, + 1806 + ], + "score": 0.92, + "latex": "p ( x )" + }, + { + "category_id": 13, + "poly": [ + 895, + 1772, + 948, + 1772, + 948, + 1806, + 895, + 1806 + ], + "score": 0.91, + "latex": "p ( y )" + }, + { + "category_id": 13, + "poly": [ + 1253, + 1772, + 1336, + 1772, + 1336, + 1806, + 1253, + 1806 + ], + "score": 0.91, + "latex": "p ( x , y ) \\mathrm { . }" + }, + { + "category_id": 13, + "poly": [ + 298, + 2003, + 351, + 2003, + 351, + 2036, + 298, + 2036 + ], + "score": 0.91, + "latex": "p ( x )" + }, + { + "category_id": 13, + "poly": [ + 649, + 1924, + 768, + 1924, + 768, + 1957, + 649, + 1957 + ], + "score": 0.91, + "latex": "\\nu , \\nu ^ { \\prime } \\in V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 644, + 1742, + 683, + 1742, + 683, + 1772, + 644, + 1772 + ], + "score": 0.89, + "latex": "V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 733, + 1742, + 770, + 1742, + 770, + 1772, + 733, + 1772 + ], + "score": 0.89, + "latex": "V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 342, + 1634, + 382, + 1634, + 382, + 1664, + 342, + 1664 + ], + "score": 0.89, + "latex": "V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 1289, + 1972, + 1406, + 1972, + 1406, + 2006, + 1289, + 2006 + ], + "score": 0.88, + "latex": "( p _ { X } ( x ) \\equiv" + }, + { + "category_id": 13, + "poly": [ + 436, + 1634, + 474, + 1634, + 474, + 1664, + 436, + 1664 + ], + "score": 0.88, + "latex": "V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 670, + 1978, + 707, + 1978, + 707, + 2005, + 670, + 2005 + ], + "score": 0.86, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 655, + 1666, + 682, + 1666, + 682, + 1692, + 655, + 1692 + ], + "score": 0.85, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 759, + 1978, + 795, + 1978, + 795, + 2005, + 759, + 2005 + ], + "score": 0.85, + "latex": "p _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 684, + 1588, + 712, + 1588, + 712, + 1614, + 684, + 1614 + ], + "score": 0.84, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1256, + 1635, + 1284, + 1635, + 1284, + 1661, + 1256, + 1661 + ], + "score": 0.83, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 762, + 1587, + 787, + 1587, + 787, + 1614, + 762, + 1614 + ], + "score": 0.81, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1336, + 1635, + 1362, + 1635, + 1362, + 1661, + 1336, + 1661 + ], + "score": 0.81, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 470, + 613, + 489, + 613, + 489, + 641, + 470, + 641 + ], + "score": 0.81, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 851, + 268, + 879, + 268, + 879, + 296, + 851, + 296 + ], + "score": 0.8, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 731, + 1666, + 756, + 1666, + 756, + 1692, + 731, + 1692 + ], + "score": 0.79, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 372, + 722, + 390, + 722, + 390, + 748, + 372, + 748 + ], + "score": 0.74, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 812, + 267, + 840, + 267, + 840, + 293, + 812, + 293 + ], + "score": 0.74, + "latex": "u _ { i }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 472.0, + 586.0, + 472.0, + 586.0, + 517.0, + 291.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1477.0, + 478.0, + 1477.0, + 478.0, + 1528.0, + 290.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 863.0, + 2085.0, + 863.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 822.0, + 1404.0, + 822.0, + 1404.0, + 858.0, + 294.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 853.0, + 1403.0, + 853.0, + 1403.0, + 889.0, + 294.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 882.0, + 1405.0, + 882.0, + 1405.0, + 920.0, + 293.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 912.0, + 1405.0, + 912.0, + 1405.0, + 951.0, + 292.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 943.0, + 387.0, + 943.0, + 387.0, + 983.0, + 292.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 309.0, + 1404.0, + 309.0, + 1404.0, + 342.0, + 293.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 339.0, + 1404.0, + 339.0, + 1404.0, + 372.0, + 294.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 369.0, + 1406.0, + 369.0, + 1406.0, + 405.0, + 294.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 400.0, + 767.0, + 400.0, + 767.0, + 432.0, + 296.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 546.0, + 1404.0, + 546.0, + 1404.0, + 583.0, + 296.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 577.0, + 1405.0, + 577.0, + 1405.0, + 614.0, + 293.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 607.0, + 469.0, + 607.0, + 469.0, + 643.0, + 292.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 607.0, + 1404.0, + 607.0, + 1404.0, + 643.0, + 490.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 638.0, + 1405.0, + 638.0, + 1405.0, + 675.0, + 293.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 666.0, + 389.0, + 666.0, + 389.0, + 705.0, + 293.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1179.0, + 1404.0, + 1179.0, + 1404.0, + 1212.0, + 294.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1209.0, + 1403.0, + 1209.0, + 1403.0, + 1243.0, + 296.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1239.0, + 1402.0, + 1239.0, + 1402.0, + 1272.0, + 294.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1268.0, + 1404.0, + 1268.0, + 1404.0, + 1302.0, + 293.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1301.0, + 838.0, + 1301.0, + 838.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 714.0, + 371.0, + 714.0, + 371.0, + 750.0, + 295.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 714.0, + 1404.0, + 714.0, + 1404.0, + 750.0, + 391.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 744.0, + 1405.0, + 744.0, + 1405.0, + 781.0, + 293.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 779.0, + 638.0, + 779.0, + 638.0, + 811.0, + 294.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1346.0, + 1403.0, + 1346.0, + 1403.0, + 1383.0, + 294.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1376.0, + 1404.0, + 1376.0, + 1404.0, + 1412.0, + 293.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1408.0, + 826.0, + 1408.0, + 826.0, + 1443.0, + 293.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1068.0, + 1405.0, + 1068.0, + 1405.0, + 1103.0, + 292.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1101.0, + 1404.0, + 1101.0, + 1404.0, + 1135.0, + 294.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1130.0, + 453.0, + 1130.0, + 453.0, + 1169.0, + 293.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1632.0, + 341.0, + 1632.0, + 341.0, + 1669.0, + 293.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1632.0, + 435.0, + 1632.0, + 435.0, + 1669.0, + 383.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1632.0, + 1255.0, + 1632.0, + 1255.0, + 1669.0, + 475.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1632.0, + 1335.0, + 1632.0, + 1335.0, + 1669.0, + 1285.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 1632.0, + 1404.0, + 1632.0, + 1404.0, + 1669.0, + 1363.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1663.0, + 654.0, + 1663.0, + 654.0, + 1700.0, + 293.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1663.0, + 730.0, + 1663.0, + 730.0, + 1700.0, + 683.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1663.0, + 1405.0, + 1663.0, + 1405.0, + 1700.0, + 757.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1694.0, + 1103.0, + 1694.0, + 1103.0, + 1731.0, + 294.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1738.0, + 643.0, + 1738.0, + 643.0, + 1775.0, + 295.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1738.0, + 732.0, + 1738.0, + 732.0, + 1775.0, + 684.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 1738.0, + 1405.0, + 1738.0, + 1405.0, + 1775.0, + 771.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1770.0, + 786.0, + 1770.0, + 786.0, + 1810.0, + 291.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1770.0, + 894.0, + 1770.0, + 894.0, + 1810.0, + 842.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1770.0, + 1252.0, + 1770.0, + 1252.0, + 1810.0, + 949.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 1770.0, + 1408.0, + 1770.0, + 1408.0, + 1810.0, + 1337.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1799.0, + 336.0, + 1799.0, + 336.0, + 1840.0, + 290.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1553.0, + 1405.0, + 1553.0, + 1405.0, + 1592.0, + 295.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1587.0, + 297.0, + 1587.0, + 297.0, + 1619.0, + 292.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1587.0, + 683.0, + 1587.0, + 683.0, + 1619.0, + 379.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1587.0, + 761.0, + 1587.0, + 761.0, + 1619.0, + 713.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1587.0, + 798.0, + 1587.0, + 798.0, + 1619.0, + 788.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 866.0, + 229.0, + 866.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 958.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 258.0, + 811.0, + 258.0, + 811.0, + 298.0, + 295.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 258.0, + 850.0, + 258.0, + 850.0, + 298.0, + 841.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 258.0, + 1074.0, + 258.0, + 1074.0, + 298.0, + 880.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1966.0, + 669.0, + 1966.0, + 669.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1966.0, + 758.0, + 1966.0, + 758.0, + 2011.0, + 708.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1966.0, + 1288.0, + 1966.0, + 1288.0, + 2011.0, + 796.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 297.0, + 2001.0, + 297.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 2001.0, + 401.0, + 2001.0, + 401.0, + 2039.0, + 352.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 2001.0, + 582.0, + 2001.0, + 582.0, + 2039.0, + 567.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 989.0, + 1405.0, + 989.0, + 1405.0, + 1032.0, + 293.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1023.0, + 1324.0, + 1023.0, + 1324.0, + 1058.0, + 293.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1923.0, + 469.0, + 1923.0, + 469.0, + 1963.0, + 295.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1923.0, + 648.0, + 1923.0, + 648.0, + 1963.0, + 597.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1923.0, + 779.0, + 1923.0, + 779.0, + 1963.0, + 769.0, + 1963.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 337, + 1404, + 337, + 1404, + 462, + 297, + 462 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 295, + 1487, + 1406, + 1487, + 1406, + 1583, + 295, + 1583 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 687, + 1405, + 687, + 1405, + 810, + 297, + 810 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1086, + 1405, + 1086, + 1405, + 1211, + 297, + 1211 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 474, + 1404, + 474, + 1404, + 599, + 297, + 599 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 880, + 1405, + 880, + 1405, + 975, + 297, + 975 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 603, + 992, + 1096, + 992, + 1096, + 1073, + 603, + 1073 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 442, + 1904, + 1251, + 1904, + 1251, + 1982, + 442, + 1982 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 1224, + 1400, + 1224, + 1400, + 1290, + 296, + 1290 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 697, + 1770, + 1000, + 1770, + 1000, + 1839, + 697, + 1839 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 476, + 603, + 1219, + 603, + 1219, + 674, + 476, + 674 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 659, + 281, + 1041, + 281, + 1041, + 322, + 659, + 322 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 722, + 1600, + 975, + 1600, + 975, + 1641, + 722, + 1641 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 686, + 814, + 1011, + 814, + 1011, + 855, + 686, + 855 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 296, + 1851, + 1183, + 1851, + 1183, + 1887, + 296, + 1887 + ], + "score": 0.939 + }, + { + "category_id": 8, + "poly": [ + 605, + 1307, + 1092, + 1307, + 1092, + 1348, + 605, + 1348 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 297, + 1363, + 718, + 1363, + 718, + 1397, + 297, + 1397 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 297, + 227, + 950, + 227, + 950, + 263, + 297, + 263 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 297, + 2003, + 675, + 2003, + 675, + 2037, + 297, + 2037 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.924 + }, + { + "category_id": 8, + "poly": [ + 588, + 1414, + 1112, + 1414, + 1112, + 1456, + 588, + 1456 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1366, + 285, + 1400, + 285, + 1400, + 316, + 1366, + 316 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 297, + 1703, + 1402, + 1703, + 1402, + 1765, + 297, + 1765 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1017, + 1400, + 1017, + 1400, + 1046, + 1366, + 1046 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1365, + 819, + 1400, + 819, + 1400, + 848, + 1365, + 848 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 295, + 1655, + 1399, + 1655, + 1399, + 1691, + 295, + 1691 + ], + "score": 0.872 + }, + { + "category_id": 9, + "poly": [ + 1366, + 614, + 1400, + 614, + 1400, + 643, + 1366, + 643 + ], + "score": 0.787 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.698 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.37 + }, + { + "category_id": 9, + "poly": [ + 1365, + 614, + 1400, + 614, + 1400, + 644, + 1365, + 644 + ], + "score": 0.14 + }, + { + "category_id": 14, + "poly": [ + 445, + 1902, + 1250, + 1902, + 1250, + 1984, + 445, + 1984 + ], + "score": 0.94, + "latex": "p ( y | x ) = { \\mathcal { N } } ( \\delta _ { x } ) ( y ) = { \\frac { 1 } { p ( x ) } } \\sum _ { i } \\eta _ { i } \\nu _ { i } ( y ) \\mu _ { i } ( x ) = p ( y ) \\sum _ { i } \\eta _ { i } v _ { i } ( y ) u _ { i } ( x ) ," + }, + { + "category_id": 14, + "poly": [ + 603, + 991, + 1095, + 991, + 1095, + 1071, + 603, + 1071 + ], + "score": 0.94, + "latex": "\\eta ( q ) = \\frac { \\chi ^ { 2 } ( \\mathcal { N } ( q ) , p _ { Y } ) } { \\chi ^ { 2 } ( q , p _ { X } ) } = \\frac { \\langle \\mathcal { N } ( \\mu ) , \\mathcal { N } ( \\mu ) \\rangle _ { Y } } { \\langle \\mu , \\mu \\rangle _ { X } } ," + }, + { + "category_id": 14, + "poly": [ + 696, + 1768, + 1000, + 1768, + 1000, + 1841, + 696, + 1841 + ], + "score": 0.94, + "latex": "\\mathcal { N } ( \\mu ) = \\sum _ { i } \\eta _ { i } \\nu _ { i } \\langle \\mu _ { i } , \\mu \\rangle _ { X } ," + }, + { + "category_id": 13, + "poly": [ + 543, + 476, + 623, + 476, + 623, + 510, + 543, + 510 + ], + "score": 0.93, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1547, + 524, + 1547, + 524, + 1583, + 297, + 1583 + ], + "score": 0.93, + "latex": "\\mathbb { E } ( f ^ { 2 } ) = \\mathbb { E } ( g ^ { 2 } ) = 1" + }, + { + "category_id": 13, + "poly": [ + 656, + 1551, + 765, + 1551, + 765, + 1582, + 656, + 1582 + ], + "score": 0.93, + "latex": "\\mu = p _ { X } f" + }, + { + "category_id": 13, + "poly": [ + 999, + 911, + 1160, + 911, + 1160, + 946, + 999, + 946 + ], + "score": 0.93, + "latex": "\\chi ^ { 2 } ( \\mathcal { N } ( q ) , p _ { Y } )" + }, + { + "category_id": 13, + "poly": [ + 642, + 1706, + 775, + 1706, + 775, + 1737, + 642, + 1737 + ], + "score": 0.93, + "latex": "\\mu _ { i } = p _ { X } f _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1163, + 720, + 1243, + 720, + 1243, + 753, + 1163, + 753 + ], + "score": 0.92, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 982, + 476, + 1058, + 476, + 1058, + 509, + 982, + 509 + ], + "score": 0.92, + "latex": "p ( y | x )" + }, + { + "category_id": 13, + "poly": [ + 601, + 538, + 691, + 538, + 691, + 570, + 601, + 570 + ], + "score": 0.92, + "latex": "\\mu \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 759, + 1518, + 976, + 1518, + 976, + 1552, + 759, + 1552 + ], + "score": 0.92, + "latex": "\\operatorname { c o r r } ( f , g ) = \\mathbb { E } ( f g )" + }, + { + "category_id": 13, + "poly": [ + 1231, + 912, + 1393, + 912, + 1393, + 945, + 1231, + 945 + ], + "score": 0.92, + "latex": "p _ { Y } = \\mathcal { N } ( p _ { X } )" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1226, + 1352, + 1226, + 1352, + 1258, + 1255, + 1258 + ], + "score": 0.92, + "latex": "\\mu \\in V _ { X }" + }, + { + "category_id": 14, + "poly": [ + 658, + 278, + 1039, + 278, + 1039, + 320, + 658, + 320 + ], + "score": 0.92, + "latex": "\\chi ^ { 2 } ( q , p _ { X } ) = \\langle q - p _ { X } , q - p _ { X } \\rangle _ { X } ," + }, + { + "category_id": 13, + "poly": [ + 988, + 720, + 1043, + 720, + 1043, + 753, + 988, + 753 + ], + "score": 0.92, + "latex": "p ( x )" + }, + { + "category_id": 13, + "poly": [ + 1135, + 1519, + 1191, + 1519, + 1191, + 1552, + 1135, + 1552 + ], + "score": 0.92, + "latex": "f ( x )" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1117, + 1089, + 1117, + 1089, + 1147, + 1020, + 1147 + ], + "score": 0.92, + "latex": "\\mathcal { N } ^ { * } \\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1242, + 1519, + 1294, + 1519, + 1294, + 1552, + 1242, + 1552 + ], + "score": 0.91, + "latex": "g ( y )" + }, + { + "category_id": 14, + "poly": [ + 684, + 813, + 1016, + 813, + 1016, + 853, + 684, + 853 + ], + "score": 0.91, + "latex": "\\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } = \\langle \\mathcal { N } ^ { * } ( \\nu ) , \\mu \\rangle _ { X } ." + }, + { + "category_id": 13, + "poly": [ + 1209, + 538, + 1359, + 538, + 1359, + 571, + 1209, + 571 + ], + "score": 0.91, + "latex": "\\mathcal { N } ^ { * } ( \\nu ) \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 743, + 1258, + 852, + 1258, + 852, + 1289, + 743, + 1289 + ], + "score": 0.91, + "latex": "\\mu = p _ { X } f" + }, + { + "category_id": 13, + "poly": [ + 1076, + 720, + 1129, + 720, + 1129, + 754, + 1076, + 754 + ], + "score": 0.91, + "latex": "p ( y )" + }, + { + "category_id": 13, + "poly": [ + 1022, + 538, + 1158, + 538, + 1158, + 571, + 1022, + 571 + ], + "score": 0.91, + "latex": "\\mathcal { N } ( \\mu ) \\in V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 398, + 913, + 491, + 913, + 491, + 944, + 398, + 944 + ], + "score": 0.91, + "latex": "q \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 1112, + 476, + 1188, + 476, + 1188, + 509, + 1112, + 509 + ], + "score": 0.91, + "latex": "p ( x | y )" + }, + { + "category_id": 14, + "poly": [ + 477, + 600, + 1221, + 600, + 1221, + 677, + 477, + 677 + ], + "score": 0.91, + "latex": "{ \\mathcal { N } } ( \\mu ) ( y ) = \\sum _ { x } p ( y | x ) \\mu ( x ) \\quad { \\mathrm { a n d } } \\quad { \\mathcal { N } } ^ { * } ( \\nu ) ( x ) = \\sum _ { y } p ( x | y ) \\nu ( y ) ." + }, + { + "category_id": 13, + "poly": [ + 743, + 538, + 829, + 538, + 829, + 568, + 743, + 568 + ], + "score": 0.91, + "latex": "\\nu \\in V _ { Y }" + }, + { + "category_id": 14, + "poly": [ + 585, + 1412, + 1113, + 1412, + 1113, + 1454, + 585, + 1454 + ], + "score": 0.91, + "latex": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } = \\operatorname { \\mathbb { E } } ( f f ^ { \\prime } ) \\quad { \\mathrm { a n d } } \\quad \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } = \\operatorname { \\mathbb { E } } ( g g ^ { \\prime } ) ." + }, + { + "category_id": 13, + "poly": [ + 298, + 1257, + 384, + 1257, + 384, + 1287, + 298, + 1287 + ], + "score": 0.91, + "latex": "\\nu \\in V _ { Y }" + }, + { + "category_id": 14, + "poly": [ + 605, + 1306, + 1095, + 1306, + 1095, + 1346, + 605, + 1346 + ], + "score": 0.9, + "latex": "\\mu ( x ) = p ( x ) f ( x ) \\quad { \\mathrm { a n d } } \\quad \\nu ( y ) = p ( y ) g ( y )" + }, + { + "category_id": 13, + "poly": [ + 372, + 2008, + 497, + 2008, + 497, + 2035, + 372, + 2035 + ], + "score": 0.9, + "latex": "\\mu _ { i } = p _ { X } u _ { i }" + }, + { + "category_id": 13, + "poly": [ + 375, + 1091, + 531, + 1091, + 531, + 1120, + 375, + 1120 + ], + "score": 0.9, + "latex": "\\mu = q - p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 1207, + 507, + 1399, + 507, + 1399, + 538, + 1207, + 538 + ], + "score": 0.9, + "latex": "\\mathcal { N ^ { * } } : V _ { X } V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 902, + 1260, + 1005, + 1260, + 1005, + 1289, + 902, + 1289 + ], + "score": 0.9, + "latex": "\\nu = p _ { Y } g" + }, + { + "category_id": 13, + "poly": [ + 779, + 229, + 812, + 229, + 812, + 263, + 779, + 263 + ], + "score": 0.9, + "latex": "\\chi ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 627, + 689, + 669, + 689, + 669, + 718, + 627, + 718 + ], + "score": 0.9, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 816, + 1555, + 917, + 1555, + 917, + 1582, + 816, + 1582 + ], + "score": 0.9, + "latex": "\\nu = p _ { Y } g" + }, + { + "category_id": 13, + "poly": [ + 972, + 510, + 1153, + 510, + 1153, + 538, + 972, + 538 + ], + "score": 0.89, + "latex": "\\mathcal { N } : V _ { X } V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 549, + 2008, + 666, + 2008, + 666, + 2035, + 549, + 2035 + ], + "score": 0.89, + "latex": "\\nu _ { i } = p _ { Y } v _ { i }" + }, + { + "category_id": 14, + "poly": [ + 722, + 1600, + 974, + 1600, + 974, + 1639, + 722, + 1639 + ], + "score": 0.89, + "latex": "\\mathbb { E } ( f g ) = \\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } ," + }, + { + "category_id": 13, + "poly": [ + 778, + 751, + 819, + 751, + 819, + 778, + 778, + 778 + ], + "score": 0.89, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 831, + 1710, + 959, + 1710, + 959, + 1737, + 831, + 1737 + ], + "score": 0.89, + "latex": "\\nu _ { i } = p _ { Y } g _ { i }" + }, + { + "category_id": 13, + "poly": [ + 999, + 1855, + 1028, + 1855, + 1028, + 1884, + 999, + 1884 + ], + "score": 0.88, + "latex": "\\delta _ { x }" + }, + { + "category_id": 13, + "poly": [ + 876, + 373, + 914, + 373, + 914, + 401, + 876, + 401 + ], + "score": 0.88, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 546, + 917, + 584, + 917, + 584, + 944, + 546, + 944 + ], + "score": 0.86, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 1015, + 343, + 1052, + 343, + 1052, + 371, + 1015, + 371 + ], + "score": 0.86, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1370, + 417, + 1370, + 417, + 1396, + 372, + 1396 + ], + "score": 0.86, + "latex": "x , y" + }, + { + "category_id": 13, + "poly": [ + 622, + 1258, + 642, + 1258, + 642, + 1289, + 622, + 1289 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1203, + 1712, + 1228, + 1712, + 1228, + 1737, + 1203, + 1737 + ], + "score": 0.84, + "latex": "\\eta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1036, + 753, + 1065, + 753, + 1065, + 777, + 1036, + 777 + ], + "score": 0.84, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 936, + 913, + 967, + 913, + 967, + 940, + 936, + 940 + ], + "score": 0.83, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1068, + 1658, + 1098, + 1658, + 1098, + 1684, + 1068, + 1684 + ], + "score": 0.83, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 947, + 1664, + 968, + 1664, + 968, + 1690, + 947, + 1690 + ], + "score": 0.83, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 549, + 689, + 578, + 689, + 578, + 717, + 549, + 717 + ], + "score": 0.82, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 710, + 1149, + 739, + 1149, + 739, + 1176, + 710, + 1176 + ], + "score": 0.81, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 946, + 345, + 964, + 345, + 964, + 371, + 946, + 371 + ], + "score": 0.81, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 739, + 948, + 756, + 948, + 756, + 975, + 739, + 975 + ], + "score": 0.81, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 692, + 1261, + 711, + 1261, + 711, + 1289, + 692, + 1289 + ], + "score": 0.79, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 1123, + 374, + 1140, + 374, + 1140, + 401, + 1123, + 401 + ], + "score": 0.79, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 1017, + 1663, + 1036, + 1663, + 1036, + 1684, + 1017, + 1684 + ], + "score": 0.77, + "latex": "\\nu" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 336.0, + 945.0, + 336.0, + 945.0, + 376.0, + 294.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 336.0, + 1014.0, + 336.0, + 1014.0, + 376.0, + 965.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 336.0, + 1406.0, + 336.0, + 1406.0, + 376.0, + 1053.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 368.0, + 875.0, + 368.0, + 875.0, + 403.0, + 292.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 368.0, + 1122.0, + 368.0, + 1122.0, + 403.0, + 915.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 368.0, + 1405.0, + 368.0, + 1405.0, + 403.0, + 1141.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 397.0, + 1405.0, + 397.0, + 1405.0, + 436.0, + 292.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 427.0, + 628.0, + 427.0, + 628.0, + 465.0, + 294.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1487.0, + 1404.0, + 1487.0, + 1404.0, + 1522.0, + 295.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1517.0, + 758.0, + 1517.0, + 758.0, + 1555.0, + 293.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 1517.0, + 1134.0, + 1517.0, + 1134.0, + 1555.0, + 977.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1517.0, + 1241.0, + 1517.0, + 1241.0, + 1555.0, + 1192.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1517.0, + 1406.0, + 1517.0, + 1406.0, + 1555.0, + 1295.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1546.0, + 296.0, + 1546.0, + 296.0, + 1586.0, + 292.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1546.0, + 655.0, + 1546.0, + 655.0, + 1586.0, + 525.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1546.0, + 815.0, + 1546.0, + 815.0, + 1586.0, + 766.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1546.0, + 1028.0, + 1546.0, + 1028.0, + 1586.0, + 918.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 688.0, + 548.0, + 688.0, + 548.0, + 724.0, + 295.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 688.0, + 626.0, + 688.0, + 626.0, + 724.0, + 579.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 688.0, + 1405.0, + 688.0, + 1405.0, + 724.0, + 670.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 718.0, + 987.0, + 718.0, + 987.0, + 757.0, + 292.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 718.0, + 1075.0, + 718.0, + 1075.0, + 757.0, + 1044.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 718.0, + 1162.0, + 718.0, + 1162.0, + 757.0, + 1130.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 718.0, + 1405.0, + 718.0, + 1405.0, + 757.0, + 1244.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 747.0, + 777.0, + 747.0, + 777.0, + 786.0, + 292.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 747.0, + 1035.0, + 747.0, + 1035.0, + 786.0, + 820.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 747.0, + 1406.0, + 747.0, + 1406.0, + 786.0, + 1066.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 780.0, + 466.0, + 780.0, + 466.0, + 813.0, + 295.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1085.0, + 374.0, + 1085.0, + 374.0, + 1125.0, + 294.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1085.0, + 1407.0, + 1085.0, + 1407.0, + 1125.0, + 532.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1117.0, + 1019.0, + 1117.0, + 1019.0, + 1153.0, + 294.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1117.0, + 1406.0, + 1117.0, + 1406.0, + 1153.0, + 1090.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1148.0, + 709.0, + 1148.0, + 709.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1148.0, + 1406.0, + 1148.0, + 1406.0, + 1185.0, + 740.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1178.0, + 949.0, + 1178.0, + 949.0, + 1214.0, + 294.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 473.0, + 542.0, + 473.0, + 542.0, + 513.0, + 294.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 473.0, + 981.0, + 473.0, + 981.0, + 513.0, + 624.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 473.0, + 1111.0, + 473.0, + 1111.0, + 513.0, + 1059.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 473.0, + 1406.0, + 473.0, + 1406.0, + 513.0, + 1189.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 505.0, + 971.0, + 505.0, + 971.0, + 541.0, + 295.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 505.0, + 1206.0, + 505.0, + 1206.0, + 541.0, + 1154.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 505.0, + 1404.0, + 505.0, + 1404.0, + 541.0, + 1400.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 536.0, + 600.0, + 536.0, + 600.0, + 573.0, + 292.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 536.0, + 742.0, + 536.0, + 742.0, + 573.0, + 692.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 536.0, + 1021.0, + 536.0, + 1021.0, + 573.0, + 830.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 536.0, + 1208.0, + 536.0, + 1208.0, + 573.0, + 1159.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 536.0, + 1405.0, + 536.0, + 1405.0, + 573.0, + 1360.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 563.0, + 423.0, + 563.0, + 423.0, + 606.0, + 292.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 876.0, + 1405.0, + 876.0, + 1405.0, + 918.0, + 294.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 906.0, + 397.0, + 906.0, + 397.0, + 951.0, + 291.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 906.0, + 545.0, + 906.0, + 545.0, + 951.0, + 492.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 906.0, + 935.0, + 906.0, + 935.0, + 951.0, + 585.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 906.0, + 998.0, + 906.0, + 998.0, + 951.0, + 968.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 906.0, + 1230.0, + 906.0, + 1230.0, + 951.0, + 1161.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 906.0, + 1406.0, + 906.0, + 1406.0, + 951.0, + 1394.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 941.0, + 738.0, + 941.0, + 738.0, + 977.0, + 294.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 941.0, + 1388.0, + 941.0, + 1388.0, + 977.0, + 757.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1225.0, + 1254.0, + 1225.0, + 1254.0, + 1261.0, + 295.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1225.0, + 1405.0, + 1225.0, + 1405.0, + 1261.0, + 1353.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1254.0, + 297.0, + 1254.0, + 297.0, + 1292.0, + 292.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1254.0, + 621.0, + 1254.0, + 621.0, + 1292.0, + 385.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1254.0, + 691.0, + 1254.0, + 691.0, + 1292.0, + 643.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1254.0, + 742.0, + 1254.0, + 742.0, + 1292.0, + 712.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1254.0, + 901.0, + 1254.0, + 901.0, + 1292.0, + 853.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1254.0, + 1047.0, + 1254.0, + 1047.0, + 1292.0, + 1006.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 998.0, + 1849.0, + 998.0, + 1890.0, + 294.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1849.0, + 1185.0, + 1849.0, + 1185.0, + 1890.0, + 1029.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1358.0, + 371.0, + 1358.0, + 371.0, + 1404.0, + 293.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1358.0, + 720.0, + 1358.0, + 720.0, + 1404.0, + 418.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 778.0, + 225.0, + 778.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 225.0, + 951.0, + 225.0, + 951.0, + 269.0, + 813.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1999.0, + 371.0, + 1999.0, + 371.0, + 2042.0, + 295.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1999.0, + 548.0, + 1999.0, + 548.0, + 2042.0, + 498.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1999.0, + 678.0, + 1999.0, + 678.0, + 2042.0, + 667.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1700.0, + 641.0, + 1700.0, + 641.0, + 1742.0, + 293.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1700.0, + 830.0, + 1700.0, + 830.0, + 1742.0, + 776.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1700.0, + 1202.0, + 1700.0, + 1202.0, + 1742.0, + 960.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1700.0, + 1406.0, + 1700.0, + 1406.0, + 1742.0, + 1229.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1736.0, + 462.0, + 1736.0, + 462.0, + 1768.0, + 291.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1655.0, + 946.0, + 1655.0, + 946.0, + 1694.0, + 295.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1655.0, + 1016.0, + 1655.0, + 1016.0, + 1694.0, + 969.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1655.0, + 1067.0, + 1655.0, + 1067.0, + 1694.0, + 1037.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 1655.0, + 1403.0, + 1655.0, + 1403.0, + 1694.0, + 1099.0, + 1694.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 541, + 1405, + 541, + 1405, + 680, + 297, + 680 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 286, + 1405, + 286, + 1405, + 413, + 297, + 413 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 296, + 1098, + 1405, + 1098, + 1405, + 1222, + 296, + 1222 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 490, + 1396, + 1209, + 1396, + 1209, + 1593, + 490, + 1593 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 1235, + 1405, + 1235, + 1405, + 1331, + 297, + 1331 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1793, + 1403, + 1793, + 1403, + 1888, + 298, + 1888 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 299, + 1685, + 1400, + 1685, + 1400, + 1780, + 299, + 1780 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 292, + 1607, + 1401, + 1607, + 1401, + 1672, + 292, + 1672 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 512, + 937, + 1182, + 937, + 1182, + 1030, + 512, + 1030 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 599, + 699, + 1097, + 699, + 1097, + 790, + 599, + 790 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 294, + 805, + 1401, + 805, + 1401, + 871, + 294, + 871 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 299, + 1049, + 1122, + 1049, + 1122, + 1088, + 299, + 1088 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 295, + 2001, + 994, + 2001, + 994, + 2036, + 295, + 2036 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 297, + 1343, + 1268, + 1343, + 1268, + 1379, + 297, + 1379 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 291, + 884, + 1265, + 884, + 1265, + 918, + 291, + 918 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 106, + 297, + 106 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 298, + 1932, + 532, + 1932, + 532, + 1968, + 298, + 1968 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 301, + 425, + 1201, + 425, + 1201, + 459, + 301, + 459 + ], + "score": 0.917 + }, + { + "category_id": 8, + "poly": [ + 322, + 473, + 1379, + 473, + 1379, + 515, + 322, + 515 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.794 + }, + { + "category_id": 2, + "poly": [ + 293, + 231, + 961, + 231, + 961, + 261, + 293, + 261 + ], + "score": 0.733 + }, + { + "category_id": 1, + "poly": [ + 293, + 231, + 961, + 231, + 961, + 261, + 293, + 261 + ], + "score": 0.138 + }, + { + "category_id": 14, + "poly": [ + 487, + 1393, + 1209, + 1393, + 1209, + 1596, + 487, + 1596 + ], + "score": 0.95, + "latex": "\\begin{array} { l } { { \\displaystyle \\sum _ { y } g _ { k } ( y ) p ( y | x ) = \\sum _ { i , j = 1 } ^ { D } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } \\mathbb { E } ( g _ { k } g _ { i } ) f _ { j } ( x ) } } \\\\ { { \\displaystyle \\qquad = \\sum _ { j = 1 } ^ { D } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) = \\sum _ { j = 1 } ^ { k _ { 0 } } ( A K ^ { - 1 } ) _ { k j } f _ { j } ( x ) } , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 596, + 695, + 1097, + 695, + 1097, + 793, + 596, + 793 + ], + "score": 0.93, + "latex": "\\sum _ { i = 1 } ^ { k _ { 0 } } \\eta _ { i } ^ { 2 } = \\mathrm { T r } \\left( N ^ { \\ast } N \\right) = \\mathrm { T r } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A ) ." + }, + { + "category_id": 14, + "poly": [ + 515, + 932, + 1183, + 932, + 1183, + 1032, + 515, + 1032 + ], + "score": 0.93, + "latex": "p ( \\boldsymbol { y } | \\boldsymbol { x } ) = \\mathcal { N } ( \\delta _ { \\boldsymbol { x } } ) ( \\boldsymbol { y } ) \\simeq p ( \\boldsymbol { y } ) \\sum _ { i , j = 1 } ^ { k _ { 0 } } ( L ^ { - 1 } A K ^ { - 1 } ) _ { i j } g _ { i } ( \\boldsymbol { y } ) f _ { j } ( \\boldsymbol { x } ) ," + }, + { + "category_id": 13, + "poly": [ + 499, + 319, + 638, + 319, + 638, + 352, + 499, + 352 + ], + "score": 0.93, + "latex": "f _ { i } = \\mu _ { i } / p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 482, + 1686, + 563, + 1686, + 563, + 1719, + 482, + 1719 + ], + "score": 0.93, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 892, + 1048, + 1114, + 1048, + 1114, + 1091, + 892, + 1091 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\delta _ { j } = \\sum _ { i } K _ { j i } ^ { - 1 } f _ { i } ( x ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 485, + 1824, + 623, + 1824, + 623, + 1858, + 485, + 1858 + ], + "score": 0.93, + "latex": "f _ { n } ( x ) = x ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 698, + 1608, + 782, + 1608, + 782, + 1640, + 698, + 1640 + ], + "score": 0.93, + "latex": "k \\leq k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 912, + 543, + 1211, + 543, + 1211, + 580, + 912, + 580 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\mathcal { N } ( p _ { X } f _ { j } ) = \\sum _ { i } N _ { i j } p _ { Y } g _ { i } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 676, + 1824, + 813, + 1824, + 813, + 1858, + 676, + 1858 + ], + "score": 0.92, + "latex": "g _ { n } ( y ) = y ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 454, + 648, + 515, + 648, + 515, + 679, + 454, + 679 + ], + "score": 0.92, + "latex": "p _ { X } f _ { i }" + }, + { + "category_id": 13, + "poly": [ + 750, + 576, + 888, + 576, + 888, + 607, + 750, + 607 + ], + "score": 0.91, + "latex": "N = L ^ { - 1 } A" + }, + { + "category_id": 13, + "poly": [ + 1325, + 578, + 1404, + 578, + 1404, + 616, + 1325, + 616 + ], + "score": 0.91, + "latex": "N _ { i j } ^ { * } =" + }, + { + "category_id": 13, + "poly": [ + 858, + 1825, + 1056, + 1825, + 1056, + 1856, + 858, + 1856 + ], + "score": 0.91, + "latex": "n = 0 , \\ldots , k - 1" + }, + { + "category_id": 13, + "poly": [ + 575, + 579, + 619, + 579, + 619, + 613, + 575, + 613 + ], + "score": 0.91, + "latex": "N _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 1186, + 353, + 1314, + 353, + 1314, + 382, + 1186, + 382 + ], + "score": 0.91, + "latex": "\\eta _ { 1 } , \\ldots , \\eta _ { k _ { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 612, + 394, + 612, + 394, + 644, + 297, + 644 + ], + "score": 0.91, + "latex": "K ^ { - 1 } A ^ { \\top }" + }, + { + "category_id": 13, + "poly": [ + 1057, + 807, + 1308, + 807, + 1308, + 841, + 1057, + 841 + ], + "score": 0.91, + "latex": "C = k _ { 0 } - \\mathrm { T r } \\left( N ^ { \\ast } N \\right)" + }, + { + "category_id": 13, + "poly": [ + 326, + 546, + 370, + 546, + 370, + 580, + 326, + 580 + ], + "score": 0.91, + "latex": "N _ { i j }" + }, + { + "category_id": 14, + "poly": [ + 315, + 474, + 1380, + 474, + 1380, + 514, + 315, + 514 + ], + "score": 0.9, + "latex": "K _ { i j } = \\langle \\mu _ { i } , \\mu _ { j } \\rangle _ { X } = \\mathbb { E } ( f _ { i } f _ { j } ) , \\quad L _ { i j } = \\langle \\nu _ { i } , \\nu _ { j } \\rangle _ { X } = \\mathbb { E } ( g _ { i } g _ { j } ) , \\quad A _ { i j } = \\langle \\nu _ { i } , N ( \\mu _ { j } ) \\rangle _ { Y } = \\mathbb { E } ( g _ { i } f _ { j } ) ." + }, + { + "category_id": 13, + "poly": [ + 822, + 351, + 1006, + 351, + 1006, + 380, + 822, + 380 + ], + "score": 0.9, + "latex": "i , j = 1 , \\ldots , k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 565, + 650, + 627, + 650, + 627, + 681, + 565, + 681 + ], + "score": 0.9, + "latex": "p _ { Y } g _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1194, + 1610, + 1219, + 1610, + 1219, + 1641, + 1194, + 1641 + ], + "score": 0.89, + "latex": "f _ { i }" + }, + { + "category_id": 13, + "poly": [ + 829, + 1856, + 895, + 1856, + 895, + 1884, + 829, + 1884 + ], + "score": 0.89, + "latex": "k + 1" + }, + { + "category_id": 13, + "poly": [ + 637, + 1131, + 666, + 1131, + 666, + 1161, + 637, + 1161 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 818, + 1053, + 846, + 1053, + 846, + 1083, + 818, + 1083 + ], + "score": 0.88, + "latex": "\\delta _ { x }" + }, + { + "category_id": 13, + "poly": [ + 601, + 1239, + 631, + 1239, + 631, + 1269, + 601, + 1269 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 443, + 1192, + 473, + 1192, + 473, + 1221, + 443, + 1221 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 773, + 381, + 798, + 381, + 798, + 412, + 773, + 412 + ], + "score": 0.88, + "latex": "f _ { i }" + }, + { + "category_id": 13, + "poly": [ + 489, + 1641, + 519, + 1641, + 519, + 1670, + 489, + 1670 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 705, + 648, + 742, + 648, + 742, + 679, + 705, + 679 + ], + "score": 0.87, + "latex": "i , j" + }, + { + "category_id": 13, + "poly": [ + 1235, + 578, + 1277, + 578, + 1277, + 607, + 1235, + 607 + ], + "score": 0.87, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 604, + 355, + 632, + 355, + 632, + 383, + 604, + 383 + ], + "score": 0.87, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 951, + 382, + 981, + 382, + 981, + 410, + 951, + 410 + ], + "score": 0.86, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 849, + 384, + 877, + 384, + 877, + 414, + 849, + 414 + ], + "score": 0.86, + "latex": "g _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1272, + 1613, + 1300, + 1613, + 1300, + 1643, + 1272, + 1643 + ], + "score": 0.86, + "latex": "g _ { j }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1350, + 817, + 1350, + 817, + 1378, + 788, + 1378 + ], + "score": 0.86, + "latex": "g _ { k }" + }, + { + "category_id": 13, + "poly": [ + 521, + 355, + 549, + 355, + 549, + 380, + 521, + 380 + ], + "score": 0.85, + "latex": "u _ { i }" + }, + { + "category_id": 13, + "poly": [ + 968, + 1348, + 994, + 1348, + 994, + 1373, + 968, + 1373 + ], + "score": 0.85, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 922, + 1797, + 939, + 1797, + 939, + 1822, + 922, + 1822 + ], + "score": 0.84, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 680, + 545, + 711, + 545, + 711, + 572, + 680, + 572 + ], + "score": 0.83, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1122, + 615, + 1153, + 615, + 1153, + 643, + 1122, + 643 + ], + "score": 0.83, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 750, + 1238, + 781, + 1238, + 781, + 1265, + 750, + 1265 + ], + "score": 0.82, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 967, + 1640, + 996, + 1640, + 996, + 1666, + 967, + 1666 + ], + "score": 0.82, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1342, + 1100, + 1372, + 1100, + 1372, + 1127, + 1342, + 1127 + ], + "score": 0.81, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1224, + 1347, + 1252, + 1347, + 1252, + 1373, + 1224, + 1373 + ], + "score": 0.8, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1382, + 1826, + 1401, + 1826, + 1401, + 1852, + 1382, + 1852 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 690, + 318, + 1023, + 318, + 1023, + 352, + 690, + 352 + ], + "score": 0.75, + "latex": "g _ { j } = \\nu _ { j } / p _ { Y } , i , j = 1 , \\ldots , k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 837, + 1131, + 859, + 1131, + 859, + 1157, + 837, + 1157 + ], + "score": 0.4, + "latex": "\\mathbf { B }" + }, + { + "category_id": 13, + "poly": [ + 691, + 319, + 830, + 319, + 830, + 352, + 691, + 352 + ], + "score": 0.32, + "latex": "g _ { j } = \\nu _ { j } / p _ { Y }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1930.0, + 535.0, + 1930.0, + 535.0, + 1974.0, + 292.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 968.0, + 230.0, + 968.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 541.0, + 325.0, + 541.0, + 325.0, + 582.0, + 291.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 541.0, + 679.0, + 541.0, + 679.0, + 582.0, + 371.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 541.0, + 911.0, + 541.0, + 911.0, + 582.0, + 712.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 541.0, + 1406.0, + 541.0, + 1406.0, + 582.0, + 1212.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 573.0, + 574.0, + 573.0, + 574.0, + 619.0, + 290.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 573.0, + 749.0, + 573.0, + 749.0, + 619.0, + 620.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 573.0, + 1234.0, + 573.0, + 1234.0, + 619.0, + 889.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 573.0, + 1324.0, + 573.0, + 1324.0, + 619.0, + 1278.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 612.0, + 1121.0, + 612.0, + 1121.0, + 650.0, + 395.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 612.0, + 1405.0, + 612.0, + 1405.0, + 650.0, + 1154.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 643.0, + 453.0, + 643.0, + 453.0, + 684.0, + 292.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 643.0, + 564.0, + 643.0, + 564.0, + 684.0, + 516.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 643.0, + 704.0, + 643.0, + 704.0, + 684.0, + 628.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 643.0, + 1339.0, + 643.0, + 1339.0, + 684.0, + 743.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 286.0, + 1403.0, + 286.0, + 1403.0, + 324.0, + 292.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 316.0, + 498.0, + 316.0, + 498.0, + 357.0, + 292.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 316.0, + 689.0, + 316.0, + 689.0, + 357.0, + 639.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 316.0, + 1406.0, + 316.0, + 1406.0, + 357.0, + 1024.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 348.0, + 520.0, + 348.0, + 520.0, + 385.0, + 294.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 348.0, + 603.0, + 348.0, + 603.0, + 385.0, + 550.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 348.0, + 821.0, + 348.0, + 821.0, + 385.0, + 633.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 348.0, + 1185.0, + 348.0, + 1185.0, + 385.0, + 1007.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 348.0, + 1406.0, + 348.0, + 1406.0, + 385.0, + 1315.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 379.0, + 772.0, + 379.0, + 772.0, + 416.0, + 294.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 379.0, + 848.0, + 379.0, + 848.0, + 416.0, + 799.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 379.0, + 950.0, + 379.0, + 950.0, + 416.0, + 878.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 379.0, + 1260.0, + 379.0, + 1260.0, + 416.0, + 982.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1098.0, + 1341.0, + 1098.0, + 1341.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1098.0, + 1406.0, + 1098.0, + 1406.0, + 1134.0, + 1373.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1131.0, + 636.0, + 1131.0, + 636.0, + 1164.0, + 295.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1131.0, + 836.0, + 1131.0, + 836.0, + 1164.0, + 667.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1131.0, + 1405.0, + 1131.0, + 1405.0, + 1164.0, + 860.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1194.0, + 293.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1192.0, + 442.0, + 1192.0, + 442.0, + 1224.0, + 294.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1192.0, + 646.0, + 1192.0, + 646.0, + 1224.0, + 474.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1233.0, + 600.0, + 1233.0, + 600.0, + 1274.0, + 292.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1233.0, + 749.0, + 1233.0, + 749.0, + 1274.0, + 632.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1233.0, + 1405.0, + 1233.0, + 1405.0, + 1274.0, + 782.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1267.0, + 1405.0, + 1267.0, + 1405.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1299.0, + 654.0, + 1299.0, + 654.0, + 1331.0, + 293.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1794.0, + 921.0, + 1794.0, + 921.0, + 1828.0, + 294.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1794.0, + 1405.0, + 1794.0, + 1405.0, + 1828.0, + 940.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1825.0, + 484.0, + 1825.0, + 484.0, + 1859.0, + 294.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1825.0, + 675.0, + 1825.0, + 675.0, + 1859.0, + 624.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1825.0, + 857.0, + 1825.0, + 857.0, + 1859.0, + 814.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 1825.0, + 1381.0, + 1825.0, + 1381.0, + 1859.0, + 1057.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1825.0, + 1405.0, + 1825.0, + 1405.0, + 1859.0, + 1402.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1855.0, + 828.0, + 1855.0, + 828.0, + 1889.0, + 294.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1855.0, + 1374.0, + 1855.0, + 1374.0, + 1889.0, + 896.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1684.0, + 481.0, + 1684.0, + 481.0, + 1722.0, + 294.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1684.0, + 1406.0, + 1684.0, + 1406.0, + 1722.0, + 564.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1715.0, + 1405.0, + 1715.0, + 1405.0, + 1751.0, + 294.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1748.0, + 689.0, + 1748.0, + 689.0, + 1782.0, + 295.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1607.0, + 697.0, + 1607.0, + 697.0, + 1643.0, + 295.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1607.0, + 1193.0, + 1607.0, + 1193.0, + 1643.0, + 783.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 1607.0, + 1271.0, + 1607.0, + 1271.0, + 1643.0, + 1220.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 1607.0, + 1403.0, + 1607.0, + 1403.0, + 1643.0, + 1301.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1637.0, + 488.0, + 1637.0, + 488.0, + 1677.0, + 294.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1637.0, + 966.0, + 1637.0, + 966.0, + 1677.0, + 520.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1637.0, + 1144.0, + 1637.0, + 1144.0, + 1677.0, + 997.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 806.0, + 1056.0, + 806.0, + 1056.0, + 842.0, + 296.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 806.0, + 1403.0, + 806.0, + 1403.0, + 842.0, + 1309.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 838.0, + 809.0, + 838.0, + 809.0, + 874.0, + 294.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1047.0, + 817.0, + 1047.0, + 817.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1047.0, + 891.0, + 1047.0, + 891.0, + 1091.0, + 847.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1047.0, + 1123.0, + 1047.0, + 1123.0, + 1091.0, + 1115.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1998.0, + 998.0, + 1998.0, + 998.0, + 2041.0, + 292.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1342.0, + 787.0, + 1342.0, + 787.0, + 1382.0, + 294.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1342.0, + 967.0, + 1342.0, + 967.0, + 1382.0, + 818.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1342.0, + 1223.0, + 1342.0, + 1223.0, + 1382.0, + 995.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1342.0, + 1273.0, + 1342.0, + 1273.0, + 1382.0, + 1253.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 1262.0, + 882.0, + 1262.0, + 921.0, + 294.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 422.0, + 1200.0, + 422.0, + 1200.0, + 463.0, + 294.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 968.0, + 230.0, + 968.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 841, + 1403, + 841, + 1403, + 967, + 298, + 967 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 671, + 1404, + 671, + 1404, + 828, + 297, + 828 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 1157, + 1405, + 1157, + 1405, + 1313, + 296, + 1313 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1833, + 1404, + 1833, + 1404, + 1958, + 298, + 1958 + ], + "score": 0.977 + }, + { + "category_id": 8, + "poly": [ + 671, + 980, + 1027, + 980, + 1027, + 1070, + 671, + 1070 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 528, + 1373, + 1166, + 1373, + 1166, + 1467, + 528, + 1467 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 297, + 1754, + 1401, + 1754, + 1401, + 1819, + 297, + 1819 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 293, + 1080, + 1402, + 1080, + 1402, + 1144, + 293, + 1144 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 299, + 1493, + 1402, + 1493, + 1402, + 1567, + 299, + 1567 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 294, + 228, + 1400, + 228, + 1400, + 296, + 294, + 296 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 319, + 556, + 1339, + 556, + 1339, + 648, + 319, + 648 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 296, + 1579, + 1404, + 1579, + 1404, + 1646, + 296, + 1646 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 298, + 306, + 1402, + 306, + 1402, + 403, + 298, + 403 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1399, + 1971, + 1399, + 2036, + 297, + 2036 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 672, + 458, + 1025, + 458, + 1025, + 499, + 672, + 499 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 298, + 1325, + 892, + 1325, + 892, + 1358, + 298, + 1358 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 292, + 414, + 1259, + 414, + 1259, + 447, + 292, + 447 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 299, + 1686, + 559, + 1686, + 559, + 1723, + 299, + 1723 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 105, + 298, + 105 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 301, + 509, + 1356, + 509, + 1356, + 544, + 301, + 544 + ], + "score": 0.916 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1402, + 1400, + 1402, + 1400, + 1433, + 1352, + 1433 + ], + "score": 0.909 + }, + { + "category_id": 9, + "poly": [ + 1366, + 465, + 1400, + 465, + 1400, + 496, + 1366, + 496 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1009, + 1400, + 1009, + 1400, + 1039, + 1366, + 1039 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1366, + 587, + 1400, + 587, + 1400, + 617, + 1366, + 617 + ], + "score": 0.859 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.747 + }, + { + "category_id": 14, + "poly": [ + 672, + 976, + 1025, + 976, + 1025, + 1072, + 672, + 1072 + ], + "score": 0.95, + "latex": "\\Theta _ { j } = \\frac { 1 } { N _ { \\mathrm { f u l l } } } \\sum _ { n = 1 } ^ { N _ { \\mathrm { f u l l } } } \\Theta ( x _ { n } ) f _ { j } ( x _ { n } ) ," + }, + { + "category_id": 14, + "poly": [ + 318, + 552, + 1344, + 552, + 1344, + 649, + 318, + 649 + ], + "score": 0.94, + "latex": "K _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } f _ { i } ( x _ { n } ) f _ { j } ( x _ { n } ) , \\quad L _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) g _ { j } ( y _ { n } ) , \\quad A _ { i j } = { \\frac { 1 } { N } } \\sum _ { n = 1 } ^ { N } g _ { i } ( y _ { n } ) f _ { j } ( x _ { n } ) ." + }, + { + "category_id": 13, + "poly": [ + 625, + 1525, + 875, + 1525, + 875, + 1566, + 625, + 1566 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\Theta ( x ) = \\sum _ { i = 1 } ^ { k } c _ { i } f _ { i } ( x ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 530, + 1368, + 1169, + 1368, + 1169, + 1469, + 530, + 1469 + ], + "score": 0.94, + "latex": "\\overline { { { \\Theta } } } = \\sum _ { x } p ( x | y ) \\Theta ( x ) \\approx \\sum _ { i , j = 1 } ^ { k _ { 0 } } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } ) _ { j i } \\Theta _ { j } g _ { i } ( y ) ." + }, + { + "category_id": 13, + "poly": [ + 1255, + 905, + 1392, + 905, + 1392, + 937, + 1255, + 937 + ], + "score": 0.93, + "latex": "x \\mapsto p ( x | y )" + }, + { + "category_id": 14, + "poly": [ + 673, + 457, + 1023, + 457, + 1023, + 499, + 673, + 499 + ], + "score": 0.93, + "latex": "C = k _ { 0 } - \\operatorname { T r } { \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) } ," + }, + { + "category_id": 13, + "poly": [ + 865, + 1279, + 998, + 1279, + 998, + 1313, + 865, + 1313 + ], + "score": 0.93, + "latex": "\\Theta ^ { \\prime } ( \\bar { x } ) = x ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 875, + 369, + 1073, + 369, + 1073, + 403, + 875, + 403 + ], + "score": 0.93, + "latex": "g _ { 1 } ( y ) , \\ldots , g _ { k _ { 0 } } ( y )" + }, + { + "category_id": 13, + "poly": [ + 571, + 261, + 651, + 261, + 651, + 295, + 571, + 295 + ], + "score": 0.93, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 889, + 230, + 1121, + 230, + 1121, + 264, + 889, + 264 + ], + "score": 0.92, + "latex": "( x _ { 1 } , y _ { 1 } ) , ( x _ { 1 } , y _ { 2 } ) , . . ." + }, + { + "category_id": 13, + "poly": [ + 749, + 935, + 902, + 935, + 902, + 967, + 749, + 967 + ], + "score": 0.92, + "latex": "j = 1 , \\ldots , k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1249, + 428, + 1249, + 428, + 1283, + 298, + 1283 + ], + "score": 0.92, + "latex": "\\Theta ( x ) = x ^ { a }" + }, + { + "category_id": 13, + "poly": [ + 1000, + 1526, + 1207, + 1526, + 1207, + 1566, + 1000, + 1566 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\Theta _ { j } = \\sum _ { i = 1 } ^ { k _ { 0 } } c _ { i } K _ { i j } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1613, + 388, + 1613, + 388, + 1645, + 298, + 1645 + ], + "score": 0.92, + "latex": "g _ { i } f _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1192, + 338, + 1393, + 338, + 1393, + 372, + 1192, + 372 + ], + "score": 0.92, + "latex": "f _ { 1 } ( x ) , \\ldots , f _ { k _ { 0 } } ( x )" + }, + { + "category_id": 13, + "poly": [ + 513, + 511, + 604, + 511, + 604, + 542, + 513, + 542 + ], + "score": 0.91, + "latex": "K , L , A" + }, + { + "category_id": 13, + "poly": [ + 1073, + 764, + 1124, + 764, + 1124, + 793, + 1073, + 793 + ], + "score": 0.9, + "latex": "L ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 961, + 764, + 1019, + 764, + 1019, + 793, + 961, + 793 + ], + "score": 0.9, + "latex": "K ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1284, + 873, + 1344, + 873, + 1344, + 906, + 1284, + 906 + ], + "score": 0.9, + "latex": "\\Theta ( x )" + }, + { + "category_id": 13, + "poly": [ + 489, + 2003, + 525, + 2003, + 525, + 2038, + 489, + 2038 + ], + "score": 0.89, + "latex": "\\Theta _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1307, + 1972, + 1402, + 1972, + 1402, + 2004, + 1307, + 2004 + ], + "score": 0.89, + "latex": "A , L , K" + }, + { + "category_id": 13, + "poly": [ + 1049, + 510, + 1143, + 510, + 1143, + 544, + 1049, + 544 + ], + "score": 0.89, + "latex": "( x _ { n } , y _ { n } )" + }, + { + "category_id": 13, + "poly": [ + 735, + 370, + 764, + 370, + 764, + 400, + 735, + 400 + ], + "score": 0.89, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1057, + 1082, + 1119, + 1082, + 1119, + 1112, + 1057, + 1112 + ], + "score": 0.89, + "latex": "N _ { \\mathrm { f u l l } } ," + }, + { + "category_id": 13, + "poly": [ + 1111, + 675, + 1141, + 675, + 1141, + 705, + 1111, + 705 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 579, + 1113, + 607, + 1113, + 607, + 1146, + 579, + 1146 + ], + "score": 0.88, + "latex": "f _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1102, + 1496, + 1131, + 1496, + 1131, + 1525, + 1102, + 1525 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 914, + 340, + 944, + 340, + 944, + 369, + 914, + 369 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 594, + 735, + 624, + 735, + 624, + 766, + 594, + 766 + ], + "score": 0.86, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 666, + 1115, + 695, + 1115, + 695, + 1146, + 666, + 1146 + ], + "score": 0.86, + "latex": "g _ { j }" + }, + { + "category_id": 13, + "poly": [ + 994, + 1161, + 1025, + 1161, + 1025, + 1186, + 994, + 1186 + ], + "score": 0.86, + "latex": "x ^ { a }" + }, + { + "category_id": 13, + "poly": [ + 1246, + 1187, + 1270, + 1187, + 1270, + 1216, + 1246, + 1216 + ], + "score": 0.85, + "latex": "l ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1160, + 1496, + 1184, + 1496, + 1184, + 1522, + 1160, + 1522 + ], + "score": 0.84, + "latex": "\\Theta" + }, + { + "category_id": 13, + "poly": [ + 1156, + 511, + 1311, + 511, + 1311, + 543, + 1156, + 543 + ], + "score": 0.82, + "latex": "n = 1 , \\ldots , N" + }, + { + "category_id": 13, + "poly": [ + 1223, + 1578, + 1341, + 1578, + 1341, + 1610, + 1223, + 1610 + ], + "score": 0.82, + "latex": "A A ^ { \\top }" + }, + { + "category_id": 13, + "poly": [ + 510, + 1116, + 529, + 1116, + 529, + 1144, + 510, + 1144 + ], + "score": 0.81, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 497, + 374, + 515, + 374, + 515, + 401, + 497, + 401 + ], + "score": 0.79, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 765, + 344, + 785, + 344, + 785, + 366, + 765, + 366 + ], + "score": 0.78, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 607, + 1165, + 626, + 1165, + 626, + 1186, + 607, + 1186 + ], + "score": 0.77, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 431, + 1117, + 451, + 1117, + 451, + 1139, + 431, + 1139 + ], + "score": 0.75, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 509, + 1256, + 526, + 1256, + 526, + 1277, + 509, + 1277 + ], + "score": 0.72, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 1105, + 1581, + 1208, + 1581, + 1208, + 1610, + 1105, + 1610 + ], + "score": 0.68, + "latex": "K L" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1684.0, + 561.0, + 1684.0, + 561.0, + 1728.0, + 292.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 840.0, + 1405.0, + 840.0, + 1405.0, + 875.0, + 294.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 871.0, + 1283.0, + 871.0, + 1283.0, + 909.0, + 292.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 871.0, + 1405.0, + 871.0, + 1405.0, + 909.0, + 1345.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 903.0, + 1254.0, + 903.0, + 1254.0, + 939.0, + 293.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 903.0, + 1404.0, + 903.0, + 1404.0, + 939.0, + 1393.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 933.0, + 748.0, + 933.0, + 748.0, + 969.0, + 293.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 933.0, + 1069.0, + 933.0, + 1069.0, + 969.0, + 903.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 671.0, + 1110.0, + 671.0, + 1110.0, + 709.0, + 295.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 671.0, + 1405.0, + 671.0, + 1405.0, + 709.0, + 1142.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 704.0, + 1406.0, + 704.0, + 1406.0, + 740.0, + 294.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 733.0, + 593.0, + 733.0, + 593.0, + 768.0, + 292.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 733.0, + 1406.0, + 733.0, + 1406.0, + 768.0, + 625.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 763.0, + 960.0, + 763.0, + 960.0, + 799.0, + 292.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 763.0, + 1072.0, + 763.0, + 1072.0, + 799.0, + 1020.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 763.0, + 1406.0, + 763.0, + 1406.0, + 799.0, + 1125.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 794.0, + 1120.0, + 794.0, + 1120.0, + 831.0, + 295.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1156.0, + 606.0, + 1156.0, + 606.0, + 1192.0, + 294.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1156.0, + 993.0, + 1156.0, + 993.0, + 1192.0, + 627.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 1156.0, + 1403.0, + 1156.0, + 1403.0, + 1192.0, + 1026.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1184.0, + 1245.0, + 1184.0, + 1245.0, + 1223.0, + 293.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1184.0, + 1405.0, + 1184.0, + 1405.0, + 1223.0, + 1271.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1405.0, + 1216.0, + 1405.0, + 1254.0, + 294.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1246.0, + 297.0, + 1246.0, + 297.0, + 1286.0, + 294.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1246.0, + 508.0, + 1246.0, + 508.0, + 1286.0, + 429.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1246.0, + 1405.0, + 1246.0, + 1405.0, + 1286.0, + 527.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1278.0, + 864.0, + 1278.0, + 864.0, + 1315.0, + 293.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1278.0, + 1056.0, + 1278.0, + 1056.0, + 1315.0, + 999.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1833.0, + 1405.0, + 1833.0, + 1405.0, + 1869.0, + 293.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1863.0, + 1408.0, + 1863.0, + 1408.0, + 1900.0, + 292.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1895.0, + 1404.0, + 1895.0, + 1404.0, + 1928.0, + 296.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1927.0, + 1256.0, + 1927.0, + 1256.0, + 1960.0, + 294.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1757.0, + 1403.0, + 1757.0, + 1403.0, + 1789.0, + 296.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1788.0, + 905.0, + 1788.0, + 905.0, + 1821.0, + 293.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1079.0, + 1056.0, + 1079.0, + 1056.0, + 1115.0, + 295.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1079.0, + 1405.0, + 1079.0, + 1405.0, + 1115.0, + 1120.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1113.0, + 430.0, + 1113.0, + 430.0, + 1145.0, + 296.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1113.0, + 509.0, + 1113.0, + 509.0, + 1145.0, + 452.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1113.0, + 578.0, + 1113.0, + 578.0, + 1145.0, + 530.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1113.0, + 665.0, + 1113.0, + 665.0, + 1145.0, + 608.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1113.0, + 980.0, + 1113.0, + 980.0, + 1145.0, + 696.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1492.0, + 624.0, + 1492.0, + 624.0, + 1583.0, + 284.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1492.0, + 999.0, + 1492.0, + 999.0, + 1583.0, + 876.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1492.0, + 1404.0, + 1492.0, + 1404.0, + 1583.0, + 1208.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 888.0, + 227.0, + 888.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 1122.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 255.0, + 570.0, + 255.0, + 570.0, + 299.0, + 293.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 255.0, + 663.0, + 255.0, + 663.0, + 299.0, + 652.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1575.0, + 1104.0, + 1575.0, + 1104.0, + 1619.0, + 293.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1575.0, + 1222.0, + 1575.0, + 1222.0, + 1619.0, + 1209.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1575.0, + 1407.0, + 1575.0, + 1407.0, + 1619.0, + 1342.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1609.0, + 297.0, + 1609.0, + 297.0, + 1650.0, + 288.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1609.0, + 402.0, + 1609.0, + 402.0, + 1650.0, + 389.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 306.0, + 1402.0, + 306.0, + 1402.0, + 341.0, + 296.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 335.0, + 764.0, + 335.0, + 764.0, + 375.0, + 292.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 335.0, + 913.0, + 335.0, + 913.0, + 375.0, + 786.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 335.0, + 1191.0, + 335.0, + 1191.0, + 375.0, + 945.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 335.0, + 1404.0, + 335.0, + 1404.0, + 375.0, + 1394.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 367.0, + 496.0, + 367.0, + 496.0, + 404.0, + 293.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 367.0, + 734.0, + 367.0, + 734.0, + 404.0, + 516.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 367.0, + 874.0, + 367.0, + 874.0, + 404.0, + 765.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 367.0, + 1083.0, + 367.0, + 1083.0, + 404.0, + 1074.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1306.0, + 1971.0, + 1306.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2002.0, + 488.0, + 2002.0, + 488.0, + 2038.0, + 296.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 2002.0, + 1011.0, + 2002.0, + 1011.0, + 2038.0, + 526.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1320.0, + 894.0, + 1320.0, + 894.0, + 1365.0, + 292.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 412.0, + 1258.0, + 412.0, + 1258.0, + 452.0, + 293.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 508.0, + 512.0, + 508.0, + 512.0, + 547.0, + 294.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 508.0, + 1048.0, + 508.0, + 1048.0, + 547.0, + 605.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 508.0, + 1155.0, + 508.0, + 1155.0, + 547.0, + 1144.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 508.0, + 1358.0, + 508.0, + 1358.0, + 547.0, + 1312.0, + 547.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1772, + 1402, + 1772, + 1402, + 1957, + 298, + 1957 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 464, + 1403, + 464, + 1403, + 617, + 298, + 617 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1478, + 1404, + 1478, + 1404, + 1633, + 299, + 1633 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 710, + 1403, + 710, + 1403, + 864, + 299, + 864 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1017, + 1403, + 1017, + 1403, + 1139, + 299, + 1139 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1310, + 1403, + 1310, + 1403, + 1463, + 299, + 1463 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 879, + 1403, + 879, + 1403, + 1001, + 298, + 1001 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 301, + 633, + 1399, + 633, + 1399, + 695, + 301, + 695 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 297, + 308, + 1399, + 308, + 1399, + 371, + 297, + 371 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1400, + 1972, + 1400, + 2034, + 298, + 2034 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 300, + 386, + 1398, + 386, + 1398, + 448, + 300, + 448 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 302, + 1694, + 1397, + 1694, + 1397, + 1758, + 302, + 1758 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 302, + 1647, + 1259, + 1647, + 1259, + 1679, + 302, + 1679 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 300, + 1232, + 647, + 1232, + 647, + 1263, + 300, + 1263 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 780, + 229, + 780, + 261, + 300, + 261 + ], + "score": 0.794 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.764 + }, + { + "category_id": 13, + "poly": [ + 901, + 1479, + 1021, + 1479, + 1021, + 1513, + 901, + 1513 + ], + "score": 0.94, + "latex": "\\{ 1 , \\ldots , k \\}" + }, + { + "category_id": 13, + "poly": [ + 897, + 801, + 1017, + 801, + 1017, + 835, + 897, + 835 + ], + "score": 0.92, + "latex": "x _ { i } \\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1572, + 1346, + 1572, + 1346, + 1605, + 1235, + 1605 + ], + "score": 0.92, + "latex": "\\Theta ( y ) = y" + }, + { + "category_id": 13, + "poly": [ + 515, + 800, + 810, + 800, + 810, + 837, + 515, + 837 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { d _ { l ^ { 2 } } ^ { 2 } ( x , \\overline { { y } } ) = \\sum _ { i } ( x _ { i } - y _ { i } ) ^ { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1144, + 1541, + 1260, + 1541, + 1260, + 1572, + 1144, + 1572 + ], + "score": 0.91, + "latex": "f _ { 1 } , \\ldots , f _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1310, + 1510, + 1405, + 1510, + 1405, + 1543, + 1310, + 1543 + ], + "score": 0.91, + "latex": "g _ { i } ( j ) =" + }, + { + "category_id": 13, + "poly": [ + 298, + 1541, + 331, + 1541, + 331, + 1575, + 298, + 1575 + ], + "score": 0.88, + "latex": "\\delta _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 1137, + 772, + 1173, + 772, + 1173, + 803, + 1137, + 803 + ], + "score": 0.88, + "latex": "d _ { l ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 698, + 1572, + 722, + 1572, + 722, + 1604, + 698, + 1604 + ], + "score": 0.87, + "latex": "f _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1227, + 800, + 1262, + 800, + 1262, + 830, + 1227, + 830 + ], + "score": 0.85, + "latex": "\\mathrm { i } ^ { t h }" + }, + { + "category_id": 13, + "poly": [ + 1299, + 1543, + 1326, + 1543, + 1326, + 1568, + 1299, + 1568 + ], + "score": 0.83, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 832, + 1605, + 851, + 1605, + 851, + 1634, + 832, + 1634 + ], + "score": 0.82, + "latex": "\\overline { y }" + }, + { + "category_id": 13, + "poly": [ + 593, + 742, + 621, + 742, + 621, + 768, + 593, + 768 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1015, + 1542, + 1034, + 1542, + 1034, + 1568, + 1015, + 1568 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 298, + 341, + 325, + 341, + 325, + 367, + 298, + 367 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 760, + 1898, + 776, + 1898, + 776, + 1922, + 760, + 1922 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1376, + 739, + 1400, + 739, + 1400, + 768, + 1376, + 768 + ], + "score": 0.81, + "latex": "l ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 442, + 1512, + 461, + 1512, + 461, + 1538, + 442, + 1538 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 695, + 741, + 721, + 741, + 721, + 768, + 695, + 768 + ], + "score": 0.77, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 476, + 1480, + 502, + 1480, + 502, + 1507, + 476, + 1507 + ], + "score": 0.75, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 297, + 1927, + 316, + 1927, + 316, + 1953, + 297, + 1953 + ], + "score": 0.72, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 376, + 340, + 401, + 340, + 401, + 367, + 376, + 367 + ], + "score": 0.45, + "latex": "Y" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1231.0, + 650.0, + 1231.0, + 650.0, + 1267.0, + 295.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 230.0, + 786.0, + 230.0, + 786.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1774.0, + 1403.0, + 1774.0, + 1403.0, + 1806.0, + 297.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1404.0, + 1803.0, + 1404.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 1403.0, + 1835.0, + 1403.0, + 1867.0, + 294.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1861.0, + 1403.0, + 1861.0, + 1403.0, + 1903.0, + 291.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1892.0, + 759.0, + 1892.0, + 759.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1892.0, + 1406.0, + 1892.0, + 1406.0, + 1929.0, + 777.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1923.0, + 296.0, + 1923.0, + 296.0, + 1961.0, + 293.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1923.0, + 1315.0, + 1923.0, + 1315.0, + 1961.0, + 317.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 463.0, + 1405.0, + 463.0, + 1405.0, + 500.0, + 296.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 493.0, + 1404.0, + 493.0, + 1404.0, + 529.0, + 294.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 526.0, + 1404.0, + 526.0, + 1404.0, + 559.0, + 296.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 552.0, + 1405.0, + 552.0, + 1405.0, + 591.0, + 293.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 583.0, + 918.0, + 583.0, + 918.0, + 622.0, + 295.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1477.0, + 475.0, + 1477.0, + 475.0, + 1514.0, + 294.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1477.0, + 900.0, + 1477.0, + 900.0, + 1514.0, + 503.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 1477.0, + 1406.0, + 1477.0, + 1406.0, + 1514.0, + 1022.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1508.0, + 441.0, + 1508.0, + 441.0, + 1549.0, + 293.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1508.0, + 1309.0, + 1508.0, + 1309.0, + 1549.0, + 462.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1540.0, + 297.0, + 1540.0, + 297.0, + 1574.0, + 292.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1540.0, + 1014.0, + 1540.0, + 1014.0, + 1574.0, + 332.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1540.0, + 1143.0, + 1540.0, + 1143.0, + 1574.0, + 1035.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 1540.0, + 1298.0, + 1540.0, + 1298.0, + 1574.0, + 1261.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 1540.0, + 1406.0, + 1540.0, + 1406.0, + 1574.0, + 1327.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1569.0, + 697.0, + 1569.0, + 697.0, + 1606.0, + 293.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1569.0, + 1234.0, + 1569.0, + 1234.0, + 1606.0, + 723.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 1569.0, + 1406.0, + 1569.0, + 1406.0, + 1606.0, + 1347.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1602.0, + 831.0, + 1602.0, + 831.0, + 1635.0, + 294.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1602.0, + 1214.0, + 1602.0, + 1214.0, + 1635.0, + 852.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 709.0, + 1404.0, + 709.0, + 1404.0, + 746.0, + 297.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 738.0, + 592.0, + 738.0, + 592.0, + 772.0, + 294.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 738.0, + 694.0, + 738.0, + 694.0, + 772.0, + 622.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 738.0, + 1375.0, + 738.0, + 1375.0, + 772.0, + 722.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 738.0, + 1404.0, + 738.0, + 1404.0, + 772.0, + 1401.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 771.0, + 1136.0, + 771.0, + 1136.0, + 804.0, + 297.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 771.0, + 1404.0, + 771.0, + 1404.0, + 804.0, + 1174.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 798.0, + 514.0, + 798.0, + 514.0, + 839.0, + 293.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 798.0, + 896.0, + 798.0, + 896.0, + 839.0, + 811.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 798.0, + 1226.0, + 798.0, + 1226.0, + 839.0, + 1018.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 798.0, + 1407.0, + 798.0, + 1407.0, + 839.0, + 1263.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 832.0, + 866.0, + 832.0, + 866.0, + 865.0, + 295.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1017.0, + 1404.0, + 1017.0, + 1404.0, + 1050.0, + 294.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1049.0, + 1405.0, + 1049.0, + 1405.0, + 1081.0, + 295.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1079.0, + 1404.0, + 1079.0, + 1404.0, + 1112.0, + 295.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1108.0, + 917.0, + 1108.0, + 917.0, + 1143.0, + 294.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1310.0, + 1403.0, + 1310.0, + 1403.0, + 1343.0, + 295.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1340.0, + 1404.0, + 1340.0, + 1404.0, + 1376.0, + 295.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1369.0, + 1407.0, + 1369.0, + 1407.0, + 1408.0, + 293.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1401.0, + 1407.0, + 1401.0, + 1407.0, + 1441.0, + 293.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1430.0, + 561.0, + 1430.0, + 561.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 880.0, + 1405.0, + 880.0, + 1405.0, + 912.0, + 296.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 910.0, + 1405.0, + 910.0, + 1405.0, + 942.0, + 293.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 940.0, + 1407.0, + 940.0, + 1407.0, + 975.0, + 292.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 968.0, + 972.0, + 968.0, + 972.0, + 1006.0, + 292.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 630.0, + 1405.0, + 630.0, + 1405.0, + 669.0, + 296.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 660.0, + 924.0, + 660.0, + 924.0, + 698.0, + 295.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 310.0, + 1403.0, + 310.0, + 1403.0, + 342.0, + 296.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 336.0, + 297.0, + 336.0, + 297.0, + 378.0, + 293.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 336.0, + 375.0, + 336.0, + 375.0, + 378.0, + 326.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 336.0, + 1321.0, + 336.0, + 1321.0, + 378.0, + 402.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2002.0, + 1335.0, + 2002.0, + 1335.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 384.0, + 1402.0, + 384.0, + 1402.0, + 420.0, + 296.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 417.0, + 1046.0, + 417.0, + 1046.0, + 452.0, + 296.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1694.0, + 1402.0, + 1694.0, + 1402.0, + 1730.0, + 298.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1726.0, + 757.0, + 1726.0, + 757.0, + 1762.0, + 297.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1647.0, + 1263.0, + 1647.0, + 1263.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 4, + "poly": [ + 295, + 1721, + 1406, + 1721, + 1406, + 1999, + 295, + 1999 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 342, + 1257, + 1392, + 1257, + 1392, + 1682, + 342, + 1682 + ], + "score": 0.962 + }, + { + "category_id": 4, + "poly": [ + 295, + 961, + 1406, + 961, + 1406, + 1177, + 295, + 1177 + ], + "score": 0.95 + }, + { + "category_id": 3, + "poly": [ + 312, + 258, + 1416, + 258, + 1416, + 922, + 312, + 922 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.659 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.258 + }, + { + "category_id": 13, + "poly": [ + 746, + 994, + 858, + 994, + 858, + 1024, + 746, + 1024 + ], + "score": 0.92, + "latex": "k _ { 0 } = 2 0 0" + }, + { + "category_id": 13, + "poly": [ + 1053, + 1116, + 1174, + 1116, + 1174, + 1147, + 1053, + 1147 + ], + "score": 0.91, + "latex": "k _ { 0 } ~ = ~ 2 0 0" + }, + { + "category_id": 13, + "poly": [ + 831, + 1147, + 861, + 1147, + 861, + 1176, + 831, + 1176 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 686, + 1259, + 729, + 1259, + 729, + 1290, + 686, + 1290 + ], + "score": 0.61, + "latex": "( \\% )" + }, + { + "category_id": 13, + "poly": [ + 1257, + 1259, + 1302, + 1259, + 1302, + 1290, + 1257, + 1290 + ], + "score": 0.54, + "latex": "( \\% )" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1720.0, + 1406.0, + 1720.0, + 1406.0, + 1758.0, + 293.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1752.0, + 1404.0, + 1752.0, + 1404.0, + 1788.0, + 293.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1781.0, + 1406.0, + 1781.0, + 1406.0, + 1819.0, + 293.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1810.0, + 1404.0, + 1810.0, + 1404.0, + 1851.0, + 292.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1844.0, + 1404.0, + 1844.0, + 1404.0, + 1878.0, + 292.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1876.0, + 1404.0, + 1876.0, + 1404.0, + 1909.0, + 295.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1905.0, + 1403.0, + 1905.0, + 1403.0, + 1937.0, + 295.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1935.0, + 1406.0, + 1935.0, + 1406.0, + 1972.0, + 294.0, + 1972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1962.0, + 401.0, + 1962.0, + 401.0, + 2006.0, + 292.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1255.0, + 685.0, + 1255.0, + 685.0, + 1293.0, + 427.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1255.0, + 733.0, + 1255.0, + 733.0, + 1293.0, + 730.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1255.0, + 1256.0, + 1255.0, + 1256.0, + 1293.0, + 978.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1293.0, + 395.0, + 1293.0, + 395.0, + 1332.0, + 340.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1295.0, + 951.0, + 1295.0, + 951.0, + 1330.0, + 905.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1322.0, + 716.0, + 1322.0, + 716.0, + 1364.0, + 642.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1323.0, + 1236.0, + 1323.0, + 1236.0, + 1363.0, + 1147.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1349.0, + 805.0, + 1349.0, + 805.0, + 1387.0, + 649.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1348.0, + 1259.0, + 1348.0, + 1259.0, + 1387.0, + 1152.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1365.0, + 394.0, + 1365.0, + 394.0, + 1405.0, + 341.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1375.0, + 739.0, + 1375.0, + 739.0, + 1413.0, + 644.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1365.0, + 946.0, + 1365.0, + 946.0, + 1404.0, + 904.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1373.0, + 1353.0, + 1373.0, + 1353.0, + 1411.0, + 1180.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1403.0, + 1375.0, + 1403.0, + 1375.0, + 1437.0, + 1186.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1438.0, + 400.0, + 1438.0, + 400.0, + 1477.0, + 341.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1438.0, + 951.0, + 1438.0, + 951.0, + 1474.0, + 904.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1452.0, + 1268.0, + 1452.0, + 1268.0, + 1467.0, + 1253.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1511.0, + 396.0, + 1511.0, + 396.0, + 1550.0, + 340.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1511.0, + 954.0, + 1511.0, + 954.0, + 1550.0, + 904.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1600.0, + 491.0, + 1600.0, + 491.0, + 1634.0, + 453.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1600.0, + 578.0, + 1600.0, + 578.0, + 1636.0, + 535.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1601.0, + 662.0, + 1601.0, + 662.0, + 1634.0, + 621.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1600.0, + 750.0, + 1600.0, + 750.0, + 1635.0, + 705.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1600.0, + 831.0, + 1600.0, + 831.0, + 1632.0, + 792.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1601.0, + 1047.0, + 1601.0, + 1047.0, + 1634.0, + 1010.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1600.0, + 1135.0, + 1600.0, + 1135.0, + 1636.0, + 1093.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1601.0, + 1220.0, + 1601.0, + 1220.0, + 1634.0, + 1180.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1600.0, + 1309.0, + 1600.0, + 1309.0, + 1635.0, + 1265.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1601.0, + 1393.0, + 1601.0, + 1393.0, + 1632.0, + 1353.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1648.0, + 619.0, + 1648.0, + 619.0, + 1688.0, + 538.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 1648.0, + 1181.0, + 1648.0, + 1181.0, + 1688.0, + 1099.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1425.0, + 1030.0, + 1425.0, + 1030.0, + 1450.0, + 981.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 1581.5, + 397.0, + 1581.5, + 397.0, + 1630.5, + 345.0, + 1630.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.75, + 1588.0, + 961.75, + 1588.0, + 961.75, + 1630.0, + 908.75, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 960.0, + 1404.0, + 960.0, + 1404.0, + 997.0, + 294.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 995.0, + 745.0, + 995.0, + 745.0, + 1026.0, + 295.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 995.0, + 1403.0, + 995.0, + 1403.0, + 1026.0, + 859.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1023.0, + 1404.0, + 1023.0, + 1404.0, + 1057.0, + 293.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1054.0, + 1406.0, + 1054.0, + 1406.0, + 1089.0, + 294.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1083.0, + 1403.0, + 1083.0, + 1403.0, + 1120.0, + 293.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1116.0, + 1052.0, + 1116.0, + 1052.0, + 1150.0, + 293.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1116.0, + 1404.0, + 1116.0, + 1404.0, + 1150.0, + 1175.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1144.0, + 830.0, + 1144.0, + 830.0, + 1183.0, + 294.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1144.0, + 1210.0, + 1144.0, + 1210.0, + 1183.0, + 862.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 307.0, + 1367.0, + 307.0, + 1367.0, + 340.0, + 1265.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 332.0, + 1397.0, + 332.0, + 1397.0, + 365.0, + 1250.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 364.0, + 1062.0, + 364.0, + 1062.0, + 404.0, + 1017.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 429.0, + 1065.0, + 429.0, + 1065.0, + 471.0, + 1001.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 485.0, + 1062.0, + 485.0, + 1062.0, + 518.0, + 997.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 531.0, + 1088.0, + 531.0, + 1088.0, + 551.0, + 1073.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 526.0, + 1201.0, + 526.0, + 1201.0, + 559.0, + 1169.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 525.0, + 1312.0, + 525.0, + 1312.0, + 559.0, + 1265.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 525.0, + 1416.0, + 525.0, + 1416.0, + 557.0, + 1369.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 557.0, + 1267.0, + 557.0, + 1267.0, + 595.0, + 1189.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 621.0, + 1063.0, + 621.0, + 1063.0, + 654.0, + 1021.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 632.0, + 1397.0, + 632.0, + 1397.0, + 662.0, + 1310.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 668.0, + 1063.0, + 668.0, + 1063.0, + 709.0, + 1020.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 657.0, + 1398.0, + 657.0, + 1398.0, + 711.0, + 1311.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 721.0, + 1066.0, + 721.0, + 1066.0, + 761.0, + 1020.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 771.0, + 1065.0, + 771.0, + 1065.0, + 811.0, + 1020.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 824.0, + 1063.0, + 824.0, + 1063.0, + 856.0, + 1021.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 855.0, + 1090.0, + 855.0, + 1090.0, + 876.0, + 1071.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 850.0, + 1162.0, + 850.0, + 1162.0, + 880.0, + 1127.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 845.0, + 1413.0, + 845.0, + 1413.0, + 882.0, + 1173.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 887.0, + 1310.0, + 887.0, + 1310.0, + 927.0, + 1147.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 314, + 223, + 1406, + 223, + 1406, + 650, + 314, + 650 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1479, + 1403, + 1479, + 1403, + 1571, + 299, + 1571 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1341, + 1403, + 1341, + 1403, + 1463, + 299, + 1463 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1726, + 1402, + 1726, + 1402, + 1849, + 299, + 1849 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1587, + 1403, + 1587, + 1403, + 1711, + 299, + 1711 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1864, + 1402, + 1864, + 1402, + 1957, + 299, + 1957 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1202, + 1403, + 1202, + 1403, + 1326, + 299, + 1326 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1014, + 1404, + 1014, + 1404, + 1107, + 299, + 1107 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 296, + 712, + 1405, + 712, + 1405, + 958, + 296, + 958 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1398, + 1972, + 1398, + 2033, + 299, + 2033 + ], + "score": 0.944 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 298, + 1147, + 875, + 1147, + 875, + 1176, + 298, + 1176 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.788 + }, + { + "category_id": 13, + "poly": [ + 1087, + 1293, + 1253, + 1293, + 1253, + 1328, + 1087, + 1328 + ], + "score": 0.92, + "latex": "\\{ 1 , x , x ^ { 2 } , \\bar { . . . } \\bar \\} \\}" + }, + { + "category_id": 13, + "poly": [ + 702, + 1203, + 783, + 1203, + 783, + 1236, + 702, + 1236 + ], + "score": 0.92, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 793, + 1235, + 821, + 1235, + 821, + 1261, + 793, + 1261 + ], + "score": 0.84, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1105, + 1235, + 1134, + 1235, + 1134, + 1261, + 1105, + 1261 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 599, + 840, + 637, + 840, + 637, + 869, + 599, + 869 + ], + "score": 0.82, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1265, + 1401, + 1265, + 1401, + 1291, + 1373, + 1291 + ], + "score": 0.81, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 887, + 1620, + 916, + 1620, + 916, + 1646, + 887, + 1646 + ], + "score": 0.79, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 297, + 1619, + 323, + 1619, + 323, + 1646, + 297, + 1646 + ], + "score": 0.77, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 366, + 1588, + 395, + 1588, + 395, + 1615, + 366, + 1615 + ], + "score": 0.77, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1372, + 715, + 1401, + 715, + 1401, + 742, + 1372, + 742 + ], + "score": 0.74, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 961, + 1239, + 982, + 1239, + 982, + 1261, + 961, + 1261 + ], + "score": 0.71, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 856, + 716, + 885, + 716, + 885, + 743, + 856, + 743 + ], + "score": 0.67, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1280, + 746, + 1306, + 746, + 1306, + 773, + 1280, + 773 + ], + "score": 0.66, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 471, + 1239, + 491, + 1239, + 491, + 1261, + 471, + 1261 + ], + "score": 0.66, + "latex": "n" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 208.0, + 856.0, + 208.0, + 856.0, + 351.0, + 306.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 211.0, + 1415.0, + 211.0, + 1415.0, + 347.0, + 861.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 324.0, + 837.0, + 324.0, + 837.0, + 440.0, + 315.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 322.0, + 1408.0, + 322.0, + 1408.0, + 444.0, + 877.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 434.0, + 843.0, + 434.0, + 843.0, + 545.0, + 326.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 426.0, + 1408.0, + 426.0, + 1408.0, + 550.0, + 877.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 537.0, + 858.0, + 537.0, + 858.0, + 655.0, + 313.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 531.0, + 1414.0, + 531.0, + 1414.0, + 661.0, + 859.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 714.0, + 855.0, + 714.0, + 855.0, + 747.0, + 295.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 714.0, + 1371.0, + 714.0, + 1371.0, + 747.0, + 886.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 745.0, + 1279.0, + 745.0, + 1279.0, + 779.0, + 294.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 745.0, + 1405.0, + 745.0, + 1405.0, + 779.0, + 1307.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 776.0, + 1405.0, + 776.0, + 1405.0, + 810.0, + 295.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 806.0, + 1405.0, + 806.0, + 1405.0, + 840.0, + 295.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 837.0, + 598.0, + 837.0, + 598.0, + 871.0, + 295.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 837.0, + 1405.0, + 837.0, + 1405.0, + 871.0, + 638.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 864.0, + 1405.0, + 864.0, + 1405.0, + 902.0, + 291.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 895.0, + 1405.0, + 895.0, + 1405.0, + 931.0, + 293.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 924.0, + 401.0, + 924.0, + 401.0, + 963.0, + 293.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1145.0, + 878.0, + 1145.0, + 878.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1477.0, + 1405.0, + 1477.0, + 1405.0, + 1514.0, + 294.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1511.0, + 1405.0, + 1511.0, + 1405.0, + 1544.0, + 295.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1542.0, + 1004.0, + 1542.0, + 1004.0, + 1575.0, + 295.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1341.0, + 1405.0, + 1341.0, + 1405.0, + 1375.0, + 293.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1373.0, + 1407.0, + 1373.0, + 1407.0, + 1405.0, + 294.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1404.0, + 1405.0, + 1404.0, + 1405.0, + 1437.0, + 294.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1433.0, + 1019.0, + 1433.0, + 1019.0, + 1466.0, + 297.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1725.0, + 1406.0, + 1725.0, + 1406.0, + 1761.0, + 294.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1757.0, + 1406.0, + 1757.0, + 1406.0, + 1790.0, + 295.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1785.0, + 1406.0, + 1785.0, + 1406.0, + 1824.0, + 292.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1815.0, + 381.0, + 1815.0, + 381.0, + 1855.0, + 290.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1587.0, + 365.0, + 1587.0, + 365.0, + 1620.0, + 294.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1587.0, + 1405.0, + 1587.0, + 1405.0, + 1620.0, + 396.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1615.0, + 296.0, + 1615.0, + 296.0, + 1654.0, + 292.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1615.0, + 886.0, + 1615.0, + 886.0, + 1654.0, + 324.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1615.0, + 1407.0, + 1615.0, + 1407.0, + 1654.0, + 917.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1649.0, + 1404.0, + 1649.0, + 1404.0, + 1682.0, + 294.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1678.0, + 1326.0, + 1678.0, + 1326.0, + 1713.0, + 294.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1863.0, + 1404.0, + 1863.0, + 1404.0, + 1897.0, + 295.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1892.0, + 1403.0, + 1892.0, + 1403.0, + 1931.0, + 292.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1921.0, + 886.0, + 1921.0, + 886.0, + 1964.0, + 293.0, + 1964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1202.0, + 701.0, + 1202.0, + 701.0, + 1238.0, + 294.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1202.0, + 1407.0, + 1202.0, + 1407.0, + 1238.0, + 784.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1234.0, + 470.0, + 1234.0, + 470.0, + 1266.0, + 294.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1234.0, + 792.0, + 1234.0, + 792.0, + 1266.0, + 492.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1234.0, + 960.0, + 1234.0, + 960.0, + 1266.0, + 822.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1234.0, + 1104.0, + 1234.0, + 1104.0, + 1266.0, + 983.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1234.0, + 1407.0, + 1234.0, + 1407.0, + 1266.0, + 1135.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1262.0, + 1372.0, + 1262.0, + 1372.0, + 1297.0, + 293.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1293.0, + 1086.0, + 1293.0, + 1086.0, + 1329.0, + 294.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1293.0, + 1267.0, + 1293.0, + 1267.0, + 1329.0, + 1254.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1013.0, + 1405.0, + 1013.0, + 1405.0, + 1051.0, + 294.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1046.0, + 1404.0, + 1046.0, + 1404.0, + 1080.0, + 294.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1074.0, + 1119.0, + 1074.0, + 1119.0, + 1115.0, + 293.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 1403.0, + 2001.0, + 1403.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1326, + 1404, + 1326, + 1404, + 1541, + 298, + 1541 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1557, + 1403, + 1557, + 1403, + 1680, + 300, + 1680 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 1881, + 1402, + 1881, + 1402, + 2033, + 300, + 2033 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 890, + 1403, + 890, + 1403, + 1046, + 299, + 1046 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1060, + 1403, + 1060, + 1403, + 1182, + 299, + 1182 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 324, + 219, + 1379, + 219, + 1379, + 627, + 324, + 627 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 300, + 1773, + 1399, + 1773, + 1399, + 1865, + 300, + 1865 + ], + "score": 0.963 + }, + { + "category_id": 4, + "poly": [ + 295, + 665, + 1405, + 665, + 1405, + 819, + 295, + 819 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 300, + 1695, + 1399, + 1695, + 1399, + 1758, + 300, + 1758 + ], + "score": 0.948 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.859 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.786 + }, + { + "category_id": 0, + "poly": [ + 297, + 1241, + 987, + 1241, + 987, + 1270, + 297, + 1270 + ], + "score": 0.757 + }, + { + "category_id": 1, + "poly": [ + 297, + 1241, + 987, + 1241, + 987, + 1270, + 297, + 1270 + ], + "score": 0.137 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 856, + 76, + 856, + 104, + 300, + 104 + ], + "score": 0.094 + }, + { + "category_id": 13, + "poly": [ + 769, + 1559, + 798, + 1559, + 798, + 1585, + 769, + 1585 + ], + "score": 0.8, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 477, + 698, + 496, + 698, + 496, + 724, + 477, + 724 + ], + "score": 0.78, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1116, + 668, + 1136, + 668, + 1136, + 694, + 1116, + 694 + ], + "score": 0.75, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 733, + 1481, + 761, + 1481, + 761, + 1507, + 733, + 1507 + ], + "score": 0.74, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 352, + 1649, + 382, + 1649, + 382, + 1676, + 352, + 1676 + ], + "score": 0.74, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1337, + 1558, + 1364, + 1558, + 1364, + 1585, + 1337, + 1585 + ], + "score": 0.56, + "latex": "Y" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 219.0, + 779.0, + 219.0, + 779.0, + 259.0, + 323.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 243.0, + 1156.0, + 243.0, + 1156.0, + 401.0, + 1096.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 245.0, + 1122.0, + 245.0, + 1122.0, + 401.0, + 1067.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 246.0, + 1317.0, + 246.0, + 1317.0, + 363.0, + 1268.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 247.0, + 1086.0, + 247.0, + 1086.0, + 374.0, + 1034.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 247.0, + 1350.0, + 247.0, + 1350.0, + 368.0, + 1305.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 248.0, + 889.0, + 248.0, + 889.0, + 363.0, + 835.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 248.0, + 1050.0, + 248.0, + 1050.0, + 433.0, + 1007.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 248.0, + 1248.0, + 248.0, + 1248.0, + 396.0, + 1201.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 250.0, + 953.0, + 250.0, + 953.0, + 333.0, + 907.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 250.0, + 1283.0, + 250.0, + 1283.0, + 363.0, + 1239.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 251.0, + 918.0, + 251.0, + 918.0, + 334.0, + 876.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 251.0, + 983.0, + 251.0, + 983.0, + 328.0, + 939.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 251.0, + 1182.0, + 251.0, + 1182.0, + 396.0, + 1139.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 251.0, + 1374.0, + 251.0, + 1374.0, + 330.0, + 1338.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 257.0, + 850.0, + 257.0, + 850.0, + 294.0, + 814.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 286.0, + 848.0, + 286.0, + 848.0, + 357.0, + 814.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 316.0, + 980.0, + 316.0, + 980.0, + 358.0, + 878.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 317.0, + 1376.0, + 317.0, + 1376.0, + 358.0, + 1344.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 352.0, + 879.0, + 352.0, + 879.0, + 391.0, + 814.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 349.0, + 984.0, + 349.0, + 984.0, + 396.0, + 883.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 351.0, + 1081.0, + 351.0, + 1081.0, + 392.0, + 1042.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 352.0, + 1374.0, + 352.0, + 1374.0, + 394.0, + 1240.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 383.0, + 980.0, + 383.0, + 980.0, + 495.0, + 815.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 388.0, + 1176.0, + 388.0, + 1176.0, + 427.0, + 1043.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 385.0, + 1309.0, + 385.0, + 1309.0, + 463.0, + 1211.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 385.0, + 1375.0, + 385.0, + 1375.0, + 424.0, + 1310.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 421.0, + 847.0, + 421.0, + 847.0, + 461.0, + 814.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 421.0, + 880.0, + 421.0, + 880.0, + 460.0, + 848.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 414.0, + 980.0, + 414.0, + 980.0, + 457.0, + 906.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 412.0, + 1115.0, + 412.0, + 1115.0, + 500.0, + 1009.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 419.0, + 1175.0, + 419.0, + 1175.0, + 456.0, + 1139.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 418.0, + 1375.0, + 418.0, + 1375.0, + 457.0, + 1239.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 450.0, + 882.0, + 450.0, + 882.0, + 489.0, + 813.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 451.0, + 979.0, + 451.0, + 979.0, + 490.0, + 907.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 446.0, + 1122.0, + 446.0, + 1122.0, + 494.0, + 1034.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 454.0, + 1176.0, + 454.0, + 1176.0, + 491.0, + 1140.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 451.0, + 1250.0, + 451.0, + 1250.0, + 493.0, + 1209.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 450.0, + 1375.0, + 450.0, + 1375.0, + 493.0, + 1273.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 487.0, + 979.0, + 487.0, + 979.0, + 524.0, + 814.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 477.0, + 1178.0, + 477.0, + 1178.0, + 596.0, + 1013.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 479.0, + 1379.0, + 479.0, + 1379.0, + 594.0, + 1206.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 518.0, + 978.0, + 518.0, + 978.0, + 562.0, + 813.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 512.0, + 1058.0, + 512.0, + 1058.0, + 593.0, + 1007.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 518.0, + 1117.0, + 518.0, + 1117.0, + 559.0, + 1076.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 513.0, + 1182.0, + 513.0, + 1182.0, + 592.0, + 1135.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 516.0, + 1312.0, + 516.0, + 1312.0, + 562.0, + 1238.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 540.0, + 717.0, + 540.0, + 717.0, + 588.0, + 369.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 545.0, + 979.0, + 545.0, + 979.0, + 590.0, + 814.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 551.0, + 1114.0, + 551.0, + 1114.0, + 589.0, + 1038.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 550.0, + 1313.0, + 550.0, + 1313.0, + 589.0, + 1268.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 590.0, + 642.0, + 590.0, + 642.0, + 629.0, + 463.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 589.0, + 910.0, + 589.0, + 910.0, + 627.0, + 883.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 587.0, + 1110.0, + 587.0, + 1110.0, + 628.0, + 1079.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 584.0, + 1315.0, + 584.0, + 1315.0, + 633.0, + 1271.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 273.5, + 400.0, + 273.5, + 400.0, + 321.5, + 313.0, + 321.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.75, + 315.5, + 398.75, + 315.5, + 398.75, + 363.5, + 313.75, + 363.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 358.5, + 400.0, + 358.5, + 400.0, + 406.5, + 313.0, + 406.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.75, + 399.5, + 402.75, + 399.5, + 402.75, + 451.5, + 310.75, + 451.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 442.0, + 401.0, + 442.0, + 401.0, + 493.5, + 311.0, + 493.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.75, + 449.5, + 1289.75, + 449.5, + 1289.75, + 490.5, + 1233.75, + 490.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.75, + 459.5, + 573.75, + 459.5, + 573.75, + 474.5, + 557.75, + 474.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 492.0, + 620.0, + 492.0, + 620.0, + 509.5, + 583.0, + 509.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 484.5, + 400.0, + 484.5, + 400.0, + 533.5, + 313.0, + 533.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 521.0, + 1089.0, + 521.0, + 1089.0, + 556.0, + 1030.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.25, + 554.0, + 1262.25, + 554.0, + 1262.25, + 586.5, + 1256.25, + 586.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 665.0, + 1115.0, + 665.0, + 1115.0, + 701.0, + 294.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 665.0, + 1403.0, + 665.0, + 1403.0, + 701.0, + 1137.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 694.0, + 476.0, + 694.0, + 476.0, + 733.0, + 292.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 694.0, + 1405.0, + 694.0, + 1405.0, + 733.0, + 497.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 725.0, + 1405.0, + 725.0, + 1405.0, + 763.0, + 293.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 757.0, + 1407.0, + 757.0, + 1407.0, + 794.0, + 293.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 786.0, + 568.0, + 786.0, + 568.0, + 823.0, + 294.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1240.0, + 991.0, + 1240.0, + 991.0, + 1272.0, + 296.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1324.0, + 1404.0, + 1324.0, + 1404.0, + 1363.0, + 292.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1357.0, + 1406.0, + 1357.0, + 1406.0, + 1395.0, + 293.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1388.0, + 1406.0, + 1388.0, + 1406.0, + 1422.0, + 293.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1452.0, + 293.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1449.0, + 1404.0, + 1449.0, + 1404.0, + 1484.0, + 294.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1480.0, + 732.0, + 1480.0, + 732.0, + 1515.0, + 294.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1480.0, + 1406.0, + 1480.0, + 1406.0, + 1515.0, + 762.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1512.0, + 1342.0, + 1512.0, + 1342.0, + 1543.0, + 296.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1555.0, + 768.0, + 1555.0, + 768.0, + 1593.0, + 294.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1555.0, + 1336.0, + 1555.0, + 1336.0, + 1593.0, + 799.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 1555.0, + 1404.0, + 1555.0, + 1404.0, + 1593.0, + 1365.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1587.0, + 1407.0, + 1587.0, + 1407.0, + 1623.0, + 294.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1617.0, + 1407.0, + 1617.0, + 1407.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1649.0, + 351.0, + 1649.0, + 351.0, + 1685.0, + 295.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1649.0, + 1266.0, + 1649.0, + 1266.0, + 1685.0, + 383.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1915.0, + 298.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1911.0, + 1403.0, + 1911.0, + 1403.0, + 1944.0, + 298.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 910.0, + 2003.0, + 910.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 891.0, + 1404.0, + 891.0, + 1404.0, + 924.0, + 295.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 920.0, + 1405.0, + 920.0, + 1405.0, + 957.0, + 293.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 951.0, + 1405.0, + 951.0, + 1405.0, + 988.0, + 293.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 984.0, + 1404.0, + 984.0, + 1404.0, + 1017.0, + 295.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1013.0, + 578.0, + 1013.0, + 578.0, + 1049.0, + 294.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1061.0, + 1402.0, + 1061.0, + 1402.0, + 1093.0, + 295.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1091.0, + 1404.0, + 1091.0, + 1404.0, + 1123.0, + 294.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1123.0, + 1404.0, + 1123.0, + 1404.0, + 1156.0, + 294.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1153.0, + 777.0, + 1153.0, + 777.0, + 1182.0, + 297.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1772.0, + 1405.0, + 1772.0, + 1405.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1839.0, + 292.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1835.0, + 1048.0, + 1835.0, + 1048.0, + 1868.0, + 295.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1695.0, + 1404.0, + 1695.0, + 1404.0, + 1730.0, + 296.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1727.0, + 1232.0, + 1727.0, + 1232.0, + 1762.0, + 295.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1240.0, + 991.0, + 1240.0, + 991.0, + 1272.0, + 296.0, + 1272.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 857, + 1402, + 857, + 1402, + 1012, + 299, + 1012 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1195, + 1403, + 1195, + 1403, + 1378, + 298, + 1378 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1027, + 1403, + 1027, + 1403, + 1180, + 299, + 1180 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 718, + 1402, + 718, + 1402, + 842, + 299, + 842 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1395, + 1402, + 1395, + 1402, + 1517, + 300, + 1517 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 338, + 1402, + 338, + 1402, + 491, + 299, + 491 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1532, + 1403, + 1532, + 1403, + 1625, + 300, + 1625 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 1746, + 1403, + 1746, + 1403, + 1868, + 299, + 1868 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 302, + 506, + 1400, + 506, + 1400, + 600, + 302, + 600 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 300, + 229, + 1401, + 229, + 1401, + 323, + 300, + 323 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 300, + 1972, + 1402, + 1972, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 298, + 647, + 495, + 647, + 495, + 684, + 298, + 684 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.882 + }, + { + "category_id": 0, + "poly": [ + 301, + 1676, + 607, + 1676, + 607, + 1709, + 301, + 1709 + ], + "score": 0.88 + }, + { + "category_id": 0, + "poly": [ + 300, + 1919, + 488, + 1919, + 488, + 1951, + 300, + 1951 + ], + "score": 0.844 + }, + { + "category_id": 2, + "poly": [ + 837, + 2089, + 865, + 2089, + 865, + 2112, + 837, + 2112 + ], + "score": 0.821 + }, + { + "category_id": 15, + "poly": [ + 291.0, + 642.0, + 503.0, + 642.0, + 503.0, + 692.0, + 291.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1676.0, + 611.0, + 1676.0, + 611.0, + 1715.0, + 298.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1919.0, + 490.0, + 1919.0, + 490.0, + 1955.0, + 296.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 859.0, + 1404.0, + 859.0, + 1404.0, + 892.0, + 297.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 889.0, + 1404.0, + 889.0, + 1404.0, + 923.0, + 295.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 920.0, + 1404.0, + 920.0, + 1404.0, + 956.0, + 293.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 948.0, + 1406.0, + 948.0, + 1406.0, + 985.0, + 295.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 980.0, + 1179.0, + 980.0, + 1179.0, + 1017.0, + 294.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1230.0, + 296.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1227.0, + 1405.0, + 1227.0, + 1405.0, + 1259.0, + 296.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1255.0, + 1405.0, + 1255.0, + 1405.0, + 1292.0, + 293.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1286.0, + 1405.0, + 1286.0, + 1405.0, + 1323.0, + 293.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1319.0, + 1404.0, + 1319.0, + 1404.0, + 1351.0, + 294.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1348.0, + 413.0, + 1348.0, + 413.0, + 1380.0, + 293.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1028.0, + 1403.0, + 1028.0, + 1403.0, + 1061.0, + 295.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1059.0, + 1404.0, + 1059.0, + 1404.0, + 1092.0, + 297.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1088.0, + 1405.0, + 1088.0, + 1405.0, + 1121.0, + 294.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1119.0, + 1404.0, + 1119.0, + 1404.0, + 1152.0, + 297.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1231.0, + 1146.0, + 1231.0, + 1184.0, + 294.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 718.0, + 1406.0, + 718.0, + 1406.0, + 754.0, + 295.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 750.0, + 1407.0, + 750.0, + 1407.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 780.0, + 1404.0, + 780.0, + 1404.0, + 815.0, + 293.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 502.0, + 812.0, + 502.0, + 845.0, + 295.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1396.0, + 1404.0, + 1396.0, + 1404.0, + 1428.0, + 296.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1425.0, + 1406.0, + 1425.0, + 1406.0, + 1461.0, + 294.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1455.0, + 1404.0, + 1455.0, + 1404.0, + 1490.0, + 294.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1487.0, + 1025.0, + 1487.0, + 1025.0, + 1521.0, + 294.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 338.0, + 1406.0, + 338.0, + 1406.0, + 371.0, + 298.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 367.0, + 1406.0, + 367.0, + 1406.0, + 404.0, + 295.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 397.0, + 1407.0, + 397.0, + 1407.0, + 434.0, + 292.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 426.0, + 1404.0, + 426.0, + 1404.0, + 467.0, + 293.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 462.0, + 571.0, + 462.0, + 571.0, + 495.0, + 297.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1531.0, + 1405.0, + 1531.0, + 1405.0, + 1569.0, + 294.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1563.0, + 1407.0, + 1563.0, + 1407.0, + 1598.0, + 294.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1598.0, + 504.0, + 1598.0, + 504.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1745.0, + 1404.0, + 1745.0, + 1404.0, + 1780.0, + 294.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1777.0, + 1405.0, + 1777.0, + 1405.0, + 1809.0, + 295.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1805.0, + 1407.0, + 1805.0, + 1407.0, + 1841.0, + 293.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 607.0, + 1835.0, + 607.0, + 1871.0, + 294.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 506.0, + 1405.0, + 506.0, + 1405.0, + 542.0, + 294.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 536.0, + 1404.0, + 536.0, + 1404.0, + 571.0, + 296.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 568.0, + 501.0, + 568.0, + 501.0, + 604.0, + 293.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 983.0, + 292.0, + 983.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1999.0, + 1128.0, + 1999.0, + 1128.0, + 2037.0, + 320.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 297, + 74, + 859, + 74, + 859, + 106, + 297, + 106 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2113, + 835, + 2113 + ], + "score": 0.817 + }, + { + "category_id": 1, + "poly": [ + 292, + 159, + 1408, + 159, + 1408, + 2039, + 292, + 2039 + ], + "score": 0.585 + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 227.0, + 1403.0, + 227.0, + 1403.0, + 270.0, + 291.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 260.0, + 1251.0, + 260.0, + 1251.0, + 299.0, + 320.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 324.0, + 1403.0, + 324.0, + 1403.0, + 363.0, + 295.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 354.0, + 563.0, + 354.0, + 563.0, + 394.0, + 320.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 418.0, + 1403.0, + 418.0, + 1403.0, + 458.0, + 295.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 447.0, + 884.0, + 447.0, + 884.0, + 486.0, + 318.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 509.0, + 1405.0, + 509.0, + 1405.0, + 552.0, + 293.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 540.0, + 400.0, + 540.0, + 400.0, + 581.0, + 318.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 606.0, + 1405.0, + 606.0, + 1405.0, + 645.0, + 295.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 633.0, + 1270.0, + 633.0, + 1270.0, + 672.0, + 322.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 697.0, + 1405.0, + 697.0, + 1405.0, + 738.0, + 293.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 730.0, + 654.0, + 730.0, + 654.0, + 769.0, + 318.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 792.0, + 1401.0, + 792.0, + 1401.0, + 833.0, + 293.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 825.0, + 624.0, + 825.0, + 624.0, + 862.0, + 318.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 887.0, + 1407.0, + 887.0, + 1407.0, + 926.0, + 293.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 917.0, + 1405.0, + 917.0, + 1405.0, + 957.0, + 322.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 946.0, + 884.0, + 946.0, + 884.0, + 990.0, + 316.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1006.0, + 1405.0, + 1006.0, + 1405.0, + 1054.0, + 291.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1043.0, + 930.0, + 1043.0, + 930.0, + 1083.0, + 320.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1105.0, + 1348.0, + 1105.0, + 1348.0, + 1144.0, + 295.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1163.0, + 1407.0, + 1163.0, + 1407.0, + 1210.0, + 289.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1200.0, + 622.0, + 1200.0, + 622.0, + 1239.0, + 320.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1260.0, + 1405.0, + 1260.0, + 1405.0, + 1301.0, + 293.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1295.0, + 983.0, + 1295.0, + 983.0, + 1332.0, + 318.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1357.0, + 1397.0, + 1357.0, + 1397.0, + 1396.0, + 295.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1423.0, + 1403.0, + 1423.0, + 1403.0, + 1456.0, + 297.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1452.0, + 545.0, + 1452.0, + 545.0, + 1486.0, + 318.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 1405.0, + 1511.0, + 1405.0, + 1555.0, + 293.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1547.0, + 1143.0, + 1547.0, + 1143.0, + 1586.0, + 320.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1606.0, + 1407.0, + 1606.0, + 1407.0, + 1650.0, + 293.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1639.0, + 599.0, + 1639.0, + 599.0, + 1676.0, + 318.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1703.0, + 1139.0, + 1703.0, + 1139.0, + 1742.0, + 293.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1765.0, + 1405.0, + 1765.0, + 1405.0, + 1804.0, + 295.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1794.0, + 1194.0, + 1794.0, + 1194.0, + 1837.0, + 316.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1860.0, + 1403.0, + 1860.0, + 1403.0, + 1899.0, + 295.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1891.0, + 932.0, + 1891.0, + 932.0, + 1930.0, + 320.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1953.0, + 1403.0, + 1953.0, + 1403.0, + 1992.0, + 293.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1984.0, + 776.0, + 1984.0, + 776.0, + 2023.0, + 320.0, + 2023.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1641, + 1404, + 1641, + 1404, + 1867, + 297, + 1867 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1298, + 1404, + 1298, + 1404, + 1422, + 297, + 1422 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 943, + 1404, + 943, + 1404, + 1097, + 298, + 1097 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 751, + 1403, + 751, + 1403, + 905, + 298, + 905 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1136, + 1403, + 1136, + 1403, + 1259, + 298, + 1259 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1403, + 1881, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 499, + 1405, + 499, + 1405, + 627, + 298, + 627 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1460, + 1400, + 1460, + 1400, + 1522, + 299, + 1522 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 297, + 353, + 1403, + 353, + 1403, + 425, + 297, + 425 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 670, + 441, + 1031, + 441, + 1031, + 485, + 670, + 485 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 300, + 668, + 506, + 668, + 506, + 701, + 300, + 701 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 302, + 1571, + 736, + 1571, + 736, + 1606, + 302, + 1606 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 0, + "poly": [ + 298, + 227, + 1027, + 227, + 1027, + 262, + 298, + 262 + ], + "score": 0.796 + }, + { + "category_id": 0, + "poly": [ + 298, + 297, + 982, + 297, + 982, + 330, + 298, + 330 + ], + "score": 0.72 + }, + { + "category_id": 1, + "poly": [ + 298, + 297, + 982, + 297, + 982, + 330, + 298, + 330 + ], + "score": 0.233 + }, + { + "category_id": 1, + "poly": [ + 298, + 227, + 1027, + 227, + 1027, + 262, + 298, + 262 + ], + "score": 0.125 + }, + { + "category_id": 13, + "poly": [ + 675, + 501, + 920, + 501, + 920, + 539, + 675, + 539 + ], + "score": 0.93, + "latex": "Q = G ( G ^ { \\top } G ) ^ { - 1 } G ^ { \\top }" + }, + { + "category_id": 13, + "poly": [ + 374, + 502, + 618, + 502, + 618, + 539, + 374, + 539 + ], + "score": 0.93, + "latex": "P = F ( F ^ { \\top } F ) ^ { - 1 } F ^ { \\top }" + }, + { + "category_id": 13, + "poly": [ + 434, + 356, + 600, + 356, + 600, + 390, + 434, + 390 + ], + "score": 0.92, + "latex": "F _ { i j } : = f _ { j } ( x _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1312, + 1643, + 1392, + 1643, + 1392, + 1676, + 1312, + 1676 + ], + "score": 0.92, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 661, + 390, + 808, + 390, + 808, + 427, + 661, + 427 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { A = \\frac { 1 } { N } G ^ { \\top } F } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 735, + 1766, + 992, + 1766, + 992, + 1806, + 735, + 1806 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { p _ { X } ( x ) \\ = \\ \\sum _ { y } p ( x , y ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1114, + 1804, + 1172, + 1804, + 1172, + 1837, + 1114, + 1837 + ], + "score": 0.91, + "latex": "Y | X" + }, + { + "category_id": 13, + "poly": [ + 297, + 1767, + 637, + 1767, + 637, + 1805, + 297, + 1805 + ], + "score": 0.91, + "latex": "p _ { Y | X } ( y | x ) ~ = ~ p ( x , y ) / p _ { Y } ( y )" + }, + { + "category_id": 13, + "poly": [ + 1263, + 1196, + 1400, + 1196, + 1400, + 1231, + 1263, + 1231 + ], + "score": 0.91, + "latex": "( K + \\epsilon { \\bf 1 } ) ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1054, + 1770, + 1308, + 1770, + 1308, + 1803, + 1054, + 1803 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { p _ { Y } ( y ) \\ = \\ \\sum _ { x } p ( x , y ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1013, + 1734, + 1350, + 1734, + 1350, + 1768, + 1013, + 1768 + ], + "score": 0.91, + "latex": "p _ { X | Y } ( x | y ) \\ : = \\ : \\dot { p } ( x , y ) / p _ { X } ( x )" + }, + { + "category_id": 13, + "poly": [ + 982, + 1227, + 1033, + 1227, + 1033, + 1256, + 982, + 1256 + ], + "score": 0.91, + "latex": "L ^ { - 1 }" + }, + { + "category_id": 14, + "poly": [ + 667, + 441, + 1030, + 441, + 1030, + 484, + 667, + 484 + ], + "score": 0.91, + "latex": "\\operatorname { T r } \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) = \\operatorname { T r } \\left( P Q \\right)" + }, + { + "category_id": 13, + "poly": [ + 657, + 356, + 825, + 356, + 825, + 390, + 657, + 390 + ], + "score": 0.91, + "latex": "G _ { i j } : = g _ { j } ( y _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 838, + 1166, + 895, + 1166, + 895, + 1196, + 838, + 1196 + ], + "score": 0.9, + "latex": "K ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1083, + 945, + 1163, + 945, + 1163, + 975, + 1083, + 975 + ], + "score": 0.9, + "latex": "k _ { 0 } - 1" + }, + { + "category_id": 13, + "poly": [ + 809, + 1917, + 872, + 1917, + 872, + 1947, + 809, + 1947 + ], + "score": 0.9, + "latex": "p _ { X | Y }" + }, + { + "category_id": 13, + "poly": [ + 924, + 1166, + 975, + 1166, + 975, + 1196, + 924, + 1196 + ], + "score": 0.9, + "latex": "L ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1165, + 1196, + 1223, + 1196, + 1223, + 1227, + 1165, + 1227 + ], + "score": 0.89, + "latex": "K ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 462, + 390, + 609, + 390, + 609, + 427, + 462, + 427 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { L = \\frac { 1 } { N } G ^ { \\top } G } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1289, + 1301, + 1315, + 1301, + 1315, + 1332, + 1289, + 1332 + ], + "score": 0.88, + "latex": "f _ { i }" + }, + { + "category_id": 13, + "poly": [ + 297, + 388, + 450, + 388, + 450, + 427, + 297, + 427 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { K = \\frac { 1 } { N } F ^ { \\top } F } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1885, + 1328, + 1885, + 1328, + 1916, + 1266, + 1916 + ], + "score": 0.86, + "latex": "p _ { Y \\mid X }" + }, + { + "category_id": 13, + "poly": [ + 1365, + 1304, + 1393, + 1304, + 1393, + 1333, + 1365, + 1333 + ], + "score": 0.84, + "latex": "g _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1327, + 507, + 1352, + 507, + 1352, + 533, + 1327, + 533 + ], + "score": 0.84, + "latex": "F" + }, + { + "category_id": 13, + "poly": [ + 807, + 1645, + 835, + 1645, + 835, + 1671, + 807, + 1671 + ], + "score": 0.83, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 976, + 1804, + 1076, + 1804, + 1076, + 1837, + 976, + 1837 + ], + "score": 0.82, + "latex": "X , X | Y" + }, + { + "category_id": 13, + "poly": [ + 1263, + 1974, + 1292, + 1974, + 1292, + 2000, + 1263, + 2000 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 821, + 1139, + 850, + 1139, + 850, + 1165, + 821, + 1165 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 897, + 1944, + 925, + 1944, + 925, + 1970, + 897, + 1970 + ], + "score": 0.81, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 795, + 1301, + 820, + 1301, + 820, + 1327, + 795, + 1327 + ], + "score": 0.8, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 298, + 538, + 323, + 538, + 323, + 565, + 298, + 565 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 751, + 946, + 776, + 946, + 776, + 972, + 751, + 972 + ], + "score": 0.8, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 816, + 1888, + 834, + 1888, + 834, + 1914, + 816, + 1914 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 902, + 1139, + 924, + 1139, + 924, + 1165, + 902, + 1165 + ], + "score": 0.8, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 447, + 1810, + 464, + 1810, + 464, + 1835, + 447, + 1835 + ], + "score": 0.79, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 886, + 1644, + 912, + 1644, + 912, + 1671, + 886, + 1671 + ], + "score": 0.78, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 902, + 1888, + 920, + 1888, + 920, + 1909, + 902, + 1909 + ], + "score": 0.77, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 773, + 1235, + 788, + 1235, + 788, + 1256, + 773, + 1256 + ], + "score": 0.67, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 667.0, + 509.0, + 667.0, + 509.0, + 706.0, + 295.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1568.0, + 740.0, + 1568.0, + 740.0, + 1612.0, + 293.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 225.0, + 1029.0, + 225.0, + 1029.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 299.0, + 983.0, + 299.0, + 983.0, + 331.0, + 296.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1638.0, + 806.0, + 1638.0, + 806.0, + 1680.0, + 294.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1638.0, + 885.0, + 1638.0, + 885.0, + 1680.0, + 836.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1638.0, + 1311.0, + 1638.0, + 1311.0, + 1680.0, + 913.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1638.0, + 1404.0, + 1638.0, + 1404.0, + 1680.0, + 1393.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1672.0, + 1404.0, + 1672.0, + 1404.0, + 1706.0, + 295.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1703.0, + 1402.0, + 1703.0, + 1402.0, + 1739.0, + 295.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1732.0, + 1012.0, + 1732.0, + 1012.0, + 1772.0, + 291.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1732.0, + 1405.0, + 1732.0, + 1405.0, + 1772.0, + 1351.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1767.0, + 296.0, + 1767.0, + 296.0, + 1806.0, + 291.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1767.0, + 734.0, + 1767.0, + 734.0, + 1806.0, + 638.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1767.0, + 1053.0, + 1767.0, + 1053.0, + 1806.0, + 993.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 1767.0, + 1405.0, + 1767.0, + 1405.0, + 1806.0, + 1309.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1802.0, + 446.0, + 1802.0, + 446.0, + 1837.0, + 292.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1802.0, + 975.0, + 1802.0, + 975.0, + 1837.0, + 465.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1802.0, + 1113.0, + 1802.0, + 1113.0, + 1837.0, + 1077.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1802.0, + 1402.0, + 1802.0, + 1402.0, + 1837.0, + 1173.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1832.0, + 370.0, + 1832.0, + 370.0, + 1875.0, + 291.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1293.0, + 794.0, + 1293.0, + 794.0, + 1339.0, + 291.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1293.0, + 1288.0, + 1293.0, + 1288.0, + 1339.0, + 821.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1293.0, + 1364.0, + 1293.0, + 1364.0, + 1339.0, + 1316.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1293.0, + 1408.0, + 1293.0, + 1408.0, + 1339.0, + 1394.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1329.0, + 1406.0, + 1329.0, + 1406.0, + 1363.0, + 291.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1358.0, + 1409.0, + 1358.0, + 1409.0, + 1396.0, + 292.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1392.0, + 1313.0, + 1392.0, + 1313.0, + 1424.0, + 296.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 945.0, + 750.0, + 945.0, + 750.0, + 978.0, + 297.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 945.0, + 1082.0, + 945.0, + 1082.0, + 978.0, + 777.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 945.0, + 1405.0, + 945.0, + 1405.0, + 978.0, + 1164.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 975.0, + 1401.0, + 975.0, + 1401.0, + 1008.0, + 296.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1003.0, + 1405.0, + 1003.0, + 1405.0, + 1043.0, + 293.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1033.0, + 1406.0, + 1033.0, + 1406.0, + 1073.0, + 293.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1065.0, + 1129.0, + 1065.0, + 1129.0, + 1103.0, + 292.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 751.0, + 1404.0, + 751.0, + 1404.0, + 787.0, + 294.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 783.0, + 1405.0, + 783.0, + 1405.0, + 816.0, + 296.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 813.0, + 1404.0, + 813.0, + 1404.0, + 847.0, + 294.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 844.0, + 1403.0, + 844.0, + 1403.0, + 877.0, + 294.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 873.0, + 534.0, + 873.0, + 534.0, + 909.0, + 293.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1132.0, + 820.0, + 1132.0, + 820.0, + 1173.0, + 293.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1132.0, + 901.0, + 1132.0, + 901.0, + 1173.0, + 851.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1132.0, + 1407.0, + 1132.0, + 1407.0, + 1173.0, + 925.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1165.0, + 837.0, + 1165.0, + 837.0, + 1201.0, + 293.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1165.0, + 923.0, + 1165.0, + 923.0, + 1201.0, + 896.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1165.0, + 1407.0, + 1165.0, + 1407.0, + 1201.0, + 976.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1196.0, + 1164.0, + 1196.0, + 1164.0, + 1231.0, + 292.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1196.0, + 1262.0, + 1196.0, + 1262.0, + 1231.0, + 1224.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1228.0, + 772.0, + 1228.0, + 772.0, + 1260.0, + 292.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1228.0, + 981.0, + 1228.0, + 981.0, + 1260.0, + 789.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1228.0, + 1042.0, + 1228.0, + 1042.0, + 1260.0, + 1034.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 815.0, + 1881.0, + 815.0, + 1918.0, + 293.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1881.0, + 901.0, + 1881.0, + 901.0, + 1918.0, + 835.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1881.0, + 1265.0, + 1881.0, + 1265.0, + 1918.0, + 921.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1918.0, + 1329.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 808.0, + 1909.0, + 808.0, + 1950.0, + 293.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1950.0, + 873.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 896.0, + 1941.0, + 896.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1941.0, + 1407.0, + 1941.0, + 1407.0, + 1978.0, + 926.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1262.0, + 1971.0, + 1262.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1971.0, + 1407.0, + 1971.0, + 1407.0, + 2008.0, + 1293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 585.0, + 1999.0, + 585.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 499.0, + 373.0, + 499.0, + 373.0, + 543.0, + 292.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 499.0, + 674.0, + 499.0, + 674.0, + 543.0, + 619.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 499.0, + 1326.0, + 499.0, + 1326.0, + 543.0, + 921.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 499.0, + 1406.0, + 499.0, + 1406.0, + 543.0, + 1353.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 534.0, + 297.0, + 534.0, + 297.0, + 573.0, + 293.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 534.0, + 1405.0, + 534.0, + 1405.0, + 573.0, + 324.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 564.0, + 1406.0, + 564.0, + 1406.0, + 602.0, + 292.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 594.0, + 503.0, + 594.0, + 503.0, + 631.0, + 294.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1458.0, + 1405.0, + 1458.0, + 1405.0, + 1497.0, + 295.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1491.0, + 1028.0, + 1491.0, + 1028.0, + 1526.0, + 295.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 350.0, + 433.0, + 350.0, + 433.0, + 394.0, + 292.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 350.0, + 656.0, + 350.0, + 656.0, + 394.0, + 601.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 350.0, + 1407.0, + 350.0, + 1407.0, + 394.0, + 826.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 384.0, + 296.0, + 384.0, + 296.0, + 429.0, + 293.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 384.0, + 461.0, + 384.0, + 461.0, + 429.0, + 451.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 384.0, + 660.0, + 384.0, + 660.0, + 429.0, + 610.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 384.0, + 1315.0, + 384.0, + 1315.0, + 429.0, + 809.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 299.0, + 983.0, + 299.0, + 983.0, + 331.0, + 296.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 225.0, + 1029.0, + 225.0, + 1029.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 286, + 1404, + 286, + 1404, + 411, + 297, + 411 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 525, + 1404, + 525, + 1404, + 652, + 297, + 652 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 296, + 665, + 1404, + 665, + 1404, + 787, + 296, + 787 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1568, + 1404, + 1568, + 1404, + 1664, + 297, + 1664 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1432, + 1404, + 1432, + 1404, + 1523, + 298, + 1523 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 694, + 426, + 1005, + 426, + 1005, + 508, + 694, + 508 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 667, + 932, + 1031, + 932, + 1031, + 1002, + 667, + 1002 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 296, + 848, + 1397, + 848, + 1397, + 915, + 296, + 915 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1402, + 1971, + 1402, + 2035, + 297, + 2035 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 760, + 1525, + 937, + 1525, + 937, + 1562, + 760, + 1562 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 298, + 1198, + 1403, + 1198, + 1403, + 1261, + 298, + 1261 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 753, + 1680, + 942, + 1680, + 942, + 1720, + 753, + 1720 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 663, + 1113, + 1036, + 1113, + 1036, + 1183, + 663, + 1183 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 299, + 1731, + 1404, + 1731, + 1404, + 1802, + 299, + 1802 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 297, + 1813, + 1405, + 1813, + 1405, + 1877, + 297, + 1877 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 659, + 786, + 1038, + 786, + 1038, + 828, + 659, + 828 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 295, + 1064, + 1051, + 1064, + 1051, + 1097, + 295, + 1097 + ], + "score": 0.932 + }, + { + "category_id": 8, + "poly": [ + 687, + 1261, + 1011, + 1261, + 1011, + 1297, + 687, + 1297 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 297, + 1018, + 481, + 1018, + 481, + 1051, + 297, + 1051 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 106, + 297, + 106 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 300, + 1915, + 710, + 1915, + 710, + 1947, + 300, + 1947 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 296, + 1303, + 928, + 1303, + 928, + 1340, + 296, + 1340 + ], + "score": 0.914 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1124, + 1400, + 1124, + 1400, + 1155, + 1352, + 1155 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 300, + 1374, + 805, + 1374, + 805, + 1407, + 300, + 1407 + ], + "score": 0.911 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1683, + 1399, + 1683, + 1399, + 1715, + 1352, + 1715 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1526, + 1399, + 1526, + 1399, + 1556, + 1352, + 1556 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1263, + 1400, + 1263, + 1400, + 1294, + 1352, + 1294 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1352, + 941, + 1400, + 941, + 1400, + 973, + 1352, + 973 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1352, + 790, + 1400, + 790, + 1400, + 820, + 1352, + 820 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1352, + 447, + 1400, + 447, + 1400, + 479, + 1352, + 479 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 864, + 2087, + 864, + 2113, + 836, + 2113 + ], + "score": 0.866 + }, + { + "category_id": 0, + "poly": [ + 304, + 230, + 891, + 230, + 891, + 261, + 304, + 261 + ], + "score": 0.84 + }, + { + "category_id": 14, + "poly": [ + 692, + 425, + 1007, + 425, + 1007, + 509, + 692, + 509 + ], + "score": 0.95, + "latex": "\\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } : = \\sum _ { x } { \\frac { \\mu ( x ) \\mu ^ { \\prime } ( x ) } { p _ { X } ( x ) } }" + }, + { + "category_id": 14, + "poly": [ + 661, + 1111, + 1038, + 1111, + 1038, + 1186, + 661, + 1186 + ], + "score": 0.94, + "latex": "\\mathcal { N ^ { * } } ( \\nu ) ( x ) = \\sum _ { x } p _ { X | Y } ( x | y ) \\nu ( x )" + }, + { + "category_id": 13, + "poly": [ + 1057, + 559, + 1138, + 559, + 1138, + 592, + 1057, + 592 + ], + "score": 0.94, + "latex": "p ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 339, + 1570, + 489, + 1570, + 489, + 1601, + 339, + 1601 + ], + "score": 0.94, + "latex": "j = 1 , \\dotsc , k" + }, + { + "category_id": 14, + "poly": [ + 667, + 930, + 1032, + 930, + 1032, + 1003, + 667, + 1003 + ], + "score": 0.94, + "latex": "\\mathcal { N } ( \\mu ) ( y ) = \\sum _ { x } p _ { Y | X } ( y | x ) \\mu ( x )" + }, + { + "category_id": 14, + "poly": [ + 754, + 1680, + 946, + 1680, + 946, + 1720, + 754, + 1720 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\mathcal { N } ^ { * } ( v _ { j } ) = \\eta _ { j } u _ { j } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 704, + 526, + 902, + 526, + 902, + 562, + 704, + 562 + ], + "score": 0.93, + "latex": "\\| \\mu \\| _ { X } ^ { 2 } = \\langle \\mu , \\mu \\rangle _ { X }" + }, + { + "category_id": 14, + "poly": [ + 683, + 1259, + 1016, + 1259, + 1016, + 1298, + 683, + 1298 + ], + "score": 0.93, + "latex": "\\langle \\nu , \\mathcal { N } ( \\mu ) \\rangle _ { Y } = \\langle \\mathcal { N } ^ { * } ( \\nu ) , \\mu \\rangle _ { X } ." + }, + { + "category_id": 14, + "poly": [ + 760, + 1522, + 937, + 1522, + 937, + 1561, + 760, + 1561 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\mathcal { N } ( u _ { j } ) = \\eta _ { j } v _ { j } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 544, + 1305, + 699, + 1305, + 699, + 1339, + 544, + 1339 + ], + "score": 0.93, + "latex": "\\mathcal { N } ( p _ { X } ) = p _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 494, + 619, + 586, + 619, + 586, + 653, + 494, + 653 + ], + "score": 0.93, + "latex": "\\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 564, + 559, + 637, + 559, + 637, + 591, + 564, + 591 + ], + "score": 0.92, + "latex": "p _ { X } ( x )" + }, + { + "category_id": 13, + "poly": [ + 387, + 528, + 520, + 528, + 520, + 561, + 387, + 561 + ], + "score": 0.92, + "latex": "\\mu , \\mu ^ { \\prime } \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 630, + 619, + 750, + 619, + 750, + 651, + 630, + 651 + ], + "score": 0.92, + "latex": "\\nu , \\nu ^ { \\prime } \\in V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 569, + 1603, + 686, + 1603, + 686, + 1636, + 569, + 1636 + ], + "score": 0.91, + "latex": "\\eta _ { j } \\in [ 0 , 1 ]" + }, + { + "category_id": 14, + "poly": [ + 658, + 783, + 1041, + 783, + 1041, + 826, + 658, + 826 + ], + "score": 0.91, + "latex": "\\chi ^ { 2 } ( q , p _ { X } ) = \\langle q - p _ { X } , q - p _ { X } \\rangle _ { X } ." + }, + { + "category_id": 13, + "poly": [ + 371, + 1232, + 458, + 1232, + 458, + 1261, + 371, + 1261 + ], + "score": 0.91, + "latex": "\\nu \\in V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 298, + 379, + 409, + 379, + 409, + 410, + 298, + 410 + ], + "score": 0.91, + "latex": "V _ { X } = \\mathbb { R } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 510, + 1231, + 599, + 1231, + 599, + 1262, + 510, + 1262 + ], + "score": 0.91, + "latex": "\\mu \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 729, + 1733, + 832, + 1733, + 832, + 1762, + 729, + 1762 + ], + "score": 0.91, + "latex": "\\mathcal { N } ^ { \\ast } \\circ \\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 826, + 856, + 888, + 856, + 888, + 886, + 826, + 886 + ], + "score": 0.91, + "latex": "p _ { Y \\mid X }" + }, + { + "category_id": 13, + "poly": [ + 383, + 1020, + 473, + 1020, + 473, + 1051, + 383, + 1051 + ], + "score": 0.91, + "latex": "\\mu \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 751, + 1304, + 919, + 1304, + 919, + 1340, + 751, + 1340 + ], + "score": 0.91, + "latex": "\\mathcal { N } ^ { * } ( p _ { Y } ) = p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1498, + 414, + 1498, + 414, + 1525, + 297, + 1525 + ], + "score": 0.9, + "latex": "v _ { 1 } , \\ldots , v _ { k }" + }, + { + "category_id": 13, + "poly": [ + 888, + 1469, + 1011, + 1469, + 1011, + 1495, + 888, + 1495 + ], + "score": 0.9, + "latex": "u _ { 1 } , \\ldots , u _ { k }" + }, + { + "category_id": 13, + "poly": [ + 794, + 1600, + 827, + 1600, + 827, + 1634, + 794, + 1634 + ], + "score": 0.9, + "latex": "\\chi ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1049, + 1465, + 1089, + 1465, + 1089, + 1494, + 1049, + 1494 + ], + "score": 0.9, + "latex": "V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 424, + 590, + 464, + 590, + 464, + 620, + 424, + 620 + ], + "score": 0.9, + "latex": "V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 1172, + 1733, + 1272, + 1733, + 1272, + 1762, + 1172, + 1762 + ], + "score": 0.9, + "latex": "\\mathcal { N } \\circ \\mathcal { N } ^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 751, + 1435, + 788, + 1435, + 788, + 1464, + 751, + 1464 + ], + "score": 0.9, + "latex": "V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 422, + 1766, + 452, + 1766, + 452, + 1805, + 422, + 1805 + ], + "score": 0.89, + "latex": "\\eta _ { j } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 662, + 1435, + 702, + 1435, + 702, + 1464, + 662, + 1464 + ], + "score": 0.89, + "latex": "V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 449, + 1495, + 486, + 1495, + 486, + 1525, + 449, + 1525 + ], + "score": 0.89, + "latex": "V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 1108, + 2003, + 1205, + 2003, + 1205, + 2032, + 1108, + 2032 + ], + "score": 0.89, + "latex": "2 8 \\times 2 8" + }, + { + "category_id": 13, + "poly": [ + 1233, + 1820, + 1342, + 1820, + 1342, + 1847, + 1233, + 1847 + ], + "score": 0.89, + "latex": "v _ { 0 } = p _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 1065, + 1821, + 1177, + 1821, + 1177, + 1847, + 1065, + 1847 + ], + "score": 0.89, + "latex": "u _ { 0 } ~ = ~ p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 338, + 756, + 370, + 756, + 370, + 790, + 338, + 790 + ], + "score": 0.89, + "latex": "\\chi ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 860, + 667, + 899, + 667, + 899, + 697, + 860, + 697 + ], + "score": 0.89, + "latex": "V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 591, + 666, + 619, + 666, + 619, + 698, + 591, + 698 + ], + "score": 0.88, + "latex": "\\mu ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 481, + 1200, + 523, + 1200, + 523, + 1227, + 481, + 1227 + ], + "score": 0.86, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 427, + 882, + 554, + 882, + 554, + 916, + 427, + 916 + ], + "score": 0.86, + "latex": "\\mu \\mapsto { \\mathcal { N } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 298, + 882, + 415, + 882, + 415, + 914, + 298, + 914 + ], + "score": 0.86, + "latex": "V _ { X } \\to V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 884, + 1065, + 926, + 1065, + 926, + 1093, + 884, + 1093 + ], + "score": 0.86, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1176, + 561, + 1214, + 561, + 1214, + 591, + 1176, + 591 + ], + "score": 0.86, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 1022, + 732, + 1060, + 732, + 1060, + 760, + 1022, + 760 + ], + "score": 0.86, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 888, + 1739, + 917, + 1739, + 917, + 1768, + 888, + 1768 + ], + "score": 0.85, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 578, + 1821, + 614, + 1821, + 614, + 1847, + 578, + 1847 + ], + "score": 0.85, + "latex": "p _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 419, + 1633, + 461, + 1633, + 461, + 1660, + 419, + 1660 + ], + "score": 0.85, + "latex": "{ \\ddot { \\mathcal { N } } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1349, + 2003, + 1404, + 2003, + 1404, + 2033, + 1349, + 2033 + ], + "score": 0.85, + "latex": "k =" + }, + { + "category_id": 13, + "poly": [ + 674, + 1634, + 703, + 1634, + 703, + 1660, + 674, + 1660 + ], + "score": 0.84, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 503, + 1821, + 541, + 1821, + 541, + 1847, + 503, + 1847 + ], + "score": 0.84, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 586, + 2004, + 614, + 2004, + 614, + 2030, + 586, + 2030 + ], + "score": 0.83, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 831, + 320, + 858, + 320, + 858, + 346, + 831, + 346 + ], + "score": 0.83, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 413, + 1607, + 441, + 1607, + 441, + 1635, + 413, + 1635 + ], + "score": 0.83, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 453, + 1739, + 485, + 1739, + 485, + 1767, + 453, + 1767 + ], + "score": 0.83, + "latex": "u _ { j }" + }, + { + "category_id": 13, + "poly": [ + 588, + 1974, + 607, + 1974, + 607, + 2000, + 588, + 2000 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 923, + 1569, + 953, + 1569, + 953, + 1596, + 923, + 1596 + ], + "score": 0.83, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 560, + 1201, + 590, + 1201, + 590, + 1227, + 560, + 1227 + ], + "score": 0.81, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 391, + 1464, + 420, + 1464, + 420, + 1491, + 391, + 1491 + ], + "score": 0.8, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1360, + 734, + 1376, + 734, + 1376, + 760, + 1360, + 760 + ], + "score": 0.8, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 519, + 671, + 540, + 671, + 540, + 698, + 519, + 698 + ], + "score": 0.8, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 1297, + 289, + 1325, + 289, + 1325, + 315, + 1297, + 315 + ], + "score": 0.8, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 399, + 1815, + 429, + 1815, + 429, + 1842, + 399, + 1842 + ], + "score": 0.79, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1269, + 590, + 1294, + 590, + 1294, + 616, + 1269, + 616 + ], + "score": 0.78, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1375, + 288, + 1401, + 288, + 1401, + 316, + 1375, + 316 + ], + "score": 0.75, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 643, + 1574, + 671, + 1574, + 671, + 1602, + 643, + 1602 + ], + "score": 0.67, + "latex": "\\eta _ { j }" + }, + { + "category_id": 13, + "poly": [ + 613, + 1571, + 671, + 1571, + 671, + 1602, + 613, + 1602 + ], + "score": 0.46, + "latex": "j , \\eta _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1354, + 851, + 1384, + 851, + 1384, + 877, + 1354, + 877 + ], + "score": 0.45, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 614, + 1571, + 630, + 1571, + 630, + 1601, + 614, + 1601 + ], + "score": 0.39, + "latex": "j" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1914.0, + 713.0, + 1914.0, + 713.0, + 1949.0, + 295.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1373.0, + 809.0, + 1373.0, + 809.0, + 1410.0, + 293.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 896.0, + 228.0, + 896.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 285.0, + 1296.0, + 285.0, + 1296.0, + 322.0, + 292.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 285.0, + 1374.0, + 285.0, + 1374.0, + 322.0, + 1326.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 285.0, + 1405.0, + 285.0, + 1405.0, + 322.0, + 1402.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 320.0, + 830.0, + 320.0, + 830.0, + 350.0, + 296.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 320.0, + 1404.0, + 320.0, + 1404.0, + 350.0, + 859.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 350.0, + 1404.0, + 350.0, + 1404.0, + 382.0, + 295.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 380.0, + 871.0, + 380.0, + 871.0, + 413.0, + 410.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 523.0, + 386.0, + 523.0, + 386.0, + 565.0, + 294.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 523.0, + 703.0, + 523.0, + 703.0, + 565.0, + 521.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 523.0, + 1405.0, + 523.0, + 1405.0, + 565.0, + 903.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 555.0, + 563.0, + 555.0, + 563.0, + 594.0, + 292.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 555.0, + 1056.0, + 555.0, + 1056.0, + 594.0, + 638.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 555.0, + 1175.0, + 555.0, + 1175.0, + 594.0, + 1139.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 555.0, + 1405.0, + 555.0, + 1405.0, + 594.0, + 1215.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 586.0, + 423.0, + 586.0, + 423.0, + 625.0, + 292.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 586.0, + 1268.0, + 586.0, + 1268.0, + 625.0, + 465.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 586.0, + 1405.0, + 586.0, + 1405.0, + 625.0, + 1295.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 620.0, + 493.0, + 620.0, + 493.0, + 654.0, + 295.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 620.0, + 629.0, + 620.0, + 629.0, + 654.0, + 587.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 620.0, + 762.0, + 620.0, + 762.0, + 654.0, + 751.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 667.0, + 518.0, + 667.0, + 518.0, + 699.0, + 295.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 667.0, + 590.0, + 667.0, + 590.0, + 699.0, + 541.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 667.0, + 859.0, + 667.0, + 859.0, + 699.0, + 620.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 667.0, + 1402.0, + 667.0, + 1402.0, + 699.0, + 900.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 697.0, + 1404.0, + 697.0, + 1404.0, + 729.0, + 294.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 726.0, + 1021.0, + 726.0, + 1021.0, + 764.0, + 291.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 726.0, + 1359.0, + 726.0, + 1359.0, + 764.0, + 1061.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 726.0, + 1405.0, + 726.0, + 1405.0, + 764.0, + 1377.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 754.0, + 337.0, + 754.0, + 337.0, + 794.0, + 294.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 754.0, + 513.0, + 754.0, + 513.0, + 794.0, + 371.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1566.0, + 338.0, + 1566.0, + 338.0, + 1604.0, + 294.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 1566.0, + 612.0, + 1566.0, + 612.0, + 1604.0, + 490.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1566.0, + 922.0, + 1566.0, + 922.0, + 1604.0, + 672.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1566.0, + 1407.0, + 1566.0, + 1407.0, + 1604.0, + 954.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1600.0, + 412.0, + 1600.0, + 412.0, + 1637.0, + 294.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1600.0, + 568.0, + 1600.0, + 568.0, + 1637.0, + 442.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1600.0, + 793.0, + 1600.0, + 793.0, + 1637.0, + 687.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 1600.0, + 1404.0, + 1600.0, + 1404.0, + 1637.0, + 828.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1629.0, + 418.0, + 1629.0, + 418.0, + 1667.0, + 296.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1629.0, + 673.0, + 1629.0, + 673.0, + 1667.0, + 462.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1629.0, + 715.0, + 1629.0, + 715.0, + 1667.0, + 704.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1431.0, + 661.0, + 1431.0, + 661.0, + 1467.0, + 295.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1431.0, + 750.0, + 1431.0, + 750.0, + 1467.0, + 703.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1431.0, + 1403.0, + 1431.0, + 1403.0, + 1467.0, + 789.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1461.0, + 390.0, + 1461.0, + 390.0, + 1500.0, + 292.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1461.0, + 887.0, + 1461.0, + 887.0, + 1500.0, + 421.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1461.0, + 1048.0, + 1461.0, + 1048.0, + 1500.0, + 1012.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1461.0, + 1404.0, + 1461.0, + 1404.0, + 1500.0, + 1090.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1496.0, + 296.0, + 1496.0, + 296.0, + 1526.0, + 293.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1496.0, + 448.0, + 1496.0, + 448.0, + 1526.0, + 415.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1496.0, + 604.0, + 1496.0, + 604.0, + 1526.0, + 487.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 842.0, + 825.0, + 842.0, + 825.0, + 887.0, + 294.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 842.0, + 1353.0, + 842.0, + 1353.0, + 887.0, + 889.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1385.0, + 842.0, + 1404.0, + 842.0, + 1404.0, + 887.0, + 1385.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 882.0, + 426.0, + 882.0, + 426.0, + 918.0, + 416.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 882.0, + 640.0, + 882.0, + 640.0, + 918.0, + 555.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1974.0, + 587.0, + 1974.0, + 587.0, + 2006.0, + 297.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2006.0, + 608.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 585.0, + 2000.0, + 585.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 2000.0, + 1107.0, + 2000.0, + 1107.0, + 2038.0, + 615.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 2000.0, + 1348.0, + 2000.0, + 1348.0, + 2038.0, + 1206.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1194.0, + 480.0, + 1194.0, + 480.0, + 1237.0, + 292.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1194.0, + 559.0, + 1194.0, + 559.0, + 1237.0, + 524.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1194.0, + 1407.0, + 1194.0, + 1407.0, + 1237.0, + 591.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1230.0, + 370.0, + 1230.0, + 370.0, + 1266.0, + 294.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1230.0, + 509.0, + 1230.0, + 509.0, + 1266.0, + 459.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1230.0, + 612.0, + 1230.0, + 612.0, + 1266.0, + 600.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1726.0, + 452.0, + 1726.0, + 452.0, + 1769.0, + 292.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1726.0, + 728.0, + 1726.0, + 728.0, + 1769.0, + 486.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1726.0, + 887.0, + 1726.0, + 887.0, + 1769.0, + 833.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1726.0, + 1171.0, + 1726.0, + 1171.0, + 1769.0, + 918.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1726.0, + 1406.0, + 1726.0, + 1406.0, + 1769.0, + 1273.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1762.0, + 421.0, + 1762.0, + 421.0, + 1809.0, + 292.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1762.0, + 465.0, + 1762.0, + 465.0, + 1809.0, + 453.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1808.0, + 398.0, + 1808.0, + 398.0, + 1853.0, + 291.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 1808.0, + 502.0, + 1808.0, + 502.0, + 1853.0, + 430.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1808.0, + 577.0, + 1808.0, + 577.0, + 1853.0, + 542.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1808.0, + 1064.0, + 1808.0, + 1064.0, + 1853.0, + 615.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1808.0, + 1232.0, + 1808.0, + 1232.0, + 1853.0, + 1178.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1808.0, + 1406.0, + 1808.0, + 1406.0, + 1853.0, + 1343.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1843.0, + 453.0, + 1843.0, + 453.0, + 1878.0, + 292.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1055.0, + 883.0, + 1055.0, + 883.0, + 1108.0, + 292.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1055.0, + 1053.0, + 1055.0, + 1053.0, + 1108.0, + 927.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1016.0, + 382.0, + 1016.0, + 382.0, + 1055.0, + 293.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1016.0, + 485.0, + 1016.0, + 485.0, + 1055.0, + 474.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1296.0, + 543.0, + 1296.0, + 543.0, + 1347.0, + 293.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1296.0, + 750.0, + 1296.0, + 750.0, + 1347.0, + 700.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1296.0, + 932.0, + 1296.0, + 932.0, + 1347.0, + 920.0, + 1347.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1227, + 1404, + 1227, + 1404, + 1383, + 296, + 1383 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 306, + 1404, + 306, + 1404, + 401, + 297, + 401 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 599, + 1390, + 1096, + 1390, + 1096, + 1535, + 599, + 1535 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 300, + 884, + 1403, + 884, + 1403, + 979, + 300, + 979 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 302, + 776, + 1402, + 776, + 1402, + 869, + 302, + 869 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 518, + 1025, + 1178, + 1025, + 1178, + 1109, + 518, + 1109 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 294, + 227, + 1402, + 227, + 1402, + 294, + 294, + 294 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 626, + 1955, + 1063, + 1955, + 1063, + 2027, + 626, + 2027 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 683, + 700, + 1014, + 700, + 1014, + 768, + 683, + 768 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 722, + 1182, + 977, + 1182, + 977, + 1222, + 722, + 1222 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 297, + 1140, + 1224, + 1140, + 1224, + 1176, + 297, + 1176 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 687, + 489, + 1007, + 489, + 1007, + 565, + 687, + 565 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 690, + 408, + 1010, + 408, + 1010, + 482, + 690, + 482 + ], + "score": 0.932 + }, + { + "category_id": 8, + "poly": [ + 590, + 1726, + 1112, + 1726, + 1112, + 1762, + 590, + 1762 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 291, + 1665, + 1403, + 1665, + 1403, + 1726, + 291, + 1726 + ], + "score": 0.929 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 106, + 297, + 106 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 296, + 1764, + 917, + 1764, + 917, + 1797, + 296, + 1797 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 299, + 984, + 1151, + 984, + 1151, + 1016, + 299, + 1016 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 298, + 1609, + 642, + 1609, + 642, + 1641, + 298, + 1641 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1186, + 1400, + 1186, + 1400, + 1217, + 1352, + 1217 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 295, + 1810, + 1405, + 1810, + 1405, + 1873, + 295, + 1873 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1050, + 1400, + 1050, + 1400, + 1082, + 1351, + 1082 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1728, + 1400, + 1728, + 1400, + 1758, + 1351, + 1758 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1351, + 418, + 1400, + 418, + 1400, + 449, + 1351, + 449 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1352, + 570, + 1400, + 570, + 1400, + 600, + 1352, + 600 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1965, + 1401, + 1965, + 1401, + 1996, + 1351, + 1996 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1890, + 1400, + 1890, + 1400, + 1921, + 1352, + 1921 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1351, + 498, + 1401, + 498, + 1401, + 529, + 1351, + 529 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1351, + 708, + 1400, + 708, + 1400, + 740, + 1351, + 740 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 293, + 610, + 1371, + 610, + 1371, + 646, + 293, + 646 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 298, + 655, + 1071, + 655, + 1071, + 692, + 298, + 692 + ], + "score": 0.834 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1542, + 1403, + 1542, + 1403, + 1571, + 1374, + 1571 + ], + "score": 0.828 + }, + { + "category_id": 8, + "poly": [ + 621, + 1880, + 1077, + 1880, + 1077, + 1948, + 621, + 1948 + ], + "score": 0.715 + }, + { + "category_id": 8, + "poly": [ + 623, + 1880, + 1077, + 1880, + 1077, + 1942, + 623, + 1942 + ], + "score": 0.45 + }, + { + "category_id": 9, + "poly": [ + 1351, + 708, + 1400, + 708, + 1400, + 740, + 1351, + 740 + ], + "score": 0.106 + }, + { + "category_id": 14, + "poly": [ + 683, + 698, + 1017, + 698, + 1017, + 770, + 683, + 770 + ], + "score": 0.94, + "latex": "{ \\mathcal { N } } _ { 0 } ( \\mu ) ( y ) = \\sum _ { x } q ( y | x ) \\mu ( x ) ." + }, + { + "category_id": 14, + "poly": [ + 515, + 1022, + 1181, + 1022, + 1181, + 1111, + 515, + 1111 + ], + "score": 0.93, + "latex": "\\sum _ { x } p ( x ) \\| \\mathcal { N } _ { 0 } ( \\delta _ { x } ) - \\mathcal { N } ( \\delta _ { x } ) \\| _ { Y } ^ { 2 } = \\sum _ { x y } \\frac { ( q ( x , y ) - p ( x , y ) ) ^ { 2 } } { p ( x ) p ( y ) } ." + }, + { + "category_id": 13, + "poly": [ + 1223, + 1811, + 1352, + 1811, + 1352, + 1844, + 1223, + 1844 + ], + "score": 0.93, + "latex": "\\mu ^ { \\prime } = p _ { X } f ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 813, + 777, + 1273, + 777, + 1273, + 811, + 813, + 811 + ], + "score": 0.93, + "latex": "q ( x , y ) = q ( x | y ) p _ { Y } ( y ) = q ( y | x ) p _ { X } ( x )" + }, + { + "category_id": 14, + "poly": [ + 687, + 407, + 1012, + 407, + 1012, + 567, + 687, + 567 + ], + "score": 0.93, + "latex": "\\begin{array} { l } { { \\displaystyle \\mathcal { N } _ { 0 } ( \\mu ) = \\sum _ { j \\leq k _ { 0 } } \\eta _ { j } \\langle u _ { j } , \\mu \\rangle _ { X } v _ { j } } } \\\\ { { \\displaystyle \\mathcal { N } _ { 0 } ^ { * } ( \\nu ) = \\sum _ { j \\leq k _ { 0 } } \\eta _ { j } \\langle v _ { j } , \\nu \\rangle _ { Y } u _ { j } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1006, + 1142, + 1146, + 1142, + 1146, + 1176, + 1006, + 1176 + ], + "score": 0.93, + "latex": "\\| \\mathcal { N } _ { 0 } - \\mathcal { N } \\| _ { \\mathrm { F } }" + }, + { + "category_id": 14, + "poly": [ + 597, + 1389, + 1100, + 1389, + 1100, + 1538, + 597, + 1538 + ], + "score": 0.93, + "latex": "\\begin{array} { l } { \\displaystyle \\operatorname { T r } \\left( \\mathcal { M ^ { * } M } \\right) = \\sum _ { x } p _ { X } ( x ) \\langle \\delta _ { x } , \\mathcal { M ^ { * } M } ( \\delta _ { x } ) \\rangle _ { Y } } \\\\ { \\displaystyle \\quad = \\sum _ { x } p _ { X } ( x ) \\| \\mathcal { M } ( \\delta _ { x } ) \\| _ { Y } ^ { 2 } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 721, + 1181, + 979, + 1181, + 979, + 1221, + 721, + 1221 + ], + "score": 0.93, + "latex": "\\| \\mathcal { M } \\| _ { F } ^ { 2 } = \\mathrm { T r } \\left( \\mathcal { M } ^ { * } \\mathcal { M } \\right)" + }, + { + "category_id": 13, + "poly": [ + 491, + 809, + 571, + 809, + 571, + 841, + 491, + 841 + ], + "score": 0.93, + "latex": "\\dot { \\boldsymbol { q } } ( \\boldsymbol { x } , \\boldsymbol { y } )" + }, + { + "category_id": 13, + "poly": [ + 433, + 1351, + 659, + 1351, + 659, + 1384, + 433, + 1384 + ], + "score": 0.93, + "latex": "\\langle \\delta _ { x } , \\delta _ { x } \\rangle = 1 / p _ { X } ( x )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1841, + 414, + 1841, + 414, + 1874, + 298, + 1874 + ], + "score": 0.92, + "latex": "\\nu ^ { \\prime } = p _ { Y } g ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1233, + 809, + 1312, + 809, + 1312, + 841, + 1233, + 841 + ], + "score": 0.92, + "latex": "q ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1319, + 414, + 1319, + 414, + 1353, + 297, + 1353 + ], + "score": 0.92, + "latex": "\\delta _ { x } ( y ) = 1" + }, + { + "category_id": 13, + "poly": [ + 803, + 658, + 878, + 658, + 878, + 692, + 803, + 692 + ], + "score": 0.92, + "latex": "q ( y | x )" + }, + { + "category_id": 13, + "poly": [ + 298, + 370, + 441, + 370, + 441, + 401, + 298, + 401 + ], + "score": 0.92, + "latex": "j = 1 , \\dots , k" + }, + { + "category_id": 13, + "poly": [ + 588, + 1667, + 690, + 1667, + 690, + 1699, + 588, + 1699 + ], + "score": 0.92, + "latex": "\\mu \\in \\ V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 747, + 1667, + 846, + 1667, + 846, + 1697, + 747, + 1697 + ], + "score": 0.91, + "latex": "\\nu \\in \\ V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 930, + 657, + 1005, + 657, + 1005, + 692, + 930, + 692 + ], + "score": 0.91, + "latex": "q ( x | y )" + }, + { + "category_id": 13, + "poly": [ + 802, + 946, + 843, + 946, + 843, + 980, + 802, + 980 + ], + "score": 0.91, + "latex": "\\mathcal { N } _ { 0 } ^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 297, + 224, + 490, + 224, + 490, + 260, + 297, + 260 + ], + "score": 0.91, + "latex": "2 5 6 ^ { 2 8 ^ { 2 } } \\simeq 1 0 ^ { 1 8 8 8 }" + }, + { + "category_id": 13, + "poly": [ + 722, + 658, + 763, + 658, + 763, + 692, + 722, + 692 + ], + "score": 0.9, + "latex": "\\mathcal { N } _ { 0 } ^ { \\ast }" + }, + { + "category_id": 14, + "poly": [ + 620, + 1878, + 1076, + 1878, + 1076, + 2032, + 620, + 2032 + ], + "score": 0.9, + "latex": "\\begin{array} { l } { { \\langle \\mu , \\mu ^ { \\prime } \\rangle _ { X } = \\displaystyle \\sum _ { x } p _ { X } ( x ) f ( x ) f ^ { \\prime } ( x ) = \\overline { { { f f ^ { \\prime } } } } , } } \\\\ { { \\langle \\nu , \\nu ^ { \\prime } \\rangle _ { Y } = \\displaystyle \\sum _ { y } p _ { Y } ( y ) g ( y ) g ^ { \\prime } ( y ) = \\overline { { { g g ^ { \\prime } } } } . } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 462, + 777, + 504, + 777, + 504, + 810, + 462, + 810 + ], + "score": 0.9, + "latex": "\\mathcal { N } _ { 0 } ^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 713, + 947, + 750, + 947, + 750, + 978, + 713, + 978 + ], + "score": 0.9, + "latex": "\\mathcal { N } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 702, + 1143, + 740, + 1143, + 740, + 1173, + 702, + 1173 + ], + "score": 0.9, + "latex": "\\mathcal { N } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 633, + 658, + 671, + 658, + 671, + 689, + 633, + 689 + ], + "score": 0.89, + "latex": "\\mathcal { N } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1308, + 1320, + 1346, + 1320, + 1346, + 1350, + 1308, + 1350 + ], + "score": 0.89, + "latex": "V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 702, + 613, + 732, + 613, + 732, + 642, + 702, + 642 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 368, + 778, + 406, + 778, + 406, + 808, + 368, + 808 + ], + "score": 0.88, + "latex": "\\mathcal { N } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 483, + 1323, + 554, + 1323, + 554, + 1351, + 483, + 1351 + ], + "score": 0.88, + "latex": "x = y" + }, + { + "category_id": 14, + "poly": [ + 585, + 1723, + 1114, + 1723, + 1114, + 1763, + 585, + 1763 + ], + "score": 0.88, + "latex": "\\mu ( x ) = p _ { X } ( x ) f ( x ) \\quad { \\mathrm { a n d } } \\quad \\nu ( y ) = p _ { Y } ( y ) g ( y )" + }, + { + "category_id": 13, + "poly": [ + 695, + 985, + 724, + 985, + 724, + 1015, + 695, + 1015 + ], + "score": 0.87, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 941, + 814, + 979, + 814, + 979, + 840, + 941, + 840 + ], + "score": 0.87, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 1371, + 309, + 1400, + 309, + 1400, + 339, + 1371, + 339 + ], + "score": 0.87, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1371, + 917, + 1400, + 917, + 1400, + 946, + 1371, + 946 + ], + "score": 0.87, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 808, + 887, + 839, + 887, + 839, + 916, + 808, + 916 + ], + "score": 0.87, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 407, + 612, + 448, + 612, + 448, + 639, + 407, + 639 + ], + "score": 0.87, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 863, + 308, + 904, + 308, + 904, + 336, + 863, + 336 + ], + "score": 0.87, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1152, + 1673, + 1190, + 1673, + 1190, + 1699, + 1152, + 1699 + ], + "score": 0.87, + "latex": "p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 876, + 344, + 904, + 344, + 904, + 373, + 876, + 373 + ], + "score": 0.86, + "latex": "\\eta _ { j }" + }, + { + "category_id": 13, + "poly": [ + 528, + 342, + 556, + 342, + 556, + 373, + 528, + 373 + ], + "score": 0.86, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 502, + 1766, + 522, + 1766, + 522, + 1797, + 502, + 1797 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1035, + 814, + 1070, + 814, + 1070, + 840, + 1035, + 840 + ], + "score": 0.86, + "latex": "p _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 1173, + 612, + 1263, + 612, + 1263, + 644, + 1173, + 644 + ], + "score": 0.86, + "latex": "\\mu \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1770, + 417, + 1770, + 417, + 1797, + 371, + 1797 + ], + "score": 0.86, + "latex": "x , y" + }, + { + "category_id": 13, + "poly": [ + 445, + 984, + 484, + 984, + 484, + 1015, + 445, + 1015 + ], + "score": 0.86, + "latex": "\\mathcal { N } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1248, + 1673, + 1283, + 1673, + 1283, + 1699, + 1248, + 1699 + ], + "score": 0.86, + "latex": "p _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 444, + 344, + 475, + 344, + 475, + 372, + 444, + 372 + ], + "score": 0.85, + "latex": "u _ { j }" + }, + { + "category_id": 13, + "poly": [ + 523, + 1257, + 546, + 1257, + 546, + 1286, + 523, + 1286 + ], + "score": 0.85, + "latex": "l ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1365, + 345, + 1392, + 345, + 1392, + 373, + 1365, + 373 + ], + "score": 0.84, + "latex": "\\eta _ { j }" + }, + { + "category_id": 13, + "poly": [ + 876, + 1290, + 905, + 1290, + 905, + 1317, + 876, + 1317 + ], + "score": 0.84, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1260, + 1055, + 1260, + 1055, + 1286, + 1019, + 1286 + ], + "score": 0.83, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 298, + 918, + 325, + 918, + 325, + 943, + 298, + 943 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 572, + 1770, + 591, + 1770, + 591, + 1797, + 572, + 1797 + ], + "score": 0.82, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 327, + 612, + 356, + 612, + 356, + 639, + 327, + 639 + ], + "score": 0.82, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1275, + 612, + 1362, + 612, + 1362, + 642, + 1275, + 642 + ], + "score": 0.81, + "latex": "\\nu \\in V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 782, + 309, + 812, + 309, + 812, + 336, + 782, + 336 + ], + "score": 0.8, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1051, + 613, + 1069, + 613, + 1069, + 639, + 1051, + 639 + ], + "score": 0.79, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 892, + 1770, + 910, + 1770, + 910, + 1797, + 892, + 1797 + ], + "score": 0.78, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 679, + 1295, + 699, + 1295, + 699, + 1317, + 679, + 1317 + ], + "score": 0.77, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 823, + 1771, + 842, + 1771, + 842, + 1792, + 823, + 1792 + ], + "score": 0.77, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 377, + 917, + 402, + 917, + 402, + 943, + 377, + 943 + ], + "score": 0.75, + "latex": "Y" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1610.0, + 643.0, + 1610.0, + 643.0, + 1641.0, + 296.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 1547.0, + 1402.0, + 1547.0, + 1402.0, + 1570.0, + 1379.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1228.0, + 1405.0, + 1228.0, + 1405.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1254.0, + 522.0, + 1254.0, + 522.0, + 1293.0, + 293.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 1254.0, + 1018.0, + 1254.0, + 1018.0, + 1293.0, + 547.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1254.0, + 1404.0, + 1254.0, + 1404.0, + 1293.0, + 1056.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1286.0, + 678.0, + 1286.0, + 678.0, + 1325.0, + 293.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1286.0, + 875.0, + 1286.0, + 875.0, + 1325.0, + 700.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1286.0, + 1405.0, + 1286.0, + 1405.0, + 1325.0, + 906.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1320.0, + 482.0, + 1320.0, + 482.0, + 1353.0, + 415.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1320.0, + 1307.0, + 1320.0, + 1307.0, + 1353.0, + 555.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 1320.0, + 1405.0, + 1320.0, + 1405.0, + 1353.0, + 1347.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1349.0, + 432.0, + 1349.0, + 432.0, + 1386.0, + 294.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1349.0, + 793.0, + 1349.0, + 793.0, + 1386.0, + 660.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 306.0, + 781.0, + 306.0, + 781.0, + 344.0, + 295.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 306.0, + 862.0, + 306.0, + 862.0, + 344.0, + 813.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 306.0, + 1370.0, + 306.0, + 1370.0, + 344.0, + 905.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 306.0, + 1404.0, + 306.0, + 1404.0, + 344.0, + 1401.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 333.0, + 443.0, + 333.0, + 443.0, + 379.0, + 291.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 333.0, + 527.0, + 333.0, + 527.0, + 379.0, + 476.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 333.0, + 875.0, + 333.0, + 875.0, + 379.0, + 557.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 333.0, + 1364.0, + 333.0, + 1364.0, + 379.0, + 905.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 333.0, + 1409.0, + 333.0, + 1409.0, + 379.0, + 1393.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 369.0, + 297.0, + 369.0, + 297.0, + 403.0, + 292.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 369.0, + 1096.0, + 369.0, + 1096.0, + 403.0, + 442.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 884.0, + 807.0, + 884.0, + 807.0, + 922.0, + 295.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 884.0, + 1407.0, + 884.0, + 1407.0, + 922.0, + 840.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 916.0, + 376.0, + 916.0, + 376.0, + 950.0, + 326.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 916.0, + 1370.0, + 916.0, + 1370.0, + 950.0, + 403.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 946.0, + 712.0, + 946.0, + 712.0, + 984.0, + 295.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 946.0, + 801.0, + 946.0, + 801.0, + 984.0, + 751.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 946.0, + 1159.0, + 946.0, + 1159.0, + 984.0, + 844.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 773.0, + 367.0, + 773.0, + 367.0, + 814.0, + 294.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 773.0, + 461.0, + 773.0, + 461.0, + 814.0, + 407.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 773.0, + 812.0, + 773.0, + 812.0, + 814.0, + 505.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 773.0, + 1405.0, + 773.0, + 1405.0, + 814.0, + 1274.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 804.0, + 490.0, + 804.0, + 490.0, + 844.0, + 295.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 804.0, + 940.0, + 804.0, + 940.0, + 844.0, + 572.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 804.0, + 1034.0, + 804.0, + 1034.0, + 844.0, + 980.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 804.0, + 1232.0, + 804.0, + 1232.0, + 844.0, + 1071.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 804.0, + 1406.0, + 804.0, + 1406.0, + 844.0, + 1313.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 839.0, + 527.0, + 839.0, + 527.0, + 872.0, + 296.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 215.0, + 296.0, + 215.0, + 296.0, + 272.0, + 288.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 215.0, + 1409.0, + 215.0, + 1409.0, + 272.0, + 491.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 259.0, + 742.0, + 259.0, + 742.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1139.0, + 701.0, + 1139.0, + 701.0, + 1179.0, + 295.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1139.0, + 1005.0, + 1139.0, + 1005.0, + 1179.0, + 741.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1139.0, + 1225.0, + 1139.0, + 1225.0, + 1179.0, + 1147.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1662.0, + 587.0, + 1662.0, + 587.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1662.0, + 746.0, + 1662.0, + 746.0, + 1704.0, + 691.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1662.0, + 1151.0, + 1662.0, + 1151.0, + 1704.0, + 847.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1662.0, + 1247.0, + 1662.0, + 1247.0, + 1704.0, + 1191.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1662.0, + 1405.0, + 1662.0, + 1405.0, + 1704.0, + 1284.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1697.0, + 410.0, + 1697.0, + 410.0, + 1732.0, + 292.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1759.0, + 370.0, + 1759.0, + 370.0, + 1803.0, + 293.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1759.0, + 501.0, + 1759.0, + 501.0, + 1803.0, + 418.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1759.0, + 571.0, + 1759.0, + 571.0, + 1803.0, + 523.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 1759.0, + 822.0, + 1759.0, + 822.0, + 1803.0, + 592.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1759.0, + 891.0, + 1759.0, + 891.0, + 1803.0, + 843.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1759.0, + 921.0, + 1759.0, + 921.0, + 1803.0, + 911.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 980.0, + 444.0, + 980.0, + 444.0, + 1021.0, + 295.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 980.0, + 694.0, + 980.0, + 694.0, + 1021.0, + 485.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 980.0, + 1154.0, + 980.0, + 1154.0, + 1021.0, + 725.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1806.0, + 1222.0, + 1806.0, + 1222.0, + 1848.0, + 293.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1806.0, + 1406.0, + 1806.0, + 1406.0, + 1848.0, + 1353.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1838.0, + 297.0, + 1838.0, + 297.0, + 1877.0, + 293.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1838.0, + 544.0, + 1838.0, + 544.0, + 1877.0, + 415.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 604.0, + 326.0, + 604.0, + 326.0, + 651.0, + 293.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 604.0, + 406.0, + 604.0, + 406.0, + 651.0, + 357.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 604.0, + 701.0, + 604.0, + 701.0, + 651.0, + 449.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 604.0, + 1050.0, + 604.0, + 1050.0, + 651.0, + 733.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 604.0, + 1172.0, + 604.0, + 1172.0, + 651.0, + 1070.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 604.0, + 1274.0, + 604.0, + 1274.0, + 651.0, + 1264.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 604.0, + 1373.0, + 604.0, + 1373.0, + 651.0, + 1363.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 651.0, + 632.0, + 651.0, + 632.0, + 699.0, + 294.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 651.0, + 721.0, + 651.0, + 721.0, + 699.0, + 672.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 651.0, + 802.0, + 651.0, + 802.0, + 699.0, + 764.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 651.0, + 929.0, + 651.0, + 929.0, + 699.0, + 879.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 651.0, + 1072.0, + 651.0, + 1072.0, + 699.0, + 1006.0, + 699.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 816, + 1404, + 816, + 1404, + 969, + 298, + 969 + ], + "score": 0.979 + }, + { + "category_id": 8, + "poly": [ + 589, + 1729, + 1111, + 1729, + 1111, + 1936, + 589, + 1936 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 295, + 228, + 1402, + 228, + 1402, + 296, + 295, + 296 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 298, + 1064, + 1397, + 1064, + 1397, + 1129, + 298, + 1129 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 698, + 1460, + 998, + 1460, + 998, + 1529, + 698, + 1529 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 300, + 1142, + 1395, + 1142, + 1395, + 1209, + 300, + 1209 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 293, + 1219, + 1400, + 1219, + 1400, + 1283, + 293, + 1283 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 294, + 308, + 1402, + 308, + 1402, + 372, + 294, + 372 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 297, + 439, + 1407, + 439, + 1407, + 500, + 297, + 500 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 662, + 1594, + 1037, + 1594, + 1037, + 1665, + 662, + 1665 + ], + "score": 0.939 + }, + { + "category_id": 8, + "poly": [ + 719, + 386, + 978, + 386, + 978, + 426, + 719, + 426 + ], + "score": 0.937 + }, + { + "category_id": 8, + "poly": [ + 684, + 714, + 1012, + 714, + 1012, + 757, + 684, + 757 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 776, + 575, + 919, + 575, + 919, + 611, + 776, + 611 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 296, + 1681, + 1035, + 1681, + 1035, + 1715, + 296, + 1715 + ], + "score": 0.934 + }, + { + "category_id": 8, + "poly": [ + 778, + 1998, + 920, + 1998, + 920, + 2034, + 778, + 2034 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 300, + 1547, + 819, + 1547, + 819, + 1580, + 300, + 1580 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 301, + 1409, + 763, + 1409, + 763, + 1443, + 301, + 1443 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 542, + 715, + 542, + 715, + 575, + 297, + 575 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 297, + 1951, + 916, + 1951, + 916, + 1985, + 297, + 1985 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 297, + 768, + 467, + 768, + 467, + 802, + 297, + 802 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 301, + 1007, + 813, + 1007, + 813, + 1039, + 301, + 1039 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 858, + 74, + 858, + 106, + 297, + 106 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 297, + 616, + 467, + 616, + 467, + 650, + 297, + 650 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 296, + 661, + 1004, + 661, + 1004, + 698, + 296, + 698 + ], + "score": 0.915 + }, + { + "category_id": 8, + "poly": [ + 808, + 503, + 891, + 503, + 891, + 533, + 808, + 533 + ], + "score": 0.91 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1606, + 1401, + 1606, + 1401, + 1637, + 1351, + 1637 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1470, + 1400, + 1470, + 1400, + 1502, + 1352, + 1502 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1352, + 576, + 1400, + 576, + 1400, + 606, + 1352, + 606 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1352, + 720, + 1400, + 720, + 1400, + 750, + 1352, + 750 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1351, + 2003, + 1401, + 2003, + 1401, + 2034, + 1351, + 2034 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1817, + 1401, + 1817, + 1401, + 1849, + 1351, + 1849 + ], + "score": 0.895 + }, + { + "category_id": 8, + "poly": [ + 680, + 1345, + 1019, + 1345, + 1019, + 1381, + 680, + 1381 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1345, + 1400, + 1345, + 1400, + 1374, + 1352, + 1374 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1352, + 388, + 1399, + 388, + 1399, + 419, + 1352, + 419 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1305, + 1400, + 1305, + 1400, + 1335, + 1352, + 1335 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1352, + 502, + 1400, + 502, + 1400, + 532, + 1352, + 532 + ], + "score": 0.885 + }, + { + "category_id": 8, + "poly": [ + 677, + 1300, + 1024, + 1300, + 1024, + 1340, + 677, + 1340 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.878 + }, + { + "category_id": 14, + "poly": [ + 584, + 1726, + 1112, + 1726, + 1112, + 1940, + 584, + 1940 + ], + "score": 0.94, + "latex": "\\begin{array} { l } { { \\displaystyle { \\cal A } _ { k j } = \\langle p _ { Y } g _ { k } , { \\cal N } ( p _ { X } f _ { j } ) \\rangle } } \\\\ { { \\displaystyle ~ = \\sum _ { x , y } \\frac { p _ { Y } ( y ) g _ { k } ( y ) p _ { Y | X } ( y | x ) p _ { X } ( x ) f _ { j } ( x ) } { p _ { Y } ( y ) } } } \\\\ { { \\displaystyle ~ = \\sum _ { x , y } p ( x , y ) g _ { k } ( y ) f _ { j } ( x ) = \\overline { { g _ { k } f _ { j } } } . } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 660, + 1593, + 1040, + 1593, + 1040, + 1667, + 660, + 1667 + ], + "score": 0.94, + "latex": "\\langle p _ { Y } g _ { k } , \\mathcal { N } ( p _ { X } f _ { j } ) \\rangle = \\sum _ { i } N _ { i j } L _ { k i } ." + }, + { + "category_id": 14, + "poly": [ + 697, + 1458, + 1002, + 1458, + 1002, + 1533, + 697, + 1533 + ], + "score": 0.94, + "latex": "\\mathcal { N } ( p _ { X } f _ { j } ) = \\sum _ { i } N _ { i j } p _ { Y } g _ { i } ." + }, + { + "category_id": 13, + "poly": [ + 692, + 260, + 806, + 260, + 806, + 295, + 692, + 295 + ], + "score": 0.93, + "latex": "f , f ^ { \\prime } , g , \\bar { g ^ { \\prime } }" + }, + { + "category_id": 13, + "poly": [ + 672, + 1174, + 806, + 1174, + 806, + 1208, + 672, + 1208 + ], + "score": 0.93, + "latex": "p _ { X } f _ { j } \\in V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 835, + 338, + 1071, + 338, + 1071, + 373, + 835, + 373 + ], + "score": 0.93, + "latex": "{ u } _ { i } \\bar { ( x ) } = p _ { X } ( x ) \\bar { a _ { i } } ( x )" + }, + { + "category_id": 14, + "poly": [ + 719, + 385, + 979, + 385, + 979, + 425, + 719, + 425 + ], + "score": 0.92, + "latex": "\\overline { { a _ { i } a _ { j } } } = \\langle u _ { i } , u _ { j } \\rangle _ { X } = 0 ," + }, + { + "category_id": 13, + "poly": [ + 298, + 470, + 416, + 470, + 416, + 503, + 298, + 503 + ], + "score": 0.92, + "latex": "a _ { 0 } ( x ) = 1" + }, + { + "category_id": 14, + "poly": [ + 672, + 1297, + 1024, + 1297, + 1024, + 1385, + 672, + 1385 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { K _ { i j } = \\langle p _ { X } f _ { i } , p _ { X } f _ { j } \\rangle _ { X } = \\overline { { f _ { i } f _ { j } } } , } \\\\ { L _ { i j } = \\langle p _ { Y } g _ { i } , p _ { Y } g _ { j } \\rangle _ { Y } = \\overline { { g _ { i } g _ { j } } } . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 684, + 712, + 1016, + 712, + 1016, + 755, + 684, + 755 + ], + "score": 0.91, + "latex": "\\overline { { b _ { i } b _ { j } } } = \\langle v _ { i } , v _ { j } \\rangle _ { Y } = 0 = \\overline { { b } } _ { i } \\overline { { b } } _ { j } ." + }, + { + "category_id": 14, + "poly": [ + 776, + 576, + 922, + 576, + 922, + 611, + 776, + 611 + ], + "score": 0.91, + "latex": "\\overline { { a _ { i } a _ { j } } } = \\overline { { a } } _ { i } \\overline { { a } } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 488, + 1412, + 532, + 1412, + 532, + 1445, + 488, + 1445 + ], + "score": 0.91, + "latex": "N _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 877, + 445, + 979, + 445, + 979, + 472, + 877, + 472 + ], + "score": 0.91, + "latex": "u _ { 0 } = p _ { X }" + }, + { + "category_id": 13, + "poly": [ + 372, + 618, + 459, + 618, + 459, + 650, + 372, + 650 + ], + "score": 0.9, + "latex": "i , j \\neq 0" + }, + { + "category_id": 13, + "poly": [ + 371, + 544, + 435, + 544, + 435, + 576, + 371, + 576 + ], + "score": 0.9, + "latex": "i \\neq 0" + }, + { + "category_id": 13, + "poly": [ + 858, + 1176, + 987, + 1176, + 987, + 1208, + 858, + 1208 + ], + "score": 0.9, + "latex": "p _ { Y } g _ { j } \\in V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 768, + 663, + 992, + 663, + 992, + 699, + 768, + 699 + ], + "score": 0.9, + "latex": "v _ { i } ( y ) = p _ { Y } ( y ) b _ { i } ( y )" + }, + { + "category_id": 13, + "poly": [ + 372, + 770, + 459, + 770, + 459, + 802, + 372, + 802 + ], + "score": 0.9, + "latex": "i , j \\neq 0" + }, + { + "category_id": 14, + "poly": [ + 777, + 1998, + 922, + 1998, + 922, + 2031, + 777, + 2031 + ], + "score": 0.9, + "latex": "N = L ^ { - 1 } A ." + }, + { + "category_id": 13, + "poly": [ + 770, + 1954, + 873, + 1954, + 873, + 1980, + 770, + 1980 + ], + "score": 0.9, + "latex": "A = L N" + }, + { + "category_id": 14, + "poly": [ + 807, + 500, + 892, + 500, + 892, + 533, + 807, + 533 + ], + "score": 0.89, + "latex": "\\overline { { a } } _ { i } = 0" + }, + { + "category_id": 13, + "poly": [ + 1185, + 1222, + 1224, + 1222, + 1224, + 1251, + 1185, + 1251 + ], + "score": 0.89, + "latex": "V _ { X }" + }, + { + "category_id": 13, + "poly": [ + 855, + 1144, + 982, + 1144, + 982, + 1176, + 855, + 1176 + ], + "score": 0.89, + "latex": "f _ { 1 } , \\ldots , f _ { k _ { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 1277, + 1221, + 1314, + 1221, + 1314, + 1251, + 1277, + 1251 + ], + "score": 0.88, + "latex": "V _ { Y }" + }, + { + "category_id": 13, + "poly": [ + 661, + 664, + 728, + 664, + 728, + 693, + 661, + 693 + ], + "score": 0.88, + "latex": "\\mathcal { N N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 633, + 1554, + 695, + 1554, + 695, + 1580, + 633, + 1580 + ], + "score": 0.87, + "latex": "p _ { Y } g _ { k }" + }, + { + "category_id": 13, + "poly": [ + 1042, + 1148, + 1168, + 1148, + 1168, + 1177, + 1042, + 1177 + ], + "score": 0.87, + "latex": "g _ { 1 } , \\ldots , g _ { k _ { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 1018, + 823, + 1130, + 823, + 1130, + 848, + 1018, + 848 + ], + "score": 0.87, + "latex": "v _ { 1 } , v _ { 2 } , \\ldots" + }, + { + "category_id": 13, + "poly": [ + 1143, + 315, + 1171, + 315, + 1171, + 339, + 1143, + 339 + ], + "score": 0.87, + "latex": "u _ { i }" + }, + { + "category_id": 13, + "poly": [ + 849, + 822, + 966, + 822, + 966, + 848, + 849, + 848 + ], + "score": 0.86, + "latex": "u _ { 1 } , u _ { 2 } , \\ldots" + }, + { + "category_id": 13, + "poly": [ + 681, + 342, + 708, + 342, + 708, + 369, + 681, + 369 + ], + "score": 0.85, + "latex": "a _ { i }" + }, + { + "category_id": 13, + "poly": [ + 368, + 308, + 439, + 308, + 439, + 337, + 368, + 337 + ], + "score": 0.85, + "latex": "\\mathcal { N } ^ { * } \\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 567, + 1412, + 596, + 1412, + 596, + 1438, + 567, + 1438 + ], + "score": 0.84, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1358, + 1065, + 1400, + 1065, + 1400, + 1093, + 1358, + 1093 + ], + "score": 0.84, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1277, + 1066, + 1306, + 1066, + 1306, + 1093, + 1277, + 1093 + ], + "score": 0.84, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 369, + 442, + 406, + 442, + 406, + 471, + 369, + 471 + ], + "score": 0.82, + "latex": "i , j" + }, + { + "category_id": 13, + "poly": [ + 873, + 238, + 890, + 238, + 890, + 262, + 873, + 262 + ], + "score": 0.78, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 493, + 474, + 512, + 474, + 512, + 497, + 493, + 497 + ], + "score": 0.73, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 529, + 1009, + 559, + 1009, + 559, + 1036, + 529, + 1036 + ], + "score": 0.69, + "latex": "\\mathcal { N }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1002.0, + 528.0, + 1002.0, + 528.0, + 1045.0, + 292.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1002.0, + 820.0, + 1002.0, + 820.0, + 1045.0, + 560.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 813.0, + 848.0, + 813.0, + 848.0, + 853.0, + 292.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 813.0, + 1017.0, + 813.0, + 1017.0, + 853.0, + 967.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 813.0, + 1407.0, + 813.0, + 1407.0, + 853.0, + 1131.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 848.0, + 1402.0, + 848.0, + 1402.0, + 881.0, + 296.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 874.0, + 1406.0, + 874.0, + 1406.0, + 914.0, + 293.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 904.0, + 1406.0, + 904.0, + 1406.0, + 943.0, + 293.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 937.0, + 434.0, + 937.0, + 434.0, + 968.0, + 294.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 225.0, + 872.0, + 225.0, + 872.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 225.0, + 1406.0, + 225.0, + 1406.0, + 269.0, + 891.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 691.0, + 261.0, + 691.0, + 297.0, + 295.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 261.0, + 817.0, + 261.0, + 817.0, + 297.0, + 807.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1062.0, + 1276.0, + 1062.0, + 1276.0, + 1102.0, + 294.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1062.0, + 1357.0, + 1062.0, + 1357.0, + 1102.0, + 1307.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1095.0, + 967.0, + 1095.0, + 967.0, + 1131.0, + 294.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1138.0, + 854.0, + 1138.0, + 854.0, + 1181.0, + 292.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1138.0, + 1041.0, + 1138.0, + 1041.0, + 1181.0, + 983.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1138.0, + 1401.0, + 1138.0, + 1401.0, + 1181.0, + 1169.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1172.0, + 671.0, + 1172.0, + 671.0, + 1211.0, + 292.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1172.0, + 857.0, + 1172.0, + 857.0, + 1211.0, + 807.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1172.0, + 998.0, + 1172.0, + 998.0, + 1211.0, + 988.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1219.0, + 1184.0, + 1219.0, + 1184.0, + 1255.0, + 295.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1219.0, + 1276.0, + 1219.0, + 1276.0, + 1255.0, + 1225.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1219.0, + 1403.0, + 1219.0, + 1403.0, + 1255.0, + 1315.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1251.0, + 603.0, + 1251.0, + 603.0, + 1283.0, + 296.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 307.0, + 367.0, + 307.0, + 367.0, + 343.0, + 297.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 307.0, + 1142.0, + 307.0, + 1142.0, + 343.0, + 440.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 307.0, + 1405.0, + 307.0, + 1405.0, + 343.0, + 1172.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 339.0, + 680.0, + 339.0, + 680.0, + 372.0, + 294.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 339.0, + 834.0, + 339.0, + 834.0, + 372.0, + 709.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 339.0, + 1358.0, + 339.0, + 1358.0, + 372.0, + 1072.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 436.0, + 368.0, + 436.0, + 368.0, + 477.0, + 292.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 436.0, + 876.0, + 436.0, + 876.0, + 477.0, + 407.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 436.0, + 1407.0, + 436.0, + 1407.0, + 477.0, + 980.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 467.0, + 297.0, + 467.0, + 297.0, + 505.0, + 292.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 467.0, + 492.0, + 467.0, + 492.0, + 505.0, + 417.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 467.0, + 534.0, + 467.0, + 534.0, + 505.0, + 513.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1681.0, + 1037.0, + 1681.0, + 1037.0, + 1717.0, + 296.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1544.0, + 632.0, + 1544.0, + 632.0, + 1584.0, + 295.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1544.0, + 822.0, + 1544.0, + 822.0, + 1584.0, + 696.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1401.0, + 487.0, + 1401.0, + 487.0, + 1453.0, + 292.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 1401.0, + 566.0, + 1401.0, + 566.0, + 1453.0, + 533.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1401.0, + 764.0, + 1401.0, + 764.0, + 1453.0, + 597.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 370.0, + 539.0, + 370.0, + 580.0, + 294.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 539.0, + 718.0, + 539.0, + 718.0, + 580.0, + 436.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1950.0, + 769.0, + 1950.0, + 769.0, + 1987.0, + 296.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1950.0, + 915.0, + 1950.0, + 915.0, + 1987.0, + 874.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 767.0, + 371.0, + 767.0, + 371.0, + 804.0, + 294.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 767.0, + 471.0, + 767.0, + 471.0, + 804.0, + 460.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 615.0, + 371.0, + 615.0, + 371.0, + 652.0, + 294.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 615.0, + 471.0, + 615.0, + 471.0, + 652.0, + 460.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 658.0, + 660.0, + 658.0, + 660.0, + 702.0, + 293.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 658.0, + 767.0, + 658.0, + 767.0, + 702.0, + 729.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 658.0, + 1003.0, + 658.0, + 1003.0, + 702.0, + 993.0, + 702.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 366, + 1405, + 366, + 1405, + 559, + 296, + 559 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 828, + 1405, + 828, + 1405, + 959, + 297, + 959 + ], + "score": 0.978 + }, + { + "category_id": 8, + "poly": [ + 610, + 1601, + 1085, + 1601, + 1085, + 1722, + 610, + 1722 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 620, + 985, + 1078, + 985, + 1078, + 1076, + 620, + 1076 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 690, + 1160, + 1010, + 1160, + 1010, + 1250, + 690, + 1250 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 298, + 660, + 1404, + 660, + 1404, + 725, + 298, + 725 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 606, + 1966, + 1088, + 1966, + 1088, + 2041, + 606, + 2041 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 594, + 1813, + 1101, + 1813, + 1101, + 1890, + 594, + 1890 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 701, + 1441, + 997, + 1441, + 997, + 1515, + 701, + 1515 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 726, + 1326, + 972, + 1326, + 972, + 1367, + 726, + 1367 + ], + "score": 0.939 + }, + { + "category_id": 8, + "poly": [ + 563, + 581, + 1133, + 581, + 1133, + 623, + 563, + 623 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 295, + 1097, + 1169, + 1097, + 1169, + 1132, + 295, + 1132 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 297, + 1555, + 457, + 1555, + 457, + 1586, + 297, + 1586 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 298, + 1270, + 691, + 1270, + 691, + 1302, + 298, + 1302 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 298, + 1385, + 808, + 1385, + 808, + 1418, + 298, + 1418 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 297, + 1760, + 905, + 1760, + 905, + 1794, + 297, + 1794 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 297, + 1914, + 739, + 1914, + 739, + 1946, + 297, + 1946 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 227, + 1161, + 227, + 1161, + 266, + 297, + 266 + ], + "score": 0.927 + }, + { + "category_id": 8, + "poly": [ + 759, + 288, + 939, + 288, + 939, + 328, + 759, + 328 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 299, + 768, + 509, + 768, + 509, + 800, + 299, + 800 + ], + "score": 0.917 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1328, + 1400, + 1328, + 1400, + 1360, + 1352, + 1360 + ], + "score": 0.909 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1012, + 1400, + 1012, + 1400, + 1043, + 1351, + 1043 + ], + "score": 0.908 + }, + { + "category_id": 9, + "poly": [ + 1351, + 587, + 1400, + 587, + 1400, + 619, + 1351, + 619 + ], + "score": 0.908 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1827, + 1401, + 1827, + 1401, + 1858, + 1352, + 1858 + ], + "score": 0.907 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1649, + 1401, + 1649, + 1401, + 1680, + 1352, + 1680 + ], + "score": 0.906 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1452, + 1401, + 1452, + 1401, + 1483, + 1351, + 1483 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1977, + 1401, + 1977, + 1401, + 2008, + 1351, + 2008 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1351, + 295, + 1400, + 295, + 1400, + 326, + 1351, + 326 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1187, + 1400, + 1187, + 1400, + 1219, + 1352, + 1219 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 866, + 2087, + 866, + 2113, + 836, + 2113 + ], + "score": 0.873 + }, + { + "category_id": 14, + "poly": [ + 621, + 980, + 1076, + 980, + 1076, + 1076, + 621, + 1076 + ], + "score": 0.95, + "latex": "\\delta _ { y } ( y ^ { \\prime } ) = p _ { Y } ( y ^ { \\prime } ) \\sum _ { i = 1 } ^ { k _ { 0 } } ( \\delta _ { y } ) _ { i } g _ { i } ( y ^ { \\prime } ) + r ( y ^ { \\prime } ) ," + }, + { + "category_id": 14, + "poly": [ + 688, + 1154, + 1009, + 1154, + 1009, + 1250, + 688, + 1250 + ], + "score": 0.94, + "latex": "\\langle p _ { Y } g _ { j } , \\delta _ { y } \\rangle _ { Y } = \\sum _ { i = 1 } ^ { k _ { 0 } } ( \\delta _ { y } ) _ { i } L _ { j i } ," + }, + { + "category_id": 14, + "poly": [ + 608, + 1965, + 1090, + 1965, + 1090, + 2043, + 608, + 2043 + ], + "score": 0.94, + "latex": "p _ { Y | X } ( y | x ) \\approx \\sum _ { i j k } N _ { k i } ( K ^ { - 1 } ) _ { i j } f _ { j } ( x ) g _ { k } ( y ) ." + }, + { + "category_id": 14, + "poly": [ + 700, + 1440, + 1000, + 1440, + 1000, + 1518, + 700, + 1518 + ], + "score": 0.94, + "latex": "( \\delta _ { y } ) _ { i } = \\sum _ { j } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) ." + }, + { + "category_id": 13, + "poly": [ + 386, + 693, + 512, + 693, + 512, + 725, + 386, + 725 + ], + "score": 0.93, + "latex": "f _ { 1 } , \\ldots , f _ { k _ { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1098, + 545, + 1098, + 545, + 1132, + 373, + 1132 + ], + "score": 0.93, + "latex": "\\langle r , p _ { Y } \\delta _ { i } \\rangle _ { Y } = 0" + }, + { + "category_id": 14, + "poly": [ + 611, + 1599, + 1088, + 1599, + 1088, + 1725, + 611, + 1725 + ], + "score": 0.93, + "latex": "\\begin{array} { l } { { \\displaystyle p _ { X | Y } ( x | y ) = \\mathcal { N } ^ { * } ( \\delta _ { y } ) ( x ) \\approx \\mathcal { N } _ { 0 } ^ { * } ( \\delta _ { y } ) ( x ) } } \\\\ { { \\displaystyle ~ = \\sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) f _ { k } ( x ) . } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 595, + 1813, + 1105, + 1813, + 1105, + 1892, + 595, + 1892 + ], + "score": 0.93, + "latex": "\\overline { { { x } } } = \\sum _ { i j k } N _ { k i } ^ { * } ( L ^ { - 1 } ) _ { i j } g _ { j } ( y ) \\sum _ { x } p _ { X } ( x ) x f _ { k } ( x ) ." + }, + { + "category_id": 13, + "poly": [ + 1302, + 860, + 1374, + 860, + 1374, + 895, + 1302, + 895 + ], + "score": 0.92, + "latex": "\\delta _ { y } ( \\bar { y } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 849, + 495, + 970, + 495, + 970, + 528, + 849, + 528 + ], + "score": 0.92, + "latex": "\\mathrm { T r } \\left( N ^ { * } N \\right)" + }, + { + "category_id": 13, + "poly": [ + 1155, + 924, + 1215, + 924, + 1215, + 959, + 1155, + 959 + ], + "score": 0.92, + "latex": "( \\delta _ { y } ) _ { j }" + }, + { + "category_id": 13, + "poly": [ + 932, + 860, + 1215, + 860, + 1215, + 897, + 932, + 897 + ], + "score": 0.92, + "latex": "p _ { X | Y } ( x | y ) = \\mathcal { N } ^ { * } ( \\delta _ { y } ) ( x )" + }, + { + "category_id": 13, + "poly": [ + 388, + 893, + 470, + 893, + 470, + 925, + 388, + 925 + ], + "score": 0.91, + "latex": "y = y ^ { \\prime }" + }, + { + "category_id": 14, + "poly": [ + 724, + 1325, + 975, + 1325, + 975, + 1365, + 724, + 1365 + ], + "score": 0.91, + "latex": "\\langle p _ { Y } g _ { j } , \\delta _ { y } \\rangle _ { Y } = g _ { j } ( y ) ." + }, + { + "category_id": 14, + "poly": [ + 757, + 288, + 943, + 288, + 943, + 327, + 757, + 327 + ], + "score": 0.91, + "latex": "N ^ { * } = K ^ { - 1 } A ^ { \\top } ." + }, + { + "category_id": 13, + "poly": [ + 488, + 232, + 533, + 232, + 533, + 268, + 488, + 268 + ], + "score": 0.91, + "latex": "N _ { i j } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 700, + 400, + 968, + 400, + 968, + 432, + 700, + 432 + ], + "score": 0.91, + "latex": "N ^ { * } N = K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A" + }, + { + "category_id": 14, + "poly": [ + 564, + 580, + 1134, + 580, + 1134, + 623, + 564, + 623 + ], + "score": 0.9, + "latex": "C = k _ { 0 } - \\operatorname { T r } \\left( N ^ { \\ast } N \\right) = k _ { 0 } - \\operatorname { T r } \\left( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A \\right) ." + }, + { + "category_id": 13, + "poly": [ + 997, + 465, + 1063, + 465, + 1063, + 493, + 997, + 493 + ], + "score": 0.9, + "latex": "N ^ { * } N" + }, + { + "category_id": 13, + "poly": [ + 1036, + 693, + 1102, + 693, + 1102, + 720, + 1036, + 720 + ], + "score": 0.9, + "latex": "N ^ { * } N" + }, + { + "category_id": 13, + "poly": [ + 710, + 929, + 903, + 929, + 903, + 958, + 710, + 958 + ], + "score": 0.9, + "latex": "p _ { Y } g _ { 1 } , \\ldots , p _ { Y } g _ { k _ { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 625, + 1387, + 653, + 1387, + 653, + 1419, + 625, + 1419 + ], + "score": 0.89, + "latex": "\\delta _ { y }" + }, + { + "category_id": 13, + "poly": [ + 327, + 465, + 394, + 465, + 394, + 493, + 327, + 493 + ], + "score": 0.89, + "latex": "N ^ { * } N" + }, + { + "category_id": 13, + "poly": [ + 479, + 527, + 507, + 527, + 507, + 560, + 479, + 560 + ], + "score": 0.89, + "latex": "f _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1075, + 371, + 1103, + 371, + 1103, + 403, + 1075, + 403 + ], + "score": 0.89, + "latex": "f _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1295, + 404, + 1323, + 404, + 1323, + 438, + 1295, + 438 + ], + "score": 0.88, + "latex": "f _ { j }" + }, + { + "category_id": 13, + "poly": [ + 983, + 1104, + 1043, + 1104, + 1043, + 1133, + 983, + 1133 + ], + "score": 0.88, + "latex": "p _ { Y } g _ { j }" + }, + { + "category_id": 13, + "poly": [ + 432, + 927, + 460, + 927, + 460, + 958, + 432, + 958 + ], + "score": 0.88, + "latex": "\\delta _ { y }" + }, + { + "category_id": 13, + "poly": [ + 639, + 436, + 670, + 436, + 670, + 465, + 639, + 465 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 567, + 231, + 609, + 231, + 609, + 258, + 567, + 258 + ], + "score": 0.87, + "latex": "\\mathcal { N } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 648, + 527, + 685, + 527, + 685, + 557, + 648, + 557 + ], + "score": 0.87, + "latex": "i , j" + }, + { + "category_id": 13, + "poly": [ + 1154, + 374, + 1182, + 374, + 1182, + 404, + 1154, + 404 + ], + "score": 0.86, + "latex": "g _ { j }" + }, + { + "category_id": 13, + "poly": [ + 849, + 1763, + 876, + 1763, + 876, + 1789, + 849, + 1789 + ], + "score": 0.86, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 524, + 1916, + 548, + 1916, + 548, + 1941, + 524, + 1941 + ], + "score": 0.85, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1372, + 407, + 1400, + 407, + 1400, + 438, + 1372, + 438 + ], + "score": 0.84, + "latex": "g _ { j }" + }, + { + "category_id": 13, + "poly": [ + 544, + 529, + 569, + 529, + 569, + 558, + 544, + 558 + ], + "score": 0.84, + "latex": "g _ { i }" + }, + { + "category_id": 13, + "poly": [ + 950, + 232, + 979, + 232, + 979, + 258, + 950, + 258 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 770, + 370, + 800, + 370, + 800, + 397, + 770, + 397 + ], + "score": 0.81, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 594, + 832, + 619, + 832, + 619, + 859, + 594, + 859 + ], + "score": 0.8, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1029, + 232, + 1054, + 232, + 1054, + 258, + 1029, + 258 + ], + "score": 0.79, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 781, + 866, + 800, + 866, + 800, + 888, + 781, + 888 + ], + "score": 0.78, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 614, + 1923, + 631, + 1923, + 631, + 1941, + 614, + 1941 + ], + "score": 0.76, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 623, + 1101, + 635, + 1101, + 635, + 1126, + 623, + 1126 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 298, + 870, + 314, + 870, + 314, + 893, + 298, + 893 + ], + "score": 0.71, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 819, + 440, + 889, + 440, + 889, + 469, + 819, + 469 + ], + "score": 0.7, + "latex": "u _ { j } , v _ { j }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 765.0, + 512.0, + 765.0, + 512.0, + 804.0, + 294.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 366.0, + 769.0, + 366.0, + 769.0, + 405.0, + 293.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 366.0, + 1074.0, + 366.0, + 1074.0, + 405.0, + 801.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 366.0, + 1153.0, + 366.0, + 1153.0, + 405.0, + 1104.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 366.0, + 1403.0, + 366.0, + 1403.0, + 405.0, + 1183.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 393.0, + 699.0, + 393.0, + 699.0, + 446.0, + 289.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 393.0, + 1294.0, + 393.0, + 1294.0, + 446.0, + 969.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 393.0, + 1371.0, + 393.0, + 1371.0, + 446.0, + 1324.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 393.0, + 1406.0, + 393.0, + 1406.0, + 446.0, + 1401.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 432.0, + 638.0, + 432.0, + 638.0, + 473.0, + 293.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 432.0, + 818.0, + 432.0, + 818.0, + 473.0, + 671.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 432.0, + 1406.0, + 432.0, + 1406.0, + 473.0, + 890.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 461.0, + 326.0, + 461.0, + 326.0, + 500.0, + 293.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 461.0, + 996.0, + 461.0, + 996.0, + 500.0, + 395.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 461.0, + 1405.0, + 461.0, + 1405.0, + 500.0, + 1064.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 494.0, + 848.0, + 494.0, + 848.0, + 531.0, + 294.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 494.0, + 1405.0, + 494.0, + 1405.0, + 531.0, + 971.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 525.0, + 478.0, + 525.0, + 478.0, + 561.0, + 294.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 525.0, + 543.0, + 525.0, + 543.0, + 561.0, + 508.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 525.0, + 647.0, + 525.0, + 647.0, + 561.0, + 570.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 525.0, + 1076.0, + 525.0, + 1076.0, + 561.0, + 686.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 827.0, + 593.0, + 827.0, + 593.0, + 866.0, + 294.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 827.0, + 1405.0, + 827.0, + 1405.0, + 866.0, + 620.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 859.0, + 297.0, + 859.0, + 297.0, + 897.0, + 291.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 859.0, + 780.0, + 859.0, + 780.0, + 897.0, + 315.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 859.0, + 931.0, + 859.0, + 931.0, + 897.0, + 801.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 859.0, + 1301.0, + 859.0, + 1301.0, + 897.0, + 1216.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1375.0, + 859.0, + 1406.0, + 859.0, + 1406.0, + 897.0, + 1375.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 894.0, + 387.0, + 894.0, + 387.0, + 927.0, + 295.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 894.0, + 1405.0, + 894.0, + 1405.0, + 927.0, + 471.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 924.0, + 431.0, + 924.0, + 431.0, + 961.0, + 294.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 924.0, + 709.0, + 924.0, + 709.0, + 961.0, + 461.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 924.0, + 1154.0, + 924.0, + 1154.0, + 961.0, + 904.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 924.0, + 1328.0, + 924.0, + 1328.0, + 961.0, + 1216.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 696.0, + 297.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 693.0, + 385.0, + 693.0, + 385.0, + 726.0, + 294.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 693.0, + 1035.0, + 693.0, + 1035.0, + 726.0, + 513.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 693.0, + 1112.0, + 693.0, + 1112.0, + 726.0, + 1103.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1094.0, + 372.0, + 1094.0, + 372.0, + 1136.0, + 296.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1094.0, + 622.0, + 1094.0, + 622.0, + 1136.0, + 546.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1094.0, + 982.0, + 1094.0, + 982.0, + 1136.0, + 636.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 1094.0, + 1170.0, + 1094.0, + 1170.0, + 1136.0, + 1044.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1551.0, + 461.0, + 1551.0, + 461.0, + 1589.0, + 293.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1266.0, + 694.0, + 1266.0, + 694.0, + 1305.0, + 295.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1377.0, + 624.0, + 1377.0, + 624.0, + 1428.0, + 292.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1377.0, + 809.0, + 1377.0, + 809.0, + 1428.0, + 654.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1759.0, + 848.0, + 1759.0, + 848.0, + 1796.0, + 295.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1759.0, + 907.0, + 1759.0, + 907.0, + 1796.0, + 877.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 523.0, + 1910.0, + 523.0, + 1950.0, + 295.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1910.0, + 613.0, + 1910.0, + 613.0, + 1950.0, + 549.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1910.0, + 739.0, + 1910.0, + 739.0, + 1950.0, + 632.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 487.0, + 224.0, + 487.0, + 273.0, + 293.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 224.0, + 566.0, + 224.0, + 566.0, + 273.0, + 534.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 224.0, + 949.0, + 224.0, + 949.0, + 273.0, + 610.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 224.0, + 1028.0, + 224.0, + 1028.0, + 273.0, + 980.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 224.0, + 1165.0, + 224.0, + 1165.0, + 273.0, + 1055.0, + 273.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 294, + 1405, + 294, + 1405, + 425, + 296, + 425 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 879, + 1405, + 879, + 1405, + 976, + 298, + 976 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1406, + 1403, + 1406, + 1403, + 1499, + 297, + 1499 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1513, + 1404, + 1513, + 1404, + 1608, + 298, + 1608 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 773, + 1405, + 773, + 1405, + 866, + 298, + 866 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 536, + 1625, + 1158, + 1625, + 1158, + 1717, + 536, + 1717 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 576, + 1290, + 1119, + 1290, + 1119, + 1397, + 576, + 1397 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 696, + 1404, + 696, + 1404, + 758, + 296, + 758 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 295, + 1733, + 1404, + 1733, + 1404, + 1795, + 295, + 1795 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 617, + 1810, + 1083, + 1810, + 1083, + 1854, + 617, + 1854 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 294, + 520, + 1399, + 520, + 1399, + 583, + 294, + 583 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 486, + 425, + 1208, + 425, + 1208, + 501, + 486, + 501 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 776, + 654, + 920, + 654, + 920, + 689, + 776, + 689 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 298, + 1912, + 1405, + 1912, + 1405, + 1978, + 298, + 1978 + ], + "score": 0.93 + }, + { + "category_id": 8, + "poly": [ + 777, + 582, + 923, + 582, + 923, + 617, + 777, + 617 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 859, + 73, + 859, + 106, + 297, + 106 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 297, + 1269, + 416, + 1269, + 416, + 1300, + 297, + 1300 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 297, + 988, + 1024, + 988, + 1024, + 1022, + 297, + 1022 + ], + "score": 0.917 + }, + { + "category_id": 9, + "poly": [ + 1351, + 655, + 1400, + 655, + 1400, + 685, + 1351, + 685 + ], + "score": 0.907 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1328, + 1400, + 1328, + 1400, + 1360, + 1351, + 1360 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 693, + 226, + 693, + 262, + 300, + 262 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 292, + 1866, + 1401, + 1866, + 1401, + 1901, + 292, + 1901 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1654, + 1400, + 1654, + 1400, + 1685, + 1351, + 1685 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1817, + 1400, + 1817, + 1400, + 1848, + 1351, + 1848 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1186, + 1400, + 1186, + 1400, + 1216, + 1351, + 1216 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1352, + 446, + 1400, + 446, + 1400, + 476, + 1352, + 476 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1352, + 583, + 1400, + 583, + 1400, + 613, + 1352, + 613 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1075, + 1400, + 1075, + 1400, + 1105, + 1352, + 1105 + ], + "score": 0.874 + }, + { + "category_id": 1, + "poly": [ + 297, + 622, + 702, + 622, + 702, + 658, + 297, + 658 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2112, + 836, + 2112 + ], + "score": 0.869 + }, + { + "category_id": 8, + "poly": [ + 827, + 1035, + 1311, + 1035, + 1311, + 1145, + 827, + 1145 + ], + "score": 0.651 + }, + { + "category_id": 8, + "poly": [ + 337, + 1034, + 1315, + 1034, + 1315, + 1256, + 337, + 1256 + ], + "score": 0.314 + }, + { + "category_id": 8, + "poly": [ + 327, + 1034, + 1314, + 1034, + 1314, + 1144, + 327, + 1144 + ], + "score": 0.175 + }, + { + "category_id": 8, + "poly": [ + 339, + 1034, + 791, + 1034, + 791, + 1256, + 339, + 1256 + ], + "score": 0.11 + }, + { + "category_id": 14, + "poly": [ + 576, + 1289, + 1123, + 1289, + 1123, + 1400, + 576, + 1400 + ], + "score": 0.95, + "latex": "M = K ^ { - 1 } A ^ { \\top } L ^ { - 1 } A = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } ( 1 - \\gamma ^ { 2 } ) } } \\\\ { { 0 } } & { { \\gamma } } & { { 0 } } \\\\ { { 0 } } & { { 0 } } & { { \\gamma ^ { 2 } } } \\end{array} \\right) ." + }, + { + "category_id": 14, + "poly": [ + 336, + 1034, + 1319, + 1034, + 1319, + 1261, + 336, + 1261 + ], + "score": 0.95, + "latex": "\\begin{array} { l c c } { { K = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } } } & { { 0 } } & { { 3 \\tau ^ { 4 } } } \\end{array} \\right) } } & { { L = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { 3 ( \\tau ^ { 2 } + \\sigma ^ { 2 } ) ^ { 2 } } } \\end{array} \\right) } } \\\\ { { A = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { \\tau ^ { 2 } ( \\sigma ^ { 2 } + 3 \\tau ^ { 2 } ) } } \\end{array} \\right) . } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 538, + 1620, + 1161, + 1620, + 1161, + 1720, + 538, + 1720 + ], + "score": 0.94, + "latex": "\\mathcal { N } _ { 0 } ^ { * } ( \\delta _ { y } ) ( x ) = p _ { X } ( x ) p _ { Y } ( y ) \\sum _ { j , k = 0 } ^ { 2 } ( K ^ { - 1 } A ^ { \\top } L ^ { - 1 } ) _ { k j } y ^ { j } x ^ { k } ." + }, + { + "category_id": 13, + "poly": [ + 369, + 296, + 449, + 296, + 449, + 330, + 369, + 330 + ], + "score": 0.94, + "latex": "p ( x , y )" + }, + { + "category_id": 14, + "poly": [ + 489, + 421, + 1211, + 421, + 1211, + 500, + 489, + 500 + ], + "score": 0.93, + "latex": "p _ { X | Y } ( x | y ) \\propto \\exp \\left( - \\frac { ( x - \\gamma y ) ^ { 2 } } { 2 \\tau ^ { 2 } ( 1 - \\gamma ) } \\right) , \\quad \\mathrm { w h e r e } \\quad \\gamma = \\frac { \\tau ^ { 2 } } { \\sigma ^ { 2 } + \\tau ^ { 2 } } ." + }, + { + "category_id": 13, + "poly": [ + 443, + 1436, + 566, + 1436, + 566, + 1469, + 443, + 1469 + ], + "score": 0.92, + "latex": "( - \\tau ^ { 2 } , 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 453, + 1576, + 534, + 1576, + 534, + 1606, + 453, + 1606 + ], + "score": 0.92, + "latex": "k _ { 0 } = 3" + }, + { + "category_id": 13, + "poly": [ + 878, + 327, + 1170, + 327, + 1170, + 366, + 878, + 366 + ], + "score": 0.92, + "latex": "p ( x ) \\ \\propto \\ \\exp { \\left( - x ^ { 2 } / 2 \\tau ^ { 2 } \\right) }" + }, + { + "category_id": 14, + "poly": [ + 616, + 1810, + 1080, + 1810, + 1080, + 1852, + 616, + 1852 + ], + "score": 0.92, + "latex": "{ \\overline { { x } } } = \\gamma y \\quad { \\mathrm { a n d } } \\quad { \\overline { { x ^ { 2 } } } } = \\gamma ^ { 2 } y ^ { 2 } + ( 1 - \\gamma ) \\tau ^ { 2 } ," + }, + { + "category_id": 13, + "poly": [ + 1087, + 943, + 1168, + 943, + 1168, + 973, + 1087, + 973 + ], + "score": 0.91, + "latex": "k _ { 0 } = 3" + }, + { + "category_id": 13, + "poly": [ + 940, + 1916, + 1017, + 1916, + 1017, + 1945, + 940, + 1945 + ], + "score": 0.91, + "latex": "k _ { 0 } - 1" + }, + { + "category_id": 13, + "poly": [ + 1137, + 1436, + 1261, + 1436, + 1261, + 1470, + 1137, + 1470 + ], + "score": 0.91, + "latex": "\\bar { u } _ { 0 } ( x ) = 1" + }, + { + "category_id": 13, + "poly": [ + 346, + 1468, + 539, + 1468, + 539, + 1501, + 346, + 1501 + ], + "score": 0.91, + "latex": "u _ { 2 } ( x ) = x ^ { 2 } - \\tau ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 299, + 1436, + 387, + 1436, + 387, + 1469, + 299, + 1469 + ], + "score": 0.9, + "latex": "( 0 , 1 , \\bar { 0 } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 805, + 378, + 805, + 378, + 836, + 297, + 836 + ], + "score": 0.9, + "latex": "Y = y" + }, + { + "category_id": 13, + "poly": [ + 296, + 362, + 688, + 362, + 688, + 402, + 296, + 402 + ], + "score": 0.9, + "latex": "\\underline { { { p } } } ( y | x ) ~ \\propto ~ \\exp \\left( - ( y - x ) ^ { 2 } / 2 \\sigma ^ { 2 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 297, + 625, + 504, + 625, + 504, + 656, + 297, + 656 + ], + "score": 0.9, + "latex": "n = 0 , \\ldots , k _ { 0 } - 1" + }, + { + "category_id": 13, + "poly": [ + 1305, + 1406, + 1392, + 1406, + 1392, + 1439, + 1305, + 1439 + ], + "score": 0.9, + "latex": "( 1 , 0 , 0 )" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1873, + 1177, + 1873, + 1177, + 1903, + 1115, + 1903 + ], + "score": 0.89, + "latex": "p _ { X | Y }" + }, + { + "category_id": 14, + "poly": [ + 774, + 579, + 923, + 579, + 923, + 617, + 774, + 617 + ], + "score": 0.89, + "latex": "f _ { n } ( x ) = x ^ { n } ," + }, + { + "category_id": 13, + "poly": [ + 1275, + 1439, + 1401, + 1439, + 1401, + 1470, + 1275, + 1470 + ], + "score": 0.89, + "latex": "u _ { 1 } ( x ) = x" + }, + { + "category_id": 13, + "poly": [ + 1169, + 523, + 1198, + 523, + 1198, + 553, + 1169, + 553 + ], + "score": 0.89, + "latex": "k _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 722, + 944, + 753, + 944, + 753, + 975, + 722, + 975 + ], + "score": 0.89, + "latex": "f _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1310, + 1916, + 1339, + 1916, + 1339, + 1945, + 1310, + 1945 + ], + "score": 0.88, + "latex": "k _ { 0 }" + }, + { + "category_id": 14, + "poly": [ + 776, + 652, + 923, + 652, + 923, + 690, + 776, + 690 + ], + "score": 0.88, + "latex": "g _ { n } ( y ) = y ^ { n } ." + }, + { + "category_id": 13, + "poly": [ + 930, + 918, + 964, + 918, + 964, + 943, + 930, + 943 + ], + "score": 0.87, + "latex": "u _ { n }" + }, + { + "category_id": 13, + "poly": [ + 921, + 775, + 995, + 775, + 995, + 803, + 921, + 803 + ], + "score": 0.86, + "latex": "( n = 1" + }, + { + "category_id": 13, + "poly": [ + 804, + 946, + 834, + 946, + 834, + 975, + 804, + 975 + ], + "score": 0.86, + "latex": "g _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1016, + 917, + 1047, + 917, + 1047, + 943, + 1016, + 943 + ], + "score": 0.86, + "latex": "v _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1148, + 1577, + 1175, + 1577, + 1175, + 1602, + 1148, + 1602 + ], + "score": 0.85, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1031, + 1870, + 1058, + 1870, + 1058, + 1895, + 1031, + 1895 + ], + "score": 0.85, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 616, + 836, + 643, + 836, + 643, + 862, + 616, + 862 + ], + "score": 0.85, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 934, + 1735, + 962, + 1735, + 962, + 1761, + 934, + 1761 + ], + "score": 0.84, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1047, + 1516, + 1077, + 1516, + 1077, + 1542, + 1047, + 1542 + ], + "score": 0.84, + "latex": "\\mathcal { N }" + }, + { + "category_id": 13, + "poly": [ + 1305, + 775, + 1333, + 775, + 1333, + 801, + 1305, + 801 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 513, + 1408, + 545, + 1408, + 545, + 1433, + 513, + 1433 + ], + "score": 0.82, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 788, + 1577, + 812, + 1577, + 812, + 1603, + 788, + 1603 + ], + "score": 0.82, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1117, + 703, + 1135, + 703, + 1135, + 729, + 1117, + 729 + ], + "score": 0.81, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 391, + 1763, + 422, + 1763, + 422, + 1791, + 391, + 1791 + ], + "score": 0.81, + "latex": "x ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 952, + 371, + 972, + 371, + 972, + 393, + 952, + 393 + ], + "score": 0.81, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 727, + 1581, + 744, + 1581, + 744, + 1608, + 727, + 1608 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1373, + 523, + 1401, + 523, + 1401, + 549, + 1373, + 549 + ], + "score": 0.8, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 802, + 372, + 821, + 372, + 821, + 398, + 802, + 398 + ], + "score": 0.79, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 491, + 805, + 517, + 805, + 517, + 832, + 491, + 832 + ], + "score": 0.79, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 491, + 1769, + 509, + 1769, + 509, + 1790, + 491, + 1790 + ], + "score": 0.76, + "latex": "x" + }, + { + "category_id": 14, + "poly": [ + 827, + 1034, + 1316, + 1034, + 1316, + 1146, + 827, + 1146 + ], + "score": 0.72, + "latex": "L = \\left( \\begin{array} { c c c } { { 1 } } & { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } \\\\ { { 0 } } & { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } \\\\ { { \\tau ^ { 2 } + \\sigma ^ { 2 } } } & { { 0 } } & { { 3 ( \\tau ^ { 2 } + \\sigma ^ { 2 } ) ^ { 2 } } } \\end{array} \\right)" + }, + { + "category_id": 13, + "poly": [ + 456, + 918, + 477, + 918, + 477, + 940, + 456, + 940 + ], + "score": 0.72, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 298, + 1773, + 316, + 1773, + 316, + 1791, + 298, + 1791 + ], + "score": 0.67, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 665, + 625, + 690, + 625, + 690, + 652, + 665, + 652 + ], + "score": 0.34, + "latex": "Y" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 222.0, + 697.0, + 222.0, + 697.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 295.0, + 368.0, + 295.0, + 368.0, + 332.0, + 295.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 295.0, + 1406.0, + 295.0, + 1406.0, + 332.0, + 450.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 327.0, + 877.0, + 327.0, + 877.0, + 364.0, + 293.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 327.0, + 1405.0, + 327.0, + 1405.0, + 364.0, + 1171.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 361.0, + 295.0, + 361.0, + 295.0, + 405.0, + 290.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 361.0, + 801.0, + 361.0, + 801.0, + 405.0, + 689.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 361.0, + 951.0, + 361.0, + 951.0, + 405.0, + 822.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 361.0, + 1407.0, + 361.0, + 1407.0, + 405.0, + 973.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 390.0, + 423.0, + 390.0, + 423.0, + 433.0, + 292.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 881.0, + 1402.0, + 881.0, + 1402.0, + 915.0, + 295.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 913.0, + 455.0, + 913.0, + 455.0, + 947.0, + 295.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 913.0, + 929.0, + 913.0, + 929.0, + 947.0, + 478.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 913.0, + 1015.0, + 913.0, + 1015.0, + 947.0, + 965.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 913.0, + 1405.0, + 913.0, + 1405.0, + 947.0, + 1048.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 944.0, + 721.0, + 944.0, + 721.0, + 978.0, + 296.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 944.0, + 803.0, + 944.0, + 803.0, + 978.0, + 754.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 944.0, + 1086.0, + 944.0, + 1086.0, + 978.0, + 835.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 944.0, + 1264.0, + 944.0, + 1264.0, + 978.0, + 1169.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1399.0, + 512.0, + 1399.0, + 512.0, + 1444.0, + 291.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1399.0, + 1304.0, + 1399.0, + 1304.0, + 1444.0, + 546.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1399.0, + 1405.0, + 1399.0, + 1405.0, + 1444.0, + 1393.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1431.0, + 298.0, + 1431.0, + 298.0, + 1474.0, + 294.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1431.0, + 442.0, + 1431.0, + 442.0, + 1474.0, + 388.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1431.0, + 1136.0, + 1431.0, + 1136.0, + 1474.0, + 567.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1431.0, + 1274.0, + 1431.0, + 1274.0, + 1474.0, + 1262.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1431.0, + 1409.0, + 1431.0, + 1409.0, + 1474.0, + 1402.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.25, + 1460.5, + 555.25, + 1460.5, + 555.25, + 1506.0, + 289.25, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1512.0, + 1046.0, + 1512.0, + 1046.0, + 1550.0, + 293.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1512.0, + 1405.0, + 1512.0, + 1405.0, + 1550.0, + 1078.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1543.0, + 1404.0, + 1543.0, + 1404.0, + 1578.0, + 294.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1575.0, + 452.0, + 1575.0, + 452.0, + 1609.0, + 296.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1575.0, + 726.0, + 1575.0, + 726.0, + 1609.0, + 535.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1575.0, + 787.0, + 1575.0, + 787.0, + 1609.0, + 745.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1575.0, + 1147.0, + 1575.0, + 1147.0, + 1609.0, + 813.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1575.0, + 1208.0, + 1575.0, + 1208.0, + 1609.0, + 1176.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 769.0, + 920.0, + 769.0, + 920.0, + 810.0, + 293.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 769.0, + 1304.0, + 769.0, + 1304.0, + 810.0, + 996.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 769.0, + 1405.0, + 769.0, + 1405.0, + 810.0, + 1334.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 803.0, + 296.0, + 803.0, + 296.0, + 839.0, + 293.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 803.0, + 490.0, + 803.0, + 490.0, + 839.0, + 379.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 803.0, + 1403.0, + 803.0, + 1403.0, + 839.0, + 518.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 835.0, + 615.0, + 835.0, + 615.0, + 868.0, + 293.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 835.0, + 654.0, + 835.0, + 654.0, + 868.0, + 644.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 692.0, + 1116.0, + 692.0, + 1116.0, + 733.0, + 295.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 692.0, + 1406.0, + 692.0, + 1406.0, + 733.0, + 1136.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 724.0, + 616.0, + 724.0, + 616.0, + 763.0, + 295.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1730.0, + 933.0, + 1730.0, + 933.0, + 1769.0, + 294.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1730.0, + 1402.0, + 1730.0, + 1402.0, + 1769.0, + 963.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1762.0, + 297.0, + 1762.0, + 297.0, + 1797.0, + 293.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1762.0, + 390.0, + 1762.0, + 390.0, + 1797.0, + 317.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1762.0, + 490.0, + 1762.0, + 490.0, + 1797.0, + 423.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1762.0, + 642.0, + 1762.0, + 642.0, + 1797.0, + 510.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1168.0, + 520.0, + 1168.0, + 555.0, + 294.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 520.0, + 1372.0, + 520.0, + 1372.0, + 555.0, + 1199.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 553.0, + 683.0, + 553.0, + 683.0, + 585.0, + 294.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 939.0, + 1910.0, + 939.0, + 1949.0, + 292.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1910.0, + 1309.0, + 1910.0, + 1309.0, + 1949.0, + 1018.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1949.0, + 1340.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1945.0, + 506.0, + 1945.0, + 506.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1266.0, + 417.0, + 1266.0, + 417.0, + 1303.0, + 295.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 983.0, + 1026.0, + 983.0, + 1026.0, + 1029.0, + 294.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1862.0, + 1030.0, + 1862.0, + 1030.0, + 1905.0, + 294.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1862.0, + 1114.0, + 1862.0, + 1114.0, + 1905.0, + 1059.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1862.0, + 1403.0, + 1862.0, + 1403.0, + 1905.0, + 1178.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 621.0, + 296.0, + 621.0, + 296.0, + 659.0, + 292.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 621.0, + 664.0, + 621.0, + 664.0, + 659.0, + 505.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 621.0, + 703.0, + 621.0, + 703.0, + 659.0, + 691.0, + 659.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/SJx9ngStPH/SJx9ngStPH.md b/parse/train/SJx9ngStPH/SJx9ngStPH.md new file mode 100644 index 0000000000000000000000000000000000000000..b8dde6a099379a555bef62adc61014670b81d052 --- /dev/null +++ b/parse/train/SJx9ngStPH/SJx9ngStPH.md @@ -0,0 +1,436 @@ +# NAS-BENCH-1SHOT1: BENCHMARKING AND DISSECTING ONE-SHOT NEURAL ARCHITECTURE SEARCH + +Arber $\mathbf { Z e l a ^ { 1 * } }$ , Julien Siems1∗, & Frank Hutter1,2 1Department of Computer Science, University of Freiburg 2Bosch Center for Artificial Intelligence {zelaa, siemsj, fh}@cs.uni-freiburg.de + +# ABSTRACT + +One-shot neural architecture search (NAS) has played a crucial role in making NAS methods computationally feasible in practice. Nevertheless, there is still a lack of understanding on how these weight-sharing algorithms exactly work due to the many factors controlling the dynamics of the process. In order to allow a scientific study of these components, we introduce a general framework for one-shot NAS that can be instantiated to many recently-introduced variants and introduce a general benchmarking framework that draws on the recent large-scale tabular benchmark NAS-Bench-101 for cheap anytime evaluations of one-shot NAS methods. To showcase the framework, we compare several state-of-the-art one-shot NAS methods, examine how sensitive they are to their hyperparameters and how they can be improved by tuning their hyperparameters, and compare their performance to that of blackbox optimizers for NAS-Bench-101. + +# 1 INTRODUCTION + +While neural architecture search (NAS) has attracted a lot of attention due to the effectiveness in automatically designing state-of-the-art neural networks (Zoph & Le, 2017; Zoph et al., 2018; Real et al., 2017; 2019), the focus has recently shifted to making the search process more efficient (Pham et al., 2018; Elsken et al., 2019; Liu et al., 2019; Xie et al., 2019; Cai et al., 2019; Casale et al., 2019). The most crucial concept which led to a reduction in search costs to the order of a single function evaluation is certainly the weight-sharing paradigm: Training only a single large architecture (the one-shot model) subsuming all the possible architectures in the search space (Brock et al., 2018; Pham et al., 2018). + +Despite the great advancements of these methods, the exact results of many NAS papers are often hard to reproduce (Li & Talwalkar, 2019; Yu et al., 2020; Yang et al., 2020). This is a result of several factors, such as unavailable original implementations, differences in the employed search spaces, training or evaluation pipelines, hyperparameter settings, and even pseudorandom number seeds (Lindauer & Hutter, 2019). One solution to guard against these problems would be a common library of NAS methods that provides primitives to construct different algorithm variants, similar to what as RLlib (Liang et al., 2017) offers for the field of reinforcement learning. Our paper makes a first step into this direction. + +Furthermore, experiments in NAS can be computationally extremely costly, making it virtually impossible to perform proper scientific evaluations with many repeated runs to draw statistically robust conclusions. To address this issue, Ying et al. (2019) introduced NAS-Bench-101, a large tabular benchmark with $4 2 3 \mathrm { k }$ unique cell architectures, trained and fully evaluated using a one-time extreme amount of compute power (several months on thousands of TPUs), which now allows to cheaply simulate an arbitrary number of runs of NAS methods, even on a laptop. NAS-Bench-101 enabled a comprehensive benchmarking of many discrete NAS optimizers (Zoph & Le, 2017; Real et al., 2019), using the exact same settings. However, the discrete nature of this benchmark does not allow to directly benchmark one-shot NAS optimizers (Pham et al., 2018; Liu et al., 2019; Xie et al., 2019; Cai et al., 2019). In this paper, we introduce the first method for making this possible. + +Specifically, after providing some background (Section 2), we make the following contributions: + +1. We introduce NAS-Bench-1Shot1, a novel benchmarking framework that allows us to reuse the extreme amount of compute time that went into generating NAS-Bench-101 (Ying et al., 2019) to cheaply benchmark one-shot NAS methods. Our mapping between search space representations is novel to the best of our knowledge and it allows querying the performance of found architectures from one-shot NAS methods, contrary to what is claimed by Ying et al. (2019). Specifically, it allows us to follow the full trajectory of architectures found by arbitrary one-shot NAS methods at each search epoch without the need for retraining them individually, allowing for a careful and statistically sound analysis (Section 3). +2. We introduce a general framework for one-shot NAS methods that can be instantiated to many recent one-shot NAS variants, enabling fair head-to-head evaluations based on a single code base (Section 4). +3. We use the above to compare several state-of-the-art one-shot NAS methods, assess the correlation between their one-shot model performance and final test performance, examine how sensitive they are to their hyperparameters, and compare their performance to that of black-box optimizers used in NAS-Bench-101 (Section 5). + +We provide our open-source implementation1, which we expect will also facilitate the reproducibility and benchmarking of other one-shot NAS methods in the future. + +# 2 BACKGROUND AND RELATED WORK + +# 2.1 NAS-BENCH-101 + +NAS-Bench-101 (Ying et al., 2019) is a database of an exhaustive evaluation of all architectures in a constrained cell-structured space on CIFAR-10 (Krizhevsky, 2009). Each cell is represented as a directed acyclic graph (DAG) where the nodes represent operation choices and the edges represent the information flow through the neural network (see also Figure 1 and Section 3.1). To limit the number of architectures in the search space, the authors used the following constraints on the cell: 3 operations in the operation set $\mathcal { O } = \{ 3 \mathrm { x } 3 $ convolution, 1x1 convolution, $3 { \mathrm { x } } 3 { \mathrm { ~ m a x } } { \mathrm { - p o o l } } \}$ , at most 7 nodes (this includes input and output node, therefore 5 choice nodes) and at most 9 edges. + +These constraints, and exploiting symmetries, reduced the search space to $4 2 3 \mathrm { k }$ unique valid architectures. Each architecture was trained from scratch three times to also obtain a measure of variance. In addition, each architecture was trained for 4, 12, 36 and 108 epochs; for our analysis, we mainly used the results for models trained for 108 epochs, if not stated otherwise. + +# 2.2 NAS-BENCH-102 + +Concurrently to this work, Dong & Yang (2020) released NAS-Bench-102, which is another NAS benchmark that, differently from NAS-Bench-101, enables the evaluation of weight-sharing NAS methods. Their search space consists of a total of 15,625 architectures, which is exhaustively evaluated on 3 image classification datasets. Similarly to Zela et al. (2020) and this work, Dong & Yang (2020) found that architectural overfitting occurs for DARTS for all their datasets. + +While NAS-Bench-102 and this work go towards the same direction, they differ in many ways: + +1. They use extensive computation to create a new benchmark (with 15.625 architectures), while we devise a novel reformulation to reuse the even much more extensive computation of the NASBench-101 dataset ( 120 TPU years) to create three new one-shot search spaces with the larges one containing 363.648 architectures. This required zero additional computational cost. +2. We show that it is possible to reuse the graph representation in NAS-Bench-101 to run one-shot NAS methods; this requires changes to the one-shot search space, but allows a mapping which can be used for architecture evaluation. + +![](images/282ee12d642c910c0d623c37f9afca75b123967986e2ff858c8e7937724695a4.jpg) +Figure 1: Overview of the NAS-Bench-1Shot1 analysis strategy. The one-shot model we construct only contains discrete architectures that are elements of NAS-Bench-101 (Ying et al., 2019). The cell architecture chosen is similar to that of Bender et al. (2018), with each choice block containing an operation decision. Note that NAS-Bench-101 does not contain a separate reduction cell type. Plot on the right from Ying et al. (2019) (Best viewed in color). + +3. They evaluate their search space on 3 image classification datasets, while we introduce 3 different search spaces (as sub-spaces of NAS-Bench-101) with growing complexity. + +# 2.3 ONE-SHOT NEURAL ARCHITECTURE SEARCH + +The NAS problem can be defined as searching for the optimal operation (e.g. in terms of validation error of architectures) out of the operation set $\mathcal { O }$ in each node of the DAG and for the best connectivity pattern between these nodes. + +Designing architectures for optimized accuracy or to comply with resource constraints led to significant breakthroughs on many standard benchmarks (Pham et al., 2018; Zoph & Le, 2017; Brock et al., 2018; Liu et al., 2019; Cai et al., 2019; Elsken et al., 2019). While early methods were computationally extremely expensive (Zoph & Le, 2017), the weight-sharing paradigm (Brock et al., 2018; Pham et al., 2018) led to a significant increase in search efficiency. Here, the weights of the operations in each architecture are shared in a supermodel (the so-called one-shot model or convolutional neural fabric (Saxena & Verbeek, 2016)), which contains an exponential number of sub-networks, each of which represents a discrete architecture. Architectures whose sub-networks share components (nodes/edges) also share the weights for these components’ operations; therefore, in analogy to DropOut (Srivastava et al., 2014), training one architecture implicitly also trains (parts of) an exponential number of related architectures. There are a variety of methods on how to conduct NAS by means of the one-shot model (Brock et al., 2018; Pham et al., 2018; Bender et al., 2018; Liu et al., 2019; Li & Talwalkar, 2019) (see also Appendix B), but the final problem is to find the optimal sub-network in this one-shot model. + +The weight sharing method was used to great effect in DARTS (Liu et al., 2019), where it allows a gradient based optimization of both the architectural and the one-shot weights. Subsequent work on DARTS has addressed further lowering the computational and the memory requirements (Dong & Yang, 2019; Xu et al., 2020; Cai et al., 2019; Casale et al., 2019). + +One fundamental drawback of the weight sharing method is the fact that the architecture search typically takes place in a lower fidelity model (e.g., using less cells and/or cheaper operations): the so-called proxy model. After the search, a discrete architecture is derived from the proxy model which is then trained with more parameters — a stage often referred to as architecture evaluation. This poses the question whether the architecture found in the proxy model is also a good architecture in the bigger model, a question studied by several recent works (Bender et al., 2018; Yu et al., 2020). + +# 3 A GENERAL FRAMEWORK FOR BENCHMARKING ONE-SHOT NAS + +We will now introduce our framework for cheaply benchmarking the anytime performance of oneshot NAS methods. Our main analysis strategy is the following: First, we run the search procedure of various methods and save the architecture weights of the one-shot models for each epoch. Second, we find the discrete architecture at each epoch and query it in NAS-Bench-101. The last step is not trivial due to the different representations of the search space used in NAS-Bench-101 and standard one-shot methods. Ying et al. (2019) state that one-shot methods cannot be directly evaluated on NAS-Bench-101. In the following sections we present a mapping between these different search space representations, which eventually enable us to evaluate one-shot methods on NAS-Bench101. To the best of our knowledge this is a novel contribution of this paper. + +# 3.1 SEARCH SPACE REPRESENTATION + +In order to carry out the analysis we propose in this work, we had to construct a search space that only contains discrete architectures that are also contained in NAS-Bench-101. This allows us to look up any discrete architectures’ performance in NAS-Bench-101 when the larger model is trained from scratch. Unfortunately, this is non-trivial since the NAS-Bench-101 space does not match the typical space used in one-shot NAS methods. We separately consider the various parts of the search space. + +Network-Level Topology. In terms of network-level topology, our search spaces closely resemble the models which were evaluated in NAS-Bench-101. We used the same macro architecture as in NAS-Bench-101, i.e., 3 stacked blocks with a max-pooling operation in-between, where each block consists of 3 stacked cells (see Figure 1). While our final evaluation models exactly follow NAS-Bench-101 in order to be able to reuse its evaluations, our one-shot model only has 16 initial convolution filters, rather than the 128 used in NAS-Bench-101. This is a common practice to accelerate NAS and used similarly in, e.g., Liu et al. (2019). + +Cell-Level Topology. The cell-level structure is represented as a DAG, where the input node is the output of a previous cell or the convolutional stem, and the output node is a concatenation of all the previous nodes. In order to have the operation choices still in the intermediate nodes of the DAG, we adapt the choice block motif from Bender et al. (2018) as depicted in Figure 1. The edges connecting input, output nodes and choice blocks represent only the information flow in the graph. To have a large and expressive enough search space(s), we introduce the following architectural weights in the DAG edges: + +• $\alpha ^ { i , j }$ to edges connecting nodes $\textit { i } < \textit { j }$ to choice block $j$ . The input of choice block $j$ is then +computed as $\begin{array} { r } { I ^ { j } = \sum _ { i < j } \frac { \exp ( \alpha ^ { i , j } ) } { \sum _ { i ^ { \prime } < j } \exp ( \alpha ^ { i ^ { \prime } , j } ) } x ^ { i } } \end{array}$ , where $x ^ { i }$ is the output tensor of node $i$ (either input node or choice block). +• $\gamma ^ { j , k }$ to the edges connecting the input node or choice blocks $j < k$ to the output node $k$ of the cell, +where the corresponding feature maps are concatenated: Ok = ⊕jSearch spaceNo. parents123Node 1111Node 2211Node 3221Node 4222Node 5 Output-- 32 2No.archs.w/ loose ends2w/o loose ends6240 248729160 3609363648 24066 + +Given the architectural weights of the cell shown in Figure 1 we query the test and validation error of the discrete architecture from NAS-Bench-101 as follows. + +1. We determine the operation chosen in each choice block by choosing the operation with the highest architectural weight. +2. We determine the parents of each choice block and the output by choosing the top- $k$ edges according to Table 1 (e.g. for choice block 4 in search space 3 we would choose the top-2 edges as parents). +3. From 1. we construct the operation list and from 2. the adjacency matrix of the cell which we use to query NAS-Bench-101 for the test and validation error. + +Each node in the graph chooses its parents during evaluation following e.g. DARTS (Liu et al., 2019). However, because edges model information flow and the output edges are also architectural decisions there is possibility of a node being a loose end. These are nodes whose output does not contribute to the output of the discrete cell, as seen in the upper cell under evaluation of Figure 1. As a result, we can count the number of architectures with or without loose ends. Note, that had we chosen the children of each node we could have invalid architectures where a node has an output but no input. + +# 4 A GENERAL FRAMEWORK FOR ONE-SHOT NAS METHODS + +Most of the follow-up works of DARTS (Algorithm 1), which focus on making the search even more efficient and effective, started from the original DARTS codebase2, and each of them only change very few components compared to DARTS. + +# Algorithm 1 DARTS + +# Algorithm 2 PC-DARTS + +# Algorithm 3 GDAS + +1: $\begin{array} { r } { I ^ { j } = \sum _ { i < j } S ( \alpha ^ { i , j } ) x ^ { i } } \end{array}$ +2: Ok = ⊕jOptimizerTest regretSearch space 1Search Space 2DARTSDefault config.1.252e-2 ± 0.0 0.864e-2 ± 0.023e-2Transferred config. 1.045e-2 ± 0.171e-20.992e-2 ± 0.211e-2GDASDefault config.1.252e-2 ± 0.0 0.871e-2 ± 0.0Transferred config. 3.093e-2 ± 3.975e-20.831e-2 ± 0.045e-2PC-DARTSDefault config.1.104e-2 ± 0.088e-2 1.133e-2 ± 0.404e-2Transferred config. 5.843e-2 ± 4.118e-20.992e-2 ± 0.087e-2 + +Figure 19 shows the interaction between the Cutout (CO) and $L _ { 2 }$ factor when optimizing on CS2, DARTS 1st order, for search space 1, 2 and 3. It should be noted that the best found configuration involves at least one relatively high value of one of the regularizers in CS2. + +![](images/0fa3c20d8d35b26e8f9dddb8adea3e24e09fcef0af9ac544ad16f7d02ec7a7c7.jpg) +Figure 19: Parameter importance for two hyperparameters, Cutout (CO) and $L _ { 2 }$ regularization (CS2) across different training epochs and search spaces (SS). + +# H CORRELATION BETWEEN THE ARCHITECTURE SEARCH MODEL AND THE ARCHITECTURE EVALUATION MODEL + +As a further experiment we wanted to test how strongly the validation error of the models used in architectures search and architecture evaluation are correlated. For this we sampled 150 architectures from search space 3 and trained this architecture in the proxy model. During training we compute the Spearman rank correlation between the validation error of the proxy model and the full architecture evaluation as queried from NAS-Bench-101. The results are shown in Figure 20. Note that 9 cells in the proxy was used for all of our previous experiments as it is also used by the NAS-Bench-101 models. + +For 9 cells (Figure 20c) in the proxy model, we find that increasing the total number of channels leads to stronger correlation between the proxy model and the full architecture. However, increasing it beyond 16 channels leads to a decrease in correlation. For 3 cells (Figure 20a) the strongest anytime correlation was interestingly found using only 2 initial channels, with more channels leading to worse performance at the beginning and no better final performance. The results suggest that there exists a set of good combinations between the number of cells and the initial number of channels to get the maximum correlation between the search and evaluation model. As a future work we plan to investigate this relationship, which could eventually lead to a more effective bandit-based NAS method. + +![](images/9eeb51d8a400b4c88d9e404eccca67fd3f7bd8a3ca37da9e0c5b70cd829ea444.jpg) +Figure 20: In this experiment we varied the total number of cells within [3, 6, 9] and the number of initial channels of the proxy model within [2, 4, 8, 16, 36]. \ No newline at end of file diff --git a/parse/train/SJx9ngStPH/SJx9ngStPH_content_list.json b/parse/train/SJx9ngStPH/SJx9ngStPH_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2c9e285496dede6d224bd4d3b213e396459559f4 --- /dev/null +++ b/parse/train/SJx9ngStPH/SJx9ngStPH_content_list.json @@ -0,0 +1,2166 @@ +[ + { + "type": "text", + "text": "NAS-BENCH-1SHOT1: BENCHMARKING AND DISSECTING ONE-SHOT NEURAL ARCHITECTURE SEARCH ", + "text_level": 1, + "bbox": [ + 174, + 98, + 723, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arber $\\mathbf { Z e l a ^ { 1 * } }$ , Julien Siems1∗, & Frank Hutter1,2 1Department of Computer Science, University of Freiburg 2Bosch Center for Artificial Intelligence {zelaa, siemsj, fh}@cs.uni-freiburg.de ", + "bbox": [ + 184, + 194, + 565, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 289, + 544, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One-shot neural architecture search (NAS) has played a crucial role in making NAS methods computationally feasible in practice. Nevertheless, there is still a lack of understanding on how these weight-sharing algorithms exactly work due to the many factors controlling the dynamics of the process. In order to allow a scientific study of these components, we introduce a general framework for one-shot NAS that can be instantiated to many recently-introduced variants and introduce a general benchmarking framework that draws on the recent large-scale tabular benchmark NAS-Bench-101 for cheap anytime evaluations of one-shot NAS methods. To showcase the framework, we compare several state-of-the-art one-shot NAS methods, examine how sensitive they are to their hyperparameters and how they can be improved by tuning their hyperparameters, and compare their performance to that of blackbox optimizers for NAS-Bench-101. ", + "bbox": [ + 233, + 321, + 764, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 517, + 336, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While neural architecture search (NAS) has attracted a lot of attention due to the effectiveness in automatically designing state-of-the-art neural networks (Zoph & Le, 2017; Zoph et al., 2018; Real et al., 2017; 2019), the focus has recently shifted to making the search process more efficient (Pham et al., 2018; Elsken et al., 2019; Liu et al., 2019; Xie et al., 2019; Cai et al., 2019; Casale et al., 2019). The most crucial concept which led to a reduction in search costs to the order of a single function evaluation is certainly the weight-sharing paradigm: Training only a single large architecture (the one-shot model) subsuming all the possible architectures in the search space (Brock et al., 2018; Pham et al., 2018). ", + "bbox": [ + 174, + 550, + 825, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the great advancements of these methods, the exact results of many NAS papers are often hard to reproduce (Li & Talwalkar, 2019; Yu et al., 2020; Yang et al., 2020). This is a result of several factors, such as unavailable original implementations, differences in the employed search spaces, training or evaluation pipelines, hyperparameter settings, and even pseudorandom number seeds (Lindauer & Hutter, 2019). One solution to guard against these problems would be a common library of NAS methods that provides primitives to construct different algorithm variants, similar to what as RLlib (Liang et al., 2017) offers for the field of reinforcement learning. Our paper makes a first step into this direction. ", + "bbox": [ + 174, + 669, + 825, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, experiments in NAS can be computationally extremely costly, making it virtually impossible to perform proper scientific evaluations with many repeated runs to draw statistically robust conclusions. To address this issue, Ying et al. (2019) introduced NAS-Bench-101, a large tabular benchmark with $4 2 3 \\mathrm { k }$ unique cell architectures, trained and fully evaluated using a one-time extreme amount of compute power (several months on thousands of TPUs), which now allows to cheaply simulate an arbitrary number of runs of NAS methods, even on a laptop. NAS-Bench-101 enabled a comprehensive benchmarking of many discrete NAS optimizers (Zoph & Le, 2017; Real et al., 2019), using the exact same settings. However, the discrete nature of this benchmark does not allow to directly benchmark one-shot NAS optimizers (Pham et al., 2018; Liu et al., 2019; Xie et al., 2019; Cai et al., 2019). In this paper, we introduce the first method for making this possible. ", + "bbox": [ + 174, + 786, + 823, + 898 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 821, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, after providing some background (Section 2), we make the following contributions: ", + "bbox": [ + 171, + 138, + 803, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We introduce NAS-Bench-1Shot1, a novel benchmarking framework that allows us to reuse the extreme amount of compute time that went into generating NAS-Bench-101 (Ying et al., 2019) to cheaply benchmark one-shot NAS methods. Our mapping between search space representations is novel to the best of our knowledge and it allows querying the performance of found architectures from one-shot NAS methods, contrary to what is claimed by Ying et al. (2019). Specifically, it allows us to follow the full trajectory of architectures found by arbitrary one-shot NAS methods at each search epoch without the need for retraining them individually, allowing for a careful and statistically sound analysis (Section 3). \n2. We introduce a general framework for one-shot NAS methods that can be instantiated to many recent one-shot NAS variants, enabling fair head-to-head evaluations based on a single code base (Section 4). \n3. We use the above to compare several state-of-the-art one-shot NAS methods, assess the correlation between their one-shot model performance and final test performance, examine how sensitive they are to their hyperparameters, and compare their performance to that of black-box optimizers used in NAS-Bench-101 (Section 5). ", + "bbox": [ + 173, + 162, + 825, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide our open-source implementation1, which we expect will also facilitate the reproducibility and benchmarking of other one-shot NAS methods in the future. ", + "bbox": [ + 173, + 392, + 821, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 BACKGROUND AND RELATED WORK ", + "text_level": 1, + "bbox": [ + 174, + 440, + 511, + 457 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 NAS-BENCH-101 ", + "text_level": 1, + "bbox": [ + 176, + 472, + 338, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NAS-Bench-101 (Ying et al., 2019) is a database of an exhaustive evaluation of all architectures in a constrained cell-structured space on CIFAR-10 (Krizhevsky, 2009). Each cell is represented as a directed acyclic graph (DAG) where the nodes represent operation choices and the edges represent the information flow through the neural network (see also Figure 1 and Section 3.1). To limit the number of architectures in the search space, the authors used the following constraints on the cell: 3 operations in the operation set $\\mathcal { O } = \\{ 3 \\mathrm { x } 3 $ convolution, 1x1 convolution, $3 { \\mathrm { x } } 3 { \\mathrm { ~ m a x } } { \\mathrm { - p o o l } } \\}$ , at most 7 nodes (this includes input and output node, therefore 5 choice nodes) and at most 9 edges. ", + "bbox": [ + 174, + 497, + 825, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These constraints, and exploiting symmetries, reduced the search space to $4 2 3 \\mathrm { k }$ unique valid architectures. Each architecture was trained from scratch three times to also obtain a measure of variance. In addition, each architecture was trained for 4, 12, 36 and 108 epochs; for our analysis, we mainly used the results for models trained for 108 epochs, if not stated otherwise. ", + "bbox": [ + 174, + 602, + 823, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 NAS-BENCH-102 ", + "text_level": 1, + "bbox": [ + 176, + 674, + 339, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concurrently to this work, Dong & Yang (2020) released NAS-Bench-102, which is another NAS benchmark that, differently from NAS-Bench-101, enables the evaluation of weight-sharing NAS methods. Their search space consists of a total of 15,625 architectures, which is exhaustively evaluated on 3 image classification datasets. Similarly to Zela et al. (2020) and this work, Dong & Yang (2020) found that architectural overfitting occurs for DARTS for all their datasets. ", + "bbox": [ + 174, + 700, + 825, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While NAS-Bench-102 and this work go towards the same direction, they differ in many ways: ", + "bbox": [ + 171, + 777, + 795, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. They use extensive computation to create a new benchmark (with 15.625 architectures), while we devise a novel reformulation to reuse the even much more extensive computation of the NASBench-101 dataset ( 120 TPU years) to create three new one-shot search spaces with the larges one containing 363.648 architectures. This required zero additional computational cost. \n2. We show that it is possible to reuse the graph representation in NAS-Bench-101 to run one-shot NAS methods; this requires changes to the one-shot search space, but allows a mapping which can be used for architecture evaluation. ", + "bbox": [ + 173, + 801, + 826, + 904 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/282ee12d642c910c0d623c37f9afca75b123967986e2ff858c8e7937724695a4.jpg", + "image_caption": [ + "Figure 1: Overview of the NAS-Bench-1Shot1 analysis strategy. The one-shot model we construct only contains discrete architectures that are elements of NAS-Bench-101 (Ying et al., 2019). The cell architecture chosen is similar to that of Bender et al. (2018), with each choice block containing an operation decision. Note that NAS-Bench-101 does not contain a separate reduction cell type. Plot on the right from Ying et al. (2019) (Best viewed in color). " + ], + "image_footnote": [], + "bbox": [ + 171, + 99, + 823, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. They evaluate their search space on 3 image classification datasets, while we introduce 3 different search spaces (as sub-spaces of NAS-Bench-101) with growing complexity. ", + "bbox": [ + 173, + 417, + 823, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 ONE-SHOT NEURAL ARCHITECTURE SEARCH", + "text_level": 1, + "bbox": [ + 174, + 489, + 531, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The NAS problem can be defined as searching for the optimal operation (e.g. in terms of validation error of architectures) out of the operation set $\\mathcal { O }$ in each node of the DAG and for the best connectivity pattern between these nodes. ", + "bbox": [ + 174, + 526, + 825, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Designing architectures for optimized accuracy or to comply with resource constraints led to significant breakthroughs on many standard benchmarks (Pham et al., 2018; Zoph & Le, 2017; Brock et al., 2018; Liu et al., 2019; Cai et al., 2019; Elsken et al., 2019). While early methods were computationally extremely expensive (Zoph & Le, 2017), the weight-sharing paradigm (Brock et al., 2018; Pham et al., 2018) led to a significant increase in search efficiency. Here, the weights of the operations in each architecture are shared in a supermodel (the so-called one-shot model or convolutional neural fabric (Saxena & Verbeek, 2016)), which contains an exponential number of sub-networks, each of which represents a discrete architecture. Architectures whose sub-networks share components (nodes/edges) also share the weights for these components’ operations; therefore, in analogy to DropOut (Srivastava et al., 2014), training one architecture implicitly also trains (parts of) an exponential number of related architectures. There are a variety of methods on how to conduct NAS by means of the one-shot model (Brock et al., 2018; Pham et al., 2018; Bender et al., 2018; Liu et al., 2019; Li & Talwalkar, 2019) (see also Appendix B), but the final problem is to find the optimal sub-network in this one-shot model. ", + "bbox": [ + 174, + 575, + 825, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The weight sharing method was used to great effect in DARTS (Liu et al., 2019), where it allows a gradient based optimization of both the architectural and the one-shot weights. Subsequent work on DARTS has addressed further lowering the computational and the memory requirements (Dong & Yang, 2019; Xu et al., 2020; Cai et al., 2019; Casale et al., 2019). ", + "bbox": [ + 176, + 777, + 823, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One fundamental drawback of the weight sharing method is the fact that the architecture search typically takes place in a lower fidelity model (e.g., using less cells and/or cheaper operations): the so-called proxy model. After the search, a discrete architecture is derived from the proxy model which is then trained with more parameters — a stage often referred to as architecture evaluation. This poses the question whether the architecture found in the proxy model is also a good architecture in the bigger model, a question studied by several recent works (Bender et al., 2018; Yu et al., 2020). ", + "bbox": [ + 174, + 840, + 825, + 922 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 A GENERAL FRAMEWORK FOR BENCHMARKING ONE-SHOT NAS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 740, + 117 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We will now introduce our framework for cheaply benchmarking the anytime performance of oneshot NAS methods. Our main analysis strategy is the following: First, we run the search procedure of various methods and save the architecture weights of the one-shot models for each epoch. Second, we find the discrete architecture at each epoch and query it in NAS-Bench-101. The last step is not trivial due to the different representations of the search space used in NAS-Bench-101 and standard one-shot methods. Ying et al. (2019) state that one-shot methods cannot be directly evaluated on NAS-Bench-101. In the following sections we present a mapping between these different search space representations, which eventually enable us to evaluate one-shot methods on NAS-Bench101. To the best of our knowledge this is a novel contribution of this paper. ", + "bbox": [ + 174, + 133, + 825, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 SEARCH SPACE REPRESENTATION ", + "text_level": 1, + "bbox": [ + 176, + 275, + 446, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to carry out the analysis we propose in this work, we had to construct a search space that only contains discrete architectures that are also contained in NAS-Bench-101. This allows us to look up any discrete architectures’ performance in NAS-Bench-101 when the larger model is trained from scratch. Unfortunately, this is non-trivial since the NAS-Bench-101 space does not match the typical space used in one-shot NAS methods. We separately consider the various parts of the search space. ", + "bbox": [ + 174, + 301, + 825, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network-Level Topology. In terms of network-level topology, our search spaces closely resemble the models which were evaluated in NAS-Bench-101. We used the same macro architecture as in NAS-Bench-101, i.e., 3 stacked blocks with a max-pooling operation in-between, where each block consists of 3 stacked cells (see Figure 1). While our final evaluation models exactly follow NAS-Bench-101 in order to be able to reuse its evaluations, our one-shot model only has 16 initial convolution filters, rather than the 128 used in NAS-Bench-101. This is a common practice to accelerate NAS and used similarly in, e.g., Liu et al. (2019). ", + "bbox": [ + 173, + 377, + 825, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cell-Level Topology. The cell-level structure is represented as a DAG, where the input node is the output of a previous cell or the convolutional stem, and the output node is a concatenation of all the previous nodes. In order to have the operation choices still in the intermediate nodes of the DAG, we adapt the choice block motif from Bender et al. (2018) as depicted in Figure 1. The edges connecting input, output nodes and choice blocks represent only the information flow in the graph. To have a large and expressive enough search space(s), we introduce the following architectural weights in the DAG edges: ", + "bbox": [ + 173, + 482, + 825, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "• $\\alpha ^ { i , j }$ to edges connecting nodes $\\textit { i } < \\textit { j }$ to choice block $j$ . The input of choice block $j$ is then \ncomputed as $\\begin{array} { r } { I ^ { j } = \\sum _ { i < j } \\frac { \\exp ( \\alpha ^ { i , j } ) } { \\sum _ { i ^ { \\prime } < j } \\exp ( \\alpha ^ { i ^ { \\prime } , j } ) } x ^ { i } } \\end{array}$ , where $x ^ { i }$ is the output tensor of node $i$ (either input node or choice block). \n• $\\gamma ^ { j , k }$ to the edges connecting the input node or choice blocks $j < k$ to the output node $k$ of the cell, \nwhere the corresponding feature maps are concatenated: Ok = ⊕jSearch spaceNo. parents123Node 1111Node 2211Node 3221Node 4222Node 5 Output-- 32 2No.archs.w/ loose ends2w/o loose ends6240 248729160 3609363648 24066", + "bbox": [ + 529, + 306, + 821, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 446, + 821, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the architectural weights of the cell shown in Figure 1 we query the test and validation error of the discrete architecture from NAS-Bench-101 as follows. ", + "bbox": [ + 173, + 482, + 820, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. We determine the operation chosen in each choice block by choosing the operation with the highest architectural weight. \n2. We determine the parents of each choice block and the output by choosing the top- $k$ edges according to Table 1 (e.g. for choice block 4 in search space 3 we would choose the top-2 edges as parents). \n3. From 1. we construct the operation list and from 2. the adjacency matrix of the cell which we use to query NAS-Bench-101 for the test and validation error. ", + "bbox": [ + 173, + 526, + 825, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each node in the graph chooses its parents during evaluation following e.g. DARTS (Liu et al., 2019). However, because edges model information flow and the output edges are also architectural decisions there is possibility of a node being a loose end. These are nodes whose output does not contribute to the output of the discrete cell, as seen in the upper cell under evaluation of Figure 1. As a result, we can count the number of architectures with or without loose ends. Note, that had we chosen the children of each node we could have invalid architectures where a node has an output but no input. ", + "bbox": [ + 173, + 671, + 825, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 A GENERAL FRAMEWORK FOR ONE-SHOT NAS METHODS", + "text_level": 1, + "bbox": [ + 174, + 799, + 694, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most of the follow-up works of DARTS (Algorithm 1), which focus on making the search even more efficient and effective, started from the original DARTS codebase2, and each of them only change very few components compared to DARTS. ", + "bbox": [ + 174, + 838, + 823, + 880 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 DARTS ", + "text_level": 1, + "bbox": [ + 174, + 118, + 316, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 2 PC-DARTS ", + "text_level": 1, + "bbox": [ + 341, + 118, + 508, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 3 GDAS ", + "text_level": 1, + "bbox": [ + 614, + 118, + 750, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1: $\\begin{array} { r } { I ^ { j } = \\sum _ { i < j } S ( \\alpha ^ { i , j } ) x ^ { i } } \\end{array}$ \n2: Ok = ⊕jOptimizerTest regretSearch space 1Search Space 2DARTSDefault config.1.252e-2 ± 0.0 0.864e-2 ± 0.023e-2Transferred config. 1.045e-2 ± 0.171e-20.992e-2 ± 0.211e-2GDASDefault config.1.252e-2 ± 0.0 0.871e-2 ± 0.0Transferred config. 3.093e-2 ± 3.975e-20.831e-2 ± 0.045e-2PC-DARTSDefault config.1.104e-2 ± 0.088e-2 1.133e-2 ± 0.404e-2Transferred config. 5.843e-2 ± 4.118e-20.992e-2 ± 0.087e-2", + "bbox": [ + 228, + 140, + 766, + 281 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 320, + 823, + 349 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Figure 19 shows the interaction between the Cutout (CO) and $L _ { 2 }$ factor when optimizing on CS2, DARTS 1st order, for search space 1, 2 and 3. It should be noted that the best found configuration involves at least one relatively high value of one of the regularizers in CS2. ", + "bbox": [ + 173, + 356, + 825, + 398 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/0fa3c20d8d35b26e8f9dddb8adea3e24e09fcef0af9ac544ad16f7d02ec7a7c7.jpg", + "image_caption": [ + "Figure 19: Parameter importance for two hyperparameters, Cutout (CO) and $L _ { 2 }$ regularization (CS2) across different training epochs and search spaces (SS). " + ], + "image_footnote": [], + "bbox": [ + 189, + 431, + 826, + 858 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "H CORRELATION BETWEEN THE ARCHITECTURE SEARCH MODEL AND THE ARCHITECTURE EVALUATION MODEL ", + "text_level": 1, + "bbox": [ + 174, + 103, + 808, + 136 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "As a further experiment we wanted to test how strongly the validation error of the models used in architectures search and architecture evaluation are correlated. For this we sampled 150 architectures from search space 3 and trained this architecture in the proxy model. During training we compute the Spearman rank correlation between the validation error of the proxy model and the full architecture evaluation as queried from NAS-Bench-101. The results are shown in Figure 20. Note that 9 cells in the proxy was used for all of our previous experiments as it is also used by the NAS-Bench-101 models. ", + "bbox": [ + 173, + 151, + 825, + 248 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "For 9 cells (Figure 20c) in the proxy model, we find that increasing the total number of channels leads to stronger correlation between the proxy model and the full architecture. However, increasing it beyond 16 channels leads to a decrease in correlation. For 3 cells (Figure 20a) the strongest anytime correlation was interestingly found using only 2 initial channels, with more channels leading to worse performance at the beginning and no better final performance. The results suggest that there exists a set of good combinations between the number of cells and the initial number of channels to get the maximum correlation between the search and evaluation model. As a future work we plan to investigate this relationship, which could eventually lead to a more effective bandit-based NAS method. ", + "bbox": [ + 173, + 256, + 825, + 381 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/9eeb51d8a400b4c88d9e404eccca67fd3f7bd8a3ca37da9e0c5b70cd829ea444.jpg", + "image_caption": [ + "Figure 20: In this experiment we varied the total number of cells within [3, 6, 9] and the number of initial channels of the proxy model within [2, 4, 8, 16, 36]. " + ], + "image_footnote": [], + "bbox": [ + 179, + 393, + 818, + 536 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/parse/train/SJx9ngStPH/SJx9ngStPH_middle.json b/parse/train/SJx9ngStPH/SJx9ngStPH_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..25d0c606ef8416d849386c888bb232dc2c102227 --- /dev/null +++ b/parse/train/SJx9ngStPH/SJx9ngStPH_middle.json @@ -0,0 +1,46838 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 443, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 277, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 277, + 96 + ], + "score": 1.0, + "content": "NAS-BENCH-1SHOT1:", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 367, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 367, + 117 + ], + "score": 1.0, + "content": "BENCHMARKING AND DISSECTING", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 118, + 444, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 118, + 444, + 137 + ], + "score": 1.0, + "content": "ONE-SHOT NEURAL ARCHITECTURE SEARCH", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 113, + 154, + 346, + 201 + ], + "lines": [ + { + "bbox": [ + 112, + 154, + 321, + 167 + ], + "spans": [ + { + "bbox": [ + 112, + 154, + 141, + 167 + ], + "score": 1.0, + "content": "Arber", + "type": "text" + }, + { + "bbox": [ + 141, + 154, + 169, + 166 + ], + "score": 0.27, + "content": "\\mathbf { Z e l a ^ { 1 * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 154, + 321, + 167 + ], + "score": 1.0, + "content": ", Julien Siems1∗, & Frank Hutter1,2", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 165, + 347, + 180 + ], + "spans": [ + { + "bbox": [ + 112, + 165, + 347, + 180 + ], + "score": 1.0, + "content": "1Department of Computer Science, University of Freiburg", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 176, + 276, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 176, + 276, + 191 + ], + "score": 1.0, + "content": "2Bosch Center for Artificial Intelligence", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 189, + 340, + 202 + ], + "spans": [ + { + "bbox": [ + 113, + 189, + 340, + 202 + ], + "score": 1.0, + "content": "{zelaa, siemsj, fh}@cs.uni-freiburg.de", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 278, + 229, + 333, + 241 + ], + "lines": [ + { + "bbox": [ + 276, + 228, + 336, + 243 + ], + "spans": [ + { + "bbox": [ + 276, + 228, + 336, + 243 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 255, + 468, + 387 + ], + "lines": [ + { + "bbox": [ + 141, + 254, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 470, + 268 + ], + "score": 1.0, + "content": "One-shot neural architecture search (NAS) has played a crucial role in making", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "NAS methods computationally feasible in practice. Nevertheless, there is still a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "score": 1.0, + "content": "lack of understanding on how these weight-sharing algorithms exactly work due", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 142, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "to the many factors controlling the dynamics of the process. In order to allow", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 140, + 299, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 140, + 299, + 470, + 312 + ], + "score": 1.0, + "content": "a scientific study of these components, we introduce a general framework for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 310, + 470, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 470, + 322 + ], + "score": 1.0, + "content": "one-shot NAS that can be instantiated to many recently-introduced variants and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 322, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 142, + 322, + 469, + 333 + ], + "score": 1.0, + "content": "introduce a general benchmarking framework that draws on the recent large-scale", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 332, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 469, + 344 + ], + "score": 1.0, + "content": "tabular benchmark NAS-Bench-101 for cheap anytime evaluations of one-shot", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 343, + 470, + 355 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 355 + ], + "score": 1.0, + "content": "NAS methods. To showcase the framework, we compare several state-of-the-art", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "score": 1.0, + "content": "one-shot NAS methods, examine how sensitive they are to their hyperparameters", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 142, + 365, + 469, + 377 + ], + "spans": [ + { + "bbox": [ + 142, + 365, + 469, + 377 + ], + "score": 1.0, + "content": "and how they can be improved by tuning their hyperparameters, and compare their", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 376, + 403, + 387 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 403, + 387 + ], + "score": 1.0, + "content": "performance to that of blackbox optimizers for NAS-Bench-101.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 108, + 410, + 206, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 208, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 208, + 426 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 436, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "While neural architecture search (NAS) has attracted a lot of attention due to the effectiveness in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "automatically designing state-of-the-art neural networks (Zoph & Le, 2017; Zoph et al., 2018; Real", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "et al., 2017; 2019), the focus has recently shifted to making the search process more efficient (Pham", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "et al., 2018; Elsken et al., 2019; Liu et al., 2019; Xie et al., 2019; Cai et al., 2019; Casale et al., 2019).", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "The most crucial concept which led to a reduction in search costs to the order of a single function", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "evaluation is certainly the weight-sharing paradigm: Training only a single large architecture (the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "one-shot model) subsuming all the possible architectures in the search space (Brock et al., 2018;", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 512, + 184, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 184, + 526 + ], + "score": 1.0, + "content": "Pham et al., 2018).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 530, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "Despite the great advancements of these methods, the exact results of many NAS papers are often", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 541, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 553 + ], + "score": 1.0, + "content": "hard to reproduce (Li & Talwalkar, 2019; Yu et al., 2020; Yang et al., 2020). This is a result of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "several factors, such as unavailable original implementations, differences in the employed search", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "spaces, training or evaluation pipelines, hyperparameter settings, and even pseudorandom number", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "seeds (Lindauer & Hutter, 2019). One solution to guard against these problems would be a common", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "library of NAS methods that provides primitives to construct different algorithm variants, similar to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "score": 1.0, + "content": "what as RLlib (Liang et al., 2017) offers for the field of reinforcement learning. Our paper makes a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 606, + 219, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 219, + 619 + ], + "score": 1.0, + "content": "first step into this direction.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 504, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 637 + ], + "score": 1.0, + "content": "Furthermore, experiments in NAS can be computationally extremely costly, making it virtually im-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 635, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 647 + ], + "score": 1.0, + "content": "possible to perform proper scientific evaluations with many repeated runs to draw statistically robust", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 646, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 658 + ], + "score": 1.0, + "content": "conclusions. To address this issue, Ying et al. (2019) introduced NAS-Bench-101, a large tabular", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 175, + 669 + ], + "score": 1.0, + "content": "benchmark with", + "type": "text" + }, + { + "bbox": [ + 175, + 657, + 198, + 667 + ], + "score": 0.72, + "content": "4 2 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "unique cell architectures, trained and fully evaluated using a one-time ex-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "treme amount of compute power (several months on thousands of TPUs), which now allows to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 678, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 504, + 690 + ], + "score": 1.0, + "content": "cheaply simulate an arbitrary number of runs of NAS methods, even on a laptop. NAS-Bench-101", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 690, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 701 + ], + "score": 1.0, + "content": "enabled a comprehensive benchmarking of many discrete NAS optimizers (Zoph & Le, 2017; Real", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "et al., 2019), using the exact same settings. However, the discrete nature of this benchmark does not", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 722, + 190, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 192, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 192, + 733 + ], + "score": 1.0, + "content": "∗Equal contribution", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 443, + 137 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 277, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 277, + 96 + ], + "score": 1.0, + "content": "NAS-BENCH-1SHOT1:", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 367, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 367, + 117 + ], + "score": 1.0, + "content": "BENCHMARKING AND DISSECTING", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 118, + 444, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 118, + 444, + 137 + ], + "score": 1.0, + "content": "ONE-SHOT NEURAL ARCHITECTURE SEARCH", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 113, + 154, + 346, + 201 + ], + "lines": [ + { + "bbox": [ + 112, + 154, + 321, + 167 + ], + "spans": [ + { + "bbox": [ + 112, + 154, + 141, + 167 + ], + "score": 1.0, + "content": "Arber", + "type": "text" + }, + { + "bbox": [ + 141, + 154, + 169, + 166 + ], + "score": 0.27, + "content": "\\mathbf { Z e l a ^ { 1 * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 154, + 321, + 167 + ], + "score": 1.0, + "content": ", Julien Siems1∗, & Frank Hutter1,2", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 165, + 347, + 180 + ], + "spans": [ + { + "bbox": [ + 112, + 165, + 347, + 180 + ], + "score": 1.0, + "content": "1Department of Computer Science, University of Freiburg", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 176, + 276, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 176, + 276, + 191 + ], + "score": 1.0, + "content": "2Bosch Center for Artificial Intelligence", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 189, + 340, + 202 + ], + "spans": [ + { + "bbox": [ + 113, + 189, + 340, + 202 + ], + "score": 1.0, + "content": "{zelaa, siemsj, fh}@cs.uni-freiburg.de", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 112, + 154, + 347, + 202 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 229, + 333, + 241 + ], + "lines": [ + { + "bbox": [ + 276, + 228, + 336, + 243 + ], + "spans": [ + { + "bbox": [ + 276, + 228, + 336, + 243 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 255, + 468, + 387 + ], + "lines": [ + { + "bbox": [ + 141, + 254, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 470, + 268 + ], + "score": 1.0, + "content": "One-shot neural architecture search (NAS) has played a crucial role in making", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 470, + 279 + ], + "score": 1.0, + "content": "NAS methods computationally feasible in practice. Nevertheless, there is still a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 470, + 290 + ], + "score": 1.0, + "content": "lack of understanding on how these weight-sharing algorithms exactly work due", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 142, + 289, + 469, + 300 + ], + "score": 1.0, + "content": "to the many factors controlling the dynamics of the process. In order to allow", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 140, + 299, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 140, + 299, + 470, + 312 + ], + "score": 1.0, + "content": "a scientific study of these components, we introduce a general framework for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 310, + 470, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 310, + 470, + 322 + ], + "score": 1.0, + "content": "one-shot NAS that can be instantiated to many recently-introduced variants and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 322, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 142, + 322, + 469, + 333 + ], + "score": 1.0, + "content": "introduce a general benchmarking framework that draws on the recent large-scale", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 332, + 469, + 344 + ], + "spans": [ + { + "bbox": [ + 141, + 332, + 469, + 344 + ], + "score": 1.0, + "content": "tabular benchmark NAS-Bench-101 for cheap anytime evaluations of one-shot", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 343, + 470, + 355 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 470, + 355 + ], + "score": 1.0, + "content": "NAS methods. To showcase the framework, we compare several state-of-the-art", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 366 + ], + "score": 1.0, + "content": "one-shot NAS methods, examine how sensitive they are to their hyperparameters", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 142, + 365, + 469, + 377 + ], + "spans": [ + { + "bbox": [ + 142, + 365, + 469, + 377 + ], + "score": 1.0, + "content": "and how they can be improved by tuning their hyperparameters, and compare their", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 376, + 403, + 387 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 403, + 387 + ], + "score": 1.0, + "content": "performance to that of blackbox optimizers for NAS-Bench-101.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13.5, + "bbox_fs": [ + 140, + 254, + 470, + 387 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 410, + 206, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 208, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 208, + 426 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 436, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "While neural architecture search (NAS) has attracted a lot of attention due to the effectiveness in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "automatically designing state-of-the-art neural networks (Zoph & Le, 2017; Zoph et al., 2018; Real", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "et al., 2017; 2019), the focus has recently shifted to making the search process more efficient (Pham", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "et al., 2018; Elsken et al., 2019; Liu et al., 2019; Xie et al., 2019; Cai et al., 2019; Casale et al., 2019).", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "The most crucial concept which led to a reduction in search costs to the order of a single function", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "evaluation is certainly the weight-sharing paradigm: Training only a single large architecture (the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "one-shot model) subsuming all the possible architectures in the search space (Brock et al., 2018;", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 512, + 184, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 184, + 526 + ], + "score": 1.0, + "content": "Pham et al., 2018).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 437, + 506, + 526 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 530, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "Despite the great advancements of these methods, the exact results of many NAS papers are often", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 541, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 553 + ], + "score": 1.0, + "content": "hard to reproduce (Li & Talwalkar, 2019; Yu et al., 2020; Yang et al., 2020). This is a result of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "several factors, such as unavailable original implementations, differences in the employed search", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "spaces, training or evaluation pipelines, hyperparameter settings, and even pseudorandom number", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "seeds (Lindauer & Hutter, 2019). One solution to guard against these problems would be a common", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "library of NAS methods that provides primitives to construct different algorithm variants, similar to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "score": 1.0, + "content": "what as RLlib (Liang et al., 2017) offers for the field of reinforcement learning. Our paper makes a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 606, + 219, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 219, + 619 + ], + "score": 1.0, + "content": "first step into this direction.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 530, + 506, + 619 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 504, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 637 + ], + "score": 1.0, + "content": "Furthermore, experiments in NAS can be computationally extremely costly, making it virtually im-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 635, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 647 + ], + "score": 1.0, + "content": "possible to perform proper scientific evaluations with many repeated runs to draw statistically robust", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 646, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 658 + ], + "score": 1.0, + "content": "conclusions. To address this issue, Ying et al. (2019) introduced NAS-Bench-101, a large tabular", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 175, + 669 + ], + "score": 1.0, + "content": "benchmark with", + "type": "text" + }, + { + "bbox": [ + 175, + 657, + 198, + 667 + ], + "score": 0.72, + "content": "4 2 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "unique cell architectures, trained and fully evaluated using a one-time ex-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "treme amount of compute power (several months on thousands of TPUs), which now allows to", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 678, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 504, + 690 + ], + "score": 1.0, + "content": "cheaply simulate an arbitrary number of runs of NAS methods, even on a laptop. NAS-Bench-101", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 690, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 701 + ], + "score": 1.0, + "content": "enabled a comprehensive benchmarking of many discrete NAS optimizers (Zoph & Le, 2017; Real", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "et al., 2019), using the exact same settings. However, the discrete nature of this benchmark does not", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "allow to directly benchmark one-shot NAS optimizers (Pham et al., 2018; Liu et al., 2019; Xie et al.,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 475, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 475, + 106 + ], + "score": 1.0, + "content": "2019; Cai et al., 2019). In this paper, we introduce the first method for making this possible.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 623, + 506, + 712 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "allow to directly benchmark one-shot NAS optimizers (Pham et al., 2018; Liu et al., 2019; Xie et al.,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 475, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 475, + 106 + ], + "score": 1.0, + "content": "2019; Cai et al., 2019). In this paper, we introduce the first method for making this possible.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 110, + 492, + 122 + ], + "lines": [ + { + "bbox": [ + 106, + 109, + 493, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 493, + 124 + ], + "score": 1.0, + "content": "Specifically, after providing some background (Section 2), we make the following contributions:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "1. We introduce NAS-Bench-1Shot1, a novel benchmarking framework that allows us to reuse the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 142, + 504, + 154 + ], + "spans": [ + { + "bbox": [ + 118, + 142, + 504, + 154 + ], + "score": 1.0, + "content": "extreme amount of compute time that went into generating NAS-Bench-101 (Ying et al., 2019) to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 118, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "cheaply benchmark one-shot NAS methods. Our mapping between search space representations", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 117, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "is novel to the best of our knowledge and it allows querying the performance of found architec-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 118, + 175, + 505, + 187 + ], + "score": 1.0, + "content": "tures from one-shot NAS methods, contrary to what is claimed by Ying et al. (2019). Specifically,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 117, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "it allows us to follow the full trajectory of architectures found by arbitrary one-shot NAS methods", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 118, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "at each search epoch without the need for retraining them individually, allowing for a careful and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 207, + 277, + 221 + ], + "spans": [ + { + "bbox": [ + 117, + 207, + 277, + 221 + ], + "score": 1.0, + "content": "statistically sound analysis (Section 3).", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 220, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 107, + 220, + 505, + 236 + ], + "score": 1.0, + "content": "2. We introduce a general framework for one-shot NAS methods that can be instantiated to many", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 117, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "recent one-shot NAS variants, enabling fair head-to-head evaluations based on a single code base", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 118, + 243, + 168, + 256 + ], + "spans": [ + { + "bbox": [ + 118, + 243, + 168, + 256 + ], + "score": 1.0, + "content": "(Section 4).", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 258, + 504, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 504, + 270 + ], + "score": 1.0, + "content": "3. We use the above to compare several state-of-the-art one-shot NAS methods, assess the correla-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 119, + 270, + 504, + 280 + ], + "spans": [ + { + "bbox": [ + 119, + 270, + 504, + 280 + ], + "score": 1.0, + "content": "tion between their one-shot model performance and final test performance, examine how sensitive", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 118, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 118, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "they are to their hyperparameters, and compare their performance to that of black-box optimizers", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 118, + 291, + 268, + 304 + ], + "spans": [ + { + "bbox": [ + 118, + 291, + 268, + 304 + ], + "score": 1.0, + "content": "used in NAS-Bench-101 (Section 5).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 311, + 503, + 334 + ], + "lines": [ + { + "bbox": [ + 107, + 312, + 504, + 323 + ], + "spans": [ + { + "bbox": [ + 107, + 312, + 504, + 323 + ], + "score": 1.0, + "content": "We provide our open-source implementation1, which we expect will also facilitate the reproducibil-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 379, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 379, + 335 + ], + "score": 1.0, + "content": "ity and benchmarking of other one-shot NAS methods in the future.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 107, + 349, + 313, + 362 + ], + "lines": [ + { + "bbox": [ + 104, + 348, + 315, + 364 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 315, + 364 + ], + "score": 1.0, + "content": "2 BACKGROUND AND RELATED WORK", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 374, + 207, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 208, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 208, + 387 + ], + "score": 1.0, + "content": "2.1 NAS-BENCH-101", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 406 + ], + "score": 1.0, + "content": "NAS-Bench-101 (Ying et al., 2019) is a database of an exhaustive evaluation of all architectures in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 405, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 417 + ], + "score": 1.0, + "content": "a constrained cell-structured space on CIFAR-10 (Krizhevsky, 2009). Each cell is represented as a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "directed acyclic graph (DAG) where the nodes represent operation choices and the edges represent", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 428, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 439 + ], + "score": 1.0, + "content": "the information flow through the neural network (see also Figure 1 and Section 3.1). To limit the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "number of architectures in the search space, the authors used the following constraints on the cell:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 237, + 462 + ], + "score": 1.0, + "content": "3 operations in the operation set", + "type": "text" + }, + { + "bbox": [ + 237, + 449, + 277, + 461 + ], + "score": 0.92, + "content": "\\mathcal { O } = \\{ 3 \\mathrm { x } 3 ", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 448, + 399, + 462 + ], + "score": 1.0, + "content": "convolution, 1x1 convolution,", + "type": "text" + }, + { + "bbox": [ + 400, + 449, + 462, + 461 + ], + "score": 0.73, + "content": "3 { \\mathrm { x } } 3 { \\mathrm { ~ m a x } } { \\mathrm { - p o o l } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 448, + 505, + 462 + ], + "score": 1.0, + "content": ", at most 7", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 461, + 467, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 467, + 473 + ], + "score": 1.0, + "content": "nodes (this includes input and output node, therefore 5 choice nodes) and at most 9 edges.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 504, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 404, + 489 + ], + "score": 1.0, + "content": "These constraints, and exploiting symmetries, reduced the search space to", + "type": "text" + }, + { + "bbox": [ + 405, + 477, + 426, + 488 + ], + "score": 0.49, + "content": "4 2 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "unique valid archi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "tectures. Each architecture was trained from scratch three times to also obtain a measure of variance.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "In addition, each architecture was trained for 4, 12, 36 and 108 epochs; for our analysis, we mainly", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 404, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 404, + 523 + ], + "score": 1.0, + "content": "used the results for models trained for 108 epochs, if not stated otherwise.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 534, + 208, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 209, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 209, + 547 + ], + "score": 1.0, + "content": "2.2 NAS-BENCH-102", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 555, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "Concurrently to this work, Dong & Yang (2020) released NAS-Bench-102, which is another NAS", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "benchmark that, differently from NAS-Bench-101, enables the evaluation of weight-sharing NAS", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "methods. Their search space consists of a total of 15,625 architectures, which is exhaustively evalu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 586, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 505, + 603 + ], + "score": 1.0, + "content": "ated on 3 image classification datasets. Similarly to Zela et al. (2020) and this work, Dong & Yang", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 599, + 434, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 434, + 610 + ], + "score": 1.0, + "content": "(2020) found that architectural overfitting occurs for DARTS for all their datasets.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 616, + 487, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 487, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 487, + 630 + ], + "score": 1.0, + "content": "While NAS-Bench-102 and this work go towards the same direction, they differ in many ways:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 506, + 716 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "1. They use extensive computation to create a new benchmark (with 15.625 architectures), while", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 118, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 118, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "we devise a novel reformulation to reuse the even much more extensive computation of the NAS-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 118, + 657, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 118, + 657, + 506, + 670 + ], + "score": 1.0, + "content": "Bench-101 dataset ( 120 TPU years) to create three new one-shot search spaces with the larges", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 118, + 669, + 471, + 681 + ], + "spans": [ + { + "bbox": [ + 118, + 669, + 471, + 681 + ], + "score": 1.0, + "content": "one containing 363.648 architectures. This required zero additional computational cost.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 682, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 506, + 696 + ], + "score": 1.0, + "content": "2. We show that it is possible to reuse the graph representation in NAS-Bench-101 to run one-shot", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 118, + 693, + 506, + 707 + ], + "spans": [ + { + "bbox": [ + 118, + 693, + 506, + 707 + ], + "score": 1.0, + "content": "NAS methods; this requires changes to the one-shot search space, but allows a mapping which", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 118, + 705, + 277, + 717 + ], + "spans": [ + { + "bbox": [ + 118, + 705, + 277, + 717 + ], + "score": 1.0, + "content": "can be used for architecture evaluation.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 121, + 722, + 278, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 721, + 280, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 721, + 280, + 733 + ], + "score": 1.0, + "content": "1https://github.com/automl/nasbench-1shot1", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 503, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 105, + 110, + 492, + 122 + ], + "lines": [ + { + "bbox": [ + 106, + 109, + 493, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 493, + 124 + ], + "score": 1.0, + "content": "Specifically, after providing some background (Section 2), we make the following contributions:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 109, + 493, + 124 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 129, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 143 + ], + "score": 1.0, + "content": "1. We introduce NAS-Bench-1Shot1, a novel benchmarking framework that allows us to reuse the", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 142, + 504, + 154 + ], + "spans": [ + { + "bbox": [ + 118, + 142, + 504, + 154 + ], + "score": 1.0, + "content": "extreme amount of compute time that went into generating NAS-Bench-101 (Ying et al., 2019) to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 118, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "cheaply benchmark one-shot NAS methods. Our mapping between search space representations", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 117, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "is novel to the best of our knowledge and it allows querying the performance of found architec-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 118, + 175, + 505, + 187 + ], + "score": 1.0, + "content": "tures from one-shot NAS methods, contrary to what is claimed by Ying et al. (2019). Specifically,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 185, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 117, + 185, + 505, + 199 + ], + "score": 1.0, + "content": "it allows us to follow the full trajectory of architectures found by arbitrary one-shot NAS methods", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 118, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "at each search epoch without the need for retraining them individually, allowing for a careful and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 207, + 277, + 221 + ], + "spans": [ + { + "bbox": [ + 117, + 207, + 277, + 221 + ], + "score": 1.0, + "content": "statistically sound analysis (Section 3).", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 220, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 107, + 220, + 505, + 236 + ], + "score": 1.0, + "content": "2. We introduce a general framework for one-shot NAS methods that can be instantiated to many", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 117, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 117, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "recent one-shot NAS variants, enabling fair head-to-head evaluations based on a single code base", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 118, + 243, + 168, + 256 + ], + "spans": [ + { + "bbox": [ + 118, + 243, + 168, + 256 + ], + "score": 1.0, + "content": "(Section 4).", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 258, + 504, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 504, + 270 + ], + "score": 1.0, + "content": "3. We use the above to compare several state-of-the-art one-shot NAS methods, assess the correla-", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 119, + 270, + 504, + 280 + ], + "spans": [ + { + "bbox": [ + 119, + 270, + 504, + 280 + ], + "score": 1.0, + "content": "tion between their one-shot model performance and final test performance, examine how sensitive", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 118, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 118, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "they are to their hyperparameters, and compare their performance to that of black-box optimizers", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 118, + 291, + 268, + 304 + ], + "spans": [ + { + "bbox": [ + 118, + 291, + 268, + 304 + ], + "score": 1.0, + "content": "used in NAS-Bench-101 (Section 5).", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + } + ], + "index": 10, + "bbox_fs": [ + 106, + 131, + 505, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 311, + 503, + 334 + ], + "lines": [ + { + "bbox": [ + 107, + 312, + 504, + 323 + ], + "spans": [ + { + "bbox": [ + 107, + 312, + 504, + 323 + ], + "score": 1.0, + "content": "We provide our open-source implementation1, which we expect will also facilitate the reproducibil-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 379, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 379, + 335 + ], + "score": 1.0, + "content": "ity and benchmarking of other one-shot NAS methods in the future.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 312, + 504, + 335 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 349, + 313, + 362 + ], + "lines": [ + { + "bbox": [ + 104, + 348, + 315, + 364 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 315, + 364 + ], + "score": 1.0, + "content": "2 BACKGROUND AND RELATED WORK", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 374, + 207, + 385 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 208, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 208, + 387 + ], + "score": 1.0, + "content": "2.1 NAS-BENCH-101", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 406 + ], + "score": 1.0, + "content": "NAS-Bench-101 (Ying et al., 2019) is a database of an exhaustive evaluation of all architectures in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 405, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 417 + ], + "score": 1.0, + "content": "a constrained cell-structured space on CIFAR-10 (Krizhevsky, 2009). Each cell is represented as a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "directed acyclic graph (DAG) where the nodes represent operation choices and the edges represent", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 428, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 439 + ], + "score": 1.0, + "content": "the information flow through the neural network (see also Figure 1 and Section 3.1). To limit the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "number of architectures in the search space, the authors used the following constraints on the cell:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 237, + 462 + ], + "score": 1.0, + "content": "3 operations in the operation set", + "type": "text" + }, + { + "bbox": [ + 237, + 449, + 277, + 461 + ], + "score": 0.92, + "content": "\\mathcal { O } = \\{ 3 \\mathrm { x } 3 ", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 448, + 399, + 462 + ], + "score": 1.0, + "content": "convolution, 1x1 convolution,", + "type": "text" + }, + { + "bbox": [ + 400, + 449, + 462, + 461 + ], + "score": 0.73, + "content": "3 { \\mathrm { x } } 3 { \\mathrm { ~ m a x } } { \\mathrm { - p o o l } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 448, + 505, + 462 + ], + "score": 1.0, + "content": ", at most 7", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 461, + 467, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 467, + 473 + ], + "score": 1.0, + "content": "nodes (this includes input and output node, therefore 5 choice nodes) and at most 9 edges.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 395, + 506, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 477, + 504, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 404, + 489 + ], + "score": 1.0, + "content": "These constraints, and exploiting symmetries, reduced the search space to", + "type": "text" + }, + { + "bbox": [ + 405, + 477, + 426, + 488 + ], + "score": 0.49, + "content": "4 2 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "unique valid archi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "tectures. Each architecture was trained from scratch three times to also obtain a measure of variance.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "In addition, each architecture was trained for 4, 12, 36 and 108 epochs; for our analysis, we mainly", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 510, + 404, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 404, + 523 + ], + "score": 1.0, + "content": "used the results for models trained for 108 epochs, if not stated otherwise.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 477, + 505, + 523 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 534, + 208, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 209, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 209, + 547 + ], + "score": 1.0, + "content": "2.2 NAS-BENCH-102", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 555, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "Concurrently to this work, Dong & Yang (2020) released NAS-Bench-102, which is another NAS", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "benchmark that, differently from NAS-Bench-101, enables the evaluation of weight-sharing NAS", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "methods. Their search space consists of a total of 15,625 architectures, which is exhaustively evalu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 586, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 505, + 603 + ], + "score": 1.0, + "content": "ated on 3 image classification datasets. Similarly to Zela et al. (2020) and this work, Dong & Yang", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 599, + 434, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 434, + 610 + ], + "score": 1.0, + "content": "(2020) found that architectural overfitting occurs for DARTS for all their datasets.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 104, + 555, + 505, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 616, + 487, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 487, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 487, + 630 + ], + "score": 1.0, + "content": "While NAS-Bench-102 and this work go towards the same direction, they differ in many ways:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 614, + 487, + 630 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 635, + 506, + 716 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "1. They use extensive computation to create a new benchmark (with 15.625 architectures), while", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 118, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "we devise a novel reformulation to reuse the even much more extensive computation of the NAS-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 118, + 657, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 118, + 657, + 506, + 670 + ], + "score": 1.0, + "content": "Bench-101 dataset ( 120 TPU years) to create three new one-shot search spaces with the larges", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 118, + 669, + 471, + 681 + ], + "spans": [ + { + "bbox": [ + 118, + 669, + 471, + 681 + ], + "score": 1.0, + "content": "one containing 363.648 architectures. This required zero additional computational cost.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 682, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 506, + 696 + ], + "score": 1.0, + "content": "2. We show that it is possible to reuse the graph representation in NAS-Bench-101 to run one-shot", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 693, + 506, + 707 + ], + "spans": [ + { + "bbox": [ + 118, + 693, + 506, + 707 + ], + "score": 1.0, + "content": "NAS methods; this requires changes to the one-shot search space, but allows a mapping which", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 118, + 705, + 277, + 717 + ], + "spans": [ + { + "bbox": [ + 118, + 705, + 277, + 717 + ], + "score": 1.0, + "content": "can be used for architecture evaluation.", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + } + ], + "index": 43, + "bbox_fs": [ + 104, + 635, + 506, + 717 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 79, + 504, + 228 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 79, + 504, + 228 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 79, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 504, + 228 + ], + "score": 0.968, + "type": "image", + "image_path": "282ee12d642c910c0d623c37f9afca75b123967986e2ff858c8e7937724695a4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 79, + 504, + 128.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 128.66666666666666, + 504, + 178.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 178.33333333333331, + 504, + 227.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 235, + 505, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "Figure 1: Overview of the NAS-Bench-1Shot1 analysis strategy. The one-shot model we construct", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "only contains discrete architectures that are elements of NAS-Bench-101 (Ying et al., 2019). The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 255, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 104, + 255, + 506, + 271 + ], + "score": 1.0, + "content": "cell architecture chosen is similar to that of Bender et al. (2018), with each choice block containing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 267, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 104, + 267, + 505, + 283 + ], + "score": 1.0, + "content": "an operation decision. Note that NAS-Bench-101 does not contain a separate reduction cell type.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 279, + 362, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 362, + 293 + ], + "score": 1.0, + "content": "Plot on the right from Ying et al. (2019) (Best viewed in color).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "3. They evaluate their search space on 3 image classification datasets, while we introduce 3 different", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 342, + 423, + 355 + ], + "spans": [ + { + "bbox": [ + 118, + 342, + 423, + 355 + ], + "score": 1.0, + "content": "search spaces (as sub-spaces of NAS-Bench-101) with growing complexity.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 107, + 388, + 325, + 400 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 327, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 327, + 401 + ], + "score": 1.0, + "content": "2.3 ONE-SHOT NEURAL ARCHITECTURE SEARCH", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "The NAS problem can be defined as searching for the optimal operation (e.g. in terms of valida-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 318, + 441 + ], + "score": 1.0, + "content": "tion error of architectures) out of the operation set", + "type": "text" + }, + { + "bbox": [ + 318, + 429, + 327, + 439 + ], + "score": 0.74, + "content": "\\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "in each node of the DAG and for the best", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 440, + 275, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 275, + 451 + ], + "score": 1.0, + "content": "connectivity pattern between these nodes.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "Designing architectures for optimized accuracy or to comply with resource constraints led to sig-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "nificant breakthroughs on many standard benchmarks (Pham et al., 2018; Zoph & Le, 2017; Brock", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "score": 1.0, + "content": "et al., 2018; Liu et al., 2019; Cai et al., 2019; Elsken et al., 2019). While early methods were compu-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "tationally extremely expensive (Zoph & Le, 2017), the weight-sharing paradigm (Brock et al., 2018;", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "Pham et al., 2018) led to a significant increase in search efficiency. Here, the weights of the opera-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "tions in each architecture are shared in a supermodel (the so-called one-shot model or convolutional", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "neural fabric (Saxena & Verbeek, 2016)), which contains an exponential number of sub-networks,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "each of which represents a discrete architecture. Architectures whose sub-networks share compo-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "nents (nodes/edges) also share the weights for these components’ operations; therefore, in analogy", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "to DropOut (Srivastava et al., 2014), training one architecture implicitly also trains (parts of) an ex-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 565, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 579 + ], + "score": 1.0, + "content": "ponential number of related architectures. There are a variety of methods on how to conduct NAS by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "means of the one-shot model (Brock et al., 2018; Pham et al., 2018; Bender et al., 2018; Liu et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "2019; Li & Talwalkar, 2019) (see also Appendix B), but the final problem is to find the optimal", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 600, + 252, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 252, + 610 + ], + "score": 1.0, + "content": "sub-network in this one-shot model.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 616, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "The weight sharing method was used to great effect in DARTS (Liu et al., 2019), where it allows a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "gradient based optimization of both the architectural and the one-shot weights. Subsequent work on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "DARTS has addressed further lowering the computational and the memory requirements (Dong &", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 649, + 369, + 661 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 369, + 661 + ], + "score": 1.0, + "content": "Yang, 2019; Xu et al., 2020; Cai et al., 2019; Casale et al., 2019).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "One fundamental drawback of the weight sharing method is the fact that the architecture search", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "typically takes place in a lower fidelity model (e.g., using less cells and/or cheaper operations): the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "so-called proxy model. After the search, a discrete architecture is derived from the proxy model", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "which is then trained with more parameters — a stage often referred to as architecture evaluation.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "This poses the question whether the architecture found in the proxy model is also a good architecture", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "in the bigger model, a question studied by several recent works (Bender et al., 2018; Yu et al., 2020).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 79, + 504, + 228 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 105, + 79, + 504, + 228 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 79, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 504, + 228 + ], + "score": 0.968, + "type": "image", + "image_path": "282ee12d642c910c0d623c37f9afca75b123967986e2ff858c8e7937724695a4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 105, + 79, + 504, + 128.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 105, + 128.66666666666666, + 504, + 178.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 105, + 178.33333333333331, + 504, + 227.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 235, + 505, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "Figure 1: Overview of the NAS-Bench-1Shot1 analysis strategy. The one-shot model we construct", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "only contains discrete architectures that are elements of NAS-Bench-101 (Ying et al., 2019). The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 255, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 104, + 255, + 506, + 271 + ], + "score": 1.0, + "content": "cell architecture chosen is similar to that of Bender et al. (2018), with each choice block containing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 267, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 104, + 267, + 505, + 283 + ], + "score": 1.0, + "content": "an operation decision. Note that NAS-Bench-101 does not contain a separate reduction cell type.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 279, + 362, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 362, + 293 + ], + "score": 1.0, + "content": "Plot on the right from Ying et al. (2019) (Best viewed in color).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "3. They evaluate their search space on 3 image classification datasets, while we introduce 3 different", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 342, + 423, + 355 + ], + "spans": [ + { + "bbox": [ + 118, + 342, + 423, + 355 + ], + "score": 1.0, + "content": "search spaces (as sub-spaces of NAS-Bench-101) with growing complexity.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 331, + 505, + 355 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 388, + 325, + 400 + ], + "lines": [ + { + "bbox": [ + 105, + 388, + 327, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 327, + 401 + ], + "score": 1.0, + "content": "2.3 ONE-SHOT NEURAL ARCHITECTURE SEARCH", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 430 + ], + "score": 1.0, + "content": "The NAS problem can be defined as searching for the optimal operation (e.g. in terms of valida-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 318, + 441 + ], + "score": 1.0, + "content": "tion error of architectures) out of the operation set", + "type": "text" + }, + { + "bbox": [ + 318, + 429, + 327, + 439 + ], + "score": 0.74, + "content": "\\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "in each node of the DAG and for the best", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 440, + 275, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 275, + 451 + ], + "score": 1.0, + "content": "connectivity pattern between these nodes.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 416, + 505, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "Designing architectures for optimized accuracy or to comply with resource constraints led to sig-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "nificant breakthroughs on many standard benchmarks (Pham et al., 2018; Zoph & Le, 2017; Brock", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 492 + ], + "score": 1.0, + "content": "et al., 2018; Liu et al., 2019; Cai et al., 2019; Elsken et al., 2019). While early methods were compu-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "tationally extremely expensive (Zoph & Le, 2017), the weight-sharing paradigm (Brock et al., 2018;", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "Pham et al., 2018) led to a significant increase in search efficiency. Here, the weights of the opera-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "tions in each architecture are shared in a supermodel (the so-called one-shot model or convolutional", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "neural fabric (Saxena & Verbeek, 2016)), which contains an exponential number of sub-networks,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "each of which represents a discrete architecture. Architectures whose sub-networks share compo-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "nents (nodes/edges) also share the weights for these components’ operations; therefore, in analogy", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "to DropOut (Srivastava et al., 2014), training one architecture implicitly also trains (parts of) an ex-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 565, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 579 + ], + "score": 1.0, + "content": "ponential number of related architectures. There are a variety of methods on how to conduct NAS by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "means of the one-shot model (Brock et al., 2018; Pham et al., 2018; Bender et al., 2018; Liu et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "2019; Li & Talwalkar, 2019) (see also Appendix B), but the final problem is to find the optimal", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 600, + 252, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 252, + 610 + ], + "score": 1.0, + "content": "sub-network in this one-shot model.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 456, + 506, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 616, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "The weight sharing method was used to great effect in DARTS (Liu et al., 2019), where it allows a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 639 + ], + "score": 1.0, + "content": "gradient based optimization of both the architectural and the one-shot weights. Subsequent work on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "DARTS has addressed further lowering the computational and the memory requirements (Dong &", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 649, + 369, + 661 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 369, + 661 + ], + "score": 1.0, + "content": "Yang, 2019; Xu et al., 2020; Cai et al., 2019; Casale et al., 2019).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 615, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "One fundamental drawback of the weight sharing method is the fact that the architecture search", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "typically takes place in a lower fidelity model (e.g., using less cells and/or cheaper operations): the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "so-called proxy model. After the search, a discrete architecture is derived from the proxy model", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "which is then trained with more parameters — a stage often referred to as architecture evaluation.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "This poses the question whether the architecture found in the proxy model is also a good architecture", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "in the bigger model, a question studied by several recent works (Bender et al., 2018; Yu et al., 2020).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 665, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 453, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 81, + 454, + 95 + ], + "spans": [ + { + "bbox": [ + 104, + 81, + 454, + 95 + ], + "score": 1.0, + "content": "3 A GENERAL FRAMEWORK FOR BENCHMARKING ONE-SHOT NAS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "We will now introduce our framework for cheaply benchmarking the anytime performance of one-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 116, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 130 + ], + "score": 1.0, + "content": "shot NAS methods. Our main analysis strategy is the following: First, we run the search procedure", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "score": 1.0, + "content": "of various methods and save the architecture weights of the one-shot models for each epoch. Second,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 506, + 152 + ], + "score": 1.0, + "content": "we find the discrete architecture at each epoch and query it in NAS-Bench-101. The last step is not", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "trivial due to the different representations of the search space used in NAS-Bench-101 and standard", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "score": 1.0, + "content": "one-shot methods. Ying et al. (2019) state that one-shot methods cannot be directly evaluated on", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "NAS-Bench-101. In the following sections we present a mapping between these different search", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "space representations, which eventually enable us to evaluate one-shot methods on NAS-Bench-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 192, + 408, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 408, + 207 + ], + "score": 1.0, + "content": "101. To the best of our knowledge this is a novel contribution of this paper.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 218, + 273, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 218, + 275, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 275, + 231 + ], + "score": 1.0, + "content": "3.1 SEARCH SPACE REPRESENTATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "In order to carry out the analysis we propose in this work, we had to construct a search space that only", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 249, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 264 + ], + "score": 1.0, + "content": "contains discrete architectures that are also contained in NAS-Bench-101. This allows us to look up", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 261, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 273 + ], + "score": 1.0, + "content": "any discrete architectures’ performance in NAS-Bench-101 when the larger model is trained from", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "scratch. Unfortunately, this is non-trivial since the NAS-Bench-101 space does not match the typical", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 283, + 504, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 504, + 296 + ], + "score": 1.0, + "content": "space used in one-shot NAS methods. We separately consider the various parts of the search space.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "Network-Level Topology. In terms of network-level topology, our search spaces closely resemble", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 310, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 323 + ], + "score": 1.0, + "content": "the models which were evaluated in NAS-Bench-101. We used the same macro architecture as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "in NAS-Bench-101, i.e., 3 stacked blocks with a max-pooling operation in-between, where each", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "block consists of 3 stacked cells (see Figure 1). While our final evaluation models exactly follow", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "NAS-Bench-101 in order to be able to reuse its evaluations, our one-shot model only has 16 initial", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "convolution filters, rather than the 128 used in NAS-Bench-101. This is a common practice to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 365, + 348, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 348, + 378 + ], + "score": 1.0, + "content": "accelerate NAS and used similarly in, e.g., Liu et al. (2019).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 382, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "score": 1.0, + "content": "Cell-Level Topology. The cell-level structure is represented as a DAG, where the input node is the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "output of a previous cell or the convolutional stem, and the output node is a concatenation of all the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 506, + 418 + ], + "score": 1.0, + "content": "previous nodes. In order to have the operation choices still in the intermediate nodes of the DAG, we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 414, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 429 + ], + "score": 1.0, + "content": "adapt the choice block motif from Bender et al. (2018) as depicted in Figure 1. The edges connecting", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "input, output nodes and choice blocks represent only the information flow in the graph. To have a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "large and expressive enough search space(s), we introduce the following architectural weights in the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 447, + 159, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 159, + 462 + ], + "score": 1.0, + "content": "DAG edges:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 466, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 114, + 483 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 115, + 468, + 132, + 479 + ], + "score": 0.87, + "content": "\\alpha ^ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 466, + 246, + 483 + ], + "score": 1.0, + "content": "to edges connecting nodes", + "type": "text" + }, + { + "bbox": [ + 246, + 470, + 273, + 480 + ], + "score": 0.89, + "content": "\\textit { i } < \\textit { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 466, + 342, + 483 + ], + "score": 1.0, + "content": "to choice block", + "type": "text" + }, + { + "bbox": [ + 342, + 470, + 348, + 481 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 466, + 466, + 483 + ], + "score": 1.0, + "content": ". The input of choice block", + "type": "text" + }, + { + "bbox": [ + 466, + 470, + 473, + 481 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 466, + 506, + 483 + ], + "score": 1.0, + "content": "is then", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 110, + 478, + 509, + 506 + ], + "spans": [ + { + "bbox": [ + 110, + 478, + 168, + 506 + ], + "score": 1.0, + "content": "computed as", + "type": "text" + }, + { + "bbox": [ + 169, + 480, + 289, + 500 + ], + "score": 0.94, + "content": "\\begin{array} { r } { I ^ { j } = \\sum _ { i < j } \\frac { \\exp ( \\alpha ^ { i , j } ) } { \\sum _ { i ^ { \\prime } < j } \\exp ( \\alpha ^ { i ^ { \\prime } , j } ) } x ^ { i } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 478, + 321, + 506 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 321, + 483, + 331, + 493 + ], + "score": 0.87, + "content": "x ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 478, + 446, + 506 + ], + "score": 1.0, + "content": "is the output tensor of node", + "type": "text" + }, + { + "bbox": [ + 447, + 484, + 452, + 493 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 478, + 509, + 506 + ], + "score": 1.0, + "content": "(either input", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 498, + 207, + 511 + ], + "spans": [ + { + "bbox": [ + 114, + 498, + 207, + 511 + ], + "score": 1.0, + "content": "node or choice block).", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 511, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 104, + 511, + 114, + 527 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 115, + 513, + 132, + 525 + ], + "score": 0.9, + "content": "\\gamma ^ { j , k }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 511, + 354, + 527 + ], + "score": 1.0, + "content": "to the edges connecting the input node or choice blocks", + "type": "text" + }, + { + "bbox": [ + 354, + 514, + 379, + 525 + ], + "score": 0.9, + "content": "j < k", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 511, + 453, + 527 + ], + "score": 1.0, + "content": "to the output node", + "type": "text" + }, + { + "bbox": [ + 453, + 514, + 460, + 524 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 511, + 505, + 527 + ], + "score": 1.0, + "content": "of the cell,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 110, + 522, + 504, + 548 + ], + "spans": [ + { + "bbox": [ + 110, + 522, + 459, + 548 + ], + "score": 1.0, + "content": "where the corresponding feature maps are concatenated: Ok = ⊕jSearch spaceNo. parents123Node 1111Node 2211Node 3221Node 4222Node 5 Output-- 32 2No.archs.w/ loose ends2w/o loose ends6240 248729160 3609363648 24066", + "type": "table", + "image_path": "dac3da73efd354eb0f2a3c37ef91f090e87952c73b9d80bd91ab521c04c2573b.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 324, + 243, + 503, + 257.7142857142857 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 324, + 257.7142857142857, + 503, + 272.42857142857144 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 324, + 272.42857142857144, + 503, + 287.14285714285717 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 324, + 287.14285714285717, + 503, + 301.8571428571429 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 324, + 301.8571428571429, + 503, + 316.5714285714286 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 324, + 316.5714285714286, + 503, + 331.28571428571433 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 324, + 331.28571428571433, + 503, + 346.00000000000006 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 23.75 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 503, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "only available tabular benchmark for one-shot NAS. For details on each search space see Appendix", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 363, + 121, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 121, + 379 + ], + "score": 1.0, + "content": "A.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 382, + 502, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 382, + 504, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 504, + 394 + ], + "score": 1.0, + "content": "Given the architectural weights of the cell shown in Figure 1 we query the test and validation error", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 392, + 352, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 352, + 405 + ], + "score": 1.0, + "content": "of the discrete architecture from NAS-Bench-101 as follows.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 417, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "1. We determine the operation chosen in each choice block by choosing the operation with the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 118, + 429, + 234, + 442 + ], + "spans": [ + { + "bbox": [ + 118, + 429, + 234, + 442 + ], + "score": 1.0, + "content": "highest architectural weight.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 450, + 504, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 456, + 465 + ], + "score": 1.0, + "content": "2. We determine the parents of each choice block and the output by choosing the top-", + "type": "text" + }, + { + "bbox": [ + 456, + 452, + 463, + 461 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 450, + 504, + 465 + ], + "score": 1.0, + "content": "edges ac-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 119, + 463, + 504, + 475 + ], + "spans": [ + { + "bbox": [ + 119, + 463, + 504, + 475 + ], + "score": 1.0, + "content": "cording to Table 1 (e.g. for choice block 4 in search space 3 we would choose the top-2 edges as", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 473, + 156, + 486 + ], + "spans": [ + { + "bbox": [ + 117, + 473, + 156, + 486 + ], + "score": 1.0, + "content": "parents).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "3. From 1. we construct the operation list and from 2. the adjacency matrix of the cell which we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 118, + 507, + 367, + 520 + ], + "spans": [ + { + "bbox": [ + 118, + 507, + 367, + 520 + ], + "score": 1.0, + "content": "use to query NAS-Bench-101 for the test and validation error.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "Each node in the graph chooses its parents during evaluation following e.g. DARTS (Liu et al.,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "2019). However, because edges model information flow and the output edges are also architectural", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 554, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 506, + 566 + ], + "score": 1.0, + "content": "decisions there is possibility of a node being a loose end. These are nodes whose output does not", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 566, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 504, + 577 + ], + "score": 1.0, + "content": "contribute to the output of the discrete cell, as seen in the upper cell under evaluation of Figure 1.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "As a result, we can count the number of architectures with or without loose ends. Note, that had we", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "chosen the children of each node we could have invalid architectures where a node has an output but", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 596, + 145, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 596, + 145, + 612 + ], + "score": 1.0, + "content": "no input.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + }, + { + "type": "title", + "bbox": [ + 107, + 633, + 425, + 647 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 426, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 426, + 648 + ], + "score": 1.0, + "content": "4 A GENERAL FRAMEWORK FOR ONE-SHOT NAS METHODS", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 664, + 504, + 697 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 676 + ], + "score": 1.0, + "content": "Most of the follow-up works of DARTS (Algorithm 1), which focus on making the search even more", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 674, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 505, + 689 + ], + "score": 1.0, + "content": "efficient and effective, started from the original DARTS codebase2, and each of them only change", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 686, + 283, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 283, + 699 + ], + "score": 1.0, + "content": "very few components compared to DARTS.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 722, + 290, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 293, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 293, + 734 + ], + "score": 1.0, + "content": "2https://github.com/quark0/darts", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "input. This adaptive choice for the number of channels is incompatible with the one-shot model due", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "to the different tensor dimensionality coming from previous choice blocks. We used 1x1 projections", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 270, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 270, + 115 + ], + "score": 1.0, + "content": "with a fixed number of channels instead.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 506, + 115 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 138, + 241, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 138, + 243, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 243, + 151 + ], + "score": 1.0, + "content": "3.2 EVALUATION PROCEDURE", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "score": 1.0, + "content": "By means of these additional weights we do not restrict the possible architectures in the search", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "space to contain only a fixed number of edges per cell, as done for example in Zoph et al. (2018),", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 184, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 505, + 197 + ], + "score": 1.0, + "content": "Pham et al. (2018), Liu et al. (2019), etc. This requirement would have restricted our architectural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 195, + 325, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 325, + 208 + ], + "score": 1.0, + "content": "decisions heavily, leading to only small search spaces.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 162, + 506, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 212, + 316, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 317, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 317, + 223 + ], + "score": 1.0, + "content": "Table 1 shows the characteristics of each search", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 223, + 317, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 317, + 235 + ], + "score": 1.0, + "content": "space. We propose three different search spaces by", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 234, + 317, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 317, + 246 + ], + "score": 1.0, + "content": "making different decisions on the number of par-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 244, + 317, + 256 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 317, + 256 + ], + "score": 1.0, + "content": "ents each choice block has. The decisions affect the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 256, + 317, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 317, + 267 + ], + "score": 1.0, + "content": "quality and quantity of the architectures contained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 266, + 317, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 317, + 279 + ], + "score": 1.0, + "content": "in each search space. For all search spaces note", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 277, + 317, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 317, + 289 + ], + "score": 1.0, + "content": "that the sum of the number of parents of all nodes", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 288, + 317, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 317, + 300 + ], + "score": 1.0, + "content": "in the search space is chosen to be 9, to match the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 317, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 317, + 311 + ], + "score": 1.0, + "content": "NAS-Bench-101 requirement. Search space 1, 2 and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 310, + 317, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 317, + 322 + ], + "score": 1.0, + "content": "3 have 6240, 29160 and 363648 architectures with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 321, + 317, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 321, + 317, + 334 + ], + "score": 1.0, + "content": "loose ends respectively, making search space 3 the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 332, + 317, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 317, + 344 + ], + "score": 1.0, + "content": "largest investigated search space. To the best of our", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 343, + 317, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 317, + 356 + ], + "score": 1.0, + "content": "knowledge search space 3 is currently the largest and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "only available tabular benchmark for one-shot NAS. For details on each search space see Appendix", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 363, + 121, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 121, + 379 + ], + "score": 1.0, + "content": "A.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 212, + 317, + 356 + ] + }, + { + "type": "table", + "bbox": [ + 324, + 243, + 503, + 346 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 324, + 212, + 504, + 234 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 324, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 324, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "Table 1: Characteristic information of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 324, + 223, + 383, + 235 + ], + "spans": [ + { + "bbox": [ + 324, + 223, + 383, + 235 + ], + "score": 1.0, + "content": "search spaces.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "table_body", + "bbox": [ + 324, + 243, + 503, + 346 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 324, + 243, + 503, + 346 + ], + "spans": [ + { + "bbox": [ + 324, + 243, + 503, + 346 + ], + "score": 0.977, + "html": "
Search space
No. parents123
Node 1111
Node 2211
Node 3221
Node 4222
Node 5 Output-- 32 2
No.archs.w/ loose ends2
w/o loose ends6240 248729160 3609363648 24066
", + "type": "table", + "image_path": "dac3da73efd354eb0f2a3c37ef91f090e87952c73b9d80bd91ab521c04c2573b.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 324, + 243, + 503, + 257.7142857142857 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 324, + 257.7142857142857, + 503, + 272.42857142857144 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 324, + 272.42857142857144, + 503, + 287.14285714285717 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 324, + 287.14285714285717, + 503, + 301.8571428571429 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 324, + 301.8571428571429, + 503, + 316.5714285714286 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 324, + 316.5714285714286, + 503, + 331.28571428571433 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 324, + 331.28571428571433, + 503, + 346.00000000000006 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 23.75 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 503, + 376 + ], + "lines": [], + "index": 30.5, + "bbox_fs": [ + 105, + 353, + 505, + 379 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 382, + 502, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 382, + 504, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 504, + 394 + ], + "score": 1.0, + "content": "Given the architectural weights of the cell shown in Figure 1 we query the test and validation error", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 392, + 352, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 352, + 405 + ], + "score": 1.0, + "content": "of the discrete architecture from NAS-Bench-101 as follows.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 382, + 504, + 405 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 417, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "1. We determine the operation chosen in each choice block by choosing the operation with the", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 429, + 234, + 442 + ], + "spans": [ + { + "bbox": [ + 118, + 429, + 234, + 442 + ], + "score": 1.0, + "content": "highest architectural weight.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 450, + 504, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 456, + 465 + ], + "score": 1.0, + "content": "2. We determine the parents of each choice block and the output by choosing the top-", + "type": "text" + }, + { + "bbox": [ + 456, + 452, + 463, + 461 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 450, + 504, + 465 + ], + "score": 1.0, + "content": "edges ac-", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 119, + 463, + 504, + 475 + ], + "spans": [ + { + "bbox": [ + 119, + 463, + 504, + 475 + ], + "score": 1.0, + "content": "cording to Table 1 (e.g. for choice block 4 in search space 3 we would choose the top-2 edges as", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 473, + 156, + 486 + ], + "spans": [ + { + "bbox": [ + 117, + 473, + 156, + 486 + ], + "score": 1.0, + "content": "parents).", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "3. From 1. we construct the operation list and from 2. the adjacency matrix of the cell which we", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 118, + 507, + 367, + 520 + ], + "spans": [ + { + "bbox": [ + 118, + 507, + 367, + 520 + ], + "score": 1.0, + "content": "use to query NAS-Bench-101 for the test and validation error.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + } + ], + "index": 37, + "bbox_fs": [ + 105, + 417, + 505, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "Each node in the graph chooses its parents during evaluation following e.g. DARTS (Liu et al.,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "2019). However, because edges model information flow and the output edges are also architectural", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 554, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 506, + 566 + ], + "score": 1.0, + "content": "decisions there is possibility of a node being a loose end. These are nodes whose output does not", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 566, + 504, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 504, + 577 + ], + "score": 1.0, + "content": "contribute to the output of the discrete cell, as seen in the upper cell under evaluation of Figure 1.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 588 + ], + "score": 1.0, + "content": "As a result, we can count the number of architectures with or without loose ends. Note, that had we", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "chosen the children of each node we could have invalid architectures where a node has an output but", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 596, + 145, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 596, + 145, + 612 + ], + "score": 1.0, + "content": "no input.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44, + "bbox_fs": [ + 104, + 532, + 506, + 612 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 633, + 425, + 647 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 426, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 426, + 648 + ], + "score": 1.0, + "content": "4 A GENERAL FRAMEWORK FOR ONE-SHOT NAS METHODS", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 664, + 504, + 697 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 676 + ], + "score": 1.0, + "content": "Most of the follow-up works of DARTS (Algorithm 1), which focus on making the search even more", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 674, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 505, + 689 + ], + "score": 1.0, + "content": "efficient and effective, started from the original DARTS codebase2, and each of them only change", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 686, + 283, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 283, + 699 + ], + "score": 1.0, + "content": "very few components compared to DARTS.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50, + "bbox_fs": [ + 105, + 664, + 505, + 699 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 94, + 194, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 93, + 195, + 108 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 195, + 108 + ], + "score": 1.0, + "content": "Algorithm 1 DARTS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 209, + 94, + 311, + 105 + ], + "lines": [ + { + "bbox": [ + 207, + 93, + 312, + 108 + ], + "spans": [ + { + "bbox": [ + 207, + 93, + 312, + 108 + ], + "score": 1.0, + "content": "Algorithm 2 PC-DARTS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 376, + 94, + 459, + 105 + ], + "lines": [ + { + "bbox": [ + 375, + 93, + 459, + 108 + ], + "spans": [ + { + "bbox": [ + 375, + 93, + 459, + 108 + ], + "score": 1.0, + "content": "Algorithm 3 GDAS", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 109, + 110, + 201, + 190 + ], + "lines": [ + { + "bbox": [ + 108, + 107, + 191, + 123 + ], + "spans": [ + { + "bbox": [ + 108, + 107, + 122, + 123 + ], + "score": 1.0, + "content": "1:", + "type": "text" + }, + { + "bbox": [ + 122, + 111, + 191, + 121 + ], + "score": 0.38, + "content": "\\begin{array} { r } { I ^ { j } = \\sum _ { i < j } S ( \\alpha ^ { i , j } ) x ^ { i } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 117, + 196, + 132 + ], + "spans": [ + { + "bbox": [ + 108, + 117, + 196, + 132 + ], + "score": 1.0, + "content": "2: Ok = ⊕jOptimizerTest regretSearch space 1Search Space 2DARTSDefault config.1.252e-2 ± 0.0 0.864e-2 ± 0.023e-2Transferred config. 1.045e-2 ± 0.171e-20.992e-2 ± 0.211e-2GDASDefault config.1.252e-2 ± 0.0 0.871e-2 ± 0.0Transferred config. 3.093e-2 ± 3.975e-20.831e-2 ± 0.045e-2PC-DARTSDefault config.1.104e-2 ± 0.088e-2 1.133e-2 ± 0.404e-2Transferred config. 5.843e-2 ± 4.118e-20.992e-2 ± 0.087e-2", + "type": "table", + "image_path": "04ff47956bae289cb669cd3433a8432687d3aeede6dd7ac5836cb618c11d41b2.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 140, + 111, + 469, + 148.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 140, + 148.33333333333334, + 469, + 185.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 140, + 185.66666666666669, + 469, + 223.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 108, + 254, + 504, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 504, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 504, + 268 + ], + "score": 1.0, + "content": "was trained on all configurations belonging to specific budgets during the BOHB optimization pro-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 266, + 138, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 138, + 278 + ], + "score": 1.0, + "content": "cedure.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 282, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 358, + 296 + ], + "score": 1.0, + "content": "Figure 19 shows the interaction between the Cutout (CO) and", + "type": "text" + }, + { + "bbox": [ + 359, + 283, + 371, + 294 + ], + "score": 0.88, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "factor when optimizing on CS2,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "DARTS 1st order, for search space 1, 2 and 3. It should be noted that the best found configuration", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 304, + 408, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 408, + 317 + ], + "score": 1.0, + "content": "involves at least one relatively high value of one of the regularizers in CS2.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 116, + 342, + 506, + 680 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 342, + 506, + 680 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 342, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 116, + 342, + 506, + 680 + ], + "score": 0.974, + "type": "image", + "image_path": "0fa3c20d8d35b26e8f9dddb8adea3e24e09fcef0af9ac544ad16f7d02ec7a7c7.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 116, + 342, + 506, + 454.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 116, + 454.6666666666667, + 506, + 567.3333333333334 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 116, + 567.3333333333334, + 506, + 680.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 689, + 503, + 711 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 688, + 504, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 408, + 701 + ], + "score": 1.0, + "content": "Figure 19: Parameter importance for two hyperparameters, Cutout (CO) and", + "type": "text" + }, + { + "bbox": [ + 408, + 689, + 421, + 700 + ], + "score": 0.87, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 688, + 504, + 701 + ], + "score": 1.0, + "content": "regularization (CS2)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 699, + 330, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 330, + 712 + ], + "score": 1.0, + "content": "across different training epochs and search spaces (SS).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 12.25 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 111, + 469, + 223 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 80, + 505, + 103 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 505, + 93 + ], + "score": 1.0, + "content": "Table 2: Results of architectures found on search space 1 and 2 with the best found configuration", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 91, + 281, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 91, + 281, + 104 + ], + "score": 1.0, + "content": "for 50 epochs by BOHB on search space 3.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 140, + 111, + 469, + 223 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 469, + 223 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 469, + 223 + ], + "score": 0.983, + "html": "
OptimizerTest regret
Search space 1Search Space 2
DARTSDefault config.1.252e-2 ± 0.0 0.864e-2 ± 0.023e-2
Transferred config. 1.045e-2 ± 0.171e-20.992e-2 ± 0.211e-2
GDASDefault config.1.252e-2 ± 0.0 0.871e-2 ± 0.0
Transferred config. 3.093e-2 ± 3.975e-20.831e-2 ± 0.045e-2
PC-DARTSDefault config.1.104e-2 ± 0.088e-2 1.133e-2 ± 0.404e-2
Transferred config. 5.843e-2 ± 4.118e-20.992e-2 ± 0.087e-2
", + "type": "table", + "image_path": "04ff47956bae289cb669cd3433a8432687d3aeede6dd7ac5836cb618c11d41b2.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 140, + 111, + 469, + 148.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 140, + 148.33333333333334, + 469, + 185.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 140, + 185.66666666666669, + 469, + 223.00000000000003 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "text", + "bbox": [ + 108, + 254, + 504, + 277 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 253, + 504, + 278 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 282, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 358, + 296 + ], + "score": 1.0, + "content": "Figure 19 shows the interaction between the Cutout (CO) and", + "type": "text" + }, + { + "bbox": [ + 359, + 283, + 371, + 294 + ], + "score": 0.88, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "factor when optimizing on CS2,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "DARTS 1st order, for search space 1, 2 and 3. It should be noted that the best found configuration", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 304, + 408, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 408, + 317 + ], + "score": 1.0, + "content": "involves at least one relatively high value of one of the regularizers in CS2.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 282, + 505, + 317 + ] + }, + { + "type": "image", + "bbox": [ + 116, + 342, + 506, + 680 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 342, + 506, + 680 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 342, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 116, + 342, + 506, + 680 + ], + "score": 0.974, + "type": "image", + "image_path": "0fa3c20d8d35b26e8f9dddb8adea3e24e09fcef0af9ac544ad16f7d02ec7a7c7.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 116, + 342, + 506, + 454.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 116, + 454.6666666666667, + 506, + 567.3333333333334 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 116, + 567.3333333333334, + 506, + 680.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 689, + 503, + 711 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 688, + 504, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 408, + 701 + ], + "score": 1.0, + "content": "Figure 19: Parameter importance for two hyperparameters, Cutout (CO) and", + "type": "text" + }, + { + "bbox": [ + 408, + 689, + 421, + 700 + ], + "score": 0.87, + "content": "L _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 688, + 504, + 701 + ], + "score": 1.0, + "content": "regularization (CS2)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 699, + 330, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 330, + 712 + ], + "score": 1.0, + "content": "across different training epochs and search spaces (SS).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 12.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 495, + 108 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 496, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 496, + 95 + ], + "score": 1.0, + "content": "H CORRELATION BETWEEN THE ARCHITECTURE SEARCH MODEL AND THE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 128, + 97, + 318, + 109 + ], + "spans": [ + { + "bbox": [ + 128, + 97, + 318, + 109 + ], + "score": 1.0, + "content": "ARCHITECTURE EVALUATION MODEL", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 120, + 505, + 197 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 132 + ], + "score": 1.0, + "content": "As a further experiment we wanted to test how strongly the validation error of the models used in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "architectures search and architecture evaluation are correlated. For this we sampled 150 architectures", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 155 + ], + "score": 1.0, + "content": "from search space 3 and trained this architecture in the proxy model. During training we compute the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "Spearman rank correlation between the validation error of the proxy model and the full architecture", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 176 + ], + "score": 1.0, + "content": "evaluation as queried from NAS-Bench-101. The results are shown in Figure 20. Note that 9 cells", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "in the proxy was used for all of our previous experiments as it is also used by the NAS-Bench-101", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 141, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 141, + 198 + ], + "score": 1.0, + "content": "models.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "For 9 cells (Figure 20c) in the proxy model, we find that increasing the total number of channels leads", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "to stronger correlation between the proxy model and the full architecture. However, increasing it", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 224, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 505, + 237 + ], + "score": 1.0, + "content": "beyond 16 channels leads to a decrease in correlation. For 3 cells (Figure 20a) the strongest anytime", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "score": 1.0, + "content": "correlation was interestingly found using only 2 initial channels, with more channels leading to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "worse performance at the beginning and no better final performance. The results suggest that there", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "score": 1.0, + "content": "exists a set of good combinations between the number of cells and the initial number of channels to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 268, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 281 + ], + "score": 1.0, + "content": "get the maximum correlation between the search and evaluation model. As a future work we plan", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 104, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "to investigate this relationship, which could eventually lead to a more effective bandit-based NAS", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 141, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 141, + 303 + ], + "score": 1.0, + "content": "method.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 110, + 312, + 501, + 425 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 312, + 501, + 425 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 312, + 501, + 425 + ], + "spans": [ + { + "bbox": [ + 110, + 312, + 501, + 425 + ], + "score": 0.969, + "type": "image", + "image_path": "9eeb51d8a400b4c88d9e404eccca67fd3f7bd8a3ca37da9e0c5b70cd829ea444.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 110, + 312, + 501, + 349.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 110, + 349.6666666666667, + 501, + 387.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 110, + 387.33333333333337, + 501, + 425.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 433, + 505, + 456 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "Figure 20: In this experiment we varied the total number of cells within [3, 6, 9] and the number of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 444, + 339, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 339, + 457 + ], + "score": 1.0, + "content": "initial channels of the proxy model within [2, 4, 8, 16, 36].", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "index": 20.25 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 495, + 108 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 496, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 496, + 95 + ], + "score": 1.0, + "content": "H CORRELATION BETWEEN THE ARCHITECTURE SEARCH MODEL AND THE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 128, + 97, + 318, + 109 + ], + "spans": [ + { + "bbox": [ + 128, + 97, + 318, + 109 + ], + "score": 1.0, + "content": "ARCHITECTURE EVALUATION MODEL", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 120, + 505, + 197 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 132 + ], + "score": 1.0, + "content": "As a further experiment we wanted to test how strongly the validation error of the models used in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 144 + ], + "score": 1.0, + "content": "architectures search and architecture evaluation are correlated. For this we sampled 150 architectures", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 155 + ], + "score": 1.0, + "content": "from search space 3 and trained this architecture in the proxy model. During training we compute the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "Spearman rank correlation between the validation error of the proxy model and the full architecture", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 176 + ], + "score": 1.0, + "content": "evaluation as queried from NAS-Bench-101. The results are shown in Figure 20. Note that 9 cells", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "in the proxy was used for all of our previous experiments as it is also used by the NAS-Bench-101", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 141, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 141, + 198 + ], + "score": 1.0, + "content": "models.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 120, + 506, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "For 9 cells (Figure 20c) in the proxy model, we find that increasing the total number of channels leads", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "to stronger correlation between the proxy model and the full architecture. However, increasing it", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 224, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 505, + 237 + ], + "score": 1.0, + "content": "beyond 16 channels leads to a decrease in correlation. For 3 cells (Figure 20a) the strongest anytime", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 249 + ], + "score": 1.0, + "content": "correlation was interestingly found using only 2 initial channels, with more channels leading to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "worse performance at the beginning and no better final performance. The results suggest that there", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 270 + ], + "score": 1.0, + "content": "exists a set of good combinations between the number of cells and the initial number of channels to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 268, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 281 + ], + "score": 1.0, + "content": "get the maximum correlation between the search and evaluation model. As a future work we plan", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 104, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "to investigate this relationship, which could eventually lead to a more effective bandit-based NAS", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 141, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 141, + 303 + ], + "score": 1.0, + "content": "method.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13, + "bbox_fs": [ + 104, + 202, + 506, + 303 + ] + }, + { + "type": "image", + "bbox": [ + 110, + 312, + 501, + 425 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 312, + 501, + 425 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 312, + 501, + 425 + ], + "spans": [ + { + "bbox": [ + 110, + 312, + 501, + 425 + ], + "score": 0.969, + "type": "image", + "image_path": "9eeb51d8a400b4c88d9e404eccca67fd3f7bd8a3ca37da9e0c5b70cd829ea444.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 110, + 312, + 501, + 349.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 110, + 349.6666666666667, + 501, + 387.33333333333337 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 110, + 387.33333333333337, + 501, + 425.00000000000006 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 433, + 505, + 456 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "Figure 20: In this experiment we varied the total number of cells within [3, 6, 9] and the number of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 444, + 339, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 339, + 457 + ], + "score": 1.0, + "content": "initial channels of the proxy model within [2, 4, 8, 16, 36].", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "index": 20.25 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/SJx9ngStPH/SJx9ngStPH_model.json b/parse/train/SJx9ngStPH/SJx9ngStPH_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d1b000dc9b76ff8902f07f585c83dd61c311ed7 --- /dev/null +++ b/parse/train/SJx9ngStPH/SJx9ngStPH_model.json @@ -0,0 +1,37226 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 709, + 1302, + 709, + 1302, + 1076, + 398, + 1076 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1212, + 1404, + 1212, + 1404, + 1457, + 298, + 1457 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1473, + 1404, + 1473, + 1404, + 1718, + 298, + 1718 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1733, + 1402, + 1733, + 1402, + 1978, + 298, + 1978 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 298, + 218, + 1231, + 218, + 1231, + 381, + 298, + 381 + ], + "score": 0.962 + }, + { + "category_id": 0, + "poly": [ + 302, + 1141, + 573, + 1141, + 573, + 1176, + 302, + 1176 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 334, + 2006, + 530, + 2006, + 530, + 2033, + 334, + 2033 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.879 + }, + { + "category_id": 0, + "poly": [ + 773, + 638, + 927, + 638, + 927, + 671, + 773, + 671 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.715 + }, + { + "category_id": 1, + "poly": [ + 315, + 428, + 962, + 428, + 962, + 559, + 315, + 559 + ], + "score": 0.646 + }, + { + "category_id": 13, + "poly": [ + 488, + 1826, + 550, + 1826, + 550, + 1855, + 488, + 1855 + ], + "score": 0.72, + "latex": "4 2 3 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 393, + 430, + 472, + 430, + 472, + 462, + 393, + 462 + ], + "score": 0.27, + "latex": "\\mathbf { Z e l a ^ { 1 * } }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 218.0, + 771.0, + 218.0, + 771.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 276.0, + 1020.0, + 276.0, + 1020.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 330.0, + 1234.0, + 330.0, + 1234.0, + 382.0, + 295.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1138.0, + 579.0, + 1138.0, + 579.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2001.0, + 536.0, + 2001.0, + 536.0, + 2038.0, + 330.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 634.0, + 934.0, + 634.0, + 934.0, + 677.0, + 768.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 707.0, + 1306.0, + 707.0, + 1306.0, + 747.0, + 394.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 740.0, + 1307.0, + 740.0, + 1307.0, + 776.0, + 393.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 772.0, + 1306.0, + 772.0, + 1306.0, + 807.0, + 394.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 803.0, + 1305.0, + 803.0, + 1305.0, + 835.0, + 395.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 832.0, + 1306.0, + 832.0, + 1306.0, + 867.0, + 391.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 863.0, + 1306.0, + 863.0, + 1306.0, + 896.0, + 394.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 896.0, + 1305.0, + 896.0, + 1305.0, + 925.0, + 395.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 923.0, + 1305.0, + 923.0, + 1305.0, + 956.0, + 394.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 953.0, + 1306.0, + 953.0, + 1306.0, + 987.0, + 393.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 984.0, + 1305.0, + 984.0, + 1305.0, + 1019.0, + 393.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1016.0, + 1304.0, + 1016.0, + 1304.0, + 1048.0, + 395.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1047.0, + 1121.0, + 1047.0, + 1121.0, + 1077.0, + 394.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1214.0, + 1405.0, + 1214.0, + 1405.0, + 1247.0, + 296.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1245.0, + 1405.0, + 1245.0, + 1405.0, + 1278.0, + 294.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1275.0, + 1405.0, + 1275.0, + 1405.0, + 1308.0, + 294.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1303.0, + 1404.0, + 1303.0, + 1404.0, + 1338.0, + 292.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1336.0, + 1404.0, + 1336.0, + 1404.0, + 1369.0, + 294.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1364.0, + 1405.0, + 1364.0, + 1405.0, + 1402.0, + 293.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1406.0, + 1396.0, + 1406.0, + 1433.0, + 293.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1424.0, + 513.0, + 1424.0, + 513.0, + 1462.0, + 294.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1473.0, + 1404.0, + 1473.0, + 1404.0, + 1507.0, + 296.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1504.0, + 1406.0, + 1504.0, + 1406.0, + 1538.0, + 294.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1537.0, + 1404.0, + 1537.0, + 1404.0, + 1567.0, + 296.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1566.0, + 1406.0, + 1566.0, + 1406.0, + 1600.0, + 292.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1595.0, + 1404.0, + 1595.0, + 1404.0, + 1629.0, + 294.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1624.0, + 1405.0, + 1624.0, + 1405.0, + 1660.0, + 293.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1691.0, + 295.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1685.0, + 609.0, + 1685.0, + 609.0, + 1721.0, + 293.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1731.0, + 1403.0, + 1731.0, + 1403.0, + 1770.0, + 292.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1764.0, + 1406.0, + 1764.0, + 1406.0, + 1799.0, + 292.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1796.0, + 1406.0, + 1796.0, + 1406.0, + 1829.0, + 294.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1824.0, + 487.0, + 1824.0, + 487.0, + 1860.0, + 293.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1824.0, + 1406.0, + 1824.0, + 1406.0, + 1860.0, + 551.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1856.0, + 1406.0, + 1856.0, + 1406.0, + 1889.0, + 293.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1885.0, + 1402.0, + 1885.0, + 1402.0, + 1919.0, + 294.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1917.0, + 1404.0, + 1917.0, + 1404.0, + 1949.0, + 293.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1947.0, + 1406.0, + 1947.0, + 1406.0, + 1980.0, + 293.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 429.0, + 392.0, + 429.0, + 392.0, + 464.0, + 312.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 429.0, + 893.0, + 429.0, + 893.0, + 464.0, + 473.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 459.0, + 966.0, + 459.0, + 966.0, + 502.0, + 312.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 490.0, + 769.0, + 490.0, + 769.0, + 532.0, + 312.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 526.0, + 945.0, + 526.0, + 945.0, + 562.0, + 315.0, + 562.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1542, + 1403, + 1542, + 1403, + 1696, + 298, + 1696 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1096, + 1403, + 1096, + 1403, + 1311, + 298, + 1311 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1326, + 1401, + 1326, + 1401, + 1449, + 299, + 1449 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 295, + 865, + 1399, + 865, + 1399, + 928, + 295, + 928 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 295, + 229, + 1399, + 229, + 1399, + 293, + 295, + 293 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 294, + 307, + 1369, + 307, + 1369, + 340, + 294, + 340 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 294, + 1712, + 1353, + 1712, + 1353, + 1744, + 294, + 1744 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 337, + 2006, + 774, + 2006, + 774, + 2033, + 337, + 2033 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 300, + 1485, + 578, + 1485, + 578, + 1516, + 300, + 1516 + ], + "score": 0.876 + }, + { + "category_id": 0, + "poly": [ + 300, + 1039, + 577, + 1039, + 577, + 1071, + 300, + 1071 + ], + "score": 0.842 + }, + { + "category_id": 0, + "poly": [ + 298, + 970, + 872, + 970, + 872, + 1007, + 298, + 1007 + ], + "score": 0.784 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 295, + 361, + 1405, + 361, + 1405, + 840, + 295, + 840 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 297, + 1766, + 1406, + 1766, + 1406, + 1990, + 297, + 1990 + ], + "score": 0.52 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.15 + }, + { + "category_id": 0, + "poly": [ + 301, + 971, + 870, + 971, + 870, + 1006, + 301, + 1006 + ], + "score": 0.109 + }, + { + "category_id": 13, + "poly": [ + 661, + 1249, + 771, + 1249, + 771, + 1282, + 661, + 1282 + ], + "score": 0.92, + "latex": "\\mathcal { O } = \\{ 3 \\mathrm { x } 3 " + }, + { + "category_id": 13, + "poly": [ + 1112, + 1249, + 1284, + 1249, + 1284, + 1283, + 1112, + 1283 + ], + "score": 0.73, + "latex": "3 { \\mathrm { x } } 3 { \\mathrm { ~ m a x } } { \\mathrm { - p o o l } } \\}" + }, + { + "category_id": 13, + "poly": [ + 1125, + 1327, + 1185, + 1327, + 1185, + 1356, + 1125, + 1356 + ], + "score": 0.49, + "latex": "4 2 3 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 2003.0, + 779.0, + 2003.0, + 779.0, + 2037.0, + 333.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1482.0, + 582.0, + 1482.0, + 582.0, + 1520.0, + 293.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1036.0, + 580.0, + 1036.0, + 580.0, + 1076.0, + 294.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 968.0, + 877.0, + 968.0, + 877.0, + 1013.0, + 291.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 971.0, + 876.0, + 971.0, + 876.0, + 1011.0, + 293.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1544.0, + 1404.0, + 1544.0, + 1404.0, + 1577.0, + 296.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1574.0, + 1403.0, + 1574.0, + 1403.0, + 1607.0, + 296.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1603.0, + 1404.0, + 1603.0, + 1404.0, + 1640.0, + 293.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1628.0, + 1405.0, + 1628.0, + 1405.0, + 1675.0, + 291.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1664.0, + 1208.0, + 1664.0, + 1208.0, + 1697.0, + 296.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1099.0, + 1404.0, + 1099.0, + 1404.0, + 1130.0, + 296.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1126.0, + 1407.0, + 1126.0, + 1407.0, + 1160.0, + 293.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1159.0, + 1405.0, + 1159.0, + 1405.0, + 1194.0, + 296.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1190.0, + 1404.0, + 1190.0, + 1404.0, + 1221.0, + 296.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1218.0, + 1403.0, + 1218.0, + 1403.0, + 1253.0, + 294.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1247.0, + 660.0, + 1247.0, + 660.0, + 1286.0, + 293.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 1247.0, + 1111.0, + 1247.0, + 1111.0, + 1286.0, + 772.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1247.0, + 1405.0, + 1247.0, + 1405.0, + 1286.0, + 1285.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1281.0, + 1299.0, + 1281.0, + 1299.0, + 1315.0, + 294.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1327.0, + 1124.0, + 1327.0, + 1124.0, + 1360.0, + 295.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1327.0, + 1403.0, + 1327.0, + 1403.0, + 1360.0, + 1186.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1391.0, + 294.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1386.0, + 1403.0, + 1386.0, + 1403.0, + 1421.0, + 294.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1417.0, + 1123.0, + 1417.0, + 1123.0, + 1453.0, + 294.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 867.0, + 1401.0, + 867.0, + 1401.0, + 899.0, + 298.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 895.0, + 1054.0, + 895.0, + 1054.0, + 933.0, + 293.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1322.0, + 261.0, + 1322.0, + 297.0, + 295.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 304.0, + 1370.0, + 304.0, + 1370.0, + 347.0, + 295.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1707.0, + 1354.0, + 1707.0, + 1354.0, + 1750.0, + 295.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 365.0, + 1405.0, + 365.0, + 1405.0, + 398.0, + 297.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 396.0, + 1402.0, + 396.0, + 1402.0, + 429.0, + 329.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 425.0, + 1405.0, + 425.0, + 1405.0, + 462.0, + 328.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 457.0, + 1403.0, + 457.0, + 1403.0, + 493.0, + 327.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 488.0, + 1403.0, + 488.0, + 1403.0, + 520.0, + 329.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 515.0, + 1405.0, + 515.0, + 1405.0, + 553.0, + 327.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 548.0, + 1405.0, + 548.0, + 1405.0, + 584.0, + 328.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 576.0, + 771.0, + 576.0, + 771.0, + 615.0, + 327.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 612.0, + 1405.0, + 612.0, + 1405.0, + 656.0, + 298.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 648.0, + 1405.0, + 648.0, + 1405.0, + 685.0, + 327.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 677.0, + 467.0, + 677.0, + 467.0, + 713.0, + 330.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 719.0, + 1402.0, + 719.0, + 1402.0, + 752.0, + 297.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 752.0, + 1401.0, + 752.0, + 1401.0, + 780.0, + 331.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 779.0, + 1405.0, + 779.0, + 1405.0, + 816.0, + 328.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 809.0, + 745.0, + 809.0, + 745.0, + 845.0, + 328.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1764.0, + 1404.0, + 1764.0, + 1404.0, + 1800.0, + 296.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1797.0, + 1404.0, + 1797.0, + 1404.0, + 1831.0, + 328.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1825.0, + 1406.0, + 1825.0, + 1406.0, + 1863.0, + 328.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1859.0, + 1309.0, + 1859.0, + 1309.0, + 1893.0, + 328.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1896.0, + 1406.0, + 1896.0, + 1406.0, + 1934.0, + 291.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1926.0, + 1406.0, + 1926.0, + 1406.0, + 1965.0, + 328.0, + 1965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1959.0, + 772.0, + 1959.0, + 772.0, + 1992.0, + 329.0, + 1992.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1269, + 1403, + 1269, + 1403, + 1695, + 298, + 1695 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2033, + 298, + 2033 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1712, + 1401, + 1712, + 1401, + 1834, + 300, + 1834 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 294, + 222, + 1402, + 222, + 1402, + 634, + 294, + 634 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 297, + 654, + 1405, + 654, + 1405, + 809, + 297, + 809 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 1160, + 1404, + 1160, + 1404, + 1252, + 298, + 1252 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 922, + 1400, + 922, + 1400, + 985, + 296, + 985 + ], + "score": 0.948 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 299, + 1080, + 905, + 1080, + 905, + 1113, + 299, + 1113 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2112, + 842, + 2112 + ], + "score": 0.579 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.419 + }, + { + "category_id": 13, + "poly": [ + 885, + 1193, + 911, + 1193, + 911, + 1220, + 885, + 1220 + ], + "score": 0.74, + "latex": "\\mathcal { O }" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 251.0, + 527.0, + 251.0, + 527.0, + 272.0, + 486.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 241.0, + 750.0, + 241.0, + 750.0, + 278.0, + 645.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 252.0, + 824.0, + 252.0, + 824.0, + 274.0, + 768.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 252.0, + 1005.0, + 252.0, + 1005.0, + 269.0, + 966.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 243.0, + 1347.0, + 243.0, + 1347.0, + 285.0, + 1191.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 261.0, + 365.0, + 261.0, + 365.0, + 282.0, + 313.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 291.0, + 370.0, + 291.0, + 370.0, + 315.0, + 306.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 310.0, + 430.0, + 310.0, + 430.0, + 323.0, + 401.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 280.0, + 535.0, + 280.0, + 535.0, + 333.0, + 491.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 275.0, + 591.0, + 275.0, + 591.0, + 299.0, + 557.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 286.0, + 657.0, + 286.0, + 657.0, + 304.0, + 618.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 286.0, + 698.0, + 286.0, + 698.0, + 304.0, + 659.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 285.0, + 771.0, + 285.0, + 771.0, + 339.0, + 714.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 279.0, + 1014.0, + 279.0, + 1014.0, + 333.0, + 966.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 277.0, + 1071.0, + 277.0, + 1071.0, + 299.0, + 1035.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 325.0, + 365.0, + 325.0, + 365.0, + 346.0, + 312.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 331.0, + 440.0, + 331.0, + 440.0, + 343.0, + 401.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 334.0, + 530.0, + 334.0, + 530.0, + 348.0, + 510.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 327.0, + 762.0, + 327.0, + 762.0, + 352.0, + 716.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 350.0, + 429.0, + 350.0, + 429.0, + 364.0, + 403.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 349.0, + 760.0, + 349.0, + 760.0, + 365.0, + 624.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 347.0, + 928.0, + 347.0, + 928.0, + 373.0, + 846.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 357.0, + 375.0, + 357.0, + 375.0, + 376.0, + 303.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 365.0, + 577.0, + 365.0, + 577.0, + 378.0, + 541.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 362.0, + 1059.0, + 362.0, + 1059.0, + 382.0, + 1017.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 386.0, + 366.0, + 386.0, + 366.0, + 407.0, + 312.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 391.0, + 549.0, + 391.0, + 549.0, + 408.0, + 511.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 376.0, + 711.0, + 376.0, + 711.0, + 392.0, + 668.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 391.0, + 1029.0, + 391.0, + 1029.0, + 408.0, + 989.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 384.0, + 1383.0, + 384.0, + 1383.0, + 390.0, + 1376.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 417.0, + 375.0, + 417.0, + 375.0, + 438.0, + 302.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 414.0, + 650.0, + 414.0, + 650.0, + 451.0, + 634.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 413.0, + 1014.0, + 413.0, + 1014.0, + 426.0, + 1005.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 414.0, + 1384.0, + 414.0, + 1384.0, + 423.0, + 1374.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 445.0, + 366.0, + 445.0, + 366.0, + 470.0, + 311.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 446.0, + 648.0, + 446.0, + 648.0, + 456.0, + 637.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 465.0, + 523.0, + 465.0, + 523.0, + 481.0, + 491.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 462.0, + 835.0, + 462.0, + 835.0, + 483.0, + 768.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 465.0, + 1004.0, + 465.0, + 1004.0, + 481.0, + 969.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 477.0, + 379.0, + 477.0, + 379.0, + 501.0, + 297.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 499.0, + 359.0, + 499.0, + 359.0, + 520.0, + 318.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 493.0, + 517.0, + 493.0, + 517.0, + 508.0, + 496.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 488.0, + 591.0, + 488.0, + 591.0, + 510.0, + 556.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 497.0, + 658.0, + 497.0, + 658.0, + 518.0, + 616.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 499.0, + 697.0, + 499.0, + 697.0, + 515.0, + 659.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 497.0, + 762.0, + 497.0, + 762.0, + 518.0, + 721.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 492.0, + 1007.0, + 492.0, + 1007.0, + 509.0, + 969.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 488.0, + 1069.0, + 488.0, + 1069.0, + 510.0, + 1035.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 528.0, + 342.0, + 528.0, + 342.0, + 538.0, + 332.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 528.0, + 533.0, + 528.0, + 533.0, + 545.0, + 498.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 525.0, + 771.0, + 525.0, + 771.0, + 565.0, + 716.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 545.0, + 929.0, + 545.0, + 929.0, + 568.0, + 847.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 526.0, + 1015.0, + 526.0, + 1015.0, + 547.0, + 970.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 538.0, + 1385.0, + 538.0, + 1385.0, + 547.0, + 1374.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 570.0, + 384.0, + 570.0, + 384.0, + 587.0, + 339.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 577.0, + 576.0, + 577.0, + 576.0, + 589.0, + 542.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 561.0, + 761.0, + 561.0, + 761.0, + 581.0, + 623.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 576.0, + 1058.0, + 576.0, + 1058.0, + 593.0, + 1018.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 584.0, + 448.0, + 584.0, + 448.0, + 620.0, + 323.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 604.0, + 549.0, + 604.0, + 549.0, + 620.0, + 511.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 592.0, + 702.0, + 592.0, + 702.0, + 604.0, + 682.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 604.0, + 1028.0, + 604.0, + 1028.0, + 620.0, + 989.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 592.0, + 1310.0, + 592.0, + 1310.0, + 616.0, + 1208.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 538.0, + 1134.0, + 538.0, + 1134.0, + 563.0, + 1074.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 656.0, + 1405.0, + 656.0, + 1405.0, + 689.0, + 296.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 686.0, + 1403.0, + 686.0, + 1403.0, + 720.0, + 295.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 710.0, + 1406.0, + 710.0, + 1406.0, + 755.0, + 291.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 743.0, + 1405.0, + 743.0, + 1405.0, + 787.0, + 291.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 775.0, + 1006.0, + 775.0, + 1006.0, + 814.0, + 294.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1078.0, + 909.0, + 1078.0, + 909.0, + 1116.0, + 292.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1267.0, + 1404.0, + 1267.0, + 1404.0, + 1308.0, + 292.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1299.0, + 1407.0, + 1299.0, + 1407.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1328.0, + 1407.0, + 1328.0, + 1407.0, + 1368.0, + 292.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1361.0, + 1405.0, + 1361.0, + 1405.0, + 1396.0, + 294.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1426.0, + 292.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1457.0, + 294.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1450.0, + 1407.0, + 1450.0, + 1407.0, + 1488.0, + 292.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1481.0, + 1404.0, + 1481.0, + 1404.0, + 1519.0, + 293.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 1403.0, + 1511.0, + 1403.0, + 1550.0, + 293.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1544.0, + 1405.0, + 1544.0, + 1405.0, + 1579.0, + 294.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1572.0, + 1405.0, + 1572.0, + 1405.0, + 1611.0, + 292.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1603.0, + 1405.0, + 1603.0, + 1405.0, + 1639.0, + 293.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1632.0, + 1405.0, + 1632.0, + 1405.0, + 1669.0, + 293.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1667.0, + 702.0, + 1667.0, + 702.0, + 1697.0, + 294.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1885.0, + 296.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1407.0, + 1881.0, + 1407.0, + 1917.0, + 292.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1945.0, + 293.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 1405.0, + 1939.0, + 1405.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2006.0, + 297.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1711.0, + 1407.0, + 1711.0, + 1407.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1776.0, + 294.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 1406.0, + 1772.0, + 1406.0, + 1810.0, + 293.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1804.0, + 1026.0, + 1804.0, + 1026.0, + 1837.0, + 298.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1158.0, + 1404.0, + 1158.0, + 1404.0, + 1196.0, + 293.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1192.0, + 884.0, + 1192.0, + 884.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1192.0, + 1405.0, + 1192.0, + 1405.0, + 1225.0, + 912.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1224.0, + 765.0, + 1224.0, + 765.0, + 1255.0, + 294.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 922.0, + 1405.0, + 922.0, + 1405.0, + 957.0, + 295.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 952.0, + 1175.0, + 952.0, + 1175.0, + 988.0, + 329.0, + 988.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 295, + 1404, + 295, + 1404, + 572, + 298, + 572 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 833, + 1403, + 833, + 1403, + 1050, + 297, + 1050 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1063, + 1404, + 1063, + 1404, + 1279, + 297, + 1279 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 664, + 1403, + 664, + 1403, + 819, + 298, + 819 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1675, + 1402, + 1675, + 1402, + 1830, + 298, + 1830 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1941, + 1402, + 1941, + 1402, + 2035, + 300, + 2035 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1567, + 1403, + 1567, + 1403, + 1661, + 298, + 1661 + ], + "score": 0.965 + }, + { + "category_id": 8, + "poly": [ + 657, + 1845, + 1039, + 1845, + 1039, + 1927, + 657, + 1927 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 296, + 1302, + 1405, + 1302, + 1405, + 1545, + 296, + 1545 + ], + "score": 0.94 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1871, + 1399, + 1871, + 1399, + 1901, + 1366, + 1901 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 302, + 608, + 759, + 608, + 759, + 639, + 302, + 639 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.773 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 1260, + 229, + 1260, + 261, + 300, + 261 + ], + "score": 0.673 + }, + { + "category_id": 14, + "poly": [ + 658, + 1844, + 1039, + 1844, + 1039, + 1928, + 658, + 1928 + ], + "score": 0.95, + "latex": "x ^ { j } = \\sum _ { o \\in \\mathcal { O } } \\frac { \\exp ( \\beta ^ { o } ) } { \\sum _ { o ^ { \\prime } \\in \\mathcal { O } } \\exp ( \\beta ^ { o ^ { \\prime } } ) } o ( I ^ { j } ) ," + }, + { + "category_id": 13, + "poly": [ + 470, + 1336, + 805, + 1336, + 805, + 1391, + 470, + 1391 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { I ^ { j } = \\sum _ { i < j } \\frac { \\exp ( \\alpha ^ { i , j } ) } { \\sum _ { i ^ { \\prime } < j } \\exp ( \\alpha ^ { i ^ { \\prime } , j } ) } x ^ { i } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 949, + 1459, + 1291, + 1459, + 1291, + 1516, + 949, + 1516 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { O ^ { k } = \\oplus _ { j < k } \\frac { \\exp ( \\gamma ^ { j , k } ) } { \\sum _ { j ^ { \\prime } < k } \\exp ( \\gamma ^ { j ^ { \\prime } , k } ) } x ^ { j } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 498, + 1769, + 574, + 1769, + 574, + 1797, + 498, + 1797 + ], + "score": 0.91, + "latex": "o \\in \\mathcal { O }" + }, + { + "category_id": 13, + "poly": [ + 985, + 1430, + 1053, + 1430, + 1053, + 1461, + 985, + 1461 + ], + "score": 0.9, + "latex": "j < k" + }, + { + "category_id": 13, + "poly": [ + 320, + 1425, + 369, + 1425, + 369, + 1461, + 320, + 1461 + ], + "score": 0.9, + "latex": "\\gamma ^ { j , k }" + }, + { + "category_id": 13, + "poly": [ + 685, + 1306, + 760, + 1306, + 760, + 1336, + 685, + 1336 + ], + "score": 0.89, + "latex": "\\textit { i } < \\textit { j }" + }, + { + "category_id": 13, + "poly": [ + 1369, + 1744, + 1400, + 1744, + 1400, + 1769, + 1369, + 1769 + ], + "score": 0.89, + "latex": "\\beta ^ { o }" + }, + { + "category_id": 13, + "poly": [ + 893, + 1342, + 922, + 1342, + 922, + 1372, + 893, + 1372 + ], + "score": 0.87, + "latex": "x ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 320, + 1302, + 368, + 1302, + 368, + 1333, + 320, + 1333 + ], + "score": 0.87, + "latex": "\\alpha ^ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 1260, + 1430, + 1279, + 1430, + 1279, + 1456, + 1260, + 1456 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1297, + 1307, + 1315, + 1307, + 1315, + 1337, + 1297, + 1337 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1206, + 1770, + 1221, + 1770, + 1221, + 1800, + 1206, + 1800 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 952, + 1307, + 969, + 1307, + 969, + 1337, + 952, + 1337 + ], + "score": 0.8, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 693, + 1709, + 719, + 1709, + 719, + 1736, + 693, + 1736 + ], + "score": 0.79, + "latex": "\\mathcal { O }" + }, + { + "category_id": 13, + "poly": [ + 1242, + 1347, + 1256, + 1347, + 1256, + 1372, + 1242, + 1372 + ], + "score": 0.76, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1376, + 1471, + 1401, + 1471, + 1401, + 1498, + 1376, + 1498 + ], + "score": 0.75, + "latex": "\\oplus" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 607.0, + 764.0, + 607.0, + 764.0, + 642.0, + 295.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 225.0, + 1263.0, + 225.0, + 1263.0, + 266.0, + 291.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 295.0, + 1403.0, + 295.0, + 1403.0, + 330.0, + 296.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 324.0, + 1404.0, + 324.0, + 1404.0, + 362.0, + 293.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 354.0, + 1405.0, + 354.0, + 1405.0, + 393.0, + 293.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 386.0, + 1406.0, + 386.0, + 1406.0, + 423.0, + 294.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 416.0, + 1406.0, + 416.0, + 1406.0, + 452.0, + 293.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 448.0, + 1404.0, + 448.0, + 1404.0, + 481.0, + 296.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 477.0, + 1405.0, + 477.0, + 1405.0, + 513.0, + 293.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 509.0, + 1405.0, + 509.0, + 1405.0, + 543.0, + 293.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 536.0, + 1135.0, + 536.0, + 1135.0, + 576.0, + 295.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 833.0, + 1405.0, + 833.0, + 1405.0, + 870.0, + 294.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 863.0, + 1407.0, + 863.0, + 1407.0, + 899.0, + 294.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 893.0, + 1405.0, + 893.0, + 1405.0, + 929.0, + 292.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 924.0, + 1405.0, + 924.0, + 1405.0, + 960.0, + 292.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 958.0, + 1404.0, + 958.0, + 1404.0, + 989.0, + 296.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 985.0, + 1407.0, + 985.0, + 1407.0, + 1021.0, + 294.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1016.0, + 968.0, + 1016.0, + 968.0, + 1052.0, + 294.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1066.0, + 1404.0, + 1066.0, + 1404.0, + 1097.0, + 297.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1094.0, + 1405.0, + 1094.0, + 1405.0, + 1129.0, + 294.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1124.0, + 1406.0, + 1124.0, + 1406.0, + 1162.0, + 292.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1151.0, + 1406.0, + 1151.0, + 1406.0, + 1194.0, + 292.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1183.0, + 1407.0, + 1183.0, + 1407.0, + 1222.0, + 294.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1215.0, + 1406.0, + 1215.0, + 1406.0, + 1254.0, + 294.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1244.0, + 442.0, + 1244.0, + 442.0, + 1286.0, + 293.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 663.0, + 1405.0, + 663.0, + 1405.0, + 702.0, + 293.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 692.0, + 1405.0, + 692.0, + 1405.0, + 734.0, + 293.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 726.0, + 1407.0, + 726.0, + 1407.0, + 761.0, + 293.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 755.0, + 1404.0, + 755.0, + 1404.0, + 793.0, + 293.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 787.0, + 1400.0, + 787.0, + 1400.0, + 824.0, + 294.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1675.0, + 1403.0, + 1675.0, + 1403.0, + 1709.0, + 297.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1705.0, + 692.0, + 1705.0, + 692.0, + 1743.0, + 294.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1705.0, + 1406.0, + 1705.0, + 1406.0, + 1743.0, + 720.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1734.0, + 1368.0, + 1734.0, + 1368.0, + 1774.0, + 293.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1734.0, + 1406.0, + 1734.0, + 1406.0, + 1774.0, + 1401.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1764.0, + 497.0, + 1764.0, + 497.0, + 1806.0, + 293.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1764.0, + 1205.0, + 1764.0, + 1205.0, + 1806.0, + 575.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1764.0, + 1403.0, + 1764.0, + 1403.0, + 1806.0, + 1222.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1796.0, + 1228.0, + 1796.0, + 1228.0, + 1836.0, + 293.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2036.0, + 296.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1567.0, + 1405.0, + 1567.0, + 1405.0, + 1604.0, + 293.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1599.0, + 1403.0, + 1599.0, + 1403.0, + 1633.0, + 294.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1630.0, + 788.0, + 1630.0, + 788.0, + 1664.0, + 296.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1296.0, + 319.0, + 1296.0, + 319.0, + 1343.0, + 293.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1296.0, + 684.0, + 1296.0, + 684.0, + 1343.0, + 369.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1296.0, + 951.0, + 1296.0, + 951.0, + 1343.0, + 761.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1296.0, + 1296.0, + 1296.0, + 1296.0, + 1343.0, + 970.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1296.0, + 1406.0, + 1296.0, + 1406.0, + 1343.0, + 1316.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1330.0, + 469.0, + 1330.0, + 469.0, + 1406.0, + 306.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1330.0, + 892.0, + 1330.0, + 892.0, + 1406.0, + 806.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1330.0, + 1241.0, + 1330.0, + 1241.0, + 1406.0, + 923.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1330.0, + 1415.0, + 1330.0, + 1415.0, + 1406.0, + 1257.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1385.0, + 576.0, + 1385.0, + 576.0, + 1422.0, + 317.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1421.0, + 319.0, + 1421.0, + 319.0, + 1466.0, + 291.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1421.0, + 984.0, + 1421.0, + 984.0, + 1466.0, + 370.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1421.0, + 1259.0, + 1421.0, + 1259.0, + 1466.0, + 1054.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1466.0, + 1280.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1465.0, + 1375.0, + 1465.0, + 1375.0, + 1509.0, + 1292.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1465.0, + 1406.0, + 1465.0, + 1406.0, + 1509.0, + 1402.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1509.0, + 650.0, + 1509.0, + 650.0, + 1545.0, + 317.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.25, + 1450.5, + 1275.25, + 1450.5, + 1275.25, + 1522.5, + 307.25, + 1522.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 900, + 676, + 1398, + 676, + 1398, + 962, + 900, + 962 + ], + "score": 0.977, + "html": "
Search space
No. parents123
Node 1111
Node 2211
Node 3221
Node 4222
Node 5 Output-- 32 2
No.archs.w/ loose ends2
w/o loose ends6240 248729160 3609363648 24066
" + }, + { + "category_id": 1, + "poly": [ + 299, + 590, + 879, + 590, + 879, + 987, + 299, + 987 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1479, + 1405, + 1479, + 1405, + 1693, + 297, + 1693 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 451, + 1404, + 451, + 1404, + 575, + 298, + 575 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1846, + 1400, + 1846, + 1400, + 1938, + 299, + 1938 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1402, + 229, + 1402, + 323, + 298, + 323 + ], + "score": 0.958 + }, + { + "category_id": 6, + "poly": [ + 900, + 590, + 1402, + 590, + 1402, + 652, + 900, + 652 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 295, + 1063, + 1397, + 1063, + 1397, + 1124, + 295, + 1124 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 330, + 2006, + 807, + 2006, + 807, + 2034, + 330, + 2034 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 1761, + 1181, + 1761, + 1181, + 1798, + 299, + 1798 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 300, + 385, + 672, + 385, + 672, + 416, + 300, + 416 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 298, + 986, + 1399, + 986, + 1399, + 1046, + 298, + 1046 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 295, + 1161, + 1405, + 1161, + 1405, + 1441, + 295, + 1441 + ], + "score": 0.768 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.736 + }, + { + "category_id": 13, + "poly": [ + 1269, + 1257, + 1288, + 1257, + 1288, + 1283, + 1269, + 1283 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 588.0, + 1404.0, + 588.0, + 1404.0, + 621.0, + 901.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 621.0, + 1066.0, + 621.0, + 1066.0, + 653.0, + 901.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1999.0, + 814.0, + 1999.0, + 814.0, + 2041.0, + 329.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1760.0, + 1186.0, + 1760.0, + 1186.0, + 1802.0, + 293.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 384.0, + 676.0, + 384.0, + 676.0, + 420.0, + 295.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 589.0, + 881.0, + 589.0, + 881.0, + 622.0, + 296.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 620.0, + 881.0, + 620.0, + 881.0, + 655.0, + 294.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 651.0, + 881.0, + 651.0, + 881.0, + 684.0, + 296.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 680.0, + 882.0, + 680.0, + 882.0, + 713.0, + 295.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 712.0, + 883.0, + 712.0, + 883.0, + 744.0, + 292.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 741.0, + 882.0, + 741.0, + 882.0, + 776.0, + 293.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 772.0, + 882.0, + 772.0, + 882.0, + 805.0, + 294.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 802.0, + 881.0, + 802.0, + 881.0, + 836.0, + 294.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 831.0, + 882.0, + 831.0, + 882.0, + 866.0, + 293.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 863.0, + 881.0, + 863.0, + 881.0, + 896.0, + 294.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 892.0, + 882.0, + 892.0, + 882.0, + 928.0, + 291.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 923.0, + 882.0, + 923.0, + 882.0, + 957.0, + 292.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 954.0, + 882.0, + 954.0, + 882.0, + 991.0, + 295.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1403.0, + 1479.0, + 1403.0, + 1514.0, + 295.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1507.0, + 1405.0, + 1507.0, + 1405.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1539.0, + 1406.0, + 1539.0, + 1406.0, + 1574.0, + 295.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1573.0, + 1402.0, + 1573.0, + 1402.0, + 1603.0, + 295.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1598.0, + 1405.0, + 1598.0, + 1405.0, + 1635.0, + 294.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1629.0, + 1405.0, + 1629.0, + 1405.0, + 1666.0, + 294.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1658.0, + 405.0, + 1658.0, + 405.0, + 1702.0, + 291.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 451.0, + 1406.0, + 451.0, + 1406.0, + 485.0, + 292.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 483.0, + 1405.0, + 483.0, + 1405.0, + 518.0, + 292.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 512.0, + 1405.0, + 512.0, + 1405.0, + 548.0, + 293.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 543.0, + 904.0, + 543.0, + 904.0, + 578.0, + 293.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1846.0, + 1404.0, + 1846.0, + 1404.0, + 1880.0, + 295.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1873.0, + 1405.0, + 1873.0, + 1405.0, + 1914.0, + 293.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1908.0, + 788.0, + 1908.0, + 788.0, + 1942.0, + 294.0, + 1942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 265.0, + 292.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 292.0, + 751.0, + 292.0, + 751.0, + 322.0, + 297.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1062.0, + 1402.0, + 1062.0, + 1402.0, + 1097.0, + 296.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1091.0, + 979.0, + 1091.0, + 979.0, + 1127.0, + 294.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 982.0, + 1404.0, + 982.0, + 1404.0, + 1021.0, + 293.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1010.0, + 337.0, + 1010.0, + 337.0, + 1055.0, + 292.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1161.0, + 1403.0, + 1161.0, + 1403.0, + 1198.0, + 297.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1193.0, + 651.0, + 1193.0, + 651.0, + 1228.0, + 330.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1252.0, + 1268.0, + 1252.0, + 1268.0, + 1292.0, + 293.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 1252.0, + 1402.0, + 1252.0, + 1402.0, + 1292.0, + 1289.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1288.0, + 1402.0, + 1288.0, + 1402.0, + 1321.0, + 331.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1316.0, + 435.0, + 1316.0, + 435.0, + 1352.0, + 327.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1376.0, + 1405.0, + 1376.0, + 1405.0, + 1414.0, + 292.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1409.0, + 1021.0, + 1409.0, + 1021.0, + 1446.0, + 328.0, + 1446.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 978, + 1405, + 978, + 1405, + 1343, + 297, + 1343 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1360, + 1403, + 1360, + 1403, + 1573, + 298, + 1573 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1587, + 1402, + 1587, + 1402, + 1742, + 298, + 1742 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1757, + 1402, + 1757, + 1402, + 1882, + 297, + 1882 + ], + "score": 0.972 + }, + { + "category_id": 0, + "poly": [ + 301, + 592, + 834, + 592, + 834, + 624, + 301, + 624 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 305, + 306, + 559, + 306, + 559, + 528, + 305, + 528 + ], + "score": 0.872 + }, + { + "category_id": 0, + "poly": [ + 859, + 590, + 1084, + 590, + 1084, + 620, + 859, + 620 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 588, + 306, + 1021, + 306, + 1021, + 531, + 588, + 531 + ], + "score": 0.863 + }, + { + "category_id": 0, + "poly": [ + 1047, + 263, + 1275, + 263, + 1275, + 293, + 1047, + 293 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 1053, + 306, + 1390, + 306, + 1390, + 530, + 1053, + 530 + ], + "score": 0.773 + }, + { + "category_id": 1, + "poly": [ + 303, + 636, + 809, + 636, + 809, + 952, + 303, + 952 + ], + "score": 0.772 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2111, + 840, + 2111 + ], + "score": 0.767 + }, + { + "category_id": 0, + "poly": [ + 582, + 263, + 866, + 263, + 866, + 294, + 582, + 294 + ], + "score": 0.707 + }, + { + "category_id": 1, + "poly": [ + 864, + 633, + 1370, + 633, + 1370, + 954, + 864, + 954 + ], + "score": 0.563 + }, + { + "category_id": 0, + "poly": [ + 299, + 263, + 541, + 263, + 541, + 294, + 299, + 294 + ], + "score": 0.543 + }, + { + "category_id": 3, + "poly": [ + 864, + 633, + 1370, + 633, + 1370, + 954, + 864, + 954 + ], + "score": 0.339 + }, + { + "category_id": 3, + "poly": [ + 303, + 636, + 809, + 636, + 809, + 952, + 303, + 952 + ], + "score": 0.122 + }, + { + "category_id": 13, + "poly": [ + 1238, + 1129, + 1400, + 1129, + 1400, + 1163, + 1238, + 1163 + ], + "score": 0.92, + "latex": "\\Lambda = \\{ \\alpha , \\beta , \\gamma \\}" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1069, + 1282, + 1069, + 1282, + 1097, + 1236, + 1097 + ], + "score": 0.87, + "latex": "M ^ { o }" + }, + { + "category_id": 13, + "poly": [ + 612, + 1099, + 641, + 1099, + 641, + 1127, + 612, + 1127 + ], + "score": 0.85, + "latex": "I ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 397, + 770, + 422, + 770, + 422, + 787, + 397, + 787 + ], + "score": 0.72, + "latex": "" + }, + { + "category_id": 13, + "poly": [ + 459, + 885, + 483, + 885, + 483, + 902, + 459, + 902 + ], + "score": 0.5, + "latex": "" + }, + { + "category_id": 13, + "poly": [ + 429, + 930, + 446, + 930, + 446, + 948, + 429, + 948 + ], + "score": 0.45, + "latex": "\\in" + }, + { + "category_id": 13, + "poly": [ + 341, + 311, + 533, + 311, + 533, + 337, + 341, + 337 + ], + "score": 0.38, + "latex": "\\begin{array} { r } { I ^ { j } = \\sum _ { i < j } S ( \\alpha ^ { i , j } ) x ^ { i } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 647, + 1131, + 701, + 1131, + 701, + 1161, + 647, + 1161 + ], + "score": 0.28, + "latex": "( G S )" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 585.0, + 842.0, + 585.0, + 842.0, + 632.0, + 295.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 586.0, + 1087.0, + 586.0, + 1087.0, + 627.0, + 855.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 259.0, + 1277.0, + 259.0, + 1277.0, + 300.0, + 1042.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 863.0, + 2086.0, + 863.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 260.0, + 867.0, + 260.0, + 867.0, + 302.0, + 577.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 260.0, + 543.0, + 260.0, + 543.0, + 302.0, + 295.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 638.0, + 1031.0, + 638.0, + 1031.0, + 662.0, + 1003.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 662.0, + 899.0, + 662.0, + 899.0, + 684.0, + 870.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 669.0, + 1012.0, + 669.0, + 1012.0, + 689.0, + 971.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 669.0, + 1034.0, + 669.0, + 1034.0, + 684.0, + 1013.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 690.0, + 895.0, + 690.0, + 895.0, + 710.0, + 871.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 690.0, + 1068.0, + 690.0, + 1068.0, + 718.0, + 1010.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 712.0, + 1087.0, + 712.0, + 1087.0, + 744.0, + 869.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 738.0, + 1065.0, + 738.0, + 1065.0, + 764.0, + 869.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 758.0, + 895.0, + 758.0, + 895.0, + 788.0, + 867.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 763.0, + 1174.0, + 763.0, + 1174.0, + 789.0, + 924.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 782.0, + 894.0, + 782.0, + 894.0, + 814.0, + 867.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 786.0, + 1370.0, + 786.0, + 1370.0, + 818.0, + 922.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 808.0, + 894.0, + 808.0, + 894.0, + 840.0, + 867.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 815.0, + 1089.0, + 815.0, + 1089.0, + 841.0, + 924.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 833.0, + 895.0, + 833.0, + 895.0, + 863.0, + 866.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 838.0, + 978.0, + 838.0, + 978.0, + 865.0, + 898.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 855.0, + 898.0, + 855.0, + 898.0, + 882.0, + 861.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 859.0, + 1025.0, + 859.0, + 1025.0, + 886.0, + 900.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 879.0, + 898.0, + 879.0, + 898.0, + 909.0, + 858.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 884.0, + 1234.0, + 884.0, + 1234.0, + 911.0, + 924.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 904.0, + 898.0, + 904.0, + 898.0, + 934.0, + 860.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 908.0, + 963.0, + 908.0, + 963.0, + 935.0, + 898.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 924.0, + 1291.0, + 924.0, + 1291.0, + 956.0, + 861.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 631.0, + 1009.0, + 631.0, + 1009.0, + 663.5, + 866.0, + 663.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 670.0, + 977.0, + 670.0, + 977.0, + 684.0, + 942.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 640.0, + 470.0, + 640.0, + 470.0, + 664.0, + 442.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 664.0, + 339.0, + 664.0, + 339.0, + 687.0, + 309.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 672.0, + 365.0, + 672.0, + 365.0, + 683.0, + 344.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 671.0, + 473.0, + 671.0, + 473.0, + 691.0, + 451.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 691.0, + 340.0, + 691.0, + 340.0, + 714.0, + 308.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 692.0, + 507.0, + 692.0, + 507.0, + 720.0, + 449.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 715.0, + 525.0, + 715.0, + 525.0, + 747.0, + 307.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 739.0, + 337.0, + 739.0, + 337.0, + 767.0, + 307.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 740.0, + 506.0, + 740.0, + 506.0, + 771.0, + 334.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 752.0, + 343.0, + 752.0, + 343.0, + 886.0, + 293.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 766.0, + 396.0, + 766.0, + 396.0, + 792.0, + 362.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 766.0, + 646.0, + 766.0, + 646.0, + 792.0, + 423.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 789.0, + 809.0, + 789.0, + 809.0, + 820.0, + 361.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 837.0, + 415.0, + 837.0, + 415.0, + 860.0, + 337.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 857.0, + 471.0, + 857.0, + 471.0, + 880.0, + 339.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 875.0, + 337.0, + 875.0, + 337.0, + 906.0, + 297.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 881.0, + 458.0, + 881.0, + 458.0, + 908.0, + 364.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 881.0, + 709.0, + 881.0, + 709.0, + 908.0, + 484.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 898.0, + 339.0, + 898.0, + 339.0, + 931.0, + 298.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 905.0, + 401.0, + 905.0, + 401.0, + 931.0, + 336.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 920.0, + 428.0, + 920.0, + 428.0, + 953.0, + 297.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 920.0, + 729.0, + 920.0, + 729.0, + 953.0, + 447.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 635.0, + 447.0, + 635.0, + 447.0, + 667.0, + 306.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 667.0, + 453.0, + 667.0, + 453.0, + 694.0, + 369.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 694.5, + 425.0, + 694.5, + 425.0, + 717.5, + 333.0, + 717.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 978.0, + 1403.0, + 978.0, + 1403.0, + 1012.0, + 296.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1007.0, + 1405.0, + 1007.0, + 1405.0, + 1044.0, + 292.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1037.0, + 1405.0, + 1037.0, + 1405.0, + 1072.0, + 294.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1066.0, + 1235.0, + 1066.0, + 1235.0, + 1105.0, + 291.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1066.0, + 1406.0, + 1066.0, + 1406.0, + 1105.0, + 1283.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1100.0, + 611.0, + 1100.0, + 611.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1100.0, + 1405.0, + 1100.0, + 1405.0, + 1134.0, + 642.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1129.0, + 646.0, + 1129.0, + 646.0, + 1167.0, + 292.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1129.0, + 1237.0, + 1129.0, + 1237.0, + 1167.0, + 702.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1129.0, + 1405.0, + 1129.0, + 1405.0, + 1167.0, + 1401.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1160.0, + 1405.0, + 1160.0, + 1405.0, + 1194.0, + 295.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1225.0, + 295.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1219.0, + 1405.0, + 1219.0, + 1405.0, + 1255.0, + 294.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1249.0, + 1406.0, + 1249.0, + 1406.0, + 1288.0, + 291.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1282.0, + 1405.0, + 1282.0, + 1405.0, + 1315.0, + 295.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1314.0, + 986.0, + 1314.0, + 986.0, + 1345.0, + 296.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1358.0, + 1403.0, + 1358.0, + 1403.0, + 1394.0, + 294.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1391.0, + 1400.0, + 1391.0, + 1400.0, + 1421.0, + 296.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1420.0, + 1407.0, + 1420.0, + 1407.0, + 1455.0, + 293.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1451.0, + 1405.0, + 1451.0, + 1405.0, + 1486.0, + 292.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 1405.0, + 1481.0, + 1405.0, + 1515.0, + 294.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 1402.0, + 1511.0, + 1402.0, + 1547.0, + 293.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1544.0, + 494.0, + 1544.0, + 494.0, + 1575.0, + 296.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1582.0, + 1407.0, + 1582.0, + 1407.0, + 1627.0, + 293.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1617.0, + 1404.0, + 1617.0, + 1404.0, + 1655.0, + 293.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1649.0, + 1404.0, + 1649.0, + 1404.0, + 1682.0, + 296.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1404.0, + 1679.0, + 1404.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1710.0, + 1101.0, + 1710.0, + 1101.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1756.0, + 1404.0, + 1756.0, + 1404.0, + 1793.0, + 295.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1788.0, + 1404.0, + 1788.0, + 1404.0, + 1822.0, + 293.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1821.0, + 1404.0, + 1821.0, + 1404.0, + 1851.0, + 293.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1846.0, + 393.0, + 1846.0, + 393.0, + 1889.0, + 292.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 299.0, + 340.0, + 299.0, + 340.0, + 344.0, + 302.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 299.0, + 540.0, + 299.0, + 540.0, + 344.0, + 534.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 327.0, + 545.0, + 327.0, + 545.0, + 369.0, + 302.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 356.0, + 561.0, + 356.0, + 561.0, + 397.0, + 301.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 384.0, + 527.0, + 384.0, + 527.0, + 418.0, + 302.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 410.0, + 505.0, + 410.0, + 505.0, + 440.0, + 306.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 425.0, + 342.0, + 425.0, + 342.0, + 533.0, + 299.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 437.0, + 560.0, + 437.0, + 560.0, + 465.0, + 361.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 458.0, + 565.0, + 458.0, + 565.0, + 489.0, + 360.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 482.0, + 417.0, + 482.0, + 417.0, + 509.0, + 334.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 503.0, + 416.0, + 503.0, + 416.0, + 529.0, + 333.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 301.0, + 824.0, + 301.0, + 824.0, + 342.0, + 585.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 327.0, + 827.0, + 327.0, + 827.0, + 369.0, + 584.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 355.0, + 1024.0, + 355.0, + 1024.0, + 399.0, + 584.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 382.0, + 811.0, + 382.0, + 811.0, + 421.0, + 584.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 410.0, + 787.0, + 410.0, + 787.0, + 440.0, + 589.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 433.0, + 616.0, + 433.0, + 616.0, + 464.0, + 589.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 436.0, + 842.0, + 436.0, + 842.0, + 464.0, + 644.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 449.0, + 620.0, + 449.0, + 620.0, + 512.0, + 586.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 460.0, + 848.0, + 460.0, + 848.0, + 489.0, + 644.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 482.0, + 699.0, + 482.0, + 699.0, + 509.0, + 617.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 496.0, + 624.0, + 496.0, + 624.0, + 527.0, + 588.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 503.0, + 699.0, + 503.0, + 699.0, + 529.0, + 617.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 300.0, + 1293.0, + 300.0, + 1293.0, + 341.0, + 1053.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 325.0, + 1304.0, + 325.0, + 1304.0, + 370.0, + 1048.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 351.0, + 1315.0, + 351.0, + 1315.0, + 401.0, + 1047.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 385.0, + 1275.0, + 385.0, + 1275.0, + 417.0, + 1055.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 408.0, + 1253.0, + 408.0, + 1253.0, + 439.0, + 1055.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 432.0, + 1082.0, + 432.0, + 1082.0, + 464.0, + 1054.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 434.0, + 1307.0, + 434.0, + 1307.0, + 465.0, + 1109.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 451.0, + 1083.0, + 451.0, + 1083.0, + 489.0, + 1053.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 459.0, + 1392.0, + 459.0, + 1392.0, + 488.0, + 1109.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 472.0, + 1089.0, + 472.0, + 1089.0, + 531.0, + 1050.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 482.0, + 1166.0, + 482.0, + 1166.0, + 511.0, + 1081.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 502.0, + 1165.0, + 502.0, + 1165.0, + 531.0, + 1081.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 640.0, + 470.0, + 640.0, + 470.0, + 664.0, + 442.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 664.0, + 339.0, + 664.0, + 339.0, + 687.0, + 309.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 672.0, + 365.0, + 672.0, + 365.0, + 683.0, + 344.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 671.0, + 473.0, + 671.0, + 473.0, + 691.0, + 451.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 691.0, + 340.0, + 691.0, + 340.0, + 714.0, + 308.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 692.0, + 507.0, + 692.0, + 507.0, + 720.0, + 449.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 715.0, + 525.0, + 715.0, + 525.0, + 747.0, + 307.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 739.0, + 337.0, + 739.0, + 337.0, + 767.0, + 307.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 740.0, + 506.0, + 740.0, + 506.0, + 771.0, + 334.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 752.0, + 343.0, + 752.0, + 343.0, + 886.0, + 293.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 766.0, + 396.0, + 766.0, + 396.0, + 792.0, + 362.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 766.0, + 646.0, + 766.0, + 646.0, + 792.0, + 423.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 789.0, + 809.0, + 789.0, + 809.0, + 820.0, + 361.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 837.0, + 415.0, + 837.0, + 415.0, + 860.0, + 337.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 857.0, + 471.0, + 857.0, + 471.0, + 880.0, + 339.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 875.0, + 337.0, + 875.0, + 337.0, + 906.0, + 297.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 881.0, + 458.0, + 881.0, + 458.0, + 908.0, + 364.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 881.0, + 709.0, + 881.0, + 709.0, + 908.0, + 484.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 898.0, + 339.0, + 898.0, + 339.0, + 931.0, + 298.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 905.0, + 401.0, + 905.0, + 401.0, + 931.0, + 336.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 920.0, + 428.0, + 920.0, + 428.0, + 953.0, + 297.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 920.0, + 729.0, + 920.0, + 729.0, + 953.0, + 447.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 635.0, + 447.0, + 635.0, + 447.0, + 667.0, + 306.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 667.0, + 453.0, + 667.0, + 453.0, + 694.0, + 369.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 694.5, + 425.0, + 694.5, + 425.0, + 717.5, + 333.0, + 717.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 638.0, + 1031.0, + 638.0, + 1031.0, + 662.0, + 1003.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 662.0, + 899.0, + 662.0, + 899.0, + 684.0, + 870.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 669.0, + 1012.0, + 669.0, + 1012.0, + 689.0, + 971.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 669.0, + 1034.0, + 669.0, + 1034.0, + 684.0, + 1013.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 690.0, + 895.0, + 690.0, + 895.0, + 710.0, + 871.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 690.0, + 1068.0, + 690.0, + 1068.0, + 718.0, + 1010.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 712.0, + 1087.0, + 712.0, + 1087.0, + 744.0, + 869.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 738.0, + 1065.0, + 738.0, + 1065.0, + 764.0, + 869.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 758.0, + 895.0, + 758.0, + 895.0, + 788.0, + 867.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 763.0, + 1174.0, + 763.0, + 1174.0, + 789.0, + 924.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 782.0, + 894.0, + 782.0, + 894.0, + 814.0, + 867.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 786.0, + 1370.0, + 786.0, + 1370.0, + 818.0, + 922.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 808.0, + 894.0, + 808.0, + 894.0, + 840.0, + 867.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 815.0, + 1089.0, + 815.0, + 1089.0, + 841.0, + 924.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 833.0, + 895.0, + 833.0, + 895.0, + 863.0, + 866.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 838.0, + 978.0, + 838.0, + 978.0, + 865.0, + 898.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 855.0, + 898.0, + 855.0, + 898.0, + 882.0, + 861.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 859.0, + 1025.0, + 859.0, + 1025.0, + 886.0, + 900.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 879.0, + 898.0, + 879.0, + 898.0, + 909.0, + 858.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 884.0, + 1234.0, + 884.0, + 1234.0, + 911.0, + 924.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 904.0, + 898.0, + 904.0, + 898.0, + 934.0, + 860.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 908.0, + 963.0, + 908.0, + 963.0, + 935.0, + 898.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 924.0, + 1291.0, + 924.0, + 1291.0, + 956.0, + 861.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 631.0, + 1009.0, + 631.0, + 1009.0, + 663.5, + 866.0, + 663.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 670.0, + 977.0, + 670.0, + 977.0, + 684.0, + 942.0, + 684.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 294, + 1404, + 294, + 1404, + 509, + 298, + 509 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1649, + 1404, + 1649, + 1404, + 1925, + 297, + 1925 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 601, + 1404, + 601, + 1404, + 818, + 298, + 818 + ], + "score": 0.975 + }, + { + "category_id": 2, + "poly": [ + 298, + 1947, + 1402, + 1947, + 1402, + 2033, + 298, + 2033 + ], + "score": 0.948 + }, + { + "category_id": 0, + "poly": [ + 299, + 1593, + 664, + 1593, + 664, + 1624, + 299, + 1624 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 817, + 76, + 817, + 104, + 300, + 104 + ], + "score": 0.879 + }, + { + "category_id": 0, + "poly": [ + 299, + 546, + 1051, + 546, + 1051, + 576, + 299, + 576 + ], + "score": 0.813 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.696 + }, + { + "category_id": 1, + "poly": [ + 299, + 1453, + 1403, + 1453, + 1403, + 1545, + 299, + 1545 + ], + "score": 0.568 + }, + { + "category_id": 0, + "poly": [ + 293, + 228, + 1348, + 228, + 1348, + 261, + 293, + 261 + ], + "score": 0.543 + }, + { + "category_id": 1, + "poly": [ + 302, + 1006, + 1399, + 1006, + 1399, + 1129, + 302, + 1129 + ], + "score": 0.316 + }, + { + "category_id": 1, + "poly": [ + 298, + 1139, + 1403, + 1139, + 1403, + 1442, + 298, + 1442 + ], + "score": 0.23 + }, + { + "category_id": 1, + "poly": [ + 293, + 228, + 1348, + 228, + 1348, + 261, + 293, + 261 + ], + "score": 0.228 + }, + { + "category_id": 1, + "poly": [ + 298, + 845, + 1405, + 845, + 1405, + 997, + 298, + 997 + ], + "score": 0.211 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.146 + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1942.0, + 913.0, + 1942.0, + 913.0, + 1980.0, + 330.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1971.0, + 1407.0, + 1971.0, + 1407.0, + 2011.0, + 328.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2004.0, + 772.0, + 2004.0, + 772.0, + 2033.0, + 296.0, + 2033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 668.0, + 1590.0, + 668.0, + 1629.0, + 294.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 543.0, + 1056.0, + 543.0, + 1056.0, + 582.0, + 293.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 222.0, + 1352.0, + 222.0, + 1352.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 294.0, + 1406.0, + 294.0, + 1406.0, + 329.0, + 296.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 325.0, + 1404.0, + 325.0, + 1404.0, + 360.0, + 293.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 392.0, + 292.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 384.0, + 1406.0, + 384.0, + 1406.0, + 422.0, + 293.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 417.0, + 1408.0, + 417.0, + 1408.0, + 452.0, + 294.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 445.0, + 1405.0, + 445.0, + 1405.0, + 486.0, + 292.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 478.0, + 904.0, + 478.0, + 904.0, + 515.0, + 293.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1650.0, + 1405.0, + 1650.0, + 1405.0, + 1686.0, + 295.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1679.0, + 1404.0, + 1679.0, + 1404.0, + 1718.0, + 292.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1745.0, + 296.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1741.0, + 1405.0, + 1741.0, + 1405.0, + 1777.0, + 292.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1774.0, + 1405.0, + 1774.0, + 1405.0, + 1807.0, + 295.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1801.0, + 1405.0, + 1801.0, + 1405.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1834.0, + 1405.0, + 1834.0, + 1405.0, + 1867.0, + 295.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1860.0, + 1405.0, + 1860.0, + 1405.0, + 1901.0, + 292.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 490.0, + 1893.0, + 490.0, + 1929.0, + 294.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 598.0, + 1405.0, + 598.0, + 1405.0, + 640.0, + 293.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 636.0, + 1404.0, + 636.0, + 1404.0, + 667.0, + 297.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 664.0, + 1406.0, + 664.0, + 1406.0, + 700.0, + 292.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 693.0, + 1404.0, + 693.0, + 1404.0, + 728.0, + 291.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 726.0, + 1404.0, + 726.0, + 1404.0, + 760.0, + 296.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 754.0, + 1405.0, + 754.0, + 1405.0, + 790.0, + 293.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 788.0, + 1163.0, + 788.0, + 1163.0, + 822.0, + 294.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1449.0, + 1404.0, + 1449.0, + 1404.0, + 1491.0, + 295.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1482.0, + 1404.0, + 1482.0, + 1404.0, + 1518.0, + 317.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1517.0, + 441.0, + 1517.0, + 441.0, + 1545.0, + 319.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1004.0, + 1404.0, + 1004.0, + 1404.0, + 1041.0, + 296.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1037.0, + 1407.0, + 1037.0, + 1407.0, + 1071.0, + 316.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1068.0, + 1405.0, + 1068.0, + 1405.0, + 1104.0, + 317.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1099.0, + 595.0, + 1099.0, + 595.0, + 1131.0, + 319.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1136.0, + 1405.0, + 1136.0, + 1405.0, + 1172.0, + 294.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1168.0, + 1404.0, + 1168.0, + 1404.0, + 1201.0, + 317.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1199.0, + 1403.0, + 1199.0, + 1403.0, + 1234.0, + 318.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1230.0, + 1404.0, + 1230.0, + 1404.0, + 1262.0, + 321.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1259.0, + 1405.0, + 1259.0, + 1405.0, + 1293.0, + 317.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1291.0, + 1405.0, + 1291.0, + 1405.0, + 1326.0, + 318.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1321.0, + 1405.0, + 1321.0, + 1405.0, + 1356.0, + 317.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1353.0, + 1404.0, + 1353.0, + 1404.0, + 1385.0, + 320.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1382.0, + 1405.0, + 1382.0, + 1405.0, + 1418.0, + 318.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1413.0, + 1402.0, + 1413.0, + 1402.0, + 1447.0, + 317.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 222.0, + 1352.0, + 222.0, + 1352.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 842.0, + 1405.0, + 842.0, + 1405.0, + 879.0, + 295.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 874.0, + 1406.0, + 874.0, + 1406.0, + 909.0, + 317.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 904.0, + 1407.0, + 904.0, + 1407.0, + 943.0, + 316.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 937.0, + 1403.0, + 937.0, + 1403.0, + 970.0, + 318.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 967.0, + 896.0, + 967.0, + 896.0, + 1004.0, + 318.0, + 1004.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1505, + 1404, + 1505, + 1404, + 1871, + 298, + 1871 + ], + "score": 0.982 + }, + { + "category_id": 3, + "poly": [ + 305, + 698, + 1396, + 698, + 1396, + 1353, + 305, + 1353 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 303, + 226, + 1396, + 226, + 1396, + 545, + 303, + 545 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 296, + 1378, + 1399, + 1378, + 1399, + 1441, + 296, + 1441 + ], + "score": 0.944 + }, + { + "category_id": 4, + "poly": [ + 296, + 568, + 1403, + 568, + 1403, + 662, + 296, + 662 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 301, + 1972, + 1401, + 1972, + 1401, + 2034, + 301, + 2034 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 298, + 1913, + 916, + 1913, + 916, + 1944, + 298, + 1944 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.785 + }, + { + "category_id": 13, + "poly": [ + 406, + 2003, + 482, + 2003, + 482, + 2035, + 406, + 2035 + ], + "score": 0.92, + "latex": "\\mathcal { L } _ { t r a i n }" + }, + { + "category_id": 13, + "poly": [ + 1125, + 600, + 1155, + 600, + 1155, + 628, + 1125, + 628 + ], + "score": 0.64, + "latex": "\\pm" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 706.0, + 333.0, + 706.0, + 333.0, + 808.0, + 313.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 706.0, + 579.0, + 706.0, + 579.0, + 728.0, + 420.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 705.0, + 942.0, + 705.0, + 942.0, + 721.0, + 784.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 701.0, + 1055.0, + 701.0, + 1055.0, + 804.0, + 1035.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 705.0, + 1317.0, + 705.0, + 1317.0, + 721.0, + 1161.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 725.0, + 348.0, + 725.0, + 348.0, + 745.0, + 326.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 729.0, + 421.0, + 729.0, + 421.0, + 745.0, + 372.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 729.0, + 1089.0, + 729.0, + 1089.0, + 750.0, + 1054.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 752.0, + 348.0, + 752.0, + 348.0, + 771.0, + 326.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 739.0, + 428.0, + 739.0, + 428.0, + 771.0, + 370.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 746.0, + 1089.0, + 746.0, + 1089.0, + 767.0, + 1054.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 763.0, + 432.0, + 763.0, + 432.0, + 783.0, + 370.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 762.0, + 707.0, + 762.0, + 707.0, + 781.0, + 684.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 766.0, + 1087.0, + 766.0, + 1087.0, + 783.0, + 1053.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 790.0, + 370.0, + 790.0, + 370.0, + 804.0, + 355.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 790.0, + 406.0, + 790.0, + 406.0, + 804.0, + 388.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 790.0, + 440.0, + 790.0, + 440.0, + 804.0, + 422.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 790.0, + 475.0, + 790.0, + 475.0, + 804.0, + 455.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 790.0, + 509.0, + 790.0, + 509.0, + 804.0, + 489.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 790.0, + 542.0, + 790.0, + 542.0, + 804.0, + 524.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 790.0, + 575.0, + 790.0, + 575.0, + 804.0, + 558.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 790.0, + 610.0, + 790.0, + 610.0, + 804.0, + 591.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 790.0, + 643.0, + 790.0, + 643.0, + 804.0, + 626.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 787.0, + 728.0, + 787.0, + 728.0, + 799.0, + 717.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 786.0, + 766.0, + 786.0, + 766.0, + 801.0, + 748.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 786.0, + 802.0, + 786.0, + 802.0, + 801.0, + 782.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 786.0, + 837.0, + 786.0, + 837.0, + 801.0, + 818.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 786.0, + 873.0, + 786.0, + 873.0, + 801.0, + 853.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 786.0, + 908.0, + 786.0, + 908.0, + 801.0, + 889.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 786.0, + 942.0, + 786.0, + 942.0, + 801.0, + 924.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 786.0, + 978.0, + 786.0, + 978.0, + 801.0, + 960.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 786.0, + 1013.0, + 786.0, + 1013.0, + 801.0, + 994.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 787.0, + 1109.0, + 787.0, + 1109.0, + 799.0, + 1097.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 786.0, + 1147.0, + 786.0, + 1147.0, + 801.0, + 1127.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 786.0, + 1181.0, + 786.0, + 1181.0, + 801.0, + 1162.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 786.0, + 1214.0, + 786.0, + 1214.0, + 801.0, + 1195.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 786.0, + 1249.0, + 786.0, + 1249.0, + 801.0, + 1229.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 786.0, + 1282.0, + 786.0, + 1282.0, + 801.0, + 1264.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 786.0, + 1316.0, + 786.0, + 1316.0, + 801.0, + 1297.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 786.0, + 1349.0, + 786.0, + 1349.0, + 801.0, + 1332.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 786.0, + 1383.0, + 786.0, + 1383.0, + 801.0, + 1365.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 799.0, + 517.0, + 799.0, + 517.0, + 817.0, + 482.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 797.0, + 880.0, + 797.0, + 880.0, + 814.0, + 846.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 797.0, + 1256.0, + 797.0, + 1256.0, + 814.0, + 1221.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 827.0, + 326.0, + 827.0, + 326.0, + 927.0, + 306.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 829.0, + 517.0, + 829.0, + 517.0, + 847.0, + 480.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 824.0, + 692.0, + 824.0, + 692.0, + 927.0, + 672.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 824.0, + 883.0, + 824.0, + 883.0, + 845.0, + 843.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 824.0, + 1061.0, + 824.0, + 1061.0, + 927.0, + 1042.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 824.0, + 1252.0, + 824.0, + 1252.0, + 845.0, + 1213.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 843.0, + 347.0, + 843.0, + 347.0, + 862.0, + 326.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 847.0, + 705.0, + 847.0, + 705.0, + 865.0, + 684.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 855.0, + 1076.0, + 855.0, + 1076.0, + 875.0, + 1053.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 867.0, + 347.0, + 867.0, + 347.0, + 885.0, + 326.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 889.0, + 348.0, + 889.0, + 348.0, + 908.0, + 322.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 883.0, + 707.0, + 883.0, + 707.0, + 901.0, + 683.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 876.0, + 1076.0, + 876.0, + 1076.0, + 896.0, + 1051.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 910.0, + 370.0, + 910.0, + 370.0, + 925.0, + 355.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 910.0, + 406.0, + 910.0, + 406.0, + 925.0, + 388.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 910.0, + 440.0, + 910.0, + 440.0, + 925.0, + 421.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 910.0, + 475.0, + 910.0, + 475.0, + 925.0, + 455.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 910.0, + 507.0, + 910.0, + 507.0, + 925.0, + 489.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 910.0, + 542.0, + 910.0, + 542.0, + 925.0, + 524.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 910.0, + 575.0, + 910.0, + 575.0, + 925.0, + 558.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 910.0, + 610.0, + 910.0, + 610.0, + 925.0, + 591.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 910.0, + 643.0, + 910.0, + 643.0, + 925.0, + 626.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 908.0, + 729.0, + 908.0, + 729.0, + 925.0, + 715.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 909.0, + 767.0, + 909.0, + 767.0, + 924.0, + 748.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 909.0, + 802.0, + 909.0, + 802.0, + 924.0, + 784.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 909.0, + 837.0, + 909.0, + 837.0, + 924.0, + 818.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 909.0, + 873.0, + 909.0, + 873.0, + 924.0, + 853.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 909.0, + 908.0, + 909.0, + 908.0, + 924.0, + 889.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 909.0, + 942.0, + 909.0, + 942.0, + 924.0, + 925.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 909.0, + 978.0, + 909.0, + 978.0, + 924.0, + 958.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 909.0, + 1013.0, + 909.0, + 1013.0, + 924.0, + 994.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 910.0, + 1097.0, + 910.0, + 1097.0, + 922.0, + 1086.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 909.0, + 1136.0, + 909.0, + 1136.0, + 924.0, + 1118.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 909.0, + 1172.0, + 909.0, + 1172.0, + 924.0, + 1153.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 909.0, + 1207.0, + 909.0, + 1207.0, + 924.0, + 1188.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 909.0, + 1243.0, + 909.0, + 1243.0, + 924.0, + 1223.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 906.0, + 1280.0, + 906.0, + 1280.0, + 925.0, + 1257.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 908.0, + 1312.0, + 908.0, + 1312.0, + 924.0, + 1295.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 909.0, + 1348.0, + 909.0, + 1348.0, + 924.0, + 1329.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 909.0, + 1383.0, + 909.0, + 1383.0, + 924.0, + 1364.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 920.0, + 517.0, + 920.0, + 517.0, + 937.0, + 482.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 919.0, + 880.0, + 919.0, + 880.0, + 936.0, + 846.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 919.0, + 1250.0, + 919.0, + 1250.0, + 936.0, + 1215.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 949.0, + 529.0, + 949.0, + 529.0, + 968.0, + 468.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 947.0, + 703.0, + 947.0, + 703.0, + 1050.0, + 683.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 949.0, + 899.0, + 949.0, + 899.0, + 968.0, + 838.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 949.0, + 1064.0, + 949.0, + 1064.0, + 1048.0, + 1048.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 949.0, + 1271.0, + 949.0, + 1271.0, + 968.0, + 1212.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 966.0, + 348.0, + 966.0, + 348.0, + 986.0, + 326.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 963.0, + 718.0, + 963.0, + 718.0, + 985.0, + 694.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 967.0, + 1118.0, + 967.0, + 1118.0, + 982.0, + 1102.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 985.0, + 1094.0, + 985.0, + 1094.0, + 1002.0, + 1065.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 993.0, + 719.0, + 993.0, + 719.0, + 1013.0, + 695.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1011.0, + 1092.0, + 1011.0, + 1092.0, + 1028.0, + 1061.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1030.0, + 370.0, + 1030.0, + 370.0, + 1045.0, + 355.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1030.0, + 406.0, + 1030.0, + 406.0, + 1045.0, + 388.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1029.0, + 441.0, + 1029.0, + 441.0, + 1048.0, + 420.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1029.0, + 476.0, + 1029.0, + 476.0, + 1048.0, + 454.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1030.0, + 508.0, + 1030.0, + 508.0, + 1045.0, + 489.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1029.0, + 544.0, + 1029.0, + 544.0, + 1048.0, + 523.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1029.0, + 578.0, + 1029.0, + 578.0, + 1048.0, + 556.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1030.0, + 610.0, + 1030.0, + 610.0, + 1045.0, + 591.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1030.0, + 643.0, + 1030.0, + 643.0, + 1045.0, + 626.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1033.0, + 739.0, + 1033.0, + 739.0, + 1045.0, + 728.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1029.0, + 779.0, + 1029.0, + 779.0, + 1048.0, + 756.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1029.0, + 812.0, + 1029.0, + 812.0, + 1048.0, + 791.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1029.0, + 846.0, + 1029.0, + 846.0, + 1048.0, + 823.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1029.0, + 879.0, + 1029.0, + 879.0, + 1048.0, + 857.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1029.0, + 914.0, + 1029.0, + 914.0, + 1048.0, + 893.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1029.0, + 947.0, + 1029.0, + 947.0, + 1048.0, + 926.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1032.0, + 981.0, + 1032.0, + 981.0, + 1047.0, + 962.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1030.0, + 1013.0, + 1030.0, + 1013.0, + 1047.0, + 996.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 1032.0, + 1116.0, + 1032.0, + 1116.0, + 1047.0, + 1101.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1032.0, + 1151.0, + 1032.0, + 1151.0, + 1047.0, + 1132.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1029.0, + 1185.0, + 1029.0, + 1185.0, + 1048.0, + 1164.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1032.0, + 1218.0, + 1032.0, + 1218.0, + 1047.0, + 1198.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1030.0, + 1250.0, + 1030.0, + 1250.0, + 1045.0, + 1231.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1032.0, + 1285.0, + 1032.0, + 1285.0, + 1047.0, + 1266.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 1032.0, + 1318.0, + 1032.0, + 1318.0, + 1047.0, + 1298.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 1032.0, + 1350.0, + 1032.0, + 1350.0, + 1047.0, + 1333.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 1029.0, + 1385.0, + 1029.0, + 1385.0, + 1048.0, + 1364.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1040.0, + 517.0, + 1040.0, + 517.0, + 1058.0, + 482.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1042.0, + 886.0, + 1042.0, + 886.0, + 1059.0, + 852.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1042.0, + 1259.0, + 1042.0, + 1259.0, + 1059.0, + 1224.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1069.0, + 324.0, + 1069.0, + 324.0, + 1168.0, + 308.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 1070.0, + 516.0, + 1070.0, + 516.0, + 1088.0, + 481.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1059.0, + 717.0, + 1059.0, + 717.0, + 1182.0, + 662.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1070.0, + 882.0, + 1070.0, + 882.0, + 1091.0, + 843.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1066.0, + 1059.0, + 1066.0, + 1059.0, + 1171.0, + 1039.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1070.0, + 1255.0, + 1070.0, + 1255.0, + 1091.0, + 1216.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1093.0, + 367.0, + 1093.0, + 367.0, + 1099.0, + 360.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1090.0, + 1082.0, + 1090.0, + 1082.0, + 1110.0, + 1060.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1096.0, + 348.0, + 1096.0, + 348.0, + 1115.0, + 326.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1110.0, + 1082.0, + 1110.0, + 1082.0, + 1129.0, + 1060.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1120.0, + 347.0, + 1120.0, + 347.0, + 1140.0, + 322.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1127.0, + 743.0, + 1127.0, + 743.0, + 1142.0, + 728.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1130.0, + 940.0, + 1130.0, + 940.0, + 1141.0, + 929.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1127.0, + 1081.0, + 1127.0, + 1081.0, + 1146.0, + 1056.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1136.0, + 728.0, + 1136.0, + 728.0, + 1148.0, + 718.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1151.0, + 372.0, + 1151.0, + 372.0, + 1167.0, + 355.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1150.0, + 408.0, + 1150.0, + 408.0, + 1168.0, + 387.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1150.0, + 441.0, + 1150.0, + 441.0, + 1168.0, + 420.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1150.0, + 476.0, + 1150.0, + 476.0, + 1168.0, + 454.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1151.0, + 508.0, + 1151.0, + 508.0, + 1166.0, + 489.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1151.0, + 542.0, + 1151.0, + 542.0, + 1166.0, + 524.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1152.0, + 575.0, + 1152.0, + 575.0, + 1167.0, + 558.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1151.0, + 610.0, + 1151.0, + 610.0, + 1166.0, + 591.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1151.0, + 645.0, + 1151.0, + 645.0, + 1166.0, + 626.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1156.0, + 726.0, + 1156.0, + 726.0, + 1167.0, + 717.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1153.0, + 766.0, + 1153.0, + 766.0, + 1168.0, + 748.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1152.0, + 803.0, + 1152.0, + 803.0, + 1171.0, + 781.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1152.0, + 838.0, + 1152.0, + 838.0, + 1171.0, + 817.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1153.0, + 872.0, + 1153.0, + 872.0, + 1168.0, + 853.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1153.0, + 908.0, + 1153.0, + 908.0, + 1168.0, + 889.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1153.0, + 942.0, + 1153.0, + 942.0, + 1168.0, + 925.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1153.0, + 978.0, + 1153.0, + 978.0, + 1168.0, + 960.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1153.0, + 1013.0, + 1153.0, + 1013.0, + 1168.0, + 994.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1153.0, + 1106.0, + 1153.0, + 1106.0, + 1170.0, + 1091.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1152.0, + 1143.0, + 1152.0, + 1143.0, + 1171.0, + 1122.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1153.0, + 1177.0, + 1153.0, + 1177.0, + 1168.0, + 1157.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1152.0, + 1213.0, + 1152.0, + 1213.0, + 1171.0, + 1190.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1153.0, + 1245.0, + 1153.0, + 1245.0, + 1168.0, + 1226.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1153.0, + 1280.0, + 1153.0, + 1280.0, + 1170.0, + 1262.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1153.0, + 1313.0, + 1153.0, + 1313.0, + 1168.0, + 1296.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1153.0, + 1349.0, + 1153.0, + 1349.0, + 1168.0, + 1331.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 1153.0, + 1383.0, + 1153.0, + 1383.0, + 1168.0, + 1365.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1162.0, + 517.0, + 1162.0, + 517.0, + 1179.0, + 482.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1165.0, + 880.0, + 1165.0, + 880.0, + 1182.0, + 846.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1165.0, + 1254.0, + 1165.0, + 1254.0, + 1182.0, + 1218.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1177.0, + 360.0, + 1177.0, + 360.0, + 1301.0, + 302.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1191.0, + 534.0, + 1191.0, + 534.0, + 1211.0, + 463.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1192.0, + 697.0, + 1192.0, + 697.0, + 1295.0, + 677.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1193.0, + 901.0, + 1193.0, + 901.0, + 1213.0, + 831.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1193.0, + 1059.0, + 1193.0, + 1059.0, + 1292.0, + 1043.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1193.0, + 1269.0, + 1193.0, + 1269.0, + 1213.0, + 1198.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1203.0, + 348.0, + 1203.0, + 348.0, + 1222.0, + 324.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 1220.0, + 713.0, + 1220.0, + 713.0, + 1239.0, + 692.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1229.0, + 1076.0, + 1229.0, + 1076.0, + 1249.0, + 1053.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1249.0, + 713.0, + 1249.0, + 713.0, + 1269.0, + 690.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1259.0, + 1076.0, + 1259.0, + 1076.0, + 1279.0, + 1053.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1273.0, + 370.0, + 1273.0, + 370.0, + 1288.0, + 355.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1273.0, + 406.0, + 1273.0, + 406.0, + 1288.0, + 388.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1273.0, + 440.0, + 1273.0, + 440.0, + 1288.0, + 421.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1270.0, + 476.0, + 1270.0, + 476.0, + 1289.0, + 454.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1271.0, + 509.0, + 1271.0, + 509.0, + 1286.0, + 489.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1273.0, + 542.0, + 1273.0, + 542.0, + 1288.0, + 524.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1273.0, + 575.0, + 1273.0, + 575.0, + 1286.0, + 558.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1271.0, + 610.0, + 1271.0, + 610.0, + 1286.0, + 591.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1270.0, + 646.0, + 1270.0, + 646.0, + 1289.0, + 625.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1276.0, + 735.0, + 1276.0, + 735.0, + 1292.0, + 720.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1276.0, + 772.0, + 1276.0, + 772.0, + 1291.0, + 754.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1276.0, + 807.0, + 1276.0, + 807.0, + 1291.0, + 787.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1276.0, + 842.0, + 1276.0, + 842.0, + 1291.0, + 822.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1276.0, + 877.0, + 1276.0, + 877.0, + 1291.0, + 857.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1276.0, + 911.0, + 1276.0, + 911.0, + 1291.0, + 891.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1276.0, + 945.0, + 1276.0, + 945.0, + 1291.0, + 925.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1276.0, + 980.0, + 1276.0, + 980.0, + 1291.0, + 960.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1276.0, + 1013.0, + 1276.0, + 1013.0, + 1291.0, + 994.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1278.0, + 1097.0, + 1278.0, + 1097.0, + 1290.0, + 1086.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1276.0, + 1137.0, + 1276.0, + 1137.0, + 1291.0, + 1117.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1276.0, + 1172.0, + 1276.0, + 1172.0, + 1291.0, + 1152.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1276.0, + 1207.0, + 1276.0, + 1207.0, + 1291.0, + 1188.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 1276.0, + 1243.0, + 1276.0, + 1243.0, + 1291.0, + 1223.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 1276.0, + 1279.0, + 1276.0, + 1279.0, + 1291.0, + 1259.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1276.0, + 1313.0, + 1276.0, + 1313.0, + 1291.0, + 1293.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 1276.0, + 1348.0, + 1276.0, + 1348.0, + 1291.0, + 1329.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 1276.0, + 1383.0, + 1276.0, + 1383.0, + 1291.0, + 1364.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1283.0, + 517.0, + 1283.0, + 517.0, + 1300.0, + 482.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1286.0, + 884.0, + 1286.0, + 884.0, + 1304.0, + 849.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1286.0, + 1251.0, + 1286.0, + 1251.0, + 1304.0, + 1215.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1316.0, + 575.0, + 1316.0, + 575.0, + 1349.0, + 383.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1321.0, + 944.0, + 1321.0, + 944.0, + 1354.0, + 749.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1321.0, + 1313.0, + 1321.0, + 1313.0, + 1354.0, + 1120.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.25, + 761.5, + 566.25, + 761.5, + 566.25, + 770.0, + 561.25, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1119.0, + 963.0, + 1119.0, + 963.0, + 1128.0, + 939.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1251.0, + 565.0, + 1251.0, + 565.0, + 1263.0, + 552.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 234.0, + 342.0, + 234.0, + 342.0, + 251.0, + 318.0, + 251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 238.0, + 615.0, + 238.0, + 615.0, + 253.0, + 528.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 237.0, + 649.0, + 237.0, + 649.0, + 255.0, + 625.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 232.0, + 709.0, + 232.0, + 709.0, + 250.0, + 680.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 238.0, + 975.0, + 238.0, + 975.0, + 253.0, + 891.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 236.0, + 1012.0, + 236.0, + 1012.0, + 253.0, + 989.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 238.0, + 1273.0, + 238.0, + 1273.0, + 253.0, + 1189.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 238.0, + 1381.0, + 238.0, + 1381.0, + 252.0, + 1361.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 246.0, + 625.0, + 246.0, + 625.0, + 268.0, + 526.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 246.0, + 989.0, + 246.0, + 989.0, + 268.0, + 888.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 246.0, + 1285.0, + 246.0, + 1285.0, + 268.0, + 1187.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 261.0, + 559.0, + 261.0, + 559.0, + 276.0, + 526.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 261.0, + 923.0, + 261.0, + 923.0, + 276.0, + 891.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 261.0, + 1092.0, + 261.0, + 1092.0, + 278.0, + 1047.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 261.0, + 1220.0, + 261.0, + 1220.0, + 277.0, + 1189.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 271.0, + 576.0, + 271.0, + 576.0, + 290.0, + 525.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 272.0, + 938.0, + 272.0, + 938.0, + 288.0, + 889.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 271.0, + 1238.0, + 271.0, + 1238.0, + 290.0, + 1188.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 286.0, + 556.0, + 286.0, + 556.0, + 299.0, + 528.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 286.0, + 919.0, + 286.0, + 919.0, + 299.0, + 891.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 283.0, + 1219.0, + 283.0, + 1219.0, + 300.0, + 1189.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 307.0, + 325.0, + 307.0, + 325.0, + 395.0, + 305.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 296.0, + 582.0, + 296.0, + 582.0, + 312.0, + 528.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 299.0, + 661.0, + 299.0, + 661.0, + 405.0, + 644.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 307.0, + 687.0, + 307.0, + 687.0, + 395.0, + 667.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 297.0, + 945.0, + 297.0, + 945.0, + 309.0, + 892.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 297.0, + 1022.0, + 297.0, + 1022.0, + 403.0, + 1006.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 332.0, + 1091.0, + 332.0, + 1091.0, + 349.0, + 1050.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 297.0, + 1242.0, + 297.0, + 1242.0, + 309.0, + 1221.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 297.0, + 1393.0, + 297.0, + 1393.0, + 403.0, + 1376.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 427.0, + 708.0, + 427.0, + 708.0, + 444.0, + 680.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 449.0, + 345.0, + 449.0, + 345.0, + 465.0, + 318.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 445.0, + 1089.0, + 445.0, + 1089.0, + 460.0, + 1062.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 469.0, + 351.0, + 469.0, + 351.0, + 480.0, + 341.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 468.0, + 412.0, + 468.0, + 412.0, + 483.0, + 395.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 468.0, + 469.0, + 468.0, + 469.0, + 482.0, + 451.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 468.0, + 526.0, + 468.0, + 526.0, + 482.0, + 509.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 468.0, + 584.0, + 468.0, + 584.0, + 482.0, + 566.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 469.0, + 715.0, + 469.0, + 715.0, + 480.0, + 704.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 468.0, + 775.0, + 468.0, + 775.0, + 482.0, + 758.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 468.0, + 832.0, + 468.0, + 832.0, + 482.0, + 813.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 468.0, + 889.0, + 468.0, + 889.0, + 482.0, + 872.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 468.0, + 946.0, + 468.0, + 946.0, + 482.0, + 929.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 469.0, + 1098.0, + 469.0, + 1098.0, + 480.0, + 1088.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 468.0, + 1155.0, + 468.0, + 1155.0, + 482.0, + 1139.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 468.0, + 1210.0, + 468.0, + 1210.0, + 482.0, + 1192.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 468.0, + 1265.0, + 468.0, + 1265.0, + 482.0, + 1247.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 468.0, + 1319.0, + 468.0, + 1319.0, + 482.0, + 1302.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 478.0, + 503.0, + 478.0, + 503.0, + 495.0, + 470.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 478.0, + 866.0, + 478.0, + 866.0, + 495.0, + 832.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 478.0, + 1242.0, + 478.0, + 1242.0, + 494.0, + 1209.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 511.0, + 578.0, + 511.0, + 578.0, + 545.0, + 387.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 511.0, + 943.0, + 511.0, + 943.0, + 545.0, + 750.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 509.0, + 1314.0, + 509.0, + 1314.0, + 547.0, + 1118.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1376.0, + 1400.0, + 1376.0, + 1400.0, + 1412.0, + 295.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1410.0, + 911.0, + 1410.0, + 911.0, + 1441.0, + 296.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 565.0, + 1404.0, + 565.0, + 1404.0, + 604.0, + 294.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 598.0, + 1124.0, + 598.0, + 1124.0, + 632.0, + 295.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 598.0, + 1405.0, + 598.0, + 1405.0, + 632.0, + 1156.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 627.0, + 1023.0, + 627.0, + 1023.0, + 664.0, + 294.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 920.0, + 1910.0, + 920.0, + 1949.0, + 294.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1506.0, + 1404.0, + 1506.0, + 1404.0, + 1540.0, + 296.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1535.0, + 1404.0, + 1535.0, + 1404.0, + 1569.0, + 293.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1565.0, + 1403.0, + 1565.0, + 1403.0, + 1601.0, + 295.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1597.0, + 1405.0, + 1597.0, + 1405.0, + 1631.0, + 293.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1626.0, + 1405.0, + 1626.0, + 1405.0, + 1660.0, + 294.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1658.0, + 1406.0, + 1658.0, + 1406.0, + 1692.0, + 294.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1688.0, + 1404.0, + 1688.0, + 1404.0, + 1722.0, + 293.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1717.0, + 1406.0, + 1717.0, + 1406.0, + 1754.0, + 292.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1749.0, + 1405.0, + 1749.0, + 1405.0, + 1783.0, + 293.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1779.0, + 1406.0, + 1779.0, + 1406.0, + 1814.0, + 293.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1807.0, + 1408.0, + 1807.0, + 1408.0, + 1847.0, + 292.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1839.0, + 488.0, + 1839.0, + 488.0, + 1874.0, + 293.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 405.0, + 2002.0, + 405.0, + 2041.0, + 295.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2041.0, + 483.0, + 2041.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1469, + 769, + 1469, + 769, + 1895, + 298, + 1895 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 721, + 1404, + 721, + 1404, + 906, + 297, + 906 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 302, + 226, + 1396, + 226, + 1396, + 542, + 302, + 542 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 802, + 1471, + 1388, + 1471, + 1388, + 1789, + 802, + 1789 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 300, + 1360, + 1405, + 1360, + 1405, + 1453, + 300, + 1453 + ], + "score": 0.964 + }, + { + "category_id": 4, + "poly": [ + 297, + 568, + 1404, + 568, + 1404, + 661, + 297, + 661 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 790, + 1823, + 1402, + 1823, + 1402, + 1947, + 790, + 1947 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 296, + 1974, + 1404, + 1974, + 1404, + 2035, + 296, + 2035 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 296, + 1912, + 769, + 1912, + 769, + 1974, + 296, + 1974 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 296, + 934, + 1406, + 934, + 1406, + 1266, + 296, + 1266 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.785 + }, + { + "category_id": 0, + "poly": [ + 295, + 1302, + 1003, + 1302, + 1003, + 1335, + 295, + 1335 + ], + "score": 0.742 + }, + { + "category_id": 0, + "poly": [ + 299, + 1303, + 1002, + 1303, + 1002, + 1335, + 299, + 1335 + ], + "score": 0.647 + }, + { + "category_id": 13, + "poly": [ + 880, + 785, + 916, + 785, + 916, + 814, + 880, + 814 + ], + "score": 0.9, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1198, + 845, + 1233, + 845, + 1233, + 875, + 1198, + 875 + ], + "score": 0.88, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 614, + 1235, + 648, + 1235, + 648, + 1264, + 614, + 1264 + ], + "score": 0.88, + "latex": "L _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 238.0, + 592.0, + 238.0, + 592.0, + 255.0, + 465.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 231.0, + 649.0, + 231.0, + 649.0, + 249.0, + 604.0, + 249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 232.0, + 1013.0, + 232.0, + 1013.0, + 257.0, + 894.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 231.0, + 1094.0, + 231.0, + 1094.0, + 249.0, + 1049.0, + 249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 232.0, + 1383.0, + 232.0, + 1383.0, + 255.0, + 1197.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 246.0, + 356.0, + 246.0, + 356.0, + 264.0, + 315.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 251.0, + 587.0, + 251.0, + 587.0, + 264.0, + 466.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 250.0, + 1335.0, + 250.0, + 1335.0, + 267.0, + 1197.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 263.0, + 601.0, + 263.0, + 601.0, + 276.0, + 466.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 288.0, + 357.0, + 288.0, + 357.0, + 307.0, + 314.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 272.0, + 601.0, + 272.0, + 601.0, + 289.0, + 465.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 281.0, + 661.0, + 281.0, + 661.0, + 405.0, + 605.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 294.0, + 729.0, + 294.0, + 729.0, + 346.0, + 675.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 296.0, + 1028.0, + 296.0, + 1028.0, + 406.0, + 968.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 276.0, + 1090.0, + 276.0, + 1090.0, + 294.0, + 1050.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 286.0, + 1397.0, + 286.0, + 1397.0, + 407.0, + 1338.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 361.0, + 359.0, + 361.0, + 359.0, + 380.0, + 314.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 367.0, + 646.0, + 367.0, + 646.0, + 385.0, + 605.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 381.0, + 1013.0, + 381.0, + 1013.0, + 400.0, + 973.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 379.0, + 1381.0, + 379.0, + 1381.0, + 396.0, + 1342.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 433.0, + 1092.0, + 433.0, + 1092.0, + 452.0, + 1064.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 442.0, + 607.0, + 442.0, + 607.0, + 459.0, + 554.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 452.0, + 635.0, + 452.0, + 635.0, + 470.0, + 601.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 468.0, + 364.0, + 468.0, + 364.0, + 480.0, + 354.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 467.0, + 417.0, + 467.0, + 417.0, + 483.0, + 399.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 467.0, + 467.0, + 467.0, + 467.0, + 483.0, + 449.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 467.0, + 517.0, + 467.0, + 517.0, + 483.0, + 500.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 467.0, + 567.0, + 467.0, + 567.0, + 483.0, + 548.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 466.0, + 620.0, + 466.0, + 620.0, + 484.0, + 595.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 465.0, + 736.0, + 465.0, + 736.0, + 482.0, + 721.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 466.0, + 787.0, + 466.0, + 787.0, + 482.0, + 768.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 465.0, + 834.0, + 465.0, + 834.0, + 480.0, + 817.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 463.0, + 885.0, + 463.0, + 885.0, + 483.0, + 864.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 463.0, + 935.0, + 463.0, + 935.0, + 483.0, + 912.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 463.0, + 986.0, + 463.0, + 986.0, + 483.0, + 960.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 468.0, + 1099.0, + 468.0, + 1099.0, + 480.0, + 1088.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 467.0, + 1151.0, + 467.0, + 1151.0, + 483.0, + 1134.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 467.0, + 1201.0, + 467.0, + 1201.0, + 483.0, + 1184.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 467.0, + 1251.0, + 467.0, + 1251.0, + 483.0, + 1233.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 467.0, + 1300.0, + 467.0, + 1300.0, + 483.0, + 1283.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 466.0, + 1354.0, + 466.0, + 1354.0, + 484.0, + 1329.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 476.0, + 500.0, + 476.0, + 500.0, + 494.0, + 466.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 475.0, + 868.0, + 475.0, + 868.0, + 493.0, + 834.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 476.0, + 1235.0, + 476.0, + 1235.0, + 494.0, + 1201.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 512.0, + 544.0, + 512.0, + 544.0, + 542.0, + 418.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 509.0, + 905.0, + 509.0, + 905.0, + 540.0, + 791.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 512.0, + 1297.0, + 512.0, + 1297.0, + 541.0, + 1135.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 241.0, + 727.0, + 241.0, + 727.0, + 260.0, + 696.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 249.5, + 963.0, + 249.5, + 963.0, + 266.5, + 894.0, + 266.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 356.5, + 734.0, + 356.5, + 734.0, + 389.5, + 663.0, + 389.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 424.5, + 549.0, + 424.5, + 549.0, + 449.0, + 466.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.75, + 405.0, + 728.75, + 405.0, + 728.75, + 423.0, + 682.75, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 433.0, + 358.0, + 433.0, + 358.0, + 452.0, + 327.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 1468.0, + 1169.0, + 1468.0, + 1169.0, + 1495.0, + 1087.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1494.0, + 1370.0, + 1494.0, + 1370.0, + 1513.0, + 1302.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 1508.0, + 1381.0, + 1508.0, + 1381.0, + 1524.0, + 1303.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 1518.0, + 1364.0, + 1518.0, + 1364.0, + 1538.0, + 1301.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 1533.0, + 1322.0, + 1533.0, + 1322.0, + 1549.0, + 1303.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1545.0, + 1323.0, + 1545.0, + 1323.0, + 1561.0, + 1302.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1560.0, + 868.0, + 1560.0, + 868.0, + 1664.0, + 804.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1557.0, + 1336.0, + 1557.0, + 1336.0, + 1595.0, + 1302.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1591.0, + 1337.0, + 1591.0, + 1337.0, + 1608.0, + 1302.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1604.0, + 1327.0, + 1604.0, + 1327.0, + 1618.0, + 1302.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1670.0, + 870.0, + 1670.0, + 870.0, + 1694.0, + 821.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1742.0, + 939.0, + 1742.0, + 939.0, + 1769.0, + 902.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1742.0, + 1062.0, + 1742.0, + 1062.0, + 1769.0, + 1025.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1741.0, + 1187.0, + 1741.0, + 1187.0, + 1770.0, + 1146.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1742.0, + 1308.0, + 1742.0, + 1308.0, + 1769.0, + 1271.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1764.0, + 1237.0, + 1764.0, + 1237.0, + 1789.0, + 1019.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.25, + 1553.5, + 973.25, + 1553.5, + 973.25, + 1565.0, + 955.25, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1622.5, + 1027.0, + 1622.5, + 1027.0, + 1645.5, + 958.0, + 1645.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 567.0, + 1405.0, + 567.0, + 1405.0, + 601.0, + 296.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 598.0, + 1406.0, + 598.0, + 1406.0, + 631.0, + 296.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 629.0, + 558.0, + 629.0, + 558.0, + 663.0, + 296.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1823.0, + 1404.0, + 1823.0, + 1404.0, + 1858.0, + 791.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1853.0, + 1404.0, + 1853.0, + 1404.0, + 1889.0, + 790.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1919.0, + 788.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1913.0, + 941.0, + 1913.0, + 941.0, + 1949.0, + 789.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1302.0, + 1007.0, + 1302.0, + 1007.0, + 1337.0, + 294.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1301.0, + 1006.0, + 1301.0, + 1006.0, + 1338.0, + 295.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1466.0, + 771.0, + 1466.0, + 771.0, + 1503.0, + 295.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1498.0, + 770.0, + 1498.0, + 770.0, + 1532.0, + 295.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1529.0, + 772.0, + 1529.0, + 772.0, + 1564.0, + 296.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1561.0, + 770.0, + 1561.0, + 770.0, + 1593.0, + 296.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1592.0, + 771.0, + 1592.0, + 771.0, + 1623.0, + 296.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1621.0, + 771.0, + 1621.0, + 771.0, + 1653.0, + 296.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 773.0, + 1650.0, + 773.0, + 1684.0, + 294.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1682.0, + 773.0, + 1682.0, + 773.0, + 1714.0, + 295.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 772.0, + 1712.0, + 772.0, + 1748.0, + 295.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1744.0, + 771.0, + 1744.0, + 771.0, + 1775.0, + 295.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1772.0, + 771.0, + 1772.0, + 771.0, + 1807.0, + 295.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1802.0, + 771.0, + 1802.0, + 771.0, + 1837.0, + 295.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1835.0, + 771.0, + 1835.0, + 771.0, + 1867.0, + 297.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1867.0, + 496.0, + 1867.0, + 496.0, + 1900.0, + 296.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 718.0, + 1405.0, + 718.0, + 1405.0, + 761.0, + 294.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 749.0, + 1405.0, + 749.0, + 1405.0, + 791.0, + 291.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 779.0, + 879.0, + 779.0, + 879.0, + 821.0, + 292.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 779.0, + 1408.0, + 779.0, + 1408.0, + 821.0, + 917.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 815.0, + 1401.0, + 815.0, + 1401.0, + 847.0, + 296.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 841.0, + 1197.0, + 841.0, + 1197.0, + 877.0, + 295.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 841.0, + 1404.0, + 841.0, + 1404.0, + 877.0, + 1234.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 873.0, + 736.0, + 873.0, + 736.0, + 908.0, + 295.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1396.0, + 294.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 1405.0, + 1392.0, + 1405.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1422.0, + 987.0, + 1422.0, + 987.0, + 1456.0, + 295.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2010.0, + 295.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1909.0, + 770.0, + 1909.0, + 770.0, + 1945.0, + 297.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 771.0, + 1942.0, + 771.0, + 1977.0, + 296.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 929.0, + 1406.0, + 929.0, + 1406.0, + 974.0, + 293.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 965.0, + 1404.0, + 965.0, + 1404.0, + 1002.0, + 317.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 998.0, + 933.0, + 998.0, + 933.0, + 1029.0, + 319.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1037.0, + 1406.0, + 1037.0, + 1406.0, + 1075.0, + 294.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1069.0, + 1404.0, + 1069.0, + 1404.0, + 1104.0, + 318.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1100.0, + 1404.0, + 1100.0, + 1404.0, + 1134.0, + 318.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1131.0, + 532.0, + 1131.0, + 532.0, + 1165.0, + 317.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1169.0, + 1406.0, + 1169.0, + 1406.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1203.0, + 1406.0, + 1203.0, + 1406.0, + 1236.0, + 318.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1236.0, + 613.0, + 1236.0, + 613.0, + 1267.0, + 320.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1236.0, + 817.0, + 1236.0, + 817.0, + 1267.0, + 649.0, + 1267.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1443, + 1404, + 1443, + 1404, + 1689, + 297, + 1689 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 568, + 1404, + 568, + 1404, + 934, + 298, + 934 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 337, + 1403, + 337, + 1403, + 552, + 298, + 552 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1803, + 1403, + 1803, + 1403, + 1896, + 299, + 1896 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 302, + 229, + 1398, + 229, + 1398, + 322, + 302, + 322 + ], + "score": 0.955 + }, + { + "category_id": 0, + "poly": [ + 298, + 1375, + 928, + 1375, + 928, + 1410, + 298, + 1410 + ], + "score": 0.912 + }, + { + "category_id": 1, + "poly": [ + 295, + 961, + 1407, + 961, + 1407, + 1330, + 295, + 1330 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 301, + 1737, + 607, + 1737, + 607, + 1769, + 301, + 1769 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 837, + 2089, + 865, + 2089, + 865, + 2112, + 837, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 13, + "poly": [ + 942, + 845, + 1009, + 845, + 1009, + 873, + 942, + 873 + ], + "score": 0.9, + "latex": "t _ { t r a i n }" + }, + { + "category_id": 13, + "poly": [ + 972, + 813, + 1235, + 813, + 1235, + 843, + 972, + 843 + ], + "score": 0.9, + "latex": "t _ { s i m } = t _ { s e a r c h } + t _ { t r a i n }" + }, + { + "category_id": 13, + "poly": [ + 911, + 782, + 957, + 782, + 957, + 810, + 911, + 810 + ], + "score": 0.46, + "latex": "2 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 1357, + 721, + 1403, + 721, + 1403, + 750, + 1357, + 750 + ], + "score": 0.44, + "latex": "1 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 568, + 782, + 615, + 782, + 615, + 810, + 568, + 810 + ], + "score": 0.4, + "latex": "2 0 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1372.0, + 931.0, + 1372.0, + 931.0, + 1417.0, + 293.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1734.0, + 612.0, + 1734.0, + 612.0, + 1777.0, + 297.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1446.0, + 1402.0, + 1446.0, + 1402.0, + 1476.0, + 297.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 1404.0, + 1473.0, + 1404.0, + 1511.0, + 294.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1505.0, + 1406.0, + 1505.0, + 1406.0, + 1542.0, + 294.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1571.0, + 294.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1567.0, + 1404.0, + 1567.0, + 1404.0, + 1601.0, + 294.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1595.0, + 1405.0, + 1595.0, + 1405.0, + 1632.0, + 292.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1628.0, + 1404.0, + 1628.0, + 1404.0, + 1662.0, + 294.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1657.0, + 1220.0, + 1657.0, + 1220.0, + 1691.0, + 295.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 569.0, + 1406.0, + 569.0, + 1406.0, + 603.0, + 296.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 598.0, + 1405.0, + 598.0, + 1405.0, + 632.0, + 293.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 630.0, + 1404.0, + 630.0, + 1404.0, + 664.0, + 294.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 694.0, + 294.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 690.0, + 1405.0, + 690.0, + 1405.0, + 724.0, + 294.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 721.0, + 1356.0, + 721.0, + 1356.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 751.0, + 1404.0, + 751.0, + 1404.0, + 785.0, + 293.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 779.0, + 567.0, + 779.0, + 567.0, + 818.0, + 293.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 779.0, + 910.0, + 779.0, + 910.0, + 818.0, + 616.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 779.0, + 1403.0, + 779.0, + 1403.0, + 818.0, + 958.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 808.0, + 971.0, + 808.0, + 971.0, + 851.0, + 292.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 808.0, + 1405.0, + 808.0, + 1405.0, + 851.0, + 1236.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 842.0, + 941.0, + 842.0, + 941.0, + 880.0, + 292.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 842.0, + 1405.0, + 842.0, + 1405.0, + 880.0, + 1010.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 871.0, + 1405.0, + 871.0, + 1405.0, + 909.0, + 293.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 904.0, + 848.0, + 904.0, + 848.0, + 938.0, + 294.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 339.0, + 1404.0, + 339.0, + 1404.0, + 373.0, + 294.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 367.0, + 1405.0, + 367.0, + 1405.0, + 404.0, + 293.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 1405.0, + 400.0, + 1405.0, + 435.0, + 294.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 429.0, + 1407.0, + 429.0, + 1407.0, + 467.0, + 293.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 460.0, + 1405.0, + 460.0, + 1405.0, + 495.0, + 294.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 523.0, + 293.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 522.0, + 602.0, + 522.0, + 602.0, + 558.0, + 294.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1404.0, + 1802.0, + 1404.0, + 1840.0, + 294.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1833.0, + 1403.0, + 1833.0, + 1403.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1866.0, + 821.0, + 1866.0, + 821.0, + 1898.0, + 294.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 462.0, + 290.0, + 462.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 961.0, + 1404.0, + 961.0, + 1404.0, + 995.0, + 294.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 992.0, + 975.0, + 992.0, + 975.0, + 1026.0, + 317.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1071.0, + 291.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1064.0, + 1404.0, + 1064.0, + 1404.0, + 1098.0, + 317.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 1093.0, + 978.0, + 1093.0, + 978.0, + 1130.0, + 314.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1131.0, + 1404.0, + 1131.0, + 1404.0, + 1172.0, + 296.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1166.0, + 1403.0, + 1166.0, + 1403.0, + 1200.0, + 317.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1199.0, + 1404.0, + 1199.0, + 1404.0, + 1230.0, + 318.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1227.0, + 997.0, + 1227.0, + 997.0, + 1260.0, + 316.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1265.0, + 1404.0, + 1265.0, + 1404.0, + 1307.0, + 291.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1297.0, + 1403.0, + 1297.0, + 1403.0, + 1334.0, + 317.0, + 1334.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 296, + 279, + 1400, + 279, + 1400, + 373, + 296, + 373 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 296, + 1589, + 1402, + 1589, + 1402, + 1654, + 296, + 1654 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 300, + 395, + 1399, + 395, + 1399, + 491, + 300, + 491 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 296, + 686, + 1403, + 686, + 1403, + 750, + 296, + 750 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1401, + 1942, + 1401, + 2035, + 299, + 2035 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 304, + 1355, + 1401, + 1355, + 1401, + 1452, + 304, + 1452 + ], + "score": 0.824 + }, + { + "category_id": 1, + "poly": [ + 299, + 1091, + 1399, + 1091, + 1399, + 1157, + 299, + 1157 + ], + "score": 0.823 + }, + { + "category_id": 1, + "poly": [ + 298, + 1472, + 1406, + 1472, + 1406, + 1568, + 298, + 1568 + ], + "score": 0.816 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.813 + }, + { + "category_id": 1, + "poly": [ + 300, + 888, + 1399, + 888, + 1399, + 985, + 300, + 985 + ], + "score": 0.808 + }, + { + "category_id": 1, + "poly": [ + 297, + 1005, + 1399, + 1005, + 1399, + 1070, + 297, + 1070 + ], + "score": 0.796 + }, + { + "category_id": 1, + "poly": [ + 297, + 1178, + 1403, + 1178, + 1403, + 1335, + 297, + 1335 + ], + "score": 0.788 + }, + { + "category_id": 1, + "poly": [ + 298, + 1793, + 1404, + 1793, + 1404, + 1920, + 298, + 1920 + ], + "score": 0.785 + }, + { + "category_id": 1, + "poly": [ + 297, + 1675, + 1403, + 1675, + 1403, + 1771, + 297, + 1771 + ], + "score": 0.785 + }, + { + "category_id": 1, + "poly": [ + 297, + 598, + 1398, + 598, + 1398, + 664, + 297, + 664 + ], + "score": 0.778 + }, + { + "category_id": 1, + "poly": [ + 297, + 771, + 1406, + 771, + 1406, + 866, + 297, + 866 + ], + "score": 0.767 + }, + { + "category_id": 1, + "poly": [ + 296, + 512, + 1397, + 512, + 1397, + 577, + 296, + 577 + ], + "score": 0.737 + }, + { + "category_id": 0, + "poly": [ + 299, + 227, + 489, + 227, + 489, + 261, + 299, + 261 + ], + "score": 0.716 + }, + { + "category_id": 1, + "poly": [ + 299, + 227, + 489, + 227, + 489, + 261, + 299, + 261 + ], + "score": 0.164 + }, + { + "category_id": 13, + "poly": [ + 904, + 463, + 922, + 463, + 922, + 485, + 904, + 485 + ], + "score": 0.64, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 742, + 956, + 760, + 956, + 760, + 978, + 742, + 978 + ], + "score": 0.56, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 228.0, + 491.0, + 228.0, + 491.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 277.0, + 1402.0, + 277.0, + 1402.0, + 314.0, + 295.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 311.0, + 1404.0, + 311.0, + 1404.0, + 345.0, + 323.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 340.0, + 446.0, + 340.0, + 446.0, + 377.0, + 318.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1585.0, + 1403.0, + 1585.0, + 1403.0, + 1629.0, + 293.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1620.0, + 524.0, + 1620.0, + 524.0, + 1654.0, + 324.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 393.0, + 1405.0, + 393.0, + 1405.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 430.0, + 1403.0, + 430.0, + 1403.0, + 461.0, + 321.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 458.0, + 903.0, + 458.0, + 903.0, + 494.0, + 325.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 458.0, + 1083.0, + 458.0, + 1083.0, + 494.0, + 923.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 686.0, + 1403.0, + 686.0, + 1403.0, + 722.0, + 296.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 718.0, + 917.0, + 718.0, + 917.0, + 749.0, + 324.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1405.0, + 1939.0, + 1405.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2008.0, + 323.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2003.0, + 1178.0, + 2003.0, + 1178.0, + 2037.0, + 324.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1353.0, + 1405.0, + 1353.0, + 1405.0, + 1394.0, + 294.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1423.0, + 321.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1417.0, + 1032.0, + 1417.0, + 1032.0, + 1452.0, + 321.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1090.0, + 1405.0, + 1090.0, + 1405.0, + 1129.0, + 293.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1123.0, + 1015.0, + 1123.0, + 1015.0, + 1158.0, + 323.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1471.0, + 1402.0, + 1471.0, + 1402.0, + 1509.0, + 295.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1505.0, + 1404.0, + 1505.0, + 1404.0, + 1540.0, + 322.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1535.0, + 633.0, + 1535.0, + 633.0, + 1569.0, + 324.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 890.0, + 1403.0, + 890.0, + 1403.0, + 924.0, + 296.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 917.0, + 1403.0, + 917.0, + 1403.0, + 957.0, + 320.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 951.0, + 741.0, + 951.0, + 741.0, + 986.0, + 326.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 951.0, + 938.0, + 951.0, + 938.0, + 986.0, + 761.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1003.0, + 1404.0, + 1003.0, + 1404.0, + 1044.0, + 295.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1036.0, + 1404.0, + 1036.0, + 1404.0, + 1072.0, + 323.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1176.0, + 1403.0, + 1176.0, + 1403.0, + 1218.0, + 294.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1211.0, + 1404.0, + 1211.0, + 1404.0, + 1244.0, + 324.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1242.0, + 1404.0, + 1242.0, + 1404.0, + 1275.0, + 322.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1271.0, + 1405.0, + 1271.0, + 1405.0, + 1305.0, + 324.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1301.0, + 1136.0, + 1301.0, + 1136.0, + 1338.0, + 322.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1794.0, + 1405.0, + 1794.0, + 1405.0, + 1827.0, + 296.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1826.0, + 1402.0, + 1826.0, + 1402.0, + 1859.0, + 323.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1857.0, + 1402.0, + 1857.0, + 1402.0, + 1890.0, + 323.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1887.0, + 772.0, + 1887.0, + 772.0, + 1920.0, + 321.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1674.0, + 1403.0, + 1674.0, + 1403.0, + 1713.0, + 292.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1706.0, + 1404.0, + 1706.0, + 1404.0, + 1746.0, + 321.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1738.0, + 587.0, + 1738.0, + 587.0, + 1772.0, + 322.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 597.0, + 1402.0, + 597.0, + 1402.0, + 635.0, + 295.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 631.0, + 775.0, + 631.0, + 775.0, + 663.0, + 323.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 771.0, + 1402.0, + 771.0, + 1402.0, + 809.0, + 295.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 805.0, + 1404.0, + 805.0, + 1404.0, + 839.0, + 323.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 830.0, + 465.0, + 830.0, + 465.0, + 869.0, + 322.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 515.0, + 1402.0, + 515.0, + 1402.0, + 547.0, + 296.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 543.0, + 1212.0, + 543.0, + 1212.0, + 578.0, + 322.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 228.0, + 491.0, + 228.0, + 491.0, + 264.0, + 297.0, + 264.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.89 + }, + { + "category_id": 1, + "poly": [ + 295, + 1103, + 1402, + 1103, + 1402, + 1169, + 295, + 1169 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 297, + 1024, + 1402, + 1024, + 1402, + 1087, + 297, + 1087 + ], + "score": 0.843 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 298, + 1325, + 1399, + 1325, + 1399, + 1421, + 298, + 1421 + ], + "score": 0.825 + }, + { + "category_id": 1, + "poly": [ + 296, + 912, + 1404, + 912, + 1404, + 1007, + 296, + 1007 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 302, + 802, + 1403, + 802, + 1403, + 897, + 302, + 897 + ], + "score": 0.809 + }, + { + "category_id": 1, + "poly": [ + 300, + 1184, + 1405, + 1184, + 1405, + 1308, + 300, + 1308 + ], + "score": 0.806 + }, + { + "category_id": 1, + "poly": [ + 292, + 721, + 1400, + 721, + 1400, + 786, + 292, + 786 + ], + "score": 0.804 + }, + { + "category_id": 1, + "poly": [ + 298, + 1436, + 1405, + 1436, + 1405, + 1623, + 298, + 1623 + ], + "score": 0.788 + }, + { + "category_id": 1, + "poly": [ + 298, + 1749, + 1403, + 1749, + 1403, + 1874, + 298, + 1874 + ], + "score": 0.787 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1403, + 229, + 1403, + 293, + 294, + 293 + ], + "score": 0.783 + }, + { + "category_id": 1, + "poly": [ + 295, + 439, + 1401, + 439, + 1401, + 504, + 295, + 504 + ], + "score": 0.774 + }, + { + "category_id": 1, + "poly": [ + 295, + 389, + 1398, + 389, + 1398, + 424, + 295, + 424 + ], + "score": 0.754 + }, + { + "category_id": 1, + "poly": [ + 296, + 309, + 1400, + 309, + 1400, + 374, + 296, + 374 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1399, + 1971, + 1399, + 2034, + 299, + 2034 + ], + "score": 0.744 + }, + { + "category_id": 1, + "poly": [ + 298, + 1639, + 1403, + 1639, + 1403, + 1734, + 298, + 1734 + ], + "score": 0.729 + }, + { + "category_id": 1, + "poly": [ + 297, + 1891, + 1398, + 1891, + 1398, + 1955, + 297, + 1955 + ], + "score": 0.727 + }, + { + "category_id": 1, + "poly": [ + 296, + 519, + 1404, + 519, + 1404, + 705, + 296, + 705 + ], + "score": 0.629 + }, + { + "category_id": 13, + "poly": [ + 525, + 1392, + 544, + 1392, + 544, + 1415, + 525, + 1415 + ], + "score": 0.56, + "latex": "{ . } = { }" + }, + { + "category_id": 13, + "poly": [ + 1102, + 1705, + 1119, + 1705, + 1119, + 1728, + 1102, + 1728 + ], + "score": 0.5, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } }" + }, + { + "category_id": 13, + "poly": [ + 1382, + 1815, + 1402, + 1815, + 1402, + 1838, + 1382, + 1838 + ], + "score": 0.37, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =" + }, + { + "category_id": 13, + "poly": [ + 360, + 1281, + 378, + 1281, + 378, + 1304, + 360, + 1304 + ], + "score": 0.29, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1104.0, + 1402.0, + 1104.0, + 1402.0, + 1140.0, + 297.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1135.0, + 1045.0, + 1135.0, + 1045.0, + 1170.0, + 322.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1020.0, + 1406.0, + 1020.0, + 1406.0, + 1063.0, + 292.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1056.0, + 927.0, + 1056.0, + 927.0, + 1087.0, + 324.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1324.0, + 1405.0, + 1324.0, + 1405.0, + 1363.0, + 293.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1356.0, + 1403.0, + 1356.0, + 1403.0, + 1395.0, + 321.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1390.0, + 524.0, + 1390.0, + 524.0, + 1424.0, + 324.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1390.0, + 722.0, + 1390.0, + 722.0, + 1424.0, + 545.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 911.0, + 1405.0, + 911.0, + 1405.0, + 950.0, + 293.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 946.0, + 1404.0, + 946.0, + 1404.0, + 980.0, + 323.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 975.0, + 532.0, + 975.0, + 532.0, + 1007.0, + 324.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 801.0, + 1405.0, + 801.0, + 1405.0, + 839.0, + 296.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 832.0, + 1403.0, + 832.0, + 1403.0, + 869.0, + 321.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 864.0, + 886.0, + 864.0, + 886.0, + 898.0, + 323.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1182.0, + 1406.0, + 1182.0, + 1406.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1216.0, + 1405.0, + 1216.0, + 1405.0, + 1250.0, + 320.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1245.0, + 1405.0, + 1245.0, + 1405.0, + 1282.0, + 323.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1277.0, + 359.0, + 1277.0, + 359.0, + 1310.0, + 324.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1277.0, + 555.0, + 1277.0, + 555.0, + 1310.0, + 379.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 720.0, + 1403.0, + 720.0, + 1403.0, + 759.0, + 294.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 752.0, + 776.0, + 752.0, + 776.0, + 784.0, + 325.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1439.0, + 1403.0, + 1439.0, + 1403.0, + 1471.0, + 297.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1471.0, + 1402.0, + 1471.0, + 1402.0, + 1499.0, + 326.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1494.0, + 1406.0, + 1494.0, + 1406.0, + 1537.0, + 320.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1527.0, + 1405.0, + 1527.0, + 1405.0, + 1567.0, + 322.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1559.0, + 1403.0, + 1559.0, + 1403.0, + 1595.0, + 321.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1592.0, + 539.0, + 1592.0, + 539.0, + 1622.0, + 321.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1749.0, + 1402.0, + 1749.0, + 1402.0, + 1782.0, + 296.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1782.0, + 1404.0, + 1782.0, + 1404.0, + 1814.0, + 322.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1810.0, + 1381.0, + 1810.0, + 1381.0, + 1847.0, + 321.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1810.0, + 1407.0, + 1810.0, + 1407.0, + 1847.0, + 1403.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1843.0, + 507.0, + 1843.0, + 507.0, + 1877.0, + 323.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 261.0, + 862.0, + 261.0, + 862.0, + 294.0, + 321.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 439.0, + 1405.0, + 439.0, + 1405.0, + 477.0, + 295.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 472.0, + 1298.0, + 472.0, + 1298.0, + 505.0, + 322.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 388.0, + 1403.0, + 388.0, + 1403.0, + 426.0, + 294.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 309.0, + 1405.0, + 309.0, + 1405.0, + 345.0, + 295.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 340.0, + 1015.0, + 340.0, + 1015.0, + 375.0, + 322.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 1403.0, + 2003.0, + 1403.0, + 2038.0, + 323.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1635.0, + 1405.0, + 1635.0, + 1405.0, + 1678.0, + 293.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1669.0, + 1403.0, + 1669.0, + 1403.0, + 1706.0, + 323.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1702.0, + 1101.0, + 1702.0, + 1101.0, + 1735.0, + 322.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1702.0, + 1299.0, + 1702.0, + 1299.0, + 1735.0, + 1120.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1890.0, + 1403.0, + 1890.0, + 1403.0, + 1928.0, + 295.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1922.0, + 1284.0, + 1922.0, + 1284.0, + 1956.0, + 322.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 517.0, + 1405.0, + 517.0, + 1405.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 551.0, + 1405.0, + 551.0, + 1405.0, + 586.0, + 323.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 577.0, + 1406.0, + 577.0, + 1406.0, + 621.0, + 322.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 612.0, + 1404.0, + 612.0, + 1404.0, + 648.0, + 323.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 643.0, + 1402.0, + 643.0, + 1402.0, + 678.0, + 323.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 673.0, + 705.0, + 673.0, + 705.0, + 706.0, + 319.0, + 706.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 857, + 1403, + 857, + 1403, + 981, + 299, + 981 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1242, + 1404, + 1242, + 1404, + 1488, + 297, + 1488 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 296, + 1403, + 296, + 1403, + 450, + 297, + 450 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1104, + 1403, + 1104, + 1403, + 1227, + 298, + 1227 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 647, + 1403, + 647, + 1403, + 740, + 299, + 740 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 487, + 1403, + 487, + 1403, + 611, + 298, + 611 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1503, + 1403, + 1503, + 1403, + 1596, + 298, + 1596 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 301, + 996, + 1398, + 996, + 1398, + 1089, + 301, + 1089 + ], + "score": 0.961 + }, + { + "category_id": 3, + "poly": [ + 305, + 1650, + 1394, + 1650, + 1394, + 1970, + 305, + 1970 + ], + "score": 0.96 + }, + { + "category_id": 4, + "poly": [ + 451, + 1994, + 1247, + 1994, + 1247, + 2026, + 451, + 2026 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 300, + 787, + 535, + 787, + 535, + 823, + 300, + 823 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.865 + }, + { + "category_id": 0, + "poly": [ + 304, + 227, + 767, + 227, + 767, + 261, + 304, + 261 + ], + "score": 0.848 + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1654.0, + 581.0, + 1654.0, + 581.0, + 1674.0, + 402.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1654.0, + 944.0, + 1654.0, + 944.0, + 1674.0, + 767.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1654.0, + 1313.0, + 1654.0, + 1313.0, + 1674.0, + 1134.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 1688.0, + 1056.0, + 1688.0, + 1056.0, + 1699.0, + 1044.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1707.0, + 690.0, + 1707.0, + 690.0, + 1722.0, + 675.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1718.0, + 327.0, + 1718.0, + 327.0, + 1732.0, + 311.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1748.0, + 1058.0, + 1748.0, + 1058.0, + 1762.0, + 1043.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1786.0, + 691.0, + 1786.0, + 691.0, + 1802.0, + 675.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1798.0, + 328.0, + 1798.0, + 328.0, + 1814.0, + 311.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1812.0, + 1055.0, + 1812.0, + 1055.0, + 1822.0, + 1045.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1865.0, + 691.0, + 1865.0, + 691.0, + 1883.0, + 670.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1871.0, + 1059.0, + 1871.0, + 1059.0, + 1885.0, + 1043.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1889.0, + 421.0, + 1889.0, + 421.0, + 1908.0, + 397.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1891.0, + 658.0, + 1891.0, + 658.0, + 1907.0, + 640.0, + 1907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1889.0, + 779.0, + 1889.0, + 779.0, + 1908.0, + 757.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1891.0, + 1022.0, + 1891.0, + 1022.0, + 1907.0, + 1004.0, + 1907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1889.0, + 1149.0, + 1889.0, + 1149.0, + 1908.0, + 1127.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1375.0, + 1892.0, + 1391.0, + 1892.0, + 1391.0, + 1905.0, + 1375.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1903.0, + 514.0, + 1903.0, + 514.0, + 1920.0, + 472.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1903.0, + 877.0, + 1903.0, + 877.0, + 1920.0, + 833.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1903.0, + 1246.0, + 1903.0, + 1246.0, + 1920.0, + 1204.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1933.0, + 582.0, + 1933.0, + 582.0, + 1975.0, + 385.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1933.0, + 946.0, + 1933.0, + 946.0, + 1975.0, + 748.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1933.0, + 1315.0, + 1933.0, + 1315.0, + 1975.0, + 1118.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.25, + 1879.5, + 329.25, + 1879.5, + 329.25, + 1893.0, + 303.25, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1990.0, + 1247.0, + 1990.0, + 1247.0, + 2030.0, + 450.0, + 2030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 783.0, + 541.0, + 783.0, + 541.0, + 829.0, + 291.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 224.0, + 772.0, + 224.0, + 772.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 856.0, + 1404.0, + 856.0, + 1404.0, + 895.0, + 294.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 891.0, + 1403.0, + 891.0, + 1403.0, + 924.0, + 295.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 920.0, + 1404.0, + 920.0, + 1404.0, + 956.0, + 294.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 951.0, + 1196.0, + 951.0, + 1196.0, + 985.0, + 294.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1241.0, + 1405.0, + 1241.0, + 1405.0, + 1279.0, + 292.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1272.0, + 1406.0, + 1272.0, + 1406.0, + 1307.0, + 294.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1304.0, + 1405.0, + 1304.0, + 1405.0, + 1339.0, + 292.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1334.0, + 1405.0, + 1334.0, + 1405.0, + 1369.0, + 294.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1364.0, + 1406.0, + 1364.0, + 1406.0, + 1405.0, + 292.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1396.0, + 1405.0, + 1396.0, + 1405.0, + 1430.0, + 294.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1424.0, + 1405.0, + 1424.0, + 1405.0, + 1464.0, + 294.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1453.0, + 420.0, + 1453.0, + 420.0, + 1498.0, + 292.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 298.0, + 1404.0, + 298.0, + 1404.0, + 331.0, + 296.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 328.0, + 1405.0, + 328.0, + 1405.0, + 361.0, + 296.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 358.0, + 1405.0, + 358.0, + 1405.0, + 392.0, + 296.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 388.0, + 1407.0, + 388.0, + 1407.0, + 425.0, + 295.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 419.0, + 864.0, + 419.0, + 864.0, + 452.0, + 295.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1106.0, + 1402.0, + 1106.0, + 1402.0, + 1139.0, + 296.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1135.0, + 1403.0, + 1135.0, + 1403.0, + 1168.0, + 294.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1167.0, + 1407.0, + 1167.0, + 1407.0, + 1203.0, + 294.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1196.0, + 1253.0, + 1196.0, + 1253.0, + 1232.0, + 293.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 644.0, + 1405.0, + 644.0, + 1405.0, + 684.0, + 293.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 679.0, + 1403.0, + 679.0, + 1403.0, + 713.0, + 295.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 707.0, + 839.0, + 707.0, + 839.0, + 743.0, + 294.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 487.0, + 1404.0, + 487.0, + 1404.0, + 523.0, + 294.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 517.0, + 1403.0, + 517.0, + 1403.0, + 553.0, + 293.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 548.0, + 1405.0, + 548.0, + 1405.0, + 583.0, + 294.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 578.0, + 682.0, + 578.0, + 682.0, + 613.0, + 295.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1500.0, + 1405.0, + 1500.0, + 1405.0, + 1540.0, + 293.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1535.0, + 1405.0, + 1535.0, + 1405.0, + 1569.0, + 294.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1565.0, + 1405.0, + 1565.0, + 1405.0, + 1599.0, + 296.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 997.0, + 1403.0, + 997.0, + 1403.0, + 1030.0, + 296.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1027.0, + 1402.0, + 1027.0, + 1402.0, + 1061.0, + 296.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1059.0, + 663.0, + 1059.0, + 663.0, + 1091.0, + 296.0, + 1091.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 739, + 1403, + 739, + 1403, + 894, + 298, + 894 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 428, + 1404, + 428, + 1404, + 583, + 298, + 583 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 413, + 298, + 413 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 306, + 1084, + 1391, + 1084, + 1391, + 1400, + 306, + 1400 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 304, + 1583, + 1393, + 1583, + 1393, + 1898, + 304, + 1898 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 292, + 1922, + 1407, + 1922, + 1407, + 1986, + 292, + 1986 + ], + "score": 0.938 + }, + { + "category_id": 4, + "poly": [ + 292, + 1422, + 1401, + 1422, + 1401, + 1486, + 292, + 1486 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 300, + 654, + 647, + 654, + 647, + 689, + 300, + 689 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.853 + }, + { + "category_id": 0, + "poly": [ + 298, + 967, + 1171, + 967, + 1171, + 1000, + 298, + 1000 + ], + "score": 0.352 + }, + { + "category_id": 4, + "poly": [ + 298, + 967, + 1171, + 967, + 1171, + 1000, + 298, + 1000 + ], + "score": 0.24 + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1088.0, + 352.0, + 1088.0, + 352.0, + 1108.0, + 318.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1095.0, + 467.0, + 1095.0, + 467.0, + 1125.0, + 370.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1094.0, + 644.0, + 1094.0, + 644.0, + 1112.0, + 620.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1090.0, + 710.0, + 1090.0, + 710.0, + 1107.0, + 682.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1095.0, + 982.0, + 1095.0, + 982.0, + 1125.0, + 883.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1094.0, + 1006.0, + 1094.0, + 1006.0, + 1112.0, + 983.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1098.0, + 1079.0, + 1098.0, + 1079.0, + 1116.0, + 1052.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1095.0, + 1352.0, + 1095.0, + 1352.0, + 1125.0, + 1254.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1094.0, + 1376.0, + 1094.0, + 1376.0, + 1112.0, + 1353.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1118.0, + 404.0, + 1118.0, + 404.0, + 1133.0, + 372.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1118.0, + 919.0, + 1118.0, + 919.0, + 1133.0, + 886.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1117.0, + 1294.0, + 1117.0, + 1294.0, + 1137.0, + 1253.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1130.0, + 420.0, + 1130.0, + 420.0, + 1145.0, + 372.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1127.0, + 935.0, + 1127.0, + 935.0, + 1146.0, + 884.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1127.0, + 1305.0, + 1127.0, + 1305.0, + 1146.0, + 1253.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1140.0, + 403.0, + 1140.0, + 403.0, + 1157.0, + 372.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1140.0, + 915.0, + 1140.0, + 915.0, + 1157.0, + 884.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1140.0, + 1285.0, + 1140.0, + 1285.0, + 1157.0, + 1253.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1165.0, + 324.0, + 1165.0, + 324.0, + 1250.0, + 308.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1153.0, + 428.0, + 1153.0, + 428.0, + 1169.0, + 373.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1153.0, + 661.0, + 1153.0, + 661.0, + 1261.0, + 641.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1164.0, + 689.0, + 1164.0, + 689.0, + 1251.0, + 670.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1150.0, + 941.0, + 1150.0, + 941.0, + 1169.0, + 884.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1153.0, + 1024.0, + 1153.0, + 1024.0, + 1262.0, + 1004.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1163.0, + 1061.0, + 1163.0, + 1061.0, + 1252.0, + 1040.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 1153.0, + 1310.0, + 1153.0, + 1310.0, + 1169.0, + 1255.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1153.0, + 1393.0, + 1153.0, + 1393.0, + 1261.0, + 1373.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1281.0, + 708.0, + 1281.0, + 708.0, + 1298.0, + 682.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1295.0, + 1079.0, + 1295.0, + 1079.0, + 1308.0, + 1053.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1302.0, + 344.0, + 1302.0, + 344.0, + 1316.0, + 317.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1309.0, + 647.0, + 1309.0, + 647.0, + 1327.0, + 620.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1309.0, + 1010.0, + 1309.0, + 1010.0, + 1326.0, + 983.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1307.0, + 1381.0, + 1307.0, + 1381.0, + 1324.0, + 1352.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1325.0, + 352.0, + 1325.0, + 352.0, + 1335.0, + 342.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1321.0, + 413.0, + 1321.0, + 413.0, + 1339.0, + 392.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1322.0, + 466.0, + 1322.0, + 466.0, + 1338.0, + 450.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1324.0, + 521.0, + 1324.0, + 521.0, + 1338.0, + 504.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1324.0, + 576.0, + 1324.0, + 576.0, + 1338.0, + 558.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1325.0, + 715.0, + 1325.0, + 715.0, + 1335.0, + 707.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1324.0, + 774.0, + 1324.0, + 774.0, + 1338.0, + 757.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1324.0, + 830.0, + 1324.0, + 830.0, + 1338.0, + 813.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1324.0, + 884.0, + 1324.0, + 884.0, + 1338.0, + 867.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1324.0, + 940.0, + 1324.0, + 940.0, + 1338.0, + 921.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1324.0, + 1145.0, + 1324.0, + 1145.0, + 1338.0, + 1126.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1324.0, + 1200.0, + 1324.0, + 1200.0, + 1338.0, + 1182.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1324.0, + 1255.0, + 1324.0, + 1255.0, + 1338.0, + 1237.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1324.0, + 1310.0, + 1324.0, + 1310.0, + 1338.0, + 1292.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1333.0, + 500.0, + 1333.0, + 500.0, + 1350.0, + 467.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1333.0, + 863.0, + 1333.0, + 863.0, + 1350.0, + 830.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1333.0, + 1234.0, + 1333.0, + 1234.0, + 1350.0, + 1200.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1364.0, + 580.0, + 1364.0, + 580.0, + 1401.0, + 386.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1365.0, + 944.0, + 1365.0, + 944.0, + 1400.0, + 748.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1364.0, + 1314.0, + 1364.0, + 1314.0, + 1401.0, + 1117.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.75, + 1125.5, + 1259.75, + 1125.5, + 1259.75, + 1138.5, + 1233.75, + 1138.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1617.0, + 375.0, + 1617.0, + 375.0, + 1638.0, + 361.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1594.0, + 649.0, + 1594.0, + 649.0, + 1625.0, + 549.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1611.0, + 723.0, + 1611.0, + 723.0, + 1633.0, + 679.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1594.0, + 1010.0, + 1594.0, + 1010.0, + 1622.0, + 914.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1586.0, + 1094.0, + 1586.0, + 1094.0, + 1608.0, + 1049.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1597.0, + 1381.0, + 1597.0, + 1381.0, + 1625.0, + 1283.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1629.0, + 358.0, + 1629.0, + 358.0, + 1649.0, + 315.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1633.0, + 1094.0, + 1633.0, + 1094.0, + 1655.0, + 1050.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1643.0, + 725.0, + 1643.0, + 725.0, + 1665.0, + 679.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1772.0, + 721.0, + 1772.0, + 721.0, + 1791.0, + 695.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1803.0, + 358.0, + 1803.0, + 358.0, + 1822.0, + 333.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1798.0, + 1091.0, + 1798.0, + 1091.0, + 1817.0, + 1065.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1825.0, + 365.0, + 1825.0, + 365.0, + 1835.0, + 356.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1822.0, + 426.0, + 1822.0, + 426.0, + 1838.0, + 411.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1822.0, + 486.0, + 1822.0, + 486.0, + 1838.0, + 468.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1822.0, + 544.0, + 1822.0, + 544.0, + 1838.0, + 525.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1822.0, + 602.0, + 1822.0, + 602.0, + 1838.0, + 582.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1820.0, + 660.0, + 1820.0, + 660.0, + 1839.0, + 639.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1822.0, + 731.0, + 1822.0, + 731.0, + 1838.0, + 716.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 1823.0, + 787.0, + 1823.0, + 787.0, + 1836.0, + 775.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1822.0, + 850.0, + 1822.0, + 850.0, + 1838.0, + 830.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1822.0, + 907.0, + 1822.0, + 907.0, + 1838.0, + 889.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1822.0, + 965.0, + 1822.0, + 965.0, + 1838.0, + 945.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1822.0, + 1022.0, + 1822.0, + 1022.0, + 1838.0, + 1003.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1825.0, + 1099.0, + 1825.0, + 1099.0, + 1835.0, + 1089.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1823.0, + 1157.0, + 1823.0, + 1157.0, + 1836.0, + 1146.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1822.0, + 1219.0, + 1822.0, + 1219.0, + 1838.0, + 1200.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1822.0, + 1277.0, + 1822.0, + 1277.0, + 1838.0, + 1257.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1822.0, + 1335.0, + 1822.0, + 1335.0, + 1838.0, + 1315.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 1822.0, + 1391.0, + 1822.0, + 1391.0, + 1838.0, + 1374.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1831.0, + 523.0, + 1831.0, + 523.0, + 1849.0, + 488.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1831.0, + 886.0, + 1831.0, + 886.0, + 1849.0, + 851.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 1831.0, + 1255.0, + 1831.0, + 1255.0, + 1849.0, + 1220.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1863.0, + 581.0, + 1863.0, + 581.0, + 1901.0, + 384.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1864.0, + 945.0, + 1864.0, + 945.0, + 1901.0, + 749.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1863.0, + 1314.0, + 1863.0, + 1314.0, + 1901.0, + 1118.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1922.0, + 1405.0, + 1922.0, + 1405.0, + 1958.0, + 295.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1953.0, + 689.0, + 1953.0, + 689.0, + 1989.0, + 295.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1422.0, + 1404.0, + 1422.0, + 1404.0, + 1458.0, + 295.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1453.0, + 962.0, + 1453.0, + 962.0, + 1489.0, + 295.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 654.0, + 651.0, + 654.0, + 651.0, + 694.0, + 295.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 964.0, + 1174.0, + 964.0, + 1174.0, + 1006.0, + 295.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 964.0, + 1174.0, + 964.0, + 1174.0, + 1006.0, + 295.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 739.0, + 1405.0, + 739.0, + 1405.0, + 777.0, + 293.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 771.0, + 1404.0, + 771.0, + 1404.0, + 805.0, + 294.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 799.0, + 1404.0, + 799.0, + 1404.0, + 838.0, + 293.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 828.0, + 1404.0, + 828.0, + 1404.0, + 868.0, + 293.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 863.0, + 865.0, + 863.0, + 865.0, + 893.0, + 293.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 430.0, + 1402.0, + 430.0, + 1402.0, + 463.0, + 296.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 462.0, + 1404.0, + 462.0, + 1404.0, + 495.0, + 296.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 490.0, + 1406.0, + 490.0, + 1406.0, + 527.0, + 293.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 520.0, + 1405.0, + 520.0, + 1405.0, + 557.0, + 293.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 549.0, + 914.0, + 549.0, + 914.0, + 589.0, + 293.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1403.0, + 261.0, + 1403.0, + 293.0, + 296.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 289.0, + 1404.0, + 289.0, + 1404.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 322.0, + 1405.0, + 322.0, + 1405.0, + 354.0, + 296.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 353.0, + 1404.0, + 353.0, + 1404.0, + 384.0, + 296.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 383.0, + 566.0, + 383.0, + 566.0, + 415.0, + 296.0, + 415.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 352, + 1405, + 352, + 1405, + 659, + 297, + 659 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1577, + 1404, + 1577, + 1404, + 1732, + 298, + 1732 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 490, + 694, + 1208, + 694, + 1208, + 1004, + 490, + 1004 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 489, + 1143, + 1208, + 1143, + 1208, + 1456, + 489, + 1456 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 299, + 1825, + 1400, + 1825, + 1400, + 1919, + 299, + 1919 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 298, + 1933, + 1401, + 1933, + 1401, + 2027, + 298, + 2027 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 294, + 1481, + 1402, + 1481, + 1402, + 1543, + 294, + 1543 + ], + "score": 0.933 + }, + { + "category_id": 4, + "poly": [ + 286, + 1029, + 1403, + 1029, + 1403, + 1092, + 286, + 1092 + ], + "score": 0.931 + }, + { + "category_id": 0, + "poly": [ + 300, + 1768, + 619, + 1768, + 619, + 1801, + 300, + 1801 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 299, + 226, + 613, + 226, + 613, + 262, + 299, + 262 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 0, + "poly": [ + 298, + 295, + 472, + 295, + 472, + 327, + 298, + 327 + ], + "score": 0.843 + }, + { + "category_id": 13, + "poly": [ + 626, + 1936, + 660, + 1936, + 660, + 1965, + 626, + 1965 + ], + "score": 0.89, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1327, + 1856, + 1400, + 1856, + 1400, + 1886, + 1327, + 1886 + ], + "score": 0.88, + "latex": "8 1 e ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 460, + 1828, + 494, + 1828, + 494, + 1858, + 460, + 1858 + ], + "score": 0.88, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 366, + 1770, + 401, + 1770, + 401, + 1801, + 366, + 1801 + ], + "score": 0.85, + "latex": "L _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 704.0, + 789.0, + 704.0, + 789.0, + 720.0, + 714.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 694.0, + 832.0, + 694.0, + 832.0, + 711.0, + 790.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 698.0, + 896.0, + 698.0, + 896.0, + 715.0, + 867.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 695.0, + 1194.0, + 695.0, + 1194.0, + 719.0, + 1076.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 713.0, + 1148.0, + 713.0, + 1148.0, + 732.0, + 1076.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 743.0, + 545.0, + 743.0, + 545.0, + 762.0, + 503.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 755.0, + 907.0, + 755.0, + 907.0, + 767.0, + 900.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 769.0, + 546.0, + 769.0, + 546.0, + 859.0, + 490.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 761.0, + 845.0, + 761.0, + 845.0, + 867.0, + 792.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 771.0, + 874.0, + 771.0, + 874.0, + 857.0, + 855.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 760.0, + 1207.0, + 760.0, + 1207.0, + 867.0, + 1154.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 814.0, + 547.0, + 814.0, + 547.0, + 833.0, + 503.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 835.0, + 1192.0, + 835.0, + 1192.0, + 849.0, + 1156.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 856.0, + 546.0, + 856.0, + 546.0, + 873.0, + 502.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 858.0, + 1041.0, + 858.0, + 1041.0, + 868.0, + 1025.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 877.0, + 832.0, + 877.0, + 832.0, + 893.0, + 791.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 889.0, + 896.0, + 889.0, + 896.0, + 906.0, + 865.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 904.0, + 911.0, + 904.0, + 911.0, + 913.0, + 902.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 918.0, + 1196.0, + 918.0, + 1196.0, + 934.0, + 1150.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 929.0, + 552.0, + 929.0, + 552.0, + 943.0, + 540.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 928.0, + 604.0, + 928.0, + 604.0, + 944.0, + 586.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 928.0, + 653.0, + 928.0, + 653.0, + 944.0, + 636.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 928.0, + 703.0, + 928.0, + 703.0, + 944.0, + 684.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 928.0, + 752.0, + 928.0, + 752.0, + 944.0, + 733.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 927.0, + 804.0, + 927.0, + 804.0, + 945.0, + 779.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 929.0, + 901.0, + 929.0, + 901.0, + 943.0, + 889.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 928.0, + 957.0, + 928.0, + 957.0, + 944.0, + 937.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 929.0, + 1008.0, + 929.0, + 1008.0, + 944.0, + 991.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 928.0, + 1060.0, + 928.0, + 1060.0, + 944.0, + 1042.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 928.0, + 1112.0, + 928.0, + 1112.0, + 944.0, + 1093.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 927.0, + 1169.0, + 927.0, + 1169.0, + 945.0, + 1143.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 938.0, + 685.0, + 938.0, + 685.0, + 954.0, + 652.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 940.0, + 1041.0, + 940.0, + 1041.0, + 953.0, + 1009.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 973.0, + 763.0, + 973.0, + 763.0, + 1003.0, + 573.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 972.0, + 1128.0, + 972.0, + 1128.0, + 1003.0, + 935.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 790.0, + 941.0, + 790.0, + 941.0, + 815.0, + 910.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 835.0, + 949.0, + 835.0, + 949.0, + 853.0, + 942.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.75, + 858.5, + 986.75, + 858.5, + 986.75, + 875.5, + 965.75, + 875.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 759.5, + 1172.0, + 759.5, + 1172.0, + 809.5, + 984.0, + 809.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1152.0, + 833.0, + 1152.0, + 833.0, + 1169.0, + 649.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1153.0, + 1196.0, + 1153.0, + 1196.0, + 1169.0, + 1012.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1163.0, + 785.0, + 1163.0, + 785.0, + 1179.0, + 649.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1164.0, + 1148.0, + 1164.0, + 1148.0, + 1180.0, + 1013.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1172.0, + 907.0, + 1172.0, + 907.0, + 1188.0, + 864.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 1182.0, + 545.0, + 1182.0, + 545.0, + 1199.0, + 501.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1220.0, + 543.0, + 1220.0, + 543.0, + 1309.0, + 491.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1206.0, + 846.0, + 1206.0, + 846.0, + 1318.0, + 789.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1204.0, + 910.0, + 1204.0, + 910.0, + 1309.0, + 853.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1205.0, + 1209.0, + 1205.0, + 1209.0, + 1319.0, + 1150.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1270.0, + 544.0, + 1270.0, + 544.0, + 1286.0, + 502.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1299.0, + 833.0, + 1299.0, + 833.0, + 1314.0, + 789.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1297.0, + 1194.0, + 1297.0, + 1194.0, + 1313.0, + 1153.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1328.0, + 648.0, + 1328.0, + 648.0, + 1339.0, + 630.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1326.0, + 740.0, + 1326.0, + 740.0, + 1334.0, + 728.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1328.0, + 1002.0, + 1328.0, + 1002.0, + 1335.0, + 992.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1331.0, + 908.0, + 1331.0, + 908.0, + 1346.0, + 878.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1341.0, + 957.0, + 1341.0, + 957.0, + 1351.0, + 929.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1343.0, + 1071.0, + 1343.0, + 1071.0, + 1357.0, + 1049.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1356.0, + 741.0, + 1356.0, + 741.0, + 1366.0, + 724.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1382.0, + 552.0, + 1382.0, + 552.0, + 1395.0, + 539.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1380.0, + 604.0, + 1380.0, + 604.0, + 1396.0, + 586.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1380.0, + 654.0, + 1380.0, + 654.0, + 1396.0, + 635.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1380.0, + 703.0, + 1380.0, + 703.0, + 1396.0, + 684.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1380.0, + 752.0, + 1380.0, + 752.0, + 1396.0, + 732.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1379.0, + 804.0, + 1379.0, + 804.0, + 1397.0, + 779.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1381.0, + 915.0, + 1381.0, + 915.0, + 1395.0, + 903.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1380.0, + 968.0, + 1380.0, + 968.0, + 1396.0, + 949.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1380.0, + 1017.0, + 1380.0, + 1017.0, + 1396.0, + 998.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1380.0, + 1066.0, + 1380.0, + 1066.0, + 1396.0, + 1046.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 1380.0, + 1115.0, + 1380.0, + 1115.0, + 1396.0, + 1096.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1379.0, + 1167.0, + 1379.0, + 1167.0, + 1397.0, + 1143.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1392.0, + 685.0, + 1392.0, + 685.0, + 1406.0, + 653.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1392.0, + 1047.0, + 1392.0, + 1047.0, + 1406.0, + 1016.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1425.0, + 764.0, + 1425.0, + 764.0, + 1457.0, + 572.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1425.0, + 1129.0, + 1425.0, + 1129.0, + 1457.0, + 934.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1304.0, + 959.0, + 1304.0, + 959.0, + 1312.0, + 951.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1318.0, + 982.0, + 1318.0, + 982.0, + 1328.5, + 969.0, + 1328.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1342.0, + 976.0, + 1342.0, + 976.0, + 1352.0, + 964.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1357.0, + 546.0, + 1357.0, + 546.0, + 1374.0, + 515.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1480.0, + 1405.0, + 1480.0, + 1405.0, + 1516.0, + 294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 539.0, + 1511.0, + 539.0, + 1543.0, + 293.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1026.0, + 1406.0, + 1026.0, + 1406.0, + 1066.0, + 293.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1059.0, + 464.0, + 1059.0, + 464.0, + 1092.0, + 291.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 365.0, + 1768.0, + 365.0, + 1804.0, + 294.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1768.0, + 623.0, + 1768.0, + 623.0, + 1804.0, + 402.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 617.0, + 224.0, + 617.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 292.0, + 475.0, + 292.0, + 475.0, + 333.0, + 294.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 352.0, + 1406.0, + 352.0, + 1406.0, + 389.0, + 292.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 414.0, + 1406.0, + 414.0, + 1406.0, + 450.0, + 295.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 444.0, + 1405.0, + 444.0, + 1405.0, + 480.0, + 295.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 475.0, + 1406.0, + 475.0, + 1406.0, + 509.0, + 292.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 506.0, + 1405.0, + 506.0, + 1405.0, + 541.0, + 294.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 533.0, + 1405.0, + 533.0, + 1405.0, + 572.0, + 292.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 568.0, + 1403.0, + 568.0, + 1403.0, + 600.0, + 295.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 598.0, + 1403.0, + 598.0, + 1403.0, + 630.0, + 295.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 626.0, + 681.0, + 626.0, + 681.0, + 662.0, + 294.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1577.0, + 1402.0, + 1577.0, + 1402.0, + 1611.0, + 294.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1611.0, + 1404.0, + 1611.0, + 1404.0, + 1644.0, + 296.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1637.0, + 1405.0, + 1637.0, + 1405.0, + 1677.0, + 292.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1669.0, + 1404.0, + 1669.0, + 1404.0, + 1706.0, + 293.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1702.0, + 713.0, + 1702.0, + 713.0, + 1736.0, + 296.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1826.0, + 459.0, + 1826.0, + 459.0, + 1860.0, + 295.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1826.0, + 1404.0, + 1826.0, + 1404.0, + 1860.0, + 495.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1855.0, + 1326.0, + 1855.0, + 1326.0, + 1893.0, + 291.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1855.0, + 1406.0, + 1855.0, + 1406.0, + 1893.0, + 1401.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1888.0, + 794.0, + 1888.0, + 794.0, + 1922.0, + 295.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1930.0, + 625.0, + 1930.0, + 625.0, + 1970.0, + 292.0, + 1970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1930.0, + 1405.0, + 1930.0, + 1405.0, + 1970.0, + 661.0, + 1970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1964.0, + 1402.0, + 1964.0, + 1402.0, + 1998.0, + 296.0, + 1998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1994.0, + 443.0, + 1994.0, + 443.0, + 2029.0, + 292.0, + 2029.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 304, + 241, + 1395, + 241, + 1395, + 552, + 304, + 552 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 493, + 717, + 1208, + 717, + 1208, + 1031, + 493, + 1031 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 492, + 1615, + 1208, + 1615, + 1208, + 1927, + 492, + 1927 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 492, + 1165, + 1207, + 1165, + 1207, + 1479, + 492, + 1479 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 297, + 577, + 1405, + 577, + 1405, + 670, + 297, + 670 + ], + "score": 0.946 + }, + { + "category_id": 4, + "poly": [ + 296, + 1502, + 1401, + 1502, + 1401, + 1566, + 296, + 1566 + ], + "score": 0.943 + }, + { + "category_id": 4, + "poly": [ + 296, + 1950, + 1403, + 1950, + 1403, + 2015, + 296, + 2015 + ], + "score": 0.94 + }, + { + "category_id": 4, + "poly": [ + 285, + 1054, + 1405, + 1054, + 1405, + 1120, + 285, + 1120 + ], + "score": 0.938 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.86 + }, + { + "category_id": 13, + "poly": [ + 1091, + 1057, + 1125, + 1057, + 1125, + 1086, + 1091, + 1086 + ], + "score": 0.89, + "latex": "L _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 254.0, + 456.0, + 254.0, + 456.0, + 285.0, + 370.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 254.0, + 714.0, + 254.0, + 714.0, + 272.0, + 688.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 251.0, + 893.0, + 251.0, + 893.0, + 278.0, + 810.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 248.0, + 1011.0, + 248.0, + 1011.0, + 264.0, + 969.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 249.0, + 1093.0, + 249.0, + 1093.0, + 264.0, + 1050.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 249.0, + 1203.0, + 249.0, + 1203.0, + 277.0, + 1119.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 243.0, + 1382.0, + 243.0, + 1382.0, + 261.0, + 1335.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 274.0, + 459.0, + 274.0, + 459.0, + 314.0, + 371.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 287.0, + 647.0, + 287.0, + 647.0, + 305.0, + 605.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 273.0, + 898.0, + 273.0, + 898.0, + 313.0, + 811.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 289.0, + 1093.0, + 289.0, + 1093.0, + 304.0, + 1050.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 294.0, + 1110.0, + 294.0, + 1110.0, + 303.0, + 1099.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 271.0, + 1208.0, + 271.0, + 1208.0, + 312.0, + 1117.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 292.0, + 1382.0, + 292.0, + 1382.0, + 310.0, + 1337.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 313.0, + 1025.0, + 313.0, + 1025.0, + 404.0, + 970.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 317.0, + 1094.0, + 317.0, + 1094.0, + 335.0, + 1053.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 316.0, + 1394.0, + 316.0, + 1394.0, + 408.0, + 1374.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 327.0, + 323.0, + 327.0, + 323.0, + 400.0, + 307.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 326.0, + 693.0, + 326.0, + 693.0, + 397.0, + 677.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 343.0, + 928.0, + 343.0, + 928.0, + 353.0, + 918.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 325.0, + 1058.0, + 325.0, + 1058.0, + 398.0, + 1042.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 326.0, + 1380.0, + 326.0, + 1380.0, + 344.0, + 1340.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 366.0, + 644.0, + 366.0, + 644.0, + 382.0, + 605.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 358.0, + 1094.0, + 358.0, + 1094.0, + 376.0, + 1052.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 366.0, + 1158.0, + 366.0, + 1158.0, + 375.0, + 1139.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 376.0, + 1377.0, + 376.0, + 1377.0, + 393.0, + 1342.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 422.0, + 505.0, + 422.0, + 505.0, + 431.0, + 495.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 429.0, + 343.0, + 429.0, + 343.0, + 447.0, + 319.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 431.0, + 549.0, + 431.0, + 549.0, + 440.0, + 534.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 427.0, + 1093.0, + 427.0, + 1093.0, + 445.0, + 1067.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 443.0, + 634.0, + 443.0, + 634.0, + 461.0, + 605.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 447.0, + 771.0, + 447.0, + 771.0, + 457.0, + 758.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 452.0, + 601.0, + 452.0, + 601.0, + 466.0, + 578.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 459.0, + 1368.0, + 459.0, + 1368.0, + 477.0, + 1341.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 479.0, + 352.0, + 479.0, + 352.0, + 490.0, + 341.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 476.0, + 408.0, + 476.0, + 408.0, + 492.0, + 389.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 476.0, + 459.0, + 476.0, + 459.0, + 492.0, + 441.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 476.0, + 511.0, + 476.0, + 511.0, + 492.0, + 493.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 476.0, + 563.0, + 476.0, + 563.0, + 492.0, + 546.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 475.0, + 619.0, + 475.0, + 619.0, + 493.0, + 594.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 476.0, + 722.0, + 476.0, + 722.0, + 489.0, + 711.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 475.0, + 776.0, + 475.0, + 776.0, + 490.0, + 758.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 475.0, + 827.0, + 475.0, + 827.0, + 490.0, + 810.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 475.0, + 878.0, + 475.0, + 878.0, + 490.0, + 861.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 475.0, + 930.0, + 475.0, + 930.0, + 490.0, + 910.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 474.0, + 985.0, + 474.0, + 985.0, + 492.0, + 959.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 479.0, + 1100.0, + 479.0, + 1100.0, + 490.0, + 1090.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 476.0, + 1153.0, + 476.0, + 1153.0, + 492.0, + 1135.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 476.0, + 1201.0, + 476.0, + 1201.0, + 492.0, + 1183.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 476.0, + 1251.0, + 476.0, + 1251.0, + 492.0, + 1233.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 476.0, + 1300.0, + 476.0, + 1300.0, + 492.0, + 1282.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 475.0, + 1353.0, + 475.0, + 1353.0, + 493.0, + 1328.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 486.0, + 493.0, + 486.0, + 493.0, + 503.0, + 460.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 485.0, + 861.0, + 485.0, + 861.0, + 502.0, + 828.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 486.0, + 1235.0, + 486.0, + 1235.0, + 503.0, + 1201.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 520.0, + 544.0, + 520.0, + 544.0, + 552.0, + 419.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 518.0, + 907.0, + 518.0, + 907.0, + 552.0, + 791.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 519.0, + 1298.0, + 519.0, + 1298.0, + 553.0, + 1133.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 241.0, + 648.0, + 241.0, + 648.0, + 259.0, + 603.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 267.5, + 982.0, + 267.5, + 982.0, + 302.0, + 889.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 316.0, + 1186.0, + 316.0, + 1186.0, + 342.0, + 1131.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.25, + 386.0, + 1011.25, + 386.0, + 1011.25, + 404.0, + 969.25, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.25, + 461.0, + 716.25, + 461.0, + 716.25, + 480.0, + 686.25, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 723.0, + 534.0, + 723.0, + 534.0, + 740.0, + 503.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 730.0, + 640.0, + 730.0, + 640.0, + 743.0, + 557.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 721.0, + 832.0, + 721.0, + 832.0, + 738.0, + 787.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 722.0, + 895.0, + 722.0, + 895.0, + 736.0, + 864.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 729.0, + 1003.0, + 729.0, + 1003.0, + 745.0, + 920.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 722.0, + 1196.0, + 722.0, + 1196.0, + 739.0, + 1151.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 740.0, + 641.0, + 740.0, + 641.0, + 756.0, + 556.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 742.0, + 1003.0, + 742.0, + 1003.0, + 758.0, + 920.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 750.0, + 643.0, + 750.0, + 643.0, + 769.0, + 555.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 753.0, + 1004.0, + 753.0, + 1004.0, + 769.0, + 920.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 763.0, + 646.0, + 763.0, + 646.0, + 789.0, + 556.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 768.0, + 833.0, + 768.0, + 833.0, + 784.0, + 788.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 765.0, + 1010.0, + 765.0, + 1010.0, + 793.0, + 919.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 767.0, + 1196.0, + 767.0, + 1196.0, + 784.0, + 1152.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 803.0, + 509.0, + 803.0, + 509.0, + 876.0, + 493.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 850.0, + 618.0, + 850.0, + 618.0, + 865.0, + 598.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 794.0, + 846.0, + 794.0, + 846.0, + 886.0, + 787.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 803.0, + 875.0, + 803.0, + 875.0, + 878.0, + 856.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 795.0, + 1209.0, + 795.0, + 1209.0, + 887.0, + 1152.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 907.0, + 1097.0, + 907.0, + 1097.0, + 914.0, + 1089.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 911.0, + 895.0, + 911.0, + 895.0, + 925.0, + 865.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 918.0, + 577.0, + 918.0, + 577.0, + 935.0, + 540.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 936.0, + 785.0, + 936.0, + 785.0, + 944.0, + 773.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 926.0, + 818.0, + 926.0, + 818.0, + 941.0, + 791.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 924.0, + 1183.0, + 924.0, + 1183.0, + 942.0, + 1155.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 958.0, + 537.0, + 958.0, + 537.0, + 968.0, + 528.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 955.0, + 594.0, + 955.0, + 594.0, + 971.0, + 574.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 955.0, + 646.0, + 955.0, + 646.0, + 970.0, + 627.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 955.0, + 698.0, + 955.0, + 698.0, + 970.0, + 678.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 955.0, + 749.0, + 955.0, + 749.0, + 970.0, + 730.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 954.0, + 805.0, + 954.0, + 805.0, + 972.0, + 779.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 958.0, + 900.0, + 958.0, + 900.0, + 968.0, + 891.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 955.0, + 956.0, + 955.0, + 956.0, + 971.0, + 938.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 955.0, + 1008.0, + 955.0, + 1008.0, + 970.0, + 990.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 955.0, + 1060.0, + 955.0, + 1060.0, + 970.0, + 1042.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 955.0, + 1111.0, + 955.0, + 1111.0, + 970.0, + 1093.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 954.0, + 1167.0, + 954.0, + 1167.0, + 972.0, + 1143.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 967.0, + 677.0, + 967.0, + 677.0, + 980.0, + 647.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 967.0, + 1041.0, + 967.0, + 1041.0, + 980.0, + 1009.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 999.0, + 762.0, + 999.0, + 762.0, + 1031.0, + 574.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 998.0, + 1127.0, + 998.0, + 1127.0, + 1030.0, + 935.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 789.0, + 637.0, + 789.0, + 637.0, + 826.0, + 550.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 792.5, + 993.0, + 792.5, + 993.0, + 819.0, + 913.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 821.0, + 598.0, + 821.0, + 598.0, + 849.5, + 552.0, + 849.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 861.5, + 597.0, + 861.5, + 597.0, + 879.5, + 563.0, + 879.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.75, + 846.0, + 584.75, + 846.0, + 584.75, + 854.0, + 580.75, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.75, + 865.0, + 652.75, + 865.0, + 652.75, + 879.0, + 629.75, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 918.5, + 535.0, + 918.5, + 535.0, + 936.0, + 501.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1623.0, + 784.0, + 1623.0, + 784.0, + 1639.0, + 701.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1622.0, + 832.0, + 1622.0, + 832.0, + 1639.0, + 789.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1617.0, + 894.0, + 1617.0, + 894.0, + 1634.0, + 867.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1624.0, + 1145.0, + 1624.0, + 1145.0, + 1640.0, + 1063.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 1620.0, + 1195.0, + 1620.0, + 1195.0, + 1637.0, + 1152.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1632.0, + 531.0, + 1632.0, + 531.0, + 1649.0, + 504.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1634.0, + 784.0, + 1634.0, + 784.0, + 1650.0, + 701.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1636.0, + 1146.0, + 1636.0, + 1146.0, + 1652.0, + 1064.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1645.0, + 783.0, + 1645.0, + 783.0, + 1661.0, + 699.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1648.0, + 1147.0, + 1648.0, + 1147.0, + 1664.0, + 1064.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1658.0, + 788.0, + 1658.0, + 788.0, + 1674.0, + 700.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1661.0, + 1150.0, + 1661.0, + 1150.0, + 1674.0, + 1065.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1669.0, + 833.0, + 1669.0, + 833.0, + 1689.0, + 700.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1671.0, + 1196.0, + 1671.0, + 1196.0, + 1687.0, + 1064.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1699.0, + 509.0, + 1699.0, + 509.0, + 1771.0, + 493.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1691.0, + 845.0, + 1691.0, + 845.0, + 1779.0, + 790.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1700.0, + 873.0, + 1700.0, + 873.0, + 1771.0, + 857.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1691.0, + 1208.0, + 1691.0, + 1208.0, + 1782.0, + 1154.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1759.0, + 831.0, + 1759.0, + 831.0, + 1774.0, + 789.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1756.0, + 993.0, + 1756.0, + 993.0, + 1767.0, + 976.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1757.0, + 1192.0, + 1757.0, + 1192.0, + 1774.0, + 1154.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1784.0, + 920.0, + 1784.0, + 920.0, + 1799.0, + 902.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1794.0, + 894.0, + 1794.0, + 894.0, + 1811.0, + 867.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1816.0, + 531.0, + 1816.0, + 531.0, + 1830.0, + 502.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1823.0, + 1151.0, + 1823.0, + 1151.0, + 1838.0, + 1129.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1852.0, + 538.0, + 1852.0, + 538.0, + 1865.0, + 526.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1850.0, + 594.0, + 1850.0, + 594.0, + 1866.0, + 575.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1851.0, + 646.0, + 1851.0, + 646.0, + 1866.0, + 627.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1850.0, + 697.0, + 1850.0, + 697.0, + 1866.0, + 678.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1850.0, + 749.0, + 1850.0, + 749.0, + 1866.0, + 730.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1849.0, + 804.0, + 1849.0, + 804.0, + 1867.0, + 779.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1852.0, + 902.0, + 1852.0, + 902.0, + 1865.0, + 890.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1850.0, + 957.0, + 1850.0, + 957.0, + 1866.0, + 938.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1850.0, + 1009.0, + 1850.0, + 1009.0, + 1866.0, + 990.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1850.0, + 1060.0, + 1850.0, + 1060.0, + 1866.0, + 1041.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1850.0, + 1113.0, + 1850.0, + 1113.0, + 1866.0, + 1093.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1849.0, + 1168.0, + 1849.0, + 1168.0, + 1867.0, + 1142.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1862.0, + 677.0, + 1862.0, + 677.0, + 1876.0, + 646.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1862.0, + 1041.0, + 1862.0, + 1041.0, + 1876.0, + 1010.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1896.0, + 763.0, + 1896.0, + 763.0, + 1927.0, + 573.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1895.0, + 1127.0, + 1895.0, + 1127.0, + 1926.0, + 935.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.25, + 1700.5, + 712.25, + 1700.5, + 712.25, + 1738.5, + 545.25, + 1738.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1680.0, + 940.0, + 1680.0, + 940.0, + 1697.0, + 904.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.75, + 1698.0, + 970.75, + 1698.0, + 970.75, + 1712.0, + 945.75, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1721.0, + 1089.0, + 1721.0, + 1089.0, + 1752.0, + 976.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1729.5, + 592.0, + 1729.5, + 592.0, + 1750.5, + 561.0, + 1750.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1739.5, + 966.0, + 1739.5, + 966.0, + 1753.5, + 927.0, + 1753.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1747.0, + 626.0, + 1747.0, + 626.0, + 1767.0, + 597.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1765.5, + 1025.0, + 1765.5, + 1025.0, + 1781.5, + 1000.0, + 1781.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1181.0, + 528.0, + 1181.0, + 528.0, + 1196.0, + 505.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1176.0, + 712.0, + 1176.0, + 712.0, + 1192.0, + 628.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1171.0, + 832.0, + 1171.0, + 832.0, + 1187.0, + 789.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1172.0, + 894.0, + 1172.0, + 894.0, + 1187.0, + 869.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1174.0, + 1194.0, + 1174.0, + 1194.0, + 1188.0, + 1154.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1187.0, + 712.0, + 1187.0, + 712.0, + 1203.0, + 629.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1199.0, + 712.0, + 1199.0, + 712.0, + 1215.0, + 629.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1210.0, + 717.0, + 1210.0, + 717.0, + 1225.0, + 629.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1221.0, + 719.0, + 1221.0, + 719.0, + 1237.0, + 628.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1251.0, + 509.0, + 1251.0, + 509.0, + 1323.0, + 493.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1276.0, + 765.0, + 1276.0, + 765.0, + 1286.0, + 757.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1242.0, + 845.0, + 1242.0, + 845.0, + 1333.0, + 787.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1251.0, + 872.0, + 1251.0, + 872.0, + 1323.0, + 855.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1237.0, + 1207.0, + 1237.0, + 1207.0, + 1332.0, + 1155.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1306.0, + 1194.0, + 1306.0, + 1194.0, + 1323.0, + 1154.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1337.0, + 832.0, + 1337.0, + 832.0, + 1354.0, + 790.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1344.0, + 895.0, + 1344.0, + 895.0, + 1362.0, + 867.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1358.0, + 593.0, + 1358.0, + 593.0, + 1378.0, + 577.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1370.0, + 529.0, + 1370.0, + 529.0, + 1385.0, + 505.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1370.0, + 922.0, + 1370.0, + 922.0, + 1383.0, + 901.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1388.0, + 1196.0, + 1388.0, + 1196.0, + 1405.0, + 1153.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1407.0, + 537.0, + 1407.0, + 537.0, + 1415.0, + 529.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1403.0, + 594.0, + 1403.0, + 594.0, + 1419.0, + 574.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1403.0, + 646.0, + 1403.0, + 646.0, + 1419.0, + 627.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1403.0, + 698.0, + 1403.0, + 698.0, + 1418.0, + 678.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1403.0, + 749.0, + 1403.0, + 749.0, + 1419.0, + 730.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1403.0, + 804.0, + 1403.0, + 804.0, + 1418.0, + 780.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1406.0, + 900.0, + 1406.0, + 900.0, + 1416.0, + 891.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1403.0, + 956.0, + 1403.0, + 956.0, + 1419.0, + 938.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1403.0, + 1008.0, + 1403.0, + 1008.0, + 1418.0, + 990.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1403.0, + 1060.0, + 1403.0, + 1060.0, + 1419.0, + 1041.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1403.0, + 1112.0, + 1403.0, + 1112.0, + 1419.0, + 1094.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1402.0, + 1168.0, + 1402.0, + 1168.0, + 1420.0, + 1142.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1413.0, + 678.0, + 1413.0, + 678.0, + 1429.0, + 645.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1413.0, + 1042.0, + 1413.0, + 1042.0, + 1429.0, + 1008.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1446.0, + 764.0, + 1446.0, + 764.0, + 1481.0, + 572.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1446.0, + 1129.0, + 1446.0, + 1129.0, + 1479.0, + 934.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1197.0, + 818.0, + 1197.0, + 818.0, + 1239.0, + 698.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 575.0, + 1406.0, + 575.0, + 1406.0, + 613.0, + 294.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 607.0, + 1405.0, + 607.0, + 1405.0, + 640.0, + 294.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 638.0, + 558.0, + 638.0, + 558.0, + 672.0, + 296.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1542.0, + 293.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1535.0, + 771.0, + 1535.0, + 771.0, + 1567.0, + 296.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1946.0, + 1404.0, + 1946.0, + 1404.0, + 1990.0, + 293.0, + 1990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1981.0, + 827.0, + 1981.0, + 827.0, + 2015.0, + 292.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1053.0, + 1090.0, + 1053.0, + 1090.0, + 1090.0, + 295.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1053.0, + 1403.0, + 1053.0, + 1403.0, + 1090.0, + 1126.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1086.0, + 463.0, + 1086.0, + 463.0, + 1119.0, + 292.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 695, + 1405, + 695, + 1405, + 1061, + 297, + 1061 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 295, + 1404, + 295, + 1404, + 601, + 297, + 601 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1788, + 1403, + 1788, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 305, + 1383, + 1393, + 1383, + 1393, + 1632, + 305, + 1632 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 297, + 1165, + 1382, + 1165, + 1382, + 1284, + 297, + 1284 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 296, + 1655, + 1400, + 1655, + 1400, + 1750, + 296, + 1750 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 296, + 1078, + 1399, + 1078, + 1399, + 1140, + 296, + 1140 + ], + "score": 0.944 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 580, + 226, + 580, + 261, + 300, + 261 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 299, + 639, + 445, + 639, + 445, + 669, + 299, + 669 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 0, + "poly": [ + 298, + 1318, + 474, + 1318, + 474, + 1348, + 298, + 1348 + ], + "score": 0.827 + }, + { + "category_id": 13, + "poly": [ + 967, + 819, + 1039, + 819, + 1039, + 850, + 967, + 850 + ], + "score": 0.91, + "latex": "\\eta = 2" + }, + { + "category_id": 13, + "poly": [ + 1058, + 942, + 1320, + 942, + 1320, + 972, + 1058, + 972 + ], + "score": 0.91, + "latex": "t _ { s i m } = t _ { s e a r c h } + t _ { t r a i n }" + }, + { + "category_id": 13, + "poly": [ + 1199, + 974, + 1267, + 974, + 1267, + 1001, + 1199, + 1001 + ], + "score": 0.9, + "latex": "t _ { t r a i n }" + }, + { + "category_id": 13, + "poly": [ + 561, + 388, + 624, + 388, + 624, + 421, + 561, + 421 + ], + "score": 0.89, + "latex": "( 1 / \\eta )" + }, + { + "category_id": 13, + "poly": [ + 331, + 1248, + 471, + 1248, + 471, + 1283, + 331, + 1283 + ], + "score": 0.87, + "latex": "C S 3 = \\{ L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 972, + 379, + 972, + 379, + 1002, + 297, + 1002 + ], + "score": 0.84, + "latex": "t _ { s e a r c h }" + }, + { + "category_id": 13, + "poly": [ + 973, + 391, + 993, + 391, + 993, + 420, + 973, + 420 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 332, + 1206, + 471, + 1206, + 471, + 1241, + 332, + 1241 + ], + "score": 0.75, + "latex": "C S 2 = \\{ L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 440, + 819, + 508, + 819, + 508, + 848, + 440, + 848 + ], + "score": 0.64, + "latex": "= 2 5" + }, + { + "category_id": 13, + "poly": [ + 742, + 819, + 832, + 819, + 832, + 848, + 742, + 848 + ], + "score": 0.59, + "latex": "t = 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 332, + 1165, + 471, + 1165, + 471, + 1200, + 332, + 1200 + ], + "score": 0.57, + "latex": "C S 1 = \\{ L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 845, + 910, + 891, + 910, + 891, + 938, + 845, + 938 + ], + "score": 0.42, + "latex": "2 0 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 297, + 944, + 315, + 944, + 315, + 968, + 297, + 968 + ], + "score": 0.41, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 487, + 1248, + 534, + 1248, + 534, + 1280, + 487, + 1280 + ], + "score": 0.38, + "latex": "C O" + }, + { + "category_id": 13, + "poly": [ + 331, + 1206, + 654, + 1206, + 654, + 1242, + 331, + 1242 + ], + "score": 0.32, + "latex": "C S 2 = \\{ L _ { 2 } , C O _ { - } p r o b , l r \\}" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1389.0, + 526.0, + 1389.0, + 526.0, + 1408.0, + 475.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1389.0, + 891.0, + 1389.0, + 891.0, + 1408.0, + 838.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1389.0, + 1261.0, + 1389.0, + 1261.0, + 1408.0, + 1208.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1406.0, + 656.0, + 1406.0, + 656.0, + 1421.0, + 601.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1406.0, + 1017.0, + 1406.0, + 1017.0, + 1421.0, + 967.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1406.0, + 1388.0, + 1406.0, + 1388.0, + 1421.0, + 1335.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1419.0, + 616.0, + 1419.0, + 616.0, + 1430.0, + 604.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1417.0, + 980.0, + 1417.0, + 980.0, + 1442.0, + 952.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 1419.0, + 1349.0, + 1419.0, + 1349.0, + 1430.0, + 1339.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1425.0, + 620.0, + 1425.0, + 620.0, + 1451.0, + 594.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1426.0, + 981.0, + 1426.0, + 981.0, + 1441.0, + 967.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1425.0, + 1353.0, + 1425.0, + 1353.0, + 1451.0, + 1335.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1436.0, + 967.0, + 1436.0, + 967.0, + 1455.0, + 952.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1437.0, + 980.0, + 1437.0, + 980.0, + 1448.0, + 968.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1450.0, + 342.0, + 1450.0, + 342.0, + 1509.0, + 308.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1443.0, + 629.0, + 1443.0, + 629.0, + 1477.0, + 593.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1446.0, + 708.0, + 1446.0, + 708.0, + 1509.0, + 670.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 1461.0, + 783.0, + 1461.0, + 783.0, + 1471.0, + 772.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1458.0, + 964.0, + 1458.0, + 964.0, + 1468.0, + 953.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1446.0, + 991.0, + 1446.0, + 991.0, + 1477.0, + 967.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1450.0, + 1078.0, + 1450.0, + 1078.0, + 1509.0, + 1041.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1448.0, + 1140.0, + 1448.0, + 1140.0, + 1458.0, + 1131.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 1443.0, + 1361.0, + 1443.0, + 1361.0, + 1477.0, + 1337.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1475.0, + 619.0, + 1475.0, + 619.0, + 1486.0, + 606.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1476.0, + 983.0, + 1476.0, + 983.0, + 1486.0, + 970.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 1474.0, + 1354.0, + 1474.0, + 1354.0, + 1487.0, + 1338.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1497.0, + 710.0, + 1497.0, + 710.0, + 1514.0, + 683.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1507.0, + 347.0, + 1507.0, + 347.0, + 1524.0, + 321.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1509.0, + 491.0, + 1509.0, + 491.0, + 1521.0, + 467.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1521.0, + 1080.0, + 1521.0, + 1080.0, + 1535.0, + 1051.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1523.0, + 1238.0, + 1523.0, + 1238.0, + 1532.0, + 1224.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1527.0, + 1266.0, + 1527.0, + 1266.0, + 1537.0, + 1254.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1540.0, + 948.0, + 1540.0, + 948.0, + 1550.0, + 934.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1555.0, + 390.0, + 1555.0, + 390.0, + 1569.0, + 369.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1555.0, + 464.0, + 1555.0, + 464.0, + 1569.0, + 443.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1555.0, + 538.0, + 1555.0, + 538.0, + 1569.0, + 519.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1555.0, + 614.0, + 1555.0, + 614.0, + 1569.0, + 593.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1554.0, + 754.0, + 1554.0, + 754.0, + 1572.0, + 730.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1555.0, + 827.0, + 1555.0, + 827.0, + 1569.0, + 806.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1555.0, + 901.0, + 1555.0, + 901.0, + 1569.0, + 880.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1555.0, + 975.0, + 1555.0, + 975.0, + 1569.0, + 955.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 1555.0, + 1122.0, + 1555.0, + 1122.0, + 1569.0, + 1101.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1555.0, + 1198.0, + 1555.0, + 1198.0, + 1569.0, + 1177.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1555.0, + 1272.0, + 1555.0, + 1272.0, + 1569.0, + 1251.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 1555.0, + 1346.0, + 1555.0, + 1346.0, + 1569.0, + 1325.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1564.0, + 568.0, + 1564.0, + 568.0, + 1584.0, + 436.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1564.0, + 932.0, + 1564.0, + 932.0, + 1584.0, + 798.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1564.0, + 1301.0, + 1564.0, + 1301.0, + 1584.0, + 1170.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1600.0, + 579.0, + 1600.0, + 579.0, + 1632.0, + 386.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1600.0, + 943.0, + 1600.0, + 943.0, + 1632.0, + 749.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1600.0, + 1313.0, + 1600.0, + 1313.0, + 1632.0, + 1120.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1475.0, + 444.0, + 1475.0, + 444.0, + 1500.0, + 413.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1484.5, + 839.0, + 1484.5, + 839.0, + 1505.5, + 782.0, + 1505.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1493.0, + 1179.0, + 1493.0, + 1179.0, + 1509.0, + 1141.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1510.0, + 869.0, + 1510.0, + 869.0, + 1518.5, + 854.0, + 1518.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1518.5, + 910.0, + 1518.5, + 910.0, + 1531.5, + 890.0, + 1531.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1654.0, + 1404.0, + 1654.0, + 1404.0, + 1693.0, + 292.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1685.0, + 1402.0, + 1685.0, + 1402.0, + 1721.0, + 294.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1718.0, + 950.0, + 1718.0, + 950.0, + 1752.0, + 295.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 223.0, + 584.0, + 223.0, + 584.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 637.0, + 447.0, + 637.0, + 447.0, + 673.0, + 293.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1314.0, + 477.0, + 1314.0, + 477.0, + 1354.0, + 292.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 696.0, + 1402.0, + 696.0, + 1402.0, + 730.0, + 295.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 726.0, + 1406.0, + 726.0, + 1406.0, + 760.0, + 295.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 758.0, + 1405.0, + 758.0, + 1405.0, + 792.0, + 295.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 785.0, + 1407.0, + 785.0, + 1407.0, + 825.0, + 292.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 819.0, + 439.0, + 819.0, + 439.0, + 853.0, + 295.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 819.0, + 741.0, + 819.0, + 741.0, + 853.0, + 509.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 819.0, + 966.0, + 819.0, + 966.0, + 853.0, + 833.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 819.0, + 1405.0, + 819.0, + 1405.0, + 853.0, + 1040.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 849.0, + 1405.0, + 849.0, + 1405.0, + 883.0, + 295.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 880.0, + 1403.0, + 880.0, + 1403.0, + 912.0, + 294.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 907.0, + 844.0, + 907.0, + 844.0, + 944.0, + 294.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 907.0, + 1405.0, + 907.0, + 1405.0, + 944.0, + 892.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 937.0, + 296.0, + 937.0, + 296.0, + 978.0, + 291.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 937.0, + 1057.0, + 937.0, + 1057.0, + 978.0, + 316.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 937.0, + 1406.0, + 937.0, + 1406.0, + 978.0, + 1321.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 970.0, + 296.0, + 970.0, + 296.0, + 1005.0, + 291.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 970.0, + 1198.0, + 970.0, + 1198.0, + 1005.0, + 380.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 970.0, + 1405.0, + 970.0, + 1405.0, + 1005.0, + 1268.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1001.0, + 1405.0, + 1001.0, + 1405.0, + 1035.0, + 294.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1032.0, + 431.0, + 1032.0, + 431.0, + 1067.0, + 294.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 294.0, + 1405.0, + 294.0, + 1405.0, + 333.0, + 294.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 322.0, + 1406.0, + 322.0, + 1406.0, + 366.0, + 292.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 355.0, + 1407.0, + 355.0, + 1407.0, + 392.0, + 291.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 390.0, + 560.0, + 390.0, + 560.0, + 421.0, + 296.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 390.0, + 972.0, + 390.0, + 972.0, + 421.0, + 625.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 390.0, + 1402.0, + 390.0, + 1402.0, + 421.0, + 994.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 418.0, + 1406.0, + 418.0, + 1406.0, + 453.0, + 291.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 449.0, + 1404.0, + 449.0, + 1404.0, + 485.0, + 294.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 477.0, + 1404.0, + 477.0, + 1404.0, + 514.0, + 292.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 509.0, + 1404.0, + 509.0, + 1404.0, + 543.0, + 294.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1404.0, + 539.0, + 1404.0, + 575.0, + 294.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 571.0, + 596.0, + 571.0, + 596.0, + 603.0, + 293.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1826.0, + 294.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1404.0, + 1821.0, + 1404.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1915.0, + 296.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1404.0, + 2001.0, + 1404.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1162.0, + 331.0, + 1162.0, + 331.0, + 1202.0, + 293.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1162.0, + 615.0, + 1162.0, + 615.0, + 1202.0, + 472.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1202.0, + 330.0, + 1202.0, + 330.0, + 1244.0, + 290.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1245.0, + 330.0, + 1245.0, + 330.0, + 1286.0, + 290.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1245.0, + 486.0, + 1245.0, + 486.0, + 1286.0, + 472.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1245.0, + 1379.0, + 1245.0, + 1379.0, + 1286.0, + 535.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1077.0, + 1403.0, + 1077.0, + 1403.0, + 1113.0, + 296.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1106.0, + 1032.0, + 1106.0, + 1032.0, + 1142.0, + 295.0, + 1142.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1674, + 1404, + 1674, + 1404, + 1797, + 299, + 1797 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 308, + 331, + 1390, + 331, + 1390, + 675, + 308, + 675 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 307, + 874, + 1393, + 874, + 1393, + 1121, + 307, + 1121 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 306, + 1228, + 1394, + 1228, + 1394, + 1475, + 306, + 1475 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 296, + 698, + 1405, + 698, + 1405, + 822, + 296, + 822 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1402, + 229, + 1402, + 293, + 294, + 293 + ], + "score": 0.936 + }, + { + "category_id": 4, + "poly": [ + 286, + 1498, + 1403, + 1498, + 1403, + 1560, + 286, + 1560 + ], + "score": 0.914 + }, + { + "category_id": 4, + "poly": [ + 301, + 1144, + 1390, + 1144, + 1390, + 1178, + 301, + 1178 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 300, + 1842, + 824, + 1842, + 824, + 1876, + 300, + 1876 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 864, + 2088, + 864, + 2112, + 837, + 2112 + ], + "score": 0.85 + }, + { + "category_id": 0, + "poly": [ + 303, + 1617, + 814, + 1617, + 814, + 1648, + 303, + 1648 + ], + "score": 0.756 + }, + { + "category_id": 1, + "poly": [ + 303, + 1617, + 814, + 1617, + 814, + 1648, + 303, + 1648 + ], + "score": 0.151 + }, + { + "category_id": 15, + "poly": [ + 565.0, + 330.0, + 640.0, + 330.0, + 640.0, + 359.0, + 565.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 330.0, + 1190.0, + 330.0, + 1190.0, + 359.0, + 1114.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 358.0, + 815.0, + 358.0, + 815.0, + 378.0, + 739.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 357.0, + 1365.0, + 357.0, + 1365.0, + 377.0, + 1291.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 374.0, + 829.0, + 374.0, + 829.0, + 391.0, + 743.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 373.0, + 1380.0, + 373.0, + 1380.0, + 389.0, + 1292.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 387.0, + 441.0, + 387.0, + 441.0, + 397.0, + 430.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 387.0, + 810.0, + 387.0, + 810.0, + 407.0, + 741.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 388.0, + 1359.0, + 388.0, + 1359.0, + 405.0, + 1292.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 402.0, + 763.0, + 402.0, + 763.0, + 421.0, + 742.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 402.0, + 1313.0, + 402.0, + 1313.0, + 421.0, + 1292.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 424.0, + 368.0, + 424.0, + 368.0, + 509.0, + 311.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 416.0, + 764.0, + 416.0, + 764.0, + 447.0, + 741.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 418.0, + 917.0, + 418.0, + 917.0, + 509.0, + 861.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 416.0, + 1313.0, + 416.0, + 1313.0, + 447.0, + 1292.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 445.0, + 778.0, + 445.0, + 778.0, + 463.0, + 742.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 445.0, + 1328.0, + 445.0, + 1328.0, + 463.0, + 1292.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 460.0, + 764.0, + 460.0, + 764.0, + 475.0, + 743.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 460.0, + 1314.0, + 460.0, + 1314.0, + 475.0, + 1293.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 474.0, + 778.0, + 474.0, + 778.0, + 492.0, + 743.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 474.0, + 1328.0, + 474.0, + 1328.0, + 492.0, + 1292.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 488.0, + 768.0, + 488.0, + 768.0, + 506.0, + 742.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 488.0, + 1318.0, + 488.0, + 1318.0, + 506.0, + 1292.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 496.0, + 918.0, + 496.0, + 918.0, + 515.0, + 878.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 516.0, + 1230.0, + 516.0, + 1230.0, + 540.0, + 1193.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 531.0, + 1257.0, + 531.0, + 1257.0, + 548.0, + 1229.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 580.0, + 434.0, + 580.0, + 434.0, + 600.0, + 399.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 579.0, + 794.0, + 579.0, + 794.0, + 603.0, + 759.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 580.0, + 984.0, + 580.0, + 984.0, + 600.0, + 949.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 580.0, + 1103.0, + 580.0, + 1103.0, + 600.0, + 1068.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 576.0, + 1226.0, + 576.0, + 1226.0, + 604.0, + 1190.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 598.0, + 699.0, + 598.0, + 699.0, + 622.0, + 505.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 598.0, + 1250.0, + 598.0, + 1250.0, + 622.0, + 1055.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 642.0, + 670.0, + 642.0, + 670.0, + 676.0, + 477.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 642.0, + 1223.0, + 642.0, + 1223.0, + 676.0, + 1027.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 508.5, + 371.0, + 508.5, + 371.0, + 531.5, + 327.0, + 531.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1152.0, + 510.0, + 1167.0, + 510.0, + 1167.0, + 526.0, + 1152.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.75, + 578.0, + 556.75, + 578.0, + 556.75, + 602.5, + 517.75, + 602.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 577.0, + 679.0, + 577.0, + 679.0, + 603.0, + 636.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 577.0, + 1347.0, + 577.0, + 1347.0, + 603.0, + 1305.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 876.0, + 528.0, + 876.0, + 528.0, + 896.0, + 476.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 876.0, + 887.0, + 876.0, + 887.0, + 896.0, + 835.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 876.0, + 1261.0, + 876.0, + 1261.0, + 896.0, + 1208.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 891.0, + 655.0, + 891.0, + 655.0, + 911.0, + 588.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 895.0, + 724.0, + 895.0, + 724.0, + 904.0, + 712.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 894.0, + 1009.0, + 894.0, + 1009.0, + 908.0, + 946.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 894.0, + 1389.0, + 894.0, + 1389.0, + 910.0, + 1337.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 902.0, + 623.0, + 902.0, + 623.0, + 931.0, + 587.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 907.0, + 961.0, + 907.0, + 961.0, + 917.0, + 949.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 902.0, + 1354.0, + 902.0, + 1354.0, + 940.0, + 1332.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 914.0, + 975.0, + 914.0, + 975.0, + 929.0, + 949.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 916.0, + 1337.0, + 916.0, + 1337.0, + 926.0, + 1327.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 923.0, + 619.0, + 923.0, + 619.0, + 937.0, + 586.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 923.0, + 977.0, + 923.0, + 977.0, + 940.0, + 944.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 924.0, + 1337.0, + 924.0, + 1337.0, + 938.0, + 1322.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 931.0, + 630.0, + 931.0, + 630.0, + 958.0, + 586.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 932.0, + 711.0, + 932.0, + 711.0, + 998.0, + 670.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 931.0, + 986.0, + 931.0, + 986.0, + 958.0, + 943.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 938.0, + 1080.0, + 938.0, + 1080.0, + 996.0, + 1040.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 931.0, + 1361.0, + 931.0, + 1361.0, + 957.0, + 1323.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 952.0, + 629.0, + 952.0, + 629.0, + 965.0, + 604.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 952.0, + 985.0, + 952.0, + 985.0, + 965.0, + 959.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 952.0, + 1360.0, + 952.0, + 1360.0, + 965.0, + 1337.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 961.0, + 621.0, + 961.0, + 621.0, + 975.0, + 596.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 961.0, + 1354.0, + 961.0, + 1354.0, + 975.0, + 1336.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 983.0, + 709.0, + 983.0, + 709.0, + 1002.0, + 683.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 984.0, + 856.0, + 984.0, + 856.0, + 1001.0, + 830.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 994.0, + 347.0, + 994.0, + 347.0, + 1012.0, + 320.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1011.0, + 638.0, + 1011.0, + 638.0, + 1022.0, + 625.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1007.0, + 1080.0, + 1007.0, + 1080.0, + 1025.0, + 1054.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1031.0, + 583.0, + 1031.0, + 583.0, + 1037.0, + 577.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1041.0, + 389.0, + 1041.0, + 389.0, + 1059.0, + 366.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1041.0, + 461.0, + 1041.0, + 461.0, + 1059.0, + 436.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1042.0, + 530.0, + 1042.0, + 530.0, + 1057.0, + 510.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1042.0, + 600.0, + 1042.0, + 600.0, + 1057.0, + 581.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 1041.0, + 751.0, + 1041.0, + 751.0, + 1059.0, + 726.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1041.0, + 820.0, + 1041.0, + 820.0, + 1058.0, + 796.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1042.0, + 885.0, + 1042.0, + 885.0, + 1057.0, + 864.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1042.0, + 953.0, + 1042.0, + 953.0, + 1057.0, + 932.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1041.0, + 1020.0, + 1041.0, + 1020.0, + 1059.0, + 998.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1041.0, + 1122.0, + 1041.0, + 1122.0, + 1059.0, + 1098.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1042.0, + 1193.0, + 1042.0, + 1193.0, + 1057.0, + 1172.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1038.0, + 1265.0, + 1038.0, + 1265.0, + 1058.0, + 1240.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1042.0, + 1336.0, + 1042.0, + 1336.0, + 1057.0, + 1315.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1052.0, + 567.0, + 1052.0, + 567.0, + 1071.0, + 436.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1052.0, + 927.0, + 1052.0, + 927.0, + 1071.0, + 794.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1052.0, + 1301.0, + 1052.0, + 1301.0, + 1071.0, + 1169.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1087.0, + 578.0, + 1087.0, + 578.0, + 1122.0, + 387.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1088.0, + 944.0, + 1088.0, + 944.0, + 1120.0, + 749.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1087.0, + 1313.0, + 1087.0, + 1313.0, + 1122.0, + 1119.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.75, + 904.5, + 976.75, + 904.5, + 976.75, + 917.5, + 956.75, + 917.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.75, + 951.0, + 818.75, + 951.0, + 818.75, + 988.5, + 763.75, + 988.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.25, + 971.0, + 476.25, + 971.0, + 476.25, + 1001.5, + 408.25, + 1001.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 951.5, + 978.0, + 951.5, + 978.0, + 976.5, + 943.0, + 976.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.25, + 991.0, + 1213.25, + 991.0, + 1213.25, + 1015.0, + 1140.25, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 998.0, + 901.0, + 998.0, + 901.0, + 1017.5, + 852.0, + 1017.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1230.0, + 527.0, + 1230.0, + 527.0, + 1250.0, + 475.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1230.0, + 891.0, + 1230.0, + 891.0, + 1250.0, + 838.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1230.0, + 1261.0, + 1230.0, + 1261.0, + 1250.0, + 1208.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1250.0, + 607.0, + 1250.0, + 607.0, + 1301.0, + 590.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1249.0, + 640.0, + 1249.0, + 640.0, + 1258.0, + 625.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1248.0, + 1012.0, + 1248.0, + 1012.0, + 1305.0, + 950.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 1249.0, + 1387.0, + 1249.0, + 1387.0, + 1311.0, + 1330.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1291.0, + 344.0, + 1291.0, + 344.0, + 1351.0, + 307.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1298.0, + 604.0, + 1298.0, + 604.0, + 1308.0, + 594.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1286.0, + 710.0, + 1286.0, + 710.0, + 1352.0, + 669.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1297.0, + 968.0, + 1297.0, + 968.0, + 1309.0, + 956.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 1291.0, + 981.0, + 1291.0, + 981.0, + 1301.0, + 973.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1292.0, + 1079.0, + 1292.0, + 1079.0, + 1351.0, + 1042.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1291.0, + 1346.0, + 1291.0, + 1346.0, + 1301.0, + 1335.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1337.0, + 711.0, + 1337.0, + 711.0, + 1354.0, + 683.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1348.0, + 346.0, + 1348.0, + 346.0, + 1366.0, + 320.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1361.0, + 1080.0, + 1361.0, + 1080.0, + 1379.0, + 1053.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1361.0, + 1240.0, + 1361.0, + 1240.0, + 1374.0, + 1221.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1395.0, + 390.0, + 1395.0, + 390.0, + 1413.0, + 366.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1396.0, + 536.0, + 1396.0, + 536.0, + 1411.0, + 515.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1394.0, + 610.0, + 1394.0, + 610.0, + 1412.0, + 585.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1395.0, + 752.0, + 1395.0, + 752.0, + 1413.0, + 729.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1396.0, + 825.0, + 1396.0, + 825.0, + 1411.0, + 804.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1394.0, + 900.0, + 1394.0, + 900.0, + 1412.0, + 876.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1395.0, + 971.0, + 1395.0, + 971.0, + 1413.0, + 949.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 1395.0, + 1123.0, + 1395.0, + 1123.0, + 1413.0, + 1099.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1394.0, + 1197.0, + 1394.0, + 1197.0, + 1412.0, + 1173.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1396.0, + 1268.0, + 1396.0, + 1268.0, + 1411.0, + 1247.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 1394.0, + 1343.0, + 1394.0, + 1343.0, + 1412.0, + 1319.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1406.0, + 568.0, + 1406.0, + 568.0, + 1425.0, + 435.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1406.0, + 930.0, + 1406.0, + 930.0, + 1425.0, + 799.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1406.0, + 1300.0, + 1406.0, + 1300.0, + 1425.0, + 1169.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1441.0, + 579.0, + 1441.0, + 579.0, + 1476.0, + 386.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1442.0, + 944.0, + 1442.0, + 944.0, + 1474.0, + 748.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1441.0, + 1313.0, + 1441.0, + 1313.0, + 1476.0, + 1120.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1318.5, + 430.0, + 1318.5, + 430.0, + 1335.5, + 414.0, + 1335.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.75, + 1317.0, + 835.75, + 1317.0, + 835.75, + 1345.0, + 774.75, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.75, + 1329.5, + 1157.75, + 1329.5, + 1157.75, + 1345.0, + 1140.75, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1361.5, + 1290.0, + 1361.5, + 1290.0, + 1382.5, + 1244.0, + 1382.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1394.0, + 465.0, + 1394.0, + 465.0, + 1412.0, + 437.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 697.0, + 1405.0, + 697.0, + 1405.0, + 734.0, + 293.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 727.0, + 1405.0, + 727.0, + 1405.0, + 763.0, + 294.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 760.0, + 1405.0, + 760.0, + 1405.0, + 795.0, + 293.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 793.0, + 490.0, + 793.0, + 490.0, + 823.0, + 296.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1495.0, + 1405.0, + 1495.0, + 1405.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 776.0, + 1529.0, + 776.0, + 1562.0, + 294.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1140.0, + 1392.0, + 1140.0, + 1392.0, + 1182.0, + 307.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1839.0, + 830.0, + 1839.0, + 830.0, + 1881.0, + 295.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1615.0, + 820.0, + 1615.0, + 820.0, + 1652.0, + 294.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1672.0, + 1406.0, + 1672.0, + 1406.0, + 1709.0, + 293.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1705.0, + 1405.0, + 1705.0, + 1405.0, + 1738.0, + 295.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1737.0, + 1404.0, + 1737.0, + 1404.0, + 1769.0, + 295.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1766.0, + 1217.0, + 1766.0, + 1217.0, + 1801.0, + 293.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1404.0, + 1909.0, + 1404.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1975.0, + 1404.0, + 1975.0, + 1404.0, + 2007.0, + 297.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 582.0, + 261.0, + 582.0, + 297.0, + 296.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1615.0, + 820.0, + 1615.0, + 820.0, + 1652.0, + 294.0, + 1652.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 389, + 311, + 1303, + 311, + 1303, + 621, + 389, + 621 + ], + "score": 0.983, + "html": "
OptimizerTest regret
Search space 1Search Space 2
DARTSDefault config.1.252e-2 ± 0.0 0.864e-2 ± 0.023e-2
Transferred config. 1.045e-2 ± 0.171e-20.992e-2 ± 0.211e-2
GDASDefault config.1.252e-2 ± 0.0 0.871e-2 ± 0.0
Transferred config. 3.093e-2 ± 3.975e-20.831e-2 ± 0.045e-2
PC-DARTSDefault config.1.104e-2 ± 0.088e-2 1.133e-2 ± 0.404e-2
Transferred config. 5.843e-2 ± 4.118e-20.992e-2 ± 0.087e-2
" + }, + { + "category_id": 3, + "poly": [ + 324, + 950, + 1407, + 950, + 1407, + 1891, + 324, + 1891 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 785, + 1405, + 785, + 1405, + 880, + 297, + 880 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 302, + 707, + 1401, + 707, + 1401, + 770, + 302, + 770 + ], + "score": 0.944 + }, + { + "category_id": 6, + "poly": [ + 297, + 223, + 1403, + 223, + 1403, + 287, + 297, + 287 + ], + "score": 0.926 + }, + { + "category_id": 4, + "poly": [ + 296, + 1914, + 1399, + 1914, + 1399, + 1977, + 296, + 1977 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.852 + }, + { + "category_id": 13, + "poly": [ + 998, + 788, + 1033, + 788, + 1033, + 817, + 998, + 817 + ], + "score": 0.88, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1136, + 1916, + 1170, + 1916, + 1170, + 1945, + 1136, + 1945 + ], + "score": 0.87, + "latex": "L _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 990.0, + 592.0, + 990.0, + 592.0, + 1006.0, + 560.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 993.0, + 955.0, + 993.0, + 955.0, + 1010.0, + 925.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1011.0, + 593.0, + 1011.0, + 593.0, + 1051.0, + 559.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1015.0, + 670.0, + 1015.0, + 670.0, + 1050.0, + 636.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1011.0, + 954.0, + 1011.0, + 954.0, + 1054.0, + 925.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1028.0, + 1029.0, + 1028.0, + 1029.0, + 1041.0, + 1004.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1005.0, + 1321.0, + 1005.0, + 1321.0, + 1055.0, + 1296.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1021.0, + 1398.0, + 1021.0, + 1398.0, + 1039.0, + 1372.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1049.0, + 671.0, + 1049.0, + 671.0, + 1070.0, + 636.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1053.0, + 1032.0, + 1053.0, + 1032.0, + 1070.0, + 1001.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1054.0, + 1398.0, + 1054.0, + 1398.0, + 1071.0, + 1371.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1073.0, + 591.0, + 1073.0, + 591.0, + 1133.0, + 552.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1069.0, + 671.0, + 1069.0, + 671.0, + 1123.0, + 634.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1068.0, + 953.0, + 1068.0, + 953.0, + 1129.0, + 918.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1078.0, + 1033.0, + 1078.0, + 1033.0, + 1121.0, + 999.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1069.0, + 1320.0, + 1069.0, + 1320.0, + 1119.0, + 1292.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1086.0, + 1398.0, + 1086.0, + 1398.0, + 1104.0, + 1371.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1121.0, + 670.0, + 1121.0, + 670.0, + 1139.0, + 638.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1119.0, + 1398.0, + 1119.0, + 1398.0, + 1136.0, + 1371.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1134.0, + 572.0, + 1134.0, + 572.0, + 1149.0, + 551.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1134.0, + 936.0, + 1134.0, + 936.0, + 1149.0, + 915.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1133.0, + 1032.0, + 1133.0, + 1032.0, + 1150.0, + 1000.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1133.0, + 1313.0, + 1133.0, + 1313.0, + 1150.0, + 1276.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1145.0, + 569.0, + 1145.0, + 569.0, + 1197.0, + 519.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1162.0, + 720.0, + 1162.0, + 720.0, + 1180.0, + 698.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1146.0, + 922.0, + 1146.0, + 922.0, + 1173.0, + 891.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1163.0, + 1089.0, + 1163.0, + 1089.0, + 1178.0, + 1069.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1145.0, + 1299.0, + 1145.0, + 1299.0, + 1175.0, + 1252.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1172.0, + 536.0, + 1172.0, + 536.0, + 1199.0, + 502.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1184.0, + 638.0, + 1184.0, + 638.0, + 1199.0, + 618.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1172.0, + 741.0, + 1172.0, + 741.0, + 1187.0, + 721.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1177.0, + 767.0, + 1177.0, + 767.0, + 1195.0, + 744.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1187.0, + 790.0, + 1187.0, + 790.0, + 1202.0, + 769.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1172.0, + 896.0, + 1172.0, + 896.0, + 1199.0, + 864.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1184.0, + 1001.0, + 1184.0, + 1001.0, + 1199.0, + 981.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1169.0, + 1161.0, + 1169.0, + 1161.0, + 1203.0, + 1091.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1170.0, + 1275.0, + 1170.0, + 1275.0, + 1202.0, + 1224.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1185.0, + 1371.0, + 1185.0, + 1371.0, + 1199.0, + 1352.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1198.0, + 507.0, + 1198.0, + 507.0, + 1213.0, + 486.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1194.0, + 816.0, + 1194.0, + 816.0, + 1209.0, + 795.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1199.0, + 843.0, + 1199.0, + 843.0, + 1218.0, + 819.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1198.0, + 870.0, + 1198.0, + 870.0, + 1213.0, + 849.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1194.0, + 1186.0, + 1194.0, + 1186.0, + 1209.0, + 1165.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1224.0, + 630.0, + 1224.0, + 630.0, + 1260.0, + 339.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1224.0, + 992.0, + 1224.0, + 992.0, + 1259.0, + 700.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1224.0, + 1362.0, + 1224.0, + 1362.0, + 1258.0, + 1070.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1309.0, + 594.0, + 1309.0, + 594.0, + 1350.0, + 557.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1323.0, + 671.0, + 1323.0, + 671.0, + 1344.0, + 634.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1317.0, + 958.0, + 1317.0, + 958.0, + 1350.0, + 925.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1330.0, + 1032.0, + 1330.0, + 1032.0, + 1349.0, + 1001.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1305.0, + 1323.0, + 1305.0, + 1323.0, + 1350.0, + 1293.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1323.0, + 1403.0, + 1323.0, + 1403.0, + 1344.0, + 1368.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1353.0, + 603.0, + 1353.0, + 603.0, + 1434.0, + 552.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1343.0, + 671.0, + 1343.0, + 671.0, + 1434.0, + 635.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1348.0, + 967.0, + 1348.0, + 967.0, + 1428.0, + 918.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1348.0, + 1034.0, + 1348.0, + 1034.0, + 1425.0, + 999.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1387.0, + 1319.0, + 1387.0, + 1319.0, + 1431.0, + 1292.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1345.0, + 1404.0, + 1345.0, + 1404.0, + 1424.0, + 1368.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1423.0, + 953.0, + 1423.0, + 953.0, + 1441.0, + 918.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1423.0, + 1032.0, + 1423.0, + 1032.0, + 1441.0, + 999.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1423.0, + 1404.0, + 1423.0, + 1404.0, + 1444.0, + 1368.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1449.0, + 573.0, + 1449.0, + 573.0, + 1468.0, + 550.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1432.0, + 671.0, + 1432.0, + 671.0, + 1471.0, + 634.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1449.0, + 943.0, + 1449.0, + 943.0, + 1467.0, + 905.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1441.0, + 1032.0, + 1441.0, + 1032.0, + 1458.0, + 1001.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1449.0, + 1307.0, + 1449.0, + 1307.0, + 1468.0, + 1284.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1444.0, + 1402.0, + 1444.0, + 1402.0, + 1462.0, + 1371.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1463.0, + 561.0, + 1463.0, + 561.0, + 1478.0, + 540.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1478.0, + 721.0, + 1478.0, + 721.0, + 1497.0, + 697.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1462.0, + 938.0, + 1462.0, + 938.0, + 1521.0, + 858.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 1463.0, + 1299.0, + 1463.0, + 1299.0, + 1511.0, + 1255.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 1489.0, + 534.0, + 1489.0, + 534.0, + 1516.0, + 501.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1501.0, + 638.0, + 1501.0, + 638.0, + 1516.0, + 618.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1486.0, + 768.0, + 1486.0, + 768.0, + 1512.0, + 720.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1503.0, + 790.0, + 1503.0, + 790.0, + 1518.0, + 769.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1501.0, + 893.0, + 1501.0, + 893.0, + 1517.0, + 853.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1501.0, + 1001.0, + 1501.0, + 1001.0, + 1516.0, + 981.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1489.0, + 1268.0, + 1489.0, + 1268.0, + 1516.0, + 1234.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 1503.0, + 1368.0, + 1503.0, + 1368.0, + 1513.0, + 1355.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1516.0, + 507.0, + 1516.0, + 507.0, + 1534.0, + 450.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1511.0, + 816.0, + 1511.0, + 816.0, + 1526.0, + 795.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1516.0, + 1240.0, + 1516.0, + 1240.0, + 1534.0, + 1183.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1537.0, + 630.0, + 1537.0, + 630.0, + 1576.0, + 338.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1539.0, + 992.0, + 1539.0, + 992.0, + 1573.0, + 700.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1538.0, + 1361.0, + 1538.0, + 1361.0, + 1573.0, + 1070.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1622.0, + 954.0, + 1622.0, + 954.0, + 1641.0, + 923.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1633.0, + 589.0, + 1633.0, + 589.0, + 1651.0, + 564.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1637.0, + 1325.0, + 1637.0, + 1325.0, + 1654.0, + 1294.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1638.0, + 1400.0, + 1638.0, + 1400.0, + 1652.0, + 1372.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1652.0, + 589.0, + 1652.0, + 589.0, + 1688.0, + 561.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1647.0, + 666.0, + 1647.0, + 666.0, + 1688.0, + 636.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1644.0, + 952.0, + 1644.0, + 952.0, + 1685.0, + 925.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1659.0, + 1028.0, + 1659.0, + 1028.0, + 1677.0, + 1001.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1658.0, + 1324.0, + 1658.0, + 1324.0, + 1676.0, + 1294.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1663.0, + 1400.0, + 1663.0, + 1400.0, + 1677.0, + 1372.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1706.0, + 587.0, + 1706.0, + 587.0, + 1741.0, + 560.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1693.0, + 666.0, + 1693.0, + 666.0, + 1735.0, + 636.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1683.0, + 964.0, + 1683.0, + 964.0, + 1726.0, + 926.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1690.0, + 1028.0, + 1690.0, + 1028.0, + 1735.0, + 1001.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1678.0, + 1336.0, + 1678.0, + 1336.0, + 1735.0, + 1292.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1686.0, + 1402.0, + 1686.0, + 1402.0, + 1728.0, + 1371.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1733.0, + 428.0, + 1733.0, + 428.0, + 1743.0, + 419.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1740.0, + 585.0, + 1740.0, + 585.0, + 1758.0, + 559.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1740.0, + 666.0, + 1740.0, + 666.0, + 1758.0, + 636.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1732.0, + 946.0, + 1732.0, + 946.0, + 1747.0, + 925.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 1738.0, + 1321.0, + 1738.0, + 1321.0, + 1756.0, + 1289.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 1735.0, + 1403.0, + 1735.0, + 1403.0, + 1752.0, + 1370.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1750.0, + 1027.0, + 1750.0, + 1027.0, + 1767.0, + 1001.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 1766.0, + 580.0, + 1766.0, + 580.0, + 1783.0, + 544.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1762.0, + 665.0, + 1762.0, + 665.0, + 1781.0, + 638.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1767.0, + 935.0, + 1767.0, + 935.0, + 1782.0, + 916.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1767.0, + 1305.0, + 1767.0, + 1305.0, + 1782.0, + 1286.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 1758.0, + 1404.0, + 1758.0, + 1404.0, + 1780.0, + 1370.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1796.0, + 355.0, + 1796.0, + 355.0, + 1811.0, + 336.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1777.0, + 573.0, + 1777.0, + 573.0, + 1832.0, + 497.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1780.0, + 922.0, + 1780.0, + 922.0, + 1795.0, + 904.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 1780.0, + 1292.0, + 1780.0, + 1292.0, + 1806.0, + 1261.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1805.0, + 898.0, + 1805.0, + 898.0, + 1820.0, + 877.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 1805.0, + 1268.0, + 1805.0, + 1268.0, + 1820.0, + 1248.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1812.0, + 403.0, + 1812.0, + 403.0, + 1826.0, + 384.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1817.0, + 428.0, + 1817.0, + 428.0, + 1836.0, + 407.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1814.0, + 533.0, + 1814.0, + 533.0, + 1835.0, + 493.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1817.0, + 638.0, + 1817.0, + 638.0, + 1832.0, + 618.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1817.0, + 1001.0, + 1817.0, + 1001.0, + 1832.0, + 981.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1818.0, + 1371.0, + 1818.0, + 1371.0, + 1832.0, + 1352.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1827.0, + 453.0, + 1827.0, + 453.0, + 1842.0, + 433.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1830.0, + 513.0, + 1830.0, + 513.0, + 1850.0, + 458.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1831.0, + 869.0, + 1831.0, + 869.0, + 1846.0, + 848.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1830.0, + 1241.0, + 1830.0, + 1241.0, + 1847.0, + 1217.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1856.0, + 635.0, + 1856.0, + 635.0, + 1891.0, + 332.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1856.0, + 996.0, + 1856.0, + 996.0, + 1892.0, + 697.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1857.0, + 1366.0, + 1857.0, + 1366.0, + 1892.0, + 1068.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.25, + 1047.0, + 607.25, + 1047.0, + 607.25, + 1073.0, + 553.25, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.25, + 1047.0, + 971.25, + 1047.0, + 971.25, + 1073.0, + 915.25, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1048.0, + 1339.0, + 1048.0, + 1339.0, + 1071.0, + 1290.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1161.5, + 384.0, + 1161.5, + 384.0, + 1186.5, + 328.0, + 1186.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.25, + 1176.0, + 434.25, + 1176.0, + 434.25, + 1202.5, + 376.25, + 1202.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.25, + 1191.0, + 428.25, + 1191.0, + 428.25, + 1211.0, + 371.25, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.75, + 1190.0, + 487.75, + 1190.0, + 487.75, + 1219.0, + 422.75, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1189.5, + 787.0, + 1189.5, + 787.0, + 1211.0, + 737.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.75, + 1168.0, + 929.75, + 1168.0, + 929.75, + 1181.5, + 888.75, + 1181.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.75, + 1189.5, + 1156.75, + 1189.5, + 1156.75, + 1211.0, + 1106.75, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.75, + 1194.0, + 1250.75, + 1194.0, + 1250.75, + 1218.0, + 1184.75, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1290.5, + 816.0, + 1290.5, + 816.0, + 1334.5, + 710.0, + 1334.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1359.0, + 1334.0, + 1359.0, + 1334.0, + 1380.5, + 1295.0, + 1380.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.75, + 1486.5, + 460.75, + 1486.5, + 460.75, + 1518.5, + 325.75, + 1518.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1486.5, + 1195.0, + 1486.5, + 1195.0, + 1518.5, + 1060.0, + 1518.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.25, + 1507.5, + 427.25, + 1507.5, + 427.25, + 1527.5, + 373.25, + 1527.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1505.5, + 789.0, + 1505.5, + 789.0, + 1528.0, + 736.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1504.5, + 1160.0, + 1504.5, + 1160.0, + 1529.5, + 1105.0, + 1529.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.25, + 1500.0, + 541.25, + 1500.0, + 541.25, + 1517.5, + 513.25, + 1517.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 1512.0, + 878.0, + 1512.0, + 878.0, + 1534.0, + 815.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1680.0, + 608.0, + 1680.0, + 608.0, + 1706.5, + 554.0, + 1706.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.75, + 1731.0, + 903.75, + 1731.0, + 903.75, + 1753.0, + 866.75, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1799.5, + 795.0, + 1799.5, + 795.0, + 1828.0, + 690.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1800.0, + 1166.0, + 1800.0, + 1166.0, + 1828.0, + 1060.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1802.0, + 378.0, + 1802.0, + 378.0, + 1820.0, + 359.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1822.5, + 425.0, + 1822.5, + 425.0, + 1843.5, + 373.0, + 1843.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1822.0, + 790.0, + 1822.0, + 790.0, + 1843.5, + 734.0, + 1843.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1822.5, + 1161.0, + 1822.5, + 1161.0, + 1844.0, + 1105.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1824.5, + 849.0, + 1824.5, + 849.0, + 1851.5, + 787.0, + 1851.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1792.0, + 944.0, + 1792.0, + 944.0, + 1831.5, + 843.0, + 1831.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.25, + 1823.5, + 1219.25, + 1823.5, + 1219.25, + 1852.5, + 1157.25, + 1852.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 223.0, + 1403.0, + 223.0, + 1403.0, + 259.0, + 295.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 254.0, + 783.0, + 254.0, + 783.0, + 290.0, + 295.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1135.0, + 1912.0, + 1135.0, + 1948.0, + 295.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1912.0, + 1401.0, + 1912.0, + 1401.0, + 1948.0, + 1171.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1944.0, + 918.0, + 1944.0, + 918.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 784.0, + 997.0, + 784.0, + 997.0, + 823.0, + 294.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 784.0, + 1403.0, + 784.0, + 1403.0, + 823.0, + 1034.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 814.0, + 1405.0, + 814.0, + 1405.0, + 851.0, + 294.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 847.0, + 1134.0, + 847.0, + 1134.0, + 881.0, + 295.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 705.0, + 1402.0, + 705.0, + 1402.0, + 746.0, + 293.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 739.0, + 386.0, + 739.0, + 386.0, + 773.0, + 296.0, + 773.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 564, + 1404, + 564, + 1404, + 839, + 296, + 839 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 335, + 1404, + 335, + 1404, + 548, + 297, + 548 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 306, + 869, + 1394, + 869, + 1394, + 1181, + 306, + 1181 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 295, + 1204, + 1403, + 1204, + 1403, + 1269, + 295, + 1269 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 866, + 2088, + 866, + 2112, + 835, + 2112 + ], + "score": 0.846 + }, + { + "category_id": 0, + "poly": [ + 298, + 229, + 1377, + 229, + 1377, + 300, + 298, + 300 + ], + "score": 0.6 + }, + { + "category_id": 15, + "poly": [ + 318.0, + 869.0, + 359.0, + 869.0, + 359.0, + 887.0, + 318.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 880.0, + 646.0, + 880.0, + 646.0, + 896.0, + 531.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 876.0, + 1008.0, + 876.0, + 1008.0, + 892.0, + 895.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 874.0, + 1380.0, + 874.0, + 1380.0, + 891.0, + 1265.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 890.0, + 647.0, + 890.0, + 647.0, + 907.0, + 532.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 887.0, + 1008.0, + 887.0, + 1008.0, + 904.0, + 895.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 887.0, + 1380.0, + 887.0, + 1380.0, + 904.0, + 1265.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 900.0, + 650.0, + 900.0, + 650.0, + 930.0, + 531.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 904.0, + 723.0, + 904.0, + 723.0, + 922.0, + 680.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 899.0, + 1013.0, + 899.0, + 1013.0, + 927.0, + 895.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 899.0, + 1383.0, + 899.0, + 1383.0, + 927.0, + 1265.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 932.0, + 359.0, + 932.0, + 359.0, + 1048.0, + 306.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 925.0, + 651.0, + 925.0, + 651.0, + 941.0, + 532.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 922.0, + 1013.0, + 922.0, + 1013.0, + 939.0, + 895.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 930.0, + 1059.0, + 930.0, + 1059.0, + 1047.0, + 1039.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 922.0, + 1385.0, + 922.0, + 1385.0, + 939.0, + 1265.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 973.0, + 724.0, + 973.0, + 724.0, + 995.0, + 680.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 993.0, + 1200.0, + 993.0, + 1200.0, + 1007.0, + 1178.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1002.0, + 356.0, + 1002.0, + 356.0, + 1020.0, + 319.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1019.0, + 424.0, + 1019.0, + 424.0, + 1034.0, + 406.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1013.0, + 1078.0, + 1013.0, + 1078.0, + 1033.0, + 1052.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 1025.0, + 634.0, + 1025.0, + 634.0, + 1083.0, + 531.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1046.0, + 721.0, + 1046.0, + 721.0, + 1065.0, + 695.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1049.0, + 1209.0, + 1049.0, + 1209.0, + 1061.0, + 1199.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1080.0, + 357.0, + 1080.0, + 357.0, + 1098.0, + 318.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1104.0, + 382.0, + 1104.0, + 382.0, + 1119.0, + 366.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1104.0, + 438.0, + 1104.0, + 438.0, + 1120.0, + 421.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1104.0, + 492.0, + 1104.0, + 492.0, + 1119.0, + 474.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1104.0, + 547.0, + 1104.0, + 547.0, + 1119.0, + 529.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1104.0, + 602.0, + 1104.0, + 602.0, + 1119.0, + 584.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1104.0, + 656.0, + 1104.0, + 656.0, + 1119.0, + 638.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1104.0, + 745.0, + 1104.0, + 745.0, + 1119.0, + 730.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1104.0, + 802.0, + 1104.0, + 802.0, + 1119.0, + 782.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1104.0, + 855.0, + 1104.0, + 855.0, + 1119.0, + 836.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1104.0, + 911.0, + 1104.0, + 911.0, + 1119.0, + 892.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1104.0, + 965.0, + 1104.0, + 965.0, + 1119.0, + 947.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1104.0, + 1020.0, + 1104.0, + 1020.0, + 1119.0, + 1001.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1104.0, + 1102.0, + 1104.0, + 1102.0, + 1120.0, + 1086.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1102.0, + 1163.0, + 1102.0, + 1163.0, + 1122.0, + 1142.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1102.0, + 1219.0, + 1102.0, + 1219.0, + 1122.0, + 1198.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1102.0, + 1277.0, + 1102.0, + 1277.0, + 1122.0, + 1256.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1102.0, + 1334.0, + 1102.0, + 1334.0, + 1122.0, + 1313.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1104.0, + 1390.0, + 1104.0, + 1390.0, + 1119.0, + 1371.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1113.0, + 525.0, + 1113.0, + 525.0, + 1131.0, + 492.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1114.0, + 887.0, + 1114.0, + 887.0, + 1131.0, + 854.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1113.0, + 1251.0, + 1113.0, + 1251.0, + 1131.0, + 1218.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1146.0, + 623.0, + 1146.0, + 623.0, + 1179.0, + 343.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 1145.0, + 991.0, + 1145.0, + 991.0, + 1181.0, + 700.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1147.0, + 1360.0, + 1147.0, + 1360.0, + 1178.0, + 1071.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 959.0, + 469.0, + 959.0, + 469.0, + 1015.0, + 394.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 978.0, + 1141.0, + 978.0, + 1141.0, + 993.5, + 1135.0, + 993.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 993.0, + 776.0, + 993.0, + 776.0, + 1009.0, + 764.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1003.0, + 1377.0, + 1003.0, + 1377.0, + 1057.0, + 1142.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.25, + 1012.0, + 807.25, + 1012.0, + 807.25, + 1030.5, + 782.25, + 1030.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.25, + 1050.5, + 1323.25, + 1050.5, + 1323.25, + 1094.5, + 1228.25, + 1094.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1203.0, + 1404.0, + 1203.0, + 1404.0, + 1235.0, + 297.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1234.0, + 942.0, + 1234.0, + 942.0, + 1270.0, + 294.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 872.0, + 2084.0, + 872.0, + 2123.0, + 829.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 1380.0, + 227.0, + 1380.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 271.0, + 884.0, + 271.0, + 884.0, + 305.0, + 356.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 563.0, + 1406.0, + 563.0, + 1406.0, + 600.0, + 294.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 593.0, + 1406.0, + 593.0, + 1406.0, + 631.0, + 293.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 623.0, + 1405.0, + 623.0, + 1405.0, + 660.0, + 294.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 653.0, + 1406.0, + 653.0, + 1406.0, + 692.0, + 293.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 688.0, + 1405.0, + 688.0, + 1405.0, + 721.0, + 295.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 716.0, + 1406.0, + 716.0, + 1406.0, + 751.0, + 293.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 747.0, + 1406.0, + 747.0, + 1406.0, + 782.0, + 293.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 776.0, + 1405.0, + 776.0, + 1405.0, + 811.0, + 291.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 808.0, + 394.0, + 808.0, + 394.0, + 842.0, + 294.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 335.0, + 1405.0, + 335.0, + 1405.0, + 369.0, + 295.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 364.0, + 1406.0, + 364.0, + 1406.0, + 400.0, + 294.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 394.0, + 1404.0, + 394.0, + 1404.0, + 431.0, + 292.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 427.0, + 1405.0, + 427.0, + 1405.0, + 461.0, + 295.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 456.0, + 1404.0, + 456.0, + 1404.0, + 489.0, + 292.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 487.0, + 1404.0, + 487.0, + 1404.0, + 525.0, + 291.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 519.0, + 392.0, + 519.0, + 392.0, + 550.0, + 294.0, + 550.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/Uq_tGs7N54M/Uq_tGs7N54M.md b/parse/train/Uq_tGs7N54M/Uq_tGs7N54M.md new file mode 100644 index 0000000000000000000000000000000000000000..71bc0d79218c897de3535217f73a8d44d807c7f8 --- /dev/null +++ b/parse/train/Uq_tGs7N54M/Uq_tGs7N54M.md @@ -0,0 +1,301 @@ +# Robust Counterfactual Explanations on Graph Neural Networks + +# Mohit Bajaj1\* Lingyang $\mathbf { C h u ^ { 2 * } }$ Zi Yu Xue1,3Jian Pei4 Lanjun Wang1Peter Cho-Ho Lam1Yong Zhang1 + +1Huawei Technologies Canada Co., Ltd. 2McMaster University 3 The University of British Columbia 4Simon Fraser University {mohit.bajaj1，zi.yu.xue，lanjun.wang，cho.ho.lam，yong.zhang3}@huawei.com chul9@mcmaster.ca,jpei@cs.sfu.ca + +# Abstract + +Massive deployment of Graph Neural Networks (GNNs) in high-stake applications generates a strong demand for explanations that are robust to noise and align well with human intuition. Most existing methods generate explanations by identifying a subgraph of an input graph that has a strong correlation with the prediction. These explanations are not robust to noise because independently optimizing the correlation for a single input can easily overfit noise.Moreover, they are not counterfactual because removing an identified subgraph from an input graph does not necessarily change the prediction result. In this paper, we propose a novel method to generate robust counterfactual explanations on GNNs by explicitly modelling the common decision logic of GNNs on similar input graphs. Our explanations are naturally robust to noise because they are produced from the common decision boundaries of a GNN that govern the predictions of many similar input graphs. The explanations are also counterfactual because removing the set of edges identified by an explanation from the input graph changes the prediction significantly. Exhaustive experiments on many public datasets demonstrate the superior performance of our method. + +# 1 Introduction + +Graph Neural Networks (GNNs) [22,37,50] have achieved great practical successes in many realworld applications,such as chemistry [31],molecular biology[17],social networks [3] and epidemic modelling [34]. For most of these applications, explaining predictions made by a GNN model is crucial for establishing trust with end-users,identifying the cause of a prediction,and even discovering potential deficiencies of a GNN model before massive deployment. Ideally,an explanation should be able to answer questions like “Would the prediction of the GNN model change if a certain part of an input molecule is removed?” in the context of predicting whether an artificial molecule is active for acertain type of proteins [19,41],“Would an item recommended stillbe recommended if a customer had not purchased some other items in the past?” for a GNN built for recommendation systems [9, 44]. + +Counterfactual explanations [28] in the form of “If X had not occurred, Y would not have occurred”[26] are the principled way to answer such questions and thus are highly desirable for GNNs. In the context of GNNs,a counterfactual explanation identifies a small subset of edges of the input graph instance such that removing those edges significantly changes the prediction made by the GNN. Counterfactual explanations are usually concise and easy to understand [28,36] because they align well with the human intuition to describe a causal situation [26]. To make explanations more trustworthy, the counterfactual explanation should be robust to noise,that is,some slight changes on an input graph do not change the explanation significantly. This idea aligns well with the notion of robustness discussed for DNN explanations in computer vision domain [11]. According to Ghorbani et al.[11] many interpretations on neural networks are fragile as it is easier to generate adversarial perturbations that produce perceptively indistinguishable inputs that are assigned the same predicted label,yet have very different interpretations.Here,the concepts of “fragile”“robustness”describe the same concept from opposite perspectives. An interpretation is said to be fragile if systematic perturbations can lead to dramatically different interpretations without changing the label. Otherwise, the interpretation is said to be robust. + +How to produce robust counterfactual explanations on predictions made by general graph neural networks is a novel problem that has not been systematically studied before.As to be discussed in Section 2, most GNN explanation methods [45,25,46,37,32] are neither counterfactual nor robust. These methods mostly focus on identifying a subgraph of an input graph that achieves a high correlation with the prediction result. Such explanations are usually not counterfactual because, due to the high non-convexity of GNNs, removing a subgraph that achieves a high correlation does not necessarily change the prediction result. Moreover, many existing methods [45,25,37,32] are not robust to noise and may change significantly upon slight modifications on input graphs, because the explanation of every single input graph prediction is independently optimized to maximize the correlation with the prediction, thus an explanation can easily overfit the noise in the data. + +In this paper², we develop RCExplainer, a novel method to produce robust counterfactual explanations on GNNs. The key idea is to first model the common decision logic of a GNN by set of decision regions where each decision region governs the predictions on a large number of graphs,and then extract robust counterfactual explanations by a deep neural network that explores the decision logic carried by the linear decision boundaries of the decision regions. We make the following contributions. + +First, we model the decision logic of a GNN by a set of decision regions, where each decision region is induced by a set of linear decision boundaries of the GNN. We propose an unsupervised method to find decision regions for each class such that each decision region governs the prediction of multiple graph samples predicted to be the same class. The linear decision boundaries of the decision region capture the common decision logic on all the graph instances inside the decision region, thus do not easily overfit the noise of an individual graph instance. By exploring the common decision logic encoded in the linear boundaries, we are able to produce counterfactual explanations that are inherently robust to noise. + +Second, based on the linear boundaries of the decision region,we propose a novel loss function to train a neural network that produces a robust counterfactual explanation as a small subset of edges of an input graph. The loss function is designed to directly optimize the explainability and counterfactual property of the subset of edges,such that: 1) the subgraph induced by the edges lies within the decision region,thus has a prediction consistent with the input graph; and 2) deleting the subset of edges from the input graph produces a remainder subgraph that lies outside the decision region, thus the prediction on the remainder subgraph changes significantly. + +Last, we conduct comprehensive experimental study to compare our method with the state-of-the-art methods on fidelity, robustness,accuracy and efficiency. All the results solidly demonstrate the superior performance of our approach. + +# 2Related work + +The existing GNN explanation methods [46,37, 45,32, 25] generally fall into two categories: model level explanation [46] and instance level explanation [37,45,32, 25]. + +A model level explanation method [46] produces a high-level explanation about the general behaviors of a GNN independent from input examples. This may be achieved by synthesizing a set of artificial graph instances such that each artificial graph instance maximizes the prediction score on a certain class. The weakness of model level explanation methods is that an input graph instance may not contain an artificial graph instance,and removing an artificial graph from an input graph does not necessarily change the prediction. As a result, model level explanations are substantially different from counterfactual explanations, because the synthesized artificial graphs do not provide insights into how the GNN makes its prediction on a specific input graph instance. + +The instance level explanation methods [37,45,32, 25] explain the prediction(s) made by a GNN on a specific input graph instance or multiple instances by identifying a subgraph of an input graph instance that achieves a high correlation with the prediction on the input graph. GNNExplainer [45] removes redundant edges from an input graph instance to produce an explanation that maximizes the mutual information between the distribution of subgraphs of the input graph and the GNN's prediction. Following the same idea by Ying et al. [45],PGExplainer [25] parameterizes the generation process of explanations by a deep neural network,and trains it to maximize a similar mutual information based loss used by GNNExplainer [45]. The trained deep neural network is then applied to generate explanations for a single input graph instance or a group of input graphs. MEG [30] incorporates strong domain knowledge in chemistry with a reinforcement learning framework to produce counterfactual explanations on GNNs specifically built for compound prediction, but the heavy reliance on domain knowledge largely limits its applicability on general GNNs. The recently proposed CF-GNNExplainer [24] independently optimizes the counterfactual property for each explanation but ignores the correlation between the prediction and the explanation. + +Some studies [32,37] also adapt the existing explanation methods of image-oriented deep neural networks to produce instance level explanations for GNNs.Pope et al.[32] extend several gradient based methods [33,35,49] to explain predictions made by GNNs. The explanations are prone to gradient saturation[12] and may also be misleading [1] due to the heavy reliance on noisy gradients. Velickovic et al.[37] extend the atention mechanism[7,8] to identify the nodes in an input graph that contribute the most to the prediction.This method has to retrain the GNN with the altered architecture and the inserted attntion layers.Thus,the explanations may not be faithful to the original GNN. + +Instance level explanations from most of the methods are usually not counterfactual because, due to the non-convexity of GNNs, removing an explanation subgraph from the input graph does not necessarily change the prediction result. Moreover, those methods [45,25,37,32, 24] are usually not robust to noise because the explanation of every single input graph prediction is independently optimized. Thus,an explanation can easily overfit the noise inside input graphs and may change significantly upon slight modifications on input graphs. + +To tackle the weaknesses in the existing methods, in this paper, we directly optimize the counterfactual property of an explanation along with the correlation between the explanation and the prediction. Our explanations are also much more robust to modifications on input graphs,because they are produced from the common decision logic on a large group of similar input graphs, which do not easily overfit the noise of an individual graph sample. + +Please note that our study is substantially different from adversarial attacks on GNNs.The adversarial attacking methods [51,53,42,43,20] use adversarial examples to change the predictions of GNNs but ignore the explainability of the generated adversarial examples [1O]. Thus, the adversarial examples generated by adversarial attcks may not explain the original prediction. + +Our method is substantially different from the above works because we focus on explaining the prediction by directly optimizing the counterfactual property of an explanation along with correlation of the explanation with the prediction. We also require that the explanation is generally valid for a large set of similar graph instances by extracting it from the common linear decision boundaries of a large decision region. + +# 3Problem Formulation + +Denote by $G = \{ V , E \}$ a graph where $V = \{ v _ { 1 } , v _ { 2 } , \ldots , v _ { n } \}$ is the set of $n$ nodes and $E \subseteq V \times V$ is the set of edges. The edge structure of a graph $G$ is described by an adjacency matrix $\mathbf { A } \in \{ 0 , 1 \} ^ { n \times n }$ where ${ \bf A } _ { i j } = 1$ if there is an edge between node $v _ { i }$ and $v _ { j }$ ; and ${ \bf A } _ { i j } = 0$ otherwise. + +Denote by $\phi$ a GNN model that maps a graph to a probability distribution over a set of classes denoted by $C$ . Let $D$ denote the set of graphs that are used to train the GNN model $\phi$ .We focus on GNNs that adopt piecewise linear activation functions,such as MaxOut[14] and the family of ReLU[13,15,29]. + +The robust counterfactual explanation problem is defined as follows. + +Definition 1 (Robust Counterfactual Explanation Problem) Given a GNN model $\phi$ trained on a set of graphs $D$ ， for an input graph $G = \{ V , E \}$ , our goal is to explain why $G$ is predicted by the GNN model as $\phi ( G )$ by identifying a small subset of edges $S \subseteq E$ ,such that $( l )$ removing the set of edges in $S$ from $G$ that causes the maximum drop in the confidence of the original prediction; and (2) $S$ is stable and doesn't change when the edges and the feature representations of the nodes of $G$ are perturbed by random noise. + +In the definition,the first requirement requires that the explanation $S$ is counterfactual, and the second requirement requires that the explanation is robust to noisy changes on the edges and nodes of $G$ ： + +# 4Method + +In this section, we first introduce how to extract the common decision logic of a GNN on a large set of graphs with the same predicted class. This is achieved by a decision region induced by a set of linear decision boundaries of the GNN.Then, based on the linear boundaries of the decision region, we propose a novel lossfunction to train a neural network that produces robust counterfactual explanations.Last, we discuss the time complexity of our method when generating explanations. + +# 4.1Modelling Decision Regions + +Following the routines of many deep neural network explanation methods [33, 48], we extract the decision region of a GNN in the $d$ -dimensional output space $\mathbb { O } ^ { d }$ of the last convolution layer of the GNN. The features generated by the last convolution layer are more conceptually meaningful and more robust to noise than those raw features of input graphs,such as vertices and edges [52, 2]. Denote by $\phi _ { g c }$ the mapping function realized by the graph convolution layers that maps an input graph $G$ to its graph embedding $\phi _ { g c } ( G ) \in \mathbb { O } ^ { d }$ ,and by $\phi _ { f c }$ the mapping function realized by the fully connected layers that maps the graph embedding $\phi _ { g c } ( G )$ to a predicted distribution over the classes in $C$ . The overall prediction $\phi ( { \bar { G } } )$ made by the GNN can be writtn as $\phi ( G ) = \phi _ { f c } ( \phi _ { g c } ( G ) )$ + +For the GNNs that adopt piecewise linear activation functions for the hidden neurons, such as MaxOut [14] and the family of ReLU [13,15,29], the decision logic of $\phi _ { f c }$ in the space $\mathbb { O } ^ { d }$ is characterized by a piecewise linear decision boundary formed by connected pieces of decision hyperplanes in $\mathbb { O } ^ { d }$ [1]. We call these hyperplanes linear decision boundaries (LDBs), and denote by $\mathcal { H }$ the set of LDBs induced by $\phi _ { f c }$ . The set of LDBs in $\mathcal { H }$ partitions the space $\mathbb { O } ^ { d }$ into a large number of convex polytopes. A convex polytope is formed by a subset of LDBs in $\mathcal { H }$ . All the graphs whose graph embeddings are contained in the same convex polytope are predicted as the same class [4]. Therefore, the LDBs of a convex polytope encode the common decision logic of $\phi _ { f c }$ on all the graphs whose graph embeddings lie within the convex polytope [4]. Here,a graph $G$ is covered by a convex polytope if the graph embedding $\phi _ { g c } ( G )$ is contained in the convex polytope. + +Based on the above insight, we model the decision region for a set of graph instances as a convex polytope that satisfies the following two properties. First, the decision region should be induced by a subset of the LDBs in $\mathcal { H }$ .In this way, when we extract counterfactual explanations from the LDBs, the explanations are loyal to the real decision logic_of the GNN. Second, the decision region should cover many graph instances in the training dataset $D$ ,and all the covered graphs should be predicted as the same class. In this way, the LDBs of the decision region capture the common decision logic on all the graphs covered by the decision region. Here, the requirement of covering a larger number of graphs ensures that the common decision logic is general,and thus it is lesslikely to overfit the noise of an individual graph instance. As a result, the counterfactual explanations extracted from the LDBs of the decision region are insensitive to slight changes in the input graphs. Our method can be easily generalized to incorporate prediction confidence in the coverage measure,such as considering the count of graphs weighted by prediction confidence. To keep our discusson simple, we do not pursue this detail further in the paper. + +Next, we illustrate how to extract a decision region satisfying the above two requirements. The key idea is to find a convex polytope covering a large set of graph instances in $D$ that are predicted as the same class $c \in C$ ： + +Denote by $D _ { c } \subseteq D$ the set of graphs in $D$ predicted as a class $c \in C$ ,by $\mathcal { P } \subseteq \mathcal { H }$ a set of LDBs that partition the space $\mathbb { O } ^ { d }$ into a set of convex polytopes, and by $r ( \mathcal { P } , c )$ the convex polytope induced by $\mathcal { P }$ that covers the largest number of graphs in $D _ { c }$ . Denote by $g ( \mathcal { P } , c )$ the number of graphs in $D _ { c }$ covered by $r ( \mathcal { P } , c )$ ,and by $h ( \mathcal { P } , c )$ the number of graphs in $D$ that are covered by $r ( \mathcal { P } , \bar { c } )$ but are not predicted as class $c$ . We extract a decision region covering a large set of graph instances in $D _ { c }$ by solving the following constrained optimization problem. + +$$ +\operatorname* { m a x } _ { \mathcal { P } \subseteq \mathcal { H } } g ( \mathcal { P } , c ) , \mathrm { ~ s . t . ~ } h ( \mathcal { P } , c ) = 0 +$$ + +This formulation realizes the two properties of decision regions because $\mathcal { P } \subseteq \mathcal { H }$ ensures that the decision region is induced by a subset of LDBs in $\mathcal { H }$ ,maximizing $g ( \mathcal { P } , c )$ requires that $r ( \mathcal { P } , c )$ covers + +a large number of graphs in $D _ { c }$ , and the constraint $h ( \mathcal { P } , c ) = 0$ ensures that allthe graphs covered by $r ( \mathcal { P } , c )$ are predicted as the same class $c$ ： + +Once we find a solution $\mathcal { P }$ to the above problem, the decision region $r ( \mathcal { P } , c )$ can be easily obtained by first counting the number of graphs in $D _ { c }$ covered by each convex polytope induced by $\mathcal { P }$ ,and then select the convex polytope that covers the largest number of graphs in $D _ { c }$ ： + +# 4.2Extracting Decision Regions + +The optimization problem in Equation (1) is intractable for standard GNNs, mainly because it is impractical to compute $\mathcal { H }$ , all the LDBs of a GNN. The number of LDBs in $\mathcal { H }$ of a GNN is exponential with respect to the number of neurons in the worst case [27].To address this challenge,we substitute $\mathcal { H }$ by a sample $\tilde { \mathcal { H } }$ of LDBs from $\tilde { \mathcal { H } }$ ： + +A LDB in the space $\mathbb { O } ^ { d }$ can be written as $\mathbf { w } ^ { \top } \mathbf { x } + b = 0$ ,where is $\mathbf { x } \in \mathbb { O } ^ { d }$ is a variable, w is the basis term, and $b$ corresponds to the bias.Following [4], for any input graph $G$ ,a linear boundary can be sampled from $\mathcal { H }$ by computing + +$$ +\mathbf { w } = \frac { \partial \left( \operatorname* { m a x } _ { 1 } ( \phi _ { f c } ( \alpha ) ) - \operatorname* { m a x } _ { 2 } ( \phi _ { f c } ( \alpha ) ) \right) } { \partial \alpha } | _ { \alpha = \phi _ { g c } ( G ) } , +$$ + +and + +$$ +b = \mathrm { m a x } _ { 1 } ( \phi _ { f c } ( \alpha ) ) - \mathrm { m a x } _ { 2 } ( \phi _ { f c } ( \alpha ) ) - \mathbf { w } ^ { T } { \alpha } | _ { \alpha = \phi _ { g c } ( G ) } , +$$ + +where $\operatorname* { m a x } _ { 1 } ( \phi _ { f c } ( \alpha ) ) )$ and $\operatorname* { m a x } _ { 2 } \bigl ( \phi _ { f c } ( \pmb { \alpha } ) \bigr )$ are the largest and the second largest values in the vector $\phi _ { f c } ( \alpha )$ ,respectively. Given an input graph $G$ ,Equations (2) and (3) identify one LDB from $\mathcal { H }$ .Thus, we can sample a subset of input graphs uniformly from $D$ ,and use Equations (2) and (3) to derive a sample of LDBs as $\tilde { \mathcal { H } } \subset \mathcal { H }$ ： + +Now, we substitute $\mathcal { H }$ in Equation (1) by $\tilde { \mathcal { H } }$ to produce the following problem. + +$$ +\operatorname* { m a x } _ { \mathcal { P } \subseteq \tilde { \mathcal { H } } } g ( \mathcal { P } , c ) , \mathrm { ~ s . t . ~ } h ( \mathcal { P } , c ) \leq \delta , +$$ + +where $\delta \geq 0$ is a tolerance parameter to keep this problem feasible. The parameter $\delta$ is required because substituting $\mathcal { H }$ by $\tilde { \mathcal { H } }$ ignores the LDBs in $\mathcal { H } \backslash \tilde { \mathcal { H } }$ . Thus, the convex polytope $r ( \mathcal { P } , c )$ induced by subset of boundaries in $\tilde { \mathcal { H } }$ may contain instances that are not predicted as class $c$ . We directly set $\boldsymbol { \delta } \dot { = } \boldsymbol { h } ( \tilde { \mathcal { H } } , \boldsymbol { c } )$ , which is the smallest value of $\delta$ that keeps the practical problem feasible. + +The problem in Equation (4) can be proven to be a Submodular Cost Submodular Cover (SCSC) problem [18] (see Appendix $\mathbf { D }$ for proof) that is well known to be NP-hard [5]. We adopt a greedy boundary selection method to find a good solution to this problem [40]. Specifically, we initialize $\mathcal { P }$ as an empty set,and then iteratively select a new boundary $h$ from $\tilde { \mathcal { H } }$ by + +$$ +h = \arg \operatorname* { m i n } _ { h \in \tilde { \mathcal { H } } \backslash \mathcal { P } } \frac { g ( \mathcal { P } , c ) - g ( \mathcal { P } \cup \{ h \} , c ) + \epsilon } { h ( \mathcal { P } , c ) - h ( \mathcal { P } \cup \{ h \} , c ) } , +$$ + +where $g ( \mathcal { P } , c ) - g ( \mathcal { P } \cup \{ h \} , c )$ is the decrease of $g ( \mathcal { P } , c )$ when adding $h$ into $\mathcal { P }$ ,and $h ( \mathcal { P } , c ) - h ( \mathcal { P } \cup$ $\{ h \} , c \bar { ) }$ is the decrease of $h ( \mathcal { P } , c )$ when adding $h$ into $\mathcal { P }$ .Both $g ( \mathcal { P } , c )$ and $h ( \mathcal { P } , c )$ are non-increasing when adding $h \in \tilde { \mathcal { H } }$ into $\mathcal { P }$ because adding a new boundary $h$ may only exclude some graphs from the convex polytope $r ( \mathcal { P } , c )$ ： + +Intuitively, in each iteration,Equation (5) selects a boundary $h \in \tilde { \mathcal { H } }$ such that adding $h$ into $\mathcal { P }$ reduces $g ( \mathcal { P } , c )$ the least and reduces $\bar { h } ( \mathcal { P } , c )$ the most. In this way,we can quickly reduce $h ( \mathcal { P } , c )$ to be smaller than $\delta$ without decreasing $g ( \mathcal { P } , c )$ too much, which produces a good feasible solution to the practical problem.We add a small constant $\epsilon$ to the numerator such that, when there are multiple candidates of $h$ that do not decrease $g ( \mathcal { P } , c )$ , we can still select the $h$ that reduces $h ( \mathcal { P } , c )$ the most. + +We apply a peeling-off strategy to iteratively extract multiple decision regions.For each class $c \in C$ we first solve the practical problem once to find a decision region $r ( \mathcal { P } , c )$ , then we remove the graphs covered by $r ( \mathcal { P } , \bar { c } )$ from $D _ { c }$ . If there are remaining graphs predicted as the class $c$ , we continue finding the decision regions using the remaining graphs until all the graphs in $D _ { c }$ are removed. When all the graphs in $D _ { c }$ are removed for each class $c \in C$ , we stop the iteration and return the set of decision regions we found. + +# 4.3Producing Explanations + +In this section, we introduce how to use the LDBs of decision regions to train a neural network that produces a robust counterfactual explanation as a small subset of edges of an input graph. We form explanations as a subset of edges because GNNs make decisions by aggregating messages passed on edges. Using edges instead of vertices as explanations can provide beter insights on the decision logic of GNNs. + +# 4.3.1The Neural Network Model + +Denote by $f _ { \theta }$ the neural network to generate a subset of edges of an input graph $G$ as the robust counterfactual explanation on the prediction $\phi ( G )$ ： $\theta$ represents the set of parameters of the neural network. For experiments, our explanation network $f$ consists of 2 fully connected layers with a ReLU activation and the hidden dimension of 64. + +For any two connected vertices $v _ { i }$ and $v _ { j }$ of $G$ ,denote by $\mathbf { z } _ { i }$ and $\mathbf { z } _ { j }$ the embeddings produced by the last convolution layer of the GNN for the two vertices, respectively. The neural network $f _ { \theta }$ takes $\mathbf { z } _ { i }$ and $\mathbf { z } _ { j }$ as the input and outputs the probability for the edge between $v _ { i }$ and $v _ { j }$ to be part of the explanation. This can be written as + +$$ +\begin{array} { r } { { \bf M } _ { i j } = f _ { \theta } ( { \bf z } _ { i } , { \bf z } _ { j } ) , } \end{array} +$$ + +where $\mathbf { M } _ { i j }$ denotes the probability that the edge between $v _ { i }$ and $v _ { j }$ is contained in the explanation. When there is no edge between $v _ { i }$ and $v _ { j }$ ,that is, ${ \bf A } _ { i j } = 0$ ,we set $\mathbf { \check { M } } _ { i j } = 0$ + +For an input graph $G = \{ V , E \}$ with $n$ vertices and a trained neural network $f _ { \boldsymbol { \theta } } , \mathbf { M }$ is an $n$ -by- $\mathbf { \nabla } \cdot n$ matrix that carries the complete information to generate a robust counterfactual explanation as a subset of edges, denoted by $S \subseteq E$ . Concretely,we obtain $S$ by selecting all the edges in $E$ whose corresponding entries in $\mathbf { M }$ are larger than O.5. + +# 4.3.2Training Model $f _ { \theta }$ + +For an input graph $G = ( V , E )$ , denote by $S \subseteq E$ the subset of edges produced by $f _ { \theta }$ to explain the prediction $\phi ( G )$ ,our goal is to train a good model $f _ { \theta }$ such that the prediction on the subgraph $G _ { S }$ induced by $S$ from $G$ is consistent with $\phi ( G )$ ; and deleting the edges in $S$ from $G$ produces a remainder subgraph $G _ { E \backslash S }$ such that the prediction on $G _ { E \backslash S }$ changes significantly from ${ \bar { \phi } } ( G )$ ： + +Since producing $S$ by $f _ { \theta }$ is a discrete operation that is hard to incorporate in an end-to-end training process, we define two proxy graphs to approximate $G _ { S }$ and $G _ { E \backslash S }$ ,respectively,such that the proxy graphs are determined by $\theta$ through continuous functions that can be smoothly incorporated into an end-to-end training process. + +The proxy graph of $G _ { S }$ , denoted by $G _ { \theta }$ , is defined by regarding $\mathbf { M }$ instead of $\mathbf { A }$ as the adjacency matrix. That is, $G _ { \theta }$ has exactly the same graph structure as $G$ , but the edge weights of $G _ { \theta }$ is given by the entries in $\mathbf { M }$ instead of $\mathbf { A }$ .Here,the subscript $\theta$ means $G _ { \theta }$ is determined by $\theta$ + +The proxy graph of $G _ { E \backslash S }$ , denoted by $G _ { \theta } ^ { \prime }$ , also have the same graph structure as $G$ , but the edge weight between each pair of vertices $v _ { i }$ and $v _ { j }$ is defined as + +$$ +\begin{array} { r } { \mathbf { M } _ { i j } ^ { \prime } = \left\{ { 1 - \mathbf { M } _ { i j } \quad \mathrm { ~ i f ~ } \mathbf { A } _ { i j } = 1 } \atop { 0 } \right. } \end{array} +$$ + +The edge weights of both $G _ { \theta }$ and $G _ { \theta } ^ { \prime }$ are determined by $\theta$ through continuous functions, thus we can smoothly incorporate $G _ { \theta }$ and $G _ { \theta } ^ { \prime }$ into an end-to-end training framework. + +As discussed later in this section, we use a regularization term to force the value of each entry in $\mathbf { M } _ { i j }$ to be close to eitherOor 1,such that $G _ { \theta }$ and $\operatorname { \mathrm { \Sigma } } _ { G _ { \theta } ^ { \prime } } ^ { G \prime }$ better approximate $G _ { S }$ and $G _ { E \backslash S }$ respectively. + +We formulate our loss function as + +$$ +\mathcal { L } ( \theta ) = \sum _ { G \in D } \left\{ \lambda \mathcal { L } _ { s a m e } ( \theta , G ) + ( 1 - \lambda ) \mathcal { L } _ { o p p } ( \theta , G ) + \beta \mathcal { R } _ { s p a r s e } ( \theta , G ) + \mu \mathcal { R } _ { d i s c r e t e } ( \theta , G ) \right\} , +$$ + +where $\lambda \in [ 0 , 1 ]$ $\beta \geq 0$ and $\mu \geq 0$ are the hyperparameters controlling the importance of each term. The influence of these parameters is discussed in Appendix $\mathbf { G }$ .The first term of our loss function requires that the prediction of the GNN on $G _ { \theta }$ is consistent with the prediction on $G$ . Intuitively, this means that the edges with larger weights in $G _ { \theta }$ dominate the prediction on $G$ .We formulate this term by requiring $G _ { \theta }$ to be covered by the same decision region covering $G$ ： + +Denote by $\mathcal { H } _ { G }$ the set of LDBs that induce the decision region covering $G$ ,and by $| \mathcal { H } _ { G } |$ the number of LDBs in $\mathcal { H } _ { G }$ . For the $i$ -th LDB $h _ { i } \in \mathcal { H } _ { G }$ , denote by $B _ { i } ( \mathbf { x } ) = \mathbf { w } _ { i } ^ { \top } \mathbf { x } + b _ { i }$ , where $\mathbf { w } _ { i }$ and $b _ { i }$ are the basis and bias of $h _ { i }$ , respectively, and $\mathbf { x } \in \mathbb { O } ^ { d }$ is a point in the space $\mathbb { O } ^ { d }$ . The sign of $B _ { i } ( { \bf x } )$ indicates whether a point $\mathbf { x }$ lies on the positive side or the negative side of $h _ { i }$ , and the absolute value $| B _ { i } ( { \bf x } ) |$ is proportional to the distance of a point $\mathbf { x }$ from $h _ { i }$ . Denote by $\sigma ( \cdot )$ the standard sigmoid function, we formulate the first term of our loss function as + +$$ +\mathcal { L } _ { s a m e } ( \theta , G ) = \frac { 1 } { | \mathcal { H } _ { G } | } \sum _ { h _ { i } \in \mathcal { H } _ { G } } \sigma \left( - \mathcal { B } _ { i } ( \phi _ { g c } ( G ) ) * \mathcal { B } _ { i } ( \phi _ { g c } ( G _ { \theta } ) ) \right) , +$$ + +such that minimizing $\mathcal { L } _ { s a m e } ( \theta , G )$ encourages the graph embeddings $\phi _ { g c } ( G )$ and $\phi _ { g c } ( G _ { \theta } )$ to lie on the same side of every LDB in $\mathcal { H } _ { G }$ . Thus, $G _ { \theta }$ is encouraged to be covered by the same decision region covering $G$ ： + +The second term of our loss function optimizes the counterfactual property of the explanations by requiring the prediction on $G _ { \theta } ^ { \prime }$ to be significantly different from the prediction on $G$ . Intuitively, this means that the set of edges with larger weights in $G _ { \theta }$ are good counterfactual explanations because reducing the weights of these edges significantly changes the prediction. Following the above intuition,we formulate the second term as + +$$ +\mathcal { L } _ { o p p } ( \theta , G ) = \operatorname* { m i n } _ { h _ { i } \in \mathcal { H } _ { G } } \sigma \left( \mathcal { B } _ { i } ( \phi _ { g c } ( G ) ) * \mathcal { B } _ { i } ( \phi _ { g c } ( G _ { \theta } ^ { \prime } ) ) \right) , +$$ + +such that minimizing $\mathcal { L } _ { o p p } ( \theta , G )$ encourages the graph embeddings $\phi _ { g c } ( G )$ and $\phi _ { g c } ( G _ { \theta } ^ { \prime } )$ to lie on the opposite sides of at least one LDB in $\mathcal { H } _ { G }$ . This further means that $G _ { \theta } ^ { \prime }$ is encouraged not to be covered by the decision region covering $G$ , thus the prediction on $G _ { \theta } ^ { \prime }$ can be changed significantly from the prediction on $G$ ： + +Similar to [45], we use a L1 regularization $\mathcal { R } _ { s p a r s e } ( \theta , G ) = \| \mathbf { M } \| _ { 1 }$ on the matrix $\mathbf { M }$ produced by $f _ { \theta }$ on an input graph $G$ to produce a sparse matrix M, such that only a small number of edges in $G$ are selected as the counterfactual explanation. We also follow [45] to use an entropy regularization + +$$ +\mathcal { R } _ { d i s c r e t e } ( \theta , G ) = - \frac { 1 } { | \mathbf { M } | } \sum _ { i , j } ( \mathbf { M } _ { i j } \log ( \mathbf { M } _ { i j } ) + ( 1 - \mathbf { M } _ { i j } ) \log ( 1 - \mathbf { M } _ { i j } ) ) +$$ + +to push the value of each entry in $\mathbf { M } _ { i j }$ to be close to either Oor 1,such that $G _ { \theta }$ and $G _ { \theta } ^ { \prime }$ approximate $G _ { S }$ and $G _ { E \backslash S }$ well, respectively. + +Now we can use the graphs in $D$ and the extracted decision regions to train the neural network $f _ { \theta }$ in an end-to-end manner by minimizing $\mathcal { L } ( \boldsymbol { \theta } )$ over $\theta$ using back propagation. Once we finish training $f _ { \theta }$ , we can first apply $f _ { \theta }$ to produce the matrix $\mathbf { M }$ for an input graph $G = ( V , E )$ ,and then obtain the explanation $S$ by selecting all the edges in $E$ whose corresponding entries in $\mathbf { M }$ are larger than 0.5. We do not need the extracted boundaries for inference as the the decision logic of GNN is already distilled into the explanation network $f$ during the training. + +As discussed in Appendix B,our method can be easily extended to generate robust counterfactual explanations for node classification tasks. + +Our method is highly efficient with a time complexity $O ( | E | )$ for explaining the prediction on an input graph $G$ , where $| E |$ is the total number of edges in $G$ : Additionally, the neural network $f _ { \theta }$ can be directly used without retraining to predict explanations on unseen graphs. Thus our method is significantly faster than the other methods [45,32,47,38] that require retraining each time when generating explanations on a new input graph. + +# 5Experiments + +We conduct series of experiments to compare our method with the state-of-the-art methods including GNNExplainer [45],PGExplainer [25],PGM-Explainer [38], SubgraphX [47] and CFGNNExplainer [24]. For the methods that identify a set of vertices as an explanation, we use the set of vertices to induce a subgraph from the input graph,and then use the set of edges of the induced subgraph as the explanation. For the methods that identify a subgraph as an explanation, we directly use the set of edges of the identified subgraph as the explanation. + +![](images/b7fd56d0939a4af57dc11439ffc5d8bfba231262c91ca025fc019f5b6e33dc49.jpg) +Figure 1: Fidelity performance averaged across 10 runs for the datasets at different levels of sparsity. + +To demonstrate the effectiveness of the decision regions,we derive another baseline method named RCExp-NoLDB that adopts the general framework of RCExplainer but does not use the LDBs of decision regions to generate explanations.Instead, RCExp-NoLDB directly maximizes the prediction confidence on class $c$ for $G _ { \theta }$ and minimizes the prediction confidence of class $c$ for $G _ { \theta } ^ { \prime }$ ： + +We evaluate the explanation performance on two typical tasks: the graph clasification task that uses a GNN to predict the class of an input graph,and the node classification task that uses a GNN to predict the class of a graph node. For the graph classification task, we use one synthetic dataset, BA-2motifs [25],and two real-world datasets,Mutagenicity [21] and NCI1 [39].For the node classification task, we use the same four synthetic datasets as used by GNNExplainer [45], namely, BA-SHAPES,BA-COMMUNITY,TREE-CYCLES and TREE-GRID. + +Limited by space, we only report here the key results_on the graph classification task for fidelity, robustness and efficiency. Please refer to Appendix E for details on datasets, baselines and the experiment setups.Detailed experimental comparison on the node classification task willbe discussed in Appendix F where we show that our method produces extremely accurate explanations. The code3 is publicly available. + +# 5.1Fidelity + +Fidelity is measured by the decrease of prediction confidence after removing the explanation (i.e., a set of edges) from the input graph [32]. We use fidelity to evaluate how counterfactual the generated explanations are on the datasets Mutagenicity,NCI1 and BA-2motifs.A large fidelity score indicates stronger counterfactual characteristics. It is important to note that fidelity may be sensitive to sparsity of explanations. The sparsity of an explanation $S$ with respect to an input graph $G = ( \bar { V , } E )$ is $\begin{array} { r } { s p a r s i t y ( S , G ) = 1 - \frac { | S | } { | E | } } \end{array}$ ,thatis,tpeeefdgafrta from $G$ . We only compare explanations with the same level of sparsity. + +Figure 1 shows the results about fidelity. Our approach achieves the best fidelity performance at all levels of sparsity. The results validate the effectiveness of our method in producing highly counterfactual explanations. RCExplainer also significantly outperforms RCExp-NoLDB. This confirms that using LDBs of decision regions extracted from GNNs produces more faithful counterfactual explanations. + +CF-GNNExplainer performs the best among the rest of the methods.This is expected as it optimizes the counterfactual behavior of the explanations which results in higher fidelity for the explanations in comparison to those produced by other methods such as GNNExplainer and PGExplainer. + +The fidelity performance of SubgraphX reported in [47] was obtained by seting the features of nodes that are part of the explanation to Obut not removing the explanation edges from the input graph. This does not remove the message passing roles of the explanation nodes from the input graph because the edges connected to those nodes still can pass messages. In our experiments, we directly block the messages that are passed on the edges in the explanation, which completely prevents the explanation nodes in the input graph to participate in the message passing. As a result, the performance of SubgraphX drops significantly. + +![](images/1aa166e7cf33380e4add83effc9273d8301d54224aa1859e6e8d4478bf17806a.jpg) +Figure 2: Noise robustness (AUC) averaged across 10 runs for the datasets at diferent levels of noise. + +# 5.2Robustness Performance + +In this experiment, we evaluate the robustness of all methods by quantifying how much an explanation changes after adding noise to the input graph. For an input graph $G$ and the explanation $S$ ,we produce a perturbed graph $G ^ { \prime }$ by adding random noise to the node features and randomly adding or deleting some edges of the input graph such that the prediction on $G ^ { \prime }$ is consistent with the prediction on $G$ ： Using the same method we obtain the explanation $S ^ { \prime }$ on $G ^ { \prime }$ . Considering top- $k$ edges of $S$ as the ground-truth and comparing $S ^ { \prime }$ against them, we compute a receiver operating characteristic (ROC) curve and evaluate the robustnessby the area under curve (AUC) of the ROC curve. We report results for $k = 8$ in Figure 2. Results for other values of $k$ are included in Appendix F where we observe similar trend. + +Figure 2 shows the AUC of GNNExplainer, PGExplainer, RCExp-NoLDB and RCExplainer at different levels of noise. A higher AUC indicates better robustness. The percentage of noise shows the proportion of nodes and edges that are modified. Baselines such as PGM-Explainer and SubgraphX are not included in this experiment as they do not output the edge weights that are required for computing AUC.We present additional robustness experiments in Appendix F where we extend all the baselines to report node and edge level accuracy. + +GNNExplainer performs the worst on most of the datasets,since it optimizes each graph independently without considering other graphs in the training set. Even when no noise is added,the AUC of GNNExplainer is significantly lower than 1 because different runs produce different explanations for the same graph prediction. PGExplainer is generally more robust than GNNExplainer because the neural network they trained to produce explanations implicitly considers all the graphs used for training. CF-GNNExplainer also performs worse than RCExplainer, which means it is more susceptible to the noise as compared to RCExplainer. + +Our method achieves the best AUC on allthe datasets, because the common decision logic carried by the decision regions of a GNN is highly robust to noise. PGExplainer achieves a comparable performance as our method on the Mutagenicity dataset, because the samples of this dataset share a lot of common structures such as carbon rings, which makes it easier for the neural network trained by PGExplainer to identify these structures in presence of noise. However, for BA-2motifs and NCI1, this is harder as samples share very few structures and thus the AUC of PGExplainer drops significantly. RCExplainer also significantly outperforms RCExp-NoLDB on these datasets which highlights the role of decision boundaries in making our method highly robust. + +Table 1: Average time cost for producing an explanation on a single graph sample. + +
MethodGNNExplainerPGExplainerPGM-ExplainerSubgraphXCF-GNNExplainerRCExplainer
Time1.2s ± 0.20.01s ± 0.0313.1s ± 3.977.8s ± 4.54.6s± 0.20.01s ± 0.02
+ +Effciency. We evaluate efficiency by comparing the average computation time taken for inference on unseen graph samples. Table 1 shows the results on the Mutagenicity dataset. Since our method also can be directly used for unseen data without any retraining,it is as eficient as PGExplainer and significantly faster than GNNExplainer, PGM-Explainer, SubgraphX and CF-GNNExplainer. + +# 6Conclusion + +In this paper, we develop a novel method for producing counterfactual explanations on GNNs. We extract decision boundaries from the given GNN model to formulate an intuitive and effective counterfactual loss function. We optimize this loss to train a neural network to produce explanations with strong counterfactual characteristics. Since the decision boundaries are shared by multiple samples of the same predicted class, explanations produced by our method are robust and do not overfit the noise. Our experiments on synthetic and real-life benchmark datasets strongly validate the efficacy of our method. In this work,we focus on GNNs that belong to Piecewise Linear Neural Networks (PLNNs). Extending our method to other families of GNNs and tasks such as link prediction,remains an interesting future direction. + +Our method will benefit multiple fields where GNNs are intensively used. By allowing the users to interpret the predictions of complex GNNs better, it will promote transparency, trust and fairness in the society. However, there also exist some inherent risks. A generated explanation may expose private information if our method is not coupled with an adequate privacy protection technique. Also, some of the ideas presented in this paper may be adopted and extended to improve adversarial attacks. Without appropriate defense mechanisms,the misuse of such attcks poses a risk of disruption in the functionality of GNNs deployed in the real world.That said, we firmly believe that these risks can be mitigated through increased awareness and proactive measures. + +# Acknowledgments and Disclosure of Funding + +Lingyang Chu's research is supported in part by the startup grant provided by the Department of Computing and Software of McMaster University. All opinions, findings,conclusions,and recommendations in this paper are those of the authors and do not necessarily reflect the views of the funding agencies. + +# References + +[1] J. Adebayo,J. Gilmer,M.Muelly,I. Goodfellow,M. Hardt,and B. Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems. +[2] A.Bojchevski and S. Gunnemann. Certifiable robustness to graph perturbations.arXiv preprint arXiv:1910.14356,2019. +[3]E. Cho,S.A. Myers,and J.Leskovec.Friendship and mobility: user movement in location-based social networks. In Proceedings of the 17th ACM SIGkDD international conference on Knowledge discovery and data mining,pages 1082-1090,2011. +[4]L.Chu, X. Hu,J. Hu,L. Wang,and J.Pei. Exact and consistent interpretation for piecewise linear neural networks: A closed form solution. In Proceedings of the 24th ACM SiGKDD International Conference on Knowledge Discovery & Data Mining,pages 1244-1253,2018. +[5]V. Crawford,A. Kuhnle,and M. Thai. Submodular cost submodular cover with an approximate oracle.In International Conference on Machine Learning, pages 1426-1435.PMLR,2019. +[6] A. K. Debnath, R.L.Lopez de Compadre, G. Debnath, A. J. Shusterman,and C. Hansch. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds.correlation with molecular orbital energies and hydrophobicity. Journal of medicinal chemistry,34(2):786-797,1991. +[7]M.Denil,S.G.Colmenarejo,S. Cabi,D.Saxton,and N. de Freitas. Programmable agents. arXiv preprint arXiv:1706.06383,2017. +[8]Y. Duan, M. Andrychowicz,B.C. Stadie,J. Ho,J.Schneider,I.Sutskever,P.Abbeel,and W. Zaremba. One-shot imitation learning. arXiv preprint arXiv:1703.07326,2017. +[9]W.Fan,Y.Ma,Q.Li,Y. He,E. Zhao,J.Tang,and D.Yin. Graph neural networks for social recommendation. In The World Wide Web Conference,pages 417-426,2019. +[10]T.Freiesleben.Counterfactualexplanations &adversarialexamplescommongrounds,essntialdifferences, and potential transfers. arXiv preprint arXiv:2009.05487, 2020. +[11]A. Ghorbani,A. Abid,and J. Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3681-3688,2019. +[12] X. Glorot and Y. Bengio. Understanding the dificulty of training deep feedforward neural networks. InProceedingsofthethirteenth international conference onartificial intelligence and statistics,pages 249-256. JMLR Workshop and Conference Proceedings, 2010. +[13] X.Glorot,A.Bordes,and Y.Bengio.Deep sparse rectifier neural networks.In Proceedings ofthe fourteenth international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and Conference Proceedings,2011. +[14] I. Goodfelow,D. Warde-Farley,M. Mirza,A. Courvile,and Y.Bengio. Maxout networks.In International conference on machine learning, pages 1319-1327.PMLR,2013. +[15]K. He, X. Zhang, S. Ren,and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings ofthe IEEE international conferenceon computer vision, pages 1026-1034,2015. +[16] L.Holdijk,M. Boon,S.Henckens,and L.de Jong. [re] parameterized explainer for graph neural network. In ML Reproducibility Challenge 2020,2021. +[17] W. Huber, V.J. Carey,L.Long, S. Falcon,and R. Gentleman. Graphs in molecular biology. BMC bioinformatics,8(6):1-14,2007. +[18] R. K. Iyer and J. A. Bilmes. Submodular optimization with submodular cover and submodular knapsack constraints.In Advances in Neural Information Processing Systems. +[19]M. Jiang,Z. Li, S. Zhang,S. Wang,X. Wang, Q. Yuan,and Z. Wei. Drug-target affinity prediction using graph neural network and contact maps.RsC Advances,10(35):20701-20712, 2020. +[20] H. Jin and X. Zhang. Latent adversarial training of graph convolution networks. In ICML Workshop on Learning and Reasoning with Graph-Structured Representations,2019. +[21] J. Kazius,R. McGuire,and R. Bursi. Derivation and validation of toxicophores for mutagenicity prediction. Journal of medicinal chemistry,48(1):312-320,2005. +[22] T.N. Kipf and M. Welling. Semi-supervised classfication with graph convolutional networks. In 5th International Conference on Learning Representations,ICLR 2Ol7, Conference Track Proceedings. +[23]M. Liu, Y. Luo,L. Wang, Y. Xie, H. Yuan,S. Gui, H. Yu,Z. Xu, J. Zhang, Y. Liu, K. Yan, H. Liu, C. Fu, B. Oztekin,X. Zhang,and S.Ji. DIG: A turnkey library for diving into graph deep learning research. arXiv preprint arXiv:2103.12608,2021. +[24] A.Lucic,M.ter Hoeve,G. Tolomei,M.de Rijke,and F. Silvestri. Cf-gnnexplainer: Counterfactual explanations for graph neural networks.arXiv preprint arXiv:2102.03322,2021. +[25] D.Luo, W. Cheng,D. Xu, W. Yu, B. Zong, H. Chen,and X. Zhang. Parameterized explainer for graph neural network. In Advances in Neural Information Processing Systems. +[26] C. Molnar.Interpretable Machine Learning.2019.https://christophm.github.io/ interpretable-ml-book/. +[27] G.Montufar,R.Pascanu,K.Cho,and Y.Bengio.On the numberof linear regions of deep neural networks. arXiv preprint arXiv:1402.1869,2014. +[28]R.Moraffah,M. Karami,R.Guo,A. Raglin,and H.Liu.Causal interpretability for machine learningproblems,methods and evaluation. ACM SIGKDD Explorations Newsleter,22(1):18-33,2020. +[29] V.Nair and G.E.Hinton.Rectified linear units improve restricted boltzmann machines.In Icml,2010. +[30] D. Numeroso and D. Bacciu. Meg: Generating molecular counterfactual explanations for deep graph networks. arXiv preprint arXiv:2104.08060,2021. +[31] D.E. Pires,T. L. Blundel, and D. B. Ascher. pkcsm: predicting small-molecule pharmacokinetic and toxicity properties using graph-based signatures. Journal of medicinal chemistry,58(9):4066-4072,2015. +[32]P.E.Pope,S. Kolouri, M.Rostami, C.E. Martin,and H. Hoffmann.Explainability methods for graph convolutional neural networks.In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10772-10781,2019. +[33]R.R.Selvaraju,M.Cogswell,A.Das,R. Vedantam,D.Parikh,and D.Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. +[34] P.L. Simon, M.Taylor,and I. Z. Kiss. Exact epidemic models on graphs using graph-automorphism driven lumping. Journal of mathematical biology,62(4):479-508,2011. +[35] K. Simonyan, A. Vedaldi,and A. Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. ICLR,2014. +[36]K.Sokol and P.A.Flach. Counterfactual explanations of machine learning predictions: opportunities and challenges for ai safety. In SafeAI@ AAAI, 2019. +[37]P. Velickovic, G. Cucurul,A. Casanova,A. Romero,P.Lio,and Y. Bengio. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2O18, Conference Track Proceedings. +[38]M. Vu and M. T. Thai.Pgm-explainer: Probabilistic graphical model explanations for graph neural networks.In Advances in Neural Information Processing Systems. +[39] N. Wale and G. Karypis. Comparison of descriptor spaces for chemical compound retrieval and classification. In Sixth International Conference on Data Mining (ICDM'06),pages 678-689,2006.doi: 10.1109/ICDM.2006.39. +[40] L.A.Wolsey. An analysis of the greedy algorithm for the submodular setcovering problem. Combinatorica, 2(4):385-393,1982. +[41] J. Xiong,Z. Xiong, K. Chen,H. Jiang,and M. Zheng. Graph neural networks for automated de novo drug design. Drug Discovery Today,2021. +[42]H. Xu, Y. Ma, H.-C.Liu,D. Deb, H. Liu, J.-L. Tang,and A. K. Jain. Adversarial atacks and defenses in images, graphs and text: A review. International Journal of Automation and Computing,17(2):151-178, 2020. +[43] K. Xu, H. Chen,S.Liu,P.-Y. Chen, T.-W. Weng,M. Hong,and X. Lin. Topology attack and defense for graph neural networks: An optimization perspective. arXiv preprint arXiv:1906.04214, 2019. +[44] R. Yin,K.Li, G. Zhang,and J.Lu. A deeper graph neural network for recommender systems. KnowledgeBased Systems,185:105020,2019. +[45] Z. Ying,D. Bourgeois, J. You, M. Zitnik,and J.Leskovec. Gnnexplainer: Generating explanations for graph neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019. +[46] H.Yuan,J.Tang,X.Hu,and S.Ji. Xgnn: Towards model-level explanations of graph neural networks.In Proceedings ofthe 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 430-438,2020. +[47] H. Yuan, H. Yu, J. Wang, K. Li, and S.Ji._On explainability of graph neural networks via subgraph explorations. arXiv preprint arXiv:2102.05152,2021. +[48] M.D. Zeiler and R.Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818-833. Springer,2014. +[49] J. Zhang,S.A. Bargal,Z.Lin,J.Brandt,X. Shen,and S.Sclarof. Top-down neural attention by excitation backprop. International Journal of Computer Vision,126(10):1084-1102,2018. +[50] M. Zhang and Y. Chen. Link prediction based on graph neural networks. In Advances in Neural Information Processing Systems. +[51] D. Zugner and S. Gunnemann. Adversarial attacks on graph neural networks via meta learning. arXiv preprint arXiv:1902.08412, 2019. +[52] D.Zügner and S.Gunnemann.Certifable robustness and robust training for graph convolutional networks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,pages 246-256,2019. +[53] D. Zugner, A. Akbarnejad,and S. Gunnemann. Adversarial atacks on neural networks for graph data. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2847-2856,2018. \ No newline at end of file diff --git a/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_content_list.json b/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..39115f917e4d9323db3b22bec74d7da5f47a453c --- /dev/null +++ b/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_content_list.json @@ -0,0 +1,1353 @@ +[ + { + "type": "text", + "text": "Robust Counterfactual Explanations on Graph Neural Networks ", + "text_level": 1, + "bbox": [ + 215, + 122, + 782, + 170 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohit Bajaj1\\* Lingyang $\\mathbf { C h u ^ { 2 * } }$ Zi Yu Xue1,3Jian Pei4 Lanjun Wang1Peter Cho-Ho Lam1Yong Zhang1 ", + "text_level": 1, + "bbox": [ + 289, + 218, + 709, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Huawei Technologies Canada Co., Ltd. 2McMaster University 3 The University of British Columbia 4Simon Fraser University {mohit.bajaj1，zi.yu.xue，lanjun.wang，cho.ho.lam，yong.zhang3}@huawei.com chul9@mcmaster.ca,jpei@cs.sfu.ca ", + "bbox": [ + 214, + 261, + 784, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 352, + 535, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Massive deployment of Graph Neural Networks (GNNs) in high-stake applications generates a strong demand for explanations that are robust to noise and align well with human intuition. Most existing methods generate explanations by identifying a subgraph of an input graph that has a strong correlation with the prediction. These explanations are not robust to noise because independently optimizing the correlation for a single input can easily overfit noise.Moreover, they are not counterfactual because removing an identified subgraph from an input graph does not necessarily change the prediction result. In this paper, we propose a novel method to generate robust counterfactual explanations on GNNs by explicitly modelling the common decision logic of GNNs on similar input graphs. Our explanations are naturally robust to noise because they are produced from the common decision boundaries of a GNN that govern the predictions of many similar input graphs. The explanations are also counterfactual because removing the set of edges identified by an explanation from the input graph changes the prediction significantly. Exhaustive experiments on many public datasets demonstrate the superior performance of our method. ", + "bbox": [ + 233, + 382, + 766, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction ", + "text_level": 1, + "bbox": [ + 174, + 617, + 310, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graph Neural Networks (GNNs) [22,37,50] have achieved great practical successes in many realworld applications,such as chemistry [31],molecular biology[17],social networks [3] and epidemic modelling [34]. For most of these applications, explaining predictions made by a GNN model is crucial for establishing trust with end-users,identifying the cause of a prediction,and even discovering potential deficiencies of a GNN model before massive deployment. Ideally,an explanation should be able to answer questions like “Would the prediction of the GNN model change if a certain part of an input molecule is removed?” in the context of predicting whether an artificial molecule is active for acertain type of proteins [19,41],“Would an item recommended stillbe recommended if a customer had not purchased some other items in the past?” for a GNN built for recommendation systems [9, 44]. ", + "bbox": [ + 174, + 647, + 825, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Counterfactual explanations [28] in the form of “If X had not occurred, Y would not have occurred”[26] are the principled way to answer such questions and thus are highly desirable for GNNs. In the context of GNNs,a counterfactual explanation identifies a small subset of edges of the input graph instance such that removing those edges significantly changes the prediction made by the GNN. Counterfactual explanations are usually concise and easy to understand [28,36] because they align well with the human intuition to describe a causal situation [26]. To make explanations more trustworthy, the counterfactual explanation should be robust to noise,that is,some slight changes on an input graph do not change the explanation significantly. This idea aligns well with the notion of robustness discussed for DNN explanations in computer vision domain [11]. According to Ghorbani et al.[11] many interpretations on neural networks are fragile as it is easier to generate adversarial perturbations that produce perceptively indistinguishable inputs that are assigned the same predicted label,yet have very different interpretations.Here,the concepts of “fragile”“robustness”describe the same concept from opposite perspectives. An interpretation is said to be fragile if systematic perturbations can lead to dramatically different interpretations without changing the label. Otherwise, the interpretation is said to be robust. ", + "bbox": [ + 174, + 785, + 825, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 90, + 825, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How to produce robust counterfactual explanations on predictions made by general graph neural networks is a novel problem that has not been systematically studied before.As to be discussed in Section 2, most GNN explanation methods [45,25,46,37,32] are neither counterfactual nor robust. These methods mostly focus on identifying a subgraph of an input graph that achieves a high correlation with the prediction result. Such explanations are usually not counterfactual because, due to the high non-convexity of GNNs, removing a subgraph that achieves a high correlation does not necessarily change the prediction result. Moreover, many existing methods [45,25,37,32] are not robust to noise and may change significantly upon slight modifications on input graphs, because the explanation of every single input graph prediction is independently optimized to maximize the correlation with the prediction, thus an explanation can easily overfit the noise in the data. ", + "bbox": [ + 174, + 204, + 825, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper², we develop RCExplainer, a novel method to produce robust counterfactual explanations on GNNs. The key idea is to first model the common decision logic of a GNN by set of decision regions where each decision region governs the predictions on a large number of graphs,and then extract robust counterfactual explanations by a deep neural network that explores the decision logic carried by the linear decision boundaries of the decision regions. We make the following contributions. ", + "bbox": [ + 174, + 343, + 825, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we model the decision logic of a GNN by a set of decision regions, where each decision region is induced by a set of linear decision boundaries of the GNN. We propose an unsupervised method to find decision regions for each class such that each decision region governs the prediction of multiple graph samples predicted to be the same class. The linear decision boundaries of the decision region capture the common decision logic on all the graph instances inside the decision region, thus do not easily overfit the noise of an individual graph instance. By exploring the common decision logic encoded in the linear boundaries, we are able to produce counterfactual explanations that are inherently robust to noise. ", + "bbox": [ + 174, + 415, + 825, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, based on the linear boundaries of the decision region,we propose a novel loss function to train a neural network that produces a robust counterfactual explanation as a small subset of edges of an input graph. The loss function is designed to directly optimize the explainability and counterfactual property of the subset of edges,such that: 1) the subgraph induced by the edges lies within the decision region,thus has a prediction consistent with the input graph; and 2) deleting the subset of edges from the input graph produces a remainder subgraph that lies outside the decision region, thus the prediction on the remainder subgraph changes significantly. ", + "bbox": [ + 174, + 527, + 825, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Last, we conduct comprehensive experimental study to compare our method with the state-of-the-art methods on fidelity, robustness,accuracy and efficiency. All the results solidly demonstrate the superior performance of our approach. ", + "bbox": [ + 176, + 626, + 821, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2Related work ", + "text_level": 1, + "bbox": [ + 174, + 685, + 316, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing GNN explanation methods [46,37, 45,32, 25] generally fall into two categories: model level explanation [46] and instance level explanation [37,45,32, 25]. ", + "bbox": [ + 174, + 715, + 823, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A model level explanation method [46] produces a high-level explanation about the general behaviors of a GNN independent from input examples. This may be achieved by synthesizing a set of artificial graph instances such that each artificial graph instance maximizes the prediction score on a certain class. The weakness of model level explanation methods is that an input graph instance may not contain an artificial graph instance,and removing an artificial graph from an input graph does not necessarily change the prediction. As a result, model level explanations are substantially different from counterfactual explanations, because the synthesized artificial graphs do not provide insights into how the GNN makes its prediction on a specific input graph instance. ", + "bbox": [ + 174, + 748, + 825, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The instance level explanation methods [37,45,32, 25] explain the prediction(s) made by a GNN on a specific input graph instance or multiple instances by identifying a subgraph of an input graph instance that achieves a high correlation with the prediction on the input graph. GNNExplainer [45] removes redundant edges from an input graph instance to produce an explanation that maximizes the mutual information between the distribution of subgraphs of the input graph and the GNN's prediction. Following the same idea by Ying et al. [45],PGExplainer [25] parameterizes the generation process of explanations by a deep neural network,and trains it to maximize a similar mutual information based loss used by GNNExplainer [45]. The trained deep neural network is then applied to generate explanations for a single input graph instance or a group of input graphs. MEG [30] incorporates strong domain knowledge in chemistry with a reinforcement learning framework to produce counterfactual explanations on GNNs specifically built for compound prediction, but the heavy reliance on domain knowledge largely limits its applicability on general GNNs. The recently proposed CF-GNNExplainer [24] independently optimizes the counterfactual property for each explanation but ignores the correlation between the prediction and the explanation. ", + "bbox": [ + 178, + 861, + 825, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 90, + 825, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some studies [32,37] also adapt the existing explanation methods of image-oriented deep neural networks to produce instance level explanations for GNNs.Pope et al.[32] extend several gradient based methods [33,35,49] to explain predictions made by GNNs. The explanations are prone to gradient saturation[12] and may also be misleading [1] due to the heavy reliance on noisy gradients. Velickovic et al.[37] extend the atention mechanism[7,8] to identify the nodes in an input graph that contribute the most to the prediction.This method has to retrain the GNN with the altered architecture and the inserted attntion layers.Thus,the explanations may not be faithful to the original GNN. ", + "bbox": [ + 173, + 256, + 825, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance level explanations from most of the methods are usually not counterfactual because, due to the non-convexity of GNNs, removing an explanation subgraph from the input graph does not necessarily change the prediction result. Moreover, those methods [45,25,37,32, 24] are usually not robust to noise because the explanation of every single input graph prediction is independently optimized. Thus,an explanation can easily overfit the noise inside input graphs and may change significantly upon slight modifications on input graphs. ", + "bbox": [ + 174, + 354, + 825, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the weaknesses in the existing methods, in this paper, we directly optimize the counterfactual property of an explanation along with the correlation between the explanation and the prediction. Our explanations are also much more robust to modifications on input graphs,because they are produced from the common decision logic on a large group of similar input graphs, which do not easily overfit the noise of an individual graph sample. ", + "bbox": [ + 174, + 440, + 825, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Please note that our study is substantially different from adversarial attacks on GNNs.The adversarial attacking methods [51,53,42,43,20] use adversarial examples to change the predictions of GNNs but ignore the explainability of the generated adversarial examples [1O]. Thus, the adversarial examples generated by adversarial attcks may not explain the original prediction. ", + "bbox": [ + 174, + 512, + 825, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method is substantially different from the above works because we focus on explaining the prediction by directly optimizing the counterfactual property of an explanation along with correlation of the explanation with the prediction. We also require that the explanation is generally valid for a large set of similar graph instances by extracting it from the common linear decision boundaries of a large decision region. ", + "bbox": [ + 174, + 571, + 825, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3Problem Formulation ", + "text_level": 1, + "bbox": [ + 174, + 667, + 387, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denote by $G = \\{ V , E \\}$ a graph where $V = \\{ v _ { 1 } , v _ { 2 } , \\ldots , v _ { n } \\}$ is the set of $n$ nodes and $E \\subseteq V \\times V$ is the set of edges. The edge structure of a graph $G$ is described by an adjacency matrix $\\mathbf { A } \\in \\{ 0 , 1 \\} ^ { n \\times n }$ where ${ \\bf A } _ { i j } = 1$ if there is an edge between node $v _ { i }$ and $v _ { j }$ ; and ${ \\bf A } _ { i j } = 0$ otherwise. ", + "bbox": [ + 174, + 704, + 825, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denote by $\\phi$ a GNN model that maps a graph to a probability distribution over a set of classes denoted by $C$ . Let $D$ denote the set of graphs that are used to train the GNN model $\\phi$ .We focus on GNNs that adopt piecewise linear activation functions,such as MaxOut[14] and the family of ReLU[13,15,29]. ", + "bbox": [ + 176, + 752, + 823, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The robust counterfactual explanation problem is defined as follows. ", + "bbox": [ + 174, + 797, + 620, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Robust Counterfactual Explanation Problem) Given a GNN model $\\phi$ trained on a set of graphs $D$ ， for an input graph $G = \\{ V , E \\}$ , our goal is to explain why $G$ is predicted by the GNN model as $\\phi ( G )$ by identifying a small subset of edges $S \\subseteq E$ ,such that $( l )$ removing the set of edges in $S$ from $G$ that causes the maximum drop in the confidence of the original prediction; and (2) $S$ is stable and doesn't change when the edges and the feature representations of the nodes of $G$ are perturbed by random noise. ", + "bbox": [ + 174, + 830, + 825, + 911 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the definition,the first requirement requires that the explanation $S$ is counterfactual, and the second requirement requires that the explanation is robust to noisy changes on the edges and nodes of $G$ ： ", + "bbox": [ + 171, + 90, + 825, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4Method ", + "text_level": 1, + "bbox": [ + 174, + 137, + 271, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first introduce how to extract the common decision logic of a GNN on a large set of graphs with the same predicted class. This is achieved by a decision region induced by a set of linear decision boundaries of the GNN.Then, based on the linear boundaries of the decision region, we propose a novel lossfunction to train a neural network that produces robust counterfactual explanations.Last, we discuss the time complexity of our method when generating explanations. ", + "bbox": [ + 174, + 167, + 825, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1Modelling Decision Regions ", + "text_level": 1, + "bbox": [ + 176, + 250, + 405, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the routines of many deep neural network explanation methods [33, 48], we extract the decision region of a GNN in the $d$ -dimensional output space $\\mathbb { O } ^ { d }$ of the last convolution layer of the GNN. The features generated by the last convolution layer are more conceptually meaningful and more robust to noise than those raw features of input graphs,such as vertices and edges [52, 2]. Denote by $\\phi _ { g c }$ the mapping function realized by the graph convolution layers that maps an input graph $G$ to its graph embedding $\\phi _ { g c } ( G ) \\in \\mathbb { O } ^ { d }$ ,and by $\\phi _ { f c }$ the mapping function realized by the fully connected layers that maps the graph embedding $\\phi _ { g c } ( G )$ to a predicted distribution over the classes in $C$ . The overall prediction $\\phi ( { \\bar { G } } )$ made by the GNN can be writtn as $\\phi ( G ) = \\phi _ { f c } ( \\phi _ { g c } ( G ) )$ ", + "bbox": [ + 173, + 275, + 825, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the GNNs that adopt piecewise linear activation functions for the hidden neurons, such as MaxOut [14] and the family of ReLU [13,15,29], the decision logic of $\\phi _ { f c }$ in the space $\\mathbb { O } ^ { d }$ is characterized by a piecewise linear decision boundary formed by connected pieces of decision hyperplanes in $\\mathbb { O } ^ { d }$ [1]. We call these hyperplanes linear decision boundaries (LDBs), and denote by $\\mathcal { H }$ the set of LDBs induced by $\\phi _ { f c }$ . The set of LDBs in $\\mathcal { H }$ partitions the space $\\mathbb { O } ^ { d }$ into a large number of convex polytopes. A convex polytope is formed by a subset of LDBs in $\\mathcal { H }$ . All the graphs whose graph embeddings are contained in the same convex polytope are predicted as the same class [4]. Therefore, the LDBs of a convex polytope encode the common decision logic of $\\phi _ { f c }$ on all the graphs whose graph embeddings lie within the convex polytope [4]. Here,a graph $G$ is covered by a convex polytope if the graph embedding $\\phi _ { g c } ( G )$ is contained in the convex polytope. ", + "bbox": [ + 173, + 391, + 825, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above insight, we model the decision region for a set of graph instances as a convex polytope that satisfies the following two properties. First, the decision region should be induced by a subset of the LDBs in $\\mathcal { H }$ .In this way, when we extract counterfactual explanations from the LDBs, the explanations are loyal to the real decision logic_of the GNN. Second, the decision region should cover many graph instances in the training dataset $D$ ,and all the covered graphs should be predicted as the same class. In this way, the LDBs of the decision region capture the common decision logic on all the graphs covered by the decision region. Here, the requirement of covering a larger number of graphs ensures that the common decision logic is general,and thus it is lesslikely to overfit the noise of an individual graph instance. As a result, the counterfactual explanations extracted from the LDBs of the decision region are insensitive to slight changes in the input graphs. Our method can be easily generalized to incorporate prediction confidence in the coverage measure,such as considering the count of graphs weighted by prediction confidence. To keep our discusson simple, we do not pursue this detail further in the paper. ", + "bbox": [ + 174, + 534, + 825, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we illustrate how to extract a decision region satisfying the above two requirements. The key idea is to find a convex polytope covering a large set of graph instances in $D$ that are predicted as the same class $c \\in C$ ： ", + "bbox": [ + 174, + 712, + 825, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denote by $D _ { c } \\subseteq D$ the set of graphs in $D$ predicted as a class $c \\in C$ ,by $\\mathcal { P } \\subseteq \\mathcal { H }$ a set of LDBs that partition the space $\\mathbb { O } ^ { d }$ into a set of convex polytopes, and by $r ( \\mathcal { P } , c )$ the convex polytope induced by $\\mathcal { P }$ that covers the largest number of graphs in $D _ { c }$ . Denote by $g ( \\mathcal { P } , c )$ the number of graphs in $D _ { c }$ covered by $r ( \\mathcal { P } , c )$ ,and by $h ( \\mathcal { P } , c )$ the number of graphs in $D$ that are covered by $r ( \\mathcal { P } , \\bar { c } )$ but are not predicted as class $c$ . We extract a decision region covering a large set of graph instances in $D _ { c }$ by solving the following constrained optimization problem. ", + "bbox": [ + 174, + 757, + 825, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/199dfcc9238c4541a86b7079313a4ecb03e682dc957ef9622a8d982bbb04d186.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\mathcal { P } \\subseteq \\mathcal { H } } g ( \\mathcal { P } , c ) , \\mathrm { ~ s . t . ~ } h ( \\mathcal { P } , c ) = 0\n$$", + "text_format": "latex", + "bbox": [ + 398, + 845, + 599, + 871 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This formulation realizes the two properties of decision regions because $\\mathcal { P } \\subseteq \\mathcal { H }$ ensures that the decision region is induced by a subset of LDBs in $\\mathcal { H }$ ,maximizing $g ( \\mathcal { P } , c )$ requires that $r ( \\mathcal { P } , c )$ covers ", + "bbox": [ + 173, + 883, + 823, + 912 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a large number of graphs in $D _ { c }$ , and the constraint $h ( \\mathcal { P } , c ) = 0$ ensures that allthe graphs covered by $r ( \\mathcal { P } , c )$ are predicted as the same class $c$ ： ", + "bbox": [ + 173, + 90, + 823, + 118 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once we find a solution $\\mathcal { P }$ to the above problem, the decision region $r ( \\mathcal { P } , c )$ can be easily obtained by first counting the number of graphs in $D _ { c }$ covered by each convex polytope induced by $\\mathcal { P }$ ,and then select the convex polytope that covers the largest number of graphs in $D _ { c }$ ： ", + "bbox": [ + 173, + 125, + 825, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2Extracting Decision Regions ", + "text_level": 1, + "bbox": [ + 174, + 180, + 410, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimization problem in Equation (1) is intractable for standard GNNs, mainly because it is impractical to compute $\\mathcal { H }$ , all the LDBs of a GNN. The number of LDBs in $\\mathcal { H }$ of a GNN is exponential with respect to the number of neurons in the worst case [27].To address this challenge,we substitute $\\mathcal { H }$ by a sample $\\tilde { \\mathcal { H } }$ of LDBs from $\\tilde { \\mathcal { H } }$ ： ", + "bbox": [ + 173, + 205, + 825, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A LDB in the space $\\mathbb { O } ^ { d }$ can be written as $\\mathbf { w } ^ { \\top } \\mathbf { x } + b = 0$ ,where is $\\mathbf { x } \\in \\mathbb { O } ^ { d }$ is a variable, w is the basis term, and $b$ corresponds to the bias.Following [4], for any input graph $G$ ,a linear boundary can be sampled from $\\mathcal { H }$ by computing ", + "bbox": [ + 174, + 267, + 825, + 310 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/2dbc5d9de809bb0aacc312e490149627bcf7576859a17ad5d232a303bbbb6bf6.jpg", + "text": "$$\n\\mathbf { w } = \\frac { \\partial \\left( \\operatorname* { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) - \\operatorname* { m a x } _ { 2 } ( \\phi _ { f c } ( \\alpha ) ) \\right) } { \\partial \\alpha } | _ { \\alpha = \\phi _ { g c } ( G ) } ,\n$$", + "text_format": "latex", + "bbox": [ + 318, + 315, + 676, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and ", + "bbox": [ + 173, + 353, + 200, + 366 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/79c9953979154b44111c9874d086d71cda854549201b41bbe069b902cc4cfe90.jpg", + "text": "$$\nb = \\mathrm { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) - \\mathrm { m a x } _ { 2 } ( \\phi _ { f c } ( \\alpha ) ) - \\mathbf { w } ^ { T } { \\alpha } | _ { \\alpha = \\phi _ { g c } ( G ) } ,\n$$", + "text_format": "latex", + "bbox": [ + 308, + 375, + 686, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname* { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) )$ and $\\operatorname* { m a x } _ { 2 } \\bigl ( \\phi _ { f c } ( \\pmb { \\alpha } ) \\bigr )$ are the largest and the second largest values in the vector $\\phi _ { f c } ( \\alpha )$ ,respectively. Given an input graph $G$ ,Equations (2) and (3) identify one LDB from $\\mathcal { H }$ .Thus, we can sample a subset of input graphs uniformly from $D$ ,and use Equations (2) and (3) to derive a sample of LDBs as $\\tilde { \\mathcal { H } } \\subset \\mathcal { H }$ ： ", + "bbox": [ + 173, + 400, + 826, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, we substitute $\\mathcal { H }$ in Equation (1) by $\\tilde { \\mathcal { H } }$ to produce the following problem. ", + "bbox": [ + 173, + 463, + 686, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/f52df860ca40a98e65cc25516b494f56cafd46c079811fa063e656225dbb0792.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\mathcal { P } \\subseteq \\tilde { \\mathcal { H } } } g ( \\mathcal { P } , c ) , \\mathrm { ~ s . t . ~ } h ( \\mathcal { P } , c ) \\leq \\delta ,\n$$", + "text_format": "latex", + "bbox": [ + 395, + 484, + 599, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta \\geq 0$ is a tolerance parameter to keep this problem feasible. The parameter $\\delta$ is required because substituting $\\mathcal { H }$ by $\\tilde { \\mathcal { H } }$ ignores the LDBs in $\\mathcal { H } \\backslash \\tilde { \\mathcal { H } }$ . Thus, the convex polytope $r ( \\mathcal { P } , c )$ induced by subset of boundaries in $\\tilde { \\mathcal { H } }$ may contain instances that are not predicted as class $c$ . We directly set $\\boldsymbol { \\delta } \\dot { = } \\boldsymbol { h } ( \\tilde { \\mathcal { H } } , \\boldsymbol { c } )$ , which is the smallest value of $\\delta$ that keeps the practical problem feasible. ", + "bbox": [ + 173, + 517, + 825, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The problem in Equation (4) can be proven to be a Submodular Cost Submodular Cover (SCSC) problem [18] (see Appendix $\\mathbf { D }$ for proof) that is well known to be NP-hard [5]. We adopt a greedy boundary selection method to find a good solution to this problem [40]. Specifically, we initialize $\\mathcal { P }$ as an empty set,and then iteratively select a new boundary $h$ from $\\tilde { \\mathcal { H } }$ by ", + "bbox": [ + 173, + 585, + 825, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/1f9f845a8a01b5db73f4894fc93b9fa1e911127e6b892e443b6edc0c7061df65.jpg", + "text": "$$\nh = \\arg \\operatorname* { m i n } _ { h \\in \\tilde { \\mathcal { H } } \\backslash \\mathcal { P } } \\frac { g ( \\mathcal { P } , c ) - g ( \\mathcal { P } \\cup \\{ h \\} , c ) + \\epsilon } { h ( \\mathcal { P } , c ) - h ( \\mathcal { P } \\cup \\{ h \\} , c ) } ,\n$$", + "text_format": "latex", + "bbox": [ + 357, + 648, + 638, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g ( \\mathcal { P } , c ) - g ( \\mathcal { P } \\cup \\{ h \\} , c )$ is the decrease of $g ( \\mathcal { P } , c )$ when adding $h$ into $\\mathcal { P }$ ,and $h ( \\mathcal { P } , c ) - h ( \\mathcal { P } \\cup$ $\\{ h \\} , c \\bar { ) }$ is the decrease of $h ( \\mathcal { P } , c )$ when adding $h$ into $\\mathcal { P }$ .Both $g ( \\mathcal { P } , c )$ and $h ( \\mathcal { P } , c )$ are non-increasing when adding $h \\in \\tilde { \\mathcal { H } }$ into $\\mathcal { P }$ because adding a new boundary $h$ may only exclude some graphs from the convex polytope $r ( \\mathcal { P } , c )$ ： ", + "bbox": [ + 173, + 693, + 825, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, in each iteration,Equation (5) selects a boundary $h \\in \\tilde { \\mathcal { H } }$ such that adding $h$ into $\\mathcal { P }$ reduces $g ( \\mathcal { P } , c )$ the least and reduces $\\bar { h } ( \\mathcal { P } , c )$ the most. In this way,we can quickly reduce $h ( \\mathcal { P } , c )$ to be smaller than $\\delta$ without decreasing $g ( \\mathcal { P } , c )$ too much, which produces a good feasible solution to the practical problem.We add a small constant $\\epsilon$ to the numerator such that, when there are multiple candidates of $h$ that do not decrease $g ( \\mathcal { P } , c )$ , we can still select the $h$ that reduces $h ( \\mathcal { P } , c )$ the most. ", + "bbox": [ + 173, + 757, + 825, + 827 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply a peeling-off strategy to iteratively extract multiple decision regions.For each class $c \\in C$ we first solve the practical problem once to find a decision region $r ( \\mathcal { P } , c )$ , then we remove the graphs covered by $r ( \\mathcal { P } , \\bar { c } )$ from $D _ { c }$ . If there are remaining graphs predicted as the class $c$ , we continue finding the decision regions using the remaining graphs until all the graphs in $D _ { c }$ are removed. When all the graphs in $D _ { c }$ are removed for each class $c \\in C$ , we stop the iteration and return the set of decision regions we found. ", + "bbox": [ + 173, + 830, + 825, + 911 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3Producing Explanations ", + "text_level": 1, + "bbox": [ + 174, + 90, + 380, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce how to use the LDBs of decision regions to train a neural network that produces a robust counterfactual explanation as a small subset of edges of an input graph. We form explanations as a subset of edges because GNNs make decisions by aggregating messages passed on edges. Using edges instead of vertices as explanations can provide beter insights on the decision logic of GNNs. ", + "bbox": [ + 173, + 116, + 825, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1The Neural Network Model ", + "text_level": 1, + "bbox": [ + 174, + 196, + 416, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Denote by $f _ { \\theta }$ the neural network to generate a subset of edges of an input graph $G$ as the robust counterfactual explanation on the prediction $\\phi ( G )$ ： $\\theta$ represents the set of parameters of the neural network. For experiments, our explanation network $f$ consists of 2 fully connected layers with a ReLU activation and the hidden dimension of 64. ", + "bbox": [ + 173, + 219, + 825, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For any two connected vertices $v _ { i }$ and $v _ { j }$ of $G$ ,denote by $\\mathbf { z } _ { i }$ and $\\mathbf { z } _ { j }$ the embeddings produced by the last convolution layer of the GNN for the two vertices, respectively. The neural network $f _ { \\theta }$ takes $\\mathbf { z } _ { i }$ and $\\mathbf { z } _ { j }$ as the input and outputs the probability for the edge between $v _ { i }$ and $v _ { j }$ to be part of the explanation. This can be written as ", + "bbox": [ + 173, + 279, + 825, + 333 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/dd0dc1b00ca37ef79ae9976d880a0ec02cf8d2aade732ccacacb9c981b208e29.jpg", + "text": "$$\n\\begin{array} { r } { { \\bf M } _ { i j } = f _ { \\theta } ( { \\bf z } _ { i } , { \\bf z } _ { j } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 436, + 342, + 560, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf { M } _ { i j }$ denotes the probability that the edge between $v _ { i }$ and $v _ { j }$ is contained in the explanation. When there is no edge between $v _ { i }$ and $v _ { j }$ ,that is, ${ \\bf A } _ { i j } = 0$ ,we set $\\mathbf { \\check { M } } _ { i j } = 0$ ", + "bbox": [ + 169, + 367, + 825, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For an input graph $G = \\{ V , E \\}$ with $n$ vertices and a trained neural network $f _ { \\boldsymbol { \\theta } } , \\mathbf { M }$ is an $n$ -by- $\\mathbf { \\nabla } \\cdot n$ matrix that carries the complete information to generate a robust counterfactual explanation as a subset of edges, denoted by $S \\subseteq E$ . Concretely,we obtain $S$ by selecting all the edges in $E$ whose corresponding entries in $\\mathbf { M }$ are larger than O.5. ", + "bbox": [ + 173, + 400, + 825, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2Training Model $f _ { \\theta }$ ", + "text_level": 1, + "bbox": [ + 174, + 468, + 352, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For an input graph $G = ( V , E )$ , denote by $S \\subseteq E$ the subset of edges produced by $f _ { \\theta }$ to explain the prediction $\\phi ( G )$ ,our goal is to train a good model $f _ { \\theta }$ such that the prediction on the subgraph $G _ { S }$ induced by $S$ from $G$ is consistent with $\\phi ( G )$ ; and deleting the edges in $S$ from $G$ produces a remainder subgraph $G _ { E \\backslash S }$ such that the prediction on $G _ { E \\backslash S }$ changes significantly from ${ \\bar { \\phi } } ( G )$ ： ", + "bbox": [ + 174, + 491, + 825, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since producing $S$ by $f _ { \\theta }$ is a discrete operation that is hard to incorporate in an end-to-end training process, we define two proxy graphs to approximate $G _ { S }$ and $G _ { E \\backslash S }$ ,respectively,such that the proxy graphs are determined by $\\theta$ through continuous functions that can be smoothly incorporated into an end-to-end training process. ", + "bbox": [ + 174, + 553, + 825, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proxy graph of $G _ { S }$ , denoted by $G _ { \\theta }$ , is defined by regarding $\\mathbf { M }$ instead of $\\mathbf { A }$ as the adjacency matrix. That is, $G _ { \\theta }$ has exactly the same graph structure as $G$ , but the edge weights of $G _ { \\theta }$ is given by the entries in $\\mathbf { M }$ instead of $\\mathbf { A }$ .Here,the subscript $\\theta$ means $G _ { \\theta }$ is determined by $\\theta$ ", + "bbox": [ + 173, + 613, + 825, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proxy graph of $G _ { E \\backslash S }$ , denoted by $G _ { \\theta } ^ { \\prime }$ , also have the same graph structure as $G$ , but the edge weight between each pair of vertices $v _ { i }$ and $v _ { j }$ is defined as ", + "bbox": [ + 173, + 660, + 821, + 688 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/afaae5efb14f27284c6a4b7446792ec5cf01118d4490a60597c99ebbd88ab0af.jpg", + "text": "$$\n\\begin{array} { r } { \\mathbf { M } _ { i j } ^ { \\prime } = \\left\\{ { 1 - \\mathbf { M } _ { i j } \\quad \\mathrm { ~ i f ~ } \\mathbf { A } _ { i j } = 1 } \\atop { 0 } \\right. } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 390, + 694, + 607, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The edge weights of both $G _ { \\theta }$ and $G _ { \\theta } ^ { \\prime }$ are determined by $\\theta$ through continuous functions, thus we can smoothly incorporate $G _ { \\theta }$ and $G _ { \\theta } ^ { \\prime }$ into an end-to-end training framework. ", + "bbox": [ + 169, + 743, + 823, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As discussed later in this section, we use a regularization term to force the value of each entry in $\\mathbf { M } _ { i j }$ to be close to eitherOor 1,such that $G _ { \\theta }$ and $\\operatorname { \\mathrm { \\Sigma } } _ { G _ { \\theta } ^ { \\prime } } ^ { G \\prime }$ better approximate $G _ { S }$ and $G _ { E \\backslash S }$ respectively. ", + "bbox": [ + 171, + 777, + 823, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We formulate our loss function as ", + "bbox": [ + 174, + 810, + 397, + 824 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/cfc41b2d76850caa450e5da3d6381caaa5c3ed3056969b987fa0fd0bf367d283.jpg", + "text": "$$\n\\mathcal { L } ( \\theta ) = \\sum _ { G \\in D } \\left\\{ \\lambda \\mathcal { L } _ { s a m e } ( \\theta , G ) + ( 1 - \\lambda ) \\mathcal { L } _ { o p p } ( \\theta , G ) + \\beta \\mathcal { R } _ { s p a r s e } ( \\theta , G ) + \\mu \\mathcal { R } _ { d i s c r e t e } ( \\theta , G ) \\right\\} ,\n$$", + "text_format": "latex", + "bbox": [ + 184, + 829, + 790, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda \\in [ 0 , 1 ]$ $\\beta \\geq 0$ and $\\mu \\geq 0$ are the hyperparameters controlling the importance of each term. The influence of these parameters is discussed in Appendix $\\mathbf { G }$ .The first term of our loss function requires that the prediction of the GNN on $G _ { \\theta }$ is consistent with the prediction on $G$ . Intuitively, this means that the edges with larger weights in $G _ { \\theta }$ dominate the prediction on $G$ .We formulate this term by requiring $G _ { \\theta }$ to be covered by the same decision region covering $G$ ： ", + "bbox": [ + 176, + 871, + 825, + 911 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 90, + 823, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Denote by $\\mathcal { H } _ { G }$ the set of LDBs that induce the decision region covering $G$ ,and by $| \\mathcal { H } _ { G } |$ the number of LDBs in $\\mathcal { H } _ { G }$ . For the $i$ -th LDB $h _ { i } \\in \\mathcal { H } _ { G }$ , denote by $B _ { i } ( \\mathbf { x } ) = \\mathbf { w } _ { i } ^ { \\top } \\mathbf { x } + b _ { i }$ , where $\\mathbf { w } _ { i }$ and $b _ { i }$ are the basis and bias of $h _ { i }$ , respectively, and $\\mathbf { x } \\in \\mathbb { O } ^ { d }$ is a point in the space $\\mathbb { O } ^ { d }$ . The sign of $B _ { i } ( { \\bf x } )$ indicates whether a point $\\mathbf { x }$ lies on the positive side or the negative side of $h _ { i }$ , and the absolute value $| B _ { i } ( { \\bf x } ) |$ is proportional to the distance of a point $\\mathbf { x }$ from $h _ { i }$ . Denote by $\\sigma ( \\cdot )$ the standard sigmoid function, we formulate the first term of our loss function as ", + "bbox": [ + 173, + 123, + 825, + 208 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/5f3b7a1f439726d3306abd8a0a28ca52ec5b541263656487dafdb921adbdf0a2.jpg", + "text": "$$\n\\mathcal { L } _ { s a m e } ( \\theta , G ) = \\frac { 1 } { | \\mathcal { H } _ { G } | } \\sum _ { h _ { i } \\in \\mathcal { H } _ { G } } \\sigma \\left( - \\mathcal { B } _ { i } ( \\phi _ { g c } ( G ) ) * \\mathcal { B } _ { i } ( \\phi _ { g c } ( G _ { \\theta } ) ) \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 289, + 212, + 707, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "such that minimizing $\\mathcal { L } _ { s a m e } ( \\theta , G )$ encourages the graph embeddings $\\phi _ { g c } ( G )$ and $\\phi _ { g c } ( G _ { \\theta } )$ to lie on the same side of every LDB in $\\mathcal { H } _ { G }$ . Thus, $G _ { \\theta }$ is encouraged to be covered by the same decision region covering $G$ ： ", + "bbox": [ + 174, + 257, + 825, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The second term of our loss function optimizes the counterfactual property of the explanations by requiring the prediction on $G _ { \\theta } ^ { \\prime }$ to be significantly different from the prediction on $G$ . Intuitively, this means that the set of edges with larger weights in $G _ { \\theta }$ are good counterfactual explanations because reducing the weights of these edges significantly changes the prediction. Following the above intuition,we formulate the second term as ", + "bbox": [ + 173, + 304, + 825, + 371 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/4a6e9260c897b62607fd0281735b7c87296a87e988d8b6aacd13c47e8cb7d8c9.jpg", + "text": "$$\n\\mathcal { L } _ { o p p } ( \\theta , G ) = \\operatorname* { m i n } _ { h _ { i } \\in \\mathcal { H } _ { G } } \\sigma \\left( \\mathcal { B } _ { i } ( \\phi _ { g c } ( G ) ) * \\mathcal { B } _ { i } ( \\phi _ { g c } ( G _ { \\theta } ^ { \\prime } ) ) \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 321, + 377, + 674, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "such that minimizing $\\mathcal { L } _ { o p p } ( \\theta , G )$ encourages the graph embeddings $\\phi _ { g c } ( G )$ and $\\phi _ { g c } ( G _ { \\theta } ^ { \\prime } )$ to lie on the opposite sides of at least one LDB in $\\mathcal { H } _ { G }$ . This further means that $G _ { \\theta } ^ { \\prime }$ is encouraged not to be covered by the decision region covering $G$ , thus the prediction on $G _ { \\theta } ^ { \\prime }$ can be changed significantly from the prediction on $G$ ： ", + "bbox": [ + 173, + 409, + 825, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to [45], we use a L1 regularization $\\mathcal { R } _ { s p a r s e } ( \\theta , G ) = \\| \\mathbf { M } \\| _ { 1 }$ on the matrix $\\mathbf { M }$ produced by $f _ { \\theta }$ on an input graph $G$ to produce a sparse matrix M, such that only a small number of edges in $G$ are selected as the counterfactual explanation. We also follow [45] to use an entropy regularization ", + "bbox": [ + 176, + 468, + 825, + 510 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/92c85db5def2d86edeb67eb745a7f02fb8e21bbf2fd6168f0abe3416cc9d7222.jpg", + "text": "$$\n\\mathcal { R } _ { d i s c r e t e } ( \\theta , G ) = - \\frac { 1 } { | \\mathbf { M } | } \\sum _ { i , j } ( \\mathbf { M } _ { i j } \\log ( \\mathbf { M } _ { i j } ) + ( 1 - \\mathbf { M } _ { i j } ) \\log ( 1 - \\mathbf { M } _ { i j } ) )\n$$", + "text_format": "latex", + "bbox": [ + 254, + 515, + 743, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to push the value of each entry in $\\mathbf { M } _ { i j }$ to be close to either Oor 1,such that $G _ { \\theta }$ and $G _ { \\theta } ^ { \\prime }$ approximate $G _ { S }$ and $G _ { E \\backslash S }$ well, respectively. ", + "bbox": [ + 174, + 561, + 825, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Now we can use the graphs in $D$ and the extracted decision regions to train the neural network $f _ { \\theta }$ in an end-to-end manner by minimizing $\\mathcal { L } ( \\boldsymbol { \\theta } )$ over $\\theta$ using back propagation. Once we finish training $f _ { \\theta }$ , we can first apply $f _ { \\theta }$ to produce the matrix $\\mathbf { M }$ for an input graph $G = ( V , E )$ ,and then obtain the explanation $S$ by selecting all the edges in $E$ whose corresponding entries in $\\mathbf { M }$ are larger than 0.5. We do not need the extracted boundaries for inference as the the decision logic of GNN is already distilled into the explanation network $f$ during the training. ", + "bbox": [ + 173, + 595, + 825, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed in Appendix B,our method can be easily extended to generate robust counterfactual explanations for node classification tasks. ", + "bbox": [ + 174, + 681, + 823, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method is highly efficient with a time complexity $O ( | E | )$ for explaining the prediction on an input graph $G$ , where $| E |$ is the total number of edges in $G$ : Additionally, the neural network $f _ { \\theta }$ can be directly used without retraining to predict explanations on unseen graphs. Thus our method is significantly faster than the other methods [45,32,47,38] that require retraining each time when generating explanations on a new input graph. ", + "bbox": [ + 174, + 714, + 825, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5Experiments ", + "text_level": 1, + "bbox": [ + 174, + 800, + 312, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct series of experiments to compare our method with the state-of-the-art methods including GNNExplainer [45],PGExplainer [25],PGM-Explainer [38], SubgraphX [47] and CFGNNExplainer [24]. For the methods that identify a set of vertices as an explanation, we use the set of vertices to induce a subgraph from the input graph,and then use the set of edges of the induced subgraph as the explanation. For the methods that identify a subgraph as an explanation, we directly use the set of edges of the identified subgraph as the explanation. ", + "bbox": [ + 174, + 832, + 825, + 911 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7fd56d0939a4af57dc11439ffc5d8bfba231262c91ca025fc019f5b6e33dc49.jpg", + "image_caption": [ + "Figure 1: Fidelity performance averaged across 10 runs for the datasets at different levels of sparsity. " + ], + "image_footnote": [], + "bbox": [ + 271, + 88, + 725, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the decision regions,we derive another baseline method named RCExp-NoLDB that adopts the general framework of RCExplainer but does not use the LDBs of decision regions to generate explanations.Instead, RCExp-NoLDB directly maximizes the prediction confidence on class $c$ for $G _ { \\theta }$ and minimizes the prediction confidence of class $c$ for $G _ { \\theta } ^ { \\prime }$ ： ", + "bbox": [ + 174, + 263, + 825, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the explanation performance on two typical tasks: the graph clasification task that uses a GNN to predict the class of an input graph,and the node classification task that uses a GNN to predict the class of a graph node. For the graph classification task, we use one synthetic dataset, BA-2motifs [25],and two real-world datasets,Mutagenicity [21] and NCI1 [39].For the node classification task, we use the same four synthetic datasets as used by GNNExplainer [45], namely, BA-SHAPES,BA-COMMUNITY,TREE-CYCLES and TREE-GRID. ", + "bbox": [ + 173, + 324, + 825, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limited by space, we only report here the key results_on the graph classification task for fidelity, robustness and efficiency. Please refer to Appendix E for details on datasets, baselines and the experiment setups.Detailed experimental comparison on the node classification task willbe discussed in Appendix F where we show that our method produces extremely accurate explanations. The code3 is publicly available. ", + "bbox": [ + 174, + 410, + 825, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1Fidelity ", + "text_level": 1, + "bbox": [ + 173, + 506, + 264, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fidelity is measured by the decrease of prediction confidence after removing the explanation (i.e., a set of edges) from the input graph [32]. We use fidelity to evaluate how counterfactual the generated explanations are on the datasets Mutagenicity,NCI1 and BA-2motifs.A large fidelity score indicates stronger counterfactual characteristics. It is important to note that fidelity may be sensitive to sparsity of explanations. The sparsity of an explanation $S$ with respect to an input graph $G = ( \\bar { V , } E )$ is $\\begin{array} { r } { s p a r s i t y ( S , G ) = 1 - \\frac { | S | } { | E | } } \\end{array}$ ,thatis,tpeeefdgafrta from $G$ . We only compare explanations with the same level of sparsity. ", + "bbox": [ + 174, + 535, + 825, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 1 shows the results about fidelity. Our approach achieves the best fidelity performance at all levels of sparsity. The results validate the effectiveness of our method in producing highly counterfactual explanations. RCExplainer also significantly outperforms RCExp-NoLDB. This confirms that using LDBs of decision regions extracted from GNNs produces more faithful counterfactual explanations. ", + "bbox": [ + 174, + 642, + 825, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CF-GNNExplainer performs the best among the rest of the methods.This is expected as it optimizes the counterfactual behavior of the explanations which results in higher fidelity for the explanations in comparison to those produced by other methods such as GNNExplainer and PGExplainer. ", + "bbox": [ + 176, + 714, + 823, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The fidelity performance of SubgraphX reported in [47] was obtained by seting the features of nodes that are part of the explanation to Obut not removing the explanation edges from the input graph. This does not remove the message passing roles of the explanation nodes from the input graph because the edges connected to those nodes still can pass messages. In our experiments, we directly block the messages that are passed on the edges in the explanation, which completely prevents the explanation nodes in the input graph to participate in the message passing. As a result, the performance of SubgraphX drops significantly. ", + "bbox": [ + 174, + 760, + 825, + 854 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1aa166e7cf33380e4add83effc9273d8301d54224aa1859e6e8d4478bf17806a.jpg", + "image_caption": [ + "Figure 2: Noise robustness (AUC) averaged across 10 runs for the datasets at diferent levels of noise. " + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 725, + 188 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2Robustness Performance ", + "text_level": 1, + "bbox": [ + 174, + 247, + 385, + 261 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this experiment, we evaluate the robustness of all methods by quantifying how much an explanation changes after adding noise to the input graph. For an input graph $G$ and the explanation $S$ ,we produce a perturbed graph $G ^ { \\prime }$ by adding random noise to the node features and randomly adding or deleting some edges of the input graph such that the prediction on $G ^ { \\prime }$ is consistent with the prediction on $G$ ： Using the same method we obtain the explanation $S ^ { \\prime }$ on $G ^ { \\prime }$ . Considering top- $k$ edges of $S$ as the ground-truth and comparing $S ^ { \\prime }$ against them, we compute a receiver operating characteristic (ROC) curve and evaluate the robustnessby the area under curve (AUC) of the ROC curve. We report results for $k = 8$ in Figure 2. Results for other values of $k$ are included in Appendix F where we observe similar trend. ", + "bbox": [ + 173, + 271, + 826, + 391 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 2 shows the AUC of GNNExplainer, PGExplainer, RCExp-NoLDB and RCExplainer at different levels of noise. A higher AUC indicates better robustness. The percentage of noise shows the proportion of nodes and edges that are modified. Baselines such as PGM-Explainer and SubgraphX are not included in this experiment as they do not output the edge weights that are required for computing AUC.We present additional robustness experiments in Appendix F where we extend all the baselines to report node and edge level accuracy. ", + "bbox": [ + 173, + 396, + 825, + 477 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "GNNExplainer performs the worst on most of the datasets,since it optimizes each graph independently without considering other graphs in the training set. Even when no noise is added,the AUC of GNNExplainer is significantly lower than 1 because different runs produce different explanations for the same graph prediction. PGExplainer is generally more robust than GNNExplainer because the neural network they trained to produce explanations implicitly considers all the graphs used for training. CF-GNNExplainer also performs worse than RCExplainer, which means it is more susceptible to the noise as compared to RCExplainer. ", + "bbox": [ + 173, + 483, + 825, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our method achieves the best AUC on allthe datasets, because the common decision logic carried by the decision regions of a GNN is highly robust to noise. PGExplainer achieves a comparable performance as our method on the Mutagenicity dataset, because the samples of this dataset share a lot of common structures such as carbon rings, which makes it easier for the neural network trained by PGExplainer to identify these structures in presence of noise. However, for BA-2motifs and NCI1, this is harder as samples share very few structures and thus the AUC of PGExplainer drops significantly. RCExplainer also significantly outperforms RCExp-NoLDB on these datasets which highlights the role of decision boundaries in making our method highly robust. ", + "bbox": [ + 173, + 582, + 825, + 689 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/747a5f1226941eaba594797ad5d8f60f7224157469cea7e8953b63d2c1a53516.jpg", + "table_caption": [ + "Table 1: Average time cost for producing an explanation on a single graph sample. " + ], + "table_footnote": [], + "table_body": "
MethodGNNExplainerPGExplainerPGM-ExplainerSubgraphXCF-GNNExplainerRCExplainer
Time1.2s ± 0.20.01s ± 0.0313.1s ± 3.977.8s ± 4.54.6s± 0.20.01s ± 0.02
", + "bbox": [ + 173, + 702, + 823, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effciency. We evaluate efficiency by comparing the average computation time taken for inference on unseen graph samples. Table 1 shows the results on the Mutagenicity dataset. Since our method also can be directly used for unseen data without any retraining,it is as eficient as PGExplainer and significantly faster than GNNExplainer, PGM-Explainer, SubgraphX and CF-GNNExplainer. ", + "bbox": [ + 173, + 781, + 825, + 835 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6Conclusion ", + "text_level": 1, + "bbox": [ + 174, + 853, + 299, + 869 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we develop a novel method for producing counterfactual explanations on GNNs. We extract decision boundaries from the given GNN model to formulate an intuitive and effective counterfactual loss function. We optimize this loss to train a neural network to produce explanations with strong counterfactual characteristics. Since the decision boundaries are shared by multiple samples of the same predicted class, explanations produced by our method are robust and do not overfit the noise. Our experiments on synthetic and real-life benchmark datasets strongly validate the efficacy of our method. In this work,we focus on GNNs that belong to Piecewise Linear Neural Networks (PLNNs). Extending our method to other families of GNNs and tasks such as link prediction,remains an interesting future direction. ", + "bbox": [ + 173, + 883, + 823, + 911 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 90, + 825, + 184 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our method will benefit multiple fields where GNNs are intensively used. By allowing the users to interpret the predictions of complex GNNs better, it will promote transparency, trust and fairness in the society. However, there also exist some inherent risks. A generated explanation may expose private information if our method is not coupled with an adequate privacy protection technique. Also, some of the ideas presented in this paper may be adopted and extended to improve adversarial attacks. Without appropriate defense mechanisms,the misuse of such attcks poses a risk of disruption in the functionality of GNNs deployed in the real world.That said, we firmly believe that these risks can be mitigated through increased awareness and proactive measures. ", + "bbox": [ + 173, + 190, + 825, + 296 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Acknowledgments and Disclosure of Funding ", + "text_level": 1, + "bbox": [ + 174, + 315, + 553, + 333 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lingyang Chu's research is supported in part by the startup grant provided by the Department of Computing and Software of McMaster University. All opinions, findings,conclusions,and recommendations in this paper are those of the authors and do not necessarily reflect the views of the funding agencies. ", + "bbox": [ + 174, + 347, + 825, + 401 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "References ", + "text_level": 1, + "bbox": [ + 174, + 420, + 266, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "[1] J. Adebayo,J. Gilmer,M.Muelly,I. Goodfellow,M. Hardt,and B. Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems. \n[2] A.Bojchevski and S. Gunnemann. Certifiable robustness to graph perturbations.arXiv preprint arXiv:1910.14356,2019. \n[3]E. Cho,S.A. Myers,and J.Leskovec.Friendship and mobility: user movement in location-based social networks. In Proceedings of the 17th ACM SIGkDD international conference on Knowledge discovery and data mining,pages 1082-1090,2011. \n[4]L.Chu, X. Hu,J. Hu,L. Wang,and J.Pei. Exact and consistent interpretation for piecewise linear neural networks: A closed form solution. In Proceedings of the 24th ACM SiGKDD International Conference on Knowledge Discovery & Data Mining,pages 1244-1253,2018. \n[5]V. Crawford,A. Kuhnle,and M. Thai. Submodular cost submodular cover with an approximate oracle.In International Conference on Machine Learning, pages 1426-1435.PMLR,2019. \n[6] A. K. Debnath, R.L.Lopez de Compadre, G. Debnath, A. J. Shusterman,and C. Hansch. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds.correlation with molecular orbital energies and hydrophobicity. Journal of medicinal chemistry,34(2):786-797,1991. \n[7]M.Denil,S.G.Colmenarejo,S. Cabi,D.Saxton,and N. de Freitas. Programmable agents. arXiv preprint arXiv:1706.06383,2017. \n[8]Y. Duan, M. Andrychowicz,B.C. Stadie,J. Ho,J.Schneider,I.Sutskever,P.Abbeel,and W. Zaremba. One-shot imitation learning. arXiv preprint arXiv:1703.07326,2017. \n[9]W.Fan,Y.Ma,Q.Li,Y. He,E. Zhao,J.Tang,and D.Yin. Graph neural networks for social recommendation. In The World Wide Web Conference,pages 417-426,2019. \n[10]T.Freiesleben.Counterfactualexplanations &adversarialexamplescommongrounds,essntialdifferences, and potential transfers. arXiv preprint arXiv:2009.05487, 2020. \n[11]A. Ghorbani,A. Abid,and J. Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3681-3688,2019. \n[12] X. Glorot and Y. Bengio. Understanding the dificulty of training deep feedforward neural networks. InProceedingsofthethirteenth international conference onartificial intelligence and statistics,pages 249-256. JMLR Workshop and Conference Proceedings, 2010. \n[13] X.Glorot,A.Bordes,and Y.Bengio.Deep sparse rectifier neural networks.In Proceedings ofthe fourteenth international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and Conference Proceedings,2011. \n[14] I. Goodfelow,D. Warde-Farley,M. Mirza,A. Courvile,and Y.Bengio. Maxout networks.In International conference on machine learning, pages 1319-1327.PMLR,2013. \n[15]K. He, X. Zhang, S. Ren,and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings ofthe IEEE international conferenceon computer vision, pages 1026-1034,2015. \n[16] L.Holdijk,M. Boon,S.Henckens,and L.de Jong. [re] parameterized explainer for graph neural network. In ML Reproducibility Challenge 2020,2021. \n[17] W. Huber, V.J. Carey,L.Long, S. Falcon,and R. Gentleman. Graphs in molecular biology. BMC bioinformatics,8(6):1-14,2007. \n[18] R. K. Iyer and J. A. Bilmes. Submodular optimization with submodular cover and submodular knapsack constraints.In Advances in Neural Information Processing Systems. \n[19]M. Jiang,Z. Li, S. Zhang,S. Wang,X. Wang, Q. Yuan,and Z. Wei. Drug-target affinity prediction using graph neural network and contact maps.RsC Advances,10(35):20701-20712, 2020. \n[20] H. Jin and X. Zhang. Latent adversarial training of graph convolution networks. In ICML Workshop on Learning and Reasoning with Graph-Structured Representations,2019. \n[21] J. Kazius,R. McGuire,and R. Bursi. Derivation and validation of toxicophores for mutagenicity prediction. Journal of medicinal chemistry,48(1):312-320,2005. \n[22] T.N. Kipf and M. Welling. Semi-supervised classfication with graph convolutional networks. In 5th International Conference on Learning Representations,ICLR 2Ol7, Conference Track Proceedings. \n[23]M. Liu, Y. Luo,L. Wang, Y. Xie, H. Yuan,S. Gui, H. Yu,Z. Xu, J. Zhang, Y. Liu, K. Yan, H. Liu, C. Fu, B. Oztekin,X. Zhang,and S.Ji. DIG: A turnkey library for diving into graph deep learning research. arXiv preprint arXiv:2103.12608,2021. \n[24] A.Lucic,M.ter Hoeve,G. Tolomei,M.de Rijke,and F. Silvestri. Cf-gnnexplainer: Counterfactual explanations for graph neural networks.arXiv preprint arXiv:2102.03322,2021. \n[25] D.Luo, W. Cheng,D. Xu, W. Yu, B. Zong, H. Chen,and X. Zhang. Parameterized explainer for graph neural network. In Advances in Neural Information Processing Systems. \n[26] C. Molnar.Interpretable Machine Learning.2019.https://christophm.github.io/ interpretable-ml-book/. \n[27] G.Montufar,R.Pascanu,K.Cho,and Y.Bengio.On the numberof linear regions of deep neural networks. arXiv preprint arXiv:1402.1869,2014. \n[28]R.Moraffah,M. Karami,R.Guo,A. Raglin,and H.Liu.Causal interpretability for machine learningproblems,methods and evaluation. ACM SIGKDD Explorations Newsleter,22(1):18-33,2020. \n[29] V.Nair and G.E.Hinton.Rectified linear units improve restricted boltzmann machines.In Icml,2010. \n[30] D. Numeroso and D. Bacciu. Meg: Generating molecular counterfactual explanations for deep graph networks. arXiv preprint arXiv:2104.08060,2021. \n[31] D.E. Pires,T. L. Blundel, and D. B. Ascher. pkcsm: predicting small-molecule pharmacokinetic and toxicity properties using graph-based signatures. Journal of medicinal chemistry,58(9):4066-4072,2015. \n[32]P.E.Pope,S. Kolouri, M.Rostami, C.E. Martin,and H. Hoffmann.Explainability methods for graph convolutional neural networks.In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10772-10781,2019. \n[33]R.R.Selvaraju,M.Cogswell,A.Das,R. Vedantam,D.Parikh,and D.Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. \n[34] P.L. Simon, M.Taylor,and I. Z. Kiss. Exact epidemic models on graphs using graph-automorphism driven lumping. Journal of mathematical biology,62(4):479-508,2011. \n[35] K. Simonyan, A. Vedaldi,and A. Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. ICLR,2014. \n[36]K.Sokol and P.A.Flach. Counterfactual explanations of machine learning predictions: opportunities and challenges for ai safety. In SafeAI@ AAAI, 2019. \n[37]P. Velickovic, G. Cucurul,A. Casanova,A. Romero,P.Lio,and Y. Bengio. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2O18, Conference Track Proceedings. \n[38]M. Vu and M. T. Thai.Pgm-explainer: Probabilistic graphical model explanations for graph neural networks.In Advances in Neural Information Processing Systems. \n[39] N. Wale and G. Karypis. Comparison of descriptor spaces for chemical compound retrieval and classification. In Sixth International Conference on Data Mining (ICDM'06),pages 678-689,2006.doi: 10.1109/ICDM.2006.39. \n[40] L.A.Wolsey. An analysis of the greedy algorithm for the submodular setcovering problem. Combinatorica, 2(4):385-393,1982. \n[41] J. Xiong,Z. Xiong, K. Chen,H. Jiang,and M. Zheng. Graph neural networks for automated de novo drug design. Drug Discovery Today,2021. \n[42]H. Xu, Y. Ma, H.-C.Liu,D. Deb, H. Liu, J.-L. Tang,and A. K. Jain. Adversarial atacks and defenses in images, graphs and text: A review. International Journal of Automation and Computing,17(2):151-178, 2020. \n[43] K. Xu, H. Chen,S.Liu,P.-Y. Chen, T.-W. Weng,M. Hong,and X. Lin. Topology attack and defense for graph neural networks: An optimization perspective. arXiv preprint arXiv:1906.04214, 2019. \n[44] R. Yin,K.Li, G. Zhang,and J.Lu. A deeper graph neural network for recommender systems. KnowledgeBased Systems,185:105020,2019. \n[45] Z. Ying,D. Bourgeois, J. You, M. Zitnik,and J.Leskovec. Gnnexplainer: Generating explanations for graph neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019. \n[46] H.Yuan,J.Tang,X.Hu,and S.Ji. Xgnn: Towards model-level explanations of graph neural networks.In Proceedings ofthe 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 430-438,2020. \n[47] H. Yuan, H. Yu, J. Wang, K. Li, and S.Ji._On explainability of graph neural networks via subgraph explorations. arXiv preprint arXiv:2102.05152,2021. \n[48] M.D. Zeiler and R.Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818-833. Springer,2014. \n[49] J. Zhang,S.A. Bargal,Z.Lin,J.Brandt,X. Shen,and S.Sclarof. Top-down neural attention by excitation backprop. International Journal of Computer Vision,126(10):1084-1102,2018. \n[50] M. Zhang and Y. Chen. Link prediction based on graph neural networks. In Advances in Neural Information Processing Systems. \n[51] D. Zugner and S. Gunnemann. Adversarial attacks on graph neural networks via meta learning. arXiv preprint arXiv:1902.08412, 2019. \n[52] D.Zügner and S.Gunnemann.Certifable robustness and robust training for graph convolutional networks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,pages 246-256,2019. \n[53] D. Zugner, A. Akbarnejad,and S. Gunnemann. Adversarial atacks on neural networks for graph data. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2847-2856,2018. ", + "bbox": [ + 178, + 438, + 826, + 912 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 83, + 828, + 916 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 92, + 828, + 765 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_middle.json b/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..a3e7a76661a5fbdc1f1783c1b3c6fe36c45d6b82 --- /dev/null +++ b/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_middle.json @@ -0,0 +1,43510 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 132, + 97, + 479, + 135 + ], + "lines": [ + { + "bbox": [ + 129, + 95, + 480, + 119 + ], + "spans": [ + { + "bbox": [ + 129, + 95, + 480, + 119 + ], + "score": 0.999, + "content": "Robust Counterfactual Explanations on Graph", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 240, + 116, + 370, + 136 + ], + "spans": [ + { + "bbox": [ + 240, + 116, + 370, + 136 + ], + "score": 0.999, + "content": "Neural Networks", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 177, + 173, + 434, + 198 + ], + "lines": [ + { + "bbox": [ + 176, + 172, + 437, + 188 + ], + "spans": [ + { + "bbox": [ + 176, + 172, + 289, + 188 + ], + "score": 0.976, + "content": "Mohit Bajaj1* Lingyang", + "type": "text" + }, + { + "bbox": [ + 289, + 174, + 318, + 186 + ], + "score": 0.46, + "content": "\\mathbf { C h u ^ { 2 * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 172, + 437, + 188 + ], + "score": 0.975, + "content": "Zi Yu Xue1,3Jian Pei4", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 184, + 422, + 200 + ], + "spans": [ + { + "bbox": [ + 187, + 184, + 422, + 200 + ], + "score": 0.996, + "content": "Lanjun Wang1Peter Cho-Ho Lam1Yong Zhang1", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 131, + 207, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 166, + 206, + 438, + 221 + ], + "spans": [ + { + "bbox": [ + 166, + 207, + 330, + 221 + ], + "score": 0.967, + "content": "1Huawei Technologies Canada Co., Ltd.", + "type": "text" + }, + { + "bbox": [ + 345, + 206, + 438, + 221 + ], + "score": 0.999, + "content": "2McMaster University", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 217, + 444, + 233 + ], + "spans": [ + { + "bbox": [ + 171, + 217, + 326, + 233 + ], + "score": 0.977, + "content": "3 The University of British Columbia", + "type": "text" + }, + { + "bbox": [ + 339, + 218, + 444, + 232 + ], + "score": 0.991, + "content": "4Simon Fraser University", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 230, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 243 + ], + "score": 0.984, + "content": "{mohit.bajaj1，zi.yu.xue，lanjun.wang，cho.ho.lam，yong.zhang3}@huawei.com", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 241, + 383, + 252 + ], + "spans": [ + { + "bbox": [ + 229, + 241, + 383, + 252 + ], + "score": 0.999, + "content": "chul9@mcmaster.ca,jpei@cs.sfu.ca", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 283, + 279, + 328, + 292 + ], + "lines": [ + { + "bbox": [ + 282, + 279, + 330, + 293 + ], + "spans": [ + { + "bbox": [ + 282, + 279, + 330, + 293 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 143, + 303, + 469, + 470 + ], + "lines": [ + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "score": 0.987, + "content": "Massive deployment of Graph Neural Networks (GNNs) in high-stake applications", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 313, + 470, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 470, + 326 + ], + "score": 0.998, + "content": " generates a strong demand for explanations that are robust to noise and align well", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 323, + 469, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 337 + ], + "score": 0.997, + "content": "with human intuition. Most existing methods generate explanations by identifying", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 335, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 470, + 347 + ], + "score": 0.989, + "content": "a subgraph of an input graph that has a strong correlation with the prediction.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 345, + 470, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 470, + 357 + ], + "score": 0.997, + "content": "These explanations are not robust to noise because independently optimizing the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 356, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 356, + 470, + 367 + ], + "score": 0.98, + "content": "correlation for a single input can easily overfit noise.Moreover, they are not", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 366, + 470, + 378 + ], + "spans": [ + { + "bbox": [ + 142, + 366, + 470, + 378 + ], + "score": 0.997, + "content": "counterfactual because removing an identified subgraph from an input graph does", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 376, + 470, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 470, + 389 + ], + "score": 0.984, + "content": " not necessarily change the prediction result. In this paper, we propose a novel", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 385, + 470, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 385, + 470, + 399 + ], + "score": 0.985, + "content": "method to generate robust counterfactual explanations on GNNs by explicitly", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 396, + 470, + 409 + ], + "spans": [ + { + "bbox": [ + 141, + 396, + 470, + 409 + ], + "score": 0.99, + "content": "modelling the common decision logic of GNNs on similar input graphs. Our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 407, + 470, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 470, + 419 + ], + "score": 0.998, + "content": "explanations are naturally robust to noise because they are produced from the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 418, + 470, + 429 + ], + "spans": [ + { + "bbox": [ + 141, + 418, + 470, + 429 + ], + "score": 0.994, + "content": "common decision boundaries of a GNN that govern the predictions of many similar", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 428, + 470, + 440 + ], + "spans": [ + { + "bbox": [ + 142, + 428, + 470, + 440 + ], + "score": 0.992, + "content": "input graphs. The explanations are also counterfactual because removing the set", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 438, + 470, + 451 + ], + "spans": [ + { + "bbox": [ + 141, + 438, + 470, + 451 + ], + "score": 0.99, + "content": "of edges identified by an explanation from the input graph changes the prediction", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 449, + 470, + 461 + ], + "spans": [ + { + "bbox": [ + 141, + 449, + 470, + 461 + ], + "score": 0.995, + "content": "significantly. Exhaustive experiments on many public datasets demonstrate the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 460, + 292, + 470 + ], + "spans": [ + { + "bbox": [ + 141, + 460, + 292, + 470 + ], + "score": 0.99, + "content": "superior performance of our method.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 107, + 489, + 190, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 192, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 192, + 504 + ], + "score": 0.969, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 506, + 525 + ], + "score": 0.983, + "content": "Graph Neural Networks (GNNs) [22,37,50] have achieved great practical successes in many real-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 505, + 536 + ], + "score": 0.983, + "content": "world applications,such as chemistry [31],molecular biology[17],social networks [3] and epidemic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 0.986, + "content": "modelling [34]. For most of these applications, explaining predictions made by a GNN model is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 0.988, + "content": "crucial for establishing trust with end-users,identifying the cause of a prediction,and even discovering", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 555, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 566 + ], + "score": 0.992, + "content": "potential deficiencies of a GNN model before massive deployment. Ideally,an explanation should", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "score": 0.992, + "content": "be able to answer questions like “Would the prediction of the GNN model change if a certain part", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 575, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 588 + ], + "score": 0.981, + "content": "of an input molecule is removed?” in the context of predicting whether an artificial molecule is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 586, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 597 + ], + "score": 0.981, + "content": "active for acertain type of proteins [19,41],“Would an item recommended stillbe recommended if a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 0.988, + "content": "customer had not purchased some other items in the past?” for a GNN built for recommendation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 607, + 172, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 172, + 619 + ], + "score": 0.969, + "content": "systems [9, 44].", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 507, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 507, + 636 + ], + "score": 0.988, + "content": "Counterfactual explanations [28] in the form of “If X had not occurred, Y would not have oc-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 0.983, + "content": "curred”[26] are the principled way to answer such questions and thus are highly desirable for GNNs.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "score": 0.997, + "content": "In the context of GNNs,a counterfactual explanation identifies a small subset of edges of the input", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 506, + 666 + ], + "score": 0.991, + "content": " graph instance such that removing those edges significantly changes the prediction made by the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "score": 0.988, + "content": "GNN. Counterfactual explanations are usually concise and easy to understand [28,36] because they", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 505, + 686 + ], + "score": 0.984, + "content": " align well with the human intuition to describe a causal situation [26]. To make explanations more", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 684, + 505, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 505, + 698 + ], + "score": 0.986, + "content": " trustworthy, the counterfactual explanation should be robust to noise,that is,some slight changes on", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 741 + ], + "lines": [ + { + "bbox": [ + 105, + 729, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 105, + 729, + 386, + 743 + ], + "score": 0.995, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 120, + 704, + 192, + 713 + ], + "lines": [ + { + "bbox": [ + 118, + 701, + 194, + 715 + ], + "spans": [ + { + "bbox": [ + 118, + 701, + 194, + 715 + ], + "score": 0.999, + "content": "*Equal contribution.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 132, + 97, + 479, + 135 + ], + "lines": [ + { + "bbox": [ + 129, + 95, + 480, + 119 + ], + "spans": [ + { + "bbox": [ + 129, + 95, + 480, + 119 + ], + "score": 0.999, + "content": "Robust Counterfactual Explanations on Graph", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 240, + 116, + 370, + 136 + ], + "spans": [ + { + "bbox": [ + 240, + 116, + 370, + 136 + ], + "score": 0.999, + "content": "Neural Networks", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 177, + 173, + 434, + 198 + ], + "lines": [ + { + "bbox": [ + 176, + 172, + 437, + 188 + ], + "spans": [ + { + "bbox": [ + 176, + 172, + 289, + 188 + ], + "score": 0.976, + "content": "Mohit Bajaj1* Lingyang", + "type": "text" + }, + { + "bbox": [ + 289, + 174, + 318, + 186 + ], + "score": 0.46, + "content": "\\mathbf { C h u ^ { 2 * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 172, + 437, + 188 + ], + "score": 0.975, + "content": "Zi Yu Xue1,3Jian Pei4", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 184, + 422, + 200 + ], + "spans": [ + { + "bbox": [ + 187, + 184, + 422, + 200 + ], + "score": 0.996, + "content": "Lanjun Wang1Peter Cho-Ho Lam1Yong Zhang1", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 131, + 207, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 166, + 206, + 438, + 221 + ], + "spans": [ + { + "bbox": [ + 166, + 207, + 330, + 221 + ], + "score": 0.967, + "content": "1Huawei Technologies Canada Co., Ltd.", + "type": "text" + }, + { + "bbox": [ + 345, + 206, + 438, + 221 + ], + "score": 0.999, + "content": "2McMaster University", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 217, + 444, + 233 + ], + "spans": [ + { + "bbox": [ + 171, + 217, + 326, + 233 + ], + "score": 0.977, + "content": "3 The University of British Columbia", + "type": "text" + }, + { + "bbox": [ + 339, + 218, + 444, + 232 + ], + "score": 0.991, + "content": "4Simon Fraser University", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 230, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 243 + ], + "score": 0.984, + "content": "{mohit.bajaj1，zi.yu.xue，lanjun.wang，cho.ho.lam，yong.zhang3}@huawei.com", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 241, + 383, + 252 + ], + "spans": [ + { + "bbox": [ + 229, + 241, + 383, + 252 + ], + "score": 0.999, + "content": "chul9@mcmaster.ca,jpei@cs.sfu.ca", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 130, + 206, + 482, + 252 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 279, + 328, + 292 + ], + "lines": [ + { + "bbox": [ + 282, + 279, + 330, + 293 + ], + "spans": [ + { + "bbox": [ + 282, + 279, + 330, + 293 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 143, + 303, + 469, + 470 + ], + "lines": [ + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "score": 0.987, + "content": "Massive deployment of Graph Neural Networks (GNNs) in high-stake applications", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 313, + 470, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 470, + 326 + ], + "score": 0.998, + "content": " generates a strong demand for explanations that are robust to noise and align well", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 323, + 469, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 337 + ], + "score": 0.997, + "content": "with human intuition. Most existing methods generate explanations by identifying", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 335, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 470, + 347 + ], + "score": 0.989, + "content": "a subgraph of an input graph that has a strong correlation with the prediction.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 345, + 470, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 470, + 357 + ], + "score": 0.997, + "content": "These explanations are not robust to noise because independently optimizing the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 356, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 356, + 470, + 367 + ], + "score": 0.98, + "content": "correlation for a single input can easily overfit noise.Moreover, they are not", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 366, + 470, + 378 + ], + "spans": [ + { + "bbox": [ + 142, + 366, + 470, + 378 + ], + "score": 0.997, + "content": "counterfactual because removing an identified subgraph from an input graph does", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 376, + 470, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 470, + 389 + ], + "score": 0.984, + "content": " not necessarily change the prediction result. In this paper, we propose a novel", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 385, + 470, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 385, + 470, + 399 + ], + "score": 0.985, + "content": "method to generate robust counterfactual explanations on GNNs by explicitly", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 396, + 470, + 409 + ], + "spans": [ + { + "bbox": [ + 141, + 396, + 470, + 409 + ], + "score": 0.99, + "content": "modelling the common decision logic of GNNs on similar input graphs. Our", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 407, + 470, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 470, + 419 + ], + "score": 0.998, + "content": "explanations are naturally robust to noise because they are produced from the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 418, + 470, + 429 + ], + "spans": [ + { + "bbox": [ + 141, + 418, + 470, + 429 + ], + "score": 0.994, + "content": "common decision boundaries of a GNN that govern the predictions of many similar", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 428, + 470, + 440 + ], + "spans": [ + { + "bbox": [ + 142, + 428, + 470, + 440 + ], + "score": 0.992, + "content": "input graphs. The explanations are also counterfactual because removing the set", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 438, + 470, + 451 + ], + "spans": [ + { + "bbox": [ + 141, + 438, + 470, + 451 + ], + "score": 0.99, + "content": "of edges identified by an explanation from the input graph changes the prediction", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 449, + 470, + 461 + ], + "spans": [ + { + "bbox": [ + 141, + 449, + 470, + 461 + ], + "score": 0.995, + "content": "significantly. Exhaustive experiments on many public datasets demonstrate the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 460, + 292, + 470 + ], + "spans": [ + { + "bbox": [ + 141, + 460, + 292, + 470 + ], + "score": 0.99, + "content": "superior performance of our method.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16.5, + "bbox_fs": [ + 141, + 303, + 470, + 470 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 489, + 190, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 192, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 192, + 504 + ], + "score": 0.969, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 513, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 506, + 525 + ], + "score": 0.983, + "content": "Graph Neural Networks (GNNs) [22,37,50] have achieved great practical successes in many real-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 505, + 536 + ], + "score": 0.983, + "content": "world applications,such as chemistry [31],molecular biology[17],social networks [3] and epidemic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 0.986, + "content": "modelling [34]. For most of these applications, explaining predictions made by a GNN model is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 0.988, + "content": "crucial for establishing trust with end-users,identifying the cause of a prediction,and even discovering", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 555, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 566 + ], + "score": 0.992, + "content": "potential deficiencies of a GNN model before massive deployment. Ideally,an explanation should", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "score": 0.992, + "content": "be able to answer questions like “Would the prediction of the GNN model change if a certain part", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 575, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 588 + ], + "score": 0.981, + "content": "of an input molecule is removed?” in the context of predicting whether an artificial molecule is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 586, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 597 + ], + "score": 0.981, + "content": "active for acertain type of proteins [19,41],“Would an item recommended stillbe recommended if a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 0.988, + "content": "customer had not purchased some other items in the past?” for a GNN built for recommendation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 607, + 172, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 172, + 619 + ], + "score": 0.969, + "content": "systems [9, 44].", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 513, + 506, + 619 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 507, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 507, + 636 + ], + "score": 0.988, + "content": "Counterfactual explanations [28] in the form of “If X had not occurred, Y would not have oc-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 0.983, + "content": "curred”[26] are the principled way to answer such questions and thus are highly desirable for GNNs.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 656 + ], + "score": 0.997, + "content": "In the context of GNNs,a counterfactual explanation identifies a small subset of edges of the input", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 654, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 506, + 666 + ], + "score": 0.991, + "content": " graph instance such that removing those edges significantly changes the prediction made by the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "score": 0.988, + "content": "GNN. Counterfactual explanations are usually concise and easy to understand [28,36] because they", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 505, + 686 + ], + "score": 0.984, + "content": " align well with the human intuition to describe a causal situation [26]. To make explanations more", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 684, + 505, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 505, + 698 + ], + "score": 0.986, + "content": " trustworthy, the counterfactual explanation should be robust to noise,that is,some slight changes on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 73, + 505, + 84 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 84 + ], + "score": 0.98, + "content": " an input graph do not change the explanation significantly. This idea aligns well with the notion of", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 0.994, + "content": "robustness discussed for DNN explanations in computer vision domain [11]. According to Ghorbani", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 0.985, + "content": "et al.[11] many interpretations on neural networks are fragile as it is easier to generate adversarial", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "score": 0.997, + "content": "perturbations that produce perceptively indistinguishable inputs that are assigned the same predicted", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "score": 0.982, + "content": "label,yet have very different interpretations.Here,the concepts of “fragile”“robustness”describe", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 124, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 505, + 137 + ], + "score": 0.987, + "content": " the same concept from opposite perspectives. An interpretation is said to be fragile if systematic", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 135, + 507, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 507, + 149 + ], + "score": 0.985, + "content": " perturbations can lead to dramatically different interpretations without changing the label. Otherwise,", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 146, + 256, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 256, + 158 + ], + "score": 0.974, + "content": "the interpretation is said to be robust.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 39, + "bbox_fs": [ + 104, + 621, + 507, + 698 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 156 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 505, + 84 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 84 + ], + "score": 0.98, + "content": " an input graph do not change the explanation significantly. This idea aligns well with the notion of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 0.994, + "content": "robustness discussed for DNN explanations in computer vision domain [11]. According to Ghorbani", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 0.985, + "content": "et al.[11] many interpretations on neural networks are fragile as it is easier to generate adversarial", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "score": 0.997, + "content": "perturbations that produce perceptively indistinguishable inputs that are assigned the same predicted", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "score": 0.982, + "content": "label,yet have very different interpretations.Here,the concepts of “fragile”“robustness”describe", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 124, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 505, + 137 + ], + "score": 0.987, + "content": " the same concept from opposite perspectives. An interpretation is said to be fragile if systematic", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 135, + 507, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 507, + 149 + ], + "score": 0.985, + "content": " perturbations can lead to dramatically different interpretations without changing the label. Otherwise,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 146, + 256, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 256, + 158 + ], + "score": 0.974, + "content": "the interpretation is said to be robust.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "score": 0.99, + "content": "How to produce robust counterfactual explanations on predictions made by general graph neural", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 173, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 183 + ], + "score": 0.99, + "content": "networks is a novel problem that has not been systematically studied before.As to be discussed", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 182, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 194 + ], + "score": 0.986, + "content": " in Section 2, most GNN explanation methods [45,25,46,37,32] are neither counterfactual nor", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 0.989, + "content": " robust. These methods mostly focus on identifying a subgraph of an input graph that achieves a high", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 204, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 215 + ], + "score": 0.992, + "content": "correlation with the prediction result. Such explanations are usually not counterfactual because, due", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 213, + 507, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 507, + 227 + ], + "score": 0.983, + "content": " to the high non-convexity of GNNs, removing a subgraph that achieves a high correlation does not", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 0.987, + "content": " necessarily change the prediction result. Moreover, many existing methods [45,25,37,32] are not", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 234, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 247 + ], + "score": 0.991, + "content": "robust to noise and may change significantly upon slight modifications on input graphs, because", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 0.987, + "content": " the explanation of every single input graph prediction is independently optimized to maximize the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 255, + 467, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 467, + 267 + ], + "score": 0.99, + "content": "correlation with the prediction, thus an explanation can easily overfit the noise in the data.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 272, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 0.986, + "content": "In this paper², we develop RCExplainer, a novel method to produce robust counterfactual explanations", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 0.985, + "content": "on GNNs. The key idea is to first model the common decision logic of a GNN by set of decision", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 293, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 305 + ], + "score": 0.996, + "content": "regions where each decision region governs the predictions on a large number of graphs,and then", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 302, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 317 + ], + "score": 0.987, + "content": " extract robust counterfactual explanations by a deep neural network that explores the decision logic", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "score": 0.983, + "content": "carried by the linear decision boundaries of the decision regions. We make the following contributions.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 329, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "score": 0.979, + "content": "First, we model the decision logic of a GNN by a set of decision regions, where each decision region", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 104, + 339, + 506, + 352 + ], + "score": 0.991, + "content": " is induced by a set of linear decision boundaries of the GNN. We propose an unsupervised method to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 0.995, + "content": " find decision regions for each class such that each decision region governs the prediction of multiple", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 359, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 104, + 359, + 506, + 374 + ], + "score": 0.994, + "content": " graph samples predicted to be the same class. The linear decision boundaries of the decision region", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 371, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 383 + ], + "score": 0.983, + "content": "capture the common decision logic on all the graph instances inside the decision region, thus do", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 0.982, + "content": "not easily overfit the noise of an individual graph instance. By exploring the common decision", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "score": 0.994, + "content": "logic encoded in the linear boundaries, we are able to produce counterfactual explanations that are", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 402, + 213, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 213, + 415 + ], + "score": 0.998, + "content": "inherently robust to noise.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 0.986, + "content": "Second, based on the linear boundaries of the decision region,we propose a novel loss function", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 104, + 428, + 506, + 441 + ], + "score": 0.992, + "content": "to train a neural network that produces a robust counterfactual explanation as a small subset of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 440, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 451 + ], + "score": 0.989, + "content": "edges of an input graph. The loss function is designed to directly optimize the explainability and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "score": 0.984, + "content": "counterfactual property of the subset of edges,such that: 1) the subgraph induced by the edges lies", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 460, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 471 + ], + "score": 0.998, + "content": "within the decision region,thus has a prediction consistent with the input graph; and 2) deleting the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 470, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 482 + ], + "score": 0.992, + "content": " subset of edges from the input graph produces a remainder subgraph that lies outside the decision", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 481, + 411, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 411, + 493 + ], + "score": 0.992, + "content": "region, thus the prediction on the remainder subgraph changes significantly.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 108, + 496, + 503, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "score": 0.987, + "content": "Last, we conduct comprehensive experimental study to compare our method with the state-of-the-art", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 505, + 519 + ], + "score": 0.989, + "content": "methods on fidelity, robustness,accuracy and efficiency. All the results solidly demonstrate the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 518, + 262, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 262, + 529 + ], + "score": 0.993, + "content": "superior performance of our approach.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "title", + "bbox": [ + 107, + 543, + 194, + 556 + ], + "lines": [ + { + "bbox": [ + 104, + 542, + 196, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 542, + 196, + 558 + ], + "score": 1.0, + "content": "2Related work", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 504, + 589 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 0.985, + "content": "The existing GNN explanation methods [46,37, 45,32, 25] generally fall into two categories: model", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 577, + 384, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 384, + 590 + ], + "score": 0.985, + "content": "level explanation [46] and instance level explanation [37,45,32, 25].", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 605 + ], + "score": 0.987, + "content": "A model level explanation method [46] produces a high-level explanation about the general behaviors", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "score": 0.978, + "content": "of a GNN independent from input examples. This may be achieved by synthesizing a set of artificial", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 614, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 104, + 614, + 506, + 627 + ], + "score": 0.986, + "content": " graph instances such that each artificial graph instance maximizes the prediction score on a certain", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "score": 0.985, + "content": "class. The weakness of model level explanation methods is that an input graph instance may not", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "score": 0.993, + "content": "contain an artificial graph instance,and removing an artificial graph from an input graph does not", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 0.992, + "content": "necessarily change the prediction. As a result, model level explanations are substantially different", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 655, + 504, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 504, + 668 + ], + "score": 0.995, + "content": "from counterfactual explanations, because the synthesized artificial graphs do not provide insights", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 665, + 403, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 403, + 680 + ], + "score": 0.988, + "content": " into how the GNN makes its prediction on a specific input graph instance.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47.5 + }, + { + "type": "text", + "bbox": [ + 109, + 682, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 107, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 107, + 681, + 505, + 694 + ], + "score": 0.988, + "content": "The instance level explanation methods [37,45,32, 25] explain the prediction(s) made by a GNN", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 692, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 506, + 705 + ], + "score": 0.99, + "content": " on a specific input graph instance or multiple instances by identifying a subgraph of an input graph", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 52.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 712, + 397, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 398, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 398, + 725 + ], + "score": 0.98, + "content": "2Other versions of the paper are available at htps://arxiv.org/abs/2107.04086", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 156 + ], + "lines": [], + "index": 3.5, + "bbox_fs": [ + 105, + 73, + 507, + 158 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 174 + ], + "score": 0.99, + "content": "How to produce robust counterfactual explanations on predictions made by general graph neural", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 173, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 183 + ], + "score": 0.99, + "content": "networks is a novel problem that has not been systematically studied before.As to be discussed", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 182, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 194 + ], + "score": 0.986, + "content": " in Section 2, most GNN explanation methods [45,25,46,37,32] are neither counterfactual nor", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 0.989, + "content": " robust. These methods mostly focus on identifying a subgraph of an input graph that achieves a high", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 204, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 505, + 215 + ], + "score": 0.992, + "content": "correlation with the prediction result. Such explanations are usually not counterfactual because, due", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 213, + 507, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 507, + 227 + ], + "score": 0.983, + "content": " to the high non-convexity of GNNs, removing a subgraph that achieves a high correlation does not", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 0.987, + "content": " necessarily change the prediction result. Moreover, many existing methods [45,25,37,32] are not", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 234, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 247 + ], + "score": 0.991, + "content": "robust to noise and may change significantly upon slight modifications on input graphs, because", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 0.987, + "content": " the explanation of every single input graph prediction is independently optimized to maximize the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 255, + 467, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 467, + 267 + ], + "score": 0.99, + "content": "correlation with the prediction, thus an explanation can easily overfit the noise in the data.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 162, + 507, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 272, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 0.986, + "content": "In this paper², we develop RCExplainer, a novel method to produce robust counterfactual explanations", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 294 + ], + "score": 0.985, + "content": "on GNNs. The key idea is to first model the common decision logic of a GNN by set of decision", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 293, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 305 + ], + "score": 0.996, + "content": "regions where each decision region governs the predictions on a large number of graphs,and then", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 302, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 317 + ], + "score": 0.987, + "content": " extract robust counterfactual explanations by a deep neural network that explores the decision logic", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "score": 0.983, + "content": "carried by the linear decision boundaries of the decision regions. We make the following contributions.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 271, + 506, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 329, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "score": 0.979, + "content": "First, we model the decision logic of a GNN by a set of decision regions, where each decision region", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 104, + 339, + 506, + 352 + ], + "score": 0.991, + "content": " is induced by a set of linear decision boundaries of the GNN. We propose an unsupervised method to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 0.995, + "content": " find decision regions for each class such that each decision region governs the prediction of multiple", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 359, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 104, + 359, + 506, + 374 + ], + "score": 0.994, + "content": " graph samples predicted to be the same class. The linear decision boundaries of the decision region", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 371, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 383 + ], + "score": 0.983, + "content": "capture the common decision logic on all the graph instances inside the decision region, thus do", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 0.982, + "content": "not easily overfit the noise of an individual graph instance. By exploring the common decision", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "score": 0.994, + "content": "logic encoded in the linear boundaries, we are able to produce counterfactual explanations that are", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 402, + 213, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 213, + 415 + ], + "score": 0.998, + "content": "inherently robust to noise.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 330, + 506, + 415 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 418, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 0.986, + "content": "Second, based on the linear boundaries of the decision region,we propose a novel loss function", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 104, + 428, + 506, + 441 + ], + "score": 0.992, + "content": "to train a neural network that produces a robust counterfactual explanation as a small subset of", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 440, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 451 + ], + "score": 0.989, + "content": "edges of an input graph. The loss function is designed to directly optimize the explainability and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "score": 0.984, + "content": "counterfactual property of the subset of edges,such that: 1) the subgraph induced by the edges lies", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 460, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 505, + 471 + ], + "score": 0.998, + "content": "within the decision region,thus has a prediction consistent with the input graph; and 2) deleting the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 470, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 482 + ], + "score": 0.992, + "content": " subset of edges from the input graph produces a remainder subgraph that lies outside the decision", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 481, + 411, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 411, + 493 + ], + "score": 0.992, + "content": "region, thus the prediction on the remainder subgraph changes significantly.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 104, + 418, + 506, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 496, + 503, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 508 + ], + "score": 0.987, + "content": "Last, we conduct comprehensive experimental study to compare our method with the state-of-the-art", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 505, + 519 + ], + "score": 0.989, + "content": "methods on fidelity, robustness,accuracy and efficiency. All the results solidly demonstrate the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 518, + 262, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 262, + 529 + ], + "score": 0.993, + "content": "superior performance of our approach.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 496, + 505, + 529 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 543, + 194, + 556 + ], + "lines": [ + { + "bbox": [ + 104, + 542, + 196, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 542, + 196, + 558 + ], + "score": 1.0, + "content": "2Related work", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 504, + 589 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 0.985, + "content": "The existing GNN explanation methods [46,37, 45,32, 25] generally fall into two categories: model", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 577, + 384, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 384, + 590 + ], + "score": 0.985, + "content": "level explanation [46] and instance level explanation [37,45,32, 25].", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 567, + 505, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 605 + ], + "score": 0.987, + "content": "A model level explanation method [46] produces a high-level explanation about the general behaviors", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 616 + ], + "score": 0.978, + "content": "of a GNN independent from input examples. This may be achieved by synthesizing a set of artificial", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 614, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 104, + 614, + 506, + 627 + ], + "score": 0.986, + "content": " graph instances such that each artificial graph instance maximizes the prediction score on a certain", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "score": 0.985, + "content": "class. The weakness of model level explanation methods is that an input graph instance may not", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "score": 0.993, + "content": "contain an artificial graph instance,and removing an artificial graph from an input graph does not", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 0.992, + "content": "necessarily change the prediction. As a result, model level explanations are substantially different", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 655, + 504, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 504, + 668 + ], + "score": 0.995, + "content": "from counterfactual explanations, because the synthesized artificial graphs do not provide insights", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 665, + 403, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 403, + 680 + ], + "score": 0.988, + "content": " into how the GNN makes its prediction on a specific input graph instance.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47.5, + "bbox_fs": [ + 104, + 594, + 506, + 680 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 682, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 107, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 107, + 681, + 505, + 694 + ], + "score": 0.988, + "content": "The instance level explanation methods [37,45,32, 25] explain the prediction(s) made by a GNN", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 692, + 506, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 506, + 705 + ], + "score": 0.99, + "content": " on a specific input graph instance or multiple instances by identifying a subgraph of an input graph", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "score": 0.987, + "content": " instance that achieves a high correlation with the prediction on the input graph. GNNExplainer [45]", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 0.993, + "content": "removes redundant edges from an input graph instance to produce an explanation that maximizes", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 105 + ], + "score": 0.996, + "content": "the mutual information between the distribution of subgraphs of the input graph and the GNN's", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "score": 0.981, + "content": "prediction. Following the same idea by Ying et al. [45],PGExplainer [25] parameterizes the genera-", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "score": 0.993, + "content": "tion process of explanations by a deep neural network,and trains it to maximize a similar mutual", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "score": 0.984, + "content": " information based loss used by GNNExplainer [45]. The trained deep neural network is then applied", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 135, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 506, + 148 + ], + "score": 0.99, + "content": "to generate explanations for a single input graph instance or a group of input graphs. MEG [30]", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 146, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 506, + 158 + ], + "score": 0.99, + "content": " incorporates strong domain knowledge in chemistry with a reinforcement learning framework to", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 505, + 169 + ], + "score": 0.985, + "content": " produce counterfactual explanations on GNNs specifically built for compound prediction, but the", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 166, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 505, + 179 + ], + "score": 0.993, + "content": "heavy reliance on domain knowledge largely limits its applicability on general GNNs. The recently", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "score": 0.992, + "content": " proposed CF-GNNExplainer [24] independently optimizes the counterfactual property for each", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 187, + 438, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 438, + 200 + ], + "score": 0.994, + "content": "explanation but ignores the correlation between the prediction and the explanation.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 52.5, + "bbox_fs": [ + 106, + 681, + 506, + 705 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "score": 0.987, + "content": " instance that achieves a high correlation with the prediction on the input graph. GNNExplainer [45]", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 0.993, + "content": "removes redundant edges from an input graph instance to produce an explanation that maximizes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 105 + ], + "score": 0.996, + "content": "the mutual information between the distribution of subgraphs of the input graph and the GNN's", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "score": 0.981, + "content": "prediction. Following the same idea by Ying et al. [45],PGExplainer [25] parameterizes the genera-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "score": 0.993, + "content": "tion process of explanations by a deep neural network,and trains it to maximize a similar mutual", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "score": 0.984, + "content": " information based loss used by GNNExplainer [45]. The trained deep neural network is then applied", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 135, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 506, + 148 + ], + "score": 0.99, + "content": "to generate explanations for a single input graph instance or a group of input graphs. MEG [30]", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 146, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 506, + 158 + ], + "score": 0.99, + "content": " incorporates strong domain knowledge in chemistry with a reinforcement learning framework to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 505, + 169 + ], + "score": 0.985, + "content": " produce counterfactual explanations on GNNs specifically built for compound prediction, but the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 166, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 505, + 179 + ], + "score": 0.993, + "content": "heavy reliance on domain knowledge largely limits its applicability on general GNNs. The recently", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 190 + ], + "score": 0.992, + "content": " proposed CF-GNNExplainer [24] independently optimizes the counterfactual property for each", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 187, + 438, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 438, + 200 + ], + "score": 0.994, + "content": "explanation but ignores the correlation between the prediction and the explanation.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 505, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "score": 0.997, + "content": "Some studies [32,37] also adapt the existing explanation methods of image-oriented deep neural", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "score": 0.996, + "content": "networks to produce instance level explanations for GNNs.Pope et al.[32] extend several gradient", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 0.988, + "content": "based methods [33,35,49] to explain predictions made by GNNs. The explanations are prone to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 234, + 507, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 507, + 247 + ], + "score": 0.981, + "content": " gradient saturation[12] and may also be misleading [1] due to the heavy reliance on noisy gradients.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 0.975, + "content": "Velickovic et al.[37] extend the atention mechanism[7,8] to identify the nodes in an input graph that", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "score": 0.995, + "content": "contribute the most to the prediction.This method has to retrain the GNN with the altered architecture", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 266, + 495, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 495, + 277 + ], + "score": 0.989, + "content": "and the inserted attntion layers.Thus,the explanations may not be faithful to the original GNN.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 281, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 293 + ], + "score": 0.993, + "content": "Instance level explanations from most of the methods are usually not counterfactual because, due", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 292, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 304 + ], + "score": 0.988, + "content": "to the non-convexity of GNNs, removing an explanation subgraph from the input graph does not", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 301, + 506, + 315 + ], + "score": 0.988, + "content": " necessarily change the prediction result. Moreover, those methods [45,25,37,32, 24] are usually", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 0.99, + "content": "not robust to noise because the explanation of every single input graph prediction is independently", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "score": 0.989, + "content": "optimized. Thus,an explanation can easily overfit the noise inside input graphs and may change", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 333, + 329, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 329, + 346 + ], + "score": 0.988, + "content": " significantly upon slight modifications on input graphs.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 0.985, + "content": "To tackle the weaknesses in the existing methods, in this paper, we directly optimize the counterfactual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "score": 0.996, + "content": "property of an explanation along with the correlation between the explanation and the prediction. Our", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 382 + ], + "score": 0.994, + "content": "explanations are also much more robust to modifications on input graphs,because they are produced", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "score": 0.995, + "content": "from the common decision logic on a large group of similar input graphs, which do not easily overfit", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 390, + 267, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 267, + 403 + ], + "score": 0.987, + "content": "the noise of an individual graph sample.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 0.991, + "content": "Please note that our study is substantially different from adversarial attacks on GNNs.The adversarial", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 0.987, + "content": "attacking methods [51,53,42,43,20] use adversarial examples to change the predictions of GNNs but", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 426, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 440 + ], + "score": 0.987, + "content": "ignore the explainability of the generated adversarial examples [1O]. Thus, the adversarial examples", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 437, + 397, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 397, + 452 + ], + "score": 0.982, + "content": " generated by adversarial attcks may not explain the original prediction.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 505, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "score": 0.988, + "content": "Our method is substantially different from the above works because we focus on explaining the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 0.996, + "content": " prediction by directly optimizing the counterfactual property of an explanation along with correlation", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 0.99, + "content": "of the explanation with the prediction. We also require that the explanation is generally valid for a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "score": 0.994, + "content": "large set of similar graph instances by extracting it from the common linear decision boundaries of a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 494, + 195, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 195, + 509 + ], + "score": 0.985, + "content": "large decision region.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 107, + 529, + 237, + 542 + ], + "lines": [ + { + "bbox": [ + 104, + 528, + 238, + 544 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 238, + 544 + ], + "score": 0.996, + "content": "3Problem Formulation", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 149, + 571 + ], + "score": 0.999, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 150, + 558, + 200, + 570 + ], + "score": 0.93, + "content": "G = \\{ V , E \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 558, + 259, + 571 + ], + "score": 0.998, + "content": "a graph where", + "type": "text" + }, + { + "bbox": [ + 259, + 559, + 347, + 570 + ], + "score": 0.92, + "content": "V = \\{ v _ { 1 } , v _ { 2 } , \\ldots , v _ { n } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 558, + 395, + 571 + ], + "score": 0.999, + "content": "is the set of", + "type": "text" + }, + { + "bbox": [ + 395, + 561, + 402, + 568 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 558, + 445, + 571 + ], + "score": 1.0, + "content": "nodes and", + "type": "text" + }, + { + "bbox": [ + 446, + 558, + 495, + 569 + ], + "score": 0.91, + "content": "E \\subseteq V \\times V", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 558, + 506, + 571 + ], + "score": 0.999, + "content": "is", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 568, + 502, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 288, + 582 + ], + "score": 0.977, + "content": " the set of edges. The edge structure of a graph", + "type": "text" + }, + { + "bbox": [ + 289, + 570, + 297, + 579 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 568, + 440, + 582 + ], + "score": 0.995, + "content": "is described by an adjacency matrix", + "type": "text" + }, + { + "bbox": [ + 440, + 569, + 502, + 582 + ], + "score": 0.92, + "content": "\\mathbf { A } \\in \\{ 0 , 1 \\} ^ { n \\times n }", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 579, + 439, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 133, + 593 + ], + "score": 0.999, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 581, + 168, + 592 + ], + "score": 0.93, + "content": "{ \\bf A } _ { i j } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 579, + 300, + 593 + ], + "score": 0.994, + "content": "if there is an edge between node", + "type": "text" + }, + { + "bbox": [ + 300, + 582, + 309, + 591 + ], + "score": 0.86, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 579, + 327, + 593 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 582, + 338, + 592 + ], + "score": 0.87, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 579, + 359, + 593 + ], + "score": 0.927, + "content": "; and", + "type": "text" + }, + { + "bbox": [ + 359, + 581, + 394, + 592 + ], + "score": 0.94, + "content": "{ \\bf A } _ { i j } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 579, + 439, + 593 + ], + "score": 0.996, + "content": "otherwise.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 504, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 149, + 608 + ], + "score": 0.999, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 149, + 596, + 156, + 607 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 595, + 506, + 608 + ], + "score": 0.991, + "content": "a GNN model that maps a graph to a probability distribution over a set of classes denoted", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 606, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 118, + 618 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 607, + 127, + 616 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 606, + 147, + 618 + ], + "score": 0.918, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 147, + 607, + 157, + 616 + ], + "score": 0.8, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 606, + 398, + 618 + ], + "score": 0.977, + "content": "denote the set of graphs that are used to train the GNN model", + "type": "text" + }, + { + "bbox": [ + 399, + 607, + 406, + 618 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 606, + 506, + 618 + ], + "score": 0.968, + "content": ".We focus on GNNs that", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 0.989, + "content": "adopt piecewise linear activation functions,such as MaxOut[14] and the family of ReLU[13,15,29].", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 380, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 382, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 382, + 645 + ], + "score": 0.999, + "content": "The robust counterfactual explanation problem is defined as follows.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 445, + 671 + ], + "score": 0.98, + "content": "Definition 1 (Robust Counterfactual Explanation Problem) Given a GNN model", + "type": "text" + }, + { + "bbox": [ + 445, + 660, + 452, + 670 + ], + "score": 0.82, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 659, + 506, + 671 + ], + "score": 0.98, + "content": "trained on a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 162, + 682 + ], + "score": 0.972, + "content": "set of graphs", + "type": "text" + }, + { + "bbox": [ + 162, + 670, + 171, + 680 + ], + "score": 0.7, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 669, + 253, + 682 + ], + "score": 0.962, + "content": "， for an input graph", + "type": "text" + }, + { + "bbox": [ + 253, + 670, + 304, + 681 + ], + "score": 0.92, + "content": "G = \\{ V , E \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 669, + 417, + 682 + ], + "score": 0.958, + "content": ", our goal is to explain why", + "type": "text" + }, + { + "bbox": [ + 417, + 670, + 426, + 679 + ], + "score": 0.78, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 669, + 506, + 682 + ], + "score": 0.987, + "content": "is predicted by the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 104, + 678, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 167, + 693 + ], + "score": 0.991, + "content": "GNN model as", + "type": "text" + }, + { + "bbox": [ + 167, + 680, + 190, + 691 + ], + "score": 0.9, + "content": "\\phi ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 678, + 342, + 693 + ], + "score": 0.984, + "content": "by identifying a small subset of edges", + "type": "text" + }, + { + "bbox": [ + 342, + 682, + 370, + 691 + ], + "score": 0.9, + "content": "S \\subseteq E", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 678, + 413, + 693 + ], + "score": 0.947, + "content": ",such that", + "type": "text" + }, + { + "bbox": [ + 414, + 681, + 426, + 691 + ], + "score": 0.38, + "content": "( l )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 678, + 506, + 693 + ], + "score": 0.97, + "content": "removing the set of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 141, + 703 + ], + "score": 0.958, + "content": "edges in", + "type": "text" + }, + { + "bbox": [ + 141, + 691, + 149, + 700 + ], + "score": 0.67, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 690, + 171, + 703 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 171, + 691, + 180, + 700 + ], + "score": 0.76, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 690, + 506, + 703 + ], + "score": 0.985, + "content": "that causes the maximum drop in the confidence of the original prediction; and (2)", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 107, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 107, + 702, + 114, + 711 + ], + "score": 0.77, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 699, + 479, + 714 + ], + "score": 0.99, + "content": "is stable and doesn't change when the edges and the feature representations of the nodes of", + "type": "text" + }, + { + "bbox": [ + 480, + 702, + 488, + 711 + ], + "score": 0.66, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 699, + 506, + 714 + ], + "score": 0.999, + "content": "are", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 104, + 711, + 219, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 711, + 219, + 723 + ], + "score": 0.992, + "content": "perturbed by random noise.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "score": 0.999, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 198 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 72, + 506, + 200 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 505, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 216 + ], + "score": 0.997, + "content": "Some studies [32,37] also adapt the existing explanation methods of image-oriented deep neural", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "score": 0.996, + "content": "networks to produce instance level explanations for GNNs.Pope et al.[32] extend several gradient", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 0.988, + "content": "based methods [33,35,49] to explain predictions made by GNNs. The explanations are prone to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 234, + 507, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 507, + 247 + ], + "score": 0.981, + "content": " gradient saturation[12] and may also be misleading [1] due to the heavy reliance on noisy gradients.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 0.975, + "content": "Velickovic et al.[37] extend the atention mechanism[7,8] to identify the nodes in an input graph that", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "score": 0.995, + "content": "contribute the most to the prediction.This method has to retrain the GNN with the altered architecture", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 266, + 495, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 495, + 277 + ], + "score": 0.989, + "content": "and the inserted attntion layers.Thus,the explanations may not be faithful to the original GNN.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 202, + 507, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 281, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 293 + ], + "score": 0.993, + "content": "Instance level explanations from most of the methods are usually not counterfactual because, due", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 292, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 304 + ], + "score": 0.988, + "content": "to the non-convexity of GNNs, removing an explanation subgraph from the input graph does not", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 301, + 506, + 315 + ], + "score": 0.988, + "content": " necessarily change the prediction result. Moreover, those methods [45,25,37,32, 24] are usually", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 0.99, + "content": "not robust to noise because the explanation of every single input graph prediction is independently", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "score": 0.989, + "content": "optimized. Thus,an explanation can easily overfit the noise inside input graphs and may change", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 333, + 329, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 329, + 346 + ], + "score": 0.988, + "content": " significantly upon slight modifications on input graphs.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 282, + 506, + 346 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 0.985, + "content": "To tackle the weaknesses in the existing methods, in this paper, we directly optimize the counterfactual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "score": 0.996, + "content": "property of an explanation along with the correlation between the explanation and the prediction. Our", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 382 + ], + "score": 0.994, + "content": "explanations are also much more robust to modifications on input graphs,because they are produced", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "score": 0.995, + "content": "from the common decision logic on a large group of similar input graphs, which do not easily overfit", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 390, + 267, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 267, + 403 + ], + "score": 0.987, + "content": "the noise of an individual graph sample.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 349, + 506, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 0.991, + "content": "Please note that our study is substantially different from adversarial attacks on GNNs.The adversarial", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 0.987, + "content": "attacking methods [51,53,42,43,20] use adversarial examples to change the predictions of GNNs but", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 426, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 440 + ], + "score": 0.987, + "content": "ignore the explainability of the generated adversarial examples [1O]. Thus, the adversarial examples", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 437, + 397, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 397, + 452 + ], + "score": 0.982, + "content": " generated by adversarial attcks may not explain the original prediction.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 406, + 505, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 505, + 507 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "score": 0.988, + "content": "Our method is substantially different from the above works because we focus on explaining the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 0.996, + "content": " prediction by directly optimizing the counterfactual property of an explanation along with correlation", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 0.99, + "content": "of the explanation with the prediction. We also require that the explanation is generally valid for a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "score": 0.994, + "content": "large set of similar graph instances by extracting it from the common linear decision boundaries of a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 494, + 195, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 195, + 509 + ], + "score": 0.985, + "content": "large decision region.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 453, + 506, + 509 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 529, + 237, + 542 + ], + "lines": [ + { + "bbox": [ + 104, + 528, + 238, + 544 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 238, + 544 + ], + "score": 0.996, + "content": "3Problem Formulation", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 149, + 571 + ], + "score": 0.999, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 150, + 558, + 200, + 570 + ], + "score": 0.93, + "content": "G = \\{ V , E \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 558, + 259, + 571 + ], + "score": 0.998, + "content": "a graph where", + "type": "text" + }, + { + "bbox": [ + 259, + 559, + 347, + 570 + ], + "score": 0.92, + "content": "V = \\{ v _ { 1 } , v _ { 2 } , \\ldots , v _ { n } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 558, + 395, + 571 + ], + "score": 0.999, + "content": "is the set of", + "type": "text" + }, + { + "bbox": [ + 395, + 561, + 402, + 568 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 558, + 445, + 571 + ], + "score": 1.0, + "content": "nodes and", + "type": "text" + }, + { + "bbox": [ + 446, + 558, + 495, + 569 + ], + "score": 0.91, + "content": "E \\subseteq V \\times V", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 558, + 506, + 571 + ], + "score": 0.999, + "content": "is", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 568, + 502, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 288, + 582 + ], + "score": 0.977, + "content": " the set of edges. The edge structure of a graph", + "type": "text" + }, + { + "bbox": [ + 289, + 570, + 297, + 579 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 568, + 440, + 582 + ], + "score": 0.995, + "content": "is described by an adjacency matrix", + "type": "text" + }, + { + "bbox": [ + 440, + 569, + 502, + 582 + ], + "score": 0.92, + "content": "\\mathbf { A } \\in \\{ 0 , 1 \\} ^ { n \\times n }", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 579, + 439, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 133, + 593 + ], + "score": 0.999, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 581, + 168, + 592 + ], + "score": 0.93, + "content": "{ \\bf A } _ { i j } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 579, + 300, + 593 + ], + "score": 0.994, + "content": "if there is an edge between node", + "type": "text" + }, + { + "bbox": [ + 300, + 582, + 309, + 591 + ], + "score": 0.86, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 579, + 327, + 593 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 582, + 338, + 592 + ], + "score": 0.87, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 579, + 359, + 593 + ], + "score": 0.927, + "content": "; and", + "type": "text" + }, + { + "bbox": [ + 359, + 581, + 394, + 592 + ], + "score": 0.94, + "content": "{ \\bf A } _ { i j } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 579, + 439, + 593 + ], + "score": 0.996, + "content": "otherwise.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 558, + 506, + 593 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 596, + 504, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 149, + 608 + ], + "score": 0.999, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 149, + 596, + 156, + 607 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 595, + 506, + 608 + ], + "score": 0.991, + "content": "a GNN model that maps a graph to a probability distribution over a set of classes denoted", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 606, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 118, + 618 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 607, + 127, + 616 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 606, + 147, + 618 + ], + "score": 0.918, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 147, + 607, + 157, + 616 + ], + "score": 0.8, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 606, + 398, + 618 + ], + "score": 0.977, + "content": "denote the set of graphs that are used to train the GNN model", + "type": "text" + }, + { + "bbox": [ + 399, + 607, + 406, + 618 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 606, + 506, + 618 + ], + "score": 0.968, + "content": ".We focus on GNNs that", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 0.989, + "content": "adopt piecewise linear activation functions,such as MaxOut[14] and the family of ReLU[13,15,29].", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 106, + 595, + 506, + 629 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 380, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 382, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 382, + 645 + ], + "score": 0.999, + "content": "The robust counterfactual explanation problem is defined as follows.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 632, + 382, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 445, + 671 + ], + "score": 0.98, + "content": "Definition 1 (Robust Counterfactual Explanation Problem) Given a GNN model", + "type": "text" + }, + { + "bbox": [ + 445, + 660, + 452, + 670 + ], + "score": 0.82, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 659, + 506, + 671 + ], + "score": 0.98, + "content": "trained on a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 162, + 682 + ], + "score": 0.972, + "content": "set of graphs", + "type": "text" + }, + { + "bbox": [ + 162, + 670, + 171, + 680 + ], + "score": 0.7, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 669, + 253, + 682 + ], + "score": 0.962, + "content": "， for an input graph", + "type": "text" + }, + { + "bbox": [ + 253, + 670, + 304, + 681 + ], + "score": 0.92, + "content": "G = \\{ V , E \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 669, + 417, + 682 + ], + "score": 0.958, + "content": ", our goal is to explain why", + "type": "text" + }, + { + "bbox": [ + 417, + 670, + 426, + 679 + ], + "score": 0.78, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 669, + 506, + 682 + ], + "score": 0.987, + "content": "is predicted by the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 104, + 678, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 167, + 693 + ], + "score": 0.991, + "content": "GNN model as", + "type": "text" + }, + { + "bbox": [ + 167, + 680, + 190, + 691 + ], + "score": 0.9, + "content": "\\phi ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 678, + 342, + 693 + ], + "score": 0.984, + "content": "by identifying a small subset of edges", + "type": "text" + }, + { + "bbox": [ + 342, + 682, + 370, + 691 + ], + "score": 0.9, + "content": "S \\subseteq E", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 678, + 413, + 693 + ], + "score": 0.947, + "content": ",such that", + "type": "text" + }, + { + "bbox": [ + 414, + 681, + 426, + 691 + ], + "score": 0.38, + "content": "( l )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 678, + 506, + 693 + ], + "score": 0.97, + "content": "removing the set of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 141, + 703 + ], + "score": 0.958, + "content": "edges in", + "type": "text" + }, + { + "bbox": [ + 141, + 691, + 149, + 700 + ], + "score": 0.67, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 690, + 171, + 703 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 171, + 691, + 180, + 700 + ], + "score": 0.76, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 690, + 506, + 703 + ], + "score": 0.985, + "content": "that causes the maximum drop in the confidence of the original prediction; and (2)", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 107, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 107, + 702, + 114, + 711 + ], + "score": 0.77, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 699, + 479, + 714 + ], + "score": 0.99, + "content": "is stable and doesn't change when the edges and the feature representations of the nodes of", + "type": "text" + }, + { + "bbox": [ + 480, + 702, + 488, + 711 + ], + "score": 0.66, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 699, + 506, + 714 + ], + "score": 0.999, + "content": "are", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 104, + 711, + 219, + 723 + ], + "spans": [ + { + "bbox": [ + 104, + 711, + 219, + 723 + ], + "score": 0.992, + "content": "perturbed by random noise.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5, + "bbox_fs": [ + 104, + 659, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 367, + 85 + ], + "score": 0.993, + "content": "In the definition,the first requirement requires that the explanation", + "type": "text" + }, + { + "bbox": [ + 367, + 73, + 375, + 82 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 72, + 505, + 85 + ], + "score": 0.979, + "content": "is counterfactual, and the second", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 497, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 484, + 95 + ], + "score": 0.997, + "content": "requirement requires that the explanation is robust to noisy changes on the edges and nodes of", + "type": "text" + }, + { + "bbox": [ + 484, + 83, + 493, + 93 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 83, + 497, + 95 + ], + "score": 0.511, + "content": "：", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 109, + 166, + 122 + ], + "lines": [ + { + "bbox": [ + 104, + 108, + 168, + 124 + ], + "spans": [ + { + "bbox": [ + 104, + 108, + 168, + 124 + ], + "score": 1.0, + "content": "4Method", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 133, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 145 + ], + "score": 0.983, + "content": "In this section, we first introduce how to extract the common decision logic of a GNN on a large", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 0.99, + "content": " set of graphs with the same predicted class. This is achieved by a decision region induced by a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 154, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 165 + ], + "score": 0.983, + "content": "set of linear decision boundaries of the GNN.Then, based on the linear boundaries of the decision", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 176 + ], + "score": 0.983, + "content": "region, we propose a novel lossfunction to train a neural network that produces robust counterfactual", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 175, + 495, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 495, + 187 + ], + "score": 0.987, + "content": "explanations.Last, we discuss the time complexity of our method when generating explanations.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 198, + 248, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 249, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 249, + 212 + ], + "score": 0.991, + "content": "4.1Modelling Decision Regions", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 218, + 505, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 231 + ], + "score": 0.985, + "content": "Following the routines of many deep neural network explanation methods [33, 48], we extract the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 243, + 243 + ], + "score": 0.985, + "content": "decision region of a GNN in the", + "type": "text" + }, + { + "bbox": [ + 244, + 231, + 250, + 240 + ], + "score": 0.81, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 230, + 359, + 243 + ], + "score": 0.981, + "content": "-dimensional output space", + "type": "text" + }, + { + "bbox": [ + 360, + 230, + 373, + 240 + ], + "score": 0.9, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 230, + 506, + 243 + ], + "score": 0.982, + "content": "of the last convolution layer of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 0.989, + "content": " the GNN. The features generated by the last convolution layer are more conceptually meaningful", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 251, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 263 + ], + "score": 0.992, + "content": "and more robust to noise than those raw features of input graphs,such as vertices and edges [52, 2].", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 262, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 151, + 273 + ], + "score": 0.999, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 152, + 262, + 167, + 274 + ], + "score": 0.9, + "content": "\\phi _ { g c }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 262, + 506, + 273 + ], + "score": 0.989, + "content": "the mapping function realized by the graph convolution layers that maps an input", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 273, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 104, + 273, + 131, + 287 + ], + "score": 1.0, + "content": "graph", + "type": "text" + }, + { + "bbox": [ + 131, + 274, + 140, + 284 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 273, + 234, + 287 + ], + "score": 0.999, + "content": "to its graph embedding", + "type": "text" + }, + { + "bbox": [ + 234, + 273, + 289, + 286 + ], + "score": 0.93, + "content": "\\phi _ { g c } ( G ) \\in \\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 273, + 321, + 287 + ], + "score": 0.941, + "content": ",and by", + "type": "text" + }, + { + "bbox": [ + 322, + 274, + 338, + 285 + ], + "score": 0.87, + "content": "\\phi _ { f c }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 273, + 506, + 287 + ], + "score": 0.99, + "content": "the mapping function realized by the fully", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 303, + 298 + ], + "score": 0.993, + "content": "connected layers that maps the graph embedding", + "type": "text" + }, + { + "bbox": [ + 303, + 285, + 334, + 296 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 284, + 506, + 298 + ], + "score": 0.998, + "content": "to a predicted distribution over the classes", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 294, + 478, + 309 + ], + "spans": [ + { + "bbox": [ + 104, + 294, + 117, + 309 + ], + "score": 0.997, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 117, + 295, + 126, + 304 + ], + "score": 0.74, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 294, + 221, + 309 + ], + "score": 0.994, + "content": ". The overall prediction", + "type": "text" + }, + { + "bbox": [ + 221, + 295, + 243, + 307 + ], + "score": 0.92, + "content": "\\phi ( { \\bar { G } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 294, + 390, + 309 + ], + "score": 0.973, + "content": "made by the GNN can be writtn as", + "type": "text" + }, + { + "bbox": [ + 390, + 294, + 478, + 307 + ], + "score": 0.93, + "content": "\\phi ( G ) = \\phi _ { f c } ( \\phi _ { g c } ( G ) )", + "type": "inline_equation" + } + ], + "index": 16 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 323 + ], + "score": 0.983, + "content": "For the GNNs that adopt piecewise linear activation functions for the hidden neurons, such as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 320, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 320, + 409, + 335 + ], + "score": 0.984, + "content": "MaxOut [14] and the family of ReLU [13,15,29], the decision logic of", + "type": "text" + }, + { + "bbox": [ + 410, + 322, + 425, + 334 + ], + "score": 0.9, + "content": "\\phi _ { f c }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 320, + 480, + 335 + ], + "score": 0.965, + "content": " in the space", + "type": "text" + }, + { + "bbox": [ + 480, + 321, + 494, + 332 + ], + "score": 0.88, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 320, + 506, + 335 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 0.993, + "content": "characterized by a piecewise linear decision boundary formed by connected pieces of decision", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 343, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 168, + 357 + ], + "score": 0.996, + "content": "hyperplanes in", + "type": "text" + }, + { + "bbox": [ + 168, + 343, + 182, + 354 + ], + "score": 0.88, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 343, + 506, + 357 + ], + "score": 0.99, + "content": "[1]. We call these hyperplanes linear decision boundaries (LDBs), and denote", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 353, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 104, + 353, + 119, + 369 + ], + "score": 0.999, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 356, + 129, + 366 + ], + "score": 0.76, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 353, + 246, + 369 + ], + "score": 0.978, + "content": "the set of LDBs induced by", + "type": "text" + }, + { + "bbox": [ + 246, + 356, + 262, + 367 + ], + "score": 0.9, + "content": "\\phi _ { f c }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 353, + 348, + 369 + ], + "score": 0.99, + "content": ". The set of LDBs in", + "type": "text" + }, + { + "bbox": [ + 348, + 356, + 358, + 366 + ], + "score": 0.76, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 353, + 441, + 369 + ], + "score": 0.966, + "content": " partitions the space", + "type": "text" + }, + { + "bbox": [ + 442, + 355, + 455, + 365 + ], + "score": 0.87, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 353, + 506, + 369 + ], + "score": 0.999, + "content": "into a large", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 366, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 457, + 378 + ], + "score": 0.985, + "content": "number of convex polytopes. A convex polytope is formed by a subset of LDBs in", + "type": "text" + }, + { + "bbox": [ + 457, + 366, + 467, + 376 + ], + "score": 0.73, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 366, + 506, + 378 + ], + "score": 0.905, + "content": ". All the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 375, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 390 + ], + "score": 0.995, + "content": " graphs whose graph embeddings are contained in the same convex polytope are predicted as the same", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 463, + 399 + ], + "score": 0.997, + "content": "class [4]. Therefore, the LDBs of a convex polytope encode the common decision logic of", + "type": "text" + }, + { + "bbox": [ + 464, + 387, + 479, + 399 + ], + "score": 0.9, + "content": "\\phi _ { f c }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 387, + 505, + 399 + ], + "score": 0.975, + "content": "on all", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 450, + 410 + ], + "score": 0.982, + "content": " the graphs whose graph embeddings lie within the convex polytope [4]. Here,a graph", + "type": "text" + }, + { + "bbox": [ + 451, + 398, + 460, + 407 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 397, + 506, + 410 + ], + "score": 0.999, + "content": "is covered", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 407, + 467, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 407, + 289, + 421 + ], + "score": 0.987, + "content": " by a convex polytope if the graph embedding", + "type": "text" + }, + { + "bbox": [ + 289, + 408, + 320, + 420 + ], + "score": 0.93, + "content": "\\phi _ { g c } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 407, + 467, + 421 + ], + "score": 0.992, + "content": "is contained in the convex polytope.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 436 + ], + "score": 0.991, + "content": "Based on the above insight, we model the decision region for a set of graph instances as a convex", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 506, + 447 + ], + "score": 0.99, + "content": "polytope that satisfies the following two properties. First, the decision region should be induced by a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 196, + 457 + ], + "score": 0.966, + "content": "subset of the LDBs in", + "type": "text" + }, + { + "bbox": [ + 196, + 445, + 206, + 454 + ], + "score": 0.59, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 444, + 506, + 457 + ], + "score": 0.984, + "content": ".In this way, when we extract counterfactual explanations from the LDBs,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 467 + ], + "score": 0.99, + "content": "the explanations are loyal to the real decision logic_of the GNN. Second, the decision region should", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 306, + 478 + ], + "score": 0.974, + "content": "cover many graph instances in the training dataset", + "type": "text" + }, + { + "bbox": [ + 306, + 465, + 316, + 475 + ], + "score": 0.78, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 465, + 506, + 478 + ], + "score": 0.991, + "content": ",and all the covered graphs should be predicted", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "score": 0.979, + "content": " as the same class. In this way, the LDBs of the decision region capture the common decision logic on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "score": 0.987, + "content": " all the graphs covered by the decision region. Here, the requirement of covering a larger number of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 0.985, + "content": " graphs ensures that the common decision logic is general,and thus it is lesslikely to overfit the noise", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 0.988, + "content": "of an individual graph instance. As a result, the counterfactual explanations extracted from the LDBs", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "score": 0.986, + "content": "of the decision region are insensitive to slight changes in the input graphs. Our method can be easily", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "score": 0.988, + "content": " generalized to incorporate prediction confidence in the coverage measure,such as considering the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "score": 0.983, + "content": "count of graphs weighted by prediction confidence. To keep our discusson simple, we do not pursue", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 547, + 229, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 229, + 561 + ], + "score": 0.984, + "content": "this detail further in the paper.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 0.986, + "content": " Next, we illustrate how to extract a decision region satisfying the above two requirements. The key", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 398, + 587 + ], + "score": 0.989, + "content": " idea is to find a convex polytope covering a large set of graph instances in", + "type": "text" + }, + { + "bbox": [ + 398, + 575, + 408, + 585 + ], + "score": 0.83, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 574, + 505, + 587 + ], + "score": 0.999, + "content": "that are predicted as the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 585, + 181, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 151, + 596 + ], + "score": 0.986, + "content": "same class", + "type": "text" + }, + { + "bbox": [ + 151, + 586, + 177, + 595 + ], + "score": 0.9, + "content": "c \\in C", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 585, + 181, + 596 + ], + "score": 0.849, + "content": "：", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 665 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 150, + 613 + ], + "score": 0.997, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 150, + 601, + 186, + 612 + ], + "score": 0.91, + "content": "D _ { c } \\subseteq D", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 600, + 265, + 613 + ], + "score": 0.993, + "content": "the set of graphs in", + "type": "text" + }, + { + "bbox": [ + 266, + 601, + 275, + 611 + ], + "score": 0.79, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 600, + 356, + 613 + ], + "score": 0.993, + "content": "predicted as a class", + "type": "text" + }, + { + "bbox": [ + 356, + 601, + 381, + 611 + ], + "score": 0.9, + "content": "c \\in C", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 600, + 398, + 613 + ], + "score": 0.98, + "content": ",by", + "type": "text" + }, + { + "bbox": [ + 398, + 601, + 429, + 612 + ], + "score": 0.9, + "content": "\\mathcal { P } \\subseteq \\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 600, + 506, + 613 + ], + "score": 0.975, + "content": "a set of LDBs that", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 611, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 181, + 626 + ], + "score": 0.983, + "content": " partition the space", + "type": "text" + }, + { + "bbox": [ + 181, + 612, + 195, + 622 + ], + "score": 0.89, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 611, + 347, + 626 + ], + "score": 0.99, + "content": "into a set of convex polytopes, and by", + "type": "text" + }, + { + "bbox": [ + 347, + 612, + 376, + 624 + ], + "score": 0.92, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 611, + 506, + 626 + ], + "score": 0.992, + "content": " the convex polytope induced by", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 622, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 107, + 623, + 115, + 633 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 622, + 294, + 636 + ], + "score": 0.986, + "content": " that covers the largest number of graphs in", + "type": "text" + }, + { + "bbox": [ + 294, + 623, + 307, + 634 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 622, + 357, + 636 + ], + "score": 0.98, + "content": ". Denote by", + "type": "text" + }, + { + "bbox": [ + 358, + 623, + 388, + 635 + ], + "score": 0.92, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 622, + 490, + 636 + ], + "score": 0.98, + "content": " the number of graphs in", + "type": "text" + }, + { + "bbox": [ + 491, + 624, + 504, + 634 + ], + "score": 0.86, + "content": "D _ { c }", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 151, + 646 + ], + "score": 0.998, + "content": "covered by", + "type": "text" + }, + { + "bbox": [ + 152, + 633, + 181, + 645 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 633, + 214, + 646 + ], + "score": 0.981, + "content": ",and by", + "type": "text" + }, + { + "bbox": [ + 214, + 633, + 245, + 645 + ], + "score": 0.93, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 633, + 342, + 646 + ], + "score": 0.996, + "content": "the number of graphs in", + "type": "text" + }, + { + "bbox": [ + 342, + 634, + 352, + 643 + ], + "score": 0.81, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 633, + 430, + 646 + ], + "score": 0.975, + "content": "that are covered by", + "type": "text" + }, + { + "bbox": [ + 430, + 633, + 460, + 645 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , \\bar { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 633, + 506, + 646 + ], + "score": 0.996, + "content": "but are not", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 180, + 658 + ], + "score": 0.997, + "content": "predicted as class", + "type": "text" + }, + { + "bbox": [ + 181, + 646, + 186, + 654 + ], + "score": 0.64, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 642, + 478, + 658 + ], + "score": 0.995, + "content": ". We extract a decision region covering a large set of graph instances in", + "type": "text" + }, + { + "bbox": [ + 478, + 644, + 491, + 655 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 654, + 333, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 333, + 667 + ], + "score": 0.999, + "content": "solving the following constrained optimization problem.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 670, + 367, + 690 + ], + "lines": [ + { + "bbox": [ + 244, + 670, + 367, + 690 + ], + "spans": [ + { + "bbox": [ + 244, + 670, + 367, + 690 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { \\mathcal { P } \\subseteq \\mathcal { H } } g ( \\mathcal { P } , c ) , \\mathrm { ~ s . t . ~ } h ( \\mathcal { P } , c ) = 0", + "type": "interline_equation", + "image_path": "199dfcc9238c4541a86b7079313a4ecb03e682dc957ef9622a8d982bbb04d186.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 244, + 670, + 367, + 690 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 701, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 405, + 712 + ], + "score": 0.988, + "content": "This formulation realizes the two properties of decision regions because", + "type": "text" + }, + { + "bbox": [ + 405, + 701, + 437, + 712 + ], + "score": 0.91, + "content": "\\mathcal { P } \\subseteq \\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 701, + 505, + 712 + ], + "score": 0.971, + "content": "ensures that the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 302, + 723 + ], + "score": 0.986, + "content": "decision region is induced by a subset of LDBs in", + "type": "text" + }, + { + "bbox": [ + 302, + 711, + 311, + 721 + ], + "score": 0.81, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 711, + 364, + 723 + ], + "score": 0.997, + "content": ",maximizing", + "type": "text" + }, + { + "bbox": [ + 364, + 711, + 394, + 723 + ], + "score": 0.93, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 711, + 446, + 723 + ], + "score": 1.0, + "content": "requires that", + "type": "text" + }, + { + "bbox": [ + 447, + 711, + 476, + 723 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "covers", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "4", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 367, + 85 + ], + "score": 0.993, + "content": "In the definition,the first requirement requires that the explanation", + "type": "text" + }, + { + "bbox": [ + 367, + 73, + 375, + 82 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 72, + 505, + 85 + ], + "score": 0.979, + "content": "is counterfactual, and the second", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 497, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 484, + 95 + ], + "score": 0.997, + "content": "requirement requires that the explanation is robust to noisy changes on the edges and nodes of", + "type": "text" + }, + { + "bbox": [ + 484, + 83, + 493, + 93 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 83, + 497, + 95 + ], + "score": 0.511, + "content": "：", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 72, + 505, + 95 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 109, + 166, + 122 + ], + "lines": [ + { + "bbox": [ + 104, + 108, + 168, + 124 + ], + "spans": [ + { + "bbox": [ + 104, + 108, + 168, + 124 + ], + "score": 1.0, + "content": "4Method", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 133, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 145 + ], + "score": 0.983, + "content": "In this section, we first introduce how to extract the common decision logic of a GNN on a large", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 0.99, + "content": " set of graphs with the same predicted class. This is achieved by a decision region induced by a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 154, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 165 + ], + "score": 0.983, + "content": "set of linear decision boundaries of the GNN.Then, based on the linear boundaries of the decision", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 506, + 176 + ], + "score": 0.983, + "content": "region, we propose a novel lossfunction to train a neural network that produces robust counterfactual", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 175, + 495, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 495, + 187 + ], + "score": 0.987, + "content": "explanations.Last, we discuss the time complexity of our method when generating explanations.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 133, + 506, + 187 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 198, + 248, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 197, + 249, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 249, + 212 + ], + "score": 0.991, + "content": "4.1Modelling Decision Regions", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 218, + 505, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 231 + ], + "score": 0.985, + "content": "Following the routines of many deep neural network explanation methods [33, 48], we extract the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 243, + 243 + ], + "score": 0.985, + "content": "decision region of a GNN in the", + "type": "text" + }, + { + "bbox": [ + 244, + 231, + 250, + 240 + ], + "score": 0.81, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 230, + 359, + 243 + ], + "score": 0.981, + "content": "-dimensional output space", + "type": "text" + }, + { + "bbox": [ + 360, + 230, + 373, + 240 + ], + "score": 0.9, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 230, + 506, + 243 + ], + "score": 0.982, + "content": "of the last convolution layer of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 0.989, + "content": " the GNN. The features generated by the last convolution layer are more conceptually meaningful", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 251, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 263 + ], + "score": 0.992, + "content": "and more robust to noise than those raw features of input graphs,such as vertices and edges [52, 2].", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 262, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 262, + 151, + 273 + ], + "score": 0.999, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 152, + 262, + 167, + 274 + ], + "score": 0.9, + "content": "\\phi _ { g c }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 262, + 506, + 273 + ], + "score": 0.989, + "content": "the mapping function realized by the graph convolution layers that maps an input", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 273, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 104, + 273, + 131, + 287 + ], + "score": 1.0, + "content": "graph", + "type": "text" + }, + { + "bbox": [ + 131, + 274, + 140, + 284 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 273, + 234, + 287 + ], + "score": 0.999, + "content": "to its graph embedding", + "type": "text" + }, + { + "bbox": [ + 234, + 273, + 289, + 286 + ], + "score": 0.93, + "content": "\\phi _ { g c } ( G ) \\in \\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 273, + 321, + 287 + ], + "score": 0.941, + "content": ",and by", + "type": "text" + }, + { + "bbox": [ + 322, + 274, + 338, + 285 + ], + "score": 0.87, + "content": "\\phi _ { f c }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 273, + 506, + 287 + ], + "score": 0.99, + "content": "the mapping function realized by the fully", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 104, + 284, + 303, + 298 + ], + "score": 0.993, + "content": "connected layers that maps the graph embedding", + "type": "text" + }, + { + "bbox": [ + 303, + 285, + 334, + 296 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 284, + 506, + 298 + ], + "score": 0.998, + "content": "to a predicted distribution over the classes", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 294, + 478, + 309 + ], + "spans": [ + { + "bbox": [ + 104, + 294, + 117, + 309 + ], + "score": 0.997, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 117, + 295, + 126, + 304 + ], + "score": 0.74, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 294, + 221, + 309 + ], + "score": 0.994, + "content": ". The overall prediction", + "type": "text" + }, + { + "bbox": [ + 221, + 295, + 243, + 307 + ], + "score": 0.92, + "content": "\\phi ( { \\bar { G } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 294, + 390, + 309 + ], + "score": 0.973, + "content": "made by the GNN can be writtn as", + "type": "text" + }, + { + "bbox": [ + 390, + 294, + 478, + 307 + ], + "score": 0.93, + "content": "\\phi ( G ) = \\phi _ { f c } ( \\phi _ { g c } ( G ) )", + "type": "inline_equation" + } + ], + "index": 16 + } + ], + "index": 12.5, + "bbox_fs": [ + 104, + 219, + 506, + 309 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 323 + ], + "score": 0.983, + "content": "For the GNNs that adopt piecewise linear activation functions for the hidden neurons, such as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 320, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 320, + 409, + 335 + ], + "score": 0.984, + "content": "MaxOut [14] and the family of ReLU [13,15,29], the decision logic of", + "type": "text" + }, + { + "bbox": [ + 410, + 322, + 425, + 334 + ], + "score": 0.9, + "content": "\\phi _ { f c }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 320, + 480, + 335 + ], + "score": 0.965, + "content": " in the space", + "type": "text" + }, + { + "bbox": [ + 480, + 321, + 494, + 332 + ], + "score": 0.88, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 320, + 506, + 335 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 0.993, + "content": "characterized by a piecewise linear decision boundary formed by connected pieces of decision", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 343, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 168, + 357 + ], + "score": 0.996, + "content": "hyperplanes in", + "type": "text" + }, + { + "bbox": [ + 168, + 343, + 182, + 354 + ], + "score": 0.88, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 343, + 506, + 357 + ], + "score": 0.99, + "content": "[1]. We call these hyperplanes linear decision boundaries (LDBs), and denote", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 353, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 104, + 353, + 119, + 369 + ], + "score": 0.999, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 356, + 129, + 366 + ], + "score": 0.76, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 353, + 246, + 369 + ], + "score": 0.978, + "content": "the set of LDBs induced by", + "type": "text" + }, + { + "bbox": [ + 246, + 356, + 262, + 367 + ], + "score": 0.9, + "content": "\\phi _ { f c }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 353, + 348, + 369 + ], + "score": 0.99, + "content": ". The set of LDBs in", + "type": "text" + }, + { + "bbox": [ + 348, + 356, + 358, + 366 + ], + "score": 0.76, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 353, + 441, + 369 + ], + "score": 0.966, + "content": " partitions the space", + "type": "text" + }, + { + "bbox": [ + 442, + 355, + 455, + 365 + ], + "score": 0.87, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 353, + 506, + 369 + ], + "score": 0.999, + "content": "into a large", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 366, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 457, + 378 + ], + "score": 0.985, + "content": "number of convex polytopes. A convex polytope is formed by a subset of LDBs in", + "type": "text" + }, + { + "bbox": [ + 457, + 366, + 467, + 376 + ], + "score": 0.73, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 366, + 506, + 378 + ], + "score": 0.905, + "content": ". All the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 375, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 390 + ], + "score": 0.995, + "content": " graphs whose graph embeddings are contained in the same convex polytope are predicted as the same", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 463, + 399 + ], + "score": 0.997, + "content": "class [4]. Therefore, the LDBs of a convex polytope encode the common decision logic of", + "type": "text" + }, + { + "bbox": [ + 464, + 387, + 479, + 399 + ], + "score": 0.9, + "content": "\\phi _ { f c }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 387, + 505, + 399 + ], + "score": 0.975, + "content": "on all", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 450, + 410 + ], + "score": 0.982, + "content": " the graphs whose graph embeddings lie within the convex polytope [4]. Here,a graph", + "type": "text" + }, + { + "bbox": [ + 451, + 398, + 460, + 407 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 397, + 506, + 410 + ], + "score": 0.999, + "content": "is covered", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 407, + 467, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 407, + 289, + 421 + ], + "score": 0.987, + "content": " by a convex polytope if the graph embedding", + "type": "text" + }, + { + "bbox": [ + 289, + 408, + 320, + 420 + ], + "score": 0.93, + "content": "\\phi _ { g c } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 407, + 467, + 421 + ], + "score": 0.992, + "content": "is contained in the convex polytope.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 309, + 506, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 560 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 436 + ], + "score": 0.991, + "content": "Based on the above insight, we model the decision region for a set of graph instances as a convex", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 506, + 447 + ], + "score": 0.99, + "content": "polytope that satisfies the following two properties. First, the decision region should be induced by a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 196, + 457 + ], + "score": 0.966, + "content": "subset of the LDBs in", + "type": "text" + }, + { + "bbox": [ + 196, + 445, + 206, + 454 + ], + "score": 0.59, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 444, + 506, + 457 + ], + "score": 0.984, + "content": ".In this way, when we extract counterfactual explanations from the LDBs,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 467 + ], + "score": 0.99, + "content": "the explanations are loyal to the real decision logic_of the GNN. Second, the decision region should", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 465, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 306, + 478 + ], + "score": 0.974, + "content": "cover many graph instances in the training dataset", + "type": "text" + }, + { + "bbox": [ + 306, + 465, + 316, + 475 + ], + "score": 0.78, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 465, + 506, + 478 + ], + "score": 0.991, + "content": ",and all the covered graphs should be predicted", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "score": 0.979, + "content": " as the same class. In this way, the LDBs of the decision region capture the common decision logic on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "score": 0.987, + "content": " all the graphs covered by the decision region. Here, the requirement of covering a larger number of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 0.985, + "content": " graphs ensures that the common decision logic is general,and thus it is lesslikely to overfit the noise", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 0.988, + "content": "of an individual graph instance. As a result, the counterfactual explanations extracted from the LDBs", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 530 + ], + "score": 0.986, + "content": "of the decision region are insensitive to slight changes in the input graphs. Our method can be easily", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "score": 0.988, + "content": " generalized to incorporate prediction confidence in the coverage measure,such as considering the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 550 + ], + "score": 0.983, + "content": "count of graphs weighted by prediction confidence. To keep our discusson simple, we do not pursue", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 547, + 229, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 229, + 561 + ], + "score": 0.984, + "content": "this detail further in the paper.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 423, + 506, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 0.986, + "content": " Next, we illustrate how to extract a decision region satisfying the above two requirements. The key", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 398, + 587 + ], + "score": 0.989, + "content": " idea is to find a convex polytope covering a large set of graph instances in", + "type": "text" + }, + { + "bbox": [ + 398, + 575, + 408, + 585 + ], + "score": 0.83, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 574, + 505, + 587 + ], + "score": 0.999, + "content": "that are predicted as the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 585, + 181, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 151, + 596 + ], + "score": 0.986, + "content": "same class", + "type": "text" + }, + { + "bbox": [ + 151, + 586, + 177, + 595 + ], + "score": 0.9, + "content": "c \\in C", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 585, + 181, + 596 + ], + "score": 0.849, + "content": "：", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 563, + 505, + 596 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 665 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 150, + 613 + ], + "score": 0.997, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 150, + 601, + 186, + 612 + ], + "score": 0.91, + "content": "D _ { c } \\subseteq D", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 600, + 265, + 613 + ], + "score": 0.993, + "content": "the set of graphs in", + "type": "text" + }, + { + "bbox": [ + 266, + 601, + 275, + 611 + ], + "score": 0.79, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 600, + 356, + 613 + ], + "score": 0.993, + "content": "predicted as a class", + "type": "text" + }, + { + "bbox": [ + 356, + 601, + 381, + 611 + ], + "score": 0.9, + "content": "c \\in C", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 600, + 398, + 613 + ], + "score": 0.98, + "content": ",by", + "type": "text" + }, + { + "bbox": [ + 398, + 601, + 429, + 612 + ], + "score": 0.9, + "content": "\\mathcal { P } \\subseteq \\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 600, + 506, + 613 + ], + "score": 0.975, + "content": "a set of LDBs that", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 611, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 181, + 626 + ], + "score": 0.983, + "content": " partition the space", + "type": "text" + }, + { + "bbox": [ + 181, + 612, + 195, + 622 + ], + "score": 0.89, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 611, + 347, + 626 + ], + "score": 0.99, + "content": "into a set of convex polytopes, and by", + "type": "text" + }, + { + "bbox": [ + 347, + 612, + 376, + 624 + ], + "score": 0.92, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 611, + 506, + 626 + ], + "score": 0.992, + "content": " the convex polytope induced by", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 622, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 107, + 623, + 115, + 633 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 622, + 294, + 636 + ], + "score": 0.986, + "content": " that covers the largest number of graphs in", + "type": "text" + }, + { + "bbox": [ + 294, + 623, + 307, + 634 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 622, + 357, + 636 + ], + "score": 0.98, + "content": ". Denote by", + "type": "text" + }, + { + "bbox": [ + 358, + 623, + 388, + 635 + ], + "score": 0.92, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 622, + 490, + 636 + ], + "score": 0.98, + "content": " the number of graphs in", + "type": "text" + }, + { + "bbox": [ + 491, + 624, + 504, + 634 + ], + "score": 0.86, + "content": "D _ { c }", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 151, + 646 + ], + "score": 0.998, + "content": "covered by", + "type": "text" + }, + { + "bbox": [ + 152, + 633, + 181, + 645 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 633, + 214, + 646 + ], + "score": 0.981, + "content": ",and by", + "type": "text" + }, + { + "bbox": [ + 214, + 633, + 245, + 645 + ], + "score": 0.93, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 633, + 342, + 646 + ], + "score": 0.996, + "content": "the number of graphs in", + "type": "text" + }, + { + "bbox": [ + 342, + 634, + 352, + 643 + ], + "score": 0.81, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 633, + 430, + 646 + ], + "score": 0.975, + "content": "that are covered by", + "type": "text" + }, + { + "bbox": [ + 430, + 633, + 460, + 645 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , \\bar { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 633, + 506, + 646 + ], + "score": 0.996, + "content": "but are not", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 180, + 658 + ], + "score": 0.997, + "content": "predicted as class", + "type": "text" + }, + { + "bbox": [ + 181, + 646, + 186, + 654 + ], + "score": 0.64, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 642, + 478, + 658 + ], + "score": 0.995, + "content": ". We extract a decision region covering a large set of graph instances in", + "type": "text" + }, + { + "bbox": [ + 478, + 644, + 491, + 655 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 654, + 333, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 333, + 667 + ], + "score": 0.999, + "content": "solving the following constrained optimization problem.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 600, + 506, + 667 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 670, + 367, + 690 + ], + "lines": [ + { + "bbox": [ + 244, + 670, + 367, + 690 + ], + "spans": [ + { + "bbox": [ + 244, + 670, + 367, + 690 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { \\mathcal { P } \\subseteq \\mathcal { H } } g ( \\mathcal { P } , c ) , \\mathrm { ~ s . t . ~ } h ( \\mathcal { P } , c ) = 0", + "type": "interline_equation", + "image_path": "199dfcc9238c4541a86b7079313a4ecb03e682dc957ef9622a8d982bbb04d186.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 244, + 670, + 367, + 690 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 701, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 405, + 712 + ], + "score": 0.988, + "content": "This formulation realizes the two properties of decision regions because", + "type": "text" + }, + { + "bbox": [ + 405, + 701, + 437, + 712 + ], + "score": 0.91, + "content": "\\mathcal { P } \\subseteq \\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 701, + 505, + 712 + ], + "score": 0.971, + "content": "ensures that the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 302, + 723 + ], + "score": 0.986, + "content": "decision region is induced by a subset of LDBs in", + "type": "text" + }, + { + "bbox": [ + 302, + 711, + 311, + 721 + ], + "score": 0.81, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 711, + 364, + 723 + ], + "score": 0.997, + "content": ",maximizing", + "type": "text" + }, + { + "bbox": [ + 364, + 711, + 394, + 723 + ], + "score": 0.93, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 711, + 446, + 723 + ], + "score": 1.0, + "content": "requires that", + "type": "text" + }, + { + "bbox": [ + 447, + 711, + 476, + 723 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "covers", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50.5, + "bbox_fs": [ + 106, + 701, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 94 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 216, + 86 + ], + "score": 0.957, + "content": " a large number of graphs in", + "type": "text" + }, + { + "bbox": [ + 216, + 73, + 230, + 83 + ], + "score": 0.88, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 70, + 306, + 86 + ], + "score": 0.989, + "content": ", and the constraint", + "type": "text" + }, + { + "bbox": [ + 306, + 72, + 355, + 84 + ], + "score": 0.93, + "content": "h ( \\mathcal { P } , c ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 70, + 506, + 86 + ], + "score": 0.982, + "content": "ensures that allthe graphs covered by", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 83, + 272, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 136, + 95 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 83, + 262, + 95 + ], + "score": 0.998, + "content": "are predicted as the same class", + "type": "text" + }, + { + "bbox": [ + 263, + 86, + 268, + 93 + ], + "score": 0.63, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 83, + 272, + 95 + ], + "score": 0.807, + "content": "：", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 99, + 505, + 131 + ], + "lines": [ + { + "bbox": [ + 106, + 98, + 506, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 204, + 111 + ], + "score": 0.995, + "content": "Once we find a solution", + "type": "text" + }, + { + "bbox": [ + 204, + 99, + 213, + 109 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 98, + 382, + 111 + ], + "score": 0.989, + "content": "to the above problem, the decision region", + "type": "text" + }, + { + "bbox": [ + 383, + 99, + 412, + 111 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 98, + 506, + 111 + ], + "score": 0.983, + "content": "can be easily obtained", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 109, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 275, + 121 + ], + "score": 0.985, + "content": "by first counting the number of graphs in", + "type": "text" + }, + { + "bbox": [ + 275, + 110, + 289, + 120 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 109, + 475, + 121 + ], + "score": 0.978, + "content": "covered by each convex polytope induced by", + "type": "text" + }, + { + "bbox": [ + 475, + 110, + 484, + 119 + ], + "score": 0.83, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 109, + 506, + 121 + ], + "score": 0.991, + "content": ",and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 119, + 424, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 406, + 132 + ], + "score": 0.992, + "content": " then select the convex polytope that covers the largest number of graphs in", + "type": "text" + }, + { + "bbox": [ + 406, + 120, + 419, + 131 + ], + "score": 0.88, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 119, + 424, + 132 + ], + "score": 0.862, + "content": "：", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 107, + 143, + 251, + 155 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 252, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 252, + 158 + ], + "score": 0.997, + "content": "4.2Extracting Decision Regions", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 505, + 175 + ], + "score": 0.995, + "content": "The optimization problem in Equation (1) is intractable for standard GNNs, mainly because it is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 197, + 185 + ], + "score": 0.997, + "content": "impractical to compute", + "type": "text" + }, + { + "bbox": [ + 198, + 174, + 207, + 184 + ], + "score": 0.8, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 173, + 398, + 185 + ], + "score": 0.97, + "content": ", all the LDBs of a GNN. The number of LDBs in", + "type": "text" + }, + { + "bbox": [ + 398, + 174, + 408, + 184 + ], + "score": 0.83, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 173, + 505, + 185 + ], + "score": 0.985, + "content": "of a GNN is exponential", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "score": 0.991, + "content": "with respect to the number of neurons in the worst case [27].To address this challenge,we substitute", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 195, + 251, + 208 + ], + "spans": [ + { + "bbox": [ + 107, + 197, + 116, + 207 + ], + "score": 0.77, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 195, + 167, + 208 + ], + "score": 0.989, + "content": "by a sample", + "type": "text" + }, + { + "bbox": [ + 168, + 195, + 177, + 207 + ], + "score": 0.86, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 195, + 237, + 208 + ], + "score": 0.996, + "content": "of LDBs from", + "type": "text" + }, + { + "bbox": [ + 238, + 195, + 247, + 207 + ], + "score": 0.85, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 195, + 251, + 208 + ], + "score": 0.831, + "content": "：", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 212, + 505, + 246 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 186, + 225 + ], + "score": 0.991, + "content": "A LDB in the space", + "type": "text" + }, + { + "bbox": [ + 186, + 212, + 200, + 223 + ], + "score": 0.88, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 212, + 268, + 225 + ], + "score": 0.996, + "content": "can be written as", + "type": "text" + }, + { + "bbox": [ + 269, + 212, + 325, + 224 + ], + "score": 0.91, + "content": "\\mathbf { w } ^ { \\top } \\mathbf { x } + b = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 212, + 364, + 225 + ], + "score": 0.968, + "content": ",where is", + "type": "text" + }, + { + "bbox": [ + 364, + 212, + 396, + 223 + ], + "score": 0.92, + "content": "\\mathbf { x } \\in \\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 212, + 506, + 225 + ], + "score": 0.979, + "content": "is a variable, w is the basis", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 146, + 236 + ], + "score": 0.963, + "content": "term, and", + "type": "text" + }, + { + "bbox": [ + 147, + 224, + 153, + 234 + ], + "score": 0.77, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 223, + 391, + 236 + ], + "score": 0.985, + "content": "corresponds to the bias.Following [4], for any input graph", + "type": "text" + }, + { + "bbox": [ + 391, + 224, + 400, + 234 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 223, + 505, + 236 + ], + "score": 0.991, + "content": ",a linear boundary can be", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 234, + 233, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 164, + 247 + ], + "score": 1.0, + "content": "sampled from", + "type": "text" + }, + { + "bbox": [ + 164, + 235, + 174, + 244 + ], + "score": 0.81, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 234, + 233, + 247 + ], + "score": 0.977, + "content": "by computing", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 250, + 414, + 276 + ], + "lines": [ + { + "bbox": [ + 195, + 250, + 414, + 276 + ], + "spans": [ + { + "bbox": [ + 195, + 250, + 414, + 276 + ], + "score": 0.92, + "content": "\\mathbf { w } = \\frac { \\partial \\left( \\operatorname* { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) - \\operatorname* { m a x } _ { 2 } ( \\phi _ { f c } ( \\alpha ) ) \\right) } { \\partial \\alpha } | _ { \\alpha = \\phi _ { g c } ( G ) } ,", + "type": "interline_equation", + "image_path": "2dbc5d9de809bb0aacc312e490149627bcf7576859a17ad5d232a303bbbb6bf6.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 195, + 250, + 414, + 276 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 280, + 123, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 123, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 123, + 291 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 297, + 420, + 312 + ], + "lines": [ + { + "bbox": [ + 189, + 297, + 420, + 312 + ], + "spans": [ + { + "bbox": [ + 189, + 297, + 420, + 312 + ], + "score": 0.88, + "content": "b = \\mathrm { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) - \\mathrm { m a x } _ { 2 } ( \\phi _ { f c } ( \\alpha ) ) - \\mathbf { w } ^ { T } { \\alpha } | _ { \\alpha = \\phi _ { g c } ( G ) } ,", + "type": "interline_equation", + "image_path": "79c9953979154b44111c9874d086d71cda854549201b41bbe069b902cc4cfe90.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 189, + 297, + 420, + 312 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 506, + 362 + ], + "lines": [ + { + "bbox": [ + 104, + 316, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 132, + 331 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 317, + 198, + 330 + ], + "score": 0.91, + "content": "\\operatorname* { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 316, + 217, + 331 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 317, + 278, + 330 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { 2 } \\bigl ( \\phi _ { f c } ( \\pmb { \\alpha } ) \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 316, + 506, + 331 + ], + "score": 0.992, + "content": " are the largest and the second largest values in the vector", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 328, + 507, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 137, + 340 + ], + "score": 0.92, + "content": "\\phi _ { f c } ( \\alpha )", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 328, + 277, + 341 + ], + "score": 0.996, + "content": ",respectively. Given an input graph", + "type": "text" + }, + { + "bbox": [ + 277, + 329, + 286, + 338 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 328, + 468, + 341 + ], + "score": 0.99, + "content": ",Equations (2) and (3) identify one LDB from", + "type": "text" + }, + { + "bbox": [ + 468, + 329, + 478, + 338 + ], + "score": 0.77, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 328, + 507, + 341 + ], + "score": 0.976, + "content": ".Thus,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 328, + 351 + ], + "score": 0.983, + "content": "we can sample a subset of input graphs uniformly from", + "type": "text" + }, + { + "bbox": [ + 328, + 339, + 337, + 348 + ], + "score": 0.84, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 339, + 506, + 351 + ], + "score": 0.99, + "content": ",and use Equations (2) and (3) to derive a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 349, + 221, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 185, + 362 + ], + "score": 0.972, + "content": "sample of LDBs as", + "type": "text" + }, + { + "bbox": [ + 185, + 349, + 216, + 361 + ], + "score": 0.92, + "content": "\\tilde { \\mathcal { H } } \\subset \\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 351, + 221, + 362 + ], + "score": 0.813, + "content": "：", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 367, + 420, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 420, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 185, + 383 + ], + "score": 0.978, + "content": "Now, we substitute", + "type": "text" + }, + { + "bbox": [ + 185, + 369, + 195, + 378 + ], + "score": 0.82, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 366, + 271, + 383 + ], + "score": 0.991, + "content": "in Equation (1) by", + "type": "text" + }, + { + "bbox": [ + 271, + 367, + 281, + 379 + ], + "score": 0.87, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 366, + 420, + 383 + ], + "score": 0.982, + "content": "to produce the following problem.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 384, + 367, + 406 + ], + "lines": [ + { + "bbox": [ + 242, + 384, + 367, + 406 + ], + "spans": [ + { + "bbox": [ + 242, + 384, + 367, + 406 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { \\mathcal { P } \\subseteq \\tilde { \\mathcal { H } } } g ( \\mathcal { P } , c ) , \\mathrm { ~ s . t . ~ } h ( \\mathcal { P } , c ) \\leq \\delta ,", + "type": "interline_equation", + "image_path": "f52df860ca40a98e65cc25516b494f56cafd46c079811fa063e656225dbb0792.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 242, + 384, + 367, + 406 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 133, + 424 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 411, + 160, + 422 + ], + "score": 0.91, + "content": "\\delta \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 410, + 451, + 424 + ], + "score": 0.983, + "content": "is a tolerance parameter to keep this problem feasible. The parameter", + "type": "text" + }, + { + "bbox": [ + 451, + 411, + 457, + 421 + ], + "score": 0.82, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 410, + 506, + 424 + ], + "score": 0.982, + "content": "is required", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 422, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 187, + 436 + ], + "score": 0.998, + "content": "because substituting", + "type": "text" + }, + { + "bbox": [ + 188, + 423, + 198, + 434 + ], + "score": 0.78, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 423, + 210, + 436 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 211, + 422, + 221, + 434 + ], + "score": 0.87, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 423, + 303, + 436 + ], + "score": 0.991, + "content": "ignores the LDBs in", + "type": "text" + }, + { + "bbox": [ + 303, + 422, + 330, + 435 + ], + "score": 0.92, + "content": "\\mathcal { H } \\backslash \\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 423, + 440, + 436 + ], + "score": 0.951, + "content": ". Thus, the convex polytope", + "type": "text" + }, + { + "bbox": [ + 440, + 423, + 470, + 435 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 423, + 506, + 436 + ], + "score": 0.944, + "content": " induced", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 434, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 213, + 450 + ], + "score": 0.981, + "content": "by subset of boundaries in", + "type": "text" + }, + { + "bbox": [ + 213, + 434, + 223, + 446 + ], + "score": 0.85, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 435, + 433, + 450 + ], + "score": 0.989, + "content": "may contain instances that are not predicted as class", + "type": "text" + }, + { + "bbox": [ + 434, + 438, + 439, + 446 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 435, + 506, + 450 + ], + "score": 0.988, + "content": ". We directly set", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 446, + 453, + 461 + ], + "spans": [ + { + "bbox": [ + 107, + 446, + 156, + 460 + ], + "score": 0.93, + "content": "\\boldsymbol { \\delta } \\dot { = } \\boldsymbol { h } ( \\tilde { \\mathcal { H } } , \\boldsymbol { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 447, + 281, + 461 + ], + "score": 0.995, + "content": ", which is the smallest value of", + "type": "text" + }, + { + "bbox": [ + 281, + 448, + 287, + 458 + ], + "score": 0.78, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 447, + 453, + 461 + ], + "score": 0.999, + "content": "that keeps the practical problem feasible.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 464, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "score": 0.997, + "content": "The problem in Equation (4) can be proven to be a Submodular Cost Submodular Cover (SCSC)", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 222, + 487 + ], + "score": 0.978, + "content": "problem [18] (see Appendix", + "type": "text" + }, + { + "bbox": [ + 223, + 475, + 232, + 485 + ], + "score": 0.44, + "content": "\\mathbf { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 474, + 506, + 487 + ], + "score": 0.976, + "content": "for proof) that is well known to be NP-hard [5]. We adopt a greedy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 495, + 497 + ], + "score": 0.989, + "content": "boundary selection method to find a good solution to this problem [40]. Specifically, we initialize", + "type": "text" + }, + { + "bbox": [ + 495, + 486, + 504, + 495 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 496, + 397, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 343, + 510 + ], + "score": 0.98, + "content": " as an empty set,and then iteratively select a new boundary", + "type": "text" + }, + { + "bbox": [ + 343, + 497, + 350, + 507 + ], + "score": 0.8, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 497, + 373, + 510 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 373, + 496, + 383, + 507 + ], + "score": 0.85, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 497, + 397, + 510 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 514, + 391, + 544 + ], + "lines": [ + { + "bbox": [ + 219, + 514, + 391, + 544 + ], + "spans": [ + { + "bbox": [ + 219, + 514, + 391, + 544 + ], + "score": 0.94, + "content": "h = \\arg \\operatorname* { m i n } _ { h \\in \\tilde { \\mathcal { H } } \\backslash \\mathcal { P } } \\frac { g ( \\mathcal { P } , c ) - g ( \\mathcal { P } \\cup \\{ h \\} , c ) + \\epsilon } { h ( \\mathcal { P } , c ) - h ( \\mathcal { P } \\cup \\{ h \\} , c ) } ,", + "type": "interline_equation", + "image_path": "1f9f845a8a01b5db73f4894fc93b9fa1e911127e6b892e443b6edc0c7061df65.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 514, + 391, + 529.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 219, + 529.0, + 391, + 544.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 133, + 562 + ], + "score": 0.999, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 550, + 227, + 562 + ], + "score": 0.91, + "content": "g ( \\mathcal { P } , c ) - g ( \\mathcal { P } \\cup \\{ h \\} , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 549, + 298, + 562 + ], + "score": 0.984, + "content": " is the decrease of", + "type": "text" + }, + { + "bbox": [ + 299, + 550, + 329, + 561 + ], + "score": 0.91, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 549, + 382, + 562 + ], + "score": 0.964, + "content": "when adding", + "type": "text" + }, + { + "bbox": [ + 382, + 550, + 389, + 560 + ], + "score": 0.82, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 549, + 407, + 562 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 408, + 550, + 416, + 560 + ], + "score": 0.83, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 549, + 437, + 562 + ], + "score": 0.992, + "content": ",and", + "type": "text" + }, + { + "bbox": [ + 437, + 550, + 506, + 562 + ], + "score": 0.9, + "content": "h ( \\mathcal { P } , c ) - h ( \\mathcal { P } \\cup", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 107, + 560, + 136, + 572 + ], + "score": 0.91, + "content": "\\{ h \\} , c \\bar { ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 559, + 206, + 573 + ], + "score": 0.999, + "content": "is the decrease of", + "type": "text" + }, + { + "bbox": [ + 207, + 561, + 237, + 572 + ], + "score": 0.91, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 559, + 290, + 573 + ], + "score": 1.0, + "content": "when adding", + "type": "text" + }, + { + "bbox": [ + 291, + 561, + 298, + 570 + ], + "score": 0.81, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 559, + 316, + 573 + ], + "score": 0.999, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 316, + 562, + 325, + 570 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 559, + 350, + 573 + ], + "score": 0.925, + "content": ".Both", + "type": "text" + }, + { + "bbox": [ + 351, + 560, + 381, + 572 + ], + "score": 0.93, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 559, + 398, + 573 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 560, + 429, + 572 + ], + "score": 0.92, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 559, + 506, + 573 + ], + "score": 0.999, + "content": "are non-increasing", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 160, + 586 + ], + "score": 1.0, + "content": "when adding", + "type": "text" + }, + { + "bbox": [ + 160, + 572, + 188, + 583 + ], + "score": 0.89, + "content": "h \\in \\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 572, + 207, + 586 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 208, + 573, + 216, + 583 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 572, + 348, + 586 + ], + "score": 0.999, + "content": "because adding a new boundary", + "type": "text" + }, + { + "bbox": [ + 348, + 573, + 355, + 583 + ], + "score": 0.79, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 572, + 505, + 586 + ], + "score": 0.998, + "content": "may only exclude some graphs from", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 582, + 224, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 189, + 597 + ], + "score": 0.999, + "content": "the convex polytope", + "type": "text" + }, + { + "bbox": [ + 190, + 584, + 219, + 595 + ], + "score": 0.92, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 582, + 224, + 597 + ], + "score": 0.535, + "content": "：", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 343, + 614 + ], + "score": 0.982, + "content": "Intuitively, in each iteration,Equation (5) selects a boundary", + "type": "text" + }, + { + "bbox": [ + 343, + 600, + 371, + 612 + ], + "score": 0.92, + "content": "h \\in \\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 600, + 437, + 614 + ], + "score": 0.981, + "content": " such that adding", + "type": "text" + }, + { + "bbox": [ + 438, + 602, + 444, + 611 + ], + "score": 0.81, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 600, + 463, + 614 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 463, + 602, + 471, + 611 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 600, + 506, + 614 + ], + "score": 1.0, + "content": "reduces", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 107, + 612, + 137, + 624 + ], + "score": 0.91, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 611, + 227, + 624 + ], + "score": 0.982, + "content": " the least and reduces", + "type": "text" + }, + { + "bbox": [ + 227, + 612, + 258, + 623 + ], + "score": 0.91, + "content": "\\bar { h } ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 611, + 449, + 624 + ], + "score": 0.993, + "content": "the most. In this way,we can quickly reduce", + "type": "text" + }, + { + "bbox": [ + 449, + 612, + 480, + 624 + ], + "score": 0.92, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 611, + 506, + 624 + ], + "score": 0.992, + "content": "to be", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 158, + 635 + ], + "score": 0.999, + "content": "smaller than", + "type": "text" + }, + { + "bbox": [ + 158, + 623, + 164, + 632 + ], + "score": 0.8, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 622, + 243, + 635 + ], + "score": 0.999, + "content": "without decreasing", + "type": "text" + }, + { + "bbox": [ + 244, + 623, + 273, + 634 + ], + "score": 0.91, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 622, + 506, + 635 + ], + "score": 0.987, + "content": " too much, which produces a good feasible solution to the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 286, + 645 + ], + "score": 0.991, + "content": " practical problem.We add a small constant", + "type": "text" + }, + { + "bbox": [ + 286, + 635, + 292, + 642 + ], + "score": 0.74, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 632, + 506, + 645 + ], + "score": 0.986, + "content": "to the numerator such that, when there are multiple", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 162, + 655 + ], + "score": 0.999, + "content": "candidates of", + "type": "text" + }, + { + "bbox": [ + 162, + 644, + 169, + 653 + ], + "score": 0.81, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 642, + 252, + 655 + ], + "score": 0.999, + "content": "that do not decrease", + "type": "text" + }, + { + "bbox": [ + 252, + 643, + 282, + 655 + ], + "score": 0.92, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 642, + 374, + 655 + ], + "score": 0.967, + "content": ", we can still select the", + "type": "text" + }, + { + "bbox": [ + 374, + 644, + 381, + 653 + ], + "score": 0.8, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 642, + 432, + 655 + ], + "score": 0.999, + "content": "that reduces", + "type": "text" + }, + { + "bbox": [ + 433, + 643, + 464, + 655 + ], + "score": 0.93, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 642, + 505, + 655 + ], + "score": 0.96, + "content": "the most.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 658, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 502, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 477, + 671 + ], + "score": 0.982, + "content": "We apply a peeling-off strategy to iteratively extract multiple decision regions.For each class", + "type": "text" + }, + { + "bbox": [ + 477, + 659, + 502, + 669 + ], + "score": 0.89, + "content": "c \\in C", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 363, + 681 + ], + "score": 0.984, + "content": "we first solve the practical problem once to find a decision region", + "type": "text" + }, + { + "bbox": [ + 364, + 670, + 393, + 681 + ], + "score": 0.92, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 669, + 505, + 681 + ], + "score": 0.99, + "content": ", then we remove the graphs", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 679, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 154, + 693 + ], + "score": 0.998, + "content": "covered by", + "type": "text" + }, + { + "bbox": [ + 155, + 680, + 185, + 692 + ], + "score": 0.92, + "content": "r ( \\mathcal { P } , \\bar { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 679, + 209, + 693 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 210, + 681, + 223, + 691 + ], + "score": 0.87, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 679, + 442, + 693 + ], + "score": 0.994, + "content": ". If there are remaining graphs predicted as the class", + "type": "text" + }, + { + "bbox": [ + 443, + 682, + 448, + 690 + ], + "score": 0.69, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 679, + 506, + 693 + ], + "score": 0.94, + "content": ", we continue", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 411, + 703 + ], + "score": 0.984, + "content": " finding the decision regions using the remaining graphs until all the graphs in", + "type": "text" + }, + { + "bbox": [ + 411, + 691, + 424, + 701 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 690, + 506, + 703 + ], + "score": 0.973, + "content": "are removed. When", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 175, + 714 + ], + "score": 0.999, + "content": "all the graphs in", + "type": "text" + }, + { + "bbox": [ + 176, + 701, + 189, + 712 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 699, + 303, + 714 + ], + "score": 0.976, + "content": "are removed for each class", + "type": "text" + }, + { + "bbox": [ + 303, + 703, + 329, + 711 + ], + "score": 0.91, + "content": "c \\in C", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 699, + 507, + 714 + ], + "score": 0.972, + "content": ", we stop the iteration and return the set of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 711, + 217, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 217, + 723 + ], + "score": 0.996, + "content": "decision regions we found.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 740, + 309, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 309, + 753 + ], + "score": 0.994, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 94 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 216, + 86 + ], + "score": 0.957, + "content": " a large number of graphs in", + "type": "text" + }, + { + "bbox": [ + 216, + 73, + 230, + 83 + ], + "score": 0.88, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 70, + 306, + 86 + ], + "score": 0.989, + "content": ", and the constraint", + "type": "text" + }, + { + "bbox": [ + 306, + 72, + 355, + 84 + ], + "score": 0.93, + "content": "h ( \\mathcal { P } , c ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 70, + 506, + 86 + ], + "score": 0.982, + "content": "ensures that allthe graphs covered by", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 83, + 272, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 136, + 95 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 83, + 262, + 95 + ], + "score": 0.998, + "content": "are predicted as the same class", + "type": "text" + }, + { + "bbox": [ + 263, + 86, + 268, + 93 + ], + "score": 0.63, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 83, + 272, + 95 + ], + "score": 0.807, + "content": "：", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 104, + 70, + 506, + 95 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 99, + 505, + 131 + ], + "lines": [ + { + "bbox": [ + 106, + 98, + 506, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 204, + 111 + ], + "score": 0.995, + "content": "Once we find a solution", + "type": "text" + }, + { + "bbox": [ + 204, + 99, + 213, + 109 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 98, + 382, + 111 + ], + "score": 0.989, + "content": "to the above problem, the decision region", + "type": "text" + }, + { + "bbox": [ + 383, + 99, + 412, + 111 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 98, + 506, + 111 + ], + "score": 0.983, + "content": "can be easily obtained", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 109, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 275, + 121 + ], + "score": 0.985, + "content": "by first counting the number of graphs in", + "type": "text" + }, + { + "bbox": [ + 275, + 110, + 289, + 120 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 109, + 475, + 121 + ], + "score": 0.978, + "content": "covered by each convex polytope induced by", + "type": "text" + }, + { + "bbox": [ + 475, + 110, + 484, + 119 + ], + "score": 0.83, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 109, + 506, + 121 + ], + "score": 0.991, + "content": ",and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 119, + 424, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 406, + 132 + ], + "score": 0.992, + "content": " then select the convex polytope that covers the largest number of graphs in", + "type": "text" + }, + { + "bbox": [ + 406, + 120, + 419, + 131 + ], + "score": 0.88, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 119, + 424, + 132 + ], + "score": 0.862, + "content": "：", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 98, + 506, + 132 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 143, + 251, + 155 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 252, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 252, + 158 + ], + "score": 0.997, + "content": "4.2Extracting Decision Regions", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 505, + 175 + ], + "score": 0.995, + "content": "The optimization problem in Equation (1) is intractable for standard GNNs, mainly because it is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 197, + 185 + ], + "score": 0.997, + "content": "impractical to compute", + "type": "text" + }, + { + "bbox": [ + 198, + 174, + 207, + 184 + ], + "score": 0.8, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 173, + 398, + 185 + ], + "score": 0.97, + "content": ", all the LDBs of a GNN. The number of LDBs in", + "type": "text" + }, + { + "bbox": [ + 398, + 174, + 408, + 184 + ], + "score": 0.83, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 173, + 505, + 185 + ], + "score": 0.985, + "content": "of a GNN is exponential", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "score": 0.991, + "content": "with respect to the number of neurons in the worst case [27].To address this challenge,we substitute", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 195, + 251, + 208 + ], + "spans": [ + { + "bbox": [ + 107, + 197, + 116, + 207 + ], + "score": 0.77, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 195, + 167, + 208 + ], + "score": 0.989, + "content": "by a sample", + "type": "text" + }, + { + "bbox": [ + 168, + 195, + 177, + 207 + ], + "score": 0.86, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 195, + 237, + 208 + ], + "score": 0.996, + "content": "of LDBs from", + "type": "text" + }, + { + "bbox": [ + 238, + 195, + 247, + 207 + ], + "score": 0.85, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 195, + 251, + 208 + ], + "score": 0.831, + "content": "：", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 163, + 505, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 212, + 505, + 246 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 186, + 225 + ], + "score": 0.991, + "content": "A LDB in the space", + "type": "text" + }, + { + "bbox": [ + 186, + 212, + 200, + 223 + ], + "score": 0.88, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 212, + 268, + 225 + ], + "score": 0.996, + "content": "can be written as", + "type": "text" + }, + { + "bbox": [ + 269, + 212, + 325, + 224 + ], + "score": 0.91, + "content": "\\mathbf { w } ^ { \\top } \\mathbf { x } + b = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 212, + 364, + 225 + ], + "score": 0.968, + "content": ",where is", + "type": "text" + }, + { + "bbox": [ + 364, + 212, + 396, + 223 + ], + "score": 0.92, + "content": "\\mathbf { x } \\in \\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 212, + 506, + 225 + ], + "score": 0.979, + "content": "is a variable, w is the basis", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 146, + 236 + ], + "score": 0.963, + "content": "term, and", + "type": "text" + }, + { + "bbox": [ + 147, + 224, + 153, + 234 + ], + "score": 0.77, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 223, + 391, + 236 + ], + "score": 0.985, + "content": "corresponds to the bias.Following [4], for any input graph", + "type": "text" + }, + { + "bbox": [ + 391, + 224, + 400, + 234 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 223, + 505, + 236 + ], + "score": 0.991, + "content": ",a linear boundary can be", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 234, + 233, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 164, + 247 + ], + "score": 1.0, + "content": "sampled from", + "type": "text" + }, + { + "bbox": [ + 164, + 235, + 174, + 244 + ], + "score": 0.81, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 234, + 233, + 247 + ], + "score": 0.977, + "content": "by computing", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 212, + 506, + 247 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 250, + 414, + 276 + ], + "lines": [ + { + "bbox": [ + 195, + 250, + 414, + 276 + ], + "spans": [ + { + "bbox": [ + 195, + 250, + 414, + 276 + ], + "score": 0.92, + "content": "\\mathbf { w } = \\frac { \\partial \\left( \\operatorname* { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) - \\operatorname* { m a x } _ { 2 } ( \\phi _ { f c } ( \\alpha ) ) \\right) } { \\partial \\alpha } | _ { \\alpha = \\phi _ { g c } ( G ) } ,", + "type": "interline_equation", + "image_path": "2dbc5d9de809bb0aacc312e490149627bcf7576859a17ad5d232a303bbbb6bf6.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 195, + 250, + 414, + 276 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 280, + 123, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 123, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 123, + 291 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 280, + 123, + 291 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 297, + 420, + 312 + ], + "lines": [ + { + "bbox": [ + 189, + 297, + 420, + 312 + ], + "spans": [ + { + "bbox": [ + 189, + 297, + 420, + 312 + ], + "score": 0.88, + "content": "b = \\mathrm { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) - \\mathrm { m a x } _ { 2 } ( \\phi _ { f c } ( \\alpha ) ) - \\mathbf { w } ^ { T } { \\alpha } | _ { \\alpha = \\phi _ { g c } ( G ) } ,", + "type": "interline_equation", + "image_path": "79c9953979154b44111c9874d086d71cda854549201b41bbe069b902cc4cfe90.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 189, + 297, + 420, + 312 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 506, + 362 + ], + "lines": [ + { + "bbox": [ + 104, + 316, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 132, + 331 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 317, + 198, + 330 + ], + "score": 0.91, + "content": "\\operatorname* { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 316, + 217, + 331 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 317, + 278, + 330 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { 2 } \\bigl ( \\phi _ { f c } ( \\pmb { \\alpha } ) \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 316, + 506, + 331 + ], + "score": 0.992, + "content": " are the largest and the second largest values in the vector", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 328, + 507, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 137, + 340 + ], + "score": 0.92, + "content": "\\phi _ { f c } ( \\alpha )", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 328, + 277, + 341 + ], + "score": 0.996, + "content": ",respectively. Given an input graph", + "type": "text" + }, + { + "bbox": [ + 277, + 329, + 286, + 338 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 328, + 468, + 341 + ], + "score": 0.99, + "content": ",Equations (2) and (3) identify one LDB from", + "type": "text" + }, + { + "bbox": [ + 468, + 329, + 478, + 338 + ], + "score": 0.77, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 328, + 507, + 341 + ], + "score": 0.976, + "content": ".Thus,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 328, + 351 + ], + "score": 0.983, + "content": "we can sample a subset of input graphs uniformly from", + "type": "text" + }, + { + "bbox": [ + 328, + 339, + 337, + 348 + ], + "score": 0.84, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 339, + 506, + 351 + ], + "score": 0.99, + "content": ",and use Equations (2) and (3) to derive a", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 349, + 221, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 185, + 362 + ], + "score": 0.972, + "content": "sample of LDBs as", + "type": "text" + }, + { + "bbox": [ + 185, + 349, + 216, + 361 + ], + "score": 0.92, + "content": "\\tilde { \\mathcal { H } } \\subset \\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 351, + 221, + 362 + ], + "score": 0.813, + "content": "：", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 316, + 507, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 367, + 420, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 420, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 185, + 383 + ], + "score": 0.978, + "content": "Now, we substitute", + "type": "text" + }, + { + "bbox": [ + 185, + 369, + 195, + 378 + ], + "score": 0.82, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 366, + 271, + 383 + ], + "score": 0.991, + "content": "in Equation (1) by", + "type": "text" + }, + { + "bbox": [ + 271, + 367, + 281, + 379 + ], + "score": 0.87, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 366, + 420, + 383 + ], + "score": 0.982, + "content": "to produce the following problem.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 366, + 420, + 383 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 384, + 367, + 406 + ], + "lines": [ + { + "bbox": [ + 242, + 384, + 367, + 406 + ], + "spans": [ + { + "bbox": [ + 242, + 384, + 367, + 406 + ], + "score": 0.92, + "content": "\\operatorname* { m a x } _ { \\mathcal { P } \\subseteq \\tilde { \\mathcal { H } } } g ( \\mathcal { P } , c ) , \\mathrm { ~ s . t . ~ } h ( \\mathcal { P } , c ) \\leq \\delta ,", + "type": "interline_equation", + "image_path": "f52df860ca40a98e65cc25516b494f56cafd46c079811fa063e656225dbb0792.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 242, + 384, + 367, + 406 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 133, + 424 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 411, + 160, + 422 + ], + "score": 0.91, + "content": "\\delta \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 410, + 451, + 424 + ], + "score": 0.983, + "content": "is a tolerance parameter to keep this problem feasible. The parameter", + "type": "text" + }, + { + "bbox": [ + 451, + 411, + 457, + 421 + ], + "score": 0.82, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 410, + 506, + 424 + ], + "score": 0.982, + "content": "is required", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 422, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 187, + 436 + ], + "score": 0.998, + "content": "because substituting", + "type": "text" + }, + { + "bbox": [ + 188, + 423, + 198, + 434 + ], + "score": 0.78, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 423, + 210, + 436 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 211, + 422, + 221, + 434 + ], + "score": 0.87, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 423, + 303, + 436 + ], + "score": 0.991, + "content": "ignores the LDBs in", + "type": "text" + }, + { + "bbox": [ + 303, + 422, + 330, + 435 + ], + "score": 0.92, + "content": "\\mathcal { H } \\backslash \\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 423, + 440, + 436 + ], + "score": 0.951, + "content": ". Thus, the convex polytope", + "type": "text" + }, + { + "bbox": [ + 440, + 423, + 470, + 435 + ], + "score": 0.93, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 423, + 506, + 436 + ], + "score": 0.944, + "content": " induced", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 434, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 213, + 450 + ], + "score": 0.981, + "content": "by subset of boundaries in", + "type": "text" + }, + { + "bbox": [ + 213, + 434, + 223, + 446 + ], + "score": 0.85, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 435, + 433, + 450 + ], + "score": 0.989, + "content": "may contain instances that are not predicted as class", + "type": "text" + }, + { + "bbox": [ + 434, + 438, + 439, + 446 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 435, + 506, + 450 + ], + "score": 0.988, + "content": ". We directly set", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 446, + 453, + 461 + ], + "spans": [ + { + "bbox": [ + 107, + 446, + 156, + 460 + ], + "score": 0.93, + "content": "\\boldsymbol { \\delta } \\dot { = } \\boldsymbol { h } ( \\tilde { \\mathcal { H } } , \\boldsymbol { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 447, + 281, + 461 + ], + "score": 0.995, + "content": ", which is the smallest value of", + "type": "text" + }, + { + "bbox": [ + 281, + 448, + 287, + 458 + ], + "score": 0.78, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 447, + 453, + 461 + ], + "score": 0.999, + "content": "that keeps the practical problem feasible.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 410, + 506, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 464, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "score": 0.997, + "content": "The problem in Equation (4) can be proven to be a Submodular Cost Submodular Cover (SCSC)", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 222, + 487 + ], + "score": 0.978, + "content": "problem [18] (see Appendix", + "type": "text" + }, + { + "bbox": [ + 223, + 475, + 232, + 485 + ], + "score": 0.44, + "content": "\\mathbf { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 474, + 506, + 487 + ], + "score": 0.976, + "content": "for proof) that is well known to be NP-hard [5]. We adopt a greedy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 495, + 497 + ], + "score": 0.989, + "content": "boundary selection method to find a good solution to this problem [40]. Specifically, we initialize", + "type": "text" + }, + { + "bbox": [ + 495, + 486, + 504, + 495 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 496, + 397, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 343, + 510 + ], + "score": 0.98, + "content": " as an empty set,and then iteratively select a new boundary", + "type": "text" + }, + { + "bbox": [ + 343, + 497, + 350, + 507 + ], + "score": 0.8, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 497, + 373, + 510 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 373, + 496, + 383, + 507 + ], + "score": 0.85, + "content": "\\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 497, + 397, + 510 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 464, + 506, + 510 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 514, + 391, + 544 + ], + "lines": [ + { + "bbox": [ + 219, + 514, + 391, + 544 + ], + "spans": [ + { + "bbox": [ + 219, + 514, + 391, + 544 + ], + "score": 0.94, + "content": "h = \\arg \\operatorname* { m i n } _ { h \\in \\tilde { \\mathcal { H } } \\backslash \\mathcal { P } } \\frac { g ( \\mathcal { P } , c ) - g ( \\mathcal { P } \\cup \\{ h \\} , c ) + \\epsilon } { h ( \\mathcal { P } , c ) - h ( \\mathcal { P } \\cup \\{ h \\} , c ) } ,", + "type": "interline_equation", + "image_path": "1f9f845a8a01b5db73f4894fc93b9fa1e911127e6b892e443b6edc0c7061df65.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 514, + 391, + 529.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 219, + 529.0, + 391, + 544.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 133, + 562 + ], + "score": 0.999, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 550, + 227, + 562 + ], + "score": 0.91, + "content": "g ( \\mathcal { P } , c ) - g ( \\mathcal { P } \\cup \\{ h \\} , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 549, + 298, + 562 + ], + "score": 0.984, + "content": " is the decrease of", + "type": "text" + }, + { + "bbox": [ + 299, + 550, + 329, + 561 + ], + "score": 0.91, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 549, + 382, + 562 + ], + "score": 0.964, + "content": "when adding", + "type": "text" + }, + { + "bbox": [ + 382, + 550, + 389, + 560 + ], + "score": 0.82, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 549, + 407, + 562 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 408, + 550, + 416, + 560 + ], + "score": 0.83, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 549, + 437, + 562 + ], + "score": 0.992, + "content": ",and", + "type": "text" + }, + { + "bbox": [ + 437, + 550, + 506, + 562 + ], + "score": 0.9, + "content": "h ( \\mathcal { P } , c ) - h ( \\mathcal { P } \\cup", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 559, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 107, + 560, + 136, + 572 + ], + "score": 0.91, + "content": "\\{ h \\} , c \\bar { ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 559, + 206, + 573 + ], + "score": 0.999, + "content": "is the decrease of", + "type": "text" + }, + { + "bbox": [ + 207, + 561, + 237, + 572 + ], + "score": 0.91, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 559, + 290, + 573 + ], + "score": 1.0, + "content": "when adding", + "type": "text" + }, + { + "bbox": [ + 291, + 561, + 298, + 570 + ], + "score": 0.81, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 559, + 316, + 573 + ], + "score": 0.999, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 316, + 562, + 325, + 570 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 559, + 350, + 573 + ], + "score": 0.925, + "content": ".Both", + "type": "text" + }, + { + "bbox": [ + 351, + 560, + 381, + 572 + ], + "score": 0.93, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 559, + 398, + 573 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 560, + 429, + 572 + ], + "score": 0.92, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 559, + 506, + 573 + ], + "score": 0.999, + "content": "are non-increasing", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 160, + 586 + ], + "score": 1.0, + "content": "when adding", + "type": "text" + }, + { + "bbox": [ + 160, + 572, + 188, + 583 + ], + "score": 0.89, + "content": "h \\in \\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 572, + 207, + 586 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 208, + 573, + 216, + 583 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 572, + 348, + 586 + ], + "score": 0.999, + "content": "because adding a new boundary", + "type": "text" + }, + { + "bbox": [ + 348, + 573, + 355, + 583 + ], + "score": 0.79, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 572, + 505, + 586 + ], + "score": 0.998, + "content": "may only exclude some graphs from", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 582, + 224, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 189, + 597 + ], + "score": 0.999, + "content": "the convex polytope", + "type": "text" + }, + { + "bbox": [ + 190, + 584, + 219, + 595 + ], + "score": 0.92, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 582, + 224, + 597 + ], + "score": 0.535, + "content": "：", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 549, + 506, + 597 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 343, + 614 + ], + "score": 0.982, + "content": "Intuitively, in each iteration,Equation (5) selects a boundary", + "type": "text" + }, + { + "bbox": [ + 343, + 600, + 371, + 612 + ], + "score": 0.92, + "content": "h \\in \\tilde { \\mathcal { H } }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 600, + 437, + 614 + ], + "score": 0.981, + "content": " such that adding", + "type": "text" + }, + { + "bbox": [ + 438, + 602, + 444, + 611 + ], + "score": 0.81, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 600, + 463, + 614 + ], + "score": 1.0, + "content": "into", + "type": "text" + }, + { + "bbox": [ + 463, + 602, + 471, + 611 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 600, + 506, + 614 + ], + "score": 1.0, + "content": "reduces", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 107, + 612, + 137, + 624 + ], + "score": 0.91, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 611, + 227, + 624 + ], + "score": 0.982, + "content": " the least and reduces", + "type": "text" + }, + { + "bbox": [ + 227, + 612, + 258, + 623 + ], + "score": 0.91, + "content": "\\bar { h } ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 611, + 449, + 624 + ], + "score": 0.993, + "content": "the most. In this way,we can quickly reduce", + "type": "text" + }, + { + "bbox": [ + 449, + 612, + 480, + 624 + ], + "score": 0.92, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 611, + 506, + 624 + ], + "score": 0.992, + "content": "to be", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 158, + 635 + ], + "score": 0.999, + "content": "smaller than", + "type": "text" + }, + { + "bbox": [ + 158, + 623, + 164, + 632 + ], + "score": 0.8, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 622, + 243, + 635 + ], + "score": 0.999, + "content": "without decreasing", + "type": "text" + }, + { + "bbox": [ + 244, + 623, + 273, + 634 + ], + "score": 0.91, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 622, + 506, + 635 + ], + "score": 0.987, + "content": " too much, which produces a good feasible solution to the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 286, + 645 + ], + "score": 0.991, + "content": " practical problem.We add a small constant", + "type": "text" + }, + { + "bbox": [ + 286, + 635, + 292, + 642 + ], + "score": 0.74, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 632, + 506, + 645 + ], + "score": 0.986, + "content": "to the numerator such that, when there are multiple", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 642, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 162, + 655 + ], + "score": 0.999, + "content": "candidates of", + "type": "text" + }, + { + "bbox": [ + 162, + 644, + 169, + 653 + ], + "score": 0.81, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 642, + 252, + 655 + ], + "score": 0.999, + "content": "that do not decrease", + "type": "text" + }, + { + "bbox": [ + 252, + 643, + 282, + 655 + ], + "score": 0.92, + "content": "g ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 642, + 374, + 655 + ], + "score": 0.967, + "content": ", we can still select the", + "type": "text" + }, + { + "bbox": [ + 374, + 644, + 381, + 653 + ], + "score": 0.8, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 642, + 432, + 655 + ], + "score": 0.999, + "content": "that reduces", + "type": "text" + }, + { + "bbox": [ + 433, + 643, + 464, + 655 + ], + "score": 0.93, + "content": "h ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 642, + 505, + 655 + ], + "score": 0.96, + "content": "the most.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 600, + 506, + 655 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 658, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 502, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 477, + 671 + ], + "score": 0.982, + "content": "We apply a peeling-off strategy to iteratively extract multiple decision regions.For each class", + "type": "text" + }, + { + "bbox": [ + 477, + 659, + 502, + 669 + ], + "score": 0.89, + "content": "c \\in C", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 363, + 681 + ], + "score": 0.984, + "content": "we first solve the practical problem once to find a decision region", + "type": "text" + }, + { + "bbox": [ + 364, + 670, + 393, + 681 + ], + "score": 0.92, + "content": "r ( \\mathcal { P } , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 669, + 505, + 681 + ], + "score": 0.99, + "content": ", then we remove the graphs", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 679, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 154, + 693 + ], + "score": 0.998, + "content": "covered by", + "type": "text" + }, + { + "bbox": [ + 155, + 680, + 185, + 692 + ], + "score": 0.92, + "content": "r ( \\mathcal { P } , \\bar { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 679, + 209, + 693 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 210, + 681, + 223, + 691 + ], + "score": 0.87, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 679, + 442, + 693 + ], + "score": 0.994, + "content": ". If there are remaining graphs predicted as the class", + "type": "text" + }, + { + "bbox": [ + 443, + 682, + 448, + 690 + ], + "score": 0.69, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 679, + 506, + 693 + ], + "score": 0.94, + "content": ", we continue", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 411, + 703 + ], + "score": 0.984, + "content": " finding the decision regions using the remaining graphs until all the graphs in", + "type": "text" + }, + { + "bbox": [ + 411, + 691, + 424, + 701 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 690, + 506, + 703 + ], + "score": 0.973, + "content": "are removed. When", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 175, + 714 + ], + "score": 0.999, + "content": "all the graphs in", + "type": "text" + }, + { + "bbox": [ + 176, + 701, + 189, + 712 + ], + "score": 0.89, + "content": "D _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 699, + 303, + 714 + ], + "score": 0.976, + "content": "are removed for each class", + "type": "text" + }, + { + "bbox": [ + 303, + 703, + 329, + 711 + ], + "score": 0.91, + "content": "c \\in C", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 699, + 507, + 714 + ], + "score": 0.972, + "content": ", we stop the iteration and return the set of", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 711, + 217, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 217, + 723 + ], + "score": 0.996, + "content": "decision regions we found.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 659, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 233, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 234, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 234, + 86 + ], + "score": 0.999, + "content": "4.3Producing Explanations", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 92, + 505, + 145 + ], + "lines": [ + { + "bbox": [ + 105, + 91, + 506, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 506, + 104 + ], + "score": 0.99, + "content": "In this section, we introduce how to use the LDBs of decision regions to train a neural network that", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "score": 0.988, + "content": "produces a robust counterfactual explanation as a small subset of edges of an input graph. We form", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "score": 0.994, + "content": " explanations as a subset of edges because GNNs make decisions by aggregating messages passed on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "score": 0.982, + "content": "edges. Using edges instead of vertices as explanations can provide beter insights on the decision", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 133, + 170, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 170, + 146 + ], + "score": 0.978, + "content": "logic of GNNs.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 107, + 156, + 255, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 256, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 256, + 169 + ], + "score": 0.974, + "content": "4.3.1The Neural Network Model", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 151, + 187 + ], + "score": 0.987, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 152, + 175, + 162, + 186 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 175, + 439, + 187 + ], + "score": 0.995, + "content": "the neural network to generate a subset of edges of an input graph", + "type": "text" + }, + { + "bbox": [ + 440, + 175, + 449, + 185 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 175, + 505, + 187 + ], + "score": 0.972, + "content": "as the robust", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 287, + 198 + ], + "score": 0.998, + "content": "counterfactual explanation on the prediction", + "type": "text" + }, + { + "bbox": [ + 288, + 185, + 309, + 197 + ], + "score": 0.9, + "content": "\\phi ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 185, + 314, + 198 + ], + "score": 0.887, + "content": "：", + "type": "text" + }, + { + "bbox": [ + 314, + 186, + 321, + 195 + ], + "score": 0.69, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 185, + 505, + 198 + ], + "score": 0.977, + "content": "represents the set of parameters of the neural", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 321, + 208 + ], + "score": 0.985, + "content": "network. For experiments, our explanation network", + "type": "text" + }, + { + "bbox": [ + 321, + 196, + 328, + 207 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 195, + 506, + 208 + ], + "score": 0.974, + "content": "consists of 2 fully connected layers with a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 206, + 305, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 305, + 217 + ], + "score": 0.981, + "content": "ReLU activation and the hidden dimension of 64.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 221, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 231, + 235 + ], + "score": 0.993, + "content": "For any two connected vertices", + "type": "text" + }, + { + "bbox": [ + 231, + 224, + 241, + 232 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 221, + 258, + 235 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 259, + 223, + 269, + 234 + ], + "score": 0.85, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 221, + 281, + 235 + ], + "score": 0.998, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 281, + 222, + 290, + 232 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 221, + 334, + 235 + ], + "score": 0.986, + "content": ",denote by", + "type": "text" + }, + { + "bbox": [ + 335, + 224, + 345, + 233 + ], + "score": 0.86, + "content": "\\mathbf { z } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 221, + 362, + 235 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 362, + 223, + 373, + 234 + ], + "score": 0.87, + "content": "\\mathbf { z } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 221, + 505, + 235 + ], + "score": 0.987, + "content": "the embeddings produced by the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 469, + 245 + ], + "score": 0.981, + "content": "last convolution layer of the GNN for the two vertices, respectively. The neural network", + "type": "text" + }, + { + "bbox": [ + 470, + 233, + 480, + 244 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 231, + 505, + 245 + ], + "score": 1.0, + "content": "takes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 116, + 254 + ], + "score": 0.84, + "content": "\\mathbf { z } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 243, + 134, + 255 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 135, + 244, + 145, + 255 + ], + "score": 0.86, + "content": "\\mathbf { z } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 243, + 397, + 255 + ], + "score": 0.974, + "content": " as the input and outputs the probability for the edge between", + "type": "text" + }, + { + "bbox": [ + 398, + 245, + 407, + 254 + ], + "score": 0.84, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 243, + 425, + 255 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 426, + 244, + 436, + 255 + ], + "score": 0.86, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 243, + 504, + 255 + ], + "score": 0.996, + "content": "to be part of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 253, + 249, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 249, + 266 + ], + "score": 0.991, + "content": "explanation. This can be written as", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 267, + 271, + 343, + 285 + ], + "lines": [ + { + "bbox": [ + 267, + 271, + 343, + 285 + ], + "spans": [ + { + "bbox": [ + 267, + 271, + 343, + 285 + ], + "score": 0.92, + "content": "\\begin{array} { r } { { \\bf M } _ { i j } = f _ { \\theta } ( { \\bf z } _ { i } , { \\bf z } _ { j } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "dd0dc1b00ca37ef79ae9976d880a0ec02cf8d2aade732ccacacb9c981b208e29.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 267, + 271, + 343, + 285 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 291, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 133, + 304 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 291, + 152, + 303 + ], + "score": 0.9, + "content": "\\mathbf { M } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 291, + 339, + 304 + ], + "score": 0.996, + "content": "denotes the probability that the edge between", + "type": "text" + }, + { + "bbox": [ + 339, + 294, + 348, + 302 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 291, + 367, + 304 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 293, + 377, + 303 + ], + "score": 0.84, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 291, + 506, + 304 + ], + "score": 0.981, + "content": "is contained in the explanation.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 301, + 408, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 233, + 314 + ], + "score": 0.972, + "content": "When there is no edge between", + "type": "text" + }, + { + "bbox": [ + 233, + 304, + 243, + 313 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 301, + 261, + 314 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 261, + 303, + 271, + 314 + ], + "score": 0.87, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 301, + 304, + 314 + ], + "score": 0.983, + "content": ",that is,", + "type": "text" + }, + { + "bbox": [ + 304, + 302, + 339, + 314 + ], + "score": 0.92, + "content": "{ \\bf A } _ { i j } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 301, + 370, + 314 + ], + "score": 0.942, + "content": ",we set", + "type": "text" + }, + { + "bbox": [ + 371, + 302, + 408, + 314 + ], + "score": 0.92, + "content": "\\mathbf { \\check { M } } _ { i j } = 0", + "type": "inline_equation" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 504, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 185, + 330 + ], + "score": 0.999, + "content": "For an input graph", + "type": "text" + }, + { + "bbox": [ + 186, + 317, + 237, + 330 + ], + "score": 0.93, + "content": "G = \\{ V , E \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 317, + 260, + 330 + ], + "score": 0.998, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 260, + 320, + 267, + 327 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 317, + 422, + 330 + ], + "score": 0.997, + "content": "vertices and a trained neural network", + "type": "text" + }, + { + "bbox": [ + 423, + 318, + 451, + 329 + ], + "score": 0.28, + "content": "f _ { \\boldsymbol { \\theta } } , \\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 317, + 473, + 330 + ], + "score": 0.992, + "content": "is an", + "type": "text" + }, + { + "bbox": [ + 474, + 320, + 481, + 327 + ], + "score": 0.75, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 317, + 496, + 330 + ], + "score": 0.957, + "content": "-by-", + "type": "text" + }, + { + "bbox": [ + 497, + 320, + 504, + 327 + ], + "score": 0.72, + "content": "\\mathbf { \\nabla } \\cdot n", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 328, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 506, + 340 + ], + "score": 0.996, + "content": "matrix that carries the complete information to generate a robust counterfactual explanation as a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 220, + 351 + ], + "score": 0.982, + "content": "subset of edges, denoted by", + "type": "text" + }, + { + "bbox": [ + 220, + 339, + 249, + 350 + ], + "score": 0.92, + "content": "S \\subseteq E", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 339, + 343, + 351 + ], + "score": 0.986, + "content": ". Concretely,we obtain", + "type": "text" + }, + { + "bbox": [ + 344, + 339, + 352, + 348 + ], + "score": 0.83, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 339, + 466, + 351 + ], + "score": 0.994, + "content": "by selecting all the edges in", + "type": "text" + }, + { + "bbox": [ + 467, + 339, + 476, + 348 + ], + "score": 0.84, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 339, + 505, + 351 + ], + "score": 1.0, + "content": "whose", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 349, + 297, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 205, + 361 + ], + "score": 0.998, + "content": "corresponding entries in", + "type": "text" + }, + { + "bbox": [ + 205, + 349, + 218, + 359 + ], + "score": 0.44, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 349, + 297, + 361 + ], + "score": 0.976, + "content": "are larger than O.5.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 107, + 371, + 216, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 216, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 205, + 386 + ], + "score": 0.998, + "content": "4.3.2Training Model", + "type": "text" + }, + { + "bbox": [ + 206, + 372, + 216, + 383 + ], + "score": 0.82, + "content": "f _ { \\theta }", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 186, + 403 + ], + "score": 0.999, + "content": "For an input graph", + "type": "text" + }, + { + "bbox": [ + 186, + 390, + 235, + 402 + ], + "score": 0.92, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 389, + 283, + 403 + ], + "score": 0.977, + "content": ", denote by", + "type": "text" + }, + { + "bbox": [ + 283, + 390, + 313, + 401 + ], + "score": 0.9, + "content": "S \\subseteq E", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 389, + 450, + 403 + ], + "score": 0.977, + "content": " the subset of edges produced by", + "type": "text" + }, + { + "bbox": [ + 450, + 390, + 460, + 401 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 389, + 505, + 403 + ], + "score": 0.999, + "content": "to explain", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 401, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 165, + 412 + ], + "score": 0.998, + "content": "the prediction", + "type": "text" + }, + { + "bbox": [ + 166, + 401, + 187, + 412 + ], + "score": 0.91, + "content": "\\phi ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 401, + 327, + 412 + ], + "score": 0.993, + "content": ",our goal is to train a good model", + "type": "text" + }, + { + "bbox": [ + 328, + 401, + 338, + 412 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 401, + 505, + 412 + ], + "score": 0.992, + "content": "such that the prediction on the subgraph", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 411, + 507, + 424 + ], + "spans": [ + { + "bbox": [ + 107, + 411, + 121, + 422 + ], + "score": 0.88, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 411, + 170, + 424 + ], + "score": 0.998, + "content": "induced by", + "type": "text" + }, + { + "bbox": [ + 171, + 412, + 178, + 421 + ], + "score": 0.83, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 411, + 202, + 424 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 202, + 412, + 211, + 421 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 411, + 285, + 424 + ], + "score": 0.995, + "content": "is consistent with", + "type": "text" + }, + { + "bbox": [ + 286, + 411, + 308, + 423 + ], + "score": 0.92, + "content": "\\phi ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 411, + 416, + 424 + ], + "score": 0.974, + "content": "; and deleting the edges in", + "type": "text" + }, + { + "bbox": [ + 417, + 411, + 425, + 421 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 411, + 448, + 424 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 449, + 411, + 457, + 421 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 411, + 507, + 424 + ], + "score": 0.986, + "content": "produces a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 421, + 484, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 188, + 434 + ], + "score": 0.985, + "content": "remainder subgraph", + "type": "text" + }, + { + "bbox": [ + 189, + 422, + 213, + 434 + ], + "score": 0.92, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 421, + 323, + 434 + ], + "score": 0.971, + "content": "such that the prediction on", + "type": "text" + }, + { + "bbox": [ + 323, + 422, + 348, + 434 + ], + "score": 0.94, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 421, + 457, + 434 + ], + "score": 0.998, + "content": "changes significantly from", + "type": "text" + }, + { + "bbox": [ + 458, + 421, + 480, + 433 + ], + "score": 0.92, + "content": "{ \\bar { \\phi } } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 421, + 484, + 434 + ], + "score": 0.629, + "content": "：", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 174, + 451 + ], + "score": 0.989, + "content": "Since producing", + "type": "text" + }, + { + "bbox": [ + 174, + 439, + 182, + 448 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 438, + 196, + 451 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 196, + 438, + 207, + 450 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 438, + 506, + 451 + ], + "score": 0.989, + "content": " is a discrete operation that is hard to incorporate in an end-to-end training", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 315, + 461 + ], + "score": 0.98, + "content": " process, we define two proxy graphs to approximate", + "type": "text" + }, + { + "bbox": [ + 315, + 449, + 329, + 460 + ], + "score": 0.89, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 448, + 347, + 461 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 348, + 449, + 372, + 461 + ], + "score": 0.91, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 448, + 505, + 461 + ], + "score": 0.991, + "content": ",respectively,such that the proxy", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 210, + 473 + ], + "score": 0.983, + "content": " graphs are determined by", + "type": "text" + }, + { + "bbox": [ + 210, + 461, + 216, + 470 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 460, + 506, + 473 + ], + "score": 0.999, + "content": "through continuous functions that can be smoothly incorporated into an", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 471, + 220, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 220, + 483 + ], + "score": 0.995, + "content": "end-to-end training process.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 188, + 499 + ], + "score": 0.988, + "content": "The proxy graph of", + "type": "text" + }, + { + "bbox": [ + 188, + 487, + 203, + 497 + ], + "score": 0.87, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 485, + 254, + 499 + ], + "score": 0.975, + "content": ", denoted by", + "type": "text" + }, + { + "bbox": [ + 255, + 487, + 268, + 497 + ], + "score": 0.87, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 485, + 368, + 499 + ], + "score": 0.98, + "content": ", is defined by regarding", + "type": "text" + }, + { + "bbox": [ + 369, + 487, + 381, + 497 + ], + "score": 0.39, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 485, + 425, + 499 + ], + "score": 0.997, + "content": "instead of", + "type": "text" + }, + { + "bbox": [ + 425, + 487, + 435, + 497 + ], + "score": 0.44, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 485, + 506, + 499 + ], + "score": 0.982, + "content": " as the adjacency", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 495, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 104, + 495, + 169, + 510 + ], + "score": 0.989, + "content": "matrix. That is,", + "type": "text" + }, + { + "bbox": [ + 170, + 497, + 182, + 508 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 495, + 339, + 510 + ], + "score": 0.991, + "content": " has exactly the same graph structure as", + "type": "text" + }, + { + "bbox": [ + 339, + 497, + 348, + 507 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 495, + 446, + 510 + ], + "score": 0.99, + "content": ", but the edge weights of", + "type": "text" + }, + { + "bbox": [ + 446, + 497, + 459, + 508 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 495, + 506, + 510 + ], + "score": 0.984, + "content": "is given by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 507, + 432, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 160, + 519 + ], + "score": 0.995, + "content": "the entries in", + "type": "text" + }, + { + "bbox": [ + 160, + 507, + 173, + 517 + ], + "score": 0.32, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 507, + 215, + 519 + ], + "score": 0.998, + "content": "instead of", + "type": "text" + }, + { + "bbox": [ + 216, + 507, + 225, + 517 + ], + "score": 0.47, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 507, + 307, + 519 + ], + "score": 0.973, + "content": ".Here,the subscript", + "type": "text" + }, + { + "bbox": [ + 307, + 507, + 313, + 517 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 507, + 342, + 519 + ], + "score": 1.0, + "content": "means", + "type": "text" + }, + { + "bbox": [ + 343, + 507, + 356, + 518 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 507, + 426, + 519 + ], + "score": 0.996, + "content": "is determined by", + "type": "text" + }, + { + "bbox": [ + 426, + 508, + 432, + 517 + ], + "score": 0.81, + "content": "\\theta", + "type": "inline_equation" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 106, + 523, + 503, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 504, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 189, + 537 + ], + "score": 0.958, + "content": "The proxy graph of", + "type": "text" + }, + { + "bbox": [ + 189, + 524, + 213, + 536 + ], + "score": 0.91, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 522, + 266, + 537 + ], + "score": 0.981, + "content": ", denoted by", + "type": "text" + }, + { + "bbox": [ + 266, + 523, + 279, + 535 + ], + "score": 0.92, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 522, + 439, + 537 + ], + "score": 0.986, + "content": ", also have the same graph structure as", + "type": "text" + }, + { + "bbox": [ + 439, + 524, + 448, + 533 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 522, + 504, + 537 + ], + "score": 0.972, + "content": ", but the edge", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 533, + 346, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 254, + 546 + ], + "score": 0.986, + "content": "weight between each pair of vertices", + "type": "text" + }, + { + "bbox": [ + 254, + 536, + 263, + 545 + ], + "score": 0.87, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 533, + 281, + 546 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 282, + 535, + 292, + 546 + ], + "score": 0.87, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 533, + 346, + 546 + ], + "score": 0.981, + "content": "is defined as", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 550, + 372, + 579 + ], + "lines": [ + { + "bbox": [ + 239, + 550, + 372, + 579 + ], + "spans": [ + { + "bbox": [ + 239, + 550, + 372, + 579 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathbf { M } _ { i j } ^ { \\prime } = \\left\\{ { 1 - \\mathbf { M } _ { i j } \\quad \\mathrm { ~ i f ~ } \\mathbf { A } _ { i j } = 1 } \\atop { 0 } \\right. } \\end{array}", + "type": "interline_equation", + "image_path": "afaae5efb14f27284c6a4b7446792ec5cf01118d4490a60597c99ebbd88ab0af.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 550, + 372, + 564.5 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 239, + 564.5, + 372, + 579.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 589, + 504, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 208, + 604 + ], + "score": 0.986, + "content": "The edge weights of both", + "type": "text" + }, + { + "bbox": [ + 209, + 590, + 222, + 600 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 588, + 240, + 604 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 590, + 253, + 601 + ], + "score": 0.9, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 588, + 328, + 604 + ], + "score": 0.958, + "content": " are determined by", + "type": "text" + }, + { + "bbox": [ + 329, + 590, + 335, + 599 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 588, + 506, + 604 + ], + "score": 0.986, + "content": "through continuous functions, thus we can", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 600, + 398, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 194, + 613 + ], + "score": 0.993, + "content": "smoothly incorporate", + "type": "text" + }, + { + "bbox": [ + 194, + 600, + 207, + 611 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 600, + 226, + 613 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 226, + 600, + 239, + 612 + ], + "score": 0.92, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 600, + 398, + 613 + ], + "score": 0.992, + "content": "into an end-to-end training framework.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 105, + 616, + 504, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 484, + 630 + ], + "score": 0.975, + "content": " As discussed later in this section, we use a regularization term to force the value of each entry in", + "type": "text" + }, + { + "bbox": [ + 485, + 616, + 504, + 628 + ], + "score": 0.89, + "content": "\\mathbf { M } _ { i j }", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 624, + 488, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 253, + 640 + ], + "score": 0.979, + "content": "to be close to eitherOor 1,such that", + "type": "text" + }, + { + "bbox": [ + 254, + 627, + 267, + 637 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 624, + 285, + 640 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 285, + 626, + 298, + 639 + ], + "score": 0.91, + "content": "\\operatorname { \\mathrm { \\Sigma } } _ { G _ { \\theta } ^ { \\prime } } ^ { G \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 624, + 377, + 640 + ], + "score": 0.997, + "content": "better approximate", + "type": "text" + }, + { + "bbox": [ + 378, + 627, + 392, + 637 + ], + "score": 0.89, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 624, + 410, + 640 + ], + "score": 0.999, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 411, + 627, + 435, + 639 + ], + "score": 0.93, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 624, + 488, + 640 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 642, + 243, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 243, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 243, + 655 + ], + "score": 0.987, + "content": "We formulate our loss function as", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 657, + 484, + 684 + ], + "lines": [ + { + "bbox": [ + 113, + 657, + 484, + 684 + ], + "spans": [ + { + "bbox": [ + 113, + 657, + 484, + 684 + ], + "score": 0.91, + "content": "\\mathcal { L } ( \\theta ) = \\sum _ { G \\in D } \\left\\{ \\lambda \\mathcal { L } _ { s a m e } ( \\theta , G ) + ( 1 - \\lambda ) \\mathcal { L } _ { o p p } ( \\theta , G ) + \\beta \\mathcal { R } _ { s p a r s e } ( \\theta , G ) + \\mu \\mathcal { R } _ { d i s c r e t e } ( \\theta , G ) \\right\\} ,", + "type": "interline_equation", + "image_path": "cfc41b2d76850caa450e5da3d6381caaa5c3ed3056969b987fa0fd0bf367d283.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 113, + 657, + 484, + 666.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 113, + 666.0, + 484, + 675.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 113, + 675.0, + 484, + 684.0 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 690, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 133, + 702 + ], + "score": 0.923, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 690, + 172, + 702 + ], + "score": 0.86, + "content": "\\lambda \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 690, + 202, + 702 + ], + "score": 0.82, + "content": "\\beta \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 690, + 219, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 220, + 691, + 245, + 702 + ], + "score": 0.9, + "content": "\\mu \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 690, + 506, + 702 + ], + "score": 0.991, + "content": "are the hyperparameters controlling the importance of each term.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 351, + 713 + ], + "score": 0.975, + "content": "The influence of these parameters is discussed in Appendix", + "type": "text" + }, + { + "bbox": [ + 352, + 703, + 360, + 711 + ], + "score": 0.27, + "content": "\\mathbf { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 700, + 506, + 713 + ], + "score": 0.994, + "content": ".The first term of our loss function", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 276, + 723 + ], + "score": 0.99, + "content": "requires that the prediction of the GNN on", + "type": "text" + }, + { + "bbox": [ + 276, + 712, + 289, + 722 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 710, + 430, + 723 + ], + "score": 0.996, + "content": "is consistent with the prediction on", + "type": "text" + }, + { + "bbox": [ + 430, + 712, + 439, + 721 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 710, + 506, + 723 + ], + "score": 0.972, + "content": ". Intuitively, this", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 233, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 234, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 234, + 86 + ], + "score": 0.999, + "content": "4.3Producing Explanations", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 92, + 505, + 145 + ], + "lines": [ + { + "bbox": [ + 105, + 91, + 506, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 506, + 104 + ], + "score": 0.99, + "content": "In this section, we introduce how to use the LDBs of decision regions to train a neural network that", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "score": 0.988, + "content": "produces a robust counterfactual explanation as a small subset of edges of an input graph. We form", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "score": 0.994, + "content": " explanations as a subset of edges because GNNs make decisions by aggregating messages passed on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "score": 0.982, + "content": "edges. Using edges instead of vertices as explanations can provide beter insights on the decision", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 133, + 170, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 170, + 146 + ], + "score": 0.978, + "content": "logic of GNNs.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 91, + 506, + 146 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 156, + 255, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 256, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 256, + 169 + ], + "score": 0.974, + "content": "4.3.1The Neural Network Model", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 175, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 151, + 187 + ], + "score": 0.987, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 152, + 175, + 162, + 186 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 175, + 439, + 187 + ], + "score": 0.995, + "content": "the neural network to generate a subset of edges of an input graph", + "type": "text" + }, + { + "bbox": [ + 440, + 175, + 449, + 185 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 175, + 505, + 187 + ], + "score": 0.972, + "content": "as the robust", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 287, + 198 + ], + "score": 0.998, + "content": "counterfactual explanation on the prediction", + "type": "text" + }, + { + "bbox": [ + 288, + 185, + 309, + 197 + ], + "score": 0.9, + "content": "\\phi ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 185, + 314, + 198 + ], + "score": 0.887, + "content": "：", + "type": "text" + }, + { + "bbox": [ + 314, + 186, + 321, + 195 + ], + "score": 0.69, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 185, + 505, + 198 + ], + "score": 0.977, + "content": "represents the set of parameters of the neural", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 321, + 208 + ], + "score": 0.985, + "content": "network. For experiments, our explanation network", + "type": "text" + }, + { + "bbox": [ + 321, + 196, + 328, + 207 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 195, + 506, + 208 + ], + "score": 0.974, + "content": "consists of 2 fully connected layers with a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 206, + 305, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 305, + 217 + ], + "score": 0.981, + "content": "ReLU activation and the hidden dimension of 64.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 175, + 506, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 221, + 505, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 231, + 235 + ], + "score": 0.993, + "content": "For any two connected vertices", + "type": "text" + }, + { + "bbox": [ + 231, + 224, + 241, + 232 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 221, + 258, + 235 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 259, + 223, + 269, + 234 + ], + "score": 0.85, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 221, + 281, + 235 + ], + "score": 0.998, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 281, + 222, + 290, + 232 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 221, + 334, + 235 + ], + "score": 0.986, + "content": ",denote by", + "type": "text" + }, + { + "bbox": [ + 335, + 224, + 345, + 233 + ], + "score": 0.86, + "content": "\\mathbf { z } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 221, + 362, + 235 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 362, + 223, + 373, + 234 + ], + "score": 0.87, + "content": "\\mathbf { z } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 221, + 505, + 235 + ], + "score": 0.987, + "content": "the embeddings produced by the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 469, + 245 + ], + "score": 0.981, + "content": "last convolution layer of the GNN for the two vertices, respectively. The neural network", + "type": "text" + }, + { + "bbox": [ + 470, + 233, + 480, + 244 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 231, + 505, + 245 + ], + "score": 1.0, + "content": "takes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 116, + 254 + ], + "score": 0.84, + "content": "\\mathbf { z } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 243, + 134, + 255 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 135, + 244, + 145, + 255 + ], + "score": 0.86, + "content": "\\mathbf { z } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 243, + 397, + 255 + ], + "score": 0.974, + "content": " as the input and outputs the probability for the edge between", + "type": "text" + }, + { + "bbox": [ + 398, + 245, + 407, + 254 + ], + "score": 0.84, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 243, + 425, + 255 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 426, + 244, + 436, + 255 + ], + "score": 0.86, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 243, + 504, + 255 + ], + "score": 0.996, + "content": "to be part of the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 253, + 249, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 249, + 266 + ], + "score": 0.991, + "content": "explanation. This can be written as", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 221, + 505, + 266 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 267, + 271, + 343, + 285 + ], + "lines": [ + { + "bbox": [ + 267, + 271, + 343, + 285 + ], + "spans": [ + { + "bbox": [ + 267, + 271, + 343, + 285 + ], + "score": 0.92, + "content": "\\begin{array} { r } { { \\bf M } _ { i j } = f _ { \\theta } ( { \\bf z } _ { i } , { \\bf z } _ { j } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "dd0dc1b00ca37ef79ae9976d880a0ec02cf8d2aade732ccacacb9c981b208e29.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 267, + 271, + 343, + 285 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 291, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 133, + 304 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 291, + 152, + 303 + ], + "score": 0.9, + "content": "\\mathbf { M } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 291, + 339, + 304 + ], + "score": 0.996, + "content": "denotes the probability that the edge between", + "type": "text" + }, + { + "bbox": [ + 339, + 294, + 348, + 302 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 291, + 367, + 304 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 293, + 377, + 303 + ], + "score": 0.84, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 291, + 506, + 304 + ], + "score": 0.981, + "content": "is contained in the explanation.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 301, + 408, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 233, + 314 + ], + "score": 0.972, + "content": "When there is no edge between", + "type": "text" + }, + { + "bbox": [ + 233, + 304, + 243, + 313 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 301, + 261, + 314 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 261, + 303, + 271, + 314 + ], + "score": 0.87, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 301, + 304, + 314 + ], + "score": 0.983, + "content": ",that is,", + "type": "text" + }, + { + "bbox": [ + 304, + 302, + 339, + 314 + ], + "score": 0.92, + "content": "{ \\bf A } _ { i j } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 301, + 370, + 314 + ], + "score": 0.942, + "content": ",we set", + "type": "text" + }, + { + "bbox": [ + 371, + 302, + 408, + 314 + ], + "score": 0.92, + "content": "\\mathbf { \\check { M } } _ { i j } = 0", + "type": "inline_equation" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 291, + 506, + 314 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 504, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 185, + 330 + ], + "score": 0.999, + "content": "For an input graph", + "type": "text" + }, + { + "bbox": [ + 186, + 317, + 237, + 330 + ], + "score": 0.93, + "content": "G = \\{ V , E \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 317, + 260, + 330 + ], + "score": 0.998, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 260, + 320, + 267, + 327 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 317, + 422, + 330 + ], + "score": 0.997, + "content": "vertices and a trained neural network", + "type": "text" + }, + { + "bbox": [ + 423, + 318, + 451, + 329 + ], + "score": 0.28, + "content": "f _ { \\boldsymbol { \\theta } } , \\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 317, + 473, + 330 + ], + "score": 0.992, + "content": "is an", + "type": "text" + }, + { + "bbox": [ + 474, + 320, + 481, + 327 + ], + "score": 0.75, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 317, + 496, + 330 + ], + "score": 0.957, + "content": "-by-", + "type": "text" + }, + { + "bbox": [ + 497, + 320, + 504, + 327 + ], + "score": 0.72, + "content": "\\mathbf { \\nabla } \\cdot n", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 328, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 506, + 340 + ], + "score": 0.996, + "content": "matrix that carries the complete information to generate a robust counterfactual explanation as a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 220, + 351 + ], + "score": 0.982, + "content": "subset of edges, denoted by", + "type": "text" + }, + { + "bbox": [ + 220, + 339, + 249, + 350 + ], + "score": 0.92, + "content": "S \\subseteq E", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 339, + 343, + 351 + ], + "score": 0.986, + "content": ". Concretely,we obtain", + "type": "text" + }, + { + "bbox": [ + 344, + 339, + 352, + 348 + ], + "score": 0.83, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 339, + 466, + 351 + ], + "score": 0.994, + "content": "by selecting all the edges in", + "type": "text" + }, + { + "bbox": [ + 467, + 339, + 476, + 348 + ], + "score": 0.84, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 339, + 505, + 351 + ], + "score": 1.0, + "content": "whose", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 349, + 297, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 205, + 361 + ], + "score": 0.998, + "content": "corresponding entries in", + "type": "text" + }, + { + "bbox": [ + 205, + 349, + 218, + 359 + ], + "score": 0.44, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 349, + 297, + 361 + ], + "score": 0.976, + "content": "are larger than O.5.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 317, + 506, + 361 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 371, + 216, + 383 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 216, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 205, + 386 + ], + "score": 0.998, + "content": "4.3.2Training Model", + "type": "text" + }, + { + "bbox": [ + 206, + 372, + 216, + 383 + ], + "score": 0.82, + "content": "f _ { \\theta }", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 186, + 403 + ], + "score": 0.999, + "content": "For an input graph", + "type": "text" + }, + { + "bbox": [ + 186, + 390, + 235, + 402 + ], + "score": 0.92, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 389, + 283, + 403 + ], + "score": 0.977, + "content": ", denote by", + "type": "text" + }, + { + "bbox": [ + 283, + 390, + 313, + 401 + ], + "score": 0.9, + "content": "S \\subseteq E", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 389, + 450, + 403 + ], + "score": 0.977, + "content": " the subset of edges produced by", + "type": "text" + }, + { + "bbox": [ + 450, + 390, + 460, + 401 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 389, + 505, + 403 + ], + "score": 0.999, + "content": "to explain", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 401, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 165, + 412 + ], + "score": 0.998, + "content": "the prediction", + "type": "text" + }, + { + "bbox": [ + 166, + 401, + 187, + 412 + ], + "score": 0.91, + "content": "\\phi ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 401, + 327, + 412 + ], + "score": 0.993, + "content": ",our goal is to train a good model", + "type": "text" + }, + { + "bbox": [ + 328, + 401, + 338, + 412 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 401, + 505, + 412 + ], + "score": 0.992, + "content": "such that the prediction on the subgraph", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 411, + 507, + 424 + ], + "spans": [ + { + "bbox": [ + 107, + 411, + 121, + 422 + ], + "score": 0.88, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 411, + 170, + 424 + ], + "score": 0.998, + "content": "induced by", + "type": "text" + }, + { + "bbox": [ + 171, + 412, + 178, + 421 + ], + "score": 0.83, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 411, + 202, + 424 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 202, + 412, + 211, + 421 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 411, + 285, + 424 + ], + "score": 0.995, + "content": "is consistent with", + "type": "text" + }, + { + "bbox": [ + 286, + 411, + 308, + 423 + ], + "score": 0.92, + "content": "\\phi ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 411, + 416, + 424 + ], + "score": 0.974, + "content": "; and deleting the edges in", + "type": "text" + }, + { + "bbox": [ + 417, + 411, + 425, + 421 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 411, + 448, + 424 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 449, + 411, + 457, + 421 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 411, + 507, + 424 + ], + "score": 0.986, + "content": "produces a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 421, + 484, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 188, + 434 + ], + "score": 0.985, + "content": "remainder subgraph", + "type": "text" + }, + { + "bbox": [ + 189, + 422, + 213, + 434 + ], + "score": 0.92, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 421, + 323, + 434 + ], + "score": 0.971, + "content": "such that the prediction on", + "type": "text" + }, + { + "bbox": [ + 323, + 422, + 348, + 434 + ], + "score": 0.94, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 421, + 457, + 434 + ], + "score": 0.998, + "content": "changes significantly from", + "type": "text" + }, + { + "bbox": [ + 458, + 421, + 480, + 433 + ], + "score": 0.92, + "content": "{ \\bar { \\phi } } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 421, + 484, + 434 + ], + "score": 0.629, + "content": "：", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 389, + 507, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 438, + 505, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 174, + 451 + ], + "score": 0.989, + "content": "Since producing", + "type": "text" + }, + { + "bbox": [ + 174, + 439, + 182, + 448 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 438, + 196, + 451 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 196, + 438, + 207, + 450 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 438, + 506, + 451 + ], + "score": 0.989, + "content": " is a discrete operation that is hard to incorporate in an end-to-end training", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 315, + 461 + ], + "score": 0.98, + "content": " process, we define two proxy graphs to approximate", + "type": "text" + }, + { + "bbox": [ + 315, + 449, + 329, + 460 + ], + "score": 0.89, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 448, + 347, + 461 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 348, + 449, + 372, + 461 + ], + "score": 0.91, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 448, + 505, + 461 + ], + "score": 0.991, + "content": ",respectively,such that the proxy", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 210, + 473 + ], + "score": 0.983, + "content": " graphs are determined by", + "type": "text" + }, + { + "bbox": [ + 210, + 461, + 216, + 470 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 460, + 506, + 473 + ], + "score": 0.999, + "content": "through continuous functions that can be smoothly incorporated into an", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 471, + 220, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 220, + 483 + ], + "score": 0.995, + "content": "end-to-end training process.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 438, + 506, + 483 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 505, + 519 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 188, + 499 + ], + "score": 0.988, + "content": "The proxy graph of", + "type": "text" + }, + { + "bbox": [ + 188, + 487, + 203, + 497 + ], + "score": 0.87, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 485, + 254, + 499 + ], + "score": 0.975, + "content": ", denoted by", + "type": "text" + }, + { + "bbox": [ + 255, + 487, + 268, + 497 + ], + "score": 0.87, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 485, + 368, + 499 + ], + "score": 0.98, + "content": ", is defined by regarding", + "type": "text" + }, + { + "bbox": [ + 369, + 487, + 381, + 497 + ], + "score": 0.39, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 485, + 425, + 499 + ], + "score": 0.997, + "content": "instead of", + "type": "text" + }, + { + "bbox": [ + 425, + 487, + 435, + 497 + ], + "score": 0.44, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 485, + 506, + 499 + ], + "score": 0.982, + "content": " as the adjacency", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 495, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 104, + 495, + 169, + 510 + ], + "score": 0.989, + "content": "matrix. That is,", + "type": "text" + }, + { + "bbox": [ + 170, + 497, + 182, + 508 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 495, + 339, + 510 + ], + "score": 0.991, + "content": " has exactly the same graph structure as", + "type": "text" + }, + { + "bbox": [ + 339, + 497, + 348, + 507 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 495, + 446, + 510 + ], + "score": 0.99, + "content": ", but the edge weights of", + "type": "text" + }, + { + "bbox": [ + 446, + 497, + 459, + 508 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 495, + 506, + 510 + ], + "score": 0.984, + "content": "is given by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 507, + 432, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 160, + 519 + ], + "score": 0.995, + "content": "the entries in", + "type": "text" + }, + { + "bbox": [ + 160, + 507, + 173, + 517 + ], + "score": 0.32, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 507, + 215, + 519 + ], + "score": 0.998, + "content": "instead of", + "type": "text" + }, + { + "bbox": [ + 216, + 507, + 225, + 517 + ], + "score": 0.47, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 507, + 307, + 519 + ], + "score": 0.973, + "content": ".Here,the subscript", + "type": "text" + }, + { + "bbox": [ + 307, + 507, + 313, + 517 + ], + "score": 0.79, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 507, + 342, + 519 + ], + "score": 1.0, + "content": "means", + "type": "text" + }, + { + "bbox": [ + 343, + 507, + 356, + 518 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 507, + 426, + 519 + ], + "score": 0.996, + "content": "is determined by", + "type": "text" + }, + { + "bbox": [ + 426, + 508, + 432, + 517 + ], + "score": 0.81, + "content": "\\theta", + "type": "inline_equation" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 104, + 485, + 506, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 523, + 503, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 504, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 189, + 537 + ], + "score": 0.958, + "content": "The proxy graph of", + "type": "text" + }, + { + "bbox": [ + 189, + 524, + 213, + 536 + ], + "score": 0.91, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 522, + 266, + 537 + ], + "score": 0.981, + "content": ", denoted by", + "type": "text" + }, + { + "bbox": [ + 266, + 523, + 279, + 535 + ], + "score": 0.92, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 522, + 439, + 537 + ], + "score": 0.986, + "content": ", also have the same graph structure as", + "type": "text" + }, + { + "bbox": [ + 439, + 524, + 448, + 533 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 522, + 504, + 537 + ], + "score": 0.972, + "content": ", but the edge", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 533, + 346, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 254, + 546 + ], + "score": 0.986, + "content": "weight between each pair of vertices", + "type": "text" + }, + { + "bbox": [ + 254, + 536, + 263, + 545 + ], + "score": 0.87, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 533, + 281, + 546 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 282, + 535, + 292, + 546 + ], + "score": 0.87, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 533, + 346, + 546 + ], + "score": 0.981, + "content": "is defined as", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 522, + 504, + 546 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 550, + 372, + 579 + ], + "lines": [ + { + "bbox": [ + 239, + 550, + 372, + 579 + ], + "spans": [ + { + "bbox": [ + 239, + 550, + 372, + 579 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathbf { M } _ { i j } ^ { \\prime } = \\left\\{ { 1 - \\mathbf { M } _ { i j } \\quad \\mathrm { ~ i f ~ } \\mathbf { A } _ { i j } = 1 } \\atop { 0 } \\right. } \\end{array}", + "type": "interline_equation", + "image_path": "afaae5efb14f27284c6a4b7446792ec5cf01118d4490a60597c99ebbd88ab0af.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 550, + 372, + 564.5 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 239, + 564.5, + 372, + 579.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 589, + 504, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 208, + 604 + ], + "score": 0.986, + "content": "The edge weights of both", + "type": "text" + }, + { + "bbox": [ + 209, + 590, + 222, + 600 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 588, + 240, + 604 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 590, + 253, + 601 + ], + "score": 0.9, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 588, + 328, + 604 + ], + "score": 0.958, + "content": " are determined by", + "type": "text" + }, + { + "bbox": [ + 329, + 590, + 335, + 599 + ], + "score": 0.78, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 588, + 506, + 604 + ], + "score": 0.986, + "content": "through continuous functions, thus we can", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 600, + 398, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 194, + 613 + ], + "score": 0.993, + "content": "smoothly incorporate", + "type": "text" + }, + { + "bbox": [ + 194, + 600, + 207, + 611 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 600, + 226, + 613 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 226, + 600, + 239, + 612 + ], + "score": 0.92, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 600, + 398, + 613 + ], + "score": 0.992, + "content": "into an end-to-end training framework.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 588, + 506, + 613 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 616, + 504, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 484, + 630 + ], + "score": 0.975, + "content": " As discussed later in this section, we use a regularization term to force the value of each entry in", + "type": "text" + }, + { + "bbox": [ + 485, + 616, + 504, + 628 + ], + "score": 0.89, + "content": "\\mathbf { M } _ { i j }", + "type": "inline_equation" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 624, + 488, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 253, + 640 + ], + "score": 0.979, + "content": "to be close to eitherOor 1,such that", + "type": "text" + }, + { + "bbox": [ + 254, + 627, + 267, + 637 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 624, + 285, + 640 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 285, + 626, + 298, + 639 + ], + "score": 0.91, + "content": "\\operatorname { \\mathrm { \\Sigma } } _ { G _ { \\theta } ^ { \\prime } } ^ { G \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 624, + 377, + 640 + ], + "score": 0.997, + "content": "better approximate", + "type": "text" + }, + { + "bbox": [ + 378, + 627, + 392, + 637 + ], + "score": 0.89, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 624, + 410, + 640 + ], + "score": 0.999, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 411, + 627, + 435, + 639 + ], + "score": 0.93, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 624, + 488, + 640 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 613, + 504, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 642, + 243, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 243, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 243, + 655 + ], + "score": 0.987, + "content": "We formulate our loss function as", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 641, + 243, + 655 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 657, + 484, + 684 + ], + "lines": [ + { + "bbox": [ + 113, + 657, + 484, + 684 + ], + "spans": [ + { + "bbox": [ + 113, + 657, + 484, + 684 + ], + "score": 0.91, + "content": "\\mathcal { L } ( \\theta ) = \\sum _ { G \\in D } \\left\\{ \\lambda \\mathcal { L } _ { s a m e } ( \\theta , G ) + ( 1 - \\lambda ) \\mathcal { L } _ { o p p } ( \\theta , G ) + \\beta \\mathcal { R } _ { s p a r s e } ( \\theta , G ) + \\mu \\mathcal { R } _ { d i s c r e t e } ( \\theta , G ) \\right\\} ,", + "type": "interline_equation", + "image_path": "cfc41b2d76850caa450e5da3d6381caaa5c3ed3056969b987fa0fd0bf367d283.jpg" + } + ] + } + ], + "index": 44, + "virtual_lines": [ + { + "bbox": [ + 113, + 657, + 484, + 666.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 113, + 666.0, + 484, + 675.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 113, + 675.0, + 484, + 684.0 + ], + "spans": [], + "index": 45 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 690, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 690, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 133, + 702 + ], + "score": 0.923, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 690, + 172, + 702 + ], + "score": 0.86, + "content": "\\lambda \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 690, + 202, + 702 + ], + "score": 0.82, + "content": "\\beta \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 690, + 219, + 702 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 220, + 691, + 245, + 702 + ], + "score": 0.9, + "content": "\\mu \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 690, + 506, + 702 + ], + "score": 0.991, + "content": "are the hyperparameters controlling the importance of each term.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 351, + 713 + ], + "score": 0.975, + "content": "The influence of these parameters is discussed in Appendix", + "type": "text" + }, + { + "bbox": [ + 352, + 703, + 360, + 711 + ], + "score": 0.27, + "content": "\\mathbf { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 700, + 506, + 713 + ], + "score": 0.994, + "content": ".The first term of our loss function", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 276, + 723 + ], + "score": 0.99, + "content": "requires that the prediction of the GNN on", + "type": "text" + }, + { + "bbox": [ + 276, + 712, + 289, + 722 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 710, + 430, + 723 + ], + "score": 0.996, + "content": "is consistent with the prediction on", + "type": "text" + }, + { + "bbox": [ + 430, + 712, + 439, + 721 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 710, + 506, + 723 + ], + "score": 0.972, + "content": ". Intuitively, this", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 277, + 85 + ], + "score": 0.985, + "content": "means that the edges with larger weights in", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 277, + 73, + 290, + 83 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 291, + 72, + 399, + 85 + ], + "score": 0.997, + "content": "dominate the prediction on", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 399, + 73, + 408, + 82 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 409, + 72, + 505, + 85 + ], + "score": 0.994, + "content": ".We formulate this term", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 393, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 158, + 95 + ], + "score": 0.986, + "content": "by requiring", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 158, + 83, + 171, + 94 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 172, + 83, + 379, + 95 + ], + "score": 0.991, + "content": "to be covered by the same decision region covering", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 380, + 83, + 388, + 93 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 389, + 83, + 393, + 95 + ], + "score": 0.624, + "content": "：", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 690, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 277, + 85 + ], + "score": 0.985, + "content": "means that the edges with larger weights in", + "type": "text" + }, + { + "bbox": [ + 277, + 73, + 290, + 83 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 72, + 399, + 85 + ], + "score": 0.997, + "content": "dominate the prediction on", + "type": "text" + }, + { + "bbox": [ + 399, + 73, + 408, + 82 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 72, + 505, + 85 + ], + "score": 0.994, + "content": ".We formulate this term", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 393, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 158, + 95 + ], + "score": 0.986, + "content": "by requiring", + "type": "text" + }, + { + "bbox": [ + 158, + 83, + 171, + 94 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 83, + 379, + 95 + ], + "score": 0.991, + "content": "to be covered by the same decision region covering", + "type": "text" + }, + { + "bbox": [ + 380, + 83, + 388, + 93 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 83, + 393, + 95 + ], + "score": 0.624, + "content": "：", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 98, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 98, + 506, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 150, + 111 + ], + "score": 0.996, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 150, + 99, + 166, + 110 + ], + "score": 0.89, + "content": "\\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 98, + 393, + 111 + ], + "score": 0.975, + "content": " the set of LDBs that induce the decision region covering", + "type": "text" + }, + { + "bbox": [ + 393, + 99, + 402, + 109 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 98, + 435, + 111 + ], + "score": 0.963, + "content": ",and by", + "type": "text" + }, + { + "bbox": [ + 435, + 99, + 457, + 111 + ], + "score": 0.92, + "content": "| \\mathcal { H } _ { G } |", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 98, + 506, + 111 + ], + "score": 1.0, + "content": "the number", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 153, + 123 + ], + "score": 0.971, + "content": "of LDBs in", + "type": "text" + }, + { + "bbox": [ + 153, + 111, + 169, + 122 + ], + "score": 0.89, + "content": "\\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 110, + 204, + 123 + ], + "score": 0.992, + "content": ". For the", + "type": "text" + }, + { + "bbox": [ + 204, + 111, + 209, + 120 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 110, + 243, + 123 + ], + "score": 0.965, + "content": "-th LDB", + "type": "text" + }, + { + "bbox": [ + 243, + 110, + 281, + 122 + ], + "score": 0.91, + "content": "h _ { i } \\in \\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 110, + 326, + 123 + ], + "score": 0.971, + "content": ", denote by", + "type": "text" + }, + { + "bbox": [ + 326, + 110, + 405, + 123 + ], + "score": 0.92, + "content": "B _ { i } ( \\mathbf { x } ) = \\mathbf { w } _ { i } ^ { \\top } \\mathbf { x } + b _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 110, + 435, + 123 + ], + "score": 0.898, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 435, + 112, + 448, + 122 + ], + "score": 0.87, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 110, + 466, + 123 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 466, + 111, + 474, + 122 + ], + "score": 0.87, + "content": "b _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 110, + 505, + 123 + ], + "score": 0.999, + "content": "are the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 121, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 104, + 121, + 174, + 136 + ], + "score": 0.971, + "content": " basis and bias of", + "type": "text" + }, + { + "bbox": [ + 174, + 123, + 184, + 134 + ], + "score": 0.87, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 121, + 256, + 136 + ], + "score": 0.964, + "content": ", respectively, and", + "type": "text" + }, + { + "bbox": [ + 256, + 122, + 288, + 133 + ], + "score": 0.9, + "content": "\\mathbf { x } \\in \\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 121, + 376, + 136 + ], + "score": 0.981, + "content": "is a point in the space", + "type": "text" + }, + { + "bbox": [ + 377, + 123, + 390, + 133 + ], + "score": 0.87, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 121, + 442, + 136 + ], + "score": 0.985, + "content": ". The sign of", + "type": "text" + }, + { + "bbox": [ + 442, + 123, + 466, + 135 + ], + "score": 0.91, + "content": "B _ { i } ( { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 121, + 506, + 136 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 169, + 146 + ], + "score": 0.999, + "content": "whether a point", + "type": "text" + }, + { + "bbox": [ + 169, + 135, + 177, + 143 + ], + "score": 0.5, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 133, + 361, + 146 + ], + "score": 0.99, + "content": "lies on the positive side or the negative side of", + "type": "text" + }, + { + "bbox": [ + 361, + 134, + 371, + 144 + ], + "score": 0.87, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 133, + 465, + 146 + ], + "score": 0.991, + "content": ", and the absolute value", + "type": "text" + }, + { + "bbox": [ + 465, + 134, + 495, + 145 + ], + "score": 0.92, + "content": "| B _ { i } ( { \\bf x } ) |", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 133, + 505, + 146 + ], + "score": 0.999, + "content": "is", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 259, + 157 + ], + "score": 0.978, + "content": " proportional to the distance of a point", + "type": "text" + }, + { + "bbox": [ + 259, + 145, + 267, + 154 + ], + "score": 0.69, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 143, + 289, + 157 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 290, + 144, + 299, + 155 + ], + "score": 0.87, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 143, + 348, + 157 + ], + "score": 0.953, + "content": ". Denote by", + "type": "text" + }, + { + "bbox": [ + 348, + 144, + 366, + 156 + ], + "score": 0.93, + "content": "\\sigma ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 143, + 506, + 157 + ], + "score": 0.999, + "content": "the standard sigmoid function, we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 154, + 292, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 292, + 166 + ], + "score": 0.991, + "content": "formulate the first term of our loss function as", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 168, + 433, + 199 + ], + "lines": [ + { + "bbox": [ + 177, + 168, + 433, + 199 + ], + "spans": [ + { + "bbox": [ + 177, + 168, + 433, + 199 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { s a m e } ( \\theta , G ) = \\frac { 1 } { | \\mathcal { H } _ { G } | } \\sum _ { h _ { i } \\in \\mathcal { H } _ { G } } \\sigma \\left( - \\mathcal { B } _ { i } ( \\phi _ { g c } ( G ) ) * \\mathcal { B } _ { i } ( \\phi _ { g c } ( G _ { \\theta } ) ) \\right) ,", + "type": "interline_equation", + "image_path": "5f3b7a1f439726d3306abd8a0a28ca52ec5b541263656487dafdb921adbdf0a2.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 177, + 168, + 433, + 178.33333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 177, + 178.33333333333334, + 433, + 188.66666666666669 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 177, + 188.66666666666669, + 433, + 199.00000000000003 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 193, + 218 + ], + "score": 0.982, + "content": " such that minimizing", + "type": "text" + }, + { + "bbox": [ + 193, + 205, + 245, + 216 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { s a m e } ( \\theta , G )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 204, + 384, + 218 + ], + "score": 0.999, + "content": "encourages the graph embeddings", + "type": "text" + }, + { + "bbox": [ + 384, + 205, + 415, + 217 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 204, + 433, + 218 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 434, + 205, + 469, + 217 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G _ { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 204, + 505, + 218 + ], + "score": 0.999, + "content": "to lie on", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 235, + 227 + ], + "score": 0.988, + "content": "the same side of every LDB in", + "type": "text" + }, + { + "bbox": [ + 236, + 216, + 252, + 226 + ], + "score": 0.89, + "content": "\\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 214, + 283, + 227 + ], + "score": 0.965, + "content": ". Thus,", + "type": "text" + }, + { + "bbox": [ + 284, + 216, + 297, + 226 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 214, + 505, + 227 + ], + "score": 0.969, + "content": "is encouraged to be covered by the same decision", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 226, + 184, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 171, + 238 + ], + "score": 0.993, + "content": "region covering", + "type": "text" + }, + { + "bbox": [ + 172, + 226, + 180, + 235 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 226, + 184, + 238 + ], + "score": 0.743, + "content": "：", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 241, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 0.985, + "content": "The second term of our loss function optimizes the counterfactual property of the explanations by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 250, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 219, + 265 + ], + "score": 0.998, + "content": "requiring the prediction on", + "type": "text" + }, + { + "bbox": [ + 220, + 252, + 233, + 264 + ], + "score": 0.9, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 250, + 445, + 265 + ], + "score": 0.993, + "content": "to be significantly different from the prediction on", + "type": "text" + }, + { + "bbox": [ + 445, + 252, + 454, + 262 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 250, + 506, + 265 + ], + "score": 0.963, + "content": ". Intuitively,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 262, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 335, + 275 + ], + "score": 0.993, + "content": "this means that the set of edges with larger weights in", + "type": "text" + }, + { + "bbox": [ + 335, + 263, + 348, + 273 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 262, + 506, + 275 + ], + "score": 0.981, + "content": " are good counterfactual explanations", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 285 + ], + "score": 0.993, + "content": "because reducing the weights of these edges significantly changes the prediction. Following the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 284, + 303, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 303, + 295 + ], + "score": 0.996, + "content": "above intuition,we formulate the second term as", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 197, + 299, + 413, + 318 + ], + "lines": [ + { + "bbox": [ + 197, + 299, + 413, + 318 + ], + "spans": [ + { + "bbox": [ + 197, + 299, + 413, + 318 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { o p p } ( \\theta , G ) = \\operatorname* { m i n } _ { h _ { i } \\in \\mathcal { H } _ { G } } \\sigma \\left( \\mathcal { B } _ { i } ( \\phi _ { g c } ( G ) ) * \\mathcal { B } _ { i } ( \\phi _ { g c } ( G _ { \\theta } ^ { \\prime } ) ) \\right) ,", + "type": "interline_equation", + "image_path": "4a6e9260c897b62607fd0281735b7c87296a87e988d8b6aacd13c47e8cb7d8c9.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 197, + 299, + 413, + 318 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 195, + 338 + ], + "score": 0.983, + "content": " such that minimizing", + "type": "text" + }, + { + "bbox": [ + 195, + 324, + 240, + 337 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { o p p } ( \\theta , G )", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 323, + 383, + 338 + ], + "score": 0.985, + "content": " encourages the graph embeddings", + "type": "text" + }, + { + "bbox": [ + 383, + 325, + 414, + 336 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 323, + 432, + 338 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 433, + 324, + 468, + 337 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G _ { \\theta } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 323, + 506, + 338 + ], + "score": 0.989, + "content": "to lie on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 334, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 274, + 349 + ], + "score": 0.975, + "content": " the opposite sides of at least one LDB in", + "type": "text" + }, + { + "bbox": [ + 275, + 335, + 291, + 346 + ], + "score": 0.89, + "content": "\\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 334, + 393, + 349 + ], + "score": 0.974, + "content": ". This further means that", + "type": "text" + }, + { + "bbox": [ + 393, + 336, + 406, + 347 + ], + "score": 0.86, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 334, + 506, + 349 + ], + "score": 0.997, + "content": "is encouraged not to be", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 270, + 358 + ], + "score": 0.996, + "content": "covered by the decision region covering", + "type": "text" + }, + { + "bbox": [ + 271, + 346, + 279, + 355 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 345, + 373, + 358 + ], + "score": 0.984, + "content": ", thus the prediction on", + "type": "text" + }, + { + "bbox": [ + 374, + 345, + 387, + 358 + ], + "score": 0.89, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 345, + 505, + 358 + ], + "score": 0.994, + "content": "can be changed significantly", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 356, + 211, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 199, + 368 + ], + "score": 0.985, + "content": "from the prediction on", + "type": "text" + }, + { + "bbox": [ + 199, + 358, + 207, + 366 + ], + "score": 0.85, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 356, + 211, + 368 + ], + "score": 0.644, + "content": "：", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 108, + 371, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 504, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 276, + 385 + ], + "score": 0.984, + "content": "Similar to [45], we use a L1 regularization", + "type": "text" + }, + { + "bbox": [ + 276, + 371, + 373, + 384 + ], + "score": 0.94, + "content": "\\mathcal { R } _ { s p a r s e } ( \\theta , G ) = \\| \\mathbf { M } \\| _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 370, + 428, + 385 + ], + "score": 0.997, + "content": "on the matrix", + "type": "text" + }, + { + "bbox": [ + 429, + 372, + 441, + 382 + ], + "score": 0.32, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 370, + 493, + 385 + ], + "score": 1.0, + "content": "produced by", + "type": "text" + }, + { + "bbox": [ + 494, + 372, + 504, + 383 + ], + "score": 0.85, + "content": "f _ { \\theta }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 179, + 394 + ], + "score": 0.996, + "content": "on an input graph", + "type": "text" + }, + { + "bbox": [ + 179, + 384, + 187, + 392 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 382, + 480, + 394 + ], + "score": 0.976, + "content": "to produce a sparse matrix M, such that only a small number of edges in", + "type": "text" + }, + { + "bbox": [ + 480, + 383, + 489, + 392 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 382, + 505, + 394 + ], + "score": 0.999, + "content": "are", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 488, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 488, + 406 + ], + "score": 0.986, + "content": " selected as the counterfactual explanation. We also follow [45] to use an entropy regularization", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 408, + 455, + 439 + ], + "lines": [ + { + "bbox": [ + 156, + 408, + 455, + 439 + ], + "spans": [ + { + "bbox": [ + 156, + 408, + 455, + 439 + ], + "score": 0.94, + "content": "\\mathcal { R } _ { d i s c r e t e } ( \\theta , G ) = - \\frac { 1 } { | \\mathbf { M } | } \\sum _ { i , j } ( \\mathbf { M } _ { i j } \\log ( \\mathbf { M } _ { i j } ) + ( 1 - \\mathbf { M } _ { i j } ) \\log ( 1 - \\mathbf { M } _ { i j } ) )", + "type": "interline_equation", + "image_path": "92c85db5def2d86edeb67eb745a7f02fb8e21bbf2fd6168f0abe3416cc9d7222.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 156, + 408, + 455, + 418.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 156, + 418.3333333333333, + 455, + 428.66666666666663 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 156, + 428.66666666666663, + 455, + 438.99999999999994 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 445, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 240, + 459 + ], + "score": 0.972, + "content": " to push the value of each entry in", + "type": "text" + }, + { + "bbox": [ + 240, + 446, + 259, + 458 + ], + "score": 0.93, + "content": "\\mathbf { M } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 444, + 407, + 459 + ], + "score": 0.988, + "content": "to be close to either Oor 1,such that", + "type": "text" + }, + { + "bbox": [ + 407, + 446, + 420, + 456 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 444, + 438, + 459 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 439, + 446, + 451, + 458 + ], + "score": 0.9, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "approximate", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 456, + 240, + 469 + ], + "spans": [ + { + "bbox": [ + 107, + 458, + 121, + 467 + ], + "score": 0.86, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 456, + 140, + 469 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 140, + 456, + 164, + 469 + ], + "score": 0.91, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 456, + 240, + 469 + ], + "score": 0.995, + "content": "well, respectively.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 227, + 485 + ], + "score": 0.999, + "content": "Now we can use the graphs in", + "type": "text" + }, + { + "bbox": [ + 228, + 473, + 237, + 483 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 473, + 483, + 485 + ], + "score": 0.996, + "content": "and the extracted decision regions to train the neural network", + "type": "text" + }, + { + "bbox": [ + 483, + 473, + 493, + 484 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 473, + 505, + 485 + ], + "score": 0.997, + "content": "in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 482, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 259, + 497 + ], + "score": 0.982, + "content": " an end-to-end manner by minimizing", + "type": "text" + }, + { + "bbox": [ + 259, + 483, + 280, + 495 + ], + "score": 0.92, + "content": "\\mathcal { L } ( \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 482, + 301, + 497 + ], + "score": 0.996, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 301, + 484, + 307, + 493 + ], + "score": 0.81, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 482, + 506, + 497 + ], + "score": 0.975, + "content": "using back propagation. Once we finish training", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 107, + 495, + 117, + 505 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 493, + 192, + 506 + ], + "score": 0.979, + "content": ", we can first apply", + "type": "text" + }, + { + "bbox": [ + 192, + 494, + 202, + 505 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 493, + 290, + 506 + ], + "score": 0.999, + "content": "to produce the matrix", + "type": "text" + }, + { + "bbox": [ + 290, + 494, + 302, + 504 + ], + "score": 0.42, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 493, + 376, + 506 + ], + "score": 0.979, + "content": "for an input graph", + "type": "text" + }, + { + "bbox": [ + 376, + 495, + 424, + 505 + ], + "score": 0.91, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 493, + 506, + 506 + ], + "score": 0.984, + "content": ",and then obtain the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 504, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 156, + 516 + ], + "score": 1.0, + "content": "explanation", + "type": "text" + }, + { + "bbox": [ + 156, + 506, + 163, + 514 + ], + "score": 0.86, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 504, + 278, + 516 + ], + "score": 0.998, + "content": "by selecting all the edges in", + "type": "text" + }, + { + "bbox": [ + 278, + 505, + 287, + 514 + ], + "score": 0.85, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 504, + 416, + 516 + ], + "score": 0.996, + "content": "whose corresponding entries in", + "type": "text" + }, + { + "bbox": [ + 416, + 505, + 428, + 514 + ], + "score": 0.35, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 504, + 506, + 516 + ], + "score": 0.969, + "content": "are larger than 0.5.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 513, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 104, + 513, + 506, + 528 + ], + "score": 0.989, + "content": "We do not need the extracted boundaries for inference as the the decision logic of GNN is already", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 523, + 345, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 258, + 539 + ], + "score": 0.985, + "content": "distilled into the explanation network", + "type": "text" + }, + { + "bbox": [ + 258, + 525, + 264, + 536 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 523, + 345, + 539 + ], + "score": 0.998, + "content": "during the training.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 540, + 504, + 562 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 504, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 504, + 552 + ], + "score": 0.989, + "content": "As discussed in Appendix B,our method can be easily extended to generate robust counterfactual", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 551, + 274, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 274, + 563 + ], + "score": 0.999, + "content": "explanations for node classification tasks.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 566, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 329, + 580 + ], + "score": 0.998, + "content": "Our method is highly efficient with a time complexity", + "type": "text" + }, + { + "bbox": [ + 330, + 567, + 359, + 578 + ], + "score": 0.92, + "content": "O ( | E | )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 565, + 506, + 580 + ], + "score": 0.98, + "content": " for explaining the prediction on an", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 155, + 590 + ], + "score": 0.971, + "content": " input graph", + "type": "text" + }, + { + "bbox": [ + 155, + 579, + 163, + 587 + ], + "score": 0.87, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 576, + 194, + 590 + ], + "score": 0.926, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 194, + 579, + 208, + 589 + ], + "score": 0.93, + "content": "| E |", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 576, + 332, + 590 + ], + "score": 0.964, + "content": " is the total number of edges in", + "type": "text" + }, + { + "bbox": [ + 333, + 578, + 341, + 587 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 576, + 477, + 590 + ], + "score": 0.979, + "content": ": Additionally, the neural network", + "type": "text" + }, + { + "bbox": [ + 477, + 578, + 487, + 588 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "can", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 0.99, + "content": "be directly used without retraining to predict explanations on unseen graphs. Thus our method is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 597, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 611 + ], + "score": 0.99, + "content": " significantly faster than the other methods [45,32,47,38] that require retraining each time when", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 609, + 293, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 293, + 621 + ], + "score": 0.982, + "content": " generating explanations on a new input graph.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + }, + { + "type": "title", + "bbox": [ + 107, + 634, + 191, + 648 + ], + "lines": [ + { + "bbox": [ + 103, + 632, + 193, + 651 + ], + "spans": [ + { + "bbox": [ + 103, + 632, + 193, + 651 + ], + "score": 1.0, + "content": "5Experiments", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 507, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 507, + 672 + ], + "score": 0.992, + "content": "We conduct series of experiments to compare our method with the state-of-the-art methods in-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "score": 0.988, + "content": "cluding GNNExplainer [45],PGExplainer [25],PGM-Explainer [38], SubgraphX [47] and CF-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 679, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 692 + ], + "score": 0.985, + "content": "GNNExplainer [24]. For the methods that identify a set of vertices as an explanation, we use the set", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "score": 0.996, + "content": "of vertices to induce a subgraph from the input graph,and then use the set of edges of the induced", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 505, + 713 + ], + "score": 0.975, + "content": " subgraph as the explanation. For the methods that identify a subgraph as an explanation, we directly", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 712, + 367, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 367, + 723 + ], + "score": 0.985, + "content": "use the set of edges of the identified subgraph as the explanation.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 48.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 753 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 504, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 72, + 505, + 95 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 98, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 98, + 506, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 150, + 111 + ], + "score": 0.996, + "content": "Denote by", + "type": "text" + }, + { + "bbox": [ + 150, + 99, + 166, + 110 + ], + "score": 0.89, + "content": "\\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 98, + 393, + 111 + ], + "score": 0.975, + "content": " the set of LDBs that induce the decision region covering", + "type": "text" + }, + { + "bbox": [ + 393, + 99, + 402, + 109 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 98, + 435, + 111 + ], + "score": 0.963, + "content": ",and by", + "type": "text" + }, + { + "bbox": [ + 435, + 99, + 457, + 111 + ], + "score": 0.92, + "content": "| \\mathcal { H } _ { G } |", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 98, + 506, + 111 + ], + "score": 1.0, + "content": "the number", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 153, + 123 + ], + "score": 0.971, + "content": "of LDBs in", + "type": "text" + }, + { + "bbox": [ + 153, + 111, + 169, + 122 + ], + "score": 0.89, + "content": "\\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 110, + 204, + 123 + ], + "score": 0.992, + "content": ". For the", + "type": "text" + }, + { + "bbox": [ + 204, + 111, + 209, + 120 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 110, + 243, + 123 + ], + "score": 0.965, + "content": "-th LDB", + "type": "text" + }, + { + "bbox": [ + 243, + 110, + 281, + 122 + ], + "score": 0.91, + "content": "h _ { i } \\in \\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 110, + 326, + 123 + ], + "score": 0.971, + "content": ", denote by", + "type": "text" + }, + { + "bbox": [ + 326, + 110, + 405, + 123 + ], + "score": 0.92, + "content": "B _ { i } ( \\mathbf { x } ) = \\mathbf { w } _ { i } ^ { \\top } \\mathbf { x } + b _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 110, + 435, + 123 + ], + "score": 0.898, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 435, + 112, + 448, + 122 + ], + "score": 0.87, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 110, + 466, + 123 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 466, + 111, + 474, + 122 + ], + "score": 0.87, + "content": "b _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 110, + 505, + 123 + ], + "score": 0.999, + "content": "are the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 121, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 104, + 121, + 174, + 136 + ], + "score": 0.971, + "content": " basis and bias of", + "type": "text" + }, + { + "bbox": [ + 174, + 123, + 184, + 134 + ], + "score": 0.87, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 121, + 256, + 136 + ], + "score": 0.964, + "content": ", respectively, and", + "type": "text" + }, + { + "bbox": [ + 256, + 122, + 288, + 133 + ], + "score": 0.9, + "content": "\\mathbf { x } \\in \\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 121, + 376, + 136 + ], + "score": 0.981, + "content": "is a point in the space", + "type": "text" + }, + { + "bbox": [ + 377, + 123, + 390, + 133 + ], + "score": 0.87, + "content": "\\mathbb { O } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 121, + 442, + 136 + ], + "score": 0.985, + "content": ". The sign of", + "type": "text" + }, + { + "bbox": [ + 442, + 123, + 466, + 135 + ], + "score": 0.91, + "content": "B _ { i } ( { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 121, + 506, + 136 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 169, + 146 + ], + "score": 0.999, + "content": "whether a point", + "type": "text" + }, + { + "bbox": [ + 169, + 135, + 177, + 143 + ], + "score": 0.5, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 133, + 361, + 146 + ], + "score": 0.99, + "content": "lies on the positive side or the negative side of", + "type": "text" + }, + { + "bbox": [ + 361, + 134, + 371, + 144 + ], + "score": 0.87, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 133, + 465, + 146 + ], + "score": 0.991, + "content": ", and the absolute value", + "type": "text" + }, + { + "bbox": [ + 465, + 134, + 495, + 145 + ], + "score": 0.92, + "content": "| B _ { i } ( { \\bf x } ) |", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 133, + 505, + 146 + ], + "score": 0.999, + "content": "is", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 259, + 157 + ], + "score": 0.978, + "content": " proportional to the distance of a point", + "type": "text" + }, + { + "bbox": [ + 259, + 145, + 267, + 154 + ], + "score": 0.69, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 143, + 289, + 157 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 290, + 144, + 299, + 155 + ], + "score": 0.87, + "content": "h _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 143, + 348, + 157 + ], + "score": 0.953, + "content": ". Denote by", + "type": "text" + }, + { + "bbox": [ + 348, + 144, + 366, + 156 + ], + "score": 0.93, + "content": "\\sigma ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 143, + 506, + 157 + ], + "score": 0.999, + "content": "the standard sigmoid function, we", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 154, + 292, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 292, + 166 + ], + "score": 0.991, + "content": "formulate the first term of our loss function as", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5, + "bbox_fs": [ + 104, + 98, + 506, + 166 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 168, + 433, + 199 + ], + "lines": [ + { + "bbox": [ + 177, + 168, + 433, + 199 + ], + "spans": [ + { + "bbox": [ + 177, + 168, + 433, + 199 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { s a m e } ( \\theta , G ) = \\frac { 1 } { | \\mathcal { H } _ { G } | } \\sum _ { h _ { i } \\in \\mathcal { H } _ { G } } \\sigma \\left( - \\mathcal { B } _ { i } ( \\phi _ { g c } ( G ) ) * \\mathcal { B } _ { i } ( \\phi _ { g c } ( G _ { \\theta } ) ) \\right) ,", + "type": "interline_equation", + "image_path": "5f3b7a1f439726d3306abd8a0a28ca52ec5b541263656487dafdb921adbdf0a2.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 177, + 168, + 433, + 178.33333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 177, + 178.33333333333334, + 433, + 188.66666666666669 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 177, + 188.66666666666669, + 433, + 199.00000000000003 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 193, + 218 + ], + "score": 0.982, + "content": " such that minimizing", + "type": "text" + }, + { + "bbox": [ + 193, + 205, + 245, + 216 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { s a m e } ( \\theta , G )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 204, + 384, + 218 + ], + "score": 0.999, + "content": "encourages the graph embeddings", + "type": "text" + }, + { + "bbox": [ + 384, + 205, + 415, + 217 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 204, + 433, + 218 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 434, + 205, + 469, + 217 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G _ { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 204, + 505, + 218 + ], + "score": 0.999, + "content": "to lie on", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 235, + 227 + ], + "score": 0.988, + "content": "the same side of every LDB in", + "type": "text" + }, + { + "bbox": [ + 236, + 216, + 252, + 226 + ], + "score": 0.89, + "content": "\\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 214, + 283, + 227 + ], + "score": 0.965, + "content": ". Thus,", + "type": "text" + }, + { + "bbox": [ + 284, + 216, + 297, + 226 + ], + "score": 0.89, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 214, + 505, + 227 + ], + "score": 0.969, + "content": "is encouraged to be covered by the same decision", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 226, + 184, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 171, + 238 + ], + "score": 0.993, + "content": "region covering", + "type": "text" + }, + { + "bbox": [ + 172, + 226, + 180, + 235 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 226, + 184, + 238 + ], + "score": 0.743, + "content": "：", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 204, + 505, + 238 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 241, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 0.985, + "content": "The second term of our loss function optimizes the counterfactual property of the explanations by", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 250, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 219, + 265 + ], + "score": 0.998, + "content": "requiring the prediction on", + "type": "text" + }, + { + "bbox": [ + 220, + 252, + 233, + 264 + ], + "score": 0.9, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 250, + 445, + 265 + ], + "score": 0.993, + "content": "to be significantly different from the prediction on", + "type": "text" + }, + { + "bbox": [ + 445, + 252, + 454, + 262 + ], + "score": 0.8, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 250, + 506, + 265 + ], + "score": 0.963, + "content": ". Intuitively,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 262, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 335, + 275 + ], + "score": 0.993, + "content": "this means that the set of edges with larger weights in", + "type": "text" + }, + { + "bbox": [ + 335, + 263, + 348, + 273 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 262, + 506, + 275 + ], + "score": 0.981, + "content": " are good counterfactual explanations", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 285 + ], + "score": 0.993, + "content": "because reducing the weights of these edges significantly changes the prediction. Following the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 284, + 303, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 303, + 295 + ], + "score": 0.996, + "content": "above intuition,we formulate the second term as", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 240, + 506, + 295 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 197, + 299, + 413, + 318 + ], + "lines": [ + { + "bbox": [ + 197, + 299, + 413, + 318 + ], + "spans": [ + { + "bbox": [ + 197, + 299, + 413, + 318 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { o p p } ( \\theta , G ) = \\operatorname* { m i n } _ { h _ { i } \\in \\mathcal { H } _ { G } } \\sigma \\left( \\mathcal { B } _ { i } ( \\phi _ { g c } ( G ) ) * \\mathcal { B } _ { i } ( \\phi _ { g c } ( G _ { \\theta } ^ { \\prime } ) ) \\right) ,", + "type": "interline_equation", + "image_path": "4a6e9260c897b62607fd0281735b7c87296a87e988d8b6aacd13c47e8cb7d8c9.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 197, + 299, + 413, + 318 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 505, + 367 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 195, + 338 + ], + "score": 0.983, + "content": " such that minimizing", + "type": "text" + }, + { + "bbox": [ + 195, + 324, + 240, + 337 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { o p p } ( \\theta , G )", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 323, + 383, + 338 + ], + "score": 0.985, + "content": " encourages the graph embeddings", + "type": "text" + }, + { + "bbox": [ + 383, + 325, + 414, + 336 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G )", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 323, + 432, + 338 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 433, + 324, + 468, + 337 + ], + "score": 0.92, + "content": "\\phi _ { g c } ( G _ { \\theta } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 323, + 506, + 338 + ], + "score": 0.989, + "content": "to lie on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 334, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 274, + 349 + ], + "score": 0.975, + "content": " the opposite sides of at least one LDB in", + "type": "text" + }, + { + "bbox": [ + 275, + 335, + 291, + 346 + ], + "score": 0.89, + "content": "\\mathcal { H } _ { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 334, + 393, + 349 + ], + "score": 0.974, + "content": ". This further means that", + "type": "text" + }, + { + "bbox": [ + 393, + 336, + 406, + 347 + ], + "score": 0.86, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 334, + 506, + 349 + ], + "score": 0.997, + "content": "is encouraged not to be", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 270, + 358 + ], + "score": 0.996, + "content": "covered by the decision region covering", + "type": "text" + }, + { + "bbox": [ + 271, + 346, + 279, + 355 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 345, + 373, + 358 + ], + "score": 0.984, + "content": ", thus the prediction on", + "type": "text" + }, + { + "bbox": [ + 374, + 345, + 387, + 358 + ], + "score": 0.89, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 345, + 505, + 358 + ], + "score": 0.994, + "content": "can be changed significantly", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 356, + 211, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 199, + 368 + ], + "score": 0.985, + "content": "from the prediction on", + "type": "text" + }, + { + "bbox": [ + 199, + 358, + 207, + 366 + ], + "score": 0.85, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 356, + 211, + 368 + ], + "score": 0.644, + "content": "：", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 323, + 506, + 368 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 371, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 504, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 276, + 385 + ], + "score": 0.984, + "content": "Similar to [45], we use a L1 regularization", + "type": "text" + }, + { + "bbox": [ + 276, + 371, + 373, + 384 + ], + "score": 0.94, + "content": "\\mathcal { R } _ { s p a r s e } ( \\theta , G ) = \\| \\mathbf { M } \\| _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 370, + 428, + 385 + ], + "score": 0.997, + "content": "on the matrix", + "type": "text" + }, + { + "bbox": [ + 429, + 372, + 441, + 382 + ], + "score": 0.32, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 370, + 493, + 385 + ], + "score": 1.0, + "content": "produced by", + "type": "text" + }, + { + "bbox": [ + 494, + 372, + 504, + 383 + ], + "score": 0.85, + "content": "f _ { \\theta }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 179, + 394 + ], + "score": 0.996, + "content": "on an input graph", + "type": "text" + }, + { + "bbox": [ + 179, + 384, + 187, + 392 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 382, + 480, + 394 + ], + "score": 0.976, + "content": "to produce a sparse matrix M, such that only a small number of edges in", + "type": "text" + }, + { + "bbox": [ + 480, + 383, + 489, + 392 + ], + "score": 0.83, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 382, + 505, + 394 + ], + "score": 0.999, + "content": "are", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 488, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 488, + 406 + ], + "score": 0.986, + "content": " selected as the counterfactual explanation. We also follow [45] to use an entropy regularization", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 370, + 505, + 406 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 408, + 455, + 439 + ], + "lines": [ + { + "bbox": [ + 156, + 408, + 455, + 439 + ], + "spans": [ + { + "bbox": [ + 156, + 408, + 455, + 439 + ], + "score": 0.94, + "content": "\\mathcal { R } _ { d i s c r e t e } ( \\theta , G ) = - \\frac { 1 } { | \\mathbf { M } | } \\sum _ { i , j } ( \\mathbf { M } _ { i j } \\log ( \\mathbf { M } _ { i j } ) + ( 1 - \\mathbf { M } _ { i j } ) \\log ( 1 - \\mathbf { M } _ { i j } ) )", + "type": "interline_equation", + "image_path": "92c85db5def2d86edeb67eb745a7f02fb8e21bbf2fd6168f0abe3416cc9d7222.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 156, + 408, + 455, + 418.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 156, + 418.3333333333333, + 455, + 428.66666666666663 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 156, + 428.66666666666663, + 455, + 438.99999999999994 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 445, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 240, + 459 + ], + "score": 0.972, + "content": " to push the value of each entry in", + "type": "text" + }, + { + "bbox": [ + 240, + 446, + 259, + 458 + ], + "score": 0.93, + "content": "\\mathbf { M } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 444, + 407, + 459 + ], + "score": 0.988, + "content": "to be close to either Oor 1,such that", + "type": "text" + }, + { + "bbox": [ + 407, + 446, + 420, + 456 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 444, + 438, + 459 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 439, + 446, + 451, + 458 + ], + "score": 0.9, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "approximate", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 456, + 240, + 469 + ], + "spans": [ + { + "bbox": [ + 107, + 458, + 121, + 467 + ], + "score": 0.86, + "content": "G _ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 456, + 140, + 469 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 140, + 456, + 164, + 469 + ], + "score": 0.91, + "content": "G _ { E \\backslash S }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 456, + 240, + 469 + ], + "score": 0.995, + "content": "well, respectively.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 444, + 505, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 227, + 485 + ], + "score": 0.999, + "content": "Now we can use the graphs in", + "type": "text" + }, + { + "bbox": [ + 228, + 473, + 237, + 483 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 473, + 483, + 485 + ], + "score": 0.996, + "content": "and the extracted decision regions to train the neural network", + "type": "text" + }, + { + "bbox": [ + 483, + 473, + 493, + 484 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 473, + 505, + 485 + ], + "score": 0.997, + "content": "in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 482, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 259, + 497 + ], + "score": 0.982, + "content": " an end-to-end manner by minimizing", + "type": "text" + }, + { + "bbox": [ + 259, + 483, + 280, + 495 + ], + "score": 0.92, + "content": "\\mathcal { L } ( \\boldsymbol { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 482, + 301, + 497 + ], + "score": 0.996, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 301, + 484, + 307, + 493 + ], + "score": 0.81, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 482, + 506, + 497 + ], + "score": 0.975, + "content": "using back propagation. Once we finish training", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 107, + 495, + 117, + 505 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 493, + 192, + 506 + ], + "score": 0.979, + "content": ", we can first apply", + "type": "text" + }, + { + "bbox": [ + 192, + 494, + 202, + 505 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 493, + 290, + 506 + ], + "score": 0.999, + "content": "to produce the matrix", + "type": "text" + }, + { + "bbox": [ + 290, + 494, + 302, + 504 + ], + "score": 0.42, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 493, + 376, + 506 + ], + "score": 0.979, + "content": "for an input graph", + "type": "text" + }, + { + "bbox": [ + 376, + 495, + 424, + 505 + ], + "score": 0.91, + "content": "G = ( V , E )", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 493, + 506, + 506 + ], + "score": 0.984, + "content": ",and then obtain the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 504, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 156, + 516 + ], + "score": 1.0, + "content": "explanation", + "type": "text" + }, + { + "bbox": [ + 156, + 506, + 163, + 514 + ], + "score": 0.86, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 504, + 278, + 516 + ], + "score": 0.998, + "content": "by selecting all the edges in", + "type": "text" + }, + { + "bbox": [ + 278, + 505, + 287, + 514 + ], + "score": 0.85, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 504, + 416, + 516 + ], + "score": 0.996, + "content": "whose corresponding entries in", + "type": "text" + }, + { + "bbox": [ + 416, + 505, + 428, + 514 + ], + "score": 0.35, + "content": "\\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 504, + 506, + 516 + ], + "score": 0.969, + "content": "are larger than 0.5.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 513, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 104, + 513, + 506, + 528 + ], + "score": 0.989, + "content": "We do not need the extracted boundaries for inference as the the decision logic of GNN is already", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 523, + 345, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 258, + 539 + ], + "score": 0.985, + "content": "distilled into the explanation network", + "type": "text" + }, + { + "bbox": [ + 258, + 525, + 264, + 536 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 523, + 345, + 539 + ], + "score": 0.998, + "content": "during the training.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 473, + 506, + 539 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 540, + 504, + 562 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 504, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 504, + 552 + ], + "score": 0.989, + "content": "As discussed in Appendix B,our method can be easily extended to generate robust counterfactual", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 551, + 274, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 274, + 563 + ], + "score": 0.999, + "content": "explanations for node classification tasks.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 541, + 504, + 563 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 566, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 329, + 580 + ], + "score": 0.998, + "content": "Our method is highly efficient with a time complexity", + "type": "text" + }, + { + "bbox": [ + 330, + 567, + 359, + 578 + ], + "score": 0.92, + "content": "O ( | E | )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 565, + 506, + 580 + ], + "score": 0.98, + "content": " for explaining the prediction on an", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 155, + 590 + ], + "score": 0.971, + "content": " input graph", + "type": "text" + }, + { + "bbox": [ + 155, + 579, + 163, + 587 + ], + "score": 0.87, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 576, + 194, + 590 + ], + "score": 0.926, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 194, + 579, + 208, + 589 + ], + "score": 0.93, + "content": "| E |", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 576, + 332, + 590 + ], + "score": 0.964, + "content": " is the total number of edges in", + "type": "text" + }, + { + "bbox": [ + 333, + 578, + 341, + 587 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 576, + 477, + 590 + ], + "score": 0.979, + "content": ": Additionally, the neural network", + "type": "text" + }, + { + "bbox": [ + 477, + 578, + 487, + 588 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "can", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 0.99, + "content": "be directly used without retraining to predict explanations on unseen graphs. Thus our method is", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 597, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 611 + ], + "score": 0.99, + "content": " significantly faster than the other methods [45,32,47,38] that require retraining each time when", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 609, + 293, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 293, + 621 + ], + "score": 0.982, + "content": " generating explanations on a new input graph.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 565, + 506, + 621 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 634, + 191, + 648 + ], + "lines": [ + { + "bbox": [ + 103, + 632, + 193, + 651 + ], + "spans": [ + { + "bbox": [ + 103, + 632, + 193, + 651 + ], + "score": 1.0, + "content": "5Experiments", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 659, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 507, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 507, + 672 + ], + "score": 0.992, + "content": "We conduct series of experiments to compare our method with the state-of-the-art methods in-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 682 + ], + "score": 0.988, + "content": "cluding GNNExplainer [45],PGExplainer [25],PGM-Explainer [38], SubgraphX [47] and CF-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 679, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 692 + ], + "score": 0.985, + "content": "GNNExplainer [24]. For the methods that identify a set of vertices as an explanation, we use the set", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 690, + 506, + 703 + ], + "score": 0.996, + "content": "of vertices to induce a subgraph from the input graph,and then use the set of edges of the induced", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 505, + 713 + ], + "score": 0.975, + "content": " subgraph as the explanation. For the methods that identify a subgraph as an explanation, we directly", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 712, + 367, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 367, + 723 + ], + "score": 0.985, + "content": "use the set of edges of the identified subgraph as the explanation.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 658, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 70, + 444, + 155 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 166, + 70, + 444, + 155 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 166, + 70, + 444, + 155 + ], + "spans": [ + { + "bbox": [ + 166, + 70, + 444, + 155 + ], + "score": 0.965, + "type": "image", + "image_path": "b7fd56d0939a4af57dc11439ffc5d8bfba231262c91ca025fc019f5b6e33dc49.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 166, + 70, + 444, + 98.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 166, + 98.33333333333333, + 444, + 126.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 166, + 126.66666666666666, + 444, + 155.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 169, + 505, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 168, + 507, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 507, + 183 + ], + "score": 0.989, + "content": "Figure 1: Fidelity performance averaged across 10 runs for the datasets at different levels of sparsity.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 0.997, + "content": "To demonstrate the effectiveness of the decision regions,we derive another baseline method named", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "score": 0.984, + "content": "RCExp-NoLDB that adopts the general framework of RCExplainer but does not use the LDBs of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 0.997, + "content": "decision regions to generate explanations.Instead, RCExp-NoLDB directly maximizes the prediction", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 238, + 458, + 257 + ], + "spans": [ + { + "bbox": [ + 104, + 238, + 186, + 257 + ], + "score": 0.975, + "content": "confidence on class", + "type": "text" + }, + { + "bbox": [ + 186, + 243, + 192, + 251 + ], + "score": 0.71, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 238, + 207, + 257 + ], + "score": 0.999, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 208, + 242, + 221, + 252 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 238, + 419, + 257 + ], + "score": 0.981, + "content": "and minimizes the prediction confidence of class", + "type": "text" + }, + { + "bbox": [ + 419, + 243, + 425, + 251 + ], + "score": 0.69, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 238, + 440, + 257 + ], + "score": 0.999, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 440, + 241, + 453, + 253 + ], + "score": 0.9, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 238, + 458, + 257 + ], + "score": 0.813, + "content": "：", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 257, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 506, + 269 + ], + "score": 0.994, + "content": "We evaluate the explanation performance on two typical tasks: the graph clasification task that uses", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 268, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 280 + ], + "score": 0.991, + "content": "a GNN to predict the class of an input graph,and the node classification task that uses a GNN to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 277, + 507, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 507, + 291 + ], + "score": 0.991, + "content": " predict the class of a graph node. For the graph classification task, we use one synthetic dataset,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "score": 0.984, + "content": "BA-2motifs [25],and two real-world datasets,Mutagenicity [21] and NCI1 [39].For the node", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 0.991, + "content": "classification task, we use the same four synthetic datasets as used by GNNExplainer [45], namely,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 308, + 369, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 369, + 321 + ], + "score": 0.986, + "content": "BA-SHAPES,BA-COMMUNITY,TREE-CYCLES and TREE-GRID.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 0.99, + "content": "Limited by space, we only report here the key results_on the graph classification task for fidelity,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "score": 0.986, + "content": "robustness and efficiency. Please refer to Appendix E for details on datasets, baselines and the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 346, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 358 + ], + "score": 0.989, + "content": "experiment setups.Detailed experimental comparison on the node classification task willbe discussed", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 104, + 356, + 505, + 370 + ], + "score": 0.986, + "content": " in Appendix F where we show that our method produces extremely accurate explanations. The code3", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 367, + 191, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 191, + 381 + ], + "score": 0.998, + "content": "is publicly available.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 106, + 401, + 162, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 164, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 164, + 416 + ], + "score": 1.0, + "content": "5.1Fidelity", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 0.994, + "content": "Fidelity is measured by the decrease of prediction confidence after removing the explanation (i.e., a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 0.99, + "content": " set of edges) from the input graph [32]. We use fidelity to evaluate how counterfactual the generated", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 0.987, + "content": "explanations are on the datasets Mutagenicity,NCI1 and BA-2motifs.A large fidelity score indicates", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "score": 0.986, + "content": " stronger counterfactual characteristics. It is important to note that fidelity may be sensitive to sparsity", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 305, + 479 + ], + "score": 0.986, + "content": "of explanations. The sparsity of an explanation", + "type": "text" + }, + { + "bbox": [ + 306, + 467, + 313, + 477 + ], + "score": 0.79, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 465, + 442, + 479 + ], + "score": 0.989, + "content": "with respect to an input graph", + "type": "text" + }, + { + "bbox": [ + 443, + 466, + 494, + 478 + ], + "score": 0.93, + "content": "G = ( \\bar { V , } E )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 465, + 506, + 479 + ], + "score": 0.997, + "content": "is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 475, + 507, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 212, + 495 + ], + "score": 0.8, + "content": "\\begin{array} { r } { s p a r s i t y ( S , G ) = 1 - \\frac { | S | } { | E | } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 475, + 507, + 497 + ], + "score": 0.727, + "content": ",thatis,tpeeefdgafrta", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 491, + 392, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 128, + 506 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 128, + 493, + 137, + 502 + ], + "score": 0.73, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 491, + 392, + 506 + ], + "score": 0.978, + "content": ". We only compare explanations with the same level of sparsity.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 509, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 0.989, + "content": "Figure 1 shows the results about fidelity. Our approach achieves the best fidelity performance at all", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 0.98, + "content": " levels of sparsity. The results validate the effectiveness of our method in producing highly counter-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 528, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 506, + 543 + ], + "score": 0.987, + "content": "factual explanations. RCExplainer also significantly outperforms RCExp-NoLDB. This confirms", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 0.993, + "content": " that using LDBs of decision regions extracted from GNNs produces more faithful counterfactual", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 550, + 162, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 162, + 563 + ], + "score": 1.0, + "content": "explanations.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 108, + 566, + 504, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "score": 0.984, + "content": "CF-GNNExplainer performs the best among the rest of the methods.This is expected as it optimizes", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 0.995, + "content": "the counterfactual behavior of the explanations which results in higher fidelity for the explanations in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 587, + 468, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 468, + 599 + ], + "score": 0.979, + "content": "comparison to those produced by other methods such as GNNExplainer and PGExplainer.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 602, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 615 + ], + "score": 0.99, + "content": "The fidelity performance of SubgraphX reported in [47] was obtained by seting the features of nodes", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 614, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 626 + ], + "score": 0.987, + "content": "that are part of the explanation to Obut not removing the explanation edges from the input graph. This", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "score": 0.991, + "content": " does not remove the message passing roles of the explanation nodes from the input graph because the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 646 + ], + "score": 0.988, + "content": "edges connected to those nodes still can pass messages. In our experiments, we directly block the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 657 + ], + "score": 0.995, + "content": "messages that are passed on the edges in the explanation, which completely prevents the explanation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 0.99, + "content": " nodes in the input graph to participate in the message passing. As a result, the performance of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 665, + 233, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 233, + 678 + ], + "score": 0.998, + "content": "SubgraphX drops significantly.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 703, + 502, + 721 + ], + "lines": [ + { + "bbox": [ + 118, + 700, + 504, + 714 + ], + "spans": [ + { + "bbox": [ + 118, + 700, + 504, + 714 + ], + "score": 0.994, + "content": "3Code available at https: //marketplace.huaweicloud.com/markets/aihub/notebook/detail/", + "type": "text" + } + ] + }, + { + "bbox": [ + 107, + 712, + 299, + 721 + ], + "spans": [ + { + "bbox": [ + 107, + 712, + 299, + 721 + ], + "score": 0.999, + "content": "?id=e41f63d3-e346-4891-bf6a-40e64b4a3278", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 0.999, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 70, + 444, + 155 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 166, + 70, + 444, + 155 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 166, + 70, + 444, + 155 + ], + "spans": [ + { + "bbox": [ + 166, + 70, + 444, + 155 + ], + "score": 0.965, + "type": "image", + "image_path": "b7fd56d0939a4af57dc11439ffc5d8bfba231262c91ca025fc019f5b6e33dc49.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 166, + 70, + 444, + 98.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 166, + 98.33333333333333, + 444, + 126.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 166, + 126.66666666666666, + 444, + 155.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 169, + 505, + 181 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 168, + 507, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 507, + 183 + ], + "score": 0.989, + "content": "Figure 1: Fidelity performance averaged across 10 runs for the datasets at different levels of sparsity.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 0.997, + "content": "To demonstrate the effectiveness of the decision regions,we derive another baseline method named", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "score": 0.984, + "content": "RCExp-NoLDB that adopts the general framework of RCExplainer but does not use the LDBs of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 505, + 243 + ], + "score": 0.997, + "content": "decision regions to generate explanations.Instead, RCExp-NoLDB directly maximizes the prediction", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 238, + 458, + 257 + ], + "spans": [ + { + "bbox": [ + 104, + 238, + 186, + 257 + ], + "score": 0.975, + "content": "confidence on class", + "type": "text" + }, + { + "bbox": [ + 186, + 243, + 192, + 251 + ], + "score": 0.71, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 238, + 207, + 257 + ], + "score": 0.999, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 208, + 242, + 221, + 252 + ], + "score": 0.88, + "content": "G _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 238, + 419, + 257 + ], + "score": 0.981, + "content": "and minimizes the prediction confidence of class", + "type": "text" + }, + { + "bbox": [ + 419, + 243, + 425, + 251 + ], + "score": 0.69, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 238, + 440, + 257 + ], + "score": 0.999, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 440, + 241, + 453, + 253 + ], + "score": 0.9, + "content": "G _ { \\theta } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 238, + 458, + 257 + ], + "score": 0.813, + "content": "：", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 210, + 505, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 257, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 506, + 269 + ], + "score": 0.994, + "content": "We evaluate the explanation performance on two typical tasks: the graph clasification task that uses", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 268, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 280 + ], + "score": 0.991, + "content": "a GNN to predict the class of an input graph,and the node classification task that uses a GNN to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 277, + 507, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 507, + 291 + ], + "score": 0.991, + "content": " predict the class of a graph node. For the graph classification task, we use one synthetic dataset,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "score": 0.984, + "content": "BA-2motifs [25],and two real-world datasets,Mutagenicity [21] and NCI1 [39].For the node", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 0.991, + "content": "classification task, we use the same four synthetic datasets as used by GNNExplainer [45], namely,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 308, + 369, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 369, + 321 + ], + "score": 0.986, + "content": "BA-SHAPES,BA-COMMUNITY,TREE-CYCLES and TREE-GRID.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 257, + 507, + 321 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 0.99, + "content": "Limited by space, we only report here the key results_on the graph classification task for fidelity,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "score": 0.986, + "content": "robustness and efficiency. Please refer to Appendix E for details on datasets, baselines and the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 346, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 358 + ], + "score": 0.989, + "content": "experiment setups.Detailed experimental comparison on the node classification task willbe discussed", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 104, + 356, + 505, + 370 + ], + "score": 0.986, + "content": " in Appendix F where we show that our method produces extremely accurate explanations. The code3", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 367, + 191, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 191, + 381 + ], + "score": 0.998, + "content": "is publicly available.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 324, + 506, + 381 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 401, + 162, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 164, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 164, + 416 + ], + "score": 1.0, + "content": "5.1Fidelity", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 0.994, + "content": "Fidelity is measured by the decrease of prediction confidence after removing the explanation (i.e., a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 0.99, + "content": " set of edges) from the input graph [32]. We use fidelity to evaluate how counterfactual the generated", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 0.987, + "content": "explanations are on the datasets Mutagenicity,NCI1 and BA-2motifs.A large fidelity score indicates", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "score": 0.986, + "content": " stronger counterfactual characteristics. It is important to note that fidelity may be sensitive to sparsity", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 465, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 305, + 479 + ], + "score": 0.986, + "content": "of explanations. The sparsity of an explanation", + "type": "text" + }, + { + "bbox": [ + 306, + 467, + 313, + 477 + ], + "score": 0.79, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 465, + 442, + 479 + ], + "score": 0.989, + "content": "with respect to an input graph", + "type": "text" + }, + { + "bbox": [ + 443, + 466, + 494, + 478 + ], + "score": 0.93, + "content": "G = ( \\bar { V , } E )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 465, + 506, + 479 + ], + "score": 0.997, + "content": "is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 475, + 507, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 212, + 495 + ], + "score": 0.8, + "content": "\\begin{array} { r } { s p a r s i t y ( S , G ) = 1 - \\frac { | S | } { | E | } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 475, + 507, + 497 + ], + "score": 0.727, + "content": ",thatis,tpeeefdgafrta", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 491, + 392, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 128, + 506 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 128, + 493, + 137, + 502 + ], + "score": 0.73, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 491, + 392, + 506 + ], + "score": 0.978, + "content": ". We only compare explanations with the same level of sparsity.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 424, + 507, + 506 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 509, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 0.989, + "content": "Figure 1 shows the results about fidelity. Our approach achieves the best fidelity performance at all", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 0.98, + "content": " levels of sparsity. The results validate the effectiveness of our method in producing highly counter-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 528, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 506, + 543 + ], + "score": 0.987, + "content": "factual explanations. RCExplainer also significantly outperforms RCExp-NoLDB. This confirms", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 0.993, + "content": " that using LDBs of decision regions extracted from GNNs produces more faithful counterfactual", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 550, + 162, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 162, + 563 + ], + "score": 1.0, + "content": "explanations.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 104, + 508, + 506, + 563 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 566, + 504, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "score": 0.984, + "content": "CF-GNNExplainer performs the best among the rest of the methods.This is expected as it optimizes", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 0.995, + "content": "the counterfactual behavior of the explanations which results in higher fidelity for the explanations in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 587, + 468, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 468, + 599 + ], + "score": 0.979, + "content": "comparison to those produced by other methods such as GNNExplainer and PGExplainer.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 565, + 505, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 602, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 615 + ], + "score": 0.99, + "content": "The fidelity performance of SubgraphX reported in [47] was obtained by seting the features of nodes", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 614, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 505, + 626 + ], + "score": 0.987, + "content": "that are part of the explanation to Obut not removing the explanation edges from the input graph. This", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "score": 0.991, + "content": " does not remove the message passing roles of the explanation nodes from the input graph because the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 646 + ], + "score": 0.988, + "content": "edges connected to those nodes still can pass messages. In our experiments, we directly block the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 657 + ], + "score": 0.995, + "content": "messages that are passed on the edges in the explanation, which completely prevents the explanation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 0.99, + "content": " nodes in the input graph to participate in the message passing. As a result, the performance of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 665, + 233, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 233, + 678 + ], + "score": 0.998, + "content": "SubgraphX drops significantly.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 603, + 506, + 678 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 70, + 444, + 149 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 167, + 70, + 444, + 149 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 70, + 444, + 149 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 444, + 149 + ], + "score": 0.965, + "type": "image", + "image_path": "1aa166e7cf33380e4add83effc9273d8301d54224aa1859e6e8d4478bf17806a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 167, + 70, + 444, + 96.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 167, + 96.33333333333333, + 444, + 122.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 167, + 122.66666666666666, + 444, + 149.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 164, + 505, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 163, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 507, + 178 + ], + "score": 0.992, + "content": "Figure 2: Noise robustness (AUC) averaged across 10 runs for the datasets at diferent levels of noise.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 196, + 236, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 237, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 237, + 209 + ], + "score": 1.0, + "content": "5.2Robustness Performance", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 215, + 506, + 310 + ], + "lines": [ + { + "bbox": [ + 104, + 214, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 506, + 229 + ], + "score": 0.988, + "content": " In this experiment, we evaluate the robustness of all methods by quantifying how much an explanation", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 358, + 239 + ], + "score": 0.992, + "content": "changes after adding noise to the input graph. For an input graph", + "type": "text" + }, + { + "bbox": [ + 359, + 227, + 368, + 236 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 226, + 446, + 239 + ], + "score": 0.995, + "content": "and the explanation", + "type": "text" + }, + { + "bbox": [ + 447, + 227, + 454, + 236 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 226, + 505, + 239 + ], + "score": 0.959, + "content": ",we produce", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 235, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 104, + 235, + 179, + 251 + ], + "score": 0.948, + "content": " a perturbed graph", + "type": "text" + }, + { + "bbox": [ + 180, + 237, + 191, + 246 + ], + "score": 0.85, + "content": "G ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 235, + 506, + 251 + ], + "score": 0.988, + "content": "by adding random noise to the node features and randomly adding or deleting", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 337, + 259 + ], + "score": 0.986, + "content": " some edges of the input graph such that the prediction on", + "type": "text" + }, + { + "bbox": [ + 337, + 247, + 348, + 257 + ], + "score": 0.87, + "content": "G ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 246, + 491, + 259 + ], + "score": 0.997, + "content": "is consistent with the prediction on", + "type": "text" + }, + { + "bbox": [ + 492, + 248, + 501, + 257 + ], + "score": 0.78, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 246, + 506, + 259 + ], + "score": 0.794, + "content": "：", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 313, + 271 + ], + "score": 0.981, + "content": "Using the same method we obtain the explanation", + "type": "text" + }, + { + "bbox": [ + 314, + 258, + 324, + 267 + ], + "score": 0.86, + "content": "S ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 257, + 338, + 271 + ], + "score": 0.997, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 339, + 258, + 350, + 268 + ], + "score": 0.84, + "content": "G ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 257, + 424, + 271 + ], + "score": 0.975, + "content": ". Considering top-", + "type": "text" + }, + { + "bbox": [ + 424, + 258, + 431, + 267 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 257, + 469, + 271 + ], + "score": 0.997, + "content": "edges of", + "type": "text" + }, + { + "bbox": [ + 470, + 258, + 478, + 267 + ], + "score": 0.82, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 257, + 506, + 271 + ], + "score": 0.997, + "content": "as the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 267, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 221, + 280 + ], + "score": 0.989, + "content": " ground-truth and comparing", + "type": "text" + }, + { + "bbox": [ + 221, + 268, + 232, + 278 + ], + "score": 0.85, + "content": "S ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 267, + 506, + 280 + ], + "score": 0.987, + "content": " against them, we compute a receiver operating characteristic (ROC)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 0.992, + "content": "curve and evaluate the robustnessby the area under curve (AUC) of the ROC curve. We report results", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 120, + 301 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 289, + 146, + 299 + ], + "score": 0.9, + "content": "k = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 288, + 307, + 301 + ], + "score": 0.985, + "content": "in Figure 2. Results for other values of", + "type": "text" + }, + { + "bbox": [ + 308, + 289, + 314, + 298 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 288, + 506, + 301 + ], + "score": 0.99, + "content": "are included in Appendix F where we observe", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 299, + 162, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 162, + 311 + ], + "score": 0.998, + "content": "similar trend.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 0.984, + "content": "Figure 2 shows the AUC of GNNExplainer, PGExplainer, RCExp-NoLDB and RCExplainer at", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 0.987, + "content": "different levels of noise. A higher AUC indicates better robustness. The percentage of noise", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "score": 0.985, + "content": " shows the proportion of nodes and edges that are modified. Baselines such as PGM-Explainer and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "score": 0.986, + "content": " SubgraphX are not included in this experiment as they do not output the edge weights that are required", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 357, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 369 + ], + "score": 0.991, + "content": "for computing AUC.We present additional robustness experiments in Appendix F where we extend", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 330, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 330, + 380 + ], + "score": 0.988, + "content": " all the baselines to report node and edge level accuracy.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "score": 0.996, + "content": "GNNExplainer performs the worst on most of the datasets,since it optimizes each graph independently", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 393, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 405 + ], + "score": 0.994, + "content": "without considering other graphs in the training set. Even when no noise is added,the AUC of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "score": 0.994, + "content": " GNNExplainer is significantly lower than 1 because different runs produce different explanations", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 413, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 427 + ], + "score": 0.998, + "content": "for the same graph prediction. PGExplainer is generally more robust than GNNExplainer because", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 0.989, + "content": " the neural network they trained to produce explanations implicitly considers all the graphs used", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 0.991, + "content": "for training. CF-GNNExplainer also performs worse than RCExplainer, which means it is more", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 446, + 321, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 321, + 458 + ], + "score": 0.997, + "content": "susceptible to the noise as compared to RCExplainer.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 461, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 0.99, + "content": "Our method achieves the best AUC on allthe datasets, because the common decision logic carried", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "score": 0.987, + "content": "by the decision regions of a GNN is highly robust to noise. PGExplainer achieves a comparable", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 482, + 507, + 494 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 507, + 494 + ], + "score": 0.985, + "content": "performance as our method on the Mutagenicity dataset, because the samples of this dataset share a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "score": 0.986, + "content": "lot of common structures such as carbon rings, which makes it easier for the neural network trained", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 0.984, + "content": "by PGExplainer to identify these structures in presence of noise. However, for BA-2motifs and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 513, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 526 + ], + "score": 0.989, + "content": "NCI1, this is harder as samples share very few structures and thus the AUC of PGExplainer drops", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 0.985, + "content": " significantly. RCExplainer also significantly outperforms RCExp-NoLDB on these datasets which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 534, + 423, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 423, + 546 + ], + "score": 0.989, + "content": "highlights the role of decision boundaries in making our method highly robust.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5 + }, + { + "type": "table", + "bbox": [ + 106, + 556, + 504, + 586 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 556, + 504, + 586 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 556, + 504, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 504, + 586 + ], + "score": 0.95, + "html": "
MethodGNNExplainerPGExplainerPGM-ExplainerSubgraphXCF-GNNExplainerRCExplainer
Time1.2s ± 0.20.01s ± 0.0313.1s ± 3.977.8s ± 4.54.6s± 0.20.01s ± 0.02
", + "type": "table", + "image_path": "747a5f1226941eaba594797ad5d8f60f7224157469cea7e8953b63d2c1a53516.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 106, + 556, + 504, + 566.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 106, + 566.0, + 504, + 576.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 106, + 576.0, + 504, + 586.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 139, + 590, + 470, + 601 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 471, + 604 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 471, + 604 + ], + "score": 0.997, + "content": "Table 1: Average time cost for producing an explanation on a single graph sample.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + } + ], + "index": 37.0 + }, + { + "type": "text", + "bbox": [ + 106, + 619, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 0.998, + "content": "Effciency. We evaluate efficiency by comparing the average computation time taken for inference", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 642 + ], + "score": 0.98, + "content": "on unseen graph samples. Table 1 shows the results on the Mutagenicity dataset. Since our method", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "score": 0.978, + "content": "also can be directly used for unseen data without any retraining,it is as eficient as PGExplainer and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 650, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 481, + 663 + ], + "score": 0.978, + "content": "significantly faster than GNNExplainer, PGM-Explainer, SubgraphX and CF-GNNExplainer.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "title", + "bbox": [ + 107, + 676, + 183, + 689 + ], + "lines": [ + { + "bbox": [ + 104, + 674, + 184, + 692 + ], + "spans": [ + { + "bbox": [ + 104, + 674, + 184, + 692 + ], + "score": 1.0, + "content": "6Conclusion", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 0.988, + "content": " In this paper, we develop a novel method for producing counterfactual explanations on GNNs.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "score": 0.989, + "content": "We extract decision boundaries from the given GNN model to formulate an intuitive and effective", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 70, + 444, + 149 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 167, + 70, + 444, + 149 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 70, + 444, + 149 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 444, + 149 + ], + "score": 0.965, + "type": "image", + "image_path": "1aa166e7cf33380e4add83effc9273d8301d54224aa1859e6e8d4478bf17806a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 167, + 70, + 444, + 96.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 167, + 96.33333333333333, + 444, + 122.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 167, + 122.66666666666666, + 444, + 149.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 164, + 505, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 163, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 507, + 178 + ], + "score": 0.992, + "content": "Figure 2: Noise robustness (AUC) averaged across 10 runs for the datasets at diferent levels of noise.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 196, + 236, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 194, + 237, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 237, + 209 + ], + "score": 1.0, + "content": "5.2Robustness Performance", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 215, + 506, + 310 + ], + "lines": [ + { + "bbox": [ + 104, + 214, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 506, + 229 + ], + "score": 0.988, + "content": " In this experiment, we evaluate the robustness of all methods by quantifying how much an explanation", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 358, + 239 + ], + "score": 0.992, + "content": "changes after adding noise to the input graph. For an input graph", + "type": "text" + }, + { + "bbox": [ + 359, + 227, + 368, + 236 + ], + "score": 0.81, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 226, + 446, + 239 + ], + "score": 0.995, + "content": "and the explanation", + "type": "text" + }, + { + "bbox": [ + 447, + 227, + 454, + 236 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 226, + 505, + 239 + ], + "score": 0.959, + "content": ",we produce", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 235, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 104, + 235, + 179, + 251 + ], + "score": 0.948, + "content": " a perturbed graph", + "type": "text" + }, + { + "bbox": [ + 180, + 237, + 191, + 246 + ], + "score": 0.85, + "content": "G ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 235, + 506, + 251 + ], + "score": 0.988, + "content": "by adding random noise to the node features and randomly adding or deleting", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 337, + 259 + ], + "score": 0.986, + "content": " some edges of the input graph such that the prediction on", + "type": "text" + }, + { + "bbox": [ + 337, + 247, + 348, + 257 + ], + "score": 0.87, + "content": "G ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 246, + 491, + 259 + ], + "score": 0.997, + "content": "is consistent with the prediction on", + "type": "text" + }, + { + "bbox": [ + 492, + 248, + 501, + 257 + ], + "score": 0.78, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 246, + 506, + 259 + ], + "score": 0.794, + "content": "：", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 257, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 313, + 271 + ], + "score": 0.981, + "content": "Using the same method we obtain the explanation", + "type": "text" + }, + { + "bbox": [ + 314, + 258, + 324, + 267 + ], + "score": 0.86, + "content": "S ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 257, + 338, + 271 + ], + "score": 0.997, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 339, + 258, + 350, + 268 + ], + "score": 0.84, + "content": "G ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 257, + 424, + 271 + ], + "score": 0.975, + "content": ". Considering top-", + "type": "text" + }, + { + "bbox": [ + 424, + 258, + 431, + 267 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 257, + 469, + 271 + ], + "score": 0.997, + "content": "edges of", + "type": "text" + }, + { + "bbox": [ + 470, + 258, + 478, + 267 + ], + "score": 0.82, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 257, + 506, + 271 + ], + "score": 0.997, + "content": "as the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 267, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 221, + 280 + ], + "score": 0.989, + "content": " ground-truth and comparing", + "type": "text" + }, + { + "bbox": [ + 221, + 268, + 232, + 278 + ], + "score": 0.85, + "content": "S ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 267, + 506, + 280 + ], + "score": 0.987, + "content": " against them, we compute a receiver operating characteristic (ROC)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 0.992, + "content": "curve and evaluate the robustnessby the area under curve (AUC) of the ROC curve. We report results", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 120, + 301 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 289, + 146, + 299 + ], + "score": 0.9, + "content": "k = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 288, + 307, + 301 + ], + "score": 0.985, + "content": "in Figure 2. Results for other values of", + "type": "text" + }, + { + "bbox": [ + 308, + 289, + 314, + 298 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 288, + 506, + 301 + ], + "score": 0.99, + "content": "are included in Appendix F where we observe", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 299, + 162, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 162, + 311 + ], + "score": 0.998, + "content": "similar trend.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9, + "bbox_fs": [ + 104, + 214, + 506, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 0.984, + "content": "Figure 2 shows the AUC of GNNExplainer, PGExplainer, RCExp-NoLDB and RCExplainer at", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 0.987, + "content": "different levels of noise. A higher AUC indicates better robustness. The percentage of noise", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "score": 0.985, + "content": " shows the proportion of nodes and edges that are modified. Baselines such as PGM-Explainer and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "score": 0.986, + "content": " SubgraphX are not included in this experiment as they do not output the edge weights that are required", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 357, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 369 + ], + "score": 0.991, + "content": "for computing AUC.We present additional robustness experiments in Appendix F where we extend", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 330, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 330, + 380 + ], + "score": 0.988, + "content": " all the baselines to report node and edge level accuracy.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 315, + 506, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 396 + ], + "score": 0.996, + "content": "GNNExplainer performs the worst on most of the datasets,since it optimizes each graph independently", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 393, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 405 + ], + "score": 0.994, + "content": "without considering other graphs in the training set. Even when no noise is added,the AUC of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "score": 0.994, + "content": " GNNExplainer is significantly lower than 1 because different runs produce different explanations", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 413, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 427 + ], + "score": 0.998, + "content": "for the same graph prediction. PGExplainer is generally more robust than GNNExplainer because", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 0.989, + "content": " the neural network they trained to produce explanations implicitly considers all the graphs used", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 0.991, + "content": "for training. CF-GNNExplainer also performs worse than RCExplainer, which means it is more", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 446, + 321, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 321, + 458 + ], + "score": 0.997, + "content": "susceptible to the noise as compared to RCExplainer.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 382, + 506, + 458 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 461, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 0.99, + "content": "Our method achieves the best AUC on allthe datasets, because the common decision logic carried", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 484 + ], + "score": 0.987, + "content": "by the decision regions of a GNN is highly robust to noise. PGExplainer achieves a comparable", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 482, + 507, + 494 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 507, + 494 + ], + "score": 0.985, + "content": "performance as our method on the Mutagenicity dataset, because the samples of this dataset share a", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "score": 0.986, + "content": "lot of common structures such as carbon rings, which makes it easier for the neural network trained", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 0.984, + "content": "by PGExplainer to identify these structures in presence of noise. However, for BA-2motifs and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 513, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 526 + ], + "score": 0.989, + "content": "NCI1, this is harder as samples share very few structures and thus the AUC of PGExplainer drops", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 0.985, + "content": " significantly. RCExplainer also significantly outperforms RCExp-NoLDB on these datasets which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 534, + 423, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 423, + 546 + ], + "score": 0.989, + "content": "highlights the role of decision boundaries in making our method highly robust.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5, + "bbox_fs": [ + 104, + 461, + 507, + 546 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 556, + 504, + 586 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 556, + 504, + 586 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 556, + 504, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 504, + 586 + ], + "score": 0.95, + "html": "
MethodGNNExplainerPGExplainerPGM-ExplainerSubgraphXCF-GNNExplainerRCExplainer
Time1.2s ± 0.20.01s ± 0.0313.1s ± 3.977.8s ± 4.54.6s± 0.20.01s ± 0.02
", + "type": "table", + "image_path": "747a5f1226941eaba594797ad5d8f60f7224157469cea7e8953b63d2c1a53516.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 106, + 556, + 504, + 566.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 106, + 566.0, + 504, + 576.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 106, + 576.0, + 504, + 586.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 139, + 590, + 470, + 601 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 471, + 604 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 471, + 604 + ], + "score": 0.997, + "content": "Table 1: Average time cost for producing an explanation on a single graph sample.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + } + ], + "index": 37.0 + }, + { + "type": "text", + "bbox": [ + 106, + 619, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 0.998, + "content": "Effciency. We evaluate efficiency by comparing the average computation time taken for inference", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 629, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 642 + ], + "score": 0.98, + "content": "on unseen graph samples. Table 1 shows the results on the Mutagenicity dataset. Since our method", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 651 + ], + "score": 0.978, + "content": "also can be directly used for unseen data without any retraining,it is as eficient as PGExplainer and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 650, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 481, + 663 + ], + "score": 0.978, + "content": "significantly faster than GNNExplainer, PGM-Explainer, SubgraphX and CF-GNNExplainer.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 618, + 506, + 663 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 676, + 183, + 689 + ], + "lines": [ + { + "bbox": [ + 104, + 674, + 184, + 692 + ], + "spans": [ + { + "bbox": [ + 104, + 674, + 184, + 692 + ], + "score": 1.0, + "content": "6Conclusion", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 0.988, + "content": " In this paper, we develop a novel method for producing counterfactual explanations on GNNs.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "score": 0.989, + "content": "We extract decision boundaries from the given GNN model to formulate an intuitive and effective", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "score": 0.989, + "content": "counterfactual loss function. We optimize this loss to train a neural network to produce explanations", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 0.989, + "content": "with strong counterfactual characteristics. Since the decision boundaries are shared by multiple", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 105 + ], + "score": 0.986, + "content": " samples of the same predicted class, explanations produced by our method are robust and do not", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "score": 0.987, + "content": "overfit the noise. Our experiments on synthetic and real-life benchmark datasets strongly validate", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 114, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 126 + ], + "score": 0.982, + "content": "the efficacy of our method. In this work,we focus on GNNs that belong to Piecewise Linear", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "score": 0.994, + "content": "Neural Networks (PLNNs). Extending our method to other families of GNNs and tasks such as link", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 135, + 309, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 309, + 147 + ], + "score": 0.999, + "content": "prediction,remains an interesting future direction.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 700, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "score": 0.989, + "content": "counterfactual loss function. We optimize this loss to train a neural network to produce explanations", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 0.989, + "content": "with strong counterfactual characteristics. Since the decision boundaries are shared by multiple", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 105 + ], + "score": 0.986, + "content": " samples of the same predicted class, explanations produced by our method are robust and do not", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "score": 0.987, + "content": "overfit the noise. Our experiments on synthetic and real-life benchmark datasets strongly validate", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 114, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 126 + ], + "score": 0.982, + "content": "the efficacy of our method. In this work,we focus on GNNs that belong to Piecewise Linear", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "score": 0.994, + "content": "Neural Networks (PLNNs). Extending our method to other families of GNNs and tasks such as link", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 135, + 309, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 309, + 147 + ], + "score": 0.999, + "content": "prediction,remains an interesting future direction.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "score": 0.985, + "content": "Our method will benefit multiple fields where GNNs are intensively used. By allowing the users to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 162, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 174 + ], + "score": 0.996, + "content": " interpret the predictions of complex GNNs better, it will promote transparency, trust and fairness", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "score": 0.991, + "content": " in the society. However, there also exist some inherent risks. A generated explanation may expose", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 182, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 104, + 182, + 506, + 195 + ], + "score": 0.985, + "content": " private information if our method is not coupled with an adequate privacy protection technique. Also,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 193, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 506, + 205 + ], + "score": 0.989, + "content": " some of the ideas presented in this paper may be adopted and extended to improve adversarial attacks.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "score": 0.993, + "content": "Without appropriate defense mechanisms,the misuse of such attcks poses a risk of disruption in the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 225 + ], + "score": 0.979, + "content": "functionality of GNNs deployed in the real world.That said, we firmly believe that these risks can be", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 223, + 362, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 362, + 236 + ], + "score": 0.989, + "content": "mitigated through increased awareness and proactive measures.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 107, + 250, + 339, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 341, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 341, + 267 + ], + "score": 0.996, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 275, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "score": 0.993, + "content": "Lingyang Chu's research is supported in part by the startup grant provided by the Department", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "score": 0.984, + "content": "of Computing and Software of McMaster University. All opinions, findings,conclusions,and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 295, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 309 + ], + "score": 0.989, + "content": "recommendations in this paper are those of the authors and do not necessarily reflect the views of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 306, + 179, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 179, + 320 + ], + "score": 1.0, + "content": "funding agencies.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 333, + 163, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 165, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 165, + 347 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 109, + 347, + 506, + 723 + ], + "lines": [ + { + "bbox": [ + 109, + 350, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 109, + 350, + 507, + 365 + ], + "score": 0.961, + "content": "[1] J. Adebayo,J. Gilmer,M.Muelly,I. Goodfellow,M. Hardt,and B. Kim. Sanity checks for saliency maps.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 361, + 329, + 374 + ], + "spans": [ + { + "bbox": [ + 125, + 361, + 329, + 374 + ], + "score": 0.982, + "content": "In Advances in Neural Information Processing Systems.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 109, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 109, + 379, + 506, + 393 + ], + "score": 0.969, + "content": "[2] A.Bojchevski and S. Gunnemann. Certifiable robustness to graph perturbations.arXiv preprint", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 389, + 219, + 401 + ], + "spans": [ + { + "bbox": [ + 125, + 389, + 219, + 401 + ], + "score": 0.998, + "content": "arXiv:1910.14356,2019.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 109, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 109, + 408, + 506, + 421 + ], + "score": 0.982, + "content": "[3]E. Cho,S.A. Myers,and J.Leskovec.Friendship and mobility: user movement in location-based social", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 124, + 417, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 124, + 417, + 506, + 432 + ], + "score": 0.98, + "content": "networks. In Proceedings of the 17th ACM SIGkDD international conference on Knowledge discovery", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 125, + 427, + 280, + 439 + ], + "spans": [ + { + "bbox": [ + 125, + 427, + 280, + 439 + ], + "score": 0.983, + "content": "and data mining,pages 1082-1090,2011.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 110, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 110, + 446, + 505, + 459 + ], + "score": 0.978, + "content": "[4]L.Chu, X. Hu,J. Hu,L. Wang,and J.Pei. Exact and consistent interpretation for piecewise linear neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 124, + 455, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 124, + 455, + 506, + 469 + ], + "score": 0.978, + "content": "networks: A closed form solution. In Proceedings of the 24th ACM SiGKDD International Conference on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 465, + 357, + 478 + ], + "spans": [ + { + "bbox": [ + 125, + 465, + 357, + 478 + ], + "score": 0.986, + "content": "Knowledge Discovery & Data Mining,pages 1244-1253,2018.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 109, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 109, + 483, + 505, + 497 + ], + "score": 0.984, + "content": "[5]V. Crawford,A. Kuhnle,and M. Thai. Submodular cost submodular cover with an approximate oracle.In", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 124, + 493, + 421, + 506 + ], + "spans": [ + { + "bbox": [ + 124, + 493, + 421, + 506 + ], + "score": 0.989, + "content": " International Conference on Machine Learning, pages 1426-1435.PMLR,2019.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 109, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 109, + 512, + 505, + 525 + ], + "score": 0.97, + "content": "[6] A. K. Debnath, R.L.Lopez de Compadre, G. Debnath, A. J. Shusterman,and C. Hansch. Structure-activity", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 125, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 125, + 522, + 505, + 534 + ], + "score": 0.994, + "content": "relationship of mutagenic aromatic and heteroaromatic nitro compounds.correlation with molecular orbital", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 532, + 430, + 545 + ], + "spans": [ + { + "bbox": [ + 125, + 532, + 430, + 545 + ], + "score": 0.993, + "content": "energies and hydrophobicity. Journal of medicinal chemistry,34(2):786-797,1991.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 109, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 109, + 550, + 506, + 564 + ], + "score": 0.984, + "content": "[7]M.Denil,S.G.Colmenarejo,S. Cabi,D.Saxton,and N. de Freitas. Programmable agents. arXiv preprint", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 125, + 561, + 219, + 572 + ], + "spans": [ + { + "bbox": [ + 125, + 561, + 219, + 572 + ], + "score": 0.998, + "content": "arXiv:1706.06383,2017.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 109, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 109, + 578, + 506, + 592 + ], + "score": 0.985, + "content": "[8]Y. Duan, M. Andrychowicz,B.C. Stadie,J. Ho,J.Schneider,I.Sutskever,P.Abbeel,and W. Zaremba.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 125, + 588, + 377, + 601 + ], + "spans": [ + { + "bbox": [ + 125, + 588, + 377, + 601 + ], + "score": 0.986, + "content": "One-shot imitation learning. arXiv preprint arXiv:1703.07326,2017.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 109, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 109, + 607, + 506, + 621 + ], + "score": 0.968, + "content": "[9]W.Fan,Y.Ma,Q.Li,Y. He,E. Zhao,J.Tang,and D.Yin. Graph neural networks for social recommendation.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 617, + 340, + 630 + ], + "spans": [ + { + "bbox": [ + 125, + 617, + 340, + 630 + ], + "score": 0.992, + "content": "In The World Wide Web Conference,pages 417-426,2019.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 634, + 507, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 507, + 650 + ], + "score": 0.93, + "content": "[10]T.Freiesleben.Counterfactualexplanations &adversarialexamplescommongrounds,essntialdifferences,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 125, + 645, + 357, + 658 + ], + "spans": [ + { + "bbox": [ + 125, + 645, + 357, + 658 + ], + "score": 0.981, + "content": "and potential transfers. arXiv preprint arXiv:2009.05487, 2020.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 678 + ], + "score": 0.973, + "content": "[11]A. Ghorbani,A. Abid,and J. Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 125, + 674, + 394, + 686 + ], + "spans": [ + { + "bbox": [ + 125, + 674, + 394, + 686 + ], + "score": 0.98, + "content": "Conference on Artificial Intelligence, volume 33, pages 3681-3688,2019.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 692, + 507, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 507, + 706 + ], + "score": 0.98, + "content": "[12] X. Glorot and Y. Bengio. Understanding the dificulty of training deep feedforward neural networks.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 124, + 701, + 507, + 717 + ], + "spans": [ + { + "bbox": [ + 124, + 701, + 507, + 717 + ], + "score": 0.967, + "content": "InProceedingsofthethirteenth international conference onartificial intelligence and statistics,pages", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 711, + 357, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 711, + 357, + 725 + ], + "score": 0.994, + "content": "249-256. JMLR Workshop and Conference Proceedings, 2010.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 34.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 146 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 105, + 72, + 506, + 147 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "score": 0.985, + "content": "Our method will benefit multiple fields where GNNs are intensively used. By allowing the users to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 162, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 174 + ], + "score": 0.996, + "content": " interpret the predictions of complex GNNs better, it will promote transparency, trust and fairness", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "score": 0.991, + "content": " in the society. However, there also exist some inherent risks. A generated explanation may expose", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 182, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 104, + 182, + 506, + 195 + ], + "score": 0.985, + "content": " private information if our method is not coupled with an adequate privacy protection technique. Also,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 193, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 506, + 205 + ], + "score": 0.989, + "content": " some of the ideas presented in this paper may be adopted and extended to improve adversarial attacks.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "score": 0.993, + "content": "Without appropriate defense mechanisms,the misuse of such attcks poses a risk of disruption in the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 225 + ], + "score": 0.979, + "content": "functionality of GNNs deployed in the real world.That said, we firmly believe that these risks can be", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 223, + 362, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 362, + 236 + ], + "score": 0.989, + "content": "mitigated through increased awareness and proactive measures.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 151, + 506, + 236 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 250, + 339, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 341, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 341, + 267 + ], + "score": 0.996, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 275, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "score": 0.993, + "content": "Lingyang Chu's research is supported in part by the startup grant provided by the Department", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 298 + ], + "score": 0.984, + "content": "of Computing and Software of McMaster University. All opinions, findings,conclusions,and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 295, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 309 + ], + "score": 0.989, + "content": "recommendations in this paper are those of the authors and do not necessarily reflect the views of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 306, + 179, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 179, + 320 + ], + "score": 1.0, + "content": "funding agencies.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 275, + 506, + 320 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 333, + 163, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 165, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 165, + 347 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "list", + "bbox": [ + 109, + 347, + 506, + 723 + ], + "lines": [ + { + "bbox": [ + 109, + 350, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 109, + 350, + 507, + 365 + ], + "score": 0.961, + "content": "[1] J. Adebayo,J. Gilmer,M.Muelly,I. Goodfellow,M. Hardt,and B. Kim. Sanity checks for saliency maps.", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 361, + 329, + 374 + ], + "spans": [ + { + "bbox": [ + 125, + 361, + 329, + 374 + ], + "score": 0.982, + "content": "In Advances in Neural Information Processing Systems.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 109, + 379, + 506, + 393 + ], + "score": 0.969, + "content": "[2] A.Bojchevski and S. Gunnemann. Certifiable robustness to graph perturbations.arXiv preprint", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 389, + 219, + 401 + ], + "spans": [ + { + "bbox": [ + 125, + 389, + 219, + 401 + ], + "score": 0.998, + "content": "arXiv:1910.14356,2019.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 109, + 408, + 506, + 421 + ], + "score": 0.982, + "content": "[3]E. Cho,S.A. Myers,and J.Leskovec.Friendship and mobility: user movement in location-based social", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 417, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 124, + 417, + 506, + 432 + ], + "score": 0.98, + "content": "networks. In Proceedings of the 17th ACM SIGkDD international conference on Knowledge discovery", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 125, + 427, + 280, + 439 + ], + "spans": [ + { + "bbox": [ + 125, + 427, + 280, + 439 + ], + "score": 0.983, + "content": "and data mining,pages 1082-1090,2011.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 110, + 446, + 505, + 459 + ], + "score": 0.978, + "content": "[4]L.Chu, X. Hu,J. Hu,L. Wang,and J.Pei. Exact and consistent interpretation for piecewise linear neural", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 455, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 124, + 455, + 506, + 469 + ], + "score": 0.978, + "content": "networks: A closed form solution. In Proceedings of the 24th ACM SiGKDD International Conference on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 465, + 357, + 478 + ], + "spans": [ + { + "bbox": [ + 125, + 465, + 357, + 478 + ], + "score": 0.986, + "content": "Knowledge Discovery & Data Mining,pages 1244-1253,2018.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 109, + 483, + 505, + 497 + ], + "score": 0.984, + "content": "[5]V. Crawford,A. Kuhnle,and M. Thai. Submodular cost submodular cover with an approximate oracle.In", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 493, + 421, + 506 + ], + "spans": [ + { + "bbox": [ + 124, + 493, + 421, + 506 + ], + "score": 0.989, + "content": " International Conference on Machine Learning, pages 1426-1435.PMLR,2019.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 109, + 512, + 505, + 525 + ], + "score": 0.97, + "content": "[6] A. K. Debnath, R.L.Lopez de Compadre, G. Debnath, A. J. Shusterman,and C. Hansch. Structure-activity", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 125, + 522, + 505, + 534 + ], + "score": 0.994, + "content": "relationship of mutagenic aromatic and heteroaromatic nitro compounds.correlation with molecular orbital", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 532, + 430, + 545 + ], + "spans": [ + { + "bbox": [ + 125, + 532, + 430, + 545 + ], + "score": 0.993, + "content": "energies and hydrophobicity. Journal of medicinal chemistry,34(2):786-797,1991.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 109, + 550, + 506, + 564 + ], + "score": 0.984, + "content": "[7]M.Denil,S.G.Colmenarejo,S. Cabi,D.Saxton,and N. de Freitas. Programmable agents. arXiv preprint", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 561, + 219, + 572 + ], + "spans": [ + { + "bbox": [ + 125, + 561, + 219, + 572 + ], + "score": 0.998, + "content": "arXiv:1706.06383,2017.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 109, + 578, + 506, + 592 + ], + "score": 0.985, + "content": "[8]Y. Duan, M. Andrychowicz,B.C. Stadie,J. Ho,J.Schneider,I.Sutskever,P.Abbeel,and W. Zaremba.", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 588, + 377, + 601 + ], + "spans": [ + { + "bbox": [ + 125, + 588, + 377, + 601 + ], + "score": 0.986, + "content": "One-shot imitation learning. arXiv preprint arXiv:1703.07326,2017.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 109, + 607, + 506, + 621 + ], + "score": 0.968, + "content": "[9]W.Fan,Y.Ma,Q.Li,Y. He,E. Zhao,J.Tang,and D.Yin. Graph neural networks for social recommendation.", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 617, + 340, + 630 + ], + "spans": [ + { + "bbox": [ + 125, + 617, + 340, + 630 + ], + "score": 0.992, + "content": "In The World Wide Web Conference,pages 417-426,2019.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 634, + 507, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 507, + 650 + ], + "score": 0.93, + "content": "[10]T.Freiesleben.Counterfactualexplanations &adversarialexamplescommongrounds,essntialdifferences,", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 645, + 357, + 658 + ], + "spans": [ + { + "bbox": [ + 125, + 645, + 357, + 658 + ], + "score": 0.981, + "content": "and potential transfers. arXiv preprint arXiv:2009.05487, 2020.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 505, + 678 + ], + "score": 0.973, + "content": "[11]A. Ghorbani,A. Abid,and J. Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 674, + 394, + 686 + ], + "spans": [ + { + "bbox": [ + 125, + 674, + 394, + 686 + ], + "score": 0.98, + "content": "Conference on Artificial Intelligence, volume 33, pages 3681-3688,2019.", + "type": "text" + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 692, + 507, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 507, + 706 + ], + "score": 0.98, + "content": "[12] X. Glorot and Y. Bengio. Understanding the dificulty of training deep feedforward neural networks.", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 701, + 507, + 717 + ], + "spans": [ + { + "bbox": [ + 124, + 701, + 507, + 717 + ], + "score": 0.967, + "content": "InProceedingsofthethirteenth international conference onartificial intelligence and statistics,pages", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 711, + 357, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 711, + 357, + 725 + ], + "score": 0.994, + "content": "249-256. JMLR Workshop and Conference Proceedings, 2010.", + "type": "text" + } + ], + "index": 48, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "score": 0.977, + "content": "[13] X.Glorot,A.Bordes,and Y.Bengio.Deep sparse rectifier neural networks.In Proceedings ofthe fourteenth", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 506, + 96 + ], + "score": 0.989, + "content": "international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 92, + 241, + 104 + ], + "spans": [ + { + "bbox": [ + 126, + 92, + 241, + 104 + ], + "score": 0.998, + "content": "Conference Proceedings,2011.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 109, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 506, + 122 + ], + "score": 0.968, + "content": "[14] I. Goodfelow,D. Warde-Farley,M. Mirza,A. Courvile,and Y.Bengio. Maxout networks.In International", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 123, + 118, + 366, + 132 + ], + "spans": [ + { + "bbox": [ + 123, + 118, + 366, + 132 + ], + "score": 0.994, + "content": "conference on machine learning, pages 1319-1327.PMLR,2013.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 135, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 135, + 506, + 150 + ], + "score": 0.982, + "content": "[15]K. He, X. Zhang, S. Ren,and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 145, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 124, + 145, + 506, + 160 + ], + "score": 0.968, + "content": " imagenet classification. In Proceedings ofthe IEEE international conferenceon computer vision, pages", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 156, + 192, + 165 + ], + "spans": [ + { + "bbox": [ + 126, + 156, + 192, + 165 + ], + "score": 0.997, + "content": "1026-1034,2015.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "score": 0.985, + "content": "[16] L.Holdijk,M. Boon,S.Henckens,and L.de Jong. [re] parameterized explainer for graph neural network.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 181, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 125, + 181, + 294, + 195 + ], + "score": 0.957, + "content": "In ML Reproducibility Challenge 2020,2021.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 200, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 506, + 211 + ], + "score": 0.963, + "content": "[17] W. Huber, V.J. Carey,L.Long, S. Falcon,and R. Gentleman. Graphs in molecular biology. BMC", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 210, + 247, + 221 + ], + "spans": [ + { + "bbox": [ + 126, + 210, + 247, + 221 + ], + "score": 0.998, + "content": "bioinformatics,8(6):1-14,2007.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "score": 0.98, + "content": "[18] R. K. Iyer and J. A. Bilmes. Submodular optimization with submodular cover and submodular knapsack", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 236, + 374, + 248 + ], + "spans": [ + { + "bbox": [ + 126, + 236, + 374, + 248 + ], + "score": 0.974, + "content": "constraints.In Advances in Neural Information Processing Systems.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "score": 0.969, + "content": "[19]M. Jiang,Z. Li, S. Zhang,S. Wang,X. Wang, Q. Yuan,and Z. Wei. Drug-target affinity prediction using", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 123, + 262, + 434, + 276 + ], + "spans": [ + { + "bbox": [ + 123, + 262, + 434, + 276 + ], + "score": 0.981, + "content": "graph neural network and contact maps.RsC Advances,10(35):20701-20712, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "score": 0.977, + "content": "[20] H. Jin and X. Zhang. Latent adversarial training of graph convolution networks. In ICML Workshop on", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 290, + 385, + 302 + ], + "spans": [ + { + "bbox": [ + 125, + 290, + 385, + 302 + ], + "score": 0.988, + "content": "Learning and Reasoning with Graph-Structured Representations,2019.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 307, + 507, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 507, + 321 + ], + "score": 0.975, + "content": "[21] J. Kazius,R. McGuire,and R. Bursi. Derivation and validation of toxicophores for mutagenicity prediction.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 317, + 323, + 329 + ], + "spans": [ + { + "bbox": [ + 125, + 317, + 323, + 329 + ], + "score": 0.989, + "content": "Journal of medicinal chemistry,48(1):312-320,2005.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "score": 0.976, + "content": "[22] T.N. Kipf and M. Welling. Semi-supervised classfication with graph convolutional networks. In 5th", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 343, + 487, + 356 + ], + "spans": [ + { + "bbox": [ + 125, + 343, + 487, + 356 + ], + "score": 0.976, + "content": "International Conference on Learning Representations,ICLR 2Ol7, Conference Track Proceedings.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 361, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 506, + 373 + ], + "score": 0.964, + "content": "[23]M. Liu, Y. Luo,L. Wang, Y. Xie, H. Yuan,S. Gui, H. Yu,Z. Xu, J. Zhang, Y. Liu, K. Yan, H. Liu, C. Fu,", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 371, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 126, + 371, + 506, + 382 + ], + "score": 0.976, + "content": "B. Oztekin,X. Zhang,and S.Ji. DIG: A turnkey library for diving into graph deep learning research. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 123, + 379, + 251, + 392 + ], + "spans": [ + { + "bbox": [ + 123, + 379, + 251, + 392 + ], + "score": 0.999, + "content": "preprint arXiv:2103.12608,2021.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 398, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 506, + 409 + ], + "score": 0.98, + "content": "[24] A.Lucic,M.ter Hoeve,G. Tolomei,M.de Rijke,and F. Silvestri. Cf-gnnexplainer: Counterfactual", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 407, + 419, + 419 + ], + "spans": [ + { + "bbox": [ + 126, + 407, + 419, + 419 + ], + "score": 0.99, + "content": "explanations for graph neural networks.arXiv preprint arXiv:2102.03322,2021.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 0.981, + "content": "[25] D.Luo, W. Cheng,D. Xu, W. Yu, B. Zong, H. Chen,and X. Zhang. Parameterized explainer for graph", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 432, + 389, + 447 + ], + "spans": [ + { + "bbox": [ + 124, + 432, + 389, + 447 + ], + "score": 0.983, + "content": "neural network. In Advances in Neural Information Processing Systems.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 448, + 507, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 507, + 465 + ], + "score": 0.989, + "content": "[26] C. Molnar.Interpretable Machine Learning.2019.https://christophm.github.io/", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 461, + 237, + 473 + ], + "spans": [ + { + "bbox": [ + 126, + 461, + 237, + 473 + ], + "score": 1.0, + "content": "interpretable-ml-book/.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 478, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 506, + 489 + ], + "score": 0.988, + "content": "[27] G.Montufar,R.Pascanu,K.Cho,and Y.Bengio.On the numberof linear regions of deep neural networks.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 487, + 268, + 498 + ], + "spans": [ + { + "bbox": [ + 125, + 487, + 268, + 498 + ], + "score": 0.997, + "content": "arXiv preprint arXiv:1402.1869,2014.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 502, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 104, + 502, + 507, + 519 + ], + "score": 0.954, + "content": "[28]R.Moraffah,M. Karami,R.Guo,A. Raglin,and H.Liu.Causal interpretability for machine learning-", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 514, + 475, + 527 + ], + "spans": [ + { + "bbox": [ + 124, + 514, + 475, + 527 + ], + "score": 0.992, + "content": "problems,methods and evaluation. ACM SIGKDD Explorations Newsleter,22(1):18-33,2020.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 532, + 498, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 498, + 543 + ], + "score": 0.994, + "content": "[29] V.Nair and G.E.Hinton.Rectified linear units improve restricted boltzmann machines.In Icml,2010.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 547, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 562 + ], + "score": 0.986, + "content": "[30] D. Numeroso and D. Bacciu. Meg: Generating molecular counterfactual explanations for deep graph", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 559, + 312, + 570 + ], + "spans": [ + { + "bbox": [ + 124, + 559, + 312, + 570 + ], + "score": 0.988, + "content": "networks. arXiv preprint arXiv:2104.08060,2021.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "score": 0.971, + "content": "[31] D.E. Pires,T. L. Blundel, and D. B. Ascher. pkcsm: predicting small-molecule pharmacokinetic and", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 585, + 507, + 598 + ], + "spans": [ + { + "bbox": [ + 124, + 585, + 507, + 598 + ], + "score": 0.992, + "content": "toxicity properties using graph-based signatures. Journal of medicinal chemistry,58(9):4066-4072,2015.", + "type": "text", + "cross_page": true + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 603, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 506, + 614 + ], + "score": 0.976, + "content": "[32]P.E.Pope,S. Kolouri, M.Rostami, C.E. Martin,and H. Hoffmann.Explainability methods for graph", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 613, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 126, + 613, + 506, + 624 + ], + "score": 0.991, + "content": "convolutional neural networks.In Proceedings of the IEEE/CVF Conference on Computer Vision and", + "type": "text", + "cross_page": true + } + ], + "index": 41 + }, + { + "bbox": [ + 125, + 621, + 303, + 635 + ], + "spans": [ + { + "bbox": [ + 125, + 621, + 303, + 635 + ], + "score": 0.991, + "content": "Pattern Recognition, pages 10772-10781,2019.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 639, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 650 + ], + "score": 0.988, + "content": "[33]R.R.Selvaraju,M.Cogswell,A.Das,R. Vedantam,D.Parikh,and D.Batra. Grad-cam: Visual explanations", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 649, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 126, + 649, + 506, + 660 + ], + "score": 0.977, + "content": "from deep networks via gradient-based localization. In Proceedings of the IEEE international conference", + "type": "text", + "cross_page": true + } + ], + "index": 44 + }, + { + "bbox": [ + 125, + 658, + 282, + 672 + ], + "spans": [ + { + "bbox": [ + 125, + 658, + 282, + 672 + ], + "score": 0.993, + "content": "on computer vision, pages 618-626, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 674, + 506, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 506, + 688 + ], + "score": 0.984, + "content": "[34] P.L. Simon, M.Taylor,and I. Z. Kiss. Exact epidemic models on graphs using graph-automorphism driven", + "type": "text", + "cross_page": true + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 685, + 365, + 697 + ], + "spans": [ + { + "bbox": [ + 125, + 685, + 365, + 697 + ], + "score": 0.991, + "content": "lumping. Journal of mathematical biology,62(4):479-508,2011.", + "type": "text", + "cross_page": true + } + ], + "index": 47, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 700, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 716 + ], + "score": 0.987, + "content": "[35] K. Simonyan, A. Vedaldi,and A. Zisserman. Deep inside convolutional networks: Visualising image", + "type": "text", + "cross_page": true + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 712, + 322, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 712, + 322, + 724 + ], + "score": 0.992, + "content": "classification models and saliency maps. ICLR,2014.", + "type": "text", + "cross_page": true + } + ], + "index": 49, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 86 + ], + "score": 0.986, + "content": "[36]K.Sokol and P.A.Flach. Counterfactual explanations of machine learning predictions: opportunities and", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 84, + 305, + 95 + ], + "spans": [ + { + "bbox": [ + 126, + 84, + 305, + 95 + ], + "score": 0.963, + "content": "challenges for ai safety. In SafeAI@ AAAI, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 101, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 505, + 112 + ], + "score": 0.963, + "content": "[37]P. Velickovic, G. Cucurul,A. Casanova,A. Romero,P.Lio,and Y. Bengio. Graph attention networks. In", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 110, + 501, + 123 + ], + "spans": [ + { + "bbox": [ + 125, + 110, + 501, + 123 + ], + "score": 0.977, + "content": "6th International Conference on Learning Representations, ICLR 2O18, Conference Track Proceedings.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 140 + ], + "score": 0.985, + "content": "[38]M. Vu and M. T. Thai.Pgm-explainer: Probabilistic graphical model explanations for graph neural", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 138, + 367, + 149 + ], + "spans": [ + { + "bbox": [ + 126, + 138, + 367, + 149 + ], + "score": 0.973, + "content": "networks.In Advances in Neural Information Processing Systems.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 154, + 507, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 507, + 167 + ], + "score": 0.984, + "content": "[39] N. Wale and G. Karypis. Comparison of descriptor spaces for chemical compound retrieval and clas-", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 125, + 164, + 506, + 176 + ], + "score": 0.975, + "content": "sification. In Sixth International Conference on Data Mining (ICDM'06),pages 678-689,2006.doi:", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 174, + 217, + 185 + ], + "spans": [ + { + "bbox": [ + 125, + 174, + 217, + 185 + ], + "score": 0.999, + "content": "10.1109/ICDM.2006.39.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "score": 0.983, + "content": "[40] L.A.Wolsey. An analysis of the greedy algorithm for the submodular setcovering problem. Combinatorica,", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 200, + 203, + 212 + ], + "spans": [ + { + "bbox": [ + 125, + 200, + 203, + 212 + ], + "score": 0.977, + "content": "2(4):385-393,1982.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 217, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 231 + ], + "score": 0.987, + "content": "[41] J. Xiong,Z. Xiong, K. Chen,H. Jiang,and M. Zheng. Graph neural networks for automated de novo drug", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 227, + 264, + 240 + ], + "spans": [ + { + "bbox": [ + 124, + 227, + 264, + 240 + ], + "score": 0.991, + "content": "design. Drug Discovery Today,2021.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 0.959, + "content": "[42]H. Xu, Y. Ma, H.-C.Liu,D. Deb, H. Liu, J.-L. Tang,and A. K. Jain. Adversarial atacks and defenses in", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 254, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 125, + 254, + 506, + 267 + ], + "score": 0.991, + "content": " images, graphs and text: A review. International Journal of Automation and Computing,17(2):151-178,", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 263, + 150, + 276 + ], + "spans": [ + { + "bbox": [ + 125, + 263, + 150, + 276 + ], + "score": 0.999, + "content": "2020.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "score": 0.966, + "content": "[43] K. Xu, H. Chen,S.Liu,P.-Y. Chen, T.-W. Weng,M. Hong,and X. Lin. Topology attack and defense for", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 291, + 465, + 304 + ], + "spans": [ + { + "bbox": [ + 124, + 291, + 465, + 304 + ], + "score": 0.982, + "content": " graph neural networks: An optimization perspective. arXiv preprint arXiv:1906.04214, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 306, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 306, + 507, + 322 + ], + "score": 0.98, + "content": "[44] R. Yin,K.Li, G. Zhang,and J.Lu. A deeper graph neural network for recommender systems. Knowledge-", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 318, + 253, + 329 + ], + "spans": [ + { + "bbox": [ + 126, + 318, + 253, + 329 + ], + "score": 0.986, + "content": "Based Systems,185:105020,2019.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "score": 0.98, + "content": "[45] Z. Ying,D. Bourgeois, J. You, M. Zitnik,and J.Leskovec. Gnnexplainer: Generating explanations for", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 344, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 124, + 344, + 480, + 358 + ], + "score": 0.984, + "content": " graph neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "score": 0.983, + "content": "[46] H.Yuan,J.Tang,X.Hu,and S.Ji. Xgnn: Towards model-level explanations of graph neural networks.In", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 370, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 124, + 370, + 507, + 386 + ], + "score": 0.98, + "content": "Proceedings ofthe 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 124, + 381, + 208, + 393 + ], + "spans": [ + { + "bbox": [ + 124, + 381, + 208, + 393 + ], + "score": 0.995, + "content": "pages 430-438,2020.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "score": 0.967, + "content": "[47] H. Yuan, H. Yu, J. Wang, K. Li, and S.Ji._On explainability of graph neural networks via subgraph", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 408, + 323, + 420 + ], + "spans": [ + { + "bbox": [ + 125, + 408, + 323, + 420 + ], + "score": 0.989, + "content": "explorations. arXiv preprint arXiv:2102.05152,2021.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 424, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 438 + ], + "score": 0.986, + "content": "[48] M.D. Zeiler and R.Fergus. Visualizing and understanding convolutional networks. In European conference", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 435, + 317, + 447 + ], + "spans": [ + { + "bbox": [ + 124, + 435, + 317, + 447 + ], + "score": 0.978, + "content": "on computer vision, pages 818-833. Springer,2014.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "score": 0.966, + "content": "[49] J. Zhang,S.A. Bargal,Z.Lin,J.Brandt,X. Shen,and S.Sclarof. Top-down neural attention by excitation", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 462, + 418, + 475 + ], + "spans": [ + { + "bbox": [ + 125, + 462, + 418, + 475 + ], + "score": 0.993, + "content": "backprop. International Journal of Computer Vision,126(10):1084-1102,2018.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 0.981, + "content": "[50] M. Zhang and Y. Chen. Link prediction based on graph neural networks. In Advances in Neural Information", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 488, + 201, + 502 + ], + "spans": [ + { + "bbox": [ + 126, + 488, + 201, + 502 + ], + "score": 0.999, + "content": "Processing Systems.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "score": 0.983, + "content": "[51] D. Zugner and S. Gunnemann. Adversarial attacks on graph neural networks via meta learning. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 516, + 251, + 529 + ], + "spans": [ + { + "bbox": [ + 124, + 516, + 251, + 529 + ], + "score": 0.98, + "content": "preprint arXiv:1902.08412, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 534, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 506, + 545 + ], + "score": 0.978, + "content": "[52] D.Zügner and S.Gunnemann.Certifable robustness and robust training for graph convolutional networks.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 542, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 124, + 542, + 506, + 556 + ], + "score": 0.99, + "content": "In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 125, + 552, + 239, + 565 + ], + "spans": [ + { + "bbox": [ + 125, + 552, + 239, + 565 + ], + "score": 0.998, + "content": "Mining,pages 246-256,2019.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 0.962, + "content": "[53] D. Zugner, A. Akbarnejad,and S. Gunnemann. Adversarial atacks on neural networks for graph data. In", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 578, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 124, + 578, + 506, + 593 + ], + "score": 0.987, + "content": "Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,", + "type": "text", + "cross_page": true + } + ], + "index": 39 + }, + { + "bbox": [ + 123, + 588, + 217, + 601 + ], + "spans": [ + { + "bbox": [ + 123, + 588, + 217, + 601 + ], + "score": 0.999, + "content": "pages 2847-2856,2018.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_end_line": true + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 350, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 66, + 507, + 726 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "score": 0.977, + "content": "[13] X.Glorot,A.Bordes,and Y.Bengio.Deep sparse rectifier neural networks.In Proceedings ofthe fourteenth", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 506, + 96 + ], + "score": 0.989, + "content": "international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 92, + 241, + 104 + ], + "spans": [ + { + "bbox": [ + 126, + 92, + 241, + 104 + ], + "score": 0.998, + "content": "Conference Proceedings,2011.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 109, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 506, + 122 + ], + "score": 0.968, + "content": "[14] I. Goodfelow,D. Warde-Farley,M. Mirza,A. Courvile,and Y.Bengio. Maxout networks.In International", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 118, + 366, + 132 + ], + "spans": [ + { + "bbox": [ + 123, + 118, + 366, + 132 + ], + "score": 0.994, + "content": "conference on machine learning, pages 1319-1327.PMLR,2013.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 135, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 135, + 506, + 150 + ], + "score": 0.982, + "content": "[15]K. He, X. Zhang, S. Ren,and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 145, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 124, + 145, + 506, + 160 + ], + "score": 0.968, + "content": " imagenet classification. In Proceedings ofthe IEEE international conferenceon computer vision, pages", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 156, + 192, + 165 + ], + "spans": [ + { + "bbox": [ + 126, + 156, + 192, + 165 + ], + "score": 0.997, + "content": "1026-1034,2015.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 185 + ], + "score": 0.985, + "content": "[16] L.Holdijk,M. Boon,S.Henckens,and L.de Jong. [re] parameterized explainer for graph neural network.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 181, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 125, + 181, + 294, + 195 + ], + "score": 0.957, + "content": "In ML Reproducibility Challenge 2020,2021.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 200, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 506, + 211 + ], + "score": 0.963, + "content": "[17] W. Huber, V.J. Carey,L.Long, S. Falcon,and R. Gentleman. Graphs in molecular biology. BMC", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 210, + 247, + 221 + ], + "spans": [ + { + "bbox": [ + 126, + 210, + 247, + 221 + ], + "score": 0.998, + "content": "bioinformatics,8(6):1-14,2007.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "score": 0.98, + "content": "[18] R. K. Iyer and J. A. Bilmes. Submodular optimization with submodular cover and submodular knapsack", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 236, + 374, + 248 + ], + "spans": [ + { + "bbox": [ + 126, + 236, + 374, + 248 + ], + "score": 0.974, + "content": "constraints.In Advances in Neural Information Processing Systems.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "score": 0.969, + "content": "[19]M. Jiang,Z. Li, S. Zhang,S. Wang,X. Wang, Q. Yuan,and Z. Wei. Drug-target affinity prediction using", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 123, + 262, + 434, + 276 + ], + "spans": [ + { + "bbox": [ + 123, + 262, + 434, + 276 + ], + "score": 0.981, + "content": "graph neural network and contact maps.RsC Advances,10(35):20701-20712, 2020.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "score": 0.977, + "content": "[20] H. Jin and X. Zhang. Latent adversarial training of graph convolution networks. In ICML Workshop on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 290, + 385, + 302 + ], + "spans": [ + { + "bbox": [ + 125, + 290, + 385, + 302 + ], + "score": 0.988, + "content": "Learning and Reasoning with Graph-Structured Representations,2019.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 307, + 507, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 507, + 321 + ], + "score": 0.975, + "content": "[21] J. Kazius,R. McGuire,and R. Bursi. Derivation and validation of toxicophores for mutagenicity prediction.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 125, + 317, + 323, + 329 + ], + "spans": [ + { + "bbox": [ + 125, + 317, + 323, + 329 + ], + "score": 0.989, + "content": "Journal of medicinal chemistry,48(1):312-320,2005.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "score": 0.976, + "content": "[22] T.N. Kipf and M. Welling. Semi-supervised classfication with graph convolutional networks. In 5th", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 343, + 487, + 356 + ], + "spans": [ + { + "bbox": [ + 125, + 343, + 487, + 356 + ], + "score": 0.976, + "content": "International Conference on Learning Representations,ICLR 2Ol7, Conference Track Proceedings.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 361, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 506, + 373 + ], + "score": 0.964, + "content": "[23]M. Liu, Y. Luo,L. Wang, Y. Xie, H. Yuan,S. Gui, H. Yu,Z. Xu, J. Zhang, Y. Liu, K. Yan, H. Liu, C. Fu,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 371, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 126, + 371, + 506, + 382 + ], + "score": 0.976, + "content": "B. Oztekin,X. Zhang,and S.Ji. DIG: A turnkey library for diving into graph deep learning research. arXiv", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 123, + 379, + 251, + 392 + ], + "spans": [ + { + "bbox": [ + 123, + 379, + 251, + 392 + ], + "score": 0.999, + "content": "preprint arXiv:2103.12608,2021.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 398, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 506, + 409 + ], + "score": 0.98, + "content": "[24] A.Lucic,M.ter Hoeve,G. Tolomei,M.de Rijke,and F. Silvestri. Cf-gnnexplainer: Counterfactual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 407, + 419, + 419 + ], + "spans": [ + { + "bbox": [ + 126, + 407, + 419, + 419 + ], + "score": 0.99, + "content": "explanations for graph neural networks.arXiv preprint arXiv:2102.03322,2021.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 0.981, + "content": "[25] D.Luo, W. Cheng,D. Xu, W. Yu, B. Zong, H. Chen,and X. Zhang. Parameterized explainer for graph", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 124, + 432, + 389, + 447 + ], + "spans": [ + { + "bbox": [ + 124, + 432, + 389, + 447 + ], + "score": 0.983, + "content": "neural network. In Advances in Neural Information Processing Systems.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 448, + 507, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 507, + 465 + ], + "score": 0.989, + "content": "[26] C. Molnar.Interpretable Machine Learning.2019.https://christophm.github.io/", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 461, + 237, + 473 + ], + "spans": [ + { + "bbox": [ + 126, + 461, + 237, + 473 + ], + "score": 1.0, + "content": "interpretable-ml-book/.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 478, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 506, + 489 + ], + "score": 0.988, + "content": "[27] G.Montufar,R.Pascanu,K.Cho,and Y.Bengio.On the numberof linear regions of deep neural networks.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 487, + 268, + 498 + ], + "spans": [ + { + "bbox": [ + 125, + 487, + 268, + 498 + ], + "score": 0.997, + "content": "arXiv preprint arXiv:1402.1869,2014.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 502, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 104, + 502, + 507, + 519 + ], + "score": 0.954, + "content": "[28]R.Moraffah,M. Karami,R.Guo,A. Raglin,and H.Liu.Causal interpretability for machine learning-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 124, + 514, + 475, + 527 + ], + "spans": [ + { + "bbox": [ + 124, + 514, + 475, + 527 + ], + "score": 0.992, + "content": "problems,methods and evaluation. ACM SIGKDD Explorations Newsleter,22(1):18-33,2020.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 532, + 498, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 498, + 543 + ], + "score": 0.994, + "content": "[29] V.Nair and G.E.Hinton.Rectified linear units improve restricted boltzmann machines.In Icml,2010.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 547, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 562 + ], + "score": 0.986, + "content": "[30] D. Numeroso and D. Bacciu. Meg: Generating molecular counterfactual explanations for deep graph", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 559, + 312, + 570 + ], + "spans": [ + { + "bbox": [ + 124, + 559, + 312, + 570 + ], + "score": 0.988, + "content": "networks. arXiv preprint arXiv:2104.08060,2021.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 589 + ], + "score": 0.971, + "content": "[31] D.E. Pires,T. L. Blundel, and D. B. Ascher. pkcsm: predicting small-molecule pharmacokinetic and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 124, + 585, + 507, + 598 + ], + "spans": [ + { + "bbox": [ + 124, + 585, + 507, + 598 + ], + "score": 0.992, + "content": "toxicity properties using graph-based signatures. Journal of medicinal chemistry,58(9):4066-4072,2015.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 603, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 506, + 614 + ], + "score": 0.976, + "content": "[32]P.E.Pope,S. Kolouri, M.Rostami, C.E. Martin,and H. Hoffmann.Explainability methods for graph", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 126, + 613, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 126, + 613, + 506, + 624 + ], + "score": 0.991, + "content": "convolutional neural networks.In Proceedings of the IEEE/CVF Conference on Computer Vision and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 125, + 621, + 303, + 635 + ], + "spans": [ + { + "bbox": [ + 125, + 621, + 303, + 635 + ], + "score": 0.991, + "content": "Pattern Recognition, pages 10772-10781,2019.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 639, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 650 + ], + "score": 0.988, + "content": "[33]R.R.Selvaraju,M.Cogswell,A.Das,R. Vedantam,D.Parikh,and D.Batra. Grad-cam: Visual explanations", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 126, + 649, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 126, + 649, + 506, + 660 + ], + "score": 0.977, + "content": "from deep networks via gradient-based localization. In Proceedings of the IEEE international conference", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 125, + 658, + 282, + 672 + ], + "spans": [ + { + "bbox": [ + 125, + 658, + 282, + 672 + ], + "score": 0.993, + "content": "on computer vision, pages 618-626, 2017.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 674, + 506, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 506, + 688 + ], + "score": 0.984, + "content": "[34] P.L. Simon, M.Taylor,and I. Z. Kiss. Exact epidemic models on graphs using graph-automorphism driven", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 125, + 685, + 365, + 697 + ], + "spans": [ + { + "bbox": [ + 125, + 685, + 365, + 697 + ], + "score": 0.991, + "content": "lumping. Journal of mathematical biology,62(4):479-508,2011.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 700, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 716 + ], + "score": 0.987, + "content": "[35] K. Simonyan, A. Vedaldi,and A. Zisserman. Deep inside convolutional networks: Visualising image", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 126, + 712, + 322, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 712, + 322, + 724 + ], + "score": 0.992, + "content": "classification models and saliency maps. ICLR,2014.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 24.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 0.999, + "content": "11", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 66, + 507, + 726 + ], + "lines": [], + "index": 24.5, + "bbox_fs": [ + 104, + 73, + 507, + 724 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 73, + 507, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 86 + ], + "score": 0.986, + "content": "[36]K.Sokol and P.A.Flach. Counterfactual explanations of machine learning predictions: opportunities and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 84, + 305, + 95 + ], + "spans": [ + { + "bbox": [ + 126, + 84, + 305, + 95 + ], + "score": 0.963, + "content": "challenges for ai safety. In SafeAI@ AAAI, 2019.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 101, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 505, + 112 + ], + "score": 0.963, + "content": "[37]P. Velickovic, G. Cucurul,A. Casanova,A. Romero,P.Lio,and Y. Bengio. Graph attention networks. In", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 110, + 501, + 123 + ], + "spans": [ + { + "bbox": [ + 125, + 110, + 501, + 123 + ], + "score": 0.977, + "content": "6th International Conference on Learning Representations, ICLR 2O18, Conference Track Proceedings.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 140 + ], + "score": 0.985, + "content": "[38]M. Vu and M. T. Thai.Pgm-explainer: Probabilistic graphical model explanations for graph neural", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 138, + 367, + 149 + ], + "spans": [ + { + "bbox": [ + 126, + 138, + 367, + 149 + ], + "score": 0.973, + "content": "networks.In Advances in Neural Information Processing Systems.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 507, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 507, + 167 + ], + "score": 0.984, + "content": "[39] N. Wale and G. Karypis. Comparison of descriptor spaces for chemical compound retrieval and clas-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 164, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 125, + 164, + 506, + 176 + ], + "score": 0.975, + "content": "sification. In Sixth International Conference on Data Mining (ICDM'06),pages 678-689,2006.doi:", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 174, + 217, + 185 + ], + "spans": [ + { + "bbox": [ + 125, + 174, + 217, + 185 + ], + "score": 0.999, + "content": "10.1109/ICDM.2006.39.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 203 + ], + "score": 0.983, + "content": "[40] L.A.Wolsey. An analysis of the greedy algorithm for the submodular setcovering problem. Combinatorica,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 200, + 203, + 212 + ], + "spans": [ + { + "bbox": [ + 125, + 200, + 203, + 212 + ], + "score": 0.977, + "content": "2(4):385-393,1982.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 217, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 231 + ], + "score": 0.987, + "content": "[41] J. Xiong,Z. Xiong, K. Chen,H. Jiang,and M. Zheng. Graph neural networks for automated de novo drug", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 124, + 227, + 264, + 240 + ], + "spans": [ + { + "bbox": [ + 124, + 227, + 264, + 240 + ], + "score": 0.991, + "content": "design. Drug Discovery Today,2021.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 0.959, + "content": "[42]H. Xu, Y. Ma, H.-C.Liu,D. Deb, H. Liu, J.-L. Tang,and A. K. Jain. Adversarial atacks and defenses in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 254, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 125, + 254, + 506, + 267 + ], + "score": 0.991, + "content": " images, graphs and text: A review. International Journal of Automation and Computing,17(2):151-178,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 263, + 150, + 276 + ], + "spans": [ + { + "bbox": [ + 125, + 263, + 150, + 276 + ], + "score": 0.999, + "content": "2020.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 506, + 294 + ], + "score": 0.966, + "content": "[43] K. Xu, H. Chen,S.Liu,P.-Y. Chen, T.-W. Weng,M. Hong,and X. Lin. Topology attack and defense for", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 124, + 291, + 465, + 304 + ], + "spans": [ + { + "bbox": [ + 124, + 291, + 465, + 304 + ], + "score": 0.982, + "content": " graph neural networks: An optimization perspective. arXiv preprint arXiv:1906.04214, 2019.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 306, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 306, + 507, + 322 + ], + "score": 0.98, + "content": "[44] R. Yin,K.Li, G. Zhang,and J.Lu. A deeper graph neural network for recommender systems. Knowledge-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 318, + 253, + 329 + ], + "spans": [ + { + "bbox": [ + 126, + 318, + 253, + 329 + ], + "score": 0.986, + "content": "Based Systems,185:105020,2019.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "score": 0.98, + "content": "[45] Z. Ying,D. Bourgeois, J. You, M. Zitnik,and J.Leskovec. Gnnexplainer: Generating explanations for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 124, + 344, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 124, + 344, + 480, + 358 + ], + "score": 0.984, + "content": " graph neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "score": 0.983, + "content": "[46] H.Yuan,J.Tang,X.Hu,and S.Ji. Xgnn: Towards model-level explanations of graph neural networks.In", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 124, + 370, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 124, + 370, + 507, + 386 + ], + "score": 0.98, + "content": "Proceedings ofthe 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 124, + 381, + 208, + 393 + ], + "spans": [ + { + "bbox": [ + 124, + 381, + 208, + 393 + ], + "score": 0.995, + "content": "pages 430-438,2020.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "score": 0.967, + "content": "[47] H. Yuan, H. Yu, J. Wang, K. Li, and S.Ji._On explainability of graph neural networks via subgraph", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 408, + 323, + 420 + ], + "spans": [ + { + "bbox": [ + 125, + 408, + 323, + 420 + ], + "score": 0.989, + "content": "explorations. arXiv preprint arXiv:2102.05152,2021.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 424, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 438 + ], + "score": 0.986, + "content": "[48] M.D. Zeiler and R.Fergus. Visualizing and understanding convolutional networks. In European conference", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 124, + 435, + 317, + 447 + ], + "spans": [ + { + "bbox": [ + 124, + 435, + 317, + 447 + ], + "score": 0.978, + "content": "on computer vision, pages 818-833. Springer,2014.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "score": 0.966, + "content": "[49] J. Zhang,S.A. Bargal,Z.Lin,J.Brandt,X. Shen,and S.Sclarof. Top-down neural attention by excitation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 462, + 418, + 475 + ], + "spans": [ + { + "bbox": [ + 125, + 462, + 418, + 475 + ], + "score": 0.993, + "content": "backprop. International Journal of Computer Vision,126(10):1084-1102,2018.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 0.981, + "content": "[50] M. Zhang and Y. Chen. Link prediction based on graph neural networks. In Advances in Neural Information", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 488, + 201, + 502 + ], + "spans": [ + { + "bbox": [ + 126, + 488, + 201, + 502 + ], + "score": 0.999, + "content": "Processing Systems.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 519 + ], + "score": 0.983, + "content": "[51] D. Zugner and S. Gunnemann. Adversarial attacks on graph neural networks via meta learning. arXiv", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 124, + 516, + 251, + 529 + ], + "spans": [ + { + "bbox": [ + 124, + 516, + 251, + 529 + ], + "score": 0.98, + "content": "preprint arXiv:1902.08412, 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 534, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 506, + 545 + ], + "score": 0.978, + "content": "[52] D.Zügner and S.Gunnemann.Certifable robustness and robust training for graph convolutional networks.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 124, + 542, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 124, + 542, + 506, + 556 + ], + "score": 0.99, + "content": "In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 125, + 552, + 239, + 565 + ], + "spans": [ + { + "bbox": [ + 125, + 552, + 239, + 565 + ], + "score": 0.998, + "content": "Mining,pages 246-256,2019.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 0.962, + "content": "[53] D. Zugner, A. Akbarnejad,and S. Gunnemann. Adversarial atacks on neural networks for graph data. In", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 124, + 578, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 124, + 578, + 506, + 593 + ], + "score": 0.987, + "content": "Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 123, + 588, + 217, + 601 + ], + "spans": [ + { + "bbox": [ + 123, + 588, + 217, + 601 + ], + "score": 0.999, + "content": "pages 2847-2856,2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 20 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 73, + 507, + 606 + ], + "lines": [], + "index": 20, + "bbox_fs": [ + 104, + 73, + 507, + 601 + ], + "lines_deleted": true + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_model.json b/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_model.json new file mode 100644 index 0000000000000000000000000000000000000000..664139d1c8082ee8b7729cbd2ea04389be82286e --- /dev/null +++ b/parse/train/Uq_tGs7N54M/Uq_tGs7N54M_model.json @@ -0,0 +1,22429 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 843, + 1303, + 843, + 1303, + 1306, + 398, + 1306 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1425, + 1405, + 1425, + 1405, + 1718, + 298, + 1718 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1730, + 1405, + 1730, + 1405, + 1934, + 298, + 1934 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 366, + 577, + 1334, + 577, + 1334, + 698, + 366, + 698 + ], + "score": 0.973 + }, + { + "category_id": 0, + "poly": [ + 369, + 271, + 1332, + 271, + 1332, + 377, + 369, + 377 + ], + "score": 0.953 + }, + { + "category_id": 0, + "poly": [ + 299, + 1359, + 529, + 1359, + 529, + 1396, + 299, + 1396 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 297, + 2031, + 1070, + 2031, + 1070, + 2060, + 297, + 2060 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 788, + 775, + 912, + 775, + 912, + 812, + 788, + 812 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 334, + 1956, + 535, + 1956, + 535, + 1982, + 334, + 1982 + ], + "score": 0.876 + }, + { + "category_id": 0, + "poly": [ + 494, + 483, + 1207, + 483, + 1207, + 551, + 494, + 551 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 494, + 483, + 1207, + 483, + 1207, + 551, + 494, + 551 + ], + "score": 0.158 + }, + { + "category_id": 13, + "poly": [ + 805, + 485, + 884, + 485, + 884, + 517, + 805, + 517 + ], + "score": 0.46, + "latex": "\\mathbf { C h u ^ { 2 * } }" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 266.0, + 1334.0, + 266.0, + 1334.0, + 332.0, + 361.0, + 332.0 + ], + "score": 0.999, + "text": "Robust Counterfactual Explanations on Graph" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 324.0, + 1030.0, + 324.0, + 1030.0, + 378.0, + 669.0, + 378.0 + ], + "score": 0.999, + "text": "Neural Networks" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1355.0, + 535.0, + 1355.0, + 535.0, + 1402.0, + 292.0, + 1402.0 + ], + "score": 0.969, + "text": "1 Introduction" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2026.0, + 1074.0, + 2026.0, + 1074.0, + 2064.0, + 294.0, + 2064.0 + ], + "score": 0.995, + "text": "35th Conference on Neural Information Processing Systems (NeurIPS 2021)." + }, + { + "category_id": 15, + "poly": [ + 784.0, + 775.0, + 917.0, + 775.0, + 917.0, + 814.0, + 784.0, + 814.0 + ], + "score": 1.0, + "text": "Abstract" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1949.0, + 541.0, + 1949.0, + 541.0, + 1987.0, + 329.0, + 1987.0 + ], + "score": 0.999, + "text": "*Equal contribution." + }, + { + "category_id": 15, + "poly": [ + 490.0, + 478.0, + 804.0, + 478.0, + 804.0, + 523.0, + 490.0, + 523.0 + ], + "score": 0.976, + "text": "Mohit Bajaj1* Lingyang" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 478.0, + 1216.0, + 478.0, + 1216.0, + 523.0, + 885.0, + 523.0 + ], + "score": 0.956, + "text": "Zi Yu Xue1,3Jian Pei4" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 512.0, + 1173.0, + 512.0, + 1173.0, + 557.0, + 522.0, + 557.0 + ], + "score": 0.996, + "text": "Lanjun Wang1Peter Cho-Ho Lam1Yong Zhang1" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 843.0, + 1305.0, + 843.0, + 1305.0, + 878.0, + 394.0, + 878.0 + ], + "score": 0.987, + "text": "Massive deployment of Graph Neural Networks (GNNs) in high-stake applications" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 871.0, + 1306.0, + 871.0, + 1306.0, + 908.0, + 392.0, + 908.0 + ], + "score": 0.998, + "text": " generates a strong demand for explanations that are robust to noise and align well" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 899.0, + 1305.0, + 899.0, + 1305.0, + 938.0, + 392.0, + 938.0 + ], + "score": 0.997, + "text": "with human intuition. Most existing methods generate explanations by identifying" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 932.0, + 1307.0, + 932.0, + 1307.0, + 964.0, + 394.0, + 964.0 + ], + "score": 0.989, + "text": "a subgraph of an input graph that has a strong correlation with the prediction." + }, + { + "category_id": 15, + "poly": [ + 393.0, + 959.0, + 1307.0, + 959.0, + 1307.0, + 994.0, + 393.0, + 994.0 + ], + "score": 0.997, + "text": "These explanations are not robust to noise because independently optimizing the" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 990.0, + 1306.0, + 990.0, + 1306.0, + 1022.0, + 395.0, + 1022.0 + ], + "score": 0.98, + "text": "correlation for a single input can easily overfit noise.Moreover, they are not" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1019.0, + 1306.0, + 1019.0, + 1306.0, + 1051.0, + 395.0, + 1051.0 + ], + "score": 0.997, + "text": "counterfactual because removing an identified subgraph from an input graph does" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1046.0, + 1307.0, + 1046.0, + 1307.0, + 1081.0, + 393.0, + 1081.0 + ], + "score": 0.984, + "text": " not necessarily change the prediction result. In this paper, we propose a novel" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1072.0, + 1306.0, + 1072.0, + 1306.0, + 1110.0, + 393.0, + 1110.0 + ], + "score": 0.985, + "text": "method to generate robust counterfactual explanations on GNNs by explicitly" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1101.0, + 1306.0, + 1101.0, + 1306.0, + 1137.0, + 394.0, + 1137.0 + ], + "score": 0.99, + "text": "modelling the common decision logic of GNNs on similar input graphs. Our" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1131.0, + 1306.0, + 1131.0, + 1306.0, + 1166.0, + 394.0, + 1166.0 + ], + "score": 0.998, + "text": "explanations are naturally robust to noise because they are produced from the" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1162.0, + 1306.0, + 1162.0, + 1306.0, + 1192.0, + 393.0, + 1192.0 + ], + "score": 0.994, + "text": "common decision boundaries of a GNN that govern the predictions of many similar" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1191.0, + 1306.0, + 1191.0, + 1306.0, + 1224.0, + 395.0, + 1224.0 + ], + "score": 0.992, + "text": "input graphs. The explanations are also counterfactual because removing the set" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1217.0, + 1306.0, + 1217.0, + 1306.0, + 1254.0, + 394.0, + 1254.0 + ], + "score": 0.99, + "text": "of edges identified by an explanation from the input graph changes the prediction" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1248.0, + 1307.0, + 1248.0, + 1307.0, + 1281.0, + 393.0, + 1281.0 + ], + "score": 0.995, + "text": "significantly. Exhaustive experiments on many public datasets demonstrate the" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1279.0, + 812.0, + 1279.0, + 812.0, + 1308.0, + 394.0, + 1308.0 + ], + "score": 0.99, + "text": "superior performance of our method." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1427.0, + 1407.0, + 1427.0, + 1407.0, + 1461.0, + 296.0, + 1461.0 + ], + "score": 0.983, + "text": "Graph Neural Networks (GNNs) [22,37,50] have achieved great practical successes in many real-" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1456.0, + 1405.0, + 1456.0, + 1405.0, + 1490.0, + 295.0, + 1490.0 + ], + "score": 0.983, + "text": "world applications,such as chemistry [31],molecular biology[17],social networks [3] and epidemic" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1483.0, + 1407.0, + 1483.0, + 1407.0, + 1521.0, + 293.0, + 1521.0 + ], + "score": 0.986, + "text": "modelling [34]. For most of these applications, explaining predictions made by a GNN model is" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1510.0, + 1407.0, + 1510.0, + 1407.0, + 1549.0, + 292.0, + 1549.0 + ], + "score": 0.988, + "text": "crucial for establishing trust with end-users,identifying the cause of a prediction,and even discovering" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1542.0, + 1406.0, + 1542.0, + 1406.0, + 1574.0, + 292.0, + 1574.0 + ], + "score": 0.992, + "text": "potential deficiencies of a GNN model before massive deployment. Ideally,an explanation should" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1569.0, + 1407.0, + 1569.0, + 1407.0, + 1606.0, + 293.0, + 1606.0 + ], + "score": 0.992, + "text": "be able to answer questions like “Would the prediction of the GNN model change if a certain part" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1599.0, + 1407.0, + 1599.0, + 1407.0, + 1635.0, + 292.0, + 1635.0 + ], + "score": 0.981, + "text": "of an input molecule is removed?” in the context of predicting whether an artificial molecule is" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1661.0, + 296.0, + 1661.0 + ], + "score": 0.981, + "text": "active for acertain type of proteins [19,41],“Would an item recommended stillbe recommended if a" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1692.0, + 292.0, + 1692.0 + ], + "score": 0.988, + "text": "customer had not purchased some other items in the past?” for a GNN built for recommendation" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1688.0, + 479.0, + 1688.0, + 479.0, + 1720.0, + 293.0, + 1720.0 + ], + "score": 0.969, + "text": "systems [9, 44]." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1726.0, + 1410.0, + 1726.0, + 1410.0, + 1767.0, + 293.0, + 1767.0 + ], + "score": 0.988, + "text": "Counterfactual explanations [28] in the form of “If X had not occurred, Y would not have oc-" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1408.0, + 1760.0, + 1408.0, + 1792.0, + 292.0, + 1792.0 + ], + "score": 0.983, + "text": "curred”[26] are the principled way to answer such questions and thus are highly desirable for GNNs." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1787.0, + 1406.0, + 1787.0, + 1406.0, + 1823.0, + 292.0, + 1823.0 + ], + "score": 0.997, + "text": "In the context of GNNs,a counterfactual explanation identifies a small subset of edges of the input" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1817.0, + 1406.0, + 1817.0, + 1406.0, + 1852.0, + 291.0, + 1852.0 + ], + "score": 0.991, + "text": " graph instance such that removing those edges significantly changes the prediction made by the" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1845.0, + 1405.0, + 1845.0, + 1405.0, + 1881.0, + 293.0, + 1881.0 + ], + "score": 0.988, + "text": "GNN. Counterfactual explanations are usually concise and easy to understand [28,36] because they" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1874.0, + 1405.0, + 1874.0, + 1405.0, + 1908.0, + 293.0, + 1908.0 + ], + "score": 0.984, + "text": " align well with the human intuition to describe a causal situation [26]. To make explanations more" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1901.0, + 1405.0, + 1901.0, + 1405.0, + 1940.0, + 292.0, + 1940.0 + ], + "score": 0.986, + "text": " trustworthy, the counterfactual explanation should be robust to noise,that is,some slight changes on" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 575.0, + 918.0, + 575.0, + 918.0, + 615.0, + 463.0, + 615.0 + ], + "score": 0.967, + "text": "1Huawei Technologies Canada Co., Ltd." + }, + { + "category_id": 15, + "poly": [ + 961.0, + 574.0, + 1219.0, + 574.0, + 1219.0, + 616.0, + 961.0, + 616.0 + ], + "score": 0.999, + "text": "2McMaster University" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 603.0, + 906.0, + 603.0, + 906.0, + 649.0, + 477.0, + 649.0 + ], + "score": 0.977, + "text": "3 The University of British Columbia" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 606.0, + 1235.0, + 606.0, + 1235.0, + 646.0, + 943.0, + 646.0 + ], + "score": 0.991, + "text": "4Simon Fraser University" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 639.0, + 1340.0, + 639.0, + 1340.0, + 677.0, + 362.0, + 677.0 + ], + "score": 0.984, + "text": "{mohit.bajaj1，zi.yu.xue，lanjun.wang，cho.ho.lam，yong.zhang3}@huawei.com" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 671.0, + 1064.0, + 671.0, + 1064.0, + 702.0, + 638.0, + 702.0 + ], + "score": 0.999, + "text": "chul9@mcmaster.ca,jpei@cs.sfu.ca" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 478.0, + 804.0, + 478.0, + 804.0, + 523.0, + 490.0, + 523.0 + ], + "score": 0.965, + "text": "Mohit Bajaj1* Lingyang" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 478.0, + 1216.0, + 478.0, + 1216.0, + 523.0, + 885.0, + 523.0 + ], + "score": 0.975, + "text": "Zi Yu Xue1,3Jian Pei4" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 512.0, + 1173.0, + 512.0, + 1173.0, + 557.0, + 522.0, + 557.0 + ], + "score": 0.996, + "text": "Lanjun Wang1Peter Cho-Ho Lam1Yong Zhang1" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 450, + 1404, + 450, + 1404, + 741, + 298, + 741 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1162, + 1403, + 1162, + 1403, + 1367, + 298, + 1367 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 202, + 1404, + 202, + 1404, + 436, + 298, + 436 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1649, + 1404, + 1649, + 1404, + 1883, + 298, + 1883 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 915, + 1404, + 915, + 1404, + 1149, + 298, + 1149 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 756, + 1403, + 756, + 1403, + 902, + 298, + 902 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 301, + 1380, + 1398, + 1380, + 1398, + 1469, + 301, + 1469 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 298, + 1575, + 1400, + 1575, + 1400, + 1637, + 298, + 1637 + ], + "score": 0.953 + }, + { + "category_id": 0, + "poly": [ + 299, + 1509, + 541, + 1509, + 541, + 1546, + 299, + 1546 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 328, + 1978, + 1104, + 1978, + 1104, + 2007, + 328, + 2007 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 303, + 1895, + 1403, + 1895, + 1403, + 1957, + 303, + 1957 + ], + "score": 0.857 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2085, + 841, + 2085 + ], + "score": 0.732 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 859, + 2062, + 859, + 2085, + 841, + 2085 + ], + "score": 0.093 + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1506.0, + 545.0, + 1506.0, + 545.0, + 1551.0, + 291.0, + 1551.0 + ], + "score": 1.0, + "text": "2Related work" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1970.0, + 1106.0, + 1970.0, + 1106.0, + 2014.0, + 328.0, + 2014.0 + ], + "score": 0.98, + "text": "2Other versions of the paper are available at htps://arxiv.org/abs/2107.04086" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2094.0, + 839.0, + 2094.0 + ], + "score": 1.0, + "text": "2" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2094.0, + 838.0, + 2094.0 + ], + "score": 1.0, + "text": "2" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 450.0, + 1406.0, + 450.0, + 1406.0, + 484.0, + 294.0, + 484.0 + ], + "score": 0.99, + "text": "How to produce robust counterfactual explanations on predictions made by general graph neural" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 481.0, + 1405.0, + 481.0, + 1405.0, + 511.0, + 294.0, + 511.0 + ], + "score": 0.99, + "text": "networks is a novel problem that has not been systematically studied before.As to be discussed" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 506.0, + 1406.0, + 506.0, + 1406.0, + 541.0, + 293.0, + 541.0 + ], + "score": 0.986, + "text": " in Section 2, most GNN explanation methods [45,25,46,37,32] are neither counterfactual nor" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 533.0, + 1404.0, + 533.0, + 1404.0, + 572.0, + 292.0, + 572.0 + ], + "score": 0.989, + "text": " robust. These methods mostly focus on identifying a subgraph of an input graph that achieves a high" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 567.0, + 1405.0, + 567.0, + 1405.0, + 598.0, + 296.0, + 598.0 + ], + "score": 0.992, + "text": "correlation with the prediction result. Such explanations are usually not counterfactual because, due" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 593.0, + 1409.0, + 593.0, + 1409.0, + 631.0, + 293.0, + 631.0 + ], + "score": 0.983, + "text": " to the high non-convexity of GNNs, removing a subgraph that achieves a high correlation does not" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 621.0, + 1406.0, + 621.0, + 1406.0, + 658.0, + 292.0, + 658.0 + ], + "score": 0.987, + "text": " necessarily change the prediction result. Moreover, many existing methods [45,25,37,32] are not" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 650.0, + 1406.0, + 650.0, + 1406.0, + 687.0, + 293.0, + 687.0 + ], + "score": 0.991, + "text": "robust to noise and may change significantly upon slight modifications on input graphs, because" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 680.0, + 1406.0, + 680.0, + 1406.0, + 716.0, + 293.0, + 716.0 + ], + "score": 0.987, + "text": " the explanation of every single input graph prediction is independently optimized to maximize the" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 709.0, + 1299.0, + 709.0, + 1299.0, + 743.0, + 294.0, + 743.0 + ], + "score": 0.99, + "text": "correlation with the prediction, thus an explanation can easily overfit the noise in the data." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1162.0, + 1405.0, + 1162.0, + 1405.0, + 1195.0, + 296.0, + 1195.0 + ], + "score": 0.986, + "text": "Second, based on the linear boundaries of the decision region,we propose a novel loss function" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1191.0, + 1408.0, + 1191.0, + 1408.0, + 1225.0, + 291.0, + 1225.0 + ], + "score": 0.992, + "text": "to train a neural network that produces a robust counterfactual explanation as a small subset of" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1223.0, + 1403.0, + 1223.0, + 1403.0, + 1253.0, + 296.0, + 1253.0 + ], + "score": 0.989, + "text": "edges of an input graph. The loss function is designed to directly optimize the explainability and" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1250.0, + 1405.0, + 1250.0, + 1405.0, + 1284.0, + 294.0, + 1284.0 + ], + "score": 0.984, + "text": "counterfactual property of the subset of edges,such that: 1) the subgraph induced by the edges lies" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1280.0, + 1404.0, + 1280.0, + 1404.0, + 1310.0, + 296.0, + 1310.0 + ], + "score": 0.998, + "text": "within the decision region,thus has a prediction consistent with the input graph; and 2) deleting the" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1307.0, + 1407.0, + 1307.0, + 1407.0, + 1341.0, + 294.0, + 1341.0 + ], + "score": 0.992, + "text": " subset of edges from the input graph produces a remainder subgraph that lies outside the decision" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1337.0, + 1142.0, + 1337.0, + 1142.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 0.992, + "text": "region, thus the prediction on the remainder subgraph changes significantly." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 203.0, + 1405.0, + 203.0, + 1405.0, + 235.0, + 294.0, + 235.0 + ], + "score": 0.98, + "text": " an input graph do not change the explanation significantly. This idea aligns well with the notion of" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 233.0, + 1404.0, + 233.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 0.994, + "text": "robustness discussed for DNN explanations in computer vision domain [11]. According to Ghorbani" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 294.0, + 294.0, + 294.0 + ], + "score": 0.985, + "text": "et al.[11] many interpretations on neural networks are fragile as it is easier to generate adversarial" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 289.0, + 1405.0, + 289.0, + 1405.0, + 323.0, + 292.0, + 323.0 + ], + "score": 0.997, + "text": "perturbations that produce perceptively indistinguishable inputs that are assigned the same predicted" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 316.0, + 1405.0, + 316.0, + 1405.0, + 353.0, + 293.0, + 353.0 + ], + "score": 0.982, + "text": "label,yet have very different interpretations.Here,the concepts of “fragile”“robustness”describe" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 346.0, + 1404.0, + 346.0, + 1404.0, + 383.0, + 292.0, + 383.0 + ], + "score": 0.987, + "text": " the same concept from opposite perspectives. An interpretation is said to be fragile if systematic" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 375.0, + 1410.0, + 375.0, + 1410.0, + 416.0, + 293.0, + 416.0 + ], + "score": 0.985, + "text": " perturbations can lead to dramatically different interpretations without changing the label. Otherwise," + }, + { + "category_id": 15, + "poly": [ + 296.0, + 406.0, + 713.0, + 406.0, + 713.0, + 439.0, + 296.0, + 439.0 + ], + "score": 0.974, + "text": "the interpretation is said to be robust." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 1405.0, + 1650.0, + 1405.0, + 1683.0, + 294.0, + 1683.0 + ], + "score": 0.987, + "text": "A model level explanation method [46] produces a high-level explanation about the general behaviors" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1677.0, + 1404.0, + 1677.0, + 1404.0, + 1713.0, + 293.0, + 1713.0 + ], + "score": 0.978, + "text": "of a GNN independent from input examples. This may be achieved by synthesizing a set of artificial" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1706.0, + 1407.0, + 1706.0, + 1407.0, + 1744.0, + 291.0, + 1744.0 + ], + "score": 0.986, + "text": " graph instances such that each artificial graph instance maximizes the prediction score on a certain" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1733.0, + 1406.0, + 1733.0, + 1406.0, + 1771.0, + 293.0, + 1771.0 + ], + "score": 0.985, + "text": "class. The weakness of model level explanation methods is that an input graph instance may not" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1763.0, + 1408.0, + 1763.0, + 1408.0, + 1800.0, + 294.0, + 1800.0 + ], + "score": 0.993, + "text": "contain an artificial graph instance,and removing an artificial graph from an input graph does not" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1792.0, + 1406.0, + 1792.0, + 1406.0, + 1829.0, + 292.0, + 1829.0 + ], + "score": 0.992, + "text": "necessarily change the prediction. As a result, model level explanations are substantially different" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1402.0, + 1821.0, + 1402.0, + 1857.0, + 294.0, + 1857.0 + ], + "score": 0.995, + "text": "from counterfactual explanations, because the synthesized artificial graphs do not provide insights" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 1120.0, + 1849.0, + 1120.0, + 1889.0, + 293.0, + 1889.0 + ], + "score": 0.988, + "text": " into how the GNN makes its prediction on a specific input graph instance." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 917.0, + 1405.0, + 917.0, + 1405.0, + 950.0, + 294.0, + 950.0 + ], + "score": 0.979, + "text": "First, we model the decision logic of a GNN by a set of decision regions, where each decision region" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 942.0, + 1406.0, + 942.0, + 1406.0, + 980.0, + 291.0, + 980.0 + ], + "score": 0.991, + "text": " is induced by a set of linear decision boundaries of the GNN. We propose an unsupervised method to" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 971.0, + 1406.0, + 971.0, + 1406.0, + 1010.0, + 292.0, + 1010.0 + ], + "score": 0.995, + "text": " find decision regions for each class such that each decision region governs the prediction of multiple" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 999.0, + 1407.0, + 999.0, + 1407.0, + 1040.0, + 291.0, + 1040.0 + ], + "score": 0.994, + "text": " graph samples predicted to be the same class. The linear decision boundaries of the decision region" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1031.0, + 1406.0, + 1031.0, + 1406.0, + 1066.0, + 292.0, + 1066.0 + ], + "score": 0.983, + "text": "capture the common decision logic on all the graph instances inside the decision region, thus do" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1059.0, + 1405.0, + 1059.0, + 1405.0, + 1095.0, + 292.0, + 1095.0 + ], + "score": 0.982, + "text": "not easily overfit the noise of an individual graph instance. By exploring the common decision" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1088.0, + 1406.0, + 1088.0, + 1406.0, + 1126.0, + 292.0, + 1126.0 + ], + "score": 0.994, + "text": "logic encoded in the linear boundaries, we are able to produce counterfactual explanations that are" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1117.0, + 592.0, + 1117.0, + 592.0, + 1155.0, + 295.0, + 1155.0 + ], + "score": 0.998, + "text": "inherently robust to noise." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 755.0, + 1405.0, + 755.0, + 1405.0, + 791.0, + 294.0, + 791.0 + ], + "score": 0.986, + "text": "In this paper², we develop RCExplainer, a novel method to produce robust counterfactual explanations" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 783.0, + 1405.0, + 783.0, + 1405.0, + 819.0, + 294.0, + 819.0 + ], + "score": 0.985, + "text": "on GNNs. The key idea is to first model the common decision logic of a GNN by set of decision" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 815.0, + 1407.0, + 815.0, + 1407.0, + 848.0, + 292.0, + 848.0 + ], + "score": 0.996, + "text": "regions where each decision region governs the predictions on a large number of graphs,and then" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 841.0, + 1407.0, + 841.0, + 1407.0, + 882.0, + 292.0, + 882.0 + ], + "score": 0.987, + "text": " extract robust counterfactual explanations by a deep neural network that explores the decision logic" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 871.0, + 1408.0, + 871.0, + 1408.0, + 907.0, + 294.0, + 907.0 + ], + "score": 0.983, + "text": "carried by the linear decision boundaries of the decision regions. We make the following contributions." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1380.0, + 1403.0, + 1380.0, + 1403.0, + 1413.0, + 295.0, + 1413.0 + ], + "score": 0.987, + "text": "Last, we conduct comprehensive experimental study to compare our method with the state-of-the-art" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1409.0, + 1403.0, + 1409.0, + 1403.0, + 1442.0, + 296.0, + 1442.0 + ], + "score": 0.989, + "text": "methods on fidelity, robustness,accuracy and efficiency. All the results solidly demonstrate the" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1439.0, + 728.0, + 1439.0, + 728.0, + 1471.0, + 296.0, + 1471.0 + ], + "score": 0.993, + "text": "superior performance of our approach." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1576.0, + 1404.0, + 1576.0, + 1404.0, + 1611.0, + 294.0, + 1611.0 + ], + "score": 0.985, + "text": "The existing GNN explanation methods [46,37, 45,32, 25] generally fall into two categories: model" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1069.0, + 1605.0, + 1069.0, + 1640.0, + 294.0, + 1640.0 + ], + "score": 0.985, + "text": "level explanation [46] and instance level explanation [37,45,32, 25]." + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1894.0, + 1403.0, + 1894.0, + 1403.0, + 1930.0, + 299.0, + 1930.0 + ], + "score": 0.988, + "text": "The instance level explanation methods [37,45,32, 25] explain the prediction(s) made by a GNN" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1923.0, + 1407.0, + 1923.0, + 1407.0, + 1961.0, + 295.0, + 1961.0 + ], + "score": 0.99, + "text": " on a specific input graph instance or multiple instances by identifying a subgraph of an input graph" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 201, + 1405, + 201, + 1405, + 552, + 298, + 552 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 564, + 1404, + 564, + 1404, + 770, + 297, + 770 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 782, + 1403, + 782, + 1403, + 959, + 298, + 959 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1261, + 1404, + 1261, + 1404, + 1409, + 298, + 1409 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 970, + 1403, + 970, + 1403, + 1118, + 299, + 1118 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1130, + 1403, + 1130, + 1403, + 1248, + 299, + 1248 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1830, + 1405, + 1830, + 1405, + 2007, + 298, + 2007 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1550, + 1403, + 1550, + 1403, + 1645, + 298, + 1645 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 301, + 1656, + 1401, + 1656, + 1401, + 1746, + 301, + 1746 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 298, + 1471, + 659, + 1471, + 659, + 1508, + 298, + 1508 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 298, + 1757, + 1058, + 1757, + 1058, + 1790, + 298, + 1790 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.624 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2084, + 841, + 2084 + ], + "score": 0.346 + }, + { + "category_id": 13, + "poly": [ + 999, + 1614, + 1097, + 1614, + 1097, + 1647, + 999, + 1647 + ], + "score": 0.94, + "latex": "{ \\bf A } _ { i j } = 0" + }, + { + "category_id": 13, + "poly": [ + 371, + 1614, + 468, + 1614, + 468, + 1647, + 371, + 1647 + ], + "score": 0.93, + "latex": "{ \\bf A } _ { i j } = 1" + }, + { + "category_id": 13, + "poly": [ + 417, + 1552, + 557, + 1552, + 557, + 1586, + 417, + 1586 + ], + "score": 0.93, + "latex": "G = \\{ V , E \\}" + }, + { + "category_id": 13, + "poly": [ + 1224, + 1583, + 1397, + 1583, + 1397, + 1617, + 1224, + 1617 + ], + "score": 0.92, + "latex": "\\mathbf { A } \\in \\{ 0 , 1 \\} ^ { n \\times n }" + }, + { + "category_id": 13, + "poly": [ + 721, + 1553, + 965, + 1553, + 965, + 1586, + 721, + 1586 + ], + "score": 0.92, + "latex": "V = \\{ v _ { 1 } , v _ { 2 } , \\ldots , v _ { n } \\}" + }, + { + "category_id": 13, + "poly": [ + 705, + 1863, + 846, + 1863, + 846, + 1894, + 705, + 1894 + ], + "score": 0.92, + "latex": "G = \\{ V , E \\}" + }, + { + "category_id": 13, + "poly": [ + 1239, + 1552, + 1376, + 1552, + 1376, + 1582, + 1239, + 1582 + ], + "score": 0.91, + "latex": "E \\subseteq V \\times V" + }, + { + "category_id": 13, + "poly": [ + 466, + 1891, + 528, + 1891, + 528, + 1922, + 466, + 1922 + ], + "score": 0.9, + "latex": "\\phi ( G )" + }, + { + "category_id": 13, + "poly": [ + 952, + 1895, + 1030, + 1895, + 1030, + 1920, + 952, + 1920 + ], + "score": 0.9, + "latex": "S \\subseteq E" + }, + { + "category_id": 13, + "poly": [ + 912, + 1617, + 940, + 1617, + 940, + 1647, + 912, + 1647 + ], + "score": 0.87, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 835, + 1618, + 861, + 1618, + 861, + 1643, + 835, + 1643 + ], + "score": 0.86, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 1687, + 1129, + 1687, + 1129, + 1717, + 1109, + 1717 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 415, + 1657, + 436, + 1657, + 436, + 1688, + 415, + 1688 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 803, + 1586, + 827, + 1586, + 827, + 1611, + 803, + 1611 + ], + "score": 0.83, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1238, + 1834, + 1258, + 1834, + 1258, + 1863, + 1238, + 1863 + ], + "score": 0.82, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 331, + 1687, + 355, + 1687, + 355, + 1713, + 331, + 1713 + ], + "score": 0.81, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 410, + 1688, + 437, + 1688, + 437, + 1713, + 410, + 1713 + ], + "score": 0.8, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1161, + 1862, + 1186, + 1862, + 1186, + 1888, + 1161, + 1888 + ], + "score": 0.78, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 298, + 1951, + 318, + 1951, + 318, + 1975, + 298, + 1975 + ], + "score": 0.77, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1559, + 1119, + 1559, + 1119, + 1580, + 1099, + 1580 + ], + "score": 0.77, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 476, + 1922, + 501, + 1922, + 501, + 1946, + 476, + 1946 + ], + "score": 0.76, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 451, + 1862, + 477, + 1862, + 477, + 1889, + 451, + 1889 + ], + "score": 0.7, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 394, + 1920, + 415, + 1920, + 415, + 1947, + 394, + 1947 + ], + "score": 0.67, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 1334, + 1950, + 1358, + 1950, + 1358, + 1975, + 1334, + 1975 + ], + "score": 0.66, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1150, + 1892, + 1184, + 1892, + 1184, + 1921, + 1150, + 1921 + ], + "score": 0.38, + "latex": "( l )" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1469.0, + 662.0, + 1469.0, + 662.0, + 1513.0, + 291.0, + 1513.0 + ], + "score": 0.996, + "text": "3Problem Formulation" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 861.0, + 2058.0, + 861.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 0.999, + "text": "3" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "3" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 238.0, + 293.0, + 238.0 + ], + "score": 0.987, + "text": " instance that achieves a high correlation with the prediction on the input graph. GNNExplainer [45]" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 233.0, + 1405.0, + 233.0, + 1405.0, + 265.0, + 295.0, + 265.0 + ], + "score": 0.993, + "text": "removes redundant edges from an input graph instance to produce an explanation that maximizes" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 259.0, + 1406.0, + 259.0, + 1406.0, + 294.0, + 295.0, + 294.0 + ], + "score": 0.996, + "text": "the mutual information between the distribution of subgraphs of the input graph and the GNN's" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 289.0, + 1408.0, + 289.0, + 1408.0, + 325.0, + 295.0, + 325.0 + ], + "score": 0.981, + "text": "prediction. Following the same idea by Ying et al. [45],PGExplainer [25] parameterizes the genera-" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 318.0, + 1406.0, + 318.0, + 1406.0, + 353.0, + 295.0, + 353.0 + ], + "score": 0.993, + "text": "tion process of explanations by a deep neural network,and trains it to maximize a similar mutual" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 347.0, + 1406.0, + 347.0, + 1406.0, + 383.0, + 293.0, + 383.0 + ], + "score": 0.984, + "text": " information based loss used by GNNExplainer [45]. The trained deep neural network is then applied" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 375.0, + 1406.0, + 375.0, + 1406.0, + 412.0, + 292.0, + 412.0 + ], + "score": 0.99, + "text": "to generate explanations for a single input graph instance or a group of input graphs. MEG [30]" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 406.0, + 1406.0, + 406.0, + 1406.0, + 441.0, + 292.0, + 441.0 + ], + "score": 0.99, + "text": " incorporates strong domain knowledge in chemistry with a reinforcement learning framework to" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 434.0, + 1405.0, + 434.0, + 1405.0, + 470.0, + 295.0, + 470.0 + ], + "score": 0.985, + "text": " produce counterfactual explanations on GNNs specifically built for compound prediction, but the" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 462.0, + 1403.0, + 462.0, + 1403.0, + 498.0, + 295.0, + 498.0 + ], + "score": 0.993, + "text": "heavy reliance on domain knowledge largely limits its applicability on general GNNs. The recently" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 491.0, + 1406.0, + 491.0, + 1406.0, + 528.0, + 292.0, + 528.0 + ], + "score": 0.992, + "text": " proposed CF-GNNExplainer [24] independently optimizes the counterfactual property for each" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 520.0, + 1217.0, + 520.0, + 1217.0, + 556.0, + 295.0, + 556.0 + ], + "score": 0.994, + "text": "explanation but ignores the correlation between the prediction and the explanation." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 563.0, + 1406.0, + 563.0, + 1406.0, + 600.0, + 294.0, + 600.0 + ], + "score": 0.997, + "text": "Some studies [32,37] also adapt the existing explanation methods of image-oriented deep neural" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 594.0, + 1406.0, + 594.0, + 1406.0, + 628.0, + 292.0, + 628.0 + ], + "score": 0.996, + "text": "networks to produce instance level explanations for GNNs.Pope et al.[32] extend several gradient" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 620.0, + 1406.0, + 620.0, + 1406.0, + 657.0, + 292.0, + 657.0 + ], + "score": 0.988, + "text": "based methods [33,35,49] to explain predictions made by GNNs. The explanations are prone to" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 650.0, + 1410.0, + 650.0, + 1410.0, + 687.0, + 294.0, + 687.0 + ], + "score": 0.981, + "text": " gradient saturation[12] and may also be misleading [1] due to the heavy reliance on noisy gradients." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 679.0, + 1406.0, + 679.0, + 1406.0, + 715.0, + 294.0, + 715.0 + ], + "score": 0.975, + "text": "Velickovic et al.[37] extend the atention mechanism[7,8] to identify the nodes in an input graph that" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 711.0, + 1405.0, + 711.0, + 1405.0, + 741.0, + 295.0, + 741.0 + ], + "score": 0.995, + "text": "contribute the most to the prediction.This method has to retrain the GNN with the altered architecture" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 739.0, + 1375.0, + 739.0, + 1375.0, + 772.0, + 295.0, + 772.0 + ], + "score": 0.989, + "text": "and the inserted attntion layers.Thus,the explanations may not be faithful to the original GNN." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 784.0, + 1404.0, + 784.0, + 1404.0, + 815.0, + 296.0, + 815.0 + ], + "score": 0.993, + "text": "Instance level explanations from most of the methods are usually not counterfactual because, due" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 812.0, + 1407.0, + 812.0, + 1407.0, + 847.0, + 294.0, + 847.0 + ], + "score": 0.988, + "text": "to the non-convexity of GNNs, removing an explanation subgraph from the input graph does not" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 837.0, + 1407.0, + 837.0, + 1407.0, + 875.0, + 291.0, + 875.0 + ], + "score": 0.988, + "text": " necessarily change the prediction result. Moreover, those methods [45,25,37,32, 24] are usually" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 869.0, + 1405.0, + 869.0, + 1405.0, + 904.0, + 293.0, + 904.0 + ], + "score": 0.99, + "text": "not robust to noise because the explanation of every single input graph prediction is independently" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 895.0, + 1407.0, + 895.0, + 1407.0, + 936.0, + 292.0, + 936.0 + ], + "score": 0.989, + "text": "optimized. Thus,an explanation can easily overfit the noise inside input graphs and may change" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 927.0, + 915.0, + 927.0, + 915.0, + 962.0, + 294.0, + 962.0 + ], + "score": 0.988, + "text": " significantly upon slight modifications on input graphs." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1260.0, + 1405.0, + 1260.0, + 1405.0, + 1296.0, + 294.0, + 1296.0 + ], + "score": 0.988, + "text": "Our method is substantially different from the above works because we focus on explaining the" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1290.0, + 1405.0, + 1290.0, + 1405.0, + 1326.0, + 294.0, + 1326.0 + ], + "score": 0.996, + "text": " prediction by directly optimizing the counterfactual property of an explanation along with correlation" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1317.0, + 1408.0, + 1317.0, + 1408.0, + 1354.0, + 292.0, + 1354.0 + ], + "score": 0.99, + "text": "of the explanation with the prediction. We also require that the explanation is generally valid for a" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1347.0, + 1408.0, + 1347.0, + 1408.0, + 1383.0, + 293.0, + 1383.0 + ], + "score": 0.994, + "text": "large set of similar graph instances by extracting it from the common linear decision boundaries of a" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1374.0, + 542.0, + 1374.0, + 542.0, + 1414.0, + 292.0, + 1414.0 + ], + "score": 0.985, + "text": "large decision region." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 970.0, + 1404.0, + 970.0, + 1404.0, + 1003.0, + 295.0, + 1003.0 + ], + "score": 0.985, + "text": "To tackle the weaknesses in the existing methods, in this paper, we directly optimize the counterfactual" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 999.0, + 1407.0, + 999.0, + 1407.0, + 1034.0, + 293.0, + 1034.0 + ], + "score": 0.996, + "text": "property of an explanation along with the correlation between the explanation and the prediction. Our" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1063.0, + 293.0, + 1063.0 + ], + "score": 0.994, + "text": "explanations are also much more robust to modifications on input graphs,because they are produced" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1057.0, + 1407.0, + 1057.0, + 1407.0, + 1093.0, + 294.0, + 1093.0 + ], + "score": 0.995, + "text": "from the common decision logic on a large group of similar input graphs, which do not easily overfit" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1085.0, + 743.0, + 1085.0, + 743.0, + 1120.0, + 294.0, + 1120.0 + ], + "score": 0.987, + "text": "the noise of an individual graph sample." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1128.0, + 1405.0, + 1128.0, + 1405.0, + 1164.0, + 293.0, + 1164.0 + ], + "score": 0.991, + "text": "Please note that our study is substantially different from adversarial attacks on GNNs.The adversarial" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1159.0, + 1405.0, + 1159.0, + 1405.0, + 1193.0, + 295.0, + 1193.0 + ], + "score": 0.987, + "text": "attacking methods [51,53,42,43,20] use adversarial examples to change the predictions of GNNs but" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1186.0, + 1405.0, + 1186.0, + 1405.0, + 1224.0, + 294.0, + 1224.0 + ], + "score": 0.987, + "text": "ignore the explainability of the generated adversarial examples [1O]. Thus, the adversarial examples" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1215.0, + 1103.0, + 1215.0, + 1103.0, + 1256.0, + 292.0, + 1256.0 + ], + "score": 0.982, + "text": " generated by adversarial attcks may not explain the original prediction." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1831.0, + 1237.0, + 1831.0, + 1237.0, + 1865.0, + 295.0, + 1865.0 + ], + "score": 0.98, + "text": "Definition 1 (Robust Counterfactual Explanation Problem) Given a GNN model" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 1831.0, + 1407.0, + 1831.0, + 1407.0, + 1865.0, + 1259.0, + 1865.0 + ], + "score": 0.98, + "text": "trained on a" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1859.0, + 450.0, + 1859.0, + 450.0, + 1897.0, + 292.0, + 1897.0 + ], + "score": 0.972, + "text": "set of graphs" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1859.0, + 704.0, + 1859.0, + 704.0, + 1897.0, + 478.0, + 1897.0 + ], + "score": 0.962, + "text": "， for an input graph" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1859.0, + 1160.0, + 1859.0, + 1160.0, + 1897.0, + 847.0, + 1897.0 + ], + "score": 0.958, + "text": ", our goal is to explain why" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1859.0, + 1408.0, + 1859.0, + 1408.0, + 1897.0, + 1187.0, + 1897.0 + ], + "score": 0.987, + "text": "is predicted by the" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1886.0, + 465.0, + 1886.0, + 465.0, + 1926.0, + 291.0, + 1926.0 + ], + "score": 0.991, + "text": "GNN model as" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1886.0, + 951.0, + 1886.0, + 951.0, + 1926.0, + 529.0, + 1926.0 + ], + "score": 0.984, + "text": "by identifying a small subset of edges" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1886.0, + 1149.0, + 1886.0, + 1149.0, + 1926.0, + 1031.0, + 1926.0 + ], + "score": 0.947, + "text": ",such that" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1886.0, + 1408.0, + 1886.0, + 1408.0, + 1926.0, + 1185.0, + 1926.0 + ], + "score": 0.97, + "text": "removing the set of" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1918.0, + 393.0, + 1918.0, + 393.0, + 1953.0, + 295.0, + 1953.0 + ], + "score": 0.958, + "text": "edges in" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1918.0, + 475.0, + 1918.0, + 475.0, + 1953.0, + 416.0, + 1953.0 + ], + "score": 1.0, + "text": "from" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1918.0, + 1407.0, + 1918.0, + 1407.0, + 1953.0, + 502.0, + 1953.0 + ], + "score": 0.985, + "text": "that causes the maximum drop in the confidence of the original prediction; and (2)" + }, + { + "category_id": 16, + "poly": [ + 292.0, + 1944.0, + 297.0, + 1944.0, + 297.0, + 1984.0, + 292.0, + 1984.0 + ], + "score": 0.745, + "text": "（20" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1944.0, + 1333.0, + 1944.0, + 1333.0, + 1984.0, + 319.0, + 1984.0 + ], + "score": 0.99, + "text": "is stable and doesn't change when the edges and the feature representations of the nodes of" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1944.0, + 1406.0, + 1944.0, + 1406.0, + 1984.0, + 1359.0, + 1984.0 + ], + "score": 0.999, + "text": "are" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1977.0, + 610.0, + 1977.0, + 610.0, + 2011.0, + 291.0, + 2011.0 + ], + "score": 0.992, + "text": "perturbed by random noise." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1550.0, + 416.0, + 1550.0, + 416.0, + 1588.0, + 294.0, + 1588.0 + ], + "score": 0.999, + "text": "Denote by" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1550.0, + 720.0, + 1550.0, + 720.0, + 1588.0, + 558.0, + 1588.0 + ], + "score": 0.998, + "text": "a graph where" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1550.0, + 1098.0, + 1550.0, + 1098.0, + 1588.0, + 966.0, + 1588.0 + ], + "score": 0.999, + "text": "is the set of" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1550.0, + 1238.0, + 1550.0, + 1238.0, + 1588.0, + 1120.0, + 1588.0 + ], + "score": 1.0, + "text": "nodes and" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1550.0, + 1407.0, + 1550.0, + 1407.0, + 1588.0, + 1377.0, + 1588.0 + ], + "score": 0.999, + "text": "is" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1579.0, + 802.0, + 1579.0, + 802.0, + 1618.0, + 292.0, + 1618.0 + ], + "score": 0.977, + "text": " the set of edges. The edge structure of a graph" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 1579.0, + 1223.0, + 1579.0, + 1223.0, + 1618.0, + 828.0, + 1618.0 + ], + "score": 0.995, + "text": "is described by an adjacency matrix" + }, + { + "category_id": 16, + "poly": [ + 1398.0, + 1579.0, + 1408.0, + 1579.0, + 1408.0, + 1618.0, + 1398.0, + 1618.0 + ], + "score": 0.243, + "text": "，" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1610.0, + 370.0, + 1610.0, + 370.0, + 1649.0, + 293.0, + 1649.0 + ], + "score": 0.999, + "text": "where" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1610.0, + 834.0, + 1610.0, + 834.0, + 1649.0, + 469.0, + 1649.0 + ], + "score": 0.994, + "text": "if there is an edge between node" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1610.0, + 911.0, + 1610.0, + 911.0, + 1649.0, + 862.0, + 1649.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1610.0, + 998.0, + 1610.0, + 998.0, + 1649.0, + 941.0, + 1649.0 + ], + "score": 0.927, + "text": "; and" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1610.0, + 1222.0, + 1610.0, + 1222.0, + 1649.0, + 1098.0, + 1649.0 + ], + "score": 0.996, + "text": "otherwise." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1655.0, + 414.0, + 1655.0, + 414.0, + 1691.0, + 295.0, + 1691.0 + ], + "score": 0.999, + "text": "Denote by" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1691.0, + 437.0, + 1691.0 + ], + "score": 0.991, + "text": "a GNN model that maps a graph to a probability distribution over a set of classes denoted" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1686.0, + 330.0, + 1686.0, + 330.0, + 1719.0, + 295.0, + 1719.0 + ], + "score": 1.0, + "text": "by" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1686.0, + 409.0, + 1686.0, + 409.0, + 1719.0, + 356.0, + 1719.0 + ], + "score": 0.918, + "text": ". Let" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1686.0, + 1108.0, + 1686.0, + 1108.0, + 1719.0, + 438.0, + 1719.0 + ], + "score": 0.977, + "text": "denote the set of graphs that are used to train the GNN model" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1686.0, + 1406.0, + 1686.0, + 1406.0, + 1719.0, + 1130.0, + 1719.0 + ], + "score": 0.968, + "text": ".We focus on GNNs that" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1715.0, + 1405.0, + 1715.0, + 1405.0, + 1748.0, + 296.0, + 1748.0 + ], + "score": 0.989, + "text": "adopt piecewise linear activation functions,such as MaxOut[14] and the family of ReLU[13,15,29]." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1756.0, + 1062.0, + 1756.0, + 1062.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 0.999, + "text": "The robust counterfactual explanation problem is defined as follows." + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1177, + 1405, + 1177, + 1405, + 1556, + 298, + 1556 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 862, + 1404, + 862, + 1404, + 1167, + 297, + 1167 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 607, + 1405, + 607, + 1405, + 853, + 297, + 853 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1669, + 1403, + 1669, + 1403, + 1849, + 298, + 1849 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 371, + 1404, + 371, + 1404, + 519, + 298, + 519 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1568, + 1404, + 1568, + 1404, + 1657, + 298, + 1657 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 294, + 201, + 1404, + 201, + 1404, + 264, + 294, + 264 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 297, + 1946, + 1402, + 1946, + 1402, + 2009, + 297, + 2009 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 680, + 1863, + 1021, + 1863, + 1021, + 1915, + 680, + 1915 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 300, + 552, + 690, + 552, + 690, + 586, + 300, + 586 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 298, + 303, + 462, + 303, + 462, + 340, + 298, + 340 + ], + "score": 0.907 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1868, + 1400, + 1868, + 1400, + 1897, + 1366, + 1897 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.785 + }, + { + "category_id": 13, + "poly": [ + 1196, + 1761, + 1278, + 1761, + 1278, + 1793, + 1196, + 1793 + ], + "score": 0.93, + "latex": "r ( \\mathcal { P } , \\bar { c } )" + }, + { + "category_id": 13, + "poly": [ + 1085, + 819, + 1329, + 819, + 1329, + 854, + 1085, + 854 + ], + "score": 0.93, + "latex": "\\phi ( G ) = \\phi _ { f c } ( \\phi _ { g c } ( G ) )" + }, + { + "category_id": 13, + "poly": [ + 1013, + 1976, + 1096, + 1976, + 1096, + 2009, + 1013, + 2009 + ], + "score": 0.93, + "latex": "g ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 651, + 760, + 804, + 760, + 804, + 795, + 651, + 795 + ], + "score": 0.93, + "latex": "\\phi _ { g c } ( G ) \\in \\mathbb { O } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 597, + 1761, + 682, + 1761, + 682, + 1793, + 597, + 1793 + ], + "score": 0.93, + "latex": "h ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 423, + 1761, + 505, + 1761, + 505, + 1793, + 423, + 1793 + ], + "score": 0.93, + "latex": "r ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 804, + 1134, + 889, + 1134, + 889, + 1168, + 804, + 1168 + ], + "score": 0.93, + "latex": "\\phi _ { g c } ( G )" + }, + { + "category_id": 13, + "poly": [ + 1242, + 1976, + 1324, + 1976, + 1324, + 2009, + 1242, + 2009 + ], + "score": 0.93, + "latex": "r ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 615, + 820, + 677, + 820, + 677, + 853, + 615, + 853 + ], + "score": 0.92, + "latex": "\\phi ( { \\bar { G } } )" + }, + { + "category_id": 13, + "poly": [ + 995, + 1733, + 1079, + 1733, + 1079, + 1764, + 995, + 1764 + ], + "score": 0.92, + "latex": "g ( \\mathcal { P } , c )" + }, + { + "category_id": 14, + "poly": [ + 678, + 1863, + 1021, + 1863, + 1021, + 1917, + 678, + 1917 + ], + "score": 0.92, + "latex": "\\operatorname* { m a x } _ { \\mathcal { P } \\subseteq \\mathcal { H } } g ( \\mathcal { P } , c ) , \\mathrm { ~ s . t . ~ } h ( \\mathcal { P } , c ) = 0" + }, + { + "category_id": 13, + "poly": [ + 965, + 1702, + 1047, + 1702, + 1047, + 1734, + 965, + 1734 + ], + "score": 0.92, + "latex": "r ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 844, + 792, + 929, + 792, + 929, + 823, + 844, + 823 + ], + "score": 0.92, + "latex": "\\phi _ { g c } ( G )" + }, + { + "category_id": 13, + "poly": [ + 419, + 1670, + 517, + 1670, + 517, + 1702, + 419, + 1702 + ], + "score": 0.91, + "latex": "D _ { c } \\subseteq D" + }, + { + "category_id": 13, + "poly": [ + 1126, + 1948, + 1215, + 1948, + 1215, + 1978, + 1126, + 1978 + ], + "score": 0.91, + "latex": "\\mathcal { P } \\subseteq \\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 1289, + 1077, + 1333, + 1077, + 1333, + 1109, + 1289, + 1109 + ], + "score": 0.9, + "latex": "\\phi _ { f c }" + }, + { + "category_id": 13, + "poly": [ + 1000, + 639, + 1038, + 639, + 1038, + 669, + 1000, + 669 + ], + "score": 0.9, + "latex": "\\mathbb { O } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 423, + 729, + 466, + 729, + 466, + 762, + 423, + 762 + ], + "score": 0.9, + "latex": "\\phi _ { g c }" + }, + { + "category_id": 13, + "poly": [ + 1107, + 1671, + 1192, + 1671, + 1192, + 1701, + 1107, + 1701 + ], + "score": 0.9, + "latex": "\\mathcal { P } \\subseteq \\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 422, + 1628, + 492, + 1628, + 492, + 1655, + 422, + 1655 + ], + "score": 0.9, + "latex": "c \\in C" + }, + { + "category_id": 13, + "poly": [ + 1139, + 896, + 1183, + 896, + 1183, + 929, + 1139, + 929 + ], + "score": 0.9, + "latex": "\\phi _ { f c }" + }, + { + "category_id": 13, + "poly": [ + 685, + 989, + 728, + 989, + 728, + 1022, + 685, + 1022 + ], + "score": 0.9, + "latex": "\\phi _ { f c }" + }, + { + "category_id": 13, + "poly": [ + 991, + 1672, + 1060, + 1672, + 1060, + 1699, + 991, + 1699 + ], + "score": 0.9, + "latex": "c \\in C" + }, + { + "category_id": 13, + "poly": [ + 1329, + 1791, + 1366, + 1791, + 1366, + 1820, + 1329, + 1820 + ], + "score": 0.89, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 505, + 1701, + 542, + 1701, + 542, + 1730, + 505, + 1730 + ], + "score": 0.89, + "latex": "\\mathbb { O } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 818, + 1733, + 855, + 1733, + 855, + 1762, + 818, + 1762 + ], + "score": 0.89, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1335, + 894, + 1373, + 894, + 1373, + 924, + 1335, + 924 + ], + "score": 0.88, + "latex": "\\mathbb { O } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 469, + 955, + 507, + 955, + 507, + 985, + 469, + 985 + ], + "score": 0.88, + "latex": "\\mathbb { O } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 1228, + 987, + 1266, + 987, + 1266, + 1016, + 1228, + 1016 + ], + "score": 0.87, + "latex": "\\mathbb { O } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 895, + 762, + 939, + 762, + 939, + 793, + 895, + 793 + ], + "score": 0.87, + "latex": "\\phi _ { f c }" + }, + { + "category_id": 13, + "poly": [ + 1364, + 1734, + 1401, + 1734, + 1401, + 1762, + 1364, + 1762 + ], + "score": 0.86, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1108, + 1599, + 1135, + 1599, + 1135, + 1625, + 1108, + 1625 + ], + "score": 0.83, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 1346, + 233, + 1370, + 233, + 1370, + 259, + 1346, + 259 + ], + "score": 0.83, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 298, + 1733, + 322, + 1733, + 322, + 1760, + 298, + 1760 + ], + "score": 0.82, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 1253, + 1107, + 1278, + 1107, + 1278, + 1133, + 1253, + 1133 + ], + "score": 0.81, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 952, + 1762, + 979, + 1762, + 979, + 1788, + 952, + 1788 + ], + "score": 0.81, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 678, + 642, + 696, + 642, + 696, + 668, + 678, + 668 + ], + "score": 0.81, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 840, + 1977, + 866, + 1977, + 866, + 2003, + 840, + 2003 + ], + "score": 0.81, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 1021, + 204, + 1042, + 204, + 1042, + 230, + 1021, + 230 + ], + "score": 0.81, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 366, + 762, + 391, + 762, + 391, + 790, + 366, + 790 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 739, + 1672, + 766, + 1672, + 766, + 1698, + 739, + 1698 + ], + "score": 0.79, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 852, + 1294, + 878, + 1294, + 878, + 1320, + 852, + 1320 + ], + "score": 0.78, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 332, + 989, + 360, + 989, + 360, + 1017, + 332, + 1017 + ], + "score": 0.76, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 969, + 989, + 997, + 989, + 997, + 1017, + 969, + 1017 + ], + "score": 0.76, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 326, + 821, + 350, + 821, + 350, + 847, + 326, + 847 + ], + "score": 0.74, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1271, + 1019, + 1298, + 1019, + 1298, + 1045, + 1271, + 1045 + ], + "score": 0.73, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 503, + 1797, + 518, + 1797, + 518, + 1817, + 503, + 1817 + ], + "score": 0.64, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 547, + 1237, + 574, + 1237, + 574, + 1263, + 547, + 1263 + ], + "score": 0.59, + "latex": "\\mathcal { H }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 548.0, + 694.0, + 548.0, + 694.0, + 591.0, + 293.0, + 591.0 + ], + "score": 0.991, + "text": "4.1Modelling Decision Regions" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 300.0, + 467.0, + 300.0, + 467.0, + 345.0, + 291.0, + 345.0 + ], + "score": 1.0, + "text": "4Method" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "4" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1176.0, + 1407.0, + 1176.0, + 1407.0, + 1212.0, + 293.0, + 1212.0 + ], + "score": 0.991, + "text": "Based on the above insight, we model the decision region for a set of graph instances as a convex" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1207.0, + 1407.0, + 1207.0, + 1407.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 0.99, + "text": "polytope that satisfies the following two properties. First, the decision region should be induced by a" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1236.0, + 546.0, + 1236.0, + 546.0, + 1270.0, + 293.0, + 1270.0 + ], + "score": 0.966, + "text": "subset of the LDBs in" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 1236.0, + 1408.0, + 1236.0, + 1408.0, + 1270.0, + 575.0, + 1270.0 + ], + "score": 0.984, + "text": ".In this way, when we extract counterfactual explanations from the LDBs," + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1264.0, + 1405.0, + 1264.0, + 1405.0, + 1299.0, + 295.0, + 1299.0 + ], + "score": 0.99, + "text": "the explanations are loyal to the real decision logic_of the GNN. Second, the decision region should" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1293.0, + 851.0, + 1293.0, + 851.0, + 1328.0, + 292.0, + 1328.0 + ], + "score": 0.974, + "text": "cover many graph instances in the training dataset" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1293.0, + 1406.0, + 1293.0, + 1406.0, + 1328.0, + 879.0, + 1328.0 + ], + "score": 0.991, + "text": ",and all the covered graphs should be predicted" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1319.0, + 1406.0, + 1319.0, + 1406.0, + 1358.0, + 292.0, + 1358.0 + ], + "score": 0.979, + "text": " as the same class. In this way, the LDBs of the decision region capture the common decision logic on" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1349.0, + 1407.0, + 1349.0, + 1407.0, + 1386.0, + 292.0, + 1386.0 + ], + "score": 0.987, + "text": " all the graphs covered by the decision region. Here, the requirement of covering a larger number of" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1379.0, + 1407.0, + 1379.0, + 1407.0, + 1415.0, + 292.0, + 1415.0 + ], + "score": 0.985, + "text": " graphs ensures that the common decision logic is general,and thus it is lesslikely to overfit the noise" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1409.0, + 1405.0, + 1409.0, + 1405.0, + 1444.0, + 293.0, + 1444.0 + ], + "score": 0.988, + "text": "of an individual graph instance. As a result, the counterfactual explanations extracted from the LDBs" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1438.0, + 1405.0, + 1438.0, + 1405.0, + 1473.0, + 293.0, + 1473.0 + ], + "score": 0.986, + "text": "of the decision region are insensitive to slight changes in the input graphs. Our method can be easily" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1465.0, + 1406.0, + 1465.0, + 1406.0, + 1502.0, + 292.0, + 1502.0 + ], + "score": 0.988, + "text": " generalized to incorporate prediction confidence in the coverage measure,such as considering the" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1495.0, + 1403.0, + 1495.0, + 1403.0, + 1530.0, + 293.0, + 1530.0 + ], + "score": 0.983, + "text": "count of graphs weighted by prediction confidence. To keep our discusson simple, we do not pursue" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1521.0, + 638.0, + 1521.0, + 638.0, + 1561.0, + 294.0, + 1561.0 + ], + "score": 0.984, + "text": "this detail further in the paper." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 861.0, + 1406.0, + 861.0, + 1406.0, + 898.0, + 292.0, + 898.0 + ], + "score": 0.983, + "text": "For the GNNs that adopt piecewise linear activation functions for the hidden neurons, such as" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 890.0, + 1138.0, + 890.0, + 1138.0, + 932.0, + 291.0, + 932.0 + ], + "score": 0.984, + "text": "MaxOut [14] and the family of ReLU [13,15,29], the decision logic of" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 890.0, + 1334.0, + 890.0, + 1334.0, + 932.0, + 1184.0, + 932.0 + ], + "score": 0.965, + "text": " in the space" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 890.0, + 1407.0, + 890.0, + 1407.0, + 932.0, + 1374.0, + 932.0 + ], + "score": 1.0, + "text": "is" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 922.0, + 1406.0, + 922.0, + 1406.0, + 960.0, + 294.0, + 960.0 + ], + "score": 0.993, + "text": "characterized by a piecewise linear decision boundary formed by connected pieces of decision" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 954.0, + 468.0, + 954.0, + 468.0, + 993.0, + 292.0, + 993.0 + ], + "score": 0.996, + "text": "hyperplanes in" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 954.0, + 1406.0, + 954.0, + 1406.0, + 993.0, + 508.0, + 993.0 + ], + "score": 0.99, + "text": "[1]. We call these hyperplanes linear decision boundaries (LDBs), and denote" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 983.0, + 331.0, + 983.0, + 331.0, + 1027.0, + 291.0, + 1027.0 + ], + "score": 0.999, + "text": "by" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 983.0, + 684.0, + 983.0, + 684.0, + 1027.0, + 361.0, + 1027.0 + ], + "score": 0.978, + "text": "the set of LDBs induced by" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 983.0, + 968.0, + 983.0, + 968.0, + 1027.0, + 729.0, + 1027.0 + ], + "score": 0.99, + "text": ". The set of LDBs in" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 983.0, + 1227.0, + 983.0, + 1227.0, + 1027.0, + 998.0, + 1027.0 + ], + "score": 0.966, + "text": " partitions the space" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 983.0, + 1406.0, + 983.0, + 1406.0, + 1027.0, + 1267.0, + 1027.0 + ], + "score": 0.999, + "text": "into a large" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1017.0, + 1270.0, + 1017.0, + 1270.0, + 1052.0, + 292.0, + 1052.0 + ], + "score": 0.985, + "text": "number of convex polytopes. A convex polytope is formed by a subset of LDBs in" + }, + { + "category_id": 15, + "poly": [ + 1299.0, + 1017.0, + 1406.0, + 1017.0, + 1406.0, + 1052.0, + 1299.0, + 1052.0 + ], + "score": 0.905, + "text": ". All the" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1044.0, + 1406.0, + 1044.0, + 1406.0, + 1084.0, + 292.0, + 1084.0 + ], + "score": 0.995, + "text": " graphs whose graph embeddings are contained in the same convex polytope are predicted as the same" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1075.0, + 1288.0, + 1075.0, + 1288.0, + 1111.0, + 295.0, + 1111.0 + ], + "score": 0.997, + "text": "class [4]. Therefore, the LDBs of a convex polytope encode the common decision logic of" + }, + { + "category_id": 15, + "poly": [ + 1334.0, + 1075.0, + 1405.0, + 1075.0, + 1405.0, + 1111.0, + 1334.0, + 1111.0 + ], + "score": 0.975, + "text": "on all" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1104.0, + 1252.0, + 1104.0, + 1252.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 0.982, + "text": " the graphs whose graph embeddings lie within the convex polytope [4]. Here,a graph" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1104.0, + 1406.0, + 1104.0, + 1406.0, + 1140.0, + 1279.0, + 1140.0 + ], + "score": 0.999, + "text": "is covered" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1131.0, + 803.0, + 1131.0, + 803.0, + 1171.0, + 291.0, + 1171.0 + ], + "score": 0.987, + "text": " by a convex polytope if the graph embedding" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1131.0, + 1299.0, + 1131.0, + 1299.0, + 1171.0, + 890.0, + 1171.0 + ], + "score": 0.992, + "text": "is contained in the convex polytope." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 609.0, + 1403.0, + 609.0, + 1403.0, + 642.0, + 294.0, + 642.0 + ], + "score": 0.985, + "text": "Following the routines of many deep neural network explanation methods [33, 48], we extract the" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 639.0, + 677.0, + 639.0, + 677.0, + 675.0, + 292.0, + 675.0 + ], + "score": 0.985, + "text": "decision region of a GNN in the" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 639.0, + 999.0, + 639.0, + 999.0, + 675.0, + 697.0, + 675.0 + ], + "score": 0.981, + "text": "-dimensional output space" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 639.0, + 1408.0, + 639.0, + 1408.0, + 675.0, + 1039.0, + 675.0 + ], + "score": 0.982, + "text": "of the last convolution layer of" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 669.0, + 1406.0, + 669.0, + 1406.0, + 706.0, + 294.0, + 706.0 + ], + "score": 0.989, + "text": " the GNN. The features generated by the last convolution layer are more conceptually meaningful" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 699.0, + 1407.0, + 699.0, + 1407.0, + 732.0, + 294.0, + 732.0 + ], + "score": 0.992, + "text": "and more robust to noise than those raw features of input graphs,such as vertices and edges [52, 2]." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 728.0, + 422.0, + 728.0, + 422.0, + 761.0, + 295.0, + 761.0 + ], + "score": 0.999, + "text": "Denote by" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 728.0, + 1406.0, + 728.0, + 1406.0, + 761.0, + 467.0, + 761.0 + ], + "score": 0.989, + "text": "the mapping function realized by the graph convolution layers that maps an input" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 759.0, + 365.0, + 759.0, + 365.0, + 798.0, + 291.0, + 798.0 + ], + "score": 1.0, + "text": "graph" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 759.0, + 650.0, + 759.0, + 650.0, + 798.0, + 392.0, + 798.0 + ], + "score": 0.999, + "text": "to its graph embedding" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 759.0, + 894.0, + 759.0, + 894.0, + 798.0, + 805.0, + 798.0 + ], + "score": 0.941, + "text": ",and by" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 759.0, + 1407.0, + 759.0, + 1407.0, + 798.0, + 940.0, + 798.0 + ], + "score": 0.99, + "text": "the mapping function realized by the fully" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 789.0, + 843.0, + 789.0, + 843.0, + 830.0, + 291.0, + 830.0 + ], + "score": 0.993, + "text": "connected layers that maps the graph embedding" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 789.0, + 1408.0, + 789.0, + 1408.0, + 830.0, + 930.0, + 830.0 + ], + "score": 0.998, + "text": "to a predicted distribution over the classes" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 818.0, + 325.0, + 818.0, + 325.0, + 859.0, + 291.0, + 859.0 + ], + "score": 0.997, + "text": "in" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 818.0, + 614.0, + 818.0, + 614.0, + 859.0, + 351.0, + 859.0 + ], + "score": 0.994, + "text": ". The overall prediction" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 818.0, + 1084.0, + 818.0, + 1084.0, + 859.0, + 678.0, + 859.0 + ], + "score": 0.973, + "text": "made by the GNN can be writtn as" + }, + { + "category_id": 16, + "poly": [ + 1330.0, + 818.0, + 1340.0, + 818.0, + 1340.0, + 859.0, + 1330.0, + 859.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1669.0, + 418.0, + 1669.0, + 418.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 0.997, + "text": "Denote by" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1669.0, + 738.0, + 1669.0, + 738.0, + 1704.0, + 518.0, + 1704.0 + ], + "score": 0.993, + "text": "the set of graphs in" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1669.0, + 990.0, + 1669.0, + 990.0, + 1704.0, + 767.0, + 1704.0 + ], + "score": 0.993, + "text": "predicted as a class" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1669.0, + 1106.0, + 1669.0, + 1106.0, + 1704.0, + 1061.0, + 1704.0 + ], + "score": 0.98, + "text": ",by" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1669.0, + 1407.0, + 1669.0, + 1407.0, + 1704.0, + 1193.0, + 1704.0 + ], + "score": 0.975, + "text": "a set of LDBs that" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1698.0, + 504.0, + 1698.0, + 504.0, + 1740.0, + 292.0, + 1740.0 + ], + "score": 0.983, + "text": " partition the space" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1698.0, + 964.0, + 1698.0, + 964.0, + 1740.0, + 543.0, + 1740.0 + ], + "score": 0.99, + "text": "into a set of convex polytopes, and by" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1698.0, + 1408.0, + 1698.0, + 1408.0, + 1740.0, + 1048.0, + 1740.0 + ], + "score": 0.992, + "text": " the convex polytope induced by" + }, + { + "category_id": 16, + "poly": [ + 292.0, + 1729.0, + 297.0, + 1729.0, + 297.0, + 1767.0, + 292.0, + 1767.0 + ], + "score": 0.734, + "text": "（20" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1729.0, + 817.0, + 1729.0, + 817.0, + 1767.0, + 323.0, + 1767.0 + ], + "score": 0.986, + "text": " that covers the largest number of graphs in" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1729.0, + 994.0, + 1729.0, + 994.0, + 1767.0, + 856.0, + 1767.0 + ], + "score": 0.98, + "text": ". Denote by" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1729.0, + 1363.0, + 1729.0, + 1363.0, + 1767.0, + 1080.0, + 1767.0 + ], + "score": 0.98, + "text": " the number of graphs in" + }, + { + "category_id": 16, + "poly": [ + 1402.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1767.0, + 1402.0, + 1767.0 + ], + "score": 0.681, + "text": "（204号" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1760.0, + 422.0, + 1760.0, + 422.0, + 1795.0, + 294.0, + 1795.0 + ], + "score": 0.998, + "text": "covered by" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1760.0, + 596.0, + 1760.0, + 596.0, + 1795.0, + 506.0, + 1795.0 + ], + "score": 0.981, + "text": ",and by" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1760.0, + 951.0, + 1760.0, + 951.0, + 1795.0, + 683.0, + 1795.0 + ], + "score": 0.996, + "text": "the number of graphs in" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1760.0, + 1195.0, + 1760.0, + 1195.0, + 1795.0, + 980.0, + 1795.0 + ], + "score": 0.975, + "text": "that are covered by" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1760.0, + 1407.0, + 1760.0, + 1407.0, + 1795.0, + 1279.0, + 1795.0 + ], + "score": 0.996, + "text": "but are not" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1786.0, + 502.0, + 1786.0, + 502.0, + 1828.0, + 292.0, + 1828.0 + ], + "score": 0.997, + "text": "predicted as class" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1786.0, + 1328.0, + 1786.0, + 1328.0, + 1828.0, + 519.0, + 1828.0 + ], + "score": 0.995, + "text": ". We extract a decision region covering a large set of graph instances in" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 1786.0, + 1408.0, + 1786.0, + 1408.0, + 1828.0, + 1367.0, + 1828.0 + ], + "score": 1.0, + "text": "by" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 927.0, + 1818.0, + 927.0, + 1853.0, + 294.0, + 1853.0 + ], + "score": 0.999, + "text": "solving the following constrained optimization problem." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 370.0, + 1405.0, + 370.0, + 1405.0, + 405.0, + 293.0, + 405.0 + ], + "score": 0.983, + "text": "In this section, we first introduce how to extract the common decision logic of a GNN on a large" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 398.0, + 1408.0, + 398.0, + 1408.0, + 436.0, + 292.0, + 436.0 + ], + "score": 0.99, + "text": " set of graphs with the same predicted class. This is achieved by a decision region induced by a" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 428.0, + 1404.0, + 428.0, + 1404.0, + 460.0, + 294.0, + 460.0 + ], + "score": 0.983, + "text": "set of linear decision boundaries of the GNN.Then, based on the linear boundaries of the decision" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 458.0, + 1406.0, + 458.0, + 1406.0, + 491.0, + 292.0, + 491.0 + ], + "score": 0.983, + "text": "region, we propose a novel lossfunction to train a neural network that produces robust counterfactual" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 488.0, + 1376.0, + 488.0, + 1376.0, + 522.0, + 293.0, + 522.0 + ], + "score": 0.987, + "text": "explanations.Last, we discuss the time complexity of our method when generating explanations." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1564.0, + 1405.0, + 1564.0, + 1405.0, + 1604.0, + 292.0, + 1604.0 + ], + "score": 0.986, + "text": " Next, we illustrate how to extract a decision region satisfying the above two requirements. The key" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1595.0, + 1107.0, + 1595.0, + 1107.0, + 1633.0, + 292.0, + 1633.0 + ], + "score": 0.989, + "text": " idea is to find a convex polytope covering a large set of graph instances in" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1595.0, + 1405.0, + 1595.0, + 1405.0, + 1633.0, + 1136.0, + 1633.0 + ], + "score": 0.999, + "text": "that are predicted as the" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1627.0, + 421.0, + 1627.0, + 421.0, + 1657.0, + 293.0, + 1657.0 + ], + "score": 0.986, + "text": "same class" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1627.0, + 504.0, + 1627.0, + 504.0, + 1657.0, + 493.0, + 1657.0 + ], + "score": 0.849, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 202.0, + 1020.0, + 202.0, + 1020.0, + 238.0, + 294.0, + 238.0 + ], + "score": 0.993, + "text": "In the definition,the first requirement requires that the explanation" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 238.0, + 1043.0, + 238.0 + ], + "score": 0.979, + "text": "is counterfactual, and the second" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 231.0, + 1345.0, + 231.0, + 1345.0, + 266.0, + 293.0, + 266.0 + ], + "score": 0.997, + "text": "requirement requires that the explanation is robust to noisy changes on the edges and nodes of" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 231.0, + 1381.0, + 231.0, + 1381.0, + 266.0, + 1371.0, + 266.0 + ], + "score": 0.511, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1948.0, + 1125.0, + 1948.0, + 1125.0, + 1980.0, + 296.0, + 1980.0 + ], + "score": 0.988, + "text": "This formulation realizes the two properties of decision regions because" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1948.0, + 1403.0, + 1948.0, + 1403.0, + 1980.0, + 1216.0, + 1980.0 + ], + "score": 0.971, + "text": "ensures that the" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1975.0, + 839.0, + 1975.0, + 839.0, + 2011.0, + 295.0, + 2011.0 + ], + "score": 0.986, + "text": "decision region is induced by a subset of LDBs in" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1975.0, + 1012.0, + 1975.0, + 1012.0, + 2011.0, + 867.0, + 2011.0 + ], + "score": 0.997, + "text": ",maximizing" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 1975.0, + 1241.0, + 1975.0, + 1241.0, + 2011.0, + 1097.0, + 2011.0 + ], + "score": 1.0, + "text": "requires that" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 1975.0, + 1406.0, + 1975.0, + 1406.0, + 2011.0, + 1325.0, + 2011.0 + ], + "score": 1.0, + "text": "covers" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1669, + 1404, + 1669, + 1404, + 1821, + 297, + 1821 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1289, + 1405, + 1289, + 1405, + 1414, + 297, + 1414 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1830, + 1404, + 1830, + 1404, + 2008, + 297, + 2008 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 590, + 1404, + 590, + 1404, + 684, + 298, + 684 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 453, + 1405, + 453, + 1405, + 577, + 297, + 577 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1141, + 1405, + 1141, + 1405, + 1279, + 297, + 1279 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1526, + 1404, + 1526, + 1404, + 1655, + 297, + 1655 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 296, + 882, + 1406, + 882, + 1406, + 1007, + 296, + 1007 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 275, + 1405, + 275, + 1405, + 366, + 297, + 366 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 609, + 1428, + 1091, + 1428, + 1091, + 1510, + 609, + 1510 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 296, + 201, + 1402, + 201, + 1402, + 263, + 296, + 263 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 546, + 697, + 1153, + 697, + 1153, + 767, + 546, + 767 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 677, + 1068, + 1022, + 1068, + 1022, + 1126, + 677, + 1126 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 526, + 826, + 1170, + 826, + 1170, + 869, + 526, + 869 + ], + "score": 0.935 + }, + { + "category_id": 0, + "poly": [ + 299, + 398, + 698, + 398, + 698, + 433, + 299, + 433 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 297, + 1020, + 1168, + 1020, + 1168, + 1058, + 297, + 1058 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 297, + 780, + 344, + 780, + 344, + 808, + 297, + 808 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1456, + 1400, + 1456, + 1400, + 1486, + 1366, + 1486 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1074, + 1400, + 1074, + 1400, + 1103, + 1366, + 1103 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 831, + 1400, + 831, + 1400, + 861, + 1366, + 861 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 716, + 1400, + 716, + 1400, + 746, + 1366, + 746 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2085, + 841, + 2085 + ], + "score": 0.74 + }, + { + "category_id": 14, + "poly": [ + 609, + 1428, + 1088, + 1428, + 1088, + 1512, + 609, + 1512 + ], + "score": 0.94, + "latex": "h = \\arg \\operatorname* { m i n } _ { h \\in \\tilde { \\mathcal { H } } \\backslash \\mathcal { P } } \\frac { g ( \\mathcal { P } , c ) - g ( \\mathcal { P } \\cup \\{ h \\} , c ) + \\epsilon } { h ( \\mathcal { P } , c ) - h ( \\mathcal { P } \\cup \\{ h \\} , c ) } ," + }, + { + "category_id": 13, + "poly": [ + 298, + 1241, + 435, + 1241, + 435, + 1280, + 298, + 1280 + ], + "score": 0.93, + "latex": "\\boldsymbol { \\delta } \\dot { = } \\boldsymbol { h } ( \\tilde { \\mathcal { H } } , \\boldsymbol { c } )" + }, + { + "category_id": 13, + "poly": [ + 851, + 202, + 987, + 202, + 987, + 236, + 851, + 236 + ], + "score": 0.93, + "latex": "h ( \\mathcal { P } , c ) = 0" + }, + { + "category_id": 13, + "poly": [ + 1224, + 1176, + 1307, + 1176, + 1307, + 1211, + 1224, + 1211 + ], + "score": 0.93, + "latex": "r ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 1203, + 1788, + 1289, + 1788, + 1289, + 1821, + 1203, + 1821 + ], + "score": 0.93, + "latex": "h ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 975, + 1558, + 1059, + 1558, + 1059, + 1591, + 975, + 1591 + ], + "score": 0.93, + "latex": "g ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 298, + 232, + 380, + 232, + 380, + 265, + 298, + 265 + ], + "score": 0.93, + "latex": "r ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 1064, + 275, + 1147, + 275, + 1147, + 309, + 1064, + 309 + ], + "score": 0.93, + "latex": "r ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 1108, + 1558, + 1193, + 1558, + 1193, + 1591, + 1108, + 1591 + ], + "score": 0.92, + "latex": "h ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 604, + 883, + 773, + 883, + 773, + 917, + 604, + 917 + ], + "score": 0.92, + "latex": "\\operatorname* { m a x } _ { 2 } \\bigl ( \\phi _ { f c } ( \\pmb { \\alpha } ) \\bigr )" + }, + { + "category_id": 13, + "poly": [ + 702, + 1788, + 785, + 1788, + 785, + 1821, + 702, + 1821 + ], + "score": 0.92, + "latex": "g ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 431, + 1890, + 514, + 1890, + 514, + 1923, + 431, + 1923 + ], + "score": 0.92, + "latex": "r ( \\mathcal { P } , \\bar { c } )" + }, + { + "category_id": 13, + "poly": [ + 1012, + 1863, + 1094, + 1863, + 1094, + 1894, + 1012, + 1894 + ], + "score": 0.92, + "latex": "r ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 1249, + 1701, + 1335, + 1701, + 1335, + 1734, + 1249, + 1734 + ], + "score": 0.92, + "latex": "h ( \\mathcal { P } , c )" + }, + { + "category_id": 14, + "poly": [ + 674, + 1068, + 1021, + 1068, + 1021, + 1128, + 674, + 1128 + ], + "score": 0.92, + "latex": "\\operatorname* { m a x } _ { \\mathcal { P } \\subseteq \\tilde { \\mathcal { H } } } g ( \\mathcal { P } , c ) , \\mathrm { ~ s . t . ~ } h ( \\mathcal { P } , c ) \\leq \\delta ," + }, + { + "category_id": 13, + "poly": [ + 528, + 1623, + 610, + 1623, + 610, + 1655, + 528, + 1655 + ], + "score": 0.92, + "latex": "r ( \\mathcal { P } , c )" + }, + { + "category_id": 14, + "poly": [ + 544, + 696, + 1151, + 696, + 1151, + 767, + 544, + 767 + ], + "score": 0.92, + "latex": "\\mathbf { w } = \\frac { \\partial \\left( \\operatorname* { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) - \\operatorname* { m a x } _ { 2 } ( \\phi _ { f c } ( \\alpha ) ) \\right) } { \\partial \\alpha } | _ { \\alpha = \\phi _ { g c } ( G ) } ," + }, + { + "category_id": 13, + "poly": [ + 844, + 1173, + 919, + 1173, + 919, + 1211, + 844, + 1211 + ], + "score": 0.92, + "latex": "\\mathcal { H } \\backslash \\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 516, + 971, + 602, + 971, + 602, + 1004, + 516, + 1004 + ], + "score": 0.92, + "latex": "\\tilde { \\mathcal { H } } \\subset \\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 955, + 1667, + 1031, + 1667, + 1031, + 1700, + 955, + 1700 + ], + "score": 0.92, + "latex": "h \\in \\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 1013, + 591, + 1101, + 591, + 1101, + 622, + 1013, + 622 + ], + "score": 0.92, + "latex": "\\mathbf { x } \\in \\mathbb { O } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 297, + 913, + 383, + 913, + 383, + 946, + 297, + 946 + ], + "score": 0.92, + "latex": "\\phi _ { f c } ( \\alpha )" + }, + { + "category_id": 13, + "poly": [ + 831, + 1528, + 914, + 1528, + 914, + 1561, + 831, + 1561 + ], + "score": 0.91, + "latex": "g ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 369, + 882, + 550, + 882, + 550, + 917, + 369, + 917 + ], + "score": 0.91, + "latex": "\\operatorname* { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1701, + 381, + 1701, + 381, + 1734, + 298, + 1734 + ], + "score": 0.91, + "latex": "g ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 373, + 1143, + 446, + 1143, + 446, + 1173, + 373, + 1173 + ], + "score": 0.91, + "latex": "\\delta \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 633, + 1701, + 719, + 1701, + 719, + 1733, + 633, + 1733 + ], + "score": 0.91, + "latex": "\\bar { h } ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 575, + 1560, + 660, + 1560, + 660, + 1591, + 575, + 1591 + ], + "score": 0.91, + "latex": "h ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 678, + 1731, + 761, + 1731, + 761, + 1763, + 678, + 1763 + ], + "score": 0.91, + "latex": "g ( \\mathcal { P } , c )" + }, + { + "category_id": 13, + "poly": [ + 371, + 1528, + 633, + 1528, + 633, + 1562, + 371, + 1562 + ], + "score": 0.91, + "latex": "g ( \\mathcal { P } , c ) - g ( \\mathcal { P } \\cup \\{ h \\} , c )" + }, + { + "category_id": 13, + "poly": [ + 844, + 1954, + 915, + 1954, + 915, + 1976, + 844, + 1976 + ], + "score": 0.91, + "latex": "c \\in C" + }, + { + "category_id": 13, + "poly": [ + 299, + 1558, + 378, + 1558, + 378, + 1591, + 299, + 1591 + ], + "score": 0.91, + "latex": "\\{ h \\} , c \\bar { ) }" + }, + { + "category_id": 13, + "poly": [ + 748, + 590, + 905, + 590, + 905, + 623, + 748, + 623 + ], + "score": 0.91, + "latex": "\\mathbf { w } ^ { \\top } \\mathbf { x } + b = 0" + }, + { + "category_id": 13, + "poly": [ + 1216, + 1528, + 1406, + 1528, + 1406, + 1562, + 1216, + 1562 + ], + "score": 0.9, + "latex": "h ( \\mathcal { P } , c ) - h ( \\mathcal { P } \\cup" + }, + { + "category_id": 13, + "poly": [ + 447, + 1590, + 523, + 1590, + 523, + 1621, + 447, + 1621 + ], + "score": 0.89, + "latex": "h \\in \\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 489, + 1949, + 526, + 1949, + 526, + 1978, + 489, + 1978 + ], + "score": 0.89, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1143, + 1920, + 1180, + 1920, + 1180, + 1949, + 1143, + 1949 + ], + "score": 0.89, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 766, + 306, + 804, + 306, + 804, + 335, + 766, + 335 + ], + "score": 0.89, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1327, + 1833, + 1397, + 1833, + 1397, + 1860, + 1327, + 1860 + ], + "score": 0.89, + "latex": "c \\in C" + }, + { + "category_id": 13, + "poly": [ + 1130, + 335, + 1166, + 335, + 1166, + 364, + 1130, + 364 + ], + "score": 0.88, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 518, + 591, + 556, + 591, + 556, + 621, + 518, + 621 + ], + "score": 0.88, + "latex": "\\mathbb { O } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 527, + 825, + 1169, + 825, + 1169, + 869, + 527, + 869 + ], + "score": 0.88, + "latex": "b = \\mathrm { m a x } _ { 1 } ( \\phi _ { f c } ( \\alpha ) ) - \\mathrm { m a x } _ { 2 } ( \\phi _ { f c } ( \\alpha ) ) - \\mathbf { w } ^ { T } { \\alpha } | _ { \\alpha = \\phi _ { g c } ( G ) } ," + }, + { + "category_id": 13, + "poly": [ + 602, + 204, + 639, + 204, + 639, + 233, + 602, + 233 + ], + "score": 0.88, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 587, + 1173, + 615, + 1173, + 615, + 1206, + 587, + 1206 + ], + "score": 0.87, + "latex": "\\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 584, + 1892, + 620, + 1892, + 620, + 1920, + 584, + 1920 + ], + "score": 0.87, + "latex": "D _ { c }" + }, + { + "category_id": 13, + "poly": [ + 754, + 1020, + 781, + 1020, + 781, + 1053, + 754, + 1053 + ], + "score": 0.87, + "latex": "\\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 467, + 542, + 494, + 542, + 494, + 575, + 467, + 575 + ], + "score": 0.86, + "latex": "\\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 1038, + 1379, + 1064, + 1379, + 1064, + 1410, + 1038, + 1410 + ], + "score": 0.85, + "latex": "\\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 662, + 542, + 687, + 542, + 687, + 575, + 662, + 575 + ], + "score": 0.85, + "latex": "\\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 593, + 1208, + 621, + 1208, + 621, + 1241, + 593, + 1241 + ], + "score": 0.85, + "latex": "\\tilde { \\mathcal { H } }" + }, + { + "category_id": 13, + "poly": [ + 1288, + 1674, + 1311, + 1674, + 1311, + 1699, + 1288, + 1699 + ], + "score": 0.84, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 913, + 943, + 938, + 943, + 938, + 968, + 913, + 968 + ], + "score": 0.84, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 568, + 277, + 593, + 277, + 593, + 303, + 568, + 303 + ], + "score": 0.84, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 880, + 1562, + 903, + 1562, + 903, + 1586, + 880, + 1586 + ], + "score": 0.84, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 1134, + 1530, + 1158, + 1530, + 1158, + 1556, + 1134, + 1556 + ], + "score": 0.83, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 1108, + 485, + 1135, + 485, + 1135, + 512, + 1108, + 512 + ], + "score": 0.83, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 1321, + 306, + 1345, + 306, + 1345, + 333, + 1321, + 333 + ], + "score": 0.83, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 515, + 1026, + 542, + 1026, + 542, + 1052, + 515, + 1052 + ], + "score": 0.82, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 1063, + 1530, + 1083, + 1530, + 1083, + 1556, + 1063, + 1556 + ], + "score": 0.82, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 1376, + 1350, + 1401, + 1350, + 1401, + 1376, + 1376, + 1376 + ], + "score": 0.82, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1144, + 1272, + 1144, + 1272, + 1170, + 1255, + 1170 + ], + "score": 0.82, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 578, + 1594, + 601, + 1594, + 601, + 1620, + 578, + 1620 + ], + "score": 0.82, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 451, + 1789, + 470, + 1789, + 470, + 1814, + 451, + 1814 + ], + "score": 0.81, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1673, + 1236, + 1673, + 1236, + 1699, + 1217, + 1699 + ], + "score": 0.81, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 1088, + 624, + 1113, + 624, + 1113, + 651, + 1088, + 651 + ], + "score": 0.81, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 457, + 653, + 484, + 653, + 484, + 679, + 457, + 679 + ], + "score": 0.81, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 809, + 1560, + 828, + 1560, + 828, + 1585, + 809, + 1585 + ], + "score": 0.81, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 1040, + 1789, + 1060, + 1789, + 1060, + 1815, + 1040, + 1815 + ], + "score": 0.8, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 441, + 1732, + 456, + 1732, + 456, + 1757, + 441, + 1757 + ], + "score": 0.8, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 772, + 915, + 796, + 915, + 796, + 940, + 772, + 940 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 954, + 1383, + 973, + 1383, + 973, + 1410, + 954, + 1410 + ], + "score": 0.8, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 550, + 485, + 577, + 485, + 577, + 512, + 550, + 512 + ], + "score": 0.8, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 968, + 1594, + 987, + 1594, + 987, + 1620, + 968, + 1620 + ], + "score": 0.79, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 523, + 1177, + 550, + 1177, + 550, + 1206, + 523, + 1206 + ], + "score": 0.78, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 782, + 1247, + 798, + 1247, + 798, + 1274, + 782, + 1274 + ], + "score": 0.78, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 298, + 549, + 323, + 549, + 323, + 575, + 298, + 575 + ], + "score": 0.77, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 409, + 624, + 425, + 624, + 425, + 650, + 409, + 650 + ], + "score": 0.77, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 1302, + 914, + 1328, + 914, + 1328, + 940, + 1302, + 940 + ], + "score": 0.77, + "latex": "\\mathcal { H }" + }, + { + "category_id": 13, + "poly": [ + 797, + 1765, + 813, + 1765, + 813, + 1785, + 797, + 1785 + ], + "score": 0.74, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1206, + 1219, + 1221, + 1219, + 1221, + 1240, + 1206, + 1240 + ], + "score": 0.7, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 1231, + 1896, + 1246, + 1896, + 1246, + 1917, + 1231, + 1917 + ], + "score": 0.69, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 731, + 239, + 745, + 239, + 745, + 259, + 731, + 259 + ], + "score": 0.63, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 620, + 1322, + 645, + 1322, + 645, + 1348, + 620, + 1348 + ], + "score": 0.44, + "latex": "\\mathbf { D }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 701.0, + 397.0, + 701.0, + 440.0, + 293.0, + 440.0 + ], + "score": 0.997, + "text": "4.2Extracting Decision Regions" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2058.0, + 861.0, + 2058.0, + 861.0, + 2093.0, + 839.0, + 2093.0 + ], + "score": 0.994, + "text": "5" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1668.0, + 954.0, + 1668.0, + 954.0, + 1707.0, + 292.0, + 1707.0 + ], + "score": 0.982, + "text": "Intuitively, in each iteration,Equation (5) selects a boundary" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1668.0, + 1216.0, + 1668.0, + 1216.0, + 1707.0, + 1032.0, + 1707.0 + ], + "score": 0.981, + "text": " such that adding" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1668.0, + 1287.0, + 1668.0, + 1287.0, + 1707.0, + 1237.0, + 1707.0 + ], + "score": 1.0, + "text": "into" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 1668.0, + 1406.0, + 1668.0, + 1406.0, + 1707.0, + 1312.0, + 1707.0 + ], + "score": 1.0, + "text": "reduces" + }, + { + "category_id": 16, + "poly": [ + 292.0, + 1698.0, + 297.0, + 1698.0, + 297.0, + 1736.0, + 292.0, + 1736.0 + ], + "score": 0.734, + "text": "（20" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1698.0, + 632.0, + 1698.0, + 632.0, + 1736.0, + 382.0, + 1736.0 + ], + "score": 0.982, + "text": " the least and reduces" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1698.0, + 1248.0, + 1698.0, + 1248.0, + 1736.0, + 720.0, + 1736.0 + ], + "score": 0.993, + "text": "the most. In this way,we can quickly reduce" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 1698.0, + 1406.0, + 1698.0, + 1406.0, + 1736.0, + 1336.0, + 1736.0 + ], + "score": 0.992, + "text": "to be" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1728.0, + 440.0, + 1728.0, + 440.0, + 1765.0, + 292.0, + 1765.0 + ], + "score": 0.999, + "text": "smaller than" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1728.0, + 677.0, + 1728.0, + 677.0, + 1765.0, + 457.0, + 1765.0 + ], + "score": 0.999, + "text": "without decreasing" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1728.0, + 1406.0, + 1728.0, + 1406.0, + 1765.0, + 762.0, + 1765.0 + ], + "score": 0.987, + "text": " too much, which produces a good feasible solution to the" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1757.0, + 796.0, + 1757.0, + 796.0, + 1792.0, + 292.0, + 1792.0 + ], + "score": 0.991, + "text": " practical problem.We add a small constant" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1757.0, + 1406.0, + 1757.0, + 1406.0, + 1792.0, + 814.0, + 1792.0 + ], + "score": 0.986, + "text": "to the numerator such that, when there are multiple" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1784.0, + 450.0, + 1784.0, + 450.0, + 1822.0, + 292.0, + 1822.0 + ], + "score": 0.999, + "text": "candidates of" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1784.0, + 701.0, + 1784.0, + 701.0, + 1822.0, + 471.0, + 1822.0 + ], + "score": 0.999, + "text": "that do not decrease" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 1784.0, + 1039.0, + 1784.0, + 1039.0, + 1822.0, + 786.0, + 1822.0 + ], + "score": 0.967, + "text": ", we can still select the" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1784.0, + 1202.0, + 1784.0, + 1202.0, + 1822.0, + 1061.0, + 1822.0 + ], + "score": 0.999, + "text": "that reduces" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1784.0, + 1404.0, + 1784.0, + 1404.0, + 1822.0, + 1290.0, + 1822.0 + ], + "score": 0.96, + "text": "the most." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1289.0, + 1407.0, + 1289.0, + 1407.0, + 1326.0, + 294.0, + 1326.0 + ], + "score": 0.997, + "text": "The problem in Equation (4) can be proven to be a Submodular Cost Submodular Cover (SCSC)" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1319.0, + 619.0, + 1319.0, + 619.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 0.978, + "text": "problem [18] (see Appendix" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 1319.0, + 1406.0, + 1319.0, + 1406.0, + 1355.0, + 646.0, + 1355.0 + ], + "score": 0.976, + "text": "for proof) that is well known to be NP-hard [5]. We adopt a greedy" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1350.0, + 1375.0, + 1350.0, + 1375.0, + 1383.0, + 295.0, + 1383.0 + ], + "score": 0.989, + "text": "boundary selection method to find a good solution to this problem [40]. Specifically, we initialize" + }, + { + "category_id": 16, + "poly": [ + 1402.0, + 1350.0, + 1405.0, + 1350.0, + 1405.0, + 1383.0, + 1402.0, + 1383.0 + ], + "score": 0.688, + "text": "（204" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1382.0, + 953.0, + 1382.0, + 953.0, + 1418.0, + 292.0, + 1418.0 + ], + "score": 0.98, + "text": " as an empty set,and then iteratively select a new boundary" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1382.0, + 1037.0, + 1382.0, + 1037.0, + 1418.0, + 974.0, + 1418.0 + ], + "score": 1.0, + "text": "from" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1382.0, + 1103.0, + 1382.0, + 1103.0, + 1418.0, + 1065.0, + 1418.0 + ], + "score": 1.0, + "text": "by" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1831.0, + 1326.0, + 1831.0, + 1326.0, + 1866.0, + 295.0, + 1866.0 + ], + "score": 0.982, + "text": "We apply a peeling-off strategy to iteratively extract multiple decision regions.For each class" + }, + { + "category_id": 16, + "poly": [ + 1398.0, + 1831.0, + 1407.0, + 1831.0, + 1407.0, + 1866.0, + 1398.0, + 1866.0 + ], + "score": 0.41, + "text": "，" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1859.0, + 1011.0, + 1859.0, + 1011.0, + 1894.0, + 294.0, + 1894.0 + ], + "score": 0.984, + "text": "we first solve the practical problem once to find a decision region" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1859.0, + 1405.0, + 1859.0, + 1405.0, + 1894.0, + 1095.0, + 1894.0 + ], + "score": 0.99, + "text": ", then we remove the graphs" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1888.0, + 430.0, + 1888.0, + 430.0, + 1926.0, + 294.0, + 1926.0 + ], + "score": 0.998, + "text": "covered by" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1888.0, + 583.0, + 1888.0, + 583.0, + 1926.0, + 515.0, + 1926.0 + ], + "score": 1.0, + "text": "from" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1888.0, + 1230.0, + 1888.0, + 1230.0, + 1926.0, + 621.0, + 1926.0 + ], + "score": 0.994, + "text": ". If there are remaining graphs predicted as the class" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1888.0, + 1407.0, + 1888.0, + 1407.0, + 1926.0, + 1247.0, + 1926.0 + ], + "score": 0.94, + "text": ", we continue" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1917.0, + 1142.0, + 1917.0, + 1142.0, + 1955.0, + 292.0, + 1955.0 + ], + "score": 0.984, + "text": " finding the decision regions using the remaining graphs until all the graphs in" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1917.0, + 1407.0, + 1917.0, + 1407.0, + 1955.0, + 1181.0, + 1955.0 + ], + "score": 0.973, + "text": "are removed. When" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1944.0, + 488.0, + 1944.0, + 488.0, + 1984.0, + 292.0, + 1984.0 + ], + "score": 0.999, + "text": "all the graphs in" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1944.0, + 843.0, + 1944.0, + 843.0, + 1984.0, + 527.0, + 1984.0 + ], + "score": 0.976, + "text": "are removed for each class" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1944.0, + 1409.0, + 1944.0, + 1409.0, + 1984.0, + 916.0, + 1984.0 + ], + "score": 0.972, + "text": ", we stop the iteration and return the set of" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1975.0, + 603.0, + 1975.0, + 603.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 0.996, + "text": "decision regions we found." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 590.0, + 517.0, + 590.0, + 517.0, + 627.0, + 292.0, + 627.0 + ], + "score": 0.991, + "text": "A LDB in the space" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 590.0, + 747.0, + 590.0, + 747.0, + 627.0, + 557.0, + 627.0 + ], + "score": 0.996, + "text": "can be written as" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 590.0, + 1012.0, + 590.0, + 1012.0, + 627.0, + 906.0, + 627.0 + ], + "score": 0.968, + "text": ",where is" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 590.0, + 1406.0, + 590.0, + 1406.0, + 627.0, + 1102.0, + 627.0 + ], + "score": 0.979, + "text": "is a variable, w is the basis" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 622.0, + 408.0, + 622.0, + 408.0, + 656.0, + 294.0, + 656.0 + ], + "score": 0.963, + "text": "term, and" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 622.0, + 1087.0, + 622.0, + 1087.0, + 656.0, + 426.0, + 656.0 + ], + "score": 0.985, + "text": "corresponds to the bias.Following [4], for any input graph" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 622.0, + 1404.0, + 622.0, + 1404.0, + 656.0, + 1114.0, + 656.0 + ], + "score": 0.991, + "text": ",a linear boundary can be" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 650.0, + 456.0, + 650.0, + 456.0, + 687.0, + 295.0, + 687.0 + ], + "score": 1.0, + "text": "sampled from" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 650.0, + 649.0, + 650.0, + 649.0, + 687.0, + 485.0, + 687.0 + ], + "score": 0.977, + "text": "by computing" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 454.0, + 1405.0, + 454.0, + 1405.0, + 487.0, + 295.0, + 487.0 + ], + "score": 0.995, + "text": "The optimization problem in Equation (1) is intractable for standard GNNs, mainly because it is" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 483.0, + 549.0, + 483.0, + 549.0, + 516.0, + 295.0, + 516.0 + ], + "score": 0.997, + "text": "impractical to compute" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 483.0, + 1107.0, + 483.0, + 1107.0, + 516.0, + 578.0, + 516.0 + ], + "score": 0.97, + "text": ", all the LDBs of a GNN. The number of LDBs in" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 483.0, + 1405.0, + 483.0, + 1405.0, + 516.0, + 1136.0, + 516.0 + ], + "score": 0.985, + "text": "of a GNN is exponential" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 513.0, + 1405.0, + 513.0, + 1405.0, + 545.0, + 295.0, + 545.0 + ], + "score": 0.991, + "text": "with respect to the number of neurons in the worst case [27].To address this challenge,we substitute" + }, + { + "category_id": 16, + "poly": [ + 293.0, + 544.0, + 297.0, + 544.0, + 297.0, + 579.0, + 293.0, + 579.0 + ], + "score": 0.751, + "text": "（20" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 544.0, + 466.0, + 544.0, + 466.0, + 579.0, + 324.0, + 579.0 + ], + "score": 0.989, + "text": "by a sample" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 544.0, + 661.0, + 544.0, + 661.0, + 579.0, + 495.0, + 579.0 + ], + "score": 0.996, + "text": "of LDBs from" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 544.0, + 699.0, + 544.0, + 699.0, + 579.0, + 688.0, + 579.0 + ], + "score": 0.831, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1141.0, + 372.0, + 1141.0, + 372.0, + 1178.0, + 292.0, + 1178.0 + ], + "score": 1.0, + "text": "where" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1141.0, + 1254.0, + 1141.0, + 1254.0, + 1178.0, + 447.0, + 1178.0 + ], + "score": 0.983, + "text": "is a tolerance parameter to keep this problem feasible. The parameter" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1141.0, + 1406.0, + 1141.0, + 1406.0, + 1178.0, + 1273.0, + 1178.0 + ], + "score": 0.982, + "text": "is required" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1175.0, + 522.0, + 1175.0, + 522.0, + 1212.0, + 292.0, + 1212.0 + ], + "score": 0.998, + "text": "because substituting" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1175.0, + 586.0, + 1175.0, + 586.0, + 1212.0, + 551.0, + 1212.0 + ], + "score": 1.0, + "text": "by" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1175.0, + 843.0, + 1175.0, + 843.0, + 1212.0, + 616.0, + 1212.0 + ], + "score": 0.991, + "text": "ignores the LDBs in" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1175.0, + 1223.0, + 1175.0, + 1223.0, + 1212.0, + 920.0, + 1212.0 + ], + "score": 0.951, + "text": ". Thus, the convex polytope" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1175.0, + 1406.0, + 1175.0, + 1406.0, + 1212.0, + 1308.0, + 1212.0 + ], + "score": 0.944, + "text": " induced" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1210.0, + 592.0, + 1210.0, + 592.0, + 1250.0, + 292.0, + 1250.0 + ], + "score": 0.981, + "text": "by subset of boundaries in" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1210.0, + 1205.0, + 1210.0, + 1205.0, + 1250.0, + 622.0, + 1250.0 + ], + "score": 0.989, + "text": "may contain instances that are not predicted as class" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1210.0, + 1406.0, + 1210.0, + 1406.0, + 1250.0, + 1222.0, + 1250.0 + ], + "score": 0.988, + "text": ". We directly set" + }, + { + "category_id": 16, + "poly": [ + 292.0, + 1243.0, + 297.0, + 1243.0, + 297.0, + 1283.0, + 292.0, + 1283.0 + ], + "score": 0.745, + "text": "（20" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1243.0, + 781.0, + 1243.0, + 781.0, + 1283.0, + 436.0, + 1283.0 + ], + "score": 0.995, + "text": ", which is the smallest value of" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1243.0, + 1261.0, + 1243.0, + 1261.0, + 1283.0, + 799.0, + 1283.0 + ], + "score": 0.999, + "text": "that keeps the practical problem feasible." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1526.0, + 370.0, + 1526.0, + 370.0, + 1563.0, + 294.0, + 1563.0 + ], + "score": 0.999, + "text": "where" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1526.0, + 830.0, + 1526.0, + 830.0, + 1563.0, + 634.0, + 1563.0 + ], + "score": 0.984, + "text": " is the decrease of" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1526.0, + 1062.0, + 1526.0, + 1062.0, + 1563.0, + 915.0, + 1563.0 + ], + "score": 0.964, + "text": "when adding" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1526.0, + 1133.0, + 1526.0, + 1133.0, + 1563.0, + 1084.0, + 1563.0 + ], + "score": 1.0, + "text": "into" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1526.0, + 1215.0, + 1526.0, + 1215.0, + 1563.0, + 1159.0, + 1563.0 + ], + "score": 0.992, + "text": ",and" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1555.0, + 574.0, + 1555.0, + 574.0, + 1594.0, + 379.0, + 1594.0 + ], + "score": 0.999, + "text": "is the decrease of" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1555.0, + 808.0, + 1555.0, + 808.0, + 1594.0, + 661.0, + 1594.0 + ], + "score": 1.0, + "text": "when adding" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 1555.0, + 879.0, + 1555.0, + 879.0, + 1594.0, + 829.0, + 1594.0 + ], + "score": 0.999, + "text": "into" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1555.0, + 974.0, + 1555.0, + 974.0, + 1594.0, + 904.0, + 1594.0 + ], + "score": 0.925, + "text": ".Both" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1555.0, + 1107.0, + 1555.0, + 1107.0, + 1594.0, + 1060.0, + 1594.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1555.0, + 1406.0, + 1555.0, + 1406.0, + 1594.0, + 1194.0, + 1594.0 + ], + "score": 0.999, + "text": "are non-increasing" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1589.0, + 446.0, + 1589.0, + 446.0, + 1630.0, + 292.0, + 1630.0 + ], + "score": 1.0, + "text": "when adding" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1589.0, + 577.0, + 1589.0, + 577.0, + 1630.0, + 524.0, + 1630.0 + ], + "score": 1.0, + "text": "into" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1589.0, + 967.0, + 1589.0, + 967.0, + 1630.0, + 602.0, + 1630.0 + ], + "score": 0.999, + "text": "because adding a new boundary" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1589.0, + 1405.0, + 1589.0, + 1405.0, + 1630.0, + 988.0, + 1630.0 + ], + "score": 0.998, + "text": "may only exclude some graphs from" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1618.0, + 527.0, + 1618.0, + 527.0, + 1659.0, + 292.0, + 1659.0 + ], + "score": 0.999, + "text": "the convex polytope" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1618.0, + 623.0, + 1618.0, + 623.0, + 1659.0, + 611.0, + 1659.0 + ], + "score": 0.535, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 879.0, + 368.0, + 879.0, + 368.0, + 921.0, + 291.0, + 921.0 + ], + "score": 1.0, + "text": "where" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 879.0, + 603.0, + 879.0, + 603.0, + 921.0, + 551.0, + 921.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 879.0, + 1408.0, + 879.0, + 1408.0, + 921.0, + 774.0, + 921.0 + ], + "score": 0.992, + "text": " are the largest and the second largest values in the vector" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 912.0, + 771.0, + 912.0, + 771.0, + 948.0, + 384.0, + 948.0 + ], + "score": 0.996, + "text": ",respectively. Given an input graph" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 912.0, + 1301.0, + 912.0, + 1301.0, + 948.0, + 797.0, + 948.0 + ], + "score": 0.99, + "text": ",Equations (2) and (3) identify one LDB from" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 912.0, + 1409.0, + 912.0, + 1409.0, + 948.0, + 1329.0, + 948.0 + ], + "score": 0.976, + "text": ".Thus," + }, + { + "category_id": 15, + "poly": [ + 294.0, + 942.0, + 912.0, + 942.0, + 912.0, + 975.0, + 294.0, + 975.0 + ], + "score": 0.983, + "text": "we can sample a subset of input graphs uniformly from" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 942.0, + 1408.0, + 942.0, + 1408.0, + 975.0, + 939.0, + 975.0 + ], + "score": 0.99, + "text": ",and use Equations (2) and (3) to derive a" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 976.0, + 515.0, + 976.0, + 515.0, + 1007.0, + 294.0, + 1007.0 + ], + "score": 0.972, + "text": "sample of LDBs as" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 976.0, + 614.0, + 976.0, + 614.0, + 1007.0, + 603.0, + 1007.0 + ], + "score": 0.813, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 273.0, + 567.0, + 273.0, + 567.0, + 310.0, + 295.0, + 310.0 + ], + "score": 0.995, + "text": "Once we find a solution" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 273.0, + 1063.0, + 273.0, + 1063.0, + 310.0, + 594.0, + 310.0 + ], + "score": 0.989, + "text": "to the above problem, the decision region" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 273.0, + 1406.0, + 273.0, + 1406.0, + 310.0, + 1148.0, + 310.0 + ], + "score": 0.983, + "text": "can be easily obtained" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 303.0, + 765.0, + 303.0, + 765.0, + 338.0, + 292.0, + 338.0 + ], + "score": 0.985, + "text": "by first counting the number of graphs in" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 303.0, + 1320.0, + 303.0, + 1320.0, + 338.0, + 805.0, + 338.0 + ], + "score": 0.978, + "text": "covered by each convex polytope induced by" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 303.0, + 1406.0, + 303.0, + 1406.0, + 338.0, + 1346.0, + 338.0 + ], + "score": 0.991, + "text": ",and" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 332.0, + 1129.0, + 332.0, + 1129.0, + 368.0, + 294.0, + 368.0 + ], + "score": 0.992, + "text": " then select the convex polytope that covers the largest number of graphs in" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 332.0, + 1178.0, + 332.0, + 1178.0, + 368.0, + 1167.0, + 368.0 + ], + "score": 0.862, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 197.0, + 601.0, + 197.0, + 601.0, + 240.0, + 291.0, + 240.0 + ], + "score": 0.957, + "text": " a large number of graphs in" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 197.0, + 850.0, + 197.0, + 850.0, + 240.0, + 640.0, + 240.0 + ], + "score": 0.989, + "text": ", and the constraint" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 197.0, + 1406.0, + 197.0, + 1406.0, + 240.0, + 988.0, + 240.0 + ], + "score": 0.982, + "text": "ensures that allthe graphs covered by" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 231.0, + 730.0, + 231.0, + 730.0, + 266.0, + 381.0, + 266.0 + ], + "score": 0.998, + "text": "are predicted as the same class" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 231.0, + 757.0, + 231.0, + 757.0, + 266.0, + 746.0, + 266.0 + ], + "score": 0.807, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1017.0, + 514.0, + 1017.0, + 514.0, + 1064.0, + 292.0, + 1064.0 + ], + "score": 0.978, + "text": "Now, we substitute" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1017.0, + 753.0, + 1017.0, + 753.0, + 1064.0, + 543.0, + 1064.0 + ], + "score": 0.991, + "text": "in Equation (1) by" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1017.0, + 1169.0, + 1017.0, + 1169.0, + 1064.0, + 782.0, + 1064.0 + ], + "score": 0.982, + "text": "to produce the following problem." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 779.0, + 344.0, + 779.0, + 344.0, + 810.0, + 293.0, + 810.0 + ], + "score": 1.0, + "text": "and" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 256, + 1405, + 256, + 1405, + 404, + 297, + 404 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1082, + 1404, + 1082, + 1404, + 1205, + 298, + 1205 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 616, + 1404, + 616, + 1404, + 736, + 297, + 736 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 486, + 1404, + 486, + 1404, + 605, + 297, + 605 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 881, + 1404, + 881, + 1404, + 1003, + 297, + 1003 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1917, + 1405, + 1917, + 1405, + 2008, + 300, + 2008 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1217, + 1403, + 1217, + 1403, + 1340, + 298, + 1340 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1351, + 1405, + 1351, + 1405, + 1442, + 297, + 1442 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 291, + 1637, + 1402, + 1637, + 1402, + 1700, + 291, + 1700 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 290, + 810, + 1404, + 810, + 1404, + 873, + 290, + 873 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 665, + 1530, + 1032, + 1530, + 1032, + 1608, + 665, + 1608 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 294, + 1712, + 1400, + 1712, + 1400, + 1774, + 294, + 1774 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 296, + 1453, + 1398, + 1453, + 1398, + 1515, + 296, + 1515 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 744, + 754, + 954, + 754, + 954, + 793, + 744, + 793 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 298, + 201, + 649, + 201, + 649, + 235, + 298, + 235 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 298, + 1032, + 602, + 1032, + 602, + 1066, + 298, + 1066 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 298, + 434, + 711, + 434, + 711, + 468, + 298, + 468 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 298, + 1785, + 675, + 1785, + 675, + 1816, + 298, + 1816 + ], + "score": 0.922 + }, + { + "category_id": 8, + "poly": [ + 313, + 1829, + 1344, + 1829, + 1344, + 1900, + 313, + 1900 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1554, + 1400, + 1554, + 1400, + 1584, + 1366, + 1584 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1367, + 759, + 1400, + 759, + 1400, + 788, + 1367, + 788 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1839, + 1400, + 1839, + 1400, + 1869, + 1366, + 1869 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2085, + 840, + 2085 + ], + "score": 0.812 + }, + { + "category_id": 13, + "poly": [ + 899, + 1173, + 968, + 1173, + 968, + 1208, + 899, + 1208 + ], + "score": 0.94, + "latex": "G _ { E \\backslash S }" + }, + { + "category_id": 14, + "poly": [ + 664, + 1530, + 1036, + 1530, + 1036, + 1609, + 664, + 1609 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\mathbf { M } _ { i j } ^ { \\prime } = \\left\\{ { 1 - \\mathbf { M } _ { i j } \\quad \\mathrm { ~ i f ~ } \\mathbf { A } _ { i j } = 1 } \\atop { 0 } \\right. } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 517, + 883, + 661, + 883, + 661, + 917, + 517, + 917 + ], + "score": 0.93, + "latex": "G = \\{ V , E \\}" + }, + { + "category_id": 13, + "poly": [ + 1142, + 1742, + 1211, + 1742, + 1211, + 1777, + 1142, + 1777 + ], + "score": 0.93, + "latex": "G _ { E \\backslash S }" + }, + { + "category_id": 13, + "poly": [ + 629, + 1669, + 665, + 1669, + 665, + 1702, + 629, + 1702 + ], + "score": 0.92, + "latex": "G _ { \\theta } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 525, + 1174, + 594, + 1174, + 594, + 1208, + 525, + 1208 + ], + "score": 0.92, + "latex": "G _ { E \\backslash S }" + }, + { + "category_id": 14, + "poly": [ + 744, + 754, + 954, + 754, + 954, + 792, + 744, + 792 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { { \\bf M } _ { i j } = f _ { \\theta } ( { \\bf z } _ { i } , { \\bf z } _ { j } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 795, + 1143, + 858, + 1143, + 858, + 1176, + 795, + 1176 + ], + "score": 0.92, + "latex": "\\phi ( G )" + }, + { + "category_id": 13, + "poly": [ + 846, + 841, + 943, + 841, + 943, + 874, + 846, + 874 + ], + "score": 0.92, + "latex": "{ \\bf A } _ { i j } = 0" + }, + { + "category_id": 13, + "poly": [ + 1273, + 1172, + 1335, + 1172, + 1335, + 1205, + 1273, + 1205 + ], + "score": 0.92, + "latex": "{ \\bar { \\phi } } ( G )" + }, + { + "category_id": 13, + "poly": [ + 518, + 1085, + 655, + 1085, + 655, + 1118, + 518, + 1118 + ], + "score": 0.92, + "latex": "G = ( V , E )" + }, + { + "category_id": 13, + "poly": [ + 740, + 1454, + 777, + 1454, + 777, + 1487, + 740, + 1487 + ], + "score": 0.92, + "latex": "G _ { \\theta } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 613, + 943, + 692, + 943, + 692, + 973, + 613, + 973 + ], + "score": 0.92, + "latex": "S \\subseteq E" + }, + { + "category_id": 13, + "poly": [ + 1031, + 840, + 1134, + 840, + 1134, + 874, + 1031, + 874 + ], + "score": 0.92, + "latex": "\\mathbf { \\check { M } } _ { i j } = 0" + }, + { + "category_id": 13, + "poly": [ + 526, + 1456, + 594, + 1456, + 594, + 1490, + 526, + 1490 + ], + "score": 0.91, + "latex": "G _ { E \\backslash S }" + }, + { + "category_id": 13, + "poly": [ + 462, + 1115, + 522, + 1115, + 522, + 1146, + 462, + 1146 + ], + "score": 0.91, + "latex": "\\phi ( G )" + }, + { + "category_id": 13, + "poly": [ + 967, + 1249, + 1035, + 1249, + 1035, + 1283, + 967, + 1283 + ], + "score": 0.91, + "latex": "G _ { E \\backslash S }" + }, + { + "category_id": 14, + "poly": [ + 316, + 1826, + 1345, + 1826, + 1345, + 1902, + 316, + 1902 + ], + "score": 0.91, + "latex": "\\mathcal { L } ( \\theta ) = \\sum _ { G \\in D } \\left\\{ \\lambda \\mathcal { L } _ { s a m e } ( \\theta , G ) + ( 1 - \\lambda ) \\mathcal { L } _ { o p p } ( \\theta , G ) + \\beta \\mathcal { R } _ { s p a r s e } ( \\theta , G ) + \\mu \\mathcal { R } _ { d i s c r e t e } ( \\theta , G ) \\right\\} ," + }, + { + "category_id": 13, + "poly": [ + 794, + 1741, + 830, + 1741, + 830, + 1775, + 794, + 1775 + ], + "score": 0.91, + "latex": "\\operatorname { \\mathrm { \\Sigma } } _ { G _ { \\theta } ^ { \\prime } } ^ { G \\prime }" + }, + { + "category_id": 13, + "poly": [ + 372, + 811, + 424, + 811, + 424, + 844, + 372, + 844 + ], + "score": 0.9, + "latex": "\\mathbf { M } _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 669, + 1639, + 705, + 1639, + 705, + 1672, + 669, + 1672 + ], + "score": 0.9, + "latex": "G _ { \\theta } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1085, + 872, + 1085, + 872, + 1115, + 788, + 1115 + ], + "score": 0.9, + "latex": "S \\subseteq E" + }, + { + "category_id": 13, + "poly": [ + 800, + 516, + 860, + 516, + 860, + 548, + 800, + 548 + ], + "score": 0.9, + "latex": "\\phi ( G )" + }, + { + "category_id": 13, + "poly": [ + 612, + 1920, + 682, + 1920, + 682, + 1951, + 612, + 1951 + ], + "score": 0.9, + "latex": "\\mu \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 706, + 1742, + 742, + 1742, + 742, + 1772, + 706, + 1772 + ], + "score": 0.89, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1240, + 1383, + 1276, + 1383, + 1276, + 1412, + 1240, + 1412 + ], + "score": 0.89, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 581, + 1640, + 618, + 1640, + 618, + 1669, + 581, + 1669 + ], + "score": 0.89, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1348, + 1713, + 1400, + 1713, + 1400, + 1747, + 1348, + 1747 + ], + "score": 0.89, + "latex": "\\mathbf { M } _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 473, + 1383, + 508, + 1383, + 508, + 1412, + 473, + 1412 + ], + "score": 0.89, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1050, + 1742, + 1091, + 1742, + 1091, + 1772, + 1050, + 1772 + ], + "score": 0.89, + "latex": "G _ { S }" + }, + { + "category_id": 13, + "poly": [ + 541, + 1669, + 577, + 1669, + 577, + 1699, + 541, + 1699 + ], + "score": 0.89, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 953, + 1411, + 990, + 1411, + 990, + 1441, + 953, + 1441 + ], + "score": 0.89, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 876, + 1249, + 916, + 1249, + 916, + 1279, + 876, + 1279 + ], + "score": 0.89, + "latex": "G _ { S }" + }, + { + "category_id": 13, + "poly": [ + 768, + 1978, + 803, + 1978, + 803, + 2007, + 768, + 2007 + ], + "score": 0.88, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 423, + 487, + 452, + 487, + 452, + 519, + 423, + 519 + ], + "score": 0.88, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1144, + 338, + 1144, + 338, + 1173, + 298, + 1173 + ], + "score": 0.88, + "latex": "G _ { S }" + }, + { + "category_id": 13, + "poly": [ + 912, + 1115, + 940, + 1115, + 940, + 1145, + 912, + 1145 + ], + "score": 0.88, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1251, + 1086, + 1280, + 1086, + 1280, + 1116, + 1251, + 1116 + ], + "score": 0.88, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 546, + 1219, + 575, + 1219, + 575, + 1251, + 546, + 1251 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 708, + 1489, + 733, + 1489, + 733, + 1514, + 708, + 1514 + ], + "score": 0.87, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 524, + 1353, + 564, + 1353, + 564, + 1383, + 524, + 1383 + ], + "score": 0.87, + "latex": "G _ { S }" + }, + { + "category_id": 13, + "poly": [ + 709, + 1353, + 746, + 1353, + 746, + 1383, + 709, + 1383 + ], + "score": 0.87, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1306, + 648, + 1335, + 648, + 1335, + 678, + 1306, + 678 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 726, + 844, + 753, + 844, + 753, + 874, + 726, + 874 + ], + "score": 0.87, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 784, + 1488, + 812, + 1488, + 812, + 1518, + 784, + 1518 + ], + "score": 0.87, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1008, + 622, + 1038, + 622, + 1038, + 651, + 1008, + 651 + ], + "score": 0.87, + "latex": "\\mathbf { z } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1184, + 680, + 1212, + 680, + 1212, + 710, + 1184, + 710 + ], + "score": 0.86, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 375, + 679, + 405, + 679, + 405, + 710, + 375, + 710 + ], + "score": 0.86, + "latex": "\\mathbf { z } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1917, + 479, + 1917, + 479, + 1951, + 371, + 1951 + ], + "score": 0.86, + "latex": "\\lambda \\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 893, + 546, + 913, + 546, + 913, + 577, + 893, + 577 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 931, + 623, + 959, + 623, + 959, + 648, + 931, + 648 + ], + "score": 0.86, + "latex": "\\mathbf { z } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 943, + 817, + 969, + 817, + 969, + 841, + 943, + 841 + ], + "score": 0.85, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 720, + 622, + 749, + 622, + 749, + 651, + 720, + 651 + ], + "score": 0.85, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 649, + 845, + 675, + 845, + 675, + 870, + 649, + 870 + ], + "score": 0.85, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 644, + 623, + 670, + 623, + 670, + 647, + 644, + 647 + ], + "score": 0.85, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1106, + 681, + 1132, + 681, + 1132, + 706, + 1106, + 706 + ], + "score": 0.84, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 296, + 680, + 323, + 680, + 323, + 706, + 296, + 706 + ], + "score": 0.84, + "latex": "\\mathbf { z } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1021, + 816, + 1048, + 816, + 1048, + 843, + 1021, + 843 + ], + "score": 0.84, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1298, + 943, + 1323, + 943, + 1323, + 969, + 1298, + 969 + ], + "score": 0.84, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 563, + 1146, + 587, + 1146, + 587, + 1170, + 563, + 1170 + ], + "score": 0.83, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 944, + 1383, + 968, + 1383, + 968, + 1409, + 944, + 1409 + ], + "score": 0.83, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 956, + 943, + 978, + 943, + 978, + 969, + 956, + 969 + ], + "score": 0.83, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 475, + 1147, + 495, + 1147, + 495, + 1170, + 475, + 1170 + ], + "score": 0.83, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 783, + 619, + 807, + 619, + 807, + 645, + 783, + 645 + ], + "score": 0.82, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 490, + 1918, + 562, + 1918, + 562, + 1950, + 490, + 1950 + ], + "score": 0.82, + "latex": "\\beta \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 573, + 1035, + 600, + 1035, + 600, + 1066, + 573, + 1066 + ], + "score": 0.82, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1159, + 1144, + 1181, + 1144, + 1181, + 1170, + 1159, + 1170 + ], + "score": 0.81, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 1186, + 1412, + 1202, + 1412, + 1202, + 1437, + 1186, + 1437 + ], + "score": 0.81, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1197, + 1978, + 1221, + 1978, + 1221, + 2003, + 1197, + 2003 + ], + "score": 0.81, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1222, + 1456, + 1246, + 1456, + 1246, + 1482, + 1222, + 1482 + ], + "score": 0.81, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1223, + 488, + 1249, + 488, + 1249, + 515, + 1223, + 515 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1248, + 1144, + 1272, + 1144, + 1272, + 1170, + 1248, + 1170 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1177, + 884, + 1205, + 884, + 1205, + 915, + 1177, + 915 + ], + "score": 0.8, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 585, + 1281, + 602, + 1281, + 602, + 1307, + 585, + 1307 + ], + "score": 0.79, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 855, + 1411, + 872, + 1411, + 872, + 1437, + 855, + 1437 + ], + "score": 0.79, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 486, + 1220, + 507, + 1220, + 507, + 1247, + 486, + 1247 + ], + "score": 0.78, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 914, + 1640, + 931, + 1640, + 931, + 1666, + 914, + 1666 + ], + "score": 0.78, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1317, + 890, + 1337, + 890, + 1337, + 911, + 1317, + 911 + ], + "score": 0.75, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 724, + 890, + 744, + 890, + 744, + 911, + 724, + 911 + ], + "score": 0.74, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 1381, + 890, + 1401, + 890, + 1401, + 911, + 1381, + 911 + ], + "score": 0.72, + "latex": "\\mathbf { \\nabla } \\cdot n" + }, + { + "category_id": 13, + "poly": [ + 874, + 517, + 892, + 517, + 892, + 543, + 874, + 543 + ], + "score": 0.69, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 600, + 1411, + 626, + 1411, + 626, + 1437, + 600, + 1437 + ], + "score": 0.47, + "latex": "\\mathbf { A }" + }, + { + "category_id": 13, + "poly": [ + 572, + 972, + 606, + 972, + 606, + 998, + 572, + 998 + ], + "score": 0.44, + "latex": "\\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 1183, + 1354, + 1209, + 1354, + 1209, + 1381, + 1183, + 1381 + ], + "score": 0.44, + "latex": "\\mathbf { A }" + }, + { + "category_id": 13, + "poly": [ + 1025, + 1353, + 1060, + 1353, + 1060, + 1381, + 1025, + 1381 + ], + "score": 0.39, + "latex": "\\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 447, + 1411, + 481, + 1411, + 481, + 1438, + 447, + 1438 + ], + "score": 0.32, + "latex": "\\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 1175, + 884, + 1253, + 884, + 1253, + 915, + 1175, + 915 + ], + "score": 0.28, + "latex": "f _ { \\boldsymbol { \\theta } } , \\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 1218, + 884, + 1253, + 884, + 1253, + 913, + 1218, + 913 + ], + "score": 0.28, + "latex": "\\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 978, + 1953, + 1002, + 1953, + 1002, + 1975, + 978, + 1975 + ], + "score": 0.27, + "latex": "\\mathbf { G }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 197.0, + 652.0, + 197.0, + 652.0, + 240.0, + 293.0, + 240.0 + ], + "score": 0.999, + "text": "4.3Producing Explanations" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1025.0, + 572.0, + 1025.0, + 572.0, + 1074.0, + 292.0, + 1074.0 + ], + "score": 0.998, + "text": "4.3.2Training Model" + }, + { + "category_id": 16, + "poly": [ + 601.0, + 1025.0, + 605.0, + 1025.0, + 605.0, + 1074.0, + 601.0, + 1074.0 + ], + "score": 0.691, + "text": "（204号" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 434.0, + 713.0, + 434.0, + 713.0, + 471.0, + 295.0, + 471.0 + ], + "score": 0.974, + "text": "4.3.1The Neural Network Model" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2089.0, + 840.0, + 2089.0 + ], + "score": 1.0, + "text": "6" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 255.0, + 1406.0, + 255.0, + 1406.0, + 291.0, + 294.0, + 291.0 + ], + "score": 0.99, + "text": "In this section, we introduce how to use the LDBs of decision regions to train a neural network that" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 287.0, + 1403.0, + 287.0, + 1403.0, + 320.0, + 295.0, + 320.0 + ], + "score": 0.988, + "text": "produces a robust counterfactual explanation as a small subset of edges of an input graph. We form" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 313.0, + 1405.0, + 313.0, + 1405.0, + 351.0, + 292.0, + 351.0 + ], + "score": 0.994, + "text": " explanations as a subset of edges because GNNs make decisions by aggregating messages passed on" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 343.0, + 1405.0, + 343.0, + 1405.0, + 379.0, + 294.0, + 379.0 + ], + "score": 0.982, + "text": "edges. Using edges instead of vertices as explanations can provide beter insights on the decision" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 372.0, + 474.0, + 372.0, + 474.0, + 406.0, + 292.0, + 406.0 + ], + "score": 0.978, + "text": "logic of GNNs." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1082.0, + 517.0, + 1082.0, + 517.0, + 1120.0, + 292.0, + 1120.0 + ], + "score": 0.999, + "text": "For an input graph" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1082.0, + 787.0, + 1082.0, + 787.0, + 1120.0, + 656.0, + 1120.0 + ], + "score": 0.977, + "text": ", denote by" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1082.0, + 1250.0, + 1082.0, + 1250.0, + 1120.0, + 873.0, + 1120.0 + ], + "score": 0.977, + "text": " the subset of edges produced by" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1082.0, + 1405.0, + 1082.0, + 1405.0, + 1120.0, + 1281.0, + 1120.0 + ], + "score": 0.999, + "text": "to explain" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1114.0, + 461.0, + 1114.0, + 461.0, + 1147.0, + 294.0, + 1147.0 + ], + "score": 0.998, + "text": "the prediction" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1114.0, + 911.0, + 1114.0, + 911.0, + 1147.0, + 523.0, + 1147.0 + ], + "score": 0.993, + "text": ",our goal is to train a good model" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1114.0, + 1404.0, + 1114.0, + 1404.0, + 1147.0, + 941.0, + 1147.0 + ], + "score": 0.992, + "text": "such that the prediction on the subgraph" + }, + { + "category_id": 16, + "poly": [ + 294.0, + 1142.0, + 297.0, + 1142.0, + 297.0, + 1178.0, + 294.0, + 1178.0 + ], + "score": 0.728, + "text": "（204" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1142.0, + 474.0, + 1142.0, + 474.0, + 1178.0, + 339.0, + 1178.0 + ], + "score": 0.998, + "text": "induced by" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1142.0, + 562.0, + 1142.0, + 562.0, + 1178.0, + 496.0, + 1178.0 + ], + "score": 1.0, + "text": "from" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1142.0, + 794.0, + 1142.0, + 794.0, + 1178.0, + 588.0, + 1178.0 + ], + "score": 0.995, + "text": "is consistent with" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 1142.0, + 1158.0, + 1142.0, + 1158.0, + 1178.0, + 859.0, + 1178.0 + ], + "score": 0.974, + "text": "; and deleting the edges in" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1142.0, + 1247.0, + 1142.0, + 1247.0, + 1178.0, + 1182.0, + 1178.0 + ], + "score": 1.0, + "text": "from" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1142.0, + 1409.0, + 1142.0, + 1409.0, + 1178.0, + 1273.0, + 1178.0 + ], + "score": 0.986, + "text": "produces a" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1171.0, + 524.0, + 1171.0, + 524.0, + 1207.0, + 292.0, + 1207.0 + ], + "score": 0.985, + "text": "remainder subgraph" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 1171.0, + 898.0, + 1171.0, + 898.0, + 1207.0, + 595.0, + 1207.0 + ], + "score": 0.971, + "text": "such that the prediction on" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1171.0, + 1272.0, + 1171.0, + 1272.0, + 1207.0, + 969.0, + 1207.0 + ], + "score": 0.998, + "text": "changes significantly from" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 1171.0, + 1346.0, + 1171.0, + 1346.0, + 1207.0, + 1336.0, + 1207.0 + ], + "score": 0.629, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 615.0, + 643.0, + 615.0, + 643.0, + 654.0, + 294.0, + 654.0 + ], + "score": 0.993, + "text": "For any two connected vertices" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 615.0, + 719.0, + 615.0, + 719.0, + 654.0, + 671.0, + 654.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 615.0, + 782.0, + 615.0, + 782.0, + 654.0, + 750.0, + 654.0 + ], + "score": 0.998, + "text": "of" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 615.0, + 930.0, + 615.0, + 930.0, + 654.0, + 808.0, + 654.0 + ], + "score": 0.986, + "text": ",denote by" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 615.0, + 1007.0, + 615.0, + 1007.0, + 654.0, + 960.0, + 654.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 615.0, + 1405.0, + 615.0, + 1405.0, + 654.0, + 1039.0, + 654.0 + ], + "score": 0.987, + "text": "the embeddings produced by the" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 644.0, + 1305.0, + 644.0, + 1305.0, + 682.0, + 292.0, + 682.0 + ], + "score": 0.981, + "text": "last convolution layer of the GNN for the two vertices, respectively. The neural network" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 644.0, + 1404.0, + 644.0, + 1404.0, + 682.0, + 1336.0, + 682.0 + ], + "score": 1.0, + "text": "takes" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 675.0, + 374.0, + 675.0, + 374.0, + 709.0, + 324.0, + 709.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 675.0, + 1105.0, + 675.0, + 1105.0, + 709.0, + 406.0, + 709.0 + ], + "score": 0.974, + "text": " as the input and outputs the probability for the edge between" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 675.0, + 1183.0, + 675.0, + 1183.0, + 709.0, + 1133.0, + 709.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 675.0, + 1401.0, + 675.0, + 1401.0, + 709.0, + 1213.0, + 709.0 + ], + "score": 0.996, + "text": "to be part of the" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 704.0, + 693.0, + 704.0, + 693.0, + 741.0, + 292.0, + 741.0 + ], + "score": 0.991, + "text": "explanation. This can be written as" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 487.0, + 422.0, + 487.0, + 422.0, + 521.0, + 296.0, + 521.0 + ], + "score": 0.987, + "text": "Denote by" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 487.0, + 1222.0, + 487.0, + 1222.0, + 521.0, + 453.0, + 521.0 + ], + "score": 0.995, + "text": "the neural network to generate a subset of edges of an input graph" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 487.0, + 1405.0, + 487.0, + 1405.0, + 521.0, + 1250.0, + 521.0 + ], + "score": 0.972, + "text": "as the robust" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 515.0, + 799.0, + 515.0, + 799.0, + 550.0, + 295.0, + 550.0 + ], + "score": 0.998, + "text": "counterfactual explanation on the prediction" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 515.0, + 873.0, + 515.0, + 873.0, + 550.0, + 861.0, + 550.0 + ], + "score": 0.887, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 515.0, + 1405.0, + 515.0, + 1405.0, + 550.0, + 893.0, + 550.0 + ], + "score": 0.977, + "text": "represents the set of parameters of the neural" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 544.0, + 892.0, + 544.0, + 892.0, + 578.0, + 295.0, + 578.0 + ], + "score": 0.985, + "text": "network. For experiments, our explanation network" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 544.0, + 1406.0, + 544.0, + 1406.0, + 578.0, + 914.0, + 578.0 + ], + "score": 0.974, + "text": "consists of 2 fully connected layers with a" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 574.0, + 849.0, + 574.0, + 849.0, + 604.0, + 296.0, + 604.0 + ], + "score": 0.981, + "text": "ReLU activation and the hidden dimension of 64." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 881.0, + 516.0, + 881.0, + 516.0, + 919.0, + 292.0, + 919.0 + ], + "score": 0.999, + "text": "For an input graph" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 881.0, + 723.0, + 881.0, + 723.0, + 919.0, + 662.0, + 919.0 + ], + "score": 0.998, + "text": "with" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 881.0, + 1174.0, + 881.0, + 1174.0, + 919.0, + 745.0, + 919.0 + ], + "score": 0.997, + "text": "vertices and a trained neural network" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 881.0, + 1316.0, + 881.0, + 1316.0, + 919.0, + 1254.0, + 919.0 + ], + "score": 0.992, + "text": "is an" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 881.0, + 1380.0, + 881.0, + 1380.0, + 919.0, + 1338.0, + 919.0 + ], + "score": 0.957, + "text": "-by-" + }, + { + "category_id": 16, + "poly": [ + 1402.0, + 881.0, + 1405.0, + 881.0, + 1405.0, + 919.0, + 1402.0, + 919.0 + ], + "score": 0.681, + "text": "（204号" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 913.0, + 1407.0, + 913.0, + 1407.0, + 947.0, + 294.0, + 947.0 + ], + "score": 0.996, + "text": "matrix that carries the complete information to generate a robust counterfactual explanation as a" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 943.0, + 612.0, + 943.0, + 612.0, + 976.0, + 295.0, + 976.0 + ], + "score": 0.982, + "text": "subset of edges, denoted by" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 943.0, + 955.0, + 943.0, + 955.0, + 976.0, + 693.0, + 976.0 + ], + "score": 0.986, + "text": ". Concretely,we obtain" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 943.0, + 1297.0, + 943.0, + 1297.0, + 976.0, + 979.0, + 976.0 + ], + "score": 0.994, + "text": "by selecting all the edges in" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 943.0, + 1404.0, + 943.0, + 1404.0, + 976.0, + 1324.0, + 976.0 + ], + "score": 1.0, + "text": "whose" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 971.0, + 571.0, + 971.0, + 571.0, + 1004.0, + 295.0, + 1004.0 + ], + "score": 0.998, + "text": "corresponding entries in" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 971.0, + 825.0, + 971.0, + 825.0, + 1004.0, + 607.0, + 1004.0 + ], + "score": 0.976, + "text": "are larger than O.5." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1918.0, + 370.0, + 1918.0, + 370.0, + 1952.0, + 295.0, + 1952.0 + ], + "score": 0.923, + "text": "where" + }, + { + "category_id": 16, + "poly": [ + 480.0, + 1918.0, + 489.0, + 1918.0, + 489.0, + 1952.0, + 480.0, + 1952.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 1918.0, + 611.0, + 1918.0, + 611.0, + 1952.0, + 563.0, + 1952.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1918.0, + 1407.0, + 1918.0, + 1407.0, + 1952.0, + 683.0, + 1952.0 + ], + "score": 0.991, + "text": "are the hyperparameters controlling the importance of each term." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1945.0, + 977.0, + 1945.0, + 977.0, + 1983.0, + 294.0, + 1983.0 + ], + "score": 0.975, + "text": "The influence of these parameters is discussed in Appendix" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1945.0, + 1406.0, + 1945.0, + 1406.0, + 1983.0, + 1003.0, + 1983.0 + ], + "score": 0.994, + "text": ".The first term of our loss function" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1974.0, + 767.0, + 1974.0, + 767.0, + 2010.0, + 294.0, + 2010.0 + ], + "score": 0.99, + "text": "requires that the prediction of the GNN on" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1974.0, + 1196.0, + 1974.0, + 1196.0, + 2010.0, + 804.0, + 2010.0 + ], + "score": 0.996, + "text": "is consistent with the prediction on" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2010.0, + 1222.0, + 2010.0 + ], + "score": 0.972, + "text": ". Intuitively, this" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1218.0, + 485.0, + 1218.0, + 485.0, + 1254.0, + 294.0, + 1254.0 + ], + "score": 0.989, + "text": "Since producing" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1218.0, + 545.0, + 1218.0, + 545.0, + 1254.0, + 508.0, + 1254.0 + ], + "score": 1.0, + "text": "by" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1218.0, + 1407.0, + 1218.0, + 1407.0, + 1254.0, + 576.0, + 1254.0 + ], + "score": 0.989, + "text": " is a discrete operation that is hard to incorporate in an end-to-end training" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1247.0, + 875.0, + 1247.0, + 875.0, + 1283.0, + 293.0, + 1283.0 + ], + "score": 0.98, + "text": " process, we define two proxy graphs to approximate" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1247.0, + 966.0, + 1247.0, + 966.0, + 1283.0, + 917.0, + 1283.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 1247.0, + 1404.0, + 1247.0, + 1404.0, + 1283.0, + 1036.0, + 1283.0 + ], + "score": 0.991, + "text": ",respectively,such that the proxy" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1279.0, + 584.0, + 1279.0, + 584.0, + 1316.0, + 292.0, + 1316.0 + ], + "score": 0.983, + "text": " graphs are determined by" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1279.0, + 1408.0, + 1279.0, + 1408.0, + 1316.0, + 603.0, + 1316.0 + ], + "score": 0.999, + "text": "through continuous functions that can be smoothly incorporated into an" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1310.0, + 613.0, + 1310.0, + 613.0, + 1342.0, + 296.0, + 1342.0 + ], + "score": 0.995, + "text": "end-to-end training process." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1349.0, + 523.0, + 1349.0, + 523.0, + 1388.0, + 292.0, + 1388.0 + ], + "score": 0.988, + "text": "The proxy graph of" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1349.0, + 708.0, + 1349.0, + 708.0, + 1388.0, + 565.0, + 1388.0 + ], + "score": 0.975, + "text": ", denoted by" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1349.0, + 1024.0, + 1349.0, + 1024.0, + 1388.0, + 747.0, + 1388.0 + ], + "score": 0.98, + "text": ", is defined by regarding" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1349.0, + 1182.0, + 1349.0, + 1182.0, + 1388.0, + 1061.0, + 1388.0 + ], + "score": 0.997, + "text": "instead of" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1349.0, + 1406.0, + 1349.0, + 1406.0, + 1388.0, + 1210.0, + 1388.0 + ], + "score": 0.982, + "text": " as the adjacency" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1377.0, + 472.0, + 1377.0, + 472.0, + 1418.0, + 291.0, + 1418.0 + ], + "score": 0.989, + "text": "matrix. That is," + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1377.0, + 943.0, + 1377.0, + 943.0, + 1418.0, + 509.0, + 1418.0 + ], + "score": 0.991, + "text": " has exactly the same graph structure as" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1377.0, + 1239.0, + 1377.0, + 1239.0, + 1418.0, + 969.0, + 1418.0 + ], + "score": 0.99, + "text": ", but the edge weights of" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1377.0, + 1406.0, + 1377.0, + 1406.0, + 1418.0, + 1277.0, + 1418.0 + ], + "score": 0.984, + "text": "is given by" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1410.0, + 446.0, + 1410.0, + 446.0, + 1443.0, + 295.0, + 1443.0 + ], + "score": 0.995, + "text": "the entries in" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1410.0, + 599.0, + 1410.0, + 599.0, + 1443.0, + 482.0, + 1443.0 + ], + "score": 0.998, + "text": "instead of" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1410.0, + 854.0, + 1410.0, + 854.0, + 1443.0, + 627.0, + 1443.0 + ], + "score": 0.973, + "text": ".Here,the subscript" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1410.0, + 952.0, + 1410.0, + 952.0, + 1443.0, + 873.0, + 1443.0 + ], + "score": 1.0, + "text": "means" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1410.0, + 1185.0, + 1410.0, + 1185.0, + 1443.0, + 991.0, + 1443.0 + ], + "score": 0.996, + "text": "is determined by" + }, + { + "category_id": 16, + "poly": [ + 1203.0, + 1410.0, + 1212.0, + 1410.0, + 1212.0, + 1443.0, + 1203.0, + 1443.0 + ], + "score": 0.378, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1635.0, + 580.0, + 1635.0, + 580.0, + 1679.0, + 293.0, + 1679.0 + ], + "score": 0.986, + "text": "The edge weights of both" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1635.0, + 668.0, + 1635.0, + 668.0, + 1679.0, + 619.0, + 1679.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1635.0, + 913.0, + 1635.0, + 913.0, + 1679.0, + 706.0, + 1679.0 + ], + "score": 0.958, + "text": " are determined by" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1635.0, + 1407.0, + 1635.0, + 1407.0, + 1679.0, + 932.0, + 1679.0 + ], + "score": 0.986, + "text": "through continuous functions, thus we can" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1667.0, + 540.0, + 1667.0, + 540.0, + 1703.0, + 293.0, + 1703.0 + ], + "score": 0.993, + "text": "smoothly incorporate" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1667.0, + 628.0, + 1667.0, + 628.0, + 1703.0, + 578.0, + 1703.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1667.0, + 1106.0, + 1667.0, + 1106.0, + 1703.0, + 666.0, + 1703.0 + ], + "score": 0.992, + "text": "into an end-to-end training framework." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 810.0, + 371.0, + 810.0, + 371.0, + 845.0, + 296.0, + 845.0 + ], + "score": 1.0, + "text": "where" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 810.0, + 942.0, + 810.0, + 942.0, + 845.0, + 425.0, + 845.0 + ], + "score": 0.996, + "text": "denotes the probability that the edge between" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 810.0, + 1020.0, + 810.0, + 1020.0, + 845.0, + 970.0, + 845.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 810.0, + 1408.0, + 810.0, + 1408.0, + 845.0, + 1049.0, + 845.0 + ], + "score": 0.981, + "text": "is contained in the explanation." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 838.0, + 648.0, + 838.0, + 648.0, + 873.0, + 294.0, + 873.0 + ], + "score": 0.972, + "text": "When there is no edge between" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 838.0, + 725.0, + 838.0, + 725.0, + 873.0, + 676.0, + 873.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 838.0, + 845.0, + 838.0, + 845.0, + 873.0, + 754.0, + 873.0 + ], + "score": 0.983, + "text": ",that is," + }, + { + "category_id": 15, + "poly": [ + 944.0, + 838.0, + 1030.0, + 838.0, + 1030.0, + 873.0, + 944.0, + 873.0 + ], + "score": 0.942, + "text": ",we set" + }, + { + "category_id": 16, + "poly": [ + 1135.0, + 838.0, + 1143.0, + 838.0, + 1143.0, + 873.0, + 1135.0, + 873.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1705.0, + 1347.0, + 1705.0, + 1347.0, + 1751.0, + 292.0, + 1751.0 + ], + "score": 0.975, + "text": " As discussed later in this section, we use a regularization term to force the value of each entry in" + }, + { + "category_id": 16, + "poly": [ + 1401.0, + 1705.0, + 1405.0, + 1705.0, + 1405.0, + 1751.0, + 1401.0, + 1751.0 + ], + "score": 0.747, + "text": "（20" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1735.0, + 705.0, + 1735.0, + 705.0, + 1778.0, + 293.0, + 1778.0 + ], + "score": 0.979, + "text": "to be close to eitherOor 1,such that" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1735.0, + 793.0, + 1735.0, + 793.0, + 1778.0, + 743.0, + 1778.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1735.0, + 1049.0, + 1735.0, + 1049.0, + 1778.0, + 831.0, + 1778.0 + ], + "score": 0.997, + "text": "better approximate" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 1735.0, + 1141.0, + 1735.0, + 1141.0, + 1778.0, + 1092.0, + 1778.0 + ], + "score": 0.999, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1735.0, + 1358.0, + 1735.0, + 1358.0, + 1778.0, + 1212.0, + 1778.0 + ], + "score": 1.0, + "text": "respectively." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1452.0, + 525.0, + 1452.0, + 525.0, + 1492.0, + 294.0, + 1492.0 + ], + "score": 0.958, + "text": "The proxy graph of" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 1452.0, + 739.0, + 1452.0, + 739.0, + 1492.0, + 595.0, + 1492.0 + ], + "score": 0.981, + "text": ", denoted by" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1452.0, + 1221.0, + 1452.0, + 1221.0, + 1492.0, + 778.0, + 1492.0 + ], + "score": 0.986, + "text": ", also have the same graph structure as" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1452.0, + 1402.0, + 1452.0, + 1402.0, + 1492.0, + 1247.0, + 1492.0 + ], + "score": 0.972, + "text": ", but the edge" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1483.0, + 707.0, + 1483.0, + 707.0, + 1518.0, + 296.0, + 1518.0 + ], + "score": 0.986, + "text": "weight between each pair of vertices" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1483.0, + 783.0, + 1483.0, + 783.0, + 1518.0, + 734.0, + 1518.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1483.0, + 962.0, + 1483.0, + 962.0, + 1518.0, + 813.0, + 1518.0 + ], + "score": 0.981, + "text": "is defined as" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1782.0, + 677.0, + 1782.0, + 677.0, + 1820.0, + 295.0, + 1820.0 + ], + "score": 0.987, + "text": "We formulate our loss function as" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1313, + 1405, + 1313, + 1405, + 1491, + 296, + 1491 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 274, + 1405, + 274, + 1405, + 461, + 297, + 461 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1574, + 1404, + 1574, + 1404, + 1725, + 298, + 1725 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1831, + 1405, + 1831, + 1405, + 2008, + 298, + 2008 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 670, + 1405, + 670, + 1405, + 819, + 297, + 819 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 901, + 1404, + 901, + 1404, + 1020, + 297, + 1020 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 569, + 1404, + 569, + 1404, + 660, + 298, + 660 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1237, + 1403, + 1237, + 1403, + 1300, + 298, + 1300 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 494, + 471, + 1205, + 471, + 1205, + 554, + 494, + 554 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 294, + 200, + 1400, + 200, + 1400, + 264, + 294, + 264 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 298, + 1502, + 1400, + 1502, + 1400, + 1562, + 298, + 1562 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 433, + 1139, + 1268, + 1139, + 1268, + 1218, + 433, + 1218 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 550, + 831, + 1151, + 831, + 1151, + 885, + 550, + 885 + ], + "score": 0.948 + }, + { + "category_id": 0, + "poly": [ + 298, + 1763, + 532, + 1763, + 532, + 1801, + 298, + 1801 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 300, + 1031, + 1405, + 1031, + 1405, + 1124, + 300, + 1124 + ], + "score": 0.914 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1156, + 1400, + 1156, + 1400, + 1187, + 1352, + 1187 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1353, + 844, + 1400, + 844, + 1400, + 874, + 1353, + 874 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 497, + 1400, + 497, + 1400, + 527, + 1366, + 527 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 842, + 2061, + 858, + 2061, + 858, + 2084, + 842, + 2084 + ], + "score": 0.714 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2084, + 841, + 2084 + ], + "score": 0.127 + }, + { + "category_id": 14, + "poly": [ + 494, + 468, + 1203, + 468, + 1203, + 555, + 494, + 555 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { s a m e } ( \\theta , G ) = \\frac { 1 } { | \\mathcal { H } _ { G } | } \\sum _ { h _ { i } \\in \\mathcal { H } _ { G } } \\sigma \\left( - \\mathcal { B } _ { i } ( \\phi _ { g c } ( G ) ) * \\mathcal { B } _ { i } ( \\phi _ { g c } ( G _ { \\theta } ) ) \\right) ," + }, + { + "category_id": 14, + "poly": [ + 434, + 1136, + 1266, + 1136, + 1266, + 1222, + 434, + 1222 + ], + "score": 0.94, + "latex": "\\mathcal { R } _ { d i s c r e t e } ( \\theta , G ) = - \\frac { 1 } { | \\mathbf { M } | } \\sum _ { i , j } ( \\mathbf { M } _ { i j } \\log ( \\mathbf { M } _ { i j } ) + ( 1 - \\mathbf { M } _ { i j } ) \\log ( 1 - \\mathbf { M } _ { i j } ) )" + }, + { + "category_id": 13, + "poly": [ + 768, + 1033, + 1037, + 1033, + 1037, + 1067, + 768, + 1067 + ], + "score": 0.94, + "latex": "\\mathcal { R } _ { s p a r s e } ( \\theta , G ) = \\| \\mathbf { M } \\| _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 541, + 1610, + 580, + 1610, + 580, + 1638, + 541, + 1638 + ], + "score": 0.93, + "latex": "| E |" + }, + { + "category_id": 13, + "poly": [ + 543, + 902, + 669, + 902, + 669, + 937, + 543, + 937 + ], + "score": 0.93, + "latex": "\\mathcal { L } _ { o p p } ( \\theta , G )" + }, + { + "category_id": 13, + "poly": [ + 969, + 402, + 1017, + 402, + 1017, + 434, + 969, + 434 + ], + "score": 0.93, + "latex": "\\sigma ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 669, + 1239, + 721, + 1239, + 721, + 1273, + 669, + 1273 + ], + "score": 0.93, + "latex": "\\mathbf { M } _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 1069, + 571, + 1154, + 571, + 1154, + 604, + 1069, + 604 + ], + "score": 0.92, + "latex": "\\phi _ { g c } ( G )" + }, + { + "category_id": 13, + "poly": [ + 721, + 1344, + 778, + 1344, + 778, + 1377, + 721, + 1377 + ], + "score": 0.92, + "latex": "\\mathcal { L } ( \\boldsymbol { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 1206, + 570, + 1303, + 570, + 1303, + 604, + 1206, + 604 + ], + "score": 0.92, + "latex": "\\phi _ { g c } ( G _ { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 1203, + 902, + 1300, + 902, + 1300, + 937, + 1203, + 937 + ], + "score": 0.92, + "latex": "\\phi _ { g c } ( G _ { \\theta } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 1293, + 373, + 1375, + 373, + 1375, + 405, + 1293, + 405 + ], + "score": 0.92, + "latex": "| B _ { i } ( { \\bf x } ) |" + }, + { + "category_id": 13, + "poly": [ + 907, + 306, + 1125, + 306, + 1125, + 342, + 907, + 342 + ], + "score": 0.92, + "latex": "B _ { i } ( \\mathbf { x } ) = \\mathbf { w } _ { i } ^ { \\top } \\mathbf { x } + b _ { i }" + }, + { + "category_id": 13, + "poly": [ + 917, + 1575, + 999, + 1575, + 999, + 1608, + 917, + 1608 + ], + "score": 0.92, + "latex": "O ( | E | )" + }, + { + "category_id": 13, + "poly": [ + 1211, + 276, + 1270, + 276, + 1270, + 309, + 1211, + 309 + ], + "score": 0.92, + "latex": "| \\mathcal { H } _ { G } |" + }, + { + "category_id": 13, + "poly": [ + 1065, + 903, + 1150, + 903, + 1150, + 935, + 1065, + 935 + ], + "score": 0.92, + "latex": "\\phi _ { g c } ( G )" + }, + { + "category_id": 13, + "poly": [ + 390, + 1269, + 458, + 1269, + 458, + 1304, + 390, + 1304 + ], + "score": 0.91, + "latex": "G _ { E \\backslash S }" + }, + { + "category_id": 13, + "poly": [ + 538, + 570, + 682, + 570, + 682, + 602, + 538, + 602 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { s a m e } ( \\theta , G )" + }, + { + "category_id": 13, + "poly": [ + 1229, + 343, + 1297, + 343, + 1297, + 375, + 1229, + 375 + ], + "score": 0.91, + "latex": "B _ { i } ( { \\bf x } )" + }, + { + "category_id": 13, + "poly": [ + 1046, + 1375, + 1180, + 1375, + 1180, + 1405, + 1046, + 1405 + ], + "score": 0.91, + "latex": "G = ( V , E )" + }, + { + "category_id": 13, + "poly": [ + 677, + 308, + 781, + 308, + 781, + 340, + 677, + 340 + ], + "score": 0.91, + "latex": "h _ { i } \\in \\mathcal { H } _ { G }" + }, + { + "category_id": 13, + "poly": [ + 713, + 341, + 801, + 341, + 801, + 371, + 713, + 371 + ], + "score": 0.9, + "latex": "\\mathbf { x } \\in \\mathbb { O } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 549, + 831, + 1148, + 831, + 1148, + 886, + 549, + 886 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { o p p } ( \\theta , G ) = \\operatorname* { m i n } _ { h _ { i } \\in \\mathcal { H } _ { G } } \\sigma \\left( \\mathcal { B } _ { i } ( \\phi _ { g c } ( G ) ) * \\mathcal { B } _ { i } ( \\phi _ { g c } ( G _ { \\theta } ^ { \\prime } ) ) \\right) ," + }, + { + "category_id": 13, + "poly": [ + 1220, + 1239, + 1255, + 1239, + 1255, + 1273, + 1220, + 1273 + ], + "score": 0.9, + "latex": "G _ { \\theta } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 612, + 701, + 648, + 701, + 648, + 734, + 612, + 734 + ], + "score": 0.9, + "latex": "G _ { \\theta } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 789, + 600, + 826, + 600, + 826, + 630, + 789, + 630 + ], + "score": 0.89, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1039, + 961, + 1075, + 961, + 1075, + 995, + 1039, + 995 + ], + "score": 0.89, + "latex": "G _ { \\theta } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 764, + 933, + 809, + 933, + 809, + 962, + 764, + 962 + ], + "score": 0.89, + "latex": "\\mathcal { H } _ { G }" + }, + { + "category_id": 13, + "poly": [ + 427, + 310, + 471, + 310, + 471, + 340, + 427, + 340 + ], + "score": 0.89, + "latex": "\\mathcal { H } _ { G }" + }, + { + "category_id": 13, + "poly": [ + 656, + 602, + 700, + 602, + 700, + 630, + 656, + 630 + ], + "score": 0.89, + "latex": "\\mathcal { H } _ { G }" + }, + { + "category_id": 13, + "poly": [ + 418, + 276, + 462, + 276, + 462, + 307, + 418, + 307 + ], + "score": 0.89, + "latex": "\\mathcal { H } _ { G }" + }, + { + "category_id": 13, + "poly": [ + 1132, + 1239, + 1168, + 1239, + 1168, + 1269, + 1132, + 1269 + ], + "score": 0.88, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 932, + 731, + 969, + 731, + 969, + 760, + 932, + 760 + ], + "score": 0.88, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 441, + 232, + 477, + 232, + 477, + 262, + 441, + 262 + ], + "score": 0.88, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 772, + 204, + 808, + 204, + 808, + 233, + 772, + 233 + ], + "score": 0.88, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 536, + 1374, + 563, + 1374, + 563, + 1404, + 536, + 1404 + ], + "score": 0.88, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1004, + 373, + 1031, + 373, + 1031, + 402, + 1004, + 402 + ], + "score": 0.87, + "latex": "h _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1296, + 310, + 1319, + 310, + 1319, + 340, + 1296, + 340 + ], + "score": 0.87, + "latex": "b _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1344, + 1316, + 1371, + 1316, + 1371, + 1347, + 1344, + 1347 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1211, + 313, + 1245, + 313, + 1245, + 340, + 1211, + 340 + ], + "score": 0.87, + "latex": "\\mathbf { w } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 485, + 343, + 513, + 343, + 513, + 373, + 485, + 373 + ], + "score": 0.87, + "latex": "h _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1048, + 342, + 1084, + 342, + 1084, + 370, + 1048, + 370 + ], + "score": 0.87, + "latex": "\\mathbb { O } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 806, + 402, + 833, + 402, + 833, + 431, + 806, + 431 + ], + "score": 0.87, + "latex": "h _ { i }" + }, + { + "category_id": 13, + "poly": [ + 432, + 1611, + 454, + 1611, + 454, + 1632, + 432, + 1632 + ], + "score": 0.87, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1094, + 934, + 1130, + 934, + 1130, + 965, + 1094, + 965 + ], + "score": 0.86, + "latex": "G _ { \\theta } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1273, + 337, + 1273, + 337, + 1298, + 298, + 1298 + ], + "score": 0.86, + "latex": "G _ { S }" + }, + { + "category_id": 13, + "poly": [ + 299, + 1376, + 325, + 1376, + 325, + 1404, + 299, + 1404 + ], + "score": 0.86, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1327, + 1606, + 1355, + 1606, + 1355, + 1636, + 1327, + 1636 + ], + "score": 0.86, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 435, + 1406, + 455, + 1406, + 455, + 1428, + 435, + 1428 + ], + "score": 0.86, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1035, + 1400, + 1035, + 1400, + 1065, + 1373, + 1065 + ], + "score": 0.85, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 718, + 1461, + 736, + 1461, + 736, + 1491, + 718, + 1491 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 554, + 995, + 577, + 995, + 577, + 1017, + 554, + 1017 + ], + "score": 0.85, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 774, + 1404, + 799, + 1404, + 799, + 1428, + 774, + 1428 + ], + "score": 0.85, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 499, + 1067, + 522, + 1067, + 522, + 1090, + 499, + 1090 + ], + "score": 0.83, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 478, + 630, + 502, + 630, + 502, + 655, + 478, + 655 + ], + "score": 0.83, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1336, + 1065, + 1360, + 1065, + 1360, + 1090, + 1336, + 1090 + ], + "score": 0.83, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 753, + 963, + 776, + 963, + 776, + 988, + 753, + 988 + ], + "score": 0.83, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1056, + 233, + 1080, + 233, + 1080, + 259, + 1056, + 259 + ], + "score": 0.82, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 634, + 1316, + 660, + 1316, + 660, + 1342, + 634, + 1342 + ], + "score": 0.82, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 925, + 1608, + 949, + 1608, + 949, + 1632, + 925, + 1632 + ], + "score": 0.82, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1093, + 277, + 1118, + 277, + 1118, + 304, + 1093, + 304 + ], + "score": 0.81, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 838, + 1345, + 855, + 1345, + 855, + 1371, + 838, + 1371 + ], + "score": 0.81, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1238, + 702, + 1262, + 702, + 1262, + 728, + 1238, + 728 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1111, + 204, + 1136, + 204, + 1136, + 230, + 1111, + 230 + ], + "score": 0.8, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 569, + 310, + 582, + 310, + 582, + 336, + 569, + 336 + ], + "score": 0.76, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 721, + 405, + 742, + 405, + 742, + 429, + 721, + 429 + ], + "score": 0.69, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 472, + 376, + 493, + 376, + 493, + 399, + 472, + 399 + ], + "score": 0.5, + "latex": "\\mathbf { x }" + }, + { + "category_id": 13, + "poly": [ + 807, + 1373, + 841, + 1373, + 841, + 1400, + 807, + 1400 + ], + "score": 0.42, + "latex": "\\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 1157, + 1405, + 1190, + 1405, + 1190, + 1428, + 1157, + 1428 + ], + "score": 0.35, + "latex": "\\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 1192, + 1034, + 1227, + 1034, + 1227, + 1062, + 1192, + 1062 + ], + "score": 0.32, + "latex": "\\mathbf { M }" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1757.0, + 538.0, + 1757.0, + 538.0, + 1811.0, + 288.0, + 1811.0 + ], + "score": 1.0, + "text": "5Experiments" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2092.0, + 840.0, + 2092.0 + ], + "score": 1.0, + "text": "7" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2092.0, + 840.0, + 2092.0 + ], + "score": 1.0, + "text": "7" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1314.0, + 633.0, + 1314.0, + 633.0, + 1349.0, + 294.0, + 1349.0 + ], + "score": 0.999, + "text": "Now we can use the graphs in" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1314.0, + 1343.0, + 1314.0, + 1343.0, + 1349.0, + 661.0, + 1349.0 + ], + "score": 0.996, + "text": "and the extracted decision regions to train the neural network" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1314.0, + 1405.0, + 1314.0, + 1405.0, + 1349.0, + 1372.0, + 1349.0 + ], + "score": 0.997, + "text": "in" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1340.0, + 720.0, + 1340.0, + 720.0, + 1382.0, + 291.0, + 1382.0 + ], + "score": 0.982, + "text": " an end-to-end manner by minimizing" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1340.0, + 837.0, + 1340.0, + 837.0, + 1382.0, + 779.0, + 1382.0 + ], + "score": 0.996, + "text": "over" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1340.0, + 1407.0, + 1340.0, + 1407.0, + 1382.0, + 856.0, + 1382.0 + ], + "score": 0.975, + "text": "using back propagation. Once we finish training" + }, + { + "category_id": 16, + "poly": [ + 293.0, + 1372.0, + 298.0, + 1372.0, + 298.0, + 1407.0, + 293.0, + 1407.0 + ], + "score": 0.662, + "text": "（204" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1372.0, + 535.0, + 1372.0, + 535.0, + 1407.0, + 326.0, + 1407.0 + ], + "score": 0.979, + "text": ", we can first apply" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1372.0, + 806.0, + 1372.0, + 806.0, + 1407.0, + 564.0, + 1407.0 + ], + "score": 0.999, + "text": "to produce the matrix" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1372.0, + 1045.0, + 1372.0, + 1045.0, + 1407.0, + 842.0, + 1407.0 + ], + "score": 0.979, + "text": "for an input graph" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1372.0, + 1406.0, + 1372.0, + 1406.0, + 1407.0, + 1181.0, + 1407.0 + ], + "score": 0.984, + "text": ",and then obtain the" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1401.0, + 434.0, + 1401.0, + 434.0, + 1436.0, + 294.0, + 1436.0 + ], + "score": 1.0, + "text": "explanation" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1401.0, + 773.0, + 1401.0, + 773.0, + 1436.0, + 456.0, + 1436.0 + ], + "score": 0.998, + "text": "by selecting all the edges in" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1401.0, + 1156.0, + 1401.0, + 1156.0, + 1436.0, + 800.0, + 1436.0 + ], + "score": 0.996, + "text": "whose corresponding entries in" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1401.0, + 1408.0, + 1401.0, + 1408.0, + 1436.0, + 1191.0, + 1436.0 + ], + "score": 0.969, + "text": "are larger than 0.5." + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1426.0, + 1406.0, + 1426.0, + 1406.0, + 1467.0, + 291.0, + 1467.0 + ], + "score": 0.989, + "text": "We do not need the extracted boundaries for inference as the the decision logic of GNN is already" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1454.0, + 717.0, + 1454.0, + 717.0, + 1498.0, + 292.0, + 1498.0 + ], + "score": 0.985, + "text": "distilled into the explanation network" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1454.0, + 960.0, + 1454.0, + 960.0, + 1498.0, + 737.0, + 1498.0 + ], + "score": 0.998, + "text": "during the training." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 274.0, + 417.0, + 274.0, + 417.0, + 310.0, + 295.0, + 310.0 + ], + "score": 0.996, + "text": "Denote by" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 274.0, + 1092.0, + 274.0, + 1092.0, + 310.0, + 463.0, + 310.0 + ], + "score": 0.975, + "text": " the set of LDBs that induce the decision region covering" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 274.0, + 1210.0, + 274.0, + 1210.0, + 310.0, + 1119.0, + 310.0 + ], + "score": 0.963, + "text": ",and by" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 274.0, + 1407.0, + 274.0, + 1407.0, + 310.0, + 1271.0, + 310.0 + ], + "score": 1.0, + "text": "the number" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 307.0, + 426.0, + 307.0, + 426.0, + 343.0, + 294.0, + 343.0 + ], + "score": 0.971, + "text": "of LDBs in" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 307.0, + 568.0, + 307.0, + 568.0, + 343.0, + 472.0, + 343.0 + ], + "score": 0.992, + "text": ". For the" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 307.0, + 676.0, + 307.0, + 676.0, + 343.0, + 583.0, + 343.0 + ], + "score": 0.965, + "text": "-th LDB" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 307.0, + 906.0, + 307.0, + 906.0, + 343.0, + 782.0, + 343.0 + ], + "score": 0.971, + "text": ", denote by" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 307.0, + 1210.0, + 307.0, + 1210.0, + 343.0, + 1126.0, + 343.0 + ], + "score": 0.898, + "text": ", where" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 307.0, + 1295.0, + 307.0, + 1295.0, + 343.0, + 1246.0, + 343.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 307.0, + 1405.0, + 307.0, + 1405.0, + 343.0, + 1320.0, + 343.0 + ], + "score": 0.999, + "text": "are the" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 338.0, + 484.0, + 338.0, + 484.0, + 379.0, + 291.0, + 379.0 + ], + "score": 0.971, + "text": " basis and bias of" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 338.0, + 712.0, + 338.0, + 712.0, + 379.0, + 514.0, + 379.0 + ], + "score": 0.964, + "text": ", respectively, and" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 338.0, + 1047.0, + 338.0, + 1047.0, + 379.0, + 802.0, + 379.0 + ], + "score": 0.981, + "text": "is a point in the space" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 338.0, + 1228.0, + 338.0, + 1228.0, + 379.0, + 1085.0, + 379.0 + ], + "score": 0.985, + "text": ". The sign of" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 338.0, + 1406.0, + 338.0, + 1406.0, + 379.0, + 1298.0, + 379.0 + ], + "score": 1.0, + "text": "indicates" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 371.0, + 471.0, + 371.0, + 471.0, + 407.0, + 294.0, + 407.0 + ], + "score": 0.999, + "text": "whether a point" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 371.0, + 1003.0, + 371.0, + 1003.0, + 407.0, + 494.0, + 407.0 + ], + "score": 0.99, + "text": "lies on the positive side or the negative side of" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 371.0, + 1292.0, + 371.0, + 1292.0, + 407.0, + 1032.0, + 407.0 + ], + "score": 0.991, + "text": ", and the absolute value" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 371.0, + 1405.0, + 371.0, + 1405.0, + 407.0, + 1376.0, + 407.0 + ], + "score": 0.999, + "text": "is" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 398.0, + 720.0, + 398.0, + 720.0, + 438.0, + 292.0, + 438.0 + ], + "score": 0.978, + "text": " proportional to the distance of a point" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 398.0, + 805.0, + 398.0, + 805.0, + 438.0, + 743.0, + 438.0 + ], + "score": 1.0, + "text": "from" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 398.0, + 968.0, + 398.0, + 968.0, + 438.0, + 834.0, + 438.0 + ], + "score": 0.953, + "text": ". Denote by" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 398.0, + 1407.0, + 398.0, + 1407.0, + 438.0, + 1018.0, + 438.0 + ], + "score": 0.999, + "text": "the standard sigmoid function, we" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 429.0, + 813.0, + 429.0, + 813.0, + 462.0, + 294.0, + 462.0 + ], + "score": 0.991, + "text": "formulate the first term of our loss function as" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1571.0, + 916.0, + 1571.0, + 916.0, + 1612.0, + 293.0, + 1612.0 + ], + "score": 0.998, + "text": "Our method is highly efficient with a time complexity" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1571.0, + 1406.0, + 1571.0, + 1406.0, + 1612.0, + 1000.0, + 1612.0 + ], + "score": 0.98, + "text": " for explaining the prediction on an" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1602.0, + 431.0, + 1602.0, + 431.0, + 1640.0, + 293.0, + 1640.0 + ], + "score": 0.971, + "text": " input graph" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1602.0, + 540.0, + 1602.0, + 540.0, + 1640.0, + 455.0, + 1640.0 + ], + "score": 0.926, + "text": ", where" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1602.0, + 924.0, + 1602.0, + 924.0, + 1640.0, + 581.0, + 1640.0 + ], + "score": 0.964, + "text": " is the total number of edges in" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1602.0, + 1326.0, + 1602.0, + 1326.0, + 1640.0, + 950.0, + 1640.0 + ], + "score": 0.979, + "text": ": Additionally, the neural network" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1602.0, + 1405.0, + 1602.0, + 1405.0, + 1640.0, + 1356.0, + 1640.0 + ], + "score": 1.0, + "text": "can" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1405.0, + 1633.0, + 1405.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 0.99, + "text": "be directly used without retraining to predict explanations on unseen graphs. Thus our method is" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1661.0, + 1404.0, + 1661.0, + 1404.0, + 1698.0, + 294.0, + 1698.0 + ], + "score": 0.99, + "text": " significantly faster than the other methods [45,32,47,38] that require retraining each time when" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1693.0, + 814.0, + 1693.0, + 814.0, + 1727.0, + 293.0, + 1727.0 + ], + "score": 0.982, + "text": " generating explanations on a new input graph." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1829.0, + 1410.0, + 1829.0, + 1410.0, + 1868.0, + 293.0, + 1868.0 + ], + "score": 0.992, + "text": "We conduct series of experiments to compare our method with the state-of-the-art methods in-" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1860.0, + 1408.0, + 1860.0, + 1408.0, + 1895.0, + 293.0, + 1895.0 + ], + "score": 0.988, + "text": "cluding GNNExplainer [45],PGExplainer [25],PGM-Explainer [38], SubgraphX [47] and CF-" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1887.0, + 1407.0, + 1887.0, + 1407.0, + 1923.0, + 293.0, + 1923.0 + ], + "score": 0.985, + "text": "GNNExplainer [24]. For the methods that identify a set of vertices as an explanation, we use the set" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1918.0, + 1406.0, + 1918.0, + 1406.0, + 1953.0, + 295.0, + 1953.0 + ], + "score": 0.996, + "text": "of vertices to induce a subgraph from the input graph,and then use the set of edges of the induced" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1948.0, + 1405.0, + 1948.0, + 1405.0, + 1983.0, + 295.0, + 1983.0 + ], + "score": 0.975, + "text": " subgraph as the explanation. For the methods that identify a subgraph as an explanation, we directly" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1978.0, + 1021.0, + 1978.0, + 1021.0, + 2009.0, + 296.0, + 2009.0 + ], + "score": 0.985, + "text": "use the set of edges of the identified subgraph as the explanation." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 668.0, + 1406.0, + 668.0, + 1406.0, + 707.0, + 294.0, + 707.0 + ], + "score": 0.985, + "text": "The second term of our loss function optimizes the counterfactual property of the explanations by" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 697.0, + 611.0, + 697.0, + 611.0, + 737.0, + 292.0, + 737.0 + ], + "score": 0.998, + "text": "requiring the prediction on" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 697.0, + 1237.0, + 697.0, + 1237.0, + 737.0, + 649.0, + 737.0 + ], + "score": 0.993, + "text": "to be significantly different from the prediction on" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 697.0, + 1408.0, + 697.0, + 1408.0, + 737.0, + 1263.0, + 737.0 + ], + "score": 0.963, + "text": ". Intuitively," + }, + { + "category_id": 15, + "poly": [ + 294.0, + 729.0, + 931.0, + 729.0, + 931.0, + 765.0, + 294.0, + 765.0 + ], + "score": 0.993, + "text": "this means that the set of edges with larger weights in" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 729.0, + 1407.0, + 729.0, + 1407.0, + 765.0, + 970.0, + 765.0 + ], + "score": 0.981, + "text": " are good counterfactual explanations" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 756.0, + 1403.0, + 756.0, + 1403.0, + 794.0, + 292.0, + 794.0 + ], + "score": 0.993, + "text": "because reducing the weights of these edges significantly changes the prediction. Following the" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 789.0, + 843.0, + 789.0, + 843.0, + 821.0, + 295.0, + 821.0 + ], + "score": 0.996, + "text": "above intuition,we formulate the second term as" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 899.0, + 542.0, + 899.0, + 542.0, + 941.0, + 292.0, + 941.0 + ], + "score": 0.983, + "text": " such that minimizing" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 899.0, + 1064.0, + 899.0, + 1064.0, + 941.0, + 670.0, + 941.0 + ], + "score": 0.985, + "text": " encourages the graph embeddings" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 899.0, + 1202.0, + 899.0, + 1202.0, + 941.0, + 1151.0, + 941.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 899.0, + 1406.0, + 899.0, + 1406.0, + 941.0, + 1301.0, + 941.0 + ], + "score": 0.989, + "text": "to lie on" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 928.0, + 763.0, + 928.0, + 763.0, + 970.0, + 292.0, + 970.0 + ], + "score": 0.975, + "text": " the opposite sides of at least one LDB in" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 928.0, + 1093.0, + 928.0, + 1093.0, + 970.0, + 810.0, + 970.0 + ], + "score": 0.974, + "text": ". This further means that" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 928.0, + 1406.0, + 928.0, + 1406.0, + 970.0, + 1131.0, + 970.0 + ], + "score": 0.997, + "text": "is encouraged not to be" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 959.0, + 752.0, + 959.0, + 752.0, + 997.0, + 295.0, + 997.0 + ], + "score": 0.996, + "text": "covered by the decision region covering" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 959.0, + 1038.0, + 959.0, + 1038.0, + 997.0, + 777.0, + 997.0 + ], + "score": 0.984, + "text": ", thus the prediction on" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 959.0, + 1404.0, + 959.0, + 1404.0, + 997.0, + 1076.0, + 997.0 + ], + "score": 0.994, + "text": "can be changed significantly" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 989.0, + 553.0, + 989.0, + 553.0, + 1023.0, + 295.0, + 1023.0 + ], + "score": 0.985, + "text": "from the prediction on" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 989.0, + 588.0, + 989.0, + 588.0, + 1023.0, + 578.0, + 1023.0 + ], + "score": 0.644, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 567.0, + 537.0, + 567.0, + 537.0, + 607.0, + 293.0, + 607.0 + ], + "score": 0.982, + "text": " such that minimizing" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 567.0, + 1068.0, + 567.0, + 1068.0, + 607.0, + 683.0, + 607.0 + ], + "score": 0.999, + "text": "encourages the graph embeddings" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 567.0, + 1205.0, + 567.0, + 1205.0, + 607.0, + 1155.0, + 607.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 567.0, + 1405.0, + 567.0, + 1405.0, + 607.0, + 1304.0, + 607.0 + ], + "score": 0.999, + "text": "to lie on" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 597.0, + 655.0, + 597.0, + 655.0, + 632.0, + 293.0, + 632.0 + ], + "score": 0.988, + "text": "the same side of every LDB in" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 597.0, + 788.0, + 597.0, + 788.0, + 632.0, + 701.0, + 632.0 + ], + "score": 0.965, + "text": ". Thus," + }, + { + "category_id": 15, + "poly": [ + 827.0, + 597.0, + 1405.0, + 597.0, + 1405.0, + 632.0, + 827.0, + 632.0 + ], + "score": 0.969, + "text": "is encouraged to be covered by the same decision" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 629.0, + 477.0, + 629.0, + 477.0, + 662.0, + 293.0, + 662.0 + ], + "score": 0.993, + "text": "region covering" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 629.0, + 513.0, + 629.0, + 513.0, + 662.0, + 503.0, + 662.0 + ], + "score": 0.743, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1235.0, + 668.0, + 1235.0, + 668.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 0.972, + "text": " to push the value of each entry in" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1235.0, + 1131.0, + 1235.0, + 1131.0, + 1275.0, + 722.0, + 1275.0 + ], + "score": 0.988, + "text": "to be close to either Oor 1,such that" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1235.0, + 1219.0, + 1235.0, + 1219.0, + 1275.0, + 1169.0, + 1275.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1235.0, + 1405.0, + 1235.0, + 1405.0, + 1275.0, + 1256.0, + 1275.0 + ], + "score": 1.0, + "text": "approximate" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1267.0, + 389.0, + 1267.0, + 389.0, + 1303.0, + 338.0, + 1303.0 + ], + "score": 1.0, + "text": "and" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1267.0, + 667.0, + 1267.0, + 667.0, + 1303.0, + 459.0, + 1303.0 + ], + "score": 0.995, + "text": "well, respectively." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 201.0, + 771.0, + 201.0, + 771.0, + 237.0, + 294.0, + 237.0 + ], + "score": 0.985, + "text": "means that the edges with larger weights in" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 201.0, + 1110.0, + 201.0, + 1110.0, + 237.0, + 809.0, + 237.0 + ], + "score": 0.997, + "text": "dominate the prediction on" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 201.0, + 1405.0, + 201.0, + 1405.0, + 237.0, + 1137.0, + 237.0 + ], + "score": 0.994, + "text": ".We formulate this term" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 232.0, + 440.0, + 232.0, + 440.0, + 265.0, + 294.0, + 265.0 + ], + "score": 0.986, + "text": "by requiring" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 232.0, + 1055.0, + 232.0, + 1055.0, + 265.0, + 478.0, + 265.0 + ], + "score": 0.991, + "text": "to be covered by the same decision region covering" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 232.0, + 1092.0, + 232.0, + 1092.0, + 265.0, + 1081.0, + 265.0 + ], + "score": 0.624, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1503.0, + 1402.0, + 1503.0, + 1402.0, + 1534.0, + 296.0, + 1534.0 + ], + "score": 0.989, + "text": "As discussed in Appendix B,our method can be easily extended to generate robust counterfactual" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1531.0, + 762.0, + 1531.0, + 762.0, + 1564.0, + 293.0, + 1564.0 + ], + "score": 0.999, + "text": "explanations for node classification tasks." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1030.0, + 767.0, + 1030.0, + 767.0, + 1071.0, + 294.0, + 1071.0 + ], + "score": 0.984, + "text": "Similar to [45], we use a L1 regularization" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1030.0, + 1191.0, + 1030.0, + 1191.0, + 1071.0, + 1038.0, + 1071.0 + ], + "score": 0.997, + "text": "on the matrix" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 1030.0, + 1372.0, + 1030.0, + 1372.0, + 1071.0, + 1228.0, + 1071.0 + ], + "score": 1.0, + "text": "produced by" + }, + { + "category_id": 16, + "poly": [ + 1401.0, + 1030.0, + 1405.0, + 1030.0, + 1405.0, + 1071.0, + 1401.0, + 1071.0 + ], + "score": 0.666, + "text": "（204" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1063.0, + 498.0, + 1063.0, + 498.0, + 1097.0, + 295.0, + 1097.0 + ], + "score": 0.996, + "text": "on an input graph" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1063.0, + 1335.0, + 1063.0, + 1335.0, + 1097.0, + 523.0, + 1097.0 + ], + "score": 0.976, + "text": "to produce a sparse matrix M, such that only a small number of edges in" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 1063.0, + 1405.0, + 1063.0, + 1405.0, + 1097.0, + 1361.0, + 1097.0 + ], + "score": 0.999, + "text": "are" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1088.0, + 1357.0, + 1088.0, + 1357.0, + 1128.0, + 293.0, + 1128.0 + ], + "score": 0.986, + "text": " selected as the counterfactual explanation. We also follow [45] to use an entropy regularization" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1674, + 1404, + 1674, + 1404, + 1881, + 298, + 1881 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1180, + 1404, + 1180, + 1404, + 1401, + 298, + 1401 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 715, + 1404, + 715, + 1404, + 889, + 297, + 889 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 583, + 1404, + 583, + 1404, + 702, + 298, + 702 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 903, + 1404, + 903, + 1404, + 1054, + 298, + 1054 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1414, + 1404, + 1414, + 1404, + 1560, + 298, + 1560 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 463, + 197, + 1235, + 197, + 1235, + 432, + 463, + 432 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 300, + 1573, + 1401, + 1573, + 1401, + 1663, + 300, + 1663 + ], + "score": 0.959 + }, + { + "category_id": 2, + "poly": [ + 295, + 1953, + 1397, + 1953, + 1397, + 2005, + 295, + 2005 + ], + "score": 0.936 + }, + { + "category_id": 4, + "poly": [ + 293, + 471, + 1403, + 471, + 1403, + 503, + 293, + 503 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 297, + 1114, + 452, + 1114, + 452, + 1147, + 297, + 1147 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 858, + 2062, + 858, + 2085, + 840, + 2085 + ], + "score": 0.809 + }, + { + "category_id": 13, + "poly": [ + 1231, + 1297, + 1373, + 1297, + 1373, + 1330, + 1231, + 1330 + ], + "score": 0.93, + "latex": "G = ( \\bar { V , } E )" + }, + { + "category_id": 13, + "poly": [ + 1224, + 672, + 1260, + 672, + 1260, + 705, + 1224, + 705 + ], + "score": 0.9, + "latex": "G _ { \\theta } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 578, + 673, + 615, + 673, + 615, + 702, + 578, + 702 + ], + "score": 0.88, + "latex": "G _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1328, + 589, + 1328, + 589, + 1375, + 296, + 1375 + ], + "score": 0.8, + "latex": "\\begin{array} { r } { s p a r s i t y ( S , G ) = 1 - \\frac { | S | } { | E | } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 850, + 1298, + 872, + 1298, + 872, + 1325, + 850, + 1325 + ], + "score": 0.79, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 358, + 1371, + 383, + 1371, + 383, + 1397, + 358, + 1397 + ], + "score": 0.73, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 519, + 677, + 536, + 677, + 536, + 699, + 519, + 699 + ], + "score": 0.71, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 1165, + 677, + 1182, + 677, + 1182, + 699, + 1165, + 699 + ], + "score": 0.69, + "latex": "c" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 198.0, + 678.0, + 198.0, + 678.0, + 221.0, + 551.0, + 221.0 + ], + "score": 0.982, + "text": "(a) Mutagenicity" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 198.0, + 926.0, + 198.0, + 926.0, + 221.0, + 813.0, + 221.0 + ], + "score": 0.956, + "text": "(b) BA-2motifs" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 199.0, + 1155.0, + 199.0, + 1155.0, + 220.0, + 1090.0, + 220.0 + ], + "score": 0.935, + "text": "(c) NCI1" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 221.0, + 511.0, + 221.0, + 511.0, + 263.0, + 480.0, + 263.0 + ], + "score": 0.812, + "text": "0.7" + }, + { + "category_id": 16, + "poly": [ + 719.0, + 221.0, + 765.0, + 221.0, + 765.0, + 317.0, + 719.0, + 317.0 + ], + "score": 0.292, + "text": "8" + }, + { + "category_id": 16, + "poly": [ + 777.0, + 262.0, + 791.0, + 262.0, + 791.0, + 274.0, + 777.0, + 274.0 + ], + "score": 0.159, + "text": "1" + }, + { + "category_id": 16, + "poly": [ + 972.0, + 231.0, + 1018.0, + 231.0, + 1018.0, + 317.0, + 972.0, + 317.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 234.0, + 1042.0, + 234.0, + 1042.0, + 251.0, + 1030.0, + 251.0 + ], + "score": 0.984, + "text": "F" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 275.0, + 510.0, + 275.0, + 510.0, + 299.0, + 479.0, + 299.0 + ], + "score": 0.999, + "text": "0.4" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 272.0, + 764.0, + 272.0, + 764.0, + 296.0, + 734.0, + 296.0 + ], + "score": 0.999, + "text": "0.3" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 279.0, + 778.0, + 279.0, + 778.0, + 287.0, + 769.0, + 287.0 + ], + "score": 0.65, + "text": "1" + }, + { + "category_id": 16, + "poly": [ + 878.0, + 279.0, + 896.0, + 279.0, + 896.0, + 289.0, + 878.0, + 289.0 + ], + "score": 0.344, + "text": "41" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 279.0, + 918.0, + 279.0, + 918.0, + 287.0, + 908.0, + 287.0 + ], + "score": 0.565, + "text": "I-." + }, + { + "category_id": 15, + "poly": [ + 987.0, + 276.0, + 1019.0, + 276.0, + 1019.0, + 299.0, + 987.0, + 299.0 + ], + "score": 0.999, + "text": "0.3" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 269.0, + 1051.0, + 269.0, + 1051.0, + 281.0, + 1037.0, + 281.0 + ], + "score": 0.92, + "text": "J" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 288.0, + 589.0, + 288.0, + 589.0, + 296.0, + 579.0, + 296.0 + ], + "score": 0.86, + "text": "1" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 284.0, + 1049.0, + 284.0, + 1049.0, + 298.0, + 1037.0, + 298.0 + ], + "score": 0.563, + "text": "F" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 283.0, + 1089.0, + 283.0, + 1089.0, + 290.0, + 1076.0, + 290.0 + ], + "score": 0.548, + "text": "\"-" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 293.0, + 511.0, + 293.0, + 511.0, + 317.0, + 479.0, + 317.0 + ], + "score": 0.998, + "text": "0.3" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 298.0, + 763.0, + 298.0, + 763.0, + 321.0, + 734.0, + 321.0 + ], + "score": 1.0, + "text": "0.2" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 299.0, + 1018.0, + 299.0, + 1018.0, + 319.0, + 988.0, + 319.0 + ], + "score": 0.999, + "text": "0.2" + }, + { + "category_id": 16, + "poly": [ + 1159.0, + 299.0, + 1170.0, + 299.0, + 1170.0, + 306.0, + 1159.0, + 306.0 + ], + "score": 0.198, + "text": "5" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 311.0, + 510.0, + 311.0, + 510.0, + 334.0, + 480.0, + 334.0 + ], + "score": 1.0, + "text": "0.2" + }, + { + "category_id": 16, + "poly": [ + 774.0, + 306.0, + 782.0, + 306.0, + 782.0, + 316.0, + 774.0, + 316.0 + ], + "score": 0.493, + "text": "F" + }, + { + "category_id": 16, + "poly": [ + 1150.0, + 308.0, + 1158.0, + 308.0, + 1158.0, + 316.0, + 1150.0, + 316.0 + ], + "score": 0.435, + "text": "1" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 329.0, + 510.0, + 329.0, + 510.0, + 352.0, + 480.0, + 352.0 + ], + "score": 0.999, + "text": "0.1" + }, + { + "category_id": 16, + "poly": [ + 520.0, + 325.0, + 541.0, + 325.0, + 541.0, + 343.0, + 520.0, + 343.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 324.0, + 762.0, + 324.0, + 762.0, + 344.0, + 734.0, + 344.0 + ], + "score": 0.999, + "text": "0.1" + }, + { + "category_id": 16, + "poly": [ + 955.0, + 328.0, + 971.0, + 328.0, + 971.0, + 340.0, + 955.0, + 340.0 + ], + "score": 0.298, + "text": "1" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 321.0, + 1018.0, + 321.0, + 1018.0, + 344.0, + 987.0, + 344.0 + ], + "score": 0.999, + "text": "0.1" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 330.0, + 1178.0, + 330.0, + 1178.0, + 338.0, + 1170.0, + 338.0 + ], + "score": 0.525, + "text": "1" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 349.0, + 521.0, + 349.0, + 521.0, + 379.0, + 481.0, + 379.0 + ], + "score": 0.891, + "text": "0.075" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 358.0, + 566.0, + 358.0, + 566.0, + 379.0, + 541.0, + 379.0 + ], + "score": 0.993, + "text": "80" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 356.0, + 612.0, + 356.0, + 612.0, + 380.0, + 584.0, + 380.0 + ], + "score": 0.999, + "text": "85" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 357.0, + 656.0, + 357.0, + 656.0, + 380.0, + 628.0, + 380.0 + ], + "score": 0.98, + "text": "90" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 348.0, + 707.0, + 348.0, + 707.0, + 379.0, + 673.0, + 379.0 + ], + "score": 0.962, + "text": "95" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 348.0, + 775.0, + 348.0, + 775.0, + 379.0, + 733.0, + 379.0 + ], + "score": 0.963, + "text": "0.075" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 357.0, + 824.0, + 357.0, + 824.0, + 380.0, + 796.0, + 380.0 + ], + "score": 0.998, + "text": "80" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 356.0, + 869.0, + 356.0, + 869.0, + 380.0, + 842.0, + 380.0 + ], + "score": 0.999, + "text": "85" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 358.0, + 915.0, + 358.0, + 915.0, + 379.0, + 889.0, + 379.0 + ], + "score": 0.991, + "text": "90" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 358.0, + 960.0, + 358.0, + 960.0, + 379.0, + 935.0, + 379.0 + ], + "score": 0.992, + "text": "95" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 344.0, + 1030.0, + 344.0, + 1030.0, + 379.0, + 987.0, + 379.0 + ], + "score": 0.935, + "text": "0.05" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 356.0, + 1077.0, + 356.0, + 1077.0, + 380.0, + 1050.0, + 380.0 + ], + "score": 0.992, + "text": "80" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 356.0, + 1123.0, + 356.0, + 1123.0, + 380.0, + 1095.0, + 380.0 + ], + "score": 0.999, + "text": "85" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 358.0, + 1168.0, + 358.0, + 1168.0, + 379.0, + 1143.0, + 379.0 + ], + "score": 0.998, + "text": "90" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 356.0, + 1214.0, + 356.0, + 1214.0, + 380.0, + 1187.0, + 380.0 + ], + "score": 0.995, + "text": "95" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 372.0, + 657.0, + 372.0, + 657.0, + 396.0, + 573.0, + 396.0 + ], + "score": 0.993, + "text": "Sparsity (%)" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 374.0, + 912.0, + 374.0, + 912.0, + 395.0, + 829.0, + 395.0 + ], + "score": 0.978, + "text": "Sparsity (%)" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 374.0, + 1165.0, + 374.0, + 1165.0, + 395.0, + 1082.0, + 395.0 + ], + "score": 0.988, + "text": "Sparsity (%)" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 394.0, + 709.0, + 394.0, + 709.0, + 414.0, + 611.0, + 414.0 + ], + "score": 0.92, + "text": "-I-GNNExplainer" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 396.0, + 827.0, + 396.0, + 827.0, + 413.0, + 723.0, + 413.0 + ], + "score": 0.929, + "text": "--I-- PGM-Explainer" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 394.0, + 959.0, + 394.0, + 959.0, + 414.0, + 865.0, + 414.0 + ], + "score": 0.999, + "text": "CF-GNNExplainer" + }, + { + "category_id": 16, + "poly": [ + 995.0, + 397.0, + 1009.0, + 397.0, + 1009.0, + 411.0, + 995.0, + 411.0 + ], + "score": 0.348, + "text": "+" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 396.0, + 1118.0, + 396.0, + 1118.0, + 413.0, + 1016.0, + 413.0 + ], + "score": 0.998, + "text": "RCExplainer (Ours)" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 408.0, + 699.0, + 408.0, + 699.0, + 428.0, + 606.0, + 428.0 + ], + "score": 0.965, + "text": "-- PGExplainer" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 412.0, + 744.0, + 412.0, + 744.0, + 422.0, + 727.0, + 422.0 + ], + "score": 0.6, + "text": ".I.." + }, + { + "category_id": 15, + "poly": [ + 746.0, + 406.0, + 811.0, + 406.0, + 811.0, + 430.0, + 746.0, + 430.0 + ], + "score": 0.961, + "text": "SubgraphX" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 409.0, + 976.0, + 409.0, + 976.0, + 428.0, + 839.0, + 428.0 + ], + "score": 0.94, + "text": "--I-- RCExp-NoLDB (Ours)" + }, + { + "category_id": 16, + "poly": [ + 643.75, + 290.0, + 654.75, + 290.0, + 654.75, + 300.5, + 643.75, + 300.5 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 16, + "poly": [ + 947.0, + 287.5, + 970.0, + 287.5, + 970.0, + 303.0, + 947.0, + 303.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 16, + "poly": [ + 908.0, + 301.0, + 928.0, + 301.0, + 928.0, + 312.5, + 908.0, + 312.5 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 16, + "poly": [ + 1173.0, + 303.5, + 1195.0, + 303.5, + 1195.0, + 318.0, + 1173.0, + 318.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1947.0, + 1402.0, + 1947.0, + 1402.0, + 1984.0, + 330.0, + 1984.0 + ], + "score": 0.994, + "text": "3Code available at https: //marketplace.huaweicloud.com/markets/aihub/notebook/detail/" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1979.0, + 831.0, + 1979.0, + 831.0, + 2005.0, + 298.0, + 2005.0 + ], + "score": 0.999, + "text": "?id=e41f63d3-e346-4891-bf6a-40e64b4a3278" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 468.0, + 1409.0, + 468.0, + 1409.0, + 510.0, + 293.0, + 510.0 + ], + "score": 0.989, + "text": "Figure 1: Fidelity performance averaged across 10 runs for the datasets at different levels of sparsity." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1108.0, + 458.0, + 1108.0, + 458.0, + 1158.0, + 292.0, + 1158.0 + ], + "score": 1.0, + "text": "5.1Fidelity" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2090.0, + 838.0, + 2090.0 + ], + "score": 0.999, + "text": "8" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1676.0, + 1405.0, + 1676.0, + 1405.0, + 1709.0, + 294.0, + 1709.0 + ], + "score": 0.99, + "text": "The fidelity performance of SubgraphX reported in [47] was obtained by seting the features of nodes" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1706.0, + 1405.0, + 1706.0, + 1405.0, + 1739.0, + 294.0, + 1739.0 + ], + "score": 0.987, + "text": "that are part of the explanation to Obut not removing the explanation edges from the input graph. This" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1731.0, + 1406.0, + 1731.0, + 1406.0, + 1769.0, + 292.0, + 1769.0 + ], + "score": 0.991, + "text": " does not remove the message passing roles of the explanation nodes from the input graph because the" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1763.0, + 1404.0, + 1763.0, + 1404.0, + 1797.0, + 294.0, + 1797.0 + ], + "score": 0.988, + "text": "edges connected to those nodes still can pass messages. In our experiments, we directly block the" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1792.0, + 1405.0, + 1792.0, + 1405.0, + 1825.0, + 293.0, + 1825.0 + ], + "score": 0.995, + "text": "messages that are passed on the edges in the explanation, which completely prevents the explanation" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1821.0, + 1408.0, + 1821.0, + 1408.0, + 1858.0, + 292.0, + 1858.0 + ], + "score": 0.99, + "text": " nodes in the input graph to participate in the message passing. As a result, the performance of" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1848.0, + 648.0, + 1848.0, + 648.0, + 1886.0, + 295.0, + 1886.0 + ], + "score": 0.998, + "text": "SubgraphX drops significantly." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1178.0, + 1406.0, + 1178.0, + 1406.0, + 1215.0, + 293.0, + 1215.0 + ], + "score": 0.994, + "text": "Fidelity is measured by the decrease of prediction confidence after removing the explanation (i.e., a" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1209.0, + 1405.0, + 1209.0, + 1405.0, + 1244.0, + 294.0, + 1244.0 + ], + "score": 0.99, + "text": " set of edges) from the input graph [32]. We use fidelity to evaluate how counterfactual the generated" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1273.0, + 294.0, + 1273.0 + ], + "score": 0.987, + "text": "explanations are on the datasets Mutagenicity,NCI1 and BA-2motifs.A large fidelity score indicates" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1264.0, + 1405.0, + 1264.0, + 1405.0, + 1304.0, + 292.0, + 1304.0 + ], + "score": 0.986, + "text": " stronger counterfactual characteristics. It is important to note that fidelity may be sensitive to sparsity" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1294.0, + 849.0, + 1294.0, + 849.0, + 1332.0, + 292.0, + 1332.0 + ], + "score": 0.986, + "text": "of explanations. The sparsity of an explanation" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1294.0, + 1230.0, + 1294.0, + 1230.0, + 1332.0, + 873.0, + 1332.0 + ], + "score": 0.989, + "text": "with respect to an input graph" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 1294.0, + 1406.0, + 1294.0, + 1406.0, + 1332.0, + 1374.0, + 1332.0 + ], + "score": 0.997, + "text": "is" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1322.0, + 1411.0, + 1322.0, + 1411.0, + 1382.0, + 590.0, + 1382.0 + ], + "score": 0.727, + "text": ",thatis,tpeeefdgafrta" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1366.0, + 357.0, + 1366.0, + 357.0, + 1406.0, + 293.0, + 1406.0 + ], + "score": 1.0, + "text": "from" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1366.0, + 1091.0, + 1366.0, + 1091.0, + 1406.0, + 384.0, + 1406.0 + ], + "score": 0.978, + "text": ". We only compare explanations with the same level of sparsity." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 715.0, + 1406.0, + 715.0, + 1406.0, + 749.0, + 295.0, + 749.0 + ], + "score": 0.994, + "text": "We evaluate the explanation performance on two typical tasks: the graph clasification task that uses" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 745.0, + 1406.0, + 745.0, + 1406.0, + 779.0, + 294.0, + 779.0 + ], + "score": 0.991, + "text": "a GNN to predict the class of an input graph,and the node classification task that uses a GNN to" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 771.0, + 1409.0, + 771.0, + 1409.0, + 809.0, + 292.0, + 809.0 + ], + "score": 0.991, + "text": " predict the class of a graph node. For the graph classification task, we use one synthetic dataset," + }, + { + "category_id": 15, + "poly": [ + 292.0, + 800.0, + 1407.0, + 800.0, + 1407.0, + 837.0, + 292.0, + 837.0 + ], + "score": 0.984, + "text": "BA-2motifs [25],and two real-world datasets,Mutagenicity [21] and NCI1 [39].For the node" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 828.0, + 1408.0, + 828.0, + 1408.0, + 866.0, + 294.0, + 866.0 + ], + "score": 0.991, + "text": "classification task, we use the same four synthetic datasets as used by GNNExplainer [45], namely," + }, + { + "category_id": 15, + "poly": [ + 294.0, + 858.0, + 1025.0, + 858.0, + 1025.0, + 893.0, + 294.0, + 893.0 + ], + "score": 0.986, + "text": "BA-SHAPES,BA-COMMUNITY,TREE-CYCLES and TREE-GRID." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 585.0, + 1405.0, + 585.0, + 1405.0, + 619.0, + 296.0, + 619.0 + ], + "score": 0.997, + "text": "To demonstrate the effectiveness of the decision regions,we derive another baseline method named" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 614.0, + 1405.0, + 614.0, + 1405.0, + 648.0, + 294.0, + 648.0 + ], + "score": 0.984, + "text": "RCExp-NoLDB that adopts the general framework of RCExplainer but does not use the LDBs of" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 642.0, + 1405.0, + 642.0, + 1405.0, + 677.0, + 296.0, + 677.0 + ], + "score": 0.997, + "text": "decision regions to generate explanations.Instead, RCExp-NoLDB directly maximizes the prediction" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 663.0, + 518.0, + 663.0, + 518.0, + 714.0, + 290.0, + 714.0 + ], + "score": 0.975, + "text": "confidence on class" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 663.0, + 577.0, + 663.0, + 577.0, + 714.0, + 537.0, + 714.0 + ], + "score": 0.999, + "text": "for" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 663.0, + 1164.0, + 663.0, + 1164.0, + 714.0, + 616.0, + 714.0 + ], + "score": 0.981, + "text": "and minimizes the prediction confidence of class" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 663.0, + 1223.0, + 663.0, + 1223.0, + 714.0, + 1183.0, + 714.0 + ], + "score": 0.999, + "text": "for" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 663.0, + 1274.0, + 663.0, + 1274.0, + 714.0, + 1261.0, + 714.0 + ], + "score": 0.813, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 900.0, + 1408.0, + 900.0, + 1408.0, + 939.0, + 293.0, + 939.0 + ], + "score": 0.99, + "text": "Limited by space, we only report here the key results_on the graph classification task for fidelity," + }, + { + "category_id": 15, + "poly": [ + 296.0, + 933.0, + 1405.0, + 933.0, + 1405.0, + 965.0, + 296.0, + 965.0 + ], + "score": 0.986, + "text": "robustness and efficiency. Please refer to Appendix E for details on datasets, baselines and the" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 962.0, + 1406.0, + 962.0, + 1406.0, + 995.0, + 292.0, + 995.0 + ], + "score": 0.989, + "text": "experiment setups.Detailed experimental comparison on the node classification task willbe discussed" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 989.0, + 1405.0, + 989.0, + 1405.0, + 1028.0, + 291.0, + 1028.0 + ], + "score": 0.986, + "text": " in Appendix F where we show that our method produces extremely accurate explanations. The code3" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1022.0, + 533.0, + 1022.0, + 533.0, + 1060.0, + 293.0, + 1060.0 + ], + "score": 0.998, + "text": "is publicly available." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1413.0, + 1404.0, + 1413.0, + 1404.0, + 1449.0, + 294.0, + 1449.0 + ], + "score": 0.989, + "text": "Figure 1 shows the results about fidelity. Our approach achieves the best fidelity performance at all" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1441.0, + 1408.0, + 1441.0, + 1408.0, + 1479.0, + 292.0, + 1479.0 + ], + "score": 0.98, + "text": " levels of sparsity. The results validate the effectiveness of our method in producing highly counter-" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1468.0, + 1406.0, + 1468.0, + 1406.0, + 1509.0, + 289.0, + 1509.0 + ], + "score": 0.987, + "text": "factual explanations. RCExplainer also significantly outperforms RCExp-NoLDB. This confirms" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1499.0, + 1408.0, + 1499.0, + 1408.0, + 1538.0, + 292.0, + 1538.0 + ], + "score": 0.993, + "text": " that using LDBs of decision regions extracted from GNNs produces more faithful counterfactual" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1529.0, + 450.0, + 1529.0, + 450.0, + 1566.0, + 292.0, + 1566.0 + ], + "score": 1.0, + "text": "explanations." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1572.0, + 1405.0, + 1572.0, + 1405.0, + 1608.0, + 295.0, + 1608.0 + ], + "score": 0.984, + "text": "CF-GNNExplainer performs the best among the rest of the methods.This is expected as it optimizes" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1637.0, + 295.0, + 1637.0 + ], + "score": 0.995, + "text": "the counterfactual behavior of the explanations which results in higher fidelity for the explanations in" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1632.0, + 1300.0, + 1632.0, + 1300.0, + 1665.0, + 295.0, + 1665.0 + ], + "score": 0.979, + "text": "comparison to those produced by other methods such as GNNExplainer and PGExplainer." + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 599, + 1406, + 599, + 1406, + 862, + 297, + 862 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1064, + 1404, + 1064, + 1404, + 1269, + 297, + 1269 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1720, + 1404, + 1720, + 1404, + 1839, + 297, + 1839 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 874, + 1404, + 874, + 1404, + 1052, + 297, + 1052 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1281, + 1404, + 1281, + 1404, + 1517, + 297, + 1517 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 464, + 196, + 1235, + 196, + 1235, + 415, + 464, + 415 + ], + "score": 0.965 + }, + { + "category_id": 5, + "poly": [ + 296, + 1547, + 1402, + 1547, + 1402, + 1630, + 296, + 1630 + ], + "score": 0.95, + "html": "
MethodGNNExplainerPGExplainerPGM-ExplainerSubgraphXCF-GNNExplainerRCExplainer
Time1.2s ± 0.20.01s ± 0.0313.1s ± 3.977.8s ± 4.54.6s± 0.20.01s ± 0.02
" + }, + { + "category_id": 1, + "poly": [ + 296, + 1947, + 1401, + 1947, + 1401, + 2007, + 296, + 2007 + ], + "score": 0.945 + }, + { + "category_id": 0, + "poly": [ + 298, + 545, + 656, + 545, + 656, + 577, + 298, + 577 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 298, + 1879, + 509, + 1879, + 509, + 1916, + 298, + 1916 + ], + "score": 0.906 + }, + { + "category_id": 4, + "poly": [ + 296, + 457, + 1403, + 457, + 1403, + 490, + 296, + 490 + ], + "score": 0.851 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 858, + 2061, + 858, + 2084, + 840, + 2084 + ], + "score": 0.802 + }, + { + "category_id": 6, + "poly": [ + 388, + 1640, + 1307, + 1640, + 1307, + 1672, + 388, + 1672 + ], + "score": 0.577 + }, + { + "category_id": 13, + "poly": [ + 337, + 803, + 407, + 803, + 407, + 831, + 337, + 831 + ], + "score": 0.9, + "latex": "k = 8" + }, + { + "category_id": 13, + "poly": [ + 938, + 688, + 969, + 688, + 969, + 715, + 938, + 715 + ], + "score": 0.87, + "latex": "G ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 873, + 717, + 901, + 717, + 901, + 744, + 873, + 744 + ], + "score": 0.86, + "latex": "S ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 500, + 660, + 532, + 660, + 532, + 686, + 500, + 686 + ], + "score": 0.85, + "latex": "G ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 616, + 745, + 645, + 745, + 645, + 773, + 616, + 773 + ], + "score": 0.85, + "latex": "S ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 942, + 717, + 973, + 717, + 973, + 745, + 942, + 745 + ], + "score": 0.84, + "latex": "G ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 856, + 805, + 874, + 805, + 874, + 830, + 856, + 830 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1306, + 718, + 1328, + 718, + 1328, + 744, + 1306, + 744 + ], + "score": 0.82, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 998, + 631, + 1023, + 631, + 1023, + 657, + 998, + 657 + ], + "score": 0.81, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 1180, + 718, + 1199, + 718, + 1199, + 744, + 1180, + 744 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1242, + 631, + 1263, + 631, + 1263, + 657, + 1242, + 657 + ], + "score": 0.78, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 1367, + 689, + 1394, + 689, + 1394, + 715, + 1367, + 715 + ], + "score": 0.78, + "latex": "G" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 199.0, + 679.0, + 199.0, + 679.0, + 222.0, + 552.0, + 222.0 + ], + "score": 0.974, + "text": "(a) Mutagenicity" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 199.0, + 926.0, + 199.0, + 926.0, + 221.0, + 813.0, + 221.0 + ], + "score": 0.984, + "text": "(b) BA-2motifs" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 197.0, + 1157.0, + 197.0, + 1157.0, + 223.0, + 1087.0, + 223.0 + ], + "score": 0.952, + "text": "(c) NCI1" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 215.0, + 516.0, + 215.0, + 516.0, + 237.0, + 480.0, + 237.0 + ], + "score": 1.0, + "text": "1.0" + }, + { + "category_id": 16, + "poly": [ + 518.0, + 224.0, + 529.0, + 224.0, + 529.0, + 232.0, + 518.0, + 232.0 + ], + "score": 0.341, + "text": "1" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 216.0, + 768.0, + 216.0, + 768.0, + 237.0, + 733.0, + 237.0 + ], + "score": 0.999, + "text": "1.0" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 216.0, + 1020.0, + 216.0, + 1020.0, + 237.0, + 988.0, + 237.0 + ], + "score": 0.999, + "text": "1.0" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 240.0, + 513.0, + 240.0, + 513.0, + 266.0, + 478.0, + 266.0 + ], + "score": 1.0, + "text": "0.9" + }, + { + "category_id": 16, + "poly": [ + 549.0, + 258.0, + 556.0, + 258.0, + 556.0, + 267.0, + 549.0, + 267.0 + ], + "score": 0.424, + "text": "I" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 242.0, + 765.0, + 242.0, + 765.0, + 264.0, + 733.0, + 264.0 + ], + "score": 1.0, + "text": "0.9" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 241.0, + 1019.0, + 241.0, + 1019.0, + 264.0, + 987.0, + 264.0 + ], + "score": 1.0, + "text": "0.9" + }, + { + "category_id": 16, + "poly": [ + 712.0, + 263.0, + 772.0, + 263.0, + 772.0, + 314.0, + 712.0, + 314.0 + ], + "score": 0.47, + "text": "0" + }, + { + "category_id": 16, + "poly": [ + 895.0, + 268.0, + 911.0, + 268.0, + 911.0, + 280.0, + 895.0, + 280.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 293.0, + 512.0, + 293.0, + 512.0, + 370.0, + 478.0, + 370.0 + ], + "score": 0.653, + "text": "5" + }, + { + "category_id": 16, + "poly": [ + 517.0, + 281.0, + 525.0, + 281.0, + 525.0, + 290.0, + 517.0, + 290.0 + ], + "score": 0.382, + "text": "F." + }, + { + "category_id": 16, + "poly": [ + 577.0, + 311.0, + 589.0, + 311.0, + 589.0, + 324.0, + 577.0, + 324.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 307.0, + 623.0, + 307.0, + 623.0, + 334.0, + 610.0, + 334.0 + ], + "score": 0.717, + "text": "小" + }, + { + "category_id": 16, + "poly": [ + 641.0, + 315.0, + 664.0, + 315.0, + 664.0, + 344.0, + 641.0, + 344.0 + ], + "score": 0.291, + "text": "+" + }, + { + "category_id": 16, + "poly": [ + 687.0, + 334.0, + 718.0, + 334.0, + 718.0, + 351.0, + 687.0, + 351.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 293.0, + 766.0, + 293.0, + 766.0, + 370.0, + 732.0, + 370.0 + ], + "score": 0.678, + "text": "5" + }, + { + "category_id": 16, + "poly": [ + 798.0, + 303.0, + 811.0, + 303.0, + 811.0, + 313.0, + 798.0, + 313.0 + ], + "score": 0.177, + "text": "1" + }, + { + "category_id": 16, + "poly": [ + 830.0, + 306.0, + 844.0, + 306.0, + 844.0, + 317.0, + 830.0, + 317.0 + ], + "score": 0.275, + "text": "+" + }, + { + "category_id": 16, + "poly": [ + 861.0, + 274.0, + 948.0, + 274.0, + 948.0, + 347.0, + 861.0, + 347.0 + ], + "score": 0.438, + "text": "N" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 285.0, + 975.0, + 285.0, + 975.0, + 357.0, + 949.0, + 357.0 + ], + "score": 0.581, + "text": "IT.TT" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 290.0, + 1022.0, + 290.0, + 1022.0, + 365.0, + 986.0, + 365.0 + ], + "score": 0.557, + "text": "65" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 283.0, + 1065.0, + 283.0, + 1065.0, + 292.0, + 1055.0, + 292.0 + ], + "score": 0.884, + "text": "1" + }, + { + "category_id": 16, + "poly": [ + 1086.0, + 288.0, + 1099.0, + 288.0, + 1099.0, + 320.0, + 1086.0, + 320.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 325.0, + 1130.0, + 325.0, + 1130.0, + 334.0, + 1121.0, + 334.0 + ], + "score": 0.632, + "text": "王-" + }, + { + "category_id": 16, + "poly": [ + 1148.0, + 330.0, + 1168.0, + 330.0, + 1168.0, + 345.0, + 1148.0, + 345.0 + ], + "score": 0.316, + "text": "中" + }, + { + "category_id": 16, + "poly": [ + 1182.0, + 341.0, + 1193.0, + 341.0, + 1193.0, + 348.0, + 1182.0, + 348.0 + ], + "score": 0.288, + "text": "于" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 345.0, + 1224.0, + 345.0, + 1224.0, + 354.0, + 1215.0, + 354.0 + ], + "score": 0.681, + "text": "4" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 359.0, + 527.0, + 359.0, + 527.0, + 378.0, + 511.0, + 378.0 + ], + "score": 0.956, + "text": "0" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 359.0, + 596.0, + 359.0, + 596.0, + 379.0, + 571.0, + 379.0 + ], + "score": 0.996, + "text": "10" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 357.0, + 662.0, + 357.0, + 662.0, + 380.0, + 634.0, + 380.0 + ], + "score": 0.997, + "text": "20" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 359.0, + 726.0, + 359.0, + 726.0, + 379.0, + 700.0, + 379.0 + ], + "score": 0.995, + "text": "30" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 359.0, + 781.0, + 359.0, + 781.0, + 378.0, + 765.0, + 378.0 + ], + "score": 0.949, + "text": "0" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 359.0, + 851.0, + 359.0, + 851.0, + 379.0, + 825.0, + 379.0 + ], + "score": 0.998, + "text": "10" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 359.0, + 914.0, + 359.0, + 914.0, + 379.0, + 889.0, + 379.0 + ], + "score": 0.998, + "text": "20" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 359.0, + 980.0, + 359.0, + 980.0, + 379.0, + 954.0, + 379.0 + ], + "score": 0.993, + "text": "30" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 359.0, + 1035.0, + 359.0, + 1035.0, + 378.0, + 1019.0, + 378.0 + ], + "score": 0.956, + "text": "0" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 359.0, + 1105.0, + 359.0, + 1105.0, + 379.0, + 1079.0, + 379.0 + ], + "score": 0.998, + "text": "10" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 357.0, + 1169.0, + 357.0, + 1169.0, + 381.0, + 1142.0, + 381.0 + ], + "score": 0.998, + "text": "20" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 359.0, + 1233.0, + 359.0, + 1233.0, + 379.0, + 1207.0, + 379.0 + ], + "score": 0.997, + "text": "30" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 373.0, + 651.0, + 373.0, + 651.0, + 396.0, + 580.0, + 396.0 + ], + "score": 0.991, + "text": "Noise (%)" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 373.0, + 905.0, + 373.0, + 905.0, + 396.0, + 833.0, + 396.0 + ], + "score": 0.995, + "text": "Noise (%)" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 373.0, + 1158.0, + 373.0, + 1158.0, + 396.0, + 1087.0, + 396.0 + ], + "score": 0.997, + "text": "Noise (%)" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 395.0, + 658.0, + 395.0, + 658.0, + 414.0, + 552.0, + 414.0 + ], + "score": 0.893, + "text": "-I-- GNNExplainer" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 393.0, + 765.0, + 393.0, + 765.0, + 415.0, + 668.0, + 415.0 + ], + "score": 0.898, + "text": "-- PGExplainer" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 395.0, + 897.0, + 395.0, + 897.0, + 414.0, + 776.0, + 414.0 + ], + "score": 0.913, + "text": "-I- CF-GNNExplainer" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 395.0, + 1047.0, + 395.0, + 1047.0, + 414.0, + 908.0, + 414.0 + ], + "score": 0.949, + "text": "--I-- RCExp-NoLDB (Ours)" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 395.0, + 1188.0, + 395.0, + 1188.0, + 414.0, + 1061.0, + 414.0 + ], + "score": 0.994, + "text": "RCExplainer (Ours)" + }, + { + "category_id": 16, + "poly": [ + 536.0, + 245.0, + 717.0, + 245.0, + 717.0, + 274.5, + 536.0, + 274.5 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 16, + "poly": [ + 1038.0, + 241.0, + 1070.0, + 241.0, + 1070.0, + 257.0, + 1038.0, + 257.0 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 16, + "poly": [ + 1083.0, + 274.5, + 1099.0, + 274.5, + 1099.0, + 289.5, + 1083.0, + 289.5 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 262.0, + 527.0, + 262.0, + 527.0, + 303.0, + 440.0, + 303.0 + ], + "score": 0.732, + "text": "8" + }, + { + "category_id": 16, + "poly": [ + 948.0, + 262.5, + 1035.0, + 262.5, + 1035.0, + 302.5, + 948.0, + 302.5 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 16, + "poly": [ + 1114.25, + 302.5, + 1223.25, + 302.5, + 1223.25, + 328.5, + 1114.25, + 328.5 + ], + "score": 0.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 541.0, + 660.0, + 541.0, + 660.0, + 583.0, + 293.0, + 583.0 + ], + "score": 1.0, + "text": "5.2Robustness Performance" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1873.0, + 513.0, + 1873.0, + 513.0, + 1924.0, + 291.0, + 1924.0 + ], + "score": 1.0, + "text": "6Conclusion" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 453.0, + 1409.0, + 453.0, + 1409.0, + 496.0, + 294.0, + 496.0 + ], + "score": 0.992, + "text": "Figure 2: Noise robustness (AUC) averaged across 10 runs for the datasets at diferent levels of noise." + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 860.0, + 2060.0, + 860.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "9" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1634.0, + 1310.0, + 1634.0, + 1310.0, + 1679.0, + 386.0, + 1679.0 + ], + "score": 0.997, + "text": "Table 1: Average time cost for producing an explanation on a single graph sample." + }, + { + "category_id": 15, + "poly": [ + 291.0, + 597.0, + 1406.0, + 597.0, + 1406.0, + 637.0, + 291.0, + 637.0 + ], + "score": 0.988, + "text": " In this experiment, we evaluate the robustness of all methods by quantifying how much an explanation" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 630.0, + 997.0, + 630.0, + 997.0, + 665.0, + 295.0, + 665.0 + ], + "score": 0.992, + "text": "changes after adding noise to the input graph. For an input graph" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 630.0, + 1241.0, + 630.0, + 1241.0, + 665.0, + 1024.0, + 665.0 + ], + "score": 0.995, + "text": "and the explanation" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 630.0, + 1404.0, + 630.0, + 1404.0, + 665.0, + 1264.0, + 665.0 + ], + "score": 0.959, + "text": ",we produce" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 655.0, + 499.0, + 655.0, + 499.0, + 699.0, + 290.0, + 699.0 + ], + "score": 0.948, + "text": " a perturbed graph" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 655.0, + 1408.0, + 655.0, + 1408.0, + 699.0, + 533.0, + 699.0 + ], + "score": 0.988, + "text": "by adding random noise to the node features and randomly adding or deleting" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 686.0, + 937.0, + 686.0, + 937.0, + 722.0, + 292.0, + 722.0 + ], + "score": 0.986, + "text": " some edges of the input graph such that the prediction on" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 686.0, + 1366.0, + 686.0, + 1366.0, + 722.0, + 970.0, + 722.0 + ], + "score": 0.997, + "text": "is consistent with the prediction on" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 686.0, + 1408.0, + 686.0, + 1408.0, + 722.0, + 1395.0, + 722.0 + ], + "score": 0.794, + "text": "：" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 714.0, + 872.0, + 714.0, + 872.0, + 753.0, + 294.0, + 753.0 + ], + "score": 0.981, + "text": "Using the same method we obtain the explanation" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 714.0, + 941.0, + 714.0, + 941.0, + 753.0, + 902.0, + 753.0 + ], + "score": 0.997, + "text": "on" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 714.0, + 1179.0, + 714.0, + 1179.0, + 753.0, + 974.0, + 753.0 + ], + "score": 0.975, + "text": ". Considering top-" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 714.0, + 1305.0, + 714.0, + 1305.0, + 753.0, + 1200.0, + 753.0 + ], + "score": 0.997, + "text": "edges of" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 714.0, + 1408.0, + 714.0, + 1408.0, + 753.0, + 1329.0, + 753.0 + ], + "score": 0.997, + "text": "as the" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 744.0, + 615.0, + 744.0, + 615.0, + 780.0, + 294.0, + 780.0 + ], + "score": 0.989, + "text": " ground-truth and comparing" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 744.0, + 1406.0, + 744.0, + 1406.0, + 780.0, + 646.0, + 780.0 + ], + "score": 0.987, + "text": " against them, we compute a receiver operating characteristic (ROC)" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 773.0, + 1404.0, + 773.0, + 1404.0, + 809.0, + 294.0, + 809.0 + ], + "score": 0.992, + "text": "curve and evaluate the robustnessby the area under curve (AUC) of the ROC curve. We report results" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 802.0, + 336.0, + 802.0, + 336.0, + 838.0, + 294.0, + 838.0 + ], + "score": 1.0, + "text": "for" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 802.0, + 855.0, + 802.0, + 855.0, + 838.0, + 408.0, + 838.0 + ], + "score": 0.985, + "text": "in Figure 2. Results for other values of" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 802.0, + 1406.0, + 802.0, + 1406.0, + 838.0, + 875.0, + 838.0 + ], + "score": 0.99, + "text": "are included in Appendix F where we observe" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 831.0, + 452.0, + 831.0, + 452.0, + 864.0, + 295.0, + 864.0 + ], + "score": 0.998, + "text": "similar trend." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1063.0, + 1406.0, + 1063.0, + 1406.0, + 1100.0, + 294.0, + 1100.0 + ], + "score": 0.996, + "text": "GNNExplainer performs the worst on most of the datasets,since it optimizes each graph independently" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1094.0, + 1406.0, + 1094.0, + 1406.0, + 1127.0, + 292.0, + 1127.0 + ], + "score": 0.994, + "text": "without considering other graphs in the training set. Even when no noise is added,the AUC of" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1120.0, + 1406.0, + 1120.0, + 1406.0, + 1158.0, + 292.0, + 1158.0 + ], + "score": 0.994, + "text": " GNNExplainer is significantly lower than 1 because different runs produce different explanations" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1149.0, + 1408.0, + 1149.0, + 1408.0, + 1188.0, + 292.0, + 1188.0 + ], + "score": 0.998, + "text": "for the same graph prediction. PGExplainer is generally more robust than GNNExplainer because" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1178.0, + 1406.0, + 1178.0, + 1406.0, + 1214.0, + 292.0, + 1214.0 + ], + "score": 0.989, + "text": " the neural network they trained to produce explanations implicitly considers all the graphs used" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1207.0, + 1406.0, + 1207.0, + 1406.0, + 1245.0, + 292.0, + 1245.0 + ], + "score": 0.991, + "text": "for training. CF-GNNExplainer also performs worse than RCExplainer, which means it is more" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1239.0, + 892.0, + 1239.0, + 892.0, + 1273.0, + 295.0, + 1273.0 + ], + "score": 0.997, + "text": "susceptible to the noise as compared to RCExplainer." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1718.0, + 1406.0, + 1718.0, + 1406.0, + 1758.0, + 294.0, + 1758.0 + ], + "score": 0.998, + "text": "Effciency. We evaluate efficiency by comparing the average computation time taken for inference" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1749.0, + 1405.0, + 1749.0, + 1405.0, + 1784.0, + 295.0, + 1784.0 + ], + "score": 0.98, + "text": "on unseen graph samples. Table 1 shows the results on the Mutagenicity dataset. Since our method" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1777.0, + 1404.0, + 1777.0, + 1404.0, + 1811.0, + 296.0, + 1811.0 + ], + "score": 0.978, + "text": "also can be directly used for unseen data without any retraining,it is as eficient as PGExplainer and" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1806.0, + 1337.0, + 1806.0, + 1337.0, + 1844.0, + 292.0, + 1844.0 + ], + "score": 0.978, + "text": "significantly faster than GNNExplainer, PGM-Explainer, SubgraphX and CF-GNNExplainer." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 910.0, + 294.0, + 910.0 + ], + "score": 0.984, + "text": "Figure 2 shows the AUC of GNNExplainer, PGExplainer, RCExp-NoLDB and RCExplainer at" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 902.0, + 1406.0, + 902.0, + 1406.0, + 941.0, + 294.0, + 941.0 + ], + "score": 0.987, + "text": "different levels of noise. A higher AUC indicates better robustness. The percentage of noise" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 933.0, + 1406.0, + 933.0, + 1406.0, + 968.0, + 295.0, + 968.0 + ], + "score": 0.985, + "text": " shows the proportion of nodes and edges that are modified. Baselines such as PGM-Explainer and" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 959.0, + 1406.0, + 959.0, + 1406.0, + 999.0, + 294.0, + 999.0 + ], + "score": 0.986, + "text": " SubgraphX are not included in this experiment as they do not output the edge weights that are required" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 992.0, + 1406.0, + 992.0, + 1406.0, + 1026.0, + 295.0, + 1026.0 + ], + "score": 0.991, + "text": "for computing AUC.We present additional robustness experiments in Appendix F where we extend" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1020.0, + 918.0, + 1020.0, + 918.0, + 1057.0, + 294.0, + 1057.0 + ], + "score": 0.988, + "text": " all the baselines to report node and edge level accuracy." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1282.0, + 1404.0, + 1282.0, + 1404.0, + 1315.0, + 295.0, + 1315.0 + ], + "score": 0.99, + "text": "Our method achieves the best AUC on allthe datasets, because the common decision logic carried" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1310.0, + 1406.0, + 1310.0, + 1406.0, + 1345.0, + 292.0, + 1345.0 + ], + "score": 0.987, + "text": "by the decision regions of a GNN is highly robust to noise. PGExplainer achieves a comparable" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1340.0, + 1409.0, + 1340.0, + 1409.0, + 1374.0, + 291.0, + 1374.0 + ], + "score": 0.985, + "text": "performance as our method on the Mutagenicity dataset, because the samples of this dataset share a" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1367.0, + 1406.0, + 1367.0, + 1406.0, + 1404.0, + 294.0, + 1404.0 + ], + "score": 0.986, + "text": "lot of common structures such as carbon rings, which makes it easier for the neural network trained" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1396.0, + 1406.0, + 1396.0, + 1406.0, + 1433.0, + 294.0, + 1433.0 + ], + "score": 0.984, + "text": "by PGExplainer to identify these structures in presence of noise. However, for BA-2motifs and" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1425.0, + 1406.0, + 1425.0, + 1406.0, + 1462.0, + 294.0, + 1462.0 + ], + "score": 0.989, + "text": "NCI1, this is harder as samples share very few structures and thus the AUC of PGExplainer drops" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1454.0, + 1406.0, + 1454.0, + 1406.0, + 1490.0, + 292.0, + 1490.0 + ], + "score": 0.985, + "text": " significantly. RCExplainer also significantly outperforms RCExp-NoLDB on these datasets which" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1486.0, + 1175.0, + 1486.0, + 1175.0, + 1518.0, + 295.0, + 1518.0 + ], + "score": 0.989, + "text": "highlights the role of decision boundaries in making our method highly robust." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1947.0, + 1406.0, + 1947.0, + 1406.0, + 1982.0, + 294.0, + 1982.0 + ], + "score": 0.988, + "text": " In this paper, we develop a novel method for producing counterfactual explanations on GNNs." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 0.989, + "text": "We extract decision boundaries from the given GNN model to formulate an intuitive and effective" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 420, + 1405, + 420, + 1405, + 654, + 297, + 654 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 202, + 1403, + 202, + 1403, + 407, + 298, + 407 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 766, + 1404, + 766, + 1404, + 884, + 298, + 884 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 304, + 965, + 1408, + 965, + 1408, + 2011, + 304, + 2011 + ], + "score": 0.966 + }, + { + "category_id": 0, + "poly": [ + 298, + 696, + 944, + 696, + 944, + 734, + 298, + 734 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 299, + 925, + 454, + 925, + 454, + 960, + 299, + 960 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.822 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 691.0, + 948.0, + 691.0, + 948.0, + 744.0, + 293.0, + 744.0 + ], + "score": 0.996, + "text": "Acknowledgments and Disclosure of Funding" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 922.0, + 459.0, + 922.0, + 459.0, + 966.0, + 294.0, + 966.0 + ], + "score": 1.0, + "text": "References" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "10" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 420.0, + 1405.0, + 420.0, + 1405.0, + 455.0, + 295.0, + 455.0 + ], + "score": 0.985, + "text": "Our method will benefit multiple fields where GNNs are intensively used. By allowing the users to" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 450.0, + 1405.0, + 450.0, + 1405.0, + 486.0, + 294.0, + 486.0 + ], + "score": 0.996, + "text": " interpret the predictions of complex GNNs better, it will promote transparency, trust and fairness" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 476.0, + 1407.0, + 476.0, + 1407.0, + 515.0, + 292.0, + 515.0 + ], + "score": 0.991, + "text": " in the society. However, there also exist some inherent risks. A generated explanation may expose" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 507.0, + 1408.0, + 507.0, + 1408.0, + 542.0, + 291.0, + 542.0 + ], + "score": 0.985, + "text": " private information if our method is not coupled with an adequate privacy protection technique. Also," + }, + { + "category_id": 15, + "poly": [ + 295.0, + 537.0, + 1407.0, + 537.0, + 1407.0, + 570.0, + 295.0, + 570.0 + ], + "score": 0.989, + "text": " some of the ideas presented in this paper may be adopted and extended to improve adversarial attacks." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 564.0, + 1405.0, + 564.0, + 1405.0, + 601.0, + 295.0, + 601.0 + ], + "score": 0.993, + "text": "Without appropriate defense mechanisms,the misuse of such attcks poses a risk of disruption in the" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 594.0, + 1403.0, + 594.0, + 1403.0, + 627.0, + 295.0, + 627.0 + ], + "score": 0.979, + "text": "functionality of GNNs deployed in the real world.That said, we firmly believe that these risks can be" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 622.0, + 1006.0, + 622.0, + 1006.0, + 658.0, + 295.0, + 658.0 + ], + "score": 0.989, + "text": "mitigated through increased awareness and proactive measures." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 201.0, + 1407.0, + 201.0, + 1407.0, + 237.0, + 292.0, + 237.0 + ], + "score": 0.989, + "text": "counterfactual loss function. We optimize this loss to train a neural network to produce explanations" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 232.0, + 1405.0, + 232.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 0.989, + "text": "with strong counterfactual characteristics. Since the decision boundaries are shared by multiple" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1407.0, + 260.0, + 1407.0, + 294.0, + 294.0, + 294.0 + ], + "score": 0.986, + "text": " samples of the same predicted class, explanations produced by our method are robust and do not" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 289.0, + 1405.0, + 289.0, + 1405.0, + 324.0, + 292.0, + 324.0 + ], + "score": 0.987, + "text": "overfit the noise. Our experiments on synthetic and real-life benchmark datasets strongly validate" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 318.0, + 1407.0, + 318.0, + 1407.0, + 352.0, + 293.0, + 352.0 + ], + "score": 0.982, + "text": "the efficacy of our method. In this work,we focus on GNNs that belong to Piecewise Linear" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 347.0, + 1407.0, + 347.0, + 1407.0, + 381.0, + 293.0, + 381.0 + ], + "score": 0.994, + "text": "Neural Networks (PLNNs). Extending our method to other families of GNNs and tasks such as link" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 377.0, + 859.0, + 377.0, + 859.0, + 411.0, + 294.0, + 411.0 + ], + "score": 0.999, + "text": "prediction,remains an interesting future direction." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 764.0, + 1408.0, + 764.0, + 1408.0, + 806.0, + 293.0, + 806.0 + ], + "score": 0.993, + "text": "Lingyang Chu's research is supported in part by the startup grant provided by the Department" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 795.0, + 1405.0, + 795.0, + 1405.0, + 829.0, + 296.0, + 829.0 + ], + "score": 0.984, + "text": "of Computing and Software of McMaster University. All opinions, findings,conclusions,and" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 821.0, + 1405.0, + 821.0, + 1405.0, + 859.0, + 293.0, + 859.0 + ], + "score": 0.989, + "text": "recommendations in this paper are those of the authors and do not necessarily reflect the views of the" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 851.0, + 498.0, + 851.0, + 498.0, + 889.0, + 294.0, + 889.0 + ], + "score": 1.0, + "text": "funding agencies." + }, + { + "category_id": 15, + "poly": [ + 304.0, + 974.0, + 1409.0, + 974.0, + 1409.0, + 1016.0, + 304.0, + 1016.0 + ], + "score": 0.961, + "text": "[1] J. Adebayo,J. Gilmer,M.Muelly,I. Goodfellow,M. Hardt,and B. Kim. Sanity checks for saliency maps." + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1003.0, + 916.0, + 1003.0, + 916.0, + 1040.0, + 348.0, + 1040.0 + ], + "score": 0.982, + "text": "In Advances in Neural Information Processing Systems." + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1055.0, + 1407.0, + 1055.0, + 1407.0, + 1093.0, + 305.0, + 1093.0 + ], + "score": 0.969, + "text": "[2] A.Bojchevski and S. Gunnemann. Certifiable robustness to graph perturbations.arXiv preprint" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1083.0, + 610.0, + 1083.0, + 610.0, + 1116.0, + 348.0, + 1116.0 + ], + "score": 0.998, + "text": "arXiv:1910.14356,2019." + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1134.0, + 1407.0, + 1134.0, + 1407.0, + 1171.0, + 305.0, + 1171.0 + ], + "score": 0.982, + "text": "[3]E. Cho,S.A. Myers,and J.Leskovec.Friendship and mobility: user movement in location-based social" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1159.0, + 1407.0, + 1159.0, + 1407.0, + 1200.0, + 347.0, + 1200.0 + ], + "score": 0.98, + "text": "networks. In Proceedings of the 17th ACM SIGkDD international conference on Knowledge discovery" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1188.0, + 779.0, + 1188.0, + 779.0, + 1222.0, + 349.0, + 1222.0 + ], + "score": 0.983, + "text": "and data mining,pages 1082-1090,2011." + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1241.0, + 1405.0, + 1241.0, + 1405.0, + 1276.0, + 308.0, + 1276.0 + ], + "score": 0.978, + "text": "[4]L.Chu, X. Hu,J. Hu,L. Wang,and J.Pei. Exact and consistent interpretation for piecewise linear neural" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1264.0, + 1407.0, + 1264.0, + 1407.0, + 1305.0, + 347.0, + 1305.0 + ], + "score": 0.978, + "text": "networks: A closed form solution. In Proceedings of the 24th ACM SiGKDD International Conference on" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1294.0, + 993.0, + 1294.0, + 993.0, + 1329.0, + 349.0, + 1329.0 + ], + "score": 0.986, + "text": "Knowledge Discovery & Data Mining,pages 1244-1253,2018." + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1382.0, + 305.0, + 1382.0 + ], + "score": 0.984, + "text": "[5]V. Crawford,A. Kuhnle,and M. Thai. Submodular cost submodular cover with an approximate oracle.In" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1372.0, + 1171.0, + 1372.0, + 1171.0, + 1408.0, + 346.0, + 1408.0 + ], + "score": 0.989, + "text": " International Conference on Machine Learning, pages 1426-1435.PMLR,2019." + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1424.0, + 1405.0, + 1424.0, + 1405.0, + 1461.0, + 305.0, + 1461.0 + ], + "score": 0.97, + "text": "[6] A. K. Debnath, R.L.Lopez de Compadre, G. Debnath, A. J. Shusterman,and C. Hansch. Structure-activity" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1451.0, + 1404.0, + 1451.0, + 1404.0, + 1486.0, + 349.0, + 1486.0 + ], + "score": 0.994, + "text": "relationship of mutagenic aromatic and heteroaromatic nitro compounds.correlation with molecular orbital" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1478.0, + 1197.0, + 1478.0, + 1197.0, + 1514.0, + 348.0, + 1514.0 + ], + "score": 0.993, + "text": "energies and hydrophobicity. Journal of medicinal chemistry,34(2):786-797,1991." + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1530.0, + 1407.0, + 1530.0, + 1407.0, + 1568.0, + 305.0, + 1568.0 + ], + "score": 0.984, + "text": "[7]M.Denil,S.G.Colmenarejo,S. Cabi,D.Saxton,and N. de Freitas. Programmable agents. arXiv preprint" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1559.0, + 611.0, + 1559.0, + 611.0, + 1590.0, + 348.0, + 1590.0 + ], + "score": 0.998, + "text": "arXiv:1706.06383,2017." + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1608.0, + 1408.0, + 1608.0, + 1408.0, + 1646.0, + 305.0, + 1646.0 + ], + "score": 0.985, + "text": "[8]Y. Duan, M. Andrychowicz,B.C. Stadie,J. Ho,J.Schneider,I.Sutskever,P.Abbeel,and W. Zaremba." + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1636.0, + 1049.0, + 1636.0, + 1049.0, + 1671.0, + 349.0, + 1671.0 + ], + "score": 0.986, + "text": "One-shot imitation learning. arXiv preprint arXiv:1703.07326,2017." + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1687.0, + 1408.0, + 1687.0, + 1408.0, + 1725.0, + 305.0, + 1725.0 + ], + "score": 0.968, + "text": "[9]W.Fan,Y.Ma,Q.Li,Y. He,E. Zhao,J.Tang,and D.Yin. Graph neural networks for social recommendation." + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1716.0, + 945.0, + 1716.0, + 945.0, + 1751.0, + 348.0, + 1751.0 + ], + "score": 0.992, + "text": "In The World Wide Web Conference,pages 417-426,2019." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1762.0, + 1409.0, + 1762.0, + 1409.0, + 1808.0, + 295.0, + 1808.0 + ], + "score": 0.93, + "text": "[10]T.Freiesleben.Counterfactualexplanations &adversarialexamplescommongrounds,essntialdifferences," + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1794.0, + 994.0, + 1794.0, + 994.0, + 1829.0, + 349.0, + 1829.0 + ], + "score": 0.981, + "text": "and potential transfers. arXiv preprint arXiv:2009.05487, 2020." + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1845.0, + 1405.0, + 1845.0, + 1405.0, + 1884.0, + 297.0, + 1884.0 + ], + "score": 0.973, + "text": "[11]A. Ghorbani,A. Abid,and J. Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1873.0, + 1096.0, + 1873.0, + 1096.0, + 1908.0, + 349.0, + 1908.0 + ], + "score": 0.98, + "text": "Conference on Artificial Intelligence, volume 33, pages 3681-3688,2019." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1924.0, + 1409.0, + 1924.0, + 1409.0, + 1963.0, + 295.0, + 1963.0 + ], + "score": 0.98, + "text": "[12] X. Glorot and Y. Bengio. Understanding the dificulty of training deep feedforward neural networks." + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1949.0, + 1409.0, + 1949.0, + 1409.0, + 1993.0, + 346.0, + 1993.0 + ], + "score": 0.967, + "text": "InProceedingsofthethirteenth international conference onartificial intelligence and statistics,pages" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1977.0, + 993.0, + 1977.0, + 993.0, + 2015.0, + 348.0, + 2015.0 + ], + "score": 0.994, + "text": "249-256. JMLR Workshop and Conference Proceedings, 2010." + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 292, + 185, + 1410, + 185, + 1410, + 2018, + 292, + 2018 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 835, + 2061, + 864, + 2061, + 864, + 2087, + 835, + 2087 + ], + "score": 0.833 + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2058.0, + 869.0, + 2058.0, + 869.0, + 2097.0, + 831.0, + 2097.0 + ], + "score": 0.999, + "text": "11" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 205.0, + 1406.0, + 205.0, + 1406.0, + 238.0, + 296.0, + 238.0 + ], + "score": 0.977, + "text": "[13] X.Glorot,A.Bordes,and Y.Bengio.Deep sparse rectifier neural networks.In Proceedings ofthe fourteenth" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 230.0, + 1408.0, + 230.0, + 1408.0, + 268.0, + 348.0, + 268.0 + ], + "score": 0.989, + "text": "international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 258.0, + 671.0, + 258.0, + 671.0, + 290.0, + 350.0, + 290.0 + ], + "score": 0.998, + "text": "Conference Proceedings,2011." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 304.0, + 1408.0, + 304.0, + 1408.0, + 340.0, + 292.0, + 340.0 + ], + "score": 0.968, + "text": "[14] I. Goodfelow,D. Warde-Farley,M. Mirza,A. Courvile,and Y.Bengio. Maxout networks.In International" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 330.0, + 1019.0, + 330.0, + 1019.0, + 367.0, + 344.0, + 367.0 + ], + "score": 0.994, + "text": "conference on machine learning, pages 1319-1327.PMLR,2013." + }, + { + "category_id": 15, + "poly": [ + 290.0, + 377.0, + 1408.0, + 377.0, + 1408.0, + 417.0, + 290.0, + 417.0 + ], + "score": 0.982, + "text": "[15]K. He, X. Zhang, S. Ren,and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 403.0, + 1406.0, + 403.0, + 1406.0, + 445.0, + 346.0, + 445.0 + ], + "score": 0.968, + "text": " imagenet classification. In Proceedings ofthe IEEE international conferenceon computer vision, pages" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 435.0, + 536.0, + 435.0, + 536.0, + 461.0, + 352.0, + 461.0 + ], + "score": 0.997, + "text": "1026-1034,2015." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 477.0, + 1408.0, + 477.0, + 1408.0, + 516.0, + 292.0, + 516.0 + ], + "score": 0.985, + "text": "[16] L.Holdijk,M. Boon,S.Henckens,and L.de Jong. [re] parameterized explainer for graph neural network." + }, + { + "category_id": 15, + "poly": [ + 348.0, + 505.0, + 817.0, + 505.0, + 817.0, + 544.0, + 348.0, + 544.0 + ], + "score": 0.957, + "text": "In ML Reproducibility Challenge 2020,2021." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 556.0, + 1406.0, + 556.0, + 1406.0, + 588.0, + 294.0, + 588.0 + ], + "score": 0.963, + "text": "[17] W. Huber, V.J. Carey,L.Long, S. Falcon,and R. Gentleman. Graphs in molecular biology. BMC" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 584.0, + 687.0, + 584.0, + 687.0, + 616.0, + 350.0, + 616.0 + ], + "score": 0.998, + "text": "bioinformatics,8(6):1-14,2007." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 628.0, + 1408.0, + 628.0, + 1408.0, + 667.0, + 292.0, + 667.0 + ], + "score": 0.98, + "text": "[18] R. K. Iyer and J. A. Bilmes. Submodular optimization with submodular cover and submodular knapsack" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 657.0, + 1039.0, + 657.0, + 1039.0, + 689.0, + 350.0, + 689.0 + ], + "score": 0.974, + "text": "constraints.In Advances in Neural Information Processing Systems." + }, + { + "category_id": 15, + "poly": [ + 290.0, + 703.0, + 1408.0, + 703.0, + 1408.0, + 743.0, + 290.0, + 743.0 + ], + "score": 0.969, + "text": "[19]M. Jiang,Z. Li, S. Zhang,S. Wang,X. Wang, Q. Yuan,and Z. Wei. Drug-target affinity prediction using" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 729.0, + 1207.0, + 729.0, + 1207.0, + 767.0, + 344.0, + 767.0 + ], + "score": 0.981, + "text": "graph neural network and contact maps.RsC Advances,10(35):20701-20712, 2020." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 781.0, + 1406.0, + 781.0, + 1406.0, + 814.0, + 294.0, + 814.0 + ], + "score": 0.977, + "text": "[20] H. Jin and X. Zhang. Latent adversarial training of graph convolution networks. In ICML Workshop on" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 808.0, + 1071.0, + 808.0, + 1071.0, + 840.0, + 348.0, + 840.0 + ], + "score": 0.988, + "text": "Learning and Reasoning with Graph-Structured Representations,2019." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 854.0, + 1410.0, + 854.0, + 1410.0, + 892.0, + 294.0, + 892.0 + ], + "score": 0.975, + "text": "[21] J. Kazius,R. McGuire,and R. Bursi. Derivation and validation of toxicophores for mutagenicity prediction." + }, + { + "category_id": 15, + "poly": [ + 348.0, + 882.0, + 899.0, + 882.0, + 899.0, + 914.0, + 348.0, + 914.0 + ], + "score": 0.989, + "text": "Journal of medicinal chemistry,48(1):312-320,2005." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 926.0, + 1408.0, + 926.0, + 1408.0, + 965.0, + 294.0, + 965.0 + ], + "score": 0.976, + "text": "[22] T.N. Kipf and M. Welling. Semi-supervised classfication with graph convolutional networks. In 5th" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 954.0, + 1354.0, + 954.0, + 1354.0, + 991.0, + 348.0, + 991.0 + ], + "score": 0.976, + "text": "International Conference on Learning Representations,ICLR 2Ol7, Conference Track Proceedings." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1005.0, + 1408.0, + 1005.0, + 1408.0, + 1037.0, + 296.0, + 1037.0 + ], + "score": 0.964, + "text": "[23]M. Liu, Y. Luo,L. Wang, Y. Xie, H. Yuan,S. Gui, H. Yu,Z. Xu, J. Zhang, Y. Liu, K. Yan, H. Liu, C. Fu," + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1031.0, + 1406.0, + 1031.0, + 1406.0, + 1063.0, + 350.0, + 1063.0 + ], + "score": 0.976, + "text": "B. Oztekin,X. Zhang,and S.Ji. DIG: A turnkey library for diving into graph deep learning research. arXiv" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1055.0, + 699.0, + 1055.0, + 699.0, + 1091.0, + 344.0, + 1091.0 + ], + "score": 0.999, + "text": "preprint arXiv:2103.12608,2021." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1106.0, + 1406.0, + 1106.0, + 1406.0, + 1138.0, + 296.0, + 1138.0 + ], + "score": 0.98, + "text": "[24] A.Lucic,M.ter Hoeve,G. Tolomei,M.de Rijke,and F. Silvestri. Cf-gnnexplainer: Counterfactual" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1132.0, + 1166.0, + 1132.0, + 1166.0, + 1164.0, + 350.0, + 1164.0 + ], + "score": 0.99, + "text": "explanations for graph neural networks.arXiv preprint arXiv:2102.03322,2021." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1178.0, + 1408.0, + 1178.0, + 1408.0, + 1216.0, + 292.0, + 1216.0 + ], + "score": 0.981, + "text": "[25] D.Luo, W. Cheng,D. Xu, W. Yu, B. Zong, H. Chen,and X. Zhang. Parameterized explainer for graph" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1202.0, + 1083.0, + 1202.0, + 1083.0, + 1242.0, + 346.0, + 1242.0 + ], + "score": 0.983, + "text": "neural network. In Advances in Neural Information Processing Systems." + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1246.0, + 1410.0, + 1246.0, + 1410.0, + 1293.0, + 290.0, + 1293.0 + ], + "score": 0.989, + "text": "[26] C. Molnar.Interpretable Machine Learning.2019.https://christophm.github.io/" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1283.0, + 659.0, + 1283.0, + 659.0, + 1315.0, + 350.0, + 1315.0 + ], + "score": 1.0, + "text": "interpretable-ml-book/." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1329.0, + 1408.0, + 1329.0, + 1408.0, + 1361.0, + 296.0, + 1361.0 + ], + "score": 0.988, + "text": "[27] G.Montufar,R.Pascanu,K.Cho,and Y.Bengio.On the numberof linear regions of deep neural networks." + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1353.0, + 745.0, + 1353.0, + 745.0, + 1385.0, + 348.0, + 1385.0 + ], + "score": 0.997, + "text": "arXiv preprint arXiv:1402.1869,2014." + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1397.0, + 1410.0, + 1397.0, + 1410.0, + 1444.0, + 290.0, + 1444.0 + ], + "score": 0.954, + "text": "[28]R.Moraffah,M. Karami,R.Guo,A. Raglin,and H.Liu.Causal interpretability for machine learning-" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1430.0, + 1320.0, + 1430.0, + 1320.0, + 1464.0, + 346.0, + 1464.0 + ], + "score": 0.992, + "text": "problems,methods and evaluation. ACM SIGKDD Explorations Newsleter,22(1):18-33,2020." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1478.0, + 1384.0, + 1478.0, + 1384.0, + 1510.0, + 296.0, + 1510.0 + ], + "score": 0.994, + "text": "[29] V.Nair and G.E.Hinton.Rectified linear units improve restricted boltzmann machines.In Icml,2010." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1522.0, + 1408.0, + 1522.0, + 1408.0, + 1563.0, + 292.0, + 1563.0 + ], + "score": 0.986, + "text": "[30] D. Numeroso and D. Bacciu. Meg: Generating molecular counterfactual explanations for deep graph" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1553.0, + 867.0, + 1553.0, + 867.0, + 1585.0, + 346.0, + 1585.0 + ], + "score": 0.988, + "text": "networks. arXiv preprint arXiv:2104.08060,2021." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1599.0, + 1408.0, + 1599.0, + 1408.0, + 1637.0, + 292.0, + 1637.0 + ], + "score": 0.971, + "text": "[31] D.E. Pires,T. L. Blundel, and D. B. Ascher. pkcsm: predicting small-molecule pharmacokinetic and" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1625.0, + 1410.0, + 1625.0, + 1410.0, + 1663.0, + 346.0, + 1663.0 + ], + "score": 0.992, + "text": "toxicity properties using graph-based signatures. Journal of medicinal chemistry,58(9):4066-4072,2015." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1675.0, + 1406.0, + 1675.0, + 1406.0, + 1708.0, + 296.0, + 1708.0 + ], + "score": 0.976, + "text": "[32]P.E.Pope,S. Kolouri, M.Rostami, C.E. Martin,and H. Hoffmann.Explainability methods for graph" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1704.0, + 1406.0, + 1704.0, + 1406.0, + 1736.0, + 350.0, + 1736.0 + ], + "score": 0.991, + "text": "convolutional neural networks.In Proceedings of the IEEE/CVF Conference on Computer Vision and" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1726.0, + 843.0, + 1726.0, + 843.0, + 1764.0, + 348.0, + 1764.0 + ], + "score": 0.991, + "text": "Pattern Recognition, pages 10772-10781,2019." + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1776.0, + 1406.0, + 1776.0, + 1406.0, + 1808.0, + 296.0, + 1808.0 + ], + "score": 0.988, + "text": "[33]R.R.Selvaraju,M.Cogswell,A.Das,R. Vedantam,D.Parikh,and D.Batra. Grad-cam: Visual explanations" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1804.0, + 1406.0, + 1804.0, + 1406.0, + 1836.0, + 350.0, + 1836.0 + ], + "score": 0.977, + "text": "from deep networks via gradient-based localization. In Proceedings of the IEEE international conference" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1828.0, + 785.0, + 1828.0, + 785.0, + 1867.0, + 348.0, + 1867.0 + ], + "score": 0.993, + "text": "on computer vision, pages 618-626, 2017." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1873.0, + 1408.0, + 1873.0, + 1408.0, + 1913.0, + 292.0, + 1913.0 + ], + "score": 0.984, + "text": "[34] P.L. Simon, M.Taylor,and I. Z. Kiss. Exact epidemic models on graphs using graph-automorphism driven" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1905.0, + 1015.0, + 1905.0, + 1015.0, + 1937.0, + 348.0, + 1937.0 + ], + "score": 0.991, + "text": "lumping. Journal of mathematical biology,62(4):479-508,2011." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1947.0, + 1406.0, + 1947.0, + 1406.0, + 1989.0, + 292.0, + 1989.0 + ], + "score": 0.987, + "text": "[35] K. Simonyan, A. Vedaldi,and A. Zisserman. Deep inside convolutional networks: Visualising image" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1979.0, + 895.0, + 1979.0, + 895.0, + 2012.0, + 350.0, + 2012.0 + ], + "score": 0.992, + "text": "classification models and saliency maps. ICLR,2014." + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 292, + 203, + 1409, + 203, + 1409, + 1684, + 292, + 1684 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 866, + 2061, + 866, + 2086, + 836, + 2086 + ], + "score": 0.582 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.536 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 869.0, + 2058.0, + 869.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "12" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "12" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 204.0, + 1406.0, + 204.0, + 1406.0, + 239.0, + 293.0, + 239.0 + ], + "score": 0.986, + "text": "[36]K.Sokol and P.A.Flach. Counterfactual explanations of machine learning predictions: opportunities and" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 234.0, + 849.0, + 234.0, + 849.0, + 265.0, + 351.0, + 265.0 + ], + "score": 0.963, + "text": "challenges for ai safety. In SafeAI@ AAAI, 2019." + }, + { + "category_id": 15, + "poly": [ + 297.0, + 281.0, + 1404.0, + 281.0, + 1404.0, + 313.0, + 297.0, + 313.0 + ], + "score": 0.963, + "text": "[37]P. Velickovic, G. Cucurul,A. Casanova,A. Romero,P.Lio,and Y. Bengio. Graph attention networks. In" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 306.0, + 1393.0, + 306.0, + 1393.0, + 342.0, + 349.0, + 342.0 + ], + "score": 0.977, + "text": "6th International Conference on Learning Representations, ICLR 2O18, Conference Track Proceedings." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 354.0, + 1406.0, + 354.0, + 1406.0, + 390.0, + 295.0, + 390.0 + ], + "score": 0.985, + "text": "[38]M. Vu and M. T. Thai.Pgm-explainer: Probabilistic graphical model explanations for graph neural" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 384.0, + 1021.0, + 384.0, + 1021.0, + 415.0, + 351.0, + 415.0 + ], + "score": 0.973, + "text": "networks.In Advances in Neural Information Processing Systems." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 430.0, + 1409.0, + 430.0, + 1409.0, + 466.0, + 293.0, + 466.0 + ], + "score": 0.984, + "text": "[39] N. Wale and G. Karypis. Comparison of descriptor spaces for chemical compound retrieval and clas-" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 458.0, + 1408.0, + 458.0, + 1408.0, + 489.0, + 349.0, + 489.0 + ], + "score": 0.975, + "text": "sification. In Sixth International Conference on Data Mining (ICDM'06),pages 678-689,2006.doi:" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 484.0, + 604.0, + 484.0, + 604.0, + 515.0, + 349.0, + 515.0 + ], + "score": 0.999, + "text": "10.1109/ICDM.2006.39." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 530.0, + 1408.0, + 530.0, + 1408.0, + 566.0, + 293.0, + 566.0 + ], + "score": 0.983, + "text": "[40] L.A.Wolsey. An analysis of the greedy algorithm for the submodular setcovering problem. Combinatorica," + }, + { + "category_id": 15, + "poly": [ + 348.0, + 557.0, + 564.0, + 557.0, + 564.0, + 589.0, + 348.0, + 589.0 + ], + "score": 0.977, + "text": "2(4):385-393,1982." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 603.0, + 1408.0, + 603.0, + 1408.0, + 644.0, + 292.0, + 644.0 + ], + "score": 0.987, + "text": "[41] J. Xiong,Z. Xiong, K. Chen,H. Jiang,and M. Zheng. Graph neural networks for automated de novo drug" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 631.0, + 735.0, + 631.0, + 735.0, + 667.0, + 346.0, + 667.0 + ], + "score": 0.991, + "text": "design. Drug Discovery Today,2021." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 680.0, + 1406.0, + 680.0, + 1406.0, + 716.0, + 293.0, + 716.0 + ], + "score": 0.959, + "text": "[42]H. Xu, Y. Ma, H.-C.Liu,D. Deb, H. Liu, J.-L. Tang,and A. K. Jain. Adversarial atacks and defenses in" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 708.0, + 1408.0, + 708.0, + 1408.0, + 744.0, + 348.0, + 744.0 + ], + "score": 0.991, + "text": " images, graphs and text: A review. International Journal of Automation and Computing,17(2):151-178," + }, + { + "category_id": 15, + "poly": [ + 349.0, + 733.0, + 417.0, + 733.0, + 417.0, + 767.0, + 349.0, + 767.0 + ], + "score": 0.999, + "text": "2020." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 780.0, + 1406.0, + 780.0, + 1406.0, + 817.0, + 293.0, + 817.0 + ], + "score": 0.966, + "text": "[43] K. Xu, H. Chen,S.Liu,P.-Y. Chen, T.-W. Weng,M. Hong,and X. Lin. Topology attack and defense for" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 809.0, + 1294.0, + 809.0, + 1294.0, + 845.0, + 346.0, + 845.0 + ], + "score": 0.982, + "text": " graph neural networks: An optimization perspective. arXiv preprint arXiv:1906.04214, 2019." + }, + { + "category_id": 15, + "poly": [ + 290.0, + 851.0, + 1409.0, + 851.0, + 1409.0, + 896.0, + 290.0, + 896.0 + ], + "score": 0.98, + "text": "[44] R. Yin,K.Li, G. Zhang,and J.Lu. A deeper graph neural network for recommender systems. Knowledge-" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 884.0, + 705.0, + 884.0, + 705.0, + 916.0, + 351.0, + 916.0 + ], + "score": 0.986, + "text": "Based Systems,185:105020,2019." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 930.0, + 1404.0, + 930.0, + 1404.0, + 967.0, + 293.0, + 967.0 + ], + "score": 0.98, + "text": "[45] Z. Ying,D. Bourgeois, J. You, M. Zitnik,and J.Leskovec. Gnnexplainer: Generating explanations for" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 957.0, + 1335.0, + 957.0, + 1335.0, + 995.0, + 346.0, + 995.0 + ], + "score": 0.984, + "text": " graph neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1008.0, + 1406.0, + 1008.0, + 1406.0, + 1039.0, + 295.0, + 1039.0 + ], + "score": 0.983, + "text": "[46] H.Yuan,J.Tang,X.Hu,and S.Ji. Xgnn: Towards model-level explanations of graph neural networks.In" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 1029.0, + 1411.0, + 1029.0, + 1411.0, + 1074.0, + 345.0, + 1074.0 + ], + "score": 0.98, + "text": "Proceedings ofthe 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining," + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1061.0, + 579.0, + 1061.0, + 579.0, + 1093.0, + 346.0, + 1093.0 + ], + "score": 0.995, + "text": "pages 430-438,2020." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1108.0, + 1406.0, + 1108.0, + 1406.0, + 1144.0, + 293.0, + 1144.0 + ], + "score": 0.967, + "text": "[47] H. Yuan, H. Yu, J. Wang, K. Li, and S.Ji._On explainability of graph neural networks via subgraph" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1135.0, + 898.0, + 1135.0, + 898.0, + 1169.0, + 348.0, + 1169.0 + ], + "score": 0.989, + "text": "explorations. arXiv preprint arXiv:2102.05152,2021." + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1179.0, + 1406.0, + 1179.0, + 1406.0, + 1219.0, + 292.0, + 1219.0 + ], + "score": 0.986, + "text": "[48] M.D. Zeiler and R.Fergus. Visualizing and understanding convolutional networks. In European conference" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1210.0, + 883.0, + 1210.0, + 883.0, + 1243.0, + 346.0, + 1243.0 + ], + "score": 0.978, + "text": "on computer vision, pages 818-833. Springer,2014." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1256.0, + 1406.0, + 1256.0, + 1406.0, + 1293.0, + 293.0, + 1293.0 + ], + "score": 0.966, + "text": "[49] J. Zhang,S.A. Bargal,Z.Lin,J.Brandt,X. Shen,and S.Sclarof. Top-down neural attention by excitation" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1284.0, + 1163.0, + 1284.0, + 1163.0, + 1321.0, + 349.0, + 1321.0 + ], + "score": 0.993, + "text": "backprop. International Journal of Computer Vision,126(10):1084-1102,2018." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1332.0, + 1408.0, + 1332.0, + 1408.0, + 1368.0, + 293.0, + 1368.0 + ], + "score": 0.981, + "text": "[50] M. Zhang and Y. Chen. Link prediction based on graph neural networks. In Advances in Neural Information" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1358.0, + 560.0, + 1358.0, + 560.0, + 1397.0, + 350.0, + 1397.0 + ], + "score": 0.999, + "text": "Processing Systems." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1406.0, + 1406.0, + 1406.0, + 1406.0, + 1443.0, + 293.0, + 1443.0 + ], + "score": 0.983, + "text": "[51] D. Zugner and S. Gunnemann. Adversarial attacks on graph neural networks via meta learning. arXiv" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1434.0, + 698.0, + 1434.0, + 698.0, + 1470.0, + 346.0, + 1470.0 + ], + "score": 0.98, + "text": "preprint arXiv:1902.08412, 2019." + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1484.0, + 1408.0, + 1484.0, + 1408.0, + 1515.0, + 295.0, + 1515.0 + ], + "score": 0.978, + "text": "[52] D.Zügner and S.Gunnemann.Certifable robustness and robust training for graph convolutional networks." + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1508.0, + 1408.0, + 1508.0, + 1408.0, + 1546.0, + 346.0, + 1546.0 + ], + "score": 0.99, + "text": "In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1535.0, + 664.0, + 1535.0, + 664.0, + 1571.0, + 348.0, + 1571.0 + ], + "score": 0.998, + "text": "Mining,pages 246-256,2019." + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1582.0, + 1406.0, + 1582.0, + 1406.0, + 1619.0, + 293.0, + 1619.0 + ], + "score": 0.962, + "text": "[53] D. Zugner, A. Akbarnejad,and S. Gunnemann. Adversarial atacks on neural networks for graph data. In" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1607.0, + 1408.0, + 1607.0, + 1408.0, + 1648.0, + 346.0, + 1648.0 + ], + "score": 0.987, + "text": "Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining," + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1635.0, + 605.0, + 1635.0, + 605.0, + 1671.0, + 344.0, + 1671.0 + ], + "score": 0.999, + "text": "pages 2847-2856,2018." + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/dV19Yyi1fS3/dV19Yyi1fS3.md b/parse/train/dV19Yyi1fS3/dV19Yyi1fS3.md new file mode 100644 index 0000000000000000000000000000000000000000..7c0312fe3571a20d674df1718d8952b08ee17eb9 --- /dev/null +++ b/parse/train/dV19Yyi1fS3/dV19Yyi1fS3.md @@ -0,0 +1,496 @@ +# TRAINING WITH QUANTIZATION NOISE FOREXTREME MODEL COMPRESSION + +Pierre Stock ∗ † Facebook AI Research, Inria + +Angela Fan∗ Facebook AI Research, LORIA + +Benjamin Graham Facebook AI Research + +Edouard Grave Facebook AI Research + +Rémi Gribonval Inria + +Hervé Jégou Facebook AI Research + +Armand Joulin Facebook AI Research + +# ABSTRACT + +We tackle the problem of producing compact models, maximizing their accuracy for a given model size. A standard solution is to train networks with Quantization Aware Training (Jacob et al., 2018), where the weights are quantized during training and the gradients approximated with the Straight-Through Estimator (Bengio et al., 2013). In this paper, we extend this approach to work beyond int8 fixedpoint quantization with extreme compression methods where the approximations introduced by STE are severe, such as Product Quantization. Our proposal is to only quantize a different random subset of weights during each forward, allowing for unbiased gradients to flow through the other weights. Controlling the amount of noise and its form allows for extreme compression rates while maintaining the performance of the original model. As a result we establish new state-of-the-art compromises between accuracy and model size both in natural language processing and image classification. For example, applying our method to state-of-the-art Transformer and ConvNet architectures, we can achieve $8 2 . 5 \%$ accuracy on MNLI by compressing RoBERTa to $1 4 \mathbf { M B }$ and $8 0 . 0 \%$ top-1 accuracy on ImageNet by compressing an EfficientNet-B3 to $3 . 3 \mathrm { M B }$ . 1 + +# 1 INTRODUCTION + +Many of the best performing neural network architectures in real-world applications have a large number of parameters. For example, the current standard machine translation architecture, Transformer (Vaswani et al., 2017), has layers that contain millions of parameters. Even models that are designed to jointly optimize the performance and the parameter efficiency, such as EfficientNets (Tan & Le, 2019), still require dozens to hundreds of megabytes, which limits their applications to domains like robotics or virtual assistants. + +Model compression schemes reduce the memory footprint of overparametrized models. Pruning (LeCun et al., 1990) and distillation (Hinton et al., 2015) remove parameters by reducing the number of network weights. In contrast, quantization focuses on reducing the bits per weight. This makes quantization particularly interesting when compressing models that have already been carefully optimized in terms of network architecture. Whereas deleting weights or whole hidden units will inevitably lead to a drop in performance, we demonstrate that quantizing the weights can be performed with little to no loss in accuracy. + +Popular postprocessing quantization methods, like scalar quantization, replace the floating-point weights of a trained network by a lower-precision representation, like fixed-width integers (Vanhoucke et al., 2011). These approaches achieve a good compression rate with the additional benefit of accelerating inference on supporting hardware. However, the errors made by these approximations accumulate in the computations operated during the forward pass, inducing a significant drop in performance (Stock et al., 2019). + +![](images/e8fa608fc575221a61a670b21fc8efe8983cacc3f544405b658da957ea4c631e.jpg) +Figure 1: Quant-Noise trains models to be resilient to inference-time quantization by mimicking the effect of the quantization method during training time. This allows for extreme compression rates without much loss in accuracy on a variety of tasks and benchmarks. + +A solution to address this drifting effect is to directly quantize the network during training. This raises two challenges. First, the discretization operators have a null gradient — the derivative with respect to the input is zero almost everywhere. This requires special workarounds to train a network with these operators. The second challenge that often comes with these workarounds is the discrepancy that appears between the train and test functions implemented by the network. Quantization Aware Training (QAT) (Jacob et al., 2018) resolves these issues by quantizing all the weights during the forward and using a straight through estimator (STE) (Bengio et al., 2013) to compute the gradient. This works when the error introduced by STE is small, like with int8 quantization, but does not suffice in compression regimes where the approximation made by the compression is more severe. + +In this work, we show that quantizing only a subset of weights instead of the entire network during training is more stable for high compression schemes. Indeed, by quantizing only a random fraction of the network at each forward, most the weights are updated with unbiased gradients. Interestingly, we show that our method can employ a simpler quantization scheme during the training. This is particularly useful for quantizers with trainable parameters, such as Product Quantizer (PQ), for which our quantization proxy is not parametrized. Our approach simply applies a quantization noise, called Quant-Noise, to a random subset of the weights, see Figure 1. We observe that this makes a network resilient to various types of discretization methods: it significantly improves the accuracy associated with (a) low precision representation of weights like int8; and (b) state-of-the-art PQ. Further, we demonstrate that Quant-Noise can be applied to existing trained networks as a post-processing step, to improve the performance network after quantization. + +In summary, this paper makes the following contributions: + +• We introduce the Quant-Noise technique to learn networks that are more resilient to a variety of quantization methods such as int4, int8, and PQ; • Adding Quant-Noise to PQ leads to new state-of-the-art trade-offs between accuracy and model size. For instance, for natural language processing (NLP), we reach $8 2 . 5 \%$ accuracy on MNLI by compressing RoBERTa to 14 MB. Similarly for computer vision, we report $8 0 . 0 \%$ top-1 accuracy on ImageNet by compressing an EfficientNet-B3 to $3 . 3 { \mathrm { M B } }$ ; • By combining PQ and int8 to quantize weights and activations for networks trained with Quant-Noise, we obtain extreme compression with fixed-precision computation and achieve $7 9 . 8 \%$ top-1 accuracy on ImageNet and 21.1 perplexity on WikiText-103. + +# 2 RELATED WORK + +Model compression. Many compression methods focus on efficient parameterization, via weight pruning (LeCun et al., 1990; Li et al., 2016; Huang et al., 2018; Mittal et al., 2018), weight sharing (Dehghani et al., 2018; Turc et al., 2019; Lan et al., 2019) or with dedicated architectures (Tan & Le, 2019; Zhang et al., 2017; Howard et al., 2019). Weight pruning is implemented during training (Louizos et al., 2017) or as a fine-tuning post-processing step (Han et al., 2015; 2016). Many pruning methods are unstructured, i.e., remove individual weights (LeCun et al., 1990; Molchanov et al., 2017). On the other hand, structured pruning methods follow the structure of the weights to + +reduce both the memory footprint and the inference time of a model (Li et al., 2016; Luo et al., 2017; +Fan et al., 2019). We refer the reader to Liu et al. (2018) for a review of different pruning strategies. + +Other authors have worked on lightweight architectures, by modifying existing models (Zhang et al., 2018; Wu et al., 2019; Sukhbaatar et al., 2019a) or developing new networks, such as MobileNet (Howard et al., 2019), ShuffleNet (Zhang et al., 2017), and EfficientNet (Tan & Le, 2019) in vision. + +Finally, knowledge distillation (Hinton et al., 2015) has been applied to sentence representation (Turc et al., 2019; Sanh et al., 2019a; Sun et al., 2019; Zhao et al., 2019; Jiao et al., 2019), to reduce the size of a BERT model (Devlin et al., 2018). + +Quantization. There are extensive studies of scalar quantization to train networks with lowprecision weights and activations (Courbariaux et al., 2015; Courbariaux & Bengio, 2016; Rastegari et al., 2016; McDonnell, 2018). These methods benefit from specialized hardware to also improve the runtime during inference (Vanhoucke et al., 2011). Other quantization methods such as Vector Quantization (VQ) and PQ (Jegou et al., 2011) quantize blocks of weights simultaneously to achieve higher compression rate (Stock et al., 2019; Gong et al., 2014; Joulin et al., 2016; Carreira-Perpiñán & Idelbayev, 2017). Closer to our work, several works have focused at simultaneously training and quantizing a network (Jacob et al., 2018; Krishnamoorthi, 2018; Gupta et al., 2015; Dong et al., 2019). Gupta et al. (2015) assigns weights to a quantized bin stochastically which is specific to scalar quantization, but allows training with fixed point arithmetic. Finally, our method can be interpreted as a form of Bayesian compression (Louizos et al., 2017), using the Bayesian interpretation of Dropout (Gal & Ghahramani, 2016). As opposed to their work, we select our noise to match the weight transformation of a target quantization methods without restricting it to a scale mixture prior. + +# 3 QUANTIZING NEURAL NETWORKS + +In this section, we present the principles of quantization, several standard quantization methods, and describe how to combine scalar and product quantization. For clarity, we focus on the case of a fixed real matrix $\mathbf { W } \in \mathbf { R } ^ { n \times p }$ . We suppose that this matrix is split into $m \times q$ blocks $\mathbf { b } _ { k l }$ : + +$$ +\mathbf { W } = \left( \sum _ { \vdots } ^ { \mathbf { b } _ { 1 1 } } \mathrm { ~ ~ \cdots ~ } \sum _ { \mathbf { b } _ { m q } } \right) , +$$ + +where the nature of these blocks is determined by the quantization method. A codebook is a set of $K$ vectors, i.e., ${ \mathcal { C } } = \{ \mathbf { c } [ 1 ] , \dots , \mathbf { c } [ K ] \}$ . Quantization methods compress the matrix W by assigning to each block $\mathbf { b } _ { k l }$ an index that points to a codeword $\mathbf { c }$ in a codebook $\mathcal { C }$ , and storing the codebook $\mathcal { C }$ and the resulting indices (as the entries ${ \bf { I } } _ { k l }$ of an index matrix $\mathbf { I }$ ) instead of the real weights. During the inference, they reconstruct an approximation $\widehat { \bf W }$ of the original matrix W such that $\widehat { \mathbf { b } } _ { k l } = \mathbf { c } [ \mathbf { I } _ { k l } ]$ . + +We distinguish scalar quantization, such as int8, where each block $\mathbf { b } _ { k l }$ consists of a single weight, from vector quantization, where several weights are quantized jointly. + +# 3.1 FIXED-POINT SCALAR QUANTIZATION + +Fixed-point scalar quantization methods replace floating-point number representations by lowprecision fixed-point representations. They simultaneously reduce a model’s memory footprint and accelerate inference by using fixed-point arithmetic on supporting hardware. + +Fixed-point scalar quantization operates on blocks that represent a single weight, i.e., $\mathbf { b } _ { k l } = \mathbf { W } _ { k l }$ Floating-point weights are replaced by $N$ bit fixed-point numbers (Gupta et al., 2015), with the extreme case of binarization where $N = 1$ (Courbariaux et al., 2015). More precisely, the weights are rounded to one of $2 ^ { N }$ possible codewords. These codewords correspond to bins evenly spaced by a scale factor $s$ and shifted by a bias $z$ . Each weight ${ \bf W } _ { k l }$ is mapped to its nearest codeword $c$ by successively quantizing with $z \mapsto$ round $( \mathbf { W } _ { k l } / s + z )$ and dequantizing with the inverse operation: + +$$ +\begin{array} { r } { \mathbf { c } = ( \mathrm { r o u n d } ( \mathbf { W } _ { k l } / s + z ) - z ) \times s , } \end{array} +$$ + +where we compute the scale and bias as: + +$$ +s = { \frac { \operatorname* { m a x } \mathbf { W } - \operatorname* { m i n } \mathbf { W } } { 2 ^ { N } - 1 } } \quad { \mathrm { a n d } } \quad z = \operatorname { r o u n d } ( \operatorname* { m i n } \mathbf { W } / s ) . +$$ + +We focus on this uniform rounding scheme instead of other non-uniform schemes (Choi et al., 2018; Li et al., 2019), because it allows for fixed-point arithmetic with implementations in PyTorch and Tensorflow (see Appendix). The compression rate is $\times 3 2 / N$ . The activations are also rounded to $N$ -bit fixed-point numbers. With int8 for instance, this leads to $\times 2$ to $\times 4$ faster inference on dedicated hardware. In this work, we consider both int4 and int8 quantization. + +# 3.2 PRODUCT QUANTIZATION + +Several quantization methods work on groups of weights, such as vectors, to benefit from the correlation induced by the structure of the network. In this work, we focus on Product Quantization for its good performance at extreme compression ratio (Stock et al., 2019). + +Traditional PQ. In vector quantization methods, the blocks are predefined groups of weights instead of single weights. The codewords are groups of values, and the index matrix I maps groups of weights from the matrix W to these codewords. In this section, we present the Product Quantization framework as it generalizes both scalar and vector quantization. We consider the case where we apply PQ to the columns of W and thus assume that $q = p$ . + +Traditional vector quantization techniques split the matrix $\mathbf { W }$ into its $p$ columns and learn a codebook on the resulting $p$ vectors. Instead, Product Quantization splits each column into $m$ subvectors and learns the same codebook for each of the resulting $m \times p$ subvectors. Each quantized vector is subsequently obtained by assigning its subvectors to the nearest codeword in the codebook. Learning the codebook is traditionally done using $k$ -means with a fixed number $K$ of centroids, typically $K = 2 5 6$ to store the index matrix I using int8. Thus, the objective function is written as: + +$$ +\Vert \mathbf { W } - \widehat { \mathbf { W } } \Vert _ { 2 } ^ { 2 } = \sum _ { k , l } \Vert \mathbf { b } _ { k l } - \mathbf { c } [ \mathbf { I } _ { k l } ] \Vert _ { 2 } ^ { 2 } . +$$ + +PQ shares representations between subvectors, which allows for higher compression rates than intN. + +Iterative PQ. When quantizing a full network rather than a single matrix, extreme compression with PQ induces a quantization drift as reconstruction error accumulates (Stock et al., 2019). Indeed, subsequent layers take as input the output of preceding layers, which are modified by the quantization of the preceding layers. This creates a drift in the network activations, resulting in large losses of performance. A solution proposed by Stock et al. (2019), which we call iterative PQ (iPQ), is to quantize layers sequentially from the lowest to the highest, and finetune the upper layers as the lower layers are quantized, under the supervision of the uncompressed (teacher) model. Codewords of each layer are finetuned by averaging the gradients of their assigned elements with gradient steps: + +$$ +\mathbf { c } \gets \mathbf { c } - \eta \frac { 1 } { | J _ { \mathbf { c } } | } \sum _ { ( k , l ) \in J _ { \mathbf { c } } } \frac { \partial \mathcal { L } } { \partial \mathbf { b } _ { k l } } , +$$ + +where $J _ { \mathbf { c } } = \{ ( k , l ) | \mathbf { c } [ \mathbf { I } _ { k l } ] = \mathbf { c } \}$ , $\mathcal { L }$ is the loss function and $\eta > 0$ is a learning rate. This adapts the upper layers to the drift appearing in their inputs, reducing the impact of the quantization approximation on the overall performance. + +# 3.3 COMBINING FIXED-POINT WITH PRODUCT QUANTIZATION + +Fixed-point quantization and Product Quantization are often regarded as competing choices, but can be advantageously combined. Indeed, PQ/iPQ compresses the network by replacing vectors of weights by their assigned centroids, but these centroids are in floating-point precision. Fixed-point quantization compresses both activations and weights to fixed-point representations. Combining both approaches means that the vectors of weights are mapped to centroids that are compressed to fixed-point representations, along with the activations. This benefits from the extreme compression ratio of iPQ and the finite-precision arithmetics of intN quantization. + +More precisely, for a given matrix, we store the $\dot { \mathtt { 1 } } \mathtt { n t } \mathtt { 8 }$ representation of the $K$ centroids of dimension $d$ along with the $\log _ { 2 } K$ representations of the centroid assignments of the $m \times p$ subvectors. The int8 representation of the centroids is obtained with Eq. (2). The overall storage of the matrix and activations during a forward pass with batch size 1 (recalling that the input dimension is n) writes + +$$ +M = 8 \times K d + \log _ { 2 } { K } \times m p + 8 \times n \mathrm { b i t s } . +$$ + +In particular, when $K = 2 5 6$ , the centroid assignments are also stored in int8, which means that every value required for a forward pass is stored in an int8 format. We divide by 4 the float32 overhead of storing the centroids, although the storage requirement associated with the centroids is small compared to the cost of indexing the subvectors for standard networks. In contrast to iPQ alone where we only quantize the weights, we also quantize the activations using int8. We evaluate this approach on both natural language processing and computer vision tasks in Section 5. + +# 4 METHOD + +Deep networks are not exposed to the noise caused by the quantization drift during training, leading to suboptimal performance. A solution to make the network robust to quantization is to introduce it during training. Quantization Aware Training (QAT) (Jacob et al., 2018) exposes the network during training by quantizing weights during the forward pass. This transformation is not differentiable and gradients are approximated with a straight through estimator (STE) (Bengio et al., 2013; Courbariaux & Bengio, 2016). STE introduces a bias in the gradients that depends on level of quantization of the weights, and thus, the compression ratio. In this section, we propose a simple modification to control this induced bias with a stochastic amelioration of QAT, called Quant-Noise. The idea is to quantize a randomly selected fraction of the weights instead of the full network as in QAT, leaving some unbiased gradients flow through unquantized weights. Our general formulation can simulate the effect of both quantization and of pruning during training. + +# 4.1 TRAINING NETWORKS WITH QUANTIZATION NOISE + +We consider the case of a real matrix W as in Section 3. During the training of a network, our proposed Quant-Noise method works as follows: first, we compute blocks $\mathbf { b } _ { k l }$ related to a target quantization method. Then, during each forward pass, we randomly select a subset of these blocks and apply some distortion to them. + +More formally, given a set of tuples of indices $J \subset \{ ( k , l ) \}$ for $1 \leq k \leq m$ , $1 \leq l \leq q$ and a distortion or noise function $\varphi$ acting on a block, we define an operator $\psi ( \cdot \mid J )$ such that, for each block $\mathbf { b } _ { k l }$ , we apply the following transformation: + +$$ +\psi ( \mathbf { b } _ { k l } \mid J ) = { \left\{ \begin{array} { l l } { \varphi ( \mathbf { b } _ { k l } ) } & { { \mathrm { i f ~ } } ( k , l ) \in J , } \\ { \mathbf { b } _ { k l } } & { { \mathrm { o t h e r w i s e } } . } \end{array} \right. } +$$ + +The noise function $\varphi$ simulates the change in the weights produced by the target quantization method (see Section 4.2 for details). We replace the matrix W by the resulting noisy matrix $\mathbf { W } _ { \mathrm { n o i s e } }$ during the forward pass to compute a noisy output $\mathbf { y } _ { \mathrm { n o i s e } }$ , i.e., + +$$ +\mathbf { W } _ { \mathrm { n o i s e } } = ( \psi ( \mathbf { b } _ { k l } \mid J ) ) _ { k l } \mathrm { a n d } \mathbf { y } _ { \mathrm { n o i s e } } = \mathbf { x } \mathbf { W } _ { \mathrm { n o i s e } } +$$ + +where $\mathbf { x }$ is an input vector. During the backward pass, we apply STE, which amounts to replacing the distorted weights $\mathbf { W _ { \mathrm { n o i s e } } }$ by their non-distorted counterparts. Note that our approach is equivalent to QAT when $J$ containts all the tuples of indices. However, an advantage of Quant-Noise over QAT is that unbiased gradients continue to flow via blocks unaffected by the noise. As these blocks are randomly selected for each forward, we guarantee that each weight regularly sees gradients that are not affected by the nature of the function $\varphi$ . As a side effect, our quantization noise regularizes the network in a similar way as DropConnect (Wan et al., 2013) or LayerDrop (Fan et al., 2019). + +Composing quantization noises. As noise operators are compositionally commutative, we can make a network robust to a combination of quantization methods by composing their noise operators: + +$$ +\psi ( \mathbf { b } _ { k l } \mid J ) = \psi _ { 1 } \circ \psi _ { 2 } ( \mathbf { b } _ { k l } \mid J ) . +$$ + +This property is particularly useful to combine quantization with pruning operators during training, as well as combining scalar quantization with product quantization. + +# 4.2 ADDING NOISE TO SPECIFIC QUANTIZATION METHODS + +In this section, we propose several implementations of the noise function $\varphi$ for the quantization methods described in Section 3. We also show how to handle pruning with it. + +Table 1: Comparison of different quantization schemes with and without Quant-Noise on language modeling and image classification. For language modeling, we train a Transformer on the Wikitext-103 benchmark and report perplexity (PPL) on test. For image classification, we train a EfficientNet-B3 on the ImageNet-1k benchmark and report top-1 accuracy on validation and use our re-implementation of EfficientNet-B3. The original implementation of Tan & Le (2019) achieves an uncompressed Top-1 accuracy of $8 1 . 9 \%$ . For both settings, we report model size in megabyte (MB) and the compression ratio compared to the original model. + +
Quantization SchemeLanguage Modeling 16-layer Transformer Wikitext-103Image Classification EfficientNet-B3 ImageNet-1k
SizeCompressionPPLSizeCompressionTop-1
Uncompressed model942×118.346.7×181.5
int 4 quantization118×839.45.8×845.3
- trained with QAT118×834.15.8×859.4
- trained with Quant-Noise118×821.85.8×867.8
int 8 quantization236×419.611.7×480.7
- trained with QAT236×421.011.7×480.8
- trained with Quant-Noise236×418.711.7×480.9
iPQ38×2525.23.3×1479.0
- trained with QAT38×2541.23.3×1455.7
- trained with Quant-Noise38×2520.73.3×1480.0
iPQ&int8 +Quant-Noise38×2521.13.1×1579.8
+ +Fixed-point scalar quantization. In intN quantization, the blocks are atomic and weights are rounded to their nearest neighbor in the codebook. The function $\varphi$ replaces weight ${ \bf W } _ { k l }$ with the output of the rounding function defined in Eq. (2), i.e., + +$$ +\varphi _ { \mathrm { i n t N } } ( w ) = ( \mathrm { r o u n d } ( w / s + z ) - z ) \times s , +$$ + +where $s$ and $z$ are updated during training. In particular, the application of Quant-Noise to int8 scalar quantization is a stochastic amelioration of QAT. + +Product quantization. As opposed to intN, codebooks in PQ require a clustering step based on weight values. During training, we learn codewords online and use the resulting centroids to implement the quantization noise. More precisely, the noise function $\varphi _ { \mathrm { P Q } }$ assigns a selected block b to its nearest codeword in the associated codebook $\mathcal { C }$ : + +$$ +\begin{array} { r } { \varphi _ { \mathrm { P Q } } ( \mathbf { v } ) = \operatorname * { a r g m i n } _ { \mathbf { c } \in \mathcal { C } } \| \mathbf { b } - \mathbf { c } \| _ { 2 } ^ { 2 } . } \end{array} +$$ + +Updating the codebooks online works well. However, empirically, running $k$ -means once per epoch is faster and does not noticeably modify the resulting accuracy. + +Note that computing the exact noise function for PQ is computationally demanding. We propose a simpler and faster alternative approximation $\varphi _ { \mathrm { p r o x y } }$ to the operational transformation of PQ and iPQ. The noise function simply zeroes out the subvectors of the selected blocks, i.e., $\varphi _ { \mathrm { p r o x y } } ( \mathbf { v } ) = 0$ . As a sidenote, we considered other alternatives, for instance one where the subvectors are mapped to the mean subvector. In practice, we found that these approximations lead to similar performance, see Section 7.2. This proxy noise function is a form of Structured Dropout and encourages correlations between the subvectors. This correlation is beneficial to the subsequent clustering involved in PQ/iPQ. + +Adding pruning to the quantization noise. The specific form of quantization noise can be adjusted to incorporate additional noise specific to pruning. We simply combine the noise operators of quantization and pruning by composing them following Eq. (8). We consider the pruning noise function of Fan et al. (2019) where they randomly drop predefined structures during training. In particular, we focus on LayerDrop, where the structures are the residual blocks of highway-like layers (Srivastava et al., 2015), as most modern architectures, such as ResNet or Transformer, are composed of this structure. More precisely, the corresponding noise operator over residual blocks $\mathbf { v }$ is $\dot { \varphi } _ { \mathrm { L a y e r D r o p } } ( \mathbf { v } ) = 0 .$ . For pruning, we do not use STE to backpropagate the gradient of pruned weights, as dropping them entirely during training has the benefit of speeding convergence (Huang et al., 2016). Once a model is trained with LayerDrop, the number of layers kept at inference can be adapted to match computation budget or time constraint. + +![](images/fb191eba2d0809d5fb2682e2e7926abb2bbee75ac52ef027ff222f8961c973a6.jpg) +Figure 2: Performance as a function of model size. We compare models quantized with PQ and trained with the related Quant-Noise to the state of the art. (a) Test perplexity on Wikitext-103 (b) Dev Accuracy on MNLI (c) ImageNet Top-1 accuracy. Model size is shown in megabytes on a log scale. Red and gray coloring indicates existing work, with different colors for visual distinction. + +Table 2: Decomposing the impact of the different compression schemes. (a) we train Transformers with Adaptive Input and LayerDrop on Wikitext-103 (b) we pre-train RoBERTA base models with LayerDrop and then finetune on MNLI (c) we train an EfficientNet-B3 on ImageNet. We report the compression ratio w.r.t. to the original model (“comp.”) and the resulting size in MB. + +
Language modelingSentence RepresentationImage Classification
Comp.SizePPLComp.SizeAcc.Comp.SizeAcc.
Unquantized models
Original model×194218.3×148084.8×146.781.5
+ Sharing×1.851018.7×1.925084.0×1.434.280.1
+ Pruning×3.725522.5×3.812581.3×1.629.578.5
Quantized models
iPQ×24.83825.2×12.63882.5× 14.13.379.0
+ Quant-Noise×24.83820.7×12.63883.6×14.13.380.0
+ Sharing× 49.51922.0×34.31482.5×182.678.9
+ Pruning×94.21024.7×58.5878.8×202.377.8
+ +# 5 RESULTS + +We demonstrate the impact of Quant-Noise on the performance of several quantization schemes in a variety of settings (see Appendix - Sec. 7.5). + +# 5.1 IMPROVING COMPRESSION WITH QUANT-NOISE + +Quant-Noise is a regularization method that makes networks more robust to the target quantization scheme or combination of quantization schemes during training. We show the impact of Quant-Noise in Table 1 for a variety of quantization methods: int8/int4 and iPQ. + +We experiment in 2 different settings: a Transformer network trained for language modeling on WikiText-103 and a EfficientNet-B3 convolutional network trained for image classification on ImageNet-1k. Our quantization noise framework is general and flexible — Quant-Noise improves the performance of quantized models for every quantization scheme in both experimental settings. Importantly, Quant-Noise only changes model training by adding a regularization noise similar to dropout, with no impact on convergence and very limited impact on training speed $( < 5 \%$ slower). + +This comparison of different quantization schemes shows that Quant-Noise works particularly well with high performance quantization methods, like iPQ, where QAT tends to degrade the performances, even compared to quantizing as a post-processing step. In subsequent experiments in this section, we focus on applications with iPQ because it offers the best trade-off between model performance and compression, and has little negative impact on FLOPS. + +
Language ModelingPPLRoBERTaAcc.
Train without Quant-Noise25.2Train without Quant-Noise82.5
+ Finetune with Quant-Noise20.9+ Finetune with Quant-Noise83.4
Train with Quant-Noise20.7Train with Quant-Noise83.6
+ +Table 3: Quant-Noise: Finetuning vs training. We report performance after iPQ quantization. We train with the $\phi _ { \mathrm { p r o x y } }$ noise and finetune with Quant-Noise, and use it during the transfer to MNLI for each RoBERTa model. + +Fixed-Point Product Quantization. Combining iPQ and int8 as described in Section 3.3 allows us to take advantage of the high compression rate of iPQ with a fixed-point representation of both centroids and activations. As shown in Table 1, this combination incurs little loss in accuracy with respect to $\mathrm { i } \mathrm { P Q } +$ Quant-Noise. Most of the memory footprint of iPQ comes from indexing and not storing centroids, so the compression ratios are comparable. + +Complementarity with Weight Pruning and Sharing. We analyze how Quant-Noise is compatible and complementary with pruning (“+Prune”) and weight sharing ( $^ { 6 6 } { + } \mathrm { S l }$ hare”), see Appendix for details on weight sharing. We report results for Language modeling on WikiText-103, pre-trained sentence representations on MNLI and object classification on ImageNet-1k in Table 2. The conclusions are remarkably consistent across tasks and benchmarks: Quant-Noise gives a large improvement over strong iPQ baselines. Combining it with sharing and pruning offers additional interesting operating points of performance vs size. + +# 5.2 COMPARISON WITH THE STATE OF THE ART + +We now compare our approach on the same tasks against the state of the art. We compare $\mathrm { i } \mathrm { P Q } +$ Quant-Noise with 6 methods of network compression for Language modeling, 8 state-of-the-art methods for Text classification, and 8 recent methods evaluate image classification on Imagenet with compressed models. These comparisons demonstrate that Quant-Noise leads to extreme compression rates at a reasonable cost in accuracy. We apply our best quantization setup on competitive models and reduce their memory footprint by $\times 2 0 - 9 4 $ when combining with weight sharing and pruning, offering extreme compression for good performance. + +Natural Language Processing. In Figure 2, we examine the trade-off between performance and model size. Our quantized RoBERTa offers a competitive trade-off between size and performance with memory reduction methods dedicated to BERT, like TinyBERT, MobileBERT, or AdaBERT. + +Image Classification. We compress EfficientNet-B3 from 46.7Mb to $3 . 3 { \mathrm { M b } }$ ( $\times 1 4$ compression) while maintaining high top-1 accuracy $( 7 8 . 5 \%$ versus $8 0 \%$ for the original model). As shown in Figure 2, our quantized EfficientNet-B3 is smaller and more accurate than architectures dedicated to optimize on-device performance with limited size like MobileNet or ShuffleNet. We further evaluate the beneficial effect of Quant-Noise on ResNet-50 to compare directly with Stock et al. (2019). Results shown in Table 4 indicate improvement with Quant-Noise compared to previous work. + +Incorporating pruning noise into quantization is also beneficial. For example, with pruning $\mathrm { i P Q + }$ Quant-Noise reduces size by $\times 2 5$ with only a drop of $2 . 4 \ : \mathrm { P P L }$ in language modeling. Further, pruning reduces FLOPS by the same ratio as its compression factor, in our case, $\times 2$ . By adding sharing with pruning, in language modeling, we achieve an extreme compression ratio of $\times 9 4$ with a drop of $6 . 4 \ : \mathrm { P P L }$ with FLOPS reduction from pruning entire shared chunks of layers. For comparison, our $1 0 \mathbf { M B }$ model has the same performance as the $5 7 0 \mathrm { M B }$ Transformer-XL base. + +# 5.3 FINETUNING WITH QUANT-NOISE FOR POST-PROCESSING QUANTIZATION + +We explore taking existing models and post-processing with Quant-Noise instead of training from scratch. For language modeling, we train for 10 additional epochs. For RoBERTa, we train for $2 5 \mathrm { k }$ additional updates. Finetuning with Quant-Noise incorporates the benefits and almost matches training from scratch (Table 3). In language modeling, there is only a $0 . 2 \ : \mathrm { P P L }$ difference. We further examine how to incorporate Quant-Noise more flexibly into pretraining RoBERTa. We take an already trained RoBERTa model and incorporate Quant-Noise during sentence classification finetuning. This is effective at compressing while retaining accuracy after quantization. + +# 6 CONCLUSION + +We show that quantizing a random subset of weights during training maintains performance in the high quantization regime. We validate that Quant-Noise works with a variety of different quantization schemes on several applications in text and vision. Our method can be applied to a combination of iPQ and int8 to benefit from extreme compression ratio and fixed-point arithmetic. Finally, we show that Quant-Noise can be used as a post-processing step to prepare already trained networks for subsequent quantization, to improve the performance of the compressed model. + +# REFERENCES + +A. Adcock, V. Reis, M. Singh, Z. Yan, L. van der Maaten, K. Zhang, S. Motwani, J. Guerin, N. Goyal, I. Misra, L. Gustafson, C. Changhan, and P. Goyal. Classy vision. 2019. +Alexei Baevski and Michael Auli. Adaptive input representations for neural language modeling. arXiv preprint arXiv:1809.10853, 2018. +Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. +James Bradbury, Stephen Merity, Caiming Xiong, and Richard Socher. Quasi-recurrent neural networks. arXiv preprint arXiv:1611.01576, 2016. +Qingqing Cao, Harsh Trivedi, Aruna Balasubramanian, and Niranjan Balasubramanian. Faster and just as accurate: A simple decomposition for transformer models. +Miguel A. Carreira-Perpiñán and Yerlan Idelbayev. Model compression as constrained optimization, with application to neural nets. part ii: quantization, 2017. +Daoyuan Chen, Yaliang Li, Minghui Qiu, Zhen Wang, Bofang Li, Bolin Ding, Hongbo Deng, Jun Huang, Wei Lin, and Jingren Zhou. Adabert: Task-adaptive bert compression with differentiable neural architecture search. arXiv preprint arXiv:2001.04246, 2020. +Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. +Matthieu Courbariaux and Yoshua Bengio. Binarynet: Training deep neural networks with weights and activations constrained to $+ 1$ or -1. CoRR, 2016. +Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. CoRR, 2015. +Zihang Dai, Zhilin Yang, Yiming Yang, William W Cohen, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. +Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In Proc. of ICML, 2017. +Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Łukasz Kaiser. Universal transformers, 2018. +J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. + +Yinpeng Dong, Renkun Ni, Jianguo Li, Yurong Chen, Hang Su, and Jun Zhu. Stochastic quantization for learning accurate low-bit deep neural networks. International Journal of Computer Vision, 127 (11-12):1629–1642, 2019. + +Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019. + +Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pp. 1050–1059, 2016. + +Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bourdev. Compressing deep convolutional networks using vector quantization. arXiv preprint arXiv:1412.6115, 2014. + +Edouard Grave, Armand Joulin, Moustapha Cisse, David Grangier, and Herve Jegou. Efficient softmax approximation for gpus. arXiv, abs/1609.04309, 2016. + +Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning with limited numerical precision. In ICML, 2015. + +Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In NIPS, pp. 1135–1143, 2015. + +Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding. ICLR, 2016. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, 2015. + +Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. + +Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. arXiv e-prints, 2019. + +Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016. + +Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In CVPR, 2018. + +Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2704–2713, 2018. + +Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. PAMI, 2011. + +Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351, 2019. + +Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas Mikolov. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651, 2016. + +Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. + +Guillaume Lample and Alexis Conneau. Cross-lingual language model pretraining. arXiv preprint arXiv:1901.07291, 2019. + +Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations, 2019. + +Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In NIPS, 1990. + +Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. + +Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: A non-uniform discretization for neural networks. arXiv preprint arXiv:1909.13144, 2019. + +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. + +Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. + +Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. + +Christos Louizos, Max Welling, and Diederik P Kingma. Learning sparse neural networks through l_0 regularization. arXiv preprint arXiv:1712.01312, 2017. + +Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. Thinet: A filter level pruning method for deep neural network compression. In ICCV, 2017. + +Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet V2: practical guidelines for efficient CNN architecture design. CoRR, 2018. + +Xindian Ma, Peng Zhang, Shuai Zhang, Nan Duan, Yuexian Hou, Dawei Song, and Ming Zhou. A tensorized transformer for language modeling. arXiv preprint arXiv:1906.09777, 2019. + +Mark D. McDonnell. Training wide residual networks for deployment using a single bit for each weight, 2018. + +Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer Sentinel Mixture Models. arXiv, abs/1609.07843, 2016. + +Deepak Mittal, Shweta Bhardwaj, Mitesh M Khapra, and Balaraman Ravindran. Recovering from random pruning: On the plasticity of deep convolutional neural networks. In WACV, 2018. + +Dmitry Molchanov, Arsenii Ashukha, and Dmitry Vetrov. Variational dropout sparsifies deep neural networks. In ICML, 2017. + +Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of NAACL-HLT 2019: Demonstrations, 2019. + +Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. How to construct deep recurrent neural networks. In Proceedings of the Second International Conference on Learning Representations (ICLR 2014), 2014. + +Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. + +Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. + +Jack W Rae, Anna Potapenko, Siddhant M Jayakumar, and Timothy P Lillicrap. Compressive transformers for long-range sequence modelling. arXiv preprint arXiv:1911.05507, 2019. + +Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV, 2016. + +Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Conference on Computer Vision and Pattern Recognition, pp. 4510–4520, 2018. + +Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. 2019a. + +Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019b. + +Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Highway networks. arXiv preprint arXiv:1505.00387, 2015. + +Pierre Stock, Armand Joulin, Rémi Gribonval, Benjamin Graham, and Hervé Jégou. And the bit goes down: Revisiting the quantization of neural networks. CoRR, abs/1907.05686, 2019. + +Sainbayar Sukhbaatar, Edouard Grave, Piotr Bojanowski, and Armand Joulin. Adaptive attention span in transformers. arXiv preprint arXiv:1905.07799, 2019a. + +Sainbayar Sukhbaatar, Edouard Grave, Guillaume Lample, Herve Jegou, and Armand Joulin. Augmenting self-attention with persistent memory. arXiv preprint arXiv:1907.01470, 2019b. + +Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. Patient knowledge distillation for bert model compression. EMNLP, 2019. + +Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. Mobilebert: Task-agnostic compression of bert for resource limited devices. + +Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization and momentum in deep learning. In International conference on machine learning, pp. 1139–1147, 2013. + +Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks, 2019. + +Iulia Turc, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Well-read students learn better: The impact of student initialization on knowledge distillation. arXiv preprint arXiv:1908.08962, 2019. + +Vincent Vanhoucke, Andrew Senior, and Mark Z Mao. Improving the speed of neural networks on cpus. 2011. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. + +Li Wan, Matthew Zeiler, Sixin Zhang, Yann Le Cun, and Rob Fergus. Regularization of neural networks using DropConnect. In ICML, 2013. + +Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. 2019. ICLR. + +Kuan Wang, Zhijian Liu, Yujun Lin andx Ji Lin, and Song Han. HAQ: hardware-aware automated quantization. CoRR, 2018. + +Adina Williams, Nikita Nangia, and Samuel R. Bowman. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of NAACL-HLT, 2018. + +Felix Wu, Angela Fan, Alexei Baevski, Yann Dauphin, and Michael Auli. Pay less attention with lightweight and dynamic convolutions. In ICLR, 2019. + +Biao Zhang, Deyi Xiong, and Jinsong Su. Accelerating neural transformer via an average attention network. arXiv preprint arXiv:1805.00631, 2018. + +Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. CoRR, 2017. + +Sanqiang Zhao, Raghav Gupta, Yang Song, and Denny Zhou. Extreme language model compression with optimal subwords and shared projections. arXiv preprint arXiv:1909.11687, 2019. + +
SettingModelCompressionTop-1 Accuracy
Small BlocksStock et al. (2019)19x73.8
Quant-Noise19x74.3
Large BlocksStock et al. (2019)32x68.2
Quant-Noise32x68.8
+ +![](images/ba35cf41421a36478cb5c76c461ca17166f7ed7e069174cbe097a3f7e7e6cbf8.jpg) +Table 4: Compression of ResNet-50 with Quant-Noise. We compare to Stock et al. (2019) in both the small and large blocks regime. For fair comparison, we hold the compression rate constant. Quant-Noise provides improved performance in both settings. +Figure 3: Effect of Quantization Parameters. We report the influence of the proportion of blocks to which we apply the noise. We focus on Transformer for Wikitext-103 language modeling. We explore two settings: iPQ and int8. For iPQ, we use $\varphi _ { \mathrm { p r o x y } }$ . + +# 7 APPENDIX + +# 7.1 QUANTIZATION OF ADDITIONAL ARCHITECTURES + +ResNet-50. We explore the compression of ResNet-50, a standard architecture used Computer Vision. In Table 4, we compare Quant-Noise to iPQ Compression from Stock et al. (2019) and show that Quant-Noise provide consistent additional improvement. + +# 7.2 ABLATIONS + +In this section, we examine the impact of the level of noise during training as well as the impact of approximating iPQ during training. + +# 7.3 IMPACT OF NOISE RATE + +We analyze the performance for various values of Quant-Noise in Figure 3 on a Transformer for language modeling. For iPQ, performance is impacted by high rates of quantization noise. For example, a Transformer with the noise function $\varphi _ { \mathrm { p r o x y } }$ degrades with rate higher than 0.5, i.e., when half of the weights are passed through the noise function $\varphi _ { \mathrm { p r o x y } }$ . We hypothesize that for large quantities of noise, a larger effect of using proxy rather than the exact PQ noise is observed. For int8 quantization and its noise function, higher rates of noise are slightly worse but not as severe. A rate of 1 for int8 quantization is equivalent to the Quantization Aware Training of (Krishnamoorthi, 2018), as the full matrix is quantized with STE, showing the potential benefit of partial quantization during training. + +# 7.4 IMPACT OF APPROXIMATING THE NOISE FUNCTION + +We study the impact of approximating quantization noise during training. We focus on the case of iPQ with the approximation described in Section 4.2. In Table 5, we compare the correct noise function for iPQ with its approximation $\varphi _ { \mathrm { p r o x y } }$ . This approximate noise function does not consider cluster assignments or centroid values and simply zeroes out the selected blocks. For completeness, we include an intermediate approximation where we consider cluster assignments to apply noise within each cluster, but still zero-out the vectors. These approximations do not affect the performance of the quantized models. This suggests that increasing the correlation between subvectors that are jointly clustered is enough to maintain the performance of a model quantized with iPQ. Since PQ tends to work well on highly correlated vectors, such as activations in convolutional networks, this is not surprising. Using the approximation $\varphi _ { \mathrm { p r o x y } }$ presents the advantage of speed and practicality. Indeed, one does not need to compute cluster assignments and centroids for every layer in the network after each epoch. Moreover, the approach $\varphi _ { \mathrm { p r o x y } }$ is less involved in terms of code. + +Table 5: Exact versus proxy noise function for different block selections with iPQ. We compare exact $\phi _ { \mathrm { P Q } }$ and the approximation $\phi _ { \mathrm { p r o x y } }$ with blocks selected from all subvectors or subvectors from the same cluster. + +
NoiseBlocksPPLQuant PPL
PQSubvectors18.321.1
PQClusters18.321.2
proxySubvectors18.321.0
proxyClusters18.421.1
+ +# 7.5 EXPERIMENTAL SETTING + +We assess the effectiveness of Quant-Noise on competitive language and vision benchmarks. We consider Transformers for language modeling, RoBERTa for pre-training sentence representations, and EfficientNet for image classification. Our models are implemented in PyTorch (Paszke et al., 2017). We use fairseq (Ott et al., 2019) for language modeling and pre-training for sentence representation tasks and Classy Vision (Adcock et al., 2019) for EfficientNet. + +Language Modeling. We experiment on the Wikitext-103 benchmark (Merity et al., 2016) that contains 100M tokens and a vocabulary of 260k words. We train a 16 layer Transformer following Baevski & Auli (2018) with a LayerDrop rate of 0.2 (Fan et al., 2019). We report perplexity (PPL) on the test set. + +Pre-Training of Sentence Representations. We pre-train the base BERT model (Devlin et al., 2018) on the BooksCorpus $^ +$ Wiki dataset with a LayerDrop rate of 0.2. We finetune the pre-trained models on the MNLI task (Williams et al., 2018) from the GLUE Benchmark (Wang et al., 2019) and report accuracy. We follow the parameters in Liu et al. (2019) training and finetuning. + +Image Classification. We train an EfficientNet-B3 model (Tan & Le, 2019) on the ImageNet object classification benchmark (Deng et al., 2009). The EfficientNet-B3 of Classy Vision achieves a Top-1 accuracy of $8 1 . 5 \%$ , which is slightly below than the performance of $8 1 . 9 \%$ reported by Tan & Le (2019). + +# 7.6 TRAINING DETAILS + +Language Modeling To handle the large vocabulary of Wikitext-103, we follow (Dauphin et al., 2017) and (Baevski & Auli, 2018) in using adaptive softmax (Grave et al., 2016) and adaptive input for computational efficiency. For both input and output embeddings, we use dimension size 1024 and three adaptive bands: 20K, 40K, and 200K. We use a cosine learning rate schedule (Baevski & Auli, 2018; Loshchilov & Hutter, 2016) and train with Nesterov’s accelerated gradient (Sutskever et al., 2013). We set the momentum to 0.99 and renormalize gradients if the norm exceeds 0.1 (Pascanu et al., 2014). During training, we partition the data into blocks of contiguous tokens that ignore document boundaries. At test time, we respect sentence boundaries. We set LayerDrop to 0.2. We set Quant-Noise value to 0.05. During training time, we searched over the parameters (0.05, 0.1, 0.2) to determine the optimal value of Quant-Noise. During training time, the block size of Quant-Noise is 8. + +RoBERTa The base architecture is a 12 layer model with embedding size 768 and FFN size 3072. We follow (Liu et al., 2019) in using the subword tokenization scheme from (Radford et al., 2019), which uses bytes as subword units. This eliminates unknown tokens. We train with large batches of size 8192 and maintain this batch size using gradient accumulation. We do not use next sentence prediction (Lample & Conneau, 2019). We optimize with Adam with a polynomial decay learning rate schedule. We set LayerDrop to 0.2. We set Quant-Noise value to 0.1. We did not hyperparameter + +
ModelMBPPL
Trans XL Large (Dai et al., 2019) Compressive Trans (Rae et al.,2019) GCNN (Dauphin et al., 2017) 4 Layer QRNN (Bradbury et al., 2016)970 970 87018.3 17.1 37.2 33.0
Trans XL Base (Dai et al.,2019) Persis Mem (Sukhbaatar etal.,2019b) Tensorized core-2 (Ma et al.,2019)570 506 32524.0 20.6
18.9
Quant-Noise3820.7
Quant-Noise + Share + Prune
10
24.2
+ +Table 6: Performance on Wikitext-103. We report test set perplexity and model size in megabytes. +Lower perplexity is better. + +search to determine the optimal value of Quant-Noise as training RoBERTa is computationally intensive. During training time, the block size of Quant-Noise is 8. + +During finetuning, we hyperparameter search over three learning rate options (1e-5, 2e-5, 3e-5) and batchsize (16 or 32 sentences). The other parameters are set following (Liu et al., 2019). We do single task finetuning, meaning we only tune on the data provided for the given natural language understanding task. We do not perform ensembling. When finetuning models trained with LayerDrop, we apply LayerDrop and Quant-Noise during finetuning time as well. + +EfficientNet We use the architecture of EfficientNet-B3 defined in Classy Vision (Adcock et al., 2019) and follow the default hyperparameters for training. We set Quant-Noise value to 0.1. During training time, we searched over the parameters (0.05, 0.1, 0.2) to determine the optimal value of Quant-Noise. During training time, the block size of Quant-Noise is set to 4 for all $1 \times 1$ convolutions, 9 for depth-wise $3 \times 3$ convolutions, 5 for depth-wise $5 \times 5$ convolutions and 4 for the classifier. For sharing, we shared weights between blocks 9-10, 11-12, 14-15, 16-17, 19-20-21, 22-23 and refer to blocks that share the same weights as a chunk. For LayerDrop, we drop the chunks of blocks defined previously with probability 0.2 and evaluate only with chunks 9-10, 14-15 and 19-20-21. + +# 7.7 SCALAR QUANTIZATION DETAILS + +We closely follow the methodology of PyTorch 1.4. We emulate scalar quantization by quantizing the weights and the activations. The scales and zero points of activations are determined by doing a few forward passes ahead of the evaluation and then fixed. We use the Histogram method to compute $s$ and $z$ , which aims at approximately minimizing the $L _ { 2 }$ quantization error by adjusting $s$ and $z$ . This scheme is a refinement of the MinMax scheme. Per channel quantization is also discussed in Table 10. + +# 7.8 IPQ QUANTIZATION DETAILS + +Language Modeling We quantize FFN with block size 8, embeddings with block size 8, and attention with block size 4. We tuned the block size for attention between the values (4, 8) to find the best performance. Note that during training with apply Quant-Noise to all the layers. + +RoBERTa We quantize FFN with block size 4, embeddings with block size 4, and attention with block size 4. We tuned the block size between the values (4, 8) to find the best performance. Note that during training with apply Quant-Noise to all the layers. + +EfficientNet We quantize blocks sequentially and end up with the classifier. The block sizes are 4 for all $1 \times 1$ convolutions, 9 for depth-wise $3 \times 3$ convolutions, 5 for depth-wise $5 \times 5$ convolutions and 4 for the classifier. Note that during training with apply Quant-Noise to all the weights in InvertedResidual Blocks (except the Squeeze-Excitation subblocks), the head convolution and the classifier. + +Table 7: Performance on MNLI. We report accuracy and size in megabytes. \* indicates distillation using BERT Large. $\dagger$ indicates training with data augmentation. Work from Sun et al. (2019) and Zhao et al. (2019) do not report results on the dev set. Cao et al. do not report model size. Higher accuracy is better. + +
ModelMBMNLI
RoBERTa Base +LD (Fan et al., 2019)48084.8
BERT Base (Devlin et al.,2018)42084.4
PreTrained Distil (Turc et al., 2019)25782.5
DistilBERT (Sanh et al.,2019b)25081.8
MobileBERT* (Sun et al.)9684.4
TinyBERTt (Jiao et al.,2019)5582.8
ALBERT Base (Lan et al.,2019)4581.6
AdaBERTt (Chen et al.,2020)3681.6
Quant-Noise3883.6
Quant-Noise+ Share +Prune1482.5
+ +Table 8: Performance on ImageNet. We report accuracy and size in megabytes. Higher accuracy is better. + +
ModelMBAcc.
EfficientNet-B7 (Tan &Le,2019) ResNet-50 (He et al., 2015) DenseNet-169 (Huang et al., 2018) EfficientNet-BO (Tan &Le,2019) MobileNet-v2 (Sandler et al., 2018)260 97.5 53.4 20.2 13.484.4 76.1 76.2 77.3 71.9
Shufflenet-v2 ×1 (Ma et al.,2018) HAQ 4 bits (Wang et al., 2018) iPQ ResNet-50 (Stock et al.,2019)8.7 12.469.4 76.2
Quant-Noise5.09 3.376.1 80.0
Quant-Noise + Share + Prune2.377.8
+ +# 7.9 DETAILS OF PRUNING AND LAYER SHARING + +We apply the Every Other Layer strategy from Fan et al. (2019). When combining layer sharing with pruning, we train models with shared layers and then prune chunks of shared layers. When sharing layers, the weights of adjacent layers are shared in chunks of two. For a concrete example, imagine we have a model with layers A, B, C, D, E, F, G, H. We share layers A and B, C and D, E and F, G and H. To prune, every other chunk would be pruned away, for example we could prune A, B, E, F. + +# 7.10 NUMERICAL RESULTS FOR GRAPHICAL DIAGRAMS + +We report the numerical values displayed in Figures 2 in Table 6 for language modeling, Table 7 for BERT, and Table 8 for ImageNet. + +# 7.11 FURTHER ABLATIONS + +# 7.11.1 IMPACT OF QUANT-NOISE FOR THE VISION SETUP + +We provide another study showing the impact of the proportion of elements on which to apply Quant-Noise in Table 9. + +# 7.11.2 IMPACT OF THE NUMBER OF CENTROIDS + +We quantize with 256 centroids which represents a balance between size and representation capacity. The effect of the number of centroids on performance and size is shown in Figure 4 (a). Quantizing + +
p00.20.40.60.81
Top-180.6680.8380.8280.8880.9280.64
+ +Table 9: Effect of Quantization Parameters. We report the influence of the Quant-Noise rate $p$ with Scalar Quantization (int8). We focus on EfficientNet for ImageNet classification. + +![](images/cf825d2cf66ad807058d5706153a322d0dca0137253c54a0e00ac009d8f56a9a.jpg) +Figure 4: Quantizing with a larger number of centroids. Results are shown on Wikitext-103 valid. + +with more centroids improves perplexity — this parameter could be adjusted based on the practical storage constraints. + +# 7.11.3 EFFECT OF INITIAL MODEL SIZE + +Large, overparameterized models are more easily compressed. In Figure 5, we explore quantizing both shallower and skinnier models. For shallow models, the gap between quantized and non-quantized perplexity does not increase as layers are removed (Figure 5, left). In contrast, there is a larger gap in performance for models with smaller FFN (Figure 5, right). As the FFN size decreases, the weights are less redundant and more difficult to quantize with iPQ. + +# 7.11.4 DIFFICULTY OF QUANTIZING DIFFERENT MODEL STRUCTURES + +Quantization is applied to various portions of the Transformer architecture — the embedding, attention, feedforward, and classifier output. We compare the quantizability of various portions of the network in this section. + +Is the order of structures important? We quantize specific network structures first — this is important as quantizing weight matrices can accumulate reconstruction error. Some structures of the network should be quantized last so the finetuning process can better adjust the centroids. We find that there are small variations in performance based on quantization order (see Figure 6). We choose to quantize FFN, then embeddings, and finally the attention matrices in Transformer networks. + +Which structures can be compressed the most? Finally, we analyze which network structures can be most compressed. During quantization, various matrix block sizes can be chosen as a parameter — the larger the block size, the more compression, but also the larger the potential reduction of performance. Thus, it is important to understand how much each network structure can be compressed to reduce the memory footprint of the final model as much as possible. In Figure 6, we quantize two model structures with a fixed block size and vary the block size of the third between 4 and 32. As shown, the FFN and embedding structures are more robust to aggressive compression, while the attention drastically loses performance as larger block sizes are used. + +# 7.11.5 APPROACH TO I N TN SCALAR QUANTIZATION + +We compare quantizing per-channel to using a histogram quantizer in Table 10. The histogram quantizer maintains a running min/max and minimizes L2 distance between quantized and nonquantized values to find the optimal min/max. Quantizing per channel learns scales and offsets as vectors along the channel dimension, which provides more flexibility since scales and offsets can be different. + +![](images/89d51efee35f6dc30bb5ffaa4bb5b370bcc344263f7bf64ce5139ec44a66a0da.jpg) +Figure 5: (a) Effect of Initial Model Size for more shallow models (b) Effect of Initial Model Size more skinny models + +![](images/8012673edf3e43ed6c0eaae7f98a58780eb2aa1fbaa3915613839ed665d88982.jpg) +Figure 6: Effect of Quantization on Model Structures. Results are shown on the validation set of Wikitext-103. (a) Quantizing Attention, FFN, and Embeddings in different order. (b) More Extreme compression of different structures. + +# 7.11.6 LAYERDROP WITH STE + +For quantization noise, we apply the straight through estimator (STE) to remaining weights in the backward pass. We experiment with applying STE to the backward pass of LayerDrop’s pruning noise. Results are shown in Table 11 and find slightly worse results. + +Table 10: Comparison of different approaches to int4 and int8 with and without QuantNoise on language modeling and image classification. For language modeling, we train a Transformer on the Wikitext-103 benchmark. We report perplexity (PPL) on the test set. For image classification, we train a EfficientNet-B3 on the ImageNet-1K benchmark. We report top-1 accuracy on the validation set. For both setting, we also report model size in megabyte (MB) and the compression ratio compared to the original model. + +
Quantization SchemeLanguage Modeling 16-layer Transformer Wikitext-103Image Classification EfficientNet-B3 ImageNet-1K
SizeCompressTest PPLSizeCompressTop-1 Acc.
Uncompressed model942×118.346.7×181.5
Int4 Quant Histogram118×839.45.8×845.3
+ Quant-Noise118×821.85.8×867.8
Int4 Quant Channel118×821.25.8×868.2
+ Quant-Noise118×819.55.8×872.3
Int8 Quant Histogram236×419.611.7×480.7
+ Quant-Noise236×418.711.7×480.9
Int8 Quant Channel236×418.511.7×481.1
+ Quant-Noise236×418.311.7×481.2
+ +Table 11: Performance on Wikitext-103 when using STE in the backward pass of the LayerDrop pruning noise. + +
ModelMBPPL
Quant-Noise + Share + Prune1024.2
Quant-Noise + Share + Prune with STE1024.5
\ No newline at end of file diff --git a/parse/train/dV19Yyi1fS3/dV19Yyi1fS3_content_list.json b/parse/train/dV19Yyi1fS3/dV19Yyi1fS3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..48d8a1abe5b6b8cfd63e20bde597b0ec5baf70d6 --- /dev/null +++ b/parse/train/dV19Yyi1fS3/dV19Yyi1fS3_content_list.json @@ -0,0 +1,2575 @@ +[ + { + "type": "text", + "text": "TRAINING WITH QUANTIZATION NOISE FOREXTREME MODEL COMPRESSION", + "text_level": 1, + "bbox": [ + 174, + 99, + 704, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pierre Stock ∗ † Facebook AI Research, Inria ", + "bbox": [ + 184, + 170, + 374, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Angela Fan∗ Facebook AI Research, LORIA ", + "bbox": [ + 415, + 170, + 620, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benjamin Graham Facebook AI Research ", + "bbox": [ + 663, + 170, + 813, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Edouard Grave Facebook AI Research ", + "bbox": [ + 184, + 219, + 333, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rémi Gribonval Inria ", + "bbox": [ + 354, + 219, + 473, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hervé Jégou Facebook AI Research ", + "bbox": [ + 490, + 219, + 640, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Armand Joulin Facebook AI Research ", + "bbox": [ + 661, + 219, + 812, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 285, + 544, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We tackle the problem of producing compact models, maximizing their accuracy for a given model size. A standard solution is to train networks with Quantization Aware Training (Jacob et al., 2018), where the weights are quantized during training and the gradients approximated with the Straight-Through Estimator (Bengio et al., 2013). In this paper, we extend this approach to work beyond int8 fixedpoint quantization with extreme compression methods where the approximations introduced by STE are severe, such as Product Quantization. Our proposal is to only quantize a different random subset of weights during each forward, allowing for unbiased gradients to flow through the other weights. Controlling the amount of noise and its form allows for extreme compression rates while maintaining the performance of the original model. As a result we establish new state-of-the-art compromises between accuracy and model size both in natural language processing and image classification. For example, applying our method to state-of-the-art Transformer and ConvNet architectures, we can achieve $8 2 . 5 \\%$ accuracy on MNLI by compressing RoBERTa to $1 4 \\mathbf { M B }$ and $8 0 . 0 \\%$ top-1 accuracy on ImageNet by compressing an EfficientNet-B3 to $3 . 3 \\mathrm { M B }$ . 1 ", + "bbox": [ + 232, + 318, + 766, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 571, + 336, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many of the best performing neural network architectures in real-world applications have a large number of parameters. For example, the current standard machine translation architecture, Transformer (Vaswani et al., 2017), has layers that contain millions of parameters. Even models that are designed to jointly optimize the performance and the parameter efficiency, such as EfficientNets (Tan & Le, 2019), still require dozens to hundreds of megabytes, which limits their applications to domains like robotics or virtual assistants. ", + "bbox": [ + 174, + 604, + 825, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model compression schemes reduce the memory footprint of overparametrized models. Pruning (LeCun et al., 1990) and distillation (Hinton et al., 2015) remove parameters by reducing the number of network weights. In contrast, quantization focuses on reducing the bits per weight. This makes quantization particularly interesting when compressing models that have already been carefully optimized in terms of network architecture. Whereas deleting weights or whole hidden units will inevitably lead to a drop in performance, we demonstrate that quantizing the weights can be performed with little to no loss in accuracy. ", + "bbox": [ + 174, + 695, + 825, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Popular postprocessing quantization methods, like scalar quantization, replace the floating-point weights of a trained network by a lower-precision representation, like fixed-width integers (Vanhoucke et al., 2011). These approaches achieve a good compression rate with the additional benefit of accelerating inference on supporting hardware. However, the errors made by these approximations accumulate in the computations operated during the forward pass, inducing a significant drop in performance (Stock et al., 2019). ", + "bbox": [ + 173, + 800, + 825, + 856 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8fa608fc575221a61a670b21fc8efe8983cacc3f544405b658da957ea4c631e.jpg", + "image_caption": [ + "Figure 1: Quant-Noise trains models to be resilient to inference-time quantization by mimicking the effect of the quantization method during training time. This allows for extreme compression rates without much loss in accuracy on a variety of tasks and benchmarks. " + ], + "image_footnote": [], + "bbox": [ + 184, + 101, + 820, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 282, + 821, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A solution to address this drifting effect is to directly quantize the network during training. This raises two challenges. First, the discretization operators have a null gradient — the derivative with respect to the input is zero almost everywhere. This requires special workarounds to train a network with these operators. The second challenge that often comes with these workarounds is the discrepancy that appears between the train and test functions implemented by the network. Quantization Aware Training (QAT) (Jacob et al., 2018) resolves these issues by quantizing all the weights during the forward and using a straight through estimator (STE) (Bengio et al., 2013) to compute the gradient. This works when the error introduced by STE is small, like with int8 quantization, but does not suffice in compression regimes where the approximation made by the compression is more severe. ", + "bbox": [ + 173, + 318, + 825, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we show that quantizing only a subset of weights instead of the entire network during training is more stable for high compression schemes. Indeed, by quantizing only a random fraction of the network at each forward, most the weights are updated with unbiased gradients. Interestingly, we show that our method can employ a simpler quantization scheme during the training. This is particularly useful for quantizers with trainable parameters, such as Product Quantizer (PQ), for which our quantization proxy is not parametrized. Our approach simply applies a quantization noise, called Quant-Noise, to a random subset of the weights, see Figure 1. We observe that this makes a network resilient to various types of discretization methods: it significantly improves the accuracy associated with (a) low precision representation of weights like int8; and (b) state-of-the-art PQ. Further, we demonstrate that Quant-Noise can be applied to existing trained networks as a post-processing step, to improve the performance network after quantization. ", + "bbox": [ + 173, + 449, + 825, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this paper makes the following contributions: ", + "bbox": [ + 174, + 609, + 555, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We introduce the Quant-Noise technique to learn networks that are more resilient to a variety of quantization methods such as int4, int8, and PQ; • Adding Quant-Noise to PQ leads to new state-of-the-art trade-offs between accuracy and model size. For instance, for natural language processing (NLP), we reach $8 2 . 5 \\%$ accuracy on MNLI by compressing RoBERTa to 14 MB. Similarly for computer vision, we report $8 0 . 0 \\%$ top-1 accuracy on ImageNet by compressing an EfficientNet-B3 to $3 . 3 { \\mathrm { M B } }$ ; • By combining PQ and int8 to quantize weights and activations for networks trained with Quant-Noise, we obtain extreme compression with fixed-precision computation and achieve $7 9 . 8 \\%$ top-1 accuracy on ImageNet and 21.1 perplexity on WikiText-103. ", + "bbox": [ + 215, + 636, + 825, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 795, + 344, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model compression. Many compression methods focus on efficient parameterization, via weight pruning (LeCun et al., 1990; Li et al., 2016; Huang et al., 2018; Mittal et al., 2018), weight sharing (Dehghani et al., 2018; Turc et al., 2019; Lan et al., 2019) or with dedicated architectures (Tan & Le, 2019; Zhang et al., 2017; Howard et al., 2019). Weight pruning is implemented during training (Louizos et al., 2017) or as a fine-tuning post-processing step (Han et al., 2015; 2016). Many pruning methods are unstructured, i.e., remove individual weights (LeCun et al., 1990; Molchanov et al., 2017). On the other hand, structured pruning methods follow the structure of the weights to ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reduce both the memory footprint and the inference time of a model (Li et al., 2016; Luo et al., 2017; \nFan et al., 2019). We refer the reader to Liu et al. (2018) for a review of different pruning strategies. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other authors have worked on lightweight architectures, by modifying existing models (Zhang et al., 2018; Wu et al., 2019; Sukhbaatar et al., 2019a) or developing new networks, such as MobileNet (Howard et al., 2019), ShuffleNet (Zhang et al., 2017), and EfficientNet (Tan & Le, 2019) in vision. ", + "bbox": [ + 176, + 138, + 823, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, knowledge distillation (Hinton et al., 2015) has been applied to sentence representation (Turc et al., 2019; Sanh et al., 2019a; Sun et al., 2019; Zhao et al., 2019; Jiao et al., 2019), to reduce the size of a BERT model (Devlin et al., 2018). ", + "bbox": [ + 174, + 188, + 825, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quantization. There are extensive studies of scalar quantization to train networks with lowprecision weights and activations (Courbariaux et al., 2015; Courbariaux & Bengio, 2016; Rastegari et al., 2016; McDonnell, 2018). These methods benefit from specialized hardware to also improve the runtime during inference (Vanhoucke et al., 2011). Other quantization methods such as Vector Quantization (VQ) and PQ (Jegou et al., 2011) quantize blocks of weights simultaneously to achieve higher compression rate (Stock et al., 2019; Gong et al., 2014; Joulin et al., 2016; Carreira-Perpiñán & Idelbayev, 2017). Closer to our work, several works have focused at simultaneously training and quantizing a network (Jacob et al., 2018; Krishnamoorthi, 2018; Gupta et al., 2015; Dong et al., 2019). Gupta et al. (2015) assigns weights to a quantized bin stochastically which is specific to scalar quantization, but allows training with fixed point arithmetic. Finally, our method can be interpreted as a form of Bayesian compression (Louizos et al., 2017), using the Bayesian interpretation of Dropout (Gal & Ghahramani, 2016). As opposed to their work, we select our noise to match the weight transformation of a target quantization methods without restricting it to a scale mixture prior. ", + "bbox": [ + 173, + 244, + 826, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 QUANTIZING NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 174, + 444, + 495, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the principles of quantization, several standard quantization methods, and describe how to combine scalar and product quantization. For clarity, we focus on the case of a fixed real matrix $\\mathbf { W } \\in \\mathbf { R } ^ { n \\times p }$ . We suppose that this matrix is split into $m \\times q$ blocks $\\mathbf { b } _ { k l }$ : ", + "bbox": [ + 174, + 474, + 825, + 517 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/964a7fabb8aac9db9e4fe8cf40dd2b4eb13eeffda3a3ea97ea73cd443e5369fb.jpg", + "text": "$$\n\\mathbf { W } = \\left( \\sum _ { \\vdots } ^ { \\mathbf { b } _ { 1 1 } } \\mathrm { ~ ~ \\cdots ~ } \\sum _ { \\mathbf { b } _ { m q } } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 400, + 518, + 594, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the nature of these blocks is determined by the quantization method. A codebook is a set of $K$ vectors, i.e., ${ \\mathcal { C } } = \\{ \\mathbf { c } [ 1 ] , \\dots , \\mathbf { c } [ K ] \\}$ . Quantization methods compress the matrix W by assigning to each block $\\mathbf { b } _ { k l }$ an index that points to a codeword $\\mathbf { c }$ in a codebook $\\mathcal { C }$ , and storing the codebook $\\mathcal { C }$ and the resulting indices (as the entries ${ \\bf { I } } _ { k l }$ of an index matrix $\\mathbf { I }$ ) instead of the real weights. During the inference, they reconstruct an approximation $\\widehat { \\bf W }$ of the original matrix W such that $\\widehat { \\mathbf { b } } _ { k l } = \\mathbf { c } [ \\mathbf { I } _ { k l } ]$ . ", + "bbox": [ + 173, + 577, + 825, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We distinguish scalar quantization, such as int8, where each block $\\mathbf { b } _ { k l }$ consists of a single weight, from vector quantization, where several weights are quantized jointly. ", + "bbox": [ + 174, + 656, + 821, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 FIXED-POINT SCALAR QUANTIZATION ", + "text_level": 1, + "bbox": [ + 176, + 700, + 482, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fixed-point scalar quantization methods replace floating-point number representations by lowprecision fixed-point representations. They simultaneously reduce a model’s memory footprint and accelerate inference by using fixed-point arithmetic on supporting hardware. ", + "bbox": [ + 176, + 727, + 823, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fixed-point scalar quantization operates on blocks that represent a single weight, i.e., $\\mathbf { b } _ { k l } = \\mathbf { W } _ { k l }$ Floating-point weights are replaced by $N$ bit fixed-point numbers (Gupta et al., 2015), with the extreme case of binarization where $N = 1$ (Courbariaux et al., 2015). More precisely, the weights are rounded to one of $2 ^ { N }$ possible codewords. These codewords correspond to bins evenly spaced by a scale factor $s$ and shifted by a bias $z$ . Each weight ${ \\bf W } _ { k l }$ is mapped to its nearest codeword $c$ by successively quantizing with $z \\mapsto$ round $( \\mathbf { W } _ { k l } / s + z )$ and dequantizing with the inverse operation: ", + "bbox": [ + 173, + 776, + 825, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/e16750d411177f96ab39549c5c49a3cceefbac25e39f103287de9eb36a9932f0.jpg", + "text": "$$\n\\begin{array} { r } { \\mathbf { c } = ( \\mathrm { r o u n d } ( \\mathbf { W } _ { k l } / s + z ) - z ) \\times s , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 382, + 862, + 614, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we compute the scale and bias as: ", + "bbox": [ + 176, + 881, + 439, + 895 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/e2852789dc75d687ebb7ab34653d54b82cca508c4d7bdfb0192012cda505ddbc.jpg", + "text": "$$\ns = { \\frac { \\operatorname* { m a x } \\mathbf { W } - \\operatorname* { m i n } \\mathbf { W } } { 2 ^ { N } - 1 } } \\quad { \\mathrm { a n d } } \\quad z = \\operatorname { r o u n d } ( \\operatorname* { m i n } \\mathbf { W } / s ) .\n$$", + "text_format": "latex", + "bbox": [ + 320, + 896, + 676, + 929 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus on this uniform rounding scheme instead of other non-uniform schemes (Choi et al., 2018; Li et al., 2019), because it allows for fixed-point arithmetic with implementations in PyTorch and Tensorflow (see Appendix). The compression rate is $\\times 3 2 / N$ . The activations are also rounded to $N$ -bit fixed-point numbers. With int8 for instance, this leads to $\\times 2$ to $\\times 4$ faster inference on dedicated hardware. In this work, we consider both int4 and int8 quantization. ", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 PRODUCT QUANTIZATION ", + "text_level": 1, + "bbox": [ + 176, + 190, + 395, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Several quantization methods work on groups of weights, such as vectors, to benefit from the correlation induced by the structure of the network. In this work, we focus on Product Quantization for its good performance at extreme compression ratio (Stock et al., 2019). ", + "bbox": [ + 176, + 215, + 825, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditional PQ. In vector quantization methods, the blocks are predefined groups of weights instead of single weights. The codewords are groups of values, and the index matrix I maps groups of weights from the matrix W to these codewords. In this section, we present the Product Quantization framework as it generalizes both scalar and vector quantization. We consider the case where we apply PQ to the columns of W and thus assume that $q = p$ . ", + "bbox": [ + 174, + 272, + 825, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditional vector quantization techniques split the matrix $\\mathbf { W }$ into its $p$ columns and learn a codebook on the resulting $p$ vectors. Instead, Product Quantization splits each column into $m$ subvectors and learns the same codebook for each of the resulting $m \\times p$ subvectors. Each quantized vector is subsequently obtained by assigning its subvectors to the nearest codeword in the codebook. Learning the codebook is traditionally done using $k$ -means with a fixed number $K$ of centroids, typically $K = 2 5 6$ to store the index matrix I using int8. Thus, the objective function is written as: ", + "bbox": [ + 173, + 348, + 825, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/e1589c4b824758805134b321a9fdc122c101ee402493d4ab7d6951e5c3ea9bf2.jpg", + "text": "$$\n\\Vert \\mathbf { W } - \\widehat { \\mathbf { W } } \\Vert _ { 2 } ^ { 2 } = \\sum _ { k , l } \\Vert \\mathbf { b } _ { k l } - \\mathbf { c } [ \\mathbf { I } _ { k l } ] \\Vert _ { 2 } ^ { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 377, + 435, + 620, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PQ shares representations between subvectors, which allows for higher compression rates than intN. ", + "bbox": [ + 173, + 473, + 825, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Iterative PQ. When quantizing a full network rather than a single matrix, extreme compression with PQ induces a quantization drift as reconstruction error accumulates (Stock et al., 2019). Indeed, subsequent layers take as input the output of preceding layers, which are modified by the quantization of the preceding layers. This creates a drift in the network activations, resulting in large losses of performance. A solution proposed by Stock et al. (2019), which we call iterative PQ (iPQ), is to quantize layers sequentially from the lowest to the highest, and finetune the upper layers as the lower layers are quantized, under the supervision of the uncompressed (teacher) model. Codewords of each layer are finetuned by averaging the gradients of their assigned elements with gradient steps: ", + "bbox": [ + 173, + 501, + 825, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/32aaf674e40606ef3ee77a071678bcc3936a2211c93f534e1722ca553c9309eb.jpg", + "text": "$$\n\\mathbf { c } \\gets \\mathbf { c } - \\eta \\frac { 1 } { | J _ { \\mathbf { c } } | } \\sum _ { ( k , l ) \\in J _ { \\mathbf { c } } } \\frac { \\partial \\mathcal { L } } { \\partial \\mathbf { b } _ { k l } } ,\n$$", + "text_format": "latex", + "bbox": [ + 398, + 616, + 596, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $J _ { \\mathbf { c } } = \\{ ( k , l ) | \\mathbf { c } [ \\mathbf { I } _ { k l } ] = \\mathbf { c } \\}$ , $\\mathcal { L }$ is the loss function and $\\eta > 0$ is a learning rate. This adapts the upper layers to the drift appearing in their inputs, reducing the impact of the quantization approximation on the overall performance. ", + "bbox": [ + 174, + 659, + 825, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 COMBINING FIXED-POINT WITH PRODUCT QUANTIZATION ", + "text_level": 1, + "bbox": [ + 174, + 718, + 624, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fixed-point quantization and Product Quantization are often regarded as competing choices, but can be advantageously combined. Indeed, PQ/iPQ compresses the network by replacing vectors of weights by their assigned centroids, but these centroids are in floating-point precision. Fixed-point quantization compresses both activations and weights to fixed-point representations. Combining both approaches means that the vectors of weights are mapped to centroids that are compressed to fixed-point representations, along with the activations. This benefits from the extreme compression ratio of iPQ and the finite-precision arithmetics of intN quantization. ", + "bbox": [ + 173, + 743, + 825, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More precisely, for a given matrix, we store the $\\dot { \\mathtt { 1 } } \\mathtt { n t } \\mathtt { 8 }$ representation of the $K$ centroids of dimension $d$ along with the $\\log _ { 2 } K$ representations of the centroid assignments of the $m \\times p$ subvectors. The int8 representation of the centroids is obtained with Eq. (2). The overall storage of the matrix and activations during a forward pass with batch size 1 (recalling that the input dimension is n) writes ", + "bbox": [ + 174, + 849, + 825, + 905 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/86485819a57974e638e685055efe01a32d93daf59595e61fd68de224d3f7a79f.jpg", + "text": "$$\nM = 8 \\times K d + \\log _ { 2 } { K } \\times m p + 8 \\times n \\mathrm { b i t s } .\n$$", + "text_format": "latex", + "bbox": [ + 352, + 909, + 647, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, when $K = 2 5 6$ , the centroid assignments are also stored in int8, which means that every value required for a forward pass is stored in an int8 format. We divide by 4 the float32 overhead of storing the centroids, although the storage requirement associated with the centroids is small compared to the cost of indexing the subvectors for standard networks. In contrast to iPQ alone where we only quantize the weights, we also quantize the activations using int8. We evaluate this approach on both natural language processing and computer vision tasks in Section 5. ", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 METHOD ", + "text_level": 1, + "bbox": [ + 174, + 207, + 282, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deep networks are not exposed to the noise caused by the quantization drift during training, leading to suboptimal performance. A solution to make the network robust to quantization is to introduce it during training. Quantization Aware Training (QAT) (Jacob et al., 2018) exposes the network during training by quantizing weights during the forward pass. This transformation is not differentiable and gradients are approximated with a straight through estimator (STE) (Bengio et al., 2013; Courbariaux & Bengio, 2016). STE introduces a bias in the gradients that depends on level of quantization of the weights, and thus, the compression ratio. In this section, we propose a simple modification to control this induced bias with a stochastic amelioration of QAT, called Quant-Noise. The idea is to quantize a randomly selected fraction of the weights instead of the full network as in QAT, leaving some unbiased gradients flow through unquantized weights. Our general formulation can simulate the effect of both quantization and of pruning during training. ", + "bbox": [ + 174, + 238, + 825, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 TRAINING NETWORKS WITH QUANTIZATION NOISE", + "text_level": 1, + "bbox": [ + 174, + 407, + 575, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider the case of a real matrix W as in Section 3. During the training of a network, our proposed Quant-Noise method works as follows: first, we compute blocks $\\mathbf { b } _ { k l }$ related to a target quantization method. Then, during each forward pass, we randomly select a subset of these blocks and apply some distortion to them. ", + "bbox": [ + 174, + 433, + 825, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More formally, given a set of tuples of indices $J \\subset \\{ ( k , l ) \\}$ for $1 \\leq k \\leq m$ , $1 \\leq l \\leq q$ and a distortion or noise function $\\varphi$ acting on a block, we define an operator $\\psi ( \\cdot \\mid J )$ such that, for each block $\\mathbf { b } _ { k l }$ , we apply the following transformation: ", + "bbox": [ + 174, + 496, + 825, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/10ee719ee23e228ab52579347f884b2d067f62e786a1fef94a0d603043df124f.jpg", + "text": "$$\n\\psi ( \\mathbf { b } _ { k l } \\mid J ) = { \\left\\{ \\begin{array} { l l } { \\varphi ( \\mathbf { b } _ { k l } ) } & { { \\mathrm { i f ~ } } ( k , l ) \\in J , } \\\\ { \\mathbf { b } _ { k l } } & { { \\mathrm { o t h e r w i s e } } . } \\end{array} \\right. }\n$$", + "text_format": "latex", + "bbox": [ + 370, + 545, + 624, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The noise function $\\varphi$ simulates the change in the weights produced by the target quantization method (see Section 4.2 for details). We replace the matrix W by the resulting noisy matrix $\\mathbf { W } _ { \\mathrm { n o i s e } }$ during the forward pass to compute a noisy output $\\mathbf { y } _ { \\mathrm { n o i s e } }$ , i.e., ", + "bbox": [ + 173, + 587, + 825, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/24084491a569e90db26327de49d058bc0584d8efdcaebbd29ff4bd3dd4eec519.jpg", + "text": "$$\n\\mathbf { W } _ { \\mathrm { n o i s e } } = ( \\psi ( \\mathbf { b } _ { k l } \\mid J ) ) _ { k l } \\mathrm { a n d } \\mathbf { y } _ { \\mathrm { n o i s e } } = \\mathbf { x } \\mathbf { W } _ { \\mathrm { n o i s e } }\n$$", + "text_format": "latex", + "bbox": [ + 326, + 633, + 671, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf { x }$ is an input vector. During the backward pass, we apply STE, which amounts to replacing the distorted weights $\\mathbf { W _ { \\mathrm { n o i s e } } }$ by their non-distorted counterparts. Note that our approach is equivalent to QAT when $J$ containts all the tuples of indices. However, an advantage of Quant-Noise over QAT is that unbiased gradients continue to flow via blocks unaffected by the noise. As these blocks are randomly selected for each forward, we guarantee that each weight regularly sees gradients that are not affected by the nature of the function $\\varphi$ . As a side effect, our quantization noise regularizes the network in a similar way as DropConnect (Wan et al., 2013) or LayerDrop (Fan et al., 2019). ", + "bbox": [ + 173, + 655, + 825, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Composing quantization noises. As noise operators are compositionally commutative, we can make a network robust to a combination of quantization methods by composing their noise operators: ", + "bbox": [ + 173, + 768, + 823, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/9bee3ba948403dbd6beab1b7a7c77744b4c7439e50c7b1f8b1055560b5e7373d.jpg", + "text": "$$\n\\psi ( \\mathbf { b } _ { k l } \\mid J ) = \\psi _ { 1 } \\circ \\psi _ { 2 } ( \\mathbf { b } _ { k l } \\mid J ) .\n$$", + "text_format": "latex", + "bbox": [ + 393, + 801, + 604, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This property is particularly useful to combine quantization with pruning operators during training, as well as combining scalar quantization with product quantization. ", + "bbox": [ + 174, + 824, + 828, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 ADDING NOISE TO SPECIFIC QUANTIZATION METHODS", + "text_level": 1, + "bbox": [ + 174, + 869, + 601, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we propose several implementations of the noise function $\\varphi$ for the quantization methods described in Section 3. We also show how to handle pruning with it. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5d264ad90c488b99c0c3d77e3d45008a2d34a1f260089c587f6dd7bd47a555b5.jpg", + "table_caption": [ + "Table 1: Comparison of different quantization schemes with and without Quant-Noise on language modeling and image classification. For language modeling, we train a Transformer on the Wikitext-103 benchmark and report perplexity (PPL) on test. For image classification, we train a EfficientNet-B3 on the ImageNet-1k benchmark and report top-1 accuracy on validation and use our re-implementation of EfficientNet-B3. The original implementation of Tan & Le (2019) achieves an uncompressed Top-1 accuracy of $8 1 . 9 \\%$ . For both settings, we report model size in megabyte (MB) and the compression ratio compared to the original model. " + ], + "table_footnote": [], + "table_body": "
Quantization SchemeLanguage Modeling 16-layer Transformer Wikitext-103Image Classification EfficientNet-B3 ImageNet-1k
SizeCompressionPPLSizeCompressionTop-1
Uncompressed model942×118.346.7×181.5
int 4 quantization118×839.45.8×845.3
- trained with QAT118×834.15.8×859.4
- trained with Quant-Noise118×821.85.8×867.8
int 8 quantization236×419.611.7×480.7
- trained with QAT236×421.011.7×480.8
- trained with Quant-Noise236×418.711.7×480.9
iPQ38×2525.23.3×1479.0
- trained with QAT38×2541.23.3×1455.7
- trained with Quant-Noise38×2520.73.3×1480.0
iPQ&int8 +Quant-Noise38×2521.13.1×1579.8
", + "bbox": [ + 173, + 101, + 826, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fixed-point scalar quantization. In intN quantization, the blocks are atomic and weights are rounded to their nearest neighbor in the codebook. The function $\\varphi$ replaces weight ${ \\bf W } _ { k l }$ with the output of the rounding function defined in Eq. (2), i.e., ", + "bbox": [ + 173, + 453, + 823, + 496 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/4eede465b895d5a0de674fafccfdd74e0c3a29c6654534873f22a4d66c143d09.jpg", + "text": "$$\n\\varphi _ { \\mathrm { i n t N } } ( w ) = ( \\mathrm { r o u n d } ( w / s + z ) - z ) \\times s ,\n$$", + "text_format": "latex", + "bbox": [ + 364, + 501, + 632, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s$ and $z$ are updated during training. In particular, the application of Quant-Noise to int8 scalar quantization is a stochastic amelioration of QAT. ", + "bbox": [ + 176, + 523, + 825, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Product quantization. As opposed to intN, codebooks in PQ require a clustering step based on weight values. During training, we learn codewords online and use the resulting centroids to implement the quantization noise. More precisely, the noise function $\\varphi _ { \\mathrm { P Q } }$ assigns a selected block b to its nearest codeword in the associated codebook $\\mathcal { C }$ : ", + "bbox": [ + 173, + 566, + 825, + 622 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/bddf3500b182bea57c4ce85ebf76ff5d4b8b43f674f4d9ba0d6ac97ba254d700.jpg", + "text": "$$\n\\begin{array} { r } { \\varphi _ { \\mathrm { P Q } } ( \\mathbf { v } ) = \\operatorname * { a r g m i n } _ { \\mathbf { c } \\in \\mathcal { C } } \\| \\mathbf { b } - \\mathbf { c } \\| _ { 2 } ^ { 2 } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 392, + 627, + 604, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Updating the codebooks online works well. However, empirically, running $k$ -means once per epoch is faster and does not noticeably modify the resulting accuracy. ", + "bbox": [ + 174, + 650, + 823, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that computing the exact noise function for PQ is computationally demanding. We propose a simpler and faster alternative approximation $\\varphi _ { \\mathrm { p r o x y } }$ to the operational transformation of PQ and iPQ. The noise function simply zeroes out the subvectors of the selected blocks, i.e., $\\varphi _ { \\mathrm { p r o x y } } ( \\mathbf { v } ) = 0$ . As a sidenote, we considered other alternatives, for instance one where the subvectors are mapped to the mean subvector. In practice, we found that these approximations lead to similar performance, see Section 7.2. This proxy noise function is a form of Structured Dropout and encourages correlations between the subvectors. This correlation is beneficial to the subsequent clustering involved in PQ/iPQ. ", + "bbox": [ + 174, + 685, + 825, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adding pruning to the quantization noise. The specific form of quantization noise can be adjusted to incorporate additional noise specific to pruning. We simply combine the noise operators of quantization and pruning by composing them following Eq. (8). We consider the pruning noise function of Fan et al. (2019) where they randomly drop predefined structures during training. In particular, we focus on LayerDrop, where the structures are the residual blocks of highway-like layers (Srivastava et al., 2015), as most modern architectures, such as ResNet or Transformer, are composed of this structure. More precisely, the corresponding noise operator over residual blocks $\\mathbf { v }$ is $\\dot { \\varphi } _ { \\mathrm { L a y e r D r o p } } ( \\mathbf { v } ) = 0 .$ . For pruning, we do not use STE to backpropagate the gradient of pruned weights, as dropping them entirely during training has the benefit of speeding convergence (Huang et al., 2016). Once a model is trained with LayerDrop, the number of layers kept at inference can be adapted to match computation budget or time constraint. ", + "bbox": [ + 173, + 797, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb191eba2d0809d5fb2682e2e7926abb2bbee75ac52ef027ff222f8961c973a6.jpg", + "image_caption": [ + "Figure 2: Performance as a function of model size. We compare models quantized with PQ and trained with the related Quant-Noise to the state of the art. (a) Test perplexity on Wikitext-103 (b) Dev Accuracy on MNLI (c) ImageNet Top-1 accuracy. Model size is shown in megabytes on a log scale. Red and gray coloring indicates existing work, with different colors for visual distinction. " + ], + "image_footnote": [], + "bbox": [ + 173, + 109, + 818, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/047fa78600dde1ff2aa05dc217c54159eb54aa183372c3e855fcdd8d283de36a.jpg", + "table_caption": [ + "Table 2: Decomposing the impact of the different compression schemes. (a) we train Transformers with Adaptive Input and LayerDrop on Wikitext-103 (b) we pre-train RoBERTA base models with LayerDrop and then finetune on MNLI (c) we train an EfficientNet-B3 on ImageNet. We report the compression ratio w.r.t. to the original model (“comp.”) and the resulting size in MB. " + ], + "table_footnote": [], + "table_body": "
Language modelingSentence RepresentationImage Classification
Comp.SizePPLComp.SizeAcc.Comp.SizeAcc.
Unquantized models
Original model×194218.3×148084.8×146.781.5
+ Sharing×1.851018.7×1.925084.0×1.434.280.1
+ Pruning×3.725522.5×3.812581.3×1.629.578.5
Quantized models
iPQ×24.83825.2×12.63882.5× 14.13.379.0
+ Quant-Noise×24.83820.7×12.63883.6×14.13.380.0
+ Sharing× 49.51922.0×34.31482.5×182.678.9
+ Pruning×94.21024.7×58.5878.8×202.377.8
", + "bbox": [ + 173, + 314, + 826, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 554, + 823, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 606, + 281, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the impact of Quant-Noise on the performance of several quantization schemes in a variety of settings (see Appendix - Sec. 7.5). ", + "bbox": [ + 174, + 638, + 825, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 IMPROVING COMPRESSION WITH QUANT-NOISE", + "text_level": 1, + "bbox": [ + 174, + 686, + 549, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quant-Noise is a regularization method that makes networks more robust to the target quantization scheme or combination of quantization schemes during training. We show the impact of Quant-Noise in Table 1 for a variety of quantization methods: int8/int4 and iPQ. ", + "bbox": [ + 176, + 714, + 825, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We experiment in 2 different settings: a Transformer network trained for language modeling on WikiText-103 and a EfficientNet-B3 convolutional network trained for image classification on ImageNet-1k. Our quantization noise framework is general and flexible — Quant-Noise improves the performance of quantized models for every quantization scheme in both experimental settings. Importantly, Quant-Noise only changes model training by adding a regularization noise similar to dropout, with no impact on convergence and very limited impact on training speed $( < 5 \\%$ slower). ", + "bbox": [ + 174, + 762, + 825, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This comparison of different quantization schemes shows that Quant-Noise works particularly well with high performance quantization methods, like iPQ, where QAT tends to degrade the performances, even compared to quantizing as a post-processing step. In subsequent experiments in this section, we focus on applications with iPQ because it offers the best trade-off between model performance and compression, and has little negative impact on FLOPS. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0bcbe8eb7024030a040baf34b938f5c85fe829c5bb1511e052acbe66bc7ad954.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Quant-Noise: Finetuning vs training. We report performance after iPQ quantization. We train with the $\\phi _ { \\mathrm { p r o x y } }$ noise and finetune with Quant-Noise, and use it during the transfer to MNLI for each RoBERTa model. " + ], + "table_body": "
Language ModelingPPLRoBERTaAcc.
Train without Quant-Noise25.2Train without Quant-Noise82.5
+ Finetune with Quant-Noise20.9+ Finetune with Quant-Noise83.4
Train with Quant-Noise20.7Train with Quant-Noise83.6
", + "bbox": [ + 225, + 101, + 769, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fixed-Point Product Quantization. Combining iPQ and int8 as described in Section 3.3 allows us to take advantage of the high compression rate of iPQ with a fixed-point representation of both centroids and activations. As shown in Table 1, this combination incurs little loss in accuracy with respect to $\\mathrm { i } \\mathrm { P Q } +$ Quant-Noise. Most of the memory footprint of iPQ comes from indexing and not storing centroids, so the compression ratios are comparable. ", + "bbox": [ + 174, + 231, + 825, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complementarity with Weight Pruning and Sharing. We analyze how Quant-Noise is compatible and complementary with pruning (“+Prune”) and weight sharing ( $^ { 6 6 } { + } \\mathrm { S l }$ hare”), see Appendix for details on weight sharing. We report results for Language modeling on WikiText-103, pre-trained sentence representations on MNLI and object classification on ImageNet-1k in Table 2. The conclusions are remarkably consistent across tasks and benchmarks: Quant-Noise gives a large improvement over strong iPQ baselines. Combining it with sharing and pruning offers additional interesting operating points of performance vs size. ", + "bbox": [ + 174, + 318, + 825, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 COMPARISON WITH THE STATE OF THE ART ", + "text_level": 1, + "bbox": [ + 174, + 434, + 514, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now compare our approach on the same tasks against the state of the art. We compare $\\mathrm { i } \\mathrm { P Q } +$ Quant-Noise with 6 methods of network compression for Language modeling, 8 state-of-the-art methods for Text classification, and 8 recent methods evaluate image classification on Imagenet with compressed models. These comparisons demonstrate that Quant-Noise leads to extreme compression rates at a reasonable cost in accuracy. We apply our best quantization setup on competitive models and reduce their memory footprint by $\\times 2 0 - 9 4 $ when combining with weight sharing and pruning, offering extreme compression for good performance. ", + "bbox": [ + 174, + 460, + 825, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Natural Language Processing. In Figure 2, we examine the trade-off between performance and model size. Our quantized RoBERTa offers a competitive trade-off between size and performance with memory reduction methods dedicated to BERT, like TinyBERT, MobileBERT, or AdaBERT. ", + "bbox": [ + 174, + 575, + 825, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image Classification. We compress EfficientNet-B3 from 46.7Mb to $3 . 3 { \\mathrm { M b } }$ ( $\\times 1 4$ compression) while maintaining high top-1 accuracy $( 7 8 . 5 \\%$ versus $8 0 \\%$ for the original model). As shown in Figure 2, our quantized EfficientNet-B3 is smaller and more accurate than architectures dedicated to optimize on-device performance with limited size like MobileNet or ShuffleNet. We further evaluate the beneficial effect of Quant-Noise on ResNet-50 to compare directly with Stock et al. (2019). Results shown in Table 4 indicate improvement with Quant-Noise compared to previous work. ", + "bbox": [ + 174, + 633, + 825, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Incorporating pruning noise into quantization is also beneficial. For example, with pruning $\\mathrm { i P Q + }$ Quant-Noise reduces size by $\\times 2 5$ with only a drop of $2 . 4 \\ : \\mathrm { P P L }$ in language modeling. Further, pruning reduces FLOPS by the same ratio as its compression factor, in our case, $\\times 2$ . By adding sharing with pruning, in language modeling, we achieve an extreme compression ratio of $\\times 9 4$ with a drop of $6 . 4 \\ : \\mathrm { P P L }$ with FLOPS reduction from pruning entire shared chunks of layers. For comparison, our $1 0 \\mathbf { M B }$ model has the same performance as the $5 7 0 \\mathrm { M B }$ Transformer-XL base. ", + "bbox": [ + 174, + 724, + 825, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 FINETUNING WITH QUANT-NOISE FOR POST-PROCESSING QUANTIZATION", + "text_level": 1, + "bbox": [ + 176, + 827, + 727, + 842 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explore taking existing models and post-processing with Quant-Noise instead of training from scratch. For language modeling, we train for 10 additional epochs. For RoBERTa, we train for $2 5 \\mathrm { k }$ additional updates. Finetuning with Quant-Noise incorporates the benefits and almost matches training from scratch (Table 3). In language modeling, there is only a $0 . 2 \\ : \\mathrm { P P L }$ difference. We further examine how to incorporate Quant-Noise more flexibly into pretraining RoBERTa. We take an already trained RoBERTa model and incorporate Quant-Noise during sentence classification finetuning. This is effective at compressing while retaining accuracy after quantization. ", + "bbox": [ + 174, + 853, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 152, + 318, + 169 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We show that quantizing a random subset of weights during training maintains performance in the high quantization regime. We validate that Quant-Noise works with a variety of different quantization schemes on several applications in text and vision. Our method can be applied to a combination of iPQ and int8 to benefit from extreme compression ratio and fixed-point arithmetic. Finally, we show that Quant-Noise can be used as a post-processing step to prepare already trained networks for subsequent quantization, to improve the performance of the compressed model. ", + "bbox": [ + 174, + 184, + 825, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 290, + 285, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A. Adcock, V. Reis, M. Singh, Z. Yan, L. van der Maaten, K. Zhang, S. Motwani, J. Guerin, N. Goyal, I. Misra, L. Gustafson, C. Changhan, and P. Goyal. Classy vision. 2019. \nAlexei Baevski and Michael Auli. Adaptive input representations for neural language modeling. arXiv preprint arXiv:1809.10853, 2018. \nYoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. \nJames Bradbury, Stephen Merity, Caiming Xiong, and Richard Socher. Quasi-recurrent neural networks. arXiv preprint arXiv:1611.01576, 2016. \nQingqing Cao, Harsh Trivedi, Aruna Balasubramanian, and Niranjan Balasubramanian. Faster and just as accurate: A simple decomposition for transformer models. \nMiguel A. Carreira-Perpiñán and Yerlan Idelbayev. Model compression as constrained optimization, with application to neural nets. part ii: quantization, 2017. \nDaoyuan Chen, Yaliang Li, Minghui Qiu, Zhen Wang, Bofang Li, Bolin Ding, Hongbo Deng, Jun Huang, Wei Lin, and Jingren Zhou. Adabert: Task-adaptive bert compression with differentiable neural architecture search. arXiv preprint arXiv:2001.04246, 2020. \nJungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. \nMatthieu Courbariaux and Yoshua Bengio. Binarynet: Training deep neural networks with weights and activations constrained to $+ 1$ or -1. CoRR, 2016. \nMatthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. CoRR, 2015. \nZihang Dai, Zhilin Yang, Yiming Yang, William W Cohen, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. \nYann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In Proc. of ICML, 2017. \nMostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Łukasz Kaiser. Universal transformers, 2018. \nJ. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. \nJacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. ", + "bbox": [ + 171, + 306, + 828, + 917 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yinpeng Dong, Renkun Ni, Jianguo Li, Yurong Chen, Hang Su, and Jun Zhu. Stochastic quantization for learning accurate low-bit deep neural networks. International Journal of Computer Vision, 127 (11-12):1629–1642, 2019. ", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019. ", + "bbox": [ + 171, + 155, + 823, + 184 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pp. 1050–1059, 2016. ", + "bbox": [ + 176, + 191, + 823, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bourdev. Compressing deep convolutional networks using vector quantization. arXiv preprint arXiv:1412.6115, 2014. ", + "bbox": [ + 173, + 242, + 823, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Edouard Grave, Armand Joulin, Moustapha Cisse, David Grangier, and Herve Jegou. Efficient softmax approximation for gpus. arXiv, abs/1609.04309, 2016. ", + "bbox": [ + 173, + 280, + 823, + 309 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning with limited numerical precision. In ICML, 2015. ", + "bbox": [ + 174, + 316, + 823, + 347 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In NIPS, pp. 1135–1143, 2015. ", + "bbox": [ + 173, + 354, + 823, + 385 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding. ICLR, 2016. ", + "bbox": [ + 171, + 392, + 823, + 421 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, 2015. ", + "bbox": [ + 173, + 429, + 823, + 458 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. ", + "bbox": [ + 171, + 467, + 823, + 496 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. arXiv e-prints, 2019. ", + "bbox": [ + 174, + 503, + 825, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016. ", + "bbox": [ + 171, + 554, + 823, + 584 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In CVPR, 2018. ", + "bbox": [ + 171, + 592, + 823, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2704–2713, 2018. ", + "bbox": [ + 174, + 630, + 825, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. PAMI, 2011. ", + "bbox": [ + 173, + 694, + 823, + 724 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. arXiv preprint arXiv:1909.10351, 2019. ", + "bbox": [ + 173, + 731, + 826, + 775 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas Mikolov. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651, 2016. ", + "bbox": [ + 174, + 784, + 823, + 813 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. ", + "bbox": [ + 173, + 820, + 823, + 849 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Guillaume Lample and Alexis Conneau. Cross-lingual language model pretraining. arXiv preprint arXiv:1901.07291, 2019. ", + "bbox": [ + 174, + 858, + 823, + 887 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations, 2019. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In NIPS, 1990. ", + "bbox": [ + 173, + 103, + 750, + 119 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. ", + "bbox": [ + 173, + 128, + 823, + 159 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: A non-uniform discretization for neural networks. arXiv preprint arXiv:1909.13144, 2019. ", + "bbox": [ + 171, + 167, + 823, + 198 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. ", + "bbox": [ + 176, + 207, + 823, + 251 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. ", + "bbox": [ + 173, + 261, + 823, + 291 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. ", + "bbox": [ + 174, + 300, + 823, + 330 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Christos Louizos, Max Welling, and Diederik P Kingma. Learning sparse neural networks through l_0 regularization. arXiv preprint arXiv:1712.01312, 2017. ", + "bbox": [ + 171, + 339, + 823, + 369 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. Thinet: A filter level pruning method for deep neural network compression. In ICCV, 2017. ", + "bbox": [ + 173, + 378, + 823, + 409 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet V2: practical guidelines for efficient CNN architecture design. CoRR, 2018. ", + "bbox": [ + 173, + 419, + 823, + 448 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xindian Ma, Peng Zhang, Shuai Zhang, Nan Duan, Yuexian Hou, Dawei Song, and Ming Zhou. A tensorized transformer for language modeling. arXiv preprint arXiv:1906.09777, 2019. ", + "bbox": [ + 173, + 458, + 821, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mark D. McDonnell. Training wide residual networks for deployment using a single bit for each weight, 2018. ", + "bbox": [ + 174, + 497, + 823, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer Sentinel Mixture Models. arXiv, abs/1609.07843, 2016. ", + "bbox": [ + 173, + 537, + 825, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Deepak Mittal, Shweta Bhardwaj, Mitesh M Khapra, and Balaraman Ravindran. Recovering from random pruning: On the plasticity of deep convolutional neural networks. In WACV, 2018. ", + "bbox": [ + 169, + 577, + 823, + 607 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dmitry Molchanov, Arsenii Ashukha, and Dmitry Vetrov. Variational dropout sparsifies deep neural networks. In ICML, 2017. ", + "bbox": [ + 171, + 616, + 823, + 645 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of NAACL-HLT 2019: Demonstrations, 2019. ", + "bbox": [ + 176, + 655, + 823, + 699 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. How to construct deep recurrent neural networks. In Proceedings of the Second International Conference on Learning Representations (ICLR 2014), 2014. ", + "bbox": [ + 173, + 709, + 823, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. ", + "bbox": [ + 174, + 762, + 825, + 805 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. ", + "bbox": [ + 171, + 815, + 823, + 845 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jack W Rae, Anna Potapenko, Siddhant M Jayakumar, and Timothy P Lillicrap. Compressive transformers for long-range sequence modelling. arXiv preprint arXiv:1911.05507, 2019. ", + "bbox": [ + 171, + 854, + 823, + 886 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV, 2016. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Conference on Computer Vision and Pattern Recognition, pp. 4510–4520, 2018. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. 2019a. ", + "bbox": [ + 171, + 155, + 823, + 185 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019b. ", + "bbox": [ + 173, + 194, + 821, + 223 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Highway networks. arXiv preprint arXiv:1505.00387, 2015. ", + "bbox": [ + 173, + 233, + 823, + 262 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pierre Stock, Armand Joulin, Rémi Gribonval, Benjamin Graham, and Hervé Jégou. And the bit goes down: Revisiting the quantization of neural networks. CoRR, abs/1907.05686, 2019. ", + "bbox": [ + 173, + 272, + 823, + 301 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sainbayar Sukhbaatar, Edouard Grave, Piotr Bojanowski, and Armand Joulin. Adaptive attention span in transformers. arXiv preprint arXiv:1905.07799, 2019a. ", + "bbox": [ + 174, + 310, + 821, + 340 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sainbayar Sukhbaatar, Edouard Grave, Guillaume Lample, Herve Jegou, and Armand Joulin. Augmenting self-attention with persistent memory. arXiv preprint arXiv:1907.01470, 2019b. ", + "bbox": [ + 176, + 349, + 823, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. Patient knowledge distillation for bert model compression. EMNLP, 2019. ", + "bbox": [ + 176, + 387, + 823, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. Mobilebert: Task-agnostic compression of bert for resource limited devices. ", + "bbox": [ + 171, + 426, + 823, + 457 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization and momentum in deep learning. In International conference on machine learning, pp. 1139–1147, 2013. ", + "bbox": [ + 176, + 465, + 823, + 508 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks, 2019. ", + "bbox": [ + 171, + 518, + 823, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Iulia Turc, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Well-read students learn better: The impact of student initialization on knowledge distillation. arXiv preprint arXiv:1908.08962, 2019. ", + "bbox": [ + 176, + 556, + 825, + 599 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Vincent Vanhoucke, Andrew Senior, and Mark Z Mao. Improving the speed of neural networks on cpus. 2011. ", + "bbox": [ + 173, + 609, + 823, + 640 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. ", + "bbox": [ + 173, + 648, + 823, + 678 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Li Wan, Matthew Zeiler, Sixin Zhang, Yann Le Cun, and Rob Fergus. Regularization of neural networks using DropConnect. In ICML, 2013. ", + "bbox": [ + 174, + 686, + 825, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. 2019. ICLR. ", + "bbox": [ + 176, + 726, + 825, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Kuan Wang, Zhijian Liu, Yujun Lin andx Ji Lin, and Song Han. HAQ: hardware-aware automated quantization. CoRR, 2018. ", + "bbox": [ + 173, + 779, + 823, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Adina Williams, Nikita Nangia, and Samuel R. Bowman. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of NAACL-HLT, 2018. ", + "bbox": [ + 171, + 816, + 823, + 847 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Felix Wu, Angela Fan, Alexei Baevski, Yann Dauphin, and Michael Auli. Pay less attention with lightweight and dynamic convolutions. In ICLR, 2019. ", + "bbox": [ + 169, + 856, + 823, + 886 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Biao Zhang, Deyi Xiong, and Jinsong Su. Accelerating neural transformer via an average attention network. arXiv preprint arXiv:1805.00631, 2018. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. CoRR, 2017. ", + "bbox": [ + 173, + 103, + 825, + 132 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sanqiang Zhao, Raghav Gupta, Yang Song, and Denny Zhou. Extreme language model compression with optimal subwords and shared projections. arXiv preprint arXiv:1909.11687, 2019. ", + "bbox": [ + 176, + 141, + 823, + 170 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0a442ac3dc3b45fe6eaac79158a639f7f0aa94ef732ecfb3cebc872eb5c320d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingModelCompressionTop-1 Accuracy
Small BlocksStock et al. (2019)19x73.8
Quant-Noise19x74.3
Large BlocksStock et al. (2019)32x68.2
Quant-Noise32x68.8
", + "bbox": [ + 251, + 101, + 745, + 193 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ba35cf41421a36478cb5c76c461ca17166f7ed7e069174cbe097a3f7e7e6cbf8.jpg", + "image_caption": [ + "Table 4: Compression of ResNet-50 with Quant-Noise. We compare to Stock et al. (2019) in both the small and large blocks regime. For fair comparison, we hold the compression rate constant. Quant-Noise provides improved performance in both settings. ", + "Figure 3: Effect of Quantization Parameters. We report the influence of the proportion of blocks to which we apply the noise. We focus on Transformer for Wikitext-103 language modeling. We explore two settings: iPQ and int8. For iPQ, we use $\\varphi _ { \\mathrm { p r o x y } }$ . " + ], + "image_footnote": [], + "bbox": [ + 181, + 265, + 812, + 390 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 469, + 294, + 486 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7.1 QUANTIZATION OF ADDITIONAL ARCHITECTURES ", + "text_level": 1, + "bbox": [ + 174, + 502, + 563, + 516 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ResNet-50. We explore the compression of ResNet-50, a standard architecture used Computer Vision. In Table 4, we compare Quant-Noise to iPQ Compression from Stock et al. (2019) and show that Quant-Noise provide consistent additional improvement. ", + "bbox": [ + 174, + 527, + 825, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7.2 ABLATIONS ", + "text_level": 1, + "bbox": [ + 176, + 588, + 297, + 602 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this section, we examine the impact of the level of noise during training as well as the impact of approximating iPQ during training. ", + "bbox": [ + 173, + 613, + 825, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7.3 IMPACT OF NOISE RATE ", + "text_level": 1, + "bbox": [ + 176, + 659, + 382, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We analyze the performance for various values of Quant-Noise in Figure 3 on a Transformer for language modeling. For iPQ, performance is impacted by high rates of quantization noise. For example, a Transformer with the noise function $\\varphi _ { \\mathrm { p r o x y } }$ degrades with rate higher than 0.5, i.e., when half of the weights are passed through the noise function $\\varphi _ { \\mathrm { p r o x y } }$ . We hypothesize that for large quantities of noise, a larger effect of using proxy rather than the exact PQ noise is observed. For int8 quantization and its noise function, higher rates of noise are slightly worse but not as severe. A rate of 1 for int8 quantization is equivalent to the Quantization Aware Training of (Krishnamoorthi, 2018), as the full matrix is quantized with STE, showing the potential benefit of partial quantization during training. ", + "bbox": [ + 174, + 685, + 825, + 810 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7.4 IMPACT OF APPROXIMATING THE NOISE FUNCTION", + "text_level": 1, + "bbox": [ + 174, + 828, + 571, + 842 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We study the impact of approximating quantization noise during training. We focus on the case of iPQ with the approximation described in Section 4.2. In Table 5, we compare the correct noise function for iPQ with its approximation $\\varphi _ { \\mathrm { p r o x y } }$ . This approximate noise function does not consider cluster assignments or centroid values and simply zeroes out the selected blocks. For completeness, we include an intermediate approximation where we consider cluster assignments to apply noise within each cluster, but still zero-out the vectors. These approximations do not affect the performance of the quantized models. This suggests that increasing the correlation between subvectors that are jointly clustered is enough to maintain the performance of a model quantized with iPQ. Since PQ tends to work well on highly correlated vectors, such as activations in convolutional networks, this is not surprising. Using the approximation $\\varphi _ { \\mathrm { p r o x y } }$ presents the advantage of speed and practicality. Indeed, one does not need to compute cluster assignments and centroids for every layer in the network after each epoch. Moreover, the approach $\\varphi _ { \\mathrm { p r o x y } }$ is less involved in terms of code. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/8e7a95ca5b6cecb24a3f95f96e0a12b874baa5b5aaf2cd47aa7818f7128b38c8.jpg", + "table_caption": [ + "Table 5: Exact versus proxy noise function for different block selections with iPQ. We compare exact $\\phi _ { \\mathrm { P Q } }$ and the approximation $\\phi _ { \\mathrm { p r o x y } }$ with blocks selected from all subvectors or subvectors from the same cluster. " + ], + "table_footnote": [], + "table_body": "
NoiseBlocksPPLQuant PPL
PQSubvectors18.321.1
PQClusters18.321.2
proxySubvectors18.321.0
proxyClusters18.421.1
", + "bbox": [ + 173, + 101, + 475, + 193 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 219, + 825, + 318 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "7.5 EXPERIMENTAL SETTING ", + "text_level": 1, + "bbox": [ + 176, + 335, + 392, + 349 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We assess the effectiveness of Quant-Noise on competitive language and vision benchmarks. We consider Transformers for language modeling, RoBERTa for pre-training sentence representations, and EfficientNet for image classification. Our models are implemented in PyTorch (Paszke et al., 2017). We use fairseq (Ott et al., 2019) for language modeling and pre-training for sentence representation tasks and Classy Vision (Adcock et al., 2019) for EfficientNet. ", + "bbox": [ + 174, + 363, + 825, + 433 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Language Modeling. We experiment on the Wikitext-103 benchmark (Merity et al., 2016) that contains 100M tokens and a vocabulary of 260k words. We train a 16 layer Transformer following Baevski & Auli (2018) with a LayerDrop rate of 0.2 (Fan et al., 2019). We report perplexity (PPL) on the test set. ", + "bbox": [ + 173, + 449, + 825, + 505 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Pre-Training of Sentence Representations. We pre-train the base BERT model (Devlin et al., 2018) on the BooksCorpus $^ +$ Wiki dataset with a LayerDrop rate of 0.2. We finetune the pre-trained models on the MNLI task (Williams et al., 2018) from the GLUE Benchmark (Wang et al., 2019) and report accuracy. We follow the parameters in Liu et al. (2019) training and finetuning. ", + "bbox": [ + 174, + 516, + 825, + 571 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Image Classification. We train an EfficientNet-B3 model (Tan & Le, 2019) on the ImageNet object classification benchmark (Deng et al., 2009). The EfficientNet-B3 of Classy Vision achieves a Top-1 accuracy of $8 1 . 5 \\%$ , which is slightly below than the performance of $8 1 . 9 \\%$ reported by Tan & Le (2019). ", + "bbox": [ + 174, + 582, + 825, + 638 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "7.6 TRAINING DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 657, + 351, + 671 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Language Modeling To handle the large vocabulary of Wikitext-103, we follow (Dauphin et al., 2017) and (Baevski & Auli, 2018) in using adaptive softmax (Grave et al., 2016) and adaptive input for computational efficiency. For both input and output embeddings, we use dimension size 1024 and three adaptive bands: 20K, 40K, and 200K. We use a cosine learning rate schedule (Baevski & Auli, 2018; Loshchilov & Hutter, 2016) and train with Nesterov’s accelerated gradient (Sutskever et al., 2013). We set the momentum to 0.99 and renormalize gradients if the norm exceeds 0.1 (Pascanu et al., 2014). During training, we partition the data into blocks of contiguous tokens that ignore document boundaries. At test time, we respect sentence boundaries. We set LayerDrop to 0.2. We set Quant-Noise value to 0.05. During training time, we searched over the parameters (0.05, 0.1, 0.2) to determine the optimal value of Quant-Noise. During training time, the block size of Quant-Noise is 8. ", + "bbox": [ + 173, + 684, + 825, + 824 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "RoBERTa The base architecture is a 12 layer model with embedding size 768 and FFN size 3072. We follow (Liu et al., 2019) in using the subword tokenization scheme from (Radford et al., 2019), which uses bytes as subword units. This eliminates unknown tokens. We train with large batches of size 8192 and maintain this batch size using gradient accumulation. We do not use next sentence prediction (Lample & Conneau, 2019). We optimize with Adam with a polynomial decay learning rate schedule. We set LayerDrop to 0.2. We set Quant-Noise value to 0.1. We did not hyperparameter ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/a38db0da2911b409673013500e1d3b6aaa867fb375b658d04d1fe70d247a44b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMBPPL
Trans XL Large (Dai et al., 2019) Compressive Trans (Rae et al.,2019) GCNN (Dauphin et al., 2017) 4 Layer QRNN (Bradbury et al., 2016)970 970 87018.3 17.1 37.2 33.0
Trans XL Base (Dai et al.,2019) Persis Mem (Sukhbaatar etal.,2019b) Tensorized core-2 (Ma et al.,2019)570 506 32524.0 20.6
18.9
Quant-Noise3820.7
Quant-Noise + Share + Prune
10
24.2
", + "bbox": [ + 326, + 101, + 668, + 250 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Table 6: Performance on Wikitext-103. We report test set perplexity and model size in megabytes. \nLower perplexity is better. ", + "bbox": [ + 173, + 267, + 825, + 295 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "search to determine the optimal value of Quant-Noise as training RoBERTa is computationally intensive. During training time, the block size of Quant-Noise is 8. ", + "bbox": [ + 176, + 321, + 823, + 351 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "During finetuning, we hyperparameter search over three learning rate options (1e-5, 2e-5, 3e-5) and batchsize (16 or 32 sentences). The other parameters are set following (Liu et al., 2019). We do single task finetuning, meaning we only tune on the data provided for the given natural language understanding task. We do not perform ensembling. When finetuning models trained with LayerDrop, we apply LayerDrop and Quant-Noise during finetuning time as well. ", + "bbox": [ + 174, + 358, + 825, + 428 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "EfficientNet We use the architecture of EfficientNet-B3 defined in Classy Vision (Adcock et al., 2019) and follow the default hyperparameters for training. We set Quant-Noise value to 0.1. During training time, we searched over the parameters (0.05, 0.1, 0.2) to determine the optimal value of Quant-Noise. During training time, the block size of Quant-Noise is set to 4 for all $1 \\times 1$ convolutions, 9 for depth-wise $3 \\times 3$ convolutions, 5 for depth-wise $5 \\times 5$ convolutions and 4 for the classifier. For sharing, we shared weights between blocks 9-10, 11-12, 14-15, 16-17, 19-20-21, 22-23 and refer to blocks that share the same weights as a chunk. For LayerDrop, we drop the chunks of blocks defined previously with probability 0.2 and evaluate only with chunks 9-10, 14-15 and 19-20-21. ", + "bbox": [ + 173, + 443, + 825, + 568 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "7.7 SCALAR QUANTIZATION DETAILS ", + "text_level": 1, + "bbox": [ + 178, + 585, + 449, + 599 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We closely follow the methodology of PyTorch 1.4. We emulate scalar quantization by quantizing the weights and the activations. The scales and zero points of activations are determined by doing a few forward passes ahead of the evaluation and then fixed. We use the Histogram method to compute $s$ and $z$ , which aims at approximately minimizing the $L _ { 2 }$ quantization error by adjusting $s$ and $z$ . This scheme is a refinement of the MinMax scheme. Per channel quantization is also discussed in Table 10. ", + "bbox": [ + 174, + 611, + 825, + 694 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "7.8 IPQ QUANTIZATION DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 712, + 419, + 727 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Language Modeling We quantize FFN with block size 8, embeddings with block size 8, and attention with block size 4. We tuned the block size for attention between the values (4, 8) to find the best performance. Note that during training with apply Quant-Noise to all the layers. ", + "bbox": [ + 174, + 738, + 825, + 781 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "RoBERTa We quantize FFN with block size 4, embeddings with block size 4, and attention with block size 4. We tuned the block size between the values (4, 8) to find the best performance. Note that during training with apply Quant-Noise to all the layers. ", + "bbox": [ + 174, + 796, + 825, + 838 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "EfficientNet We quantize blocks sequentially and end up with the classifier. The block sizes are 4 for all $1 \\times 1$ convolutions, 9 for depth-wise $3 \\times 3$ convolutions, 5 for depth-wise $5 \\times 5$ convolutions and 4 for the classifier. Note that during training with apply Quant-Noise to all the weights in InvertedResidual Blocks (except the Squeeze-Excitation subblocks), the head convolution and the classifier. ", + "bbox": [ + 176, + 854, + 823, + 922 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/646a4a280b70108d73b471db8a5e1cfb1f443f9da04faba4e93f86c67e85c7f9.jpg", + "table_caption": [ + "Table 7: Performance on MNLI. We report accuracy and size in megabytes. \\* indicates distillation using BERT Large. $\\dagger$ indicates training with data augmentation. Work from Sun et al. (2019) and Zhao et al. (2019) do not report results on the dev set. Cao et al. do not report model size. Higher accuracy is better. " + ], + "table_footnote": [], + "table_body": "
ModelMBMNLI
RoBERTa Base +LD (Fan et al., 2019)48084.8
BERT Base (Devlin et al.,2018)42084.4
PreTrained Distil (Turc et al., 2019)25782.5
DistilBERT (Sanh et al.,2019b)25081.8
MobileBERT* (Sun et al.)9684.4
TinyBERTt (Jiao et al.,2019)5582.8
ALBERT Base (Lan et al.,2019)4581.6
AdaBERTt (Chen et al.,2020)3681.6
Quant-Noise3883.6
Quant-Noise+ Share +Prune1482.5
", + "bbox": [ + 318, + 101, + 676, + 262 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/2015213e70459a921a8d164d59e3c6ae0b4e29ed3dbe57eb40889aa85bcec1b0.jpg", + "table_caption": [ + "Table 8: Performance on ImageNet. We report accuracy and size in megabytes. Higher accuracy is better. " + ], + "table_footnote": [], + "table_body": "
ModelMBAcc.
EfficientNet-B7 (Tan &Le,2019) ResNet-50 (He et al., 2015) DenseNet-169 (Huang et al., 2018) EfficientNet-BO (Tan &Le,2019) MobileNet-v2 (Sandler et al., 2018)260 97.5 53.4 20.2 13.484.4 76.1 76.2 77.3 71.9
Shufflenet-v2 ×1 (Ma et al.,2018) HAQ 4 bits (Wang et al., 2018) iPQ ResNet-50 (Stock et al.,2019)8.7 12.469.4 76.2
Quant-Noise5.09 3.376.1 80.0
Quant-Noise + Share + Prune2.377.8
", + "bbox": [ + 333, + 349, + 660, + 517 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "7.9 DETAILS OF PRUNING AND LAYER SHARING ", + "text_level": 1, + "bbox": [ + 174, + 590, + 524, + 604 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We apply the Every Other Layer strategy from Fan et al. (2019). When combining layer sharing with pruning, we train models with shared layers and then prune chunks of shared layers. When sharing layers, the weights of adjacent layers are shared in chunks of two. For a concrete example, imagine we have a model with layers A, B, C, D, E, F, G, H. We share layers A and B, C and D, E and F, G and H. To prune, every other chunk would be pruned away, for example we could prune A, B, E, F. ", + "bbox": [ + 173, + 616, + 825, + 686 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "7.10 NUMERICAL RESULTS FOR GRAPHICAL DIAGRAMS ", + "text_level": 1, + "bbox": [ + 174, + 704, + 581, + 719 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We report the numerical values displayed in Figures 2 in Table 6 for language modeling, Table 7 for BERT, and Table 8 for ImageNet. ", + "bbox": [ + 173, + 729, + 823, + 758 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "7.11 FURTHER ABLATIONS ", + "text_level": 1, + "bbox": [ + 176, + 776, + 377, + 791 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "7.11.1 IMPACT OF QUANT-NOISE FOR THE VISION SETUP ", + "text_level": 1, + "bbox": [ + 174, + 801, + 586, + 816 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We provide another study showing the impact of the proportion of elements on which to apply Quant-Noise in Table 9. ", + "bbox": [ + 173, + 827, + 823, + 854 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "7.11.2 IMPACT OF THE NUMBER OF CENTROIDS ", + "text_level": 1, + "bbox": [ + 174, + 871, + 516, + 886 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We quantize with 256 centroids which represents a balance between size and representation capacity. The effect of the number of centroids on performance and size is shown in Figure 4 (a). Quantizing ", + "bbox": [ + 174, + 895, + 825, + 924 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/6625f75b2c7be88a96a61fe5e571cb9c50197ba7200f19ce956b5e5446ce47e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
p00.20.40.60.81
Top-180.6680.8380.8280.8880.9280.64
", + "bbox": [ + 310, + 102, + 683, + 142 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Table 9: Effect of Quantization Parameters. We report the influence of the Quant-Noise rate $p$ with Scalar Quantization (int8). We focus on EfficientNet for ImageNet classification. ", + "bbox": [ + 173, + 160, + 826, + 189 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/cf825d2cf66ad807058d5706153a322d0dca0137253c54a0e00ac009d8f56a9a.jpg", + "image_caption": [ + "Figure 4: Quantizing with a larger number of centroids. Results are shown on Wikitext-103 valid. " + ], + "image_footnote": [], + "bbox": [ + 408, + 209, + 589, + 319 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "with more centroids improves perplexity — this parameter could be adjusted based on the practical storage constraints. ", + "bbox": [ + 173, + 378, + 823, + 407 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "7.11.3 EFFECT OF INITIAL MODEL SIZE ", + "text_level": 1, + "bbox": [ + 176, + 422, + 465, + 438 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Large, overparameterized models are more easily compressed. In Figure 5, we explore quantizing both shallower and skinnier models. For shallow models, the gap between quantized and non-quantized perplexity does not increase as layers are removed (Figure 5, left). In contrast, there is a larger gap in performance for models with smaller FFN (Figure 5, right). As the FFN size decreases, the weights are less redundant and more difficult to quantize with iPQ. ", + "bbox": [ + 174, + 448, + 825, + 517 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "7.11.4 DIFFICULTY OF QUANTIZING DIFFERENT MODEL STRUCTURES ", + "text_level": 1, + "bbox": [ + 176, + 534, + 674, + 547 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Quantization is applied to various portions of the Transformer architecture — the embedding, attention, feedforward, and classifier output. We compare the quantizability of various portions of the network in this section. ", + "bbox": [ + 174, + 558, + 825, + 599 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Is the order of structures important? We quantize specific network structures first — this is important as quantizing weight matrices can accumulate reconstruction error. Some structures of the network should be quantized last so the finetuning process can better adjust the centroids. We find that there are small variations in performance based on quantization order (see Figure 6). We choose to quantize FFN, then embeddings, and finally the attention matrices in Transformer networks. ", + "bbox": [ + 174, + 616, + 825, + 686 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Which structures can be compressed the most? Finally, we analyze which network structures can be most compressed. During quantization, various matrix block sizes can be chosen as a parameter — the larger the block size, the more compression, but also the larger the potential reduction of performance. Thus, it is important to understand how much each network structure can be compressed to reduce the memory footprint of the final model as much as possible. In Figure 6, we quantize two model structures with a fixed block size and vary the block size of the third between 4 and 32. As shown, the FFN and embedding structures are more robust to aggressive compression, while the attention drastically loses performance as larger block sizes are used. ", + "bbox": [ + 174, + 702, + 825, + 814 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "7.11.5 APPROACH TO I N TN SCALAR QUANTIZATION", + "text_level": 1, + "bbox": [ + 174, + 829, + 555, + 844 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "We compare quantizing per-channel to using a histogram quantizer in Table 10. The histogram quantizer maintains a running min/max and minimizes L2 distance between quantized and nonquantized values to find the optimal min/max. Quantizing per channel learns scales and offsets as vectors along the channel dimension, which provides more flexibility since scales and offsets can be different. ", + "bbox": [ + 174, + 853, + 825, + 922 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/89d51efee35f6dc30bb5ffaa4bb5b370bcc344263f7bf64ce5139ec44a66a0da.jpg", + "image_caption": [ + "Figure 5: (a) Effect of Initial Model Size for more shallow models (b) Effect of Initial Model Size more skinny models " + ], + "image_footnote": [], + "bbox": [ + 281, + 106, + 717, + 267 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/8012673edf3e43ed6c0eaae7f98a58780eb2aa1fbaa3915613839ed665d88982.jpg", + "image_caption": [ + "Figure 6: Effect of Quantization on Model Structures. Results are shown on the validation set of Wikitext-103. (a) Quantizing Attention, FFN, and Embeddings in different order. (b) More Extreme compression of different structures. " + ], + "image_footnote": [], + "bbox": [ + 279, + 335, + 717, + 484 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "7.11.6 LAYERDROP WITH STE ", + "text_level": 1, + "bbox": [ + 176, + 571, + 401, + 587 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "For quantization noise, we apply the straight through estimator (STE) to remaining weights in the backward pass. We experiment with applying STE to the backward pass of LayerDrop’s pruning noise. Results are shown in Table 11 and find slightly worse results. ", + "bbox": [ + 174, + 597, + 825, + 638 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/a0b27ff2b8235f8e3d7baa72f239c15d78108896149885e7378f8fe820ae675a.jpg", + "table_caption": [ + "Table 10: Comparison of different approaches to int4 and int8 with and without QuantNoise on language modeling and image classification. For language modeling, we train a Transformer on the Wikitext-103 benchmark. We report perplexity (PPL) on the test set. For image classification, we train a EfficientNet-B3 on the ImageNet-1K benchmark. We report top-1 accuracy on the validation set. For both setting, we also report model size in megabyte (MB) and the compression ratio compared to the original model. " + ], + "table_footnote": [], + "table_body": "
Quantization SchemeLanguage Modeling 16-layer Transformer Wikitext-103Image Classification EfficientNet-B3 ImageNet-1K
SizeCompressTest PPLSizeCompressTop-1 Acc.
Uncompressed model942×118.346.7×181.5
Int4 Quant Histogram118×839.45.8×845.3
+ Quant-Noise118×821.85.8×867.8
Int4 Quant Channel118×821.25.8×868.2
+ Quant-Noise118×819.55.8×872.3
Int8 Quant Histogram236×419.611.7×480.7
+ Quant-Noise236×418.711.7×480.9
Int8 Quant Channel236×418.511.7×481.1
+ Quant-Noise236×418.311.7×481.2
", + "bbox": [ + 194, + 202, + 803, + 414 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/b2546aff33fa9ce235993179cb151ccff04c5f22d2fca32774da112a5e3a2f88.jpg", + "table_caption": [ + "Table 11: Performance on Wikitext-103 when using STE in the backward pass of the LayerDrop pruning noise. " + ], + "table_footnote": [], + "table_body": "
ModelMBPPL
Quant-Noise + Share + Prune1024.2
Quant-Noise + Share + Prune with STE1024.5
", + "bbox": [ + 321, + 719, + 671, + 773 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/parse/train/dV19Yyi1fS3/dV19Yyi1fS3_model.json b/parse/train/dV19Yyi1fS3/dV19Yyi1fS3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ee17e1690e7a276c12f8b3cb5ad116abc0fd7c27 --- /dev/null +++ b/parse/train/dV19Yyi1fS3/dV19Yyi1fS3_model.json @@ -0,0 +1,22825 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 397, + 701, + 1303, + 701, + 1303, + 1190, + 397, + 1190 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 299, + 1332, + 1404, + 1332, + 1404, + 1516, + 299, + 1516 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1532, + 1404, + 1532, + 1404, + 1747, + 298, + 1747 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1762, + 1403, + 1762, + 1403, + 1885, + 297, + 1885 + ], + "score": 0.972 + }, + { + "category_id": 0, + "poly": [ + 298, + 220, + 1198, + 220, + 1198, + 324, + 298, + 324 + ], + "score": 0.963 + }, + { + "category_id": 2, + "poly": [ + 298, + 1916, + 1407, + 1916, + 1407, + 2034, + 298, + 2034 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 1129, + 377, + 1385, + 377, + 1385, + 438, + 1129, + 438 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 1125, + 486, + 1382, + 486, + 1382, + 545, + 1125, + 545 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 315, + 375, + 637, + 375, + 637, + 438, + 315, + 438 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 707, + 376, + 1058, + 376, + 1058, + 438, + 707, + 438 + ], + "score": 0.921 + }, + { + "category_id": 0, + "poly": [ + 302, + 1259, + 573, + 1259, + 573, + 1294, + 302, + 1294 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.877 + }, + { + "category_id": 0, + "poly": [ + 773, + 628, + 927, + 628, + 927, + 660, + 773, + 660 + ], + "score": 0.876 + }, + { + "category_id": 1, + "poly": [ + 835, + 484, + 1091, + 484, + 1091, + 545, + 835, + 545 + ], + "score": 0.861 + }, + { + "category_id": 1, + "poly": [ + 314, + 485, + 569, + 485, + 569, + 545, + 314, + 545 + ], + "score": 0.819 + }, + { + "category_id": 1, + "poly": [ + 604, + 484, + 806, + 484, + 806, + 544, + 604, + 544 + ], + "score": 0.773 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 856, + 2088, + 856, + 2112, + 840, + 2112 + ], + "score": 0.739 + }, + { + "category_id": 13, + "poly": [ + 1014, + 1098, + 1088, + 1098, + 1088, + 1127, + 1014, + 1127 + ], + "score": 0.87, + "latex": "8 2 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 864, + 1129, + 940, + 1129, + 940, + 1157, + 864, + 1157 + ], + "score": 0.86, + "latex": "8 0 . 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 733, + 1128, + 812, + 1128, + 812, + 1158, + 733, + 1158 + ], + "score": 0.4, + "latex": "1 4 \\mathbf { M B }" + }, + { + "category_id": 13, + "poly": [ + 788, + 1159, + 873, + 1159, + 873, + 1188, + 788, + 1188 + ], + "score": 0.39, + "latex": "3 . 3 \\mathrm { M B }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 218.0, + 1202.0, + 218.0, + 1202.0, + 274.0, + 293.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 274.0, + 989.0, + 274.0, + 989.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1914.0, + 1238.0, + 1914.0, + 1238.0, + 1949.0, + 330.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1938.0, + 1291.0, + 1938.0, + 1291.0, + 1986.0, + 329.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1968.0, + 1408.0, + 1968.0, + 1408.0, + 2016.0, + 327.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2006.0, + 468.0, + 2006.0, + 468.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1256.0, + 579.0, + 1256.0, + 579.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 626.0, + 932.0, + 626.0, + 932.0, + 664.0, + 770.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 859.0, + 2087.0, + 859.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 700.0, + 1305.0, + 700.0, + 1305.0, + 738.0, + 395.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 733.0, + 1305.0, + 733.0, + 1305.0, + 768.0, + 393.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 758.0, + 1306.0, + 758.0, + 1306.0, + 801.0, + 391.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 792.0, + 1307.0, + 792.0, + 1307.0, + 830.0, + 393.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 823.0, + 1308.0, + 823.0, + 1308.0, + 858.0, + 393.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 856.0, + 1304.0, + 856.0, + 1304.0, + 889.0, + 393.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 883.0, + 1306.0, + 883.0, + 1306.0, + 919.0, + 392.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 915.0, + 1305.0, + 915.0, + 1305.0, + 952.0, + 393.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 945.0, + 1306.0, + 945.0, + 1306.0, + 979.0, + 393.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 974.0, + 1306.0, + 974.0, + 1306.0, + 1011.0, + 393.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1007.0, + 1307.0, + 1007.0, + 1307.0, + 1042.0, + 393.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1033.0, + 1306.0, + 1033.0, + 1306.0, + 1074.0, + 391.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1068.0, + 1306.0, + 1068.0, + 1306.0, + 1100.0, + 394.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1097.0, + 1013.0, + 1097.0, + 1013.0, + 1131.0, + 394.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1097.0, + 1306.0, + 1097.0, + 1306.0, + 1131.0, + 1089.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1128.0, + 732.0, + 1128.0, + 732.0, + 1166.0, + 393.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1128.0, + 863.0, + 1128.0, + 863.0, + 1166.0, + 813.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1128.0, + 1306.0, + 1128.0, + 1306.0, + 1166.0, + 941.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1160.0, + 787.0, + 1160.0, + 787.0, + 1190.0, + 393.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1160.0, + 902.0, + 1160.0, + 902.0, + 1190.0, + 874.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 1404.0, + 1330.0, + 1404.0, + 1370.0, + 294.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1363.0, + 1409.0, + 1363.0, + 1409.0, + 1399.0, + 294.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1394.0, + 1405.0, + 1394.0, + 1405.0, + 1428.0, + 294.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1424.0, + 1404.0, + 1424.0, + 1404.0, + 1460.0, + 295.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1490.0, + 292.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1485.0, + 667.0, + 1485.0, + 667.0, + 1518.0, + 296.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1527.0, + 1405.0, + 1527.0, + 1405.0, + 1571.0, + 291.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1562.0, + 1405.0, + 1562.0, + 1405.0, + 1599.0, + 293.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1594.0, + 1404.0, + 1594.0, + 1404.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1624.0, + 1404.0, + 1624.0, + 1404.0, + 1659.0, + 292.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1653.0, + 1406.0, + 1653.0, + 1406.0, + 1689.0, + 292.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1684.0, + 1406.0, + 1684.0, + 1406.0, + 1722.0, + 293.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 661.0, + 1713.0, + 661.0, + 1753.0, + 294.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1761.0, + 1405.0, + 1761.0, + 1405.0, + 1798.0, + 294.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1793.0, + 1403.0, + 1793.0, + 1403.0, + 1826.0, + 293.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1824.0, + 1405.0, + 1824.0, + 1405.0, + 1860.0, + 293.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1856.0, + 1404.0, + 1856.0, + 1404.0, + 1889.0, + 295.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 376.0, + 1357.0, + 376.0, + 1357.0, + 409.0, + 1129.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 407.0, + 1386.0, + 407.0, + 1386.0, + 440.0, + 1128.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 482.0, + 1313.0, + 482.0, + 1313.0, + 519.0, + 1123.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 516.0, + 1383.0, + 516.0, + 1383.0, + 547.0, + 1124.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 374.0, + 500.0, + 374.0, + 500.0, + 405.0, + 309.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 404.0, + 638.0, + 404.0, + 638.0, + 439.0, + 311.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 374.0, + 858.0, + 374.0, + 858.0, + 409.0, + 703.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 405.0, + 1061.0, + 405.0, + 1061.0, + 439.0, + 702.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 481.0, + 993.0, + 481.0, + 993.0, + 521.0, + 834.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 514.0, + 1094.0, + 514.0, + 1094.0, + 547.0, + 835.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 482.0, + 506.0, + 482.0, + 506.0, + 518.0, + 312.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 513.0, + 572.0, + 513.0, + 572.0, + 547.0, + 312.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 484.0, + 808.0, + 484.0, + 808.0, + 515.0, + 600.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 514.0, + 665.0, + 514.0, + 665.0, + 547.0, + 599.0, + 547.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 702, + 1404, + 702, + 1404, + 977, + 297, + 977 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 991, + 1405, + 991, + 1405, + 1329, + 297, + 1329 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1403, + 1819, + 1403, + 2034, + 297, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 315, + 223, + 1396, + 223, + 1396, + 494, + 315, + 494 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 300, + 623, + 1398, + 623, + 1398, + 687, + 300, + 687 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 368, + 1401, + 1404, + 1401, + 1404, + 1699, + 368, + 1699 + ], + "score": 0.941 + }, + { + "category_id": 4, + "poly": [ + 298, + 515, + 1405, + 515, + 1405, + 608, + 298, + 608 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 300, + 1751, + 588, + 1751, + 588, + 1787, + 300, + 1787 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 298, + 1344, + 947, + 1344, + 947, + 1376, + 298, + 1376 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.714 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.122 + }, + { + "category_id": 13, + "poly": [ + 1223, + 1503, + 1298, + 1503, + 1298, + 1532, + 1223, + 1532 + ], + "score": 0.86, + "latex": "8 2 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 397, + 1564, + 472, + 1564, + 472, + 1594, + 397, + 1594 + ], + "score": 0.84, + "latex": "8 0 . 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 396, + 1666, + 471, + 1666, + 471, + 1695, + 396, + 1695 + ], + "score": 0.82, + "latex": "7 9 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1221, + 1564, + 1307, + 1564, + 1307, + 1594, + 1221, + 1594 + ], + "score": 0.45, + "latex": "3 . 3 { \\mathrm { M B } }" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 232.0, + 413.0, + 232.0, + 413.0, + 259.0, + 365.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 234.0, + 580.0, + 234.0, + 580.0, + 262.0, + 416.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 233.0, + 961.0, + 233.0, + 961.0, + 268.0, + 656.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 235.0, + 1333.0, + 235.0, + 1333.0, + 266.0, + 1056.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 267.0, + 803.0, + 267.0, + 803.0, + 296.0, + 692.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 270.0, + 977.0, + 270.0, + 977.0, + 294.0, + 875.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 262.0, + 1195.0, + 262.0, + 1195.0, + 294.0, + 1081.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 266.0, + 1367.0, + 266.0, + 1367.0, + 292.0, + 1266.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 282.0, + 354.0, + 282.0, + 354.0, + 421.0, + 322.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 281.0, + 394.0, + 281.0, + 394.0, + 307.0, + 365.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 286.0, + 806.0, + 286.0, + 806.0, + 315.0, + 693.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 286.0, + 981.0, + 286.0, + 981.0, + 319.0, + 866.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 286.0, + 1194.0, + 286.0, + 1194.0, + 319.0, + 1080.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 287.0, + 1375.0, + 287.0, + 1375.0, + 320.0, + 1259.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 326.0, + 394.0, + 326.0, + 394.0, + 352.0, + 365.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 352.0, + 582.0, + 352.0, + 582.0, + 379.0, + 425.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 335.0, + 1177.0, + 335.0, + 1177.0, + 379.0, + 1168.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 361.0, + 1213.0, + 361.0, + 1213.0, + 374.0, + 1205.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 369.0, + 395.0, + 369.0, + 395.0, + 398.0, + 362.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 415.0, + 394.0, + 415.0, + 394.0, + 440.0, + 366.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 397.0, + 598.0, + 397.0, + 598.0, + 431.0, + 402.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 431.0, + 422.0, + 431.0, + 422.0, + 469.0, + 395.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 454.0, + 608.0, + 454.0, + 608.0, + 467.0, + 597.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 468.0, + 384.0, + 468.0, + 384.0, + 494.0, + 309.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 469.0, + 432.0, + 469.0, + 432.0, + 494.0, + 404.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 467.0, + 502.0, + 467.0, + 502.0, + 495.0, + 462.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 467.0, + 625.0, + 467.0, + 625.0, + 494.0, + 580.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 516.0, + 1403.0, + 516.0, + 1403.0, + 549.0, + 296.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 546.0, + 1405.0, + 546.0, + 1405.0, + 580.0, + 295.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 577.0, + 1061.0, + 577.0, + 1061.0, + 611.0, + 296.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1747.0, + 594.0, + 1747.0, + 594.0, + 1793.0, + 290.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 701.0, + 1405.0, + 701.0, + 1405.0, + 737.0, + 294.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 763.0, + 1405.0, + 763.0, + 1405.0, + 799.0, + 294.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 787.0, + 1405.0, + 787.0, + 1405.0, + 833.0, + 291.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 823.0, + 1405.0, + 823.0, + 1405.0, + 862.0, + 292.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 854.0, + 1404.0, + 854.0, + 1404.0, + 890.0, + 295.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 882.0, + 1406.0, + 882.0, + 1406.0, + 922.0, + 292.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 915.0, + 1405.0, + 915.0, + 1405.0, + 951.0, + 295.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 944.0, + 1392.0, + 944.0, + 1392.0, + 984.0, + 294.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 988.0, + 1406.0, + 988.0, + 1406.0, + 1031.0, + 292.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1024.0, + 1406.0, + 1024.0, + 1406.0, + 1059.0, + 295.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1049.0, + 1407.0, + 1049.0, + 1407.0, + 1092.0, + 292.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1085.0, + 1407.0, + 1085.0, + 1407.0, + 1120.0, + 294.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1114.0, + 1406.0, + 1114.0, + 1406.0, + 1149.0, + 292.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1407.0, + 1146.0, + 1407.0, + 1180.0, + 294.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1175.0, + 1405.0, + 1175.0, + 1405.0, + 1210.0, + 294.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1206.0, + 1406.0, + 1206.0, + 1406.0, + 1241.0, + 294.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1235.0, + 1405.0, + 1235.0, + 1405.0, + 1272.0, + 292.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1266.0, + 1408.0, + 1266.0, + 1408.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1298.0, + 916.0, + 1298.0, + 916.0, + 1333.0, + 295.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1848.0, + 1407.0, + 1848.0, + 1407.0, + 1888.0, + 292.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1915.0, + 295.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1906.0, + 1407.0, + 1906.0, + 1407.0, + 1952.0, + 292.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1407.0, + 1941.0, + 1407.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2006.0, + 293.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2037.0, + 292.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 622.0, + 1403.0, + 622.0, + 1403.0, + 660.0, + 294.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 656.0, + 667.0, + 656.0, + 667.0, + 688.0, + 295.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1399.0, + 1404.0, + 1399.0, + 1404.0, + 1437.0, + 372.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1431.0, + 1014.0, + 1431.0, + 1014.0, + 1465.0, + 395.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1472.0, + 1405.0, + 1472.0, + 1405.0, + 1510.0, + 375.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1499.0, + 1222.0, + 1499.0, + 1222.0, + 1541.0, + 393.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1299.0, + 1499.0, + 1406.0, + 1499.0, + 1406.0, + 1541.0, + 1299.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1531.0, + 1407.0, + 1531.0, + 1407.0, + 1571.0, + 393.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1561.0, + 396.0, + 1561.0, + 396.0, + 1600.0, + 393.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 1561.0, + 1220.0, + 1561.0, + 1220.0, + 1600.0, + 473.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1561.0, + 1319.0, + 1561.0, + 1319.0, + 1600.0, + 1308.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1603.0, + 1404.0, + 1603.0, + 1404.0, + 1638.0, + 367.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1636.0, + 1405.0, + 1636.0, + 1405.0, + 1671.0, + 395.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1666.0, + 1219.0, + 1666.0, + 1219.0, + 1700.0, + 472.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1341.0, + 951.0, + 1341.0, + 951.0, + 1384.0, + 294.0, + 1384.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 539, + 1406, + 539, + 1406, + 937, + 297, + 937 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1709, + 1404, + 1709, + 1404, + 1895, + 297, + 1895 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1046, + 1405, + 1046, + 1405, + 1140, + 298, + 1140 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1272, + 1404, + 1272, + 1404, + 1433, + 297, + 1433 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 414, + 1404, + 414, + 1404, + 507, + 299, + 507 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 300, + 1601, + 1402, + 1601, + 1402, + 1696, + 300, + 1696 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 684, + 1146, + 1012, + 1146, + 1012, + 1268, + 684, + 1268 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 298, + 1446, + 1398, + 1446, + 1398, + 1510, + 298, + 1510 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 301, + 307, + 1401, + 307, + 1401, + 401, + 301, + 401 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1404, + 229, + 1404, + 293, + 294, + 293 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 544, + 1975, + 1152, + 1975, + 1152, + 2043, + 544, + 2043 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 654, + 1900, + 1047, + 1900, + 1047, + 1936, + 654, + 1936 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 299, + 979, + 843, + 979, + 843, + 1016, + 299, + 1016 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 300, + 1941, + 748, + 1941, + 748, + 1972, + 300, + 1972 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 301, + 1544, + 822, + 1544, + 822, + 1577, + 301, + 1577 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1192, + 1400, + 1192, + 1400, + 1223, + 1366, + 1223 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1903, + 1400, + 1903, + 1400, + 1932, + 1366, + 1932 + ], + "score": 0.867 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.673 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.325 + }, + { + "category_id": 14, + "poly": [ + 683, + 1143, + 1012, + 1143, + 1012, + 1271, + 683, + 1271 + ], + "score": 0.94, + "latex": "\\mathbf { W } = \\left( \\sum _ { \\vdots } ^ { \\mathbf { b } _ { 1 1 } } \\mathrm { ~ ~ \\cdots ~ } \\sum _ { \\mathbf { b } _ { m q } } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 425, + 1108, + 562, + 1108, + 562, + 1137, + 425, + 1137 + ], + "score": 0.93, + "latex": "\\mathbf { W } \\in \\mathbf { R } ^ { n \\times p }" + }, + { + "category_id": 13, + "poly": [ + 1230, + 1395, + 1369, + 1395, + 1369, + 1434, + 1230, + 1434 + ], + "score": 0.93, + "latex": "\\widehat { \\mathbf { b } } _ { k l } = \\mathbf { c } [ \\mathbf { I } _ { k l } ]" + }, + { + "category_id": 13, + "poly": [ + 441, + 1302, + 683, + 1302, + 683, + 1336, + 441, + 1336 + ], + "score": 0.92, + "latex": "{ \\mathcal { C } } = \\{ \\mathbf { c } [ 1 ] , \\dots , \\mathbf { c } [ K ] \\}" + }, + { + "category_id": 13, + "poly": [ + 697, + 1773, + 776, + 1773, + 776, + 1800, + 697, + 1800 + ], + "score": 0.92, + "latex": "N = 1" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1112, + 1094, + 1112, + 1094, + 1141, + 1019, + 1141 + ], + "score": 0.91, + "latex": "m \\times q" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1710, + 1398, + 1710, + 1398, + 1741, + 1266, + 1741 + ], + "score": 0.9, + "latex": "\\mathbf { b } _ { k l } = \\mathbf { W } _ { k l }" + }, + { + "category_id": 13, + "poly": [ + 1056, + 1449, + 1098, + 1449, + 1098, + 1478, + 1056, + 1478 + ], + "score": 0.9, + "latex": "\\mathbf { b } _ { k l }" + }, + { + "category_id": 13, + "poly": [ + 897, + 1833, + 954, + 1833, + 954, + 1863, + 897, + 1863 + ], + "score": 0.89, + "latex": "{ \\bf W } _ { k l }" + }, + { + "category_id": 13, + "poly": [ + 689, + 1365, + 726, + 1365, + 726, + 1395, + 689, + 1395 + ], + "score": 0.89, + "latex": "{ \\bf { I } } _ { k l }" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1110, + 1218, + 1110, + 1218, + 1140, + 1176, + 1140 + ], + "score": 0.89, + "latex": "\\mathbf { b } _ { k l }" + }, + { + "category_id": 14, + "poly": [ + 545, + 1974, + 1150, + 1974, + 1150, + 2046, + 545, + 2046 + ], + "score": 0.89, + "latex": "s = { \\frac { \\operatorname* { m a x } \\mathbf { W } - \\operatorname* { m i n } \\mathbf { W } } { 2 ^ { N } - 1 } } \\quad { \\mathrm { a n d } } \\quad z = \\operatorname { r o u n d } ( \\operatorname* { m i n } \\mathbf { W } / s ) ." + }, + { + "category_id": 13, + "poly": [ + 420, + 1335, + 462, + 1335, + 462, + 1364, + 420, + 1364 + ], + "score": 0.88, + "latex": "\\mathbf { b } _ { k l }" + }, + { + "category_id": 13, + "poly": [ + 538, + 1799, + 575, + 1799, + 575, + 1830, + 538, + 1830 + ], + "score": 0.87, + "latex": "2 ^ { N }" + }, + { + "category_id": 14, + "poly": [ + 651, + 1899, + 1046, + 1899, + 1046, + 1936, + 651, + 1936 + ], + "score": 0.84, + "latex": "\\begin{array} { r } { \\mathbf { c } = ( \\mathrm { r o u n d } ( \\mathbf { W } _ { k l } / s + z ) - z ) \\times s , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 800, + 1394, + 836, + 1394, + 836, + 1429, + 800, + 1429 + ], + "score": 0.83, + "latex": "\\widehat { \\bf W }" + }, + { + "category_id": 13, + "poly": [ + 745, + 1862, + 895, + 1862, + 895, + 1896, + 745, + 1896 + ], + "score": 0.83, + "latex": "( \\mathbf { W } _ { k l } / s + z )" + }, + { + "category_id": 13, + "poly": [ + 1372, + 1274, + 1401, + 1274, + 1401, + 1300, + 1372, + 1300 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 750, + 1742, + 778, + 1742, + 778, + 1768, + 750, + 1768 + ], + "score": 0.79, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1335, + 1335, + 1355, + 1335, + 1355, + 1361, + 1335, + 1361 + ], + "score": 0.79, + "latex": "\\mathcal { C }" + }, + { + "category_id": 13, + "poly": [ + 717, + 1838, + 734, + 1838, + 734, + 1860, + 717, + 1860 + ], + "score": 0.77, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1025, + 1335, + 1044, + 1335, + 1044, + 1361, + 1025, + 1361 + ], + "score": 0.76, + "latex": "\\mathcal { C }" + }, + { + "category_id": 13, + "poly": [ + 454, + 1838, + 472, + 1838, + 472, + 1860, + 454, + 1860 + ], + "score": 0.73, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1351, + 1838, + 1367, + 1838, + 1367, + 1860, + 1351, + 1860 + ], + "score": 0.7, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 621, + 1866, + 677, + 1866, + 677, + 1891, + 621, + 1891 + ], + "score": 0.57, + "latex": "z \\mapsto" + }, + { + "category_id": 13, + "poly": [ + 845, + 1339, + 863, + 1339, + 863, + 1361, + 845, + 1361 + ], + "score": 0.53, + "latex": "\\mathbf { c }" + }, + { + "category_id": 13, + "poly": [ + 940, + 1365, + 958, + 1365, + 958, + 1391, + 940, + 1391 + ], + "score": 0.36, + "latex": "\\mathbf { I }" + }, + { + "category_id": 14, + "poly": [ + 890, + 1991, + 1147, + 1991, + 1147, + 2030, + 890, + 2030 + ], + "score": 0.29, + "latex": "z = \\operatorname { r o u n d } ( \\operatorname* { m i n } \\mathbf { W } / s ) ." + }, + { + "category_id": 15, + "poly": [ + 288.0, + 975.0, + 847.0, + 975.0, + 847.0, + 1023.0, + 288.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1539.0, + 826.0, + 1539.0, + 826.0, + 1584.0, + 292.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 540.0, + 1409.0, + 540.0, + 1409.0, + 576.0, + 297.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 569.0, + 1406.0, + 569.0, + 1406.0, + 608.0, + 292.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 601.0, + 1406.0, + 601.0, + 1406.0, + 637.0, + 294.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 629.0, + 1406.0, + 629.0, + 1406.0, + 667.0, + 292.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 664.0, + 1403.0, + 664.0, + 1403.0, + 696.0, + 296.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 694.0, + 1403.0, + 694.0, + 1403.0, + 726.0, + 295.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 721.0, + 1406.0, + 721.0, + 1406.0, + 760.0, + 292.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 751.0, + 1408.0, + 751.0, + 1408.0, + 791.0, + 292.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 785.0, + 1407.0, + 785.0, + 1407.0, + 821.0, + 295.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 815.0, + 1406.0, + 815.0, + 1406.0, + 850.0, + 294.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 844.0, + 1406.0, + 844.0, + 1406.0, + 881.0, + 292.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 1404.0, + 872.0, + 1404.0, + 913.0, + 294.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 905.0, + 1394.0, + 905.0, + 1394.0, + 942.0, + 295.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1705.0, + 1265.0, + 1705.0, + 1265.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1705.0, + 1410.0, + 1705.0, + 1410.0, + 1748.0, + 1399.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1737.0, + 749.0, + 1737.0, + 749.0, + 1778.0, + 292.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1737.0, + 1407.0, + 1737.0, + 1407.0, + 1778.0, + 779.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1773.0, + 696.0, + 1773.0, + 696.0, + 1805.0, + 296.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1773.0, + 1402.0, + 1773.0, + 1402.0, + 1805.0, + 777.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1797.0, + 537.0, + 1797.0, + 537.0, + 1839.0, + 292.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1797.0, + 1405.0, + 1797.0, + 1405.0, + 1839.0, + 576.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1833.0, + 453.0, + 1833.0, + 453.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 1833.0, + 716.0, + 1833.0, + 716.0, + 1868.0, + 473.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1833.0, + 896.0, + 1833.0, + 896.0, + 1868.0, + 735.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1833.0, + 1350.0, + 1833.0, + 1350.0, + 1868.0, + 955.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1833.0, + 1404.0, + 1833.0, + 1404.0, + 1868.0, + 1368.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1862.0, + 620.0, + 1862.0, + 620.0, + 1899.0, + 292.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1862.0, + 744.0, + 1862.0, + 744.0, + 1899.0, + 678.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1862.0, + 1400.0, + 1862.0, + 1400.0, + 1899.0, + 896.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1046.0, + 1406.0, + 1046.0, + 1406.0, + 1083.0, + 293.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1078.0, + 1405.0, + 1078.0, + 1405.0, + 1112.0, + 296.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1104.0, + 424.0, + 1104.0, + 424.0, + 1146.0, + 291.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 1104.0, + 1018.0, + 1104.0, + 1018.0, + 1146.0, + 563.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1104.0, + 1175.0, + 1104.0, + 1175.0, + 1146.0, + 1095.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1104.0, + 1231.0, + 1104.0, + 1231.0, + 1146.0, + 1219.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1272.0, + 1371.0, + 1272.0, + 1371.0, + 1306.0, + 296.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1299.0, + 440.0, + 1299.0, + 440.0, + 1339.0, + 294.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1299.0, + 1406.0, + 1299.0, + 1406.0, + 1339.0, + 684.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1332.0, + 419.0, + 1332.0, + 419.0, + 1366.0, + 296.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1332.0, + 844.0, + 1332.0, + 844.0, + 1366.0, + 463.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1332.0, + 1024.0, + 1332.0, + 1024.0, + 1366.0, + 864.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1332.0, + 1334.0, + 1332.0, + 1334.0, + 1366.0, + 1045.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1332.0, + 1402.0, + 1332.0, + 1402.0, + 1366.0, + 1356.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1363.0, + 688.0, + 1363.0, + 688.0, + 1397.0, + 296.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1363.0, + 939.0, + 1363.0, + 939.0, + 1397.0, + 727.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1363.0, + 1405.0, + 1363.0, + 1405.0, + 1397.0, + 959.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1396.0, + 799.0, + 1396.0, + 799.0, + 1437.0, + 294.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1396.0, + 1229.0, + 1396.0, + 1229.0, + 1437.0, + 837.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 1396.0, + 1382.0, + 1396.0, + 1382.0, + 1437.0, + 1370.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 412.0, + 1404.0, + 412.0, + 1404.0, + 451.0, + 293.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 446.0, + 1404.0, + 446.0, + 1404.0, + 480.0, + 294.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 476.0, + 785.0, + 476.0, + 785.0, + 510.0, + 294.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1601.0, + 1407.0, + 1601.0, + 1407.0, + 1639.0, + 295.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1406.0, + 1633.0, + 1406.0, + 1668.0, + 294.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1663.0, + 1198.0, + 1663.0, + 1198.0, + 1701.0, + 295.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1442.0, + 1055.0, + 1442.0, + 1055.0, + 1484.0, + 293.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 1442.0, + 1402.0, + 1442.0, + 1402.0, + 1484.0, + 1099.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1475.0, + 1074.0, + 1475.0, + 1074.0, + 1514.0, + 294.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 307.0, + 1407.0, + 307.0, + 1407.0, + 344.0, + 295.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 337.0, + 1406.0, + 337.0, + 1406.0, + 375.0, + 295.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 365.0, + 1405.0, + 365.0, + 1405.0, + 405.0, + 294.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1407.0, + 229.0, + 1407.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 258.0, + 1409.0, + 258.0, + 1409.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1939.0, + 751.0, + 1939.0, + 751.0, + 1975.0, + 294.0, + 1975.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1105, + 1405, + 1105, + 1405, + 1354, + 297, + 1354 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1638, + 1404, + 1638, + 1404, + 1856, + 297, + 1856 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 769, + 1404, + 769, + 1404, + 956, + 297, + 956 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 601, + 1404, + 601, + 1404, + 757, + 298, + 757 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1404, + 228, + 1404, + 384, + 298, + 384 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1870, + 1403, + 1870, + 1403, + 1994, + 298, + 1994 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1452, + 1403, + 1452, + 1403, + 1548, + 298, + 1548 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 476, + 1404, + 476, + 1404, + 570, + 300, + 570 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 681, + 1360, + 1017, + 1360, + 1017, + 1445, + 681, + 1445 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 297, + 1043, + 1405, + 1043, + 1405, + 1076, + 297, + 1076 + ], + "score": 0.937 + }, + { + "category_id": 8, + "poly": [ + 604, + 2001, + 1094, + 2001, + 1094, + 2036, + 604, + 2036 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 299, + 1581, + 1063, + 1581, + 1063, + 1615, + 299, + 1615 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 300, + 420, + 674, + 420, + 674, + 452, + 300, + 452 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1366, + 972, + 1400, + 972, + 1400, + 1003, + 1366, + 1003 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1380, + 1399, + 1380, + 1399, + 1409, + 1366, + 1409 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 2005, + 1400, + 2005, + 1400, + 2033, + 1366, + 2033 + ], + "score": 0.868 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.794 + }, + { + "category_id": 8, + "poly": [ + 646, + 962, + 1054, + 962, + 1054, + 1034, + 646, + 1034 + ], + "score": 0.367 + }, + { + "category_id": 8, + "poly": [ + 642, + 961, + 1056, + 961, + 1056, + 1034, + 642, + 1034 + ], + "score": 0.312 + }, + { + "category_id": 14, + "poly": [ + 680, + 1358, + 1016, + 1358, + 1016, + 1447, + 680, + 1447 + ], + "score": 0.95, + "latex": "\\mathbf { c } \\gets \\mathbf { c } - \\eta \\frac { 1 } { | J _ { \\mathbf { c } } | } \\sum _ { ( k , l ) \\in J _ { \\mathbf { c } } } \\frac { \\partial \\mathcal { L } } { \\partial \\mathbf { b } _ { k l } } ," + }, + { + "category_id": 14, + "poly": [ + 644, + 961, + 1056, + 961, + 1056, + 1039, + 644, + 1039 + ], + "score": 0.94, + "latex": "\\Vert \\mathbf { W } - \\widehat { \\mathbf { W } } \\Vert _ { 2 } ^ { 2 } = \\sum _ { k , l } \\Vert \\mathbf { b } _ { k l } - \\mathbf { c } [ \\mathbf { I } _ { k l } ] \\Vert _ { 2 } ^ { 2 } ." + }, + { + "category_id": 13, + "poly": [ + 889, + 835, + 969, + 835, + 969, + 864, + 889, + 864 + ], + "score": 0.92, + "latex": "m \\times p" + }, + { + "category_id": 13, + "poly": [ + 898, + 291, + 987, + 291, + 987, + 324, + 898, + 324 + ], + "score": 0.91, + "latex": "\\times 3 2 / N" + }, + { + "category_id": 13, + "poly": [ + 486, + 1901, + 569, + 1901, + 569, + 1933, + 486, + 1933 + ], + "score": 0.91, + "latex": "\\log _ { 2 } K" + }, + { + "category_id": 13, + "poly": [ + 1138, + 1903, + 1215, + 1903, + 1215, + 1932, + 1138, + 1932 + ], + "score": 0.91, + "latex": "m \\times p" + }, + { + "category_id": 13, + "poly": [ + 298, + 923, + 405, + 923, + 405, + 952, + 298, + 952 + ], + "score": 0.9, + "latex": "K = 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 1034, + 1456, + 1117, + 1456, + 1117, + 1487, + 1034, + 1487 + ], + "score": 0.9, + "latex": "\\eta > 0" + }, + { + "category_id": 13, + "poly": [ + 376, + 1453, + 703, + 1453, + 703, + 1488, + 376, + 1488 + ], + "score": 0.9, + "latex": "J _ { \\mathbf { c } } = \\{ ( k , l ) | \\mathbf { c } [ \\mathbf { I } _ { k l } ] = \\mathbf { c } \\}" + }, + { + "category_id": 13, + "poly": [ + 814, + 728, + 882, + 728, + 882, + 757, + 814, + 757 + ], + "score": 0.89, + "latex": "q = p" + }, + { + "category_id": 13, + "poly": [ + 1063, + 323, + 1103, + 323, + 1103, + 351, + 1063, + 351 + ], + "score": 0.86, + "latex": "\\times 2" + }, + { + "category_id": 13, + "poly": [ + 1139, + 323, + 1179, + 323, + 1179, + 351, + 1139, + 351 + ], + "score": 0.86, + "latex": "\\times 4" + }, + { + "category_id": 13, + "poly": [ + 767, + 894, + 786, + 894, + 786, + 921, + 767, + 921 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1113, + 894, + 1143, + 894, + 1143, + 921, + 1113, + 921 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1049, + 777, + 1067, + 777, + 1067, + 804, + 1049, + 804 + ], + "score": 0.81, + "latex": "p" + }, + { + "category_id": 14, + "poly": [ + 600, + 2001, + 1101, + 2001, + 1101, + 2036, + 600, + 2036 + ], + "score": 0.81, + "latex": "M = 8 \\times K d + \\log _ { 2 } { K } \\times m p + 8 \\times n \\mathrm { b i t s } ." + }, + { + "category_id": 13, + "poly": [ + 298, + 324, + 326, + 324, + 326, + 349, + 298, + 349 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1117, + 1872, + 1146, + 1872, + 1146, + 1898, + 1117, + 1898 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 298, + 1902, + 316, + 1902, + 316, + 1928, + 298, + 1928 + ], + "score": 0.77, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 477, + 807, + 496, + 807, + 496, + 834, + 477, + 834 + ], + "score": 0.74, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1200, + 807, + 1228, + 807, + 1228, + 830, + 1200, + 830 + ], + "score": 0.74, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 601, + 2001, + 1058, + 2001, + 1058, + 2036, + 601, + 2036 + ], + "score": 0.63, + "latex": "M = 8 \\times K d + \\log _ { 2 } { K } \\times m p + 8 \\times n \\mathrm { t }" + }, + { + "category_id": 13, + "poly": [ + 720, + 1455, + 743, + 1455, + 743, + 1483, + 720, + 1483 + ], + "score": 0.63, + "latex": "\\mathcal { L }" + }, + { + "category_id": 13, + "poly": [ + 929, + 772, + 966, + 772, + 966, + 800, + 929, + 800 + ], + "score": 0.34, + "latex": "\\mathbf { W }" + }, + { + "category_id": 13, + "poly": [ + 817, + 1873, + 885, + 1873, + 885, + 1899, + 817, + 1899 + ], + "score": 0.32, + "latex": "\\dot { \\mathtt { 1 } } \\mathtt { n t } \\mathtt { 8 }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1577.0, + 1067.0, + 1577.0, + 1067.0, + 1622.0, + 292.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 416.0, + 678.0, + 416.0, + 678.0, + 459.0, + 293.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1106.0, + 1406.0, + 1106.0, + 1406.0, + 1146.0, + 292.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1139.0, + 1407.0, + 1139.0, + 1407.0, + 1173.0, + 295.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1169.0, + 1406.0, + 1169.0, + 1406.0, + 1205.0, + 294.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1199.0, + 1405.0, + 1199.0, + 1405.0, + 1233.0, + 295.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1228.0, + 1406.0, + 1228.0, + 1406.0, + 1266.0, + 292.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1260.0, + 1408.0, + 1260.0, + 1408.0, + 1296.0, + 292.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1293.0, + 1405.0, + 1293.0, + 1405.0, + 1324.0, + 292.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1320.0, + 1328.0, + 1320.0, + 1328.0, + 1359.0, + 292.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1640.0, + 1405.0, + 1640.0, + 1405.0, + 1675.0, + 295.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1672.0, + 1406.0, + 1672.0, + 1406.0, + 1707.0, + 295.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1701.0, + 1406.0, + 1701.0, + 1406.0, + 1738.0, + 292.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1732.0, + 1405.0, + 1732.0, + 1405.0, + 1771.0, + 292.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1763.0, + 1404.0, + 1763.0, + 1404.0, + 1798.0, + 295.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1793.0, + 1404.0, + 1793.0, + 1404.0, + 1828.0, + 295.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1824.0, + 1081.0, + 1824.0, + 1081.0, + 1859.0, + 294.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 770.0, + 928.0, + 770.0, + 928.0, + 806.0, + 295.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 770.0, + 1048.0, + 770.0, + 1048.0, + 806.0, + 967.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 770.0, + 1405.0, + 770.0, + 1405.0, + 806.0, + 1068.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 801.0, + 476.0, + 801.0, + 476.0, + 836.0, + 294.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 801.0, + 1199.0, + 801.0, + 1199.0, + 836.0, + 497.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 801.0, + 1405.0, + 801.0, + 1405.0, + 836.0, + 1229.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 830.0, + 888.0, + 830.0, + 888.0, + 865.0, + 294.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 830.0, + 1405.0, + 830.0, + 1405.0, + 865.0, + 970.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 859.0, + 1408.0, + 859.0, + 1408.0, + 901.0, + 292.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 889.0, + 766.0, + 889.0, + 766.0, + 928.0, + 294.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 889.0, + 1112.0, + 889.0, + 1112.0, + 928.0, + 787.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 889.0, + 1404.0, + 889.0, + 1404.0, + 928.0, + 1144.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 923.0, + 1317.0, + 923.0, + 1317.0, + 957.0, + 406.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 599.0, + 1405.0, + 599.0, + 1405.0, + 638.0, + 295.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 631.0, + 1406.0, + 631.0, + 1406.0, + 668.0, + 294.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 664.0, + 1404.0, + 664.0, + 1404.0, + 698.0, + 296.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 687.0, + 1405.0, + 687.0, + 1405.0, + 732.0, + 292.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 722.0, + 813.0, + 722.0, + 813.0, + 762.0, + 295.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 722.0, + 893.0, + 722.0, + 893.0, + 762.0, + 883.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1406.0, + 230.0, + 1406.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 897.0, + 290.0, + 897.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 290.0, + 1406.0, + 290.0, + 1406.0, + 326.0, + 988.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 321.0, + 1062.0, + 321.0, + 1062.0, + 355.0, + 327.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 321.0, + 1138.0, + 321.0, + 1138.0, + 355.0, + 1104.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 355.0, + 1180.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 350.0, + 1214.0, + 350.0, + 1214.0, + 389.0, + 295.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1871.0, + 816.0, + 1871.0, + 816.0, + 1904.0, + 294.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1871.0, + 1116.0, + 1871.0, + 1116.0, + 1904.0, + 886.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1871.0, + 1404.0, + 1871.0, + 1404.0, + 1904.0, + 1147.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1898.0, + 297.0, + 1898.0, + 297.0, + 1936.0, + 292.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1898.0, + 485.0, + 1898.0, + 485.0, + 1936.0, + 317.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1898.0, + 1137.0, + 1898.0, + 1137.0, + 1936.0, + 570.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1898.0, + 1405.0, + 1898.0, + 1405.0, + 1936.0, + 1216.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1932.0, + 1404.0, + 1932.0, + 1404.0, + 1965.0, + 294.0, + 1965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1964.0, + 1381.0, + 1964.0, + 1381.0, + 1996.0, + 294.0, + 1996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1450.0, + 375.0, + 1450.0, + 375.0, + 1491.0, + 293.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1450.0, + 719.0, + 1450.0, + 719.0, + 1491.0, + 704.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1450.0, + 1033.0, + 1450.0, + 1033.0, + 1491.0, + 744.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1450.0, + 1405.0, + 1450.0, + 1405.0, + 1491.0, + 1118.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1485.0, + 1403.0, + 1485.0, + 1403.0, + 1520.0, + 296.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1517.0, + 779.0, + 1517.0, + 779.0, + 1550.0, + 294.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 477.0, + 1403.0, + 477.0, + 1403.0, + 511.0, + 295.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 510.0, + 1403.0, + 510.0, + 1403.0, + 540.0, + 296.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 538.0, + 1132.0, + 538.0, + 1132.0, + 575.0, + 294.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 1411.0, + 1039.0, + 1411.0, + 1080.0, + 294.0, + 1080.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 526, + 1405, + 526, + 1405, + 863, + 298, + 863 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 1444, + 1404, + 1444, + 1404, + 1661, + 297, + 1661 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1404, + 228, + 1404, + 415, + 299, + 415 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 955, + 1404, + 955, + 1404, + 1081, + 298, + 1081 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1094, + 1404, + 1094, + 1404, + 1188, + 298, + 1188 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1292, + 1404, + 1292, + 1404, + 1386, + 297, + 1386 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 633, + 1203, + 1063, + 1203, + 1063, + 1279, + 633, + 1279 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 295, + 1692, + 1401, + 1692, + 1401, + 1755, + 295, + 1755 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 298, + 1816, + 1410, + 1816, + 1410, + 1879, + 298, + 1879 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 295, + 1971, + 1400, + 1971, + 1400, + 2035, + 295, + 2035 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 672, + 1767, + 1028, + 1767, + 1028, + 1806, + 672, + 1806 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 557, + 1397, + 1140, + 1397, + 1140, + 1436, + 557, + 1436 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 298, + 1914, + 1023, + 1914, + 1023, + 1947, + 298, + 1947 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 457, + 481, + 457, + 481, + 493, + 299, + 493 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 298, + 899, + 979, + 899, + 979, + 931, + 298, + 931 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1401, + 1399, + 1401, + 1399, + 1431, + 1366, + 1431 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1771, + 1399, + 1771, + 1399, + 1800, + 1366, + 1800 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1367, + 1227, + 1400, + 1227, + 1400, + 1256, + 1367, + 1256 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.747 + }, + { + "category_id": 14, + "poly": [ + 633, + 1202, + 1062, + 1202, + 1062, + 1280, + 633, + 1280 + ], + "score": 0.94, + "latex": "\\psi ( \\mathbf { b } _ { k l } \\mid J ) = { \\left\\{ \\begin{array} { l l } { \\varphi ( \\mathbf { b } _ { k l } ) } & { { \\mathrm { i f ~ } } ( k , l ) \\in J , } \\\\ { \\mathbf { b } _ { k l } } & { { \\mathrm { o t h e r w i s e } } . } \\end{array} \\right. }" + }, + { + "category_id": 13, + "poly": [ + 843, + 1096, + 994, + 1096, + 994, + 1130, + 843, + 1130 + ], + "score": 0.93, + "latex": "J \\subset \\{ ( k , l ) \\}" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1128, + 1186, + 1128, + 1186, + 1160, + 1095, + 1160 + ], + "score": 0.92, + "latex": "\\psi ( \\cdot \\mid J )" + }, + { + "category_id": 14, + "poly": [ + 671, + 1766, + 1028, + 1766, + 1028, + 1805, + 671, + 1805 + ], + "score": 0.92, + "latex": "\\psi ( \\mathbf { b } _ { k l } \\mid J ) = \\psi _ { 1 } \\circ \\psi _ { 2 } ( \\mathbf { b } _ { k l } \\mid J ) ." + }, + { + "category_id": 13, + "poly": [ + 493, + 1476, + 571, + 1476, + 571, + 1507, + 493, + 1507 + ], + "score": 0.9, + "latex": "\\mathbf { W _ { \\mathrm { n o i s e } } }" + }, + { + "category_id": 13, + "poly": [ + 517, + 231, + 624, + 231, + 624, + 260, + 517, + 260 + ], + "score": 0.9, + "latex": "K = 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 1151, + 989, + 1194, + 989, + 1194, + 1020, + 1151, + 1020 + ], + "score": 0.89, + "latex": "\\mathbf { b } _ { k l }" + }, + { + "category_id": 13, + "poly": [ + 1242, + 1323, + 1320, + 1323, + 1320, + 1354, + 1242, + 1354 + ], + "score": 0.88, + "latex": "\\mathbf { W } _ { \\mathrm { n o i s e } }" + }, + { + "category_id": 13, + "poly": [ + 365, + 1157, + 406, + 1157, + 406, + 1188, + 365, + 1188 + ], + "score": 0.88, + "latex": "\\mathbf { b } _ { k l }" + }, + { + "category_id": 14, + "poly": [ + 557, + 1396, + 1143, + 1396, + 1143, + 1435, + 557, + 1435 + ], + "score": 0.85, + "latex": "\\mathbf { W } _ { \\mathrm { n o i s e } } = ( \\psi ( \\mathbf { b } _ { k l } \\mid J ) ) _ { k l } \\mathrm { a n d } \\mathbf { y } _ { \\mathrm { n o i s e } } = \\mathbf { x } \\mathbf { W } _ { \\mathrm { n o i s e } }" + }, + { + "category_id": 13, + "poly": [ + 507, + 1298, + 529, + 1298, + 529, + 1324, + 507, + 1324 + ], + "score": 0.82, + "latex": "\\varphi" + }, + { + "category_id": 13, + "poly": [ + 781, + 1357, + 843, + 1357, + 843, + 1385, + 781, + 1385 + ], + "score": 0.82, + "latex": "\\mathbf { y } _ { \\mathrm { n o i s e } }" + }, + { + "category_id": 13, + "poly": [ + 1146, + 1978, + 1168, + 1978, + 1168, + 2005, + 1146, + 2005 + ], + "score": 0.82, + "latex": "\\varphi" + }, + { + "category_id": 13, + "poly": [ + 612, + 1132, + 634, + 1132, + 634, + 1159, + 612, + 1159 + ], + "score": 0.82, + "latex": "\\varphi" + }, + { + "category_id": 13, + "poly": [ + 760, + 1603, + 781, + 1603, + 781, + 1629, + 760, + 1629 + ], + "score": 0.81, + "latex": "\\varphi" + }, + { + "category_id": 13, + "poly": [ + 428, + 1508, + 449, + 1508, + 449, + 1535, + 428, + 1535 + ], + "score": 0.8, + "latex": "J" + }, + { + "category_id": 13, + "poly": [ + 1201, + 1097, + 1331, + 1097, + 1331, + 1128, + 1201, + 1128 + ], + "score": 0.76, + "latex": "1 \\leq l \\leq q" + }, + { + "category_id": 13, + "poly": [ + 1043, + 1097, + 1186, + 1097, + 1186, + 1127, + 1043, + 1127 + ], + "score": 0.72, + "latex": "1 \\leq k \\leq m" + }, + { + "category_id": 13, + "poly": [ + 369, + 1451, + 390, + 1451, + 390, + 1474, + 369, + 1474 + ], + "score": 0.69, + "latex": "\\mathbf { x }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1914.0, + 1025.0, + 1914.0, + 1025.0, + 1950.0, + 295.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 454.0, + 486.0, + 454.0, + 486.0, + 500.0, + 291.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 896.0, + 982.0, + 896.0, + 982.0, + 937.0, + 293.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 524.0, + 1405.0, + 524.0, + 1405.0, + 565.0, + 292.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 558.0, + 1407.0, + 558.0, + 1407.0, + 593.0, + 295.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 586.0, + 1406.0, + 586.0, + 1406.0, + 626.0, + 292.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 619.0, + 1405.0, + 619.0, + 1405.0, + 654.0, + 293.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 650.0, + 1406.0, + 650.0, + 1406.0, + 683.0, + 293.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 678.0, + 1407.0, + 678.0, + 1407.0, + 717.0, + 293.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 711.0, + 1406.0, + 711.0, + 1406.0, + 745.0, + 295.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 740.0, + 1406.0, + 740.0, + 1406.0, + 775.0, + 295.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 769.0, + 1406.0, + 769.0, + 1406.0, + 810.0, + 292.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 803.0, + 1405.0, + 803.0, + 1405.0, + 837.0, + 295.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 830.0, + 985.0, + 830.0, + 985.0, + 870.0, + 293.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1446.0, + 368.0, + 1446.0, + 368.0, + 1481.0, + 295.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1446.0, + 1404.0, + 1446.0, + 1404.0, + 1481.0, + 391.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 492.0, + 1473.0, + 492.0, + 1512.0, + 294.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1473.0, + 1406.0, + 1473.0, + 1406.0, + 1512.0, + 572.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1507.0, + 427.0, + 1507.0, + 427.0, + 1541.0, + 295.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1507.0, + 1406.0, + 1507.0, + 1406.0, + 1541.0, + 450.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1536.0, + 1406.0, + 1536.0, + 1406.0, + 1571.0, + 294.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1566.0, + 1406.0, + 1566.0, + 1406.0, + 1605.0, + 292.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1598.0, + 759.0, + 1598.0, + 759.0, + 1633.0, + 294.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1598.0, + 1405.0, + 1598.0, + 1405.0, + 1633.0, + 782.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1629.0, + 1331.0, + 1629.0, + 1331.0, + 1664.0, + 295.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 516.0, + 229.0, + 516.0, + 264.0, + 294.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 264.0, + 625.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1406.0, + 291.0, + 1406.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 323.0, + 1403.0, + 323.0, + 1403.0, + 355.0, + 297.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 382.0, + 1256.0, + 382.0, + 1256.0, + 418.0, + 294.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 956.0, + 1406.0, + 956.0, + 1406.0, + 991.0, + 295.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 987.0, + 1150.0, + 987.0, + 1150.0, + 1024.0, + 292.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 987.0, + 1406.0, + 987.0, + 1406.0, + 1024.0, + 1195.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1019.0, + 1406.0, + 1019.0, + 1406.0, + 1052.0, + 293.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1050.0, + 688.0, + 1050.0, + 688.0, + 1083.0, + 294.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1091.0, + 842.0, + 1091.0, + 842.0, + 1134.0, + 292.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1091.0, + 1042.0, + 1091.0, + 1042.0, + 1134.0, + 995.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1091.0, + 1200.0, + 1091.0, + 1200.0, + 1134.0, + 1187.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 1091.0, + 1405.0, + 1091.0, + 1405.0, + 1134.0, + 1332.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1125.0, + 611.0, + 1125.0, + 611.0, + 1163.0, + 294.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1125.0, + 1094.0, + 1125.0, + 1094.0, + 1163.0, + 635.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1125.0, + 1405.0, + 1125.0, + 1405.0, + 1163.0, + 1187.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1153.0, + 364.0, + 1153.0, + 364.0, + 1192.0, + 293.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1153.0, + 859.0, + 1153.0, + 859.0, + 1192.0, + 407.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1290.0, + 506.0, + 1290.0, + 506.0, + 1328.0, + 294.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1290.0, + 1405.0, + 1290.0, + 1405.0, + 1328.0, + 530.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1321.0, + 1241.0, + 1321.0, + 1241.0, + 1360.0, + 294.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 1321.0, + 1404.0, + 1321.0, + 1404.0, + 1360.0, + 1321.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1352.0, + 780.0, + 1352.0, + 780.0, + 1389.0, + 295.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1352.0, + 903.0, + 1352.0, + 903.0, + 1389.0, + 844.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1693.0, + 1405.0, + 1693.0, + 1405.0, + 1729.0, + 297.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1722.0, + 1406.0, + 1722.0, + 1406.0, + 1758.0, + 295.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1813.0, + 1408.0, + 1813.0, + 1408.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1845.0, + 1050.0, + 1845.0, + 1050.0, + 1883.0, + 293.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1145.0, + 1970.0, + 1145.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1970.0, + 1404.0, + 1970.0, + 1404.0, + 2008.0, + 1169.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1161.0, + 2002.0, + 1161.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 296, + 223, + 1406, + 223, + 1406, + 752, + 296, + 752 + ], + "score": 0.985, + "html": "
Quantization SchemeLanguage Modeling 16-layer Transformer Wikitext-103Image Classification EfficientNet-B3 ImageNet-1k
SizeCompressionPPLSizeCompressionTop-1
Uncompressed model942×118.346.7×181.5
int 4 quantization118×839.45.8×845.3
- trained with QAT118×834.15.8×859.4
- trained with Quant-Noise118×821.85.8×867.8
int 8 quantization236×419.611.7×480.7
- trained with QAT236×421.011.7×480.8
- trained with Quant-Noise236×418.711.7×480.9
iPQ38×2525.23.3×1479.0
- trained with QAT38×2541.23.3×1455.7
- trained with Quant-Noise38×2520.73.3×1480.0
iPQ&int8 +Quant-Noise38×2521.13.1×1579.8
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1757, + 1405, + 1757, + 1405, + 2035, + 297, + 2035 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1510, + 1405, + 1510, + 1405, + 1727, + 298, + 1727 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1248, + 1403, + 1248, + 1403, + 1372, + 297, + 1372 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 999, + 1400, + 999, + 1400, + 1094, + 296, + 1094 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 298, + 1433, + 1400, + 1433, + 1400, + 1497, + 298, + 1497 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 300, + 1154, + 1404, + 1154, + 1404, + 1217, + 300, + 1217 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 671, + 1384, + 1029, + 1384, + 1029, + 1424, + 671, + 1424 + ], + "score": 0.939 + }, + { + "category_id": 8, + "poly": [ + 621, + 1105, + 1078, + 1105, + 1078, + 1143, + 621, + 1143 + ], + "score": 0.935 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1388, + 1399, + 1388, + 1399, + 1418, + 1353, + 1418 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1108, + 1399, + 1108, + 1399, + 1138, + 1366, + 1138 + ], + "score": 0.885 + }, + { + "category_id": 6, + "poly": [ + 296, + 774, + 1406, + 774, + 1406, + 944, + 296, + 944 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.791 + }, + { + "category_id": 14, + "poly": [ + 621, + 1104, + 1077, + 1104, + 1077, + 1142, + 621, + 1142 + ], + "score": 0.93, + "latex": "\\varphi _ { \\mathrm { i n t N } } ( w ) = ( \\mathrm { r o u n d } ( w / s + z ) - z ) \\times s ," + }, + { + "category_id": 14, + "poly": [ + 669, + 1381, + 1030, + 1381, + 1030, + 1425, + 669, + 1425 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\varphi _ { \\mathrm { P Q } } ( \\mathbf { v } ) = \\operatorname * { a r g m i n } _ { \\mathbf { c } \\in \\mathcal { C } } \\| \\mathbf { b } - \\mathbf { c } \\| _ { 2 } ^ { 2 } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1179, + 1573, + 1336, + 1573, + 1336, + 1607, + 1179, + 1607 + ], + "score": 0.92, + "latex": "\\varphi _ { \\mathrm { p r o x y } } ( \\mathbf { v } ) = 0" + }, + { + "category_id": 13, + "poly": [ + 1060, + 1316, + 1108, + 1316, + 1108, + 1344, + 1060, + 1344 + ], + "score": 0.9, + "latex": "\\varphi _ { \\mathrm { P Q } }" + }, + { + "category_id": 13, + "poly": [ + 792, + 1547, + 860, + 1547, + 860, + 1577, + 792, + 1577 + ], + "score": 0.9, + "latex": "\\varphi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 1244, + 1032, + 1301, + 1032, + 1301, + 1063, + 1244, + 1063 + ], + "score": 0.9, + "latex": "{ \\bf W } _ { k l }" + }, + { + "category_id": 13, + "poly": [ + 350, + 1972, + 553, + 1972, + 553, + 2007, + 350, + 2007 + ], + "score": 0.9, + "latex": "\\dot { \\varphi } _ { \\mathrm { L a y e r D r o p } } ( \\mathbf { v } ) = 0 ." + }, + { + "category_id": 13, + "poly": [ + 1227, + 886, + 1297, + 886, + 1297, + 914, + 1227, + 914 + ], + "score": 0.89, + "latex": "8 1 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1034, + 1036, + 1056, + 1036, + 1056, + 1064, + 1034, + 1064 + ], + "score": 0.82, + "latex": "\\varphi" + }, + { + "category_id": 13, + "poly": [ + 1128, + 1436, + 1147, + 1436, + 1147, + 1462, + 1128, + 1462 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 863, + 1343, + 883, + 1343, + 883, + 1369, + 863, + 1369 + ], + "score": 0.79, + "latex": "\\mathcal { C }" + }, + { + "category_id": 13, + "poly": [ + 442, + 1161, + 460, + 1161, + 460, + 1182, + 442, + 1182 + ], + "score": 0.76, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 373, + 1161, + 389, + 1161, + 389, + 1182, + 373, + 1182 + ], + "score": 0.61, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 296, + 1977, + 318, + 1977, + 318, + 2001, + 296, + 2001 + ], + "score": 0.43, + "latex": "\\mathbf { v }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 772.0, + 1408.0, + 772.0, + 1408.0, + 807.0, + 293.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 804.0, + 1407.0, + 804.0, + 1407.0, + 835.0, + 293.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 830.0, + 1406.0, + 830.0, + 1406.0, + 863.0, + 293.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 858.0, + 1404.0, + 858.0, + 1404.0, + 892.0, + 294.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 886.0, + 1226.0, + 886.0, + 1226.0, + 919.0, + 294.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 886.0, + 1406.0, + 886.0, + 1406.0, + 919.0, + 1298.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 916.0, + 1378.0, + 916.0, + 1378.0, + 946.0, + 294.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1758.0, + 1408.0, + 1758.0, + 1408.0, + 1795.0, + 295.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1787.0, + 1406.0, + 1787.0, + 1406.0, + 1826.0, + 292.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1816.0, + 1405.0, + 1816.0, + 1405.0, + 1859.0, + 292.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1887.0, + 294.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1918.0, + 294.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1949.0, + 295.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 1403.0, + 1941.0, + 1403.0, + 1977.0, + 295.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1972.0, + 295.0, + 1972.0, + 295.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1972.0, + 349.0, + 1972.0, + 349.0, + 2009.0, + 319.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2009.0, + 554.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 1406.0, + 2000.0, + 1406.0, + 2043.0, + 292.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1508.0, + 1409.0, + 1508.0, + 1409.0, + 1549.0, + 292.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1540.0, + 791.0, + 1540.0, + 791.0, + 1583.0, + 292.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1540.0, + 1410.0, + 1540.0, + 1410.0, + 1583.0, + 861.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1568.0, + 1178.0, + 1568.0, + 1178.0, + 1612.0, + 291.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 1568.0, + 1410.0, + 1568.0, + 1410.0, + 1612.0, + 1337.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1638.0, + 293.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1631.0, + 1407.0, + 1631.0, + 1407.0, + 1671.0, + 291.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1664.0, + 1406.0, + 1664.0, + 1406.0, + 1699.0, + 295.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1693.0, + 1406.0, + 1693.0, + 1406.0, + 1730.0, + 293.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1247.0, + 1407.0, + 1247.0, + 1407.0, + 1284.0, + 292.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1281.0, + 1407.0, + 1281.0, + 1407.0, + 1317.0, + 293.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1310.0, + 1059.0, + 1310.0, + 1059.0, + 1346.0, + 292.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1310.0, + 1406.0, + 1310.0, + 1406.0, + 1346.0, + 1109.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1343.0, + 862.0, + 1343.0, + 862.0, + 1373.0, + 293.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1343.0, + 895.0, + 1343.0, + 895.0, + 1373.0, + 884.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 998.0, + 1405.0, + 998.0, + 1405.0, + 1037.0, + 294.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1032.0, + 1033.0, + 1032.0, + 1033.0, + 1066.0, + 295.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 1032.0, + 1243.0, + 1032.0, + 1243.0, + 1066.0, + 1057.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1032.0, + 1404.0, + 1032.0, + 1404.0, + 1066.0, + 1302.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1062.0, + 908.0, + 1062.0, + 908.0, + 1096.0, + 296.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 1127.0, + 1432.0, + 1127.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1432.0, + 1402.0, + 1432.0, + 1402.0, + 1470.0, + 1148.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1461.0, + 1001.0, + 1461.0, + 1001.0, + 1502.0, + 292.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1154.0, + 372.0, + 1154.0, + 372.0, + 1189.0, + 295.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1154.0, + 441.0, + 1154.0, + 441.0, + 1189.0, + 390.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1154.0, + 1404.0, + 1154.0, + 1404.0, + 1189.0, + 461.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1183.0, + 916.0, + 1183.0, + 916.0, + 1219.0, + 296.0, + 1219.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 297, + 693, + 1407, + 693, + 1407, + 1062, + 297, + 1062 + ], + "score": 0.982, + "html": "
Language modelingSentence RepresentationImage Classification
Comp.SizePPLComp.SizeAcc.Comp.SizeAcc.
Unquantized models
Original model×194218.3×148084.8×146.781.5
+ Sharing×1.851018.7×1.925084.0×1.434.280.1
+ Pruning×3.725522.5×3.812581.3×1.629.578.5
Quantized models
iPQ×24.83825.2×12.63882.5× 14.13.379.0
+ Quant-Noise×24.83820.7×12.63883.6×14.13.380.0
+ Sharing× 49.51922.0×34.31482.5×182.678.9
+ Pruning×94.21024.7×58.5878.8×202.377.8
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1680, + 1404, + 1680, + 1404, + 1866, + 298, + 1866 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1404, + 1881, + 1404, + 2034, + 299, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 296, + 243, + 1394, + 243, + 1394, + 551, + 296, + 551 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 301, + 1573, + 1403, + 1573, + 1403, + 1665, + 301, + 1665 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 296, + 1220, + 1401, + 1220, + 1401, + 1283, + 296, + 1283 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 299, + 1407, + 1403, + 1407, + 1403, + 1470, + 299, + 1470 + ], + "score": 0.949 + }, + { + "category_id": 4, + "poly": [ + 296, + 559, + 1405, + 559, + 1405, + 672, + 296, + 672 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 299, + 1334, + 480, + 1334, + 480, + 1369, + 299, + 1369 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 298, + 1513, + 934, + 1513, + 934, + 1545, + 298, + 1545 + ], + "score": 0.886 + }, + { + "category_id": 6, + "poly": [ + 298, + 1070, + 1406, + 1070, + 1406, + 1183, + 298, + 1183 + ], + "score": 0.75 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2111, + 841, + 2111 + ], + "score": 0.736 + }, + { + "category_id": 4, + "poly": [ + 298, + 1070, + 1406, + 1070, + 1406, + 1183, + 298, + 1183 + ], + "score": 0.16 + }, + { + "category_id": 13, + "poly": [ + 1225, + 1835, + 1290, + 1835, + 1290, + 1864, + 1225, + 1864 + ], + "score": 0.85, + "latex": "( < 5 \\%" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 241.0, + 594.0, + 241.0, + 594.0, + 272.0, + 416.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 241.0, + 907.0, + 241.0, + 907.0, + 272.0, + 843.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 247.0, + 1312.0, + 247.0, + 1312.0, + 276.0, + 1171.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 273.0, + 361.0, + 273.0, + 361.0, + 303.0, + 319.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 275.0, + 620.0, + 275.0, + 620.0, + 300.0, + 546.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 289.0, + 729.0, + 289.0, + 729.0, + 319.0, + 694.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 277.0, + 796.0, + 277.0, + 796.0, + 306.0, + 730.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 299.0, + 956.0, + 299.0, + 956.0, + 327.0, + 839.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 269.0, + 1165.0, + 269.0, + 1165.0, + 309.0, + 1065.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 290.0, + 1364.0, + 290.0, + 1364.0, + 319.0, + 1254.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 329.0, + 358.0, + 329.0, + 358.0, + 438.0, + 295.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 329.0, + 500.0, + 329.0, + 500.0, + 362.0, + 377.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 311.0, + 728.0, + 311.0, + 728.0, + 436.0, + 664.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 354.0, + 856.0, + 354.0, + 856.0, + 384.0, + 790.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 315.0, + 997.0, + 315.0, + 997.0, + 345.0, + 925.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 321.0, + 1097.0, + 321.0, + 1097.0, + 438.0, + 1037.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 322.0, + 1388.0, + 322.0, + 1388.0, + 406.0, + 1102.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 386.0, + 358.0, + 386.0, + 358.0, + 416.0, + 325.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 369.0, + 648.0, + 369.0, + 648.0, + 432.0, + 481.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 380.0, + 729.0, + 380.0, + 729.0, + 416.0, + 693.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 394.0, + 964.0, + 394.0, + 964.0, + 437.0, + 735.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 366.0, + 1348.0, + 366.0, + 1348.0, + 444.0, + 1066.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 442.0, + 358.0, + 442.0, + 358.0, + 472.0, + 325.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 433.0, + 641.0, + 433.0, + 641.0, + 463.0, + 479.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 431.0, + 729.0, + 431.0, + 729.0, + 462.0, + 694.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 448.0, + 879.0, + 448.0, + 879.0, + 477.0, + 790.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 442.0, + 1013.0, + 442.0, + 1013.0, + 464.0, + 892.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 435.0, + 1100.0, + 435.0, + 1100.0, + 466.0, + 1066.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 443.0, + 1299.0, + 443.0, + 1299.0, + 471.0, + 1174.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 462.0, + 431.0, + 462.0, + 431.0, + 496.0, + 360.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 467.0, + 593.0, + 467.0, + 593.0, + 496.0, + 492.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 466.0, + 887.0, + 466.0, + 887.0, + 491.0, + 793.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 491.0, + 610.0, + 491.0, + 610.0, + 527.0, + 371.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 480.0, + 749.0, + 480.0, + 749.0, + 525.0, + 694.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 495.0, + 810.0, + 495.0, + 810.0, + 525.0, + 771.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 494.0, + 881.0, + 494.0, + 881.0, + 526.0, + 834.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 494.0, + 941.0, + 494.0, + 941.0, + 525.0, + 890.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 496.0, + 1007.0, + 496.0, + 1007.0, + 523.0, + 955.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 503.0, + 1108.0, + 503.0, + 1108.0, + 521.0, + 1093.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 498.0, + 1156.0, + 498.0, + 1156.0, + 526.0, + 1130.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 495.0, + 1395.0, + 495.0, + 1395.0, + 527.0, + 1172.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 522.0, + 555.0, + 522.0, + 555.0, + 555.0, + 455.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 522.0, + 926.0, + 522.0, + 926.0, + 558.0, + 823.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 523.0, + 1292.0, + 523.0, + 1292.0, + 557.0, + 1194.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.25, + 318.0, + 957.25, + 318.0, + 957.25, + 369.0, + 837.25, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 369.0, + 495.0, + 369.0, + 495.0, + 404.5, + 433.0, + 404.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 421.0, + 562.0, + 421.0, + 562.0, + 444.5, + 525.0, + 444.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 558.0, + 1405.0, + 558.0, + 1405.0, + 592.0, + 296.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 588.0, + 1403.0, + 588.0, + 1403.0, + 617.0, + 296.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 613.0, + 1402.0, + 613.0, + 1402.0, + 646.0, + 295.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 639.0, + 873.0, + 639.0, + 873.0, + 675.0, + 294.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1331.0, + 484.0, + 1331.0, + 484.0, + 1376.0, + 292.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1508.0, + 939.0, + 1508.0, + 939.0, + 1553.0, + 292.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1067.0, + 1404.0, + 1067.0, + 1404.0, + 1105.0, + 292.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1100.0, + 1403.0, + 1100.0, + 1403.0, + 1129.0, + 296.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1122.0, + 1404.0, + 1122.0, + 1404.0, + 1158.0, + 293.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1153.0, + 885.0, + 1153.0, + 885.0, + 1186.0, + 296.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1067.0, + 1404.0, + 1067.0, + 1404.0, + 1105.0, + 292.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1100.0, + 1403.0, + 1100.0, + 1403.0, + 1129.0, + 296.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1122.0, + 1404.0, + 1122.0, + 1404.0, + 1158.0, + 293.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1153.0, + 885.0, + 1153.0, + 885.0, + 1186.0, + 296.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1404.0, + 1680.0, + 1404.0, + 1717.0, + 293.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 1404.0, + 1712.0, + 1404.0, + 1746.0, + 295.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 1406.0, + 1740.0, + 1406.0, + 1778.0, + 293.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 1406.0, + 1772.0, + 1406.0, + 1810.0, + 293.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1804.0, + 1406.0, + 1804.0, + 1406.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1834.0, + 1224.0, + 1834.0, + 1224.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 1834.0, + 1389.0, + 1834.0, + 1389.0, + 1870.0, + 1291.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 1401.0, + 1881.0, + 1401.0, + 1914.0, + 295.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1408.0, + 1910.0, + 1408.0, + 1948.0, + 293.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1980.0, + 294.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 910.0, + 2002.0, + 910.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1572.0, + 1404.0, + 1572.0, + 1404.0, + 1609.0, + 296.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1606.0, + 1403.0, + 1606.0, + 1403.0, + 1639.0, + 296.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1099.0, + 1633.0, + 1099.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1220.0, + 1403.0, + 1220.0, + 1403.0, + 1255.0, + 295.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1252.0, + 929.0, + 1252.0, + 929.0, + 1287.0, + 296.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1403.0, + 1408.0, + 1403.0, + 1408.0, + 1445.0, + 293.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1439.0, + 795.0, + 1439.0, + 795.0, + 1474.0, + 295.0, + 1474.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1015, + 1405, + 1015, + 1405, + 1230, + 298, + 1230 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 384, + 224, + 1310, + 224, + 1310, + 384, + 384, + 384 + ], + "score": 0.979, + "html": "
Language ModelingPPLRoBERTaAcc.
Train without Quant-Noise25.2Train without Quant-Noise82.5
+ Finetune with Quant-Noise20.9+ Finetune with Quant-Noise83.4
Train with Quant-Noise20.7Train with Quant-Noise83.6
" + }, + { + "category_id": 1, + "poly": [ + 299, + 701, + 1405, + 701, + 1405, + 915, + 299, + 915 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1397, + 1404, + 1397, + 1404, + 1581, + 298, + 1581 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1597, + 1405, + 1597, + 1405, + 1781, + 298, + 1781 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 509, + 1403, + 509, + 1403, + 663, + 298, + 663 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1402, + 1880, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1267, + 1405, + 1267, + 1405, + 1359, + 299, + 1359 + ], + "score": 0.971 + }, + { + "category_id": 7, + "poly": [ + 295, + 390, + 1401, + 390, + 1401, + 447, + 295, + 447 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 105, + 300, + 105 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 299, + 958, + 876, + 958, + 876, + 988, + 299, + 988 + ], + "score": 0.879 + }, + { + "category_id": 0, + "poly": [ + 301, + 1821, + 1238, + 1821, + 1238, + 1854, + 301, + 1854 + ], + "score": 0.818 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.788 + }, + { + "category_id": 13, + "poly": [ + 333, + 419, + 392, + 419, + 392, + 448, + 333, + 448 + ], + "score": 0.91, + "latex": "\\phi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 722, + 1169, + 837, + 1169, + 837, + 1198, + 722, + 1198 + ], + "score": 0.89, + "latex": "\\times 2 0 - 9 4 " + }, + { + "category_id": 13, + "poly": [ + 915, + 1428, + 970, + 1428, + 970, + 1458, + 915, + 1458 + ], + "score": 0.88, + "latex": "8 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 683, + 1629, + 736, + 1629, + 736, + 1658, + 683, + 1658 + ], + "score": 0.88, + "latex": "\\times 2 5" + }, + { + "category_id": 13, + "poly": [ + 1274, + 1690, + 1327, + 1690, + 1327, + 1718, + 1274, + 1718 + ], + "score": 0.87, + "latex": "\\times 9 4" + }, + { + "category_id": 13, + "poly": [ + 1224, + 1659, + 1264, + 1659, + 1264, + 1689, + 1224, + 1689 + ], + "score": 0.85, + "latex": "\\times 2" + }, + { + "category_id": 13, + "poly": [ + 1193, + 1399, + 1247, + 1399, + 1247, + 1429, + 1193, + 1429 + ], + "score": 0.85, + "latex": "\\times 1 4" + }, + { + "category_id": 13, + "poly": [ + 750, + 1428, + 831, + 1428, + 831, + 1459, + 750, + 1459 + ], + "score": 0.84, + "latex": "( 7 8 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 1942, + 344, + 1942, + 344, + 1972, + 296, + 1972 + ], + "score": 0.69, + "latex": "2 5 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 1333, + 1016, + 1408, + 1016, + 1408, + 1048, + 1333, + 1048 + ], + "score": 0.58, + "latex": "\\mathrm { i } \\mathrm { P Q } +" + }, + { + "category_id": 13, + "poly": [ + 462, + 605, + 486, + 605, + 486, + 630, + 462, + 630 + ], + "score": 0.48, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 1081, + 733, + 1127, + 733, + 1127, + 761, + 1081, + 761 + ], + "score": 0.47, + "latex": "^ { 6 6 } { + } \\mathrm { S l }" + }, + { + "category_id": 13, + "poly": [ + 413, + 603, + 486, + 603, + 486, + 633, + 413, + 633 + ], + "score": 0.45, + "latex": "\\mathrm { i } \\mathrm { P Q } +" + }, + { + "category_id": 13, + "poly": [ + 296, + 1628, + 360, + 1628, + 360, + 1659, + 296, + 1659 + ], + "score": 0.42, + "latex": "\\mathrm { i P Q + }" + }, + { + "category_id": 13, + "poly": [ + 384, + 1720, + 476, + 1720, + 476, + 1749, + 384, + 1749 + ], + "score": 0.41, + "latex": "6 . 4 \\ : \\mathrm { P P L }" + }, + { + "category_id": 13, + "poly": [ + 1103, + 1399, + 1182, + 1399, + 1182, + 1429, + 1103, + 1429 + ], + "score": 0.4, + "latex": "3 . 3 { \\mathrm { M b } }" + }, + { + "category_id": 13, + "poly": [ + 1060, + 1972, + 1153, + 1972, + 1153, + 2002, + 1060, + 2002 + ], + "score": 0.4, + "latex": "0 . 2 \\ : \\mathrm { P P L }" + }, + { + "category_id": 13, + "poly": [ + 867, + 1750, + 962, + 1750, + 962, + 1779, + 867, + 1779 + ], + "score": 0.37, + "latex": "5 7 0 \\mathrm { M B }" + }, + { + "category_id": 13, + "poly": [ + 959, + 1628, + 1053, + 1628, + 1053, + 1658, + 959, + 1658 + ], + "score": 0.36, + "latex": "2 . 4 \\ : \\mathrm { P P L }" + }, + { + "category_id": 13, + "poly": [ + 342, + 1750, + 422, + 1750, + 422, + 1779, + 342, + 1779 + ], + "score": 0.36, + "latex": "1 0 \\mathbf { M B }" + }, + { + "category_id": 13, + "poly": [ + 1383, + 1018, + 1407, + 1018, + 1407, + 1044, + 1383, + 1044 + ], + "score": 0.31, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 385.0, + 1403.0, + 385.0, + 1403.0, + 423.0, + 294.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 417.0, + 332.0, + 417.0, + 332.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 417.0, + 1404.0, + 417.0, + 1404.0, + 449.0, + 393.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 294.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 956.0, + 881.0, + 956.0, + 881.0, + 991.0, + 294.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1816.0, + 1247.0, + 1816.0, + 1247.0, + 1863.0, + 291.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1015.0, + 1332.0, + 1015.0, + 1332.0, + 1050.0, + 295.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1046.0, + 1406.0, + 1046.0, + 1406.0, + 1081.0, + 296.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1077.0, + 1403.0, + 1077.0, + 1403.0, + 1111.0, + 295.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1109.0, + 1403.0, + 1109.0, + 1403.0, + 1140.0, + 295.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1136.0, + 1406.0, + 1136.0, + 1406.0, + 1174.0, + 292.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1163.0, + 721.0, + 1163.0, + 721.0, + 1206.0, + 292.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1163.0, + 1409.0, + 1163.0, + 1409.0, + 1206.0, + 838.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1198.0, + 891.0, + 1198.0, + 891.0, + 1235.0, + 295.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 700.0, + 1409.0, + 700.0, + 1409.0, + 738.0, + 295.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 731.0, + 1080.0, + 731.0, + 1080.0, + 769.0, + 293.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 731.0, + 1407.0, + 731.0, + 1407.0, + 769.0, + 1128.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 763.0, + 1407.0, + 763.0, + 1407.0, + 797.0, + 295.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 793.0, + 1405.0, + 793.0, + 1405.0, + 826.0, + 293.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 821.0, + 1406.0, + 821.0, + 1406.0, + 861.0, + 292.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 851.0, + 1405.0, + 851.0, + 1405.0, + 891.0, + 293.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 637.0, + 887.0, + 637.0, + 921.0, + 295.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1102.0, + 1396.0, + 1102.0, + 1433.0, + 293.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1396.0, + 1192.0, + 1396.0, + 1192.0, + 1433.0, + 1183.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 1396.0, + 1406.0, + 1396.0, + 1406.0, + 1433.0, + 1248.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1427.0, + 749.0, + 1427.0, + 749.0, + 1464.0, + 293.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 1427.0, + 914.0, + 1427.0, + 914.0, + 1464.0, + 832.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 1427.0, + 1406.0, + 1427.0, + 1406.0, + 1464.0, + 971.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1457.0, + 1406.0, + 1457.0, + 1406.0, + 1493.0, + 294.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1488.0, + 1407.0, + 1488.0, + 1407.0, + 1524.0, + 292.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1521.0, + 1408.0, + 1521.0, + 1408.0, + 1552.0, + 296.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1550.0, + 1348.0, + 1550.0, + 1348.0, + 1584.0, + 295.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1593.0, + 1406.0, + 1593.0, + 1406.0, + 1636.0, + 292.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1626.0, + 682.0, + 1626.0, + 682.0, + 1664.0, + 361.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1626.0, + 958.0, + 1626.0, + 958.0, + 1664.0, + 737.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1626.0, + 1409.0, + 1626.0, + 1409.0, + 1664.0, + 1054.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1656.0, + 1223.0, + 1656.0, + 1223.0, + 1697.0, + 292.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1656.0, + 1406.0, + 1656.0, + 1406.0, + 1697.0, + 1265.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1688.0, + 1273.0, + 1688.0, + 1273.0, + 1723.0, + 293.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1688.0, + 1407.0, + 1688.0, + 1407.0, + 1723.0, + 1328.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1717.0, + 383.0, + 1717.0, + 383.0, + 1755.0, + 295.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1717.0, + 1408.0, + 1717.0, + 1408.0, + 1755.0, + 477.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1750.0, + 341.0, + 1750.0, + 341.0, + 1782.0, + 295.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1750.0, + 866.0, + 1750.0, + 866.0, + 1782.0, + 423.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1750.0, + 1217.0, + 1750.0, + 1217.0, + 1782.0, + 963.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 511.0, + 1405.0, + 511.0, + 1405.0, + 544.0, + 297.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 540.0, + 1405.0, + 540.0, + 1405.0, + 575.0, + 293.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 570.0, + 1404.0, + 570.0, + 1404.0, + 607.0, + 293.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 599.0, + 412.0, + 599.0, + 412.0, + 639.0, + 292.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 599.0, + 1407.0, + 599.0, + 1407.0, + 639.0, + 487.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 631.0, + 966.0, + 631.0, + 966.0, + 667.0, + 293.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1879.0, + 1407.0, + 1879.0, + 1407.0, + 1920.0, + 294.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 345.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 1059.0, + 1973.0, + 1059.0, + 2006.0, + 293.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1973.0, + 1407.0, + 1973.0, + 1407.0, + 2006.0, + 1154.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1266.0, + 1405.0, + 1266.0, + 1405.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1405.0, + 1300.0, + 1405.0, + 1332.0, + 294.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1327.0, + 1385.0, + 1327.0, + 1385.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 408, + 1405, + 408, + 1405, + 592, + 298, + 592 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1402, + 229, + 1402, + 293, + 294, + 293 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 292, + 675, + 1409, + 675, + 1409, + 2022, + 292, + 2022 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 299, + 338, + 544, + 338, + 544, + 374, + 299, + 374 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 299, + 640, + 488, + 640, + 488, + 673, + 299, + 673 + ], + "score": 0.864 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.772 + }, + { + "category_id": 13, + "poly": [ + 659, + 1464, + 693, + 1464, + 693, + 1491, + 659, + 1491 + ], + "score": 0.83, + "latex": "+ 1" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 331.0, + 550.0, + 331.0, + 550.0, + 383.0, + 292.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 640.0, + 490.0, + 640.0, + 490.0, + 677.0, + 297.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 406.0, + 1405.0, + 406.0, + 1405.0, + 443.0, + 293.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 439.0, + 1406.0, + 439.0, + 1406.0, + 475.0, + 295.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 471.0, + 1405.0, + 471.0, + 1405.0, + 503.0, + 296.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 500.0, + 1405.0, + 500.0, + 1405.0, + 536.0, + 295.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 530.0, + 1407.0, + 530.0, + 1407.0, + 566.0, + 295.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 560.0, + 1183.0, + 560.0, + 1183.0, + 595.0, + 293.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1083.0, + 261.0, + 1083.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 691.0, + 1409.0, + 691.0, + 1409.0, + 734.0, + 292.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 721.0, + 1135.0, + 721.0, + 1135.0, + 764.0, + 318.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 776.0, + 1406.0, + 776.0, + 1406.0, + 816.0, + 292.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 807.0, + 776.0, + 807.0, + 776.0, + 845.0, + 320.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 860.0, + 1407.0, + 860.0, + 1407.0, + 902.0, + 294.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 893.0, + 1301.0, + 893.0, + 1301.0, + 931.0, + 321.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 946.0, + 1406.0, + 946.0, + 1406.0, + 988.0, + 292.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 979.0, + 894.0, + 979.0, + 894.0, + 1014.0, + 320.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1030.0, + 1407.0, + 1030.0, + 1407.0, + 1073.0, + 294.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1062.0, + 1059.0, + 1062.0, + 1059.0, + 1103.0, + 315.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1118.0, + 1407.0, + 1118.0, + 1407.0, + 1156.0, + 294.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1148.0, + 975.0, + 1148.0, + 975.0, + 1186.0, + 323.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1199.0, + 1407.0, + 1199.0, + 1407.0, + 1243.0, + 292.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1234.0, + 1406.0, + 1234.0, + 1406.0, + 1272.0, + 323.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1261.0, + 1077.0, + 1261.0, + 1077.0, + 1302.0, + 320.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1315.0, + 1409.0, + 1315.0, + 1409.0, + 1359.0, + 291.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1345.0, + 1410.0, + 1345.0, + 1410.0, + 1389.0, + 318.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1379.0, + 776.0, + 1379.0, + 776.0, + 1416.0, + 321.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1428.0, + 1404.0, + 1428.0, + 1404.0, + 1474.0, + 291.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1463.0, + 658.0, + 1463.0, + 658.0, + 1501.0, + 323.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1463.0, + 916.0, + 1463.0, + 916.0, + 1501.0, + 694.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1514.0, + 1407.0, + 1514.0, + 1407.0, + 1560.0, + 291.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1547.0, + 1048.0, + 1547.0, + 1048.0, + 1585.0, + 321.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1599.0, + 1407.0, + 1599.0, + 1407.0, + 1642.0, + 291.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1632.0, + 1404.0, + 1632.0, + 1404.0, + 1674.0, + 321.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1664.0, + 710.0, + 1664.0, + 710.0, + 1704.0, + 316.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1713.0, + 1406.0, + 1713.0, + 1406.0, + 1760.0, + 292.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1749.0, + 878.0, + 1749.0, + 878.0, + 1787.0, + 321.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1799.0, + 1406.0, + 1799.0, + 1406.0, + 1843.0, + 291.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1834.0, + 549.0, + 1834.0, + 549.0, + 1872.0, + 321.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1888.0, + 1406.0, + 1888.0, + 1406.0, + 1927.0, + 288.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1918.0, + 704.0, + 1918.0, + 704.0, + 1956.0, + 321.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1969.0, + 1406.0, + 1969.0, + 1406.0, + 2013.0, + 291.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 2000.0, + 1384.0, + 2000.0, + 1384.0, + 2036.0, + 317.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 292, + 1028, + 1402, + 1028, + 1402, + 1093, + 292, + 1093 + ], + "score": 0.802 + }, + { + "category_id": 1, + "poly": [ + 299, + 1110, + 1403, + 1110, + 1403, + 1205, + 299, + 1205 + ], + "score": 0.768 + }, + { + "category_id": 1, + "poly": [ + 296, + 946, + 1401, + 946, + 1401, + 1010, + 296, + 1010 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 296, + 781, + 1401, + 781, + 1401, + 848, + 296, + 848 + ], + "score": 0.754 + }, + { + "category_id": 1, + "poly": [ + 294, + 1222, + 1401, + 1222, + 1401, + 1288, + 294, + 1288 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 294, + 864, + 1402, + 864, + 1402, + 930, + 294, + 930 + ], + "score": 0.751 + }, + { + "category_id": 1, + "poly": [ + 293, + 1304, + 1402, + 1304, + 1402, + 1370, + 293, + 1370 + ], + "score": 0.739 + }, + { + "category_id": 1, + "poly": [ + 298, + 699, + 1401, + 699, + 1401, + 765, + 298, + 765 + ], + "score": 0.736 + }, + { + "category_id": 1, + "poly": [ + 298, + 1725, + 1402, + 1725, + 1402, + 1789, + 298, + 1789 + ], + "score": 0.735 + }, + { + "category_id": 1, + "poly": [ + 295, + 1529, + 1400, + 1529, + 1400, + 1595, + 295, + 1595 + ], + "score": 0.691 + }, + { + "category_id": 1, + "poly": [ + 297, + 1807, + 1400, + 1807, + 1400, + 1871, + 297, + 1871 + ], + "score": 0.688 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1401, + 229, + 1401, + 323, + 299, + 323 + ], + "score": 0.684 + }, + { + "category_id": 1, + "poly": [ + 297, + 1611, + 1408, + 1611, + 1408, + 1706, + 297, + 1706 + ], + "score": 0.682 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1399, + 1971, + 1399, + 2035, + 299, + 2035 + ], + "score": 0.672 + }, + { + "category_id": 1, + "poly": [ + 295, + 536, + 1402, + 536, + 1402, + 601, + 295, + 601 + ], + "score": 0.667 + }, + { + "category_id": 1, + "poly": [ + 295, + 617, + 1401, + 617, + 1401, + 683, + 295, + 683 + ], + "score": 0.663 + }, + { + "category_id": 1, + "poly": [ + 302, + 423, + 1400, + 423, + 1400, + 517, + 302, + 517 + ], + "score": 0.65 + }, + { + "category_id": 1, + "poly": [ + 298, + 1889, + 1401, + 1889, + 1401, + 1953, + 298, + 1953 + ], + "score": 0.65 + }, + { + "category_id": 1, + "poly": [ + 298, + 1387, + 1405, + 1387, + 1405, + 1513, + 298, + 1513 + ], + "score": 0.643 + }, + { + "category_id": 1, + "poly": [ + 292, + 342, + 1402, + 342, + 1402, + 406, + 292, + 406 + ], + "score": 0.626 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 830.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1030.0, + 1404.0, + 1030.0, + 1404.0, + 1066.0, + 297.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1061.0, + 708.0, + 1061.0, + 708.0, + 1094.0, + 320.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1106.0, + 1403.0, + 1106.0, + 1403.0, + 1149.0, + 293.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1138.0, + 1405.0, + 1138.0, + 1405.0, + 1181.0, + 322.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1172.0, + 758.0, + 1172.0, + 758.0, + 1206.0, + 323.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 940.0, + 1406.0, + 940.0, + 1406.0, + 987.0, + 293.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 977.0, + 617.0, + 977.0, + 617.0, + 1011.0, + 320.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 782.0, + 1405.0, + 782.0, + 1405.0, + 818.0, + 296.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 815.0, + 955.0, + 815.0, + 955.0, + 847.0, + 324.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1221.0, + 1404.0, + 1221.0, + 1404.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1255.0, + 703.0, + 1255.0, + 703.0, + 1287.0, + 321.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 863.0, + 1403.0, + 863.0, + 1403.0, + 900.0, + 297.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 894.0, + 1093.0, + 894.0, + 1093.0, + 931.0, + 323.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1306.0, + 1404.0, + 1306.0, + 1404.0, + 1342.0, + 296.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1336.0, + 1093.0, + 1336.0, + 1093.0, + 1371.0, + 324.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 700.0, + 1402.0, + 700.0, + 1402.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 731.0, + 824.0, + 731.0, + 824.0, + 766.0, + 323.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1725.0, + 1408.0, + 1725.0, + 1408.0, + 1761.0, + 296.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1757.0, + 1367.0, + 1757.0, + 1367.0, + 1790.0, + 322.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1530.0, + 1402.0, + 1530.0, + 1402.0, + 1566.0, + 295.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1561.0, + 472.0, + 1561.0, + 472.0, + 1594.0, + 323.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1807.0, + 1405.0, + 1807.0, + 1405.0, + 1843.0, + 296.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1838.0, + 912.0, + 1838.0, + 912.0, + 1872.0, + 322.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 266.0, + 297.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 295.0, + 323.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 291.0, + 622.0, + 291.0, + 622.0, + 324.0, + 322.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1611.0, + 1408.0, + 1611.0, + 1408.0, + 1649.0, + 295.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1640.0, + 1409.0, + 1640.0, + 1409.0, + 1680.0, + 319.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1668.0, + 399.0, + 1668.0, + 399.0, + 1709.0, + 320.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2002.0, + 1327.0, + 2002.0, + 1327.0, + 2038.0, + 324.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 536.0, + 1404.0, + 536.0, + 1404.0, + 572.0, + 298.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 569.0, + 1168.0, + 569.0, + 1168.0, + 602.0, + 322.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 616.0, + 1405.0, + 616.0, + 1405.0, + 655.0, + 294.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 651.0, + 1031.0, + 651.0, + 1031.0, + 682.0, + 323.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 421.0, + 1405.0, + 421.0, + 1405.0, + 462.0, + 296.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 454.0, + 1408.0, + 454.0, + 1408.0, + 491.0, + 319.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 480.0, + 397.0, + 480.0, + 397.0, + 521.0, + 319.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1885.0, + 1405.0, + 1885.0, + 1405.0, + 1929.0, + 293.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1919.0, + 612.0, + 1919.0, + 612.0, + 1953.0, + 322.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1384.0, + 1406.0, + 1384.0, + 1406.0, + 1425.0, + 293.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1454.0, + 322.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1451.0, + 1403.0, + 1451.0, + 1403.0, + 1484.0, + 322.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1481.0, + 807.0, + 1481.0, + 807.0, + 1514.0, + 321.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 341.0, + 1404.0, + 341.0, + 1404.0, + 380.0, + 295.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 373.0, + 997.0, + 373.0, + 997.0, + 407.0, + 321.0, + 407.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 862, + 2088, + 862, + 2113, + 835, + 2113 + ], + "score": 0.808 + }, + { + "category_id": 1, + "poly": [ + 290, + 1271, + 1401, + 1271, + 1401, + 1337, + 290, + 1337 + ], + "score": 0.523 + }, + { + "category_id": 1, + "poly": [ + 297, + 1010, + 1399, + 1010, + 1399, + 1076, + 297, + 1076 + ], + "score": 0.504 + }, + { + "category_id": 1, + "poly": [ + 296, + 923, + 1400, + 923, + 1400, + 988, + 296, + 988 + ], + "score": 0.503 + }, + { + "category_id": 1, + "poly": [ + 301, + 1444, + 1401, + 1444, + 1401, + 1540, + 301, + 1540 + ], + "score": 0.5 + }, + { + "category_id": 1, + "poly": [ + 296, + 1184, + 1403, + 1184, + 1403, + 1249, + 296, + 1249 + ], + "score": 0.499 + }, + { + "category_id": 1, + "poly": [ + 297, + 1562, + 1402, + 1562, + 1402, + 1657, + 297, + 1657 + ], + "score": 0.493 + }, + { + "category_id": 1, + "poly": [ + 292, + 1357, + 1401, + 1357, + 1401, + 1422, + 292, + 1422 + ], + "score": 0.487 + }, + { + "category_id": 1, + "poly": [ + 295, + 228, + 1276, + 228, + 1276, + 264, + 295, + 264 + ], + "score": 0.485 + }, + { + "category_id": 1, + "poly": [ + 299, + 1097, + 1401, + 1097, + 1401, + 1162, + 299, + 1162 + ], + "score": 0.48 + }, + { + "category_id": 1, + "poly": [ + 298, + 1679, + 1404, + 1679, + 1404, + 1774, + 298, + 1774 + ], + "score": 0.46 + }, + { + "category_id": 1, + "poly": [ + 292, + 749, + 1402, + 749, + 1402, + 816, + 292, + 816 + ], + "score": 0.459 + }, + { + "category_id": 1, + "poly": [ + 295, + 836, + 1401, + 836, + 1401, + 902, + 295, + 902 + ], + "score": 0.454 + }, + { + "category_id": 1, + "poly": [ + 298, + 662, + 1402, + 662, + 1402, + 728, + 298, + 728 + ], + "score": 0.419 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1399, + 1971, + 1399, + 2035, + 299, + 2035 + ], + "score": 0.419 + }, + { + "category_id": 1, + "poly": [ + 293, + 1883, + 1402, + 1883, + 1402, + 1950, + 293, + 1950 + ], + "score": 0.409 + }, + { + "category_id": 1, + "poly": [ + 292, + 1796, + 1402, + 1796, + 1402, + 1862, + 292, + 1862 + ], + "score": 0.397 + }, + { + "category_id": 1, + "poly": [ + 302, + 458, + 1400, + 458, + 1400, + 554, + 302, + 554 + ], + "score": 0.389 + }, + { + "category_id": 1, + "poly": [ + 296, + 577, + 1401, + 577, + 1401, + 642, + 296, + 642 + ], + "score": 0.379 + }, + { + "category_id": 1, + "poly": [ + 295, + 284, + 1400, + 284, + 1400, + 350, + 295, + 350 + ], + "score": 0.361 + }, + { + "category_id": 1, + "poly": [ + 293, + 370, + 1402, + 370, + 1402, + 437, + 293, + 437 + ], + "score": 0.36 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1274.0, + 1403.0, + 1274.0, + 1403.0, + 1307.0, + 297.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1303.0, + 1336.0, + 1303.0, + 1336.0, + 1336.0, + 321.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1009.0, + 1405.0, + 1009.0, + 1405.0, + 1047.0, + 293.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1042.0, + 1302.0, + 1042.0, + 1302.0, + 1077.0, + 322.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 921.0, + 1406.0, + 921.0, + 1406.0, + 962.0, + 292.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 955.0, + 861.0, + 955.0, + 861.0, + 988.0, + 325.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1444.0, + 1406.0, + 1444.0, + 1406.0, + 1482.0, + 295.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1474.0, + 1406.0, + 1474.0, + 1406.0, + 1515.0, + 320.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1508.0, + 807.0, + 1508.0, + 807.0, + 1539.0, + 323.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1184.0, + 1405.0, + 1184.0, + 1405.0, + 1220.0, + 296.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1216.0, + 758.0, + 1216.0, + 758.0, + 1249.0, + 324.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1559.0, + 1406.0, + 1559.0, + 1406.0, + 1601.0, + 294.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1590.0, + 1406.0, + 1590.0, + 1406.0, + 1633.0, + 319.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1624.0, + 734.0, + 1624.0, + 734.0, + 1658.0, + 324.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1355.0, + 1404.0, + 1355.0, + 1404.0, + 1396.0, + 292.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1391.0, + 622.0, + 1391.0, + 622.0, + 1422.0, + 321.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 1279.0, + 227.0, + 1279.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1095.0, + 1405.0, + 1095.0, + 1405.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1127.0, + 483.0, + 1127.0, + 483.0, + 1164.0, + 320.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1675.0, + 1408.0, + 1675.0, + 1408.0, + 1718.0, + 293.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1713.0, + 1404.0, + 1713.0, + 1404.0, + 1744.0, + 323.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1744.0, + 497.0, + 1744.0, + 497.0, + 1774.0, + 319.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 750.0, + 1404.0, + 750.0, + 1404.0, + 786.0, + 297.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 781.0, + 989.0, + 781.0, + 989.0, + 816.0, + 323.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 837.0, + 1403.0, + 837.0, + 1403.0, + 873.0, + 295.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 869.0, + 755.0, + 869.0, + 755.0, + 903.0, + 320.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 663.0, + 1406.0, + 663.0, + 1406.0, + 699.0, + 296.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 695.0, + 707.0, + 695.0, + 707.0, + 727.0, + 319.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1968.0, + 1404.0, + 1968.0, + 1404.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 1158.0, + 2003.0, + 1158.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1919.0, + 295.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1915.0, + 1326.0, + 1915.0, + 1326.0, + 1952.0, + 322.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1790.0, + 1407.0, + 1790.0, + 1407.0, + 1839.0, + 293.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1829.0, + 891.0, + 1829.0, + 891.0, + 1863.0, + 323.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 458.0, + 1403.0, + 458.0, + 1403.0, + 496.0, + 297.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 489.0, + 1405.0, + 489.0, + 1405.0, + 529.0, + 321.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 523.0, + 895.0, + 523.0, + 895.0, + 555.0, + 322.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 577.0, + 1405.0, + 577.0, + 1405.0, + 613.0, + 296.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 610.0, + 976.0, + 610.0, + 976.0, + 642.0, + 321.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 285.0, + 1405.0, + 285.0, + 1405.0, + 321.0, + 295.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 316.0, + 986.0, + 316.0, + 986.0, + 352.0, + 325.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 370.0, + 1404.0, + 370.0, + 1404.0, + 407.0, + 296.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 405.0, + 1162.0, + 405.0, + 1162.0, + 437.0, + 325.0, + 437.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 294, + 1142, + 1401, + 1142, + 1401, + 1207, + 294, + 1207 + ], + "score": 0.646 + }, + { + "category_id": 1, + "poly": [ + 300, + 1227, + 1403, + 1227, + 1403, + 1321, + 300, + 1321 + ], + "score": 0.62 + }, + { + "category_id": 1, + "poly": [ + 297, + 1343, + 1401, + 1343, + 1401, + 1409, + 297, + 1409 + ], + "score": 0.614 + }, + { + "category_id": 1, + "poly": [ + 300, + 1026, + 1401, + 1026, + 1401, + 1120, + 300, + 1120 + ], + "score": 0.602 + }, + { + "category_id": 1, + "poly": [ + 294, + 941, + 1402, + 941, + 1402, + 1006, + 294, + 1006 + ], + "score": 0.589 + }, + { + "category_id": 1, + "poly": [ + 298, + 1513, + 1403, + 1513, + 1403, + 1579, + 298, + 1579 + ], + "score": 0.579 + }, + { + "category_id": 1, + "poly": [ + 297, + 1428, + 1401, + 1428, + 1401, + 1493, + 297, + 1493 + ], + "score": 0.579 + }, + { + "category_id": 1, + "poly": [ + 300, + 1598, + 1404, + 1598, + 1404, + 1693, + 300, + 1693 + ], + "score": 0.577 + }, + { + "category_id": 1, + "poly": [ + 295, + 1714, + 1401, + 1714, + 1401, + 1780, + 295, + 1780 + ], + "score": 0.533 + }, + { + "category_id": 1, + "poly": [ + 299, + 685, + 1399, + 685, + 1399, + 750, + 299, + 750 + ], + "score": 0.528 + }, + { + "category_id": 1, + "poly": [ + 300, + 855, + 1401, + 855, + 1401, + 920, + 300, + 920 + ], + "score": 0.516 + }, + { + "category_id": 1, + "poly": [ + 296, + 600, + 1401, + 600, + 1401, + 665, + 296, + 665 + ], + "score": 0.506 + }, + { + "category_id": 1, + "poly": [ + 300, + 770, + 1401, + 770, + 1401, + 837, + 300, + 837 + ], + "score": 0.498 + }, + { + "category_id": 1, + "poly": [ + 302, + 229, + 1401, + 229, + 1401, + 323, + 302, + 323 + ], + "score": 0.472 + }, + { + "category_id": 1, + "poly": [ + 296, + 514, + 1400, + 514, + 1400, + 579, + 296, + 579 + ], + "score": 0.448 + }, + { + "category_id": 1, + "poly": [ + 293, + 1799, + 1401, + 1799, + 1401, + 1865, + 293, + 1865 + ], + "score": 0.448 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1398, + 1972, + 1398, + 2034, + 299, + 2034 + ], + "score": 0.448 + }, + { + "category_id": 1, + "poly": [ + 291, + 1886, + 1401, + 1886, + 1401, + 1950, + 291, + 1950 + ], + "score": 0.428 + }, + { + "category_id": 1, + "poly": [ + 297, + 430, + 1399, + 430, + 1399, + 494, + 297, + 494 + ], + "score": 0.393 + }, + { + "category_id": 1, + "poly": [ + 293, + 344, + 1402, + 344, + 1402, + 409, + 293, + 409 + ], + "score": 0.386 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1142.0, + 1404.0, + 1142.0, + 1404.0, + 1178.0, + 296.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1172.0, + 511.0, + 1172.0, + 511.0, + 1207.0, + 321.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1229.0, + 1407.0, + 1229.0, + 1407.0, + 1263.0, + 295.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1259.0, + 1408.0, + 1259.0, + 1408.0, + 1293.0, + 321.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1284.0, + 400.0, + 1284.0, + 400.0, + 1325.0, + 319.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1341.0, + 1405.0, + 1341.0, + 1405.0, + 1381.0, + 295.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1375.0, + 463.0, + 1375.0, + 463.0, + 1408.0, + 321.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1028.0, + 1405.0, + 1028.0, + 1405.0, + 1062.0, + 296.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1058.0, + 1408.0, + 1058.0, + 1408.0, + 1092.0, + 321.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1086.0, + 399.0, + 1086.0, + 399.0, + 1123.0, + 320.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 943.0, + 1407.0, + 943.0, + 1407.0, + 979.0, + 296.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 975.0, + 1030.0, + 975.0, + 1030.0, + 1007.0, + 323.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 1404.0, + 1511.0, + 1404.0, + 1551.0, + 293.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1547.0, + 844.0, + 1547.0, + 844.0, + 1578.0, + 322.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1431.0, + 1403.0, + 1431.0, + 1403.0, + 1463.0, + 297.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1459.0, + 1112.0, + 1459.0, + 1112.0, + 1495.0, + 324.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1598.0, + 1408.0, + 1598.0, + 1408.0, + 1636.0, + 295.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1630.0, + 1409.0, + 1630.0, + 1409.0, + 1667.0, + 323.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1656.0, + 406.0, + 1656.0, + 406.0, + 1696.0, + 320.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1715.0, + 1405.0, + 1715.0, + 1405.0, + 1751.0, + 297.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1749.0, + 629.0, + 1749.0, + 629.0, + 1779.0, + 321.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 687.0, + 1403.0, + 687.0, + 1403.0, + 723.0, + 295.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 718.0, + 1031.0, + 718.0, + 1031.0, + 751.0, + 323.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 854.0, + 1406.0, + 854.0, + 1406.0, + 894.0, + 295.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 889.0, + 657.0, + 889.0, + 657.0, + 921.0, + 323.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 600.0, + 1405.0, + 600.0, + 1405.0, + 638.0, + 294.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 633.0, + 1270.0, + 633.0, + 1270.0, + 666.0, + 321.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 767.0, + 1406.0, + 767.0, + 1406.0, + 811.0, + 294.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 804.0, + 1313.0, + 804.0, + 1313.0, + 837.0, + 323.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 297.0, + 322.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 292.0, + 720.0, + 292.0, + 720.0, + 326.0, + 323.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 512.0, + 1404.0, + 512.0, + 1404.0, + 553.0, + 294.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 547.0, + 610.0, + 547.0, + 610.0, + 577.0, + 322.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1798.0, + 1404.0, + 1798.0, + 1404.0, + 1837.0, + 295.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1832.0, + 1233.0, + 1832.0, + 1233.0, + 1867.0, + 322.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2003.0, + 881.0, + 2003.0, + 881.0, + 2036.0, + 321.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1886.0, + 1403.0, + 1886.0, + 1403.0, + 1922.0, + 295.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1919.0, + 938.0, + 1919.0, + 938.0, + 1949.0, + 323.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 430.0, + 1403.0, + 430.0, + 1403.0, + 462.0, + 298.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 461.0, + 1242.0, + 461.0, + 1242.0, + 495.0, + 322.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 344.0, + 1405.0, + 344.0, + 1405.0, + 380.0, + 298.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 375.0, + 860.0, + 375.0, + 860.0, + 411.0, + 322.0, + 411.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 814, + 75, + 814, + 105, + 300, + 105 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2112, + 835, + 2112 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 300, + 312, + 1402, + 312, + 1402, + 376, + 300, + 376 + ], + "score": 0.819 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1403, + 229, + 1403, + 293, + 296, + 293 + ], + "score": 0.809 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 311.0, + 1406.0, + 311.0, + 1406.0, + 349.0, + 294.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 343.0, + 1300.0, + 343.0, + 1300.0, + 379.0, + 321.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 261.0, + 1029.0, + 261.0, + 1029.0, + 294.0, + 321.0, + 294.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1509, + 1405, + 1509, + 1405, + 1786, + 298, + 1786 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 430, + 223, + 1269, + 223, + 1269, + 427, + 430, + 427 + ], + "score": 0.976, + "html": "
SettingModelCompressionTop-1 Accuracy
Small BlocksStock et al. (2019)19x73.8
Quant-Noise19x74.3
Large BlocksStock et al. (2019)32x68.2
Quant-Noise32x68.8
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1403, + 1881, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 311, + 586, + 1381, + 586, + 1381, + 860, + 311, + 860 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 1163, + 1403, + 1163, + 1403, + 1257, + 298, + 1257 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 295, + 1352, + 1404, + 1352, + 1404, + 1414, + 295, + 1414 + ], + "score": 0.952 + }, + { + "category_id": 4, + "poly": [ + 297, + 887, + 1401, + 887, + 1401, + 983, + 297, + 983 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 301, + 1452, + 651, + 1452, + 651, + 1483, + 301, + 1483 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 299, + 1823, + 974, + 1823, + 974, + 1854, + 299, + 1854 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 299, + 1106, + 961, + 1106, + 961, + 1137, + 299, + 1137 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 812, + 76, + 812, + 105, + 298, + 105 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 300, + 1036, + 500, + 1036, + 500, + 1071, + 300, + 1071 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 300, + 1295, + 507, + 1295, + 507, + 1325, + 300, + 1325 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 863, + 2089, + 863, + 2112, + 836, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 4, + "poly": [ + 297, + 447, + 1404, + 447, + 1404, + 542, + 297, + 542 + ], + "score": 0.802 + }, + { + "category_id": 4, + "poly": [ + 296, + 887, + 1402, + 887, + 1402, + 983, + 296, + 983 + ], + "score": 0.358 + }, + { + "category_id": 4, + "poly": [ + 296, + 447, + 1406, + 447, + 1406, + 543, + 296, + 543 + ], + "score": 0.117 + }, + { + "category_id": 13, + "poly": [ + 829, + 1575, + 897, + 1575, + 897, + 1605, + 829, + 1605 + ], + "score": 0.88, + "latex": "\\varphi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 968, + 1606, + 1035, + 1606, + 1035, + 1635, + 968, + 1635 + ], + "score": 0.88, + "latex": "\\varphi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 749, + 1946, + 817, + 1946, + 817, + 1976, + 749, + 1976 + ], + "score": 0.88, + "latex": "\\varphi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 899, + 954, + 967, + 954, + 967, + 984, + 899, + 984 + ], + "score": 0.84, + "latex": "\\varphi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 586.0, + 619.0, + 586.0, + 619.0, + 620.0, + 569.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 584.0, + 1186.0, + 584.0, + 1186.0, + 624.0, + 1132.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 627.0, + 380.0, + 627.0, + 380.0, + 728.0, + 339.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 634.0, + 594.0, + 634.0, + 594.0, + 668.0, + 397.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 635.0, + 800.0, + 635.0, + 800.0, + 662.0, + 781.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 620.0, + 948.0, + 620.0, + 948.0, + 765.0, + 901.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 634.0, + 1158.0, + 634.0, + 1158.0, + 668.0, + 963.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 644.0, + 1349.0, + 644.0, + 1349.0, + 666.0, + 1326.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 662.0, + 344.0, + 662.0, + 344.0, + 765.0, + 309.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 667.0, + 549.0, + 667.0, + 549.0, + 697.0, + 398.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 654.0, + 802.0, + 654.0, + 802.0, + 688.0, + 780.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 662.0, + 908.0, + 662.0, + 908.0, + 765.0, + 873.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 667.0, + 1115.0, + 667.0, + 1115.0, + 697.0, + 965.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 702.0, + 408.0, + 702.0, + 408.0, + 721.0, + 388.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 696.0, + 1277.0, + 696.0, + 1277.0, + 716.0, + 1256.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 715.0, + 378.0, + 715.0, + 378.0, + 794.0, + 341.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 759.0, + 407.0, + 759.0, + 407.0, + 780.0, + 387.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 735.0, + 485.0, + 735.0, + 485.0, + 799.0, + 467.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 723.0, + 566.0, + 723.0, + 566.0, + 800.0, + 544.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 734.0, + 642.0, + 734.0, + 642.0, + 786.0, + 623.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 739.0, + 729.0, + 739.0, + 729.0, + 787.0, + 700.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 751.0, + 942.0, + 751.0, + 942.0, + 788.0, + 907.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 718.0, + 992.0, + 718.0, + 992.0, + 797.0, + 970.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 781.0, + 1019.0, + 781.0, + 1019.0, + 791.0, + 1009.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 749.0, + 1136.0, + 749.0, + 1136.0, + 800.0, + 1110.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 741.0, + 1207.0, + 741.0, + 1207.0, + 781.0, + 1185.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 763.0, + 1273.0, + 763.0, + 1273.0, + 773.0, + 1260.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 738.0, + 1348.0, + 738.0, + 1348.0, + 759.0, + 1329.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 800.0, + 421.0, + 800.0, + 421.0, + 832.0, + 377.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 800.0, + 499.0, + 800.0, + 499.0, + 832.0, + 455.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 800.0, + 577.0, + 800.0, + 577.0, + 832.0, + 534.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 800.0, + 653.0, + 800.0, + 653.0, + 832.0, + 611.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 800.0, + 733.0, + 800.0, + 733.0, + 832.0, + 690.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 800.0, + 811.0, + 800.0, + 811.0, + 832.0, + 769.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 802.0, + 1001.0, + 802.0, + 1001.0, + 829.0, + 961.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 800.0, + 1071.0, + 800.0, + 1071.0, + 831.0, + 1007.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 800.0, + 1145.0, + 800.0, + 1145.0, + 832.0, + 1101.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 800.0, + 1216.0, + 800.0, + 1216.0, + 831.0, + 1173.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 800.0, + 1358.0, + 800.0, + 1358.0, + 832.0, + 1245.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 832.0, + 672.0, + 832.0, + 672.0, + 862.0, + 518.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 832.0, + 1237.0, + 832.0, + 1237.0, + 862.0, + 1082.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 1403.0, + 887.0, + 1403.0, + 922.0, + 295.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 917.0, + 1405.0, + 917.0, + 1405.0, + 955.0, + 293.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 941.0, + 898.0, + 941.0, + 898.0, + 994.0, + 291.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 941.0, + 981.0, + 941.0, + 981.0, + 994.0, + 968.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1450.0, + 654.0, + 1450.0, + 654.0, + 1486.0, + 295.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 977.0, + 1821.0, + 977.0, + 1858.0, + 294.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1105.0, + 963.0, + 1105.0, + 963.0, + 1141.0, + 295.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 70.0, + 816.0, + 70.0, + 816.0, + 110.0, + 293.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1031.0, + 508.0, + 1031.0, + 508.0, + 1079.0, + 291.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1291.0, + 510.0, + 1291.0, + 510.0, + 1332.0, + 293.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 448.0, + 1404.0, + 448.0, + 1404.0, + 482.0, + 294.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 480.0, + 1407.0, + 480.0, + 1407.0, + 514.0, + 295.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 508.0, + 989.0, + 508.0, + 989.0, + 547.0, + 296.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 1404.0, + 887.0, + 1404.0, + 922.0, + 295.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 918.0, + 1404.0, + 918.0, + 1404.0, + 952.0, + 295.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 941.0, + 898.0, + 941.0, + 898.0, + 994.0, + 292.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 941.0, + 981.0, + 941.0, + 981.0, + 994.0, + 968.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 448.0, + 1404.0, + 448.0, + 1404.0, + 483.0, + 294.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 479.0, + 1407.0, + 479.0, + 1407.0, + 513.0, + 295.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 508.0, + 991.0, + 508.0, + 991.0, + 548.0, + 295.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1510.0, + 1403.0, + 1510.0, + 1403.0, + 1543.0, + 297.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1541.0, + 1405.0, + 1541.0, + 1405.0, + 1574.0, + 296.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1570.0, + 828.0, + 1570.0, + 828.0, + 1611.0, + 293.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1570.0, + 1406.0, + 1570.0, + 1406.0, + 1611.0, + 898.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1596.0, + 967.0, + 1596.0, + 967.0, + 1644.0, + 291.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 1596.0, + 1407.0, + 1596.0, + 1407.0, + 1644.0, + 1036.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1631.0, + 1406.0, + 1631.0, + 1406.0, + 1667.0, + 293.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1662.0, + 1406.0, + 1662.0, + 1406.0, + 1699.0, + 296.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1693.0, + 1406.0, + 1693.0, + 1406.0, + 1726.0, + 295.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1719.0, + 1405.0, + 1719.0, + 1405.0, + 1760.0, + 293.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1751.0, + 476.0, + 1751.0, + 476.0, + 1793.0, + 294.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1915.0, + 297.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1939.0, + 748.0, + 1939.0, + 748.0, + 1983.0, + 290.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1939.0, + 1408.0, + 1939.0, + 1408.0, + 1983.0, + 818.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1403.0, + 2002.0, + 1403.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1164.0, + 1405.0, + 1164.0, + 1405.0, + 1198.0, + 294.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1193.0, + 1407.0, + 1193.0, + 1407.0, + 1231.0, + 294.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1226.0, + 980.0, + 1226.0, + 980.0, + 1260.0, + 296.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1350.0, + 1407.0, + 1350.0, + 1407.0, + 1389.0, + 293.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1382.0, + 695.0, + 1382.0, + 695.0, + 1421.0, + 294.0, + 1421.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1506, + 1405, + 1506, + 1405, + 1814, + 297, + 1814 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 485, + 1405, + 485, + 1405, + 702, + 298, + 702 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1405, + 1850, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1283, + 1404, + 1283, + 1404, + 1406, + 299, + 1406 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 800, + 1405, + 800, + 1405, + 953, + 298, + 953 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1137, + 1405, + 1137, + 1405, + 1261, + 298, + 1261 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 296, + 224, + 809, + 224, + 809, + 426, + 296, + 426 + ], + "score": 0.976, + "html": "
NoiseBlocksPPLQuant PPL
PQSubvectors18.321.1
PQClusters18.321.2
proxySubvectors18.321.0
proxyClusters18.421.1
" + }, + { + "category_id": 1, + "poly": [ + 297, + 991, + 1404, + 991, + 1404, + 1113, + 297, + 1113 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 301, + 1448, + 599, + 1448, + 599, + 1480, + 301, + 1480 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 300, + 740, + 667, + 740, + 667, + 772, + 300, + 772 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.85 + }, + { + "category_id": 6, + "poly": [ + 830, + 223, + 1368, + 223, + 1368, + 376, + 830, + 376 + ], + "score": 0.608 + }, + { + "category_id": 4, + "poly": [ + 830, + 223, + 1368, + 223, + 1368, + 376, + 830, + 376 + ], + "score": 0.476 + }, + { + "category_id": 13, + "poly": [ + 1243, + 286, + 1310, + 286, + 1310, + 320, + 1243, + 320 + ], + "score": 0.91, + "latex": "\\phi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 764, + 672, + 832, + 672, + 832, + 703, + 764, + 703 + ], + "score": 0.9, + "latex": "\\varphi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 947, + 286, + 993, + 286, + 993, + 318, + 947, + 318 + ], + "score": 0.88, + "latex": "\\phi _ { \\mathrm { P Q } }" + }, + { + "category_id": 13, + "poly": [ + 502, + 1345, + 577, + 1345, + 577, + 1376, + 502, + 1376 + ], + "score": 0.87, + "latex": "8 1 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1119, + 1345, + 1195, + 1345, + 1195, + 1376, + 1119, + 1376 + ], + "score": 0.86, + "latex": "8 1 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 785, + 612, + 853, + 612, + 853, + 642, + 785, + 642 + ], + "score": 0.85, + "latex": "\\varphi _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 653, + 1172, + 678, + 1172, + 678, + 1196, + 653, + 1196 + ], + "score": 0.69, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1444.0, + 605.0, + 1444.0, + 605.0, + 1486.0, + 294.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 737.0, + 671.0, + 737.0, + 671.0, + 777.0, + 295.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 222.0, + 1368.0, + 222.0, + 1368.0, + 258.0, + 829.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 251.0, + 1370.0, + 251.0, + 1370.0, + 289.0, + 830.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 287.0, + 946.0, + 287.0, + 946.0, + 320.0, + 829.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 287.0, + 1242.0, + 287.0, + 1242.0, + 320.0, + 994.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 287.0, + 1367.0, + 287.0, + 1367.0, + 320.0, + 1311.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 314.0, + 1366.0, + 314.0, + 1366.0, + 347.0, + 830.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 345.0, + 1082.0, + 345.0, + 1082.0, + 378.0, + 831.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 222.0, + 1368.0, + 222.0, + 1368.0, + 258.0, + 829.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 251.0, + 1370.0, + 251.0, + 1370.0, + 289.0, + 830.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 287.0, + 946.0, + 287.0, + 946.0, + 320.0, + 829.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 287.0, + 1242.0, + 287.0, + 1242.0, + 320.0, + 994.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 287.0, + 1367.0, + 287.0, + 1367.0, + 320.0, + 1311.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 314.0, + 1366.0, + 314.0, + 1366.0, + 347.0, + 830.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 345.0, + 1082.0, + 345.0, + 1082.0, + 378.0, + 831.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1508.0, + 1407.0, + 1508.0, + 1407.0, + 1540.0, + 295.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1534.0, + 1407.0, + 1534.0, + 1407.0, + 1573.0, + 294.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1569.0, + 1406.0, + 1569.0, + 1406.0, + 1601.0, + 296.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1598.0, + 1407.0, + 1598.0, + 1407.0, + 1633.0, + 295.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1627.0, + 1408.0, + 1627.0, + 1408.0, + 1664.0, + 294.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1657.0, + 1407.0, + 1657.0, + 1407.0, + 1696.0, + 294.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1687.0, + 1407.0, + 1687.0, + 1407.0, + 1728.0, + 291.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1720.0, + 1407.0, + 1720.0, + 1407.0, + 1756.0, + 295.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1752.0, + 1403.0, + 1752.0, + 1403.0, + 1784.0, + 296.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1780.0, + 1408.0, + 1780.0, + 1408.0, + 1816.0, + 295.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 487.0, + 1405.0, + 487.0, + 1405.0, + 522.0, + 295.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 515.0, + 1406.0, + 515.0, + 1406.0, + 551.0, + 293.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 548.0, + 1405.0, + 548.0, + 1405.0, + 580.0, + 293.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 576.0, + 1406.0, + 576.0, + 1406.0, + 612.0, + 293.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 606.0, + 784.0, + 606.0, + 784.0, + 648.0, + 291.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 606.0, + 1410.0, + 606.0, + 1410.0, + 648.0, + 854.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 637.0, + 1406.0, + 637.0, + 1406.0, + 674.0, + 292.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 668.0, + 763.0, + 668.0, + 763.0, + 706.0, + 293.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 668.0, + 1207.0, + 668.0, + 1207.0, + 706.0, + 833.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1851.0, + 1407.0, + 1851.0, + 1407.0, + 1882.0, + 296.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1879.0, + 1408.0, + 1879.0, + 1408.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1913.0, + 1406.0, + 1913.0, + 1406.0, + 1945.0, + 296.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1406.0, + 1969.0, + 1406.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2002.0, + 1407.0, + 2002.0, + 1407.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1281.0, + 1405.0, + 1281.0, + 1405.0, + 1320.0, + 293.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1312.0, + 1408.0, + 1312.0, + 1408.0, + 1350.0, + 293.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1345.0, + 501.0, + 1345.0, + 501.0, + 1381.0, + 293.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1345.0, + 1118.0, + 1345.0, + 1118.0, + 1381.0, + 578.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1345.0, + 1408.0, + 1345.0, + 1408.0, + 1381.0, + 1196.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1374.0, + 421.0, + 1374.0, + 421.0, + 1408.0, + 292.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 834.0, + 295.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 829.0, + 1408.0, + 829.0, + 1408.0, + 866.0, + 295.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 857.0, + 1409.0, + 857.0, + 1409.0, + 898.0, + 292.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 888.0, + 1406.0, + 888.0, + 1406.0, + 928.0, + 293.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 921.0, + 1221.0, + 921.0, + 1221.0, + 956.0, + 293.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1135.0, + 1410.0, + 1135.0, + 1410.0, + 1172.0, + 292.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1166.0, + 652.0, + 1166.0, + 652.0, + 1204.0, + 293.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1166.0, + 1407.0, + 1166.0, + 1407.0, + 1204.0, + 679.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1199.0, + 1408.0, + 1199.0, + 1408.0, + 1234.0, + 292.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1227.0, + 1371.0, + 1227.0, + 1371.0, + 1267.0, + 293.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 988.0, + 1408.0, + 988.0, + 1408.0, + 1027.0, + 292.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1022.0, + 1406.0, + 1022.0, + 1406.0, + 1054.0, + 295.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1051.0, + 1404.0, + 1051.0, + 1404.0, + 1089.0, + 292.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1083.0, + 536.0, + 1083.0, + 536.0, + 1116.0, + 295.0, + 1116.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 976, + 1405, + 976, + 1405, + 1250, + 297, + 1250 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 556, + 223, + 1137, + 223, + 1137, + 551, + 556, + 551 + ], + "score": 0.978, + "html": "
ModelMBPPL
Trans XL Large (Dai et al., 2019) Compressive Trans (Rae et al.,2019) GCNN (Dauphin et al., 2017) 4 Layer QRNN (Bradbury et al., 2016)970 970 87018.3 17.1 37.2 33.0
Trans XL Base (Dai et al.,2019) Persis Mem (Sukhbaatar etal.,2019b) Tensorized core-2 (Ma et al.,2019)570 506 32524.0 20.6
18.9
Quant-Noise3820.7
Quant-Noise + Share + Prune
10
24.2
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1346, + 1405, + 1346, + 1405, + 1530, + 298, + 1530 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 789, + 1404, + 789, + 1404, + 943, + 299, + 943 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1754, + 1403, + 1754, + 1403, + 1847, + 299, + 1847 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1881, + 1402, + 1881, + 1402, + 2033, + 300, + 2033 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1626, + 1403, + 1626, + 1403, + 1720, + 299, + 1720 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 711, + 1400, + 711, + 1400, + 773, + 300, + 773 + ], + "score": 0.951 + }, + { + "category_id": 0, + "poly": [ + 302, + 1568, + 714, + 1568, + 714, + 1601, + 302, + 1601 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 303, + 1289, + 765, + 1289, + 765, + 1321, + 303, + 1321 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 863, + 2089, + 863, + 2112, + 836, + 2112 + ], + "score": 0.84 + }, + { + "category_id": 1, + "poly": [ + 297, + 589, + 1403, + 589, + 1403, + 650, + 297, + 650 + ], + "score": 0.618 + }, + { + "category_id": 6, + "poly": [ + 297, + 589, + 1403, + 589, + 1403, + 650, + 297, + 650 + ], + "score": 0.334 + }, + { + "category_id": 13, + "poly": [ + 1338, + 1069, + 1404, + 1069, + 1404, + 1098, + 1338, + 1098 + ], + "score": 0.9, + "latex": "1 \\times 1" + }, + { + "category_id": 13, + "poly": [ + 649, + 1100, + 717, + 1100, + 717, + 1129, + 649, + 1129 + ], + "score": 0.89, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 371, + 1912, + 436, + 1912, + 436, + 1941, + 371, + 1941 + ], + "score": 0.89, + "latex": "1 \\times 1" + }, + { + "category_id": 13, + "poly": [ + 1189, + 1912, + 1254, + 1912, + 1254, + 1941, + 1189, + 1941 + ], + "score": 0.89, + "latex": "5 \\times 5" + }, + { + "category_id": 13, + "poly": [ + 1072, + 1100, + 1139, + 1100, + 1139, + 1129, + 1072, + 1129 + ], + "score": 0.89, + "latex": "5 \\times 5" + }, + { + "category_id": 13, + "poly": [ + 780, + 1913, + 845, + 1913, + 845, + 1941, + 780, + 1941 + ], + "score": 0.88, + "latex": "3 \\times 3" + }, + { + "category_id": 13, + "poly": [ + 914, + 1441, + 949, + 1441, + 949, + 1470, + 914, + 1470 + ], + "score": 0.88, + "latex": "L _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 366, + 1444, + 384, + 1444, + 384, + 1467, + 366, + 1467 + ], + "score": 0.76, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1381, + 1444, + 1398, + 1444, + 1398, + 1467, + 1381, + 1467 + ], + "score": 0.76, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1311, + 1445, + 1329, + 1445, + 1329, + 1468, + 1311, + 1468 + ], + "score": 0.54, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 298, + 1445, + 314, + 1445, + 314, + 1467, + 298, + 1467 + ], + "score": 0.47, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1568.0, + 719.0, + 1568.0, + 719.0, + 1604.0, + 296.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1284.0, + 772.0, + 1284.0, + 772.0, + 1329.0, + 294.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 584.0, + 1408.0, + 584.0, + 1408.0, + 626.0, + 291.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 619.0, + 596.0, + 619.0, + 596.0, + 655.0, + 293.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 978.0, + 1406.0, + 978.0, + 1406.0, + 1010.0, + 295.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1005.0, + 1408.0, + 1005.0, + 1408.0, + 1042.0, + 294.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1038.0, + 1407.0, + 1038.0, + 1407.0, + 1073.0, + 292.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1069.0, + 1337.0, + 1069.0, + 1337.0, + 1101.0, + 296.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1100.0, + 648.0, + 1100.0, + 648.0, + 1131.0, + 294.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1100.0, + 1071.0, + 1100.0, + 1071.0, + 1131.0, + 718.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1100.0, + 1406.0, + 1100.0, + 1406.0, + 1131.0, + 1140.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1129.0, + 1408.0, + 1129.0, + 1408.0, + 1165.0, + 295.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1158.0, + 1403.0, + 1158.0, + 1403.0, + 1195.0, + 295.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1190.0, + 1406.0, + 1190.0, + 1406.0, + 1226.0, + 294.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1219.0, + 412.0, + 1219.0, + 412.0, + 1253.0, + 294.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1344.0, + 1403.0, + 1344.0, + 1403.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1377.0, + 1405.0, + 1377.0, + 1405.0, + 1414.0, + 295.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1405.0, + 1405.0, + 1405.0, + 1405.0, + 1444.0, + 293.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1438.0, + 297.0, + 1438.0, + 297.0, + 1475.0, + 292.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1438.0, + 365.0, + 1438.0, + 365.0, + 1475.0, + 315.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1438.0, + 913.0, + 1438.0, + 913.0, + 1475.0, + 385.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1438.0, + 1310.0, + 1438.0, + 1310.0, + 1475.0, + 950.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 1438.0, + 1380.0, + 1438.0, + 1380.0, + 1475.0, + 1330.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1438.0, + 1408.0, + 1438.0, + 1408.0, + 1475.0, + 1399.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1468.0, + 1405.0, + 1468.0, + 1405.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1499.0, + 405.0, + 1499.0, + 405.0, + 1532.0, + 293.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 788.0, + 1405.0, + 788.0, + 1405.0, + 825.0, + 294.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 817.0, + 1404.0, + 817.0, + 1404.0, + 854.0, + 295.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 846.0, + 1405.0, + 846.0, + 1405.0, + 890.0, + 292.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 875.0, + 1409.0, + 875.0, + 1409.0, + 920.0, + 293.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 911.0, + 1072.0, + 911.0, + 1072.0, + 948.0, + 294.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1753.0, + 1403.0, + 1753.0, + 1403.0, + 1787.0, + 295.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1782.0, + 1404.0, + 1782.0, + 1404.0, + 1819.0, + 294.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1816.0, + 975.0, + 1816.0, + 975.0, + 1850.0, + 295.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1878.0, + 1407.0, + 1878.0, + 1407.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 370.0, + 1910.0, + 370.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1910.0, + 779.0, + 1910.0, + 779.0, + 1945.0, + 437.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1910.0, + 1188.0, + 1910.0, + 1188.0, + 1945.0, + 846.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1945.0, + 1255.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1999.0, + 408.0, + 1999.0, + 408.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1627.0, + 1405.0, + 1627.0, + 1405.0, + 1661.0, + 295.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1658.0, + 1403.0, + 1658.0, + 1403.0, + 1689.0, + 297.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1688.0, + 1241.0, + 1688.0, + 1241.0, + 1725.0, + 295.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 710.0, + 1404.0, + 710.0, + 1404.0, + 748.0, + 294.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 743.0, + 1045.0, + 743.0, + 1045.0, + 776.0, + 295.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 584.0, + 1408.0, + 584.0, + 1408.0, + 626.0, + 291.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 619.0, + 596.0, + 619.0, + 596.0, + 655.0, + 293.0, + 655.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 569, + 772, + 1123, + 772, + 1123, + 1141, + 569, + 1141 + ], + "score": 0.98, + "html": "
ModelMBAcc.
EfficientNet-B7 (Tan &Le,2019) ResNet-50 (He et al., 2015) DenseNet-169 (Huang et al., 2018) EfficientNet-BO (Tan &Le,2019) MobileNet-v2 (Sandler et al., 2018)260 97.5 53.4 20.2 13.484.4 76.1 76.2 77.3 71.9
Shufflenet-v2 ×1 (Ma et al.,2018) HAQ 4 bits (Wang et al., 2018) iPQ ResNet-50 (Stock et al.,2019)8.7 12.469.4 76.2
Quant-Noise5.09 3.376.1 80.0
Quant-Noise + Share + Prune2.377.8
" + }, + { + "category_id": 5, + "poly": [ + 543, + 224, + 1150, + 224, + 1150, + 579, + 543, + 579 + ], + "score": 0.98, + "html": "
ModelMBMNLI
RoBERTa Base +LD (Fan et al., 2019)48084.8
BERT Base (Devlin et al.,2018)42084.4
PreTrained Distil (Turc et al., 2019)25782.5
DistilBERT (Sanh et al.,2019b)25081.8
MobileBERT* (Sun et al.)9684.4
TinyBERTt (Jiao et al.,2019)5582.8
ALBERT Base (Lan et al.,2019)4581.6
AdaBERTt (Chen et al.,2020)3681.6
Quant-Noise3883.6
Quant-Noise+ Share +Prune1482.5
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1358, + 1405, + 1358, + 1405, + 1513, + 297, + 1513 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 295, + 1608, + 1401, + 1608, + 1401, + 1672, + 295, + 1672 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1403, + 1971, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 296, + 1820, + 1401, + 1820, + 1401, + 1883, + 296, + 1883 + ], + "score": 0.935 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 299, + 1919, + 880, + 1919, + 880, + 1950, + 299, + 1950 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 299, + 1300, + 893, + 1300, + 893, + 1333, + 299, + 1333 + ], + "score": 0.879 + }, + { + "category_id": 0, + "poly": [ + 300, + 1709, + 643, + 1709, + 643, + 1742, + 300, + 1742 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.845 + }, + { + "category_id": 6, + "poly": [ + 298, + 615, + 1405, + 615, + 1405, + 739, + 298, + 739 + ], + "score": 0.675 + }, + { + "category_id": 0, + "poly": [ + 299, + 1765, + 998, + 1765, + 998, + 1799, + 299, + 1799 + ], + "score": 0.661 + }, + { + "category_id": 1, + "poly": [ + 298, + 615, + 1405, + 615, + 1405, + 739, + 298, + 739 + ], + "score": 0.6 + }, + { + "category_id": 6, + "poly": [ + 299, + 1179, + 1404, + 1179, + 1404, + 1241, + 299, + 1241 + ], + "score": 0.594 + }, + { + "category_id": 0, + "poly": [ + 299, + 1551, + 989, + 1551, + 989, + 1584, + 299, + 1584 + ], + "score": 0.558 + }, + { + "category_id": 1, + "poly": [ + 299, + 1551, + 989, + 1551, + 989, + 1584, + 299, + 1584 + ], + "score": 0.373 + }, + { + "category_id": 1, + "poly": [ + 299, + 1765, + 998, + 1765, + 998, + 1799, + 299, + 1799 + ], + "score": 0.288 + }, + { + "category_id": 1, + "poly": [ + 299, + 1179, + 1404, + 1179, + 1404, + 1241, + 299, + 1241 + ], + "score": 0.198 + }, + { + "category_id": 13, + "poly": [ + 538, + 649, + 556, + 649, + 556, + 680, + 538, + 680 + ], + "score": 0.65, + "latex": "\\dagger" + }, + { + "category_id": 13, + "poly": [ + 777, + 542, + 797, + 542, + 797, + 564, + 777, + 564 + ], + "score": 0.36, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 803, + 1105, + 823, + 1105, + 823, + 1128, + 803, + 1128 + ], + "score": 0.3, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 729, + 964, + 765, + 964, + 765, + 990, + 729, + 990 + ], + "score": 0.28, + "latex": "\\times 1" + }, + { + "category_id": 13, + "poly": [ + 721, + 279, + 742, + 279, + 742, + 301, + 721, + 301 + ], + "score": 0.28, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1917.0, + 881.0, + 1917.0, + 881.0, + 1952.0, + 296.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1300.0, + 896.0, + 1300.0, + 896.0, + 1336.0, + 295.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1708.0, + 644.0, + 1708.0, + 644.0, + 1744.0, + 295.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 615.0, + 1406.0, + 615.0, + 1406.0, + 651.0, + 293.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 647.0, + 537.0, + 647.0, + 537.0, + 679.0, + 296.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 647.0, + 1405.0, + 647.0, + 1405.0, + 679.0, + 557.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 677.0, + 1406.0, + 677.0, + 1406.0, + 713.0, + 295.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 710.0, + 503.0, + 710.0, + 503.0, + 743.0, + 296.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1765.0, + 1000.0, + 1765.0, + 1000.0, + 1802.0, + 296.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1175.0, + 1405.0, + 1175.0, + 1405.0, + 1216.0, + 292.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1206.0, + 376.0, + 1206.0, + 376.0, + 1245.0, + 293.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1550.0, + 991.0, + 1550.0, + 991.0, + 1587.0, + 294.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1396.0, + 295.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1386.0, + 1406.0, + 1386.0, + 1406.0, + 1429.0, + 291.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1420.0, + 1407.0, + 1420.0, + 1407.0, + 1457.0, + 294.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1450.0, + 1405.0, + 1450.0, + 1405.0, + 1487.0, + 294.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1481.0, + 1401.0, + 1481.0, + 1401.0, + 1518.0, + 295.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1607.0, + 1407.0, + 1607.0, + 1407.0, + 1646.0, + 294.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1639.0, + 675.0, + 1639.0, + 675.0, + 1673.0, + 295.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1968.0, + 1408.0, + 1968.0, + 1408.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1997.0, + 1405.0, + 1997.0, + 1405.0, + 2041.0, + 293.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1815.0, + 1405.0, + 1815.0, + 1405.0, + 1860.0, + 293.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1852.0, + 571.0, + 1852.0, + 571.0, + 1886.0, + 295.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 615.0, + 1406.0, + 615.0, + 1406.0, + 651.0, + 293.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 647.0, + 537.0, + 647.0, + 537.0, + 679.0, + 296.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 647.0, + 1405.0, + 647.0, + 1405.0, + 679.0, + 557.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 677.0, + 1406.0, + 677.0, + 1406.0, + 713.0, + 295.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 710.0, + 503.0, + 710.0, + 503.0, + 743.0, + 296.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1550.0, + 991.0, + 1550.0, + 991.0, + 1587.0, + 294.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1765.0, + 1000.0, + 1765.0, + 1000.0, + 1802.0, + 296.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1175.0, + 1405.0, + 1175.0, + 1405.0, + 1216.0, + 292.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1206.0, + 376.0, + 1206.0, + 376.0, + 1245.0, + 293.0, + 1245.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1546, + 1405, + 1546, + 1405, + 1792, + 298, + 1792 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 987, + 1404, + 987, + 1404, + 1141, + 298, + 1141 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1358, + 1403, + 1358, + 1403, + 1512, + 298, + 1512 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1404, + 1880, + 1404, + 2033, + 299, + 2033 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1229, + 1404, + 1229, + 1404, + 1322, + 299, + 1322 + ], + "score": 0.971 + }, + { + "category_id": 5, + "poly": [ + 529, + 225, + 1162, + 225, + 1162, + 315, + 529, + 315 + ], + "score": 0.963, + "html": "
p00.20.40.60.81
Top-180.6680.8380.8280.8880.9280.64
" + }, + { + "category_id": 3, + "poly": [ + 695, + 462, + 1004, + 462, + 1004, + 703, + 695, + 703 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 295, + 835, + 1401, + 835, + 1401, + 899, + 295, + 899 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 296, + 353, + 1407, + 353, + 1407, + 417, + 296, + 417 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 813, + 75, + 813, + 105, + 298, + 105 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 301, + 933, + 792, + 933, + 792, + 965, + 301, + 965 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 298, + 1827, + 946, + 1827, + 946, + 1859, + 298, + 1859 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 4, + "poly": [ + 292, + 742, + 1400, + 742, + 1400, + 776, + 292, + 776 + ], + "score": 0.778 + }, + { + "category_id": 0, + "poly": [ + 302, + 1175, + 1149, + 1175, + 1149, + 1207, + 302, + 1207 + ], + "score": 0.661 + }, + { + "category_id": 1, + "poly": [ + 302, + 1175, + 1149, + 1175, + 1149, + 1207, + 302, + 1207 + ], + "score": 0.259 + }, + { + "category_id": 1, + "poly": [ + 292, + 742, + 1400, + 742, + 1400, + 776, + 292, + 776 + ], + "score": 0.136 + }, + { + "category_id": 13, + "poly": [ + 1383, + 360, + 1402, + 360, + 1402, + 386, + 1383, + 386 + ], + "score": 0.68, + "latex": "p" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 461.0, + 995.0, + 461.0, + 995.0, + 489.0, + 757.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 499.0, + 758.0, + 499.0, + 758.0, + 627.0, + 716.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 504.0, + 728.0, + 504.0, + 728.0, + 640.0, + 696.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 586.0, + 878.0, + 586.0, + 878.0, + 604.0, + 862.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 594.0, + 993.0, + 594.0, + 993.0, + 612.0, + 976.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 634.0, + 764.0, + 634.0, + 764.0, + 676.0, + 724.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 650.0, + 893.0, + 650.0, + 893.0, + 678.0, + 844.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 649.0, + 1006.0, + 649.0, + 1006.0, + 680.0, + 954.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 678.0, + 960.0, + 678.0, + 960.0, + 704.0, + 792.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.75, + 558.5, + 809.75, + 558.5, + 809.75, + 572.5, + 784.75, + 572.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 814.0, + 73.0, + 814.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 930.0, + 798.0, + 930.0, + 798.0, + 970.0, + 295.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1822.0, + 949.0, + 1822.0, + 949.0, + 1865.0, + 294.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2122.0, + 831.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 740.0, + 1405.0, + 740.0, + 1405.0, + 778.0, + 294.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1173.0, + 1154.0, + 1173.0, + 1154.0, + 1212.0, + 295.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1548.0, + 1406.0, + 1548.0, + 1406.0, + 1581.0, + 296.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1579.0, + 1407.0, + 1579.0, + 1407.0, + 1613.0, + 295.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1609.0, + 1406.0, + 1609.0, + 1406.0, + 1644.0, + 292.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1639.0, + 1405.0, + 1639.0, + 1405.0, + 1673.0, + 293.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1669.0, + 1408.0, + 1669.0, + 1408.0, + 1703.0, + 295.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1698.0, + 1406.0, + 1698.0, + 1406.0, + 1736.0, + 292.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1729.0, + 1409.0, + 1729.0, + 1409.0, + 1765.0, + 292.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1759.0, + 1176.0, + 1759.0, + 1176.0, + 1795.0, + 293.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 986.0, + 1404.0, + 986.0, + 1404.0, + 1023.0, + 293.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1016.0, + 1404.0, + 1016.0, + 1404.0, + 1052.0, + 295.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1046.0, + 1407.0, + 1046.0, + 1407.0, + 1088.0, + 292.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1076.0, + 1405.0, + 1076.0, + 1405.0, + 1117.0, + 293.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1109.0, + 948.0, + 1109.0, + 948.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1357.0, + 1405.0, + 1357.0, + 1405.0, + 1393.0, + 293.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1390.0, + 1404.0, + 1390.0, + 1404.0, + 1423.0, + 296.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1418.0, + 1404.0, + 1418.0, + 1404.0, + 1455.0, + 293.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1446.0, + 1405.0, + 1446.0, + 1405.0, + 1487.0, + 293.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1479.0, + 1349.0, + 1479.0, + 1349.0, + 1516.0, + 294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1916.0, + 297.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1408.0, + 1910.0, + 1408.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1408.0, + 1941.0, + 1408.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1974.0, + 1403.0, + 1974.0, + 1403.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 406.0, + 2000.0, + 406.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1226.0, + 1408.0, + 1226.0, + 1408.0, + 1267.0, + 294.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1261.0, + 1403.0, + 1261.0, + 1403.0, + 1295.0, + 294.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1291.0, + 563.0, + 1291.0, + 563.0, + 1325.0, + 294.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 835.0, + 1403.0, + 835.0, + 1403.0, + 871.0, + 294.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 870.0, + 519.0, + 870.0, + 519.0, + 900.0, + 294.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 346.0, + 1382.0, + 346.0, + 1382.0, + 393.0, + 293.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 346.0, + 1407.0, + 346.0, + 1407.0, + 393.0, + 1403.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 381.0, + 1278.0, + 381.0, + 1278.0, + 422.0, + 294.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1173.0, + 1154.0, + 1173.0, + 1154.0, + 1212.0, + 295.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 740.0, + 1405.0, + 740.0, + 1405.0, + 778.0, + 294.0, + 778.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1314, + 1405, + 1314, + 1405, + 1408, + 298, + 1408 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 478, + 236, + 1220, + 236, + 1220, + 591, + 478, + 591 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 477, + 740, + 1221, + 740, + 1221, + 1069, + 477, + 1069 + ], + "score": 0.962 + }, + { + "category_id": 4, + "poly": [ + 297, + 1107, + 1398, + 1107, + 1398, + 1202, + 297, + 1202 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 297, + 631, + 1402, + 631, + 1402, + 696, + 297, + 696 + ], + "score": 0.951 + }, + { + "category_id": 0, + "poly": [ + 301, + 1260, + 686, + 1260, + 686, + 1292, + 301, + 1292 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.846 + }, + { + "category_id": 15, + "poly": [ + 614.0, + 239.0, + 744.0, + 239.0, + 744.0, + 272.0, + 614.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 240.0, + 1135.0, + 240.0, + 1135.0, + 269.0, + 1007.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 278.0, + 553.0, + 278.0, + 553.0, + 438.0, + 504.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 289.0, + 514.0, + 289.0, + 514.0, + 453.0, + 478.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 291.0, + 906.0, + 291.0, + 906.0, + 451.0, + 869.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 300.0, + 591.0, + 300.0, + 591.0, + 317.0, + 573.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 304.0, + 970.0, + 304.0, + 970.0, + 333.0, + 940.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 315.0, + 938.0, + 315.0, + 938.0, + 349.0, + 902.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 353.0, + 656.0, + 353.0, + 656.0, + 373.0, + 634.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 394.0, + 956.0, + 394.0, + 956.0, + 404.0, + 947.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 393.0, + 1123.0, + 393.0, + 1123.0, + 416.0, + 1103.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 405.0, + 937.0, + 405.0, + 937.0, + 434.0, + 903.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 420.0, + 1117.0, + 420.0, + 1117.0, + 430.0, + 1108.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 446.0, + 546.0, + 446.0, + 546.0, + 480.0, + 512.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 456.0, + 1198.0, + 456.0, + 1198.0, + 464.0, + 1190.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 469.0, + 591.0, + 469.0, + 591.0, + 495.0, + 571.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 464.0, + 700.0, + 464.0, + 700.0, + 500.0, + 630.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 465.0, + 796.0, + 465.0, + 796.0, + 500.0, + 726.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 465.0, + 983.0, + 465.0, + 983.0, + 498.0, + 924.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 466.0, + 1063.0, + 466.0, + 1063.0, + 498.0, + 1002.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 466.0, + 1142.0, + 466.0, + 1142.0, + 498.0, + 1082.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 466.0, + 1222.0, + 466.0, + 1222.0, + 498.0, + 1162.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 497.0, + 768.0, + 497.0, + 768.0, + 531.0, + 591.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 496.0, + 1122.0, + 496.0, + 1122.0, + 532.0, + 1020.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 549.0, + 827.0, + 549.0, + 827.0, + 581.0, + 687.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 550.0, + 1059.0, + 550.0, + 1059.0, + 582.0, + 862.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.75, + 369.0, + 1040.75, + 369.0, + 1040.75, + 389.0, + 1018.75, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 738.0, + 794.0, + 738.0, + 794.0, + 776.0, + 591.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 742.0, + 1167.0, + 742.0, + 1167.0, + 775.0, + 943.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 761.0, + 565.0, + 761.0, + 565.0, + 798.0, + 507.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 762.0, + 927.0, + 762.0, + 927.0, + 799.0, + 869.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 791.0, + 565.0, + 791.0, + 565.0, + 950.0, + 478.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 791.0, + 925.0, + 791.0, + 925.0, + 950.0, + 838.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 821.0, + 1179.0, + 821.0, + 1179.0, + 835.0, + 1165.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 850.0, + 565.0, + 850.0, + 565.0, + 892.0, + 509.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 852.0, + 924.0, + 852.0, + 924.0, + 889.0, + 869.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 844.0, + 981.0, + 844.0, + 981.0, + 878.0, + 959.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 894.0, + 565.0, + 894.0, + 565.0, + 936.0, + 509.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 895.0, + 926.0, + 895.0, + 926.0, + 936.0, + 870.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 942.0, + 564.0, + 942.0, + 564.0, + 979.0, + 508.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 943.0, + 943.0, + 943.0, + 943.0, + 980.0, + 869.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 953.0, + 1042.0, + 953.0, + 1042.0, + 963.0, + 1034.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 952.0, + 1181.0, + 952.0, + 1181.0, + 963.0, + 1167.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 963.0, + 814.0, + 963.0, + 814.0, + 993.0, + 565.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 960.0, + 986.0, + 960.0, + 986.0, + 996.0, + 921.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 961.0, + 1057.0, + 961.0, + 1057.0, + 995.0, + 1019.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 960.0, + 1192.0, + 960.0, + 1192.0, + 997.0, + 1153.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 987.0, + 819.0, + 987.0, + 819.0, + 1017.0, + 567.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 991.0, + 1111.0, + 991.0, + 1111.0, + 1027.0, + 999.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1009.0, + 811.0, + 1009.0, + 811.0, + 1039.0, + 571.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1024.0, + 982.0, + 1024.0, + 982.0, + 1059.0, + 893.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1027.0, + 1099.0, + 1027.0, + 1099.0, + 1059.0, + 1011.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1024.0, + 1215.0, + 1024.0, + 1215.0, + 1061.0, + 1127.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.75, + 831.0, + 1043.75, + 831.0, + 1043.75, + 859.0, + 1029.75, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 866.0, + 955.0, + 866.0, + 955.0, + 888.5, + 929.0, + 888.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.25, + 886.0, + 949.25, + 886.0, + 949.25, + 902.0, + 927.25, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1108.0, + 1404.0, + 1108.0, + 1404.0, + 1142.0, + 296.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1137.0, + 1403.0, + 1137.0, + 1403.0, + 1174.0, + 295.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1170.0, + 697.0, + 1170.0, + 697.0, + 1204.0, + 294.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 630.0, + 1404.0, + 630.0, + 1404.0, + 666.0, + 293.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 664.0, + 529.0, + 664.0, + 529.0, + 697.0, + 293.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1260.0, + 689.0, + 1260.0, + 689.0, + 1296.0, + 296.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1311.0, + 1406.0, + 1311.0, + 1406.0, + 1353.0, + 292.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1339.0, + 1406.0, + 1339.0, + 1406.0, + 1387.0, + 291.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1374.0, + 1055.0, + 1374.0, + 1055.0, + 1410.0, + 293.0, + 1410.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 333, + 445, + 1369, + 445, + 1369, + 913, + 333, + 913 + ], + "score": 0.984, + "html": "
Quantization SchemeLanguage Modeling 16-layer Transformer Wikitext-103Image Classification EfficientNet-B3 ImageNet-1K
SizeCompressTest PPLSizeCompressTop-1 Acc.
Uncompressed model942×118.346.7×181.5
Int4 Quant Histogram118×839.45.8×845.3
+ Quant-Noise118×821.85.8×867.8
Int4 Quant Channel118×821.25.8×868.2
+ Quant-Noise118×819.55.8×872.3
Int8 Quant Histogram236×419.611.7×480.7
+ Quant-Noise236×418.711.7×480.9
Int8 Quant Channel236×418.511.7×481.1
+ Quant-Noise236×418.311.7×481.2
" + }, + { + "category_id": 5, + "poly": [ + 549, + 1585, + 1144, + 1585, + 1144, + 1703, + 549, + 1703 + ], + "score": 0.969, + "html": "
ModelMBPPL
Quant-Noise + Share + Prune1024.2
Quant-Noise + Share + Prune with STE1024.5
" + }, + { + "category_id": 6, + "poly": [ + 292, + 1739, + 1406, + 1739, + 1406, + 1804, + 292, + 1804 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 866, + 2088, + 866, + 2113, + 834, + 2113 + ], + "score": 0.854 + }, + { + "category_id": 6, + "poly": [ + 296, + 933, + 1408, + 933, + 1408, + 1119, + 296, + 1119 + ], + "score": 0.629 + }, + { + "category_id": 1, + "poly": [ + 296, + 933, + 1408, + 933, + 1408, + 1119, + 296, + 1119 + ], + "score": 0.345 + }, + { + "category_id": 13, + "poly": [ + 701, + 1639, + 722, + 1639, + 722, + 1660, + 701, + 1660 + ], + "score": 0.46, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 785, + 1639, + 806, + 1639, + 806, + 1660, + 785, + 1660 + ], + "score": 0.39, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 785, + 1667, + 806, + 1667, + 806, + 1690, + 785, + 1690 + ], + "score": 0.36, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 702, + 1668, + 722, + 1668, + 722, + 1690, + 702, + 1690 + ], + "score": 0.25, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1735.0, + 1409.0, + 1735.0, + 1409.0, + 1779.0, + 291.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1771.0, + 543.0, + 1771.0, + 543.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2123.0, + 828.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 932.0, + 1409.0, + 932.0, + 1409.0, + 969.0, + 293.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 962.0, + 1408.0, + 962.0, + 1408.0, + 1000.0, + 293.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 993.0, + 1408.0, + 993.0, + 1408.0, + 1031.0, + 293.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1023.0, + 1405.0, + 1023.0, + 1405.0, + 1060.0, + 293.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1057.0, + 1406.0, + 1057.0, + 1406.0, + 1092.0, + 294.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 716.0, + 1087.0, + 716.0, + 1119.0, + 294.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 932.0, + 1409.0, + 932.0, + 1409.0, + 969.0, + 293.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 962.0, + 1408.0, + 962.0, + 1408.0, + 1000.0, + 293.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 993.0, + 1408.0, + 993.0, + 1408.0, + 1031.0, + 293.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1023.0, + 1405.0, + 1023.0, + 1405.0, + 1060.0, + 293.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1057.0, + 1406.0, + 1057.0, + 1406.0, + 1092.0, + 294.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 716.0, + 1087.0, + 716.0, + 1119.0, + 294.0, + 1119.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/hfkER_KJiNw/hfkER_KJiNw.md b/parse/train/hfkER_KJiNw/hfkER_KJiNw.md new file mode 100644 index 0000000000000000000000000000000000000000..d63d5c14f35c40b55b61a825ddd60410e9495169 --- /dev/null +++ b/parse/train/hfkER_KJiNw/hfkER_KJiNw.md @@ -0,0 +1,358 @@ +# Graph Neural Networks with Adaptive Residual + +Xiaorui Liu1 xiaorui@msu.edu + +Jiayuan Ding1 dingjia5@msu.edu + +Wei Jin1 jinwei2@msu.edu + +Han Xu1 xuhan1@msu.edu + +Yao Ma2 yao.ma@njit.edu + +Zitao Liu3 liuzitao@100tal.com + +Jiliang Tang1 tangjili@msu.edu + +1Michigan State University, East Lansing, MI, USA 2New Jersey Institute of Technology, Newark, NJ, USA 3TAL Education Group, Beijing, China + +# Abstract + +Graph neural networks (GNNs) have shown the power in graph representation learning for numerous tasks. In this work, we discover an interesting phenomenon that although residual connections in the message passing of GNNs help improve the performance, they immensely amplify GNNs’ vulnerability against abnormal node features. This is undesirable because in real-world applications, node features in graphs could often be abnormal such as being naturally noisy or adversarially manipulated. We analyze possible reasons to understand this phenomenon and aim to design GNNs with stronger resilience to abnormal features. Our understandings motivate us to propose and derive a simple, efficient, interpretable, and adaptive message passing scheme, leading to a novel GNN with Adaptive residual, AirGNN1. Extensive experiments under various abnormal feature scenarios demonstrate the effectiveness of the proposed algorithm. + +# 1 Introduction + +Recent years have witnessed the great success of graph neural networks (GNNs) in representation learning for graph structure data [1]. Essentially, GNNs generalize deep neural networks (DNNs) from regular grids, such as image, video and text, to irregular data such as social, energy, transportation, citation, and biological networks. Such data can be naturally represented as graphs with nodes and edges. The key building block for such generalization is the neural message passing framework [2]: + +$$ +\mathbf { x } _ { u } ^ { ( k + 1 ) } = \mathrm { U P D A T E } ^ { ( k ) } \left( \mathbf { x } _ { u } ^ { ( k ) } , \mathbf { m } _ { \mathcal { N } ( u ) } ^ { ( k ) } \right) +$$ + +where x(ku $\mathbf { x } _ { u } ^ { ( k ) } \in \mathbb { R } ^ { d }$ denotes the feature vector of node $u$ in the $k$ -th iteration of message passing, and $\mathbf { m } _ { \mathcal { N } ( u ) } ^ { ( k ) }$ is the message aggregated from $u$ ’s neighborhood $\mathcal { N } ( u )$ . The specific design of message passing scheme can be motivated from spectral domain [3, 4] or spatial domain [5, 6, 7, 2]. It usually linearly smooths the features in a local neighborhood on the graph. + +GNNs have achieved superior performance in a large number of benchmark datasets [8] where the node features are assumed to be complete and informative. However, in real-world applications, some node features could be abnormal from various aspects. For instance, in social networks, new users might not have complete profile before they make connections with others, leading to missing user features. In transportation networks, node features can be noisy since there exist certain uncertainty and dynamics in the observation of the traffic information. What is worse, node features can be adversarially chosen by the attacker to maliciously manipulate the prediction made by GNNs. Therefore, it is greatly desired to design GNN models with stronger resilience to abnormal node features. + +In this work, we first perform empirical investigations on how representative GNN models behave on graphs with abnormal features. Specifically, based upon standard benchmark datasets, we simulate the abnormal features by replacing the features of randomly selected nodes with random Gaussian noise. Then the performance of node classification on abnormal features and normal features are examined separately. From our preliminary study in Section 2, we reveal two interesting observations: (1) Feature aggregation can boost the resilience to abnormal features, but too many aggregations could hurt the performance on both normal and abnormal features; and (2) Residual connection helps GNNs benefit from more layers for normal features, while making GNNs more fragile to abnormal features. We then provide possible explanations to understand these observed phenomena from the perspective of graph Laplacian smoothing. Our analyses imply that there might exist an intrinsic tension between feature aggregation and residual connection, which results in a performance tradeoff between normal features and abnormal features. + +Motivated by these findings and understandings, we aim to design new GNNs with stronger resilience to abnormal features while largely maintaining the performance on normal features. Our contributions can be summarized as follows: + +• We discover an intrinsic tension between feature aggregation and residual connection in GNNs, and the corresponding performance tradeoff between abnormal and normal features. We also analyze possible reasons to explain and understand these findings. +• We propose a simple, efficient, principled and adaptive message passing scheme, which leads to a novel GNN model with adaptive residual, named as AirGNN. +• Extensive experiments under various abnormal feature scenarios demonstrate the superiority of the proposed algorithm. The ablation study demonstrates how the adaptive residuals mitigate the impact of abnormal features. + +# 2 Preliminary + +Before introducing the preliminary study, we first define the notations used throughout the paper. + +Notations. We use bold upper-case letters such as $\mathbf { X }$ to denote matrices. Given a matrix $\mathbf { X } \in \mathbb { R } ^ { n \times d }$ , we use $\mathbf { X } _ { i }$ to denote its $i$ -th row and $\mathbf { X } _ { i j }$ to denote its element in $i$ -th row and $j$ -th column. The Frobenius norm and $\ell _ { 2 1 }$ norm of a matrix $\mathbf { X }$ are defined as $\Vert \mathbf { X } \Vert _ { F } = \sqrt { \sum _ { i j } \mathbf { X } _ { i j } ^ { 2 } }$ and $\| \mathbf { X } \| _ { 2 1 } =$ $\begin{array} { r } { \sum _ { i } \| \mathbf { X } _ { i } \| _ { 2 } = \sum _ { i } \sqrt { \sum _ { j } \mathbf { X } _ { i j } ^ { 2 } } } \end{array}$ , respectively. We define $\| \mathbf { X } \| _ { 2 } = \sigma _ { \operatorname* { m a x } } ( \mathbf { X } )$ where $\sigma _ { \mathrm { m a x } } ( \mathbf { X } )$ is the largest singular value of $\mathbf { X }$ . + +Let $\mathcal { G } = \{ \nu , \mathcal { E } \}$ be a graph with the node set $\mathcal { V } = \{ v _ { 1 } , \ldots , v _ { n } \}$ and the undirected edge set $\mathcal { E } = \{ e _ { 1 } , \ldots , e _ { m } \}$ . We use $\mathcal { N } ( v _ { i } )$ to denote the neighboring nodes of node $v _ { i }$ , including $v _ { i }$ itself. Suppose that each node is associated with a $d$ -dimensional feature vector, and the features for all nodes are denoted as $\mathbf { X } _ { \mathrm { f e a } } \in \mathbb { R } ^ { n \times d }$ . The graph structure $\mathcal { G }$ can be represented as an adjacent matrix $\mathbf { A } \in \mathbb { R } ^ { n \times n }$ , where $\mathbf { A } _ { i j } ~ = ~ 1$ when there exists an edge between nodes $v _ { i }$ and $v _ { j }$ , and $\mathbf { A } _ { i j } \ = \ 0$ otherwise. The graph Laplacian matrix is defined as $\mathbf { L } = \mathbf { D } - \mathbf { A }$ , where $\mathbf { D }$ is the diagonal degree matrix. Let us denote the commonly used feature aggregation matrix in GNNs [3] as $\tilde { \bf A } = \hat { \bf D } ^ { - \frac { 1 } { 2 } } \bar { \hat { \bf A } } \hat { \bf D } ^ { - \frac { 1 } { 2 } }$ where $\hat { \mathbf { A } } = \mathbf { A } + \mathbf { I }$ is the adjacent matrix with self-loop and its degree matrix is $\hat { \bf D }$ . The corresponding Laplacian matrix is defined as $\tilde { \mathbf { L } } = \mathbf { I } - \tilde { \mathbf { A } }$ . + +In this work, we focus on the setting where a subset of nodes in the graph contain abnormal features, while the remaining nodes have normal features. In the remaining of this paper, we use abnormal/normal features to denote nodes with abnormal/normal features, for simplicity. + +# 2.1 Preliminary Study + +Experimental setup. To investigate how GNNs behave on abnormal and normal node features, we design semi-supervised node classification experiments on three common datasets (i.e., Cora, + +CiteSeer and PubMed), following the data splits in the work [3]. Moreover, we simulate the abnormal features by assigning $1 0 \%$ of the nodes with random features sampled from a standard Gaussian distribution. The experiments are performed on representative GNN models covering coupled and decoupled architectures, including GCN [3], GCNII [9], APPNP [10], and their variants with or without residual connections in feature aggregations, denoted as w/Res and wo/Res. All methods follow the hyperparameter settings in their original papers. We examine how these models perform when the number of layers increases. Note that for the decoupled architectures such as APPNP, we fix the 2-layer MLP and increase the number of propagation layers. While for the coupled architectures such as GCN and GCNII, we increase the number of feature transformation and propagation layers simultaneously. We report the average performance over 10 times of random selection of the noise node sets. The node classification accuracy (mean and standard variance) on nodes with abnormal and normal features is illustrated in Figure 1 and Figure. 2, separately. + +![](images/00d120485edd7da85f1457dd4c00230b18d99c238c177b5cb16e29a69a61afb9.jpg) +Figure 1: Node classification accuracy on abnormal nodes (Cora) + +![](images/08499e09f3e29a9b457146ea9c203108450fafd94446152fa1b7aada0f5024ce.jpg) +Figure 2: Node classification accuracy on normal nodes (Cora) + +Observations. From Figure 1 and Figure 2, we can make the following observations: (1) Without residual connection, more layers (e.g., $> 2$ for GCN and GCNII, $> 1 0$ for APPNP) hurt the accuracy on nodes with normal features. However, more layers boost the accuracy on nodes with abnormal features significantly, before finally starting to decrease; (2) With residual connection, the accuracy on nodes with normal features keeps increasing with more layers2. However, the accuracy on nodes with abnormal features only increases marginally when stacking more layers, and then starts to decrease. While we only present the experiments on Cora, we defer the results on other datasets to Appendix C, which provide similar observations. To conclude, we can summarize these observations into two major findings: + +• Finding I: Feature aggregation can boost the resilience to abnormal features, but too many aggregations could hurt the performance on both normal and abnormal nodes; + +• Finding II: Residual connection helps GNNs benefit from more layers for nodes with normal features, while making GNNs more fragile to abnormal features. + +# 2.2 Understandings + +In this subsection, we provide the understanding and explanation for aforementioned findings, from the perspective of graph Laplacian smoothing. + +# Understanding Finding I: Feature aggregation as Laplacian smoothing + +The message passing in GCN [3], GCNII wo/ residual and APPNP wo/ residual (as well as many popular GNN models), follows the feature aggregation + +$$ +\mathbf { X } _ { \mathrm { o u t } } = \tilde { \mathbf { A } } \mathbf { X } _ { \mathrm { i n } } , +$$ + +where ${ \bf X } _ { \mathrm { i n } }$ and $\mathbf { X _ { o u t } }$ represent the features before and after message passing layer, respectively. It can be interpreted as one gradient descent step for the Laplacian smoothing problem [11] + +$$ +\underset { \mathbf { X } \in \mathbb { R } ^ { n \times d } } { \arg \operatorname* { m i n } } \mathcal { L } _ { 1 } ( \mathbf { X } ) : = \frac { 1 } { 2 } \mathrm { t r } \Big ( \mathbf { X } ^ { \top } ( \mathbf { I } - \tilde { \mathbf { A } } ) \mathbf { X } \Big ) = \frac { 1 } { 2 } \sum _ { ( v _ { i } , v _ { j } ) \in \mathcal { E } } \| \frac { \mathbf { X } _ { i } } { \sqrt { d _ { i } + 1 } } - \frac { \mathbf { X } _ { j } } { \sqrt { d _ { j } + 1 } } \| _ { 2 } ^ { 2 } , +$$ + +where $d _ { i }$ is the node degree of node $v _ { i }$ . Eq. (2) can be derived from $\mathbf { X } _ { \mathrm { o u t } } = \mathbf { X } _ { \mathrm { i n } } - ( \mathbf { I } - \tilde { \mathbf { A } } ) \mathbf { X } _ { \mathrm { i n } } = \tilde { \mathbf { A } } \mathbf { X } _ { \mathrm { i n } }$ , with the initialization $\mathbf { X } = \mathbf { X } _ { \mathrm { i n } }$ and stepsize $\gamma = 1$ . The Laplacian smoothing problem penalizes the feature difference between neighboring nodes. To reduce this penalty, the feature aggregation in Eq. (2) smooths the node features by taking the average of local neighbors, and thus can be considered as low-pass filter which gradually filters out high-frequency signals [12, 13]. Therefore, it increases the resilience to abnormal features which are likely to be high-frequency signals. In other words, the local neighboring nodes help to correct the abnormal features. Unfortunately, if applied too many times, these low-pass filters could overly smooth the features (well-known as oversmoothing [14, 15]) such that nodes are not distinguishable enough, providing an explanation to the degraded performance on both abnormal and normal features when stacking too many layers. + +# Understanding Finding II: Residual connection maintains feature proximity + +To adjust the feature smoothness for better performance, APPNP [10] utilizes residual connections in message passing as follows + +$$ +\mathbf { X } ^ { k + 1 } = ( 1 - \alpha ) \tilde { \mathbf { A } } \mathbf { X } ^ { k } + \alpha \mathbf { X _ { \mathrm { i n } } } , +$$ + +where ${ \bf X } ^ { 0 } = { \bf X } _ { \mathrm { i n } }$ . It can be considered as an iterative solution for the regularized Laplacian smoothing problem [11] + +$$ +\underset { { \substack { \mathbf { X } \in \mathbb { R } ^ { n \times d } } } } { \arg \operatorname* { m i n } } \ \mathcal { L } _ { 2 } ( { \mathbf { X } } ) : = \frac { \alpha } { 2 ( 1 - \alpha ) } \| \mathbf { X } - \mathbf { X } _ { \mathrm { i n } } \| _ { F } ^ { 2 } + \frac { 1 } { 2 } \mathrm { t r } \Big ( \mathbf { X } ^ { \top } ( { \mathbf { I } } - \tilde { \mathbf { A } } ) \mathbf { X } \Big ) , +$$ + +with initialization $\mathbf { X } = \mathbf { X } _ { \mathrm { i n } }$ and stepsize $\gamma = 1 - \alpha$ due to + +$$ +\mathbf { X } ^ { k + 1 } = \mathbf { X } ^ { k } - ( 1 - \alpha ) \bigg ( \frac { \alpha } { 1 - \alpha } ( \mathbf { X } ^ { k } - \mathbf { X _ { \mathrm { i n } } } ) + ( \mathbf { I } - \tilde { \mathbf { A } } ) \mathbf { X } ^ { k } \bigg ) = ( 1 - \alpha ) \tilde { \mathbf { A } } \mathbf { X } ^ { k } + \alpha \mathbf { X _ { \mathrm { i n } } } . +$$ + +GCNII [9] adopts a similar message passing but further combines a feature transformation layer in each message passing step, which leads to a coupled architecture, as contrast to the decoupled architecture of APPNP. The residual connection naturally arises when regularizing the proximity between input and output features, as showed in the first term of $\mathcal { L } _ { 2 } ( \mathbf { X } )$ . Such proximity can help avoid the trivial solution for the problem in Eq. (3), i.e., totally oversmoothed features only depending on node degrees, and consequently mitigates the oversmoothing issue. More intuitively, residual connections in GNNs provide direct information flows between layers that can preserve some necessary high-frequency signals for better discrimination between classes. More layers with residual provide a more accurate solution to Eq. (5), which explains the performance gain from deeper GNNs. Unfortunately, these residual connections also undesirably carry on abnormal features which are detrimental, leading to the inferior performance on abnormal features. + +# 3 The Proposed Framework + +In this section, we first motivate the proposed adaptive message passing scheme (AMP) with further discussions on our preliminary study. We then introduce more details about AMP, its interpretations, convergence guarantee and computation complexity, as well as the model architecture of AirGNN. + +# 3.1 Design Motivation + +Our preliminary study in Section 2 reveals an intrinsic tension between feature aggregation and residual connection: (1) feature aggregation helps smooth out abnormal features, while it could cause inappropriate smoothing for normal features; (2) residual connection is essential for adjusting the feature smoothness, but it could be detrimental for abnormal features. Although this conflict can be partially mitigated by adjusting the residual connection such as the residual weight $\alpha$ in GCNII [9] and APPNP [10], such global adjustment cannot be adaptive to a subset of the nodes, e.g., the nodes with abnormal features. This is crucial because in practice we often encounter the scenario where only a subset of nodes contain abnormal features. Therefore, how to reconcile this dilemma still desires dedicated efforts. We then naturally ask a question: Can we design a better message passing scheme with node-wise adaptive feature aggregation and residual connection? + +The motivation of the proposed idea builds upon the following intuition: while it is important to maintain the proximity between input and output features as in Eq. (5), it could be over aggressive to penalize their deviations by the square of Frobenius norm, i.e., $\begin{array} { r } { \| \dot { \bf X } - \dot { \bf X } _ { \mathrm { i n } } \| _ { F } ^ { 2 } = \sum _ { i = 1 } ^ { n } \| \dot { \bf X } _ { i } - ( { \bf X } _ { \mathrm { i n } } ) _ { i } \| _ { 2 } ^ { 2 } } \end{array}$ The fact that this penalty does not tolerate large deviations weakens the capability to remove abnormal features through Laplacian smoothing. This motivates us to consider an alternative proximity penalty + +$$ +\| { \bf X } - { \bf X } _ { \mathrm { i n } } \| _ { 2 1 } : = \sum _ { i = 1 } ^ { n } \| { \bf X } _ { i } - ( { \bf X } _ { \mathrm { i n } } ) _ { i } \| _ { 2 } , +$$ + +which instead penalizes the deviations by the $\ell _ { 1 }$ norm of row-wise $\ell _ { 2 }$ norms, namely $\ell _ { 2 1 }$ norm. The $\ell _ { 2 1 }$ norm promotes row sparsity in $\mathbf { X } - \mathbf { X } _ { \mathrm { i n } }$ , and it also allows large deviations because the penalty on large values is less aggressive, leading to the potential removal of abnormal features. Therefore, we propose the following Laplacian smoothing problem regularized by $\ell _ { 2 1 }$ norm proximity control: + +$$ +\underset { \mathbf { X } \in \mathbb { R } ^ { n \times d } } { \arg \operatorname* { m i n } } \lambda \| \mathbf { X } - \mathbf { X } _ { \mathrm { i n } } \| _ { 2 1 } + \frac { 1 } { 2 } \mathrm { t r } ( \mathbf { X } ^ { \top } ( \mathbf { I } - \tilde { \mathbf { A } } ) \mathbf { X } ) , +$$ + +where $\lambda \in [ 0 , \infty )$ is a parameter to adjust the balance between proximity and Laplacian smoothing. In order to easy the tuning of $\lambda$ , we made a modification of Eq. (7): + +$$ +\underset { \mathbf { X } \in \mathbb { R } ^ { n \times d } } { \arg \operatorname* { m i n } } \ \mathcal { L } ( \mathbf { X } ) : = \lambda \| \mathbf { X } - \mathbf { X _ { \mathrm { i n } } } \| _ { 2 1 } + ( 1 - \lambda ) \mathrm { t r } ( \mathbf { X } ^ { \top } ( \mathbf { I } - \tilde { \mathbf { A } } ) \mathbf { X } ) , +$$ + +where $\lambda \in [ 0 , 1 ]$ controls the balance. + +# 3.2 Adaptive Message Passing + +![](images/b9c08a90235cf9dc7aa82eaeface75a6b91e84c8306c75e6acc04630e39b84db.jpg) +Figure 3: Diagram of Adaptive Message Passing + +$\mathcal { L } ( \mathbf { X } )$ is a composite objective with non-smooth and smooth components. We optimize it by proximal gradient descent [16] and obtain the following iterations as the adaptive message passing (AMP): + +$$ +\begin{array} { r l } { { \displaystyle { \bf Y } ^ { k } = { \bf X } ^ { k } - 2 \gamma ( 1 - \lambda ) ( { \bf I } - \tilde { \bf A } ) { \bf X } ^ { k } = \left( 1 - 2 \gamma ( 1 - \lambda ) \right) { \bf X } ^ { k } + 2 \gamma ( 1 - \lambda ) \tilde { \bf A } { \bf X } ^ { k } } } & { { } { } { } } \\ { { \displaystyle { \bf X } ^ { k + 1 } = \arg \operatorname* { m i n } \left\{ \lambda \| { \bf X } - { \bf X } _ { \mathrm { i n } } \| _ { 2 1 } + \frac { 1 } { 2 \gamma } \| { \bf X } - { \bf Y } ^ { k } \| _ { F } ^ { 2 } \right\} } } & { { } { } { } } \end{array} +$$ + +where ${ \bf X } ^ { 0 } = { \bf X } _ { \mathrm { i n } }$ and $\gamma$ is the stepsize to be specified later. Let $\mathbf { Z } = \mathbf { X } - \mathbf { X } _ { \mathrm { i n } }$ , and Eq. (10) can be rewritten as: + +$$ +\begin{array} { r l } { { \mathbf { Z } ^ { k + 1 } = \arg \operatorname* { m i n } \{ \lambda \| \mathbf { Z } \| _ { 2 1 } + \frac { 1 } { 2 \gamma } \| \mathbf { Z } - ( \mathbf { Y } ^ { k } - \mathbf { X } _ { \mathrm { i n } } ) \| _ { F } ^ { 2 } \} } } \\ & { = \mathbf { p r o x } _ { \gamma \lambda \| \cdot \| _ { 2 1 } } ( \mathbf { Y } ^ { k } - \mathbf { X } _ { \mathrm { i n } } ) } \\ & { \mathbf { X } ^ { k + 1 } = \mathbf { X } _ { \mathrm { i n } } + \mathbf { Z } ^ { k + 1 } . } \end{array} +$$ + +The $i$ -th row of the proximal operator in Eq. (11) can be computed analytically + +$$ +\left( \mathbf { p r o x } _ { \gamma \lambda \parallel \cdot \parallel _ { 2 1 } } ( \mathbf { X } ) \right) _ { i } = \frac { \mathbf { X } _ { i } } { \parallel \mathbf { X } _ { i } \parallel _ { 2 } } \operatorname* { m a x } ( \parallel \mathbf { X } _ { i } \parallel _ { 2 } - \gamma \lambda , 0 ) = \operatorname* { m a x } ( 1 - \frac { \gamma \lambda } { \parallel \mathbf { X } _ { i } \parallel _ { 2 } } , 0 ) \cdot \mathbf { X } _ { i } . +$$ + +Note that the proximal operator returns 0 if the input vector is 0. Substituting $\mathbf { X }$ in Eq. (13) with $\mathbf { Y } ^ { k } - \mathbf { X } _ { \mathrm { i n } }$ and combining Eq. (11) and Eq. (12), then Eq. (12) becomes + +$$ +\mathbf { X } _ { i } ^ { k + 1 } = ( \mathbf { X } _ { \mathrm { i n } } ) _ { i } + \beta _ { i } ( \mathbf { Y } _ { i } ^ { k } - ( \mathbf { X } _ { \mathrm { i n } } ) _ { i } ) = ( 1 - \beta _ { i } ) ( \mathbf { X } _ { \mathrm { i n } } ) _ { i } + \beta _ { i } \mathbf { Y } _ { i } ^ { k } , \quad \forall i \in [ n ] , +$$ + +where $\begin{array} { r } { \beta _ { i } : = \operatorname* { m a x } ( 1 - \frac { \gamma \lambda } { \| \mathbf { Y } _ { i } ^ { k } - ( \mathbf { X } _ { \mathrm { i n } } ) _ { i } \| _ { 2 } } , 0 ) } \end{array}$ . To summarize, the proposed adaptive message passing (AMP) scheme is showed in Figure 4, and a diagram is showed in Figure 3. In detail, AMP works as follows: + +• The first step takes a feature aggregation within the local neighbors with a self-loop weighted by $1 - 2 \gamma ( 1 - \lambda )$ ; +• The second step computes a weight $\beta _ { i } \in [ 0 , 1 ]$ for each node $v _ { i }$ depending on the local deviation $\| { \bf Y } _ { i } ^ { k } - ( { \bf X } _ { \mathrm { i n } } ) _ { i } \| _ { 2 }$ . +• The final step takes a linear combination of input features $\mathbf { X } _ { \mathrm { i n } }$ and the aggregated features $\mathbf { Y } ^ { k }$ , where the node-wise residual is adaptively weighted by $1 - \beta _ { i }$ for each node $v _ { i }$ . + +![](images/31b59335ec89ac9c9b7a99814a2a6a89348cd1cc224bfd947556babaafd3c58b.jpg) +Figure 4: Adaptive Message Passing (AMP) + +The convergence guarantee of AMP and parameter setting for the stepsize $\gamma$ are illustrated in Theorem 1 and proved in Appendix A. According to Theorem 1, if we set γ = 14(1−λ) or γ = 12(1−λ) , then the first step of AMP can be simplified as $\begin{array} { r } { \mathbf { Y } ^ { k } = \frac { 1 } { 2 } \mathbf { X } ^ { k } + \frac { 1 } { 2 } \tilde { \mathbf { A } } \mathbf { X } ^ { k } } \end{array}$ and $\mathbf { Y } ^ { k } = \tilde { \mathbf { A } } \mathbf { X } ^ { k }$ , respectively. The choice of stepsize will only impact the convergence speed but not the ultimate effect of AMP when it convergences to the fixed point solution. We also discuss the computation complexity per iteration of AMP in Remark 1. + +Theorem 1 (Convergence of AMP). Under the stepsize setting < 1(1−λ)kL˜k , the proposed adaptive message passing scheme (AMP) in Eq. (9) and Eq. (10) converges to the optimal solution of the problem defined in Eq. (8). In practice, it is sufficient to choose any $\begin{array} { r } { \gamma < \frac { 1 } { 2 ( 1 - \lambda ) } } \end{array}$ since $\Vert \tilde { \mathbf { L } } \Vert _ { 2 } \leq 2$ Moreover, if the connected components of the graph $\mathcal { G }$ are not bipartite graphs, it is sufficient to choose $\begin{array} { r } { \gamma = \frac { 1 } { 2 ( 1 - \lambda ) } } \end{array}$ since $\Vert \tilde { \mathbf { L } } \Vert _ { 2 } < 2$ . + +Remark 1 (Computation complexity). AMP is as efficient as simple feature aggregation $\mathbf { X } _ { o u t } =$ $\tilde { \mathbf { A } } \mathbf { X } _ { i n }$ because the additional computation cost from the second and third steps in Figure 4 is in the order $\mathcal { O } ( n d )$ , where $n$ is the number of nodes and $d$ is the feature dimension. This is negligible compared with the computation cost $\mathcal { O } ( m d )$ in feature aggregation, where m is the number of edges, due to the fact that usually there are many more edges than nodes in real-world graphs, i.e., $m \gg n$ + +# 3.3 Interpretation of AMP + +Interestingly, the proposed AMP has a simple and intuitive interpretation as adaptive residual connection, which aligns well with our design motivation: + +• If the feature of node $v _ { i }$ , i.e., $( \mathbf { X } _ { \mathrm { i n } } ) _ { i }$ , is significantly inconsistent with its local neighbors, i.e., the aggregated feature $\mathbf { Y } _ { i } ^ { k }$ , then the local deviation $\| { \bf Y } _ { i } ^ { k } - ( { \bf X } _ { \mathrm { i n } } ) _ { i } \| _ { 2 }$ will be large, which leads to a $\beta _ { i }$ close to 1. Therefore, the final step will assign a small weight to the residual, i.e., $( 1 - \beta _ { i } ) ( \mathbf { X } _ { \mathrm { i n } } ) _ { i }$ , and the aggregated feature $\mathbf { Y } _ { i } ^ { k }$ will dominate. + +• On the contrary, if $( \mathbf { X } _ { \mathrm { i n } } ) _ { i }$ is already consistent with its local neighbors, $\| { \bf Y } _ { i } ^ { k } - ( { \bf X } _ { \mathrm { i n } } ) _ { i } \| _ { 2 }$ will be small, which leads to a $\beta _ { i }$ close to 0. Thus, the residual will dominate, which is reasonable since there is less need to aggregate features in this case. + +• To summarize, the local deviation $\| { \bf Y } _ { i } ^ { k } - ( { \bf X } _ { \mathrm { i n } } ) _ { i } \| _ { 2 }$ provides a natural transition from $\beta _ { i } \to 1$ to $\beta _ { i } \to 0$ , and the transition can be modulated by $\lambda$ which can be either learned or tuned as a hyperparameter through cross-validation. This transition provides an node-wise adaptive residual connection for the message passing scheme. + +Adaptivity for abnormal $\pmb { \& }$ normal features. According to the homophily assumption on graph structure data [17, 18, 19, 3], the feature representations of normal features should be more consistent with local neighbors than abnormal features. As a result, AMP will assign more residual (i.e., smaller $\beta$ ) to normal features but less residual (i.e., larger $\beta$ ) to abnormal features, providing a customized tradeoff between feature aggregation and residual connection. Consequently, it can promote both the resilience to abnormal features and the performance on normal features. Above discussion also implies a clear physical meaning for $\beta$ in AMP, and we formally define it as the adaptive score. + +Definition 1 (Adaptive score). The variables $\{ \beta _ { 1 } , \cdot \cdot \cdot , \beta _ { n } \}$ in the adaptive message passing scheme (AMP) are defined as the adaptive scores for nodes $\{ v _ { 1 } , \cdots , v _ { n } \}$ respectively in graph $\mathcal { G }$ . In particular, the larger $\beta _ { i }$ is, the more likely the feature of node $v _ { i }$ is abnormal. + +Remark 2 (Nonlinear smoother). Different from most existing message passing scheme which are linear smoothers, AMP is a nonlinear smoother because the weights $\{ \beta _ { i } \}$ are computed from $\mathbf { Y } ^ { k }$ and ${ \bf X } _ { i n }$ . This nonlinearity is the key to achieve adaptive residual connection for different nodes. + +# 3.4 The Model Architecture + +The proposed adaptive message passing (AMP) can be used as a building block in many GNN models to improve the resilience to abnormal node features. In this work, we choose the the decoupled architectures as APPNP [10] and DAGNN [20], and propose the Adaptive residual GNN (AirGNN): + +$$ +\begin{array} { r l } & { \mathbf { X } _ { \mathrm { i n } } = h _ { \theta } ( \mathbf { X } _ { \mathrm { f e a } } ) , } \\ & { \mathbf { Y } _ { \mathrm { p r e } } = \mathbf { A } \mathbf { M } \mathbf { P } \left( \mathbf { X } _ { \mathrm { i n } } , K , \lambda \right) . } \end{array} +$$ + +$h _ { \theta } ( \cdot )$ is any machine learning model parameterized by learnable parameters $\theta$ , such as multilayer perceptrons (MLPs). $\mathbf { X } _ { \mathrm { f e a } } \in \mathbb { R } ^ { n \times d }$ denotes the initial node features. The model $h _ { \theta } ( \cdot )$ will first transform the initial node features as $\mathbf { X } _ { \mathrm { i n } } = h _ { \theta } ( \mathbf { X } _ { \mathrm { f e a } } )$ . AMP takes $h _ { \theta } ( \mathbf { X } _ { \mathrm { f e a } } )$ as input, and performs $K$ steps of AMP with the hyperparameter $\lambda$ . Similar to the majority of existing GNN models, the training objective is the cross-entropy classification loss on the labeled nodes, and the whole model is trained in an end-to-end way. Note that AirGNN is very efficient as explained in Remark 1, and it only requires two hyperparameters $K$ and $\lambda$ without introducing additional parameters to learn, which could reduce the risk of overfitting. + +# 4 Experiment + +In this section, we aim to verify the effective of the proposed adaptive message passing scheme (AMP) and the AirGNN model through the semi-supervised node classification tasks. Specifically, we try to answer the following questions: (1) How does AirGNN perform on abnormal and normal features? (Section 4.2 and 4.3) and (2) How does AirGNN work by adjusting the adaptive residual? (Section 4.4) + +# 4.1 Experimental Settings + +Datasets and baselines. We conduct experiments on 8 real-world datasets including three citation graphs, i.e., Cora, Citeseer, Pubmed [21], two co-authorship graphs, i.e., Coauthor CS and Coauthor Physics [22], two co-purchase graphs, i.e., Amazon Computers and Amazon Photo [22], and one OGB dataset, i.e., ogbn-arxiv [23]. Due to the space limit, we only present the results on Cora, Citeseer, and Pubmed in this section, but defer the results on other datasets to Appendix D.1. More details about the data statistics and data splits are summarized in Appendix B. The proposed AirGNN is compared with representative GNNs, including GCN [3], GAT [6], APPNP [10] and GCNII [9]. We defer the comparison with the variants of APPNP and Robust GCN [24] to Appendix D.3 and D.4 respectively. + +Parameter settings. For all baselines, we follow the best hyperparameter settings in their original papers. Additionally, we tune a best residual weight $\alpha$ for APPNP and GCNII in the range $[ 0 , 1 ]$ . For AirGNN, we use a two-layer MLP as the base model $h _ { \theta } ( \cdot )$ , following APPNP. We fix the learning rate 0.01, dropout 0.8, and weight decay 0.0005. Moreover, we set $\begin{array} { r } { \gamma = \frac { 1 } { 2 ( 1 - \lambda ) } } \end{array}$ as suggested by Theorem 1. We choose $K = 1 0$ and tune $\lambda$ in the range $[ 0 , 1 ]$ . Adam optimizer [25] is used in all experiments. We run all experiments by 10 times, and report the mean and variance. + +Evaluation setting. We assess the performance of all models under two types of abnormal feature scenarios, including noisy features and adversarial features. The abnormal features are injected to randomly selected test nodes after model training. By default, all hyperparameters are tuned according to the performance on validation sets when the dataset is clean. If tuning the hyperparameter $\lambda$ of AirGNN according to the validation sets after injecting abnormal features, the performance will be even better, as discussed in Appendix D.2. The performance on clean data are showed in Appendix D.5 to demonstrate that AirGNN doesn’t need to sacrifice accuracy for better robustness against abnormal features. + +# 4.2 Performance Comparison with Noisy Features + +In this subsection, we consider the abnormal features in the noisy feature scenario. Specifically, we simulate the noisy features by assigning a subset of the nodes with random features sampled from a multivariate standard Gaussian distribution. Note that the selection of noise subsets has a apparent impact on the performance since some nodes are less vulnerable to abnormal features while others are more vulnerable. To reduce such variance, we report the average performance over 10 times of random selection of the noise node sets, similar to the settings in the preliminary study in Section 2. We report the node classification test accuracy on abnormal (noisy) features and normal features in Figure 5 and Figure 6, separately, under varying noisy ratio. From these figures, we can observe: + +• Figure 5 shows that AirGNN significantly outperforms all baselines on all datasets in terms of the performance on noisy nodes. This verifies that AMP is able to improve the resilience to noisy features, aligning well with the design motivation. +• Figure 6 shows that AirGNN promotes the performance on normal nodes when abnormal nodes exist. This is because AMP can remove some abnormal features which are detrimental to normal nodes. + +![](images/cb981f452a6f87e6cccc28935d8e1aa89643dc8d85a1008f917f9b4fedf8a508.jpg) +Figure 5: Node classification accuracy on abnormal (noisy) nodes + +# 4.3 Performance Comparison with Adversarial Features + +In this subsection, we consider the abnormal feature scenario when the node features are maliciously attacked by the attacker to manipulate the prediction of GNNs. We use the Nettack [26] implemented in DeepRobust3 [27], a PyTorch library for adversarial attacks and defenses, to generate the adversarial features. We randomly choose 40 test nodes as the targeted nodes, and assess the performance under increasing perturbation budgets $\{ 0 , 5 , 1 0 , 2 0 , 5 0 , 8 0 \}$ , where the perturbation numbers denote the number of feature dimensions that can be manipulated. The node classification accuracy on these attacked nodes are showed in Figure 7. From these figures, we can make the following observations: + +![](images/521af9135379cb9d9a4e04bfd77d43979bb91c8146837931dfc5f87ae829d691.jpg) +Figure 6: Node classification accuracy on normal nodes + +• AirGNN is significantly more robust against adversarially attacked features than all baselines. MLP is the most vulnerable model, which demonstrates the usefulness of graph structure information in combating against abnormal node features. + +• The advantages of AirGNN over the baselines become much stronger with larger perturbation budgets. This suggests that AMP can significantly improve the resilience to abnormal features. + +![](images/b4a84b017c1ab3aee2d1fd5a97a59af3692c895e3672220068ba8be20564b118.jpg) +Figure 7: Node classification accuracy on adversarial nodes + +# 4.4 Adaptive Residual for Abnormal & Normal Nodes + +To further understand and verify how AMP and AirGNN work, we investigate the adaptive score $\beta _ { i }$ for each node $v _ { i }$ . Specifically, the average adaptive scores for abnormal nodes and normal nodes in the last layer of AMP are computed separately. In the noisy feature scenario, we fix ratio of noisy nodes as $10 \%$ . In the adversarial feature scenario, we choose 40 target nodes and fix the perturbation number as 80. The results in noisy and adversarial feature scenarios are showed in Table 1 and Table 2, respectively. From these tables, we can observe: + +• On the one hand, it can be clearly observed that in both scenarios, the average adaptive scores for abnormal nodes are significantly higher than those for normal nodes. Therefore, it verifies our intuition that large adaptive scores are strongly related to abnormal features. • On the other hand, it also implies that the residual weights (i.e., $1 - \beta _ { i } )$ for abnormal nodes are much lower than those of normal nodes. This perfectly aligns with our motivation to remove abnormal features by reducing their residual connections. + +The study on adaptive scores verifies how the adaptive residuals in AMP and AirGNN work as designed. It corroborates that AirGNN not only tremendously boosts the resilience to abnormal features but also provides interpretable information for anomaly detection that will be useful in many security-critical scenarios since the adaptive score serves as a good indicator of abnormal nodes. Morever, it is expected that APPNP without residual will perform well on abnormal nodes but it will sacrifice the performance on normal nodes. We provide detailed comparison with APPNP w/Res and APPNP wo/Res in Appendix D.3 to show the advantages of adaptive residual of AirGNN. + +Table 1: Average adaptive score $( \beta )$ and residual weight $( 1 - \beta )$ in the noisy feature scenario. + +
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes0.998 ± 0.0000.988 ± 0.0000.996 ± 0.000
Average adaptive score for normal nodes Average residual weight for abnormal nodes0.924 ± 0.002 0.002 ±0.0000.807 ± 0.005 0.012 ± 0.0000.869 ± 0.006
Average residual weight for normal nodes0.076 ± 0.0020.193 ± 0.0050.004±0.000 0.131 ± 0.006
+ +Table 2: Average adaptive score $( \beta )$ and residual weight $( 1 - \beta )$ in the adversarial feature scenario. + +
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes Average adaptive score for normal nodes0.987 ±0.0000.930 ± 0.0070.959 ± 0.005
Average residual weight for abnormal nodes0.922 ± 0.0040.689 ± 0.0240.826 ± 0.016
0.013 ±0.0000.070±0.0070.041 ±0.005
Average residual weight for normal nodes0.078 ± 0.0040.311 ± 0.0240.174 ± 0.016
+ +# 5 Related Work + +GNNs generalize convolutional neural networks (CNN) to graph structure data through the message passing framework [1, 2, 7]. The design of message passing and GNN architectures are majorly motivated in spectral domain [3, 4] and spatial domain [5, 6, 7, 2]. Recent works have shown that the message passing in GNNs can be regarded as low-pass graph filters [12, 13]. More generally, it has been proven that message passing in many GNNs can be uniformly derived from graph signal denoising [11, 28, 29, 30]. Classic GNNs such as GCN [3] and GAT [6] achieve their best performance with shallow models, but their performance degrades when stacking more layers, which can be partially explained through oversmoothing analyses [14, 15]. Recent works propose to use residual connections or skip connections to mitigate the oversmoothing issues, and they demonstrate the potential benefits from more feature aggregations. Examples include but not limited to DeepGCNs [31], JKNet [32], GCNII [9], APPNP [10] and DeeperGNN [20]. These models use global residual connection that can not be adaptive for each node, which significantly differ from the proposed AirGNN. + +Recently, there are growing interests in reducing GNNs’ vulnerability to the graph structure noise, such as Robust GCN [24], GCN-SVD [33], Pro-GNN [34], IDGL [35], ElasticGNN [36], etc. Please refer to the comprehensive surveys [37, 38] for more details. However, how to design GNNs with strong resilience to abnormal node features remains to be developed. To the best of our knowledge, AirGNN is the first GNN model that is intrinsically robust to many types of abnormal node features by design. It improves the performance in various kinds of abnormal scenarios without needing to sacrifice clean accuracy in normal settings. + +# 6 Conclusion + +In this work, we discover an intrinsic tension between feature aggregation and residual connection in the message passing scheme of GNNs, as well as the corresponding performance tradeoff between nodes with abnormal and normal features. We analyze possible reasons to explain these findings from the perspective of graph Laplacian smoothing. Our understandings further motivate us to propose a simple, efficient, interpretable and adaptive message passing scheme as well as a new GNN model with adaptive residual, named AirGNN. AirGNN provides a node-wise adaptive transition between feature aggregation and residual connection, and the significant advantages of AirGNN are demonstrated through extensive experiments. + +# Acknowledgments and Disclosure of Funding + +This research is supported by the National Science Foundation (NSF) under grant numbers IIS1714741, CNS1815636, IIS1845081, IIS1907704, DRL2025244, IIS1928278, IIS1955285, + +IOS2107215, IOS2035472 and Army Research Office (ARO) under grant number W911NF-21- 1-0198. + +# Societal Impact and Limitations + +The methodology proposed in this paper might have significant positive societal impact since it reduces machine learning models’ vulnerability to abnormal datasets that generally exist in real-world applications, especially in many security-critical scenarios. In practice, for a given graph, we do not have the prior knowledge about if the graph is clean, has noisy features or adversarial features. Therefore, algorithms like the proposed AirGNN that can work under both the clean and various abnormal feature settings are appealing. While we are unaware of any potential negative society impact, we point out two limitations of this work: (1) this paper focuses on abnormal node features and has not evaluated the performance of the proposed method when the dataset contains both abnormal node features and edges; (2) it is unclear how it performs on heterophilic graphs. It will be interesting to investigate these problems in future works. + +# References + +[1] Yao Ma and Jiliang Tang. Deep Learning on Graphs. Cambridge University Press, 2020. +[2] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International Conference on Machine Learning, pages 1263–1272. PMLR, 2017. +[3] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. +[4] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 3844–3852, 2016. +[5] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. arXiv preprint arXiv:1706.02216, 2017. +[6] Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua ´ Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. +[7] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61–80, 2008. +[8] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 2020. +[9] Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1725–1735. PMLR, 13–18 Jul 2020. +[10] Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate: Graph neural networks meet personalized pagerank. arXiv preprint arXiv:1810.05997, 2018. +[11] Yao Ma, Xiaorui Liu, Tong Zhao, Yozen Liu, Jiliang Tang, and Neil Shah. A unified view on graph neural networks as graph signal denoising. Proceedings of the 30th ACM International Conference on Information and Knowledge Management, 2021. +[12] Hoang Nt and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass filters. arXiv preprint arXiv:1905.09550, 2019. +[13] Lingxiao Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. In International Conference on Learning Representations, 2019. +[14] Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. +[15] Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. In International Conference on Learning Representations, 2020. +[16] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. +[17] Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in social networks. Annual review of sociology, 27(1):415–444, 2001. +[18] Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propagation. +[19] Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. 2004. +[20] Meng Liu, Hongyang Gao, and Shuiwang Ji. Towards deeper graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, 2020. +[21] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina EliassiRad. Collective classification in network data. AI magazine, 29(3):93–93, 2008. +[22] Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868, 2018. +[23] Marinka Zitnik Yuxiao Dong Hongyu Ren Bowen Liu Michele Catasta Jure Leskovec Weihua Hu, Matthias Fey. Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687, 2020. +[24] Dingyuan Zhu, Ziwei Zhang, Peng Cui, and Wenwu Zhu. Robust graph convolutional networks against adversarial attacks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1399–1407, 2019. +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[26] Daniel Zügner, Amir Akbarnejad, and Stephan Günnemann. Adversarial attacks on neural networks for graph data. In KDD. ACM, 2018. +[27] Yaxin Li, Wei Jin, Han Xu, and Jiliang Tang. Deeprobust: A pytorch library for adversarial attacks and defenses, 2020. +[28] Xuran Pan, Song Shiji, and Huang Gao. A unified framework for convolution-based graph neural networks. https://openreview.net/forum?id=zUMD–Fb9Bt, 2020. +[29] Meiqi Zhu, Xiao Wang, Chuan Shi, Houye Ji, and Peng Cui. Interpreting and unifying graph neural networks with an optimization framework, 2021. +[30] Siheng Chen, Yonina C. Eldar, and Lingxiao Zhao. Graph unrolling networks: Interpretable neural networks for graph signal denoising, 2020. +[31] Guohao Li, Matthias Müller, Ali Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep as cnns? In The IEEE International Conference on Computer Vision (ICCV), 2019. +[32] Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 5453–5462. PMLR, 10–15 Jul 2018. +[33] Negin Entezari, Saba A Al-Sayouri, Amirali Darvishzadeh, and Evangelos E Papalexakis. All you need is low (rank) defending against adversarial attacks on graphs. In Proceedings of the 13th International Conference on Web Search and Data Mining, pages 169–177, 2020. +[34] Wei Jin, Yao Ma, Xiaorui Liu, Xianfeng Tang, Suhang Wang, and Jiliang Tang. Graph structure learning for robust graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 66–74, 2020. +[35] Yu Chen, Lingfei Wu, and Mohammed Zaki. Iterative deep graph learning for graph neural networks: Better and robust node embeddings. Advances in Neural Information Processing Systems, 33, 2020. +[36] Xiaorui Liu, Wei Jin, Yao Ma, Yaxin Li, Hua Liu, Yiqi Wang, Ming Yan, and Jiliang Tang. Elastic graph neural networks. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 6837–6849. PMLR, 18–24 Jul 2021. +[37] Wei Jin, Yaxing Li, Han Xu, Yiqi Wang, Shuiwang Ji, Charu Aggarwal, and Jiliang Tang. Adversarial attacks and defenses on graphs. ACM SIGKDD Explorations Newsletter, 22(2):19– 34, 2021. +[38] Yanqiao Zhu, Weizhi Xu, Jinghao Zhang, Qiang Liu, Shu Wu, and Liang Wang. Deep graph structure learning for robust representations: A survey. arXiv preprint arXiv:2103.03036, 2021. +[39] Laurent Condat. A primal–dual splitting method for convex optimization involving lipschitzian, proximable and linear composite terms. Journal of optimization theory and applications, 158(2):460–479, 2013. +[40] Fan RK Chung and Fan Chung Graham. Spectral graph theory. Number 92. American Mathematical Soc., 1997. + +# Checklist + +1. For all authors... + +(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] +(b) Did you describe the limitations of your work? [Yes] +(c) Did you discuss any potential negative societal impacts of your work? [Yes] +(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] + +2. If you are including theoretical results... + +(a) Did you state the full set of assumptions of all theoretical results? [Yes] (b) Did you include complete proofs of all theoretical results? [Yes] + +3. If you ran experiments... + +(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] +(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] +(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] +(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No] + +4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... + +(a) If your work uses existing assets, did you cite the creators? [Yes] +(b) Did you mention the license of the assets? [No] +(c) Did you include any new assets either in the supplemental material or as a URL? [No] +(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [No] +(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [No] + +5. If you used crowdsourcing or conducted research with human subjects... + +(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] +(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] +(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] \ No newline at end of file diff --git a/parse/train/hfkER_KJiNw/hfkER_KJiNw_content_list.json b/parse/train/hfkER_KJiNw/hfkER_KJiNw_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f4fa5c900c429baa11b13c3f95211c1fdc79dca7 --- /dev/null +++ b/parse/train/hfkER_KJiNw/hfkER_KJiNw_content_list.json @@ -0,0 +1,1577 @@ +[ + { + "type": "text", + "text": "Graph Neural Networks with Adaptive Residual ", + "text_level": 1, + "bbox": [ + 205, + 122, + 794, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaorui Liu1 xiaorui@msu.edu ", + "bbox": [ + 232, + 200, + 364, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiayuan Ding1 dingjia5@msu.edu ", + "bbox": [ + 428, + 200, + 568, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Jin1 jinwei2@msu.edu ", + "bbox": [ + 633, + 200, + 764, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Han Xu1 xuhan1@msu.edu ", + "bbox": [ + 187, + 250, + 308, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yao Ma2 yao.ma@njit.edu ", + "bbox": [ + 331, + 250, + 460, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zitao Liu3 liuzitao@100tal.com ", + "bbox": [ + 488, + 250, + 643, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiliang Tang1 tangjili@msu.edu ", + "bbox": [ + 669, + 250, + 808, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Michigan State University, East Lansing, MI, USA 2New Jersey Institute of Technology, Newark, NJ, USA 3TAL Education Group, Beijing, China ", + "bbox": [ + 315, + 296, + 681, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 375, + 535, + 391 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graph neural networks (GNNs) have shown the power in graph representation learning for numerous tasks. In this work, we discover an interesting phenomenon that although residual connections in the message passing of GNNs help improve the performance, they immensely amplify GNNs’ vulnerability against abnormal node features. This is undesirable because in real-world applications, node features in graphs could often be abnormal such as being naturally noisy or adversarially manipulated. We analyze possible reasons to understand this phenomenon and aim to design GNNs with stronger resilience to abnormal features. Our understandings motivate us to propose and derive a simple, efficient, interpretable, and adaptive message passing scheme, leading to a novel GNN with Adaptive residual, AirGNN1. Extensive experiments under various abnormal feature scenarios demonstrate the effectiveness of the proposed algorithm. ", + "bbox": [ + 233, + 406, + 766, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction ", + "text_level": 1, + "bbox": [ + 174, + 597, + 310, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed the great success of graph neural networks (GNNs) in representation learning for graph structure data [1]. Essentially, GNNs generalize deep neural networks (DNNs) from regular grids, such as image, video and text, to irregular data such as social, energy, transportation, citation, and biological networks. Such data can be naturally represented as graphs with nodes and edges. The key building block for such generalization is the neural message passing framework [2]: ", + "bbox": [ + 173, + 628, + 825, + 699 + ], + "page_idx": 0 + }, + { + "type": "equation", + "img_path": "images/6f9e56a3d9b3fdbb7f89623ce9d7ae036869e671c7b87fc24582e7a11728195c.jpg", + "text": "$$\n\\mathbf { x } _ { u } ^ { ( k + 1 ) } = \\mathrm { U P D A T E } ^ { ( k ) } \\left( \\mathbf { x } _ { u } ^ { ( k ) } , \\mathbf { m } _ { \\mathcal { N } ( u ) } ^ { ( k ) } \\right)\n$$", + "text_format": "latex", + "bbox": [ + 374, + 704, + 620, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where x(ku $\\mathbf { x } _ { u } ^ { ( k ) } \\in \\mathbb { R } ^ { d }$ denotes the feature vector of node $u$ in the $k$ -th iteration of message passing, and $\\mathbf { m } _ { \\mathcal { N } ( u ) } ^ { ( k ) }$ is the message aggregated from $u$ ’s neighborhood $\\mathcal { N } ( u )$ . The specific design of message passing scheme can be motivated from spectral domain [3, 4] or spatial domain [5, 6, 7, 2]. It usually linearly smooths the features in a local neighborhood on the graph. ", + "bbox": [ + 174, + 736, + 825, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GNNs have achieved superior performance in a large number of benchmark datasets [8] where the node features are assumed to be complete and informative. However, in real-world applications, some node features could be abnormal from various aspects. For instance, in social networks, new users might not have complete profile before they make connections with others, leading to missing user features. In transportation networks, node features can be noisy since there exist certain uncertainty and dynamics in the observation of the traffic information. What is worse, node features can be adversarially chosen by the attacker to maliciously manipulate the prediction made by GNNs. Therefore, it is greatly desired to design GNN models with stronger resilience to abnormal node features. ", + "bbox": [ + 174, + 806, + 825, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 90, + 825, + 147 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we first perform empirical investigations on how representative GNN models behave on graphs with abnormal features. Specifically, based upon standard benchmark datasets, we simulate the abnormal features by replacing the features of randomly selected nodes with random Gaussian noise. Then the performance of node classification on abnormal features and normal features are examined separately. From our preliminary study in Section 2, we reveal two interesting observations: (1) Feature aggregation can boost the resilience to abnormal features, but too many aggregations could hurt the performance on both normal and abnormal features; and (2) Residual connection helps GNNs benefit from more layers for normal features, while making GNNs more fragile to abnormal features. We then provide possible explanations to understand these observed phenomena from the perspective of graph Laplacian smoothing. Our analyses imply that there might exist an intrinsic tension between feature aggregation and residual connection, which results in a performance tradeoff between normal features and abnormal features. ", + "bbox": [ + 173, + 154, + 825, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these findings and understandings, we aim to design new GNNs with stronger resilience to abnormal features while largely maintaining the performance on normal features. Our contributions can be summarized as follows: ", + "bbox": [ + 176, + 325, + 825, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We discover an intrinsic tension between feature aggregation and residual connection in GNNs, and the corresponding performance tradeoff between abnormal and normal features. We also analyze possible reasons to explain and understand these findings. \n• We propose a simple, efficient, principled and adaptive message passing scheme, which leads to a novel GNN model with adaptive residual, named as AirGNN. \n• Extensive experiments under various abnormal feature scenarios demonstrate the superiority of the proposed algorithm. The ablation study demonstrates how the adaptive residuals mitigate the impact of abnormal features. ", + "bbox": [ + 194, + 378, + 826, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Preliminary ", + "text_level": 1, + "bbox": [ + 174, + 516, + 305, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before introducing the preliminary study, we first define the notations used throughout the paper. ", + "bbox": [ + 173, + 546, + 805, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notations. We use bold upper-case letters such as $\\mathbf { X }$ to denote matrices. Given a matrix $\\mathbf { X } \\in \\mathbb { R } ^ { n \\times d }$ , we use $\\mathbf { X } _ { i }$ to denote its $i$ -th row and $\\mathbf { X } _ { i j }$ to denote its element in $i$ -th row and $j$ -th column. The Frobenius norm and $\\ell _ { 2 1 }$ norm of a matrix $\\mathbf { X }$ are defined as $\\Vert \\mathbf { X } \\Vert _ { F } = \\sqrt { \\sum _ { i j } \\mathbf { X } _ { i j } ^ { 2 } }$ and $\\| \\mathbf { X } \\| _ { 2 1 } =$ $\\begin{array} { r } { \\sum _ { i } \\| \\mathbf { X } _ { i } \\| _ { 2 } = \\sum _ { i } \\sqrt { \\sum _ { j } \\mathbf { X } _ { i j } ^ { 2 } } } \\end{array}$ , respectively. We define $\\| \\mathbf { X } \\| _ { 2 } = \\sigma _ { \\operatorname* { m a x } } ( \\mathbf { X } )$ where $\\sigma _ { \\mathrm { m a x } } ( \\mathbf { X } )$ is the largest singular value of $\\mathbf { X }$ . ", + "bbox": [ + 173, + 568, + 825, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $\\mathcal { G } = \\{ \\nu , \\mathcal { E } \\}$ be a graph with the node set $\\mathcal { V } = \\{ v _ { 1 } , \\ldots , v _ { n } \\}$ and the undirected edge set $\\mathcal { E } = \\{ e _ { 1 } , \\ldots , e _ { m } \\}$ . We use $\\mathcal { N } ( v _ { i } )$ to denote the neighboring nodes of node $v _ { i }$ , including $v _ { i }$ itself. Suppose that each node is associated with a $d$ -dimensional feature vector, and the features for all nodes are denoted as $\\mathbf { X } _ { \\mathrm { f e a } } \\in \\mathbb { R } ^ { n \\times d }$ . The graph structure $\\mathcal { G }$ can be represented as an adjacent matrix $\\mathbf { A } \\in \\mathbb { R } ^ { n \\times n }$ , where $\\mathbf { A } _ { i j } ~ = ~ 1$ when there exists an edge between nodes $v _ { i }$ and $v _ { j }$ , and $\\mathbf { A } _ { i j } \\ = \\ 0$ otherwise. The graph Laplacian matrix is defined as $\\mathbf { L } = \\mathbf { D } - \\mathbf { A }$ , where $\\mathbf { D }$ is the diagonal degree matrix. Let us denote the commonly used feature aggregation matrix in GNNs [3] as $\\tilde { \\bf A } = \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } \\bar { \\hat { \\bf A } } \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } }$ where $\\hat { \\mathbf { A } } = \\mathbf { A } + \\mathbf { I }$ is the adjacent matrix with self-loop and its degree matrix is $\\hat { \\bf D }$ . The corresponding Laplacian matrix is defined as $\\tilde { \\mathbf { L } } = \\mathbf { I } - \\tilde { \\mathbf { A } }$ . ", + "bbox": [ + 173, + 664, + 825, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on the setting where a subset of nodes in the graph contain abnormal features, while the remaining nodes have normal features. In the remaining of this paper, we use abnormal/normal features to denote nodes with abnormal/normal features, for simplicity. ", + "bbox": [ + 174, + 800, + 825, + 842 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Preliminary Study ", + "text_level": 1, + "bbox": [ + 174, + 857, + 341, + 872 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental setup. To investigate how GNNs behave on abnormal and normal node features, we design semi-supervised node classification experiments on three common datasets (i.e., Cora, ", + "bbox": [ + 173, + 882, + 823, + 911 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CiteSeer and PubMed), following the data splits in the work [3]. Moreover, we simulate the abnormal features by assigning $1 0 \\%$ of the nodes with random features sampled from a standard Gaussian distribution. The experiments are performed on representative GNN models covering coupled and decoupled architectures, including GCN [3], GCNII [9], APPNP [10], and their variants with or without residual connections in feature aggregations, denoted as w/Res and wo/Res. All methods follow the hyperparameter settings in their original papers. We examine how these models perform when the number of layers increases. Note that for the decoupled architectures such as APPNP, we fix the 2-layer MLP and increase the number of propagation layers. While for the coupled architectures such as GCN and GCNII, we increase the number of feature transformation and propagation layers simultaneously. We report the average performance over 10 times of random selection of the noise node sets. The node classification accuracy (mean and standard variance) on nodes with abnormal and normal features is illustrated in Figure 1 and Figure. 2, separately. ", + "bbox": [ + 173, + 90, + 825, + 257 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/00d120485edd7da85f1457dd4c00230b18d99c238c177b5cb16e29a69a61afb9.jpg", + "image_caption": [ + "Figure 1: Node classification accuracy on abnormal nodes (Cora) " + ], + "image_footnote": [], + "bbox": [ + 184, + 280, + 812, + 415 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08499e09f3e29a9b457146ea9c203108450fafd94446152fa1b7aada0f5024ce.jpg", + "image_caption": [ + "Figure 2: Node classification accuracy on normal nodes (Cora) " + ], + "image_footnote": [], + "bbox": [ + 186, + 474, + 812, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observations. From Figure 1 and Figure 2, we can make the following observations: (1) Without residual connection, more layers (e.g., $> 2$ for GCN and GCNII, $> 1 0$ for APPNP) hurt the accuracy on nodes with normal features. However, more layers boost the accuracy on nodes with abnormal features significantly, before finally starting to decrease; (2) With residual connection, the accuracy on nodes with normal features keeps increasing with more layers2. However, the accuracy on nodes with abnormal features only increases marginally when stacking more layers, and then starts to decrease. While we only present the experiments on Cora, we defer the results on other datasets to Appendix C, which provide similar observations. To conclude, we can summarize these observations into two major findings: ", + "bbox": [ + 173, + 657, + 825, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "• Finding I: Feature aggregation can boost the resilience to abnormal features, but too many aggregations could hurt the performance on both normal and abnormal nodes; ", + "bbox": [ + 191, + 797, + 823, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "• Finding II: Residual connection helps GNNs benefit from more layers for nodes with normal features, while making GNNs more fragile to abnormal features. ", + "bbox": [ + 192, + 837, + 820, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Understandings ", + "text_level": 1, + "bbox": [ + 174, + 90, + 323, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we provide the understanding and explanation for aforementioned findings, from the perspective of graph Laplacian smoothing. ", + "bbox": [ + 173, + 116, + 823, + 145 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Understanding Finding I: Feature aggregation as Laplacian smoothing ", + "text_level": 1, + "bbox": [ + 176, + 150, + 669, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The message passing in GCN [3], GCNII wo/ residual and APPNP wo/ residual (as well as many popular GNN models), follows the feature aggregation ", + "bbox": [ + 174, + 171, + 825, + 200 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/5be41b7d9771905b48713d3d8dca4d437d51c9c6e6be8231f4869bf05b49f235.jpg", + "text": "$$\n\\mathbf { X } _ { \\mathrm { o u t } } = \\tilde { \\mathbf { A } } \\mathbf { X } _ { \\mathrm { i n } } ,\n$$", + "text_format": "latex", + "bbox": [ + 449, + 205, + 547, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where ${ \\bf X } _ { \\mathrm { i n } }$ and $\\mathbf { X _ { o u t } }$ represent the features before and after message passing layer, respectively. It can be interpreted as one gradient descent step for the Laplacian smoothing problem [11] ", + "bbox": [ + 166, + 231, + 823, + 260 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/a8ba35e5f2a5c87dc5a4f1ef62b9e0fcb76a09d3a2e3b4ba8193f99725d8ad32.jpg", + "text": "$$\n\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\mathcal { L } _ { 1 } ( \\mathbf { X } ) : = \\frac { 1 } { 2 } \\mathrm { t r } \\Big ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } \\Big ) = \\frac { 1 } { 2 } \\sum _ { ( v _ { i } , v _ { j } ) \\in \\mathcal { E } } \\| \\frac { \\mathbf { X } _ { i } } { \\sqrt { d _ { i } + 1 } } - \\frac { \\mathbf { X } _ { j } } { \\sqrt { d _ { j } + 1 } } \\| _ { 2 } ^ { 2 } ,\n$$", + "text_format": "latex", + "bbox": [ + 225, + 263, + 769, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d _ { i }$ is the node degree of node $v _ { i }$ . Eq. (2) can be derived from $\\mathbf { X } _ { \\mathrm { o u t } } = \\mathbf { X } _ { \\mathrm { i n } } - ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } _ { \\mathrm { i n } } = \\tilde { \\mathbf { A } } \\mathbf { X } _ { \\mathrm { i n } }$ , with the initialization $\\mathbf { X } = \\mathbf { X } _ { \\mathrm { i n } }$ and stepsize $\\gamma = 1$ . The Laplacian smoothing problem penalizes the feature difference between neighboring nodes. To reduce this penalty, the feature aggregation in Eq. (2) smooths the node features by taking the average of local neighbors, and thus can be considered as low-pass filter which gradually filters out high-frequency signals [12, 13]. Therefore, it increases the resilience to abnormal features which are likely to be high-frequency signals. In other words, the local neighboring nodes help to correct the abnormal features. Unfortunately, if applied too many times, these low-pass filters could overly smooth the features (well-known as oversmoothing [14, 15]) such that nodes are not distinguishable enough, providing an explanation to the degraded performance on both abnormal and normal features when stacking too many layers. ", + "bbox": [ + 173, + 313, + 825, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Understanding Finding II: Residual connection maintains feature proximity ", + "text_level": 1, + "bbox": [ + 176, + 458, + 704, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To adjust the feature smoothness for better performance, APPNP [10] utilizes residual connections in message passing as follows ", + "bbox": [ + 173, + 478, + 823, + 507 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/4bd63d5a2bf0fef5bca13df83fe1b77403e60662228148abd44885dde98f7b30.jpg", + "text": "$$\n\\mathbf { X } ^ { k + 1 } = ( 1 - \\alpha ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } + \\alpha \\mathbf { X _ { \\mathrm { i n } } } ,\n$$", + "text_format": "latex", + "bbox": [ + 392, + 512, + 602, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where ${ \\bf X } ^ { 0 } = { \\bf X } _ { \\mathrm { i n } }$ . It can be considered as an iterative solution for the regularized Laplacian smoothing problem [11] ", + "bbox": [ + 173, + 539, + 825, + 569 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/dc684fa92c8954def72ee007cc1a50155f45d52e5b35867cbf9e129b802e67a4.jpg", + "text": "$$\n\\underset { { \\substack { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } _ { 2 } ( { \\mathbf { X } } ) : = \\frac { \\alpha } { 2 ( 1 - \\alpha ) } \\| \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } } \\| _ { F } ^ { 2 } + \\frac { 1 } { 2 } \\mathrm { t r } \\Big ( \\mathbf { X } ^ { \\top } ( { \\mathbf { I } } - \\tilde { \\mathbf { A } } ) \\mathbf { X } \\Big ) ,\n$$", + "text_format": "latex", + "bbox": [ + 272, + 573, + 722, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with initialization $\\mathbf { X } = \\mathbf { X } _ { \\mathrm { i n } }$ and stepsize $\\gamma = 1 - \\alpha$ due to ", + "bbox": [ + 173, + 613, + 557, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/b21d10dc68f05ed9a814f9fe8eb81efce214601248cf9f4bc15d2afa481cfc3d.jpg", + "text": "$$\n\\mathbf { X } ^ { k + 1 } = \\mathbf { X } ^ { k } - ( 1 - \\alpha ) \\bigg ( \\frac { \\alpha } { 1 - \\alpha } ( \\mathbf { X } ^ { k } - \\mathbf { X _ { \\mathrm { i n } } } ) + ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ^ { k } \\bigg ) = ( 1 - \\alpha ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } + \\alpha \\mathbf { X _ { \\mathrm { i n } } } .\n$$", + "text_format": "latex", + "bbox": [ + 215, + 632, + 782, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GCNII [9] adopts a similar message passing but further combines a feature transformation layer in each message passing step, which leads to a coupled architecture, as contrast to the decoupled architecture of APPNP. The residual connection naturally arises when regularizing the proximity between input and output features, as showed in the first term of $\\mathcal { L } _ { 2 } ( \\mathbf { X } )$ . Such proximity can help avoid the trivial solution for the problem in Eq. (3), i.e., totally oversmoothed features only depending on node degrees, and consequently mitigates the oversmoothing issue. More intuitively, residual connections in GNNs provide direct information flows between layers that can preserve some necessary high-frequency signals for better discrimination between classes. More layers with residual provide a more accurate solution to Eq. (5), which explains the performance gain from deeper GNNs. Unfortunately, these residual connections also undesirably carry on abnormal features which are detrimental, leading to the inferior performance on abnormal features. ", + "bbox": [ + 173, + 666, + 826, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 The Proposed Framework ", + "text_level": 1, + "bbox": [ + 176, + 838, + 424, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first motivate the proposed adaptive message passing scheme (AMP) with further discussions on our preliminary study. We then introduce more details about AMP, its interpretations, convergence guarantee and computation complexity, as well as the model architecture of AirGNN. ", + "bbox": [ + 176, + 869, + 825, + 911 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Design Motivation ", + "text_level": 1, + "bbox": [ + 174, + 90, + 341, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our preliminary study in Section 2 reveals an intrinsic tension between feature aggregation and residual connection: (1) feature aggregation helps smooth out abnormal features, while it could cause inappropriate smoothing for normal features; (2) residual connection is essential for adjusting the feature smoothness, but it could be detrimental for abnormal features. Although this conflict can be partially mitigated by adjusting the residual connection such as the residual weight $\\alpha$ in GCNII [9] and APPNP [10], such global adjustment cannot be adaptive to a subset of the nodes, e.g., the nodes with abnormal features. This is crucial because in practice we often encounter the scenario where only a subset of nodes contain abnormal features. Therefore, how to reconcile this dilemma still desires dedicated efforts. We then naturally ask a question: Can we design a better message passing scheme with node-wise adaptive feature aggregation and residual connection? ", + "bbox": [ + 173, + 116, + 825, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The motivation of the proposed idea builds upon the following intuition: while it is important to maintain the proximity between input and output features as in Eq. (5), it could be over aggressive to penalize their deviations by the square of Frobenius norm, i.e., $\\begin{array} { r } { \\| \\dot { \\bf X } - \\dot { \\bf X } _ { \\mathrm { i n } } \\| _ { F } ^ { 2 } = \\sum _ { i = 1 } ^ { n } \\| \\dot { \\bf X } _ { i } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } ^ { 2 } } \\end{array}$ The fact that this penalty does not tolerate large deviations weakens the capability to remove abnormal features through Laplacian smoothing. This motivates us to consider an alternative proximity penalty ", + "bbox": [ + 173, + 260, + 825, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/2f28e684de6375dcb669923823ae95ee31b28bb7d9ad48bed4a2a0043c2958b9.jpg", + "text": "$$\n\\| { \\bf X } - { \\bf X } _ { \\mathrm { i n } } \\| _ { 2 1 } : = \\sum _ { i = 1 } ^ { n } \\| { \\bf X } _ { i } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } ,\n$$", + "text_format": "latex", + "bbox": [ + 372, + 334, + 624, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which instead penalizes the deviations by the $\\ell _ { 1 }$ norm of row-wise $\\ell _ { 2 }$ norms, namely $\\ell _ { 2 1 }$ norm. The $\\ell _ { 2 1 }$ norm promotes row sparsity in $\\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } }$ , and it also allows large deviations because the penalty on large values is less aggressive, leading to the potential removal of abnormal features. Therefore, we propose the following Laplacian smoothing problem regularized by $\\ell _ { 2 1 }$ norm proximity control: ", + "bbox": [ + 173, + 378, + 825, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/b26e0b4245ae411a1f22ecbedf9e93e2abbac5e9c18379c4b68de77e7bd76391.jpg", + "text": "$$\n\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\lambda \\| \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } } \\| _ { 2 1 } + \\frac { 1 } { 2 } \\mathrm { t r } ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 341, + 438, + 655, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda \\in [ 0 , \\infty )$ is a parameter to adjust the balance between proximity and Laplacian smoothing. In order to easy the tuning of $\\lambda$ , we made a modification of Eq. (7): ", + "bbox": [ + 171, + 477, + 823, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/1e2687fdb8986a4f9b310d05308855779adba69c619b7a25128f76c35752f2d4.jpg", + "text": "$$\n\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } ( \\mathbf { X } ) : = \\lambda \\| \\mathbf { X } - \\mathbf { X _ { \\mathrm { i n } } } \\| _ { 2 1 } + ( 1 - \\lambda ) \\mathrm { t r } ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 284, + 510, + 707, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda \\in [ 0 , 1 ]$ controls the balance. ", + "bbox": [ + 174, + 542, + 419, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Adaptive Message Passing ", + "text_level": 1, + "bbox": [ + 173, + 573, + 397, + 588 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b9c08a90235cf9dc7aa82eaeface75a6b91e84c8306c75e6acc04630e39b84db.jpg", + "image_caption": [ + "Figure 3: Diagram of Adaptive Message Passing " + ], + "image_footnote": [], + "bbox": [ + 240, + 601, + 754, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal { L } ( \\mathbf { X } )$ is a composite objective with non-smooth and smooth components. We optimize it by proximal gradient descent [16] and obtain the following iterations as the adaptive message passing (AMP): ", + "bbox": [ + 168, + 713, + 823, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/04047c793794adee67acacd26cda21357adc84912531b7a797e778030c221957.jpg", + "text": "$$\n\\begin{array} { r l } { { \\displaystyle { \\bf Y } ^ { k } = { \\bf X } ^ { k } - 2 \\gamma ( 1 - \\lambda ) ( { \\bf I } - \\tilde { \\bf A } ) { \\bf X } ^ { k } = \\left( 1 - 2 \\gamma ( 1 - \\lambda ) \\right) { \\bf X } ^ { k } + 2 \\gamma ( 1 - \\lambda ) \\tilde { \\bf A } { \\bf X } ^ { k } } } & { { } { } { } } \\\\ { { \\displaystyle { \\bf X } ^ { k + 1 } = \\arg \\operatorname* { m i n } \\left\\{ \\lambda \\| { \\bf X } - { \\bf X } _ { \\mathrm { i n } } \\| _ { 2 1 } + \\frac { 1 } { 2 \\gamma } \\| { \\bf X } - { \\bf Y } ^ { k } \\| _ { F } ^ { 2 } \\right\\} } } & { { } { } { } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 233, + 746, + 764, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where ${ \\bf X } ^ { 0 } = { \\bf X } _ { \\mathrm { i n } }$ and $\\gamma$ is the stepsize to be specified later. Let $\\mathbf { Z } = \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } }$ , and Eq. (10) can be rewritten as: ", + "bbox": [ + 174, + 804, + 825, + 833 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/68fda391e9e91a1ee7340bf5b8ef9ec4653a491d169bc42f8f524ca1be439353.jpg", + "text": "$$\n\\begin{array} { r l } { { \\mathbf { Z } ^ { k + 1 } = \\arg \\operatorname* { m i n } \\{ \\lambda \\| \\mathbf { Z } \\| _ { 2 1 } + \\frac { 1 } { 2 \\gamma } \\| \\mathbf { Z } - ( \\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } } ) \\| _ { F } ^ { 2 } \\} } } \\\\ & { = \\mathbf { p r o x } _ { \\gamma \\lambda \\| \\cdot \\| _ { 2 1 } } ( \\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } } ) } \\\\ & { \\mathbf { X } ^ { k + 1 } = \\mathbf { X } _ { \\mathrm { i n } } + \\mathbf { Z } ^ { k + 1 } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 312, + 833, + 684, + 910 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $i$ -th row of the proximal operator in Eq. (11) can be computed analytically ", + "bbox": [ + 173, + 90, + 691, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/2f99296051acdd82ba233470d0b60b0f2eed6592b1b372eecf8968426e2607f2.jpg", + "text": "$$\n\\left( \\mathbf { p r o x } _ { \\gamma \\lambda \\parallel \\cdot \\parallel _ { 2 1 } } ( \\mathbf { X } ) \\right) _ { i } = \\frac { \\mathbf { X } _ { i } } { \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } } \\operatorname* { m a x } ( \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } - \\gamma \\lambda , 0 ) = \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } } , 0 ) \\cdot \\mathbf { X } _ { i } .\n$$", + "text_format": "latex", + "bbox": [ + 232, + 109, + 766, + 143 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the proximal operator returns 0 if the input vector is 0. Substituting $\\mathbf { X }$ in Eq. (13) with $\\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } }$ and combining Eq. (11) and Eq. (12), then Eq. (12) becomes ", + "bbox": [ + 171, + 147, + 830, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/ddd1e17c5d829c98938c8b1388cd54b3e08aa5daae1060e5eac9a190d0f7b2f6.jpg", + "text": "$$\n\\mathbf { X } _ { i } ^ { k + 1 } = ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } ( \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } ) = ( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } \\mathbf { Y } _ { i } ^ { k } , \\quad \\forall i \\in [ n ] ,\n$$", + "text_format": "latex", + "bbox": [ + 251, + 180, + 746, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\begin{array} { r } { \\beta _ { i } : = \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\| \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } } , 0 ) } \\end{array}$ . To summarize, the proposed adaptive message passing (AMP) scheme is showed in Figure 4, and a diagram is showed in Figure 3. In detail, AMP works as follows: ", + "bbox": [ + 174, + 205, + 825, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "• The first step takes a feature aggregation within the local neighbors with a self-loop weighted by $1 - 2 \\gamma ( 1 - \\lambda )$ ; \n• The second step computes a weight $\\beta _ { i } \\in [ 0 , 1 ]$ for each node $v _ { i }$ depending on the local deviation $\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }$ . \n• The final step takes a linear combination of input features $\\mathbf { X } _ { \\mathrm { i n } }$ and the aggregated features $\\mathbf { Y } ^ { k }$ , where the node-wise residual is adaptively weighted by $1 - \\beta _ { i }$ for each node $v _ { i }$ . ", + "bbox": [ + 194, + 263, + 826, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31b59335ec89ac9c9b7a99814a2a6a89348cd1cc224bfd947556babaafd3c58b.jpg", + "image_caption": [ + "Figure 4: Adaptive Message Passing (AMP) " + ], + "image_footnote": [], + "bbox": [ + 199, + 385, + 679, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The convergence guarantee of AMP and parameter setting for the stepsize $\\gamma$ are illustrated in Theorem 1 and proved in Appendix A. According to Theorem 1, if we set γ = 14(1−λ) or γ = 12(1−λ) , then the first step of AMP can be simplified as $\\begin{array} { r } { \\mathbf { Y } ^ { k } = \\frac { 1 } { 2 } \\mathbf { X } ^ { k } + \\frac { 1 } { 2 } \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } } \\end{array}$ and $\\mathbf { Y } ^ { k } = \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k }$ , respectively. The choice of stepsize will only impact the convergence speed but not the ultimate effect of AMP when it convergences to the fixed point solution. We also discuss the computation complexity per iteration of AMP in Remark 1. ", + "bbox": [ + 173, + 522, + 826, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 1 (Convergence of AMP). Under the stepsize setting < 1(1−λ)kL˜k , the proposed adaptive message passing scheme (AMP) in Eq. (9) and Eq. (10) converges to the optimal solution of the problem defined in Eq. (8). In practice, it is sufficient to choose any $\\begin{array} { r } { \\gamma < \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}$ since $\\Vert \\tilde { \\mathbf { L } } \\Vert _ { 2 } \\leq 2$ Moreover, if the connected components of the graph $\\mathcal { G }$ are not bipartite graphs, it is sufficient to choose $\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}$ since $\\Vert \\tilde { \\mathbf { L } } \\Vert _ { 2 } < 2$ . ", + "bbox": [ + 173, + 614, + 825, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 1 (Computation complexity). AMP is as efficient as simple feature aggregation $\\mathbf { X } _ { o u t } =$ $\\tilde { \\mathbf { A } } \\mathbf { X } _ { i n }$ because the additional computation cost from the second and third steps in Figure 4 is in the order $\\mathcal { O } ( n d )$ , where $n$ is the number of nodes and $d$ is the feature dimension. This is negligible compared with the computation cost $\\mathcal { O } ( m d )$ in feature aggregation, where m is the number of edges, due to the fact that usually there are many more edges than nodes in real-world graphs, i.e., $m \\gg n$ ", + "bbox": [ + 173, + 700, + 825, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Interpretation of AMP ", + "text_level": 1, + "bbox": [ + 174, + 789, + 370, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interestingly, the proposed AMP has a simple and intuitive interpretation as adaptive residual connection, which aligns well with our design motivation: ", + "bbox": [ + 173, + 814, + 825, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "• If the feature of node $v _ { i }$ , i.e., $( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }$ , is significantly inconsistent with its local neighbors, i.e., the aggregated feature $\\mathbf { Y } _ { i } ^ { k }$ , then the local deviation $\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }$ will be large, which leads to a $\\beta _ { i }$ close to 1. Therefore, the final step will assign a small weight to the residual, i.e., $( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }$ , and the aggregated feature $\\mathbf { Y } _ { i } ^ { k }$ will dominate. ", + "bbox": [ + 183, + 852, + 825, + 912 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "• On the contrary, if $( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }$ is already consistent with its local neighbors, $\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }$ will be small, which leads to a $\\beta _ { i }$ close to 0. Thus, the residual will dominate, which is reasonable since there is less need to aggregate features in this case. ", + "bbox": [ + 183, + 90, + 823, + 132 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "• To summarize, the local deviation $\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }$ provides a natural transition from $\\beta _ { i } \\to 1$ to $\\beta _ { i } \\to 0$ , and the transition can be modulated by $\\lambda$ which can be either learned or tuned as a hyperparameter through cross-validation. This transition provides an node-wise adaptive residual connection for the message passing scheme. ", + "bbox": [ + 183, + 136, + 825, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptivity for abnormal $\\pmb { \\& }$ normal features. According to the homophily assumption on graph structure data [17, 18, 19, 3], the feature representations of normal features should be more consistent with local neighbors than abnormal features. As a result, AMP will assign more residual (i.e., smaller $\\beta$ ) to normal features but less residual (i.e., larger $\\beta$ ) to abnormal features, providing a customized tradeoff between feature aggregation and residual connection. Consequently, it can promote both the resilience to abnormal features and the performance on normal features. Above discussion also implies a clear physical meaning for $\\beta$ in AMP, and we formally define it as the adaptive score. ", + "bbox": [ + 173, + 202, + 825, + 300 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 1 (Adaptive score). The variables $\\{ \\beta _ { 1 } , \\cdot \\cdot \\cdot , \\beta _ { n } \\}$ in the adaptive message passing scheme (AMP) are defined as the adaptive scores for nodes $\\{ v _ { 1 } , \\cdots , v _ { n } \\}$ respectively in graph $\\mathcal { G }$ . In particular, the larger $\\beta _ { i }$ is, the more likely the feature of node $v _ { i }$ is abnormal. ", + "bbox": [ + 173, + 303, + 823, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Remark 2 (Nonlinear smoother). Different from most existing message passing scheme which are linear smoothers, AMP is a nonlinear smoother because the weights $\\{ \\beta _ { i } \\}$ are computed from $\\mathbf { Y } ^ { k }$ and ${ \\bf X } _ { i n }$ . This nonlinearity is the key to achieve adaptive residual connection for different nodes. ", + "bbox": [ + 174, + 348, + 825, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 The Model Architecture ", + "text_level": 1, + "bbox": [ + 174, + 405, + 380, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The proposed adaptive message passing (AMP) can be used as a building block in many GNN models to improve the resilience to abnormal node features. In this work, we choose the the decoupled architectures as APPNP [10] and DAGNN [20], and propose the Adaptive residual GNN (AirGNN): ", + "bbox": [ + 174, + 431, + 825, + 473 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/f62553c679b7785ed1d60b6a031508aaff7f04a8c520ca11429283397931d80c.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbf { X } _ { \\mathrm { i n } } = h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } ) , } \\\\ & { \\mathbf { Y } _ { \\mathrm { p r e } } = \\mathbf { A } \\mathbf { M } \\mathbf { P } \\left( \\mathbf { X } _ { \\mathrm { i n } } , K , \\lambda \\right) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 411, + 476, + 586, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$h _ { \\theta } ( \\cdot )$ is any machine learning model parameterized by learnable parameters $\\theta$ , such as multilayer perceptrons (MLPs). $\\mathbf { X } _ { \\mathrm { f e a } } \\in \\mathbb { R } ^ { n \\times d }$ denotes the initial node features. The model $h _ { \\theta } ( \\cdot )$ will first transform the initial node features as $\\mathbf { X } _ { \\mathrm { i n } } = h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } )$ . AMP takes $h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } )$ as input, and performs $K$ steps of AMP with the hyperparameter $\\lambda$ . Similar to the majority of existing GNN models, the training objective is the cross-entropy classification loss on the labeled nodes, and the whole model is trained in an end-to-end way. Note that AirGNN is very efficient as explained in Remark 1, and it only requires two hyperparameters $K$ and $\\lambda$ without introducing additional parameters to learn, which could reduce the risk of overfitting. ", + "bbox": [ + 174, + 513, + 825, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiment ", + "text_level": 1, + "bbox": [ + 174, + 645, + 303, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we aim to verify the effective of the proposed adaptive message passing scheme (AMP) and the AirGNN model through the semi-supervised node classification tasks. Specifically, we try to answer the following questions: (1) How does AirGNN perform on abnormal and normal features? (Section 4.2 and 4.3) and (2) How does AirGNN work by adjusting the adaptive residual? (Section 4.4) ", + "bbox": [ + 173, + 675, + 825, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Settings ", + "text_level": 1, + "bbox": [ + 174, + 761, + 367, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and baselines. We conduct experiments on 8 real-world datasets including three citation graphs, i.e., Cora, Citeseer, Pubmed [21], two co-authorship graphs, i.e., Coauthor CS and Coauthor Physics [22], two co-purchase graphs, i.e., Amazon Computers and Amazon Photo [22], and one OGB dataset, i.e., ogbn-arxiv [23]. Due to the space limit, we only present the results on Cora, Citeseer, and Pubmed in this section, but defer the results on other datasets to Appendix D.1. More details about the data statistics and data splits are summarized in Appendix B. The proposed AirGNN is compared with representative GNNs, including GCN [3], GAT [6], APPNP [10] and GCNII [9]. We defer the comparison with the variants of APPNP and Robust GCN [24] to Appendix D.3 and D.4 respectively. ", + "bbox": [ + 174, + 786, + 825, + 911 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameter settings. For all baselines, we follow the best hyperparameter settings in their original papers. Additionally, we tune a best residual weight $\\alpha$ for APPNP and GCNII in the range $[ 0 , 1 ]$ . For AirGNN, we use a two-layer MLP as the base model $h _ { \\theta } ( \\cdot )$ , following APPNP. We fix the learning rate 0.01, dropout 0.8, and weight decay 0.0005. Moreover, we set $\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}$ as suggested by Theorem 1. We choose $K = 1 0$ and tune $\\lambda$ in the range $[ 0 , 1 ]$ . Adam optimizer [25] is used in all experiments. We run all experiments by 10 times, and report the mean and variance. ", + "bbox": [ + 173, + 90, + 825, + 178 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation setting. We assess the performance of all models under two types of abnormal feature scenarios, including noisy features and adversarial features. The abnormal features are injected to randomly selected test nodes after model training. By default, all hyperparameters are tuned according to the performance on validation sets when the dataset is clean. If tuning the hyperparameter $\\lambda$ of AirGNN according to the validation sets after injecting abnormal features, the performance will be even better, as discussed in Appendix D.2. The performance on clean data are showed in Appendix D.5 to demonstrate that AirGNN doesn’t need to sacrifice accuracy for better robustness against abnormal features. ", + "bbox": [ + 173, + 184, + 825, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Performance Comparison with Noisy Features ", + "text_level": 1, + "bbox": [ + 174, + 313, + 534, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this subsection, we consider the abnormal features in the noisy feature scenario. Specifically, we simulate the noisy features by assigning a subset of the nodes with random features sampled from a multivariate standard Gaussian distribution. Note that the selection of noise subsets has a apparent impact on the performance since some nodes are less vulnerable to abnormal features while others are more vulnerable. To reduce such variance, we report the average performance over 10 times of random selection of the noise node sets, similar to the settings in the preliminary study in Section 2. We report the node classification test accuracy on abnormal (noisy) features and normal features in Figure 5 and Figure 6, separately, under varying noisy ratio. From these figures, we can observe: ", + "bbox": [ + 173, + 338, + 825, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "• Figure 5 shows that AirGNN significantly outperforms all baselines on all datasets in terms of the performance on noisy nodes. This verifies that AMP is able to improve the resilience to noisy features, aligning well with the design motivation. \n• Figure 6 shows that AirGNN promotes the performance on normal nodes when abnormal nodes exist. This is because AMP can remove some abnormal features which are detrimental to normal nodes. ", + "bbox": [ + 189, + 462, + 825, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb981f452a6f87e6cccc28935d8e1aa89643dc8d85a1008f917f9b4fedf8a508.jpg", + "image_caption": [ + "Figure 5: Node classification accuracy on abnormal (noisy) nodes " + ], + "image_footnote": [], + "bbox": [ + 165, + 564, + 812, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Performance Comparison with Adversarial Features ", + "text_level": 1, + "bbox": [ + 173, + 762, + 578, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this subsection, we consider the abnormal feature scenario when the node features are maliciously attacked by the attacker to manipulate the prediction of GNNs. We use the Nettack [26] implemented in DeepRobust3 [27], a PyTorch library for adversarial attacks and defenses, to generate the adversarial features. We randomly choose 40 test nodes as the targeted nodes, and assess the performance under increasing perturbation budgets $\\{ 0 , 5 , 1 0 , 2 0 , 5 0 , 8 0 \\}$ , where the perturbation numbers denote the number of feature dimensions that can be manipulated. The node classification accuracy on these attacked nodes are showed in Figure 7. From these figures, we can make the following observations: ", + "bbox": [ + 174, + 787, + 825, + 885 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/521af9135379cb9d9a4e04bfd77d43979bb91c8146837931dfc5f87ae829d691.jpg", + "image_caption": [ + "Figure 6: Node classification accuracy on normal nodes " + ], + "image_footnote": [], + "bbox": [ + 163, + 84, + 836, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "• AirGNN is significantly more robust against adversarially attacked features than all baselines. MLP is the most vulnerable model, which demonstrates the usefulness of graph structure information in combating against abnormal node features. ", + "bbox": [ + 194, + 286, + 826, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "• The advantages of AirGNN over the baselines become much stronger with larger perturbation budgets. This suggests that AMP can significantly improve the resilience to abnormal features. ", + "bbox": [ + 192, + 332, + 825, + 361 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b4a84b017c1ab3aee2d1fd5a97a59af3692c895e3672220068ba8be20564b118.jpg", + "image_caption": [ + "Figure 7: Node classification accuracy on adversarial nodes " + ], + "image_footnote": [], + "bbox": [ + 165, + 393, + 816, + 541 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Adaptive Residual for Abnormal & Normal Nodes ", + "text_level": 1, + "bbox": [ + 173, + 588, + 562, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To further understand and verify how AMP and AirGNN work, we investigate the adaptive score $\\beta _ { i }$ for each node $v _ { i }$ . Specifically, the average adaptive scores for abnormal nodes and normal nodes in the last layer of AMP are computed separately. In the noisy feature scenario, we fix ratio of noisy nodes as $10 \\%$ . In the adversarial feature scenario, we choose 40 target nodes and fix the perturbation number as 80. The results in noisy and adversarial feature scenarios are showed in Table 1 and Table 2, respectively. From these tables, we can observe: ", + "bbox": [ + 173, + 613, + 825, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "• On the one hand, it can be clearly observed that in both scenarios, the average adaptive scores for abnormal nodes are significantly higher than those for normal nodes. Therefore, it verifies our intuition that large adaptive scores are strongly related to abnormal features. • On the other hand, it also implies that the residual weights (i.e., $1 - \\beta _ { i } )$ for abnormal nodes are much lower than those of normal nodes. This perfectly aligns with our motivation to remove abnormal features by reducing their residual connections. ", + "bbox": [ + 194, + 708, + 825, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The study on adaptive scores verifies how the adaptive residuals in AMP and AirGNN work as designed. It corroborates that AirGNN not only tremendously boosts the resilience to abnormal features but also provides interpretable information for anomaly detection that will be useful in many security-critical scenarios since the adaptive score serves as a good indicator of abnormal nodes. Morever, it is expected that APPNP without residual will perform well on abnormal nodes but it will sacrifice the performance on normal nodes. We provide detailed comparison with APPNP w/Res and APPNP wo/Res in Appendix D.3 to show the advantages of adaptive residual of AirGNN. ", + "bbox": [ + 174, + 808, + 825, + 906 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d49a0a661a725cae61dfea576b2654826b85c9df4b3a4674a943e47200a74fd9.jpg", + "table_caption": [ + "Table 1: Average adaptive score $( \\beta )$ and residual weight $( 1 - \\beta )$ in the noisy feature scenario. " + ], + "table_footnote": [], + "table_body": "
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes0.998 ± 0.0000.988 ± 0.0000.996 ± 0.000
Average adaptive score for normal nodes Average residual weight for abnormal nodes0.924 ± 0.002 0.002 ±0.0000.807 ± 0.005 0.012 ± 0.0000.869 ± 0.006
Average residual weight for normal nodes0.076 ± 0.0020.193 ± 0.0050.004±0.000 0.131 ± 0.006
", + "bbox": [ + 209, + 114, + 790, + 195 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bd195752ecb387a8a737f39dd60085cba6c171a851c3d77460489987040b8bfc.jpg", + "table_caption": [ + "Table 2: Average adaptive score $( \\beta )$ and residual weight $( 1 - \\beta )$ in the adversarial feature scenario. " + ], + "table_footnote": [], + "table_body": "
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes Average adaptive score for normal nodes0.987 ±0.0000.930 ± 0.0070.959 ± 0.005
Average residual weight for abnormal nodes0.922 ± 0.0040.689 ± 0.0240.826 ± 0.016
0.013 ±0.0000.070±0.0070.041 ±0.005
Average residual weight for normal nodes0.078 ± 0.0040.311 ± 0.0240.174 ± 0.016
", + "bbox": [ + 209, + 237, + 789, + 319 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Related Work ", + "text_level": 1, + "bbox": [ + 174, + 344, + 321, + 362 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "GNNs generalize convolutional neural networks (CNN) to graph structure data through the message passing framework [1, 2, 7]. The design of message passing and GNN architectures are majorly motivated in spectral domain [3, 4] and spatial domain [5, 6, 7, 2]. Recent works have shown that the message passing in GNNs can be regarded as low-pass graph filters [12, 13]. More generally, it has been proven that message passing in many GNNs can be uniformly derived from graph signal denoising [11, 28, 29, 30]. Classic GNNs such as GCN [3] and GAT [6] achieve their best performance with shallow models, but their performance degrades when stacking more layers, which can be partially explained through oversmoothing analyses [14, 15]. Recent works propose to use residual connections or skip connections to mitigate the oversmoothing issues, and they demonstrate the potential benefits from more feature aggregations. Examples include but not limited to DeepGCNs [31], JKNet [32], GCNII [9], APPNP [10] and DeeperGNN [20]. These models use global residual connection that can not be adaptive for each node, which significantly differ from the proposed AirGNN. ", + "bbox": [ + 174, + 377, + 825, + 556 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Recently, there are growing interests in reducing GNNs’ vulnerability to the graph structure noise, such as Robust GCN [24], GCN-SVD [33], Pro-GNN [34], IDGL [35], ElasticGNN [36], etc. Please refer to the comprehensive surveys [37, 38] for more details. However, how to design GNNs with strong resilience to abnormal node features remains to be developed. To the best of our knowledge, AirGNN is the first GNN model that is intrinsically robust to many types of abnormal node features by design. It improves the performance in various kinds of abnormal scenarios without needing to sacrifice clean accuracy in normal settings. ", + "bbox": [ + 174, + 564, + 825, + 661 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Conclusion ", + "text_level": 1, + "bbox": [ + 174, + 683, + 299, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this work, we discover an intrinsic tension between feature aggregation and residual connection in the message passing scheme of GNNs, as well as the corresponding performance tradeoff between nodes with abnormal and normal features. We analyze possible reasons to explain these findings from the perspective of graph Laplacian smoothing. Our understandings further motivate us to propose a simple, efficient, interpretable and adaptive message passing scheme as well as a new GNN model with adaptive residual, named AirGNN. AirGNN provides a node-wise adaptive transition between feature aggregation and residual connection, and the significant advantages of AirGNN are demonstrated through extensive experiments. ", + "bbox": [ + 174, + 715, + 825, + 828 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Acknowledgments and Disclosure of Funding ", + "text_level": 1, + "bbox": [ + 174, + 849, + 553, + 867 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This research is supported by the National Science Foundation (NSF) under grant numbers IIS1714741, CNS1815636, IIS1845081, IIS1907704, DRL2025244, IIS1928278, IIS1955285, ", + "bbox": [ + 176, + 882, + 823, + 911 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "IOS2107215, IOS2035472 and Army Research Office (ARO) under grant number W911NF-21- 1-0198. ", + "bbox": [ + 173, + 92, + 823, + 119 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Societal Impact and Limitations ", + "text_level": 1, + "bbox": [ + 176, + 138, + 442, + 156 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The methodology proposed in this paper might have significant positive societal impact since it reduces machine learning models’ vulnerability to abnormal datasets that generally exist in real-world applications, especially in many security-critical scenarios. In practice, for a given graph, we do not have the prior knowledge about if the graph is clean, has noisy features or adversarial features. Therefore, algorithms like the proposed AirGNN that can work under both the clean and various abnormal feature settings are appealing. While we are unaware of any potential negative society impact, we point out two limitations of this work: (1) this paper focuses on abnormal node features and has not evaluated the performance of the proposed method when the dataset contains both abnormal node features and edges; (2) it is unclear how it performs on heterophilic graphs. It will be interesting to investigate these problems in future works. ", + "bbox": [ + 174, + 171, + 825, + 309 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "References ", + "text_level": 1, + "bbox": [ + 174, + 329, + 266, + 345 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[1] Yao Ma and Jiliang Tang. Deep Learning on Graphs. Cambridge University Press, 2020. \n[2] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International Conference on Machine Learning, pages 1263–1272. PMLR, 2017. \n[3] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. \n[4] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 3844–3852, 2016. \n[5] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. arXiv preprint arXiv:1706.02216, 2017. \n[6] Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua ´ Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. \n[7] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61–80, 2008. \n[8] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 2020. \n[9] Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1725–1735. PMLR, 13–18 Jul 2020. \n[10] Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate: Graph neural networks meet personalized pagerank. arXiv preprint arXiv:1810.05997, 2018. \n[11] Yao Ma, Xiaorui Liu, Tong Zhao, Yozen Liu, Jiliang Tang, and Neil Shah. A unified view on graph neural networks as graph signal denoising. Proceedings of the 30th ACM International Conference on Information and Knowledge Management, 2021. \n[12] Hoang Nt and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass filters. arXiv preprint arXiv:1905.09550, 2019. \n[13] Lingxiao Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. In International Conference on Learning Representations, 2019. \n[14] Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. \n[15] Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. In International Conference on Learning Representations, 2020. \n[16] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. \n[17] Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in social networks. Annual review of sociology, 27(1):415–444, 2001. \n[18] Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propagation. \n[19] Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. 2004. \n[20] Meng Liu, Hongyang Gao, and Shuiwang Ji. Towards deeper graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, 2020. \n[21] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina EliassiRad. Collective classification in network data. AI magazine, 29(3):93–93, 2008. \n[22] Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868, 2018. \n[23] Marinka Zitnik Yuxiao Dong Hongyu Ren Bowen Liu Michele Catasta Jure Leskovec Weihua Hu, Matthias Fey. Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687, 2020. \n[24] Dingyuan Zhu, Ziwei Zhang, Peng Cui, and Wenwu Zhu. Robust graph convolutional networks against adversarial attacks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1399–1407, 2019. \n[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. \n[26] Daniel Zügner, Amir Akbarnejad, and Stephan Günnemann. Adversarial attacks on neural networks for graph data. In KDD. ACM, 2018. \n[27] Yaxin Li, Wei Jin, Han Xu, and Jiliang Tang. Deeprobust: A pytorch library for adversarial attacks and defenses, 2020. \n[28] Xuran Pan, Song Shiji, and Huang Gao. A unified framework for convolution-based graph neural networks. https://openreview.net/forum?id=zUMD–Fb9Bt, 2020. \n[29] Meiqi Zhu, Xiao Wang, Chuan Shi, Houye Ji, and Peng Cui. Interpreting and unifying graph neural networks with an optimization framework, 2021. \n[30] Siheng Chen, Yonina C. Eldar, and Lingxiao Zhao. Graph unrolling networks: Interpretable neural networks for graph signal denoising, 2020. \n[31] Guohao Li, Matthias Müller, Ali Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep as cnns? In The IEEE International Conference on Computer Vision (ICCV), 2019. \n[32] Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 5453–5462. PMLR, 10–15 Jul 2018. \n[33] Negin Entezari, Saba A Al-Sayouri, Amirali Darvishzadeh, and Evangelos E Papalexakis. All you need is low (rank) defending against adversarial attacks on graphs. In Proceedings of the 13th International Conference on Web Search and Data Mining, pages 169–177, 2020. \n[34] Wei Jin, Yao Ma, Xiaorui Liu, Xianfeng Tang, Suhang Wang, and Jiliang Tang. Graph structure learning for robust graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 66–74, 2020. \n[35] Yu Chen, Lingfei Wu, and Mohammed Zaki. Iterative deep graph learning for graph neural networks: Better and robust node embeddings. Advances in Neural Information Processing Systems, 33, 2020. \n[36] Xiaorui Liu, Wei Jin, Yao Ma, Yaxin Li, Hua Liu, Yiqi Wang, Ming Yan, and Jiliang Tang. Elastic graph neural networks. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 6837–6849. PMLR, 18–24 Jul 2021. \n[37] Wei Jin, Yaxing Li, Han Xu, Yiqi Wang, Shuiwang Ji, Charu Aggarwal, and Jiliang Tang. Adversarial attacks and defenses on graphs. ACM SIGKDD Explorations Newsletter, 22(2):19– 34, 2021. \n[38] Yanqiao Zhu, Weizhi Xu, Jinghao Zhang, Qiang Liu, Shu Wu, and Liang Wang. Deep graph structure learning for robust representations: A survey. arXiv preprint arXiv:2103.03036, 2021. \n[39] Laurent Condat. A primal–dual splitting method for convex optimization involving lipschitzian, proximable and linear composite terms. Journal of optimization theory and applications, 158(2):460–479, 2013. \n[40] Fan RK Chung and Fan Chung Graham. Spectral graph theory. Number 92. American Mathematical Soc., 1997. ", + "bbox": [ + 173, + 349, + 826, + 916 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 70, + 828, + 921 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 90, + 828, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Checklist ", + "text_level": 1, + "bbox": [ + 174, + 89, + 254, + 106 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "1. For all authors... ", + "bbox": [ + 214, + 116, + 339, + 131 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ", + "bbox": [ + 238, + 135, + 825, + 227 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "2. If you are including theoretical results... ", + "bbox": [ + 214, + 231, + 493, + 244 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] (b) Did you include complete proofs of all theoretical results? [Yes] ", + "bbox": [ + 236, + 248, + 735, + 280 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "3. If you ran experiments... ", + "bbox": [ + 212, + 285, + 393, + 299 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No] ", + "bbox": [ + 238, + 303, + 826, + 421 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ", + "bbox": [ + 214, + 426, + 823, + 440 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) If your work uses existing assets, did you cite the creators? [Yes] \n(b) Did you mention the license of the assets? [No] \n(c) Did you include any new assets either in the supplemental material or as a URL? [No] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [No] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [No] ", + "bbox": [ + 238, + 444, + 823, + 551 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5. If you used crowdsourcing or conducted research with human subjects... ", + "bbox": [ + 214, + 556, + 705, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ", + "bbox": [ + 238, + 574, + 825, + 665 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/parse/train/hfkER_KJiNw/hfkER_KJiNw_middle.json b/parse/train/hfkER_KJiNw/hfkER_KJiNw_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..851c4672024323cdc275030c893c696bbe043bdf --- /dev/null +++ b/parse/train/hfkER_KJiNw/hfkER_KJiNw_middle.json @@ -0,0 +1,36363 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 126, + 97, + 486, + 117 + ], + "lines": [ + { + "bbox": [ + 126, + 97, + 485, + 119 + ], + "spans": [ + { + "bbox": [ + 126, + 97, + 485, + 119 + ], + "score": 1.0, + "content": "Graph Neural Networks with Adaptive Residual", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 142, + 159, + 223, + 181 + ], + "lines": [ + { + "bbox": [ + 155, + 159, + 212, + 169 + ], + "spans": [ + { + "bbox": [ + 155, + 159, + 212, + 169 + ], + "score": 1.0, + "content": "Xiaorui Liu1", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 142, + 171, + 223, + 181 + ], + "spans": [ + { + "bbox": [ + 142, + 171, + 223, + 181 + ], + "score": 1.0, + "content": "xiaorui@msu.edu", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 262, + 159, + 348, + 182 + ], + "lines": [ + { + "bbox": [ + 273, + 159, + 336, + 173 + ], + "spans": [ + { + "bbox": [ + 273, + 159, + 336, + 173 + ], + "score": 1.0, + "content": "Jiayuan Ding1", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 263, + 171, + 348, + 182 + ], + "spans": [ + { + "bbox": [ + 263, + 171, + 348, + 182 + ], + "score": 1.0, + "content": "dingjia5@msu.edu", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 388, + 159, + 468, + 182 + ], + "lines": [ + { + "bbox": [ + 407, + 158, + 448, + 171 + ], + "spans": [ + { + "bbox": [ + 407, + 158, + 448, + 171 + ], + "score": 1.0, + "content": "Wei Jin1", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 387, + 171, + 469, + 182 + ], + "spans": [ + { + "bbox": [ + 387, + 171, + 469, + 182 + ], + "score": 1.0, + "content": "jinwei2@msu.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 115, + 198, + 189, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 197, + 173, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 173, + 209 + ], + "score": 1.0, + "content": "Han Xu1", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 210, + 190, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 190, + 221 + ], + "score": 1.0, + "content": "xuhan1@msu.edu", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 203, + 198, + 282, + 221 + ], + "lines": [ + { + "bbox": [ + 221, + 196, + 264, + 210 + ], + "spans": [ + { + "bbox": [ + 221, + 196, + 264, + 210 + ], + "score": 1.0, + "content": "Yao Ma2", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 201, + 210, + 284, + 221 + ], + "spans": [ + { + "bbox": [ + 201, + 210, + 284, + 221 + ], + "score": 1.0, + "content": "yao.ma@njit.edu", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.0 + }, + { + "type": "text", + "bbox": [ + 299, + 198, + 394, + 220 + ], + "lines": [ + { + "bbox": [ + 322, + 198, + 370, + 209 + ], + "spans": [ + { + "bbox": [ + 322, + 198, + 370, + 209 + ], + "score": 1.0, + "content": "Zitao Liu3", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 209, + 397, + 221 + ], + "spans": [ + { + "bbox": [ + 297, + 209, + 397, + 221 + ], + "score": 1.0, + "content": "liuzitao@100tal.com", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.0 + }, + { + "type": "text", + "bbox": [ + 410, + 198, + 495, + 221 + ], + "lines": [ + { + "bbox": [ + 423, + 197, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 423, + 197, + 482, + 212 + ], + "score": 1.0, + "content": "Jiliang Tang1", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 410, + 210, + 496, + 222 + ], + "spans": [ + { + "bbox": [ + 410, + 210, + 496, + 222 + ], + "score": 1.0, + "content": "tangjili@msu.edu", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.0 + }, + { + "type": "text", + "bbox": [ + 193, + 235, + 417, + 269 + ], + "lines": [ + { + "bbox": [ + 200, + 233, + 410, + 248 + ], + "spans": [ + { + "bbox": [ + 200, + 233, + 410, + 248 + ], + "score": 1.0, + "content": "1Michigan State University, East Lansing, MI, USA", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 192, + 244, + 419, + 260 + ], + "spans": [ + { + "bbox": [ + 192, + 244, + 419, + 260 + ], + "score": 1.0, + "content": "2New Jersey Institute of Technology, Newark, NJ, USA", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 225, + 255, + 386, + 272 + ], + "spans": [ + { + "bbox": [ + 225, + 255, + 386, + 272 + ], + "score": 1.0, + "content": "3TAL Education Group, Beijing, China", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 283, + 297, + 328, + 310 + ], + "lines": [ + { + "bbox": [ + 281, + 296, + 330, + 311 + ], + "spans": [ + { + "bbox": [ + 281, + 296, + 330, + 311 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 143, + 322, + 469, + 453 + ], + "lines": [ + { + "bbox": [ + 141, + 322, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 335 + ], + "score": 1.0, + "content": "Graph neural networks (GNNs) have shown the power in graph representation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 333, + 469, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 346 + ], + "score": 1.0, + "content": "learning for numerous tasks. In this work, we discover an interesting phenomenon", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 344, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 469, + 356 + ], + "score": 1.0, + "content": "that although residual connections in the message passing of GNNs help improve", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 354, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 367 + ], + "score": 1.0, + "content": "the performance, they immensely amplify GNNs’ vulnerability against abnormal", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 366, + 469, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 469, + 378 + ], + "score": 1.0, + "content": "node features. This is undesirable because in real-world applications, node features", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 376, + 469, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 469, + 390 + ], + "score": 1.0, + "content": "in graphs could often be abnormal such as being naturally noisy or adversarially", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 388, + 469, + 400 + ], + "spans": [ + { + "bbox": [ + 142, + 388, + 469, + 400 + ], + "score": 1.0, + "content": "manipulated. We analyze possible reasons to understand this phenomenon and aim", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 142, + 399, + 469, + 411 + ], + "spans": [ + { + "bbox": [ + 142, + 399, + 469, + 411 + ], + "score": 1.0, + "content": "to design GNNs with stronger resilience to abnormal features. Our understandings", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 410, + 469, + 421 + ], + "spans": [ + { + "bbox": [ + 142, + 410, + 469, + 421 + ], + "score": 1.0, + "content": "motivate us to propose and derive a simple, efficient, interpretable, and adaptive", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 421, + 470, + 432 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 470, + 432 + ], + "score": 1.0, + "content": "message passing scheme, leading to a novel GNN with Adaptive residual, AirGNN1.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 432, + 469, + 443 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 443 + ], + "score": 1.0, + "content": "Extensive experiments under various abnormal feature scenarios demonstrate the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 442, + 304, + 455 + ], + "spans": [ + { + "bbox": [ + 141, + 442, + 304, + 455 + ], + "score": 1.0, + "content": "effectiveness of the proposed algorithm.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 473, + 190, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 192, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 192, + 489 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 498, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "Recent years have witnessed the great success of graph neural networks (GNNs) in representation", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 509, + 504, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 504, + 522 + ], + "score": 1.0, + "content": "learning for graph structure data [1]. Essentially, GNNs generalize deep neural networks (DNNs) from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "score": 1.0, + "content": "regular grids, such as image, video and text, to irregular data such as social, energy, transportation,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "citation, and biological networks. Such data can be naturally represented as graphs with nodes and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "score": 1.0, + "content": "edges. The key building block for such generalization is the neural message passing framework [2]:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 558, + 380, + 577 + ], + "lines": [ + { + "bbox": [ + 229, + 558, + 380, + 577 + ], + "spans": [ + { + "bbox": [ + 229, + 558, + 380, + 577 + ], + "score": 0.92, + "content": "\\mathbf { x } _ { u } ^ { ( k + 1 ) } = \\mathrm { U P D A T E } ^ { ( k ) } \\left( \\mathbf { x } _ { u } ^ { ( k ) } , \\mathbf { m } _ { \\mathcal { N } ( u ) } ^ { ( k ) } \\right)", + "type": "interline_equation", + "image_path": "6f9e56a3d9b3fdbb7f89623ce9d7ae036869e671c7b87fc24582e7a11728195c.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 229, + 558, + 380, + 577 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 583, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 148, + 598 + ], + "score": 1.0, + "content": "where x(ku", + "type": "text" + }, + { + "bbox": [ + 133, + 583, + 176, + 596 + ], + "score": 0.93, + "content": "\\mathbf { x } _ { u } ^ { ( k ) } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 583, + 316, + 599 + ], + "score": 1.0, + "content": "denotes the feature vector of node", + "type": "text" + }, + { + "bbox": [ + 316, + 588, + 324, + 595 + ], + "score": 0.79, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 583, + 349, + 599 + ], + "score": 1.0, + "content": "in the", + "type": "text" + }, + { + "bbox": [ + 350, + 586, + 356, + 595 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 583, + 505, + 599 + ], + "score": 1.0, + "content": "-th iteration of message passing, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 596, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 136, + 614 + ], + "score": 0.93, + "content": "\\mathbf { m } _ { \\mathcal { N } ( u ) } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 599, + 270, + 613 + ], + "score": 1.0, + "content": "is the message aggregated from", + "type": "text" + }, + { + "bbox": [ + 270, + 601, + 277, + 609 + ], + "score": 0.68, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 599, + 344, + 613 + ], + "score": 1.0, + "content": "’s neighborhood", + "type": "text" + }, + { + "bbox": [ + 345, + 599, + 369, + 611 + ], + "score": 0.92, + "content": "\\mathcal { N } ( u )", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 599, + 505, + 613 + ], + "score": 1.0, + "content": ". The specific design of message", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "passing scheme can be motivated from spectral domain [3, 4] or spatial domain [5, 6, 7, 2]. It usually", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 622, + 375, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 375, + 636 + ], + "score": 1.0, + "content": "linearly smooths the features in a local neighborhood on the graph.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 107, + 639, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "score": 1.0, + "content": "GNNs have achieved superior performance in a large number of benchmark datasets [8] where the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 649, + 507, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 507, + 663 + ], + "score": 1.0, + "content": "node features are assumed to be complete and informative. However, in real-world applications,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 660, + 507, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 507, + 674 + ], + "score": 1.0, + "content": "some node features could be abnormal from various aspects. For instance, in social networks,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 671, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 685 + ], + "score": 1.0, + "content": "new users might not have complete profile before they make connections with others, leading to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 684, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 505, + 695 + ], + "score": 1.0, + "content": "missing user features. In transportation networks, node features can be noisy since there exist certain", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 703, + 412, + 713 + ], + "lines": [ + { + "bbox": [ + 120, + 700, + 413, + 715 + ], + "spans": [ + { + "bbox": [ + 120, + 700, + 413, + 715 + ], + "score": 1.0, + "content": "1The implementation is available at https://github.com/lxiaorui/AirGNN.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 106, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 126, + 97, + 486, + 117 + ], + "lines": [ + { + "bbox": [ + 126, + 97, + 485, + 119 + ], + "spans": [ + { + "bbox": [ + 126, + 97, + 485, + 119 + ], + "score": 1.0, + "content": "Graph Neural Networks with Adaptive Residual", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 142, + 159, + 223, + 181 + ], + "lines": [ + { + "bbox": [ + 155, + 159, + 212, + 169 + ], + "spans": [ + { + "bbox": [ + 155, + 159, + 212, + 169 + ], + "score": 1.0, + "content": "Xiaorui Liu1", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 142, + 171, + 223, + 181 + ], + "spans": [ + { + "bbox": [ + 142, + 171, + 223, + 181 + ], + "score": 1.0, + "content": "xiaorui@msu.edu", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 142, + 159, + 223, + 181 + ] + }, + { + "type": "text", + "bbox": [ + 262, + 159, + 348, + 182 + ], + "lines": [ + { + "bbox": [ + 273, + 159, + 336, + 173 + ], + "spans": [ + { + "bbox": [ + 273, + 159, + 336, + 173 + ], + "score": 1.0, + "content": "Jiayuan Ding1", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 263, + 171, + 348, + 182 + ], + "spans": [ + { + "bbox": [ + 263, + 171, + 348, + 182 + ], + "score": 1.0, + "content": "dingjia5@msu.edu", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 263, + 159, + 348, + 182 + ] + }, + { + "type": "text", + "bbox": [ + 388, + 159, + 468, + 182 + ], + "lines": [ + { + "bbox": [ + 407, + 158, + 448, + 171 + ], + "spans": [ + { + "bbox": [ + 407, + 158, + 448, + 171 + ], + "score": 1.0, + "content": "Wei Jin1", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 387, + 171, + 469, + 182 + ], + "spans": [ + { + "bbox": [ + 387, + 171, + 469, + 182 + ], + "score": 1.0, + "content": "jinwei2@msu.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 387, + 158, + 469, + 182 + ] + }, + { + "type": "text", + "bbox": [ + 115, + 198, + 189, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 197, + 173, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 173, + 209 + ], + "score": 1.0, + "content": "Han Xu1", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 210, + 190, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 190, + 221 + ], + "score": 1.0, + "content": "xuhan1@msu.edu", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.0, + "bbox_fs": [ + 115, + 197, + 190, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 203, + 198, + 282, + 221 + ], + "lines": [ + { + "bbox": [ + 221, + 196, + 264, + 210 + ], + "spans": [ + { + "bbox": [ + 221, + 196, + 264, + 210 + ], + "score": 1.0, + "content": "Yao Ma2", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 201, + 210, + 284, + 221 + ], + "spans": [ + { + "bbox": [ + 201, + 210, + 284, + 221 + ], + "score": 1.0, + "content": "yao.ma@njit.edu", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.0, + "bbox_fs": [ + 201, + 196, + 284, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 299, + 198, + 394, + 220 + ], + "lines": [ + { + "bbox": [ + 322, + 198, + 370, + 209 + ], + "spans": [ + { + "bbox": [ + 322, + 198, + 370, + 209 + ], + "score": 1.0, + "content": "Zitao Liu3", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 209, + 397, + 221 + ], + "spans": [ + { + "bbox": [ + 297, + 209, + 397, + 221 + ], + "score": 1.0, + "content": "liuzitao@100tal.com", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.0, + "bbox_fs": [ + 297, + 198, + 397, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 410, + 198, + 495, + 221 + ], + "lines": [ + { + "bbox": [ + 423, + 197, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 423, + 197, + 482, + 212 + ], + "score": 1.0, + "content": "Jiliang Tang1", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 410, + 210, + 496, + 222 + ], + "spans": [ + { + "bbox": [ + 410, + 210, + 496, + 222 + ], + "score": 1.0, + "content": "tangjili@msu.edu", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.0, + "bbox_fs": [ + 410, + 197, + 496, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 193, + 235, + 417, + 269 + ], + "lines": [ + { + "bbox": [ + 200, + 233, + 410, + 248 + ], + "spans": [ + { + "bbox": [ + 200, + 233, + 410, + 248 + ], + "score": 1.0, + "content": "1Michigan State University, East Lansing, MI, USA", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 192, + 244, + 419, + 260 + ], + "spans": [ + { + "bbox": [ + 192, + 244, + 419, + 260 + ], + "score": 1.0, + "content": "2New Jersey Institute of Technology, Newark, NJ, USA", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 225, + 255, + 386, + 272 + ], + "spans": [ + { + "bbox": [ + 225, + 255, + 386, + 272 + ], + "score": 1.0, + "content": "3TAL Education Group, Beijing, China", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 192, + 233, + 419, + 272 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 297, + 328, + 310 + ], + "lines": [ + { + "bbox": [ + 281, + 296, + 330, + 311 + ], + "spans": [ + { + "bbox": [ + 281, + 296, + 330, + 311 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 143, + 322, + 469, + 453 + ], + "lines": [ + { + "bbox": [ + 141, + 322, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 335 + ], + "score": 1.0, + "content": "Graph neural networks (GNNs) have shown the power in graph representation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 333, + 469, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 346 + ], + "score": 1.0, + "content": "learning for numerous tasks. In this work, we discover an interesting phenomenon", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 344, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 469, + 356 + ], + "score": 1.0, + "content": "that although residual connections in the message passing of GNNs help improve", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 354, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 367 + ], + "score": 1.0, + "content": "the performance, they immensely amplify GNNs’ vulnerability against abnormal", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 366, + 469, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 469, + 378 + ], + "score": 1.0, + "content": "node features. This is undesirable because in real-world applications, node features", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 376, + 469, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 469, + 390 + ], + "score": 1.0, + "content": "in graphs could often be abnormal such as being naturally noisy or adversarially", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 388, + 469, + 400 + ], + "spans": [ + { + "bbox": [ + 142, + 388, + 469, + 400 + ], + "score": 1.0, + "content": "manipulated. We analyze possible reasons to understand this phenomenon and aim", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 142, + 399, + 469, + 411 + ], + "spans": [ + { + "bbox": [ + 142, + 399, + 469, + 411 + ], + "score": 1.0, + "content": "to design GNNs with stronger resilience to abnormal features. Our understandings", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 410, + 469, + 421 + ], + "spans": [ + { + "bbox": [ + 142, + 410, + 469, + 421 + ], + "score": 1.0, + "content": "motivate us to propose and derive a simple, efficient, interpretable, and adaptive", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 421, + 470, + 432 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 470, + 432 + ], + "score": 1.0, + "content": "message passing scheme, leading to a novel GNN with Adaptive residual, AirGNN1.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 432, + 469, + 443 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 443 + ], + "score": 1.0, + "content": "Extensive experiments under various abnormal feature scenarios demonstrate the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 442, + 304, + 455 + ], + "spans": [ + { + "bbox": [ + 141, + 442, + 304, + 455 + ], + "score": 1.0, + "content": "effectiveness of the proposed algorithm.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24.5, + "bbox_fs": [ + 141, + 322, + 470, + 455 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 473, + 190, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 192, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 192, + 489 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 498, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "Recent years have witnessed the great success of graph neural networks (GNNs) in representation", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 509, + 504, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 504, + 522 + ], + "score": 1.0, + "content": "learning for graph structure data [1]. Essentially, GNNs generalize deep neural networks (DNNs) from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "score": 1.0, + "content": "regular grids, such as image, video and text, to irregular data such as social, energy, transportation,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "citation, and biological networks. Such data can be naturally represented as graphs with nodes and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "score": 1.0, + "content": "edges. The key building block for such generalization is the neural message passing framework [2]:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 498, + 506, + 555 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 558, + 380, + 577 + ], + "lines": [ + { + "bbox": [ + 229, + 558, + 380, + 577 + ], + "spans": [ + { + "bbox": [ + 229, + 558, + 380, + 577 + ], + "score": 0.92, + "content": "\\mathbf { x } _ { u } ^ { ( k + 1 ) } = \\mathrm { U P D A T E } ^ { ( k ) } \\left( \\mathbf { x } _ { u } ^ { ( k ) } , \\mathbf { m } _ { \\mathcal { N } ( u ) } ^ { ( k ) } \\right)", + "type": "interline_equation", + "image_path": "6f9e56a3d9b3fdbb7f89623ce9d7ae036869e671c7b87fc24582e7a11728195c.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 229, + 558, + 380, + 577 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 583, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 148, + 598 + ], + "score": 1.0, + "content": "where x(ku", + "type": "text" + }, + { + "bbox": [ + 133, + 583, + 176, + 596 + ], + "score": 0.93, + "content": "\\mathbf { x } _ { u } ^ { ( k ) } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 583, + 316, + 599 + ], + "score": 1.0, + "content": "denotes the feature vector of node", + "type": "text" + }, + { + "bbox": [ + 316, + 588, + 324, + 595 + ], + "score": 0.79, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 583, + 349, + 599 + ], + "score": 1.0, + "content": "in the", + "type": "text" + }, + { + "bbox": [ + 350, + 586, + 356, + 595 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 583, + 505, + 599 + ], + "score": 1.0, + "content": "-th iteration of message passing, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 596, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 136, + 614 + ], + "score": 0.93, + "content": "\\mathbf { m } _ { \\mathcal { N } ( u ) } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 599, + 270, + 613 + ], + "score": 1.0, + "content": "is the message aggregated from", + "type": "text" + }, + { + "bbox": [ + 270, + 601, + 277, + 609 + ], + "score": 0.68, + "content": "u", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 599, + 344, + 613 + ], + "score": 1.0, + "content": "’s neighborhood", + "type": "text" + }, + { + "bbox": [ + 345, + 599, + 369, + 611 + ], + "score": 0.92, + "content": "\\mathcal { N } ( u )", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 599, + 505, + 613 + ], + "score": 1.0, + "content": ". The specific design of message", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "passing scheme can be motivated from spectral domain [3, 4] or spatial domain [5, 6, 7, 2]. It usually", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 622, + 375, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 375, + 636 + ], + "score": 1.0, + "content": "linearly smooths the features in a local neighborhood on the graph.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 583, + 505, + 636 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 639, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "score": 1.0, + "content": "GNNs have achieved superior performance in a large number of benchmark datasets [8] where the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 649, + 507, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 507, + 663 + ], + "score": 1.0, + "content": "node features are assumed to be complete and informative. However, in real-world applications,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 660, + 507, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 507, + 674 + ], + "score": 1.0, + "content": "some node features could be abnormal from various aspects. For instance, in social networks,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 671, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 685 + ], + "score": 1.0, + "content": "new users might not have complete profile before they make connections with others, leading to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 684, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 505, + 695 + ], + "score": 1.0, + "content": "missing user features. In transportation networks, node features can be noisy since there exist certain", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "uncertainty and dynamics in the observation of the traffic information. What is worse, node features", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 506, + 96 + ], + "score": 1.0, + "content": "can be adversarially chosen by the attacker to maliciously manipulate the prediction made by GNNs.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "Therefore, it is greatly desired to design GNN models with stronger resilience to abnormal node", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 144, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 144, + 118 + ], + "score": 1.0, + "content": "features.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 640, + 507, + 695 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "uncertainty and dynamics in the observation of the traffic information. What is worse, node features", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 506, + 96 + ], + "score": 1.0, + "content": "can be adversarially chosen by the attacker to maliciously manipulate the prediction made by GNNs.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "Therefore, it is greatly desired to design GNN models with stronger resilience to abnormal node", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 144, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 144, + 118 + ], + "score": 1.0, + "content": "features.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 506, + 134 + ], + "score": 1.0, + "content": "In this work, we first perform empirical investigations on how representative GNN models behave on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "graphs with abnormal features. Specifically, based upon standard benchmark datasets, we simulate", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "the abnormal features by replacing the features of randomly selected nodes with random Gaussian", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "noise. Then the performance of node classification on abnormal features and normal features are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "examined separately. From our preliminary study in Section 2, we reveal two interesting observations:", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 176, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 190 + ], + "score": 1.0, + "content": "(1) Feature aggregation can boost the resilience to abnormal features, but too many aggregations", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 504, + 199 + ], + "score": 1.0, + "content": "could hurt the performance on both normal and abnormal features; and (2) Residual connection helps", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 197, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 211 + ], + "score": 1.0, + "content": "GNNs benefit from more layers for normal features, while making GNNs more fragile to abnormal", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "features. We then provide possible explanations to understand these observed phenomena from the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 219, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 104, + 219, + 506, + 233 + ], + "score": 1.0, + "content": "perspective of graph Laplacian smoothing. Our analyses imply that there might exist an intrinsic", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "tension between feature aggregation and residual connection, which results in a performance tradeoff", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 242, + 300, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 300, + 254 + ], + "score": 1.0, + "content": "between normal features and abnormal features.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 108, + 258, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "Motivated by these findings and understandings, we aim to design new GNNs with stronger resilience", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "to abnormal features while largely maintaining the performance on normal features. Our contributions", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 280, + 231, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 231, + 292 + ], + "score": 1.0, + "content": "can be summarized as follows:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 119, + 300, + 506, + 395 + ], + "lines": [ + { + "bbox": [ + 119, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 119, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "• We discover an intrinsic tension between feature aggregation and residual connection in GNNs,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 128, + 311, + 504, + 322 + ], + "spans": [ + { + "bbox": [ + 128, + 311, + 504, + 322 + ], + "score": 1.0, + "content": "and the corresponding performance tradeoff between abnormal and normal features. We also", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 321, + 393, + 335 + ], + "spans": [ + { + "bbox": [ + 127, + 321, + 393, + 335 + ], + "score": 1.0, + "content": "analyze possible reasons to explain and understand these findings.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 119, + 335, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 119, + 335, + 506, + 350 + ], + "score": 1.0, + "content": "• We propose a simple, efficient, principled and adaptive message passing scheme, which leads to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 347, + 383, + 358 + ], + "spans": [ + { + "bbox": [ + 127, + 347, + 383, + 358 + ], + "score": 1.0, + "content": "a novel GNN model with adaptive residual, named as AirGNN.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 119, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 119, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "• Extensive experiments under various abnormal feature scenarios demonstrate the superiority of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 127, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "the proposed algorithm. The ablation study demonstrates how the adaptive residuals mitigate", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 383, + 259, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 383, + 259, + 396 + ], + "score": 1.0, + "content": "the impact of abnormal features.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 107, + 409, + 187, + 423 + ], + "lines": [ + { + "bbox": [ + 104, + 407, + 189, + 428 + ], + "spans": [ + { + "bbox": [ + 104, + 407, + 189, + 428 + ], + "score": 1.0, + "content": "2 Preliminary", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 493, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 495, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 495, + 448 + ], + "score": 1.0, + "content": "Before introducing the preliminary study, we first define the notations used throughout the paper.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 450, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 307, + 463 + ], + "score": 1.0, + "content": "Notations. We use bold upper-case letters such as", + "type": "text" + }, + { + "bbox": [ + 307, + 451, + 317, + 460 + ], + "score": 0.58, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 448, + 457, + 463 + ], + "score": 1.0, + "content": "to denote matrices. Given a matrix", + "type": "text" + }, + { + "bbox": [ + 457, + 449, + 502, + 461 + ], + "score": 0.91, + "content": "\\mathbf { X } \\in \\mathbb { R } ^ { n \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 448, + 506, + 463 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 137, + 474 + ], + "score": 1.0, + "content": "we use", + "type": "text" + }, + { + "bbox": [ + 137, + 461, + 150, + 472 + ], + "score": 0.89, + "content": "\\mathbf { X } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 460, + 205, + 474 + ], + "score": 1.0, + "content": "to denote its", + "type": "text" + }, + { + "bbox": [ + 205, + 462, + 210, + 471 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 460, + 259, + 474 + ], + "score": 1.0, + "content": "-th row and", + "type": "text" + }, + { + "bbox": [ + 259, + 461, + 276, + 474 + ], + "score": 0.91, + "content": "\\mathbf { X } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 460, + 376, + 474 + ], + "score": 1.0, + "content": "to denote its element in", + "type": "text" + }, + { + "bbox": [ + 377, + 462, + 381, + 471 + ], + "score": 0.68, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 460, + 430, + 474 + ], + "score": 1.0, + "content": "-th row and", + "type": "text" + }, + { + "bbox": [ + 431, + 462, + 437, + 473 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "-th column. The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 473, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 474, + 193, + 492 + ], + "score": 1.0, + "content": "Frobenius norm and", + "type": "text" + }, + { + "bbox": [ + 193, + 477, + 207, + 488 + ], + "score": 0.89, + "content": "\\ell _ { 2 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 474, + 283, + 492 + ], + "score": 1.0, + "content": "norm of a matrix", + "type": "text" + }, + { + "bbox": [ + 283, + 477, + 293, + 487 + ], + "score": 0.68, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 474, + 356, + 492 + ], + "score": 1.0, + "content": "are defined as", + "type": "text" + }, + { + "bbox": [ + 356, + 473, + 443, + 493 + ], + "score": 0.95, + "content": "\\Vert \\mathbf { X } \\Vert _ { F } = \\sqrt { \\sum _ { i j } \\mathbf { X } _ { i j } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 474, + 464, + 492 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 464, + 476, + 505, + 489 + ], + "score": 0.9, + "content": "\\| \\mathbf { X } \\| _ { 2 1 } =", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 492, + 507, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 224, + 512 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\sum _ { i } \\| \\mathbf { X } _ { i } \\| _ { 2 } = \\sum _ { i } \\sqrt { \\sum _ { j } \\mathbf { X } _ { i j } ^ { 2 } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 492, + 330, + 513 + ], + "score": 1.0, + "content": ", respectively. We define", + "type": "text" + }, + { + "bbox": [ + 331, + 496, + 408, + 508 + ], + "score": 0.93, + "content": "\\| \\mathbf { X } \\| _ { 2 } = \\sigma _ { \\operatorname* { m a x } } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 492, + 439, + 513 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 439, + 496, + 477, + 508 + ], + "score": 0.93, + "content": "\\sigma _ { \\mathrm { m a x } } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 492, + 507, + 513 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 510, + 218, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 204, + 523 + ], + "score": 1.0, + "content": "largest singular value of", + "type": "text" + }, + { + "bbox": [ + 205, + 511, + 214, + 520 + ], + "score": 0.47, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 510, + 218, + 523 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 526, + 505, + 629 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 123, + 540 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 124, + 526, + 177, + 538 + ], + "score": 0.9, + "content": "\\mathcal { G } = \\{ \\nu , \\mathcal { E } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 525, + 306, + 540 + ], + "score": 1.0, + "content": "be a graph with the node set", + "type": "text" + }, + { + "bbox": [ + 306, + 527, + 384, + 539 + ], + "score": 0.93, + "content": "\\mathcal { V } = \\{ v _ { 1 } , \\ldots , v _ { n } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 525, + 506, + 540 + ], + "score": 1.0, + "content": "and the undirected edge set", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 536, + 507, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 182, + 550 + ], + "score": 0.9, + "content": "\\mathcal { E } = \\{ e _ { 1 } , \\ldots , e _ { m } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 536, + 219, + 552 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 219, + 538, + 246, + 550 + ], + "score": 0.92, + "content": "\\mathcal { N } ( v _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 536, + 415, + 552 + ], + "score": 1.0, + "content": "to denote the neighboring nodes of node", + "type": "text" + }, + { + "bbox": [ + 416, + 540, + 425, + 549 + ], + "score": 0.83, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 536, + 470, + 552 + ], + "score": 1.0, + "content": ", including", + "type": "text" + }, + { + "bbox": [ + 470, + 539, + 480, + 549 + ], + "score": 0.84, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 536, + 507, + 552 + ], + "score": 1.0, + "content": "itself.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 549, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 295, + 560 + ], + "score": 1.0, + "content": "Suppose that each node is associated with a", + "type": "text" + }, + { + "bbox": [ + 295, + 549, + 302, + 559 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 549, + 506, + 560 + ], + "score": 1.0, + "content": "-dimensional feature vector, and the features for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 208, + 572 + ], + "score": 1.0, + "content": "all nodes are denoted as", + "type": "text" + }, + { + "bbox": [ + 209, + 559, + 265, + 570 + ], + "score": 0.91, + "content": "\\mathbf { X } _ { \\mathrm { f e a } } \\in \\mathbb { R } ^ { n \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 558, + 354, + 572 + ], + "score": 1.0, + "content": ". The graph structure", + "type": "text" + }, + { + "bbox": [ + 355, + 560, + 363, + 570 + ], + "score": 0.81, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "can be represented as an adjacent", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 136, + 583 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 137, + 570, + 187, + 581 + ], + "score": 0.9, + "content": "\\mathbf { A } \\in \\mathbb { R } ^ { n \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 569, + 221, + 583 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 221, + 571, + 261, + 583 + ], + "score": 0.91, + "content": "\\mathbf { A } _ { i j } ~ = ~ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 569, + 441, + 583 + ], + "score": 1.0, + "content": "when there exists an edge between nodes", + "type": "text" + }, + { + "bbox": [ + 442, + 572, + 451, + 582 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 569, + 471, + 583 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 472, + 572, + 482, + 583 + ], + "score": 0.85, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 569, + 506, + 583 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 146, + 594 + ], + "score": 0.91, + "content": "\\mathbf { A } _ { i j } \\ = \\ 0", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 581, + 375, + 595 + ], + "score": 1.0, + "content": "otherwise. The graph Laplacian matrix is defined as", + "type": "text" + }, + { + "bbox": [ + 375, + 582, + 433, + 592 + ], + "score": 0.89, + "content": "\\mathbf { L } = \\mathbf { D } - \\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 581, + 466, + 595 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 466, + 582, + 477, + 592 + ], + "score": 0.55, + "content": "\\mathbf { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 592, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 604 + ], + "score": 1.0, + "content": "diagonal degree matrix. Let us denote the commonly used feature aggregation matrix in GNNs [3] as", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 603, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 107, + 603, + 181, + 615 + ], + "score": 0.88, + "content": "\\tilde { \\bf A } = \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } \\bar { \\hat { \\bf A } } \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 603, + 209, + 618 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 210, + 603, + 258, + 615 + ], + "score": 0.92, + "content": "\\hat { \\mathbf { A } } = \\mathbf { A } + \\mathbf { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 603, + 506, + 618 + ], + "score": 1.0, + "content": "is the adjacent matrix with self-loop and its degree matrix is", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 615, + 371, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 117, + 628 + ], + "score": 0.62, + "content": "\\hat { \\bf D }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 616, + 320, + 630 + ], + "score": 1.0, + "content": ". The corresponding Laplacian matrix is defined as", + "type": "text" + }, + { + "bbox": [ + 320, + 615, + 367, + 628 + ], + "score": 0.92, + "content": "\\tilde { \\mathbf { L } } = \\mathbf { I } - \\tilde { \\mathbf { A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 616, + 371, + 630 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 634, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 647 + ], + "score": 1.0, + "content": "In this work, we focus on the setting where a subset of nodes in the graph contain abnormal", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "score": 1.0, + "content": "features, while the remaining nodes have normal features. In the remaining of this paper, we use", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 655, + 466, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 466, + 669 + ], + "score": 1.0, + "content": "abnormal/normal features to denote nodes with abnormal/normal features, for simplicity.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "title", + "bbox": [ + 107, + 679, + 209, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 210, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 210, + 694 + ], + "score": 1.0, + "content": "2.1 Preliminary Study", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 507, + 713 + ], + "score": 1.0, + "content": "Experimental setup. To investigate how GNNs behave on abnormal and normal node features,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "score": 1.0, + "content": "we design semi-supervised node classification experiments on three common datasets (i.e., Cora,", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 117 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 72, + 506, + 118 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 506, + 134 + ], + "score": 1.0, + "content": "In this work, we first perform empirical investigations on how representative GNN models behave on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "graphs with abnormal features. Specifically, based upon standard benchmark datasets, we simulate", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "the abnormal features by replacing the features of randomly selected nodes with random Gaussian", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "noise. Then the performance of node classification on abnormal features and normal features are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "examined separately. From our preliminary study in Section 2, we reveal two interesting observations:", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 176, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 190 + ], + "score": 1.0, + "content": "(1) Feature aggregation can boost the resilience to abnormal features, but too many aggregations", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 187, + 504, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 504, + 199 + ], + "score": 1.0, + "content": "could hurt the performance on both normal and abnormal features; and (2) Residual connection helps", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 197, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 211 + ], + "score": 1.0, + "content": "GNNs benefit from more layers for normal features, while making GNNs more fragile to abnormal", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "features. We then provide possible explanations to understand these observed phenomena from the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 219, + 506, + 233 + ], + "spans": [ + { + "bbox": [ + 104, + 219, + 506, + 233 + ], + "score": 1.0, + "content": "perspective of graph Laplacian smoothing. Our analyses imply that there might exist an intrinsic", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "tension between feature aggregation and residual connection, which results in a performance tradeoff", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 242, + 300, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 300, + 254 + ], + "score": 1.0, + "content": "between normal features and abnormal features.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 122, + 506, + 254 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 258, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "Motivated by these findings and understandings, we aim to design new GNNs with stronger resilience", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "to abnormal features while largely maintaining the performance on normal features. Our contributions", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 280, + 231, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 231, + 292 + ], + "score": 1.0, + "content": "can be summarized as follows:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 257, + 505, + 292 + ] + }, + { + "type": "list", + "bbox": [ + 119, + 300, + 506, + 395 + ], + "lines": [ + { + "bbox": [ + 119, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 119, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "• We discover an intrinsic tension between feature aggregation and residual connection in GNNs,", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 311, + 504, + 322 + ], + "spans": [ + { + "bbox": [ + 128, + 311, + 504, + 322 + ], + "score": 1.0, + "content": "and the corresponding performance tradeoff between abnormal and normal features. We also", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 321, + 393, + 335 + ], + "spans": [ + { + "bbox": [ + 127, + 321, + 393, + 335 + ], + "score": 1.0, + "content": "analyze possible reasons to explain and understand these findings.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 119, + 335, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 119, + 335, + 506, + 350 + ], + "score": 1.0, + "content": "• We propose a simple, efficient, principled and adaptive message passing scheme, which leads to", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 347, + 383, + 358 + ], + "spans": [ + { + "bbox": [ + 127, + 347, + 383, + 358 + ], + "score": 1.0, + "content": "a novel GNN model with adaptive residual, named as AirGNN.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 119, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 119, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "• Extensive experiments under various abnormal feature scenarios demonstrate the superiority of", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 127, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "the proposed algorithm. The ablation study demonstrates how the adaptive residuals mitigate", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 383, + 259, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 383, + 259, + 396 + ], + "score": 1.0, + "content": "the impact of abnormal features.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + } + ], + "index": 22.5, + "bbox_fs": [ + 119, + 299, + 506, + 396 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 409, + 187, + 423 + ], + "lines": [ + { + "bbox": [ + 104, + 407, + 189, + 428 + ], + "spans": [ + { + "bbox": [ + 104, + 407, + 189, + 428 + ], + "score": 1.0, + "content": "2 Preliminary", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 433, + 493, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 495, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 495, + 448 + ], + "score": 1.0, + "content": "Before introducing the preliminary study, we first define the notations used throughout the paper.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 432, + 495, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 450, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 307, + 463 + ], + "score": 1.0, + "content": "Notations. We use bold upper-case letters such as", + "type": "text" + }, + { + "bbox": [ + 307, + 451, + 317, + 460 + ], + "score": 0.58, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 448, + 457, + 463 + ], + "score": 1.0, + "content": "to denote matrices. Given a matrix", + "type": "text" + }, + { + "bbox": [ + 457, + 449, + 502, + 461 + ], + "score": 0.91, + "content": "\\mathbf { X } \\in \\mathbb { R } ^ { n \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 448, + 506, + 463 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 137, + 474 + ], + "score": 1.0, + "content": "we use", + "type": "text" + }, + { + "bbox": [ + 137, + 461, + 150, + 472 + ], + "score": 0.89, + "content": "\\mathbf { X } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 460, + 205, + 474 + ], + "score": 1.0, + "content": "to denote its", + "type": "text" + }, + { + "bbox": [ + 205, + 462, + 210, + 471 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 460, + 259, + 474 + ], + "score": 1.0, + "content": "-th row and", + "type": "text" + }, + { + "bbox": [ + 259, + 461, + 276, + 474 + ], + "score": 0.91, + "content": "\\mathbf { X } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 460, + 376, + 474 + ], + "score": 1.0, + "content": "to denote its element in", + "type": "text" + }, + { + "bbox": [ + 377, + 462, + 381, + 471 + ], + "score": 0.68, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 460, + 430, + 474 + ], + "score": 1.0, + "content": "-th row and", + "type": "text" + }, + { + "bbox": [ + 431, + 462, + 437, + 473 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "-th column. The", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 473, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 474, + 193, + 492 + ], + "score": 1.0, + "content": "Frobenius norm and", + "type": "text" + }, + { + "bbox": [ + 193, + 477, + 207, + 488 + ], + "score": 0.89, + "content": "\\ell _ { 2 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 474, + 283, + 492 + ], + "score": 1.0, + "content": "norm of a matrix", + "type": "text" + }, + { + "bbox": [ + 283, + 477, + 293, + 487 + ], + "score": 0.68, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 474, + 356, + 492 + ], + "score": 1.0, + "content": "are defined as", + "type": "text" + }, + { + "bbox": [ + 356, + 473, + 443, + 493 + ], + "score": 0.95, + "content": "\\Vert \\mathbf { X } \\Vert _ { F } = \\sqrt { \\sum _ { i j } \\mathbf { X } _ { i j } ^ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 474, + 464, + 492 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 464, + 476, + 505, + 489 + ], + "score": 0.9, + "content": "\\| \\mathbf { X } \\| _ { 2 1 } =", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 492, + 507, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 224, + 512 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\sum _ { i } \\| \\mathbf { X } _ { i } \\| _ { 2 } = \\sum _ { i } \\sqrt { \\sum _ { j } \\mathbf { X } _ { i j } ^ { 2 } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 492, + 330, + 513 + ], + "score": 1.0, + "content": ", respectively. We define", + "type": "text" + }, + { + "bbox": [ + 331, + 496, + 408, + 508 + ], + "score": 0.93, + "content": "\\| \\mathbf { X } \\| _ { 2 } = \\sigma _ { \\operatorname* { m a x } } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 492, + 439, + 513 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 439, + 496, + 477, + 508 + ], + "score": 0.93, + "content": "\\sigma _ { \\mathrm { m a x } } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 492, + 507, + 513 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 510, + 218, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 204, + 523 + ], + "score": 1.0, + "content": "largest singular value of", + "type": "text" + }, + { + "bbox": [ + 205, + 511, + 214, + 520 + ], + "score": 0.47, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 510, + 218, + 523 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 448, + 507, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 526, + 505, + 629 + ], + "lines": [ + { + "bbox": [ + 105, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 123, + 540 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 124, + 526, + 177, + 538 + ], + "score": 0.9, + "content": "\\mathcal { G } = \\{ \\nu , \\mathcal { E } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 525, + 306, + 540 + ], + "score": 1.0, + "content": "be a graph with the node set", + "type": "text" + }, + { + "bbox": [ + 306, + 527, + 384, + 539 + ], + "score": 0.93, + "content": "\\mathcal { V } = \\{ v _ { 1 } , \\ldots , v _ { n } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 525, + 506, + 540 + ], + "score": 1.0, + "content": "and the undirected edge set", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 536, + 507, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 182, + 550 + ], + "score": 0.9, + "content": "\\mathcal { E } = \\{ e _ { 1 } , \\ldots , e _ { m } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 536, + 219, + 552 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 219, + 538, + 246, + 550 + ], + "score": 0.92, + "content": "\\mathcal { N } ( v _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 536, + 415, + 552 + ], + "score": 1.0, + "content": "to denote the neighboring nodes of node", + "type": "text" + }, + { + "bbox": [ + 416, + 540, + 425, + 549 + ], + "score": 0.83, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 536, + 470, + 552 + ], + "score": 1.0, + "content": ", including", + "type": "text" + }, + { + "bbox": [ + 470, + 539, + 480, + 549 + ], + "score": 0.84, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 536, + 507, + 552 + ], + "score": 1.0, + "content": "itself.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 549, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 295, + 560 + ], + "score": 1.0, + "content": "Suppose that each node is associated with a", + "type": "text" + }, + { + "bbox": [ + 295, + 549, + 302, + 559 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 549, + 506, + 560 + ], + "score": 1.0, + "content": "-dimensional feature vector, and the features for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 208, + 572 + ], + "score": 1.0, + "content": "all nodes are denoted as", + "type": "text" + }, + { + "bbox": [ + 209, + 559, + 265, + 570 + ], + "score": 0.91, + "content": "\\mathbf { X } _ { \\mathrm { f e a } } \\in \\mathbb { R } ^ { n \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 558, + 354, + 572 + ], + "score": 1.0, + "content": ". The graph structure", + "type": "text" + }, + { + "bbox": [ + 355, + 560, + 363, + 570 + ], + "score": 0.81, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "can be represented as an adjacent", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 136, + 583 + ], + "score": 1.0, + "content": "matrix", + "type": "text" + }, + { + "bbox": [ + 137, + 570, + 187, + 581 + ], + "score": 0.9, + "content": "\\mathbf { A } \\in \\mathbb { R } ^ { n \\times n }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 569, + 221, + 583 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 221, + 571, + 261, + 583 + ], + "score": 0.91, + "content": "\\mathbf { A } _ { i j } ~ = ~ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 569, + 441, + 583 + ], + "score": 1.0, + "content": "when there exists an edge between nodes", + "type": "text" + }, + { + "bbox": [ + 442, + 572, + 451, + 582 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 569, + 471, + 583 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 472, + 572, + 482, + 583 + ], + "score": 0.85, + "content": "v _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 569, + 506, + 583 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 146, + 594 + ], + "score": 0.91, + "content": "\\mathbf { A } _ { i j } \\ = \\ 0", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 581, + 375, + 595 + ], + "score": 1.0, + "content": "otherwise. The graph Laplacian matrix is defined as", + "type": "text" + }, + { + "bbox": [ + 375, + 582, + 433, + 592 + ], + "score": 0.89, + "content": "\\mathbf { L } = \\mathbf { D } - \\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 581, + 466, + 595 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 466, + 582, + 477, + 592 + ], + "score": 0.55, + "content": "\\mathbf { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 592, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 604 + ], + "score": 1.0, + "content": "diagonal degree matrix. Let us denote the commonly used feature aggregation matrix in GNNs [3] as", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 603, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 107, + 603, + 181, + 615 + ], + "score": 0.88, + "content": "\\tilde { \\bf A } = \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } \\bar { \\hat { \\bf A } } \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 603, + 209, + 618 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 210, + 603, + 258, + 615 + ], + "score": 0.92, + "content": "\\hat { \\mathbf { A } } = \\mathbf { A } + \\mathbf { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 603, + 506, + 618 + ], + "score": 1.0, + "content": "is the adjacent matrix with self-loop and its degree matrix is", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 615, + 371, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 117, + 628 + ], + "score": 0.62, + "content": "\\hat { \\bf D }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 616, + 320, + 630 + ], + "score": 1.0, + "content": ". The corresponding Laplacian matrix is defined as", + "type": "text" + }, + { + "bbox": [ + 320, + 615, + 367, + 628 + ], + "score": 0.92, + "content": "\\tilde { \\mathbf { L } } = \\mathbf { I } - \\tilde { \\mathbf { A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 616, + 371, + 630 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 525, + 507, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 634, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 647 + ], + "score": 1.0, + "content": "In this work, we focus on the setting where a subset of nodes in the graph contain abnormal", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 658 + ], + "score": 1.0, + "content": "features, while the remaining nodes have normal features. In the remaining of this paper, we use", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 655, + 466, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 466, + 669 + ], + "score": 1.0, + "content": "abnormal/normal features to denote nodes with abnormal/normal features, for simplicity.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 633, + 506, + 669 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 679, + 209, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 210, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 210, + 694 + ], + "score": 1.0, + "content": "2.1 Preliminary Study", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 507, + 713 + ], + "score": 1.0, + "content": "Experimental setup. To investigate how GNNs behave on abnormal and normal node features,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "score": 1.0, + "content": "we design semi-supervised node classification experiments on three common datasets (i.e., Cora,", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 698, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "CiteSeer and PubMed), following the data splits in the work [3]. Moreover, we simulate the abnormal", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 196, + 97 + ], + "score": 1.0, + "content": "features by assigning", + "type": "text" + }, + { + "bbox": [ + 196, + 83, + 216, + 95 + ], + "score": 0.9, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "of the nodes with random features sampled from a standard Gaussian", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "distribution. The experiments are performed on representative GNN models covering coupled and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "decoupled architectures, including GCN [3], GCNII [9], APPNP [10], and their variants with or", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "without residual connections in feature aggregations, denoted as w/Res and wo/Res. All methods", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "follow the hyperparameter settings in their original papers. We examine how these models perform", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "when the number of layers increases. Note that for the decoupled architectures such as APPNP, we fix", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "the 2-layer MLP and increase the number of propagation layers. While for the coupled architectures", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "such as GCN and GCNII, we increase the number of feature transformation and propagation layers", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "simultaneously. We report the average performance over 10 times of random selection of the noise", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "node sets. The node classification accuracy (mean and standard variance) on nodes with abnormal", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 192, + 388, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 388, + 206 + ], + "score": 1.0, + "content": "and normal features is illustrated in Figure 1 and Figure. 2, separately.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "image", + "bbox": [ + 113, + 222, + 497, + 329 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 222, + 497, + 329 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 222, + 497, + 329 + ], + "spans": [ + { + "bbox": [ + 113, + 222, + 497, + 329 + ], + "score": 0.97, + "type": "image", + "image_path": "00d120485edd7da85f1457dd4c00230b18d99c238c177b5cb16e29a69a61afb9.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 113, + 222, + 497, + 257.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 113, + 257.6666666666667, + 497, + 293.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 113, + 293.33333333333337, + 497, + 329.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 172, + 335, + 437, + 347 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 173, + 334, + 437, + 348 + ], + "spans": [ + { + "bbox": [ + 173, + 334, + 437, + 348 + ], + "score": 1.0, + "content": "Figure 1: Node classification accuracy on abnormal nodes (Cora)", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 14.0 + }, + { + "type": "image", + "bbox": [ + 114, + 376, + 497, + 486 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 376, + 497, + 486 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 376, + 497, + 486 + ], + "spans": [ + { + "bbox": [ + 114, + 376, + 497, + 486 + ], + "score": 0.966, + "type": "image", + "image_path": "08499e09f3e29a9b457146ea9c203108450fafd94446152fa1b7aada0f5024ce.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 114, + 376, + 497, + 412.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 114, + 412.6666666666667, + 497, + 449.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 114, + 449.33333333333337, + 497, + 486.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 178, + 492, + 432, + 504 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 177, + 490, + 433, + 507 + ], + "spans": [ + { + "bbox": [ + 177, + 490, + 433, + 507 + ], + "score": 1.0, + "content": "Figure 2: Node classification accuracy on normal nodes (Cora)", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "index": 18.0 + }, + { + "type": "text", + "bbox": [ + 106, + 521, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "Observations. From Figure 1 and Figure 2, we can make the following observations: (1) Without", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 531, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 258, + 546 + ], + "score": 1.0, + "content": "residual connection, more layers (e.g.,", + "type": "text" + }, + { + "bbox": [ + 258, + 533, + 276, + 543 + ], + "score": 0.83, + "content": "> 2", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 531, + 362, + 546 + ], + "score": 1.0, + "content": "for GCN and GCNII,", + "type": "text" + }, + { + "bbox": [ + 362, + 533, + 384, + 543 + ], + "score": 0.84, + "content": "> 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 531, + 505, + 546 + ], + "score": 1.0, + "content": "for APPNP) hurt the accuracy", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "on nodes with normal features. However, more layers boost the accuracy on nodes with abnormal", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "score": 1.0, + "content": "features significantly, before finally starting to decrease; (2) With residual connection, the accuracy on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "nodes with normal features keeps increasing with more layers2. However, the accuracy on nodes with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "score": 1.0, + "content": "abnormal features only increases marginally when stacking more layers, and then starts to decrease.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "While we only present the experiments on Cora, we defer the results on other datasets to Appendix C,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "which provide similar observations. To conclude, we can summarize these observations into two", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 608, + 170, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 170, + 623 + ], + "score": 1.0, + "content": "major findings:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 117, + 632, + 504, + 654 + ], + "lines": [ + { + "bbox": [ + 119, + 630, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 119, + 630, + 505, + 646 + ], + "score": 1.0, + "content": "• Finding I: Feature aggregation can boost the resilience to abnormal features, but too many", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 642, + 442, + 655 + ], + "spans": [ + { + "bbox": [ + 127, + 642, + 442, + 655 + ], + "score": 1.0, + "content": "aggregations could hurt the performance on both normal and abnormal nodes;", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 118, + 663, + 502, + 685 + ], + "lines": [ + { + "bbox": [ + 118, + 662, + 504, + 675 + ], + "spans": [ + { + "bbox": [ + 118, + 662, + 504, + 675 + ], + "score": 1.0, + "content": "• Finding II: Residual connection helps GNNs benefit from more layers for nodes with normal", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 673, + 388, + 686 + ], + "spans": [ + { + "bbox": [ + 127, + 673, + 388, + 686 + ], + "score": 1.0, + "content": "features, while making GNNs more fragile to abnormal features.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 701, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 117, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 117, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "2GCN w/Res is an exception because its residual is not appropriate, which is consistent with the experiments", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 710, + 165, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 165, + 723 + ], + "score": 1.0, + "content": "in the work [3].", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "CiteSeer and PubMed), following the data splits in the work [3]. Moreover, we simulate the abnormal", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 196, + 97 + ], + "score": 1.0, + "content": "features by assigning", + "type": "text" + }, + { + "bbox": [ + 196, + 83, + 216, + 95 + ], + "score": 0.9, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "of the nodes with random features sampled from a standard Gaussian", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "distribution. The experiments are performed on representative GNN models covering coupled and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "decoupled architectures, including GCN [3], GCNII [9], APPNP [10], and their variants with or", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "without residual connections in feature aggregations, denoted as w/Res and wo/Res. All methods", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "follow the hyperparameter settings in their original papers. We examine how these models perform", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "when the number of layers increases. Note that for the decoupled architectures such as APPNP, we fix", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "the 2-layer MLP and increase the number of propagation layers. While for the coupled architectures", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "such as GCN and GCNII, we increase the number of feature transformation and propagation layers", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "simultaneously. We report the average performance over 10 times of random selection of the noise", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "node sets. The node classification accuracy (mean and standard variance) on nodes with abnormal", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 192, + 388, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 388, + 206 + ], + "score": 1.0, + "content": "and normal features is illustrated in Figure 1 and Figure. 2, separately.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 72, + 506, + 206 + ] + }, + { + "type": "image", + "bbox": [ + 113, + 222, + 497, + 329 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 113, + 222, + 497, + 329 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 222, + 497, + 329 + ], + "spans": [ + { + "bbox": [ + 113, + 222, + 497, + 329 + ], + "score": 0.97, + "type": "image", + "image_path": "00d120485edd7da85f1457dd4c00230b18d99c238c177b5cb16e29a69a61afb9.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 113, + 222, + 497, + 257.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 113, + 257.6666666666667, + 497, + 293.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 113, + 293.33333333333337, + 497, + 329.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 172, + 335, + 437, + 347 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 173, + 334, + 437, + 348 + ], + "spans": [ + { + "bbox": [ + 173, + 334, + 437, + 348 + ], + "score": 1.0, + "content": "Figure 1: Node classification accuracy on abnormal nodes (Cora)", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 14.0 + }, + { + "type": "image", + "bbox": [ + 114, + 376, + 497, + 486 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 376, + 497, + 486 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 376, + 497, + 486 + ], + "spans": [ + { + "bbox": [ + 114, + 376, + 497, + 486 + ], + "score": 0.966, + "type": "image", + "image_path": "08499e09f3e29a9b457146ea9c203108450fafd94446152fa1b7aada0f5024ce.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 114, + 376, + 497, + 412.6666666666667 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 114, + 412.6666666666667, + 497, + 449.33333333333337 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 114, + 449.33333333333337, + 497, + 486.00000000000006 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 178, + 492, + 432, + 504 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 177, + 490, + 433, + 507 + ], + "spans": [ + { + "bbox": [ + 177, + 490, + 433, + 507 + ], + "score": 1.0, + "content": "Figure 2: Node classification accuracy on normal nodes (Cora)", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "index": 18.0 + }, + { + "type": "text", + "bbox": [ + 106, + 521, + 505, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 521, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 506, + 534 + ], + "score": 1.0, + "content": "Observations. From Figure 1 and Figure 2, we can make the following observations: (1) Without", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 531, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 258, + 546 + ], + "score": 1.0, + "content": "residual connection, more layers (e.g.,", + "type": "text" + }, + { + "bbox": [ + 258, + 533, + 276, + 543 + ], + "score": 0.83, + "content": "> 2", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 531, + 362, + 546 + ], + "score": 1.0, + "content": "for GCN and GCNII,", + "type": "text" + }, + { + "bbox": [ + 362, + 533, + 384, + 543 + ], + "score": 0.84, + "content": "> 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 531, + 505, + 546 + ], + "score": 1.0, + "content": "for APPNP) hurt the accuracy", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "on nodes with normal features. However, more layers boost the accuracy on nodes with abnormal", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 506, + 567 + ], + "score": 1.0, + "content": "features significantly, before finally starting to decrease; (2) With residual connection, the accuracy on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "nodes with normal features keeps increasing with more layers2. However, the accuracy on nodes with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "score": 1.0, + "content": "abnormal features only increases marginally when stacking more layers, and then starts to decrease.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "While we only present the experiments on Cora, we defer the results on other datasets to Appendix C,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "which provide similar observations. To conclude, we can summarize these observations into two", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 608, + 170, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 170, + 623 + ], + "score": 1.0, + "content": "major findings:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 521, + 506, + 623 + ] + }, + { + "type": "text", + "bbox": [ + 117, + 632, + 504, + 654 + ], + "lines": [ + { + "bbox": [ + 119, + 630, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 119, + 630, + 505, + 646 + ], + "score": 1.0, + "content": "• Finding I: Feature aggregation can boost the resilience to abnormal features, but too many", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 642, + 442, + 655 + ], + "spans": [ + { + "bbox": [ + 127, + 642, + 442, + 655 + ], + "score": 1.0, + "content": "aggregations could hurt the performance on both normal and abnormal nodes;", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 119, + 630, + 505, + 655 + ] + }, + { + "type": "text", + "bbox": [ + 118, + 663, + 502, + 685 + ], + "lines": [ + { + "bbox": [ + 118, + 662, + 504, + 675 + ], + "spans": [ + { + "bbox": [ + 118, + 662, + 504, + 675 + ], + "score": 1.0, + "content": "• Finding II: Residual connection helps GNNs benefit from more layers for nodes with normal", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 673, + 388, + 686 + ], + "spans": [ + { + "bbox": [ + 127, + 673, + 388, + 686 + ], + "score": 1.0, + "content": "features, while making GNNs more fragile to abnormal features.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 118, + 662, + 504, + 686 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 198, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 200, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 200, + 87 + ], + "score": 1.0, + "content": "2.2 Understandings", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 92, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 92, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 505, + 105 + ], + "score": 1.0, + "content": "In this subsection, we provide the understanding and explanation for aforementioned findings, from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 293, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 293, + 117 + ], + "score": 1.0, + "content": "the perspective of graph Laplacian smoothing.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 108, + 119, + 410, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 117, + 410, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 410, + 136 + ], + "score": 1.0, + "content": "Understanding Finding I: Feature aggregation as Laplacian smoothing", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 136, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 505, + 150 + ], + "score": 1.0, + "content": "The message passing in GCN [3], GCNII wo/ residual and APPNP wo/ residual (as well as many", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 327, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 327, + 160 + ], + "score": 1.0, + "content": "popular GNN models), follows the feature aggregation", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 163, + 335, + 178 + ], + "lines": [ + { + "bbox": [ + 275, + 163, + 335, + 178 + ], + "spans": [ + { + "bbox": [ + 275, + 163, + 335, + 178 + ], + "score": 0.93, + "content": "\\mathbf { X } _ { \\mathrm { o u t } } = \\tilde { \\mathbf { A } } \\mathbf { X } _ { \\mathrm { i n } } ,", + "type": "interline_equation", + "image_path": "5be41b7d9771905b48713d3d8dca4d437d51c9c6e6be8231f4869bf05b49f235.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 275, + 163, + 335, + 178 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 102, + 183, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 132, + 196 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 183, + 149, + 194 + ], + "score": 0.89, + "content": "{ \\bf X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 182, + 166, + 196 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 166, + 183, + 186, + 194 + ], + "score": 0.9, + "content": "\\mathbf { X _ { o u t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "represent the features before and after message passing layer, respectively. It can", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 448, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 448, + 207 + ], + "score": 1.0, + "content": "be interpreted as one gradient descent step for the Laplacian smoothing problem [11]", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 209, + 471, + 242 + ], + "lines": [ + { + "bbox": [ + 138, + 209, + 471, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 471, + 242 + ], + "score": 0.93, + "content": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\mathcal { L } _ { 1 } ( \\mathbf { X } ) : = \\frac { 1 } { 2 } \\mathrm { t r } \\Big ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } \\Big ) = \\frac { 1 } { 2 } \\sum _ { ( v _ { i } , v _ { j } ) \\in \\mathcal { E } } \\| \\frac { \\mathbf { X } _ { i } } { \\sqrt { d _ { i } + 1 } } - \\frac { \\mathbf { X } _ { j } } { \\sqrt { d _ { j } + 1 } } \\| _ { 2 } ^ { 2 } ,", + "type": "interline_equation", + "image_path": "a8ba35e5f2a5c87dc5a4f1ef62b9e0fcb76a09d3a2e3b4ba8193f99725d8ad32.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 138, + 209, + 471, + 220.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 138, + 220.0, + 471, + 231.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 138, + 231.0, + 471, + 242.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 248, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 132, + 262 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 249, + 142, + 260 + ], + "score": 0.88, + "content": "d _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 248, + 245, + 262 + ], + "score": 1.0, + "content": "is the node degree of node", + "type": "text" + }, + { + "bbox": [ + 245, + 250, + 254, + 260 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 248, + 366, + 262 + ], + "score": 1.0, + "content": ". Eq. (2) can be derived from", + "type": "text" + }, + { + "bbox": [ + 367, + 247, + 503, + 261 + ], + "score": 0.92, + "content": "\\mathbf { X } _ { \\mathrm { o u t } } = \\mathbf { X } _ { \\mathrm { i n } } - ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } _ { \\mathrm { i n } } = \\tilde { \\mathbf { A } } \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 248, + 506, + 262 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 259, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 196, + 273 + ], + "score": 1.0, + "content": "with the initialization", + "type": "text" + }, + { + "bbox": [ + 196, + 260, + 235, + 271 + ], + "score": 0.92, + "content": "\\mathbf { X } = \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 259, + 289, + 273 + ], + "score": 1.0, + "content": "and stepsize", + "type": "text" + }, + { + "bbox": [ + 289, + 260, + 315, + 271 + ], + "score": 0.91, + "content": "\\gamma = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 259, + 505, + 273 + ], + "score": 1.0, + "content": ". The Laplacian smoothing problem penalizes", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "the feature difference between neighboring nodes. To reduce this penalty, the feature aggregation in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Eq. (2) smooths the node features by taking the average of local neighbors, and thus can be considered", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "as low-pass filter which gradually filters out high-frequency signals [12, 13]. Therefore, it increases", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "the resilience to abnormal features which are likely to be high-frequency signals. In other words, the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "local neighboring nodes help to correct the abnormal features. Unfortunately, if applied too many", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 339 + ], + "score": 1.0, + "content": "times, these low-pass filters could overly smooth the features (well-known as oversmoothing [14, 15])", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "such that nodes are not distinguishable enough, providing an explanation to the degraded performance", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 346, + 389, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 389, + 360 + ], + "score": 1.0, + "content": "on both abnormal and normal features when stacking too many layers.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 363, + 431, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 431, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 431, + 378 + ], + "score": 1.0, + "content": "Understanding Finding II: Residual connection maintains feature proximity", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "To adjust the feature smoothness for better performance, APPNP [10] utilizes residual connections in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 391, + 218, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 218, + 403 + ], + "score": 1.0, + "content": "message passing as follows", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 406, + 369, + 421 + ], + "lines": [ + { + "bbox": [ + 240, + 406, + 369, + 421 + ], + "spans": [ + { + "bbox": [ + 240, + 406, + 369, + 421 + ], + "score": 0.92, + "content": "\\mathbf { X } ^ { k + 1 } = ( 1 - \\alpha ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } + \\alpha \\mathbf { X _ { \\mathrm { i n } } } ,", + "type": "interline_equation", + "image_path": "4bd63d5a2bf0fef5bca13df83fe1b77403e60662228148abd44885dde98f7b30.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 240, + 406, + 369, + 421 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 427, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 104, + 424, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 104, + 424, + 132, + 442 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 427, + 174, + 439 + ], + "score": 0.92, + "content": "{ \\bf X } ^ { 0 } = { \\bf X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 424, + 506, + 442 + ], + "score": 1.0, + "content": ". It can be considered as an iterative solution for the regularized Laplacian smoothing", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 438, + 163, + 452 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 163, + 452 + ], + "score": 1.0, + "content": "problem [11]", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 454, + 442, + 482 + ], + "lines": [ + { + "bbox": [ + 167, + 454, + 442, + 482 + ], + "spans": [ + { + "bbox": [ + 167, + 454, + 442, + 482 + ], + "score": 0.91, + "content": "\\underset { { \\substack { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } _ { 2 } ( { \\mathbf { X } } ) : = \\frac { \\alpha } { 2 ( 1 - \\alpha ) } \\| \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } } \\| _ { F } ^ { 2 } + \\frac { 1 } { 2 } \\mathrm { t r } \\Big ( \\mathbf { X } ^ { \\top } ( { \\mathbf { I } } - \\tilde { \\mathbf { A } } ) \\mathbf { X } \\Big ) ,", + "type": "interline_equation", + "image_path": "dc684fa92c8954def72ee007cc1a50155f45d52e5b35867cbf9e129b802e67a4.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 167, + 454, + 442, + 482 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 341, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 342, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 179, + 499 + ], + "score": 1.0, + "content": "with initialization", + "type": "text" + }, + { + "bbox": [ + 179, + 486, + 217, + 497 + ], + "score": 0.92, + "content": "\\mathbf { X } = \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 484, + 269, + 499 + ], + "score": 1.0, + "content": "and stepsize", + "type": "text" + }, + { + "bbox": [ + 269, + 486, + 313, + 498 + ], + "score": 0.92, + "content": "\\gamma = 1 - \\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 484, + 342, + 499 + ], + "score": 1.0, + "content": "due to", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 132, + 501, + 479, + 525 + ], + "lines": [ + { + "bbox": [ + 132, + 501, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 479, + 525 + ], + "score": 0.93, + "content": "\\mathbf { X } ^ { k + 1 } = \\mathbf { X } ^ { k } - ( 1 - \\alpha ) \\bigg ( \\frac { \\alpha } { 1 - \\alpha } ( \\mathbf { X } ^ { k } - \\mathbf { X _ { \\mathrm { i n } } } ) + ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ^ { k } \\bigg ) = ( 1 - \\alpha ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } + \\alpha \\mathbf { X _ { \\mathrm { i n } } } .", + "type": "interline_equation", + "image_path": "b21d10dc68f05ed9a814f9fe8eb81efce214601248cf9f4bc15d2afa481cfc3d.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 132, + 501, + 479, + 525 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 506, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "GCNII [9] adopts a similar message passing but further combines a feature transformation layer", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "score": 1.0, + "content": "in each message passing step, which leads to a coupled architecture, as contrast to the decoupled", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "architecture of APPNP. The residual connection naturally arises when regularizing the proximity", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 383, + 575 + ], + "score": 1.0, + "content": "between input and output features, as showed in the first term of", + "type": "text" + }, + { + "bbox": [ + 384, + 561, + 412, + 573 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { 2 } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 560, + 506, + 575 + ], + "score": 1.0, + "content": ". Such proximity can", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "help avoid the trivial solution for the problem in Eq. (3), i.e., totally oversmoothed features only", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "depending on node degrees, and consequently mitigates the oversmoothing issue. More intuitively,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "residual connections in GNNs provide direct information flows between layers that can preserve some", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "necessary high-frequency signals for better discrimination between classes. More layers with residual", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "provide a more accurate solution to Eq. (5), which explains the performance gain from deeper GNNs.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "Unfortunately, these residual connections also undesirably carry on abnormal features which are", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 637, + 388, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 388, + 650 + ], + "score": 1.0, + "content": "detrimental, leading to the inferior performance on abnormal features.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 664, + 260, + 678 + ], + "lines": [ + { + "bbox": [ + 104, + 663, + 262, + 681 + ], + "spans": [ + { + "bbox": [ + 104, + 663, + 262, + 681 + ], + "score": 1.0, + "content": "3 The Proposed Framework", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "In this section, we first motivate the proposed adaptive message passing scheme (AMP) with further", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "discussions on our preliminary study. We then introduce more details about AMP, its interpretations,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 711, + 502, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 502, + 722 + ], + "score": 1.0, + "content": "convergence guarantee and computation complexity, as well as the model architecture of AirGNN.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "4", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 198, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 200, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 200, + 87 + ], + "score": 1.0, + "content": "2.2 Understandings", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 92, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 106, + 92, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 505, + 105 + ], + "score": 1.0, + "content": "In this subsection, we provide the understanding and explanation for aforementioned findings, from", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 293, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 293, + 117 + ], + "score": 1.0, + "content": "the perspective of graph Laplacian smoothing.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 92, + 505, + 117 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 119, + 410, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 117, + 410, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 410, + 136 + ], + "score": 1.0, + "content": "Understanding Finding I: Feature aggregation as Laplacian smoothing", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 136, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 505, + 150 + ], + "score": 1.0, + "content": "The message passing in GCN [3], GCNII wo/ residual and APPNP wo/ residual (as well as many", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 327, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 327, + 160 + ], + "score": 1.0, + "content": "popular GNN models), follows the feature aggregation", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 135, + 505, + 160 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 275, + 163, + 335, + 178 + ], + "lines": [ + { + "bbox": [ + 275, + 163, + 335, + 178 + ], + "spans": [ + { + "bbox": [ + 275, + 163, + 335, + 178 + ], + "score": 0.93, + "content": "\\mathbf { X } _ { \\mathrm { o u t } } = \\tilde { \\mathbf { A } } \\mathbf { X } _ { \\mathrm { i n } } ,", + "type": "interline_equation", + "image_path": "5be41b7d9771905b48713d3d8dca4d437d51c9c6e6be8231f4869bf05b49f235.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 275, + 163, + 335, + 178 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 102, + 183, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 106, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 132, + 196 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 183, + 149, + 194 + ], + "score": 0.89, + "content": "{ \\bf X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 182, + 166, + 196 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 166, + 183, + 186, + 194 + ], + "score": 0.9, + "content": "\\mathbf { X _ { o u t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "represent the features before and after message passing layer, respectively. It can", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 448, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 448, + 207 + ], + "score": 1.0, + "content": "be interpreted as one gradient descent step for the Laplacian smoothing problem [11]", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 182, + 505, + 207 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 209, + 471, + 242 + ], + "lines": [ + { + "bbox": [ + 138, + 209, + 471, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 471, + 242 + ], + "score": 0.93, + "content": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\mathcal { L } _ { 1 } ( \\mathbf { X } ) : = \\frac { 1 } { 2 } \\mathrm { t r } \\Big ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } \\Big ) = \\frac { 1 } { 2 } \\sum _ { ( v _ { i } , v _ { j } ) \\in \\mathcal { E } } \\| \\frac { \\mathbf { X } _ { i } } { \\sqrt { d _ { i } + 1 } } - \\frac { \\mathbf { X } _ { j } } { \\sqrt { d _ { j } + 1 } } \\| _ { 2 } ^ { 2 } ,", + "type": "interline_equation", + "image_path": "a8ba35e5f2a5c87dc5a4f1ef62b9e0fcb76a09d3a2e3b4ba8193f99725d8ad32.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 138, + 209, + 471, + 220.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 138, + 220.0, + 471, + 231.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 138, + 231.0, + 471, + 242.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 248, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 132, + 262 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 249, + 142, + 260 + ], + "score": 0.88, + "content": "d _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 248, + 245, + 262 + ], + "score": 1.0, + "content": "is the node degree of node", + "type": "text" + }, + { + "bbox": [ + 245, + 250, + 254, + 260 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 248, + 366, + 262 + ], + "score": 1.0, + "content": ". Eq. (2) can be derived from", + "type": "text" + }, + { + "bbox": [ + 367, + 247, + 503, + 261 + ], + "score": 0.92, + "content": "\\mathbf { X } _ { \\mathrm { o u t } } = \\mathbf { X } _ { \\mathrm { i n } } - ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } _ { \\mathrm { i n } } = \\tilde { \\mathbf { A } } \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 248, + 506, + 262 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 259, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 196, + 273 + ], + "score": 1.0, + "content": "with the initialization", + "type": "text" + }, + { + "bbox": [ + 196, + 260, + 235, + 271 + ], + "score": 0.92, + "content": "\\mathbf { X } = \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 259, + 289, + 273 + ], + "score": 1.0, + "content": "and stepsize", + "type": "text" + }, + { + "bbox": [ + 289, + 260, + 315, + 271 + ], + "score": 0.91, + "content": "\\gamma = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 259, + 505, + 273 + ], + "score": 1.0, + "content": ". The Laplacian smoothing problem penalizes", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "the feature difference between neighboring nodes. To reduce this penalty, the feature aggregation in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Eq. (2) smooths the node features by taking the average of local neighbors, and thus can be considered", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "as low-pass filter which gradually filters out high-frequency signals [12, 13]. Therefore, it increases", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "the resilience to abnormal features which are likely to be high-frequency signals. In other words, the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "local neighboring nodes help to correct the abnormal features. Unfortunately, if applied too many", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 339 + ], + "score": 1.0, + "content": "times, these low-pass filters could overly smooth the features (well-known as oversmoothing [14, 15])", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "such that nodes are not distinguishable enough, providing an explanation to the degraded performance", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 346, + 389, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 389, + 360 + ], + "score": 1.0, + "content": "on both abnormal and normal features when stacking too many layers.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 247, + 506, + 360 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 363, + 431, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 431, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 431, + 378 + ], + "score": 1.0, + "content": "Understanding Finding II: Residual connection maintains feature proximity", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "To adjust the feature smoothness for better performance, APPNP [10] utilizes residual connections in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 391, + 218, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 218, + 403 + ], + "score": 1.0, + "content": "message passing as follows", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 379, + 505, + 403 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 240, + 406, + 369, + 421 + ], + "lines": [ + { + "bbox": [ + 240, + 406, + 369, + 421 + ], + "spans": [ + { + "bbox": [ + 240, + 406, + 369, + 421 + ], + "score": 0.92, + "content": "\\mathbf { X } ^ { k + 1 } = ( 1 - \\alpha ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } + \\alpha \\mathbf { X _ { \\mathrm { i n } } } ,", + "type": "interline_equation", + "image_path": "4bd63d5a2bf0fef5bca13df83fe1b77403e60662228148abd44885dde98f7b30.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 240, + 406, + 369, + 421 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 427, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 104, + 424, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 104, + 424, + 132, + 442 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 427, + 174, + 439 + ], + "score": 0.92, + "content": "{ \\bf X } ^ { 0 } = { \\bf X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 424, + 506, + 442 + ], + "score": 1.0, + "content": ". It can be considered as an iterative solution for the regularized Laplacian smoothing", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 438, + 163, + 452 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 163, + 452 + ], + "score": 1.0, + "content": "problem [11]", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 424, + 506, + 452 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 454, + 442, + 482 + ], + "lines": [ + { + "bbox": [ + 167, + 454, + 442, + 482 + ], + "spans": [ + { + "bbox": [ + 167, + 454, + 442, + 482 + ], + "score": 0.91, + "content": "\\underset { { \\substack { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } _ { 2 } ( { \\mathbf { X } } ) : = \\frac { \\alpha } { 2 ( 1 - \\alpha ) } \\| \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } } \\| _ { F } ^ { 2 } + \\frac { 1 } { 2 } \\mathrm { t r } \\Big ( \\mathbf { X } ^ { \\top } ( { \\mathbf { I } } - \\tilde { \\mathbf { A } } ) \\mathbf { X } \\Big ) ,", + "type": "interline_equation", + "image_path": "dc684fa92c8954def72ee007cc1a50155f45d52e5b35867cbf9e129b802e67a4.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 167, + 454, + 442, + 482 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 341, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 342, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 179, + 499 + ], + "score": 1.0, + "content": "with initialization", + "type": "text" + }, + { + "bbox": [ + 179, + 486, + 217, + 497 + ], + "score": 0.92, + "content": "\\mathbf { X } = \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 484, + 269, + 499 + ], + "score": 1.0, + "content": "and stepsize", + "type": "text" + }, + { + "bbox": [ + 269, + 486, + 313, + 498 + ], + "score": 0.92, + "content": "\\gamma = 1 - \\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 484, + 342, + 499 + ], + "score": 1.0, + "content": "due to", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 484, + 342, + 499 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 132, + 501, + 479, + 525 + ], + "lines": [ + { + "bbox": [ + 132, + 501, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 479, + 525 + ], + "score": 0.93, + "content": "\\mathbf { X } ^ { k + 1 } = \\mathbf { X } ^ { k } - ( 1 - \\alpha ) \\bigg ( \\frac { \\alpha } { 1 - \\alpha } ( \\mathbf { X } ^ { k } - \\mathbf { X _ { \\mathrm { i n } } } ) + ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ^ { k } \\bigg ) = ( 1 - \\alpha ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } + \\alpha \\mathbf { X _ { \\mathrm { i n } } } .", + "type": "interline_equation", + "image_path": "b21d10dc68f05ed9a814f9fe8eb81efce214601248cf9f4bc15d2afa481cfc3d.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 132, + 501, + 479, + 525 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 506, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "GCNII [9] adopts a similar message passing but further combines a feature transformation layer", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "score": 1.0, + "content": "in each message passing step, which leads to a coupled architecture, as contrast to the decoupled", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "architecture of APPNP. The residual connection naturally arises when regularizing the proximity", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 383, + 575 + ], + "score": 1.0, + "content": "between input and output features, as showed in the first term of", + "type": "text" + }, + { + "bbox": [ + 384, + 561, + 412, + 573 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { 2 } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 560, + 506, + 575 + ], + "score": 1.0, + "content": ". Such proximity can", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "help avoid the trivial solution for the problem in Eq. (3), i.e., totally oversmoothed features only", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "depending on node degrees, and consequently mitigates the oversmoothing issue. More intuitively,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 506, + 607 + ], + "score": 1.0, + "content": "residual connections in GNNs provide direct information flows between layers that can preserve some", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 617 + ], + "score": 1.0, + "content": "necessary high-frequency signals for better discrimination between classes. More layers with residual", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 506, + 628 + ], + "score": 1.0, + "content": "provide a more accurate solution to Eq. (5), which explains the performance gain from deeper GNNs.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "Unfortunately, these residual connections also undesirably carry on abnormal features which are", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 637, + 388, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 388, + 650 + ], + "score": 1.0, + "content": "detrimental, leading to the inferior performance on abnormal features.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 529, + 506, + 650 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 664, + 260, + 678 + ], + "lines": [ + { + "bbox": [ + 104, + 663, + 262, + 681 + ], + "spans": [ + { + "bbox": [ + 104, + 663, + 262, + 681 + ], + "score": 1.0, + "content": "3 The Proposed Framework", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "In this section, we first motivate the proposed adaptive message passing scheme (AMP) with further", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "discussions on our preliminary study. We then introduce more details about AMP, its interpretations,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 711, + 502, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 502, + 722 + ], + "score": 1.0, + "content": "convergence guarantee and computation complexity, as well as the model architecture of AirGNN.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 688, + 506, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 209, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 210, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 210, + 87 + ], + "score": 1.0, + "content": "3.1 Design Motivation", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 92, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "Our preliminary study in Section 2 reveals an intrinsic tension between feature aggregation and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 115 + ], + "score": 1.0, + "content": "residual connection: (1) feature aggregation helps smooth out abnormal features, while it could cause", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 127 + ], + "score": 1.0, + "content": "inappropriate smoothing for normal features; (2) residual connection is essential for adjusting the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "score": 1.0, + "content": "feature smoothness, but it could be detrimental for abnormal features. Although this conflict can be", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 136, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 136, + 441, + 150 + ], + "score": 1.0, + "content": "partially mitigated by adjusting the residual connection such as the residual weight", + "type": "text" + }, + { + "bbox": [ + 441, + 138, + 449, + 146 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 136, + 506, + 150 + ], + "score": 1.0, + "content": "in GCNII [9]", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 159 + ], + "score": 1.0, + "content": "and APPNP [10], such global adjustment cannot be adaptive to a subset of the nodes, e.g., the nodes", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "with abnormal features. This is crucial because in practice we often encounter the scenario where", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 169, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 505, + 181 + ], + "score": 1.0, + "content": "only a subset of nodes contain abnormal features. Therefore, how to reconcile this dilemma still", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "score": 1.0, + "content": "desires dedicated efforts. We then naturally ask a question: Can we design a better message passing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 422, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 422, + 203 + ], + "score": 1.0, + "content": "scheme with node-wise adaptive feature aggregation and residual connection?", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 206, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "The motivation of the proposed idea builds upon the following intuition: while it is important to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "maintain the proximity between input and output features as in Eq. (5), it could be over aggressive to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 103, + 225, + 503, + 245 + ], + "spans": [ + { + "bbox": [ + 103, + 225, + 352, + 245 + ], + "score": 1.0, + "content": "penalize their deviations by the square of Frobenius norm, i.e.,", + "type": "text" + }, + { + "bbox": [ + 352, + 228, + 503, + 241 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\| \\dot { \\bf X } - \\dot { \\bf X } _ { \\mathrm { i n } } \\| _ { F } ^ { 2 } = \\sum _ { i = 1 } ^ { n } \\| \\dot { \\bf X } _ { i } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } ^ { 2 } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "The fact that this penalty does not tolerate large deviations weakens the capability to remove abnormal", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 250, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 265 + ], + "score": 1.0, + "content": "features through Laplacian smoothing. This motivates us to consider an alternative proximity penalty", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 265, + 382, + 298 + ], + "lines": [ + { + "bbox": [ + 228, + 265, + 382, + 298 + ], + "spans": [ + { + "bbox": [ + 228, + 265, + 382, + 298 + ], + "score": 0.95, + "content": "\\| { \\bf X } - { \\bf X } _ { \\mathrm { i n } } \\| _ { 2 1 } : = \\sum _ { i = 1 } ^ { n } \\| { \\bf X } _ { i } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } ,", + "type": "interline_equation", + "image_path": "2f28e684de6375dcb669923823ae95ee31b28bb7d9ad48bed4a2a0043c2958b9.jpg" + } + ] + } + ], + "index": 16.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 265, + 382, + 281.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 228, + 281.5, + 382, + 298.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 300, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 288, + 313 + ], + "score": 1.0, + "content": "which instead penalizes the deviations by the", + "type": "text" + }, + { + "bbox": [ + 288, + 301, + 298, + 312 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 300, + 372, + 313 + ], + "score": 1.0, + "content": "norm of row-wise", + "type": "text" + }, + { + "bbox": [ + 372, + 301, + 383, + 312 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 300, + 445, + 313 + ], + "score": 1.0, + "content": "norms, namely", + "type": "text" + }, + { + "bbox": [ + 445, + 301, + 459, + 312 + ], + "score": 0.88, + "content": "\\ell _ { 2 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "norm. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 107, + 312, + 120, + 323 + ], + "score": 0.88, + "content": "\\ell _ { 2 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 312, + 246, + 324 + ], + "score": 1.0, + "content": "norm promotes row sparsity in", + "type": "text" + }, + { + "bbox": [ + 246, + 312, + 283, + 323 + ], + "score": 0.91, + "content": "\\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 312, + 505, + 324 + ], + "score": 1.0, + "content": ", and it also allows large deviations because the penalty", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 322, + 507, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 507, + 336 + ], + "score": 1.0, + "content": "on large values is less aggressive, leading to the potential removal of abnormal features. Therefore,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 334, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 391, + 346 + ], + "score": 1.0, + "content": "we propose the following Laplacian smoothing problem regularized by", + "type": "text" + }, + { + "bbox": [ + 392, + 334, + 405, + 345 + ], + "score": 0.89, + "content": "\\ell _ { 2 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 334, + 506, + 346 + ], + "score": 1.0, + "content": "norm proximity control:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 347, + 401, + 375 + ], + "lines": [ + { + "bbox": [ + 209, + 347, + 401, + 375 + ], + "spans": [ + { + "bbox": [ + 209, + 347, + 401, + 375 + ], + "score": 0.94, + "content": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\lambda \\| \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } } \\| _ { 2 1 } + \\frac { 1 } { 2 } \\mathrm { t r } ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ) ,", + "type": "interline_equation", + "image_path": "b26e0b4245ae411a1f22ecbedf9e93e2abbac5e9c18379c4b68de77e7bd76391.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 347, + 401, + 361.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 209, + 361.0, + 401, + 375.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 378, + 504, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 133, + 392 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 378, + 178, + 390 + ], + "score": 0.93, + "content": "\\lambda \\in [ 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 376, + 506, + 392 + ], + "score": 1.0, + "content": "is a parameter to adjust the balance between proximity and Laplacian smoothing.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 389, + 377, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 225, + 402 + ], + "score": 1.0, + "content": "In order to easy the tuning of", + "type": "text" + }, + { + "bbox": [ + 225, + 390, + 231, + 399 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 389, + 377, + 402 + ], + "score": 1.0, + "content": ", we made a modification of Eq. (7):", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 174, + 404, + 433, + 426 + ], + "lines": [ + { + "bbox": [ + 174, + 404, + 433, + 426 + ], + "spans": [ + { + "bbox": [ + 174, + 404, + 433, + 426 + ], + "score": 0.9, + "content": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } ( \\mathbf { X } ) : = \\lambda \\| \\mathbf { X } - \\mathbf { X _ { \\mathrm { i n } } } \\| _ { 2 1 } + ( 1 - \\lambda ) \\mathrm { t r } ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ) ,", + "type": "interline_equation", + "image_path": "1e2687fdb8986a4f9b310d05308855779adba69c619b7a25128f76c35752f2d4.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 174, + 404, + 433, + 426 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 257, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 258, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 133, + 443 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 430, + 172, + 442 + ], + "score": 0.92, + "content": "\\lambda \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 429, + 258, + 443 + ], + "score": 1.0, + "content": "controls the balance.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 106, + 454, + 243, + 466 + ], + "lines": [ + { + "bbox": [ + 104, + 452, + 244, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 244, + 469 + ], + "score": 1.0, + "content": "3.2 Adaptive Message Passing", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 147, + 476, + 462, + 533 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 147, + 476, + 462, + 533 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 476, + 462, + 533 + ], + "spans": [ + { + "bbox": [ + 147, + 476, + 462, + 533 + ], + "score": 0.963, + "type": "image", + "image_path": "b9c08a90235cf9dc7aa82eaeface75a6b91e84c8306c75e6acc04630e39b84db.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 147, + 476, + 462, + 495.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 147, + 495.0, + 462, + 514.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 147, + 514.0, + 462, + 533.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 207, + 542, + 404, + 555 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 206, + 541, + 405, + 557 + ], + "spans": [ + { + "bbox": [ + 206, + 541, + 405, + 557 + ], + "score": 1.0, + "content": "Figure 3: Diagram of Adaptive Message Passing", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + } + ], + "index": 31.0 + }, + { + "type": "text", + "bbox": [ + 103, + 565, + 504, + 588 + ], + "lines": [ + { + "bbox": [ + 107, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 107, + 566, + 131, + 578 + ], + "score": 0.91, + "content": "\\mathcal { L } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "is a composite objective with non-smooth and smooth components. We optimize it by proximal", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 576, + 496, + 590 + ], + "spans": [ + { + "bbox": [ + 104, + 576, + 496, + 590 + ], + "score": 1.0, + "content": "gradient descent [16] and obtain the following iterations as the adaptive message passing (AMP):", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "interline_equation", + "bbox": [ + 143, + 591, + 468, + 635 + ], + "lines": [ + { + "bbox": [ + 143, + 591, + 468, + 635 + ], + "spans": [ + { + "bbox": [ + 143, + 591, + 468, + 635 + ], + "score": 0.92, + "content": "\\begin{array} { r l } { { \\displaystyle { \\bf Y } ^ { k } = { \\bf X } ^ { k } - 2 \\gamma ( 1 - \\lambda ) ( { \\bf I } - \\tilde { \\bf A } ) { \\bf X } ^ { k } = \\left( 1 - 2 \\gamma ( 1 - \\lambda ) \\right) { \\bf X } ^ { k } + 2 \\gamma ( 1 - \\lambda ) \\tilde { \\bf A } { \\bf X } ^ { k } } } & { { } { } { } } \\\\ { { \\displaystyle { \\bf X } ^ { k + 1 } = \\arg \\operatorname* { m i n } \\left\\{ \\lambda \\| { \\bf X } - { \\bf X } _ { \\mathrm { i n } } \\| _ { 2 1 } + \\frac { 1 } { 2 \\gamma } \\| { \\bf X } - { \\bf Y } ^ { k } \\| _ { F } ^ { 2 } \\right\\} } } & { { } { } { } } \\end{array}", + "type": "interline_equation", + "image_path": "04047c793794adee67acacd26cda21357adc84912531b7a797e778030c221957.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 143, + 591, + 468, + 605.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 143, + 605.6666666666666, + 468, + 620.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 143, + 620.3333333333333, + 468, + 634.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 133, + 651 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 637, + 176, + 649 + ], + "score": 0.92, + "content": "{ \\bf X } ^ { 0 } = { \\bf X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 636, + 194, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 194, + 640, + 202, + 649 + ], + "score": 0.83, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 636, + 362, + 651 + ], + "score": 1.0, + "content": "is the stepsize to be specified later. Let", + "type": "text" + }, + { + "bbox": [ + 363, + 638, + 420, + 649 + ], + "score": 0.91, + "content": "\\mathbf { Z } = \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 636, + 505, + 651 + ], + "score": 1.0, + "content": ", and Eq. (10) can be", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 649, + 159, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 159, + 661 + ], + "score": 1.0, + "content": "rewritten as:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 660, + 419, + 721 + ], + "lines": [ + { + "bbox": [ + 191, + 660, + 419, + 721 + ], + "spans": [ + { + "bbox": [ + 191, + 660, + 419, + 721 + ], + "score": 0.94, + "content": "\\begin{array} { r l } { { \\mathbf { Z } ^ { k + 1 } = \\arg \\operatorname* { m i n } \\{ \\lambda \\| \\mathbf { Z } \\| _ { 2 1 } + \\frac { 1 } { 2 \\gamma } \\| \\mathbf { Z } - ( \\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } } ) \\| _ { F } ^ { 2 } \\} } } \\\\ & { = \\mathbf { p r o x } _ { \\gamma \\lambda \\| \\cdot \\| _ { 2 1 } } ( \\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } } ) } \\\\ & { \\mathbf { X } ^ { k + 1 } = \\mathbf { X } _ { \\mathrm { i n } } + \\mathbf { Z } ^ { k + 1 } . } \\end{array}", + "type": "interline_equation", + "image_path": "68fda391e9e91a1ee7340bf5b8ef9ec4653a491d169bc42f8f524ca1be439353.jpg" + } + ] + } + ], + "index": 41.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 660, + 419, + 675.25 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 191, + 675.25, + 419, + 690.5 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 191, + 690.5, + 419, + 705.75 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 191, + 705.75, + 419, + 721.0 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 209, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 210, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 210, + 87 + ], + "score": 1.0, + "content": "3.1 Design Motivation", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 92, + 505, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "Our preliminary study in Section 2 reveals an intrinsic tension between feature aggregation and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 115 + ], + "score": 1.0, + "content": "residual connection: (1) feature aggregation helps smooth out abnormal features, while it could cause", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 127 + ], + "score": 1.0, + "content": "inappropriate smoothing for normal features; (2) residual connection is essential for adjusting the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "score": 1.0, + "content": "feature smoothness, but it could be detrimental for abnormal features. Although this conflict can be", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 136, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 136, + 441, + 150 + ], + "score": 1.0, + "content": "partially mitigated by adjusting the residual connection such as the residual weight", + "type": "text" + }, + { + "bbox": [ + 441, + 138, + 449, + 146 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 136, + 506, + 150 + ], + "score": 1.0, + "content": "in GCNII [9]", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 159 + ], + "score": 1.0, + "content": "and APPNP [10], such global adjustment cannot be adaptive to a subset of the nodes, e.g., the nodes", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "with abnormal features. This is crucial because in practice we often encounter the scenario where", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 169, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 169, + 505, + 181 + ], + "score": 1.0, + "content": "only a subset of nodes contain abnormal features. Therefore, how to reconcile this dilemma still", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 505, + 193 + ], + "score": 1.0, + "content": "desires dedicated efforts. We then naturally ask a question: Can we design a better message passing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 191, + 422, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 422, + 203 + ], + "score": 1.0, + "content": "scheme with node-wise adaptive feature aggregation and residual connection?", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 92, + 506, + 203 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 206, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "The motivation of the proposed idea builds upon the following intuition: while it is important to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "maintain the proximity between input and output features as in Eq. (5), it could be over aggressive to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 103, + 225, + 503, + 245 + ], + "spans": [ + { + "bbox": [ + 103, + 225, + 352, + 245 + ], + "score": 1.0, + "content": "penalize their deviations by the square of Frobenius norm, i.e.,", + "type": "text" + }, + { + "bbox": [ + 352, + 228, + 503, + 241 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\| \\dot { \\bf X } - \\dot { \\bf X } _ { \\mathrm { i n } } \\| _ { F } ^ { 2 } = \\sum _ { i = 1 } ^ { n } \\| \\dot { \\bf X } _ { i } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } ^ { 2 } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "The fact that this penalty does not tolerate large deviations weakens the capability to remove abnormal", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 250, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 265 + ], + "score": 1.0, + "content": "features through Laplacian smoothing. This motivates us to consider an alternative proximity penalty", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 103, + 207, + 506, + 265 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 265, + 382, + 298 + ], + "lines": [ + { + "bbox": [ + 228, + 265, + 382, + 298 + ], + "spans": [ + { + "bbox": [ + 228, + 265, + 382, + 298 + ], + "score": 0.95, + "content": "\\| { \\bf X } - { \\bf X } _ { \\mathrm { i n } } \\| _ { 2 1 } : = \\sum _ { i = 1 } ^ { n } \\| { \\bf X } _ { i } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } ,", + "type": "interline_equation", + "image_path": "2f28e684de6375dcb669923823ae95ee31b28bb7d9ad48bed4a2a0043c2958b9.jpg" + } + ] + } + ], + "index": 16.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 265, + 382, + 281.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 228, + 281.5, + 382, + 298.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 300, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 288, + 313 + ], + "score": 1.0, + "content": "which instead penalizes the deviations by the", + "type": "text" + }, + { + "bbox": [ + 288, + 301, + 298, + 312 + ], + "score": 0.87, + "content": "\\ell _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 300, + 372, + 313 + ], + "score": 1.0, + "content": "norm of row-wise", + "type": "text" + }, + { + "bbox": [ + 372, + 301, + 383, + 312 + ], + "score": 0.87, + "content": "\\ell _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 300, + 445, + 313 + ], + "score": 1.0, + "content": "norms, namely", + "type": "text" + }, + { + "bbox": [ + 445, + 301, + 459, + 312 + ], + "score": 0.88, + "content": "\\ell _ { 2 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "norm. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 107, + 312, + 120, + 323 + ], + "score": 0.88, + "content": "\\ell _ { 2 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 312, + 246, + 324 + ], + "score": 1.0, + "content": "norm promotes row sparsity in", + "type": "text" + }, + { + "bbox": [ + 246, + 312, + 283, + 323 + ], + "score": 0.91, + "content": "\\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 312, + 505, + 324 + ], + "score": 1.0, + "content": ", and it also allows large deviations because the penalty", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 322, + 507, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 507, + 336 + ], + "score": 1.0, + "content": "on large values is less aggressive, leading to the potential removal of abnormal features. Therefore,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 334, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 391, + 346 + ], + "score": 1.0, + "content": "we propose the following Laplacian smoothing problem regularized by", + "type": "text" + }, + { + "bbox": [ + 392, + 334, + 405, + 345 + ], + "score": 0.89, + "content": "\\ell _ { 2 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 334, + 506, + 346 + ], + "score": 1.0, + "content": "norm proximity control:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 300, + 507, + 346 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 347, + 401, + 375 + ], + "lines": [ + { + "bbox": [ + 209, + 347, + 401, + 375 + ], + "spans": [ + { + "bbox": [ + 209, + 347, + 401, + 375 + ], + "score": 0.94, + "content": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\lambda \\| \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } } \\| _ { 2 1 } + \\frac { 1 } { 2 } \\mathrm { t r } ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ) ,", + "type": "interline_equation", + "image_path": "b26e0b4245ae411a1f22ecbedf9e93e2abbac5e9c18379c4b68de77e7bd76391.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 347, + 401, + 361.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 209, + 361.0, + 401, + 375.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 378, + 504, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 133, + 392 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 378, + 178, + 390 + ], + "score": 0.93, + "content": "\\lambda \\in [ 0 , \\infty )", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 376, + 506, + 392 + ], + "score": 1.0, + "content": "is a parameter to adjust the balance between proximity and Laplacian smoothing.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 389, + 377, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 225, + 402 + ], + "score": 1.0, + "content": "In order to easy the tuning of", + "type": "text" + }, + { + "bbox": [ + 225, + 390, + 231, + 399 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 389, + 377, + 402 + ], + "score": 1.0, + "content": ", we made a modification of Eq. (7):", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 376, + 506, + 402 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 174, + 404, + 433, + 426 + ], + "lines": [ + { + "bbox": [ + 174, + 404, + 433, + 426 + ], + "spans": [ + { + "bbox": [ + 174, + 404, + 433, + 426 + ], + "score": 0.9, + "content": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } ( \\mathbf { X } ) : = \\lambda \\| \\mathbf { X } - \\mathbf { X _ { \\mathrm { i n } } } \\| _ { 2 1 } + ( 1 - \\lambda ) \\mathrm { t r } ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ) ,", + "type": "interline_equation", + "image_path": "1e2687fdb8986a4f9b310d05308855779adba69c619b7a25128f76c35752f2d4.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 174, + 404, + 433, + 426 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 257, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 429, + 258, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 133, + 443 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 430, + 172, + 442 + ], + "score": 0.92, + "content": "\\lambda \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 429, + 258, + 443 + ], + "score": 1.0, + "content": "controls the balance.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 429, + 258, + 443 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 454, + 243, + 466 + ], + "lines": [ + { + "bbox": [ + 104, + 452, + 244, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 244, + 469 + ], + "score": 1.0, + "content": "3.2 Adaptive Message Passing", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 147, + 476, + 462, + 533 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 147, + 476, + 462, + 533 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 476, + 462, + 533 + ], + "spans": [ + { + "bbox": [ + 147, + 476, + 462, + 533 + ], + "score": 0.963, + "type": "image", + "image_path": "b9c08a90235cf9dc7aa82eaeface75a6b91e84c8306c75e6acc04630e39b84db.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 147, + 476, + 462, + 495.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 147, + 495.0, + 462, + 514.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 147, + 514.0, + 462, + 533.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 207, + 542, + 404, + 555 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 206, + 541, + 405, + 557 + ], + "spans": [ + { + "bbox": [ + 206, + 541, + 405, + 557 + ], + "score": 1.0, + "content": "Figure 3: Diagram of Adaptive Message Passing", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + } + ], + "index": 31.0 + }, + { + "type": "text", + "bbox": [ + 103, + 565, + 504, + 588 + ], + "lines": [ + { + "bbox": [ + 107, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 107, + 566, + 131, + 578 + ], + "score": 0.91, + "content": "\\mathcal { L } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "is a composite objective with non-smooth and smooth components. We optimize it by proximal", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 576, + 496, + 590 + ], + "spans": [ + { + "bbox": [ + 104, + 576, + 496, + 590 + ], + "score": 1.0, + "content": "gradient descent [16] and obtain the following iterations as the adaptive message passing (AMP):", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 565, + 505, + 590 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 143, + 591, + 468, + 635 + ], + "lines": [ + { + "bbox": [ + 143, + 591, + 468, + 635 + ], + "spans": [ + { + "bbox": [ + 143, + 591, + 468, + 635 + ], + "score": 0.92, + "content": "\\begin{array} { r l } { { \\displaystyle { \\bf Y } ^ { k } = { \\bf X } ^ { k } - 2 \\gamma ( 1 - \\lambda ) ( { \\bf I } - \\tilde { \\bf A } ) { \\bf X } ^ { k } = \\left( 1 - 2 \\gamma ( 1 - \\lambda ) \\right) { \\bf X } ^ { k } + 2 \\gamma ( 1 - \\lambda ) \\tilde { \\bf A } { \\bf X } ^ { k } } } & { { } { } { } } \\\\ { { \\displaystyle { \\bf X } ^ { k + 1 } = \\arg \\operatorname* { m i n } \\left\\{ \\lambda \\| { \\bf X } - { \\bf X } _ { \\mathrm { i n } } \\| _ { 2 1 } + \\frac { 1 } { 2 \\gamma } \\| { \\bf X } - { \\bf Y } ^ { k } \\| _ { F } ^ { 2 } \\right\\} } } & { { } { } { } } \\end{array}", + "type": "interline_equation", + "image_path": "04047c793794adee67acacd26cda21357adc84912531b7a797e778030c221957.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 143, + 591, + 468, + 605.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 143, + 605.6666666666666, + 468, + 620.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 143, + 620.3333333333333, + 468, + 634.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 133, + 651 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 637, + 176, + 649 + ], + "score": 0.92, + "content": "{ \\bf X } ^ { 0 } = { \\bf X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 636, + 194, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 194, + 640, + 202, + 649 + ], + "score": 0.83, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 636, + 362, + 651 + ], + "score": 1.0, + "content": "is the stepsize to be specified later. Let", + "type": "text" + }, + { + "bbox": [ + 363, + 638, + 420, + 649 + ], + "score": 0.91, + "content": "\\mathbf { Z } = \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 636, + 505, + 651 + ], + "score": 1.0, + "content": ", and Eq. (10) can be", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 649, + 159, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 159, + 661 + ], + "score": 1.0, + "content": "rewritten as:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 636, + 505, + 661 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 660, + 419, + 721 + ], + "lines": [ + { + "bbox": [ + 191, + 660, + 419, + 721 + ], + "spans": [ + { + "bbox": [ + 191, + 660, + 419, + 721 + ], + "score": 0.94, + "content": "\\begin{array} { r l } { { \\mathbf { Z } ^ { k + 1 } = \\arg \\operatorname* { m i n } \\{ \\lambda \\| \\mathbf { Z } \\| _ { 2 1 } + \\frac { 1 } { 2 \\gamma } \\| \\mathbf { Z } - ( \\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } } ) \\| _ { F } ^ { 2 } \\} } } \\\\ & { = \\mathbf { p r o x } _ { \\gamma \\lambda \\| \\cdot \\| _ { 2 1 } } ( \\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } } ) } \\\\ & { \\mathbf { X } ^ { k + 1 } = \\mathbf { X } _ { \\mathrm { i n } } + \\mathbf { Z } ^ { k + 1 } . } \\end{array}", + "type": "interline_equation", + "image_path": "68fda391e9e91a1ee7340bf5b8ef9ec4653a491d169bc42f8f524ca1be439353.jpg" + } + ] + } + ], + "index": 41.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 660, + 419, + 675.25 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 191, + 675.25, + 419, + 690.5 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 191, + 690.5, + 419, + 705.75 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 191, + 705.75, + 419, + 721.0 + ], + "spans": [], + "index": 43 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 423, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 423, + 88 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 124, + 88 + ], + "score": 1.0, + "content": "The", + "type": "text" + }, + { + "bbox": [ + 124, + 74, + 129, + 83 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 69, + 423, + 88 + ], + "score": 1.0, + "content": "-th row of the proximal operator in Eq. (11) can be computed analytically", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 87, + 469, + 114 + ], + "lines": [ + { + "bbox": [ + 142, + 87, + 469, + 114 + ], + "spans": [ + { + "bbox": [ + 142, + 87, + 469, + 114 + ], + "score": 0.93, + "content": "\\left( \\mathbf { p r o x } _ { \\gamma \\lambda \\parallel \\cdot \\parallel _ { 2 1 } } ( \\mathbf { X } ) \\right) _ { i } = \\frac { \\mathbf { X } _ { i } } { \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } } \\operatorname* { m a x } ( \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } - \\gamma \\lambda , 0 ) = \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } } , 0 ) \\cdot \\mathbf { X } _ { i } .", + "type": "interline_equation", + "image_path": "2f99296051acdd82ba233470d0b60b0f2eed6592b1b372eecf8968426e2607f2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 142, + 87, + 469, + 114 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 117, + 508, + 140 + ], + "lines": [ + { + "bbox": [ + 105, + 117, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 426, + 130 + ], + "score": 1.0, + "content": "Note that the proximal operator returns 0 if the input vector is 0. Substituting", + "type": "text" + }, + { + "bbox": [ + 427, + 118, + 437, + 127 + ], + "score": 0.77, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 117, + 505, + 130 + ], + "score": 1.0, + "content": "in Eq. (13) with", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 394, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 148, + 139 + ], + "score": 0.91, + "content": "\\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 126, + 394, + 142 + ], + "score": 1.0, + "content": "and combining Eq. (11) and Eq. (12), then Eq. (12) becomes", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 143, + 457, + 159 + ], + "lines": [ + { + "bbox": [ + 154, + 143, + 457, + 159 + ], + "spans": [ + { + "bbox": [ + 154, + 143, + 457, + 159 + ], + "score": 0.88, + "content": "\\mathbf { X } _ { i } ^ { k + 1 } = ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } ( \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } ) = ( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } \\mathbf { Y } _ { i } ^ { k } , \\quad \\forall i \\in [ n ] ,", + "type": "interline_equation", + "image_path": "ddd1e17c5d829c98938c8b1388cd54b3e08aa5daae1060e5eac9a190d0f7b2f6.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 154, + 143, + 457, + 159 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 163, + 505, + 201 + ], + "lines": [ + { + "bbox": [ + 102, + 160, + 508, + 185 + ], + "spans": [ + { + "bbox": [ + 102, + 160, + 134, + 185 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 163, + 270, + 180 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\beta _ { i } : = \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\| \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } } , 0 ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 160, + 508, + 185 + ], + "score": 1.0, + "content": ". To summarize, the proposed adaptive message passing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "score": 1.0, + "content": "(AMP) scheme is showed in Figure 4, and a diagram is showed in Figure 3. In detail, AMP works as", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 188, + 143, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 143, + 202 + ], + "score": 1.0, + "content": "follows:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 119, + 209, + 506, + 286 + ], + "lines": [ + { + "bbox": [ + 117, + 207, + 506, + 224 + ], + "spans": [ + { + "bbox": [ + 117, + 207, + 506, + 224 + ], + "score": 1.0, + "content": "• The first step takes a feature aggregation within the local neighbors with a self-loop weighted", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 219, + 205, + 234 + ], + "spans": [ + { + "bbox": [ + 126, + 219, + 140, + 234 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 141, + 221, + 200, + 233 + ], + "score": 0.92, + "content": "1 - 2 \\gamma ( 1 - \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 219, + 205, + 234 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 118, + 232, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 118, + 232, + 268, + 250 + ], + "score": 1.0, + "content": "• The second step computes a weight", + "type": "text" + }, + { + "bbox": [ + 268, + 235, + 309, + 248 + ], + "score": 0.92, + "content": "\\beta _ { i } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 232, + 366, + 250 + ], + "score": 1.0, + "content": "for each node", + "type": "text" + }, + { + "bbox": [ + 366, + 237, + 375, + 246 + ], + "score": 0.84, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 232, + 506, + 250 + ], + "score": 1.0, + "content": "depending on the local deviation", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 128, + 245, + 200, + 261 + ], + "spans": [ + { + "bbox": [ + 128, + 246, + 195, + 260 + ], + "score": 0.9, + "content": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 245, + 200, + 261 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 261, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 117, + 261, + 359, + 276 + ], + "score": 1.0, + "content": "• The final step takes a linear combination of input features", + "type": "text" + }, + { + "bbox": [ + 360, + 264, + 375, + 274 + ], + "score": 0.89, + "content": "\\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 261, + 487, + 276 + ], + "score": 1.0, + "content": "and the aggregated features", + "type": "text" + }, + { + "bbox": [ + 488, + 262, + 502, + 273 + ], + "score": 0.88, + "content": "\\mathbf { Y } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 261, + 505, + 276 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 273, + 450, + 287 + ], + "spans": [ + { + "bbox": [ + 127, + 273, + 351, + 287 + ], + "score": 1.0, + "content": "where the node-wise residual is adaptively weighted by", + "type": "text" + }, + { + "bbox": [ + 351, + 275, + 378, + 285 + ], + "score": 0.91, + "content": "1 - \\beta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 273, + 436, + 287 + ], + "score": 1.0, + "content": "for each node", + "type": "text" + }, + { + "bbox": [ + 436, + 276, + 445, + 285 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 273, + 450, + 287 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "image", + "bbox": [ + 122, + 305, + 416, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 305, + 416, + 385 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 305, + 416, + 385 + ], + "spans": [ + { + "bbox": [ + 122, + 305, + 416, + 385 + ], + "score": 0.393, + "type": "image", + "image_path": "31b59335ec89ac9c9b7a99814a2a6a89348cd1cc224bfd947556babaafd3c58b.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 122, + 305, + 416, + 331.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 122, + 331.6666666666667, + 416, + 358.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 122, + 358.33333333333337, + 416, + 385.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 217, + 391, + 395, + 404 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 389, + 395, + 407 + ], + "spans": [ + { + "bbox": [ + 215, + 389, + 395, + 407 + ], + "score": 1.0, + "content": "Figure 4: Adaptive Message Passing (AMP)", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + } + ], + "index": 16.0 + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 506, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 424, + 429 + ], + "score": 1.0, + "content": "The convergence guarantee of AMP and parameter setting for the stepsize", + "type": "text" + }, + { + "bbox": [ + 424, + 417, + 432, + 426 + ], + "score": 0.79, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 414, + 506, + 429 + ], + "score": 1.0, + "content": "are illustrated in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 424, + 507, + 441 + ], + "spans": [ + { + "bbox": [ + 104, + 424, + 507, + 441 + ], + "score": 1.0, + "content": "Theorem 1 and proved in Appendix A. According to Theorem 1, if we set γ = 14(1−λ) or γ = 12(1−λ) ,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 438, + 507, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 293, + 455 + ], + "score": 1.0, + "content": "then the first step of AMP can be simplified as", + "type": "text" + }, + { + "bbox": [ + 294, + 439, + 382, + 454 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\mathbf { Y } ^ { k } = \\frac { 1 } { 2 } \\mathbf { X } ^ { k } + \\frac { 1 } { 2 } \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 438, + 400, + 455 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 440, + 451, + 452 + ], + "score": 0.9, + "content": "\\mathbf { Y } ^ { k } = \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 438, + 507, + 455 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "The choice of stepsize will only impact the convergence speed but not the ultimate effect of AMP", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "score": 1.0, + "content": "when it convergences to the fixed point solution. We also discuss the computation complexity per", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 474, + 231, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 231, + 486 + ], + "score": 1.0, + "content": "iteration of AMP in Remark 1.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 505, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 507, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 352, + 504 + ], + "score": 1.0, + "content": "Theorem 1 (Convergence of AMP). Under the stepsize setting", + "type": "text" + }, + { + "bbox": [ + 367, + 485, + 507, + 505 + ], + "score": 1.0, + "content": "< 1(1−λ)kL˜k , the proposed adaptive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 501, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 506, + 515 + ], + "score": 1.0, + "content": "message passing scheme (AMP) in Eq. (9) and Eq. (10) converges to the optimal solution of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 102, + 512, + 503, + 532 + ], + "spans": [ + { + "bbox": [ + 102, + 512, + 387, + 532 + ], + "score": 1.0, + "content": "problem defined in Eq. (8). In practice, it is sufficient to choose any", + "type": "text" + }, + { + "bbox": [ + 387, + 514, + 435, + 529 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\gamma < \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 512, + 461, + 532 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 461, + 513, + 503, + 527 + ], + "score": 0.92, + "content": "\\Vert \\tilde { \\mathbf { L } } \\Vert _ { 2 } \\leq 2", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 323, + 540 + ], + "score": 1.0, + "content": "Moreover, if the connected components of the graph", + "type": "text" + }, + { + "bbox": [ + 324, + 528, + 331, + 538 + ], + "score": 0.73, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "are not bipartite graphs, it is sufficient to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 538, + 254, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 538, + 136, + 556 + ], + "score": 1.0, + "content": "choose", + "type": "text" + }, + { + "bbox": [ + 137, + 539, + 184, + 555 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 538, + 209, + 556 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 209, + 538, + 249, + 552 + ], + "score": 0.92, + "content": "\\Vert \\tilde { \\mathbf { L } } \\Vert _ { 2 } < 2", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 538, + 254, + 556 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 473, + 569 + ], + "score": 1.0, + "content": "Remark 1 (Computation complexity). AMP is as efficient as simple feature aggregation", + "type": "text" + }, + { + "bbox": [ + 473, + 556, + 505, + 568 + ], + "score": 0.87, + "content": "\\mathbf { X } _ { o u t } =", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 567, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 107, + 567, + 131, + 579 + ], + "score": 0.89, + "content": "\\tilde { \\mathbf { A } } \\mathbf { X } _ { i n }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 567, + 505, + 582 + ], + "score": 1.0, + "content": "because the additional computation cost from the second and third steps in Figure 4 is in the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 132, + 592 + ], + "score": 1.0, + "content": "order", + "type": "text" + }, + { + "bbox": [ + 132, + 579, + 160, + 591 + ], + "score": 0.93, + "content": "\\mathcal { O } ( n d )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 578, + 192, + 592 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 192, + 582, + 200, + 589 + ], + "score": 0.35, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 578, + 317, + 592 + ], + "score": 1.0, + "content": "is the number of nodes and", + "type": "text" + }, + { + "bbox": [ + 317, + 580, + 324, + 589 + ], + "score": 0.51, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "is the feature dimension. This is negligible", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 589, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 252, + 604 + ], + "score": 1.0, + "content": "compared with the computation cost", + "type": "text" + }, + { + "bbox": [ + 252, + 590, + 282, + 602 + ], + "score": 0.92, + "content": "\\mathcal { O } ( m d )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 589, + 506, + 604 + ], + "score": 1.0, + "content": "in feature aggregation, where m is the number of edges,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 600, + 503, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 471, + 614 + ], + "score": 1.0, + "content": "due to the fact that usually there are many more edges than nodes in real-world graphs, i.e.,", + "type": "text" + }, + { + "bbox": [ + 472, + 603, + 503, + 612 + ], + "score": 0.85, + "content": "m \\gg n", + "type": "inline_equation" + } + ], + "index": 33 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 107, + 625, + 227, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 228, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 228, + 639 + ], + "score": 1.0, + "content": "3.3 Interpretation of AMP", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 645, + 505, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "score": 1.0, + "content": "Interestingly, the proposed AMP has a simple and intuitive interpretation as adaptive residual connec-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 656, + 312, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 312, + 669 + ], + "score": 1.0, + "content": "tion, which aligns well with our design motivation:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 112, + 675, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 111, + 674, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 111, + 674, + 206, + 688 + ], + "score": 1.0, + "content": "• If the feature of node", + "type": "text" + }, + { + "bbox": [ + 207, + 677, + 216, + 687 + ], + "score": 0.83, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 674, + 237, + 688 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 237, + 676, + 263, + 688 + ], + "score": 0.91, + "content": "( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 674, + 505, + 688 + ], + "score": 1.0, + "content": ", is significantly inconsistent with its local neighbors, i.e., the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 118, + 686, + 504, + 702 + ], + "spans": [ + { + "bbox": [ + 118, + 686, + 195, + 702 + ], + "score": 1.0, + "content": "aggregated feature", + "type": "text" + }, + { + "bbox": [ + 195, + 687, + 210, + 700 + ], + "score": 0.91, + "content": "\\mathbf { Y } _ { i } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 686, + 308, + 702 + ], + "score": 1.0, + "content": ", then the local deviation", + "type": "text" + }, + { + "bbox": [ + 308, + 687, + 374, + 700 + ], + "score": 0.93, + "content": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 686, + 493, + 702 + ], + "score": 1.0, + "content": "will be large, which leads to a", + "type": "text" + }, + { + "bbox": [ + 494, + 688, + 504, + 699 + ], + "score": 0.87, + "content": "\\beta _ { i }", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 120, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 120, + 698, + 442, + 712 + ], + "score": 1.0, + "content": "close to 1. Therefore, the final step will assign a small weight to the residual, i.e.,", + "type": "text" + }, + { + "bbox": [ + 442, + 699, + 502, + 711 + ], + "score": 0.92, + "content": "( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 698, + 506, + 712 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 120, + 710, + 305, + 723 + ], + "spans": [ + { + "bbox": [ + 120, + 710, + 228, + 723 + ], + "score": 1.0, + "content": "and the aggregated feature", + "type": "text" + }, + { + "bbox": [ + 228, + 710, + 244, + 723 + ], + "score": 0.92, + "content": "\\mathbf { Y } _ { i } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 710, + 305, + 723 + ], + "score": 1.0, + "content": "will dominate.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 423, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 423, + 88 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 124, + 88 + ], + "score": 1.0, + "content": "The", + "type": "text" + }, + { + "bbox": [ + 124, + 74, + 129, + 83 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 69, + 423, + 88 + ], + "score": 1.0, + "content": "-th row of the proximal operator in Eq. (11) can be computed analytically", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 69, + 423, + 88 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 87, + 469, + 114 + ], + "lines": [ + { + "bbox": [ + 142, + 87, + 469, + 114 + ], + "spans": [ + { + "bbox": [ + 142, + 87, + 469, + 114 + ], + "score": 0.93, + "content": "\\left( \\mathbf { p r o x } _ { \\gamma \\lambda \\parallel \\cdot \\parallel _ { 2 1 } } ( \\mathbf { X } ) \\right) _ { i } = \\frac { \\mathbf { X } _ { i } } { \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } } \\operatorname* { m a x } ( \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } - \\gamma \\lambda , 0 ) = \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } } , 0 ) \\cdot \\mathbf { X } _ { i } .", + "type": "interline_equation", + "image_path": "2f99296051acdd82ba233470d0b60b0f2eed6592b1b372eecf8968426e2607f2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 142, + 87, + 469, + 114 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 117, + 508, + 140 + ], + "lines": [ + { + "bbox": [ + 105, + 117, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 426, + 130 + ], + "score": 1.0, + "content": "Note that the proximal operator returns 0 if the input vector is 0. Substituting", + "type": "text" + }, + { + "bbox": [ + 427, + 118, + 437, + 127 + ], + "score": 0.77, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 117, + 505, + 130 + ], + "score": 1.0, + "content": "in Eq. (13) with", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 394, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 148, + 139 + ], + "score": 0.91, + "content": "\\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 126, + 394, + 142 + ], + "score": 1.0, + "content": "and combining Eq. (11) and Eq. (12), then Eq. (12) becomes", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 117, + 505, + 142 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 143, + 457, + 159 + ], + "lines": [ + { + "bbox": [ + 154, + 143, + 457, + 159 + ], + "spans": [ + { + "bbox": [ + 154, + 143, + 457, + 159 + ], + "score": 0.88, + "content": "\\mathbf { X } _ { i } ^ { k + 1 } = ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } ( \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } ) = ( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } \\mathbf { Y } _ { i } ^ { k } , \\quad \\forall i \\in [ n ] ,", + "type": "interline_equation", + "image_path": "ddd1e17c5d829c98938c8b1388cd54b3e08aa5daae1060e5eac9a190d0f7b2f6.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 154, + 143, + 457, + 159 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 163, + 505, + 201 + ], + "lines": [ + { + "bbox": [ + 102, + 160, + 508, + 185 + ], + "spans": [ + { + "bbox": [ + 102, + 160, + 134, + 185 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 163, + 270, + 180 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\beta _ { i } : = \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\| \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } } , 0 ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 160, + 508, + 185 + ], + "score": 1.0, + "content": ". To summarize, the proposed adaptive message passing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "score": 1.0, + "content": "(AMP) scheme is showed in Figure 4, and a diagram is showed in Figure 3. In detail, AMP works as", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 188, + 143, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 143, + 202 + ], + "score": 1.0, + "content": "follows:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 102, + 160, + 508, + 202 + ] + }, + { + "type": "list", + "bbox": [ + 119, + 209, + 506, + 286 + ], + "lines": [ + { + "bbox": [ + 117, + 207, + 506, + 224 + ], + "spans": [ + { + "bbox": [ + 117, + 207, + 506, + 224 + ], + "score": 1.0, + "content": "• The first step takes a feature aggregation within the local neighbors with a self-loop weighted", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 219, + 205, + 234 + ], + "spans": [ + { + "bbox": [ + 126, + 219, + 140, + 234 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 141, + 221, + 200, + 233 + ], + "score": 0.92, + "content": "1 - 2 \\gamma ( 1 - \\lambda )", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 219, + 205, + 234 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 118, + 232, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 118, + 232, + 268, + 250 + ], + "score": 1.0, + "content": "• The second step computes a weight", + "type": "text" + }, + { + "bbox": [ + 268, + 235, + 309, + 248 + ], + "score": 0.92, + "content": "\\beta _ { i } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 232, + 366, + 250 + ], + "score": 1.0, + "content": "for each node", + "type": "text" + }, + { + "bbox": [ + 366, + 237, + 375, + 246 + ], + "score": 0.84, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 232, + 506, + 250 + ], + "score": 1.0, + "content": "depending on the local deviation", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 245, + 200, + 261 + ], + "spans": [ + { + "bbox": [ + 128, + 246, + 195, + 260 + ], + "score": 0.9, + "content": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 245, + 200, + 261 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 117, + 261, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 117, + 261, + 359, + 276 + ], + "score": 1.0, + "content": "• The final step takes a linear combination of input features", + "type": "text" + }, + { + "bbox": [ + 360, + 264, + 375, + 274 + ], + "score": 0.89, + "content": "\\mathbf { X } _ { \\mathrm { i n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 261, + 487, + 276 + ], + "score": 1.0, + "content": "and the aggregated features", + "type": "text" + }, + { + "bbox": [ + 488, + 262, + 502, + 273 + ], + "score": 0.88, + "content": "\\mathbf { Y } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 261, + 505, + 276 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 273, + 450, + 287 + ], + "spans": [ + { + "bbox": [ + 127, + 273, + 351, + 287 + ], + "score": 1.0, + "content": "where the node-wise residual is adaptively weighted by", + "type": "text" + }, + { + "bbox": [ + 351, + 275, + 378, + 285 + ], + "score": 0.91, + "content": "1 - \\beta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 273, + 436, + 287 + ], + "score": 1.0, + "content": "for each node", + "type": "text" + }, + { + "bbox": [ + 436, + 276, + 445, + 285 + ], + "score": 0.85, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 273, + 450, + 287 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + } + ], + "index": 10.5, + "bbox_fs": [ + 117, + 207, + 506, + 287 + ] + }, + { + "type": "image", + "bbox": [ + 122, + 305, + 416, + 385 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 305, + 416, + 385 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 305, + 416, + 385 + ], + "spans": [ + { + "bbox": [ + 122, + 305, + 416, + 385 + ], + "score": 0.393, + "type": "image", + "image_path": "31b59335ec89ac9c9b7a99814a2a6a89348cd1cc224bfd947556babaafd3c58b.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 122, + 305, + 416, + 331.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 122, + 331.6666666666667, + 416, + 358.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 122, + 358.33333333333337, + 416, + 385.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 217, + 391, + 395, + 404 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 389, + 395, + 407 + ], + "spans": [ + { + "bbox": [ + 215, + 389, + 395, + 407 + ], + "score": 1.0, + "content": "Figure 4: Adaptive Message Passing (AMP)", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + } + ], + "index": 16.0 + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 506, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 424, + 429 + ], + "score": 1.0, + "content": "The convergence guarantee of AMP and parameter setting for the stepsize", + "type": "text" + }, + { + "bbox": [ + 424, + 417, + 432, + 426 + ], + "score": 0.79, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 414, + 506, + 429 + ], + "score": 1.0, + "content": "are illustrated in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 424, + 507, + 441 + ], + "spans": [ + { + "bbox": [ + 104, + 424, + 507, + 441 + ], + "score": 1.0, + "content": "Theorem 1 and proved in Appendix A. According to Theorem 1, if we set γ = 14(1−λ) or γ = 12(1−λ) ,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 438, + 507, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 293, + 455 + ], + "score": 1.0, + "content": "then the first step of AMP can be simplified as", + "type": "text" + }, + { + "bbox": [ + 294, + 439, + 382, + 454 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\mathbf { Y } ^ { k } = \\frac { 1 } { 2 } \\mathbf { X } ^ { k } + \\frac { 1 } { 2 } \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 438, + 400, + 455 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 440, + 451, + 452 + ], + "score": 0.9, + "content": "\\mathbf { Y } ^ { k } = \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 438, + 507, + 455 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "The choice of stepsize will only impact the convergence speed but not the ultimate effect of AMP", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "score": 1.0, + "content": "when it convergences to the fixed point solution. We also discuss the computation complexity per", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 474, + 231, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 231, + 486 + ], + "score": 1.0, + "content": "iteration of AMP in Remark 1.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 414, + 507, + 486 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 505, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 507, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 352, + 504 + ], + "score": 1.0, + "content": "Theorem 1 (Convergence of AMP). Under the stepsize setting", + "type": "text" + }, + { + "bbox": [ + 367, + 485, + 507, + 505 + ], + "score": 1.0, + "content": "< 1(1−λ)kL˜k , the proposed adaptive", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 501, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 506, + 515 + ], + "score": 1.0, + "content": "message passing scheme (AMP) in Eq. (9) and Eq. (10) converges to the optimal solution of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 102, + 512, + 503, + 532 + ], + "spans": [ + { + "bbox": [ + 102, + 512, + 387, + 532 + ], + "score": 1.0, + "content": "problem defined in Eq. (8). In practice, it is sufficient to choose any", + "type": "text" + }, + { + "bbox": [ + 387, + 514, + 435, + 529 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\gamma < \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 512, + 461, + 532 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 461, + 513, + 503, + 527 + ], + "score": 0.92, + "content": "\\Vert \\tilde { \\mathbf { L } } \\Vert _ { 2 } \\leq 2", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 323, + 540 + ], + "score": 1.0, + "content": "Moreover, if the connected components of the graph", + "type": "text" + }, + { + "bbox": [ + 324, + 528, + 331, + 538 + ], + "score": 0.73, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "are not bipartite graphs, it is sufficient to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 538, + 254, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 538, + 136, + 556 + ], + "score": 1.0, + "content": "choose", + "type": "text" + }, + { + "bbox": [ + 137, + 539, + 184, + 555 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 538, + 209, + 556 + ], + "score": 1.0, + "content": "since", + "type": "text" + }, + { + "bbox": [ + 209, + 538, + 249, + 552 + ], + "score": 0.92, + "content": "\\Vert \\tilde { \\mathbf { L } } \\Vert _ { 2 } < 2", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 538, + 254, + 556 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 102, + 485, + 507, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 473, + 569 + ], + "score": 1.0, + "content": "Remark 1 (Computation complexity). AMP is as efficient as simple feature aggregation", + "type": "text" + }, + { + "bbox": [ + 473, + 556, + 505, + 568 + ], + "score": 0.87, + "content": "\\mathbf { X } _ { o u t } =", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 567, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 107, + 567, + 131, + 579 + ], + "score": 0.89, + "content": "\\tilde { \\mathbf { A } } \\mathbf { X } _ { i n }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 567, + 505, + 582 + ], + "score": 1.0, + "content": "because the additional computation cost from the second and third steps in Figure 4 is in the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 132, + 592 + ], + "score": 1.0, + "content": "order", + "type": "text" + }, + { + "bbox": [ + 132, + 579, + 160, + 591 + ], + "score": 0.93, + "content": "\\mathcal { O } ( n d )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 578, + 192, + 592 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 192, + 582, + 200, + 589 + ], + "score": 0.35, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 578, + 317, + 592 + ], + "score": 1.0, + "content": "is the number of nodes and", + "type": "text" + }, + { + "bbox": [ + 317, + 580, + 324, + 589 + ], + "score": 0.51, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "is the feature dimension. This is negligible", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 589, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 252, + 604 + ], + "score": 1.0, + "content": "compared with the computation cost", + "type": "text" + }, + { + "bbox": [ + 252, + 590, + 282, + 602 + ], + "score": 0.92, + "content": "\\mathcal { O } ( m d )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 589, + 506, + 604 + ], + "score": 1.0, + "content": "in feature aggregation, where m is the number of edges,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 600, + 503, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 471, + 614 + ], + "score": 1.0, + "content": "due to the fact that usually there are many more edges than nodes in real-world graphs, i.e.,", + "type": "text" + }, + { + "bbox": [ + 472, + 603, + 503, + 612 + ], + "score": 0.85, + "content": "m \\gg n", + "type": "inline_equation" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 555, + 506, + 614 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 625, + 227, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 228, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 228, + 639 + ], + "score": 1.0, + "content": "3.3 Interpretation of AMP", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 645, + 505, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "score": 1.0, + "content": "Interestingly, the proposed AMP has a simple and intuitive interpretation as adaptive residual connec-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 656, + 312, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 312, + 669 + ], + "score": 1.0, + "content": "tion, which aligns well with our design motivation:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 644, + 506, + 669 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 675, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 111, + 674, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 111, + 674, + 206, + 688 + ], + "score": 1.0, + "content": "• If the feature of node", + "type": "text" + }, + { + "bbox": [ + 207, + 677, + 216, + 687 + ], + "score": 0.83, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 674, + 237, + 688 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 237, + 676, + 263, + 688 + ], + "score": 0.91, + "content": "( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 674, + 505, + 688 + ], + "score": 1.0, + "content": ", is significantly inconsistent with its local neighbors, i.e., the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 118, + 686, + 504, + 702 + ], + "spans": [ + { + "bbox": [ + 118, + 686, + 195, + 702 + ], + "score": 1.0, + "content": "aggregated feature", + "type": "text" + }, + { + "bbox": [ + 195, + 687, + 210, + 700 + ], + "score": 0.91, + "content": "\\mathbf { Y } _ { i } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 686, + 308, + 702 + ], + "score": 1.0, + "content": ", then the local deviation", + "type": "text" + }, + { + "bbox": [ + 308, + 687, + 374, + 700 + ], + "score": 0.93, + "content": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 686, + 493, + 702 + ], + "score": 1.0, + "content": "will be large, which leads to a", + "type": "text" + }, + { + "bbox": [ + 494, + 688, + 504, + 699 + ], + "score": 0.87, + "content": "\\beta _ { i }", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 120, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 120, + 698, + 442, + 712 + ], + "score": 1.0, + "content": "close to 1. Therefore, the final step will assign a small weight to the residual, i.e.,", + "type": "text" + }, + { + "bbox": [ + 442, + 699, + 502, + 711 + ], + "score": 0.92, + "content": "( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 698, + 506, + 712 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 120, + 710, + 305, + 723 + ], + "spans": [ + { + "bbox": [ + 120, + 710, + 228, + 723 + ], + "score": 1.0, + "content": "and the aggregated feature", + "type": "text" + }, + { + "bbox": [ + 228, + 710, + 244, + 723 + ], + "score": 0.92, + "content": "\\mathbf { Y } _ { i } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 710, + 305, + 723 + ], + "score": 1.0, + "content": "will dominate.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5, + "bbox_fs": [ + 111, + 674, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 112, + 72, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 197, + 86 + ], + "score": 1.0, + "content": "• On the contrary, if", + "type": "text" + }, + { + "bbox": [ + 197, + 73, + 223, + 84 + ], + "score": 0.91, + "content": "( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 70, + 407, + 86 + ], + "score": 1.0, + "content": "is already consistent with its local neighbors,", + "type": "text" + }, + { + "bbox": [ + 407, + 72, + 474, + 85 + ], + "score": 0.92, + "content": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 70, + 506, + 86 + ], + "score": 1.0, + "content": "will be", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 120, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 120, + 83, + 214, + 96 + ], + "score": 1.0, + "content": "small, which leads to a", + "type": "text" + }, + { + "bbox": [ + 214, + 85, + 224, + 95 + ], + "score": 0.89, + "content": "\\beta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "close to 0. Thus, the residual will dominate, which is reasonable since", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 120, + 94, + 327, + 107 + ], + "spans": [ + { + "bbox": [ + 120, + 94, + 327, + 107 + ], + "score": 1.0, + "content": "there is less need to aggregate features in this case.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 112, + 108, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 110, + 106, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 110, + 106, + 263, + 122 + ], + "score": 1.0, + "content": "• To summarize, the local deviation", + "type": "text" + }, + { + "bbox": [ + 263, + 108, + 330, + 121 + ], + "score": 0.93, + "content": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 106, + 472, + 122 + ], + "score": 1.0, + "content": "provides a natural transition from", + "type": "text" + }, + { + "bbox": [ + 473, + 109, + 505, + 120 + ], + "score": 0.9, + "content": "\\beta _ { i } \\to 1", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 120, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 120, + 120, + 131, + 131 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 131, + 120, + 163, + 131 + ], + "score": 0.91, + "content": "\\beta _ { i } \\to 0", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 120, + 327, + 131 + ], + "score": 1.0, + "content": ", and the transition can be modulated by", + "type": "text" + }, + { + "bbox": [ + 327, + 120, + 335, + 129 + ], + "score": 0.71, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 120, + 506, + 131 + ], + "score": 1.0, + "content": "which can be either learned or tuned as a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 131, + 506, + 143 + ], + "spans": [ + { + "bbox": [ + 119, + 131, + 506, + 143 + ], + "score": 1.0, + "content": "hyperparameter through cross-validation. This transition provides an node-wise adaptive residual", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 142, + 299, + 154 + ], + "spans": [ + { + "bbox": [ + 120, + 142, + 299, + 154 + ], + "score": 1.0, + "content": "connection for the message passing scheme.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 238 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 214, + 173 + ], + "score": 1.0, + "content": "Adaptivity for abnormal", + "type": "text" + }, + { + "bbox": [ + 214, + 162, + 223, + 171 + ], + "score": 0.38, + "content": "\\pmb { \\& }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 161, + 504, + 173 + ], + "score": 1.0, + "content": "normal features. According to the homophily assumption on graph", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "score": 1.0, + "content": "structure data [17, 18, 19, 3], the feature representations of normal features should be more consistent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "with local neighbors than abnormal features. As a result, AMP will assign more residual (i.e., smaller", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 114, + 205 + ], + "score": 0.72, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 194, + 308, + 207 + ], + "score": 1.0, + "content": ") to normal features but less residual (i.e., larger", + "type": "text" + }, + { + "bbox": [ + 308, + 194, + 316, + 205 + ], + "score": 0.78, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 194, + 505, + 207 + ], + "score": 1.0, + "content": ") to abnormal features, providing a customized", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "tradeoff between feature aggregation and residual connection. Consequently, it can promote both", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "the resilience to abnormal features and the performance on normal features. Above discussion also", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 226, + 488, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 253, + 239 + ], + "score": 1.0, + "content": "implies a clear physical meaning for", + "type": "text" + }, + { + "bbox": [ + 254, + 227, + 261, + 238 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 226, + 488, + 239 + ], + "score": 1.0, + "content": "in AMP, and we formally define it as the adaptive score.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 504, + 274 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 288, + 254 + ], + "score": 1.0, + "content": "Definition 1 (Adaptive score). The variables", + "type": "text" + }, + { + "bbox": [ + 288, + 240, + 343, + 252 + ], + "score": 0.9, + "content": "\\{ \\beta _ { 1 } , \\cdot \\cdot \\cdot , \\beta _ { n } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 238, + 505, + 254 + ], + "score": 1.0, + "content": "in the adaptive message passing scheme", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 329, + 264 + ], + "score": 1.0, + "content": "(AMP) are defined as the adaptive scores for nodes", + "type": "text" + }, + { + "bbox": [ + 329, + 252, + 383, + 263 + ], + "score": 0.93, + "content": "\\{ v _ { 1 } , \\cdots , v _ { n } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 250, + 477, + 264 + ], + "score": 1.0, + "content": "respectively in graph", + "type": "text" + }, + { + "bbox": [ + 478, + 252, + 485, + 262 + ], + "score": 0.57, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 250, + 505, + 264 + ], + "score": 1.0, + "content": ". In", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 262, + 417, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 192, + 275 + ], + "score": 1.0, + "content": "particular, the larger", + "type": "text" + }, + { + "bbox": [ + 192, + 263, + 202, + 274 + ], + "score": 0.86, + "content": "\\beta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 262, + 353, + 275 + ], + "score": 1.0, + "content": "is, the more likely the feature of node", + "type": "text" + }, + { + "bbox": [ + 354, + 264, + 363, + 273 + ], + "score": 0.83, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 262, + 417, + 275 + ], + "score": 1.0, + "content": "is abnormal.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 289 + ], + "score": 1.0, + "content": "Remark 2 (Nonlinear smoother). Different from most existing message passing scheme which are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 285, + 504, + 300 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 388, + 300 + ], + "score": 1.0, + "content": "linear smoothers, AMP is a nonlinear smoother because the weights", + "type": "text" + }, + { + "bbox": [ + 388, + 287, + 408, + 299 + ], + "score": 0.92, + "content": "\\{ \\beta _ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 285, + 488, + 300 + ], + "score": 1.0, + "content": "are computed from", + "type": "text" + }, + { + "bbox": [ + 489, + 286, + 504, + 297 + ], + "score": 0.86, + "content": "\\mathbf { Y } ^ { k }", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 298, + 491, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 124, + 310 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 298, + 139, + 309 + ], + "score": 0.86, + "content": "{ \\bf X } _ { i n }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 298, + 491, + 310 + ], + "score": 1.0, + "content": ". This nonlinearity is the key to achieve adaptive residual connection for different nodes.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 321, + 233, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 234, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 234, + 334 + ], + "score": 1.0, + "content": "3.4 The Model Architecture", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 105, + 341, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 356 + ], + "score": 1.0, + "content": "The proposed adaptive message passing (AMP) can be used as a building block in many GNN models", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "to improve the resilience to abnormal node features. In this work, we choose the the decoupled", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 364, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 376 + ], + "score": 1.0, + "content": "architectures as APPNP [10] and DAGNN [20], and propose the Adaptive residual GNN (AirGNN):", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 377, + 359, + 406 + ], + "lines": [ + { + "bbox": [ + 252, + 377, + 359, + 406 + ], + "spans": [ + { + "bbox": [ + 252, + 377, + 359, + 406 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\mathbf { X } _ { \\mathrm { i n } } = h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } ) , } \\\\ & { \\mathbf { Y } _ { \\mathrm { p r e } } = \\mathbf { A } \\mathbf { M } \\mathbf { P } \\left( \\mathbf { X } _ { \\mathrm { i n } } , K , \\lambda \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "f62553c679b7785ed1d60b6a031508aaff7f04a8c520ca11429283397931d80c.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 252, + 377, + 359, + 391.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 252, + 391.5, + 359, + 406.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 107, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 107, + 408, + 128, + 420 + ], + "score": 0.9, + "content": "h _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 408, + 418, + 420 + ], + "score": 1.0, + "content": "is any machine learning model parameterized by learnable parameters", + "type": "text" + }, + { + "bbox": [ + 418, + 408, + 424, + 418 + ], + "score": 0.76, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 408, + 505, + 420 + ], + "score": 1.0, + "content": ", such as multilayer", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 417, + 507, + 433 + ], + "spans": [ + { + "bbox": [ + 104, + 417, + 196, + 433 + ], + "score": 1.0, + "content": "perceptrons (MLPs).", + "type": "text" + }, + { + "bbox": [ + 197, + 419, + 254, + 431 + ], + "score": 0.92, + "content": "\\mathbf { X } _ { \\mathrm { f e a } } \\in \\mathbb { R } ^ { n \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 417, + 443, + 433 + ], + "score": 1.0, + "content": "denotes the initial node features. The model", + "type": "text" + }, + { + "bbox": [ + 443, + 420, + 465, + 432 + ], + "score": 0.92, + "content": "h _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 417, + 507, + 433 + ], + "score": 1.0, + "content": "will first", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 257, + 443 + ], + "score": 1.0, + "content": "transform the initial node features as", + "type": "text" + }, + { + "bbox": [ + 257, + 430, + 322, + 443 + ], + "score": 0.92, + "content": "\\mathbf { X } _ { \\mathrm { i n } } = h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 430, + 374, + 443 + ], + "score": 1.0, + "content": ". AMP takes", + "type": "text" + }, + { + "bbox": [ + 374, + 431, + 411, + 443 + ], + "score": 0.95, + "content": "h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "as input, and performs", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 107, + 443, + 117, + 452 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 441, + 279, + 455 + ], + "score": 1.0, + "content": "steps of AMP with the hyperparameter", + "type": "text" + }, + { + "bbox": [ + 279, + 443, + 286, + 452 + ], + "score": 0.75, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 441, + 505, + 455 + ], + "score": 1.0, + "content": ". Similar to the majority of existing GNN models, the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "training objective is the cross-entropy classification loss on the labeled nodes, and the whole model", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "is trained in an end-to-end way. Note that AirGNN is very efficient as explained in Remark 1, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 259, + 488 + ], + "score": 1.0, + "content": "it only requires two hyperparameters", + "type": "text" + }, + { + "bbox": [ + 259, + 475, + 270, + 484 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 474, + 288, + 488 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 288, + 475, + 295, + 484 + ], + "score": 0.82, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 474, + 507, + 488 + ], + "score": 1.0, + "content": "without introducing additional parameters to learn,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 484, + 275, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 275, + 498 + ], + "score": 1.0, + "content": "which could reduce the risk of overfitting.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 107, + 511, + 186, + 525 + ], + "lines": [ + { + "bbox": [ + 104, + 509, + 189, + 529 + ], + "spans": [ + { + "bbox": [ + 104, + 509, + 189, + 529 + ], + "score": 1.0, + "content": "4 Experiment", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "In this section, we aim to verify the effective of the proposed adaptive message passing scheme", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "(AMP) and the AirGNN model through the semi-supervised node classification tasks. Specifically,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 104, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "we try to answer the following questions: (1) How does AirGNN perform on abnormal and normal", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 567, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 582 + ], + "score": 1.0, + "content": "features? (Section 4.2 and 4.3) and (2) How does AirGNN work by adjusting the adaptive residual?", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 578, + 161, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 161, + 592 + ], + "score": 1.0, + "content": "(Section 4.4)", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 107, + 603, + 225, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 226, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 226, + 619 + ], + "score": 1.0, + "content": "4.1 Experimental Settings", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "Datasets and baselines. We conduct experiments on 8 real-world datasets including three citation", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 634, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 647 + ], + "score": 1.0, + "content": "graphs, i.e., Cora, Citeseer, Pubmed [21], two co-authorship graphs, i.e., Coauthor CS and Coauthor", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "Physics [22], two co-purchase graphs, i.e., Amazon Computers and Amazon Photo [22], and one", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "OGB dataset, i.e., ogbn-arxiv [23]. Due to the space limit, we only present the results on Cora,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "Citeseer, and Pubmed in this section, but defer the results on other datasets to Appendix D.1. More", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "details about the data statistics and data splits are summarized in Appendix B. The proposed AirGNN", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 687, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 507, + 702 + ], + "score": 1.0, + "content": "is compared with representative GNNs, including GCN [3], GAT [6], APPNP [10] and GCNII [9].", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "We defer the comparison with the variants of APPNP and Robust GCN [24] to Appendix D.3 and D.4", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 711, + 159, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 159, + 725 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 45 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 112, + 72, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 197, + 86 + ], + "score": 1.0, + "content": "• On the contrary, if", + "type": "text" + }, + { + "bbox": [ + 197, + 73, + 223, + 84 + ], + "score": 0.91, + "content": "( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 70, + 407, + 86 + ], + "score": 1.0, + "content": "is already consistent with its local neighbors,", + "type": "text" + }, + { + "bbox": [ + 407, + 72, + 474, + 85 + ], + "score": 0.92, + "content": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 70, + 506, + 86 + ], + "score": 1.0, + "content": "will be", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 120, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 120, + 83, + 214, + 96 + ], + "score": 1.0, + "content": "small, which leads to a", + "type": "text" + }, + { + "bbox": [ + 214, + 85, + 224, + 95 + ], + "score": 0.89, + "content": "\\beta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "close to 0. Thus, the residual will dominate, which is reasonable since", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 120, + 94, + 327, + 107 + ], + "spans": [ + { + "bbox": [ + 120, + 94, + 327, + 107 + ], + "score": 1.0, + "content": "there is less need to aggregate features in this case.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 111, + 70, + 506, + 107 + ] + }, + { + "type": "text", + "bbox": [ + 112, + 108, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 110, + 106, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 110, + 106, + 263, + 122 + ], + "score": 1.0, + "content": "• To summarize, the local deviation", + "type": "text" + }, + { + "bbox": [ + 263, + 108, + 330, + 121 + ], + "score": 0.93, + "content": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 106, + 472, + 122 + ], + "score": 1.0, + "content": "provides a natural transition from", + "type": "text" + }, + { + "bbox": [ + 473, + 109, + 505, + 120 + ], + "score": 0.9, + "content": "\\beta _ { i } \\to 1", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 120, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 120, + 120, + 131, + 131 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 131, + 120, + 163, + 131 + ], + "score": 0.91, + "content": "\\beta _ { i } \\to 0", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 120, + 327, + 131 + ], + "score": 1.0, + "content": ", and the transition can be modulated by", + "type": "text" + }, + { + "bbox": [ + 327, + 120, + 335, + 129 + ], + "score": 0.71, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 120, + 506, + 131 + ], + "score": 1.0, + "content": "which can be either learned or tuned as a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 131, + 506, + 143 + ], + "spans": [ + { + "bbox": [ + 119, + 131, + 506, + 143 + ], + "score": 1.0, + "content": "hyperparameter through cross-validation. This transition provides an node-wise adaptive residual", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 142, + 299, + 154 + ], + "spans": [ + { + "bbox": [ + 120, + 142, + 299, + 154 + ], + "score": 1.0, + "content": "connection for the message passing scheme.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 110, + 106, + 506, + 154 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 238 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 214, + 173 + ], + "score": 1.0, + "content": "Adaptivity for abnormal", + "type": "text" + }, + { + "bbox": [ + 214, + 162, + 223, + 171 + ], + "score": 0.38, + "content": "\\pmb { \\& }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 161, + 504, + 173 + ], + "score": 1.0, + "content": "normal features. According to the homophily assumption on graph", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 184 + ], + "score": 1.0, + "content": "structure data [17, 18, 19, 3], the feature representations of normal features should be more consistent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "with local neighbors than abnormal features. As a result, AMP will assign more residual (i.e., smaller", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 114, + 205 + ], + "score": 0.72, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 194, + 308, + 207 + ], + "score": 1.0, + "content": ") to normal features but less residual (i.e., larger", + "type": "text" + }, + { + "bbox": [ + 308, + 194, + 316, + 205 + ], + "score": 0.78, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 194, + 505, + 207 + ], + "score": 1.0, + "content": ") to abnormal features, providing a customized", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "tradeoff between feature aggregation and residual connection. Consequently, it can promote both", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "the resilience to abnormal features and the performance on normal features. Above discussion also", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 226, + 488, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 253, + 239 + ], + "score": 1.0, + "content": "implies a clear physical meaning for", + "type": "text" + }, + { + "bbox": [ + 254, + 227, + 261, + 238 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 226, + 488, + 239 + ], + "score": 1.0, + "content": "in AMP, and we formally define it as the adaptive score.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 161, + 506, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 504, + 274 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 288, + 254 + ], + "score": 1.0, + "content": "Definition 1 (Adaptive score). The variables", + "type": "text" + }, + { + "bbox": [ + 288, + 240, + 343, + 252 + ], + "score": 0.9, + "content": "\\{ \\beta _ { 1 } , \\cdot \\cdot \\cdot , \\beta _ { n } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 238, + 505, + 254 + ], + "score": 1.0, + "content": "in the adaptive message passing scheme", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 329, + 264 + ], + "score": 1.0, + "content": "(AMP) are defined as the adaptive scores for nodes", + "type": "text" + }, + { + "bbox": [ + 329, + 252, + 383, + 263 + ], + "score": 0.93, + "content": "\\{ v _ { 1 } , \\cdots , v _ { n } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 250, + 477, + 264 + ], + "score": 1.0, + "content": "respectively in graph", + "type": "text" + }, + { + "bbox": [ + 478, + 252, + 485, + 262 + ], + "score": 0.57, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 250, + 505, + 264 + ], + "score": 1.0, + "content": ". In", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 262, + 417, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 192, + 275 + ], + "score": 1.0, + "content": "particular, the larger", + "type": "text" + }, + { + "bbox": [ + 192, + 263, + 202, + 274 + ], + "score": 0.86, + "content": "\\beta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 262, + 353, + 275 + ], + "score": 1.0, + "content": "is, the more likely the feature of node", + "type": "text" + }, + { + "bbox": [ + 354, + 264, + 363, + 273 + ], + "score": 0.83, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 262, + 417, + 275 + ], + "score": 1.0, + "content": "is abnormal.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 238, + 505, + 275 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 276, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 289 + ], + "score": 1.0, + "content": "Remark 2 (Nonlinear smoother). Different from most existing message passing scheme which are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 285, + 504, + 300 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 388, + 300 + ], + "score": 1.0, + "content": "linear smoothers, AMP is a nonlinear smoother because the weights", + "type": "text" + }, + { + "bbox": [ + 388, + 287, + 408, + 299 + ], + "score": 0.92, + "content": "\\{ \\beta _ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 285, + 488, + 300 + ], + "score": 1.0, + "content": "are computed from", + "type": "text" + }, + { + "bbox": [ + 489, + 286, + 504, + 297 + ], + "score": 0.86, + "content": "\\mathbf { Y } ^ { k }", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 298, + 491, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 124, + 310 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 298, + 139, + 309 + ], + "score": 0.86, + "content": "{ \\bf X } _ { i n }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 298, + 491, + 310 + ], + "score": 1.0, + "content": ". This nonlinearity is the key to achieve adaptive residual connection for different nodes.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 275, + 505, + 310 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 321, + 233, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 234, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 234, + 334 + ], + "score": 1.0, + "content": "3.4 The Model Architecture", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 342, + 505, + 375 + ], + "lines": [ + { + "bbox": [ + 105, + 341, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 356 + ], + "score": 1.0, + "content": "The proposed adaptive message passing (AMP) can be used as a building block in many GNN models", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "to improve the resilience to abnormal node features. In this work, we choose the the decoupled", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 364, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 376 + ], + "score": 1.0, + "content": "architectures as APPNP [10] and DAGNN [20], and propose the Adaptive residual GNN (AirGNN):", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 341, + 506, + 376 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 252, + 377, + 359, + 406 + ], + "lines": [ + { + "bbox": [ + 252, + 377, + 359, + 406 + ], + "spans": [ + { + "bbox": [ + 252, + 377, + 359, + 406 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\mathbf { X } _ { \\mathrm { i n } } = h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } ) , } \\\\ & { \\mathbf { Y } _ { \\mathrm { p r e } } = \\mathbf { A } \\mathbf { M } \\mathbf { P } \\left( \\mathbf { X } _ { \\mathrm { i n } } , K , \\lambda \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "f62553c679b7785ed1d60b6a031508aaff7f04a8c520ca11429283397931d80c.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 252, + 377, + 359, + 391.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 252, + 391.5, + 359, + 406.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 107, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 107, + 408, + 128, + 420 + ], + "score": 0.9, + "content": "h _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 408, + 418, + 420 + ], + "score": 1.0, + "content": "is any machine learning model parameterized by learnable parameters", + "type": "text" + }, + { + "bbox": [ + 418, + 408, + 424, + 418 + ], + "score": 0.76, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 408, + 505, + 420 + ], + "score": 1.0, + "content": ", such as multilayer", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 417, + 507, + 433 + ], + "spans": [ + { + "bbox": [ + 104, + 417, + 196, + 433 + ], + "score": 1.0, + "content": "perceptrons (MLPs).", + "type": "text" + }, + { + "bbox": [ + 197, + 419, + 254, + 431 + ], + "score": 0.92, + "content": "\\mathbf { X } _ { \\mathrm { f e a } } \\in \\mathbb { R } ^ { n \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 417, + 443, + 433 + ], + "score": 1.0, + "content": "denotes the initial node features. The model", + "type": "text" + }, + { + "bbox": [ + 443, + 420, + 465, + 432 + ], + "score": 0.92, + "content": "h _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 417, + 507, + 433 + ], + "score": 1.0, + "content": "will first", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 257, + 443 + ], + "score": 1.0, + "content": "transform the initial node features as", + "type": "text" + }, + { + "bbox": [ + 257, + 430, + 322, + 443 + ], + "score": 0.92, + "content": "\\mathbf { X } _ { \\mathrm { i n } } = h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 430, + 374, + 443 + ], + "score": 1.0, + "content": ". AMP takes", + "type": "text" + }, + { + "bbox": [ + 374, + 431, + 411, + 443 + ], + "score": 0.95, + "content": "h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "as input, and performs", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 107, + 443, + 117, + 452 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 441, + 279, + 455 + ], + "score": 1.0, + "content": "steps of AMP with the hyperparameter", + "type": "text" + }, + { + "bbox": [ + 279, + 443, + 286, + 452 + ], + "score": 0.75, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 441, + 505, + 455 + ], + "score": 1.0, + "content": ". Similar to the majority of existing GNN models, the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "training objective is the cross-entropy classification loss on the labeled nodes, and the whole model", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "is trained in an end-to-end way. Note that AirGNN is very efficient as explained in Remark 1, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 259, + 488 + ], + "score": 1.0, + "content": "it only requires two hyperparameters", + "type": "text" + }, + { + "bbox": [ + 259, + 475, + 270, + 484 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 474, + 288, + 488 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 288, + 475, + 295, + 484 + ], + "score": 0.82, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 474, + 507, + 488 + ], + "score": 1.0, + "content": "without introducing additional parameters to learn,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 484, + 275, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 275, + 498 + ], + "score": 1.0, + "content": "which could reduce the risk of overfitting.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 408, + 507, + 498 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 511, + 186, + 525 + ], + "lines": [ + { + "bbox": [ + 104, + 509, + 189, + 529 + ], + "spans": [ + { + "bbox": [ + 104, + 509, + 189, + 529 + ], + "score": 1.0, + "content": "4 Experiment", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "In this section, we aim to verify the effective of the proposed adaptive message passing scheme", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "(AMP) and the AirGNN model through the semi-supervised node classification tasks. Specifically,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 104, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "we try to answer the following questions: (1) How does AirGNN perform on abnormal and normal", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 567, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 582 + ], + "score": 1.0, + "content": "features? (Section 4.2 and 4.3) and (2) How does AirGNN work by adjusting the adaptive residual?", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 578, + 161, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 161, + 592 + ], + "score": 1.0, + "content": "(Section 4.4)", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37, + "bbox_fs": [ + 104, + 535, + 506, + 592 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 603, + 225, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 600, + 226, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 226, + 619 + ], + "score": 1.0, + "content": "4.1 Experimental Settings", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "Datasets and baselines. We conduct experiments on 8 real-world datasets including three citation", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 634, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 647 + ], + "score": 1.0, + "content": "graphs, i.e., Cora, Citeseer, Pubmed [21], two co-authorship graphs, i.e., Coauthor CS and Coauthor", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "Physics [22], two co-purchase graphs, i.e., Amazon Computers and Amazon Photo [22], and one", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "OGB dataset, i.e., ogbn-arxiv [23]. Due to the space limit, we only present the results on Cora,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "Citeseer, and Pubmed in this section, but defer the results on other datasets to Appendix D.1. More", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "details about the data statistics and data splits are summarized in Appendix B. The proposed AirGNN", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 687, + 507, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 507, + 702 + ], + "score": 1.0, + "content": "is compared with representative GNNs, including GCN [3], GAT [6], APPNP [10] and GCNII [9].", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 698, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 506, + 713 + ], + "score": 1.0, + "content": "We defer the comparison with the variants of APPNP and Robust GCN [24] to Appendix D.3 and D.4", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 711, + 159, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 159, + 725 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 45, + "bbox_fs": [ + 104, + 624, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "Parameter settings. For all baselines, we follow the best hyperparameter settings in their original", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 83, + 312, + 96 + ], + "score": 1.0, + "content": "papers. Additionally, we tune a best residual weight", + "type": "text" + }, + { + "bbox": [ + 313, + 85, + 320, + 93 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 83, + 464, + 96 + ], + "score": 1.0, + "content": "for APPNP and GCNII in the range", + "type": "text" + }, + { + "bbox": [ + 464, + 84, + 484, + 96 + ], + "score": 0.54, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 83, + 506, + 96 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 92, + 506, + 109 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 322, + 109 + ], + "score": 1.0, + "content": "AirGNN, we use a two-layer MLP as the base model", + "type": "text" + }, + { + "bbox": [ + 322, + 95, + 344, + 106 + ], + "score": 0.91, + "content": "h _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 92, + 506, + 109 + ], + "score": 1.0, + "content": ", following APPNP. We fix the learning", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 102, + 507, + 124 + ], + "spans": [ + { + "bbox": [ + 103, + 102, + 386, + 124 + ], + "score": 1.0, + "content": "rate 0.01, dropout 0.8, and weight decay 0.0005. Moreover, we set", + "type": "text" + }, + { + "bbox": [ + 387, + 105, + 436, + 120 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 102, + 507, + 124 + ], + "score": 1.0, + "content": "as suggested by", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 118, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 203, + 133 + ], + "score": 1.0, + "content": "Theorem 1. We choose", + "type": "text" + }, + { + "bbox": [ + 204, + 119, + 238, + 129 + ], + "score": 0.9, + "content": "K = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 118, + 277, + 133 + ], + "score": 1.0, + "content": "and tune", + "type": "text" + }, + { + "bbox": [ + 277, + 119, + 284, + 129 + ], + "score": 0.76, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 118, + 336, + 133 + ], + "score": 1.0, + "content": "in the range", + "type": "text" + }, + { + "bbox": [ + 337, + 118, + 357, + 131 + ], + "score": 0.25, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 118, + 506, + 133 + ], + "score": 1.0, + "content": ". Adam optimizer [25] is used in all", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 130, + 444, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 130, + 444, + 144 + ], + "score": 1.0, + "content": "experiments. We run all experiments by 10 times, and report the mean and variance.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 146, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "Evaluation setting. We assess the performance of all models under two types of abnormal feature", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 169 + ], + "score": 1.0, + "content": "scenarios, including noisy features and adversarial features. The abnormal features are injected to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "randomly selected test nodes after model training. By default, all hyperparameters are tuned according", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 179, + 504, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 496, + 191 + ], + "score": 1.0, + "content": "to the performance on validation sets when the dataset is clean. If tuning the hyperparameter", + "type": "text" + }, + { + "bbox": [ + 497, + 179, + 504, + 189 + ], + "score": 0.76, + "content": "\\lambda", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 189, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 203 + ], + "score": 1.0, + "content": "of AirGNN according to the validation sets after injecting abnormal features, the performance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "will be even better, as discussed in Appendix D.2. The performance on clean data are showed in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "Appendix D.5 to demonstrate that AirGNN doesn’t need to sacrifice accuracy for better robustness", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 223, + 214, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 214, + 234 + ], + "score": 1.0, + "content": "against abnormal features.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 107, + 248, + 327, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 329, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 329, + 262 + ], + "score": 1.0, + "content": "4.2 Performance Comparison with Noisy Features", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "In this subsection, we consider the abnormal features in the noisy feature scenario. Specifically, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 280, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 291 + ], + "score": 1.0, + "content": "simulate the noisy features by assigning a subset of the nodes with random features sampled from a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "multivariate standard Gaussian distribution. Note that the selection of noise subsets has a apparent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 302, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 313 + ], + "score": 1.0, + "content": "impact on the performance since some nodes are less vulnerable to abnormal features while others", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 506, + 325 + ], + "score": 1.0, + "content": "are more vulnerable. To reduce such variance, we report the average performance over 10 times of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 322, + 507, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 322, + 507, + 335 + ], + "score": 1.0, + "content": "random selection of the noise node sets, similar to the settings in the preliminary study in Section 2.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "We report the node classification test accuracy on abnormal (noisy) features and normal features in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 494, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 494, + 358 + ], + "score": 1.0, + "content": "Figure 5 and Figure 6, separately, under varying noisy ratio. From these figures, we can observe:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 116, + 366, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 120, + 366, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 120, + 366, + 506, + 377 + ], + "score": 1.0, + "content": "• Figure 5 shows that AirGNN significantly outperforms all baselines on all datasets in terms of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 127, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "the performance on noisy nodes. This verifies that AMP is able to improve the resilience to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 388, + 354, + 400 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 354, + 400 + ], + "score": 1.0, + "content": "noisy features, aligning well with the design motivation.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 121, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 121, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "• Figure 6 shows that AirGNN promotes the performance on normal nodes when abnormal nodes", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 128, + 415, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 128, + 415, + 505, + 426 + ], + "score": 1.0, + "content": "exist. This is because AMP can remove some abnormal features which are detrimental to normal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 425, + 157, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 425, + 157, + 437 + ], + "score": 1.0, + "content": "nodes.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + }, + { + "type": "image", + "bbox": [ + 101, + 447, + 497, + 561 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 101, + 447, + 497, + 561 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 101, + 447, + 497, + 561 + ], + "spans": [ + { + "bbox": [ + 101, + 447, + 497, + 561 + ], + "score": 0.973, + "type": "image", + "image_path": "cb981f452a6f87e6cccc28935d8e1aa89643dc8d85a1008f917f9b4fedf8a508.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 101, + 447, + 497, + 485.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 101, + 485.0, + 497, + 523.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 101, + 523.0, + 497, + 561.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 172, + 568, + 438, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 172, + 567, + 438, + 581 + ], + "spans": [ + { + "bbox": [ + 172, + 567, + 438, + 581 + ], + "score": 1.0, + "content": "Figure 5: Node classification accuracy on abnormal (noisy) nodes", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + } + ], + "index": 31.0 + }, + { + "type": "title", + "bbox": [ + 106, + 604, + 354, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 355, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 355, + 617 + ], + "score": 1.0, + "content": "4.3 Performance Comparison with Adversarial Features", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 504, + 636 + ], + "score": 1.0, + "content": "In this subsection, we consider the abnormal feature scenario when the node features are maliciously", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 636, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 647 + ], + "score": 1.0, + "content": "attacked by the attacker to manipulate the prediction of GNNs. We use the Nettack [26] implemented", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 645, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 659 + ], + "score": 1.0, + "content": "in DeepRobust3 [27], a PyTorch library for adversarial attacks and defenses, to generate the adversarial", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 658, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 669 + ], + "score": 1.0, + "content": "features. We randomly choose 40 test nodes as the targeted nodes, and assess the performance under", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 668, + 504, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 239, + 680 + ], + "score": 1.0, + "content": "increasing perturbation budgets", + "type": "text" + }, + { + "bbox": [ + 239, + 668, + 322, + 680 + ], + "score": 0.91, + "content": "\\{ 0 , 5 , 1 0 , 2 0 , 5 0 , 8 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 668, + 504, + 680 + ], + "score": 1.0, + "content": ", where the perturbation numbers denote the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "number of feature dimensions that can be manipulated. The node classification accuracy on these", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 506, + 703 + ], + "score": 1.0, + "content": "attacked nodes are showed in Figure 7. From these figures, we can make the following observations:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 712, + 298, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 300, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 300, + 725 + ], + "score": 1.0, + "content": "3https://github.com/DSE-MSU/DeepRobust", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "Parameter settings. For all baselines, we follow the best hyperparameter settings in their original", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 83, + 312, + 96 + ], + "score": 1.0, + "content": "papers. Additionally, we tune a best residual weight", + "type": "text" + }, + { + "bbox": [ + 313, + 85, + 320, + 93 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 83, + 464, + 96 + ], + "score": 1.0, + "content": "for APPNP and GCNII in the range", + "type": "text" + }, + { + "bbox": [ + 464, + 84, + 484, + 96 + ], + "score": 0.54, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 83, + 506, + 96 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 92, + 506, + 109 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 322, + 109 + ], + "score": 1.0, + "content": "AirGNN, we use a two-layer MLP as the base model", + "type": "text" + }, + { + "bbox": [ + 322, + 95, + 344, + 106 + ], + "score": 0.91, + "content": "h _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 92, + 506, + 109 + ], + "score": 1.0, + "content": ", following APPNP. We fix the learning", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 102, + 507, + 124 + ], + "spans": [ + { + "bbox": [ + 103, + 102, + 386, + 124 + ], + "score": 1.0, + "content": "rate 0.01, dropout 0.8, and weight decay 0.0005. Moreover, we set", + "type": "text" + }, + { + "bbox": [ + 387, + 105, + 436, + 120 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 102, + 507, + 124 + ], + "score": 1.0, + "content": "as suggested by", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 118, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 203, + 133 + ], + "score": 1.0, + "content": "Theorem 1. We choose", + "type": "text" + }, + { + "bbox": [ + 204, + 119, + 238, + 129 + ], + "score": 0.9, + "content": "K = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 118, + 277, + 133 + ], + "score": 1.0, + "content": "and tune", + "type": "text" + }, + { + "bbox": [ + 277, + 119, + 284, + 129 + ], + "score": 0.76, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 118, + 336, + 133 + ], + "score": 1.0, + "content": "in the range", + "type": "text" + }, + { + "bbox": [ + 337, + 118, + 357, + 131 + ], + "score": 0.25, + "content": "[ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 118, + 506, + 133 + ], + "score": 1.0, + "content": ". Adam optimizer [25] is used in all", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 130, + 444, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 130, + 444, + 144 + ], + "score": 1.0, + "content": "experiments. We run all experiments by 10 times, and report the mean and variance.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 103, + 73, + 507, + 144 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 146, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "Evaluation setting. We assess the performance of all models under two types of abnormal feature", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 169 + ], + "score": 1.0, + "content": "scenarios, including noisy features and adversarial features. The abnormal features are injected to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "randomly selected test nodes after model training. By default, all hyperparameters are tuned according", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 179, + 504, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 496, + 191 + ], + "score": 1.0, + "content": "to the performance on validation sets when the dataset is clean. If tuning the hyperparameter", + "type": "text" + }, + { + "bbox": [ + 497, + 179, + 504, + 189 + ], + "score": 0.76, + "content": "\\lambda", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 189, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 505, + 203 + ], + "score": 1.0, + "content": "of AirGNN according to the validation sets after injecting abnormal features, the performance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "will be even better, as discussed in Appendix D.2. The performance on clean data are showed in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "Appendix D.5 to demonstrate that AirGNN doesn’t need to sacrifice accuracy for better robustness", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 223, + 214, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 214, + 234 + ], + "score": 1.0, + "content": "against abnormal features.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 146, + 506, + 234 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 248, + 327, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 329, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 329, + 262 + ], + "score": 1.0, + "content": "4.2 Performance Comparison with Noisy Features", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "In this subsection, we consider the abnormal features in the noisy feature scenario. Specifically, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 280, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 291 + ], + "score": 1.0, + "content": "simulate the noisy features by assigning a subset of the nodes with random features sampled from a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "multivariate standard Gaussian distribution. Note that the selection of noise subsets has a apparent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 302, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 313 + ], + "score": 1.0, + "content": "impact on the performance since some nodes are less vulnerable to abnormal features while others", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 312, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 506, + 325 + ], + "score": 1.0, + "content": "are more vulnerable. To reduce such variance, we report the average performance over 10 times of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 322, + 507, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 322, + 507, + 335 + ], + "score": 1.0, + "content": "random selection of the noise node sets, similar to the settings in the preliminary study in Section 2.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "We report the node classification test accuracy on abnormal (noisy) features and normal features in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 345, + 494, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 494, + 358 + ], + "score": 1.0, + "content": "Figure 5 and Figure 6, separately, under varying noisy ratio. From these figures, we can observe:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 268, + 507, + 358 + ] + }, + { + "type": "list", + "bbox": [ + 116, + 366, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 120, + 366, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 120, + 366, + 506, + 377 + ], + "score": 1.0, + "content": "• Figure 5 shows that AirGNN significantly outperforms all baselines on all datasets in terms of", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 127, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "the performance on noisy nodes. This verifies that AMP is able to improve the resilience to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 388, + 354, + 400 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 354, + 400 + ], + "score": 1.0, + "content": "noisy features, aligning well with the design motivation.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 121, + 402, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 121, + 402, + 506, + 416 + ], + "score": 1.0, + "content": "• Figure 6 shows that AirGNN promotes the performance on normal nodes when abnormal nodes", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 415, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 128, + 415, + 505, + 426 + ], + "score": 1.0, + "content": "exist. This is because AMP can remove some abnormal features which are detrimental to normal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 425, + 157, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 425, + 157, + 437 + ], + "score": 1.0, + "content": "nodes.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + } + ], + "index": 25.5, + "bbox_fs": [ + 120, + 366, + 506, + 437 + ] + }, + { + "type": "image", + "bbox": [ + 101, + 447, + 497, + 561 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 101, + 447, + 497, + 561 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 101, + 447, + 497, + 561 + ], + "spans": [ + { + "bbox": [ + 101, + 447, + 497, + 561 + ], + "score": 0.973, + "type": "image", + "image_path": "cb981f452a6f87e6cccc28935d8e1aa89643dc8d85a1008f917f9b4fedf8a508.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 101, + 447, + 497, + 485.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 101, + 485.0, + 497, + 523.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 101, + 523.0, + 497, + 561.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 172, + 568, + 438, + 580 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 172, + 567, + 438, + 581 + ], + "spans": [ + { + "bbox": [ + 172, + 567, + 438, + 581 + ], + "score": 1.0, + "content": "Figure 5: Node classification accuracy on abnormal (noisy) nodes", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + } + ], + "index": 31.0 + }, + { + "type": "title", + "bbox": [ + 106, + 604, + 354, + 615 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 355, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 355, + 617 + ], + "score": 1.0, + "content": "4.3 Performance Comparison with Adversarial Features", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 504, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 504, + 636 + ], + "score": 1.0, + "content": "In this subsection, we consider the abnormal feature scenario when the node features are maliciously", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 636, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 505, + 647 + ], + "score": 1.0, + "content": "attacked by the attacker to manipulate the prediction of GNNs. We use the Nettack [26] implemented", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 645, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 659 + ], + "score": 1.0, + "content": "in DeepRobust3 [27], a PyTorch library for adversarial attacks and defenses, to generate the adversarial", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 658, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 669 + ], + "score": 1.0, + "content": "features. We randomly choose 40 test nodes as the targeted nodes, and assess the performance under", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 668, + 504, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 239, + 680 + ], + "score": 1.0, + "content": "increasing perturbation budgets", + "type": "text" + }, + { + "bbox": [ + 239, + 668, + 322, + 680 + ], + "score": 0.91, + "content": "\\{ 0 , 5 , 1 0 , 2 0 , 5 0 , 8 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 668, + 504, + 680 + ], + "score": 1.0, + "content": ", where the perturbation numbers denote the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "number of feature dimensions that can be manipulated. The node classification accuracy on these", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 690, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 506, + 703 + ], + "score": 1.0, + "content": "attacked nodes are showed in Figure 7. From these figures, we can make the following observations:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 625, + 506, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 67, + 512, + 194 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 100, + 67, + 512, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 100, + 67, + 512, + 194 + ], + "spans": [ + { + "bbox": [ + 100, + 67, + 512, + 194 + ], + "score": 0.963, + "type": "image", + "image_path": "521af9135379cb9d9a4e04bfd77d43979bb91c8146837931dfc5f87ae829d691.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 100, + 67, + 512, + 109.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 100, + 109.33333333333334, + 512, + 151.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 100, + 151.66666666666669, + 512, + 194.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 193, + 201, + 417, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 193, + 200, + 418, + 214 + ], + "spans": [ + { + "bbox": [ + 193, + 200, + 418, + 214 + ], + "score": 1.0, + "content": "Figure 6: Node classification accuracy on normal nodes", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 119, + 227, + 506, + 259 + ], + "lines": [ + { + "bbox": [ + 119, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 119, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "• AirGNN is significantly more robust against adversarially attacked features than all baselines.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 127, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "MLP is the most vulnerable model, which demonstrates the usefulness of graph structure", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 249, + 362, + 261 + ], + "spans": [ + { + "bbox": [ + 127, + 249, + 362, + 261 + ], + "score": 1.0, + "content": "information in combating against abnormal node features.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 118, + 263, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 118, + 262, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 118, + 262, + 505, + 276 + ], + "score": 1.0, + "content": "• The advantages of AirGNN over the baselines become much stronger with larger perturbation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 127, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "budgets. This suggests that AMP can significantly improve the resilience to abnormal features.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "image", + "bbox": [ + 101, + 312, + 500, + 429 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 101, + 312, + 500, + 429 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 101, + 312, + 500, + 429 + ], + "spans": [ + { + "bbox": [ + 101, + 312, + 500, + 429 + ], + "score": 0.972, + "type": "image", + "image_path": "b4a84b017c1ab3aee2d1fd5a97a59af3692c895e3672220068ba8be20564b118.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 101, + 312, + 500, + 351.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 101, + 351.0, + 500, + 390.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 101, + 390.0, + 500, + 429.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 185, + 435, + 425, + 447 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 185, + 434, + 426, + 448 + ], + "spans": [ + { + "bbox": [ + 185, + 434, + 426, + 448 + ], + "score": 1.0, + "content": "Figure 7: Node classification accuracy on adversarial nodes", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + } + ], + "index": 11.0 + }, + { + "type": "title", + "bbox": [ + 106, + 466, + 344, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 345, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 345, + 480 + ], + "score": 1.0, + "content": "4.4 Adaptive Residual for Abnormal & Normal Nodes", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 504, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 493, + 499 + ], + "score": 1.0, + "content": "To further understand and verify how AMP and AirGNN work, we investigate the adaptive score", + "type": "text" + }, + { + "bbox": [ + 494, + 487, + 504, + 498 + ], + "score": 0.87, + "content": "\\beta _ { i }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 163, + 510 + ], + "score": 1.0, + "content": "for each node", + "type": "text" + }, + { + "bbox": [ + 163, + 499, + 173, + 509 + ], + "score": 0.84, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 497, + 506, + 510 + ], + "score": 1.0, + "content": ". Specifically, the average adaptive scores for abnormal nodes and normal nodes in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 508, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 522 + ], + "score": 1.0, + "content": "the last layer of AMP are computed separately. In the noisy feature scenario, we fix ratio of noisy", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 142, + 532 + ], + "score": 1.0, + "content": "nodes as", + "type": "text" + }, + { + "bbox": [ + 143, + 520, + 162, + 530 + ], + "score": 0.88, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 519, + 506, + 532 + ], + "score": 1.0, + "content": ". In the adversarial feature scenario, we choose 40 target nodes and fix the perturbation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 530, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 542 + ], + "score": 1.0, + "content": "number as 80. The results in noisy and adversarial feature scenarios are showed in Table 1 and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 542, + 335, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 335, + 554 + ], + "score": 1.0, + "content": "Table 2, respectively. From these tables, we can observe:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 119, + 561, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 119, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 119, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "• On the one hand, it can be clearly observed that in both scenarios, the average adaptive scores", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 573, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 127, + 573, + 505, + 585 + ], + "score": 1.0, + "content": "for abnormal nodes are significantly higher than those for normal nodes. Therefore, it verifies", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 585, + 449, + 596 + ], + "spans": [ + { + "bbox": [ + 127, + 585, + 449, + 596 + ], + "score": 1.0, + "content": "our intuition that large adaptive scores are strongly related to abnormal features.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 119, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 119, + 598, + 379, + 611 + ], + "score": 1.0, + "content": "• On the other hand, it also implies that the residual weights (i.e.,", + "type": "text" + }, + { + "bbox": [ + 380, + 599, + 410, + 610 + ], + "score": 0.87, + "content": "1 - \\beta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "for abnormal nodes are", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 128, + 610, + 504, + 621 + ], + "spans": [ + { + "bbox": [ + 128, + 610, + 504, + 621 + ], + "score": 1.0, + "content": "much lower than those of normal nodes. This perfectly aligns with our motivation to remove", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 128, + 621, + 359, + 633 + ], + "spans": [ + { + "bbox": [ + 128, + 621, + 359, + 633 + ], + "score": 1.0, + "content": "abnormal features by reducing their residual connections.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 718 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "The study on adaptive scores verifies how the adaptive residuals in AMP and AirGNN work as", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 651, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 506, + 664 + ], + "score": 1.0, + "content": "designed. It corroborates that AirGNN not only tremendously boosts the resilience to abnormal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 661, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 676 + ], + "score": 1.0, + "content": "features but also provides interpretable information for anomaly detection that will be useful in many", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 674, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 506, + 685 + ], + "score": 1.0, + "content": "security-critical scenarios since the adaptive score serves as a good indicator of abnormal nodes.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 684, + 506, + 697 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 506, + 697 + ], + "score": 1.0, + "content": "Morever, it is expected that APPNP without residual will perform well on abnormal nodes but it will", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 695, + 506, + 708 + ], + "spans": [ + { + "bbox": [ + 106, + 695, + 506, + 708 + ], + "score": 1.0, + "content": "sacrifice the performance on normal nodes. We provide detailed comparison with APPNP w/Res and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 706, + 467, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 467, + 718 + ], + "score": 1.0, + "content": "APPNP wo/Res in Appendix D.3 to show the advantages of adaptive residual of AirGNN.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 67, + 512, + 194 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 100, + 67, + 512, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 100, + 67, + 512, + 194 + ], + "spans": [ + { + "bbox": [ + 100, + 67, + 512, + 194 + ], + "score": 0.963, + "type": "image", + "image_path": "521af9135379cb9d9a4e04bfd77d43979bb91c8146837931dfc5f87ae829d691.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 100, + 67, + 512, + 109.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 100, + 109.33333333333334, + 512, + 151.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 100, + 151.66666666666669, + 512, + 194.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 193, + 201, + 417, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 193, + 200, + 418, + 214 + ], + "spans": [ + { + "bbox": [ + 193, + 200, + 418, + 214 + ], + "score": 1.0, + "content": "Figure 6: Node classification accuracy on normal nodes", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 119, + 227, + 506, + 259 + ], + "lines": [ + { + "bbox": [ + 119, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 119, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "• AirGNN is significantly more robust against adversarially attacked features than all baselines.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 237, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 127, + 237, + 505, + 250 + ], + "score": 1.0, + "content": "MLP is the most vulnerable model, which demonstrates the usefulness of graph structure", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 249, + 362, + 261 + ], + "spans": [ + { + "bbox": [ + 127, + 249, + 362, + 261 + ], + "score": 1.0, + "content": "information in combating against abnormal node features.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 119, + 226, + 506, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 118, + 263, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 118, + 262, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 118, + 262, + 505, + 276 + ], + "score": 1.0, + "content": "• The advantages of AirGNN over the baselines become much stronger with larger perturbation", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 127, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "budgets. This suggests that AMP can significantly improve the resilience to abnormal features.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 118, + 262, + 506, + 287 + ] + }, + { + "type": "image", + "bbox": [ + 101, + 312, + 500, + 429 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 101, + 312, + 500, + 429 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 101, + 312, + 500, + 429 + ], + "spans": [ + { + "bbox": [ + 101, + 312, + 500, + 429 + ], + "score": 0.972, + "type": "image", + "image_path": "b4a84b017c1ab3aee2d1fd5a97a59af3692c895e3672220068ba8be20564b118.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 101, + 312, + 500, + 351.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 101, + 351.0, + 500, + 390.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 101, + 390.0, + 500, + 429.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 185, + 435, + 425, + 447 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 185, + 434, + 426, + 448 + ], + "spans": [ + { + "bbox": [ + 185, + 434, + 426, + 448 + ], + "score": 1.0, + "content": "Figure 7: Node classification accuracy on adversarial nodes", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + } + ], + "index": 11.0 + }, + { + "type": "title", + "bbox": [ + 106, + 466, + 344, + 478 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 345, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 345, + 480 + ], + "score": 1.0, + "content": "4.4 Adaptive Residual for Abnormal & Normal Nodes", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 504, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 493, + 499 + ], + "score": 1.0, + "content": "To further understand and verify how AMP and AirGNN work, we investigate the adaptive score", + "type": "text" + }, + { + "bbox": [ + 494, + 487, + 504, + 498 + ], + "score": 0.87, + "content": "\\beta _ { i }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 163, + 510 + ], + "score": 1.0, + "content": "for each node", + "type": "text" + }, + { + "bbox": [ + 163, + 499, + 173, + 509 + ], + "score": 0.84, + "content": "v _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 497, + 506, + 510 + ], + "score": 1.0, + "content": ". Specifically, the average adaptive scores for abnormal nodes and normal nodes in", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 508, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 522 + ], + "score": 1.0, + "content": "the last layer of AMP are computed separately. In the noisy feature scenario, we fix ratio of noisy", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 142, + 532 + ], + "score": 1.0, + "content": "nodes as", + "type": "text" + }, + { + "bbox": [ + 143, + 520, + 162, + 530 + ], + "score": 0.88, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 519, + 506, + 532 + ], + "score": 1.0, + "content": ". In the adversarial feature scenario, we choose 40 target nodes and fix the perturbation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 530, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 542 + ], + "score": 1.0, + "content": "number as 80. The results in noisy and adversarial feature scenarios are showed in Table 1 and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 542, + 335, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 335, + 554 + ], + "score": 1.0, + "content": "Table 2, respectively. From these tables, we can observe:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 487, + 506, + 554 + ] + }, + { + "type": "text", + "bbox": [ + 119, + 561, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 119, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 119, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "• On the one hand, it can be clearly observed that in both scenarios, the average adaptive scores", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 573, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 127, + 573, + 505, + 585 + ], + "score": 1.0, + "content": "for abnormal nodes are significantly higher than those for normal nodes. Therefore, it verifies", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 585, + 449, + 596 + ], + "spans": [ + { + "bbox": [ + 127, + 585, + 449, + 596 + ], + "score": 1.0, + "content": "our intuition that large adaptive scores are strongly related to abnormal features.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 119, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 119, + 598, + 379, + 611 + ], + "score": 1.0, + "content": "• On the other hand, it also implies that the residual weights (i.e.,", + "type": "text" + }, + { + "bbox": [ + 380, + 599, + 410, + 610 + ], + "score": 0.87, + "content": "1 - \\beta _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "for abnormal nodes are", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 128, + 610, + 504, + 621 + ], + "spans": [ + { + "bbox": [ + 128, + 610, + 504, + 621 + ], + "score": 1.0, + "content": "much lower than those of normal nodes. This perfectly aligns with our motivation to remove", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 128, + 621, + 359, + 633 + ], + "spans": [ + { + "bbox": [ + 128, + 621, + 359, + 633 + ], + "score": 1.0, + "content": "abnormal features by reducing their residual connections.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 119, + 561, + 505, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 718 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 653 + ], + "score": 1.0, + "content": "The study on adaptive scores verifies how the adaptive residuals in AMP and AirGNN work as", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 651, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 506, + 664 + ], + "score": 1.0, + "content": "designed. It corroborates that AirGNN not only tremendously boosts the resilience to abnormal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 661, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 676 + ], + "score": 1.0, + "content": "features but also provides interpretable information for anomaly detection that will be useful in many", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 674, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 506, + 685 + ], + "score": 1.0, + "content": "security-critical scenarios since the adaptive score serves as a good indicator of abnormal nodes.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 684, + 506, + 697 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 506, + 697 + ], + "score": 1.0, + "content": "Morever, it is expected that APPNP without residual will perform well on abnormal nodes but it will", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 695, + 506, + 708 + ], + "spans": [ + { + "bbox": [ + 106, + 695, + 506, + 708 + ], + "score": 1.0, + "content": "sacrifice the performance on normal nodes. We provide detailed comparison with APPNP w/Res and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 706, + 467, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 467, + 718 + ], + "score": 1.0, + "content": "APPNP wo/Res in Appendix D.3 to show the advantages of adaptive residual of AirGNN.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 640, + 506, + 718 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 128, + 91, + 484, + 155 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 115, + 77, + 492, + 89 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 75, + 493, + 91 + ], + "spans": [ + { + "bbox": [ + 117, + 75, + 248, + 91 + ], + "score": 1.0, + "content": "Table 1: Average adaptive score", + "type": "text" + }, + { + "bbox": [ + 248, + 78, + 262, + 89 + ], + "score": 0.73, + "content": "( \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 75, + 343, + 91 + ], + "score": 1.0, + "content": "and residual weight", + "type": "text" + }, + { + "bbox": [ + 344, + 78, + 375, + 89 + ], + "score": 0.88, + "content": "( 1 - \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 75, + 493, + 91 + ], + "score": 1.0, + "content": "in the noisy feature scenario.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 128, + 91, + 484, + 155 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 91, + 484, + 155 + ], + "spans": [ + { + "bbox": [ + 128, + 91, + 484, + 155 + ], + "score": 0.978, + "html": "
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes0.998 ± 0.0000.988 ± 0.0000.996 ± 0.000
Average adaptive score for normal nodes Average residual weight for abnormal nodes0.924 ± 0.002 0.002 ±0.0000.807 ± 0.005 0.012 ± 0.0000.869 ± 0.006
Average residual weight for normal nodes0.076 ± 0.0020.193 ± 0.0050.004±0.000 0.131 ± 0.006
", + "type": "table", + "image_path": "d49a0a661a725cae61dfea576b2654826b85c9df4b3a4674a943e47200a74fd9.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 128, + 91, + 484, + 112.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 112.33333333333333, + 484, + 133.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 128, + 133.66666666666666, + 484, + 155.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 128, + 188, + 483, + 253 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 109, + 174, + 503, + 186 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 173, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 107, + 173, + 237, + 187 + ], + "score": 1.0, + "content": "Table 2: Average adaptive score", + "type": "text" + }, + { + "bbox": [ + 237, + 175, + 251, + 186 + ], + "score": 0.79, + "content": "( \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 173, + 333, + 187 + ], + "score": 1.0, + "content": "and residual weight", + "type": "text" + }, + { + "bbox": [ + 333, + 175, + 364, + 186 + ], + "score": 0.88, + "content": "( 1 - \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 173, + 504, + 187 + ], + "score": 1.0, + "content": "in the adversarial feature scenario.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 128, + 188, + 483, + 253 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 128, + 188, + 483, + 253 + ], + "spans": [ + { + "bbox": [ + 128, + 188, + 483, + 253 + ], + "score": 0.98, + "html": "
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes Average adaptive score for normal nodes0.987 ±0.0000.930 ± 0.0070.959 ± 0.005
Average residual weight for abnormal nodes0.922 ± 0.0040.689 ± 0.0240.826 ± 0.016
0.013 ±0.0000.070±0.0070.041 ±0.005
Average residual weight for normal nodes0.078 ± 0.0040.311 ± 0.0240.174 ± 0.016
", + "type": "table", + "image_path": "bd195752ecb387a8a737f39dd60085cba6c171a851c3d77460489987040b8bfc.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 128, + 188, + 483, + 209.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 128, + 209.66666666666666, + 483, + 231.33333333333331 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 128, + 231.33333333333331, + 483, + 252.99999999999997 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "title", + "bbox": [ + 107, + 273, + 197, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 272, + 198, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 198, + 288 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 299, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "GNNs generalize convolutional neural networks (CNN) to graph structure data through the message", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "score": 1.0, + "content": "passing framework [1, 2, 7]. The design of message passing and GNN architectures are majorly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "motivated in spectral domain [3, 4] and spatial domain [5, 6, 7, 2]. Recent works have shown that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "score": 1.0, + "content": "the message passing in GNNs can be regarded as low-pass graph filters [12, 13]. More generally,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "it has been proven that message passing in many GNNs can be uniformly derived from graph", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "signal denoising [11, 28, 29, 30]. Classic GNNs such as GCN [3] and GAT [6] achieve their best", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "performance with shallow models, but their performance degrades when stacking more layers,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 375, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 506, + 390 + ], + "score": 1.0, + "content": "which can be partially explained through oversmoothing analyses [14, 15]. Recent works propose", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 386, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 104, + 386, + 506, + 401 + ], + "score": 1.0, + "content": "to use residual connections or skip connections to mitigate the oversmoothing issues, and they", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "demonstrate the potential benefits from more feature aggregations. Examples include but not limited", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "score": 1.0, + "content": "to DeepGCNs [31], JKNet [32], GCNII [9], APPNP [10] and DeeperGNN [20]. These models use", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "global residual connection that can not be adaptive for each node, which significantly differ from the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 185, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 185, + 442 + ], + "score": 1.0, + "content": "proposed AirGNN.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 447, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 507, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 507, + 460 + ], + "score": 1.0, + "content": "Recently, there are growing interests in reducing GNNs’ vulnerability to the graph structure noise,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "such as Robust GCN [24], GCN-SVD [33], Pro-GNN [34], IDGL [35], ElasticGNN [36], etc. Please", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 481 + ], + "score": 1.0, + "content": "refer to the comprehensive surveys [37, 38] for more details. However, how to design GNNs with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "strong resilience to abnormal node features remains to be developed. To the best of our knowledge,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "AirGNN is the first GNN model that is intrinsically robust to many types of abnormal node features", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "by design. It improves the performance in various kinds of abnormal scenarios without needing to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 513, + 280, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 280, + 525 + ], + "score": 1.0, + "content": "sacrifice clean accuracy in normal settings.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 107, + 541, + 183, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 185, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 185, + 558 + ], + "score": 1.0, + "content": "6 Conclusion", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 505, + 656 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "In this work, we discover an intrinsic tension between feature aggregation and residual connection in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "the message passing scheme of GNNs, as well as the corresponding performance tradeoff between", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "nodes with abnormal and normal features. We analyze possible reasons to explain these findings", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "from the perspective of graph Laplacian smoothing. Our understandings further motivate us to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "propose a simple, efficient, interpretable and adaptive message passing scheme as well as a new GNN", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 622, + 504, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 504, + 635 + ], + "score": 1.0, + "content": "model with adaptive residual, named AirGNN. AirGNN provides a node-wise adaptive transition", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "between feature aggregation and residual connection, and the significant advantages of AirGNN are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 644, + 289, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 289, + 657 + ], + "score": 1.0, + "content": "demonstrated through extensive experiments.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 107, + 673, + 339, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 671, + 341, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 341, + 691 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "This research is supported by the National Science Foundation (NSF) under grant numbers", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "IIS1714741, CNS1815636, IIS1845081, IIS1907704, DRL2025244, IIS1928278, IIS1955285,", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 755 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 128, + 91, + 484, + 155 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 115, + 77, + 492, + 89 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 75, + 493, + 91 + ], + "spans": [ + { + "bbox": [ + 117, + 75, + 248, + 91 + ], + "score": 1.0, + "content": "Table 1: Average adaptive score", + "type": "text" + }, + { + "bbox": [ + 248, + 78, + 262, + 89 + ], + "score": 0.73, + "content": "( \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 75, + 343, + 91 + ], + "score": 1.0, + "content": "and residual weight", + "type": "text" + }, + { + "bbox": [ + 344, + 78, + 375, + 89 + ], + "score": 0.88, + "content": "( 1 - \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 75, + 493, + 91 + ], + "score": 1.0, + "content": "in the noisy feature scenario.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 128, + 91, + 484, + 155 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 91, + 484, + 155 + ], + "spans": [ + { + "bbox": [ + 128, + 91, + 484, + 155 + ], + "score": 0.978, + "html": "
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes0.998 ± 0.0000.988 ± 0.0000.996 ± 0.000
Average adaptive score for normal nodes Average residual weight for abnormal nodes0.924 ± 0.002 0.002 ±0.0000.807 ± 0.005 0.012 ± 0.0000.869 ± 0.006
Average residual weight for normal nodes0.076 ± 0.0020.193 ± 0.0050.004±0.000 0.131 ± 0.006
", + "type": "table", + "image_path": "d49a0a661a725cae61dfea576b2654826b85c9df4b3a4674a943e47200a74fd9.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 128, + 91, + 484, + 112.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 128, + 112.33333333333333, + 484, + 133.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 128, + 133.66666666666666, + 484, + 155.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "table", + "bbox": [ + 128, + 188, + 483, + 253 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 109, + 174, + 503, + 186 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 173, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 107, + 173, + 237, + 187 + ], + "score": 1.0, + "content": "Table 2: Average adaptive score", + "type": "text" + }, + { + "bbox": [ + 237, + 175, + 251, + 186 + ], + "score": 0.79, + "content": "( \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 173, + 333, + 187 + ], + "score": 1.0, + "content": "and residual weight", + "type": "text" + }, + { + "bbox": [ + 333, + 175, + 364, + 186 + ], + "score": 0.88, + "content": "( 1 - \\beta )", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 173, + 504, + 187 + ], + "score": 1.0, + "content": "in the adversarial feature scenario.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table_body", + "bbox": [ + 128, + 188, + 483, + 253 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 128, + 188, + 483, + 253 + ], + "spans": [ + { + "bbox": [ + 128, + 188, + 483, + 253 + ], + "score": 0.98, + "html": "
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes Average adaptive score for normal nodes0.987 ±0.0000.930 ± 0.0070.959 ± 0.005
Average residual weight for abnormal nodes0.922 ± 0.0040.689 ± 0.0240.826 ± 0.016
0.013 ±0.0000.070±0.0070.041 ±0.005
Average residual weight for normal nodes0.078 ± 0.0040.311 ± 0.0240.174 ± 0.016
", + "type": "table", + "image_path": "bd195752ecb387a8a737f39dd60085cba6c171a851c3d77460489987040b8bfc.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 128, + 188, + 483, + 209.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 128, + 209.66666666666666, + 483, + 231.33333333333331 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 128, + 231.33333333333331, + 483, + 252.99999999999997 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 5.0 + }, + { + "type": "title", + "bbox": [ + 107, + 273, + 197, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 272, + 198, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 198, + 288 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 299, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "GNNs generalize convolutional neural networks (CNN) to graph structure data through the message", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "score": 1.0, + "content": "passing framework [1, 2, 7]. The design of message passing and GNN architectures are majorly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "motivated in spectral domain [3, 4] and spatial domain [5, 6, 7, 2]. Recent works have shown that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 506, + 345 + ], + "score": 1.0, + "content": "the message passing in GNNs can be regarded as low-pass graph filters [12, 13]. More generally,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 343, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 506, + 358 + ], + "score": 1.0, + "content": "it has been proven that message passing in many GNNs can be uniformly derived from graph", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "signal denoising [11, 28, 29, 30]. Classic GNNs such as GCN [3] and GAT [6] achieve their best", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "performance with shallow models, but their performance degrades when stacking more layers,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 375, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 506, + 390 + ], + "score": 1.0, + "content": "which can be partially explained through oversmoothing analyses [14, 15]. Recent works propose", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 386, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 104, + 386, + 506, + 401 + ], + "score": 1.0, + "content": "to use residual connections or skip connections to mitigate the oversmoothing issues, and they", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 410 + ], + "score": 1.0, + "content": "demonstrate the potential benefits from more feature aggregations. Examples include but not limited", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "score": 1.0, + "content": "to DeepGCNs [31], JKNet [32], GCNII [9], APPNP [10] and DeeperGNN [20]. These models use", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "global residual connection that can not be adaptive for each node, which significantly differ from the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 432, + 185, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 185, + 442 + ], + "score": 1.0, + "content": "proposed AirGNN.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 299, + 506, + 442 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 447, + 505, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 507, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 507, + 460 + ], + "score": 1.0, + "content": "Recently, there are growing interests in reducing GNNs’ vulnerability to the graph structure noise,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "such as Robust GCN [24], GCN-SVD [33], Pro-GNN [34], IDGL [35], ElasticGNN [36], etc. Please", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 481 + ], + "score": 1.0, + "content": "refer to the comprehensive surveys [37, 38] for more details. However, how to design GNNs with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "strong resilience to abnormal node features remains to be developed. To the best of our knowledge,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "AirGNN is the first GNN model that is intrinsically robust to many types of abnormal node features", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "by design. It improves the performance in various kinds of abnormal scenarios without needing to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 513, + 280, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 280, + 525 + ], + "score": 1.0, + "content": "sacrifice clean accuracy in normal settings.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 446, + 507, + 525 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 541, + 183, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 185, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 185, + 558 + ], + "score": 1.0, + "content": "6 Conclusion", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 505, + 656 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "In this work, we discover an intrinsic tension between feature aggregation and residual connection in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "the message passing scheme of GNNs, as well as the corresponding performance tradeoff between", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "nodes with abnormal and normal features. We analyze possible reasons to explain these findings", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "from the perspective of graph Laplacian smoothing. Our understandings further motivate us to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "propose a simple, efficient, interpretable and adaptive message passing scheme as well as a new GNN", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 622, + 504, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 504, + 635 + ], + "score": 1.0, + "content": "model with adaptive residual, named AirGNN. AirGNN provides a node-wise adaptive transition", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "between feature aggregation and residual connection, and the significant advantages of AirGNN are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 644, + 289, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 289, + 657 + ], + "score": 1.0, + "content": "demonstrated through extensive experiments.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 567, + 506, + 657 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 673, + 339, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 671, + 341, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 341, + 691 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "This research is supported by the National Science Foundation (NSF) under grant numbers", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "IIS1714741, CNS1815636, IIS1845081, IIS1907704, DRL2025244, IIS1928278, IIS1955285,", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 700, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "IOS2107215, IOS2035472 and Army Research Office (ARO) under grant number W911NF-21-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 140, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 140, + 95 + ], + "score": 1.0, + "content": "1-0198.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 108, + 110, + 271, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 273, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 273, + 126 + ], + "score": 1.0, + "content": "Societal Impact and Limitations", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 136, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "score": 1.0, + "content": "The methodology proposed in this paper might have significant positive societal impact since it", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 146, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 505, + 158 + ], + "score": 1.0, + "content": "reduces machine learning models’ vulnerability to abnormal datasets that generally exist in real-world", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 157, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 506, + 170 + ], + "score": 1.0, + "content": "applications, especially in many security-critical scenarios. In practice, for a given graph, we do", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "not have the prior knowledge about if the graph is clean, has noisy features or adversarial features.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 178, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 506, + 192 + ], + "score": 1.0, + "content": "Therefore, algorithms like the proposed AirGNN that can work under both the clean and various", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 189, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 506, + 204 + ], + "score": 1.0, + "content": "abnormal feature settings are appealing. While we are unaware of any potential negative society", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "impact, we point out two limitations of this work: (1) this paper focuses on abnormal node features", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "and has not evaluated the performance of the proposed method when the dataset contains both", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "abnormal node features and edges; (2) it is unclear how it performs on heterophilic graphs. It will be", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 235, + 334, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 334, + 246 + ], + "score": 1.0, + "content": "interesting to investigate these problems in future works.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 107, + 261, + 163, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 260, + 165, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 165, + 276 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 506, + 726 + ], + "lines": [ + { + "bbox": [ + 110, + 281, + 487, + 294 + ], + "spans": [ + { + "bbox": [ + 110, + 281, + 487, + 294 + ], + "score": 1.0, + "content": "[1] Yao Ma and Jiliang Tang. Deep Learning on Graphs. Cambridge University Press, 2020.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 110, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 110, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "[2] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 126, + 311, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 126, + 311, + 507, + 326 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. In International Conference on Machine Learning,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 322, + 261, + 335 + ], + "spans": [ + { + "bbox": [ + 126, + 322, + 261, + 335 + ], + "score": 1.0, + "content": "pages 1263–1272. PMLR, 2017.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 110, + 341, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 110, + 341, + 505, + 355 + ], + "score": 1.0, + "content": "[3] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 352, + 333, + 366 + ], + "spans": [ + { + "bbox": [ + 127, + 352, + 333, + 366 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1609.02907, 2016.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 110, + 371, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 110, + 371, + 506, + 385 + ], + "score": 1.0, + "content": "[4] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 383, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 126, + 383, + 506, + 396 + ], + "score": 1.0, + "content": "on graphs with fast localized spectral filtering. In Proceedings of the 30th International", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 393, + 452, + 408 + ], + "spans": [ + { + "bbox": [ + 127, + 393, + 452, + 408 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems, pages 3844–3852, 2016.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 110, + 412, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 110, + 412, + 505, + 428 + ], + "score": 1.0, + "content": "[5] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 425, + 324, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 425, + 324, + 437 + ], + "score": 1.0, + "content": "graphs. arXiv preprint arXiv:1706.02216, 2017.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 110, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 110, + 443, + 506, + 457 + ], + "score": 1.0, + "content": "[6] Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua ´", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 454, + 433, + 468 + ], + "spans": [ + { + "bbox": [ + 126, + 454, + 433, + 468 + ], + "score": 1.0, + "content": "Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 110, + 474, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 110, + 474, + 507, + 488 + ], + "score": 1.0, + "content": "[7] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 484, + 502, + 499 + ], + "spans": [ + { + "bbox": [ + 127, + 484, + 502, + 499 + ], + "score": 1.0, + "content": "The graph neural network model. IEEE transactions on neural networks, 20(1):61–80, 2008.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 110, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 110, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "[8] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 127, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "comprehensive survey on graph neural networks. IEEE transactions on neural networks and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 526, + 225, + 540 + ], + "spans": [ + { + "bbox": [ + 127, + 526, + 225, + 540 + ], + "score": 1.0, + "content": "learning systems, 2020.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 109, + 544, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 109, + 544, + 506, + 561 + ], + "score": 1.0, + "content": "[9] Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 556, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 127, + 556, + 505, + 570 + ], + "score": 1.0, + "content": "graph convolutional networks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 567, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 126, + 567, + 505, + 582 + ], + "score": 1.0, + "content": "37th International Conference on Machine Learning, volume 119 of Proceedings of Machine", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 578, + 382, + 592 + ], + "spans": [ + { + "bbox": [ + 126, + 578, + 382, + 592 + ], + "score": 1.0, + "content": "Learning Research, pages 1725–1735. PMLR, 13–18 Jul 2020.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "[10] Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate:", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 128, + 609, + 502, + 622 + ], + "spans": [ + { + "bbox": [ + 128, + 609, + 502, + 622 + ], + "score": 1.0, + "content": "Graph neural networks meet personalized pagerank. arXiv preprint arXiv:1810.05997, 2018.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "[11] Yao Ma, Xiaorui Liu, Tong Zhao, Yozen Liu, Jiliang Tang, and Neil Shah. A unified view on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 639, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 127, + 639, + 506, + 654 + ], + "score": 1.0, + "content": "graph neural networks as graph signal denoising. Proceedings of the 30th ACM International", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 649, + 385, + 664 + ], + "spans": [ + { + "bbox": [ + 127, + 649, + 385, + 664 + ], + "score": 1.0, + "content": "Conference on Information and Knowledge Management, 2021.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "[12] Hoang Nt and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 680, + 320, + 694 + ], + "spans": [ + { + "bbox": [ + 126, + 680, + 320, + 694 + ], + "score": 1.0, + "content": "filters. arXiv preprint arXiv:1905.09550, 2019.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 700, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 714 + ], + "score": 1.0, + "content": "[13] Lingxiao Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. In International", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 710, + 321, + 725 + ], + "spans": [ + { + "bbox": [ + 127, + 710, + 321, + 725 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 29 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "IOS2107215, IOS2035472 and Army Research Office (ARO) under grant number W911NF-21-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 140, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 140, + 95 + ], + "score": 1.0, + "content": "1-0198.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 72, + 506, + 95 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 110, + 271, + 124 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 273, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 273, + 126 + ], + "score": 1.0, + "content": "Societal Impact and Limitations", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 136, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "score": 1.0, + "content": "The methodology proposed in this paper might have significant positive societal impact since it", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 146, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 505, + 158 + ], + "score": 1.0, + "content": "reduces machine learning models’ vulnerability to abnormal datasets that generally exist in real-world", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 157, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 506, + 170 + ], + "score": 1.0, + "content": "applications, especially in many security-critical scenarios. In practice, for a given graph, we do", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "not have the prior knowledge about if the graph is clean, has noisy features or adversarial features.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 178, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 506, + 192 + ], + "score": 1.0, + "content": "Therefore, algorithms like the proposed AirGNN that can work under both the clean and various", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 189, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 506, + 204 + ], + "score": 1.0, + "content": "abnormal feature settings are appealing. While we are unaware of any potential negative society", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "impact, we point out two limitations of this work: (1) this paper focuses on abnormal node features", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "and has not evaluated the performance of the proposed method when the dataset contains both", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "abnormal node features and edges; (2) it is unclear how it performs on heterophilic graphs. It will be", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 235, + 334, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 334, + 246 + ], + "score": 1.0, + "content": "interesting to investigate these problems in future works.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 135, + 506, + 246 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 261, + 163, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 260, + 165, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 165, + 276 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "list", + "bbox": [ + 106, + 277, + 506, + 726 + ], + "lines": [ + { + "bbox": [ + 110, + 281, + 487, + 294 + ], + "spans": [ + { + "bbox": [ + 110, + 281, + 487, + 294 + ], + "score": 1.0, + "content": "[1] Yao Ma and Jiliang Tang. Deep Learning on Graphs. Cambridge University Press, 2020.", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 110, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "[2] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 311, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 126, + 311, + 507, + 326 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. In International Conference on Machine Learning,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 322, + 261, + 335 + ], + "spans": [ + { + "bbox": [ + 126, + 322, + 261, + 335 + ], + "score": 1.0, + "content": "pages 1263–1272. PMLR, 2017.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 341, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 110, + 341, + 505, + 355 + ], + "score": 1.0, + "content": "[3] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 352, + 333, + 366 + ], + "spans": [ + { + "bbox": [ + 127, + 352, + 333, + 366 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1609.02907, 2016.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 371, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 110, + 371, + 506, + 385 + ], + "score": 1.0, + "content": "[4] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 383, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 126, + 383, + 506, + 396 + ], + "score": 1.0, + "content": "on graphs with fast localized spectral filtering. In Proceedings of the 30th International", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 393, + 452, + 408 + ], + "spans": [ + { + "bbox": [ + 127, + 393, + 452, + 408 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems, pages 3844–3852, 2016.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 412, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 110, + 412, + 505, + 428 + ], + "score": 1.0, + "content": "[5] William L Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 425, + 324, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 425, + 324, + 437 + ], + "score": 1.0, + "content": "graphs. arXiv preprint arXiv:1706.02216, 2017.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 443, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 110, + 443, + 506, + 457 + ], + "score": 1.0, + "content": "[6] Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua ´", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 454, + 433, + 468 + ], + "spans": [ + { + "bbox": [ + 126, + 454, + 433, + 468 + ], + "score": 1.0, + "content": "Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 474, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 110, + 474, + 507, + 488 + ], + "score": 1.0, + "content": "[7] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini.", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 484, + 502, + 499 + ], + "spans": [ + { + "bbox": [ + 127, + 484, + 502, + 499 + ], + "score": 1.0, + "content": "The graph neural network model. IEEE transactions on neural networks, 20(1):61–80, 2008.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 110, + 504, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 110, + 504, + 506, + 518 + ], + "score": 1.0, + "content": "[8] Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 516, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 127, + 516, + 506, + 528 + ], + "score": 1.0, + "content": "comprehensive survey on graph neural networks. IEEE transactions on neural networks and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 526, + 225, + 540 + ], + "spans": [ + { + "bbox": [ + 127, + 526, + 225, + 540 + ], + "score": 1.0, + "content": "learning systems, 2020.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 544, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 109, + 544, + 506, + 561 + ], + "score": 1.0, + "content": "[9] Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 556, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 127, + 556, + 505, + 570 + ], + "score": 1.0, + "content": "graph convolutional networks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 567, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 126, + 567, + 505, + 582 + ], + "score": 1.0, + "content": "37th International Conference on Machine Learning, volume 119 of Proceedings of Machine", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 578, + 382, + 592 + ], + "spans": [ + { + "bbox": [ + 126, + 578, + 382, + 592 + ], + "score": 1.0, + "content": "Learning Research, pages 1725–1735. PMLR, 13–18 Jul 2020.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 612 + ], + "score": 1.0, + "content": "[10] Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate:", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 609, + 502, + 622 + ], + "spans": [ + { + "bbox": [ + 128, + 609, + 502, + 622 + ], + "score": 1.0, + "content": "Graph neural networks meet personalized pagerank. arXiv preprint arXiv:1810.05997, 2018.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "[11] Yao Ma, Xiaorui Liu, Tong Zhao, Yozen Liu, Jiliang Tang, and Neil Shah. A unified view on", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 639, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 127, + 639, + 506, + 654 + ], + "score": 1.0, + "content": "graph neural networks as graph signal denoising. Proceedings of the 30th ACM International", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 649, + 385, + 664 + ], + "spans": [ + { + "bbox": [ + 127, + 649, + 385, + 664 + ], + "score": 1.0, + "content": "Conference on Information and Knowledge Management, 2021.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "[12] Hoang Nt and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 680, + 320, + 694 + ], + "spans": [ + { + "bbox": [ + 126, + 680, + 320, + 694 + ], + "score": 1.0, + "content": "filters. arXiv preprint arXiv:1905.09550, 2019.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 700, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 714 + ], + "score": 1.0, + "content": "[13] Lingxiao Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. In International", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 710, + 321, + 725 + ], + "spans": [ + { + "bbox": [ + 127, + 710, + 321, + 725 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2019.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "score": 1.0, + "content": "[14] Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 82, + 507, + 99 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 507, + 99 + ], + "score": 1.0, + "content": "for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 128, + 94, + 199, + 106 + ], + "spans": [ + { + "bbox": [ + 128, + 94, + 199, + 106 + ], + "score": 1.0, + "content": "volume 32, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "[15] Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 124, + 469, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 124, + 469, + 138 + ], + "score": 1.0, + "content": "node classification. In International Conference on Learning Representations, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 143, + 507, + 160 + ], + "spans": [ + { + "bbox": [ + 103, + 143, + 507, + 160 + ], + "score": 1.0, + "content": "[16] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press,", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 156, + 154, + 168 + ], + "spans": [ + { + "bbox": [ + 127, + 156, + 154, + 168 + ], + "score": 1.0, + "content": "2004.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "score": 1.0, + "content": "[17] Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 186, + 399, + 200 + ], + "spans": [ + { + "bbox": [ + 126, + 186, + 399, + 200 + ], + "score": 1.0, + "content": "social networks. Annual review of sociology, 27(1):415–444, 2001.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 206, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 506, + 220 + ], + "score": 1.0, + "content": "[18] Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 218, + 181, + 230 + ], + "spans": [ + { + "bbox": [ + 126, + 218, + 181, + 230 + ], + "score": 1.0, + "content": "propagation.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "score": 1.0, + "content": "[19] Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Schölkopf.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 247, + 331, + 261 + ], + "spans": [ + { + "bbox": [ + 126, + 247, + 331, + 261 + ], + "score": 1.0, + "content": "Learning with local and global consistency. 2004.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 266, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 104, + 266, + 507, + 282 + ], + "score": 1.0, + "content": "[20] Meng Liu, Hongyang Gao, and Shuiwang Ji. Towards deeper graph neural networks. In", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 278, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 127, + 278, + 506, + 292 + ], + "score": 1.0, + "content": "Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery &", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 289, + 238, + 302 + ], + "spans": [ + { + "bbox": [ + 125, + 289, + 238, + 302 + ], + "score": 1.0, + "content": "Data Mining. ACM, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 308, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 507, + 322 + ], + "score": 1.0, + "content": "[21] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 321, + 449, + 333 + ], + "spans": [ + { + "bbox": [ + 128, + 321, + 449, + 333 + ], + "score": 1.0, + "content": "Rad. Collective classification in network data. AI magazine, 29(3):93–93, 2008.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 338, + 507, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 507, + 353 + ], + "score": 1.0, + "content": "[22] Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 350, + 466, + 364 + ], + "spans": [ + { + "bbox": [ + 127, + 350, + 466, + 364 + ], + "score": 1.0, + "content": "Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 369, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 383 + ], + "score": 1.0, + "content": "[23] Marinka Zitnik Yuxiao Dong Hongyu Ren Bowen Liu Michele Catasta Jure Leskovec Wei-", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 126, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "hua Hu, Matthias Fey. Open graph benchmark: Datasets for machine learning on graphs. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 392, + 267, + 405 + ], + "spans": [ + { + "bbox": [ + 125, + 392, + 267, + 405 + ], + "score": 1.0, + "content": "preprint arXiv:2005.00687, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 506, + 425 + ], + "score": 1.0, + "content": "[24] Dingyuan Zhu, Ziwei Zhang, Peng Cui, and Wenwu Zhu. Robust graph convolutional networks", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 421, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 126, + 421, + 507, + 437 + ], + "score": 1.0, + "content": "against adversarial attacks. In Proceedings of the 25th ACM SIGKDD International Conference", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 432, + 399, + 448 + ], + "spans": [ + { + "bbox": [ + 126, + 432, + 399, + 448 + ], + "score": 1.0, + "content": "on Knowledge Discovery & Data Mining, pages 1399–1407, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 452, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 467 + ], + "score": 1.0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 463, + 226, + 477 + ], + "spans": [ + { + "bbox": [ + 127, + 463, + 226, + 477 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "[26] Daniel Zügner, Amir Akbarnejad, and Stephan Günnemann. Adversarial attacks on neural", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 494, + 318, + 508 + ], + "spans": [ + { + "bbox": [ + 126, + 494, + 318, + 508 + ], + "score": 1.0, + "content": "networks for graph data. In KDD. ACM, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "[27] Yaxin Li, Wei Jin, Han Xu, and Jiliang Tang. Deeprobust: A pytorch library for adversarial", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 525, + 241, + 538 + ], + "spans": [ + { + "bbox": [ + 125, + 525, + 241, + 538 + ], + "score": 1.0, + "content": "attacks and defenses, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "[28] Xuran Pan, Song Shiji, and Huang Gao. A unified framework for convolution-based graph", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 556, + 417, + 568 + ], + "spans": [ + { + "bbox": [ + 127, + 556, + 417, + 568 + ], + "score": 1.0, + "content": "neural networks. https://openreview.net/forum?id=zUMD–Fb9Bt, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 574, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 507, + 590 + ], + "score": 1.0, + "content": "[29] Meiqi Zhu, Xiao Wang, Chuan Shi, Houye Ji, and Peng Cui. Interpreting and unifying graph", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 586, + 354, + 600 + ], + "spans": [ + { + "bbox": [ + 126, + 586, + 354, + 600 + ], + "score": 1.0, + "content": "neural networks with an optimization framework, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "[30] Siheng Chen, Yonina C. Eldar, and Lingxiao Zhao. Graph unrolling networks: Interpretable", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 617, + 330, + 630 + ], + "spans": [ + { + "bbox": [ + 126, + 617, + 330, + 630 + ], + "score": 1.0, + "content": "neural networks for graph signal denoising, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "[31] Guohao Li, Matthias Müller, Ali Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 648, + 463, + 660 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 463, + 660 + ], + "score": 1.0, + "content": "as cnns? In The IEEE International Conference on Computer Vision (ICCV), 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "[32] Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 678, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 127, + 678, + 506, + 692 + ], + "score": 1.0, + "content": "Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. In", + "type": "text", + "cross_page": true + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 690, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 127, + 690, + 506, + 702 + ], + "score": 1.0, + "content": "Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on", + "type": "text", + "cross_page": true + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 700, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 700, + 507, + 714 + ], + "score": 1.0, + "content": "Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 5453–5462.", + "type": "text", + "cross_page": true + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 711, + 227, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 227, + 723 + ], + "score": 1.0, + "content": "PMLR, 10–15 Jul 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "[33] Negin Entezari, Saba A Al-Sayouri, Amirali Darvishzadeh, and Evangelos E Papalexakis. All", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 84, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 126, + 84, + 505, + 97 + ], + "score": 1.0, + "content": "you need is low (rank) defending against adversarial attacks on graphs. In Proceedings of the", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 95, + 475, + 108 + ], + "spans": [ + { + "bbox": [ + 127, + 95, + 475, + 108 + ], + "score": 1.0, + "content": "13th International Conference on Web Search and Data Mining, pages 169–177, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "score": 1.0, + "content": "[34] Wei Jin, Yao Ma, Xiaorui Liu, Xianfeng Tang, Suhang Wang, and Jiliang Tang. Graph", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 125, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 127, + 125, + 505, + 137 + ], + "score": 1.0, + "content": "structure learning for robust graph neural networks. In Proceedings of the 26th ACM SIGKDD", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 135, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 126, + 135, + 479, + 148 + ], + "score": 1.0, + "content": "International Conference on Knowledge Discovery & Data Mining, pages 66–74, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "[35] Yu Chen, Lingfei Wu, and Mohammed Zaki. Iterative deep graph learning for graph neural", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 164, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 126, + 164, + 506, + 179 + ], + "score": 1.0, + "content": "networks: Better and robust node embeddings. Advances in Neural Information Processing", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 176, + 203, + 188 + ], + "spans": [ + { + "bbox": [ + 127, + 176, + 203, + 188 + ], + "score": 1.0, + "content": "Systems, 33, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 194, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 209 + ], + "score": 1.0, + "content": "[36] Xiaorui Liu, Wei Jin, Yao Ma, Yaxin Li, Hua Liu, Yiqi Wang, Ming Yan, and Jiliang Tang.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 205, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 127, + 205, + 505, + 219 + ], + "score": 1.0, + "content": "Elastic graph neural networks. In Marina Meila and Tong Zhang, editors, Proceedings of the", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 216, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 506, + 230 + ], + "score": 1.0, + "content": "38th International Conference on Machine Learning, volume 139 of Proceedings of Machine", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 228, + 381, + 241 + ], + "spans": [ + { + "bbox": [ + 127, + 228, + 381, + 241 + ], + "score": 1.0, + "content": "Learning Research, pages 6837–6849. PMLR, 18–24 Jul 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 245, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 261 + ], + "score": 1.0, + "content": "[37] Wei Jin, Yaxing Li, Han Xu, Yiqi Wang, Shuiwang Ji, Charu Aggarwal, and Jiliang Tang.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 257, + 507, + 271 + ], + "spans": [ + { + "bbox": [ + 126, + 257, + 507, + 271 + ], + "score": 1.0, + "content": "Adversarial attacks and defenses on graphs. ACM SIGKDD Explorations Newsletter, 22(2):19–", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 269, + 168, + 280 + ], + "spans": [ + { + "bbox": [ + 127, + 269, + 168, + 280 + ], + "score": 1.0, + "content": "34, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 287, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 506, + 301 + ], + "score": 1.0, + "content": "[38] Yanqiao Zhu, Weizhi Xu, Jinghao Zhang, Qiang Liu, Shu Wu, and Liang Wang. Deep graph", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 298, + 507, + 312 + ], + "spans": [ + { + "bbox": [ + 127, + 298, + 507, + 312 + ], + "score": 1.0, + "content": "structure learning for robust representations: A survey. arXiv preprint arXiv:2103.03036, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 317, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 506, + 330 + ], + "score": 1.0, + "content": "[39] Laurent Condat. A primal–dual splitting method for convex optimization involving lipschitzian,", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 328, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 127, + 328, + 507, + 342 + ], + "score": 1.0, + "content": "proximable and linear composite terms. Journal of optimization theory and applications,", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 339, + 221, + 352 + ], + "spans": [ + { + "bbox": [ + 127, + 339, + 221, + 352 + ], + "score": 1.0, + "content": "158(2):460–479, 2013.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "[40] Fan RK Chung and Fan Chung Graham. Spectral graph theory. Number 92. American", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 369, + 233, + 380 + ], + "spans": [ + { + "bbox": [ + 127, + 369, + 233, + 380 + ], + "score": 1.0, + "content": "Mathematical Soc., 1997.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + } + ], + "index": 29, + "bbox_fs": [ + 105, + 281, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 56, + 507, + 730 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "score": 1.0, + "content": "[14] Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 82, + 507, + 99 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 507, + 99 + ], + "score": 1.0, + "content": "for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 128, + 94, + 199, + 106 + ], + "spans": [ + { + "bbox": [ + 128, + 94, + 199, + 106 + ], + "score": 1.0, + "content": "volume 32, 2018.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "[15] Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 124, + 469, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 124, + 469, + 138 + ], + "score": 1.0, + "content": "node classification. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 143, + 507, + 160 + ], + "spans": [ + { + "bbox": [ + 103, + 143, + 507, + 160 + ], + "score": 1.0, + "content": "[16] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 156, + 154, + 168 + ], + "spans": [ + { + "bbox": [ + 127, + 156, + 154, + 168 + ], + "score": 1.0, + "content": "2004.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 506, + 189 + ], + "score": 1.0, + "content": "[17] Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 186, + 399, + 200 + ], + "spans": [ + { + "bbox": [ + 126, + 186, + 399, + 200 + ], + "score": 1.0, + "content": "social networks. Annual review of sociology, 27(1):415–444, 2001.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 206, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 506, + 220 + ], + "score": 1.0, + "content": "[18] Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 218, + 181, + 230 + ], + "spans": [ + { + "bbox": [ + 126, + 218, + 181, + 230 + ], + "score": 1.0, + "content": "propagation.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "score": 1.0, + "content": "[19] Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Schölkopf.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 247, + 331, + 261 + ], + "spans": [ + { + "bbox": [ + 126, + 247, + 331, + 261 + ], + "score": 1.0, + "content": "Learning with local and global consistency. 2004.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 266, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 104, + 266, + 507, + 282 + ], + "score": 1.0, + "content": "[20] Meng Liu, Hongyang Gao, and Shuiwang Ji. Towards deeper graph neural networks. In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 278, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 127, + 278, + 506, + 292 + ], + "score": 1.0, + "content": "Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery &", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 289, + 238, + 302 + ], + "spans": [ + { + "bbox": [ + 125, + 289, + 238, + 302 + ], + "score": 1.0, + "content": "Data Mining. ACM, 2020.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 308, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 507, + 322 + ], + "score": 1.0, + "content": "[21] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 128, + 321, + 449, + 333 + ], + "spans": [ + { + "bbox": [ + 128, + 321, + 449, + 333 + ], + "score": 1.0, + "content": "Rad. Collective classification in network data. AI magazine, 29(3):93–93, 2008.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 338, + 507, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 507, + 353 + ], + "score": 1.0, + "content": "[22] Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 350, + 466, + 364 + ], + "spans": [ + { + "bbox": [ + 127, + 350, + 466, + 364 + ], + "score": 1.0, + "content": "Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868, 2018.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 369, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 383 + ], + "score": 1.0, + "content": "[23] Marinka Zitnik Yuxiao Dong Hongyu Ren Bowen Liu Michele Catasta Jure Leskovec Wei-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 380, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 126, + 380, + 506, + 395 + ], + "score": 1.0, + "content": "hua Hu, Matthias Fey. Open graph benchmark: Datasets for machine learning on graphs. arXiv", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 392, + 267, + 405 + ], + "spans": [ + { + "bbox": [ + 125, + 392, + 267, + 405 + ], + "score": 1.0, + "content": "preprint arXiv:2005.00687, 2020.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 506, + 425 + ], + "score": 1.0, + "content": "[24] Dingyuan Zhu, Ziwei Zhang, Peng Cui, and Wenwu Zhu. Robust graph convolutional networks", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 421, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 126, + 421, + 507, + 437 + ], + "score": 1.0, + "content": "against adversarial attacks. In Proceedings of the 25th ACM SIGKDD International Conference", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 432, + 399, + 448 + ], + "spans": [ + { + "bbox": [ + 126, + 432, + 399, + 448 + ], + "score": 1.0, + "content": "on Knowledge Discovery & Data Mining, pages 1399–1407, 2019.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 452, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 467 + ], + "score": 1.0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 463, + 226, + 477 + ], + "spans": [ + { + "bbox": [ + 127, + 463, + 226, + 477 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 483, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 497 + ], + "score": 1.0, + "content": "[26] Daniel Zügner, Amir Akbarnejad, and Stephan Günnemann. Adversarial attacks on neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 494, + 318, + 508 + ], + "spans": [ + { + "bbox": [ + 126, + 494, + 318, + 508 + ], + "score": 1.0, + "content": "networks for graph data. In KDD. ACM, 2018.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "[27] Yaxin Li, Wei Jin, Han Xu, and Jiliang Tang. Deeprobust: A pytorch library for adversarial", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 525, + 241, + 538 + ], + "spans": [ + { + "bbox": [ + 125, + 525, + 241, + 538 + ], + "score": 1.0, + "content": "attacks and defenses, 2020.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "[28] Xuran Pan, Song Shiji, and Huang Gao. A unified framework for convolution-based graph", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 556, + 417, + 568 + ], + "spans": [ + { + "bbox": [ + 127, + 556, + 417, + 568 + ], + "score": 1.0, + "content": "neural networks. https://openreview.net/forum?id=zUMD–Fb9Bt, 2020.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 574, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 507, + 590 + ], + "score": 1.0, + "content": "[29] Meiqi Zhu, Xiao Wang, Chuan Shi, Houye Ji, and Peng Cui. Interpreting and unifying graph", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 586, + 354, + 600 + ], + "spans": [ + { + "bbox": [ + 126, + 586, + 354, + 600 + ], + "score": 1.0, + "content": "neural networks with an optimization framework, 2021.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "[30] Siheng Chen, Yonina C. Eldar, and Lingxiao Zhao. Graph unrolling networks: Interpretable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 126, + 617, + 330, + 630 + ], + "spans": [ + { + "bbox": [ + 126, + 617, + 330, + 630 + ], + "score": 1.0, + "content": "neural networks for graph signal denoising, 2020.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "[31] Guohao Li, Matthias Müller, Ali Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 648, + 463, + 660 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 463, + 660 + ], + "score": 1.0, + "content": "as cnns? In The IEEE International Conference on Computer Vision (ICCV), 2019.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "[32] Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 678, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 127, + 678, + 506, + 692 + ], + "score": 1.0, + "content": "Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. In", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 690, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 127, + 690, + 506, + 702 + ], + "score": 1.0, + "content": "Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 700, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 700, + 507, + 714 + ], + "score": 1.0, + "content": "Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 5453–5462.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 711, + 227, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 227, + 723 + ], + "score": 1.0, + "content": "PMLR, 10–15 Jul 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 22 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 56, + 507, + 730 + ], + "lines": [], + "index": 22, + "bbox_fs": [ + 103, + 73, + 507, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 507, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "[33] Negin Entezari, Saba A Al-Sayouri, Amirali Darvishzadeh, and Evangelos E Papalexakis. All", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 84, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 126, + 84, + 505, + 97 + ], + "score": 1.0, + "content": "you need is low (rank) defending against adversarial attacks on graphs. In Proceedings of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 95, + 475, + 108 + ], + "spans": [ + { + "bbox": [ + 127, + 95, + 475, + 108 + ], + "score": 1.0, + "content": "13th International Conference on Web Search and Data Mining, pages 169–177, 2020.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 506, + 127 + ], + "score": 1.0, + "content": "[34] Wei Jin, Yao Ma, Xiaorui Liu, Xianfeng Tang, Suhang Wang, and Jiliang Tang. Graph", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 125, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 127, + 125, + 505, + 137 + ], + "score": 1.0, + "content": "structure learning for robust graph neural networks. In Proceedings of the 26th ACM SIGKDD", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 135, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 126, + 135, + 479, + 148 + ], + "score": 1.0, + "content": "International Conference on Knowledge Discovery & Data Mining, pages 66–74, 2020.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "[35] Yu Chen, Lingfei Wu, and Mohammed Zaki. Iterative deep graph learning for graph neural", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 164, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 126, + 164, + 506, + 179 + ], + "score": 1.0, + "content": "networks: Better and robust node embeddings. Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 176, + 203, + 188 + ], + "spans": [ + { + "bbox": [ + 127, + 176, + 203, + 188 + ], + "score": 1.0, + "content": "Systems, 33, 2020.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 194, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 209 + ], + "score": 1.0, + "content": "[36] Xiaorui Liu, Wei Jin, Yao Ma, Yaxin Li, Hua Liu, Yiqi Wang, Ming Yan, and Jiliang Tang.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 205, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 127, + 205, + 505, + 219 + ], + "score": 1.0, + "content": "Elastic graph neural networks. In Marina Meila and Tong Zhang, editors, Proceedings of the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 216, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 506, + 230 + ], + "score": 1.0, + "content": "38th International Conference on Machine Learning, volume 139 of Proceedings of Machine", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 228, + 381, + 241 + ], + "spans": [ + { + "bbox": [ + 127, + 228, + 381, + 241 + ], + "score": 1.0, + "content": "Learning Research, pages 6837–6849. PMLR, 18–24 Jul 2021.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 245, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 506, + 261 + ], + "score": 1.0, + "content": "[37] Wei Jin, Yaxing Li, Han Xu, Yiqi Wang, Shuiwang Ji, Charu Aggarwal, and Jiliang Tang.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 257, + 507, + 271 + ], + "spans": [ + { + "bbox": [ + 126, + 257, + 507, + 271 + ], + "score": 1.0, + "content": "Adversarial attacks and defenses on graphs. ACM SIGKDD Explorations Newsletter, 22(2):19–", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 269, + 168, + 280 + ], + "spans": [ + { + "bbox": [ + 127, + 269, + 168, + 280 + ], + "score": 1.0, + "content": "34, 2021.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 287, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 506, + 301 + ], + "score": 1.0, + "content": "[38] Yanqiao Zhu, Weizhi Xu, Jinghao Zhang, Qiang Liu, Shu Wu, and Liang Wang. Deep graph", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 298, + 507, + 312 + ], + "spans": [ + { + "bbox": [ + 127, + 298, + 507, + 312 + ], + "score": 1.0, + "content": "structure learning for robust representations: A survey. arXiv preprint arXiv:2103.03036, 2021.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 317, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 506, + 330 + ], + "score": 1.0, + "content": "[39] Laurent Condat. A primal–dual splitting method for convex optimization involving lipschitzian,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 328, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 127, + 328, + 507, + 342 + ], + "score": 1.0, + "content": "proximable and linear composite terms. Journal of optimization theory and applications,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 339, + 221, + 352 + ], + "spans": [ + { + "bbox": [ + 127, + 339, + 221, + 352 + ], + "score": 1.0, + "content": "158(2):460–479, 2013.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "[40] Fan RK Chung and Fan Chung Graham. Spectral graph theory. Number 92. American", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 369, + 233, + 380 + ], + "spans": [ + { + "bbox": [ + 127, + 369, + 233, + 380 + ], + "score": 1.0, + "content": "Mathematical Soc., 1997.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 11 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 72, + 507, + 384 + ], + "lines": [], + "index": 11, + "bbox_fs": [ + 105, + 73, + 507, + 380 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 156, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 131, + 92, + 208, + 104 + ], + "lines": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 146, + 107, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 145, + 106, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 145, + 106, + 505, + 120 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 118, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 162, + 118, + 288, + 130 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 131, + 376, + 144 + ], + "spans": [ + { + "bbox": [ + 145, + 131, + 376, + 144 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes]", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 144, + 468, + 157 + ], + "spans": [ + { + "bbox": [ + 145, + 144, + 468, + 157 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes]", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 145, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 161, + 167, + 214, + 181 + ], + "spans": [ + { + "bbox": [ + 161, + 167, + 214, + 181 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 131, + 183, + 302, + 194 + ], + "lines": [ + { + "bbox": [ + 129, + 182, + 304, + 196 + ], + "spans": [ + { + "bbox": [ + 129, + 182, + 304, + 196 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 145, + 197, + 450, + 222 + ], + "lines": [ + { + "bbox": [ + 145, + 197, + 450, + 209 + ], + "spans": [ + { + "bbox": [ + 145, + 197, + 450, + 209 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [Yes]", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 144, + 210, + 421, + 223 + ], + "spans": [ + { + "bbox": [ + 144, + 210, + 421, + 223 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [Yes]", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 130, + 226, + 241, + 237 + ], + "lines": [ + { + "bbox": [ + 128, + 223, + 243, + 239 + ], + "spans": [ + { + "bbox": [ + 128, + 223, + 243, + 239 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 146, + 240, + 506, + 334 + ], + "lines": [ + { + "bbox": [ + 145, + 239, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 145, + 239, + 507, + 253 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 161, + 251, + 447, + 263 + ], + "spans": [ + { + "bbox": [ + 161, + 251, + 447, + 263 + ], + "score": 1.0, + "content": "mental results (either in the supplemental material or as a URL)? [Yes]", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 144, + 263, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 144, + 263, + 506, + 277 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 160, + 273, + 246, + 288 + ], + "spans": [ + { + "bbox": [ + 160, + 273, + 246, + 288 + ], + "score": 1.0, + "content": "were chosen)? [Yes]", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 146, + 287, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 146, + 287, + 506, + 300 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 161, + 298, + 280, + 311 + ], + "spans": [ + { + "bbox": [ + 161, + 298, + 280, + 311 + ], + "score": 1.0, + "content": "ments multiple times)? [Yes]", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 146, + 309, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 146, + 309, + 506, + 325 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 162, + 323, + 367, + 335 + ], + "spans": [ + { + "bbox": [ + 162, + 323, + 367, + 335 + ], + "score": 1.0, + "content": "of GPUs, internal cluster, or cloud provider)? [No]", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 131, + 338, + 504, + 349 + ], + "lines": [ + { + "bbox": [ + 129, + 336, + 507, + 351 + ], + "spans": [ + { + "bbox": [ + 129, + 336, + 507, + 351 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 146, + 352, + 504, + 437 + ], + "lines": [ + { + "bbox": [ + 145, + 352, + 424, + 365 + ], + "spans": [ + { + "bbox": [ + 145, + 352, + 424, + 365 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes]", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 145, + 364, + 355, + 378 + ], + "spans": [ + { + "bbox": [ + 145, + 364, + 355, + 378 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [No]", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 145, + 377, + 504, + 391 + ], + "spans": [ + { + "bbox": [ + 145, + 377, + 504, + 391 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [No]", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 146, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 146, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 162, + 402, + 248, + 414 + ], + "spans": [ + { + "bbox": [ + 162, + 402, + 248, + 414 + ], + "score": 1.0, + "content": "using/curating? [No]", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 146, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 146, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 161, + 425, + 320, + 438 + ], + "spans": [ + { + "bbox": [ + 161, + 425, + 320, + 438 + ], + "score": 1.0, + "content": "information or offensive content? [No]", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 131, + 441, + 432, + 452 + ], + "lines": [ + { + "bbox": [ + 128, + 439, + 433, + 455 + ], + "spans": [ + { + "bbox": [ + 128, + 439, + 433, + 455 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 146, + 455, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 145, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 145, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 161, + 466, + 237, + 479 + ], + "spans": [ + { + "bbox": [ + 161, + 466, + 237, + 479 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 145, + 478, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 145, + 478, + 506, + 493 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 162, + 490, + 342, + 503 + ], + "spans": [ + { + "bbox": [ + 162, + 490, + 342, + 503 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 146, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 146, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 161, + 515, + 333, + 527 + ], + "spans": [ + { + "bbox": [ + 161, + 515, + 333, + 527 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 156, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 131, + 92, + 208, + 104 + ], + "lines": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 129, + 91, + 210, + 105 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 107, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 145, + 106, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 145, + 106, + 505, + 120 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 118, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 162, + 118, + 288, + 130 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 131, + 376, + 144 + ], + "spans": [ + { + "bbox": [ + 145, + 131, + 376, + 144 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes]", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 144, + 468, + 157 + ], + "spans": [ + { + "bbox": [ + 145, + 144, + 468, + 157 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes]", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 145, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 167, + 214, + 181 + ], + "spans": [ + { + "bbox": [ + 161, + 167, + 214, + 181 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + } + ], + "index": 4.5, + "bbox_fs": [ + 145, + 106, + 505, + 181 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 183, + 302, + 194 + ], + "lines": [ + { + "bbox": [ + 129, + 182, + 304, + 196 + ], + "spans": [ + { + "bbox": [ + 129, + 182, + 304, + 196 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 129, + 182, + 304, + 196 + ] + }, + { + "type": "text", + "bbox": [ + 145, + 197, + 450, + 222 + ], + "lines": [ + { + "bbox": [ + 145, + 197, + 450, + 209 + ], + "spans": [ + { + "bbox": [ + 145, + 197, + 450, + 209 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [Yes]", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 144, + 210, + 421, + 223 + ], + "spans": [ + { + "bbox": [ + 144, + 210, + 421, + 223 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [Yes]", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 144, + 197, + 450, + 223 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 226, + 241, + 237 + ], + "lines": [ + { + "bbox": [ + 128, + 223, + 243, + 239 + ], + "spans": [ + { + "bbox": [ + 128, + 223, + 243, + 239 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 128, + 223, + 243, + 239 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 240, + 506, + 334 + ], + "lines": [ + { + "bbox": [ + 145, + 239, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 145, + 239, + 507, + 253 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 251, + 447, + 263 + ], + "spans": [ + { + "bbox": [ + 161, + 251, + 447, + 263 + ], + "score": 1.0, + "content": "mental results (either in the supplemental material or as a URL)? [Yes]", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 144, + 263, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 144, + 263, + 506, + 277 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 160, + 273, + 246, + 288 + ], + "spans": [ + { + "bbox": [ + 160, + 273, + 246, + 288 + ], + "score": 1.0, + "content": "were chosen)? [Yes]", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 287, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 146, + 287, + 506, + 300 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 298, + 280, + 311 + ], + "spans": [ + { + "bbox": [ + 161, + 298, + 280, + 311 + ], + "score": 1.0, + "content": "ments multiple times)? [Yes]", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 309, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 146, + 309, + 506, + 325 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 323, + 367, + 335 + ], + "spans": [ + { + "bbox": [ + 162, + 323, + 367, + 335 + ], + "score": 1.0, + "content": "of GPUs, internal cluster, or cloud provider)? [No]", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + } + ], + "index": 15.5, + "bbox_fs": [ + 144, + 239, + 507, + 335 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 338, + 504, + 349 + ], + "lines": [ + { + "bbox": [ + 129, + 336, + 507, + 351 + ], + "spans": [ + { + "bbox": [ + 129, + 336, + 507, + 351 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 129, + 336, + 507, + 351 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 352, + 504, + 437 + ], + "lines": [ + { + "bbox": [ + 145, + 352, + 424, + 365 + ], + "spans": [ + { + "bbox": [ + 145, + 352, + 424, + 365 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes]", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 364, + 355, + 378 + ], + "spans": [ + { + "bbox": [ + 145, + 364, + 355, + 378 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [No]", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 377, + 504, + 391 + ], + "spans": [ + { + "bbox": [ + 145, + 377, + 504, + 391 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [No]", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 146, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 146, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 402, + 248, + 414 + ], + "spans": [ + { + "bbox": [ + 162, + 402, + 248, + 414 + ], + "score": 1.0, + "content": "using/curating? [No]", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 146, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 425, + 320, + 438 + ], + "spans": [ + { + "bbox": [ + 161, + 425, + 320, + 438 + ], + "score": 1.0, + "content": "information or offensive content? [No]", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + } + ], + "index": 24, + "bbox_fs": [ + 145, + 352, + 505, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 441, + 432, + 452 + ], + "lines": [ + { + "bbox": [ + 128, + 439, + 433, + 455 + ], + "spans": [ + { + "bbox": [ + 128, + 439, + 433, + 455 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 128, + 439, + 433, + 455 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 455, + 505, + 527 + ], + "lines": [ + { + "bbox": [ + 145, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 145, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 466, + 237, + 479 + ], + "spans": [ + { + "bbox": [ + 161, + 466, + 237, + 479 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 478, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 145, + 478, + 506, + 493 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 490, + 342, + 503 + ], + "spans": [ + { + "bbox": [ + 162, + 490, + 342, + 503 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 146, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 515, + 333, + 527 + ], + "spans": [ + { + "bbox": [ + 161, + 515, + 333, + 527 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + } + ], + "index": 31.5, + "bbox_fs": [ + 145, + 455, + 506, + 527 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/hfkER_KJiNw/hfkER_KJiNw_model.json b/parse/train/hfkER_KJiNw/hfkER_KJiNw_model.json new file mode 100644 index 0000000000000000000000000000000000000000..829e9b49529ec9aed28e3ae664993a4a61161f1d --- /dev/null +++ b/parse/train/hfkER_KJiNw/hfkER_KJiNw_model.json @@ -0,0 +1,21032 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 896, + 1304, + 896, + 1304, + 1261, + 398, + 1261 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1384, + 1404, + 1384, + 1404, + 1539, + 297, + 1539 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1622, + 1404, + 1622, + 1404, + 1764, + 298, + 1764 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1776, + 1405, + 1776, + 1405, + 1931, + 299, + 1931 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 538, + 653, + 1159, + 653, + 1159, + 748, + 538, + 748 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 730, + 443, + 969, + 443, + 969, + 506, + 730, + 506 + ], + "score": 0.928 + }, + { + "category_id": 8, + "poly": [ + 639, + 1552, + 1057, + 1552, + 1057, + 1603, + 639, + 1603 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 1080, + 443, + 1301, + 443, + 1301, + 506, + 1080, + 506 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 396, + 443, + 620, + 443, + 620, + 504, + 396, + 504 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 299, + 1316, + 530, + 1316, + 530, + 1353, + 299, + 1353 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 351, + 271, + 1350, + 271, + 1350, + 327, + 351, + 327 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 332, + 1953, + 1146, + 1953, + 1146, + 1982, + 332, + 1982 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 298, + 2033, + 1071, + 2033, + 1071, + 2062, + 298, + 2062 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 788, + 825, + 913, + 825, + 913, + 862, + 788, + 862 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1565, + 1399, + 1565, + 1399, + 1595, + 1366, + 1595 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 1139, + 551, + 1376, + 551, + 1376, + 615, + 1139, + 615 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 321, + 551, + 525, + 551, + 525, + 613, + 321, + 613 + ], + "score": 0.828 + }, + { + "category_id": 1, + "poly": [ + 566, + 550, + 785, + 550, + 785, + 614, + 566, + 614 + ], + "score": 0.78 + }, + { + "category_id": 1, + "poly": [ + 833, + 551, + 1097, + 551, + 1097, + 613, + 833, + 613 + ], + "score": 0.271 + }, + { + "category_id": 13, + "poly": [ + 372, + 1620, + 489, + 1620, + 489, + 1658, + 372, + 1658 + ], + "score": 0.93, + "latex": "\\mathbf { x } _ { u } ^ { ( k ) } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1658, + 379, + 1658, + 379, + 1706, + 297, + 1706 + ], + "score": 0.93, + "latex": "\\mathbf { m } _ { \\mathcal { N } ( u ) } ^ { ( k ) }" + }, + { + "category_id": 13, + "poly": [ + 959, + 1664, + 1026, + 1664, + 1026, + 1699, + 959, + 1699 + ], + "score": 0.92, + "latex": "\\mathcal { N } ( u )" + }, + { + "category_id": 14, + "poly": [ + 638, + 1552, + 1058, + 1552, + 1058, + 1605, + 638, + 1605 + ], + "score": 0.92, + "latex": "\\mathbf { x } _ { u } ^ { ( k + 1 ) } = \\mathrm { U P D A T E } ^ { ( k ) } \\left( \\mathbf { x } _ { u } ^ { ( k ) } , \\mathbf { m } _ { \\mathcal { N } ( u ) } ^ { ( k ) } \\right)" + }, + { + "category_id": 13, + "poly": [ + 973, + 1629, + 991, + 1629, + 991, + 1655, + 973, + 1655 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 880, + 1634, + 900, + 1634, + 900, + 1655, + 880, + 1655 + ], + "score": 0.79, + "latex": "u" + }, + { + "category_id": 13, + "poly": [ + 752, + 1670, + 772, + 1670, + 772, + 1693, + 752, + 1693 + ], + "score": 0.68, + "latex": "u" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1312.0, + 536.0, + 1312.0, + 536.0, + 1359.0, + 292.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 270.0, + 1349.0, + 270.0, + 1349.0, + 332.0, + 351.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1947.0, + 1148.0, + 1947.0, + 1148.0, + 1987.0, + 334.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2028.0, + 1074.0, + 2028.0, + 1074.0, + 2066.0, + 295.0, + 2066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 824.0, + 919.0, + 824.0, + 919.0, + 866.0, + 783.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 897.0, + 1304.0, + 897.0, + 1304.0, + 931.0, + 394.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 925.0, + 1305.0, + 925.0, + 1305.0, + 962.0, + 392.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 956.0, + 1304.0, + 956.0, + 1304.0, + 991.0, + 393.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 985.0, + 1305.0, + 985.0, + 1305.0, + 1020.0, + 394.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1017.0, + 1305.0, + 1017.0, + 1305.0, + 1052.0, + 393.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1046.0, + 1305.0, + 1046.0, + 1305.0, + 1084.0, + 392.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1080.0, + 1305.0, + 1080.0, + 1305.0, + 1112.0, + 395.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1111.0, + 1304.0, + 1111.0, + 1304.0, + 1143.0, + 395.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1139.0, + 1304.0, + 1139.0, + 1304.0, + 1171.0, + 395.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1171.0, + 1308.0, + 1171.0, + 1308.0, + 1201.0, + 392.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1200.0, + 1305.0, + 1200.0, + 1305.0, + 1232.0, + 394.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1229.0, + 846.0, + 1229.0, + 846.0, + 1264.0, + 394.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1384.0, + 1405.0, + 1384.0, + 1405.0, + 1422.0, + 295.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1416.0, + 1402.0, + 1416.0, + 1402.0, + 1450.0, + 296.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1446.0, + 1407.0, + 1446.0, + 1407.0, + 1483.0, + 294.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1477.0, + 1405.0, + 1477.0, + 1405.0, + 1511.0, + 296.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1504.0, + 1406.0, + 1504.0, + 1406.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 1620.0, + 879.0, + 1620.0, + 879.0, + 1665.0, + 490.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1620.0, + 972.0, + 1620.0, + 972.0, + 1665.0, + 901.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1665.0, + 992.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1664.0, + 751.0, + 1664.0, + 751.0, + 1703.0, + 380.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1664.0, + 958.0, + 1664.0, + 958.0, + 1703.0, + 773.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1664.0, + 1405.0, + 1664.0, + 1405.0, + 1703.0, + 1027.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1702.0, + 1404.0, + 1702.0, + 1404.0, + 1736.0, + 293.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1730.0, + 1042.0, + 1730.0, + 1042.0, + 1767.0, + 293.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.25, + 1624.0, + 413.25, + 1624.0, + 413.25, + 1662.0, + 292.25, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1668.0, + 384.0, + 1668.0, + 384.0, + 1711.0, + 292.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1778.0, + 1405.0, + 1778.0, + 1405.0, + 1811.0, + 297.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1804.0, + 1409.0, + 1804.0, + 1409.0, + 1844.0, + 294.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1835.0, + 1409.0, + 1835.0, + 1409.0, + 1873.0, + 293.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1865.0, + 1405.0, + 1865.0, + 1405.0, + 1904.0, + 293.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1900.0, + 1403.0, + 1900.0, + 1403.0, + 1933.0, + 297.0, + 1933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 649.0, + 1141.0, + 649.0, + 1141.0, + 690.0, + 557.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 678.0, + 1164.0, + 678.0, + 1164.0, + 724.0, + 534.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 710.0, + 1074.0, + 710.0, + 1074.0, + 757.0, + 626.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 442.0, + 936.0, + 442.0, + 936.0, + 481.0, + 760.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 476.0, + 968.0, + 476.0, + 968.0, + 508.0, + 731.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 439.0, + 1247.0, + 439.0, + 1247.0, + 476.0, + 1133.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 475.0, + 1305.0, + 475.0, + 1305.0, + 506.0, + 1076.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 443.0, + 589.0, + 443.0, + 589.0, + 472.0, + 431.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 476.0, + 621.0, + 476.0, + 621.0, + 505.0, + 396.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 549.0, + 1341.0, + 549.0, + 1341.0, + 589.0, + 1175.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 585.0, + 1378.0, + 585.0, + 1378.0, + 618.0, + 1140.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 549.0, + 481.0, + 549.0, + 481.0, + 582.0, + 369.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 585.0, + 529.0, + 585.0, + 529.0, + 614.0, + 320.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 546.0, + 736.0, + 546.0, + 736.0, + 584.0, + 616.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 584.0, + 791.0, + 584.0, + 791.0, + 616.0, + 561.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 550.0, + 1030.0, + 550.0, + 1030.0, + 582.0, + 895.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 583.0, + 1103.0, + 583.0, + 1103.0, + 614.0, + 827.0, + 614.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 340, + 1405, + 340, + 1405, + 703, + 297, + 703 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1251, + 1405, + 1251, + 1405, + 1451, + 297, + 1451 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 295, + 1462, + 1405, + 1462, + 1405, + 1749, + 295, + 1749 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 332, + 834, + 1406, + 834, + 1406, + 1099, + 332, + 1099 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 297, + 202, + 1405, + 202, + 1405, + 325, + 297, + 325 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 298, + 1763, + 1404, + 1763, + 1404, + 1855, + 298, + 1855 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 300, + 718, + 1405, + 718, + 1405, + 811, + 300, + 811 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 297, + 1944, + 1402, + 1944, + 1402, + 2007, + 297, + 2007 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 298, + 1888, + 582, + 1888, + 582, + 1922, + 298, + 1922 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 297, + 1205, + 1371, + 1205, + 1371, + 1239, + 297, + 1239 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 298, + 1137, + 522, + 1137, + 522, + 1176, + 298, + 1176 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2084, + 841, + 2084 + ], + "score": 0.727 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2084, + 841, + 2084 + ], + "score": 0.117 + }, + { + "category_id": 13, + "poly": [ + 991, + 1314, + 1233, + 1314, + 1233, + 1372, + 991, + 1372 + ], + "score": 0.95, + "latex": "\\Vert \\mathbf { X } \\Vert _ { F } = \\sqrt { \\sum _ { i j } \\mathbf { X } _ { i j } ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 920, + 1378, + 1136, + 1378, + 1136, + 1412, + 920, + 1412 + ], + "score": 0.93, + "latex": "\\| \\mathbf { X } \\| _ { 2 } = \\sigma _ { \\operatorname* { m a x } } ( \\mathbf { X } )" + }, + { + "category_id": 13, + "poly": [ + 1221, + 1378, + 1327, + 1378, + 1327, + 1412, + 1221, + 1412 + ], + "score": 0.93, + "latex": "\\sigma _ { \\mathrm { m a x } } ( \\mathbf { X } )" + }, + { + "category_id": 13, + "poly": [ + 851, + 1464, + 1067, + 1464, + 1067, + 1498, + 851, + 1498 + ], + "score": 0.93, + "latex": "\\mathcal { V } = \\{ v _ { 1 } , \\ldots , v _ { n } \\}" + }, + { + "category_id": 13, + "poly": [ + 611, + 1495, + 685, + 1495, + 685, + 1529, + 611, + 1529 + ], + "score": 0.92, + "latex": "\\mathcal { N } ( v _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 296, + 1367, + 624, + 1367, + 624, + 1423, + 296, + 1423 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\sum _ { i } \\| \\mathbf { X } _ { i } \\| _ { 2 } = \\sum _ { i } \\sqrt { \\sum _ { j } \\mathbf { X } _ { i j } ^ { 2 } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 584, + 1676, + 719, + 1676, + 719, + 1711, + 584, + 1711 + ], + "score": 0.92, + "latex": "\\hat { \\mathbf { A } } = \\mathbf { A } + \\mathbf { I }" + }, + { + "category_id": 13, + "poly": [ + 891, + 1711, + 1020, + 1711, + 1020, + 1746, + 891, + 1746 + ], + "score": 0.92, + "latex": "\\tilde { \\mathbf { L } } = \\mathbf { I } - \\tilde { \\mathbf { A } }" + }, + { + "category_id": 13, + "poly": [ + 581, + 1554, + 737, + 1554, + 737, + 1586, + 581, + 1586 + ], + "score": 0.91, + "latex": "\\mathbf { X } _ { \\mathrm { f e a } } \\in \\mathbb { R } ^ { n \\times d }" + }, + { + "category_id": 13, + "poly": [ + 616, + 1588, + 727, + 1588, + 727, + 1620, + 616, + 1620 + ], + "score": 0.91, + "latex": "\\mathbf { A } _ { i j } ~ = ~ 1" + }, + { + "category_id": 13, + "poly": [ + 296, + 1617, + 408, + 1617, + 408, + 1650, + 296, + 1650 + ], + "score": 0.91, + "latex": "\\mathbf { A } _ { i j } \\ = \\ 0" + }, + { + "category_id": 13, + "poly": [ + 1272, + 1249, + 1397, + 1249, + 1397, + 1281, + 1272, + 1281 + ], + "score": 0.91, + "latex": "\\mathbf { X } \\in \\mathbb { R } ^ { n \\times d }" + }, + { + "category_id": 13, + "poly": [ + 721, + 1283, + 768, + 1283, + 768, + 1317, + 721, + 1317 + ], + "score": 0.91, + "latex": "\\mathbf { X } _ { i j }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1497, + 506, + 1497, + 506, + 1529, + 297, + 1529 + ], + "score": 0.9, + "latex": "\\mathcal { E } = \\{ e _ { 1 } , \\ldots , e _ { m } \\}" + }, + { + "category_id": 13, + "poly": [ + 345, + 1463, + 494, + 1463, + 494, + 1497, + 345, + 1497 + ], + "score": 0.9, + "latex": "\\mathcal { G } = \\{ \\nu , \\mathcal { E } \\}" + }, + { + "category_id": 13, + "poly": [ + 381, + 1586, + 521, + 1586, + 521, + 1616, + 381, + 1616 + ], + "score": 0.9, + "latex": "\\mathbf { A } \\in \\mathbb { R } ^ { n \\times n }" + }, + { + "category_id": 13, + "poly": [ + 1290, + 1323, + 1403, + 1323, + 1403, + 1360, + 1290, + 1360 + ], + "score": 0.9, + "latex": "\\| \\mathbf { X } \\| _ { 2 1 } =" + }, + { + "category_id": 13, + "poly": [ + 538, + 1326, + 576, + 1326, + 576, + 1357, + 538, + 1357 + ], + "score": 0.89, + "latex": "\\ell _ { 2 1 }" + }, + { + "category_id": 13, + "poly": [ + 1043, + 1617, + 1204, + 1617, + 1204, + 1647, + 1043, + 1647 + ], + "score": 0.89, + "latex": "\\mathbf { L } = \\mathbf { D } - \\mathbf { A }" + }, + { + "category_id": 13, + "poly": [ + 382, + 1283, + 419, + 1283, + 419, + 1313, + 382, + 1313 + ], + "score": 0.89, + "latex": "\\mathbf { X } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1676, + 503, + 1676, + 503, + 1710, + 298, + 1710 + ], + "score": 0.88, + "latex": "\\tilde { \\bf A } = \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } \\bar { \\hat { \\bf A } } \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 1228, + 1591, + 1255, + 1591, + 1255, + 1617, + 1228, + 1617 + ], + "score": 0.85, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1312, + 1591, + 1341, + 1591, + 1341, + 1620, + 1312, + 1620 + ], + "score": 0.85, + "latex": "v _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1308, + 1499, + 1334, + 1499, + 1334, + 1526, + 1308, + 1526 + ], + "score": 0.84, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1156, + 1500, + 1182, + 1500, + 1182, + 1526, + 1156, + 1526 + ], + "score": 0.83, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 987, + 1557, + 1010, + 1557, + 1010, + 1586, + 987, + 1586 + ], + "score": 0.81, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 1198, + 1285, + 1214, + 1285, + 1214, + 1314, + 1198, + 1314 + ], + "score": 0.81, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 821, + 1527, + 840, + 1527, + 840, + 1553, + 821, + 1553 + ], + "score": 0.79, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 572, + 1285, + 585, + 1285, + 585, + 1310, + 572, + 1310 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 788, + 1325, + 816, + 1325, + 816, + 1354, + 788, + 1354 + ], + "score": 0.68, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 1048, + 1285, + 1061, + 1285, + 1061, + 1310, + 1048, + 1310 + ], + "score": 0.68, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 296, + 1712, + 325, + 1712, + 325, + 1745, + 296, + 1745 + ], + "score": 0.62, + "latex": "\\hat { \\bf D }" + }, + { + "category_id": 13, + "poly": [ + 854, + 1253, + 882, + 1253, + 882, + 1280, + 854, + 1280 + ], + "score": 0.58, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 1297, + 1619, + 1327, + 1619, + 1327, + 1645, + 1297, + 1645 + ], + "score": 0.55, + "latex": "\\mathbf { D }" + }, + { + "category_id": 13, + "poly": [ + 570, + 1422, + 597, + 1422, + 597, + 1447, + 570, + 1447 + ], + "score": 0.47, + "latex": "\\mathbf { X }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1883.0, + 585.0, + 1883.0, + 585.0, + 1929.0, + 292.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1131.0, + 526.0, + 1131.0, + 526.0, + 1190.0, + 289.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2093.0, + 839.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2093.0, + 838.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 340.0, + 1406.0, + 340.0, + 1406.0, + 373.0, + 295.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 368.0, + 1406.0, + 368.0, + 1406.0, + 406.0, + 292.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 402.0, + 1405.0, + 402.0, + 1405.0, + 435.0, + 296.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 431.0, + 1406.0, + 431.0, + 1406.0, + 466.0, + 292.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 462.0, + 1407.0, + 462.0, + 1407.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 528.0, + 294.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 522.0, + 1402.0, + 522.0, + 1402.0, + 554.0, + 295.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 549.0, + 1406.0, + 549.0, + 1406.0, + 587.0, + 294.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 583.0, + 1403.0, + 583.0, + 1403.0, + 616.0, + 296.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 611.0, + 1407.0, + 611.0, + 1407.0, + 649.0, + 291.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 641.0, + 1407.0, + 641.0, + 1407.0, + 677.0, + 292.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 835.0, + 673.0, + 835.0, + 707.0, + 294.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1245.0, + 853.0, + 1245.0, + 853.0, + 1287.0, + 292.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1245.0, + 1271.0, + 1245.0, + 1271.0, + 1287.0, + 883.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1245.0, + 1407.0, + 1245.0, + 1407.0, + 1287.0, + 1398.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1280.0, + 381.0, + 1280.0, + 381.0, + 1318.0, + 295.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1280.0, + 571.0, + 1280.0, + 571.0, + 1318.0, + 420.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1280.0, + 720.0, + 1280.0, + 720.0, + 1318.0, + 586.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1280.0, + 1047.0, + 1280.0, + 1047.0, + 1318.0, + 769.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1280.0, + 1197.0, + 1280.0, + 1197.0, + 1318.0, + 1062.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1280.0, + 1405.0, + 1280.0, + 1405.0, + 1318.0, + 1215.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1319.0, + 537.0, + 1319.0, + 537.0, + 1368.0, + 291.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1319.0, + 787.0, + 1319.0, + 787.0, + 1368.0, + 577.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1319.0, + 990.0, + 1319.0, + 990.0, + 1368.0, + 817.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1319.0, + 1289.0, + 1319.0, + 1289.0, + 1368.0, + 1234.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1367.0, + 295.0, + 1367.0, + 295.0, + 1426.0, + 290.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1367.0, + 919.0, + 1367.0, + 919.0, + 1426.0, + 625.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1367.0, + 1220.0, + 1367.0, + 1220.0, + 1426.0, + 1137.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1367.0, + 1410.0, + 1367.0, + 1410.0, + 1426.0, + 1328.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1417.0, + 569.0, + 1417.0, + 569.0, + 1454.0, + 293.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1417.0, + 607.0, + 1417.0, + 607.0, + 1454.0, + 598.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1460.0, + 344.0, + 1460.0, + 344.0, + 1500.0, + 292.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1460.0, + 850.0, + 1460.0, + 850.0, + 1500.0, + 495.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1460.0, + 1407.0, + 1460.0, + 1407.0, + 1500.0, + 1068.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1489.0, + 296.0, + 1489.0, + 296.0, + 1534.0, + 290.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 1489.0, + 610.0, + 1489.0, + 610.0, + 1534.0, + 507.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1489.0, + 1155.0, + 1489.0, + 1155.0, + 1534.0, + 686.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1489.0, + 1307.0, + 1489.0, + 1307.0, + 1534.0, + 1183.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1489.0, + 1411.0, + 1489.0, + 1411.0, + 1534.0, + 1335.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1527.0, + 820.0, + 1527.0, + 820.0, + 1557.0, + 297.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 1527.0, + 1406.0, + 1527.0, + 1406.0, + 1557.0, + 841.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1550.0, + 580.0, + 1550.0, + 580.0, + 1591.0, + 292.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1550.0, + 986.0, + 1550.0, + 986.0, + 1591.0, + 738.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1550.0, + 1408.0, + 1550.0, + 1408.0, + 1591.0, + 1011.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1581.0, + 380.0, + 1581.0, + 380.0, + 1622.0, + 292.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1581.0, + 615.0, + 1581.0, + 615.0, + 1622.0, + 522.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1581.0, + 1227.0, + 1581.0, + 1227.0, + 1622.0, + 728.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1581.0, + 1311.0, + 1581.0, + 1311.0, + 1622.0, + 1256.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1581.0, + 1407.0, + 1581.0, + 1407.0, + 1622.0, + 1342.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1615.0, + 1042.0, + 1615.0, + 1042.0, + 1653.0, + 409.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1615.0, + 1296.0, + 1615.0, + 1296.0, + 1653.0, + 1205.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1615.0, + 1407.0, + 1615.0, + 1407.0, + 1653.0, + 1328.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1647.0, + 1406.0, + 1647.0, + 1406.0, + 1680.0, + 294.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1676.0, + 297.0, + 1676.0, + 297.0, + 1717.0, + 293.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1676.0, + 583.0, + 1676.0, + 583.0, + 1717.0, + 504.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1676.0, + 1406.0, + 1676.0, + 1406.0, + 1717.0, + 720.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1713.0, + 890.0, + 1713.0, + 890.0, + 1751.0, + 326.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1713.0, + 1031.0, + 1713.0, + 1031.0, + 1751.0, + 1021.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 833.0, + 1407.0, + 833.0, + 1407.0, + 869.0, + 331.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 865.0, + 1402.0, + 865.0, + 1402.0, + 897.0, + 357.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 894.0, + 1094.0, + 894.0, + 1094.0, + 931.0, + 354.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 931.0, + 1407.0, + 931.0, + 1407.0, + 973.0, + 331.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 965.0, + 1066.0, + 965.0, + 1066.0, + 997.0, + 353.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1003.0, + 1405.0, + 1003.0, + 1405.0, + 1039.0, + 333.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1035.0, + 1405.0, + 1035.0, + 1405.0, + 1070.0, + 354.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1065.0, + 721.0, + 1065.0, + 721.0, + 1102.0, + 355.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 238.0, + 295.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 234.0, + 1408.0, + 234.0, + 1408.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 298.0, + 292.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 293.0, + 400.0, + 293.0, + 400.0, + 329.0, + 293.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1761.0, + 1405.0, + 1761.0, + 1405.0, + 1798.0, + 293.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1790.0, + 1406.0, + 1790.0, + 1406.0, + 1829.0, + 292.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1821.0, + 1297.0, + 1821.0, + 1297.0, + 1861.0, + 292.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 716.0, + 1404.0, + 716.0, + 1404.0, + 752.0, + 294.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 748.0, + 1404.0, + 748.0, + 1404.0, + 781.0, + 295.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 779.0, + 644.0, + 779.0, + 644.0, + 812.0, + 295.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1409.0, + 1941.0, + 1409.0, + 1983.0, + 293.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1974.0, + 1409.0, + 1974.0, + 1409.0, + 2012.0, + 292.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1200.0, + 1377.0, + 1200.0, + 1377.0, + 1246.0, + 294.0, + 1246.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 202, + 1405, + 202, + 1405, + 569, + 297, + 569 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1449, + 1405, + 1449, + 1405, + 1724, + 297, + 1724 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 316, + 618, + 1382, + 618, + 1382, + 916, + 316, + 916 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 317, + 1047, + 1382, + 1047, + 1382, + 1351, + 317, + 1351 + ], + "score": 0.966 + }, + { + "category_id": 2, + "poly": [ + 297, + 1949, + 1403, + 1949, + 1403, + 2008, + 297, + 2008 + ], + "score": 0.926 + }, + { + "category_id": 4, + "poly": [ + 496, + 1369, + 1202, + 1369, + 1202, + 1402, + 496, + 1402 + ], + "score": 0.923 + }, + { + "category_id": 4, + "poly": [ + 480, + 932, + 1215, + 932, + 1215, + 965, + 480, + 965 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.697 + }, + { + "category_id": 1, + "poly": [ + 330, + 1842, + 1396, + 1842, + 1396, + 1904, + 330, + 1904 + ], + "score": 0.512 + }, + { + "category_id": 1, + "poly": [ + 326, + 1756, + 1400, + 1756, + 1400, + 1818, + 326, + 1818 + ], + "score": 0.469 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.233 + }, + { + "category_id": 13, + "poly": [ + 546, + 233, + 601, + 233, + 601, + 264, + 546, + 264 + ], + "score": 0.9, + "latex": "1 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1007, + 1481, + 1069, + 1481, + 1069, + 1510, + 1007, + 1510 + ], + "score": 0.84, + "latex": "> 1 0" + }, + { + "category_id": 13, + "poly": [ + 718, + 1482, + 767, + 1482, + 767, + 1511, + 718, + 1511 + ], + "score": 0.83, + "latex": "> 2" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 620.0, + 363.0, + 620.0, + 363.0, + 641.0, + 329.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 621.0, + 735.0, + 621.0, + 735.0, + 644.0, + 701.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 621.0, + 1106.0, + 621.0, + 1106.0, + 644.0, + 1071.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 636.0, + 364.0, + 636.0, + 364.0, + 659.0, + 329.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 641.0, + 526.0, + 641.0, + 526.0, + 650.0, + 516.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 635.0, + 621.0, + 635.0, + 621.0, + 655.0, + 533.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 638.0, + 989.0, + 638.0, + 989.0, + 658.0, + 909.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 643.0, + 1279.0, + 643.0, + 1279.0, + 651.0, + 1264.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 638.0, + 1362.0, + 638.0, + 1362.0, + 658.0, + 1286.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 654.0, + 364.0, + 654.0, + 364.0, + 676.0, + 329.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 655.0, + 532.0, + 655.0, + 532.0, + 669.0, + 510.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 650.0, + 629.0, + 650.0, + 629.0, + 674.0, + 533.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 656.0, + 736.0, + 656.0, + 736.0, + 677.0, + 701.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 653.0, + 998.0, + 653.0, + 998.0, + 676.0, + 907.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 656.0, + 1106.0, + 656.0, + 1106.0, + 679.0, + 1071.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 653.0, + 1370.0, + 653.0, + 1370.0, + 676.0, + 1263.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 671.0, + 377.0, + 671.0, + 377.0, + 783.0, + 308.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 686.0, + 604.0, + 686.0, + 604.0, + 695.0, + 594.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 692.0, + 734.0, + 692.0, + 734.0, + 711.0, + 703.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 691.0, + 1105.0, + 691.0, + 1105.0, + 714.0, + 1072.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 723.0, + 363.0, + 723.0, + 363.0, + 746.0, + 330.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 721.0, + 503.0, + 721.0, + 503.0, + 732.0, + 492.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 704.0, + 735.0, + 704.0, + 735.0, + 770.0, + 689.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 704.0, + 1105.0, + 704.0, + 1105.0, + 770.0, + 1059.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 740.0, + 363.0, + 740.0, + 363.0, + 762.0, + 329.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 745.0, + 470.0, + 745.0, + 470.0, + 755.0, + 460.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 757.0, + 364.0, + 757.0, + 364.0, + 779.0, + 330.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 763.0, + 732.0, + 763.0, + 732.0, + 782.0, + 703.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 762.0, + 1106.0, + 762.0, + 1106.0, + 783.0, + 1072.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 776.0, + 364.0, + 776.0, + 364.0, + 797.0, + 329.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 792.0, + 363.0, + 792.0, + 363.0, + 814.0, + 329.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 797.0, + 735.0, + 797.0, + 735.0, + 818.0, + 701.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 797.0, + 1105.0, + 797.0, + 1105.0, + 818.0, + 1071.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 809.0, + 364.0, + 809.0, + 364.0, + 832.0, + 329.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 809.0, + 374.0, + 809.0, + 374.0, + 819.0, + 366.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 827.0, + 363.0, + 827.0, + 363.0, + 849.0, + 329.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 843.0, + 374.0, + 843.0, + 374.0, + 858.0, + 360.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 843.0, + 407.0, + 843.0, + 407.0, + 859.0, + 392.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 843.0, + 440.0, + 843.0, + 440.0, + 858.0, + 425.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 842.0, + 474.0, + 842.0, + 474.0, + 858.0, + 458.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 842.0, + 508.0, + 842.0, + 508.0, + 858.0, + 491.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 840.0, + 544.0, + 840.0, + 544.0, + 859.0, + 521.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 840.0, + 577.0, + 840.0, + 577.0, + 860.0, + 554.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 840.0, + 611.0, + 840.0, + 611.0, + 860.0, + 587.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 840.0, + 643.0, + 840.0, + 643.0, + 860.0, + 619.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 843.0, + 772.0, + 843.0, + 772.0, + 858.0, + 758.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 843.0, + 806.0, + 843.0, + 806.0, + 858.0, + 792.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 844.0, + 839.0, + 844.0, + 839.0, + 855.0, + 827.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 842.0, + 875.0, + 842.0, + 875.0, + 858.0, + 860.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 840.0, + 912.0, + 840.0, + 912.0, + 859.0, + 890.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 840.0, + 946.0, + 840.0, + 946.0, + 859.0, + 923.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 840.0, + 981.0, + 840.0, + 981.0, + 859.0, + 957.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 840.0, + 1013.0, + 840.0, + 1013.0, + 859.0, + 990.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 843.0, + 1144.0, + 843.0, + 1144.0, + 858.0, + 1128.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 843.0, + 1177.0, + 843.0, + 1177.0, + 858.0, + 1162.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 843.0, + 1212.0, + 843.0, + 1212.0, + 858.0, + 1196.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 842.0, + 1245.0, + 842.0, + 1245.0, + 858.0, + 1230.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 840.0, + 1284.0, + 840.0, + 1284.0, + 859.0, + 1261.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 840.0, + 1318.0, + 840.0, + 1318.0, + 859.0, + 1294.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 840.0, + 1351.0, + 840.0, + 1351.0, + 859.0, + 1326.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 840.0, + 1385.0, + 840.0, + 1385.0, + 859.0, + 1360.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 853.0, + 555.0, + 853.0, + 555.0, + 876.0, + 442.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 853.0, + 924.0, + 853.0, + 924.0, + 876.0, + 810.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 854.0, + 1294.0, + 854.0, + 1294.0, + 874.0, + 1183.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 883.0, + 542.0, + 883.0, + 542.0, + 915.0, + 415.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 883.0, + 908.0, + 883.0, + 908.0, + 915.0, + 790.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 883.0, + 1269.0, + 883.0, + 1269.0, + 916.0, + 1169.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 778.5, + 452.0, + 778.5, + 452.0, + 788.5, + 421.0, + 788.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 828.5, + 741.0, + 828.5, + 741.0, + 858.5, + 702.0, + 858.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 828.0, + 1111.0, + 828.0, + 1111.0, + 859.0, + 1073.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1055.0, + 358.0, + 1055.0, + 358.0, + 1074.0, + 331.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1055.0, + 728.0, + 1055.0, + 728.0, + 1074.0, + 703.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1071.0, + 1099.0, + 1071.0, + 1099.0, + 1090.0, + 1073.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1090.0, + 728.0, + 1090.0, + 728.0, + 1109.0, + 703.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 1104.0, + 466.0, + 1104.0, + 466.0, + 1113.0, + 444.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1104.0, + 536.0, + 1104.0, + 536.0, + 1114.0, + 522.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1104.0, + 1099.0, + 1104.0, + 1099.0, + 1123.0, + 1073.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1127.0, + 357.0, + 1127.0, + 357.0, + 1145.0, + 333.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1127.0, + 728.0, + 1127.0, + 728.0, + 1145.0, + 704.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1122.0, + 1112.0, + 1122.0, + 1112.0, + 1216.0, + 1049.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1138.0, + 337.0, + 1138.0, + 337.0, + 1204.0, + 317.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 1135.0, + 728.0, + 1135.0, + 728.0, + 1206.0, + 689.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1170.0, + 1099.0, + 1170.0, + 1099.0, + 1189.0, + 1073.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1198.0, + 359.0, + 1198.0, + 359.0, + 1217.0, + 333.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1199.0, + 728.0, + 1199.0, + 728.0, + 1217.0, + 703.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1204.0, + 1099.0, + 1204.0, + 1099.0, + 1222.0, + 1074.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1202.0, + 1204.0, + 1202.0, + 1204.0, + 1208.0, + 1198.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1232.0, + 482.0, + 1232.0, + 482.0, + 1255.0, + 394.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1233.0, + 728.0, + 1233.0, + 728.0, + 1252.0, + 703.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1234.0, + 848.0, + 1234.0, + 848.0, + 1255.0, + 766.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1234.0, + 1101.0, + 1234.0, + 1101.0, + 1257.0, + 1072.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1239.0, + 1124.0, + 1239.0, + 1124.0, + 1248.0, + 1114.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1230.0, + 1215.0, + 1230.0, + 1215.0, + 1257.0, + 1135.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1251.0, + 489.0, + 1251.0, + 489.0, + 1271.0, + 396.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1249.0, + 857.0, + 1249.0, + 857.0, + 1273.0, + 765.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1257.0, + 1127.0, + 1257.0, + 1127.0, + 1266.0, + 1114.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1249.0, + 1221.0, + 1249.0, + 1221.0, + 1273.0, + 1136.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1279.0, + 398.0, + 1279.0, + 398.0, + 1294.0, + 382.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1280.0, + 430.0, + 1280.0, + 430.0, + 1292.0, + 419.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 1280.0, + 465.0, + 1280.0, + 465.0, + 1292.0, + 454.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1279.0, + 501.0, + 1279.0, + 501.0, + 1294.0, + 487.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1276.0, + 539.0, + 1276.0, + 539.0, + 1295.0, + 517.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1279.0, + 573.0, + 1279.0, + 573.0, + 1294.0, + 552.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 1277.0, + 609.0, + 1277.0, + 609.0, + 1295.0, + 585.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1279.0, + 641.0, + 1279.0, + 641.0, + 1294.0, + 620.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1279.0, + 768.0, + 1279.0, + 768.0, + 1294.0, + 754.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1279.0, + 802.0, + 1279.0, + 802.0, + 1294.0, + 788.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1279.0, + 837.0, + 1279.0, + 837.0, + 1294.0, + 823.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1279.0, + 873.0, + 1279.0, + 873.0, + 1294.0, + 857.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1279.0, + 909.0, + 1279.0, + 909.0, + 1294.0, + 888.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1279.0, + 943.0, + 1279.0, + 943.0, + 1294.0, + 922.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1276.0, + 979.0, + 1276.0, + 979.0, + 1295.0, + 955.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1276.0, + 1013.0, + 1276.0, + 1013.0, + 1295.0, + 990.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1277.0, + 1140.0, + 1277.0, + 1140.0, + 1294.0, + 1124.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1279.0, + 1174.0, + 1279.0, + 1174.0, + 1294.0, + 1159.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1279.0, + 1207.0, + 1279.0, + 1207.0, + 1294.0, + 1193.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 1279.0, + 1241.0, + 1279.0, + 1241.0, + 1294.0, + 1227.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1279.0, + 1280.0, + 1279.0, + 1280.0, + 1294.0, + 1260.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1279.0, + 1314.0, + 1279.0, + 1314.0, + 1294.0, + 1294.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1279.0, + 1348.0, + 1279.0, + 1348.0, + 1294.0, + 1328.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1360.0, + 1276.0, + 1383.0, + 1276.0, + 1383.0, + 1295.0, + 1360.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1286.0, + 551.0, + 1286.0, + 551.0, + 1313.0, + 436.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1286.0, + 923.0, + 1286.0, + 923.0, + 1313.0, + 806.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1286.0, + 1293.0, + 1286.0, + 1293.0, + 1313.0, + 1178.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1319.0, + 542.0, + 1319.0, + 542.0, + 1353.0, + 415.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 1321.0, + 908.0, + 1321.0, + 908.0, + 1350.0, + 791.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1319.0, + 1269.0, + 1319.0, + 1269.0, + 1352.0, + 1168.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.75, + 1225.5, + 897.75, + 1225.5, + 897.75, + 1232.5, + 886.75, + 1232.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1266.5, + 733.0, + 1266.5, + 733.0, + 1293.5, + 706.0, + 1293.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1268.5, + 362.0, + 1268.5, + 362.0, + 1293.0, + 335.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 1268.5, + 1104.0, + 1268.5, + 1104.0, + 1293.0, + 1076.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1942.0, + 1407.0, + 1942.0, + 1407.0, + 1985.0, + 326.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 459.0, + 1973.0, + 459.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1363.0, + 1204.0, + 1363.0, + 1204.0, + 1409.0, + 493.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 930.0, + 1215.0, + 930.0, + 1215.0, + 969.0, + 483.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 202.0, + 1406.0, + 202.0, + 1406.0, + 240.0, + 295.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 232.0, + 545.0, + 232.0, + 545.0, + 270.0, + 294.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 232.0, + 1407.0, + 232.0, + 1407.0, + 270.0, + 602.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 260.0, + 1406.0, + 260.0, + 1406.0, + 299.0, + 292.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 294.0, + 1407.0, + 294.0, + 1407.0, + 328.0, + 295.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 323.0, + 1406.0, + 323.0, + 1406.0, + 361.0, + 294.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 389.0, + 294.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 384.0, + 1405.0, + 384.0, + 1405.0, + 417.0, + 294.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 416.0, + 1405.0, + 416.0, + 1405.0, + 450.0, + 295.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 444.0, + 1406.0, + 444.0, + 1406.0, + 482.0, + 294.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 476.0, + 1405.0, + 476.0, + 1405.0, + 510.0, + 294.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 541.0, + 295.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 534.0, + 1078.0, + 534.0, + 1078.0, + 573.0, + 294.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1448.0, + 1406.0, + 1448.0, + 1406.0, + 1486.0, + 295.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1476.0, + 717.0, + 1476.0, + 717.0, + 1519.0, + 292.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1476.0, + 1006.0, + 1476.0, + 1006.0, + 1519.0, + 768.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1476.0, + 1405.0, + 1476.0, + 1405.0, + 1519.0, + 1070.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1508.0, + 1405.0, + 1508.0, + 1405.0, + 1545.0, + 294.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1538.0, + 1406.0, + 1538.0, + 1406.0, + 1577.0, + 292.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1571.0, + 1405.0, + 1571.0, + 1405.0, + 1606.0, + 292.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1600.0, + 1408.0, + 1600.0, + 1408.0, + 1637.0, + 294.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1629.0, + 1407.0, + 1629.0, + 1407.0, + 1667.0, + 292.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1662.0, + 1405.0, + 1662.0, + 1405.0, + 1697.0, + 295.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1691.0, + 473.0, + 1691.0, + 473.0, + 1732.0, + 294.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1839.0, + 1402.0, + 1839.0, + 1402.0, + 1877.0, + 330.0, + 1877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1872.0, + 1078.0, + 1872.0, + 1078.0, + 1907.0, + 355.0, + 1907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1752.0, + 1405.0, + 1752.0, + 1405.0, + 1795.0, + 331.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1786.0, + 1229.0, + 1786.0, + 1229.0, + 1822.0, + 353.0, + 1822.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1468, + 1406, + 1468, + 1406, + 1805, + 297, + 1805 + ], + "score": 0.987 + }, + { + "category_id": 1, + "poly": [ + 297, + 689, + 1405, + 689, + 1405, + 997, + 297, + 997 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 300, + 1914, + 1405, + 1914, + 1405, + 2007, + 300, + 2007 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 286, + 509, + 1402, + 509, + 1402, + 573, + 286, + 573 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 299, + 379, + 1404, + 379, + 1404, + 443, + 299, + 443 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 365, + 1396, + 1325, + 1396, + 1325, + 1457, + 365, + 1457 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 297, + 1054, + 1401, + 1054, + 1401, + 1118, + 297, + 1118 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 295, + 1188, + 1403, + 1188, + 1403, + 1253, + 295, + 1253 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 389, + 584, + 1310, + 584, + 1310, + 672, + 389, + 672 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 672, + 1131, + 1029, + 1131, + 1029, + 1172, + 672, + 1172 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 296, + 258, + 1402, + 258, + 1402, + 321, + 296, + 321 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 467, + 1268, + 1228, + 1268, + 1228, + 1336, + 467, + 1336 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 296, + 1350, + 949, + 1350, + 949, + 1384, + 296, + 1384 + ], + "score": 0.939 + }, + { + "category_id": 8, + "poly": [ + 767, + 457, + 929, + 457, + 929, + 497, + 767, + 497 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 298, + 202, + 552, + 202, + 552, + 236, + 298, + 236 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 301, + 333, + 1141, + 333, + 1141, + 368, + 301, + 368 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 300, + 1845, + 723, + 1845, + 723, + 1885, + 300, + 1885 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 300, + 1009, + 1198, + 1009, + 1198, + 1043, + 300, + 1043 + ], + "score": 0.916 + }, + { + "category_id": 9, + "poly": [ + 1366, + 464, + 1400, + 464, + 1400, + 493, + 1366, + 493 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1139, + 1400, + 1139, + 1400, + 1168, + 1366, + 1168 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1283, + 1400, + 1283, + 1400, + 1313, + 1366, + 1313 + ], + "score": 0.857 + }, + { + "category_id": 9, + "poly": [ + 1366, + 604, + 1400, + 604, + 1400, + 635, + 1366, + 635 + ], + "score": 0.846 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.771 + }, + { + "category_id": 13, + "poly": [ + 1067, + 1561, + 1147, + 1561, + 1147, + 1594, + 1067, + 1594 + ], + "score": 0.93, + "latex": "\\mathcal { L } _ { 2 } ( \\mathbf { X } )" + }, + { + "category_id": 14, + "poly": [ + 385, + 582, + 1309, + 582, + 1309, + 673, + 385, + 673 + ], + "score": 0.93, + "latex": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\mathcal { L } _ { 1 } ( \\mathbf { X } ) : = \\frac { 1 } { 2 } \\mathrm { t r } \\Big ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } \\Big ) = \\frac { 1 } { 2 } \\sum _ { ( v _ { i } , v _ { j } ) \\in \\mathcal { E } } \\| \\frac { \\mathbf { X } _ { i } } { \\sqrt { d _ { i } + 1 } } - \\frac { \\mathbf { X } _ { j } } { \\sqrt { d _ { j } + 1 } } \\| _ { 2 } ^ { 2 } ," + }, + { + "category_id": 14, + "poly": [ + 367, + 1394, + 1332, + 1394, + 1332, + 1459, + 367, + 1459 + ], + "score": 0.93, + "latex": "\\mathbf { X } ^ { k + 1 } = \\mathbf { X } ^ { k } - ( 1 - \\alpha ) \\bigg ( \\frac { \\alpha } { 1 - \\alpha } ( \\mathbf { X } ^ { k } - \\mathbf { X _ { \\mathrm { i n } } } ) + ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ^ { k } \\bigg ) = ( 1 - \\alpha ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } + \\alpha \\mathbf { X _ { \\mathrm { i n } } } ." + }, + { + "category_id": 14, + "poly": [ + 766, + 453, + 931, + 453, + 931, + 496, + 766, + 496 + ], + "score": 0.93, + "latex": "\\mathbf { X } _ { \\mathrm { o u t } } = \\tilde { \\mathbf { A } } \\mathbf { X } _ { \\mathrm { i n } } ," + }, + { + "category_id": 14, + "poly": [ + 669, + 1129, + 1027, + 1129, + 1027, + 1172, + 669, + 1172 + ], + "score": 0.92, + "latex": "\\mathbf { X } ^ { k + 1 } = ( 1 - \\alpha ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } + \\alpha \\mathbf { X _ { \\mathrm { i n } } } ," + }, + { + "category_id": 13, + "poly": [ + 1020, + 687, + 1398, + 687, + 1398, + 726, + 1020, + 726 + ], + "score": 0.92, + "latex": "\\mathbf { X } _ { \\mathrm { o u t } } = \\mathbf { X } _ { \\mathrm { i n } } - ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } _ { \\mathrm { i n } } = \\tilde { \\mathbf { A } } \\mathbf { X } _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 499, + 1351, + 604, + 1351, + 604, + 1383, + 499, + 1383 + ], + "score": 0.92, + "latex": "\\mathbf { X } = \\mathbf { X } _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 749, + 1352, + 871, + 1352, + 871, + 1384, + 749, + 1384 + ], + "score": 0.92, + "latex": "\\gamma = 1 - \\alpha" + }, + { + "category_id": 13, + "poly": [ + 369, + 1188, + 485, + 1188, + 485, + 1220, + 369, + 1220 + ], + "score": 0.92, + "latex": "{ \\bf X } ^ { 0 } = { \\bf X } _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 547, + 723, + 655, + 723, + 655, + 753, + 547, + 753 + ], + "score": 0.92, + "latex": "\\mathbf { X } = \\mathbf { X } _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 804, + 724, + 877, + 724, + 877, + 754, + 804, + 754 + ], + "score": 0.91, + "latex": "\\gamma = 1" + }, + { + "category_id": 14, + "poly": [ + 465, + 1263, + 1230, + 1263, + 1230, + 1339, + 465, + 1339 + ], + "score": 0.91, + "latex": "\\underset { { \\substack { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } _ { 2 } ( { \\mathbf { X } } ) : = \\frac { \\alpha } { 2 ( 1 - \\alpha ) } \\| \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } } \\| _ { F } ^ { 2 } + \\frac { 1 } { 2 } \\mathrm { t r } \\Big ( \\mathbf { X } ^ { \\top } ( { \\mathbf { I } } - \\tilde { \\mathbf { A } } ) \\mathbf { X } \\Big ) ," + }, + { + "category_id": 13, + "poly": [ + 463, + 511, + 517, + 511, + 517, + 541, + 463, + 541 + ], + "score": 0.9, + "latex": "\\mathbf { X _ { o u t } }" + }, + { + "category_id": 13, + "poly": [ + 369, + 511, + 414, + 511, + 414, + 540, + 369, + 540 + ], + "score": 0.89, + "latex": "{ \\bf X } _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 368, + 694, + 395, + 694, + 395, + 723, + 368, + 723 + ], + "score": 0.88, + "latex": "d _ { i }" + }, + { + "category_id": 13, + "poly": [ + 682, + 697, + 707, + 697, + 707, + 723, + 682, + 723 + ], + "score": 0.85, + "latex": "v _ { i }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 196.0, + 556.0, + 196.0, + 556.0, + 243.0, + 293.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 327.0, + 1140.0, + 327.0, + 1140.0, + 379.0, + 294.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1842.0, + 728.0, + 1842.0, + 728.0, + 1893.0, + 290.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1005.0, + 1198.0, + 1005.0, + 1198.0, + 1050.0, + 295.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1470.0, + 1404.0, + 1470.0, + 1404.0, + 1505.0, + 296.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1406.0, + 1500.0, + 1406.0, + 1535.0, + 294.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 1403.0, + 1528.0, + 1403.0, + 1567.0, + 294.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1557.0, + 1066.0, + 1557.0, + 1066.0, + 1598.0, + 292.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1557.0, + 1406.0, + 1557.0, + 1406.0, + 1598.0, + 1148.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1588.0, + 1404.0, + 1588.0, + 1404.0, + 1627.0, + 292.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1620.0, + 1406.0, + 1620.0, + 1406.0, + 1658.0, + 294.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1649.0, + 1406.0, + 1649.0, + 1406.0, + 1687.0, + 292.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1682.0, + 1406.0, + 1682.0, + 1406.0, + 1715.0, + 292.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 1407.0, + 1712.0, + 1407.0, + 1747.0, + 295.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 1406.0, + 1741.0, + 1406.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1772.0, + 1078.0, + 1772.0, + 1078.0, + 1808.0, + 295.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 689.0, + 367.0, + 689.0, + 367.0, + 728.0, + 292.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 689.0, + 681.0, + 689.0, + 681.0, + 728.0, + 396.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 689.0, + 1019.0, + 689.0, + 1019.0, + 728.0, + 708.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 689.0, + 1406.0, + 689.0, + 1406.0, + 728.0, + 1399.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 721.0, + 546.0, + 721.0, + 546.0, + 759.0, + 292.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 721.0, + 803.0, + 721.0, + 803.0, + 759.0, + 656.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 721.0, + 1405.0, + 721.0, + 1405.0, + 759.0, + 878.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 750.0, + 1405.0, + 750.0, + 1405.0, + 788.0, + 294.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 783.0, + 1405.0, + 783.0, + 1405.0, + 819.0, + 295.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 811.0, + 1403.0, + 811.0, + 1403.0, + 847.0, + 294.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 842.0, + 1405.0, + 842.0, + 1405.0, + 879.0, + 294.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 872.0, + 1405.0, + 872.0, + 1405.0, + 911.0, + 292.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 900.0, + 1406.0, + 900.0, + 1406.0, + 942.0, + 292.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 935.0, + 1405.0, + 935.0, + 1405.0, + 971.0, + 295.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 963.0, + 1083.0, + 963.0, + 1083.0, + 1001.0, + 294.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 1407.0, + 1913.0, + 1407.0, + 1951.0, + 294.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1945.0, + 1407.0, + 1945.0, + 1407.0, + 1979.0, + 296.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1977.0, + 1397.0, + 1977.0, + 1397.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 506.0, + 368.0, + 506.0, + 368.0, + 545.0, + 296.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 506.0, + 462.0, + 506.0, + 462.0, + 545.0, + 415.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 506.0, + 1404.0, + 506.0, + 1404.0, + 545.0, + 518.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 540.0, + 1246.0, + 540.0, + 1246.0, + 576.0, + 294.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 375.0, + 1405.0, + 375.0, + 1405.0, + 419.0, + 293.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 408.0, + 910.0, + 408.0, + 910.0, + 447.0, + 293.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1054.0, + 1405.0, + 1054.0, + 1405.0, + 1090.0, + 295.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1087.0, + 608.0, + 1087.0, + 608.0, + 1122.0, + 293.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1180.0, + 368.0, + 1180.0, + 368.0, + 1230.0, + 290.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1180.0, + 1407.0, + 1180.0, + 1407.0, + 1230.0, + 486.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1218.0, + 453.0, + 1218.0, + 453.0, + 1256.0, + 291.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 258.0, + 1404.0, + 258.0, + 1404.0, + 293.0, + 295.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 287.0, + 815.0, + 287.0, + 815.0, + 325.0, + 295.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1347.0, + 498.0, + 1347.0, + 498.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 1347.0, + 748.0, + 1347.0, + 748.0, + 1388.0, + 605.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1347.0, + 950.0, + 1347.0, + 950.0, + 1388.0, + 872.0, + 1388.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 256, + 1405, + 256, + 1405, + 565, + 297, + 565 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 835, + 1404, + 835, + 1404, + 960, + 297, + 960 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 574, + 1404, + 574, + 1404, + 732, + 297, + 732 + ], + "score": 0.98 + }, + { + "category_id": 3, + "poly": [ + 411, + 1324, + 1286, + 1324, + 1286, + 1482, + 411, + 1482 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 634, + 743, + 1064, + 743, + 1064, + 827, + 634, + 827 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 287, + 1572, + 1402, + 1572, + 1402, + 1636, + 287, + 1636 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 298, + 1771, + 1405, + 1771, + 1405, + 1836, + 298, + 1836 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 292, + 1052, + 1402, + 1052, + 1402, + 1115, + 292, + 1115 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 584, + 970, + 1115, + 970, + 1115, + 1036, + 584, + 1036 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 536, + 1841, + 1164, + 1841, + 1164, + 1956, + 536, + 1956 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 489, + 1125, + 1207, + 1125, + 1207, + 1185, + 489, + 1185 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 400, + 1694, + 1021, + 1694, + 1021, + 1759, + 400, + 1759 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 297, + 1262, + 676, + 1262, + 676, + 1296, + 297, + 1296 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 298, + 202, + 581, + 202, + 581, + 235, + 298, + 235 + ], + "score": 0.931 + }, + { + "category_id": 4, + "poly": [ + 577, + 1507, + 1123, + 1507, + 1123, + 1543, + 577, + 1543 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 298, + 1195, + 715, + 1195, + 715, + 1230, + 298, + 1230 + ], + "score": 0.917 + }, + { + "category_id": 8, + "poly": [ + 423, + 1645, + 1300, + 1645, + 1300, + 1687, + 423, + 1687 + ], + "score": 0.911 + }, + { + "category_id": 9, + "poly": [ + 1366, + 767, + 1400, + 767, + 1400, + 797, + 1366, + 797 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1366, + 987, + 1400, + 987, + 1400, + 1018, + 1366, + 1018 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1967, + 1401, + 1967, + 1401, + 1998, + 1351, + 1998 + ], + "score": 0.884 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1709, + 1400, + 1709, + 1400, + 1739, + 1351, + 1739 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1132, + 1401, + 1132, + 1401, + 1161, + 1365, + 1161 + ], + "score": 0.864 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1653, + 1400, + 1653, + 1400, + 1682, + 1365, + 1682 + ], + "score": 0.86 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1918, + 1401, + 1918, + 1401, + 1950, + 1351, + 1950 + ], + "score": 0.847 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2085, + 840, + 2085 + ], + "score": 0.768 + }, + { + "category_id": 8, + "poly": [ + 535, + 1961, + 783, + 1961, + 783, + 2000, + 535, + 2000 + ], + "score": 0.725 + }, + { + "category_id": 14, + "poly": [ + 634, + 738, + 1063, + 738, + 1063, + 829, + 634, + 829 + ], + "score": 0.95, + "latex": "\\| { \\bf X } - { \\bf X } _ { \\mathrm { i n } } \\| _ { 2 1 } : = \\sum _ { i = 1 } ^ { n } \\| { \\bf X } _ { i } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } ," + }, + { + "category_id": 14, + "poly": [ + 582, + 966, + 1114, + 966, + 1114, + 1042, + 582, + 1042 + ], + "score": 0.94, + "latex": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\lambda \\| \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } } \\| _ { 2 1 } + \\frac { 1 } { 2 } \\mathrm { t r } ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ) ," + }, + { + "category_id": 14, + "poly": [ + 533, + 1836, + 1166, + 1836, + 1166, + 2005, + 533, + 2005 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } { { \\mathbf { Z } ^ { k + 1 } = \\arg \\operatorname* { m i n } \\{ \\lambda \\| \\mathbf { Z } \\| _ { 2 1 } + \\frac { 1 } { 2 \\gamma } \\| \\mathbf { Z } - ( \\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } } ) \\| _ { F } ^ { 2 } \\} } } \\\\ & { = \\mathbf { p r o x } _ { \\gamma \\lambda \\| \\cdot \\| _ { 2 1 } } ( \\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } } ) } \\\\ & { \\mathbf { X } ^ { k + 1 } = \\mathbf { X } _ { \\mathrm { i n } } + \\mathbf { Z } ^ { k + 1 } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 372, + 1052, + 496, + 1052, + 496, + 1085, + 372, + 1085 + ], + "score": 0.93, + "latex": "\\lambda \\in [ 0 , \\infty )" + }, + { + "category_id": 13, + "poly": [ + 373, + 1772, + 490, + 1772, + 490, + 1804, + 373, + 1804 + ], + "score": 0.92, + "latex": "{ \\bf X } ^ { 0 } = { \\bf X } _ { \\mathrm { i n } }" + }, + { + "category_id": 14, + "poly": [ + 399, + 1642, + 1300, + 1642, + 1300, + 1765, + 399, + 1765 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } { { \\displaystyle { \\bf Y } ^ { k } = { \\bf X } ^ { k } - 2 \\gamma ( 1 - \\lambda ) ( { \\bf I } - \\tilde { \\bf A } ) { \\bf X } ^ { k } = \\left( 1 - 2 \\gamma ( 1 - \\lambda ) \\right) { \\bf X } ^ { k } + 2 \\gamma ( 1 - \\lambda ) \\tilde { \\bf A } { \\bf X } ^ { k } } } & { { } { } { } } \\\\ { { \\displaystyle { \\bf X } ^ { k + 1 } = \\arg \\operatorname* { m i n } \\left\\{ \\lambda \\| { \\bf X } - { \\bf X } _ { \\mathrm { i n } } \\| _ { 2 1 } + \\frac { 1 } { 2 \\gamma } \\| { \\bf X } - { \\bf Y } ^ { k } \\| _ { F } ^ { 2 } \\right\\} } } & { { } { } { } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 371, + 1196, + 479, + 1196, + 479, + 1230, + 371, + 1230 + ], + "score": 0.92, + "latex": "\\lambda \\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1009, + 1774, + 1167, + 1774, + 1167, + 1804, + 1009, + 1804 + ], + "score": 0.91, + "latex": "\\mathbf { Z } = \\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1573, + 365, + 1573, + 365, + 1607, + 298, + 1607 + ], + "score": 0.91, + "latex": "\\mathcal { L } ( \\mathbf { X } )" + }, + { + "category_id": 13, + "poly": [ + 686, + 868, + 788, + 868, + 788, + 898, + 686, + 898 + ], + "score": 0.91, + "latex": "\\mathbf { X } - \\mathbf { X } _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 979, + 636, + 1398, + 636, + 1398, + 672, + 979, + 672 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\| \\dot { \\bf X } - \\dot { \\bf X } _ { \\mathrm { i n } } \\| _ { F } ^ { 2 } = \\sum _ { i = 1 } ^ { n } \\| \\dot { \\bf X } _ { i } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } ^ { 2 } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 486, + 1124, + 1205, + 1124, + 1205, + 1186, + 486, + 1186 + ], + "score": 0.9, + "latex": "\\underset { \\mathbf { X } \\in \\mathbb { R } ^ { n \\times d } } { \\arg \\operatorname* { m i n } } \\ \\mathcal { L } ( \\mathbf { X } ) : = \\lambda \\| \\mathbf { X } - \\mathbf { X _ { \\mathrm { i n } } } \\| _ { 2 1 } + ( 1 - \\lambda ) \\mathrm { t r } ( \\mathbf { X } ^ { \\top } ( \\mathbf { I } - \\tilde { \\mathbf { A } } ) \\mathbf { X } ) ," + }, + { + "category_id": 13, + "poly": [ + 1089, + 929, + 1127, + 929, + 1127, + 959, + 1089, + 959 + ], + "score": 0.89, + "latex": "\\ell _ { 2 1 }" + }, + { + "category_id": 13, + "poly": [ + 1238, + 838, + 1276, + 838, + 1276, + 868, + 1238, + 868 + ], + "score": 0.88, + "latex": "\\ell _ { 2 1 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 868, + 335, + 868, + 335, + 898, + 298, + 898 + ], + "score": 0.88, + "latex": "\\ell _ { 2 1 }" + }, + { + "category_id": 13, + "poly": [ + 1036, + 838, + 1064, + 838, + 1064, + 867, + 1036, + 867 + ], + "score": 0.87, + "latex": "\\ell _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 801, + 837, + 828, + 837, + 828, + 867, + 801, + 867 + ], + "score": 0.87, + "latex": "\\ell _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 541, + 1778, + 562, + 1778, + 562, + 1805, + 541, + 1805 + ], + "score": 0.83, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 626, + 1084, + 644, + 1084, + 644, + 1110, + 626, + 1110 + ], + "score": 0.81, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 1227, + 384, + 1249, + 384, + 1249, + 407, + 1227, + 407 + ], + "score": 0.73, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1394.0, + 481.0, + 1394.0, + 481.0, + 1416.0, + 446.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1388.0, + 684.0, + 1388.0, + 684.0, + 1430.0, + 549.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1380.0, + 956.0, + 1380.0, + 956.0, + 1428.0, + 899.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1386.0, + 1127.0, + 1386.0, + 1127.0, + 1421.0, + 1034.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 1349.5, + 805.0, + 1349.5, + 805.0, + 1376.0, + 759.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 1382.0, + 1276.0, + 1382.0, + 1276.0, + 1420.5, + 1195.0, + 1420.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.25, + 1424.5, + 801.25, + 1424.5, + 801.25, + 1463.0, + 754.25, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1256.0, + 679.0, + 1256.0, + 679.0, + 1305.0, + 291.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 197.0, + 584.0, + 197.0, + 584.0, + 243.0, + 292.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 1503.0, + 1125.0, + 1503.0, + 1125.0, + 1548.0, + 574.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 861.0, + 2059.0, + 861.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 257.0, + 1406.0, + 257.0, + 1406.0, + 293.0, + 295.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 322.0, + 294.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 318.0, + 1406.0, + 318.0, + 1406.0, + 354.0, + 292.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 350.0, + 1405.0, + 350.0, + 1405.0, + 382.0, + 296.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 379.0, + 1226.0, + 379.0, + 1226.0, + 417.0, + 291.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 379.0, + 1406.0, + 379.0, + 1406.0, + 417.0, + 1250.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 412.0, + 1403.0, + 412.0, + 1403.0, + 444.0, + 295.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 439.0, + 1403.0, + 439.0, + 1403.0, + 476.0, + 294.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 472.0, + 1403.0, + 472.0, + 1403.0, + 504.0, + 296.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 499.0, + 1405.0, + 499.0, + 1405.0, + 538.0, + 294.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 531.0, + 1173.0, + 531.0, + 1173.0, + 565.0, + 292.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 835.0, + 800.0, + 835.0, + 800.0, + 872.0, + 294.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 835.0, + 1035.0, + 835.0, + 1035.0, + 872.0, + 829.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 835.0, + 1237.0, + 835.0, + 1237.0, + 872.0, + 1065.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 835.0, + 1405.0, + 835.0, + 1405.0, + 872.0, + 1277.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 868.0, + 685.0, + 868.0, + 685.0, + 901.0, + 336.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 868.0, + 1404.0, + 868.0, + 1404.0, + 901.0, + 789.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 897.0, + 1409.0, + 897.0, + 1409.0, + 934.0, + 295.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 928.0, + 1088.0, + 928.0, + 1088.0, + 962.0, + 292.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 928.0, + 1407.0, + 928.0, + 1407.0, + 962.0, + 1128.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 576.0, + 1405.0, + 576.0, + 1405.0, + 610.0, + 296.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 606.0, + 1407.0, + 606.0, + 1407.0, + 644.0, + 295.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 626.0, + 978.0, + 626.0, + 978.0, + 681.0, + 288.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 626.0, + 1411.0, + 626.0, + 1411.0, + 681.0, + 1399.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 668.0, + 1405.0, + 668.0, + 1405.0, + 701.0, + 296.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 695.0, + 1405.0, + 695.0, + 1405.0, + 738.0, + 292.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1572.0, + 1404.0, + 1572.0, + 1404.0, + 1608.0, + 366.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1601.0, + 1380.0, + 1601.0, + 1380.0, + 1640.0, + 291.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1767.0, + 372.0, + 1767.0, + 372.0, + 1810.0, + 292.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1767.0, + 540.0, + 1767.0, + 540.0, + 1810.0, + 491.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 1767.0, + 1008.0, + 1767.0, + 1008.0, + 1810.0, + 563.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1767.0, + 1405.0, + 1767.0, + 1405.0, + 1810.0, + 1168.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1803.0, + 443.0, + 1803.0, + 443.0, + 1837.0, + 295.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1047.0, + 371.0, + 1047.0, + 371.0, + 1091.0, + 293.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1047.0, + 1408.0, + 1047.0, + 1408.0, + 1091.0, + 497.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1082.0, + 625.0, + 1082.0, + 625.0, + 1118.0, + 295.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1082.0, + 1049.0, + 1082.0, + 1049.0, + 1118.0, + 645.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1192.0, + 370.0, + 1192.0, + 370.0, + 1232.0, + 296.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1192.0, + 718.0, + 1192.0, + 718.0, + 1232.0, + 480.0, + 1232.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1355, + 1405, + 1355, + 1405, + 1537, + 296, + 1537 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 313, + 1876, + 1405, + 1876, + 1405, + 2009, + 313, + 2009 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 295, + 1151, + 1407, + 1151, + 1407, + 1350, + 295, + 1350 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 299, + 454, + 1405, + 454, + 1405, + 559, + 299, + 559 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 296, + 1792, + 1404, + 1792, + 1404, + 1857, + 296, + 1857 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 293, + 327, + 1413, + 327, + 1413, + 390, + 293, + 390 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 332, + 582, + 1407, + 582, + 1407, + 796, + 332, + 796 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 297, + 1544, + 1405, + 1544, + 1405, + 1704, + 297, + 1704 + ], + "score": 0.928 + }, + { + "category_id": 8, + "poly": [ + 395, + 246, + 1304, + 246, + 1304, + 318, + 395, + 318 + ], + "score": 0.923 + }, + { + "category_id": 4, + "poly": [ + 603, + 1088, + 1098, + 1088, + 1098, + 1123, + 603, + 1123 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 295, + 201, + 1175, + 201, + 1175, + 237, + 295, + 237 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 299, + 1737, + 632, + 1737, + 632, + 1771, + 299, + 1771 + ], + "score": 0.91 + }, + { + "category_id": 8, + "poly": [ + 426, + 400, + 1275, + 400, + 1275, + 444, + 426, + 444 + ], + "score": 0.908 + }, + { + "category_id": 9, + "poly": [ + 1352, + 265, + 1400, + 265, + 1400, + 295, + 1352, + 295 + ], + "score": 0.856 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 859, + 2062, + 859, + 2085, + 840, + 2085 + ], + "score": 0.789 + }, + { + "category_id": 9, + "poly": [ + 1352, + 407, + 1400, + 407, + 1400, + 438, + 1352, + 438 + ], + "score": 0.77 + }, + { + "category_id": 3, + "poly": [ + 339, + 848, + 1157, + 848, + 1157, + 1071, + 339, + 1071 + ], + "score": 0.393 + }, + { + "category_id": 3, + "poly": [ + 380, + 859, + 1158, + 859, + 1158, + 1071, + 380, + 1071 + ], + "score": 0.352 + }, + { + "category_id": 9, + "poly": [ + 1352, + 407, + 1400, + 407, + 1400, + 437, + 1352, + 437 + ], + "score": 0.208 + }, + { + "category_id": 13, + "poly": [ + 381, + 1498, + 512, + 1498, + 512, + 1542, + 381, + 1542 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1076, + 1428, + 1211, + 1428, + 1211, + 1471, + 1076, + 1471 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\gamma < \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 546, + 868, + 1154, + 868, + 1154, + 1040, + 546, + 1040 + ], + "score": 0.94, + "latex": "\\left\\{ \\begin{array} { l l } { \\mathbf { Y } ^ { k } \\ = \\ \\big ( 1 - 2 \\gamma ( 1 - \\lambda ) \\big ) \\mathbf { X } ^ { k } + 2 \\gamma ( 1 - \\lambda ) \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } } \\\\ { \\beta _ { i } \\ = \\ \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\| \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } } , 0 ) \\quad \\forall i \\in [ n ] } \\\\ { \\mathbf { X } _ { i } ^ { k + 1 } = \\ ( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } \\mathbf { Y } _ { i } ^ { k } \\qquad \\forall i \\in [ n ] } \\end{array} \\right." + }, + { + "category_id": 13, + "poly": [ + 817, + 1221, + 1062, + 1221, + 1062, + 1263, + 817, + 1263 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\mathbf { Y } ^ { k } = \\frac { 1 } { 2 } \\mathbf { X } ^ { k } + \\frac { 1 } { 2 } \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 981, + 1354, + 1148, + 1354, + 1148, + 1402, + 981, + 1402 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\gamma < \\frac { 1 } { ( 1 - \\lambda ) \\| \\tilde { \\mathbf { L } } \\| _ { 2 } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 396, + 243, + 1305, + 243, + 1305, + 318, + 396, + 318 + ], + "score": 0.93, + "latex": "\\left( \\mathbf { p r o x } _ { \\gamma \\lambda \\parallel \\cdot \\parallel _ { 2 1 } } ( \\mathbf { X } ) \\right) _ { i } = \\frac { \\mathbf { X } _ { i } } { \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } } \\operatorname* { m a x } ( \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } - \\gamma \\lambda , 0 ) = \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\parallel \\mathbf { X } _ { i } \\parallel _ { 2 } } , 0 ) \\cdot \\mathbf { X } _ { i } ." + }, + { + "category_id": 13, + "poly": [ + 369, + 1611, + 445, + 1611, + 445, + 1643, + 369, + 1643 + ], + "score": 0.93, + "latex": "\\mathcal { O } ( n d )" + }, + { + "category_id": 13, + "poly": [ + 1265, + 1182, + 1396, + 1182, + 1396, + 1225, + 1265, + 1225 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 857, + 1909, + 1039, + 1909, + 1039, + 1946, + 857, + 1946 + ], + "score": 0.93, + "latex": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 375, + 454, + 750, + 454, + 750, + 502, + 375, + 502 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\beta _ { i } : = \\operatorname* { m a x } ( 1 - \\frac { \\gamma \\lambda } { \\| \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 } } , 0 ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 392, + 615, + 557, + 615, + 557, + 649, + 392, + 649 + ], + "score": 0.92, + "latex": "1 - 2 \\gamma ( 1 - \\lambda )" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1184, + 1232, + 1184, + 1232, + 1223, + 1100, + 1223 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\gamma = \\frac { 1 } { 4 ( 1 - \\lambda ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1283, + 1426, + 1398, + 1426, + 1398, + 1465, + 1283, + 1465 + ], + "score": 0.92, + "latex": "\\Vert \\tilde { \\mathbf { L } } \\Vert _ { 2 } \\leq 2" + }, + { + "category_id": 13, + "poly": [ + 746, + 655, + 861, + 655, + 861, + 689, + 746, + 689 + ], + "score": 0.92, + "latex": "\\beta _ { i } \\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1230, + 1943, + 1396, + 1943, + 1396, + 1976, + 1230, + 1976 + ], + "score": 0.92, + "latex": "( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }" + }, + { + "category_id": 13, + "poly": [ + 582, + 1496, + 693, + 1496, + 693, + 1535, + 582, + 1535 + ], + "score": 0.92, + "latex": "\\Vert \\tilde { \\mathbf { L } } \\Vert _ { 2 } < 2" + }, + { + "category_id": 13, + "poly": [ + 636, + 1973, + 678, + 1973, + 678, + 2009, + 636, + 2009 + ], + "score": 0.92, + "latex": "\\mathbf { Y } _ { i } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 701, + 1641, + 786, + 1641, + 786, + 1674, + 701, + 1674 + ], + "score": 0.92, + "latex": "\\mathcal { O } ( m d )" + }, + { + "category_id": 13, + "poly": [ + 660, + 1879, + 732, + 1879, + 732, + 1912, + 660, + 1912 + ], + "score": 0.91, + "latex": "( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }" + }, + { + "category_id": 13, + "poly": [ + 544, + 1910, + 585, + 1910, + 585, + 1946, + 544, + 1946 + ], + "score": 0.91, + "latex": "\\mathbf { Y } _ { i } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 977, + 765, + 1052, + 765, + 1052, + 794, + 977, + 794 + ], + "score": 0.91, + "latex": "1 - \\beta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 296, + 356, + 412, + 356, + 412, + 388, + 296, + 388 + ], + "score": 0.91, + "latex": "\\mathbf { Y } ^ { k } - \\mathbf { X } _ { \\mathrm { i n } }" + }, + { + "category_id": 13, + "poly": [ + 358, + 686, + 544, + 686, + 544, + 724, + 358, + 724 + ], + "score": 0.9, + "latex": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1114, + 1223, + 1253, + 1223, + 1253, + 1256, + 1114, + 1256 + ], + "score": 0.9, + "latex": "\\mathbf { Y } ^ { k } = \\tilde { \\mathbf { A } } \\mathbf { X } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1576, + 365, + 1576, + 365, + 1611, + 298, + 1611 + ], + "score": 0.89, + "latex": "\\tilde { \\mathbf { A } } \\mathbf { X } _ { i n }" + }, + { + "category_id": 13, + "poly": [ + 1000, + 734, + 1044, + 734, + 1044, + 763, + 1000, + 763 + ], + "score": 0.89, + "latex": "\\mathbf { X } _ { \\mathrm { i n } }" + }, + { + "category_id": 14, + "poly": [ + 428, + 398, + 1270, + 398, + 1270, + 442, + 428, + 442 + ], + "score": 0.88, + "latex": "\\mathbf { X } _ { i } ^ { k + 1 } = ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } ( \\mathbf { Y } _ { i } ^ { k } - ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } ) = ( 1 - \\beta _ { i } ) ( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i } + \\beta _ { i } \\mathbf { Y } _ { i } ^ { k } , \\quad \\forall i \\in [ n ] ," + }, + { + "category_id": 13, + "poly": [ + 1356, + 730, + 1396, + 730, + 1396, + 761, + 1356, + 761 + ], + "score": 0.88, + "latex": "\\mathbf { Y } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 1316, + 1547, + 1404, + 1547, + 1404, + 1578, + 1316, + 1578 + ], + "score": 0.87, + "latex": "\\mathbf { X } _ { o u t } =" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1913, + 1400, + 1913, + 1400, + 1944, + 1373, + 1944 + ], + "score": 0.87, + "latex": "\\beta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1312, + 1675, + 1398, + 1675, + 1398, + 1701, + 1312, + 1701 + ], + "score": 0.85, + "latex": "m \\gg n" + }, + { + "category_id": 13, + "poly": [ + 1213, + 768, + 1238, + 768, + 1238, + 793, + 1213, + 793 + ], + "score": 0.85, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1018, + 660, + 1043, + 660, + 1043, + 686, + 1018, + 686 + ], + "score": 0.84, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 575, + 1883, + 600, + 1883, + 600, + 1909, + 575, + 1909 + ], + "score": 0.83, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1179, + 1160, + 1200, + 1160, + 1200, + 1184, + 1179, + 1184 + ], + "score": 0.79, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 347, + 207, + 359, + 207, + 359, + 231, + 347, + 231 + ], + "score": 0.78, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1187, + 328, + 1214, + 328, + 1214, + 355, + 1187, + 355 + ], + "score": 0.77, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 900, + 1467, + 921, + 1467, + 921, + 1497, + 900, + 1497 + ], + "score": 0.73, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 883, + 1613, + 901, + 1613, + 901, + 1638, + 883, + 1638 + ], + "score": 0.51, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 536, + 1617, + 556, + 1617, + 556, + 1638, + 536, + 1638 + ], + "score": 0.35, + "latex": "n" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1082.0, + 1099.0, + 1082.0, + 1099.0, + 1132.0, + 598.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1734.0, + 635.0, + 1734.0, + 635.0, + 1777.0, + 292.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2060.0, + 862.0, + 2060.0, + 862.0, + 2091.0, + 840.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 926.0, + 1006.0, + 926.0, + 1006.0, + 1000.0, + 575.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 926.0, + 1003.0, + 926.0, + 1003.0, + 1001.0, + 575.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1350.0, + 980.0, + 1350.0, + 980.0, + 1401.0, + 293.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1394.0, + 1406.0, + 1394.0, + 1406.0, + 1432.0, + 291.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1423.0, + 1075.0, + 1423.0, + 1075.0, + 1479.0, + 285.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1423.0, + 1282.0, + 1423.0, + 1282.0, + 1479.0, + 1212.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1423.0, + 1415.0, + 1423.0, + 1415.0, + 1479.0, + 1399.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1464.0, + 899.0, + 1464.0, + 899.0, + 1501.0, + 293.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1464.0, + 1405.0, + 1464.0, + 1405.0, + 1501.0, + 922.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1495.0, + 380.0, + 1495.0, + 380.0, + 1545.0, + 289.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 1495.0, + 581.0, + 1495.0, + 581.0, + 1545.0, + 513.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1495.0, + 708.0, + 1495.0, + 708.0, + 1545.0, + 694.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1349.0, + 1411.0, + 1349.0, + 1411.0, + 1405.5, + 1021.0, + 1405.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1873.0, + 574.0, + 1873.0, + 574.0, + 1913.0, + 309.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1873.0, + 659.0, + 1873.0, + 659.0, + 1913.0, + 601.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1873.0, + 1405.0, + 1873.0, + 1405.0, + 1913.0, + 733.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1906.0, + 543.0, + 1906.0, + 543.0, + 1950.0, + 330.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1906.0, + 856.0, + 1906.0, + 856.0, + 1950.0, + 586.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1906.0, + 1372.0, + 1906.0, + 1372.0, + 1950.0, + 1040.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1906.0, + 1407.0, + 1906.0, + 1407.0, + 1950.0, + 1401.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1941.0, + 1229.0, + 1941.0, + 1229.0, + 1979.0, + 335.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1941.0, + 1408.0, + 1941.0, + 1408.0, + 1979.0, + 1397.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1973.0, + 635.0, + 1973.0, + 635.0, + 2010.0, + 335.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 1973.0, + 849.0, + 1973.0, + 849.0, + 2010.0, + 679.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1150.0, + 1178.0, + 1150.0, + 1178.0, + 1192.0, + 293.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1150.0, + 1406.0, + 1150.0, + 1406.0, + 1192.0, + 1201.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1218.0, + 816.0, + 1218.0, + 816.0, + 1265.0, + 290.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1218.0, + 1113.0, + 1218.0, + 1113.0, + 1265.0, + 1063.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1218.0, + 1410.0, + 1218.0, + 1410.0, + 1265.0, + 1254.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1293.0, + 293.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1284.0, + 1407.0, + 1284.0, + 1407.0, + 1325.0, + 293.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1317.0, + 644.0, + 1317.0, + 644.0, + 1351.0, + 296.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1179.5, + 1411.0, + 1179.5, + 1411.0, + 1226.5, + 291.0, + 1226.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 446.0, + 374.0, + 446.0, + 374.0, + 514.0, + 286.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 446.0, + 1412.0, + 446.0, + 1412.0, + 514.0, + 751.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 495.0, + 1405.0, + 495.0, + 1405.0, + 532.0, + 293.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 523.0, + 398.0, + 523.0, + 398.0, + 563.0, + 292.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 1406.0, + 1790.0, + 1406.0, + 1831.0, + 294.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1823.0, + 869.0, + 1823.0, + 869.0, + 1861.0, + 294.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 327.0, + 1186.0, + 327.0, + 1186.0, + 362.0, + 294.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 327.0, + 1403.0, + 327.0, + 1403.0, + 362.0, + 1215.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 351.0, + 1096.0, + 351.0, + 1096.0, + 397.0, + 413.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 577.0, + 1407.0, + 577.0, + 1407.0, + 623.0, + 327.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 610.0, + 391.0, + 610.0, + 391.0, + 652.0, + 351.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 610.0, + 571.0, + 610.0, + 571.0, + 652.0, + 558.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 647.0, + 745.0, + 647.0, + 745.0, + 695.0, + 329.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 647.0, + 1017.0, + 647.0, + 1017.0, + 695.0, + 862.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 647.0, + 1406.0, + 647.0, + 1406.0, + 695.0, + 1044.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 681.0, + 357.0, + 681.0, + 357.0, + 727.0, + 352.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 681.0, + 556.0, + 681.0, + 556.0, + 727.0, + 545.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 726.0, + 999.0, + 726.0, + 999.0, + 768.0, + 327.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 726.0, + 1355.0, + 726.0, + 1355.0, + 768.0, + 1045.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 726.0, + 1404.0, + 726.0, + 1404.0, + 768.0, + 1397.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 760.0, + 976.0, + 760.0, + 976.0, + 798.0, + 353.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 760.0, + 1212.0, + 760.0, + 1212.0, + 798.0, + 1053.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 760.0, + 1250.0, + 760.0, + 1250.0, + 798.0, + 1239.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1543.0, + 1315.0, + 1543.0, + 1315.0, + 1583.0, + 294.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1617.0, + 366.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1608.0, + 368.0, + 1608.0, + 368.0, + 1647.0, + 294.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1608.0, + 535.0, + 1608.0, + 535.0, + 1647.0, + 446.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1608.0, + 882.0, + 1608.0, + 882.0, + 1647.0, + 557.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1608.0, + 1406.0, + 1608.0, + 1406.0, + 1647.0, + 902.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1638.0, + 700.0, + 1638.0, + 700.0, + 1678.0, + 292.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1638.0, + 1408.0, + 1638.0, + 1408.0, + 1678.0, + 787.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1669.0, + 1311.0, + 1669.0, + 1311.0, + 1708.0, + 292.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1669.0, + 1410.0, + 1669.0, + 1410.0, + 1708.0, + 1399.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 194.0, + 346.0, + 194.0, + 346.0, + 246.0, + 292.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 194.0, + 1176.0, + 194.0, + 1176.0, + 246.0, + 360.0, + 246.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1732, + 1405, + 1732, + 1405, + 2008, + 298, + 2008 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1133, + 1404, + 1133, + 1404, + 1382, + 298, + 1382 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 447, + 1404, + 447, + 1404, + 663, + 297, + 663 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1488, + 1405, + 1488, + 1405, + 1640, + 297, + 1640 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 950, + 1403, + 950, + 1403, + 1042, + 298, + 1042 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 767, + 1403, + 767, + 1403, + 860, + 298, + 860 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 667, + 1400, + 667, + 1400, + 762, + 297, + 762 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 299, + 893, + 649, + 893, + 649, + 926, + 299, + 926 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 299, + 1676, + 625, + 1676, + 625, + 1709, + 299, + 1709 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 298, + 1421, + 519, + 1421, + 519, + 1459, + 298, + 1459 + ], + "score": 0.915 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1055, + 1399, + 1055, + 1399, + 1084, + 1353, + 1084 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1093, + 1399, + 1093, + 1399, + 1122, + 1353, + 1122 + ], + "score": 0.867 + }, + { + "category_id": 8, + "poly": [ + 703, + 1090, + 995, + 1090, + 995, + 1126, + 703, + 1126 + ], + "score": 0.785 + }, + { + "category_id": 8, + "poly": [ + 707, + 1052, + 896, + 1052, + 896, + 1086, + 707, + 1086 + ], + "score": 0.73 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.699 + }, + { + "category_id": 1, + "poly": [ + 312, + 201, + 1402, + 201, + 1402, + 294, + 312, + 294 + ], + "score": 0.634 + }, + { + "category_id": 1, + "poly": [ + 313, + 301, + 1404, + 301, + 1404, + 426, + 313, + 426 + ], + "score": 0.625 + }, + { + "category_id": 8, + "poly": [ + 702, + 1051, + 997, + 1051, + 997, + 1126, + 702, + 1126 + ], + "score": 0.272 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2084, + 841, + 2084 + ], + "score": 0.127 + }, + { + "category_id": 13, + "poly": [ + 1040, + 1198, + 1142, + 1198, + 1142, + 1231, + 1040, + 1231 + ], + "score": 0.95, + "latex": "h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } )" + }, + { + "category_id": 13, + "poly": [ + 732, + 300, + 919, + 300, + 919, + 337, + 732, + 337 + ], + "score": 0.93, + "latex": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 916, + 700, + 1066, + 700, + 1066, + 732, + 916, + 732 + ], + "score": 0.93, + "latex": "\\{ v _ { 1 } , \\cdots , v _ { n } \\}" + }, + { + "category_id": 13, + "poly": [ + 1132, + 201, + 1317, + 201, + 1317, + 237, + 1132, + 237 + ], + "score": 0.92, + "latex": "\\| { \\bf Y } _ { i } ^ { k } - ( { \\bf X } _ { \\mathrm { i n } } ) _ { i } \\| _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 548, + 1165, + 706, + 1165, + 706, + 1198, + 548, + 1198 + ], + "score": 0.92, + "latex": "\\mathbf { X } _ { \\mathrm { f e a } } \\in \\mathbb { R } ^ { n \\times d }" + }, + { + "category_id": 13, + "poly": [ + 1079, + 798, + 1134, + 798, + 1134, + 831, + 1079, + 831 + ], + "score": 0.92, + "latex": "\\{ \\beta _ { i } \\}" + }, + { + "category_id": 13, + "poly": [ + 716, + 1197, + 895, + 1197, + 895, + 1231, + 716, + 1231 + ], + "score": 0.92, + "latex": "\\mathbf { X } _ { \\mathrm { i n } } = h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } )" + }, + { + "category_id": 13, + "poly": [ + 1233, + 1167, + 1294, + 1167, + 1294, + 1201, + 1233, + 1201 + ], + "score": 0.92, + "latex": "h _ { \\theta } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 549, + 203, + 621, + 203, + 621, + 236, + 549, + 236 + ], + "score": 0.91, + "latex": "( \\mathbf { X } _ { \\mathrm { i n } } ) _ { i }" + }, + { + "category_id": 13, + "poly": [ + 366, + 334, + 453, + 334, + 453, + 365, + 366, + 365 + ], + "score": 0.91, + "latex": "\\beta _ { i } \\to 0" + }, + { + "category_id": 13, + "poly": [ + 802, + 668, + 955, + 668, + 955, + 702, + 802, + 702 + ], + "score": 0.9, + "latex": "\\{ \\beta _ { 1 } , \\cdot \\cdot \\cdot , \\beta _ { n } \\}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1136, + 358, + 1136, + 358, + 1168, + 298, + 1168 + ], + "score": 0.9, + "latex": "h _ { \\theta } ( \\cdot )" + }, + { + "category_id": 14, + "poly": [ + 700, + 1048, + 999, + 1048, + 999, + 1130, + 700, + 1130 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } & { \\mathbf { X } _ { \\mathrm { i n } } = h _ { \\theta } ( \\mathbf { X } _ { \\mathrm { f e a } } ) , } \\\\ & { \\mathbf { Y } _ { \\mathrm { p r e } } = \\mathbf { A } \\mathbf { M } \\mathbf { P } \\left( \\mathbf { X } _ { \\mathrm { i n } } , K , \\lambda \\right) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1314, + 304, + 1404, + 304, + 1404, + 335, + 1314, + 335 + ], + "score": 0.9, + "latex": "\\beta _ { i } \\to 1" + }, + { + "category_id": 13, + "poly": [ + 596, + 237, + 624, + 237, + 624, + 265, + 596, + 265 + ], + "score": 0.89, + "latex": "\\beta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 346, + 829, + 388, + 829, + 388, + 859, + 346, + 859 + ], + "score": 0.86, + "latex": "{ \\bf X } _ { i n }" + }, + { + "category_id": 13, + "poly": [ + 1359, + 796, + 1400, + 796, + 1400, + 826, + 1359, + 826 + ], + "score": 0.86, + "latex": "\\mathbf { Y } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 536, + 731, + 563, + 731, + 563, + 762, + 536, + 762 + ], + "score": 0.86, + "latex": "\\beta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 706, + 632, + 727, + 632, + 727, + 663, + 706, + 663 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 984, + 735, + 1010, + 735, + 1010, + 760, + 984, + 760 + ], + "score": 0.83, + "latex": "v _ { i }" + }, + { + "category_id": 13, + "poly": [ + 721, + 1321, + 750, + 1321, + 750, + 1347, + 721, + 1347 + ], + "score": 0.82, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 802, + 1321, + 822, + 1321, + 822, + 1347, + 802, + 1347 + ], + "score": 0.82, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 858, + 541, + 879, + 541, + 879, + 571, + 858, + 571 + ], + "score": 0.78, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 298, + 1231, + 326, + 1231, + 326, + 1256, + 298, + 1256 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1163, + 1136, + 1180, + 1136, + 1180, + 1162, + 1163, + 1162 + ], + "score": 0.76, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 776, + 1231, + 795, + 1231, + 795, + 1256, + 776, + 1256 + ], + "score": 0.75, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 297, + 541, + 319, + 541, + 319, + 571, + 297, + 571 + ], + "score": 0.72, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 911, + 336, + 931, + 336, + 931, + 361, + 911, + 361 + ], + "score": 0.71, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 1328, + 700, + 1349, + 700, + 1349, + 728, + 1328, + 728 + ], + "score": 0.57, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 596, + 450, + 622, + 450, + 622, + 477, + 596, + 477 + ], + "score": 0.38, + "latex": "\\pmb { \\& }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 891.0, + 652.0, + 891.0, + 652.0, + 930.0, + 294.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1669.0, + 630.0, + 1669.0, + 630.0, + 1720.0, + 292.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1414.0, + 525.0, + 1414.0, + 525.0, + 1470.0, + 290.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2092.0, + 840.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1734.0, + 1405.0, + 1734.0, + 1405.0, + 1767.0, + 296.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1763.0, + 1407.0, + 1763.0, + 1407.0, + 1798.0, + 293.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1793.0, + 1406.0, + 1793.0, + 1406.0, + 1830.0, + 293.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 1408.0, + 1820.0, + 1408.0, + 1860.0, + 293.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1852.0, + 1407.0, + 1852.0, + 1407.0, + 1891.0, + 293.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1406.0, + 1883.0, + 1406.0, + 1920.0, + 295.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1910.0, + 1410.0, + 1910.0, + 1410.0, + 1952.0, + 291.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1940.0, + 1408.0, + 1940.0, + 1408.0, + 1982.0, + 291.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1976.0, + 443.0, + 1976.0, + 443.0, + 2014.0, + 293.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1135.0, + 1162.0, + 1135.0, + 1162.0, + 1169.0, + 359.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1135.0, + 1404.0, + 1135.0, + 1404.0, + 1169.0, + 1181.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1161.0, + 547.0, + 1161.0, + 547.0, + 1205.0, + 289.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1161.0, + 1232.0, + 1161.0, + 1232.0, + 1205.0, + 707.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1161.0, + 1410.0, + 1161.0, + 1410.0, + 1205.0, + 1295.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1196.0, + 715.0, + 1196.0, + 715.0, + 1233.0, + 293.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1196.0, + 1039.0, + 1196.0, + 1039.0, + 1233.0, + 896.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1196.0, + 1404.0, + 1196.0, + 1404.0, + 1233.0, + 1143.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1227.0, + 297.0, + 1227.0, + 297.0, + 1265.0, + 294.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1227.0, + 775.0, + 1227.0, + 775.0, + 1265.0, + 327.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1227.0, + 1405.0, + 1227.0, + 1405.0, + 1265.0, + 796.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1259.0, + 1405.0, + 1259.0, + 1405.0, + 1293.0, + 294.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1285.0, + 1405.0, + 1285.0, + 1405.0, + 1324.0, + 292.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1317.0, + 720.0, + 1317.0, + 720.0, + 1357.0, + 292.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1317.0, + 801.0, + 1317.0, + 801.0, + 1357.0, + 751.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1317.0, + 1409.0, + 1317.0, + 1409.0, + 1357.0, + 823.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1347.0, + 766.0, + 1347.0, + 766.0, + 1385.0, + 293.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 449.0, + 595.0, + 449.0, + 595.0, + 483.0, + 295.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 449.0, + 1402.0, + 449.0, + 1402.0, + 483.0, + 623.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 1406.0, + 478.0, + 1406.0, + 513.0, + 294.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 508.0, + 1405.0, + 508.0, + 1405.0, + 541.0, + 295.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 540.0, + 857.0, + 540.0, + 857.0, + 575.0, + 320.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 540.0, + 1405.0, + 540.0, + 1405.0, + 575.0, + 880.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 570.0, + 1404.0, + 570.0, + 1404.0, + 604.0, + 295.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 598.0, + 1405.0, + 598.0, + 1405.0, + 634.0, + 295.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 630.0, + 705.0, + 630.0, + 705.0, + 666.0, + 295.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 630.0, + 1357.0, + 630.0, + 1357.0, + 666.0, + 728.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1488.0, + 1405.0, + 1488.0, + 1405.0, + 1525.0, + 294.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1519.0, + 1407.0, + 1519.0, + 1407.0, + 1556.0, + 294.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1550.0, + 1405.0, + 1550.0, + 1405.0, + 1585.0, + 291.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1577.0, + 1407.0, + 1577.0, + 1407.0, + 1617.0, + 292.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1606.0, + 449.0, + 1606.0, + 449.0, + 1647.0, + 293.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 948.0, + 1408.0, + 948.0, + 1408.0, + 989.0, + 292.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 982.0, + 1405.0, + 982.0, + 1405.0, + 1014.0, + 293.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1012.0, + 1407.0, + 1012.0, + 1407.0, + 1045.0, + 294.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 766.0, + 1404.0, + 766.0, + 1404.0, + 804.0, + 294.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 794.0, + 1078.0, + 794.0, + 1078.0, + 834.0, + 291.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 794.0, + 1358.0, + 794.0, + 1358.0, + 834.0, + 1135.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 794.0, + 1406.0, + 794.0, + 1406.0, + 834.0, + 1401.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 829.0, + 345.0, + 829.0, + 345.0, + 863.0, + 296.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 829.0, + 1366.0, + 829.0, + 1366.0, + 863.0, + 389.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 663.0, + 801.0, + 663.0, + 801.0, + 706.0, + 292.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 663.0, + 1404.0, + 663.0, + 1404.0, + 706.0, + 956.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 697.0, + 915.0, + 697.0, + 915.0, + 734.0, + 292.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 697.0, + 1327.0, + 697.0, + 1327.0, + 734.0, + 1067.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 697.0, + 1404.0, + 697.0, + 1404.0, + 734.0, + 1350.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 729.0, + 535.0, + 729.0, + 535.0, + 764.0, + 291.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 729.0, + 983.0, + 729.0, + 983.0, + 764.0, + 564.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 729.0, + 1160.0, + 729.0, + 1160.0, + 764.0, + 1011.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 197.0, + 548.0, + 197.0, + 548.0, + 240.0, + 309.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 197.0, + 1131.0, + 197.0, + 1131.0, + 240.0, + 622.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 197.0, + 1407.0, + 197.0, + 1407.0, + 240.0, + 1318.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 233.0, + 595.0, + 233.0, + 595.0, + 267.0, + 336.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 233.0, + 1405.0, + 233.0, + 1405.0, + 267.0, + 625.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 262.0, + 909.0, + 262.0, + 909.0, + 298.0, + 336.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 297.0, + 731.0, + 297.0, + 731.0, + 341.0, + 308.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 297.0, + 1313.0, + 297.0, + 1313.0, + 341.0, + 920.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1405.0, + 297.0, + 1411.0, + 297.0, + 1411.0, + 341.0, + 1405.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 335.0, + 365.0, + 335.0, + 365.0, + 365.0, + 335.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 335.0, + 910.0, + 335.0, + 910.0, + 365.0, + 454.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 335.0, + 1406.0, + 335.0, + 1406.0, + 365.0, + 932.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 365.0, + 1407.0, + 365.0, + 1407.0, + 399.0, + 332.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 396.0, + 832.0, + 396.0, + 832.0, + 429.0, + 335.0, + 429.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 407, + 1404, + 407, + 1404, + 651, + 297, + 651 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1735, + 1404, + 1735, + 1404, + 1949, + 298, + 1949 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 202, + 1404, + 202, + 1404, + 394, + 297, + 394 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 746, + 1405, + 746, + 1405, + 992, + 297, + 992 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 283, + 1244, + 1382, + 1244, + 1382, + 1561, + 283, + 1561 + ], + "score": 0.973 + }, + { + "category_id": 0, + "poly": [ + 299, + 689, + 910, + 689, + 910, + 722, + 299, + 722 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 297, + 1678, + 984, + 1678, + 984, + 1711, + 297, + 1711 + ], + "score": 0.924 + }, + { + "category_id": 4, + "poly": [ + 480, + 1579, + 1218, + 1579, + 1218, + 1612, + 480, + 1612 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 333, + 1978, + 829, + 1978, + 829, + 2007, + 333, + 2007 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 858, + 2062, + 858, + 2085, + 840, + 2085 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 323, + 1017, + 1405, + 1017, + 1405, + 1214, + 323, + 1214 + ], + "score": 0.608 + }, + { + "category_id": 13, + "poly": [ + 1075, + 292, + 1213, + 292, + 1213, + 334, + 1075, + 334 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\gamma = \\frac { 1 } { 2 ( 1 - \\lambda ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 665, + 1857, + 895, + 1857, + 895, + 1890, + 665, + 1890 + ], + "score": 0.91, + "latex": "\\{ 0 , 5 , 1 0 , 2 0 , 5 0 , 8 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 897, + 264, + 958, + 264, + 958, + 297, + 897, + 297 + ], + "score": 0.91, + "latex": "h _ { \\theta } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 567, + 331, + 663, + 331, + 663, + 360, + 567, + 360 + ], + "score": 0.9, + "latex": "K = 1 0" + }, + { + "category_id": 13, + "poly": [ + 870, + 238, + 891, + 238, + 891, + 261, + 870, + 261 + ], + "score": 0.8, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 771, + 332, + 791, + 332, + 791, + 360, + 771, + 360 + ], + "score": 0.76, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 1381, + 499, + 1401, + 499, + 1401, + 526, + 1381, + 526 + ], + "score": 0.76, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 1291, + 234, + 1347, + 234, + 1347, + 267, + 1291, + 267 + ], + "score": 0.54, + "latex": "[ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 937, + 330, + 994, + 330, + 994, + 365, + 937, + 365 + ], + "score": 0.25, + "latex": "[ 0 , 1 ]" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 322.0, + 1251.0, + 322.0, + 1271.0, + 295.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1259.0, + 595.0, + 1259.0, + 595.0, + 1279.0, + 538.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1246.0, + 700.0, + 1246.0, + 700.0, + 1264.0, + 673.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1245.0, + 1076.0, + 1245.0, + 1076.0, + 1264.0, + 1050.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1281.0, + 324.0, + 1281.0, + 324.0, + 1305.0, + 294.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1279.0, + 563.0, + 1279.0, + 563.0, + 1306.0, + 538.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1272.0, + 604.0, + 1272.0, + 604.0, + 1309.0, + 564.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1273.0, + 1076.0, + 1273.0, + 1076.0, + 1293.0, + 1050.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1307.0, + 563.0, + 1307.0, + 563.0, + 1322.0, + 539.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1305.0, + 608.0, + 1305.0, + 608.0, + 1323.0, + 565.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1302.0, + 700.0, + 1302.0, + 700.0, + 1322.0, + 673.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1302.0, + 1077.0, + 1302.0, + 1077.0, + 1322.0, + 1050.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1313.0, + 324.0, + 1313.0, + 324.0, + 1336.0, + 294.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1323.0, + 561.0, + 1323.0, + 561.0, + 1336.0, + 541.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 1320.0, + 615.0, + 1320.0, + 615.0, + 1341.0, + 563.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1330.0, + 322.0, + 1330.0, + 322.0, + 1406.0, + 278.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1333.0, + 681.0, + 1333.0, + 681.0, + 1404.0, + 659.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1330.0, + 970.0, + 1330.0, + 970.0, + 1339.0, + 960.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1361.0, + 699.0, + 1361.0, + 699.0, + 1380.0, + 673.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1358.0, + 1079.0, + 1358.0, + 1079.0, + 1382.0, + 1049.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1376.0, + 323.0, + 1376.0, + 323.0, + 1401.0, + 294.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1376.0, + 984.0, + 1376.0, + 984.0, + 1386.0, + 940.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1388.0, + 765.0, + 1388.0, + 765.0, + 1397.0, + 753.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1388.0, + 1077.0, + 1388.0, + 1077.0, + 1408.0, + 1050.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1388.0, + 1105.0, + 1388.0, + 1105.0, + 1397.0, + 1089.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1400.0, + 730.0, + 1400.0, + 730.0, + 1409.0, + 715.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1402.0, + 935.0, + 1402.0, + 935.0, + 1413.0, + 922.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1397.0, + 972.0, + 1397.0, + 972.0, + 1417.0, + 941.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1397.0, + 1350.0, + 1397.0, + 1350.0, + 1417.0, + 1296.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1412.0, + 348.0, + 1412.0, + 348.0, + 1444.0, + 295.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1418.0, + 699.0, + 1418.0, + 699.0, + 1438.0, + 673.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1412.0, + 981.0, + 1412.0, + 981.0, + 1447.0, + 916.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1418.0, + 1077.0, + 1418.0, + 1077.0, + 1438.0, + 1050.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1431.0, + 1316.0, + 1431.0, + 1316.0, + 1446.0, + 1293.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 1412.0, + 1360.0, + 1412.0, + 1360.0, + 1449.0, + 1317.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1443.0, + 322.0, + 1443.0, + 322.0, + 1464.0, + 295.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1443.0, + 986.0, + 1443.0, + 986.0, + 1465.0, + 916.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1447.0, + 1077.0, + 1447.0, + 1077.0, + 1466.0, + 1050.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1443.0, + 1363.0, + 1443.0, + 1363.0, + 1465.0, + 1293.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1464.0, + 935.0, + 1464.0, + 935.0, + 1474.0, + 920.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1459.0, + 990.0, + 1459.0, + 990.0, + 1481.0, + 941.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1464.0, + 1310.0, + 1464.0, + 1310.0, + 1474.0, + 1295.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 1459.0, + 1367.0, + 1459.0, + 1367.0, + 1481.0, + 1317.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 1486.0, + 359.0, + 1486.0, + 359.0, + 1502.0, + 344.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1487.0, + 387.0, + 1487.0, + 387.0, + 1500.0, + 375.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1487.0, + 417.0, + 1487.0, + 417.0, + 1499.0, + 405.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 1487.0, + 445.0, + 1487.0, + 445.0, + 1499.0, + 434.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1487.0, + 477.0, + 1487.0, + 477.0, + 1499.0, + 464.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1486.0, + 513.0, + 1486.0, + 513.0, + 1500.0, + 488.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1485.0, + 629.0, + 1485.0, + 629.0, + 1502.0, + 515.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1487.0, + 734.0, + 1487.0, + 734.0, + 1500.0, + 723.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1487.0, + 764.0, + 1487.0, + 764.0, + 1500.0, + 753.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1487.0, + 794.0, + 1487.0, + 794.0, + 1499.0, + 783.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 1487.0, + 823.0, + 1487.0, + 823.0, + 1499.0, + 811.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1486.0, + 854.0, + 1486.0, + 854.0, + 1502.0, + 839.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1486.0, + 889.0, + 1486.0, + 889.0, + 1500.0, + 866.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1485.0, + 949.0, + 1485.0, + 949.0, + 1502.0, + 894.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1486.0, + 976.0, + 1486.0, + 976.0, + 1500.0, + 952.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1483.0, + 1006.0, + 1483.0, + 1006.0, + 1503.0, + 981.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1487.0, + 1111.0, + 1487.0, + 1111.0, + 1500.0, + 1100.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1489.0, + 1140.0, + 1489.0, + 1140.0, + 1500.0, + 1130.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1487.0, + 1170.0, + 1487.0, + 1170.0, + 1499.0, + 1158.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1486.0, + 1201.0, + 1486.0, + 1201.0, + 1502.0, + 1186.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1486.0, + 1231.0, + 1486.0, + 1231.0, + 1502.0, + 1216.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1486.0, + 1265.0, + 1486.0, + 1265.0, + 1500.0, + 1242.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1485.0, + 1325.0, + 1485.0, + 1325.0, + 1502.0, + 1272.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1485.0, + 1355.0, + 1485.0, + 1355.0, + 1503.0, + 1328.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1358.0, + 1483.0, + 1383.0, + 1483.0, + 1383.0, + 1503.0, + 1358.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1499.0, + 553.0, + 1499.0, + 553.0, + 1520.0, + 387.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1499.0, + 930.0, + 1499.0, + 930.0, + 1520.0, + 763.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1499.0, + 1307.0, + 1499.0, + 1307.0, + 1520.0, + 1140.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1531.0, + 505.0, + 1531.0, + 505.0, + 1562.0, + 409.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1530.0, + 901.0, + 1530.0, + 901.0, + 1561.0, + 767.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1531.0, + 1276.0, + 1531.0, + 1276.0, + 1560.0, + 1146.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.75, + 1314.0, + 1054.75, + 1314.0, + 1054.75, + 1410.0, + 1039.75, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1472.5, + 327.0, + 1472.5, + 327.0, + 1501.0, + 299.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1472.5, + 704.0, + 1472.5, + 704.0, + 1501.5, + 675.0, + 1501.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1473.0, + 1081.0, + 1473.0, + 1081.0, + 1500.0, + 1052.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 686.0, + 915.0, + 686.0, + 915.0, + 729.0, + 292.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1676.0, + 988.0, + 1676.0, + 988.0, + 1716.0, + 293.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 1576.0, + 1218.0, + 1576.0, + 1218.0, + 1615.0, + 480.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1970.0, + 834.0, + 1970.0, + 834.0, + 2014.0, + 330.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2090.0, + 838.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 406.0, + 1404.0, + 406.0, + 1404.0, + 442.0, + 294.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 439.0, + 1405.0, + 439.0, + 1405.0, + 472.0, + 294.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 467.0, + 1406.0, + 467.0, + 1406.0, + 504.0, + 294.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 499.0, + 1380.0, + 499.0, + 1380.0, + 533.0, + 294.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 499.0, + 1405.0, + 499.0, + 1405.0, + 533.0, + 1402.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 527.0, + 1404.0, + 527.0, + 1404.0, + 566.0, + 294.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 560.0, + 1404.0, + 560.0, + 1404.0, + 593.0, + 294.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 590.0, + 1405.0, + 590.0, + 1405.0, + 623.0, + 295.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 621.0, + 595.0, + 621.0, + 595.0, + 652.0, + 293.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1737.0, + 1402.0, + 1737.0, + 1402.0, + 1767.0, + 296.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1768.0, + 1404.0, + 1768.0, + 1404.0, + 1799.0, + 296.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1794.0, + 1405.0, + 1794.0, + 1405.0, + 1831.0, + 293.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1829.0, + 1404.0, + 1829.0, + 1404.0, + 1859.0, + 296.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1857.0, + 664.0, + 1857.0, + 664.0, + 1891.0, + 294.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1857.0, + 1402.0, + 1857.0, + 1402.0, + 1891.0, + 896.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1886.0, + 1405.0, + 1886.0, + 1405.0, + 1922.0, + 293.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1917.0, + 1408.0, + 1917.0, + 1408.0, + 1954.0, + 293.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 203.0, + 1405.0, + 203.0, + 1405.0, + 239.0, + 295.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 232.0, + 869.0, + 232.0, + 869.0, + 269.0, + 291.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 232.0, + 1290.0, + 232.0, + 1290.0, + 269.0, + 892.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 232.0, + 1407.0, + 232.0, + 1407.0, + 269.0, + 1348.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 258.0, + 896.0, + 258.0, + 896.0, + 303.0, + 291.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 258.0, + 1406.0, + 258.0, + 1406.0, + 303.0, + 959.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 286.0, + 1074.0, + 286.0, + 1074.0, + 345.0, + 287.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 286.0, + 1409.0, + 286.0, + 1409.0, + 345.0, + 1214.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 329.0, + 566.0, + 329.0, + 566.0, + 370.0, + 294.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 329.0, + 770.0, + 329.0, + 770.0, + 370.0, + 664.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 329.0, + 936.0, + 329.0, + 936.0, + 370.0, + 792.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 329.0, + 1406.0, + 329.0, + 1406.0, + 370.0, + 995.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 362.0, + 1236.0, + 362.0, + 1236.0, + 401.0, + 294.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 745.0, + 1405.0, + 745.0, + 1405.0, + 783.0, + 294.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 778.0, + 1407.0, + 778.0, + 1407.0, + 811.0, + 295.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 807.0, + 1405.0, + 807.0, + 1405.0, + 844.0, + 294.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 840.0, + 1403.0, + 840.0, + 1403.0, + 870.0, + 296.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 869.0, + 1406.0, + 869.0, + 1406.0, + 903.0, + 295.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 897.0, + 1410.0, + 897.0, + 1410.0, + 933.0, + 291.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 930.0, + 1405.0, + 930.0, + 1405.0, + 964.0, + 295.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 959.0, + 1374.0, + 959.0, + 1374.0, + 996.0, + 294.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 1017.0, + 1406.0, + 1017.0, + 1406.0, + 1049.0, + 335.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1048.0, + 1405.0, + 1048.0, + 1405.0, + 1081.0, + 355.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1078.0, + 986.0, + 1078.0, + 986.0, + 1113.0, + 353.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 1119.0, + 1406.0, + 1119.0, + 1406.0, + 1156.0, + 337.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1155.0, + 1403.0, + 1155.0, + 1403.0, + 1184.0, + 357.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1183.0, + 437.0, + 1183.0, + 437.0, + 1214.0, + 354.0, + 1214.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1352, + 1404, + 1352, + 1404, + 1536, + 297, + 1536 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1780, + 1405, + 1780, + 1405, + 1995, + 298, + 1995 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 282, + 868, + 1390, + 868, + 1390, + 1192, + 282, + 1192 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 280, + 188, + 1424, + 188, + 1424, + 540, + 280, + 540 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 332, + 1561, + 1404, + 1561, + 1404, + 1756, + 332, + 1756 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 297, + 1296, + 958, + 1296, + 958, + 1330, + 297, + 1330 + ], + "score": 0.929 + }, + { + "category_id": 4, + "poly": [ + 514, + 1209, + 1183, + 1209, + 1183, + 1242, + 514, + 1242 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 329, + 733, + 1404, + 733, + 1404, + 796, + 329, + 796 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2084, + 840, + 2084 + ], + "score": 0.798 + }, + { + "category_id": 1, + "poly": [ + 332, + 631, + 1407, + 631, + 1407, + 722, + 332, + 722 + ], + "score": 0.759 + }, + { + "category_id": 4, + "poly": [ + 537, + 559, + 1160, + 559, + 1160, + 592, + 537, + 592 + ], + "score": 0.413 + }, + { + "category_id": 4, + "poly": [ + 533, + 559, + 1163, + 559, + 1163, + 592, + 533, + 592 + ], + "score": 0.108 + }, + { + "category_id": 13, + "poly": [ + 398, + 1445, + 450, + 1445, + 450, + 1473, + 398, + 1473 + ], + "score": 0.88, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 1056, + 1664, + 1140, + 1664, + 1140, + 1695, + 1056, + 1695 + ], + "score": 0.87, + "latex": "1 - \\beta _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1373, + 1354, + 1400, + 1354, + 1400, + 1385, + 1373, + 1385 + ], + "score": 0.87, + "latex": "\\beta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 455, + 1388, + 481, + 1388, + 481, + 1415, + 455, + 1415 + ], + "score": 0.84, + "latex": "v _ { i }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 868.0, + 323.0, + 868.0, + 323.0, + 887.0, + 296.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 870.0, + 1078.0, + 870.0, + 1078.0, + 888.0, + 1050.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 894.0, + 323.0, + 894.0, + 323.0, + 913.0, + 296.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 891.0, + 700.0, + 891.0, + 700.0, + 909.0, + 674.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 895.0, + 733.0, + 895.0, + 733.0, + 904.0, + 713.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 895.0, + 776.0, + 895.0, + 776.0, + 909.0, + 747.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 894.0, + 1091.0, + 894.0, + 1091.0, + 911.0, + 1075.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 908.0, + 1078.0, + 908.0, + 1078.0, + 928.0, + 1051.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 912.0, + 1125.0, + 912.0, + 1125.0, + 923.0, + 1113.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 919.0, + 323.0, + 919.0, + 323.0, + 940.0, + 296.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 935.0, + 700.0, + 935.0, + 700.0, + 953.0, + 674.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 939.0, + 835.0, + 939.0, + 835.0, + 950.0, + 823.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 279.0, + 945.0, + 323.0, + 945.0, + 323.0, + 1030.0, + 279.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 958.0, + 700.0, + 958.0, + 700.0, + 1034.0, + 656.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 948.0, + 919.0, + 948.0, + 919.0, + 957.0, + 909.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 948.0, + 1078.0, + 948.0, + 1078.0, + 1032.0, + 1034.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 944.0, + 1127.0, + 944.0, + 1127.0, + 960.0, + 1110.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 994.0, + 1152.0, + 994.0, + 1152.0, + 1010.0, + 1137.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 985.0, + 1164.0, + 985.0, + 1164.0, + 997.0, + 1156.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1022.0, + 323.0, + 1022.0, + 323.0, + 1069.0, + 295.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1025.0, + 383.0, + 1025.0, + 383.0, + 1055.0, + 334.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1019.0, + 700.0, + 1019.0, + 700.0, + 1039.0, + 672.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1029.0, + 755.0, + 1029.0, + 755.0, + 1060.0, + 706.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1026.0, + 1078.0, + 1026.0, + 1078.0, + 1046.0, + 1051.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1034.0, + 1100.0, + 1034.0, + 1100.0, + 1045.0, + 1085.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1029.0, + 1132.0, + 1029.0, + 1132.0, + 1060.0, + 1103.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1050.0, + 388.0, + 1050.0, + 388.0, + 1068.0, + 334.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1055.0, + 760.0, + 1055.0, + 760.0, + 1072.0, + 706.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1056.0, + 1102.0, + 1056.0, + 1102.0, + 1071.0, + 1084.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1055.0, + 1137.0, + 1055.0, + 1137.0, + 1072.0, + 1103.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1063.0, + 353.0, + 1063.0, + 353.0, + 1132.0, + 295.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1062.0, + 396.0, + 1062.0, + 396.0, + 1132.0, + 354.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1116.0, + 421.0, + 1116.0, + 421.0, + 1132.0, + 406.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1115.0, + 462.0, + 1115.0, + 462.0, + 1132.0, + 448.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1113.0, + 510.0, + 1113.0, + 510.0, + 1133.0, + 486.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1113.0, + 553.0, + 1113.0, + 553.0, + 1133.0, + 528.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1113.0, + 596.0, + 1113.0, + 596.0, + 1133.0, + 570.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1113.0, + 637.0, + 1113.0, + 637.0, + 1133.0, + 612.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1062.0, + 700.0, + 1062.0, + 700.0, + 1082.0, + 672.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1071.0, + 723.0, + 1071.0, + 723.0, + 1107.0, + 708.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1067.0, + 767.0, + 1067.0, + 767.0, + 1132.0, + 725.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1117.0, + 792.0, + 1117.0, + 792.0, + 1131.0, + 781.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1117.0, + 835.0, + 1117.0, + 835.0, + 1129.0, + 823.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1100.0, + 889.0, + 1100.0, + 889.0, + 1133.0, + 861.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1113.0, + 928.0, + 1113.0, + 928.0, + 1133.0, + 904.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1112.0, + 971.0, + 1112.0, + 971.0, + 1133.0, + 947.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1113.0, + 1013.0, + 1113.0, + 1013.0, + 1133.0, + 990.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1066.0, + 1100.0, + 1066.0, + 1100.0, + 1136.0, + 1046.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1067.0, + 1144.0, + 1067.0, + 1144.0, + 1132.0, + 1104.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1115.0, + 1169.0, + 1115.0, + 1169.0, + 1132.0, + 1156.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1070.0, + 1211.0, + 1070.0, + 1211.0, + 1129.0, + 1201.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1090.0, + 1391.0, + 1090.0, + 1391.0, + 1133.0, + 1237.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1128.0, + 549.0, + 1128.0, + 549.0, + 1149.0, + 404.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1128.0, + 924.0, + 1128.0, + 924.0, + 1149.0, + 778.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1128.0, + 1300.0, + 1128.0, + 1300.0, + 1149.0, + 1154.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1160.0, + 506.0, + 1160.0, + 506.0, + 1194.0, + 411.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1160.0, + 901.0, + 1160.0, + 901.0, + 1194.0, + 768.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1163.0, + 1273.0, + 1163.0, + 1273.0, + 1190.0, + 1147.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 878.0, + 344.0, + 878.0, + 344.0, + 888.0, + 329.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.75, + 903.5, + 833.75, + 903.5, + 833.75, + 917.0, + 797.75, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 907.5, + 1130.0, + 907.5, + 1130.0, + 971.0, + 1071.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.75, + 914.0, + 1082.75, + 914.0, + 1082.75, + 924.0, + 1078.75, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 973.5, + 1135.0, + 973.5, + 1135.0, + 990.0, + 1120.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 973.0, + 879.0, + 973.0, + 879.0, + 986.0, + 862.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.25, + 993.5, + 889.25, + 993.5, + 889.25, + 1001.5, + 883.25, + 1001.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.75, + 1010.5, + 1176.75, + 1010.5, + 1176.75, + 1026.0, + 1164.75, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1026.0, + 923.0, + 1026.0, + 923.0, + 1042.0, + 905.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1065.0, + 875.0, + 1065.0, + 875.0, + 1077.0, + 864.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1039.0, + 1197.0, + 1039.0, + 1197.0, + 1067.0, + 1171.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1075.0, + 945.0, + 1075.0, + 945.0, + 1081.5, + 934.0, + 1081.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.25, + 1101.5, + 706.25, + 1101.5, + 706.25, + 1130.5, + 675.25, + 1130.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 188.0, + 327.0, + 188.0, + 327.0, + 211.0, + 295.0, + 211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 209.0, + 1081.0, + 209.0, + 1081.0, + 232.0, + 1050.0, + 232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 327.0, + 227.0, + 327.0, + 246.0, + 296.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 223.0, + 704.0, + 223.0, + 704.0, + 246.0, + 673.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 243.0, + 741.0, + 243.0, + 741.0, + 252.0, + 730.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 238.0, + 1093.0, + 238.0, + 1093.0, + 250.0, + 1080.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 225.0, + 1281.0, + 225.0, + 1281.0, + 238.0, + 1258.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 247.0, + 374.0, + 247.0, + 374.0, + 261.0, + 356.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 248.0, + 772.0, + 248.0, + 772.0, + 257.0, + 761.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 251.0, + 1151.0, + 251.0, + 1151.0, + 265.0, + 1133.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 327.0, + 261.0, + 327.0, + 284.0, + 295.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 270.0, + 404.0, + 270.0, + 404.0, + 279.0, + 393.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 264.0, + 803.0, + 264.0, + 803.0, + 274.0, + 792.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 270.0, + 834.0, + 270.0, + 834.0, + 281.0, + 822.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 263.0, + 1182.0, + 263.0, + 1182.0, + 277.0, + 1164.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 264.0, + 1215.0, + 264.0, + 1215.0, + 278.0, + 1195.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 281.0, + 436.0, + 281.0, + 436.0, + 291.0, + 424.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 288.0, + 327.0, + 288.0, + 327.0, + 367.0, + 278.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 287.0, + 706.0, + 287.0, + 706.0, + 368.0, + 657.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 304.0, + 896.0, + 304.0, + 896.0, + 318.0, + 879.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 286.0, + 1062.0, + 286.0, + 1062.0, + 368.0, + 1033.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 333.0, + 327.0, + 333.0, + 327.0, + 358.0, + 295.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 354.0, + 1018.0, + 354.0, + 1018.0, + 360.0, + 1011.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 326.0, + 1083.0, + 326.0, + 1083.0, + 353.0, + 1051.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 358.0, + 405.0, + 358.0, + 405.0, + 381.0, + 341.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 364.0, + 602.0, + 364.0, + 602.0, + 378.0, + 585.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 360.0, + 779.0, + 360.0, + 779.0, + 382.0, + 716.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 360.0, + 1157.0, + 360.0, + 1157.0, + 382.0, + 1090.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 370.0, + 327.0, + 370.0, + 327.0, + 394.0, + 296.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 376.0, + 406.0, + 376.0, + 406.0, + 397.0, + 369.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 383.0, + 636.0, + 383.0, + 636.0, + 399.0, + 615.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 376.0, + 704.0, + 376.0, + 704.0, + 395.0, + 676.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 378.0, + 781.0, + 378.0, + 781.0, + 400.0, + 743.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 381.0, + 988.0, + 381.0, + 988.0, + 391.0, + 975.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 372.0, + 1019.0, + 372.0, + 1019.0, + 390.0, + 1004.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 383.0, + 1109.0, + 383.0, + 1109.0, + 394.0, + 1098.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 377.0, + 1158.0, + 377.0, + 1158.0, + 400.0, + 1118.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 406.0, + 327.0, + 406.0, + 327.0, + 431.0, + 295.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 392.0, + 418.0, + 392.0, + 418.0, + 431.0, + 343.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 394.0, + 794.0, + 394.0, + 794.0, + 435.0, + 712.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 404.0, + 1019.0, + 404.0, + 1019.0, + 414.0, + 1006.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 394.0, + 1172.0, + 394.0, + 1172.0, + 437.0, + 1089.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 427.0, + 423.0, + 427.0, + 423.0, + 449.0, + 370.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 438.0, + 731.0, + 438.0, + 731.0, + 447.0, + 720.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 429.0, + 799.0, + 429.0, + 799.0, + 455.0, + 742.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 436.0, + 1109.0, + 436.0, + 1109.0, + 447.0, + 1098.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 429.0, + 1176.0, + 429.0, + 1176.0, + 455.0, + 1119.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 444.0, + 327.0, + 444.0, + 327.0, + 467.0, + 295.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 463.0, + 340.0, + 463.0, + 340.0, + 480.0, + 325.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 460.0, + 375.0, + 460.0, + 375.0, + 481.0, + 354.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 460.0, + 408.0, + 460.0, + 408.0, + 481.0, + 387.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 460.0, + 441.0, + 460.0, + 441.0, + 480.0, + 419.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 462.0, + 471.0, + 462.0, + 471.0, + 478.0, + 453.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 458.0, + 641.0, + 458.0, + 641.0, + 482.0, + 485.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 460.0, + 742.0, + 460.0, + 742.0, + 478.0, + 726.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 462.0, + 774.0, + 462.0, + 774.0, + 478.0, + 757.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 462.0, + 805.0, + 462.0, + 805.0, + 477.0, + 787.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 462.0, + 844.0, + 462.0, + 844.0, + 477.0, + 818.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 459.0, + 1026.0, + 459.0, + 1026.0, + 480.0, + 847.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 462.0, + 1120.0, + 462.0, + 1120.0, + 478.0, + 1103.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 460.0, + 1154.0, + 460.0, + 1154.0, + 480.0, + 1132.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 462.0, + 1184.0, + 462.0, + 1184.0, + 476.0, + 1163.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 462.0, + 1215.0, + 462.0, + 1215.0, + 477.0, + 1194.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 459.0, + 1401.0, + 459.0, + 1401.0, + 480.0, + 1223.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 477.0, + 586.0, + 477.0, + 586.0, + 498.0, + 404.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 477.0, + 967.0, + 477.0, + 967.0, + 498.0, + 778.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 477.0, + 1344.0, + 477.0, + 1344.0, + 498.0, + 1154.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 511.0, + 504.0, + 511.0, + 504.0, + 542.0, + 407.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 510.0, + 901.0, + 510.0, + 901.0, + 544.0, + 768.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 511.0, + 1275.0, + 511.0, + 1275.0, + 541.0, + 1143.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.25, + 226.5, + 426.25, + 226.5, + 426.25, + 243.0, + 389.25, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 236.0, + 346.0, + 236.0, + 346.0, + 250.0, + 330.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.25, + 244.0, + 1120.25, + 244.0, + 1120.25, + 263.0, + 1096.25, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.25, + 446.5, + 709.25, + 446.5, + 709.25, + 478.5, + 677.25, + 478.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.75, + 448.0, + 1084.75, + 448.0, + 1084.75, + 476.0, + 1055.75, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1296.0, + 961.0, + 1296.0, + 961.0, + 1336.0, + 295.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1207.0, + 1184.0, + 1207.0, + 1184.0, + 1245.0, + 514.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 861.0, + 2060.0, + 861.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 557.0, + 1163.0, + 557.0, + 1163.0, + 596.0, + 537.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 557.0, + 1162.0, + 557.0, + 1162.0, + 596.0, + 537.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1353.0, + 1372.0, + 1353.0, + 1372.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1353.0, + 1404.0, + 1353.0, + 1404.0, + 1388.0, + 1401.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1383.0, + 454.0, + 1383.0, + 454.0, + 1419.0, + 294.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1383.0, + 1406.0, + 1383.0, + 1406.0, + 1419.0, + 482.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1412.0, + 1405.0, + 1412.0, + 1405.0, + 1450.0, + 294.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1443.0, + 397.0, + 1443.0, + 397.0, + 1480.0, + 294.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1443.0, + 1406.0, + 1443.0, + 1406.0, + 1480.0, + 451.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 1406.0, + 1473.0, + 1406.0, + 1507.0, + 294.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1507.0, + 931.0, + 1507.0, + 931.0, + 1539.0, + 296.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1780.0, + 1405.0, + 1780.0, + 1405.0, + 1815.0, + 295.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1811.0, + 1406.0, + 1811.0, + 1406.0, + 1846.0, + 295.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1838.0, + 1406.0, + 1838.0, + 1406.0, + 1879.0, + 292.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1874.0, + 1406.0, + 1874.0, + 1406.0, + 1905.0, + 296.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1902.0, + 1406.0, + 1902.0, + 1406.0, + 1937.0, + 295.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1933.0, + 1406.0, + 1933.0, + 1406.0, + 1967.0, + 295.0, + 1967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1963.0, + 1299.0, + 1963.0, + 1299.0, + 1996.0, + 293.0, + 1996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1559.0, + 1404.0, + 1559.0, + 1404.0, + 1595.0, + 332.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1592.0, + 1405.0, + 1592.0, + 1405.0, + 1627.0, + 354.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1625.0, + 1249.0, + 1625.0, + 1249.0, + 1657.0, + 355.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1662.0, + 1055.0, + 1662.0, + 1055.0, + 1698.0, + 332.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1662.0, + 1404.0, + 1662.0, + 1404.0, + 1698.0, + 1141.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1697.0, + 1402.0, + 1697.0, + 1402.0, + 1725.0, + 358.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1726.0, + 999.0, + 1726.0, + 999.0, + 1759.0, + 356.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 730.0, + 1404.0, + 730.0, + 1404.0, + 768.0, + 330.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 762.0, + 1407.0, + 762.0, + 1407.0, + 798.0, + 355.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 630.0, + 1408.0, + 630.0, + 1408.0, + 667.0, + 332.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 696.0, + 353.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 693.0, + 1007.0, + 693.0, + 1007.0, + 727.0, + 355.0, + 727.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 833, + 1405, + 833, + 1405, + 1227, + 298, + 1227 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 299, + 1242, + 1405, + 1242, + 1405, + 1456, + 299, + 1456 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1577, + 1404, + 1577, + 1404, + 1823, + 298, + 1823 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 358, + 523, + 1344, + 523, + 1344, + 703, + 358, + 703 + ], + "score": 0.98, + "html": "
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes Average adaptive score for normal nodes0.987 ±0.0000.930 ± 0.0070.959 ± 0.005
Average residual weight for abnormal nodes0.922 ± 0.0040.689 ± 0.0240.826 ± 0.016
0.013 ±0.0000.070±0.0070.041 ±0.005
Average residual weight for normal nodes0.078 ± 0.0040.311 ± 0.0240.174 ± 0.016
" + }, + { + "category_id": 5, + "poly": [ + 356, + 253, + 1345, + 253, + 1345, + 432, + 356, + 432 + ], + "score": 0.978, + "html": "
MeasureCoraCiteSeerPubMed
Average adaptive score for abnormal nodes0.998 ± 0.0000.988 ± 0.0000.996 ± 0.000
Average adaptive score for normal nodes Average residual weight for abnormal nodes0.924 ± 0.002 0.002 ±0.0000.807 ± 0.005 0.012 ± 0.0000.869 ± 0.006
Average residual weight for normal nodes0.076 ± 0.0020.193 ± 0.0050.004±0.000 0.131 ± 0.006
" + }, + { + "category_id": 1, + "poly": [ + 301, + 1944, + 1402, + 1944, + 1402, + 2006, + 301, + 2006 + ], + "score": 0.945 + }, + { + "category_id": 0, + "poly": [ + 299, + 1870, + 944, + 1870, + 944, + 1910, + 299, + 1910 + ], + "score": 0.931 + }, + { + "category_id": 0, + "poly": [ + 298, + 759, + 549, + 759, + 549, + 798, + 298, + 798 + ], + "score": 0.915 + }, + { + "category_id": 6, + "poly": [ + 322, + 216, + 1368, + 216, + 1368, + 249, + 322, + 249 + ], + "score": 0.915 + }, + { + "category_id": 6, + "poly": [ + 304, + 486, + 1398, + 486, + 1398, + 519, + 304, + 519 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 298, + 1505, + 509, + 1505, + 509, + 1543, + 298, + 1543 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 837, + 2061, + 865, + 2061, + 865, + 2085, + 837, + 2085 + ], + "score": 0.846 + }, + { + "category_id": 13, + "poly": [ + 926, + 487, + 1012, + 487, + 1012, + 519, + 926, + 519 + ], + "score": 0.88, + "latex": "( 1 - \\beta )" + }, + { + "category_id": 13, + "poly": [ + 956, + 217, + 1042, + 217, + 1042, + 249, + 956, + 249 + ], + "score": 0.88, + "latex": "( 1 - \\beta )" + }, + { + "category_id": 13, + "poly": [ + 660, + 487, + 698, + 487, + 698, + 519, + 660, + 519 + ], + "score": 0.79, + "latex": "( \\beta )" + }, + { + "category_id": 13, + "poly": [ + 691, + 217, + 729, + 217, + 729, + 248, + 691, + 248 + ], + "score": 0.73, + "latex": "( \\beta )" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1866.0, + 949.0, + 1866.0, + 949.0, + 1920.0, + 293.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 757.0, + 552.0, + 757.0, + 552.0, + 802.0, + 292.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 210.0, + 690.0, + 210.0, + 690.0, + 255.0, + 325.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 210.0, + 955.0, + 210.0, + 955.0, + 255.0, + 730.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 210.0, + 1371.0, + 210.0, + 1371.0, + 255.0, + 1043.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 483.0, + 659.0, + 483.0, + 659.0, + 522.0, + 299.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 483.0, + 925.0, + 483.0, + 925.0, + 522.0, + 699.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 483.0, + 1401.0, + 483.0, + 1401.0, + 522.0, + 1013.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1501.0, + 514.0, + 1501.0, + 514.0, + 1550.0, + 292.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2098.0, + 831.0, + 2098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 831.0, + 1405.0, + 831.0, + 1405.0, + 871.0, + 295.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 863.0, + 1406.0, + 863.0, + 1406.0, + 901.0, + 292.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 895.0, + 1407.0, + 895.0, + 1407.0, + 931.0, + 293.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 925.0, + 1407.0, + 925.0, + 1407.0, + 961.0, + 295.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 953.0, + 1407.0, + 953.0, + 1407.0, + 997.0, + 289.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 985.0, + 1406.0, + 985.0, + 1406.0, + 1021.0, + 295.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1016.0, + 1408.0, + 1016.0, + 1408.0, + 1052.0, + 293.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1042.0, + 1407.0, + 1042.0, + 1407.0, + 1084.0, + 291.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1074.0, + 1406.0, + 1074.0, + 1406.0, + 1115.0, + 290.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1106.0, + 1403.0, + 1106.0, + 1403.0, + 1141.0, + 295.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1135.0, + 1405.0, + 1135.0, + 1405.0, + 1173.0, + 292.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1164.0, + 1405.0, + 1164.0, + 1405.0, + 1204.0, + 293.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1200.0, + 515.0, + 1200.0, + 515.0, + 1228.0, + 294.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 1409.0, + 1241.0, + 1409.0, + 1279.0, + 294.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1272.0, + 1407.0, + 1272.0, + 1407.0, + 1310.0, + 293.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1301.0, + 1406.0, + 1301.0, + 1406.0, + 1338.0, + 292.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1333.0, + 1407.0, + 1333.0, + 1407.0, + 1369.0, + 293.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1362.0, + 1405.0, + 1362.0, + 1405.0, + 1399.0, + 293.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1393.0, + 1403.0, + 1393.0, + 1403.0, + 1428.0, + 295.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1426.0, + 778.0, + 1426.0, + 778.0, + 1461.0, + 294.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1577.0, + 1406.0, + 1577.0, + 1406.0, + 1615.0, + 293.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1610.0, + 1405.0, + 1610.0, + 1405.0, + 1644.0, + 294.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1639.0, + 1404.0, + 1639.0, + 1404.0, + 1675.0, + 293.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1670.0, + 1405.0, + 1670.0, + 1405.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1701.0, + 1405.0, + 1701.0, + 1405.0, + 1736.0, + 292.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1730.0, + 1402.0, + 1730.0, + 1402.0, + 1764.0, + 294.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1797.0, + 292.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1790.0, + 803.0, + 1790.0, + 803.0, + 1827.0, + 295.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1945.0, + 1404.0, + 1945.0, + 1404.0, + 1980.0, + 296.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1407.0, + 1974.0, + 1407.0, + 2010.0, + 295.0, + 2010.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 378, + 1405, + 378, + 1405, + 681, + 298, + 681 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 295, + 772, + 1408, + 772, + 1408, + 2018, + 295, + 2018 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 300, + 307, + 755, + 307, + 755, + 346, + 300, + 346 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 299, + 727, + 455, + 727, + 455, + 762, + 299, + 762 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 296, + 203, + 1402, + 203, + 1402, + 266, + 296, + 266 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 863, + 2061, + 863, + 2085, + 836, + 2085 + ], + "score": 0.803 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 307.0, + 760.0, + 307.0, + 760.0, + 351.0, + 296.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 723.0, + 460.0, + 723.0, + 460.0, + 768.0, + 295.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 868.0, + 2058.0, + 868.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 376.0, + 1406.0, + 376.0, + 1406.0, + 414.0, + 293.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 407.0, + 1405.0, + 407.0, + 1405.0, + 441.0, + 293.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 438.0, + 1406.0, + 438.0, + 1406.0, + 474.0, + 292.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 469.0, + 1408.0, + 469.0, + 1408.0, + 504.0, + 293.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 497.0, + 1407.0, + 497.0, + 1407.0, + 535.0, + 293.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 526.0, + 1406.0, + 526.0, + 1406.0, + 567.0, + 292.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 560.0, + 1405.0, + 560.0, + 1405.0, + 595.0, + 295.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 589.0, + 1405.0, + 589.0, + 1405.0, + 625.0, + 293.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 619.0, + 1405.0, + 619.0, + 1405.0, + 655.0, + 293.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 653.0, + 929.0, + 653.0, + 929.0, + 684.0, + 296.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 781.0, + 1353.0, + 781.0, + 1353.0, + 819.0, + 307.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 834.0, + 1406.0, + 834.0, + 1406.0, + 872.0, + 307.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 864.0, + 1410.0, + 864.0, + 1410.0, + 907.0, + 352.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 896.0, + 725.0, + 896.0, + 725.0, + 931.0, + 352.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 949.0, + 1405.0, + 949.0, + 1405.0, + 988.0, + 307.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 979.0, + 926.0, + 979.0, + 926.0, + 1018.0, + 353.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1033.0, + 1406.0, + 1033.0, + 1406.0, + 1071.0, + 307.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1064.0, + 1406.0, + 1064.0, + 1406.0, + 1101.0, + 352.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1092.0, + 1256.0, + 1092.0, + 1256.0, + 1136.0, + 353.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1145.0, + 1405.0, + 1145.0, + 1405.0, + 1189.0, + 307.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1181.0, + 900.0, + 1181.0, + 900.0, + 1215.0, + 353.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1231.0, + 1406.0, + 1231.0, + 1406.0, + 1270.0, + 308.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1263.0, + 1204.0, + 1263.0, + 1204.0, + 1302.0, + 352.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1318.0, + 1409.0, + 1318.0, + 1409.0, + 1357.0, + 308.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1347.0, + 1395.0, + 1347.0, + 1395.0, + 1387.0, + 355.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1402.0, + 1407.0, + 1402.0, + 1407.0, + 1441.0, + 308.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1434.0, + 1407.0, + 1434.0, + 1407.0, + 1469.0, + 353.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1462.0, + 627.0, + 1462.0, + 627.0, + 1502.0, + 353.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 1513.0, + 1407.0, + 1513.0, + 1407.0, + 1559.0, + 305.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1546.0, + 1405.0, + 1546.0, + 1405.0, + 1586.0, + 353.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1576.0, + 1405.0, + 1576.0, + 1405.0, + 1618.0, + 350.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1607.0, + 1062.0, + 1607.0, + 1062.0, + 1645.0, + 350.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1661.0, + 1407.0, + 1661.0, + 1407.0, + 1701.0, + 294.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1694.0, + 1395.0, + 1694.0, + 1395.0, + 1729.0, + 356.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1746.0, + 1406.0, + 1746.0, + 1406.0, + 1785.0, + 293.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1775.0, + 1407.0, + 1775.0, + 1407.0, + 1818.0, + 353.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1804.0, + 1071.0, + 1804.0, + 1071.0, + 1846.0, + 353.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1860.0, + 1406.0, + 1860.0, + 1406.0, + 1900.0, + 294.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1891.0, + 889.0, + 1891.0, + 889.0, + 1929.0, + 352.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1946.0, + 1406.0, + 1946.0, + 1406.0, + 1985.0, + 294.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1974.0, + 894.0, + 1974.0, + 894.0, + 2016.0, + 353.0, + 2016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 200.0, + 1408.0, + 200.0, + 1408.0, + 240.0, + 292.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 232.0, + 389.0, + 232.0, + 389.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 294, + 157, + 1410, + 157, + 1410, + 2029, + 294, + 2029 + ], + "score": 0.78 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.727 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.327 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 204.0, + 1407.0, + 204.0, + 1407.0, + 243.0, + 293.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 228.0, + 1411.0, + 228.0, + 1411.0, + 275.0, + 348.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 263.0, + 554.0, + 263.0, + 554.0, + 296.0, + 356.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 319.0, + 1407.0, + 319.0, + 1407.0, + 358.0, + 293.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 347.0, + 1304.0, + 347.0, + 1304.0, + 386.0, + 352.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 399.0, + 1411.0, + 399.0, + 1411.0, + 446.0, + 288.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 434.0, + 430.0, + 434.0, + 430.0, + 469.0, + 354.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 487.0, + 1407.0, + 487.0, + 1407.0, + 526.0, + 293.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 518.0, + 1110.0, + 518.0, + 1110.0, + 557.0, + 352.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 573.0, + 1407.0, + 573.0, + 1407.0, + 612.0, + 293.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 607.0, + 505.0, + 607.0, + 505.0, + 641.0, + 351.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 658.0, + 1409.0, + 658.0, + 1409.0, + 697.0, + 293.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 688.0, + 920.0, + 688.0, + 920.0, + 727.0, + 352.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 740.0, + 1409.0, + 740.0, + 1409.0, + 785.0, + 290.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 773.0, + 1407.0, + 773.0, + 1407.0, + 812.0, + 354.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 803.0, + 662.0, + 803.0, + 662.0, + 840.0, + 349.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 857.0, + 1409.0, + 857.0, + 1409.0, + 896.0, + 295.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 892.0, + 1249.0, + 892.0, + 1249.0, + 925.0, + 356.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 941.0, + 1409.0, + 941.0, + 1409.0, + 982.0, + 295.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 974.0, + 1297.0, + 974.0, + 1297.0, + 1013.0, + 354.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1027.0, + 1405.0, + 1027.0, + 1405.0, + 1066.0, + 295.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1056.0, + 1407.0, + 1056.0, + 1407.0, + 1099.0, + 350.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1089.0, + 742.0, + 1089.0, + 742.0, + 1126.0, + 349.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1144.0, + 1407.0, + 1144.0, + 1407.0, + 1183.0, + 295.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1171.0, + 1409.0, + 1171.0, + 1409.0, + 1216.0, + 352.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1202.0, + 1110.0, + 1202.0, + 1110.0, + 1247.0, + 350.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1257.0, + 1407.0, + 1257.0, + 1407.0, + 1298.0, + 293.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1288.0, + 630.0, + 1288.0, + 630.0, + 1327.0, + 354.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1383.0, + 295.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1374.0, + 886.0, + 1374.0, + 886.0, + 1413.0, + 352.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1428.0, + 1407.0, + 1428.0, + 1407.0, + 1467.0, + 293.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1459.0, + 670.0, + 1459.0, + 670.0, + 1496.0, + 349.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1512.0, + 1403.0, + 1512.0, + 1403.0, + 1551.0, + 295.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1547.0, + 1160.0, + 1547.0, + 1160.0, + 1580.0, + 354.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1596.0, + 1409.0, + 1596.0, + 1409.0, + 1641.0, + 293.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1629.0, + 985.0, + 1629.0, + 985.0, + 1668.0, + 352.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1687.0, + 1403.0, + 1687.0, + 1403.0, + 1720.0, + 297.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1716.0, + 917.0, + 1716.0, + 917.0, + 1750.0, + 352.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1769.0, + 1407.0, + 1769.0, + 1407.0, + 1808.0, + 293.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1802.0, + 1287.0, + 1802.0, + 1287.0, + 1835.0, + 354.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1853.0, + 1405.0, + 1853.0, + 1405.0, + 1892.0, + 295.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1884.0, + 1407.0, + 1884.0, + 1407.0, + 1923.0, + 354.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1917.0, + 1407.0, + 1917.0, + 1407.0, + 1950.0, + 354.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1945.0, + 1409.0, + 1945.0, + 1409.0, + 1985.0, + 350.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1976.0, + 632.0, + 1976.0, + 632.0, + 2009.0, + 354.0, + 2009.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.828 + }, + { + "category_id": 1, + "poly": [ + 294, + 202, + 1410, + 202, + 1410, + 1067, + 294, + 1067 + ], + "score": 0.719 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 870.0, + 2057.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 204.0, + 1403.0, + 204.0, + 1403.0, + 239.0, + 297.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 234.0, + 1405.0, + 234.0, + 1405.0, + 271.0, + 350.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 265.0, + 1321.0, + 265.0, + 1321.0, + 300.0, + 354.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 318.0, + 1406.0, + 318.0, + 1406.0, + 353.0, + 296.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 348.0, + 1405.0, + 348.0, + 1405.0, + 381.0, + 353.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 376.0, + 1333.0, + 376.0, + 1333.0, + 413.0, + 352.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 428.0, + 1406.0, + 428.0, + 1406.0, + 466.0, + 293.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 457.0, + 1406.0, + 457.0, + 1406.0, + 499.0, + 350.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 491.0, + 566.0, + 491.0, + 566.0, + 524.0, + 355.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 540.0, + 1408.0, + 540.0, + 1408.0, + 582.0, + 293.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 572.0, + 1404.0, + 572.0, + 1404.0, + 609.0, + 353.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 601.0, + 1406.0, + 601.0, + 1406.0, + 641.0, + 352.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 634.0, + 1061.0, + 634.0, + 1061.0, + 670.0, + 354.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 683.0, + 1408.0, + 683.0, + 1408.0, + 725.0, + 293.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 715.0, + 1409.0, + 715.0, + 1409.0, + 754.0, + 352.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 748.0, + 468.0, + 748.0, + 468.0, + 780.0, + 354.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 1407.0, + 798.0, + 1407.0, + 837.0, + 295.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 829.0, + 1409.0, + 829.0, + 1409.0, + 867.0, + 353.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 882.0, + 1407.0, + 882.0, + 1407.0, + 917.0, + 296.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 912.0, + 1409.0, + 912.0, + 1409.0, + 951.0, + 353.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 942.0, + 616.0, + 942.0, + 616.0, + 978.0, + 353.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 991.0, + 1407.0, + 991.0, + 1407.0, + 1034.0, + 293.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1027.0, + 648.0, + 1027.0, + 648.0, + 1058.0, + 354.0, + 1058.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 406, + 669, + 1406, + 669, + 1406, + 930, + 406, + 930 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 407, + 979, + 1402, + 979, + 1402, + 1216, + 407, + 1216 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 407, + 1266, + 1404, + 1266, + 1404, + 1464, + 407, + 1464 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 406, + 299, + 1404, + 299, + 1404, + 500, + 406, + 500 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 405, + 549, + 1251, + 549, + 1251, + 618, + 405, + 618 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 365, + 509, + 840, + 509, + 840, + 541, + 365, + 541 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 363, + 628, + 671, + 628, + 671, + 659, + 363, + 659 + ], + "score": 0.913 + }, + { + "category_id": 1, + "poly": [ + 364, + 257, + 580, + 257, + 580, + 289, + 364, + 289 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 298, + 199, + 435, + 199, + 435, + 235, + 298, + 235 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 365, + 1226, + 1200, + 1226, + 1200, + 1257, + 365, + 1257 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 366, + 939, + 1402, + 939, + 1402, + 971, + 366, + 971 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.844 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 197.0, + 440.0, + 197.0, + 440.0, + 238.0, + 296.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2095.0, + 832.0, + 2095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 666.0, + 1410.0, + 666.0, + 1410.0, + 704.0, + 404.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 698.0, + 1244.0, + 698.0, + 1244.0, + 733.0, + 449.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 731.0, + 1407.0, + 731.0, + 1407.0, + 772.0, + 401.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 761.0, + 685.0, + 761.0, + 685.0, + 800.0, + 447.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 798.0, + 1408.0, + 798.0, + 1408.0, + 836.0, + 408.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 830.0, + 779.0, + 830.0, + 779.0, + 865.0, + 449.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 861.0, + 1406.0, + 861.0, + 1406.0, + 904.0, + 408.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 898.0, + 1022.0, + 898.0, + 1022.0, + 933.0, + 450.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 978.0, + 1178.0, + 978.0, + 1178.0, + 1014.0, + 404.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1013.0, + 987.0, + 1013.0, + 987.0, + 1050.0, + 404.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1048.0, + 1400.0, + 1048.0, + 1400.0, + 1088.0, + 403.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1088.0, + 1404.0, + 1088.0, + 1404.0, + 1121.0, + 406.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1119.0, + 689.0, + 1119.0, + 689.0, + 1152.0, + 451.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1152.0, + 1404.0, + 1152.0, + 1404.0, + 1189.0, + 406.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1183.0, + 891.0, + 1183.0, + 891.0, + 1219.0, + 449.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1264.0, + 1408.0, + 1264.0, + 1408.0, + 1302.0, + 404.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1296.0, + 661.0, + 1296.0, + 661.0, + 1331.0, + 448.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1330.0, + 1406.0, + 1330.0, + 1406.0, + 1370.0, + 405.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1363.0, + 951.0, + 1363.0, + 951.0, + 1399.0, + 450.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1398.0, + 1404.0, + 1398.0, + 1404.0, + 1434.0, + 408.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1431.0, + 926.0, + 1431.0, + 926.0, + 1464.0, + 449.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 296.0, + 1405.0, + 296.0, + 1405.0, + 335.0, + 404.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 330.0, + 801.0, + 330.0, + 801.0, + 363.0, + 450.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 364.0, + 1046.0, + 364.0, + 1046.0, + 401.0, + 404.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 401.0, + 1301.0, + 401.0, + 1301.0, + 437.0, + 405.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 436.0, + 1404.0, + 436.0, + 1404.0, + 470.0, + 404.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 465.0, + 596.0, + 465.0, + 596.0, + 503.0, + 449.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 548.0, + 1251.0, + 548.0, + 1251.0, + 583.0, + 404.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 584.0, + 1172.0, + 584.0, + 1172.0, + 621.0, + 402.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 506.0, + 846.0, + 506.0, + 846.0, + 546.0, + 359.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 622.0, + 675.0, + 622.0, + 675.0, + 666.0, + 357.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 253.0, + 586.0, + 253.0, + 586.0, + 292.0, + 361.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1221.0, + 1205.0, + 1221.0, + 1205.0, + 1264.0, + 357.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 934.0, + 1409.0, + 934.0, + 1409.0, + 977.0, + 360.0, + 977.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/rJgJDAVKvB/rJgJDAVKvB.md b/parse/train/rJgJDAVKvB/rJgJDAVKvB.md new file mode 100644 index 0000000000000000000000000000000000000000..aae06fb5e41c21881684ef05e45b8659a38180c5 --- /dev/null +++ b/parse/train/rJgJDAVKvB/rJgJDAVKvB.md @@ -0,0 +1,570 @@ +# LEARNING TO PLAN IN HIGH DIMENSIONS VIA NEURAL EXPLORATION-EXPLOITATION TREES + +∗Binghong Chen1, $\mathbf { \mathbf { \mathbf { \mathbf { \mathbf { \mathbf { B } } } } } } \mathbf { \mathbf { 0 } } \mathbf { \mathbf { \mathbf { \mathbf { D } } } } \mathbf { \mathbf { a } } \mathbf { i } ^ { 2 }$ , Qinjie $\mathbf { L i n ^ { 3 } }$ , Guo $\mathbf { Y e ^ { 3 } }$ , Han Liu3, Le Song1,4 1Georgia Institute of Technology 2Google Research, Brain Team 3Northwestern University 4Ant Financial + +# ABSTRACT + +We propose a meta path planning algorithm named Neural Exploration-Exploitation Trees (NEXT) for learning from prior experience for solving new path planning problems in high dimensional continuous state and action spaces. Compared to more classical sampling-based methods like RRT, our approach achieves much better sample efficiency in high-dimensions and can benefit from prior experience of planning in similar environments. More specifically, NEXT exploits a novel neural architecture which can learn promising search directions from problem structures. The learned prior is then integrated into a UCB-type algorithm to achieve an online balance between exploration and exploitation when solving a new problem. We conduct thorough experiments to show that NEXT accomplishes new planning problems with more compact search trees and significantly outperforms state-of-the-art methods on several benchmarks. + +# 1 INTRODUCTION + +Path planning is a fundamental problem with many real-world applications, such as robot manipulation and autonomous driving. A simple planning problem within low-dimensional state space can be solved by first discretizing the continuous state space into a grid, and then searching for a path on top of it using graph search algorithms such as $A ^ { * }$ (Hart et al., 1968). However, due to the curse of dimensionality, these approaches do not scale well with the number of dimensions of the state space. For high-dimensional planning problems, people often resort to sampling-based approaches to avoid explicit discretization. Sampling-based planning algorithms, such as probabilistic roadmaps (PRM) (Kavraki et al., 1996), rapidly-exploring random trees (RRT) (LaValle, 1998), and their variants (Karaman & Frazzoli, 2011) incrementally build an implicit representation of the state space using probing samples. These generic algorithms typically employ a uniform sampler which does not make use of the structures of the problem. Therefore they may require lots of samples to obtain a feasible solution for complicated problems. To improve the sample efficiency, heuristic biased samplers, such as Gaussian sampler (Boor et al., 1999), bridge test (Hsu et al., 2003) and reachability-guided sampler (Shkolnik et al., 2009) have been proposed. All these sampling heuristics are designed manually to address specific structural properties, which may or may not be valid for a new problem, and may lead to even worse performance compared to the uniform proposal. + +Online adaptation in path planning has also been investigated for improving sample efficiency in current planning problem. Specifically, Hsu et al. (2005) exploits online algorithms to dynamically adapts the mixture weights of several manually designed biased samplers. Burns & Brock (2005a;b) fit a model for the planning environment incrementally and use the model for planning. Yee et al. (2016) mimics the Monte-Carlos tree search (MCTS) for problems with continuous state and action spaces. These algorithms treat each planning problem independently, and the collected data from previous experiences and built model will be simply discarded when solving a new problem. However, in practice, similar planning problems may be solved again and again, where the problems are different but sharing common structures. For instance, grabbing a coffee cup on a table at different time are different problems, since the layout of paper and pens, the position and orientation of coffee cups may be different every time; however, all these problems show common structures of handling similar objects which are placed in similar fashions. Intuitively, if the common characteristics across problems can be learned via some shared latent representation, a planner based on such representation can then be transferred to new problems with improved sample efficiency. + +Several methods have been proposed recently to learn from past planning experiences to conduct more efficient and generalizable planning for future problems. These works are limited in one way or the other. Zucker et al. (2008); Zhang et al. (2018); Huh & Lee (2018) treat the sampler in the sampling-based planner as a stochastic policy to be learned and apply policy gradient or TD-algorithm to improve the policy. Finney et al. (2007); Bowen & Alterovitz (2014); Ye & Alterovitz (2017); Ichter et al. (2018); Kim et al. (2018); Kuo et al. (2018) apply imitation learning based on the collected demonstrations to bias for better sampler via variants of probabilistic models, e.g., (mixture of) Gaussians, conditional VAE, GAN, HMM and RNN. However, many of these approaches either rely on specially designed local features or assume the problems are indexed by special parameters, which limits the generalization ability. Deep representation learning provides a promising direction to extract the common structure among the planning problems, and thus mitigate such limitation on handdesigned features. However, existing work, e.g., motion planning networks (Qureshi et al., 2019), value iteration networks (VIN) (Tamar et al., 2016), and gated path planning networks (GPPN) (Lee et al., 2018), either apply off-the-shelf MLP architecture ignoring special structures in planning problems or can only deal with discrete state and action spaces in low-dimensional settings. + +In this paper, we present Neural EXploration-EXploitation Tree (NEXT), a meta neural path planning algorithm for high-dimensional continuous state space problems. The core contribution is a novel attention-based neural architecture that is capable of learning generalizable problem structures from previous experiences and produce promising search directions with automatic online exploration-exploitation balance adaption. Compared to existing learning-based planners, + +• NEXT is more generic. We propose an architecture that can embed high dimensional continuous state spaces into low dimensional discrete spaces, on which a neural planning module is used to extract planning representation. These module will be learned end-to-end. • NEXT balances exploration-exploitation trade-off. We integrate the learned neural prior into an upper confidence bound (UCB) style algorithm to achieve an online balance between exploration and exploitation when solving a new problem. + +Empirically, we show that NEXT can exploit past experiences to reduce the number of required samples drastically for solving new planning problems, and significantly outperforms previous state-of-the-arts on several benchmark tasks. + +# 1.1 RELATED WORKS + +Designing non-uniform sampling strategies for random search to improve the planning efficiency has been considered as we discussed above. Besides the mentioned algorithms, there are other works along this line, including informed RRT∗ (Gammell et al., 2014) and batch informed Trees $\mathrm { ( B I T ^ { * } ) }$ (Gammell et al., 2015) as the representative work. Randomized $A ^ { * }$ (Diankov & Kuffner, 2007) and sampling-based $A ^ { * }$ (Persson & Sharf, 2014) expand the search tree with hand-designed heuristics. These methods incorporate the human prior knowledge via hard-coded rules, which is fixed and unable to adapt to problems, and thus, may not universally applicable. Choudhury et al. (2018); Song et al. (2018) attempt to learn search heuristics. However, both methods are restricted to planning on discrete domains. Meanwhile, the latter one always employs an unnecessary hierarchical structure for path planning, which leads to inferior sample efficiency and extra computation. + +The online exploration-exploitation trade-off is also an important issue in planning. For instance, Rickert et al. (2009) constructs a potential field sampler and tuned the sampler variance based on collision rate for the trade-off heuristically. Paxton et al. (2017) separates the action space into high-level discrete options and low-level continuous actions, and only considered the trade-off at the discrete option level, ignoring the exploration-exploitation in the fine action space. These existing works address the trade-off in an ad-hoc way, which may be inferior for the balance. + +There have been sevearl non-learning-based planning methods that can also leverage experiences (Kavraki et al., 1996; Phillips et al., 2012) by utilizing search graphs created in previous problems. However, they are designed for largely fixed obstacles and cannot be generalized to unseen tasks from the same planning problems distribution. + +# 2 SETTINGS FOR LEARNING TO PLAN + +Let $S \subseteq \mathbb { R } ^ { q }$ be the state space of the problem, e.g., all the configurations of a robot and its base location in the workspace, $S _ { o b s } \subsetneq S$ be the obstacles set, $S _ { f r e e } : = \mathcal { S } \setminus \mathcal { S } _ { o b s }$ be the free space, $s _ { i n i t } \in S _ { f r e e }$ be the initial state and $\mathcal { S } _ { g o a l } \subsetneq S _ { f r e e }$ be the goal region. Then the space of all collisionfree paths can be defined as a continuous function $\Xi : = \{ \bar { \xi } ( \cdot ) : \bar { [ 0 , 1 ] } S _ { f r e e } \}$ . Let $c \left( \cdot \right) : \Xi \mapsto \mathbb { R }$ be the cost functional over a path. The optimal path planning problem is to find the optimal path in terms of cost $c ( \cdot )$ from start $s _ { i n i t }$ to goal $\mathcal { S } _ { g o a l }$ in free space $\boldsymbol { S } _ { f r e e }$ , i.e., + +$$ +\begin{array} { r } { \xi ^ { * } = \mathrm { a r g m i n } _ { \xi \in \Xi } \ c \left( \xi \right) , \quad \mathrm { s . t . } \xi ( 0 ) = s _ { i n i t } , \xi ( 1 ) \in S _ { g o a l } . } \end{array} +$$ + +Traditionally (Karaman & Frazzoli, 2011), the planner has direct access to $( s _ { i n i t } , S _ { g o a l } , c \left( \cdot \right) )$ and the workspace map (Ichter et al., 2018; Tamar et al., 2016; Lee et al., 2018), $\mathtt { m a p } ( \cdot ) : \mathbb { R } ^ { 2 }$ or $\mathbb { R } ^ { 3 } \mapsto \{ 0 , 1 \}$ , (0: free spaces and 1: obstacles). Since $\boldsymbol { S } _ { f r e e }$ often has a very irregular geometry (illustrated in Figure 10 in Appendix A), it is usually represented via a collision detection module which is able to detect the obstacles in a path segment. For the same reason, the feasible paths in $\Xi$ are hard to be described in parametric forms, and thus, the nonparametric $\xi$ , such as a sequence of interconnected path segments $\left[ s _ { 0 } , s _ { 1 } \right] , \left[ s _ { 1 } , s _ { 2 } \right] , \ldots , \left[ s _ { T - 1 } , s _ { T } \right] \subset \mathcal { S }$ with $\xi ( 0 ) = s _ { 0 } = s _ { i n i t }$ and $\xi ( 1 ) = s _ { T }$ , is used with an additive cost $\textstyle \sum _ { i = 1 } ^ { T } c \left( \left[ s _ { i - 1 } , s _ { i } \right] \right)$ −. + +Assuming given th planning problems $\{ U _ { i } : = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \mathtt { m a p } , c \left( \cdot \right) ) \} _ { i = 1 } ^ { N }$ sampled from some distribution $\mathcal { U }$ , we are interested in learning an algorithm $\mathsf { a l g } \left( \cdot \right)$ , which can produce the (nearly)-optimal path efficiently from the observed planning problems. Formally, the learning to plan is defined as + +$$ +\begin{array} { r } { \mathbf { a } \mathbf { 1 g } ^ { * } \left( \cdot \right) = \operatorname * { a r g m i n } _ { \mathbf { a } \mathbf { 1 g } \in \mathcal { A } } \mathbb { E } _ { U \in \mathcal { U } } \left[ \ell \left( \mathbf { a } \mathbf { 1 g } ( U ) \right) \right] , } \end{array} +$$ + +where $\mathcal { A }$ denotes the planning algorithm family, and $\ell \left( \cdot \right)$ denotes some loss function which evaluates the quality of the generated path and the efficiency of the $\mathsf { a l g } \left( \cdot \right)$ , e.g., size of the search tree. We elaborate each component in Eq (2) in the following sections. We first introduce the tree-based sampling algorithm template in Section 3, upon which we instantiate the $\mathsf { a l g } \left( \cdot \right)$ via a novel attention-based neural parametrization in Section 4.2 with exploration-exploitation balance mechanism in Section 4.1. We design the $\ell$ -loss function and the meta learning algorithm in Section 4.3. Composing every component together, we obtain the neural exploration-exploitation trees (NEXT) which achieves outstanding performances in Section 5. Repeat + +![](images/c81f5245215344ac0251524c1e65b7b0d359bb22ec9a016d0918f1939b8a9b94.jpg) +Figure 1: Illustration of NEXT. In each epoch, NEXT is executed on a randomly generated planning problem. The search tree grows with $\tilde { V } ^ { * }$ and $\tilde { \pi } ^ { * }$ guidance. $\{ \tilde { V } ^ { * } , \tilde { \pi } ^ { * } \}$ will be updated according to the successful path. Such planning and learning iteration is continued interactively. + +# 3 PRELIMINARIES + +# 1 Problem: + +# Algorithm 1: Tree-based Sampling Algorithm + +The sampling-based planners are more practical and become dominant for high-dimensional path planning problems (Elbanhawi & Simic, 3 2014). We describe a unifying view for many 4 existing tree-based sampling algorithms (TSA), 5 which we will also base our algorithm upon. More specifically, this family of algorithms 6 maintain a search tree $\tau$ rooted at the initial point $s _ { i n i t }$ and connecting all sampled points $\nu$ 7 in the configuration space with edge set $\mathcal { E }$ . The 8 tree will be expanded by incorporating more sampled states until some leaf reaches $\mathcal { S } _ { g o a l }$ 9. Then, a feasible solution for the path planning + +$U = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \mathrm { { m a p } , \it { c ( \cdot ) } ) }$ ; +2 Initialize $\boldsymbol { \mathcal { T } } = ( \boldsymbol { \mathcal { V } } , \boldsymbol { \mathcal { E } } )$ with $\mathcal { V } \{ s _ { i n i t } \}$ , $\mathcal { E } \emptyset$ ; +for $t \gets 0$ to $T$ do sparent, $s _ { n e w } \mathrm { E }$ xpand $( \mathcal { T } , U )$ ; if ObstacleFree( $\prime s _ { p a r e n t } , s _ { n e w } )$ ) then $\mathcal { V } \mathcal { V } \cup \{ s _ { n e w } \}$ and $\mathcal { E } \mathcal { E } \cup \{ [ s _ { p a r e n t } , s _ { n e w } ] \}$ ; T ← Postprocess(T , U) ; $\triangleright$ Optional if $s _ { n e w } \in S _ { g o a l }$ then return T ; + +problem will be extracted based on the tree $\tau$ . The template of tree-based sampling algorithms is summarized in Algorithm 1 and illustrated in Fig. 1(c). A key component of the algorithm is the Expand operator, which generates the next exploration point $s _ { n e w }$ and its parent $s _ { p a r e n t } \in \mathcal { V }$ . To ensure the feasibility of the solution, the $s _ { n e w }$ must be reachable from $\tau$ , i.e., $[ \bar { s _ { p a r e n t } } , s _ { n e w } ]$ is collision-free, which is checked by a collision detection function. As we will discuss in Appendix B, by instantiating different Expand operators, we will arrive at many existing algorithms, such as RRT (LaValle, 1998) and EST (Hsu et al., 1997; Phillips et al., 2004). + +One major limitation of existing TSAs is that they solve each problem independently from scratch and ignore past planning experiences in similar environments. We introduce the neural components into TSA template to form the learnable planning algorithm family $\mathcal { A }$ , which can explicitly take advantages of the past successful experiences to bias the Expand towards more promising regions. + +# 4 NEURAL EXPLORATION-EXPLOITATION TREES + +Based on the TSA framework, we introduce a learnable neural based Expand operator, which can balance between exploration and exploitation, to instantiate $\mathcal { A }$ in Eq (2). With the self-improving training, we obtain the meta NEXT algorithm illustrated in Figure 1. + +# 4.1 GUIDED PROGRESSIVE EXPANSION + +We start with our design of the Expand. We assume having an estimated value function ${ \tilde { V } } ^ { * } ( s | U )$ , which stands for the optimal cost from $s$ to target in planning problem $U$ , and a policy $\tilde { \pi } ^ { * } \left( s ^ { \prime } | s , U \right)$ , which generates the promising action $s ^ { \prime }$ from state $s$ . The concrete parametriza + +
Algorithm 2: NExT :: Expand(T = (V,ε),U)
Sparent ← argmaxs∈vΦ(s) ; Selection
{s1,*..,Sk} id π*(s'sparent,U)； >Candidates
Snew ← argmaxs'∈{s1,.,sk}(s')； DExpansion
4 return Sparent, Snew;
+ +tion of ${ \tilde { V } } ^ { * }$ and $\tilde { \pi } ^ { * }$ will be explained in Section 4.2 and learned in Section 4.3. We will use these functions to construct the learnable Expand with explicit exploration-exploitation balancing. + +The Expand operator will expand the current search tree $\tau$ by a new neighboring state $s _ { n e w }$ around $\tau$ . We design the expansion as a two-step procedure: (i) select a state $s _ { p a r e n t }$ from existing tree $\tau$ ; (ii) expand a state $s _ { n e w }$ in the neighborhood of $s _ { p a r e n t }$ . More specifically, + +Selecting $s _ { p a r e n t }$ from $\tau$ in step (i). Consider the negative value function $- \tilde { V } ^ { \ast } \left( s | U \right)$ as the rewards $r \left( s \right)$ , step (i) shares some similarity with the multi-armed bandit problem by viewing existing nodes $s \in \mathcal V$ as arms. However, the vanilla UCB algorithm is not directly applicable, since the number of states is increasing as the algorithm proceeds and the value of these adjacent states are naturally correlated. We address this challenge by modeling the correlation explicitly as in contextual bandits. Specifically, we parametrize the UCB of the reward function as $\phi \left( s \right)$ , and select a node from $\tau$ according to $\phi ( s )$ + +$$ +\begin{array} { r } { s _ { p a r e n t } = \mathrm { a r g m a x } _ { s \in \mathcal { V } } \phi ( s ) : = \bar { r } _ { t } \left( s \right) + \lambda \sigma _ { t } \left( s \right) , } \end{array} +$$ + +t t the sequence of where and denote the average reward and variance estimator after $t$ selected tree nodes so far as $S _ { t } = \{ s _ { p a r e n t } ^ { 1 } , . . . , s _ { p a r e n t } ^ { t } \}$ $t$ -calls to Expand. Denote , then we can use kernel smoothing estimator for $\begin{array} { r } { \bar { r } _ { t } ( s ) = \frac { \sum _ { s ^ { \prime } \in { \mathcal S } _ { t } } k \left( s ^ { \prime } , s \right) r \left( s ^ { \prime } \right) } { \sum _ { s ^ { \prime } \in { \mathcal S } _ { t } } k \left( s ^ { \prime } , s \right) } } \end{array}$ and $\begin{array} { r } { \sigma _ { t } ( s ) = \sqrt { \frac { \log { \sum _ { s ^ { \prime } \in { \mathcal S } _ { t } } w ( s ^ { \prime } ) } } { w ( s ) } } } \end{array}$ where $k ( s ^ { \prime } , s )$ is a kernel function and $\begin{array} { r } { w ( s ) = \sum _ { s ^ { \prime } \in S _ { t } } k ( s ^ { \prime } , s ) } \end{array}$ . Other parametrizations of $\bar { r } _ { t }$ and $\sigma _ { t }$ are also possible, such as Gaussian Process parametrization in Appendix C. The average reward exploits more promising states, while the variance promotes exploration towards less frequently visited states; and the exploration versus exploitation is balanced by a tunable weight $\lambda > 0$ . + +Expanding a reachable $s _ { n e w }$ in step (ii). Given the selected $s _ { p a r e n t }$ , we consider expanding a reachable state in the neighborhood $s _ { p a r e n t }$ as an infinite-armed bandit problem. Although one can first samples $k$ arms uniformly from a neighborhood around $s _ { p a r e n t }$ and runs a UCB algorithm on the randomly generated finite arms (Wang et al., 2009), such uniform sampler ignores problem structures, and will lead to unnecessary samples. Instead we will employ a policy $\bar { \tilde { \pi } } ^ { * } ( s ^ { \prime } \vert s , \bar { U } ) ^ { 1 }$ for guidance when generating the candidates. The final choice for next move will be selected from these candidates with max $\phi \left( s \right)$ defined in (3). As explained in more details in Section 4.2, $\tilde { \pi } ^ { * }$ will be trained to mimic previous successful planning experiences across different problems, that is, biasing the sampling towards the states with higher successful probability. + +![](images/659fdd9f888e78d83e6532bb1c8c3abf8ec8cc26808c966e6aef6311a6ac402b.jpg) +Figure 2: Our neural network model maps a $N$ -link robot from the original planning space (a $( N + 2 )$ -d configuration space) to a 3d discrete latent planning space in which we plan a path using value iteration. The result of value iteration is then used as features for defining $\tilde { V } ^ { * } \left( s | U \right)$ and $\tilde { \pi } ^ { * } \left( s ^ { \prime } | s , U \right)$ . + +With these detailed step (i) and (ii), we obtain NEXT :: Expand in Algorithm 2 (illustrated in Figure 1(b) and (c)). Plugging it into the TSA in 1, we construct $\tt a l g \left( \cdot \right) \in \mathcal { A }$ which will be learned. + +The guided progressive expansion bears similarity to MCTS but deals with high dimensional continuous spaces. Moreover, the essential difference lies in the way to select state in $\tau$ for expansion: the MCTS only expands the leaf states in $\tau$ due to the hierarchical assumption, limiting the exploration ability and incurring extra unnecessary UCB sampling for internal traversal in $\tau$ ; while the proposed operation enables expansion from each visited state, particularly suitable for path planning problems. + +# 4.2 NEURAL ARCHITECTURE FOR VALUE FUNCTION AND EXPANSION POLICY + +In this section, we will introduce our neural architectures for $\tilde { V } ^ { * } \left( s | U \right)$ and $\tilde { \pi } ^ { * } \left( s ^ { \prime } | s , U \right)$ used in NEXT :: Expand. The proposed neural architectures can be understood as first embedding the state and problem into a discrete latent representation via an attention-based module in Section 4.2.1, upon which the neuralized value iteration, introduced in Section 4.2.2, is performed to extract features for defining $\tilde { V } ^ { * } \left( s | U \right)$ and $\tilde { \pi } ^ { * } \left( s ^ { \prime } | s , U \right)$ , as illustrated in Figure 2. + +# 4.2.1 CONFIGURATION SPACE EMBEDDING + +Our network for embedding high-dimension configuration space into a latent representation is designed based on an attention mechanism. More specifically, let $s ^ { w }$ denote the workspace in state and ${ \bf \bar { \Phi } } _ { s } h$ denote the remaining dimensions of the state, i.e. $s = \mathbf { \bar { \rho } } ( s ^ { w } , s ^ { h } )$ . $s ^ { w }$ and $s ^ { h }$ will be embedded by different sub-neural networks and combined for the final representation, as illustrated in Figure 3. For simplicity of exposition, we will focus on the 2d workspace, i.e., $s ^ { w } \in \mathbb { R } ^ { 2 }$ . However, our method applies to 3d workspace as well. + +![](images/1f9670de482ed991ba97956d7ebba8ef1a8b77c6a4180f100db30cdc84a6e86c.jpg) +Figure 3: Attention-based state embedding module. $\boldsymbol { s } ^ { w } = ( x , y )$ and $\mathbf { z } = s ^ { h }$ . The upper part is spatial attention, with the first two channels being $_ x$ and $_ y$ , and the last two channels being constant templates with the row and column coordinates, as shown with a $d$ set to 3. The bottom module learns the representation for $\mathbf { z }$ . The final embedding is obtained by outer-product of these two attention parts. + +• Spatial attention. The workspace information $s ^ { w }$ will be embedded as $\mu ^ { w } ( s ^ { w } ) \in \mathbb { R } ^ { d \times d }$ , $d$ is a hyperparameter related to map (see remark below). The spatial embedding module (upper part in Figure 3) is composed of $k _ { w }$ convolution layers, i.e., + +$$ +\mu ^ { w } ( s ^ { w } ) = { \tt s o f t m a x 2 d } ( f _ { k _ { w } } ^ { w } ( s ^ { w } ) ) , \qquad f _ { i + 1 } ^ { w } ( s ^ { w } ) = { \tt r e l u } ( \theta _ { i } ^ { w } \oplus f _ { i } ^ { w } ( s ^ { w } ) ) , +$$ + +where $\theta _ { i } ^ { w }$ denotes the convolution kernels, $\oplus$ denotes the convolution operator and $f _ { i } ^ { w } ( s ^ { w } ) \in$ $\mathbb { R } ^ { d \times d \times d _ { i } }$ with $d _ { i }$ channels. The first layer $f _ { 0 } ^ { w }$ is designed to represent $s ^ { w }$ into a $d \times d \times 4$ tensor as $f _ { 0 } ^ { w } ( s ^ { w } ) _ { i j } = [ s _ { 1 } ^ { w } , s _ { 2 } ^ { w } , i , j ] , i , j = 1 , \ldots , a$ 0, without any loss of information. + +![](images/e690ebd6ddc08b67ba7f0c4c102d2ba25dbff64c677153964eecb7a40240d491.jpg) +Figure 4: Overall model architecture. Current and goal states are embedded through attention module. Then the embedding of the goal state is concatenated with the map to produce $\nu ^ { * ( 0 ) }$ and $\bar { \tilde { \boldsymbol { R } } }$ as the input to the planning module. The output of the planning module is aggregated with the embedding of the current state to produce feature $\psi ( s )$ for defining $\tilde { V } ^ { \ast }$ and $\tilde { \pi } ^ { * }$ . + +• Configuration attention. The remaining configuration state information will be embedded as $\mu ^ { h } ( s ^ { h } )$ through $k _ { h }$ fully-connected layers (bottom-left part in Figure 3), i.e. + +$$ +\mu ^ { h } ( s ^ { h } ) = \mathsf { s o f t m a x } ( f _ { k _ { h } } ^ { h } ( s ^ { h } ) ) , \qquad f _ { i + 1 } ^ { h } ( s ^ { h } ) = \mathtt { r e l u } ( \theta _ { i } ^ { h } f _ { i } ^ { h } ( s ^ { h } ) + b _ { i } ) , +$$ + +where $\mu ^ { h } ( s ^ { h } ) \in \mathbb { R } ^ { d _ { a } }$ and $f _ { 0 } ^ { h } ( s ^ { h } ) = s ^ { h }$ . + +The final representation $\mu _ { \boldsymbol { \theta } } ( s )$ will be obtained by multiplying $\mu ^ { w } ( s ^ { w } )$ with $\mu ^ { h } ( s ^ { h } )$ element-wisely, $\mu _ { \boldsymbol \theta } ( s ) _ { i j l } = \bar { \mu } ^ { w } ( s ^ { w } ) _ { i j } \cdot \bar { \mu } ^ { h } ( \bar { s } ^ { h } ) _ { l }$ , which is a $d \times d \times d _ { a }$ tensor attention map with $\mu _ { \theta } ( s ) _ { i j l } \geqslant$ 0, and $\begin{array} { r } { \sum _ { i j l } \mu _ { \theta } ( s ) _ { i j l } = 1 } \end{array}$ (bottom-right part in Figure 3). Intuitively, one can think of $d _ { a }$ as the level of the learned discretization of the configuration space $s ^ { h }$ , and the entries in softly assign the actual state $s$ to these discretized locations. $\theta : = ( \{ \theta _ { i } ^ { w } \} _ { i = 0 } ^ { k _ { w } - 1 } , \{ \theta _ { i } ^ { h } , b _ { i } \} _ { i = 0 } ^ { k _ { h } - 1 } )$ are the parameters to be + +Remark (different map size): To process map using convolution neural networks, we resize it to a $d \times d$ image, with the same size as the spatial attention in Eq (4), where $d$ is a hyperparameter. + +# 4.2.2 NEURAL VALUE ITERATION + +We then apply neuralized value iteration on top of the configuration space embedding to extract further planning features (Tamar et al., 2016; Lee et al., 2018). Specifically, we first produce the embedding $\mu _ { \theta } \big ( s _ { g o a l } \big )$ of the center of the goal region $s _ { g o a l }$ . We execute $T$ steps of neuralized Bellman updates (planning module in Figure 4) in the embedding space, + +$$ +\nu ^ { * ( t ) } = \operatorname* { m i n } \left( W _ { 1 } \oplus \left[ \nu ^ { * ( t - 1 ) } , \tilde { R } \right] \right) , \quad \mathrm { w i t h } \quad \left( \nu ^ { * ( 0 ) } , \tilde { R } \right) = \sigma \left( W _ { 0 } \oplus \left[ \mu _ { \theta } \big ( s _ { g o a l } \big ) , \mathtt { m a p l } \right) \right) , +$$ + +and obtain $\nu ^ { * ( T ) } \in \mathbb { R } ^ { d \times d \times d _ { a } \times p }$ . Both $W _ { 0 }$ and $W _ { 1 }$ are 3d convolution kernels, min implements the pooling across channels. Accordingly, $\nu ^ { * ( T ) }$ now can be understood as a latent representation of the value function $\tilde { V } ^ { * } ( \cdot )$ in learned embedding space for the problem $U$ with $s _ { g o a l }$ in map. + +To define the value function for particular state $s$ , i.e., $\tilde { V } ^ { * } \left( s | U \right)$ , from the latent representation $\nu ^ { * ( T ) }$ , we first construct another attention model between the embedding of state $s$ using $\mu _ { \boldsymbol { \theta } } ( s )$ and $\nu ^ { * ( T ) }$ , i.e., $\begin{array} { r } { \psi \left( s \right) _ { k } = \sum _ { i j l } \nu _ { i j l k } ^ { * ( T ) } \cdot \mu _ { \theta } \left( s \right) _ { i j l } } \end{array}$ , for $k = 1 , \hdots , p$ . Finally we define + +$$ +\tilde { V } ^ { \ast } \left( s | U \right) = h _ { W _ { 2 } } \left( \psi ( s ) \right) , \quad \mathrm { a n d } \quad \tilde { \pi } ^ { \ast } \left( s ^ { \prime } | s , U \right) = \mathcal { N } \left( h _ { W _ { 3 } } \left( \psi ( s ) \right) , \sigma ^ { 2 } \right) +$$ + +where $h _ { W _ { 2 } }$ and $h _ { W _ { 3 } }$ are fully connected dense layers with parameters $W _ { 2 }$ and $W _ { 3 }$ respectively, and $\mathcal { N } ( h _ { W _ { 3 } } \left( \psi ( s ) \right) , \sigma ^ { 2 } )$ is a Gaussian distribution with variance $\sigma ^ { 2 }$ . Note that we also parametrize the policy $\tilde { \pi } ^ { * } \left( s ^ { \prime } | s , U \right)$ using the embedding $\nu ^ { * ( T ) }$ , since the policy is connected to the value function via $\begin{array} { r } { \pi ^ { * } \left( s ^ { \prime } | s , U \right) = \mathrm { a r g m i n } _ { s ^ { \prime } \in S } c \left( \left[ s , s ^ { \prime } \right] \right) + V ^ { * } \left( s ^ { \prime } | U \right. } \end{array}$ ). It should also be emphasized that in our parametrization, the calculation of $\nu ^ { * ( T ) }$ only relies on the $\mu _ { \theta } \left( s _ { g o a l } \right)$ , which can be reused for evaluating $\tilde { V } ^ { * } \left( s | U \right)$ and $\tilde { \pi } ^ { * } \left( s ^ { \prime } | s , U \right)$ over different $s$ , saving computational resources. Using this trick the algorithm runs $1 0 \times - 1 0 0 \times$ faster empirically. + +The overall model architecture in $\mathsf { a l g } \left( \cdot \right)$ is illustrated in Figure 4. The parameters $W =$ $( W _ { 0 } , W _ { 1 } , W _ { 2 } , W _ { 3 } , \theta )$ will be learned together by our meta self-improving learning. For the details of the parameterization and the size of convolution kernels in our implementation, please refer to Figure 12 in Appendix D. + +# 4.3 META SELF-IMPROVING LEARNING + +The learning of the planner $\mathsf { a l g } \left( \cdot \right)$ reduces to learning the parameters in $\tilde { V } ^ { * } \left( s | U \right)$ and $\tilde { \pi } ^ { * } \left( s ^ { \prime } | s , U \right)$ and is carried out while planning experiences accumulate. We do not have an explicit training and testing phase separation. Particularly, we use a mixture of RRT :: Expand and NEXT :: Expand with probability $\epsilon$ and $1 - \epsilon$ , respectively, inside the TSA framework in Algorithm 1. The RRT∗ postprocessing step is used in the template. The $\epsilon$ is set to be 1 initially since $\{ \tilde { V } ^ { * } , \tilde { \pi } ^ { * } \}$ are not well-trained, and thus, the algorithm behaves like $\mathbf { R R T ^ { * } }$ . As the training proceeds, we anneal $\epsilon$ gradually as the sampler becomes more efficient. + +The dataset $\mathcal { D } _ { n } = \{ ( \mathcal { T } _ { j } , U _ { j } ) \} _ { j = 1 } ^ { n }$ for the $n$ -th training epoch is collected from the previous successful planning experiences across multiple random problems. We fix the size of dataset and update $\mathcal { D }$ in the same way as experience reply buffer (Lin, 1992; Schaul et al., 2015). For an experience $( { \mathcal { T } } , U ) \in { \mathcal { D } } _ { n }$ , we can reconstruct the successful path $\{ s ^ { i } \} _ { i = 1 } ^ { m }$ from the search tree $\tau$ ( $\mathrm { ~ m ~ }$ is the region, i.e., number of segments), and the value of each state $\begin{array} { r } { y ^ { i } : = \sum _ { l = i } ^ { m - 1 } c ( [ s ^ { l } , s ^ { l + 1 } ] ) } \end{array}$ . We learn $\{ \tilde { V } ^ { * } , \tilde { \pi } ^ { * } \}$ $s ^ { i }$ in the path will be the sum of cost to the goal by optimizing objective + +$$ +\operatorname* { m i n } _ { W } \sum _ { ( { \cal T } , { \cal U } ) \in { \mathcal D } _ { n } } \ell ( \tilde { V } ^ { * } , \tilde { \pi } ^ { * } ; { \cal T } , { \cal U } ) : = - \sum _ { { \mathcal D } _ { n } } \sum _ { i = 1 } ^ { m - 1 } \log \tilde { \pi } ^ { * } \left( s ^ { i + 1 } | s ^ { i } \right) + \sum _ { i = 1 } ^ { m } ( \tilde { V } ^ { * } \left( s ^ { i } \right) - y ^ { i } ) _ { 2 } ^ { 2 } + \lambda \left\| { \cal W } \right\| ^ { 2 } . +$$ + +The loss (6) pushes the ${ \tilde { V } } ^ { * }$ and $\tilde { \pi } ^ { * }$ to chase the successful trajectories, providing effective guidance in $\mathsf { a l g } \left( \cdot \right)$ for searching, and therefore leading to efficient searching procedure with less sample complexity and better solution. On one hand, the value function and policy estimation $\{ \tilde { V } ^ { * } , \tilde { \pi } ^ { * } \}$ is improved based upon the successful outcomes from NEXT itself on previous problems. On the other hand, the updated $\{ \tilde { V } ^ { * } , \bar { \pi } ^ { * } \}$ will be applied in the next epoch to improve the performance. Therefore, the training is named as Meta Self-Improving Learning (MSIL). Since all the trajectories we collected for learning are feasible, the reachability of the proposed samples is enforced implicitly via imitating these successful paths. + +By putting every components together into the 10 learning to plan framework in Eq (2), the over + +# Algorithm 3: Meta Self-Improving Learning + +1 Initialize dataset $\mathcal { D } _ { 0 }$ ; +$^ 2$ for epoch $n \gets 1$ to $N$ do +$^ 3$ Sample a planning problem $U$ ; +4 $\mathcal { T } \mathrm { T S A } ( U )$ with $\epsilon \sim \mathcal { U } n i f [ 0 , 1 ]$ , and $\epsilon$ ·RRT :: Expand $+ ( 1 - \epsilon )$ ·NEXT :: Expand; Postprocessing with $\mathbb { R R T } ^ { * }$ :: Postprocess; +5 ${ \mathcal { D } } _ { n } \gets { \mathcal { D } } _ { n - 1 } \bar { \cup } \{ ( { \mathcal { T } } , U ) \}$ if successful else $\mathcal { D } _ { n - 1 }$ ; +6 for $j 0$ to $L$ do +7 Sample Reconst $( \tau _ { j } , U _ { j } )$ from -optim $\mathcal { D } _ { n }$ ;path $\{ s ^ { i } \} _ { i = 1 } ^ { m }$ and the cost of paths based on $\tau _ { j }$ ; +9 Update parameters $W W - \eta \nabla _ { W } \ell ( \tilde { V } ^ { * } , \tilde { \pi } ^ { * } ; T _ { j } , U _ { j } ) ;$ Anneal $\epsilon = \alpha \epsilon$ , $\alpha \in ( 0 , 1 )$ ; + +all procedure is summarized in Algorithm 3 and illustrated in Figure 1. + +# 5 EXPERIMENTS + +In this section, we evaluate the proposed NEXT empirically on different planning tasks in a variety of environments. Comparing to the existing planning algorithms, NEXT achieves the state-of-the-art performances, in terms of both success rate and the quality of the found solutions. We further demonstrate the power of the proposed two components by the corresponding ablation study. We also include a case study on a real-world robot arm control problem at the end of the section. + +# 5.1 EXPERIMENT SETUP + +Benchmark environments. We designed four benchmark tasks to demonstrate the effectiveness of our algorithm for high-dimensional planning. The first three involve planning in a 2d workspace with a 2 DoF (degrees of freedom) point robot, a 3 DoF stick robot and a 5 DoF snake robot, respectively. The last one involves planning a 7 DoF spacecraft in a 3d workspace. For all problems in each benchmark task, the workspace maps were randomly generated from a fixed distribution; the initial and goal states were sampled uniformly randomly in the free space; the cost function $c \left( \cdot \right)$ was set as the sum of the Euclidean path length and the control effort penalty of rotating the robot joints. + +Baselines. We compared NEXT with RRT∗ (Karaman & Frazzoli, 2011), $\mathrm { B I T ^ { * } }$ (Gammell et al., 2015), CVAE-plan (Ichter et al., 2018), Reinforce-plan (Zhang et al., 2018), and an improved version of GPPN (Lee et al., 2018) in terms of both planning time and solution quality. RRT∗ and $\mathrm { B I T ^ { * } }$ are two widely used effective instances of TSA in Algorithm 1. In our experiments, we equipped $\mathrm { R R T ^ { * } }$ with the goal biasing heuristic to improve its performance. $\mathrm { B I T ^ { * } }$ adopts the informed search strategy (Gammell et al., 2015) to accelerate planning. CVAE-plan and Reinforce-plan are two learning-enhanced TSA planners proposed recently. CVAE-plan learns a conditional VAE as the sampler (Sohn et al., 2015), which will be trained by near-optimal paths produced by $\mathbf { R R T ^ { * } }$ . Reinforce-plan learns to do rejection sampling with policy gradient methods. For the improved GPPN, we combined its architecture for map with a fully-connected MLP for the rest state, such that it can be applied to high-dimensional continuous spaces. Please refer to Appendix E for more details. + +![](images/b0b18ba9b0a2a21dab4fe0925d12969b8bbd06feb8546ddafdca1a81131ba122.jpg) +Figure 5: Search trees and the learned ${ \tilde { V } } ^ { * }$ and $\tilde { \pi } ^ { * }$ produced by NEXT. Obstacles are colored in blue. The start and goal locations are denoted by orange and brown dots. In (a) to (c), samples are represented with yellow circles. In (d), the level of redness denotes the value of the cost-to-go estimate $\tilde { V } ^ { * }$ . The cyan arrows point from a given state $s$ to the mean of the learned policy $\tilde { \pi } ^ { * } ( s ^ { \prime } | s , U )$ . + +![](images/e9b64d33b209ec3b631b087dbd2e5d1f20da83a228d5b5ee93466c67d2fe104a.jpg) +Figure 6: Search trees and a solution path produced in an instance of spacecraft planning. The 7 DOF spacecraft has a yellow body and two 2 DOF red arms. NEXT-KS produced a nearly minimum viable search tree while ${ \mathrm { R R T } } ^ { * }$ failed to find a path within limited trials. + +Settings. For each task, we randomly generated 3000 different problems from the same distribution without duplicated maps. We trained all learning-based baselines using the first 2000 problems, and reserved the rest for testing. The parameters for $\mathrm { { R R T ^ { * } } }$ and $\mathrm { B I T ^ { * } }$ are also tuned using the first 2000 problems. For NEXT, we let it improve itself using MSIL over the first 2000 problems. In this period, for every 200 problems, we updated its parameters and annealed $\epsilon$ once. + +# 5.2 RESULTS AND ANALYSIS + +Comparison results. Examples of all four environments are illustrated in Appendix F.1 and Figure 6, where NEXT finds high-quality solutions as shown. We also illustrated the comparison of the search trees on two 2d and 7d planning tasks between NEXT and RRT\* in Figure 5 (a)-(c) and Figure 6 (b) and (c). Obviously, the proposed NEXT algorithm explores with guidance and achieves better quality solutions with fewer samples, while the RRT\* expands randomly which may fail to find a solution. The learned ${ \tilde { V } } ^ { * }$ and $\tilde { \pi } ^ { * }$ in the 2d task are also shown in Figure 5(d). As we can see, they are consistent with our expectation, towards the ultimate target in the map. For more search tree comparisons for all four experiments, please check Figure 18, 19, 20 and 21 in Appendix F. + +To systematically evaluate the algorithms, we recorded the cost of time (measured by the number of collision checks used) to find a collision-free path, the success rate within time limits, and the cost of the solution path for each run. The results of the reserved 1000 test problems of each environment are shown in the top row of Figure 7. We set the maximal number of samples as 500 for all algorithms. Both the kernel smoothing (NEXT-KS) and the Gaussian process (NEXT-GP) version of NEXT achieves the state-of-the-art performances, under all three criteria in all test environments. Although the $\mathrm { B I T ^ { * } }$ utilizes the heuristic particularly suitable for 3d maze in 7d task and performs quite well, the NEXT algorithm still outperform every competitor by a large margin, no matter learning-based or prefixed heuristic planner, demonstrating the advantages of the proposed NEXT algorithm. + +![](images/e1b7307ae6db488a13797cccd31ec60ba79f87159ac956c545b9f9d1f3fe33ce.jpg) +Figure 7: First row: histograms of results, in terms of success rate, average collision checks, and average cost of the solution paths; Second row: NEXT improvement curves in the 5d experiments. All algorithms are set to use up to 500 samples, except $\mathrm { R R T ^ { * } }$ -10k, which uses 10,000 samples. The value of collision checks and path costs are normalized w.r.t. the performance of $\mathrm { R R T ^ { * } }$ . + +Self-improving. We plot the performance improvement curves of our algorithms on the 5d planning task in the bottom row of Figure 7. For comparison, we also plot the performance of RRT∗. At the beginning phase of self-improving, our algorithms are comparable to $\mathrm { R R T ^ { * } }$ . They then gradually learn from previous experiences and improve themselves as they see more problems and better solutions. In the end, NEXT-KS is able to match the performance of RRT∗-10k using only one-twentieth of its samples, while the competitors perform consistently without any improvements. + +Due to the space limits, we put improvement curves on other environments in Figure 17 and the quantitative evaluation in Table 1, 2, and 3 in Appendix F. Please refer to the details there. + +# 5.3 ABLATION STUDIES + +Ablation study I: guided progressive expansion. To demonstrate the power of NEXT :: Expand, we replace it with breadth-first search (BFS) (Kim et al., 2018), another expanding strategy, while keeping other components the same. Specifically, BFS uses a search queue in planning. It repeatedly pops a state $s$ out from the search queue, samples $k$ states from $\pi \left( \cdot | s \right)$ , and pushes all generated samples and state $s$ back to the queue, until the goal is reached. For fairness, we use the learned sampling policy $\pi \left( s ^ { \prime } | s , U \right)$ by NEXT-KS in BFS. As shown in Figure 7, BFS obtained worse paths with a much larger number of collision checks and far lower success rate, which justifies the importance of the balance between exploration versus exploitation achieved by the proposed NEXT :: Expand. + +Ablation study II: neural architecture. To further demonstrate the benefits of the proposed neural architecture for learning generalizable representations in high-dimension planning problems, We replaced our attention-based neural architecture with an improved GPPN, as explained in Appendix E, for ablation study. We extended the GPPN for continuous space by adding an extra reactive policy network to its final layers. We emphasize the original GPPN is not applicable to the tasks in our experiments. Intuitively, the improved GPPN first produces a ‘rough plan’ by processing the robot’s discretized workspace positions. Then the reactive policy network predicts a continuous action from both the workspace feature and the full configuration state of the robot. We provide more preference to the improved GPPN by training it to imitate the near-optimal paths produced by RRT\* in the training problems. During test time it is also combined with both versions of the guided progressive expansion operators. As we can see, both GPPN-KS and GPPN-GP are clearly much worse than NEXT-KS and NEXT-GP, demonstrating the advantage of our proposed attention-based neural architecture in high-dimensional planning tasks. + +![](images/a7f19f8a9284429626e2ad6f69c231d6aa6a4e1e1021040a5d3d6919d297a607.jpg) +Figure 8: The success rate and average path cost of the different planners under varying time limits. Running NEXT for 1 second achieves the same success rate as running $\mathrm { B I T ^ { * } }$ for 50 seconds. + +Ablation study III: learning versus heuristic. The NEXT algorithm in Figure 5 shows similar behavior as the Dijkstra heuristic, i.e. sampling on the shortest path connecting the start and the goal in workspace. However, in higher dimensional space, the Dijkstra heuristic will fail. To demonstrate that, we replace the policy and value network with Dijkstra heuristic, using the best direction in workspace to guide sampling. NEXT performs much better than Dijkstra in all but the 2d case, in which the workspace happens to be the state space. + +# 5.4 CASE STUDY: ROBOT ARM CONTROL + +We conduct a real-world case study on controlling robot arms to move objects on a shelf. On this representative real-time task, we demonstrate the advantages of the NEXT in terms of the wall-clock. + +In each planning task, there is a shelf of multiple levels, with each level horizontally divided into multiple bins. The task is to plan a path from a location in one bin to another, i.e., the end effectors of the start and goal configurations are in different bins. The heights of levels, widths of bins, and the start and goal are randomly drawn from some fixed distribution. Different from previous experiments, the base of the robot is fixed. We consider the $\mathrm { B I T ^ { * } }$ instead of $\mathrm { R R T ^ { * } }$ as the imperfect expert in 3000 training problems. We then evaluate the algorithm on a separated 1000 testing problems. We compare NEXT(- KS) with the highly tuned $\mathrm { B I T ^ { * } }$ and RRT\* in OMPL, and also CVAE-plan and Reinforce-plan in Figure 8. As seen from the visualization of the found paths in Figure 9, this is a very difficult task. Our NEXT outperforms the baselines by a large margin, requiring only 1 second to reach the same success rate as running 50 seconds of $\mathrm { B I T ^ { * } }$ . + +![](images/2c27831a81842e20acfb22f33643b38f7e0d636dac989887fae6e6036e1293d0.jpg) +Figure 9: The collision-free path produced by NEXT for robot arm planning. The start and goal configurations have end-effectors in different bins of the shelf. + +Due to space limits, we put details of the experiment setups, more results and analysis in Appendix F.4. + +# 6 CONCLUSION + +In this paper, we propose a self-improving planner, Neural EXploration-EXploitation Trees (NEXT), which can generalize and achieve better performance with experiences accumulated. The algorithm achieves a delicate balance between exploration-exploitation via our carefully designed UCB-type expansion operation. To obtain the generalizable ability across different problems, we proposed a new parametrization for the value function and policy, which captures the Bellman recursive structure in the high-dimensional continuous state and action space. We demonstrate the power of the proposed algorithm by outperforming previous state-of-the-art planners with significant margins on planning problems in a variety of different environments. + +# ACKNOWLEDGEMENT + +We thank the Google Research Brain team members for helpful thoughts and discussions as well as the anonymous reviewers for their insightful comments and suggestions. This work is supported in part by NSF grants CDS&E-1900017 D3SC, CCF-1836936 FMitF, IIS-1841351, CAREER IIS1350983 to L.S, and by NSF grants BIGDATA 1840866, CAREER 1841569, TRIPODS 1740735, DARPA-PA-18-02-09-QED-RML-FP-003, an Alfred P Sloan Fellowship, a PECASE award to H.L. + +# REFERENCES + +Auer, P., Cesa-Bianchi, N., and Fischer, P. Finite-time analysis of the multiarmed bandit problem. Machine learning, 47(2-3):235–256, 2002. + +Mohak Bhardwaj, Sanjiban Choudhury, and Sebastian Scherer. Learning heuristic search via imitation. arXiv preprint arXiv:1707.03034, 2017. + +Boor, V., Overmars, M. H., and Van Der Stappen, A. F. The gaussian sampling strategy for probabilistic roadmap planners. In Robotics and automation, 1999. proceedings. 1999 ieee international conference on, volume 2, pp. 1018–1023. IEEE, 1999. + +Burns, B. and Brock, O. Sampling-based motion planning using predictive models. In Robotics and Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference on, pp. 3120–3125. IEEE, 2005a. + +Burns, B. and Brock, O. Toward optimal configuration space sampling. In Robotics: Science and Systems, pp. 105–112. Citeseer, 2005b. + +Bowen, Chris, and Ron Alterovitz. Closed-loop global motion planning for reactive execution of learned tasks. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1754-1760, 2014. + +Cadena Cesar, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J. Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on Robotics, 32(6), 2016. + +Sanjiban Choudhury, Mohak Bhardwaj, Sankalp Arora, Ashish Kapoor, Gireeja Ranade, Sebastian Scherer, and Debadeepta Dey. Data-driven planning via imitation learning. The International Journal of Robotics Research, 37(13-14):1632–1672, 2018. + +Chu, W., Li, L., Reyzin, L., and Schapire, R. Contextual bandits with linear payoff functions. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp. 208–214, 2011. + +Couëtoux, A., Hoock, J.-B., Sokolovska, N., Teytaud, O., and Bonnard, N. Continuous upper confidence trees. In International Conference on Learning and Intelligent Optimization, pp. 433–445. Springer, 2011. + +Rosen Diankov. Automated Construction of Robotic Manipulation Programs. PhD thesis, Carnegie Mellon University, Robotics Institute, August 2010. + +Rosen Diankov and James Kuffner. Randomized statistical path planning. In 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1–6. IEEE, 2007. + +Elbanhawi, M. and Simic, M. Sampling-based robot motion planning: A review. Ieee access, 2: 56–77, 2014. + +Finney, S., Kaelbling, L. P., and Lozano-Perez, T. Predicting partial paths from planning problem parameters. In Robotics Science and Systems, 2007. + +Fulgenzi, Chiara and Tay, Christopher and Spalanzani, Anne and Laugier, Christian Probabilistic navigation in dynamic environment using rapidly-exploring random trees and gaussian processes In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1056–1062, 2008. + +Gammell, J. D., Srinivasa, S. S., and Barfoot, T. D. Informed $\mathrm { { r r t } ^ { * } }$ : Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic. arXiv preprint arXiv:1404.2334, 2014. +Gammell, J. D., Srinivasa, S. S., and Barfoot, T. D. Batch informed trees (bit\*): Sampling-based optimal planning via the heuristically guided search of implicit random geometric graphs. In Robotics and Automation (ICRA), 2015 IEEE International Conference on, pp. 3067–3074. IEEE, 2015. +Guez, Arthur and Weber, Théophane and Antonoglou, Ioannis and Simonyan, Karen and Vinyals, Oriol and Wierstra, Daan and Munos, Rémi and Silver, David Learning to search with mctsnets arXiv preprint arXiv:1802.04697, 2018. +Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100–107, 1968. +Hsu, D., Latombe, J.-C., and Motwani, R. Path planning in expansive configuration spaces. In Robotics and Automation, 1997. Proceedings., 1997 IEEE International Conference on, volume 3, pp. 2719–2726. IEEE, 1997. +Hsu, D., Jiang, T., Reif, J., and Sun, Z. The bridge test for sampling narrow passages with probabilistic roadmap planners. In Robotics and Automation, 2003. Proceedings. ICRA’03. IEEE International Conference on, volume 3, pp. 4420–4426. IEEE, 2003. +Hsu, D., Sánchez-Ante, G., and Sun, Z. Hybrid prm sampling with a cost-sensitive adaptive strategy. In Robotics and Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference on, pp. 3874–3880. IEEE, 2005. +Huh, Jinwook and Lee, Daniel Efficient Sampling With Q-Learning to Guide Rapidly Exploring Random Trees. IEEE Robotics and Automation Letters, 3:3868–3875, 2018. +Ichter, B., Harrison, J., and Pavone, M. Learning sampling distributions for robot motion planning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 7087–7094. IEEE, 2018. +Karaman, S. and Frazzoli, E. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846–894, 2011. +Karkus, P., Hsu, D., and Lee, W. S. Qmdp-net: Deep learning for planning under partial observability. In Advances in Neural Information Processing Systems, pp. 4694–4704, 2017. +Kavraki, L. E., Svestka, P., Latombe, J.-C., and Overmars, M. H. Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics and Automation, 12(4), 1996. +Kim, B., Kaelbling, L. P., and Lozano-Pérez, T. Guiding search in continuous state-action spaces by learning an action sampler from off-target search experience. 2018. +Kocsis, L. and Szepesvári, C. Bandit based monte-carlo planning. In European conference on machine learning, pp. 282–293. Springer, 2006. +Krause, A. and Ong, C. S. Contextual gaussian process bandit optimization. In Advances in Neural Information Processing Systems, pp. 2447–2455, 2011. +Kuo, Yen-Ling and Barbu, Andrei and Katz, Boris Deep sequential models for sampling-based planning In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 6490–6497, 2018. +Langford, J. and Zhang, T. The epoch-greedy algorithm for multi-armed bandits with side information. In Advances in neural information processing systems, pp. 817–824, 2008. +LaValle, S. M. Rapidly-exploring random trees: A new tool for path planning. 1998. +Lee, L., Parisotto, E., Chaplot, D. S., Xing, E., and Salakhutdinov, R. Gated path planning networks. arXiv preprint arXiv:1806.06408, 2018. + +Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine learning, 8(3-4):293–321, 1992. + +Paxton, Chris and Raman, Vasumathi and Hager, Gregory D and Kobilarov, Marin Combining neural networks and tree search for task and motion planning in challenging environments In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 6059–6066, 2017. + +Sven Mikael Persson and Inna Sharf. Sampling-based a\* algorithm for robot path-planning. The International Journal of Robotics Research, 33(13):1683–1708, 2014. + +Phillips, J. M., Bedrossian, N., and Kavraki, L. E. Guided expansive spaces trees: A search strategy for motion-and cost-constrained state spaces. In IEEE International Conference on Robotics and Automation, pp. 3968–3973, 2004. + +Mike Phillips, Benjamin J Cohen, Sachin Chitta, and Maxim Likhachev. E-graphs: Bootstrapping planning with experience graphs. In Robotics: Science and Systems, volume 5, pp. 110, 2012. + +Qureshi, Ahmed H and Simeonov, Anthony and Bency, Mayur J and Yip, Michael C. Motion planning networks. In IEEE International Conference on Robotics and Automation, 2019. + +Reif, J. H. Complexity of the mover’s problem and generalizations. In Foundations of Computer Science, 1979., 20th Annual Symposium on, pp. 421–427. IEEE, 1979. + +Rickert, Markus and Brock, Oliver and Knoll, Alois. Balancing exploration and exploitation in motion planning In IEEE International Conference on Robotics and Automation, pp. 2812–2817. IEEE, 2008. + +Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. + +Shkolnik, A., Walter, M., and Tedrake, R. Reachability-guided sampling for planning under differential constraints. In Robotics and Automation, 2009. ICRA’09. IEEE International Conference on, pp. 2859–2865. IEEE, 2009. + +Silver, David and Hubert, Thomas and Schrittwieser, Julian and Antonoglou, Ioannis and Lai, Matthew and Guez, Arthur and Lanctot, Marc and Sifre, Laurent and Kumaran, Dharshan and Graepel, Thore and others Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint arXiv:1712.01815, 2017. + +Sohn, K., Lee, H., and Yan, X. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems, pp. 3483–3491, 2015. + +Jialin Song, Ravi Lanka, Albert Zhao, Yisong Yue, and Masahiro Ono. Learning to search via retrospective imitation. arXiv preprint arXiv:1804.00846, 2018. + +Srinivas, N., Krause, A., Kakade, S. M., and Seeger, M. Gaussian process optimization in the bandit setting: No regret and experimental design. arXiv preprint arXiv:0912.3995, 2009. + +Ioan A. ¸Sucan, Mark Moll, and Lydia E. Kavraki. The Open Motion Planning Library. IEEE Robotics & Automation Magazine, 19(4):72–82, December 2012. doi: 10.1109/MRA.2012.2205651. http://ompl.kavrakilab.org. + +Tamar, A., Wu, Y., Thomas, G., Levine, S., and Abbeel, P. Value iteration networks. In Advances in Neural Information Processing Systems, pp. 2154–2162, 2016. + +Wang, Y., Audibert, J.-Y., and Munos, R. Algorithms for infinitely many-armed bandits. In Advances in Neural Information Processing Systems, pp. 1729–1736, 2009. + +Ye, Gu and Alterovitz, Ron Guided motion planning. In Robotics research, pp. 291–6307. Springer, 2017. + +Yee, T., Lisy, V., and Bowling, M. Monte carlo tree search in continuous action spaces with execution uncertainty. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, pp. 690–696. AAAI Press, 2016. + +Zhang, C., Huh, J., and Lee, D. D. Learning implicit sampling distributions for motion planning. arXiv preprint arXiv:1806.01968, 2018. + +Zucker, M., Kuffner, J., and Bagnell, J. A. Adaptive workspace biasing for sampling-based planners. In Robotics and Automation, 2008. ICRA 2008. IEEE International Conference on, pp. 3757–3762. IEEE, 2008. + +# Appendix + +# A ILLUSTRATION OF THE DIFFICULTY IN PLANNING PROBLEMS + +The Figure 10(a) illustrates a concrete planning problem for a stick robot in 2d workspace. With one extra continuous action for rotation, the configuration state is visualized in Figure 10(b), which is highly irregular and unknown to the planner. + +![](images/03f4d274b006a2d945a961a75df61f65f470c9506bae0cc5429f4ef3d09ead9c.jpg) +Figure 10: Different views of the same planning problem. In (a) we color the obstacles, the starting and the goal position of the robot in deep blue, orange and brown, respectively. The stick robot can move and rotate. The corresponding configuration space is 3d, as visualized in (b), with the extra dimension being the rotation angle w.r.t. the $\mathbf { X }$ -axis. The blue region indicates the feasible state space, i.e., the set of collision-free states. The starting and the goal position are denoted with an orange and a brown dot, respectively. Although the workspace looks trivial, the configuration space is irregular, which makes the planning difficult. + +# B MORE PRELIMINARIES + +Tree-based sampling planner The tree-based sampling planner algorithm is illustrated in Figure 11. The Expand in Algorithm 1 operator returns an existing node in the tree $s _ { p a r e n t } \in \mathcal { V }$ and a new state $s _ { n e w } \in S$ sampled from the neighborhood of $s _ { p a r e n t }$ . Then the line segment $[ s _ { p a r e n t } , s _ { n e w } ]$ is passed to function ObstacleFree for collision checking. If the line segment $[ s _ { p a r e n t } , s _ { n e w } ]$ is collision-free (no obstacle in the middle, or called reachable from $\tau$ ), then $s _ { n e w }$ is added to the tree vertex set $\nu$ , and the line segment is added to the tree edge set $\mathcal { E }$ . If the newly added node $s _ { n e w }$ has reached the target $\mathcal { S } _ { g o a l }$ , the algorithm will return. Optionally, some concrete algorithms can define a Postprocess operator to refine the search tree. For an example of the Expand operator, as shown in Figure 1 (c), since there is no obstacle on the dotted edge $[ s _ { p a r e n t } , s _ { n e w } ]$ , i.e., $s _ { n e w }$ is reachable, the new state and edge will be added to the search tree (connected by the solid edges). + +![](images/be00082f31c012e5df15315075e706a96419afcb7dbc1ba05cf0204fbd263e3d.jpg) +Figure 11: Illustration for one iteration of Algorithm 1. The left and right figures illustrate two different cases where the sample returned by the Expand operator is unreachable and reachable from the search tree. + +Now we will provide two concrete algorithm examples. For instance, • If we instantiate the Expand operator as Algorithm 4, then we obtain the rapidly-exploring random trees (RRT) algorithm (LaValle, 1998), which first samples a state $s$ from the configuration space + +# $\mathbf { A l g o r i t h m 4 : R R T : : E x p a n d } ( { T } , { U } )$ + +Data: $\mathcal { T } = ( \mathcal { V } , \mathcal { E } ) , U = \left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \mathtt { m a p } , c ( \cdot ) \right)$ 1 $s _ { r a n d } \gets \mathcal { U } n i f ( \mathcal { S } )$ ; 2 $\begin{array} { r } { s _ { p a r e n t } \gets \operatorname * { a r g m i n } _ { s \in \mathcal { V } } \left\| s _ { r a n d } - s \right\| } \end{array}$ ; 3 $s _ { n e w } \gets \mathrm { a r g m i n } _ { s \in \mathcal { B } ( s _ { p a r e n t } , \eta ) } \left\| s - s _ { r a n d } \right\|$ ; + +4 return $s _ { p a r e n t } , s _ { n e w }$ . Sample configuration space . Pull to a tree node + +# $\mathbf { A l g o r i t h m 5 : E S T : : E x p a n d } ( T , U )$ + +Data: $\mathcal { T } = ( \mathcal { V } , \mathcal { E } ) , U = \left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \mathtt { m a p } , c ( \cdot ) \right)$ ，1 sparent $\sim \phi ( s ) , s \in \mathcal { V }$ ;2 $\hat { s _ { n e w } } \mathcal { U } n i f ( B ( s _ { p a r e n t } ) )$ ;3 return snearest, snew; + +. Sample a tree node . Sample neighborhood + +$s$ and then pulls it toward the neighborhood of current tree $\tau$ measured by a ball of radius $\eta$ + +Moreover, if the Postprocess operator is introduced to modify the maintained search tree as in $\mathrm { { R R T ^ { * } } }$ (Karaman & Frazzoli, 2011), the algorithm is provable to obtain the optimal path asymptotically. + +• If we instantiate the Expand operator as Algorithm 5, then we obtain the expansive-space trees (EST) algorithm (Hsu et al., 1997; Phillips et al., 2004), which samples a state $s$ from the nodes of the existing tree, and then draw a sample from the neighborhood of $s$ . + +UCB-based algorithms Specifically, in a $K$ -armed bandit problem, the UCB algorithm will first play each of the arms once, and then keep track of the average reward $\bar { r } _ { i }$ and the visitation count $n _ { i }$ for each arm. After $T$ rounds of trials, the UCB algorithm will maintain a set of information $\{ ( \bar { r } _ { i } , n _ { i } ) \} _ { i = 1 } ^ { K }$ with $\textstyle \sum _ { i = 1 } ^ { K } n _ { i } = T$ . Then, for the next round, the UCB algorithm will select the next arm based on the one-sided confidence interval estimation provided by the Chernoff-Hoeffding bound, + +$$ +\begin{array} { r } { a _ { T + 1 } ^ { * } = \operatorname * { a r g m a x } _ { i \in \{ 1 , \ldots , K \} } \bar { r } _ { i } + \lambda \sqrt { \frac { \log T } { n _ { i } } } , } \end{array} +$$ + +where $\lambda$ controls the exploration-exploration trade-off. It has been shown that the UCB algorithm achieves $\mathcal { O } \left( \log T \right)$ regret. However, the MCTS is not directly applicable to continuous state-action spaces. + +There have been many attempts to generalize the UCB and UCT algorithms to continuous state-action spaces (Chu et al., 2011; Krause & Ong, 2011; Couëtoux et al., 2011; Yee et al., 2016). For instance, contextual bandit algorithms allow continuous arms but involve a non-trivial high dimensional non-convex optimization to select the next arm. In UCT, the progressive widening technique has been designed to deal with continuous actions (Wang et al., 2009). Even with these extensions, the MCTS restricts the exploration only from leaves states, implicitly adding an unnecessary hierarchical structure for path planning, resulting inferior exploration efficiency and extra computation in path planning tasks. + +Although these off-the-shelf algorithms are not directly applicable to our path planning setting, their successes show the importance of exploration-exploitation trade-off and will provide the principles for our algorithm for continuous state-action planning problems. + +Planning networks Value iteration networks (Tamar et al., 2016) employ neural networks to embed the value iteration algorithm from planning and then use this embedded algorithm to extract input features and define downstream models such as value functions and policies. + +Specifically, VIN mimics the following recursive application of Bellman update operator $\mathcal { G }$ to value function $V ^ { * }$ , + +$$ +V ^ { * } \left( s | U \right) = \left( \mathcal { G } V ^ { * } \right) ( s ) : = \operatorname* { m i n } _ { a } \sum _ { s ^ { \prime } } P ( s ^ { \prime } | s , a ) ( c \left( \left[ s , s ^ { \prime } \right] \right) + V ^ { * } \left( s ^ { \prime } | U \right) ) . +$$ + +where $P ( s ^ { \prime } | s , a )$ is the state transition model. When the state space for $s$ and action space for $a$ are low dimensional, these spaces can be discretized into grids. Then, the local cost function $c ( [ s , s ^ { \prime } ] )$ and the value function $V ^ { * } ( s ^ { \prime } | U )$ can be represented as matrices (2d) or tensors (3d) with each entry indexed by grid locations. Furthermore, if the transition model $P ( s ^ { \prime } | s , a )$ is local, that is $P ( s ^ { \prime } | \dot { s } , a ) = 0$ for $\bar { \boldsymbol { s } } ^ { \prime } \notin B ( \boldsymbol { s } )$ , it resembles a set of convolution kernels, each indexed by a discrete action $a$ . And the Bellman update operator essentially convolves $P ( s ^ { \prime } | s , a )$ with $c ( [ s , s ^ { \bar { \prime } } ] )$ and $V ^ { * } ( s ^ { \prime } | U )$ , and then performs a min-pooling operation across the convolution channels. + +Inspired by the above computation pattern of the Bellman operator, value iteration networks design the neural architecture as follows, + +$$ +\begin{array} { r l } & { \tilde { V } ^ { * 0 } = \operatorname* { m i n } \ \left( W _ { 1 } \oplus \left[ \tt { m a p } , \tilde { R } \right] \right) } \\ & { \tilde { V } ^ { * t } = \operatorname* { m i n } \ \left( W _ { 1 } \oplus \left[ \tilde { V } ^ { * t - 1 } , \tilde { R } \right] \right) } \end{array} +$$ + +where $\oplus$ is the convolution operation, both map, $\tilde { V } ^ { * t }$ and $\tilde { R }$ are $d \times d$ matrices, and the parameter $W _ { 1 }$ are $k _ { c }$ convolution kernels of size $k \times k$ . The min implements the pooling across $k _ { c }$ convolution channels. + +The gated path planning networks (GPPN) (Lee et al., 2018) improves the VIN by replacing the VIN cell (9) with the well-established LSTM update, i.e., + +$$ +\tilde { V } ^ { t } , \tilde { c } ^ { t } = \mathrm { L S T M } \left( \sum \left( W _ { 1 } \oplus \left[ \tilde { V } ^ { t } , \tilde { R } \right] \right) , \tilde { c } ^ { t } \right) , +$$ + +where the summation is taking over all the $k _ { c }$ convolution channels. + +After constructing the VIN and GPPN, the parameters of the model, i.e., $\left\{ \tilde { R } , W _ { 1 } \right\}$ can be learned by imitation learning or reinforcement learning. + +The application of planning networks are restricted in low-dimension tasks. However, their success enlightens our neural architecture for generalizable representation for high-dimension planning tasks. + +# C PARAMETRIZED UCB ALGORITHMS + +We list two examples of parametrized UCB as the instantiation of (3) used in GPE: + +• GP-UCB: The GP-UCB Chu et al. (2011) is derived by parameterizing via Gaussian Processes (GP) with kernel $k \left( s , s ^ { \prime } \right)$ , i.e., E $[ r ( s ) | \mathcal { T } , U ] \sim \mathcal { \bar { G } } \mathcal { P } ( 0 , k )$ , GP-UCB maintains an UCB of the reward after t-step asφ (s) := ¯rt (s) + λσt (s) , (12) + +where + +$$ +\begin{array} { r c l } { { { \bar { r } } _ { t } } \left( s \right) } & { { = } } & { { { k _ { t } } \left( s \right) \left( K _ { t } + \alpha I \right) ^ { - 1 } { r _ { t } } , } } \\ { { \sigma _ { t } ^ { 2 } \left( s \right) } } & { { = } } & { { { k \left( s , s \right) - k _ { t } } \left( s \right) ^ { \top } \left( K _ { t } + \alpha I \right) ^ { - 1 } k _ { t } \left( s \right) , } } \end{array} +$$ + +with $\boldsymbol { k } _ { t } \left( \boldsymbol { s } \right) = \left[ k \left( s _ { i } , \boldsymbol { s } \right) \right] _ { s _ { i } \in \boldsymbol { S } _ { t } }$ , $K _ { t } = [ k \left( s , s ^ { \prime } \right) ] _ { s , s ^ { \prime } \in S _ { t } }$ , and ${ \cal { S } } _ { t } = \{ s _ { 1 } , s _ { 2 } , \ldots , s _ { t } \}$ denotes the sequence of selected nodes in current trees. The variance estimation $\sigma _ { t } ^ { 2 } \left( s \right)$ takes the number of visits into account in an implicit way: the variance will reduce, as the neighborhood of $s$ is visited more frequently (Srinivas et al., 2009). + +• KS-UCB: We can also use kernel regression as an alternative parametrization for (7) (Yee et al., 2016), which leads to an UCB of the reward after $t$ -step as + +$$ +\phi \left( s \right) : = { \bar { r } } _ { t } \left( s \right) + \lambda \sigma _ { t } \left( s \right) , +$$ + +where + +$$ +\begin{array} { r l r } { \bar { r } _ { t } \left[ s \right] } & { = } & { \frac { \sum _ { s ^ { \prime } \in S _ { t } } k \left( s ^ { \prime } , s \right) r \left( s ^ { \prime } \right) } { \sum _ { s ^ { \prime } \in S _ { t } } k \left( s ^ { \prime } , s \right) } , } \\ { \sigma _ { t } \left( s \right) } & { = } & { \sqrt { \frac { \log \sum _ { s ^ { \prime } \in S _ { t } } w \left( s ^ { \prime } \right) } { w \left( s \right) } } , } \end{array} +$$ + +with $\begin{array} { r } { w \left( s \right) = \sum _ { s ^ { \prime } \in S _ { t } } k \left( s ^ { \prime } , s \right) } \end{array}$ . Clearly, the variance estimation is to promote exploration towards less frequently visited states. + +As we can see, in both two examples of the parametrized UCB, we parametrize the observed rewards, leading to generalizable UCB for increased states by considering the correlations. + +# D POLICY AND VALUE NETWORK ARCHITECTURE + +We explain the implementation details of the proposed parametrization for policy and value function. Figure 12 and Figure 13 are neural architectures for the attention module, the policy/value network, and the planning module, respectively. + +![](images/fb3272bba1630226b3a65a134ae6d6e537e4da69dcd4e45215968c4d3b3ba712.jpg) +Figure 12: Left: attention module, instantiating the Figure 3; Right: policy/value network, instantiating the Figure 4. + +In the figures, we use rectangle blocks to denote inputs, intermediate results and outputs, stadium shape blocks to denote operations, and rounded rectangle blocks to denote modules. We use different colors for different operations. In particular, we use blue for convolutional/LSTM layers, green for dense layers, and orange for anything else. For convolutional layers, "Conv $1 \times 1$ , 32, relu" denotes a layer with $1 \times 1$ kernels, 32 channels, followed by a rectified linear unit; for dense layers, "Dense, 64, relu" denotes a layer of size 64, followed by a rectified linear unit. + +The attention module (Figure 12-left) embeds a state to a $d \times d \times d _ { a }$ tensor. The planning module (Figure 13) is a one-step LSTM update which takes the result of a convolutional layer as input. Both the input and hidden size of the LSTM cell are $d _ { e }$ . All $d \times d$ locations share one set of parameters and are processed by the LSTM in one batch. + +The main architecture is illustrated in Figure 12-right. It takes maze map, state and goal as input, and outputs the action and the value. Refer to Section 4.2 for details for computing $\psi ( s )$ . In our experiments, we set the values of the hyper-parameters to be $( d , d _ { e } , d _ { a } , p ) = ( 1 \bar { 5 } , 6 4 , \bar { 8 } , 8 )$ . + +![](images/093f34c3bdcd9d32756f1eaf06749627ebe9c0724c310e6d0259da319ce67d5a.jpg) +Figure 13: planning module + +# E EXPERIMENT DETAILS + +# E.1 BENCHMARK ENVIRONMENTS + +We used four benchmark environments in our experiment. For the first three, the workspace dimension is 2d. We generated the maze maps with the recursive backtracker algorithm using the following implementation: https://github.com/lileee/gated-path-planning-networks/blob/master/generate_dataset.py. Examples of the workspace are shown in Figure 15. Three environments differ in the choice of robots: + +• Workspace planning (2d). The robot is abstracted with a point mass moving in the plane. Without higher dimensions, this problem reduces to planning in the workspace. +• Rigid body navigation (3d). A rigid body robot, abstracted as a thin rectangle, is used here. The extra rotation dimension is added to the planning problem. This robot can rotate and move freely without any constraints in the free space. +• 3-link snake (5d). The robot is a 5 DoF snake with two joints. Two more angle dimensions are added to the planning task. To prevent links from folding, we restrict the angles to the range of $[ - \pi / 4 , \pi / 4 ]$ . + +The fourth environment has a 3d workspace. Cuboid obstacles were generated uniformly randomly in space with density $\approx 2 0 \%$ . Example of the workspace is shown in Figure 6 and 16, where the blue cuboids are obstacles. The environment is described below: + +• Spacecraft planning (7d). The robot is a spacecraft with a cuboid body and two 2 DoF arms connecting to two opposite sides of the body. There is a joint in the middle of each arm. The outer arm can rotate around this joint. Each arm can also rotate as a whole around its connection point with the body. All rotation angles are restricted in the range of $[ 0 , \pi / 2 ]$ . The spacecraft itself cannot rotate. + +# E.2 HYPERPARAMETER FOR MSIL + +During self-improving over the first 2000 problems, NEXT updated its parameters and annealed $\epsilon$ once for every 200 problems. The value of the annealing $\epsilon$ was set as the following: + +$$ +\epsilon = \left\{ \begin{array} { l l } { 1 , } & { \mathrm { i f ~ } i < 1 0 0 0 , } \\ { 0 . 5 - 0 . 1 \cdot \lfloor ( i - 1 0 0 0 ) / 2 0 0 \rfloor , } & { \mathrm { i f ~ } 1 0 0 0 \leqslant i < 2 0 0 0 , } \\ { 0 . 1 , } & { \mathrm { o t h e r w i s e , } } \end{array} \right. +$$ + +with $i$ denoting the problem number. + +# E.3 BASELINE: THE IMPROVED GPPN + +The original GPPN is not directly applicable to our experiments. Inspired by Tamar et al. (2016), we add a fully-connected MLP to its final layers, so that the improved architecture can be applied to high-dimensional continuous domain. As shown in Figure 14, the GPPN first processes the discretized + +![](images/1197e1a0f579ba28f10b818c4b2e442327f2cb38b96543bbc6dbcfb1e5ec6325.jpg) +Figure 14: Improved GPPN architecture + +workspace locations. Its output and the full robot configurations are processed together by the MLP, which then produces the current value and action estimates. The improved GPPN is trained using supervisions from the near-optimal paths produced by RRT\*. + +# F EXPERIMENT RESULTS + +![](images/45835db4516f023ab21e7a5a2bb8b4907c34df7064909a1410c42cc4f9f92003.jpg) +Figure 15: The solution path produced by NEXT in a workspace planning task (2d), rigid body navigation task (3d), 3-link snake task (5d) from left to right. The orange dot and the brown dot are starting and goal locations, respectively. + +![](images/28558e709732ca4be6503115df6e181bb722fcd64cd5ffe0ac9a8fec10b06a19.jpg) +Figure 16: The solution paths produced by NEXT in spacecraft planning task (7d). Spacecraft has a yellow body and two 2 DoF (red) arms. Blue cuboids are obstacles. + +# F.2 DETAILS OF QUANTITATIVE EVALUATION + +More detailed results are shown in Table 1, 2, 3, including learning-based and non-learning-based ones, on the last 1000 problems in each experiment. We normalized the number of collision checks and the cost of paths based on the solution of RRT∗. The success rate result is not normalized. The best planners in each experiment are in bold. NEXT-KS and NEXT-GP outperform the current state-of-the-art planning algorithm with large margins. + +Table 1: Success rate results. The higher the better. NEXT-KS performs the best. + +
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.9880.9810.7180.6320.7350.7100.1850.5350.720
3d0.9430.8410.6890.5540.4900.5140.1210.1140.498
5d0.8830.7680.6330.5150.4550.4970.0300.4760.444
7d0.9310.9060.6340.3690.3610.8140.2880.2720.370
+ +Table 2: Average number of collision checks results. The lower the better. The score is normalized based on the solution of $\mathbf { R R T ^ { * } }$ . NEXT-KS performs the best in 3 benchmarks. + +
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1770.2432.3423.4841.0005.9459.2471.9831.011
3d0.6941.3342.2143.1251.0007.9247.2922.1620.988
5d0.8881.5201.8002.7061.0007.4835.7581.1880.997
7d0.6530.5021.8771.3131.0004.6833.8561.5910.987
+ +Table 3: Average cost of paths. The lower the better. The score is normalized based on the solution of $\mathrm { R R T ^ { * } }$ . The NEXT-KS achieves the best solutions. + +
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1720.1931.0491.3331.0001.1402.8111.6491.050
3d0.1160.3150.6120.8751.0000.9551.7201.7340.984
5d0.2150.4260.6730.8901.0000.9231.7800.9611.020
7d0.1080.1470.5730.9871.0000.2911.1141.1390.986
+ +We demonstrated the performance improvement curves for 2d workspace planning, 3d rigid body navigation in Figure 17. As we can see, similar to the performances on 5d 3-link snake planning task in Figure 7, in these tasks, the NEXT-KS and NEXT-GP improve the performances along with more and more experiences collected, justified the self-improvement ability by learning ${ \tilde { V } } ^ { * }$ and $\tilde { \pi } ^ { * }$ . + +# F.3 SEARCH TREES COMPARISON + +We illustrate the search trees generated by $\mathrm { R R T ^ { * } }$ and the proposed NEXT algorithms with 500 samples in Figure 18, Figure 19, Figure 20 and Figure 21 on several 2d, 3d, 5d and 7d planning tasks, respectively. To help readers better understand how the trees were expanded, we actually visualize the RRT\* search trees without edge rewiring, which is equivalent to the RRT search trees, however the vertex set is the same. Comparing to the search trees generated by $\mathbf { R R T ^ { * } }$ side by side, we can clearly see the advantages and the efficiency of the proposed NEXT algorithms. In all the tasks, even in 2d workspace planning task, the $\mathrm { R R T ^ { * } }$ indeed randomly searches without realizing the goals, and thus cannot complete the missions, while the NEXT algorithms search towards the goals with the guidance from ${ \tilde { V } } ^ { * }$ and $\tilde { \pi } ^ { * }$ , therefore, successfully provides high-quality solutions. + +![](images/4c7753cd5f57fdbee146ea4f7042a6906667f2b11da11f1627ca15e30997b485.jpg) +Figure 17: The first and second rows display the improvement curves of our algorithms on all 3000 problems of the 2d workspace planning and 3d rigid body navigation problems. We compare our algorithms with $\mathrm { R R T ^ { * } }$ . Three columns correspond to the success rate, the average collision checks, and the average cost of the solution paths for each algorithm. + +![](images/e12728f87cbc3a880f7afe9bb7ba0f37c6e5a599ae9915a27fe8f8e4dd07bd7e.jpg) + +(a) $\mathrm { R R T ^ { * } }$ (w/o rewiring) (b) NEXT-KS search tree (c) NEXT-GP search tree (d) learned ${ \tilde { V } } ^ { * }$ and $\tilde { \pi } ^ { * }$ + +Figure 18: Column (a) to (c) are the search trees produced by the $\mathrm { R R T ^ { * } }$ , NEXT-KS, and NEXT-GP on the same workspace planning task (2d). The learned $\tilde { V } ^ { \ast }$ and $\tilde { \pi } ^ { * }$ from NEXT-KS are plotted in column (d). In the figures, obstacles are colored in deep blue, the starting and goal locations are denoted by orange and brown dots, respectively. In column (a) to (c), samples are represented with hollow yellow circles, and edges are colored in green. In column (d), the level of redness denotes the value of the cost-to-go estimate $\tilde { V } ^ { \ast }$ , and the cyan arrows point from a given state $s$ to the center of the proposal distribution $\tilde { \pi } ^ { * } \bigl ( s ^ { \prime } | s , U \bigr )$ . We set the maximum number of samples to be 500. + +![](images/e0735fc244ea4c94bcec57f2daf0ecaa89ac9d792df1ae99e175f050da010796.jpg) +Figure 19: Each column corresponds to one example from the rigid body navigation problem (3d). The top and the bottom rows are the search trees produced by the RRT\* and NEXT-KS, respectively. In the figures, obstacles are colored in deep blue, and the rigid bodies are represented with matchsticks. The samples, starting states, and goal states are denoted by yellow, orange, and brown matchsticks, respectively. Edges are colored in green. We set the maximum number of samples to be 500. + +![](images/cfa7ea4cfcbf4aafa0403606ed187484f2a7a371df073fcdcf9d9b4c8e59fe6c.jpg) +Figure 20: Each column corresponds to one example from the 3-link snake problem (5d). The top and the bottom rows are the search trees produced by the RRT\* and NEXT-KS, respectively. In the figures, obstacles are colored in deep blue, and the rigid bodies are represented with matchsticks. The samples, starting states, and goal states are denoted by yellow, orange, and brown matchsticks, respectively. Edges are colored in green. We set the maximum number of samples to be 500. + +![](images/acacf1ccb84450bee6725175dc3f6794eeb8b2a6eca9fd60b1ef87035d99e36d.jpg) +Figure 21: Each column corresponds to one example from the spacecraft planning problem (7d). The top and the bottom rows are the search trees produced by the $\mathrm { R R T ^ { * } }$ and NEXT-KS, respectively. In the figures, obstacles are colored in blue, and each spacecraft has a yellow body and two 2 DoF red arms. We set the maximum number of samples to be 500. + +# F.4 CASE STUDY DETAILS + +We conduct a real-world case study on controlling robot arms to move objects on a shelf. This is a representative of common scenarios in practice where the robot needs to plan its motion in real-time to reach the inside of some narrow space. For this case study, we focus more on the practical aspect to evaluate how much can we improve on the wall-clock time by learning from similar planning problems. + +# F.4.1 TASK DESCRIPTION + +We generate planning problems randomly to form the training set and test set. In each planning task, there is a shelf of multiple levels, with each level horizontally divided into multiple bins. The heights of levels and widths of bins are randomly drawn from some fixed distribution. Samples of shelves are shown in Figure 22. Both the start and goal configurations are randomly sampled from a distribution within the reachable region of the robot arm. The planning environment is created with the OpenRave simulator (Diankov, 2010). + +The task is to find a path for the 7 DoF robot arm to move from a location in one bin to another, i.e., the end effectors of the start and goal configurations are in different bins, as illustrated in Figure 23. In this case, the base of the robot is fixed and we are planning the movement of arm. We generated 3000 problems for training and 1000 problems for testing. + +# F.4.2 BASELINES AND TRAINING + +For traditional planners, we include $\mathrm { C } + +$ OMPL ( ¸Sucan et al., 2012) implementation of $\mathrm { B I T ^ { * } }$ (Gammell et al., 2015) and $\mathrm { R R T ^ { * } }$ (Karaman & Frazzoli, 2011) as baselines. The hyperparameters of RRT\* and $\mathbf { B I T ^ { * } }$ are specially tuned for this experiment. We also compare with learning-based planners CVAE-plan (Ichter et al., 2018) and Reinforce-plan (Zhang et al., 2018). The supervisions for CVAE-plan are produced by the well-tuned $\mathrm { B I T ^ { * } }$ on the training set. To train NEXT, we consider the $\mathrm { B I T ^ { * } }$ instead of RRT\* as the imperfect expert in training problems. + +# F.4.3 RESULTS + +We evaluate the algorithms on the separated testing problems, and record the success rate using 10 different time limits. The success rate and average path quality are plot in Figure 8 and recorded in Table 4 and Table 5. The solution paths found by NEXT and $\mathrm { B I T ^ { * } }$ are illustrated in Figure 23. In terms of both success rate and solution path quality, NEXT dominates all the planners under all time limits. + +![](images/1a68418f9fc5b8db4c0728ef4e4b915a1ac7b12ca2d6815c4396b38f554580ed.jpg) +Figure 22: Examples of different shelves sampled from the distribution. + +![](images/4918f228ef3bb16fc560cb6a309ed28b7f1abfcfd2648d63098ec7552a0228e2.jpg) +Figure 23: First row: robot arm solution trajectories produced by NEXT(-KS) in four planning problems; Second row: $\mathrm { B I T ^ { * } }$ solutions on the same planning problems. NEXT only takes 5 seconds to complete each problem while $\mathrm { B I T ^ { * } }$ needs 250 seconds to find a solution for the hardest problems (last two columns). + +Table 4: Success rates of different planners under varying time limits, the higher the better. + +
5s10s15s20s25s30s35s40s45s50s
NEXT0.5790.6570.7030.7090.7450.7430.7460.7520.7720.763
CVAE0.3540.4370.4820.5090.5070.5390.5530.5510.5790.580
Reinforce0.1600.1700.2000.1500.1800.1900.1750.1800.2250.175
BIT*0.2260.2880.3200.3650.3640.4290.4250.4220.4430.475
RRT*0.1350.1480.1440.1360.1470.1590.1520.1580.1570.165
+ +Table 5: Average path costs of different planners under varying time limits, the lower the better. + +
5s10s15s20s25s30s35s40s45s50s
NEXT25.14321.87320.05219.77318.31818.38118.27418.04717.22417.587
CVAE34.41430.81528.90927.71727.78426.40825.77625.88324.66524.659
Reinforce42.78142.47141.14243.29542.12941.65742.23542.09340.26942.282
BIT*39.54337.08635.74434.00234.00731.41431.58731.65130.84129.601
RRT*43.36842.85542.98143.35542.91942.41042.68042.44142.50442.165
\ No newline at end of file diff --git a/parse/train/rJgJDAVKvB/rJgJDAVKvB_content_list.json b/parse/train/rJgJDAVKvB/rJgJDAVKvB_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9a35ec74011afb1152cb25bde16cc585fdc144c0 --- /dev/null +++ b/parse/train/rJgJDAVKvB/rJgJDAVKvB_content_list.json @@ -0,0 +1,2921 @@ +[ + { + "type": "text", + "text": "LEARNING TO PLAN IN HIGH DIMENSIONS VIA NEURAL EXPLORATION-EXPLOITATION TREES ", + "text_level": 1, + "bbox": [ + 176, + 98, + 738, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "∗Binghong Chen1, $\\mathbf { \\mathbf { \\mathbf { \\mathbf { \\mathbf { \\mathbf { B } } } } } } \\mathbf { \\mathbf { 0 } } \\mathbf { \\mathbf { \\mathbf { \\mathbf { D } } } } \\mathbf { \\mathbf { a } } \\mathbf { i } ^ { 2 }$ , Qinjie $\\mathbf { L i n ^ { 3 } }$ , Guo $\\mathbf { Y e ^ { 3 } }$ , Han Liu3, Le Song1,4 1Georgia Institute of Technology 2Google Research, Brain Team 3Northwestern University 4Ant Financial ", + "bbox": [ + 184, + 167, + 692, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 239, + 544, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a meta path planning algorithm named Neural Exploration-Exploitation Trees (NEXT) for learning from prior experience for solving new path planning problems in high dimensional continuous state and action spaces. Compared to more classical sampling-based methods like RRT, our approach achieves much better sample efficiency in high-dimensions and can benefit from prior experience of planning in similar environments. More specifically, NEXT exploits a novel neural architecture which can learn promising search directions from problem structures. The learned prior is then integrated into a UCB-type algorithm to achieve an online balance between exploration and exploitation when solving a new problem. We conduct thorough experiments to show that NEXT accomplishes new planning problems with more compact search trees and significantly outperforms state-of-the-art methods on several benchmarks. ", + "bbox": [ + 212, + 266, + 787, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 444, + 336, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Path planning is a fundamental problem with many real-world applications, such as robot manipulation and autonomous driving. A simple planning problem within low-dimensional state space can be solved by first discretizing the continuous state space into a grid, and then searching for a path on top of it using graph search algorithms such as $A ^ { * }$ (Hart et al., 1968). However, due to the curse of dimensionality, these approaches do not scale well with the number of dimensions of the state space. For high-dimensional planning problems, people often resort to sampling-based approaches to avoid explicit discretization. Sampling-based planning algorithms, such as probabilistic roadmaps (PRM) (Kavraki et al., 1996), rapidly-exploring random trees (RRT) (LaValle, 1998), and their variants (Karaman & Frazzoli, 2011) incrementally build an implicit representation of the state space using probing samples. These generic algorithms typically employ a uniform sampler which does not make use of the structures of the problem. Therefore they may require lots of samples to obtain a feasible solution for complicated problems. To improve the sample efficiency, heuristic biased samplers, such as Gaussian sampler (Boor et al., 1999), bridge test (Hsu et al., 2003) and reachability-guided sampler (Shkolnik et al., 2009) have been proposed. All these sampling heuristics are designed manually to address specific structural properties, which may or may not be valid for a new problem, and may lead to even worse performance compared to the uniform proposal. ", + "bbox": [ + 174, + 470, + 825, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Online adaptation in path planning has also been investigated for improving sample efficiency in current planning problem. Specifically, Hsu et al. (2005) exploits online algorithms to dynamically adapts the mixture weights of several manually designed biased samplers. Burns & Brock (2005a;b) fit a model for the planning environment incrementally and use the model for planning. Yee et al. (2016) mimics the Monte-Carlos tree search (MCTS) for problems with continuous state and action spaces. These algorithms treat each planning problem independently, and the collected data from previous experiences and built model will be simply discarded when solving a new problem. However, in practice, similar planning problems may be solved again and again, where the problems are different but sharing common structures. For instance, grabbing a coffee cup on a table at different time are different problems, since the layout of paper and pens, the position and orientation of coffee cups may be different every time; however, all these problems show common structures of handling similar objects which are placed in similar fashions. Intuitively, if the common characteristics across problems can be learned via some shared latent representation, a planner based on such representation can then be transferred to new problems with improved sample efficiency. ", + "bbox": [ + 174, + 699, + 825, + 893 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several methods have been proposed recently to learn from past planning experiences to conduct more efficient and generalizable planning for future problems. These works are limited in one way or the other. Zucker et al. (2008); Zhang et al. (2018); Huh & Lee (2018) treat the sampler in the sampling-based planner as a stochastic policy to be learned and apply policy gradient or TD-algorithm to improve the policy. Finney et al. (2007); Bowen & Alterovitz (2014); Ye & Alterovitz (2017); Ichter et al. (2018); Kim et al. (2018); Kuo et al. (2018) apply imitation learning based on the collected demonstrations to bias for better sampler via variants of probabilistic models, e.g., (mixture of) Gaussians, conditional VAE, GAN, HMM and RNN. However, many of these approaches either rely on specially designed local features or assume the problems are indexed by special parameters, which limits the generalization ability. Deep representation learning provides a promising direction to extract the common structure among the planning problems, and thus mitigate such limitation on handdesigned features. However, existing work, e.g., motion planning networks (Qureshi et al., 2019), value iteration networks (VIN) (Tamar et al., 2016), and gated path planning networks (GPPN) (Lee et al., 2018), either apply off-the-shelf MLP architecture ignoring special structures in planning problems or can only deal with discrete state and action spaces in low-dimensional settings. ", + "bbox": [ + 173, + 104, + 825, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present Neural EXploration-EXploitation Tree (NEXT), a meta neural path planning algorithm for high-dimensional continuous state space problems. The core contribution is a novel attention-based neural architecture that is capable of learning generalizable problem structures from previous experiences and produce promising search directions with automatic online exploration-exploitation balance adaption. Compared to existing learning-based planners, ", + "bbox": [ + 173, + 319, + 825, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• NEXT is more generic. We propose an architecture that can embed high dimensional continuous state spaces into low dimensional discrete spaces, on which a neural planning module is used to extract planning representation. These module will be learned end-to-end. • NEXT balances exploration-exploitation trade-off. We integrate the learned neural prior into an upper confidence bound (UCB) style algorithm to achieve an online balance between exploration and exploitation when solving a new problem. ", + "bbox": [ + 173, + 392, + 825, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirically, we show that NEXT can exploit past experiences to reduce the number of required samples drastically for solving new planning problems, and significantly outperforms previous state-of-the-arts on several benchmark tasks. ", + "bbox": [ + 176, + 479, + 821, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1 RELATED WORKS ", + "text_level": 1, + "bbox": [ + 176, + 530, + 339, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Designing non-uniform sampling strategies for random search to improve the planning efficiency has been considered as we discussed above. Besides the mentioned algorithms, there are other works along this line, including informed RRT∗ (Gammell et al., 2014) and batch informed Trees $\\mathrm { ( B I T ^ { * } ) }$ (Gammell et al., 2015) as the representative work. Randomized $A ^ { * }$ (Diankov & Kuffner, 2007) and sampling-based $A ^ { * }$ (Persson & Sharf, 2014) expand the search tree with hand-designed heuristics. These methods incorporate the human prior knowledge via hard-coded rules, which is fixed and unable to adapt to problems, and thus, may not universally applicable. Choudhury et al. (2018); Song et al. (2018) attempt to learn search heuristics. However, both methods are restricted to planning on discrete domains. Meanwhile, the latter one always employs an unnecessary hierarchical structure for path planning, which leads to inferior sample efficiency and extra computation. ", + "bbox": [ + 174, + 546, + 825, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The online exploration-exploitation trade-off is also an important issue in planning. For instance, Rickert et al. (2009) constructs a potential field sampler and tuned the sampler variance based on collision rate for the trade-off heuristically. Paxton et al. (2017) separates the action space into high-level discrete options and low-level continuous actions, and only considered the trade-off at the discrete option level, ignoring the exploration-exploitation in the fine action space. These existing works address the trade-off in an ad-hoc way, which may be inferior for the balance. ", + "bbox": [ + 174, + 691, + 825, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There have been sevearl non-learning-based planning methods that can also leverage experiences (Kavraki et al., 1996; Phillips et al., 2012) by utilizing search graphs created in previous problems. However, they are designed for largely fixed obstacles and cannot be generalized to unseen tasks from the same planning problems distribution. ", + "bbox": [ + 174, + 781, + 825, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 SETTINGS FOR LEARNING TO PLAN ", + "text_level": 1, + "bbox": [ + 174, + 845, + 501, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $S \\subseteq \\mathbb { R } ^ { q }$ be the state space of the problem, e.g., all the configurations of a robot and its base location in the workspace, $S _ { o b s } \\subsetneq S$ be the obstacles set, $S _ { f r e e } : = \\mathcal { S } \\setminus \\mathcal { S } _ { o b s }$ be the free space, $s _ { i n i t } \\in S _ { f r e e }$ be the initial state and $\\mathcal { S } _ { g o a l } \\subsetneq S _ { f r e e }$ be the goal region. Then the space of all collisionfree paths can be defined as a continuous function $\\Xi : = \\{ \\bar { \\xi } ( \\cdot ) : \\bar { [ 0 , 1 ] } S _ { f r e e } \\}$ . Let $c \\left( \\cdot \\right) : \\Xi \\mapsto \\mathbb { R }$ be the cost functional over a path. The optimal path planning problem is to find the optimal path in terms of cost $c ( \\cdot )$ from start $s _ { i n i t }$ to goal $\\mathcal { S } _ { g o a l }$ in free space $\\boldsymbol { S } _ { f r e e }$ , i.e., ", + "bbox": [ + 174, + 867, + 825, + 925 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/6219e9be4af3c9da34e4d21a4c18ba35346a4c909f89f4775907c40c6f1e2b0f.jpg", + "text": "$$\n\\begin{array} { r } { \\xi ^ { * } = \\mathrm { a r g m i n } _ { \\xi \\in \\Xi } \\ c \\left( \\xi \\right) , \\quad \\mathrm { s . t . } \\xi ( 0 ) = s _ { i n i t } , \\xi ( 1 ) \\in S _ { g o a l } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 308, + 133, + 687, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditionally (Karaman & Frazzoli, 2011), the planner has direct access to $( s _ { i n i t } , S _ { g o a l } , c \\left( \\cdot \\right) )$ and the workspace map (Ichter et al., 2018; Tamar et al., 2016; Lee et al., 2018), $\\mathtt { m a p } ( \\cdot ) : \\mathbb { R } ^ { 2 }$ or $\\mathbb { R } ^ { 3 } \\mapsto \\{ 0 , 1 \\}$ , (0: free spaces and 1: obstacles). Since $\\boldsymbol { S } _ { f r e e }$ often has a very irregular geometry (illustrated in Figure 10 in Appendix A), it is usually represented via a collision detection module which is able to detect the obstacles in a path segment. For the same reason, the feasible paths in $\\Xi$ are hard to be described in parametric forms, and thus, the nonparametric $\\xi$ , such as a sequence of interconnected path segments $\\left[ s _ { 0 } , s _ { 1 } \\right] , \\left[ s _ { 1 } , s _ { 2 } \\right] , \\ldots , \\left[ s _ { T - 1 } , s _ { T } \\right] \\subset \\mathcal { S }$ with $\\xi ( 0 ) = s _ { 0 } = s _ { i n i t }$ and $\\xi ( 1 ) = s _ { T }$ , is used with an additive cost $\\textstyle \\sum _ { i = 1 } ^ { T } c \\left( \\left[ s _ { i - 1 } , s _ { i } \\right] \\right)$ −. ", + "bbox": [ + 173, + 151, + 825, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming given th planning problems $\\{ U _ { i } : = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c \\left( \\cdot \\right) ) \\} _ { i = 1 } ^ { N }$ sampled from some distribution $\\mathcal { U }$ , we are interested in learning an algorithm $\\mathsf { a l g } \\left( \\cdot \\right)$ , which can produce the (nearly)-optimal path efficiently from the observed planning problems. Formally, the learning to plan is defined as ", + "bbox": [ + 173, + 276, + 825, + 330 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/8d470ec5867b80166a3f19c2e34b8a822874d68e6287bf6444650bf2bae8b2b9.jpg", + "text": "$$\n\\begin{array} { r } { \\mathbf { a } \\mathbf { 1 g } ^ { * } \\left( \\cdot \\right) = \\operatorname * { a r g m i n } _ { \\mathbf { a } \\mathbf { 1 g } \\in \\mathcal { A } } \\mathbb { E } _ { U \\in \\mathcal { U } } \\left[ \\ell \\left( \\mathbf { a } \\mathbf { 1 g } ( U ) \\right) \\right] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 348, + 327, + 648, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal { A }$ denotes the planning algorithm family, and $\\ell \\left( \\cdot \\right)$ denotes some loss function which evaluates the quality of the generated path and the efficiency of the $\\mathsf { a l g } \\left( \\cdot \\right)$ , e.g., size of the search tree. We elaborate each component in Eq (2) in the following sections. We first introduce the tree-based sampling algorithm template in Section 3, upon which we instantiate the $\\mathsf { a l g } \\left( \\cdot \\right)$ via a novel attention-based neural parametrization in Section 4.2 with exploration-exploitation balance mechanism in Section 4.1. We design the $\\ell$ -loss function and the meta learning algorithm in Section 4.3. Composing every component together, we obtain the neural exploration-exploitation trees (NEXT) which achieves outstanding performances in Section 5. Repeat ", + "bbox": [ + 173, + 347, + 826, + 459 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c81f5245215344ac0251524c1e65b7b0d359bb22ec9a016d0918f1939b8a9b94.jpg", + "image_caption": [ + "Figure 1: Illustration of NEXT. In each epoch, NEXT is executed on a randomly generated planning problem. The search tree grows with $\\tilde { V } ^ { * }$ and $\\tilde { \\pi } ^ { * }$ guidance. $\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}$ will be updated according to the successful path. Such planning and learning iteration is continued interactively. " + ], + "image_footnote": [], + "bbox": [ + 209, + 459, + 784, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 176, + 717, + 339, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1 Problem: ", + "text_level": 1, + "bbox": [ + 488, + 731, + 563, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1: Tree-based Sampling Algorithm ", + "text_level": 1, + "bbox": [ + 498, + 712, + 799, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sampling-based planners are more practical and become dominant for high-dimensional path planning problems (Elbanhawi & Simic, 3 2014). We describe a unifying view for many 4 existing tree-based sampling algorithms (TSA), 5 which we will also base our algorithm upon. More specifically, this family of algorithms 6 maintain a search tree $\\tau$ rooted at the initial point $s _ { i n i t }$ and connecting all sampled points $\\nu$ 7 in the configuration space with edge set $\\mathcal { E }$ . The 8 tree will be expanded by incorporating more sampled states until some leaf reaches $\\mathcal { S } _ { g o a l }$ 9. Then, a feasible solution for the path planning ", + "bbox": [ + 174, + 744, + 490, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$U = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathrm { { m a p } , \\it { c ( \\cdot ) } ) }$ ; \n2 Initialize $\\boldsymbol { \\mathcal { T } } = ( \\boldsymbol { \\mathcal { V } } , \\boldsymbol { \\mathcal { E } } )$ with $\\mathcal { V } \\{ s _ { i n i t } \\}$ , $\\mathcal { E } \\emptyset$ ; \nfor $t \\gets 0$ to $T$ do sparent, $s _ { n e w } \\mathrm { E }$ xpand $( \\mathcal { T } , U )$ ; if ObstacleFree( $\\prime s _ { p a r e n t } , s _ { n e w } )$ ) then $\\mathcal { V } \\mathcal { V } \\cup \\{ s _ { n e w } \\}$ and $\\mathcal { E } \\mathcal { E } \\cup \\{ [ s _ { p a r e n t } , s _ { n e w } ] \\}$ ; T ← Postprocess(T , U) ; $\\triangleright$ Optional if $s _ { n e w } \\in S _ { g o a l }$ then return T ; ", + "bbox": [ + 485, + 742, + 799, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "problem will be extracted based on the tree $\\tau$ . The template of tree-based sampling algorithms is summarized in Algorithm 1 and illustrated in Fig. 1(c). A key component of the algorithm is the Expand operator, which generates the next exploration point $s _ { n e w }$ and its parent $s _ { p a r e n t } \\in \\mathcal { V }$ . To ensure the feasibility of the solution, the $s _ { n e w }$ must be reachable from $\\tau$ , i.e., $[ \\bar { s _ { p a r e n t } } , s _ { n e w } ]$ is collision-free, which is checked by a collision detection function. As we will discuss in Appendix B, by instantiating different Expand operators, we will arrive at many existing algorithms, such as RRT (LaValle, 1998) and EST (Hsu et al., 1997; Phillips et al., 2004). ", + "bbox": [ + 174, + 103, + 825, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One major limitation of existing TSAs is that they solve each problem independently from scratch and ignore past planning experiences in similar environments. We introduce the neural components into TSA template to form the learnable planning algorithm family $\\mathcal { A }$ , which can explicitly take advantages of the past successful experiences to bias the Expand towards more promising regions. ", + "bbox": [ + 174, + 208, + 825, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 NEURAL EXPLORATION-EXPLOITATION TREES", + "text_level": 1, + "bbox": [ + 174, + 276, + 596, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the TSA framework, we introduce a learnable neural based Expand operator, which can balance between exploration and exploitation, to instantiate $\\mathcal { A }$ in Eq (2). With the self-improving training, we obtain the meta NEXT algorithm illustrated in Figure 1. ", + "bbox": [ + 174, + 301, + 825, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 GUIDED PROGRESSIVE EXPANSION ", + "text_level": 1, + "bbox": [ + 174, + 352, + 460, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start with our design of the Expand. We assume having an estimated value function ${ \\tilde { V } } ^ { * } ( s | U )$ , which stands for the optimal cost from $s$ to target in planning problem $U$ , and a policy $\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)$ , which generates the promising action $s ^ { \\prime }$ from state $s$ . The concrete parametriza", + "bbox": [ + 174, + 371, + 446, + 470 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a8c886f8d6118bbe9fbb4275382954ce28939501d27215b83992823a8b9fe93d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Algorithm 2: NExT :: Expand(T = (V,ε),U)
Sparent ← argmaxs∈vΦ(s) ; Selection
{s1,*..,Sk} id π*(s'sparent,U)； >Candidates
Snew ← argmaxs'∈{s1,.,sk}(s')； DExpansion
4 return Sparent, Snew;
", + "bbox": [ + 455, + 375, + 823, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion of ${ \\tilde { V } } ^ { * }$ and $\\tilde { \\pi } ^ { * }$ will be explained in Section 4.2 and learned in Section 4.3. We will use these functions to construct the learnable Expand with explicit exploration-exploitation balancing. ", + "bbox": [ + 174, + 472, + 823, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Expand operator will expand the current search tree $\\tau$ by a new neighboring state $s _ { n e w }$ around $\\tau$ . We design the expansion as a two-step procedure: (i) select a state $s _ { p a r e n t }$ from existing tree $\\tau$ ; (ii) expand a state $s _ { n e w }$ in the neighborhood of $s _ { p a r e n t }$ . More specifically, ", + "bbox": [ + 174, + 506, + 825, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Selecting $s _ { p a r e n t }$ from $\\tau$ in step (i). Consider the negative value function $- \\tilde { V } ^ { \\ast } \\left( s | U \\right)$ as the rewards $r \\left( s \\right)$ , step (i) shares some similarity with the multi-armed bandit problem by viewing existing nodes $s \\in \\mathcal V$ as arms. However, the vanilla UCB algorithm is not directly applicable, since the number of states is increasing as the algorithm proceeds and the value of these adjacent states are naturally correlated. We address this challenge by modeling the correlation explicitly as in contextual bandits. Specifically, we parametrize the UCB of the reward function as $\\phi \\left( s \\right)$ , and select a node from $\\tau$ according to $\\phi ( s )$ ", + "bbox": [ + 173, + 558, + 825, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/737553e4156cfdfc16e92c6b0157200463f7e7d79262d5451b425025895670c0.jpg", + "text": "$$\n\\begin{array} { r } { s _ { p a r e n t } = \\mathrm { a r g m a x } _ { s \\in \\mathcal { V } } \\phi ( s ) : = \\bar { r } _ { t } \\left( s \\right) + \\lambda \\sigma _ { t } \\left( s \\right) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 333, + 656, + 661, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "t t the sequence of where and denote the average reward and variance estimator after $t$ selected tree nodes so far as $S _ { t } = \\{ s _ { p a r e n t } ^ { 1 } , . . . , s _ { p a r e n t } ^ { t } \\}$ $t$ -calls to Expand. Denote , then we can use kernel smoothing estimator for $\\begin{array} { r } { \\bar { r } _ { t } ( s ) = \\frac { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } k \\left( s ^ { \\prime } , s \\right) r \\left( s ^ { \\prime } \\right) } { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } k \\left( s ^ { \\prime } , s \\right) } } \\end{array}$ and $\\begin{array} { r } { \\sigma _ { t } ( s ) = \\sqrt { \\frac { \\log { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } w ( s ^ { \\prime } ) } } { w ( s ) } } } \\end{array}$ where $k ( s ^ { \\prime } , s )$ is a kernel function and $\\begin{array} { r } { w ( s ) = \\sum _ { s ^ { \\prime } \\in S _ { t } } k ( s ^ { \\prime } , s ) } \\end{array}$ . Other parametrizations of $\\bar { r } _ { t }$ and $\\sigma _ { t }$ are also possible, such as Gaussian Process parametrization in Appendix C. The average reward exploits more promising states, while the variance promotes exploration towards less frequently visited states; and the exploration versus exploitation is balanced by a tunable weight $\\lambda > 0$ . ", + "bbox": [ + 173, + 674, + 825, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Expanding a reachable $s _ { n e w }$ in step (ii). Given the selected $s _ { p a r e n t }$ , we consider expanding a reachable state in the neighborhood $s _ { p a r e n t }$ as an infinite-armed bandit problem. Although one can first samples $k$ arms uniformly from a neighborhood around $s _ { p a r e n t }$ and runs a UCB algorithm on the randomly generated finite arms (Wang et al., 2009), such uniform sampler ignores problem structures, and will lead to unnecessary samples. Instead we will employ a policy $\\bar { \\tilde { \\pi } } ^ { * } ( s ^ { \\prime } \\vert s , \\bar { U } ) ^ { 1 }$ for guidance when generating the candidates. The final choice for next move will be selected from these candidates with max $\\phi \\left( s \\right)$ defined in (3). As explained in more details in Section 4.2, $\\tilde { \\pi } ^ { * }$ will be trained to mimic previous successful planning experiences across different problems, that is, biasing the sampling towards the states with higher successful probability. ", + "bbox": [ + 173, + 797, + 825, + 896 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/659fdd9f888e78d83e6532bb1c8c3abf8ec8cc26808c966e6aef6311a6ac402b.jpg", + "image_caption": [ + "Figure 2: Our neural network model maps a $N$ -link robot from the original planning space (a $( N + 2 )$ -d configuration space) to a 3d discrete latent planning space in which we plan a path using value iteration. The result of value iteration is then used as features for defining $\\tilde { V } ^ { * } \\left( s | U \\right)$ and $\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)$ . " + ], + "image_footnote": [], + "bbox": [ + 181, + 90, + 537, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 238, + 823, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With these detailed step (i) and (ii), we obtain NEXT :: Expand in Algorithm 2 (illustrated in Figure 1(b) and (c)). Plugging it into the TSA in 1, we construct $\\tt a l g \\left( \\cdot \\right) \\in \\mathcal { A }$ which will be learned. ", + "bbox": [ + 174, + 273, + 823, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The guided progressive expansion bears similarity to MCTS but deals with high dimensional continuous spaces. Moreover, the essential difference lies in the way to select state in $\\tau$ for expansion: the MCTS only expands the leaf states in $\\tau$ due to the hierarchical assumption, limiting the exploration ability and incurring extra unnecessary UCB sampling for internal traversal in $\\tau$ ; while the proposed operation enables expansion from each visited state, particularly suitable for path planning problems. ", + "bbox": [ + 173, + 308, + 826, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 NEURAL ARCHITECTURE FOR VALUE FUNCTION AND EXPANSION POLICY ", + "text_level": 1, + "bbox": [ + 176, + 385, + 730, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we will introduce our neural architectures for $\\tilde { V } ^ { * } \\left( s | U \\right)$ and $\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)$ used in NEXT :: Expand. The proposed neural architectures can be understood as first embedding the state and problem into a discrete latent representation via an attention-based module in Section 4.2.1, upon which the neuralized value iteration, introduced in Section 4.2.2, is performed to extract features for defining $\\tilde { V } ^ { * } \\left( s | U \\right)$ and $\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)$ , as illustrated in Figure 2. ", + "bbox": [ + 173, + 406, + 825, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 CONFIGURATION SPACE EMBEDDING ", + "text_level": 1, + "bbox": [ + 176, + 487, + 486, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our network for embedding high-dimension configuration space into a latent representation is designed based on an attention mechanism. More specifically, let $s ^ { w }$ denote the workspace in state and ${ \\bf \\bar { \\Phi } } _ { s } h$ denote the remaining dimensions of the state, i.e. $s = \\mathbf { \\bar { \\rho } } ( s ^ { w } , s ^ { h } )$ . $s ^ { w }$ and $s ^ { h }$ will be embedded by different sub-neural networks and combined for the final representation, as illustrated in Figure 3. For simplicity of exposition, we will focus on the 2d workspace, i.e., $s ^ { w } \\in \\mathbb { R } ^ { 2 }$ . However, our method applies to 3d workspace as well. ", + "bbox": [ + 174, + 505, + 825, + 588 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1f9670de482ed991ba97956d7ebba8ef1a8b77c6a4180f100db30cdc84a6e86c.jpg", + "image_caption": [ + "Figure 3: Attention-based state embedding module. $\\boldsymbol { s } ^ { w } = ( x , y )$ and $\\mathbf { z } = s ^ { h }$ . The upper part is spatial attention, with the first two channels being $_ x$ and $_ y$ , and the last two channels being constant templates with the row and column coordinates, as shown with a $d$ set to 3. The bottom module learns the representation for $\\mathbf { z }$ . The final embedding is obtained by outer-product of these two attention parts. " + ], + "image_footnote": [], + "bbox": [ + 205, + 599, + 772, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "• Spatial attention. The workspace information $s ^ { w }$ will be embedded as $\\mu ^ { w } ( s ^ { w } ) \\in \\mathbb { R } ^ { d \\times d }$ , $d$ is a hyperparameter related to map (see remark below). The spatial embedding module (upper part in Figure 3) is composed of $k _ { w }$ convolution layers, i.e., ", + "bbox": [ + 174, + 820, + 825, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/317548e75c0ae16c1cddef18d446e697154769250eeb70f2004ecf9a97134885.jpg", + "text": "$$\n\\mu ^ { w } ( s ^ { w } ) = { \\tt s o f t m a x 2 d } ( f _ { k _ { w } } ^ { w } ( s ^ { w } ) ) , \\qquad f _ { i + 1 } ^ { w } ( s ^ { w } ) = { \\tt r e l u } ( \\theta _ { i } ^ { w } \\oplus f _ { i } ^ { w } ( s ^ { w } ) ) ,\n$$", + "text_format": "latex", + "bbox": [ + 264, + 863, + 748, + 881 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta _ { i } ^ { w }$ denotes the convolution kernels, $\\oplus$ denotes the convolution operator and $f _ { i } ^ { w } ( s ^ { w } ) \\in$ $\\mathbb { R } ^ { d \\times d \\times d _ { i } }$ with $d _ { i }$ channels. The first layer $f _ { 0 } ^ { w }$ is designed to represent $s ^ { w }$ into a $d \\times d \\times 4$ tensor as $f _ { 0 } ^ { w } ( s ^ { w } ) _ { i j } = [ s _ { 1 } ^ { w } , s _ { 2 } ^ { w } , i , j ] , i , j = 1 , \\ldots , a$ 0, without any loss of information. ", + "bbox": [ + 187, + 881, + 825, + 925 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e690ebd6ddc08b67ba7f0c4c102d2ba25dbff64c677153964eecb7a40240d491.jpg", + "image_caption": [ + "Figure 4: Overall model architecture. Current and goal states are embedded through attention module. Then the embedding of the goal state is concatenated with the map to produce $\\nu ^ { * ( 0 ) }$ and $\\bar { \\tilde { \\boldsymbol { R } } }$ as the input to the planning module. The output of the planning module is aggregated with the embedding of the current state to produce feature $\\psi ( s )$ for defining $\\tilde { V } ^ { \\ast }$ and $\\tilde { \\pi } ^ { * }$ . " + ], + "image_footnote": [], + "bbox": [ + 207, + 97, + 789, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "• Configuration attention. The remaining configuration state information will be embedded as $\\mu ^ { h } ( s ^ { h } )$ through $k _ { h }$ fully-connected layers (bottom-left part in Figure 3), i.e. ", + "bbox": [ + 174, + 296, + 823, + 324 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/b9f52b29aa72f5ea2e0438622f43cd032a335db9b8db958f90e8a3b48576656e.jpg", + "text": "$$\n\\mu ^ { h } ( s ^ { h } ) = \\mathsf { s o f t m a x } ( f _ { k _ { h } } ^ { h } ( s ^ { h } ) ) , \\qquad f _ { i + 1 } ^ { h } ( s ^ { h } ) = \\mathtt { r e l u } ( \\theta _ { i } ^ { h } f _ { i } ^ { h } ( s ^ { h } ) + b _ { i } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 272, + 325, + 738, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mu ^ { h } ( s ^ { h } ) \\in \\mathbb { R } ^ { d _ { a } }$ and $f _ { 0 } ^ { h } ( s ^ { h } ) = s ^ { h }$ . ", + "bbox": [ + 191, + 347, + 449, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final representation $\\mu _ { \\boldsymbol { \\theta } } ( s )$ will be obtained by multiplying $\\mu ^ { w } ( s ^ { w } )$ with $\\mu ^ { h } ( s ^ { h } )$ element-wisely, $\\mu _ { \\boldsymbol \\theta } ( s ) _ { i j l } = \\bar { \\mu } ^ { w } ( s ^ { w } ) _ { i j } \\cdot \\bar { \\mu } ^ { h } ( \\bar { s } ^ { h } ) _ { l }$ , which is a $d \\times d \\times d _ { a }$ tensor attention map with $\\mu _ { \\theta } ( s ) _ { i j l } \\geqslant$ 0, and $\\begin{array} { r } { \\sum _ { i j l } \\mu _ { \\theta } ( s ) _ { i j l } = 1 } \\end{array}$ (bottom-right part in Figure 3). Intuitively, one can think of $d _ { a }$ as the level of the learned discretization of the configuration space $s ^ { h }$ , and the entries in softly assign the actual state $s$ to these discretized locations. $\\theta : = ( \\{ \\theta _ { i } ^ { w } \\} _ { i = 0 } ^ { k _ { w } - 1 } , \\{ \\theta _ { i } ^ { h } , b _ { i } \\} _ { i = 0 } ^ { k _ { h } - 1 } )$ are the parameters to be ", + "bbox": [ + 173, + 373, + 825, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark (different map size): To process map using convolution neural networks, we resize it to a $d \\times d$ image, with the same size as the spatial attention in Eq (4), where $d$ is a hyperparameter. ", + "bbox": [ + 173, + 470, + 825, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 NEURAL VALUE ITERATION ", + "text_level": 1, + "bbox": [ + 176, + 525, + 421, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then apply neuralized value iteration on top of the configuration space embedding to extract further planning features (Tamar et al., 2016; Lee et al., 2018). Specifically, we first produce the embedding $\\mu _ { \\theta } \\big ( s _ { g o a l } \\big )$ of the center of the goal region $s _ { g o a l }$ . We execute $T$ steps of neuralized Bellman updates (planning module in Figure 4) in the embedding space, ", + "bbox": [ + 173, + 550, + 826, + 606 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/031e99ba81d68aaa517f693a0ebc251fd46e6b8dba4e8c0d754e254cbb1c0902.jpg", + "text": "$$\n\\nu ^ { * ( t ) } = \\operatorname* { m i n } \\left( W _ { 1 } \\oplus \\left[ \\nu ^ { * ( t - 1 ) } , \\tilde { R } \\right] \\right) , \\quad \\mathrm { w i t h } \\quad \\left( \\nu ^ { * ( 0 ) } , \\tilde { R } \\right) = \\sigma \\left( W _ { 0 } \\oplus \\left[ \\mu _ { \\theta } \\big ( s _ { g o a l } \\big ) , \\mathtt { m a p l } \\right) \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 214, + 607, + 782, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and obtain $\\nu ^ { * ( T ) } \\in \\mathbb { R } ^ { d \\times d \\times d _ { a } \\times p }$ . Both $W _ { 0 }$ and $W _ { 1 }$ are 3d convolution kernels, min implements the pooling across channels. Accordingly, $\\nu ^ { * ( T ) }$ now can be understood as a latent representation of the value function $\\tilde { V } ^ { * } ( \\cdot )$ in learned embedding space for the problem $U$ with $s _ { g o a l }$ in map. ", + "bbox": [ + 173, + 635, + 825, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To define the value function for particular state $s$ , i.e., $\\tilde { V } ^ { * } \\left( s | U \\right)$ , from the latent representation $\\nu ^ { * ( T ) }$ , we first construct another attention model between the embedding of state $s$ using $\\mu _ { \\boldsymbol { \\theta } } ( s )$ and $\\nu ^ { * ( T ) }$ , i.e., $\\begin{array} { r } { \\psi \\left( s \\right) _ { k } = \\sum _ { i j l } \\nu _ { i j l k } ^ { * ( T ) } \\cdot \\mu _ { \\theta } \\left( s \\right) _ { i j l } } \\end{array}$ , for $k = 1 , \\hdots , p$ . Finally we define ", + "bbox": [ + 173, + 689, + 826, + 741 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/0b18fc86bde7bbd5b1268f8408e733fa5a87511bda64f1095ef6cc2258a4ab25.jpg", + "text": "$$\n\\tilde { V } ^ { \\ast } \\left( s | U \\right) = h _ { W _ { 2 } } \\left( \\psi ( s ) \\right) , \\quad \\mathrm { a n d } \\quad \\tilde { \\pi } ^ { \\ast } \\left( s ^ { \\prime } | s , U \\right) = \\mathcal { N } \\left( h _ { W _ { 3 } } \\left( \\psi ( s ) \\right) , \\sigma ^ { 2 } \\right)\n$$", + "text_format": "latex", + "bbox": [ + 267, + 743, + 727, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $h _ { W _ { 2 } }$ and $h _ { W _ { 3 } }$ are fully connected dense layers with parameters $W _ { 2 }$ and $W _ { 3 }$ respectively, and $\\mathcal { N } ( h _ { W _ { 3 } } \\left( \\psi ( s ) \\right) , \\sigma ^ { 2 } )$ is a Gaussian distribution with variance $\\sigma ^ { 2 }$ . Note that we also parametrize the policy $\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)$ using the embedding $\\nu ^ { * ( T ) }$ , since the policy is connected to the value function via $\\begin{array} { r } { \\pi ^ { * } \\left( s ^ { \\prime } | s , U \\right) = \\mathrm { a r g m i n } _ { s ^ { \\prime } \\in S } c \\left( \\left[ s , s ^ { \\prime } \\right] \\right) + V ^ { * } \\left( s ^ { \\prime } | U \\right. } \\end{array}$ ). It should also be emphasized that in our parametrization, the calculation of $\\nu ^ { * ( T ) }$ only relies on the $\\mu _ { \\theta } \\left( s _ { g o a l } \\right)$ , which can be reused for evaluating $\\tilde { V } ^ { * } \\left( s | U \\right)$ and $\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)$ over different $s$ , saving computational resources. Using this trick the algorithm runs $1 0 \\times - 1 0 0 \\times$ faster empirically. ", + "bbox": [ + 173, + 761, + 825, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall model architecture in $\\mathsf { a l g } \\left( \\cdot \\right)$ is illustrated in Figure 4. The parameters $W =$ $( W _ { 0 } , W _ { 1 } , W _ { 2 } , W _ { 3 } , \\theta )$ will be learned together by our meta self-improving learning. For the details of the parameterization and the size of convolution kernels in our implementation, please refer to Figure 12 in Appendix D. ", + "bbox": [ + 173, + 871, + 825, + 929 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 META SELF-IMPROVING LEARNING ", + "text_level": 1, + "bbox": [ + 176, + 103, + 457, + 117 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The learning of the planner $\\mathsf { a l g } \\left( \\cdot \\right)$ reduces to learning the parameters in $\\tilde { V } ^ { * } \\left( s | U \\right)$ and $\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)$ and is carried out while planning experiences accumulate. We do not have an explicit training and testing phase separation. Particularly, we use a mixture of RRT :: Expand and NEXT :: Expand with probability $\\epsilon$ and $1 - \\epsilon$ , respectively, inside the TSA framework in Algorithm 1. The RRT∗ postprocessing step is used in the template. The $\\epsilon$ is set to be 1 initially since $\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}$ are not well-trained, and thus, the algorithm behaves like $\\mathbf { R R T ^ { * } }$ . As the training proceeds, we anneal $\\epsilon$ gradually as the sampler becomes more efficient. ", + "bbox": [ + 173, + 122, + 825, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The dataset $\\mathcal { D } _ { n } = \\{ ( \\mathcal { T } _ { j } , U _ { j } ) \\} _ { j = 1 } ^ { n }$ for the $n$ -th training epoch is collected from the previous successful planning experiences across multiple random problems. We fix the size of dataset and update $\\mathcal { D }$ in the same way as experience reply buffer (Lin, 1992; Schaul et al., 2015). For an experience $( { \\mathcal { T } } , U ) \\in { \\mathcal { D } } _ { n }$ , we can reconstruct the successful path $\\{ s ^ { i } \\} _ { i = 1 } ^ { m }$ from the search tree $\\tau$ ( $\\mathrm { ~ m ~ }$ is the region, i.e., number of segments), and the value of each state $\\begin{array} { r } { y ^ { i } : = \\sum _ { l = i } ^ { m - 1 } c ( [ s ^ { l } , s ^ { l + 1 } ] ) } \\end{array}$ . We learn $\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}$ $s ^ { i }$ in the path will be the sum of cost to the goal by optimizing objective ", + "bbox": [ + 174, + 228, + 825, + 316 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/73a67d0d8a8b20fc2ce7c14d50c96eb2453ad6e3db342b8b427574a485703b12.jpg", + "text": "$$\n\\operatorname* { m i n } _ { W } \\sum _ { ( { \\cal T } , { \\cal U } ) \\in { \\mathcal D } _ { n } } \\ell ( \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } ; { \\cal T } , { \\cal U } ) : = - \\sum _ { { \\mathcal D } _ { n } } \\sum _ { i = 1 } ^ { m - 1 } \\log \\tilde { \\pi } ^ { * } \\left( s ^ { i + 1 } | s ^ { i } \\right) + \\sum _ { i = 1 } ^ { m } ( \\tilde { V } ^ { * } \\left( s ^ { i } \\right) - y ^ { i } ) _ { 2 } ^ { 2 } + \\lambda \\left\\| { \\cal W } \\right\\| ^ { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 178, + 318, + 799, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The loss (6) pushes the ${ \\tilde { V } } ^ { * }$ and $\\tilde { \\pi } ^ { * }$ to chase the successful trajectories, providing effective guidance in $\\mathsf { a l g } \\left( \\cdot \\right)$ for searching, and therefore leading to efficient searching procedure with less sample complexity and better solution. On one hand, the value function and policy estimation $\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}$ is improved based upon the successful outcomes from NEXT itself on previous problems. On the other hand, the updated $\\{ \\tilde { V } ^ { * } , \\bar { \\pi } ^ { * } \\}$ will be applied in the next epoch to improve the performance. Therefore, the training is named as Meta Self-Improving Learning (MSIL). Since all the trajectories we collected for learning are feasible, the reachability of the proposed samples is enforced implicitly via imitating these successful paths. ", + "bbox": [ + 174, + 368, + 480, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By putting every components together into the 10 learning to plan framework in Eq (2), the over", + "bbox": [ + 173, + 602, + 491, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Algorithm 3: Meta Self-Improving Learning ", + "text_level": 1, + "bbox": [ + 500, + 368, + 792, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1 Initialize dataset $\\mathcal { D } _ { 0 }$ ; \n$^ 2$ for epoch $n \\gets 1$ to $N$ do \n$^ 3$ Sample a planning problem $U$ ; \n4 $\\mathcal { T } \\mathrm { T S A } ( U )$ with $\\epsilon \\sim \\mathcal { U } n i f [ 0 , 1 ]$ , and $\\epsilon$ ·RRT :: Expand $+ ( 1 - \\epsilon )$ ·NEXT :: Expand; Postprocessing with $\\mathbb { R R T } ^ { * }$ :: Postprocess; \n5 ${ \\mathcal { D } } _ { n } \\gets { \\mathcal { D } } _ { n - 1 } \\bar { \\cup } \\{ ( { \\mathcal { T } } , U ) \\}$ if successful else $\\mathcal { D } _ { n - 1 }$ ; \n6 for $j 0$ to $L$ do \n7 Sample Reconst $( \\tau _ { j } , U _ { j } )$ from -optim $\\mathcal { D } _ { n }$ ;path $\\{ s ^ { i } \\} _ { i = 1 } ^ { m }$ and the cost of paths based on $\\tau _ { j }$ ; \n9 Update parameters $W W - \\eta \\nabla _ { W } \\ell ( \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } ; T _ { j } , U _ { j } ) ;$ Anneal $\\epsilon = \\alpha \\epsilon$ , $\\alpha \\in ( 0 , 1 )$ ; ", + "bbox": [ + 483, + 388, + 805, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "all procedure is summarized in Algorithm 3 and illustrated in Figure 1. ", + "bbox": [ + 176, + 630, + 640, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 659, + 326, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed NEXT empirically on different planning tasks in a variety of environments. Comparing to the existing planning algorithms, NEXT achieves the state-of-the-art performances, in terms of both success rate and the quality of the found solutions. We further demonstrate the power of the proposed two components by the corresponding ablation study. We also include a case study on a real-world robot arm control problem at the end of the section. ", + "bbox": [ + 174, + 683, + 825, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 EXPERIMENT SETUP ", + "text_level": 1, + "bbox": [ + 174, + 768, + 356, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmark environments. We designed four benchmark tasks to demonstrate the effectiveness of our algorithm for high-dimensional planning. The first three involve planning in a 2d workspace with a 2 DoF (degrees of freedom) point robot, a 3 DoF stick robot and a 5 DoF snake robot, respectively. The last one involves planning a 7 DoF spacecraft in a 3d workspace. For all problems in each benchmark task, the workspace maps were randomly generated from a fixed distribution; the initial and goal states were sampled uniformly randomly in the free space; the cost function $c \\left( \\cdot \\right)$ was set as the sum of the Euclidean path length and the control effort penalty of rotating the robot joints. ", + "bbox": [ + 174, + 791, + 825, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compared NEXT with RRT∗ (Karaman & Frazzoli, 2011), $\\mathrm { B I T ^ { * } }$ (Gammell et al., 2015), CVAE-plan (Ichter et al., 2018), Reinforce-plan (Zhang et al., 2018), and an improved version of GPPN (Lee et al., 2018) in terms of both planning time and solution quality. RRT∗ and $\\mathrm { B I T ^ { * } }$ are two widely used effective instances of TSA in Algorithm 1. In our experiments, we equipped $\\mathrm { R R T ^ { * } }$ with the goal biasing heuristic to improve its performance. $\\mathrm { B I T ^ { * } }$ adopts the informed search strategy (Gammell et al., 2015) to accelerate planning. CVAE-plan and Reinforce-plan are two learning-enhanced TSA planners proposed recently. CVAE-plan learns a conditional VAE as the sampler (Sohn et al., 2015), which will be trained by near-optimal paths produced by $\\mathbf { R R T ^ { * } }$ . Reinforce-plan learns to do rejection sampling with policy gradient methods. For the improved GPPN, we combined its architecture for map with a fully-connected MLP for the rest state, such that it can be applied to high-dimensional continuous spaces. Please refer to Appendix E for more details. ", + "bbox": [ + 173, + 895, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0b18ba9b0a2a21dab4fe0925d12969b8bbd06feb8546ddafdca1a81131ba122.jpg", + "image_caption": [ + "Figure 5: Search trees and the learned ${ \\tilde { V } } ^ { * }$ and $\\tilde { \\pi } ^ { * }$ produced by NEXT. Obstacles are colored in blue. The start and goal locations are denoted by orange and brown dots. In (a) to (c), samples are represented with yellow circles. In (d), the level of redness denotes the value of the cost-to-go estimate $\\tilde { V } ^ { * }$ . The cyan arrows point from a given state $s$ to the mean of the learned policy $\\tilde { \\pi } ^ { * } ( s ^ { \\prime } | s , U )$ . " + ], + "image_footnote": [], + "bbox": [ + 187, + 104, + 810, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9b64d33b209ec3b631b087dbd2e5d1f20da83a228d5b5ee93466c67d2fe104a.jpg", + "image_caption": [ + "Figure 6: Search trees and a solution path produced in an instance of spacecraft planning. The 7 DOF spacecraft has a yellow body and two 2 DOF red arms. NEXT-KS produced a nearly minimum viable search tree while ${ \\mathrm { R R T } } ^ { * }$ failed to find a path within limited trials. " + ], + "image_footnote": [], + "bbox": [ + 207, + 297, + 821, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 464, + 825, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Settings. For each task, we randomly generated 3000 different problems from the same distribution without duplicated maps. We trained all learning-based baselines using the first 2000 problems, and reserved the rest for testing. The parameters for $\\mathrm { { R R T ^ { * } } }$ and $\\mathrm { B I T ^ { * } }$ are also tuned using the first 2000 problems. For NEXT, we let it improve itself using MSIL over the first 2000 problems. In this period, for every 200 problems, we updated its parameters and annealed $\\epsilon$ once. ", + "bbox": [ + 174, + 597, + 825, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 RESULTS AND ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 684, + 387, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison results. Examples of all four environments are illustrated in Appendix F.1 and Figure 6, where NEXT finds high-quality solutions as shown. We also illustrated the comparison of the search trees on two 2d and 7d planning tasks between NEXT and RRT\\* in Figure 5 (a)-(c) and Figure 6 (b) and (c). Obviously, the proposed NEXT algorithm explores with guidance and achieves better quality solutions with fewer samples, while the RRT\\* expands randomly which may fail to find a solution. The learned ${ \\tilde { V } } ^ { * }$ and $\\tilde { \\pi } ^ { * }$ in the 2d task are also shown in Figure 5(d). As we can see, they are consistent with our expectation, towards the ultimate target in the map. For more search tree comparisons for all four experiments, please check Figure 18, 19, 20 and 21 in Appendix F. ", + "bbox": [ + 174, + 707, + 825, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To systematically evaluate the algorithms, we recorded the cost of time (measured by the number of collision checks used) to find a collision-free path, the success rate within time limits, and the cost of the solution path for each run. The results of the reserved 1000 test problems of each environment are shown in the top row of Figure 7. We set the maximal number of samples as 500 for all algorithms. Both the kernel smoothing (NEXT-KS) and the Gaussian process (NEXT-GP) version of NEXT achieves the state-of-the-art performances, under all three criteria in all test environments. Although the $\\mathrm { B I T ^ { * } }$ utilizes the heuristic particularly suitable for 3d maze in 7d task and performs quite well, the NEXT algorithm still outperform every competitor by a large margin, no matter learning-based or prefixed heuristic planner, demonstrating the advantages of the proposed NEXT algorithm. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1b7307ae6db488a13797cccd31ec60ba79f87159ac956c545b9f9d1f3fe33ce.jpg", + "image_caption": [ + "Figure 7: First row: histograms of results, in terms of success rate, average collision checks, and average cost of the solution paths; Second row: NEXT improvement curves in the 5d experiments. All algorithms are set to use up to 500 samples, except $\\mathrm { R R T ^ { * } }$ -10k, which uses 10,000 samples. The value of collision checks and path costs are normalized w.r.t. the performance of $\\mathrm { R R T ^ { * } }$ . " + ], + "image_footnote": [], + "bbox": [ + 173, + 102, + 820, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 470, + 820, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Self-improving. We plot the performance improvement curves of our algorithms on the 5d planning task in the bottom row of Figure 7. For comparison, we also plot the performance of RRT∗. At the beginning phase of self-improving, our algorithms are comparable to $\\mathrm { R R T ^ { * } }$ . They then gradually learn from previous experiences and improve themselves as they see more problems and better solutions. In the end, NEXT-KS is able to match the performance of RRT∗-10k using only one-twentieth of its samples, while the competitors perform consistently without any improvements. ", + "bbox": [ + 174, + 506, + 825, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Due to the space limits, we put improvement curves on other environments in Figure 17 and the quantitative evaluation in Table 1, 2, and 3 in Appendix F. Please refer to the details there. ", + "bbox": [ + 176, + 597, + 821, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3 ABLATION STUDIES ", + "text_level": 1, + "bbox": [ + 174, + 633, + 352, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ablation study I: guided progressive expansion. To demonstrate the power of NEXT :: Expand, we replace it with breadth-first search (BFS) (Kim et al., 2018), another expanding strategy, while keeping other components the same. Specifically, BFS uses a search queue in planning. It repeatedly pops a state $s$ out from the search queue, samples $k$ states from $\\pi \\left( \\cdot | s \\right)$ , and pushes all generated samples and state $s$ back to the queue, until the goal is reached. For fairness, we use the learned sampling policy $\\pi \\left( s ^ { \\prime } | s , U \\right)$ by NEXT-KS in BFS. As shown in Figure 7, BFS obtained worse paths with a much larger number of collision checks and far lower success rate, which justifies the importance of the balance between exploration versus exploitation achieved by the proposed NEXT :: Expand. ", + "bbox": [ + 174, + 652, + 825, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ablation study II: neural architecture. To further demonstrate the benefits of the proposed neural architecture for learning generalizable representations in high-dimension planning problems, We replaced our attention-based neural architecture with an improved GPPN, as explained in Appendix E, for ablation study. We extended the GPPN for continuous space by adding an extra reactive policy network to its final layers. We emphasize the original GPPN is not applicable to the tasks in our experiments. Intuitively, the improved GPPN first produces a ‘rough plan’ by processing the robot’s discretized workspace positions. Then the reactive policy network predicts a continuous action from both the workspace feature and the full configuration state of the robot. We provide more preference to the improved GPPN by training it to imitate the near-optimal paths produced by RRT\\* in the training problems. During test time it is also combined with both versions of the guided progressive expansion operators. As we can see, both GPPN-KS and GPPN-GP are clearly much worse than NEXT-KS and NEXT-GP, demonstrating the advantage of our proposed attention-based neural architecture in high-dimensional planning tasks. ", + "bbox": [ + 174, + 771, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a7f19f8a9284429626e2ad6f69c231d6aa6a4e1e1021040a5d3d6919d297a607.jpg", + "image_caption": [ + "Figure 8: The success rate and average path cost of the different planners under varying time limits. Running NEXT for 1 second achieves the same success rate as running $\\mathrm { B I T ^ { * } }$ for 50 seconds. " + ], + "image_footnote": [], + "bbox": [ + 287, + 102, + 709, + 242 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 290, + 821, + 319 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ablation study III: learning versus heuristic. The NEXT algorithm in Figure 5 shows similar behavior as the Dijkstra heuristic, i.e. sampling on the shortest path connecting the start and the goal in workspace. However, in higher dimensional space, the Dijkstra heuristic will fail. To demonstrate that, we replace the policy and value network with Dijkstra heuristic, using the best direction in workspace to guide sampling. NEXT performs much better than Dijkstra in all but the 2d case, in which the workspace happens to be the state space. ", + "bbox": [ + 174, + 325, + 825, + 409 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.4 CASE STUDY: ROBOT ARM CONTROL ", + "text_level": 1, + "bbox": [ + 176, + 430, + 473, + 444 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct a real-world case study on controlling robot arms to move objects on a shelf. On this representative real-time task, we demonstrate the advantages of the NEXT in terms of the wall-clock. ", + "bbox": [ + 174, + 455, + 825, + 484 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In each planning task, there is a shelf of multiple levels, with each level horizontally divided into multiple bins. The task is to plan a path from a location in one bin to another, i.e., the end effectors of the start and goal configurations are in different bins. The heights of levels, widths of bins, and the start and goal are randomly drawn from some fixed distribution. Different from previous experiments, the base of the robot is fixed. We consider the $\\mathrm { B I T ^ { * } }$ instead of $\\mathrm { R R T ^ { * } }$ as the imperfect expert in 3000 training problems. We then evaluate the algorithm on a separated 1000 testing problems. We compare NEXT(- KS) with the highly tuned $\\mathrm { B I T ^ { * } }$ and RRT\\* in OMPL, and also CVAE-plan and Reinforce-plan in Figure 8. As seen from the visualization of the found paths in Figure 9, this is a very difficult task. Our NEXT outperforms the baselines by a large margin, requiring only 1 second to reach the same success rate as running 50 seconds of $\\mathrm { B I T ^ { * } }$ . ", + "bbox": [ + 174, + 491, + 544, + 739 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2c27831a81842e20acfb22f33643b38f7e0d636dac989887fae6e6036e1293d0.jpg", + "image_caption": [ + "Figure 9: The collision-free path produced by NEXT for robot arm planning. The start and goal configurations have end-effectors in different bins of the shelf. " + ], + "image_footnote": [], + "bbox": [ + 604, + 491, + 750, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Due to space limits, we put details of the experiment setups, more results and analysis in Appendix F.4. ", + "bbox": [ + 173, + 747, + 823, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 784, + 318, + 799 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this paper, we propose a self-improving planner, Neural EXploration-EXploitation Trees (NEXT), which can generalize and achieve better performance with experiences accumulated. The algorithm achieves a delicate balance between exploration-exploitation via our carefully designed UCB-type expansion operation. To obtain the generalizable ability across different problems, we proposed a new parametrization for the value function and policy, which captures the Bellman recursive structure in the high-dimensional continuous state and action space. We demonstrate the power of the proposed algorithm by outperforming previous state-of-the-art planners with significant margins on planning problems in a variety of different environments. ", + "bbox": [ + 174, + 811, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ACKNOWLEDGEMENT ", + "text_level": 1, + "bbox": [ + 176, + 103, + 357, + 117 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We thank the Google Research Brain team members for helpful thoughts and discussions as well as the anonymous reviewers for their insightful comments and suggestions. This work is supported in part by NSF grants CDS&E-1900017 D3SC, CCF-1836936 FMitF, IIS-1841351, CAREER IIS1350983 to L.S, and by NSF grants BIGDATA 1840866, CAREER 1841569, TRIPODS 1740735, DARPA-PA-18-02-09-QED-RML-FP-003, an Alfred P Sloan Fellowship, a PECASE award to H.L. ", + "bbox": [ + 174, + 133, + 825, + 204 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 224, + 285, + 239 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Auer, P., Cesa-Bianchi, N., and Fischer, P. Finite-time analysis of the multiarmed bandit problem. Machine learning, 47(2-3):235–256, 2002. ", + "bbox": [ + 176, + 247, + 823, + 276 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mohak Bhardwaj, Sanjiban Choudhury, and Sebastian Scherer. Learning heuristic search via imitation. arXiv preprint arXiv:1707.03034, 2017. ", + "bbox": [ + 173, + 286, + 823, + 315 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Boor, V., Overmars, M. H., and Van Der Stappen, A. F. The gaussian sampling strategy for probabilistic roadmap planners. In Robotics and automation, 1999. proceedings. 1999 ieee international conference on, volume 2, pp. 1018–1023. IEEE, 1999. ", + "bbox": [ + 173, + 324, + 825, + 367 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Burns, B. and Brock, O. Sampling-based motion planning using predictive models. In Robotics and Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference on, pp. 3120–3125. IEEE, 2005a. ", + "bbox": [ + 176, + 377, + 825, + 420 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Burns, B. and Brock, O. Toward optimal configuration space sampling. In Robotics: Science and Systems, pp. 105–112. Citeseer, 2005b. ", + "bbox": [ + 173, + 428, + 825, + 458 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Bowen, Chris, and Ron Alterovitz. Closed-loop global motion planning for reactive execution of learned tasks. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1754-1760, 2014. ", + "bbox": [ + 173, + 467, + 825, + 510 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cadena Cesar, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J. Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on Robotics, 32(6), 2016. ", + "bbox": [ + 174, + 518, + 825, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sanjiban Choudhury, Mohak Bhardwaj, Sankalp Arora, Ashish Kapoor, Gireeja Ranade, Sebastian Scherer, and Debadeepta Dey. Data-driven planning via imitation learning. The International Journal of Robotics Research, 37(13-14):1632–1672, 2018. ", + "bbox": [ + 174, + 570, + 825, + 614 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Chu, W., Li, L., Reyzin, L., and Schapire, R. Contextual bandits with linear payoff functions. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp. 208–214, 2011. ", + "bbox": [ + 173, + 622, + 826, + 666 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Couëtoux, A., Hoock, J.-B., Sokolovska, N., Teytaud, O., and Bonnard, N. Continuous upper confidence trees. In International Conference on Learning and Intelligent Optimization, pp. 433–445. Springer, 2011. ", + "bbox": [ + 173, + 675, + 823, + 718 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rosen Diankov. Automated Construction of Robotic Manipulation Programs. PhD thesis, Carnegie Mellon University, Robotics Institute, August 2010. ", + "bbox": [ + 169, + 727, + 825, + 757 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rosen Diankov and James Kuffner. Randomized statistical path planning. In 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1–6. IEEE, 2007. ", + "bbox": [ + 173, + 766, + 823, + 796 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Elbanhawi, M. and Simic, M. Sampling-based robot motion planning: A review. Ieee access, 2: 56–77, 2014. ", + "bbox": [ + 173, + 804, + 825, + 833 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Finney, S., Kaelbling, L. P., and Lozano-Perez, T. Predicting partial paths from planning problem parameters. In Robotics Science and Systems, 2007. ", + "bbox": [ + 171, + 843, + 823, + 872 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fulgenzi, Chiara and Tay, Christopher and Spalanzani, Anne and Laugier, Christian Probabilistic navigation in dynamic environment using rapidly-exploring random trees and gaussian processes In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1056–1062, 2008. ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gammell, J. D., Srinivasa, S. S., and Barfoot, T. D. Informed $\\mathrm { { r r t } ^ { * } }$ : Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic. arXiv preprint arXiv:1404.2334, 2014. \nGammell, J. D., Srinivasa, S. S., and Barfoot, T. D. Batch informed trees (bit\\*): Sampling-based optimal planning via the heuristically guided search of implicit random geometric graphs. In Robotics and Automation (ICRA), 2015 IEEE International Conference on, pp. 3067–3074. IEEE, 2015. \nGuez, Arthur and Weber, Théophane and Antonoglou, Ioannis and Simonyan, Karen and Vinyals, Oriol and Wierstra, Daan and Munos, Rémi and Silver, David Learning to search with mctsnets arXiv preprint arXiv:1802.04697, 2018. \nPeter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100–107, 1968. \nHsu, D., Latombe, J.-C., and Motwani, R. Path planning in expansive configuration spaces. In Robotics and Automation, 1997. Proceedings., 1997 IEEE International Conference on, volume 3, pp. 2719–2726. IEEE, 1997. \nHsu, D., Jiang, T., Reif, J., and Sun, Z. The bridge test for sampling narrow passages with probabilistic roadmap planners. In Robotics and Automation, 2003. Proceedings. ICRA’03. IEEE International Conference on, volume 3, pp. 4420–4426. IEEE, 2003. \nHsu, D., Sánchez-Ante, G., and Sun, Z. Hybrid prm sampling with a cost-sensitive adaptive strategy. In Robotics and Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference on, pp. 3874–3880. IEEE, 2005. \nHuh, Jinwook and Lee, Daniel Efficient Sampling With Q-Learning to Guide Rapidly Exploring Random Trees. IEEE Robotics and Automation Letters, 3:3868–3875, 2018. \nIchter, B., Harrison, J., and Pavone, M. Learning sampling distributions for robot motion planning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 7087–7094. IEEE, 2018. \nKaraman, S. and Frazzoli, E. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846–894, 2011. \nKarkus, P., Hsu, D., and Lee, W. S. Qmdp-net: Deep learning for planning under partial observability. In Advances in Neural Information Processing Systems, pp. 4694–4704, 2017. \nKavraki, L. E., Svestka, P., Latombe, J.-C., and Overmars, M. H. Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics and Automation, 12(4), 1996. \nKim, B., Kaelbling, L. P., and Lozano-Pérez, T. Guiding search in continuous state-action spaces by learning an action sampler from off-target search experience. 2018. \nKocsis, L. and Szepesvári, C. Bandit based monte-carlo planning. In European conference on machine learning, pp. 282–293. Springer, 2006. \nKrause, A. and Ong, C. S. Contextual gaussian process bandit optimization. In Advances in Neural Information Processing Systems, pp. 2447–2455, 2011. \nKuo, Yen-Ling and Barbu, Andrei and Katz, Boris Deep sequential models for sampling-based planning In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 6490–6497, 2018. \nLangford, J. and Zhang, T. The epoch-greedy algorithm for multi-armed bandits with side information. In Advances in neural information processing systems, pp. 817–824, 2008. \nLaValle, S. M. Rapidly-exploring random trees: A new tool for path planning. 1998. \nLee, L., Parisotto, E., Chaplot, D. S., Xing, E., and Salakhutdinov, R. Gated path planning networks. arXiv preprint arXiv:1806.06408, 2018. ", + "bbox": [ + 171, + 80, + 828, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine learning, 8(3-4):293–321, 1992. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Paxton, Chris and Raman, Vasumathi and Hager, Gregory D and Kobilarov, Marin Combining neural networks and tree search for task and motion planning in challenging environments In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 6059–6066, 2017. ", + "bbox": [ + 174, + 140, + 823, + 184 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sven Mikael Persson and Inna Sharf. Sampling-based a\\* algorithm for robot path-planning. The International Journal of Robotics Research, 33(13):1683–1708, 2014. ", + "bbox": [ + 173, + 193, + 823, + 222 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Phillips, J. M., Bedrossian, N., and Kavraki, L. E. Guided expansive spaces trees: A search strategy for motion-and cost-constrained state spaces. In IEEE International Conference on Robotics and Automation, pp. 3968–3973, 2004. ", + "bbox": [ + 176, + 229, + 823, + 273 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Mike Phillips, Benjamin J Cohen, Sachin Chitta, and Maxim Likhachev. E-graphs: Bootstrapping planning with experience graphs. In Robotics: Science and Systems, volume 5, pp. 110, 2012. ", + "bbox": [ + 173, + 281, + 821, + 311 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qureshi, Ahmed H and Simeonov, Anthony and Bency, Mayur J and Yip, Michael C. Motion planning networks. In IEEE International Conference on Robotics and Automation, 2019. ", + "bbox": [ + 174, + 319, + 821, + 349 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reif, J. H. Complexity of the mover’s problem and generalizations. In Foundations of Computer Science, 1979., 20th Annual Symposium on, pp. 421–427. IEEE, 1979. ", + "bbox": [ + 174, + 357, + 823, + 387 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Rickert, Markus and Brock, Oliver and Knoll, Alois. Balancing exploration and exploitation in motion planning In IEEE International Conference on Robotics and Automation, pp. 2812–2817. IEEE, 2008. ", + "bbox": [ + 173, + 395, + 823, + 439 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. ", + "bbox": [ + 173, + 446, + 823, + 477 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Shkolnik, A., Walter, M., and Tedrake, R. Reachability-guided sampling for planning under differential constraints. In Robotics and Automation, 2009. ICRA’09. IEEE International Conference on, pp. 2859–2865. IEEE, 2009. ", + "bbox": [ + 171, + 484, + 825, + 527 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Silver, David and Hubert, Thomas and Schrittwieser, Julian and Antonoglou, Ioannis and Lai, Matthew and Guez, Arthur and Lanctot, Marc and Sifre, Laurent and Kumaran, Dharshan and Graepel, Thore and others Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint arXiv:1712.01815, 2017. ", + "bbox": [ + 174, + 536, + 826, + 594 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sohn, K., Lee, H., and Yan, X. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems, pp. 3483–3491, 2015. ", + "bbox": [ + 169, + 602, + 825, + 632 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jialin Song, Ravi Lanka, Albert Zhao, Yisong Yue, and Masahiro Ono. Learning to search via retrospective imitation. arXiv preprint arXiv:1804.00846, 2018. ", + "bbox": [ + 171, + 640, + 823, + 670 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Srinivas, N., Krause, A., Kakade, S. M., and Seeger, M. Gaussian process optimization in the bandit setting: No regret and experimental design. arXiv preprint arXiv:0912.3995, 2009. ", + "bbox": [ + 169, + 678, + 823, + 708 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ioan A. ¸Sucan, Mark Moll, and Lydia E. Kavraki. The Open Motion Planning Library. IEEE Robotics & Automation Magazine, 19(4):72–82, December 2012. doi: 10.1109/MRA.2012.2205651. http://ompl.kavrakilab.org. ", + "bbox": [ + 173, + 715, + 825, + 760 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tamar, A., Wu, Y., Thomas, G., Levine, S., and Abbeel, P. Value iteration networks. In Advances in Neural Information Processing Systems, pp. 2154–2162, 2016. ", + "bbox": [ + 171, + 767, + 825, + 797 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Wang, Y., Audibert, J.-Y., and Munos, R. Algorithms for infinitely many-armed bandits. In Advances in Neural Information Processing Systems, pp. 1729–1736, 2009. ", + "bbox": [ + 171, + 805, + 823, + 835 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ye, Gu and Alterovitz, Ron Guided motion planning. In Robotics research, pp. 291–6307. Springer, 2017. ", + "bbox": [ + 171, + 843, + 825, + 872 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Yee, T., Lisy, V., and Bowling, M. Monte carlo tree search in continuous action spaces with execution uncertainty. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, pp. 690–696. AAAI Press, 2016. ", + "bbox": [ + 176, + 881, + 825, + 924 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Zhang, C., Huh, J., and Lee, D. D. Learning implicit sampling distributions for motion planning. arXiv preprint arXiv:1806.01968, 2018. ", + "bbox": [ + 169, + 103, + 825, + 132 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Zucker, M., Kuffner, J., and Bagnell, J. A. Adaptive workspace biasing for sampling-based planners. In Robotics and Automation, 2008. ICRA 2008. IEEE International Conference on, pp. 3757–3762. IEEE, 2008. ", + "bbox": [ + 173, + 140, + 826, + 183 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Appendix ", + "text_level": 1, + "bbox": [ + 433, + 99, + 566, + 127 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A ILLUSTRATION OF THE DIFFICULTY IN PLANNING PROBLEMS ", + "text_level": 1, + "bbox": [ + 174, + 143, + 718, + 161 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The Figure 10(a) illustrates a concrete planning problem for a stick robot in 2d workspace. With one extra continuous action for rotation, the configuration state is visualized in Figure 10(b), which is highly irregular and unknown to the planner. ", + "bbox": [ + 173, + 176, + 826, + 219 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/03f4d274b006a2d945a961a75df61f65f470c9506bae0cc5429f4ef3d09ead9c.jpg", + "image_caption": [ + "Figure 10: Different views of the same planning problem. In (a) we color the obstacles, the starting and the goal position of the robot in deep blue, orange and brown, respectively. The stick robot can move and rotate. The corresponding configuration space is 3d, as visualized in (b), with the extra dimension being the rotation angle w.r.t. the $\\mathbf { X }$ -axis. The blue region indicates the feasible state space, i.e., the set of collision-free states. The starting and the goal position are denoted with an orange and a brown dot, respectively. Although the workspace looks trivial, the configuration space is irregular, which makes the planning difficult. " + ], + "image_footnote": [], + "bbox": [ + 258, + 229, + 761, + 391 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B MORE PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 176, + 518, + 400, + 535 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Tree-based sampling planner The tree-based sampling planner algorithm is illustrated in Figure 11. The Expand in Algorithm 1 operator returns an existing node in the tree $s _ { p a r e n t } \\in \\mathcal { V }$ and a new state $s _ { n e w } \\in S$ sampled from the neighborhood of $s _ { p a r e n t }$ . Then the line segment $[ s _ { p a r e n t } , s _ { n e w } ]$ is passed to function ObstacleFree for collision checking. If the line segment $[ s _ { p a r e n t } , s _ { n e w } ]$ is collision-free (no obstacle in the middle, or called reachable from $\\tau$ ), then $s _ { n e w }$ is added to the tree vertex set $\\nu$ , and the line segment is added to the tree edge set $\\mathcal { E }$ . If the newly added node $s _ { n e w }$ has reached the target $\\mathcal { S } _ { g o a l }$ , the algorithm will return. Optionally, some concrete algorithms can define a Postprocess operator to refine the search tree. For an example of the Expand operator, as shown in Figure 1 (c), since there is no obstacle on the dotted edge $[ s _ { p a r e n t } , s _ { n e w } ]$ , i.e., $s _ { n e w }$ is reachable, the new state and edge will be added to the search tree (connected by the solid edges). ", + "bbox": [ + 173, + 550, + 826, + 690 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/be00082f31c012e5df15315075e706a96419afcb7dbc1ba05cf0204fbd263e3d.jpg", + "image_caption": [ + "Figure 11: Illustration for one iteration of Algorithm 1. The left and right figures illustrate two different cases where the sample returned by the Expand operator is unreachable and reachable from the search tree. " + ], + "image_footnote": [], + "bbox": [ + 312, + 698, + 686, + 810 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Now we will provide two concrete algorithm examples. For instance, • If we instantiate the Expand operator as Algorithm 4, then we obtain the rapidly-exploring random trees (RRT) algorithm (LaValle, 1998), which first samples a state $s$ from the configuration space ", + "bbox": [ + 174, + 868, + 627, + 883 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 895, + 820, + 924 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "$\\mathbf { A l g o r i t h m 4 : R R T : : E x p a n d } ( { T } , { U } )$ ", + "text_level": 1, + "bbox": [ + 174, + 108, + 410, + 122 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Data: $\\mathcal { T } = ( \\mathcal { V } , \\mathcal { E } ) , U = \\left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c ( \\cdot ) \\right)$ 1 $s _ { r a n d } \\gets \\mathcal { U } n i f ( \\mathcal { S } )$ ; 2 $\\begin{array} { r } { s _ { p a r e n t } \\gets \\operatorname * { a r g m i n } _ { s \\in \\mathcal { V } } \\left\\| s _ { r a n d } - s \\right\\| } \\end{array}$ ; 3 $s _ { n e w } \\gets \\mathrm { a r g m i n } _ { s \\in \\mathcal { B } ( s _ { p a r e n t } , \\eta ) } \\left\\| s - s _ { r a n d } \\right\\|$ ; ", + "bbox": [ + 161, + 126, + 550, + 186 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "4 return $s _ { p a r e n t } , s _ { n e w }$ . Sample configuration space . Pull to a tree node ", + "bbox": [ + 161, + 188, + 316, + 200 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "", + "bbox": [ + 604, + 141, + 800, + 169 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "$\\mathbf { A l g o r i t h m 5 : E S T : : E x p a n d } ( T , U )$ ", + "text_level": 1, + "bbox": [ + 174, + 218, + 410, + 233 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Data: $\\mathcal { T } = ( \\mathcal { V } , \\mathcal { E } ) , U = \\left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c ( \\cdot ) \\right)$ ，1 sparent $\\sim \\phi ( s ) , s \\in \\mathcal { V }$ ;2 $\\hat { s _ { n e w } } \\mathcal { U } n i f ( B ( s _ { p a r e n t } ) )$ ;3 return snearest, snew;", + "bbox": [ + 160, + 236, + 552, + 295 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": ". Sample a tree node . Sample neighborhood ", + "bbox": [ + 640, + 252, + 800, + 280 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "$s$ and then pulls it toward the neighborhood of current tree $\\tau$ measured by a ball of radius $\\eta$ ", + "bbox": [ + 184, + 308, + 797, + 323 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Moreover, if the Postprocess operator is introduced to modify the maintained search tree as in $\\mathrm { { R R T ^ { * } } }$ (Karaman & Frazzoli, 2011), the algorithm is provable to obtain the optimal path asymptotically. ", + "bbox": [ + 187, + 342, + 825, + 383 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "• If we instantiate the Expand operator as Algorithm 5, then we obtain the expansive-space trees (EST) algorithm (Hsu et al., 1997; Phillips et al., 2004), which samples a state $s$ from the nodes of the existing tree, and then draw a sample from the neighborhood of $s$ . ", + "bbox": [ + 176, + 387, + 825, + 429 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "UCB-based algorithms Specifically, in a $K$ -armed bandit problem, the UCB algorithm will first play each of the arms once, and then keep track of the average reward $\\bar { r } _ { i }$ and the visitation count $n _ { i }$ for each arm. After $T$ rounds of trials, the UCB algorithm will maintain a set of information $\\{ ( \\bar { r } _ { i } , n _ { i } ) \\} _ { i = 1 } ^ { K }$ with $\\textstyle \\sum _ { i = 1 } ^ { K } n _ { i } = T$ . Then, for the next round, the UCB algorithm will select the next arm based on the one-sided confidence interval estimation provided by the Chernoff-Hoeffding bound, ", + "bbox": [ + 173, + 448, + 825, + 521 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/bb94fadba4fdb19b851b3da9d5f8f9946e77c39171639687da54124e30089e17.jpg", + "text": "$$\n\\begin{array} { r } { a _ { T + 1 } ^ { * } = \\operatorname * { a r g m a x } _ { i \\in \\{ 1 , \\ldots , K \\} } \\bar { r } _ { i } + \\lambda \\sqrt { \\frac { \\log T } { n _ { i } } } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 359, + 532, + 635, + 559 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "where $\\lambda$ controls the exploration-exploration trade-off. It has been shown that the UCB algorithm achieves $\\mathcal { O } \\left( \\log T \\right)$ regret. However, the MCTS is not directly applicable to continuous state-action spaces. ", + "bbox": [ + 174, + 559, + 825, + 602 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "There have been many attempts to generalize the UCB and UCT algorithms to continuous state-action spaces (Chu et al., 2011; Krause & Ong, 2011; Couëtoux et al., 2011; Yee et al., 2016). For instance, contextual bandit algorithms allow continuous arms but involve a non-trivial high dimensional non-convex optimization to select the next arm. In UCT, the progressive widening technique has been designed to deal with continuous actions (Wang et al., 2009). Even with these extensions, the MCTS restricts the exploration only from leaves states, implicitly adding an unnecessary hierarchical structure for path planning, resulting inferior exploration efficiency and extra computation in path planning tasks. ", + "bbox": [ + 173, + 608, + 825, + 720 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Although these off-the-shelf algorithms are not directly applicable to our path planning setting, their successes show the importance of exploration-exploitation trade-off and will provide the principles for our algorithm for continuous state-action planning problems. ", + "bbox": [ + 176, + 727, + 821, + 770 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Planning networks Value iteration networks (Tamar et al., 2016) employ neural networks to embed the value iteration algorithm from planning and then use this embedded algorithm to extract input features and define downstream models such as value functions and policies. ", + "bbox": [ + 174, + 784, + 825, + 827 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Specifically, VIN mimics the following recursive application of Bellman update operator $\\mathcal { G }$ to value function $V ^ { * }$ , ", + "bbox": [ + 176, + 833, + 825, + 861 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/4226ddc32831387ca8faaca832db0dc985d7476a84f1955aad452ab384d5c4ea.jpg", + "text": "$$\nV ^ { * } \\left( s | U \\right) = \\left( \\mathcal { G } V ^ { * } \\right) ( s ) : = \\operatorname* { m i n } _ { a } \\sum _ { s ^ { \\prime } } P ( s ^ { \\prime } | s , a ) ( c \\left( \\left[ s , s ^ { \\prime } \\right] \\right) + V ^ { * } \\left( s ^ { \\prime } | U \\right) ) .\n$$", + "text_format": "latex", + "bbox": [ + 266, + 861, + 730, + 893 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "where $P ( s ^ { \\prime } | s , a )$ is the state transition model. When the state space for $s$ and action space for $a$ are low dimensional, these spaces can be discretized into grids. Then, the local cost function $c ( [ s , s ^ { \\prime } ] )$ and the value function $V ^ { * } ( s ^ { \\prime } | U )$ can be represented as matrices (2d) or tensors (3d) with each entry indexed by grid locations. Furthermore, if the transition model $P ( s ^ { \\prime } | s , a )$ is local, that is $P ( s ^ { \\prime } | \\dot { s } , a ) = 0$ for $\\bar { \\boldsymbol { s } } ^ { \\prime } \\notin B ( \\boldsymbol { s } )$ , it resembles a set of convolution kernels, each indexed by a discrete action $a$ . And the Bellman update operator essentially convolves $P ( s ^ { \\prime } | s , a )$ with $c ( [ s , s ^ { \\bar { \\prime } } ] )$ and $V ^ { * } ( s ^ { \\prime } | U )$ , and then performs a min-pooling operation across the convolution channels. ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 102, + 826, + 174 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Inspired by the above computation pattern of the Bellman operator, value iteration networks design the neural architecture as follows, ", + "bbox": [ + 173, + 179, + 823, + 208 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/dee3f8bfd36cb521afcd22f462a62dffa0835e6ce5931e74fd22a401561b036a.jpg", + "text": "$$\n\\begin{array} { r l } & { \\tilde { V } ^ { * 0 } = \\operatorname* { m i n } \\ \\left( W _ { 1 } \\oplus \\left[ \\tt { m a p } , \\tilde { R } \\right] \\right) } \\\\ & { \\tilde { V } ^ { * t } = \\operatorname* { m i n } \\ \\left( W _ { 1 } \\oplus \\left[ \\tilde { V } ^ { * t - 1 } , \\tilde { R } \\right] \\right) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 383, + 208, + 612, + 261 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where $\\oplus$ is the convolution operation, both map, $\\tilde { V } ^ { * t }$ and $\\tilde { R }$ are $d \\times d$ matrices, and the parameter $W _ { 1 }$ are $k _ { c }$ convolution kernels of size $k \\times k$ . The min implements the pooling across $k _ { c }$ convolution channels. ", + "bbox": [ + 173, + 262, + 825, + 305 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The gated path planning networks (GPPN) (Lee et al., 2018) improves the VIN by replacing the VIN cell (9) with the well-established LSTM update, i.e., ", + "bbox": [ + 174, + 310, + 823, + 340 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/29d45bb013ca6188267f130a8fe3ab6c4c91b7fa1a03866c1230f433c3afccf3.jpg", + "text": "$$\n\\tilde { V } ^ { t } , \\tilde { c } ^ { t } = \\mathrm { L S T M } \\left( \\sum \\left( W _ { 1 } \\oplus \\left[ \\tilde { V } ^ { t } , \\tilde { R } \\right] \\right) , \\tilde { c } ^ { t } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 352, + 340, + 643, + 367 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where the summation is taking over all the $k _ { c }$ convolution channels. ", + "bbox": [ + 174, + 367, + 617, + 382 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "After constructing the VIN and GPPN, the parameters of the model, i.e., $\\left\\{ \\tilde { R } , W _ { 1 } \\right\\}$ can be learned by imitation learning or reinforcement learning. ", + "bbox": [ + 176, + 390, + 821, + 426 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "The application of planning networks are restricted in low-dimension tasks. However, their success enlightens our neural architecture for generalizable representation for high-dimension planning tasks. ", + "bbox": [ + 174, + 433, + 823, + 462 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "C PARAMETRIZED UCB ALGORITHMS ", + "text_level": 1, + "bbox": [ + 173, + 484, + 511, + 502 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We list two examples of parametrized UCB as the instantiation of (3) used in GPE: ", + "bbox": [ + 173, + 518, + 715, + 535 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• GP-UCB: The GP-UCB Chu et al. (2011) is derived by parameterizing via Gaussian Processes (GP) with kernel $k \\left( s , s ^ { \\prime } \\right)$ , i.e., E $[ r ( s ) | \\mathcal { T } , U ] \\sim \\mathcal { \\bar { G } } \\mathcal { P } ( 0 , k )$ , GP-UCB maintains an UCB of the reward after t-step asφ (s) := ¯rt (s) + λσt (s) , (12) ", + "bbox": [ + 217, + 542, + 825, + 597 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 232, + 598, + 274, + 612 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/fbd68378e007a673e477047db489ceaba05b0f68da9498aa6be27ea6816d48ad.jpg", + "text": "$$\n\\begin{array} { r c l } { { { \\bar { r } } _ { t } } \\left( s \\right) } & { { = } } & { { { k _ { t } } \\left( s \\right) \\left( K _ { t } + \\alpha I \\right) ^ { - 1 } { r _ { t } } , } } \\\\ { { \\sigma _ { t } ^ { 2 } \\left( s \\right) } } & { { = } } & { { { k \\left( s , s \\right) - k _ { t } } \\left( s \\right) ^ { \\top } \\left( K _ { t } + \\alpha I \\right) ^ { - 1 } k _ { t } \\left( s \\right) , } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 354, + 611, + 699, + 648 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "with $\\boldsymbol { k } _ { t } \\left( \\boldsymbol { s } \\right) = \\left[ k \\left( s _ { i } , \\boldsymbol { s } \\right) \\right] _ { s _ { i } \\in \\boldsymbol { S } _ { t } }$ , $K _ { t } = [ k \\left( s , s ^ { \\prime } \\right) ] _ { s , s ^ { \\prime } \\in S _ { t } }$ , and ${ \\cal { S } } _ { t } = \\{ s _ { 1 } , s _ { 2 } , \\ldots , s _ { t } \\}$ denotes the sequence of selected nodes in current trees. The variance estimation $\\sigma _ { t } ^ { 2 } \\left( s \\right)$ takes the number of visits into account in an implicit way: the variance will reduce, as the neighborhood of $s$ is visited more frequently (Srinivas et al., 2009). ", + "bbox": [ + 230, + 648, + 825, + 708 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• KS-UCB: We can also use kernel regression as an alternative parametrization for (7) (Yee et al., 2016), which leads to an UCB of the reward after $t$ -step as ", + "bbox": [ + 215, + 715, + 823, + 744 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/b215a1dbae611adf3ab8688a0d0700b31f4578d9a215b8ac318d12ccd63dc469.jpg", + "text": "$$\n\\phi \\left( s \\right) : = { \\bar { r } } _ { t } \\left( s \\right) + \\lambda \\sigma _ { t } \\left( s \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 441, + 744, + 614, + 762 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "where ", + "bbox": [ + 232, + 762, + 274, + 776 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/f1483a983d3b9d895751aa88d2e56eed7a6c2602aaadb294a53ab188a30bd0d3.jpg", + "text": "$$\n\\begin{array} { r l r } { \\bar { r } _ { t } \\left[ s \\right] } & { = } & { \\frac { \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) r \\left( s ^ { \\prime } \\right) } { \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) } , } \\\\ { \\sigma _ { t } \\left( s \\right) } & { = } & { \\sqrt { \\frac { \\log \\sum _ { s ^ { \\prime } \\in S _ { t } } w \\left( s ^ { \\prime } \\right) } { w \\left( s \\right) } } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 408, + 773, + 647, + 853 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "with $\\begin{array} { r } { w \\left( s \\right) = \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) } \\end{array}$ . Clearly, the variance estimation is to promote exploration towards less frequently visited states. ", + "bbox": [ + 230, + 853, + 823, + 882 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "As we can see, in both two examples of the parametrized UCB, we parametrize the observed rewards, leading to generalizable UCB for increased states by considering the correlations. ", + "bbox": [ + 174, + 895, + 825, + 924 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D POLICY AND VALUE NETWORK ARCHITECTURE ", + "text_level": 1, + "bbox": [ + 173, + 102, + 611, + 118 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "We explain the implementation details of the proposed parametrization for policy and value function. Figure 12 and Figure 13 are neural architectures for the attention module, the policy/value network, and the planning module, respectively. ", + "bbox": [ + 174, + 162, + 826, + 204 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/fb3272bba1630226b3a65a134ae6d6e537e4da69dcd4e45215968c4d3b3ba712.jpg", + "image_caption": [ + "Figure 12: Left: attention module, instantiating the Figure 3; Right: policy/value network, instantiating the Figure 4. " + ], + "image_footnote": [], + "bbox": [ + 192, + 207, + 807, + 685 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In the figures, we use rectangle blocks to denote inputs, intermediate results and outputs, stadium shape blocks to denote operations, and rounded rectangle blocks to denote modules. We use different colors for different operations. In particular, we use blue for convolutional/LSTM layers, green for dense layers, and orange for anything else. For convolutional layers, \"Conv $1 \\times 1$ , 32, relu\" denotes a layer with $1 \\times 1$ kernels, 32 channels, followed by a rectified linear unit; for dense layers, \"Dense, 64, relu\" denotes a layer of size 64, followed by a rectified linear unit. ", + "bbox": [ + 173, + 728, + 825, + 811 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "The attention module (Figure 12-left) embeds a state to a $d \\times d \\times d _ { a }$ tensor. The planning module (Figure 13) is a one-step LSTM update which takes the result of a convolutional layer as input. Both the input and hidden size of the LSTM cell are $d _ { e }$ . All $d \\times d$ locations share one set of parameters and are processed by the LSTM in one batch. ", + "bbox": [ + 174, + 819, + 825, + 875 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "The main architecture is illustrated in Figure 12-right. It takes maze map, state and goal as input, and outputs the action and the value. Refer to Section 4.2 for details for computing $\\psi ( s )$ . In our experiments, we set the values of the hyper-parameters to be $( d , d _ { e } , d _ { a } , p ) = ( 1 \\bar { 5 } , 6 4 , \\bar { 8 } , 8 )$ . ", + "bbox": [ + 174, + 882, + 825, + 924 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/093f34c3bdcd9d32756f1eaf06749627ebe9c0724c310e6d0259da319ce67d5a.jpg", + "image_caption": [ + "Figure 13: planning module " + ], + "image_footnote": [], + "bbox": [ + 374, + 102, + 622, + 301 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "E EXPERIMENT DETAILS ", + "text_level": 1, + "bbox": [ + 173, + 330, + 398, + 348 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "E.1 BENCHMARK ENVIRONMENTS ", + "text_level": 1, + "bbox": [ + 176, + 364, + 426, + 378 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "We used four benchmark environments in our experiment. For the first three, the workspace dimension is 2d. We generated the maze maps with the recursive backtracker algorithm using the following implementation: https://github.com/lileee/gated-path-planning-networks/blob/master/generate_dataset.py. Examples of the workspace are shown in Figure 15. Three environments differ in the choice of robots: ", + "bbox": [ + 173, + 391, + 826, + 448 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "• Workspace planning (2d). The robot is abstracted with a point mass moving in the plane. Without higher dimensions, this problem reduces to planning in the workspace. \n• Rigid body navigation (3d). A rigid body robot, abstracted as a thin rectangle, is used here. The extra rotation dimension is added to the planning problem. This robot can rotate and move freely without any constraints in the free space. \n• 3-link snake (5d). The robot is a 5 DoF snake with two joints. Two more angle dimensions are added to the planning task. To prevent links from folding, we restrict the angles to the range of $[ - \\pi / 4 , \\pi / 4 ]$ . ", + "bbox": [ + 173, + 454, + 826, + 568 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "The fourth environment has a 3d workspace. Cuboid obstacles were generated uniformly randomly in space with density $\\approx 2 0 \\%$ . Example of the workspace is shown in Figure 6 and 16, where the blue cuboids are obstacles. The environment is described below: ", + "bbox": [ + 173, + 573, + 825, + 614 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "• Spacecraft planning (7d). The robot is a spacecraft with a cuboid body and two 2 DoF arms connecting to two opposite sides of the body. There is a joint in the middle of each arm. The outer arm can rotate around this joint. Each arm can also rotate as a whole around its connection point with the body. All rotation angles are restricted in the range of $[ 0 , \\pi / 2 ]$ . The spacecraft itself cannot rotate. ", + "bbox": [ + 174, + 621, + 825, + 691 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "E.2 HYPERPARAMETER FOR MSIL ", + "text_level": 1, + "bbox": [ + 176, + 710, + 428, + 726 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "During self-improving over the first 2000 problems, NEXT updated its parameters and annealed $\\epsilon$ once for every 200 problems. The value of the annealing $\\epsilon$ was set as the following: ", + "bbox": [ + 173, + 738, + 823, + 766 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/8d768de638d1253288962d6fc415fd345b9b974c6cbfafddd2aa591a51b77228.jpg", + "text": "$$\n\\epsilon = \\left\\{ \\begin{array} { l l } { 1 , } & { \\mathrm { i f ~ } i < 1 0 0 0 , } \\\\ { 0 . 5 - 0 . 1 \\cdot \\lfloor ( i - 1 0 0 0 ) / 2 0 0 \\rfloor , } & { \\mathrm { i f ~ } 1 0 0 0 \\leqslant i < 2 0 0 0 , } \\\\ { 0 . 1 , } & { \\mathrm { o t h e r w i s e , } } \\end{array} \\right.\n$$", + "text_format": "latex", + "bbox": [ + 302, + 768, + 692, + 819 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "with $i$ denoting the problem number. ", + "bbox": [ + 174, + 820, + 415, + 835 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "E.3 BASELINE: THE IMPROVED GPPN", + "text_level": 1, + "bbox": [ + 176, + 854, + 452, + 869 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "The original GPPN is not directly applicable to our experiments. Inspired by Tamar et al. (2016), we add a fully-connected MLP to its final layers, so that the improved architecture can be applied to high-dimensional continuous domain. As shown in Figure 14, the GPPN first processes the discretized ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/1197e1a0f579ba28f10b818c4b2e442327f2cb38b96543bbc6dbcfb1e5ec6325.jpg", + "image_caption": [ + "Figure 14: Improved GPPN architecture " + ], + "image_footnote": [], + "bbox": [ + 210, + 106, + 790, + 314 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "workspace locations. Its output and the full robot configurations are processed together by the MLP, which then produces the current value and action estimates. The improved GPPN is trained using supervisions from the near-optimal paths produced by RRT\\*. ", + "bbox": [ + 174, + 353, + 825, + 395 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "F EXPERIMENT RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 414, + 400, + 430 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/45835db4516f023ab21e7a5a2bb8b4907c34df7064909a1410c42cc4f9f92003.jpg", + "image_caption": [ + "Figure 15: The solution path produced by NEXT in a workspace planning task (2d), rigid body navigation task (3d), 3-link snake task (5d) from left to right. The orange dot and the brown dot are starting and goal locations, respectively. " + ], + "image_footnote": [], + "bbox": [ + 192, + 463, + 807, + 622 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/28558e709732ca4be6503115df6e181bb722fcd64cd5ffe0ac9a8fec10b06a19.jpg", + "image_caption": [ + "Figure 16: The solution paths produced by NEXT in spacecraft planning task (7d). Spacecraft has a yellow body and two 2 DoF (red) arms. Blue cuboids are obstacles. " + ], + "image_footnote": [], + "bbox": [ + 200, + 685, + 800, + 791 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "F.2 DETAILS OF QUANTITATIVE EVALUATION ", + "text_level": 1, + "bbox": [ + 173, + 856, + 503, + 871 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "More detailed results are shown in Table 1, 2, 3, including learning-based and non-learning-based ones, on the last 1000 problems in each experiment. We normalized the number of collision checks and the cost of paths based on the solution of RRT∗. The success rate result is not normalized. The best planners in each experiment are in bold. NEXT-KS and NEXT-GP outperform the current state-of-the-art planning algorithm with large margins. ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/8ab1e479cbe5f5dc1a3e294cb9eca675bb59491b24d146a5477f9e869df469a1.jpg", + "table_caption": [ + "Table 1: Success rate results. The higher the better. NEXT-KS performs the best. " + ], + "table_footnote": [], + "table_body": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.9880.9810.7180.6320.7350.7100.1850.5350.720
3d0.9430.8410.6890.5540.4900.5140.1210.1140.498
5d0.8830.7680.6330.5150.4550.4970.0300.4760.444
7d0.9310.9060.6340.3690.3610.8140.2880.2720.370
", + "bbox": [ + 179, + 170, + 815, + 237 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/73ec7d6b0b7a0c1eb7fab7d9ec2322bcece0b1e2ae2b0a48342a4450d9623596.jpg", + "table_caption": [ + "Table 2: Average number of collision checks results. The lower the better. The score is normalized based on the solution of $\\mathbf { R R T ^ { * } }$ . NEXT-KS performs the best in 3 benchmarks. " + ], + "table_footnote": [], + "table_body": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1770.2432.3423.4841.0005.9459.2471.9831.011
3d0.6941.3342.2143.1251.0007.9247.2922.1620.988
5d0.8881.5201.8002.7061.0007.4835.7581.1880.997
7d0.6530.5021.8771.3131.0004.6833.8561.5910.987
", + "bbox": [ + 179, + 289, + 815, + 356 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/87367e14b8aa88a514f45838acfb7a2ff1b4118f39278d228592551eef4e197f.jpg", + "table_caption": [ + "Table 3: Average cost of paths. The lower the better. The score is normalized based on the solution of $\\mathrm { R R T ^ { * } }$ . The NEXT-KS achieves the best solutions. " + ], + "table_footnote": [], + "table_body": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1720.1931.0491.3331.0001.1402.8111.6491.050
3d0.1160.3150.6120.8751.0000.9551.7201.7340.984
5d0.2150.4260.6730.8901.0000.9231.7800.9611.020
7d0.1080.1470.5730.9871.0000.2911.1141.1390.986
", + "bbox": [ + 179, + 405, + 818, + 473 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "We demonstrated the performance improvement curves for 2d workspace planning, 3d rigid body navigation in Figure 17. As we can see, similar to the performances on 5d 3-link snake planning task in Figure 7, in these tasks, the NEXT-KS and NEXT-GP improve the performances along with more and more experiences collected, justified the self-improvement ability by learning ${ \\tilde { V } } ^ { * }$ and $\\tilde { \\pi } ^ { * }$ . ", + "bbox": [ + 174, + 484, + 825, + 541 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "F.3 SEARCH TREES COMPARISON ", + "text_level": 1, + "bbox": [ + 176, + 558, + 419, + 573 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "We illustrate the search trees generated by $\\mathrm { R R T ^ { * } }$ and the proposed NEXT algorithms with 500 samples in Figure 18, Figure 19, Figure 20 and Figure 21 on several 2d, 3d, 5d and 7d planning tasks, respectively. To help readers better understand how the trees were expanded, we actually visualize the RRT\\* search trees without edge rewiring, which is equivalent to the RRT search trees, however the vertex set is the same. Comparing to the search trees generated by $\\mathbf { R R T ^ { * } }$ side by side, we can clearly see the advantages and the efficiency of the proposed NEXT algorithms. In all the tasks, even in 2d workspace planning task, the $\\mathrm { R R T ^ { * } }$ indeed randomly searches without realizing the goals, and thus cannot complete the missions, while the NEXT algorithms search towards the goals with the guidance from ${ \\tilde { V } } ^ { * }$ and $\\tilde { \\pi } ^ { * }$ , therefore, successfully provides high-quality solutions. ", + "bbox": [ + 174, + 584, + 825, + 713 + ], + "page_idx": 20 + }, + { + "type": "image", + "img_path": "images/4c7753cd5f57fdbee146ea4f7042a6906667f2b11da11f1627ca15e30997b485.jpg", + "image_caption": [ + "Figure 17: The first and second rows display the improvement curves of our algorithms on all 3000 problems of the 2d workspace planning and 3d rigid body navigation problems. We compare our algorithms with $\\mathrm { R R T ^ { * } }$ . Three columns correspond to the success rate, the average collision checks, and the average cost of the solution paths for each algorithm. " + ], + "image_footnote": [], + "bbox": [ + 178, + 135, + 800, + 377 + ], + "page_idx": 21 + }, + { + "type": "image", + "img_path": "images/e12728f87cbc3a880f7afe9bb7ba0f37c6e5a599ae9915a27fe8f8e4dd07bd7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 517, + 831, + 756 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "(a) $\\mathrm { R R T ^ { * } }$ (w/o rewiring) (b) NEXT-KS search tree (c) NEXT-GP search tree (d) learned ${ \\tilde { V } } ^ { * }$ and $\\tilde { \\pi } ^ { * }$ ", + "bbox": [ + 181, + 763, + 834, + 780 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Figure 18: Column (a) to (c) are the search trees produced by the $\\mathrm { R R T ^ { * } }$ , NEXT-KS, and NEXT-GP on the same workspace planning task (2d). The learned $\\tilde { V } ^ { \\ast }$ and $\\tilde { \\pi } ^ { * }$ from NEXT-KS are plotted in column (d). In the figures, obstacles are colored in deep blue, the starting and goal locations are denoted by orange and brown dots, respectively. In column (a) to (c), samples are represented with hollow yellow circles, and edges are colored in green. In column (d), the level of redness denotes the value of the cost-to-go estimate $\\tilde { V } ^ { \\ast }$ , and the cyan arrows point from a given state $s$ to the center of the proposal distribution $\\tilde { \\pi } ^ { * } \\bigl ( s ^ { \\prime } | s , U \\bigr )$ . We set the maximum number of samples to be 500. ", + "bbox": [ + 173, + 789, + 826, + 892 + ], + "page_idx": 21 + }, + { + "type": "image", + "img_path": "images/e0735fc244ea4c94bcec57f2daf0ecaa89ac9d792df1ae99e175f050da010796.jpg", + "image_caption": [ + "Figure 19: Each column corresponds to one example from the rigid body navigation problem (3d). The top and the bottom rows are the search trees produced by the RRT\\* and NEXT-KS, respectively. In the figures, obstacles are colored in deep blue, and the rigid bodies are represented with matchsticks. The samples, starting states, and goal states are denoted by yellow, orange, and brown matchsticks, respectively. Edges are colored in green. We set the maximum number of samples to be 500. " + ], + "image_footnote": [], + "bbox": [ + 189, + 128, + 815, + 401 + ], + "page_idx": 22 + }, + { + "type": "image", + "img_path": "images/cfa7ea4cfcbf4aafa0403606ed187484f2a7a371df073fcdcf9d9b4c8e59fe6c.jpg", + "image_caption": [ + "Figure 20: Each column corresponds to one example from the 3-link snake problem (5d). The top and the bottom rows are the search trees produced by the RRT\\* and NEXT-KS, respectively. In the figures, obstacles are colored in deep blue, and the rigid bodies are represented with matchsticks. The samples, starting states, and goal states are denoted by yellow, orange, and brown matchsticks, respectively. Edges are colored in green. We set the maximum number of samples to be 500. " + ], + "image_footnote": [], + "bbox": [ + 187, + 542, + 813, + 815 + ], + "page_idx": 22 + }, + { + "type": "image", + "img_path": "images/acacf1ccb84450bee6725175dc3f6794eeb8b2a6eca9fd60b1ef87035d99e36d.jpg", + "image_caption": [ + "Figure 21: Each column corresponds to one example from the spacecraft planning problem (7d). The top and the bottom rows are the search trees produced by the $\\mathrm { R R T ^ { * } }$ and NEXT-KS, respectively. In the figures, obstacles are colored in blue, and each spacecraft has a yellow body and two 2 DoF red arms. We set the maximum number of samples to be 500. " + ], + "image_footnote": [], + "bbox": [ + 202, + 347, + 802, + 612 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "F.4 CASE STUDY DETAILS ", + "text_level": 1, + "bbox": [ + 174, + 103, + 370, + 117 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "We conduct a real-world case study on controlling robot arms to move objects on a shelf. This is a representative of common scenarios in practice where the robot needs to plan its motion in real-time to reach the inside of some narrow space. For this case study, we focus more on the practical aspect to evaluate how much can we improve on the wall-clock time by learning from similar planning problems. ", + "bbox": [ + 174, + 128, + 825, + 199 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "F.4.1 TASK DESCRIPTION ", + "text_level": 1, + "bbox": [ + 174, + 214, + 366, + 229 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "We generate planning problems randomly to form the training set and test set. In each planning task, there is a shelf of multiple levels, with each level horizontally divided into multiple bins. The heights of levels and widths of bins are randomly drawn from some fixed distribution. Samples of shelves are shown in Figure 22. Both the start and goal configurations are randomly sampled from a distribution within the reachable region of the robot arm. The planning environment is created with the OpenRave simulator (Diankov, 2010). ", + "bbox": [ + 174, + 239, + 825, + 323 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "The task is to find a path for the 7 DoF robot arm to move from a location in one bin to another, i.e., the end effectors of the start and goal configurations are in different bins, as illustrated in Figure 23. In this case, the base of the robot is fixed and we are planning the movement of arm. We generated 3000 problems for training and 1000 problems for testing. ", + "bbox": [ + 174, + 329, + 825, + 386 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "F.4.2 BASELINES AND TRAINING", + "text_level": 1, + "bbox": [ + 176, + 401, + 416, + 415 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "For traditional planners, we include $\\mathrm { C } + +$ OMPL ( ¸Sucan et al., 2012) implementation of $\\mathrm { B I T ^ { * } }$ (Gammell et al., 2015) and $\\mathrm { R R T ^ { * } }$ (Karaman & Frazzoli, 2011) as baselines. The hyperparameters of RRT\\* and $\\mathbf { B I T ^ { * } }$ are specially tuned for this experiment. We also compare with learning-based planners CVAE-plan (Ichter et al., 2018) and Reinforce-plan (Zhang et al., 2018). The supervisions for CVAE-plan are produced by the well-tuned $\\mathrm { B I T ^ { * } }$ on the training set. To train NEXT, we consider the $\\mathrm { B I T ^ { * } }$ instead of RRT\\* as the imperfect expert in training problems. ", + "bbox": [ + 174, + 424, + 825, + 510 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "F.4.3 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 523, + 292, + 537 + ], + "page_idx": 24 + }, + { + "type": "text", + "text": "We evaluate the algorithms on the separated testing problems, and record the success rate using 10 different time limits. The success rate and average path quality are plot in Figure 8 and recorded in Table 4 and Table 5. The solution paths found by NEXT and $\\mathrm { B I T ^ { * } }$ are illustrated in Figure 23. In terms of both success rate and solution path quality, NEXT dominates all the planners under all time limits. ", + "bbox": [ + 174, + 549, + 825, + 617 + ], + "page_idx": 24 + }, + { + "type": "image", + "img_path": "images/1a68418f9fc5b8db4c0728ef4e4b915a1ac7b12ca2d6815c4396b38f554580ed.jpg", + "image_caption": [ + "Figure 22: Examples of different shelves sampled from the distribution. " + ], + "image_footnote": [], + "bbox": [ + 196, + 623, + 805, + 794 + ], + "page_idx": 24 + }, + { + "type": "image", + "img_path": "images/4918f228ef3bb16fc560cb6a309ed28b7f1abfcfd2648d63098ec7552a0228e2.jpg", + "image_caption": [ + "Figure 23: First row: robot arm solution trajectories produced by NEXT(-KS) in four planning problems; Second row: $\\mathrm { B I T ^ { * } }$ solutions on the same planning problems. NEXT only takes 5 seconds to complete each problem while $\\mathrm { B I T ^ { * } }$ needs 250 seconds to find a solution for the hardest problems (last two columns). " + ], + "image_footnote": [], + "bbox": [ + 191, + 130, + 794, + 462 + ], + "page_idx": 25 + }, + { + "type": "table", + "img_path": "images/948be226a6bad3dfcd234d9761507c5f0ee44a0c74d5bfc76f2f1e7f3c70aac5.jpg", + "table_caption": [ + "Table 4: Success rates of different planners under varying time limits, the higher the better. " + ], + "table_footnote": [], + "table_body": "
5s10s15s20s25s30s35s40s45s50s
NEXT0.5790.6570.7030.7090.7450.7430.7460.7520.7720.763
CVAE0.3540.4370.4820.5090.5070.5390.5530.5510.5790.580
Reinforce0.1600.1700.2000.1500.1800.1900.1750.1800.2250.175
BIT*0.2260.2880.3200.3650.3640.4290.4250.4220.4430.475
RRT*0.1350.1480.1440.1360.1470.1590.1520.1580.1570.165
", + "bbox": [ + 196, + 631, + 802, + 712 + ], + "page_idx": 25 + }, + { + "type": "table", + "img_path": "images/17b397e3af9f47e2586529a736802b6c4ddb66c41d7cfd6009886c510b1b4236.jpg", + "table_caption": [ + "Table 5: Average path costs of different planners under varying time limits, the lower the better. " + ], + "table_footnote": [], + "table_body": "
5s10s15s20s25s30s35s40s45s50s
NEXT25.14321.87320.05219.77318.31818.38118.27418.04717.22417.587
CVAE34.41430.81528.90927.71727.78426.40825.77625.88324.66524.659
Reinforce42.78142.47141.14243.29542.12941.65742.23542.09340.26942.282
BIT*39.54337.08635.74434.00234.00731.41431.58731.65130.84129.601
RRT*43.36842.85542.98143.35542.91942.41042.68042.44142.50442.165
", + "bbox": [ + 174, + 806, + 852, + 887 + ], + "page_idx": 25 + } +] \ No newline at end of file diff --git a/parse/train/rJgJDAVKvB/rJgJDAVKvB_middle.json b/parse/train/rJgJDAVKvB/rJgJDAVKvB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..7974fa02c4f63c7e1614d3fc20203120297de54a --- /dev/null +++ b/parse/train/rJgJDAVKvB/rJgJDAVKvB_middle.json @@ -0,0 +1,69333 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 452, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 455, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 455, + 96 + ], + "score": 1.0, + "content": "LEARNING TO PLAN IN HIGH DIMENSIONS VIA", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 98, + 452, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 452, + 117 + ], + "score": 1.0, + "content": "NEURAL EXPLORATION-EXPLOITATION TREES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 133, + 424, + 169 + ], + "lines": [ + { + "bbox": [ + 113, + 132, + 426, + 148 + ], + "spans": [ + { + "bbox": [ + 113, + 132, + 195, + 148 + ], + "score": 1.0, + "content": "∗Binghong Chen1,", + "type": "text" + }, + { + "bbox": [ + 195, + 134, + 234, + 146 + ], + "score": 0.47, + "content": "\\mathbf { \\mathbf { \\mathbf { \\mathbf { \\mathbf { \\mathbf { B } } } } } } \\mathbf { \\mathbf { 0 } } \\mathbf { \\mathbf { \\mathbf { \\mathbf { D } } } } \\mathbf { \\mathbf { a } } \\mathbf { i } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 132, + 267, + 148 + ], + "score": 1.0, + "content": ", Qinjie", + "type": "text" + }, + { + "bbox": [ + 267, + 134, + 288, + 146 + ], + "score": 0.77, + "content": "\\mathbf { L i n ^ { 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 132, + 312, + 148 + ], + "score": 1.0, + "content": ", Guo", + "type": "text" + }, + { + "bbox": [ + 312, + 134, + 328, + 146 + ], + "score": 0.76, + "content": "\\mathbf { Y e ^ { 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 132, + 426, + 148 + ], + "score": 1.0, + "content": ", Han Liu3, Le Song1,4", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 144, + 383, + 160 + ], + "spans": [ + { + "bbox": [ + 111, + 144, + 383, + 160 + ], + "score": 1.0, + "content": "1Georgia Institute of Technology 2Google Research, Brain Team", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 156, + 290, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 156, + 290, + 171 + ], + "score": 1.0, + "content": "3Northwestern University 4Ant Financial", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 278, + 190, + 333, + 201 + ], + "lines": [ + { + "bbox": [ + 276, + 189, + 335, + 203 + ], + "spans": [ + { + "bbox": [ + 276, + 189, + 335, + 203 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 130, + 211, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 128, + 211, + 483, + 225 + ], + "spans": [ + { + "bbox": [ + 128, + 211, + 483, + 225 + ], + "score": 1.0, + "content": "We propose a meta path planning algorithm named Neural Exploration-Exploitation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 128, + 222, + 483, + 235 + ], + "spans": [ + { + "bbox": [ + 128, + 222, + 483, + 235 + ], + "score": 1.0, + "content": "Trees (NEXT) for learning from prior experience for solving new path planning problems", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 128, + 233, + 483, + 246 + ], + "spans": [ + { + "bbox": [ + 128, + 233, + 483, + 246 + ], + "score": 1.0, + "content": "in high dimensional continuous state and action spaces. Compared to more classical", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 128, + 245, + 483, + 257 + ], + "spans": [ + { + "bbox": [ + 128, + 245, + 483, + 257 + ], + "score": 1.0, + "content": "sampling-based methods like RRT, our approach achieves much better sample efficiency", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 256, + 483, + 267 + ], + "spans": [ + { + "bbox": [ + 128, + 256, + 483, + 267 + ], + "score": 1.0, + "content": "in high-dimensions and can benefit from prior experience of planning in similar envi-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 128, + 267, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 128, + 267, + 482, + 278 + ], + "score": 1.0, + "content": "ronments. More specifically, NEXT exploits a novel neural architecture which can learn", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 128, + 277, + 483, + 289 + ], + "spans": [ + { + "bbox": [ + 128, + 277, + 483, + 289 + ], + "score": 1.0, + "content": "promising search directions from problem structures. The learned prior is then integrated", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 128, + 288, + 484, + 301 + ], + "spans": [ + { + "bbox": [ + 128, + 288, + 484, + 301 + ], + "score": 1.0, + "content": "into a UCB-type algorithm to achieve an online balance between exploration and exploita-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 128, + 298, + 483, + 311 + ], + "spans": [ + { + "bbox": [ + 128, + 298, + 483, + 311 + ], + "score": 1.0, + "content": "tion when solving a new problem. We conduct thorough experiments to show that NEXT", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 128, + 310, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 128, + 310, + 482, + 323 + ], + "score": 1.0, + "content": "accomplishes new planning problems with more compact search trees and significantly", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 128, + 321, + 374, + 333 + ], + "spans": [ + { + "bbox": [ + 128, + 321, + 374, + 333 + ], + "score": 1.0, + "content": "outperforms state-of-the-art methods on several benchmarks.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 352, + 206, + 364 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 208, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 208, + 367 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 373, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "Path planning is a fundamental problem with many real-world applications, such as robot manipulation", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "and autonomous driving. A simple planning problem within low-dimensional state space can be", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "solved by first discretizing the continuous state space into a grid, and then searching for a path", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 319, + 418 + ], + "score": 1.0, + "content": "on top of it using graph search algorithms such as", + "type": "text" + }, + { + "bbox": [ + 320, + 406, + 333, + 416 + ], + "score": 0.86, + "content": "A ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 407, + 505, + 418 + ], + "score": 1.0, + "content": "(Hart et al., 1968). However, due to the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 417, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 429 + ], + "score": 1.0, + "content": "curse of dimensionality, these approaches do not scale well with the number of dimensions of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "the state space. For high-dimensional planning problems, people often resort to sampling-based", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "approaches to avoid explicit discretization. Sampling-based planning algorithms, such as probabilistic", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "roadmaps (PRM) (Kavraki et al., 1996), rapidly-exploring random trees (RRT) (LaValle, 1998), and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "their variants (Karaman & Frazzoli, 2011) incrementally build an implicit representation of the state", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 472, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 485 + ], + "score": 1.0, + "content": "space using probing samples. These generic algorithms typically employ a uniform sampler which", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "does not make use of the structures of the problem. Therefore they may require lots of samples to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "obtain a feasible solution for complicated problems. To improve the sample efficiency, heuristic", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "biased samplers, such as Gaussian sampler (Boor et al., 1999), bridge test (Hsu et al., 2003) and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "score": 1.0, + "content": "reachability-guided sampler (Shkolnik et al., 2009) have been proposed. All these sampling heuristics", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "are designed manually to address specific structural properties, which may or may not be valid for a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 538, + 469, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 469, + 550 + ], + "score": 1.0, + "content": "new problem, and may lead to even worse performance compared to the uniform proposal.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 708 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "Online adaptation in path planning has also been investigated for improving sample efficiency in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 566, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 578 + ], + "score": 1.0, + "content": "current planning problem. Specifically, Hsu et al. (2005) exploits online algorithms to dynamically", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "score": 1.0, + "content": "adapts the mixture weights of several manually designed biased samplers. Burns & Brock (2005a;b) fit", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "a model for the planning environment incrementally and use the model for planning. Yee et al. (2016)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 598, + 507, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 507, + 612 + ], + "score": 1.0, + "content": "mimics the Monte-Carlos tree search (MCTS) for problems with continuous state and action spaces.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "These algorithms treat each planning problem independently, and the collected data from previous", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 633 + ], + "score": 1.0, + "content": "experiences and built model will be simply discarded when solving a new problem. However, in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "practice, similar planning problems may be solved again and again, where the problems are different", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 643, + 504, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 504, + 654 + ], + "score": 1.0, + "content": "but sharing common structures. For instance, grabbing a coffee cup on a table at different time", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "score": 1.0, + "content": "are different problems, since the layout of paper and pens, the position and orientation of coffee", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "cups may be different every time; however, all these problems show common structures of handling", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 675, + 506, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 506, + 688 + ], + "score": 1.0, + "content": "similar objects which are placed in similar fashions. Intuitively, if the common characteristics across", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 686, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 699 + ], + "score": 1.0, + "content": "problems can be learned via some shared latent representation, a planner based on such representation", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 696, + 403, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 403, + 711 + ], + "score": 1.0, + "content": "can then be transferred to new problems with improved sample efficiency.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 40.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 225, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 227, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 227, + 733 + ], + "score": 1.0, + "content": "∗indicates equal contribution.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 452, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 455, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 455, + 96 + ], + "score": 1.0, + "content": "LEARNING TO PLAN IN HIGH DIMENSIONS VIA", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 98, + 452, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 452, + 117 + ], + "score": 1.0, + "content": "NEURAL EXPLORATION-EXPLOITATION TREES", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 133, + 424, + 169 + ], + "lines": [ + { + "bbox": [ + 113, + 132, + 426, + 148 + ], + "spans": [ + { + "bbox": [ + 113, + 132, + 195, + 148 + ], + "score": 1.0, + "content": "∗Binghong Chen1,", + "type": "text" + }, + { + "bbox": [ + 195, + 134, + 234, + 146 + ], + "score": 0.47, + "content": "\\mathbf { \\mathbf { \\mathbf { \\mathbf { \\mathbf { \\mathbf { B } } } } } } \\mathbf { \\mathbf { 0 } } \\mathbf { \\mathbf { \\mathbf { \\mathbf { D } } } } \\mathbf { \\mathbf { a } } \\mathbf { i } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 132, + 267, + 148 + ], + "score": 1.0, + "content": ", Qinjie", + "type": "text" + }, + { + "bbox": [ + 267, + 134, + 288, + 146 + ], + "score": 0.77, + "content": "\\mathbf { L i n ^ { 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 132, + 312, + 148 + ], + "score": 1.0, + "content": ", Guo", + "type": "text" + }, + { + "bbox": [ + 312, + 134, + 328, + 146 + ], + "score": 0.76, + "content": "\\mathbf { Y e ^ { 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 132, + 426, + 148 + ], + "score": 1.0, + "content": ", Han Liu3, Le Song1,4", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 144, + 383, + 160 + ], + "spans": [ + { + "bbox": [ + 111, + 144, + 383, + 160 + ], + "score": 1.0, + "content": "1Georgia Institute of Technology 2Google Research, Brain Team", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 156, + 290, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 156, + 290, + 171 + ], + "score": 1.0, + "content": "3Northwestern University 4Ant Financial", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 111, + 132, + 426, + 171 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 190, + 333, + 201 + ], + "lines": [ + { + "bbox": [ + 276, + 189, + 335, + 203 + ], + "spans": [ + { + "bbox": [ + 276, + 189, + 335, + 203 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 130, + 211, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 128, + 211, + 483, + 225 + ], + "spans": [ + { + "bbox": [ + 128, + 211, + 483, + 225 + ], + "score": 1.0, + "content": "We propose a meta path planning algorithm named Neural Exploration-Exploitation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 128, + 222, + 483, + 235 + ], + "spans": [ + { + "bbox": [ + 128, + 222, + 483, + 235 + ], + "score": 1.0, + "content": "Trees (NEXT) for learning from prior experience for solving new path planning problems", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 128, + 233, + 483, + 246 + ], + "spans": [ + { + "bbox": [ + 128, + 233, + 483, + 246 + ], + "score": 1.0, + "content": "in high dimensional continuous state and action spaces. Compared to more classical", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 128, + 245, + 483, + 257 + ], + "spans": [ + { + "bbox": [ + 128, + 245, + 483, + 257 + ], + "score": 1.0, + "content": "sampling-based methods like RRT, our approach achieves much better sample efficiency", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 256, + 483, + 267 + ], + "spans": [ + { + "bbox": [ + 128, + 256, + 483, + 267 + ], + "score": 1.0, + "content": "in high-dimensions and can benefit from prior experience of planning in similar envi-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 128, + 267, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 128, + 267, + 482, + 278 + ], + "score": 1.0, + "content": "ronments. More specifically, NEXT exploits a novel neural architecture which can learn", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 128, + 277, + 483, + 289 + ], + "spans": [ + { + "bbox": [ + 128, + 277, + 483, + 289 + ], + "score": 1.0, + "content": "promising search directions from problem structures. The learned prior is then integrated", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 128, + 288, + 484, + 301 + ], + "spans": [ + { + "bbox": [ + 128, + 288, + 484, + 301 + ], + "score": 1.0, + "content": "into a UCB-type algorithm to achieve an online balance between exploration and exploita-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 128, + 298, + 483, + 311 + ], + "spans": [ + { + "bbox": [ + 128, + 298, + 483, + 311 + ], + "score": 1.0, + "content": "tion when solving a new problem. We conduct thorough experiments to show that NEXT", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 128, + 310, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 128, + 310, + 482, + 323 + ], + "score": 1.0, + "content": "accomplishes new planning problems with more compact search trees and significantly", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 128, + 321, + 374, + 333 + ], + "spans": [ + { + "bbox": [ + 128, + 321, + 374, + 333 + ], + "score": 1.0, + "content": "outperforms state-of-the-art methods on several benchmarks.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11, + "bbox_fs": [ + 128, + 211, + 484, + 333 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 352, + 206, + 364 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 208, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 208, + 367 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 373, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "Path planning is a fundamental problem with many real-world applications, such as robot manipulation", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "and autonomous driving. A simple planning problem within low-dimensional state space can be", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "solved by first discretizing the continuous state space into a grid, and then searching for a path", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 319, + 418 + ], + "score": 1.0, + "content": "on top of it using graph search algorithms such as", + "type": "text" + }, + { + "bbox": [ + 320, + 406, + 333, + 416 + ], + "score": 0.86, + "content": "A ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 407, + 505, + 418 + ], + "score": 1.0, + "content": "(Hart et al., 1968). However, due to the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 417, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 429 + ], + "score": 1.0, + "content": "curse of dimensionality, these approaches do not scale well with the number of dimensions of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "the state space. For high-dimensional planning problems, people often resort to sampling-based", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "approaches to avoid explicit discretization. Sampling-based planning algorithms, such as probabilistic", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "roadmaps (PRM) (Kavraki et al., 1996), rapidly-exploring random trees (RRT) (LaValle, 1998), and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 473 + ], + "score": 1.0, + "content": "their variants (Karaman & Frazzoli, 2011) incrementally build an implicit representation of the state", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 472, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 485 + ], + "score": 1.0, + "content": "space using probing samples. These generic algorithms typically employ a uniform sampler which", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "does not make use of the structures of the problem. Therefore they may require lots of samples to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "obtain a feasible solution for complicated problems. To improve the sample efficiency, heuristic", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "biased samplers, such as Gaussian sampler (Boor et al., 1999), bridge test (Hsu et al., 2003) and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "score": 1.0, + "content": "reachability-guided sampler (Shkolnik et al., 2009) have been proposed. All these sampling heuristics", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "are designed manually to address specific structural properties, which may or may not be valid for a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 538, + 469, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 469, + 550 + ], + "score": 1.0, + "content": "new problem, and may lead to even worse performance compared to the uniform proposal.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 374, + 506, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 708 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "Online adaptation in path planning has also been investigated for improving sample efficiency in", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 566, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 578 + ], + "score": 1.0, + "content": "current planning problem. Specifically, Hsu et al. (2005) exploits online algorithms to dynamically", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "score": 1.0, + "content": "adapts the mixture weights of several manually designed biased samplers. Burns & Brock (2005a;b) fit", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "a model for the planning environment incrementally and use the model for planning. Yee et al. (2016)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 598, + 507, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 507, + 612 + ], + "score": 1.0, + "content": "mimics the Monte-Carlos tree search (MCTS) for problems with continuous state and action spaces.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "These algorithms treat each planning problem independently, and the collected data from previous", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 633 + ], + "score": 1.0, + "content": "experiences and built model will be simply discarded when solving a new problem. However, in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "practice, similar planning problems may be solved again and again, where the problems are different", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 643, + 504, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 504, + 654 + ], + "score": 1.0, + "content": "but sharing common structures. For instance, grabbing a coffee cup on a table at different time", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "score": 1.0, + "content": "are different problems, since the layout of paper and pens, the position and orientation of coffee", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 677 + ], + "score": 1.0, + "content": "cups may be different every time; however, all these problems show common structures of handling", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 675, + 506, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 506, + 688 + ], + "score": 1.0, + "content": "similar objects which are placed in similar fashions. Intuitively, if the common characteristics across", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 686, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 699 + ], + "score": 1.0, + "content": "problems can be learned via some shared latent representation, a planner based on such representation", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 696, + 403, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 403, + 711 + ], + "score": 1.0, + "content": "can then be transferred to new problems with improved sample efficiency.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 555, + 507, + 711 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 83, + 505, + 247 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Several methods have been proposed recently to learn from past planning experiences to conduct", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "more efficient and generalizable planning for future problems. These works are limited in one way", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "or the other. Zucker et al. (2008); Zhang et al. (2018); Huh & Lee (2018) treat the sampler in the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "sampling-based planner as a stochastic policy to be learned and apply policy gradient or TD-algorithm", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "to improve the policy. Finney et al. (2007); Bowen & Alterovitz (2014); Ye & Alterovitz (2017);", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "Ichter et al. (2018); Kim et al. (2018); Kuo et al. (2018) apply imitation learning based on the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "collected demonstrations to bias for better sampler via variants of probabilistic models, e.g., (mixture", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "of) Gaussians, conditional VAE, GAN, HMM and RNN. However, many of these approaches either", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 104, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "rely on specially designed local features or assume the problems are indexed by special parameters,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "score": 1.0, + "content": "which limits the generalization ability. Deep representation learning provides a promising direction to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 204 + ], + "score": 1.0, + "content": "extract the common structure among the planning problems, and thus mitigate such limitation on hand-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "designed features. However, existing work, e.g., motion planning networks (Qureshi et al., 2019),", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "value iteration networks (VIN) (Tamar et al., 2016), and gated path planning networks (GPPN) (Lee", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "score": 1.0, + "content": "et al., 2018), either apply off-the-shelf MLP architecture ignoring special structures in planning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 236, + 474, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 474, + 249 + ], + "score": 1.0, + "content": "problems or can only deal with discrete state and action spaces in low-dimensional settings.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "In this paper, we present Neural EXploration-EXploitation Tree (NEXT), a meta neural path plan-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "ning algorithm for high-dimensional continuous state space problems. The core contribution is a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 273, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 506, + 288 + ], + "score": 1.0, + "content": "novel attention-based neural architecture that is capable of learning generalizable problem struc-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "tures from previous experiences and produce promising search directions with automatic online", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 472, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 472, + 310 + ], + "score": 1.0, + "content": "exploration-exploitation balance adaption. Compared to existing learning-based planners,", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 311, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 324 + ], + "score": 1.0, + "content": "• NEXT is more generic. We propose an architecture that can embed high dimensional continuous", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "state spaces into low dimensional discrete spaces, on which a neural planning module is used to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 333, + 413, + 345 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 413, + 345 + ], + "score": 1.0, + "content": "extract planning representation. These module will be learned end-to-end.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 108, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 108, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "• NEXT balances exploration-exploitation trade-off. We integrate the learned neural prior into", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 115, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "an upper confidence bound (UCB) style algorithm to achieve an online balance between exploration", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 366, + 303, + 378 + ], + "spans": [ + { + "bbox": [ + 116, + 366, + 303, + 378 + ], + "score": 1.0, + "content": "and exploitation when solving a new problem.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 108, + 380, + 503, + 413 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 393 + ], + "score": 1.0, + "content": "Empirically, we show that NEXT can exploit past experiences to reduce the number of required", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "samples drastically for solving new planning problems, and significantly outperforms previous", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 401, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 287, + 414 + ], + "score": 1.0, + "content": "state-of-the-arts on several benchmark tasks.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 420, + 208, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 209, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 209, + 432 + ], + "score": 1.0, + "content": "1.1 RELATED WORKS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 506, + 446 + ], + "score": 1.0, + "content": "Designing non-uniform sampling strategies for random search to improve the planning efficiency", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 443, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 455 + ], + "score": 1.0, + "content": "has been considered as we discussed above. Besides the mentioned algorithms, there are other", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "score": 1.0, + "content": "works along this line, including informed RRT∗ (Gammell et al., 2014) and batch informed Trees", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 107, + 465, + 135, + 476 + ], + "score": 0.57, + "content": "\\mathrm { ( B I T ^ { * } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 466, + 402, + 477 + ], + "score": 1.0, + "content": "(Gammell et al., 2015) as the representative work. Randomized", + "type": "text" + }, + { + "bbox": [ + 403, + 465, + 416, + 475 + ], + "score": 0.87, + "content": "A ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 466, + 506, + 477 + ], + "score": 1.0, + "content": "(Diankov & Kuffner,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 216, + 489 + ], + "score": 1.0, + "content": "2007) and sampling-based", + "type": "text" + }, + { + "bbox": [ + 216, + 477, + 229, + 487 + ], + "score": 0.87, + "content": "A ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "(Persson & Sharf, 2014) expand the search tree with hand-designed", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "heuristics. These methods incorporate the human prior knowledge via hard-coded rules, which is", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "score": 1.0, + "content": "fixed and unable to adapt to problems, and thus, may not universally applicable. Choudhury et al.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 509, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 506, + 522 + ], + "score": 1.0, + "content": "(2018); Song et al. (2018) attempt to learn search heuristics. However, both methods are restricted to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 520, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 532 + ], + "score": 1.0, + "content": "planning on discrete domains. Meanwhile, the latter one always employs an unnecessary hierarchical", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 531, + 475, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 475, + 544 + ], + "score": 1.0, + "content": "structure for path planning, which leads to inferior sample efficiency and extra computation.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 548, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "The online exploration-exploitation trade-off is also an important issue in planning. For instance,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "Rickert et al. (2009) constructs a potential field sampler and tuned the sampler variance based on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "collision rate for the trade-off heuristically. Paxton et al. (2017) separates the action space into", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "high-level discrete options and low-level continuous actions, and only considered the trade-off at the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "score": 1.0, + "content": "discrete option level, ignoring the exploration-exploitation in the fine action space. These existing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 603, + 444, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 444, + 615 + ], + "score": 1.0, + "content": "works address the trade-off in an ad-hoc way, which may be inferior for the balance.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 619, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "There have been sevearl non-learning-based planning methods that can also leverage experi-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "ences (Kavraki et al., 1996; Phillips et al., 2012) by utilizing search graphs created in previous", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "problems. However, they are designed for largely fixed obstacles and cannot be generalized to unseen", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 652, + 316, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 316, + 666 + ], + "score": 1.0, + "content": "tasks from the same planning problems distribution.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5 + }, + { + "type": "title", + "bbox": [ + 107, + 670, + 307, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 309, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 309, + 684 + ], + "score": 1.0, + "content": "2 SETTINGS FOR LEARNING TO PLAN", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 123, + 700 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 687, + 158, + 699 + ], + "score": 0.91, + "content": "S \\subseteq \\mathbb { R } ^ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "be the state space of the problem, e.g., all the configurations of a robot and its base", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 698, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 219, + 712 + ], + "score": 1.0, + "content": "location in the workspace,", + "type": "text" + }, + { + "bbox": [ + 219, + 699, + 261, + 711 + ], + "score": 0.92, + "content": "S _ { o b s } \\subsetneq S", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 698, + 350, + 712 + ], + "score": 1.0, + "content": "be the obstacles set,", + "type": "text" + }, + { + "bbox": [ + 351, + 699, + 429, + 711 + ], + "score": 0.91, + "content": "S _ { f r e e } : = \\mathcal { S } \\setminus \\mathcal { S } _ { o b s }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 698, + 507, + 712 + ], + "score": 1.0, + "content": "be the free space,", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 107, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 161, + 722 + ], + "score": 0.92, + "content": "s _ { i n i t } \\in S _ { f r e e }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 709, + 249, + 722 + ], + "score": 1.0, + "content": "be the initial state and", + "type": "text" + }, + { + "bbox": [ + 250, + 710, + 308, + 722 + ], + "score": 0.92, + "content": "\\mathcal { S } _ { g o a l } \\subsetneq S _ { f r e e }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "be the goal region. Then the space of all collision-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 504, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 309, + 734 + ], + "score": 1.0, + "content": "free paths can be defined as a continuous function", + "type": "text" + }, + { + "bbox": [ + 310, + 720, + 428, + 733 + ], + "score": 0.91, + "content": "\\Xi : = \\{ \\bar { \\xi } ( \\cdot ) : \\bar { [ 0 , 1 ] } S _ { f r e e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 720, + 448, + 734 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 448, + 720, + 504, + 733 + ], + "score": 0.89, + "content": "c \\left( \\cdot \\right) : \\Xi \\mapsto \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 54 + } + ], + "index": 52.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 83, + 505, + 247 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Several methods have been proposed recently to learn from past planning experiences to conduct", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "more efficient and generalizable planning for future problems. These works are limited in one way", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "or the other. Zucker et al. (2008); Zhang et al. (2018); Huh & Lee (2018) treat the sampler in the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "sampling-based planner as a stochastic policy to be learned and apply policy gradient or TD-algorithm", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "to improve the policy. Finney et al. (2007); Bowen & Alterovitz (2014); Ye & Alterovitz (2017);", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "Ichter et al. (2018); Kim et al. (2018); Kuo et al. (2018) apply imitation learning based on the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "collected demonstrations to bias for better sampler via variants of probabilistic models, e.g., (mixture", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "of) Gaussians, conditional VAE, GAN, HMM and RNN. However, many of these approaches either", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 104, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "rely on specially designed local features or assume the problems are indexed by special parameters,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "score": 1.0, + "content": "which limits the generalization ability. Deep representation learning provides a promising direction to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 204 + ], + "score": 1.0, + "content": "extract the common structure among the planning problems, and thus mitigate such limitation on hand-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "designed features. However, existing work, e.g., motion planning networks (Qureshi et al., 2019),", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "value iteration networks (VIN) (Tamar et al., 2016), and gated path planning networks (GPPN) (Lee", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "score": 1.0, + "content": "et al., 2018), either apply off-the-shelf MLP architecture ignoring special structures in planning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 236, + 474, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 474, + 249 + ], + "score": 1.0, + "content": "problems or can only deal with discrete state and action spaces in low-dimensional settings.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7, + "bbox_fs": [ + 104, + 83, + 506, + 249 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "In this paper, we present Neural EXploration-EXploitation Tree (NEXT), a meta neural path plan-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "ning algorithm for high-dimensional continuous state space problems. The core contribution is a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 273, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 506, + 288 + ], + "score": 1.0, + "content": "novel attention-based neural architecture that is capable of learning generalizable problem struc-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 298 + ], + "score": 1.0, + "content": "tures from previous experiences and produce promising search directions with automatic online", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 472, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 472, + 310 + ], + "score": 1.0, + "content": "exploration-exploitation balance adaption. Compared to existing learning-based planners,", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 252, + 506, + 310 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 311, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 324 + ], + "score": 1.0, + "content": "• NEXT is more generic. We propose an architecture that can embed high dimensional continuous", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "state spaces into low dimensional discrete spaces, on which a neural planning module is used to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 333, + 413, + 345 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 413, + 345 + ], + "score": 1.0, + "content": "extract planning representation. These module will be learned end-to-end.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 108, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 108, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "• NEXT balances exploration-exploitation trade-off. We integrate the learned neural prior into", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 115, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "an upper confidence bound (UCB) style algorithm to achieve an online balance between exploration", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 366, + 303, + 378 + ], + "spans": [ + { + "bbox": [ + 116, + 366, + 303, + 378 + ], + "score": 1.0, + "content": "and exploitation when solving a new problem.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 309, + 506, + 378 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 380, + 503, + 413 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 393 + ], + "score": 1.0, + "content": "Empirically, we show that NEXT can exploit past experiences to reduce the number of required", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "samples drastically for solving new planning problems, and significantly outperforms previous", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 401, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 287, + 414 + ], + "score": 1.0, + "content": "state-of-the-arts on several benchmark tasks.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 379, + 505, + 414 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 420, + 208, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 209, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 209, + 432 + ], + "score": 1.0, + "content": "1.1 RELATED WORKS", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 506, + 446 + ], + "score": 1.0, + "content": "Designing non-uniform sampling strategies for random search to improve the planning efficiency", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 443, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 455 + ], + "score": 1.0, + "content": "has been considered as we discussed above. Besides the mentioned algorithms, there are other", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "score": 1.0, + "content": "works along this line, including informed RRT∗ (Gammell et al., 2014) and batch informed Trees", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 107, + 465, + 135, + 476 + ], + "score": 0.57, + "content": "\\mathrm { ( B I T ^ { * } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 466, + 402, + 477 + ], + "score": 1.0, + "content": "(Gammell et al., 2015) as the representative work. Randomized", + "type": "text" + }, + { + "bbox": [ + 403, + 465, + 416, + 475 + ], + "score": 0.87, + "content": "A ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 466, + 506, + 477 + ], + "score": 1.0, + "content": "(Diankov & Kuffner,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 216, + 489 + ], + "score": 1.0, + "content": "2007) and sampling-based", + "type": "text" + }, + { + "bbox": [ + 216, + 477, + 229, + 487 + ], + "score": 0.87, + "content": "A ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "(Persson & Sharf, 2014) expand the search tree with hand-designed", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "heuristics. These methods incorporate the human prior knowledge via hard-coded rules, which is", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "score": 1.0, + "content": "fixed and unable to adapt to problems, and thus, may not universally applicable. Choudhury et al.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 509, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 506, + 522 + ], + "score": 1.0, + "content": "(2018); Song et al. (2018) attempt to learn search heuristics. However, both methods are restricted to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 520, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 532 + ], + "score": 1.0, + "content": "planning on discrete domains. Meanwhile, the latter one always employs an unnecessary hierarchical", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 531, + 475, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 475, + 544 + ], + "score": 1.0, + "content": "structure for path planning, which leads to inferior sample efficiency and extra computation.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 431, + 507, + 544 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 548, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "The online exploration-exploitation trade-off is also an important issue in planning. For instance,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 571 + ], + "score": 1.0, + "content": "Rickert et al. (2009) constructs a potential field sampler and tuned the sampler variance based on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "collision rate for the trade-off heuristically. Paxton et al. (2017) separates the action space into", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "high-level discrete options and low-level continuous actions, and only considered the trade-off at the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "score": 1.0, + "content": "discrete option level, ignoring the exploration-exploitation in the fine action space. These existing", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 603, + 444, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 444, + 615 + ], + "score": 1.0, + "content": "works address the trade-off in an ad-hoc way, which may be inferior for the balance.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 547, + 506, + 615 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 619, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "There have been sevearl non-learning-based planning methods that can also leverage experi-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "ences (Kavraki et al., 1996; Phillips et al., 2012) by utilizing search graphs created in previous", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "problems. However, they are designed for largely fixed obstacles and cannot be generalized to unseen", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 652, + 316, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 316, + 666 + ], + "score": 1.0, + "content": "tasks from the same planning problems distribution.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 619, + 506, + 666 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 670, + 307, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 669, + 309, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 309, + 684 + ], + "score": 1.0, + "content": "2 SETTINGS FOR LEARNING TO PLAN", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 123, + 700 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 687, + 158, + 699 + ], + "score": 0.91, + "content": "S \\subseteq \\mathbb { R } ^ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "be the state space of the problem, e.g., all the configurations of a robot and its base", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 698, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 219, + 712 + ], + "score": 1.0, + "content": "location in the workspace,", + "type": "text" + }, + { + "bbox": [ + 219, + 699, + 261, + 711 + ], + "score": 0.92, + "content": "S _ { o b s } \\subsetneq S", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 698, + 350, + 712 + ], + "score": 1.0, + "content": "be the obstacles set,", + "type": "text" + }, + { + "bbox": [ + 351, + 699, + 429, + 711 + ], + "score": 0.91, + "content": "S _ { f r e e } : = \\mathcal { S } \\setminus \\mathcal { S } _ { o b s }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 698, + 507, + 712 + ], + "score": 1.0, + "content": "be the free space,", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 107, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 161, + 722 + ], + "score": 0.92, + "content": "s _ { i n i t } \\in S _ { f r e e }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 709, + 249, + 722 + ], + "score": 1.0, + "content": "be the initial state and", + "type": "text" + }, + { + "bbox": [ + 250, + 710, + 308, + 722 + ], + "score": 0.92, + "content": "\\mathcal { S } _ { g o a l } \\subsetneq S _ { f r e e }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "be the goal region. Then the space of all collision-", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 504, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 309, + 734 + ], + "score": 1.0, + "content": "free paths can be defined as a continuous function", + "type": "text" + }, + { + "bbox": [ + 310, + 720, + 428, + 733 + ], + "score": 0.91, + "content": "\\Xi : = \\{ \\bar { \\xi } ( \\cdot ) : \\bar { [ 0 , 1 ] } S _ { f r e e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 720, + 448, + 734 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 448, + 720, + 504, + 733 + ], + "score": 0.89, + "content": "c \\left( \\cdot \\right) : \\Xi \\mapsto \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "be the cost functional over a path. The optimal path planning problem is to find the optimal path in", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 393, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 160, + 108 + ], + "score": 1.0, + "content": "terms of cost", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 161, + 93, + 176, + 106 + ], + "score": 0.9, + "content": "c ( \\cdot )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 177, + 92, + 219, + 108 + ], + "score": 1.0, + "content": "from start", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 219, + 95, + 239, + 105 + ], + "score": 0.88, + "content": "s _ { i n i t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 240, + 92, + 270, + 108 + ], + "score": 1.0, + "content": "to goal", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 271, + 94, + 293, + 106 + ], + "score": 0.92, + "content": "\\mathcal { S } _ { g o a l }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 293, + 92, + 347, + 108 + ], + "score": 1.0, + "content": "in free space", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 347, + 94, + 371, + 106 + ], + "score": 0.88, + "content": "\\boldsymbol { S } _ { f r e e }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 371, + 92, + 393, + 108 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 52.5, + "bbox_fs": [ + 105, + 686, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "be the cost functional over a path. The optimal path planning problem is to find the optimal path in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 393, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 160, + 108 + ], + "score": 1.0, + "content": "terms of cost", + "type": "text" + }, + { + "bbox": [ + 161, + 93, + 176, + 106 + ], + "score": 0.9, + "content": "c ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 92, + 219, + 108 + ], + "score": 1.0, + "content": "from start", + "type": "text" + }, + { + "bbox": [ + 219, + 95, + 239, + 105 + ], + "score": 0.88, + "content": "s _ { i n i t }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 92, + 270, + 108 + ], + "score": 1.0, + "content": "to goal", + "type": "text" + }, + { + "bbox": [ + 271, + 94, + 293, + 106 + ], + "score": 0.92, + "content": "\\mathcal { S } _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 92, + 347, + 108 + ], + "score": 1.0, + "content": "in free space", + "type": "text" + }, + { + "bbox": [ + 347, + 94, + 371, + 106 + ], + "score": 0.88, + "content": "\\boldsymbol { S } _ { f r e e }", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 92, + 393, + 108 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 106, + 421, + 120 + ], + "lines": [ + { + "bbox": [ + 189, + 106, + 421, + 119 + ], + "spans": [ + { + "bbox": [ + 189, + 106, + 421, + 119 + ], + "score": 0.73, + "content": "\\begin{array} { r } { \\xi ^ { * } = \\mathrm { a r g m i n } _ { \\xi \\in \\Xi } \\ c \\left( \\xi \\right) , \\quad \\mathrm { s . t . } \\xi ( 0 ) = s _ { i n i t } , \\xi ( 1 ) \\in S _ { g o a l } . } \\end{array}", + "type": "interline_equation", + "image_path": "6219e9be4af3c9da34e4d21a4c18ba35346a4c909f89f4775907c40c6f1e2b0f.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 189, + 106, + 421, + 120 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 120, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 399, + 133 + ], + "score": 1.0, + "content": "Traditionally (Karaman & Frazzoli, 2011), the planner has direct access to", + "type": "text" + }, + { + "bbox": [ + 399, + 120, + 473, + 132 + ], + "score": 0.91, + "content": "( s _ { i n i t } , S _ { g o a l } , c \\left( \\cdot \\right) )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 120, + 506, + 133 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 130, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 104, + 130, + 389, + 146 + ], + "score": 1.0, + "content": "workspace map (Ichter et al., 2018; Tamar et al., 2016; Lee et al., 2018),", + "type": "text" + }, + { + "bbox": [ + 390, + 132, + 437, + 145 + ], + "score": 0.91, + "content": "\\mathtt { m a p } ( \\cdot ) : \\mathbb { R } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 130, + 450, + 146 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 450, + 132, + 502, + 145 + ], + "score": 0.93, + "content": "\\mathbb { R } ^ { 3 } \\mapsto \\{ 0 , 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 130, + 506, + 146 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 280, + 157 + ], + "score": 1.0, + "content": "(0: free spaces and 1: obstacles). Since", + "type": "text" + }, + { + "bbox": [ + 280, + 144, + 304, + 156 + ], + "score": 0.91, + "content": "\\boldsymbol { S } _ { f r e e }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 143, + 506, + 157 + ], + "score": 1.0, + "content": "often has a very irregular geometry (illustrated", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 166 + ], + "score": 1.0, + "content": "in Figure 10 in Appendix A), it is usually represented via a collision detection module which is able", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 166, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 438, + 177 + ], + "score": 1.0, + "content": "to detect the obstacles in a path segment. For the same reason, the feasible paths in", + "type": "text" + }, + { + "bbox": [ + 439, + 166, + 447, + 176 + ], + "score": 0.56, + "content": "\\Xi", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 166, + 505, + 177 + ], + "score": 1.0, + "content": "are hard to be", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 345, + 189 + ], + "score": 1.0, + "content": "described in parametric forms, and thus, the nonparametric", + "type": "text" + }, + { + "bbox": [ + 345, + 177, + 351, + 188 + ], + "score": 0.82, + "content": "\\xi", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 176, + 506, + 189 + ], + "score": 1.0, + "content": ", such as a sequence of interconnected", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 167, + 200 + ], + "score": 1.0, + "content": "path segments", + "type": "text" + }, + { + "bbox": [ + 167, + 187, + 313, + 199 + ], + "score": 0.87, + "content": "\\left[ s _ { 0 } , s _ { 1 } \\right] , \\left[ s _ { 1 } , s _ { 2 } \\right] , \\ldots , \\left[ s _ { T - 1 } , s _ { T } \\right] \\subset \\mathcal { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 187, + 335, + 200 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 335, + 187, + 409, + 199 + ], + "score": 0.92, + "content": "\\xi ( 0 ) = s _ { 0 } = s _ { i n i t }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 187, + 427, + 200 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 428, + 187, + 470, + 199 + ], + "score": 0.93, + "content": "\\xi ( 1 ) = s _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 187, + 506, + 200 + ], + "score": 1.0, + "content": ", is used", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 193, + 275, + 218 + ], + "spans": [ + { + "bbox": [ + 102, + 193, + 191, + 218 + ], + "score": 1.0, + "content": "with an additive cost", + "type": "text" + }, + { + "bbox": [ + 191, + 199, + 267, + 214 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { i = 1 } ^ { T } c \\left( \\left[ s _ { i - 1 } , s _ { i } \\right] \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 193, + 275, + 218 + ], + "score": 1.0, + "content": "−.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 219, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 101, + 212, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 101, + 212, + 181, + 240 + ], + "score": 1.0, + "content": "Assuming given th", + "type": "text" + }, + { + "bbox": [ + 190, + 212, + 264, + 240 + ], + "score": 1.0, + "content": "planning problems", + "type": "text" + }, + { + "bbox": [ + 264, + 218, + 446, + 233 + ], + "score": 0.86, + "content": "\\{ U _ { i } : = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c \\left( \\cdot \\right) ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 212, + 505, + 240 + ], + "score": 1.0, + "content": "sampled from", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 181, + 245 + ], + "score": 1.0, + "content": "some distribution", + "type": "text" + }, + { + "bbox": [ + 181, + 232, + 190, + 241 + ], + "score": 0.77, + "content": "\\mathcal { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 231, + 374, + 245 + ], + "score": 1.0, + "content": ", we are interested in learning an algorithm", + "type": "text" + }, + { + "bbox": [ + 374, + 232, + 402, + 243 + ], + "score": 0.82, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 231, + 505, + 245 + ], + "score": 1.0, + "content": ", which can produce the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "(nearly)-optimal path efficiently from the observed planning problems. Formally, the learning to plan", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 252, + 159, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 159, + 266 + ], + "score": 1.0, + "content": "is defined as", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 259, + 397, + 274 + ], + "lines": [ + { + "bbox": [ + 213, + 259, + 397, + 274 + ], + "spans": [ + { + "bbox": [ + 213, + 259, + 397, + 274 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\mathbf { a } \\mathbf { 1 g } ^ { * } \\left( \\cdot \\right) = \\operatorname * { a r g m i n } _ { \\mathbf { a } \\mathbf { 1 g } \\in \\mathcal { A } } \\mathbb { E } _ { U \\in \\mathcal { U } } \\left[ \\ell \\left( \\mathbf { a } \\mathbf { 1 g } ( U ) \\right) \\right] , } \\end{array}", + "type": "interline_equation", + "image_path": "8d470ec5867b80166a3f19c2e34b8a822874d68e6287bf6444650bf2bae8b2b9.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 213, + 259, + 397, + 274 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 506, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 132, + 287 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 276, + 142, + 285 + ], + "score": 0.78, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 275, + 313, + 287 + ], + "score": 1.0, + "content": "denotes the planning algorithm family, and", + "type": "text" + }, + { + "bbox": [ + 313, + 275, + 331, + 287 + ], + "score": 0.86, + "content": "\\ell \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "denotes some loss function which evaluates", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 329, + 299 + ], + "score": 1.0, + "content": "the quality of the generated path and the efficiency of the", + "type": "text" + }, + { + "bbox": [ + 329, + 286, + 358, + 298 + ], + "score": 0.87, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 286, + 507, + 299 + ], + "score": 1.0, + "content": ", e.g., size of the search tree. We elab-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "orate each component in Eq (2) in the following sections. We first introduce the tree-based sampling", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 308, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 363, + 320 + ], + "score": 1.0, + "content": "algorithm template in Section 3, upon which we instantiate the", + "type": "text" + }, + { + "bbox": [ + 363, + 308, + 392, + 320 + ], + "score": 0.87, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 308, + 506, + 320 + ], + "score": 1.0, + "content": "via a novel attention-based", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 320, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 331 + ], + "score": 1.0, + "content": "neural parametrization in Section 4.2 with exploration-exploitation balance mechanism in Section 4.1.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 329, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 167, + 344 + ], + "score": 1.0, + "content": "We design the", + "type": "text" + }, + { + "bbox": [ + 167, + 330, + 173, + 340 + ], + "score": 0.74, + "content": "\\ell", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 329, + 506, + 344 + ], + "score": 1.0, + "content": "-loss function and the meta learning algorithm in Section 4.3. Composing every", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "component together, we obtain the neural exploration-exploitation trees (NEXT) which achieves", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 325, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 266, + 364 + ], + "score": 1.0, + "content": "outstanding performances in Section 5.", + "type": "text" + }, + { + "bbox": [ + 299, + 356, + 325, + 366 + ], + "score": 1.0, + "content": "Repeat", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "image", + "bbox": [ + 128, + 364, + 480, + 522 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 364, + 480, + 522 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 364, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 128, + 364, + 480, + 522 + ], + "score": 0.969, + "type": "image", + "image_path": "c81f5245215344ac0251524c1e65b7b0d359bb22ec9a016d0918f1939b8a9b94.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 128, + 364, + 480, + 416.6666666666667 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 128, + 416.6666666666667, + 480, + 469.33333333333337 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 128, + 469.33333333333337, + 480, + 522.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 527, + 506, + 559 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 527, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 506, + 538 + ], + "score": 1.0, + "content": "Figure 1: Illustration of NEXT. In each epoch, NEXT is executed on a randomly generated planning problem.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 208, + 549 + ], + "score": 1.0, + "content": "The search tree grows with", + "type": "text" + }, + { + "bbox": [ + 208, + 536, + 221, + 547 + ], + "score": 0.87, + "content": "\\tilde { V } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 537, + 237, + 549 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 538, + 249, + 547 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 537, + 288, + 549 + ], + "score": 1.0, + "content": "guidance.", + "type": "text" + }, + { + "bbox": [ + 289, + 537, + 324, + 549 + ], + "score": 0.93, + "content": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 537, + 506, + 549 + ], + "score": 1.0, + "content": "will be updated according to the successful path.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 549, + 333, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 333, + 559 + ], + "score": 1.0, + "content": "Such planning and learning iteration is continued interactively.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 108, + 568, + 208, + 580 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 209, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 209, + 583 + ], + "score": 1.0, + "content": "3 PRELIMINARIES", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 299, + 579, + 345, + 588 + ], + "lines": [ + { + "bbox": [ + 296, + 577, + 347, + 591 + ], + "spans": [ + { + "bbox": [ + 296, + 577, + 347, + 591 + ], + "score": 1.0, + "content": "1 Problem:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 305, + 564, + 489, + 576 + ], + "lines": [ + { + "bbox": [ + 303, + 562, + 490, + 578 + ], + "spans": [ + { + "bbox": [ + 303, + 562, + 490, + 578 + ], + "score": 1.0, + "content": "Algorithm 1: Tree-based Sampling Algorithm", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 300, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 589, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 107, + 589, + 294, + 601 + ], + "score": 1.0, + "content": "The sampling-based planners are more practi-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 600, + 302, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 302, + 612 + ], + "score": 1.0, + "content": "cal and become dominant for high-dimensional", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 611, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 294, + 623 + ], + "score": 1.0, + "content": "path planning problems (Elbanhawi & Simic,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 621, + 303, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 303, + 635 + ], + "score": 1.0, + "content": "3 2014). We describe a unifying view for many", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 633, + 303, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 303, + 645 + ], + "score": 1.0, + "content": "4 existing tree-based sampling algorithms (TSA),", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 643, + 302, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 302, + 656 + ], + "score": 1.0, + "content": "5 which we will also base our algorithm upon.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 655, + 303, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 295, + 667 + ], + "score": 1.0, + "content": "More specifically, this family of algorithms", + "type": "text" + }, + { + "bbox": [ + 296, + 658, + 303, + 666 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 666, + 293, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 200, + 677 + ], + "score": 1.0, + "content": "maintain a search tree", + "type": "text" + }, + { + "bbox": [ + 200, + 666, + 210, + 676 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 666, + 293, + 677 + ], + "score": 1.0, + "content": "rooted at the initial", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 677, + 303, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 128, + 689 + ], + "score": 1.0, + "content": "point", + "type": "text" + }, + { + "bbox": [ + 129, + 678, + 149, + 688 + ], + "score": 0.9, + "content": "s _ { i n i t }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 677, + 284, + 689 + ], + "score": 1.0, + "content": "and connecting all sampled points", + "type": "text" + }, + { + "bbox": [ + 284, + 677, + 293, + 687 + ], + "score": 0.63, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 680, + 303, + 688 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 688, + 293, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 263, + 700 + ], + "score": 1.0, + "content": "in the configuration space with edge set", + "type": "text" + }, + { + "bbox": [ + 264, + 688, + 271, + 698 + ], + "score": 0.72, + "content": "\\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 688, + 293, + 700 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 302, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 302, + 712 + ], + "score": 1.0, + "content": "8 tree will be expanded by incorporating more", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 708, + 301, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 268, + 723 + ], + "score": 1.0, + "content": "sampled states until some leaf reaches", + "type": "text" + }, + { + "bbox": [ + 268, + 710, + 291, + 722 + ], + "score": 0.88, + "content": "\\mathcal { S } _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 708, + 301, + 723 + ], + "score": 1.0, + "content": "9.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 719, + 293, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 293, + 733 + ], + "score": 1.0, + "content": "Then, a feasible solution for the path planning", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 297, + 588, + 489, + 732 + ], + "lines": [ + { + "bbox": [ + 304, + 587, + 462, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 457, + 601 + ], + "score": 0.81, + "content": "U = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathrm { { m a p } , \\it { c ( \\cdot ) } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 587, + 462, + 603 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 295, + 599, + 468, + 614 + ], + "spans": [ + { + "bbox": [ + 295, + 599, + 342, + 614 + ], + "score": 1.0, + "content": "2 Initialize", + "type": "text" + }, + { + "bbox": [ + 343, + 601, + 390, + 613 + ], + "score": 0.9, + "content": "\\boldsymbol { \\mathcal { T } } = ( \\boldsymbol { \\mathcal { V } } , \\boldsymbol { \\mathcal { E } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 599, + 411, + 614 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 412, + 601, + 464, + 613 + ], + "score": 0.88, + "content": "\\mathcal { V } \\{ s _ { i n i t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 599, + 468, + 614 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 309, + 610, + 343, + 623 + ], + "spans": [ + { + "bbox": [ + 309, + 612, + 337, + 622 + ], + "score": 0.86, + "content": "\\mathcal { E } \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 610, + 343, + 623 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 294, + 622, + 380, + 633 + ], + "spans": [ + { + "bbox": [ + 294, + 622, + 319, + 633 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 319, + 623, + 345, + 633 + ], + "score": 0.87, + "content": "t \\gets 0", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 622, + 356, + 633 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 357, + 623, + 366, + 632 + ], + "score": 0.68, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 622, + 380, + 633 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 319, + 633, + 466, + 645 + ], + "spans": [ + { + "bbox": [ + 319, + 634, + 353, + 645 + ], + "score": 1.0, + "content": "sparent,", + "type": "text" + }, + { + "bbox": [ + 353, + 635, + 396, + 645 + ], + "score": 0.63, + "content": "s _ { n e w } \\mathrm { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 634, + 428, + 645 + ], + "score": 1.0, + "content": "xpand", + "type": "text" + }, + { + "bbox": [ + 428, + 633, + 457, + 644 + ], + "score": 0.49, + "content": "( \\mathcal { T } , U )", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 634, + 466, + 645 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 318, + 643, + 491, + 658 + ], + "spans": [ + { + "bbox": [ + 318, + 643, + 405, + 658 + ], + "score": 1.0, + "content": "if ObstacleFree(", + "type": "text" + }, + { + "bbox": [ + 405, + 646, + 463, + 657 + ], + "score": 0.46, + "content": "\\prime s _ { p a r e n t } , s _ { n e w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 643, + 491, + 658 + ], + "score": 1.0, + "content": ") then", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 335, + 653, + 426, + 669 + ], + "spans": [ + { + "bbox": [ + 335, + 655, + 407, + 667 + ], + "score": 0.77, + "content": "\\mathcal { V } \\mathcal { V } \\cup \\{ s _ { n e w } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 653, + 426, + 669 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 337, + 664, + 455, + 681 + ], + "spans": [ + { + "bbox": [ + 337, + 667, + 450, + 678 + ], + "score": 0.67, + "content": "\\mathcal { E } \\mathcal { E } \\cup \\{ [ s _ { p a r e n t } , s _ { n e w } ] \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 664, + 455, + 681 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 334, + 676, + 465, + 690 + ], + "spans": [ + { + "bbox": [ + 334, + 676, + 465, + 690 + ], + "score": 1.0, + "content": "T ← Postprocess(T , U) ;", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 339, + 688, + 385, + 700 + ], + "spans": [ + { + "bbox": [ + 339, + 690, + 347, + 698 + ], + "score": 0.39, + "content": "\\triangleright", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 688, + 385, + 700 + ], + "score": 1.0, + "content": "Optional", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 333, + 696, + 423, + 713 + ], + "spans": [ + { + "bbox": [ + 333, + 696, + 343, + 713 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 344, + 700, + 398, + 711 + ], + "score": 0.38, + "content": "s _ { n e w } \\in S _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 696, + 423, + 713 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 348, + 709, + 395, + 723 + ], + "spans": [ + { + "bbox": [ + 348, + 709, + 395, + 723 + ], + "score": 1.0, + "content": "return T ;", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 51.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 506, + 108 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 106, + 421, + 120 + ], + "lines": [ + { + "bbox": [ + 189, + 106, + 421, + 119 + ], + "spans": [ + { + "bbox": [ + 189, + 106, + 421, + 119 + ], + "score": 0.73, + "content": "\\begin{array} { r } { \\xi ^ { * } = \\mathrm { a r g m i n } _ { \\xi \\in \\Xi } \\ c \\left( \\xi \\right) , \\quad \\mathrm { s . t . } \\xi ( 0 ) = s _ { i n i t } , \\xi ( 1 ) \\in S _ { g o a l } . } \\end{array}", + "type": "interline_equation", + "image_path": "6219e9be4af3c9da34e4d21a4c18ba35346a4c909f89f4775907c40c6f1e2b0f.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 189, + 106, + 421, + 120 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 120, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 399, + 133 + ], + "score": 1.0, + "content": "Traditionally (Karaman & Frazzoli, 2011), the planner has direct access to", + "type": "text" + }, + { + "bbox": [ + 399, + 120, + 473, + 132 + ], + "score": 0.91, + "content": "( s _ { i n i t } , S _ { g o a l } , c \\left( \\cdot \\right) )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 120, + 506, + 133 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 130, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 104, + 130, + 389, + 146 + ], + "score": 1.0, + "content": "workspace map (Ichter et al., 2018; Tamar et al., 2016; Lee et al., 2018),", + "type": "text" + }, + { + "bbox": [ + 390, + 132, + 437, + 145 + ], + "score": 0.91, + "content": "\\mathtt { m a p } ( \\cdot ) : \\mathbb { R } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 130, + 450, + 146 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 450, + 132, + 502, + 145 + ], + "score": 0.93, + "content": "\\mathbb { R } ^ { 3 } \\mapsto \\{ 0 , 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 130, + 506, + 146 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 280, + 157 + ], + "score": 1.0, + "content": "(0: free spaces and 1: obstacles). Since", + "type": "text" + }, + { + "bbox": [ + 280, + 144, + 304, + 156 + ], + "score": 0.91, + "content": "\\boldsymbol { S } _ { f r e e }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 143, + 506, + 157 + ], + "score": 1.0, + "content": "often has a very irregular geometry (illustrated", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 166 + ], + "score": 1.0, + "content": "in Figure 10 in Appendix A), it is usually represented via a collision detection module which is able", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 166, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 438, + 177 + ], + "score": 1.0, + "content": "to detect the obstacles in a path segment. For the same reason, the feasible paths in", + "type": "text" + }, + { + "bbox": [ + 439, + 166, + 447, + 176 + ], + "score": 0.56, + "content": "\\Xi", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 166, + 505, + 177 + ], + "score": 1.0, + "content": "are hard to be", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 345, + 189 + ], + "score": 1.0, + "content": "described in parametric forms, and thus, the nonparametric", + "type": "text" + }, + { + "bbox": [ + 345, + 177, + 351, + 188 + ], + "score": 0.82, + "content": "\\xi", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 176, + 506, + 189 + ], + "score": 1.0, + "content": ", such as a sequence of interconnected", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 167, + 200 + ], + "score": 1.0, + "content": "path segments", + "type": "text" + }, + { + "bbox": [ + 167, + 187, + 313, + 199 + ], + "score": 0.87, + "content": "\\left[ s _ { 0 } , s _ { 1 } \\right] , \\left[ s _ { 1 } , s _ { 2 } \\right] , \\ldots , \\left[ s _ { T - 1 } , s _ { T } \\right] \\subset \\mathcal { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 187, + 335, + 200 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 335, + 187, + 409, + 199 + ], + "score": 0.92, + "content": "\\xi ( 0 ) = s _ { 0 } = s _ { i n i t }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 187, + 427, + 200 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 428, + 187, + 470, + 199 + ], + "score": 0.93, + "content": "\\xi ( 1 ) = s _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 187, + 506, + 200 + ], + "score": 1.0, + "content": ", is used", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 193, + 275, + 218 + ], + "spans": [ + { + "bbox": [ + 102, + 193, + 191, + 218 + ], + "score": 1.0, + "content": "with an additive cost", + "type": "text" + }, + { + "bbox": [ + 191, + 199, + 267, + 214 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { i = 1 } ^ { T } c \\left( \\left[ s _ { i - 1 } , s _ { i } \\right] \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 193, + 275, + 218 + ], + "score": 1.0, + "content": "−.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5, + "bbox_fs": [ + 102, + 120, + 506, + 218 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 219, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 101, + 212, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 101, + 212, + 181, + 240 + ], + "score": 1.0, + "content": "Assuming given th", + "type": "text" + }, + { + "bbox": [ + 190, + 212, + 264, + 240 + ], + "score": 1.0, + "content": "planning problems", + "type": "text" + }, + { + "bbox": [ + 264, + 218, + 446, + 233 + ], + "score": 0.86, + "content": "\\{ U _ { i } : = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c \\left( \\cdot \\right) ) \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 212, + 505, + 240 + ], + "score": 1.0, + "content": "sampled from", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 231, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 181, + 245 + ], + "score": 1.0, + "content": "some distribution", + "type": "text" + }, + { + "bbox": [ + 181, + 232, + 190, + 241 + ], + "score": 0.77, + "content": "\\mathcal { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 231, + 374, + 245 + ], + "score": 1.0, + "content": ", we are interested in learning an algorithm", + "type": "text" + }, + { + "bbox": [ + 374, + 232, + 402, + 243 + ], + "score": 0.82, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 231, + 505, + 245 + ], + "score": 1.0, + "content": ", which can produce the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "(nearly)-optimal path efficiently from the observed planning problems. Formally, the learning to plan", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 252, + 159, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 159, + 266 + ], + "score": 1.0, + "content": "is defined as", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 101, + 212, + 505, + 266 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 259, + 397, + 274 + ], + "lines": [ + { + "bbox": [ + 213, + 259, + 397, + 274 + ], + "spans": [ + { + "bbox": [ + 213, + 259, + 397, + 274 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\mathbf { a } \\mathbf { 1 g } ^ { * } \\left( \\cdot \\right) = \\operatorname * { a r g m i n } _ { \\mathbf { a } \\mathbf { 1 g } \\in \\mathcal { A } } \\mathbb { E } _ { U \\in \\mathcal { U } } \\left[ \\ell \\left( \\mathbf { a } \\mathbf { 1 g } ( U ) \\right) \\right] , } \\end{array}", + "type": "interline_equation", + "image_path": "8d470ec5867b80166a3f19c2e34b8a822874d68e6287bf6444650bf2bae8b2b9.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 213, + 259, + 397, + 274 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 506, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 132, + 287 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 276, + 142, + 285 + ], + "score": 0.78, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 275, + 313, + 287 + ], + "score": 1.0, + "content": "denotes the planning algorithm family, and", + "type": "text" + }, + { + "bbox": [ + 313, + 275, + 331, + 287 + ], + "score": 0.86, + "content": "\\ell \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "denotes some loss function which evaluates", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 329, + 299 + ], + "score": 1.0, + "content": "the quality of the generated path and the efficiency of the", + "type": "text" + }, + { + "bbox": [ + 329, + 286, + 358, + 298 + ], + "score": 0.87, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 286, + 507, + 299 + ], + "score": 1.0, + "content": ", e.g., size of the search tree. We elab-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 297, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 506, + 309 + ], + "score": 1.0, + "content": "orate each component in Eq (2) in the following sections. We first introduce the tree-based sampling", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 308, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 363, + 320 + ], + "score": 1.0, + "content": "algorithm template in Section 3, upon which we instantiate the", + "type": "text" + }, + { + "bbox": [ + 363, + 308, + 392, + 320 + ], + "score": 0.87, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 308, + 506, + 320 + ], + "score": 1.0, + "content": "via a novel attention-based", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 320, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 331 + ], + "score": 1.0, + "content": "neural parametrization in Section 4.2 with exploration-exploitation balance mechanism in Section 4.1.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 329, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 167, + 344 + ], + "score": 1.0, + "content": "We design the", + "type": "text" + }, + { + "bbox": [ + 167, + 330, + 173, + 340 + ], + "score": 0.74, + "content": "\\ell", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 329, + 506, + 344 + ], + "score": 1.0, + "content": "-loss function and the meta learning algorithm in Section 4.3. Composing every", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "component together, we obtain the neural exploration-exploitation trees (NEXT) which achieves", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 325, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 266, + 364 + ], + "score": 1.0, + "content": "outstanding performances in Section 5.", + "type": "text" + }, + { + "bbox": [ + 299, + 356, + 325, + 366 + ], + "score": 1.0, + "content": "Repeat", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5, + "bbox_fs": [ + 104, + 275, + 507, + 366 + ] + }, + { + "type": "image", + "bbox": [ + 128, + 364, + 480, + 522 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 128, + 364, + 480, + 522 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 128, + 364, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 128, + 364, + 480, + 522 + ], + "score": 0.969, + "type": "image", + "image_path": "c81f5245215344ac0251524c1e65b7b0d359bb22ec9a016d0918f1939b8a9b94.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 128, + 364, + 480, + 416.6666666666667 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 128, + 416.6666666666667, + 480, + 469.33333333333337 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 128, + 469.33333333333337, + 480, + 522.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 527, + 506, + 559 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 527, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 506, + 538 + ], + "score": 1.0, + "content": "Figure 1: Illustration of NEXT. In each epoch, NEXT is executed on a randomly generated planning problem.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 536, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 208, + 549 + ], + "score": 1.0, + "content": "The search tree grows with", + "type": "text" + }, + { + "bbox": [ + 208, + 536, + 221, + 547 + ], + "score": 0.87, + "content": "\\tilde { V } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 537, + 237, + 549 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 538, + 249, + 547 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 537, + 288, + 549 + ], + "score": 1.0, + "content": "guidance.", + "type": "text" + }, + { + "bbox": [ + 289, + 537, + 324, + 549 + ], + "score": 0.93, + "content": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 537, + 506, + 549 + ], + "score": 1.0, + "content": "will be updated according to the successful path.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 549, + 333, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 333, + 559 + ], + "score": 1.0, + "content": "Such planning and learning iteration is continued interactively.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 108, + 568, + 208, + 580 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 209, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 209, + 583 + ], + "score": 1.0, + "content": "3 PRELIMINARIES", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 299, + 579, + 345, + 588 + ], + "lines": [ + { + "bbox": [ + 296, + 577, + 347, + 591 + ], + "spans": [ + { + "bbox": [ + 296, + 577, + 347, + 591 + ], + "score": 1.0, + "content": "1 Problem:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 305, + 564, + 489, + 576 + ], + "lines": [ + { + "bbox": [ + 303, + 562, + 490, + 578 + ], + "spans": [ + { + "bbox": [ + 303, + 562, + 490, + 578 + ], + "score": 1.0, + "content": "Algorithm 1: Tree-based Sampling Algorithm", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 300, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 589, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 107, + 589, + 294, + 601 + ], + "score": 1.0, + "content": "The sampling-based planners are more practi-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 600, + 302, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 302, + 612 + ], + "score": 1.0, + "content": "cal and become dominant for high-dimensional", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 611, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 294, + 623 + ], + "score": 1.0, + "content": "path planning problems (Elbanhawi & Simic,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 621, + 303, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 303, + 635 + ], + "score": 1.0, + "content": "3 2014). We describe a unifying view for many", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 633, + 303, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 303, + 645 + ], + "score": 1.0, + "content": "4 existing tree-based sampling algorithms (TSA),", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 643, + 302, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 302, + 656 + ], + "score": 1.0, + "content": "5 which we will also base our algorithm upon.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 655, + 303, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 295, + 667 + ], + "score": 1.0, + "content": "More specifically, this family of algorithms", + "type": "text" + }, + { + "bbox": [ + 296, + 658, + 303, + 666 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 666, + 293, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 200, + 677 + ], + "score": 1.0, + "content": "maintain a search tree", + "type": "text" + }, + { + "bbox": [ + 200, + 666, + 210, + 676 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 666, + 293, + 677 + ], + "score": 1.0, + "content": "rooted at the initial", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 677, + 303, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 128, + 689 + ], + "score": 1.0, + "content": "point", + "type": "text" + }, + { + "bbox": [ + 129, + 678, + 149, + 688 + ], + "score": 0.9, + "content": "s _ { i n i t }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 677, + 284, + 689 + ], + "score": 1.0, + "content": "and connecting all sampled points", + "type": "text" + }, + { + "bbox": [ + 284, + 677, + 293, + 687 + ], + "score": 0.63, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 680, + 303, + 688 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 688, + 293, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 263, + 700 + ], + "score": 1.0, + "content": "in the configuration space with edge set", + "type": "text" + }, + { + "bbox": [ + 264, + 688, + 271, + 698 + ], + "score": 0.72, + "content": "\\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 688, + 293, + 700 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 302, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 302, + 712 + ], + "score": 1.0, + "content": "8 tree will be expanded by incorporating more", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 708, + 301, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 268, + 723 + ], + "score": 1.0, + "content": "sampled states until some leaf reaches", + "type": "text" + }, + { + "bbox": [ + 268, + 710, + 291, + 722 + ], + "score": 0.88, + "content": "\\mathcal { S } _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 708, + 301, + 723 + ], + "score": 1.0, + "content": "9.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 719, + 293, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 293, + 733 + ], + "score": 1.0, + "content": "Then, a feasible solution for the path planning", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 589, + 303, + 733 + ] + }, + { + "type": "list", + "bbox": [ + 297, + 588, + 489, + 732 + ], + "lines": [ + { + "bbox": [ + 304, + 587, + 462, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 457, + 601 + ], + "score": 0.81, + "content": "U = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathrm { { m a p } , \\it { c ( \\cdot ) } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 587, + 462, + 603 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 295, + 599, + 468, + 614 + ], + "spans": [ + { + "bbox": [ + 295, + 599, + 342, + 614 + ], + "score": 1.0, + "content": "2 Initialize", + "type": "text" + }, + { + "bbox": [ + 343, + 601, + 390, + 613 + ], + "score": 0.9, + "content": "\\boldsymbol { \\mathcal { T } } = ( \\boldsymbol { \\mathcal { V } } , \\boldsymbol { \\mathcal { E } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 599, + 411, + 614 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 412, + 601, + 464, + 613 + ], + "score": 0.88, + "content": "\\mathcal { V } \\{ s _ { i n i t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 599, + 468, + 614 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 309, + 610, + 343, + 623 + ], + "spans": [ + { + "bbox": [ + 309, + 612, + 337, + 622 + ], + "score": 0.86, + "content": "\\mathcal { E } \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 610, + 343, + 623 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 48, + "is_list_end_line": true + }, + { + "bbox": [ + 294, + 622, + 380, + 633 + ], + "spans": [ + { + "bbox": [ + 294, + 622, + 319, + 633 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 319, + 623, + 345, + 633 + ], + "score": 0.87, + "content": "t \\gets 0", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 622, + 356, + 633 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 357, + 623, + 366, + 632 + ], + "score": 0.68, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 622, + 380, + 633 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 49, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 319, + 633, + 466, + 645 + ], + "spans": [ + { + "bbox": [ + 319, + 634, + 353, + 645 + ], + "score": 1.0, + "content": "sparent,", + "type": "text" + }, + { + "bbox": [ + 353, + 635, + 396, + 645 + ], + "score": 0.63, + "content": "s _ { n e w } \\mathrm { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 634, + 428, + 645 + ], + "score": 1.0, + "content": "xpand", + "type": "text" + }, + { + "bbox": [ + 428, + 633, + 457, + 644 + ], + "score": 0.49, + "content": "( \\mathcal { T } , U )", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 634, + 466, + 645 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 318, + 643, + 491, + 658 + ], + "spans": [ + { + "bbox": [ + 318, + 643, + 405, + 658 + ], + "score": 1.0, + "content": "if ObstacleFree(", + "type": "text" + }, + { + "bbox": [ + 405, + 646, + 463, + 657 + ], + "score": 0.46, + "content": "\\prime s _ { p a r e n t } , s _ { n e w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 643, + 491, + 658 + ], + "score": 1.0, + "content": ") then", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 335, + 653, + 426, + 669 + ], + "spans": [ + { + "bbox": [ + 335, + 655, + 407, + 667 + ], + "score": 0.77, + "content": "\\mathcal { V } \\mathcal { V } \\cup \\{ s _ { n e w } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 653, + 426, + 669 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 52, + "is_list_end_line": true + }, + { + "bbox": [ + 337, + 664, + 455, + 681 + ], + "spans": [ + { + "bbox": [ + 337, + 667, + 450, + 678 + ], + "score": 0.67, + "content": "\\mathcal { E } \\mathcal { E } \\cup \\{ [ s _ { p a r e n t } , s _ { n e w } ] \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 664, + 455, + 681 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 53, + "is_list_end_line": true + }, + { + "bbox": [ + 334, + 676, + 465, + 690 + ], + "spans": [ + { + "bbox": [ + 334, + 676, + 465, + 690 + ], + "score": 1.0, + "content": "T ← Postprocess(T , U) ;", + "type": "text" + } + ], + "index": 54, + "is_list_end_line": true + }, + { + "bbox": [ + 339, + 688, + 385, + 700 + ], + "spans": [ + { + "bbox": [ + 339, + 690, + 347, + 698 + ], + "score": 0.39, + "content": "\\triangleright", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 688, + 385, + 700 + ], + "score": 1.0, + "content": "Optional", + "type": "text" + } + ], + "index": 55, + "is_list_end_line": true + }, + { + "bbox": [ + 333, + 696, + 423, + 713 + ], + "spans": [ + { + "bbox": [ + 333, + 696, + 343, + 713 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 344, + 700, + 398, + 711 + ], + "score": 0.38, + "content": "s _ { n e w } \\in S _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 696, + 423, + 713 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 56, + "is_list_end_line": true + }, + { + "bbox": [ + 348, + 709, + 395, + 723 + ], + "spans": [ + { + "bbox": [ + 348, + 709, + 395, + 723 + ], + "score": 1.0, + "content": "return T ;", + "type": "text" + } + ], + "index": 57, + "is_list_end_line": true + } + ], + "index": 51.5, + "bbox_fs": [ + 294, + 587, + 491, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 284, + 95 + ], + "score": 1.0, + "content": "problem will be extracted based on the tree", + "type": "text" + }, + { + "bbox": [ + 285, + 83, + 294, + 93 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 82, + 506, + 95 + ], + "score": 1.0, + "content": ". The template of tree-based sampling algorithms is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "summarized in Algorithm 1 and illustrated in Fig. 1(c). A key component of the algorithm is the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 356, + 118 + ], + "score": 1.0, + "content": "Expand operator, which generates the next exploration point", + "type": "text" + }, + { + "bbox": [ + 356, + 106, + 377, + 116 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 103, + 436, + 118 + ], + "score": 1.0, + "content": "and its parent", + "type": "text" + }, + { + "bbox": [ + 436, + 105, + 487, + 116 + ], + "score": 0.85, + "content": "s _ { p a r e n t } \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 103, + 506, + 118 + ], + "score": 1.0, + "content": ". To", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 507, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 276, + 129 + ], + "score": 1.0, + "content": "ensure the feasibility of the solution, the", + "type": "text" + }, + { + "bbox": [ + 276, + 117, + 297, + 127 + ], + "score": 0.9, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 114, + 402, + 129 + ], + "score": 1.0, + "content": "must be reachable from", + "type": "text" + }, + { + "bbox": [ + 402, + 116, + 411, + 126 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 114, + 434, + 129 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 435, + 116, + 494, + 127 + ], + "score": 0.89, + "content": "[ \\bar { s _ { p a r e n t } } , s _ { n e w } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 114, + 507, + 129 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 506, + 138 + ], + "score": 1.0, + "content": "collision-free, which is checked by a collision detection function. As we will discuss in Appendix B,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "by instantiating different Expand operators, we will arrive at many existing algorithms, such as", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 387, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 387, + 161 + ], + "score": 1.0, + "content": "RRT (LaValle, 1998) and EST (Hsu et al., 1997; Phillips et al., 2004).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "One major limitation of existing TSAs is that they solve each problem independently from scratch", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "and ignore past planning experiences in similar environments. We introduce the neural components", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 385, + 200 + ], + "score": 1.0, + "content": "into TSA template to form the learnable planning algorithm family", + "type": "text" + }, + { + "bbox": [ + 386, + 187, + 395, + 198 + ], + "score": 0.75, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 187, + 505, + 200 + ], + "score": 1.0, + "content": ", which can explicitly take", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 501, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 501, + 211 + ], + "score": 1.0, + "content": "advantages of the past successful experiences to bias the Expand towards more promising regions.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 107, + 219, + 365, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 366, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 366, + 234 + ], + "score": 1.0, + "content": "4 NEURAL EXPLORATION-EXPLOITATION TREES", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "Based on the TSA framework, we introduce a learnable neural based Expand operator, which can", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 248, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 104, + 248, + 351, + 264 + ], + "score": 1.0, + "content": "balance between exploration and exploitation, to instantiate", + "type": "text" + }, + { + "bbox": [ + 352, + 250, + 361, + 260 + ], + "score": 0.75, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 248, + 506, + 264 + ], + "score": 1.0, + "content": "in Eq (2). With the self-improving", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 261, + 382, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 382, + 273 + ], + "score": 1.0, + "content": "training, we obtain the meta NEXT algorithm illustrated in Figure 1.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 107, + 279, + 282, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 282, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 282, + 291 + ], + "score": 1.0, + "content": "4.1 GUIDED PROGRESSIVE EXPANSION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 273, + 373 + ], + "lines": [ + { + "bbox": [ + 106, + 294, + 274, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 274, + 307 + ], + "score": 1.0, + "content": "We start with our design of the Expand.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 305, + 273, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 273, + 317 + ], + "score": 1.0, + "content": "We assume having an estimated value", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 316, + 273, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 144, + 329 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 144, + 316, + 180, + 330 + ], + "score": 0.94, + "content": "{ \\tilde { V } } ^ { * } ( s | U )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 317, + 273, + 329 + ], + "score": 1.0, + "content": ", which stands for the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 329, + 275, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 188, + 340 + ], + "score": 1.0, + "content": "optimal cost from", + "type": "text" + }, + { + "bbox": [ + 189, + 331, + 195, + 339 + ], + "score": 0.68, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 329, + 275, + 340 + ], + "score": 1.0, + "content": "to target in plan-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 339, + 274, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 160, + 352 + ], + "score": 1.0, + "content": "ning problem", + "type": "text" + }, + { + "bbox": [ + 161, + 340, + 170, + 350 + ], + "score": 0.79, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 339, + 222, + 352 + ], + "score": 1.0, + "content": ", and a policy", + "type": "text" + }, + { + "bbox": [ + 222, + 340, + 270, + 352 + ], + "score": 0.9, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 339, + 274, + 352 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 351, + 272, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 263, + 363 + ], + "score": 1.0, + "content": "which generates the promising action", + "type": "text" + }, + { + "bbox": [ + 263, + 352, + 272, + 361 + ], + "score": 0.76, + "content": "s ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 360, + 275, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 151, + 374 + ], + "score": 1.0, + "content": "from state", + "type": "text" + }, + { + "bbox": [ + 151, + 364, + 158, + 371 + ], + "score": 0.67, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 360, + 275, + 374 + ], + "score": 1.0, + "content": ". The concrete parametriza-", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 279, + 297, + 504, + 367 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 279, + 297, + 504, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 279, + 297, + 504, + 367 + ], + "spans": [ + { + "bbox": [ + 279, + 297, + 504, + 367 + ], + "score": 0.205, + "html": "
Algorithm 2: NExT :: Expand(T = (V,ε),U)
Sparent ← argmaxs∈vΦ(s) ; Selection
{s1,*..,Sk} id π*(s'sparent,U)； >Candidates
Snew ← argmaxs'∈{s1,.,sk}(s')； DExpansion
4 return Sparent, Snew;
", + "type": "table", + "image_path": "a8c886f8d6118bbe9fbb4275382954ce28939501d27215b83992823a8b9fe93d.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 279, + 297, + 504, + 311.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 279, + 311.0, + 504, + 325.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 279, + 325.0, + 504, + 339.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 279, + 339.0, + 504, + 353.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 279, + 353.0, + 504, + 367.0 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 504, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 136, + 386 + ], + "score": 1.0, + "content": "tion of", + "type": "text" + }, + { + "bbox": [ + 137, + 372, + 150, + 384 + ], + "score": 0.89, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 372, + 169, + 386 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 169, + 374, + 181, + 384 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 372, + 505, + 386 + ], + "score": 1.0, + "content": "will be explained in Section 4.2 and learned in Section 4.3. We will use these", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 383, + 478, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 478, + 399 + ], + "score": 1.0, + "content": "functions to construct the learnable Expand with explicit exploration-exploitation balancing.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 333, + 415 + ], + "score": 1.0, + "content": "The Expand operator will expand the current search tree", + "type": "text" + }, + { + "bbox": [ + 333, + 402, + 343, + 412 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 401, + 452, + 415 + ], + "score": 1.0, + "content": "by a new neighboring state", + "type": "text" + }, + { + "bbox": [ + 453, + 404, + 474, + 413 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "around", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 411, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 116, + 423 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 411, + 387, + 427 + ], + "score": 1.0, + "content": ". We design the expansion as a two-step procedure: (i) select a state", + "type": "text" + }, + { + "bbox": [ + 387, + 414, + 417, + 425 + ], + "score": 0.81, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 411, + 492, + 427 + ], + "score": 1.0, + "content": "from existing tree", + "type": "text" + }, + { + "bbox": [ + 493, + 413, + 502, + 423 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 411, + 506, + 427 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 423, + 405, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 179, + 437 + ], + "score": 1.0, + "content": "(ii) expand a state", + "type": "text" + }, + { + "bbox": [ + 180, + 425, + 200, + 435 + ], + "score": 0.89, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 423, + 295, + 437 + ], + "score": 1.0, + "content": "in the neighborhood of", + "type": "text" + }, + { + "bbox": [ + 296, + 425, + 326, + 436 + ], + "score": 0.71, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 423, + 405, + 437 + ], + "score": 1.0, + "content": ". More specifically,", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 104, + 441, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 441, + 147, + 455 + ], + "score": 1.0, + "content": "Selecting", + "type": "text" + }, + { + "bbox": [ + 147, + 444, + 177, + 455 + ], + "score": 0.75, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 441, + 201, + 455 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 201, + 443, + 211, + 453 + ], + "score": 0.6, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 441, + 400, + 455 + ], + "score": 1.0, + "content": "in step (i). Consider the negative value function", + "type": "text" + }, + { + "bbox": [ + 400, + 441, + 446, + 455 + ], + "score": 0.93, + "content": "- \\tilde { V } ^ { \\ast } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 441, + 506, + 455 + ], + "score": 1.0, + "content": "as the rewards", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 107, + 453, + 126, + 465 + ], + "score": 0.88, + "content": "r \\left( s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 453, + 505, + 466 + ], + "score": 1.0, + "content": ", step (i) shares some similarity with the multi-armed bandit problem by viewing existing nodes", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 107, + 465, + 133, + 475 + ], + "score": 0.89, + "content": "s \\in \\mathcal V", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "as arms. However, the vanilla UCB algorithm is not directly applicable, since the number", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "of states is increasing as the algorithm proceeds and the value of these adjacent states are naturally", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "correlated. We address this challenge by modeling the correlation explicitly as in contextual bandits.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 497, + 504, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 371, + 510 + ], + "score": 1.0, + "content": "Specifically, we parametrize the UCB of the reward function as", + "type": "text" + }, + { + "bbox": [ + 371, + 498, + 393, + 510 + ], + "score": 0.91, + "content": "\\phi \\left( s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 497, + 495, + 510 + ], + "score": 1.0, + "content": ", and select a node from", + "type": "text" + }, + { + "bbox": [ + 495, + 498, + 504, + 508 + ], + "score": 0.81, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 508, + 178, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 158, + 521 + ], + "score": 1.0, + "content": "according to", + "type": "text" + }, + { + "bbox": [ + 159, + 509, + 178, + 520 + ], + "score": 0.91, + "content": "\\phi ( s )", + "type": "inline_equation" + } + ], + "index": 39 + } + ], + "index": 36 + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 520, + 405, + 534 + ], + "lines": [ + { + "bbox": [ + 204, + 520, + 405, + 534 + ], + "spans": [ + { + "bbox": [ + 204, + 520, + 405, + 534 + ], + "score": 0.86, + "content": "\\begin{array} { r } { s _ { p a r e n t } = \\mathrm { a r g m a x } _ { s \\in \\mathcal { V } } \\phi ( s ) : = \\bar { r } _ { t } \\left( s \\right) + \\lambda \\sigma _ { t } \\left( s \\right) , } \\end{array}", + "type": "interline_equation", + "image_path": "737553e4156cfdfc16e92c6b0157200463f7e7d79262d5451b425025895670c0.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 204, + 520, + 405, + 534 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 102, + 533, + 508, + 563 + ], + "spans": [ + { + "bbox": [ + 102, + 539, + 172, + 563 + ], + "score": 1.0, + "content": "t t the sequence of", + "type": "text" + }, + { + "bbox": [ + 105, + 533, + 133, + 547 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 143, + 533, + 161, + 547 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 172, + 533, + 397, + 547 + ], + "score": 1.0, + "content": "denote the average reward and variance estimator after", + "type": "text" + }, + { + "bbox": [ + 172, + 547, + 177, + 555 + ], + "score": 0.8, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 539, + 292, + 563 + ], + "score": 1.0, + "content": "selected tree nodes so far as", + "type": "text" + }, + { + "bbox": [ + 293, + 545, + 407, + 558 + ], + "score": 0.91, + "content": "S _ { t } = \\{ s _ { p a r e n t } ^ { 1 } , . . . , s _ { p a r e n t } ^ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 536, + 402, + 544 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "-calls to Expand. Denote", + "type": "text" + }, + { + "bbox": [ + 408, + 539, + 508, + 563 + ], + "score": 1.0, + "content": ", then we can use kernel", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 559, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 207, + 585 + ], + "score": 1.0, + "content": "smoothing estimator for", + "type": "text" + }, + { + "bbox": [ + 207, + 560, + 317, + 582 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\bar { r } _ { t } ( s ) = \\frac { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } k \\left( s ^ { \\prime } , s \\right) r \\left( s ^ { \\prime } \\right) } { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } k \\left( s ^ { \\prime } , s \\right) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 560, + 336, + 585 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 337, + 559, + 444, + 584 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\sigma _ { t } ( s ) = \\sqrt { \\frac { \\log { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } w ( s ^ { \\prime } ) } } { w ( s ) } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 559, + 474, + 583 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 474, + 565, + 505, + 578 + ], + "score": 0.91, + "content": "k ( s ^ { \\prime } , s )", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 103, + 580, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 103, + 580, + 210, + 599 + ], + "score": 1.0, + "content": "is a kernel function and", + "type": "text" + }, + { + "bbox": [ + 210, + 583, + 309, + 596 + ], + "score": 0.9, + "content": "\\begin{array} { r } { w ( s ) = \\sum _ { s ^ { \\prime } \\in S _ { t } } k ( s ^ { \\prime } , s ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 580, + 427, + 599 + ], + "score": 1.0, + "content": ". Other parametrizations of", + "type": "text" + }, + { + "bbox": [ + 428, + 584, + 437, + 594 + ], + "score": 0.85, + "content": "\\bar { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 580, + 457, + 599 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 457, + 585, + 468, + 594 + ], + "score": 0.84, + "content": "\\sigma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 580, + 506, + 599 + ], + "score": 1.0, + "content": "are also", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "possible, such as Gaussian Process parametrization in Appendix C. The average reward exploits more", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "promising states, while the variance promotes exploration towards less frequently visited states; and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 616, + 404, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 374, + 628 + ], + "score": 1.0, + "content": "the exploration versus exploitation is balanced by a tunable weight", + "type": "text" + }, + { + "bbox": [ + 375, + 616, + 400, + 626 + ], + "score": 0.9, + "content": "\\lambda > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 616, + 404, + 628 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 710 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 209, + 647 + ], + "score": 1.0, + "content": "Expanding a reachable", + "type": "text" + }, + { + "bbox": [ + 210, + 634, + 231, + 644 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 631, + 363, + 647 + ], + "score": 1.0, + "content": "in step (ii). Given the selected", + "type": "text" + }, + { + "bbox": [ + 363, + 635, + 394, + 645 + ], + "score": 0.9, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 631, + 506, + 647 + ], + "score": 1.0, + "content": ", we consider expanding a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 643, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 257, + 658 + ], + "score": 1.0, + "content": "reachable state in the neighborhood", + "type": "text" + }, + { + "bbox": [ + 257, + 645, + 288, + 656 + ], + "score": 0.8, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 643, + 506, + 658 + ], + "score": 1.0, + "content": "as an infinite-armed bandit problem. Although one", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 175, + 669 + ], + "score": 1.0, + "content": "can first samples", + "type": "text" + }, + { + "bbox": [ + 176, + 655, + 182, + 664 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 654, + 365, + 669 + ], + "score": 1.0, + "content": "arms uniformly from a neighborhood around", + "type": "text" + }, + { + "bbox": [ + 365, + 656, + 396, + 667 + ], + "score": 0.78, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 654, + 506, + 669 + ], + "score": 1.0, + "content": "and runs a UCB algorithm", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "on the randomly generated finite arms (Wang et al., 2009), such uniform sampler ignores problem", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 104, + 676, + 437, + 689 + ], + "score": 1.0, + "content": "structures, and will lead to unnecessary samples. Instead we will employ a policy", + "type": "text" + }, + { + "bbox": [ + 437, + 676, + 489, + 689 + ], + "score": 0.93, + "content": "\\bar { \\tilde { \\pi } } ^ { * } ( s ^ { \\prime } \\vert s , \\bar { U } ) ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "guidance when generating the candidates. The final choice for next move will be selected from these", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 193, + 711 + ], + "score": 1.0, + "content": "candidates with max", + "type": "text" + }, + { + "bbox": [ + 194, + 699, + 214, + 711 + ], + "score": 0.36, + "content": "\\phi \\left( s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 698, + 460, + 711 + ], + "score": 1.0, + "content": "defined in (3). As explained in more details in Section 4.2,", + "type": "text" + }, + { + "bbox": [ + 461, + 699, + 473, + 709 + ], + "score": 0.81, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "will be", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 114, + 721, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 249, + 734 + ], + "score": 1.0, + "content": "1In the path planning setting, we use", + "type": "text" + }, + { + "bbox": [ + 250, + 721, + 294, + 732 + ], + "score": 0.92, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 720, + 311, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 311, + 721, + 353, + 732 + ], + "score": 0.93, + "content": "\\tilde { \\pi } ^ { * } \\left( a | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "interchangeably as the action is next state.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 284, + 95 + ], + "score": 1.0, + "content": "problem will be extracted based on the tree", + "type": "text" + }, + { + "bbox": [ + 285, + 83, + 294, + 93 + ], + "score": 0.72, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 82, + 506, + 95 + ], + "score": 1.0, + "content": ". The template of tree-based sampling algorithms is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "summarized in Algorithm 1 and illustrated in Fig. 1(c). A key component of the algorithm is the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 103, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 356, + 118 + ], + "score": 1.0, + "content": "Expand operator, which generates the next exploration point", + "type": "text" + }, + { + "bbox": [ + 356, + 106, + 377, + 116 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 103, + 436, + 118 + ], + "score": 1.0, + "content": "and its parent", + "type": "text" + }, + { + "bbox": [ + 436, + 105, + 487, + 116 + ], + "score": 0.85, + "content": "s _ { p a r e n t } \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 103, + 506, + 118 + ], + "score": 1.0, + "content": ". To", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 507, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 276, + 129 + ], + "score": 1.0, + "content": "ensure the feasibility of the solution, the", + "type": "text" + }, + { + "bbox": [ + 276, + 117, + 297, + 127 + ], + "score": 0.9, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 114, + 402, + 129 + ], + "score": 1.0, + "content": "must be reachable from", + "type": "text" + }, + { + "bbox": [ + 402, + 116, + 411, + 126 + ], + "score": 0.74, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 114, + 434, + 129 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 435, + 116, + 494, + 127 + ], + "score": 0.89, + "content": "[ \\bar { s _ { p a r e n t } } , s _ { n e w } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 114, + 507, + 129 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 506, + 138 + ], + "score": 1.0, + "content": "collision-free, which is checked by a collision detection function. As we will discuss in Appendix B,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "by instantiating different Expand operators, we will arrive at many existing algorithms, such as", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 387, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 387, + 161 + ], + "score": 1.0, + "content": "RRT (LaValle, 1998) and EST (Hsu et al., 1997; Phillips et al., 2004).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 82, + 507, + 161 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "One major limitation of existing TSAs is that they solve each problem independently from scratch", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "and ignore past planning experiences in similar environments. We introduce the neural components", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 385, + 200 + ], + "score": 1.0, + "content": "into TSA template to form the learnable planning algorithm family", + "type": "text" + }, + { + "bbox": [ + 386, + 187, + 395, + 198 + ], + "score": 0.75, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 187, + 505, + 200 + ], + "score": 1.0, + "content": ", which can explicitly take", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 501, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 501, + 211 + ], + "score": 1.0, + "content": "advantages of the past successful experiences to bias the Expand towards more promising regions.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 165, + 505, + 211 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 219, + 365, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 366, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 366, + 234 + ], + "score": 1.0, + "content": "4 NEURAL EXPLORATION-EXPLOITATION TREES", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 239, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "Based on the TSA framework, we introduce a learnable neural based Expand operator, which can", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 248, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 104, + 248, + 351, + 264 + ], + "score": 1.0, + "content": "balance between exploration and exploitation, to instantiate", + "type": "text" + }, + { + "bbox": [ + 352, + 250, + 361, + 260 + ], + "score": 0.75, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 248, + 506, + 264 + ], + "score": 1.0, + "content": "in Eq (2). With the self-improving", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 261, + 382, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 382, + 273 + ], + "score": 1.0, + "content": "training, we obtain the meta NEXT algorithm illustrated in Figure 1.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 104, + 238, + 506, + 273 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 279, + 282, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 282, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 282, + 291 + ], + "score": 1.0, + "content": "4.1 GUIDED PROGRESSIVE EXPANSION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 273, + 373 + ], + "lines": [ + { + "bbox": [ + 106, + 294, + 274, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 274, + 307 + ], + "score": 1.0, + "content": "We start with our design of the Expand.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 305, + 273, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 273, + 317 + ], + "score": 1.0, + "content": "We assume having an estimated value", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 316, + 273, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 144, + 329 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 144, + 316, + 180, + 330 + ], + "score": 0.94, + "content": "{ \\tilde { V } } ^ { * } ( s | U )", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 317, + 273, + 329 + ], + "score": 1.0, + "content": ", which stands for the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 329, + 275, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 188, + 340 + ], + "score": 1.0, + "content": "optimal cost from", + "type": "text" + }, + { + "bbox": [ + 189, + 331, + 195, + 339 + ], + "score": 0.68, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 329, + 275, + 340 + ], + "score": 1.0, + "content": "to target in plan-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 339, + 274, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 160, + 352 + ], + "score": 1.0, + "content": "ning problem", + "type": "text" + }, + { + "bbox": [ + 161, + 340, + 170, + 350 + ], + "score": 0.79, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 339, + 222, + 352 + ], + "score": 1.0, + "content": ", and a policy", + "type": "text" + }, + { + "bbox": [ + 222, + 340, + 270, + 352 + ], + "score": 0.9, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 339, + 274, + 352 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 351, + 272, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 263, + 363 + ], + "score": 1.0, + "content": "which generates the promising action", + "type": "text" + }, + { + "bbox": [ + 263, + 352, + 272, + 361 + ], + "score": 0.76, + "content": "s ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 360, + 275, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 151, + 374 + ], + "score": 1.0, + "content": "from state", + "type": "text" + }, + { + "bbox": [ + 151, + 364, + 158, + 371 + ], + "score": 0.67, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 360, + 275, + 374 + ], + "score": 1.0, + "content": ". The concrete parametriza-", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 294, + 275, + 374 + ] + }, + { + "type": "table", + "bbox": [ + 279, + 297, + 504, + 367 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 279, + 297, + 504, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 279, + 297, + 504, + 367 + ], + "spans": [ + { + "bbox": [ + 279, + 297, + 504, + 367 + ], + "score": 0.205, + "html": "
Algorithm 2: NExT :: Expand(T = (V,ε),U)
Sparent ← argmaxs∈vΦ(s) ; Selection
{s1,*..,Sk} id π*(s'sparent,U)； >Candidates
Snew ← argmaxs'∈{s1,.,sk}(s')； DExpansion
4 return Sparent, Snew;
", + "type": "table", + "image_path": "a8c886f8d6118bbe9fbb4275382954ce28939501d27215b83992823a8b9fe93d.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 279, + 297, + 504, + 311.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 279, + 311.0, + 504, + 325.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 279, + 325.0, + 504, + 339.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 279, + 339.0, + 504, + 353.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 279, + 353.0, + 504, + 367.0 + ], + "spans": [], + "index": 27 + } + ] + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 504, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 136, + 386 + ], + "score": 1.0, + "content": "tion of", + "type": "text" + }, + { + "bbox": [ + 137, + 372, + 150, + 384 + ], + "score": 0.89, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 372, + 169, + 386 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 169, + 374, + 181, + 384 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 372, + 505, + 386 + ], + "score": 1.0, + "content": "will be explained in Section 4.2 and learned in Section 4.3. We will use these", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 383, + 478, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 478, + 399 + ], + "score": 1.0, + "content": "functions to construct the learnable Expand with explicit exploration-exploitation balancing.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 372, + 505, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 333, + 415 + ], + "score": 1.0, + "content": "The Expand operator will expand the current search tree", + "type": "text" + }, + { + "bbox": [ + 333, + 402, + 343, + 412 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 401, + 452, + 415 + ], + "score": 1.0, + "content": "by a new neighboring state", + "type": "text" + }, + { + "bbox": [ + 453, + 404, + 474, + 413 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "around", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 411, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 116, + 423 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 411, + 387, + 427 + ], + "score": 1.0, + "content": ". We design the expansion as a two-step procedure: (i) select a state", + "type": "text" + }, + { + "bbox": [ + 387, + 414, + 417, + 425 + ], + "score": 0.81, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 411, + 492, + 427 + ], + "score": 1.0, + "content": "from existing tree", + "type": "text" + }, + { + "bbox": [ + 493, + 413, + 502, + 423 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 411, + 506, + 427 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 423, + 405, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 179, + 437 + ], + "score": 1.0, + "content": "(ii) expand a state", + "type": "text" + }, + { + "bbox": [ + 180, + 425, + 200, + 435 + ], + "score": 0.89, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 423, + 295, + 437 + ], + "score": 1.0, + "content": "in the neighborhood of", + "type": "text" + }, + { + "bbox": [ + 296, + 425, + 326, + 436 + ], + "score": 0.71, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 423, + 405, + 437 + ], + "score": 1.0, + "content": ". More specifically,", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 401, + 506, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 104, + 441, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 104, + 441, + 147, + 455 + ], + "score": 1.0, + "content": "Selecting", + "type": "text" + }, + { + "bbox": [ + 147, + 444, + 177, + 455 + ], + "score": 0.75, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 441, + 201, + 455 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 201, + 443, + 211, + 453 + ], + "score": 0.6, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 441, + 400, + 455 + ], + "score": 1.0, + "content": "in step (i). Consider the negative value function", + "type": "text" + }, + { + "bbox": [ + 400, + 441, + 446, + 455 + ], + "score": 0.93, + "content": "- \\tilde { V } ^ { \\ast } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 441, + 506, + 455 + ], + "score": 1.0, + "content": "as the rewards", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 107, + 453, + 126, + 465 + ], + "score": 0.88, + "content": "r \\left( s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 453, + 505, + 466 + ], + "score": 1.0, + "content": ", step (i) shares some similarity with the multi-armed bandit problem by viewing existing nodes", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 107, + 465, + 133, + 475 + ], + "score": 0.89, + "content": "s \\in \\mathcal V", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "as arms. However, the vanilla UCB algorithm is not directly applicable, since the number", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "of states is increasing as the algorithm proceeds and the value of these adjacent states are naturally", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "correlated. We address this challenge by modeling the correlation explicitly as in contextual bandits.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 497, + 504, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 371, + 510 + ], + "score": 1.0, + "content": "Specifically, we parametrize the UCB of the reward function as", + "type": "text" + }, + { + "bbox": [ + 371, + 498, + 393, + 510 + ], + "score": 0.91, + "content": "\\phi \\left( s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 497, + 495, + 510 + ], + "score": 1.0, + "content": ", and select a node from", + "type": "text" + }, + { + "bbox": [ + 495, + 498, + 504, + 508 + ], + "score": 0.81, + "content": "\\tau", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 508, + 178, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 158, + 521 + ], + "score": 1.0, + "content": "according to", + "type": "text" + }, + { + "bbox": [ + 159, + 509, + 178, + 520 + ], + "score": 0.91, + "content": "\\phi ( s )", + "type": "inline_equation" + } + ], + "index": 39 + } + ], + "index": 36, + "bbox_fs": [ + 104, + 441, + 506, + 521 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 520, + 405, + 534 + ], + "lines": [ + { + "bbox": [ + 204, + 520, + 405, + 534 + ], + "spans": [ + { + "bbox": [ + 204, + 520, + 405, + 534 + ], + "score": 0.86, + "content": "\\begin{array} { r } { s _ { p a r e n t } = \\mathrm { a r g m a x } _ { s \\in \\mathcal { V } } \\phi ( s ) : = \\bar { r } _ { t } \\left( s \\right) + \\lambda \\sigma _ { t } \\left( s \\right) , } \\end{array}", + "type": "interline_equation", + "image_path": "737553e4156cfdfc16e92c6b0157200463f7e7d79262d5451b425025895670c0.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 204, + 520, + 405, + 534 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 102, + 533, + 508, + 563 + ], + "spans": [ + { + "bbox": [ + 102, + 539, + 172, + 563 + ], + "score": 1.0, + "content": "t t the sequence of", + "type": "text" + }, + { + "bbox": [ + 105, + 533, + 133, + 547 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 143, + 533, + 161, + 547 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 172, + 533, + 397, + 547 + ], + "score": 1.0, + "content": "denote the average reward and variance estimator after", + "type": "text" + }, + { + "bbox": [ + 172, + 547, + 177, + 555 + ], + "score": 0.8, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 539, + 292, + 563 + ], + "score": 1.0, + "content": "selected tree nodes so far as", + "type": "text" + }, + { + "bbox": [ + 293, + 545, + 407, + 558 + ], + "score": 0.91, + "content": "S _ { t } = \\{ s _ { p a r e n t } ^ { 1 } , . . . , s _ { p a r e n t } ^ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 536, + 402, + 544 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "-calls to Expand. Denote", + "type": "text" + }, + { + "bbox": [ + 408, + 539, + 508, + 563 + ], + "score": 1.0, + "content": ", then we can use kernel", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 559, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 207, + 585 + ], + "score": 1.0, + "content": "smoothing estimator for", + "type": "text" + }, + { + "bbox": [ + 207, + 560, + 317, + 582 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\bar { r } _ { t } ( s ) = \\frac { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } k \\left( s ^ { \\prime } , s \\right) r \\left( s ^ { \\prime } \\right) } { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } k \\left( s ^ { \\prime } , s \\right) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 560, + 336, + 585 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 337, + 559, + 444, + 584 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\sigma _ { t } ( s ) = \\sqrt { \\frac { \\log { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } w ( s ^ { \\prime } ) } } { w ( s ) } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 559, + 474, + 583 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 474, + 565, + 505, + 578 + ], + "score": 0.91, + "content": "k ( s ^ { \\prime } , s )", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 103, + 580, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 103, + 580, + 210, + 599 + ], + "score": 1.0, + "content": "is a kernel function and", + "type": "text" + }, + { + "bbox": [ + 210, + 583, + 309, + 596 + ], + "score": 0.9, + "content": "\\begin{array} { r } { w ( s ) = \\sum _ { s ^ { \\prime } \\in S _ { t } } k ( s ^ { \\prime } , s ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 580, + 427, + 599 + ], + "score": 1.0, + "content": ". Other parametrizations of", + "type": "text" + }, + { + "bbox": [ + 428, + 584, + 437, + 594 + ], + "score": 0.85, + "content": "\\bar { r } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 580, + 457, + 599 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 457, + 585, + 468, + 594 + ], + "score": 0.84, + "content": "\\sigma _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 580, + 506, + 599 + ], + "score": 1.0, + "content": "are also", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "possible, such as Gaussian Process parametrization in Appendix C. The average reward exploits more", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "promising states, while the variance promotes exploration towards less frequently visited states; and", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 616, + 404, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 374, + 628 + ], + "score": 1.0, + "content": "the exploration versus exploitation is balanced by a tunable weight", + "type": "text" + }, + { + "bbox": [ + 375, + 616, + 400, + 626 + ], + "score": 0.9, + "content": "\\lambda > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 616, + 404, + 628 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 102, + 533, + 508, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 710 + ], + "lines": [ + { + "bbox": [ + 105, + 631, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 209, + 647 + ], + "score": 1.0, + "content": "Expanding a reachable", + "type": "text" + }, + { + "bbox": [ + 210, + 634, + 231, + 644 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 631, + 363, + 647 + ], + "score": 1.0, + "content": "in step (ii). Given the selected", + "type": "text" + }, + { + "bbox": [ + 363, + 635, + 394, + 645 + ], + "score": 0.9, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 631, + 506, + 647 + ], + "score": 1.0, + "content": ", we consider expanding a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 643, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 257, + 658 + ], + "score": 1.0, + "content": "reachable state in the neighborhood", + "type": "text" + }, + { + "bbox": [ + 257, + 645, + 288, + 656 + ], + "score": 0.8, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 643, + 506, + 658 + ], + "score": 1.0, + "content": "as an infinite-armed bandit problem. Although one", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 175, + 669 + ], + "score": 1.0, + "content": "can first samples", + "type": "text" + }, + { + "bbox": [ + 176, + 655, + 182, + 664 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 654, + 365, + 669 + ], + "score": 1.0, + "content": "arms uniformly from a neighborhood around", + "type": "text" + }, + { + "bbox": [ + 365, + 656, + 396, + 667 + ], + "score": 0.78, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 654, + 506, + 669 + ], + "score": 1.0, + "content": "and runs a UCB algorithm", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "on the randomly generated finite arms (Wang et al., 2009), such uniform sampler ignores problem", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 104, + 676, + 437, + 689 + ], + "score": 1.0, + "content": "structures, and will lead to unnecessary samples. Instead we will employ a policy", + "type": "text" + }, + { + "bbox": [ + 437, + 676, + 489, + 689 + ], + "score": 0.93, + "content": "\\bar { \\tilde { \\pi } } ^ { * } ( s ^ { \\prime } \\vert s , \\bar { U } ) ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "guidance when generating the candidates. The final choice for next move will be selected from these", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 193, + 711 + ], + "score": 1.0, + "content": "candidates with max", + "type": "text" + }, + { + "bbox": [ + 194, + 699, + 214, + 711 + ], + "score": 0.36, + "content": "\\phi \\left( s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 698, + 460, + 711 + ], + "score": 1.0, + "content": "defined in (3). As explained in more details in Section 4.2,", + "type": "text" + }, + { + "bbox": [ + 461, + 699, + 473, + 709 + ], + "score": 0.81, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "will be", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 187, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 203 + ], + "score": 1.0, + "content": "trained to mimic previous successful planning experiences across different problems, that is, biasing", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 200, + 372, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 372, + 213 + ], + "score": 1.0, + "content": "the sampling towards the states with higher successful probability.", + "type": "text", + "cross_page": true + } + ], + "index": 16 + } + ], + "index": 50, + "bbox_fs": [ + 104, + 631, + 506, + 711 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 72, + 329, + 180 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 72, + 329, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 72, + 329, + 180 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 329, + 180 + ], + "score": 0.942, + "type": "image", + "image_path": "659fdd9f888e78d83e6532bb1c8c3abf8ec8cc26808c966e6aef6311a6ac402b.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 111, + 72, + 329, + 85.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 85.5, + 329, + 99.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 99.0, + 329, + 112.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 112.5, + 329, + 126.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 126.0, + 329, + 139.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 111, + 139.5, + 329, + 153.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 111, + 153.0, + 329, + 166.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 166.5, + 329, + 180.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 343, + 89, + 505, + 160 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 343, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 343, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "Figure 2: Our neural network model maps", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 343, + 98, + 506, + 111 + ], + "spans": [ + { + "bbox": [ + 343, + 98, + 351, + 111 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 352, + 100, + 361, + 109 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 98, + 506, + 111 + ], + "score": 1.0, + "content": "-link robot from the original planning", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 110, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 343, + 110, + 375, + 121 + ], + "score": 1.0, + "content": "space (a", + "type": "text" + }, + { + "bbox": [ + 376, + 110, + 407, + 120 + ], + "score": 0.91, + "content": "( N + 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 110, + 506, + 121 + ], + "score": 1.0, + "content": "-d configuration space) to a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 119, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 343, + 119, + 506, + 131 + ], + "score": 1.0, + "content": "3d discrete latent planning space in which we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 343, + 129, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 343, + 129, + 506, + 140 + ], + "score": 1.0, + "content": "plan a path using value iteration. The result", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 140, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 344, + 140, + 505, + 149 + ], + "score": 1.0, + "content": "of value iteration is then used as features for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 148, + 476, + 161 + ], + "spans": [ + { + "bbox": [ + 343, + 149, + 376, + 161 + ], + "score": 1.0, + "content": "defining", + "type": "text" + }, + { + "bbox": [ + 376, + 148, + 411, + 160 + ], + "score": 0.9, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 149, + 428, + 161 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 428, + 149, + 473, + 160 + ], + "score": 0.93, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 149, + 476, + 161 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 107, + 189, + 504, + 212 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 505, + 203 + ], + "score": 1.0, + "content": "trained to mimic previous successful planning experiences across different problems, that is, biasing", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 200, + 372, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 372, + 213 + ], + "score": 1.0, + "content": "the sampling towards the states with higher successful probability.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 217, + 504, + 240 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 506, + 231 + ], + "score": 1.0, + "content": "With these detailed step (i) and (ii), we obtain NEXT :: Expand in Algorithm 2 (illustrated in Fig-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 228, + 493, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 352, + 241 + ], + "score": 1.0, + "content": "ure 1(b) and (c)). Plugging it into the TSA in 1, we construct", + "type": "text" + }, + { + "bbox": [ + 352, + 228, + 401, + 240 + ], + "score": 0.87, + "content": "\\tt a l g \\left( \\cdot \\right) \\in \\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 228, + 493, + 241 + ], + "score": 1.0, + "content": "which will be learned.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 506, + 300 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 506, + 257 + ], + "score": 1.0, + "content": "The guided progressive expansion bears similarity to MCTS but deals with high dimensional continu-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 255, + 506, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 419, + 268 + ], + "score": 1.0, + "content": "ous spaces. Moreover, the essential difference lies in the way to select state in", + "type": "text" + }, + { + "bbox": [ + 420, + 256, + 429, + 266 + ], + "score": 0.82, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 255, + 506, + 268 + ], + "score": 1.0, + "content": "for expansion: the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 257, + 280 + ], + "score": 1.0, + "content": "MCTS only expands the leaf states in", + "type": "text" + }, + { + "bbox": [ + 258, + 267, + 267, + 277 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 266, + 506, + 280 + ], + "score": 1.0, + "content": "due to the hierarchical assumption, limiting the exploration", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 277, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 415, + 291 + ], + "score": 1.0, + "content": "ability and incurring extra unnecessary UCB sampling for internal traversal in", + "type": "text" + }, + { + "bbox": [ + 415, + 278, + 424, + 288 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 277, + 505, + 291 + ], + "score": 1.0, + "content": "; while the proposed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 289, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 507, + 302 + ], + "score": 1.0, + "content": "operation enables expansion from each visited state, particularly suitable for path planning problems.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 108, + 305, + 447, + 317 + ], + "lines": [ + { + "bbox": [ + 105, + 304, + 449, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 449, + 318 + ], + "score": 1.0, + "content": "4.2 NEURAL ARCHITECTURE FOR VALUE FUNCTION AND EXPANSION POLICY", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 322, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 363, + 336 + ], + "score": 1.0, + "content": "In this section, we will introduce our neural architectures for", + "type": "text" + }, + { + "bbox": [ + 363, + 321, + 402, + 335 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 320, + 422, + 336 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 422, + 322, + 470, + 335 + ], + "score": 0.92, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 320, + 506, + 336 + ], + "score": 1.0, + "content": "used in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "NEXT :: Expand. The proposed neural architectures can be understood as first embedding the state", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "and problem into a discrete latent representation via an attention-based module in Section 4.2.1, upon", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 356, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 506, + 368 + ], + "score": 1.0, + "content": "which the neuralized value iteration, introduced in Section 4.2.2, is performed to extract features for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 367, + 353, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 141, + 381 + ], + "score": 1.0, + "content": "defining", + "type": "text" + }, + { + "bbox": [ + 142, + 367, + 180, + 380 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 367, + 198, + 381 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 199, + 368, + 246, + 380 + ], + "score": 0.95, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 367, + 353, + 381 + ], + "score": 1.0, + "content": ", as illustrated in Figure 2.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 386, + 298, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 299, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 299, + 398 + ], + "score": 1.0, + "content": "4.2.1 CONFIGURATION SPACE EMBEDDING", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "Our network for embedding high-dimension configuration space into a latent representation is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 370, + 423 + ], + "score": 1.0, + "content": "designed based on an attention mechanism. More specifically, let", + "type": "text" + }, + { + "bbox": [ + 371, + 411, + 383, + 421 + ], + "score": 0.86, + "content": "s ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "denote the workspace in state", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 123, + 435 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 421, + 135, + 432 + ], + "score": 0.88, + "content": "{ \\bf \\bar { \\Phi } } _ { s } h", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 421, + 333, + 435 + ], + "score": 1.0, + "content": "denote the remaining dimensions of the state, i.e.", + "type": "text" + }, + { + "bbox": [ + 333, + 421, + 384, + 434 + ], + "score": 0.92, + "content": "s = \\mathbf { \\bar { \\rho } } ( s ^ { w } , s ^ { h } )", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 421, + 389, + 435 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 389, + 422, + 401, + 432 + ], + "score": 0.82, + "content": "s ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 421, + 420, + 435 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 420, + 421, + 431, + 432 + ], + "score": 0.88, + "content": "s ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "will be embedded", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 433, + 507, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 507, + 445 + ], + "score": 1.0, + "content": "by different sub-neural networks and combined for the final representation, as illustrated in Figure 3.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 377, + 456 + ], + "score": 1.0, + "content": "For simplicity of exposition, we will focus on the 2d workspace, i.e.,", + "type": "text" + }, + { + "bbox": [ + 378, + 443, + 413, + 454 + ], + "score": 0.91, + "content": "s ^ { w } \\in \\mathbb { R } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 443, + 506, + 456 + ], + "score": 1.0, + "content": ". However, our method", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 455, + 238, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 238, + 467 + ], + "score": 1.0, + "content": "applies to 3d workspace as well.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "image", + "bbox": [ + 126, + 475, + 473, + 597 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 475, + 473, + 597 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 126, + 475, + 473, + 597 + ], + "spans": [ + { + "bbox": [ + 126, + 475, + 473, + 597 + ], + "score": 0.967, + "type": "image", + "image_path": "1f9670de482ed991ba97956d7ebba8ef1a8b77c6a4180f100db30cdc84a6e86c.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 126, + 475, + 473, + 515.6666666666666 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 126, + 515.6666666666666, + 473, + 556.3333333333333 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 126, + 556.3333333333333, + 473, + 596.9999999999999 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 601, + 505, + 642 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 293, + 614 + ], + "score": 1.0, + "content": "Figure 3: Attention-based state embedding module.", + "type": "text" + }, + { + "bbox": [ + 293, + 602, + 338, + 613 + ], + "score": 0.93, + "content": "\\boldsymbol { s } ^ { w } = ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 600, + 353, + 614 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 353, + 601, + 381, + 612 + ], + "score": 0.91, + "content": "\\mathbf { z } = s ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 600, + 506, + 614 + ], + "score": 1.0, + "content": ". The upper part is spatial attention,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 225, + 623 + ], + "score": 1.0, + "content": "with the first two channels being", + "type": "text" + }, + { + "bbox": [ + 226, + 614, + 232, + 621 + ], + "score": 0.77, + "content": "_ x", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 612, + 248, + 623 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 249, + 614, + 254, + 622 + ], + "score": 0.83, + "content": "_ y", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 612, + 505, + 623 + ], + "score": 1.0, + "content": ", and the last two channels being constant templates with the row and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 622, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 243, + 633 + ], + "score": 1.0, + "content": "column coordinates, as shown with a", + "type": "text" + }, + { + "bbox": [ + 243, + 623, + 250, + 631 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 622, + 460, + 633 + ], + "score": 1.0, + "content": "set to 3. The bottom module learns the representation for", + "type": "text" + }, + { + "bbox": [ + 460, + 624, + 466, + 631 + ], + "score": 0.65, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 622, + 505, + 633 + ], + "score": 1.0, + "content": ". The final", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 632, + 354, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 354, + 644 + ], + "score": 1.0, + "content": "embedding is obtained by outer-product of these two attention parts.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + } + ], + "index": 39.75 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 505, + 683 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 308, + 662 + ], + "score": 1.0, + "content": "• Spatial attention. The workspace information", + "type": "text" + }, + { + "bbox": [ + 309, + 651, + 321, + 660 + ], + "score": 0.87, + "content": "s ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 648, + 409, + 662 + ], + "score": 1.0, + "content": "will be embedded as", + "type": "text" + }, + { + "bbox": [ + 409, + 649, + 477, + 662 + ], + "score": 0.92, + "content": "\\mu ^ { w } ( s ^ { w } ) \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 648, + 480, + 662 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 481, + 650, + 487, + 660 + ], + "score": 0.68, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "hyperparameter related to map (see remark below). The spatial embedding module (upper part in", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 671, + 329, + 685 + ], + "spans": [ + { + "bbox": [ + 115, + 671, + 218, + 685 + ], + "score": 1.0, + "content": "Figure 3) is composed of", + "type": "text" + }, + { + "bbox": [ + 219, + 672, + 231, + 683 + ], + "score": 0.89, + "content": "k _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 671, + 329, + 685 + ], + "score": 1.0, + "content": "convolution layers, i.e.,", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 684, + 458, + 698 + ], + "lines": [ + { + "bbox": [ + 162, + 684, + 458, + 698 + ], + "spans": [ + { + "bbox": [ + 162, + 684, + 458, + 698 + ], + "score": 0.75, + "content": "\\mu ^ { w } ( s ^ { w } ) = { \\tt s o f t m a x 2 d } ( f _ { k _ { w } } ^ { w } ( s ^ { w } ) ) , \\qquad f _ { i + 1 } ^ { w } ( s ^ { w } ) = { \\tt r e l u } ( \\theta _ { i } ^ { w } \\oplus f _ { i } ^ { w } ( s ^ { w } ) ) ,", + "type": "interline_equation", + "image_path": "317548e75c0ae16c1cddef18d446e697154769250eeb70f2004ecf9a97134885.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 162, + 684, + 458, + 698 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "text", + "bbox": [ + 115, + 698, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 115, + 695, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 115, + 695, + 144, + 710 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 144, + 698, + 156, + 709 + ], + "score": 0.88, + "content": "\\theta _ { i } ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 695, + 294, + 710 + ], + "score": 1.0, + "content": "denotes the convolution kernels,", + "type": "text" + }, + { + "bbox": [ + 294, + 699, + 304, + 708 + ], + "score": 0.82, + "content": "\\oplus", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 695, + 461, + 710 + ], + "score": 1.0, + "content": "denotes the convolution operator and", + "type": "text" + }, + { + "bbox": [ + 461, + 697, + 505, + 709 + ], + "score": 0.91, + "content": "f _ { i } ^ { w } ( s ^ { w } ) \\in", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 117, + 706, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 117, + 708, + 153, + 720 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { d \\times d \\times d _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 706, + 175, + 725 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 175, + 710, + 185, + 720 + ], + "score": 0.85, + "content": "d _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 706, + 285, + 725 + ], + "score": 1.0, + "content": "channels. The first layer", + "type": "text" + }, + { + "bbox": [ + 286, + 710, + 299, + 721 + ], + "score": 0.9, + "content": "f _ { 0 } ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 706, + 397, + 725 + ], + "score": 1.0, + "content": "is designed to represent", + "type": "text" + }, + { + "bbox": [ + 397, + 710, + 410, + 720 + ], + "score": 0.87, + "content": "s ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 706, + 436, + 725 + ], + "score": 1.0, + "content": "into a", + "type": "text" + }, + { + "bbox": [ + 436, + 710, + 477, + 720 + ], + "score": 0.92, + "content": "d \\times d \\times 4", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 706, + 506, + 725 + ], + "score": 1.0, + "content": "tensor", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 113, + 716, + 429, + 736 + ], + "spans": [ + { + "bbox": [ + 113, + 716, + 127, + 736 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 127, + 721, + 290, + 733 + ], + "score": 0.82, + "content": "f _ { 0 } ^ { w } ( s ^ { w } ) _ { i j } = [ s _ { 1 } ^ { w } , s _ { 2 } ^ { w } , i , j ] , i , j = 1 , \\ldots , a", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 716, + 429, + 736 + ], + "score": 1.0, + "content": "0, without any loss of information.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 72, + 329, + 180 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 72, + 329, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 72, + 329, + 180 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 329, + 180 + ], + "score": 0.942, + "type": "image", + "image_path": "659fdd9f888e78d83e6532bb1c8c3abf8ec8cc26808c966e6aef6311a6ac402b.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 111, + 72, + 329, + 85.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 85.5, + 329, + 99.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 99.0, + 329, + 112.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 111, + 112.5, + 329, + 126.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 111, + 126.0, + 329, + 139.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 111, + 139.5, + 329, + 153.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 111, + 153.0, + 329, + 166.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 166.5, + 329, + 180.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 343, + 89, + 505, + 160 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 343, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 343, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "Figure 2: Our neural network model maps", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 343, + 98, + 506, + 111 + ], + "spans": [ + { + "bbox": [ + 343, + 98, + 351, + 111 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 352, + 100, + 361, + 109 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 98, + 506, + 111 + ], + "score": 1.0, + "content": "-link robot from the original planning", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 110, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 343, + 110, + 375, + 121 + ], + "score": 1.0, + "content": "space (a", + "type": "text" + }, + { + "bbox": [ + 376, + 110, + 407, + 120 + ], + "score": 0.91, + "content": "( N + 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 110, + 506, + 121 + ], + "score": 1.0, + "content": "-d configuration space) to a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 119, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 343, + 119, + 506, + 131 + ], + "score": 1.0, + "content": "3d discrete latent planning space in which we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 343, + 129, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 343, + 129, + 506, + 140 + ], + "score": 1.0, + "content": "plan a path using value iteration. The result", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 140, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 344, + 140, + 505, + 149 + ], + "score": 1.0, + "content": "of value iteration is then used as features for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 148, + 476, + 161 + ], + "spans": [ + { + "bbox": [ + 343, + 149, + 376, + 161 + ], + "score": 1.0, + "content": "defining", + "type": "text" + }, + { + "bbox": [ + 376, + 148, + 411, + 160 + ], + "score": 0.9, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 149, + 428, + 161 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 428, + 149, + 473, + 160 + ], + "score": 0.93, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 149, + 476, + 161 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + } + ], + "index": 7.25 + }, + { + "type": "text", + "bbox": [ + 107, + 189, + 504, + 212 + ], + "lines": [], + "index": 15.5, + "bbox_fs": [ + 105, + 187, + 505, + 213 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 217, + 504, + 240 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 506, + 231 + ], + "score": 1.0, + "content": "With these detailed step (i) and (ii), we obtain NEXT :: Expand in Algorithm 2 (illustrated in Fig-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 228, + 493, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 352, + 241 + ], + "score": 1.0, + "content": "ure 1(b) and (c)). Plugging it into the TSA in 1, we construct", + "type": "text" + }, + { + "bbox": [ + 352, + 228, + 401, + 240 + ], + "score": 0.87, + "content": "\\tt a l g \\left( \\cdot \\right) \\in \\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 228, + 493, + 241 + ], + "score": 1.0, + "content": "which will be learned.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 216, + 506, + 241 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 506, + 300 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 506, + 257 + ], + "score": 1.0, + "content": "The guided progressive expansion bears similarity to MCTS but deals with high dimensional continu-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 255, + 506, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 419, + 268 + ], + "score": 1.0, + "content": "ous spaces. Moreover, the essential difference lies in the way to select state in", + "type": "text" + }, + { + "bbox": [ + 420, + 256, + 429, + 266 + ], + "score": 0.82, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 255, + 506, + 268 + ], + "score": 1.0, + "content": "for expansion: the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 266, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 257, + 280 + ], + "score": 1.0, + "content": "MCTS only expands the leaf states in", + "type": "text" + }, + { + "bbox": [ + 258, + 267, + 267, + 277 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 266, + 506, + 280 + ], + "score": 1.0, + "content": "due to the hierarchical assumption, limiting the exploration", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 277, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 415, + 291 + ], + "score": 1.0, + "content": "ability and incurring extra unnecessary UCB sampling for internal traversal in", + "type": "text" + }, + { + "bbox": [ + 415, + 278, + 424, + 288 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 277, + 505, + 291 + ], + "score": 1.0, + "content": "; while the proposed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 289, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 507, + 302 + ], + "score": 1.0, + "content": "operation enables expansion from each visited state, particularly suitable for path planning problems.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 245, + 507, + 302 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 305, + 447, + 317 + ], + "lines": [ + { + "bbox": [ + 105, + 304, + 449, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 449, + 318 + ], + "score": 1.0, + "content": "4.2 NEURAL ARCHITECTURE FOR VALUE FUNCTION AND EXPANSION POLICY", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 322, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 320, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 363, + 336 + ], + "score": 1.0, + "content": "In this section, we will introduce our neural architectures for", + "type": "text" + }, + { + "bbox": [ + 363, + 321, + 402, + 335 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 320, + 422, + 336 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 422, + 322, + 470, + 335 + ], + "score": 0.92, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 320, + 506, + 336 + ], + "score": 1.0, + "content": "used in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "NEXT :: Expand. The proposed neural architectures can be understood as first embedding the state", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "and problem into a discrete latent representation via an attention-based module in Section 4.2.1, upon", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 356, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 506, + 368 + ], + "score": 1.0, + "content": "which the neuralized value iteration, introduced in Section 4.2.2, is performed to extract features for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 367, + 353, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 141, + 381 + ], + "score": 1.0, + "content": "defining", + "type": "text" + }, + { + "bbox": [ + 142, + 367, + 180, + 380 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 367, + 198, + 381 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 199, + 368, + 246, + 380 + ], + "score": 0.95, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 367, + 353, + 381 + ], + "score": 1.0, + "content": ", as illustrated in Figure 2.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 320, + 506, + 381 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 386, + 298, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 299, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 299, + 398 + ], + "score": 1.0, + "content": "4.2.1 CONFIGURATION SPACE EMBEDDING", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "Our network for embedding high-dimension configuration space into a latent representation is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 370, + 423 + ], + "score": 1.0, + "content": "designed based on an attention mechanism. More specifically, let", + "type": "text" + }, + { + "bbox": [ + 371, + 411, + 383, + 421 + ], + "score": 0.86, + "content": "s ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "denote the workspace in state", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 123, + 435 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 421, + 135, + 432 + ], + "score": 0.88, + "content": "{ \\bf \\bar { \\Phi } } _ { s } h", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 421, + 333, + 435 + ], + "score": 1.0, + "content": "denote the remaining dimensions of the state, i.e.", + "type": "text" + }, + { + "bbox": [ + 333, + 421, + 384, + 434 + ], + "score": 0.92, + "content": "s = \\mathbf { \\bar { \\rho } } ( s ^ { w } , s ^ { h } )", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 421, + 389, + 435 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 389, + 422, + 401, + 432 + ], + "score": 0.82, + "content": "s ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 421, + 420, + 435 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 420, + 421, + 431, + 432 + ], + "score": 0.88, + "content": "s ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "will be embedded", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 433, + 507, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 507, + 445 + ], + "score": 1.0, + "content": "by different sub-neural networks and combined for the final representation, as illustrated in Figure 3.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 377, + 456 + ], + "score": 1.0, + "content": "For simplicity of exposition, we will focus on the 2d workspace, i.e.,", + "type": "text" + }, + { + "bbox": [ + 378, + 443, + 413, + 454 + ], + "score": 0.91, + "content": "s ^ { w } \\in \\mathbb { R } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 443, + 506, + 456 + ], + "score": 1.0, + "content": ". However, our method", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 455, + 238, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 238, + 467 + ], + "score": 1.0, + "content": "applies to 3d workspace as well.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 399, + 507, + 467 + ] + }, + { + "type": "image", + "bbox": [ + 126, + 475, + 473, + 597 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 475, + 473, + 597 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 126, + 475, + 473, + 597 + ], + "spans": [ + { + "bbox": [ + 126, + 475, + 473, + 597 + ], + "score": 0.967, + "type": "image", + "image_path": "1f9670de482ed991ba97956d7ebba8ef1a8b77c6a4180f100db30cdc84a6e86c.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 126, + 475, + 473, + 515.6666666666666 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 126, + 515.6666666666666, + 473, + 556.3333333333333 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 126, + 556.3333333333333, + 473, + 596.9999999999999 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 601, + 505, + 642 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 600, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 293, + 614 + ], + "score": 1.0, + "content": "Figure 3: Attention-based state embedding module.", + "type": "text" + }, + { + "bbox": [ + 293, + 602, + 338, + 613 + ], + "score": 0.93, + "content": "\\boldsymbol { s } ^ { w } = ( x , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 600, + 353, + 614 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 353, + 601, + 381, + 612 + ], + "score": 0.91, + "content": "\\mathbf { z } = s ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 600, + 506, + 614 + ], + "score": 1.0, + "content": ". The upper part is spatial attention,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 225, + 623 + ], + "score": 1.0, + "content": "with the first two channels being", + "type": "text" + }, + { + "bbox": [ + 226, + 614, + 232, + 621 + ], + "score": 0.77, + "content": "_ x", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 612, + 248, + 623 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 249, + 614, + 254, + 622 + ], + "score": 0.83, + "content": "_ y", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 612, + 505, + 623 + ], + "score": 1.0, + "content": ", and the last two channels being constant templates with the row and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 622, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 243, + 633 + ], + "score": 1.0, + "content": "column coordinates, as shown with a", + "type": "text" + }, + { + "bbox": [ + 243, + 623, + 250, + 631 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 622, + 460, + 633 + ], + "score": 1.0, + "content": "set to 3. The bottom module learns the representation for", + "type": "text" + }, + { + "bbox": [ + 460, + 624, + 466, + 631 + ], + "score": 0.65, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 622, + 505, + 633 + ], + "score": 1.0, + "content": ". The final", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 632, + 354, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 354, + 644 + ], + "score": 1.0, + "content": "embedding is obtained by outer-product of these two attention parts.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + } + ], + "index": 39.75 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 505, + 683 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 308, + 662 + ], + "score": 1.0, + "content": "• Spatial attention. The workspace information", + "type": "text" + }, + { + "bbox": [ + 309, + 651, + 321, + 660 + ], + "score": 0.87, + "content": "s ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 648, + 409, + 662 + ], + "score": 1.0, + "content": "will be embedded as", + "type": "text" + }, + { + "bbox": [ + 409, + 649, + 477, + 662 + ], + "score": 0.92, + "content": "\\mu ^ { w } ( s ^ { w } ) \\in \\mathbb { R } ^ { d \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 648, + 480, + 662 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 481, + 650, + 487, + 660 + ], + "score": 0.68, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "hyperparameter related to map (see remark below). The spatial embedding module (upper part in", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 671, + 329, + 685 + ], + "spans": [ + { + "bbox": [ + 115, + 671, + 218, + 685 + ], + "score": 1.0, + "content": "Figure 3) is composed of", + "type": "text" + }, + { + "bbox": [ + 219, + 672, + 231, + 683 + ], + "score": 0.89, + "content": "k _ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 671, + 329, + 685 + ], + "score": 1.0, + "content": "convolution layers, i.e.,", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 648, + 506, + 685 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 162, + 684, + 458, + 698 + ], + "lines": [ + { + "bbox": [ + 162, + 684, + 458, + 698 + ], + "spans": [ + { + "bbox": [ + 162, + 684, + 458, + 698 + ], + "score": 0.75, + "content": "\\mu ^ { w } ( s ^ { w } ) = { \\tt s o f t m a x 2 d } ( f _ { k _ { w } } ^ { w } ( s ^ { w } ) ) , \\qquad f _ { i + 1 } ^ { w } ( s ^ { w } ) = { \\tt r e l u } ( \\theta _ { i } ^ { w } \\oplus f _ { i } ^ { w } ( s ^ { w } ) ) ,", + "type": "interline_equation", + "image_path": "317548e75c0ae16c1cddef18d446e697154769250eeb70f2004ecf9a97134885.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 162, + 684, + 458, + 698 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "text", + "bbox": [ + 115, + 698, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 115, + 695, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 115, + 695, + 144, + 710 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 144, + 698, + 156, + 709 + ], + "score": 0.88, + "content": "\\theta _ { i } ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 695, + 294, + 710 + ], + "score": 1.0, + "content": "denotes the convolution kernels,", + "type": "text" + }, + { + "bbox": [ + 294, + 699, + 304, + 708 + ], + "score": 0.82, + "content": "\\oplus", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 695, + 461, + 710 + ], + "score": 1.0, + "content": "denotes the convolution operator and", + "type": "text" + }, + { + "bbox": [ + 461, + 697, + 505, + 709 + ], + "score": 0.91, + "content": "f _ { i } ^ { w } ( s ^ { w } ) \\in", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 117, + 706, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 117, + 708, + 153, + 720 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { d \\times d \\times d _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 706, + 175, + 725 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 175, + 710, + 185, + 720 + ], + "score": 0.85, + "content": "d _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 706, + 285, + 725 + ], + "score": 1.0, + "content": "channels. The first layer", + "type": "text" + }, + { + "bbox": [ + 286, + 710, + 299, + 721 + ], + "score": 0.9, + "content": "f _ { 0 } ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 706, + 397, + 725 + ], + "score": 1.0, + "content": "is designed to represent", + "type": "text" + }, + { + "bbox": [ + 397, + 710, + 410, + 720 + ], + "score": 0.87, + "content": "s ^ { w }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 706, + 436, + 725 + ], + "score": 1.0, + "content": "into a", + "type": "text" + }, + { + "bbox": [ + 436, + 710, + 477, + 720 + ], + "score": 0.92, + "content": "d \\times d \\times 4", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 706, + 506, + 725 + ], + "score": 1.0, + "content": "tensor", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 113, + 716, + 429, + 736 + ], + "spans": [ + { + "bbox": [ + 113, + 716, + 127, + 736 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 127, + 721, + 290, + 733 + ], + "score": 0.82, + "content": "f _ { 0 } ^ { w } ( s ^ { w } ) _ { i j } = [ s _ { 1 } ^ { w } , s _ { 2 } ^ { w } , i , j ] , i , j = 1 , \\ldots , a", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 716, + 429, + 736 + ], + "score": 1.0, + "content": "0, without any loss of information.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49, + "bbox_fs": [ + 113, + 695, + 506, + 736 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 77, + 483, + 179 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 77, + 483, + 179 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 77, + 483, + 179 + ], + "spans": [ + { + "bbox": [ + 127, + 77, + 483, + 179 + ], + "score": 0.97, + "type": "image", + "image_path": "e690ebd6ddc08b67ba7f0c4c102d2ba25dbff64c677153964eecb7a40240d491.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 77, + 483, + 111.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 111.0, + 483, + 145.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 145.0, + 483, + 179.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 186, + 505, + 230 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "Figure 4: Overall model architecture. Current and goal states are embedded through attention module. Then the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 194, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 358, + 213 + ], + "score": 1.0, + "content": "embedding of the goal state is concatenated with the map to produce", + "type": "text" + }, + { + "bbox": [ + 358, + 197, + 378, + 208 + ], + "score": 0.9, + "content": "\\nu ^ { * ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 194, + 394, + 213 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 395, + 196, + 403, + 208 + ], + "score": 0.85, + "content": "\\bar { \\tilde { \\boldsymbol { R } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 194, + 506, + 213 + ], + "score": 1.0, + "content": "as the input to the planning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "module. The output of the planning module is aggregated with the embedding of the current state to produce", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 218, + 241, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 133, + 231 + ], + "score": 1.0, + "content": "feature", + "type": "text" + }, + { + "bbox": [ + 134, + 219, + 153, + 230 + ], + "score": 0.91, + "content": "\\psi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 219, + 198, + 231 + ], + "score": 1.0, + "content": "for defining", + "type": "text" + }, + { + "bbox": [ + 198, + 218, + 210, + 228 + ], + "score": 0.86, + "content": "\\tilde { V } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 219, + 227, + 231 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 227, + 219, + 238, + 228 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 219, + 241, + 231 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 504, + 257 + ], + "lines": [ + { + "bbox": [ + 107, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 107, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "• Configuration attention. The remaining configuration state information will be embedded as", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 245, + 421, + 259 + ], + "spans": [ + { + "bbox": [ + 117, + 246, + 146, + 258 + ], + "score": 0.92, + "content": "\\mu ^ { h } ( s ^ { h } )", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 245, + 181, + 259 + ], + "score": 1.0, + "content": "through", + "type": "text" + }, + { + "bbox": [ + 181, + 247, + 192, + 257 + ], + "score": 0.88, + "content": "k _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 245, + 421, + 259 + ], + "score": 1.0, + "content": "fully-connected layers (bottom-left part in Figure 3), i.e.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 258, + 452, + 274 + ], + "lines": [ + { + "bbox": [ + 167, + 258, + 452, + 274 + ], + "spans": [ + { + "bbox": [ + 167, + 258, + 452, + 274 + ], + "score": 0.71, + "content": "\\mu ^ { h } ( s ^ { h } ) = \\mathsf { s o f t m a x } ( f _ { k _ { h } } ^ { h } ( s ^ { h } ) ) , \\qquad f _ { i + 1 } ^ { h } ( s ^ { h } ) = \\mathtt { r e l u } ( \\theta _ { i } ^ { h } f _ { i } ^ { h } ( s ^ { h } ) + b _ { i } ) ,", + "type": "interline_equation", + "image_path": "b9f52b29aa72f5ea2e0438622f43cd032a335db9b8db958f90e8a3b48576656e.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 167, + 258, + 452, + 274 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 117, + 275, + 275, + 288 + ], + "lines": [ + { + "bbox": [ + 115, + 273, + 276, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 273, + 143, + 289 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 143, + 275, + 201, + 288 + ], + "score": 0.91, + "content": "\\mu ^ { h } ( s ^ { h } ) \\in \\mathbb { R } ^ { d _ { a } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 273, + 219, + 289 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 220, + 275, + 272, + 288 + ], + "score": 0.86, + "content": "f _ { 0 } ^ { h } ( s ^ { h } ) = s ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 273, + 276, + 289 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 203, + 311 + ], + "score": 1.0, + "content": "The final representation", + "type": "text" + }, + { + "bbox": [ + 203, + 297, + 227, + 308 + ], + "score": 0.91, + "content": "\\mu _ { \\boldsymbol { \\theta } } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 295, + 356, + 311 + ], + "score": 1.0, + "content": "will be obtained by multiplying", + "type": "text" + }, + { + "bbox": [ + 356, + 297, + 388, + 309 + ], + "score": 0.93, + "content": "\\mu ^ { w } ( s ^ { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 295, + 410, + 311 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 410, + 297, + 440, + 309 + ], + "score": 0.93, + "content": "\\mu ^ { h } ( s ^ { h } )", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 295, + 506, + 311 + ], + "score": 1.0, + "content": "element-wisely,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 305, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 238, + 320 + ], + "score": 0.88, + "content": "\\mu _ { \\boldsymbol \\theta } ( s ) _ { i j l } = \\bar { \\mu } ^ { w } ( s ^ { w } ) _ { i j } \\cdot \\bar { \\mu } ^ { h } ( \\bar { s } ^ { h } ) _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 305, + 291, + 324 + ], + "score": 1.0, + "content": ", which is a", + "type": "text" + }, + { + "bbox": [ + 292, + 308, + 342, + 320 + ], + "score": 0.92, + "content": "d \\times d \\times d _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 305, + 456, + 324 + ], + "score": 1.0, + "content": "tensor attention map with", + "type": "text" + }, + { + "bbox": [ + 457, + 308, + 505, + 321 + ], + "score": 0.9, + "content": "\\mu _ { \\theta } ( s ) _ { i j l } \\geqslant", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 318, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 318, + 142, + 334 + ], + "score": 1.0, + "content": "0, and", + "type": "text" + }, + { + "bbox": [ + 142, + 320, + 218, + 333 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\sum _ { i j l } \\mu _ { \\theta } ( s ) _ { i j l } = 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 318, + 466, + 334 + ], + "score": 1.0, + "content": "(bottom-right part in Figure 3). Intuitively, one can think of", + "type": "text" + }, + { + "bbox": [ + 466, + 321, + 477, + 330 + ], + "score": 0.88, + "content": "d _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 318, + 506, + 334 + ], + "score": 1.0, + "content": "as the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 345, + 346 + ], + "score": 1.0, + "content": "level of the learned discretization of the configuration space", + "type": "text" + }, + { + "bbox": [ + 345, + 332, + 356, + 343 + ], + "score": 0.87, + "content": "s ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 331, + 430, + 346 + ], + "score": 1.0, + "content": ", and the entries in", + "type": "text" + }, + { + "bbox": [ + 438, + 331, + 505, + 346 + ], + "score": 1.0, + "content": "softly assign the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 101, + 338, + 509, + 365 + ], + "spans": [ + { + "bbox": [ + 101, + 338, + 153, + 365 + ], + "score": 1.0, + "content": "actual state", + "type": "text" + }, + { + "bbox": [ + 153, + 348, + 159, + 356 + ], + "score": 0.76, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 338, + 279, + 365 + ], + "score": 1.0, + "content": "to these discretized locations.", + "type": "text" + }, + { + "bbox": [ + 279, + 344, + 407, + 358 + ], + "score": 0.76, + "content": "\\theta : = ( \\{ \\theta _ { i } ^ { w } \\} _ { i = 0 } ^ { k _ { w } - 1 } , \\{ \\theta _ { i } ^ { h } , b _ { i } \\} _ { i = 0 } ^ { k _ { h } - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 338, + 509, + 365 + ], + "score": 1.0, + "content": "are the parameters to be", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 372, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 506, + 386 + ], + "score": 1.0, + "content": "Remark (different map size): To process map using convolution neural networks, we resize it to a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 384, + 486, + 397 + ], + "spans": [ + { + "bbox": [ + 107, + 385, + 130, + 395 + ], + "score": 0.89, + "content": "d \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 384, + 395, + 397 + ], + "score": 1.0, + "content": "image, with the same size as the spatial attention in Eq (4), where", + "type": "text" + }, + { + "bbox": [ + 395, + 385, + 402, + 394 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 384, + 486, + 397 + ], + "score": 1.0, + "content": "is a hyperparameter.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 416, + 258, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 415, + 260, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 260, + 428 + ], + "score": 1.0, + "content": "4.2.2 NEURAL VALUE ITERATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 506, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 436, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 506, + 449 + ], + "score": 1.0, + "content": "We then apply neuralized value iteration on top of the configuration space embedding to extract", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 448, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 505, + 459 + ], + "score": 1.0, + "content": "further planning features (Tamar et al., 2016; Lee et al., 2018). Specifically, we first produce the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 152, + 472 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 153, + 459, + 191, + 471 + ], + "score": 0.92, + "content": "\\mu _ { \\theta } \\big ( s _ { g o a l } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 458, + 313, + 472 + ], + "score": 1.0, + "content": "of the center of the goal region", + "type": "text" + }, + { + "bbox": [ + 313, + 460, + 334, + 470 + ], + "score": 0.89, + "content": "s _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 458, + 385, + 472 + ], + "score": 1.0, + "content": ". We execute", + "type": "text" + }, + { + "bbox": [ + 385, + 459, + 394, + 468 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "steps of neuralized Bellman", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 469, + 362, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 362, + 483 + ], + "score": 1.0, + "content": "updates (planning module in Figure 4) in the embedding space,", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 131, + 481, + 479, + 502 + ], + "lines": [ + { + "bbox": [ + 131, + 481, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 479, + 502 + ], + "score": 0.88, + "content": "\\nu ^ { * ( t ) } = \\operatorname* { m i n } \\left( W _ { 1 } \\oplus \\left[ \\nu ^ { * ( t - 1 ) } , \\tilde { R } \\right] \\right) , \\quad \\mathrm { w i t h } \\quad \\left( \\nu ^ { * ( 0 ) } , \\tilde { R } \\right) = \\sigma \\left( W _ { 0 } \\oplus \\left[ \\mu _ { \\theta } \\big ( s _ { g o a l } \\big ) , \\mathtt { m a p l } \\right) \\right) ,", + "type": "interline_equation", + "image_path": "031e99ba81d68aaa517f693a0ebc251fd46e6b8dba4e8c0d754e254cbb1c0902.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 131, + 481, + 479, + 502 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 151, + 517 + ], + "score": 1.0, + "content": "and obtain", + "type": "text" + }, + { + "bbox": [ + 151, + 503, + 234, + 514 + ], + "score": 0.91, + "content": "\\nu ^ { * ( T ) } \\in \\mathbb { R } ^ { d \\times d \\times d _ { a } \\times p }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 501, + 260, + 517 + ], + "score": 1.0, + "content": ". Both", + "type": "text" + }, + { + "bbox": [ + 261, + 504, + 276, + 515 + ], + "score": 0.88, + "content": "W _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 501, + 294, + 517 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 294, + 504, + 309, + 515 + ], + "score": 0.9, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 501, + 506, + 517 + ], + "score": 1.0, + "content": "are 3d convolution kernels, min implements the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 514, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 260, + 530 + ], + "score": 1.0, + "content": "pooling across channels. Accordingly,", + "type": "text" + }, + { + "bbox": [ + 261, + 515, + 283, + 527 + ], + "score": 0.9, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 514, + 506, + 530 + ], + "score": 1.0, + "content": "now can be understood as a latent representation of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 527, + 453, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 165, + 542 + ], + "score": 1.0, + "content": "value function", + "type": "text" + }, + { + "bbox": [ + 166, + 527, + 190, + 541 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 527, + 369, + 542 + ], + "score": 1.0, + "content": "in learned embedding space for the problem", + "type": "text" + }, + { + "bbox": [ + 369, + 529, + 378, + 538 + ], + "score": 0.81, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 527, + 399, + 542 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 400, + 530, + 421, + 541 + ], + "score": 0.89, + "content": "s _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 527, + 453, + 542 + ], + "score": 1.0, + "content": "in map.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 506, + 587 + ], + "lines": [ + { + "bbox": [ + 104, + 545, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 292, + 560 + ], + "score": 1.0, + "content": "To define the value function for particular state", + "type": "text" + }, + { + "bbox": [ + 293, + 550, + 298, + 558 + ], + "score": 0.64, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 545, + 318, + 560 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 319, + 546, + 357, + 560 + ], + "score": 0.92, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 545, + 479, + 560 + ], + "score": 1.0, + "content": ", from the latent representation", + "type": "text" + }, + { + "bbox": [ + 479, + 546, + 502, + 558 + ], + "score": 0.9, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 545, + 505, + 560 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 557, + 507, + 573 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 405, + 573 + ], + "score": 1.0, + "content": "we first construct another attention model between the embedding of state", + "type": "text" + }, + { + "bbox": [ + 405, + 562, + 411, + 570 + ], + "score": 0.74, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 557, + 436, + 573 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 437, + 560, + 461, + 572 + ], + "score": 0.91, + "content": "\\mu _ { \\boldsymbol { \\theta } } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 557, + 479, + 573 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 479, + 559, + 502, + 570 + ], + "score": 0.89, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 557, + 507, + 573 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 570, + 397, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 123, + 591 + ], + "score": 1.0, + "content": "i.e.,", + "type": "text" + }, + { + "bbox": [ + 124, + 571, + 249, + 588 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\psi \\left( s \\right) _ { k } = \\sum _ { i j l } \\nu _ { i j l k } ^ { * ( T ) } \\cdot \\mu _ { \\theta } \\left( s \\right) _ { i j l } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 570, + 267, + 591 + ], + "score": 1.0, + "content": ", for", + "type": "text" + }, + { + "bbox": [ + 267, + 574, + 319, + 586 + ], + "score": 0.91, + "content": "k = 1 , \\hdots , p", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 570, + 397, + 591 + ], + "score": 1.0, + "content": ". Finally we define", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 589, + 445, + 603 + ], + "lines": [ + { + "bbox": [ + 164, + 589, + 445, + 603 + ], + "spans": [ + { + "bbox": [ + 164, + 589, + 445, + 603 + ], + "score": 0.7, + "content": "\\tilde { V } ^ { \\ast } \\left( s | U \\right) = h _ { W _ { 2 } } \\left( \\psi ( s ) \\right) , \\quad \\mathrm { a n d } \\quad \\tilde { \\pi } ^ { \\ast } \\left( s ^ { \\prime } | s , U \\right) = \\mathcal { N } \\left( h _ { W _ { 3 } } \\left( \\psi ( s ) \\right) , \\sigma ^ { 2 } \\right)", + "type": "interline_equation", + "image_path": "0b18fc86bde7bbd5b1268f8408e733fa5a87511bda64f1095ef6cc2258a4ab25.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 164, + 589, + 445, + 603 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 133, + 615 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 604, + 151, + 615 + ], + "score": 0.9, + "content": "h _ { W _ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 603, + 169, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 170, + 603, + 188, + 614 + ], + "score": 0.9, + "content": "h _ { W _ { 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 603, + 386, + 615 + ], + "score": 1.0, + "content": "are fully connected dense layers with parameters", + "type": "text" + }, + { + "bbox": [ + 387, + 604, + 402, + 614 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 603, + 420, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 420, + 603, + 435, + 614 + ], + "score": 0.92, + "content": "W _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "respectively, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 613, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 107, + 614, + 188, + 626 + ], + "score": 0.9, + "content": "\\mathcal { N } ( h _ { W _ { 3 } } \\left( \\psi ( s ) \\right) , \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 613, + 351, + 628 + ], + "score": 1.0, + "content": "is a Gaussian distribution with variance", + "type": "text" + }, + { + "bbox": [ + 351, + 614, + 362, + 624 + ], + "score": 0.87, + "content": "\\sigma ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 613, + 506, + 628 + ], + "score": 1.0, + "content": ". Note that we also parametrize the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 624, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 624, + 134, + 640 + ], + "score": 1.0, + "content": "policy", + "type": "text" + }, + { + "bbox": [ + 135, + 627, + 182, + 638 + ], + "score": 0.9, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 624, + 271, + 640 + ], + "score": 1.0, + "content": "using the embedding", + "type": "text" + }, + { + "bbox": [ + 272, + 625, + 295, + 636 + ], + "score": 0.9, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 624, + 506, + 640 + ], + "score": 1.0, + "content": ", since the policy is connected to the value function", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 121, + 650 + ], + "score": 1.0, + "content": "via", + "type": "text" + }, + { + "bbox": [ + 122, + 637, + 322, + 650 + ], + "score": 0.62, + "content": "\\begin{array} { r } { \\pi ^ { * } \\left( s ^ { \\prime } | s , U \\right) = \\mathrm { a r g m i n } _ { s ^ { \\prime } \\in S } c \\left( \\left[ s , s ^ { \\prime } \\right] \\right) + V ^ { * } \\left( s ^ { \\prime } | U \\right. } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "). It should also be emphasized that in our", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 648, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 104, + 648, + 252, + 664 + ], + "score": 1.0, + "content": "parametrization, the calculation of", + "type": "text" + }, + { + "bbox": [ + 253, + 650, + 276, + 660 + ], + "score": 0.89, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 648, + 355, + 664 + ], + "score": 1.0, + "content": "only relies on the", + "type": "text" + }, + { + "bbox": [ + 355, + 650, + 396, + 663 + ], + "score": 0.92, + "content": "\\mu _ { \\theta } \\left( s _ { g o a l } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 648, + 506, + 664 + ], + "score": 1.0, + "content": ", which can be reused for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 151, + 676 + ], + "score": 1.0, + "content": "evaluating", + "type": "text" + }, + { + "bbox": [ + 151, + 662, + 190, + 675 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 662, + 208, + 676 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 209, + 663, + 257, + 675 + ], + "score": 0.94, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 662, + 316, + 676 + ], + "score": 1.0, + "content": "over different", + "type": "text" + }, + { + "bbox": [ + 317, + 666, + 322, + 673 + ], + "score": 0.74, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 662, + 506, + 676 + ], + "score": 1.0, + "content": ", saving computational resources. Using this", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 674, + 323, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 203, + 687 + ], + "score": 1.0, + "content": "trick the algorithm runs", + "type": "text" + }, + { + "bbox": [ + 203, + 675, + 248, + 685 + ], + "score": 0.87, + "content": "1 0 \\times - 1 0 0 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 674, + 323, + 687 + ], + "score": 1.0, + "content": "faster empirically.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 690, + 505, + 736 + ], + "lines": [ + { + "bbox": [ + 104, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 104, + 690, + 255, + 704 + ], + "score": 1.0, + "content": "The overall model architecture in", + "type": "text" + }, + { + "bbox": [ + 255, + 691, + 284, + 703 + ], + "score": 0.87, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 690, + 477, + 704 + ], + "score": 1.0, + "content": "is illustrated in Figure 4. The parameters", + "type": "text" + }, + { + "bbox": [ + 477, + 691, + 505, + 702 + ], + "score": 0.83, + "content": "W =", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 702, + 507, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 192, + 714 + ], + "score": 0.9, + "content": "( W _ { 0 } , W _ { 1 } , W _ { 2 } , W _ { 3 } , \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 702, + 507, + 715 + ], + "score": 1.0, + "content": "will be learned together by our meta self-improving learning. For the de-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 712, + 505, + 726 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 505, + 726 + ], + "score": 1.0, + "content": "tails of the parameterization and the size of convolution kernels in our implementation, please refer", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 724, + 222, + 736 + ], + "spans": [ + { + "bbox": [ + 105, + 724, + 222, + 736 + ], + "score": 1.0, + "content": "to Figure 12 in Appendix D.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 761 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 77, + 483, + 179 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 77, + 483, + 179 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 77, + 483, + 179 + ], + "spans": [ + { + "bbox": [ + 127, + 77, + 483, + 179 + ], + "score": 0.97, + "type": "image", + "image_path": "e690ebd6ddc08b67ba7f0c4c102d2ba25dbff64c677153964eecb7a40240d491.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 127, + 77, + 483, + 111.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 127, + 111.0, + 483, + 145.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 127, + 145.0, + 483, + 179.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 186, + 505, + 230 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 186, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 199 + ], + "score": 1.0, + "content": "Figure 4: Overall model architecture. Current and goal states are embedded through attention module. Then the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 194, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 358, + 213 + ], + "score": 1.0, + "content": "embedding of the goal state is concatenated with the map to produce", + "type": "text" + }, + { + "bbox": [ + 358, + 197, + 378, + 208 + ], + "score": 0.9, + "content": "\\nu ^ { * ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 194, + 394, + 213 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 395, + 196, + 403, + 208 + ], + "score": 0.85, + "content": "\\bar { \\tilde { \\boldsymbol { R } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 194, + 506, + 213 + ], + "score": 1.0, + "content": "as the input to the planning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "module. The output of the planning module is aggregated with the embedding of the current state to produce", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 218, + 241, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 133, + 231 + ], + "score": 1.0, + "content": "feature", + "type": "text" + }, + { + "bbox": [ + 134, + 219, + 153, + 230 + ], + "score": 0.91, + "content": "\\psi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 219, + 198, + 231 + ], + "score": 1.0, + "content": "for defining", + "type": "text" + }, + { + "bbox": [ + 198, + 218, + 210, + 228 + ], + "score": 0.86, + "content": "\\tilde { V } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 219, + 227, + 231 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 227, + 219, + 238, + 228 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 219, + 241, + 231 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 504, + 257 + ], + "lines": [ + { + "bbox": [ + 107, + 235, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 107, + 235, + 506, + 248 + ], + "score": 1.0, + "content": "• Configuration attention. The remaining configuration state information will be embedded as", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 245, + 421, + 259 + ], + "spans": [ + { + "bbox": [ + 117, + 246, + 146, + 258 + ], + "score": 0.92, + "content": "\\mu ^ { h } ( s ^ { h } )", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 245, + 181, + 259 + ], + "score": 1.0, + "content": "through", + "type": "text" + }, + { + "bbox": [ + 181, + 247, + 192, + 257 + ], + "score": 0.88, + "content": "k _ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 245, + 421, + 259 + ], + "score": 1.0, + "content": "fully-connected layers (bottom-left part in Figure 3), i.e.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 107, + 235, + 506, + 259 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 167, + 258, + 452, + 274 + ], + "lines": [ + { + "bbox": [ + 167, + 258, + 452, + 274 + ], + "spans": [ + { + "bbox": [ + 167, + 258, + 452, + 274 + ], + "score": 0.71, + "content": "\\mu ^ { h } ( s ^ { h } ) = \\mathsf { s o f t m a x } ( f _ { k _ { h } } ^ { h } ( s ^ { h } ) ) , \\qquad f _ { i + 1 } ^ { h } ( s ^ { h } ) = \\mathtt { r e l u } ( \\theta _ { i } ^ { h } f _ { i } ^ { h } ( s ^ { h } ) + b _ { i } ) ,", + "type": "interline_equation", + "image_path": "b9f52b29aa72f5ea2e0438622f43cd032a335db9b8db958f90e8a3b48576656e.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 167, + 258, + 452, + 274 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 117, + 275, + 275, + 288 + ], + "lines": [ + { + "bbox": [ + 115, + 273, + 276, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 273, + 143, + 289 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 143, + 275, + 201, + 288 + ], + "score": 0.91, + "content": "\\mu ^ { h } ( s ^ { h } ) \\in \\mathbb { R } ^ { d _ { a } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 273, + 219, + 289 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 220, + 275, + 272, + 288 + ], + "score": 0.86, + "content": "f _ { 0 } ^ { h } ( s ^ { h } ) = s ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 273, + 276, + 289 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 115, + 273, + 276, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 203, + 311 + ], + "score": 1.0, + "content": "The final representation", + "type": "text" + }, + { + "bbox": [ + 203, + 297, + 227, + 308 + ], + "score": 0.91, + "content": "\\mu _ { \\boldsymbol { \\theta } } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 295, + 356, + 311 + ], + "score": 1.0, + "content": "will be obtained by multiplying", + "type": "text" + }, + { + "bbox": [ + 356, + 297, + 388, + 309 + ], + "score": 0.93, + "content": "\\mu ^ { w } ( s ^ { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 295, + 410, + 311 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 410, + 297, + 440, + 309 + ], + "score": 0.93, + "content": "\\mu ^ { h } ( s ^ { h } )", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 295, + 506, + 311 + ], + "score": 1.0, + "content": "element-wisely,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 305, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 238, + 320 + ], + "score": 0.88, + "content": "\\mu _ { \\boldsymbol \\theta } ( s ) _ { i j l } = \\bar { \\mu } ^ { w } ( s ^ { w } ) _ { i j } \\cdot \\bar { \\mu } ^ { h } ( \\bar { s } ^ { h } ) _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 305, + 291, + 324 + ], + "score": 1.0, + "content": ", which is a", + "type": "text" + }, + { + "bbox": [ + 292, + 308, + 342, + 320 + ], + "score": 0.92, + "content": "d \\times d \\times d _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 305, + 456, + 324 + ], + "score": 1.0, + "content": "tensor attention map with", + "type": "text" + }, + { + "bbox": [ + 457, + 308, + 505, + 321 + ], + "score": 0.9, + "content": "\\mu _ { \\theta } ( s ) _ { i j l } \\geqslant", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 318, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 318, + 142, + 334 + ], + "score": 1.0, + "content": "0, and", + "type": "text" + }, + { + "bbox": [ + 142, + 320, + 218, + 333 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\sum _ { i j l } \\mu _ { \\theta } ( s ) _ { i j l } = 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 318, + 466, + 334 + ], + "score": 1.0, + "content": "(bottom-right part in Figure 3). Intuitively, one can think of", + "type": "text" + }, + { + "bbox": [ + 466, + 321, + 477, + 330 + ], + "score": 0.88, + "content": "d _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 318, + 506, + 334 + ], + "score": 1.0, + "content": "as the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 331, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 345, + 346 + ], + "score": 1.0, + "content": "level of the learned discretization of the configuration space", + "type": "text" + }, + { + "bbox": [ + 345, + 332, + 356, + 343 + ], + "score": 0.87, + "content": "s ^ { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 331, + 430, + 346 + ], + "score": 1.0, + "content": ", and the entries in", + "type": "text" + }, + { + "bbox": [ + 438, + 331, + 505, + 346 + ], + "score": 1.0, + "content": "softly assign the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 101, + 338, + 509, + 365 + ], + "spans": [ + { + "bbox": [ + 101, + 338, + 153, + 365 + ], + "score": 1.0, + "content": "actual state", + "type": "text" + }, + { + "bbox": [ + 153, + 348, + 159, + 356 + ], + "score": 0.76, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 338, + 279, + 365 + ], + "score": 1.0, + "content": "to these discretized locations.", + "type": "text" + }, + { + "bbox": [ + 279, + 344, + 407, + 358 + ], + "score": 0.76, + "content": "\\theta : = ( \\{ \\theta _ { i } ^ { w } \\} _ { i = 0 } ^ { k _ { w } - 1 } , \\{ \\theta _ { i } ^ { h } , b _ { i } \\} _ { i = 0 } ^ { k _ { h } - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 338, + 509, + 365 + ], + "score": 1.0, + "content": "are the parameters to be", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 101, + 295, + 509, + 365 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 372, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 506, + 386 + ], + "score": 1.0, + "content": "Remark (different map size): To process map using convolution neural networks, we resize it to a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 384, + 486, + 397 + ], + "spans": [ + { + "bbox": [ + 107, + 385, + 130, + 395 + ], + "score": 0.89, + "content": "d \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 384, + 395, + 397 + ], + "score": 1.0, + "content": "image, with the same size as the spatial attention in Eq (4), where", + "type": "text" + }, + { + "bbox": [ + 395, + 385, + 402, + 394 + ], + "score": 0.79, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 384, + 486, + 397 + ], + "score": 1.0, + "content": "is a hyperparameter.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 372, + 506, + 397 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 416, + 258, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 415, + 260, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 260, + 428 + ], + "score": 1.0, + "content": "4.2.2 NEURAL VALUE ITERATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 506, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 436, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 506, + 449 + ], + "score": 1.0, + "content": "We then apply neuralized value iteration on top of the configuration space embedding to extract", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 448, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 505, + 459 + ], + "score": 1.0, + "content": "further planning features (Tamar et al., 2016; Lee et al., 2018). Specifically, we first produce the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 152, + 472 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 153, + 459, + 191, + 471 + ], + "score": 0.92, + "content": "\\mu _ { \\theta } \\big ( s _ { g o a l } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 458, + 313, + 472 + ], + "score": 1.0, + "content": "of the center of the goal region", + "type": "text" + }, + { + "bbox": [ + 313, + 460, + 334, + 470 + ], + "score": 0.89, + "content": "s _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 458, + 385, + 472 + ], + "score": 1.0, + "content": ". We execute", + "type": "text" + }, + { + "bbox": [ + 385, + 459, + 394, + 468 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "steps of neuralized Bellman", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 469, + 362, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 362, + 483 + ], + "score": 1.0, + "content": "updates (planning module in Figure 4) in the embedding space,", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 436, + 506, + 483 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 131, + 481, + 479, + 502 + ], + "lines": [ + { + "bbox": [ + 131, + 481, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 479, + 502 + ], + "score": 0.88, + "content": "\\nu ^ { * ( t ) } = \\operatorname* { m i n } \\left( W _ { 1 } \\oplus \\left[ \\nu ^ { * ( t - 1 ) } , \\tilde { R } \\right] \\right) , \\quad \\mathrm { w i t h } \\quad \\left( \\nu ^ { * ( 0 ) } , \\tilde { R } \\right) = \\sigma \\left( W _ { 0 } \\oplus \\left[ \\mu _ { \\theta } \\big ( s _ { g o a l } \\big ) , \\mathtt { m a p l } \\right) \\right) ,", + "type": "interline_equation", + "image_path": "031e99ba81d68aaa517f693a0ebc251fd46e6b8dba4e8c0d754e254cbb1c0902.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 131, + 481, + 479, + 502 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 151, + 517 + ], + "score": 1.0, + "content": "and obtain", + "type": "text" + }, + { + "bbox": [ + 151, + 503, + 234, + 514 + ], + "score": 0.91, + "content": "\\nu ^ { * ( T ) } \\in \\mathbb { R } ^ { d \\times d \\times d _ { a } \\times p }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 501, + 260, + 517 + ], + "score": 1.0, + "content": ". Both", + "type": "text" + }, + { + "bbox": [ + 261, + 504, + 276, + 515 + ], + "score": 0.88, + "content": "W _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 501, + 294, + 517 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 294, + 504, + 309, + 515 + ], + "score": 0.9, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 501, + 506, + 517 + ], + "score": 1.0, + "content": "are 3d convolution kernels, min implements the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 514, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 260, + 530 + ], + "score": 1.0, + "content": "pooling across channels. Accordingly,", + "type": "text" + }, + { + "bbox": [ + 261, + 515, + 283, + 527 + ], + "score": 0.9, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 514, + 506, + 530 + ], + "score": 1.0, + "content": "now can be understood as a latent representation of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 527, + 453, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 165, + 542 + ], + "score": 1.0, + "content": "value function", + "type": "text" + }, + { + "bbox": [ + 166, + 527, + 190, + 541 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 527, + 369, + 542 + ], + "score": 1.0, + "content": "in learned embedding space for the problem", + "type": "text" + }, + { + "bbox": [ + 369, + 529, + 378, + 538 + ], + "score": 0.81, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 527, + 399, + 542 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 400, + 530, + 421, + 541 + ], + "score": 0.89, + "content": "s _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 527, + 453, + 542 + ], + "score": 1.0, + "content": "in map.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 501, + 506, + 542 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 506, + 587 + ], + "lines": [ + { + "bbox": [ + 104, + 545, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 104, + 545, + 292, + 560 + ], + "score": 1.0, + "content": "To define the value function for particular state", + "type": "text" + }, + { + "bbox": [ + 293, + 550, + 298, + 558 + ], + "score": 0.64, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 545, + 318, + 560 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 319, + 546, + 357, + 560 + ], + "score": 0.92, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 545, + 479, + 560 + ], + "score": 1.0, + "content": ", from the latent representation", + "type": "text" + }, + { + "bbox": [ + 479, + 546, + 502, + 558 + ], + "score": 0.9, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 545, + 505, + 560 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 557, + 507, + 573 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 405, + 573 + ], + "score": 1.0, + "content": "we first construct another attention model between the embedding of state", + "type": "text" + }, + { + "bbox": [ + 405, + 562, + 411, + 570 + ], + "score": 0.74, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 557, + 436, + 573 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 437, + 560, + 461, + 572 + ], + "score": 0.91, + "content": "\\mu _ { \\boldsymbol { \\theta } } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 557, + 479, + 573 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 479, + 559, + 502, + 570 + ], + "score": 0.89, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 557, + 507, + 573 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 570, + 397, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 570, + 123, + 591 + ], + "score": 1.0, + "content": "i.e.,", + "type": "text" + }, + { + "bbox": [ + 124, + 571, + 249, + 588 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\psi \\left( s \\right) _ { k } = \\sum _ { i j l } \\nu _ { i j l k } ^ { * ( T ) } \\cdot \\mu _ { \\theta } \\left( s \\right) _ { i j l } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 570, + 267, + 591 + ], + "score": 1.0, + "content": ", for", + "type": "text" + }, + { + "bbox": [ + 267, + 574, + 319, + 586 + ], + "score": 0.91, + "content": "k = 1 , \\hdots , p", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 570, + 397, + 591 + ], + "score": 1.0, + "content": ". Finally we define", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 545, + 507, + 591 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 589, + 445, + 603 + ], + "lines": [ + { + "bbox": [ + 164, + 589, + 445, + 603 + ], + "spans": [ + { + "bbox": [ + 164, + 589, + 445, + 603 + ], + "score": 0.7, + "content": "\\tilde { V } ^ { \\ast } \\left( s | U \\right) = h _ { W _ { 2 } } \\left( \\psi ( s ) \\right) , \\quad \\mathrm { a n d } \\quad \\tilde { \\pi } ^ { \\ast } \\left( s ^ { \\prime } | s , U \\right) = \\mathcal { N } \\left( h _ { W _ { 3 } } \\left( \\psi ( s ) \\right) , \\sigma ^ { 2 } \\right)", + "type": "interline_equation", + "image_path": "0b18fc86bde7bbd5b1268f8408e733fa5a87511bda64f1095ef6cc2258a4ab25.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 164, + 589, + 445, + 603 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 133, + 615 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 604, + 151, + 615 + ], + "score": 0.9, + "content": "h _ { W _ { 2 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 603, + 169, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 170, + 603, + 188, + 614 + ], + "score": 0.9, + "content": "h _ { W _ { 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 603, + 386, + 615 + ], + "score": 1.0, + "content": "are fully connected dense layers with parameters", + "type": "text" + }, + { + "bbox": [ + 387, + 604, + 402, + 614 + ], + "score": 0.89, + "content": "W _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 603, + 420, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 420, + 603, + 435, + 614 + ], + "score": 0.92, + "content": "W _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "respectively, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 613, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 107, + 614, + 188, + 626 + ], + "score": 0.9, + "content": "\\mathcal { N } ( h _ { W _ { 3 } } \\left( \\psi ( s ) \\right) , \\sigma ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 613, + 351, + 628 + ], + "score": 1.0, + "content": "is a Gaussian distribution with variance", + "type": "text" + }, + { + "bbox": [ + 351, + 614, + 362, + 624 + ], + "score": 0.87, + "content": "\\sigma ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 613, + 506, + 628 + ], + "score": 1.0, + "content": ". Note that we also parametrize the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 624, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 624, + 134, + 640 + ], + "score": 1.0, + "content": "policy", + "type": "text" + }, + { + "bbox": [ + 135, + 627, + 182, + 638 + ], + "score": 0.9, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 624, + 271, + 640 + ], + "score": 1.0, + "content": "using the embedding", + "type": "text" + }, + { + "bbox": [ + 272, + 625, + 295, + 636 + ], + "score": 0.9, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 624, + 506, + 640 + ], + "score": 1.0, + "content": ", since the policy is connected to the value function", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 121, + 650 + ], + "score": 1.0, + "content": "via", + "type": "text" + }, + { + "bbox": [ + 122, + 637, + 322, + 650 + ], + "score": 0.62, + "content": "\\begin{array} { r } { \\pi ^ { * } \\left( s ^ { \\prime } | s , U \\right) = \\mathrm { a r g m i n } _ { s ^ { \\prime } \\in S } c \\left( \\left[ s , s ^ { \\prime } \\right] \\right) + V ^ { * } \\left( s ^ { \\prime } | U \\right. } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "). It should also be emphasized that in our", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 648, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 104, + 648, + 252, + 664 + ], + "score": 1.0, + "content": "parametrization, the calculation of", + "type": "text" + }, + { + "bbox": [ + 253, + 650, + 276, + 660 + ], + "score": 0.89, + "content": "\\nu ^ { * ( T ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 648, + 355, + 664 + ], + "score": 1.0, + "content": "only relies on the", + "type": "text" + }, + { + "bbox": [ + 355, + 650, + 396, + 663 + ], + "score": 0.92, + "content": "\\mu _ { \\theta } \\left( s _ { g o a l } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 648, + 506, + 664 + ], + "score": 1.0, + "content": ", which can be reused for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 151, + 676 + ], + "score": 1.0, + "content": "evaluating", + "type": "text" + }, + { + "bbox": [ + 151, + 662, + 190, + 675 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 662, + 208, + 676 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 209, + 663, + 257, + 675 + ], + "score": 0.94, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 662, + 316, + 676 + ], + "score": 1.0, + "content": "over different", + "type": "text" + }, + { + "bbox": [ + 317, + 666, + 322, + 673 + ], + "score": 0.74, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 662, + 506, + 676 + ], + "score": 1.0, + "content": ", saving computational resources. Using this", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 674, + 323, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 203, + 687 + ], + "score": 1.0, + "content": "trick the algorithm runs", + "type": "text" + }, + { + "bbox": [ + 203, + 675, + 248, + 685 + ], + "score": 0.87, + "content": "1 0 \\times - 1 0 0 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 674, + 323, + 687 + ], + "score": 1.0, + "content": "faster empirically.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 104, + 603, + 506, + 687 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 690, + 505, + 736 + ], + "lines": [ + { + "bbox": [ + 104, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 104, + 690, + 255, + 704 + ], + "score": 1.0, + "content": "The overall model architecture in", + "type": "text" + }, + { + "bbox": [ + 255, + 691, + 284, + 703 + ], + "score": 0.87, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 690, + 477, + 704 + ], + "score": 1.0, + "content": "is illustrated in Figure 4. The parameters", + "type": "text" + }, + { + "bbox": [ + 477, + 691, + 505, + 702 + ], + "score": 0.83, + "content": "W =", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 702, + 507, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 192, + 714 + ], + "score": 0.9, + "content": "( W _ { 0 } , W _ { 1 } , W _ { 2 } , W _ { 3 } , \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 702, + 507, + 715 + ], + "score": 1.0, + "content": "will be learned together by our meta self-improving learning. For the de-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 712, + 505, + 726 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 505, + 726 + ], + "score": 1.0, + "content": "tails of the parameterization and the size of convolution kernels in our implementation, please refer", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 724, + 222, + 736 + ], + "spans": [ + { + "bbox": [ + 105, + 724, + 222, + 736 + ], + "score": 1.0, + "content": "to Figure 12 in Appendix D.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 690, + 507, + 736 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 280, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 281, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 281, + 95 + ], + "score": 1.0, + "content": "4.3 META SELF-IMPROVING LEARNING", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 97, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 96, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 218, + 112 + ], + "score": 1.0, + "content": "The learning of the planner", + "type": "text" + }, + { + "bbox": [ + 218, + 97, + 247, + 110 + ], + "score": 0.82, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 96, + 399, + 112 + ], + "score": 1.0, + "content": "reduces to learning the parameters in", + "type": "text" + }, + { + "bbox": [ + 400, + 96, + 438, + 110 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 96, + 456, + 112 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 456, + 97, + 505, + 110 + ], + "score": 0.91, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 108, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 506, + 123 + ], + "score": 1.0, + "content": "and is carried out while planning experiences accumulate. We do not have an explicit training", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 119, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 132 + ], + "score": 1.0, + "content": "and testing phase separation. Particularly, we use a mixture of RRT :: Expand and NEXT :: Expand", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 504, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 174, + 143 + ], + "score": 1.0, + "content": "with probability", + "type": "text" + }, + { + "bbox": [ + 175, + 133, + 181, + 141 + ], + "score": 0.65, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 131, + 200, + 143 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 200, + 131, + 223, + 142 + ], + "score": 0.81, + "content": "1 - \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 131, + 504, + 143 + ], + "score": 1.0, + "content": ", respectively, inside the TSA framework in Algorithm 1. The RRT∗", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 309, + 155 + ], + "score": 1.0, + "content": "postprocessing step is used in the template. The", + "type": "text" + }, + { + "bbox": [ + 310, + 145, + 316, + 153 + ], + "score": 0.68, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 142, + 433, + 155 + ], + "score": 1.0, + "content": "is set to be 1 initially since", + "type": "text" + }, + { + "bbox": [ + 433, + 141, + 472, + 155 + ], + "score": 0.93, + "content": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "are not", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 314, + 167 + ], + "score": 1.0, + "content": "well-trained, and thus, the algorithm behaves like", + "type": "text" + }, + { + "bbox": [ + 315, + 154, + 339, + 165 + ], + "score": 0.56, + "content": "\\mathbf { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 154, + 498, + 167 + ], + "score": 1.0, + "content": ". As the training proceeds, we anneal", + "type": "text" + }, + { + "bbox": [ + 498, + 156, + 504, + 164 + ], + "score": 0.7, + "content": "\\epsilon", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 303, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 303, + 177 + ], + "score": 1.0, + "content": "gradually as the sampler becomes more efficient.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 181, + 505, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 153, + 196 + ], + "score": 1.0, + "content": "The dataset", + "type": "text" + }, + { + "bbox": [ + 154, + 181, + 237, + 195 + ], + "score": 0.92, + "content": "\\mathcal { D } _ { n } = \\{ ( \\mathcal { T } _ { j } , U _ { j } ) \\} _ { j = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 180, + 267, + 196 + ], + "score": 1.0, + "content": "for the", + "type": "text" + }, + { + "bbox": [ + 267, + 183, + 274, + 192 + ], + "score": 0.8, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 180, + 506, + 196 + ], + "score": 1.0, + "content": "-th training epoch is collected from the previous successful", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 504, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 494, + 204 + ], + "score": 1.0, + "content": "planning experiences across multiple random problems. We fix the size of dataset and update", + "type": "text" + }, + { + "bbox": [ + 495, + 193, + 504, + 203 + ], + "score": 0.81, + "content": "\\mathcal { D }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "score": 1.0, + "content": "in the same way as experience reply buffer (Lin, 1992; Schaul et al., 2015). For an experience", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 211, + 507, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 163, + 227 + ], + "score": 0.92, + "content": "( { \\mathcal { T } } , U ) \\in { \\mathcal { D } } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 211, + 331, + 231 + ], + "score": 1.0, + "content": ", we can reconstruct the successful path", + "type": "text" + }, + { + "bbox": [ + 332, + 214, + 364, + 227 + ], + "score": 0.94, + "content": "\\{ s ^ { i } \\} _ { i = 1 } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 211, + 452, + 231 + ], + "score": 1.0, + "content": "from the search tree", + "type": "text" + }, + { + "bbox": [ + 453, + 215, + 463, + 225 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 211, + 468, + 231 + ], + "score": 1.0, + "content": "(", + "type": "text" + }, + { + "bbox": [ + 468, + 216, + 478, + 225 + ], + "score": 0.72, + "content": "\\mathrm { ~ m ~ }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 211, + 507, + 231 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 102, + 226, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 102, + 231, + 154, + 256 + ], + "score": 1.0, + "content": "region, i.e.,", + "type": "text" + }, + { + "bbox": [ + 105, + 226, + 308, + 239 + ], + "score": 1.0, + "content": "number of segments), and the value of each state", + "type": "text" + }, + { + "bbox": [ + 154, + 236, + 257, + 251 + ], + "score": 0.91, + "content": "\\begin{array} { r } { y ^ { i } : = \\sum _ { l = i } ^ { m - 1 } c ( [ s ^ { l } , s ^ { l + 1 } ] ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 231, + 300, + 256 + ], + "score": 1.0, + "content": ". We learn", + "type": "text" + }, + { + "bbox": [ + 300, + 237, + 338, + 250 + ], + "score": 0.94, + "content": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 226, + 317, + 236 + ], + "score": 0.87, + "content": "s ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "in the path will be the sum of cost to the goal", + "type": "text" + }, + { + "bbox": [ + 339, + 231, + 441, + 256 + ], + "score": 1.0, + "content": "by optimizing objective", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 109, + 252, + 489, + 286 + ], + "lines": [ + { + "bbox": [ + 109, + 252, + 489, + 286 + ], + "spans": [ + { + "bbox": [ + 109, + 252, + 489, + 286 + ], + "score": 0.88, + "content": "\\operatorname* { m i n } _ { W } \\sum _ { ( { \\cal T } , { \\cal U } ) \\in { \\mathcal D } _ { n } } \\ell ( \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } ; { \\cal T } , { \\cal U } ) : = - \\sum _ { { \\mathcal D } _ { n } } \\sum _ { i = 1 } ^ { m - 1 } \\log \\tilde { \\pi } ^ { * } \\left( s ^ { i + 1 } | s ^ { i } \\right) + \\sum _ { i = 1 } ^ { m } ( \\tilde { V } ^ { * } \\left( s ^ { i } \\right) - y ^ { i } ) _ { 2 } ^ { 2 } + \\lambda \\left\\| { \\cal W } \\right\\| ^ { 2 } .", + "type": "interline_equation", + "image_path": "73a67d0d8a8b20fc2ce7c14d50c96eb2453ad6e3db342b8b427574a485703b12.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 109, + 252, + 489, + 263.3333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 109, + 263.3333333333333, + 489, + 274.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 109, + 274.66666666666663, + 489, + 285.99999999999994 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 292, + 294, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 207, + 306 + ], + "score": 1.0, + "content": "The loss (6) pushes the", + "type": "text" + }, + { + "bbox": [ + 208, + 291, + 222, + 303 + ], + "score": 0.87, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 291, + 242, + 306 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 242, + 293, + 254, + 303 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 291, + 294, + 306 + ], + "score": 1.0, + "content": "to chase", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 303, + 293, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 293, + 316 + ], + "score": 1.0, + "content": "the successful trajectories, providing effective", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 315, + 293, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 153, + 327 + ], + "score": 1.0, + "content": "guidance in", + "type": "text" + }, + { + "bbox": [ + 154, + 315, + 182, + 327 + ], + "score": 0.88, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 315, + 293, + 327 + ], + "score": 1.0, + "content": "for searching, and therefore", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 325, + 293, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 293, + 337 + ], + "score": 1.0, + "content": "leading to efficient searching procedure with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 336, + 293, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 293, + 348 + ], + "score": 1.0, + "content": "less sample complexity and better solution. On", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 347, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 294, + 360 + ], + "score": 1.0, + "content": "one hand, the value function and policy esti-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 137, + 372 + ], + "score": 1.0, + "content": "mation", + "type": "text" + }, + { + "bbox": [ + 138, + 358, + 176, + 372 + ], + "score": 0.94, + "content": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 358, + 294, + 372 + ], + "score": 1.0, + "content": "is improved based upon the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 371, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 294, + 383 + ], + "score": 1.0, + "content": "successful outcomes from NEXT itself on pre-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 381, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 294, + 393 + ], + "score": 1.0, + "content": "vious problems. On the other hand, the updated", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 392, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 107, + 392, + 145, + 406 + ], + "score": 0.93, + "content": "\\{ \\tilde { V } ^ { * } , \\bar { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 392, + 294, + 407 + ], + "score": 1.0, + "content": "will be applied in the next epoch to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 405, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 294, + 417 + ], + "score": 1.0, + "content": "improve the performance. Therefore, the train-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 416, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 295, + 428 + ], + "score": 1.0, + "content": "ing is named as Meta Self-Improving Learn-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 427, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 294, + 438 + ], + "score": 1.0, + "content": "ing (MSIL). Since all the trajectories we col-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 437, + 293, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 293, + 450 + ], + "score": 1.0, + "content": "lected for learning are feasible, the reachability", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 294, + 462 + ], + "score": 1.0, + "content": "of the proposed samples is enforced implicitly", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 460, + 252, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 252, + 472 + ], + "score": 1.0, + "content": "via imitating these successful paths.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 477, + 301, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 293, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 293, + 488 + ], + "score": 1.0, + "content": "By putting every components together into the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 486, + 302, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 302, + 500 + ], + "score": 1.0, + "content": "10 learning to plan framework in Eq (2), the over-", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 306, + 292, + 485, + 304 + ], + "lines": [ + { + "bbox": [ + 303, + 290, + 487, + 308 + ], + "spans": [ + { + "bbox": [ + 303, + 290, + 487, + 308 + ], + "score": 1.0, + "content": "Algorithm 3: Meta Self-Improving Learning", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 296, + 308, + 493, + 495 + ], + "lines": [ + { + "bbox": [ + 295, + 306, + 390, + 319 + ], + "spans": [ + { + "bbox": [ + 295, + 306, + 372, + 319 + ], + "score": 1.0, + "content": "1 Initialize dataset", + "type": "text" + }, + { + "bbox": [ + 373, + 307, + 386, + 318 + ], + "score": 0.85, + "content": "\\mathcal { D } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 306, + 390, + 319 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 296, + 318, + 410, + 329 + ], + "spans": [ + { + "bbox": [ + 296, + 320, + 302, + 327 + ], + "score": 0.48, + "content": "^ 2", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 318, + 345, + 329 + ], + "score": 1.0, + "content": "for epoch", + "type": "text" + }, + { + "bbox": [ + 346, + 318, + 374, + 328 + ], + "score": 0.89, + "content": "n \\gets 1", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 318, + 385, + 329 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 385, + 318, + 396, + 328 + ], + "score": 0.69, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 318, + 410, + 329 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 296, + 329, + 446, + 340 + ], + "spans": [ + { + "bbox": [ + 296, + 331, + 302, + 338 + ], + "score": 0.37, + "content": "^ 3", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 329, + 432, + 340 + ], + "score": 1.0, + "content": "Sample a planning problem", + "type": "text" + }, + { + "bbox": [ + 432, + 329, + 441, + 339 + ], + "score": 0.77, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 329, + 446, + 340 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 297, + 339, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 297, + 343, + 303, + 349 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 320, + 340, + 375, + 352 + ], + "score": 0.88, + "content": "\\mathcal { T } \\mathrm { T S A } ( U )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 339, + 397, + 352 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 397, + 340, + 458, + 352 + ], + "score": 0.9, + "content": "\\epsilon \\sim \\mathcal { U } n i f [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 339, + 479, + 352 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 325, + 352, + 493, + 365 + ], + "spans": [ + { + "bbox": [ + 325, + 354, + 330, + 363 + ], + "score": 0.54, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 352, + 391, + 365 + ], + "score": 1.0, + "content": "·RRT :: Expand", + "type": "text" + }, + { + "bbox": [ + 391, + 353, + 423, + 365 + ], + "score": 0.34, + "content": "+ ( 1 - \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 352, + 493, + 365 + ], + "score": 1.0, + "content": "·NEXT :: Expand;", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 324, + 365, + 408, + 378 + ], + "spans": [ + { + "bbox": [ + 324, + 365, + 408, + 378 + ], + "score": 1.0, + "content": "Postprocessing with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 325, + 376, + 420, + 389 + ], + "spans": [ + { + "bbox": [ + 325, + 377, + 346, + 387 + ], + "score": 0.26, + "content": "\\mathbb { R R T } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 376, + 420, + 389 + ], + "score": 1.0, + "content": ":: Postprocess;", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 297, + 387, + 475, + 399 + ], + "spans": [ + { + "bbox": [ + 297, + 390, + 302, + 398 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 320, + 387, + 421, + 399 + ], + "score": 0.9, + "content": "{ \\mathcal { D } } _ { n } \\gets { \\mathcal { D } } _ { n - 1 } \\bar { \\cup } \\{ ( { \\mathcal { T } } , U ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 387, + 475, + 399 + ], + "score": 1.0, + "content": "if successful", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 324, + 397, + 372, + 412 + ], + "spans": [ + { + "bbox": [ + 324, + 397, + 342, + 412 + ], + "score": 1.0, + "content": "else", + "type": "text" + }, + { + "bbox": [ + 343, + 399, + 367, + 410 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { n - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 397, + 372, + 412 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 297, + 409, + 396, + 421 + ], + "spans": [ + { + "bbox": [ + 297, + 413, + 303, + 420 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 319, + 409, + 335, + 421 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 335, + 410, + 361, + 421 + ], + "score": 0.9, + "content": "j 0", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 409, + 373, + 421 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 373, + 410, + 381, + 419 + ], + "score": 0.72, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 409, + 396, + 421 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 297, + 420, + 487, + 451 + ], + "spans": [ + { + "bbox": [ + 297, + 424, + 303, + 443 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 331, + 420, + 368, + 451 + ], + "score": 1.0, + "content": "Sample Reconst", + "type": "text" + }, + { + "bbox": [ + 368, + 420, + 401, + 433 + ], + "score": 0.92, + "content": "( \\tau _ { j } , U _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 420, + 425, + 451 + ], + "score": 1.0, + "content": "from -optim", + "type": "text" + }, + { + "bbox": [ + 425, + 421, + 439, + 432 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 420, + 455, + 451 + ], + "score": 1.0, + "content": ";path", + "type": "text" + }, + { + "bbox": [ + 456, + 432, + 487, + 445 + ], + "score": 0.9, + "content": "\\{ s ^ { i } \\} _ { i = 1 } ^ { m }", + "type": "inline_equation" + } + ], + "index": 45 + }, + { + "bbox": [ + 339, + 444, + 478, + 457 + ], + "spans": [ + { + "bbox": [ + 339, + 444, + 462, + 457 + ], + "score": 1.0, + "content": "and the cost of paths based on", + "type": "text" + }, + { + "bbox": [ + 462, + 445, + 473, + 456 + ], + "score": 0.87, + "content": "\\tau _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 444, + 478, + 457 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 297, + 455, + 413, + 466 + ], + "spans": [ + { + "bbox": [ + 297, + 458, + 303, + 464 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 335, + 455, + 413, + 466 + ], + "score": 1.0, + "content": "Update parameters", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 339, + 466, + 484, + 481 + ], + "spans": [ + { + "bbox": [ + 339, + 466, + 484, + 481 + ], + "score": 0.71, + "content": "W W - \\eta \\nabla _ { W } \\ell ( \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } ; T _ { j } , U _ { j } ) ;", + "type": "inline_equation", + "image_path": "a45aa838211278ffba9b38ea7cf900286b1f3f0324cacf5bec3f0c31b3e16a52.jpg" + } + ], + "index": 48 + }, + { + "bbox": [ + 319, + 483, + 430, + 496 + ], + "spans": [ + { + "bbox": [ + 319, + 483, + 351, + 496 + ], + "score": 1.0, + "content": "Anneal", + "type": "text" + }, + { + "bbox": [ + 351, + 484, + 380, + 494 + ], + "score": 0.55, + "content": "\\epsilon = \\alpha \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 483, + 384, + 496 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 384, + 483, + 426, + 495 + ], + "score": 0.82, + "content": "\\alpha \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 483, + 430, + 496 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 108, + 499, + 392, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 391, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 391, + 512 + ], + "score": 1.0, + "content": "all procedure is summarized in Algorithm 3 and illustrated in Figure 1.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "title", + "bbox": [ + 107, + 522, + 200, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 201, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 201, + 536 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "In this section, we evaluate the proposed NEXT empirically on different planning tasks in a variety of", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "score": 1.0, + "content": "environments. Comparing to the existing planning algorithms, NEXT achieves the state-of-the-art", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "performances, in terms of both success rate and the quality of the found solutions. We further", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "demonstrate the power of the proposed two components by the corresponding ablation study. We also", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 585, + 459, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 459, + 598 + ], + "score": 1.0, + "content": "include a case study on a real-world robot arm control problem at the end of the section.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 54 + }, + { + "type": "title", + "bbox": [ + 107, + 609, + 218, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 220, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 220, + 621 + ], + "score": 1.0, + "content": "5.1 EXPERIMENT SETUP", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 57 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "Benchmark environments. We designed four benchmark tasks to demonstrate the effectiveness of", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "our algorithm for high-dimensional planning. The first three involve planning in a 2d workspace with", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "a 2 DoF (degrees of freedom) point robot, a 3 DoF stick robot and a 5 DoF snake robot, respectively.", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "The last one involves planning a 7 DoF spacecraft in a 3d workspace. For all problems in each", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "benchmark task, the workspace maps were randomly generated from a fixed distribution; the initial", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 444, + 694 + ], + "score": 1.0, + "content": "and goal states were sampled uniformly randomly in the free space; the cost function", + "type": "text" + }, + { + "bbox": [ + 444, + 682, + 462, + 694 + ], + "score": 0.87, + "content": "c \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "was set as", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 693, + 483, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 483, + 705 + ], + "score": 1.0, + "content": "the sum of the Euclidean path length and the control effort penalty of rotating the robot joints.", + "type": "text" + } + ], + "index": 64 + } + ], + "index": 61 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 412, + 722 + ], + "score": 1.0, + "content": "Baselines. We compared NEXT with RRT∗ (Karaman & Frazzoli, 2011),", + "type": "text" + }, + { + "bbox": [ + 413, + 709, + 435, + 720 + ], + "score": 0.56, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "(Gammell et al.,", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "2015), CVAE-plan (Ichter et al., 2018), Reinforce-plan (Zhang et al., 2018), and an improved", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 65.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 280, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 281, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 281, + 95 + ], + "score": 1.0, + "content": "4.3 META SELF-IMPROVING LEARNING", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 97, + 505, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 96, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 218, + 112 + ], + "score": 1.0, + "content": "The learning of the planner", + "type": "text" + }, + { + "bbox": [ + 218, + 97, + 247, + 110 + ], + "score": 0.82, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 96, + 399, + 112 + ], + "score": 1.0, + "content": "reduces to learning the parameters in", + "type": "text" + }, + { + "bbox": [ + 400, + 96, + 438, + 110 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { * } \\left( s | U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 96, + 456, + 112 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 456, + 97, + 505, + 110 + ], + "score": 0.91, + "content": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 108, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 506, + 123 + ], + "score": 1.0, + "content": "and is carried out while planning experiences accumulate. We do not have an explicit training", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 119, + 505, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 505, + 132 + ], + "score": 1.0, + "content": "and testing phase separation. Particularly, we use a mixture of RRT :: Expand and NEXT :: Expand", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 504, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 174, + 143 + ], + "score": 1.0, + "content": "with probability", + "type": "text" + }, + { + "bbox": [ + 175, + 133, + 181, + 141 + ], + "score": 0.65, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 131, + 200, + 143 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 200, + 131, + 223, + 142 + ], + "score": 0.81, + "content": "1 - \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 131, + 504, + 143 + ], + "score": 1.0, + "content": ", respectively, inside the TSA framework in Algorithm 1. The RRT∗", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 309, + 155 + ], + "score": 1.0, + "content": "postprocessing step is used in the template. The", + "type": "text" + }, + { + "bbox": [ + 310, + 145, + 316, + 153 + ], + "score": 0.68, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 142, + 433, + 155 + ], + "score": 1.0, + "content": "is set to be 1 initially since", + "type": "text" + }, + { + "bbox": [ + 433, + 141, + 472, + 155 + ], + "score": 0.93, + "content": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "are not", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 314, + 167 + ], + "score": 1.0, + "content": "well-trained, and thus, the algorithm behaves like", + "type": "text" + }, + { + "bbox": [ + 315, + 154, + 339, + 165 + ], + "score": 0.56, + "content": "\\mathbf { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 154, + 498, + 167 + ], + "score": 1.0, + "content": ". As the training proceeds, we anneal", + "type": "text" + }, + { + "bbox": [ + 498, + 156, + 504, + 164 + ], + "score": 0.7, + "content": "\\epsilon", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 303, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 303, + 177 + ], + "score": 1.0, + "content": "gradually as the sampler becomes more efficient.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4, + "bbox_fs": [ + 104, + 96, + 506, + 177 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 181, + 505, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 180, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 153, + 196 + ], + "score": 1.0, + "content": "The dataset", + "type": "text" + }, + { + "bbox": [ + 154, + 181, + 237, + 195 + ], + "score": 0.92, + "content": "\\mathcal { D } _ { n } = \\{ ( \\mathcal { T } _ { j } , U _ { j } ) \\} _ { j = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 180, + 267, + 196 + ], + "score": 1.0, + "content": "for the", + "type": "text" + }, + { + "bbox": [ + 267, + 183, + 274, + 192 + ], + "score": 0.8, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 180, + 506, + 196 + ], + "score": 1.0, + "content": "-th training epoch is collected from the previous successful", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 504, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 494, + 204 + ], + "score": 1.0, + "content": "planning experiences across multiple random problems. We fix the size of dataset and update", + "type": "text" + }, + { + "bbox": [ + 495, + 193, + 504, + 203 + ], + "score": 0.81, + "content": "\\mathcal { D }", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 217 + ], + "score": 1.0, + "content": "in the same way as experience reply buffer (Lin, 1992; Schaul et al., 2015). For an experience", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 211, + 507, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 163, + 227 + ], + "score": 0.92, + "content": "( { \\mathcal { T } } , U ) \\in { \\mathcal { D } } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 211, + 331, + 231 + ], + "score": 1.0, + "content": ", we can reconstruct the successful path", + "type": "text" + }, + { + "bbox": [ + 332, + 214, + 364, + 227 + ], + "score": 0.94, + "content": "\\{ s ^ { i } \\} _ { i = 1 } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 211, + 452, + 231 + ], + "score": 1.0, + "content": "from the search tree", + "type": "text" + }, + { + "bbox": [ + 453, + 215, + 463, + 225 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 211, + 468, + 231 + ], + "score": 1.0, + "content": "(", + "type": "text" + }, + { + "bbox": [ + 468, + 216, + 478, + 225 + ], + "score": 0.72, + "content": "\\mathrm { ~ m ~ }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 211, + 507, + 231 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 102, + 226, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 102, + 231, + 154, + 256 + ], + "score": 1.0, + "content": "region, i.e.,", + "type": "text" + }, + { + "bbox": [ + 105, + 226, + 308, + 239 + ], + "score": 1.0, + "content": "number of segments), and the value of each state", + "type": "text" + }, + { + "bbox": [ + 154, + 236, + 257, + 251 + ], + "score": 0.91, + "content": "\\begin{array} { r } { y ^ { i } : = \\sum _ { l = i } ^ { m - 1 } c ( [ s ^ { l } , s ^ { l + 1 } ] ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 231, + 300, + 256 + ], + "score": 1.0, + "content": ". We learn", + "type": "text" + }, + { + "bbox": [ + 300, + 237, + 338, + 250 + ], + "score": 0.94, + "content": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 226, + 317, + 236 + ], + "score": 0.87, + "content": "s ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "in the path will be the sum of cost to the goal", + "type": "text" + }, + { + "bbox": [ + 339, + 231, + 441, + 256 + ], + "score": 1.0, + "content": "by optimizing objective", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 102, + 180, + 507, + 256 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 109, + 252, + 489, + 286 + ], + "lines": [ + { + "bbox": [ + 109, + 252, + 489, + 286 + ], + "spans": [ + { + "bbox": [ + 109, + 252, + 489, + 286 + ], + "score": 0.88, + "content": "\\operatorname* { m i n } _ { W } \\sum _ { ( { \\cal T } , { \\cal U } ) \\in { \\mathcal D } _ { n } } \\ell ( \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } ; { \\cal T } , { \\cal U } ) : = - \\sum _ { { \\mathcal D } _ { n } } \\sum _ { i = 1 } ^ { m - 1 } \\log \\tilde { \\pi } ^ { * } \\left( s ^ { i + 1 } | s ^ { i } \\right) + \\sum _ { i = 1 } ^ { m } ( \\tilde { V } ^ { * } \\left( s ^ { i } \\right) - y ^ { i } ) _ { 2 } ^ { 2 } + \\lambda \\left\\| { \\cal W } \\right\\| ^ { 2 } .", + "type": "interline_equation", + "image_path": "73a67d0d8a8b20fc2ce7c14d50c96eb2453ad6e3db342b8b427574a485703b12.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 109, + 252, + 489, + 263.3333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 109, + 263.3333333333333, + 489, + 274.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 109, + 274.66666666666663, + 489, + 285.99999999999994 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 292, + 294, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 291, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 207, + 306 + ], + "score": 1.0, + "content": "The loss (6) pushes the", + "type": "text" + }, + { + "bbox": [ + 208, + 291, + 222, + 303 + ], + "score": 0.87, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 291, + 242, + 306 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 242, + 293, + 254, + 303 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 291, + 294, + 306 + ], + "score": 1.0, + "content": "to chase", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 303, + 293, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 293, + 316 + ], + "score": 1.0, + "content": "the successful trajectories, providing effective", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 315, + 293, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 153, + 327 + ], + "score": 1.0, + "content": "guidance in", + "type": "text" + }, + { + "bbox": [ + 154, + 315, + 182, + 327 + ], + "score": 0.88, + "content": "\\mathsf { a l g } \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 315, + 293, + 327 + ], + "score": 1.0, + "content": "for searching, and therefore", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 325, + 293, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 293, + 337 + ], + "score": 1.0, + "content": "leading to efficient searching procedure with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 336, + 293, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 293, + 348 + ], + "score": 1.0, + "content": "less sample complexity and better solution. On", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 347, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 294, + 360 + ], + "score": 1.0, + "content": "one hand, the value function and policy esti-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 137, + 372 + ], + "score": 1.0, + "content": "mation", + "type": "text" + }, + { + "bbox": [ + 138, + 358, + 176, + 372 + ], + "score": 0.94, + "content": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 358, + 294, + 372 + ], + "score": 1.0, + "content": "is improved based upon the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 371, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 294, + 383 + ], + "score": 1.0, + "content": "successful outcomes from NEXT itself on pre-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 381, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 294, + 393 + ], + "score": 1.0, + "content": "vious problems. On the other hand, the updated", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 392, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 107, + 392, + 145, + 406 + ], + "score": 0.93, + "content": "\\{ \\tilde { V } ^ { * } , \\bar { \\pi } ^ { * } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 392, + 294, + 407 + ], + "score": 1.0, + "content": "will be applied in the next epoch to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 405, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 294, + 417 + ], + "score": 1.0, + "content": "improve the performance. Therefore, the train-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 416, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 295, + 428 + ], + "score": 1.0, + "content": "ing is named as Meta Self-Improving Learn-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 427, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 294, + 438 + ], + "score": 1.0, + "content": "ing (MSIL). Since all the trajectories we col-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 437, + 293, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 293, + 450 + ], + "score": 1.0, + "content": "lected for learning are feasible, the reachability", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 294, + 462 + ], + "score": 1.0, + "content": "of the proposed samples is enforced implicitly", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 460, + 252, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 252, + 472 + ], + "score": 1.0, + "content": "via imitating these successful paths.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 291, + 295, + 472 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 477, + 301, + 499 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 293, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 293, + 488 + ], + "score": 1.0, + "content": "By putting every components together into the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 486, + 302, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 302, + 500 + ], + "score": 1.0, + "content": "10 learning to plan framework in Eq (2), the over-", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 477, + 302, + 500 + ] + }, + { + "type": "title", + "bbox": [ + 306, + 292, + 485, + 304 + ], + "lines": [ + { + "bbox": [ + 303, + 290, + 487, + 308 + ], + "spans": [ + { + "bbox": [ + 303, + 290, + 487, + 308 + ], + "score": 1.0, + "content": "Algorithm 3: Meta Self-Improving Learning", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "list", + "bbox": [ + 296, + 308, + 493, + 495 + ], + "lines": [ + { + "bbox": [ + 295, + 306, + 390, + 319 + ], + "spans": [ + { + "bbox": [ + 295, + 306, + 372, + 319 + ], + "score": 1.0, + "content": "1 Initialize dataset", + "type": "text" + }, + { + "bbox": [ + 373, + 307, + 386, + 318 + ], + "score": 0.85, + "content": "\\mathcal { D } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 306, + 390, + 319 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 296, + 318, + 410, + 329 + ], + "spans": [ + { + "bbox": [ + 296, + 320, + 302, + 327 + ], + "score": 0.48, + "content": "^ 2", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 318, + 345, + 329 + ], + "score": 1.0, + "content": "for epoch", + "type": "text" + }, + { + "bbox": [ + 346, + 318, + 374, + 328 + ], + "score": 0.89, + "content": "n \\gets 1", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 318, + 385, + 329 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 385, + 318, + 396, + 328 + ], + "score": 0.69, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 318, + 410, + 329 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 296, + 329, + 446, + 340 + ], + "spans": [ + { + "bbox": [ + 296, + 331, + 302, + 338 + ], + "score": 0.37, + "content": "^ 3", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 329, + 432, + 340 + ], + "score": 1.0, + "content": "Sample a planning problem", + "type": "text" + }, + { + "bbox": [ + 432, + 329, + 441, + 339 + ], + "score": 0.77, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 329, + 446, + 340 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 297, + 339, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 297, + 343, + 303, + 349 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 320, + 340, + 375, + 352 + ], + "score": 0.88, + "content": "\\mathcal { T } \\mathrm { T S A } ( U )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 339, + 397, + 352 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 397, + 340, + 458, + 352 + ], + "score": 0.9, + "content": "\\epsilon \\sim \\mathcal { U } n i f [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 339, + 479, + 352 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 325, + 352, + 493, + 365 + ], + "spans": [ + { + "bbox": [ + 325, + 354, + 330, + 363 + ], + "score": 0.54, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 352, + 391, + 365 + ], + "score": 1.0, + "content": "·RRT :: Expand", + "type": "text" + }, + { + "bbox": [ + 391, + 353, + 423, + 365 + ], + "score": 0.34, + "content": "+ ( 1 - \\epsilon )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 352, + 493, + 365 + ], + "score": 1.0, + "content": "·NEXT :: Expand;", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 324, + 365, + 408, + 378 + ], + "spans": [ + { + "bbox": [ + 324, + 365, + 408, + 378 + ], + "score": 1.0, + "content": "Postprocessing with", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 325, + 376, + 420, + 389 + ], + "spans": [ + { + "bbox": [ + 325, + 377, + 346, + 387 + ], + "score": 0.26, + "content": "\\mathbb { R R T } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 376, + 420, + 389 + ], + "score": 1.0, + "content": ":: Postprocess;", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 297, + 387, + 475, + 399 + ], + "spans": [ + { + "bbox": [ + 297, + 390, + 302, + 398 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 320, + 387, + 421, + 399 + ], + "score": 0.9, + "content": "{ \\mathcal { D } } _ { n } \\gets { \\mathcal { D } } _ { n - 1 } \\bar { \\cup } \\{ ( { \\mathcal { T } } , U ) \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 387, + 475, + 399 + ], + "score": 1.0, + "content": "if successful", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 324, + 397, + 372, + 412 + ], + "spans": [ + { + "bbox": [ + 324, + 397, + 342, + 412 + ], + "score": 1.0, + "content": "else", + "type": "text" + }, + { + "bbox": [ + 343, + 399, + 367, + 410 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { n - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 397, + 372, + 412 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 297, + 409, + 396, + 421 + ], + "spans": [ + { + "bbox": [ + 297, + 413, + 303, + 420 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 319, + 409, + 335, + 421 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 335, + 410, + 361, + 421 + ], + "score": 0.9, + "content": "j 0", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 409, + 373, + 421 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 373, + 410, + 381, + 419 + ], + "score": 0.72, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 409, + 396, + 421 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 297, + 420, + 487, + 451 + ], + "spans": [ + { + "bbox": [ + 297, + 424, + 303, + 443 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 331, + 420, + 368, + 451 + ], + "score": 1.0, + "content": "Sample Reconst", + "type": "text" + }, + { + "bbox": [ + 368, + 420, + 401, + 433 + ], + "score": 0.92, + "content": "( \\tau _ { j } , U _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 420, + 425, + 451 + ], + "score": 1.0, + "content": "from -optim", + "type": "text" + }, + { + "bbox": [ + 425, + 421, + 439, + 432 + ], + "score": 0.89, + "content": "\\mathcal { D } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 420, + 455, + 451 + ], + "score": 1.0, + "content": ";path", + "type": "text" + }, + { + "bbox": [ + 456, + 432, + 487, + 445 + ], + "score": 0.9, + "content": "\\{ s ^ { i } \\} _ { i = 1 } ^ { m }", + "type": "inline_equation" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 339, + 444, + 478, + 457 + ], + "spans": [ + { + "bbox": [ + 339, + 444, + 462, + 457 + ], + "score": 1.0, + "content": "and the cost of paths based on", + "type": "text" + }, + { + "bbox": [ + 462, + 445, + 473, + 456 + ], + "score": 0.87, + "content": "\\tau _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 444, + 478, + 457 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 297, + 455, + 413, + 466 + ], + "spans": [ + { + "bbox": [ + 297, + 458, + 303, + 464 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 335, + 455, + 413, + 466 + ], + "score": 1.0, + "content": "Update parameters", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 339, + 466, + 484, + 481 + ], + "spans": [ + { + "bbox": [ + 339, + 466, + 484, + 481 + ], + "score": 0.71, + "content": "W W - \\eta \\nabla _ { W } \\ell ( \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } ; T _ { j } , U _ { j } ) ;", + "type": "inline_equation", + "image_path": "a45aa838211278ffba9b38ea7cf900286b1f3f0324cacf5bec3f0c31b3e16a52.jpg" + } + ], + "index": 48 + }, + { + "bbox": [ + 319, + 483, + 430, + 496 + ], + "spans": [ + { + "bbox": [ + 319, + 483, + 351, + 496 + ], + "score": 1.0, + "content": "Anneal", + "type": "text" + }, + { + "bbox": [ + 351, + 484, + 380, + 494 + ], + "score": 0.55, + "content": "\\epsilon = \\alpha \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 483, + 384, + 496 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 384, + 483, + 426, + 495 + ], + "score": 0.82, + "content": "\\alpha \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 483, + 430, + 496 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 49, + "is_list_end_line": true + } + ], + "index": 42, + "bbox_fs": [ + 295, + 306, + 493, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 499, + 392, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 391, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 391, + 512 + ], + "score": 1.0, + "content": "all procedure is summarized in Algorithm 3 and illustrated in Figure 1.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50, + "bbox_fs": [ + 106, + 498, + 391, + 512 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 522, + 200, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 201, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 201, + 536 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "In this section, we evaluate the proposed NEXT empirically on different planning tasks in a variety of", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "score": 1.0, + "content": "environments. Comparing to the existing planning algorithms, NEXT achieves the state-of-the-art", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "performances, in terms of both success rate and the quality of the found solutions. We further", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "demonstrate the power of the proposed two components by the corresponding ablation study. We also", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 585, + 459, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 459, + 598 + ], + "score": 1.0, + "content": "include a case study on a real-world robot arm control problem at the end of the section.", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 54, + "bbox_fs": [ + 105, + 541, + 506, + 598 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 609, + 218, + 620 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 220, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 220, + 621 + ], + "score": 1.0, + "content": "5.1 EXPERIMENT SETUP", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 57 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "Benchmark environments. We designed four benchmark tasks to demonstrate the effectiveness of", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "our algorithm for high-dimensional planning. The first three involve planning in a 2d workspace with", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 661 + ], + "score": 1.0, + "content": "a 2 DoF (degrees of freedom) point robot, a 3 DoF stick robot and a 5 DoF snake robot, respectively.", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "The last one involves planning a 7 DoF spacecraft in a 3d workspace. For all problems in each", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "benchmark task, the workspace maps were randomly generated from a fixed distribution; the initial", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 444, + 694 + ], + "score": 1.0, + "content": "and goal states were sampled uniformly randomly in the free space; the cost function", + "type": "text" + }, + { + "bbox": [ + 444, + 682, + 462, + 694 + ], + "score": 0.87, + "content": "c \\left( \\cdot \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "was set as", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 693, + 483, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 483, + 705 + ], + "score": 1.0, + "content": "the sum of the Euclidean path length and the control effort penalty of rotating the robot joints.", + "type": "text" + } + ], + "index": 64 + } + ], + "index": 61, + "bbox_fs": [ + 105, + 626, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 412, + 722 + ], + "score": 1.0, + "content": "Baselines. We compared NEXT with RRT∗ (Karaman & Frazzoli, 2011),", + "type": "text" + }, + { + "bbox": [ + 413, + 709, + 435, + 720 + ], + "score": 0.56, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "(Gammell et al.,", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "2015), CVAE-plan (Ichter et al., 2018), Reinforce-plan (Zhang et al., 2018), and an improved", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 368, + 504, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 504, + 381 + ], + "score": 1.0, + "content": "version of GPPN (Lee et al., 2018) in terms of both planning time and solution quality. RRT∗", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 124, + 393 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 124, + 380, + 146, + 390 + ], + "score": 0.68, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 146, + 379, + 505, + 393 + ], + "score": 1.0, + "content": "are two widely used effective instances of TSA in Algorithm 1. In our experiments, we", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 144, + 404 + ], + "score": 1.0, + "content": "equipped", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 145, + 391, + 169, + 401 + ], + "score": 0.36, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 169, + 390, + 401, + 404 + ], + "score": 1.0, + "content": "with the goal biasing heuristic to improve its performance.", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 402, + 391, + 423, + 401 + ], + "score": 0.58, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 423, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "adopts the informed", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "search strategy (Gammell et al., 2015) to accelerate planning. CVAE-plan and Reinforce-plan are", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 412, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 427 + ], + "score": 1.0, + "content": "two learning-enhanced TSA planners proposed recently. CVAE-plan learns a conditional VAE as", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 478, + 437 + ], + "score": 1.0, + "content": "the sampler (Sohn et al., 2015), which will be trained by near-optimal paths produced by", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 478, + 424, + 503, + 434 + ], + "score": 0.28, + "content": "\\mathbf { R R T ^ { * } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 503, + 423, + 506, + 437 + ], + "score": 1.0, + "content": ".", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "Reinforce-plan learns to do rejection sampling with policy gradient methods. For the improved", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "GPPN, we combined its architecture for map with a fully-connected MLP for the rest state, such that", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "it can be applied to high-dimensional continuous spaces. Please refer to Appendix E for more details.", + "type": "text", + "cross_page": true + } + ], + "index": 21 + } + ], + "index": 65.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 83, + 496, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 83, + 496, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 83, + 496, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 83, + 496, + 186 + ], + "score": 0.965, + "type": "image", + "image_path": "b0b18ba9b0a2a21dab4fe0925d12969b8bbd06feb8546ddafdca1a81131ba122.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 83, + 496, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 117.33333333333334, + 496, + 151.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 151.66666666666669, + 496, + 186.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 190, + 505, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 189, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 249, + 203 + ], + "score": 1.0, + "content": "Figure 5: Search trees and the learned", + "type": "text" + }, + { + "bbox": [ + 250, + 189, + 262, + 200 + ], + "score": 0.89, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 191, + 278, + 203 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 279, + 191, + 290, + 201 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "produced by NEXT. Obstacles are colored in blue. The start", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "and goal locations are denoted by orange and brown dots. In (a) to (c), samples are represented with yellow", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 210, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 387, + 224 + ], + "score": 1.0, + "content": "circles. In (d), the level of redness denotes the value of the cost-to-go estimate", + "type": "text" + }, + { + "bbox": [ + 387, + 210, + 399, + 221 + ], + "score": 0.87, + "content": "\\tilde { V } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 212, + 505, + 224 + ], + "score": 1.0, + "content": ". The cyan arrows point from", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 222, + 327, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 153, + 234 + ], + "score": 1.0, + "content": "a given state", + "type": "text" + }, + { + "bbox": [ + 154, + 225, + 159, + 231 + ], + "score": 0.77, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 222, + 280, + 234 + ], + "score": 1.0, + "content": "to the mean of the learned policy", + "type": "text" + }, + { + "bbox": [ + 280, + 222, + 323, + 232 + ], + "score": 0.92, + "content": "\\tilde { \\pi } ^ { * } ( s ^ { \\prime } | s , U )", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 222, + 327, + 234 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 127, + 236, + 503, + 328 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 236, + 503, + 328 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 127, + 236, + 503, + 328 + ], + "spans": [ + { + "bbox": [ + 127, + 236, + 503, + 328 + ], + "score": 0.967, + "type": "image", + "image_path": "e9b64d33b209ec3b631b087dbd2e5d1f20da83a228d5b5ee93466c67d2fe104a.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 127, + 236, + 503, + 266.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 127, + 266.6666666666667, + 503, + 297.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 127, + 297.33333333333337, + 503, + 328.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 333, + 505, + 363 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 333, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 505, + 344 + ], + "score": 1.0, + "content": "Figure 6: Search trees and a solution path produced in an instance of spacecraft planning. The 7 DOF spacecraft", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 343, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 354 + ], + "score": 1.0, + "content": "has a yellow body and two 2 DOF red arms. NEXT-KS produced a nearly minimum viable search tree while", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 352, + 277, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 129, + 363 + ], + "score": 0.35, + "content": "{ \\mathrm { R R T } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 352, + 277, + 364 + ], + "score": 1.0, + "content": "failed to find a path within limited trials.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 368, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 368, + 504, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 504, + 381 + ], + "score": 1.0, + "content": "version of GPPN (Lee et al., 2018) in terms of both planning time and solution quality. RRT∗", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 124, + 393 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 380, + 146, + 390 + ], + "score": 0.68, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 379, + 505, + 393 + ], + "score": 1.0, + "content": "are two widely used effective instances of TSA in Algorithm 1. In our experiments, we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 144, + 404 + ], + "score": 1.0, + "content": "equipped", + "type": "text" + }, + { + "bbox": [ + 145, + 391, + 169, + 401 + ], + "score": 0.36, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 390, + 401, + 404 + ], + "score": 1.0, + "content": "with the goal biasing heuristic to improve its performance.", + "type": "text" + }, + { + "bbox": [ + 402, + 391, + 423, + 401 + ], + "score": 0.58, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "adopts the informed", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "search strategy (Gammell et al., 2015) to accelerate planning. CVAE-plan and Reinforce-plan are", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 412, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 427 + ], + "score": 1.0, + "content": "two learning-enhanced TSA planners proposed recently. CVAE-plan learns a conditional VAE as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 478, + 437 + ], + "score": 1.0, + "content": "the sampler (Sohn et al., 2015), which will be trained by near-optimal paths produced by", + "type": "text" + }, + { + "bbox": [ + 478, + 424, + 503, + 434 + ], + "score": 0.28, + "content": "\\mathbf { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 423, + 506, + 437 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "Reinforce-plan learns to do rejection sampling with policy gradient methods. For the improved", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "GPPN, we combined its architecture for map with a fully-connected MLP for the rest state, such that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 104, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "it can be applied to high-dimensional continuous spaces. Please refer to Appendix E for more details.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "score": 1.0, + "content": "Settings. For each task, we randomly generated 3000 different problems from the same distribution", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "without duplicated maps. We trained all learning-based baselines using the first 2000 problems, and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 300, + 508 + ], + "score": 1.0, + "content": "reserved the rest for testing. The parameters for", + "type": "text" + }, + { + "bbox": [ + 300, + 495, + 325, + 506 + ], + "score": 0.65, + "content": "\\mathrm { { R R T ^ { * } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 495, + 343, + 508 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 344, + 495, + 365, + 506 + ], + "score": 0.69, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "are also tuned using the first 2000", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "problems. For NEXT, we let it improve itself using MSIL over the first 2000 problems. In this period,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 517, + 396, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 365, + 530 + ], + "score": 1.0, + "content": "for every 200 problems, we updated its parameters and annealed", + "type": "text" + }, + { + "bbox": [ + 365, + 519, + 371, + 527 + ], + "score": 0.74, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 517, + 396, + 530 + ], + "score": 1.0, + "content": "once.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 108, + 542, + 237, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 239, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 239, + 555 + ], + "score": 1.0, + "content": "5.2 RESULTS AND ANALYSIS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 107, + 560, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 107, + 560, + 506, + 572 + ], + "score": 1.0, + "content": "Comparison results. Examples of all four environments are illustrated in Appendix F.1 and Figure 6,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "where NEXT finds high-quality solutions as shown. We also illustrated the comparison of the search", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 582, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 594 + ], + "score": 1.0, + "content": "trees on two 2d and 7d planning tasks between NEXT and RRT* in Figure 5 (a)-(c) and Figure 6", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "(b) and (c). Obviously, the proposed NEXT algorithm explores with guidance and achieves better", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "quality solutions with fewer samples, while the RRT* expands randomly which may fail to find a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 195, + 630 + ], + "score": 1.0, + "content": "solution. The learned", + "type": "text" + }, + { + "bbox": [ + 196, + 614, + 209, + 626 + ], + "score": 0.89, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 615, + 227, + 630 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 616, + 240, + 626 + ], + "score": 0.87, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "in the 2d task are also shown in Figure 5(d). As we can see, they", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 628, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 639 + ], + "score": 1.0, + "content": "are consistent with our expectation, towards the ultimate target in the map. For more search tree", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 639, + 474, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 474, + 651 + ], + "score": 1.0, + "content": "comparisons for all four experiments, please check Figure 18, 19, 20 and 21 in Appendix F.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "To systematically evaluate the algorithms, we recorded the cost of time (measured by the number of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 666, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 677 + ], + "score": 1.0, + "content": "collision checks used) to find a collision-free path, the success rate within time limits, and the cost of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "the solution path for each run. The results of the reserved 1000 test problems of each environment are", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "shown in the top row of Figure 7. We set the maximal number of samples as 500 for all algorithms.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Both the kernel smoothing (NEXT-KS) and the Gaussian process (NEXT-GP) version of NEXT", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 710, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 504, + 721 + ], + "score": 1.0, + "content": "achieves the state-of-the-art performances, under all three criteria in all test environments. Although", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 121, + 733 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 721, + 144, + 731 + ], + "score": 0.63, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "utilizes the heuristic particularly suitable for 3d maze in 7d task and performs quite well,", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 83, + 496, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 83, + 496, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 83, + 496, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 83, + 496, + 186 + ], + "score": 0.965, + "type": "image", + "image_path": "b0b18ba9b0a2a21dab4fe0925d12969b8bbd06feb8546ddafdca1a81131ba122.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 83, + 496, + 117.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 117.33333333333334, + 496, + 151.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 151.66666666666669, + 496, + 186.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 190, + 505, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 189, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 249, + 203 + ], + "score": 1.0, + "content": "Figure 5: Search trees and the learned", + "type": "text" + }, + { + "bbox": [ + 250, + 189, + 262, + 200 + ], + "score": 0.89, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 191, + 278, + 203 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 279, + 191, + 290, + 201 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "produced by NEXT. Obstacles are colored in blue. The start", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 212 + ], + "score": 1.0, + "content": "and goal locations are denoted by orange and brown dots. In (a) to (c), samples are represented with yellow", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 210, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 387, + 224 + ], + "score": 1.0, + "content": "circles. In (d), the level of redness denotes the value of the cost-to-go estimate", + "type": "text" + }, + { + "bbox": [ + 387, + 210, + 399, + 221 + ], + "score": 0.87, + "content": "\\tilde { V } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 212, + 505, + 224 + ], + "score": 1.0, + "content": ". The cyan arrows point from", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 222, + 327, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 153, + 234 + ], + "score": 1.0, + "content": "a given state", + "type": "text" + }, + { + "bbox": [ + 154, + 225, + 159, + 231 + ], + "score": 0.77, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 222, + 280, + 234 + ], + "score": 1.0, + "content": "to the mean of the learned policy", + "type": "text" + }, + { + "bbox": [ + 280, + 222, + 323, + 232 + ], + "score": 0.92, + "content": "\\tilde { \\pi } ^ { * } ( s ^ { \\prime } | s , U )", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 222, + 327, + 234 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 127, + 236, + 503, + 328 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 236, + 503, + 328 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 127, + 236, + 503, + 328 + ], + "spans": [ + { + "bbox": [ + 127, + 236, + 503, + 328 + ], + "score": 0.967, + "type": "image", + "image_path": "e9b64d33b209ec3b631b087dbd2e5d1f20da83a228d5b5ee93466c67d2fe104a.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 127, + 236, + 503, + 266.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 127, + 266.6666666666667, + 503, + 297.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 127, + 297.33333333333337, + 503, + 328.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 333, + 505, + 363 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 333, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 505, + 344 + ], + "score": 1.0, + "content": "Figure 6: Search trees and a solution path produced in an instance of spacecraft planning. The 7 DOF spacecraft", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 343, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 354 + ], + "score": 1.0, + "content": "has a yellow body and two 2 DOF red arms. NEXT-KS produced a nearly minimum viable search tree while", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 352, + 277, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 129, + 363 + ], + "score": 0.35, + "content": "{ \\mathrm { R R T } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 352, + 277, + 364 + ], + "score": 1.0, + "content": "failed to find a path within limited trials.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 368, + 505, + 468 + ], + "lines": [], + "index": 17, + "bbox_fs": [ + 104, + 368, + 506, + 469 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 473, + 505, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 486 + ], + "score": 1.0, + "content": "Settings. For each task, we randomly generated 3000 different problems from the same distribution", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "without duplicated maps. We trained all learning-based baselines using the first 2000 problems, and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 300, + 508 + ], + "score": 1.0, + "content": "reserved the rest for testing. The parameters for", + "type": "text" + }, + { + "bbox": [ + 300, + 495, + 325, + 506 + ], + "score": 0.65, + "content": "\\mathrm { { R R T ^ { * } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 495, + 343, + 508 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 344, + 495, + 365, + 506 + ], + "score": 0.69, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "are also tuned using the first 2000", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 506, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 506, + 519 + ], + "score": 1.0, + "content": "problems. For NEXT, we let it improve itself using MSIL over the first 2000 problems. In this period,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 517, + 396, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 365, + 530 + ], + "score": 1.0, + "content": "for every 200 problems, we updated its parameters and annealed", + "type": "text" + }, + { + "bbox": [ + 365, + 519, + 371, + 527 + ], + "score": 0.74, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 517, + 396, + 530 + ], + "score": 1.0, + "content": "once.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 472, + 506, + 530 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 542, + 237, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 541, + 239, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 239, + 555 + ], + "score": 1.0, + "content": "5.2 RESULTS AND ANALYSIS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 107, + 560, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 107, + 560, + 506, + 572 + ], + "score": 1.0, + "content": "Comparison results. Examples of all four environments are illustrated in Appendix F.1 and Figure 6,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "where NEXT finds high-quality solutions as shown. We also illustrated the comparison of the search", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 582, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 594 + ], + "score": 1.0, + "content": "trees on two 2d and 7d planning tasks between NEXT and RRT* in Figure 5 (a)-(c) and Figure 6", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "(b) and (c). Obviously, the proposed NEXT algorithm explores with guidance and achieves better", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "quality solutions with fewer samples, while the RRT* expands randomly which may fail to find a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 614, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 195, + 630 + ], + "score": 1.0, + "content": "solution. The learned", + "type": "text" + }, + { + "bbox": [ + 196, + 614, + 209, + 626 + ], + "score": 0.89, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 615, + 227, + 630 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 616, + 240, + 626 + ], + "score": 0.87, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "in the 2d task are also shown in Figure 5(d). As we can see, they", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 628, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 639 + ], + "score": 1.0, + "content": "are consistent with our expectation, towards the ultimate target in the map. For more search tree", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 639, + 474, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 474, + 651 + ], + "score": 1.0, + "content": "comparisons for all four experiments, please check Figure 18, 19, 20 and 21 in Appendix F.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 560, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "To systematically evaluate the algorithms, we recorded the cost of time (measured by the number of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 666, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 677 + ], + "score": 1.0, + "content": "collision checks used) to find a collision-free path, the success rate within time limits, and the cost of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "the solution path for each run. The results of the reserved 1000 test problems of each environment are", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "shown in the top row of Figure 7. We set the maximal number of samples as 500 for all algorithms.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Both the kernel smoothing (NEXT-KS) and the Gaussian process (NEXT-GP) version of NEXT", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 710, + 504, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 504, + 721 + ], + "score": 1.0, + "content": "achieves the state-of-the-art performances, under all three criteria in all test environments. Although", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 121, + 733 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 721, + 144, + 731 + ], + "score": 0.63, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "utilizes the heuristic particularly suitable for 3d maze in 7d task and performs quite well,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "the NEXT algorithm still outperform every competitor by a large margin, no matter learning-based or", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 384, + 471, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 471, + 397 + ], + "score": 1.0, + "content": "prefixed heuristic planner, demonstrating the advantages of the proposed NEXT algorithm.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 654, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 81, + 502, + 321 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 81, + 502, + 321 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 502, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 502, + 321 + ], + "score": 0.975, + "type": "image", + "image_path": "e1b7307ae6db488a13797cccd31ec60ba79f87159ac956c545b9f9d1f3fe33ce.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 81, + 502, + 161.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 161.0, + 502, + 241.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 241.0, + 502, + 321.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 328, + 505, + 368 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "Figure 7: First row: histograms of results, in terms of success rate, average collision checks, and average cost", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "of the solution paths; Second row: NEXT improvement curves in the 5d experiments. All algorithms are set to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 348, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 217, + 359 + ], + "score": 1.0, + "content": "use up to 500 samples, except", + "type": "text" + }, + { + "bbox": [ + 217, + 348, + 239, + 357 + ], + "score": 0.42, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 348, + 505, + 359 + ], + "score": 1.0, + "content": "-10k, which uses 10,000 samples. The value of collision checks and path", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 357, + 299, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 273, + 370 + ], + "score": 1.0, + "content": "costs are normalized w.r.t. the performance of", + "type": "text" + }, + { + "bbox": [ + 273, + 358, + 295, + 368 + ], + "score": 0.6, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 357, + 299, + 370 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 373, + 502, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "the NEXT algorithm still outperform every competitor by a large margin, no matter learning-based or", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 384, + 471, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 471, + 397 + ], + "score": 1.0, + "content": "prefixed heuristic planner, demonstrating the advantages of the proposed NEXT algorithm.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 506, + 415 + ], + "score": 1.0, + "content": "Self-improving. We plot the performance improvement curves of our algorithms on the 5d planning", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "task in the bottom row of Figure 7. For comparison, we also plot the performance of RRT∗. At the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 374, + 436 + ], + "score": 1.0, + "content": "beginning phase of self-improving, our algorithms are comparable to", + "type": "text" + }, + { + "bbox": [ + 375, + 424, + 398, + 434 + ], + "score": 0.35, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 424, + 505, + 436 + ], + "score": 1.0, + "content": ". They then gradually learn", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "from previous experiences and improve themselves as they see more problems and better solutions.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "In the end, NEXT-KS is able to match the performance of RRT∗-10k using only one-twentieth of its", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 456, + 428, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 428, + 469 + ], + "score": 1.0, + "content": "samples, while the competitors perform consistently without any improvements.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 108, + 473, + 503, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "Due to the space limits, we put improvement curves on other environments in Figure 17 and the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 483, + 469, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 469, + 496 + ], + "score": 1.0, + "content": "quantitative evaluation in Table 1, 2, and 3 in Appendix F. Please refer to the details there.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 107, + 502, + 216, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 218, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 218, + 514 + ], + "score": 1.0, + "content": "5.3 ABLATION STUDIES", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "Ablation study I: guided progressive expansion. To demonstrate the power of NEXT :: Expand, we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "replace it with breadth-first search (BFS) (Kim et al., 2018), another expanding strategy, while keeping", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "other components the same. Specifically, BFS uses a search queue in planning. It repeatedly pops a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 127, + 563 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 128, + 552, + 134, + 560 + ], + "score": 0.72, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 550, + 281, + 563 + ], + "score": 1.0, + "content": "out from the search queue, samples", + "type": "text" + }, + { + "bbox": [ + 281, + 551, + 288, + 560 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 550, + 336, + 563 + ], + "score": 1.0, + "content": "states from", + "type": "text" + }, + { + "bbox": [ + 336, + 550, + 363, + 562 + ], + "score": 0.84, + "content": "\\pi \\left( \\cdot | s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 550, + 506, + 563 + ], + "score": 1.0, + "content": ", and pushes all generated samples", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 560, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 145, + 575 + ], + "score": 1.0, + "content": "and state", + "type": "text" + }, + { + "bbox": [ + 145, + 563, + 151, + 571 + ], + "score": 0.7, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 560, + 505, + 575 + ], + "score": 1.0, + "content": "back to the queue, until the goal is reached. For fairness, we use the learned sampling", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 135, + 585 + ], + "score": 1.0, + "content": "policy", + "type": "text" + }, + { + "bbox": [ + 135, + 573, + 178, + 584 + ], + "score": 0.92, + "content": "\\pi \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "by NEXT-KS in BFS. As shown in Figure 7, BFS obtained worse paths with a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "much larger number of collision checks and far lower success rate, which justifies the importance of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 594, + 487, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 487, + 606 + ], + "score": 1.0, + "content": "the balance between exploration versus exploitation achieved by the proposed NEXT :: Expand.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 611, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "Ablation study II: neural architecture. To further demonstrate the benefits of the proposed neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "architecture for learning generalizable representations in high-dimension planning problems, We", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "replaced our attention-based neural architecture with an improved GPPN, as explained in Appendix E,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "for ablation study. We extended the GPPN for continuous space by adding an extra reactive policy", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "network to its final layers. We emphasize the original GPPN is not applicable to the tasks in our", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "experiments. Intuitively, the improved GPPN first produces a ‘rough plan’ by processing the robot’s", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "discretized workspace positions. Then the reactive policy network predicts a continuous action", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "from both the workspace feature and the full configuration state of the robot. We provide more", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "preference to the improved GPPN by training it to imitate the near-optimal paths produced by RRT*", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "in the training problems. During test time it is also combined with both versions of the guided", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "progressive expansion operators. As we can see, both GPPN-KS and GPPN-GP are clearly much", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 81, + 502, + 321 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 81, + 502, + 321 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 502, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 502, + 321 + ], + "score": 0.975, + "type": "image", + "image_path": "e1b7307ae6db488a13797cccd31ec60ba79f87159ac956c545b9f9d1f3fe33ce.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 81, + 502, + 161.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 161.0, + 502, + 241.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 241.0, + 502, + 321.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 328, + 505, + 368 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "Figure 7: First row: histograms of results, in terms of success rate, average collision checks, and average cost", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "of the solution paths; Second row: NEXT improvement curves in the 5d experiments. All algorithms are set to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 348, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 217, + 359 + ], + "score": 1.0, + "content": "use up to 500 samples, except", + "type": "text" + }, + { + "bbox": [ + 217, + 348, + 239, + 357 + ], + "score": 0.42, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 348, + 505, + 359 + ], + "score": 1.0, + "content": "-10k, which uses 10,000 samples. The value of collision checks and path", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 357, + 299, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 273, + 370 + ], + "score": 1.0, + "content": "costs are normalized w.r.t. the performance of", + "type": "text" + }, + { + "bbox": [ + 273, + 358, + 295, + 368 + ], + "score": 0.6, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 357, + 299, + 370 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 373, + 502, + 396 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 373, + 505, + 397 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 506, + 415 + ], + "score": 1.0, + "content": "Self-improving. We plot the performance improvement curves of our algorithms on the 5d planning", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "task in the bottom row of Figure 7. For comparison, we also plot the performance of RRT∗. At the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 374, + 436 + ], + "score": 1.0, + "content": "beginning phase of self-improving, our algorithms are comparable to", + "type": "text" + }, + { + "bbox": [ + 375, + 424, + 398, + 434 + ], + "score": 0.35, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 424, + 505, + 436 + ], + "score": 1.0, + "content": ". They then gradually learn", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 433, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 447 + ], + "score": 1.0, + "content": "from previous experiences and improve themselves as they see more problems and better solutions.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "In the end, NEXT-KS is able to match the performance of RRT∗-10k using only one-twentieth of its", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 456, + 428, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 428, + 469 + ], + "score": 1.0, + "content": "samples, while the competitors perform consistently without any improvements.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 400, + 506, + 469 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 473, + 503, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "Due to the space limits, we put improvement curves on other environments in Figure 17 and the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 483, + 469, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 469, + 496 + ], + "score": 1.0, + "content": "quantitative evaluation in Table 1, 2, and 3 in Appendix F. Please refer to the details there.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 473, + 505, + 496 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 502, + 216, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 218, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 218, + 514 + ], + "score": 1.0, + "content": "5.3 ABLATION STUDIES", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 531 + ], + "score": 1.0, + "content": "Ablation study I: guided progressive expansion. To demonstrate the power of NEXT :: Expand, we", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "replace it with breadth-first search (BFS) (Kim et al., 2018), another expanding strategy, while keeping", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "other components the same. Specifically, BFS uses a search queue in planning. It repeatedly pops a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 127, + 563 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 128, + 552, + 134, + 560 + ], + "score": 0.72, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 550, + 281, + 563 + ], + "score": 1.0, + "content": "out from the search queue, samples", + "type": "text" + }, + { + "bbox": [ + 281, + 551, + 288, + 560 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 550, + 336, + 563 + ], + "score": 1.0, + "content": "states from", + "type": "text" + }, + { + "bbox": [ + 336, + 550, + 363, + 562 + ], + "score": 0.84, + "content": "\\pi \\left( \\cdot | s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 550, + 506, + 563 + ], + "score": 1.0, + "content": ", and pushes all generated samples", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 560, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 145, + 575 + ], + "score": 1.0, + "content": "and state", + "type": "text" + }, + { + "bbox": [ + 145, + 563, + 151, + 571 + ], + "score": 0.7, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 560, + 505, + 575 + ], + "score": 1.0, + "content": "back to the queue, until the goal is reached. For fairness, we use the learned sampling", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 135, + 585 + ], + "score": 1.0, + "content": "policy", + "type": "text" + }, + { + "bbox": [ + 135, + 573, + 178, + 584 + ], + "score": 0.92, + "content": "\\pi \\left( s ^ { \\prime } | s , U \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "by NEXT-KS in BFS. As shown in Figure 7, BFS obtained worse paths with a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "score": 1.0, + "content": "much larger number of collision checks and far lower success rate, which justifies the importance of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 594, + 487, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 487, + 606 + ], + "score": 1.0, + "content": "the balance between exploration versus exploitation achieved by the proposed NEXT :: Expand.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 516, + 506, + 606 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 611, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "Ablation study II: neural architecture. To further demonstrate the benefits of the proposed neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "architecture for learning generalizable representations in high-dimension planning problems, We", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "replaced our attention-based neural architecture with an improved GPPN, as explained in Appendix E,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 506, + 658 + ], + "score": 1.0, + "content": "for ablation study. We extended the GPPN for continuous space by adding an extra reactive policy", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "network to its final layers. We emphasize the original GPPN is not applicable to the tasks in our", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "experiments. Intuitively, the improved GPPN first produces a ‘rough plan’ by processing the robot’s", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "discretized workspace positions. Then the reactive policy network predicts a continuous action", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "from both the workspace feature and the full configuration state of the robot. We provide more", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "preference to the improved GPPN by training it to imitate the near-optimal paths produced by RRT*", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "in the training problems. During test time it is also combined with both versions of the guided", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "progressive expansion operators. As we can see, both GPPN-KS and GPPN-GP are clearly much", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "worse than NEXT-KS and NEXT-GP, demonstrating the advantage of our proposed attention-based", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 242, + 327, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 327, + 253 + ], + "score": 1.0, + "content": "neural architecture in high-dimensional planning tasks.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 611, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 81, + 434, + 192 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 176, + 81, + 434, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 81, + 434, + 192 + ], + "spans": [ + { + "bbox": [ + 176, + 81, + 434, + 192 + ], + "score": 0.969, + "type": "image", + "image_path": "a7f19f8a9284429626e2ad6f69c231d6aa6a4e1e1021040a5d3d6919d297a607.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 176, + 81, + 434, + 118.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 176, + 118.0, + 434, + 155.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 176, + 155.0, + 434, + 192.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 204, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 202, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 217 + ], + "score": 1.0, + "content": "Figure 8: The success rate and average path cost of the different planners under varying time limits. Running", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 213, + 408, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 330, + 226 + ], + "score": 1.0, + "content": "NEXT for 1 second achieves the same success rate as running", + "type": "text" + }, + { + "bbox": [ + 330, + 215, + 351, + 224 + ], + "score": 0.56, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 213, + 408, + 226 + ], + "score": 1.0, + "content": "for 50 seconds.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 105, + 230, + 503, + 253 + ], + "lines": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "worse than NEXT-KS and NEXT-GP, demonstrating the advantage of our proposed attention-based", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 242, + 327, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 327, + 253 + ], + "score": 1.0, + "content": "neural architecture in high-dimensional planning tasks.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "Ablation study III: learning versus heuristic. The NEXT algorithm in Figure 5 shows similar", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "behavior as the Dijkstra heuristic, i.e. sampling on the shortest path connecting the start and the goal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 280, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 504, + 293 + ], + "score": 1.0, + "content": "in workspace. However, in higher dimensional space, the Dijkstra heuristic will fail. To demonstrate", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "score": 1.0, + "content": "that, we replace the policy and value network with Dijkstra heuristic, using the best direction in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "workspace to guide sampling. NEXT performs much better than Dijkstra in all but the 2d case, in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 313, + 312, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 312, + 326 + ], + "score": 1.0, + "content": "which the workspace happens to be the state space.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 341, + 290, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 293, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 293, + 354 + ], + "score": 1.0, + "content": "5.4 CASE STUDY: ROBOT ARM CONTROL", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "We conduct a real-world case study on controlling robot arms to move objects on a shelf. On this", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 373, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 506, + 384 + ], + "score": 1.0, + "content": "representative real-time task, we demonstrate the advantages of the NEXT in terms of the wall-clock.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 333, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 334, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 334, + 402 + ], + "score": 1.0, + "content": "In each planning task, there is a shelf of multiple levels,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 400, + 334, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 334, + 412 + ], + "score": 1.0, + "content": "with each level horizontally divided into multiple bins.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 411, + 333, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 333, + 423 + ], + "score": 1.0, + "content": "The task is to plan a path from a location in one bin", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 422, + 333, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 333, + 434 + ], + "score": 1.0, + "content": "to another, i.e., the end effectors of the start and goal", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 433, + 334, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 334, + 446 + ], + "score": 1.0, + "content": "configurations are in different bins. The heights of levels,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 443, + 333, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 333, + 457 + ], + "score": 1.0, + "content": "widths of bins, and the start and goal are randomly drawn", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 455, + 333, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 333, + 467 + ], + "score": 1.0, + "content": "from some fixed distribution. Different from previous", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 466, + 334, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 334, + 478 + ], + "score": 1.0, + "content": "experiments, the base of the robot is fixed. We consider", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 477, + 333, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 120, + 489 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 477, + 143, + 488 + ], + "score": 0.61, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 477, + 184, + 489 + ], + "score": 1.0, + "content": "instead of", + "type": "text" + }, + { + "bbox": [ + 184, + 477, + 209, + 488 + ], + "score": 0.28, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 477, + 333, + 489 + ], + "score": 1.0, + "content": "as the imperfect expert in 3000", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 488, + 334, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 334, + 500 + ], + "score": 1.0, + "content": "training problems. We then evaluate the algorithm on a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 499, + 334, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 334, + 510 + ], + "score": 1.0, + "content": "separated 1000 testing problems. We compare NEXT(-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 510, + 334, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 218, + 522 + ], + "score": 1.0, + "content": "KS) with the highly tuned", + "type": "text" + }, + { + "bbox": [ + 219, + 510, + 242, + 520 + ], + "score": 0.52, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 510, + 334, + 522 + ], + "score": 1.0, + "content": "and RRT* in OMPL,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 521, + 334, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 334, + 533 + ], + "score": 1.0, + "content": "and also CVAE-plan and Reinforce-plan in Figure 8. As", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 532, + 333, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 333, + 544 + ], + "score": 1.0, + "content": "seen from the visualization of the found paths in Figure 9,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 542, + 333, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 333, + 555 + ], + "score": 1.0, + "content": "this is a very difficult task. Our NEXT outperforms the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 553, + 333, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 333, + 567 + ], + "score": 1.0, + "content": "baselines by a large margin, requiring only 1 second to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 565, + 333, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 333, + 577 + ], + "score": 1.0, + "content": "reach the same success rate as running 50 seconds of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 575, + 133, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 128, + 587 + ], + "score": 0.52, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 575, + 133, + 587 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 24.5 + }, + { + "type": "image", + "bbox": [ + 370, + 389, + 459, + 540 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 370, + 389, + 459, + 540 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 370, + 389, + 459, + 540 + ], + "spans": [ + { + "bbox": [ + 370, + 389, + 459, + 540 + ], + "score": 0.467, + "type": "image", + "image_path": "2c27831a81842e20acfb22f33643b38f7e0d636dac989887fae6e6036e1293d0.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 370, + 389, + 459, + 540 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 344, + 543, + 505, + 583 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 343, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 343, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "Figure 9: The collision-free path produced", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 343, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 343, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "by NEXT for robot arm planning. The start", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 343, + 563, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 343, + 563, + 505, + 574 + ], + "score": 1.0, + "content": "and goal configurations have end-effectors in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 344, + 573, + 439, + 583 + ], + "spans": [ + { + "bbox": [ + 344, + 573, + 439, + 583 + ], + "score": 1.0, + "content": "different bins of the shelf.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + } + ], + "index": 33.25 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 504, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 592, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 506, + 606 + ], + "score": 1.0, + "content": "Due to space limits, we put details of the experiment setups, more results and analysis in Appendix F.4.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "title", + "bbox": [ + 107, + 621, + 195, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 197, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 197, + 636 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "In this paper, we propose a self-improving planner, Neural EXploration-EXploitation Trees (NEXT),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "which can generalize and achieve better performance with experiences accumulated. The algorithm", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "achieves a delicate balance between exploration-exploitation via our carefully designed UCB-type", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "expansion operation. To obtain the generalizable ability across different problems, we proposed a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "new parametrization for the value function and policy, which captures the Bellman recursive structure", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "in the high-dimensional continuous state and action space. We demonstrate the power of the proposed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "algorithm by outperforming previous state-of-the-art planners with significant margins on planning", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 299, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 299, + 733 + ], + "score": 1.0, + "content": "problems in a variety of different environments.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 81, + 434, + 192 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 176, + 81, + 434, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 176, + 81, + 434, + 192 + ], + "spans": [ + { + "bbox": [ + 176, + 81, + 434, + 192 + ], + "score": 0.969, + "type": "image", + "image_path": "a7f19f8a9284429626e2ad6f69c231d6aa6a4e1e1021040a5d3d6919d297a607.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 176, + 81, + 434, + 118.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 176, + 118.0, + 434, + 155.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 176, + 155.0, + 434, + 192.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 204, + 505, + 225 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 202, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 217 + ], + "score": 1.0, + "content": "Figure 8: The success rate and average path cost of the different planners under varying time limits. Running", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 213, + 408, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 330, + 226 + ], + "score": 1.0, + "content": "NEXT for 1 second achieves the same success rate as running", + "type": "text" + }, + { + "bbox": [ + 330, + 215, + 351, + 224 + ], + "score": 0.56, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 213, + 408, + 226 + ], + "score": 1.0, + "content": "for 50 seconds.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 105, + 230, + 503, + 253 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 106, + 230, + 505, + 253 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 258, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "Ablation study III: learning versus heuristic. The NEXT algorithm in Figure 5 shows similar", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "behavior as the Dijkstra heuristic, i.e. sampling on the shortest path connecting the start and the goal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 280, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 504, + 293 + ], + "score": 1.0, + "content": "in workspace. However, in higher dimensional space, the Dijkstra heuristic will fail. To demonstrate", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 504, + 303 + ], + "score": 1.0, + "content": "that, we replace the policy and value network with Dijkstra heuristic, using the best direction in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "score": 1.0, + "content": "workspace to guide sampling. NEXT performs much better than Dijkstra in all but the 2d case, in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 313, + 312, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 312, + 326 + ], + "score": 1.0, + "content": "which the workspace happens to be the state space.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 258, + 506, + 326 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 341, + 290, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 293, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 293, + 354 + ], + "score": 1.0, + "content": "5.4 CASE STUDY: ROBOT ARM CONTROL", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "We conduct a real-world case study on controlling robot arms to move objects on a shelf. On this", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 373, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 506, + 384 + ], + "score": 1.0, + "content": "representative real-time task, we demonstrate the advantages of the NEXT in terms of the wall-clock.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 361, + 506, + 384 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 333, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 334, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 334, + 402 + ], + "score": 1.0, + "content": "In each planning task, there is a shelf of multiple levels,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 400, + 334, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 334, + 412 + ], + "score": 1.0, + "content": "with each level horizontally divided into multiple bins.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 411, + 333, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 333, + 423 + ], + "score": 1.0, + "content": "The task is to plan a path from a location in one bin", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 422, + 333, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 333, + 434 + ], + "score": 1.0, + "content": "to another, i.e., the end effectors of the start and goal", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 433, + 334, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 334, + 446 + ], + "score": 1.0, + "content": "configurations are in different bins. The heights of levels,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 443, + 333, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 333, + 457 + ], + "score": 1.0, + "content": "widths of bins, and the start and goal are randomly drawn", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 455, + 333, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 333, + 467 + ], + "score": 1.0, + "content": "from some fixed distribution. Different from previous", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 466, + 334, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 334, + 478 + ], + "score": 1.0, + "content": "experiments, the base of the robot is fixed. We consider", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 477, + 333, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 120, + 489 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 477, + 143, + 488 + ], + "score": 0.61, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 477, + 184, + 489 + ], + "score": 1.0, + "content": "instead of", + "type": "text" + }, + { + "bbox": [ + 184, + 477, + 209, + 488 + ], + "score": 0.28, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 477, + 333, + 489 + ], + "score": 1.0, + "content": "as the imperfect expert in 3000", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 488, + 334, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 334, + 500 + ], + "score": 1.0, + "content": "training problems. We then evaluate the algorithm on a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 499, + 334, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 334, + 510 + ], + "score": 1.0, + "content": "separated 1000 testing problems. We compare NEXT(-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 510, + 334, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 218, + 522 + ], + "score": 1.0, + "content": "KS) with the highly tuned", + "type": "text" + }, + { + "bbox": [ + 219, + 510, + 242, + 520 + ], + "score": 0.52, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 242, + 510, + 334, + 522 + ], + "score": 1.0, + "content": "and RRT* in OMPL,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 521, + 334, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 334, + 533 + ], + "score": 1.0, + "content": "and also CVAE-plan and Reinforce-plan in Figure 8. As", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 532, + 333, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 333, + 544 + ], + "score": 1.0, + "content": "seen from the visualization of the found paths in Figure 9,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 542, + 333, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 333, + 555 + ], + "score": 1.0, + "content": "this is a very difficult task. Our NEXT outperforms the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 553, + 333, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 333, + 567 + ], + "score": 1.0, + "content": "baselines by a large margin, requiring only 1 second to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 565, + 333, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 333, + 577 + ], + "score": 1.0, + "content": "reach the same success rate as running 50 seconds of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 575, + 133, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 128, + 587 + ], + "score": 0.52, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 575, + 133, + 587 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 389, + 334, + 587 + ] + }, + { + "type": "image", + "bbox": [ + 370, + 389, + 459, + 540 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 370, + 389, + 459, + 540 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 370, + 389, + 459, + 540 + ], + "spans": [ + { + "bbox": [ + 370, + 389, + 459, + 540 + ], + "score": 0.467, + "type": "image", + "image_path": "2c27831a81842e20acfb22f33643b38f7e0d636dac989887fae6e6036e1293d0.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 370, + 389, + 459, + 540 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 344, + 543, + 505, + 583 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 343, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 343, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "Figure 9: The collision-free path produced", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 343, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 343, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "by NEXT for robot arm planning. The start", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 343, + 563, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 343, + 563, + 505, + 574 + ], + "score": 1.0, + "content": "and goal configurations have end-effectors in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 344, + 573, + 439, + 583 + ], + "spans": [ + { + "bbox": [ + 344, + 573, + 439, + 583 + ], + "score": 1.0, + "content": "different bins of the shelf.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36.5 + } + ], + "index": 33.25 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 504, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 592, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 506, + 606 + ], + "score": 1.0, + "content": "Due to space limits, we put details of the experiment setups, more results and analysis in Appendix F.4.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 592, + 506, + 606 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 621, + 195, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 197, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 197, + 636 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "In this paper, we propose a self-improving planner, Neural EXploration-EXploitation Trees (NEXT),", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "which can generalize and achieve better performance with experiences accumulated. The algorithm", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "achieves a delicate balance between exploration-exploitation via our carefully designed UCB-type", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "expansion operation. To obtain the generalizable ability across different problems, we proposed a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "new parametrization for the value function and policy, which captures the Bellman recursive structure", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "in the high-dimensional continuous state and action space. We demonstrate the power of the proposed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "algorithm by outperforming previous state-of-the-art planners with significant margins on planning", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 721, + 299, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 299, + 733 + ], + "score": 1.0, + "content": "problems in a variety of different environments.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 643, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 219, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 220, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 82, + 220, + 96 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENT", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "We thank the Google Research Brain team members for helpful thoughts and discussions as well", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 506, + 130 + ], + "score": 1.0, + "content": "as the anonymous reviewers for their insightful comments and suggestions. This work is supported", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 506, + 140 + ], + "score": 1.0, + "content": "in part by NSF grants CDS&E-1900017 D3SC, CCF-1836936 FMitF, IIS-1841351, CAREER IIS-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 140, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 506, + 151 + ], + "score": 1.0, + "content": "1350983 to L.S, and by NSF grants BIGDATA 1840866, CAREER 1841569, TRIPODS 1740735,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 163 + ], + "score": 1.0, + "content": "DARPA-PA-18-02-09-QED-RML-FP-003, an Alfred P Sloan Fellowship, a PECASE award to H.L.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 107, + 178, + 175, + 190 + ], + "lines": [ + { + "bbox": [ + 106, + 178, + 176, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 176, + 191 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 108, + 196, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 107, + 196, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 107, + 196, + 506, + 209 + ], + "score": 1.0, + "content": "Auer, P., Cesa-Bianchi, N., and Fischer, P. Finite-time analysis of the multiarmed bandit problem.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 208, + 289, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 208, + 289, + 221 + ], + "score": 1.0, + "content": "Machine learning, 47(2-3):235–256, 2002.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 227, + 504, + 250 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 241 + ], + "score": 1.0, + "content": "Mohak Bhardwaj, Sanjiban Choudhury, and Sebastian Scherer. Learning heuristic search via imitation.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 238, + 279, + 250 + ], + "spans": [ + { + "bbox": [ + 115, + 238, + 279, + 250 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1707.03034, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 257, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 256, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 506, + 271 + ], + "score": 1.0, + "content": "Boor, V., Overmars, M. H., and Van Der Stappen, A. F. The gaussian sampling strategy for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 116, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "probabilistic roadmap planners. In Robotics and automation, 1999. proceedings. 1999 ieee", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 280, + 390, + 292 + ], + "spans": [ + { + "bbox": [ + 116, + 280, + 390, + 292 + ], + "score": 1.0, + "content": "international conference on, volume 2, pp. 1018–1023. IEEE, 1999.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 108, + 299, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 311 + ], + "score": 1.0, + "content": "Burns, B. and Brock, O. Sampling-based motion planning using predictive models. In Robotics and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 308, + 507, + 324 + ], + "spans": [ + { + "bbox": [ + 114, + 308, + 507, + 324 + ], + "score": 1.0, + "content": "Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference on, pp.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 320, + 223, + 333 + ], + "spans": [ + { + "bbox": [ + 115, + 320, + 223, + 333 + ], + "score": 1.0, + "content": "3120–3125. IEEE, 2005a.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 339, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 353 + ], + "score": 1.0, + "content": "Burns, B. and Brock, O. Toward optimal configuration space sampling. In Robotics: Science and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 351, + 275, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 275, + 363 + ], + "score": 1.0, + "content": "Systems, pp. 105–112. Citeseer, 2005b.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 370, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "Bowen, Chris, and Ron Alterovitz. Closed-loop global motion planning for reactive execution", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 380, + 507, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 507, + 396 + ], + "score": 1.0, + "content": "of learned tasks. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 393, + 189, + 404 + ], + "spans": [ + { + "bbox": [ + 116, + 393, + 189, + 404 + ], + "score": 1.0, + "content": "1754-1760, 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 506, + 424 + ], + "score": 1.0, + "content": "Cadena Cesar, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 423, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 116, + 423, + 506, + 435 + ], + "score": 1.0, + "content": "and John J. Leonard. Past, present, and future of simultaneous localization and mapping: Toward", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 434, + 405, + 446 + ], + "spans": [ + { + "bbox": [ + 116, + 434, + 405, + 446 + ], + "score": 1.0, + "content": "the robust-perception age. IEEE Transactions on Robotics, 32(6), 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "Sanjiban Choudhury, Mohak Bhardwaj, Sankalp Arora, Ashish Kapoor, Gireeja Ranade, Sebastian", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "Scherer, and Debadeepta Dey. Data-driven planning via imitation learning. The International", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 475, + 356, + 487 + ], + "spans": [ + { + "bbox": [ + 115, + 475, + 356, + 487 + ], + "score": 1.0, + "content": "Journal of Robotics Research, 37(13-14):1632–1672, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 493, + 506, + 528 + ], + "lines": [ + { + "bbox": [ + 107, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 107, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "Chu, W., Li, L., Reyzin, L., and Schapire, R. Contextual bandits with linear payoff functions. In", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 504, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 114, + 504, + 507, + 519 + ], + "score": 1.0, + "content": "Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 516, + 181, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 181, + 528 + ], + "score": 1.0, + "content": "208–214, 2011.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 504, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "Couëtoux, A., Hoock, J.-B., Sokolovska, N., Teytaud, O., and Bonnard, N. Continuous upper", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 546, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 546, + 506, + 561 + ], + "score": 1.0, + "content": "confidence trees. In International Conference on Learning and Intelligent Optimization, pp.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 558, + 221, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 221, + 570 + ], + "score": 1.0, + "content": "433–445. Springer, 2011.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 104, + 576, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 591 + ], + "score": 1.0, + "content": "Rosen Diankov. Automated Construction of Robotic Manipulation Programs. PhD thesis, Carnegie", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 588, + 325, + 600 + ], + "spans": [ + { + "bbox": [ + 116, + 588, + 325, + 600 + ], + "score": 1.0, + "content": "Mellon University, Robotics Institute, August 2010.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 607, + 504, + 631 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "Rosen Diankov and James Kuffner. Randomized statistical path planning. In 2007 IEEE/RSJ", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 618, + 446, + 631 + ], + "spans": [ + { + "bbox": [ + 115, + 618, + 446, + 631 + ], + "score": 1.0, + "content": "International Conference on Intelligent Robots and Systems, pp. 1–6. IEEE, 2007.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "Elbanhawi, M. and Simic, M. Sampling-based robot motion planning: A review. Ieee access, 2:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 649, + 172, + 659 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 172, + 659 + ], + "score": 1.0, + "content": "56–77, 2014.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 682 + ], + "score": 1.0, + "content": "Finney, S., Kaelbling, L. P., and Lozano-Perez, T. Predicting partial paths from planning problem", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 680, + 326, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 326, + 691 + ], + "score": 1.0, + "content": "parameters. In Robotics Science and Systems, 2007.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Fulgenzi, Chiara and Tay, Christopher and Spalanzani, Anne and Laugier, Christian Probabilistic", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "navigation in dynamic environment using rapidly-exploring random trees and gaussian processes", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 503, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 503, + 733 + ], + "score": 1.0, + "content": "In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1056–1062, 2008.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 219, + 93 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 220, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 82, + 220, + 96 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENT", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "We thank the Google Research Brain team members for helpful thoughts and discussions as well", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 506, + 130 + ], + "score": 1.0, + "content": "as the anonymous reviewers for their insightful comments and suggestions. This work is supported", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 506, + 140 + ], + "score": 1.0, + "content": "in part by NSF grants CDS&E-1900017 D3SC, CCF-1836936 FMitF, IIS-1841351, CAREER IIS-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 140, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 506, + 151 + ], + "score": 1.0, + "content": "1350983 to L.S, and by NSF grants BIGDATA 1840866, CAREER 1841569, TRIPODS 1740735,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 163 + ], + "score": 1.0, + "content": "DARPA-PA-18-02-09-QED-RML-FP-003, an Alfred P Sloan Fellowship, a PECASE award to H.L.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 106, + 506, + 163 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 178, + 175, + 190 + ], + "lines": [ + { + "bbox": [ + 106, + 178, + 176, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 176, + 191 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 108, + 196, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 107, + 196, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 107, + 196, + 506, + 209 + ], + "score": 1.0, + "content": "Auer, P., Cesa-Bianchi, N., and Fischer, P. Finite-time analysis of the multiarmed bandit problem.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 208, + 289, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 208, + 289, + 221 + ], + "score": 1.0, + "content": "Machine learning, 47(2-3):235–256, 2002.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 107, + 196, + 506, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 227, + 504, + 250 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 241 + ], + "score": 1.0, + "content": "Mohak Bhardwaj, Sanjiban Choudhury, and Sebastian Scherer. Learning heuristic search via imitation.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 238, + 279, + 250 + ], + "spans": [ + { + "bbox": [ + 115, + 238, + 279, + 250 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1707.03034, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 226, + 506, + 250 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 257, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 256, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 256, + 506, + 271 + ], + "score": 1.0, + "content": "Boor, V., Overmars, M. H., and Van Der Stappen, A. F. The gaussian sampling strategy for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 116, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "probabilistic roadmap planners. In Robotics and automation, 1999. proceedings. 1999 ieee", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 280, + 390, + 292 + ], + "spans": [ + { + "bbox": [ + 116, + 280, + 390, + 292 + ], + "score": 1.0, + "content": "international conference on, volume 2, pp. 1018–1023. IEEE, 1999.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 256, + 506, + 292 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 299, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 311 + ], + "score": 1.0, + "content": "Burns, B. and Brock, O. Sampling-based motion planning using predictive models. In Robotics and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 114, + 308, + 507, + 324 + ], + "spans": [ + { + "bbox": [ + 114, + 308, + 507, + 324 + ], + "score": 1.0, + "content": "Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International Conference on, pp.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 320, + 223, + 333 + ], + "spans": [ + { + "bbox": [ + 115, + 320, + 223, + 333 + ], + "score": 1.0, + "content": "3120–3125. IEEE, 2005a.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 299, + 507, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 339, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 353 + ], + "score": 1.0, + "content": "Burns, B. and Brock, O. Toward optimal configuration space sampling. In Robotics: Science and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 351, + 275, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 275, + 363 + ], + "score": 1.0, + "content": "Systems, pp. 105–112. Citeseer, 2005b.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 339, + 505, + 363 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 370, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "Bowen, Chris, and Ron Alterovitz. Closed-loop global motion planning for reactive execution", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 380, + 507, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 507, + 396 + ], + "score": 1.0, + "content": "of learned tasks. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 393, + 189, + 404 + ], + "spans": [ + { + "bbox": [ + 116, + 393, + 189, + 404 + ], + "score": 1.0, + "content": "1754-1760, 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 370, + 507, + 404 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 411, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 412, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 412, + 506, + 424 + ], + "score": 1.0, + "content": "Cadena Cesar, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 423, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 116, + 423, + 506, + 435 + ], + "score": 1.0, + "content": "and John J. Leonard. Past, present, and future of simultaneous localization and mapping: Toward", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 434, + 405, + 446 + ], + "spans": [ + { + "bbox": [ + 116, + 434, + 405, + 446 + ], + "score": 1.0, + "content": "the robust-perception age. IEEE Transactions on Robotics, 32(6), 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 412, + 506, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "Sanjiban Choudhury, Mohak Bhardwaj, Sankalp Arora, Ashish Kapoor, Gireeja Ranade, Sebastian", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 465, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 465, + 505, + 477 + ], + "score": 1.0, + "content": "Scherer, and Debadeepta Dey. Data-driven planning via imitation learning. The International", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 475, + 356, + 487 + ], + "spans": [ + { + "bbox": [ + 115, + 475, + 356, + 487 + ], + "score": 1.0, + "content": "Journal of Robotics Research, 37(13-14):1632–1672, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 453, + 505, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 493, + 506, + 528 + ], + "lines": [ + { + "bbox": [ + 107, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 107, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "Chu, W., Li, L., Reyzin, L., and Schapire, R. Contextual bandits with linear payoff functions. In", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 504, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 114, + 504, + 507, + 519 + ], + "score": 1.0, + "content": "Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 516, + 181, + 528 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 181, + 528 + ], + "score": 1.0, + "content": "208–214, 2011.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 107, + 494, + 507, + 528 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 504, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "Couëtoux, A., Hoock, J.-B., Sokolovska, N., Teytaud, O., and Bonnard, N. Continuous upper", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 546, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 546, + 506, + 561 + ], + "score": 1.0, + "content": "confidence trees. In International Conference on Learning and Intelligent Optimization, pp.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 558, + 221, + 570 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 221, + 570 + ], + "score": 1.0, + "content": "433–445. Springer, 2011.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 534, + 506, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 576, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 591 + ], + "score": 1.0, + "content": "Rosen Diankov. Automated Construction of Robotic Manipulation Programs. PhD thesis, Carnegie", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 588, + 325, + 600 + ], + "spans": [ + { + "bbox": [ + 116, + 588, + 325, + 600 + ], + "score": 1.0, + "content": "Mellon University, Robotics Institute, August 2010.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 575, + 505, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 607, + 504, + 631 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "Rosen Diankov and James Kuffner. Randomized statistical path planning. In 2007 IEEE/RSJ", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 618, + 446, + 631 + ], + "spans": [ + { + "bbox": [ + 115, + 618, + 446, + 631 + ], + "score": 1.0, + "content": "International Conference on Intelligent Robots and Systems, pp. 1–6. IEEE, 2007.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 608, + 505, + 631 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "Elbanhawi, M. and Simic, M. Sampling-based robot motion planning: A review. Ieee access, 2:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 649, + 172, + 659 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 172, + 659 + ], + "score": 1.0, + "content": "56–77, 2014.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 636, + 506, + 659 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 682 + ], + "score": 1.0, + "content": "Finney, S., Kaelbling, L. P., and Lozano-Perez, T. Predicting partial paths from planning problem", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 680, + 326, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 326, + 691 + ], + "score": 1.0, + "content": "parameters. In Robotics Science and Systems, 2007.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 666, + 505, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Fulgenzi, Chiara and Tay, Christopher and Spalanzani, Anne and Laugier, Christian Probabilistic", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "navigation in dynamic environment using rapidly-exploring random trees and gaussian processes", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 720, + 503, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 503, + 733 + ], + "score": 1.0, + "content": "In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1056–1062, 2008.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 698, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 64, + 507, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 377, + 96 + ], + "score": 1.0, + "content": "Gammell, J. D., Srinivasa, S. S., and Barfoot, T. D. Informed", + "type": "text" + }, + { + "bbox": [ + 378, + 83, + 394, + 93 + ], + "score": 0.37, + "content": "\\mathrm { { r r t } ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 82, + 506, + 96 + ], + "score": 1.0, + "content": ": Optimal sampling-based", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 113, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "path planning focused via direct sampling of an admissible ellipsoidal heuristic. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 214, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 214, + 117 + ], + "score": 1.0, + "content": "arXiv:1404.2334, 2014.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "Gammell, J. D., Srinivasa, S. S., and Barfoot, T. D. Batch informed trees (bit*): Sampling-based", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "score": 1.0, + "content": "optimal planning via the heuristically guided search of implicit random geometric graphs. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 507, + 159 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 507, + 159 + ], + "score": 1.0, + "content": "Robotics and Automation (ICRA), 2015 IEEE International Conference on, pp. 3067–3074. IEEE,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 143, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 143, + 168 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 507, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 507, + 187 + ], + "score": 1.0, + "content": "Guez, Arthur and Weber, Théophane and Antonoglou, Ioannis and Simonyan, Karen and Vinyals,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "Oriol and Wierstra, Daan and Munos, Rémi and Silver, David Learning to search with mctsnets", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 195, + 279, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 279, + 209 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1802.04697, 2018.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 226, + 507, + 240 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 507, + 240 + ], + "score": 1.0, + "content": "minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100–107, 1968.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 507, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 507, + 257 + ], + "score": 1.0, + "content": "Hsu, D., Latombe, J.-C., and Motwani, R. Path planning in expansive configuration spaces. In", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 114, + 253, + 506, + 267 + ], + "score": 1.0, + "content": "Robotics and Automation, 1997. Proceedings., 1997 IEEE International Conference on, volume 3,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 113, + 265, + 234, + 279 + ], + "spans": [ + { + "bbox": [ + 113, + 265, + 234, + 279 + ], + "score": 1.0, + "content": "pp. 2719–2726. IEEE, 1997.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "score": 1.0, + "content": "Hsu, D., Jiang, T., Reif, J., and Sun, Z. The bridge test for sampling narrow passages with probabilistic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 294, + 507, + 309 + ], + "spans": [ + { + "bbox": [ + 113, + 294, + 507, + 309 + ], + "score": 1.0, + "content": "roadmap planners. In Robotics and Automation, 2003. Proceedings. ICRA’03. IEEE International", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 113, + 305, + 339, + 320 + ], + "spans": [ + { + "bbox": [ + 113, + 305, + 339, + 320 + ], + "score": 1.0, + "content": "Conference on, volume 3, pp. 4420–4426. IEEE, 2003.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 322, + 507, + 340 + ], + "spans": [ + { + "bbox": [ + 104, + 322, + 507, + 340 + ], + "score": 1.0, + "content": "Hsu, D., Sánchez-Ante, G., and Sun, Z. Hybrid prm sampling with a cost-sensitive adaptive strategy.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 115, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "In Robotics and Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 345, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 114, + 345, + 296, + 361 + ], + "score": 1.0, + "content": "Conference on, pp. 3874–3880. IEEE, 2005.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 362, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 506, + 380 + ], + "score": 1.0, + "content": "Huh, Jinwook and Lee, Daniel Efficient Sampling With Q-Learning to Guide Rapidly Exploring", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 375, + 424, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 424, + 389 + ], + "score": 1.0, + "content": "Random Trees. IEEE Robotics and Automation Letters, 3:3868–3875, 2018.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 393, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 408 + ], + "score": 1.0, + "content": "Ichter, B., Harrison, J., and Pavone, M. Learning sampling distributions for robot motion planning. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 406, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 506, + 417 + ], + "score": 1.0, + "content": "2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 7087–7094. IEEE,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 416, + 143, + 429 + ], + "spans": [ + { + "bbox": [ + 115, + 416, + 143, + 429 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 434, + 507, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 507, + 448 + ], + "score": 1.0, + "content": "Karaman, S. and Frazzoli, E. Sampling-based algorithms for optimal motion planning. The interna-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 446, + 347, + 458 + ], + "spans": [ + { + "bbox": [ + 115, + 446, + 347, + 458 + ], + "score": 1.0, + "content": "tional journal of robotics research, 30(7):846–894, 2011.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 462, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 507, + 478 + ], + "score": 1.0, + "content": "Karkus, P., Hsu, D., and Lee, W. S. Qmdp-net: Deep learning for planning under partial observability.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 474, + 430, + 488 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 430, + 488 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 4694–4704, 2017.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "Kavraki, L. E., Svestka, P., Latombe, J.-C., and Overmars, M. H. Probabilistic roadmaps for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 113, + 504, + 507, + 517 + ], + "spans": [ + { + "bbox": [ + 113, + 504, + 507, + 517 + ], + "score": 1.0, + "content": "path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 516, + 218, + 528 + ], + "spans": [ + { + "bbox": [ + 114, + 516, + 218, + 528 + ], + "score": 1.0, + "content": "Automation, 12(4), 1996.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "Kim, B., Kaelbling, L. P., and Lozano-Pérez, T. Guiding search in continuous state-action spaces by", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 545, + 388, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 388, + 558 + ], + "score": 1.0, + "content": "learning an action sampler from off-target search experience. 2018.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "Kocsis, L. and Szepesvári, C. Bandit based monte-carlo planning. In European conference on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 573, + 309, + 587 + ], + "spans": [ + { + "bbox": [ + 114, + 573, + 309, + 587 + ], + "score": 1.0, + "content": "machine learning, pp. 282–293. Springer, 2006.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "Krause, A. and Ong, C. S. Contextual gaussian process bandit optimization. In Advances in Neural", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 603, + 340, + 617 + ], + "spans": [ + { + "bbox": [ + 114, + 603, + 340, + 617 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 2447–2455, 2011.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "Kuo, Yen-Ling and Barbu, Andrei and Katz, Boris Deep sequential models for sampling-based", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 632, + 507, + 646 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 507, + 646 + ], + "score": 1.0, + "content": "planning In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 6490–6497,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 641, + 143, + 658 + ], + "spans": [ + { + "bbox": [ + 115, + 641, + 143, + 658 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 663, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 506, + 674 + ], + "score": 1.0, + "content": "Langford, J. and Zhang, T. The epoch-greedy algorithm for multi-armed bandits with side information.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 672, + 417, + 687 + ], + "spans": [ + { + "bbox": [ + 114, + 672, + 417, + 687 + ], + "score": 1.0, + "content": "In Advances in neural information processing systems, pp. 817–824, 2008.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 692, + 446, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 446, + 704 + ], + "score": 1.0, + "content": "LaValle, S. M. Rapidly-exploring random trees: A new tool for path planning. 1998.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 709, + 507, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 507, + 723 + ], + "score": 1.0, + "content": "Lee, L., Parisotto, E., Chaplot, D. S., Xing, E., and Salakhutdinov, R. Gated path planning networks.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 279, + 734 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 279, + 734 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1806.06408, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 23 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 64, + 507, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 377, + 96 + ], + "score": 1.0, + "content": "Gammell, J. D., Srinivasa, S. S., and Barfoot, T. D. Informed", + "type": "text" + }, + { + "bbox": [ + 378, + 83, + 394, + 93 + ], + "score": 0.37, + "content": "\\mathrm { { r r t } ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 82, + 506, + 96 + ], + "score": 1.0, + "content": ": Optimal sampling-based", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 113, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "path planning focused via direct sampling of an admissible ellipsoidal heuristic. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 214, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 214, + 117 + ], + "score": 1.0, + "content": "arXiv:1404.2334, 2014.", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "Gammell, J. D., Srinivasa, S. S., and Barfoot, T. D. Batch informed trees (bit*): Sampling-based", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "score": 1.0, + "content": "optimal planning via the heuristically guided search of implicit random geometric graphs. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 145, + 507, + 159 + ], + "spans": [ + { + "bbox": [ + 115, + 145, + 507, + 159 + ], + "score": 1.0, + "content": "Robotics and Automation (ICRA), 2015 IEEE International Conference on, pp. 3067–3074. IEEE,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 143, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 143, + 168 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 172, + 507, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 507, + 187 + ], + "score": 1.0, + "content": "Guez, Arthur and Weber, Théophane and Antonoglou, Ioannis and Simonyan, Karen and Vinyals,", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "Oriol and Wierstra, Daan and Munos, Rémi and Silver, David Learning to search with mctsnets", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 195, + 279, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 279, + 209 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1802.04697, 2018.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 226 + ], + "score": 1.0, + "content": "Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 226, + 507, + 240 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 507, + 240 + ], + "score": 1.0, + "content": "minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100–107, 1968.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 507, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 507, + 257 + ], + "score": 1.0, + "content": "Hsu, D., Latombe, J.-C., and Motwani, R. Path planning in expansive configuration spaces. In", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 114, + 253, + 506, + 267 + ], + "score": 1.0, + "content": "Robotics and Automation, 1997. Proceedings., 1997 IEEE International Conference on, volume 3,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 113, + 265, + 234, + 279 + ], + "spans": [ + { + "bbox": [ + 113, + 265, + 234, + 279 + ], + "score": 1.0, + "content": "pp. 2719–2726. IEEE, 1997.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 298 + ], + "score": 1.0, + "content": "Hsu, D., Jiang, T., Reif, J., and Sun, Z. The bridge test for sampling narrow passages with probabilistic", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 294, + 507, + 309 + ], + "spans": [ + { + "bbox": [ + 113, + 294, + 507, + 309 + ], + "score": 1.0, + "content": "roadmap planners. In Robotics and Automation, 2003. Proceedings. ICRA’03. IEEE International", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 113, + 305, + 339, + 320 + ], + "spans": [ + { + "bbox": [ + 113, + 305, + 339, + 320 + ], + "score": 1.0, + "content": "Conference on, volume 3, pp. 4420–4426. IEEE, 2003.", + "type": "text" + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 322, + 507, + 340 + ], + "spans": [ + { + "bbox": [ + 104, + 322, + 507, + 340 + ], + "score": 1.0, + "content": "Hsu, D., Sánchez-Ante, G., and Sun, Z. Hybrid prm sampling with a cost-sensitive adaptive strategy.", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 115, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "In Robotics and Automation, 2005. ICRA 2005. Proceedings of the 2005 IEEE International", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 345, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 114, + 345, + 296, + 361 + ], + "score": 1.0, + "content": "Conference on, pp. 3874–3880. IEEE, 2005.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 362, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 506, + 380 + ], + "score": 1.0, + "content": "Huh, Jinwook and Lee, Daniel Efficient Sampling With Q-Learning to Guide Rapidly Exploring", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 375, + 424, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 424, + 389 + ], + "score": 1.0, + "content": "Random Trees. IEEE Robotics and Automation Letters, 3:3868–3875, 2018.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 393, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 408 + ], + "score": 1.0, + "content": "Ichter, B., Harrison, J., and Pavone, M. Learning sampling distributions for robot motion planning. In", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 406, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 506, + 417 + ], + "score": 1.0, + "content": "2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 7087–7094. IEEE,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 416, + 143, + 429 + ], + "spans": [ + { + "bbox": [ + 115, + 416, + 143, + 429 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 434, + 507, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 507, + 448 + ], + "score": 1.0, + "content": "Karaman, S. and Frazzoli, E. Sampling-based algorithms for optimal motion planning. The interna-", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 446, + 347, + 458 + ], + "spans": [ + { + "bbox": [ + 115, + 446, + 347, + 458 + ], + "score": 1.0, + "content": "tional journal of robotics research, 30(7):846–894, 2011.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 462, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 507, + 478 + ], + "score": 1.0, + "content": "Karkus, P., Hsu, D., and Lee, W. S. Qmdp-net: Deep learning for planning under partial observability.", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 474, + 430, + 488 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 430, + 488 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 4694–4704, 2017.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "Kavraki, L. E., Svestka, P., Latombe, J.-C., and Overmars, M. H. Probabilistic roadmaps for", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 504, + 507, + 517 + ], + "spans": [ + { + "bbox": [ + 113, + 504, + 507, + 517 + ], + "score": 1.0, + "content": "path planning in high-dimensional configuration spaces. IEEE Transactions on Robotics and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 516, + 218, + 528 + ], + "spans": [ + { + "bbox": [ + 114, + 516, + 218, + 528 + ], + "score": 1.0, + "content": "Automation, 12(4), 1996.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "Kim, B., Kaelbling, L. P., and Lozano-Pérez, T. Guiding search in continuous state-action spaces by", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 545, + 388, + 558 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 388, + 558 + ], + "score": 1.0, + "content": "learning an action sampler from off-target search experience. 2018.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "Kocsis, L. and Szepesvári, C. Bandit based monte-carlo planning. In European conference on", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 573, + 309, + 587 + ], + "spans": [ + { + "bbox": [ + 114, + 573, + 309, + 587 + ], + "score": 1.0, + "content": "machine learning, pp. 282–293. Springer, 2006.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 505, + 605 + ], + "score": 1.0, + "content": "Krause, A. and Ong, C. S. Contextual gaussian process bandit optimization. In Advances in Neural", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 603, + 340, + 617 + ], + "spans": [ + { + "bbox": [ + 114, + 603, + 340, + 617 + ], + "score": 1.0, + "content": "Information Processing Systems, pp. 2447–2455, 2011.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "Kuo, Yen-Ling and Barbu, Andrei and Katz, Boris Deep sequential models for sampling-based", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 632, + 507, + 646 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 507, + 646 + ], + "score": 1.0, + "content": "planning In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 6490–6497,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 641, + 143, + 658 + ], + "spans": [ + { + "bbox": [ + 115, + 641, + 143, + 658 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 663, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 506, + 674 + ], + "score": 1.0, + "content": "Langford, J. and Zhang, T. The epoch-greedy algorithm for multi-armed bandits with side information.", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 672, + 417, + 687 + ], + "spans": [ + { + "bbox": [ + 114, + 672, + 417, + 687 + ], + "score": 1.0, + "content": "In Advances in neural information processing systems, pp. 817–824, 2008.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 692, + 446, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 446, + 704 + ], + "score": 1.0, + "content": "LaValle, S. M. Rapidly-exploring random trees: A new tool for path planning. 1998.", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 709, + 507, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 507, + 723 + ], + "score": 1.0, + "content": "Lee, L., Parisotto, E., Chaplot, D. S., Xing, E., and Salakhutdinov, R. Gated path planning networks.", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 721, + 279, + 734 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 279, + 734 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1806.06408, 2018.", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + } + ], + "index": 23, + "bbox_fs": [ + 104, + 82, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 507, + 97 + ], + "score": 1.0, + "content": "Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 284, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 284, + 105 + ], + "score": 1.0, + "content": "Machine learning, 8(3-4):293–321, 1992.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 125 + ], + "score": 1.0, + "content": "Paxton, Chris and Raman, Vasumathi and Hager, Gregory D and Kobilarov, Marin Combining neural", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "networks and tree search for task and motion planning in challenging environments In IEEE/RSJ", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 480, + 147 + ], + "score": 1.0, + "content": "International Conference on Intelligent Robots and Systems (IROS), pp. 6059–6066, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 505, + 166 + ], + "score": 1.0, + "content": "Sven Mikael Persson and Inna Sharf. Sampling-based a* algorithm for robot path-planning. The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 398, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 398, + 176 + ], + "score": 1.0, + "content": "International Journal of Robotics Research, 33(13):1683–1708, 2014.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 108, + 182, + 504, + 217 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 197 + ], + "score": 1.0, + "content": "Phillips, J. M., Bedrossian, N., and Kavraki, L. E. Guided expansive spaces trees: A search strategy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 505, + 206 + ], + "score": 1.0, + "content": "for motion-and cost-constrained state spaces. In IEEE International Conference on Robotics and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 205, + 257, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 205, + 257, + 217 + ], + "score": 1.0, + "content": "Automation, pp. 3968–3973, 2004.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 503, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 238 + ], + "score": 1.0, + "content": "Mike Phillips, Benjamin J Cohen, Sachin Chitta, and Maxim Likhachev. E-graphs: Bootstrapping", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 494, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 494, + 248 + ], + "score": 1.0, + "content": "planning with experience graphs. In Robotics: Science and Systems, volume 5, pp. 110, 2012.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 253, + 503, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 268 + ], + "score": 1.0, + "content": "Qureshi, Ahmed H and Simeonov, Anthony and Bency, Mayur J and Yip, Michael C. Motion planning", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 440, + 277 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 440, + 277 + ], + "score": 1.0, + "content": "networks. In IEEE International Conference on Robotics and Automation, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 283, + 504, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 298 + ], + "score": 1.0, + "content": "Reif, J. H. Complexity of the mover’s problem and generalizations. In Foundations of Computer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 399, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 399, + 307 + ], + "score": 1.0, + "content": "Science, 1979., 20th Annual Symposium on, pp. 421–427. IEEE, 1979.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 504, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "Rickert, Markus and Brock, Oliver and Knoll, Alois. Balancing exploration and exploitation in", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 325, + 507, + 338 + ], + "spans": [ + { + "bbox": [ + 116, + 325, + 507, + 338 + ], + "score": 1.0, + "content": "motion planning In IEEE International Conference on Robotics and Automation, pp. 2812–2817.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 336, + 168, + 348 + ], + "spans": [ + { + "bbox": [ + 116, + 336, + 168, + 348 + ], + "score": 1.0, + "content": "IEEE, 2008.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 504, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 368 + ], + "score": 1.0, + "content": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 366, + 254, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 366, + 254, + 378 + ], + "score": 1.0, + "content": "preprint arXiv:1511.05952, 2015.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 105, + 384, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "Shkolnik, A., Walter, M., and Tedrake, R. Reachability-guided sampling for planning under differen-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 394, + 507, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 507, + 410 + ], + "score": 1.0, + "content": "tial constraints. In Robotics and Automation, 2009. ICRA’09. IEEE International Conference on,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 407, + 233, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 233, + 418 + ], + "score": 1.0, + "content": "pp. 2859–2865. IEEE, 2009.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 506, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "Silver, David and Hubert, Thomas and Schrittwieser, Julian and Antonoglou, Ioannis and Lai,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 438, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 116, + 438, + 505, + 448 + ], + "score": 1.0, + "content": "Matthew and Guez, Arthur and Lanctot, Marc and Sifre, Laurent and Kumaran, Dharshan and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 115, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "Graepel, Thore and others Mastering chess and shogi by self-play with a general reinforcement", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 459, + 359, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 359, + 471 + ], + "score": 1.0, + "content": "learning algorithm. arXiv preprint arXiv:1712.01815, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 104, + 477, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "Sohn, K., Lee, H., and Yan, X. Learning structured output representation using deep conditional", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 488, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 507, + 502 + ], + "score": 1.0, + "content": "generative models. In Advances in Neural Information Processing Systems, pp. 3483–3491, 2015.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 105, + 507, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 521 + ], + "score": 1.0, + "content": "Jialin Song, Ravi Lanka, Albert Zhao, Yisong Yue, and Masahiro Ono. Learning to search via", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 519, + 375, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 519, + 375, + 531 + ], + "score": 1.0, + "content": "retrospective imitation. arXiv preprint arXiv:1804.00846, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 104, + 537, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "Srinivas, N., Krause, A., Kakade, S. M., and Seeger, M. Gaussian process optimization in the bandit", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 549, + 450, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 549, + 450, + 561 + ], + "score": 1.0, + "content": "setting: No regret and experimental design. arXiv preprint arXiv:0912.3995, 2009.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 567, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "Ioan A. ¸Sucan, Mark Moll, and Lydia E. Kavraki. The Open Motion Planning Library. IEEE Robotics", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 578, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 116, + 578, + 507, + 591 + ], + "score": 1.0, + "content": "& Automation Magazine, 19(4):72–82, December 2012. doi: 10.1109/MRA.2012.2205651.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 588, + 232, + 604 + ], + "spans": [ + { + "bbox": [ + 115, + 588, + 232, + 604 + ], + "score": 1.0, + "content": "http://ompl.kavrakilab.org.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 105, + 608, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "Tamar, A., Wu, Y., Thomas, G., Levine, S., and Abbeel, P. Value iteration networks. In Advances in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 619, + 369, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 369, + 632 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 2154–2162, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 105, + 638, + 504, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "Wang, Y., Audibert, J.-Y., and Munos, R. Algorithms for infinitely many-armed bandits. In Advances", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 649, + 379, + 662 + ], + "spans": [ + { + "bbox": [ + 115, + 649, + 379, + 662 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, pp. 1729–1736, 2009.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 683 + ], + "score": 1.0, + "content": "Ye, Gu and Alterovitz, Ron Guided motion planning. In Robotics research, pp. 291–6307. Springer,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 677, + 144, + 693 + ], + "spans": [ + { + "bbox": [ + 114, + 677, + 144, + 693 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 107, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Yee, T., Lisy, V., and Bowling, M. Monte carlo tree search in continuous action spaces with execution", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "uncertainty. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 719, + 302, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 302, + 733 + ], + "score": 1.0, + "content": "Intelligence, pp. 690–696. AAAI Press, 2016.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 507, + 97 + ], + "score": 1.0, + "content": "Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 284, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 284, + 105 + ], + "score": 1.0, + "content": "Machine learning, 8(3-4):293–321, 1992.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 104, + 80, + 507, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 111, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 125 + ], + "score": 1.0, + "content": "Paxton, Chris and Raman, Vasumathi and Hager, Gregory D and Kobilarov, Marin Combining neural", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "networks and tree search for task and motion planning in challenging environments In IEEE/RSJ", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 480, + 147 + ], + "score": 1.0, + "content": "International Conference on Intelligent Robots and Systems (IROS), pp. 6059–6066, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 111, + 506, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 505, + 166 + ], + "score": 1.0, + "content": "Sven Mikael Persson and Inna Sharf. Sampling-based a* algorithm for robot path-planning. The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 398, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 398, + 176 + ], + "score": 1.0, + "content": "International Journal of Robotics Research, 33(13):1683–1708, 2014.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 152, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 182, + 504, + 217 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 197 + ], + "score": 1.0, + "content": "Phillips, J. M., Bedrossian, N., and Kavraki, L. E. Guided expansive spaces trees: A search strategy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 505, + 206 + ], + "score": 1.0, + "content": "for motion-and cost-constrained state spaces. In IEEE International Conference on Robotics and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 205, + 257, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 205, + 257, + 217 + ], + "score": 1.0, + "content": "Automation, pp. 3968–3973, 2004.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 181, + 505, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 503, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 238 + ], + "score": 1.0, + "content": "Mike Phillips, Benjamin J Cohen, Sachin Chitta, and Maxim Likhachev. E-graphs: Bootstrapping", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 494, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 494, + 248 + ], + "score": 1.0, + "content": "planning with experience graphs. In Robotics: Science and Systems, volume 5, pp. 110, 2012.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 222, + 505, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 253, + 503, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 268 + ], + "score": 1.0, + "content": "Qureshi, Ahmed H and Simeonov, Anthony and Bency, Mayur J and Yip, Michael C. Motion planning", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 440, + 277 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 440, + 277 + ], + "score": 1.0, + "content": "networks. In IEEE International Conference on Robotics and Automation, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 252, + 505, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 283, + 504, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 282, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 298 + ], + "score": 1.0, + "content": "Reif, J. H. Complexity of the mover’s problem and generalizations. In Foundations of Computer", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 294, + 399, + 307 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 399, + 307 + ], + "score": 1.0, + "content": "Science, 1979., 20th Annual Symposium on, pp. 421–427. IEEE, 1979.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 282, + 505, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 504, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "Rickert, Markus and Brock, Oliver and Knoll, Alois. Balancing exploration and exploitation in", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 325, + 507, + 338 + ], + "spans": [ + { + "bbox": [ + 116, + 325, + 507, + 338 + ], + "score": 1.0, + "content": "motion planning In IEEE International Conference on Robotics and Automation, pp. 2812–2817.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 336, + 168, + 348 + ], + "spans": [ + { + "bbox": [ + 116, + 336, + 168, + 348 + ], + "score": 1.0, + "content": "IEEE, 2008.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 313, + 507, + 348 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 354, + 504, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 368 + ], + "score": 1.0, + "content": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 366, + 254, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 366, + 254, + 378 + ], + "score": 1.0, + "content": "preprint arXiv:1511.05952, 2015.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 354, + 505, + 378 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 384, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "Shkolnik, A., Walter, M., and Tedrake, R. Reachability-guided sampling for planning under differen-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 394, + 507, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 507, + 410 + ], + "score": 1.0, + "content": "tial constraints. In Robotics and Automation, 2009. ICRA’09. IEEE International Conference on,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 407, + 233, + 418 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 233, + 418 + ], + "score": 1.0, + "content": "pp. 2859–2865. IEEE, 2009.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 384, + 507, + 418 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 506, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "Silver, David and Hubert, Thomas and Schrittwieser, Julian and Antonoglou, Ioannis and Lai,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 438, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 116, + 438, + 505, + 448 + ], + "score": 1.0, + "content": "Matthew and Guez, Arthur and Lanctot, Marc and Sifre, Laurent and Kumaran, Dharshan and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 115, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "Graepel, Thore and others Mastering chess and shogi by self-play with a general reinforcement", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 459, + 359, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 359, + 471 + ], + "score": 1.0, + "content": "learning algorithm. arXiv preprint arXiv:1712.01815, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 426, + 506, + 471 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 477, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 106, + 477, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 490 + ], + "score": 1.0, + "content": "Sohn, K., Lee, H., and Yan, X. Learning structured output representation using deep conditional", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 488, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 507, + 502 + ], + "score": 1.0, + "content": "generative models. In Advances in Neural Information Processing Systems, pp. 3483–3491, 2015.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 477, + 507, + 502 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 507, + 504, + 531 + ], + "lines": [ + { + "bbox": [ + 105, + 506, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 521 + ], + "score": 1.0, + "content": "Jialin Song, Ravi Lanka, Albert Zhao, Yisong Yue, and Masahiro Ono. Learning to search via", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 519, + 375, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 519, + 375, + 531 + ], + "score": 1.0, + "content": "retrospective imitation. arXiv preprint arXiv:1804.00846, 2018.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 506, + 505, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 537, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "Srinivas, N., Krause, A., Kakade, S. M., and Seeger, M. Gaussian process optimization in the bandit", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 549, + 450, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 549, + 450, + 561 + ], + "score": 1.0, + "content": "setting: No regret and experimental design. arXiv preprint arXiv:0912.3995, 2009.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 537, + 505, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 567, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "Ioan A. ¸Sucan, Mark Moll, and Lydia E. Kavraki. The Open Motion Planning Library. IEEE Robotics", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 578, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 116, + 578, + 507, + 591 + ], + "score": 1.0, + "content": "& Automation Magazine, 19(4):72–82, December 2012. doi: 10.1109/MRA.2012.2205651.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 588, + 232, + 604 + ], + "spans": [ + { + "bbox": [ + 115, + 588, + 232, + 604 + ], + "score": 1.0, + "content": "http://ompl.kavrakilab.org.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 567, + 507, + 604 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 608, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 608, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 621 + ], + "score": 1.0, + "content": "Tamar, A., Wu, Y., Thomas, G., Levine, S., and Abbeel, P. Value iteration networks. In Advances in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 619, + 369, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 619, + 369, + 632 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 2154–2162, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 608, + 505, + 632 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 638, + 504, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "Wang, Y., Audibert, J.-Y., and Munos, R. Algorithms for infinitely many-armed bandits. In Advances", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 649, + 379, + 662 + ], + "spans": [ + { + "bbox": [ + 115, + 649, + 379, + 662 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, pp. 1729–1736, 2009.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 638, + 505, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 683 + ], + "score": 1.0, + "content": "Ye, Gu and Alterovitz, Ron Guided motion planning. In Robotics research, pp. 291–6307. Springer,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 677, + 144, + 693 + ], + "spans": [ + { + "bbox": [ + 114, + 677, + 144, + 693 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 666, + 506, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 107, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Yee, T., Lisy, V., and Bowling, M. Monte carlo tree search in continuous action spaces with execution", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "uncertainty. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 719, + 302, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 302, + 733 + ], + "score": 1.0, + "content": "Intelligence, pp. 690–696. AAAI Press, 2016.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 107, + 699, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Zhang, C., Huh, J., and Lee, D. D. Learning implicit sampling distributions for motion planning.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 279, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 279, + 105 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1806.01968, 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 506, + 145 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 507, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 507, + 126 + ], + "score": 1.0, + "content": "Zucker, M., Kuffner, J., and Bagnell, J. A. Adaptive workspace biasing for sampling-based planners.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 506, + 136 + ], + "score": 1.0, + "content": "In Robotics and Automation, 2008. ICRA 2008. IEEE International Conference on, pp. 3757–3762.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 133, + 168, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 133, + 168, + 146 + ], + "score": 1.0, + "content": "IEEE, 2008.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Zhang, C., Huh, J., and Lee, D. D. Learning implicit sampling distributions for motion planning.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 279, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 279, + 105 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1806.01968, 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 506, + 145 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 507, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 507, + 126 + ], + "score": 1.0, + "content": "Zucker, M., Kuffner, J., and Bagnell, J. A. Adaptive workspace biasing for sampling-based planners.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 506, + 136 + ], + "score": 1.0, + "content": "In Robotics and Automation, 2008. ICRA 2008. IEEE International Conference on, pp. 3757–3762.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 133, + 168, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 133, + 168, + 146 + ], + "score": 1.0, + "content": "IEEE, 2008.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 111, + 507, + 146 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 265, + 79, + 347, + 101 + ], + "lines": [ + { + "bbox": [ + 261, + 77, + 349, + 103 + ], + "spans": [ + { + "bbox": [ + 261, + 77, + 349, + 103 + ], + "score": 1.0, + "content": "Appendix", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 114, + 440, + 128 + ], + "lines": [ + { + "bbox": [ + 105, + 114, + 442, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 442, + 129 + ], + "score": 1.0, + "content": "A ILLUSTRATION OF THE DIFFICULTY IN PLANNING PROBLEMS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 506, + 174 + ], + "lines": [ + { + "bbox": [ + 104, + 139, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 139, + 505, + 153 + ], + "score": 1.0, + "content": "The Figure 10(a) illustrates a concrete planning problem for a stick robot in 2d workspace. With one", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "score": 1.0, + "content": "extra continuous action for rotation, the configuration state is visualized in Figure 10(b), which is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 163, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 286, + 175 + ], + "score": 1.0, + "content": "highly irregular and unknown to the planner.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 158, + 182, + 466, + 310 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 158, + 182, + 466, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 158, + 182, + 466, + 310 + ], + "spans": [ + { + "bbox": [ + 158, + 182, + 466, + 310 + ], + "score": 0.972, + "type": "image", + "image_path": "03f4d274b006a2d945a961a75df61f65f470c9506bae0cc5429f4ef3d09ead9c.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 158, + 182, + 466, + 224.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 158, + 224.66666666666666, + 466, + 267.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 158, + 267.3333333333333, + 466, + 310.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 318, + 506, + 396 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 317, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 506, + 332 + ], + "score": 1.0, + "content": "Figure 10: Different views of the same planning problem. In (a) we color the obstacles, the starting", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "and the goal position of the robot in deep blue, orange and brown, respectively. The stick robot can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "move and rotate. The corresponding configuration space is 3d, as visualized in (b), with the extra", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 281, + 364 + ], + "score": 1.0, + "content": "dimension being the rotation angle w.r.t. the", + "type": "text" + }, + { + "bbox": [ + 281, + 353, + 288, + 362 + ], + "score": 0.46, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "-axis. The blue region indicates the feasible state space,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "i.e., the set of collision-free states. The starting and the goal position are denoted with an orange and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 374, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 386 + ], + "score": 1.0, + "content": "a brown dot, respectively. Although the workspace looks trivial, the configuration space is irregular,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 384, + 249, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 249, + 397 + ], + "score": 1.0, + "content": "which makes the planning difficult.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 411, + 245, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 246, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 246, + 426 + ], + "score": 1.0, + "content": "B MORE PRELIMINARIES", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 506, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 507, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 507, + 451 + ], + "score": 1.0, + "content": "Tree-based sampling planner The tree-based sampling planner algorithm is illustrated in Fig-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 430, + 461 + ], + "score": 1.0, + "content": "ure 11. The Expand in Algorithm 1 operator returns an existing node in the tree", + "type": "text" + }, + { + "bbox": [ + 430, + 448, + 480, + 460 + ], + "score": 0.89, + "content": "s _ { p a r e n t } \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 447, + 506, + 461 + ], + "score": 1.0, + "content": "and a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 457, + 504, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 144, + 473 + ], + "score": 1.0, + "content": "new state", + "type": "text" + }, + { + "bbox": [ + 145, + 459, + 185, + 470 + ], + "score": 0.92, + "content": "s _ { n e w } \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 457, + 321, + 473 + ], + "score": 1.0, + "content": "sampled from the neighborhood of", + "type": "text" + }, + { + "bbox": [ + 322, + 460, + 352, + 471 + ], + "score": 0.8, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 457, + 444, + 473 + ], + "score": 1.0, + "content": ". Then the line segment", + "type": "text" + }, + { + "bbox": [ + 444, + 460, + 504, + 470 + ], + "score": 0.72, + "content": "[ s _ { p a r e n t } , s _ { n e w } ]", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 468, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 434, + 484 + ], + "score": 1.0, + "content": "is passed to function ObstacleFree for collision checking. If the line segment", + "type": "text" + }, + { + "bbox": [ + 434, + 471, + 494, + 482 + ], + "score": 0.89, + "content": "[ s _ { p a r e n t } , s _ { n e w } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 468, + 506, + 484 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 370, + 493 + ], + "score": 1.0, + "content": "collision-free (no obstacle in the middle, or called reachable from", + "type": "text" + }, + { + "bbox": [ + 370, + 481, + 380, + 491 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 480, + 406, + 493 + ], + "score": 1.0, + "content": "), then", + "type": "text" + }, + { + "bbox": [ + 406, + 482, + 427, + 492 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "is added to the tree", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 490, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 147, + 505 + ], + "score": 1.0, + "content": "vertex set", + "type": "text" + }, + { + "bbox": [ + 147, + 492, + 155, + 502 + ], + "score": 0.7, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 490, + 356, + 505 + ], + "score": 1.0, + "content": ", and the line segment is added to the tree edge set", + "type": "text" + }, + { + "bbox": [ + 356, + 492, + 364, + 501 + ], + "score": 0.68, + "content": "\\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 490, + 466, + 505 + ], + "score": 1.0, + "content": ". If the newly added node", + "type": "text" + }, + { + "bbox": [ + 467, + 493, + 488, + 503 + ], + "score": 0.89, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 490, + 506, + 505 + ], + "score": 1.0, + "content": "has", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 178, + 516 + ], + "score": 1.0, + "content": "reached the target", + "type": "text" + }, + { + "bbox": [ + 178, + 503, + 201, + 515 + ], + "score": 0.92, + "content": "\\mathcal { S } _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 502, + 506, + 516 + ], + "score": 1.0, + "content": ", the algorithm will return. Optionally, some concrete algorithms can define a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "Postprocess operator to refine the search tree. For an example of the Expand operator, as shown in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 335, + 538 + ], + "score": 1.0, + "content": "Figure 1 (c), since there is no obstacle on the dotted edge", + "type": "text" + }, + { + "bbox": [ + 336, + 524, + 395, + 537 + ], + "score": 0.85, + "content": "[ s _ { p a r e n t } , s _ { n e w } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 524, + 416, + 538 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 416, + 526, + 437, + 536 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "is reachable, the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 535, + 437, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 437, + 548 + ], + "score": 1.0, + "content": "new state and edge will be added to the search tree (connected by the solid edges).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5 + }, + { + "type": "image", + "bbox": [ + 191, + 553, + 420, + 642 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 191, + 553, + 420, + 642 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 191, + 553, + 420, + 642 + ], + "spans": [ + { + "bbox": [ + 191, + 553, + 420, + 642 + ], + "score": 0.967, + "type": "image", + "image_path": "be00082f31c012e5df15315075e706a96419afcb7dbc1ba05cf0204fbd263e3d.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 553, + 420, + 567.8333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 191, + 567.8333333333334, + 420, + 582.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 191, + 582.6666666666667, + 420, + 597.5000000000001 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 191, + 597.5000000000001, + 420, + 612.3333333333335 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 191, + 612.3333333333335, + 420, + 627.1666666666669 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 191, + 627.1666666666669, + 420, + 642.0000000000002 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 653, + 505, + 686 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "score": 1.0, + "content": "Figure 11: Illustration for one iteration of Algorithm 1. The left and right figures illustrate two", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 663, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 676 + ], + "score": 1.0, + "content": "different cases where the sample returned by the Expand operator is unreachable and reachable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 675, + 192, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 192, + 687 + ], + "score": 1.0, + "content": "from the search tree.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "index": 30.75 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 384, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 385, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 385, + 703 + ], + "score": 1.0, + "content": "Now we will provide two concrete algorithm examples. For instance,", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "• If we instantiate the Expand operator as Algorithm 4, then we obtain the rapidly-exploring random", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 720, + 504, + 734 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 381, + 734 + ], + "score": 1.0, + "content": "trees (RRT) algorithm (LaValle, 1998), which first samples a state", + "type": "text" + }, + { + "bbox": [ + 381, + 723, + 387, + 730 + ], + "score": 0.7, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 720, + 504, + 734 + ], + "score": 1.0, + "content": "from the configuration space", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 265, + 79, + 347, + 101 + ], + "lines": [ + { + "bbox": [ + 261, + 77, + 349, + 103 + ], + "spans": [ + { + "bbox": [ + 261, + 77, + 349, + 103 + ], + "score": 1.0, + "content": "Appendix", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 114, + 440, + 128 + ], + "lines": [ + { + "bbox": [ + 105, + 114, + 442, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 442, + 129 + ], + "score": 1.0, + "content": "A ILLUSTRATION OF THE DIFFICULTY IN PLANNING PROBLEMS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 140, + 506, + 174 + ], + "lines": [ + { + "bbox": [ + 104, + 139, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 139, + 505, + 153 + ], + "score": 1.0, + "content": "The Figure 10(a) illustrates a concrete planning problem for a stick robot in 2d workspace. With one", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 164 + ], + "score": 1.0, + "content": "extra continuous action for rotation, the configuration state is visualized in Figure 10(b), which is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 163, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 286, + 175 + ], + "score": 1.0, + "content": "highly irregular and unknown to the planner.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 139, + 505, + 175 + ] + }, + { + "type": "image", + "bbox": [ + 158, + 182, + 466, + 310 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 158, + 182, + 466, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 158, + 182, + 466, + 310 + ], + "spans": [ + { + "bbox": [ + 158, + 182, + 466, + 310 + ], + "score": 0.972, + "type": "image", + "image_path": "03f4d274b006a2d945a961a75df61f65f470c9506bae0cc5429f4ef3d09ead9c.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 158, + 182, + 466, + 224.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 158, + 224.66666666666666, + 466, + 267.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 158, + 267.3333333333333, + 466, + 310.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 318, + 506, + 396 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 317, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 506, + 332 + ], + "score": 1.0, + "content": "Figure 10: Different views of the same planning problem. In (a) we color the obstacles, the starting", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "and the goal position of the robot in deep blue, orange and brown, respectively. The stick robot can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "move and rotate. The corresponding configuration space is 3d, as visualized in (b), with the extra", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 281, + 364 + ], + "score": 1.0, + "content": "dimension being the rotation angle w.r.t. the", + "type": "text" + }, + { + "bbox": [ + 281, + 353, + 288, + 362 + ], + "score": 0.46, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "-axis. The blue region indicates the feasible state space,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "i.e., the set of collision-free states. The starting and the goal position are denoted with an orange and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 374, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 386 + ], + "score": 1.0, + "content": "a brown dot, respectively. Although the workspace looks trivial, the configuration space is irregular,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 384, + 249, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 249, + 397 + ], + "score": 1.0, + "content": "which makes the planning difficult.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 108, + 411, + 245, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 246, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 246, + 426 + ], + "score": 1.0, + "content": "B MORE PRELIMINARIES", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 506, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 507, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 507, + 451 + ], + "score": 1.0, + "content": "Tree-based sampling planner The tree-based sampling planner algorithm is illustrated in Fig-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 430, + 461 + ], + "score": 1.0, + "content": "ure 11. The Expand in Algorithm 1 operator returns an existing node in the tree", + "type": "text" + }, + { + "bbox": [ + 430, + 448, + 480, + 460 + ], + "score": 0.89, + "content": "s _ { p a r e n t } \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 447, + 506, + 461 + ], + "score": 1.0, + "content": "and a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 457, + 504, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 144, + 473 + ], + "score": 1.0, + "content": "new state", + "type": "text" + }, + { + "bbox": [ + 145, + 459, + 185, + 470 + ], + "score": 0.92, + "content": "s _ { n e w } \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 457, + 321, + 473 + ], + "score": 1.0, + "content": "sampled from the neighborhood of", + "type": "text" + }, + { + "bbox": [ + 322, + 460, + 352, + 471 + ], + "score": 0.8, + "content": "s _ { p a r e n t }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 457, + 444, + 473 + ], + "score": 1.0, + "content": ". Then the line segment", + "type": "text" + }, + { + "bbox": [ + 444, + 460, + 504, + 470 + ], + "score": 0.72, + "content": "[ s _ { p a r e n t } , s _ { n e w } ]", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 468, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 434, + 484 + ], + "score": 1.0, + "content": "is passed to function ObstacleFree for collision checking. If the line segment", + "type": "text" + }, + { + "bbox": [ + 434, + 471, + 494, + 482 + ], + "score": 0.89, + "content": "[ s _ { p a r e n t } , s _ { n e w } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 468, + 506, + 484 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 370, + 493 + ], + "score": 1.0, + "content": "collision-free (no obstacle in the middle, or called reachable from", + "type": "text" + }, + { + "bbox": [ + 370, + 481, + 380, + 491 + ], + "score": 0.75, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 480, + 406, + 493 + ], + "score": 1.0, + "content": "), then", + "type": "text" + }, + { + "bbox": [ + 406, + 482, + 427, + 492 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "is added to the tree", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 490, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 147, + 505 + ], + "score": 1.0, + "content": "vertex set", + "type": "text" + }, + { + "bbox": [ + 147, + 492, + 155, + 502 + ], + "score": 0.7, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 490, + 356, + 505 + ], + "score": 1.0, + "content": ", and the line segment is added to the tree edge set", + "type": "text" + }, + { + "bbox": [ + 356, + 492, + 364, + 501 + ], + "score": 0.68, + "content": "\\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 490, + 466, + 505 + ], + "score": 1.0, + "content": ". If the newly added node", + "type": "text" + }, + { + "bbox": [ + 467, + 493, + 488, + 503 + ], + "score": 0.89, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 490, + 506, + 505 + ], + "score": 1.0, + "content": "has", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 178, + 516 + ], + "score": 1.0, + "content": "reached the target", + "type": "text" + }, + { + "bbox": [ + 178, + 503, + 201, + 515 + ], + "score": 0.92, + "content": "\\mathcal { S } _ { g o a l }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 502, + 506, + 516 + ], + "score": 1.0, + "content": ", the algorithm will return. Optionally, some concrete algorithms can define a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "Postprocess operator to refine the search tree. For an example of the Expand operator, as shown in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 335, + 538 + ], + "score": 1.0, + "content": "Figure 1 (c), since there is no obstacle on the dotted edge", + "type": "text" + }, + { + "bbox": [ + 336, + 524, + 395, + 537 + ], + "score": 0.85, + "content": "[ s _ { p a r e n t } , s _ { n e w } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 524, + 416, + 538 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 416, + 526, + 437, + 536 + ], + "score": 0.88, + "content": "s _ { n e w }", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "is reachable, the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 535, + 437, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 437, + 548 + ], + "score": 1.0, + "content": "new state and edge will be added to the search tree (connected by the solid edges).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 434, + 507, + 548 + ] + }, + { + "type": "image", + "bbox": [ + 191, + 553, + 420, + 642 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 191, + 553, + 420, + 642 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 191, + 553, + 420, + 642 + ], + "spans": [ + { + "bbox": [ + 191, + 553, + 420, + 642 + ], + "score": 0.967, + "type": "image", + "image_path": "be00082f31c012e5df15315075e706a96419afcb7dbc1ba05cf0204fbd263e3d.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 553, + 420, + 567.8333333333334 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 191, + 567.8333333333334, + 420, + 582.6666666666667 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 191, + 582.6666666666667, + 420, + 597.5000000000001 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 191, + 597.5000000000001, + 420, + 612.3333333333335 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 191, + 612.3333333333335, + 420, + 627.1666666666669 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 191, + 627.1666666666669, + 420, + 642.0000000000002 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 653, + 505, + 686 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "score": 1.0, + "content": "Figure 11: Illustration for one iteration of Algorithm 1. The left and right figures illustrate two", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 663, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 676 + ], + "score": 1.0, + "content": "different cases where the sample returned by the Expand operator is unreachable and reachable", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 675, + 192, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 192, + 687 + ], + "score": 1.0, + "content": "from the search tree.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "index": 30.75 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 384, + 700 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 385, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 385, + 703 + ], + "score": 1.0, + "content": "Now we will provide two concrete algorithm examples. For instance,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "• If we instantiate the Expand operator as Algorithm 4, then we obtain the rapidly-exploring random", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 720, + 504, + 734 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 381, + 734 + ], + "score": 1.0, + "content": "trees (RRT) algorithm (LaValle, 1998), which first samples a state", + "type": "text" + }, + { + "bbox": [ + 381, + 723, + 387, + 730 + ], + "score": 0.7, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 720, + 504, + 734 + ], + "score": 1.0, + "content": "from the configuration space", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 686, + 385, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [], + "index": 36.5, + "bbox_fs": [ + 106, + 709, + 504, + 734 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 86, + 251, + 97 + ], + "lines": [ + { + "bbox": [ + 106, + 86, + 251, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 86, + 251, + 98 + ], + "score": 0.63, + "content": "\\mathbf { A l g o r i t h m 4 : R R T : : E x p a n d } ( { T } , { U } )", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 99, + 100, + 337, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 97, + 336, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 134, + 115 + ], + "score": 1.0, + "content": "Data:", + "type": "text" + }, + { + "bbox": [ + 135, + 101, + 336, + 113 + ], + "score": 0.75, + "content": "\\mathcal { T } = ( \\mathcal { V } , \\mathcal { E } ) , U = \\left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c ( \\cdot ) \\right)", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 112, + 188, + 123 + ], + "spans": [ + { + "bbox": [ + 97, + 112, + 106, + 123 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 107, + 113, + 183, + 123 + ], + "score": 0.79, + "content": "s _ { r a n d } \\gets \\mathcal { U } n i f ( \\mathcal { S } )", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 112, + 188, + 123 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 123, + 255, + 136 + ], + "spans": [ + { + "bbox": [ + 96, + 123, + 106, + 136 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 107, + 124, + 249, + 135 + ], + "score": 0.83, + "content": "\\begin{array} { r } { s _ { p a r e n t } \\gets \\operatorname * { a r g m i n } _ { s \\in \\mathcal { V } } \\left\\| s _ { r a n d } - s \\right\\| } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 123, + 255, + 136 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 132, + 284, + 150 + ], + "spans": [ + { + "bbox": [ + 95, + 132, + 106, + 150 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 106, + 136, + 279, + 147 + ], + "score": 0.85, + "content": "s _ { n e w } \\gets \\mathrm { a r g m i n } _ { s \\in \\mathcal { B } ( s _ { p a r e n t } , \\eta ) } \\left\\| s - s _ { r a n d } \\right\\|", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 132, + 284, + 150 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 99, + 149, + 194, + 159 + ], + "lines": [ + { + "bbox": [ + 97, + 146, + 191, + 160 + ], + "spans": [ + { + "bbox": [ + 97, + 146, + 136, + 160 + ], + "score": 1.0, + "content": "4 return", + "type": "text" + }, + { + "bbox": [ + 137, + 150, + 191, + 159 + ], + "score": 0.74, + "content": "s _ { p a r e n t } , s _ { n e w }", + "type": "inline_equation" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 370, + 112, + 490, + 134 + ], + "lines": [ + { + "bbox": [ + 369, + 112, + 490, + 123 + ], + "spans": [ + { + "bbox": [ + 369, + 112, + 490, + 123 + ], + "score": 1.0, + "content": ". Sample configuration space", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 406, + 122, + 490, + 134 + ], + "spans": [ + { + "bbox": [ + 406, + 122, + 490, + 134 + ], + "score": 1.0, + "content": ". Pull to a tree node", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 173, + 251, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 251, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 251, + 186 + ], + "score": 0.51, + "content": "\\mathbf { A l g o r i t h m 5 : E S T : : E x p a n d } ( T , U )", + "type": "inline_equation" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 98, + 187, + 338, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 338, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 134, + 201 + ], + "score": 1.0, + "content": "Data:", + "type": "text" + }, + { + "bbox": [ + 135, + 188, + 336, + 200 + ], + "score": 0.78, + "content": "\\mathcal { T } = ( \\mathcal { V } , \\mathcal { E } ) , U = \\left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c ( \\cdot ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 186, + 338, + 201 + ], + "score": 0.793, + "content": "，", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 98, + 199, + 202, + 211 + ], + "spans": [ + { + "bbox": [ + 98, + 199, + 138, + 211 + ], + "score": 1.0, + "content": "1 sparent", + "type": "text" + }, + { + "bbox": [ + 138, + 200, + 197, + 211 + ], + "score": 0.73, + "content": "\\sim \\phi ( s ) , s \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 199, + 202, + 211 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 96, + 209, + 221, + 223 + ], + "spans": [ + { + "bbox": [ + 96, + 209, + 106, + 223 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 106, + 210, + 217, + 222 + ], + "score": 0.83, + "content": "\\hat { s _ { n e w } } \\mathcal { U } n i f ( B ( s _ { p a r e n t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 209, + 221, + 223 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 97, + 222, + 199, + 234 + ], + "spans": [ + { + "bbox": [ + 97, + 222, + 199, + 234 + ], + "score": 1.0, + "content": "3 return snearest, snew;", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 392, + 200, + 490, + 222 + ], + "lines": [ + { + "bbox": [ + 403, + 199, + 489, + 210 + ], + "spans": [ + { + "bbox": [ + 403, + 199, + 489, + 210 + ], + "score": 1.0, + "content": ". Sample a tree node", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 391, + 210, + 490, + 221 + ], + "spans": [ + { + "bbox": [ + 391, + 210, + 490, + 221 + ], + "score": 1.0, + "content": ". Sample neighborhood", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 113, + 244, + 488, + 256 + ], + "lines": [ + { + "bbox": [ + 117, + 241, + 486, + 259 + ], + "spans": [ + { + "bbox": [ + 117, + 245, + 124, + 255 + ], + "score": 0.71, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 241, + 353, + 259 + ], + "score": 1.0, + "content": "and then pulls it toward the neighborhood of current tree", + "type": "text" + }, + { + "bbox": [ + 353, + 245, + 363, + 255 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 241, + 479, + 259 + ], + "score": 1.0, + "content": "measured by a ball of radius", + "type": "text" + }, + { + "bbox": [ + 479, + 247, + 486, + 256 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 256, + 380, + 270 + ], + "lines": [], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 237, + 256, + 380, + 270 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 115, + 271, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 116, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 116, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "Moreover, if the Postprocess operator is introduced to modify the maintained search tree as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 280, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 128, + 294 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 128, + 281, + 153, + 292 + ], + "score": 0.54, + "content": "\\mathrm { { R R T ^ { * } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 280, + 505, + 294 + ], + "score": 1.0, + "content": "(Karaman & Frazzoli, 2011), the algorithm is provable to obtain the optimal path", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 292, + 180, + 306 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 180, + 306 + ], + "score": 1.0, + "content": "asymptotically.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 307, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "• If we instantiate the Expand operator as Algorithm 5, then we obtain the expansive-space trees", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 425, + 330 + ], + "score": 1.0, + "content": "(EST) algorithm (Hsu et al., 1997; Phillips et al., 2004), which samples a state", + "type": "text" + }, + { + "bbox": [ + 426, + 320, + 432, + 328 + ], + "score": 0.59, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "from the nodes of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 329, + 396, + 341 + ], + "spans": [ + { + "bbox": [ + 116, + 329, + 387, + 341 + ], + "score": 1.0, + "content": "the existing tree, and then draw a sample from the neighborhood of", + "type": "text" + }, + { + "bbox": [ + 387, + 331, + 392, + 339 + ], + "score": 0.7, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 329, + 396, + 341 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 355, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 106, + 355, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 282, + 369 + ], + "score": 1.0, + "content": "UCB-based algorithms Specifically, in a", + "type": "text" + }, + { + "bbox": [ + 283, + 356, + 293, + 366 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 355, + 506, + 369 + ], + "score": 1.0, + "content": "-armed bandit problem, the UCB algorithm will first", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 367, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 396, + 379 + ], + "score": 1.0, + "content": "play each of the arms once, and then keep track of the average reward", + "type": "text" + }, + { + "bbox": [ + 396, + 367, + 405, + 378 + ], + "score": 0.86, + "content": "\\bar { r } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 367, + 506, + 379 + ], + "score": 1.0, + "content": "and the visitation count", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 377, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 107, + 380, + 117, + 389 + ], + "score": 0.79, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 377, + 203, + 390 + ], + "score": 1.0, + "content": "for each arm. After", + "type": "text" + }, + { + "bbox": [ + 203, + 378, + 212, + 388 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 377, + 506, + 390 + ], + "score": 1.0, + "content": "rounds of trials, the UCB algorithm will maintain a set of information", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 388, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 107, + 388, + 160, + 403 + ], + "score": 0.92, + "content": "\\{ ( \\bar { r } _ { i } , n _ { i } ) \\} _ { i = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 388, + 182, + 407 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 182, + 389, + 239, + 403 + ], + "score": 0.91, + "content": "\\textstyle \\sum _ { i = 1 } ^ { K } n _ { i } = T", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 388, + 507, + 407 + ], + "score": 1.0, + "content": ". Then, for the next round, the UCB algorithm will select the next", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 402, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 506, + 414 + ], + "score": 1.0, + "content": "arm based on the one-sided confidence interval estimation provided by the Chernoff-Hoeffding bound,", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 422, + 389, + 443 + ], + "lines": [ + { + "bbox": [ + 220, + 422, + 389, + 443 + ], + "spans": [ + { + "bbox": [ + 220, + 422, + 389, + 443 + ], + "score": 0.93, + "content": "\\begin{array} { r } { a _ { T + 1 } ^ { * } = \\operatorname * { a r g m a x } _ { i \\in \\{ 1 , \\ldots , K \\} } \\bar { r } _ { i } + \\lambda \\sqrt { \\frac { \\log T } { n _ { i } } } , } \\end{array}", + "type": "interline_equation", + "image_path": "bb94fadba4fdb19b851b3da9d5f8f9946e77c39171639687da54124e30089e17.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 220, + 422, + 389, + 443 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 133, + 456 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 444, + 141, + 453 + ], + "score": 0.8, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "controls the exploration-exploration trade-off. It has been shown that the UCB algorithm", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 143, + 467 + ], + "score": 1.0, + "content": "achieves", + "type": "text" + }, + { + "bbox": [ + 144, + 455, + 183, + 467 + ], + "score": 0.92, + "content": "\\mathcal { O } \\left( \\log T \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "regret. However, the MCTS is not directly applicable to continuous state-action", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 466, + 138, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 138, + 479 + ], + "score": 1.0, + "content": "spaces.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "There have been many attempts to generalize the UCB and UCT algorithms to continuous state-action", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 493, + 507, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 507, + 507 + ], + "score": 1.0, + "content": "spaces (Chu et al., 2011; Krause & Ong, 2011; Couëtoux et al., 2011; Yee et al., 2016). For instance,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "contextual bandit algorithms allow continuous arms but involve a non-trivial high dimensional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "non-convex optimization to select the next arm. In UCT, the progressive widening technique has", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "been designed to deal with continuous actions (Wang et al., 2009). Even with these extensions, the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "MCTS restricts the exploration only from leaves states, implicitly adding an unnecessary hierarchical", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "structure for path planning, resulting inferior exploration efficiency and extra computation in path", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 559, + 169, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 169, + 572 + ], + "score": 1.0, + "content": "planning tasks.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 108, + 576, + 503, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "Although these off-the-shelf algorithms are not directly applicable to our path planning setting, their", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "successes show the importance of exploration-exploitation trade-off and will provide the principles", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 598, + 366, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 366, + 612 + ], + "score": 1.0, + "content": "for our algorithm for continuous state-action planning problems.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "Planning networks Value iteration networks (Tamar et al., 2016) employ neural networks to embed", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "the value iteration algorithm from planning and then use this embedded algorithm to extract input", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 643, + 414, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 414, + 656 + ], + "score": 1.0, + "content": "features and define downstream models such as value functions and policies.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 108, + 660, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 462, + 673 + ], + "score": 1.0, + "content": "Specifically, VIN mimics the following recursive application of Bellman update operator", + "type": "text" + }, + { + "bbox": [ + 462, + 661, + 470, + 671 + ], + "score": 0.85, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "to value", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 669, + 159, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 142, + 684 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 142, + 672, + 155, + 681 + ], + "score": 0.86, + "content": "V ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 669, + 159, + 684 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + }, + { + "type": "interline_equation", + "bbox": [ + 163, + 682, + 447, + 708 + ], + "lines": [ + { + "bbox": [ + 163, + 682, + 447, + 708 + ], + "spans": [ + { + "bbox": [ + 163, + 682, + 447, + 708 + ], + "score": 0.92, + "content": "V ^ { * } \\left( s | U \\right) = \\left( \\mathcal { G } V ^ { * } \\right) ( s ) : = \\operatorname* { m i n } _ { a } \\sum _ { s ^ { \\prime } } P ( s ^ { \\prime } | s , a ) ( c \\left( \\left[ s , s ^ { \\prime } \\right] \\right) + V ^ { * } \\left( s ^ { \\prime } | U \\right) ) .", + "type": "interline_equation", + "image_path": "4226ddc32831387ca8faaca832db0dc985d7476a84f1955aad452ab384d5c4ea.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 163, + 682, + 447, + 708 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 132, + 722 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 709, + 173, + 722 + ], + "score": 0.93, + "content": "P ( s ^ { \\prime } | s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 708, + 381, + 722 + ], + "score": 1.0, + "content": "is the state transition model. When the state space for", + "type": "text" + }, + { + "bbox": [ + 382, + 712, + 387, + 720 + ], + "score": 0.77, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 708, + 467, + 722 + ], + "score": 1.0, + "content": "and action space for", + "type": "text" + }, + { + "bbox": [ + 468, + 712, + 474, + 720 + ], + "score": 0.78, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "are low", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 107, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 438, + 732 + ], + "score": 1.0, + "content": "dimensional, these spaces can be discretized into grids. Then, the local cost function", + "type": "text" + }, + { + "bbox": [ + 438, + 720, + 473, + 732 + ], + "score": 0.93, + "content": "c ( [ s , s ^ { \\prime } ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 86, + 251, + 97 + ], + "lines": [ + { + "bbox": [ + 106, + 86, + 251, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 86, + 251, + 98 + ], + "score": 0.63, + "content": "\\mathbf { A l g o r i t h m 4 : R R T : : E x p a n d } ( { T } , { U } )", + "type": "inline_equation" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 99, + 100, + 337, + 148 + ], + "lines": [ + { + "bbox": [ + 105, + 97, + 336, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 134, + 115 + ], + "score": 1.0, + "content": "Data:", + "type": "text" + }, + { + "bbox": [ + 135, + 101, + 336, + 113 + ], + "score": 0.75, + "content": "\\mathcal { T } = ( \\mathcal { V } , \\mathcal { E } ) , U = \\left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c ( \\cdot ) \\right)", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 112, + 188, + 123 + ], + "spans": [ + { + "bbox": [ + 97, + 112, + 106, + 123 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 107, + 113, + 183, + 123 + ], + "score": 0.79, + "content": "s _ { r a n d } \\gets \\mathcal { U } n i f ( \\mathcal { S } )", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 112, + 188, + 123 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 123, + 255, + 136 + ], + "spans": [ + { + "bbox": [ + 96, + 123, + 106, + 136 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 107, + 124, + 249, + 135 + ], + "score": 0.83, + "content": "\\begin{array} { r } { s _ { p a r e n t } \\gets \\operatorname * { a r g m i n } _ { s \\in \\mathcal { V } } \\left\\| s _ { r a n d } - s \\right\\| } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 123, + 255, + 136 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 132, + 284, + 150 + ], + "spans": [ + { + "bbox": [ + 95, + 132, + 106, + 150 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 106, + 136, + 279, + 147 + ], + "score": 0.85, + "content": "s _ { n e w } \\gets \\mathrm { a r g m i n } _ { s \\in \\mathcal { B } ( s _ { p a r e n t } , \\eta ) } \\left\\| s - s _ { r a n d } \\right\\|", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 132, + 284, + 150 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 95, + 97, + 336, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 99, + 149, + 194, + 159 + ], + "lines": [ + { + "bbox": [ + 97, + 146, + 191, + 160 + ], + "spans": [ + { + "bbox": [ + 97, + 146, + 136, + 160 + ], + "score": 1.0, + "content": "4 return", + "type": "text" + }, + { + "bbox": [ + 137, + 150, + 191, + 159 + ], + "score": 0.74, + "content": "s _ { p a r e n t } , s _ { n e w }", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 369, + 112, + 490, + 123 + ], + "spans": [ + { + "bbox": [ + 369, + 112, + 490, + 123 + ], + "score": 1.0, + "content": ". Sample configuration space", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 406, + 122, + 490, + 134 + ], + "spans": [ + { + "bbox": [ + 406, + 122, + 490, + 134 + ], + "score": 1.0, + "content": ". Pull to a tree node", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 97, + 146, + 191, + 160 + ] + }, + { + "type": "text", + "bbox": [ + 370, + 112, + 490, + 134 + ], + "lines": [], + "index": 6.5, + "bbox_fs": [ + 369, + 112, + 490, + 134 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 173, + 251, + 185 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 251, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 251, + 186 + ], + "score": 0.51, + "content": "\\mathbf { A l g o r i t h m 5 : E S T : : E x p a n d } ( T , U )", + "type": "inline_equation" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 98, + 187, + 338, + 234 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 338, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 134, + 201 + ], + "score": 1.0, + "content": "Data:", + "type": "text" + }, + { + "bbox": [ + 135, + 188, + 336, + 200 + ], + "score": 0.78, + "content": "\\mathcal { T } = ( \\mathcal { V } , \\mathcal { E } ) , U = \\left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c ( \\cdot ) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 186, + 338, + 201 + ], + "score": 0.793, + "content": "，", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 98, + 199, + 202, + 211 + ], + "spans": [ + { + "bbox": [ + 98, + 199, + 138, + 211 + ], + "score": 1.0, + "content": "1 sparent", + "type": "text" + }, + { + "bbox": [ + 138, + 200, + 197, + 211 + ], + "score": 0.73, + "content": "\\sim \\phi ( s ) , s \\in \\mathcal { V }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 199, + 202, + 211 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 96, + 209, + 221, + 223 + ], + "spans": [ + { + "bbox": [ + 96, + 209, + 106, + 223 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 106, + 210, + 217, + 222 + ], + "score": 0.83, + "content": "\\hat { s _ { n e w } } \\mathcal { U } n i f ( B ( s _ { p a r e n t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 209, + 221, + 223 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 97, + 222, + 199, + 234 + ], + "spans": [ + { + "bbox": [ + 97, + 222, + 199, + 234 + ], + "score": 1.0, + "content": "3 return snearest, snew;", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 96, + 186, + 338, + 234 + ] + }, + { + "type": "text", + "bbox": [ + 392, + 200, + 490, + 222 + ], + "lines": [ + { + "bbox": [ + 403, + 199, + 489, + 210 + ], + "spans": [ + { + "bbox": [ + 403, + 199, + 489, + 210 + ], + "score": 1.0, + "content": ". Sample a tree node", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 391, + 210, + 490, + 221 + ], + "spans": [ + { + "bbox": [ + 391, + 210, + 490, + 221 + ], + "score": 1.0, + "content": ". Sample neighborhood", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 391, + 199, + 490, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 244, + 488, + 256 + ], + "lines": [ + { + "bbox": [ + 117, + 241, + 486, + 259 + ], + "spans": [ + { + "bbox": [ + 117, + 245, + 124, + 255 + ], + "score": 0.71, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 241, + 353, + 259 + ], + "score": 1.0, + "content": "and then pulls it toward the neighborhood of current tree", + "type": "text" + }, + { + "bbox": [ + 353, + 245, + 363, + 255 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 241, + 479, + 259 + ], + "score": 1.0, + "content": "measured by a ball of radius", + "type": "text" + }, + { + "bbox": [ + 479, + 247, + 486, + 256 + ], + "score": 0.78, + "content": "\\eta", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 117, + 241, + 486, + 259 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 256, + 380, + 270 + ], + "lines": [], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 237, + 256, + 380, + 270 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 115, + 271, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 116, + 270, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 116, + 270, + 505, + 283 + ], + "score": 1.0, + "content": "Moreover, if the Postprocess operator is introduced to modify the maintained search tree as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 280, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 128, + 294 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 128, + 281, + 153, + 292 + ], + "score": 0.54, + "content": "\\mathrm { { R R T ^ { * } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 280, + 505, + 294 + ], + "score": 1.0, + "content": "(Karaman & Frazzoli, 2011), the algorithm is provable to obtain the optimal path", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 292, + 180, + 306 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 180, + 306 + ], + "score": 1.0, + "content": "asymptotically.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 115, + 270, + 505, + 306 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 307, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "• If we instantiate the Expand operator as Algorithm 5, then we obtain the expansive-space trees", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 425, + 330 + ], + "score": 1.0, + "content": "(EST) algorithm (Hsu et al., 1997; Phillips et al., 2004), which samples a state", + "type": "text" + }, + { + "bbox": [ + 426, + 320, + 432, + 328 + ], + "score": 0.59, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "from the nodes of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 329, + 396, + 341 + ], + "spans": [ + { + "bbox": [ + 116, + 329, + 387, + 341 + ], + "score": 1.0, + "content": "the existing tree, and then draw a sample from the neighborhood of", + "type": "text" + }, + { + "bbox": [ + 387, + 331, + 392, + 339 + ], + "score": 0.7, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 329, + 396, + 341 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 306, + 505, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 355, + 505, + 413 + ], + "lines": [ + { + "bbox": [ + 106, + 355, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 282, + 369 + ], + "score": 1.0, + "content": "UCB-based algorithms Specifically, in a", + "type": "text" + }, + { + "bbox": [ + 283, + 356, + 293, + 366 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 355, + 506, + 369 + ], + "score": 1.0, + "content": "-armed bandit problem, the UCB algorithm will first", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 367, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 396, + 379 + ], + "score": 1.0, + "content": "play each of the arms once, and then keep track of the average reward", + "type": "text" + }, + { + "bbox": [ + 396, + 367, + 405, + 378 + ], + "score": 0.86, + "content": "\\bar { r } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 367, + 506, + 379 + ], + "score": 1.0, + "content": "and the visitation count", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 377, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 107, + 380, + 117, + 389 + ], + "score": 0.79, + "content": "n _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 377, + 203, + 390 + ], + "score": 1.0, + "content": "for each arm. After", + "type": "text" + }, + { + "bbox": [ + 203, + 378, + 212, + 388 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 377, + 506, + 390 + ], + "score": 1.0, + "content": "rounds of trials, the UCB algorithm will maintain a set of information", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 388, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 107, + 388, + 160, + 403 + ], + "score": 0.92, + "content": "\\{ ( \\bar { r } _ { i } , n _ { i } ) \\} _ { i = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 388, + 182, + 407 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 182, + 389, + 239, + 403 + ], + "score": 0.91, + "content": "\\textstyle \\sum _ { i = 1 } ^ { K } n _ { i } = T", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 388, + 507, + 407 + ], + "score": 1.0, + "content": ". Then, for the next round, the UCB algorithm will select the next", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 402, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 506, + 414 + ], + "score": 1.0, + "content": "arm based on the one-sided confidence interval estimation provided by the Chernoff-Hoeffding bound,", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 355, + 507, + 414 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 422, + 389, + 443 + ], + "lines": [ + { + "bbox": [ + 220, + 422, + 389, + 443 + ], + "spans": [ + { + "bbox": [ + 220, + 422, + 389, + 443 + ], + "score": 0.93, + "content": "\\begin{array} { r } { a _ { T + 1 } ^ { * } = \\operatorname * { a r g m a x } _ { i \\in \\{ 1 , \\ldots , K \\} } \\bar { r } _ { i } + \\lambda \\sqrt { \\frac { \\log T } { n _ { i } } } , } \\end{array}", + "type": "interline_equation", + "image_path": "bb94fadba4fdb19b851b3da9d5f8f9946e77c39171639687da54124e30089e17.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 220, + 422, + 389, + 443 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 133, + 456 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 444, + 141, + 453 + ], + "score": 0.8, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "controls the exploration-exploration trade-off. It has been shown that the UCB algorithm", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 143, + 467 + ], + "score": 1.0, + "content": "achieves", + "type": "text" + }, + { + "bbox": [ + 144, + 455, + 183, + 467 + ], + "score": 0.92, + "content": "\\mathcal { O } \\left( \\log T \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "regret. However, the MCTS is not directly applicable to continuous state-action", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 466, + 138, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 138, + 479 + ], + "score": 1.0, + "content": "spaces.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 442, + 505, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "There have been many attempts to generalize the UCB and UCT algorithms to continuous state-action", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 493, + 507, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 507, + 507 + ], + "score": 1.0, + "content": "spaces (Chu et al., 2011; Krause & Ong, 2011; Couëtoux et al., 2011; Yee et al., 2016). For instance,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "contextual bandit algorithms allow continuous arms but involve a non-trivial high dimensional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "non-convex optimization to select the next arm. In UCT, the progressive widening technique has", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "been designed to deal with continuous actions (Wang et al., 2009). Even with these extensions, the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "score": 1.0, + "content": "MCTS restricts the exploration only from leaves states, implicitly adding an unnecessary hierarchical", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "structure for path planning, resulting inferior exploration efficiency and extra computation in path", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 559, + 169, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 169, + 572 + ], + "score": 1.0, + "content": "planning tasks.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 482, + 507, + 572 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 576, + 503, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "Although these off-the-shelf algorithms are not directly applicable to our path planning setting, their", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "successes show the importance of exploration-exploitation trade-off and will provide the principles", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 598, + 366, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 366, + 612 + ], + "score": 1.0, + "content": "for our algorithm for continuous state-action planning problems.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 576, + 505, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "Planning networks Value iteration networks (Tamar et al., 2016) employ neural networks to embed", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "the value iteration algorithm from planning and then use this embedded algorithm to extract input", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 643, + 414, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 414, + 656 + ], + "score": 1.0, + "content": "features and define downstream models such as value functions and policies.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 621, + 506, + 656 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 660, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 462, + 673 + ], + "score": 1.0, + "content": "Specifically, VIN mimics the following recursive application of Bellman update operator", + "type": "text" + }, + { + "bbox": [ + 462, + 661, + 470, + 671 + ], + "score": 0.85, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "to value", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 669, + 159, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 142, + 684 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 142, + 672, + 155, + 681 + ], + "score": 0.86, + "content": "V ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 669, + 159, + 684 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 659, + 505, + 684 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 163, + 682, + 447, + 708 + ], + "lines": [ + { + "bbox": [ + 163, + 682, + 447, + 708 + ], + "spans": [ + { + "bbox": [ + 163, + 682, + 447, + 708 + ], + "score": 0.92, + "content": "V ^ { * } \\left( s | U \\right) = \\left( \\mathcal { G } V ^ { * } \\right) ( s ) : = \\operatorname* { m i n } _ { a } \\sum _ { s ^ { \\prime } } P ( s ^ { \\prime } | s , a ) ( c \\left( \\left[ s , s ^ { \\prime } \\right] \\right) + V ^ { * } \\left( s ^ { \\prime } | U \\right) ) .", + "type": "interline_equation", + "image_path": "4226ddc32831387ca8faaca832db0dc985d7476a84f1955aad452ab384d5c4ea.jpg" + } + ] + } + ], + "index": 48, + "virtual_lines": [ + { + "bbox": [ + 163, + 682, + 447, + 708 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 132, + 722 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 709, + 173, + 722 + ], + "score": 0.93, + "content": "P ( s ^ { \\prime } | s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 708, + 381, + 722 + ], + "score": 1.0, + "content": "is the state transition model. When the state space for", + "type": "text" + }, + { + "bbox": [ + 382, + 712, + 387, + 720 + ], + "score": 0.77, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 708, + 467, + 722 + ], + "score": 1.0, + "content": "and action space for", + "type": "text" + }, + { + "bbox": [ + 468, + 712, + 474, + 720 + ], + "score": 0.78, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "are low", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 107, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 438, + 732 + ], + "score": 1.0, + "content": "dimensional, these spaces can be discretized into grids. Then, the local cost function", + "type": "text" + }, + { + "bbox": [ + 438, + 720, + 473, + 732 + ], + "score": 0.93, + "content": "c ( [ s , s ^ { \\prime } ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 165, + 94 + ], + "score": 1.0, + "content": "value function", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 165, + 82, + 205, + 95 + ], + "score": 0.93, + "content": "V ^ { * } ( s ^ { \\prime } | U )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 205, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "can be represented as matrices (2d) or tensors (3d) with each entry indexed", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 504, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 335, + 106 + ], + "score": 1.0, + "content": "by grid locations. Furthermore, if the transition model", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 335, + 93, + 376, + 106 + ], + "score": 0.93, + "content": "P ( s ^ { \\prime } | s , a )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 376, + 93, + 442, + 106 + ], + "score": 1.0, + "content": "is local, that is", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 442, + 93, + 504, + 106 + ], + "score": 0.92, + "content": "P ( s ^ { \\prime } | \\dot { s } , a ) = 0", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 121, + 117 + ], + "score": 1.0, + "content": "for", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 121, + 104, + 162, + 117 + ], + "score": 0.92, + "content": "\\bar { \\boldsymbol { s } } ^ { \\prime } \\notin B ( \\boldsymbol { s } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 162, + 104, + 473, + 117 + ], + "score": 1.0, + "content": ", it resembles a set of convolution kernels, each indexed by a discrete action", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 474, + 107, + 480, + 114 + ], + "score": 0.73, + "content": "a", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 480, + 104, + 506, + 117 + ], + "score": 1.0, + "content": ". And", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 308, + 128 + ], + "score": 1.0, + "content": "the Bellman update operator essentially convolves", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 309, + 115, + 350, + 127 + ], + "score": 0.93, + "content": "P ( s ^ { \\prime } | s , a )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 350, + 114, + 371, + 128 + ], + "score": 1.0, + "content": "with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 372, + 115, + 406, + 127 + ], + "score": 0.93, + "content": "c ( [ s , s ^ { \\bar { \\prime } } ] )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 407, + 114, + 425, + 128 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 425, + 115, + 464, + 127 + ], + "score": 0.94, + "content": "V ^ { * } ( s ^ { \\prime } | U )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 465, + 114, + 505, + 128 + ], + "score": 1.0, + "content": ", and then", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 376, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 376, + 139 + ], + "score": 1.0, + "content": "performs a min-pooling operation across the convolution channels.", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 49.5, + "bbox_fs": [ + 106, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 506, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 165, + 94 + ], + "score": 1.0, + "content": "value function", + "type": "text" + }, + { + "bbox": [ + 165, + 82, + 205, + 95 + ], + "score": 0.93, + "content": "V ^ { * } ( s ^ { \\prime } | U )", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "can be represented as matrices (2d) or tensors (3d) with each entry indexed", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 504, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 335, + 106 + ], + "score": 1.0, + "content": "by grid locations. Furthermore, if the transition model", + "type": "text" + }, + { + "bbox": [ + 335, + 93, + 376, + 106 + ], + "score": 0.93, + "content": "P ( s ^ { \\prime } | s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 93, + 442, + 106 + ], + "score": 1.0, + "content": "is local, that is", + "type": "text" + }, + { + "bbox": [ + 442, + 93, + 504, + 106 + ], + "score": 0.92, + "content": "P ( s ^ { \\prime } | \\dot { s } , a ) = 0", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 121, + 117 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 104, + 162, + 117 + ], + "score": 0.92, + "content": "\\bar { \\boldsymbol { s } } ^ { \\prime } \\notin B ( \\boldsymbol { s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 104, + 473, + 117 + ], + "score": 1.0, + "content": ", it resembles a set of convolution kernels, each indexed by a discrete action", + "type": "text" + }, + { + "bbox": [ + 474, + 107, + 480, + 114 + ], + "score": 0.73, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 104, + 506, + 117 + ], + "score": 1.0, + "content": ". And", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 308, + 128 + ], + "score": 1.0, + "content": "the Bellman update operator essentially convolves", + "type": "text" + }, + { + "bbox": [ + 309, + 115, + 350, + 127 + ], + "score": 0.93, + "content": "P ( s ^ { \\prime } | s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 114, + 371, + 128 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 372, + 115, + 406, + 127 + ], + "score": 0.93, + "content": "c ( [ s , s ^ { \\bar { \\prime } } ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 114, + 425, + 128 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 425, + 115, + 464, + 127 + ], + "score": 0.94, + "content": "V ^ { * } ( s ^ { \\prime } | U )", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 114, + 505, + 128 + ], + "score": 1.0, + "content": ", and then", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 376, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 376, + 139 + ], + "score": 1.0, + "content": "performs a min-pooling operation across the convolution channels.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 157 + ], + "score": 1.0, + "content": "Inspired by the above computation pattern of the Bellman operator, value iteration networks design", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 244, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 244, + 166 + ], + "score": 1.0, + "content": "the neural architecture as follows,", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 165, + 375, + 207 + ], + "lines": [ + { + "bbox": [ + 235, + 165, + 375, + 207 + ], + "spans": [ + { + "bbox": [ + 235, + 165, + 375, + 207 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\tilde { V } ^ { * 0 } = \\operatorname* { m i n } \\ \\left( W _ { 1 } \\oplus \\left[ \\tt { m a p } , \\tilde { R } \\right] \\right) } \\\\ & { \\tilde { V } ^ { * t } = \\operatorname* { m i n } \\ \\left( W _ { 1 } \\oplus \\left[ \\tilde { V } ^ { * t - 1 } , \\tilde { R } \\right] \\right) } \\end{array}", + "type": "interline_equation", + "image_path": "dee3f8bfd36cb521afcd22f462a62dffa0835e6ce5931e74fd22a401561b036a.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 165, + 375, + 186.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 235, + 186.0, + 375, + 207.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 208, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 104, + 206, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 104, + 206, + 133, + 222 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 210, + 143, + 219 + ], + "score": 0.83, + "content": "\\oplus", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 206, + 304, + 222 + ], + "score": 1.0, + "content": "is the convolution operation, both map,", + "type": "text" + }, + { + "bbox": [ + 304, + 207, + 321, + 219 + ], + "score": 0.87, + "content": "\\tilde { V } ^ { * t }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 206, + 339, + 222 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 340, + 207, + 349, + 219 + ], + "score": 0.86, + "content": "\\tilde { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 206, + 365, + 222 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 365, + 208, + 389, + 219 + ], + "score": 0.89, + "content": "d \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 206, + 506, + 222 + ], + "score": 1.0, + "content": "matrices, and the parameter", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 107, + 220, + 121, + 231 + ], + "score": 0.88, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 219, + 137, + 232 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 137, + 220, + 148, + 231 + ], + "score": 0.88, + "content": "k _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 219, + 257, + 232 + ], + "score": 1.0, + "content": "convolution kernels of size", + "type": "text" + }, + { + "bbox": [ + 257, + 220, + 281, + 230 + ], + "score": 0.91, + "content": "k \\times k", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 219, + 444, + 232 + ], + "score": 1.0, + "content": ". The min implements the pooling across", + "type": "text" + }, + { + "bbox": [ + 444, + 220, + 455, + 231 + ], + "score": 0.88, + "content": "k _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "convolution", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 229, + 147, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 147, + 243 + ], + "score": 1.0, + "content": "channels.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 504, + 260 + ], + "score": 1.0, + "content": "The gated path planning networks (GPPN) (Lee et al., 2018) improves the VIN by replacing the VIN", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 258, + 317, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 317, + 272 + ], + "score": 1.0, + "content": "cell (9) with the well-established LSTM update, i.e.,", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 216, + 270, + 394, + 291 + ], + "lines": [ + { + "bbox": [ + 216, + 270, + 394, + 291 + ], + "spans": [ + { + "bbox": [ + 216, + 270, + 394, + 291 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { t } , \\tilde { c } ^ { t } = \\mathrm { L S T M } \\left( \\sum \\left( W _ { 1 } \\oplus \\left[ \\tilde { V } ^ { t } , \\tilde { R } \\right] \\right) , \\tilde { c } ^ { t } \\right) ,", + "type": "interline_equation", + "image_path": "29d45bb013ca6188267f130a8fe3ab6c4c91b7fa1a03866c1230f433c3afccf3.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 216, + 270, + 394, + 291 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 378, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 379, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 277, + 304 + ], + "score": 1.0, + "content": "where the summation is taking over all the", + "type": "text" + }, + { + "bbox": [ + 278, + 292, + 289, + 303 + ], + "score": 0.87, + "content": "k _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 291, + 379, + 304 + ], + "score": 1.0, + "content": "convolution channels.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 108, + 309, + 503, + 338 + ], + "lines": [ + { + "bbox": [ + 102, + 304, + 507, + 330 + ], + "spans": [ + { + "bbox": [ + 102, + 304, + 392, + 330 + ], + "score": 1.0, + "content": "After constructing the VIN and GPPN, the parameters of the model, i.e.,", + "type": "text" + }, + { + "bbox": [ + 393, + 308, + 433, + 329 + ], + "score": 0.93, + "content": "\\left\\{ \\tilde { R } , W _ { 1 } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 304, + 507, + 330 + ], + "score": 1.0, + "content": "can be learned by", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 324, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 287, + 340 + ], + "score": 1.0, + "content": "imitation learning or reinforcement learning.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 504, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "The application of planning networks are restricted in low-dimension tasks. However, their success", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "enlightens our neural architecture for generalizable representation for high-dimension planning tasks.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 106, + 384, + 313, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 314, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 314, + 400 + ], + "score": 1.0, + "content": "C PARAMETRIZED UCB ALGORITHMS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 438, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 439, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 439, + 424 + ], + "score": 1.0, + "content": "We list two examples of parametrized UCB as the instantiation of (3) used in GPE:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 133, + 430, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 131, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "• GP-UCB: The GP-UCB Chu et al. (2011) is derived by parameterizing via Gaussian", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 140, + 441, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 140, + 441, + 255, + 456 + ], + "score": 1.0, + "content": "Processes (GP) with kernel", + "type": "text" + }, + { + "bbox": [ + 255, + 442, + 287, + 455 + ], + "score": 0.91, + "content": "k \\left( s , s ^ { \\prime } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 441, + 316, + 456 + ], + "score": 1.0, + "content": ", i.e., E", + "type": "text" + }, + { + "bbox": [ + 316, + 442, + 419, + 455 + ], + "score": 0.88, + "content": "[ r ( s ) | \\mathcal { T } , U ] \\sim \\mathcal { \\bar { G } } \\mathcal { P } ( 0 , k )", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 441, + 506, + 456 + ], + "score": 1.0, + "content": ", GP-UCB maintains", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 451, + 507, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 451, + 379, + 475 + ], + "score": 1.0, + "content": "an UCB of the reward after t-step asφ (s) := ¯rt (s) + λσt (s) ,", + "type": "text" + }, + { + "bbox": [ + 484, + 460, + 507, + 474 + ], + "score": 1.0, + "content": "(12)", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 142, + 474, + 168, + 485 + ], + "lines": [ + { + "bbox": [ + 140, + 473, + 171, + 486 + ], + "spans": [ + { + "bbox": [ + 140, + 473, + 171, + 486 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 484, + 428, + 514 + ], + "lines": [ + { + "bbox": [ + 217, + 484, + 428, + 514 + ], + "spans": [ + { + "bbox": [ + 217, + 484, + 428, + 514 + ], + "score": 0.8, + "content": "\\begin{array} { r c l } { { { \\bar { r } } _ { t } } \\left( s \\right) } & { { = } } & { { { k _ { t } } \\left( s \\right) \\left( K _ { t } + \\alpha I \\right) ^ { - 1 } { r _ { t } } , } } \\\\ { { \\sigma _ { t } ^ { 2 } \\left( s \\right) } } & { { = } } & { { { k \\left( s , s \\right) - k _ { t } } \\left( s \\right) ^ { \\top } \\left( K _ { t } + \\alpha I \\right) ^ { - 1 } k _ { t } \\left( s \\right) , } } \\end{array}", + "type": "interline_equation", + "image_path": "fbd68378e007a673e477047db489ceaba05b0f68da9498aa6be27ea6816d48ad.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 484, + 428, + 499.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 217, + 499.0, + 428, + 514.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 141, + 514, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 140, + 511, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 140, + 511, + 161, + 529 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 162, + 514, + 257, + 528 + ], + "score": 0.75, + "content": "\\boldsymbol { k } _ { t } \\left( \\boldsymbol { s } \\right) = \\left[ k \\left( s _ { i } , \\boldsymbol { s } \\right) \\right] _ { s _ { i } \\in \\boldsymbol { S } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 511, + 260, + 529 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 261, + 514, + 351, + 528 + ], + "score": 0.57, + "content": "K _ { t } = [ k \\left( s , s ^ { \\prime } \\right) ] _ { s , s ^ { \\prime } \\in S _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 511, + 370, + 529 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 371, + 514, + 457, + 526 + ], + "score": 0.91, + "content": "{ \\cal { S } } _ { t } = \\{ s _ { 1 } , s _ { 2 } , \\ldots , s _ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 511, + 506, + 529 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 141, + 527, + 409, + 540 + ], + "score": 1.0, + "content": "sequence of selected nodes in current trees. The variance estimation", + "type": "text" + }, + { + "bbox": [ + 410, + 527, + 435, + 540 + ], + "score": 0.75, + "content": "\\sigma _ { t } ^ { 2 } \\left( s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "takes the number", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 142, + 538, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 142, + 538, + 498, + 551 + ], + "score": 1.0, + "content": "of visits into account in an implicit way: the variance will reduce, as the neighborhood of", + "type": "text" + }, + { + "bbox": [ + 498, + 541, + 504, + 548 + ], + "score": 0.75, + "content": "s", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 549, + 338, + 561 + ], + "spans": [ + { + "bbox": [ + 141, + 549, + 338, + 561 + ], + "score": 1.0, + "content": "is visited more frequently (Srinivas et al., 2009).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 132, + 567, + 504, + 590 + ], + "lines": [ + { + "bbox": [ + 132, + 566, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 505, + 580 + ], + "score": 1.0, + "content": "• KS-UCB: We can also use kernel regression as an alternative parametrization for (7) (Yee", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 577, + 403, + 591 + ], + "spans": [ + { + "bbox": [ + 141, + 577, + 366, + 591 + ], + "score": 1.0, + "content": "et al., 2016), which leads to an UCB of the reward after", + "type": "text" + }, + { + "bbox": [ + 366, + 579, + 371, + 588 + ], + "score": 0.8, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 577, + 403, + 591 + ], + "score": 1.0, + "content": "-step as", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "interline_equation", + "bbox": [ + 270, + 590, + 376, + 604 + ], + "lines": [ + { + "bbox": [ + 270, + 590, + 376, + 604 + ], + "spans": [ + { + "bbox": [ + 270, + 590, + 376, + 604 + ], + "score": 0.85, + "content": "\\phi \\left( s \\right) : = { \\bar { r } } _ { t } \\left( s \\right) + \\lambda \\sigma _ { t } \\left( s \\right) ,", + "type": "interline_equation", + "image_path": "b215a1dbae611adf3ab8688a0d0700b31f4578d9a215b8ac318d12ccd63dc469.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 270, + 590, + 376, + 604 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 142, + 604, + 168, + 615 + ], + "lines": [ + { + "bbox": [ + 141, + 603, + 170, + 617 + ], + "spans": [ + { + "bbox": [ + 141, + 603, + 170, + 617 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 613, + 396, + 676 + ], + "lines": [ + { + "bbox": [ + 250, + 613, + 396, + 676 + ], + "spans": [ + { + "bbox": [ + 250, + 613, + 396, + 676 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } { \\bar { r } _ { t } \\left[ s \\right] } & { = } & { \\frac { \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) r \\left( s ^ { \\prime } \\right) } { \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) } , } \\\\ { \\sigma _ { t } \\left( s \\right) } & { = } & { \\sqrt { \\frac { \\log \\sum _ { s ^ { \\prime } \\in S _ { t } } w \\left( s ^ { \\prime } \\right) } { w \\left( s \\right) } } , } \\end{array}", + "type": "interline_equation", + "image_path": "f1483a983d3b9d895751aa88d2e56eed7a6c2602aaadb294a53ab188a30bd0d3.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 250, + 613, + 396, + 644.5 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 250, + 644.5, + 396, + 676.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 141, + 676, + 504, + 699 + ], + "lines": [ + { + "bbox": [ + 140, + 673, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 140, + 673, + 163, + 692 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 163, + 676, + 263, + 690 + ], + "score": 0.91, + "content": "\\begin{array} { r } { w \\left( s \\right) = \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 673, + 506, + 692 + ], + "score": 1.0, + "content": ". Clearly, the variance estimation is to promote exploration", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 142, + 687, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 142, + 687, + 294, + 700 + ], + "score": 1.0, + "content": "towards less frequently visited states.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "As we can see, in both two examples of the parametrized UCB, we parametrize the observed rewards,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 720, + 433, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 433, + 732 + ], + "score": 1.0, + "content": "leading to generalizable UCB for increased states by considering the correlations.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 506, + 138 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 105, + 82, + 506, + 139 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 505, + 157 + ], + "score": 1.0, + "content": "Inspired by the above computation pattern of the Bellman operator, value iteration networks design", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 244, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 244, + 166 + ], + "score": 1.0, + "content": "the neural architecture as follows,", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 141, + 505, + 166 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 165, + 375, + 207 + ], + "lines": [ + { + "bbox": [ + 235, + 165, + 375, + 207 + ], + "spans": [ + { + "bbox": [ + 235, + 165, + 375, + 207 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\tilde { V } ^ { * 0 } = \\operatorname* { m i n } \\ \\left( W _ { 1 } \\oplus \\left[ \\tt { m a p } , \\tilde { R } \\right] \\right) } \\\\ & { \\tilde { V } ^ { * t } = \\operatorname* { m i n } \\ \\left( W _ { 1 } \\oplus \\left[ \\tilde { V } ^ { * t - 1 } , \\tilde { R } \\right] \\right) } \\end{array}", + "type": "interline_equation", + "image_path": "dee3f8bfd36cb521afcd22f462a62dffa0835e6ce5931e74fd22a401561b036a.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 165, + 375, + 186.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 235, + 186.0, + 375, + 207.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 208, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 104, + 206, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 104, + 206, + 133, + 222 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 210, + 143, + 219 + ], + "score": 0.83, + "content": "\\oplus", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 206, + 304, + 222 + ], + "score": 1.0, + "content": "is the convolution operation, both map,", + "type": "text" + }, + { + "bbox": [ + 304, + 207, + 321, + 219 + ], + "score": 0.87, + "content": "\\tilde { V } ^ { * t }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 206, + 339, + 222 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 340, + 207, + 349, + 219 + ], + "score": 0.86, + "content": "\\tilde { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 206, + 365, + 222 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 365, + 208, + 389, + 219 + ], + "score": 0.89, + "content": "d \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 206, + 506, + 222 + ], + "score": 1.0, + "content": "matrices, and the parameter", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 107, + 220, + 121, + 231 + ], + "score": 0.88, + "content": "W _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 219, + 137, + 232 + ], + "score": 1.0, + "content": "are", + "type": "text" + }, + { + "bbox": [ + 137, + 220, + 148, + 231 + ], + "score": 0.88, + "content": "k _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 219, + 257, + 232 + ], + "score": 1.0, + "content": "convolution kernels of size", + "type": "text" + }, + { + "bbox": [ + 257, + 220, + 281, + 230 + ], + "score": 0.91, + "content": "k \\times k", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 219, + 444, + 232 + ], + "score": 1.0, + "content": ". The min implements the pooling across", + "type": "text" + }, + { + "bbox": [ + 444, + 220, + 455, + 231 + ], + "score": 0.88, + "content": "k _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "convolution", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 229, + 147, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 147, + 243 + ], + "score": 1.0, + "content": "channels.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 206, + 506, + 243 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 246, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 504, + 260 + ], + "score": 1.0, + "content": "The gated path planning networks (GPPN) (Lee et al., 2018) improves the VIN by replacing the VIN", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 258, + 317, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 317, + 272 + ], + "score": 1.0, + "content": "cell (9) with the well-established LSTM update, i.e.,", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 247, + 504, + 272 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 216, + 270, + 394, + 291 + ], + "lines": [ + { + "bbox": [ + 216, + 270, + 394, + 291 + ], + "spans": [ + { + "bbox": [ + 216, + 270, + 394, + 291 + ], + "score": 0.93, + "content": "\\tilde { V } ^ { t } , \\tilde { c } ^ { t } = \\mathrm { L S T M } \\left( \\sum \\left( W _ { 1 } \\oplus \\left[ \\tilde { V } ^ { t } , \\tilde { R } \\right] \\right) , \\tilde { c } ^ { t } \\right) ,", + "type": "interline_equation", + "image_path": "29d45bb013ca6188267f130a8fe3ab6c4c91b7fa1a03866c1230f433c3afccf3.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 216, + 270, + 394, + 291 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 291, + 378, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 291, + 379, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 277, + 304 + ], + "score": 1.0, + "content": "where the summation is taking over all the", + "type": "text" + }, + { + "bbox": [ + 278, + 292, + 289, + 303 + ], + "score": 0.87, + "content": "k _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 291, + 379, + 304 + ], + "score": 1.0, + "content": "convolution channels.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 291, + 379, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 309, + 503, + 338 + ], + "lines": [ + { + "bbox": [ + 102, + 304, + 507, + 330 + ], + "spans": [ + { + "bbox": [ + 102, + 304, + 392, + 330 + ], + "score": 1.0, + "content": "After constructing the VIN and GPPN, the parameters of the model, i.e.,", + "type": "text" + }, + { + "bbox": [ + 393, + 308, + 433, + 329 + ], + "score": 0.93, + "content": "\\left\\{ \\tilde { R } , W _ { 1 } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 304, + 507, + 330 + ], + "score": 1.0, + "content": "can be learned by", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 324, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 287, + 340 + ], + "score": 1.0, + "content": "imitation learning or reinforcement learning.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 102, + 304, + 507, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 504, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "The application of planning networks are restricted in low-dimension tasks. However, their success", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "enlightens our neural architecture for generalizable representation for high-dimension planning tasks.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 342, + 506, + 367 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 384, + 313, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 314, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 314, + 400 + ], + "score": 1.0, + "content": "C PARAMETRIZED UCB ALGORITHMS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 438, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 439, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 439, + 424 + ], + "score": 1.0, + "content": "We list two examples of parametrized UCB as the instantiation of (3) used in GPE:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 411, + 439, + 424 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 430, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 131, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "• GP-UCB: The GP-UCB Chu et al. (2011) is derived by parameterizing via Gaussian", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 140, + 441, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 140, + 441, + 255, + 456 + ], + "score": 1.0, + "content": "Processes (GP) with kernel", + "type": "text" + }, + { + "bbox": [ + 255, + 442, + 287, + 455 + ], + "score": 0.91, + "content": "k \\left( s , s ^ { \\prime } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 441, + 316, + 456 + ], + "score": 1.0, + "content": ", i.e., E", + "type": "text" + }, + { + "bbox": [ + 316, + 442, + 419, + 455 + ], + "score": 0.88, + "content": "[ r ( s ) | \\mathcal { T } , U ] \\sim \\mathcal { \\bar { G } } \\mathcal { P } ( 0 , k )", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 441, + 506, + 456 + ], + "score": 1.0, + "content": ", GP-UCB maintains", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 451, + 507, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 451, + 379, + 475 + ], + "score": 1.0, + "content": "an UCB of the reward after t-step asφ (s) := ¯rt (s) + λσt (s) ,", + "type": "text" + }, + { + "bbox": [ + 484, + 460, + 507, + 474 + ], + "score": 1.0, + "content": "(12)", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 131, + 430, + 507, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 142, + 474, + 168, + 485 + ], + "lines": [ + { + "bbox": [ + 140, + 473, + 171, + 486 + ], + "spans": [ + { + "bbox": [ + 140, + 473, + 171, + 486 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 140, + 473, + 171, + 486 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 484, + 428, + 514 + ], + "lines": [ + { + "bbox": [ + 217, + 484, + 428, + 514 + ], + "spans": [ + { + "bbox": [ + 217, + 484, + 428, + 514 + ], + "score": 0.8, + "content": "\\begin{array} { r c l } { { { \\bar { r } } _ { t } } \\left( s \\right) } & { { = } } & { { { k _ { t } } \\left( s \\right) \\left( K _ { t } + \\alpha I \\right) ^ { - 1 } { r _ { t } } , } } \\\\ { { \\sigma _ { t } ^ { 2 } \\left( s \\right) } } & { { = } } & { { { k \\left( s , s \\right) - k _ { t } } \\left( s \\right) ^ { \\top } \\left( K _ { t } + \\alpha I \\right) ^ { - 1 } k _ { t } \\left( s \\right) , } } \\end{array}", + "type": "interline_equation", + "image_path": "fbd68378e007a673e477047db489ceaba05b0f68da9498aa6be27ea6816d48ad.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 484, + 428, + 499.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 217, + 499.0, + 428, + 514.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 141, + 514, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 140, + 511, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 140, + 511, + 161, + 529 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 162, + 514, + 257, + 528 + ], + "score": 0.75, + "content": "\\boldsymbol { k } _ { t } \\left( \\boldsymbol { s } \\right) = \\left[ k \\left( s _ { i } , \\boldsymbol { s } \\right) \\right] _ { s _ { i } \\in \\boldsymbol { S } _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 511, + 260, + 529 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 261, + 514, + 351, + 528 + ], + "score": 0.57, + "content": "K _ { t } = [ k \\left( s , s ^ { \\prime } \\right) ] _ { s , s ^ { \\prime } \\in S _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 511, + 370, + 529 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 371, + 514, + 457, + 526 + ], + "score": 0.91, + "content": "{ \\cal { S } } _ { t } = \\{ s _ { 1 } , s _ { 2 } , \\ldots , s _ { t } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 511, + 506, + 529 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 141, + 527, + 409, + 540 + ], + "score": 1.0, + "content": "sequence of selected nodes in current trees. The variance estimation", + "type": "text" + }, + { + "bbox": [ + 410, + 527, + 435, + 540 + ], + "score": 0.75, + "content": "\\sigma _ { t } ^ { 2 } \\left( s \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "takes the number", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 142, + 538, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 142, + 538, + 498, + 551 + ], + "score": 1.0, + "content": "of visits into account in an implicit way: the variance will reduce, as the neighborhood of", + "type": "text" + }, + { + "bbox": [ + 498, + 541, + 504, + 548 + ], + "score": 0.75, + "content": "s", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 549, + 338, + 561 + ], + "spans": [ + { + "bbox": [ + 141, + 549, + 338, + 561 + ], + "score": 1.0, + "content": "is visited more frequently (Srinivas et al., 2009).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 140, + 511, + 506, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 567, + 504, + 590 + ], + "lines": [ + { + "bbox": [ + 132, + 566, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 505, + 580 + ], + "score": 1.0, + "content": "• KS-UCB: We can also use kernel regression as an alternative parametrization for (7) (Yee", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 577, + 403, + 591 + ], + "spans": [ + { + "bbox": [ + 141, + 577, + 366, + 591 + ], + "score": 1.0, + "content": "et al., 2016), which leads to an UCB of the reward after", + "type": "text" + }, + { + "bbox": [ + 366, + 579, + 371, + 588 + ], + "score": 0.8, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 577, + 403, + 591 + ], + "score": 1.0, + "content": "-step as", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 132, + 566, + 505, + 591 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 270, + 590, + 376, + 604 + ], + "lines": [ + { + "bbox": [ + 270, + 590, + 376, + 604 + ], + "spans": [ + { + "bbox": [ + 270, + 590, + 376, + 604 + ], + "score": 0.85, + "content": "\\phi \\left( s \\right) : = { \\bar { r } } _ { t } \\left( s \\right) + \\lambda \\sigma _ { t } \\left( s \\right) ,", + "type": "interline_equation", + "image_path": "b215a1dbae611adf3ab8688a0d0700b31f4578d9a215b8ac318d12ccd63dc469.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 270, + 590, + 376, + 604 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 142, + 604, + 168, + 615 + ], + "lines": [ + { + "bbox": [ + 141, + 603, + 170, + 617 + ], + "spans": [ + { + "bbox": [ + 141, + 603, + 170, + 617 + ], + "score": 1.0, + "content": "where", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 141, + 603, + 170, + 617 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 613, + 396, + 676 + ], + "lines": [ + { + "bbox": [ + 250, + 613, + 396, + 676 + ], + "spans": [ + { + "bbox": [ + 250, + 613, + 396, + 676 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } { \\bar { r } _ { t } \\left[ s \\right] } & { = } & { \\frac { \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) r \\left( s ^ { \\prime } \\right) } { \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) } , } \\\\ { \\sigma _ { t } \\left( s \\right) } & { = } & { \\sqrt { \\frac { \\log \\sum _ { s ^ { \\prime } \\in S _ { t } } w \\left( s ^ { \\prime } \\right) } { w \\left( s \\right) } } , } \\end{array}", + "type": "interline_equation", + "image_path": "f1483a983d3b9d895751aa88d2e56eed7a6c2602aaadb294a53ab188a30bd0d3.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 250, + 613, + 396, + 644.5 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 250, + 644.5, + 396, + 676.0 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 141, + 676, + 504, + 699 + ], + "lines": [ + { + "bbox": [ + 140, + 673, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 140, + 673, + 163, + 692 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 163, + 676, + 263, + 690 + ], + "score": 0.91, + "content": "\\begin{array} { r } { w \\left( s \\right) = \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 673, + 506, + 692 + ], + "score": 1.0, + "content": ". Clearly, the variance estimation is to promote exploration", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 142, + 687, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 142, + 687, + 294, + 700 + ], + "score": 1.0, + "content": "towards less frequently visited states.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 140, + 673, + 506, + 700 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "As we can see, in both two examples of the parametrized UCB, we parametrize the observed rewards,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 720, + 433, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 433, + 732 + ], + "score": 1.0, + "content": "leading to generalizable UCB for increased states by considering the correlations.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 709, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 374, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 375, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 375, + 96 + ], + "score": 1.0, + "content": "D POLICY AND VALUE NETWORK ARCHITECTURE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 129, + 506, + 162 + ], + "lines": [ + { + "bbox": [ + 105, + 128, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 506, + 142 + ], + "score": 1.0, + "content": "We explain the implementation details of the proposed parametrization for policy and value function.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 139, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 506, + 153 + ], + "score": 1.0, + "content": "Figure 12 and Figure 13 are neural architectures for the attention module, the policy/value network,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 151, + 262, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 262, + 164 + ], + "score": 1.0, + "content": "and the planning module, respectively.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 118, + 164, + 494, + 543 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 164, + 494, + 543 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 164, + 494, + 543 + ], + "spans": [ + { + "bbox": [ + 118, + 164, + 494, + 543 + ], + "score": 0.975, + "type": "image", + "image_path": "fb3272bba1630226b3a65a134ae6d6e537e4da69dcd4e45215968c4d3b3ba712.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 118, + 164, + 494, + 290.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 118, + 290.3333333333333, + 494, + 416.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 118, + 416.66666666666663, + 494, + 543.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 550, + 505, + 573 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 550, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 507, + 564 + ], + "score": 1.0, + "content": "Figure 12: Left: attention module, instantiating the Figure 3; Right: policy/value network, instantiat-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 562, + 175, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 175, + 574 + ], + "score": 1.0, + "content": "ing the Figure 4.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "In the figures, we use rectangle blocks to denote inputs, intermediate results and outputs, stadium", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 588, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 600 + ], + "score": 1.0, + "content": "shape blocks to denote operations, and rounded rectangle blocks to denote modules. We use different", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "colors for different operations. In particular, we use blue for convolutional/LSTM layers, green for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 403, + 622 + ], + "score": 1.0, + "content": "dense layers, and orange for anything else. For convolutional layers, \"Conv", + "type": "text" + }, + { + "bbox": [ + 403, + 610, + 426, + 621 + ], + "score": 0.83, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 610, + 506, + 622 + ], + "score": 1.0, + "content": ", 32, relu\" denotes a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 621, + 507, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 148, + 634 + ], + "score": 1.0, + "content": "layer with", + "type": "text" + }, + { + "bbox": [ + 148, + 621, + 171, + 631 + ], + "score": 0.89, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 621, + 507, + 634 + ], + "score": 1.0, + "content": "kernels, 32 channels, followed by a rectified linear unit; for dense layers, \"Dense, 64,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 632, + 372, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 372, + 644 + ], + "score": 1.0, + "content": "relu\" denotes a layer of size 64, followed by a rectified linear unit.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 339, + 660 + ], + "score": 1.0, + "content": "The attention module (Figure 12-left) embeds a state to a", + "type": "text" + }, + { + "bbox": [ + 339, + 649, + 385, + 660 + ], + "score": 0.92, + "content": "d \\times d \\times d _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "tensor. The planning module", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "(Figure 13) is a one-step LSTM update which takes the result of a convolutional layer as input. Both", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 297, + 683 + ], + "score": 1.0, + "content": "the input and hidden size of the LSTM cell are", + "type": "text" + }, + { + "bbox": [ + 298, + 671, + 308, + 682 + ], + "score": 0.87, + "content": "d _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 671, + 328, + 683 + ], + "score": 1.0, + "content": ". All", + "type": "text" + }, + { + "bbox": [ + 328, + 671, + 352, + 681 + ], + "score": 0.89, + "content": "d \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "locations share one set of parameters", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 682, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 289, + 693 + ], + "score": 1.0, + "content": "and are processed by the LSTM in one batch.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "score": 1.0, + "content": "The main architecture is illustrated in Figure 12-right. It takes maze map, state and goal as input,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 451, + 722 + ], + "score": 1.0, + "content": "and outputs the action and the value. Refer to Section 4.2 for details for computing", + "type": "text" + }, + { + "bbox": [ + 452, + 710, + 472, + 721 + ], + "score": 0.9, + "content": "\\psi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 710, + 505, + 722 + ], + "score": 1.0, + "content": ". In our", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 720, + 470, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 350, + 734 + ], + "score": 1.0, + "content": "experiments, we set the values of the hyper-parameters to be", + "type": "text" + }, + { + "bbox": [ + 350, + 720, + 465, + 732 + ], + "score": 0.92, + "content": "( d , d _ { e } , d _ { a } , p ) = ( 1 \\bar { 5 } , 6 4 , \\bar { 8 } , 8 )", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 720, + 470, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 374, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 375, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 375, + 96 + ], + "score": 1.0, + "content": "D POLICY AND VALUE NETWORK ARCHITECTURE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 129, + 506, + 162 + ], + "lines": [ + { + "bbox": [ + 105, + 128, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 506, + 142 + ], + "score": 1.0, + "content": "We explain the implementation details of the proposed parametrization for policy and value function.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 139, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 506, + 153 + ], + "score": 1.0, + "content": "Figure 12 and Figure 13 are neural architectures for the attention module, the policy/value network,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 151, + 262, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 262, + 164 + ], + "score": 1.0, + "content": "and the planning module, respectively.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 128, + 506, + 164 + ] + }, + { + "type": "image", + "bbox": [ + 118, + 164, + 494, + 543 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 164, + 494, + 543 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 164, + 494, + 543 + ], + "spans": [ + { + "bbox": [ + 118, + 164, + 494, + 543 + ], + "score": 0.975, + "type": "image", + "image_path": "fb3272bba1630226b3a65a134ae6d6e537e4da69dcd4e45215968c4d3b3ba712.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 118, + 164, + 494, + 290.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 118, + 290.3333333333333, + 494, + 416.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 118, + 416.66666666666663, + 494, + 543.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 550, + 505, + 573 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 550, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 507, + 564 + ], + "score": 1.0, + "content": "Figure 12: Left: attention module, instantiating the Figure 3; Right: policy/value network, instantiat-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 562, + 175, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 175, + 574 + ], + "score": 1.0, + "content": "ing the Figure 4.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + } + ], + "index": 6.25 + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "In the figures, we use rectangle blocks to denote inputs, intermediate results and outputs, stadium", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 588, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 600 + ], + "score": 1.0, + "content": "shape blocks to denote operations, and rounded rectangle blocks to denote modules. We use different", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "colors for different operations. In particular, we use blue for convolutional/LSTM layers, green for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 403, + 622 + ], + "score": 1.0, + "content": "dense layers, and orange for anything else. For convolutional layers, \"Conv", + "type": "text" + }, + { + "bbox": [ + 403, + 610, + 426, + 621 + ], + "score": 0.83, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 610, + 506, + 622 + ], + "score": 1.0, + "content": ", 32, relu\" denotes a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 621, + 507, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 148, + 634 + ], + "score": 1.0, + "content": "layer with", + "type": "text" + }, + { + "bbox": [ + 148, + 621, + 171, + 631 + ], + "score": 0.89, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 621, + 507, + 634 + ], + "score": 1.0, + "content": "kernels, 32 channels, followed by a rectified linear unit; for dense layers, \"Dense, 64,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 632, + 372, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 372, + 644 + ], + "score": 1.0, + "content": "relu\" denotes a layer of size 64, followed by a rectified linear unit.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 576, + 507, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 339, + 660 + ], + "score": 1.0, + "content": "The attention module (Figure 12-left) embeds a state to a", + "type": "text" + }, + { + "bbox": [ + 339, + 649, + 385, + 660 + ], + "score": 0.92, + "content": "d \\times d \\times d _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "tensor. The planning module", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "(Figure 13) is a one-step LSTM update which takes the result of a convolutional layer as input. Both", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 297, + 683 + ], + "score": 1.0, + "content": "the input and hidden size of the LSTM cell are", + "type": "text" + }, + { + "bbox": [ + 298, + 671, + 308, + 682 + ], + "score": 0.87, + "content": "d _ { e }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 671, + 328, + 683 + ], + "score": 1.0, + "content": ". All", + "type": "text" + }, + { + "bbox": [ + 328, + 671, + 352, + 681 + ], + "score": 0.89, + "content": "d \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "locations share one set of parameters", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 682, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 289, + 693 + ], + "score": 1.0, + "content": "and are processed by the LSTM in one batch.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 649, + 505, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 506, + 712 + ], + "score": 1.0, + "content": "The main architecture is illustrated in Figure 12-right. It takes maze map, state and goal as input,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 451, + 722 + ], + "score": 1.0, + "content": "and outputs the action and the value. Refer to Section 4.2 for details for computing", + "type": "text" + }, + { + "bbox": [ + 452, + 710, + 472, + 721 + ], + "score": 0.9, + "content": "\\psi ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 710, + 505, + 722 + ], + "score": 1.0, + "content": ". In our", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 720, + 470, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 350, + 734 + ], + "score": 1.0, + "content": "experiments, we set the values of the hyper-parameters to be", + "type": "text" + }, + { + "bbox": [ + 350, + 720, + 465, + 732 + ], + "score": 0.92, + "content": "( d , d _ { e } , d _ { a } , p ) = ( 1 \\bar { 5 } , 6 4 , \\bar { 8 } , 8 )", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 720, + 470, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 697, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 229, + 81, + 381, + 239 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 229, + 81, + 381, + 239 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 229, + 81, + 381, + 239 + ], + "spans": [ + { + "bbox": [ + 229, + 81, + 381, + 239 + ], + "score": 0.968, + "type": "image", + "image_path": "093f34c3bdcd9d32756f1eaf06749627ebe9c0724c310e6d0259da319ce67d5a.jpg" + } + ] + } + ], + "index": 0.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 81, + 381, + 160.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 229, + 160.0, + 381, + 239.0 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 249, + 248, + 362, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 248, + 247, + 362, + 261 + ], + "spans": [ + { + "bbox": [ + 248, + 247, + 362, + 261 + ], + "score": 1.0, + "content": "Figure 13: planning module", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + } + ], + "index": 1.25 + }, + { + "type": "title", + "bbox": [ + 106, + 262, + 244, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 245, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 245, + 277 + ], + "score": 1.0, + "content": "E EXPERIMENT DETAILS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 108, + 289, + 261, + 300 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 263, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 263, + 302 + ], + "score": 1.0, + "content": "E.1 BENCHMARK ENVIRONMENTS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 506, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "We used four benchmark environments in our experiment. For the first three, the workspace dimension", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 322, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 334 + ], + "score": 1.0, + "content": "is 2d. We generated the maze maps with the recursive backtracker algorithm using the following imple-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 332, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 345 + ], + "score": 1.0, + "content": "mentation: https://github.com/lileee/gated-path-planning-networks/blob/master/generate_dataset.py.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "Examples of the workspace are shown in Figure 15. Three environments differ in the choice of robots:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 506, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 506, + 373 + ], + "score": 1.0, + "content": "• Workspace planning (2d). The robot is abstracted with a point mass moving in the plane. Without", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 371, + 400, + 384 + ], + "spans": [ + { + "bbox": [ + 114, + 371, + 400, + 384 + ], + "score": 1.0, + "content": "higher dimensions, this problem reduces to planning in the workspace.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 108, + 383, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 108, + 383, + 505, + 395 + ], + "score": 1.0, + "content": "• Rigid body navigation (3d). A rigid body robot, abstracted as a thin rectangle, is used here. The", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "extra rotation dimension is added to the planning problem. This robot can rotate and move freely", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 404, + 282, + 417 + ], + "spans": [ + { + "bbox": [ + 114, + 404, + 282, + 417 + ], + "score": 1.0, + "content": "without any constraints in the free space.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 108, + 414, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 108, + 414, + 506, + 428 + ], + "score": 1.0, + "content": "• 3-link snake (5d). The robot is a 5 DoF snake with two joints. Two more angle dimensions are", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 116, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "added to the planning task. To prevent links from folding, we restrict the angles to the range of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 437, + 171, + 451 + ], + "spans": [ + { + "bbox": [ + 117, + 437, + 167, + 450 + ], + "score": 0.91, + "content": "[ - \\pi / 4 , \\pi / 4 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 437, + 171, + 451 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 454, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "The fourth environment has a 3d workspace. Cuboid obstacles were generated uniformly randomly", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 190, + 478 + ], + "score": 1.0, + "content": "in space with density", + "type": "text" + }, + { + "bbox": [ + 191, + 465, + 221, + 476 + ], + "score": 0.9, + "content": "\\approx 2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 464, + 505, + 478 + ], + "score": 1.0, + "content": ". Example of the workspace is shown in Figure 6 and 16, where the blue", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 475, + 347, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 347, + 488 + ], + "score": 1.0, + "content": "cuboids are obstacles. The environment is described below:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 108, + 492, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 108, + 492, + 505, + 504 + ], + "score": 1.0, + "content": "• Spacecraft planning (7d). The robot is a spacecraft with a cuboid body and two 2 DoF arms", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "connecting to two opposite sides of the body. There is a joint in the middle of each arm. The", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "outer arm can rotate around this joint. Each arm can also rotate as a whole around its connection", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 525, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 114, + 525, + 387, + 539 + ], + "score": 1.0, + "content": "point with the body. All rotation angles are restricted in the range of", + "type": "text" + }, + { + "bbox": [ + 387, + 525, + 418, + 538 + ], + "score": 0.93, + "content": "[ 0 , \\pi / 2 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 525, + 506, + 539 + ], + "score": 1.0, + "content": ". The spacecraft itself", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 538, + 173, + 548 + ], + "spans": [ + { + "bbox": [ + 115, + 538, + 173, + 548 + ], + "score": 1.0, + "content": "cannot rotate.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 563, + 262, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 263, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 263, + 576 + ], + "score": 1.0, + "content": "E.2 HYPERPARAMETER FOR MSIL", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 585, + 504, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 504, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 498, + 598 + ], + "score": 1.0, + "content": "During self-improving over the first 2000 problems, NEXT updated its parameters and annealed", + "type": "text" + }, + { + "bbox": [ + 498, + 587, + 504, + 595 + ], + "score": 0.5, + "content": "\\epsilon", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 595, + 442, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 334, + 610 + ], + "score": 1.0, + "content": "once for every 200 problems. The value of the annealing", + "type": "text" + }, + { + "bbox": [ + 335, + 598, + 340, + 606 + ], + "score": 0.62, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 595, + 442, + 610 + ], + "score": 1.0, + "content": "was set as the following:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 609, + 424, + 649 + ], + "lines": [ + { + "bbox": [ + 185, + 609, + 424, + 649 + ], + "spans": [ + { + "bbox": [ + 185, + 609, + 424, + 649 + ], + "score": 0.93, + "content": "\\epsilon = \\left\\{ \\begin{array} { l l } { 1 , } & { \\mathrm { i f ~ } i < 1 0 0 0 , } \\\\ { 0 . 5 - 0 . 1 \\cdot \\lfloor ( i - 1 0 0 0 ) / 2 0 0 \\rfloor , } & { \\mathrm { i f ~ } 1 0 0 0 \\leqslant i < 2 0 0 0 , } \\\\ { 0 . 1 , } & { \\mathrm { o t h e r w i s e , } } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "8d768de638d1253288962d6fc415fd345b9b974c6cbfafddd2aa591a51b77228.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 185, + 609, + 424, + 622.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 185, + 622.3333333333334, + 424, + 635.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 185, + 635.6666666666667, + 424, + 649.0000000000001 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 254, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 254, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 126, + 664 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 651, + 131, + 660 + ], + "score": 0.79, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 648, + 254, + 664 + ], + "score": 1.0, + "content": "denoting the problem number.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 108, + 677, + 277, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 279, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 279, + 689 + ], + "score": 1.0, + "content": "E.3 BASELINE: THE IMPROVED GPPN", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 507, + 712 + ], + "score": 1.0, + "content": "The original GPPN is not directly applicable to our experiments. Inspired by Tamar et al. (2016),", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "we add a fully-connected MLP to its final layers, so that the improved architecture can be applied to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "high-dimensional continuous domain. As shown in Figure 14, the GPPN first processes the discretized", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 229, + 81, + 381, + 239 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 229, + 81, + 381, + 239 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 229, + 81, + 381, + 239 + ], + "spans": [ + { + "bbox": [ + 229, + 81, + 381, + 239 + ], + "score": 0.968, + "type": "image", + "image_path": "093f34c3bdcd9d32756f1eaf06749627ebe9c0724c310e6d0259da319ce67d5a.jpg" + } + ] + } + ], + "index": 0.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 81, + 381, + 160.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 229, + 160.0, + 381, + 239.0 + ], + "spans": [], + "index": 1 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 249, + 248, + 362, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 248, + 247, + 362, + 261 + ], + "spans": [ + { + "bbox": [ + 248, + 247, + 362, + 261 + ], + "score": 1.0, + "content": "Figure 13: planning module", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + } + ], + "index": 1.25 + }, + { + "type": "title", + "bbox": [ + 106, + 262, + 244, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 245, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 245, + 277 + ], + "score": 1.0, + "content": "E EXPERIMENT DETAILS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 108, + 289, + 261, + 300 + ], + "lines": [ + { + "bbox": [ + 105, + 288, + 263, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 263, + 302 + ], + "score": 1.0, + "content": "E.1 BENCHMARK ENVIRONMENTS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 506, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "We used four benchmark environments in our experiment. For the first three, the workspace dimension", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 322, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 334 + ], + "score": 1.0, + "content": "is 2d. We generated the maze maps with the recursive backtracker algorithm using the following imple-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 332, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 345 + ], + "score": 1.0, + "content": "mentation: https://github.com/lileee/gated-path-planning-networks/blob/master/generate_dataset.py.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "Examples of the workspace are shown in Figure 15. Three environments differ in the choice of robots:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 311, + 506, + 356 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 360, + 506, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 506, + 373 + ], + "score": 1.0, + "content": "• Workspace planning (2d). The robot is abstracted with a point mass moving in the plane. Without", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 371, + 400, + 384 + ], + "spans": [ + { + "bbox": [ + 114, + 371, + 400, + 384 + ], + "score": 1.0, + "content": "higher dimensions, this problem reduces to planning in the workspace.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 383, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 108, + 383, + 505, + 395 + ], + "score": 1.0, + "content": "• Rigid body navigation (3d). A rigid body robot, abstracted as a thin rectangle, is used here. The", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "extra rotation dimension is added to the planning problem. This robot can rotate and move freely", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 404, + 282, + 417 + ], + "spans": [ + { + "bbox": [ + 114, + 404, + 282, + 417 + ], + "score": 1.0, + "content": "without any constraints in the free space.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 414, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 108, + 414, + 506, + 428 + ], + "score": 1.0, + "content": "• 3-link snake (5d). The robot is a 5 DoF snake with two joints. Two more angle dimensions are", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 116, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "added to the planning task. To prevent links from folding, we restrict the angles to the range of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 437, + 171, + 451 + ], + "spans": [ + { + "bbox": [ + 117, + 437, + 167, + 450 + ], + "score": 0.91, + "content": "[ - \\pi / 4 , \\pi / 4 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 437, + 171, + 451 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 360, + 506, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 454, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "The fourth environment has a 3d workspace. Cuboid obstacles were generated uniformly randomly", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 190, + 478 + ], + "score": 1.0, + "content": "in space with density", + "type": "text" + }, + { + "bbox": [ + 191, + 465, + 221, + 476 + ], + "score": 0.9, + "content": "\\approx 2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 464, + 505, + 478 + ], + "score": 1.0, + "content": ". Example of the workspace is shown in Figure 6 and 16, where the blue", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 475, + 347, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 347, + 488 + ], + "score": 1.0, + "content": "cuboids are obstacles. The environment is described below:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 453, + 505, + 488 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 108, + 492, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 108, + 492, + 505, + 504 + ], + "score": 1.0, + "content": "• Spacecraft planning (7d). The robot is a spacecraft with a cuboid body and two 2 DoF arms", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "connecting to two opposite sides of the body. There is a joint in the middle of each arm. The", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "outer arm can rotate around this joint. Each arm can also rotate as a whole around its connection", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 525, + 506, + 539 + ], + "spans": [ + { + "bbox": [ + 114, + 525, + 387, + 539 + ], + "score": 1.0, + "content": "point with the body. All rotation angles are restricted in the range of", + "type": "text" + }, + { + "bbox": [ + 387, + 525, + 418, + 538 + ], + "score": 0.93, + "content": "[ 0 , \\pi / 2 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 525, + 506, + 539 + ], + "score": 1.0, + "content": ". The spacecraft itself", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 538, + 173, + 548 + ], + "spans": [ + { + "bbox": [ + 115, + 538, + 173, + 548 + ], + "score": 1.0, + "content": "cannot rotate.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 108, + 492, + 506, + 548 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 563, + 262, + 575 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 263, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 263, + 576 + ], + "score": 1.0, + "content": "E.2 HYPERPARAMETER FOR MSIL", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 585, + 504, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 504, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 498, + 598 + ], + "score": 1.0, + "content": "During self-improving over the first 2000 problems, NEXT updated its parameters and annealed", + "type": "text" + }, + { + "bbox": [ + 498, + 587, + 504, + 595 + ], + "score": 0.5, + "content": "\\epsilon", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 595, + 442, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 334, + 610 + ], + "score": 1.0, + "content": "once for every 200 problems. The value of the annealing", + "type": "text" + }, + { + "bbox": [ + 335, + 598, + 340, + 606 + ], + "score": 0.62, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 595, + 442, + 610 + ], + "score": 1.0, + "content": "was set as the following:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 585, + 504, + 610 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 185, + 609, + 424, + 649 + ], + "lines": [ + { + "bbox": [ + 185, + 609, + 424, + 649 + ], + "spans": [ + { + "bbox": [ + 185, + 609, + 424, + 649 + ], + "score": 0.93, + "content": "\\epsilon = \\left\\{ \\begin{array} { l l } { 1 , } & { \\mathrm { i f ~ } i < 1 0 0 0 , } \\\\ { 0 . 5 - 0 . 1 \\cdot \\lfloor ( i - 1 0 0 0 ) / 2 0 0 \\rfloor , } & { \\mathrm { i f ~ } 1 0 0 0 \\leqslant i < 2 0 0 0 , } \\\\ { 0 . 1 , } & { \\mathrm { o t h e r w i s e , } } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "8d768de638d1253288962d6fc415fd345b9b974c6cbfafddd2aa591a51b77228.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 185, + 609, + 424, + 622.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 185, + 622.3333333333334, + 424, + 635.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 185, + 635.6666666666667, + 424, + 649.0000000000001 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 254, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 254, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 126, + 664 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 651, + 131, + 660 + ], + "score": 0.79, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 648, + 254, + 664 + ], + "score": 1.0, + "content": "denoting the problem number.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 648, + 254, + 664 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 677, + 277, + 689 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 279, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 279, + 689 + ], + "score": 1.0, + "content": "E.3 BASELINE: THE IMPROVED GPPN", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 507, + 712 + ], + "score": 1.0, + "content": "The original GPPN is not directly applicable to our experiments. Inspired by Tamar et al. (2016),", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "we add a fully-connected MLP to its final layers, so that the improved architecture can be applied to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "high-dimensional continuous domain. As shown in Figure 14, the GPPN first processes the discretized", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 698, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 84, + 484, + 249 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 84, + 484, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 84, + 484, + 249 + ], + "spans": [ + { + "bbox": [ + 129, + 84, + 484, + 249 + ], + "score": 0.966, + "type": "image", + "image_path": "1197e1a0f579ba28f10b818c4b2e442327f2cb38b96543bbc6dbcfb1e5ec6325.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 84, + 484, + 139.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 139.0, + 484, + 194.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 194.0, + 484, + 249.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 224, + 263, + 387, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 223, + 262, + 387, + 276 + ], + "spans": [ + { + "bbox": [ + 223, + 262, + 387, + 276 + ], + "score": 1.0, + "content": "Figure 14: Improved GPPN architecture", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 280, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 292 + ], + "score": 1.0, + "content": "workspace locations. Its output and the full robot configurations are processed together by the MLP,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 289, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 304 + ], + "score": 1.0, + "content": "which then produces the current value and action estimates. The improved GPPN is trained using", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 302, + 352, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 352, + 314 + ], + "score": 1.0, + "content": "supervisions from the near-optimal paths produced by RRT*.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 328, + 245, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 327, + 245, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 245, + 342 + ], + "score": 1.0, + "content": "F EXPERIMENT RESULTS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 118, + 367, + 494, + 493 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 367, + 494, + 493 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 118, + 367, + 494, + 493 + ], + "spans": [ + { + "bbox": [ + 118, + 367, + 494, + 493 + ], + "score": 0.955, + "type": "image", + "image_path": "45835db4516f023ab21e7a5a2bb8b4907c34df7064909a1410c42cc4f9f92003.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 118, + 367, + 494, + 409.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 118, + 409.0, + 494, + 451.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 118, + 451.0, + 494, + 493.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 503, + 505, + 537 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 517 + ], + "score": 1.0, + "content": "Figure 15: The solution path produced by NEXT in a workspace planning task (2d), rigid body", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "navigation task (3d), 3-link snake task (5d) from left to right. The orange dot and the brown dot are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 525, + 269, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 269, + 539 + ], + "score": 1.0, + "content": "starting and goal locations, respectively.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "image", + "bbox": [ + 123, + 543, + 490, + 627 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 543, + 490, + 627 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 543, + 490, + 627 + ], + "spans": [ + { + "bbox": [ + 123, + 543, + 490, + 627 + ], + "score": 0.966, + "type": "image", + "image_path": "28558e709732ca4be6503115df6e181bb722fcd64cd5ffe0ac9a8fec10b06a19.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 123, + 543, + 490, + 571.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 123, + 571.0, + 490, + 599.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 123, + 599.0, + 490, + 627.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 642, + 504, + 665 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "Figure 16: The solution paths produced by NEXT in spacecraft planning task (7d). Spacecraft has a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 653, + 378, + 666 + ], + "spans": [ + { + "bbox": [ + 104, + 653, + 378, + 666 + ], + "score": 1.0, + "content": "yellow body and two 2 DoF (red) arms. Blue cuboids are obstacles.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 16.25 + }, + { + "type": "title", + "bbox": [ + 106, + 678, + 308, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 309, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 309, + 691 + ], + "score": 1.0, + "content": "F.2 DETAILS OF QUANTITATIVE EVALUATION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "More detailed results are shown in Table 1, 2, 3, including learning-based and non-learning-based", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "ones, on the last 1000 problems in each experiment. We normalized the number of collision checks", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "and the cost of paths based on the solution of RRT∗. The success rate result is not normalized. The", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 84, + 484, + 249 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 129, + 84, + 484, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 84, + 484, + 249 + ], + "spans": [ + { + "bbox": [ + 129, + 84, + 484, + 249 + ], + "score": 0.966, + "type": "image", + "image_path": "1197e1a0f579ba28f10b818c4b2e442327f2cb38b96543bbc6dbcfb1e5ec6325.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 129, + 84, + 484, + 139.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 129, + 139.0, + 484, + 194.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 129, + 194.0, + 484, + 249.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 224, + 263, + 387, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 223, + 262, + 387, + 276 + ], + "spans": [ + { + "bbox": [ + 223, + 262, + 387, + 276 + ], + "score": 1.0, + "content": "Figure 14: Improved GPPN architecture", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 280, + 505, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 506, + 292 + ], + "score": 1.0, + "content": "workspace locations. Its output and the full robot configurations are processed together by the MLP,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 289, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 304 + ], + "score": 1.0, + "content": "which then produces the current value and action estimates. The improved GPPN is trained using", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 302, + 352, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 352, + 314 + ], + "score": 1.0, + "content": "supervisions from the near-optimal paths produced by RRT*.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 280, + 506, + 314 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 328, + 245, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 327, + 245, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 245, + 342 + ], + "score": 1.0, + "content": "F EXPERIMENT RESULTS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 118, + 367, + 494, + 493 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 367, + 494, + 493 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 118, + 367, + 494, + 493 + ], + "spans": [ + { + "bbox": [ + 118, + 367, + 494, + 493 + ], + "score": 0.955, + "type": "image", + "image_path": "45835db4516f023ab21e7a5a2bb8b4907c34df7064909a1410c42cc4f9f92003.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 118, + 367, + 494, + 409.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 118, + 409.0, + 494, + 451.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 118, + 451.0, + 494, + 493.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 503, + 505, + 537 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 502, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 517 + ], + "score": 1.0, + "content": "Figure 15: The solution path produced by NEXT in a workspace planning task (2d), rigid body", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "navigation task (3d), 3-link snake task (5d) from left to right. The orange dot and the brown dot are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 525, + 269, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 269, + 539 + ], + "score": 1.0, + "content": "starting and goal locations, respectively.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "image", + "bbox": [ + 123, + 543, + 490, + 627 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 543, + 490, + 627 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 543, + 490, + 627 + ], + "spans": [ + { + "bbox": [ + 123, + 543, + 490, + 627 + ], + "score": 0.966, + "type": "image", + "image_path": "28558e709732ca4be6503115df6e181bb722fcd64cd5ffe0ac9a8fec10b06a19.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 123, + 543, + 490, + 571.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 123, + 571.0, + 490, + 599.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 123, + 599.0, + 490, + 627.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 642, + 504, + 665 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "Figure 16: The solution paths produced by NEXT in spacecraft planning task (7d). Spacecraft has a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 653, + 378, + 666 + ], + "spans": [ + { + "bbox": [ + 104, + 653, + 378, + 666 + ], + "score": 1.0, + "content": "yellow body and two 2 DoF (red) arms. Blue cuboids are obstacles.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 16.25 + }, + { + "type": "title", + "bbox": [ + 106, + 678, + 308, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 309, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 309, + 691 + ], + "score": 1.0, + "content": "F.2 DETAILS OF QUANTITATIVE EVALUATION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "More detailed results are shown in Table 1, 2, 3, including learning-based and non-learning-based", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "ones, on the last 1000 problems in each experiment. We normalized the number of collision checks", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "and the cost of paths based on the solution of RRT∗. The success rate result is not normalized. The", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "best planners in each experiment are in bold. NEXT-KS and NEXT-GP outperform the current", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 326, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 326, + 106 + ], + "score": 1.0, + "content": "state-of-the-art planning algorithm with large margins.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "best planners in each experiment are in bold. NEXT-KS and NEXT-GP outperform the current", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 326, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 326, + 106 + ], + "score": 1.0, + "content": "state-of-the-art planning algorithm with large margins.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table", + "bbox": [ + 110, + 135, + 499, + 188 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 142, + 115, + 468, + 127 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 113, + 468, + 129 + ], + "spans": [ + { + "bbox": [ + 141, + 113, + 468, + 129 + ], + "score": 1.0, + "content": "Table 1: Success rate results. The higher the better. NEXT-KS performs the best.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 110, + 135, + 499, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 135, + 499, + 188 + ], + "spans": [ + { + "bbox": [ + 110, + 135, + 499, + 188 + ], + "score": 0.978, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.9880.9810.7180.6320.7350.7100.1850.5350.720
3d0.9430.8410.6890.5540.4900.5140.1210.1140.498
5d0.8830.7680.6330.5150.4550.4970.0300.4760.444
7d0.9310.9060.6340.3690.3610.8140.2880.2720.370
", + "type": "table", + "image_path": "8ab1e479cbe5f5dc1a3e294cb9eca675bb59491b24d146a5477f9e869df469a1.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 110, + 135, + 499, + 152.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 110, + 152.66666666666666, + 499, + 170.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 170.33333333333331, + 499, + 187.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 110, + 229, + 499, + 282 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 199, + 504, + 221 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 199, + 504, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 504, + 210 + ], + "score": 1.0, + "content": "Table 2: Average number of collision checks results. The lower the better. The score is normalized", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 209, + 420, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 204, + 222 + ], + "score": 1.0, + "content": "based on the solution of", + "type": "text" + }, + { + "bbox": [ + 204, + 210, + 228, + 220 + ], + "score": 0.31, + "content": "\\mathbf { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 209, + 420, + 222 + ], + "score": 1.0, + "content": ". NEXT-KS performs the best in 3 benchmarks.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "table_body", + "bbox": [ + 110, + 229, + 499, + 282 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 229, + 499, + 282 + ], + "spans": [ + { + "bbox": [ + 110, + 229, + 499, + 282 + ], + "score": 0.976, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1770.2432.3423.4841.0005.9459.2471.9831.011
3d0.6941.3342.2143.1251.0007.9247.2922.1620.988
5d0.8881.5201.8002.7061.0007.4835.7581.1880.997
7d0.6530.5021.8771.3131.0004.6833.8561.5910.987
", + "type": "table", + "image_path": "73ec7d6b0b7a0c1eb7fab7d9ec2322bcece0b1e2ae2b0a48342a4450d9623596.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 229, + 499, + 246.66666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 246.66666666666666, + 499, + 264.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 264.3333333333333, + 499, + 282.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.75 + }, + { + "type": "table", + "bbox": [ + 110, + 321, + 501, + 375 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 293, + 503, + 316 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 294, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 505, + 305 + ], + "score": 1.0, + "content": "Table 3: Average cost of paths. The lower the better. The score is normalized based on the solution of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 303, + 308, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 131, + 315 + ], + "score": 0.47, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 303, + 308, + 316 + ], + "score": 1.0, + "content": ". The NEXT-KS achieves the best solutions.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "table_body", + "bbox": [ + 110, + 321, + 501, + 375 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 110, + 321, + 501, + 375 + ], + "spans": [ + { + "bbox": [ + 110, + 321, + 501, + 375 + ], + "score": 0.975, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1720.1931.0491.3331.0001.1402.8111.6491.050
3d0.1160.3150.6120.8751.0000.9551.7201.7340.984
5d0.2150.4260.6730.8901.0000.9231.7800.9611.020
7d0.1080.1470.5730.9871.0000.2911.1141.1390.986
", + "type": "table", + "image_path": "87367e14b8aa88a514f45838acfb7a2ff1b4118f39278d228592551eef4e197f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 110, + 321, + 501, + 339.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 110, + 339.0, + 501, + 357.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 110, + 357.0, + 501, + 375.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 12.75 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "We demonstrated the performance improvement curves for 2d workspace planning, 3d rigid body", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "navigation in Figure 17. As we can see, similar to the performances on 5d 3-link snake planning task", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 405, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 419 + ], + "score": 1.0, + "content": "in Figure 7, in these tasks, the NEXT-KS and NEXT-GP improve the performances along with more", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 416, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 434, + 431 + ], + "score": 1.0, + "content": "and more experiences collected, justified the self-improvement ability by learning", + "type": "text" + }, + { + "bbox": [ + 434, + 416, + 448, + 428 + ], + "score": 0.88, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 417, + 466, + 431 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 466, + 418, + 478, + 428 + ], + "score": 0.85, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 417, + 482, + 431 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 442, + 257, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 259, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 259, + 455 + ], + "score": 1.0, + "content": "F.3 SEARCH TREES COMPARISON", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 286, + 476 + ], + "score": 1.0, + "content": "We illustrate the search trees generated by", + "type": "text" + }, + { + "bbox": [ + 286, + 464, + 312, + 474 + ], + "score": 0.41, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 463, + 506, + 476 + ], + "score": 1.0, + "content": "and the proposed NEXT algorithms with 500", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "samples in Figure 18, Figure 19, Figure 20 and Figure 21 on several 2d, 3d, 5d and 7d planning tasks,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "respectively. To help readers better understand how the trees were expanded, we actually visualize", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "the RRT* search trees without edge rewiring, which is equivalent to the RRT search trees, however", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 394, + 520 + ], + "score": 1.0, + "content": "the vertex set is the same. Comparing to the search trees generated by", + "type": "text" + }, + { + "bbox": [ + 394, + 507, + 419, + 518 + ], + "score": 0.47, + "content": "\\mathbf { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "side by side, we can", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "clearly see the advantages and the efficiency of the proposed NEXT algorithms. In all the tasks, even", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 246, + 542 + ], + "score": 1.0, + "content": "in 2d workspace planning task, the", + "type": "text" + }, + { + "bbox": [ + 246, + 529, + 271, + 540 + ], + "score": 0.37, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "indeed randomly searches without realizing the goals, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "thus cannot complete the missions, while the NEXT algorithms search towards the goals with the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 551, + 435, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 167, + 565 + ], + "score": 1.0, + "content": "guidance from", + "type": "text" + }, + { + "bbox": [ + 167, + 551, + 181, + 563 + ], + "score": 0.88, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 552, + 198, + 565 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 199, + 552, + 210, + 563 + ], + "score": 0.85, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 552, + 435, + 565 + ], + "score": 1.0, + "content": ", therefore, successfully provides high-quality solutions.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "table", + "bbox": [ + 110, + 135, + 499, + 188 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 142, + 115, + 468, + 127 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 141, + 113, + 468, + 129 + ], + "spans": [ + { + "bbox": [ + 141, + 113, + 468, + 129 + ], + "score": 1.0, + "content": "Table 1: Success rate results. The higher the better. NEXT-KS performs the best.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 110, + 135, + 499, + 188 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 135, + 499, + 188 + ], + "spans": [ + { + "bbox": [ + 110, + 135, + 499, + 188 + ], + "score": 0.978, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.9880.9810.7180.6320.7350.7100.1850.5350.720
3d0.9430.8410.6890.5540.4900.5140.1210.1140.498
5d0.8830.7680.6330.5150.4550.4970.0300.4760.444
7d0.9310.9060.6340.3690.3610.8140.2880.2720.370
", + "type": "table", + "image_path": "8ab1e479cbe5f5dc1a3e294cb9eca675bb59491b24d146a5477f9e869df469a1.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 110, + 135, + 499, + 152.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 110, + 152.66666666666666, + 499, + 170.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 170.33333333333331, + 499, + 187.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 110, + 229, + 499, + 282 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 199, + 504, + 221 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 199, + 504, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 504, + 210 + ], + "score": 1.0, + "content": "Table 2: Average number of collision checks results. The lower the better. The score is normalized", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 209, + 420, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 204, + 222 + ], + "score": 1.0, + "content": "based on the solution of", + "type": "text" + }, + { + "bbox": [ + 204, + 210, + 228, + 220 + ], + "score": 0.31, + "content": "\\mathbf { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 209, + 420, + 222 + ], + "score": 1.0, + "content": ". NEXT-KS performs the best in 3 benchmarks.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "table_body", + "bbox": [ + 110, + 229, + 499, + 282 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 229, + 499, + 282 + ], + "spans": [ + { + "bbox": [ + 110, + 229, + 499, + 282 + ], + "score": 0.976, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1770.2432.3423.4841.0005.9459.2471.9831.011
3d0.6941.3342.2143.1251.0007.9247.2922.1620.988
5d0.8881.5201.8002.7061.0007.4835.7581.1880.997
7d0.6530.5021.8771.3131.0004.6833.8561.5910.987
", + "type": "table", + "image_path": "73ec7d6b0b7a0c1eb7fab7d9ec2322bcece0b1e2ae2b0a48342a4450d9623596.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 110, + 229, + 499, + 246.66666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 246.66666666666666, + 499, + 264.3333333333333 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 264.3333333333333, + 499, + 282.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.75 + }, + { + "type": "table", + "bbox": [ + 110, + 321, + 501, + 375 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 293, + 503, + 316 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 294, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 505, + 305 + ], + "score": 1.0, + "content": "Table 3: Average cost of paths. The lower the better. The score is normalized based on the solution of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 303, + 308, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 131, + 315 + ], + "score": 0.47, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 303, + 308, + 316 + ], + "score": 1.0, + "content": ". The NEXT-KS achieves the best solutions.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "table_body", + "bbox": [ + 110, + 321, + 501, + 375 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 110, + 321, + 501, + 375 + ], + "spans": [ + { + "bbox": [ + 110, + 321, + 501, + 375 + ], + "score": 0.975, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1720.1931.0491.3331.0001.1402.8111.6491.050
3d0.1160.3150.6120.8751.0000.9551.7201.7340.984
5d0.2150.4260.6730.8901.0000.9231.7800.9611.020
7d0.1080.1470.5730.9871.0000.2911.1141.1390.986
", + "type": "table", + "image_path": "87367e14b8aa88a514f45838acfb7a2ff1b4118f39278d228592551eef4e197f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 110, + 321, + 501, + 339.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 110, + 339.0, + 501, + 357.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 110, + 357.0, + 501, + 375.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 12.75 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "We demonstrated the performance improvement curves for 2d workspace planning, 3d rigid body", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "navigation in Figure 17. As we can see, similar to the performances on 5d 3-link snake planning task", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 405, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 419 + ], + "score": 1.0, + "content": "in Figure 7, in these tasks, the NEXT-KS and NEXT-GP improve the performances along with more", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 416, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 434, + 431 + ], + "score": 1.0, + "content": "and more experiences collected, justified the self-improvement ability by learning", + "type": "text" + }, + { + "bbox": [ + 434, + 416, + 448, + 428 + ], + "score": 0.88, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 417, + 466, + 431 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 466, + 418, + 478, + 428 + ], + "score": 0.85, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 417, + 482, + 431 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 384, + 505, + 431 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 442, + 257, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 259, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 259, + 455 + ], + "score": 1.0, + "content": "F.3 SEARCH TREES COMPARISON", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 565 + ], + "lines": [ + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 286, + 476 + ], + "score": 1.0, + "content": "We illustrate the search trees generated by", + "type": "text" + }, + { + "bbox": [ + 286, + 464, + 312, + 474 + ], + "score": 0.41, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 463, + 506, + 476 + ], + "score": 1.0, + "content": "and the proposed NEXT algorithms with 500", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "samples in Figure 18, Figure 19, Figure 20 and Figure 21 on several 2d, 3d, 5d and 7d planning tasks,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "respectively. To help readers better understand how the trees were expanded, we actually visualize", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "the RRT* search trees without edge rewiring, which is equivalent to the RRT search trees, however", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 394, + 520 + ], + "score": 1.0, + "content": "the vertex set is the same. Comparing to the search trees generated by", + "type": "text" + }, + { + "bbox": [ + 394, + 507, + 419, + 518 + ], + "score": 0.47, + "content": "\\mathbf { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "side by side, we can", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "clearly see the advantages and the efficiency of the proposed NEXT algorithms. In all the tasks, even", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 246, + 542 + ], + "score": 1.0, + "content": "in 2d workspace planning task, the", + "type": "text" + }, + { + "bbox": [ + 246, + 529, + 271, + 540 + ], + "score": 0.37, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "indeed randomly searches without realizing the goals, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "thus cannot complete the missions, while the NEXT algorithms search towards the goals with the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 551, + 435, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 167, + 565 + ], + "score": 1.0, + "content": "guidance from", + "type": "text" + }, + { + "bbox": [ + 167, + 551, + 181, + 563 + ], + "score": 0.88, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 552, + 198, + 565 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 199, + 552, + 210, + 563 + ], + "score": 0.85, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 552, + 435, + 565 + ], + "score": 1.0, + "content": ", therefore, successfully provides high-quality solutions.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 463, + 506, + 565 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 107, + 490, + 299 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 107, + 490, + 299 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 107, + 490, + 299 + ], + "spans": [ + { + "bbox": [ + 109, + 107, + 490, + 299 + ], + "score": 0.967, + "type": "image", + "image_path": "4c7753cd5f57fdbee146ea4f7042a6906667f2b11da11f1627ca15e30997b485.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 107, + 490, + 171.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 171.0, + 490, + 235.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 235.0, + 490, + 299.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 309, + 507, + 354 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "Figure 17: The first and second rows display the improvement curves of our algorithms on all 3000", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "problems of the 2d workspace planning and 3d rigid body navigation problems. We compare our", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 172, + 344 + ], + "score": 1.0, + "content": "algorithms with", + "type": "text" + }, + { + "bbox": [ + 172, + 332, + 198, + 342 + ], + "score": 0.28, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 331, + 506, + 344 + ], + "score": 1.0, + "content": ". Three columns correspond to the success rate, the average collision checks,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 343, + 352, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 352, + 355 + ], + "score": 1.0, + "content": "and the average cost of the solution paths for each algorithm.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 117, + 410, + 509, + 599 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 410, + 509, + 599 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 117, + 410, + 509, + 599 + ], + "spans": [ + { + "bbox": [ + 117, + 410, + 509, + 599 + ], + "score": 0.951, + "type": "image", + "image_path": "e12728f87cbc3a880f7afe9bb7ba0f37c6e5a599ae9915a27fe8f8e4dd07bd7e.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 117, + 410, + 509, + 473.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 117, + 473.0, + 509, + 536.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 117, + 536.0, + 509, + 599.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 111, + 605, + 511, + 618 + ], + "lines": [ + { + "bbox": [ + 111, + 604, + 512, + 618 + ], + "spans": [ + { + "bbox": [ + 111, + 604, + 126, + 618 + ], + "score": 1.0, + "content": "(a)", + "type": "text" + }, + { + "bbox": [ + 126, + 606, + 151, + 617 + ], + "score": 0.39, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 604, + 468, + 618 + ], + "score": 1.0, + "content": "(w/o rewiring) (b) NEXT-KS search tree (c) NEXT-GP search tree (d) learned", + "type": "text" + }, + { + "bbox": [ + 468, + 604, + 482, + 616 + ], + "score": 0.88, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 604, + 500, + 618 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 500, + 606, + 512, + 616 + ], + "score": 0.82, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 625, + 506, + 707 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 369, + 639 + ], + "score": 1.0, + "content": "Figure 18: Column (a) to (c) are the search trees produced by the", + "type": "text" + }, + { + "bbox": [ + 369, + 626, + 394, + 636 + ], + "score": 0.58, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 626, + 506, + 639 + ], + "score": 1.0, + "content": ", NEXT-KS, and NEXT-GP", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 335, + 651 + ], + "score": 1.0, + "content": "on the same workspace planning task (2d). The learned", + "type": "text" + }, + { + "bbox": [ + 335, + 637, + 349, + 649 + ], + "score": 0.88, + "content": "\\tilde { V } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 638, + 367, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 368, + 639, + 379, + 649 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "from NEXT-KS are plotted in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "column (d). In the figures, obstacles are colored in deep blue, the starting and goal locations are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "denoted by orange and brown dots, respectively. In column (a) to (c), samples are represented with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "hollow yellow circles, and edges are colored in green. In column (d), the level of redness denotes the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 682, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 236, + 696 + ], + "score": 1.0, + "content": "value of the cost-to-go estimate", + "type": "text" + }, + { + "bbox": [ + 236, + 682, + 249, + 693 + ], + "score": 0.87, + "content": "\\tilde { V } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 683, + 434, + 696 + ], + "score": 1.0, + "content": ", and the cyan arrows point from a given state", + "type": "text" + }, + { + "bbox": [ + 435, + 686, + 441, + 693 + ], + "score": 0.69, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 683, + 506, + 696 + ], + "score": 1.0, + "content": "to the center of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 694, + 464, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 694, + 206, + 707 + ], + "score": 1.0, + "content": "the proposal distribution", + "type": "text" + }, + { + "bbox": [ + 207, + 694, + 253, + 707 + ], + "score": 0.93, + "content": "\\tilde { \\pi } ^ { * } \\bigl ( s ^ { \\prime } | s , U \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 694, + 464, + 707 + ], + "score": 1.0, + "content": ". We set the maximum number of samples to be 500.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "22", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 107, + 490, + 299 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 107, + 490, + 299 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 107, + 490, + 299 + ], + "spans": [ + { + "bbox": [ + 109, + 107, + 490, + 299 + ], + "score": 0.967, + "type": "image", + "image_path": "4c7753cd5f57fdbee146ea4f7042a6906667f2b11da11f1627ca15e30997b485.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 107, + 490, + 171.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 171.0, + 490, + 235.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 235.0, + 490, + 299.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 309, + 507, + 354 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "Figure 17: The first and second rows display the improvement curves of our algorithms on all 3000", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "problems of the 2d workspace planning and 3d rigid body navigation problems. We compare our", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 172, + 344 + ], + "score": 1.0, + "content": "algorithms with", + "type": "text" + }, + { + "bbox": [ + 172, + 332, + 198, + 342 + ], + "score": 0.28, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 331, + 506, + 344 + ], + "score": 1.0, + "content": ". Three columns correspond to the success rate, the average collision checks,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 343, + 352, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 352, + 355 + ], + "score": 1.0, + "content": "and the average cost of the solution paths for each algorithm.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 117, + 410, + 509, + 599 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 410, + 509, + 599 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 117, + 410, + 509, + 599 + ], + "spans": [ + { + "bbox": [ + 117, + 410, + 509, + 599 + ], + "score": 0.951, + "type": "image", + "image_path": "e12728f87cbc3a880f7afe9bb7ba0f37c6e5a599ae9915a27fe8f8e4dd07bd7e.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 117, + 410, + 509, + 473.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 117, + 473.0, + 509, + 536.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 117, + 536.0, + 509, + 599.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 111, + 605, + 511, + 618 + ], + "lines": [ + { + "bbox": [ + 111, + 604, + 512, + 618 + ], + "spans": [ + { + "bbox": [ + 111, + 604, + 126, + 618 + ], + "score": 1.0, + "content": "(a)", + "type": "text" + }, + { + "bbox": [ + 126, + 606, + 151, + 617 + ], + "score": 0.39, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 604, + 468, + 618 + ], + "score": 1.0, + "content": "(w/o rewiring) (b) NEXT-KS search tree (c) NEXT-GP search tree (d) learned", + "type": "text" + }, + { + "bbox": [ + 468, + 604, + 482, + 616 + ], + "score": 0.88, + "content": "{ \\tilde { V } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 604, + 500, + 618 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 500, + 606, + 512, + 616 + ], + "score": 0.82, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 111, + 604, + 512, + 618 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 625, + 506, + 707 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 369, + 639 + ], + "score": 1.0, + "content": "Figure 18: Column (a) to (c) are the search trees produced by the", + "type": "text" + }, + { + "bbox": [ + 369, + 626, + 394, + 636 + ], + "score": 0.58, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 626, + 506, + 639 + ], + "score": 1.0, + "content": ", NEXT-KS, and NEXT-GP", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 335, + 651 + ], + "score": 1.0, + "content": "on the same workspace planning task (2d). The learned", + "type": "text" + }, + { + "bbox": [ + 335, + 637, + 349, + 649 + ], + "score": 0.88, + "content": "\\tilde { V } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 638, + 367, + 651 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 368, + 639, + 379, + 649 + ], + "score": 0.86, + "content": "\\tilde { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "from NEXT-KS are plotted in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "column (d). In the figures, obstacles are colored in deep blue, the starting and goal locations are", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "denoted by orange and brown dots, respectively. In column (a) to (c), samples are represented with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "hollow yellow circles, and edges are colored in green. In column (d), the level of redness denotes the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 682, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 236, + 696 + ], + "score": 1.0, + "content": "value of the cost-to-go estimate", + "type": "text" + }, + { + "bbox": [ + 236, + 682, + 249, + 693 + ], + "score": 0.87, + "content": "\\tilde { V } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 683, + 434, + 696 + ], + "score": 1.0, + "content": ", and the cyan arrows point from a given state", + "type": "text" + }, + { + "bbox": [ + 435, + 686, + 441, + 693 + ], + "score": 0.69, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 683, + 506, + 696 + ], + "score": 1.0, + "content": "to the center of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 694, + 464, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 694, + 206, + 707 + ], + "score": 1.0, + "content": "the proposal distribution", + "type": "text" + }, + { + "bbox": [ + 207, + 694, + 253, + 707 + ], + "score": 0.93, + "content": "\\tilde { \\pi } ^ { * } \\bigl ( s ^ { \\prime } | s , U \\bigr )", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 694, + 464, + 707 + ], + "score": 1.0, + "content": ". We set the maximum number of samples to be 500.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 626, + 506, + 707 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 102, + 499, + 318 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 102, + 499, + 318 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 102, + 499, + 318 + ], + "spans": [ + { + "bbox": [ + 116, + 102, + 499, + 318 + ], + "score": 0.974, + "type": "image", + "image_path": "e0735fc244ea4c94bcec57f2daf0ecaa89ac9d792df1ae99e175f050da010796.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 102, + 499, + 174.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 174.0, + 499, + 246.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 246.0, + 499, + 318.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 327, + 507, + 383 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 327, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 506, + 340 + ], + "score": 1.0, + "content": "Figure 19: Each column corresponds to one example from the rigid body navigation problem (3d).", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 336, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 352 + ], + "score": 1.0, + "content": "The top and the bottom rows are the search trees produced by the RRT* and NEXT-KS, respectively.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 348, + 507, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 507, + 362 + ], + "score": 1.0, + "content": "In the figures, obstacles are colored in deep blue, and the rigid bodies are represented with matchsticks.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 360, + 507, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 507, + 373 + ], + "score": 1.0, + "content": "The samples, starting states, and goal states are denoted by yellow, orange, and brown matchsticks,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 371, + 478, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 478, + 384 + ], + "score": 1.0, + "content": "respectively. Edges are colored in green. We set the maximum number of samples to be 500.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "image", + "bbox": [ + 115, + 430, + 498, + 646 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 430, + 498, + 646 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 430, + 498, + 646 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 498, + 646 + ], + "score": 0.977, + "type": "image", + "image_path": "cfa7ea4cfcbf4aafa0403606ed187484f2a7a371df073fcdcf9d9b4c8e59fe6c.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 115, + 430, + 498, + 502.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 115, + 502.0, + 498, + 574.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 115, + 574.0, + 498, + 646.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 655, + 507, + 711 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "score": 1.0, + "content": "Figure 20: Each column corresponds to one example from the 3-link snake problem (5d). The top", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "and the bottom rows are the search trees produced by the RRT* and NEXT-KS, respectively. In the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "figures, obstacles are colored in deep blue, and the rigid bodies are represented with matchsticks.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "The samples, starting states, and goal states are denoted by yellow, orange, and brown matchsticks,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 699, + 478, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 478, + 712 + ], + "score": 1.0, + "content": "respectively. Edges are colored in green. We set the maximum number of samples to be 500.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + } + ], + "index": 11.0 + } + ], + "page_idx": 22, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "23", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 102, + 499, + 318 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 102, + 499, + 318 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 102, + 499, + 318 + ], + "spans": [ + { + "bbox": [ + 116, + 102, + 499, + 318 + ], + "score": 0.974, + "type": "image", + "image_path": "e0735fc244ea4c94bcec57f2daf0ecaa89ac9d792df1ae99e175f050da010796.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 102, + 499, + 174.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 174.0, + 499, + 246.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 246.0, + 499, + 318.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 327, + 507, + 383 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 327, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 506, + 340 + ], + "score": 1.0, + "content": "Figure 19: Each column corresponds to one example from the rigid body navigation problem (3d).", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 336, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 352 + ], + "score": 1.0, + "content": "The top and the bottom rows are the search trees produced by the RRT* and NEXT-KS, respectively.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 348, + 507, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 507, + 362 + ], + "score": 1.0, + "content": "In the figures, obstacles are colored in deep blue, and the rigid bodies are represented with matchsticks.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 360, + 507, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 507, + 373 + ], + "score": 1.0, + "content": "The samples, starting states, and goal states are denoted by yellow, orange, and brown matchsticks,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 371, + 478, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 478, + 384 + ], + "score": 1.0, + "content": "respectively. Edges are colored in green. We set the maximum number of samples to be 500.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "image", + "bbox": [ + 115, + 430, + 498, + 646 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 430, + 498, + 646 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 430, + 498, + 646 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 498, + 646 + ], + "score": 0.977, + "type": "image", + "image_path": "cfa7ea4cfcbf4aafa0403606ed187484f2a7a371df073fcdcf9d9b4c8e59fe6c.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 115, + 430, + 498, + 502.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 115, + 502.0, + 498, + 574.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 115, + 574.0, + 498, + 646.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 655, + 507, + 711 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 669 + ], + "score": 1.0, + "content": "Figure 20: Each column corresponds to one example from the 3-link snake problem (5d). The top", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "and the bottom rows are the search trees produced by the RRT* and NEXT-KS, respectively. In the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "figures, obstacles are colored in deep blue, and the rigid bodies are represented with matchsticks.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "The samples, starting states, and goal states are denoted by yellow, orange, and brown matchsticks,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 699, + 478, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 478, + 712 + ], + "score": 1.0, + "content": "respectively. Edges are colored in green. We set the maximum number of samples to be 500.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + } + ], + "index": 11.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 275, + 491, + 485 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 275, + 491, + 485 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 275, + 491, + 484 + ], + "spans": [ + { + "bbox": [ + 124, + 275, + 491, + 484 + ], + "score": 0.951, + "type": "image", + "image_path": "acacf1ccb84450bee6725175dc3f6794eeb8b2a6eca9fd60b1ef87035d99e36d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 275, + 491, + 345.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 345.0, + 491, + 415.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 415.0, + 491, + 485.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 493, + 506, + 538 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 493, + 507, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 507, + 506 + ], + "score": 1.0, + "content": "Figure 21: Each column corresponds to one example from the spacecraft planning problem (7d).", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 504, + 507, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 366, + 518 + ], + "score": 1.0, + "content": "The top and the bottom rows are the search trees produced by the", + "type": "text" + }, + { + "bbox": [ + 366, + 505, + 391, + 515 + ], + "score": 0.33, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 504, + 507, + 518 + ], + "score": 1.0, + "content": "and NEXT-KS, respectively.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 515, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 528 + ], + "score": 1.0, + "content": "In the figures, obstacles are colored in blue, and each spacecraft has a yellow body and two 2 DoF", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 527, + 353, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 353, + 539 + ], + "score": 1.0, + "content": "red arms. We set the maximum number of samples to be 500.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + } + ], + "page_idx": 23, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 275, + 491, + 485 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 275, + 491, + 485 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 275, + 491, + 484 + ], + "spans": [ + { + "bbox": [ + 124, + 275, + 491, + 484 + ], + "score": 0.951, + "type": "image", + "image_path": "acacf1ccb84450bee6725175dc3f6794eeb8b2a6eca9fd60b1ef87035d99e36d.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 275, + 491, + 345.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 345.0, + 491, + 415.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 415.0, + 491, + 485.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 493, + 506, + 538 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 493, + 507, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 507, + 506 + ], + "score": 1.0, + "content": "Figure 21: Each column corresponds to one example from the spacecraft planning problem (7d).", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 504, + 507, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 366, + 518 + ], + "score": 1.0, + "content": "The top and the bottom rows are the search trees produced by the", + "type": "text" + }, + { + "bbox": [ + 366, + 505, + 391, + 515 + ], + "score": 0.33, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 504, + 507, + 518 + ], + "score": 1.0, + "content": "and NEXT-KS, respectively.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 515, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 528 + ], + "score": 1.0, + "content": "In the figures, obstacles are colored in blue, and each spacecraft has a yellow body and two 2 DoF", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 527, + 353, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 353, + 539 + ], + "score": 1.0, + "content": "red arms. We set the maximum number of samples to be 500.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 227, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 229, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 229, + 96 + ], + "score": 1.0, + "content": "F.4 CASE STUDY DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "score": 1.0, + "content": "We conduct a real-world case study on controlling robot arms to move objects on a shelf. This is a", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "representative of common scenarios in practice where the robot needs to plan its motion in real-time", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "score": 1.0, + "content": "to reach the inside of some narrow space. For this case study, we focus more on the practical aspect", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 134, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 134, + 505, + 150 + ], + "score": 1.0, + "content": "to evaluate how much can we improve on the wall-clock time by learning from similar planning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 149, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 149, + 159 + ], + "score": 1.0, + "content": "problems.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 107, + 170, + 224, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 225, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 225, + 182 + ], + "score": 1.0, + "content": "F.4.1 TASK DESCRIPTION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 190, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 189, + 506, + 202 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 506, + 202 + ], + "score": 1.0, + "content": "We generate planning problems randomly to form the training set and test set. In each planning task,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "there is a shelf of multiple levels, with each level horizontally divided into multiple bins. The heights", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 211, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 506, + 225 + ], + "score": 1.0, + "content": "of levels and widths of bins are randomly drawn from some fixed distribution. Samples of shelves are", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "shown in Figure 22. Both the start and goal configurations are randomly sampled from a distribution", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "within the reachable region of the robot arm. The planning environment is created with the OpenRave", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 216, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 216, + 256 + ], + "score": 1.0, + "content": "simulator (Diankov, 2010).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "The task is to find a path for the 7 DoF robot arm to move from a location in one bin to another, i.e., the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 273, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 284 + ], + "score": 1.0, + "content": "end effectors of the start and goal configurations are in different bins, as illustrated in Figure 23. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 283, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 506, + 296 + ], + "score": 1.0, + "content": "this case, the base of the robot is fixed and we are planning the movement of arm. We generated 3000", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 319, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 319, + 308 + ], + "score": 1.0, + "content": "problems for training and 1000 problems for testing.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 318, + 255, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 257, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 257, + 330 + ], + "score": 1.0, + "content": "F.4.2 BASELINES AND TRAINING", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 249, + 349 + ], + "score": 1.0, + "content": "For traditional planners, we include", + "type": "text" + }, + { + "bbox": [ + 250, + 338, + 269, + 348 + ], + "score": 0.78, + "content": "\\mathrm { C } + +", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 337, + 455, + 349 + ], + "score": 1.0, + "content": "OMPL ( ¸Sucan et al., 2012) implementation of", + "type": "text" + }, + { + "bbox": [ + 455, + 337, + 477, + 348 + ], + "score": 0.38, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "(Gam-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 199, + 361 + ], + "score": 1.0, + "content": "mell et al., 2015) and", + "type": "text" + }, + { + "bbox": [ + 199, + 348, + 225, + 359 + ], + "score": 0.36, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "(Karaman & Frazzoli, 2011) as baselines. The hyperparameters of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 152, + 371 + ], + "score": 1.0, + "content": "RRT* and", + "type": "text" + }, + { + "bbox": [ + 153, + 360, + 175, + 370 + ], + "score": 0.28, + "content": "\\mathbf { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "are specially tuned for this experiment. We also compare with learning-based", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "planners CVAE-plan (Ichter et al., 2018) and Reinforce-plan (Zhang et al., 2018). The supervisions", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 293, + 393 + ], + "score": 1.0, + "content": "for CVAE-plan are produced by the well-tuned", + "type": "text" + }, + { + "bbox": [ + 293, + 381, + 315, + 392 + ], + "score": 0.36, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 381, + 506, + 393 + ], + "score": 1.0, + "content": "on the training set. To train NEXT, we consider", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 392, + 391, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 121, + 404 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 392, + 144, + 402 + ], + "score": 0.32, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 392, + 391, + 404 + ], + "score": 1.0, + "content": "instead of RRT* as the imperfect expert in training problems.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 107, + 415, + 179, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 180, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 180, + 428 + ], + "score": 1.0, + "content": "F.4.3 RESULTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 435, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "We evaluate the algorithms on the separated testing problems, and record the success rate using 10", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 444, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 459 + ], + "score": 1.0, + "content": "different time limits. The success rate and average path quality are plot in Figure 8 and recorded in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 356, + 469 + ], + "score": 1.0, + "content": "Table 4 and Table 5. The solution paths found by NEXT and", + "type": "text" + }, + { + "bbox": [ + 356, + 457, + 379, + 468 + ], + "score": 0.52, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 455, + 505, + 469 + ], + "score": 1.0, + "content": "are illustrated in Figure 23. In", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "terms of both success rate and solution path quality, NEXT dominates all the planners under all time", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 477, + 135, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 135, + 491 + ], + "score": 1.0, + "content": "limits.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 120, + 494, + 493, + 629 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 494, + 493, + 629 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 494, + 493, + 629 + ], + "spans": [ + { + "bbox": [ + 120, + 494, + 493, + 629 + ], + "score": 0.975, + "type": "image", + "image_path": "1a68418f9fc5b8db4c0728ef4e4b915a1ac7b12ca2d6815c4396b38f554580ed.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 120, + 494, + 493, + 539.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 120, + 539.0, + 493, + 584.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 120, + 584.0, + 493, + 629.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 161, + 645, + 449, + 657 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 160, + 644, + 450, + 658 + ], + "spans": [ + { + "bbox": [ + 160, + 644, + 450, + 658 + ], + "score": 1.0, + "content": "Figure 22: Examples of different shelves sampled from the distribution.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "index": 32.0 + } + ], + "page_idx": 24, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 227, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 229, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 229, + 96 + ], + "score": 1.0, + "content": "F.4 CASE STUDY DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "score": 1.0, + "content": "We conduct a real-world case study on controlling robot arms to move objects on a shelf. This is a", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "representative of common scenarios in practice where the robot needs to plan its motion in real-time", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "score": 1.0, + "content": "to reach the inside of some narrow space. For this case study, we focus more on the practical aspect", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 134, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 104, + 134, + 505, + 150 + ], + "score": 1.0, + "content": "to evaluate how much can we improve on the wall-clock time by learning from similar planning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 149, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 149, + 159 + ], + "score": 1.0, + "content": "problems.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 103, + 506, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 170, + 224, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 225, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 225, + 182 + ], + "score": 1.0, + "content": "F.4.1 TASK DESCRIPTION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 190, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 189, + 506, + 202 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 506, + 202 + ], + "score": 1.0, + "content": "We generate planning problems randomly to form the training set and test set. In each planning task,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "there is a shelf of multiple levels, with each level horizontally divided into multiple bins. The heights", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 211, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 506, + 225 + ], + "score": 1.0, + "content": "of levels and widths of bins are randomly drawn from some fixed distribution. Samples of shelves are", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "shown in Figure 22. Both the start and goal configurations are randomly sampled from a distribution", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "within the reachable region of the robot arm. The planning environment is created with the OpenRave", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 216, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 216, + 256 + ], + "score": 1.0, + "content": "simulator (Diankov, 2010).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 189, + 506, + 256 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "The task is to find a path for the 7 DoF robot arm to move from a location in one bin to another, i.e., the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 273, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 505, + 284 + ], + "score": 1.0, + "content": "end effectors of the start and goal configurations are in different bins, as illustrated in Figure 23. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 283, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 506, + 296 + ], + "score": 1.0, + "content": "this case, the base of the robot is fixed and we are planning the movement of arm. We generated 3000", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 293, + 319, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 319, + 308 + ], + "score": 1.0, + "content": "problems for training and 1000 problems for testing.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 261, + 506, + 308 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 318, + 255, + 329 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 257, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 257, + 330 + ], + "score": 1.0, + "content": "F.4.2 BASELINES AND TRAINING", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 505, + 404 + ], + "lines": [ + { + "bbox": [ + 106, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 249, + 349 + ], + "score": 1.0, + "content": "For traditional planners, we include", + "type": "text" + }, + { + "bbox": [ + 250, + 338, + 269, + 348 + ], + "score": 0.78, + "content": "\\mathrm { C } + +", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 337, + 455, + 349 + ], + "score": 1.0, + "content": "OMPL ( ¸Sucan et al., 2012) implementation of", + "type": "text" + }, + { + "bbox": [ + 455, + 337, + 477, + 348 + ], + "score": 0.38, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "(Gam-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 199, + 361 + ], + "score": 1.0, + "content": "mell et al., 2015) and", + "type": "text" + }, + { + "bbox": [ + 199, + 348, + 225, + 359 + ], + "score": 0.36, + "content": "\\mathrm { R R T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "(Karaman & Frazzoli, 2011) as baselines. The hyperparameters of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 152, + 371 + ], + "score": 1.0, + "content": "RRT* and", + "type": "text" + }, + { + "bbox": [ + 153, + 360, + 175, + 370 + ], + "score": 0.28, + "content": "\\mathbf { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "are specially tuned for this experiment. We also compare with learning-based", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "planners CVAE-plan (Ichter et al., 2018) and Reinforce-plan (Zhang et al., 2018). The supervisions", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 293, + 393 + ], + "score": 1.0, + "content": "for CVAE-plan are produced by the well-tuned", + "type": "text" + }, + { + "bbox": [ + 293, + 381, + 315, + 392 + ], + "score": 0.36, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 381, + 506, + 393 + ], + "score": 1.0, + "content": "on the training set. To train NEXT, we consider", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 392, + 391, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 121, + 404 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 392, + 144, + 402 + ], + "score": 0.32, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 392, + 391, + 404 + ], + "score": 1.0, + "content": "instead of RRT* as the imperfect expert in training problems.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 337, + 506, + 404 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 415, + 179, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 180, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 180, + 428 + ], + "score": 1.0, + "content": "F.4.3 RESULTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 435, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "We evaluate the algorithms on the separated testing problems, and record the success rate using 10", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 444, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 459 + ], + "score": 1.0, + "content": "different time limits. The success rate and average path quality are plot in Figure 8 and recorded in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 455, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 356, + 469 + ], + "score": 1.0, + "content": "Table 4 and Table 5. The solution paths found by NEXT and", + "type": "text" + }, + { + "bbox": [ + 356, + 457, + 379, + 468 + ], + "score": 0.52, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 455, + 505, + 469 + ], + "score": 1.0, + "content": "are illustrated in Figure 23. In", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "terms of both success rate and solution path quality, NEXT dominates all the planners under all time", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 477, + 135, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 135, + 491 + ], + "score": 1.0, + "content": "limits.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 435, + 506, + 491 + ] + }, + { + "type": "image", + "bbox": [ + 120, + 494, + 493, + 629 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 494, + 493, + 629 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 494, + 493, + 629 + ], + "spans": [ + { + "bbox": [ + 120, + 494, + 493, + 629 + ], + "score": 0.975, + "type": "image", + "image_path": "1a68418f9fc5b8db4c0728ef4e4b915a1ac7b12ca2d6815c4396b38f554580ed.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 120, + 494, + 493, + 539.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 120, + 539.0, + 493, + 584.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 120, + 584.0, + 493, + 629.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 161, + 645, + 449, + 657 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 160, + 644, + 450, + 658 + ], + "spans": [ + { + "bbox": [ + 160, + 644, + 450, + 658 + ], + "score": 1.0, + "content": "Figure 22: Examples of different shelves sampled from the distribution.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "index": 32.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 103, + 486, + 366 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 103, + 486, + 366 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 486, + 366 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 486, + 366 + ], + "score": 0.675, + "type": "image", + "image_path": "4918f228ef3bb16fc560cb6a309ed28b7f1abfcfd2648d63098ec7552a0228e2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 103, + 486, + 190.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 190.66666666666669, + 486, + 278.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 278.33333333333337, + 486, + 366.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 386, + 506, + 430 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 399 + ], + "score": 1.0, + "content": "Figure 23: First row: robot arm solution trajectories produced by NEXT(-KS) in four planning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 200, + 410 + ], + "score": 1.0, + "content": "problems; Second row:", + "type": "text" + }, + { + "bbox": [ + 201, + 397, + 223, + 408 + ], + "score": 0.35, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "solutions on the same planning problems. NEXT only takes 5 seconds", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 408, + 236, + 421 + ], + "score": 1.0, + "content": "to complete each problem while", + "type": "text" + }, + { + "bbox": [ + 236, + 408, + 259, + 419 + ], + "score": 0.6, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "needs 250 seconds to find a solution for the hardest problems", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 420, + 184, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 184, + 431 + ], + "score": 1.0, + "content": "(last two columns).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 120, + 500, + 491, + 564 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 122, + 480, + 487, + 493 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 479, + 488, + 494 + ], + "spans": [ + { + "bbox": [ + 122, + 479, + 488, + 494 + ], + "score": 1.0, + "content": "Table 4: Success rates of different planners under varying time limits, the higher the better.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "table_body", + "bbox": [ + 120, + 500, + 491, + 564 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 500, + 491, + 564 + ], + "spans": [ + { + "bbox": [ + 120, + 500, + 491, + 564 + ], + "score": 0.982, + "html": "
5s10s15s20s25s30s35s40s45s50s
NEXT0.5790.6570.7030.7090.7450.7430.7460.7520.7720.763
CVAE0.3540.4370.4820.5090.5070.5390.5530.5510.5790.580
Reinforce0.1600.1700.2000.1500.1800.1900.1750.1800.2250.175
BIT*0.2260.2880.3200.3650.3640.4290.4250.4220.4430.475
RRT*0.1350.1480.1440.1360.1470.1590.1520.1580.1570.165
", + "type": "table", + "image_path": "948be226a6bad3dfcd234d9761507c5f0ee44a0c74d5bfc76f2f1e7f3c70aac5.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 120, + 500, + 491, + 521.3333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 120, + 521.3333333333334, + 491, + 542.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 120, + 542.6666666666667, + 491, + 564.0000000000001 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 8.0 + }, + { + "type": "table", + "bbox": [ + 107, + 639, + 522, + 703 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 111, + 618, + 496, + 631 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 618, + 498, + 632 + ], + "spans": [ + { + "bbox": [ + 114, + 618, + 498, + 632 + ], + "score": 1.0, + "content": "Table 5: Average path costs of different planners under varying time limits, the lower the better.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "table_body", + "bbox": [ + 107, + 639, + 522, + 703 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 639, + 522, + 703 + ], + "spans": [ + { + "bbox": [ + 107, + 639, + 522, + 703 + ], + "score": 0.982, + "html": "
5s10s15s20s25s30s35s40s45s50s
NEXT25.14321.87320.05219.77318.31818.38118.27418.04717.22417.587
CVAE34.41430.81528.90927.71727.78426.40825.77625.88324.66524.659
Reinforce42.78142.47141.14243.29542.12941.65742.23542.09340.26942.282
BIT*39.54337.08635.74434.00234.00731.41431.58731.65130.84129.601
RRT*43.36842.85542.98143.35542.91942.41042.68042.44142.50442.165
", + "type": "table", + "image_path": "17b397e3af9f47e2586529a736802b6c4ddb66c41d7cfd6009886c510b1b4236.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 107, + 639, + 522, + 660.3333333333334 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 107, + 660.3333333333334, + 522, + 681.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 107, + 681.6666666666667, + 522, + 703.0000000000001 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 12.0 + } + ], + "page_idx": 25, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "26", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 103, + 486, + 366 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 103, + 486, + 366 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 486, + 366 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 486, + 366 + ], + "score": 0.675, + "type": "image", + "image_path": "4918f228ef3bb16fc560cb6a309ed28b7f1abfcfd2648d63098ec7552a0228e2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 103, + 486, + 190.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 190.66666666666669, + 486, + 278.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 278.33333333333337, + 486, + 366.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 386, + 506, + 430 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 399 + ], + "score": 1.0, + "content": "Figure 23: First row: robot arm solution trajectories produced by NEXT(-KS) in four planning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 200, + 410 + ], + "score": 1.0, + "content": "problems; Second row:", + "type": "text" + }, + { + "bbox": [ + 201, + 397, + 223, + 408 + ], + "score": 0.35, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "solutions on the same planning problems. NEXT only takes 5 seconds", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 104, + 408, + 236, + 421 + ], + "score": 1.0, + "content": "to complete each problem while", + "type": "text" + }, + { + "bbox": [ + 236, + 408, + 259, + 419 + ], + "score": 0.6, + "content": "\\mathrm { B I T ^ { * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "needs 250 seconds to find a solution for the hardest problems", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 420, + 184, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 184, + 431 + ], + "score": 1.0, + "content": "(last two columns).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 120, + 500, + 491, + 564 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 122, + 480, + 487, + 493 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 479, + 488, + 494 + ], + "spans": [ + { + "bbox": [ + 122, + 479, + 488, + 494 + ], + "score": 1.0, + "content": "Table 4: Success rates of different planners under varying time limits, the higher the better.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "table_body", + "bbox": [ + 120, + 500, + 491, + 564 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 500, + 491, + 564 + ], + "spans": [ + { + "bbox": [ + 120, + 500, + 491, + 564 + ], + "score": 0.982, + "html": "
5s10s15s20s25s30s35s40s45s50s
NEXT0.5790.6570.7030.7090.7450.7430.7460.7520.7720.763
CVAE0.3540.4370.4820.5090.5070.5390.5530.5510.5790.580
Reinforce0.1600.1700.2000.1500.1800.1900.1750.1800.2250.175
BIT*0.2260.2880.3200.3650.3640.4290.4250.4220.4430.475
RRT*0.1350.1480.1440.1360.1470.1590.1520.1580.1570.165
", + "type": "table", + "image_path": "948be226a6bad3dfcd234d9761507c5f0ee44a0c74d5bfc76f2f1e7f3c70aac5.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 120, + 500, + 491, + 521.3333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 120, + 521.3333333333334, + 491, + 542.6666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 120, + 542.6666666666667, + 491, + 564.0000000000001 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 8.0 + }, + { + "type": "table", + "bbox": [ + 107, + 639, + 522, + 703 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 111, + 618, + 496, + 631 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 114, + 618, + 498, + 632 + ], + "spans": [ + { + "bbox": [ + 114, + 618, + 498, + 632 + ], + "score": 1.0, + "content": "Table 5: Average path costs of different planners under varying time limits, the lower the better.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "table_body", + "bbox": [ + 107, + 639, + 522, + 703 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 639, + 522, + 703 + ], + "spans": [ + { + "bbox": [ + 107, + 639, + 522, + 703 + ], + "score": 0.982, + "html": "
5s10s15s20s25s30s35s40s45s50s
NEXT25.14321.87320.05219.77318.31818.38118.27418.04717.22417.587
CVAE34.41430.81528.90927.71727.78426.40825.77625.88324.66524.659
Reinforce42.78142.47141.14243.29542.12941.65742.23542.09340.26942.282
BIT*39.54337.08635.74434.00234.00731.41431.58731.65130.84129.601
RRT*43.36842.85542.98143.35542.91942.41042.68042.44142.50442.165
", + "type": "table", + "image_path": "17b397e3af9f47e2586529a736802b6c4ddb66c41d7cfd6009886c510b1b4236.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 107, + 639, + 522, + 660.3333333333334 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 107, + 660.3333333333334, + 522, + 681.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 107, + 681.6666666666667, + 522, + 703.0000000000001 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 12.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/rJgJDAVKvB/rJgJDAVKvB_model.json b/parse/train/rJgJDAVKvB/rJgJDAVKvB_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7933d00c32740c1c7b5ea8e0ece980a5081e0d02 --- /dev/null +++ b/parse/train/rJgJDAVKvB/rJgJDAVKvB_model.json @@ -0,0 +1,43731 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1037, + 1403, + 1037, + 1403, + 1526, + 299, + 1526 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 362, + 588, + 1341, + 588, + 1341, + 923, + 362, + 923 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1541, + 1404, + 1541, + 1404, + 1968, + 298, + 1968 + ], + "score": 0.982 + }, + { + "category_id": 0, + "poly": [ + 301, + 218, + 1258, + 218, + 1258, + 324, + 301, + 324 + ], + "score": 0.964 + }, + { + "category_id": 2, + "poly": [ + 333, + 2006, + 627, + 2006, + 627, + 2033, + 333, + 2033 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 315, + 372, + 1179, + 372, + 1179, + 470, + 315, + 470 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 302, + 979, + 573, + 979, + 573, + 1013, + 302, + 1013 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 817, + 76, + 817, + 104, + 299, + 104 + ], + "score": 0.881 + }, + { + "category_id": 0, + "poly": [ + 773, + 528, + 926, + 528, + 926, + 561, + 773, + 561 + ], + "score": 0.868 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.737 + }, + { + "category_id": 13, + "poly": [ + 889, + 1130, + 925, + 1130, + 925, + 1158, + 889, + 1158 + ], + "score": 0.86, + "latex": "A ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 743, + 374, + 800, + 374, + 800, + 406, + 743, + 406 + ], + "score": 0.77, + "latex": "\\mathbf { L i n ^ { 3 } }" + }, + { + "category_id": 13, + "poly": [ + 869, + 374, + 913, + 374, + 913, + 406, + 869, + 406 + ], + "score": 0.76, + "latex": "\\mathbf { Y e ^ { 3 } }" + }, + { + "category_id": 13, + "poly": [ + 543, + 374, + 652, + 374, + 652, + 407, + 543, + 407 + ], + "score": 0.47, + "latex": "\\mathbf { \\mathbf { \\mathbf { \\mathbf { \\mathbf { \\mathbf { B } } } } } } \\mathbf { \\mathbf { 0 } } \\mathbf { \\mathbf { \\mathbf { \\mathbf { D } } } } \\mathbf { \\mathbf { a } } \\mathbf { i } ^ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 222.0, + 1265.0, + 222.0, + 1265.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 273.0, + 1256.0, + 273.0, + 1256.0, + 325.0, + 295.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2001.0, + 632.0, + 2001.0, + 632.0, + 2037.0, + 330.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 975.0, + 579.0, + 975.0, + 579.0, + 1021.0, + 294.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 527.0, + 932.0, + 527.0, + 932.0, + 564.0, + 769.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 1404.0, + 1039.0, + 1404.0, + 1073.0, + 294.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1069.0, + 1405.0, + 1069.0, + 1405.0, + 1104.0, + 294.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1100.0, + 1404.0, + 1100.0, + 1404.0, + 1134.0, + 295.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1131.0, + 888.0, + 1131.0, + 888.0, + 1162.0, + 293.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1131.0, + 1405.0, + 1131.0, + 1405.0, + 1162.0, + 926.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1160.0, + 1407.0, + 1160.0, + 1407.0, + 1194.0, + 294.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1221.0, + 1407.0, + 1221.0, + 1407.0, + 1257.0, + 293.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1252.0, + 1405.0, + 1252.0, + 1405.0, + 1286.0, + 294.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1281.0, + 1404.0, + 1281.0, + 1404.0, + 1316.0, + 295.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1312.0, + 1407.0, + 1312.0, + 1407.0, + 1348.0, + 293.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1342.0, + 1405.0, + 1342.0, + 1405.0, + 1378.0, + 293.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1408.0, + 295.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1404.0, + 1405.0, + 1404.0, + 1405.0, + 1439.0, + 295.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 1407.0, + 1432.0, + 1407.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1466.0, + 1408.0, + 1466.0, + 1408.0, + 1500.0, + 295.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1496.0, + 1305.0, + 1496.0, + 1305.0, + 1529.0, + 294.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 587.0, + 1342.0, + 587.0, + 1342.0, + 625.0, + 357.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 617.0, + 1342.0, + 617.0, + 1342.0, + 654.0, + 356.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 649.0, + 1343.0, + 649.0, + 1343.0, + 684.0, + 357.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 681.0, + 1342.0, + 681.0, + 1342.0, + 716.0, + 358.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 712.0, + 1344.0, + 712.0, + 1344.0, + 744.0, + 358.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 742.0, + 1340.0, + 742.0, + 1340.0, + 774.0, + 357.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 771.0, + 1343.0, + 771.0, + 1343.0, + 805.0, + 356.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 802.0, + 1346.0, + 802.0, + 1346.0, + 837.0, + 357.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 830.0, + 1344.0, + 830.0, + 1344.0, + 866.0, + 356.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 863.0, + 1340.0, + 863.0, + 1340.0, + 898.0, + 357.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 894.0, + 1040.0, + 894.0, + 1040.0, + 925.0, + 357.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1542.0, + 1405.0, + 1542.0, + 1405.0, + 1577.0, + 296.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1573.0, + 1406.0, + 1573.0, + 1406.0, + 1608.0, + 293.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1601.0, + 1406.0, + 1601.0, + 1406.0, + 1637.0, + 293.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1634.0, + 1406.0, + 1634.0, + 1406.0, + 1670.0, + 293.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1662.0, + 1409.0, + 1662.0, + 1409.0, + 1700.0, + 292.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1695.0, + 1406.0, + 1695.0, + 1406.0, + 1730.0, + 294.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1723.0, + 1406.0, + 1723.0, + 1406.0, + 1761.0, + 292.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1756.0, + 1406.0, + 1756.0, + 1406.0, + 1789.0, + 293.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1787.0, + 1402.0, + 1787.0, + 1402.0, + 1819.0, + 296.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1814.0, + 1406.0, + 1814.0, + 1406.0, + 1853.0, + 292.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1847.0, + 1405.0, + 1847.0, + 1405.0, + 1882.0, + 294.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1876.0, + 1406.0, + 1876.0, + 1406.0, + 1912.0, + 294.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1943.0, + 294.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1934.0, + 1120.0, + 1934.0, + 1120.0, + 1976.0, + 293.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 369.0, + 542.0, + 369.0, + 542.0, + 413.0, + 316.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 369.0, + 742.0, + 369.0, + 742.0, + 413.0, + 653.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 369.0, + 868.0, + 369.0, + 868.0, + 413.0, + 801.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 369.0, + 1185.0, + 369.0, + 1185.0, + 413.0, + 914.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 402.0, + 1066.0, + 402.0, + 1066.0, + 446.0, + 311.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 434.0, + 806.0, + 434.0, + 806.0, + 476.0, + 312.0, + 476.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 231, + 1405, + 231, + 1405, + 687, + 297, + 687 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1203, + 1405, + 1203, + 1405, + 1508, + 298, + 1508 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1523, + 1403, + 1523, + 1403, + 1707, + 298, + 1707 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1722, + 1404, + 1722, + 1404, + 1845, + 299, + 1845 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1404, + 1911, + 1404, + 2037, + 299, + 2037 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 703, + 1405, + 703, + 1405, + 857, + 297, + 857 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 300, + 1056, + 1398, + 1056, + 1398, + 1148, + 300, + 1148 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 297, + 864, + 1405, + 864, + 1405, + 1048, + 297, + 1048 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 298, + 1863, + 855, + 1863, + 855, + 1897, + 298, + 1897 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 300, + 1168, + 579, + 1168, + 579, + 1198, + 300, + 1198 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.692 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.27 + }, + { + "category_id": 13, + "poly": [ + 298, + 1973, + 449, + 1973, + 449, + 2007, + 298, + 2007 + ], + "score": 0.92, + "latex": "s _ { i n i t } \\in S _ { f r e e }" + }, + { + "category_id": 13, + "poly": [ + 695, + 1973, + 857, + 1973, + 857, + 2006, + 695, + 2006 + ], + "score": 0.92, + "latex": "\\mathcal { S } _ { g o a l } \\subsetneq S _ { f r e e }" + }, + { + "category_id": 13, + "poly": [ + 610, + 1943, + 727, + 1943, + 727, + 1975, + 610, + 1975 + ], + "score": 0.92, + "latex": "S _ { o b s } \\subsetneq S" + }, + { + "category_id": 13, + "poly": [ + 975, + 1942, + 1192, + 1942, + 1192, + 1976, + 975, + 1976 + ], + "score": 0.91, + "latex": "S _ { f r e e } : = \\mathcal { S } \\setminus \\mathcal { S } _ { o b s }" + }, + { + "category_id": 13, + "poly": [ + 862, + 2002, + 1190, + 2002, + 1190, + 2038, + 862, + 2038 + ], + "score": 0.91, + "latex": "\\Xi : = \\{ \\bar { \\xi } ( \\cdot ) : \\bar { [ 0 , 1 ] } S _ { f r e e } \\}" + }, + { + "category_id": 13, + "poly": [ + 343, + 1911, + 440, + 1911, + 440, + 1942, + 343, + 1942 + ], + "score": 0.91, + "latex": "S \\subseteq \\mathbb { R } ^ { q }" + }, + { + "category_id": 13, + "poly": [ + 1247, + 2002, + 1401, + 2002, + 1401, + 2037, + 1247, + 2037 + ], + "score": 0.89, + "latex": "c \\left( \\cdot \\right) : \\Xi \\mapsto \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 601, + 1325, + 637, + 1325, + 637, + 1353, + 601, + 1353 + ], + "score": 0.87, + "latex": "A ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1120, + 1294, + 1157, + 1294, + 1157, + 1322, + 1120, + 1322 + ], + "score": 0.87, + "latex": "A ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1294, + 375, + 1294, + 375, + 1324, + 298, + 1324 + ], + "score": 0.57, + "latex": "\\mathrm { ( B I T ^ { * } ) }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1859.0, + 860.0, + 1859.0, + 860.0, + 1902.0, + 293.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1164.0, + 582.0, + 1164.0, + 582.0, + 1202.0, + 295.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 259.0, + 1406.0, + 259.0, + 1406.0, + 299.0, + 292.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 1406.0, + 321.0, + 1406.0, + 357.0, + 292.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 351.0, + 1408.0, + 351.0, + 1408.0, + 389.0, + 292.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 417.0, + 294.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 412.0, + 1405.0, + 412.0, + 1405.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 444.0, + 1405.0, + 444.0, + 1405.0, + 479.0, + 295.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 473.0, + 1408.0, + 473.0, + 1408.0, + 512.0, + 291.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 504.0, + 1407.0, + 504.0, + 1407.0, + 542.0, + 292.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 535.0, + 1408.0, + 535.0, + 1408.0, + 569.0, + 294.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 564.0, + 1408.0, + 564.0, + 1408.0, + 603.0, + 294.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 593.0, + 1406.0, + 593.0, + 1406.0, + 633.0, + 291.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 624.0, + 1406.0, + 624.0, + 1406.0, + 665.0, + 292.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 658.0, + 1318.0, + 658.0, + 1318.0, + 692.0, + 295.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1199.0, + 1406.0, + 1199.0, + 1406.0, + 1240.0, + 292.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1231.0, + 1406.0, + 1231.0, + 1406.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1262.0, + 1405.0, + 1262.0, + 1405.0, + 1299.0, + 293.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1295.0, + 1119.0, + 1295.0, + 1119.0, + 1326.0, + 376.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1295.0, + 1407.0, + 1295.0, + 1407.0, + 1326.0, + 1158.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1324.0, + 600.0, + 1324.0, + 600.0, + 1359.0, + 295.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1324.0, + 1407.0, + 1324.0, + 1407.0, + 1359.0, + 638.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1353.0, + 1406.0, + 1353.0, + 1406.0, + 1388.0, + 293.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1385.0, + 1410.0, + 1385.0, + 1410.0, + 1420.0, + 293.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1415.0, + 1407.0, + 1415.0, + 1407.0, + 1450.0, + 295.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1445.0, + 1406.0, + 1445.0, + 1406.0, + 1480.0, + 293.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 1322.0, + 1477.0, + 1322.0, + 1512.0, + 293.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1520.0, + 1408.0, + 1520.0, + 1408.0, + 1561.0, + 293.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1552.0, + 1405.0, + 1552.0, + 1405.0, + 1588.0, + 292.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1582.0, + 1405.0, + 1582.0, + 1405.0, + 1619.0, + 293.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1616.0, + 1404.0, + 1616.0, + 1404.0, + 1648.0, + 296.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1644.0, + 1405.0, + 1644.0, + 1405.0, + 1682.0, + 293.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1677.0, + 1236.0, + 1677.0, + 1236.0, + 1709.0, + 296.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1720.0, + 1406.0, + 1720.0, + 1406.0, + 1758.0, + 294.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1751.0, + 1405.0, + 1751.0, + 1405.0, + 1788.0, + 293.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1786.0, + 1403.0, + 1786.0, + 1403.0, + 1818.0, + 295.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1813.0, + 879.0, + 1813.0, + 879.0, + 1850.0, + 293.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1908.0, + 342.0, + 1908.0, + 342.0, + 1946.0, + 292.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1908.0, + 1406.0, + 1908.0, + 1406.0, + 1946.0, + 441.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 609.0, + 1940.0, + 609.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1940.0, + 974.0, + 1940.0, + 974.0, + 1979.0, + 728.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1940.0, + 1409.0, + 1940.0, + 1409.0, + 1979.0, + 1193.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 297.0, + 1972.0, + 297.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1972.0, + 694.0, + 1972.0, + 694.0, + 2008.0, + 450.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1972.0, + 1408.0, + 1972.0, + 1408.0, + 2008.0, + 858.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 861.0, + 2000.0, + 861.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 2000.0, + 1246.0, + 2000.0, + 1246.0, + 2040.0, + 1191.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 2000.0, + 1406.0, + 2000.0, + 1406.0, + 2040.0, + 1402.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 701.0, + 1408.0, + 701.0, + 1408.0, + 739.0, + 294.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 735.0, + 1406.0, + 735.0, + 1406.0, + 768.0, + 296.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 761.0, + 1408.0, + 761.0, + 1408.0, + 801.0, + 292.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 795.0, + 1406.0, + 795.0, + 1406.0, + 830.0, + 294.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 826.0, + 1313.0, + 826.0, + 1313.0, + 862.0, + 295.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1055.0, + 1404.0, + 1055.0, + 1404.0, + 1092.0, + 295.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1088.0, + 1403.0, + 1088.0, + 1403.0, + 1121.0, + 296.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1116.0, + 798.0, + 1116.0, + 798.0, + 1150.0, + 296.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 861.0, + 1406.0, + 861.0, + 1406.0, + 901.0, + 294.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 894.0, + 1405.0, + 894.0, + 1405.0, + 932.0, + 320.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 926.0, + 1149.0, + 926.0, + 1149.0, + 959.0, + 320.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 956.0, + 1405.0, + 956.0, + 1405.0, + 992.0, + 301.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 986.0, + 1405.0, + 986.0, + 1405.0, + 1021.0, + 320.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1017.0, + 843.0, + 1017.0, + 843.0, + 1052.0, + 323.0, + 1052.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 336, + 1405, + 336, + 1405, + 596, + 296, + 596 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 296, + 765, + 1406, + 765, + 1406, + 1012, + 296, + 1012 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 610, + 1403, + 610, + 1403, + 730, + 297, + 730 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 357, + 1013, + 1336, + 1013, + 1336, + 1452, + 357, + 1452 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 299, + 1639, + 835, + 1639, + 835, + 2034, + 299, + 2034 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 297, + 1464, + 1406, + 1464, + 1406, + 1553, + 297, + 1553 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 292, + 229, + 1401, + 229, + 1401, + 294, + 292, + 294 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 593, + 725, + 1103, + 725, + 1103, + 762, + 593, + 762 + ], + "score": 0.93 + }, + { + "category_id": 8, + "poly": [ + 518, + 296, + 1175, + 296, + 1175, + 332, + 518, + 332 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 300, + 1579, + 578, + 1579, + 578, + 1613, + 300, + 1613 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 827, + 1636, + 1360, + 1636, + 1360, + 2035, + 827, + 2035 + ], + "score": 0.873 + }, + { + "category_id": 9, + "poly": [ + 1366, + 727, + 1399, + 727, + 1399, + 756, + 1366, + 756 + ], + "score": 0.864 + }, + { + "category_id": 9, + "poly": [ + 1366, + 298, + 1400, + 298, + 1400, + 327, + 1366, + 327 + ], + "score": 0.827 + }, + { + "category_id": 0, + "poly": [ + 848, + 1569, + 1361, + 1569, + 1361, + 1602, + 848, + 1602 + ], + "score": 0.78 + }, + { + "category_id": 0, + "poly": [ + 831, + 1611, + 960, + 1611, + 960, + 1636, + 831, + 1636 + ], + "score": 0.71 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.644 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.433 + }, + { + "category_id": 13, + "poly": [ + 533, + 555, + 742, + 555, + 742, + 595, + 533, + 595 + ], + "score": 0.93, + "latex": "\\textstyle \\sum _ { i = 1 } ^ { T } c \\left( \\left[ s _ { i - 1 } , s _ { i } \\right] \\right)" + }, + { + "category_id": 13, + "poly": [ + 803, + 1492, + 902, + 1492, + 902, + 1526, + 803, + 1526 + ], + "score": 0.93, + "latex": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}" + }, + { + "category_id": 13, + "poly": [ + 1189, + 521, + 1308, + 521, + 1308, + 555, + 1189, + 555 + ], + "score": 0.93, + "latex": "\\xi ( 1 ) = s _ { T }" + }, + { + "category_id": 13, + "poly": [ + 1252, + 369, + 1397, + 369, + 1397, + 403, + 1252, + 403 + ], + "score": 0.93, + "latex": "\\mathbb { R } ^ { 3 } \\mapsto \\{ 0 , 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 933, + 522, + 1138, + 522, + 1138, + 555, + 933, + 555 + ], + "score": 0.92, + "latex": "\\xi ( 0 ) = s _ { 0 } = s _ { i n i t }" + }, + { + "category_id": 13, + "poly": [ + 753, + 262, + 815, + 262, + 815, + 295, + 753, + 295 + ], + "score": 0.92, + "latex": "\\mathcal { S } _ { g o a l }" + }, + { + "category_id": 13, + "poly": [ + 779, + 401, + 846, + 401, + 846, + 434, + 779, + 434 + ], + "score": 0.91, + "latex": "\\boldsymbol { S } _ { f r e e }" + }, + { + "category_id": 14, + "poly": [ + 592, + 722, + 1104, + 722, + 1104, + 762, + 592, + 762 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\mathbf { a } \\mathbf { 1 g } ^ { * } \\left( \\cdot \\right) = \\operatorname * { a r g m i n } _ { \\mathbf { a } \\mathbf { 1 g } \\in \\mathcal { A } } \\mathbb { E } _ { U \\in \\mathcal { U } } \\left[ \\ell \\left( \\mathbf { a } \\mathbf { 1 g } ( U ) \\right) \\right] , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1111, + 336, + 1315, + 336, + 1315, + 369, + 1111, + 369 + ], + "score": 0.91, + "latex": "( s _ { i n i t } , S _ { g o a l } , c \\left( \\cdot \\right) )" + }, + { + "category_id": 13, + "poly": [ + 1084, + 369, + 1216, + 369, + 1216, + 403, + 1084, + 403 + ], + "score": 0.91, + "latex": "\\mathtt { m a p } ( \\cdot ) : \\mathbb { R } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 448, + 261, + 491, + 261, + 491, + 295, + 448, + 295 + ], + "score": 0.9, + "latex": "c ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 953, + 1672, + 1084, + 1672, + 1084, + 1703, + 953, + 1703 + ], + "score": 0.9, + "latex": "\\boldsymbol { \\mathcal { T } } = ( \\boldsymbol { \\mathcal { V } } , \\boldsymbol { \\mathcal { E } } )" + }, + { + "category_id": 13, + "poly": [ + 359, + 1885, + 414, + 1885, + 414, + 1913, + 359, + 1913 + ], + "score": 0.9, + "latex": "s _ { i n i t }" + }, + { + "category_id": 13, + "poly": [ + 1145, + 1671, + 1289, + 1671, + 1289, + 1703, + 1145, + 1703 + ], + "score": 0.88, + "latex": "\\mathcal { V } \\{ s _ { i n i t } \\}" + }, + { + "category_id": 13, + "poly": [ + 746, + 1973, + 809, + 1973, + 809, + 2007, + 746, + 2007 + ], + "score": 0.88, + "latex": "\\mathcal { S } _ { g o a l }" + }, + { + "category_id": 13, + "poly": [ + 966, + 263, + 1031, + 263, + 1031, + 295, + 966, + 295 + ], + "score": 0.88, + "latex": "\\boldsymbol { S } _ { f r e e }" + }, + { + "category_id": 13, + "poly": [ + 611, + 266, + 666, + 266, + 666, + 293, + 611, + 293 + ], + "score": 0.88, + "latex": "s _ { i n i t }" + }, + { + "category_id": 13, + "poly": [ + 888, + 1733, + 959, + 1733, + 959, + 1759, + 888, + 1759 + ], + "score": 0.87, + "latex": "t \\gets 0" + }, + { + "category_id": 13, + "poly": [ + 579, + 1491, + 614, + 1491, + 614, + 1522, + 579, + 1522 + ], + "score": 0.87, + "latex": "\\tilde { V } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1011, + 856, + 1091, + 856, + 1091, + 889, + 1011, + 889 + ], + "score": 0.87, + "latex": "\\mathsf { a l g } \\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 916, + 796, + 996, + 796, + 996, + 829, + 916, + 829 + ], + "score": 0.87, + "latex": "\\mathsf { a l g } \\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 465, + 522, + 872, + 522, + 872, + 555, + 465, + 555 + ], + "score": 0.87, + "latex": "\\left[ s _ { 0 } , s _ { 1 } \\right] , \\left[ s _ { 1 } , s _ { 2 } \\right] , \\ldots , \\left[ s _ { T - 1 } , s _ { T } \\right] \\subset \\mathcal { S }" + }, + { + "category_id": 13, + "poly": [ + 860, + 1700, + 938, + 1700, + 938, + 1730, + 860, + 1730 + ], + "score": 0.86, + "latex": "\\mathcal { E } \\emptyset" + }, + { + "category_id": 13, + "poly": [ + 736, + 608, + 1241, + 608, + 1241, + 648, + 736, + 648 + ], + "score": 0.86, + "latex": "\\{ U _ { i } : = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c \\left( \\cdot \\right) ) \\} _ { i = 1 } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 871, + 766, + 920, + 766, + 920, + 798, + 871, + 798 + ], + "score": 0.86, + "latex": "\\ell \\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 661, + 1495, + 692, + 1495, + 692, + 1522, + 661, + 1522 + ], + "score": 0.86, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1040, + 647, + 1119, + 647, + 1119, + 677, + 1040, + 677 + ], + "score": 0.82, + "latex": "\\mathsf { a l g } \\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 960, + 493, + 976, + 493, + 976, + 523, + 960, + 523 + ], + "score": 0.82, + "latex": "\\xi" + }, + { + "category_id": 13, + "poly": [ + 847, + 1639, + 1272, + 1639, + 1272, + 1672, + 847, + 1672 + ], + "score": 0.81, + "latex": "U = ( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathrm { { m a p } , \\it { c ( \\cdot ) } ) }" + }, + { + "category_id": 13, + "poly": [ + 558, + 1851, + 585, + 1851, + 585, + 1879, + 558, + 1879 + ], + "score": 0.8, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 370, + 767, + 395, + 767, + 395, + 793, + 370, + 793 + ], + "score": 0.78, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 931, + 1822, + 1131, + 1822, + 1131, + 1854, + 931, + 1854 + ], + "score": 0.77, + "latex": "\\mathcal { V } \\mathcal { V } \\cup \\{ s _ { n e w } \\}" + }, + { + "category_id": 13, + "poly": [ + 505, + 646, + 529, + 646, + 529, + 672, + 505, + 672 + ], + "score": 0.77, + "latex": "\\mathcal { U }" + }, + { + "category_id": 13, + "poly": [ + 466, + 919, + 482, + 919, + 482, + 945, + 466, + 945 + ], + "score": 0.74, + "latex": "\\ell" + }, + { + "category_id": 14, + "poly": [ + 525, + 296, + 1172, + 296, + 1172, + 332, + 525, + 332 + ], + "score": 0.73, + "latex": "\\begin{array} { r } { \\xi ^ { * } = \\mathrm { a r g m i n } _ { \\xi \\in \\Xi } \\ c \\left( \\xi \\right) , \\quad \\mathrm { s . t . } \\xi ( 0 ) = s _ { i n i t } , \\xi ( 1 ) \\in S _ { g o a l } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 734, + 1913, + 754, + 1913, + 754, + 1939, + 734, + 1939 + ], + "score": 0.72, + "latex": "\\mathcal { E }" + }, + { + "category_id": 13, + "poly": [ + 992, + 1732, + 1017, + 1732, + 1017, + 1758, + 992, + 1758 + ], + "score": 0.68, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 938, + 1853, + 1250, + 1853, + 1250, + 1886, + 938, + 1886 + ], + "score": 0.67, + "latex": "\\mathcal { E } \\mathcal { E } \\cup \\{ [ s _ { p a r e n t } , s _ { n e w } ] \\}" + }, + { + "category_id": 13, + "poly": [ + 790, + 1882, + 814, + 1882, + 814, + 1910, + 790, + 1910 + ], + "score": 0.63, + "latex": "\\nu" + }, + { + "category_id": 14, + "poly": [ + 525, + 296, + 788, + 296, + 788, + 334, + 525, + 334 + ], + "score": 0.63, + "latex": "\\begin{array} { r } { \\xi ^ { * } = \\operatorname * { a r g m i n } _ { \\xi \\in \\Xi } c \\left( \\xi \\right) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 983, + 1764, + 1101, + 1764, + 1101, + 1793, + 983, + 1793 + ], + "score": 0.63, + "latex": "s _ { n e w } \\mathrm { E }" + }, + { + "category_id": 13, + "poly": [ + 1220, + 462, + 1244, + 462, + 1244, + 489, + 1220, + 489 + ], + "score": 0.56, + "latex": "\\Xi" + }, + { + "category_id": 13, + "poly": [ + 1191, + 1761, + 1270, + 1761, + 1270, + 1791, + 1191, + 1791 + ], + "score": 0.49, + "latex": "( \\mathcal { T } , U )" + }, + { + "category_id": 13, + "poly": [ + 1126, + 1795, + 1288, + 1795, + 1288, + 1825, + 1126, + 1825 + ], + "score": 0.46, + "latex": "\\prime s _ { p a r e n t } , s _ { n e w } )" + }, + { + "category_id": 13, + "poly": [ + 862, + 295, + 1172, + 295, + 1172, + 332, + 862, + 332 + ], + "score": 0.42, + "latex": "\\xi ( 0 ) = s _ { i n i t } , \\xi ( 1 ) \\in S _ { g o a l } ." + }, + { + "category_id": 13, + "poly": [ + 944, + 1919, + 965, + 1919, + 965, + 1941, + 944, + 1941 + ], + "score": 0.39, + "latex": "\\triangleright" + }, + { + "category_id": 13, + "poly": [ + 956, + 1945, + 1108, + 1945, + 1108, + 1976, + 956, + 1976 + ], + "score": 0.38, + "latex": "s _ { n e w } \\in S _ { g o a l }" + }, + { + "category_id": 13, + "poly": [ + 524, + 296, + 785, + 296, + 785, + 334, + 524, + 334 + ], + "score": 0.37, + "latex": "\\begin{array} { r } { \\xi ^ { * } = \\operatorname * { a r g m i n } _ { \\xi \\in \\Xi } c \\left( \\xi \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 862, + 295, + 1003, + 295, + 1003, + 331, + 862, + 331 + ], + "score": 0.3, + "latex": "\\xi ( 0 ) = s _ { i n i t }" + }, + { + "category_id": 13, + "poly": [ + 466, + 521, + 546, + 521, + 546, + 555, + 466, + 555 + ], + "score": 0.28, + "latex": "[ s _ { 0 } , s _ { 1 } ]" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 1009.0, + 760.0, + 1009.0, + 760.0, + 1019.0, + 750.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1019.0, + 569.0, + 1019.0, + 569.0, + 1052.0, + 463.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1015.0, + 795.0, + 1015.0, + 795.0, + 1052.0, + 712.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1018.0, + 1030.0, + 1018.0, + 1030.0, + 1052.0, + 940.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1018.0, + 1272.0, + 1018.0, + 1272.0, + 1052.0, + 1181.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1056.0, + 626.0, + 1056.0, + 626.0, + 1091.0, + 573.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1057.0, + 859.0, + 1057.0, + 859.0, + 1090.0, + 806.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1057.0, + 1094.0, + 1057.0, + 1094.0, + 1090.0, + 1041.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1054.0, + 1329.0, + 1054.0, + 1329.0, + 1085.0, + 1276.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1082.0, + 612.0, + 1082.0, + 612.0, + 1106.0, + 586.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1081.0, + 848.0, + 1081.0, + 848.0, + 1108.0, + 818.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1082.0, + 1081.0, + 1082.0, + 1081.0, + 1104.0, + 1056.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 1076.0, + 1316.0, + 1076.0, + 1316.0, + 1102.0, + 1287.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1108.0, + 718.0, + 1108.0, + 718.0, + 1122.0, + 704.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1108.0, + 953.0, + 1108.0, + 953.0, + 1121.0, + 939.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1097.0, + 1096.0, + 1097.0, + 1096.0, + 1126.0, + 1009.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1103.0, + 1187.0, + 1103.0, + 1187.0, + 1116.0, + 1172.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 1111.0, + 1275.0, + 1111.0, + 1275.0, + 1127.0, + 1261.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1153.0, + 717.0, + 1153.0, + 717.0, + 1166.0, + 704.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1149.0, + 766.0, + 1149.0, + 766.0, + 1168.0, + 746.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1154.0, + 951.0, + 1154.0, + 951.0, + 1164.0, + 940.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1152.0, + 998.0, + 1152.0, + 998.0, + 1166.0, + 986.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1142.0, + 1067.0, + 1142.0, + 1067.0, + 1181.0, + 999.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1137.0, + 1160.0, + 1137.0, + 1160.0, + 1168.0, + 1106.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1151.0, + 1185.0, + 1151.0, + 1185.0, + 1158.0, + 1175.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1149.0, + 1231.0, + 1149.0, + 1231.0, + 1161.0, + 1221.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1161.0, + 388.0, + 1161.0, + 388.0, + 1276.0, + 358.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1170.0, + 459.0, + 1170.0, + 459.0, + 1217.0, + 405.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1172.0, + 691.0, + 1172.0, + 691.0, + 1216.0, + 637.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1166.0, + 795.0, + 1166.0, + 795.0, + 1226.0, + 714.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1171.0, + 928.0, + 1171.0, + 928.0, + 1218.0, + 874.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1172.0, + 1061.0, + 1172.0, + 1061.0, + 1213.0, + 1005.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1165.0, + 1190.0, + 1165.0, + 1190.0, + 1230.0, + 1118.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1170.0, + 1231.0, + 1170.0, + 1231.0, + 1203.0, + 1198.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1173.0, + 1319.0, + 1173.0, + 1319.0, + 1224.0, + 1240.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1246.0, + 463.0, + 1246.0, + 463.0, + 1282.0, + 406.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1247.0, + 699.0, + 1247.0, + 699.0, + 1282.0, + 640.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1253.0, + 743.0, + 1253.0, + 743.0, + 1262.0, + 732.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1249.0, + 936.0, + 1249.0, + 936.0, + 1281.0, + 878.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1252.0, + 979.0, + 1252.0, + 979.0, + 1262.0, + 968.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1248.0, + 1170.0, + 1248.0, + 1170.0, + 1280.0, + 1112.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1252.0, + 1214.0, + 1252.0, + 1214.0, + 1262.0, + 1204.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1257.0, + 705.0, + 1257.0, + 705.0, + 1266.0, + 696.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1258.0, + 940.0, + 1258.0, + 940.0, + 1266.0, + 930.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1257.0, + 1177.0, + 1257.0, + 1177.0, + 1266.0, + 1163.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1275.0, + 447.0, + 1275.0, + 447.0, + 1298.0, + 423.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 1277.0, + 677.0, + 1277.0, + 677.0, + 1295.0, + 658.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1277.0, + 914.0, + 1277.0, + 914.0, + 1297.0, + 895.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1276.0, + 1036.0, + 1276.0, + 1036.0, + 1322.0, + 981.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1277.0, + 1148.0, + 1277.0, + 1148.0, + 1297.0, + 1130.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1334.0, + 387.0, + 1334.0, + 387.0, + 1448.0, + 359.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1322.0, + 617.0, + 1322.0, + 617.0, + 1354.0, + 564.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1307.0, + 697.0, + 1307.0, + 697.0, + 1321.0, + 685.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1332.0, + 716.0, + 1332.0, + 716.0, + 1348.0, + 702.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1322.0, + 851.0, + 1322.0, + 851.0, + 1354.0, + 798.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1305.0, + 933.0, + 1305.0, + 933.0, + 1321.0, + 920.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1309.0, + 1086.0, + 1309.0, + 1086.0, + 1354.0, + 938.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1307.0, + 1169.0, + 1307.0, + 1169.0, + 1321.0, + 1154.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1339.0, + 1262.0, + 1339.0, + 1262.0, + 1349.0, + 1253.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 1321.0, + 1321.0, + 1321.0, + 1321.0, + 1354.0, + 1268.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1348.0, + 605.0, + 1348.0, + 605.0, + 1371.0, + 580.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1344.0, + 748.0, + 1344.0, + 748.0, + 1373.0, + 667.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1348.0, + 838.0, + 1348.0, + 838.0, + 1371.0, + 813.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1348.0, + 1072.0, + 1348.0, + 1072.0, + 1371.0, + 1048.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1348.0, + 1306.0, + 1348.0, + 1306.0, + 1371.0, + 1283.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1366.0, + 736.0, + 1366.0, + 736.0, + 1409.0, + 681.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1369.0, + 1050.0, + 1369.0, + 1050.0, + 1412.0, + 948.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 1372.0, + 1195.0, + 1372.0, + 1195.0, + 1407.0, + 1121.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1381.0, + 1226.0, + 1381.0, + 1226.0, + 1400.0, + 1206.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1381.0, + 1268.0, + 1381.0, + 1268.0, + 1400.0, + 1247.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1376.0, + 1325.0, + 1376.0, + 1325.0, + 1404.0, + 1284.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 1439.0, + 997.0, + 1439.0, + 997.0, + 1451.0, + 977.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1439.0, + 1232.0, + 1439.0, + 1232.0, + 1451.0, + 1211.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.75, + 1108.5, + 1094.75, + 1108.5, + 1094.75, + 1154.5, + 1008.75, + 1154.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1336.0, + 1185.0, + 1336.0, + 1185.0, + 1345.0, + 1171.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1464.0, + 1406.0, + 1464.0, + 1406.0, + 1497.0, + 295.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1494.0, + 578.0, + 1494.0, + 578.0, + 1527.0, + 295.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1494.0, + 660.0, + 1494.0, + 660.0, + 1527.0, + 615.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1494.0, + 802.0, + 1494.0, + 802.0, + 1527.0, + 693.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1494.0, + 1406.0, + 1494.0, + 1406.0, + 1527.0, + 903.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1525.0, + 926.0, + 1525.0, + 926.0, + 1555.0, + 296.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1574.0, + 583.0, + 1574.0, + 583.0, + 1620.0, + 292.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1563.0, + 1363.0, + 1563.0, + 1363.0, + 1607.0, + 843.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1605.0, + 965.0, + 1605.0, + 965.0, + 1642.0, + 824.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 334.0, + 1110.0, + 334.0, + 1110.0, + 372.0, + 293.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 334.0, + 1406.0, + 334.0, + 1406.0, + 372.0, + 1316.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 363.0, + 1083.0, + 363.0, + 1083.0, + 407.0, + 290.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 363.0, + 1251.0, + 363.0, + 1251.0, + 407.0, + 1217.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 363.0, + 1408.0, + 363.0, + 1408.0, + 407.0, + 1398.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 398.0, + 778.0, + 398.0, + 778.0, + 437.0, + 294.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 398.0, + 1407.0, + 398.0, + 1407.0, + 437.0, + 847.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 430.0, + 1406.0, + 430.0, + 1406.0, + 463.0, + 293.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 462.0, + 1219.0, + 462.0, + 1219.0, + 494.0, + 295.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 462.0, + 1403.0, + 462.0, + 1403.0, + 494.0, + 1245.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 491.0, + 959.0, + 491.0, + 959.0, + 526.0, + 295.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 491.0, + 1406.0, + 491.0, + 1406.0, + 526.0, + 977.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 464.0, + 520.0, + 464.0, + 558.0, + 294.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 520.0, + 932.0, + 520.0, + 932.0, + 558.0, + 873.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 520.0, + 1188.0, + 520.0, + 1188.0, + 558.0, + 1139.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 520.0, + 1406.0, + 520.0, + 1406.0, + 558.0, + 1309.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 537.0, + 532.0, + 537.0, + 532.0, + 608.0, + 284.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 537.0, + 766.0, + 537.0, + 766.0, + 608.0, + 743.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 766.0, + 369.0, + 766.0, + 369.0, + 799.0, + 295.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 766.0, + 870.0, + 766.0, + 870.0, + 799.0, + 396.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 766.0, + 1404.0, + 766.0, + 1404.0, + 799.0, + 921.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 795.0, + 915.0, + 795.0, + 915.0, + 831.0, + 293.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 795.0, + 1409.0, + 795.0, + 1409.0, + 831.0, + 997.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 827.0, + 1406.0, + 827.0, + 1406.0, + 861.0, + 295.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 857.0, + 1010.0, + 857.0, + 1010.0, + 891.0, + 295.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 857.0, + 1406.0, + 857.0, + 1406.0, + 891.0, + 1092.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 889.0, + 1408.0, + 889.0, + 1408.0, + 920.0, + 293.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 914.0, + 465.0, + 914.0, + 465.0, + 956.0, + 291.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 914.0, + 1406.0, + 914.0, + 1406.0, + 956.0, + 483.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 948.0, + 1407.0, + 948.0, + 1407.0, + 982.0, + 293.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 979.0, + 740.0, + 979.0, + 740.0, + 1013.0, + 294.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 990.0, + 903.0, + 990.0, + 903.0, + 1017.0, + 832.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 589.0, + 504.0, + 589.0, + 504.0, + 667.0, + 283.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 589.0, + 735.0, + 589.0, + 735.0, + 667.0, + 530.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 589.0, + 1404.0, + 589.0, + 1404.0, + 667.0, + 1242.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 642.0, + 504.0, + 642.0, + 504.0, + 681.0, + 293.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 642.0, + 1039.0, + 642.0, + 1039.0, + 681.0, + 530.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 642.0, + 1405.0, + 642.0, + 1405.0, + 681.0, + 1120.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 671.0, + 1405.0, + 671.0, + 1405.0, + 711.0, + 292.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 700.0, + 443.0, + 700.0, + 443.0, + 739.0, + 293.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1638.0, + 817.0, + 1638.0, + 817.0, + 1670.0, + 298.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1669.0, + 841.0, + 1669.0, + 841.0, + 1701.0, + 297.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 817.0, + 1698.0, + 817.0, + 1732.0, + 295.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1727.0, + 842.0, + 1727.0, + 842.0, + 1764.0, + 295.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1759.0, + 842.0, + 1759.0, + 842.0, + 1794.0, + 295.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1788.0, + 841.0, + 1788.0, + 841.0, + 1824.0, + 295.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 822.0, + 1820.0, + 822.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 557.0, + 1851.0, + 557.0, + 1882.0, + 295.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1851.0, + 816.0, + 1851.0, + 816.0, + 1882.0, + 586.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 358.0, + 1882.0, + 358.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1882.0, + 789.0, + 1882.0, + 789.0, + 1915.0, + 415.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 1882.0, + 832.0, + 1882.0, + 832.0, + 1915.0, + 815.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1913.0, + 733.0, + 1913.0, + 733.0, + 1945.0, + 296.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1913.0, + 815.0, + 1913.0, + 815.0, + 1945.0, + 755.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 840.0, + 1941.0, + 840.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1968.0, + 745.0, + 1968.0, + 745.0, + 2011.0, + 294.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1968.0, + 838.0, + 1968.0, + 838.0, + 2011.0, + 810.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.25, + 1999.0, + 816.25, + 1999.0, + 816.25, + 2038.0, + 293.25, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 258.0, + 447.0, + 258.0, + 447.0, + 300.0, + 292.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 258.0, + 610.0, + 258.0, + 610.0, + 300.0, + 492.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 258.0, + 752.0, + 258.0, + 752.0, + 300.0, + 667.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 258.0, + 965.0, + 258.0, + 965.0, + 300.0, + 816.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 258.0, + 1092.0, + 258.0, + 1092.0, + 300.0, + 1032.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1633.0, + 846.0, + 1633.0, + 846.0, + 1676.0, + 842.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1633.0, + 1284.0, + 1633.0, + 1284.0, + 1676.0, + 1273.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1665.0, + 952.0, + 1665.0, + 952.0, + 1706.0, + 820.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1665.0, + 1144.0, + 1665.0, + 1144.0, + 1706.0, + 1085.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1665.0, + 1300.0, + 1665.0, + 1300.0, + 1706.0, + 1290.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1695.0, + 859.0, + 1695.0, + 859.0, + 1733.0, + 855.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1695.0, + 955.0, + 1695.0, + 955.0, + 1733.0, + 939.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1728.0, + 887.0, + 1728.0, + 887.0, + 1760.0, + 818.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 1728.0, + 991.0, + 1728.0, + 991.0, + 1760.0, + 960.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1728.0, + 1057.0, + 1728.0, + 1057.0, + 1760.0, + 1018.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1769.0, + 841.0, + 1769.0, + 841.0, + 1788.0, + 824.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1762.0, + 982.0, + 1762.0, + 982.0, + 1793.0, + 887.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1762.0, + 1190.0, + 1762.0, + 1190.0, + 1793.0, + 1102.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1762.0, + 1295.0, + 1762.0, + 1295.0, + 1793.0, + 1271.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 1800.0, + 841.0, + 1800.0, + 841.0, + 1820.0, + 825.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1787.0, + 1125.0, + 1787.0, + 1125.0, + 1830.0, + 884.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 1787.0, + 1364.0, + 1787.0, + 1364.0, + 1830.0, + 1289.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1830.0, + 844.0, + 1830.0, + 844.0, + 1850.0, + 824.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1816.0, + 930.0, + 1816.0, + 930.0, + 1859.0, + 927.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1816.0, + 1186.0, + 1816.0, + 1186.0, + 1859.0, + 1132.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1846.0, + 1264.0, + 1846.0, + 1264.0, + 1892.0, + 1251.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1889.0, + 843.0, + 1889.0, + 843.0, + 1912.0, + 824.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1878.0, + 1293.0, + 1878.0, + 1293.0, + 1918.0, + 930.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1913.0, + 943.0, + 1913.0, + 943.0, + 1945.0, + 938.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1913.0, + 1072.0, + 1913.0, + 1072.0, + 1945.0, + 966.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1951.0, + 844.0, + 1951.0, + 844.0, + 1974.0, + 824.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1936.0, + 955.0, + 1936.0, + 955.0, + 1983.0, + 926.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1936.0, + 1177.0, + 1936.0, + 1177.0, + 1983.0, + 1109.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1984.0, + 843.0, + 1984.0, + 843.0, + 2001.0, + 827.0, + 2001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1972.0, + 1098.0, + 1972.0, + 1098.0, + 2010.0, + 967.0, + 2010.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1486, + 1405, + 1486, + 1405, + 1744, + 296, + 1744 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1405, + 228, + 1405, + 445, + 298, + 445 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 296, + 1228, + 1405, + 1228, + 1405, + 1445, + 296, + 1445 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1757, + 1404, + 1757, + 1404, + 1974, + 297, + 1974 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 460, + 1404, + 460, + 1404, + 583, + 298, + 583 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 819, + 761, + 819, + 761, + 1037, + 298, + 1037 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1116, + 1405, + 1116, + 1405, + 1211, + 298, + 1211 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 298, + 664, + 1405, + 664, + 1405, + 758, + 298, + 758 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 298, + 1039, + 1401, + 1039, + 1401, + 1103, + 298, + 1103 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 566, + 1448, + 1130, + 1448, + 1130, + 1484, + 566, + 1484 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 299, + 609, + 1015, + 609, + 1015, + 644, + 299, + 644 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 319, + 2003, + 1399, + 2003, + 1399, + 2036, + 319, + 2036 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 298, + 776, + 784, + 776, + 784, + 807, + 298, + 807 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1451, + 1400, + 1451, + 1400, + 1479, + 1366, + 1479 + ], + "score": 0.842 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2111, + 840, + 2111 + ], + "score": 0.789 + }, + { + "category_id": 5, + "poly": [ + 776, + 825, + 1401, + 825, + 1401, + 1021, + 776, + 1021 + ], + "score": 0.205, + "html": "
Algorithm 2: NExT :: Expand(T = (V,ε),U)
Sparent ← argmaxs∈vΦ(s) ; Selection
{s1,*..,Sk} id π*(s'sparent,U)； >Candidates
Snew ← argmaxs'∈{s1,.,sk}(s')； DExpansion
4 return Sparent, Snew;
" + }, + { + "category_id": 8, + "poly": [ + 776, + 872, + 1364, + 872, + 1364, + 1020, + 776, + 1020 + ], + "score": 0.14 + }, + { + "category_id": 8, + "poly": [ + 776, + 825, + 1401, + 825, + 1401, + 1021, + 776, + 1021 + ], + "score": 0.125 + }, + { + "category_id": 13, + "poly": [ + 401, + 879, + 502, + 879, + 502, + 917, + 401, + 917 + ], + "score": 0.94, + "latex": "{ \\tilde { V } } ^ { * } ( s | U )" + }, + { + "category_id": 13, + "poly": [ + 1216, + 1880, + 1361, + 1880, + 1361, + 1914, + 1216, + 1914 + ], + "score": 0.93, + "latex": "\\bar { \\tilde { \\pi } } ^ { * } ( s ^ { \\prime } \\vert s , \\bar { U } ) ^ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 937, + 1553, + 1235, + 1553, + 1235, + 1624, + 937, + 1624 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\sigma _ { t } ( s ) = \\sqrt { \\frac { \\log { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } w ( s ^ { \\prime } ) } } { w ( s ) } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1113, + 1226, + 1241, + 1226, + 1241, + 1264, + 1113, + 1264 + ], + "score": 0.93, + "latex": "- \\tilde { V } ^ { \\ast } \\left( s | U \\right)" + }, + { + "category_id": 13, + "poly": [ + 865, + 2004, + 982, + 2004, + 982, + 2035, + 865, + 2035 + ], + "score": 0.93, + "latex": "\\tilde { \\pi } ^ { * } \\left( a | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 695, + 2004, + 819, + 2004, + 819, + 2035, + 695, + 2035 + ], + "score": 0.92, + "latex": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 577, + 1556, + 881, + 1556, + 881, + 1619, + 577, + 1619 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\bar { r } _ { t } ( s ) = \\frac { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } k \\left( s ^ { \\prime } , s \\right) r \\left( s ^ { \\prime } \\right) } { \\sum _ { s ^ { \\prime } \\in { \\mathcal S } _ { t } } k \\left( s ^ { \\prime } , s \\right) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 442, + 1414, + 496, + 1414, + 496, + 1447, + 442, + 1447 + ], + "score": 0.91, + "latex": "\\phi ( s )" + }, + { + "category_id": 13, + "poly": [ + 1033, + 1384, + 1092, + 1384, + 1092, + 1417, + 1033, + 1417 + ], + "score": 0.91, + "latex": "\\phi \\left( s \\right)" + }, + { + "category_id": 13, + "poly": [ + 1318, + 1572, + 1403, + 1572, + 1403, + 1608, + 1318, + 1608 + ], + "score": 0.91, + "latex": "k ( s ^ { \\prime } , s )" + }, + { + "category_id": 13, + "poly": [ + 814, + 1514, + 1133, + 1514, + 1133, + 1552, + 814, + 1552 + ], + "score": 0.91, + "latex": "S _ { t } = \\{ s _ { p a r e n t } ^ { 1 } , . . . , s _ { p a r e n t } ^ { t } \\}" + }, + { + "category_id": 13, + "poly": [ + 1011, + 1764, + 1095, + 1764, + 1095, + 1793, + 1011, + 1793 + ], + "score": 0.9, + "latex": "s _ { p a r e n t }" + }, + { + "category_id": 13, + "poly": [ + 619, + 945, + 752, + 945, + 752, + 978, + 619, + 978 + ], + "score": 0.9, + "latex": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 585, + 1620, + 861, + 1620, + 861, + 1657, + 585, + 1657 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { w ( s ) = \\sum _ { s ^ { \\prime } \\in S _ { t } } k ( s ^ { \\prime } , s ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1042, + 1712, + 1112, + 1712, + 1112, + 1741, + 1042, + 1741 + ], + "score": 0.9, + "latex": "\\lambda > 0" + }, + { + "category_id": 13, + "poly": [ + 769, + 326, + 827, + 326, + 827, + 353, + 769, + 353 + ], + "score": 0.9, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 500, + 1182, + 558, + 1182, + 558, + 1210, + 500, + 1210 + ], + "score": 0.89, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 381, + 1035, + 418, + 1035, + 418, + 1067, + 381, + 1067 + ], + "score": 0.89, + "latex": "{ \\tilde { V } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1209, + 323, + 1373, + 323, + 1373, + 355, + 1209, + 355 + ], + "score": 0.89, + "latex": "[ \\bar { s _ { p a r e n t } } , s _ { n e w } ]" + }, + { + "category_id": 13, + "poly": [ + 298, + 1294, + 372, + 1294, + 372, + 1320, + 298, + 1320 + ], + "score": 0.89, + "latex": "s \\in \\mathcal V" + }, + { + "category_id": 13, + "poly": [ + 990, + 296, + 1048, + 296, + 1048, + 323, + 990, + 323 + ], + "score": 0.88, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 1259, + 1124, + 1317, + 1124, + 1317, + 1149, + 1259, + 1149 + ], + "score": 0.88, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1261, + 352, + 1261, + 352, + 1294, + 298, + 1294 + ], + "score": 0.88, + "latex": "r \\left( s \\right)" + }, + { + "category_id": 13, + "poly": [ + 584, + 1762, + 642, + 1762, + 642, + 1791, + 584, + 1791 + ], + "score": 0.88, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1489, + 397, + 1489, + 397, + 1516, + 372, + 1516 + ], + "score": 0.86, + "latex": "\\bar { r } _ { t }" + }, + { + "category_id": 14, + "poly": [ + 569, + 1447, + 1127, + 1447, + 1127, + 1485, + 569, + 1485 + ], + "score": 0.86, + "latex": "\\begin{array} { r } { s _ { p a r e n t } = \\mathrm { a r g m a x } _ { s \\in \\mathcal { V } } \\phi ( s ) : = \\bar { r } _ { t } \\left( s \\right) + \\lambda \\sigma _ { t } \\left( s \\right) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 472, + 1040, + 505, + 1040, + 505, + 1068, + 472, + 1068 + ], + "score": 0.86, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 293, + 1354, + 293, + 1354, + 324, + 1213, + 324 + ], + "score": 0.85, + "latex": "s _ { p a r e n t } \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 449, + 1490, + 478, + 1490, + 478, + 1516, + 449, + 1516 + ], + "score": 0.85, + "latex": "\\sigma _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1189, + 1624, + 1215, + 1624, + 1215, + 1651, + 1189, + 1651 + ], + "score": 0.85, + "latex": "\\bar { r } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1272, + 1626, + 1301, + 1626, + 1301, + 1651, + 1272, + 1651 + ], + "score": 0.84, + "latex": "\\sigma _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1281, + 1943, + 1314, + 1943, + 1314, + 1970, + 1281, + 1970 + ], + "score": 0.81, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1376, + 1386, + 1402, + 1386, + 1402, + 1412, + 1376, + 1412 + ], + "score": 0.81, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1077, + 1152, + 1161, + 1152, + 1161, + 1182, + 1077, + 1182 + ], + "score": 0.81, + "latex": "s _ { p a r e n t }" + }, + { + "category_id": 13, + "poly": [ + 489, + 1821, + 508, + 1821, + 508, + 1847, + 489, + 1847 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 716, + 1793, + 801, + 1793, + 801, + 1824, + 716, + 1824 + ], + "score": 0.8, + "latex": "s _ { p a r e n t }" + }, + { + "category_id": 13, + "poly": [ + 479, + 1520, + 492, + 1520, + 492, + 1544, + 479, + 1544 + ], + "score": 0.8, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 926, + 1119, + 953, + 1119, + 953, + 1146, + 926, + 1146 + ], + "score": 0.8, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1370, + 1148, + 1395, + 1148, + 1395, + 1176, + 1370, + 1176 + ], + "score": 0.8, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 448, + 946, + 473, + 946, + 473, + 973, + 448, + 973 + ], + "score": 0.79, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 1016, + 1823, + 1100, + 1823, + 1100, + 1854, + 1016, + 1854 + ], + "score": 0.78, + "latex": "s _ { p a r e n t }" + }, + { + "category_id": 13, + "poly": [ + 790, + 950, + 1180, + 950, + 1180, + 984, + 790, + 984 + ], + "score": 0.78, + "latex": "s _ { n e w } \\gets \\mathrm { a r g m a x } _ { s ^ { \\prime } \\in \\{ s _ { 1 } , \\dots , s _ { k } \\} } \\phi ( s ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 1106, + 1489, + 1119, + 1489, + 1119, + 1513, + 1106, + 1513 + ], + "score": 0.76, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 733, + 978, + 758, + 978, + 758, + 1005, + 733, + 1005 + ], + "score": 0.76, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 978, + 697, + 1004, + 697, + 1004, + 724, + 978, + 724 + ], + "score": 0.75, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 410, + 1234, + 494, + 1234, + 494, + 1265, + 410, + 1265 + ], + "score": 0.75, + "latex": "s _ { p a r e n t }" + }, + { + "category_id": 13, + "poly": [ + 1073, + 522, + 1098, + 522, + 1098, + 550, + 1073, + 550 + ], + "score": 0.75, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1148, + 323, + 1148, + 323, + 1176, + 297, + 1176 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1118, + 323, + 1143, + 323, + 1143, + 350, + 1118, + 350 + ], + "score": 0.74, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 792, + 231, + 818, + 231, + 818, + 259, + 792, + 259 + ], + "score": 0.72, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 823, + 1182, + 906, + 1182, + 906, + 1213, + 823, + 1213 + ], + "score": 0.71, + "latex": "s _ { p a r e n t }" + }, + { + "category_id": 13, + "poly": [ + 525, + 921, + 542, + 921, + 542, + 942, + 525, + 942 + ], + "score": 0.68, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 422, + 1012, + 439, + 1012, + 439, + 1033, + 422, + 1033 + ], + "score": 0.67, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 876, + 989, + 1026, + 989, + 1026, + 1015, + 876, + 1015 + ], + "score": 0.61, + "latex": "s _ { p a r e n t } , s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 788, + 828, + 1313, + 828, + 1313, + 865, + 788, + 865 + ], + "score": 0.61, + "latex": "\\mathbf { A l g o r i t h m 2 : N E X T : : E x p a n d } ( { \\mathcal { T } } = ( { \\mathcal { V } } , { \\mathcal { E } } ) , U )" + }, + { + "category_id": 13, + "poly": [ + 560, + 1231, + 587, + 1231, + 587, + 1260, + 560, + 1260 + ], + "score": 0.6, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 792, + 909, + 1181, + 909, + 1181, + 948, + 792, + 948 + ], + "score": 0.59, + "latex": "\\{ s _ { 1 } , . . . , s _ { k } \\} \\stackrel { i i d . } { \\sim } \\tilde { \\pi } ^ { * } ( s ^ { \\prime } | s _ { p a r e n t } , U )" + }, + { + "category_id": 13, + "poly": [ + 539, + 1943, + 559, + 1943, + 559, + 1973, + 539, + 1973 + ], + "score": 0.57, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 784, + 873, + 1112, + 873, + 1112, + 906, + 784, + 906 + ], + "score": 0.54, + "latex": "s _ { p a r e n t } \\gets \\operatorname { a r g m a x } _ { s \\in \\mathcal { V } } \\phi ( s )" + }, + { + "category_id": 13, + "poly": [ + 539, + 1942, + 597, + 1942, + 597, + 1975, + 539, + 1975 + ], + "score": 0.36, + "latex": "\\phi \\left( s \\right)" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 605.0, + 1017.0, + 605.0, + 1017.0, + 650.0, + 292.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 2000.0, + 694.0, + 2000.0, + 694.0, + 2039.0, + 332.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 2000.0, + 864.0, + 2000.0, + 864.0, + 2039.0, + 820.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2039.0, + 983.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 775.0, + 786.0, + 775.0, + 786.0, + 810.0, + 295.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 837.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1482.0, + 371.0, + 1482.0, + 371.0, + 1521.0, + 294.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1482.0, + 448.0, + 1482.0, + 448.0, + 1521.0, + 398.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 1482.0, + 1105.0, + 1482.0, + 1105.0, + 1521.0, + 479.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1482.0, + 1406.0, + 1482.0, + 1406.0, + 1521.0, + 1120.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1499.0, + 478.0, + 1499.0, + 478.0, + 1565.0, + 286.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1499.0, + 813.0, + 1499.0, + 813.0, + 1565.0, + 493.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1499.0, + 1413.0, + 1499.0, + 1413.0, + 1565.0, + 1134.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1557.0, + 576.0, + 1557.0, + 576.0, + 1625.0, + 294.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1557.0, + 936.0, + 1557.0, + 936.0, + 1625.0, + 882.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1554.0, + 1317.0, + 1554.0, + 1317.0, + 1621.0, + 1236.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1612.0, + 584.0, + 1612.0, + 584.0, + 1665.0, + 288.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1612.0, + 1188.0, + 1612.0, + 1188.0, + 1665.0, + 862.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1612.0, + 1271.0, + 1612.0, + 1271.0, + 1665.0, + 1216.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 1612.0, + 1408.0, + 1612.0, + 1408.0, + 1665.0, + 1302.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1405.0, + 1651.0, + 1405.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1405.0, + 1682.0, + 1405.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1712.0, + 1041.0, + 1712.0, + 1041.0, + 1747.0, + 296.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1712.0, + 1123.0, + 1712.0, + 1123.0, + 1747.0, + 1113.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 229.0, + 791.0, + 229.0, + 791.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 266.0, + 819.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 288.0, + 989.0, + 288.0, + 989.0, + 330.0, + 291.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 288.0, + 1212.0, + 288.0, + 1212.0, + 330.0, + 1049.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 288.0, + 1407.0, + 288.0, + 1407.0, + 330.0, + 1355.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 319.0, + 768.0, + 319.0, + 768.0, + 361.0, + 292.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 319.0, + 1117.0, + 319.0, + 1117.0, + 361.0, + 828.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 319.0, + 1208.0, + 319.0, + 1208.0, + 361.0, + 1144.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 319.0, + 1410.0, + 319.0, + 1410.0, + 361.0, + 1374.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 351.0, + 1406.0, + 351.0, + 1406.0, + 386.0, + 295.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 413.0, + 1077.0, + 413.0, + 1077.0, + 448.0, + 295.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1226.0, + 409.0, + 1226.0, + 409.0, + 1266.0, + 290.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1226.0, + 559.0, + 1226.0, + 559.0, + 1266.0, + 495.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1226.0, + 1112.0, + 1226.0, + 1112.0, + 1266.0, + 588.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1226.0, + 1406.0, + 1226.0, + 1406.0, + 1266.0, + 1242.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1261.0, + 1405.0, + 1261.0, + 1405.0, + 1296.0, + 353.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1289.0, + 297.0, + 1289.0, + 297.0, + 1325.0, + 293.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1289.0, + 1406.0, + 1289.0, + 1406.0, + 1325.0, + 373.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1320.0, + 1403.0, + 1320.0, + 1403.0, + 1358.0, + 294.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1351.0, + 1407.0, + 1351.0, + 1407.0, + 1386.0, + 295.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1382.0, + 1032.0, + 1382.0, + 1032.0, + 1417.0, + 296.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1382.0, + 1375.0, + 1382.0, + 1375.0, + 1417.0, + 1093.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1413.0, + 441.0, + 1413.0, + 441.0, + 1449.0, + 294.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1754.0, + 583.0, + 1754.0, + 583.0, + 1799.0, + 292.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1754.0, + 1010.0, + 1754.0, + 1010.0, + 1799.0, + 643.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 1754.0, + 1406.0, + 1754.0, + 1406.0, + 1799.0, + 1096.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1787.0, + 715.0, + 1787.0, + 715.0, + 1830.0, + 291.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 1787.0, + 1406.0, + 1787.0, + 1406.0, + 1830.0, + 802.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 488.0, + 1817.0, + 488.0, + 1859.0, + 292.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1817.0, + 1015.0, + 1817.0, + 1015.0, + 1859.0, + 509.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 1817.0, + 1406.0, + 1817.0, + 1406.0, + 1859.0, + 1101.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1878.0, + 1215.0, + 1878.0, + 1215.0, + 1916.0, + 291.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 1878.0, + 1406.0, + 1878.0, + 1406.0, + 1916.0, + 1362.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1945.0, + 293.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 538.0, + 1941.0, + 538.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1941.0, + 1280.0, + 1941.0, + 1280.0, + 1976.0, + 598.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1941.0, + 1404.0, + 1941.0, + 1404.0, + 1976.0, + 1315.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 460.0, + 1405.0, + 460.0, + 1405.0, + 496.0, + 294.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 492.0, + 1404.0, + 492.0, + 1404.0, + 525.0, + 296.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 522.0, + 1072.0, + 522.0, + 1072.0, + 558.0, + 293.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 522.0, + 1404.0, + 522.0, + 1404.0, + 558.0, + 1099.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 552.0, + 1392.0, + 552.0, + 1392.0, + 587.0, + 295.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 817.0, + 763.0, + 817.0, + 763.0, + 853.0, + 296.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 849.0, + 761.0, + 849.0, + 761.0, + 882.0, + 295.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 881.0, + 400.0, + 881.0, + 400.0, + 916.0, + 296.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 881.0, + 759.0, + 881.0, + 759.0, + 916.0, + 503.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 915.0, + 524.0, + 915.0, + 524.0, + 947.0, + 295.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 915.0, + 764.0, + 915.0, + 764.0, + 947.0, + 543.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 944.0, + 447.0, + 944.0, + 447.0, + 979.0, + 295.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 944.0, + 618.0, + 944.0, + 618.0, + 979.0, + 474.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 944.0, + 762.0, + 944.0, + 762.0, + 979.0, + 753.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 975.0, + 732.0, + 975.0, + 732.0, + 1009.0, + 295.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1002.0, + 421.0, + 1002.0, + 421.0, + 1039.0, + 294.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1002.0, + 764.0, + 1002.0, + 764.0, + 1039.0, + 440.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1116.0, + 925.0, + 1116.0, + 925.0, + 1154.0, + 293.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1116.0, + 1258.0, + 1116.0, + 1258.0, + 1154.0, + 954.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1116.0, + 1405.0, + 1116.0, + 1405.0, + 1154.0, + 1318.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1143.0, + 296.0, + 1143.0, + 296.0, + 1187.0, + 292.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1143.0, + 1076.0, + 1143.0, + 1076.0, + 1187.0, + 324.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1143.0, + 1369.0, + 1143.0, + 1369.0, + 1187.0, + 1162.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1143.0, + 1407.0, + 1143.0, + 1407.0, + 1187.0, + 1396.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1176.0, + 499.0, + 1176.0, + 499.0, + 1216.0, + 293.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1176.0, + 822.0, + 1176.0, + 822.0, + 1216.0, + 559.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1176.0, + 1126.0, + 1176.0, + 1126.0, + 1216.0, + 907.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 662.0, + 1403.0, + 662.0, + 1403.0, + 700.0, + 293.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 689.0, + 977.0, + 689.0, + 977.0, + 735.0, + 291.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 689.0, + 1406.0, + 689.0, + 1406.0, + 735.0, + 1005.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 727.0, + 1063.0, + 727.0, + 1063.0, + 761.0, + 296.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1035.0, + 380.0, + 1035.0, + 380.0, + 1074.0, + 294.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1035.0, + 471.0, + 1035.0, + 471.0, + 1074.0, + 419.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1035.0, + 1405.0, + 1035.0, + 1405.0, + 1074.0, + 506.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1065.0, + 1328.0, + 1065.0, + 1328.0, + 1110.0, + 292.0, + 1110.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1112, + 1404, + 1112, + 1404, + 1296, + 298, + 1296 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 896, + 1405, + 896, + 1405, + 1057, + 297, + 1057 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 680, + 1406, + 680, + 1406, + 836, + 297, + 836 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 351, + 1320, + 1316, + 1320, + 1316, + 1659, + 351, + 1659 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 955, + 249, + 1404, + 249, + 1404, + 447, + 955, + 447 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 604, + 1401, + 604, + 1401, + 667, + 298, + 667 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 298, + 526, + 1400, + 526, + 1400, + 590, + 298, + 590 + ], + "score": 0.943 + }, + { + "category_id": 3, + "poly": [ + 309, + 202, + 915, + 202, + 915, + 501, + 309, + 501 + ], + "score": 0.942 + }, + { + "category_id": 4, + "poly": [ + 296, + 1672, + 1404, + 1672, + 1404, + 1786, + 296, + 1786 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 299, + 1806, + 1405, + 1806, + 1405, + 1898, + 299, + 1898 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 300, + 1074, + 828, + 1074, + 828, + 1105, + 300, + 1105 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 322, + 1940, + 1403, + 1940, + 1403, + 2038, + 322, + 2038 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 301, + 848, + 1244, + 848, + 1244, + 881, + 301, + 881 + ], + "score": 0.813 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1906, + 1399, + 1906, + 1399, + 1933, + 1366, + 1933 + ], + "score": 0.767 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.747 + }, + { + "category_id": 8, + "poly": [ + 451, + 1902, + 830, + 1902, + 830, + 1937, + 451, + 1937 + ], + "score": 0.727 + }, + { + "category_id": 8, + "poly": [ + 440, + 1900, + 1273, + 1900, + 1273, + 1939, + 440, + 1939 + ], + "score": 0.394 + }, + { + "category_id": 1, + "poly": [ + 440, + 1900, + 1273, + 1900, + 1273, + 1939, + 440, + 1939 + ], + "score": 0.272 + }, + { + "category_id": 4, + "poly": [ + 689, + 475, + 864, + 475, + 864, + 500, + 689, + 500 + ], + "score": 0.154 + }, + { + "category_id": 1, + "poly": [ + 301, + 848, + 1244, + 848, + 1244, + 881, + 301, + 881 + ], + "score": 0.092 + }, + { + "category_id": 13, + "poly": [ + 553, + 1023, + 686, + 1023, + 686, + 1058, + 553, + 1058 + ], + "score": 0.95, + "latex": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 1011, + 894, + 1117, + 894, + 1117, + 932, + 1011, + 932 + ], + "score": 0.93, + "latex": "\\tilde { V } ^ { * } \\left( s | U \\right)" + }, + { + "category_id": 13, + "poly": [ + 395, + 1020, + 501, + 1020, + 501, + 1058, + 395, + 1058 + ], + "score": 0.93, + "latex": "\\tilde { V } ^ { * } \\left( s | U \\right)" + }, + { + "category_id": 13, + "poly": [ + 1191, + 415, + 1314, + 415, + 1314, + 447, + 1191, + 447 + ], + "score": 0.93, + "latex": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 816, + 1673, + 940, + 1673, + 940, + 1704, + 816, + 1704 + ], + "score": 0.93, + "latex": "\\boldsymbol { s } ^ { w } = ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 1138, + 1804, + 1325, + 1804, + 1325, + 1840, + 1138, + 1840 + ], + "score": 0.92, + "latex": "\\mu ^ { w } ( s ^ { w } ) \\in \\mathbb { R } ^ { d \\times d }" + }, + { + "category_id": 13, + "poly": [ + 1174, + 897, + 1307, + 897, + 1307, + 932, + 1174, + 932 + ], + "score": 0.92, + "latex": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1973, + 1326, + 1973, + 1326, + 2001, + 1213, + 2001 + ], + "score": 0.92, + "latex": "d \\times d \\times 4" + }, + { + "category_id": 13, + "poly": [ + 926, + 1172, + 1069, + 1172, + 1069, + 1206, + 926, + 1206 + ], + "score": 0.92, + "latex": "s = \\mathbf { \\bar { \\rho } } ( s ^ { w } , s ^ { h } )" + }, + { + "category_id": 13, + "poly": [ + 1282, + 1938, + 1403, + 1938, + 1403, + 1972, + 1282, + 1972 + ], + "score": 0.91, + "latex": "f _ { i } ^ { w } ( s ^ { w } ) \\in" + }, + { + "category_id": 13, + "poly": [ + 983, + 1670, + 1059, + 1670, + 1059, + 1700, + 983, + 1700 + ], + "score": 0.91, + "latex": "\\mathbf { z } = s ^ { h }" + }, + { + "category_id": 13, + "poly": [ + 1045, + 306, + 1132, + 306, + 1132, + 336, + 1045, + 336 + ], + "score": 0.91, + "latex": "( N + 2 )" + }, + { + "category_id": 13, + "poly": [ + 1050, + 1233, + 1149, + 1233, + 1149, + 1263, + 1050, + 1263 + ], + "score": 0.91, + "latex": "s ^ { w } \\in \\mathbb { R } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 325, + 1969, + 425, + 1969, + 425, + 2000, + 325, + 2000 + ], + "score": 0.9, + "latex": "\\mathbb { R } ^ { d \\times d \\times d _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 1046, + 413, + 1144, + 413, + 1144, + 447, + 1046, + 447 + ], + "score": 0.9, + "latex": "\\tilde { V } ^ { * } \\left( s | U \\right)" + }, + { + "category_id": 13, + "poly": [ + 795, + 1973, + 832, + 1973, + 832, + 2005, + 795, + 2005 + ], + "score": 0.9, + "latex": "f _ { 0 } ^ { w }" + }, + { + "category_id": 13, + "poly": [ + 609, + 1869, + 644, + 1869, + 644, + 1898, + 609, + 1898 + ], + "score": 0.89, + "latex": "k _ { w }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 1171, + 1198, + 1171, + 1198, + 1201, + 1168, + 1201 + ], + "score": 0.88, + "latex": "s ^ { h }" + }, + { + "category_id": 13, + "poly": [ + 344, + 1171, + 375, + 1171, + 375, + 1201, + 344, + 1201 + ], + "score": 0.88, + "latex": "{ \\bf \\bar { \\Phi } } _ { s } h" + }, + { + "category_id": 13, + "poly": [ + 401, + 1940, + 436, + 1940, + 436, + 1971, + 401, + 1971 + ], + "score": 0.88, + "latex": "\\theta _ { i } ^ { w }" + }, + { + "category_id": 13, + "poly": [ + 1105, + 1974, + 1139, + 1974, + 1139, + 2001, + 1105, + 2001 + ], + "score": 0.87, + "latex": "s ^ { w }" + }, + { + "category_id": 13, + "poly": [ + 980, + 634, + 1115, + 634, + 1115, + 668, + 980, + 668 + ], + "score": 0.87, + "latex": "\\tt a l g \\left( \\cdot \\right) \\in \\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 859, + 1809, + 892, + 1809, + 892, + 1834, + 859, + 1834 + ], + "score": 0.87, + "latex": "s ^ { w }" + }, + { + "category_id": 13, + "poly": [ + 1031, + 1144, + 1065, + 1144, + 1065, + 1170, + 1031, + 1170 + ], + "score": 0.86, + "latex": "s ^ { w }" + }, + { + "category_id": 13, + "poly": [ + 488, + 1973, + 515, + 1973, + 515, + 2002, + 488, + 2002 + ], + "score": 0.85, + "latex": "d _ { i }" + }, + { + "category_id": 13, + "poly": [ + 978, + 279, + 1004, + 279, + 1004, + 303, + 978, + 303 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 692, + 1706, + 708, + 1706, + 708, + 1729, + 692, + 1729 + ], + "score": 0.83, + "latex": "_ y" + }, + { + "category_id": 13, + "poly": [ + 1167, + 713, + 1193, + 713, + 1193, + 741, + 1167, + 741 + ], + "score": 0.82, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 819, + 1942, + 845, + 1942, + 845, + 1968, + 819, + 1968 + ], + "score": 0.82, + "latex": "\\oplus" + }, + { + "category_id": 13, + "poly": [ + 1082, + 1173, + 1116, + 1173, + 1116, + 1202, + 1082, + 1202 + ], + "score": 0.82, + "latex": "s ^ { w }" + }, + { + "category_id": 13, + "poly": [ + 355, + 2003, + 808, + 2003, + 808, + 2038, + 355, + 2038 + ], + "score": 0.82, + "latex": "f _ { 0 } ^ { w } ( s ^ { w } ) _ { i j } = [ s _ { 1 } ^ { w } , s _ { 2 } ^ { w } , i , j ] , i , j = 1 , \\ldots , a" + }, + { + "category_id": 13, + "poly": [ + 677, + 1731, + 695, + 1731, + 695, + 1754, + 677, + 1754 + ], + "score": 0.79, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 717, + 744, + 744, + 744, + 744, + 771, + 717, + 771 + ], + "score": 0.79, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1154, + 774, + 1178, + 774, + 1178, + 801, + 1154, + 801 + ], + "score": 0.79, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 628, + 1706, + 647, + 1706, + 647, + 1726, + 628, + 1726 + ], + "score": 0.77, + "latex": "_ x" + }, + { + "category_id": 14, + "poly": [ + 450, + 1901, + 1273, + 1901, + 1273, + 1939, + 450, + 1939 + ], + "score": 0.75, + "latex": "\\mu ^ { w } ( s ^ { w } ) = { \\tt s o f t m a x 2 d } ( f _ { k _ { w } } ^ { w } ( s ^ { w } ) ) , \\qquad f _ { i + 1 } ^ { w } ( s ^ { w } ) = { \\tt r e l u } ( \\theta _ { i } ^ { w } \\oplus f _ { i } ^ { w } ( s ^ { w } ) ) ," + }, + { + "category_id": 13, + "poly": [ + 1337, + 1807, + 1355, + 1807, + 1355, + 1834, + 1337, + 1834 + ], + "score": 0.68, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1280, + 1734, + 1296, + 1734, + 1296, + 1753, + 1280, + 1753 + ], + "score": 0.65, + "latex": "\\mathbf { z }" + }, + { + "category_id": 13, + "poly": [ + 649, + 2003, + 813, + 2003, + 813, + 2035, + 649, + 2035 + ], + "score": 0.46, + "latex": "i , j = 1 , \\ldots , d" + }, + { + "category_id": 14, + "poly": [ + 450, + 1902, + 827, + 1902, + 827, + 1938, + 450, + 1938 + ], + "score": 0.43, + "latex": "\\mu ^ { w } ( s ^ { w } ) = \\mathsf { s o f t m a x } 2 \\mathsf { d } ( f _ { k _ { w } } ^ { w } ( s ^ { w } ) )" + }, + { + "category_id": 13, + "poly": [ + 1232, + 1632, + 1306, + 1632, + 1306, + 1658, + 1232, + 1658 + ], + "score": 0.35, + "latex": "( \\mathbf { d } , \\mathbf { d } , \\mathbf { d } _ { \\mathbf { a } } )" + }, + { + "category_id": 13, + "poly": [ + 685, + 1632, + 761, + 1632, + 761, + 1659, + 685, + 1659 + ], + "score": 0.32, + "latex": "( \\mathbf { 1 } , \\mathbf { 1 } , \\mathbf { d } _ { \\mathbf { a } } )" + }, + { + "category_id": 13, + "poly": [ + 435, + 1387, + 447, + 1387, + 447, + 1399, + 435, + 1399 + ], + "score": 0.26, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1046, + 413, + 1080, + 413, + 1080, + 445, + 1046, + 445 + ], + "score": 0.25, + "latex": "\\tilde { V } ^ { * }" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1331.0, + 917.0, + 1331.0, + 917.0, + 1348.0, + 898.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1329.0, + 943.0, + 1329.0, + 943.0, + 1347.0, + 920.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1329.0, + 967.0, + 1329.0, + 967.0, + 1347.0, + 944.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1312.0, + 1322.0, + 1312.0, + 1322.0, + 1346.0, + 1113.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 1355.0, + 434.0, + 1355.0, + 434.0, + 1405.0, + 430.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1355.0, + 452.0, + 1355.0, + 452.0, + 1405.0, + 448.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1353.0, + 481.0, + 1353.0, + 481.0, + 1405.0, + 456.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1355.0, + 508.0, + 1355.0, + 508.0, + 1404.0, + 487.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1355.0, + 547.0, + 1355.0, + 547.0, + 1406.0, + 524.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1356.0, + 575.0, + 1356.0, + 575.0, + 1406.0, + 552.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 1353.0, + 603.0, + 1353.0, + 603.0, + 1406.0, + 579.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1355.0, + 638.0, + 1355.0, + 638.0, + 1401.0, + 621.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1355.0, + 666.0, + 1355.0, + 666.0, + 1401.0, + 647.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1355.0, + 695.0, + 1355.0, + 695.0, + 1400.0, + 676.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1353.0, + 735.0, + 1353.0, + 735.0, + 1403.0, + 713.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1357.0, + 761.0, + 1357.0, + 761.0, + 1403.0, + 743.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1355.0, + 790.0, + 1355.0, + 790.0, + 1405.0, + 768.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 1359.0, + 880.0, + 1359.0, + 880.0, + 1396.0, + 815.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1347.0, + 1011.0, + 1347.0, + 1011.0, + 1401.0, + 897.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 1351.0, + 1279.0, + 1351.0, + 1279.0, + 1412.0, + 1015.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1415.0, + 391.0, + 1415.0, + 391.0, + 1440.0, + 368.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1409.0, + 453.0, + 1409.0, + 453.0, + 1431.0, + 432.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1408.0, + 480.0, + 1408.0, + 480.0, + 1431.0, + 456.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1409.0, + 506.0, + 1409.0, + 506.0, + 1430.0, + 487.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1409.0, + 547.0, + 1409.0, + 547.0, + 1433.0, + 525.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1411.0, + 574.0, + 1411.0, + 574.0, + 1432.0, + 553.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 1409.0, + 603.0, + 1409.0, + 603.0, + 1432.0, + 579.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1409.0, + 638.0, + 1409.0, + 638.0, + 1430.0, + 621.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1409.0, + 666.0, + 1409.0, + 666.0, + 1430.0, + 648.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1411.0, + 694.0, + 1411.0, + 694.0, + 1430.0, + 676.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1411.0, + 734.0, + 1411.0, + 734.0, + 1430.0, + 715.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1414.0, + 756.0, + 1414.0, + 756.0, + 1427.0, + 747.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1408.0, + 790.0, + 1408.0, + 790.0, + 1432.0, + 768.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1399.0, + 933.0, + 1399.0, + 933.0, + 1422.0, + 916.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1392.0, + 977.0, + 1392.0, + 977.0, + 1444.0, + 934.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1393.0, + 1000.0, + 1393.0, + 1000.0, + 1439.0, + 981.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1392.0, + 1029.0, + 1392.0, + 1029.0, + 1443.0, + 1005.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1412.0, + 1222.0, + 1412.0, + 1222.0, + 1441.0, + 1196.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1412.0, + 1251.0, + 1412.0, + 1251.0, + 1441.0, + 1224.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 1412.0, + 1279.0, + 1412.0, + 1279.0, + 1441.0, + 1252.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1439.0, + 392.0, + 1439.0, + 392.0, + 1468.0, + 365.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1431.0, + 804.0, + 1431.0, + 804.0, + 1463.0, + 607.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1445.0, + 975.0, + 1445.0, + 975.0, + 1468.0, + 952.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1446.0, + 1000.0, + 1446.0, + 1000.0, + 1467.0, + 981.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1444.0, + 1028.0, + 1444.0, + 1028.0, + 1468.0, + 1005.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1476.0, + 656.0, + 1476.0, + 656.0, + 1503.0, + 624.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1481.0, + 1179.0, + 1481.0, + 1179.0, + 1511.0, + 1100.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1501.0, + 389.0, + 1501.0, + 389.0, + 1522.0, + 370.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1508.0, + 621.0, + 1508.0, + 621.0, + 1558.0, + 459.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1503.0, + 658.0, + 1503.0, + 658.0, + 1534.0, + 623.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1514.0, + 825.0, + 1514.0, + 825.0, + 1533.0, + 807.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1514.0, + 1122.0, + 1514.0, + 1122.0, + 1534.0, + 1104.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1507.0, + 1197.0, + 1507.0, + 1197.0, + 1532.0, + 1125.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1533.0, + 457.0, + 1533.0, + 457.0, + 1559.0, + 435.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 1551.0, + 613.0, + 1551.0, + 613.0, + 1560.0, + 605.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1534.0, + 652.0, + 1534.0, + 652.0, + 1559.0, + 629.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 1538.0, + 839.0, + 1538.0, + 839.0, + 1573.0, + 772.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1540.0, + 1122.0, + 1540.0, + 1122.0, + 1558.0, + 1104.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1536.0, + 1144.0, + 1536.0, + 1144.0, + 1557.0, + 1126.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1543.0, + 1185.0, + 1543.0, + 1185.0, + 1568.0, + 1162.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1544.0, + 1211.0, + 1544.0, + 1211.0, + 1568.0, + 1186.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1547.0, + 1234.0, + 1547.0, + 1234.0, + 1567.0, + 1215.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1567.0, + 409.0, + 1567.0, + 409.0, + 1599.0, + 349.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 1571.0, + 646.0, + 1571.0, + 646.0, + 1590.0, + 635.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 1565.0, + 914.0, + 1565.0, + 914.0, + 1598.0, + 717.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1560.0, + 1144.0, + 1560.0, + 1144.0, + 1582.0, + 1127.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1571.0, + 1185.0, + 1571.0, + 1185.0, + 1595.0, + 1162.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1571.0, + 1211.0, + 1571.0, + 1211.0, + 1595.0, + 1187.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 1571.0, + 1236.0, + 1571.0, + 1236.0, + 1595.0, + 1214.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1599.0, + 653.0, + 1599.0, + 653.0, + 1624.0, + 629.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1596.0, + 1238.0, + 1596.0, + 1238.0, + 1623.0, + 1162.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1628.0, + 684.0, + 1628.0, + 684.0, + 1664.0, + 526.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1628.0, + 767.0, + 1628.0, + 767.0, + 1664.0, + 762.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1627.0, + 1231.0, + 1627.0, + 1231.0, + 1662.0, + 1098.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1627.0, + 1312.0, + 1627.0, + 1312.0, + 1662.0, + 1307.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 248.0, + 1405.0, + 248.0, + 1405.0, + 283.0, + 955.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 274.0, + 977.0, + 274.0, + 977.0, + 311.0, + 954.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 274.0, + 1406.0, + 274.0, + 1406.0, + 311.0, + 1005.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 306.0, + 1044.0, + 306.0, + 1044.0, + 338.0, + 954.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 306.0, + 1407.0, + 306.0, + 1407.0, + 338.0, + 1133.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 331.0, + 1407.0, + 331.0, + 1407.0, + 366.0, + 955.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 359.0, + 1406.0, + 359.0, + 1406.0, + 391.0, + 953.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 390.0, + 1404.0, + 390.0, + 1404.0, + 416.0, + 957.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 414.0, + 1045.0, + 414.0, + 1045.0, + 449.0, + 955.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 414.0, + 1190.0, + 414.0, + 1190.0, + 449.0, + 1145.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 414.0, + 1324.0, + 414.0, + 1324.0, + 449.0, + 1315.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 270.0, + 920.0, + 270.0, + 920.0, + 306.0, + 821.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 339.0, + 491.0, + 339.0, + 491.0, + 364.0, + 477.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 384.0, + 575.0, + 384.0, + 575.0, + 418.0, + 335.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 405.0, + 898.0, + 405.0, + 898.0, + 440.0, + 768.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 472.0, + 537.0, + 472.0, + 537.0, + 502.0, + 343.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 469.0, + 868.0, + 469.0, + 868.0, + 504.0, + 684.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1669.0, + 815.0, + 1669.0, + 815.0, + 1706.0, + 295.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1669.0, + 982.0, + 1669.0, + 982.0, + 1706.0, + 941.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1669.0, + 1406.0, + 1669.0, + 1406.0, + 1706.0, + 1060.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1700.0, + 627.0, + 1700.0, + 627.0, + 1733.0, + 295.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1700.0, + 691.0, + 1700.0, + 691.0, + 1733.0, + 648.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1700.0, + 1405.0, + 1700.0, + 1405.0, + 1733.0, + 709.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1728.0, + 676.0, + 1728.0, + 676.0, + 1760.0, + 294.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1728.0, + 1279.0, + 1728.0, + 1279.0, + 1760.0, + 696.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1728.0, + 1404.0, + 1728.0, + 1404.0, + 1760.0, + 1297.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1756.0, + 985.0, + 1756.0, + 985.0, + 1789.0, + 296.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1072.0, + 831.0, + 1072.0, + 831.0, + 1108.0, + 295.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 845.0, + 1248.0, + 845.0, + 1248.0, + 885.0, + 293.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 475.0, + 863.0, + 475.0, + 863.0, + 500.0, + 689.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1110.0, + 1406.0, + 1110.0, + 1406.0, + 1147.0, + 293.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1141.0, + 1030.0, + 1141.0, + 1030.0, + 1177.0, + 293.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1141.0, + 1406.0, + 1141.0, + 1406.0, + 1177.0, + 1066.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1170.0, + 343.0, + 1170.0, + 343.0, + 1209.0, + 293.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1170.0, + 925.0, + 1170.0, + 925.0, + 1209.0, + 376.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1170.0, + 1081.0, + 1170.0, + 1081.0, + 1209.0, + 1070.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1170.0, + 1167.0, + 1170.0, + 1167.0, + 1209.0, + 1117.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1170.0, + 1408.0, + 1170.0, + 1408.0, + 1209.0, + 1199.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1203.0, + 1409.0, + 1203.0, + 1409.0, + 1238.0, + 294.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1233.0, + 1049.0, + 1233.0, + 1049.0, + 1269.0, + 293.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1233.0, + 1406.0, + 1233.0, + 1406.0, + 1269.0, + 1150.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1265.0, + 662.0, + 1265.0, + 662.0, + 1298.0, + 293.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 891.0, + 1010.0, + 891.0, + 1010.0, + 936.0, + 292.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 891.0, + 1173.0, + 891.0, + 1173.0, + 936.0, + 1118.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 891.0, + 1406.0, + 891.0, + 1406.0, + 936.0, + 1308.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 926.0, + 1406.0, + 926.0, + 1406.0, + 966.0, + 292.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 960.0, + 1403.0, + 960.0, + 1403.0, + 994.0, + 296.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 990.0, + 1406.0, + 990.0, + 1406.0, + 1024.0, + 296.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1020.0, + 394.0, + 1020.0, + 394.0, + 1061.0, + 294.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1020.0, + 552.0, + 1020.0, + 552.0, + 1061.0, + 502.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1020.0, + 981.0, + 1020.0, + 981.0, + 1061.0, + 687.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 682.0, + 1407.0, + 682.0, + 1407.0, + 715.0, + 296.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 711.0, + 1166.0, + 711.0, + 1166.0, + 747.0, + 294.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 711.0, + 1406.0, + 711.0, + 1406.0, + 747.0, + 1194.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 739.0, + 716.0, + 739.0, + 716.0, + 779.0, + 294.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 739.0, + 1406.0, + 739.0, + 1406.0, + 779.0, + 745.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 771.0, + 1153.0, + 771.0, + 1153.0, + 809.0, + 295.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 771.0, + 1404.0, + 771.0, + 1404.0, + 809.0, + 1179.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 803.0, + 1409.0, + 803.0, + 1409.0, + 839.0, + 295.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 600.0, + 1407.0, + 600.0, + 1407.0, + 642.0, + 293.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 634.0, + 979.0, + 634.0, + 979.0, + 670.0, + 296.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 634.0, + 1371.0, + 634.0, + 1371.0, + 670.0, + 1116.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 522.0, + 1405.0, + 522.0, + 1405.0, + 564.0, + 293.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 557.0, + 1036.0, + 557.0, + 1036.0, + 593.0, + 296.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1800.0, + 858.0, + 1800.0, + 858.0, + 1841.0, + 294.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1800.0, + 1137.0, + 1800.0, + 1137.0, + 1841.0, + 893.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1326.0, + 1800.0, + 1336.0, + 1800.0, + 1336.0, + 1841.0, + 1326.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1800.0, + 1406.0, + 1800.0, + 1406.0, + 1841.0, + 1356.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1835.0, + 1405.0, + 1835.0, + 1405.0, + 1872.0, + 321.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1866.0, + 608.0, + 1866.0, + 608.0, + 1905.0, + 321.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1866.0, + 916.0, + 1866.0, + 916.0, + 1905.0, + 645.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1933.0, + 400.0, + 1933.0, + 400.0, + 1974.0, + 320.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1933.0, + 818.0, + 1933.0, + 818.0, + 1974.0, + 437.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1933.0, + 1281.0, + 1933.0, + 1281.0, + 1974.0, + 846.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1962.0, + 324.0, + 1962.0, + 324.0, + 2014.0, + 319.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1962.0, + 487.0, + 1962.0, + 487.0, + 2014.0, + 426.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1962.0, + 794.0, + 1962.0, + 794.0, + 2014.0, + 516.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1962.0, + 1104.0, + 1962.0, + 1104.0, + 2014.0, + 833.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1962.0, + 1212.0, + 1962.0, + 1212.0, + 2014.0, + 1140.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 1962.0, + 1408.0, + 1962.0, + 1408.0, + 2014.0, + 1327.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1989.0, + 354.0, + 1989.0, + 354.0, + 2047.0, + 315.0, + 2047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1989.0, + 1192.0, + 1989.0, + 1192.0, + 2047.0, + 814.0, + 2047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1896.0, + 1278.0, + 1896.0, + 1278.0, + 1947.0, + 1274.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 845.0, + 1248.0, + 845.0, + 1248.0, + 885.0, + 293.0, + 885.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1677, + 1405, + 1677, + 1405, + 1908, + 296, + 1908 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 296, + 824, + 1405, + 824, + 1405, + 1024, + 296, + 1024 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 296, + 1212, + 1406, + 1212, + 1406, + 1336, + 296, + 1336 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 296, + 1919, + 1404, + 1919, + 1404, + 2046, + 296, + 2046 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 355, + 215, + 1344, + 215, + 1344, + 498, + 355, + 498 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 297, + 1399, + 1405, + 1399, + 1405, + 1503, + 297, + 1503 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 297, + 1518, + 1406, + 1518, + 1406, + 1631, + 297, + 1631 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 296, + 1037, + 1404, + 1037, + 1404, + 1103, + 296, + 1103 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 299, + 654, + 1402, + 654, + 1402, + 716, + 299, + 716 + ], + "score": 0.935 + }, + { + "category_id": 4, + "poly": [ + 296, + 519, + 1405, + 519, + 1405, + 640, + 296, + 640 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 300, + 1156, + 718, + 1156, + 718, + 1189, + 300, + 1189 + ], + "score": 0.917 + }, + { + "category_id": 8, + "poly": [ + 457, + 1637, + 1240, + 1637, + 1240, + 1675, + 457, + 1675 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.902 + }, + { + "category_id": 8, + "poly": [ + 348, + 1341, + 1333, + 1341, + 1333, + 1394, + 348, + 1394 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1642, + 1400, + 1642, + 1400, + 1671, + 1366, + 1671 + ], + "score": 0.841 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.8 + }, + { + "category_id": 1, + "poly": [ + 325, + 764, + 765, + 764, + 765, + 802, + 325, + 802 + ], + "score": 0.598 + }, + { + "category_id": 8, + "poly": [ + 454, + 719, + 1258, + 719, + 1258, + 802, + 454, + 802 + ], + "score": 0.582 + }, + { + "category_id": 8, + "poly": [ + 477, + 720, + 1249, + 720, + 1249, + 763, + 477, + 763 + ], + "score": 0.262 + }, + { + "category_id": 8, + "poly": [ + 446, + 720, + 1270, + 720, + 1270, + 765, + 446, + 765 + ], + "score": 0.105 + }, + { + "category_id": 13, + "poly": [ + 581, + 1842, + 714, + 1842, + 714, + 1876, + 581, + 1876 + ], + "score": 0.94, + "latex": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 421, + 1839, + 528, + 1839, + 528, + 1877, + 421, + 1877 + ], + "score": 0.93, + "latex": "\\tilde { V } ^ { * } \\left( s | U \\right)" + }, + { + "category_id": 13, + "poly": [ + 990, + 827, + 1080, + 827, + 1080, + 861, + 990, + 861 + ], + "score": 0.93, + "latex": "\\mu ^ { w } ( s ^ { w } )" + }, + { + "category_id": 13, + "poly": [ + 1141, + 825, + 1223, + 825, + 1223, + 861, + 1141, + 861 + ], + "score": 0.93, + "latex": "\\mu ^ { h } ( s ^ { h } )" + }, + { + "category_id": 13, + "poly": [ + 462, + 1465, + 529, + 1465, + 529, + 1503, + 462, + 1503 + ], + "score": 0.93, + "latex": "\\tilde { V } ^ { * } ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 887, + 1517, + 993, + 1517, + 993, + 1556, + 887, + 1556 + ], + "score": 0.92, + "latex": "\\tilde { V } ^ { * } \\left( s | U \\right)" + }, + { + "category_id": 13, + "poly": [ + 425, + 1275, + 533, + 1275, + 533, + 1309, + 425, + 1309 + ], + "score": 0.92, + "latex": "\\mu _ { \\theta } \\big ( s _ { g o a l } \\big )" + }, + { + "category_id": 13, + "poly": [ + 988, + 1806, + 1101, + 1806, + 1101, + 1842, + 988, + 1842 + ], + "score": 0.92, + "latex": "\\mu _ { \\theta } \\left( s _ { g o a l } \\right)" + }, + { + "category_id": 13, + "poly": [ + 325, + 684, + 407, + 684, + 407, + 719, + 325, + 719 + ], + "score": 0.92, + "latex": "\\mu ^ { h } ( s ^ { h } )" + }, + { + "category_id": 13, + "poly": [ + 812, + 858, + 952, + 858, + 952, + 889, + 812, + 889 + ], + "score": 0.92, + "latex": "d \\times d \\times d _ { a }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 1677, + 1209, + 1677, + 1209, + 1708, + 1168, + 1708 + ], + "score": 0.92, + "latex": "W _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1556, + 1281, + 1556, + 1281, + 1590, + 1214, + 1590 + ], + "score": 0.91, + "latex": "\\mu _ { \\boldsymbol { \\theta } } ( s )" + }, + { + "category_id": 13, + "poly": [ + 744, + 1595, + 888, + 1595, + 888, + 1628, + 744, + 1628 + ], + "score": 0.91, + "latex": "k = 1 , \\hdots , p" + }, + { + "category_id": 13, + "poly": [ + 399, + 764, + 559, + 764, + 559, + 801, + 399, + 801 + ], + "score": 0.91, + "latex": "\\mu ^ { h } ( s ^ { h } ) \\in \\mathbb { R } ^ { d _ { a } }" + }, + { + "category_id": 13, + "poly": [ + 422, + 1398, + 650, + 1398, + 650, + 1430, + 422, + 1430 + ], + "score": 0.91, + "latex": "\\nu ^ { * ( T ) } \\in \\mathbb { R } ^ { d \\times d \\times d _ { a } \\times p }" + }, + { + "category_id": 13, + "poly": [ + 565, + 827, + 631, + 827, + 631, + 858, + 565, + 858 + ], + "score": 0.91, + "latex": "\\mu _ { \\boldsymbol { \\theta } } ( s )" + }, + { + "category_id": 13, + "poly": [ + 373, + 610, + 425, + 610, + 425, + 641, + 373, + 641 + ], + "score": 0.91, + "latex": "\\psi ( s )" + }, + { + "category_id": 13, + "poly": [ + 473, + 1677, + 524, + 1677, + 524, + 1708, + 473, + 1708 + ], + "score": 0.9, + "latex": "h _ { W _ { 3 } }" + }, + { + "category_id": 13, + "poly": [ + 1270, + 858, + 1403, + 858, + 1403, + 892, + 1270, + 892 + ], + "score": 0.9, + "latex": "\\mu _ { \\theta } ( s ) _ { i j l } \\geqslant" + }, + { + "category_id": 13, + "poly": [ + 1333, + 1519, + 1396, + 1519, + 1396, + 1551, + 1333, + 1551 + ], + "score": 0.9, + "latex": "\\nu ^ { * ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 996, + 548, + 1050, + 548, + 1050, + 578, + 996, + 578 + ], + "score": 0.9, + "latex": "\\nu ^ { * ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1708, + 523, + 1708, + 523, + 1741, + 298, + 1741 + ], + "score": 0.9, + "latex": "\\mathcal { N } ( h _ { W _ { 3 } } \\left( \\psi ( s ) \\right) , \\sigma ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 375, + 1742, + 508, + 1742, + 508, + 1774, + 375, + 1774 + ], + "score": 0.9, + "latex": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 397, + 891, + 606, + 891, + 606, + 927, + 397, + 927 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\sum _ { i j l } \\mu _ { \\theta } ( s ) _ { i j l } = 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 756, + 1738, + 820, + 1738, + 820, + 1769, + 756, + 1769 + ], + "score": 0.9, + "latex": "\\nu ^ { * ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 725, + 1433, + 788, + 1433, + 788, + 1464, + 725, + 1464 + ], + "score": 0.9, + "latex": "\\nu ^ { * ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1678, + 422, + 1678, + 422, + 1709, + 371, + 1709 + ], + "score": 0.9, + "latex": "h _ { W _ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1951, + 535, + 1951, + 535, + 1984, + 296, + 1984 + ], + "score": 0.9, + "latex": "( W _ { 0 } , W _ { 1 } , W _ { 2 } , W _ { 3 } , \\theta )" + }, + { + "category_id": 13, + "poly": [ + 818, + 1402, + 859, + 1402, + 859, + 1433, + 818, + 1433 + ], + "score": 0.9, + "latex": "W _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1333, + 1553, + 1395, + 1553, + 1395, + 1585, + 1333, + 1585 + ], + "score": 0.89, + "latex": "\\nu ^ { * ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 703, + 1806, + 767, + 1806, + 767, + 1835, + 703, + 1835 + ], + "score": 0.89, + "latex": "\\nu ^ { * ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 1075, + 1678, + 1117, + 1678, + 1117, + 1708, + 1075, + 1708 + ], + "score": 0.89, + "latex": "W _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 872, + 1279, + 930, + 1279, + 930, + 1308, + 872, + 1308 + ], + "score": 0.89, + "latex": "s _ { g o a l }" + }, + { + "category_id": 13, + "poly": [ + 1112, + 1473, + 1171, + 1473, + 1171, + 1504, + 1112, + 1504 + ], + "score": 0.89, + "latex": "s _ { g o a l }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1070, + 363, + 1070, + 363, + 1099, + 298, + 1099 + ], + "score": 0.89, + "latex": "d \\times d" + }, + { + "category_id": 13, + "poly": [ + 345, + 1587, + 694, + 1587, + 694, + 1634, + 345, + 1634 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\psi \\left( s \\right) _ { k } = \\sum _ { i j l } \\nu _ { i j l k } ^ { * ( T ) } \\cdot \\mu _ { \\theta } \\left( s \\right) _ { i j l } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 725, + 1402, + 767, + 1402, + 767, + 1432, + 725, + 1432 + ], + "score": 0.88, + "latex": "W _ { 0 }" + }, + { + "category_id": 14, + "poly": [ + 364, + 1338, + 1332, + 1338, + 1332, + 1396, + 364, + 1396 + ], + "score": 0.88, + "latex": "\\nu ^ { * ( t ) } = \\operatorname* { m i n } \\left( W _ { 1 } \\oplus \\left[ \\nu ^ { * ( t - 1 ) } , \\tilde { R } \\right] \\right) , \\quad \\mathrm { w i t h } \\quad \\left( \\nu ^ { * ( 0 ) } , \\tilde { R } \\right) = \\sigma \\left( W _ { 0 } \\oplus \\left[ \\mu _ { \\theta } \\big ( s _ { g o a l } \\big ) , \\mathtt { m a p l } \\right) \\right) ," + }, + { + "category_id": 13, + "poly": [ + 1297, + 892, + 1327, + 892, + 1327, + 919, + 1297, + 919 + ], + "score": 0.88, + "latex": "d _ { a }" + }, + { + "category_id": 13, + "poly": [ + 505, + 687, + 536, + 687, + 536, + 716, + 505, + 716 + ], + "score": 0.88, + "latex": "k _ { h }" + }, + { + "category_id": 13, + "poly": [ + 297, + 857, + 663, + 857, + 663, + 890, + 297, + 890 + ], + "score": 0.88, + "latex": "\\mu _ { \\boldsymbol \\theta } ( s ) _ { i j l } = \\bar { \\mu } ^ { w } ( s ^ { w } ) _ { i j } \\cdot \\bar { \\mu } ^ { h } ( \\bar { s } ^ { h } ) _ { l }" + }, + { + "category_id": 13, + "poly": [ + 710, + 1921, + 791, + 1921, + 791, + 1954, + 710, + 1954 + ], + "score": 0.87, + "latex": "\\mathsf { a l g } \\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 976, + 1706, + 1008, + 1706, + 1008, + 1736, + 976, + 1736 + ], + "score": 0.87, + "latex": "\\sigma ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 565, + 1876, + 690, + 1876, + 690, + 1904, + 565, + 1904 + ], + "score": 0.87, + "latex": "1 0 \\times - 1 0 0 \\times" + }, + { + "category_id": 13, + "poly": [ + 961, + 923, + 991, + 923, + 991, + 954, + 961, + 954 + ], + "score": 0.87, + "latex": "s ^ { h }" + }, + { + "category_id": 13, + "poly": [ + 552, + 607, + 586, + 607, + 586, + 636, + 552, + 636 + ], + "score": 0.86, + "latex": "\\tilde { V } ^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 612, + 764, + 757, + 764, + 757, + 802, + 612, + 802 + ], + "score": 0.86, + "latex": "f _ { 0 } ^ { h } ( s ^ { h } ) = s ^ { h }" + }, + { + "category_id": 13, + "poly": [ + 633, + 611, + 662, + 611, + 662, + 636, + 633, + 636 + ], + "score": 0.86, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1098, + 547, + 1121, + 547, + 1121, + 578, + 1098, + 578 + ], + "score": 0.85, + "latex": "\\bar { \\tilde { \\boldsymbol { R } } }" + }, + { + "category_id": 13, + "poly": [ + 1326, + 1921, + 1403, + 1921, + 1403, + 1952, + 1326, + 1952 + ], + "score": 0.83, + "latex": "W =" + }, + { + "category_id": 13, + "poly": [ + 1072, + 1276, + 1096, + 1276, + 1096, + 1302, + 1072, + 1302 + ], + "score": 0.82, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1196, + 931, + 1217, + 931, + 1217, + 959, + 1196, + 959 + ], + "score": 0.82, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 1027, + 1470, + 1052, + 1470, + 1052, + 1497, + 1027, + 1497 + ], + "score": 0.81, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1071, + 1118, + 1071, + 1118, + 1097, + 1099, + 1097 + ], + "score": 0.79, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 777, + 957, + 1131, + 957, + 1131, + 997, + 777, + 997 + ], + "score": 0.76, + "latex": "\\theta : = ( \\{ \\theta _ { i } ^ { w } \\} _ { i = 0 } ^ { k _ { w } - 1 } , \\{ \\theta _ { i } ^ { h } , b _ { i } \\} _ { i = 0 } ^ { k _ { h } - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 427, + 968, + 443, + 968, + 443, + 989, + 427, + 989 + ], + "score": 0.76, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1563, + 1144, + 1563, + 1144, + 1584, + 1127, + 1584 + ], + "score": 0.74, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 881, + 1851, + 896, + 1851, + 896, + 1871, + 881, + 1871 + ], + "score": 0.74, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 460, + 1635, + 739, + 1635, + 739, + 1675, + 460, + 1675 + ], + "score": 0.73, + "latex": "\\tilde { V } ^ { \\ast } \\left( s | U \\right) = h _ { W _ { 2 } } \\left( \\psi ( s ) \\right)" + }, + { + "category_id": 14, + "poly": [ + 465, + 719, + 1256, + 719, + 1256, + 762, + 465, + 762 + ], + "score": 0.71, + "latex": "\\mu ^ { h } ( s ^ { h } ) = \\mathsf { s o f t m a x } ( f _ { k _ { h } } ^ { h } ( s ^ { h } ) ) , \\qquad f _ { i + 1 } ^ { h } ( s ^ { h } ) = \\mathtt { r e l u } ( \\theta _ { i } ^ { h } f _ { i } ^ { h } ( s ^ { h } ) + b _ { i } ) ," + }, + { + "category_id": 14, + "poly": [ + 457, + 1637, + 1238, + 1637, + 1238, + 1675, + 457, + 1675 + ], + "score": 0.7, + "latex": "\\tilde { V } ^ { \\ast } \\left( s | U \\right) = h _ { W _ { 2 } } \\left( \\psi ( s ) \\right) , \\quad \\mathrm { a n d } \\quad \\tilde { \\pi } ^ { \\ast } \\left( s ^ { \\prime } | s , U \\right) = \\mathcal { N } \\left( h _ { W _ { 3 } } \\left( \\psi ( s ) \\right) , \\sigma ^ { 2 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 814, + 1529, + 830, + 1529, + 830, + 1551, + 814, + 1551 + ], + "score": 0.64, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 339, + 1772, + 896, + 1772, + 896, + 1807, + 339, + 1807 + ], + "score": 0.62, + "latex": "\\begin{array} { r } { \\pi ^ { * } \\left( s ^ { \\prime } | s , U \\right) = \\mathrm { a r g m i n } _ { s ^ { \\prime } \\in S } c \\left( \\left[ s , s ^ { \\prime } \\right] \\right) + V ^ { * } \\left( s ^ { \\prime } | U \\right. } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 974, + 956, + 1131, + 956, + 1131, + 996, + 974, + 996 + ], + "score": 0.42, + "latex": "\\{ \\theta _ { i } ^ { h } , b _ { i } \\} _ { i = 0 } ^ { k _ { h } - 1 } )" + }, + { + "category_id": 14, + "poly": [ + 364, + 1338, + 766, + 1338, + 766, + 1396, + 364, + 1396 + ], + "score": 0.32, + "latex": "\\nu ^ { * ( t ) } = \\operatorname* { m i n } \\left( W _ { 1 } \\oplus \\left[ \\nu ^ { * ( t - 1 ) } , \\tilde { R } \\right] \\right) ," + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 222.0, + 1250.0, + 222.0, + 1250.0, + 246.0, + 1114.0, + 246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 238.0, + 539.0, + 238.0, + 539.0, + 267.0, + 355.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 240.0, + 787.0, + 240.0, + 787.0, + 265.0, + 678.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 284.0, + 640.0, + 284.0, + 640.0, + 337.0, + 553.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 306.0, + 684.0, + 306.0, + 684.0, + 315.0, + 673.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 294.0, + 877.0, + 294.0, + 877.0, + 323.0, + 793.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 280.0, + 1034.0, + 280.0, + 1034.0, + 312.0, + 979.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 277.0, + 1099.0, + 277.0, + 1099.0, + 305.0, + 1043.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 310.0, + 1224.0, + 310.0, + 1224.0, + 339.0, + 1143.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 305.0, + 1265.0, + 305.0, + 1265.0, + 316.0, + 1256.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 324.0, + 1084.0, + 324.0, + 1084.0, + 361.0, + 1053.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 371.0, + 530.0, + 371.0, + 530.0, + 404.0, + 444.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 374.0, + 787.0, + 374.0, + 787.0, + 400.0, + 676.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 386.0, + 1196.0, + 386.0, + 1196.0, + 444.0, + 1082.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 418.0, + 641.0, + 418.0, + 641.0, + 472.0, + 553.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 433.0, + 943.0, + 433.0, + 943.0, + 476.0, + 785.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 416.0, + 1059.0, + 416.0, + 1059.0, + 466.0, + 974.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 440.0, + 1263.0, + 440.0, + 1263.0, + 496.0, + 1079.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 471.0, + 982.0, + 471.0, + 982.0, + 495.0, + 786.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.25, + 286.5, + 1222.25, + 286.5, + 1222.25, + 315.5, + 1139.25, + 315.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 289.5, + 1333.0, + 289.5, + 1333.0, + 327.5, + 1257.0, + 327.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 519.0, + 1403.0, + 519.0, + 1403.0, + 553.0, + 296.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 541.0, + 995.0, + 541.0, + 995.0, + 592.0, + 290.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 541.0, + 1097.0, + 541.0, + 1097.0, + 592.0, + 1051.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 541.0, + 1407.0, + 541.0, + 1407.0, + 592.0, + 1122.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 578.0, + 1405.0, + 578.0, + 1405.0, + 613.0, + 295.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 609.0, + 372.0, + 609.0, + 372.0, + 642.0, + 295.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 609.0, + 551.0, + 609.0, + 551.0, + 642.0, + 426.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 609.0, + 632.0, + 609.0, + 632.0, + 642.0, + 587.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 609.0, + 672.0, + 609.0, + 672.0, + 642.0, + 663.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1153.0, + 723.0, + 1153.0, + 723.0, + 1191.0, + 295.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1675.0, + 370.0, + 1675.0, + 370.0, + 1711.0, + 295.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1675.0, + 472.0, + 1675.0, + 472.0, + 1711.0, + 423.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1675.0, + 1074.0, + 1675.0, + 1074.0, + 1711.0, + 525.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1675.0, + 1167.0, + 1675.0, + 1167.0, + 1711.0, + 1118.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1675.0, + 1405.0, + 1675.0, + 1405.0, + 1711.0, + 1210.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1705.0, + 297.0, + 1705.0, + 297.0, + 1745.0, + 293.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1705.0, + 975.0, + 1705.0, + 975.0, + 1745.0, + 524.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1705.0, + 1406.0, + 1705.0, + 1406.0, + 1745.0, + 1009.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1734.0, + 374.0, + 1734.0, + 374.0, + 1778.0, + 290.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1734.0, + 755.0, + 1734.0, + 755.0, + 1778.0, + 509.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1734.0, + 1407.0, + 1734.0, + 1407.0, + 1778.0, + 821.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1769.0, + 338.0, + 1769.0, + 338.0, + 1808.0, + 293.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1769.0, + 1407.0, + 1769.0, + 1407.0, + 1808.0, + 897.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1802.0, + 702.0, + 1802.0, + 702.0, + 1845.0, + 291.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1802.0, + 987.0, + 1802.0, + 987.0, + 1845.0, + 768.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1802.0, + 1408.0, + 1802.0, + 1408.0, + 1845.0, + 1102.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1840.0, + 420.0, + 1840.0, + 420.0, + 1880.0, + 294.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1840.0, + 580.0, + 1840.0, + 580.0, + 1880.0, + 529.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1840.0, + 880.0, + 1840.0, + 880.0, + 1880.0, + 715.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1840.0, + 1406.0, + 1840.0, + 1406.0, + 1880.0, + 897.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1874.0, + 564.0, + 1874.0, + 564.0, + 1910.0, + 295.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1874.0, + 898.0, + 1874.0, + 898.0, + 1910.0, + 691.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 821.0, + 564.0, + 821.0, + 564.0, + 864.0, + 293.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 821.0, + 989.0, + 821.0, + 989.0, + 864.0, + 632.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 821.0, + 1140.0, + 821.0, + 1140.0, + 864.0, + 1081.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 821.0, + 1408.0, + 821.0, + 1408.0, + 864.0, + 1224.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 848.0, + 296.0, + 848.0, + 296.0, + 901.0, + 289.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 848.0, + 811.0, + 848.0, + 811.0, + 901.0, + 664.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 848.0, + 1269.0, + 848.0, + 1269.0, + 901.0, + 953.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 848.0, + 1410.0, + 848.0, + 1410.0, + 901.0, + 1404.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 885.0, + 396.0, + 885.0, + 396.0, + 928.0, + 290.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 885.0, + 1296.0, + 885.0, + 1296.0, + 928.0, + 607.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 885.0, + 1407.0, + 885.0, + 1407.0, + 928.0, + 1328.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 921.0, + 960.0, + 921.0, + 960.0, + 963.0, + 293.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 921.0, + 1195.0, + 921.0, + 1195.0, + 963.0, + 992.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 921.0, + 1405.0, + 921.0, + 1405.0, + 963.0, + 1218.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 940.0, + 426.0, + 940.0, + 426.0, + 1015.0, + 283.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 940.0, + 776.0, + 940.0, + 776.0, + 1015.0, + 444.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 940.0, + 1416.0, + 940.0, + 1416.0, + 1015.0, + 1132.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 990.0, + 393.0, + 990.0, + 393.0, + 1024.0, + 293.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1212.0, + 1407.0, + 1212.0, + 1407.0, + 1248.0, + 294.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1245.0, + 1404.0, + 1245.0, + 1404.0, + 1277.0, + 295.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1273.0, + 424.0, + 1273.0, + 424.0, + 1312.0, + 294.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1273.0, + 871.0, + 1273.0, + 871.0, + 1312.0, + 534.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1273.0, + 1071.0, + 1273.0, + 1071.0, + 1312.0, + 931.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 1273.0, + 1407.0, + 1273.0, + 1407.0, + 1312.0, + 1097.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1303.0, + 1007.0, + 1303.0, + 1007.0, + 1343.0, + 295.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1917.0, + 709.0, + 1917.0, + 709.0, + 1957.0, + 291.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1917.0, + 1325.0, + 1917.0, + 1325.0, + 1957.0, + 792.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 1951.0, + 1409.0, + 1951.0, + 1409.0, + 1987.0, + 536.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1980.0, + 1405.0, + 1980.0, + 1405.0, + 2018.0, + 294.0, + 2018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2013.0, + 617.0, + 2013.0, + 617.0, + 2046.0, + 294.0, + 2046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1392.0, + 421.0, + 1392.0, + 421.0, + 1438.0, + 292.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1392.0, + 724.0, + 1392.0, + 724.0, + 1438.0, + 651.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1392.0, + 817.0, + 1392.0, + 817.0, + 1438.0, + 768.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1392.0, + 1406.0, + 1392.0, + 1406.0, + 1438.0, + 860.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1429.0, + 724.0, + 1429.0, + 724.0, + 1473.0, + 291.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1429.0, + 1406.0, + 1429.0, + 1406.0, + 1473.0, + 789.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1464.0, + 461.0, + 1464.0, + 461.0, + 1508.0, + 292.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1464.0, + 1026.0, + 1464.0, + 1026.0, + 1508.0, + 530.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1464.0, + 1111.0, + 1464.0, + 1111.0, + 1508.0, + 1053.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1464.0, + 1261.0, + 1464.0, + 1261.0, + 1508.0, + 1172.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1516.0, + 813.0, + 1516.0, + 813.0, + 1557.0, + 290.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1516.0, + 886.0, + 1516.0, + 886.0, + 1557.0, + 831.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1516.0, + 1332.0, + 1516.0, + 1332.0, + 1557.0, + 994.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1516.0, + 1405.0, + 1516.0, + 1405.0, + 1557.0, + 1397.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1549.0, + 1126.0, + 1549.0, + 1126.0, + 1593.0, + 290.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1549.0, + 1213.0, + 1549.0, + 1213.0, + 1593.0, + 1145.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 1549.0, + 1332.0, + 1549.0, + 1332.0, + 1593.0, + 1282.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1549.0, + 1409.0, + 1549.0, + 1409.0, + 1593.0, + 1396.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1584.0, + 344.0, + 1584.0, + 344.0, + 1642.0, + 290.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1584.0, + 743.0, + 1584.0, + 743.0, + 1642.0, + 695.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 1584.0, + 1105.0, + 1584.0, + 1105.0, + 1642.0, + 889.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1036.0, + 1406.0, + 1036.0, + 1406.0, + 1074.0, + 295.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1069.0, + 1098.0, + 1069.0, + 1098.0, + 1105.0, + 364.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1069.0, + 1351.0, + 1069.0, + 1351.0, + 1105.0, + 1119.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 653.0, + 1406.0, + 653.0, + 1406.0, + 689.0, + 298.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 682.0, + 504.0, + 682.0, + 504.0, + 722.0, + 408.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 682.0, + 1170.0, + 682.0, + 1170.0, + 722.0, + 537.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 759.0, + 398.0, + 759.0, + 398.0, + 804.0, + 321.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 759.0, + 611.0, + 759.0, + 611.0, + 804.0, + 560.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 759.0, + 769.0, + 759.0, + 769.0, + 804.0, + 758.0, + 804.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 270, + 1404, + 270, + 1404, + 491, + 297, + 491 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1505, + 1404, + 1505, + 1404, + 1659, + 298, + 1659 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 505, + 1403, + 505, + 1403, + 698, + 298, + 698 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 812, + 819, + 812, + 819, + 1310, + 299, + 1310 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1742, + 1404, + 1742, + 1404, + 1957, + 298, + 1957 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 824, + 856, + 1372, + 856, + 1372, + 1377, + 824, + 1377 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 296, + 1971, + 1403, + 1971, + 1403, + 2035, + 296, + 2035 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 302, + 703, + 1345, + 703, + 1345, + 794, + 302, + 794 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 299, + 1692, + 608, + 1692, + 608, + 1723, + 299, + 1723 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 299, + 1451, + 558, + 1451, + 558, + 1485, + 299, + 1485 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 300, + 230, + 779, + 230, + 779, + 261, + 300, + 261 + ], + "score": 0.856 + }, + { + "category_id": 1, + "poly": [ + 297, + 1326, + 837, + 1326, + 837, + 1387, + 297, + 1387 + ], + "score": 0.852 + }, + { + "category_id": 0, + "poly": [ + 850, + 813, + 1348, + 813, + 1348, + 846, + 850, + 846 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 302, + 1388, + 1089, + 1388, + 1089, + 1419, + 302, + 1419 + ], + "score": 0.803 + }, + { + "category_id": 9, + "poly": [ + 1368, + 731, + 1400, + 731, + 1400, + 760, + 1368, + 760 + ], + "score": 0.766 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.739 + }, + { + "category_id": 13, + "poly": [ + 923, + 596, + 1012, + 596, + 1012, + 632, + 923, + 632 + ], + "score": 0.94, + "latex": "\\{ s ^ { i } \\} _ { i = 1 } ^ { m }" + }, + { + "category_id": 13, + "poly": [ + 835, + 659, + 941, + 659, + 941, + 697, + 835, + 697 + ], + "score": 0.94, + "latex": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}" + }, + { + "category_id": 13, + "poly": [ + 384, + 996, + 491, + 996, + 491, + 1034, + 384, + 1034 + ], + "score": 0.94, + "latex": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}" + }, + { + "category_id": 13, + "poly": [ + 1205, + 394, + 1313, + 394, + 1313, + 432, + 1205, + 432 + ], + "score": 0.93, + "latex": "\\{ \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } \\}" + }, + { + "category_id": 13, + "poly": [ + 1112, + 268, + 1218, + 268, + 1218, + 307, + 1112, + 307 + ], + "score": 0.93, + "latex": "\\tilde { V } ^ { * } \\left( s | U \\right)" + }, + { + "category_id": 13, + "poly": [ + 298, + 1091, + 405, + 1091, + 405, + 1129, + 298, + 1129 + ], + "score": 0.93, + "latex": "\\{ \\tilde { V } ^ { * } , \\bar { \\pi } ^ { * } \\}" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1169, + 1116, + 1169, + 1116, + 1204, + 1024, + 1204 + ], + "score": 0.92, + "latex": "( \\tau _ { j } , U _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 428, + 505, + 661, + 505, + 661, + 544, + 428, + 544 + ], + "score": 0.92, + "latex": "\\mathcal { D } _ { n } = \\{ ( \\mathcal { T } _ { j } , U _ { j } ) \\} _ { j = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 295, + 597, + 454, + 597, + 454, + 631, + 295, + 631 + ], + "score": 0.92, + "latex": "( { \\mathcal { T } } , U ) \\in { \\mathcal { D } } _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1269, + 272, + 1403, + 272, + 1403, + 307, + 1269, + 307 + ], + "score": 0.91, + "latex": "\\tilde { \\pi } ^ { * } \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 429, + 658, + 714, + 658, + 714, + 698, + 429, + 698 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { y ^ { i } : = \\sum _ { l = i } ^ { m - 1 } c ( [ s ^ { l } , s ^ { l + 1 } ] ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 889, + 1077, + 1172, + 1077, + 1172, + 1111, + 889, + 1111 + ], + "score": 0.9, + "latex": "{ \\mathcal { D } } _ { n } \\gets { \\mathcal { D } } _ { n - 1 } \\bar { \\cup } \\{ ( { \\mathcal { T } } , U ) \\}" + }, + { + "category_id": 13, + "poly": [ + 1104, + 945, + 1273, + 945, + 1273, + 978, + 1104, + 978 + ], + "score": 0.9, + "latex": "\\epsilon \\sim \\mathcal { U } n i f [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1267, + 1202, + 1354, + 1202, + 1354, + 1238, + 1267, + 1238 + ], + "score": 0.9, + "latex": "\\{ s ^ { i } \\} _ { i = 1 } ^ { m }" + }, + { + "category_id": 13, + "poly": [ + 932, + 1141, + 1005, + 1141, + 1005, + 1170, + 932, + 1170 + ], + "score": 0.9, + "latex": "j 0" + }, + { + "category_id": 13, + "poly": [ + 1182, + 1171, + 1220, + 1171, + 1220, + 1201, + 1182, + 1201 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { n }" + }, + { + "category_id": 13, + "poly": [ + 953, + 1111, + 1021, + 1111, + 1021, + 1140, + 953, + 1140 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { n - 1 }" + }, + { + "category_id": 13, + "poly": [ + 962, + 886, + 1039, + 886, + 1039, + 913, + 962, + 913 + ], + "score": 0.89, + "latex": "n \\gets 1" + }, + { + "category_id": 13, + "poly": [ + 889, + 945, + 1044, + 945, + 1044, + 978, + 889, + 978 + ], + "score": 0.88, + "latex": "\\mathcal { T } \\mathrm { T S A } ( U )" + }, + { + "category_id": 14, + "poly": [ + 303, + 700, + 1359, + 700, + 1359, + 796, + 303, + 796 + ], + "score": 0.88, + "latex": "\\operatorname* { m i n } _ { W } \\sum _ { ( { \\cal T } , { \\cal U } ) \\in { \\mathcal D } _ { n } } \\ell ( \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } ; { \\cal T } , { \\cal U } ) : = - \\sum _ { { \\mathcal D } _ { n } } \\sum _ { i = 1 } ^ { m - 1 } \\log \\tilde { \\pi } ^ { * } \\left( s ^ { i + 1 } | s ^ { i } \\right) + \\sum _ { i = 1 } ^ { m } ( \\tilde { V } ^ { * } \\left( s ^ { i } \\right) - y ^ { i } ) _ { 2 } ^ { 2 } + \\lambda \\left\\| { \\cal W } \\right\\| ^ { 2 } ." + }, + { + "category_id": 13, + "poly": [ + 428, + 875, + 508, + 875, + 508, + 909, + 428, + 909 + ], + "score": 0.88, + "latex": "\\mathsf { a l g } \\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 857, + 628, + 882, + 628, + 882, + 656, + 857, + 656 + ], + "score": 0.87, + "latex": "s ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 578, + 809, + 617, + 809, + 617, + 843, + 578, + 843 + ], + "score": 0.87, + "latex": "{ \\tilde { V } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1895, + 1284, + 1895, + 1284, + 1928, + 1236, + 1928 + ], + "score": 0.87, + "latex": "c \\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 1286, + 1238, + 1314, + 1238, + 1314, + 1269, + 1286, + 1269 + ], + "score": 0.87, + "latex": "\\tau _ { j }" + }, + { + "category_id": 13, + "poly": [ + 674, + 815, + 707, + 815, + 707, + 843, + 674, + 843 + ], + "score": 0.86, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1037, + 855, + 1073, + 855, + 1073, + 884, + 1037, + 884 + ], + "score": 0.85, + "latex": "\\mathcal { D } _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 608, + 272, + 688, + 272, + 688, + 307, + 608, + 307 + ], + "score": 0.82, + "latex": "\\mathsf { a l g } \\left( \\cdot \\right)" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1343, + 1184, + 1343, + 1184, + 1377, + 1069, + 1377 + ], + "score": 0.82, + "latex": "\\alpha \\in ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 557, + 366, + 622, + 366, + 622, + 395, + 557, + 395 + ], + "score": 0.81, + "latex": "1 - \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1375, + 538, + 1401, + 538, + 1401, + 564, + 1375, + 564 + ], + "score": 0.81, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 743, + 511, + 763, + 511, + 763, + 534, + 743, + 534 + ], + "score": 0.8, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 1202, + 916, + 1227, + 916, + 1227, + 942, + 1202, + 942 + ], + "score": 0.77, + "latex": "U" + }, + { + "category_id": 13, + "poly": [ + 1259, + 598, + 1287, + 598, + 1287, + 626, + 1259, + 626 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1038, + 1141, + 1061, + 1141, + 1061, + 1166, + 1038, + 1166 + ], + "score": 0.72, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 1301, + 602, + 1329, + 602, + 1329, + 626, + 1301, + 626 + ], + "score": 0.72, + "latex": "\\mathrm { ~ m ~ }" + }, + { + "category_id": 14, + "poly": [ + 944, + 1296, + 1347, + 1296, + 1347, + 1337, + 944, + 1337 + ], + "score": 0.71, + "latex": "W W - \\eta \\nabla _ { W } \\ell ( \\tilde { V } ^ { * } , \\tilde { \\pi } ^ { * } ; T _ { j } , U _ { j } ) ;" + }, + { + "category_id": 13, + "poly": [ + 1386, + 436, + 1402, + 436, + 1402, + 457, + 1386, + 457 + ], + "score": 0.7, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1072, + 886, + 1101, + 886, + 1101, + 912, + 1072, + 912 + ], + "score": 0.69, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 862, + 404, + 879, + 404, + 879, + 426, + 862, + 426 + ], + "score": 0.68, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 487, + 371, + 504, + 371, + 504, + 392, + 487, + 392 + ], + "score": 0.65, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 875, + 430, + 943, + 430, + 943, + 459, + 875, + 459 + ], + "score": 0.56, + "latex": "\\mathbf { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 1148, + 1972, + 1210, + 1972, + 1210, + 2002, + 1148, + 2002 + ], + "score": 0.56, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 977, + 1347, + 1057, + 1347, + 1057, + 1374, + 977, + 1374 + ], + "score": 0.55, + "latex": "\\epsilon = \\alpha \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 904, + 986, + 919, + 986, + 919, + 1009, + 904, + 1009 + ], + "score": 0.54, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 823, + 889, + 840, + 889, + 840, + 911, + 823, + 911 + ], + "score": 0.48, + "latex": "^ 2" + }, + { + "category_id": 13, + "poly": [ + 824, + 920, + 841, + 920, + 841, + 941, + 824, + 941 + ], + "score": 0.37, + "latex": "^ 3" + }, + { + "category_id": 13, + "poly": [ + 1088, + 981, + 1177, + 981, + 1177, + 1014, + 1088, + 1014 + ], + "score": 0.34, + "latex": "+ ( 1 - \\epsilon )" + }, + { + "category_id": 13, + "poly": [ + 903, + 1049, + 962, + 1049, + 962, + 1075, + 903, + 1075 + ], + "score": 0.26, + "latex": "\\mathbb { R R T } ^ { * }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1690.0, + 612.0, + 1690.0, + 612.0, + 1726.0, + 295.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1446.0, + 561.0, + 1446.0, + 561.0, + 1491.0, + 292.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 783.0, + 228.0, + 783.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 807.0, + 1355.0, + 807.0, + 1355.0, + 856.0, + 843.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 269.0, + 607.0, + 269.0, + 607.0, + 312.0, + 292.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 269.0, + 1111.0, + 269.0, + 1111.0, + 312.0, + 689.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 269.0, + 1268.0, + 269.0, + 1268.0, + 312.0, + 1219.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 269.0, + 1407.0, + 269.0, + 1407.0, + 312.0, + 1404.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 300.0, + 1406.0, + 300.0, + 1406.0, + 342.0, + 292.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 333.0, + 1404.0, + 333.0, + 1404.0, + 368.0, + 294.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 364.0, + 486.0, + 364.0, + 486.0, + 398.0, + 293.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 364.0, + 556.0, + 364.0, + 556.0, + 398.0, + 505.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 364.0, + 1402.0, + 364.0, + 1402.0, + 398.0, + 623.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 397.0, + 861.0, + 397.0, + 861.0, + 433.0, + 291.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 397.0, + 1204.0, + 397.0, + 1204.0, + 433.0, + 880.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.0, + 397.0, + 1405.0, + 397.0, + 1405.0, + 433.0, + 1314.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 428.0, + 874.0, + 428.0, + 874.0, + 465.0, + 292.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 428.0, + 1385.0, + 428.0, + 1385.0, + 465.0, + 944.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 461.0, + 844.0, + 461.0, + 844.0, + 493.0, + 293.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1504.0, + 1406.0, + 1504.0, + 1406.0, + 1541.0, + 294.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1536.0, + 1406.0, + 1536.0, + 1406.0, + 1573.0, + 294.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1566.0, + 1407.0, + 1566.0, + 1407.0, + 1602.0, + 293.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1597.0, + 1406.0, + 1597.0, + 1406.0, + 1632.0, + 293.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1626.0, + 1277.0, + 1626.0, + 1277.0, + 1663.0, + 294.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 502.0, + 427.0, + 502.0, + 427.0, + 547.0, + 292.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 502.0, + 742.0, + 502.0, + 742.0, + 547.0, + 662.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 502.0, + 1407.0, + 502.0, + 1407.0, + 547.0, + 764.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 536.0, + 1374.0, + 536.0, + 1374.0, + 569.0, + 293.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 566.0, + 1405.0, + 566.0, + 1405.0, + 603.0, + 292.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 588.0, + 294.0, + 588.0, + 294.0, + 642.0, + 288.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 588.0, + 922.0, + 588.0, + 922.0, + 642.0, + 455.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 588.0, + 1258.0, + 588.0, + 1258.0, + 642.0, + 1013.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 588.0, + 1300.0, + 588.0, + 1300.0, + 642.0, + 1288.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 588.0, + 1410.0, + 588.0, + 1410.0, + 642.0, + 1330.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 628.0, + 856.0, + 628.0, + 856.0, + 664.0, + 293.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 628.0, + 1405.0, + 628.0, + 1405.0, + 664.0, + 883.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 643.0, + 428.0, + 643.0, + 428.0, + 713.0, + 284.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 643.0, + 834.0, + 643.0, + 834.0, + 713.0, + 715.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 643.0, + 1225.0, + 643.0, + 1225.0, + 713.0, + 942.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 809.0, + 577.0, + 809.0, + 577.0, + 850.0, + 294.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 809.0, + 673.0, + 809.0, + 673.0, + 850.0, + 618.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 809.0, + 818.0, + 809.0, + 818.0, + 850.0, + 708.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 843.0, + 816.0, + 843.0, + 816.0, + 878.0, + 295.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 875.0, + 427.0, + 875.0, + 427.0, + 910.0, + 296.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 875.0, + 815.0, + 875.0, + 815.0, + 910.0, + 509.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 905.0, + 814.0, + 905.0, + 814.0, + 937.0, + 295.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 935.0, + 816.0, + 935.0, + 816.0, + 969.0, + 293.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 966.0, + 817.0, + 966.0, + 817.0, + 1000.0, + 295.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 997.0, + 383.0, + 997.0, + 383.0, + 1035.0, + 293.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 997.0, + 817.0, + 997.0, + 817.0, + 1035.0, + 492.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1031.0, + 819.0, + 1031.0, + 819.0, + 1064.0, + 295.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1061.0, + 817.0, + 1061.0, + 817.0, + 1094.0, + 295.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1091.0, + 819.0, + 1091.0, + 819.0, + 1131.0, + 406.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1126.0, + 819.0, + 1126.0, + 819.0, + 1159.0, + 295.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1156.0, + 820.0, + 1156.0, + 820.0, + 1190.0, + 294.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1187.0, + 819.0, + 1187.0, + 819.0, + 1219.0, + 294.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1215.0, + 816.0, + 1215.0, + 816.0, + 1251.0, + 293.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1246.0, + 817.0, + 1246.0, + 817.0, + 1284.0, + 294.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1278.0, + 701.0, + 1278.0, + 701.0, + 1313.0, + 295.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 1406.0, + 1741.0, + 1406.0, + 1776.0, + 294.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 1405.0, + 1772.0, + 1405.0, + 1810.0, + 293.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1803.0, + 1408.0, + 1803.0, + 1408.0, + 1838.0, + 292.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1831.0, + 1405.0, + 1831.0, + 1405.0, + 1870.0, + 293.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1864.0, + 1405.0, + 1864.0, + 1405.0, + 1899.0, + 294.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1894.0, + 1235.0, + 1894.0, + 1235.0, + 1928.0, + 294.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1894.0, + 1405.0, + 1894.0, + 1405.0, + 1928.0, + 1285.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1925.0, + 1342.0, + 1925.0, + 1342.0, + 1961.0, + 293.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 850.0, + 1036.0, + 850.0, + 1036.0, + 888.0, + 822.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 850.0, + 1086.0, + 850.0, + 1086.0, + 888.0, + 1074.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 885.0, + 961.0, + 885.0, + 961.0, + 915.0, + 841.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 885.0, + 1071.0, + 885.0, + 1071.0, + 915.0, + 1040.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 885.0, + 1140.0, + 885.0, + 1140.0, + 915.0, + 1102.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 914.0, + 1201.0, + 914.0, + 1201.0, + 947.0, + 887.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 914.0, + 1239.0, + 914.0, + 1239.0, + 947.0, + 1228.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 954.0, + 842.0, + 954.0, + 842.0, + 972.0, + 825.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 943.0, + 1103.0, + 943.0, + 1103.0, + 979.0, + 1045.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 943.0, + 1332.0, + 943.0, + 1332.0, + 979.0, + 1274.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 979.0, + 1087.0, + 979.0, + 1087.0, + 1015.0, + 920.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 979.0, + 1371.0, + 979.0, + 1371.0, + 1015.0, + 1178.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1015.0, + 1134.0, + 1015.0, + 1134.0, + 1051.0, + 902.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1045.0, + 1167.0, + 1045.0, + 1167.0, + 1082.0, + 963.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 1086.0, + 841.0, + 1086.0, + 841.0, + 1106.0, + 825.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1076.0, + 1321.0, + 1076.0, + 1321.0, + 1111.0, + 1173.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1105.0, + 952.0, + 1105.0, + 952.0, + 1145.0, + 901.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 1105.0, + 1035.0, + 1105.0, + 1035.0, + 1145.0, + 1022.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 1148.0, + 842.0, + 1148.0, + 842.0, + 1167.0, + 825.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1138.0, + 931.0, + 1138.0, + 931.0, + 1171.0, + 888.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1138.0, + 1037.0, + 1138.0, + 1037.0, + 1171.0, + 1006.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1138.0, + 1101.0, + 1138.0, + 1101.0, + 1171.0, + 1062.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 1178.0, + 844.0, + 1178.0, + 844.0, + 1233.0, + 825.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1168.0, + 1023.0, + 1168.0, + 1023.0, + 1254.0, + 921.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1168.0, + 1181.0, + 1168.0, + 1181.0, + 1254.0, + 1117.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1168.0, + 1266.0, + 1168.0, + 1266.0, + 1254.0, + 1221.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 1168.0, + 1367.0, + 1168.0, + 1367.0, + 1254.0, + 1355.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 1234.0, + 1285.0, + 1234.0, + 1285.0, + 1270.0, + 943.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 1234.0, + 1329.0, + 1234.0, + 1329.0, + 1270.0, + 1315.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1274.0, + 842.0, + 1274.0, + 842.0, + 1291.0, + 827.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1265.0, + 1149.0, + 1265.0, + 1149.0, + 1297.0, + 932.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 1297.0, + 1352.0, + 1297.0, + 1352.0, + 1339.0, + 1348.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1351.0, + 841.0, + 1351.0, + 841.0, + 1373.0, + 820.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1342.0, + 976.0, + 1342.0, + 976.0, + 1378.0, + 887.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1342.0, + 1068.0, + 1342.0, + 1068.0, + 1378.0, + 1058.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1342.0, + 1195.0, + 1342.0, + 1195.0, + 1378.0, + 1185.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1147.0, + 1971.0, + 1147.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 1211.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1325.0, + 815.0, + 1325.0, + 815.0, + 1358.0, + 295.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1351.0, + 841.0, + 1351.0, + 841.0, + 1390.0, + 294.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1384.0, + 1088.0, + 1384.0, + 1088.0, + 1423.0, + 297.0, + 1423.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1556, + 1404, + 1556, + 1404, + 1806, + 298, + 1806 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 1024, + 1405, + 1024, + 1405, + 1300, + 297, + 1300 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1316, + 1405, + 1316, + 1405, + 1468, + 298, + 1468 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1404, + 1819, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 353, + 658, + 1398, + 658, + 1398, + 913, + 353, + 913 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 320, + 232, + 1378, + 232, + 1378, + 517, + 320, + 517 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 295, + 926, + 1403, + 926, + 1403, + 1010, + 295, + 1010 + ], + "score": 0.945 + }, + { + "category_id": 4, + "poly": [ + 298, + 529, + 1404, + 529, + 1404, + 647, + 298, + 647 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 300, + 1506, + 661, + 1506, + 661, + 1537, + 300, + 1537 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 816, + 76, + 816, + 104, + 298, + 104 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.791 + }, + { + "category_id": 13, + "poly": [ + 779, + 617, + 898, + 617, + 898, + 647, + 779, + 647 + ], + "score": 0.92, + "latex": "\\tilde { \\pi } ^ { * } ( s ^ { \\prime } | s , U )" + }, + { + "category_id": 13, + "poly": [ + 695, + 527, + 730, + 527, + 730, + 558, + 695, + 558 + ], + "score": 0.89, + "latex": "{ \\tilde { V } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 545, + 1708, + 582, + 1708, + 582, + 1741, + 545, + 1741 + ], + "score": 0.89, + "latex": "{ \\tilde { V } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1076, + 586, + 1110, + 586, + 1110, + 616, + 1076, + 616 + ], + "score": 0.87, + "latex": "\\tilde { V } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 634, + 1712, + 667, + 1712, + 667, + 1741, + 634, + 1741 + ], + "score": 0.87, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 775, + 532, + 806, + 532, + 806, + 559, + 775, + 559 + ], + "score": 0.86, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1254, + 480, + 1290, + 480, + 1290, + 511, + 1254, + 511 + ], + "score": 0.85, + "latex": "\\tilde { V } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 428, + 625, + 443, + 625, + 443, + 643, + 428, + 643 + ], + "score": 0.77, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1016, + 1442, + 1031, + 1442, + 1031, + 1465, + 1016, + 1465 + ], + "score": 0.74, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1336, + 485, + 1366, + 485, + 1366, + 511, + 1336, + 511 + ], + "score": 0.7, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 956, + 1377, + 1016, + 1377, + 1016, + 1406, + 956, + 1406 + ], + "score": 0.69, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 346, + 1056, + 407, + 1056, + 407, + 1085, + 346, + 1085 + ], + "score": 0.68, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 836, + 1377, + 904, + 1377, + 904, + 1406, + 836, + 1406 + ], + "score": 0.65, + "latex": "\\mathrm { { R R T ^ { * } } }" + }, + { + "category_id": 13, + "poly": [ + 338, + 2003, + 402, + 2003, + 402, + 2032, + 338, + 2032 + ], + "score": 0.63, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 1117, + 1087, + 1176, + 1087, + 1176, + 1116, + 1117, + 1116 + ], + "score": 0.58, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 403, + 1087, + 470, + 1087, + 470, + 1116, + 403, + 1116 + ], + "score": 0.36, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 296, + 982, + 360, + 982, + 360, + 1009, + 296, + 1009 + ], + "score": 0.35, + "latex": "{ \\mathrm { R R T } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1329, + 1178, + 1398, + 1178, + 1398, + 1208, + 1329, + 1208 + ], + "score": 0.28, + "latex": "\\mathbf { R R T ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 880.0, + 629.0, + 880.0, + 629.0, + 916.0, + 348.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 882.0, + 995.0, + 882.0, + 995.0, + 916.0, + 730.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 882.0, + 1403.0, + 882.0, + 1403.0, + 916.0, + 1060.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 291.0, + 954.0, + 291.0, + 954.0, + 306.0, + 937.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 407.0, + 416.0, + 407.0, + 416.0, + 429.0, + 380.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 425.0, + 406.0, + 425.0, + 406.0, + 435.0, + 393.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 460.0, + 1354.0, + 460.0, + 1354.0, + 479.0, + 1161.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 483.0, + 1119.0, + 483.0, + 1119.0, + 518.0, + 319.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 483.0, + 1253.0, + 483.0, + 1253.0, + 514.0, + 1134.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 483.0, + 1335.0, + 483.0, + 1335.0, + 514.0, + 1291.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 305.0, + 761.0, + 305.0, + 761.0, + 313.0, + 744.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 925.0, + 1405.0, + 925.0, + 1405.0, + 958.0, + 295.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 955.0, + 1404.0, + 955.0, + 1404.0, + 984.0, + 295.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 979.0, + 772.0, + 979.0, + 772.0, + 1013.0, + 361.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 531.0, + 694.0, + 531.0, + 694.0, + 565.0, + 296.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 531.0, + 774.0, + 531.0, + 774.0, + 565.0, + 731.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 531.0, + 1405.0, + 531.0, + 1405.0, + 565.0, + 807.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 561.0, + 1404.0, + 561.0, + 1404.0, + 591.0, + 296.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 589.0, + 1075.0, + 589.0, + 1075.0, + 623.0, + 296.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 589.0, + 1404.0, + 589.0, + 1404.0, + 623.0, + 1111.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 617.0, + 427.0, + 617.0, + 427.0, + 651.0, + 293.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 617.0, + 778.0, + 617.0, + 778.0, + 651.0, + 444.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 617.0, + 909.0, + 617.0, + 909.0, + 651.0, + 899.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1503.0, + 666.0, + 1503.0, + 666.0, + 1542.0, + 293.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1558.0, + 1406.0, + 1558.0, + 1406.0, + 1589.0, + 298.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1585.0, + 1405.0, + 1585.0, + 1405.0, + 1620.0, + 293.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1618.0, + 1406.0, + 1618.0, + 1406.0, + 1652.0, + 294.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1647.0, + 1407.0, + 1647.0, + 1407.0, + 1682.0, + 292.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1677.0, + 1407.0, + 1677.0, + 1407.0, + 1714.0, + 292.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1709.0, + 544.0, + 1709.0, + 544.0, + 1750.0, + 292.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1709.0, + 633.0, + 1709.0, + 633.0, + 1750.0, + 583.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1709.0, + 1405.0, + 1709.0, + 1405.0, + 1750.0, + 668.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1776.0, + 294.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1775.0, + 1318.0, + 1775.0, + 1318.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1024.0, + 1402.0, + 1024.0, + 1402.0, + 1060.0, + 294.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1054.0, + 345.0, + 1054.0, + 345.0, + 1092.0, + 294.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1054.0, + 1405.0, + 1054.0, + 1405.0, + 1092.0, + 408.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1085.0, + 402.0, + 1085.0, + 402.0, + 1123.0, + 292.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1085.0, + 1116.0, + 1085.0, + 1116.0, + 1123.0, + 471.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1085.0, + 1406.0, + 1085.0, + 1406.0, + 1123.0, + 1177.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1116.0, + 1403.0, + 1116.0, + 1403.0, + 1153.0, + 294.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1146.0, + 1407.0, + 1146.0, + 1407.0, + 1187.0, + 292.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1175.0, + 1328.0, + 1175.0, + 1328.0, + 1214.0, + 292.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1175.0, + 1406.0, + 1175.0, + 1406.0, + 1214.0, + 1399.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1208.0, + 1407.0, + 1208.0, + 1407.0, + 1244.0, + 294.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1274.0, + 295.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1268.0, + 1408.0, + 1268.0, + 1408.0, + 1303.0, + 291.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1313.0, + 1406.0, + 1313.0, + 1406.0, + 1351.0, + 293.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1345.0, + 1405.0, + 1345.0, + 1405.0, + 1380.0, + 295.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1375.0, + 835.0, + 1375.0, + 835.0, + 1412.0, + 293.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1375.0, + 955.0, + 1375.0, + 955.0, + 1412.0, + 905.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1375.0, + 1405.0, + 1375.0, + 1405.0, + 1412.0, + 1017.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1406.0, + 1408.0, + 1406.0, + 1408.0, + 1443.0, + 296.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1437.0, + 1015.0, + 1437.0, + 1015.0, + 1473.0, + 295.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1437.0, + 1100.0, + 1437.0, + 1100.0, + 1473.0, + 1032.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1819.0, + 1406.0, + 1819.0, + 1406.0, + 1853.0, + 293.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1883.0, + 293.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1408.0, + 1912.0, + 1408.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 1408.0, + 1939.0, + 1408.0, + 1978.0, + 292.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1401.0, + 1973.0, + 1401.0, + 2004.0, + 294.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 337.0, + 2002.0, + 337.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 2002.0, + 1408.0, + 2002.0, + 1408.0, + 2037.0, + 403.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1438, + 1404, + 1438, + 1404, + 1683, + 298, + 1683 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1698, + 1405, + 1698, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1116, + 1404, + 1116, + 1404, + 1300, + 298, + 1300 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 297, + 226, + 1397, + 226, + 1397, + 893, + 297, + 893 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 297, + 912, + 1405, + 912, + 1405, + 1024, + 297, + 1024 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 300, + 1316, + 1399, + 1316, + 1399, + 1377, + 300, + 1377 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 298, + 1038, + 1397, + 1038, + 1397, + 1101, + 298, + 1101 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 299, + 1395, + 602, + 1395, + 602, + 1425, + 299, + 1425 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 815, + 76, + 815, + 104, + 298, + 104 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.792 + }, + { + "category_id": 13, + "poly": [ + 376, + 1592, + 496, + 1592, + 496, + 1624, + 376, + 1624 + ], + "score": 0.92, + "latex": "\\pi \\left( s ^ { \\prime } | s , U \\right)" + }, + { + "category_id": 13, + "poly": [ + 936, + 1530, + 1010, + 1530, + 1010, + 1563, + 936, + 1563 + ], + "score": 0.84, + "latex": "\\pi \\left( \\cdot | s \\right)" + }, + { + "category_id": 13, + "poly": [ + 782, + 1531, + 801, + 1531, + 801, + 1557, + 782, + 1557 + ], + "score": 0.78, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 356, + 1535, + 373, + 1535, + 373, + 1557, + 356, + 1557 + ], + "score": 0.72, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 405, + 1565, + 421, + 1565, + 421, + 1587, + 405, + 1587 + ], + "score": 0.7, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 761, + 996, + 821, + 996, + 821, + 1023, + 761, + 1023 + ], + "score": 0.6, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 604, + 967, + 666, + 967, + 666, + 994, + 604, + 994 + ], + "score": 0.42, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 1042, + 1178, + 1108, + 1178, + 1108, + 1207, + 1042, + 1207 + ], + "score": 0.35, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 235.0, + 651.0, + 235.0, + 651.0, + 257.0, + 580.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 235.0, + 776.0, + 235.0, + 776.0, + 257.0, + 712.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 233.0, + 916.0, + 233.0, + 916.0, + 255.0, + 839.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 232.0, + 1017.0, + 232.0, + 1017.0, + 258.0, + 973.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 233.0, + 1162.0, + 233.0, + 1162.0, + 258.0, + 1111.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 255.0, + 653.0, + 255.0, + 653.0, + 277.0, + 577.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 253.0, + 784.0, + 253.0, + 784.0, + 278.0, + 711.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 253.0, + 889.0, + 253.0, + 889.0, + 278.0, + 841.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 254.0, + 1056.0, + 254.0, + 1056.0, + 278.0, + 973.0, + 278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 253.0, + 1152.0, + 253.0, + 1152.0, + 281.0, + 1111.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 268.0, + 356.0, + 268.0, + 356.0, + 292.0, + 321.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 281.0, + 1096.0, + 281.0, + 1096.0, + 307.0, + 1048.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 319.0, + 359.0, + 319.0, + 359.0, + 341.0, + 319.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 303.0, + 721.0, + 303.0, + 721.0, + 494.0, + 685.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 311.0, + 1095.0, + 311.0, + 1095.0, + 337.0, + 1049.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 348.0, + 355.0, + 348.0, + 355.0, + 452.0, + 302.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 360.0, + 721.0, + 360.0, + 721.0, + 379.0, + 706.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 329.0, + 1095.0, + 329.0, + 1095.0, + 408.0, + 1029.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 421.0, + 356.0, + 421.0, + 356.0, + 445.0, + 319.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 414.0, + 721.0, + 414.0, + 721.0, + 433.0, + 706.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 398.0, + 1096.0, + 398.0, + 1096.0, + 473.0, + 1034.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 469.0, + 365.0, + 469.0, + 365.0, + 495.0, + 318.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 469.0, + 721.0, + 469.0, + 721.0, + 488.0, + 706.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 460.0, + 1095.0, + 460.0, + 1095.0, + 486.0, + 1049.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 491.0, + 1093.0, + 491.0, + 1093.0, + 514.0, + 1049.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 520.0, + 358.0, + 520.0, + 358.0, + 547.0, + 318.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 522.0, + 733.0, + 522.0, + 733.0, + 544.0, + 703.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 527.0, + 751.0, + 527.0, + 751.0, + 536.0, + 735.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 522.0, + 1093.0, + 522.0, + 1093.0, + 544.0, + 1049.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 533.0, + 415.0, + 533.0, + 415.0, + 557.0, + 382.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 533.0, + 485.0, + 533.0, + 485.0, + 557.0, + 453.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 533.0, + 554.0, + 533.0, + 554.0, + 557.0, + 523.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 533.0, + 626.0, + 533.0, + 626.0, + 557.0, + 594.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 533.0, + 783.0, + 533.0, + 783.0, + 557.0, + 752.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 533.0, + 855.0, + 533.0, + 855.0, + 558.0, + 823.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 533.0, + 927.0, + 533.0, + 927.0, + 557.0, + 894.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 533.0, + 997.0, + 533.0, + 997.0, + 557.0, + 967.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 533.0, + 1150.0, + 533.0, + 1150.0, + 557.0, + 1118.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 533.0, + 1222.0, + 533.0, + 1222.0, + 557.0, + 1191.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 533.0, + 1292.0, + 533.0, + 1292.0, + 557.0, + 1262.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 533.0, + 1364.0, + 533.0, + 1364.0, + 557.0, + 1333.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 551.0, + 552.0, + 551.0, + 552.0, + 575.0, + 452.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 551.0, + 923.0, + 551.0, + 923.0, + 575.0, + 822.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 551.0, + 1291.0, + 551.0, + 1291.0, + 575.0, + 1189.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 600.0, + 361.0, + 600.0, + 361.0, + 626.0, + 318.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 590.0, + 711.0, + 590.0, + 711.0, + 615.0, + 659.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 595.0, + 767.0, + 595.0, + 767.0, + 628.0, + 732.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 590.0, + 852.0, + 590.0, + 852.0, + 629.0, + 770.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 615.0, + 882.0, + 615.0, + 882.0, + 629.0, + 865.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 590.0, + 987.0, + 590.0, + 987.0, + 629.0, + 908.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 615.0, + 1013.0, + 615.0, + 1013.0, + 629.0, + 997.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 590.0, + 1124.0, + 590.0, + 1124.0, + 628.0, + 1045.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 618.0, + 1157.0, + 618.0, + 1157.0, + 626.0, + 1147.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 616.0, + 1247.0, + 616.0, + 1247.0, + 626.0, + 1237.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 619.0, + 1335.0, + 619.0, + 1335.0, + 625.0, + 1328.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 616.0, + 1381.0, + 616.0, + 1381.0, + 626.0, + 1372.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 626.0, + 723.0, + 626.0, + 723.0, + 649.0, + 704.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 629.0, + 419.0, + 629.0, + 419.0, + 640.0, + 407.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 631.0, + 712.0, + 631.0, + 712.0, + 828.0, + 685.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 633.0, + 1097.0, + 633.0, + 1097.0, + 659.0, + 1057.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 652.0, + 358.0, + 652.0, + 358.0, + 785.0, + 299.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 683.0, + 723.0, + 683.0, + 723.0, + 705.0, + 704.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 652.0, + 1096.0, + 652.0, + 1096.0, + 806.0, + 1036.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 703.0, + 354.0, + 703.0, + 354.0, + 727.0, + 321.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 705.0, + 1096.0, + 705.0, + 1096.0, + 732.0, + 1057.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 754.0, + 356.0, + 754.0, + 356.0, + 778.0, + 319.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 742.0, + 722.0, + 742.0, + 722.0, + 759.0, + 706.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 742.0, + 1096.0, + 742.0, + 1096.0, + 769.0, + 1056.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 779.0, + 1093.0, + 779.0, + 1093.0, + 803.0, + 1058.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 804.0, + 356.0, + 804.0, + 356.0, + 828.0, + 319.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 797.0, + 723.0, + 797.0, + 723.0, + 819.0, + 704.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 816.0, + 1100.0, + 816.0, + 1100.0, + 838.0, + 1058.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 829.0, + 373.0, + 829.0, + 373.0, + 846.0, + 363.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 833.0, + 418.0, + 833.0, + 418.0, + 845.0, + 408.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 835.0, + 508.0, + 835.0, + 508.0, + 843.0, + 498.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 833.0, + 598.0, + 833.0, + 598.0, + 845.0, + 587.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 832.0, + 642.0, + 832.0, + 642.0, + 847.0, + 632.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 833.0, + 1011.0, + 833.0, + 1011.0, + 847.0, + 1002.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 842.0, + 378.0, + 842.0, + 378.0, + 872.0, + 359.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 843.0, + 467.0, + 843.0, + 467.0, + 872.0, + 448.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 851.0, + 557.0, + 851.0, + 557.0, + 872.0, + 538.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 842.0, + 646.0, + 842.0, + 646.0, + 872.0, + 628.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 851.0, + 747.0, + 851.0, + 747.0, + 872.0, + 728.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 840.0, + 836.0, + 840.0, + 836.0, + 872.0, + 818.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 842.0, + 926.0, + 842.0, + 926.0, + 872.0, + 908.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 842.0, + 1016.0, + 842.0, + 1016.0, + 874.0, + 997.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 842.0, + 1116.0, + 842.0, + 1116.0, + 872.0, + 1098.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 853.0, + 1203.0, + 853.0, + 1203.0, + 871.0, + 1189.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 853.0, + 1294.0, + 853.0, + 1294.0, + 871.0, + 1278.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 853.0, + 1383.0, + 853.0, + 1383.0, + 871.0, + 1368.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 869.0, + 563.0, + 869.0, + 563.0, + 893.0, + 443.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 869.0, + 932.0, + 869.0, + 932.0, + 893.0, + 812.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 869.0, + 1301.0, + 869.0, + 1301.0, + 893.0, + 1181.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 908.0, + 1407.0, + 908.0, + 1407.0, + 947.0, + 294.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 938.0, + 1406.0, + 938.0, + 1406.0, + 972.0, + 295.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 967.0, + 603.0, + 967.0, + 603.0, + 998.0, + 295.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 967.0, + 1405.0, + 967.0, + 1405.0, + 998.0, + 667.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 994.0, + 760.0, + 994.0, + 760.0, + 1028.0, + 295.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 994.0, + 832.0, + 994.0, + 832.0, + 1028.0, + 822.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 606.0, + 1392.0, + 606.0, + 1430.0, + 294.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1436.0, + 1405.0, + 1436.0, + 1405.0, + 1475.0, + 292.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1466.0, + 1405.0, + 1466.0, + 1405.0, + 1507.0, + 292.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1498.0, + 1406.0, + 1498.0, + 1406.0, + 1535.0, + 293.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1529.0, + 355.0, + 1529.0, + 355.0, + 1566.0, + 293.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1529.0, + 781.0, + 1529.0, + 781.0, + 1566.0, + 374.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 1529.0, + 935.0, + 1529.0, + 935.0, + 1566.0, + 802.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1529.0, + 1406.0, + 1529.0, + 1406.0, + 1566.0, + 1011.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1558.0, + 404.0, + 1558.0, + 404.0, + 1598.0, + 293.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1558.0, + 1405.0, + 1558.0, + 1405.0, + 1598.0, + 422.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1589.0, + 375.0, + 1589.0, + 375.0, + 1626.0, + 294.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1589.0, + 1408.0, + 1589.0, + 1408.0, + 1626.0, + 497.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1619.0, + 1407.0, + 1619.0, + 1407.0, + 1655.0, + 292.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1650.0, + 1354.0, + 1650.0, + 1354.0, + 1686.0, + 293.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1732.0, + 295.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1730.0, + 1405.0, + 1730.0, + 1405.0, + 1764.0, + 295.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1758.0, + 1408.0, + 1758.0, + 1408.0, + 1795.0, + 292.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1785.0, + 1406.0, + 1785.0, + 1406.0, + 1829.0, + 292.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1406.0, + 1818.0, + 1406.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1888.0, + 293.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1408.0, + 1942.0, + 1408.0, + 1977.0, + 292.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 292.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2005.0, + 1403.0, + 2005.0, + 1403.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1113.0, + 1406.0, + 1113.0, + 1406.0, + 1155.0, + 292.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1405.0, + 1146.0, + 1405.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1178.0, + 1041.0, + 1178.0, + 1041.0, + 1213.0, + 293.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1178.0, + 1404.0, + 1178.0, + 1404.0, + 1213.0, + 1109.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1205.0, + 1407.0, + 1205.0, + 1407.0, + 1243.0, + 293.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1237.0, + 1406.0, + 1237.0, + 1406.0, + 1273.0, + 294.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1269.0, + 1191.0, + 1269.0, + 1191.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1315.0, + 1404.0, + 1315.0, + 1404.0, + 1350.0, + 296.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1343.0, + 1304.0, + 1343.0, + 1304.0, + 1380.0, + 294.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1038.0, + 1403.0, + 1038.0, + 1403.0, + 1073.0, + 295.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1068.0, + 1310.0, + 1068.0, + 1310.0, + 1105.0, + 292.0, + 1105.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1081, + 925, + 1081, + 925, + 1629, + 299, + 1629 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1788, + 1404, + 1788, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 718, + 1404, + 718, + 1404, + 902, + 298, + 902 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 490, + 227, + 1206, + 227, + 1206, + 535, + 490, + 535 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 1005, + 1405, + 1005, + 1405, + 1067, + 298, + 1067 + ], + "score": 0.95 + }, + { + "category_id": 4, + "poly": [ + 957, + 1510, + 1403, + 1510, + 1403, + 1621, + 957, + 1621 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 293, + 640, + 1399, + 640, + 1399, + 703, + 293, + 703 + ], + "score": 0.948 + }, + { + "category_id": 4, + "poly": [ + 294, + 569, + 1403, + 569, + 1403, + 627, + 294, + 627 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 1647, + 1402, + 1647, + 1402, + 1679, + 297, + 1679 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 301, + 949, + 808, + 949, + 808, + 980, + 301, + 980 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 299, + 1725, + 543, + 1725, + 543, + 1760, + 299, + 1760 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 864, + 2089, + 864, + 2112, + 836, + 2112 + ], + "score": 0.848 + }, + { + "category_id": 3, + "poly": [ + 1030, + 1083, + 1276, + 1083, + 1276, + 1502, + 1030, + 1502 + ], + "score": 0.467 + }, + { + "category_id": 4, + "poly": [ + 957, + 1510, + 1403, + 1510, + 1403, + 1621, + 957, + 1621 + ], + "score": 0.127 + }, + { + "category_id": 13, + "poly": [ + 337, + 1327, + 398, + 1327, + 398, + 1356, + 337, + 1356 + ], + "score": 0.61, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 919, + 599, + 976, + 599, + 976, + 624, + 919, + 624 + ], + "score": 0.56, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 609, + 1418, + 673, + 1418, + 673, + 1447, + 609, + 1447 + ], + "score": 0.52, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1601, + 358, + 1601, + 358, + 1631, + 297, + 1631 + ], + "score": 0.52, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 513, + 1327, + 581, + 1327, + 581, + 1356, + 513, + 1356 + ], + "score": 0.28, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 232.0, + 698.0, + 232.0, + 698.0, + 255.0, + 652.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 232.0, + 794.0, + 232.0, + 794.0, + 256.0, + 748.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 231.0, + 919.0, + 231.0, + 919.0, + 256.0, + 846.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 233.0, + 1008.0, + 233.0, + 1008.0, + 255.0, + 970.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 231.0, + 1103.0, + 231.0, + 1103.0, + 256.0, + 1026.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 264.0, + 537.0, + 264.0, + 537.0, + 285.0, + 506.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 267.0, + 925.0, + 267.0, + 925.0, + 289.0, + 897.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 275.0, + 936.0, + 275.0, + 936.0, + 284.0, + 928.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 298.0, + 626.0, + 298.0, + 626.0, + 305.0, + 618.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 312.0, + 906.0, + 312.0, + 906.0, + 455.0, + 875.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 319.0, + 533.0, + 319.0, + 533.0, + 338.0, + 507.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 326.0, + 921.0, + 326.0, + 921.0, + 342.0, + 900.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 328.0, + 518.0, + 328.0, + 518.0, + 438.0, + 489.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 332.0, + 574.0, + 332.0, + 574.0, + 340.0, + 568.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 341.0, + 585.0, + 341.0, + 585.0, + 367.0, + 519.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 342.0, + 839.0, + 342.0, + 839.0, + 361.0, + 772.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 359.0, + 555.0, + 359.0, + 555.0, + 367.0, + 547.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 371.0, + 534.0, + 371.0, + 534.0, + 393.0, + 507.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 380.0, + 921.0, + 380.0, + 921.0, + 399.0, + 899.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 426.0, + 533.0, + 426.0, + 533.0, + 445.0, + 508.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 436.0, + 919.0, + 436.0, + 919.0, + 452.0, + 899.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 480.0, + 537.0, + 480.0, + 537.0, + 498.0, + 508.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 498.0, + 553.0, + 498.0, + 553.0, + 516.0, + 538.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 498.0, + 609.0, + 498.0, + 609.0, + 517.0, + 587.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 498.0, + 661.0, + 498.0, + 661.0, + 516.0, + 639.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 498.0, + 713.0, + 498.0, + 713.0, + 516.0, + 691.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 498.0, + 766.0, + 498.0, + 766.0, + 517.0, + 743.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 497.0, + 817.0, + 497.0, + 817.0, + 517.0, + 796.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 489.0, + 938.0, + 489.0, + 938.0, + 514.0, + 897.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 498.0, + 993.0, + 498.0, + 993.0, + 516.0, + 971.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 500.0, + 1043.0, + 500.0, + 1043.0, + 515.0, + 1025.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 499.0, + 1094.0, + 499.0, + 1094.0, + 515.0, + 1076.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 499.0, + 1146.0, + 499.0, + 1146.0, + 517.0, + 1125.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 498.0, + 1198.0, + 498.0, + 1198.0, + 517.0, + 1176.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 509.0, + 756.0, + 509.0, + 756.0, + 535.0, + 596.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 508.0, + 1139.0, + 508.0, + 1139.0, + 536.0, + 980.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1507.0, + 1405.0, + 1507.0, + 1405.0, + 1543.0, + 954.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1538.0, + 1405.0, + 1538.0, + 1405.0, + 1567.0, + 955.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1566.0, + 1405.0, + 1566.0, + 1405.0, + 1595.0, + 954.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1594.0, + 1220.0, + 1594.0, + 1220.0, + 1621.0, + 957.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 563.0, + 1406.0, + 563.0, + 1406.0, + 605.0, + 293.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 594.0, + 918.0, + 594.0, + 918.0, + 630.0, + 293.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 594.0, + 1136.0, + 594.0, + 1136.0, + 630.0, + 977.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 945.0, + 815.0, + 945.0, + 815.0, + 985.0, + 293.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1721.0, + 548.0, + 1721.0, + 548.0, + 1769.0, + 292.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1258.0, + 1106.0, + 1258.0, + 1106.0, + 1271.0, + 1093.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1278.0, + 1105.0, + 1278.0, + 1105.0, + 1291.0, + 1093.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1307.0, + 1129.0, + 1307.0, + 1129.0, + 1315.0, + 1120.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1340.0, + 1234.0, + 1340.0, + 1234.0, + 1347.0, + 1210.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1408.0, + 1125.0, + 1408.0, + 1125.0, + 1455.0, + 1069.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1181.5, + 1058.0, + 1181.5, + 1058.0, + 1194.5, + 1050.0, + 1194.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1337.5, + 1121.0, + 1337.5, + 1121.0, + 1365.5, + 1072.0, + 1365.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1507.0, + 1405.0, + 1507.0, + 1405.0, + 1543.0, + 954.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1538.0, + 1405.0, + 1538.0, + 1405.0, + 1567.0, + 955.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1566.0, + 1405.0, + 1566.0, + 1405.0, + 1595.0, + 954.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1594.0, + 1220.0, + 1594.0, + 1220.0, + 1621.0, + 957.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1082.0, + 929.0, + 1082.0, + 929.0, + 1118.0, + 294.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1112.0, + 929.0, + 1112.0, + 929.0, + 1147.0, + 295.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1142.0, + 926.0, + 1142.0, + 926.0, + 1177.0, + 294.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1173.0, + 925.0, + 1173.0, + 925.0, + 1207.0, + 294.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1204.0, + 928.0, + 1204.0, + 928.0, + 1239.0, + 294.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1233.0, + 925.0, + 1233.0, + 925.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1265.0, + 925.0, + 1265.0, + 925.0, + 1298.0, + 294.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 928.0, + 1297.0, + 928.0, + 1329.0, + 293.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1325.0, + 336.0, + 1325.0, + 336.0, + 1359.0, + 294.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1325.0, + 512.0, + 1325.0, + 512.0, + 1359.0, + 399.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1325.0, + 927.0, + 1325.0, + 927.0, + 1359.0, + 582.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 928.0, + 1358.0, + 928.0, + 1389.0, + 295.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1388.0, + 930.0, + 1388.0, + 930.0, + 1419.0, + 294.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1417.0, + 608.0, + 1417.0, + 608.0, + 1451.0, + 295.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1417.0, + 929.0, + 1417.0, + 929.0, + 1451.0, + 674.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1448.0, + 928.0, + 1448.0, + 928.0, + 1483.0, + 295.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1479.0, + 927.0, + 1479.0, + 927.0, + 1513.0, + 294.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1507.0, + 926.0, + 1507.0, + 926.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1538.0, + 926.0, + 1538.0, + 926.0, + 1575.0, + 294.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1572.0, + 927.0, + 1572.0, + 927.0, + 1603.0, + 294.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1599.0, + 370.0, + 1599.0, + 370.0, + 1632.0, + 359.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1788.0, + 1408.0, + 1788.0, + 1408.0, + 1826.0, + 293.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1404.0, + 1821.0, + 1404.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1847.0, + 1404.0, + 1847.0, + 1404.0, + 1888.0, + 292.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1406.0, + 1881.0, + 1406.0, + 1918.0, + 292.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2011.0, + 292.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 833.0, + 2004.0, + 833.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 717.0, + 1406.0, + 717.0, + 1406.0, + 753.0, + 294.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 747.0, + 1404.0, + 747.0, + 1404.0, + 783.0, + 293.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 780.0, + 1402.0, + 780.0, + 1402.0, + 815.0, + 293.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 811.0, + 1401.0, + 811.0, + 1401.0, + 843.0, + 296.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 840.0, + 1405.0, + 840.0, + 1405.0, + 876.0, + 294.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 870.0, + 869.0, + 870.0, + 869.0, + 908.0, + 293.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1004.0, + 1405.0, + 1004.0, + 1405.0, + 1040.0, + 296.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1037.0, + 1407.0, + 1037.0, + 1407.0, + 1069.0, + 295.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 640.0, + 1404.0, + 640.0, + 1404.0, + 675.0, + 295.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 673.0, + 911.0, + 673.0, + 911.0, + 705.0, + 296.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1645.0, + 1407.0, + 1645.0, + 1407.0, + 1684.0, + 296.0, + 1684.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 296, + 1405, + 296, + 1405, + 450, + 298, + 450 + ], + "score": 0.963 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.881 + }, + { + "category_id": 1, + "poly": [ + 300, + 547, + 1402, + 547, + 1402, + 611, + 300, + 611 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 861, + 2088, + 861, + 2113, + 836, + 2113 + ], + "score": 0.814 + }, + { + "category_id": 1, + "poly": [ + 297, + 632, + 1401, + 632, + 1401, + 695, + 297, + 695 + ], + "score": 0.784 + }, + { + "category_id": 1, + "poly": [ + 296, + 944, + 1403, + 944, + 1403, + 1009, + 296, + 1009 + ], + "score": 0.764 + }, + { + "category_id": 1, + "poly": [ + 295, + 1772, + 1403, + 1772, + 1403, + 1836, + 295, + 1836 + ], + "score": 0.751 + }, + { + "category_id": 0, + "poly": [ + 302, + 228, + 609, + 228, + 609, + 261, + 302, + 261 + ], + "score": 0.747 + }, + { + "category_id": 1, + "poly": [ + 295, + 1687, + 1401, + 1687, + 1401, + 1753, + 295, + 1753 + ], + "score": 0.728 + }, + { + "category_id": 1, + "poly": [ + 293, + 1856, + 1402, + 1856, + 1402, + 1921, + 293, + 1921 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 291, + 1602, + 1404, + 1602, + 1404, + 1669, + 291, + 1669 + ], + "score": 0.716 + }, + { + "category_id": 1, + "poly": [ + 296, + 1029, + 1404, + 1029, + 1404, + 1124, + 296, + 1124 + ], + "score": 0.711 + }, + { + "category_id": 1, + "poly": [ + 300, + 831, + 1403, + 831, + 1403, + 925, + 300, + 925 + ], + "score": 0.703 + }, + { + "category_id": 1, + "poly": [ + 297, + 1488, + 1402, + 1488, + 1402, + 1583, + 297, + 1583 + ], + "score": 0.669 + }, + { + "category_id": 1, + "poly": [ + 296, + 715, + 1404, + 715, + 1404, + 811, + 296, + 811 + ], + "score": 0.667 + }, + { + "category_id": 1, + "poly": [ + 298, + 1144, + 1405, + 1144, + 1405, + 1239, + 298, + 1239 + ], + "score": 0.644 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1404, + 1941, + 1404, + 2034, + 299, + 2034 + ], + "score": 0.564 + }, + { + "category_id": 0, + "poly": [ + 299, + 496, + 488, + 496, + 488, + 530, + 299, + 530 + ], + "score": 0.556 + }, + { + "category_id": 1, + "poly": [ + 298, + 1258, + 1404, + 1258, + 1404, + 1353, + 298, + 1353 + ], + "score": 0.541 + }, + { + "category_id": 1, + "poly": [ + 296, + 1372, + 1408, + 1372, + 1408, + 1467, + 296, + 1467 + ], + "score": 0.484 + }, + { + "category_id": 1, + "poly": [ + 299, + 496, + 488, + 496, + 488, + 530, + 299, + 530 + ], + "score": 0.345 + }, + { + "category_id": 1, + "poly": [ + 302, + 228, + 609, + 228, + 609, + 261, + 302, + 261 + ], + "score": 0.112 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 228.0, + 613.0, + 228.0, + 613.0, + 267.0, + 298.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 497.0, + 490.0, + 497.0, + 490.0, + 533.0, + 296.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 296.0, + 1405.0, + 296.0, + 1405.0, + 333.0, + 296.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 327.0, + 1406.0, + 327.0, + 1406.0, + 362.0, + 293.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 356.0, + 1408.0, + 356.0, + 1408.0, + 391.0, + 292.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 389.0, + 1406.0, + 389.0, + 1406.0, + 422.0, + 296.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 415.0, + 1408.0, + 415.0, + 1408.0, + 454.0, + 293.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 547.0, + 1406.0, + 547.0, + 1406.0, + 583.0, + 298.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 578.0, + 805.0, + 578.0, + 805.0, + 614.0, + 321.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 628.0, + 1407.0, + 628.0, + 1407.0, + 670.0, + 292.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 662.0, + 776.0, + 662.0, + 776.0, + 695.0, + 322.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 942.0, + 1405.0, + 942.0, + 1405.0, + 982.0, + 294.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 977.0, + 764.0, + 977.0, + 764.0, + 1011.0, + 320.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1769.0, + 1406.0, + 1769.0, + 1406.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1804.0, + 478.0, + 1804.0, + 478.0, + 1833.0, + 324.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1690.0, + 1403.0, + 1690.0, + 1403.0, + 1723.0, + 297.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1719.0, + 1240.0, + 1719.0, + 1240.0, + 1754.0, + 320.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1852.0, + 1404.0, + 1852.0, + 1404.0, + 1895.0, + 292.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1889.0, + 908.0, + 1889.0, + 908.0, + 1922.0, + 321.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1599.0, + 1405.0, + 1599.0, + 1405.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1634.0, + 903.0, + 1634.0, + 903.0, + 1667.0, + 324.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1028.0, + 1404.0, + 1028.0, + 1404.0, + 1068.0, + 293.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1057.0, + 1409.0, + 1057.0, + 1409.0, + 1100.0, + 322.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1092.0, + 527.0, + 1092.0, + 527.0, + 1123.0, + 323.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 832.0, + 1407.0, + 832.0, + 1407.0, + 866.0, + 295.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 857.0, + 1409.0, + 857.0, + 1409.0, + 901.0, + 318.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 890.0, + 620.0, + 890.0, + 620.0, + 927.0, + 320.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1486.0, + 1407.0, + 1486.0, + 1407.0, + 1527.0, + 295.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1517.0, + 1408.0, + 1517.0, + 1408.0, + 1559.0, + 321.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1551.0, + 614.0, + 1551.0, + 614.0, + 1585.0, + 322.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 713.0, + 1406.0, + 713.0, + 1406.0, + 754.0, + 293.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 745.0, + 1404.0, + 745.0, + 1404.0, + 783.0, + 323.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 778.0, + 1084.0, + 778.0, + 1084.0, + 812.0, + 323.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1145.0, + 1406.0, + 1145.0, + 1406.0, + 1179.0, + 296.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1177.0, + 1406.0, + 1177.0, + 1406.0, + 1211.0, + 323.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1206.0, + 1127.0, + 1206.0, + 1127.0, + 1240.0, + 323.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1404.0, + 1940.0, + 1404.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2008.0, + 322.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2000.0, + 1398.0, + 2000.0, + 1398.0, + 2038.0, + 321.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1260.0, + 1402.0, + 1260.0, + 1402.0, + 1295.0, + 297.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1292.0, + 1404.0, + 1292.0, + 1404.0, + 1326.0, + 323.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1321.0, + 990.0, + 1321.0, + 990.0, + 1355.0, + 322.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1374.0, + 1404.0, + 1374.0, + 1404.0, + 1409.0, + 298.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1401.0, + 1411.0, + 1401.0, + 1411.0, + 1444.0, + 319.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1435.0, + 505.0, + 1435.0, + 505.0, + 1467.0, + 323.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 497.0, + 490.0, + 497.0, + 490.0, + 533.0, + 296.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 228.0, + 613.0, + 228.0, + 613.0, + 267.0, + 298.0, + 267.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 817, + 75, + 817, + 105, + 300, + 105 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.813 + }, + { + "category_id": 1, + "poly": [ + 292, + 179, + 1411, + 179, + 1411, + 2034, + 292, + 2034 + ], + "score": 0.67 + }, + { + "category_id": 13, + "poly": [ + 1050, + 232, + 1095, + 232, + 1095, + 260, + 1050, + 260 + ], + "score": 0.37, + "latex": "\\mathrm { { r r t } ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1049.0, + 229.0, + 1049.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 229.0, + 1407.0, + 229.0, + 1407.0, + 267.0, + 1096.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 259.0, + 1409.0, + 259.0, + 1409.0, + 302.0, + 316.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 294.0, + 595.0, + 294.0, + 595.0, + 327.0, + 322.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 341.0, + 1407.0, + 341.0, + 1407.0, + 379.0, + 294.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 373.0, + 1407.0, + 373.0, + 1407.0, + 412.0, + 322.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 404.0, + 1409.0, + 404.0, + 1409.0, + 443.0, + 320.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 434.0, + 398.0, + 434.0, + 398.0, + 469.0, + 322.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 479.0, + 1409.0, + 479.0, + 1409.0, + 522.0, + 292.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 518.0, + 1405.0, + 518.0, + 1405.0, + 551.0, + 322.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 544.0, + 775.0, + 544.0, + 775.0, + 583.0, + 320.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 597.0, + 1405.0, + 597.0, + 1405.0, + 630.0, + 296.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 628.0, + 1409.0, + 628.0, + 1409.0, + 667.0, + 320.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 675.0, + 1409.0, + 675.0, + 1409.0, + 716.0, + 292.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 705.0, + 1407.0, + 705.0, + 1407.0, + 744.0, + 318.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 738.0, + 650.0, + 738.0, + 650.0, + 776.0, + 316.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 789.0, + 1407.0, + 789.0, + 1407.0, + 828.0, + 294.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 817.0, + 1409.0, + 817.0, + 1409.0, + 860.0, + 316.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 848.0, + 944.0, + 848.0, + 944.0, + 891.0, + 316.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 895.0, + 1411.0, + 895.0, + 1411.0, + 946.0, + 290.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 931.0, + 1407.0, + 931.0, + 1407.0, + 970.0, + 320.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 960.0, + 823.0, + 960.0, + 823.0, + 1003.0, + 318.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1007.0, + 1407.0, + 1007.0, + 1407.0, + 1058.0, + 290.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1043.0, + 1180.0, + 1043.0, + 1180.0, + 1082.0, + 320.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1094.0, + 1407.0, + 1094.0, + 1407.0, + 1135.0, + 292.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1129.0, + 1407.0, + 1129.0, + 1407.0, + 1161.0, + 322.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1157.0, + 398.0, + 1157.0, + 398.0, + 1192.0, + 322.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1206.0, + 1409.0, + 1206.0, + 1409.0, + 1247.0, + 292.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1241.0, + 966.0, + 1241.0, + 966.0, + 1273.0, + 322.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1286.0, + 1409.0, + 1286.0, + 1409.0, + 1329.0, + 292.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1318.0, + 1196.0, + 1318.0, + 1196.0, + 1357.0, + 320.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1369.0, + 1407.0, + 1369.0, + 1407.0, + 1408.0, + 294.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1402.0, + 1409.0, + 1402.0, + 1409.0, + 1438.0, + 316.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1434.0, + 607.0, + 1434.0, + 607.0, + 1467.0, + 318.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 1403.0, + 1481.0, + 1403.0, + 1520.0, + 294.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1514.0, + 1080.0, + 1514.0, + 1080.0, + 1552.0, + 320.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1565.0, + 1407.0, + 1565.0, + 1407.0, + 1603.0, + 294.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1593.0, + 860.0, + 1593.0, + 860.0, + 1632.0, + 318.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1648.0, + 1405.0, + 1648.0, + 1405.0, + 1681.0, + 296.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1677.0, + 946.0, + 1677.0, + 946.0, + 1715.0, + 318.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1730.0, + 1405.0, + 1730.0, + 1405.0, + 1762.0, + 296.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1758.0, + 1409.0, + 1758.0, + 1409.0, + 1797.0, + 320.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1782.0, + 399.0, + 1782.0, + 399.0, + 1828.0, + 320.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1842.0, + 1407.0, + 1842.0, + 1407.0, + 1874.0, + 296.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1868.0, + 1160.0, + 1868.0, + 1160.0, + 1909.0, + 318.0, + 1909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1923.0, + 1240.0, + 1923.0, + 1240.0, + 1956.0, + 296.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1409.0, + 1972.0, + 1409.0, + 2011.0, + 294.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 2003.0, + 777.0, + 2003.0, + 777.0, + 2041.0, + 320.0, + 2041.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 292, + 229, + 1404, + 229, + 1404, + 293, + 292, + 293 + ], + "score": 0.682 + }, + { + "category_id": 1, + "poly": [ + 295, + 985, + 1400, + 985, + 1400, + 1050, + 295, + 1050 + ], + "score": 0.648 + }, + { + "category_id": 1, + "poly": [ + 297, + 622, + 1399, + 622, + 1399, + 688, + 297, + 688 + ], + "score": 0.636 + }, + { + "category_id": 1, + "poly": [ + 299, + 788, + 1400, + 788, + 1400, + 855, + 299, + 855 + ], + "score": 0.63 + }, + { + "category_id": 1, + "poly": [ + 299, + 705, + 1399, + 705, + 1399, + 772, + 299, + 772 + ], + "score": 0.623 + }, + { + "category_id": 1, + "poly": [ + 292, + 1691, + 1403, + 1691, + 1403, + 1756, + 292, + 1756 + ], + "score": 0.613 + }, + { + "category_id": 1, + "poly": [ + 296, + 872, + 1402, + 872, + 1402, + 967, + 296, + 967 + ], + "score": 0.606 + }, + { + "category_id": 1, + "poly": [ + 296, + 1576, + 1404, + 1576, + 1404, + 1673, + 296, + 1673 + ], + "score": 0.598 + }, + { + "category_id": 1, + "poly": [ + 293, + 1774, + 1401, + 1774, + 1401, + 1839, + 293, + 1839 + ], + "score": 0.587 + }, + { + "category_id": 1, + "poly": [ + 291, + 1493, + 1401, + 1493, + 1401, + 1559, + 291, + 1559 + ], + "score": 0.573 + }, + { + "category_id": 1, + "poly": [ + 301, + 508, + 1400, + 508, + 1400, + 603, + 301, + 603 + ], + "score": 0.57 + }, + { + "category_id": 1, + "poly": [ + 295, + 426, + 1402, + 426, + 1402, + 490, + 295, + 490 + ], + "score": 0.552 + }, + { + "category_id": 1, + "poly": [ + 290, + 1326, + 1403, + 1326, + 1403, + 1393, + 290, + 1393 + ], + "score": 0.552 + }, + { + "category_id": 1, + "poly": [ + 298, + 311, + 1400, + 311, + 1400, + 407, + 298, + 407 + ], + "score": 0.527 + }, + { + "category_id": 1, + "poly": [ + 293, + 1410, + 1402, + 1410, + 1402, + 1476, + 293, + 1476 + ], + "score": 0.527 + }, + { + "category_id": 1, + "poly": [ + 298, + 1182, + 1406, + 1182, + 1406, + 1309, + 298, + 1309 + ], + "score": 0.483 + }, + { + "category_id": 1, + "poly": [ + 300, + 1941, + 1405, + 1941, + 1405, + 2035, + 300, + 2035 + ], + "score": 0.474 + }, + { + "category_id": 1, + "poly": [ + 294, + 1856, + 1404, + 1856, + 1404, + 1921, + 294, + 1921 + ], + "score": 0.471 + }, + { + "category_id": 1, + "poly": [ + 294, + 1069, + 1404, + 1069, + 1404, + 1163, + 294, + 1163 + ], + "score": 0.342 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 223.0, + 1410.0, + 223.0, + 1410.0, + 271.0, + 291.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 261.0, + 790.0, + 261.0, + 790.0, + 293.0, + 320.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 984.0, + 1405.0, + 984.0, + 1405.0, + 1024.0, + 295.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1019.0, + 707.0, + 1019.0, + 707.0, + 1052.0, + 320.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 619.0, + 1405.0, + 619.0, + 1405.0, + 663.0, + 292.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 654.0, + 1373.0, + 654.0, + 1373.0, + 689.0, + 322.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 785.0, + 1405.0, + 785.0, + 1405.0, + 828.0, + 292.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 819.0, + 1111.0, + 819.0, + 1111.0, + 855.0, + 322.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 702.0, + 1404.0, + 702.0, + 1404.0, + 746.0, + 294.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 737.0, + 1224.0, + 737.0, + 1224.0, + 772.0, + 321.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1691.0, + 1405.0, + 1691.0, + 1405.0, + 1727.0, + 295.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1721.0, + 1025.0, + 1721.0, + 1025.0, + 1756.0, + 321.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 1404.0, + 872.0, + 1404.0, + 909.0, + 294.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 905.0, + 1409.0, + 905.0, + 1409.0, + 939.0, + 323.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 935.0, + 467.0, + 935.0, + 467.0, + 967.0, + 324.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1575.0, + 1406.0, + 1575.0, + 1406.0, + 1615.0, + 293.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1608.0, + 1409.0, + 1608.0, + 1409.0, + 1643.0, + 323.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1635.0, + 646.0, + 1635.0, + 646.0, + 1678.0, + 322.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1774.0, + 1404.0, + 1774.0, + 1404.0, + 1810.0, + 297.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1803.0, + 1055.0, + 1803.0, + 1055.0, + 1841.0, + 321.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1494.0, + 1404.0, + 1494.0, + 1404.0, + 1530.0, + 296.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1525.0, + 1252.0, + 1525.0, + 1252.0, + 1560.0, + 322.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 505.0, + 1404.0, + 505.0, + 1404.0, + 548.0, + 294.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 543.0, + 1403.0, + 543.0, + 1403.0, + 574.0, + 323.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 571.0, + 714.0, + 571.0, + 714.0, + 605.0, + 320.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 423.0, + 1404.0, + 423.0, + 1404.0, + 463.0, + 295.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 455.0, + 1108.0, + 455.0, + 1108.0, + 491.0, + 322.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1327.0, + 1405.0, + 1327.0, + 1405.0, + 1363.0, + 296.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1358.0, + 1409.0, + 1358.0, + 1409.0, + 1395.0, + 321.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 310.0, + 1406.0, + 310.0, + 1406.0, + 349.0, + 292.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 346.0, + 1405.0, + 346.0, + 1405.0, + 380.0, + 323.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 375.0, + 1334.0, + 375.0, + 1334.0, + 409.0, + 324.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1408.0, + 1405.0, + 1408.0, + 1405.0, + 1448.0, + 293.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1443.0, + 1043.0, + 1443.0, + 1043.0, + 1475.0, + 321.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1184.0, + 1406.0, + 1184.0, + 1406.0, + 1217.0, + 296.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1217.0, + 1404.0, + 1217.0, + 1404.0, + 1247.0, + 324.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1244.0, + 1404.0, + 1244.0, + 1404.0, + 1280.0, + 322.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1276.0, + 998.0, + 1276.0, + 998.0, + 1309.0, + 321.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1976.0, + 298.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2006.0, + 322.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1999.0, + 840.0, + 1999.0, + 840.0, + 2038.0, + 320.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1407.0, + 1851.0, + 1407.0, + 1898.0, + 293.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1882.0, + 400.0, + 1882.0, + 400.0, + 1925.0, + 318.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1067.0, + 1407.0, + 1067.0, + 1407.0, + 1108.0, + 293.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1097.0, + 1409.0, + 1097.0, + 1409.0, + 1139.0, + 321.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1131.0, + 648.0, + 1131.0, + 648.0, + 1163.0, + 321.0, + 1163.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 817, + 75, + 817, + 105, + 300, + 105 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2112, + 835, + 2112 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 291, + 229, + 1405, + 229, + 1405, + 293, + 291, + 293 + ], + "score": 0.452 + }, + { + "category_id": 1, + "poly": [ + 297, + 311, + 1406, + 311, + 1406, + 405, + 297, + 405 + ], + "score": 0.435 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1408.0, + 225.0, + 1408.0, + 269.0, + 293.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 775.0, + 261.0, + 775.0, + 293.0, + 323.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 311.0, + 1409.0, + 311.0, + 1409.0, + 351.0, + 292.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 342.0, + 1408.0, + 342.0, + 1408.0, + 379.0, + 323.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 372.0, + 469.0, + 372.0, + 469.0, + 408.0, + 323.0, + 408.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1212, + 1406, + 1212, + 1406, + 1520, + 295, + 1520 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 296, + 390, + 1407, + 390, + 1407, + 484, + 296, + 484 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 441, + 506, + 1295, + 506, + 1295, + 863, + 441, + 863 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 532, + 1538, + 1168, + 1538, + 1168, + 1785, + 532, + 1785 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 295, + 885, + 1406, + 885, + 1406, + 1102, + 295, + 1102 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1396, + 1972, + 1396, + 2035, + 297, + 2035 + ], + "score": 0.932 + }, + { + "category_id": 4, + "poly": [ + 300, + 1815, + 1404, + 1815, + 1404, + 1906, + 300, + 1906 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 737, + 220, + 964, + 220, + 964, + 283, + 737, + 283 + ], + "score": 0.921 + }, + { + "category_id": 0, + "poly": [ + 298, + 319, + 1224, + 319, + 1224, + 356, + 298, + 356 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 299, + 1913, + 1068, + 1913, + 1068, + 1946, + 299, + 1946 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 300, + 1143, + 682, + 1143, + 682, + 1179, + 300, + 1179 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.862 + }, + { + "category_id": 13, + "poly": [ + 497, + 1398, + 559, + 1398, + 559, + 1431, + 497, + 1431 + ], + "score": 0.92, + "latex": "\\mathcal { S } _ { g o a l }" + }, + { + "category_id": 13, + "poly": [ + 403, + 1277, + 514, + 1277, + 514, + 1306, + 403, + 1306 + ], + "score": 0.92, + "latex": "s _ { n e w } \\in S" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1309, + 1374, + 1309, + 1374, + 1340, + 1208, + 1340 + ], + "score": 0.89, + "latex": "[ s _ { p a r e n t } , s _ { n e w } ]" + }, + { + "category_id": 13, + "poly": [ + 1298, + 1370, + 1356, + 1370, + 1356, + 1398, + 1298, + 1398 + ], + "score": 0.89, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 1196, + 1246, + 1334, + 1246, + 1334, + 1278, + 1196, + 1278 + ], + "score": 0.89, + "latex": "s _ { p a r e n t } \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 1129, + 1341, + 1187, + 1341, + 1187, + 1367, + 1129, + 1367 + ], + "score": 0.88, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 1157, + 1462, + 1215, + 1462, + 1215, + 1490, + 1157, + 1490 + ], + "score": 0.88, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 934, + 1458, + 1099, + 1458, + 1099, + 1492, + 934, + 1492 + ], + "score": 0.85, + "latex": "[ s _ { p a r e n t } , s _ { n e w } ]" + }, + { + "category_id": 13, + "poly": [ + 895, + 1279, + 978, + 1279, + 978, + 1309, + 895, + 1309 + ], + "score": 0.8, + "latex": "s _ { p a r e n t }" + }, + { + "category_id": 13, + "poly": [ + 1030, + 1338, + 1056, + 1338, + 1056, + 1366, + 1030, + 1366 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1278, + 1401, + 1278, + 1401, + 1308, + 1236, + 1308 + ], + "score": 0.72, + "latex": "[ s _ { p a r e n t } , s _ { n e w } ]" + }, + { + "category_id": 13, + "poly": [ + 1061, + 2009, + 1077, + 2009, + 1077, + 2030, + 1061, + 2030 + ], + "score": 0.7, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 410, + 1368, + 432, + 1368, + 432, + 1395, + 410, + 1395 + ], + "score": 0.7, + "latex": "\\nu" + }, + { + "category_id": 13, + "poly": [ + 991, + 1368, + 1012, + 1368, + 1012, + 1394, + 991, + 1394 + ], + "score": 0.68, + "latex": "\\mathcal { E }" + }, + { + "category_id": 13, + "poly": [ + 782, + 983, + 800, + 983, + 800, + 1006, + 782, + 1006 + ], + "score": 0.46, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 1335, + 1276, + 1400, + 1276, + 1400, + 1308, + 1335, + 1308 + ], + "score": 0.38, + "latex": "s _ { n e w } ]" + }, + { + "category_id": 13, + "poly": [ + 1269, + 600, + 1289, + 600, + 1289, + 622, + 1269, + 622 + ], + "score": 0.34, + "latex": "\\scriptstyle { \\frac { 1 } { 2 } } \\pi" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 518.0, + 462.0, + 518.0, + 462.0, + 533.0, + 450.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 549.0, + 462.0, + 549.0, + 462.0, + 562.0, + 450.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 570.0, + 1287.0, + 570.0, + 1287.0, + 579.0, + 1278.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 580.0, + 460.0, + 580.0, + 460.0, + 589.0, + 451.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 609.0, + 463.0, + 609.0, + 463.0, + 622.0, + 449.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 639.0, + 462.0, + 639.0, + 462.0, + 653.0, + 450.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 643.0, + 1033.0, + 643.0, + 1033.0, + 650.0, + 1015.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 638.0, + 1284.0, + 638.0, + 1284.0, + 650.0, + 1271.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 669.0, + 463.0, + 669.0, + 463.0, + 681.0, + 450.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 670.0, + 1292.0, + 670.0, + 1292.0, + 694.0, + 1265.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 699.0, + 462.0, + 699.0, + 462.0, + 712.0, + 449.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 705.0, + 1285.0, + 705.0, + 1285.0, + 719.0, + 1266.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 758.0, + 938.0, + 758.0, + 938.0, + 768.0, + 929.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 765.0, + 961.0, + 765.0, + 961.0, + 773.0, + 949.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 783.0, + 1049.0, + 783.0, + 1049.0, + 792.0, + 1041.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 791.0, + 461.0, + 791.0, + 461.0, + 800.0, + 451.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 788.0, + 1079.0, + 788.0, + 1079.0, + 800.0, + 1062.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 797.0, + 1123.0, + 797.0, + 1123.0, + 812.0, + 1107.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 813.0, + 485.0, + 813.0, + 485.0, + 826.0, + 472.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 804.0, + 758.0, + 804.0, + 758.0, + 832.0, + 498.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 829.0, + 700.0, + 829.0, + 700.0, + 869.0, + 528.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 827.0, + 1223.0, + 827.0, + 1223.0, + 869.0, + 950.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 767.0, + 982.0, + 767.0, + 982.0, + 778.5, + 971.0, + 778.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 748.0, + 1272.0, + 748.0, + 1272.0, + 766.0, + 1169.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 791.5, + 1100.0, + 791.5, + 1100.0, + 803.5, + 1084.0, + 803.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 785.5, + 1188.0, + 785.5, + 1188.0, + 802.0, + 1148.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1570.0, + 825.0, + 1570.0, + 825.0, + 1598.0, + 779.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1569.0, + 1154.0, + 1569.0, + 1154.0, + 1597.0, + 1108.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1605.0, + 576.0, + 1605.0, + 576.0, + 1622.0, + 560.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1592.0, + 815.0, + 1592.0, + 815.0, + 1616.0, + 789.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1605.0, + 903.0, + 1605.0, + 903.0, + 1619.0, + 890.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1592.0, + 1143.0, + 1592.0, + 1143.0, + 1616.0, + 1118.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1628.0, + 746.0, + 1628.0, + 746.0, + 1654.0, + 703.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1662.0, + 593.0, + 1662.0, + 593.0, + 1678.0, + 577.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1640.0, + 708.0, + 1640.0, + 708.0, + 1699.0, + 677.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1662.0, + 921.0, + 1662.0, + 921.0, + 1675.0, + 907.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 1667.0, + 1117.0, + 1667.0, + 1117.0, + 1691.0, + 1074.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1681.0, + 581.0, + 1681.0, + 581.0, + 1739.0, + 532.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1691.0, + 747.0, + 1691.0, + 747.0, + 1723.0, + 676.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1682.0, + 908.0, + 1682.0, + 908.0, + 1736.0, + 861.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1698.0, + 1078.0, + 1698.0, + 1078.0, + 1731.0, + 1004.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1731.0, + 637.0, + 1731.0, + 637.0, + 1747.0, + 621.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1731.0, + 965.0, + 1731.0, + 965.0, + 1747.0, + 950.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1750.0, + 702.0, + 1750.0, + 702.0, + 1766.0, + 685.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1749.0, + 1030.0, + 1749.0, + 1030.0, + 1766.0, + 1013.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1619.5, + 643.0, + 1619.5, + 643.0, + 1631.5, + 618.0, + 1631.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1671.5, + 1085.0, + 1671.5, + 1085.0, + 1695.0, + 1025.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 882.0, + 1406.0, + 882.0, + 1406.0, + 924.0, + 293.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 918.0, + 1406.0, + 918.0, + 1406.0, + 953.0, + 295.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 948.0, + 1406.0, + 948.0, + 1406.0, + 982.0, + 293.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 977.0, + 781.0, + 977.0, + 781.0, + 1013.0, + 294.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 977.0, + 1408.0, + 977.0, + 1408.0, + 1013.0, + 801.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1008.0, + 1406.0, + 1008.0, + 1406.0, + 1043.0, + 294.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1039.0, + 1407.0, + 1039.0, + 1407.0, + 1074.0, + 293.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1069.0, + 694.0, + 1069.0, + 694.0, + 1104.0, + 295.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1812.0, + 1405.0, + 1812.0, + 1405.0, + 1852.0, + 293.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1843.0, + 1404.0, + 1843.0, + 1404.0, + 1880.0, + 295.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1875.0, + 535.0, + 1875.0, + 535.0, + 1911.0, + 294.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 215.0, + 971.0, + 215.0, + 971.0, + 287.0, + 727.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1228.0, + 319.0, + 1228.0, + 361.0, + 294.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1139.0, + 686.0, + 1139.0, + 686.0, + 1184.0, + 292.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1208.0, + 1409.0, + 1208.0, + 1409.0, + 1253.0, + 293.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1242.0, + 1195.0, + 1242.0, + 1195.0, + 1282.0, + 294.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 1242.0, + 1407.0, + 1242.0, + 1407.0, + 1282.0, + 1335.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1272.0, + 402.0, + 1272.0, + 402.0, + 1315.0, + 292.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1272.0, + 894.0, + 1272.0, + 894.0, + 1315.0, + 515.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1272.0, + 1235.0, + 1272.0, + 1235.0, + 1315.0, + 979.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1302.0, + 1207.0, + 1302.0, + 1207.0, + 1346.0, + 292.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1375.0, + 1302.0, + 1407.0, + 1302.0, + 1407.0, + 1346.0, + 1375.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1334.0, + 1029.0, + 1334.0, + 1029.0, + 1372.0, + 293.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 1334.0, + 1128.0, + 1334.0, + 1128.0, + 1372.0, + 1057.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1334.0, + 1406.0, + 1334.0, + 1406.0, + 1372.0, + 1188.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1362.0, + 409.0, + 1362.0, + 409.0, + 1404.0, + 292.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1362.0, + 990.0, + 1362.0, + 990.0, + 1404.0, + 433.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1362.0, + 1297.0, + 1362.0, + 1297.0, + 1404.0, + 1013.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1357.0, + 1362.0, + 1407.0, + 1362.0, + 1407.0, + 1404.0, + 1357.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1395.0, + 496.0, + 1395.0, + 496.0, + 1435.0, + 293.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1395.0, + 1407.0, + 1395.0, + 1407.0, + 1435.0, + 560.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1427.0, + 1404.0, + 1427.0, + 1404.0, + 1463.0, + 293.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1457.0, + 933.0, + 1457.0, + 933.0, + 1496.0, + 292.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1457.0, + 1156.0, + 1457.0, + 1156.0, + 1496.0, + 1100.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 1457.0, + 1407.0, + 1457.0, + 1407.0, + 1496.0, + 1216.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1488.0, + 1215.0, + 1488.0, + 1215.0, + 1524.0, + 294.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 388.0, + 1405.0, + 388.0, + 1405.0, + 427.0, + 291.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 422.0, + 1405.0, + 422.0, + 1405.0, + 456.0, + 294.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 453.0, + 797.0, + 453.0, + 797.0, + 487.0, + 295.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 1401.0, + 1972.0, + 1401.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 2000.0, + 1060.0, + 2000.0, + 1060.0, + 2040.0, + 319.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 2000.0, + 1402.0, + 2000.0, + 1402.0, + 2040.0, + 1078.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1908.0, + 1072.0, + 1908.0, + 1072.0, + 1953.0, + 293.0, + 1953.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1339, + 1404, + 1339, + 1404, + 1587, + 297, + 1587 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 296, + 987, + 1405, + 987, + 1405, + 1149, + 296, + 1149 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1726, + 1405, + 1726, + 1405, + 1821, + 298, + 1821 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1233, + 1404, + 1233, + 1404, + 1326, + 298, + 1326 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 300, + 1600, + 1399, + 1600, + 1399, + 1695, + 300, + 1695 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 301, + 853, + 1403, + 853, + 1403, + 947, + 301, + 947 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 300, + 1972, + 1401, + 1972, + 1401, + 2035, + 300, + 2035 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 321, + 753, + 1403, + 753, + 1403, + 845, + 321, + 845 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 613, + 1176, + 1087, + 1176, + 1087, + 1230, + 613, + 1230 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 662, + 714, + 1063, + 714, + 1063, + 751, + 662, + 751 + ], + "score": 0.934 + }, + { + "category_id": 8, + "poly": [ + 453, + 1899, + 1242, + 1899, + 1242, + 1965, + 453, + 1965 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 316, + 679, + 1358, + 679, + 1358, + 712, + 316, + 712 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 816, + 74, + 816, + 105, + 299, + 105 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1190, + 1399, + 1190, + 1399, + 1220, + 1366, + 1220 + ], + "score": 0.874 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1909, + 1399, + 1909, + 1399, + 1938, + 1366, + 1938 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.867 + }, + { + "category_id": 1, + "poly": [ + 276, + 279, + 938, + 279, + 938, + 413, + 276, + 413 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 300, + 1834, + 1403, + 1834, + 1403, + 1897, + 300, + 1897 + ], + "score": 0.821 + }, + { + "category_id": 1, + "poly": [ + 1089, + 556, + 1362, + 556, + 1362, + 618, + 1089, + 618 + ], + "score": 0.76 + }, + { + "category_id": 0, + "poly": [ + 298, + 482, + 699, + 482, + 699, + 515, + 298, + 515 + ], + "score": 0.649 + }, + { + "category_id": 1, + "poly": [ + 274, + 522, + 939, + 522, + 939, + 652, + 274, + 652 + ], + "score": 0.556 + }, + { + "category_id": 0, + "poly": [ + 299, + 239, + 699, + 239, + 699, + 272, + 299, + 272 + ], + "score": 0.538 + }, + { + "category_id": 1, + "poly": [ + 276, + 414, + 540, + 414, + 540, + 442, + 276, + 442 + ], + "score": 0.507 + }, + { + "category_id": 1, + "poly": [ + 299, + 239, + 699, + 239, + 699, + 272, + 299, + 272 + ], + "score": 0.328 + }, + { + "category_id": 1, + "poly": [ + 1028, + 313, + 1362, + 313, + 1362, + 373, + 1028, + 373 + ], + "score": 0.238 + }, + { + "category_id": 1, + "poly": [ + 298, + 482, + 699, + 482, + 699, + 515, + 298, + 515 + ], + "score": 0.201 + }, + { + "category_id": 13, + "poly": [ + 369, + 1972, + 482, + 1972, + 482, + 2006, + 369, + 2006 + ], + "score": 0.93, + "latex": "P ( s ^ { \\prime } | s , a )" + }, + { + "category_id": 14, + "poly": [ + 613, + 1174, + 1083, + 1174, + 1083, + 1233, + 613, + 1233 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { a _ { T + 1 } ^ { * } = \\operatorname * { a r g m a x } _ { i \\in \\{ 1 , \\ldots , K \\} } \\bar { r } _ { i } + \\lambda \\sqrt { \\frac { \\log T } { n _ { i } } } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1218, + 2002, + 1314, + 2002, + 1314, + 2036, + 1218, + 2036 + ], + "score": 0.93, + "latex": "c ( [ s , s ^ { \\prime } ] )" + }, + { + "category_id": 13, + "poly": [ + 299, + 1080, + 446, + 1080, + 446, + 1121, + 299, + 1121 + ], + "score": 0.92, + "latex": "\\{ ( \\bar { r } _ { i } , n _ { i } ) \\} _ { i = 1 } ^ { K }" + }, + { + "category_id": 14, + "poly": [ + 455, + 1897, + 1244, + 1897, + 1244, + 1969, + 455, + 1969 + ], + "score": 0.92, + "latex": "V ^ { * } \\left( s | U \\right) = \\left( \\mathcal { G } V ^ { * } \\right) ( s ) : = \\operatorname* { m i n } _ { a } \\sum _ { s ^ { \\prime } } P ( s ^ { \\prime } | s , a ) ( c \\left( \\left[ s , s ^ { \\prime } \\right] \\right) + V ^ { * } \\left( s ^ { \\prime } | U \\right) ) ." + }, + { + "category_id": 13, + "poly": [ + 400, + 1264, + 511, + 1264, + 511, + 1298, + 400, + 1298 + ], + "score": 0.92, + "latex": "\\mathcal { O } \\left( \\log T \\right)" + }, + { + "category_id": 13, + "poly": [ + 507, + 1081, + 665, + 1081, + 665, + 1122, + 507, + 1122 + ], + "score": 0.91, + "latex": "\\textstyle \\sum _ { i = 1 } ^ { K } n _ { i } = T" + }, + { + "category_id": 13, + "poly": [ + 660, + 713, + 1057, + 713, + 1057, + 751, + 660, + 751 + ], + "score": 0.89, + "latex": "B ( s , \\eta ) = \\{ s ^ { \\prime } \\in S \\mid \\Vert s ^ { \\prime } - s \\Vert \\leqslant \\eta \\}" + }, + { + "category_id": 13, + "poly": [ + 397, + 1868, + 433, + 1868, + 433, + 1894, + 397, + 1894 + ], + "score": 0.86, + "latex": "V ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1101, + 1022, + 1126, + 1022, + 1126, + 1050, + 1101, + 1050 + ], + "score": 0.86, + "latex": "\\bar { r } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1285, + 1838, + 1306, + 1838, + 1306, + 1865, + 1285, + 1865 + ], + "score": 0.85, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 297, + 378, + 776, + 378, + 776, + 411, + 297, + 411 + ], + "score": 0.85, + "latex": "s _ { n e w } \\gets \\mathrm { a r g m i n } _ { s \\in \\mathcal { B } ( s _ { p a r e n t } , \\eta ) } \\left\\| s - s _ { r a n d } \\right\\|" + }, + { + "category_id": 13, + "poly": [ + 787, + 991, + 816, + 991, + 816, + 1017, + 787, + 1017 + ], + "score": 0.83, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 296, + 586, + 604, + 586, + 604, + 619, + 296, + 619 + ], + "score": 0.83, + "latex": "\\hat { s _ { n e w } } \\mathcal { U } n i f ( B ( s _ { p a r e n t } ) )" + }, + { + "category_id": 13, + "poly": [ + 298, + 346, + 692, + 346, + 692, + 376, + 298, + 376 + ], + "score": 0.83, + "latex": "\\begin{array} { r } { s _ { p a r e n t } \\gets \\operatorname * { a r g m i n } _ { s \\in \\mathcal { V } } \\left\\| s _ { r a n d } - s \\right\\| } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 565, + 1052, + 589, + 1052, + 589, + 1078, + 565, + 1078 + ], + "score": 0.82, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 983, + 682, + 1010, + 682, + 1010, + 709, + 983, + 709 + ], + "score": 0.8, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 372, + 1235, + 393, + 1235, + 393, + 1261, + 372, + 1261 + ], + "score": 0.8, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 298, + 1056, + 326, + 1056, + 326, + 1081, + 298, + 1081 + ], + "score": 0.79, + "latex": "n _ { i }" + }, + { + "category_id": 13, + "poly": [ + 298, + 315, + 509, + 315, + 509, + 344, + 298, + 344 + ], + "score": 0.79, + "latex": "s _ { r a n d } \\gets \\mathcal { U } n i f ( \\mathcal { S } )" + }, + { + "category_id": 13, + "poly": [ + 1332, + 687, + 1350, + 687, + 1350, + 713, + 1332, + 713 + ], + "score": 0.78, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1300, + 1980, + 1317, + 1980, + 1317, + 2000, + 1300, + 2000 + ], + "score": 0.78, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 375, + 524, + 934, + 524, + 934, + 558, + 375, + 558 + ], + "score": 0.78, + "latex": "\\mathcal { T } = ( \\mathcal { V } , \\mathcal { E } ) , U = \\left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c ( \\cdot ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 484, + 622, + 541, + 622, + 541, + 647, + 484, + 647 + ], + "score": 0.77, + "latex": "s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 1062, + 1979, + 1077, + 1979, + 1077, + 2000, + 1062, + 2000 + ], + "score": 0.77, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 375, + 282, + 934, + 282, + 934, + 315, + 375, + 315 + ], + "score": 0.75, + "latex": "\\mathcal { T } = ( \\mathcal { V } , \\mathcal { E } ) , U = \\left( s _ { i n i t } , S _ { g o a l } , S , S _ { f r e e } , \\mathtt { m a p } , c ( \\cdot ) \\right)" + }, + { + "category_id": 14, + "poly": [ + 660, + 712, + 1057, + 712, + 1057, + 751, + 660, + 751 + ], + "score": 0.74, + "latex": "B ( s , \\eta ) = \\{ s ^ { \\prime } \\in S \\mid \\Vert s ^ { \\prime } - s \\Vert \\leqslant \\eta \\}" + }, + { + "category_id": 13, + "poly": [ + 381, + 419, + 531, + 419, + 531, + 443, + 381, + 443 + ], + "score": 0.74, + "latex": "s _ { p a r e n t } , s _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 386, + 557, + 548, + 557, + 548, + 587, + 386, + 587 + ], + "score": 0.73, + "latex": "\\sim \\phi ( s ) , s \\in \\mathcal { V }" + }, + { + "category_id": 13, + "poly": [ + 325, + 681, + 347, + 681, + 347, + 709, + 325, + 709 + ], + "score": 0.71, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1076, + 921, + 1091, + 921, + 1091, + 942, + 1076, + 942 + ], + "score": 0.7, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 295, + 239, + 698, + 239, + 698, + 274, + 295, + 274 + ], + "score": 0.63, + "latex": "\\mathbf { A l g o r i t h m 4 : R R T : : E x p a n d } ( { T } , { U } )" + }, + { + "category_id": 13, + "poly": [ + 1184, + 891, + 1201, + 891, + 1201, + 912, + 1184, + 912 + ], + "score": 0.59, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 357, + 783, + 427, + 783, + 427, + 813, + 357, + 813 + ], + "score": 0.54, + "latex": "\\mathrm { { R R T ^ { * } } }" + }, + { + "category_id": 13, + "poly": [ + 296, + 483, + 699, + 483, + 699, + 517, + 296, + 517 + ], + "score": 0.51, + "latex": "\\mathbf { A l g o r i t h m 5 : E S T : : E x p a n d } ( T , U )" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 474.0, + 703.0, + 474.0, + 703.0, + 525.0, + 700.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 231.0, + 703.0, + 231.0, + 703.0, + 282.0, + 699.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1339.0, + 1406.0, + 1339.0, + 1406.0, + 1378.0, + 292.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1371.0, + 1409.0, + 1371.0, + 1409.0, + 1409.0, + 292.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1404.0, + 1405.0, + 1404.0, + 1405.0, + 1438.0, + 295.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1433.0, + 1405.0, + 1433.0, + 1405.0, + 1470.0, + 292.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1464.0, + 1404.0, + 1464.0, + 1404.0, + 1498.0, + 295.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1492.0, + 1406.0, + 1492.0, + 1406.0, + 1529.0, + 294.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1526.0, + 1405.0, + 1526.0, + 1405.0, + 1560.0, + 295.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1555.0, + 471.0, + 1555.0, + 471.0, + 1590.0, + 295.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 987.0, + 786.0, + 987.0, + 786.0, + 1026.0, + 295.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 987.0, + 1406.0, + 987.0, + 1406.0, + 1026.0, + 817.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1020.0, + 1100.0, + 1020.0, + 1100.0, + 1054.0, + 296.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1020.0, + 1406.0, + 1020.0, + 1406.0, + 1054.0, + 1127.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1049.0, + 297.0, + 1049.0, + 297.0, + 1084.0, + 293.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1049.0, + 564.0, + 1049.0, + 564.0, + 1084.0, + 327.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1049.0, + 1406.0, + 1049.0, + 1406.0, + 1084.0, + 590.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1080.0, + 298.0, + 1080.0, + 298.0, + 1132.0, + 293.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1080.0, + 506.0, + 1080.0, + 506.0, + 1132.0, + 447.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1080.0, + 1410.0, + 1080.0, + 1410.0, + 1132.0, + 666.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1117.0, + 1407.0, + 1117.0, + 1407.0, + 1151.0, + 295.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1726.0, + 1406.0, + 1726.0, + 1406.0, + 1764.0, + 295.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1759.0, + 1405.0, + 1759.0, + 1405.0, + 1793.0, + 296.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1787.0, + 1151.0, + 1787.0, + 1151.0, + 1823.0, + 293.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1230.0, + 371.0, + 1230.0, + 371.0, + 1269.0, + 295.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1230.0, + 1403.0, + 1230.0, + 1403.0, + 1269.0, + 394.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1265.0, + 399.0, + 1265.0, + 399.0, + 1299.0, + 296.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1265.0, + 1404.0, + 1265.0, + 1404.0, + 1299.0, + 512.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1297.0, + 385.0, + 1297.0, + 385.0, + 1331.0, + 292.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1600.0, + 1404.0, + 1600.0, + 1404.0, + 1638.0, + 295.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1632.0, + 1403.0, + 1632.0, + 1403.0, + 1668.0, + 294.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1662.0, + 1019.0, + 1662.0, + 1019.0, + 1700.0, + 295.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 851.0, + 1404.0, + 851.0, + 1404.0, + 888.0, + 296.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 884.0, + 1183.0, + 884.0, + 1183.0, + 918.0, + 322.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 884.0, + 1405.0, + 884.0, + 1405.0, + 918.0, + 1202.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 915.0, + 1075.0, + 915.0, + 1075.0, + 949.0, + 324.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 915.0, + 1102.0, + 915.0, + 1102.0, + 949.0, + 1092.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1969.0, + 368.0, + 1969.0, + 368.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1969.0, + 1061.0, + 1969.0, + 1061.0, + 2007.0, + 483.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1969.0, + 1299.0, + 1969.0, + 1299.0, + 2007.0, + 1078.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1969.0, + 1405.0, + 1969.0, + 1405.0, + 2007.0, + 1318.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 2004.0, + 1217.0, + 2004.0, + 1217.0, + 2035.0, + 298.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 2004.0, + 1403.0, + 2004.0, + 1403.0, + 2035.0, + 1315.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 751.0, + 1405.0, + 751.0, + 1405.0, + 788.0, + 323.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 780.0, + 356.0, + 780.0, + 356.0, + 818.0, + 320.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 780.0, + 1404.0, + 780.0, + 1404.0, + 818.0, + 428.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 813.0, + 500.0, + 813.0, + 500.0, + 852.0, + 321.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 672.0, + 324.0, + 672.0, + 324.0, + 720.0, + 321.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 672.0, + 982.0, + 672.0, + 982.0, + 720.0, + 348.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 672.0, + 1331.0, + 672.0, + 1331.0, + 720.0, + 1011.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 672.0, + 1362.0, + 672.0, + 1362.0, + 720.0, + 1351.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 272.0, + 374.0, + 272.0, + 374.0, + 320.0, + 293.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 272.0, + 940.0, + 272.0, + 940.0, + 320.0, + 935.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 272.0, + 312.0, + 297.0, + 312.0, + 297.0, + 344.0, + 272.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 312.0, + 523.0, + 312.0, + 523.0, + 344.0, + 510.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 269.0, + 342.0, + 297.0, + 342.0, + 297.0, + 379.0, + 269.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 342.0, + 709.0, + 342.0, + 709.0, + 379.0, + 693.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 264.0, + 368.0, + 296.0, + 368.0, + 296.0, + 417.0, + 264.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 368.0, + 789.0, + 368.0, + 789.0, + 417.0, + 777.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1831.0, + 1284.0, + 1831.0, + 1284.0, + 1872.0, + 294.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1831.0, + 1404.0, + 1831.0, + 1404.0, + 1872.0, + 1307.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1859.0, + 396.0, + 1859.0, + 396.0, + 1900.0, + 293.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 1859.0, + 444.0, + 1859.0, + 444.0, + 1900.0, + 434.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 554.0, + 1361.0, + 554.0, + 1361.0, + 586.0, + 1120.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1087.0, + 584.0, + 1362.0, + 584.0, + 1362.0, + 616.0, + 1087.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 517.0, + 374.0, + 517.0, + 374.0, + 561.0, + 293.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 517.0, + 940.0, + 517.0, + 940.0, + 561.0, + 935.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 273.0, + 554.0, + 385.0, + 554.0, + 385.0, + 588.0, + 273.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 554.0, + 562.0, + 554.0, + 562.0, + 588.0, + 549.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 269.0, + 582.0, + 295.0, + 582.0, + 295.0, + 622.0, + 269.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 582.0, + 616.0, + 582.0, + 616.0, + 622.0, + 605.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 271.25, + 617.5, + 554.25, + 617.5, + 554.25, + 650.5, + 271.25, + 650.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 271.0, + 408.0, + 380.0, + 408.0, + 380.0, + 447.0, + 271.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 408.0, + 545.0, + 408.0, + 545.0, + 447.0, + 532.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 231.0, + 703.0, + 231.0, + 703.0, + 282.0, + 699.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 312.0, + 1363.0, + 312.0, + 1363.0, + 344.0, + 1026.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 340.0, + 1363.0, + 340.0, + 1363.0, + 374.0, + 1129.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 474.0, + 703.0, + 474.0, + 703.0, + 525.0, + 700.0, + 525.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 227, + 1406, + 227, + 1406, + 386, + 297, + 386 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 392, + 1430, + 1404, + 1430, + 1404, + 1560, + 392, + 1560 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 296, + 579, + 1405, + 579, + 1405, + 673, + 296, + 673 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 694, + 1704, + 1099, + 1704, + 1099, + 1875, + 694, + 1875 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1405, + 1971, + 1405, + 2036, + 298, + 2036 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 298, + 953, + 1401, + 953, + 1401, + 1019, + 298, + 1019 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 301, + 859, + 1399, + 859, + 1399, + 940, + 301, + 940 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 298, + 686, + 1401, + 686, + 1401, + 750, + 298, + 750 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 394, + 1879, + 1402, + 1879, + 1402, + 1943, + 394, + 1943 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 369, + 1576, + 1401, + 1576, + 1401, + 1639, + 369, + 1639 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 599, + 753, + 1098, + 753, + 1098, + 810, + 599, + 810 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 295, + 397, + 1401, + 397, + 1401, + 460, + 295, + 460 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 607, + 1348, + 1191, + 1348, + 1191, + 1428, + 607, + 1428 + ], + "score": 0.936 + }, + { + "category_id": 8, + "poly": [ + 752, + 1642, + 1044, + 1642, + 1044, + 1680, + 752, + 1680 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 297, + 1068, + 870, + 1068, + 870, + 1107, + 297, + 1107 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 298, + 811, + 1052, + 811, + 1052, + 844, + 298, + 844 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 295, + 1144, + 1219, + 1144, + 1219, + 1180, + 295, + 1180 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 396, + 1680, + 469, + 1680, + 469, + 1710, + 396, + 1710 + ], + "score": 0.914 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 816, + 74, + 816, + 105, + 299, + 105 + ], + "score": 0.913 + }, + { + "category_id": 9, + "poly": [ + 1352, + 765, + 1400, + 765, + 1400, + 797, + 1352, + 797 + ], + "score": 0.897 + }, + { + "category_id": 8, + "poly": [ + 659, + 518, + 1042, + 518, + 1042, + 574, + 659, + 574 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1352, + 529, + 1400, + 529, + 1400, + 560, + 1352, + 560 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1282, + 1400, + 1282, + 1400, + 1312, + 1352, + 1312 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1365, + 475, + 1400, + 475, + 1400, + 504, + 1365, + 504 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 863, + 2087, + 863, + 2112, + 835, + 2112 + ], + "score": 0.871 + }, + { + "category_id": 1, + "poly": [ + 396, + 1319, + 469, + 1319, + 469, + 1348, + 396, + 1348 + ], + "score": 0.864 + }, + { + "category_id": 8, + "poly": [ + 657, + 462, + 1016, + 462, + 1016, + 516, + 657, + 516 + ], + "score": 0.813 + }, + { + "category_id": 1, + "poly": [ + 371, + 1197, + 1404, + 1197, + 1404, + 1314, + 371, + 1314 + ], + "score": 0.753 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1644, + 1400, + 1644, + 1400, + 1675, + 1351, + 1675 + ], + "score": 0.734 + }, + { + "category_id": 8, + "poly": [ + 748, + 1282, + 1044, + 1282, + 1044, + 1317, + 748, + 1317 + ], + "score": 0.363 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1644, + 1400, + 1644, + 1400, + 1675, + 1351, + 1675 + ], + "score": 0.325 + }, + { + "category_id": 8, + "poly": [ + 654, + 463, + 1041, + 463, + 1041, + 574, + 654, + 574 + ], + "score": 0.117 + }, + { + "category_id": 13, + "poly": [ + 1183, + 321, + 1291, + 321, + 1291, + 355, + 1183, + 355 + ], + "score": 0.94, + "latex": "V ^ { * } ( s ^ { \\prime } | U )" + }, + { + "category_id": 14, + "poly": [ + 600, + 751, + 1096, + 751, + 1096, + 811, + 600, + 811 + ], + "score": 0.93, + "latex": "\\tilde { V } ^ { t } , \\tilde { c } ^ { t } = \\mathrm { L S T M } \\left( \\sum \\left( W _ { 1 } \\oplus \\left[ \\tilde { V } ^ { t } , \\tilde { R } \\right] \\right) , \\tilde { c } ^ { t } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 859, + 321, + 973, + 321, + 973, + 355, + 859, + 355 + ], + "score": 0.93, + "latex": "P ( s ^ { \\prime } | s , a )" + }, + { + "category_id": 13, + "poly": [ + 933, + 261, + 1046, + 261, + 1046, + 295, + 933, + 295 + ], + "score": 0.93, + "latex": "P ( s ^ { \\prime } | s , a )" + }, + { + "category_id": 13, + "poly": [ + 461, + 230, + 571, + 230, + 571, + 264, + 461, + 264 + ], + "score": 0.93, + "latex": "V ^ { * } ( s ^ { \\prime } | U )" + }, + { + "category_id": 13, + "poly": [ + 1034, + 321, + 1130, + 321, + 1130, + 355, + 1034, + 355 + ], + "score": 0.93, + "latex": "c ( [ s , s ^ { \\bar { \\prime } } ] )" + }, + { + "category_id": 13, + "poly": [ + 1092, + 856, + 1203, + 856, + 1203, + 914, + 1092, + 914 + ], + "score": 0.93, + "latex": "\\left\\{ \\tilde { R } , W _ { 1 } \\right\\}" + }, + { + "category_id": 14, + "poly": [ + 654, + 460, + 1044, + 460, + 1044, + 576, + 654, + 576 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\tilde { V } ^ { * 0 } = \\operatorname* { m i n } \\ \\left( W _ { 1 } \\oplus \\left[ \\tt { m a p } , \\tilde { R } \\right] \\right) } \\\\ & { \\tilde { V } ^ { * t } = \\operatorname* { m i n } \\ \\left( W _ { 1 } \\oplus \\left[ \\tilde { V } ^ { * t - 1 } , \\tilde { R } \\right] \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1230, + 260, + 1402, + 260, + 1402, + 295, + 1230, + 295 + ], + "score": 0.92, + "latex": "P ( s ^ { \\prime } | \\dot { s } , a ) = 0" + }, + { + "category_id": 13, + "poly": [ + 338, + 291, + 450, + 291, + 450, + 325, + 338, + 325 + ], + "score": 0.92, + "latex": "\\bar { \\boldsymbol { s } } ^ { \\prime } \\notin B ( \\boldsymbol { s } )" + }, + { + "category_id": 14, + "poly": [ + 695, + 1703, + 1100, + 1703, + 1100, + 1879, + 695, + 1879 + ], + "score": 0.92, + "latex": "\\begin{array} { r l r } { \\bar { r } _ { t } \\left[ s \\right] } & { = } & { \\frac { \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) r \\left( s ^ { \\prime } \\right) } { \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) } , } \\\\ { \\sigma _ { t } \\left( s \\right) } & { = } & { \\sqrt { \\frac { \\log \\sum _ { s ^ { \\prime } \\in S _ { t } } w \\left( s ^ { \\prime } \\right) } { w \\left( s \\right) } } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 715, + 612, + 783, + 612, + 783, + 640, + 715, + 640 + ], + "score": 0.91, + "latex": "k \\times k" + }, + { + "category_id": 13, + "poly": [ + 710, + 1230, + 799, + 1230, + 799, + 1264, + 710, + 1264 + ], + "score": 0.91, + "latex": "k \\left( s , s ^ { \\prime } \\right)" + }, + { + "category_id": 13, + "poly": [ + 454, + 1879, + 733, + 1879, + 733, + 1918, + 454, + 1918 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { w \\left( s \\right) = \\sum _ { s ^ { \\prime } \\in S _ { t } } k \\left( s ^ { \\prime } , s \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1031, + 1430, + 1271, + 1430, + 1271, + 1463, + 1031, + 1463 + ], + "score": 0.91, + "latex": "{ \\cal { S } } _ { t } = \\{ s _ { 1 } , s _ { 2 } , \\ldots , s _ { t } \\}" + }, + { + "category_id": 13, + "poly": [ + 1015, + 580, + 1082, + 580, + 1082, + 611, + 1015, + 611 + ], + "score": 0.89, + "latex": "d \\times d" + }, + { + "category_id": 13, + "poly": [ + 1236, + 613, + 1265, + 613, + 1265, + 642, + 1236, + 642 + ], + "score": 0.88, + "latex": "k _ { c }" + }, + { + "category_id": 13, + "poly": [ + 880, + 1230, + 1166, + 1230, + 1166, + 1264, + 880, + 1264 + ], + "score": 0.88, + "latex": "[ r ( s ) | \\mathcal { T } , U ] \\sim \\mathcal { \\bar { G } } \\mathcal { P } ( 0 , k )" + }, + { + "category_id": 13, + "poly": [ + 298, + 613, + 338, + 613, + 338, + 642, + 298, + 642 + ], + "score": 0.88, + "latex": "W _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 383, + 613, + 412, + 613, + 412, + 642, + 383, + 642 + ], + "score": 0.88, + "latex": "k _ { c }" + }, + { + "category_id": 13, + "poly": [ + 773, + 813, + 803, + 813, + 803, + 842, + 773, + 842 + ], + "score": 0.87, + "latex": "k _ { c }" + }, + { + "category_id": 13, + "poly": [ + 846, + 577, + 892, + 577, + 892, + 610, + 846, + 610 + ], + "score": 0.87, + "latex": "\\tilde { V } ^ { * t }" + }, + { + "category_id": 13, + "poly": [ + 945, + 577, + 970, + 577, + 970, + 609, + 945, + 609 + ], + "score": 0.86, + "latex": "\\tilde { R }" + }, + { + "category_id": 14, + "poly": [ + 751, + 1641, + 1045, + 1641, + 1045, + 1678, + 751, + 1678 + ], + "score": 0.85, + "latex": "\\phi \\left( s \\right) : = { \\bar { r } } _ { t } \\left( s \\right) + \\lambda \\sigma _ { t } \\left( s \\right) ," + }, + { + "category_id": 13, + "poly": [ + 372, + 585, + 398, + 585, + 398, + 611, + 372, + 611 + ], + "score": 0.83, + "latex": "\\oplus" + }, + { + "category_id": 13, + "poly": [ + 1018, + 1611, + 1031, + 1611, + 1031, + 1635, + 1018, + 1635 + ], + "score": 0.8, + "latex": "t" + }, + { + "category_id": 14, + "poly": [ + 604, + 1346, + 1191, + 1346, + 1191, + 1428, + 604, + 1428 + ], + "score": 0.8, + "latex": "\\begin{array} { r c l } { { { \\bar { r } } _ { t } } \\left( s \\right) } & { { = } } & { { { k _ { t } } \\left( s \\right) \\left( K _ { t } + \\alpha I \\right) ^ { - 1 } { r _ { t } } , } } \\\\ { { \\sigma _ { t } ^ { 2 } \\left( s \\right) } } & { { = } } & { { { k \\left( s , s \\right) - k _ { t } } \\left( s \\right) ^ { \\top } \\left( K _ { t } + \\alpha I \\right) ^ { - 1 } k _ { t } \\left( s \\right) , } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1139, + 1466, + 1209, + 1466, + 1209, + 1500, + 1139, + 1500 + ], + "score": 0.75, + "latex": "\\sigma _ { t } ^ { 2 } \\left( s \\right)" + }, + { + "category_id": 13, + "poly": [ + 1385, + 1503, + 1401, + 1503, + 1401, + 1524, + 1385, + 1524 + ], + "score": 0.75, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 450, + 1429, + 714, + 1429, + 714, + 1467, + 450, + 1467 + ], + "score": 0.75, + "latex": "\\boldsymbol { k } _ { t } \\left( \\boldsymbol { s } \\right) = \\left[ k \\left( s _ { i } , \\boldsymbol { s } \\right) \\right] _ { s _ { i } \\in \\boldsymbol { S } _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1317, + 299, + 1334, + 299, + 1334, + 319, + 1317, + 319 + ], + "score": 0.73, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 705, + 1265, + 719, + 1265, + 719, + 1289, + 705, + 1289 + ], + "score": 0.73, + "latex": "t { \\cdot }" + }, + { + "category_id": 14, + "poly": [ + 753, + 1270, + 1043, + 1270, + 1043, + 1316, + 753, + 1316 + ], + "score": 0.67, + "latex": "\\begin{array} { r } { \\mathrm { \\ i p \\ a s } } \\\\ { \\phi \\left( s \\right) : = { { \\bar { r } } _ { t } } \\left( s \\right) + \\lambda { { \\sigma } _ { t } } \\left( s \\right) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 725, + 1429, + 976, + 1429, + 976, + 1467, + 725, + 1467 + ], + "score": 0.57, + "latex": "K _ { t } = [ k \\left( s , s ^ { \\prime } \\right) ] _ { s , s ^ { \\prime } \\in S _ { t } }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1066.0, + 873.0, + 1066.0, + 873.0, + 1113.0, + 295.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 460.0, + 230.0, + 460.0, + 263.0, + 295.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 263.0, + 572.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 259.0, + 932.0, + 259.0, + 932.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 259.0, + 1229.0, + 259.0, + 1229.0, + 296.0, + 1047.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 259.0, + 1407.0, + 259.0, + 1407.0, + 296.0, + 1403.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 289.0, + 337.0, + 289.0, + 337.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 289.0, + 1316.0, + 289.0, + 1316.0, + 327.0, + 451.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 289.0, + 1406.0, + 289.0, + 1406.0, + 327.0, + 1335.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 858.0, + 319.0, + 858.0, + 356.0, + 294.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 319.0, + 1033.0, + 319.0, + 1033.0, + 356.0, + 974.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 319.0, + 1182.0, + 319.0, + 1182.0, + 356.0, + 1131.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 319.0, + 1403.0, + 319.0, + 1403.0, + 356.0, + 1292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 352.0, + 1047.0, + 352.0, + 1047.0, + 387.0, + 292.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1422.0, + 449.0, + 1422.0, + 449.0, + 1472.0, + 390.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1422.0, + 724.0, + 1422.0, + 724.0, + 1472.0, + 715.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 1422.0, + 1030.0, + 1422.0, + 1030.0, + 1472.0, + 977.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1422.0, + 1408.0, + 1422.0, + 1408.0, + 1472.0, + 1272.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1464.0, + 1138.0, + 1464.0, + 1138.0, + 1501.0, + 393.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1464.0, + 1405.0, + 1464.0, + 1405.0, + 1501.0, + 1210.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1497.0, + 1384.0, + 1497.0, + 1384.0, + 1531.0, + 395.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1497.0, + 1405.0, + 1497.0, + 1405.0, + 1531.0, + 1402.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1527.0, + 939.0, + 1527.0, + 939.0, + 1561.0, + 394.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 574.0, + 371.0, + 574.0, + 371.0, + 619.0, + 291.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 574.0, + 845.0, + 574.0, + 845.0, + 619.0, + 399.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 574.0, + 944.0, + 574.0, + 944.0, + 619.0, + 893.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 971.0, + 574.0, + 1014.0, + 574.0, + 1014.0, + 619.0, + 971.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 574.0, + 1407.0, + 574.0, + 1407.0, + 619.0, + 1083.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 609.0, + 382.0, + 609.0, + 382.0, + 645.0, + 339.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 609.0, + 714.0, + 609.0, + 714.0, + 645.0, + 413.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 609.0, + 1235.0, + 609.0, + 1235.0, + 645.0, + 784.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 609.0, + 1403.0, + 609.0, + 1403.0, + 645.0, + 1266.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 637.0, + 411.0, + 637.0, + 411.0, + 675.0, + 293.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1407.0, + 1971.0, + 1407.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2002.0, + 1205.0, + 2002.0, + 1205.0, + 2034.0, + 296.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 952.0, + 1405.0, + 952.0, + 1405.0, + 990.0, + 294.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 985.0, + 1406.0, + 985.0, + 1406.0, + 1021.0, + 296.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 845.0, + 1091.0, + 845.0, + 1091.0, + 919.0, + 286.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 845.0, + 1411.0, + 845.0, + 1411.0, + 919.0, + 1204.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 902.0, + 799.0, + 902.0, + 799.0, + 946.0, + 294.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 687.0, + 1402.0, + 687.0, + 1402.0, + 723.0, + 296.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 717.0, + 883.0, + 717.0, + 883.0, + 756.0, + 294.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1871.0, + 453.0, + 1871.0, + 453.0, + 1923.0, + 390.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1871.0, + 1407.0, + 1871.0, + 1407.0, + 1923.0, + 734.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1910.0, + 817.0, + 1910.0, + 817.0, + 1946.0, + 395.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1574.0, + 1404.0, + 1574.0, + 1404.0, + 1613.0, + 369.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1605.0, + 1017.0, + 1605.0, + 1017.0, + 1643.0, + 393.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1605.0, + 1120.0, + 1605.0, + 1120.0, + 1643.0, + 1032.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 394.0, + 1405.0, + 394.0, + 1405.0, + 437.0, + 294.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 428.0, + 678.0, + 428.0, + 678.0, + 463.0, + 295.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 810.0, + 772.0, + 810.0, + 772.0, + 846.0, + 295.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 810.0, + 1053.0, + 810.0, + 1053.0, + 846.0, + 804.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1144.0, + 1220.0, + 1144.0, + 1220.0, + 1180.0, + 294.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1675.0, + 474.0, + 1675.0, + 474.0, + 1714.0, + 393.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1314.0, + 475.0, + 1314.0, + 475.0, + 1352.0, + 391.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1197.0, + 1404.0, + 1197.0, + 1404.0, + 1234.0, + 366.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1226.0, + 709.0, + 1226.0, + 709.0, + 1267.0, + 391.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1226.0, + 879.0, + 1226.0, + 879.0, + 1267.0, + 800.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1226.0, + 1407.0, + 1226.0, + 1407.0, + 1267.0, + 1167.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1278.0, + 1409.0, + 1278.0, + 1409.0, + 1319.0, + 1345.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1254.5, + 1054.0, + 1254.5, + 1054.0, + 1319.5, + 385.0, + 1319.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1604, + 1405, + 1604, + 1405, + 1788, + 297, + 1788 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 328, + 457, + 1373, + 457, + 1373, + 1509, + 328, + 1509 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1803, + 1404, + 1803, + 1404, + 1926, + 298, + 1926 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1403, + 1942, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 359, + 1407, + 359, + 1407, + 452, + 298, + 452 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 297, + 1530, + 1403, + 1530, + 1403, + 1593, + 297, + 1593 + ], + "score": 0.931 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 296, + 225, + 1040, + 225, + 1040, + 263, + 296, + 263 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.851 + }, + { + "category_id": 13, + "poly": [ + 974, + 2002, + 1293, + 2002, + 1293, + 2036, + 974, + 2036 + ], + "score": 0.92, + "latex": "( d , d _ { e } , d _ { a } , p ) = ( 1 \\bar { 5 } , 6 4 , \\bar { 8 } , 8 )" + }, + { + "category_id": 13, + "poly": [ + 944, + 1804, + 1070, + 1804, + 1070, + 1834, + 944, + 1834 + ], + "score": 0.92, + "latex": "d \\times d \\times d _ { a }" + }, + { + "category_id": 13, + "poly": [ + 1256, + 1973, + 1312, + 1973, + 1312, + 2005, + 1256, + 2005 + ], + "score": 0.9, + "latex": "\\psi ( s )" + }, + { + "category_id": 13, + "poly": [ + 413, + 1727, + 477, + 1727, + 477, + 1755, + 413, + 1755 + ], + "score": 0.89, + "latex": "1 \\times 1" + }, + { + "category_id": 13, + "poly": [ + 913, + 1865, + 979, + 1865, + 979, + 1893, + 913, + 1893 + ], + "score": 0.89, + "latex": "d \\times d" + }, + { + "category_id": 13, + "poly": [ + 828, + 1866, + 857, + 1866, + 857, + 1895, + 828, + 1895 + ], + "score": 0.87, + "latex": "d _ { e }" + }, + { + "category_id": 13, + "poly": [ + 1121, + 1696, + 1185, + 1696, + 1185, + 1725, + 1121, + 1725 + ], + "score": 0.83, + "latex": "1 \\times 1" + }, + { + "category_id": 13, + "poly": [ + 1127, + 893, + 1214, + 893, + 1214, + 912, + 1127, + 912 + ], + "score": 0.82, + "latex": "c _ { 0 } \\in \\mathbb { R } ^ { d \\times d \\times d _ { e } }" + }, + { + "category_id": 13, + "poly": [ + 382, + 706, + 501, + 706, + 501, + 727, + 382, + 727 + ], + "score": 0.37, + "latex": "f _ { 0 } ^ { w } ( s ^ { w } ) \\in \\mathbb { R } ^ { d \\times d \\times 4 }" + }, + { + "category_id": 13, + "poly": [ + 904, + 893, + 994, + 893, + 994, + 912, + 904, + 912 + ], + "score": 0.37, + "latex": "h _ { 0 } \\in \\mathbb { R } ^ { d \\times d \\times d _ { e } }" + }, + { + "category_id": 13, + "poly": [ + 390, + 1234, + 499, + 1234, + 499, + 1254, + 390, + 1254 + ], + "score": 0.3, + "latex": "\\mu _ { \\theta } ^ { w } ( s ^ { w } ) \\in \\mathbb { R } ^ { d \\times d }" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 469.0, + 648.0, + 469.0, + 648.0, + 499.0, + 449.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 471.0, + 1152.0, + 471.0, + 1152.0, + 498.0, + 968.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 473.0, + 1333.0, + 473.0, + 1333.0, + 494.0, + 1208.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 501.0, + 1069.0, + 501.0, + 1069.0, + 517.0, + 1053.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 501.0, + 1280.0, + 501.0, + 1280.0, + 517.0, + 1264.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 537.0, + 1120.0, + 537.0, + 1120.0, + 563.0, + 1003.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 538.0, + 1330.0, + 538.0, + 1330.0, + 561.0, + 1214.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 570.0, + 451.0, + 570.0, + 451.0, + 579.0, + 440.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 583.0, + 481.0, + 583.0, + 481.0, + 613.0, + 403.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 581.0, + 687.0, + 581.0, + 687.0, + 613.0, + 612.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 586.0, + 1069.0, + 586.0, + 1069.0, + 601.0, + 1054.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 585.0, + 1279.0, + 585.0, + 1279.0, + 601.0, + 1264.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 601.0, + 1120.0, + 601.0, + 1120.0, + 632.0, + 1004.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 606.0, + 1323.0, + 606.0, + 1323.0, + 630.0, + 1223.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 617.0, + 455.0, + 617.0, + 455.0, + 639.0, + 435.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 642.0, + 507.0, + 642.0, + 507.0, + 670.0, + 383.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 653.0, + 1070.0, + 653.0, + 1070.0, + 673.0, + 1052.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 687.0, + 450.0, + 687.0, + 450.0, + 697.0, + 439.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 678.0, + 1087.0, + 678.0, + 1087.0, + 707.0, + 1035.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 699.0, + 381.0, + 699.0, + 381.0, + 729.0, + 374.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 699.0, + 510.0, + 699.0, + 510.0, + 729.0, + 502.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 725.0, + 1068.0, + 725.0, + 1068.0, + 733.0, + 1058.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 742.0, + 450.0, + 742.0, + 450.0, + 751.0, + 440.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 761.0, + 513.0, + 761.0, + 513.0, + 789.0, + 376.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 801.0, + 455.0, + 801.0, + 455.0, + 820.0, + 435.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 826.0, + 514.0, + 826.0, + 514.0, + 852.0, + 376.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 818.0, + 1006.0, + 818.0, + 1006.0, + 854.0, + 897.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 825.0, + 1225.0, + 825.0, + 1225.0, + 852.0, + 1120.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 863.0, + 453.0, + 863.0, + 453.0, + 883.0, + 435.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 873.0, + 958.0, + 873.0, + 958.0, + 882.0, + 948.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 871.0, + 1178.0, + 871.0, + 1178.0, + 882.0, + 1168.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 889.0, + 513.0, + 889.0, + 513.0, + 917.0, + 375.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 933.0, + 452.0, + 933.0, + 452.0, + 948.0, + 437.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 935.0, + 1071.0, + 935.0, + 1071.0, + 957.0, + 1051.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 955.0, + 514.0, + 955.0, + 514.0, + 982.0, + 376.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 991.0, + 455.0, + 991.0, + 455.0, + 1013.0, + 435.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 993.0, + 915.0, + 993.0, + 915.0, + 1003.0, + 905.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 985.0, + 1119.0, + 985.0, + 1119.0, + 1012.0, + 1003.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1019.0, + 513.0, + 1019.0, + 513.0, + 1045.0, + 375.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1015.0, + 706.0, + 1015.0, + 706.0, + 1049.0, + 597.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1007.0, + 893.0, + 1007.0, + 893.0, + 1035.0, + 799.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1045.0, + 1070.0, + 1045.0, + 1070.0, + 1066.0, + 1052.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1057.0, + 455.0, + 1057.0, + 455.0, + 1078.0, + 435.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1055.0, + 662.0, + 1055.0, + 662.0, + 1078.0, + 644.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1080.0, + 496.0, + 1080.0, + 496.0, + 1111.0, + 394.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1079.0, + 693.0, + 1079.0, + 693.0, + 1115.0, + 612.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1072.0, + 1190.0, + 1072.0, + 1190.0, + 1099.0, + 930.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1121.0, + 455.0, + 1121.0, + 455.0, + 1141.0, + 435.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1117.0, + 664.0, + 1117.0, + 664.0, + 1143.0, + 641.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 1144.0, + 480.0, + 1144.0, + 480.0, + 1176.0, + 410.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1144.0, + 687.0, + 1144.0, + 687.0, + 1176.0, + 618.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1146.0, + 1150.0, + 1146.0, + 1150.0, + 1189.0, + 975.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1162.0, + 1176.0, + 1162.0, + 1176.0, + 1170.0, + 1167.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1185.0, + 455.0, + 1185.0, + 455.0, + 1205.0, + 435.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1185.0, + 663.0, + 1185.0, + 663.0, + 1205.0, + 641.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1208.0, + 507.0, + 1208.0, + 507.0, + 1239.0, + 380.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1208.0, + 716.0, + 1208.0, + 716.0, + 1239.0, + 588.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1205.0, + 1069.0, + 1205.0, + 1069.0, + 1221.0, + 1054.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1228.0, + 389.0, + 1228.0, + 389.0, + 1254.0, + 384.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1226.0, + 706.0, + 1226.0, + 706.0, + 1259.0, + 597.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1221.0, + 1100.0, + 1221.0, + 1100.0, + 1251.0, + 1017.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1264.0, + 1068.0, + 1264.0, + 1068.0, + 1273.0, + 1056.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1301.0, + 554.0, + 1301.0, + 554.0, + 1312.0, + 543.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1283.0, + 1114.0, + 1283.0, + 1114.0, + 1311.0, + 1008.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 1315.0, + 596.0, + 1315.0, + 596.0, + 1339.0, + 501.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1318.0, + 757.0, + 1318.0, + 757.0, + 1349.0, + 674.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1319.0, + 1070.0, + 1319.0, + 1070.0, + 1339.0, + 1052.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1334.0, + 579.0, + 1334.0, + 579.0, + 1358.0, + 515.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1341.0, + 752.0, + 1341.0, + 752.0, + 1369.0, + 681.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1347.0, + 1115.0, + 1347.0, + 1115.0, + 1374.0, + 1008.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1393.0, + 959.0, + 1393.0, + 959.0, + 1410.0, + 945.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1393.0, + 1182.0, + 1393.0, + 1182.0, + 1410.0, + 1167.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1416.0, + 987.0, + 1416.0, + 987.0, + 1448.0, + 916.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1419.0, + 1207.0, + 1419.0, + 1207.0, + 1446.0, + 1140.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1433.0, + 578.0, + 1433.0, + 578.0, + 1461.0, + 519.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 1450.0, + 659.0, + 1450.0, + 659.0, + 1478.0, + 434.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1461.0, + 959.0, + 1461.0, + 959.0, + 1475.0, + 945.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1458.0, + 1183.0, + 1458.0, + 1183.0, + 1474.0, + 1167.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 1477.0, + 1016.0, + 1477.0, + 1016.0, + 1505.0, + 886.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1475.0, + 1231.0, + 1475.0, + 1231.0, + 1506.0, + 1116.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 468.5, + 944.0, + 468.5, + 944.0, + 500.0, + 774.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 620.0, + 1136.0, + 620.0, + 1136.0, + 651.0, + 985.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.75, + 621.5, + 1332.75, + 621.5, + 1332.75, + 648.5, + 1207.75, + 648.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 738.0, + 1166.0, + 738.0, + 1166.0, + 769.0, + 955.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 883.5, + 1002.0, + 883.5, + 1002.0, + 914.5, + 895.0, + 914.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 884.0, + 1224.0, + 884.0, + 1224.0, + 914.0, + 1118.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1528.0, + 1409.0, + 1528.0, + 1409.0, + 1568.0, + 293.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1562.0, + 488.0, + 1562.0, + 488.0, + 1596.0, + 292.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 222.0, + 1043.0, + 222.0, + 1043.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2122.0, + 831.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1602.0, + 1405.0, + 1602.0, + 1405.0, + 1639.0, + 292.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1634.0, + 1406.0, + 1634.0, + 1406.0, + 1668.0, + 294.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1666.0, + 1406.0, + 1666.0, + 1406.0, + 1701.0, + 295.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1695.0, + 1120.0, + 1695.0, + 1120.0, + 1730.0, + 295.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1695.0, + 1407.0, + 1695.0, + 1407.0, + 1730.0, + 1186.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1726.0, + 412.0, + 1726.0, + 412.0, + 1762.0, + 295.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1726.0, + 1410.0, + 1726.0, + 1410.0, + 1762.0, + 478.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1758.0, + 1035.0, + 1758.0, + 1035.0, + 1790.0, + 295.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 943.0, + 1803.0, + 943.0, + 1835.0, + 294.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1803.0, + 1404.0, + 1803.0, + 1404.0, + 1835.0, + 1071.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 1404.0, + 1835.0, + 1404.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1866.0, + 827.0, + 1866.0, + 827.0, + 1898.0, + 294.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1866.0, + 912.0, + 1866.0, + 912.0, + 1898.0, + 858.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1866.0, + 1404.0, + 1866.0, + 1404.0, + 1898.0, + 980.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1896.0, + 804.0, + 1896.0, + 804.0, + 1926.0, + 294.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1938.0, + 1408.0, + 1938.0, + 1408.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1255.0, + 1973.0, + 1255.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 1313.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 973.0, + 2002.0, + 973.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 2002.0, + 1306.0, + 2002.0, + 1306.0, + 2039.0, + 1294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 357.0, + 1407.0, + 357.0, + 1407.0, + 396.0, + 292.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 387.0, + 1407.0, + 387.0, + 1407.0, + 427.0, + 292.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 422.0, + 728.0, + 422.0, + 728.0, + 456.0, + 296.0, + 456.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1405, + 1941, + 1405, + 2035, + 299, + 2035 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1369, + 1405, + 1369, + 1405, + 1523, + 298, + 1523 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 637, + 227, + 1059, + 227, + 1059, + 664, + 637, + 664 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 296, + 862, + 1406, + 862, + 1406, + 987, + 296, + 987 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 297, + 1262, + 1404, + 1262, + 1404, + 1354, + 297, + 1354 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 297, + 1001, + 1406, + 1001, + 1406, + 1250, + 297, + 1250 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 297, + 1625, + 1401, + 1625, + 1401, + 1688, + 297, + 1688 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 299, + 1807, + 706, + 1807, + 706, + 1840, + 299, + 1840 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 297, + 730, + 678, + 730, + 678, + 767, + 297, + 767 + ], + "score": 0.917 + }, + { + "category_id": 4, + "poly": [ + 692, + 690, + 1007, + 690, + 1007, + 722, + 692, + 722 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 300, + 803, + 727, + 803, + 727, + 835, + 300, + 835 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 300, + 1566, + 729, + 1566, + 729, + 1598, + 300, + 1598 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 301, + 1881, + 772, + 1881, + 772, + 1914, + 301, + 1914 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.862 + }, + { + "category_id": 8, + "poly": [ + 514, + 1695, + 1180, + 1695, + 1180, + 1803, + 514, + 1803 + ], + "score": 0.803 + }, + { + "category_id": 1, + "poly": [ + 514, + 1695, + 1180, + 1695, + 1180, + 1803, + 514, + 1803 + ], + "score": 0.121 + }, + { + "category_id": 14, + "poly": [ + 514, + 1694, + 1179, + 1694, + 1179, + 1804, + 514, + 1804 + ], + "score": 0.93, + "latex": "\\epsilon = \\left\\{ \\begin{array} { l l } { 1 , } & { \\mathrm { i f ~ } i < 1 0 0 0 , } \\\\ { 0 . 5 - 0 . 1 \\cdot \\lfloor ( i - 1 0 0 0 ) / 2 0 0 \\rfloor , } & { \\mathrm { i f ~ } 1 0 0 0 \\leqslant i < 2 0 0 0 , } \\\\ { 0 . 1 , } & { \\mathrm { o t h e r w i s e , } } \\end{array} \\right." + }, + { + "category_id": 13, + "poly": [ + 1076, + 1461, + 1163, + 1461, + 1163, + 1495, + 1076, + 1495 + ], + "score": 0.93, + "latex": "[ 0 , \\pi / 2 ]" + }, + { + "category_id": 13, + "poly": [ + 326, + 1215, + 466, + 1215, + 466, + 1250, + 326, + 1250 + ], + "score": 0.91, + "latex": "[ - \\pi / 4 , \\pi / 4 ]" + }, + { + "category_id": 13, + "poly": [ + 531, + 1292, + 616, + 1292, + 616, + 1323, + 531, + 1323 + ], + "score": 0.9, + "latex": "\\approx 2 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 353, + 1811, + 365, + 1811, + 365, + 1834, + 353, + 1834 + ], + "score": 0.79, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 931, + 1663, + 947, + 1663, + 947, + 1684, + 931, + 1684 + ], + "score": 0.62, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1386, + 1633, + 1402, + 1633, + 1402, + 1654, + 1386, + 1654 + ], + "score": 0.5, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 970, + 1695, + 1079, + 1695, + 1079, + 1727, + 970, + 1727 + ], + "score": 0.36, + "latex": "i < 1 0 0 0" + }, + { + "category_id": 13, + "poly": [ + 871, + 507, + 909, + 507, + 909, + 524, + 871, + 524 + ], + "score": 0.34, + "latex": "d \\times d" + }, + { + "category_id": 13, + "poly": [ + 962, + 489, + 984, + 489, + 984, + 508, + 962, + 508 + ], + "score": 0.33, + "latex": "d _ { e } )" + }, + { + "category_id": 13, + "poly": [ + 928, + 241, + 1012, + 241, + 1012, + 261, + 928, + 261 + ], + "score": 0.3, + "latex": "c _ { i } \\in \\mathbb { R } ^ { d \\times d \\times d _ { e } }" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 309.0, + 1040.0, + 309.0, + 1040.0, + 345.0, + 961.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 329.0, + 816.0, + 329.0, + 816.0, + 365.0, + 707.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 335.0, + 1035.0, + 335.0, + 1035.0, + 362.0, + 966.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 383.0, + 768.0, + 383.0, + 768.0, + 396.0, + 755.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 400.0, + 851.0, + 400.0, + 851.0, + 424.0, + 670.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 456.0, + 862.0, + 456.0, + 862.0, + 466.0, + 854.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 485.0, + 961.0, + 485.0, + 961.0, + 513.0, + 724.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 485.0, + 991.0, + 485.0, + 991.0, + 513.0, + 985.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 505.0, + 870.0, + 505.0, + 870.0, + 526.0, + 802.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 505.0, + 914.0, + 505.0, + 914.0, + 526.0, + 910.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 616.0, + 769.0, + 616.0, + 769.0, + 628.0, + 755.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 617.0, + 957.0, + 617.0, + 957.0, + 626.0, + 947.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 631.0, + 834.0, + 631.0, + 834.0, + 654.0, + 688.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 632.0, + 1022.0, + 632.0, + 1022.0, + 654.0, + 879.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 238.0, + 829.0, + 238.0, + 829.0, + 262.5, + 693.0, + 262.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.75, + 237.0, + 1017.75, + 237.0, + 1017.75, + 262.5, + 882.75, + 262.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 728.0, + 681.0, + 728.0, + 681.0, + 772.0, + 293.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 687.0, + 1008.0, + 687.0, + 1008.0, + 727.0, + 691.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 801.0, + 731.0, + 801.0, + 731.0, + 839.0, + 294.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1567.0, + 733.0, + 1567.0, + 733.0, + 1602.0, + 295.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 775.0, + 1881.0, + 775.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1409.0, + 1941.0, + 1409.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1369.0, + 1403.0, + 1369.0, + 1403.0, + 1402.0, + 300.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1401.0, + 1403.0, + 1401.0, + 1403.0, + 1433.0, + 321.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1431.0, + 1405.0, + 1431.0, + 1405.0, + 1465.0, + 321.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1459.0, + 1075.0, + 1459.0, + 1075.0, + 1498.0, + 318.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1459.0, + 1407.0, + 1459.0, + 1407.0, + 1498.0, + 1164.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1495.0, + 482.0, + 1495.0, + 482.0, + 1524.0, + 322.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 864.0, + 1404.0, + 864.0, + 1404.0, + 896.0, + 295.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 895.0, + 1408.0, + 895.0, + 1408.0, + 928.0, + 294.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 923.0, + 1408.0, + 923.0, + 1408.0, + 960.0, + 293.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 956.0, + 1408.0, + 956.0, + 1408.0, + 989.0, + 296.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1260.0, + 1404.0, + 1260.0, + 1404.0, + 1296.0, + 294.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1290.0, + 530.0, + 1290.0, + 530.0, + 1328.0, + 294.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1290.0, + 1404.0, + 1290.0, + 1404.0, + 1328.0, + 617.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1322.0, + 964.0, + 1322.0, + 964.0, + 1356.0, + 295.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1001.0, + 1406.0, + 1001.0, + 1406.0, + 1038.0, + 295.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1033.0, + 1112.0, + 1033.0, + 1112.0, + 1067.0, + 318.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1064.0, + 1404.0, + 1064.0, + 1404.0, + 1098.0, + 300.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1092.0, + 1404.0, + 1092.0, + 1404.0, + 1130.0, + 320.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1123.0, + 784.0, + 1123.0, + 784.0, + 1159.0, + 318.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1152.0, + 1406.0, + 1152.0, + 1406.0, + 1189.0, + 300.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1185.0, + 1406.0, + 1185.0, + 1406.0, + 1219.0, + 323.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1214.0, + 477.0, + 1214.0, + 477.0, + 1253.0, + 467.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1626.0, + 1385.0, + 1626.0, + 1385.0, + 1662.0, + 296.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1655.0, + 930.0, + 1655.0, + 930.0, + 1695.0, + 295.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1655.0, + 1230.0, + 1655.0, + 1230.0, + 1695.0, + 948.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1801.0, + 352.0, + 1801.0, + 352.0, + 1846.0, + 294.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1801.0, + 708.0, + 1801.0, + 708.0, + 1846.0, + 366.0, + 1846.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1404, + 1941, + 1404, + 2034, + 299, + 2034 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 359, + 235, + 1346, + 235, + 1346, + 693, + 359, + 693 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 344, + 1509, + 1362, + 1509, + 1362, + 1744, + 344, + 1744 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 298, + 778, + 1405, + 778, + 1405, + 872, + 298, + 872 + ], + "score": 0.957 + }, + { + "category_id": 3, + "poly": [ + 328, + 1022, + 1374, + 1022, + 1374, + 1372, + 328, + 1372 + ], + "score": 0.955 + }, + { + "category_id": 4, + "poly": [ + 623, + 732, + 1075, + 732, + 1075, + 765, + 623, + 765 + ], + "score": 0.918 + }, + { + "category_id": 4, + "poly": [ + 298, + 1784, + 1401, + 1784, + 1401, + 1849, + 298, + 1849 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 297, + 1884, + 858, + 1884, + 858, + 1917, + 297, + 1917 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 866, + 2088, + 866, + 2113, + 835, + 2113 + ], + "score": 0.86 + }, + { + "category_id": 4, + "poly": [ + 298, + 1398, + 1404, + 1398, + 1404, + 1493, + 298, + 1493 + ], + "score": 0.846 + }, + { + "category_id": 0, + "poly": [ + 300, + 912, + 681, + 912, + 681, + 948, + 300, + 948 + ], + "score": 0.204 + }, + { + "category_id": 8, + "poly": [ + 300, + 912, + 681, + 912, + 681, + 948, + 300, + 948 + ], + "score": 0.111 + }, + { + "category_id": 14, + "poly": [ + 1076, + 459, + 1340, + 459, + 1340, + 582, + 1076, + 582 + ], + "score": 0.37, + "latex": "\\begin{array} { r } { \\stackrel { \\displaystyle et { } { ' } \\tilde { V } ^ { * } ( s ) \\mathrm { V a l u e } } { \\displaystyle et { } { ' } } } \\\\ { { \\displaystyle \\tilde { \\pi } ^ { * } ( s ^ { \\prime } | \\psi ( s ) ) \\mathrm { A c t i o n } } } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 237.0, + 464.0, + 237.0, + 464.0, + 280.0, + 352.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 263.0, + 437.0, + 263.0, + 437.0, + 307.0, + 379.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 358.0, + 443.0, + 358.0, + 443.0, + 393.0, + 371.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 369.0, + 483.0, + 369.0, + 483.0, + 387.0, + 464.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 383.0, + 449.0, + 383.0, + 449.0, + 418.0, + 364.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 385.0, + 726.0, + 385.0, + 726.0, + 421.0, + 644.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 470.0, + 789.0, + 470.0, + 789.0, + 502.0, + 738.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 484.0, + 449.0, + 484.0, + 449.0, + 543.0, + 364.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 495.0, + 805.0, + 495.0, + 805.0, + 554.0, + 739.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 500.0, + 1025.0, + 500.0, + 1025.0, + 540.0, + 953.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 535.0, + 1348.0, + 535.0, + 1348.0, + 586.0, + 1341.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 586.0, + 789.0, + 586.0, + 789.0, + 620.0, + 738.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 610.0, + 796.0, + 610.0, + 796.0, + 647.0, + 740.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 636.0, + 798.0, + 636.0, + 798.0, + 671.0, + 739.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1570.0, + 1245.0, + 1570.0, + 1245.0, + 1578.0, + 1222.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1587.0, + 1228.0, + 1587.0, + 1228.0, + 1609.0, + 1199.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1079.0, + 444.0, + 1079.0, + 444.0, + 1140.0, + 391.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1169.0, + 468.0, + 1169.0, + 468.0, + 1177.0, + 451.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1178.0, + 469.0, + 1178.0, + 469.0, + 1186.0, + 457.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1195.0, + 512.0, + 1195.0, + 512.0, + 1211.0, + 485.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 1240.0, + 859.0, + 1240.0, + 859.0, + 1252.0, + 834.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1272.0, + 634.0, + 1272.0, + 634.0, + 1301.0, + 486.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1279.0, + 1122.0, + 1279.0, + 1122.0, + 1295.0, + 1080.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.75, + 1185.0, + 481.75, + 1185.0, + 481.75, + 1198.0, + 463.75, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.25, + 1178.0, + 595.25, + 1178.0, + 595.25, + 1197.5, + 501.25, + 1197.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1252.0, + 894.0, + 1252.0, + 894.0, + 1306.0, + 792.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1263.0, + 942.0, + 1263.0, + 942.0, + 1299.5, + 868.0, + 1299.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.25, + 1265.0, + 470.25, + 1265.0, + 470.25, + 1310.5, + 371.25, + 1310.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 729.0, + 1077.0, + 729.0, + 1077.0, + 768.0, + 622.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1783.0, + 1403.0, + 1783.0, + 1403.0, + 1819.0, + 294.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1815.0, + 1051.0, + 1815.0, + 1051.0, + 1851.0, + 291.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1884.0, + 859.0, + 1884.0, + 859.0, + 1920.0, + 294.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1405.0, + 1396.0, + 1405.0, + 1437.0, + 293.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1428.0, + 1404.0, + 1428.0, + 1404.0, + 1465.0, + 293.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1460.0, + 748.0, + 1460.0, + 748.0, + 1498.0, + 294.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 911.0, + 683.0, + 911.0, + 683.0, + 951.0, + 293.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 779.0, + 1407.0, + 779.0, + 1407.0, + 813.0, + 296.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 804.0, + 1406.0, + 804.0, + 1406.0, + 847.0, + 292.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 840.0, + 980.0, + 840.0, + 980.0, + 873.0, + 294.0, + 873.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1287, + 1405, + 1287, + 1405, + 1570, + 298, + 1570 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 307, + 375, + 1388, + 375, + 1388, + 524, + 307, + 524 + ], + "score": 0.978, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.9880.9810.7180.6320.7350.7100.1850.5350.720
3d0.9430.8410.6890.5540.4900.5140.1210.1140.498
5d0.8830.7680.6330.5150.4550.4970.0300.4760.444
7d0.9310.9060.6340.3690.3610.8140.2880.2720.370
" + }, + { + "category_id": 5, + "poly": [ + 307, + 638, + 1388, + 638, + 1388, + 786, + 307, + 786 + ], + "score": 0.976, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1770.2432.3423.4841.0005.9459.2471.9831.011
3d0.6941.3342.2143.1251.0007.9247.2922.1620.988
5d0.8881.5201.8002.7061.0007.4835.7581.1880.997
7d0.6530.5021.8771.3131.0004.6833.8561.5910.987
" + }, + { + "category_id": 5, + "poly": [ + 308, + 894, + 1393, + 894, + 1393, + 1043, + 308, + 1043 + ], + "score": 0.975, + "html": "
NEXT-KSNEXT-GPGPPN-KSGPPN-GPRRT*BIT*BFSCVAEReject
2d0.1720.1931.0491.3331.0001.1402.8111.6491.050
3d0.1160.3150.6120.8751.0000.9551.7201.7340.984
5d0.2150.4260.6730.8901.0000.9231.7800.9611.020
7d0.1080.1470.5730.9871.0000.2911.1141.1390.986
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1067, + 1403, + 1067, + 1403, + 1194, + 299, + 1194 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 293, + 228, + 1403, + 228, + 1403, + 294, + 293, + 294 + ], + "score": 0.949 + }, + { + "category_id": 6, + "poly": [ + 397, + 320, + 1300, + 320, + 1300, + 355, + 397, + 355 + ], + "score": 0.921 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 301, + 1230, + 716, + 1230, + 716, + 1263, + 301, + 1263 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2112, + 834, + 2112 + ], + "score": 0.819 + }, + { + "category_id": 6, + "poly": [ + 298, + 553, + 1401, + 553, + 1401, + 616, + 298, + 616 + ], + "score": 0.599 + }, + { + "category_id": 6, + "poly": [ + 297, + 816, + 1399, + 816, + 1399, + 878, + 297, + 878 + ], + "score": 0.546 + }, + { + "category_id": 1, + "poly": [ + 297, + 816, + 1399, + 816, + 1399, + 878, + 297, + 878 + ], + "score": 0.412 + }, + { + "category_id": 1, + "poly": [ + 298, + 553, + 1401, + 553, + 1401, + 616, + 298, + 616 + ], + "score": 0.311 + }, + { + "category_id": 13, + "poly": [ + 1208, + 1157, + 1247, + 1157, + 1247, + 1191, + 1208, + 1191 + ], + "score": 0.88, + "latex": "{ \\tilde { V } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 465, + 1532, + 503, + 1532, + 503, + 1565, + 465, + 1565 + ], + "score": 0.88, + "latex": "{ \\tilde { V } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 553, + 1536, + 585, + 1536, + 585, + 1565, + 553, + 1565 + ], + "score": 0.85, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1297, + 1162, + 1329, + 1162, + 1329, + 1191, + 1297, + 1191 + ], + "score": 0.85, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 297, + 847, + 364, + 847, + 364, + 876, + 297, + 876 + ], + "score": 0.47, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 1096, + 1411, + 1165, + 1411, + 1165, + 1440, + 1096, + 1440 + ], + "score": 0.47, + "latex": "\\mathbf { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 796, + 1289, + 868, + 1289, + 868, + 1318, + 796, + 1318 + ], + "score": 0.41, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 686, + 1472, + 754, + 1472, + 754, + 1501, + 686, + 1501 + ], + "score": 0.37, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 569, + 585, + 635, + 585, + 635, + 613, + 569, + 613 + ], + "score": 0.31, + "latex": "\\mathbf { R R T ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 316.0, + 1302.0, + 316.0, + 1302.0, + 359.0, + 394.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1229.0, + 721.0, + 1229.0, + 721.0, + 1265.0, + 294.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 553.0, + 1402.0, + 553.0, + 1402.0, + 584.0, + 296.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 583.0, + 568.0, + 583.0, + 568.0, + 619.0, + 294.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 583.0, + 1167.0, + 583.0, + 1167.0, + 619.0, + 636.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 817.0, + 1404.0, + 817.0, + 1404.0, + 848.0, + 296.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 844.0, + 857.0, + 844.0, + 857.0, + 879.0, + 365.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1287.0, + 795.0, + 1287.0, + 795.0, + 1324.0, + 295.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1287.0, + 1406.0, + 1287.0, + 1406.0, + 1324.0, + 869.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1318.0, + 1408.0, + 1318.0, + 1408.0, + 1355.0, + 293.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1350.0, + 1406.0, + 1350.0, + 1406.0, + 1385.0, + 292.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1379.0, + 1406.0, + 1379.0, + 1406.0, + 1415.0, + 293.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1409.0, + 1095.0, + 1409.0, + 1095.0, + 1446.0, + 293.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 1409.0, + 1406.0, + 1409.0, + 1406.0, + 1446.0, + 1166.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1439.0, + 1406.0, + 1439.0, + 1406.0, + 1476.0, + 293.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1470.0, + 685.0, + 1470.0, + 685.0, + 1507.0, + 292.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1470.0, + 1406.0, + 1470.0, + 1406.0, + 1507.0, + 755.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1499.0, + 1406.0, + 1499.0, + 1406.0, + 1537.0, + 292.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1535.0, + 464.0, + 1535.0, + 464.0, + 1572.0, + 293.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 1535.0, + 552.0, + 1535.0, + 552.0, + 1572.0, + 504.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1535.0, + 1210.0, + 1535.0, + 1210.0, + 1572.0, + 586.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1067.0, + 1404.0, + 1067.0, + 1404.0, + 1103.0, + 294.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1099.0, + 1405.0, + 1099.0, + 1405.0, + 1132.0, + 295.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1127.0, + 1405.0, + 1127.0, + 1405.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1161.0, + 1207.0, + 1161.0, + 1207.0, + 1198.0, + 293.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 1161.0, + 1296.0, + 1161.0, + 1296.0, + 1198.0, + 1248.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 1161.0, + 1340.0, + 1161.0, + 1340.0, + 1198.0, + 1330.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 906.0, + 260.0, + 906.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 817.0, + 1404.0, + 817.0, + 1404.0, + 848.0, + 296.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 844.0, + 857.0, + 844.0, + 857.0, + 879.0, + 365.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 553.0, + 1402.0, + 553.0, + 1402.0, + 584.0, + 296.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 583.0, + 568.0, + 583.0, + 568.0, + 619.0, + 294.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 583.0, + 1167.0, + 583.0, + 1167.0, + 619.0, + 636.0, + 619.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 304, + 298, + 1363, + 298, + 1363, + 831, + 304, + 831 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 326, + 1140, + 1416, + 1140, + 1416, + 1666, + 326, + 1666 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 296, + 860, + 1409, + 860, + 1409, + 985, + 296, + 985 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2113, + 835, + 2113 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 296, + 1738, + 1408, + 1738, + 1408, + 1964, + 296, + 1964 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 310, + 1682, + 1420, + 1682, + 1420, + 1718, + 310, + 1718 + ], + "score": 0.495 + }, + { + "category_id": 4, + "poly": [ + 310, + 1682, + 1420, + 1682, + 1420, + 1718, + 310, + 1718 + ], + "score": 0.464 + }, + { + "category_id": 4, + "poly": [ + 296, + 1738, + 1408, + 1738, + 1408, + 1964, + 296, + 1964 + ], + "score": 0.28 + }, + { + "category_id": 13, + "poly": [ + 575, + 1930, + 703, + 1930, + 703, + 1964, + 575, + 1964 + ], + "score": 0.93, + "latex": "\\tilde { \\pi } ^ { * } \\bigl ( s ^ { \\prime } | s , U \\bigr )" + }, + { + "category_id": 13, + "poly": [ + 1302, + 1679, + 1340, + 1679, + 1340, + 1712, + 1302, + 1712 + ], + "score": 0.88, + "latex": "{ \\tilde { V } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 932, + 1770, + 971, + 1770, + 971, + 1803, + 932, + 1803 + ], + "score": 0.88, + "latex": "\\tilde { V } ^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 657, + 1895, + 694, + 1895, + 694, + 1927, + 657, + 1927 + ], + "score": 0.87, + "latex": "\\tilde { V } ^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 1023, + 1775, + 1055, + 1775, + 1055, + 1803, + 1023, + 1803 + ], + "score": 0.86, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1390, + 1684, + 1423, + 1684, + 1423, + 1712, + 1390, + 1712 + ], + "score": 0.82, + "latex": "\\tilde { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1209, + 1906, + 1225, + 1906, + 1225, + 1927, + 1209, + 1927 + ], + "score": 0.69, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1027, + 1740, + 1096, + 1740, + 1096, + 1769, + 1027, + 1769 + ], + "score": 0.58, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 351, + 1684, + 422, + 1684, + 422, + 1715, + 351, + 1715 + ], + "score": 0.39, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 479, + 923, + 550, + 923, + 550, + 952, + 479, + 952 + ], + "score": 0.28, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 304.0, + 384.0, + 304.0, + 384.0, + 320.0, + 330.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 304.0, + 471.0, + 304.0, + 471.0, + 320.0, + 398.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 307.0, + 503.0, + 307.0, + 503.0, + 317.0, + 488.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 302.0, + 556.0, + 302.0, + 556.0, + 322.0, + 505.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 307.0, + 585.0, + 307.0, + 585.0, + 317.0, + 570.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 302.0, + 642.0, + 302.0, + 642.0, + 322.0, + 588.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 304.0, + 743.0, + 304.0, + 743.0, + 320.0, + 690.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 302.0, + 831.0, + 302.0, + 831.0, + 322.0, + 757.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 302.0, + 916.0, + 302.0, + 916.0, + 322.0, + 844.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 302.0, + 1000.0, + 302.0, + 1000.0, + 322.0, + 948.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 304.0, + 1102.0, + 304.0, + 1102.0, + 321.0, + 1046.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 304.0, + 1188.0, + 304.0, + 1188.0, + 320.0, + 1115.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 302.0, + 1274.0, + 302.0, + 1274.0, + 322.0, + 1202.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 306.0, + 1306.0, + 306.0, + 1306.0, + 318.0, + 1286.0, + 318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 302.0, + 1360.0, + 302.0, + 1360.0, + 322.0, + 1307.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 322.0, + 447.0, + 322.0, + 447.0, + 342.0, + 317.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 326.0, + 489.0, + 326.0, + 489.0, + 337.0, + 479.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 321.0, + 617.0, + 321.0, + 617.0, + 346.0, + 499.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 327.0, + 725.0, + 327.0, + 725.0, + 338.0, + 713.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 328.0, + 766.0, + 328.0, + 766.0, + 337.0, + 754.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 330.0, + 805.0, + 330.0, + 805.0, + 336.0, + 798.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 327.0, + 848.0, + 327.0, + 848.0, + 338.0, + 836.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 327.0, + 931.0, + 327.0, + 931.0, + 338.0, + 919.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 327.0, + 971.0, + 327.0, + 971.0, + 339.0, + 960.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 326.0, + 1070.0, + 326.0, + 1070.0, + 341.0, + 1050.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 327.0, + 1082.0, + 327.0, + 1082.0, + 338.0, + 1071.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 327.0, + 1123.0, + 327.0, + 1123.0, + 337.0, + 1115.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 327.0, + 1331.0, + 327.0, + 1331.0, + 338.0, + 1320.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 335.0, + 529.0, + 335.0, + 529.0, + 344.0, + 520.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 358.0, + 365.0, + 358.0, + 365.0, + 477.0, + 292.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 353.0, + 707.0, + 353.0, + 707.0, + 487.0, + 677.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 361.0, + 1073.0, + 361.0, + 1073.0, + 482.0, + 1025.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 426.0, + 349.0, + 426.0, + 349.0, + 447.0, + 316.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 435.0, + 706.0, + 435.0, + 706.0, + 450.0, + 691.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 463.0, + 352.0, + 463.0, + 352.0, + 480.0, + 317.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 473.0, + 1063.0, + 473.0, + 1063.0, + 485.0, + 1052.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 479.0, + 702.0, + 479.0, + 702.0, + 489.0, + 693.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 498.0, + 351.0, + 498.0, + 351.0, + 519.0, + 314.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 514.0, + 366.0, + 514.0, + 366.0, + 543.0, + 352.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 520.0, + 408.0, + 520.0, + 408.0, + 544.0, + 395.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 522.0, + 448.0, + 522.0, + 448.0, + 543.0, + 435.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 525.0, + 489.0, + 525.0, + 489.0, + 545.0, + 476.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 522.0, + 530.0, + 522.0, + 530.0, + 543.0, + 518.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 522.0, + 572.0, + 522.0, + 572.0, + 545.0, + 559.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 526.0, + 614.0, + 526.0, + 614.0, + 545.0, + 600.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 516.0, + 725.0, + 516.0, + 725.0, + 545.0, + 690.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 531.0, + 763.0, + 531.0, + 763.0, + 542.0, + 754.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 522.0, + 807.0, + 522.0, + 807.0, + 545.0, + 794.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 522.0, + 848.0, + 522.0, + 848.0, + 545.0, + 835.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 525.0, + 890.0, + 525.0, + 890.0, + 545.0, + 877.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 521.0, + 931.0, + 521.0, + 931.0, + 545.0, + 918.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 529.0, + 972.0, + 529.0, + 972.0, + 545.0, + 959.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 509.0, + 1069.0, + 509.0, + 1069.0, + 524.0, + 1050.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 530.0, + 1082.0, + 530.0, + 1082.0, + 542.0, + 1073.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 530.0, + 1164.0, + 530.0, + 1164.0, + 542.0, + 1155.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 522.0, + 1208.0, + 522.0, + 1208.0, + 543.0, + 1194.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 521.0, + 1249.0, + 521.0, + 1249.0, + 545.0, + 1236.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 521.0, + 1290.0, + 521.0, + 1290.0, + 546.0, + 1277.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 522.0, + 1332.0, + 522.0, + 1332.0, + 545.0, + 1319.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 537.0, + 522.0, + 537.0, + 522.0, + 560.0, + 443.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 537.0, + 882.0, + 537.0, + 882.0, + 560.0, + 803.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 537.0, + 1241.0, + 537.0, + 1241.0, + 560.0, + 1161.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 576.0, + 387.0, + 576.0, + 387.0, + 595.0, + 325.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 576.0, + 473.0, + 576.0, + 473.0, + 595.0, + 393.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 576.0, + 556.0, + 576.0, + 556.0, + 595.0, + 506.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 581.0, + 585.0, + 581.0, + 585.0, + 589.0, + 575.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 576.0, + 642.0, + 576.0, + 642.0, + 595.0, + 588.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 574.0, + 744.0, + 574.0, + 744.0, + 595.0, + 684.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 576.0, + 831.0, + 576.0, + 831.0, + 595.0, + 753.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 576.0, + 916.0, + 576.0, + 916.0, + 595.0, + 846.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 576.0, + 1000.0, + 576.0, + 1000.0, + 595.0, + 948.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 574.0, + 1103.0, + 574.0, + 1103.0, + 595.0, + 1045.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 576.0, + 1190.0, + 576.0, + 1190.0, + 595.0, + 1112.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 581.0, + 1222.0, + 581.0, + 1222.0, + 590.0, + 1208.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 576.0, + 1274.0, + 576.0, + 1274.0, + 595.0, + 1223.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 581.0, + 1304.0, + 581.0, + 1304.0, + 590.0, + 1290.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 576.0, + 1360.0, + 576.0, + 1360.0, + 595.0, + 1307.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 604.0, + 365.0, + 604.0, + 365.0, + 614.0, + 354.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 604.0, + 492.0, + 604.0, + 492.0, + 614.0, + 479.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 599.0, + 722.0, + 599.0, + 722.0, + 618.0, + 689.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 598.0, + 974.0, + 598.0, + 974.0, + 614.0, + 960.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 593.0, + 1068.0, + 593.0, + 1068.0, + 608.0, + 1050.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 599.0, + 1331.0, + 599.0, + 1331.0, + 611.0, + 1320.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 615.0, + 352.0, + 615.0, + 352.0, + 634.0, + 322.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 615.0, + 487.0, + 615.0, + 487.0, + 624.0, + 476.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 619.0, + 1069.0, + 619.0, + 1069.0, + 634.0, + 1050.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 635.0, + 1077.0, + 635.0, + 1077.0, + 757.0, + 1028.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 636.0, + 698.0, + 636.0, + 698.0, + 761.0, + 675.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 642.0, + 706.0, + 642.0, + 706.0, + 657.0, + 690.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 657.0, + 348.0, + 657.0, + 348.0, + 734.0, + 310.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 672.0, + 1063.0, + 672.0, + 1063.0, + 691.0, + 1050.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 684.0, + 706.0, + 684.0, + 706.0, + 700.0, + 691.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 697.0, + 348.0, + 697.0, + 348.0, + 715.0, + 322.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 729.0, + 706.0, + 729.0, + 706.0, + 744.0, + 691.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 728.0, + 1064.0, + 728.0, + 1064.0, + 743.0, + 1050.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 738.0, + 352.0, + 738.0, + 352.0, + 756.0, + 322.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 751.0, + 1166.0, + 751.0, + 1166.0, + 760.0, + 1155.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 755.0, + 1068.0, + 755.0, + 1068.0, + 770.0, + 1050.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 778.0, + 351.0, + 778.0, + 351.0, + 796.0, + 322.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 775.0, + 703.0, + 775.0, + 703.0, + 785.0, + 693.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 782.0, + 1067.0, + 782.0, + 1067.0, + 797.0, + 1050.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 802.0, + 366.0, + 802.0, + 366.0, + 817.0, + 352.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 802.0, + 407.0, + 802.0, + 407.0, + 818.0, + 394.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 803.0, + 447.0, + 803.0, + 447.0, + 815.0, + 436.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 804.0, + 488.0, + 804.0, + 488.0, + 815.0, + 479.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 798.0, + 530.0, + 798.0, + 530.0, + 817.0, + 517.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 802.0, + 573.0, + 802.0, + 573.0, + 818.0, + 559.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 802.0, + 614.0, + 802.0, + 614.0, + 818.0, + 600.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 802.0, + 725.0, + 802.0, + 725.0, + 818.0, + 712.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 804.0, + 765.0, + 804.0, + 765.0, + 815.0, + 754.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 801.0, + 808.0, + 801.0, + 808.0, + 818.0, + 794.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 802.0, + 849.0, + 802.0, + 849.0, + 817.0, + 835.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 804.0, + 889.0, + 804.0, + 889.0, + 815.0, + 878.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 796.0, + 931.0, + 796.0, + 931.0, + 818.0, + 918.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 804.0, + 971.0, + 804.0, + 971.0, + 817.0, + 961.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 804.0, + 1082.0, + 804.0, + 1082.0, + 817.0, + 1073.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 798.0, + 1124.0, + 798.0, + 1124.0, + 818.0, + 1112.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 804.0, + 1164.0, + 804.0, + 1164.0, + 817.0, + 1155.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 798.0, + 1208.0, + 798.0, + 1208.0, + 818.0, + 1194.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 804.0, + 1248.0, + 804.0, + 1248.0, + 815.0, + 1237.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 797.0, + 1290.0, + 797.0, + 1290.0, + 818.0, + 1277.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 804.0, + 1331.0, + 804.0, + 1331.0, + 817.0, + 1320.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 813.0, + 522.0, + 813.0, + 522.0, + 833.0, + 445.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 813.0, + 881.0, + 813.0, + 881.0, + 833.0, + 803.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 812.0, + 1241.0, + 812.0, + 1241.0, + 834.0, + 1160.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1201.0, + 441.0, + 1201.0, + 441.0, + 1223.0, + 404.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1189.0, + 1028.0, + 1189.0, + 1028.0, + 1228.0, + 964.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1223.0, + 433.0, + 1223.0, + 433.0, + 1235.0, + 401.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1243.0, + 546.0, + 1243.0, + 546.0, + 1381.0, + 359.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1249.0, + 1409.0, + 1249.0, + 1409.0, + 1260.0, + 1397.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 1266.0, + 546.0, + 1266.0, + 546.0, + 1278.0, + 536.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1262.0, + 1408.0, + 1262.0, + 1408.0, + 1275.0, + 1397.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1275.0, + 1405.0, + 1275.0, + 1405.0, + 1281.0, + 1399.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1298.0, + 447.0, + 1298.0, + 447.0, + 1308.0, + 438.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1294.0, + 1222.0, + 1294.0, + 1222.0, + 1309.0, + 1202.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1313.0, + 442.0, + 1313.0, + 442.0, + 1322.0, + 433.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1308.0, + 1227.0, + 1308.0, + 1227.0, + 1325.0, + 1199.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1313.0, + 1249.0, + 1313.0, + 1249.0, + 1322.0, + 1238.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1320.0, + 551.0, + 1320.0, + 551.0, + 1332.0, + 541.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1371.0, + 1332.0, + 1371.0, + 1332.0, + 1387.0, + 1226.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1404.0, + 1161.0, + 1404.0, + 1161.0, + 1692.0, + 864.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1424.0, + 1327.0, + 1424.0, + 1327.0, + 1440.0, + 1288.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 1425.0, + 1266.0, + 1425.0, + 1266.0, + 1438.0, + 1244.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 1460.0, + 546.0, + 1460.0, + 546.0, + 1531.0, + 434.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1448.0, + 824.0, + 1448.0, + 824.0, + 1556.0, + 642.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1459.0, + 1410.0, + 1459.0, + 1410.0, + 1487.0, + 1394.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1493.0, + 437.0, + 1493.0, + 437.0, + 1502.0, + 427.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1484.0, + 1410.0, + 1484.0, + 1410.0, + 1511.0, + 1393.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1499.0, + 565.0, + 1499.0, + 565.0, + 1661.0, + 346.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1500.0, + 852.0, + 1500.0, + 852.0, + 1665.0, + 609.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1506.0, + 1406.0, + 1506.0, + 1406.0, + 1518.0, + 1397.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1518.0, + 1410.0, + 1518.0, + 1410.0, + 1534.0, + 1394.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 1537.0, + 1408.0, + 1537.0, + 1408.0, + 1548.0, + 1397.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1586.0, + 1406.0, + 1586.0, + 1406.0, + 1596.0, + 1395.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1645.0, + 1397.0, + 1645.0, + 1397.0, + 1659.0, + 1240.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 860.0, + 1406.0, + 860.0, + 1406.0, + 897.0, + 294.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 891.0, + 1407.0, + 891.0, + 1407.0, + 927.0, + 293.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 922.0, + 478.0, + 922.0, + 478.0, + 958.0, + 295.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 922.0, + 1407.0, + 922.0, + 1407.0, + 958.0, + 551.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 954.0, + 978.0, + 954.0, + 978.0, + 987.0, + 295.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 1680.0, + 350.0, + 1680.0, + 350.0, + 1718.0, + 310.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1680.0, + 1301.0, + 1680.0, + 1301.0, + 1718.0, + 423.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 1680.0, + 1389.0, + 1680.0, + 1389.0, + 1718.0, + 1341.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1739.0, + 1026.0, + 1739.0, + 1026.0, + 1775.0, + 295.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 1739.0, + 1406.0, + 1739.0, + 1406.0, + 1775.0, + 1097.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 931.0, + 1773.0, + 931.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1773.0, + 1022.0, + 1773.0, + 1022.0, + 1809.0, + 972.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1809.0, + 1056.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1406.0, + 1803.0, + 1406.0, + 1841.0, + 294.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1836.0, + 1405.0, + 1836.0, + 1405.0, + 1871.0, + 295.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1865.0, + 1406.0, + 1865.0, + 1406.0, + 1901.0, + 294.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1899.0, + 656.0, + 1899.0, + 656.0, + 1935.0, + 294.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1899.0, + 1208.0, + 1899.0, + 1208.0, + 1935.0, + 695.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1899.0, + 1406.0, + 1899.0, + 1406.0, + 1935.0, + 1226.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1930.0, + 574.0, + 1930.0, + 574.0, + 1966.0, + 295.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1930.0, + 1290.0, + 1930.0, + 1290.0, + 1966.0, + 704.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1739.0, + 1026.0, + 1739.0, + 1026.0, + 1775.0, + 295.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 1739.0, + 1406.0, + 1739.0, + 1406.0, + 1775.0, + 1097.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 931.0, + 1773.0, + 931.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1773.0, + 1022.0, + 1773.0, + 1022.0, + 1809.0, + 972.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1809.0, + 1056.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1406.0, + 1803.0, + 1406.0, + 1841.0, + 294.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1836.0, + 1405.0, + 1836.0, + 1405.0, + 1871.0, + 295.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1865.0, + 1406.0, + 1865.0, + 1406.0, + 1901.0, + 294.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1899.0, + 656.0, + 1899.0, + 656.0, + 1935.0, + 294.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1899.0, + 1208.0, + 1899.0, + 1208.0, + 1935.0, + 695.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1899.0, + 1406.0, + 1899.0, + 1406.0, + 1935.0, + 1226.0, + 1935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1930.0, + 574.0, + 1930.0, + 574.0, + 1966.0, + 295.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1930.0, + 1290.0, + 1930.0, + 1290.0, + 1966.0, + 704.0, + 1966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 1680.0, + 350.0, + 1680.0, + 350.0, + 1718.0, + 310.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1680.0, + 1301.0, + 1680.0, + 1301.0, + 1718.0, + 423.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 1680.0, + 1389.0, + 1680.0, + 1389.0, + 1718.0, + 1341.0, + 1718.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 322, + 1195, + 1386, + 1195, + 1386, + 1795, + 322, + 1795 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 324, + 285, + 1387, + 285, + 1387, + 885, + 324, + 885 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 295, + 909, + 1410, + 909, + 1410, + 1065, + 295, + 1065 + ], + "score": 0.918 + }, + { + "category_id": 4, + "poly": [ + 294, + 1820, + 1409, + 1820, + 1409, + 1976, + 294, + 1976 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 105, + 299, + 105 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2113, + 835, + 2113 + ], + "score": 0.855 + }, + { + "category_id": 15, + "poly": [ + 347.0, + 1216.0, + 548.0, + 1216.0, + 548.0, + 1358.0, + 347.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1231.0, + 1015.0, + 1231.0, + 1015.0, + 1319.0, + 904.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1282.0, + 561.0, + 1282.0, + 561.0, + 1440.0, + 346.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1308.0, + 974.0, + 1308.0, + 974.0, + 1320.0, + 947.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1400.0, + 1081.0, + 1400.0, + 1081.0, + 1409.0, + 1059.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 1462.0, + 1036.0, + 1462.0, + 1036.0, + 1496.0, + 671.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1486.0, + 598.0, + 1486.0, + 598.0, + 1771.0, + 317.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1514.0, + 1086.0, + 1514.0, + 1086.0, + 1636.0, + 883.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1558.0, + 757.0, + 1558.0, + 757.0, + 1572.0, + 739.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1564.0, + 1327.0, + 1564.0, + 1327.0, + 1584.0, + 1277.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1575.0, + 1332.0, + 1575.0, + 1332.0, + 1600.0, + 1274.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1596.0, + 1324.0, + 1596.0, + 1324.0, + 1612.0, + 1281.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1620.0, + 808.0, + 1620.0, + 808.0, + 1727.0, + 615.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1608.0, + 1059.0, + 1608.0, + 1059.0, + 1733.0, + 884.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1627.0, + 1340.0, + 1627.0, + 1340.0, + 1704.0, + 1234.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1762.0, + 986.0, + 1762.0, + 986.0, + 1800.0, + 723.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 1236.5, + 1062.0, + 1236.5, + 1062.0, + 1314.5, + 982.0, + 1314.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 303.0, + 498.0, + 303.0, + 498.0, + 392.0, + 461.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 303.0, + 1369.0, + 303.0, + 1369.0, + 425.0, + 1143.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 380.0, + 487.0, + 380.0, + 487.0, + 400.0, + 468.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 385.0, + 1377.0, + 385.0, + 1377.0, + 528.0, + 1147.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 443.0, + 788.0, + 443.0, + 788.0, + 465.0, + 745.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 474.0, + 717.0, + 474.0, + 717.0, + 501.0, + 677.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 550.0, + 1036.0, + 550.0, + 1036.0, + 588.0, + 670.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 618.0, + 802.0, + 618.0, + 802.0, + 714.0, + 624.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 581.0, + 1392.0, + 581.0, + 1392.0, + 760.0, + 1126.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 662.0, + 830.0, + 662.0, + 830.0, + 844.0, + 600.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 668.0, + 1395.0, + 668.0, + 1395.0, + 845.0, + 1132.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 708.0, + 1014.0, + 708.0, + 1014.0, + 820.0, + 889.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 853.0, + 986.0, + 853.0, + 986.0, + 889.0, + 724.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.25, + 457.5, + 955.25, + 457.5, + 955.25, + 493.5, + 895.25, + 493.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 642.0, + 688.0, + 642.0, + 688.0, + 657.0, + 681.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 713.5, + 1191.0, + 713.5, + 1191.0, + 726.5, + 1185.0, + 726.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 784.5, + 385.0, + 784.5, + 385.0, + 792.5, + 379.0, + 792.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 910.0, + 1407.0, + 910.0, + 1407.0, + 947.0, + 293.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 936.0, + 1408.0, + 936.0, + 1408.0, + 978.0, + 293.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 968.0, + 1409.0, + 968.0, + 1409.0, + 1007.0, + 292.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1000.0, + 1409.0, + 1000.0, + 1409.0, + 1038.0, + 294.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1032.0, + 1328.0, + 1032.0, + 1328.0, + 1067.0, + 293.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 1407.0, + 1817.0, + 1407.0, + 1861.0, + 292.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1887.0, + 293.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1408.0, + 1883.0, + 1408.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1408.0, + 1911.0, + 1408.0, + 1949.0, + 295.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1943.0, + 1329.0, + 1943.0, + 1329.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 830.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 22, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 346, + 765, + 1365, + 765, + 1365, + 1345, + 346, + 1345 + ], + "score": 0.951 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 105, + 299, + 105 + ], + "score": 0.878 + }, + { + "category_id": 4, + "poly": [ + 295, + 1372, + 1408, + 1372, + 1408, + 1497, + 295, + 1497 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2112, + 834, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 4, + "poly": [ + 720, + 1318, + 978, + 1318, + 978, + 1348, + 720, + 1348 + ], + "score": 0.394 + }, + { + "category_id": 1, + "poly": [ + 295, + 1372, + 1408, + 1372, + 1408, + 1497, + 295, + 1497 + ], + "score": 0.118 + }, + { + "category_id": 13, + "poly": [ + 1018, + 1404, + 1088, + 1404, + 1088, + 1433, + 1018, + 1433 + ], + "score": 0.33, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 792.0, + 865.0, + 792.0, + 865.0, + 817.0, + 828.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 850.0, + 1223.0, + 850.0, + 1223.0, + 866.0, + 1196.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 859.0, + 760.0, + 859.0, + 760.0, + 873.0, + 744.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1010.0, + 1034.0, + 1010.0, + 1034.0, + 1061.0, + 661.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1136.0, + 1225.0, + 1136.0, + 1225.0, + 1160.0, + 1199.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1315.0, + 980.0, + 1315.0, + 980.0, + 1351.0, + 718.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1370.0, + 1409.0, + 1370.0, + 1409.0, + 1408.0, + 293.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1400.0, + 1017.0, + 1400.0, + 1017.0, + 1439.0, + 292.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1400.0, + 1409.0, + 1400.0, + 1409.0, + 1439.0, + 1089.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1432.0, + 1406.0, + 1432.0, + 1406.0, + 1468.0, + 293.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1465.0, + 983.0, + 1465.0, + 983.0, + 1498.0, + 294.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 872.0, + 2084.0, + 872.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1314.0, + 981.0, + 1314.0, + 981.0, + 1352.0, + 719.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1370.0, + 1409.0, + 1370.0, + 1409.0, + 1408.0, + 293.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1400.0, + 1017.0, + 1400.0, + 1017.0, + 1439.0, + 292.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1400.0, + 1409.0, + 1400.0, + 1409.0, + 1439.0, + 1089.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1432.0, + 1406.0, + 1432.0, + 1406.0, + 1468.0, + 293.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1465.0, + 983.0, + 1465.0, + 983.0, + 1498.0, + 294.0, + 1498.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 23, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 528, + 1405, + 528, + 1405, + 712, + 298, + 712 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 935, + 1405, + 935, + 1405, + 1123, + 298, + 1123 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 727, + 1404, + 727, + 1404, + 851, + 299, + 851 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1209, + 1405, + 1209, + 1405, + 1360, + 299, + 1360 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 286, + 1404, + 286, + 1404, + 440, + 298, + 440 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 335, + 1373, + 1370, + 1373, + 1370, + 1749, + 335, + 1749 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 448, + 1792, + 1249, + 1792, + 1249, + 1825, + 448, + 1825 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 299, + 474, + 624, + 474, + 624, + 507, + 299, + 507 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 229, + 632, + 229, + 632, + 261, + 299, + 261 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 298, + 1154, + 498, + 1154, + 498, + 1186, + 298, + 1186 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 300, + 884, + 711, + 884, + 711, + 915, + 300, + 915 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.861 + }, + { + "category_id": 13, + "poly": [ + 695, + 940, + 749, + 940, + 749, + 967, + 695, + 967 + ], + "score": 0.78, + "latex": "\\mathrm { C } + +" + }, + { + "category_id": 13, + "poly": [ + 990, + 1271, + 1054, + 1271, + 1054, + 1300, + 990, + 1300 + ], + "score": 0.52, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 938, + 1327, + 938, + 1327, + 968, + 1266, + 968 + ], + "score": 0.38, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 554, + 969, + 625, + 969, + 625, + 998, + 554, + 998 + ], + "score": 0.36, + "latex": "\\mathrm { R R T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 815, + 1060, + 877, + 1060, + 877, + 1089, + 815, + 1089 + ], + "score": 0.36, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 338, + 1091, + 401, + 1091, + 401, + 1119, + 338, + 1119 + ], + "score": 0.32, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 425, + 1000, + 488, + 1000, + 488, + 1028, + 425, + 1028 + ], + "score": 0.28, + "latex": "\\mathbf { B I T ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 1442.0, + 454.0, + 1442.0, + 454.0, + 1456.0, + 434.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 1447.0, + 1266.0, + 1447.0, + 1266.0, + 1459.0, + 1246.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1548.0, + 459.0, + 1548.0, + 459.0, + 1560.0, + 433.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1588.0, + 1219.0, + 1588.0, + 1219.0, + 1596.0, + 1208.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1634.0, + 507.0, + 1634.0, + 507.0, + 1724.0, + 365.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 1687.0, + 491.0, + 1687.0, + 491.0, + 1699.0, + 481.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1696.0, + 487.0, + 1696.0, + 487.0, + 1707.0, + 446.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1678.5, + 407.0, + 1678.5, + 407.0, + 1687.5, + 403.0, + 1687.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1789.0, + 1251.0, + 1789.0, + 1251.0, + 1829.0, + 447.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 475.0, + 627.0, + 475.0, + 627.0, + 508.0, + 295.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 637.0, + 225.0, + 637.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1151.0, + 501.0, + 1151.0, + 501.0, + 1190.0, + 293.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 881.0, + 714.0, + 881.0, + 714.0, + 919.0, + 294.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 527.0, + 1407.0, + 527.0, + 1407.0, + 563.0, + 295.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 555.0, + 1403.0, + 555.0, + 1403.0, + 593.0, + 295.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 587.0, + 1406.0, + 587.0, + 1406.0, + 625.0, + 293.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 617.0, + 1405.0, + 617.0, + 1405.0, + 654.0, + 293.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 650.0, + 1405.0, + 650.0, + 1405.0, + 686.0, + 295.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 681.0, + 602.0, + 681.0, + 602.0, + 713.0, + 294.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 938.0, + 694.0, + 938.0, + 694.0, + 970.0, + 296.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 938.0, + 1265.0, + 938.0, + 1265.0, + 970.0, + 750.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 938.0, + 1407.0, + 938.0, + 1407.0, + 970.0, + 1328.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 967.0, + 553.0, + 967.0, + 553.0, + 1003.0, + 295.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 967.0, + 1407.0, + 967.0, + 1407.0, + 1003.0, + 626.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 996.0, + 424.0, + 996.0, + 424.0, + 1033.0, + 292.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 996.0, + 1405.0, + 996.0, + 1405.0, + 1033.0, + 489.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1029.0, + 1405.0, + 1029.0, + 1405.0, + 1065.0, + 293.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1061.0, + 814.0, + 1061.0, + 814.0, + 1093.0, + 295.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1061.0, + 1406.0, + 1061.0, + 1406.0, + 1093.0, + 878.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1089.0, + 337.0, + 1089.0, + 337.0, + 1124.0, + 295.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1089.0, + 1088.0, + 1089.0, + 1088.0, + 1124.0, + 402.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 727.0, + 1403.0, + 727.0, + 1403.0, + 760.0, + 295.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 759.0, + 1404.0, + 759.0, + 1404.0, + 791.0, + 295.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 788.0, + 1406.0, + 788.0, + 1406.0, + 824.0, + 294.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 816.0, + 887.0, + 816.0, + 887.0, + 858.0, + 293.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1210.0, + 1405.0, + 1210.0, + 1405.0, + 1243.0, + 297.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1236.0, + 1406.0, + 1236.0, + 1406.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1266.0, + 989.0, + 1266.0, + 989.0, + 1305.0, + 293.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1266.0, + 1405.0, + 1266.0, + 1405.0, + 1305.0, + 1055.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1299.0, + 1405.0, + 1299.0, + 1405.0, + 1335.0, + 294.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1327.0, + 376.0, + 1327.0, + 376.0, + 1366.0, + 293.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 288.0, + 1406.0, + 288.0, + 1406.0, + 321.0, + 297.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 351.0, + 294.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 346.0, + 1406.0, + 346.0, + 1406.0, + 384.0, + 293.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 374.0, + 1405.0, + 374.0, + 1405.0, + 419.0, + 291.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 408.0, + 416.0, + 408.0, + 416.0, + 443.0, + 292.0, + 443.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 24, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 336, + 1390, + 1364, + 1390, + 1364, + 1567, + 336, + 1567 + ], + "score": 0.982, + "html": "
5s10s15s20s25s30s35s40s45s50s
NEXT0.5790.6570.7030.7090.7450.7430.7460.7520.7720.763
CVAE0.3540.4370.4820.5090.5070.5390.5530.5510.5790.580
Reinforce0.1600.1700.2000.1500.1800.1900.1750.1800.2250.175
BIT*0.2260.2880.3200.3650.3640.4290.4250.4220.4430.475
RRT*0.1350.1480.1440.1360.1470.1590.1520.1580.1570.165
" + }, + { + "category_id": 5, + "poly": [ + 298, + 1777, + 1451, + 1777, + 1451, + 1953, + 298, + 1953 + ], + "score": 0.982, + "html": "
5s10s15s20s25s30s35s40s45s50s
NEXT25.14321.87320.05219.77318.31818.38118.27418.04717.22417.587
CVAE34.41430.81528.90927.71727.78426.40825.77625.88324.66524.659
Reinforce42.78142.47141.14243.29542.12941.65742.23542.09340.26942.282
BIT*39.54337.08635.74434.00234.00731.41431.58731.65130.84129.601
RRT*43.36842.85542.98143.35542.91942.41042.68042.44142.50442.165
" + }, + { + "category_id": 4, + "poly": [ + 296, + 1073, + 1407, + 1073, + 1407, + 1197, + 296, + 1197 + ], + "score": 0.945 + }, + { + "category_id": 6, + "poly": [ + 341, + 1335, + 1353, + 1335, + 1353, + 1370, + 341, + 1370 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 105, + 299, + 105 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 866, + 2087, + 866, + 2113, + 834, + 2113 + ], + "score": 0.86 + }, + { + "category_id": 6, + "poly": [ + 310, + 1719, + 1380, + 1719, + 1380, + 1753, + 310, + 1753 + ], + "score": 0.854 + }, + { + "category_id": 3, + "poly": [ + 326, + 287, + 1350, + 287, + 1350, + 1018, + 326, + 1018 + ], + "score": 0.675 + }, + { + "category_id": 0, + "poly": [ + 326, + 287, + 1350, + 287, + 1350, + 1018, + 326, + 1018 + ], + "score": 0.211 + }, + { + "category_id": 13, + "poly": [ + 658, + 1136, + 720, + 1136, + 720, + 1165, + 658, + 1165 + ], + "score": 0.6, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 559, + 1105, + 622, + 1105, + 622, + 1134, + 559, + 1134 + ], + "score": 0.35, + "latex": "\\mathrm { B I T ^ { * } }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1071.0, + 1405.0, + 1071.0, + 1405.0, + 1111.0, + 293.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1103.0, + 558.0, + 1103.0, + 558.0, + 1140.0, + 293.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1103.0, + 1407.0, + 1103.0, + 1407.0, + 1140.0, + 623.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1135.0, + 657.0, + 1135.0, + 657.0, + 1170.0, + 291.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1135.0, + 1405.0, + 1135.0, + 1405.0, + 1170.0, + 721.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1168.0, + 513.0, + 1168.0, + 513.0, + 1198.0, + 295.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1332.0, + 1356.0, + 1332.0, + 1356.0, + 1373.0, + 341.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 817.0, + 71.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 829.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1718.0, + 1384.0, + 1718.0, + 1384.0, + 1757.0, + 317.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 349.0, + 709.0, + 349.0, + 709.0, + 357.0, + 686.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 356.0, + 723.0, + 356.0, + 723.0, + 364.0, + 710.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 363.0, + 443.0, + 363.0, + 443.0, + 376.0, + 427.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 376.0, + 749.0, + 376.0, + 749.0, + 393.0, + 724.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 385.0, + 652.0, + 385.0, + 652.0, + 406.0, + 624.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 523.0, + 500.0, + 523.0, + 500.0, + 553.0, + 459.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 546.0, + 500.0, + 546.0, + 500.0, + 571.0, + 456.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 553.0, + 1045.0, + 553.0, + 1045.0, + 571.0, + 1016.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 736.0, + 990.0, + 736.0, + 990.0, + 745.0, + 981.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 741.0, + 795.0, + 741.0, + 795.0, + 749.0, + 783.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 752.0, + 500.0, + 752.0, + 500.0, + 765.0, + 479.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 748.0, + 1022.0, + 748.0, + 1022.0, + 768.0, + 982.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 772.0, + 514.0, + 772.0, + 514.0, + 780.0, + 500.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 779.0, + 625.0, + 779.0, + 625.0, + 792.0, + 605.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 775.0, + 1078.0, + 775.0, + 1078.0, + 788.0, + 1059.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 791.0, + 507.0, + 791.0, + 507.0, + 804.0, + 483.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 896.0, + 498.0, + 896.0, + 498.0, + 905.0, + 489.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 905.0, + 1018.0, + 905.0, + 1018.0, + 956.0, + 941.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 905.0, + 1248.0, + 905.0, + 1248.0, + 913.0, + 1229.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 909.0, + 1052.0, + 909.0, + 1052.0, + 951.0, + 992.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 371.5, + 364.0, + 371.5, + 364.0, + 388.0, + 344.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 359.5, + 1033.0, + 359.5, + 1033.0, + 375.5, + 1002.0, + 375.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 387.0, + 475.0, + 387.0, + 475.0, + 408.5, + 448.0, + 408.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 396.5, + 1022.0, + 396.5, + 1022.0, + 417.0, + 1000.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 504.0, + 1290.0, + 504.0, + 1290.0, + 556.0, + 1285.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.75, + 508.5, + 405.75, + 508.5, + 405.75, + 569.5, + 388.75, + 569.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.25, + 732.5, + 983.25, + 732.5, + 983.25, + 747.5, + 948.25, + 747.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.75, + 844.5, + 1361.75, + 844.5, + 1361.75, + 1007.5, + 1141.75, + 1007.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 860.0, + 566.0, + 860.0, + 566.0, + 990.0, + 346.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 349.0, + 709.0, + 349.0, + 709.0, + 357.0, + 686.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 356.0, + 723.0, + 356.0, + 723.0, + 364.0, + 710.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 363.0, + 443.0, + 363.0, + 443.0, + 376.0, + 427.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 376.0, + 749.0, + 376.0, + 749.0, + 393.0, + 724.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 385.0, + 652.0, + 385.0, + 652.0, + 406.0, + 624.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 523.0, + 500.0, + 523.0, + 500.0, + 553.0, + 459.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 546.0, + 500.0, + 546.0, + 500.0, + 571.0, + 456.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 553.0, + 1045.0, + 553.0, + 1045.0, + 571.0, + 1016.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 736.0, + 990.0, + 736.0, + 990.0, + 745.0, + 981.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 741.0, + 795.0, + 741.0, + 795.0, + 749.0, + 783.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 752.0, + 500.0, + 752.0, + 500.0, + 765.0, + 479.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 748.0, + 1022.0, + 748.0, + 1022.0, + 768.0, + 982.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 772.0, + 514.0, + 772.0, + 514.0, + 780.0, + 500.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 779.0, + 625.0, + 779.0, + 625.0, + 792.0, + 605.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 775.0, + 1078.0, + 775.0, + 1078.0, + 788.0, + 1059.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 791.0, + 507.0, + 791.0, + 507.0, + 804.0, + 483.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 896.0, + 498.0, + 896.0, + 498.0, + 905.0, + 489.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 905.0, + 1018.0, + 905.0, + 1018.0, + 956.0, + 941.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 905.0, + 1248.0, + 905.0, + 1248.0, + 913.0, + 1229.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 909.0, + 1052.0, + 909.0, + 1052.0, + 951.0, + 992.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 371.5, + 364.0, + 371.5, + 364.0, + 388.0, + 344.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 359.5, + 1033.0, + 359.5, + 1033.0, + 375.5, + 1002.0, + 375.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 387.0, + 475.0, + 387.0, + 475.0, + 408.5, + 448.0, + 408.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 396.5, + 1022.0, + 396.5, + 1022.0, + 417.0, + 1000.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 504.0, + 1290.0, + 504.0, + 1290.0, + 556.0, + 1285.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.75, + 508.5, + 405.75, + 508.5, + 405.75, + 569.5, + 388.75, + 569.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.25, + 732.5, + 983.25, + 732.5, + 983.25, + 747.5, + 948.25, + 747.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.75, + 844.5, + 1361.75, + 844.5, + 1361.75, + 1007.5, + 1141.75, + 1007.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 860.0, + 566.0, + 860.0, + 566.0, + 990.0, + 346.0, + 990.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 25, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/rylrdxHFDr/rylrdxHFDr_model.json b/parse/train/rylrdxHFDr/rylrdxHFDr_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d545deefc6ba4628edeefd67c8303e98f163f7cd --- /dev/null +++ b/parse/train/rylrdxHFDr/rylrdxHFDr_model.json @@ -0,0 +1,18745 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1552, + 1402, + 1552, + 1402, + 1814, + 298, + 1814 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1246, + 1403, + 1246, + 1403, + 1537, + 298, + 1537 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 398, + 588, + 1302, + 588, + 1302, + 852, + 398, + 852 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 969, + 1403, + 969, + 1403, + 1232, + 298, + 1232 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1830, + 1402, + 1830, + 1402, + 2033, + 298, + 2033 + ], + "score": 0.978 + }, + { + "category_id": 0, + "poly": [ + 297, + 220, + 1309, + 220, + 1309, + 269, + 297, + 269 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 302, + 915, + 573, + 915, + 573, + 950, + 302, + 950 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 773, + 519, + 927, + 519, + 927, + 552, + 773, + 552 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 311, + 319, + 1045, + 319, + 1045, + 353, + 311, + 353 + ], + "score": 0.782 + }, + { + "category_id": 1, + "poly": [ + 314, + 355, + 928, + 355, + 928, + 442, + 314, + 442 + ], + "score": 0.766 + }, + { + "category_id": 2, + "poly": [ + 842, + 2089, + 857, + 2089, + 857, + 2112, + 842, + 2112 + ], + "score": 0.632 + }, + { + "category_id": 13, + "poly": [ + 877, + 1950, + 896, + 1950, + 896, + 1977, + 877, + 1977 + ], + "score": 0.86, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1086, + 1173, + 1086, + 1173, + 1118, + 1133, + 1118 + ], + "score": 0.38, + "latex": "\\mathrm { N g }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 220.0, + 1314.0, + 220.0, + 1314.0, + 274.0, + 296.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 912.0, + 579.0, + 912.0, + 579.0, + 959.0, + 294.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 515.0, + 934.0, + 515.0, + 934.0, + 558.0, + 768.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 860.0, + 2088.0, + 860.0, + 2118.0, + 841.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1550.0, + 1406.0, + 1550.0, + 1406.0, + 1586.0, + 294.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 1406.0, + 1581.0, + 1406.0, + 1616.0, + 294.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1611.0, + 1407.0, + 1611.0, + 1407.0, + 1647.0, + 294.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1638.0, + 1406.0, + 1638.0, + 1406.0, + 1675.0, + 293.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1668.0, + 1404.0, + 1668.0, + 1404.0, + 1702.0, + 294.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1697.0, + 1404.0, + 1697.0, + 1404.0, + 1731.0, + 294.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1727.0, + 1404.0, + 1727.0, + 1404.0, + 1760.0, + 293.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1752.0, + 1406.0, + 1752.0, + 1406.0, + 1790.0, + 293.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1785.0, + 736.0, + 1785.0, + 736.0, + 1816.0, + 296.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1245.0, + 1404.0, + 1245.0, + 1404.0, + 1282.0, + 293.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1274.0, + 1407.0, + 1274.0, + 1407.0, + 1312.0, + 293.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1305.0, + 1405.0, + 1305.0, + 1405.0, + 1339.0, + 293.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1333.0, + 1405.0, + 1333.0, + 1405.0, + 1371.0, + 292.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1363.0, + 1405.0, + 1363.0, + 1405.0, + 1398.0, + 293.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 1407.0, + 1392.0, + 1407.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1455.0, + 293.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1451.0, + 1404.0, + 1451.0, + 1404.0, + 1482.0, + 296.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 1405.0, + 1477.0, + 1405.0, + 1514.0, + 293.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1511.0, + 371.0, + 1511.0, + 371.0, + 1539.0, + 294.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 588.0, + 1303.0, + 588.0, + 1303.0, + 625.0, + 394.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 618.0, + 1305.0, + 618.0, + 1305.0, + 653.0, + 394.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 648.0, + 1306.0, + 648.0, + 1306.0, + 683.0, + 394.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 674.0, + 1307.0, + 674.0, + 1307.0, + 713.0, + 392.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 707.0, + 1305.0, + 707.0, + 1305.0, + 739.0, + 394.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 735.0, + 1306.0, + 735.0, + 1306.0, + 770.0, + 394.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 764.0, + 1307.0, + 764.0, + 1307.0, + 800.0, + 393.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 791.0, + 1306.0, + 791.0, + 1306.0, + 830.0, + 392.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 821.0, + 1095.0, + 821.0, + 1095.0, + 853.0, + 395.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 968.0, + 1405.0, + 968.0, + 1405.0, + 1003.0, + 294.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 998.0, + 1404.0, + 998.0, + 1404.0, + 1034.0, + 293.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1027.0, + 1405.0, + 1027.0, + 1405.0, + 1063.0, + 293.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1054.0, + 1405.0, + 1054.0, + 1405.0, + 1093.0, + 293.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1085.0, + 1132.0, + 1085.0, + 1132.0, + 1121.0, + 294.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1085.0, + 1407.0, + 1085.0, + 1407.0, + 1121.0, + 1174.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1112.0, + 1407.0, + 1112.0, + 1407.0, + 1150.0, + 291.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1141.0, + 1403.0, + 1141.0, + 1403.0, + 1179.0, + 292.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1171.0, + 1406.0, + 1171.0, + 1406.0, + 1208.0, + 292.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1201.0, + 835.0, + 1201.0, + 835.0, + 1237.0, + 294.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1826.0, + 1406.0, + 1826.0, + 1406.0, + 1865.0, + 293.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1859.0, + 1406.0, + 1859.0, + 1406.0, + 1892.0, + 294.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1888.0, + 1404.0, + 1888.0, + 1404.0, + 1921.0, + 294.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1918.0, + 1403.0, + 1918.0, + 1403.0, + 1948.0, + 296.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1947.0, + 876.0, + 1947.0, + 876.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1947.0, + 1404.0, + 1947.0, + 1404.0, + 1976.0, + 897.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2006.0, + 293.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1406.0, + 2004.0, + 1406.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 317.0, + 487.0, + 317.0, + 487.0, + 361.0, + 309.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 315.0, + 676.0, + 315.0, + 676.0, + 363.0, + 534.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 317.0, + 902.0, + 317.0, + 902.0, + 361.0, + 722.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 316.0, + 1049.0, + 316.0, + 1049.0, + 358.0, + 948.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 345.0, + 709.0, + 345.0, + 709.0, + 386.0, + 311.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 379.0, + 608.0, + 379.0, + 608.0, + 410.0, + 311.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 407.0, + 928.0, + 407.0, + 928.0, + 445.0, + 312.0, + 445.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 230, + 1404, + 230, + 1404, + 608, + 298, + 608 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1531, + 1404, + 1531, + 1404, + 1880, + 298, + 1880 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 623, + 1404, + 623, + 1404, + 915, + 298, + 915 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1225, + 1404, + 1225, + 1404, + 1517, + 298, + 1517 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 370, + 968, + 1404, + 968, + 1404, + 1149, + 370, + 1149 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 299, + 1973, + 1400, + 1973, + 1400, + 2035, + 299, + 2035 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 295, + 929, + 843, + 929, + 843, + 959, + 295, + 959 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 300, + 1174, + 582, + 1174, + 582, + 1208, + 300, + 1208 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 301, + 1914, + 574, + 1914, + 574, + 1950, + 301, + 1950 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.687 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.22 + }, + { + "category_id": 13, + "poly": [ + 887, + 2004, + 975, + 2004, + 975, + 2036, + 887, + 2036 + ], + "score": 0.93, + "latex": "p _ { \\boldsymbol { \\theta } } ( \\mathbf { x } | \\mathbf { z } )" + }, + { + "category_id": 13, + "poly": [ + 655, + 1590, + 674, + 1590, + 674, + 1621, + 655, + 1621 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 978, + 1028, + 999, + 1028, + 999, + 1059, + 978, + 1059 + ], + "score": 0.83, + "latex": "\\beta" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1169.0, + 586.0, + 1169.0, + 586.0, + 1214.0, + 291.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1907.0, + 579.0, + 1907.0, + 579.0, + 1959.0, + 290.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 260.0, + 1406.0, + 260.0, + 1406.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 288.0, + 1406.0, + 288.0, + 1406.0, + 323.0, + 293.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 351.0, + 296.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 347.0, + 1405.0, + 347.0, + 1405.0, + 382.0, + 293.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 375.0, + 1404.0, + 375.0, + 1404.0, + 407.0, + 296.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 404.0, + 1405.0, + 404.0, + 1405.0, + 439.0, + 293.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 434.0, + 1405.0, + 434.0, + 1405.0, + 469.0, + 293.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 460.0, + 1404.0, + 460.0, + 1404.0, + 497.0, + 292.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 490.0, + 1406.0, + 490.0, + 1406.0, + 525.0, + 293.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 519.0, + 1405.0, + 519.0, + 1405.0, + 553.0, + 293.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 548.0, + 1405.0, + 548.0, + 1405.0, + 583.0, + 294.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 577.0, + 756.0, + 577.0, + 756.0, + 612.0, + 294.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1532.0, + 1404.0, + 1532.0, + 1404.0, + 1564.0, + 297.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1558.0, + 1406.0, + 1558.0, + 1406.0, + 1595.0, + 292.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1587.0, + 654.0, + 1587.0, + 654.0, + 1623.0, + 294.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1587.0, + 1404.0, + 1587.0, + 1404.0, + 1623.0, + 675.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1619.0, + 1404.0, + 1619.0, + 1404.0, + 1651.0, + 296.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1648.0, + 1404.0, + 1648.0, + 1404.0, + 1680.0, + 294.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1674.0, + 1404.0, + 1674.0, + 1404.0, + 1709.0, + 294.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1702.0, + 1406.0, + 1702.0, + 1406.0, + 1740.0, + 293.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1735.0, + 1404.0, + 1735.0, + 1404.0, + 1767.0, + 294.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1799.0, + 292.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1791.0, + 1403.0, + 1791.0, + 1403.0, + 1825.0, + 293.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1848.0, + 612.0, + 1848.0, + 612.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 623.0, + 1406.0, + 623.0, + 1406.0, + 658.0, + 296.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 651.0, + 1406.0, + 651.0, + 1406.0, + 688.0, + 292.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 681.0, + 1406.0, + 681.0, + 1406.0, + 715.0, + 294.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 708.0, + 1409.0, + 708.0, + 1409.0, + 746.0, + 293.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 740.0, + 1404.0, + 740.0, + 1404.0, + 774.0, + 294.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 767.0, + 1408.0, + 767.0, + 1408.0, + 805.0, + 293.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 796.0, + 1407.0, + 796.0, + 1407.0, + 833.0, + 292.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 825.0, + 1407.0, + 825.0, + 1407.0, + 860.0, + 292.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 855.0, + 1407.0, + 855.0, + 1407.0, + 888.0, + 293.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 887.0, + 691.0, + 887.0, + 691.0, + 917.0, + 296.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1222.0, + 1405.0, + 1222.0, + 1405.0, + 1263.0, + 292.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1252.0, + 1405.0, + 1252.0, + 1405.0, + 1291.0, + 292.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1279.0, + 1405.0, + 1279.0, + 1405.0, + 1321.0, + 292.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1312.0, + 1404.0, + 1312.0, + 1404.0, + 1347.0, + 294.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1337.0, + 1406.0, + 1337.0, + 1406.0, + 1377.0, + 295.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1370.0, + 1406.0, + 1370.0, + 1406.0, + 1404.0, + 294.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1399.0, + 1405.0, + 1399.0, + 1405.0, + 1433.0, + 294.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1424.0, + 1406.0, + 1424.0, + 1406.0, + 1464.0, + 292.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1457.0, + 1406.0, + 1457.0, + 1406.0, + 1491.0, + 294.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1486.0, + 715.0, + 1486.0, + 715.0, + 1520.0, + 296.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 967.0, + 1404.0, + 967.0, + 1404.0, + 1002.0, + 371.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 997.0, + 992.0, + 997.0, + 992.0, + 1028.0, + 393.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1027.0, + 977.0, + 1027.0, + 977.0, + 1062.0, + 371.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1027.0, + 1404.0, + 1027.0, + 1404.0, + 1062.0, + 1000.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1056.0, + 834.0, + 1056.0, + 834.0, + 1090.0, + 392.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1085.0, + 1403.0, + 1085.0, + 1403.0, + 1122.0, + 387.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1117.0, + 1023.0, + 1117.0, + 1023.0, + 1153.0, + 393.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2010.0, + 297.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 886.0, + 2001.0, + 886.0, + 2041.0, + 294.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2041.0, + 976.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 927.0, + 847.0, + 927.0, + 847.0, + 962.0, + 298.0, + 962.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1244, + 901, + 1244, + 901, + 1737, + 298, + 1737 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 961, + 1403, + 961, + 1403, + 1078, + 298, + 1078 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1737, + 1403, + 1737, + 1403, + 1882, + 298, + 1882 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 402, + 183, + 1295, + 183, + 1295, + 495, + 402, + 495 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 923, + 1554, + 1402, + 1554, + 1402, + 1701, + 923, + 1701 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 959, + 1254, + 1347, + 1254, + 1347, + 1525, + 959, + 1525 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 605, + 900, + 1095, + 900, + 1095, + 955, + 605, + 955 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 292, + 721, + 1401, + 721, + 1401, + 784, + 292, + 784 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 300, + 813, + 1401, + 813, + 1401, + 900, + 300, + 900 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 291, + 585, + 1401, + 585, + 1401, + 648, + 291, + 648 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 298, + 1973, + 1403, + 1973, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 477, + 655, + 1219, + 655, + 1219, + 696, + 477, + 696 + ], + "score": 0.94 + }, + { + "category_id": 4, + "poly": [ + 617, + 523, + 1085, + 523, + 1085, + 558, + 617, + 558 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 298, + 1918, + 946, + 1918, + 946, + 1950, + 298, + 1950 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 302, + 1120, + 1154, + 1120, + 1154, + 1156, + 302, + 1156 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 296, + 76, + 817, + 76, + 817, + 104, + 296, + 104 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 299, + 1188, + 497, + 1188, + 497, + 1219, + 299, + 1219 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1366, + 660, + 1399, + 660, + 1399, + 690, + 1366, + 690 + ], + "score": 0.873 + }, + { + "category_id": 9, + "poly": [ + 1366, + 902, + 1399, + 902, + 1399, + 932, + 1366, + 932 + ], + "score": 0.686 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.645 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.358 + }, + { + "category_id": 9, + "poly": [ + 1366, + 903, + 1399, + 903, + 1399, + 932, + 1366, + 932 + ], + "score": 0.273 + }, + { + "category_id": 14, + "poly": [ + 605, + 897, + 1094, + 897, + 1094, + 957, + 605, + 957 + ], + "score": 0.94, + "latex": "\\mathcal { W } ( p , q ) = \\operatorname* { s u p } _ { \\phi \\in \\mathcal { L } _ { 1 } } \\mathbb { E } _ { p ( x ) } [ \\phi ( x ) ] - \\mathbb { E } _ { q ( x ) } [ \\phi ( x ) ]" + }, + { + "category_id": 13, + "poly": [ + 698, + 844, + 775, + 844, + 775, + 877, + 698, + 877 + ], + "score": 0.92, + "latex": "( M , d )" + }, + { + "category_id": 13, + "poly": [ + 431, + 586, + 521, + 586, + 521, + 620, + 431, + 620 + ], + "score": 0.92, + "latex": "q _ { \\phi } ( { \\bf z } | { \\bf x } )" + }, + { + "category_id": 13, + "poly": [ + 1190, + 815, + 1243, + 815, + 1243, + 847, + 1190, + 847 + ], + "score": 0.92, + "latex": "p ( x )" + }, + { + "category_id": 13, + "poly": [ + 1293, + 815, + 1345, + 815, + 1345, + 847, + 1293, + 847 + ], + "score": 0.91, + "latex": "q ( x )" + }, + { + "category_id": 14, + "poly": [ + 480, + 654, + 1219, + 654, + 1219, + 697, + 480, + 697 + ], + "score": 0.9, + "latex": "\\mathcal { L } ( \\boldsymbol { \\theta } , \\boldsymbol { \\phi } ; \\mathbf { x } , \\mathbf { z } , \\boldsymbol { \\beta } ) = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) } \\left[ \\log p _ { \\theta } ( \\mathbf { x } | \\mathbf { z } ) \\right] - \\beta D _ { K L } \\left( q _ { \\phi } ( \\mathbf { z } | \\mathbf { x } ) \\| p ( \\mathbf { z } ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 367, + 962, + 402, + 962, + 402, + 992, + 367, + 992 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1311, + 588, + 1332, + 588, + 1332, + 619, + 1311, + 619 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 377, + 724, + 397, + 724, + 397, + 754, + 377, + 754 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 635, + 754, + 657, + 754, + 657, + 783, + 635, + 783 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 900, + 963, + 935, + 963, + 935, + 989, + 900, + 989 + ], + "score": 0.83, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 972, + 963, + 995, + 963, + 995, + 989, + 972, + 989 + ], + "score": 0.82, + "latex": "\\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 534, + 587, + 554, + 587, + 554, + 619, + 534, + 619 + ], + "score": 0.8, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 870, + 1389, + 892, + 1389, + 892, + 1420, + 870, + 1420 + ], + "score": 0.79, + "latex": "\\beta" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 184.0, + 620.0, + 184.0, + 620.0, + 222.0, + 443.0, + 222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 211.0, + 669.0, + 211.0, + 669.0, + 219.0, + 660.0, + 219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 203.0, + 803.0, + 203.0, + 803.0, + 227.0, + 675.0, + 227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 185.0, + 1098.0, + 185.0, + 1098.0, + 222.0, + 907.0, + 222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 237.0, + 620.0, + 237.0, + 620.0, + 271.0, + 441.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 235.0, + 670.0, + 235.0, + 670.0, + 248.0, + 656.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 227.0, + 791.0, + 227.0, + 791.0, + 255.0, + 676.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 241.0, + 1088.0, + 241.0, + 1088.0, + 268.0, + 914.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 340.0, + 1107.0, + 340.0, + 1107.0, + 352.0, + 1094.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 356.0, + 970.0, + 356.0, + 970.0, + 370.0, + 957.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 368.0, + 931.0, + 368.0, + 931.0, + 382.0, + 918.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 380.0, + 1070.0, + 380.0, + 1070.0, + 392.0, + 1055.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 409.0, + 859.0, + 409.0, + 859.0, + 414.0, + 854.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 424.0, + 796.0, + 424.0, + 796.0, + 432.0, + 785.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 429.0, + 525.0, + 429.0, + 525.0, + 446.0, + 503.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 436.0, + 1191.0, + 436.0, + 1191.0, + 450.0, + 1176.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 451.0, + 771.0, + 451.0, + 771.0, + 463.0, + 759.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 452.0, + 1148.0, + 452.0, + 1148.0, + 460.0, + 1140.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 410.0, + 452.0, + 410.0, + 452.0, + 426.0, + 435.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1554.0, + 1404.0, + 1554.0, + 1404.0, + 1588.0, + 922.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1585.0, + 1404.0, + 1585.0, + 1404.0, + 1615.0, + 923.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1613.0, + 1403.0, + 1613.0, + 1403.0, + 1645.0, + 921.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1642.0, + 1404.0, + 1642.0, + 1404.0, + 1675.0, + 921.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1668.0, + 1192.0, + 1668.0, + 1192.0, + 1704.0, + 923.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1351.0, + 1328.0, + 1351.0, + 1328.0, + 1372.0, + 1235.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1375.0, + 1334.0, + 1375.0, + 1334.0, + 1400.0, + 1236.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1487.0, + 1340.0, + 1487.0, + 1340.0, + 1508.0, + 1236.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 522.0, + 1084.0, + 522.0, + 1084.0, + 562.0, + 614.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1916.0, + 948.0, + 1916.0, + 948.0, + 1953.0, + 296.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1119.0, + 1157.0, + 1119.0, + 1157.0, + 1161.0, + 293.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1186.0, + 500.0, + 1186.0, + 500.0, + 1222.0, + 293.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1242.0, + 903.0, + 1242.0, + 903.0, + 1278.0, + 295.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1273.0, + 903.0, + 1273.0, + 903.0, + 1306.0, + 295.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1302.0, + 903.0, + 1302.0, + 903.0, + 1332.0, + 296.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1329.0, + 904.0, + 1329.0, + 904.0, + 1364.0, + 294.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1359.0, + 902.0, + 1359.0, + 902.0, + 1392.0, + 295.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1388.0, + 869.0, + 1388.0, + 869.0, + 1421.0, + 295.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1388.0, + 904.0, + 1388.0, + 904.0, + 1421.0, + 893.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1415.0, + 905.0, + 1415.0, + 905.0, + 1453.0, + 294.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1446.0, + 903.0, + 1446.0, + 903.0, + 1479.0, + 296.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1477.0, + 904.0, + 1477.0, + 904.0, + 1507.0, + 294.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1504.0, + 904.0, + 1504.0, + 904.0, + 1537.0, + 294.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1532.0, + 905.0, + 1532.0, + 905.0, + 1566.0, + 294.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1564.0, + 903.0, + 1564.0, + 903.0, + 1594.0, + 294.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1593.0, + 904.0, + 1593.0, + 904.0, + 1623.0, + 296.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1621.0, + 904.0, + 1621.0, + 904.0, + 1654.0, + 295.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1649.0, + 904.0, + 1649.0, + 904.0, + 1684.0, + 294.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1676.0, + 903.0, + 1676.0, + 903.0, + 1712.0, + 293.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1706.0, + 903.0, + 1706.0, + 903.0, + 1739.0, + 295.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 960.0, + 366.0, + 960.0, + 366.0, + 998.0, + 294.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 960.0, + 899.0, + 960.0, + 899.0, + 998.0, + 403.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 960.0, + 971.0, + 960.0, + 971.0, + 998.0, + 936.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 960.0, + 1405.0, + 960.0, + 1405.0, + 998.0, + 996.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 988.0, + 1404.0, + 988.0, + 1404.0, + 1025.0, + 294.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1018.0, + 1407.0, + 1018.0, + 1407.0, + 1052.0, + 296.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1044.0, + 521.0, + 1044.0, + 521.0, + 1082.0, + 295.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1733.0, + 1404.0, + 1733.0, + 1404.0, + 1770.0, + 293.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1766.0, + 1405.0, + 1766.0, + 1405.0, + 1798.0, + 296.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1792.0, + 1406.0, + 1792.0, + 1406.0, + 1827.0, + 292.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1858.0, + 292.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1852.0, + 574.0, + 1852.0, + 574.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 723.0, + 376.0, + 723.0, + 376.0, + 755.0, + 295.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 723.0, + 1403.0, + 723.0, + 1403.0, + 755.0, + 398.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 750.0, + 634.0, + 750.0, + 634.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 750.0, + 1297.0, + 750.0, + 1297.0, + 786.0, + 658.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 1189.0, + 812.0, + 1189.0, + 847.0, + 295.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 812.0, + 1292.0, + 812.0, + 1292.0, + 847.0, + 1244.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 812.0, + 1403.0, + 812.0, + 1403.0, + 847.0, + 1346.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 845.0, + 697.0, + 845.0, + 697.0, + 875.0, + 294.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 845.0, + 1405.0, + 845.0, + 1405.0, + 875.0, + 776.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 873.0, + 559.0, + 873.0, + 559.0, + 906.0, + 295.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 585.0, + 430.0, + 585.0, + 430.0, + 620.0, + 295.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 585.0, + 533.0, + 585.0, + 533.0, + 620.0, + 522.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 585.0, + 1310.0, + 585.0, + 1310.0, + 620.0, + 555.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 585.0, + 1403.0, + 585.0, + 1403.0, + 620.0, + 1333.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 617.0, + 506.0, + 617.0, + 506.0, + 650.0, + 295.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1771, + 1402, + 1771, + 1402, + 2033, + 298, + 2033 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1058, + 1403, + 1058, + 1403, + 1262, + 298, + 1262 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1610, + 1403, + 1610, + 1403, + 1757, + 299, + 1757 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 897, + 1402, + 897, + 1402, + 1044, + 298, + 1044 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1277, + 1403, + 1277, + 1403, + 1396, + 299, + 1396 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 301, + 1409, + 1399, + 1409, + 1399, + 1498, + 301, + 1498 + ], + "score": 0.957 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 301, + 267, + 1406, + 267, + 1406, + 827, + 301, + 827 + ], + "score": 0.876 + }, + { + "category_id": 0, + "poly": [ + 298, + 1548, + 970, + 1548, + 970, + 1580, + 298, + 1580 + ], + "score": 0.849 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.761 + }, + { + "category_id": 6, + "poly": [ + 297, + 228, + 995, + 228, + 995, + 261, + 297, + 261 + ], + "score": 0.641 + }, + { + "category_id": 0, + "poly": [ + 297, + 228, + 995, + 228, + 995, + 261, + 297, + 261 + ], + "score": 0.17 + }, + { + "category_id": 5, + "poly": [ + 301, + 267, + 1406, + 267, + 1406, + 827, + 301, + 827 + ], + "score": 0.14, + "html": "
Require: Expert trajectories Te : [S1,a1, S2,a2,.]~ Te, initial policy π, inverse dynamics model
1:g, discriminatorΦ, total episode T,memory capacity S if Imitator and Expert have the same dynamics model then
2:Pre-train g using Te and transitions collected by a random policy
3: elsePre-train g using transitions collected by a random policy
4: 5:end if
6:Pre-train VAE using Te,and obtain the policy_prior >Pre-train VAE and obtain policy prior
7:Pretrain π using policy prior as described in Sec 4.5
8:while episode ≤Tdo
9:while|τ|≤S do Tis the collected trajectories
10:Collect trajectory {(s,a,s',r,done)} using π
11:Update r using (4)
12:Add {(s,a,s',r,done)} to T
13:end while
14:Train Φ using max𝜙∈L1 Es~te[𝜙(s)]-Es~r[(s)] Calculate Wasserstein Distance
15:Update inverse dynamics model g
16:Update policy using (5)
17:end while
" + }, + { + "category_id": 13, + "poly": [ + 457, + 558, + 546, + 558, + 546, + 590, + 457, + 590 + ], + "score": 0.91, + "latex": "| \\tau | \\leq S" + }, + { + "category_id": 13, + "poly": [ + 483, + 645, + 700, + 645, + 700, + 678, + 483, + 678 + ], + "score": 0.91, + "latex": "\\{ ( s , a , s ^ { \\prime } , r , d o n e ) \\}" + }, + { + "category_id": 13, + "poly": [ + 627, + 269, + 939, + 269, + 939, + 301, + 627, + 301 + ], + "score": 0.9, + "latex": "\\tau _ { e } : [ s _ { 1 } , a _ { 1 } , s _ { 2 } , a _ { 2 } , . . . ] \\sim \\pi _ { e }" + }, + { + "category_id": 13, + "poly": [ + 629, + 586, + 846, + 586, + 846, + 619, + 629, + 619 + ], + "score": 0.88, + "latex": "\\left\\{ \\left( s , a , s ^ { \\prime } , r , d o n e \\right) \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 1245, + 1440, + 1266, + 1440, + 1266, + 1471, + 1245, + 1471 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 708, + 1410, + 728, + 1410, + 728, + 1441, + 708, + 1441 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 577, + 360, + 605, + 360, + 605, + 387, + 577, + 387 + ], + "score": 0.84, + "latex": "\\tau _ { e }" + }, + { + "category_id": 13, + "poly": [ + 543, + 702, + 957, + 702, + 957, + 736, + 543, + 736 + ], + "score": 0.84, + "latex": "\\begin{array} { r } { \\operatorname* { m a x } _ { \\phi \\in \\mathcal { L } _ { 1 } } E _ { s \\sim \\tau _ { e } } [ \\phi ( s ) ] - E _ { s \\sim \\tau } [ \\phi ( s ) ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 527, + 300, + 547, + 300, + 547, + 329, + 527, + 329 + ], + "score": 0.83, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 574, + 477, + 601, + 477, + 601, + 502, + 574, + 502 + ], + "score": 0.82, + "latex": "\\tau _ { e }" + }, + { + "category_id": 13, + "poly": [ + 451, + 704, + 472, + 704, + 472, + 734, + 451, + 734 + ], + "score": 0.81, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 705, + 301, + 728, + 301, + 728, + 325, + 705, + 325 + ], + "score": 0.78, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 489, + 361, + 508, + 361, + 508, + 388, + 489, + 388 + ], + "score": 0.78, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 489, + 419, + 508, + 419, + 508, + 445, + 489, + 445 + ], + "score": 0.77, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 438, + 505, + 460, + 505, + 460, + 527, + 438, + 527 + ], + "score": 0.72, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 514, + 622, + 530, + 622, + 530, + 642, + 514, + 642 + ], + "score": 0.72, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 937, + 299, + 960, + 299, + 960, + 325, + 937, + 325 + ], + "score": 0.7, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 747, + 738, + 766, + 738, + 766, + 764, + 747, + 764 + ], + "score": 0.68, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 344, + 304, + 362, + 304, + 362, + 329, + 344, + 329 + ], + "score": 0.68, + "latex": "g" + }, + { + "category_id": 13, + "poly": [ + 1098, + 276, + 1117, + 276, + 1117, + 296, + 1098, + 296 + ], + "score": 0.66, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 733, + 651, + 753, + 651, + 753, + 672, + 733, + 672 + ], + "score": 0.65, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 505, + 530, + 556, + 530, + 556, + 558, + 505, + 558 + ], + "score": 0.63, + "latex": "\\leq \\mathrm { T }" + }, + { + "category_id": 13, + "poly": [ + 916, + 593, + 938, + 593, + 938, + 615, + 916, + 615 + ], + "score": 0.58, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1079, + 564, + 1098, + 564, + 1098, + 585, + 1079, + 585 + ], + "score": 0.35, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 957, + 475, + 979, + 475, + 979, + 499, + 957, + 499 + ], + "score": 0.31, + "latex": "\\triangleright" + }, + { + "category_id": 13, + "poly": [ + 830, + 1336, + 938, + 1336, + 938, + 1365, + 830, + 1365 + ], + "score": 0.27, + "latex": "\\mathrm { S e c } ~ 5 . 3 . 2 " + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1548.0, + 974.0, + 1548.0, + 974.0, + 1584.0, + 294.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 221.0, + 1001.0, + 221.0, + 1001.0, + 273.0, + 293.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 221.0, + 1001.0, + 221.0, + 1001.0, + 273.0, + 293.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1773.0, + 1404.0, + 1773.0, + 1404.0, + 1804.0, + 296.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1799.0, + 1404.0, + 1799.0, + 1404.0, + 1834.0, + 294.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1830.0, + 1407.0, + 1830.0, + 1407.0, + 1863.0, + 293.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1858.0, + 1406.0, + 1858.0, + 1406.0, + 1893.0, + 294.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1888.0, + 1406.0, + 1888.0, + 1406.0, + 1923.0, + 294.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1916.0, + 1404.0, + 1916.0, + 1404.0, + 1951.0, + 294.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1945.0, + 1406.0, + 1945.0, + 1406.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1012.0, + 2001.0, + 1012.0, + 2035.0, + 293.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1057.0, + 1407.0, + 1057.0, + 1407.0, + 1092.0, + 293.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1089.0, + 1404.0, + 1089.0, + 1404.0, + 1120.0, + 293.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1117.0, + 1405.0, + 1117.0, + 1405.0, + 1149.0, + 293.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1145.0, + 1407.0, + 1145.0, + 1407.0, + 1179.0, + 293.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1175.0, + 1404.0, + 1175.0, + 1404.0, + 1205.0, + 297.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1203.0, + 1405.0, + 1203.0, + 1405.0, + 1236.0, + 294.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1234.0, + 1376.0, + 1234.0, + 1376.0, + 1263.0, + 296.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1611.0, + 1402.0, + 1611.0, + 1402.0, + 1644.0, + 297.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1640.0, + 1403.0, + 1640.0, + 1403.0, + 1672.0, + 297.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1668.0, + 1407.0, + 1668.0, + 1407.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1698.0, + 1404.0, + 1698.0, + 1404.0, + 1730.0, + 297.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1727.0, + 1057.0, + 1727.0, + 1057.0, + 1759.0, + 295.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 897.0, + 1406.0, + 897.0, + 1406.0, + 933.0, + 294.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 925.0, + 1406.0, + 925.0, + 1406.0, + 961.0, + 294.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 952.0, + 1404.0, + 952.0, + 1404.0, + 991.0, + 293.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 982.0, + 1407.0, + 982.0, + 1407.0, + 1019.0, + 292.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1013.0, + 572.0, + 1013.0, + 572.0, + 1049.0, + 294.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1278.0, + 1404.0, + 1278.0, + 1404.0, + 1312.0, + 297.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1306.0, + 1404.0, + 1306.0, + 1404.0, + 1341.0, + 295.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1335.0, + 829.0, + 1335.0, + 829.0, + 1369.0, + 295.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1335.0, + 1404.0, + 1335.0, + 1404.0, + 1369.0, + 939.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1363.0, + 879.0, + 1363.0, + 879.0, + 1399.0, + 292.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1409.0, + 707.0, + 1409.0, + 707.0, + 1442.0, + 296.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1409.0, + 1403.0, + 1409.0, + 1403.0, + 1442.0, + 729.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1439.0, + 1244.0, + 1439.0, + 1244.0, + 1472.0, + 296.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1439.0, + 1404.0, + 1439.0, + 1404.0, + 1472.0, + 1267.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1470.0, + 854.0, + 1470.0, + 854.0, + 1499.0, + 296.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 265.0, + 626.0, + 265.0, + 626.0, + 305.0, + 295.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 265.0, + 1097.0, + 265.0, + 1097.0, + 305.0, + 940.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 265.0, + 1405.0, + 265.0, + 1405.0, + 305.0, + 1118.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 297.0, + 343.0, + 297.0, + 343.0, + 332.0, + 339.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 297.0, + 526.0, + 297.0, + 526.0, + 332.0, + 363.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 297.0, + 704.0, + 297.0, + 704.0, + 332.0, + 548.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 297.0, + 936.0, + 297.0, + 936.0, + 332.0, + 729.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 297.0, + 965.0, + 297.0, + 965.0, + 332.0, + 961.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 324.0, + 1003.0, + 324.0, + 1003.0, + 360.0, + 309.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 355.0, + 339.0, + 355.0, + 339.0, + 390.0, + 309.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 351.0, + 488.0, + 351.0, + 488.0, + 392.0, + 379.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 351.0, + 576.0, + 351.0, + 576.0, + 392.0, + 509.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 351.0, + 1105.0, + 351.0, + 1105.0, + 392.0, + 606.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 380.0, + 400.0, + 380.0, + 400.0, + 419.0, + 326.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 386.0, + 341.0, + 386.0, + 341.0, + 417.0, + 309.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 413.0, + 339.0, + 413.0, + 339.0, + 446.0, + 307.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 410.0, + 488.0, + 410.0, + 488.0, + 450.0, + 380.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 410.0, + 1026.0, + 410.0, + 1026.0, + 450.0, + 509.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 444.0, + 341.0, + 444.0, + 341.0, + 474.0, + 309.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 444.0, + 421.0, + 444.0, + 421.0, + 471.0, + 338.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 468.0, + 573.0, + 468.0, + 573.0, + 506.0, + 311.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 468.0, + 915.0, + 468.0, + 915.0, + 506.0, + 602.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 472.0, + 1401.0, + 472.0, + 1401.0, + 504.0, + 980.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 498.0, + 437.0, + 498.0, + 437.0, + 534.0, + 306.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 498.0, + 926.0, + 498.0, + 926.0, + 534.0, + 461.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 526.0, + 504.0, + 526.0, + 504.0, + 563.0, + 310.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 526.0, + 595.0, + 526.0, + 595.0, + 563.0, + 557.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 557.0, + 339.0, + 557.0, + 339.0, + 590.0, + 306.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 557.0, + 456.0, + 557.0, + 456.0, + 591.0, + 384.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 557.0, + 592.0, + 557.0, + 592.0, + 591.0, + 547.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 556.0, + 1078.0, + 556.0, + 1078.0, + 592.0, + 1053.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 556.0, + 1405.0, + 556.0, + 1405.0, + 592.0, + 1099.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 586.0, + 340.0, + 586.0, + 340.0, + 619.0, + 296.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 582.0, + 628.0, + 582.0, + 628.0, + 624.0, + 422.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 582.0, + 915.0, + 582.0, + 915.0, + 624.0, + 847.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 615.0, + 340.0, + 615.0, + 340.0, + 649.0, + 295.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 615.0, + 513.0, + 615.0, + 513.0, + 649.0, + 427.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 615.0, + 641.0, + 615.0, + 641.0, + 649.0, + 531.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 644.0, + 339.0, + 644.0, + 339.0, + 678.0, + 295.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 641.0, + 482.0, + 641.0, + 482.0, + 678.0, + 422.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 641.0, + 732.0, + 641.0, + 732.0, + 678.0, + 701.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 672.0, + 340.0, + 672.0, + 340.0, + 706.0, + 295.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 672.0, + 507.0, + 672.0, + 507.0, + 706.0, + 384.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 702.0, + 340.0, + 702.0, + 340.0, + 736.0, + 295.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 699.0, + 450.0, + 699.0, + 450.0, + 739.0, + 382.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 699.0, + 542.0, + 699.0, + 542.0, + 739.0, + 473.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 699.0, + 1405.0, + 699.0, + 1405.0, + 738.0, + 1024.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 731.0, + 340.0, + 731.0, + 340.0, + 764.0, + 295.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 726.0, + 746.0, + 726.0, + 746.0, + 769.0, + 382.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 726.0, + 770.0, + 726.0, + 770.0, + 769.0, + 767.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 760.0, + 340.0, + 760.0, + 340.0, + 793.0, + 295.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 757.0, + 655.0, + 757.0, + 655.0, + 797.0, + 381.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 788.0, + 467.0, + 788.0, + 467.0, + 822.0, + 296.0, + 822.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1692, + 1404, + 1692, + 1404, + 1842, + 297, + 1842 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1944, + 1402, + 1944, + 1402, + 2035, + 299, + 2035 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 750, + 1402, + 750, + 1402, + 843, + 300, + 843 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1399, + 228, + 1399, + 320, + 298, + 320 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 299, + 554, + 1399, + 554, + 1399, + 645, + 299, + 645 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 299, + 1619, + 1403, + 1619, + 1403, + 1680, + 299, + 1680 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 297, + 1082, + 1401, + 1082, + 1401, + 1146, + 297, + 1146 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 624, + 661, + 1074, + 661, + 1074, + 725, + 624, + 725 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 409, + 1522, + 1288, + 1522, + 1288, + 1603, + 409, + 1603 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 293, + 1445, + 1402, + 1445, + 1402, + 1510, + 293, + 1510 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 296, + 332, + 1402, + 332, + 1402, + 395, + 296, + 395 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 299, + 1235, + 1407, + 1235, + 1407, + 1298, + 299, + 1298 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 298, + 1008, + 1401, + 1008, + 1401, + 1070, + 298, + 1070 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 293, + 480, + 1400, + 480, + 1400, + 542, + 293, + 542 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 855, + 923, + 855, + 923, + 888, + 298, + 888 + ], + "score": 0.932 + }, + { + "category_id": 8, + "poly": [ + 593, + 1356, + 1105, + 1356, + 1105, + 1416, + 593, + 1416 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 299, + 1309, + 907, + 1309, + 907, + 1342, + 299, + 1342 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 817, + 74, + 817, + 105, + 299, + 105 + ], + "score": 0.921 + }, + { + "category_id": 0, + "poly": [ + 297, + 1178, + 948, + 1178, + 948, + 1212, + 297, + 1212 + ], + "score": 0.917 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1371, + 1400, + 1371, + 1400, + 1400, + 1366, + 1400 + ], + "score": 0.898 + }, + { + "category_id": 8, + "poly": [ + 358, + 904, + 1344, + 904, + 1344, + 995, + 358, + 995 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1366, + 411, + 1400, + 411, + 1400, + 442, + 1366, + 442 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1877, + 1400, + 1877, + 1400, + 1908, + 1365, + 1908 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 678, + 1400, + 678, + 1400, + 708, + 1366, + 708 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1547, + 1400, + 1547, + 1400, + 1576, + 1366, + 1576 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2112, + 841, + 2112 + ], + "score": 0.77 + }, + { + "category_id": 8, + "poly": [ + 598, + 407, + 1102, + 407, + 1102, + 465, + 598, + 465 + ], + "score": 0.678 + }, + { + "category_id": 8, + "poly": [ + 581, + 1854, + 1112, + 1854, + 1112, + 1930, + 581, + 1930 + ], + "score": 0.637 + }, + { + "category_id": 8, + "poly": [ + 596, + 407, + 1102, + 407, + 1102, + 465, + 596, + 465 + ], + "score": 0.58 + }, + { + "category_id": 8, + "poly": [ + 582, + 1854, + 1112, + 1854, + 1112, + 1929, + 582, + 1929 + ], + "score": 0.567 + }, + { + "category_id": 14, + "poly": [ + 584, + 1854, + 1114, + 1854, + 1114, + 1931, + 584, + 1931 + ], + "score": 0.95, + "latex": "p _ { a } ( a _ { t } | s _ { t } ) \\propto \\mathrm { e x p } \\left( - \\bigg | \\Big | \\frac { g _ { i n v } ( s _ { t } , f ( s _ { t } ) ) - a _ { t } } { \\sigma } \\Big | \\Big | ^ { 2 } \\right)" + }, + { + "category_id": 14, + "poly": [ + 627, + 658, + 1074, + 658, + 1074, + 725, + 627, + 725 + ], + "score": 0.94, + "latex": "r ( s _ { i } , s _ { i + 1 } ) = \\frac { 1 } { T } [ \\phi ( s _ { i + 1 } ) - \\mathbb { E } _ { s \\sim \\tau _ { e } } \\phi ( s ) ]" + }, + { + "category_id": 14, + "poly": [ + 599, + 406, + 1101, + 406, + 1101, + 467, + 599, + 467 + ], + "score": 0.93, + "latex": "\\mathcal { W } ( \\tau _ { e } , \\tau ) = \\operatorname* { s u p } _ { \\phi \\in \\mathcal { L } _ { 1 } } \\mathbb { E } _ { s \\sim \\tau _ { e } } [ \\phi ( s ) ] - \\mathbb { E } _ { s \\sim \\tau } [ \\phi ( s ) ]" + }, + { + "category_id": 14, + "poly": [ + 357, + 901, + 1343, + 901, + 1343, + 998, + 357, + 998 + ], + "score": 0.93, + "latex": "J ( \\pi ) = \\sum _ { t = 1 } ^ { T } \\mathbb { E } _ { s _ { t } , s _ { t + 1 } \\sim \\pi } [ r ( s _ { t } , s _ { t + 1 } ) ] = \\sum _ { t = 1 } ^ { T } \\frac { \\mathbb { E } _ { s _ { t + 1 } } [ \\phi ( s _ { t + 1 } ) - \\mathbb { E } _ { s \\sim \\tau _ { e } } [ \\phi ( s ) ] ] } { T } = - \\mathcal { W } ( \\tau _ { e } , \\tau )" + }, + { + "category_id": 13, + "poly": [ + 644, + 362, + 698, + 362, + 698, + 396, + 644, + 396 + ], + "score": 0.93, + "latex": "\\{ \\tau _ { e } \\}" + }, + { + "category_id": 14, + "poly": [ + 593, + 1354, + 1106, + 1354, + 1106, + 1416, + 593, + 1416 + ], + "score": 0.93, + "latex": "J ( \\pi _ { \\boldsymbol { \\theta } } ) = L ^ { C L I P } ( \\boldsymbol { \\theta } ) - \\lambda D _ { K L } \\left( \\pi _ { \\boldsymbol { \\theta } } ( \\cdot \\left| \\boldsymbol { s } _ { t } ) \\right| \\bigg | p _ { a } \\right)" + }, + { + "category_id": 13, + "poly": [ + 751, + 1445, + 870, + 1445, + 870, + 1482, + 751, + 1482 + ], + "score": 0.93, + "latex": "L ^ { C L I P } ( \\theta )" + }, + { + "category_id": 14, + "poly": [ + 409, + 1523, + 1287, + 1523, + 1287, + 1601, + 409, + 1601 + ], + "score": 0.92, + "latex": "L ^ { C L I P } \\left( \\theta \\right) = \\hat { \\mathbb { E } } _ { t } \\left[ \\operatorname* { m i n } \\left( \\frac { \\pi _ { \\theta } ( a \\vert s ) } { \\pi _ { \\theta _ { o l d } } ( a \\vert s ) } \\hat { A } _ { t } , \\operatorname { c l i p } \\left( \\frac { \\pi _ { \\theta } ( a \\vert s ) } { \\pi _ { \\theta _ { o l d } } ( a \\vert s ) } , 1 - \\epsilon , 1 + \\epsilon \\right) \\hat { A } _ { t } \\right) \\right] ," + }, + { + "category_id": 13, + "poly": [ + 1023, + 1752, + 1240, + 1752, + 1240, + 1785, + 1023, + 1785 + ], + "score": 0.92, + "latex": "g _ { i n v } ( s _ { t } , \\dot { s } _ { t + 1 } ) = a" + }, + { + "category_id": 13, + "poly": [ + 457, + 1752, + 616, + 1752, + 616, + 1785, + 457, + 1785 + ], + "score": 0.92, + "latex": "f ( s _ { t } ) = s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 976, + 586, + 1082, + 586, + 1082, + 618, + 976, + 618 + ], + "score": 0.92, + "latex": "( s _ { i } , s _ { i + 1 } )" + }, + { + "category_id": 13, + "poly": [ + 957, + 363, + 1002, + 363, + 1002, + 396, + 957, + 396 + ], + "score": 0.92, + "latex": "\\{ \\tau \\}" + }, + { + "category_id": 13, + "poly": [ + 1226, + 1944, + 1393, + 1944, + 1393, + 1979, + 1226, + 1979 + ], + "score": 0.91, + "latex": "g _ { i n v } { \\left( s _ { t } , f ( s _ { t } ) \\right) }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1614, + 406, + 1614, + 406, + 1650, + 373, + 1650 + ], + "score": 0.91, + "latex": "\\hat { A } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 346, + 1695, + 408, + 1695, + 408, + 1724, + 346, + 1724 + ], + "score": 0.9, + "latex": "D _ { K L }" + }, + { + "category_id": 13, + "poly": [ + 699, + 784, + 719, + 784, + 719, + 814, + 699, + 814 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 340, + 1753, + 360, + 1753, + 360, + 1783, + 340, + 1783 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 437, + 1727, + 467, + 1727, + 467, + 1754, + 437, + 1754 + ], + "score": 0.85, + "latex": "p _ { a }" + }, + { + "category_id": 13, + "poly": [ + 1306, + 1698, + 1335, + 1698, + 1335, + 1725, + 1306, + 1725 + ], + "score": 0.85, + "latex": "p _ { a }" + }, + { + "category_id": 13, + "poly": [ + 986, + 557, + 1006, + 557, + 1006, + 587, + 986, + 587 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 373, + 482, + 393, + 482, + 393, + 512, + 373, + 512 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 377, + 513, + 397, + 513, + 397, + 542, + 377, + 542 + ], + "score": 0.82, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 916, + 586, + 940, + 586, + 940, + 612, + 916, + 612 + ], + "score": 0.82, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 297, + 1980, + 318, + 1980, + 318, + 2001, + 297, + 2001 + ], + "score": 0.71, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1004, + 1623, + 1018, + 1623, + 1018, + 1647, + 1004, + 1647 + ], + "score": 0.69, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 950.0, + 1177.0, + 950.0, + 1215.0, + 294.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1690.0, + 345.0, + 1690.0, + 345.0, + 1728.0, + 295.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1690.0, + 1305.0, + 1690.0, + 1305.0, + 1728.0, + 409.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 1690.0, + 1405.0, + 1690.0, + 1405.0, + 1728.0, + 1336.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1720.0, + 436.0, + 1720.0, + 436.0, + 1757.0, + 295.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1720.0, + 1402.0, + 1720.0, + 1402.0, + 1757.0, + 468.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1749.0, + 339.0, + 1749.0, + 339.0, + 1789.0, + 294.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1749.0, + 456.0, + 1749.0, + 456.0, + 1789.0, + 361.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1749.0, + 1022.0, + 1749.0, + 1022.0, + 1789.0, + 617.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1749.0, + 1405.0, + 1749.0, + 1405.0, + 1789.0, + 1241.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1780.0, + 1405.0, + 1780.0, + 1405.0, + 1814.0, + 293.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1807.0, + 1058.0, + 1807.0, + 1058.0, + 1845.0, + 294.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1225.0, + 1941.0, + 1225.0, + 1981.0, + 293.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1941.0, + 1403.0, + 1941.0, + 1403.0, + 1981.0, + 1394.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1970.0, + 296.0, + 1970.0, + 296.0, + 2010.0, + 290.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1970.0, + 1406.0, + 1970.0, + 1406.0, + 2010.0, + 319.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1216.0, + 2001.0, + 1216.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 748.0, + 1402.0, + 748.0, + 1402.0, + 788.0, + 295.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 780.0, + 698.0, + 780.0, + 698.0, + 816.0, + 294.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 780.0, + 1406.0, + 780.0, + 1406.0, + 816.0, + 720.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 807.0, + 487.0, + 807.0, + 487.0, + 845.0, + 295.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 1401.0, + 227.0, + 1401.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 292.0, + 293.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 286.0, + 452.0, + 286.0, + 452.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 555.0, + 985.0, + 555.0, + 985.0, + 589.0, + 296.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 555.0, + 1404.0, + 555.0, + 1404.0, + 589.0, + 1007.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 584.0, + 915.0, + 584.0, + 915.0, + 621.0, + 294.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 584.0, + 975.0, + 584.0, + 975.0, + 621.0, + 941.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 584.0, + 1405.0, + 584.0, + 1405.0, + 621.0, + 1083.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 615.0, + 749.0, + 615.0, + 749.0, + 648.0, + 295.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1616.0, + 372.0, + 1616.0, + 372.0, + 1653.0, + 295.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1616.0, + 1003.0, + 1616.0, + 1003.0, + 1653.0, + 407.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1616.0, + 1403.0, + 1616.0, + 1403.0, + 1653.0, + 1019.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1648.0, + 900.0, + 1648.0, + 900.0, + 1680.0, + 297.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1079.0, + 1406.0, + 1079.0, + 1406.0, + 1122.0, + 293.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1113.0, + 1326.0, + 1113.0, + 1326.0, + 1147.0, + 293.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1438.0, + 750.0, + 1438.0, + 750.0, + 1489.0, + 291.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1438.0, + 1408.0, + 1438.0, + 1408.0, + 1489.0, + 871.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1476.0, + 611.0, + 1476.0, + 611.0, + 1514.0, + 295.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 334.0, + 1404.0, + 334.0, + 1404.0, + 366.0, + 296.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 362.0, + 643.0, + 362.0, + 643.0, + 398.0, + 295.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 362.0, + 956.0, + 362.0, + 956.0, + 398.0, + 699.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 362.0, + 1349.0, + 362.0, + 1349.0, + 398.0, + 1003.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1235.0, + 1404.0, + 1235.0, + 1404.0, + 1269.0, + 296.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1263.0, + 1288.0, + 1263.0, + 1288.0, + 1300.0, + 293.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1006.0, + 1403.0, + 1006.0, + 1403.0, + 1042.0, + 294.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1036.0, + 541.0, + 1036.0, + 541.0, + 1072.0, + 296.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 479.0, + 372.0, + 479.0, + 372.0, + 515.0, + 296.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 479.0, + 1403.0, + 479.0, + 1403.0, + 515.0, + 394.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 507.0, + 376.0, + 507.0, + 376.0, + 544.0, + 293.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 544.0, + 398.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 851.0, + 925.0, + 851.0, + 925.0, + 895.0, + 293.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1306.0, + 909.0, + 1306.0, + 909.0, + 1347.0, + 295.0, + 1347.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 514, + 1404, + 514, + 1404, + 720, + 298, + 720 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1150, + 1405, + 1150, + 1405, + 1327, + 298, + 1327 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 831, + 1403, + 831, + 1403, + 1007, + 298, + 1007 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 226, + 1403, + 226, + 1403, + 349, + 299, + 349 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 391, + 1387, + 1277, + 1387, + 1277, + 1857, + 391, + 1857 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 293, + 362, + 1401, + 362, + 1401, + 422, + 293, + 422 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 301, + 1342, + 723, + 1342, + 723, + 1370, + 301, + 1370 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 300, + 764, + 557, + 764, + 557, + 798, + 300, + 798 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 302, + 1098, + 1123, + 1098, + 1123, + 1130, + 302, + 1130 + ], + "score": 0.874 + }, + { + "category_id": 0, + "poly": [ + 299, + 459, + 544, + 459, + 544, + 490, + 299, + 490 + ], + "score": 0.852 + }, + { + "category_id": 4, + "poly": [ + 329, + 1863, + 1356, + 1863, + 1356, + 1896, + 329, + 1896 + ], + "score": 0.821 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.775 + }, + { + "category_id": 0, + "poly": [ + 302, + 1042, + 1250, + 1042, + 1250, + 1074, + 302, + 1074 + ], + "score": 0.703 + }, + { + "category_id": 1, + "poly": [ + 302, + 1042, + 1250, + 1042, + 1250, + 1074, + 302, + 1074 + ], + "score": 0.185 + }, + { + "category_id": 0, + "poly": [ + 299, + 459, + 543, + 459, + 543, + 490, + 299, + 490 + ], + "score": 0.124 + }, + { + "category_id": 13, + "poly": [ + 498, + 602, + 665, + 602, + 665, + 635, + 498, + 635 + ], + "score": 0.92, + "latex": "\\hat { g _ { i n v } } ( s _ { t } , \\mathbf { \\hat { \\boldsymbol { f } } } ( s _ { t } ) )" + }, + { + "category_id": 13, + "poly": [ + 1049, + 261, + 1111, + 261, + 1111, + 290, + 1049, + 290 + ], + "score": 0.9, + "latex": "D _ { K L }" + }, + { + "category_id": 13, + "poly": [ + 778, + 225, + 812, + 225, + 812, + 260, + 778, + 260 + ], + "score": 0.9, + "latex": "\\hat { A } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 227, + 380, + 227, + 380, + 258, + 297, + 258 + ], + "score": 0.87, + "latex": "L ^ { C L I P }" + }, + { + "category_id": 13, + "poly": [ + 766, + 261, + 786, + 261, + 786, + 291, + 766, + 291 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 535, + 578, + 564, + 578, + 564, + 604, + 535, + 604 + ], + "score": 0.82, + "latex": "p _ { a }" + }, + { + "category_id": 13, + "poly": [ + 722, + 607, + 743, + 607, + 743, + 629, + 722, + 629 + ], + "score": 0.76, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 372, + 579, + 392, + 579, + 392, + 601, + 372, + 601 + ], + "score": 0.76, + "latex": "\\pi" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1384.0, + 1040.0, + 1384.0, + 1040.0, + 1403.0, + 1014.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1401.0, + 432.0, + 1401.0, + 432.0, + 1420.0, + 406.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1399.0, + 735.0, + 1399.0, + 735.0, + 1415.0, + 711.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1403.0, + 1040.0, + 1403.0, + 1040.0, + 1422.0, + 1014.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1429.0, + 431.0, + 1429.0, + 431.0, + 1444.0, + 408.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1429.0, + 735.0, + 1429.0, + 735.0, + 1445.0, + 711.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1421.0, + 955.0, + 1421.0, + 955.0, + 1450.0, + 837.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1422.0, + 1040.0, + 1422.0, + 1040.0, + 1441.0, + 1014.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1440.0, + 1040.0, + 1440.0, + 1040.0, + 1459.0, + 1014.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1455.0, + 409.0, + 1455.0, + 409.0, + 1490.0, + 391.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1459.0, + 429.0, + 1459.0, + 429.0, + 1469.0, + 421.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1466.0, + 645.0, + 1466.0, + 645.0, + 1484.0, + 587.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1453.0, + 736.0, + 1453.0, + 736.0, + 1492.0, + 695.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1454.0, + 1017.0, + 1454.0, + 1017.0, + 1490.0, + 998.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 1461.0, + 1039.0, + 1461.0, + 1039.0, + 1474.0, + 1026.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1483.0, + 430.0, + 1483.0, + 430.0, + 1498.0, + 406.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1477.0, + 1040.0, + 1477.0, + 1040.0, + 1495.0, + 1011.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1494.0, + 733.0, + 1494.0, + 733.0, + 1503.0, + 725.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1497.0, + 1039.0, + 1497.0, + 1039.0, + 1512.0, + 1011.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1508.0, + 432.0, + 1508.0, + 432.0, + 1526.0, + 403.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1519.0, + 736.0, + 1519.0, + 736.0, + 1537.0, + 707.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1515.0, + 1040.0, + 1515.0, + 1040.0, + 1533.0, + 1010.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1525.0, + 1256.0, + 1525.0, + 1256.0, + 1543.0, + 1231.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1534.0, + 432.0, + 1534.0, + 432.0, + 1552.0, + 403.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1533.0, + 1040.0, + 1533.0, + 1040.0, + 1551.0, + 1010.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1557.0, + 446.0, + 1557.0, + 446.0, + 1565.0, + 438.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1554.0, + 499.0, + 1554.0, + 499.0, + 1568.0, + 468.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1554.0, + 542.0, + 1554.0, + 542.0, + 1568.0, + 509.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1554.0, + 583.0, + 1554.0, + 583.0, + 1568.0, + 551.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1554.0, + 626.0, + 1554.0, + 626.0, + 1568.0, + 594.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1554.0, + 668.0, + 1554.0, + 668.0, + 1568.0, + 634.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1557.0, + 750.0, + 1557.0, + 750.0, + 1565.0, + 742.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 1554.0, + 804.0, + 1554.0, + 804.0, + 1568.0, + 772.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1554.0, + 846.0, + 1554.0, + 846.0, + 1568.0, + 813.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1554.0, + 887.0, + 1554.0, + 887.0, + 1569.0, + 855.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1554.0, + 930.0, + 1554.0, + 930.0, + 1568.0, + 898.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1554.0, + 972.0, + 1554.0, + 972.0, + 1568.0, + 937.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1556.0, + 1054.0, + 1556.0, + 1054.0, + 1565.0, + 1046.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1554.0, + 1108.0, + 1554.0, + 1108.0, + 1568.0, + 1075.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1116.0, + 1554.0, + 1150.0, + 1554.0, + 1150.0, + 1568.0, + 1116.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1554.0, + 1191.0, + 1554.0, + 1191.0, + 1568.0, + 1159.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1554.0, + 1234.0, + 1554.0, + 1234.0, + 1568.0, + 1201.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1554.0, + 1276.0, + 1554.0, + 1276.0, + 1568.0, + 1242.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1561.0, + 563.0, + 1561.0, + 563.0, + 1582.0, + 530.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1561.0, + 868.0, + 1561.0, + 868.0, + 1582.0, + 833.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1561.0, + 1172.0, + 1561.0, + 1172.0, + 1582.0, + 1137.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1582.0, + 630.0, + 1582.0, + 630.0, + 1613.0, + 456.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1581.0, + 919.0, + 1581.0, + 919.0, + 1618.0, + 776.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1581.0, + 1227.0, + 1581.0, + 1227.0, + 1619.0, + 1075.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1631.0, + 735.0, + 1631.0, + 735.0, + 1647.0, + 715.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1644.0, + 431.0, + 1644.0, + 431.0, + 1680.0, + 411.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1656.0, + 477.0, + 1656.0, + 477.0, + 1674.0, + 448.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1650.0, + 735.0, + 1650.0, + 735.0, + 1685.0, + 715.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1644.0, + 1039.0, + 1644.0, + 1039.0, + 1679.0, + 1019.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1669.0, + 1100.0, + 1669.0, + 1100.0, + 1689.0, + 1083.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1642.0, + 1249.0, + 1642.0, + 1249.0, + 1662.0, + 1169.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1683.0, + 438.0, + 1683.0, + 438.0, + 1742.0, + 395.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1682.0, + 449.0, + 1682.0, + 449.0, + 1691.0, + 442.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1687.0, + 735.0, + 1687.0, + 735.0, + 1704.0, + 715.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1681.0, + 1047.0, + 1681.0, + 1047.0, + 1741.0, + 1000.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 1695.0, + 1239.0, + 1695.0, + 1239.0, + 1704.0, + 1228.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1699.0, + 725.0, + 1699.0, + 725.0, + 1737.0, + 703.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1706.0, + 735.0, + 1706.0, + 735.0, + 1723.0, + 715.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1723.0, + 431.0, + 1723.0, + 431.0, + 1738.0, + 412.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1726.0, + 735.0, + 1726.0, + 735.0, + 1741.0, + 716.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 1742.0, + 432.0, + 1742.0, + 432.0, + 1759.0, + 415.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1746.0, + 734.0, + 1746.0, + 734.0, + 1759.0, + 720.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1737.0, + 1038.0, + 1737.0, + 1038.0, + 1751.0, + 1024.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1738.0, + 1267.0, + 1738.0, + 1267.0, + 1758.0, + 1211.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1758.0, + 1037.0, + 1758.0, + 1037.0, + 1769.0, + 1029.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1753.0, + 1242.0, + 1753.0, + 1242.0, + 1771.0, + 1212.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1763.0, + 431.0, + 1763.0, + 431.0, + 1776.0, + 419.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1767.0, + 733.0, + 1767.0, + 733.0, + 1776.0, + 725.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1763.0, + 1243.0, + 1763.0, + 1243.0, + 1785.0, + 1210.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1783.0, + 430.0, + 1783.0, + 430.0, + 1796.0, + 413.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1782.0, + 735.0, + 1782.0, + 735.0, + 1798.0, + 716.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1776.0, + 1038.0, + 1776.0, + 1038.0, + 1789.0, + 1021.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1778.0, + 1236.0, + 1778.0, + 1236.0, + 1797.0, + 1211.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1801.0, + 447.0, + 1801.0, + 447.0, + 1812.0, + 438.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1799.0, + 500.0, + 1799.0, + 500.0, + 1814.0, + 468.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1799.0, + 541.0, + 1799.0, + 541.0, + 1814.0, + 509.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1799.0, + 583.0, + 1799.0, + 583.0, + 1814.0, + 552.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1799.0, + 626.0, + 1799.0, + 626.0, + 1814.0, + 594.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1799.0, + 669.0, + 1799.0, + 669.0, + 1814.0, + 634.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1802.0, + 750.0, + 1802.0, + 750.0, + 1812.0, + 742.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 1799.0, + 803.0, + 1799.0, + 803.0, + 1814.0, + 772.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1799.0, + 845.0, + 1799.0, + 845.0, + 1813.0, + 813.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1799.0, + 886.0, + 1799.0, + 886.0, + 1813.0, + 856.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1799.0, + 928.0, + 1799.0, + 928.0, + 1814.0, + 898.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1799.0, + 970.0, + 1799.0, + 970.0, + 1813.0, + 939.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1801.0, + 1054.0, + 1801.0, + 1054.0, + 1812.0, + 1046.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 1799.0, + 1108.0, + 1799.0, + 1108.0, + 1814.0, + 1076.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1799.0, + 1149.0, + 1799.0, + 1149.0, + 1814.0, + 1117.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1799.0, + 1191.0, + 1799.0, + 1191.0, + 1814.0, + 1159.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 1799.0, + 1234.0, + 1799.0, + 1234.0, + 1814.0, + 1201.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1799.0, + 1276.0, + 1799.0, + 1276.0, + 1814.0, + 1242.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1806.0, + 563.0, + 1806.0, + 563.0, + 1827.0, + 529.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1807.0, + 867.0, + 1807.0, + 867.0, + 1825.0, + 833.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1806.0, + 1171.0, + 1806.0, + 1171.0, + 1827.0, + 1137.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1827.0, + 660.0, + 1827.0, + 660.0, + 1859.0, + 426.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1829.0, + 946.0, + 1829.0, + 946.0, + 1859.0, + 748.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1828.0, + 1254.0, + 1828.0, + 1254.0, + 1862.0, + 1049.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.75, + 1398.0, + 848.75, + 1398.0, + 848.75, + 1430.0, + 771.75, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1389.5, + 1271.0, + 1389.5, + 1271.0, + 1439.5, + 1048.0, + 1439.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1399.0, + 661.0, + 1399.0, + 661.0, + 1444.0, + 437.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1419.5, + 779.0, + 1419.5, + 779.0, + 1440.5, + 744.0, + 1440.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1482.5, + 653.0, + 1482.5, + 653.0, + 1529.5, + 442.0, + 1529.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1487.5, + 1198.0, + 1487.5, + 1198.0, + 1501.5, + 1179.0, + 1501.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.25, + 1704.5, + 644.25, + 1704.5, + 644.25, + 1715.5, + 641.25, + 1715.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1767.0, + 1077.0, + 1767.0, + 1077.0, + 1785.0, + 1061.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 759.0, + 562.0, + 759.0, + 562.0, + 804.0, + 291.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1096.0, + 1125.0, + 1096.0, + 1125.0, + 1134.0, + 295.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 457.0, + 547.0, + 457.0, + 547.0, + 493.0, + 295.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1859.0, + 1363.0, + 1859.0, + 1363.0, + 1902.0, + 332.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1040.0, + 1256.0, + 1040.0, + 1256.0, + 1079.0, + 293.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 456.0, + 546.0, + 456.0, + 546.0, + 494.0, + 294.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 513.0, + 1405.0, + 513.0, + 1405.0, + 551.0, + 293.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 544.0, + 1406.0, + 544.0, + 1406.0, + 579.0, + 292.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 573.0, + 371.0, + 573.0, + 371.0, + 610.0, + 293.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 573.0, + 534.0, + 573.0, + 534.0, + 610.0, + 393.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 573.0, + 1408.0, + 573.0, + 1408.0, + 610.0, + 565.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 601.0, + 497.0, + 601.0, + 497.0, + 636.0, + 291.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 601.0, + 721.0, + 601.0, + 721.0, + 636.0, + 666.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 601.0, + 1406.0, + 601.0, + 1406.0, + 636.0, + 744.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 630.0, + 1402.0, + 630.0, + 1402.0, + 664.0, + 294.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 659.0, + 1406.0, + 659.0, + 1406.0, + 696.0, + 293.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 686.0, + 693.0, + 686.0, + 693.0, + 725.0, + 293.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1152.0, + 1405.0, + 1152.0, + 1405.0, + 1187.0, + 296.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1180.0, + 1406.0, + 1180.0, + 1406.0, + 1213.0, + 293.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1208.0, + 1406.0, + 1208.0, + 1406.0, + 1243.0, + 291.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1236.0, + 1406.0, + 1236.0, + 1406.0, + 1273.0, + 292.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1268.0, + 1405.0, + 1268.0, + 1405.0, + 1299.0, + 296.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1294.0, + 1279.0, + 1294.0, + 1279.0, + 1330.0, + 293.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 830.0, + 1405.0, + 830.0, + 1405.0, + 865.0, + 296.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 860.0, + 1405.0, + 860.0, + 1405.0, + 894.0, + 293.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 889.0, + 1405.0, + 889.0, + 1405.0, + 924.0, + 294.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 918.0, + 1407.0, + 918.0, + 1407.0, + 952.0, + 294.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 949.0, + 1403.0, + 949.0, + 1403.0, + 979.0, + 294.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 976.0, + 792.0, + 976.0, + 792.0, + 1009.0, + 293.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 221.0, + 296.0, + 221.0, + 296.0, + 270.0, + 291.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 221.0, + 777.0, + 221.0, + 777.0, + 270.0, + 381.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 221.0, + 1409.0, + 221.0, + 1409.0, + 270.0, + 813.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 765.0, + 260.0, + 765.0, + 292.0, + 294.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 260.0, + 1048.0, + 260.0, + 1048.0, + 292.0, + 787.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 292.0, + 1112.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 320.0, + 293.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 318.0, + 1379.0, + 318.0, + 1379.0, + 350.0, + 295.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 357.0, + 1404.0, + 357.0, + 1404.0, + 400.0, + 295.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 392.0, + 1150.0, + 392.0, + 1150.0, + 423.0, + 296.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1341.0, + 727.0, + 1341.0, + 727.0, + 1373.0, + 298.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1040.0, + 1256.0, + 1040.0, + 1256.0, + 1079.0, + 293.0, + 1079.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 661, + 1404, + 661, + 1404, + 837, + 297, + 837 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 851, + 1404, + 851, + 1404, + 1027, + 297, + 1027 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 420, + 1404, + 420, + 1404, + 571, + 298, + 571 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 298, + 1073, + 1382, + 1073, + 1382, + 1378, + 298, + 1378 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1403, + 229, + 1403, + 406, + 297, + 406 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1560, + 1403, + 1560, + 1403, + 1679, + 299, + 1679 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 1457, + 1403, + 1457, + 1403, + 1546, + 298, + 1546 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 304, + 1850, + 1401, + 1850, + 1401, + 1944, + 304, + 1944 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 297, + 1691, + 1402, + 1691, + 1402, + 1753, + 297, + 1753 + ], + "score": 0.942 + }, + { + "category_id": 2, + "poly": [ + 322, + 1975, + 1400, + 1975, + 1400, + 2034, + 322, + 2034 + ], + "score": 0.939 + }, + { + "category_id": 4, + "poly": [ + 412, + 1390, + 1283, + 1390, + 1283, + 1423, + 412, + 1423 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 298, + 608, + 986, + 608, + 986, + 638, + 298, + 638 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 299, + 1794, + 1199, + 1794, + 1199, + 1826, + 299, + 1826 + ], + "score": 0.833 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.701 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.146 + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1231.0, + 582.0, + 1231.0, + 582.0, + 1310.0, + 382.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1347.0, + 521.0, + 1347.0, + 521.0, + 1380.0, + 343.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1346.0, + 794.0, + 1346.0, + 794.0, + 1379.0, + 613.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1345.0, + 1053.0, + 1345.0, + 1053.0, + 1378.0, + 893.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1346.0, + 1316.0, + 1346.0, + 1316.0, + 1379.0, + 1174.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1971.0, + 1196.0, + 1971.0, + 1196.0, + 2010.0, + 331.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2000.0, + 1400.0, + 2000.0, + 1400.0, + 2038.0, + 330.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1388.0, + 1286.0, + 1388.0, + 1286.0, + 1427.0, + 412.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 605.0, + 988.0, + 605.0, + 988.0, + 644.0, + 294.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1790.0, + 1203.0, + 1790.0, + 1203.0, + 1831.0, + 293.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 695.0, + 295.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 690.0, + 1405.0, + 690.0, + 1405.0, + 723.0, + 294.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 721.0, + 1405.0, + 721.0, + 1405.0, + 751.0, + 294.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 748.0, + 1405.0, + 748.0, + 1405.0, + 782.0, + 294.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 777.0, + 1407.0, + 777.0, + 1407.0, + 812.0, + 295.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 807.0, + 1179.0, + 807.0, + 1179.0, + 838.0, + 295.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 849.0, + 1406.0, + 849.0, + 1406.0, + 887.0, + 294.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 881.0, + 1405.0, + 881.0, + 1405.0, + 916.0, + 294.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 908.0, + 1407.0, + 908.0, + 1407.0, + 945.0, + 291.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 938.0, + 1407.0, + 938.0, + 1407.0, + 973.0, + 292.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 966.0, + 1409.0, + 966.0, + 1409.0, + 1004.0, + 292.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 994.0, + 423.0, + 994.0, + 423.0, + 1032.0, + 293.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 420.0, + 1406.0, + 420.0, + 1406.0, + 456.0, + 294.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 447.0, + 1404.0, + 447.0, + 1404.0, + 485.0, + 293.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 476.0, + 1403.0, + 476.0, + 1403.0, + 514.0, + 295.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 506.0, + 1406.0, + 506.0, + 1406.0, + 543.0, + 292.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1257.0, + 539.0, + 1257.0, + 575.0, + 294.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 259.0, + 1405.0, + 259.0, + 1405.0, + 294.0, + 295.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 288.0, + 1405.0, + 288.0, + 1405.0, + 322.0, + 293.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 1407.0, + 317.0, + 1407.0, + 352.0, + 295.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 345.0, + 1405.0, + 345.0, + 1405.0, + 381.0, + 294.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 377.0, + 864.0, + 377.0, + 864.0, + 408.0, + 296.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1561.0, + 1404.0, + 1561.0, + 1404.0, + 1595.0, + 295.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1589.0, + 1404.0, + 1589.0, + 1404.0, + 1624.0, + 295.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1615.0, + 1407.0, + 1615.0, + 1407.0, + 1654.0, + 294.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1646.0, + 1361.0, + 1646.0, + 1361.0, + 1684.0, + 294.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1454.0, + 1403.0, + 1454.0, + 1403.0, + 1491.0, + 294.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1487.0, + 1403.0, + 1487.0, + 1403.0, + 1517.0, + 296.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1517.0, + 695.0, + 1517.0, + 695.0, + 1550.0, + 296.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1847.0, + 1406.0, + 1847.0, + 1406.0, + 1887.0, + 296.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1879.0, + 1404.0, + 1879.0, + 1404.0, + 1915.0, + 296.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1948.0, + 296.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1692.0, + 1404.0, + 1692.0, + 1404.0, + 1727.0, + 296.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1719.0, + 693.0, + 1719.0, + 693.0, + 1756.0, + 292.0, + 1756.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1163, + 1403, + 1163, + 1403, + 1427, + 298, + 1427 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 339, + 629, + 1352, + 629, + 1352, + 882, + 339, + 882 + ], + "score": 0.982, + "html": "
Hopper-v2HalfCheetah-v2
#Demo105010 50
Expert3566± 1.2412294.22 ± 273.59
BC1318.76± 804.363525.87 ± 160.74971.42 ± 249.624813.20± 1949.26
GAIL3372.66 ± 130.753363.97± 262.77474.42 ± 389.30-175.83± 26.76
BC-GAIL3132.11 ± 520.653130.82 ± 554.54578.85 ± 934.341597.51 ± 1173.93
AIRL3.07 ± 0.023.31 ± 0.02-146.46± 23.57755.46± 10.92
Ourinit3412.58 ± 450.973601.16± 300.141064.44 ± 227.327102.29 ± 910.54
Our final3539.56 ±130.363614.19 ± 150.741616.34 ±180.768817.32 ± 860.55
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1619, + 1404, + 1619, + 1404, + 1795, + 298, + 1795 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1402, + 229, + 1402, + 377, + 299, + 377 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1032, + 1403, + 1032, + 1403, + 1150, + 298, + 1150 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 393, + 1403, + 393, + 1403, + 570, + 299, + 570 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1441, + 1400, + 1441, + 1400, + 1531, + 300, + 1531 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 301, + 1885, + 1400, + 1885, + 1400, + 1946, + 301, + 1946 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 325, + 1975, + 1319, + 1975, + 1319, + 2034, + 325, + 2034 + ], + "score": 0.939 + }, + { + "category_id": 6, + "poly": [ + 529, + 599, + 1170, + 599, + 1170, + 629, + 529, + 629 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 299, + 922, + 578, + 922, + 578, + 953, + 299, + 953 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 300, + 977, + 696, + 977, + 696, + 1010, + 300, + 1010 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 301, + 1565, + 730, + 1565, + 730, + 1598, + 301, + 1598 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 301, + 1832, + 1148, + 1832, + 1148, + 1863, + 301, + 1863 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.803 + }, + { + "category_id": 13, + "poly": [ + 434, + 1369, + 454, + 1369, + 454, + 1400, + 434, + 1400 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 697, + 1092, + 718, + 1092, + 718, + 1122, + 697, + 1122 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 589, + 1397, + 609, + 1397, + 609, + 1428, + 589, + 1428 + ], + "score": 0.85, + "latex": "\\bar { \\boldsymbol \\beta }" + }, + { + "category_id": 13, + "poly": [ + 1283, + 1679, + 1303, + 1679, + 1303, + 1709, + 1283, + 1709 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 298, + 1033, + 318, + 1033, + 318, + 1064, + 298, + 1064 + ], + "score": 0.81, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 642, + 1571, + 662, + 1571, + 662, + 1599, + 642, + 1599 + ], + "score": 0.8, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 608, + 980, + 629, + 980, + 629, + 1012, + 608, + 1012 + ], + "score": 0.66, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 551, + 981, + 572, + 981, + 572, + 1012, + 551, + 1012 + ], + "score": 0.57, + "latex": "\\beta" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1967.0, + 1323.0, + 1967.0, + 1323.0, + 2012.0, + 329.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 2001.0, + 987.0, + 2001.0, + 987.0, + 2036.0, + 331.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 597.0, + 1169.0, + 597.0, + 1169.0, + 632.0, + 530.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 919.0, + 583.0, + 919.0, + 583.0, + 958.0, + 294.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 974.0, + 550.0, + 974.0, + 550.0, + 1016.0, + 294.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 974.0, + 607.0, + 974.0, + 607.0, + 1016.0, + 573.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 974.0, + 701.0, + 974.0, + 701.0, + 1016.0, + 630.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1564.0, + 641.0, + 1564.0, + 641.0, + 1600.0, + 296.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1564.0, + 733.0, + 1564.0, + 733.0, + 1600.0, + 663.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1830.0, + 1152.0, + 1830.0, + 1152.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1200.0, + 294.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1190.0, + 1407.0, + 1190.0, + 1407.0, + 1231.0, + 292.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1220.0, + 1404.0, + 1220.0, + 1404.0, + 1256.0, + 294.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1251.0, + 1405.0, + 1251.0, + 1405.0, + 1286.0, + 294.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1280.0, + 1405.0, + 1280.0, + 1405.0, + 1315.0, + 293.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1309.0, + 1404.0, + 1309.0, + 1404.0, + 1343.0, + 294.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1339.0, + 1404.0, + 1339.0, + 1404.0, + 1371.0, + 296.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1368.0, + 433.0, + 1368.0, + 433.0, + 1400.0, + 296.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1368.0, + 1405.0, + 1368.0, + 1405.0, + 1400.0, + 455.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1395.0, + 588.0, + 1395.0, + 588.0, + 1430.0, + 294.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1395.0, + 1370.0, + 1395.0, + 1370.0, + 1430.0, + 610.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1620.0, + 1404.0, + 1620.0, + 1404.0, + 1654.0, + 292.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1648.0, + 1405.0, + 1648.0, + 1405.0, + 1681.0, + 292.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1675.0, + 1282.0, + 1675.0, + 1282.0, + 1714.0, + 291.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 1675.0, + 1408.0, + 1675.0, + 1408.0, + 1714.0, + 1304.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1706.0, + 1405.0, + 1706.0, + 1405.0, + 1740.0, + 294.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1734.0, + 1406.0, + 1734.0, + 1406.0, + 1771.0, + 292.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1764.0, + 699.0, + 1764.0, + 699.0, + 1798.0, + 293.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 1407.0, + 227.0, + 1407.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1403.0, + 260.0, + 1403.0, + 293.0, + 294.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 288.0, + 1406.0, + 288.0, + 1406.0, + 322.0, + 292.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 315.0, + 1407.0, + 315.0, + 1407.0, + 352.0, + 294.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 348.0, + 970.0, + 348.0, + 970.0, + 381.0, + 297.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1028.0, + 297.0, + 1028.0, + 297.0, + 1068.0, + 294.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1028.0, + 1405.0, + 1028.0, + 1405.0, + 1068.0, + 319.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1058.0, + 1404.0, + 1058.0, + 1404.0, + 1096.0, + 294.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1084.0, + 696.0, + 1084.0, + 696.0, + 1131.0, + 292.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1084.0, + 1405.0, + 1084.0, + 1405.0, + 1131.0, + 719.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1118.0, + 847.0, + 1118.0, + 847.0, + 1156.0, + 294.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 391.0, + 1405.0, + 391.0, + 1405.0, + 430.0, + 293.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 420.0, + 1407.0, + 420.0, + 1407.0, + 461.0, + 292.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 452.0, + 1405.0, + 452.0, + 1405.0, + 486.0, + 294.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 479.0, + 1405.0, + 479.0, + 1405.0, + 515.0, + 293.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 509.0, + 1408.0, + 509.0, + 1408.0, + 546.0, + 292.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 540.0, + 613.0, + 540.0, + 613.0, + 571.0, + 295.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1441.0, + 1405.0, + 1441.0, + 1405.0, + 1474.0, + 295.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1469.0, + 1403.0, + 1469.0, + 1403.0, + 1503.0, + 295.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1500.0, + 831.0, + 1500.0, + 831.0, + 1533.0, + 295.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1884.0, + 1404.0, + 1884.0, + 1404.0, + 1919.0, + 296.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1948.0, + 292.0, + 1948.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 408, + 294, + 1285, + 294, + 1285, + 537, + 408, + 537 + ], + "score": 0.982, + "html": "
βEnvironments
HalfCheetah-50HalfCheetah-20HeavyAnt-InitialHeavyAnt-Final
0.22007.861289.21258.91282.13
0.152653.041151.931149.651502.68
0.17102.291797.441219.345208.45
0.055933.282215.71987.724850.62
0.015893.171982.62740.541921.26
0.0054415.041369.57320.54399.31
None4759.691123.79359.15-62.13
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1311, + 1404, + 1311, + 1404, + 1546, + 298, + 1546 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1673, + 1403, + 1673, + 1403, + 1763, + 300, + 1763 + ], + "score": 0.967 + }, + { + "category_id": 5, + "poly": [ + 584, + 609, + 1108, + 609, + 1108, + 776, + 584, + 776 + ], + "score": 0.963, + "html": "
βEnvironments
HalfCheetah-50Hopper-50
0.1230.52±13.26203.87 ±14.39
0.011320.04 ± 15.43438.10± 20.43
0.0013306.91 ± 12.513303.72±10.46
None4813.20± 1949.263525.87 ± 6.74
" + }, + { + "category_id": 3, + "poly": [ + 301, + 819, + 1370, + 819, + 1370, + 1090, + 301, + 1090 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 298, + 1237, + 1397, + 1237, + 1397, + 1298, + 298, + 1298 + ], + "score": 0.946 + }, + { + "category_id": 4, + "poly": [ + 294, + 1132, + 1404, + 1132, + 1404, + 1195, + 294, + 1195 + ], + "score": 0.93 + }, + { + "category_id": 6, + "poly": [ + 291, + 232, + 1402, + 232, + 1402, + 292, + 291, + 292 + ], + "score": 0.925 + }, + { + "category_id": 6, + "poly": [ + 478, + 578, + 1222, + 578, + 1222, + 608, + 478, + 608 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 300, + 1599, + 543, + 1599, + 543, + 1634, + 300, + 1634 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 299, + 1874, + 1403, + 1874, + 1403, + 1935, + 299, + 1935 + ], + "score": 0.86 + }, + { + "category_id": 0, + "poly": [ + 300, + 1817, + 488, + 1817, + 488, + 1851, + 300, + 1851 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 294, + 1972, + 1400, + 1972, + 1400, + 2033, + 294, + 2033 + ], + "score": 0.841 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.777 + }, + { + "category_id": 13, + "poly": [ + 622, + 623, + 641, + 623, + 641, + 650, + 622, + 650 + ], + "score": 0.55, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 604, + 1874, + 671, + 1874, + 671, + 1906, + 604, + 1906 + ], + "score": 0.39, + "latex": "\\mathrm { ~ Y ~ N ~ g ~ }" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 833.0, + 344.0, + 833.0, + 344.0, + 846.0, + 318.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 836.0, + 714.0, + 836.0, + 714.0, + 852.0, + 687.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 830.0, + 994.0, + 830.0, + 994.0, + 890.0, + 760.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 849.0, + 1081.0, + 849.0, + 1081.0, + 863.0, + 1062.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 870.0, + 345.0, + 870.0, + 345.0, + 886.0, + 317.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 868.0, + 714.0, + 868.0, + 714.0, + 884.0, + 686.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 886.0, + 1084.0, + 886.0, + 1084.0, + 903.0, + 1061.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 901.0, + 345.0, + 901.0, + 345.0, + 943.0, + 304.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 900.0, + 714.0, + 900.0, + 714.0, + 941.0, + 673.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 894.0, + 1092.0, + 894.0, + 1092.0, + 951.0, + 1042.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 931.0, + 714.0, + 931.0, + 714.0, + 948.0, + 686.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 948.0, + 346.0, + 948.0, + 346.0, + 964.0, + 317.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 963.0, + 714.0, + 963.0, + 714.0, + 980.0, + 685.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 967.0, + 1083.0, + 967.0, + 1083.0, + 981.0, + 1062.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 973.0, + 454.0, + 973.0, + 454.0, + 992.0, + 373.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 973.0, + 822.0, + 973.0, + 822.0, + 992.0, + 747.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 990.0, + 341.0, + 990.0, + 341.0, + 996.0, + 335.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 986.0, + 448.0, + 986.0, + 448.0, + 1004.0, + 373.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 983.0, + 818.0, + 983.0, + 818.0, + 1006.0, + 744.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 999.0, + 412.0, + 999.0, + 412.0, + 1019.0, + 373.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 996.0, + 713.0, + 996.0, + 713.0, + 1009.0, + 687.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 999.0, + 779.0, + 999.0, + 779.0, + 1018.0, + 743.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 1007.0, + 1083.0, + 1007.0, + 1083.0, + 1021.0, + 1062.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1025.0, + 364.0, + 1025.0, + 364.0, + 1036.0, + 353.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1021.0, + 618.0, + 1021.0, + 618.0, + 1039.0, + 370.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1025.0, + 732.0, + 1025.0, + 732.0, + 1036.0, + 721.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1022.0, + 788.0, + 1022.0, + 788.0, + 1038.0, + 748.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1022.0, + 829.0, + 1022.0, + 829.0, + 1038.0, + 790.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 1022.0, + 872.0, + 1022.0, + 872.0, + 1038.0, + 834.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1021.0, + 999.0, + 1021.0, + 999.0, + 1039.0, + 874.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1025.0, + 1101.0, + 1025.0, + 1101.0, + 1037.0, + 1090.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1022.0, + 1163.0, + 1022.0, + 1163.0, + 1038.0, + 1127.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1022.0, + 1213.0, + 1022.0, + 1213.0, + 1038.0, + 1175.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1022.0, + 1262.0, + 1022.0, + 1262.0, + 1038.0, + 1224.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1022.0, + 1312.0, + 1022.0, + 1312.0, + 1038.0, + 1273.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 1022.0, + 1362.0, + 1022.0, + 1362.0, + 1038.0, + 1320.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1033.0, + 507.0, + 1033.0, + 507.0, + 1053.0, + 467.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 1032.0, + 874.0, + 1032.0, + 874.0, + 1053.0, + 834.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1031.0, + 1244.0, + 1031.0, + 1244.0, + 1055.0, + 1202.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1054.0, + 585.0, + 1054.0, + 585.0, + 1088.0, + 376.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1054.0, + 945.0, + 1054.0, + 945.0, + 1089.0, + 754.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1054.0, + 1291.0, + 1054.0, + 1291.0, + 1091.0, + 1147.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 879.0, + 977.0, + 879.0, + 977.0, + 972.0, + 727.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1134.0, + 1404.0, + 1134.0, + 1404.0, + 1166.0, + 297.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1161.0, + 1357.0, + 1161.0, + 1357.0, + 1198.0, + 294.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 270.0, + 293.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 700.0, + 262.0, + 700.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 572.0, + 1225.0, + 572.0, + 1225.0, + 616.0, + 474.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1593.0, + 549.0, + 1593.0, + 549.0, + 1643.0, + 291.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1817.0, + 490.0, + 1817.0, + 490.0, + 1854.0, + 296.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1312.0, + 1404.0, + 1312.0, + 1404.0, + 1345.0, + 296.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1339.0, + 1403.0, + 1339.0, + 1403.0, + 1375.0, + 293.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1367.0, + 1408.0, + 1367.0, + 1408.0, + 1409.0, + 291.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1397.0, + 1406.0, + 1397.0, + 1406.0, + 1434.0, + 294.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1429.0, + 1405.0, + 1429.0, + 1405.0, + 1461.0, + 296.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1457.0, + 1405.0, + 1457.0, + 1405.0, + 1490.0, + 294.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1486.0, + 1405.0, + 1486.0, + 1405.0, + 1519.0, + 296.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1515.0, + 1247.0, + 1515.0, + 1247.0, + 1548.0, + 294.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1670.0, + 1405.0, + 1670.0, + 1405.0, + 1709.0, + 294.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1703.0, + 1404.0, + 1703.0, + 1404.0, + 1736.0, + 296.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1732.0, + 1300.0, + 1732.0, + 1300.0, + 1765.0, + 295.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1235.0, + 1402.0, + 1235.0, + 1402.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1264.0, + 863.0, + 1264.0, + 863.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1872.0, + 603.0, + 1872.0, + 603.0, + 1910.0, + 294.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1872.0, + 1403.0, + 1872.0, + 1403.0, + 1910.0, + 672.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1902.0, + 1403.0, + 1902.0, + 1403.0, + 1937.0, + 324.0, + 1937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1967.0, + 1405.0, + 1967.0, + 1405.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2000.0, + 613.0, + 2000.0, + 613.0, + 2035.0, + 321.0, + 2035.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 297, + 1620, + 1397, + 1620, + 1397, + 1683, + 297, + 1683 + ], + "score": 0.554 + }, + { + "category_id": 1, + "poly": [ + 298, + 1700, + 1399, + 1700, + 1399, + 1764, + 298, + 1764 + ], + "score": 0.527 + }, + { + "category_id": 1, + "poly": [ + 296, + 1540, + 1401, + 1540, + 1401, + 1604, + 296, + 1604 + ], + "score": 0.527 + }, + { + "category_id": 1, + "poly": [ + 296, + 1430, + 1400, + 1430, + 1400, + 1522, + 296, + 1522 + ], + "score": 0.526 + }, + { + "category_id": 1, + "poly": [ + 295, + 1319, + 1401, + 1319, + 1401, + 1413, + 295, + 1413 + ], + "score": 0.523 + }, + { + "category_id": 1, + "poly": [ + 294, + 1100, + 1400, + 1100, + 1400, + 1192, + 294, + 1192 + ], + "score": 0.509 + }, + { + "category_id": 1, + "poly": [ + 296, + 990, + 1400, + 990, + 1400, + 1084, + 296, + 1084 + ], + "score": 0.507 + }, + { + "category_id": 1, + "poly": [ + 294, + 1209, + 1401, + 1209, + 1401, + 1301, + 294, + 1301 + ], + "score": 0.498 + }, + { + "category_id": 1, + "poly": [ + 302, + 229, + 1398, + 229, + 1398, + 320, + 302, + 320 + ], + "score": 0.498 + }, + { + "category_id": 1, + "poly": [ + 298, + 719, + 1401, + 719, + 1401, + 811, + 298, + 811 + ], + "score": 0.486 + }, + { + "category_id": 1, + "poly": [ + 292, + 448, + 1402, + 448, + 1402, + 511, + 292, + 511 + ], + "score": 0.471 + }, + { + "category_id": 1, + "poly": [ + 294, + 909, + 1399, + 909, + 1399, + 973, + 294, + 973 + ], + "score": 0.466 + }, + { + "category_id": 1, + "poly": [ + 298, + 609, + 1401, + 609, + 1401, + 701, + 298, + 701 + ], + "score": 0.462 + }, + { + "category_id": 1, + "poly": [ + 301, + 339, + 1397, + 339, + 1397, + 430, + 301, + 430 + ], + "score": 0.445 + }, + { + "category_id": 1, + "poly": [ + 293, + 529, + 1402, + 529, + 1402, + 592, + 293, + 592 + ], + "score": 0.441 + }, + { + "category_id": 1, + "poly": [ + 295, + 1782, + 1401, + 1782, + 1401, + 1874, + 295, + 1874 + ], + "score": 0.436 + }, + { + "category_id": 1, + "poly": [ + 296, + 1892, + 1398, + 1892, + 1398, + 1955, + 296, + 1955 + ], + "score": 0.403 + }, + { + "category_id": 1, + "poly": [ + 294, + 828, + 1400, + 828, + 1400, + 892, + 294, + 892 + ], + "score": 0.385 + }, + { + "category_id": 1, + "poly": [ + 296, + 1972, + 1400, + 1972, + 1400, + 2035, + 296, + 2035 + ], + "score": 0.294 + }, + { + "category_id": 13, + "poly": [ + 561, + 1652, + 580, + 1652, + 580, + 1683, + 561, + 1683 + ], + "score": 0.86, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 408, + 1703, + 432, + 1703, + 432, + 1730, + 408, + 1730 + ], + "score": 0.35, + "latex": "\\mathrm { \\bf P }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1621.0, + 1402.0, + 1621.0, + 1402.0, + 1657.0, + 294.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1650.0, + 560.0, + 1650.0, + 560.0, + 1686.0, + 323.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1650.0, + 1328.0, + 1650.0, + 1328.0, + 1686.0, + 581.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1696.0, + 407.0, + 1696.0, + 407.0, + 1740.0, + 293.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1696.0, + 1404.0, + 1696.0, + 1404.0, + 1740.0, + 433.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1731.0, + 597.0, + 1731.0, + 597.0, + 1761.0, + 322.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1540.0, + 1405.0, + 1540.0, + 1405.0, + 1576.0, + 295.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1569.0, + 935.0, + 1569.0, + 935.0, + 1605.0, + 323.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1428.0, + 1405.0, + 1428.0, + 1405.0, + 1466.0, + 292.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1459.0, + 1406.0, + 1459.0, + 1406.0, + 1494.0, + 323.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1490.0, + 946.0, + 1490.0, + 946.0, + 1520.0, + 322.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1314.0, + 1406.0, + 1314.0, + 1406.0, + 1360.0, + 293.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1350.0, + 1405.0, + 1350.0, + 1405.0, + 1384.0, + 323.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1379.0, + 1041.0, + 1379.0, + 1041.0, + 1413.0, + 324.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1102.0, + 1402.0, + 1102.0, + 1402.0, + 1135.0, + 296.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1127.0, + 1406.0, + 1127.0, + 1406.0, + 1172.0, + 319.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1160.0, + 533.0, + 1160.0, + 533.0, + 1192.0, + 324.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 987.0, + 1405.0, + 987.0, + 1405.0, + 1029.0, + 292.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1021.0, + 1404.0, + 1021.0, + 1404.0, + 1055.0, + 324.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1052.0, + 882.0, + 1052.0, + 882.0, + 1084.0, + 322.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1208.0, + 1404.0, + 1208.0, + 1404.0, + 1244.0, + 293.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1241.0, + 1406.0, + 1241.0, + 1406.0, + 1274.0, + 322.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1268.0, + 614.0, + 1268.0, + 614.0, + 1302.0, + 322.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 226.0, + 1404.0, + 226.0, + 1404.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 258.0, + 1404.0, + 258.0, + 1404.0, + 296.0, + 321.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 289.0, + 688.0, + 289.0, + 688.0, + 322.0, + 323.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 721.0, + 1403.0, + 721.0, + 1403.0, + 754.0, + 296.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 747.0, + 1406.0, + 747.0, + 1406.0, + 788.0, + 319.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 779.0, + 717.0, + 779.0, + 717.0, + 813.0, + 323.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 446.0, + 1402.0, + 446.0, + 1402.0, + 482.0, + 295.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 477.0, + 504.0, + 477.0, + 504.0, + 510.0, + 321.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 908.0, + 1404.0, + 908.0, + 1404.0, + 946.0, + 293.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 938.0, + 688.0, + 938.0, + 688.0, + 973.0, + 322.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 608.0, + 1403.0, + 608.0, + 1403.0, + 645.0, + 294.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 636.0, + 1406.0, + 636.0, + 1406.0, + 680.0, + 320.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 669.0, + 534.0, + 669.0, + 534.0, + 701.0, + 323.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 336.0, + 1403.0, + 336.0, + 1403.0, + 377.0, + 296.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 367.0, + 1403.0, + 367.0, + 1403.0, + 406.0, + 320.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 398.0, + 688.0, + 398.0, + 688.0, + 431.0, + 322.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 527.0, + 1404.0, + 527.0, + 1404.0, + 566.0, + 293.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 559.0, + 903.0, + 559.0, + 903.0, + 592.0, + 322.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1779.0, + 1405.0, + 1779.0, + 1405.0, + 1821.0, + 293.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1812.0, + 1403.0, + 1812.0, + 1403.0, + 1846.0, + 323.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1841.0, + 599.0, + 1841.0, + 599.0, + 1875.0, + 323.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1889.0, + 1404.0, + 1889.0, + 1404.0, + 1930.0, + 292.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1922.0, + 831.0, + 1922.0, + 831.0, + 1957.0, + 321.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 829.0, + 1404.0, + 829.0, + 1404.0, + 865.0, + 296.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 859.0, + 1074.0, + 859.0, + 1074.0, + 893.0, + 322.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1401.0, + 1974.0, + 1401.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 2001.0, + 1391.0, + 2001.0, + 1391.0, + 2038.0, + 320.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.828 + }, + { + "category_id": 1, + "poly": [ + 299, + 610, + 1400, + 610, + 1400, + 702, + 299, + 702 + ], + "score": 0.701 + }, + { + "category_id": 1, + "poly": [ + 296, + 1022, + 1402, + 1022, + 1402, + 1085, + 296, + 1085 + ], + "score": 0.691 + }, + { + "category_id": 1, + "poly": [ + 299, + 1103, + 1400, + 1103, + 1400, + 1167, + 299, + 1167 + ], + "score": 0.685 + }, + { + "category_id": 1, + "poly": [ + 298, + 1184, + 1402, + 1184, + 1402, + 1249, + 298, + 1249 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 296, + 1265, + 1403, + 1265, + 1403, + 1329, + 296, + 1329 + ], + "score": 0.677 + }, + { + "category_id": 1, + "poly": [ + 302, + 229, + 1398, + 229, + 1398, + 321, + 302, + 321 + ], + "score": 0.676 + }, + { + "category_id": 1, + "poly": [ + 295, + 529, + 1401, + 529, + 1401, + 593, + 295, + 593 + ], + "score": 0.667 + }, + { + "category_id": 1, + "poly": [ + 293, + 830, + 1400, + 830, + 1400, + 895, + 293, + 895 + ], + "score": 0.66 + }, + { + "category_id": 1, + "poly": [ + 300, + 720, + 1399, + 720, + 1399, + 813, + 300, + 813 + ], + "score": 0.65 + }, + { + "category_id": 1, + "poly": [ + 296, + 1728, + 1400, + 1728, + 1400, + 1793, + 296, + 1793 + ], + "score": 0.648 + }, + { + "category_id": 1, + "poly": [ + 301, + 912, + 1400, + 912, + 1400, + 1004, + 301, + 1004 + ], + "score": 0.64 + }, + { + "category_id": 1, + "poly": [ + 298, + 1347, + 1400, + 1347, + 1400, + 1439, + 298, + 1439 + ], + "score": 0.637 + }, + { + "category_id": 1, + "poly": [ + 297, + 1457, + 1399, + 1457, + 1399, + 1549, + 297, + 1549 + ], + "score": 0.616 + }, + { + "category_id": 1, + "poly": [ + 300, + 420, + 1399, + 420, + 1399, + 510, + 300, + 510 + ], + "score": 0.611 + }, + { + "category_id": 1, + "poly": [ + 298, + 1648, + 1400, + 1648, + 1400, + 1711, + 298, + 1711 + ], + "score": 0.61 + }, + { + "category_id": 1, + "poly": [ + 296, + 338, + 1399, + 338, + 1399, + 402, + 296, + 402 + ], + "score": 0.596 + }, + { + "category_id": 1, + "poly": [ + 298, + 1567, + 1400, + 1567, + 1400, + 1630, + 298, + 1630 + ], + "score": 0.594 + }, + { + "category_id": 1, + "poly": [ + 295, + 1809, + 1400, + 1809, + 1400, + 1873, + 295, + 1873 + ], + "score": 0.594 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1398, + 1972, + 1398, + 2034, + 299, + 2034 + ], + "score": 0.56 + }, + { + "category_id": 1, + "poly": [ + 298, + 1891, + 1400, + 1891, + 1400, + 1954, + 298, + 1954 + ], + "score": 0.535 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 609.0, + 1404.0, + 609.0, + 1404.0, + 648.0, + 293.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 640.0, + 1405.0, + 640.0, + 1405.0, + 678.0, + 322.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 670.0, + 547.0, + 670.0, + 547.0, + 702.0, + 320.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1022.0, + 1406.0, + 1022.0, + 1406.0, + 1057.0, + 295.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1054.0, + 908.0, + 1054.0, + 908.0, + 1085.0, + 323.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1103.0, + 1405.0, + 1103.0, + 1405.0, + 1140.0, + 295.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1134.0, + 1045.0, + 1134.0, + 1045.0, + 1168.0, + 322.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1187.0, + 1404.0, + 1187.0, + 1404.0, + 1219.0, + 296.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1215.0, + 1323.0, + 1215.0, + 1323.0, + 1250.0, + 323.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1264.0, + 1405.0, + 1264.0, + 1405.0, + 1302.0, + 295.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1296.0, + 1370.0, + 1296.0, + 1370.0, + 1332.0, + 324.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 228.0, + 1402.0, + 228.0, + 1402.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 259.0, + 1403.0, + 259.0, + 1403.0, + 293.0, + 321.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 287.0, + 1154.0, + 287.0, + 1154.0, + 323.0, + 317.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 525.0, + 1403.0, + 525.0, + 1403.0, + 567.0, + 294.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 558.0, + 589.0, + 558.0, + 589.0, + 594.0, + 322.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 833.0, + 1402.0, + 833.0, + 1402.0, + 865.0, + 296.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 862.0, + 1383.0, + 862.0, + 1383.0, + 895.0, + 325.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 719.0, + 1404.0, + 719.0, + 1404.0, + 757.0, + 294.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 751.0, + 1403.0, + 751.0, + 1403.0, + 783.0, + 324.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 781.0, + 939.0, + 781.0, + 939.0, + 814.0, + 321.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1727.0, + 1405.0, + 1727.0, + 1405.0, + 1766.0, + 292.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1758.0, + 1378.0, + 1758.0, + 1378.0, + 1794.0, + 320.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 909.0, + 1405.0, + 909.0, + 1405.0, + 950.0, + 294.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 942.0, + 1405.0, + 942.0, + 1405.0, + 976.0, + 323.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 970.0, + 1263.0, + 970.0, + 1263.0, + 1007.0, + 321.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1345.0, + 1404.0, + 1345.0, + 1404.0, + 1385.0, + 293.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1380.0, + 1402.0, + 1380.0, + 1402.0, + 1410.0, + 323.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1407.0, + 1078.0, + 1407.0, + 1078.0, + 1441.0, + 321.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1454.0, + 1404.0, + 1454.0, + 1404.0, + 1495.0, + 293.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1490.0, + 1403.0, + 1490.0, + 1403.0, + 1520.0, + 323.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1515.0, + 1080.0, + 1515.0, + 1080.0, + 1551.0, + 318.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 421.0, + 1404.0, + 421.0, + 1404.0, + 454.0, + 296.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 451.0, + 1404.0, + 451.0, + 1404.0, + 484.0, + 321.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 474.0, + 412.0, + 474.0, + 412.0, + 513.0, + 320.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1646.0, + 1406.0, + 1646.0, + 1406.0, + 1686.0, + 296.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1677.0, + 944.0, + 1677.0, + 944.0, + 1713.0, + 322.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 334.0, + 1404.0, + 334.0, + 1404.0, + 379.0, + 294.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 369.0, + 1232.0, + 369.0, + 1232.0, + 405.0, + 322.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1563.0, + 1405.0, + 1563.0, + 1405.0, + 1604.0, + 292.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1593.0, + 398.0, + 1593.0, + 398.0, + 1632.0, + 319.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1805.0, + 1404.0, + 1805.0, + 1404.0, + 1850.0, + 292.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1841.0, + 1057.0, + 1841.0, + 1057.0, + 1874.0, + 323.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1970.0, + 1402.0, + 1970.0, + 1402.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 1092.0, + 2003.0, + 1092.0, + 2035.0, + 323.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1891.0, + 1404.0, + 1891.0, + 1404.0, + 1926.0, + 294.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1919.0, + 518.0, + 1919.0, + 518.0, + 1956.0, + 324.0, + 1956.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 296, + 738, + 1401, + 738, + 1401, + 801, + 296, + 801 + ], + "score": 0.559 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1402, + 229, + 1402, + 320, + 299, + 320 + ], + "score": 0.551 + }, + { + "category_id": 1, + "poly": [ + 296, + 658, + 1402, + 658, + 1402, + 720, + 296, + 720 + ], + "score": 0.534 + }, + { + "category_id": 1, + "poly": [ + 298, + 577, + 1401, + 577, + 1401, + 641, + 298, + 641 + ], + "score": 0.492 + }, + { + "category_id": 1, + "poly": [ + 299, + 497, + 1402, + 497, + 1402, + 560, + 299, + 560 + ], + "score": 0.465 + }, + { + "category_id": 1, + "poly": [ + 297, + 338, + 1399, + 338, + 1399, + 400, + 297, + 400 + ], + "score": 0.449 + }, + { + "category_id": 1, + "poly": [ + 295, + 418, + 1402, + 418, + 1402, + 480, + 295, + 480 + ], + "score": 0.386 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 736.0, + 1405.0, + 736.0, + 1405.0, + 775.0, + 294.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 767.0, + 665.0, + 767.0, + 665.0, + 803.0, + 323.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1403.0, + 230.0, + 1403.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 258.0, + 1404.0, + 258.0, + 1404.0, + 294.0, + 320.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 288.0, + 470.0, + 288.0, + 470.0, + 322.0, + 322.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 652.0, + 1406.0, + 652.0, + 1406.0, + 698.0, + 293.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 685.0, + 399.0, + 685.0, + 399.0, + 723.0, + 318.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 573.0, + 1407.0, + 573.0, + 1407.0, + 618.0, + 293.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 608.0, + 1326.0, + 608.0, + 1326.0, + 644.0, + 322.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 494.0, + 1406.0, + 494.0, + 1406.0, + 535.0, + 293.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 522.0, + 399.0, + 522.0, + 399.0, + 563.0, + 319.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 337.0, + 1403.0, + 337.0, + 1403.0, + 373.0, + 295.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 368.0, + 787.0, + 368.0, + 787.0, + 402.0, + 322.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 415.0, + 1405.0, + 415.0, + 1405.0, + 455.0, + 293.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 448.0, + 624.0, + 448.0, + 624.0, + 481.0, + 323.0, + 481.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 388, + 1657, + 1313, + 1657, + 1313, + 1854, + 388, + 1854 + ], + "score": 0.984, + "html": "
EnvironmentState DimAction DimReward thresholdExpert Performance
Swimmer-v282360332
Hopper-v211338003566
Walker2d-v2176-4924
Ant-v2111860006157
HalfCheetah-v2176480012294
Humanoid-v23761710005187
" + }, + { + "category_id": 5, + "poly": [ + 440, + 1143, + 1258, + 1143, + 1258, + 1285, + 440, + 1285 + ], + "score": 0.979, + "html": "
DisabledAntHeavyAntLightAnt
BC1042.45± 75.13550.6± 77.624936.59± 53.42
GAIL-1033.54± 254.36-1089.34± 174.13-971.74 ± 123.14
AIRL-3252.69± 153.47-62.02± 5.33-626.44 ± 104.31
SAIL(Ours)3305.71 ± 67.215608.47 ± 57.674335.46± 82.34
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1424, + 1404, + 1424, + 1404, + 1575, + 298, + 1575 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 472, + 899, + 1229, + 899, + 1229, + 1041, + 472, + 1041 + ], + "score": 0.977, + "html": "
DisabledSwimmerLightSwimmerHeavySwimmer
BC249.09±1.53277.99± 3.41255.95± 2.5
GAIL228.46±2.02-4.11 ± 0.51254.91 ± 1.35
AIRL283.42 ± 3.6967.58 ± 25.09301.27 ± 5.21
SAIL(Ours)287.71 ± 2.31342.61 ± 6.14286.4 ± 3.2
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1886, + 1404, + 1886, + 1404, + 2035, + 298, + 2035 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 296, + 1404, + 296, + 1404, + 443, + 297, + 443 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 296, + 759, + 1401, + 759, + 1401, + 821, + 296, + 821 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 298, + 685, + 1401, + 685, + 1401, + 747, + 298, + 747 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 298, + 1356, + 1266, + 1356, + 1266, + 1393, + 298, + 1393 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 286, + 457, + 1307, + 457, + 1307, + 491, + 286, + 491 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 296, + 594, + 956, + 594, + 956, + 626, + 296, + 626 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 296, + 640, + 987, + 640, + 987, + 672, + 296, + 672 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 299, + 503, + 808, + 503, + 808, + 535, + 299, + 535 + ], + "score": 0.892 + }, + { + "category_id": 6, + "poly": [ + 636, + 1119, + 1065, + 1119, + 1065, + 1145, + 636, + 1145 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.889 + }, + { + "category_id": 6, + "poly": [ + 601, + 873, + 1099, + 873, + 1099, + 900, + 601, + 900 + ], + "score": 0.884 + }, + { + "category_id": 6, + "poly": [ + 569, + 1633, + 1126, + 1633, + 1126, + 1659, + 569, + 1659 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 298, + 549, + 1022, + 549, + 1022, + 582, + 298, + 582 + ], + "score": 0.864 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.853 + }, + { + "category_id": 0, + "poly": [ + 298, + 228, + 1034, + 228, + 1034, + 262, + 298, + 262 + ], + "score": 0.834 + }, + { + "category_id": 13, + "poly": [ + 1205, + 298, + 1299, + 298, + 1299, + 327, + 1205, + 327 + ], + "score": 0.84, + "latex": "2 4 \\times 2 4" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1355.0, + 1274.0, + 1355.0, + 1274.0, + 1397.0, + 293.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1114.0, + 1068.0, + 1114.0, + 1068.0, + 1151.0, + 631.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 871.0, + 1103.0, + 871.0, + 1103.0, + 906.0, + 596.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1630.0, + 1128.0, + 1630.0, + 1128.0, + 1665.0, + 572.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 228.0, + 1035.0, + 228.0, + 1035.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1424.0, + 1404.0, + 1424.0, + 1404.0, + 1462.0, + 296.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1452.0, + 1402.0, + 1452.0, + 1402.0, + 1489.0, + 294.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1481.0, + 1406.0, + 1481.0, + 1406.0, + 1520.0, + 291.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1513.0, + 1406.0, + 1513.0, + 1406.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1540.0, + 641.0, + 1540.0, + 641.0, + 1581.0, + 293.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1887.0, + 1405.0, + 1887.0, + 1405.0, + 1923.0, + 296.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1916.0, + 1404.0, + 1916.0, + 1404.0, + 1949.0, + 296.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1946.0, + 1405.0, + 1946.0, + 1405.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2005.0, + 369.0, + 2005.0, + 369.0, + 2041.0, + 293.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 295.0, + 1204.0, + 295.0, + 1204.0, + 331.0, + 295.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 295.0, + 1402.0, + 295.0, + 1402.0, + 331.0, + 1300.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 1406.0, + 322.0, + 1406.0, + 359.0, + 294.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 353.0, + 1406.0, + 353.0, + 1406.0, + 387.0, + 292.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 384.0, + 1405.0, + 384.0, + 1405.0, + 416.0, + 295.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 412.0, + 417.0, + 412.0, + 417.0, + 445.0, + 295.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 757.0, + 1405.0, + 757.0, + 1405.0, + 794.0, + 294.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 786.0, + 403.0, + 786.0, + 403.0, + 823.0, + 294.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 682.0, + 1405.0, + 682.0, + 1405.0, + 721.0, + 293.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 708.0, + 376.0, + 708.0, + 376.0, + 751.0, + 288.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 455.0, + 1311.0, + 455.0, + 1311.0, + 495.0, + 294.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 586.0, + 950.0, + 586.0, + 950.0, + 636.0, + 293.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 634.0, + 994.0, + 634.0, + 994.0, + 680.0, + 294.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 495.0, + 810.0, + 495.0, + 810.0, + 545.0, + 293.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 543.0, + 1025.0, + 543.0, + 1025.0, + 589.0, + 294.0, + 589.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 351, + 664, + 1350, + 664, + 1350, + 916, + 351, + 916 + ], + "score": 0.984, + "html": "
Hopper-v2
#Demo5 10 20 50
Expert3566±1.24
BC1471.40 ± 637.251318.76± 804.361282.46± 772.243525.87 ± 160.74
GAIL3300.32 ± 331.613372.66 ± 130.753201.97 ± 295.273363.97± 262.77
BC-GAIL3122.23± 358.653132.11 ± 520.653111.42 ± 414.283130.82± 554.54
AIRL4.12 ± 0.013.07 ± 0.024.11 ± 0.013.31 ± 0.02
Our init2322.49 ± 300.933412.58 ± 450.973314.03± 310.323601.16± 300.14
Our final3092.26± 670.723539.56±130.363516.81± 280.983610.19± 150.74
" + }, + { + "category_id": 5, + "poly": [ + 331, + 1759, + 1368, + 1759, + 1368, + 2010, + 331, + 2010 + ], + "score": 0.984, + "html": "
HalfCheetah-v2
#Demo5 10 20 50
Expert12294.22 ± 208.41
BC225.42 ± 147.16971.42 ± 249.622782.76± 959.674813.20 ±1949.26
GAIL-84.92 ± 43.29474.42 ± 389.30-116.70± 34.14-175.83 ± 26.76
BC-GAIL1362.59 ± 1255.57578.85 ± 934.343744.32 ± 1471.901597.51 ± 1173.93
AIRL782.36± 48.98-146.46± 23.571437.25 ± 25.45755.46 ±10.92
Our init267.71± 90.381064.44± 227.323200.80± 520.047102.74 ± 910.54
Our final513.66 ± 15.311616.34 ± 180.766059.27 ± 344.418817.32 ± 860.55
" + }, + { + "category_id": 5, + "poly": [ + 337, + 1030, + 1362, + 1030, + 1362, + 1281, + 337, + 1281 + ], + "score": 0.983, + "html": "
Walker2d-v2
#Demo5 10 20 50
Expert5070.97士209.19
BC1617.34 ± 693.634425.50± 930.624689.30± 372.334796.24 ± 490.05
GAIL1307.21 ± 388.55692.16±145.341991.58 ± 446.66751.21 ± 150.18
BC-GAIL3454.91 ± 792.402094.68 ±1425.053482.31 ± 828.212896.50± 828.18
AIRL-7.13 ± 0.11-7.39 ±0.09-3.74 ± 0.13-4.64 ± 0.09
Our init1859.10± 720.442038.90±260.784509.82 ± 1470.654757.58 ± 880.45
Our final2681.20± 530.673764.14 ± 470.014778.82 ± 760.344780.73± 360.66
" + }, + { + "category_id": 5, + "poly": [ + 412, + 300, + 1287, + 300, + 1287, + 552, + 412, + 552 + ], + "score": 0.98, + "html": "
Swimmer-v2
#Demo5 10 20 50
Expert332.88 ± 1.24
BC328.85 ± 2.26331.17 ± 2.4332.17 ± 2.4330.65± 2.42
GAIL304.64 ± 3.16271.59 ± 11.7756.16 ± 5.99246.73 ± 5.76
BC-GAIL313.80± 3.42326.58 ± 7.87294.93±12.21315.68 ± 9.99
AIRL332.11 ± 2.57338.43±3.65335.67 ± 2.72340.08 ± 2.70
Our init332.36± 3.62335.78± 0.34336.23± 2.53334.03 ± 2.11
Our final332.22 ± 3.23339.67 ± 3.21336.18 ± 1.87336.31± 3.20
" + }, + { + "category_id": 5, + "poly": [ + 315, + 1394, + 1384, + 1394, + 1384, + 1646, + 315, + 1646 + ], + "score": 0.979, + "html": "
Ant-v2
#Demo5 10 20 50
Expert6190.90 ± 254.18
BC3958.20 ± 661.283948.88± 753.415424.01 ± 473.055852.79± 572.97
GAIL340.02 ± 59.02335.25 ± 89.19314.35 ± 52.13284.18 ± 32.40
BC-GAIL-1081.30± 673.65-1177.27 ± 618.67-13618.45 ± 4237.79-1166.16±1246.79
AIRL-839.32 ± -301.54-386.43±156.98-586.07 ± 145.43-393.90± 145.13
Our init1150.82 ± 200.873015.43± 300.705200.58± 870.745849.88 ± 890.56
Our final1693.59± 350.743983.34± 250.995980.37± 420.165988.65± 470.03
" + }, + { + "category_id": 6, + "poly": [ + 466, + 1727, + 1231, + 1727, + 1231, + 1757, + 466, + 1757 + ], + "score": 0.923 + }, + { + "category_id": 6, + "poly": [ + 514, + 1363, + 1184, + 1363, + 1184, + 1393, + 514, + 1393 + ], + "score": 0.923 + }, + { + "category_id": 6, + "poly": [ + 487, + 270, + 1210, + 270, + 1210, + 299, + 487, + 299 + ], + "score": 0.919 + }, + { + "category_id": 6, + "poly": [ + 500, + 634, + 1197, + 634, + 1197, + 664, + 500, + 664 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.903 + }, + { + "category_id": 6, + "poly": [ + 489, + 997, + 1208, + 997, + 1208, + 1028, + 489, + 1028 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 865, + 2087, + 865, + 2113, + 835, + 2113 + ], + "score": 0.871 + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1722.0, + 1234.0, + 1722.0, + 1234.0, + 1763.0, + 464.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1358.0, + 1185.0, + 1358.0, + 1185.0, + 1399.0, + 514.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 264.0, + 1211.0, + 264.0, + 1211.0, + 306.0, + 489.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 628.0, + 1199.0, + 628.0, + 1199.0, + 672.0, + 499.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 991.0, + 1212.0, + 991.0, + 1212.0, + 1036.0, + 487.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 337, + 280, + 1363, + 280, + 1363, + 533, + 337, + 533 + ], + "score": 0.985, + "html": "
Humanoid-v2
#Demo5 1020 50
Expert5286.21± 145.98
BC1521.55± 272.143491.07± 518.644686.05 ±355.744746.88 ±605.61
GAIL485.92± 27.59486.44 ±27.18477.15± 22.07481.14± 24.37
BC-GAIL363.68 ±44.44410.03±33.07487.99± 30.77464.91±33.21
AIRL79.72± 4.2787.15 ± 5.01-1293.86±10.7084.84±6.46
Ourinit452.31± 190.121517.63 ± 110.454610.25± 2750.864776.83 ± 1320.46
Our final1225.58± 210.882190.43± 280.184716.91 ±680.294780.07± 700.01
" + }, + { + "category_id": 1, + "poly": [ + 297, + 646, + 1405, + 646, + 1405, + 767, + 297, + 767 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 986, + 1405, + 986, + 1405, + 1105, + 297, + 1105 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1944, + 1403, + 1944, + 1403, + 2036, + 299, + 2036 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 296, + 1370, + 1405, + 1370, + 1405, + 1491, + 296, + 1491 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 677, + 1132, + 1016, + 1132, + 1016, + 1282, + 677, + 1282 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 661, + 1854, + 1038, + 1854, + 1038, + 1922, + 661, + 1922 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 295, + 1503, + 1395, + 1503, + 1395, + 1565, + 295, + 1565 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 296, + 1782, + 1396, + 1782, + 1396, + 1844, + 296, + 1844 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 298, + 1707, + 1404, + 1707, + 1404, + 1769, + 298, + 1769 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 710, + 1661, + 992, + 1661, + 992, + 1700, + 710, + 1700 + ], + "score": 0.94 + }, + { + "category_id": 4, + "poly": [ + 548, + 1311, + 1151, + 1311, + 1151, + 1345, + 548, + 1345 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 300, + 911, + 1401, + 911, + 1401, + 972, + 300, + 972 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 664, + 1573, + 1038, + 1573, + 1038, + 1612, + 664, + 1612 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 300, + 1618, + 1128, + 1618, + 1128, + 1653, + 300, + 1653 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 300, + 583, + 1098, + 583, + 1098, + 616, + 300, + 616 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1871, + 1399, + 1871, + 1399, + 1901, + 1352, + 1901 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1578, + 1399, + 1578, + 1399, + 1607, + 1366, + 1607 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1666, + 1399, + 1666, + 1399, + 1695, + 1366, + 1695 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.875 + }, + { + "category_id": 6, + "poly": [ + 475, + 250, + 1220, + 250, + 1220, + 280, + 475, + 280 + ], + "score": 0.853 + }, + { + "category_id": 0, + "poly": [ + 297, + 806, + 1236, + 806, + 1236, + 879, + 297, + 879 + ], + "score": 0.762 + }, + { + "category_id": 1, + "poly": [ + 297, + 806, + 1236, + 806, + 1236, + 879, + 297, + 879 + ], + "score": 0.16 + }, + { + "category_id": 13, + "poly": [ + 405, + 1812, + 582, + 1812, + 582, + 1845, + 405, + 1845 + ], + "score": 0.95, + "latex": "f _ { 1 2 } \\stackrel { - } { = } f ( \\stackrel { - } { s _ { 1 } } , s _ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 629, + 1619, + 853, + 1619, + 853, + 1654, + 629, + 1654 + ], + "score": 0.94, + "latex": "\\phi ( s ) = g ( s ) - h ( s )" + }, + { + "category_id": 13, + "poly": [ + 904, + 1619, + 1063, + 1619, + 1063, + 1654, + 904, + 1654 + ], + "score": 0.93, + "latex": "\\psi ( s ) = \\gamma h ( s )" + }, + { + "category_id": 13, + "poly": [ + 347, + 2005, + 445, + 2005, + 445, + 2035, + 347, + 2035 + ], + "score": 0.93, + "latex": "\\psi _ { 2 } > \\psi _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 634, + 1812, + 765, + 1812, + 765, + 1845, + 634, + 1845 + ], + "score": 0.93, + "latex": "\\phi _ { 1 } = { \\bar { \\phi } } ( s _ { 1 } )" + }, + { + "category_id": 14, + "poly": [ + 661, + 1571, + 1040, + 1571, + 1040, + 1611, + 661, + 1611 + ], + "score": 0.93, + "latex": "f _ { \\theta } ( s , s ^ { \\prime } ) = g ( s ) + \\gamma h ( s ^ { \\prime } ) - h ( s )" + }, + { + "category_id": 13, + "poly": [ + 1199, + 1975, + 1312, + 1975, + 1312, + 2006, + 1199, + 2006 + ], + "score": 0.92, + "latex": "f _ { 1 2 } > f _ { 1 1 }" + }, + { + "category_id": 13, + "poly": [ + 781, + 2003, + 881, + 2003, + 881, + 2035, + 781, + 2035 + ], + "score": 0.92, + "latex": "\\psi _ { 1 } > \\psi _ { 2 }" + }, + { + "category_id": 14, + "poly": [ + 657, + 1852, + 1041, + 1852, + 1041, + 1921, + 657, + 1921 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { f _ { 1 2 } = \\phi _ { 1 } + \\psi _ { 2 } , f _ { 1 1 } = \\phi _ { 1 } + \\psi _ { 1 } } \\\\ { f _ { 2 1 } = \\phi _ { 2 } + \\psi _ { 1 } , f _ { 2 2 } = \\phi _ { 2 } + \\psi _ { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 526, + 2005, + 639, + 2005, + 639, + 2035, + 526, + 2035 + ], + "score": 0.92, + "latex": "f _ { 2 1 } > f _ { 2 2 }" + }, + { + "category_id": 13, + "poly": [ + 346, + 1975, + 459, + 1975, + 459, + 2005, + 346, + 2005 + ], + "score": 0.91, + "latex": "f _ { 1 2 } > f _ { 1 1 }" + }, + { + "category_id": 13, + "poly": [ + 1070, + 648, + 1184, + 648, + 1184, + 679, + 1070, + 679 + ], + "score": 0.91, + "latex": "\\beta = 0 . 0 5" + }, + { + "category_id": 14, + "poly": [ + 707, + 1661, + 992, + 1661, + 992, + 1700, + 707, + 1700 + ], + "score": 0.9, + "latex": "f _ { \\theta } ( s , s ^ { \\prime } ) = \\phi ( s ) + \\psi ( s ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 510, + 1975, + 622, + 1975, + 622, + 2005, + 510, + 2005 + ], + "score": 0.9, + "latex": "f _ { 2 1 } > f _ { 2 2 }" + }, + { + "category_id": 13, + "poly": [ + 613, + 736, + 706, + 736, + 706, + 764, + 613, + 764 + ], + "score": 0.9, + "latex": "\\sigma = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 479, + 1405, + 508, + 1405, + 508, + 1431, + 479, + 1431 + ], + "score": 0.85, + "latex": "s _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 398, + 1405, + 426, + 1405, + 426, + 1431, + 398, + 1431 + ], + "score": 0.84, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1220, + 648, + 1241, + 648, + 1241, + 679, + 1220, + 679 + ], + "score": 0.83, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 830, + 1137, + 861, + 1137, + 861, + 1156, + 830, + 1156 + ], + "score": 0.65, + "latex": "a _ { 2 1 }" + }, + { + "category_id": 13, + "poly": [ + 830, + 1262, + 861, + 1262, + 861, + 1281, + 830, + 1281 + ], + "score": 0.45, + "latex": "a _ { 1 2 }" + }, + { + "category_id": 13, + "poly": [ + 681, + 1200, + 711, + 1200, + 711, + 1218, + 681, + 1218 + ], + "score": 0.43, + "latex": "a _ { 1 1 }" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1130.0, + 829.0, + 1130.0, + 829.0, + 1160.0, + 826.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1130.0, + 865.0, + 1130.0, + 865.0, + 1160.0, + 862.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1186.0, + 955.0, + 1186.0, + 955.0, + 1228.0, + 913.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1192.0, + 1018.0, + 1192.0, + 1018.0, + 1222.0, + 978.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.25, + 1191.0, + 711.25, + 1191.0, + 711.25, + 1221.0, + 679.25, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1190.0, + 763.0, + 1190.0, + 763.0, + 1226.5, + 752.0, + 1226.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1255.0, + 864.0, + 1255.0, + 864.0, + 1284.0, + 827.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1308.0, + 1154.0, + 1308.0, + 1154.0, + 1347.0, + 546.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 577.0, + 1103.0, + 577.0, + 1103.0, + 624.0, + 292.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 245.0, + 1222.0, + 245.0, + 1222.0, + 286.0, + 476.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 804.0, + 1243.0, + 804.0, + 1243.0, + 848.0, + 291.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 847.0, + 558.0, + 847.0, + 558.0, + 881.0, + 354.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 646.0, + 1069.0, + 646.0, + 1069.0, + 680.0, + 296.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 646.0, + 1219.0, + 646.0, + 1219.0, + 680.0, + 1185.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 646.0, + 1405.0, + 646.0, + 1405.0, + 680.0, + 1242.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 675.0, + 1405.0, + 675.0, + 1405.0, + 709.0, + 296.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 705.0, + 1403.0, + 705.0, + 1403.0, + 738.0, + 295.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 731.0, + 612.0, + 731.0, + 612.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 731.0, + 1192.0, + 731.0, + 1192.0, + 769.0, + 707.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 987.0, + 1406.0, + 987.0, + 1406.0, + 1021.0, + 296.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1014.0, + 1403.0, + 1014.0, + 1403.0, + 1048.0, + 295.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1040.0, + 1406.0, + 1040.0, + 1406.0, + 1081.0, + 292.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1072.0, + 577.0, + 1072.0, + 577.0, + 1110.0, + 295.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1980.0, + 294.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 345.0, + 1971.0, + 345.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 1971.0, + 509.0, + 1971.0, + 509.0, + 2010.0, + 460.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1971.0, + 1198.0, + 1971.0, + 1198.0, + 2010.0, + 623.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2010.0, + 1313.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 346.0, + 2002.0, + 346.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 2002.0, + 525.0, + 2002.0, + 525.0, + 2039.0, + 446.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 2002.0, + 780.0, + 2002.0, + 780.0, + 2039.0, + 640.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 2002.0, + 1172.0, + 2002.0, + 1172.0, + 2039.0, + 882.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1371.0, + 1403.0, + 1371.0, + 1403.0, + 1406.0, + 295.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1400.0, + 397.0, + 1400.0, + 397.0, + 1435.0, + 295.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1400.0, + 478.0, + 1400.0, + 478.0, + 1435.0, + 427.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1400.0, + 1403.0, + 1400.0, + 1403.0, + 1435.0, + 509.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1429.0, + 1403.0, + 1429.0, + 1403.0, + 1464.0, + 295.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1455.0, + 983.0, + 1455.0, + 983.0, + 1494.0, + 294.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1502.0, + 1403.0, + 1502.0, + 1403.0, + 1539.0, + 293.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1531.0, + 1319.0, + 1531.0, + 1319.0, + 1569.0, + 295.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1779.0, + 1402.0, + 1779.0, + 1402.0, + 1819.0, + 295.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1810.0, + 404.0, + 1810.0, + 404.0, + 1846.0, + 293.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1810.0, + 633.0, + 1810.0, + 633.0, + 1846.0, + 583.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1810.0, + 850.0, + 1810.0, + 850.0, + 1846.0, + 766.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1704.0, + 1406.0, + 1704.0, + 1406.0, + 1743.0, + 292.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1732.0, + 364.0, + 1732.0, + 364.0, + 1776.0, + 289.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 909.0, + 1403.0, + 909.0, + 1403.0, + 947.0, + 295.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 940.0, + 473.0, + 940.0, + 473.0, + 977.0, + 295.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1617.0, + 628.0, + 1617.0, + 628.0, + 1657.0, + 295.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1617.0, + 903.0, + 1617.0, + 903.0, + 1657.0, + 854.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1617.0, + 1129.0, + 1617.0, + 1129.0, + 1657.0, + 1064.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 804.0, + 1243.0, + 804.0, + 1243.0, + 848.0, + 291.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 847.0, + 558.0, + 847.0, + 558.0, + 881.0, + 354.0, + 881.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/ryxB2lBtvH/ryxB2lBtvH.md b/parse/train/ryxB2lBtvH/ryxB2lBtvH.md new file mode 100644 index 0000000000000000000000000000000000000000..e1688e40606fc8bb4f06b9d9a02f33c5d26e1203 --- /dev/null +++ b/parse/train/ryxB2lBtvH/ryxB2lBtvH.md @@ -0,0 +1,356 @@ +# LEARNING TO COORDINATE MANIPULATION SKILLS VIA SKILL BEHAVIOR DIVERSIFICATION + +Youngwoon Lee, Jingyun Yang, Joseph J. Lim + +Department of Computer Science University of Southern California {lee504,jingyuny,limjj}@usc.edu + +# ABSTRACT + +When mastering a complex manipulation task, humans often decompose the task into sub-skills of their body parts, practice the sub-skills independently, and then execute the sub-skills together. Similarly, a robot with multiple end-effectors can perform complex tasks by coordinating sub-skills of each end-effector. To realize temporal and behavioral coordination of skills, we propose a modular framework that first individually trains sub-skills of each end-effector with skill behavior diversification, and then learns to coordinate end-effectors using diverse behaviors of the skills. We demonstrate that our proposed framework is able to efficiently coordinate skills to solve challenging collaborative control tasks such as picking up a long bar, placing a block inside a container while pushing the container with two robot arms, and pushing a box with two ant agents. Videos and code are available at https://clvrai.com/coordination. + +# 1 INTRODUCTION + +Imagine you wish to play Chopin’s Fantaisie Impromptu on the piano. With little prior knowledge about the piece, you would first practice playing the piece with each hand separately. After independently mastering the left and right hand parts, you would move on to practicing with both hands simultaneously. To find the synchronized and non-interfering movements of two hands, you would try variable ways of playing the same melody with each hand, and eventually create a complete piece of music. Through the decomposition of skills into sub-skills of two hands and learning variations of sub-skills, humans make the learning process of manipulation skills much faster than learning everything at once. + +Can autonomous agents efficiently learn complicated tasks with coordination of different skills from multiple end-effectors like humans? Learning to perform collaborative and composite tasks from scratch requires a huge amount of environment interaction and extensive reward engineering, which often results in undesired behaviors (Riedmiller et al., 2018). Hence, instead of learning a task at once, modular approaches (Andreas et al., 2017; Oh et al., 2017; Frans et al., 2018; Lee et al., 2019; Peng et al., 2019; Goyal et al., 2020) suggest to learn reusable primitive skills and solve more complex tasks by recombining the skills. However, all these approaches either focus on working with single end-effector manipulation or single agent locomotion, and these do not scale to multi-agent problems. + +To this end, we propose a modular framework that learns to coordinate multiple end-effectors with their primitive skills for various robotics tasks, such as bimanual manipulation. The main challenge is that naive simultaneous execution of primitive skills from multiple end-effectors can often cause unintended behaviors (e.g. collisions between end-effectors). Thus, as illustrated in Figure 1, our model needs to learn to appropriately coordinate end-effectors; and hence needs a way to obtain, represent, and control detailed behaviors of each primitive skill. Inspired by these intuitions, our method consists of two parts: (1) acquiring primitive skills with diverse behaviors by mutual information maximization, and (2) learning a meta policy that selects a skill for each end-effector and coordinates the chosen skills by controlling the behavior of each skill. + +The main contribution of this paper is a modular and hierarchical approach that tackles cooperative manipulation tasks with multiple end-effectors by (1) learning primitive skills of each end-effector independently with skill behavior diversification and (2) coordinating end-effectors using diverse behaviors of the skills. Our empirical results indicate that our proposed method is able to efficiently learn primitive skills with diverse behaviors and coordinate these skills to solve challenging collaborative control tasks such as picking up a long bar, placing a block inside the container on the right side, and pushing a box with two ant agents. We provide additional qualitative results and code at https://clvrai.com/coordination. + +![](images/7e1b9f5f33a08c620e270f4735ec881e08b70e6cf015791f9524143e7c2eda6a.jpg) +Figure 1: Composing complex skills using multiple agents’ primitive skills requires proper coordination between agents since concurrent execution of primitive skills requires temporal and behavioral coordination. For example, to move a block into a container on the other end of the table, the agent needs to not only utilize pick, place, and push primitive skills at the right time but also select the appropriate behaviors for these skills, represented as latent vectors $z _ { 1 }$ , $z _ { 2 }$ , $z _ { 3 }$ , and $z _ { 4 }$ above. Naive methods neglecting either temporal or behavioral coordination will produce unintended behaviors, such as collisions between end-effectors. + +# 2 RELATED WORK + +Deep reinforcement learning (RL) for continuous control is an active research area. However, learning a complex task either from a sparse reward or a heavily engineered reward becomes computationally impractical as the target task becomes complicated. Instead of learning from scratch, complex tasks can be tackled by decomposing the tasks into easier and reusable sub-tasks. Hierarchical reinforcement learning temporally splits a task into a sequence of temporally extended meta actions. It often consists of one meta policy (high-level policy) and a set of low-level policies, such as options framework (Sutton et al., 1999). The meta policy decides which low-level policy to activate and the chosen low-level policy generates an action sequence until the meta policy switches it to another lowlevel policy. Options can be discovered without supervision (Schmidhuber, 1990; Bacon et al., 2017; Nachum et al., 2018; Levy et al., 2019), meta-learned (Frans et al., 2018), pre-defined (Kulkarni et al., 2016; Oh et al., 2017; Merel et al., 2019; Lee et al., 2019), or attained from additional supervision signals (Andreas et al., 2017; Ghosh et al., 2018). However, option frameworks are not flexible to solve a task that requires simultaneous activation or interpolation of multiple skills since only one skill can be activated at each time step. + +To solve composite tasks multiple policies can be simultaneously activated by adding Qfunctions (Haarnoja et al., 2018a), additive composition (Qureshi et al., 2020; Goyal et al., 2020), or multiplicative composition (Peng et al., 2019). As each policy takes the whole observation as input and controls the whole agent, it is not robust to changes in unrelated parts of the observation. For example, a left arm skill can be affected by the pose change in the right arm, which is not relevant to the left arm skill. Hence, these skill composition approaches can fail when an agent encounters a new combination of skills or a new skill is introduced since the agent will experience unseen observations. + +Instead of having a policy with the full observation and action space, multi-agent reinforcement learning (MARL) suggests to explicitly split the observation and action space according to agents (e.g. robots or end-effectors), which allows efficient low-level policy training as well as flexible skill composition. For cooperative tasks, communication mechanisms (Sukhbaatar et al., 2016; Peng et al., 2017; Jiang & Lu, 2018), sharing policy parameters (Gupta et al., 2017), and decentralized actor with centralized critic (Lowe et al., 2017; Foerster et al., 2018) have been actively used. However, these approaches suffer from the credit assignment problem (Sutton, 1984) among agents and the lazy agent problem (Sunehag et al., 2018). As agents have more complicated morphology and larger observation space, learning a policy for a multi-agent system from scratch requires extremely long training time. Moreover, the credit assignment problem becomes more challenging when the complexity of cooperative tasks increases and all agents need to learn completely from scratch. To resolve these issues, we propose to first train reusable skills for each agent in isolation, instead of learning primitive skills of multiple agents together. Then, we recombine these skills (Maes & Brooks, 1990) to complete more complicated tasks with learned coordination of the skills. + +![](images/63021c65e6e95a999eba964c17a1759f617f5364c3d76c2f449409b6d5dca506.jpg) +Figure 2: Our method is composed of two components: a meta policy and a set of agent-specific primitive policies relevant to task completion. The meta policy selects which primitive skill to run for each agent as well as the behavior embedding (i.e. variation in behavior) of the chosen primitive skill. Each selected primitive skill takes as input the agent observation and the behavior embedding and outputs action for that agent. + +To coordinate skills from multiple agents, the skills have to be flexible; hence, a skill can be adjusted to collaborate with other agents’ skills. Maximum entropy policies (Haarnoja et al., 2017; 2018a;b) can learn diverse ways to achieve a goal by maximizing not only reward but also entropy of the policy. In addition, Eysenbach et al. (2019) proposes to discover diverse skills without reward by maximizing entropy as well as mutual information between resulting states and latent representations of skills (i.e. skill embeddings). Our method leverages the maximum entropy policy (Haarnoja et al., 2018b) with the discriminability objective (Eysenbach et al., 2019) to learn a primitive skill with diverse behaviors conditioned on a controllable skill embedding. This controllable skill embedding will be later used as a behavior embedding for the meta policy to adjust a primitive skill’s behavior for coordination. + +# 3 METHOD + +In this paper, we address the problem of solving cooperative manipulation tasks that require collaboration between multiple end-effectors or agents. Note that we use the terms “end-effector” and “agent” interchangeably in this paper. Instead of learning a multi-agent task from scratch (Lowe et al., 2017; Gupta et al., 2017; Sunehag et al., 2018; Foerster et al., 2018), modular approaches (Andreas et al., 2017; Frans et al., 2018; Peng et al., 2019) suggest to learn reusable primitive skills and solve more complex tasks by recombining these skills. However, concurrent execution of primitive skills of multiple agents fails when agents never experienced a combination of skills during the pre-training stage, or skills require temporal or behavioral coordination. + +Therefore, we propose a modular and hierarchical framework that learns to coordinate multiple agents with primitive skills to perform a complex task. Moreover, during primitive skill training, we propose to learn a latent behavior embedding, which provides controllability of each primitive skill to the meta policy while coordinating skills. In Section 3.2, we describe our modular framework in detail. Next, in Section 3.3, we elaborate how controllable primitive skills can be acquired. Lastly, we describe how the meta policy learns to coordinate primitive skills in Section 3.4. + +# 3.1 PRELIMINARIES + +We formulate our problem as a Markov decision process defined by a tuple $\{ S , { \mathcal { A } } , { \mathcal { T } } , R , \rho , \gamma \}$ of states, actions, transition probability, reward, initial state distribution, and discount factor. In our formulation, we assume the environment includes $N$ agents. To promote consistency in our terminology, we use superscripts to denote the index of agent and subscripts to denote time or primitive skill index. Hence, the state space and action space for an agent $i$ can be represented as $S ^ { \imath }$ and $A ^ { i }$ where each element of $S ^ { i }$ is a subset of the corresponding element in $\boldsymbol { S }$ and $\mathcal { A } = \mathcal { A } ^ { 1 } \times \mathcal { A } ^ { 2 } \times \cdot \cdot \cdot \times \mathcal { A } ^ { N }$ , respectively. For each agent $i$ , we provide a set of $m ^ { i }$ skills, $\Pi ^ { i } = \{ \pi _ { 1 } ^ { i } , \ldots , \pi _ { m ^ { i } } ^ { i } \}$ . A policy of an agent $i$ is represented as $\pi _ { c _ { t } ^ { i } } ^ { i } \left( a _ { t } ^ { i } | s _ { t } ^ { i } \right) \in \Pi ^ { i }$ where is sam $c _ { t } ^ { i }$ is a skiled from index, , and t $s _ { t } ^ { i } \in S ^ { i }$ is a state, and agents take act $a _ { t } ^ { i } \in \mathcal A ^ { i }$ action at time sampled from $t$ . composite policy $\pi ( a _ { t } ^ { 1 } , \bar { a } _ { t } ^ { 2 } , \dots , a _ { t } ^ { N } | s _ { t } , c _ { t } ^ { 1 } , c _ { t } ^ { 2 } , \dots , \bar { c } _ { t } ^ { N } ) = ( \pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( a _ { t } ^ { 1 } | s _ { t } ) , \pi _ { c _ { t } ^ { 2 } } ^ { 2 } ( a _ { t } ^ { 2 } | s _ { t } ) , \dots , \pi _ { c _ { t } ^ { N } } ^ { N } ( a _ { t } ^ { N } | s _ { t } ) )$ $\rho$ $N$ $a _ { t } ^ { 1 } , a _ { t } ^ { 2 } , \ldots , a _ { t } ^ { N }$ a singl, where reward is the e $r _ { t }$ . The performance is evaluated based on a discounted return sode horizon. $R =$ $\textstyle \sum _ { t = 0 } ^ { T - 1 } \gamma ^ { t } r _ { t }$ $T$ + +![](images/852e7eff92afaaa227e4f7f1a58fdeda7147bb405d07ee79a206747f65017adf.jpg) +Figure 3: Different multi-agent architectures. (a) The vanilla RL method considers all agents as a monolithic agent; thus a single policy takes the full observation as input and outputs the full action. (b) The multi-agent RL method (MARL) consists of $N$ policies that operate on the observations and actions of corresponding agents. (c) The modular network consists of $N$ sets of skills for the $N$ agents trained in isolation and a meta policy that selects a skill for each agent. (d-f) The RL, MARL, and modular network methods augmented with skill behavior diversification (SBD) has a meta policy that outputs a skill behavior embedding vector $z$ for each skill. + +# 3.2 MODULAR FRAMEWORK + +As illustrated in Figure 2, our model is composed of two components: a meta policy $\pi _ { m e t a }$ and a set of primitive skills of $N$ agents $\Pi ^ { 1 } , \dots , \Pi ^ { N }$ . Note that each primitive skill $\pi _ { c ^ { i } } ^ { i } \in \Pi ^ { i }$ contains variants of behaviors parameterized by an $N _ { z }$ -dimensional latent behavior embedding $z ^ { i }$ (see Section 3.3). The meta policy selects a skill to execute for each agent, rather than selecting one primitive skill for the entire multi-agent system to execute. Also, we give the meta policy the capability to select which variant of the skill to execute (see Section 3.4). Then, the chosen primitive skills are simultaneously executed for $T _ { l o w }$ time steps. + +The concurrent execution of multiple skills often leads to undesired results and therefore requires coordination between the skills. For example, naively placing a block in the left hand to a container being moved by the right hand can cause collision between the two robot arms. The arms can avoid collision while performing the skills by properly adjusting their skill behaviors (e.g. the left arm leaning to the left side while placing the block and the right arm leaning to the right side while pushing the container) as shown in Figure 1. In our method, the meta policy learns to coordinate multiple agents’ skills by manipulating the behavior embeddings (i.e. selecting a proper behavior from diverse behaviors of each skill). + +# 3.3 TRAINING AGENT-SPECIFIC PRIMITIVE SKILLS WITH DIVERSE BEHAVIORS + +To adjust a primitive skill to collaborate with other agents’ skills in a new environment, the skill needs to support variations of skill behaviors when executed at a given state. Moreover, a behavioral variation of a skill should be controllable by the meta policy for skill coordination. In order to make our primitive skill policies generate diverse behaviors controlled by a latent vector $z$ , we leverage the entropy and mutual information maximization objective introduced in Eysenbach et al. (2019). + +More specifically, a primitive policy of an agent $i$ outputs an action $a \in { \mathcal { A } }$ conditioned on the current state $s \in S$ and a latent behavior embedding $z \sim p ( z )$ , where the prior distribution $p ( z )$ is Gaussian (we omit agent $i$ in this section for the simplicity of notations). Diverse behaviors conditioned on a random sample $z$ can be achieved by maximizing the mutual information between behaviors and states $M I ( s , z )$ , while minimizing the mutual information between behaviors and actions given the state $M I ( a , z | s )$ , together with maximizing the entropy of the policy $\mathcal { H } ( a | s )$ to encourage diverse behaviors. The objective can be written as follows (we refer the readers to Eysenbach et al. (2019) + +# Algorithm 1 ROLLOUT + +1: Input: Meta policy $\pi _ { m e t a }$ , sets of primitive policies $\Pi ^ { 1 } , . . . , \Pi ^ { N }$ , and meta horizon $T _ { l o w }$ +2: Initialize an episode $t \gets 0$ and receive initial state $s _ { 0 }$ +3: while episode is not terminated do +4: Sample skill indexes and behavior embeddings $( c _ { t } ^ { 1 } , \ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \ldots , z _ { t } ^ { N } ) \sim \pi _ { m e t a } ( s _ { t } )$ +5: $\tau 0$ +6: while $\tau < T _ { l o w }$ and episode is not terminated do +7: $\boldsymbol a _ { t + \tau } = ( a _ { t + \tau } ^ { 1 } , \dots , \dot { a _ { t + \tau } ^ { N } } ) \sim ( \pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( s _ { t + \tau } , z _ { t } ^ { 1 } ) , \dots , \pi _ { c _ { t } ^ { N } } ^ { N } ( s _ { t + \tau } , z _ { t } ^ { N } ) )$ +8: $\begin{array} { r l } & { s _ { t + \tau + 1 } , r _ { t + \tau } \gets \mathrm { E N V } ( s _ { t + \tau } , a _ { t + \tau } ) } \\ & { \tau \gets \tau + 1 } \end{array}$ +9: +10: end while +11: Add a transition $s _ { t } , ( c _ { t } ^ { 1 } , \ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \ldots , z _ { t } ^ { N } ) , s _ { t + \tau } , r _ { t : t + \tau - 1 }$ to the rollout buffer $\boldsymbol { B }$ +12: $t \gets t + \tau$ +13: end while + +for derivation): + +$$ +\begin{array} { r l } & { \mathcal { F } ( \theta ) = M I ( s , z ) - M I ( a , z \vert s ) + \mathcal { H } ( a \vert s ) = \mathcal { H } ( a \vert s , z ) - \mathcal { H } ( z \vert s ) + \mathcal { H } ( z ) } \\ & { \quad \quad = \mathcal { H } ( a \vert s , z ) + \mathbb { E } _ { z \sim p ( z ) , s \sim \pi ( z ) } [ \log p ( z \vert s ) ] - \mathbb { E } _ { z \sim p ( z ) } [ \log p ( z ) ] } \\ & { \quad \quad \geq \mathcal { H } ( a \vert s , z ) + \mathbb { E } _ { z \sim p ( z ) , s \sim \pi ( z ) } [ \log q _ { \phi } ( z \vert s ) - \log p ( z ) ] , } \end{array} +$$ + +where the learned discriminator $q _ { \phi } ( z | s )$ approximates the posterior $p ( z | s )$ . + +To achieve a primitive skill with diverse behaviors, we augment Equation (3) to the environment reward: + +$$ +r _ { t } + \lambda _ { 1 } \mathcal { H } ( a | s , z ) + \lambda _ { 2 } \mathbb { E } _ { z \sim p ( z ) , s \sim \pi ( z ) } [ \log q _ { \phi } ( z | s ) - \log p ( z ) ] , +$$ + +where $\lambda _ { 1 }$ is the entropy coefficient and $\lambda _ { 2 }$ is the diversity coefficient which corresponds identifiability of behaviors. Maximizing Equation (3) encourages multi-modal exploration strategies while maximizing the reward $r _ { t }$ forces to achieve its own goal. Moreover, by maximizing identifiability of behaviors, the latent vector $z$ , named behavior embedding, can represent a variation of the learned policy and thus can be used to control the behavior of the policy. For example, when training a robot to move an object, a policy learns to move the object quickly as well as slowly, and these diverse behaviors map to different latent vectors $z$ . We empirically show that the policies with diverse behaviors achieve better compositionality with other agents in our experiments. + +# 3.4 COMPOSING PRIMITIVE SKILLS WITH META POLICY + +We denote the meta policy as $\pi _ { m e t a } ( c ^ { 1 } , \ldots , c ^ { N } , z ^ { 1 } , \ldots , z ^ { N } | s _ { t } )$ , where $c ^ { i } \in [ 1 , m ^ { i } ]$ represents a skill index of an agent $i \in [ 1 , N ]$ and $z ^ { i } \in \mathbb { R } ^ { N _ { z } }$ represents a behavior embedding of the skill. Every $T _ { l o w }$ time steps, the meta policy chooses one primitive skill $\pi _ { c ^ { i } } ^ { i } \in \Pi ^ { i }$ for each agent $i$ . Also, the meta policy outputs a set of latent behavior embeddings $( z ^ { 1 } , z ^ { 2 } , \dots , z ^ { N } )$ and feeds them to the corresponding skills (i.e. $\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )$ for agent $i$ ). Once a set of primitive skills $\{ \pi _ { c ^ { 1 } } ^ { 1 } , \ldots , \pi _ { c ^ { N } } ^ { N } \}$ are chosen to be executed, each primitive skill generates an action $a ^ { i } \sim \pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )$ based on the current state $s ^ { i }$ and the latent vector $z ^ { i }$ . Algorithm 1 illustrates the overall rollout process. + +Since there are a finite number of skills for each agent to execute, the meta action space for each agent $[ 1 , m ^ { i } ]$ is discrete, while the behavior embedding space for each agent $\mathbb { R } ^ { N _ { z } }$ is continuous. Thus, the meta policy is modeled as a $( 2 \times N )$ -head neural network where the first $N$ heads represent $m ^ { i }$ -way categorical distributions for skill selection and the last $N$ heads represent $N _ { z }$ -dimensional Gaussian distributions for behavior control of the chosen skill. + +# 3.5 IMPLEMENTATION + +We model the primitive policies and posterior distributions $q _ { \phi }$ as neural networks. We train the primitive policies using soft actor-critic (Haarnoja et al., 2018b). When we train a primitive policy, we use a unit Gaussian distribution as the prior distribution of latent variables $p ( z )$ . We use 5 as the size of latent behavior embedding $N _ { z }$ . Each primitive policy outputs the mean and standard deviation of a Gaussian distribution over an action space. For a primitive policy, we apply tanh activation to normalize the action between $[ - 1 , 1 ]$ . We model the meta policy as neural network with multiple heads that output the skill index $\bar { c } ^ { i }$ and behavior embedding $z ^ { i }$ for each agent. The meta policy is trained using PPO (Schulman et al., 2017; 2016; Dhariwal et al., 2017). All policy networks in this paper consist of 3 fully connected layers of 64 hidden units with ReLU nonlinearities. The discriminator $q _ { \phi }$ in Equation (4) is a 2-layer fully connected network with 64 hidden units. + +![](images/11edbefb96b52e765ac0cdb4e36075acafa79155b0f2fabe7fb7cdafaf5487ff.jpg) +Figure 4: The composite tasks pose a challenging combination of object manipulation and locomotion skills, which requires coordination of multiple agents and temporally extended behaviors. (a) The left Jaco arm needs to pick up a block while the right Jaco arm pushes a container, and then it places the block into the container. (b) Two Jaco arms are required to pick and place a bar-shaped block together. (c) Two ants push the red box to the goal location (green circle) together. + +# 4 EXPERIMENTS + +To demonstrate the effectiveness of our framework, we compare our method to prior methods in the field of multi-agent RL and ablate the components of our framework to understand their importance. We conducted experiments on a set of challenging robot control environments that require coordination of different agents to complete collaborative robotic manipulation and locomotion tasks. + +Through our experiments, we aim to answer the following questions: (1) can our framework efficiently learn to combine primitive skills to execute a complicated task; (2) can our learned agent exhibit collaborative behaviors during task execution; and (3) can our framework leverage the controllable behavior variations of the primitive skills to achieve better coordination? + +For details about environments and training, please refer to the supplementary material. As the performance of training algorithms varies between runs, we train each method on each task with 6 different random seeds and report mean and standard deviation of each method’s success rate. + +# 4.1 BASELINES + +We compare the performance of our method with various single- and multi-agent RL methods illustrated in Figure 3: + +Single-agent RL (RL): A vanilla RL method where a single policy takes as input the full observation and outputs all agents’ actions. + +Multi-agent RL (MARL): A multi-agent RL method where each of $N$ policies takes as input the observation of the corresponding agent and outputs an action for that agent. All policies share the global critic learned from a single task reward (Lowe et al., 2017). + +Modular Framework (Modular): A modular framework composed of a meta policy and $N$ sets of primitive skills (i.e. one or more primitive skills per agent). Every $T _ { l o w }$ time steps, the meta policy selects a primitive skill for each agent based on the full observation. Then, the chosen skills are executed for $T _ { l o w }$ time steps. + +![](images/38d7a3b3c6d72759e19da2cea6bad148d9e0ead37b5a570252e66d5ccb027d90.jpg) +Figure 5: Success rates of our method (Modular-SBD) and baselines. For modular frameworks (Modular and Modular-SBD), we shift the learning curves rightwards the total number of environment steps the agent takes to learn the primitive skills (0.9 M, $1 . 2 \mathbf { M }$ , and $2 . 0 \mathbf { M }$ , respectively). Our method substantially improves learning speed and performance on JACO PICK-PUSH-PLACE and ANT PUSH. The shaded areas represent the standard deviation of results from six different seeds. The curves are smoothed using moving average over 10 runs. + +Single-agent RL with Skill Behavior Diversification (RL-SBD): An RL method augmented with the behavior diversification objective. A meta policy is employed to generate a behavior embedding for a low-level policy, and the low-level policy outputs all agents’ actions conditioned on the behavior embedding and the full observation for $T _ { l o w }$ time steps. The meta policy and the low-level policy are jointly trained with the behavior diversification objective described in Equation (4). + +Multi-agent RL with Skill Behavior Diversification (MARL-SBD): A MARL method augmented with the behavior diversification objective. A meta policy generates $N$ behavior embeddings. Then, each low-level policy outputs each agent’s action conditioned on its observation and behavior embedding for $T _ { l o w }$ time steps. All policies are jointly trained to maximize Equation (4). + +Modular Framework with Skill Behavior Diversification (Modular-SBD, Ours): Our method which coordinates primitive skills of multiple agents. The modular framework consists of a meta policy and $N$ sets of primitive skills, where each primitive skill is conditioned on a behavior embedding $z$ . The meta policy takes as input the full observation and selects both a primitive skill and a behavior embedding for each agent. Then, each primitive skill outputs action for each agent. + +# 4.2 JACO PICK-PUSH-PLACE + +We developed JACO PICK-PUSH-PLACE and JACO BAR-MOVING environments using two Kinova Jaco arms, where each Jaco arm is a 9 DoF robotic arm with 3 fingers. JACO PICK-PUSH-PLACE starts with a block on the left and a container on the right. The robotic arms need to pick up the block, push the container to the center, and place the block inside the container. For successful completion of the task, the two Jaco arms have to concurrently execute their distinct sets of skills and dynamically adjust their picking, pushing, and placing directions to avoid collision between arms. + +Primitives skills. There are three primitive skills available to each arm: Picking up, Pushing, and Placing to center (see Figure 4a). Picking up requires a robotic arm to pick up a small block, which is randomly placed on the table. If the block is not picked up after a certain amount of time or the arm drops the block, the agent fails. Pushing learns to push a big container to its opposite side (e.g. from left to the center or from right to center). The agent fails if it cannot place the container to the center. Placing to center requires placing an object in the gripper to the table. The agent only succeeds when it stably places the object at the desired location on the container. + +Composite task. Our method (Modular-SBD) can successfully perform JACO PICK-PUSH-PLACE task while all baselines fail to compose primitive skills as shown in Figure 5a. The RL and MARL baselines cannot learn the composite task mainly because the agent requires to learn the combinatorial number of skill compositions and to solve the credit assignment problem across multiple agents. Since the composite task requires multiple primitive skills of multiple agents to be performed properly at the same time, a reward signal about a failure case cannot be assigned to the correct agent or skill. By using pre-trained primitive skills, the credit assignment problem is relaxed and all agents can + +
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
RL0.000±0.0000.000±0.0000.000±0.000
MARL0.000±0.0000.000± 0.0000.000± 0.000
RL-SBD0.000± 0.0000.000± 0.0000.000±0.000
MARL-SBD0.000 ± 0.0000.000 ±0.0000.000 ± 0.000
Modular0.324± 0.4680.917± 0.2760.003± 0.058
Modular-SBD (Ours)0.902 ± 0.2980.950 ± 0.2180.323 ± 0.468
+ +Table 1: Success rates for all tasks, comparing our method against baselines. Each entry in the table represents average success rate and standard deviation over 100 runs. The baselines learning from scratch fail to learn complex tasks with multiple agents. + +perform their skills concurrently. Therefore, the Modular baseline learns to achieve success but shows significantly lower performance than our method (Modular-SBD). This is because the lack of skill behavior diversification makes it impossible to adjust pushing and placing trajectories during skill composition time, which resulting in frequent end-effector collisions. + +# 4.3 JACO BAR-MOVING + +In JACO BAR-MOVING, two Jaco arms need to pick up a long bar together, move the bar towards a target location while maintaining its rotation, and place it on the table (see Figure 4b). The initial position of the bar is randomly initialized every episode and an agent needs to find appropriate coordination between two arms for each initialization. Compared to JACO PICK-PUSH-PLACE, this task requires that the two arms synchronize their movements and perform more micro-level adjustments to their behaviors. + +Primitives skills. There are two pre-trained primitive skills available to each arm: Picking up and Placing towards arm. Picking up is same as described in Section 4.2. Placing towards arm learns to move a small block (half size of the block used in the composite task) in the hand towards the robotic arm and then place it on the table. The agent fails if it cannot place the block to the target location. + +Composite task. The JACO BAR-MOVING task requires the two arms to work very closely together. For example, the Picking up skill of both arms should be synchronized when they start to lift the bar and two arms require to lift the bar while maintaining the relative position between them since they are connected by holding the bar. The modular framework without explicit coordination of skills (Modular) can synchronize the execution of picking, moving, and placing. But the inability to micro-adjust the movement of the other arm causes instability of bar picking and moving. This results in degraded success rates compared to the modular framework with explicit coordination. Meanwhile, all baselines without pre-defined primitive skills fail to learn JACO BAR-MOVING. + +# 4.4 ANT PUSH + +We developed a multi-ant environment, ANT-PUSH, inspired from Nachum et al. (2019), simulated in the MuJoCo (Todorov et al., 2012) physics engine. We use the ant model in OpenAI Gym (Brockman et al., 2016). In this environment, two ants need to push a large object toward a green target place, collaborating with each other to keep the angle of the object as stable as possible (see Figure 4c). + +Primitives skills. We train walking skills of an ant agent in 4 directions: up, down, left, and right. During primitive skill training, a block (half size of the block used in the composite task) and an ant agent are randomly placed. Pushing the block gives an additional reward to the agent, which prevents an ant to avoid the block. The learned primitive skills have different speed and trajectories conditioned on the latent behavior embedding. + +Composite task. Our method achieves $3 2 . 3 \%$ success rate on ANT PUSH task while all baselines fail to compose primitive skills as shown in Figure 5c and Table 1. The poor performance of RL, MARL, RL-SBD, and MARL-SBD baselines shows the difficulty of credit assignment between agents, which leads one of the ants moves toward a block and pushes it but another ant does not move. Moreover, the Modular baseline with primitive skills also fails to learn the pushing task. This result illustrates the importance of coordination of agents, which helps synchronizing and controlling the velocities of both ant agents to push the block toward the goal position while maintaining its rotation. + +![](images/99d31a6b5e96ecb62faf4db19c419b513aea19dd229cabd50fa233595361425f.jpg) +Figure 6: Learning curves of our method with different diversity coefficients $\lambda _ { 2 }$ on ANT PUSH. +Figure 7: Success rates of our method with different $T _ { l o w }$ coefficients on Jaco environments. + +# 4.5 EFFECT OF DIVERSITY OF PRIMITIVE SKILLS + +To analyze the effect of the diversity of primitive skills, we compare our model with primitive skills trained with different diversity coefficients $\lambda _ { 2 } = \{ 0 . 0 , 0 . 0 5 , 0 . 1 , 0 . 5 , 1 . 0 \}$ in Equation (4) on ANT PUSH. Figure 6 shows that with small diversity coefficients $\lambda _ { 2 } = \{ 0 . 0 5 , 0 . 1 \}$ , the agent can control detailed behaviors of primitive skills while primitive skills without diversity ${ \lambda } _ { 2 } = 0$ ) cannot be coordinated. The meta policy tries to synchronize two ant agents’ positions and velocities by switching primitive skills, but it cannot achieve proper coordination without diversified skills. On the other hand, large diversity coefficients $\lambda _ { 2 } = \bar { \{ 0 . 5 , 1 . 0 \} }$ make the primitive skills often focus on demonstrating diverse behaviors and fail to achieve the goals of the skills. Hence, these primitive skills do not have enough functionality to solve the target task. The diversity coefficient needs to be carefully chosen to acquire primitive skills with good performance as well as diverse behaviors. + +# 4.6 EFFECT OF SKILL SELECTION INTERVAL $T _ { l o w }$ + +To analyze the effect of the skill selection interval hyperparameter $T _ { l o w }$ , we compare our method trained with $T _ { l o w } = \{ 1 , 2 , 3 , 5 , 1 0 \}$ on Jaco environments. The success rate curves in Figure 7 demonstrate that smaller $T _ { l o w }$ values in range [1, 3] lead to better performance. This can be because the agent can realize more flexible skill coordination by adjusting the behavior embedding frequently. + +In addition to the fixed $T _ { l o w }$ values, we also consider the variation of our method in which the skill behavior embedding is only sampled when the meta policy updates its skill selection. Concretely, we set the value of $T _ { l o w }$ to 1 but update $( z _ { t } ^ { 1 } , \ldots , z _ { t } ^ { N } )$ only if $\mathbf { \bar { \Phi } } ( c _ { t } ^ { 1 } , \ldots , c _ { t } ^ { N } ) \neq ( c _ { t - 1 } ^ { 1 } , \ldots , c _ { t - 1 } ^ { N } )$ . We observe that in this setting, the meta policy at times switch back and forth between two skills in two consecutive time steps, leading to slightly worse performance compared to our method with small $T _ { l o w }$ values. This indicates that the meta policy needs to adjust the behavior embedding in order to optimally coordinate skills of the different agents. + +# 5 CONCLUSION + +In this paper, we propose a modular framework with skill coordination to tackle challenges of composition of sub-skills with multiple agents. Specifically, we use entropy maximization with mutual information maximization to train controllable primitive skills with diverse behaviors. To coordinate learned primitive skills, the meta policy predicts not only the skill to execute for each agent (end-effector) but also the behavior embedding that controls the chosen primitive skill’s behavior. The experimental results on robotic manipulation and locomotion tasks demonstrate that the proposed framework is able to efficiently learn primitive skills with diverse behaviors and coordinate multiple agents (end-effectors) to solve challenging cooperative control tasks. Acquiring skills without supervision and extending our method to a visual domain are exciting directions for future work. + +# ACKNOWLEDGMENTS + +This project was funded by SKT. The authors would like to thank Karl Pertsch and many members of the USC CLVR lab for helpful discussion. + +# REFERENCES + +Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with policy sketches. In International Conference on Machine Learning, pp. 166–175, 2017. + +Pierre-Luc Bacon, Jean Harb, and Doina Precup. The option-critic architecture. In Association for the Advancement of Artificial Intelligence, 2017. + +Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. + +Prafulla Dhariwal, Christopher Hesse, Oleg Klimov, Alex Nichol, Matthias Plappert, Alec Radford, John Schulman, Szymon Sidor, Yuhuai Wu, and Peter Zhokhov. Openai baselines. https: //github.com/openai/baselines, 2017. + +Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=SJx63jRqFm. + +Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. In Association for the Advancement of Artificial Intelligence, 2018. + +Kevin Frans, Jonathan Ho, Xi Chen, Pieter Abbeel, and John Schulman. META LEARNING SHARED HIERARCHIES. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ SyX0IeWAW. + +Dibya Ghosh, Avi Singh, Aravind Rajeswaran, Vikash Kumar, and Sergey Levine. Divide-andconquer reinforcement learning. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ rJwelMbR-. + +Anirudh Goyal, Shagun Sodhani, Jonathan Binas, Xue Bin Peng, Sergey Levine, and Yoshua Bengio. Reinforcement learning with competitive ensembles of information-constrained primitives. In International Conference on Learning Representations, 2020. URL https://openreview. net/forum?id $=$ ryxgJTEYDr. + +Jayesh K Gupta, Maxim Egorov, and Mykel Kochenderfer. Cooperative multi-agent control using deep reinforcement learning. In International Conference on Autonomous Agents and Multi-Agent Systems, pp. 66–83, 2017. + +Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with deep energy-based policies. In International Conference on Machine Learning, pp. 1352–1361, 2017. + +Tuomas Haarnoja, Vitchyr Pong, Aurick Zhou, Murtaza Dalal, Pieter Abbeel, and Sergey Levine. Composable deep reinforcement learning for robotic manipulation. In IEEE International Conference on Robotics and Automation, pp. 6244–6251, 2018a. + +Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International Conference on Machine Learning, pp. 1856–1865, 2018b. + +Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation. In Advances in Neural Information Processing Systems, pp. 7254–7264, 2018. + +Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in Neural Information Processing Systems, pp. 3675–3683, 2016. + +Youngwoon Lee, Shao-Hua Sun, Sriram Somasundaram, Edward Hu, and Joseph J. Lim. Composing complex skills by learning transition policies. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ rygrBhC5tQ. + +Andrew Levy, Robert Platt, and Kate Saenko. Hierarchical reinforcement learning with hindsight. In International Conference on Learning Representations, 2019. URL https://openreview. net/forum?id ${ . } = { }$ ryzECoAcY7. + +Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems, pp. 6379–6390, 2017. + +Pattie Maes and Rodney A Brooks. Learning to coordinate behaviors. In Association for the Advancement of Artificial Intelligence, volume 90, pp. 796–802, 1990. + +Josh Merel, Arun Ahuja, Vu Pham, Saran Tunyasuvunakool, Siqi Liu, Dhruva Tirumala, Nicolas Heess, and Greg Wayne. Hierarchical visuomotor control of humanoids. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id $=$ BJfYvo09Y7. + +Ofir Nachum, Shixiang Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical reinforcement learning. In Advances in Neural Information Processing Systems, pp. 3303–3313, 2018. + +Ofir Nachum, Michael Ahn, Hugo Ponte, Shixiang Gu, and Vikash Kumar. Multi-agent manipulation via locomotion using hierarchical sim2real. In Conference on Robot Learning, 2019. + +Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning. In International Conference on Machine Learning, pp. 2661–2670, 2017. + +Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. In Advances in Neural Information Processing Systems Autodiff Workshop, 2017. + +Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, and Jun Wang. Multiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning to play starcraft combat games. arXiv preprint arXiv:1703.10069, 2017. + +Xue Bin Peng, Michael Chang, Grace Zhang, Pieter Abbeel, and Sergey Levine. Mcp: Learning composable hierarchical control with multiplicative compositional policies. Advances in Neural Information Processing Systems, 2019. + +Ahmed H. Qureshi, Jacob J. Johnson, Yuzhe Qin, Taylor Henderson, Byron Boots, and Michael C. Yip. Composing task-agnostic policies with deep reinforcement learning. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id $=$ H1ezFREtwH. + +Martin Riedmiller, Roland Hafner, Thomas Lampe, Michael Neunert, Jonas Degrave, Tom Van de Wiele, Volodymyr Mnih, Nicolas Heess, and Jost Tobias Springenberg. Learning by playing - solving sparse reward tasks from scratch. In International Conference on Machine Learning, 2018. + +Jurgen Schmidhuber. ¨ Towards compositional learning with dynamic neural networks. Inst. fur¨ Informatik, 1990. + +John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. In International Conference on Learning Representations, 2016. + +John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. + +Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation. In Advances in Neural Information Processing Systems, pp. 2244–2252, 2016. + +Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl Tuyls, et al. Value-decomposition networks for cooperative multi-agent learning based on team reward. In International Conference on Autonomous Agents and Multi-Agent Systems, pp. 2085–2087, 2018. + +Richard S Sutton. Temporal credit assignment in reinforcement learning. PhD thesis, University of Massachusetts, 1984. + +Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181–211, 1999. + +Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033, 2012. + +# A ENVIRONMENT DETAILS + +The details of observation spaces, action spaces, number of agents, and episode lengths are described in Table 2. All units in this section are in meters unless otherwise specified. + +Table 2: Environment details + +
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
Observation Space8888100
- Robot observation626282
- Object observation262618
Action Space181816
Number of Agents222
Episode length150100200
+ +# A.1 ENVIRONMENT DESCRIPTIONS + +In both Jaco environments, the robot works on a table with size (1.6, 1.6) and top center position $( 0 , 0 , 0 . 8 2 )$ . The two Jaco arms are initialized at positions $( - 0 . 1 6 , - 0 . 1 6 , 1 . 2 )$ and $( \bar { - } 0 . 1 6 , 0 . \bar { 2 } 4 , 1 . 2 )$ Left arm and right arm objects are initialized around $( 0 . 3 , 0 . 2 , 0 . 8 6 )$ and ( $0 . 3 , - 0 . 2 , 0 . 8 6 )$ respectively in all primitive training and composite task training environments, with small random position and rotation perturbation. + +In the Jaco Pick-Push-Place task, the right jaco arm needs to pick up the object and place it into the container initialized at the other side of the table. Success is defined by contact between the object and the inner top side of the container. + +In the Jaco Bar-Moving task, the two Jaco arms need to together pick the long bar up by height of 0.7, move it towards the arms by distance of 0.15, and place it back on the table. Success is defined by (1) the bar being placed within 0.04 away from the desired destination both in height and in xy-position and (2) the bar having been picked 0.7 above the table. + +In the Ant Push task, the two ant agents need to push a big box together to the goal position. The box has a size of $8 . 0 \times 1 . 6 \times 1 . 6$ . The distance between ants and the box is $2 0 \mathrm { c m }$ and the distance between the box and the goal is $3 0 \mathrm { c m }$ . Initial positions have $1 \mathrm { c m }$ of randomness and the agent has a randomness of 0.01 in each joint. The task is considered as success when both the distances between left and right end of the box and the goal are within $5 \mathrm { c m }$ . + +# A.2 REWARD DESIGN + +For every task, we add a control penalty, $- 0 . 0 0 1 * \| a \| ^ { 2 }$ , to regularize the magnitude of actions where $a$ is a torque action performed by an agent. + +Jaco Pick: To help the agent learn to reach, pick, and hold the picked object, we provide dense reward to the agent defined by the weighted sum of pick reward, gripper-to-cube distance reward, cube position and quaternion stability reward, hold duration reward, success reward, and robot control reward. More concretely, + +$$ +\begin{array} { r l } & { R ( s ) = \lambda _ { p i c k } \cdot ( z _ { b o x } - z _ { i n i t } ) + \lambda _ { d i s t } \cdot \mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \lambda _ { p o s } \cdot \mathrm { d i s t } ( p _ { b o x } , p _ { i n i t } ) + } \\ & { \qquad \lambda _ { q u a t } \cdot \mathrm { a b s } ( \Delta _ { q u a t } ) + \lambda _ { h o l d } \cdot t _ { h o l d } + \lambda _ { s u c c e s s } \cdot \mathbf { 1 } _ { \mathrm { s u c c e s s } } + \lambda _ { c t r l } \| a \| ^ { 2 } , } \end{array} +$$ + +where $\lambda _ { p i c k } = 5 0 0 , \lambda _ { d i s t } = 1 0 0 , \lambda _ { p o s } = 1 0 0 0 , \lambda _ { q u a t } = 1 0 0 0 , \lambda _ { h o l d } = 1 0 , \lambda _ { s u c c e s s } = 1 0 0 , \lambda _ { c t r l } = 1 0 0 0 , \lambda _ { d i s t }$ $1 \times 1 0 ^ { - \hat { 4 } }$ . + +Jaco Place: Reward for place primitive is defined by the weighted sum of xy-distance reward, height reward (larger when cube close to floor), success reward, and robot control reward. + +$$ +\begin{array} { r l } & { R ( s ) = \lambda _ { x y } \cdot \mathrm { d i s t } _ { x y } ( p _ { b o x } , p _ { g o a l } ) + \lambda _ { z } \cdot | z _ { b o x } - z _ { g o a l } | + \lambda _ { s u c c e s s } \cdot \mathbf { 1 } _ { \mathrm { s u c c e s s } } + \lambda _ { c t r l } \| a \| ^ { 2 } , } \\ & { \mathrm { ~ e \lambda _ { x y } = 5 0 0 , } \lambda _ { z } = 5 0 0 , \lambda _ { s u c c e s s } = 5 0 0 , \lambda _ { c t r l } = 1 \times 1 0 ^ { - 4 } . } \end{array} +$$ + +Jaco Push: Reward for push primitive is defined by the weighted sum of gripper reaching reward, box-to-destination distance reward, quaternion stability reward, hold duration reward, success reward, and robot control reward. + +$$ +\begin{array} { r } { R ( s ) = \lambda _ { r e a c h i n g } \cdot \mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \lambda _ { p o s } \cdot \mathrm { d i s t } ( p _ { b o x } , p _ { d e s t } ) + \phantom { x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x } } \\ { \lambda _ { q u a t } \cdot \mathsf { a b s } ( \Delta _ { q u a t } ) + \lambda _ { h o l d } \cdot t _ { h o l d } + \lambda _ { s u c c e s s } \cdot \mathbf { 1 } _ { s u c c e s s } + \lambda _ { c t r l } \| a \| ^ { 2 } , \phantom { x x x x x x x x x x x x x x } } \end{array} +$$ + +$$ +\lambda _ { r e a c h i n g } = 1 0 0 , \lambda _ { p o s } = 5 0 0 , \lambda _ { q u a t } = 3 0 , \lambda _ { h o l d } = 1 0 , \lambda _ { s u c c e s s } = 1 0 0 0 , \lambda _ { c t r l } = 1 \times 1 0 ^ { - 4 } . +$$ + +Jaco Pick-Push-Place: Reward for Pick-Push-Place is defined by the weighted sum of gripper contact reward, per-stage reach/pick/push/place rewards, success reward, and control reward. We tune the reward carefully for all baselines. + +$$ +\begin{array} { r l } & { R ( s ) = \lambda _ { c o n t a c t } \cdot \big ( { \bf 1 } _ { \scriptscriptstyle { \mathrm { l e f f ~ g r i p p e r ~ r o u c h e s ~ c o n t a i n e r } } } + { \bf 1 } _ { \scriptscriptstyle { \mathrm { r i g h t ~ g r i p p e r ~ t o u c h e s ~ b o x } } } \big ) + } \\ & { \quad \quad \lambda _ { r e a c h } \cdot { \bf 1 } _ { \scriptscriptstyle { \mathrm { r e a c h } } } \cdot \big ( \mathrm { d i s t } ( p _ { l e f t . g r i p p e r } , p _ { c o n t a i n e r } ) + \mathrm { d i s t } ( p _ { r i g h t . g r i p p e r } , p _ { b o x } ) \big ) + } \\ & { \quad \quad \lambda _ { p i c k } \cdot { \bf 1 } _ { \scriptscriptstyle { \mathrm { p i c k } } } \cdot \mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + \lambda _ { p l a c e } \cdot { \bf 1 } _ { \scriptscriptstyle { \mathrm { p l a c e } } } \cdot \mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + } \\ & { \quad \quad \lambda _ { p u s h } \cdot \mathrm { d i s t } ( p _ { c o n t a i n e r } , p _ { c o n t a i n e r . t a r g e t } ) + \lambda _ { s u c c e s s } \cdot { \bf 1 } _ { \scriptscriptstyle { \mathrm { s u c c e s s } } } + \lambda _ { c t r l } \cdot \| a \| ^ { 2 } , } \end{array} +$$ + +where $\lambda _ { r e a c h } = 1 0 , \lambda _ { c o n t a c t } = 1 0 , \lambda _ { p i c k } = \lambda _ { p l a c e } = \lambda _ { p l a c e } = 1 0 , \lambda _ { s u c c e s s } = 5 0 , \lambda _ { c t r l } = 0 ,$ , and 1reaching, $\mathbf { 1 } _ { \mathrm { p i c k } }$ , and $\mathbf { 1 } _ { \mathrm { p l a c e } }$ are indicator functions specifying whether the agent is in reaching, pick or place stage. Agent stages are determined by how many multiples of 25 steps the agent has stepped through in the environment. + +Jaco Bar-Moving: Reward for Bar-Moving is defined by the weighted sum of per-stage reach/pick/move/place rewards, success reward, and control reward. + +$$ +\begin{array} { r l } & { R ( s ) = \lambda _ { r e a c h } \cdot { \bf 1 } _ { r e a c h } \cdot \big ( \mathrm { d i s t } ( p _ { l e f t \_ g r i p p e r } , p _ { l e f t \_ h a n d l e } ) + \mathrm { d i s t } ( p _ { r i g h t \_ g r i p p e r } , p _ { r i g h t \_ h a n d l e } ) \big ) + } \\ & { \qquad \lambda _ { p i c k } \cdot { \bf 1 } _ { \mathrm { p i c k } } \cdot \mathrm { d i s t } ( p _ { b a r } , p _ { b a r \_ t a r g e t } ) + \lambda _ { m o v e } \cdot { \bf 1 } _ { \mathrm { p l a e } } \cdot \mathrm { d i s t } _ { x y } ( p _ { b a r \_ p b a r \_ t a r g e t } ) + } \\ & { \qquad \lambda _ { p l a c e } \cdot { \bf 1 } _ { \mathrm { p l a e } } \cdot \mathrm { d i s t } _ { z } ( p _ { b a r \_ p b a r \_ t a r g e t } ) + \lambda _ { s u c e s s } \cdot { \bf 1 } _ { \mathrm { s u c e s s } } + \lambda _ { c t r l } \cdot \| a \| ^ { 2 } , } \end{array} +$$ + +where $\lambda _ { r e a c h } = 1 0 , \lambda _ { p i c k } = 3 0 , \lambda _ { m o v e } = 1 0 0 , \lambda _ { p l a c e } = 1 0 0 , \lambda _ { s u c c e s s } = 1 0 0 , \lambda _ { c t r l } = 1 \times 1 0 ^ { - 4 }$ , and $\mathbf { 1 } _ { \mathrm { p i c k } }$ and $\mathbf { 1 } _ { \mathrm { p l a c e } }$ are indicator functions specifying whether the agent is in pick or place stage. Agent stages are determined by whether the pick objective is fulfilled or not. + +Ant Push & Ant Moving: Reward for ANT PUSH is defined by upright, velocity towards the desired direction. We provide a dense reward to encourage the desired locomotion behavior using velocity, stability, and posture, as following: + +$$ +\begin{array} { r l } & { R ( s ) = \lambda _ { v e l } \cdot \mathbf { a b s } ( \Delta x _ { a n t } ) + \lambda _ { b o x v e l } \cdot \mathbf { a b s } ( \Delta x _ { b o x } ) + \lambda _ { u p r i g h t } \cdot \mathbf { c o s } ( \theta ) - \lambda _ { h e i g h t } \cdot \mathbf { a b s } ( 0 . 6 - h ) + } \\ & { ~ \lambda _ { g o a l } \cdot \mathbf { d i s t } ( p _ { g o a l } , p _ { b o x } ) , } \end{array} +$$ + +where $\lambda _ { v e l } = 5 0 , \lambda _ { b o x v e l } = 2 0 , \lambda _ { u p r i g h t } = 1 , \lambda _ { h e i g h t } = 0 . 5$ . For ANT PUSH, we provide an additional reward based on distance between the box and the goal position with $\lambda _ { g o a l } = 2 0 0$ . + +# B EXPERIMENT DETAILS + +We use PyTorch (Paszke et al., 2017) for our implementation and all experiments are conducted on a workstation with Intel Xeon Gold 6154 CPU and 4 NVIDIA GeForce RTX 2080 Ti GPUs. + +# B.1 HYPERPARAMETERS + +Table 3: Hyperparameters + +
ParametersValue
learning rate3e-4
gradient steps50
batch size256
discount factor0.99
target smoothing coefficient0.005
reward scale (SAC)1.0
experience buffer size (# episodes)1000
Tow1 for JACO, 5 for ANT
Nz (dimensionality of z)5
+ +# B.2 NETWORK ARCHITECTURES + +Actor Networks: In all experiments, we model our actor network for each primitive skill as a 3-layer MLP with hidden layer size 64. The last layer of the MLP is two-headed – one for the mean of the action distribution and the other for the standard deviation of it. We use ReLU as activation function in hidden layers. We do not apply any activation function for the final output layer. The action distribution output represents per-dimension normal distribution, from which single actions can be sampled and executed in the environment. + +Critic Networks: The critic network for each primitive skill and meta policy is modeled as a 2-layer MLP with hidden layer size 128. ReLU is used as an activation function in the hidden layers. The critic network output is used to assess the value of a given state-action pair, and is trained by fitting its outputs to the target Q-value clamped by $\pm 1 0 0$ . + +Meta Policy: The meta policy is modeled as a 3-layer MLP with hidden layer size of 64. Since the actions of meta policy are sampled from $N$ categorical distributions for each end-effector/agent and $N$ normal distributions for behavior embeddings, the output dimension of the meta policy is $\begin{array} { r } { \sum _ { i = 1 } ^ { N } ( m ^ { i } + N _ { z } ) } \end{array}$ . The meta policy uses ReLU as an activation function for all layers except for the final output layer. + +# B.3 TRAINING DETAILS + +For all baselines, we train the meta policies using PPO and the low-level policies using SAC. We use the same environment configurations, composite task reward definitions, and value of $T _ { l o w }$ across all baselines. + +For Jaco tasks, we train a total of 4 primitive skills – right arm pick, right arm place-to-center, right arm place-towards-arm, and left arm push – to be composed by meta-policy. For Jaco Pick-PushPlace, we provide the meta-policy with right arm pick and right arm place-to-center as right arm primitives and left arm push as left arm primitives; for Jaco Bar-Moving, we provide the meta-policy with right arm pick and right arm place-towards-arm as both right and left arm primitives and left arm pick and right arm place-towards-arm as left arm primitives. We obtain left arm primitives for bar-moving task by using the learned right arm primitives directly. + +To obtain the 4 primitives skills described above, we train right arm pick with diversity coefficient $\lambda _ { 2 } = 0 . 0 1$ and the other three primitives with $\lambda _ { 2 } = 0 . 1$ . The destination of right arm push is set to $( 0 . 3 , - 0 . 0 3 , 0 . 8 6 )$ , which is slightly left of the center of the table. After pick primitive is trained, we train the two right arm place primitives where episodes are initialized by intermediate states of successful right arm pick episodes where the height of the box is larger than 0.94 (0.01 higher than the target pick height). The place destinations for towards-arm and to-center primitives are $( 0 . 1 5 , - 0 . 2 , 0 . 8 6 )$ and $( 0 . 3 , - 0 . 0 2 , 0 . 8 6 )$ , respectively. + +For non-modular baselines that incorporates skill behavior diversification, we use $\lambda _ { 2 } = 0 . 0 1$ for both Jaco Pick-Push-Place and Jaco Bar-Moving because both tasks require picking skills, which can only be trained with a small value of $\lambda _ { 2 }$ . \ No newline at end of file diff --git a/parse/train/ryxB2lBtvH/ryxB2lBtvH_content_list.json b/parse/train/ryxB2lBtvH/ryxB2lBtvH_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dd001300af29bc92ddc6acc8f0cc234ae9613f1a --- /dev/null +++ b/parse/train/ryxB2lBtvH/ryxB2lBtvH_content_list.json @@ -0,0 +1,1868 @@ +[ + { + "type": "text", + "text": "LEARNING TO COORDINATE MANIPULATION SKILLS VIA SKILL BEHAVIOR DIVERSIFICATION ", + "text_level": 1, + "bbox": [ + 174, + 98, + 805, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngwoon Lee, Jingyun Yang, Joseph J. Lim ", + "bbox": [ + 186, + 170, + 506, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science University of Southern California {lee504,jingyuny,limjj}@usc.edu ", + "bbox": [ + 184, + 185, + 485, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 262, + 544, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When mastering a complex manipulation task, humans often decompose the task into sub-skills of their body parts, practice the sub-skills independently, and then execute the sub-skills together. Similarly, a robot with multiple end-effectors can perform complex tasks by coordinating sub-skills of each end-effector. To realize temporal and behavioral coordination of skills, we propose a modular framework that first individually trains sub-skills of each end-effector with skill behavior diversification, and then learns to coordinate end-effectors using diverse behaviors of the skills. We demonstrate that our proposed framework is able to efficiently coordinate skills to solve challenging collaborative control tasks such as picking up a long bar, placing a block inside a container while pushing the container with two robot arms, and pushing a box with two ant agents. Videos and code are available at https://clvrai.com/coordination. ", + "bbox": [ + 233, + 292, + 764, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 178, + 482, + 336, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imagine you wish to play Chopin’s Fantaisie Impromptu on the piano. With little prior knowledge about the piece, you would first practice playing the piece with each hand separately. After independently mastering the left and right hand parts, you would move on to practicing with both hands simultaneously. To find the synchronized and non-interfering movements of two hands, you would try variable ways of playing the same melody with each hand, and eventually create a complete piece of music. Through the decomposition of skills into sub-skills of two hands and learning variations of sub-skills, humans make the learning process of manipulation skills much faster than learning everything at once. ", + "bbox": [ + 174, + 512, + 825, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can autonomous agents efficiently learn complicated tasks with coordination of different skills from multiple end-effectors like humans? Learning to perform collaborative and composite tasks from scratch requires a huge amount of environment interaction and extensive reward engineering, which often results in undesired behaviors (Riedmiller et al., 2018). Hence, instead of learning a task at once, modular approaches (Andreas et al., 2017; Oh et al., 2017; Frans et al., 2018; Lee et al., 2019; Peng et al., 2019; Goyal et al., 2020) suggest to learn reusable primitive skills and solve more complex tasks by recombining the skills. However, all these approaches either focus on working with single end-effector manipulation or single agent locomotion, and these do not scale to multi-agent problems. ", + "bbox": [ + 174, + 631, + 825, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we propose a modular framework that learns to coordinate multiple end-effectors with their primitive skills for various robotics tasks, such as bimanual manipulation. The main challenge is that naive simultaneous execution of primitive skills from multiple end-effectors can often cause unintended behaviors (e.g. collisions between end-effectors). Thus, as illustrated in Figure 1, our model needs to learn to appropriately coordinate end-effectors; and hence needs a way to obtain, represent, and control detailed behaviors of each primitive skill. Inspired by these intuitions, our method consists of two parts: (1) acquiring primitive skills with diverse behaviors by mutual information maximization, and (2) learning a meta policy that selects a skill for each end-effector and coordinates the chosen skills by controlling the behavior of each skill. ", + "bbox": [ + 174, + 750, + 825, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main contribution of this paper is a modular and hierarchical approach that tackles cooperative manipulation tasks with multiple end-effectors by (1) learning primitive skills of each end-effector independently with skill behavior diversification and (2) coordinating end-effectors using diverse behaviors of the skills. Our empirical results indicate that our proposed method is able to efficiently learn primitive skills with diverse behaviors and coordinate these skills to solve challenging collaborative control tasks such as picking up a long bar, placing a block inside the container on the right side, and pushing a box with two ant agents. We provide additional qualitative results and code at https://clvrai.com/coordination. ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e1b9f5f33a08c620e270f4735ec881e08b70e6cf015791f9524143e7c2eda6a.jpg", + "image_caption": [ + "Figure 1: Composing complex skills using multiple agents’ primitive skills requires proper coordination between agents since concurrent execution of primitive skills requires temporal and behavioral coordination. For example, to move a block into a container on the other end of the table, the agent needs to not only utilize pick, place, and push primitive skills at the right time but also select the appropriate behaviors for these skills, represented as latent vectors $z _ { 1 }$ , $z _ { 2 }$ , $z _ { 3 }$ , and $z _ { 4 }$ above. Naive methods neglecting either temporal or behavioral coordination will produce unintended behaviors, such as collisions between end-effectors. " + ], + "image_footnote": [], + "bbox": [ + 176, + 102, + 821, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 329, + 825, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 420, + 341, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep reinforcement learning (RL) for continuous control is an active research area. However, learning a complex task either from a sparse reward or a heavily engineered reward becomes computationally impractical as the target task becomes complicated. Instead of learning from scratch, complex tasks can be tackled by decomposing the tasks into easier and reusable sub-tasks. Hierarchical reinforcement learning temporally splits a task into a sequence of temporally extended meta actions. It often consists of one meta policy (high-level policy) and a set of low-level policies, such as options framework (Sutton et al., 1999). The meta policy decides which low-level policy to activate and the chosen low-level policy generates an action sequence until the meta policy switches it to another lowlevel policy. Options can be discovered without supervision (Schmidhuber, 1990; Bacon et al., 2017; Nachum et al., 2018; Levy et al., 2019), meta-learned (Frans et al., 2018), pre-defined (Kulkarni et al., 2016; Oh et al., 2017; Merel et al., 2019; Lee et al., 2019), or attained from additional supervision signals (Andreas et al., 2017; Ghosh et al., 2018). However, option frameworks are not flexible to solve a task that requires simultaneous activation or interpolation of multiple skills since only one skill can be activated at each time step. ", + "bbox": [ + 174, + 450, + 825, + 646 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve composite tasks multiple policies can be simultaneously activated by adding Qfunctions (Haarnoja et al., 2018a), additive composition (Qureshi et al., 2020; Goyal et al., 2020), or multiplicative composition (Peng et al., 2019). As each policy takes the whole observation as input and controls the whole agent, it is not robust to changes in unrelated parts of the observation. For example, a left arm skill can be affected by the pose change in the right arm, which is not relevant to the left arm skill. Hence, these skill composition approaches can fail when an agent encounters a new combination of skills or a new skill is introduced since the agent will experience unseen observations. ", + "bbox": [ + 174, + 652, + 825, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of having a policy with the full observation and action space, multi-agent reinforcement learning (MARL) suggests to explicitly split the observation and action space according to agents (e.g. robots or end-effectors), which allows efficient low-level policy training as well as flexible skill composition. For cooperative tasks, communication mechanisms (Sukhbaatar et al., 2016; Peng et al., 2017; Jiang & Lu, 2018), sharing policy parameters (Gupta et al., 2017), and decentralized actor with centralized critic (Lowe et al., 2017; Foerster et al., 2018) have been actively used. However, these approaches suffer from the credit assignment problem (Sutton, 1984) among agents and the lazy agent problem (Sunehag et al., 2018). As agents have more complicated morphology and larger observation space, learning a policy for a multi-agent system from scratch requires extremely long training time. Moreover, the credit assignment problem becomes more challenging when the complexity of cooperative tasks increases and all agents need to learn completely from scratch. To resolve these issues, we propose to first train reusable skills for each agent in isolation, instead of learning primitive skills of multiple agents together. Then, we recombine these skills (Maes & Brooks, 1990) to complete more complicated tasks with learned coordination of the skills. ", + "bbox": [ + 174, + 757, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63021c65e6e95a999eba964c17a1759f617f5364c3d76c2f449409b6d5dca506.jpg", + "image_caption": [ + "Figure 2: Our method is composed of two components: a meta policy and a set of agent-specific primitive policies relevant to task completion. The meta policy selects which primitive skill to run for each agent as well as the behavior embedding (i.e. variation in behavior) of the chosen primitive skill. Each selected primitive skill takes as input the agent observation and the behavior embedding and outputs action for that agent. " + ], + "image_footnote": [], + "bbox": [ + 173, + 102, + 823, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 362, + 823, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To coordinate skills from multiple agents, the skills have to be flexible; hence, a skill can be adjusted to collaborate with other agents’ skills. Maximum entropy policies (Haarnoja et al., 2017; 2018a;b) can learn diverse ways to achieve a goal by maximizing not only reward but also entropy of the policy. In addition, Eysenbach et al. (2019) proposes to discover diverse skills without reward by maximizing entropy as well as mutual information between resulting states and latent representations of skills (i.e. skill embeddings). Our method leverages the maximum entropy policy (Haarnoja et al., 2018b) with the discriminability objective (Eysenbach et al., 2019) to learn a primitive skill with diverse behaviors conditioned on a controllable skill embedding. This controllable skill embedding will be later used as a behavior embedding for the meta policy to adjust a primitive skill’s behavior for coordination. ", + "bbox": [ + 173, + 397, + 826, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 METHOD ", + "text_level": 1, + "bbox": [ + 176, + 545, + 281, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we address the problem of solving cooperative manipulation tasks that require collaboration between multiple end-effectors or agents. Note that we use the terms “end-effector” and “agent” interchangeably in this paper. Instead of learning a multi-agent task from scratch (Lowe et al., 2017; Gupta et al., 2017; Sunehag et al., 2018; Foerster et al., 2018), modular approaches (Andreas et al., 2017; Frans et al., 2018; Peng et al., 2019) suggest to learn reusable primitive skills and solve more complex tasks by recombining these skills. However, concurrent execution of primitive skills of multiple agents fails when agents never experienced a combination of skills during the pre-training stage, or skills require temporal or behavioral coordination. ", + "bbox": [ + 173, + 578, + 825, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, we propose a modular and hierarchical framework that learns to coordinate multiple agents with primitive skills to perform a complex task. Moreover, during primitive skill training, we propose to learn a latent behavior embedding, which provides controllability of each primitive skill to the meta policy while coordinating skills. In Section 3.2, we describe our modular framework in detail. Next, in Section 3.3, we elaborate how controllable primitive skills can be acquired. Lastly, we describe how the meta policy learns to coordinate primitive skills in Section 3.4. ", + "bbox": [ + 174, + 695, + 825, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 174, + 799, + 326, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate our problem as a Markov decision process defined by a tuple $\\{ S , { \\mathcal { A } } , { \\mathcal { T } } , R , \\rho , \\gamma \\}$ of states, actions, transition probability, reward, initial state distribution, and discount factor. In our formulation, we assume the environment includes $N$ agents. To promote consistency in our terminology, we use superscripts to denote the index of agent and subscripts to denote time or primitive skill index. Hence, the state space and action space for an agent $i$ can be represented as $S ^ { \\imath }$ and $A ^ { i }$ where each element of $S ^ { i }$ is a subset of the corresponding element in $\\boldsymbol { S }$ and $\\mathcal { A } = \\mathcal { A } ^ { 1 } \\times \\mathcal { A } ^ { 2 } \\times \\cdot \\cdot \\cdot \\times \\mathcal { A } ^ { N }$ , respectively. For each agent $i$ , we provide a set of $m ^ { i }$ skills, $\\Pi ^ { i } = \\{ \\pi _ { 1 } ^ { i } , \\ldots , \\pi _ { m ^ { i } } ^ { i } \\}$ . A policy of an agent $i$ is represented as $\\pi _ { c _ { t } ^ { i } } ^ { i } \\left( a _ { t } ^ { i } | s _ { t } ^ { i } \\right) \\in \\Pi ^ { i }$ where is sam $c _ { t } ^ { i }$ is a skiled from index, , and t $s _ { t } ^ { i } \\in S ^ { i }$ is a state, and agents take act $a _ { t } ^ { i } \\in \\mathcal A ^ { i }$ action at time sampled from $t$ . composite policy $\\pi ( a _ { t } ^ { 1 } , \\bar { a } _ { t } ^ { 2 } , \\dots , a _ { t } ^ { N } | s _ { t } , c _ { t } ^ { 1 } , c _ { t } ^ { 2 } , \\dots , \\bar { c } _ { t } ^ { N } ) = ( \\pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( a _ { t } ^ { 1 } | s _ { t } ) , \\pi _ { c _ { t } ^ { 2 } } ^ { 2 } ( a _ { t } ^ { 2 } | s _ { t } ) , \\dots , \\pi _ { c _ { t } ^ { N } } ^ { N } ( a _ { t } ^ { N } | s _ { t } ) )$ $\\rho$ $N$ $a _ { t } ^ { 1 } , a _ { t } ^ { 2 } , \\ldots , a _ { t } ^ { N }$ a singl, where reward is the e $r _ { t }$ . The performance is evaluated based on a discounted return sode horizon. $R =$ $\\textstyle \\sum _ { t = 0 } ^ { T - 1 } \\gamma ^ { t } r _ { t }$ $T$ ", + "bbox": [ + 174, + 825, + 825, + 925 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/852e7eff92afaaa227e4f7f1a58fdeda7147bb405d07ee79a206747f65017adf.jpg", + "image_caption": [ + "Figure 3: Different multi-agent architectures. (a) The vanilla RL method considers all agents as a monolithic agent; thus a single policy takes the full observation as input and outputs the full action. (b) The multi-agent RL method (MARL) consists of $N$ policies that operate on the observations and actions of corresponding agents. (c) The modular network consists of $N$ sets of skills for the $N$ agents trained in isolation and a meta policy that selects a skill for each agent. (d-f) The RL, MARL, and modular network methods augmented with skill behavior diversification (SBD) has a meta policy that outputs a skill behavior embedding vector $z$ for each skill. " + ], + "image_footnote": [], + "bbox": [ + 183, + 98, + 821, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 345, + 826, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 MODULAR FRAMEWORK ", + "text_level": 1, + "bbox": [ + 174, + 452, + 385, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, our model is composed of two components: a meta policy $\\pi _ { m e t a }$ and a set of primitive skills of $N$ agents $\\Pi ^ { 1 } , \\dots , \\Pi ^ { N }$ . Note that each primitive skill $\\pi _ { c ^ { i } } ^ { i } \\in \\Pi ^ { i }$ contains variants of behaviors parameterized by an $N _ { z }$ -dimensional latent behavior embedding $z ^ { i }$ (see Section 3.3). The meta policy selects a skill to execute for each agent, rather than selecting one primitive skill for the entire multi-agent system to execute. Also, we give the meta policy the capability to select which variant of the skill to execute (see Section 3.4). Then, the chosen primitive skills are simultaneously executed for $T _ { l o w }$ time steps. ", + "bbox": [ + 173, + 479, + 825, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The concurrent execution of multiple skills often leads to undesired results and therefore requires coordination between the skills. For example, naively placing a block in the left hand to a container being moved by the right hand can cause collision between the two robot arms. The arms can avoid collision while performing the skills by properly adjusting their skill behaviors (e.g. the left arm leaning to the left side while placing the block and the right arm leaning to the right side while pushing the container) as shown in Figure 1. In our method, the meta policy learns to coordinate multiple agents’ skills by manipulating the behavior embeddings (i.e. selecting a proper behavior from diverse behaviors of each skill). ", + "bbox": [ + 174, + 585, + 825, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 TRAINING AGENT-SPECIFIC PRIMITIVE SKILLS WITH DIVERSE BEHAVIORS ", + "text_level": 1, + "bbox": [ + 176, + 722, + 723, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To adjust a primitive skill to collaborate with other agents’ skills in a new environment, the skill needs to support variations of skill behaviors when executed at a given state. Moreover, a behavioral variation of a skill should be controllable by the meta policy for skill coordination. In order to make our primitive skill policies generate diverse behaviors controlled by a latent vector $z$ , we leverage the entropy and mutual information maximization objective introduced in Eysenbach et al. (2019). ", + "bbox": [ + 174, + 750, + 825, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, a primitive policy of an agent $i$ outputs an action $a \\in { \\mathcal { A } }$ conditioned on the current state $s \\in S$ and a latent behavior embedding $z \\sim p ( z )$ , where the prior distribution $p ( z )$ is Gaussian (we omit agent $i$ in this section for the simplicity of notations). Diverse behaviors conditioned on a random sample $z$ can be achieved by maximizing the mutual information between behaviors and states $M I ( s , z )$ , while minimizing the mutual information between behaviors and actions given the state $M I ( a , z | s )$ , together with maximizing the entropy of the policy $\\mathcal { H } ( a | s )$ to encourage diverse behaviors. The objective can be written as follows (we refer the readers to Eysenbach et al. (2019) ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 ROLLOUT ", + "text_level": 1, + "bbox": [ + 174, + 106, + 331, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1: Input: Meta policy $\\pi _ { m e t a }$ , sets of primitive policies $\\Pi ^ { 1 } , . . . , \\Pi ^ { N }$ , and meta horizon $T _ { l o w }$ \n2: Initialize an episode $t \\gets 0$ and receive initial state $s _ { 0 }$ \n3: while episode is not terminated do \n4: Sample skill indexes and behavior embeddings $( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } ) \\sim \\pi _ { m e t a } ( s _ { t } )$ \n5: $\\tau 0$ \n6: while $\\tau < T _ { l o w }$ and episode is not terminated do \n7: $\\boldsymbol a _ { t + \\tau } = ( a _ { t + \\tau } ^ { 1 } , \\dots , \\dot { a _ { t + \\tau } ^ { N } } ) \\sim ( \\pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( s _ { t + \\tau } , z _ { t } ^ { 1 } ) , \\dots , \\pi _ { c _ { t } ^ { N } } ^ { N } ( s _ { t + \\tau } , z _ { t } ^ { N } ) )$ \n8: $\\begin{array} { r l } & { s _ { t + \\tau + 1 } , r _ { t + \\tau } \\gets \\mathrm { E N V } ( s _ { t + \\tau } , a _ { t + \\tau } ) } \\\\ & { \\tau \\gets \\tau + 1 } \\end{array}$ \n9: \n10: end while \n11: Add a transition $s _ { t } , ( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } ) , s _ { t + \\tau } , r _ { t : t + \\tau - 1 }$ to the rollout buffer $\\boldsymbol { B }$ \n12: $t \\gets t + \\tau$ \n13: end while ", + "bbox": [ + 178, + 125, + 795, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for derivation): ", + "bbox": [ + 174, + 339, + 274, + 353 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/d92048ef0214b9413b930951da396cad2d15a7c9b9c6cb3e6d96816918b24d97.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathcal { F } ( \\theta ) = M I ( s , z ) - M I ( a , z \\vert s ) + \\mathcal { H } ( a \\vert s ) = \\mathcal { H } ( a \\vert s , z ) - \\mathcal { H } ( z \\vert s ) + \\mathcal { H } ( z ) } \\\\ & { \\quad \\quad = \\mathcal { H } ( a \\vert s , z ) + \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log p ( z \\vert s ) ] - \\mathbb { E } _ { z \\sim p ( z ) } [ \\log p ( z ) ] } \\\\ & { \\quad \\quad \\geq \\mathcal { H } ( a \\vert s , z ) + \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log q _ { \\phi } ( z \\vert s ) - \\log p ( z ) ] , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 253, + 358, + 743, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the learned discriminator $q _ { \\phi } ( z | s )$ approximates the posterior $p ( z | s )$ . ", + "bbox": [ + 174, + 421, + 663, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve a primitive skill with diverse behaviors, we augment Equation (3) to the environment reward: ", + "bbox": [ + 171, + 443, + 821, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/7dcfa1f10a0ed872bfb88b0fa8018a0f0cffb519a0831b833fece71eb0731dbc.jpg", + "text": "$$\nr _ { t } + \\lambda _ { 1 } \\mathcal { H } ( a | s , z ) + \\lambda _ { 2 } \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log q _ { \\phi } ( z | s ) - \\log p ( z ) ] ,\n$$", + "text_format": "latex", + "bbox": [ + 294, + 470, + 702, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda _ { 1 }$ is the entropy coefficient and $\\lambda _ { 2 }$ is the diversity coefficient which corresponds identifiability of behaviors. Maximizing Equation (3) encourages multi-modal exploration strategies while maximizing the reward $r _ { t }$ forces to achieve its own goal. Moreover, by maximizing identifiability of behaviors, the latent vector $z$ , named behavior embedding, can represent a variation of the learned policy and thus can be used to control the behavior of the policy. For example, when training a robot to move an object, a policy learns to move the object quickly as well as slowly, and these diverse behaviors map to different latent vectors $z$ . We empirically show that the policies with diverse behaviors achieve better compositionality with other agents in our experiments. ", + "bbox": [ + 174, + 489, + 825, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 COMPOSING PRIMITIVE SKILLS WITH META POLICY ", + "text_level": 1, + "bbox": [ + 174, + 619, + 571, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote the meta policy as $\\pi _ { m e t a } ( c ^ { 1 } , \\ldots , c ^ { N } , z ^ { 1 } , \\ldots , z ^ { N } | s _ { t } )$ , where $c ^ { i } \\in [ 1 , m ^ { i } ]$ represents a skill index of an agent $i \\in [ 1 , N ]$ and $z ^ { i } \\in \\mathbb { R } ^ { N _ { z } }$ represents a behavior embedding of the skill. Every $T _ { l o w }$ time steps, the meta policy chooses one primitive skill $\\pi _ { c ^ { i } } ^ { i } \\in \\Pi ^ { i }$ for each agent $i$ . Also, the meta policy outputs a set of latent behavior embeddings $( z ^ { 1 } , z ^ { 2 } , \\dots , z ^ { N } )$ and feeds them to the corresponding skills (i.e. $\\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )$ for agent $i$ ). Once a set of primitive skills $\\{ \\pi _ { c ^ { 1 } } ^ { 1 } , \\ldots , \\pi _ { c ^ { N } } ^ { N } \\}$ are chosen to be executed, each primitive skill generates an action $a ^ { i } \\sim \\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )$ based on the current state $s ^ { i }$ and the latent vector $z ^ { i }$ . Algorithm 1 illustrates the overall rollout process. ", + "bbox": [ + 173, + 642, + 825, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since there are a finite number of skills for each agent to execute, the meta action space for each agent $[ 1 , m ^ { i } ]$ is discrete, while the behavior embedding space for each agent $\\mathbb { R } ^ { N _ { z } }$ is continuous. Thus, the meta policy is modeled as a $( 2 \\times N )$ -head neural network where the first $N$ heads represent $m ^ { i }$ -way categorical distributions for skill selection and the last $N$ heads represent $N _ { z }$ -dimensional Gaussian distributions for behavior control of the chosen skill. ", + "bbox": [ + 174, + 755, + 825, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5 IMPLEMENTATION ", + "text_level": 1, + "bbox": [ + 176, + 842, + 341, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We model the primitive policies and posterior distributions $q _ { \\phi }$ as neural networks. We train the primitive policies using soft actor-critic (Haarnoja et al., 2018b). When we train a primitive policy, we use a unit Gaussian distribution as the prior distribution of latent variables $p ( z )$ . We use 5 as the size of latent behavior embedding $N _ { z }$ . Each primitive policy outputs the mean and standard deviation of a Gaussian distribution over an action space. For a primitive policy, we apply tanh activation to normalize the action between $[ - 1 , 1 ]$ . We model the meta policy as neural network with multiple heads that output the skill index $\\bar { c } ^ { i }$ and behavior embedding $z ^ { i }$ for each agent. The meta policy is trained using PPO (Schulman et al., 2017; 2016; Dhariwal et al., 2017). All policy networks in this paper consist of 3 fully connected layers of 64 hidden units with ReLU nonlinearities. The discriminator $q _ { \\phi }$ in Equation (4) is a 2-layer fully connected network with 64 hidden units. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/11edbefb96b52e765ac0cdb4e36075acafa79155b0f2fabe7fb7cdafaf5487ff.jpg", + "image_caption": [ + "Figure 4: The composite tasks pose a challenging combination of object manipulation and locomotion skills, which requires coordination of multiple agents and temporally extended behaviors. (a) The left Jaco arm needs to pick up a block while the right Jaco arm pushes a container, and then it places the block into the container. (b) Two Jaco arms are required to pick and place a bar-shaped block together. (c) Two ants push the red box to the goal location (green circle) together. " + ], + "image_footnote": [], + "bbox": [ + 205, + 99, + 794, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 377, + 825, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 489, + 326, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our framework, we compare our method to prior methods in the field of multi-agent RL and ablate the components of our framework to understand their importance. We conducted experiments on a set of challenging robot control environments that require coordination of different agents to complete collaborative robotic manipulation and locomotion tasks. ", + "bbox": [ + 174, + 526, + 825, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through our experiments, we aim to answer the following questions: (1) can our framework efficiently learn to combine primitive skills to execute a complicated task; (2) can our learned agent exhibit collaborative behaviors during task execution; and (3) can our framework leverage the controllable behavior variations of the primitive skills to achieve better coordination? ", + "bbox": [ + 174, + 589, + 825, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For details about environments and training, please refer to the supplementary material. As the performance of training algorithms varies between runs, we train each method on each task with 6 different random seeds and report mean and standard deviation of each method’s success rate. ", + "bbox": [ + 174, + 651, + 823, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 BASELINES", + "text_level": 1, + "bbox": [ + 174, + 718, + 294, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the performance of our method with various single- and multi-agent RL methods illustrated in Figure 3: ", + "bbox": [ + 176, + 747, + 823, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single-agent RL (RL): A vanilla RL method where a single policy takes as input the full observation and outputs all agents’ actions. ", + "bbox": [ + 174, + 784, + 821, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-agent RL (MARL): A multi-agent RL method where each of $N$ policies takes as input the observation of the corresponding agent and outputs an action for that agent. All policies share the global critic learned from a single task reward (Lowe et al., 2017). ", + "bbox": [ + 174, + 818, + 825, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Modular Framework (Modular): A modular framework composed of a meta policy and $N$ sets of primitive skills (i.e. one or more primitive skills per agent). Every $T _ { l o w }$ time steps, the meta policy selects a primitive skill for each agent based on the full observation. Then, the chosen skills are executed for $T _ { l o w }$ time steps. ", + "bbox": [ + 176, + 867, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38d7a3b3c6d72759e19da2cea6bad148d9e0ead37b5a570252e66d5ccb027d90.jpg", + "image_caption": [ + "Figure 5: Success rates of our method (Modular-SBD) and baselines. For modular frameworks (Modular and Modular-SBD), we shift the learning curves rightwards the total number of environment steps the agent takes to learn the primitive skills (0.9 M, $1 . 2 \\mathbf { M }$ , and $2 . 0 \\mathbf { M }$ , respectively). Our method substantially improves learning speed and performance on JACO PICK-PUSH-PLACE and ANT PUSH. The shaded areas represent the standard deviation of results from six different seeds. The curves are smoothed using moving average over 10 runs. " + ], + "image_footnote": [], + "bbox": [ + 187, + 103, + 810, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single-agent RL with Skill Behavior Diversification (RL-SBD): An RL method augmented with the behavior diversification objective. A meta policy is employed to generate a behavior embedding for a low-level policy, and the low-level policy outputs all agents’ actions conditioned on the behavior embedding and the full observation for $T _ { l o w }$ time steps. The meta policy and the low-level policy are jointly trained with the behavior diversification objective described in Equation (4). ", + "bbox": [ + 174, + 375, + 825, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-agent RL with Skill Behavior Diversification (MARL-SBD): A MARL method augmented with the behavior diversification objective. A meta policy generates $N$ behavior embeddings. Then, each low-level policy outputs each agent’s action conditioned on its observation and behavior embedding for $T _ { l o w }$ time steps. All policies are jointly trained to maximize Equation (4). ", + "bbox": [ + 174, + 452, + 825, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modular Framework with Skill Behavior Diversification (Modular-SBD, Ours): Our method which coordinates primitive skills of multiple agents. The modular framework consists of a meta policy and $N$ sets of primitive skills, where each primitive skill is conditioned on a behavior embedding $z$ . The meta policy takes as input the full observation and selects both a primitive skill and a behavior embedding for each agent. Then, each primitive skill outputs action for each agent. ", + "bbox": [ + 174, + 513, + 825, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 JACO PICK-PUSH-PLACE ", + "text_level": 1, + "bbox": [ + 174, + 603, + 387, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We developed JACO PICK-PUSH-PLACE and JACO BAR-MOVING environments using two Kinova Jaco arms, where each Jaco arm is a 9 DoF robotic arm with 3 fingers. JACO PICK-PUSH-PLACE starts with a block on the left and a container on the right. The robotic arms need to pick up the block, push the container to the center, and place the block inside the container. For successful completion of the task, the two Jaco arms have to concurrently execute their distinct sets of skills and dynamically adjust their picking, pushing, and placing directions to avoid collision between arms. ", + "bbox": [ + 174, + 631, + 825, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Primitives skills. There are three primitive skills available to each arm: Picking up, Pushing, and Placing to center (see Figure 4a). Picking up requires a robotic arm to pick up a small block, which is randomly placed on the table. If the block is not picked up after a certain amount of time or the arm drops the block, the agent fails. Pushing learns to push a big container to its opposite side (e.g. from left to the center or from right to center). The agent fails if it cannot place the container to the center. Placing to center requires placing an object in the gripper to the table. The agent only succeeds when it stably places the object at the desired location on the container. ", + "bbox": [ + 174, + 722, + 825, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Composite task. Our method (Modular-SBD) can successfully perform JACO PICK-PUSH-PLACE task while all baselines fail to compose primitive skills as shown in Figure 5a. The RL and MARL baselines cannot learn the composite task mainly because the agent requires to learn the combinatorial number of skill compositions and to solve the credit assignment problem across multiple agents. Since the composite task requires multiple primitive skills of multiple agents to be performed properly at the same time, a reward signal about a failure case cannot be assigned to the correct agent or skill. By using pre-trained primitive skills, the credit assignment problem is relaxed and all agents can ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/81b3836f62ff4588cec048ce9a9d8153404ee36c49b3e4b3237d9bad87b5e2a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
RL0.000±0.0000.000±0.0000.000±0.000
MARL0.000±0.0000.000± 0.0000.000± 0.000
RL-SBD0.000± 0.0000.000± 0.0000.000±0.000
MARL-SBD0.000 ± 0.0000.000 ±0.0000.000 ± 0.000
Modular0.324± 0.4680.917± 0.2760.003± 0.058
Modular-SBD (Ours)0.902 ± 0.2980.950 ± 0.2180.323 ± 0.468
", + "bbox": [ + 259, + 101, + 738, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1: Success rates for all tasks, comparing our method against baselines. Each entry in the table represents average success rate and standard deviation over 100 runs. The baselines learning from scratch fail to learn complex tasks with multiple agents. ", + "bbox": [ + 176, + 205, + 823, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "perform their skills concurrently. Therefore, the Modular baseline learns to achieve success but shows significantly lower performance than our method (Modular-SBD). This is because the lack of skill behavior diversification makes it impossible to adjust pushing and placing trajectories during skill composition time, which resulting in frequent end-effector collisions. ", + "bbox": [ + 174, + 275, + 825, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 JACO BAR-MOVING ", + "text_level": 1, + "bbox": [ + 174, + 349, + 352, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In JACO BAR-MOVING, two Jaco arms need to pick up a long bar together, move the bar towards a target location while maintaining its rotation, and place it on the table (see Figure 4b). The initial position of the bar is randomly initialized every episode and an agent needs to find appropriate coordination between two arms for each initialization. Compared to JACO PICK-PUSH-PLACE, this task requires that the two arms synchronize their movements and perform more micro-level adjustments to their behaviors. ", + "bbox": [ + 174, + 376, + 825, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Primitives skills. There are two pre-trained primitive skills available to each arm: Picking up and Placing towards arm. Picking up is same as described in Section 4.2. Placing towards arm learns to move a small block (half size of the block used in the composite task) in the hand towards the robotic arm and then place it on the table. The agent fails if it cannot place the block to the target location. ", + "bbox": [ + 174, + 467, + 825, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Composite task. The JACO BAR-MOVING task requires the two arms to work very closely together. For example, the Picking up skill of both arms should be synchronized when they start to lift the bar and two arms require to lift the bar while maintaining the relative position between them since they are connected by holding the bar. The modular framework without explicit coordination of skills (Modular) can synchronize the execution of picking, moving, and placing. But the inability to micro-adjust the movement of the other arm causes instability of bar picking and moving. This results in degraded success rates compared to the modular framework with explicit coordination. Meanwhile, all baselines without pre-defined primitive skills fail to learn JACO BAR-MOVING. ", + "bbox": [ + 174, + 530, + 825, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 ANT PUSH ", + "text_level": 1, + "bbox": [ + 174, + 660, + 289, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We developed a multi-ant environment, ANT-PUSH, inspired from Nachum et al. (2019), simulated in the MuJoCo (Todorov et al., 2012) physics engine. We use the ant model in OpenAI Gym (Brockman et al., 2016). In this environment, two ants need to push a large object toward a green target place, collaborating with each other to keep the angle of the object as stable as possible (see Figure 4c). ", + "bbox": [ + 174, + 686, + 825, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Primitives skills. We train walking skills of an ant agent in 4 directions: up, down, left, and right. During primitive skill training, a block (half size of the block used in the composite task) and an ant agent are randomly placed. Pushing the block gives an additional reward to the agent, which prevents an ant to avoid the block. The learned primitive skills have different speed and trajectories conditioned on the latent behavior embedding. ", + "bbox": [ + 174, + 750, + 825, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Composite task. Our method achieves $3 2 . 3 \\%$ success rate on ANT PUSH task while all baselines fail to compose primitive skills as shown in Figure 5c and Table 1. The poor performance of RL, MARL, RL-SBD, and MARL-SBD baselines shows the difficulty of credit assignment between agents, which leads one of the ants moves toward a block and pushes it but another ant does not move. Moreover, the Modular baseline with primitive skills also fails to learn the pushing task. This result illustrates the importance of coordination of agents, which helps synchronizing and controlling the velocities of both ant agents to push the block toward the goal position while maintaining its rotation. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99d31a6b5e96ecb62faf4db19c419b513aea19dd229cabd50fa233595361425f.jpg", + "image_caption": [ + "Figure 6: Learning curves of our method with different diversity coefficients $\\lambda _ { 2 }$ on ANT PUSH. ", + "Figure 7: Success rates of our method with different $T _ { l o w }$ coefficients on Jaco environments. " + ], + "image_footnote": [], + "bbox": [ + 179, + 117, + 818, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5 EFFECT OF DIVERSITY OF PRIMITIVE SKILLS ", + "text_level": 1, + "bbox": [ + 178, + 304, + 521, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To analyze the effect of the diversity of primitive skills, we compare our model with primitive skills trained with different diversity coefficients $\\lambda _ { 2 } = \\{ 0 . 0 , 0 . 0 5 , 0 . 1 , 0 . 5 , 1 . 0 \\}$ in Equation (4) on ANT PUSH. Figure 6 shows that with small diversity coefficients $\\lambda _ { 2 } = \\{ 0 . 0 5 , 0 . 1 \\}$ , the agent can control detailed behaviors of primitive skills while primitive skills without diversity ${ \\lambda } _ { 2 } = 0$ ) cannot be coordinated. The meta policy tries to synchronize two ant agents’ positions and velocities by switching primitive skills, but it cannot achieve proper coordination without diversified skills. On the other hand, large diversity coefficients $\\lambda _ { 2 } = \\bar { \\{ 0 . 5 , 1 . 0 \\} }$ make the primitive skills often focus on demonstrating diverse behaviors and fail to achieve the goals of the skills. Hence, these primitive skills do not have enough functionality to solve the target task. The diversity coefficient needs to be carefully chosen to acquire primitive skills with good performance as well as diverse behaviors. ", + "bbox": [ + 174, + 329, + 825, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6 EFFECT OF SKILL SELECTION INTERVAL $T _ { l o w }$ ", + "text_level": 1, + "bbox": [ + 174, + 487, + 524, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To analyze the effect of the skill selection interval hyperparameter $T _ { l o w }$ , we compare our method trained with $T _ { l o w } = \\{ 1 , 2 , 3 , 5 , 1 0 \\}$ on Jaco environments. The success rate curves in Figure 7 demonstrate that smaller $T _ { l o w }$ values in range [1, 3] lead to better performance. This can be because the agent can realize more flexible skill coordination by adjusting the behavior embedding frequently. ", + "bbox": [ + 174, + 513, + 825, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In addition to the fixed $T _ { l o w }$ values, we also consider the variation of our method in which the skill behavior embedding is only sampled when the meta policy updates its skill selection. Concretely, we set the value of $T _ { l o w }$ to 1 but update $( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } )$ only if $\\mathbf { \\bar { \\Phi } } ( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) \\neq ( c _ { t - 1 } ^ { 1 } , \\ldots , c _ { t - 1 } ^ { N } )$ . We observe that in this setting, the meta policy at times switch back and forth between two skills in two consecutive time steps, leading to slightly worse performance compared to our method with small $T _ { l o w }$ values. This indicates that the meta policy needs to adjust the behavior embedding in order to optimally coordinate skills of the different agents. ", + "bbox": [ + 174, + 575, + 825, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 695, + 318, + 712 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a modular framework with skill coordination to tackle challenges of composition of sub-skills with multiple agents. Specifically, we use entropy maximization with mutual information maximization to train controllable primitive skills with diverse behaviors. To coordinate learned primitive skills, the meta policy predicts not only the skill to execute for each agent (end-effector) but also the behavior embedding that controls the chosen primitive skill’s behavior. The experimental results on robotic manipulation and locomotion tasks demonstrate that the proposed framework is able to efficiently learn primitive skills with diverse behaviors and coordinate multiple agents (end-effectors) to solve challenging cooperative control tasks. Acquiring skills without supervision and extending our method to a visual domain are exciting directions for future work. ", + "bbox": [ + 174, + 728, + 825, + 853 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 872, + 326, + 883 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project was funded by SKT. The authors would like to thank Karl Pertsch and many members of the USC CLVR lab for helpful discussion. ", + "bbox": [ + 174, + 895, + 823, + 922 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 102, + 287, + 118 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with policy sketches. In International Conference on Machine Learning, pp. 166–175, 2017. ", + "bbox": [ + 171, + 126, + 823, + 156 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pierre-Luc Bacon, Jean Harb, and Doina Precup. The option-critic architecture. In Association for the Advancement of Artificial Intelligence, 2017. ", + "bbox": [ + 171, + 165, + 823, + 195 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. ", + "bbox": [ + 171, + 204, + 823, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Prafulla Dhariwal, Christopher Hesse, Oleg Klimov, Alex Nichol, Matthias Plappert, Alec Radford, John Schulman, Szymon Sidor, Yuhuai Wu, and Peter Zhokhov. Openai baselines. https: //github.com/openai/baselines, 2017. ", + "bbox": [ + 173, + 243, + 828, + 286 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you need: Learning skills without a reward function. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=SJx63jRqFm. ", + "bbox": [ + 174, + 296, + 826, + 340 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfactual multi-agent policy gradients. In Association for the Advancement of Artificial Intelligence, 2018. ", + "bbox": [ + 173, + 349, + 826, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kevin Frans, Jonathan Ho, Xi Chen, Pieter Abbeel, and John Schulman. META LEARNING SHARED HIERARCHIES. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ SyX0IeWAW. ", + "bbox": [ + 174, + 402, + 825, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dibya Ghosh, Avi Singh, Aravind Rajeswaran, Vikash Kumar, and Sergey Levine. Divide-andconquer reinforcement learning. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ rJwelMbR-. ", + "bbox": [ + 174, + 455, + 826, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Anirudh Goyal, Shagun Sodhani, Jonathan Binas, Xue Bin Peng, Sergey Levine, and Yoshua Bengio. Reinforcement learning with competitive ensembles of information-constrained primitives. In International Conference on Learning Representations, 2020. URL https://openreview. net/forum?id $=$ ryxgJTEYDr. ", + "bbox": [ + 174, + 508, + 828, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jayesh K Gupta, Maxim Egorov, and Mykel Kochenderfer. Cooperative multi-agent control using deep reinforcement learning. In International Conference on Autonomous Agents and Multi-Agent Systems, pp. 66–83, 2017. ", + "bbox": [ + 173, + 575, + 823, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with deep energy-based policies. In International Conference on Machine Learning, pp. 1352–1361, 2017. ", + "bbox": [ + 174, + 628, + 825, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tuomas Haarnoja, Vitchyr Pong, Aurick Zhou, Murtaza Dalal, Pieter Abbeel, and Sergey Levine. Composable deep reinforcement learning for robotic manipulation. In IEEE International Conference on Robotics and Automation, pp. 6244–6251, 2018a. ", + "bbox": [ + 174, + 683, + 826, + 727 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International Conference on Machine Learning, pp. 1856–1865, 2018b. ", + "bbox": [ + 174, + 736, + 826, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation. In Advances in Neural Information Processing Systems, pp. 7254–7264, 2018. ", + "bbox": [ + 169, + 789, + 825, + 819 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in Neural Information Processing Systems, pp. 3675–3683, 2016. ", + "bbox": [ + 176, + 828, + 823, + 872 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Youngwoon Lee, Shao-Hua Sun, Sriram Somasundaram, Edward Hu, and Joseph J. Lim. Composing complex skills by learning transition policies. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ rygrBhC5tQ. ", + "bbox": [ + 176, + 881, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Andrew Levy, Robert Platt, and Kate Saenko. Hierarchical reinforcement learning with hindsight. In International Conference on Learning Representations, 2019. URL https://openreview. net/forum?id ${ . } = { }$ ryzECoAcY7. ", + "bbox": [ + 174, + 103, + 825, + 146 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems, pp. 6379–6390, 2017. ", + "bbox": [ + 176, + 154, + 823, + 196 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pattie Maes and Rodney A Brooks. Learning to coordinate behaviors. In Association for the Advancement of Artificial Intelligence, volume 90, pp. 796–802, 1990. ", + "bbox": [ + 169, + 204, + 825, + 233 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Josh Merel, Arun Ahuja, Vu Pham, Saran Tunyasuvunakool, Siqi Liu, Dhruva Tirumala, Nicolas Heess, and Greg Wayne. Hierarchical visuomotor control of humanoids. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id $=$ BJfYvo09Y7. ", + "bbox": [ + 173, + 241, + 825, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ofir Nachum, Shixiang Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical reinforcement learning. In Advances in Neural Information Processing Systems, pp. 3303–3313, 2018. ", + "bbox": [ + 176, + 304, + 823, + 347 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ofir Nachum, Michael Ahn, Hugo Ponte, Shixiang Gu, and Vikash Kumar. Multi-agent manipulation via locomotion using hierarchical sim2real. In Conference on Robot Learning, 2019. ", + "bbox": [ + 171, + 354, + 825, + 383 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning. In International Conference on Machine Learning, pp. 2661–2670, 2017. ", + "bbox": [ + 173, + 391, + 825, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. In Advances in Neural Information Processing Systems Autodiff Workshop, 2017. ", + "bbox": [ + 176, + 441, + 825, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, and Jun Wang. Multiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning to play starcraft combat games. arXiv preprint arXiv:1703.10069, 2017. ", + "bbox": [ + 174, + 492, + 825, + 535 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xue Bin Peng, Michael Chang, Grace Zhang, Pieter Abbeel, and Sergey Levine. Mcp: Learning composable hierarchical control with multiplicative compositional policies. Advances in Neural Information Processing Systems, 2019. ", + "bbox": [ + 174, + 542, + 825, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ahmed H. Qureshi, Jacob J. Johnson, Yuzhe Qin, Taylor Henderson, Byron Boots, and Michael C. Yip. Composing task-agnostic policies with deep reinforcement learning. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id $=$ H1ezFREtwH. ", + "bbox": [ + 174, + 592, + 826, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Martin Riedmiller, Roland Hafner, Thomas Lampe, Michael Neunert, Jonas Degrave, Tom Van de Wiele, Volodymyr Mnih, Nicolas Heess, and Jost Tobias Springenberg. Learning by playing - solving sparse reward tasks from scratch. In International Conference on Machine Learning, 2018. ", + "bbox": [ + 174, + 657, + 825, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jurgen Schmidhuber. ¨ Towards compositional learning with dynamic neural networks. Inst. fur¨ Informatik, 1990. ", + "bbox": [ + 173, + 707, + 823, + 736 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. In International Conference on Learning Representations, 2016. ", + "bbox": [ + 173, + 744, + 825, + 786 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. ", + "bbox": [ + 173, + 794, + 823, + 824 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation. In Advances in Neural Information Processing Systems, pp. 2244–2252, 2016. ", + "bbox": [ + 173, + 830, + 825, + 861 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl Tuyls, et al. Value-decomposition networks for cooperative multi-agent learning based on team reward. In International Conference on Autonomous Agents and Multi-Agent Systems, pp. 2085–2087, 2018. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Richard S Sutton. Temporal credit assignment in reinforcement learning. PhD thesis, University of Massachusetts, 1984. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181–211, 1999. ", + "bbox": [ + 173, + 140, + 825, + 170 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033, 2012. ", + "bbox": [ + 173, + 179, + 823, + 208 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A ENVIRONMENT DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 102, + 413, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The details of observation spaces, action spaces, number of agents, and episode lengths are described in Table 2. All units in this section are in meters unless otherwise specified. ", + "bbox": [ + 173, + 133, + 826, + 162 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/28b86d0a1f27a2f11bd196806df5f0e97f1f1883aa4bb3f5ac42b0e9dd4cf42d.jpg", + "table_caption": [ + "Table 2: Environment details " + ], + "table_footnote": [], + "table_body": "
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
Observation Space8888100
- Robot observation626282
- Object observation262618
Action Space181816
Number of Agents222
Episode length150100200
", + "bbox": [ + 233, + 178, + 763, + 280 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1 ENVIRONMENT DESCRIPTIONS ", + "text_level": 1, + "bbox": [ + 176, + 332, + 429, + 347 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In both Jaco environments, the robot works on a table with size (1.6, 1.6) and top center position $( 0 , 0 , 0 . 8 2 )$ . The two Jaco arms are initialized at positions $( - 0 . 1 6 , - 0 . 1 6 , 1 . 2 )$ and $( \\bar { - } 0 . 1 6 , 0 . \\bar { 2 } 4 , 1 . 2 )$ Left arm and right arm objects are initialized around $( 0 . 3 , 0 . 2 , 0 . 8 6 )$ and ( $0 . 3 , - 0 . 2 , 0 . 8 6 )$ respectively in all primitive training and composite task training environments, with small random position and rotation perturbation. ", + "bbox": [ + 173, + 358, + 825, + 429 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the Jaco Pick-Push-Place task, the right jaco arm needs to pick up the object and place it into the container initialized at the other side of the table. Success is defined by contact between the object and the inner top side of the container. ", + "bbox": [ + 174, + 435, + 825, + 478 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the Jaco Bar-Moving task, the two Jaco arms need to together pick the long bar up by height of 0.7, move it towards the arms by distance of 0.15, and place it back on the table. Success is defined by (1) the bar being placed within 0.04 away from the desired destination both in height and in xy-position and (2) the bar having been picked 0.7 above the table. ", + "bbox": [ + 174, + 484, + 825, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the Ant Push task, the two ant agents need to push a big box together to the goal position. The box has a size of $8 . 0 \\times 1 . 6 \\times 1 . 6$ . The distance between ants and the box is $2 0 \\mathrm { c m }$ and the distance between the box and the goal is $3 0 \\mathrm { c m }$ . Initial positions have $1 \\mathrm { c m }$ of randomness and the agent has a randomness of 0.01 in each joint. The task is considered as success when both the distances between left and right end of the box and the goal are within $5 \\mathrm { c m }$ . ", + "bbox": [ + 173, + 547, + 825, + 618 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.2 REWARD DESIGN ", + "text_level": 1, + "bbox": [ + 176, + 636, + 338, + 650 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For every task, we add a control penalty, $- 0 . 0 0 1 * \\| a \\| ^ { 2 }$ , to regularize the magnitude of actions where $a$ is a torque action performed by an agent. ", + "bbox": [ + 174, + 662, + 823, + 691 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jaco Pick: To help the agent learn to reach, pick, and hold the picked object, we provide dense reward to the agent defined by the weighted sum of pick reward, gripper-to-cube distance reward, cube position and quaternion stability reward, hold duration reward, success reward, and robot control reward. More concretely, ", + "bbox": [ + 173, + 698, + 825, + 755 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/7218e69723138229a1241e98230e1161cf031d597d01762af334ae8915a44c47.jpg", + "text": "$$\n\\begin{array} { r l } & { R ( s ) = \\lambda _ { p i c k } \\cdot ( z _ { b o x } - z _ { i n i t } ) + \\lambda _ { d i s t } \\cdot \\mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \\lambda _ { p o s } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { i n i t } ) + } \\\\ & { \\qquad \\lambda _ { q u a t } \\cdot \\mathrm { a b s } ( \\Delta _ { q u a t } ) + \\lambda _ { h o l d } \\cdot t _ { h o l d } + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 217, + 762, + 782, + 803 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "where $\\lambda _ { p i c k } = 5 0 0 , \\lambda _ { d i s t } = 1 0 0 , \\lambda _ { p o s } = 1 0 0 0 , \\lambda _ { q u a t } = 1 0 0 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 0 0 0 , \\lambda _ { d i s t }$ $1 \\times 1 0 ^ { - \\hat { 4 } }$ . ", + "bbox": [ + 173, + 808, + 825, + 838 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jaco Place: Reward for place primitive is defined by the weighted sum of xy-distance reward, height reward (larger when cube close to floor), success reward, and robot control reward. ", + "bbox": [ + 173, + 844, + 823, + 873 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/b98953b49bbde6752e8d53228b0cb112649e82f672d717587232fa8845567e86.jpg", + "text": "$$\n\\begin{array} { r l } & { R ( s ) = \\lambda _ { x y } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b o x } , p _ { g o a l } ) + \\lambda _ { z } \\cdot | z _ { b o x } - z _ { g o a l } | + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , } \\\\ & { \\mathrm { ~ e \\lambda _ { x y } = 5 0 0 , } \\lambda _ { z } = 5 0 0 , \\lambda _ { s u c c e s s } = 5 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 173, + 880, + 777, + 926 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jaco Push: Reward for push primitive is defined by the weighted sum of gripper reaching reward, box-to-destination distance reward, quaternion stability reward, hold duration reward, success reward, and robot control reward. ", + "bbox": [ + 173, + 103, + 826, + 146 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/0c68286af7bd7b82024e785edfec8cf4090d625996595f95caf4394882238f62.jpg", + "text": "$$\n\\begin{array} { r } { R ( s ) = \\lambda _ { r e a c h i n g } \\cdot \\mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \\lambda _ { p o s } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { d e s t } ) + \\phantom { x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x } } \\\\ { \\lambda _ { q u a t } \\cdot \\mathsf { a b s } ( \\Delta _ { q u a t } ) + \\lambda _ { h o l d } \\cdot t _ { h o l d } + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { s u c c e s s } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , \\phantom { x x x x x x x x x x x x x x } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 250, + 165, + 750, + 205 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/4ca888393fa5398d33bb9e915293fc545ae6a42f71899ee55aace23e9f7c6ce1.jpg", + "text": "$$\n\\lambda _ { r e a c h i n g } = 1 0 0 , \\lambda _ { p o s } = 5 0 0 , \\lambda _ { q u a t } = 3 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 } .\n$$", + "text_format": "latex", + "bbox": [ + 217, + 228, + 815, + 244 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Jaco Pick-Push-Place: Reward for Pick-Push-Place is defined by the weighted sum of gripper contact reward, per-stage reach/pick/push/place rewards, success reward, and control reward. We tune the reward carefully for all baselines. ", + "bbox": [ + 173, + 250, + 825, + 292 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/b9880c8304a14716c17f00f4bae48746e17f96e2635ba93e6f38d0c38a727515.jpg", + "text": "$$\n\\begin{array} { r l } & { R ( s ) = \\lambda _ { c o n t a c t } \\cdot \\big ( { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { l e f f ~ g r i p p e r ~ r o u c h e s ~ c o n t a i n e r } } } + { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { r i g h t ~ g r i p p e r ~ t o u c h e s ~ b o x } } } \\big ) + } \\\\ & { \\quad \\quad \\lambda _ { r e a c h } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { r e a c h } } } \\cdot \\big ( \\mathrm { d i s t } ( p _ { l e f t . g r i p p e r } , p _ { c o n t a i n e r } ) + \\mathrm { d i s t } ( p _ { r i g h t . g r i p p e r } , p _ { b o x } ) \\big ) + } \\\\ & { \\quad \\quad \\lambda _ { p i c k } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { p i c k } } } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + \\lambda _ { p l a c e } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { p l a c e } } } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + } \\\\ & { \\quad \\quad \\lambda _ { p u s h } \\cdot \\mathrm { d i s t } ( p _ { c o n t a i n e r } , p _ { c o n t a i n e r . t a r g e t } ) + \\lambda _ { s u c c e s s } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { s u c c e s s } } } + \\lambda _ { c t r l } \\cdot \\| a \\| ^ { 2 } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 218, + 314, + 782, + 395 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where $\\lambda _ { r e a c h } = 1 0 , \\lambda _ { c o n t a c t } = 1 0 , \\lambda _ { p i c k } = \\lambda _ { p l a c e } = \\lambda _ { p l a c e } = 1 0 , \\lambda _ { s u c c e s s } = 5 0 , \\lambda _ { c t r l } = 0 ,$ , and 1reaching, $\\mathbf { 1 } _ { \\mathrm { p i c k } }$ , and $\\mathbf { 1 } _ { \\mathrm { p l a c e } }$ are indicator functions specifying whether the agent is in reaching, pick or place stage. Agent stages are determined by how many multiples of 25 steps the agent has stepped through in the environment. ", + "bbox": [ + 173, + 415, + 826, + 472 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Jaco Bar-Moving: Reward for Bar-Moving is defined by the weighted sum of per-stage reach/pick/move/place rewards, success reward, and control reward. ", + "bbox": [ + 171, + 478, + 823, + 508 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/b0cc74c444bcd7c674d571d64c9cf8e175f34b540d056c3ea313f5b20ca868eb.jpg", + "text": "$$\n\\begin{array} { r l } & { R ( s ) = \\lambda _ { r e a c h } \\cdot { \\bf 1 } _ { r e a c h } \\cdot \\big ( \\mathrm { d i s t } ( p _ { l e f t \\_ g r i p p e r } , p _ { l e f t \\_ h a n d l e } ) + \\mathrm { d i s t } ( p _ { r i g h t \\_ g r i p p e r } , p _ { r i g h t \\_ h a n d l e } ) \\big ) + } \\\\ & { \\qquad \\lambda _ { p i c k } \\cdot { \\bf 1 } _ { \\mathrm { p i c k } } \\cdot \\mathrm { d i s t } ( p _ { b a r } , p _ { b a r \\_ t a r g e t } ) + \\lambda _ { m o v e } \\cdot { \\bf 1 } _ { \\mathrm { p l a e } } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b a r \\_ p b a r \\_ t a r g e t } ) + } \\\\ & { \\qquad \\lambda _ { p l a c e } \\cdot { \\bf 1 } _ { \\mathrm { p l a e } } \\cdot \\mathrm { d i s t } _ { z } ( p _ { b a r \\_ p b a r \\_ t a r g e t } ) + \\lambda _ { s u c e s s } \\cdot { \\bf 1 } _ { \\mathrm { s u c e s s } } + \\lambda _ { c t r l } \\cdot \\| a \\| ^ { 2 } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 183, + 530, + 816, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where $\\lambda _ { r e a c h } = 1 0 , \\lambda _ { p i c k } = 3 0 , \\lambda _ { m o v e } = 1 0 0 , \\lambda _ { p l a c e } = 1 0 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 }$ , and $\\mathbf { 1 } _ { \\mathrm { p i c k } }$ and $\\mathbf { 1 } _ { \\mathrm { p l a c e } }$ are indicator functions specifying whether the agent is in pick or place stage. Agent stages are determined by whether the pick objective is fulfilled or not. ", + "bbox": [ + 174, + 612, + 826, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ant Push & Ant Moving: Reward for ANT PUSH is defined by upright, velocity towards the desired direction. We provide a dense reward to encourage the desired locomotion behavior using velocity, stability, and posture, as following: ", + "bbox": [ + 174, + 661, + 826, + 705 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/a1fac56114820ef7a956ab1929340bdc0bd6b2e6f6daaf866ab74929c70ad2de.jpg", + "text": "$$\n\\begin{array} { r l } & { R ( s ) = \\lambda _ { v e l } \\cdot \\mathbf { a b s } ( \\Delta x _ { a n t } ) + \\lambda _ { b o x v e l } \\cdot \\mathbf { a b s } ( \\Delta x _ { b o x } ) + \\lambda _ { u p r i g h t } \\cdot \\mathbf { c o s } ( \\theta ) - \\lambda _ { h e i g h t } \\cdot \\mathbf { a b s } ( 0 . 6 - h ) + } \\\\ & { ~ \\lambda _ { g o a l } \\cdot \\mathbf { d i s t } ( p _ { g o a l } , p _ { b o x } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 183, + 727, + 820, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where $\\lambda _ { v e l } = 5 0 , \\lambda _ { b o x v e l } = 2 0 , \\lambda _ { u p r i g h t } = 1 , \\lambda _ { h e i g h t } = 0 . 5$ . For ANT PUSH, we provide an additional reward based on distance between the box and the goal position with $\\lambda _ { g o a l } = 2 0 0$ . ", + "bbox": [ + 173, + 785, + 825, + 815 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B EXPERIMENT DETAILS ", + "text_level": 1, + "bbox": [ + 174, + 853, + 397, + 869 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We use PyTorch (Paszke et al., 2017) for our implementation and all experiments are conducted on a workstation with Intel Xeon Gold 6154 CPU and 4 NVIDIA GeForce RTX 2080 Ti GPUs. ", + "bbox": [ + 173, + 895, + 825, + 924 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B.1 HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 174, + 103, + 359, + 117 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/2e0f2d3e912b19281466669264a4f5594f304b9e72c85d327661d9f4d2d9566c.jpg", + "table_caption": [ + "Table 3: Hyperparameters " + ], + "table_footnote": [], + "table_body": "
ParametersValue
learning rate3e-4
gradient steps50
batch size256
discount factor0.99
target smoothing coefficient0.005
reward scale (SAC)1.0
experience buffer size (# episodes)1000
Tow1 for JACO, 5 for ANT
Nz (dimensionality of z)5
", + "bbox": [ + 294, + 133, + 704, + 276 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B.2 NETWORK ARCHITECTURES ", + "text_level": 1, + "bbox": [ + 176, + 327, + 411, + 340 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Actor Networks: In all experiments, we model our actor network for each primitive skill as a 3-layer MLP with hidden layer size 64. The last layer of the MLP is two-headed – one for the mean of the action distribution and the other for the standard deviation of it. We use ReLU as activation function in hidden layers. We do not apply any activation function for the final output layer. The action distribution output represents per-dimension normal distribution, from which single actions can be sampled and executed in the environment. ", + "bbox": [ + 173, + 352, + 825, + 435 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Critic Networks: The critic network for each primitive skill and meta policy is modeled as a 2-layer MLP with hidden layer size 128. ReLU is used as an activation function in the hidden layers. The critic network output is used to assess the value of a given state-action pair, and is trained by fitting its outputs to the target Q-value clamped by $\\pm 1 0 0$ . ", + "bbox": [ + 174, + 443, + 825, + 500 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Meta Policy: The meta policy is modeled as a 3-layer MLP with hidden layer size of 64. Since the actions of meta policy are sampled from $N$ categorical distributions for each end-effector/agent and $N$ normal distributions for behavior embeddings, the output dimension of the meta policy is $\\begin{array} { r } { \\sum _ { i = 1 } ^ { N } ( m ^ { i } + N _ { z } ) } \\end{array}$ . The meta policy uses ReLU as an activation function for all layers except for the final output layer. ", + "bbox": [ + 174, + 506, + 825, + 579 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B.3 TRAINING DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 595, + 352, + 609 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "For all baselines, we train the meta policies using PPO and the low-level policies using SAC. We use the same environment configurations, composite task reward definitions, and value of $T _ { l o w }$ across all baselines. ", + "bbox": [ + 174, + 621, + 825, + 662 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "For Jaco tasks, we train a total of 4 primitive skills – right arm pick, right arm place-to-center, right arm place-towards-arm, and left arm push – to be composed by meta-policy. For Jaco Pick-PushPlace, we provide the meta-policy with right arm pick and right arm place-to-center as right arm primitives and left arm push as left arm primitives; for Jaco Bar-Moving, we provide the meta-policy with right arm pick and right arm place-towards-arm as both right and left arm primitives and left arm pick and right arm place-towards-arm as left arm primitives. We obtain left arm primitives for bar-moving task by using the learned right arm primitives directly. ", + "bbox": [ + 173, + 670, + 825, + 768 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "To obtain the 4 primitives skills described above, we train right arm pick with diversity coefficient $\\lambda _ { 2 } = 0 . 0 1$ and the other three primitives with $\\lambda _ { 2 } = 0 . 1$ . The destination of right arm push is set to $( 0 . 3 , - 0 . 0 3 , 0 . 8 6 )$ , which is slightly left of the center of the table. After pick primitive is trained, we train the two right arm place primitives where episodes are initialized by intermediate states of successful right arm pick episodes where the height of the box is larger than 0.94 (0.01 higher than the target pick height). The place destinations for towards-arm and to-center primitives are $( 0 . 1 5 , - 0 . 2 , 0 . 8 6 )$ and $( 0 . 3 , - 0 . 0 2 , 0 . 8 6 )$ , respectively. ", + "bbox": [ + 173, + 773, + 825, + 873 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "For non-modular baselines that incorporates skill behavior diversification, we use $\\lambda _ { 2 } = 0 . 0 1$ for both Jaco Pick-Push-Place and Jaco Bar-Moving because both tasks require picking skills, which can only be trained with a small value of $\\lambda _ { 2 }$ . ", + "bbox": [ + 176, + 878, + 823, + 921 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/parse/train/ryxB2lBtvH/ryxB2lBtvH_middle.json b/parse/train/ryxB2lBtvH/ryxB2lBtvH_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..0956e6e0a32b6c0a6d7ddea3849095256404e759 --- /dev/null +++ b/parse/train/ryxB2lBtvH/ryxB2lBtvH_middle.json @@ -0,0 +1,39988 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 493, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 494, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 494, + 97 + ], + "score": 1.0, + "content": "LEARNING TO COORDINATE MANIPULATION SKILLS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 404, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 404, + 118 + ], + "score": 1.0, + "content": "VIA SKILL BEHAVIOR DIVERSIFICATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 135, + 310, + 147 + ], + "lines": [ + { + "bbox": [ + 112, + 134, + 311, + 150 + ], + "spans": [ + { + "bbox": [ + 112, + 134, + 311, + 150 + ], + "score": 1.0, + "content": "Youngwoon Lee, Jingyun Yang, Joseph J. Lim", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 113, + 147, + 297, + 180 + ], + "lines": [ + { + "bbox": [ + 111, + 145, + 249, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 249, + 158 + ], + "score": 1.0, + "content": "Department of Computer Science", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 157, + 249, + 168 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 249, + 168 + ], + "score": 1.0, + "content": "University of Southern California", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 168, + 298, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 298, + 181 + ], + "score": 1.0, + "content": "{lee504,jingyuny,limjj}@usc.edu", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 208, + 333, + 220 + ], + "lines": [ + { + "bbox": [ + 277, + 208, + 335, + 221 + ], + "spans": [ + { + "bbox": [ + 277, + 208, + 335, + 221 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 232, + 468, + 363 + ], + "lines": [ + { + "bbox": [ + 142, + 232, + 469, + 244 + ], + "spans": [ + { + "bbox": [ + 142, + 232, + 469, + 244 + ], + "score": 1.0, + "content": "When mastering a complex manipulation task, humans often decompose the task", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 243, + 470, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 470, + 255 + ], + "score": 1.0, + "content": "into sub-skills of their body parts, practice the sub-skills independently, and then", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 470, + 266 + ], + "score": 1.0, + "content": "execute the sub-skills together. Similarly, a robot with multiple end-effectors can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 265, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 265, + 469, + 276 + ], + "score": 1.0, + "content": "perform complex tasks by coordinating sub-skills of each end-effector. To realize", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 276, + 470, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 276, + 470, + 287 + ], + "score": 1.0, + "content": "temporal and behavioral coordination of skills, we propose a modular framework", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 286, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 286, + 470, + 298 + ], + "score": 1.0, + "content": "that first individually trains sub-skills of each end-effector with skill behavior", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 297, + 470, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 470, + 309 + ], + "score": 1.0, + "content": "diversification, and then learns to coordinate end-effectors using diverse behaviors", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "score": 1.0, + "content": "of the skills. We demonstrate that our proposed framework is able to efficiently", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 318, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 470, + 333 + ], + "score": 1.0, + "content": "coordinate skills to solve challenging collaborative control tasks such as picking up", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 331, + 470, + 343 + ], + "spans": [ + { + "bbox": [ + 142, + 331, + 470, + 343 + ], + "score": 1.0, + "content": "a long bar, placing a block inside a container while pushing the container with two", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 342, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 470, + 353 + ], + "score": 1.0, + "content": "robot arms, and pushing a box with two ant agents. Videos and code are available", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 353, + 343, + 364 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 343, + 364 + ], + "score": 1.0, + "content": "at https://clvrai.com/coordination.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 109, + 382, + 206, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 208, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 208, + 398 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "Imagine you wish to play Chopin’s Fantaisie Impromptu on the piano. With little prior knowledge", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 418, + 507, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 507, + 430 + ], + "score": 1.0, + "content": "about the piece, you would first practice playing the piece with each hand separately. After inde-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "pendently mastering the left and right hand parts, you would move on to practicing with both hands", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "simultaneously. To find the synchronized and non-interfering movements of two hands, you would", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "try variable ways of playing the same melody with each hand, and eventually create a complete piece", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "score": 1.0, + "content": "of music. Through the decomposition of skills into sub-skills of two hands and learning variations", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 471, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 486 + ], + "score": 1.0, + "content": "of sub-skills, humans make the learning process of manipulation skills much faster than learning", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 484, + 185, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 185, + 497 + ], + "score": 1.0, + "content": "everything at once.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "Can autonomous agents efficiently learn complicated tasks with coordination of different skills from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "multiple end-effectors like humans? Learning to perform collaborative and composite tasks from", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "scratch requires a huge amount of environment interaction and extensive reward engineering, which", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "score": 1.0, + "content": "often results in undesired behaviors (Riedmiller et al., 2018). Hence, instead of learning a task at once,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 543, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 543, + 506, + 558 + ], + "score": 1.0, + "content": "modular approaches (Andreas et al., 2017; Oh et al., 2017; Frans et al., 2018; Lee et al., 2019; Peng", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 104, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "et al., 2019; Goyal et al., 2020) suggest to learn reusable primitive skills and solve more complex", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "tasks by recombining the skills. However, all these approaches either focus on working with single", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "end-effector manipulation or single agent locomotion, and these do not scale to multi-agent problems.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "To this end, we propose a modular framework that learns to coordinate multiple end-effectors", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "with their primitive skills for various robotics tasks, such as bimanual manipulation. The main", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "challenge is that naive simultaneous execution of primitive skills from multiple end-effectors can", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "often cause unintended behaviors (e.g. collisions between end-effectors). Thus, as illustrated in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 652 + ], + "score": 1.0, + "content": "Figure 1, our model needs to learn to appropriately coordinate end-effectors; and hence needs a way", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "to obtain, represent, and control detailed behaviors of each primitive skill. Inspired by these intuitions,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "our method consists of two parts: (1) acquiring primitive skills with diverse behaviors by mutual", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "information maximization, and (2) learning a meta policy that selects a skill for each end-effector and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 682, + 387, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 387, + 694 + ], + "score": 1.0, + "content": "coordinates the chosen skills by controlling the behavior of each skill.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "The main contribution of this paper is a modular and hierarchical approach that tackles cooperative", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "manipulation tasks with multiple end-effectors by (1) learning primitive skills of each end-effector", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "independently with skill behavior diversification and (2) coordinating end-effectors using diverse", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 493, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 494, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 494, + 97 + ], + "score": 1.0, + "content": "LEARNING TO COORDINATE MANIPULATION SKILLS", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 404, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 404, + 118 + ], + "score": 1.0, + "content": "VIA SKILL BEHAVIOR DIVERSIFICATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 114, + 135, + 310, + 147 + ], + "lines": [ + { + "bbox": [ + 112, + 134, + 311, + 150 + ], + "spans": [ + { + "bbox": [ + 112, + 134, + 311, + 150 + ], + "score": 1.0, + "content": "Youngwoon Lee, Jingyun Yang, Joseph J. Lim", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 112, + 134, + 311, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 147, + 297, + 180 + ], + "lines": [ + { + "bbox": [ + 111, + 145, + 249, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 249, + 158 + ], + "score": 1.0, + "content": "Department of Computer Science", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 157, + 249, + 168 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 249, + 168 + ], + "score": 1.0, + "content": "University of Southern California", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 168, + 298, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 298, + 181 + ], + "score": 1.0, + "content": "{lee504,jingyuny,limjj}@usc.edu", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 111, + 145, + 298, + 181 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 208, + 333, + 220 + ], + "lines": [ + { + "bbox": [ + 277, + 208, + 335, + 221 + ], + "spans": [ + { + "bbox": [ + 277, + 208, + 335, + 221 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 143, + 232, + 468, + 363 + ], + "lines": [ + { + "bbox": [ + 142, + 232, + 469, + 244 + ], + "spans": [ + { + "bbox": [ + 142, + 232, + 469, + 244 + ], + "score": 1.0, + "content": "When mastering a complex manipulation task, humans often decompose the task", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 243, + 470, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 243, + 470, + 255 + ], + "score": 1.0, + "content": "into sub-skills of their body parts, practice the sub-skills independently, and then", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 253, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 470, + 266 + ], + "score": 1.0, + "content": "execute the sub-skills together. Similarly, a robot with multiple end-effectors can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 265, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 265, + 469, + 276 + ], + "score": 1.0, + "content": "perform complex tasks by coordinating sub-skills of each end-effector. To realize", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 276, + 470, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 276, + 470, + 287 + ], + "score": 1.0, + "content": "temporal and behavioral coordination of skills, we propose a modular framework", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 286, + 470, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 286, + 470, + 298 + ], + "score": 1.0, + "content": "that first individually trains sub-skills of each end-effector with skill behavior", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 297, + 470, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 470, + 309 + ], + "score": 1.0, + "content": "diversification, and then learns to coordinate end-effectors using diverse behaviors", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 470, + 321 + ], + "score": 1.0, + "content": "of the skills. We demonstrate that our proposed framework is able to efficiently", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 318, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 470, + 333 + ], + "score": 1.0, + "content": "coordinate skills to solve challenging collaborative control tasks such as picking up", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 331, + 470, + 343 + ], + "spans": [ + { + "bbox": [ + 142, + 331, + 470, + 343 + ], + "score": 1.0, + "content": "a long bar, placing a block inside a container while pushing the container with two", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 342, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 470, + 353 + ], + "score": 1.0, + "content": "robot arms, and pushing a box with two ant agents. Videos and code are available", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 353, + 343, + 364 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 343, + 364 + ], + "score": 1.0, + "content": "at https://clvrai.com/coordination.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 12.5, + "bbox_fs": [ + 141, + 232, + 470, + 364 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 382, + 206, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 208, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 208, + 398 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 420 + ], + "score": 1.0, + "content": "Imagine you wish to play Chopin’s Fantaisie Impromptu on the piano. With little prior knowledge", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 418, + 507, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 507, + 430 + ], + "score": 1.0, + "content": "about the piece, you would first practice playing the piece with each hand separately. After inde-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "pendently mastering the left and right hand parts, you would move on to practicing with both hands", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "simultaneously. To find the synchronized and non-interfering movements of two hands, you would", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "try variable ways of playing the same melody with each hand, and eventually create a complete piece", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 475 + ], + "score": 1.0, + "content": "of music. Through the decomposition of skills into sub-skills of two hands and learning variations", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 471, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 486 + ], + "score": 1.0, + "content": "of sub-skills, humans make the learning process of manipulation skills much faster than learning", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 484, + 185, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 185, + 497 + ], + "score": 1.0, + "content": "everything at once.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 406, + 507, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "Can autonomous agents efficiently learn complicated tasks with coordination of different skills from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "multiple end-effectors like humans? Learning to perform collaborative and composite tasks from", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 534 + ], + "score": 1.0, + "content": "scratch requires a huge amount of environment interaction and extensive reward engineering, which", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 507, + 547 + ], + "score": 1.0, + "content": "often results in undesired behaviors (Riedmiller et al., 2018). Hence, instead of learning a task at once,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 543, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 104, + 543, + 506, + 558 + ], + "score": 1.0, + "content": "modular approaches (Andreas et al., 2017; Oh et al., 2017; Frans et al., 2018; Lee et al., 2019; Peng", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 104, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "et al., 2019; Goyal et al., 2020) suggest to learn reusable primitive skills and solve more complex", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "tasks by recombining the skills. However, all these approaches either focus on working with single", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "end-effector manipulation or single agent locomotion, and these do not scale to multi-agent problems.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 104, + 500, + 507, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "To this end, we propose a modular framework that learns to coordinate multiple end-effectors", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "with their primitive skills for various robotics tasks, such as bimanual manipulation. The main", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "challenge is that naive simultaneous execution of primitive skills from multiple end-effectors can", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "often cause unintended behaviors (e.g. collisions between end-effectors). Thus, as illustrated in", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 652 + ], + "score": 1.0, + "content": "Figure 1, our model needs to learn to appropriately coordinate end-effectors; and hence needs a way", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "to obtain, represent, and control detailed behaviors of each primitive skill. Inspired by these intuitions,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "our method consists of two parts: (1) acquiring primitive skills with diverse behaviors by mutual", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 683 + ], + "score": 1.0, + "content": "information maximization, and (2) learning a meta policy that selects a skill for each end-effector and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 682, + 387, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 387, + 694 + ], + "score": 1.0, + "content": "coordinates the chosen skills by controlling the behavior of each skill.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 594, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "The main contribution of this paper is a modular and hierarchical approach that tackles cooperative", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "manipulation tasks with multiple end-effectors by (1) learning primitive skills of each end-effector", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "independently with skill behavior diversification and (2) coordinating end-effectors using diverse", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "behaviors of the skills. Our empirical results indicate that our proposed method is able to efficiently", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "learn primitive skills with diverse behaviors and coordinate these skills to solve challenging collabo-", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 284, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 295 + ], + "score": 1.0, + "content": "rative control tasks such as picking up a long bar, placing a block inside the container on the right", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 294, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 506, + 307 + ], + "score": 1.0, + "content": "side, and pushing a box with two ant agents. We provide additional qualitative results and code at", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 306, + 297, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 297, + 317 + ], + "score": 1.0, + "content": "https://clvrai.com/coordination.", + "type": "text", + "cross_page": true + } + ], + "index": 14 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 81, + 503, + 153 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 81, + 503, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 81, + 503, + 153 + ], + "spans": [ + { + "bbox": [ + 108, + 81, + 503, + 153 + ], + "score": 0.966, + "type": "image", + "image_path": "7e1b9f5f33a08c620e270f4735ec881e08b70e6cf015791f9524143e7c2eda6a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 81, + 503, + 105.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 105.0, + 503, + 129.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 129.0, + 503, + 153.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 162, + 505, + 239 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "Figure 1: Composing complex skills using multiple agents’ primitive skills requires proper coordina-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 172, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 186 + ], + "score": 1.0, + "content": "tion between agents since concurrent execution of primitive skills requires temporal and behavioral", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "coordination. For example, to move a block into a container on the other end of the table, the agent", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "needs to not only utilize pick, place, and push primitive skills at the right time but also select the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 205, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 378, + 219 + ], + "score": 1.0, + "content": "appropriate behaviors for these skills, represented as latent vectors", + "type": "text" + }, + { + "bbox": [ + 378, + 208, + 388, + 217 + ], + "score": 0.71, + "content": "z _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 205, + 392, + 219 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 392, + 208, + 402, + 217 + ], + "score": 0.59, + "content": "z _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 205, + 405, + 219 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 406, + 208, + 416, + 217 + ], + "score": 0.61, + "content": "z _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 205, + 437, + 219 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 437, + 207, + 448, + 217 + ], + "score": 0.86, + "content": "z _ { 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 205, + 506, + 219 + ], + "score": 1.0, + "content": "above. Naive", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "methods neglecting either temporal or behavioral coordination will produce unintended behaviors,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 227, + 272, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 272, + 240 + ], + "score": 1.0, + "content": "such as collisions between end-effectors.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "behaviors of the skills. Our empirical results indicate that our proposed method is able to efficiently", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "learn primitive skills with diverse behaviors and coordinate these skills to solve challenging collabo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 284, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 295 + ], + "score": 1.0, + "content": "rative control tasks such as picking up a long bar, placing a block inside the container on the right", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 294, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 506, + 307 + ], + "score": 1.0, + "content": "side, and pushing a box with two ant agents. We provide additional qualitative results and code at", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 306, + 297, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 297, + 317 + ], + "score": 1.0, + "content": "https://clvrai.com/coordination.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 333, + 209, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 210, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 210, + 347 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) for continuous control is an active research area. However, learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "score": 1.0, + "content": "a complex task either from a sparse reward or a heavily engineered reward becomes computationally", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 506, + 393 + ], + "score": 1.0, + "content": "impractical as the target task becomes complicated. Instead of learning from scratch, complex", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "tasks can be tackled by decomposing the tasks into easier and reusable sub-tasks. Hierarchical", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "reinforcement learning temporally splits a task into a sequence of temporally extended meta actions.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "It often consists of one meta policy (high-level policy) and a set of low-level policies, such as options", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "framework (Sutton et al., 1999). The meta policy decides which low-level policy to activate and the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "chosen low-level policy generates an action sequence until the meta policy switches it to another low-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "level policy. Options can be discovered without supervision (Schmidhuber, 1990; Bacon et al., 2017;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "Nachum et al., 2018; Levy et al., 2019), meta-learned (Frans et al., 2018), pre-defined (Kulkarni et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "2016; Oh et al., 2017; Merel et al., 2019; Lee et al., 2019), or attained from additional supervision", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "signals (Andreas et al., 2017; Ghosh et al., 2018). However, option frameworks are not flexible to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 490, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 501 + ], + "score": 1.0, + "content": "solve a task that requires simultaneous activation or interpolation of multiple skills since only one", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 500, + 264, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 264, + 514 + ], + "score": 1.0, + "content": "skill can be activated at each time step.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 507, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 507, + 531 + ], + "score": 1.0, + "content": "To solve composite tasks multiple policies can be simultaneously activated by adding Q-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "functions (Haarnoja et al., 2018a), additive composition (Qureshi et al., 2020; Goyal et al., 2020), or", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "score": 1.0, + "content": "multiplicative composition (Peng et al., 2019). As each policy takes the whole observation as input", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "and controls the whole agent, it is not robust to changes in unrelated parts of the observation. For", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "score": 1.0, + "content": "example, a left arm skill can be affected by the pose change in the right arm, which is not relevant to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "the left arm skill. Hence, these skill composition approaches can fail when an agent encounters a new", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 507, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 507, + 595 + ], + "score": 1.0, + "content": "combination of skills or a new skill is introduced since the agent will experience unseen observations.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "Instead of having a policy with the full observation and action space, multi-agent reinforcement", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "learning (MARL) suggests to explicitly split the observation and action space according to agents", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "(e.g. robots or end-effectors), which allows efficient low-level policy training as well as flexible skill", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "composition. For cooperative tasks, communication mechanisms (Sukhbaatar et al., 2016; Peng et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "2017; Jiang & Lu, 2018), sharing policy parameters (Gupta et al., 2017), and decentralized actor", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "with centralized critic (Lowe et al., 2017; Foerster et al., 2018) have been actively used. However,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "these approaches suffer from the credit assignment problem (Sutton, 1984) among agents and the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "lazy agent problem (Sunehag et al., 2018). As agents have more complicated morphology and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "larger observation space, learning a policy for a multi-agent system from scratch requires extremely", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "long training time. Moreover, the credit assignment problem becomes more challenging when the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "complexity of cooperative tasks increases and all agents need to learn completely from scratch. To", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "resolve these issues, we propose to first train reusable skills for each agent in isolation, instead of", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 42.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 81, + 503, + 153 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 81, + 503, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 81, + 503, + 153 + ], + "spans": [ + { + "bbox": [ + 108, + 81, + 503, + 153 + ], + "score": 0.966, + "type": "image", + "image_path": "7e1b9f5f33a08c620e270f4735ec881e08b70e6cf015791f9524143e7c2eda6a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 81, + 503, + 105.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 105.0, + 503, + 129.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 129.0, + 503, + 153.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 162, + 505, + 239 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "Figure 1: Composing complex skills using multiple agents’ primitive skills requires proper coordina-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 172, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 186 + ], + "score": 1.0, + "content": "tion between agents since concurrent execution of primitive skills requires temporal and behavioral", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "coordination. For example, to move a block into a container on the other end of the table, the agent", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "needs to not only utilize pick, place, and push primitive skills at the right time but also select the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 205, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 378, + 219 + ], + "score": 1.0, + "content": "appropriate behaviors for these skills, represented as latent vectors", + "type": "text" + }, + { + "bbox": [ + 378, + 208, + 388, + 217 + ], + "score": 0.71, + "content": "z _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 205, + 392, + 219 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 392, + 208, + 402, + 217 + ], + "score": 0.59, + "content": "z _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 205, + 405, + 219 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 406, + 208, + 416, + 217 + ], + "score": 0.61, + "content": "z _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 205, + 437, + 219 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 437, + 207, + 448, + 217 + ], + "score": 0.86, + "content": "z _ { 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 205, + 506, + 219 + ], + "score": 1.0, + "content": "above. Naive", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "methods neglecting either temporal or behavioral coordination will produce unintended behaviors,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 227, + 272, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 272, + 240 + ], + "score": 1.0, + "content": "such as collisions between end-effectors.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 316 + ], + "lines": [], + "index": 12, + "bbox_fs": [ + 105, + 261, + 506, + 317 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 333, + 209, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 210, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 210, + 347 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "Deep reinforcement learning (RL) for continuous control is an active research area. However, learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "score": 1.0, + "content": "a complex task either from a sparse reward or a heavily engineered reward becomes computationally", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 506, + 393 + ], + "score": 1.0, + "content": "impractical as the target task becomes complicated. Instead of learning from scratch, complex", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "tasks can be tackled by decomposing the tasks into easier and reusable sub-tasks. Hierarchical", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "reinforcement learning temporally splits a task into a sequence of temporally extended meta actions.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "It often consists of one meta policy (high-level policy) and a set of low-level policies, such as options", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "framework (Sutton et al., 1999). The meta policy decides which low-level policy to activate and the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "chosen low-level policy generates an action sequence until the meta policy switches it to another low-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "level policy. Options can be discovered without supervision (Schmidhuber, 1990; Bacon et al., 2017;", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 469 + ], + "score": 1.0, + "content": "Nachum et al., 2018; Levy et al., 2019), meta-learned (Frans et al., 2018), pre-defined (Kulkarni et al.,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "2016; Oh et al., 2017; Merel et al., 2019; Lee et al., 2019), or attained from additional supervision", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "signals (Andreas et al., 2017; Ghosh et al., 2018). However, option frameworks are not flexible to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 490, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 501 + ], + "score": 1.0, + "content": "solve a task that requires simultaneous activation or interpolation of multiple skills since only one", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 500, + 264, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 264, + 514 + ], + "score": 1.0, + "content": "skill can be activated at each time step.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 356, + 506, + 514 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 507, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 507, + 531 + ], + "score": 1.0, + "content": "To solve composite tasks multiple policies can be simultaneously activated by adding Q-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "functions (Haarnoja et al., 2018a), additive composition (Qureshi et al., 2020; Goyal et al., 2020), or", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 552 + ], + "score": 1.0, + "content": "multiplicative composition (Peng et al., 2019). As each policy takes the whole observation as input", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 563 + ], + "score": 1.0, + "content": "and controls the whole agent, it is not robust to changes in unrelated parts of the observation. For", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 506, + 574 + ], + "score": 1.0, + "content": "example, a left arm skill can be affected by the pose change in the right arm, which is not relevant to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "the left arm skill. Hence, these skill composition approaches can fail when an agent encounters a new", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 583, + 507, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 507, + 595 + ], + "score": 1.0, + "content": "combination of skills or a new skill is introduced since the agent will experience unseen observations.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 516, + 507, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "Instead of having a policy with the full observation and action space, multi-agent reinforcement", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "learning (MARL) suggests to explicitly split the observation and action space according to agents", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "(e.g. robots or end-effectors), which allows efficient low-level policy training as well as flexible skill", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "composition. For cooperative tasks, communication mechanisms (Sukhbaatar et al., 2016; Peng et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "2017; Jiang & Lu, 2018), sharing policy parameters (Gupta et al., 2017), and decentralized actor", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "with centralized critic (Lowe et al., 2017; Foerster et al., 2018) have been actively used. However,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 679 + ], + "score": 1.0, + "content": "these approaches suffer from the credit assignment problem (Sutton, 1984) among agents and the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "lazy agent problem (Sunehag et al., 2018). As agents have more complicated morphology and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 701 + ], + "score": 1.0, + "content": "larger observation space, learning a policy for a multi-agent system from scratch requires extremely", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "long training time. Moreover, the credit assignment problem becomes more challenging when the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "complexity of cooperative tasks increases and all agents need to learn completely from scratch. To", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "resolve these issues, we propose to first train reusable skills for each agent in isolation, instead of", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "learning primitive skills of multiple agents together. Then, we recombine these skills (Maes & Brooks,", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 298, + 433, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 433, + 311 + ], + "score": 1.0, + "content": "1990) to complete more complicated tasks with learned coordination of the skills.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 600, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 81, + 504, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 81, + 504, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 504, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 504, + 198 + ], + "score": 0.961, + "type": "image", + "image_path": "63021c65e6e95a999eba964c17a1759f617f5364c3d76c2f449409b6d5dca506.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 81, + 504, + 120.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 120.0, + 504, + 159.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 159.0, + 504, + 198.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 208, + 506, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "Figure 2: Our method is composed of two components: a meta policy and a set of agent-specific", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "primitive policies relevant to task completion. The meta policy selects which primitive skill to run for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "each agent as well as the behavior embedding (i.e. variation in behavior) of the chosen primitive skill.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Each selected primitive skill takes as input the agent observation and the behavior embedding and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 253, + 222, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 222, + 266 + ], + "score": 1.0, + "content": "outputs action for that agent.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 287, + 504, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "learning primitive skills of multiple agents together. Then, we recombine these skills (Maes & Brooks,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 298, + 433, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 433, + 311 + ], + "score": 1.0, + "content": "1990) to complete more complicated tasks with learned coordination of the skills.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 315, + 506, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 505, + 327 + ], + "score": 1.0, + "content": "To coordinate skills from multiple agents, the skills have to be flexible; hence, a skill can be adjusted", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "to collaborate with other agents’ skills. Maximum entropy policies (Haarnoja et al., 2017; 2018a;b)", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 336, + 507, + 351 + ], + "spans": [ + { + "bbox": [ + 104, + 336, + 507, + 351 + ], + "score": 1.0, + "content": "can learn diverse ways to achieve a goal by maximizing not only reward but also entropy of the policy.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 347, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 104, + 347, + 506, + 362 + ], + "score": 1.0, + "content": "In addition, Eysenbach et al. (2019) proposes to discover diverse skills without reward by maximizing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 360, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 506, + 372 + ], + "score": 1.0, + "content": "entropy as well as mutual information between resulting states and latent representations of skills (i.e.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "skill embeddings). Our method leverages the maximum entropy policy (Haarnoja et al., 2018b) with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "the discriminability objective (Eysenbach et al., 2019) to learn a primitive skill with diverse behaviors", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "score": 1.0, + "content": "conditioned on a controllable skill embedding. This controllable skill embedding will be later used as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 403, + 491, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 491, + 416 + ], + "score": 1.0, + "content": "a behavior embedding for the meta policy to adjust a primitive skill’s behavior for coordination.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 108, + 432, + 172, + 444 + ], + "lines": [ + { + "bbox": [ + 104, + 430, + 174, + 448 + ], + "spans": [ + { + "bbox": [ + 104, + 430, + 174, + 448 + ], + "score": 1.0, + "content": "3 METHOD", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 458, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "In this paper, we address the problem of solving cooperative manipulation tasks that require collab-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "score": 1.0, + "content": "oration between multiple end-effectors or agents. Note that we use the terms “end-effector” and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "“agent” interchangeably in this paper. Instead of learning a multi-agent task from scratch (Lowe et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "2017; Gupta et al., 2017; Sunehag et al., 2018; Foerster et al., 2018), modular approaches (Andreas", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "et al., 2017; Frans et al., 2018; Peng et al., 2019) suggest to learn reusable primitive skills and solve", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "more complex tasks by recombining these skills. However, concurrent execution of primitive skills of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "multiple agents fails when agents never experienced a combination of skills during the pre-training", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 536, + 345, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 345, + 547 + ], + "score": 1.0, + "content": "stage, or skills require temporal or behavioral coordination.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 551, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "Therefore, we propose a modular and hierarchical framework that learns to coordinate multiple agents", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 562, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 576 + ], + "score": 1.0, + "content": "with primitive skills to perform a complex task. Moreover, during primitive skill training, we propose", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "to learn a latent behavior embedding, which provides controllability of each primitive skill to the meta", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "policy while coordinating skills. In Section 3.2, we describe our modular framework in detail. Next,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "in Section 3.3, we elaborate how controllable primitive skills can be acquired. Lastly, we describe", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 606, + 394, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 394, + 619 + ], + "score": 1.0, + "content": "how the meta policy learns to coordinate primitive skills in Section 3.4.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 107, + 633, + 200, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 201, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 201, + 646 + ], + "score": 1.0, + "content": "3.1 PRELIMINARIES", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 507, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 394, + 667 + ], + "score": 1.0, + "content": "We formulate our problem as a Markov decision process defined by a tuple", + "type": "text" + }, + { + "bbox": [ + 395, + 655, + 469, + 667 + ], + "score": 0.91, + "content": "\\{ S , { \\mathcal { A } } , { \\mathcal { T } } , R , \\rho , \\gamma \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 655, + 507, + 667 + ], + "score": 1.0, + "content": "of states,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "actions, transition probability, reward, initial state distribution, and discount factor. In our formulation,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 255, + 690 + ], + "score": 1.0, + "content": "we assume the environment includes", + "type": "text" + }, + { + "bbox": [ + 256, + 679, + 266, + 687 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "agents. To promote consistency in our terminology, we use", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "superscripts to denote the index of agent and subscripts to denote time or primitive skill index. Hence,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 281, + 711 + ], + "score": 1.0, + "content": "the state space and action space for an agent", + "type": "text" + }, + { + "bbox": [ + 281, + 701, + 285, + 708 + ], + "score": 0.82, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 698, + 372, + 711 + ], + "score": 1.0, + "content": "can be represented as", + "type": "text" + }, + { + "bbox": [ + 372, + 700, + 383, + 708 + ], + "score": 0.89, + "content": "S ^ { \\imath }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 698, + 401, + 711 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 699, + 413, + 709 + ], + "score": 0.87, + "content": "A ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "where each element of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 709, + 118, + 720 + ], + "score": 0.87, + "content": "S ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 708, + 286, + 723 + ], + "score": 1.0, + "content": "is a subset of the corresponding element in", + "type": "text" + }, + { + "bbox": [ + 286, + 712, + 293, + 719 + ], + "score": 0.87, + "content": "\\boldsymbol { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 708, + 311, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 312, + 711, + 414, + 720 + ], + "score": 0.91, + "content": "\\mathcal { A } = \\mathcal { A } ^ { 1 } \\times \\mathcal { A } ^ { 2 } \\times \\cdot \\cdot \\cdot \\times \\mathcal { A } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 708, + 506, + 723 + ], + "score": 1.0, + "content": ", respectively. For each", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 718, + 506, + 735 + ], + "spans": [ + { + "bbox": [ + 104, + 718, + 131, + 735 + ], + "score": 1.0, + "content": "agent", + "type": "text" + }, + { + "bbox": [ + 131, + 721, + 135, + 731 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 718, + 218, + 735 + ], + "score": 1.0, + "content": ", we provide a set of", + "type": "text" + }, + { + "bbox": [ + 218, + 721, + 230, + 730 + ], + "score": 0.9, + "content": "m ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 718, + 258, + 735 + ], + "score": 1.0, + "content": "skills,", + "type": "text" + }, + { + "bbox": [ + 259, + 721, + 342, + 734 + ], + "score": 0.92, + "content": "\\Pi ^ { i } = \\{ \\pi _ { 1 } ^ { i } , \\ldots , \\pi _ { m ^ { i } } ^ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 718, + 430, + 735 + ], + "score": 1.0, + "content": ". A policy of an agent", + "type": "text" + }, + { + "bbox": [ + 430, + 722, + 435, + 730 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 718, + 506, + 735 + ], + "score": 1.0, + "content": "is represented as", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 81, + 504, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 81, + 504, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 504, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 504, + 198 + ], + "score": 0.961, + "type": "image", + "image_path": "63021c65e6e95a999eba964c17a1759f617f5364c3d76c2f449409b6d5dca506.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 81, + 504, + 120.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 120.0, + 504, + 159.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 159.0, + 504, + 198.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 208, + 506, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 221 + ], + "score": 1.0, + "content": "Figure 2: Our method is composed of two components: a meta policy and a set of agent-specific", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 232 + ], + "score": 1.0, + "content": "primitive policies relevant to task completion. The meta policy selects which primitive skill to run for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "each agent as well as the behavior embedding (i.e. variation in behavior) of the chosen primitive skill.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Each selected primitive skill takes as input the agent observation and the behavior embedding and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 253, + 222, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 222, + 266 + ], + "score": 1.0, + "content": "outputs action for that agent.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 287, + 504, + 309 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 106, + 287, + 505, + 311 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 315, + 506, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 505, + 327 + ], + "score": 1.0, + "content": "To coordinate skills from multiple agents, the skills have to be flexible; hence, a skill can be adjusted", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "to collaborate with other agents’ skills. Maximum entropy policies (Haarnoja et al., 2017; 2018a;b)", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 336, + 507, + 351 + ], + "spans": [ + { + "bbox": [ + 104, + 336, + 507, + 351 + ], + "score": 1.0, + "content": "can learn diverse ways to achieve a goal by maximizing not only reward but also entropy of the policy.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 347, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 104, + 347, + 506, + 362 + ], + "score": 1.0, + "content": "In addition, Eysenbach et al. (2019) proposes to discover diverse skills without reward by maximizing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 360, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 506, + 372 + ], + "score": 1.0, + "content": "entropy as well as mutual information between resulting states and latent representations of skills (i.e.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "skill embeddings). Our method leverages the maximum entropy policy (Haarnoja et al., 2018b) with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 505, + 393 + ], + "score": 1.0, + "content": "the discriminability objective (Eysenbach et al., 2019) to learn a primitive skill with diverse behaviors", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 405 + ], + "score": 1.0, + "content": "conditioned on a controllable skill embedding. This controllable skill embedding will be later used as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 403, + 491, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 491, + 416 + ], + "score": 1.0, + "content": "a behavior embedding for the meta policy to adjust a primitive skill’s behavior for coordination.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 316, + 507, + 416 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 432, + 172, + 444 + ], + "lines": [ + { + "bbox": [ + 104, + 430, + 174, + 448 + ], + "spans": [ + { + "bbox": [ + 104, + 430, + 174, + 448 + ], + "score": 1.0, + "content": "3 METHOD", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 458, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "In this paper, we address the problem of solving cooperative manipulation tasks that require collab-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "score": 1.0, + "content": "oration between multiple end-effectors or agents. Note that we use the terms “end-effector” and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "“agent” interchangeably in this paper. Instead of learning a multi-agent task from scratch (Lowe et al.,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "2017; Gupta et al., 2017; Sunehag et al., 2018; Foerster et al., 2018), modular approaches (Andreas", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "et al., 2017; Frans et al., 2018; Peng et al., 2019) suggest to learn reusable primitive skills and solve", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 506, + 525 + ], + "score": 1.0, + "content": "more complex tasks by recombining these skills. However, concurrent execution of primitive skills of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "multiple agents fails when agents never experienced a combination of skills during the pre-training", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 536, + 345, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 345, + 547 + ], + "score": 1.0, + "content": "stage, or skills require temporal or behavioral coordination.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 457, + 506, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 551, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 505, + 565 + ], + "score": 1.0, + "content": "Therefore, we propose a modular and hierarchical framework that learns to coordinate multiple agents", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 562, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 576 + ], + "score": 1.0, + "content": "with primitive skills to perform a complex task. Moreover, during primitive skill training, we propose", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "to learn a latent behavior embedding, which provides controllability of each primitive skill to the meta", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "policy while coordinating skills. In Section 3.2, we describe our modular framework in detail. Next,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "in Section 3.3, we elaborate how controllable primitive skills can be acquired. Lastly, we describe", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 606, + 394, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 394, + 619 + ], + "score": 1.0, + "content": "how the meta policy learns to coordinate primitive skills in Section 3.4.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 551, + 506, + 619 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 633, + 200, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 201, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 201, + 646 + ], + "score": 1.0, + "content": "3.1 PRELIMINARIES", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 507, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 394, + 667 + ], + "score": 1.0, + "content": "We formulate our problem as a Markov decision process defined by a tuple", + "type": "text" + }, + { + "bbox": [ + 395, + 655, + 469, + 667 + ], + "score": 0.91, + "content": "\\{ S , { \\mathcal { A } } , { \\mathcal { T } } , R , \\rho , \\gamma \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 655, + 507, + 667 + ], + "score": 1.0, + "content": "of states,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "actions, transition probability, reward, initial state distribution, and discount factor. In our formulation,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 255, + 690 + ], + "score": 1.0, + "content": "we assume the environment includes", + "type": "text" + }, + { + "bbox": [ + 256, + 679, + 266, + 687 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "agents. To promote consistency in our terminology, we use", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "superscripts to denote the index of agent and subscripts to denote time or primitive skill index. Hence,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 281, + 711 + ], + "score": 1.0, + "content": "the state space and action space for an agent", + "type": "text" + }, + { + "bbox": [ + 281, + 701, + 285, + 708 + ], + "score": 0.82, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 698, + 372, + 711 + ], + "score": 1.0, + "content": "can be represented as", + "type": "text" + }, + { + "bbox": [ + 372, + 700, + 383, + 708 + ], + "score": 0.89, + "content": "S ^ { \\imath }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 698, + 401, + 711 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 401, + 699, + 413, + 709 + ], + "score": 0.87, + "content": "A ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "where each element of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 709, + 118, + 720 + ], + "score": 0.87, + "content": "S ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 708, + 286, + 723 + ], + "score": 1.0, + "content": "is a subset of the corresponding element in", + "type": "text" + }, + { + "bbox": [ + 286, + 712, + 293, + 719 + ], + "score": 0.87, + "content": "\\boldsymbol { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 708, + 311, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 312, + 711, + 414, + 720 + ], + "score": 0.91, + "content": "\\mathcal { A } = \\mathcal { A } ^ { 1 } \\times \\mathcal { A } ^ { 2 } \\times \\cdot \\cdot \\cdot \\times \\mathcal { A } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 708, + 506, + 723 + ], + "score": 1.0, + "content": ", respectively. For each", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 718, + 506, + 735 + ], + "spans": [ + { + "bbox": [ + 104, + 718, + 131, + 735 + ], + "score": 1.0, + "content": "agent", + "type": "text" + }, + { + "bbox": [ + 131, + 721, + 135, + 731 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 718, + 218, + 735 + ], + "score": 1.0, + "content": ", we provide a set of", + "type": "text" + }, + { + "bbox": [ + 218, + 721, + 230, + 730 + ], + "score": 0.9, + "content": "m ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 718, + 258, + 735 + ], + "score": 1.0, + "content": "skills,", + "type": "text" + }, + { + "bbox": [ + 259, + 721, + 342, + 734 + ], + "score": 0.92, + "content": "\\Pi ^ { i } = \\{ \\pi _ { 1 } ^ { i } , \\ldots , \\pi _ { m ^ { i } } ^ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 718, + 430, + 735 + ], + "score": 1.0, + "content": ". A policy of an agent", + "type": "text" + }, + { + "bbox": [ + 430, + 722, + 435, + 730 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 718, + 506, + 735 + ], + "score": 1.0, + "content": "is represented as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 273, + 508, + 309 + ], + "spans": [ + { + "bbox": [ + 107, + 274, + 171, + 291 + ], + "score": 0.92, + "content": "\\pi _ { c _ { t } ^ { i } } ^ { i } \\left( a _ { t } ^ { i } | s _ { t } ^ { i } \\right) \\in \\Pi ^ { i }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 177, + 273, + 201, + 309 + ], + "score": 1.0, + "content": "where is sam", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 201, + 275, + 210, + 288 + ], + "score": 0.89, + "content": "c _ { t } ^ { i }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 211, + 273, + 244, + 309 + ], + "score": 1.0, + "content": "is a skiled from", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 250, + 273, + 274, + 309 + ], + "score": 1.0, + "content": "index, , and t", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 274, + 275, + 305, + 288 + ], + "score": 0.93, + "content": "s _ { t } ^ { i } \\in S ^ { i }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 306, + 273, + 362, + 309 + ], + "score": 1.0, + "content": "is a state, and agents take act", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 363, + 275, + 396, + 288 + ], + "score": 0.93, + "content": "a _ { t } ^ { i } \\in \\mathcal A ^ { i }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 441, + 273, + 498, + 309 + ], + "score": 1.0, + "content": "action at time sampled from", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 498, + 277, + 503, + 285 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 503, + 273, + 508, + 309 + ], + "score": 1.0, + "content": ".", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 102, + 289, + 504, + 318 + ], + "spans": [ + { + "bbox": [ + 102, + 296, + 179, + 318 + ], + "score": 1.0, + "content": "composite policy", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 179, + 301, + 504, + 317 + ], + "score": 0.9, + "content": "\\pi ( a _ { t } ^ { 1 } , \\bar { a } _ { t } ^ { 2 } , \\dots , a _ { t } ^ { N } | s _ { t } , c _ { t } ^ { 1 } , c _ { t } ^ { 2 } , \\dots , \\bar { c } _ { t } ^ { N } ) = ( \\pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( a _ { t } ^ { 1 } | s _ { t } ) , \\pi _ { c _ { t } ^ { 2 } } ^ { 2 } ( a _ { t } ^ { 2 } | s _ { t } ) , \\dots , \\pi _ { c _ { t } ^ { N } } ^ { N } ( a _ { t } ^ { N } | s _ { t } ) )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 244, + 293, + 250, + 302 + ], + "score": 0.71, + "content": "\\rho", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 292, + 291, + 303, + 300 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 380, + 289, + 440, + 302 + ], + "score": 0.9, + "content": "a _ { t } ^ { 1 } , a _ { t } ^ { 2 } , \\ldots , a _ { t } ^ { N }", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 154, + 315, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 154, + 315, + 184, + 348 + ], + "score": 1.0, + "content": "a singl, where", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 193, + 315, + 222, + 348 + ], + "score": 1.0, + "content": "reward is the e", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 222, + 317, + 231, + 326 + ], + "score": 0.83, + "content": "r _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 231, + 315, + 483, + 348 + ], + "score": 1.0, + "content": ". The performance is evaluated based on a discounted return sode horizon.", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 484, + 316, + 505, + 327 + ], + "score": 0.81, + "content": "R =", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 326, + 193, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 154, + 341 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { t = 0 } ^ { T - 1 } \\gamma ^ { t } r _ { t }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 184, + 328, + 193, + 338 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 38, + "bbox_fs": [ + 104, + 655, + 507, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 78, + 503, + 164 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 78, + 503, + 164 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 78, + 503, + 164 + ], + "spans": [ + { + "bbox": [ + 112, + 78, + 503, + 164 + ], + "score": 0.971, + "type": "image", + "image_path": "852e7eff92afaaa227e4f7f1a58fdeda7147bb405d07ee79a206747f65017adf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 78, + 503, + 106.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 106.66666666666667, + 503, + 135.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 135.33333333333334, + 503, + 164.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 172, + 506, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "score": 1.0, + "content": "Figure 3: Different multi-agent architectures. (a) The vanilla RL method considers all agents as a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 183, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 506, + 196 + ], + "score": 1.0, + "content": "monolithic agent; thus a single policy takes the full observation as input and outputs the full action.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 315, + 207 + ], + "score": 1.0, + "content": "(b) The multi-agent RL method (MARL) consists of", + "type": "text" + }, + { + "bbox": [ + 316, + 195, + 326, + 204 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "policies that operate on the observations and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 205, + 504, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 397, + 217 + ], + "score": 1.0, + "content": "actions of corresponding agents. (c) The modular network consists of", + "type": "text" + }, + { + "bbox": [ + 397, + 205, + 407, + 215 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 205, + 493, + 217 + ], + "score": 1.0, + "content": "sets of skills for the", + "type": "text" + }, + { + "bbox": [ + 493, + 205, + 504, + 215 + ], + "score": 0.74, + "content": "N", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 216, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 506, + 228 + ], + "score": 1.0, + "content": "agents trained in isolation and a meta policy that selects a skill for each agent. (d-f) The RL, MARL,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 226, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 506, + 241 + ], + "score": 1.0, + "content": "and modular network methods augmented with skill behavior diversification (SBD) has a meta policy", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 239, + 358, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 293, + 250 + ], + "score": 1.0, + "content": "that outputs a skill behavior embedding vector", + "type": "text" + }, + { + "bbox": [ + 294, + 240, + 300, + 248 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 239, + 358, + 250 + ], + "score": 1.0, + "content": "for each skill.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 274, + 506, + 340 + ], + "lines": [ + { + "bbox": [ + 107, + 273, + 508, + 309 + ], + "spans": [ + { + "bbox": [ + 107, + 274, + 171, + 291 + ], + "score": 0.92, + "content": "\\pi _ { c _ { t } ^ { i } } ^ { i } \\left( a _ { t } ^ { i } | s _ { t } ^ { i } \\right) \\in \\Pi ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 273, + 201, + 309 + ], + "score": 1.0, + "content": "where is sam", + "type": "text" + }, + { + "bbox": [ + 201, + 275, + 210, + 288 + ], + "score": 0.89, + "content": "c _ { t } ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 273, + 244, + 309 + ], + "score": 1.0, + "content": "is a skiled from", + "type": "text" + }, + { + "bbox": [ + 250, + 273, + 274, + 309 + ], + "score": 1.0, + "content": "index, , and t", + "type": "text" + }, + { + "bbox": [ + 274, + 275, + 305, + 288 + ], + "score": 0.93, + "content": "s _ { t } ^ { i } \\in S ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 273, + 362, + 309 + ], + "score": 1.0, + "content": "is a state, and agents take act", + "type": "text" + }, + { + "bbox": [ + 363, + 275, + 396, + 288 + ], + "score": 0.93, + "content": "a _ { t } ^ { i } \\in \\mathcal A ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 273, + 498, + 309 + ], + "score": 1.0, + "content": "action at time sampled from", + "type": "text" + }, + { + "bbox": [ + 498, + 277, + 503, + 285 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 273, + 508, + 309 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 102, + 289, + 504, + 318 + ], + "spans": [ + { + "bbox": [ + 102, + 296, + 179, + 318 + ], + "score": 1.0, + "content": "composite policy", + "type": "text" + }, + { + "bbox": [ + 179, + 301, + 504, + 317 + ], + "score": 0.9, + "content": "\\pi ( a _ { t } ^ { 1 } , \\bar { a } _ { t } ^ { 2 } , \\dots , a _ { t } ^ { N } | s _ { t } , c _ { t } ^ { 1 } , c _ { t } ^ { 2 } , \\dots , \\bar { c } _ { t } ^ { N } ) = ( \\pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( a _ { t } ^ { 1 } | s _ { t } ) , \\pi _ { c _ { t } ^ { 2 } } ^ { 2 } ( a _ { t } ^ { 2 } | s _ { t } ) , \\dots , \\pi _ { c _ { t } ^ { N } } ^ { N } ( a _ { t } ^ { N } | s _ { t } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 293, + 250, + 302 + ], + "score": 0.71, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 291, + 303, + 300 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 289, + 440, + 302 + ], + "score": 0.9, + "content": "a _ { t } ^ { 1 } , a _ { t } ^ { 2 } , \\ldots , a _ { t } ^ { N }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 154, + 315, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 154, + 315, + 184, + 348 + ], + "score": 1.0, + "content": "a singl, where", + "type": "text" + }, + { + "bbox": [ + 193, + 315, + 222, + 348 + ], + "score": 1.0, + "content": "reward is the e", + "type": "text" + }, + { + "bbox": [ + 222, + 317, + 231, + 326 + ], + "score": 0.83, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 315, + 483, + 348 + ], + "score": 1.0, + "content": ". The performance is evaluated based on a discounted return sode horizon.", + "type": "text" + }, + { + "bbox": [ + 484, + 316, + 505, + 327 + ], + "score": 0.81, + "content": "R =", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 326, + 193, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 154, + 341 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { t = 0 } ^ { T - 1 } \\gamma ^ { t } r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 328, + 193, + 338 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 107, + 358, + 236, + 369 + ], + "lines": [ + { + "bbox": [ + 106, + 358, + 237, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 237, + 370 + ], + "score": 1.0, + "content": "3.2 MODULAR FRAMEWORK", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 104, + 379, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 441, + 394 + ], + "score": 1.0, + "content": "As illustrated in Figure 2, our model is composed of two components: a meta policy", + "type": "text" + }, + { + "bbox": [ + 441, + 383, + 466, + 392 + ], + "score": 0.88, + "content": "\\pi _ { m e t a }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 379, + 506, + 394 + ], + "score": 1.0, + "content": "and a set", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 189, + 405 + ], + "score": 1.0, + "content": "of primitive skills of", + "type": "text" + }, + { + "bbox": [ + 189, + 392, + 199, + 402 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 390, + 228, + 405 + ], + "score": 1.0, + "content": "agents", + "type": "text" + }, + { + "bbox": [ + 228, + 391, + 278, + 403 + ], + "score": 0.92, + "content": "\\Pi ^ { 1 } , \\dots , \\Pi ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 390, + 398, + 405 + ], + "score": 1.0, + "content": ". Note that each primitive skill", + "type": "text" + }, + { + "bbox": [ + 399, + 391, + 435, + 404 + ], + "score": 0.92, + "content": "\\pi _ { c ^ { i } } ^ { i } \\in \\Pi ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 390, + 506, + 405 + ], + "score": 1.0, + "content": "contains variants", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 244, + 417 + ], + "score": 1.0, + "content": "of behaviors parameterized by an", + "type": "text" + }, + { + "bbox": [ + 244, + 404, + 258, + 415 + ], + "score": 0.89, + "content": "N _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 403, + 423, + 417 + ], + "score": 1.0, + "content": "-dimensional latent behavior embedding", + "type": "text" + }, + { + "bbox": [ + 423, + 404, + 433, + 414 + ], + "score": 0.85, + "content": "z ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "(see Section 3.3).", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 506, + 428 + ], + "score": 1.0, + "content": "The meta policy selects a skill to execute for each agent, rather than selecting one primitive skill for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "the entire multi-agent system to execute. Also, we give the meta policy the capability to select which", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "variant of the skill to execute (see Section 3.4). Then, the chosen primitive skills are simultaneously", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 447, + 225, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 158, + 461 + ], + "score": 1.0, + "content": "executed for", + "type": "text" + }, + { + "bbox": [ + 158, + 448, + 178, + 459 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 447, + 225, + 461 + ], + "score": 1.0, + "content": "time steps.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "The concurrent execution of multiple skills often leads to undesired results and therefore requires", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 476, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 488 + ], + "score": 1.0, + "content": "coordination between the skills. For example, naively placing a block in the left hand to a container", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "being moved by the right hand can cause collision between the two robot arms. The arms can avoid", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "collision while performing the skills by properly adjusting their skill behaviors (e.g. the left arm", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "score": 1.0, + "content": "leaning to the left side while placing the block and the right arm leaning to the right side while", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "pushing the container) as shown in Figure 1. In our method, the meta policy learns to coordinate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 531, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 544 + ], + "score": 1.0, + "content": "multiple agents’ skills by manipulating the behavior embeddings (i.e. selecting a proper behavior", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 541, + 257, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 257, + 554 + ], + "score": 1.0, + "content": "from diverse behaviors of each skill).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 572, + 443, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 445, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 445, + 583 + ], + "score": 1.0, + "content": "3.3 TRAINING AGENT-SPECIFIC PRIMITIVE SKILLS WITH DIVERSE BEHAVIORS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "To adjust a primitive skill to collaborate with other agents’ skills in a new environment, the skill", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "needs to support variations of skill behaviors when executed at a given state. Moreover, a behavioral", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "variation of a skill should be controllable by the meta policy for skill coordination. In order to make", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 431, + 640 + ], + "score": 1.0, + "content": "our primitive skill policies generate diverse behaviors controlled by a latent vector", + "type": "text" + }, + { + "bbox": [ + 431, + 629, + 438, + 637 + ], + "score": 0.7, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 627, + 505, + 640 + ], + "score": 1.0, + "content": ", we leverage the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 638, + 486, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 486, + 651 + ], + "score": 1.0, + "content": "entropy and mutual information maximization objective introduced in Eysenbach et al. (2019).", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 297, + 667 + ], + "score": 1.0, + "content": "More specifically, a primitive policy of an agent", + "type": "text" + }, + { + "bbox": [ + 297, + 657, + 301, + 664 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 655, + 371, + 667 + ], + "score": 1.0, + "content": "outputs an action", + "type": "text" + }, + { + "bbox": [ + 372, + 655, + 398, + 665 + ], + "score": 0.9, + "content": "a \\in { \\mathcal { A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "conditioned on the current", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 127, + 678 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 127, + 666, + 152, + 676 + ], + "score": 0.89, + "content": "s \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 666, + 285, + 678 + ], + "score": 1.0, + "content": "and a latent behavior embedding", + "type": "text" + }, + { + "bbox": [ + 285, + 667, + 322, + 678 + ], + "score": 0.92, + "content": "z \\sim p ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 666, + 437, + 678 + ], + "score": 1.0, + "content": ", where the prior distribution", + "type": "text" + }, + { + "bbox": [ + 437, + 666, + 456, + 678 + ], + "score": 0.92, + "content": "p ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "is Gaussian", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 170, + 690 + ], + "score": 1.0, + "content": "(we omit agent", + "type": "text" + }, + { + "bbox": [ + 170, + 678, + 175, + 687 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "in this section for the simplicity of notations). Diverse behaviors conditioned on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 177, + 700 + ], + "score": 1.0, + "content": "a random sample", + "type": "text" + }, + { + "bbox": [ + 178, + 690, + 185, + 698 + ], + "score": 0.76, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "can be achieved by maximizing the mutual information between behaviors and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 131, + 711 + ], + "score": 1.0, + "content": "states", + "type": "text" + }, + { + "bbox": [ + 131, + 699, + 170, + 710 + ], + "score": 0.89, + "content": "M I ( s , z )", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 699, + 505, + 711 + ], + "score": 1.0, + "content": ", while minimizing the mutual information between behaviors and actions given the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 127, + 722 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 128, + 711, + 174, + 721 + ], + "score": 0.9, + "content": "M I ( a , z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 710, + 388, + 722 + ], + "score": 1.0, + "content": ", together with maximizing the entropy of the policy", + "type": "text" + }, + { + "bbox": [ + 388, + 710, + 417, + 722 + ], + "score": 0.93, + "content": "\\mathcal { H } ( a | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "to encourage diverse", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "behaviors. The objective can be written as follows (we refer the readers to Eysenbach et al. (2019)", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 112, + 78, + 503, + 164 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 78, + 503, + 164 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 78, + 503, + 164 + ], + "spans": [ + { + "bbox": [ + 112, + 78, + 503, + 164 + ], + "score": 0.971, + "type": "image", + "image_path": "852e7eff92afaaa227e4f7f1a58fdeda7147bb405d07ee79a206747f65017adf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 78, + 503, + 106.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 106.66666666666667, + 503, + 135.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 135.33333333333334, + 503, + 164.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 172, + 506, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "score": 1.0, + "content": "Figure 3: Different multi-agent architectures. (a) The vanilla RL method considers all agents as a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 183, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 506, + 196 + ], + "score": 1.0, + "content": "monolithic agent; thus a single policy takes the full observation as input and outputs the full action.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 315, + 207 + ], + "score": 1.0, + "content": "(b) The multi-agent RL method (MARL) consists of", + "type": "text" + }, + { + "bbox": [ + 316, + 195, + 326, + 204 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "policies that operate on the observations and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 205, + 504, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 397, + 217 + ], + "score": 1.0, + "content": "actions of corresponding agents. (c) The modular network consists of", + "type": "text" + }, + { + "bbox": [ + 397, + 205, + 407, + 215 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 205, + 493, + 217 + ], + "score": 1.0, + "content": "sets of skills for the", + "type": "text" + }, + { + "bbox": [ + 493, + 205, + 504, + 215 + ], + "score": 0.74, + "content": "N", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 216, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 506, + 228 + ], + "score": 1.0, + "content": "agents trained in isolation and a meta policy that selects a skill for each agent. (d-f) The RL, MARL,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 226, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 506, + 241 + ], + "score": 1.0, + "content": "and modular network methods augmented with skill behavior diversification (SBD) has a meta policy", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 239, + 358, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 293, + 250 + ], + "score": 1.0, + "content": "that outputs a skill behavior embedding vector", + "type": "text" + }, + { + "bbox": [ + 294, + 240, + 300, + 248 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 239, + 358, + 250 + ], + "score": 1.0, + "content": "for each skill.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 274, + 506, + 340 + ], + "lines": [], + "index": 11.5, + "bbox_fs": [ + 102, + 273, + 508, + 348 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 358, + 236, + 369 + ], + "lines": [ + { + "bbox": [ + 106, + 358, + 237, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 237, + 370 + ], + "score": 1.0, + "content": "3.2 MODULAR FRAMEWORK", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 104, + 379, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 379, + 441, + 394 + ], + "score": 1.0, + "content": "As illustrated in Figure 2, our model is composed of two components: a meta policy", + "type": "text" + }, + { + "bbox": [ + 441, + 383, + 466, + 392 + ], + "score": 0.88, + "content": "\\pi _ { m e t a }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 379, + 506, + 394 + ], + "score": 1.0, + "content": "and a set", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 189, + 405 + ], + "score": 1.0, + "content": "of primitive skills of", + "type": "text" + }, + { + "bbox": [ + 189, + 392, + 199, + 402 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 390, + 228, + 405 + ], + "score": 1.0, + "content": "agents", + "type": "text" + }, + { + "bbox": [ + 228, + 391, + 278, + 403 + ], + "score": 0.92, + "content": "\\Pi ^ { 1 } , \\dots , \\Pi ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 390, + 398, + 405 + ], + "score": 1.0, + "content": ". Note that each primitive skill", + "type": "text" + }, + { + "bbox": [ + 399, + 391, + 435, + 404 + ], + "score": 0.92, + "content": "\\pi _ { c ^ { i } } ^ { i } \\in \\Pi ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 390, + 506, + 405 + ], + "score": 1.0, + "content": "contains variants", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 244, + 417 + ], + "score": 1.0, + "content": "of behaviors parameterized by an", + "type": "text" + }, + { + "bbox": [ + 244, + 404, + 258, + 415 + ], + "score": 0.89, + "content": "N _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 403, + 423, + 417 + ], + "score": 1.0, + "content": "-dimensional latent behavior embedding", + "type": "text" + }, + { + "bbox": [ + 423, + 404, + 433, + 414 + ], + "score": 0.85, + "content": "z ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "(see Section 3.3).", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 506, + 428 + ], + "score": 1.0, + "content": "The meta policy selects a skill to execute for each agent, rather than selecting one primitive skill for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "the entire multi-agent system to execute. Also, we give the meta policy the capability to select which", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "variant of the skill to execute (see Section 3.4). Then, the chosen primitive skills are simultaneously", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 447, + 225, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 158, + 461 + ], + "score": 1.0, + "content": "executed for", + "type": "text" + }, + { + "bbox": [ + 158, + 448, + 178, + 459 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 447, + 225, + 461 + ], + "score": 1.0, + "content": "time steps.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 379, + 506, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 477 + ], + "score": 1.0, + "content": "The concurrent execution of multiple skills often leads to undesired results and therefore requires", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 476, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 488 + ], + "score": 1.0, + "content": "coordination between the skills. For example, naively placing a block in the left hand to a container", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "being moved by the right hand can cause collision between the two robot arms. The arms can avoid", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "collision while performing the skills by properly adjusting their skill behaviors (e.g. the left arm", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "score": 1.0, + "content": "leaning to the left side while placing the block and the right arm leaning to the right side while", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "pushing the container) as shown in Figure 1. In our method, the meta policy learns to coordinate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 531, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 544 + ], + "score": 1.0, + "content": "multiple agents’ skills by manipulating the behavior embeddings (i.e. selecting a proper behavior", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 541, + 257, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 257, + 554 + ], + "score": 1.0, + "content": "from diverse behaviors of each skill).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 464, + 506, + 554 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 572, + 443, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 445, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 445, + 583 + ], + "score": 1.0, + "content": "3.3 TRAINING AGENT-SPECIFIC PRIMITIVE SKILLS WITH DIVERSE BEHAVIORS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "To adjust a primitive skill to collaborate with other agents’ skills in a new environment, the skill", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "needs to support variations of skill behaviors when executed at a given state. Moreover, a behavioral", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "variation of a skill should be controllable by the meta policy for skill coordination. In order to make", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 431, + 640 + ], + "score": 1.0, + "content": "our primitive skill policies generate diverse behaviors controlled by a latent vector", + "type": "text" + }, + { + "bbox": [ + 431, + 629, + 438, + 637 + ], + "score": 0.7, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 627, + 505, + 640 + ], + "score": 1.0, + "content": ", we leverage the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 638, + 486, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 486, + 651 + ], + "score": 1.0, + "content": "entropy and mutual information maximization objective introduced in Eysenbach et al. (2019).", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 594, + 505, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 297, + 667 + ], + "score": 1.0, + "content": "More specifically, a primitive policy of an agent", + "type": "text" + }, + { + "bbox": [ + 297, + 657, + 301, + 664 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 655, + 371, + 667 + ], + "score": 1.0, + "content": "outputs an action", + "type": "text" + }, + { + "bbox": [ + 372, + 655, + 398, + 665 + ], + "score": 0.9, + "content": "a \\in { \\mathcal { A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "conditioned on the current", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 127, + 678 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 127, + 666, + 152, + 676 + ], + "score": 0.89, + "content": "s \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 666, + 285, + 678 + ], + "score": 1.0, + "content": "and a latent behavior embedding", + "type": "text" + }, + { + "bbox": [ + 285, + 667, + 322, + 678 + ], + "score": 0.92, + "content": "z \\sim p ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 666, + 437, + 678 + ], + "score": 1.0, + "content": ", where the prior distribution", + "type": "text" + }, + { + "bbox": [ + 437, + 666, + 456, + 678 + ], + "score": 0.92, + "content": "p ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "is Gaussian", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 170, + 690 + ], + "score": 1.0, + "content": "(we omit agent", + "type": "text" + }, + { + "bbox": [ + 170, + 678, + 175, + 687 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "in this section for the simplicity of notations). Diverse behaviors conditioned on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 177, + 700 + ], + "score": 1.0, + "content": "a random sample", + "type": "text" + }, + { + "bbox": [ + 178, + 690, + 185, + 698 + ], + "score": 0.76, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "can be achieved by maximizing the mutual information between behaviors and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 131, + 711 + ], + "score": 1.0, + "content": "states", + "type": "text" + }, + { + "bbox": [ + 131, + 699, + 170, + 710 + ], + "score": 0.89, + "content": "M I ( s , z )", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 699, + 505, + 711 + ], + "score": 1.0, + "content": ", while minimizing the mutual information between behaviors and actions given the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 127, + 722 + ], + "score": 1.0, + "content": "state", + "type": "text" + }, + { + "bbox": [ + 128, + 711, + 174, + 721 + ], + "score": 0.9, + "content": "M I ( a , z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 710, + 388, + 722 + ], + "score": 1.0, + "content": ", together with maximizing the entropy of the policy", + "type": "text" + }, + { + "bbox": [ + 388, + 710, + 417, + 722 + ], + "score": 0.93, + "content": "\\mathcal { H } ( a | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "to encourage diverse", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "behaviors. The objective can be written as follows (we refer the readers to Eysenbach et al. (2019)", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 655, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 84, + 203, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 205, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 205, + 98 + ], + "score": 1.0, + "content": "Algorithm 1 ROLLOUT", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 109, + 99, + 487, + 247 + ], + "lines": [ + { + "bbox": [ + 109, + 96, + 474, + 113 + ], + "spans": [ + { + "bbox": [ + 109, + 96, + 204, + 113 + ], + "score": 1.0, + "content": "1: Input: Meta policy", + "type": "text" + }, + { + "bbox": [ + 204, + 101, + 229, + 110 + ], + "score": 0.86, + "content": "\\pi _ { m e t a }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 96, + 333, + 113 + ], + "score": 1.0, + "content": ", sets of primitive policies", + "type": "text" + }, + { + "bbox": [ + 334, + 98, + 379, + 110 + ], + "score": 0.92, + "content": "\\Pi ^ { 1 } , . . . , \\Pi ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 96, + 454, + 113 + ], + "score": 1.0, + "content": ", and meta horizon", + "type": "text" + }, + { + "bbox": [ + 455, + 100, + 474, + 110 + ], + "score": 0.87, + "content": "T _ { l o w }", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 108, + 337, + 123 + ], + "spans": [ + { + "bbox": [ + 111, + 108, + 205, + 123 + ], + "score": 1.0, + "content": "2: Initialize an episode", + "type": "text" + }, + { + "bbox": [ + 206, + 111, + 232, + 120 + ], + "score": 0.89, + "content": "t \\gets 0", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 108, + 327, + 123 + ], + "score": 1.0, + "content": "and receive initial state", + "type": "text" + }, + { + "bbox": [ + 327, + 112, + 337, + 122 + ], + "score": 0.84, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 121, + 264, + 132 + ], + "spans": [ + { + "bbox": [ + 111, + 121, + 264, + 132 + ], + "score": 1.0, + "content": "3: while episode is not terminated do", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 130, + 485, + 146 + ], + "spans": [ + { + "bbox": [ + 110, + 132, + 121, + 143 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 131, + 130, + 323, + 146 + ], + "score": 1.0, + "content": "Sample skill indexes and behavior embeddings", + "type": "text" + }, + { + "bbox": [ + 323, + 131, + 485, + 144 + ], + "score": 0.92, + "content": "( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } ) \\sim \\pi _ { m e t a } ( s _ { t } )", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 143, + 161, + 154 + ], + "spans": [ + { + "bbox": [ + 110, + 144, + 122, + 154 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 133, + 143, + 161, + 154 + ], + "score": 0.66, + "content": "\\tau 0", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 153, + 331, + 165 + ], + "spans": [ + { + "bbox": [ + 110, + 154, + 122, + 165 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 133, + 153, + 158, + 165 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 159, + 154, + 197, + 164 + ], + "score": 0.89, + "content": "\\tau < T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 153, + 331, + 165 + ], + "score": 1.0, + "content": "and episode is not terminated do", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 164, + 403, + 180 + ], + "spans": [ + { + "bbox": [ + 110, + 165, + 122, + 178 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 142, + 164, + 403, + 180 + ], + "score": 0.77, + "content": "\\boldsymbol a _ { t + \\tau } = ( a _ { t + \\tau } ^ { 1 } , \\dots , \\dot { a _ { t + \\tau } ^ { N } } ) \\sim ( \\pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( s _ { t + \\tau } , z _ { t } ^ { 1 } ) , \\dots , \\pi _ { c _ { t } ^ { N } } ^ { N } ( s _ { t + \\tau } , z _ { t } ^ { N } ) )", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 177, + 283, + 202 + ], + "spans": [ + { + "bbox": [ + 110, + 179, + 121, + 190 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 142, + 177, + 283, + 202 + ], + "score": 0.38, + "content": "\\begin{array} { r l } & { s _ { t + \\tau + 1 } , r _ { t + \\tau } \\gets \\mathrm { E N V } ( s _ { t + \\tau } , a _ { t + \\tau } ) } \\\\ & { \\tau \\gets \\tau + 1 } \\end{array}", + "type": "inline_equation", + "image_path": "416eef5398b508d7f4182c92239826f0f1381cebb57c910d7fa2e5f5a9cd6011.jpg" + } + ], + "index": 8 + }, + { + "bbox": [ + 109, + 190, + 122, + 202 + ], + "spans": [ + { + "bbox": [ + 109, + 190, + 122, + 202 + ], + "score": 1.0, + "content": "9:", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 201, + 177, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 122, + 213 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 133, + 201, + 177, + 212 + ], + "score": 1.0, + "content": "end while", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 210, + 473, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 122, + 224 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 131, + 210, + 199, + 225 + ], + "score": 1.0, + "content": "Add a transition", + "type": "text" + }, + { + "bbox": [ + 200, + 211, + 383, + 224 + ], + "score": 0.93, + "content": "s _ { t } , ( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } ) , s _ { t + \\tau } , r _ { t : t + \\tau - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 210, + 464, + 225 + ], + "score": 1.0, + "content": "to the rollout buffer", + "type": "text" + }, + { + "bbox": [ + 464, + 213, + 473, + 222 + ], + "score": 0.76, + "content": "\\boldsymbol { B }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 175, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 123, + 235 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 133, + 223, + 175, + 234 + ], + "score": 0.51, + "content": "t \\gets t + \\tau", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 234, + 167, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 234, + 167, + 245 + ], + "score": 1.0, + "content": "13: end while", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 168, + 280 + ], + "lines": [ + { + "bbox": [ + 105, + 267, + 169, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 169, + 281 + ], + "score": 1.0, + "content": "for derivation):", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 284, + 455, + 331 + ], + "lines": [ + { + "bbox": [ + 155, + 284, + 455, + 331 + ], + "spans": [ + { + "bbox": [ + 155, + 284, + 455, + 331 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathcal { F } ( \\theta ) = M I ( s , z ) - M I ( a , z \\vert s ) + \\mathcal { H } ( a \\vert s ) = \\mathcal { H } ( a \\vert s , z ) - \\mathcal { H } ( z \\vert s ) + \\mathcal { H } ( z ) } \\\\ & { \\quad \\quad = \\mathcal { H } ( a \\vert s , z ) + \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log p ( z \\vert s ) ] - \\mathbb { E } _ { z \\sim p ( z ) } [ \\log p ( z ) ] } \\\\ & { \\quad \\quad \\geq \\mathcal { H } ( a \\vert s , z ) + \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log q _ { \\phi } ( z \\vert s ) - \\log p ( z ) ] , } \\end{array}", + "type": "interline_equation", + "image_path": "d92048ef0214b9413b930951da396cad2d15a7c9b9c6cb3e6d96816918b24d97.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 155, + 284, + 455, + 299.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 155, + 299.6666666666667, + 455, + 315.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 155, + 315.33333333333337, + 455, + 331.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 334, + 406, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 406, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 235, + 349 + ], + "score": 1.0, + "content": "where the learned discriminator", + "type": "text" + }, + { + "bbox": [ + 235, + 335, + 266, + 347 + ], + "score": 0.93, + "content": "q _ { \\phi } ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 333, + 376, + 349 + ], + "score": 1.0, + "content": "approximates the posterior", + "type": "text" + }, + { + "bbox": [ + 376, + 334, + 402, + 347 + ], + "score": 0.92, + "content": "p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 333, + 406, + 349 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 105, + 351, + 503, + 373 + ], + "lines": [ + { + "bbox": [ + 104, + 350, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 104, + 350, + 505, + 365 + ], + "score": 1.0, + "content": "To achieve a primitive skill with diverse behaviors, we augment Equation (3) to the environment", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 362, + 140, + 375 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 140, + 375 + ], + "score": 1.0, + "content": "reward:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 180, + 373, + 430, + 387 + ], + "lines": [ + { + "bbox": [ + 180, + 373, + 430, + 387 + ], + "spans": [ + { + "bbox": [ + 180, + 373, + 430, + 387 + ], + "score": 0.89, + "content": "r _ { t } + \\lambda _ { 1 } \\mathcal { H } ( a | s , z ) + \\lambda _ { 2 } \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log q _ { \\phi } ( z | s ) - \\log p ( z ) ] ,", + "type": "interline_equation", + "image_path": "7dcfa1f10a0ed872bfb88b0fa8018a0f0cffb519a0831b833fece71eb0731dbc.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 180, + 373, + 430, + 387 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 388, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 133, + 401 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 389, + 145, + 399 + ], + "score": 0.88, + "content": "\\lambda _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 388, + 267, + 401 + ], + "score": 1.0, + "content": "is the entropy coefficient and", + "type": "text" + }, + { + "bbox": [ + 268, + 389, + 279, + 399 + ], + "score": 0.88, + "content": "\\lambda _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "is the diversity coefficient which corresponds identifia-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "bility of behaviors. Maximizing Equation (3) encourages multi-modal exploration strategies while", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 411, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 200, + 423 + ], + "score": 1.0, + "content": "maximizing the reward", + "type": "text" + }, + { + "bbox": [ + 200, + 412, + 209, + 421 + ], + "score": 0.85, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 411, + 506, + 423 + ], + "score": 1.0, + "content": "forces to achieve its own goal. Moreover, by maximizing identifiability of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 421, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 218, + 433 + ], + "score": 1.0, + "content": "behaviors, the latent vector", + "type": "text" + }, + { + "bbox": [ + 218, + 424, + 225, + 431 + ], + "score": 0.76, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 421, + 506, + 433 + ], + "score": 1.0, + "content": ", named behavior embedding, can represent a variation of the learned", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "policy and thus can be used to control the behavior of the policy. For example, when training a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "score": 1.0, + "content": "robot to move an object, a policy learns to move the object quickly as well as slowly, and these", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 296, + 467 + ], + "score": 1.0, + "content": "diverse behaviors map to different latent vectors", + "type": "text" + }, + { + "bbox": [ + 296, + 456, + 302, + 464 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 454, + 505, + 467 + ], + "score": 1.0, + "content": ". We empirically show that the policies with diverse", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 465, + 425, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 425, + 478 + ], + "score": 1.0, + "content": "behaviors achieve better compositionality with other agents in our experiments.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 107, + 491, + 350, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 351, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 351, + 502 + ], + "score": 1.0, + "content": "3.4 COMPOSING PRIMITIVE SKILLS WITH META POLICY", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 509, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 103, + 508, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 103, + 508, + 222, + 526 + ], + "score": 1.0, + "content": "We denote the meta policy as", + "type": "text" + }, + { + "bbox": [ + 222, + 510, + 358, + 523 + ], + "score": 0.89, + "content": "\\pi _ { m e t a } ( c ^ { 1 } , \\ldots , c ^ { N } , z ^ { 1 } , \\ldots , z ^ { N } | s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 508, + 388, + 526 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 388, + 510, + 436, + 523 + ], + "score": 0.93, + "content": "c ^ { i } \\in [ 1 , m ^ { i } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 508, + 505, + 526 + ], + "score": 1.0, + "content": "represents a skill", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 519, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 104, + 519, + 181, + 536 + ], + "score": 1.0, + "content": "index of an agent", + "type": "text" + }, + { + "bbox": [ + 182, + 522, + 224, + 534 + ], + "score": 0.93, + "content": "i \\in [ 1 , N ]", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 519, + 244, + 536 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 522, + 285, + 532 + ], + "score": 0.85, + "content": "z ^ { i } \\in \\mathbb { R } ^ { N _ { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 519, + 506, + 536 + ], + "score": 1.0, + "content": "represents a behavior embedding of the skill. Every", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 531, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 127, + 544 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 531, + 353, + 546 + ], + "score": 1.0, + "content": "time steps, the meta policy chooses one primitive skill", + "type": "text" + }, + { + "bbox": [ + 354, + 532, + 392, + 545 + ], + "score": 0.92, + "content": "\\pi _ { c ^ { i } } ^ { i } \\in \\Pi ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 531, + 455, + 546 + ], + "score": 1.0, + "content": "for each agent", + "type": "text" + }, + { + "bbox": [ + 455, + 533, + 460, + 542 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 531, + 506, + 546 + ], + "score": 1.0, + "content": ". Also, the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 542, + 507, + 561 + ], + "spans": [ + { + "bbox": [ + 104, + 542, + 343, + 561 + ], + "score": 1.0, + "content": "meta policy outputs a set of latent behavior embeddings", + "type": "text" + }, + { + "bbox": [ + 344, + 545, + 410, + 558 + ], + "score": 0.9, + "content": "( z ^ { 1 } , z ^ { 2 } , \\dots , z ^ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 542, + 507, + 561 + ], + "score": 1.0, + "content": "and feeds them to the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 101, + 552, + 504, + 574 + ], + "spans": [ + { + "bbox": [ + 101, + 552, + 210, + 574 + ], + "score": 1.0, + "content": "corresponding skills (i.e.", + "type": "text" + }, + { + "bbox": [ + 211, + 556, + 264, + 569 + ], + "score": 0.93, + "content": "\\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 552, + 306, + 574 + ], + "score": 1.0, + "content": "for agent", + "type": "text" + }, + { + "bbox": [ + 306, + 558, + 311, + 566 + ], + "score": 0.54, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 552, + 441, + 574 + ], + "score": 1.0, + "content": "). Once a set of primitive skills", + "type": "text" + }, + { + "bbox": [ + 441, + 556, + 504, + 569 + ], + "score": 0.93, + "content": "\\{ \\pi _ { c ^ { 1 } } ^ { 1 } , \\ldots , \\pi _ { c ^ { N } } ^ { N } \\}", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 375, + 582 + ], + "score": 1.0, + "content": "are chosen to be executed, each primitive skill generates an action", + "type": "text" + }, + { + "bbox": [ + 375, + 568, + 451, + 582 + ], + "score": 0.94, + "content": "a ^ { i } \\sim \\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "based on the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 581, + 467, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 158, + 594 + ], + "score": 1.0, + "content": "current state", + "type": "text" + }, + { + "bbox": [ + 158, + 581, + 167, + 591 + ], + "score": 0.86, + "content": "s ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 581, + 252, + 594 + ], + "score": 1.0, + "content": "and the latent vector", + "type": "text" + }, + { + "bbox": [ + 252, + 581, + 261, + 591 + ], + "score": 0.86, + "content": "z ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 581, + 467, + 594 + ], + "score": 1.0, + "content": ". Algorithm 1 illustrates the overall rollout process.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 598, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 611 + ], + "score": 1.0, + "content": "Since there are a finite number of skills for each agent to execute, the meta action space for each agent", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 607, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 107, + 609, + 135, + 621 + ], + "score": 0.91, + "content": "[ 1 , m ^ { i } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 607, + 387, + 622 + ], + "score": 1.0, + "content": "is discrete, while the behavior embedding space for each agent", + "type": "text" + }, + { + "bbox": [ + 387, + 609, + 406, + 619 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { N _ { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 607, + 506, + 622 + ], + "score": 1.0, + "content": "is continuous. Thus, the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 619, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 218, + 634 + ], + "score": 1.0, + "content": "meta policy is modeled as a", + "type": "text" + }, + { + "bbox": [ + 219, + 621, + 252, + 631 + ], + "score": 0.86, + "content": "( 2 \\times N )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 619, + 396, + 634 + ], + "score": 1.0, + "content": "-head neural network where the first", + "type": "text" + }, + { + "bbox": [ + 396, + 621, + 406, + 630 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 619, + 471, + 634 + ], + "score": 1.0, + "content": "heads represent", + "type": "text" + }, + { + "bbox": [ + 471, + 621, + 484, + 631 + ], + "score": 0.88, + "content": "m ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 619, + 506, + 634 + ], + "score": 1.0, + "content": "-way", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 323, + 644 + ], + "score": 1.0, + "content": "categorical distributions for skill selection and the last", + "type": "text" + }, + { + "bbox": [ + 324, + 632, + 334, + 641 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 631, + 399, + 644 + ], + "score": 1.0, + "content": "heads represent", + "type": "text" + }, + { + "bbox": [ + 399, + 632, + 413, + 642 + ], + "score": 0.89, + "content": "N _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "-dimensional Gaussian", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 642, + 318, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 318, + 654 + ], + "score": 1.0, + "content": "distributions for behavior control of the chosen skill.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "title", + "bbox": [ + 108, + 667, + 209, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 210, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 210, + 679 + ], + "score": 1.0, + "content": "3.5 IMPLEMENTATION", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 353, + 700 + ], + "score": 1.0, + "content": "We model the primitive policies and posterior distributions", + "type": "text" + }, + { + "bbox": [ + 353, + 691, + 364, + 700 + ], + "score": 0.79, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "as neural networks. We train the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "primitive policies using soft actor-critic (Haarnoja et al., 2018b). When we train a primitive policy,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 428, + 722 + ], + "score": 1.0, + "content": "we use a unit Gaussian distribution as the prior distribution of latent variables", + "type": "text" + }, + { + "bbox": [ + 428, + 710, + 447, + 722 + ], + "score": 0.92, + "content": "p ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 709, + 506, + 722 + ], + "score": 1.0, + "content": ". We use 5 as", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 265, + 733 + ], + "score": 1.0, + "content": "the size of latent behavior embedding", + "type": "text" + }, + { + "bbox": [ + 265, + 723, + 279, + 732 + ], + "score": 0.88, + "content": "N _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 720, + 505, + 733 + ], + "score": 1.0, + "content": ". Each primitive policy outputs the mean and standard", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 84, + 203, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 205, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 205, + 98 + ], + "score": 1.0, + "content": "Algorithm 1 ROLLOUT", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "index", + "bbox": [ + 109, + 99, + 487, + 247 + ], + "lines": [ + { + "bbox": [ + 109, + 96, + 474, + 113 + ], + "spans": [ + { + "bbox": [ + 109, + 96, + 204, + 113 + ], + "score": 1.0, + "content": "1: Input: Meta policy", + "type": "text" + }, + { + "bbox": [ + 204, + 101, + 229, + 110 + ], + "score": 0.86, + "content": "\\pi _ { m e t a }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 96, + 333, + 113 + ], + "score": 1.0, + "content": ", sets of primitive policies", + "type": "text" + }, + { + "bbox": [ + 334, + 98, + 379, + 110 + ], + "score": 0.92, + "content": "\\Pi ^ { 1 } , . . . , \\Pi ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 96, + 454, + 113 + ], + "score": 1.0, + "content": ", and meta horizon", + "type": "text" + }, + { + "bbox": [ + 455, + 100, + 474, + 110 + ], + "score": 0.87, + "content": "T _ { l o w }", + "type": "inline_equation" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 108, + 337, + 123 + ], + "spans": [ + { + "bbox": [ + 111, + 108, + 205, + 123 + ], + "score": 1.0, + "content": "2: Initialize an episode", + "type": "text" + }, + { + "bbox": [ + 206, + 111, + 232, + 120 + ], + "score": 0.89, + "content": "t \\gets 0", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 108, + 327, + 123 + ], + "score": 1.0, + "content": "and receive initial state", + "type": "text" + }, + { + "bbox": [ + 327, + 112, + 337, + 122 + ], + "score": 0.84, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 121, + 264, + 132 + ], + "spans": [ + { + "bbox": [ + 111, + 121, + 264, + 132 + ], + "score": 1.0, + "content": "3: while episode is not terminated do", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 130, + 485, + 146 + ], + "spans": [ + { + "bbox": [ + 110, + 132, + 121, + 143 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 131, + 130, + 323, + 146 + ], + "score": 1.0, + "content": "Sample skill indexes and behavior embeddings", + "type": "text" + }, + { + "bbox": [ + 323, + 131, + 485, + 144 + ], + "score": 0.92, + "content": "( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } ) \\sim \\pi _ { m e t a } ( s _ { t } )", + "type": "inline_equation" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 143, + 161, + 154 + ], + "spans": [ + { + "bbox": [ + 110, + 144, + 122, + 154 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 133, + 143, + 161, + 154 + ], + "score": 0.66, + "content": "\\tau 0", + "type": "inline_equation" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 153, + 331, + 165 + ], + "spans": [ + { + "bbox": [ + 110, + 154, + 122, + 165 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 133, + 153, + 158, + 165 + ], + "score": 1.0, + "content": "while", + "type": "text" + }, + { + "bbox": [ + 159, + 154, + 197, + 164 + ], + "score": 0.89, + "content": "\\tau < T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 153, + 331, + 165 + ], + "score": 1.0, + "content": "and episode is not terminated do", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 164, + 403, + 180 + ], + "spans": [ + { + "bbox": [ + 110, + 165, + 122, + 178 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 142, + 164, + 403, + 180 + ], + "score": 0.77, + "content": "\\boldsymbol a _ { t + \\tau } = ( a _ { t + \\tau } ^ { 1 } , \\dots , \\dot { a _ { t + \\tau } ^ { N } } ) \\sim ( \\pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( s _ { t + \\tau } , z _ { t } ^ { 1 } ) , \\dots , \\pi _ { c _ { t } ^ { N } } ^ { N } ( s _ { t + \\tau } , z _ { t } ^ { N } ) )", + "type": "inline_equation" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 177, + 283, + 202 + ], + "spans": [ + { + "bbox": [ + 110, + 179, + 121, + 190 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 142, + 177, + 283, + 202 + ], + "score": 0.38, + "content": "\\begin{array} { r l } & { s _ { t + \\tau + 1 } , r _ { t + \\tau } \\gets \\mathrm { E N V } ( s _ { t + \\tau } , a _ { t + \\tau } ) } \\\\ & { \\tau \\gets \\tau + 1 } \\end{array}", + "type": "inline_equation", + "image_path": "416eef5398b508d7f4182c92239826f0f1381cebb57c910d7fa2e5f5a9cd6011.jpg" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 190, + 122, + 202 + ], + "spans": [ + { + "bbox": [ + 109, + 190, + 122, + 202 + ], + "score": 1.0, + "content": "9:", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 201, + 177, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 122, + 213 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 133, + 201, + 177, + 212 + ], + "score": 1.0, + "content": "end while", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 210, + 473, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 122, + 224 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 131, + 210, + 199, + 225 + ], + "score": 1.0, + "content": "Add a transition", + "type": "text" + }, + { + "bbox": [ + 200, + 211, + 383, + 224 + ], + "score": 0.93, + "content": "s _ { t } , ( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } ) , s _ { t + \\tau } , r _ { t : t + \\tau - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 210, + 464, + 225 + ], + "score": 1.0, + "content": "to the rollout buffer", + "type": "text" + }, + { + "bbox": [ + 464, + 213, + 473, + 222 + ], + "score": 0.76, + "content": "\\boldsymbol { B }", + "type": "inline_equation" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 221, + 175, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 123, + 235 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 133, + 223, + 175, + 234 + ], + "score": 0.51, + "content": "t \\gets t + \\tau", + "type": "inline_equation" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 107, + 234, + 167, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 234, + 167, + 245 + ], + "score": 1.0, + "content": "13: end while", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + } + ], + "index": 7, + "bbox_fs": [ + 106, + 96, + 485, + 245 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 269, + 168, + 280 + ], + "lines": [ + { + "bbox": [ + 105, + 267, + 169, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 169, + 281 + ], + "score": 1.0, + "content": "for derivation):", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 267, + 169, + 281 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 284, + 455, + 331 + ], + "lines": [ + { + "bbox": [ + 155, + 284, + 455, + 331 + ], + "spans": [ + { + "bbox": [ + 155, + 284, + 455, + 331 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\mathcal { F } ( \\theta ) = M I ( s , z ) - M I ( a , z \\vert s ) + \\mathcal { H } ( a \\vert s ) = \\mathcal { H } ( a \\vert s , z ) - \\mathcal { H } ( z \\vert s ) + \\mathcal { H } ( z ) } \\\\ & { \\quad \\quad = \\mathcal { H } ( a \\vert s , z ) + \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log p ( z \\vert s ) ] - \\mathbb { E } _ { z \\sim p ( z ) } [ \\log p ( z ) ] } \\\\ & { \\quad \\quad \\geq \\mathcal { H } ( a \\vert s , z ) + \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log q _ { \\phi } ( z \\vert s ) - \\log p ( z ) ] , } \\end{array}", + "type": "interline_equation", + "image_path": "d92048ef0214b9413b930951da396cad2d15a7c9b9c6cb3e6d96816918b24d97.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 155, + 284, + 455, + 299.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 155, + 299.6666666666667, + 455, + 315.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 155, + 315.33333333333337, + 455, + 331.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 334, + 406, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 406, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 235, + 349 + ], + "score": 1.0, + "content": "where the learned discriminator", + "type": "text" + }, + { + "bbox": [ + 235, + 335, + 266, + 347 + ], + "score": 0.93, + "content": "q _ { \\phi } ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 333, + 376, + 349 + ], + "score": 1.0, + "content": "approximates the posterior", + "type": "text" + }, + { + "bbox": [ + 376, + 334, + 402, + 347 + ], + "score": 0.92, + "content": "p ( z | s )", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 333, + 406, + 349 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 333, + 406, + 349 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 351, + 503, + 373 + ], + "lines": [ + { + "bbox": [ + 104, + 350, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 104, + 350, + 505, + 365 + ], + "score": 1.0, + "content": "To achieve a primitive skill with diverse behaviors, we augment Equation (3) to the environment", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 362, + 140, + 375 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 140, + 375 + ], + "score": 1.0, + "content": "reward:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 104, + 350, + 505, + 375 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 180, + 373, + 430, + 387 + ], + "lines": [ + { + "bbox": [ + 180, + 373, + 430, + 387 + ], + "spans": [ + { + "bbox": [ + 180, + 373, + 430, + 387 + ], + "score": 0.89, + "content": "r _ { t } + \\lambda _ { 1 } \\mathcal { H } ( a | s , z ) + \\lambda _ { 2 } \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log q _ { \\phi } ( z | s ) - \\log p ( z ) ] ,", + "type": "interline_equation", + "image_path": "7dcfa1f10a0ed872bfb88b0fa8018a0f0cffb519a0831b833fece71eb0731dbc.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 180, + 373, + 430, + 387 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 388, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 133, + 401 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 389, + 145, + 399 + ], + "score": 0.88, + "content": "\\lambda _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 388, + 267, + 401 + ], + "score": 1.0, + "content": "is the entropy coefficient and", + "type": "text" + }, + { + "bbox": [ + 268, + 389, + 279, + 399 + ], + "score": 0.88, + "content": "\\lambda _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 388, + 506, + 401 + ], + "score": 1.0, + "content": "is the diversity coefficient which corresponds identifia-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "bility of behaviors. Maximizing Equation (3) encourages multi-modal exploration strategies while", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 411, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 200, + 423 + ], + "score": 1.0, + "content": "maximizing the reward", + "type": "text" + }, + { + "bbox": [ + 200, + 412, + 209, + 421 + ], + "score": 0.85, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 411, + 506, + 423 + ], + "score": 1.0, + "content": "forces to achieve its own goal. Moreover, by maximizing identifiability of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 421, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 218, + 433 + ], + "score": 1.0, + "content": "behaviors, the latent vector", + "type": "text" + }, + { + "bbox": [ + 218, + 424, + 225, + 431 + ], + "score": 0.76, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 421, + 506, + 433 + ], + "score": 1.0, + "content": ", named behavior embedding, can represent a variation of the learned", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "policy and thus can be used to control the behavior of the policy. For example, when training a", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 456 + ], + "score": 1.0, + "content": "robot to move an object, a policy learns to move the object quickly as well as slowly, and these", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 296, + 467 + ], + "score": 1.0, + "content": "diverse behaviors map to different latent vectors", + "type": "text" + }, + { + "bbox": [ + 296, + 456, + 302, + 464 + ], + "score": 0.77, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 454, + 505, + 467 + ], + "score": 1.0, + "content": ". We empirically show that the policies with diverse", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 465, + 425, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 425, + 478 + ], + "score": 1.0, + "content": "behaviors achieve better compositionality with other agents in our experiments.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 388, + 506, + 478 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 491, + 350, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 351, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 351, + 502 + ], + "score": 1.0, + "content": "3.4 COMPOSING PRIMITIVE SKILLS WITH META POLICY", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 509, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 103, + 508, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 103, + 508, + 222, + 526 + ], + "score": 1.0, + "content": "We denote the meta policy as", + "type": "text" + }, + { + "bbox": [ + 222, + 510, + 358, + 523 + ], + "score": 0.89, + "content": "\\pi _ { m e t a } ( c ^ { 1 } , \\ldots , c ^ { N } , z ^ { 1 } , \\ldots , z ^ { N } | s _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 508, + 388, + 526 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 388, + 510, + 436, + 523 + ], + "score": 0.93, + "content": "c ^ { i } \\in [ 1 , m ^ { i } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 508, + 505, + 526 + ], + "score": 1.0, + "content": "represents a skill", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 519, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 104, + 519, + 181, + 536 + ], + "score": 1.0, + "content": "index of an agent", + "type": "text" + }, + { + "bbox": [ + 182, + 522, + 224, + 534 + ], + "score": 0.93, + "content": "i \\in [ 1 , N ]", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 519, + 244, + 536 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 244, + 522, + 285, + 532 + ], + "score": 0.85, + "content": "z ^ { i } \\in \\mathbb { R } ^ { N _ { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 519, + 506, + 536 + ], + "score": 1.0, + "content": "represents a behavior embedding of the skill. Every", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 531, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 127, + 544 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 531, + 353, + 546 + ], + "score": 1.0, + "content": "time steps, the meta policy chooses one primitive skill", + "type": "text" + }, + { + "bbox": [ + 354, + 532, + 392, + 545 + ], + "score": 0.92, + "content": "\\pi _ { c ^ { i } } ^ { i } \\in \\Pi ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 531, + 455, + 546 + ], + "score": 1.0, + "content": "for each agent", + "type": "text" + }, + { + "bbox": [ + 455, + 533, + 460, + 542 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 531, + 506, + 546 + ], + "score": 1.0, + "content": ". Also, the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 542, + 507, + 561 + ], + "spans": [ + { + "bbox": [ + 104, + 542, + 343, + 561 + ], + "score": 1.0, + "content": "meta policy outputs a set of latent behavior embeddings", + "type": "text" + }, + { + "bbox": [ + 344, + 545, + 410, + 558 + ], + "score": 0.9, + "content": "( z ^ { 1 } , z ^ { 2 } , \\dots , z ^ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 542, + 507, + 561 + ], + "score": 1.0, + "content": "and feeds them to the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 101, + 552, + 504, + 574 + ], + "spans": [ + { + "bbox": [ + 101, + 552, + 210, + 574 + ], + "score": 1.0, + "content": "corresponding skills (i.e.", + "type": "text" + }, + { + "bbox": [ + 211, + 556, + 264, + 569 + ], + "score": 0.93, + "content": "\\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 552, + 306, + 574 + ], + "score": 1.0, + "content": "for agent", + "type": "text" + }, + { + "bbox": [ + 306, + 558, + 311, + 566 + ], + "score": 0.54, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 552, + 441, + 574 + ], + "score": 1.0, + "content": "). Once a set of primitive skills", + "type": "text" + }, + { + "bbox": [ + 441, + 556, + 504, + 569 + ], + "score": 0.93, + "content": "\\{ \\pi _ { c ^ { 1 } } ^ { 1 } , \\ldots , \\pi _ { c ^ { N } } ^ { N } \\}", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 375, + 582 + ], + "score": 1.0, + "content": "are chosen to be executed, each primitive skill generates an action", + "type": "text" + }, + { + "bbox": [ + 375, + 568, + 451, + 582 + ], + "score": 0.94, + "content": "a ^ { i } \\sim \\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "based on the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 581, + 467, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 158, + 594 + ], + "score": 1.0, + "content": "current state", + "type": "text" + }, + { + "bbox": [ + 158, + 581, + 167, + 591 + ], + "score": 0.86, + "content": "s ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 581, + 252, + 594 + ], + "score": 1.0, + "content": "and the latent vector", + "type": "text" + }, + { + "bbox": [ + 252, + 581, + 261, + 591 + ], + "score": 0.86, + "content": "z ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 581, + 467, + 594 + ], + "score": 1.0, + "content": ". Algorithm 1 illustrates the overall rollout process.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 101, + 508, + 507, + 594 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 598, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 506, + 611 + ], + "score": 1.0, + "content": "Since there are a finite number of skills for each agent to execute, the meta action space for each agent", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 607, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 107, + 609, + 135, + 621 + ], + "score": 0.91, + "content": "[ 1 , m ^ { i } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 607, + 387, + 622 + ], + "score": 1.0, + "content": "is discrete, while the behavior embedding space for each agent", + "type": "text" + }, + { + "bbox": [ + 387, + 609, + 406, + 619 + ], + "score": 0.9, + "content": "\\mathbb { R } ^ { N _ { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 607, + 506, + 622 + ], + "score": 1.0, + "content": "is continuous. Thus, the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 619, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 218, + 634 + ], + "score": 1.0, + "content": "meta policy is modeled as a", + "type": "text" + }, + { + "bbox": [ + 219, + 621, + 252, + 631 + ], + "score": 0.86, + "content": "( 2 \\times N )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 619, + 396, + 634 + ], + "score": 1.0, + "content": "-head neural network where the first", + "type": "text" + }, + { + "bbox": [ + 396, + 621, + 406, + 630 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 619, + 471, + 634 + ], + "score": 1.0, + "content": "heads represent", + "type": "text" + }, + { + "bbox": [ + 471, + 621, + 484, + 631 + ], + "score": 0.88, + "content": "m ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 619, + 506, + 634 + ], + "score": 1.0, + "content": "-way", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 323, + 644 + ], + "score": 1.0, + "content": "categorical distributions for skill selection and the last", + "type": "text" + }, + { + "bbox": [ + 324, + 632, + 334, + 641 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 631, + 399, + 644 + ], + "score": 1.0, + "content": "heads represent", + "type": "text" + }, + { + "bbox": [ + 399, + 632, + 413, + 642 + ], + "score": 0.89, + "content": "N _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "-dimensional Gaussian", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 642, + 318, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 318, + 654 + ], + "score": 1.0, + "content": "distributions for behavior control of the chosen skill.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 597, + 506, + 654 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 667, + 209, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 210, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 210, + 679 + ], + "score": 1.0, + "content": "3.5 IMPLEMENTATION", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 353, + 700 + ], + "score": 1.0, + "content": "We model the primitive policies and posterior distributions", + "type": "text" + }, + { + "bbox": [ + 353, + 691, + 364, + 700 + ], + "score": 0.79, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "as neural networks. We train the", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "primitive policies using soft actor-critic (Haarnoja et al., 2018b). When we train a primitive policy,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 428, + 722 + ], + "score": 1.0, + "content": "we use a unit Gaussian distribution as the prior distribution of latent variables", + "type": "text" + }, + { + "bbox": [ + 428, + 710, + 447, + 722 + ], + "score": 0.92, + "content": "p ( z )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 709, + 506, + 722 + ], + "score": 1.0, + "content": ". We use 5 as", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 265, + 733 + ], + "score": 1.0, + "content": "the size of latent behavior embedding", + "type": "text" + }, + { + "bbox": [ + 265, + 723, + 279, + 732 + ], + "score": 0.88, + "content": "N _ { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 720, + 505, + 733 + ], + "score": 1.0, + "content": ". Each primitive policy outputs the mean and standard", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "deviation of a Gaussian distribution over an action space. For a primitive policy, we apply tanh", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 275, + 323 + ], + "score": 1.0, + "content": "activation to normalize the action between", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 275, + 311, + 303, + 322 + ], + "score": 0.86, + "content": "[ - 1 , 1 ]", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 304, + 310, + 505, + 323 + ], + "score": 1.0, + "content": ". We model the meta policy as neural network with", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 276, + 334 + ], + "score": 1.0, + "content": "multiple heads that output the skill index", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 276, + 322, + 284, + 331 + ], + "score": 0.82, + "content": "\\bar { c } ^ { i }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 285, + 321, + 389, + 334 + ], + "score": 1.0, + "content": "and behavior embedding", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 389, + 321, + 399, + 332 + ], + "score": 0.86, + "content": "z ^ { i }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 399, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "for each agent. The meta", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 333, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 344 + ], + "score": 1.0, + "content": "policy is trained using PPO (Schulman et al., 2017; 2016; Dhariwal et al., 2017). All policy networks", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "in this paper consist of 3 fully connected layers of 64 hidden units with ReLU nonlinearities. The", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 354, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 162, + 366 + ], + "score": 1.0, + "content": "discriminator", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 162, + 356, + 173, + 367 + ], + "score": 0.87, + "content": "q _ { \\phi }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 173, + 354, + 470, + 366 + ], + "score": 1.0, + "content": "in Equation (4) is a 2-layer fully connected network with 64 hidden units.", + "type": "text", + "cross_page": true + } + ], + "index": 13 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 79, + 486, + 207 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 79, + 486, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 79, + 486, + 207 + ], + "spans": [ + { + "bbox": [ + 126, + 79, + 486, + 207 + ], + "score": 0.976, + "type": "image", + "image_path": "11edbefb96b52e765ac0cdb4e36075acafa79155b0f2fabe7fb7cdafaf5487ff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 126, + 79, + 486, + 121.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 126, + 121.66666666666666, + 486, + 164.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 126, + 164.33333333333331, + 486, + 206.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 217, + 506, + 272 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "Figure 4: The composite tasks pose a challenging combination of object manipulation and locomotion", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "score": 1.0, + "content": "skills, which requires coordination of multiple agents and temporally extended behaviors. (a) The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 238, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 252 + ], + "score": 1.0, + "content": "left Jaco arm needs to pick up a block while the right Jaco arm pushes a container, and then it places", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "the block into the container. (b) Two Jaco arms are required to pick and place a bar-shaped block", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 261, + 437, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 437, + 274 + ], + "score": 1.0, + "content": "together. (c) Two ants push the red box to the goal location (green circle) together.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 299, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "deviation of a Gaussian distribution over an action space. For a primitive policy, we apply tanh", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 275, + 323 + ], + "score": 1.0, + "content": "activation to normalize the action between", + "type": "text" + }, + { + "bbox": [ + 275, + 311, + 303, + 322 + ], + "score": 0.86, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 310, + 505, + 323 + ], + "score": 1.0, + "content": ". We model the meta policy as neural network with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 276, + 334 + ], + "score": 1.0, + "content": "multiple heads that output the skill index", + "type": "text" + }, + { + "bbox": [ + 276, + 322, + 284, + 331 + ], + "score": 0.82, + "content": "\\bar { c } ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 321, + 389, + 334 + ], + "score": 1.0, + "content": "and behavior embedding", + "type": "text" + }, + { + "bbox": [ + 389, + 321, + 399, + 332 + ], + "score": 0.86, + "content": "z ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "for each agent. The meta", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 333, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 344 + ], + "score": 1.0, + "content": "policy is trained using PPO (Schulman et al., 2017; 2016; Dhariwal et al., 2017). All policy networks", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "in this paper consist of 3 fully connected layers of 64 hidden units with ReLU nonlinearities. The", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 354, + 470, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 162, + 366 + ], + "score": 1.0, + "content": "discriminator", + "type": "text" + }, + { + "bbox": [ + 162, + 356, + 173, + 367 + ], + "score": 0.87, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 354, + 470, + 366 + ], + "score": 1.0, + "content": "in Equation (4) is a 2-layer fully connected network with 64 hidden units.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 388, + 200, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 387, + 201, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 201, + 403 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "To demonstrate the effectiveness of our framework, we compare our method to prior methods in the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 427, + 507, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 507, + 441 + ], + "score": 1.0, + "content": "field of multi-agent RL and ablate the components of our framework to understand their importance.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "We conducted experiments on a set of challenging robot control environments that require coordination", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 451, + 460, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 460, + 461 + ], + "score": 1.0, + "content": "of different agents to complete collaborative robotic manipulation and locomotion tasks.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "Through our experiments, we aim to answer the following questions: (1) can our framework efficiently", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 478, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 490 + ], + "score": 1.0, + "content": "learn to combine primitive skills to execute a complicated task; (2) can our learned agent exhibit", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "collaborative behaviors during task execution; and (3) can our framework leverage the controllable", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 500, + 398, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 398, + 511 + ], + "score": 1.0, + "content": "behavior variations of the primitive skills to achieve better coordination?", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "For details about environments and training, please refer to the supplementary material. As the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "performance of training algorithms varies between runs, we train each method on each task with 6", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 538, + 483, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 483, + 551 + ], + "score": 1.0, + "content": "different random seeds and report mean and standard deviation of each method’s success rate.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 107, + 569, + 180, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 181, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 181, + 582 + ], + "score": 1.0, + "content": "4.1 BASELINES", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 108, + 592, + 504, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 506, + 606 + ], + "score": 1.0, + "content": "We compare the performance of our method with various single- and multi-agent RL methods", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 604, + 198, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 198, + 617 + ], + "score": 1.0, + "content": "illustrated in Figure 3:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 503, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 635 + ], + "score": 1.0, + "content": "Single-agent RL (RL): A vanilla RL method where a single policy takes as input the full observation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 632, + 232, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 232, + 644 + ], + "score": 1.0, + "content": "and outputs all agents’ actions.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 386, + 662 + ], + "score": 1.0, + "content": "Multi-agent RL (MARL): A multi-agent RL method where each of", + "type": "text" + }, + { + "bbox": [ + 387, + 649, + 397, + 659 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "policies takes as input the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "observation of the corresponding agent and outputs an action for that agent. All policies share the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 671, + 372, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 372, + 684 + ], + "score": 1.0, + "content": "global critic learned from a single task reward (Lowe et al., 2017).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 465, + 700 + ], + "score": 1.0, + "content": "Modular Framework (Modular): A modular framework composed of a meta policy and", + "type": "text" + }, + { + "bbox": [ + 466, + 688, + 476, + 698 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "sets of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 374, + 712 + ], + "score": 1.0, + "content": "primitive skills (i.e. one or more primitive skills per agent). Every", + "type": "text" + }, + { + "bbox": [ + 374, + 700, + 394, + 710 + ], + "score": 0.89, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "time steps, the meta policy", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "selects a primitive skill for each agent based on the full observation. Then, the chosen skills are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 719, + 225, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 158, + 734 + ], + "score": 1.0, + "content": "executed for", + "type": "text" + }, + { + "bbox": [ + 158, + 721, + 178, + 732 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 719, + 225, + 734 + ], + "score": 1.0, + "content": "time steps.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 79, + 486, + 207 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 79, + 486, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 79, + 486, + 207 + ], + "spans": [ + { + "bbox": [ + 126, + 79, + 486, + 207 + ], + "score": 0.976, + "type": "image", + "image_path": "11edbefb96b52e765ac0cdb4e36075acafa79155b0f2fabe7fb7cdafaf5487ff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 126, + 79, + 486, + 121.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 126, + 121.66666666666666, + 486, + 164.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 126, + 164.33333333333331, + 486, + 206.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 217, + 506, + 272 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "Figure 4: The composite tasks pose a challenging combination of object manipulation and locomotion", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "score": 1.0, + "content": "skills, which requires coordination of multiple agents and temporally extended behaviors. (a) The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 238, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 252 + ], + "score": 1.0, + "content": "left Jaco arm needs to pick up a block while the right Jaco arm pushes a container, and then it places", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "the block into the container. (b) Two Jaco arms are required to pick and place a bar-shaped block", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 261, + 437, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 437, + 274 + ], + "score": 1.0, + "content": "together. (c) Two ants push the red box to the goal location (green circle) together.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 107, + 299, + 505, + 366 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 105, + 299, + 506, + 367 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 388, + 200, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 387, + 201, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 201, + 403 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "To demonstrate the effectiveness of our framework, we compare our method to prior methods in the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 427, + 507, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 507, + 441 + ], + "score": 1.0, + "content": "field of multi-agent RL and ablate the components of our framework to understand their importance.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 505, + 451 + ], + "score": 1.0, + "content": "We conducted experiments on a set of challenging robot control environments that require coordination", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 451, + 460, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 460, + 461 + ], + "score": 1.0, + "content": "of different agents to complete collaborative robotic manipulation and locomotion tasks.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 416, + 507, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 506, + 480 + ], + "score": 1.0, + "content": "Through our experiments, we aim to answer the following questions: (1) can our framework efficiently", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 478, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 490 + ], + "score": 1.0, + "content": "learn to combine primitive skills to execute a complicated task; (2) can our learned agent exhibit", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 489, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 501 + ], + "score": 1.0, + "content": "collaborative behaviors during task execution; and (3) can our framework leverage the controllable", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 500, + 398, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 398, + 511 + ], + "score": 1.0, + "content": "behavior variations of the primitive skills to achieve better coordination?", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 467, + 506, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 504, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "For details about environments and training, please refer to the supplementary material. As the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "performance of training algorithms varies between runs, we train each method on each task with 6", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 538, + 483, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 483, + 551 + ], + "score": 1.0, + "content": "different random seeds and report mean and standard deviation of each method’s success rate.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 515, + 505, + 551 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 569, + 180, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 181, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 181, + 582 + ], + "score": 1.0, + "content": "4.1 BASELINES", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 108, + 592, + 504, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 506, + 606 + ], + "score": 1.0, + "content": "We compare the performance of our method with various single- and multi-agent RL methods", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 604, + 198, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 198, + 617 + ], + "score": 1.0, + "content": "illustrated in Figure 3:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 591, + 506, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 503, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 619, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 635 + ], + "score": 1.0, + "content": "Single-agent RL (RL): A vanilla RL method where a single policy takes as input the full observation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 632, + 232, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 232, + 644 + ], + "score": 1.0, + "content": "and outputs all agents’ actions.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 619, + 505, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 386, + 662 + ], + "score": 1.0, + "content": "Multi-agent RL (MARL): A multi-agent RL method where each of", + "type": "text" + }, + { + "bbox": [ + 387, + 649, + 397, + 659 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "policies takes as input the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "observation of the corresponding agent and outputs an action for that agent. All policies share the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 671, + 372, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 372, + 684 + ], + "score": 1.0, + "content": "global critic learned from a single task reward (Lowe et al., 2017).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 648, + 505, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 465, + 700 + ], + "score": 1.0, + "content": "Modular Framework (Modular): A modular framework composed of a meta policy and", + "type": "text" + }, + { + "bbox": [ + 466, + 688, + 476, + 698 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "sets of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 374, + 712 + ], + "score": 1.0, + "content": "primitive skills (i.e. one or more primitive skills per agent). Every", + "type": "text" + }, + { + "bbox": [ + 374, + 700, + 394, + 710 + ], + "score": 0.89, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "time steps, the meta policy", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "selects a primitive skill for each agent based on the full observation. Then, the chosen skills are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 719, + 225, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 158, + 734 + ], + "score": 1.0, + "content": "executed for", + "type": "text" + }, + { + "bbox": [ + 158, + 721, + 178, + 732 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 719, + 225, + 734 + ], + "score": 1.0, + "content": "time steps.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 687, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 82, + 496, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 82, + 496, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 82, + 496, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 82, + 496, + 198 + ], + "score": 0.97, + "type": "image", + "image_path": "38d7a3b3c6d72759e19da2cea6bad148d9e0ead37b5a570252e66d5ccb027d90.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 82, + 496, + 120.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 120.66666666666666, + 496, + 159.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 159.33333333333331, + 496, + 197.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 207, + 506, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 208, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 219 + ], + "score": 1.0, + "content": "Figure 5: Success rates of our method (Modular-SBD) and baselines. For modular frameworks", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "(Modular and Modular-SBD), we shift the learning curves rightwards the total number of environment", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 229, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 327, + 241 + ], + "score": 1.0, + "content": "steps the agent takes to learn the primitive skills (0.9 M,", + "type": "text" + }, + { + "bbox": [ + 327, + 230, + 352, + 240 + ], + "score": 0.25, + "content": "1 . 2 \\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 229, + 371, + 241 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 372, + 230, + 397, + 240 + ], + "score": 0.68, + "content": "2 . 0 \\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 229, + 506, + 241 + ], + "score": 1.0, + "content": ", respectively). Our method", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 240, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 507, + 253 + ], + "score": 1.0, + "content": "substantially improves learning speed and performance on JACO PICK-PUSH-PLACE and ANT PUSH.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "The shaded areas represent the standard deviation of results from six different seeds. The curves are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 262, + 292, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 292, + 275 + ], + "score": 1.0, + "content": "smoothed using moving average over 10 runs.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 297, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "Single-agent RL with Skill Behavior Diversification (RL-SBD): An RL method augmented with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "the behavior diversification objective. A meta policy is employed to generate a behavior embedding", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "score": 1.0, + "content": "for a low-level policy, and the low-level policy outputs all agents’ actions conditioned on the behavior", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 261, + 343 + ], + "score": 1.0, + "content": "embedding and the full observation for", + "type": "text" + }, + { + "bbox": [ + 262, + 331, + 281, + 341 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "time steps. The meta policy and the low-level policy are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 340, + 441, + 355 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 441, + 355 + ], + "score": 1.0, + "content": "jointly trained with the behavior diversification objective described in Equation (4).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 358, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "Multi-agent RL with Skill Behavior Diversification (MARL-SBD): A MARL method augmented", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 368, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 378, + 382 + ], + "score": 1.0, + "content": "with the behavior diversification objective. A meta policy generates", + "type": "text" + }, + { + "bbox": [ + 378, + 370, + 389, + 379 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 368, + 506, + 382 + ], + "score": 1.0, + "content": "behavior embeddings. Then,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "score": 1.0, + "content": "each low-level policy outputs each agent’s action conditioned on its observation and behavior", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 391, + 463, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 167, + 403 + ], + "score": 1.0, + "content": "embedding for", + "type": "text" + }, + { + "bbox": [ + 168, + 391, + 187, + 402 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 392, + 463, + 403 + ], + "score": 1.0, + "content": "time steps. All policies are jointly trained to maximize Equation (4).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 463 + ], + "lines": [ + { + "bbox": [ + 106, + 408, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 504, + 419 + ], + "score": 1.0, + "content": "Modular Framework with Skill Behavior Diversification (Modular-SBD, Ours): Our method", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "score": 1.0, + "content": "which coordinates primitive skills of multiple agents. The modular framework consists of a meta", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 154, + 443 + ], + "score": 1.0, + "content": "policy and", + "type": "text" + }, + { + "bbox": [ + 154, + 430, + 165, + 440 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "sets of primitive skills, where each primitive skill is conditioned on a behavior", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 154, + 453 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 154, + 443, + 160, + 451 + ], + "score": 0.73, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 440, + 505, + 453 + ], + "score": 1.0, + "content": ". The meta policy takes as input the full observation and selects both a primitive skill", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 451, + 500, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 500, + 465 + ], + "score": 1.0, + "content": "and a behavior embedding for each agent. Then, each primitive skill outputs action for each agent.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 107, + 478, + 237, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 238, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 238, + 491 + ], + "score": 1.0, + "content": "4.2 JACO PICK-PUSH-PLACE", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "We developed JACO PICK-PUSH-PLACE and JACO BAR-MOVING environments using two Kinova", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "Jaco arms, where each Jaco arm is a 9 DoF robotic arm with 3 fingers. JACO PICK-PUSH-PLACE", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "starts with a block on the left and a container on the right. The robotic arms need to pick up the block,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "push the container to the center, and place the block inside the container. For successful completion", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "of the task, the two Jaco arms have to concurrently execute their distinct sets of skills and dynamically", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 555, + 447, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 447, + 568 + ], + "score": 1.0, + "content": "adjust their picking, pushing, and placing directions to avoid collision between arms.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "Primitives skills. There are three primitive skills available to each arm: Picking up, Pushing, and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "Placing to center (see Figure 4a). Picking up requires a robotic arm to pick up a small block, which is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "randomly placed on the table. If the block is not picked up after a certain amount of time or the arm", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "drops the block, the agent fails. Pushing learns to push a big container to its opposite side (e.g. from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "left to the center or from right to center). The agent fails if it cannot place the container to the center.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "Placing to center requires placing an object in the gripper to the table. The agent only succeeds when", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 638, + 369, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 369, + 650 + ], + "score": 1.0, + "content": "it stably places the object at the desired location on the container.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Composite task. Our method (Modular-SBD) can successfully perform JACO PICK-PUSH-PLACE", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "task while all baselines fail to compose primitive skills as shown in Figure 5a. The RL and MARL", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "baselines cannot learn the composite task mainly because the agent requires to learn the combinatorial", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 507, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 507, + 700 + ], + "score": 1.0, + "content": "number of skill compositions and to solve the credit assignment problem across multiple agents.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Since the composite task requires multiple primitive skills of multiple agents to be performed properly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "at the same time, a reward signal about a failure case cannot be assigned to the correct agent or skill.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "By using pre-trained primitive skills, the credit assignment problem is relaxed and all agents can", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 82, + 496, + 198 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 82, + 496, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 82, + 496, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 82, + 496, + 198 + ], + "score": 0.97, + "type": "image", + "image_path": "38d7a3b3c6d72759e19da2cea6bad148d9e0ead37b5a570252e66d5ccb027d90.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 82, + 496, + 120.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 120.66666666666666, + 496, + 159.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 159.33333333333331, + 496, + 197.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 207, + 506, + 273 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 208, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 219 + ], + "score": 1.0, + "content": "Figure 5: Success rates of our method (Modular-SBD) and baselines. For modular frameworks", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "(Modular and Modular-SBD), we shift the learning curves rightwards the total number of environment", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 229, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 327, + 241 + ], + "score": 1.0, + "content": "steps the agent takes to learn the primitive skills (0.9 M,", + "type": "text" + }, + { + "bbox": [ + 327, + 230, + 352, + 240 + ], + "score": 0.25, + "content": "1 . 2 \\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 229, + 371, + 241 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 372, + 230, + 397, + 240 + ], + "score": 0.68, + "content": "2 . 0 \\mathbf { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 229, + 506, + 241 + ], + "score": 1.0, + "content": ", respectively). Our method", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 240, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 507, + 253 + ], + "score": 1.0, + "content": "substantially improves learning speed and performance on JACO PICK-PUSH-PLACE and ANT PUSH.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "The shaded areas represent the standard deviation of results from six different seeds. The curves are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 262, + 292, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 292, + 275 + ], + "score": 1.0, + "content": "smoothed using moving average over 10 runs.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 297, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "Single-agent RL with Skill Behavior Diversification (RL-SBD): An RL method augmented with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "the behavior diversification objective. A meta policy is employed to generate a behavior embedding", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "score": 1.0, + "content": "for a low-level policy, and the low-level policy outputs all agents’ actions conditioned on the behavior", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 261, + 343 + ], + "score": 1.0, + "content": "embedding and the full observation for", + "type": "text" + }, + { + "bbox": [ + 262, + 331, + 281, + 341 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "time steps. The meta policy and the low-level policy are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 340, + 441, + 355 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 441, + 355 + ], + "score": 1.0, + "content": "jointly trained with the behavior diversification objective described in Equation (4).", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 104, + 297, + 506, + 355 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 358, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "Multi-agent RL with Skill Behavior Diversification (MARL-SBD): A MARL method augmented", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 368, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 378, + 382 + ], + "score": 1.0, + "content": "with the behavior diversification objective. A meta policy generates", + "type": "text" + }, + { + "bbox": [ + 378, + 370, + 389, + 379 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 368, + 506, + 382 + ], + "score": 1.0, + "content": "behavior embeddings. Then,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 392 + ], + "score": 1.0, + "content": "each low-level policy outputs each agent’s action conditioned on its observation and behavior", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 391, + 463, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 167, + 403 + ], + "score": 1.0, + "content": "embedding for", + "type": "text" + }, + { + "bbox": [ + 168, + 391, + 187, + 402 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 392, + 463, + 403 + ], + "score": 1.0, + "content": "time steps. All policies are jointly trained to maximize Equation (4).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 358, + 506, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 463 + ], + "lines": [ + { + "bbox": [ + 106, + 408, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 504, + 419 + ], + "score": 1.0, + "content": "Modular Framework with Skill Behavior Diversification (Modular-SBD, Ours): Our method", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "score": 1.0, + "content": "which coordinates primitive skills of multiple agents. The modular framework consists of a meta", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 154, + 443 + ], + "score": 1.0, + "content": "policy and", + "type": "text" + }, + { + "bbox": [ + 154, + 430, + 165, + 440 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "sets of primitive skills, where each primitive skill is conditioned on a behavior", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 154, + 453 + ], + "score": 1.0, + "content": "embedding", + "type": "text" + }, + { + "bbox": [ + 154, + 443, + 160, + 451 + ], + "score": 0.73, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 440, + 505, + 453 + ], + "score": 1.0, + "content": ". The meta policy takes as input the full observation and selects both a primitive skill", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 451, + 500, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 500, + 465 + ], + "score": 1.0, + "content": "and a behavior embedding for each agent. Then, each primitive skill outputs action for each agent.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 408, + 506, + 465 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 478, + 237, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 238, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 238, + 491 + ], + "score": 1.0, + "content": "4.2 JACO PICK-PUSH-PLACE", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "We developed JACO PICK-PUSH-PLACE and JACO BAR-MOVING environments using two Kinova", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "Jaco arms, where each Jaco arm is a 9 DoF robotic arm with 3 fingers. JACO PICK-PUSH-PLACE", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "starts with a block on the left and a container on the right. The robotic arms need to pick up the block,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "push the container to the center, and place the block inside the container. For successful completion", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "of the task, the two Jaco arms have to concurrently execute their distinct sets of skills and dynamically", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 555, + 447, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 447, + 568 + ], + "score": 1.0, + "content": "adjust their picking, pushing, and placing directions to avoid collision between arms.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 500, + 506, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "Primitives skills. There are three primitive skills available to each arm: Picking up, Pushing, and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "Placing to center (see Figure 4a). Picking up requires a robotic arm to pick up a small block, which is", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "randomly placed on the table. If the block is not picked up after a certain amount of time or the arm", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "drops the block, the agent fails. Pushing learns to push a big container to its opposite side (e.g. from", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "left to the center or from right to center). The agent fails if it cannot place the container to the center.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "Placing to center requires placing an object in the gripper to the table. The agent only succeeds when", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 638, + 369, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 369, + 650 + ], + "score": 1.0, + "content": "it stably places the object at the desired location on the container.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 572, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Composite task. Our method (Modular-SBD) can successfully perform JACO PICK-PUSH-PLACE", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "task while all baselines fail to compose primitive skills as shown in Figure 5a. The RL and MARL", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "baselines cannot learn the composite task mainly because the agent requires to learn the combinatorial", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 687, + 507, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 507, + 700 + ], + "score": 1.0, + "content": "number of skill compositions and to solve the credit assignment problem across multiple agents.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Since the composite task requires multiple primitive skills of multiple agents to be performed properly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "at the same time, a reward signal about a failure case cannot be assigned to the correct agent or skill.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "By using pre-trained primitive skills, the credit assignment problem is relaxed and all agents can", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 655, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 80, + 452, + 155 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 159, + 80, + 452, + 155 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 80, + 452, + 155 + ], + "spans": [ + { + "bbox": [ + 159, + 80, + 452, + 155 + ], + "score": 0.981, + "html": "
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
RL0.000±0.0000.000±0.0000.000±0.000
MARL0.000±0.0000.000± 0.0000.000± 0.000
RL-SBD0.000± 0.0000.000± 0.0000.000±0.000
MARL-SBD0.000 ± 0.0000.000 ±0.0000.000 ± 0.000
Modular0.324± 0.4680.917± 0.2760.003± 0.058
Modular-SBD (Ours)0.902 ± 0.2980.950 ± 0.2180.323 ± 0.468
", + "type": "table", + "image_path": "81b3836f62ff4588cec048ce9a9d8153404ee36c49b3e4b3237d9bad87b5e2a8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 159, + 80, + 452, + 105.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 159, + 105.0, + 452, + 130.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 159, + 130.0, + 452, + 155.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 163, + 504, + 196 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "Table 1: Success rates for all tasks, comparing our method against baselines. Each entry in the table", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "represents average success rate and standard deviation over 100 runs. The baselines learning from", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 185, + 330, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 330, + 198 + ], + "score": 1.0, + "content": "scratch fail to learn complex tasks with multiple agents.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 218, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "perform their skills concurrently. Therefore, the Modular baseline learns to achieve success but shows", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "significantly lower performance than our method (Modular-SBD). This is because the lack of skill", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 241, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 253 + ], + "score": 1.0, + "content": "behavior diversification makes it impossible to adjust pushing and placing trajectories during skill", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 252, + 385, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 385, + 264 + ], + "score": 1.0, + "content": "composition time, which resulting in frequent end-effector collisions.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 107, + 277, + 216, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 217, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 217, + 291 + ], + "score": 1.0, + "content": "4.3 JACO BAR-MOVING", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 311 + ], + "score": 1.0, + "content": "In JACO BAR-MOVING, two Jaco arms need to pick up a long bar together, move the bar towards a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "target location while maintaining its rotation, and place it on the table (see Figure 4b). The initial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "position of the bar is randomly initialized every episode and an agent needs to find appropriate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "coordination between two arms for each initialization. Compared to JACO PICK-PUSH-PLACE,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "score": 1.0, + "content": "this task requires that the two arms synchronize their movements and perform more micro-level", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 354, + 231, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 231, + 366 + ], + "score": 1.0, + "content": "adjustments to their behaviors.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "Primitives skills. There are two pre-trained primitive skills available to each arm: Picking up and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 381, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 394 + ], + "score": 1.0, + "content": "Placing towards arm. Picking up is same as described in Section 4.2. Placing towards arm learns to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "move a small block (half size of the block used in the composite task) in the hand towards the robotic", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 403, + 501, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 501, + 415 + ], + "score": 1.0, + "content": "arm and then place it on the table. The agent fails if it cannot place the block to the target location.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 420, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 106, + 421, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 506, + 433 + ], + "score": 1.0, + "content": "Composite task. The JACO BAR-MOVING task requires the two arms to work very closely together.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 431, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 506, + 444 + ], + "score": 1.0, + "content": "For example, the Picking up skill of both arms should be synchronized when they start to lift the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "bar and two arms require to lift the bar while maintaining the relative position between them since", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 453, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 506, + 466 + ], + "score": 1.0, + "content": "they are connected by holding the bar. The modular framework without explicit coordination of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "skills (Modular) can synchronize the execution of picking, moving, and placing. But the inability", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "to micro-adjust the movement of the other arm causes instability of bar picking and moving. This", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "results in degraded success rates compared to the modular framework with explicit coordination.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 497, + 488, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 488, + 509 + ], + "score": 1.0, + "content": "Meanwhile, all baselines without pre-defined primitive skills fail to learn JACO BAR-MOVING.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 523, + 177, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 178, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 178, + 535 + ], + "score": 1.0, + "content": "4.4 ANT PUSH", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "We developed a multi-ant environment, ANT-PUSH, inspired from Nachum et al. (2019), simulated in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "the MuJoCo (Todorov et al., 2012) physics engine. We use the ant model in OpenAI Gym (Brockman", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "et al., 2016). In this environment, two ants need to push a large object toward a green target place,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 578, + 495, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 495, + 590 + ], + "score": 1.0, + "content": "collaborating with each other to keep the angle of the object as stable as possible (see Figure 4c).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "Primitives skills. We train walking skills of an ant agent in 4 directions: up, down, left, and right.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "During primitive skill training, a block (half size of the block used in the composite task) and an", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "ant agent are randomly placed. Pushing the block gives an additional reward to the agent, which", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "prevents an ant to avoid the block. The learned primitive skills have different speed and trajectories", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 636, + 294, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 294, + 652 + ], + "score": 1.0, + "content": "conditioned on the latent behavior embedding.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 108, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 108, + 655, + 262, + 667 + ], + "score": 1.0, + "content": "Composite task. Our method achieves", + "type": "text" + }, + { + "bbox": [ + 262, + 655, + 289, + 666 + ], + "score": 0.87, + "content": "3 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "success rate on ANT PUSH task while all baselines fail", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 666, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 677 + ], + "score": 1.0, + "content": "to compose primitive skills as shown in Figure 5c and Table 1. The poor performance of RL, MARL,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "RL-SBD, and MARL-SBD baselines shows the difficulty of credit assignment between agents, which", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 687, + 507, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 507, + 700 + ], + "score": 1.0, + "content": "leads one of the ants moves toward a block and pushes it but another ant does not move. Moreover,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "the Modular baseline with primitive skills also fails to learn the pushing task. This result illustrates", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "the importance of coordination of agents, which helps synchronizing and controlling the velocities of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 461, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 461, + 733 + ], + "score": 1.0, + "content": "both ant agents to push the block toward the goal position while maintaining its rotation.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 80, + 452, + 155 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 159, + 80, + 452, + 155 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 80, + 452, + 155 + ], + "spans": [ + { + "bbox": [ + 159, + 80, + 452, + 155 + ], + "score": 0.981, + "html": "
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
RL0.000±0.0000.000±0.0000.000±0.000
MARL0.000±0.0000.000± 0.0000.000± 0.000
RL-SBD0.000± 0.0000.000± 0.0000.000±0.000
MARL-SBD0.000 ± 0.0000.000 ±0.0000.000 ± 0.000
Modular0.324± 0.4680.917± 0.2760.003± 0.058
Modular-SBD (Ours)0.902 ± 0.2980.950 ± 0.2180.323 ± 0.468
", + "type": "table", + "image_path": "81b3836f62ff4588cec048ce9a9d8153404ee36c49b3e4b3237d9bad87b5e2a8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 159, + 80, + 452, + 105.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 159, + 105.0, + 452, + 130.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 159, + 130.0, + 452, + 155.0 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 163, + 504, + 196 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "Table 1: Success rates for all tasks, comparing our method against baselines. Each entry in the table", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "represents average success rate and standard deviation over 100 runs. The baselines learning from", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 185, + 330, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 330, + 198 + ], + "score": 1.0, + "content": "scratch fail to learn complex tasks with multiple agents.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 162, + 505, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 218, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 231 + ], + "score": 1.0, + "content": "perform their skills concurrently. Therefore, the Modular baseline learns to achieve success but shows", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "significantly lower performance than our method (Modular-SBD). This is because the lack of skill", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 241, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 253 + ], + "score": 1.0, + "content": "behavior diversification makes it impossible to adjust pushing and placing trajectories during skill", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 252, + 385, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 385, + 264 + ], + "score": 1.0, + "content": "composition time, which resulting in frequent end-effector collisions.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 218, + 505, + 264 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 277, + 216, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 217, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 217, + 291 + ], + "score": 1.0, + "content": "4.3 JACO BAR-MOVING", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 311 + ], + "score": 1.0, + "content": "In JACO BAR-MOVING, two Jaco arms need to pick up a long bar together, move the bar towards a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "target location while maintaining its rotation, and place it on the table (see Figure 4b). The initial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 334 + ], + "score": 1.0, + "content": "position of the bar is randomly initialized every episode and an agent needs to find appropriate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 345 + ], + "score": 1.0, + "content": "coordination between two arms for each initialization. Compared to JACO PICK-PUSH-PLACE,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "score": 1.0, + "content": "this task requires that the two arms synchronize their movements and perform more micro-level", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 354, + 231, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 231, + 366 + ], + "score": 1.0, + "content": "adjustments to their behaviors.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 299, + 506, + 366 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "Primitives skills. There are two pre-trained primitive skills available to each arm: Picking up and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 381, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 394 + ], + "score": 1.0, + "content": "Placing towards arm. Picking up is same as described in Section 4.2. Placing towards arm learns to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "move a small block (half size of the block used in the composite task) in the hand towards the robotic", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 403, + 501, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 501, + 415 + ], + "score": 1.0, + "content": "arm and then place it on the table. The agent fails if it cannot place the block to the target location.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 370, + 506, + 415 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 420, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 106, + 421, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 506, + 433 + ], + "score": 1.0, + "content": "Composite task. The JACO BAR-MOVING task requires the two arms to work very closely together.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 431, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 506, + 444 + ], + "score": 1.0, + "content": "For example, the Picking up skill of both arms should be synchronized when they start to lift the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "bar and two arms require to lift the bar while maintaining the relative position between them since", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 453, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 506, + 466 + ], + "score": 1.0, + "content": "they are connected by holding the bar. The modular framework without explicit coordination of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "skills (Modular) can synchronize the execution of picking, moving, and placing. But the inability", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "to micro-adjust the movement of the other arm causes instability of bar picking and moving. This", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "results in degraded success rates compared to the modular framework with explicit coordination.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 497, + 488, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 488, + 509 + ], + "score": 1.0, + "content": "Meanwhile, all baselines without pre-defined primitive skills fail to learn JACO BAR-MOVING.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 421, + 506, + 509 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 523, + 177, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 178, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 178, + 535 + ], + "score": 1.0, + "content": "4.4 ANT PUSH", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "We developed a multi-ant environment, ANT-PUSH, inspired from Nachum et al. (2019), simulated in", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "the MuJoCo (Todorov et al., 2012) physics engine. We use the ant model in OpenAI Gym (Brockman", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "et al., 2016). In this environment, two ants need to push a large object toward a green target place,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 578, + 495, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 495, + 590 + ], + "score": 1.0, + "content": "collaborating with each other to keep the angle of the object as stable as possible (see Figure 4c).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 545, + 506, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "Primitives skills. We train walking skills of an ant agent in 4 directions: up, down, left, and right.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "During primitive skill training, a block (half size of the block used in the composite task) and an", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "ant agent are randomly placed. Pushing the block gives an additional reward to the agent, which", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "prevents an ant to avoid the block. The learned primitive skills have different speed and trajectories", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 636, + 294, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 294, + 652 + ], + "score": 1.0, + "content": "conditioned on the latent behavior embedding.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 594, + 506, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 108, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 108, + 655, + 262, + 667 + ], + "score": 1.0, + "content": "Composite task. Our method achieves", + "type": "text" + }, + { + "bbox": [ + 262, + 655, + 289, + 666 + ], + "score": 0.87, + "content": "3 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "success rate on ANT PUSH task while all baselines fail", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 666, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 677 + ], + "score": 1.0, + "content": "to compose primitive skills as shown in Figure 5c and Table 1. The poor performance of RL, MARL,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "RL-SBD, and MARL-SBD baselines shows the difficulty of credit assignment between agents, which", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 687, + 507, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 507, + 700 + ], + "score": 1.0, + "content": "leads one of the ants moves toward a block and pushes it but another ant does not move. Moreover,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "the Modular baseline with primitive skills also fails to learn the pushing task. This result illustrates", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "the importance of coordination of agents, which helps synchronizing and controlling the velocities of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 461, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 461, + 733 + ], + "score": 1.0, + "content": "both ant agents to push the block toward the goal position while maintaining its rotation.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 655, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 93, + 501, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 93, + 501, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 93, + 501, + 186 + ], + "spans": [ + { + "bbox": [ + 110, + 93, + 501, + 186 + ], + "score": 0.953, + "type": "image", + "image_path": "99d31a6b5e96ecb62faf4db19c419b513aea19dd229cabd50fa233595361425f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 93, + 501, + 124.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 124.0, + 501, + 155.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 155.0, + 501, + 186.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 195, + 301, + 217 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 195, + 302, + 207 + ], + "spans": [ + { + "bbox": [ + 107, + 195, + 302, + 207 + ], + "score": 1.0, + "content": "Figure 6: Learning curves of our method with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 205, + 302, + 218 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 227, + 218 + ], + "score": 1.0, + "content": "different diversity coefficients", + "type": "text" + }, + { + "bbox": [ + 227, + 207, + 239, + 217 + ], + "score": 0.87, + "content": "\\lambda _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 205, + 302, + 218 + ], + "score": 1.0, + "content": "on ANT PUSH.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.0 + }, + { + "type": "image_caption", + "bbox": [ + 309, + 195, + 504, + 217 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 506, + 206 + ], + "score": 1.0, + "content": "Figure 7: Success rates of our method with dif-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 309, + 205, + 496, + 218 + ], + "spans": [ + { + "bbox": [ + 309, + 205, + 335, + 218 + ], + "score": 1.0, + "content": "ferent", + "type": "text" + }, + { + "bbox": [ + 335, + 206, + 355, + 217 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 205, + 496, + 218 + ], + "score": 1.0, + "content": "coefficients on Jaco environments.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.0 + } + ], + "index": 4.0 + }, + { + "type": "title", + "bbox": [ + 109, + 241, + 319, + 251 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 321, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 321, + 253 + ], + "score": 1.0, + "content": "4.5 EFFECT OF DIVERSITY OF PRIMITIVE SKILLS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "To analyze the effect of the diversity of primitive skills, we compare our model with primitive", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 306, + 285 + ], + "score": 1.0, + "content": "skills trained with different diversity coefficients", + "type": "text" + }, + { + "bbox": [ + 306, + 272, + 427, + 284 + ], + "score": 0.92, + "content": "\\lambda _ { 2 } = \\{ 0 . 0 , 0 . 0 5 , 0 . 1 , 0 . 5 , 1 . 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "in Equation (4) on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 374, + 296 + ], + "score": 1.0, + "content": "ANT PUSH. Figure 6 shows that with small diversity coefficients", + "type": "text" + }, + { + "bbox": [ + 374, + 284, + 444, + 295 + ], + "score": 0.91, + "content": "\\lambda _ { 2 } = \\{ 0 . 0 5 , 0 . 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 282, + 505, + 296 + ], + "score": 1.0, + "content": ", the agent can", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 294, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 441, + 307 + ], + "score": 1.0, + "content": "control detailed behaviors of primitive skills while primitive skills without diversity", + "type": "text" + }, + { + "bbox": [ + 442, + 295, + 473, + 305 + ], + "score": 0.87, + "content": "{ \\lambda } _ { 2 } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 294, + 506, + 307 + ], + "score": 1.0, + "content": ") cannot", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "be coordinated. The meta policy tries to synchronize two ant agents’ positions and velocities by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "score": 1.0, + "content": "switching primitive skills, but it cannot achieve proper coordination without diversified skills. On", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 277, + 340 + ], + "score": 1.0, + "content": "the other hand, large diversity coefficients", + "type": "text" + }, + { + "bbox": [ + 277, + 327, + 342, + 339 + ], + "score": 0.93, + "content": "\\lambda _ { 2 } = \\bar { \\{ 0 . 5 , 1 . 0 \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "make the primitive skills often focus on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "demonstrating diverse behaviors and fail to achieve the goals of the skills. Hence, these primitive", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "skills do not have enough functionality to solve the target task. The diversity coefficient needs to be", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 360, + 490, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 490, + 372 + ], + "score": 1.0, + "content": "carefully chosen to acquire primitive skills with good performance as well as diverse behaviors.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 386, + 321, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 321, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 301, + 401 + ], + "score": 1.0, + "content": "4.6 EFFECT OF SKILL SELECTION INTERVAL", + "type": "text" + }, + { + "bbox": [ + 302, + 387, + 321, + 397 + ], + "score": 0.85, + "content": "T _ { l o w }", + "type": "inline_equation" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 379, + 420 + ], + "score": 1.0, + "content": "To analyze the effect of the skill selection interval hyperparameter", + "type": "text" + }, + { + "bbox": [ + 379, + 408, + 399, + 418 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 406, + 506, + 420 + ], + "score": 1.0, + "content": ", we compare our method", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 418, + 507, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 159, + 431 + ], + "score": 1.0, + "content": "trained with", + "type": "text" + }, + { + "bbox": [ + 160, + 418, + 254, + 430 + ], + "score": 0.91, + "content": "T _ { l o w } = \\{ 1 , 2 , 3 , 5 , 1 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 418, + 507, + 431 + ], + "score": 1.0, + "content": "on Jaco environments. The success rate curves in Figure 7", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 429, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 207, + 442 + ], + "score": 1.0, + "content": "demonstrate that smaller", + "type": "text" + }, + { + "bbox": [ + 207, + 430, + 226, + 440 + ], + "score": 0.89, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 429, + 506, + 442 + ], + "score": 1.0, + "content": "values in range [1, 3] lead to better performance. This can be because", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 440, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 452 + ], + "score": 1.0, + "content": "the agent can realize more flexible skill coordination by adjusting the behavior embedding frequently.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 200, + 469 + ], + "score": 1.0, + "content": "In addition to the fixed", + "type": "text" + }, + { + "bbox": [ + 200, + 457, + 220, + 468 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "values, we also consider the variation of our method in which the skill", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 467, + 507, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 507, + 481 + ], + "score": 1.0, + "content": "behavior embedding is only sampled when the meta policy updates its skill selection. Concretely,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 477, + 507, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 477, + 185, + 493 + ], + "score": 1.0, + "content": "we set the value of", + "type": "text" + }, + { + "bbox": [ + 185, + 479, + 205, + 490 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 477, + 269, + 493 + ], + "score": 1.0, + "content": "to 1 but update", + "type": "text" + }, + { + "bbox": [ + 270, + 479, + 322, + 491 + ], + "score": 0.93, + "content": "( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 477, + 354, + 493 + ], + "score": 1.0, + "content": "only if", + "type": "text" + }, + { + "bbox": [ + 354, + 478, + 484, + 491 + ], + "score": 0.91, + "content": "\\mathbf { \\bar { \\Phi } } ( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) \\neq ( c _ { t - 1 } ^ { 1 } , \\ldots , c _ { t - 1 } ^ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 477, + 507, + 493 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "score": 1.0, + "content": "observe that in this setting, the meta policy at times switch back and forth between two skills in two", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "score": 1.0, + "content": "consecutive time steps, leading to slightly worse performance compared to our method with small", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 107, + 512, + 126, + 523 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "values. This indicates that the meta policy needs to adjust the behavior embedding in order to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 522, + 308, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 308, + 536 + ], + "score": 1.0, + "content": "optimally coordinate skills of the different agents.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 108, + 551, + 195, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 197, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 197, + 567 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "In this paper, we propose a modular framework with skill coordination to tackle challenges of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "composition of sub-skills with multiple agents. Specifically, we use entropy maximization with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "mutual information maximization to train controllable primitive skills with diverse behaviors. To", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "coordinate learned primitive skills, the meta policy predicts not only the skill to execute for each agent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "(end-effector) but also the behavior embedding that controls the chosen primitive skill’s behavior.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "The experimental results on robotic manipulation and locomotion tasks demonstrate that the proposed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 642, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 656 + ], + "score": 1.0, + "content": "framework is able to efficiently learn primitive skills with diverse behaviors and coordinate multiple", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "agents (end-effectors) to solve challenging cooperative control tasks. Acquiring skills without", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 666, + 493, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 493, + 677 + ], + "score": 1.0, + "content": "supervision and extending our method to a visual domain are exciting directions for future work.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 108, + 691, + 200, + 700 + ], + "lines": [ + { + "bbox": [ + 107, + 691, + 200, + 701 + ], + "spans": [ + { + "bbox": [ + 107, + 691, + 200, + 701 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "This project was funded by SKT. The authors would like to thank Karl Pertsch and many members of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 720, + 276, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 276, + 733 + ], + "score": 1.0, + "content": "the USC CLVR lab for helpful discussion.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 93, + 501, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 93, + 501, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 93, + 501, + 186 + ], + "spans": [ + { + "bbox": [ + 110, + 93, + 501, + 186 + ], + "score": 0.953, + "type": "image", + "image_path": "99d31a6b5e96ecb62faf4db19c419b513aea19dd229cabd50fa233595361425f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 93, + 501, + 124.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 124.0, + 501, + 155.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 155.0, + 501, + 186.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 195, + 301, + 217 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 195, + 302, + 207 + ], + "spans": [ + { + "bbox": [ + 107, + 195, + 302, + 207 + ], + "score": 1.0, + "content": "Figure 6: Learning curves of our method with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 205, + 302, + 218 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 227, + 218 + ], + "score": 1.0, + "content": "different diversity coefficients", + "type": "text" + }, + { + "bbox": [ + 227, + 207, + 239, + 217 + ], + "score": 0.87, + "content": "\\lambda _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 205, + 302, + 218 + ], + "score": 1.0, + "content": "on ANT PUSH.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.0 + }, + { + "type": "image_caption", + "bbox": [ + 309, + 195, + 504, + 217 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 506, + 206 + ], + "score": 1.0, + "content": "Figure 7: Success rates of our method with dif-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 309, + 205, + 496, + 218 + ], + "spans": [ + { + "bbox": [ + 309, + 205, + 335, + 218 + ], + "score": 1.0, + "content": "ferent", + "type": "text" + }, + { + "bbox": [ + 335, + 206, + 355, + 217 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 205, + 496, + 218 + ], + "score": 1.0, + "content": "coefficients on Jaco environments.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.0 + } + ], + "index": 4.0 + }, + { + "type": "title", + "bbox": [ + 109, + 241, + 319, + 251 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 321, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 321, + 253 + ], + "score": 1.0, + "content": "4.5 EFFECT OF DIVERSITY OF PRIMITIVE SKILLS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "To analyze the effect of the diversity of primitive skills, we compare our model with primitive", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 306, + 285 + ], + "score": 1.0, + "content": "skills trained with different diversity coefficients", + "type": "text" + }, + { + "bbox": [ + 306, + 272, + 427, + 284 + ], + "score": 0.92, + "content": "\\lambda _ { 2 } = \\{ 0 . 0 , 0 . 0 5 , 0 . 1 , 0 . 5 , 1 . 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "in Equation (4) on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 374, + 296 + ], + "score": 1.0, + "content": "ANT PUSH. Figure 6 shows that with small diversity coefficients", + "type": "text" + }, + { + "bbox": [ + 374, + 284, + 444, + 295 + ], + "score": 0.91, + "content": "\\lambda _ { 2 } = \\{ 0 . 0 5 , 0 . 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 282, + 505, + 296 + ], + "score": 1.0, + "content": ", the agent can", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 294, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 441, + 307 + ], + "score": 1.0, + "content": "control detailed behaviors of primitive skills while primitive skills without diversity", + "type": "text" + }, + { + "bbox": [ + 442, + 295, + 473, + 305 + ], + "score": 0.87, + "content": "{ \\lambda } _ { 2 } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 294, + 506, + 307 + ], + "score": 1.0, + "content": ") cannot", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "be coordinated. The meta policy tries to synchronize two ant agents’ positions and velocities by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 506, + 329 + ], + "score": 1.0, + "content": "switching primitive skills, but it cannot achieve proper coordination without diversified skills. On", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 326, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 277, + 340 + ], + "score": 1.0, + "content": "the other hand, large diversity coefficients", + "type": "text" + }, + { + "bbox": [ + 277, + 327, + 342, + 339 + ], + "score": 0.93, + "content": "\\lambda _ { 2 } = \\bar { \\{ 0 . 5 , 1 . 0 \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 326, + 506, + 340 + ], + "score": 1.0, + "content": "make the primitive skills often focus on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "demonstrating diverse behaviors and fail to achieve the goals of the skills. Hence, these primitive", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "skills do not have enough functionality to solve the target task. The diversity coefficient needs to be", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 360, + 490, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 490, + 372 + ], + "score": 1.0, + "content": "carefully chosen to acquire primitive skills with good performance as well as diverse behaviors.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 260, + 506, + 372 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 386, + 321, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 321, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 301, + 401 + ], + "score": 1.0, + "content": "4.6 EFFECT OF SKILL SELECTION INTERVAL", + "type": "text" + }, + { + "bbox": [ + 302, + 387, + 321, + 397 + ], + "score": 0.85, + "content": "T _ { l o w }", + "type": "inline_equation" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 407, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 105, + 406, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 379, + 420 + ], + "score": 1.0, + "content": "To analyze the effect of the skill selection interval hyperparameter", + "type": "text" + }, + { + "bbox": [ + 379, + 408, + 399, + 418 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 406, + 506, + 420 + ], + "score": 1.0, + "content": ", we compare our method", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 418, + 507, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 159, + 431 + ], + "score": 1.0, + "content": "trained with", + "type": "text" + }, + { + "bbox": [ + 160, + 418, + 254, + 430 + ], + "score": 0.91, + "content": "T _ { l o w } = \\{ 1 , 2 , 3 , 5 , 1 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 418, + 507, + 431 + ], + "score": 1.0, + "content": "on Jaco environments. The success rate curves in Figure 7", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 429, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 207, + 442 + ], + "score": 1.0, + "content": "demonstrate that smaller", + "type": "text" + }, + { + "bbox": [ + 207, + 430, + 226, + 440 + ], + "score": 0.89, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 429, + 506, + 442 + ], + "score": 1.0, + "content": "values in range [1, 3] lead to better performance. This can be because", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 440, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 452 + ], + "score": 1.0, + "content": "the agent can realize more flexible skill coordination by adjusting the behavior embedding frequently.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 406, + 507, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 200, + 469 + ], + "score": 1.0, + "content": "In addition to the fixed", + "type": "text" + }, + { + "bbox": [ + 200, + 457, + 220, + 468 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "values, we also consider the variation of our method in which the skill", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 467, + 507, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 507, + 481 + ], + "score": 1.0, + "content": "behavior embedding is only sampled when the meta policy updates its skill selection. Concretely,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 477, + 507, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 477, + 185, + 493 + ], + "score": 1.0, + "content": "we set the value of", + "type": "text" + }, + { + "bbox": [ + 185, + 479, + 205, + 490 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 477, + 269, + 493 + ], + "score": 1.0, + "content": "to 1 but update", + "type": "text" + }, + { + "bbox": [ + 270, + 479, + 322, + 491 + ], + "score": 0.93, + "content": "( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 477, + 354, + 493 + ], + "score": 1.0, + "content": "only if", + "type": "text" + }, + { + "bbox": [ + 354, + 478, + 484, + 491 + ], + "score": 0.91, + "content": "\\mathbf { \\bar { \\Phi } } ( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) \\neq ( c _ { t - 1 } ^ { 1 } , \\ldots , c _ { t - 1 } ^ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 477, + 507, + 493 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "score": 1.0, + "content": "observe that in this setting, the meta policy at times switch back and forth between two skills in two", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "score": 1.0, + "content": "consecutive time steps, leading to slightly worse performance compared to our method with small", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 107, + 512, + 126, + 523 + ], + "score": 0.9, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "values. This indicates that the meta policy needs to adjust the behavior embedding in order to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 522, + 308, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 308, + 536 + ], + "score": 1.0, + "content": "optimally coordinate skills of the different agents.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 457, + 507, + 536 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 551, + 195, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 197, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 197, + 567 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 577, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "In this paper, we propose a modular framework with skill coordination to tackle challenges of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 506, + 601 + ], + "score": 1.0, + "content": "composition of sub-skills with multiple agents. Specifically, we use entropy maximization with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "mutual information maximization to train controllable primitive skills with diverse behaviors. To", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "coordinate learned primitive skills, the meta policy predicts not only the skill to execute for each agent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "(end-effector) but also the behavior embedding that controls the chosen primitive skill’s behavior.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "The experimental results on robotic manipulation and locomotion tasks demonstrate that the proposed", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 642, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 656 + ], + "score": 1.0, + "content": "framework is able to efficiently learn primitive skills with diverse behaviors and coordinate multiple", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 667 + ], + "score": 1.0, + "content": "agents (end-effectors) to solve challenging cooperative control tasks. Acquiring skills without", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 666, + 493, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 493, + 677 + ], + "score": 1.0, + "content": "supervision and extending our method to a visual domain are exciting directions for future work.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 577, + 506, + 677 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 691, + 200, + 700 + ], + "lines": [ + { + "bbox": [ + 107, + 691, + 200, + 701 + ], + "spans": [ + { + "bbox": [ + 107, + 691, + 200, + 701 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "This project was funded by SKT. The authors would like to thank Karl Pertsch and many members of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 720, + 276, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 276, + 733 + ], + "score": 1.0, + "content": "the USC CLVR lab for helpful discussion.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 100, + 504, + 124 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "score": 1.0, + "content": "Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with policy", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 110, + 442, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 110, + 442, + 124 + ], + "score": 1.0, + "content": "sketches. In International Conference on Machine Learning, pp. 166–175, 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 105, + 131, + 504, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 131, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 144 + ], + "score": 1.0, + "content": "Pierre-Luc Bacon, Jean Harb, and Doina Precup. The option-critic architecture. In Association for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 142, + 313, + 155 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 313, + 155 + ], + "score": 1.0, + "content": "the Advancement of Artificial Intelligence, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 105, + 162, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 505, + 175 + ], + "score": 1.0, + "content": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 173, + 414, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 414, + 186 + ], + "score": 1.0, + "content": "Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 507, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 507, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 507, + 207 + ], + "score": 1.0, + "content": "Prafulla Dhariwal, Christopher Hesse, Oleg Klimov, Alex Nichol, Matthias Plappert, Alec Radford,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 205, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 205, + 507, + 217 + ], + "score": 1.0, + "content": "John Schulman, Szymon Sidor, Yuhuai Wu, and Peter Zhokhov. Openai baselines. https:", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 216, + 319, + 227 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 319, + 227 + ], + "score": 1.0, + "content": "//github.com/openai/baselines, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 506, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "score": 1.0, + "content": "Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 506, + 261 + ], + "score": 1.0, + "content": "need: Learning skills without a reward function. In International Conference on Learning", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 258, + 489, + 270 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 489, + 270 + ], + "score": 1.0, + "content": "Representations, 2019. URL https://openreview.net/forum?id=SJx63jRqFm.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 506, + 312 + ], + "lines": [ + { + "bbox": [ + 106, + 278, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 506, + 290 + ], + "score": 1.0, + "content": "Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 116, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "Counterfactual multi-agent policy gradients. In Association for the Advancement of Artificial", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 299, + 193, + 312 + ], + "spans": [ + { + "bbox": [ + 116, + 299, + 193, + 312 + ], + "score": 1.0, + "content": "Intelligence, 2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 319, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 504, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 504, + 331 + ], + "score": 1.0, + "content": "Kevin Frans, Jonathan Ho, Xi Chen, Pieter Abbeel, and John Schulman. META LEARNING", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 331, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 116, + 331, + 506, + 343 + ], + "score": 1.0, + "content": "SHARED HIERARCHIES. In International Conference on Learning Representations, 2018. URL", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 343, + 367, + 355 + ], + "spans": [ + { + "bbox": [ + 117, + 343, + 302, + 355 + ], + "score": 1.0, + "content": "https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 302, + 343, + 309, + 352 + ], + "score": 0.42, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 343, + 367, + 355 + ], + "score": 1.0, + "content": "SyX0IeWAW.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 506, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "score": 1.0, + "content": "Dibya Ghosh, Avi Singh, Aravind Rajeswaran, Vikash Kumar, and Sergey Levine. Divide-and-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 373, + 507, + 385 + ], + "spans": [ + { + "bbox": [ + 115, + 373, + 507, + 385 + ], + "score": 1.0, + "content": "conquer reinforcement learning. In International Conference on Learning Representations, 2018.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 383, + 390, + 397 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 325, + 397 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 325, + 386, + 331, + 394 + ], + "score": 0.65, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 383, + 390, + 397 + ], + "score": 1.0, + "content": "rJwelMbR-.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 403, + 507, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 507, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 507, + 417 + ], + "score": 1.0, + "content": "Anirudh Goyal, Shagun Sodhani, Jonathan Binas, Xue Bin Peng, Sergey Levine, and Yoshua Bengio.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 115, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "Reinforcement learning with competitive ensembles of information-constrained primitives. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 426, + 509, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 426, + 509, + 440 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, 2020. URL https://openreview.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 438, + 259, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 189, + 450 + ], + "score": 1.0, + "content": "net/forum?id", + "type": "text" + }, + { + "bbox": [ + 189, + 439, + 195, + 447 + ], + "score": 0.54, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 438, + 259, + 450 + ], + "score": 1.0, + "content": "ryxgJTEYDr.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 504, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "Jayesh K Gupta, Maxim Egorov, and Mykel Kochenderfer. Cooperative multi-agent control using", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 116, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "deep reinforcement learning. In International Conference on Autonomous Agents and Multi-Agent", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 479, + 223, + 491 + ], + "spans": [ + { + "bbox": [ + 115, + 479, + 223, + 491 + ], + "score": 1.0, + "content": "Systems, pp. 66–83, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 505, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "deep energy-based policies. In International Conference on Machine Learning, pp. 1352–1361,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 520, + 142, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 520, + 142, + 533 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 506, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 555 + ], + "score": 1.0, + "content": "Tuomas Haarnoja, Vitchyr Pong, Aurick Zhou, Murtaza Dalal, Pieter Abbeel, and Sergey Levine.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 552, + 507, + 565 + ], + "spans": [ + { + "bbox": [ + 117, + 552, + 507, + 565 + ], + "score": 1.0, + "content": "Composable deep reinforcement learning for robotic manipulation. In IEEE International Confer-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 563, + 350, + 576 + ], + "spans": [ + { + "bbox": [ + 117, + 563, + 350, + 576 + ], + "score": 1.0, + "content": "ence on Robotics and Automation, pp. 6244–6251, 2018a.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 583, + 506, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 115, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "maximum entropy deep reinforcement learning with a stochastic actor. In International Conference", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 605, + 302, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 302, + 617 + ], + "score": 1.0, + "content": "on Machine Learning, pp. 1856–1865, 2018b.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 104, + 625, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 636, + 431, + 649 + ], + "spans": [ + { + "bbox": [ + 116, + 636, + 431, + 649 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 7254–7264, 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 108, + 656, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 670 + ], + "score": 1.0, + "content": "Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 116, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 679, + 369, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 369, + 691 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 3675–3683, 2016.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Youngwoon Lee, Shao-Hua Sun, Sriram Somasundaram, Edward Hu, and Joseph J. Lim. Com-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "posing complex skills by learning transition policies. In International Conference on Learning", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 721, + 489, + 734 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 419, + 734 + ], + "score": 1.0, + "content": "Representations, 2019. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 419, + 722, + 426, + 730 + ], + "score": 0.64, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 721, + 489, + 734 + ], + "score": 1.0, + "content": "rygrBhC5tQ.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 100, + 504, + 124 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "score": 1.0, + "content": "Jacob Andreas, Dan Klein, and Sergey Levine. Modular multitask reinforcement learning with policy", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 110, + 442, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 110, + 442, + 124 + ], + "score": 1.0, + "content": "sketches. In International Conference on Machine Learning, pp. 166–175, 2017.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 99, + 505, + 124 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 131, + 504, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 131, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 144 + ], + "score": 1.0, + "content": "Pierre-Luc Bacon, Jean Harb, and Doina Precup. The option-critic architecture. In Association for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 142, + 313, + 155 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 313, + 155 + ], + "score": 1.0, + "content": "the Advancement of Artificial Intelligence, 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 131, + 505, + 155 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 162, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 505, + 175 + ], + "score": 1.0, + "content": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 173, + 414, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 414, + 186 + ], + "score": 1.0, + "content": "Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 162, + 505, + 186 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 193, + 507, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 507, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 507, + 207 + ], + "score": 1.0, + "content": "Prafulla Dhariwal, Christopher Hesse, Oleg Klimov, Alex Nichol, Matthias Plappert, Alec Radford,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 205, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 205, + 507, + 217 + ], + "score": 1.0, + "content": "John Schulman, Szymon Sidor, Yuhuai Wu, and Peter Zhokhov. Openai baselines. https:", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 216, + 319, + 227 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 319, + 227 + ], + "score": 1.0, + "content": "//github.com/openai/baselines, 2017.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 193, + 507, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 235, + 506, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 249 + ], + "score": 1.0, + "content": "Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. Diversity is all you", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 246, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 506, + 261 + ], + "score": 1.0, + "content": "need: Learning skills without a reward function. In International Conference on Learning", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 258, + 489, + 270 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 489, + 270 + ], + "score": 1.0, + "content": "Representations, 2019. URL https://openreview.net/forum?id=SJx63jRqFm.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 235, + 506, + 270 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 506, + 312 + ], + "lines": [ + { + "bbox": [ + 106, + 278, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 506, + 290 + ], + "score": 1.0, + "content": "Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 289, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 116, + 289, + 505, + 302 + ], + "score": 1.0, + "content": "Counterfactual multi-agent policy gradients. In Association for the Advancement of Artificial", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 299, + 193, + 312 + ], + "spans": [ + { + "bbox": [ + 116, + 299, + 193, + 312 + ], + "score": 1.0, + "content": "Intelligence, 2018.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 278, + 506, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 319, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 320, + 504, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 504, + 331 + ], + "score": 1.0, + "content": "Kevin Frans, Jonathan Ho, Xi Chen, Pieter Abbeel, and John Schulman. META LEARNING", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 331, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 116, + 331, + 506, + 343 + ], + "score": 1.0, + "content": "SHARED HIERARCHIES. In International Conference on Learning Representations, 2018. URL", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 343, + 367, + 355 + ], + "spans": [ + { + "bbox": [ + 117, + 343, + 302, + 355 + ], + "score": 1.0, + "content": "https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 302, + 343, + 309, + 352 + ], + "score": 0.42, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 343, + 367, + 355 + ], + "score": 1.0, + "content": "SyX0IeWAW.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 320, + 506, + 355 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 506, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "score": 1.0, + "content": "Dibya Ghosh, Avi Singh, Aravind Rajeswaran, Vikash Kumar, and Sergey Levine. Divide-and-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 373, + 507, + 385 + ], + "spans": [ + { + "bbox": [ + 115, + 373, + 507, + 385 + ], + "score": 1.0, + "content": "conquer reinforcement learning. In International Conference on Learning Representations, 2018.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 383, + 390, + 397 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 325, + 397 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 325, + 386, + 331, + 394 + ], + "score": 0.65, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 383, + 390, + 397 + ], + "score": 1.0, + "content": "rJwelMbR-.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 362, + 507, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 403, + 507, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 507, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 507, + 417 + ], + "score": 1.0, + "content": "Anirudh Goyal, Shagun Sodhani, Jonathan Binas, Xue Bin Peng, Sergey Levine, and Yoshua Bengio.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 115, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "Reinforcement learning with competitive ensembles of information-constrained primitives. In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 426, + 509, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 426, + 509, + 440 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, 2020. URL https://openreview.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 438, + 259, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 189, + 450 + ], + "score": 1.0, + "content": "net/forum?id", + "type": "text" + }, + { + "bbox": [ + 189, + 439, + 195, + 447 + ], + "score": 0.54, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 438, + 259, + 450 + ], + "score": 1.0, + "content": "ryxgJTEYDr.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 404, + 509, + 450 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 504, + 491 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "Jayesh K Gupta, Maxim Egorov, and Mykel Kochenderfer. Cooperative multi-agent control using", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 116, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "deep reinforcement learning. In International Conference on Autonomous Agents and Multi-Agent", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 479, + 223, + 491 + ], + "spans": [ + { + "bbox": [ + 115, + 479, + 223, + 491 + ], + "score": 1.0, + "content": "Systems, pp. 66–83, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 456, + 506, + 491 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 505, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "deep energy-based policies. In International Conference on Machine Learning, pp. 1352–1361,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 520, + 142, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 520, + 142, + 533 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 498, + 506, + 533 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 506, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 555 + ], + "score": 1.0, + "content": "Tuomas Haarnoja, Vitchyr Pong, Aurick Zhou, Murtaza Dalal, Pieter Abbeel, and Sergey Levine.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 117, + 552, + 507, + 565 + ], + "spans": [ + { + "bbox": [ + 117, + 552, + 507, + 565 + ], + "score": 1.0, + "content": "Composable deep reinforcement learning for robotic manipulation. In IEEE International Confer-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 563, + 350, + 576 + ], + "spans": [ + { + "bbox": [ + 117, + 563, + 350, + 576 + ], + "score": 1.0, + "content": "ence on Robotics and Automation, pp. 6244–6251, 2018a.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 540, + 507, + 576 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 583, + 506, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 115, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "maximum entropy deep reinforcement learning with a stochastic actor. In International Conference", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 605, + 302, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 302, + 617 + ], + "score": 1.0, + "content": "on Machine Learning, pp. 1856–1865, 2018b.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 582, + 506, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 625, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 636, + 431, + 649 + ], + "spans": [ + { + "bbox": [ + 116, + 636, + 431, + 649 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 7254–7264, 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 625, + 506, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 656, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 670 + ], + "score": 1.0, + "content": "Tejas D Kulkarni, Karthik Narasimhan, Ardavan Saeedi, and Josh Tenenbaum. Hierarchical deep", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 116, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "reinforcement learning: Integrating temporal abstraction and intrinsic motivation. In Advances in", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 679, + 369, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 679, + 369, + 691 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 3675–3683, 2016.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 655, + 506, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Youngwoon Lee, Shao-Hua Sun, Sriram Somasundaram, Edward Hu, and Joseph J. Lim. Com-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "posing complex skills by learning transition policies. In International Conference on Learning", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 721, + 489, + 734 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 419, + 734 + ], + "score": 1.0, + "content": "Representations, 2019. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 419, + 722, + 426, + 730 + ], + "score": 0.64, + "content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 721, + 489, + 734 + ], + "score": 1.0, + "content": "rygrBhC5tQ.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 106, + 698, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Andrew Levy, Robert Platt, and Kate Saenko. Hierarchical reinforcement learning with hindsight. In", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 507, + 106 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, 2019. URL https://openreview.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 259, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 188, + 117 + ], + "score": 1.0, + "content": "net/forum?id", + "type": "text" + }, + { + "bbox": [ + 189, + 106, + 195, + 114 + ], + "score": 0.51, + "content": "{ . } = { }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 105, + 259, + 117 + ], + "score": 1.0, + "content": "ryzECoAcY7.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 122, + 504, + 156 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 288, + 157 + ], + "score": 1.0, + "content": "Processing Systems, pp. 6379–6390, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 104, + 162, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 504, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 504, + 174 + ], + "score": 1.0, + "content": "Pattie Maes and Rodney A Brooks. Learning to coordinate behaviors. In Association for the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 173, + 399, + 185 + ], + "spans": [ + { + "bbox": [ + 114, + 173, + 399, + 185 + ], + "score": 1.0, + "content": "Advancement of Artificial Intelligence, volume 90, pp. 796–802, 1990.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 191, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "Josh Merel, Arun Ahuja, Vu Pham, Saran Tunyasuvunakool, Siqi Liu, Dhruva Tirumala, Nicolas", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 202, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 506, + 214 + ], + "score": 1.0, + "content": "Heess, and Greg Wayne. Hierarchical visuomotor control of humanoids. In International Confer-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 497, + 226 + ], + "score": 1.0, + "content": "ence on Learning Representations, 2019. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 497, + 214, + 505, + 223 + ], + "score": 0.27, + "content": "=", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 225, + 181, + 235 + ], + "spans": [ + { + "bbox": [ + 116, + 225, + 181, + 235 + ], + "score": 1.0, + "content": "BJfYvo09Y7.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 108, + 241, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "Ofir Nachum, Shixiang Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 115, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "reinforcement learning. In Advances in Neural Information Processing Systems, pp. 3303–3313,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 262, + 143, + 276 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 143, + 276 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 105, + 281, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "Ofir Nachum, Michael Ahn, Hugo Ponte, Shixiang Gu, and Vikash Kumar. Multi-agent manipulation", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 293, + 457, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 457, + 305 + ], + "score": 1.0, + "content": "via locomotion using hierarchical sim2real. In Conference on Robot Learning, 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "score": 1.0, + "content": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 320, + 507, + 336 + ], + "spans": [ + { + "bbox": [ + 115, + 320, + 507, + 336 + ], + "score": 1.0, + "content": "multi-task deep reinforcement learning. In International Conference on Machine Learning, pp.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 333, + 191, + 344 + ], + "spans": [ + { + "bbox": [ + 116, + 333, + 191, + 344 + ], + "score": 1.0, + "content": "2661–2670, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 350, + 505, + 384 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 364 + ], + "score": 1.0, + "content": "Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 116, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 372, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 482, + 386 + ], + "score": 1.0, + "content": "PyTorch. In Advances in Neural Information Processing Systems Autodiff Workshop, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 507, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 507, + 405 + ], + "score": 1.0, + "content": "Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, and Jun Wang.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "Multiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 413, + 407, + 425 + ], + "spans": [ + { + "bbox": [ + 116, + 413, + 407, + 425 + ], + "score": 1.0, + "content": "to play starcraft combat games. arXiv preprint arXiv:1703.10069, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 444 + ], + "score": 1.0, + "content": "Xue Bin Peng, Michael Chang, Grace Zhang, Pieter Abbeel, and Sergey Levine. Mcp: Learning", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 116, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "composable hierarchical control with multiplicative compositional policies. Advances in Neural", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 452, + 274, + 465 + ], + "spans": [ + { + "bbox": [ + 116, + 452, + 274, + 465 + ], + "score": 1.0, + "content": "Information Processing Systems, 2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 469, + 506, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 470, + 507, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 507, + 483 + ], + "score": 1.0, + "content": "Ahmed H. Qureshi, Jacob J. Johnson, Yuzhe Qin, Taylor Henderson, Byron Boots, and Michael C.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 482, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 116, + 482, + 506, + 493 + ], + "score": 1.0, + "content": "Yip. Composing task-agnostic policies with deep reinforcement learning. In International Confer-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 497, + 505 + ], + "score": 1.0, + "content": "ence on Learning Representations, 2020. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 497, + 493, + 505, + 502 + ], + "score": 0.3, + "content": "=", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 504, + 181, + 515 + ], + "spans": [ + { + "bbox": [ + 116, + 504, + 181, + 515 + ], + "score": 1.0, + "content": "H1ezFREtwH.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "Martin Riedmiller, Roland Hafner, Thomas Lampe, Michael Neunert, Jonas Degrave, Tom Van de", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 116, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "Wiele, Volodymyr Mnih, Nicolas Heess, and Jost Tobias Springenberg. Learning by playing -", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 542, + 507, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 542, + 507, + 556 + ], + "score": 1.0, + "content": "solving sparse reward tasks from scratch. In International Conference on Machine Learning, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 504, + 583 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "Jurgen Schmidhuber. ¨ Towards compositional learning with dynamic neural networks. Inst. fur¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 572, + 189, + 583 + ], + "spans": [ + { + "bbox": [ + 116, + 572, + 189, + 583 + ], + "score": 1.0, + "content": "Informatik, 1990.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 116, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "continuous control using generalized advantage estimation. In International Conference on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 612, + 249, + 624 + ], + "spans": [ + { + "bbox": [ + 115, + 612, + 249, + 624 + ], + "score": 1.0, + "content": "Learning Representations, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 629, + 504, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 644 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 641, + 380, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 641, + 380, + 653 + ], + "score": 1.0, + "content": "optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 106, + 658, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 669, + 432, + 683 + ], + "spans": [ + { + "bbox": [ + 115, + 669, + 432, + 683 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 2244–2252, 2016.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zambaldi, Max", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl Tuyls, et al. Value-decomposition", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "networks for cooperative multi-agent learning based on team reward. In International Conference", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 721, + 405, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 405, + 733 + ], + "score": 1.0, + "content": "on Autonomous Agents and Multi-Agent Systems, pp. 2085–2087, 2018.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Andrew Levy, Robert Platt, and Kate Saenko. Hierarchical reinforcement learning with hindsight. In", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 507, + 106 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, 2019. URL https://openreview.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 259, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 188, + 117 + ], + "score": 1.0, + "content": "net/forum?id", + "type": "text" + }, + { + "bbox": [ + 189, + 106, + 195, + 114 + ], + "score": 0.51, + "content": "{ . } = { }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 105, + 259, + 117 + ], + "score": 1.0, + "content": "ryzECoAcY7.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 81, + 507, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 122, + 504, + 156 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 288, + 157 + ], + "score": 1.0, + "content": "Processing Systems, pp. 6379–6390, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 122, + 505, + 157 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 162, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 504, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 504, + 174 + ], + "score": 1.0, + "content": "Pattie Maes and Rodney A Brooks. Learning to coordinate behaviors. In Association for the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 173, + 399, + 185 + ], + "spans": [ + { + "bbox": [ + 114, + 173, + 399, + 185 + ], + "score": 1.0, + "content": "Advancement of Artificial Intelligence, volume 90, pp. 796–802, 1990.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 162, + 504, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 191, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "Josh Merel, Arun Ahuja, Vu Pham, Saran Tunyasuvunakool, Siqi Liu, Dhruva Tirumala, Nicolas", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 202, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 506, + 214 + ], + "score": 1.0, + "content": "Heess, and Greg Wayne. Hierarchical visuomotor control of humanoids. In International Confer-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 497, + 226 + ], + "score": 1.0, + "content": "ence on Learning Representations, 2019. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 497, + 214, + 505, + 223 + ], + "score": 0.27, + "content": "=", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 225, + 181, + 235 + ], + "spans": [ + { + "bbox": [ + 116, + 225, + 181, + 235 + ], + "score": 1.0, + "content": "BJfYvo09Y7.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 191, + 506, + 235 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 241, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 254 + ], + "score": 1.0, + "content": "Ofir Nachum, Shixiang Shane Gu, Honglak Lee, and Sergey Levine. Data-efficient hierarchical", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 115, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "reinforcement learning. In Advances in Neural Information Processing Systems, pp. 3303–3313,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 262, + 143, + 276 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 143, + 276 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 241, + 506, + 276 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 281, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "Ofir Nachum, Michael Ahn, Hugo Ponte, Shixiang Gu, and Vikash Kumar. Multi-agent manipulation", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 293, + 457, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 457, + 305 + ], + "score": 1.0, + "content": "via locomotion using hierarchical sim2real. In Conference on Robot Learning, 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 281, + 505, + 305 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "score": 1.0, + "content": "Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 320, + 507, + 336 + ], + "spans": [ + { + "bbox": [ + 115, + 320, + 507, + 336 + ], + "score": 1.0, + "content": "multi-task deep reinforcement learning. In International Conference on Machine Learning, pp.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 333, + 191, + 344 + ], + "spans": [ + { + "bbox": [ + 116, + 333, + 191, + 344 + ], + "score": 1.0, + "content": "2661–2670, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 310, + 507, + 344 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 350, + 505, + 384 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 364 + ], + "score": 1.0, + "content": "Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 116, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 372, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 482, + 386 + ], + "score": 1.0, + "content": "PyTorch. In Advances in Neural Information Processing Systems Autodiff Workshop, 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 349, + 506, + 386 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 507, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 507, + 405 + ], + "score": 1.0, + "content": "Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, and Jun Wang.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "Multiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 413, + 407, + 425 + ], + "spans": [ + { + "bbox": [ + 116, + 413, + 407, + 425 + ], + "score": 1.0, + "content": "to play starcraft combat games. arXiv preprint arXiv:1703.10069, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 389, + 507, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 444 + ], + "score": 1.0, + "content": "Xue Bin Peng, Michael Chang, Grace Zhang, Pieter Abbeel, and Sergey Levine. Mcp: Learning", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 116, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "composable hierarchical control with multiplicative compositional policies. Advances in Neural", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 452, + 274, + 465 + ], + "spans": [ + { + "bbox": [ + 116, + 452, + 274, + 465 + ], + "score": 1.0, + "content": "Information Processing Systems, 2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 428, + 505, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 469, + 506, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 470, + 507, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 507, + 483 + ], + "score": 1.0, + "content": "Ahmed H. Qureshi, Jacob J. Johnson, Yuzhe Qin, Taylor Henderson, Byron Boots, and Michael C.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 482, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 116, + 482, + 506, + 493 + ], + "score": 1.0, + "content": "Yip. Composing task-agnostic policies with deep reinforcement learning. In International Confer-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 497, + 505 + ], + "score": 1.0, + "content": "ence on Learning Representations, 2020. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 497, + 493, + 505, + 502 + ], + "score": 0.3, + "content": "=", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 504, + 181, + 515 + ], + "spans": [ + { + "bbox": [ + 116, + 504, + 181, + 515 + ], + "score": 1.0, + "content": "H1ezFREtwH.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 470, + 507, + 515 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "Martin Riedmiller, Roland Hafner, Thomas Lampe, Michael Neunert, Jonas Degrave, Tom Van de", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 116, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "Wiele, Volodymyr Mnih, Nicolas Heess, and Jost Tobias Springenberg. Learning by playing -", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 542, + 507, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 542, + 507, + 556 + ], + "score": 1.0, + "content": "solving sparse reward tasks from scratch. In International Conference on Machine Learning, 2018.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 520, + 507, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 504, + 583 + ], + "lines": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "Jurgen Schmidhuber. ¨ Towards compositional learning with dynamic neural networks. Inst. fur¨", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 572, + 189, + 583 + ], + "spans": [ + { + "bbox": [ + 116, + 572, + 189, + 583 + ], + "score": 1.0, + "content": "Informatik, 1990.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 560, + 505, + 583 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "score": 1.0, + "content": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 116, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "continuous control using generalized advantage estimation. In International Conference on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 612, + 249, + 624 + ], + "spans": [ + { + "bbox": [ + 115, + 612, + 249, + 624 + ], + "score": 1.0, + "content": "Learning Representations, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 590, + 506, + 624 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 629, + 504, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 644 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 641, + 380, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 641, + 380, + 653 + ], + "score": 1.0, + "content": "optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 628, + 505, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 658, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 669, + 432, + 683 + ], + "spans": [ + { + "bbox": [ + 115, + 669, + 432, + 683 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 2244–2252, 2016.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 658, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zambaldi, Max", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl Tuyls, et al. Value-decomposition", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "networks for cooperative multi-agent learning based on team reward. In International Conference", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 721, + 405, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 405, + 733 + ], + "score": 1.0, + "content": "on Autonomous Agents and Multi-Agent Systems, pp. 2085–2087, 2018.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 106, + 688, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Richard S Sutton. Temporal credit assignment in reinforcement learning. PhD thesis, University of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 205, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 205, + 106 + ], + "score": 1.0, + "content": "Massachusetts, 1984.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 505, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 125 + ], + "score": 1.0, + "content": "Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 506, + 136 + ], + "score": 1.0, + "content": "for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181–211, 1999.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 153, + 502, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 153, + 502, + 166 + ], + "score": 1.0, + "content": "In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033, 2012.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Richard S Sutton. Temporal credit assignment in reinforcement learning. PhD thesis, University of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 205, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 205, + 106 + ], + "score": 1.0, + "content": "Massachusetts, 1984.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 505, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 125 + ], + "score": 1.0, + "content": "Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 506, + 136 + ], + "score": 1.0, + "content": "for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181–211, 1999.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 506, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 153, + 502, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 153, + 502, + 166 + ], + "score": 1.0, + "content": "In IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033, 2012.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 142, + 506, + 166 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 253, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 254, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 254, + 96 + ], + "score": 1.0, + "content": "A ENVIRONMENT DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 506, + 129 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "The details of observation spaces, action spaces, number of agents, and episode lengths are described", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 409, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 409, + 131 + ], + "score": 1.0, + "content": "in Table 2. All units in this section are in meters unless otherwise specified.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "table", + "bbox": [ + 143, + 141, + 467, + 222 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 143, + 141, + 467, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 141, + 467, + 222 + ], + "spans": [ + { + "bbox": [ + 143, + 141, + 467, + 222 + ], + "score": 0.967, + "html": "
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
Observation Space8888100
- Robot observation626282
- Object observation262618
Action Space181816
Number of Agents222
Episode length150100200
", + "type": "table", + "image_path": "28b86d0a1f27a2f11bd196806df5f0e97f1f1883aa4bb3f5ac42b0e9dd4cf42d.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 143, + 141, + 467, + 168.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 143, + 168.0, + 467, + 195.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 143, + 195.0, + 467, + 222.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 247, + 230, + 364, + 242 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 246, + 230, + 364, + 242 + ], + "spans": [ + { + "bbox": [ + 246, + 230, + 364, + 242 + ], + "score": 1.0, + "content": "Table 2: Environment details", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 5.0 + }, + { + "type": "title", + "bbox": [ + 108, + 263, + 263, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 264, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 264, + 276 + ], + "score": 1.0, + "content": "A.1 ENVIRONMENT DESCRIPTIONS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 284, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "In both Jaco environments, the robot works on a table with size (1.6, 1.6) and top center position", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 295, + 503, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 150, + 307 + ], + "score": 0.86, + "content": "( 0 , 0 , 0 . 8 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 295, + 332, + 308 + ], + "score": 1.0, + "content": ". The two Jaco arms are initialized at positions", + "type": "text" + }, + { + "bbox": [ + 332, + 296, + 412, + 307 + ], + "score": 0.84, + "content": "( - 0 . 1 6 , - 0 . 1 6 , 1 . 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 295, + 430, + 308 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 430, + 295, + 503, + 307 + ], + "score": 0.65, + "content": "( \\bar { - } 0 . 1 6 , 0 . \\bar { 2 } 4 , 1 . 2 )", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 309, + 320 + ], + "score": 1.0, + "content": "Left arm and right arm objects are initialized around", + "type": "text" + }, + { + "bbox": [ + 310, + 307, + 369, + 318 + ], + "score": 0.35, + "content": "( 0 . 3 , 0 . 2 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 306, + 392, + 320 + ], + "score": 1.0, + "content": "and (", + "type": "text" + }, + { + "bbox": [ + 392, + 307, + 455, + 318 + ], + "score": 0.43, + "content": "0 . 3 , - 0 . 2 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "respectively", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "in all primitive training and composite task training environments, with small random position and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 329, + 192, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 192, + 342 + ], + "score": 1.0, + "content": "rotation perturbation.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 345, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "In the Jaco Pick-Push-Place task, the right jaco arm needs to pick up the object and place it into the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "container initialized at the other side of the table. Success is defined by contact between the object", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 367, + 262, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 262, + 380 + ], + "score": 1.0, + "content": "and the inner top side of the container.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 507, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 507, + 396 + ], + "score": 1.0, + "content": "In the Jaco Bar-Moving task, the two Jaco arms need to together pick the long bar up by height of 0.7,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 394, + 507, + 409 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 507, + 409 + ], + "score": 1.0, + "content": "move it towards the arms by distance of 0.15, and place it back on the table. Success is defined by (1)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "the bar being placed within 0.04 away from the desired destination both in height and in xy-position", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 418, + 327, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 327, + 429 + ], + "score": 1.0, + "content": "and (2) the bar having been picked 0.7 above the table.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "In the Ant Push task, the two ant agents need to push a big box together to the goal position. The", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 176, + 457 + ], + "score": 1.0, + "content": "box has a size of", + "type": "text" + }, + { + "bbox": [ + 176, + 446, + 240, + 456 + ], + "score": 0.91, + "content": "8 . 0 \\times 1 . 6 \\times 1 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 444, + 410, + 457 + ], + "score": 1.0, + "content": ". The distance between ants and the box is", + "type": "text" + }, + { + "bbox": [ + 411, + 446, + 437, + 456 + ], + "score": 0.58, + "content": "2 0 \\mathrm { c m }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "and the distance", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 232, + 469 + ], + "score": 1.0, + "content": "between the box and the goal is", + "type": "text" + }, + { + "bbox": [ + 232, + 457, + 258, + 467 + ], + "score": 0.65, + "content": "3 0 \\mathrm { c m }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 456, + 346, + 469 + ], + "score": 1.0, + "content": ". Initial positions have", + "type": "text" + }, + { + "bbox": [ + 347, + 457, + 367, + 467 + ], + "score": 0.39, + "content": "1 \\mathrm { c m }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "of randomness and the agent has a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 466, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 480 + ], + "score": 1.0, + "content": "randomness of 0.01 in each joint. The task is considered as success when both the distances between", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 478, + 338, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 314, + 491 + ], + "score": 1.0, + "content": "left and right end of the box and the goal are within", + "type": "text" + }, + { + "bbox": [ + 314, + 479, + 335, + 488 + ], + "score": 0.55, + "content": "5 \\mathrm { c m }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 478, + 338, + 491 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 504, + 207, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 208, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 208, + 517 + ], + "score": 1.0, + "content": "A.2 REWARD DESIGN", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 504, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 266, + 538 + ], + "score": 1.0, + "content": "For every task, we add a control penalty,", + "type": "text" + }, + { + "bbox": [ + 266, + 525, + 325, + 538 + ], + "score": 0.93, + "content": "- 0 . 0 0 1 * \\| a \\| ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 524, + 505, + 538 + ], + "score": 1.0, + "content": ", to regularize the magnitude of actions where", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 536, + 280, + 550 + ], + "spans": [ + { + "bbox": [ + 107, + 539, + 113, + 546 + ], + "score": 0.73, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 536, + 280, + 550 + ], + "score": 1.0, + "content": "is a torque action performed by an agent.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 553, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "Jaco Pick: To help the agent learn to reach, pick, and hold the picked object, we provide dense", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 565, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 577 + ], + "score": 1.0, + "content": "reward to the agent defined by the weighted sum of pick reward, gripper-to-cube distance reward,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 576, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 587 + ], + "score": 1.0, + "content": "cube position and quaternion stability reward, hold duration reward, success reward, and robot control", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 586, + 209, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 209, + 599 + ], + "score": 1.0, + "content": "reward. More concretely,", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 604, + 479, + 636 + ], + "lines": [ + { + "bbox": [ + 133, + 604, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 133, + 604, + 479, + 636 + ], + "score": 0.87, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { p i c k } \\cdot ( z _ { b o x } - z _ { i n i t } ) + \\lambda _ { d i s t } \\cdot \\mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \\lambda _ { p o s } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { i n i t } ) + } \\\\ & { \\qquad \\lambda _ { q u a t } \\cdot \\mathrm { a b s } ( \\Delta _ { q u a t } ) + \\lambda _ { h o l d } \\cdot t _ { h o l d } + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "7218e69723138229a1241e98230e1161cf031d597d01762af334ae8915a44c47.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 133, + 604, + 479, + 614.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 133, + 614.6666666666666, + 479, + 625.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 133, + 625.3333333333333, + 479, + 635.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 131, + 654 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 640, + 505, + 654 + ], + "score": 0.52, + "content": "\\lambda _ { p i c k } = 5 0 0 , \\lambda _ { d i s t } = 1 0 0 , \\lambda _ { p o s } = 1 0 0 0 , \\lambda _ { q u a t } = 1 0 0 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 0 0 0 , \\lambda _ { d i s t }", + "type": "inline_equation", + "image_path": "b0a62c13177af340afd116bb3df7689325aa92456495ed3f5d1df39e3720ef3f.jpg" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 649, + 149, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 652, + 145, + 663 + ], + "score": 0.9, + "content": "1 \\times 1 0 ^ { - \\hat { 4 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 649, + 149, + 666 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 669, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 104, + 668, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 104, + 668, + 505, + 683 + ], + "score": 1.0, + "content": "Jaco Place: Reward for place primitive is defined by the weighted sum of xy-distance reward, height", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 681, + 438, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 438, + 692 + ], + "score": 1.0, + "content": "reward (larger when cube close to floor), success reward, and robot control reward.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "interline_equation", + "bbox": [ + 106, + 697, + 476, + 734 + ], + "lines": [ + { + "bbox": [ + 128, + 697, + 476, + 734 + ], + "spans": [ + { + "bbox": [ + 128, + 697, + 476, + 734 + ], + "score": 0.43, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { x y } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b o x } , p _ { g o a l } ) + \\lambda _ { z } \\cdot | z _ { b o x } - z _ { g o a l } | + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , } \\\\ & { \\mathrm { ~ e \\lambda _ { x y } = 5 0 0 , } \\lambda _ { z } = 5 0 0 , \\lambda _ { s u c c e s s } = 5 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 } . } \\end{array}", + "type": "interline_equation", + "image_path": "b98953b49bbde6752e8d53228b0cb112649e82f672d717587232fa8845567e86.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 106, + 697, + 476, + 709.3333333333334 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 106, + 709.3333333333334, + 476, + 721.6666666666667 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 106, + 721.6666666666667, + 476, + 734.0000000000001 + ], + "spans": [], + "index": 41 + } + ] + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 253, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 254, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 254, + 96 + ], + "score": 1.0, + "content": "A ENVIRONMENT DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 506, + 129 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "score": 1.0, + "content": "The details of observation spaces, action spaces, number of agents, and episode lengths are described", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 409, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 409, + 131 + ], + "score": 1.0, + "content": "in Table 2. All units in this section are in meters unless otherwise specified.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 106, + 505, + 131 + ] + }, + { + "type": "table", + "bbox": [ + 143, + 141, + 467, + 222 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 143, + 141, + 467, + 222 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 141, + 467, + 222 + ], + "spans": [ + { + "bbox": [ + 143, + 141, + 467, + 222 + ], + "score": 0.967, + "html": "
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
Observation Space8888100
- Robot observation626282
- Object observation262618
Action Space181816
Number of Agents222
Episode length150100200
", + "type": "table", + "image_path": "28b86d0a1f27a2f11bd196806df5f0e97f1f1883aa4bb3f5ac42b0e9dd4cf42d.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 143, + 141, + 467, + 168.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 143, + 168.0, + 467, + 195.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 143, + 195.0, + 467, + 222.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 247, + 230, + 364, + 242 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 246, + 230, + 364, + 242 + ], + "spans": [ + { + "bbox": [ + 246, + 230, + 364, + 242 + ], + "score": 1.0, + "content": "Table 2: Environment details", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 5.0 + }, + { + "type": "title", + "bbox": [ + 108, + 263, + 263, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 264, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 264, + 276 + ], + "score": 1.0, + "content": "A.1 ENVIRONMENT DESCRIPTIONS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 284, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "In both Jaco environments, the robot works on a table with size (1.6, 1.6) and top center position", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 295, + 503, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 150, + 307 + ], + "score": 0.86, + "content": "( 0 , 0 , 0 . 8 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 295, + 332, + 308 + ], + "score": 1.0, + "content": ". The two Jaco arms are initialized at positions", + "type": "text" + }, + { + "bbox": [ + 332, + 296, + 412, + 307 + ], + "score": 0.84, + "content": "( - 0 . 1 6 , - 0 . 1 6 , 1 . 2 )", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 295, + 430, + 308 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 430, + 295, + 503, + 307 + ], + "score": 0.65, + "content": "( \\bar { - } 0 . 1 6 , 0 . \\bar { 2 } 4 , 1 . 2 )", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 309, + 320 + ], + "score": 1.0, + "content": "Left arm and right arm objects are initialized around", + "type": "text" + }, + { + "bbox": [ + 310, + 307, + 369, + 318 + ], + "score": 0.35, + "content": "( 0 . 3 , 0 . 2 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 306, + 392, + 320 + ], + "score": 1.0, + "content": "and (", + "type": "text" + }, + { + "bbox": [ + 392, + 307, + 455, + 318 + ], + "score": 0.43, + "content": "0 . 3 , - 0 . 2 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "respectively", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "in all primitive training and composite task training environments, with small random position and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 329, + 192, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 192, + 342 + ], + "score": 1.0, + "content": "rotation perturbation.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 284, + 505, + 342 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 345, + 505, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "In the Jaco Pick-Push-Place task, the right jaco arm needs to pick up the object and place it into the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "container initialized at the other side of the table. Success is defined by contact between the object", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 367, + 262, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 262, + 380 + ], + "score": 1.0, + "content": "and the inner top side of the container.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 345, + 505, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 384, + 505, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 507, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 507, + 396 + ], + "score": 1.0, + "content": "In the Jaco Bar-Moving task, the two Jaco arms need to together pick the long bar up by height of 0.7,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 394, + 507, + 409 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 507, + 409 + ], + "score": 1.0, + "content": "move it towards the arms by distance of 0.15, and place it back on the table. Success is defined by (1)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "the bar being placed within 0.04 away from the desired destination both in height and in xy-position", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 418, + 327, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 327, + 429 + ], + "score": 1.0, + "content": "and (2) the bar having been picked 0.7 above the table.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 384, + 507, + 429 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "In the Ant Push task, the two ant agents need to push a big box together to the goal position. The", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 176, + 457 + ], + "score": 1.0, + "content": "box has a size of", + "type": "text" + }, + { + "bbox": [ + 176, + 446, + 240, + 456 + ], + "score": 0.91, + "content": "8 . 0 \\times 1 . 6 \\times 1 . 6", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 444, + 410, + 457 + ], + "score": 1.0, + "content": ". The distance between ants and the box is", + "type": "text" + }, + { + "bbox": [ + 411, + 446, + 437, + 456 + ], + "score": 0.58, + "content": "2 0 \\mathrm { c m }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "and the distance", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 232, + 469 + ], + "score": 1.0, + "content": "between the box and the goal is", + "type": "text" + }, + { + "bbox": [ + 232, + 457, + 258, + 467 + ], + "score": 0.65, + "content": "3 0 \\mathrm { c m }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 456, + 346, + 469 + ], + "score": 1.0, + "content": ". Initial positions have", + "type": "text" + }, + { + "bbox": [ + 347, + 457, + 367, + 467 + ], + "score": 0.39, + "content": "1 \\mathrm { c m }", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "of randomness and the agent has a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 466, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 480 + ], + "score": 1.0, + "content": "randomness of 0.01 in each joint. The task is considered as success when both the distances between", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 478, + 338, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 314, + 491 + ], + "score": 1.0, + "content": "left and right end of the box and the goal are within", + "type": "text" + }, + { + "bbox": [ + 314, + 479, + 335, + 488 + ], + "score": 0.55, + "content": "5 \\mathrm { c m }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 478, + 338, + 491 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 433, + 506, + 491 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 504, + 207, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 208, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 208, + 517 + ], + "score": 1.0, + "content": "A.2 REWARD DESIGN", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 504, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 266, + 538 + ], + "score": 1.0, + "content": "For every task, we add a control penalty,", + "type": "text" + }, + { + "bbox": [ + 266, + 525, + 325, + 538 + ], + "score": 0.93, + "content": "- 0 . 0 0 1 * \\| a \\| ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 524, + 505, + 538 + ], + "score": 1.0, + "content": ", to regularize the magnitude of actions where", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 536, + 280, + 550 + ], + "spans": [ + { + "bbox": [ + 107, + 539, + 113, + 546 + ], + "score": 0.73, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 536, + 280, + 550 + ], + "score": 1.0, + "content": "is a torque action performed by an agent.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 524, + 505, + 550 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 553, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "Jaco Pick: To help the agent learn to reach, pick, and hold the picked object, we provide dense", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 565, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 506, + 577 + ], + "score": 1.0, + "content": "reward to the agent defined by the weighted sum of pick reward, gripper-to-cube distance reward,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 576, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 587 + ], + "score": 1.0, + "content": "cube position and quaternion stability reward, hold duration reward, success reward, and robot control", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 586, + 209, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 209, + 599 + ], + "score": 1.0, + "content": "reward. More concretely,", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 553, + 506, + 599 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 604, + 479, + 636 + ], + "lines": [ + { + "bbox": [ + 133, + 604, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 133, + 604, + 479, + 636 + ], + "score": 0.87, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { p i c k } \\cdot ( z _ { b o x } - z _ { i n i t } ) + \\lambda _ { d i s t } \\cdot \\mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \\lambda _ { p o s } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { i n i t } ) + } \\\\ & { \\qquad \\lambda _ { q u a t } \\cdot \\mathrm { a b s } ( \\Delta _ { q u a t } ) + \\lambda _ { h o l d } \\cdot t _ { h o l d } + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "7218e69723138229a1241e98230e1161cf031d597d01762af334ae8915a44c47.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 133, + 604, + 479, + 614.6666666666666 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 133, + 614.6666666666666, + 479, + 625.3333333333333 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 133, + 625.3333333333333, + 479, + 635.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 640, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 131, + 654 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 640, + 505, + 654 + ], + "score": 0.52, + "content": "\\lambda _ { p i c k } = 5 0 0 , \\lambda _ { d i s t } = 1 0 0 , \\lambda _ { p o s } = 1 0 0 0 , \\lambda _ { q u a t } = 1 0 0 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 0 0 0 , \\lambda _ { d i s t }", + "type": "inline_equation", + "image_path": "b0a62c13177af340afd116bb3df7689325aa92456495ed3f5d1df39e3720ef3f.jpg" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 649, + 149, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 652, + 145, + 663 + ], + "score": 0.9, + "content": "1 \\times 1 0 ^ { - \\hat { 4 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 649, + 149, + 666 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 639, + 505, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 669, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 104, + 668, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 104, + 668, + 505, + 683 + ], + "score": 1.0, + "content": "Jaco Place: Reward for place primitive is defined by the weighted sum of xy-distance reward, height", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 681, + 438, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 438, + 692 + ], + "score": 1.0, + "content": "reward (larger when cube close to floor), success reward, and robot control reward.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 104, + 668, + 505, + 692 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 106, + 697, + 476, + 734 + ], + "lines": [ + { + "bbox": [ + 128, + 697, + 476, + 734 + ], + "spans": [ + { + "bbox": [ + 128, + 697, + 476, + 734 + ], + "score": 0.43, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { x y } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b o x } , p _ { g o a l } ) + \\lambda _ { z } \\cdot | z _ { b o x } - z _ { g o a l } | + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , } \\\\ & { \\mathrm { ~ e \\lambda _ { x y } = 5 0 0 , } \\lambda _ { z } = 5 0 0 , \\lambda _ { s u c c e s s } = 5 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 } . } \\end{array}", + "type": "interline_equation", + "image_path": "b98953b49bbde6752e8d53228b0cb112649e82f672d717587232fa8845567e86.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 106, + 697, + 476, + 709.3333333333334 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 106, + 709.3333333333334, + 476, + 721.6666666666667 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 106, + 721.6666666666667, + 476, + 734.0000000000001 + ], + "spans": [], + "index": 41 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Jaco Push: Reward for push primitive is defined by the weighted sum of gripper reaching reward,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "box-to-destination distance reward, quaternion stability reward, hold duration reward, success reward,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 209, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 209, + 116 + ], + "score": 1.0, + "content": "and robot control reward.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 131, + 459, + 163 + ], + "lines": [ + { + "bbox": [ + 153, + 131, + 459, + 163 + ], + "spans": [ + { + "bbox": [ + 153, + 131, + 459, + 163 + ], + "score": 0.91, + "content": "\\begin{array} { r } { R ( s ) = \\lambda _ { r e a c h i n g } \\cdot \\mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \\lambda _ { p o s } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { d e s t } ) + \\phantom { x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x } } \\\\ { \\lambda _ { q u a t } \\cdot \\mathsf { a b s } ( \\Delta _ { q u a t } ) + \\lambda _ { h o l d } \\cdot t _ { h o l d } + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { s u c c e s s } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , \\phantom { x x x x x x x x x x x x x x } } \\end{array}", + "type": "interline_equation", + "image_path": "0c68286af7bd7b82024e785edfec8cf4090d625996595f95caf4394882238f62.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 153, + 131, + 459, + 141.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 153, + 141.66666666666666, + 459, + 152.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 153, + 152.33333333333331, + 459, + 162.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 181, + 499, + 194 + ], + "lines": [ + { + "bbox": [ + 133, + 181, + 499, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 181, + 499, + 194 + ], + "score": 0.65, + "content": "\\lambda _ { r e a c h i n g } = 1 0 0 , \\lambda _ { p o s } = 5 0 0 , \\lambda _ { q u a t } = 3 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 } .", + "type": "interline_equation", + "image_path": "4ca888393fa5398d33bb9e915293fc545ae6a42f71899ee55aace23e9f7c6ce1.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 133, + 181, + 499, + 194 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 104, + 196, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 506, + 212 + ], + "score": 1.0, + "content": "Jaco Pick-Push-Place: Reward for Pick-Push-Place is defined by the weighted sum of gripper", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "contact reward, per-stage reach/pick/push/place rewards, success reward, and control reward. We", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 221, + 277, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 277, + 233 + ], + "score": 1.0, + "content": "tune the reward carefully for all baselines.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 134, + 249, + 479, + 313 + ], + "lines": [ + { + "bbox": [ + 134, + 249, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 134, + 249, + 479, + 313 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { c o n t a c t } \\cdot \\big ( { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { l e f f ~ g r i p p e r ~ r o u c h e s ~ c o n t a i n e r } } } + { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { r i g h t ~ g r i p p e r ~ t o u c h e s ~ b o x } } } \\big ) + } \\\\ & { \\quad \\quad \\lambda _ { r e a c h } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { r e a c h } } } \\cdot \\big ( \\mathrm { d i s t } ( p _ { l e f t . g r i p p e r } , p _ { c o n t a i n e r } ) + \\mathrm { d i s t } ( p _ { r i g h t . g r i p p e r } , p _ { b o x } ) \\big ) + } \\\\ & { \\quad \\quad \\lambda _ { p i c k } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { p i c k } } } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + \\lambda _ { p l a c e } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { p l a c e } } } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + } \\\\ & { \\quad \\quad \\lambda _ { p u s h } \\cdot \\mathrm { d i s t } ( p _ { c o n t a i n e r } , p _ { c o n t a i n e r . t a r g e t } ) + \\lambda _ { s u c c e s s } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { s u c c e s s } } } + \\lambda _ { c t r l } \\cdot \\| a \\| ^ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "b9880c8304a14716c17f00f4bae48746e17f96e2635ba93e6f38d0c38a727515.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 134, + 249, + 479, + 270.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 134, + 270.3333333333333, + 479, + 291.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 134, + 291.66666666666663, + 479, + 312.99999999999994 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 329, + 506, + 374 + ], + "lines": [ + { + "bbox": [ + 104, + 329, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 133, + 344 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 330, + 484, + 343 + ], + "score": 0.82, + "content": "\\lambda _ { r e a c h } = 1 0 , \\lambda _ { c o n t a c t } = 1 0 , \\lambda _ { p i c k } = \\lambda _ { p l a c e } = \\lambda _ { p l a c e } = 1 0 , \\lambda _ { s u c c e s s } = 5 0 , \\lambda _ { c t r l } = 0 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 329, + 506, + 344 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 141, + 353 + ], + "score": 1.0, + "content": "1reaching,", + "type": "text" + }, + { + "bbox": [ + 141, + 342, + 160, + 353 + ], + "score": 0.79, + "content": "\\mathbf { 1 } _ { \\mathrm { p i c k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 341, + 181, + 353 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 182, + 342, + 204, + 353 + ], + "score": 0.87, + "content": "\\mathbf { 1 } _ { \\mathrm { p l a c e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "are indicator functions specifying whether the agent is in reaching, pick or", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "place stage. Agent stages are determined by how many multiples of 25 steps the agent has stepped", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 363, + 220, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 220, + 375 + ], + "score": 1.0, + "content": "through in the environment.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 105, + 379, + 504, + 403 + ], + "lines": [ + { + "bbox": [ + 104, + 378, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 378, + 505, + 394 + ], + "score": 1.0, + "content": "Jaco Bar-Moving: Reward for Bar-Moving is defined by the weighted sum of per-stage", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 391, + 380, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 380, + 403 + ], + "score": 1.0, + "content": "reach/pick/move/place rewards, success reward, and control reward.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 420, + 500, + 468 + ], + "lines": [ + { + "bbox": [ + 112, + 420, + 500, + 468 + ], + "spans": [ + { + "bbox": [ + 112, + 420, + 500, + 468 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { r e a c h } \\cdot { \\bf 1 } _ { r e a c h } \\cdot \\big ( \\mathrm { d i s t } ( p _ { l e f t \\_ g r i p p e r } , p _ { l e f t \\_ h a n d l e } ) + \\mathrm { d i s t } ( p _ { r i g h t \\_ g r i p p e r } , p _ { r i g h t \\_ h a n d l e } ) \\big ) + } \\\\ & { \\qquad \\lambda _ { p i c k } \\cdot { \\bf 1 } _ { \\mathrm { p i c k } } \\cdot \\mathrm { d i s t } ( p _ { b a r } , p _ { b a r \\_ t a r g e t } ) + \\lambda _ { m o v e } \\cdot { \\bf 1 } _ { \\mathrm { p l a e } } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b a r \\_ p b a r \\_ t a r g e t } ) + } \\\\ & { \\qquad \\lambda _ { p l a c e } \\cdot { \\bf 1 } _ { \\mathrm { p l a e } } \\cdot \\mathrm { d i s t } _ { z } ( p _ { b a r \\_ p b a r \\_ t a r g e t } ) + \\lambda _ { s u c e s s } \\cdot { \\bf 1 } _ { \\mathrm { s u c e s s } } + \\lambda _ { c t r l } \\cdot \\| a \\| ^ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "b0cc74c444bcd7c674d571d64c9cf8e175f34b540d056c3ea313f5b20ca868eb.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 112, + 420, + 500, + 436.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 112, + 436.0, + 500, + 452.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 112, + 452.0, + 500, + 468.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 506, + 520 + ], + "lines": [ + { + "bbox": [ + 104, + 483, + 507, + 500 + ], + "spans": [ + { + "bbox": [ + 104, + 483, + 131, + 500 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 486, + 503, + 498 + ], + "score": 0.81, + "content": "\\lambda _ { r e a c h } = 1 0 , \\lambda _ { p i c k } = 3 0 , \\lambda _ { m o v e } = 1 0 0 , \\lambda _ { p l a c e } = 1 0 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 483, + 507, + 500 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 496, + 507, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 124, + 510 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 498, + 144, + 509 + ], + "score": 0.88, + "content": "\\mathbf { 1 } _ { \\mathrm { p i c k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 496, + 163, + 510 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 498, + 185, + 509 + ], + "score": 0.89, + "content": "\\mathbf { 1 } _ { \\mathrm { p l a c e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 496, + 507, + 510 + ], + "score": 1.0, + "content": "are indicator functions specifying whether the agent is in pick or place stage.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 508, + 414, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 414, + 520 + ], + "score": 1.0, + "content": "Agent stages are determined by whether the pick objective is fulfilled or not.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 506, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "Ant Push & Ant Moving: Reward for ANT PUSH is defined by upright, velocity towards the desired", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "direction. We provide a dense reward to encourage the desired locomotion behavior using velocity,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 546, + 249, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 249, + 560 + ], + "score": 1.0, + "content": "stability, and posture, as following:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 576, + 502, + 605 + ], + "lines": [ + { + "bbox": [ + 112, + 576, + 502, + 605 + ], + "spans": [ + { + "bbox": [ + 112, + 576, + 502, + 605 + ], + "score": 0.87, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { v e l } \\cdot \\mathbf { a b s } ( \\Delta x _ { a n t } ) + \\lambda _ { b o x v e l } \\cdot \\mathbf { a b s } ( \\Delta x _ { b o x } ) + \\lambda _ { u p r i g h t } \\cdot \\mathbf { c o s } ( \\theta ) - \\lambda _ { h e i g h t } \\cdot \\mathbf { a b s } ( 0 . 6 - h ) + } \\\\ & { ~ \\lambda _ { g o a l } \\cdot \\mathbf { d i s t } ( p _ { g o a l } , p _ { b o x } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "a1fac56114820ef7a956ab1929340bdc0bd6b2e6f6daaf866ab74929c70ad2de.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 112, + 576, + 502, + 585.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 112, + 585.6666666666666, + 502, + 595.3333333333333 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 112, + 595.3333333333333, + 502, + 604.9999999999999 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 134, + 637 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 623, + 364, + 635 + ], + "score": 0.86, + "content": "\\lambda _ { v e l } = 5 0 , \\lambda _ { b o x v e l } = 2 0 , \\lambda _ { u p r i g h t } = 1 , \\lambda _ { h e i g h t } = 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 621, + 506, + 637 + ], + "score": 1.0, + "content": ". For ANT PUSH, we provide an", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 632, + 479, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 425, + 646 + ], + "score": 1.0, + "content": "additional reward based on distance between the box and the goal position with", + "type": "text" + }, + { + "bbox": [ + 425, + 634, + 476, + 646 + ], + "score": 0.92, + "content": "\\lambda _ { g o a l } = 2 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 632, + 479, + 646 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 107, + 676, + 243, + 689 + ], + "lines": [ + { + "bbox": [ + 105, + 675, + 244, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 244, + 690 + ], + "score": 1.0, + "content": "B EXPERIMENT DETAILS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "We use PyTorch (Paszke et al., 2017) for our implementation and all experiments are conducted on a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 721, + 470, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 470, + 732 + ], + "score": 1.0, + "content": "workstation with Intel Xeon Gold 6154 CPU and 4 NVIDIA GeForce RTX 2080 Ti GPUs.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Jaco Push: Reward for push primitive is defined by the weighted sum of gripper reaching reward,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "box-to-destination distance reward, quaternion stability reward, hold duration reward, success reward,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 209, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 209, + 116 + ], + "score": 1.0, + "content": "and robot control reward.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 506, + 116 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 153, + 131, + 459, + 163 + ], + "lines": [ + { + "bbox": [ + 153, + 131, + 459, + 163 + ], + "spans": [ + { + "bbox": [ + 153, + 131, + 459, + 163 + ], + "score": 0.91, + "content": "\\begin{array} { r } { R ( s ) = \\lambda _ { r e a c h i n g } \\cdot \\mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \\lambda _ { p o s } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { d e s t } ) + \\phantom { x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x } } \\\\ { \\lambda _ { q u a t } \\cdot \\mathsf { a b s } ( \\Delta _ { q u a t } ) + \\lambda _ { h o l d } \\cdot t _ { h o l d } + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { s u c c e s s } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , \\phantom { x x x x x x x x x x x x x x } } \\end{array}", + "type": "interline_equation", + "image_path": "0c68286af7bd7b82024e785edfec8cf4090d625996595f95caf4394882238f62.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 153, + 131, + 459, + 141.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 153, + 141.66666666666666, + 459, + 152.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 153, + 152.33333333333331, + 459, + 162.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 181, + 499, + 194 + ], + "lines": [ + { + "bbox": [ + 133, + 181, + 499, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 181, + 499, + 194 + ], + "score": 0.65, + "content": "\\lambda _ { r e a c h i n g } = 1 0 0 , \\lambda _ { p o s } = 5 0 0 , \\lambda _ { q u a t } = 3 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 } .", + "type": "interline_equation", + "image_path": "4ca888393fa5398d33bb9e915293fc545ae6a42f71899ee55aace23e9f7c6ce1.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 133, + 181, + 499, + 194 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 505, + 232 + ], + "lines": [ + { + "bbox": [ + 104, + 196, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 506, + 212 + ], + "score": 1.0, + "content": "Jaco Pick-Push-Place: Reward for Pick-Push-Place is defined by the weighted sum of gripper", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "contact reward, per-stage reach/pick/push/place rewards, success reward, and control reward. We", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 221, + 277, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 277, + 233 + ], + "score": 1.0, + "content": "tune the reward carefully for all baselines.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 104, + 196, + 506, + 233 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 134, + 249, + 479, + 313 + ], + "lines": [ + { + "bbox": [ + 134, + 249, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 134, + 249, + 479, + 313 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { c o n t a c t } \\cdot \\big ( { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { l e f f ~ g r i p p e r ~ r o u c h e s ~ c o n t a i n e r } } } + { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { r i g h t ~ g r i p p e r ~ t o u c h e s ~ b o x } } } \\big ) + } \\\\ & { \\quad \\quad \\lambda _ { r e a c h } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { r e a c h } } } \\cdot \\big ( \\mathrm { d i s t } ( p _ { l e f t . g r i p p e r } , p _ { c o n t a i n e r } ) + \\mathrm { d i s t } ( p _ { r i g h t . g r i p p e r } , p _ { b o x } ) \\big ) + } \\\\ & { \\quad \\quad \\lambda _ { p i c k } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { p i c k } } } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + \\lambda _ { p l a c e } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { p l a c e } } } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + } \\\\ & { \\quad \\quad \\lambda _ { p u s h } \\cdot \\mathrm { d i s t } ( p _ { c o n t a i n e r } , p _ { c o n t a i n e r . t a r g e t } ) + \\lambda _ { s u c c e s s } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { s u c c e s s } } } + \\lambda _ { c t r l } \\cdot \\| a \\| ^ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "b9880c8304a14716c17f00f4bae48746e17f96e2635ba93e6f38d0c38a727515.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 134, + 249, + 479, + 270.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 134, + 270.3333333333333, + 479, + 291.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 134, + 291.66666666666663, + 479, + 312.99999999999994 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 329, + 506, + 374 + ], + "lines": [ + { + "bbox": [ + 104, + 329, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 133, + 344 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 330, + 484, + 343 + ], + "score": 0.82, + "content": "\\lambda _ { r e a c h } = 1 0 , \\lambda _ { c o n t a c t } = 1 0 , \\lambda _ { p i c k } = \\lambda _ { p l a c e } = \\lambda _ { p l a c e } = 1 0 , \\lambda _ { s u c c e s s } = 5 0 , \\lambda _ { c t r l } = 0 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 329, + 506, + 344 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 141, + 353 + ], + "score": 1.0, + "content": "1reaching,", + "type": "text" + }, + { + "bbox": [ + 141, + 342, + 160, + 353 + ], + "score": 0.79, + "content": "\\mathbf { 1 } _ { \\mathrm { p i c k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 341, + 181, + 353 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 182, + 342, + 204, + 353 + ], + "score": 0.87, + "content": "\\mathbf { 1 } _ { \\mathrm { p l a c e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "are indicator functions specifying whether the agent is in reaching, pick or", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "place stage. Agent stages are determined by how many multiples of 25 steps the agent has stepped", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 363, + 220, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 220, + 375 + ], + "score": 1.0, + "content": "through in the environment.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 329, + 506, + 375 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 379, + 504, + 403 + ], + "lines": [ + { + "bbox": [ + 104, + 378, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 378, + 505, + 394 + ], + "score": 1.0, + "content": "Jaco Bar-Moving: Reward for Bar-Moving is defined by the weighted sum of per-stage", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 391, + 380, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 380, + 403 + ], + "score": 1.0, + "content": "reach/pick/move/place rewards, success reward, and control reward.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 378, + 505, + 403 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 420, + 500, + 468 + ], + "lines": [ + { + "bbox": [ + 112, + 420, + 500, + 468 + ], + "spans": [ + { + "bbox": [ + 112, + 420, + 500, + 468 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { r e a c h } \\cdot { \\bf 1 } _ { r e a c h } \\cdot \\big ( \\mathrm { d i s t } ( p _ { l e f t \\_ g r i p p e r } , p _ { l e f t \\_ h a n d l e } ) + \\mathrm { d i s t } ( p _ { r i g h t \\_ g r i p p e r } , p _ { r i g h t \\_ h a n d l e } ) \\big ) + } \\\\ & { \\qquad \\lambda _ { p i c k } \\cdot { \\bf 1 } _ { \\mathrm { p i c k } } \\cdot \\mathrm { d i s t } ( p _ { b a r } , p _ { b a r \\_ t a r g e t } ) + \\lambda _ { m o v e } \\cdot { \\bf 1 } _ { \\mathrm { p l a e } } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b a r \\_ p b a r \\_ t a r g e t } ) + } \\\\ & { \\qquad \\lambda _ { p l a c e } \\cdot { \\bf 1 } _ { \\mathrm { p l a e } } \\cdot \\mathrm { d i s t } _ { z } ( p _ { b a r \\_ p b a r \\_ t a r g e t } ) + \\lambda _ { s u c e s s } \\cdot { \\bf 1 } _ { \\mathrm { s u c e s s } } + \\lambda _ { c t r l } \\cdot \\| a \\| ^ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "b0cc74c444bcd7c674d571d64c9cf8e175f34b540d056c3ea313f5b20ca868eb.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 112, + 420, + 500, + 436.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 112, + 436.0, + 500, + 452.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 112, + 452.0, + 500, + 468.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 506, + 520 + ], + "lines": [ + { + "bbox": [ + 104, + 483, + 507, + 500 + ], + "spans": [ + { + "bbox": [ + 104, + 483, + 131, + 500 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 486, + 503, + 498 + ], + "score": 0.81, + "content": "\\lambda _ { r e a c h } = 1 0 , \\lambda _ { p i c k } = 3 0 , \\lambda _ { m o v e } = 1 0 0 , \\lambda _ { p l a c e } = 1 0 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 483, + 507, + 500 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 496, + 507, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 124, + 510 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 498, + 144, + 509 + ], + "score": 0.88, + "content": "\\mathbf { 1 } _ { \\mathrm { p i c k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 496, + 163, + 510 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 164, + 498, + 185, + 509 + ], + "score": 0.89, + "content": "\\mathbf { 1 } _ { \\mathrm { p l a c e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 496, + 507, + 510 + ], + "score": 1.0, + "content": "are indicator functions specifying whether the agent is in pick or place stage.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 508, + 414, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 414, + 520 + ], + "score": 1.0, + "content": "Agent stages are determined by whether the pick objective is fulfilled or not.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 104, + 483, + 507, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 506, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 538 + ], + "score": 1.0, + "content": "Ant Push & Ant Moving: Reward for ANT PUSH is defined by upright, velocity towards the desired", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 534, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 506, + 549 + ], + "score": 1.0, + "content": "direction. We provide a dense reward to encourage the desired locomotion behavior using velocity,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 546, + 249, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 249, + 560 + ], + "score": 1.0, + "content": "stability, and posture, as following:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 524, + 506, + 560 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 576, + 502, + 605 + ], + "lines": [ + { + "bbox": [ + 112, + 576, + 502, + 605 + ], + "spans": [ + { + "bbox": [ + 112, + 576, + 502, + 605 + ], + "score": 0.87, + "content": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { v e l } \\cdot \\mathbf { a b s } ( \\Delta x _ { a n t } ) + \\lambda _ { b o x v e l } \\cdot \\mathbf { a b s } ( \\Delta x _ { b o x } ) + \\lambda _ { u p r i g h t } \\cdot \\mathbf { c o s } ( \\theta ) - \\lambda _ { h e i g h t } \\cdot \\mathbf { a b s } ( 0 . 6 - h ) + } \\\\ & { ~ \\lambda _ { g o a l } \\cdot \\mathbf { d i s t } ( p _ { g o a l } , p _ { b o x } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "a1fac56114820ef7a956ab1929340bdc0bd6b2e6f6daaf866ab74929c70ad2de.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 112, + 576, + 502, + 585.6666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 112, + 585.6666666666666, + 502, + 595.3333333333333 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 112, + 595.3333333333333, + 502, + 604.9999999999999 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 646 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 134, + 637 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 623, + 364, + 635 + ], + "score": 0.86, + "content": "\\lambda _ { v e l } = 5 0 , \\lambda _ { b o x v e l } = 2 0 , \\lambda _ { u p r i g h t } = 1 , \\lambda _ { h e i g h t } = 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 621, + 506, + 637 + ], + "score": 1.0, + "content": ". For ANT PUSH, we provide an", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 632, + 479, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 425, + 646 + ], + "score": 1.0, + "content": "additional reward based on distance between the box and the goal position with", + "type": "text" + }, + { + "bbox": [ + 425, + 634, + 476, + 646 + ], + "score": 0.92, + "content": "\\lambda _ { g o a l } = 2 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 632, + 479, + 646 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 621, + 506, + 646 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 676, + 243, + 689 + ], + "lines": [ + { + "bbox": [ + 105, + 675, + 244, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 244, + 690 + ], + "score": 1.0, + "content": "B EXPERIMENT DETAILS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "We use PyTorch (Paszke et al., 2017) for our implementation and all experiments are conducted on a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 721, + 470, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 470, + 732 + ], + "score": 1.0, + "content": "workstation with Intel Xeon Gold 6154 CPU and 4 NVIDIA GeForce RTX 2080 Ti GPUs.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 709, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 220, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 221, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 221, + 94 + ], + "score": 1.0, + "content": "B.1 HYPERPARAMETERS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 180, + 106, + 431, + 219 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 180, + 106, + 431, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 180, + 106, + 431, + 219 + ], + "spans": [ + { + "bbox": [ + 180, + 106, + 431, + 219 + ], + "score": 0.955, + "html": "
ParametersValue
learning rate3e-4
gradient steps50
batch size256
discount factor0.99
target smoothing coefficient0.005
reward scale (SAC)1.0
experience buffer size (# episodes)1000
Tow1 for JACO, 5 for ANT
Nz (dimensionality of z)5
", + "type": "table", + "image_path": "2e0f2d3e912b19281466669264a4f5594f304b9e72c85d327661d9f4d2d9566c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 180, + 106, + 431, + 143.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 180, + 143.66666666666666, + 431, + 181.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 180, + 181.33333333333331, + 431, + 218.99999999999997 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 253, + 227, + 358, + 239 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 252, + 226, + 359, + 240 + ], + "spans": [ + { + "bbox": [ + 252, + 226, + 359, + 240 + ], + "score": 1.0, + "content": "Table 3: Hyperparameters", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 108, + 259, + 252, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 253, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 253, + 271 + ], + "score": 1.0, + "content": "B.2 NETWORK ARCHITECTURES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 506, + 292 + ], + "score": 1.0, + "content": "Actor Networks: In all experiments, we model our actor network for each primitive skill as a 3-layer", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 291, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 505, + 302 + ], + "score": 1.0, + "content": "MLP with hidden layer size 64. The last layer of the MLP is two-headed – one for the mean of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "action distribution and the other for the standard deviation of it. We use ReLU as activation function", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "in hidden layers. We do not apply any activation function for the final output layer. The action", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "distribution output represents per-dimension normal distribution, from which single actions can be", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 334, + 277, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 277, + 347 + ], + "score": 1.0, + "content": "sampled and executed in the environment.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "Critic Networks: The critic network for each primitive skill and meta policy is modeled as a 2-layer", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 362, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 504, + 374 + ], + "score": 1.0, + "content": "MLP with hidden layer size 128. ReLU is used as an activation function in the hidden layers. The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "critic network output is used to assess the value of a given state-action pair, and is trained by fitting", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 384, + 311, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 283, + 396 + ], + "score": 1.0, + "content": "its outputs to the target Q-value clamped by", + "type": "text" + }, + { + "bbox": [ + 283, + 384, + 307, + 395 + ], + "score": 0.88, + "content": "\\pm 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 384, + 311, + 396 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "Meta Policy: The meta policy is modeled as a 3-layer MLP with hidden layer size of 64. Since", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 284, + 425 + ], + "score": 1.0, + "content": "the actions of meta policy are sampled from", + "type": "text" + }, + { + "bbox": [ + 285, + 412, + 295, + 422 + ], + "score": 0.75, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "categorical distributions for each end-effector/agent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 124, + 437 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 423, + 135, + 433 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 422, + 505, + 437 + ], + "score": 1.0, + "content": "normal distributions for behavior embeddings, the output dimension of the meta policy is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 433, + 507, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 176, + 449 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { N } ( m ^ { i } + N _ { z } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 433, + 507, + 452 + ], + "score": 1.0, + "content": ". The meta policy uses ReLU as an activation function for all layers except for the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 446, + 179, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 179, + 461 + ], + "score": 1.0, + "content": "final output layer.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 472, + 216, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 217, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 217, + 484 + ], + "score": 1.0, + "content": "B.3 TRAINING DETAILS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "For all baselines, we train the meta policies using PPO and the low-level policies using SAC. We use", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 444, + 515 + ], + "score": 1.0, + "content": "the same environment configurations, composite task reward definitions, and value of", + "type": "text" + }, + { + "bbox": [ + 444, + 504, + 464, + 514 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "across all", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 513, + 149, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 149, + 526 + ], + "score": 1.0, + "content": "baselines.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "For Jaco tasks, we train a total of 4 primitive skills – right arm pick, right arm place-to-center, right", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 542, + 507, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 507, + 555 + ], + "score": 1.0, + "content": "arm place-towards-arm, and left arm push – to be composed by meta-policy. For Jaco Pick-Push-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "Place, we provide the meta-policy with right arm pick and right arm place-to-center as right arm", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "primitives and left arm push as left arm primitives; for Jaco Bar-Moving, we provide the meta-policy", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 575, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 506, + 587 + ], + "score": 1.0, + "content": "with right arm pick and right arm place-towards-arm as both right and left arm primitives and left", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "arm pick and right arm place-towards-arm as left arm primitives. We obtain left arm primitives for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 597, + 373, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 373, + 610 + ], + "score": 1.0, + "content": "bar-moving task by using the learned right arm primitives directly.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 613, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "score": 1.0, + "content": "To obtain the 4 primitives skills described above, we train right arm pick with diversity coefficient", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 625, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 149, + 636 + ], + "score": 0.89, + "content": "\\lambda _ { 2 } = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 626, + 292, + 637 + ], + "score": 1.0, + "content": "and the other three primitives with", + "type": "text" + }, + { + "bbox": [ + 292, + 625, + 329, + 636 + ], + "score": 0.91, + "content": "\\lambda _ { 2 } = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 626, + 505, + 637 + ], + "score": 1.0, + "content": ". The destination of right arm push is set to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 107, + 636, + 178, + 648 + ], + "score": 0.37, + "content": "( 0 . 3 , - 0 . 0 3 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 636, + 506, + 649 + ], + "score": 1.0, + "content": ", which is slightly left of the center of the table. After pick primitive is trained,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "we train the two right arm place primitives where episodes are initialized by intermediate states", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 658, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 506, + 671 + ], + "score": 1.0, + "content": "of successful right arm pick episodes where the height of the box is larger than 0.94 (0.01 higher", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "than the target pick height). The place destinations for towards-arm and to-center primitives are", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 680, + 326, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 179, + 692 + ], + "score": 0.78, + "content": "( 0 . 1 5 , - 0 . 2 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 680, + 198, + 693 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 199, + 680, + 270, + 691 + ], + "score": 0.68, + "content": "( 0 . 3 , - 0 . 0 2 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 680, + 326, + 693 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 108, + 696, + 504, + 730 + ], + "lines": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 427, + 709 + ], + "score": 1.0, + "content": "For non-modular baselines that incorporates skill behavior diversification, we use", + "type": "text" + }, + { + "bbox": [ + 427, + 697, + 470, + 708 + ], + "score": 0.91, + "content": "\\lambda _ { 2 } = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "for both", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 706, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 505, + 721 + ], + "score": 1.0, + "content": "Jaco Pick-Push-Place and Jaco Bar-Moving because both tasks require picking skills, which can only", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 717, + 250, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 717, + 234, + 731 + ], + "score": 1.0, + "content": "be trained with a small value of", + "type": "text" + }, + { + "bbox": [ + 235, + 719, + 246, + 730 + ], + "score": 0.88, + "content": "\\lambda _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 717, + 250, + 731 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 220, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 221, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 221, + 94 + ], + "score": 1.0, + "content": "B.1 HYPERPARAMETERS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 180, + 106, + 431, + 219 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 180, + 106, + 431, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 180, + 106, + 431, + 219 + ], + "spans": [ + { + "bbox": [ + 180, + 106, + 431, + 219 + ], + "score": 0.955, + "html": "
ParametersValue
learning rate3e-4
gradient steps50
batch size256
discount factor0.99
target smoothing coefficient0.005
reward scale (SAC)1.0
experience buffer size (# episodes)1000
Tow1 for JACO, 5 for ANT
Nz (dimensionality of z)5
", + "type": "table", + "image_path": "2e0f2d3e912b19281466669264a4f5594f304b9e72c85d327661d9f4d2d9566c.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 180, + 106, + 431, + 143.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 180, + 143.66666666666666, + 431, + 181.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 180, + 181.33333333333331, + 431, + 218.99999999999997 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 253, + 227, + 358, + 239 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 252, + 226, + 359, + 240 + ], + "spans": [ + { + "bbox": [ + 252, + 226, + 359, + 240 + ], + "score": 1.0, + "content": "Table 3: Hyperparameters", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 3.0 + }, + { + "type": "title", + "bbox": [ + 108, + 259, + 252, + 270 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 253, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 253, + 271 + ], + "score": 1.0, + "content": "B.2 NETWORK ARCHITECTURES", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 506, + 292 + ], + "score": 1.0, + "content": "Actor Networks: In all experiments, we model our actor network for each primitive skill as a 3-layer", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 291, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 505, + 302 + ], + "score": 1.0, + "content": "MLP with hidden layer size 64. The last layer of the MLP is two-headed – one for the mean of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 301, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 314 + ], + "score": 1.0, + "content": "action distribution and the other for the standard deviation of it. We use ReLU as activation function", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "in hidden layers. We do not apply any activation function for the final output layer. The action", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "distribution output represents per-dimension normal distribution, from which single actions can be", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 334, + 277, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 277, + 347 + ], + "score": 1.0, + "content": "sampled and executed in the environment.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 279, + 506, + 347 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 106, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "Critic Networks: The critic network for each primitive skill and meta policy is modeled as a 2-layer", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 362, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 504, + 374 + ], + "score": 1.0, + "content": "MLP with hidden layer size 128. ReLU is used as an activation function in the hidden layers. The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "critic network output is used to assess the value of a given state-action pair, and is trained by fitting", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 384, + 311, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 283, + 396 + ], + "score": 1.0, + "content": "its outputs to the target Q-value clamped by", + "type": "text" + }, + { + "bbox": [ + 283, + 384, + 307, + 395 + ], + "score": 0.88, + "content": "\\pm 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 384, + 311, + 396 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 351, + 506, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "Meta Policy: The meta policy is modeled as a 3-layer MLP with hidden layer size of 64. Since", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 284, + 425 + ], + "score": 1.0, + "content": "the actions of meta policy are sampled from", + "type": "text" + }, + { + "bbox": [ + 285, + 412, + 295, + 422 + ], + "score": 0.75, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "categorical distributions for each end-effector/agent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 124, + 437 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 423, + 135, + 433 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 422, + 505, + 437 + ], + "score": 1.0, + "content": "normal distributions for behavior embeddings, the output dimension of the meta policy is", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 433, + 507, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 176, + 449 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { N } ( m ^ { i } + N _ { z } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 433, + 507, + 452 + ], + "score": 1.0, + "content": ". The meta policy uses ReLU as an activation function for all layers except for the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 446, + 179, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 179, + 461 + ], + "score": 1.0, + "content": "final output layer.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 401, + 507, + 461 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 472, + 216, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 217, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 217, + 484 + ], + "score": 1.0, + "content": "B.3 TRAINING DETAILS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "For all baselines, we train the meta policies using PPO and the low-level policies using SAC. We use", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 504, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 444, + 515 + ], + "score": 1.0, + "content": "the same environment configurations, composite task reward definitions, and value of", + "type": "text" + }, + { + "bbox": [ + 444, + 504, + 464, + 514 + ], + "score": 0.91, + "content": "T _ { l o w }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 504, + 505, + 515 + ], + "score": 1.0, + "content": "across all", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 513, + 149, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 149, + 526 + ], + "score": 1.0, + "content": "baselines.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 492, + 505, + 526 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "For Jaco tasks, we train a total of 4 primitive skills – right arm pick, right arm place-to-center, right", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 542, + 507, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 507, + 555 + ], + "score": 1.0, + "content": "arm place-towards-arm, and left arm push – to be composed by meta-policy. For Jaco Pick-Push-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "Place, we provide the meta-policy with right arm pick and right arm place-to-center as right arm", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "primitives and left arm push as left arm primitives; for Jaco Bar-Moving, we provide the meta-policy", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 575, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 506, + 587 + ], + "score": 1.0, + "content": "with right arm pick and right arm place-towards-arm as both right and left arm primitives and left", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "arm pick and right arm place-towards-arm as left arm primitives. We obtain left arm primitives for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 597, + 373, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 373, + 610 + ], + "score": 1.0, + "content": "bar-moving task by using the learned right arm primitives directly.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 531, + 507, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 613, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 626 + ], + "score": 1.0, + "content": "To obtain the 4 primitives skills described above, we train right arm pick with diversity coefficient", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 625, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 149, + 636 + ], + "score": 0.89, + "content": "\\lambda _ { 2 } = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 626, + 292, + 637 + ], + "score": 1.0, + "content": "and the other three primitives with", + "type": "text" + }, + { + "bbox": [ + 292, + 625, + 329, + 636 + ], + "score": 0.91, + "content": "\\lambda _ { 2 } = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 626, + 505, + 637 + ], + "score": 1.0, + "content": ". The destination of right arm push is set to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 107, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 107, + 636, + 178, + 648 + ], + "score": 0.37, + "content": "( 0 . 3 , - 0 . 0 3 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 636, + 506, + 649 + ], + "score": 1.0, + "content": ", which is slightly left of the center of the table. After pick primitive is trained,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 506, + 660 + ], + "score": 1.0, + "content": "we train the two right arm place primitives where episodes are initialized by intermediate states", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 658, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 506, + 671 + ], + "score": 1.0, + "content": "of successful right arm pick episodes where the height of the box is larger than 0.94 (0.01 higher", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "than the target pick height). The place destinations for towards-arm and to-center primitives are", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 680, + 326, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 179, + 692 + ], + "score": 0.78, + "content": "( 0 . 1 5 , - 0 . 2 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 680, + 198, + 693 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 199, + 680, + 270, + 691 + ], + "score": 0.68, + "content": "( 0 . 3 , - 0 . 0 2 , 0 . 8 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 680, + 326, + 693 + ], + "score": 1.0, + "content": ", respectively.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 614, + 506, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 696, + 504, + 730 + ], + "lines": [ + { + "bbox": [ + 105, + 696, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 427, + 709 + ], + "score": 1.0, + "content": "For non-modular baselines that incorporates skill behavior diversification, we use", + "type": "text" + }, + { + "bbox": [ + 427, + 697, + 470, + 708 + ], + "score": 0.91, + "content": "\\lambda _ { 2 } = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 696, + 505, + 709 + ], + "score": 1.0, + "content": "for both", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 706, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 706, + 505, + 721 + ], + "score": 1.0, + "content": "Jaco Pick-Push-Place and Jaco Bar-Moving because both tasks require picking skills, which can only", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 717, + 250, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 717, + 234, + 731 + ], + "score": 1.0, + "content": "be trained with a small value of", + "type": "text" + }, + { + "bbox": [ + 235, + 719, + 246, + 730 + ], + "score": 0.88, + "content": "\\lambda _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 717, + 250, + 731 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 696, + 505, + 731 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/ryxB2lBtvH/ryxB2lBtvH_model.json b/parse/train/ryxB2lBtvH/ryxB2lBtvH_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8688bfb272993ca6397c7caa8d5e5bb0b6e3a7cc --- /dev/null +++ b/parse/train/ryxB2lBtvH/ryxB2lBtvH_model.json @@ -0,0 +1,20455 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1391, + 1404, + 1391, + 1404, + 1636, + 298, + 1636 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 398, + 645, + 1302, + 645, + 1302, + 1010, + 398, + 1010 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1650, + 1404, + 1650, + 1404, + 1926, + 298, + 1926 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1130, + 1404, + 1130, + 1404, + 1375, + 298, + 1375 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 301, + 1942, + 1402, + 1942, + 1402, + 2034, + 301, + 2034 + ], + "score": 0.969 + }, + { + "category_id": 0, + "poly": [ + 299, + 219, + 1371, + 219, + 1371, + 324, + 299, + 324 + ], + "score": 0.964 + }, + { + "category_id": 0, + "poly": [ + 303, + 1063, + 573, + 1063, + 573, + 1098, + 303, + 1098 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.877 + }, + { + "category_id": 0, + "poly": [ + 773, + 579, + 927, + 579, + 927, + 613, + 773, + 613 + ], + "score": 0.871 + }, + { + "category_id": 1, + "poly": [ + 318, + 376, + 863, + 376, + 863, + 410, + 318, + 410 + ], + "score": 0.764 + }, + { + "category_id": 1, + "poly": [ + 315, + 411, + 827, + 411, + 827, + 501, + 315, + 501 + ], + "score": 0.753 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.726 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 219.0, + 1373.0, + 219.0, + 1373.0, + 271.0, + 295.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 275.0, + 1124.0, + 275.0, + 1124.0, + 328.0, + 295.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1059.0, + 579.0, + 1059.0, + 579.0, + 1106.0, + 294.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 578.0, + 932.0, + 578.0, + 932.0, + 616.0, + 770.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1427.0, + 294.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1422.0, + 1405.0, + 1422.0, + 1405.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1452.0, + 1404.0, + 1452.0, + 1404.0, + 1486.0, + 293.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1480.0, + 1409.0, + 1480.0, + 1409.0, + 1520.0, + 292.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1509.0, + 1408.0, + 1509.0, + 1408.0, + 1552.0, + 291.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1541.0, + 1408.0, + 1541.0, + 1408.0, + 1579.0, + 291.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1573.0, + 1405.0, + 1573.0, + 1405.0, + 1611.0, + 293.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1606.0, + 1408.0, + 1606.0, + 1408.0, + 1639.0, + 294.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 647.0, + 1304.0, + 647.0, + 1304.0, + 679.0, + 397.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 675.0, + 1306.0, + 675.0, + 1306.0, + 710.0, + 394.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 705.0, + 1306.0, + 705.0, + 1306.0, + 740.0, + 393.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 737.0, + 1305.0, + 737.0, + 1305.0, + 769.0, + 394.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 767.0, + 1306.0, + 767.0, + 1306.0, + 798.0, + 393.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 796.0, + 1306.0, + 796.0, + 1306.0, + 830.0, + 393.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 826.0, + 1306.0, + 826.0, + 1306.0, + 861.0, + 394.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 857.0, + 1306.0, + 857.0, + 1306.0, + 894.0, + 393.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 886.0, + 1306.0, + 886.0, + 1306.0, + 926.0, + 392.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 921.0, + 1306.0, + 921.0, + 1306.0, + 953.0, + 395.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 951.0, + 1306.0, + 951.0, + 1306.0, + 983.0, + 394.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 981.0, + 953.0, + 981.0, + 953.0, + 1012.0, + 393.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1652.0, + 1405.0, + 1652.0, + 1405.0, + 1685.0, + 296.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1682.0, + 1404.0, + 1682.0, + 1404.0, + 1715.0, + 296.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1714.0, + 1405.0, + 1714.0, + 1405.0, + 1746.0, + 296.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 1405.0, + 1741.0, + 1405.0, + 1778.0, + 294.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1770.0, + 1405.0, + 1770.0, + 1405.0, + 1813.0, + 293.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1801.0, + 1408.0, + 1801.0, + 1408.0, + 1840.0, + 293.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1832.0, + 1406.0, + 1832.0, + 1406.0, + 1870.0, + 292.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1864.0, + 1406.0, + 1864.0, + 1406.0, + 1899.0, + 293.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1895.0, + 1076.0, + 1895.0, + 1076.0, + 1928.0, + 296.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1405.0, + 1130.0, + 1405.0, + 1167.0, + 294.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1163.0, + 1409.0, + 1163.0, + 1409.0, + 1196.0, + 294.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1227.0, + 294.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1223.0, + 1404.0, + 1223.0, + 1404.0, + 1256.0, + 294.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1254.0, + 1404.0, + 1254.0, + 1404.0, + 1287.0, + 294.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1280.0, + 1406.0, + 1280.0, + 1406.0, + 1320.0, + 292.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1311.0, + 1406.0, + 1311.0, + 1406.0, + 1351.0, + 292.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1346.0, + 514.0, + 1346.0, + 514.0, + 1381.0, + 294.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 296.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2004.0, + 1404.0, + 2004.0, + 1404.0, + 2037.0, + 297.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 374.0, + 864.0, + 374.0, + 864.0, + 417.0, + 313.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 404.0, + 692.0, + 404.0, + 692.0, + 441.0, + 310.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 437.0, + 693.0, + 437.0, + 693.0, + 469.0, + 313.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 469.0, + 828.0, + 469.0, + 828.0, + 503.0, + 314.0, + 503.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 994, + 1405, + 994, + 1405, + 1423, + 298, + 1423 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1668, + 1405, + 1668, + 1405, + 2034, + 298, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1438, + 1404, + 1438, + 1404, + 1652, + 298, + 1652 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 725, + 1405, + 725, + 1405, + 879, + 298, + 879 + ], + "score": 0.976 + }, + { + "category_id": 4, + "poly": [ + 296, + 450, + 1405, + 450, + 1405, + 665, + 296, + 665 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 301, + 225, + 1398, + 225, + 1398, + 426, + 301, + 426 + ], + "score": 0.966 + }, + { + "category_id": 0, + "poly": [ + 301, + 925, + 581, + 925, + 581, + 960, + 301, + 960 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 816, + 76, + 816, + 104, + 298, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.712 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.146 + }, + { + "category_id": 13, + "poly": [ + 1216, + 576, + 1246, + 576, + 1246, + 604, + 1216, + 604 + ], + "score": 0.86, + "latex": "z _ { 4 }" + }, + { + "category_id": 13, + "poly": [ + 1051, + 578, + 1080, + 578, + 1080, + 604, + 1051, + 604 + ], + "score": 0.71, + "latex": "z _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1128, + 578, + 1158, + 578, + 1158, + 604, + 1128, + 604 + ], + "score": 0.61, + "latex": "z _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 1091, + 579, + 1119, + 579, + 1119, + 605, + 1091, + 605 + ], + "score": 0.59, + "latex": "z _ { 2 }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 450.0, + 1408.0, + 450.0, + 1408.0, + 489.0, + 294.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 480.0, + 1405.0, + 480.0, + 1405.0, + 517.0, + 294.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 512.0, + 1405.0, + 512.0, + 1405.0, + 546.0, + 294.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 543.0, + 1405.0, + 543.0, + 1405.0, + 576.0, + 293.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 570.0, + 1050.0, + 570.0, + 1050.0, + 609.0, + 294.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 570.0, + 1090.0, + 570.0, + 1090.0, + 609.0, + 1081.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 570.0, + 1127.0, + 570.0, + 1127.0, + 609.0, + 1120.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 570.0, + 1215.0, + 570.0, + 1215.0, + 609.0, + 1159.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 570.0, + 1406.0, + 570.0, + 1406.0, + 609.0, + 1247.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 603.0, + 1408.0, + 603.0, + 1408.0, + 639.0, + 293.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 633.0, + 757.0, + 633.0, + 757.0, + 668.0, + 293.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 265.0, + 692.0, + 265.0, + 692.0, + 305.0, + 588.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 268.0, + 1149.0, + 268.0, + 1149.0, + 303.0, + 1034.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 255.0, + 1344.0, + 255.0, + 1344.0, + 329.0, + 1246.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 321.0, + 560.0, + 321.0, + 560.0, + 340.0, + 540.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 308.0, + 694.0, + 308.0, + 694.0, + 350.0, + 584.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 321.0, + 1018.0, + 321.0, + 1018.0, + 340.0, + 998.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 306.0, + 1149.0, + 306.0, + 1149.0, + 350.0, + 1035.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1299.0, + 276.5, + 1308.0, + 276.5, + 1308.0, + 303.5, + 1299.0, + 303.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 922.0, + 586.0, + 922.0, + 586.0, + 966.0, + 292.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 991.0, + 1406.0, + 991.0, + 1406.0, + 1034.0, + 292.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1027.0, + 1405.0, + 1027.0, + 1405.0, + 1062.0, + 293.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1056.0, + 1406.0, + 1056.0, + 1406.0, + 1092.0, + 295.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1086.0, + 1406.0, + 1086.0, + 1406.0, + 1122.0, + 295.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1116.0, + 1408.0, + 1116.0, + 1408.0, + 1154.0, + 292.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1147.0, + 1405.0, + 1147.0, + 1405.0, + 1183.0, + 293.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1177.0, + 1405.0, + 1177.0, + 1405.0, + 1214.0, + 292.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1208.0, + 1408.0, + 1208.0, + 1408.0, + 1244.0, + 295.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1239.0, + 1408.0, + 1239.0, + 1408.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1267.0, + 1408.0, + 1267.0, + 1408.0, + 1304.0, + 292.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1299.0, + 1405.0, + 1299.0, + 1405.0, + 1333.0, + 295.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1365.0, + 293.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1362.0, + 1403.0, + 1362.0, + 1403.0, + 1394.0, + 296.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1389.0, + 735.0, + 1389.0, + 735.0, + 1429.0, + 293.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1669.0, + 1405.0, + 1669.0, + 1405.0, + 1703.0, + 296.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1698.0, + 1405.0, + 1698.0, + 1405.0, + 1735.0, + 293.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1728.0, + 1406.0, + 1728.0, + 1406.0, + 1764.0, + 292.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1758.0, + 1408.0, + 1758.0, + 1408.0, + 1796.0, + 292.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1787.0, + 1408.0, + 1787.0, + 1408.0, + 1827.0, + 292.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1819.0, + 1408.0, + 1819.0, + 1408.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1887.0, + 292.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1406.0, + 1912.0, + 1406.0, + 1949.0, + 293.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1980.0, + 293.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2010.0, + 295.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1406.0, + 2003.0, + 1406.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1435.0, + 1409.0, + 1435.0, + 1409.0, + 1475.0, + 292.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1469.0, + 1406.0, + 1469.0, + 1406.0, + 1503.0, + 294.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1406.0, + 1500.0, + 1406.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1530.0, + 1406.0, + 1530.0, + 1406.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1558.0, + 1406.0, + 1558.0, + 1406.0, + 1595.0, + 293.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1589.0, + 1406.0, + 1589.0, + 1406.0, + 1626.0, + 293.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1620.0, + 1409.0, + 1620.0, + 1409.0, + 1655.0, + 293.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 727.0, + 1403.0, + 727.0, + 1403.0, + 760.0, + 296.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 756.0, + 1407.0, + 756.0, + 1407.0, + 791.0, + 293.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 789.0, + 1405.0, + 789.0, + 1405.0, + 822.0, + 295.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 817.0, + 1407.0, + 817.0, + 1407.0, + 854.0, + 295.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 850.0, + 826.0, + 850.0, + 826.0, + 883.0, + 296.0, + 883.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 876, + 1406, + 876, + 1406, + 1152, + 297, + 1152 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1273, + 1405, + 1273, + 1405, + 1519, + 297, + 1519 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1533, + 1404, + 1533, + 1404, + 1718, + 298, + 1718 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1405, + 1819, + 1405, + 2038, + 298, + 2038 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 297, + 226, + 1401, + 226, + 1401, + 550, + 297, + 550 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 296, + 580, + 1406, + 580, + 1406, + 735, + 296, + 735 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 299, + 799, + 1400, + 799, + 1400, + 861, + 299, + 861 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 300, + 1201, + 480, + 1201, + 480, + 1236, + 300, + 1236 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 815, + 76, + 815, + 104, + 298, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 299, + 1761, + 556, + 1761, + 556, + 1791, + 299, + 1791 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.563 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.556 + }, + { + "category_id": 13, + "poly": [ + 720, + 2005, + 950, + 2005, + 950, + 2039, + 720, + 2039 + ], + "score": 0.92, + "latex": "\\Pi ^ { i } = \\{ \\pi _ { 1 } ^ { i } , \\ldots , \\pi _ { m ^ { i } } ^ { i } \\}" + }, + { + "category_id": 13, + "poly": [ + 1098, + 1820, + 1303, + 1820, + 1303, + 1854, + 1098, + 1854 + ], + "score": 0.91, + "latex": "\\{ S , { \\mathcal { A } } , { \\mathcal { T } } , R , \\rho , \\gamma \\}" + }, + { + "category_id": 13, + "poly": [ + 867, + 1975, + 1150, + 1975, + 1150, + 2001, + 867, + 2001 + ], + "score": 0.91, + "latex": "\\mathcal { A } = \\mathcal { A } ^ { 1 } \\times \\mathcal { A } ^ { 2 } \\times \\cdot \\cdot \\cdot \\times \\mathcal { A } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 607, + 2005, + 641, + 2005, + 641, + 2030, + 607, + 2030 + ], + "score": 0.9, + "latex": "m ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 1036, + 1945, + 1065, + 1945, + 1065, + 1969, + 1036, + 1969 + ], + "score": 0.89, + "latex": "S ^ { \\imath }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1972, + 328, + 1972, + 328, + 2000, + 298, + 2000 + ], + "score": 0.87, + "latex": "S ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 1116, + 1942, + 1149, + 1942, + 1149, + 1970, + 1116, + 1970 + ], + "score": 0.87, + "latex": "A ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 796, + 1979, + 816, + 1979, + 816, + 1999, + 796, + 1999 + ], + "score": 0.87, + "latex": "\\boldsymbol { S }" + }, + { + "category_id": 13, + "poly": [ + 712, + 1887, + 739, + 1887, + 739, + 1909, + 712, + 1909 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 783, + 1948, + 794, + 1948, + 794, + 1969, + 783, + 1969 + ], + "score": 0.82, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 365, + 2005, + 377, + 2005, + 377, + 2031, + 365, + 2031 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1197, + 2006, + 1210, + 2006, + 1210, + 2030, + 1197, + 2030 + ], + "score": 0.73, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 390, + 536, + 432, + 536, + 432, + 559, + 390, + 559 + ], + "score": 0.25, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 249.0, + 531.0, + 249.0, + 531.0, + 291.0, + 458.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 287.0, + 528.0, + 287.0, + 528.0, + 348.0, + 461.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 296.0, + 739.0, + 296.0, + 739.0, + 349.0, + 573.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 276.0, + 962.0, + 276.0, + 962.0, + 349.0, + 755.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 323.0, + 1175.0, + 323.0, + 1175.0, + 349.0, + 1010.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 280.0, + 1402.0, + 280.0, + 1402.0, + 352.0, + 1188.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 369.0, + 661.0, + 369.0, + 661.0, + 380.0, + 651.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 366.0, + 912.0, + 366.0, + 912.0, + 396.0, + 859.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 365.0, + 1099.0, + 365.0, + 1099.0, + 381.0, + 1084.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 366.0, + 1347.0, + 366.0, + 1347.0, + 398.0, + 1295.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 383.0, + 720.0, + 383.0, + 720.0, + 414.0, + 592.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 386.0, + 1156.0, + 386.0, + 1156.0, + 412.0, + 1028.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 381.0, + 1278.0, + 381.0, + 1278.0, + 390.0, + 1263.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 408.0, + 738.0, + 408.0, + 738.0, + 440.0, + 573.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 406.0, + 852.0, + 406.0, + 852.0, + 440.0, + 815.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 406.0, + 917.0, + 406.0, + 917.0, + 440.0, + 853.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 407.0, + 1173.0, + 407.0, + 1173.0, + 439.0, + 1010.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 405.0, + 1352.0, + 405.0, + 1352.0, + 441.0, + 1252.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 430.0, + 515.0, + 430.0, + 515.0, + 469.0, + 473.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 449.0, + 914.0, + 449.0, + 914.0, + 480.0, + 858.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 446.0, + 1287.0, + 446.0, + 1287.0, + 482.0, + 1249.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 449.0, + 1349.0, + 449.0, + 1349.0, + 478.0, + 1293.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 475.0, + 529.0, + 475.0, + 529.0, + 502.0, + 461.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 470.0, + 788.0, + 470.0, + 788.0, + 502.0, + 607.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 486.0, + 1025.0, + 486.0, + 1025.0, + 504.0, + 1002.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 471.0, + 1224.0, + 471.0, + 1224.0, + 502.0, + 1041.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 529.0, + 389.0, + 529.0, + 389.0, + 559.0, + 295.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 529.0, + 437.0, + 529.0, + 437.0, + 559.0, + 433.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.75, + 363.0, + 404.75, + 363.0, + 404.75, + 390.5, + 328.75, + 390.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 582.0, + 1404.0, + 582.0, + 1404.0, + 615.0, + 295.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 611.0, + 1406.0, + 611.0, + 1406.0, + 646.0, + 293.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 643.0, + 1407.0, + 643.0, + 1407.0, + 676.0, + 295.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 669.0, + 1406.0, + 669.0, + 1406.0, + 708.0, + 293.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 705.0, + 619.0, + 705.0, + 619.0, + 739.0, + 296.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1195.0, + 485.0, + 1195.0, + 485.0, + 1245.0, + 291.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1760.0, + 559.0, + 1760.0, + 559.0, + 1795.0, + 295.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 878.0, + 1404.0, + 878.0, + 1404.0, + 911.0, + 296.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 907.0, + 1406.0, + 907.0, + 1406.0, + 944.0, + 294.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 934.0, + 1411.0, + 934.0, + 1411.0, + 977.0, + 291.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 964.0, + 1407.0, + 964.0, + 1407.0, + 1007.0, + 291.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1001.0, + 1408.0, + 1001.0, + 1408.0, + 1034.0, + 295.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1030.0, + 1403.0, + 1030.0, + 1403.0, + 1062.0, + 295.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1061.0, + 1404.0, + 1061.0, + 1404.0, + 1094.0, + 296.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1090.0, + 1406.0, + 1090.0, + 1406.0, + 1125.0, + 294.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1120.0, + 1364.0, + 1120.0, + 1364.0, + 1156.0, + 294.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1272.0, + 1408.0, + 1272.0, + 1408.0, + 1308.0, + 292.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1305.0, + 1406.0, + 1305.0, + 1406.0, + 1338.0, + 292.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1336.0, + 1407.0, + 1336.0, + 1407.0, + 1370.0, + 294.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1366.0, + 1405.0, + 1366.0, + 1405.0, + 1400.0, + 295.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1396.0, + 1403.0, + 1396.0, + 1403.0, + 1430.0, + 295.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1427.0, + 1406.0, + 1427.0, + 1406.0, + 1461.0, + 295.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1456.0, + 1406.0, + 1456.0, + 1406.0, + 1495.0, + 292.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 960.0, + 1490.0, + 960.0, + 1520.0, + 293.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1531.0, + 1404.0, + 1531.0, + 1404.0, + 1570.0, + 293.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1563.0, + 1405.0, + 1563.0, + 1405.0, + 1601.0, + 295.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1595.0, + 1406.0, + 1595.0, + 1406.0, + 1631.0, + 294.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1626.0, + 1407.0, + 1626.0, + 1407.0, + 1660.0, + 293.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1657.0, + 1404.0, + 1657.0, + 1404.0, + 1689.0, + 296.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1685.0, + 1096.0, + 1685.0, + 1096.0, + 1720.0, + 293.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1820.0, + 1097.0, + 1820.0, + 1097.0, + 1855.0, + 296.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 1820.0, + 1410.0, + 1820.0, + 1410.0, + 1855.0, + 1304.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1850.0, + 1408.0, + 1850.0, + 1408.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 711.0, + 1880.0, + 711.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1880.0, + 1406.0, + 1880.0, + 1406.0, + 1917.0, + 740.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1408.0, + 1912.0, + 1408.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 782.0, + 1941.0, + 782.0, + 1976.0, + 293.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1941.0, + 1035.0, + 1941.0, + 1035.0, + 1976.0, + 795.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1941.0, + 1115.0, + 1941.0, + 1115.0, + 1976.0, + 1066.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1976.0, + 1150.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1967.0, + 297.0, + 1967.0, + 297.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1967.0, + 795.0, + 1967.0, + 795.0, + 2009.0, + 329.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1967.0, + 866.0, + 1967.0, + 866.0, + 2009.0, + 817.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 1967.0, + 1407.0, + 1967.0, + 1407.0, + 2009.0, + 1151.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1997.0, + 364.0, + 1997.0, + 364.0, + 2043.0, + 291.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1997.0, + 606.0, + 1997.0, + 606.0, + 2043.0, + 378.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1997.0, + 719.0, + 1997.0, + 719.0, + 2043.0, + 642.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1997.0, + 1196.0, + 1997.0, + 1196.0, + 2043.0, + 951.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1997.0, + 1408.0, + 1997.0, + 1408.0, + 2043.0, + 1211.0, + 2043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 1404.0, + 798.0, + 1404.0, + 834.0, + 295.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 830.0, + 1205.0, + 830.0, + 1205.0, + 865.0, + 295.0, + 865.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1291, + 1404, + 1291, + 1404, + 1537, + 298, + 1537 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1057, + 1405, + 1057, + 1405, + 1277, + 297, + 1277 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 762, + 1406, + 762, + 1406, + 945, + 297, + 945 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1650, + 1403, + 1650, + 1403, + 1805, + 298, + 1805 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 313, + 219, + 1399, + 219, + 1399, + 456, + 313, + 456 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 296, + 478, + 1406, + 478, + 1406, + 695, + 296, + 695 + ], + "score": 0.962 + }, + { + "category_id": 0, + "poly": [ + 299, + 996, + 657, + 996, + 657, + 1027, + 299, + 1027 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 301, + 1590, + 1231, + 1590, + 1231, + 1619, + 301, + 1619 + ], + "score": 0.844 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.772 + }, + { + "category_id": 13, + "poly": [ + 763, + 765, + 849, + 765, + 849, + 800, + 763, + 800 + ], + "score": 0.93, + "latex": "s _ { t } ^ { i } \\in S ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 1009, + 765, + 1101, + 765, + 1101, + 800, + 1009, + 800 + ], + "score": 0.93, + "latex": "a _ { t } ^ { i } \\in \\mathcal A ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 297, + 907, + 428, + 907, + 428, + 949, + 297, + 949 + ], + "score": 0.93, + "latex": "\\textstyle \\sum _ { t = 0 } ^ { T - 1 } \\gamma ^ { t } r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1079, + 1973, + 1161, + 1973, + 1161, + 2006, + 1079, + 2006 + ], + "score": 0.93, + "latex": "\\mathcal { H } ( a | s )" + }, + { + "category_id": 13, + "poly": [ + 298, + 763, + 476, + 763, + 476, + 809, + 298, + 809 + ], + "score": 0.92, + "latex": "\\pi _ { c _ { t } ^ { i } } ^ { i } \\left( a _ { t } ^ { i } | s _ { t } ^ { i } \\right) \\in \\Pi ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 635, + 1087, + 773, + 1087, + 773, + 1121, + 635, + 1121 + ], + "score": 0.92, + "latex": "\\Pi ^ { 1 } , \\dots , \\Pi ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 1088, + 1211, + 1088, + 1211, + 1124, + 1109, + 1124 + ], + "score": 0.92, + "latex": "\\pi _ { c ^ { i } } ^ { i } \\in \\Pi ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 794, + 1853, + 895, + 1853, + 895, + 1884, + 794, + 1884 + ], + "score": 0.92, + "latex": "z \\sim p ( z )" + }, + { + "category_id": 13, + "poly": [ + 1215, + 1851, + 1267, + 1851, + 1267, + 1884, + 1215, + 1884 + ], + "score": 0.92, + "latex": "p ( z )" + }, + { + "category_id": 13, + "poly": [ + 356, + 1975, + 486, + 1975, + 486, + 2005, + 356, + 2005 + ], + "score": 0.9, + "latex": "M I ( a , z | s )" + }, + { + "category_id": 13, + "poly": [ + 1058, + 805, + 1224, + 805, + 1224, + 840, + 1058, + 840 + ], + "score": 0.9, + "latex": "a _ { t } ^ { 1 } , a _ { t } ^ { 2 } , \\ldots , a _ { t } ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 499, + 838, + 1402, + 838, + 1402, + 881, + 499, + 881 + ], + "score": 0.9, + "latex": "\\pi ( a _ { t } ^ { 1 } , \\bar { a } _ { t } ^ { 2 } , \\dots , a _ { t } ^ { N } | s _ { t } , c _ { t } ^ { 1 } , c _ { t } ^ { 2 } , \\dots , \\bar { c } _ { t } ^ { N } ) = ( \\pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( a _ { t } ^ { 1 } | s _ { t } ) , \\pi _ { c _ { t } ^ { 2 } } ^ { 2 } ( a _ { t } ^ { 2 } | s _ { t } ) , \\dots , \\pi _ { c _ { t } ^ { N } } ^ { N } ( a _ { t } ^ { N } | s _ { t } ) )" + }, + { + "category_id": 13, + "poly": [ + 441, + 1247, + 496, + 1247, + 496, + 1276, + 441, + 1276 + ], + "score": 0.9, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 1034, + 1821, + 1108, + 1821, + 1108, + 1849, + 1034, + 1849 + ], + "score": 0.9, + "latex": "a \\in { \\mathcal { A } }" + }, + { + "category_id": 13, + "poly": [ + 355, + 1852, + 424, + 1852, + 424, + 1879, + 355, + 1879 + ], + "score": 0.89, + "latex": "s \\in S" + }, + { + "category_id": 13, + "poly": [ + 366, + 1942, + 473, + 1942, + 473, + 1974, + 366, + 1974 + ], + "score": 0.89, + "latex": "M I ( s , z )" + }, + { + "category_id": 13, + "poly": [ + 680, + 1124, + 717, + 1124, + 717, + 1155, + 680, + 1155 + ], + "score": 0.89, + "latex": "N _ { z }" + }, + { + "category_id": 13, + "poly": [ + 561, + 765, + 586, + 765, + 586, + 800, + 561, + 800 + ], + "score": 0.89, + "latex": "c _ { t } ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 1227, + 1064, + 1297, + 1064, + 1297, + 1089, + 1227, + 1089 + ], + "score": 0.88, + "latex": "\\pi _ { m e t a }" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1123, + 1203, + 1123, + 1203, + 1152, + 1176, + 1152 + ], + "score": 0.85, + "latex": "z ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 464, + 814, + 492, + 814, + 492, + 839, + 464, + 839 + ], + "score": 0.84, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 618, + 883, + 643, + 883, + 643, + 908, + 618, + 908 + ], + "score": 0.83, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 526, + 1090, + 554, + 1090, + 554, + 1117, + 526, + 1117 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 513, + 913, + 537, + 913, + 537, + 940, + 513, + 940 + ], + "score": 0.82, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1345, + 880, + 1405, + 880, + 1405, + 909, + 1345, + 909 + ], + "score": 0.81, + "latex": "R =" + }, + { + "category_id": 13, + "poly": [ + 1104, + 572, + 1133, + 572, + 1133, + 599, + 1104, + 599 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 878, + 542, + 907, + 542, + 907, + 568, + 878, + 568 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 813, + 810, + 842, + 810, + 842, + 836, + 813, + 836 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 474, + 1884, + 487, + 1884, + 487, + 1909, + 474, + 1909 + ], + "score": 0.78, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 817, + 668, + 835, + 668, + 835, + 690, + 817, + 690 + ], + "score": 0.77, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1386, + 772, + 1398, + 772, + 1398, + 793, + 1386, + 793 + ], + "score": 0.76, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 495, + 1917, + 514, + 1917, + 514, + 1939, + 495, + 1939 + ], + "score": 0.76, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 826, + 1825, + 837, + 1825, + 837, + 1847, + 826, + 1847 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1372, + 572, + 1401, + 572, + 1401, + 599, + 1372, + 599 + ], + "score": 0.74, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 679, + 815, + 696, + 815, + 696, + 839, + 679, + 839 + ], + "score": 0.71, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 1199, + 1749, + 1217, + 1749, + 1217, + 1770, + 1199, + 1770 + ], + "score": 0.7, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1057, + 806, + 1088, + 806, + 1088, + 840, + 1057, + 840 + ], + "score": 0.28, + "latex": "a _ { t } ^ { 1 }" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 234.0, + 337.0, + 234.0, + 337.0, + 253.0, + 316.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 234.0, + 520.0, + 234.0, + 520.0, + 253.0, + 499.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 234.0, + 703.0, + 234.0, + 703.0, + 253.0, + 682.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 234.0, + 887.0, + 234.0, + 887.0, + 255.0, + 865.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 234.0, + 1070.0, + 234.0, + 1070.0, + 255.0, + 1047.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 234.0, + 1254.0, + 234.0, + 1254.0, + 255.0, + 1230.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 273.0, + 715.0, + 273.0, + 715.0, + 297.0, + 672.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 268.0, + 901.0, + 268.0, + 901.0, + 301.0, + 854.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 265.0, + 984.0, + 265.0, + 984.0, + 295.0, + 965.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 270.0, + 1083.0, + 270.0, + 1083.0, + 298.0, + 1036.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 262.0, + 1124.0, + 262.0, + 1124.0, + 299.0, + 1103.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 262.0, + 1169.0, + 262.0, + 1169.0, + 297.0, + 1146.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 273.0, + 1208.0, + 273.0, + 1208.0, + 294.0, + 1193.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 272.0, + 1266.0, + 272.0, + 1266.0, + 298.0, + 1219.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 264.0, + 1307.0, + 264.0, + 1307.0, + 298.0, + 1285.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 269.0, + 1349.0, + 269.0, + 1349.0, + 295.0, + 1330.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 309.0, + 770.0, + 309.0, + 770.0, + 322.0, + 758.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 307.0, + 1319.0, + 307.0, + 1319.0, + 324.0, + 1304.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 340.0, + 419.0, + 340.0, + 419.0, + 357.0, + 403.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 326.0, + 661.0, + 326.0, + 661.0, + 368.0, + 527.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 324.0, + 837.0, + 324.0, + 837.0, + 367.0, + 706.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 339.0, + 967.0, + 339.0, + 967.0, + 356.0, + 951.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 330.0, + 1207.0, + 330.0, + 1207.0, + 364.0, + 1081.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 325.0, + 1385.0, + 325.0, + 1385.0, + 367.0, + 1255.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 382.0, + 339.0, + 382.0, + 339.0, + 405.0, + 315.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 385.0, + 522.0, + 385.0, + 522.0, + 405.0, + 496.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 385.0, + 704.0, + 385.0, + 704.0, + 405.0, + 679.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 385.0, + 888.0, + 385.0, + 888.0, + 405.0, + 863.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 385.0, + 1071.0, + 385.0, + 1071.0, + 405.0, + 1046.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 385.0, + 1254.0, + 385.0, + 1254.0, + 406.0, + 1229.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 422.0, + 431.0, + 422.0, + 431.0, + 454.0, + 350.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 422.0, + 635.0, + 422.0, + 635.0, + 456.0, + 516.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 424.0, + 824.0, + 424.0, + 824.0, + 455.0, + 694.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 422.0, + 1010.0, + 422.0, + 1010.0, + 453.0, + 872.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 424.0, + 1397.0, + 424.0, + 1397.0, + 453.0, + 1038.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 1408.0, + 478.0, + 1408.0, + 516.0, + 294.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 511.0, + 1408.0, + 511.0, + 1408.0, + 546.0, + 294.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 541.0, + 877.0, + 541.0, + 877.0, + 575.0, + 295.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 541.0, + 1406.0, + 541.0, + 1406.0, + 575.0, + 908.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 572.0, + 1103.0, + 572.0, + 1103.0, + 605.0, + 293.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 572.0, + 1371.0, + 572.0, + 1371.0, + 605.0, + 1134.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 600.0, + 1407.0, + 600.0, + 1407.0, + 636.0, + 293.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 629.0, + 1406.0, + 629.0, + 1406.0, + 671.0, + 291.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 665.0, + 816.0, + 665.0, + 816.0, + 695.0, + 296.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 665.0, + 995.0, + 665.0, + 995.0, + 695.0, + 836.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 996.0, + 660.0, + 996.0, + 660.0, + 1029.0, + 295.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1587.0, + 1238.0, + 1587.0, + 1238.0, + 1622.0, + 295.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1291.0, + 1404.0, + 1291.0, + 1404.0, + 1326.0, + 293.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1324.0, + 1406.0, + 1324.0, + 1406.0, + 1358.0, + 294.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1354.0, + 1405.0, + 1354.0, + 1405.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1384.0, + 1404.0, + 1384.0, + 1404.0, + 1418.0, + 294.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1415.0, + 1404.0, + 1415.0, + 1404.0, + 1449.0, + 294.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1444.0, + 1405.0, + 1444.0, + 1405.0, + 1480.0, + 293.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1475.0, + 1408.0, + 1475.0, + 1408.0, + 1513.0, + 293.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1503.0, + 715.0, + 1503.0, + 715.0, + 1539.0, + 293.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1054.0, + 1226.0, + 1054.0, + 1226.0, + 1095.0, + 291.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 1054.0, + 1407.0, + 1054.0, + 1407.0, + 1095.0, + 1298.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1085.0, + 525.0, + 1085.0, + 525.0, + 1127.0, + 292.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1085.0, + 634.0, + 1085.0, + 634.0, + 1127.0, + 555.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1085.0, + 1108.0, + 1085.0, + 1108.0, + 1127.0, + 774.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1085.0, + 1407.0, + 1085.0, + 1407.0, + 1127.0, + 1212.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1121.0, + 679.0, + 1121.0, + 679.0, + 1159.0, + 294.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1121.0, + 1175.0, + 1121.0, + 1175.0, + 1159.0, + 718.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1121.0, + 1408.0, + 1121.0, + 1408.0, + 1159.0, + 1204.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1155.0, + 1407.0, + 1155.0, + 1407.0, + 1189.0, + 295.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1185.0, + 1405.0, + 1185.0, + 1405.0, + 1219.0, + 295.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1405.0, + 1216.0, + 1405.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1243.0, + 440.0, + 1243.0, + 440.0, + 1282.0, + 294.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1243.0, + 626.0, + 1243.0, + 626.0, + 1282.0, + 497.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 761.0, + 297.0, + 761.0, + 297.0, + 861.0, + 282.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 761.0, + 560.0, + 761.0, + 560.0, + 861.0, + 493.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 761.0, + 678.0, + 761.0, + 678.0, + 861.0, + 587.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 761.0, + 762.0, + 761.0, + 762.0, + 861.0, + 697.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 761.0, + 1008.0, + 761.0, + 1008.0, + 861.0, + 850.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 761.0, + 1385.0, + 761.0, + 1385.0, + 861.0, + 1225.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 761.0, + 1413.0, + 761.0, + 1413.0, + 861.0, + 1399.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 823.0, + 498.0, + 823.0, + 498.0, + 886.0, + 286.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 823.0, + 1413.0, + 823.0, + 1413.0, + 886.0, + 1403.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 876.0, + 296.0, + 876.0, + 296.0, + 967.0, + 282.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 876.0, + 512.0, + 876.0, + 512.0, + 967.0, + 429.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 876.0, + 617.0, + 876.0, + 617.0, + 967.0, + 538.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 876.0, + 1344.0, + 876.0, + 1344.0, + 967.0, + 644.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 825.0, + 1821.0, + 825.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1821.0, + 1033.0, + 1821.0, + 1033.0, + 1855.0, + 838.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 1109.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 354.0, + 1851.0, + 354.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 1851.0, + 793.0, + 1851.0, + 793.0, + 1886.0, + 425.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1851.0, + 1214.0, + 1851.0, + 1214.0, + 1886.0, + 896.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1886.0, + 1268.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 473.0, + 1882.0, + 473.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1917.0, + 488.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1912.0, + 494.0, + 1912.0, + 494.0, + 1945.0, + 292.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1945.0, + 515.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 365.0, + 1942.0, + 365.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1977.0, + 474.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 355.0, + 1973.0, + 355.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 1973.0, + 1078.0, + 1973.0, + 1078.0, + 2008.0, + 487.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1162.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2008.0, + 1162.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 1405.0, + 1650.0, + 1405.0, + 1688.0, + 294.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1404.0, + 1682.0, + 1404.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1403.0, + 1712.0, + 1403.0, + 1745.0, + 294.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1742.0, + 1198.0, + 1742.0, + 1198.0, + 1778.0, + 293.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1742.0, + 1404.0, + 1742.0, + 1404.0, + 1778.0, + 1218.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1351.0, + 1773.0, + 1351.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1415, + 1404, + 1415, + 1404, + 1649, + 297, + 1649 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1080, + 1405, + 1080, + 1405, + 1326, + 298, + 1326 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 1662, + 1403, + 1662, + 1403, + 1817, + 299, + 1817 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1403, + 1911, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 494, + 1039, + 1198, + 1039, + 1198, + 1076, + 494, + 1076 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 293, + 977, + 1399, + 977, + 1399, + 1037, + 293, + 1037 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 298, + 1364, + 973, + 1364, + 973, + 1394, + 298, + 1394 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 298, + 748, + 468, + 748, + 468, + 779, + 298, + 779 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 300, + 1854, + 582, + 1854, + 582, + 1885, + 300, + 1885 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 305, + 275, + 1354, + 275, + 1354, + 687, + 305, + 687 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 298, + 930, + 1128, + 930, + 1128, + 965, + 298, + 965 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 299, + 235, + 566, + 235, + 566, + 266, + 299, + 266 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1366, + 881, + 1400, + 881, + 1400, + 910, + 1366, + 910 + ], + "score": 0.869 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1041, + 1400, + 1041, + 1400, + 1070, + 1366, + 1070 + ], + "score": 0.866 + }, + { + "category_id": 9, + "poly": [ + 1366, + 839, + 1400, + 839, + 1400, + 869, + 1366, + 869 + ], + "score": 0.859 + }, + { + "category_id": 9, + "poly": [ + 1366, + 800, + 1400, + 800, + 1400, + 829, + 1366, + 829 + ], + "score": 0.859 + }, + { + "category_id": 8, + "poly": [ + 433, + 795, + 1266, + 795, + 1266, + 832, + 433, + 832 + ], + "score": 0.822 + }, + { + "category_id": 8, + "poly": [ + 475, + 837, + 1195, + 837, + 1195, + 874, + 475, + 874 + ], + "score": 0.804 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.718 + }, + { + "category_id": 8, + "poly": [ + 480, + 879, + 1116, + 879, + 1116, + 917, + 480, + 917 + ], + "score": 0.494 + }, + { + "category_id": 8, + "poly": [ + 450, + 837, + 1194, + 837, + 1194, + 874, + 450, + 874 + ], + "score": 0.277 + }, + { + "category_id": 8, + "poly": [ + 434, + 795, + 1261, + 795, + 1261, + 917, + 434, + 917 + ], + "score": 0.109 + }, + { + "category_id": 13, + "poly": [ + 1044, + 1579, + 1253, + 1579, + 1253, + 1617, + 1044, + 1617 + ], + "score": 0.94, + "latex": "a ^ { i } \\sim \\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )" + }, + { + "category_id": 13, + "poly": [ + 587, + 1545, + 736, + 1545, + 736, + 1582, + 587, + 1582 + ], + "score": 0.93, + "latex": "\\pi _ { c ^ { i } } ^ { i } ( a ^ { i } | s ^ { i } , z ^ { i } )" + }, + { + "category_id": 13, + "poly": [ + 1227, + 1546, + 1400, + 1546, + 1400, + 1582, + 1227, + 1582 + ], + "score": 0.93, + "latex": "\\{ \\pi _ { c ^ { 1 } } ^ { 1 } , \\ldots , \\pi _ { c ^ { N } } ^ { N } \\}" + }, + { + "category_id": 13, + "poly": [ + 655, + 931, + 740, + 931, + 740, + 966, + 655, + 966 + ], + "score": 0.93, + "latex": "q _ { \\phi } ( z | s )" + }, + { + "category_id": 13, + "poly": [ + 556, + 587, + 1064, + 587, + 1064, + 624, + 556, + 624 + ], + "score": 0.93, + "latex": "s _ { t } , ( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } ) , s _ { t + \\tau } , r _ { t : t + \\tau - 1 }" + }, + { + "category_id": 13, + "poly": [ + 1080, + 1418, + 1212, + 1418, + 1212, + 1453, + 1080, + 1453 + ], + "score": 0.93, + "latex": "c ^ { i } \\in [ 1 , m ^ { i } ]" + }, + { + "category_id": 13, + "poly": [ + 506, + 1451, + 624, + 1451, + 624, + 1484, + 506, + 1484 + ], + "score": 0.93, + "latex": "i \\in [ 1 , N ]" + }, + { + "category_id": 14, + "poly": [ + 431, + 791, + 1266, + 791, + 1266, + 921, + 431, + 921 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { \\mathcal { F } ( \\theta ) = M I ( s , z ) - M I ( a , z \\vert s ) + \\mathcal { H } ( a \\vert s ) = \\mathcal { H } ( a \\vert s , z ) - \\mathcal { H } ( z \\vert s ) + \\mathcal { H } ( z ) } \\\\ & { \\quad \\quad = \\mathcal { H } ( a \\vert s , z ) + \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log p ( z \\vert s ) ] - \\mathbb { E } _ { z \\sim p ( z ) } [ \\log p ( z ) ] } \\\\ & { \\quad \\quad \\geq \\mathcal { H } ( a \\vert s , z ) + \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log q _ { \\phi } ( z \\vert s ) - \\log p ( z ) ] , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 899, + 364, + 1348, + 364, + 1348, + 402, + 899, + 402 + ], + "score": 0.92, + "latex": "( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) , ( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } ) \\sim \\pi _ { m e t a } ( s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1047, + 930, + 1119, + 930, + 1119, + 965, + 1047, + 965 + ], + "score": 0.92, + "latex": "p ( z | s )" + }, + { + "category_id": 13, + "poly": [ + 984, + 1479, + 1090, + 1479, + 1090, + 1515, + 984, + 1515 + ], + "score": 0.92, + "latex": "\\pi _ { c ^ { i } } ^ { i } \\in \\Pi ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 928, + 274, + 1053, + 274, + 1053, + 308, + 928, + 308 + ], + "score": 0.92, + "latex": "\\Pi ^ { 1 } , . . . , \\Pi ^ { N }" + }, + { + "category_id": 13, + "poly": [ + 1191, + 1973, + 1244, + 1973, + 1244, + 2006, + 1191, + 2006 + ], + "score": 0.92, + "latex": "p ( z )" + }, + { + "category_id": 13, + "poly": [ + 299, + 1692, + 375, + 1692, + 375, + 1727, + 299, + 1727 + ], + "score": 0.91, + "latex": "[ 1 , m ^ { i } ]" + }, + { + "category_id": 13, + "poly": [ + 956, + 1515, + 1140, + 1515, + 1140, + 1550, + 956, + 1550 + ], + "score": 0.9, + "latex": "( z ^ { 1 } , z ^ { 2 } , \\dots , z ^ { N } )" + }, + { + "category_id": 13, + "poly": [ + 1076, + 1692, + 1128, + 1692, + 1128, + 1721, + 1076, + 1721 + ], + "score": 0.9, + "latex": "\\mathbb { R } ^ { N _ { z } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1481, + 353, + 1481, + 353, + 1513, + 298, + 1513 + ], + "score": 0.9, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 619, + 1417, + 995, + 1417, + 995, + 1453, + 619, + 1453 + ], + "score": 0.89, + "latex": "\\pi _ { m e t a } ( c ^ { 1 } , \\ldots , c ^ { N } , z ^ { 1 } , \\ldots , z ^ { N } | s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1111, + 1756, + 1148, + 1756, + 1148, + 1785, + 1111, + 1785 + ], + "score": 0.89, + "latex": "N _ { z }" + }, + { + "category_id": 13, + "poly": [ + 442, + 430, + 549, + 430, + 549, + 458, + 442, + 458 + ], + "score": 0.89, + "latex": "\\tau < T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 573, + 309, + 645, + 309, + 645, + 335, + 573, + 335 + ], + "score": 0.89, + "latex": "t \\gets 0" + }, + { + "category_id": 14, + "poly": [ + 501, + 1037, + 1197, + 1037, + 1197, + 1077, + 501, + 1077 + ], + "score": 0.89, + "latex": "r _ { t } + \\lambda _ { 1 } \\mathcal { H } ( a | s , z ) + \\lambda _ { 2 } \\mathbb { E } _ { z \\sim p ( z ) , s \\sim \\pi ( z ) } [ \\log q _ { \\phi } ( z | s ) - \\log p ( z ) ] ," + }, + { + "category_id": 13, + "poly": [ + 1311, + 1725, + 1346, + 1725, + 1346, + 1753, + 1311, + 1753 + ], + "score": 0.88, + "latex": "m ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1082, + 404, + 1082, + 404, + 1111, + 373, + 1111 + ], + "score": 0.88, + "latex": "\\lambda _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 738, + 2009, + 775, + 2009, + 775, + 2034, + 738, + 2034 + ], + "score": 0.88, + "latex": "N _ { z }" + }, + { + "category_id": 13, + "poly": [ + 745, + 1082, + 776, + 1082, + 776, + 1111, + 745, + 1111 + ], + "score": 0.88, + "latex": "\\lambda _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1264, + 278, + 1318, + 278, + 1318, + 308, + 1264, + 308 + ], + "score": 0.87, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 441, + 1615, + 465, + 1615, + 465, + 1644, + 441, + 1644 + ], + "score": 0.86, + "latex": "s ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 609, + 1725, + 701, + 1725, + 701, + 1755, + 609, + 1755 + ], + "score": 0.86, + "latex": "( 2 \\times N )" + }, + { + "category_id": 13, + "poly": [ + 702, + 1615, + 727, + 1615, + 727, + 1644, + 702, + 1644 + ], + "score": 0.86, + "latex": "z ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 568, + 283, + 638, + 283, + 638, + 307, + 568, + 307 + ], + "score": 0.86, + "latex": "\\pi _ { m e t a }" + }, + { + "category_id": 13, + "poly": [ + 557, + 1145, + 583, + 1145, + 583, + 1172, + 557, + 1172 + ], + "score": 0.85, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 680, + 1451, + 794, + 1451, + 794, + 1479, + 680, + 1479 + ], + "score": 0.85, + "latex": "z ^ { i } \\in \\mathbb { R } ^ { N _ { z } }" + }, + { + "category_id": 13, + "poly": [ + 910, + 313, + 938, + 313, + 938, + 339, + 910, + 339 + ], + "score": 0.84, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 900, + 1756, + 929, + 1756, + 929, + 1782, + 900, + 1782 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1102, + 1726, + 1130, + 1726, + 1130, + 1752, + 1102, + 1752 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 983, + 1920, + 1012, + 1920, + 1012, + 1945, + 983, + 1945 + ], + "score": 0.79, + "latex": "q _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 824, + 1269, + 841, + 1269, + 841, + 1291, + 824, + 1291 + ], + "score": 0.77, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 396, + 457, + 1120, + 457, + 1120, + 501, + 396, + 501 + ], + "score": 0.77, + "latex": "\\boldsymbol a _ { t + \\tau } = ( a _ { t + \\tau } ^ { 1 } , \\dots , \\dot { a _ { t + \\tau } ^ { N } } ) \\sim ( \\pi _ { c _ { t } ^ { 1 } } ^ { 1 } ( s _ { t + \\tau } , z _ { t } ^ { 1 } ) , \\dots , \\pi _ { c _ { t } ^ { N } } ^ { N } ( s _ { t + \\tau } , z _ { t } ^ { N } ) )" + }, + { + "category_id": 13, + "poly": [ + 607, + 1178, + 625, + 1178, + 625, + 1199, + 607, + 1199 + ], + "score": 0.76, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1291, + 592, + 1314, + 592, + 1314, + 618, + 1291, + 618 + ], + "score": 0.76, + "latex": "\\boldsymbol { B }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1483, + 1278, + 1483, + 1278, + 1508, + 1266, + 1508 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 371, + 399, + 449, + 399, + 449, + 428, + 371, + 428 + ], + "score": 0.66, + "latex": "\\tau 0" + }, + { + "category_id": 13, + "poly": [ + 851, + 1550, + 864, + 1550, + 864, + 1574, + 851, + 1574 + ], + "score": 0.54, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 371, + 622, + 488, + 622, + 488, + 650, + 371, + 650 + ], + "score": 0.51, + "latex": "t \\gets t + \\tau" + }, + { + "category_id": 14, + "poly": [ + 397, + 494, + 788, + 494, + 788, + 562, + 397, + 562 + ], + "score": 0.38, + "latex": "\\begin{array} { r l } & { s _ { t + \\tau + 1 } , r _ { t + \\tau } \\gets \\mathrm { E N V } ( s _ { t + \\tau } , a _ { t + \\tau } ) } \\\\ & { \\tau \\gets \\tau + 1 } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1363.0, + 977.0, + 1363.0, + 977.0, + 1397.0, + 294.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1854.0, + 584.0, + 1854.0, + 584.0, + 1887.0, + 296.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 232.0, + 572.0, + 232.0, + 572.0, + 274.0, + 295.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1412.0, + 618.0, + 1412.0, + 618.0, + 1463.0, + 287.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1412.0, + 1079.0, + 1412.0, + 1079.0, + 1463.0, + 996.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 1412.0, + 1405.0, + 1412.0, + 1405.0, + 1463.0, + 1213.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1442.0, + 505.0, + 1442.0, + 505.0, + 1489.0, + 291.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 1442.0, + 679.0, + 1442.0, + 679.0, + 1489.0, + 625.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1442.0, + 1406.0, + 1442.0, + 1406.0, + 1489.0, + 795.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1477.0, + 297.0, + 1477.0, + 297.0, + 1518.0, + 294.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1477.0, + 983.0, + 1477.0, + 983.0, + 1518.0, + 354.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1477.0, + 1265.0, + 1477.0, + 1265.0, + 1518.0, + 1091.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1477.0, + 1407.0, + 1477.0, + 1407.0, + 1518.0, + 1279.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1506.0, + 955.0, + 1506.0, + 955.0, + 1559.0, + 290.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1506.0, + 1411.0, + 1506.0, + 1411.0, + 1559.0, + 1141.0, + 1559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1536.0, + 586.0, + 1536.0, + 586.0, + 1597.0, + 283.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1536.0, + 850.0, + 1536.0, + 850.0, + 1597.0, + 737.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1536.0, + 1226.0, + 1536.0, + 1226.0, + 1597.0, + 865.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1536.0, + 1413.0, + 1536.0, + 1413.0, + 1597.0, + 1401.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1578.0, + 1043.0, + 1578.0, + 1043.0, + 1618.0, + 294.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1578.0, + 1407.0, + 1578.0, + 1407.0, + 1618.0, + 1254.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1614.0, + 440.0, + 1614.0, + 440.0, + 1652.0, + 294.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1614.0, + 701.0, + 1614.0, + 701.0, + 1652.0, + 466.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1614.0, + 1299.0, + 1614.0, + 1299.0, + 1652.0, + 728.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.25, + 1555.5, + 1391.25, + 1555.5, + 1391.25, + 1570.0, + 1359.25, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1079.0, + 372.0, + 1079.0, + 372.0, + 1116.0, + 295.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1079.0, + 744.0, + 1079.0, + 744.0, + 1116.0, + 405.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1079.0, + 1408.0, + 1079.0, + 1408.0, + 1116.0, + 777.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1110.0, + 1406.0, + 1110.0, + 1406.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1142.0, + 556.0, + 1142.0, + 556.0, + 1175.0, + 295.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1142.0, + 1406.0, + 1142.0, + 1406.0, + 1175.0, + 584.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1172.0, + 606.0, + 1172.0, + 606.0, + 1205.0, + 295.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1172.0, + 1406.0, + 1172.0, + 1406.0, + 1205.0, + 626.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1202.0, + 1408.0, + 1202.0, + 1408.0, + 1239.0, + 293.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1232.0, + 1406.0, + 1232.0, + 1406.0, + 1269.0, + 293.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1262.0, + 823.0, + 1262.0, + 823.0, + 1299.0, + 293.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1262.0, + 1405.0, + 1262.0, + 1405.0, + 1299.0, + 842.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1293.0, + 1182.0, + 1293.0, + 1182.0, + 1328.0, + 293.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1660.0, + 1407.0, + 1660.0, + 1407.0, + 1698.0, + 294.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1688.0, + 298.0, + 1688.0, + 298.0, + 1729.0, + 293.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1688.0, + 1075.0, + 1688.0, + 1075.0, + 1729.0, + 376.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1688.0, + 1407.0, + 1688.0, + 1407.0, + 1729.0, + 1129.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1720.0, + 608.0, + 1720.0, + 608.0, + 1763.0, + 292.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1720.0, + 1101.0, + 1720.0, + 1101.0, + 1763.0, + 702.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1720.0, + 1310.0, + 1720.0, + 1310.0, + 1763.0, + 1131.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 1720.0, + 1407.0, + 1720.0, + 1407.0, + 1763.0, + 1347.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1753.0, + 899.0, + 1753.0, + 899.0, + 1790.0, + 295.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1753.0, + 1110.0, + 1753.0, + 1110.0, + 1790.0, + 930.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 1753.0, + 1405.0, + 1753.0, + 1405.0, + 1790.0, + 1149.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1785.0, + 884.0, + 1785.0, + 884.0, + 1818.0, + 297.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 982.0, + 1909.0, + 982.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1947.0, + 1013.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1408.0, + 1940.0, + 1408.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1190.0, + 1972.0, + 1190.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2008.0, + 1245.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 737.0, + 2002.0, + 737.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2038.0, + 776.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 973.0, + 1404.0, + 973.0, + 1404.0, + 1015.0, + 291.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1007.0, + 390.0, + 1007.0, + 390.0, + 1042.0, + 290.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 744.0, + 472.0, + 744.0, + 472.0, + 783.0, + 294.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 267.0, + 567.0, + 267.0, + 567.0, + 316.0, + 304.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 267.0, + 927.0, + 267.0, + 927.0, + 316.0, + 639.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 267.0, + 1263.0, + 267.0, + 1263.0, + 316.0, + 1054.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 267.0, + 1324.0, + 267.0, + 1324.0, + 316.0, + 1319.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 300.0, + 572.0, + 300.0, + 572.0, + 344.0, + 310.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 300.0, + 909.0, + 300.0, + 909.0, + 344.0, + 646.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 300.0, + 945.0, + 300.0, + 945.0, + 344.0, + 939.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 338.0, + 735.0, + 338.0, + 735.0, + 369.0, + 311.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 368.0, + 338.0, + 368.0, + 338.0, + 399.0, + 308.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 362.0, + 898.0, + 362.0, + 898.0, + 407.0, + 364.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 400.0, + 339.0, + 400.0, + 339.0, + 429.0, + 308.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 399.0, + 454.0, + 399.0, + 454.0, + 431.0, + 450.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 429.0, + 339.0, + 429.0, + 339.0, + 460.0, + 308.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 427.0, + 441.0, + 427.0, + 441.0, + 461.0, + 371.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 427.0, + 922.0, + 427.0, + 922.0, + 461.0, + 550.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 459.0, + 340.0, + 459.0, + 340.0, + 495.0, + 306.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 450.0, + 1126.0, + 450.0, + 1126.0, + 504.0, + 1121.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 498.0, + 338.0, + 498.0, + 338.0, + 530.0, + 308.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 528.0, + 339.0, + 528.0, + 339.0, + 563.0, + 305.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 560.0, + 339.0, + 560.0, + 339.0, + 592.0, + 297.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 559.0, + 494.0, + 559.0, + 494.0, + 591.0, + 371.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 589.0, + 339.0, + 589.0, + 339.0, + 623.0, + 297.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 586.0, + 555.0, + 586.0, + 555.0, + 627.0, + 366.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 586.0, + 1290.0, + 586.0, + 1290.0, + 627.0, + 1065.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 586.0, + 1320.0, + 586.0, + 1320.0, + 627.0, + 1315.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 615.0, + 342.0, + 615.0, + 342.0, + 655.0, + 295.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 617.0, + 370.0, + 617.0, + 370.0, + 654.0, + 367.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 617.0, + 492.0, + 617.0, + 492.0, + 654.0, + 489.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 651.0, + 466.0, + 651.0, + 466.0, + 683.0, + 298.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 925.0, + 654.0, + 925.0, + 654.0, + 971.0, + 292.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 925.0, + 1046.0, + 925.0, + 1046.0, + 971.0, + 741.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 925.0, + 1130.0, + 925.0, + 1130.0, + 971.0, + 1120.0, + 971.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 352, + 221, + 1352, + 221, + 1352, + 577, + 352, + 577 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1298, + 1403, + 1298, + 1403, + 1420, + 299, + 1420 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 831, + 1404, + 831, + 1404, + 1018, + 299, + 1018 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 1159, + 1404, + 1159, + 1404, + 1282, + 298, + 1282 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1802, + 1404, + 1802, + 1404, + 1896, + 299, + 1896 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 300, + 1911, + 1402, + 1911, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 295, + 604, + 1406, + 604, + 1406, + 758, + 295, + 758 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 299, + 1436, + 1400, + 1436, + 1400, + 1529, + 299, + 1529 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 301, + 1647, + 1401, + 1647, + 1401, + 1711, + 301, + 1711 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 298, + 1725, + 1398, + 1725, + 1398, + 1787, + 298, + 1787 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 300, + 1080, + 557, + 1080, + 557, + 1114, + 300, + 1114 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 298, + 1583, + 501, + 1583, + 501, + 1614, + 298, + 1614 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.766 + }, + { + "category_id": 13, + "poly": [ + 441, + 2003, + 496, + 2003, + 496, + 2034, + 441, + 2034 + ], + "score": 0.9, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 1041, + 1946, + 1095, + 1946, + 1095, + 1973, + 1041, + 1973 + ], + "score": 0.89, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 452, + 989, + 482, + 989, + 482, + 1021, + 452, + 1021 + ], + "score": 0.87, + "latex": "q _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 1083, + 894, + 1109, + 894, + 1109, + 923, + 1083, + 923 + ], + "score": 0.86, + "latex": "z ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 766, + 864, + 844, + 864, + 844, + 897, + 766, + 897 + ], + "score": 0.86, + "latex": "[ - 1 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1075, + 1805, + 1104, + 1805, + 1104, + 1831, + 1075, + 1831 + ], + "score": 0.83, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 768, + 895, + 791, + 895, + 791, + 922, + 768, + 922 + ], + "score": 0.82, + "latex": "\\bar { c } ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 1295, + 1913, + 1323, + 1913, + 1323, + 1939, + 1295, + 1939 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 476.0, + 515.0, + 476.0, + 515.0, + 492.0, + 491.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 473.0, + 630.0, + 473.0, + 630.0, + 481.0, + 621.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 546.0, + 660.0, + 546.0, + 660.0, + 578.0, + 353.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 545.0, + 977.0, + 545.0, + 977.0, + 579.0, + 723.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 545.0, + 1272.0, + 545.0, + 1272.0, + 579.0, + 1112.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 602.0, + 1403.0, + 602.0, + 1403.0, + 639.0, + 294.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 633.0, + 1404.0, + 633.0, + 1404.0, + 670.0, + 294.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 662.0, + 1406.0, + 662.0, + 1406.0, + 702.0, + 292.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 694.0, + 1406.0, + 694.0, + 1406.0, + 731.0, + 294.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 725.0, + 1214.0, + 725.0, + 1214.0, + 762.0, + 295.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1077.0, + 561.0, + 1077.0, + 561.0, + 1120.0, + 293.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1582.0, + 505.0, + 1582.0, + 505.0, + 1618.0, + 294.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 1407.0, + 1298.0, + 1407.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1328.0, + 1407.0, + 1328.0, + 1407.0, + 1362.0, + 293.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1361.0, + 1404.0, + 1361.0, + 1404.0, + 1394.0, + 294.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1389.0, + 1106.0, + 1389.0, + 1106.0, + 1421.0, + 294.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 832.0, + 1404.0, + 832.0, + 1404.0, + 868.0, + 295.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 863.0, + 765.0, + 863.0, + 765.0, + 899.0, + 294.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 863.0, + 1404.0, + 863.0, + 1404.0, + 899.0, + 845.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 894.0, + 767.0, + 894.0, + 767.0, + 930.0, + 295.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 894.0, + 1082.0, + 894.0, + 1082.0, + 930.0, + 792.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 894.0, + 1406.0, + 894.0, + 1406.0, + 930.0, + 1110.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 926.0, + 1404.0, + 926.0, + 1404.0, + 958.0, + 294.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 955.0, + 1405.0, + 955.0, + 1405.0, + 991.0, + 294.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 985.0, + 451.0, + 985.0, + 451.0, + 1019.0, + 293.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 985.0, + 1306.0, + 985.0, + 1306.0, + 1019.0, + 483.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1158.0, + 1404.0, + 1158.0, + 1404.0, + 1193.0, + 293.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1188.0, + 1409.0, + 1188.0, + 1409.0, + 1226.0, + 292.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1218.0, + 1405.0, + 1218.0, + 1405.0, + 1255.0, + 292.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1254.0, + 1279.0, + 1254.0, + 1279.0, + 1283.0, + 296.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1800.0, + 1074.0, + 1800.0, + 1074.0, + 1839.0, + 293.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1800.0, + 1404.0, + 1800.0, + 1404.0, + 1839.0, + 1105.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1833.0, + 1403.0, + 1833.0, + 1403.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1864.0, + 1036.0, + 1864.0, + 1036.0, + 1901.0, + 294.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1294.0, + 1910.0, + 1294.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1946.0, + 1324.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1040.0, + 1941.0, + 1040.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 1096.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1406.0, + 1974.0, + 1406.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1999.0, + 440.0, + 1999.0, + 440.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1999.0, + 626.0, + 1999.0, + 626.0, + 2040.0, + 497.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1433.0, + 1405.0, + 1433.0, + 1405.0, + 1472.0, + 293.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1468.0, + 1405.0, + 1468.0, + 1405.0, + 1501.0, + 294.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1496.0, + 1342.0, + 1496.0, + 1342.0, + 1532.0, + 294.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1644.0, + 1406.0, + 1644.0, + 1406.0, + 1684.0, + 296.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1678.0, + 552.0, + 1678.0, + 552.0, + 1714.0, + 296.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1722.0, + 1404.0, + 1722.0, + 1404.0, + 1764.0, + 293.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1756.0, + 647.0, + 1756.0, + 647.0, + 1791.0, + 295.0, + 1791.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1589, + 1404, + 1589, + 1404, + 1805, + 298, + 1805 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1390, + 1404, + 1390, + 1404, + 1575, + 298, + 1575 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1404, + 1819, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1133, + 1404, + 1133, + 1404, + 1288, + 298, + 1288 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 825, + 1403, + 825, + 1403, + 981, + 298, + 981 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 995, + 1404, + 995, + 1404, + 1119, + 298, + 1119 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 322, + 229, + 1379, + 229, + 1379, + 552, + 322, + 552 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 296, + 576, + 1406, + 576, + 1406, + 761, + 296, + 761 + ], + "score": 0.968 + }, + { + "category_id": 0, + "poly": [ + 299, + 1330, + 660, + 1330, + 660, + 1362, + 299, + 1362 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.748 + }, + { + "category_id": 13, + "poly": [ + 728, + 920, + 783, + 920, + 783, + 949, + 728, + 949 + ], + "score": 0.91, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 467, + 1088, + 521, + 1088, + 521, + 1119, + 467, + 1119 + ], + "score": 0.91, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 1052, + 1028, + 1081, + 1028, + 1081, + 1054, + 1052, + 1054 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 430, + 1197, + 459, + 1197, + 459, + 1223, + 430, + 1223 + ], + "score": 0.8, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 429, + 1231, + 447, + 1231, + 447, + 1253, + 429, + 1253 + ], + "score": 0.73, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1034, + 639, + 1103, + 639, + 1103, + 668, + 1034, + 668 + ], + "score": 0.68, + "latex": "2 . 0 \\mathbf { M }" + }, + { + "category_id": 13, + "poly": [ + 911, + 639, + 979, + 639, + 979, + 668, + 911, + 668 + ], + "score": 0.25, + "latex": "1 . 2 \\mathbf { M }" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 229.0, + 370.0, + 229.0, + 370.0, + 256.0, + 335.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 246.0, + 432.0, + 246.0, + 432.0, + 269.0, + 401.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 236.0, + 726.0, + 236.0, + 726.0, + 257.0, + 693.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 242.0, + 1008.0, + 242.0, + 1008.0, + 269.0, + 905.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 231.0, + 1084.0, + 231.0, + 1084.0, + 252.0, + 1053.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 274.0, + 371.0, + 274.0, + 371.0, + 297.0, + 336.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 262.0, + 460.0, + 262.0, + 460.0, + 302.0, + 403.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 279.0, + 728.0, + 279.0, + 728.0, + 302.0, + 695.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 268.0, + 1009.0, + 268.0, + 1009.0, + 305.0, + 912.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 276.0, + 1082.0, + 276.0, + 1082.0, + 294.0, + 1055.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 292.0, + 385.0, + 292.0, + 385.0, + 410.0, + 309.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 297.0, + 478.0, + 297.0, + 478.0, + 336.0, + 405.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 304.0, + 728.0, + 304.0, + 728.0, + 393.0, + 679.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 306.0, + 1083.0, + 306.0, + 1083.0, + 399.0, + 1038.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 330.0, + 494.0, + 330.0, + 494.0, + 357.0, + 404.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 347.0, + 532.0, + 347.0, + 532.0, + 371.0, + 505.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 359.0, + 370.0, + 359.0, + 370.0, + 386.0, + 335.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 362.0, + 729.0, + 362.0, + 729.0, + 388.0, + 694.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 363.0, + 1083.0, + 363.0, + 1083.0, + 384.0, + 1054.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 405.0, + 368.0, + 405.0, + 368.0, + 428.0, + 336.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 404.0, + 728.0, + 404.0, + 728.0, + 430.0, + 694.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 407.0, + 1083.0, + 407.0, + 1083.0, + 429.0, + 1054.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 447.0, + 384.0, + 447.0, + 384.0, + 484.0, + 335.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 461.0, + 454.0, + 461.0, + 454.0, + 484.0, + 423.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 458.0, + 525.0, + 458.0, + 525.0, + 485.0, + 491.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 461.0, + 593.0, + 461.0, + 593.0, + 484.0, + 561.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 461.0, + 664.0, + 461.0, + 664.0, + 484.0, + 631.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 448.0, + 743.0, + 448.0, + 743.0, + 484.0, + 695.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 461.0, + 812.0, + 461.0, + 812.0, + 484.0, + 781.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 460.0, + 882.0, + 460.0, + 882.0, + 484.0, + 852.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 460.0, + 952.0, + 460.0, + 952.0, + 484.0, + 922.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 460.0, + 1024.0, + 460.0, + 1024.0, + 487.0, + 989.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 452.0, + 1098.0, + 452.0, + 1098.0, + 487.0, + 1054.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 464.0, + 1147.0, + 464.0, + 1147.0, + 485.0, + 1114.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 464.0, + 1193.0, + 464.0, + 1193.0, + 485.0, + 1163.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 464.0, + 1240.0, + 464.0, + 1240.0, + 485.0, + 1211.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 464.0, + 1288.0, + 464.0, + 1288.0, + 485.0, + 1258.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 464.0, + 1334.0, + 464.0, + 1334.0, + 487.0, + 1305.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 466.0, + 1381.0, + 466.0, + 1381.0, + 484.0, + 1353.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 477.0, + 583.0, + 477.0, + 583.0, + 500.0, + 431.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 477.0, + 943.0, + 477.0, + 943.0, + 500.0, + 790.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 478.0, + 1300.0, + 478.0, + 1300.0, + 501.0, + 1148.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 519.0, + 644.0, + 519.0, + 644.0, + 553.0, + 336.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 519.0, + 977.0, + 519.0, + 977.0, + 553.0, + 720.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 520.0, + 1288.0, + 520.0, + 1288.0, + 553.0, + 1129.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.75, + 291.5, + 658.75, + 291.5, + 658.75, + 344.5, + 513.75, + 344.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.25, + 389.0, + 653.25, + 389.0, + 653.25, + 440.0, + 482.25, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 578.0, + 1403.0, + 578.0, + 1403.0, + 610.0, + 295.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 606.0, + 1404.0, + 606.0, + 1404.0, + 643.0, + 294.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 638.0, + 910.0, + 638.0, + 910.0, + 672.0, + 293.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 638.0, + 1033.0, + 638.0, + 1033.0, + 672.0, + 980.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 638.0, + 1406.0, + 638.0, + 1406.0, + 672.0, + 1104.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 669.0, + 1409.0, + 669.0, + 1409.0, + 704.0, + 293.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 696.0, + 1407.0, + 696.0, + 1407.0, + 735.0, + 293.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 812.0, + 728.0, + 812.0, + 764.0, + 294.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 663.0, + 1330.0, + 663.0, + 1366.0, + 294.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 1405.0, + 1590.0, + 1405.0, + 1624.0, + 294.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1619.0, + 1406.0, + 1619.0, + 1406.0, + 1657.0, + 293.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1650.0, + 1405.0, + 1650.0, + 1405.0, + 1687.0, + 293.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1404.0, + 1682.0, + 1404.0, + 1717.0, + 294.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1711.0, + 1408.0, + 1711.0, + 1408.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1741.0, + 1405.0, + 1741.0, + 1405.0, + 1778.0, + 292.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1774.0, + 1026.0, + 1774.0, + 1026.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1424.0, + 295.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1419.0, + 1406.0, + 1419.0, + 1406.0, + 1456.0, + 292.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1485.0, + 294.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1483.0, + 1405.0, + 1483.0, + 1405.0, + 1517.0, + 293.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1513.0, + 1404.0, + 1513.0, + 1404.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1544.0, + 1243.0, + 1544.0, + 1243.0, + 1579.0, + 296.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 297.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1883.0, + 294.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1404.0, + 1882.0, + 1404.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1409.0, + 1910.0, + 1409.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1408.0, + 1973.0, + 1408.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2004.0, + 1404.0, + 2004.0, + 1404.0, + 2038.0, + 296.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1136.0, + 1402.0, + 1136.0, + 1402.0, + 1165.0, + 297.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1163.0, + 1406.0, + 1163.0, + 1406.0, + 1201.0, + 293.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1193.0, + 429.0, + 1193.0, + 429.0, + 1232.0, + 292.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 1193.0, + 1407.0, + 1193.0, + 1407.0, + 1232.0, + 460.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1224.0, + 428.0, + 1224.0, + 428.0, + 1260.0, + 293.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1224.0, + 1404.0, + 1224.0, + 1404.0, + 1260.0, + 448.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1254.0, + 1391.0, + 1254.0, + 1391.0, + 1293.0, + 293.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 827.0, + 1404.0, + 827.0, + 1404.0, + 860.0, + 296.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 856.0, + 1405.0, + 856.0, + 1405.0, + 894.0, + 293.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 887.0, + 1407.0, + 887.0, + 1407.0, + 924.0, + 294.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 918.0, + 727.0, + 918.0, + 727.0, + 955.0, + 294.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 918.0, + 1405.0, + 918.0, + 1405.0, + 955.0, + 784.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 947.0, + 1226.0, + 947.0, + 1226.0, + 987.0, + 291.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 996.0, + 1405.0, + 996.0, + 1405.0, + 1029.0, + 294.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1024.0, + 1051.0, + 1024.0, + 1051.0, + 1062.0, + 293.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1024.0, + 1408.0, + 1024.0, + 1408.0, + 1062.0, + 1082.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1058.0, + 1406.0, + 1058.0, + 1406.0, + 1091.0, + 293.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1090.0, + 466.0, + 1090.0, + 466.0, + 1122.0, + 296.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1090.0, + 1287.0, + 1090.0, + 1287.0, + 1122.0, + 522.0, + 1122.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1168, + 1405, + 1168, + 1405, + 1414, + 298, + 1414 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 443, + 224, + 1256, + 224, + 1256, + 431, + 443, + 431 + ], + "score": 0.981, + "html": "
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
RL0.000±0.0000.000±0.0000.000±0.000
MARL0.000±0.0000.000± 0.0000.000± 0.000
RL-SBD0.000± 0.0000.000± 0.0000.000±0.000
MARL-SBD0.000 ± 0.0000.000 ±0.0000.000 ± 0.000
Modular0.324± 0.4680.917± 0.2760.003± 0.058
Modular-SBD (Ours)0.902 ± 0.2980.950 ± 0.2180.323 ± 0.468
" + }, + { + "category_id": 1, + "poly": [ + 298, + 830, + 1404, + 830, + 1404, + 1014, + 298, + 1014 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1650, + 1404, + 1650, + 1404, + 1804, + 298, + 1804 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1404, + 1819, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1513, + 1404, + 1513, + 1404, + 1635, + 299, + 1635 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 608, + 1403, + 608, + 1403, + 733, + 298, + 733 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1030, + 1404, + 1030, + 1404, + 1154, + 298, + 1154 + ], + "score": 0.975 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 816, + 75, + 816, + 105, + 298, + 105 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 299, + 1453, + 493, + 1453, + 493, + 1484, + 299, + 1484 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 300, + 454, + 1401, + 454, + 1401, + 547, + 300, + 547 + ], + "score": 0.809 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.793 + }, + { + "category_id": 0, + "poly": [ + 299, + 772, + 601, + 772, + 601, + 804, + 299, + 804 + ], + "score": 0.78 + }, + { + "category_id": 0, + "poly": [ + 300, + 772, + 600, + 772, + 600, + 804, + 300, + 804 + ], + "score": 0.419 + }, + { + "category_id": 6, + "poly": [ + 300, + 454, + 1401, + 454, + 1401, + 547, + 300, + 547 + ], + "score": 0.124 + }, + { + "category_id": 13, + "poly": [ + 730, + 1821, + 804, + 1821, + 804, + 1850, + 730, + 1850 + ], + "score": 0.87, + "latex": "3 2 . 3 \\%" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1452.0, + 497.0, + 1452.0, + 497.0, + 1488.0, + 295.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 769.0, + 605.0, + 769.0, + 605.0, + 809.0, + 293.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 769.0, + 605.0, + 769.0, + 605.0, + 809.0, + 294.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 451.0, + 1405.0, + 451.0, + 1405.0, + 490.0, + 294.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 485.0, + 1405.0, + 485.0, + 1405.0, + 519.0, + 294.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 516.0, + 919.0, + 516.0, + 919.0, + 550.0, + 295.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1170.0, + 1408.0, + 1170.0, + 1408.0, + 1203.0, + 297.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1198.0, + 1406.0, + 1198.0, + 1406.0, + 1236.0, + 293.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1230.0, + 1405.0, + 1230.0, + 1405.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1261.0, + 1406.0, + 1261.0, + 1406.0, + 1295.0, + 295.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1291.0, + 1405.0, + 1291.0, + 1405.0, + 1328.0, + 293.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1322.0, + 1405.0, + 1322.0, + 1405.0, + 1356.0, + 295.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1353.0, + 1408.0, + 1353.0, + 1408.0, + 1387.0, + 293.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1382.0, + 1357.0, + 1382.0, + 1357.0, + 1416.0, + 295.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 831.0, + 1406.0, + 831.0, + 1406.0, + 866.0, + 293.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 864.0, + 1404.0, + 864.0, + 1404.0, + 895.0, + 296.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 893.0, + 1405.0, + 893.0, + 1405.0, + 928.0, + 293.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 922.0, + 1408.0, + 922.0, + 1408.0, + 959.0, + 293.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 951.0, + 1406.0, + 951.0, + 1406.0, + 988.0, + 293.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 985.0, + 642.0, + 985.0, + 642.0, + 1017.0, + 297.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1652.0, + 1406.0, + 1652.0, + 1406.0, + 1685.0, + 296.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1682.0, + 1405.0, + 1682.0, + 1405.0, + 1715.0, + 296.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1711.0, + 1405.0, + 1711.0, + 1405.0, + 1747.0, + 292.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1740.0, + 1406.0, + 1740.0, + 1406.0, + 1780.0, + 292.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1768.0, + 818.0, + 1768.0, + 818.0, + 1813.0, + 293.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1822.0, + 729.0, + 1822.0, + 729.0, + 1853.0, + 300.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 1822.0, + 1404.0, + 1822.0, + 1404.0, + 1853.0, + 805.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1407.0, + 1850.0, + 1407.0, + 1883.0, + 293.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1914.0, + 293.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 1409.0, + 1909.0, + 1409.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1282.0, + 2002.0, + 1282.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1514.0, + 1404.0, + 1514.0, + 1404.0, + 1546.0, + 295.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1542.0, + 1405.0, + 1542.0, + 1405.0, + 1579.0, + 293.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1573.0, + 1408.0, + 1573.0, + 1408.0, + 1609.0, + 293.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1606.0, + 1377.0, + 1606.0, + 1377.0, + 1639.0, + 295.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 608.0, + 1405.0, + 608.0, + 1405.0, + 644.0, + 292.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 640.0, + 1405.0, + 640.0, + 1405.0, + 673.0, + 292.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 672.0, + 1403.0, + 672.0, + 1403.0, + 704.0, + 294.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 702.0, + 1071.0, + 702.0, + 1071.0, + 736.0, + 293.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1028.0, + 1406.0, + 1028.0, + 1406.0, + 1066.0, + 292.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1059.0, + 1406.0, + 1059.0, + 1406.0, + 1097.0, + 293.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1092.0, + 1404.0, + 1092.0, + 1404.0, + 1125.0, + 294.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1122.0, + 1394.0, + 1122.0, + 1394.0, + 1155.0, + 294.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 451.0, + 1405.0, + 451.0, + 1405.0, + 490.0, + 294.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 485.0, + 1405.0, + 485.0, + 1405.0, + 519.0, + 294.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 516.0, + 919.0, + 516.0, + 919.0, + 550.0, + 295.0, + 550.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 726, + 1404, + 726, + 1404, + 1032, + 298, + 1032 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1269, + 1404, + 1269, + 1404, + 1485, + 298, + 1485 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1132, + 1405, + 1132, + 1405, + 1254, + 299, + 1254 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 298, + 543, + 837, + 543, + 837, + 605, + 298, + 605 + ], + "score": 0.954 + }, + { + "category_id": 4, + "poly": [ + 861, + 544, + 1402, + 544, + 1402, + 605, + 861, + 605 + ], + "score": 0.953 + }, + { + "category_id": 3, + "poly": [ + 307, + 261, + 1392, + 261, + 1392, + 517, + 307, + 517 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 298, + 1603, + 1405, + 1603, + 1405, + 1880, + 298, + 1880 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1401, + 1972, + 1401, + 2033, + 299, + 2033 + ], + "score": 0.945 + }, + { + "category_id": 0, + "poly": [ + 300, + 1533, + 543, + 1533, + 543, + 1568, + 300, + 1568 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 300, + 1920, + 557, + 1920, + 557, + 1947, + 300, + 1947 + ], + "score": 0.872 + }, + { + "category_id": 0, + "poly": [ + 299, + 1073, + 892, + 1073, + 892, + 1105, + 299, + 1105 + ], + "score": 0.84 + }, + { + "category_id": 0, + "poly": [ + 304, + 670, + 888, + 670, + 888, + 698, + 304, + 698 + ], + "score": 0.789 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.781 + }, + { + "category_id": 0, + "poly": [ + 303, + 670, + 886, + 670, + 886, + 698, + 303, + 698 + ], + "score": 0.136 + }, + { + "category_id": 13, + "poly": [ + 750, + 1331, + 896, + 1331, + 896, + 1365, + 750, + 1365 + ], + "score": 0.93, + "latex": "( z _ { t } ^ { 1 } , \\ldots , z _ { t } ^ { N } )" + }, + { + "category_id": 13, + "poly": [ + 772, + 909, + 951, + 909, + 951, + 943, + 772, + 943 + ], + "score": 0.93, + "latex": "\\lambda _ { 2 } = \\bar { \\{ 0 . 5 , 1 . 0 \\} }" + }, + { + "category_id": 13, + "poly": [ + 852, + 757, + 1187, + 757, + 1187, + 790, + 852, + 790 + ], + "score": 0.92, + "latex": "\\lambda _ { 2 } = \\{ 0 . 0 , 0 . 0 5 , 0 . 1 , 0 . 5 , 1 . 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 985, + 1330, + 1347, + 1330, + 1347, + 1366, + 985, + 1366 + ], + "score": 0.91, + "latex": "\\mathbf { \\bar { \\Phi } } ( c _ { t } ^ { 1 } , \\ldots , c _ { t } ^ { N } ) \\neq ( c _ { t - 1 } ^ { 1 } , \\ldots , c _ { t - 1 } ^ { N } )" + }, + { + "category_id": 13, + "poly": [ + 445, + 1162, + 707, + 1162, + 707, + 1196, + 445, + 1196 + ], + "score": 0.91, + "latex": "T _ { l o w } = \\{ 1 , 2 , 3 , 5 , 1 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 1041, + 789, + 1235, + 789, + 1235, + 821, + 1041, + 821 + ], + "score": 0.91, + "latex": "\\lambda _ { 2 } = \\{ 0 . 0 5 , 0 . 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 515, + 1333, + 570, + 1333, + 570, + 1362, + 515, + 1362 + ], + "score": 0.91, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 558, + 1271, + 613, + 1271, + 613, + 1302, + 558, + 1302 + ], + "score": 0.91, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 1055, + 1134, + 1110, + 1134, + 1110, + 1163, + 1055, + 1163 + ], + "score": 0.9, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 933, + 574, + 988, + 574, + 988, + 604, + 933, + 604 + ], + "score": 0.9, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1423, + 352, + 1423, + 352, + 1454, + 298, + 1454 + ], + "score": 0.9, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 576, + 1196, + 630, + 1196, + 630, + 1224, + 576, + 1224 + ], + "score": 0.89, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 1228, + 820, + 1314, + 820, + 1314, + 849, + 1228, + 849 + ], + "score": 0.87, + "latex": "{ \\lambda } _ { 2 } = 0" + }, + { + "category_id": 13, + "poly": [ + 633, + 575, + 665, + 575, + 665, + 604, + 633, + 604 + ], + "score": 0.87, + "latex": "\\lambda _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 839, + 1075, + 893, + 1075, + 893, + 1105, + 839, + 1105 + ], + "score": 0.85, + "latex": "T _ { l o w }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 542.0, + 840.0, + 542.0, + 840.0, + 576.0, + 298.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 572.0, + 632.0, + 572.0, + 632.0, + 608.0, + 298.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 572.0, + 840.0, + 572.0, + 840.0, + 608.0, + 666.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 540.0, + 1406.0, + 540.0, + 1406.0, + 574.0, + 858.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 572.0, + 932.0, + 572.0, + 932.0, + 608.0, + 859.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 572.0, + 1380.0, + 572.0, + 1380.0, + 608.0, + 989.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 263.0, + 346.0, + 263.0, + 346.0, + 283.0, + 320.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 273.0, + 414.0, + 273.0, + 414.0, + 294.0, + 373.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 262.0, + 624.0, + 262.0, + 624.0, + 283.0, + 587.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 267.0, + 909.0, + 267.0, + 909.0, + 285.0, + 882.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 267.0, + 1174.0, + 267.0, + 1174.0, + 285.0, + 1148.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 286.0, + 413.0, + 286.0, + 413.0, + 307.0, + 374.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 295.0, + 346.0, + 295.0, + 346.0, + 404.0, + 306.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 301.0, + 416.0, + 301.0, + 416.0, + 333.0, + 373.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 307.0, + 593.0, + 307.0, + 593.0, + 401.0, + 573.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 304.0, + 624.0, + 304.0, + 624.0, + 322.0, + 594.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 298.0, + 908.0, + 298.0, + 908.0, + 393.0, + 867.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 298.0, + 1174.0, + 298.0, + 1174.0, + 393.0, + 1131.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 326.0, + 347.0, + 326.0, + 347.0, + 348.0, + 317.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 327.0, + 414.0, + 327.0, + 414.0, + 345.0, + 376.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 330.0, + 908.0, + 330.0, + 908.0, + 349.0, + 882.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 328.0, + 1177.0, + 328.0, + 1177.0, + 351.0, + 1147.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 343.0, + 625.0, + 343.0, + 625.0, + 366.0, + 593.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 347.0, + 1378.0, + 347.0, + 1378.0, + 367.0, + 1270.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 359.0, + 345.0, + 359.0, + 345.0, + 379.0, + 320.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 358.0, + 909.0, + 358.0, + 909.0, + 384.0, + 880.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 358.0, + 1175.0, + 358.0, + 1175.0, + 383.0, + 1147.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 362.0, + 1320.0, + 362.0, + 1320.0, + 380.0, + 1272.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 373.0, + 1320.0, + 373.0, + 1320.0, + 394.0, + 1270.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 394.0, + 346.0, + 394.0, + 346.0, + 411.0, + 321.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 385.0, + 622.0, + 385.0, + 622.0, + 404.0, + 594.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 394.0, + 908.0, + 394.0, + 908.0, + 412.0, + 882.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 393.0, + 1178.0, + 393.0, + 1178.0, + 415.0, + 1147.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 387.0, + 1320.0, + 387.0, + 1320.0, + 419.0, + 1270.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 414.0, + 1324.0, + 414.0, + 1324.0, + 432.0, + 1270.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 425.0, + 347.0, + 425.0, + 347.0, + 443.0, + 320.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 422.0, + 641.0, + 422.0, + 641.0, + 459.0, + 604.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 424.0, + 912.0, + 424.0, + 912.0, + 446.0, + 879.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 424.0, + 1180.0, + 424.0, + 1180.0, + 446.0, + 1146.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 435.0, + 362.0, + 435.0, + 362.0, + 457.0, + 331.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 436.0, + 401.0, + 436.0, + 401.0, + 455.0, + 376.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 437.0, + 442.0, + 437.0, + 442.0, + 455.0, + 416.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 437.0, + 484.0, + 437.0, + 484.0, + 455.0, + 457.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 437.0, + 525.0, + 437.0, + 525.0, + 455.0, + 498.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 435.0, + 568.0, + 435.0, + 568.0, + 457.0, + 538.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 437.0, + 677.0, + 437.0, + 677.0, + 455.0, + 651.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 437.0, + 716.0, + 437.0, + 716.0, + 455.0, + 689.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 437.0, + 756.0, + 437.0, + 756.0, + 455.0, + 727.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 437.0, + 794.0, + 437.0, + 794.0, + 455.0, + 766.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 437.0, + 833.0, + 437.0, + 833.0, + 455.0, + 806.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 437.0, + 920.0, + 437.0, + 920.0, + 455.0, + 894.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 437.0, + 973.0, + 437.0, + 973.0, + 455.0, + 946.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 437.0, + 1024.0, + 437.0, + 1024.0, + 455.0, + 998.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 437.0, + 1075.0, + 437.0, + 1075.0, + 455.0, + 1049.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 437.0, + 1127.0, + 437.0, + 1127.0, + 456.0, + 1101.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 437.0, + 1188.0, + 437.0, + 1188.0, + 456.0, + 1159.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 437.0, + 1240.0, + 437.0, + 1240.0, + 455.0, + 1214.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 437.0, + 1291.0, + 437.0, + 1291.0, + 455.0, + 1264.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 436.0, + 1343.0, + 436.0, + 1343.0, + 455.0, + 1316.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 437.0, + 1394.0, + 437.0, + 1394.0, + 456.0, + 1368.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 447.0, + 521.0, + 447.0, + 521.0, + 471.0, + 379.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 447.0, + 793.0, + 447.0, + 793.0, + 471.0, + 651.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 447.0, + 1082.0, + 447.0, + 1082.0, + 471.0, + 941.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 447.0, + 1347.0, + 447.0, + 1347.0, + 471.0, + 1207.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 487.0, + 519.0, + 487.0, + 519.0, + 519.0, + 353.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 484.0, + 764.0, + 484.0, + 764.0, + 520.0, + 641.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 487.0, + 1120.0, + 487.0, + 1120.0, + 518.0, + 875.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 487.0, + 1361.0, + 487.0, + 1361.0, + 518.0, + 1167.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1529.0, + 549.0, + 1529.0, + 549.0, + 1577.0, + 292.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1921.0, + 558.0, + 1921.0, + 558.0, + 1949.0, + 299.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1065.0, + 838.0, + 1065.0, + 838.0, + 1115.0, + 293.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1065.0, + 898.0, + 1065.0, + 898.0, + 1115.0, + 894.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 667.0, + 894.0, + 667.0, + 894.0, + 703.0, + 297.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 668.0, + 892.0, + 668.0, + 892.0, + 702.0, + 296.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 724.0, + 1405.0, + 724.0, + 1405.0, + 761.0, + 293.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 753.0, + 851.0, + 753.0, + 851.0, + 794.0, + 293.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 753.0, + 1404.0, + 753.0, + 1404.0, + 794.0, + 1188.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 786.0, + 1040.0, + 786.0, + 1040.0, + 823.0, + 292.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 786.0, + 1405.0, + 786.0, + 1405.0, + 823.0, + 1236.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 818.0, + 1227.0, + 818.0, + 1227.0, + 854.0, + 293.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 818.0, + 1406.0, + 818.0, + 1406.0, + 854.0, + 1315.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 849.0, + 1405.0, + 849.0, + 1405.0, + 884.0, + 294.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 879.0, + 1406.0, + 879.0, + 1406.0, + 915.0, + 294.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 907.0, + 771.0, + 907.0, + 771.0, + 945.0, + 293.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 907.0, + 1406.0, + 907.0, + 1406.0, + 945.0, + 952.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 939.0, + 1405.0, + 939.0, + 1405.0, + 975.0, + 293.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 970.0, + 1405.0, + 970.0, + 1405.0, + 1006.0, + 294.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1001.0, + 1362.0, + 1001.0, + 1362.0, + 1036.0, + 294.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1270.0, + 557.0, + 1270.0, + 557.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1270.0, + 1405.0, + 1270.0, + 1405.0, + 1304.0, + 614.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1298.0, + 1409.0, + 1298.0, + 1409.0, + 1338.0, + 292.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1325.0, + 514.0, + 1325.0, + 514.0, + 1372.0, + 291.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1325.0, + 749.0, + 1325.0, + 749.0, + 1372.0, + 571.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1325.0, + 984.0, + 1325.0, + 984.0, + 1372.0, + 897.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 1325.0, + 1410.0, + 1325.0, + 1410.0, + 1372.0, + 1348.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1362.0, + 1406.0, + 1362.0, + 1406.0, + 1397.0, + 294.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1391.0, + 1406.0, + 1391.0, + 1406.0, + 1428.0, + 292.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1421.0, + 297.0, + 1421.0, + 297.0, + 1457.0, + 292.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1421.0, + 1407.0, + 1421.0, + 1407.0, + 1457.0, + 353.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1452.0, + 857.0, + 1452.0, + 857.0, + 1491.0, + 294.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1331.5, + 905.0, + 1331.5, + 905.0, + 1368.0, + 830.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1129.0, + 1054.0, + 1129.0, + 1054.0, + 1169.0, + 293.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1129.0, + 1407.0, + 1129.0, + 1407.0, + 1169.0, + 1111.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1162.0, + 444.0, + 1162.0, + 444.0, + 1198.0, + 294.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1162.0, + 1409.0, + 1162.0, + 1409.0, + 1198.0, + 708.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1193.0, + 575.0, + 1193.0, + 575.0, + 1229.0, + 294.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1193.0, + 1407.0, + 1193.0, + 1407.0, + 1229.0, + 631.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1223.0, + 1407.0, + 1223.0, + 1407.0, + 1258.0, + 294.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1638.0, + 296.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1634.0, + 1406.0, + 1634.0, + 1406.0, + 1671.0, + 292.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1664.0, + 1406.0, + 1664.0, + 1406.0, + 1700.0, + 293.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1696.0, + 1406.0, + 1696.0, + 1406.0, + 1732.0, + 295.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1726.0, + 1408.0, + 1726.0, + 1408.0, + 1762.0, + 293.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1755.0, + 1406.0, + 1755.0, + 1406.0, + 1793.0, + 293.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1785.0, + 1405.0, + 1785.0, + 1405.0, + 1823.0, + 293.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1819.0, + 1406.0, + 1819.0, + 1406.0, + 1853.0, + 293.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1850.0, + 1371.0, + 1850.0, + 1371.0, + 1883.0, + 295.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2000.0, + 769.0, + 2000.0, + 769.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 772, + 1406, + 772, + 1406, + 867, + 297, + 867 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 298, + 1122, + 1409, + 1122, + 1409, + 1248, + 298, + 1248 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 298, + 1005, + 1408, + 1005, + 1408, + 1101, + 298, + 1101 + ], + "score": 0.881 + }, + { + "category_id": 1, + "poly": [ + 299, + 1503, + 1407, + 1503, + 1407, + 1600, + 299, + 1600 + ], + "score": 0.872 + }, + { + "category_id": 1, + "poly": [ + 298, + 1620, + 1406, + 1620, + 1406, + 1716, + 298, + 1716 + ], + "score": 0.858 + }, + { + "category_id": 1, + "poly": [ + 299, + 1386, + 1403, + 1386, + 1403, + 1481, + 299, + 1481 + ], + "score": 0.857 + }, + { + "category_id": 1, + "poly": [ + 298, + 888, + 1404, + 888, + 1404, + 985, + 298, + 985 + ], + "score": 0.852 + }, + { + "category_id": 1, + "poly": [ + 296, + 1269, + 1401, + 1269, + 1401, + 1364, + 296, + 1364 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 298, + 654, + 1407, + 654, + 1407, + 750, + 298, + 750 + ], + "score": 0.844 + }, + { + "category_id": 1, + "poly": [ + 289, + 1737, + 1404, + 1737, + 1404, + 1803, + 289, + 1803 + ], + "score": 0.829 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.816 + }, + { + "category_id": 0, + "poly": [ + 299, + 227, + 489, + 227, + 489, + 262, + 299, + 262 + ], + "score": 0.811 + }, + { + "category_id": 1, + "poly": [ + 292, + 451, + 1402, + 451, + 1402, + 517, + 292, + 517 + ], + "score": 0.786 + }, + { + "category_id": 1, + "poly": [ + 300, + 1941, + 1404, + 1941, + 1404, + 2035, + 300, + 2035 + ], + "score": 0.771 + }, + { + "category_id": 1, + "poly": [ + 302, + 1824, + 1402, + 1824, + 1402, + 1920, + 302, + 1920 + ], + "score": 0.768 + }, + { + "category_id": 1, + "poly": [ + 297, + 538, + 1409, + 538, + 1409, + 633, + 297, + 633 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 293, + 279, + 1402, + 279, + 1402, + 345, + 293, + 345 + ], + "score": 0.764 + }, + { + "category_id": 1, + "poly": [ + 293, + 365, + 1402, + 365, + 1402, + 431, + 293, + 431 + ], + "score": 0.755 + }, + { + "category_id": 13, + "poly": [ + 904, + 1073, + 922, + 1073, + 922, + 1095, + 904, + 1095 + ], + "score": 0.65, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1166, + 2006, + 1184, + 2006, + 1184, + 2029, + 1166, + 2029 + ], + "score": 0.64, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =" + }, + { + "category_id": 13, + "poly": [ + 526, + 1220, + 544, + 1220, + 544, + 1242, + 526, + 1242 + ], + "score": 0.54, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 841, + 955, + 860, + 955, + 860, + 978, + 841, + 978 + ], + "score": 0.42, + "latex": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 773.0, + 1406.0, + 773.0, + 1406.0, + 807.0, + 295.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 805.0, + 1404.0, + 805.0, + 1404.0, + 839.0, + 324.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 833.0, + 538.0, + 833.0, + 538.0, + 868.0, + 324.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1123.0, + 1409.0, + 1123.0, + 1409.0, + 1160.0, + 293.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1156.0, + 1405.0, + 1156.0, + 1405.0, + 1189.0, + 322.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1186.0, + 1414.0, + 1186.0, + 1414.0, + 1223.0, + 322.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1217.0, + 525.0, + 1217.0, + 525.0, + 1250.0, + 322.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1217.0, + 721.0, + 1217.0, + 721.0, + 1250.0, + 545.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1006.0, + 1406.0, + 1006.0, + 1406.0, + 1041.0, + 296.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1038.0, + 1409.0, + 1038.0, + 1409.0, + 1071.0, + 321.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1065.0, + 903.0, + 1065.0, + 903.0, + 1104.0, + 322.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1065.0, + 1084.0, + 1065.0, + 1084.0, + 1104.0, + 923.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1501.0, + 1408.0, + 1501.0, + 1408.0, + 1542.0, + 293.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1535.0, + 1409.0, + 1535.0, + 1409.0, + 1570.0, + 325.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1566.0, + 973.0, + 1566.0, + 973.0, + 1601.0, + 325.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1618.0, + 1404.0, + 1618.0, + 1404.0, + 1658.0, + 293.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1652.0, + 1406.0, + 1652.0, + 1406.0, + 1689.0, + 321.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1683.0, + 841.0, + 1683.0, + 841.0, + 1716.0, + 323.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1386.0, + 1404.0, + 1386.0, + 1404.0, + 1424.0, + 295.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1419.0, + 1408.0, + 1419.0, + 1408.0, + 1453.0, + 322.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1447.0, + 396.0, + 1447.0, + 396.0, + 1481.0, + 320.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 891.0, + 1402.0, + 891.0, + 1402.0, + 922.0, + 297.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 920.0, + 1406.0, + 920.0, + 1406.0, + 955.0, + 323.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 953.0, + 840.0, + 953.0, + 840.0, + 987.0, + 325.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 953.0, + 1021.0, + 953.0, + 1021.0, + 987.0, + 861.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1268.0, + 1406.0, + 1268.0, + 1406.0, + 1309.0, + 294.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1301.0, + 1405.0, + 1301.0, + 1405.0, + 1335.0, + 323.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1332.0, + 620.0, + 1332.0, + 620.0, + 1366.0, + 321.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 653.0, + 1405.0, + 653.0, + 1405.0, + 693.0, + 293.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 685.0, + 1407.0, + 685.0, + 1407.0, + 725.0, + 320.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 718.0, + 1359.0, + 718.0, + 1359.0, + 752.0, + 324.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1738.0, + 1407.0, + 1738.0, + 1407.0, + 1774.0, + 295.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1769.0, + 1198.0, + 1769.0, + 1198.0, + 1805.0, + 323.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 452.0, + 1404.0, + 452.0, + 1404.0, + 488.0, + 296.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 483.0, + 1152.0, + 483.0, + 1152.0, + 519.0, + 321.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1941.0, + 1408.0, + 1941.0, + 1408.0, + 1978.0, + 296.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2011.0, + 320.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2003.0, + 1165.0, + 2003.0, + 1165.0, + 2040.0, + 321.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 2003.0, + 1361.0, + 2003.0, + 1361.0, + 2040.0, + 1185.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1822.0, + 1406.0, + 1822.0, + 1406.0, + 1862.0, + 296.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1856.0, + 1406.0, + 1856.0, + 1406.0, + 1890.0, + 323.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1887.0, + 1025.0, + 1887.0, + 1025.0, + 1921.0, + 322.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 537.0, + 1409.0, + 537.0, + 1409.0, + 576.0, + 292.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 570.0, + 1409.0, + 570.0, + 1409.0, + 605.0, + 320.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 600.0, + 888.0, + 600.0, + 888.0, + 633.0, + 323.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 277.0, + 1404.0, + 277.0, + 1404.0, + 317.0, + 293.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 308.0, + 1228.0, + 308.0, + 1228.0, + 347.0, + 322.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 366.0, + 1404.0, + 366.0, + 1404.0, + 402.0, + 296.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 395.0, + 870.0, + 395.0, + 870.0, + 432.0, + 322.0, + 432.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 295, + 862, + 1404, + 862, + 1404, + 957, + 295, + 957 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 300, + 671, + 1402, + 671, + 1402, + 766, + 300, + 766 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 299, + 1195, + 1403, + 1195, + 1403, + 1290, + 299, + 1290 + ], + "score": 0.828 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 861, + 2088, + 861, + 2113, + 835, + 2113 + ], + "score": 0.82 + }, + { + "category_id": 1, + "poly": [ + 296, + 1748, + 1400, + 1748, + 1400, + 1814, + 296, + 1814 + ], + "score": 0.807 + }, + { + "category_id": 1, + "poly": [ + 298, + 1448, + 1405, + 1448, + 1405, + 1542, + 298, + 1542 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 292, + 782, + 1403, + 782, + 1403, + 847, + 292, + 847 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 297, + 1639, + 1405, + 1639, + 1405, + 1733, + 297, + 1733 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 298, + 1305, + 1407, + 1305, + 1407, + 1430, + 298, + 1430 + ], + "score": 0.786 + }, + { + "category_id": 1, + "poly": [ + 295, + 1557, + 1401, + 1557, + 1401, + 1622, + 295, + 1622 + ], + "score": 0.786 + }, + { + "category_id": 1, + "poly": [ + 299, + 1084, + 1403, + 1084, + 1403, + 1180, + 299, + 1180 + ], + "score": 0.78 + }, + { + "category_id": 1, + "poly": [ + 300, + 973, + 1404, + 973, + 1404, + 1069, + 300, + 1069 + ], + "score": 0.772 + }, + { + "category_id": 1, + "poly": [ + 295, + 1830, + 1403, + 1830, + 1403, + 1895, + 295, + 1895 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 290, + 451, + 1403, + 451, + 1403, + 515, + 290, + 515 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1404, + 229, + 1404, + 323, + 298, + 323 + ], + "score": 0.749 + }, + { + "category_id": 1, + "poly": [ + 296, + 531, + 1405, + 531, + 1405, + 653, + 296, + 653 + ], + "score": 0.74 + }, + { + "category_id": 1, + "poly": [ + 299, + 1910, + 1404, + 1910, + 1404, + 2035, + 299, + 2035 + ], + "score": 0.707 + }, + { + "category_id": 1, + "poly": [ + 302, + 340, + 1400, + 340, + 1400, + 434, + 302, + 434 + ], + "score": 0.689 + }, + { + "category_id": 13, + "poly": [ + 525, + 297, + 544, + 297, + 544, + 318, + 525, + 318 + ], + "score": 0.51, + "latex": "{ . } = { }" + }, + { + "category_id": 13, + "poly": [ + 1383, + 1372, + 1403, + 1372, + 1403, + 1395, + 1383, + 1395 + ], + "score": 0.3, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1383, + 597, + 1403, + 597, + 1403, + 620, + 1383, + 620 + ], + "score": 0.27, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 862.0, + 1405.0, + 862.0, + 1405.0, + 899.0, + 293.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 891.0, + 1410.0, + 891.0, + 1410.0, + 934.0, + 321.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 925.0, + 533.0, + 925.0, + 533.0, + 957.0, + 324.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 708.0, + 294.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 703.0, + 1407.0, + 703.0, + 1407.0, + 739.0, + 320.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 728.0, + 398.0, + 728.0, + 398.0, + 769.0, + 320.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1236.0, + 292.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1228.0, + 1404.0, + 1228.0, + 1404.0, + 1262.0, + 323.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1256.0, + 763.0, + 1256.0, + 763.0, + 1294.0, + 323.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1789.0, + 292.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1781.0, + 1057.0, + 1781.0, + 1057.0, + 1815.0, + 322.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1446.0, + 1405.0, + 1446.0, + 1405.0, + 1484.0, + 293.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1478.0, + 1407.0, + 1478.0, + 1407.0, + 1516.0, + 323.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1508.0, + 1410.0, + 1508.0, + 1410.0, + 1546.0, + 323.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 782.0, + 1405.0, + 782.0, + 1405.0, + 820.0, + 295.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 815.0, + 1270.0, + 815.0, + 1270.0, + 848.0, + 321.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1639.0, + 1406.0, + 1639.0, + 1406.0, + 1675.0, + 292.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1671.0, + 1405.0, + 1671.0, + 1405.0, + 1705.0, + 324.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1701.0, + 692.0, + 1701.0, + 692.0, + 1734.0, + 322.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1307.0, + 1409.0, + 1307.0, + 1409.0, + 1343.0, + 295.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1339.0, + 1408.0, + 1339.0, + 1408.0, + 1372.0, + 324.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1368.0, + 1382.0, + 1368.0, + 1382.0, + 1404.0, + 320.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1368.0, + 1408.0, + 1368.0, + 1408.0, + 1404.0, + 1404.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1401.0, + 505.0, + 1401.0, + 505.0, + 1431.0, + 324.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1556.0, + 1405.0, + 1556.0, + 1405.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1589.0, + 525.0, + 1589.0, + 525.0, + 1622.0, + 323.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1081.0, + 1409.0, + 1081.0, + 1409.0, + 1126.0, + 292.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1114.0, + 1405.0, + 1114.0, + 1405.0, + 1154.0, + 321.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1149.0, + 1132.0, + 1149.0, + 1132.0, + 1181.0, + 323.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 972.0, + 1408.0, + 972.0, + 1408.0, + 1012.0, + 294.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1006.0, + 1405.0, + 1006.0, + 1405.0, + 1041.0, + 323.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1036.0, + 1341.0, + 1036.0, + 1341.0, + 1074.0, + 321.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1829.0, + 1406.0, + 1829.0, + 1406.0, + 1868.0, + 295.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1859.0, + 1200.0, + 1859.0, + 1200.0, + 1899.0, + 322.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 450.0, + 1402.0, + 450.0, + 1402.0, + 486.0, + 294.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 482.0, + 1109.0, + 482.0, + 1109.0, + 516.0, + 318.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1411.0, + 260.0, + 1411.0, + 297.0, + 322.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 292.0, + 524.0, + 292.0, + 524.0, + 326.0, + 323.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 292.0, + 720.0, + 292.0, + 720.0, + 326.0, + 545.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 533.0, + 1405.0, + 533.0, + 1405.0, + 565.0, + 295.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 563.0, + 1408.0, + 563.0, + 1408.0, + 595.0, + 323.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 593.0, + 1382.0, + 593.0, + 1382.0, + 629.0, + 320.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 593.0, + 1407.0, + 593.0, + 1407.0, + 629.0, + 1404.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 626.0, + 505.0, + 626.0, + 505.0, + 654.0, + 323.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1944.0, + 295.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 322.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2009.0, + 321.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2003.0, + 1126.0, + 2003.0, + 1126.0, + 2037.0, + 322.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 340.0, + 1405.0, + 340.0, + 1405.0, + 377.0, + 296.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 372.0, + 1404.0, + 372.0, + 1404.0, + 405.0, + 322.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 403.0, + 802.0, + 403.0, + 802.0, + 437.0, + 323.0, + 437.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.86 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1404, + 229, + 1404, + 293, + 294, + 293 + ], + "score": 0.833 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 297, + 395, + 1402, + 395, + 1402, + 460, + 297, + 460 + ], + "score": 0.782 + }, + { + "category_id": 1, + "poly": [ + 296, + 311, + 1404, + 311, + 1404, + 377, + 296, + 377 + ], + "score": 0.761 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1407.0, + 225.0, + 1407.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 260.0, + 570.0, + 260.0, + 570.0, + 296.0, + 321.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 395.0, + 1407.0, + 395.0, + 1407.0, + 431.0, + 296.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 426.0, + 1396.0, + 426.0, + 1396.0, + 462.0, + 324.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 310.0, + 1405.0, + 310.0, + 1405.0, + 348.0, + 294.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 343.0, + 1407.0, + 343.0, + 1407.0, + 379.0, + 323.0, + 379.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1206, + 1404, + 1206, + 1404, + 1362, + 297, + 1362 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 791, + 1405, + 791, + 1405, + 946, + 297, + 946 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1068, + 1404, + 1068, + 1404, + 1192, + 298, + 1192 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1537, + 1405, + 1537, + 1405, + 1662, + 297, + 1662 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 960, + 1404, + 960, + 1404, + 1053, + 298, + 1053 + ], + "score": 0.971 + }, + { + "category_id": 5, + "poly": [ + 398, + 393, + 1299, + 393, + 1299, + 617, + 398, + 617 + ], + "score": 0.967, + "html": "
Jaco Pick-Push-PlaceJaco Bar-MovingAnt Push
Observation Space8888100
- Robot observation626282
- Object observation262618
Action Space181816
Number of Agents222
Episode length150100200
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1459, + 1401, + 1459, + 1401, + 1524, + 298, + 1524 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 1860, + 1401, + 1860, + 1401, + 1924, + 296, + 1924 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 295, + 1779, + 1401, + 1779, + 1401, + 1845, + 295, + 1845 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 300, + 1402, + 576, + 1402, + 576, + 1433, + 300, + 1433 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 302, + 227, + 704, + 227, + 704, + 262, + 302, + 262 + ], + "score": 0.909 + }, + { + "category_id": 6, + "poly": [ + 687, + 641, + 1012, + 641, + 1012, + 674, + 687, + 674 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 301, + 733, + 732, + 733, + 732, + 764, + 301, + 764 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 817, + 74, + 817, + 105, + 299, + 105 + ], + "score": 0.904 + }, + { + "category_id": 8, + "poly": [ + 372, + 1680, + 1326, + 1680, + 1326, + 1765, + 372, + 1765 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 295, + 297, + 1406, + 297, + 1406, + 361, + 295, + 361 + ], + "score": 0.744 + }, + { + "category_id": 1, + "poly": [ + 295, + 2000, + 1015, + 2000, + 1015, + 2038, + 295, + 2038 + ], + "score": 0.647 + }, + { + "category_id": 8, + "poly": [ + 347, + 1942, + 1321, + 1942, + 1321, + 1983, + 347, + 1983 + ], + "score": 0.476 + }, + { + "category_id": 8, + "poly": [ + 303, + 1942, + 1324, + 1942, + 1324, + 2039, + 303, + 2039 + ], + "score": 0.172 + }, + { + "category_id": 1, + "poly": [ + 347, + 1942, + 1321, + 1942, + 1321, + 1983, + 347, + 1983 + ], + "score": 0.163 + }, + { + "category_id": 8, + "poly": [ + 295, + 2000, + 1015, + 2000, + 1015, + 2038, + 295, + 2038 + ], + "score": 0.144 + }, + { + "category_id": 13, + "poly": [ + 741, + 1460, + 905, + 1460, + 905, + 1495, + 741, + 1495 + ], + "score": 0.93, + "latex": "- 0 . 0 0 1 * \\| a \\| ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 490, + 1239, + 668, + 1239, + 668, + 1268, + 490, + 1268 + ], + "score": 0.91, + "latex": "8 . 0 \\times 1 . 6 \\times 1 . 6" + }, + { + "category_id": 13, + "poly": [ + 295, + 1813, + 403, + 1813, + 403, + 1844, + 295, + 1844 + ], + "score": 0.9, + "latex": "1 \\times 1 0 ^ { - \\hat { 4 } }" + }, + { + "category_id": 14, + "poly": [ + 372, + 1678, + 1332, + 1678, + 1332, + 1767, + 372, + 1767 + ], + "score": 0.87, + "latex": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { p i c k } \\cdot ( z _ { b o x } - z _ { i n i t } ) + \\lambda _ { d i s t } \\cdot \\mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \\lambda _ { p o s } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { i n i t } ) + } \\\\ & { \\qquad \\lambda _ { q u a t } \\cdot \\mathrm { a b s } ( \\Delta _ { q u a t } ) + \\lambda _ { h o l d } \\cdot t _ { h o l d } + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 296, + 822, + 418, + 822, + 418, + 855, + 296, + 855 + ], + "score": 0.86, + "latex": "( 0 , 0 , 0 . 8 2 )" + }, + { + "category_id": 13, + "poly": [ + 924, + 823, + 1146, + 823, + 1146, + 854, + 924, + 854 + ], + "score": 0.84, + "latex": "( - 0 . 1 6 , - 0 . 1 6 , 1 . 2 )" + }, + { + "category_id": 13, + "poly": [ + 372, + 2000, + 1007, + 2000, + 1007, + 2037, + 372, + 2037 + ], + "score": 0.81, + "latex": "\\lambda _ { x y } = 5 0 0 , \\lambda _ { z } = 5 0 0 , \\lambda _ { s u c c e s s } = 5 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1498, + 315, + 1498, + 315, + 1518, + 298, + 1518 + ], + "score": 0.73, + "latex": "a" + }, + { + "category_id": 14, + "poly": [ + 350, + 1941, + 1324, + 1941, + 1324, + 1982, + 350, + 1982 + ], + "score": 0.72, + "latex": "R ( s ) = \\lambda _ { x y } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b o x } , p _ { g o a l } ) + \\lambda _ { z } \\cdot | z _ { b o x } - z _ { g o a l } | + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } ," + }, + { + "category_id": 13, + "poly": [ + 1197, + 822, + 1399, + 822, + 1399, + 855, + 1197, + 855 + ], + "score": 0.65, + "latex": "( \\bar { - } 0 . 1 6 , 0 . \\bar { 2 } 4 , 1 . 2 )" + }, + { + "category_id": 13, + "poly": [ + 647, + 1270, + 718, + 1270, + 718, + 1298, + 647, + 1298 + ], + "score": 0.65, + "latex": "3 0 \\mathrm { c m }" + }, + { + "category_id": 13, + "poly": [ + 1142, + 1239, + 1216, + 1239, + 1216, + 1268, + 1142, + 1268 + ], + "score": 0.58, + "latex": "2 0 \\mathrm { c m }" + }, + { + "category_id": 13, + "poly": [ + 874, + 1331, + 932, + 1331, + 932, + 1358, + 874, + 1358 + ], + "score": 0.55, + "latex": "5 \\mathrm { c m }" + }, + { + "category_id": 14, + "poly": [ + 367, + 1779, + 1404, + 1779, + 1404, + 1818, + 367, + 1818 + ], + "score": 0.52, + "latex": "\\lambda _ { p i c k } = 5 0 0 , \\lambda _ { d i s t } = 1 0 0 , \\lambda _ { p o s } = 1 0 0 0 , \\lambda _ { q u a t } = 1 0 0 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 0 0 0 , \\lambda _ { d i s t }" + }, + { + "category_id": 13, + "poly": [ + 370, + 1780, + 1407, + 1780, + 1407, + 1817, + 370, + 1817 + ], + "score": 0.44, + "latex": "\\lambda _ { p i c k } = 5 0 0 , \\lambda _ { d i s t } = 1 0 0 , \\lambda _ { p o s } = 1 0 0 0 , \\lambda _ { q u a t } = 1 0 0 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 0 0 0 , \\lambda _ { d i s t }" + }, + { + "category_id": 13, + "poly": [ + 1090, + 855, + 1265, + 855, + 1265, + 886, + 1090, + 886 + ], + "score": 0.43, + "latex": "0 . 3 , - 0 . 2 , 0 . 8 6 )" + }, + { + "category_id": 14, + "poly": [ + 356, + 1938, + 1324, + 1938, + 1324, + 2041, + 356, + 2041 + ], + "score": 0.43, + "latex": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { x y } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b o x } , p _ { g o a l } ) + \\lambda _ { z } \\cdot | z _ { b o x } - z _ { g o a l } | + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { \\mathrm { s u c c e s s } } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , } \\\\ & { \\mathrm { ~ e \\lambda _ { x y } = 5 0 0 , } \\lambda _ { z } = 5 0 0 , \\lambda _ { s u c c e s s } = 5 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 964, + 1270, + 1022, + 1270, + 1022, + 1298, + 964, + 1298 + ], + "score": 0.39, + "latex": "1 \\mathrm { c m }" + }, + { + "category_id": 13, + "poly": [ + 862, + 854, + 1027, + 854, + 1027, + 886, + 862, + 886 + ], + "score": 0.35, + "latex": "( 0 . 3 , 0 . 2 , 0 . 8 6 )" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1397.0, + 580.0, + 1397.0, + 580.0, + 1438.0, + 294.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 224.0, + 706.0, + 224.0, + 706.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 641.0, + 1013.0, + 641.0, + 1013.0, + 674.0, + 686.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 733.0, + 736.0, + 733.0, + 736.0, + 768.0, + 296.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1205.0, + 1405.0, + 1205.0, + 1405.0, + 1243.0, + 294.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1236.0, + 489.0, + 1236.0, + 489.0, + 1272.0, + 294.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1236.0, + 1141.0, + 1236.0, + 1141.0, + 1272.0, + 669.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1236.0, + 1405.0, + 1236.0, + 1405.0, + 1272.0, + 1217.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1267.0, + 646.0, + 1267.0, + 646.0, + 1303.0, + 294.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1267.0, + 963.0, + 1267.0, + 963.0, + 1303.0, + 719.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1267.0, + 1405.0, + 1267.0, + 1405.0, + 1303.0, + 1023.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1295.0, + 1406.0, + 1295.0, + 1406.0, + 1335.0, + 292.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1328.0, + 873.0, + 1328.0, + 873.0, + 1364.0, + 294.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1328.0, + 941.0, + 1328.0, + 941.0, + 1364.0, + 933.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 790.0, + 1405.0, + 790.0, + 1405.0, + 826.0, + 294.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 821.0, + 923.0, + 821.0, + 923.0, + 858.0, + 419.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 821.0, + 1196.0, + 821.0, + 1196.0, + 858.0, + 1147.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 821.0, + 1407.0, + 821.0, + 1407.0, + 858.0, + 1400.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 850.0, + 861.0, + 850.0, + 861.0, + 890.0, + 292.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 850.0, + 1089.0, + 850.0, + 1089.0, + 890.0, + 1028.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 850.0, + 1405.0, + 850.0, + 1405.0, + 890.0, + 1266.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 884.0, + 1405.0, + 884.0, + 1405.0, + 918.0, + 295.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 916.0, + 536.0, + 916.0, + 536.0, + 950.0, + 295.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1069.0, + 1409.0, + 1069.0, + 1409.0, + 1102.0, + 294.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1095.0, + 1409.0, + 1095.0, + 1409.0, + 1137.0, + 291.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1131.0, + 1404.0, + 1131.0, + 1404.0, + 1164.0, + 296.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1162.0, + 910.0, + 1162.0, + 910.0, + 1194.0, + 296.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1574.0, + 292.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1570.0, + 1407.0, + 1570.0, + 1407.0, + 1603.0, + 295.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1633.0, + 296.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1629.0, + 583.0, + 1629.0, + 583.0, + 1665.0, + 293.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 961.0, + 1404.0, + 961.0, + 1404.0, + 994.0, + 294.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1404.0, + 991.0, + 1404.0, + 1025.0, + 294.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1022.0, + 729.0, + 1022.0, + 729.0, + 1057.0, + 295.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1457.0, + 740.0, + 1457.0, + 740.0, + 1497.0, + 293.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1457.0, + 1403.0, + 1457.0, + 1403.0, + 1497.0, + 906.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1491.0, + 297.0, + 1491.0, + 297.0, + 1528.0, + 293.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1491.0, + 778.0, + 1491.0, + 778.0, + 1528.0, + 316.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1857.0, + 1405.0, + 1857.0, + 1405.0, + 1898.0, + 291.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1893.0, + 1219.0, + 1893.0, + 1219.0, + 1923.0, + 295.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1776.0, + 366.0, + 1776.0, + 366.0, + 1819.0, + 292.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1804.0, + 294.0, + 1804.0, + 294.0, + 1851.0, + 289.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1804.0, + 416.0, + 1804.0, + 416.0, + 1851.0, + 404.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 297.0, + 1404.0, + 297.0, + 1404.0, + 333.0, + 295.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 328.0, + 1138.0, + 328.0, + 1138.0, + 364.0, + 294.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1994.0, + 355.0, + 1994.0, + 355.0, + 2045.0, + 290.0, + 2045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1325.0, + 1937.0, + 1330.0, + 1937.0, + 1330.0, + 1988.0, + 1325.0, + 1988.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 916, + 1406, + 916, + 1406, + 1041, + 297, + 1041 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1458, + 1406, + 1458, + 1406, + 1553, + 298, + 1553 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1349, + 1406, + 1349, + 1406, + 1445, + 299, + 1445 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 297, + 551, + 1403, + 551, + 1403, + 646, + 297, + 646 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 295, + 228, + 1407, + 228, + 1407, + 324, + 295, + 324 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 292, + 1054, + 1401, + 1054, + 1401, + 1120, + 292, + 1120 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 296, + 1729, + 1404, + 1729, + 1404, + 1795, + 296, + 1795 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 296, + 1971, + 1404, + 1971, + 1404, + 2035, + 296, + 2035 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 316, + 1170, + 1387, + 1170, + 1387, + 1299, + 316, + 1299 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 377, + 697, + 1326, + 697, + 1326, + 871, + 377, + 871 + ], + "score": 0.934 + }, + { + "category_id": 8, + "poly": [ + 427, + 368, + 1275, + 368, + 1275, + 453, + 427, + 453 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 299, + 1878, + 675, + 1878, + 675, + 1914, + 299, + 1914 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 817, + 74, + 817, + 106, + 299, + 106 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 295, + 502, + 1387, + 502, + 1387, + 541, + 295, + 541 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 863, + 2088, + 863, + 2113, + 835, + 2113 + ], + "score": 0.864 + }, + { + "category_id": 8, + "poly": [ + 313, + 1602, + 1393, + 1602, + 1393, + 1681, + 313, + 1681 + ], + "score": 0.797 + }, + { + "category_id": 1, + "poly": [ + 313, + 1602, + 1393, + 1602, + 1393, + 1681, + 313, + 1681 + ], + "score": 0.158 + }, + { + "category_id": 14, + "poly": [ + 373, + 694, + 1331, + 694, + 1331, + 872, + 373, + 872 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { c o n t a c t } \\cdot \\big ( { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { l e f f ~ g r i p p e r ~ r o u c h e s ~ c o n t a i n e r } } } + { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { r i g h t ~ g r i p p e r ~ t o u c h e s ~ b o x } } } \\big ) + } \\\\ & { \\quad \\quad \\lambda _ { r e a c h } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { r e a c h } } } \\cdot \\big ( \\mathrm { d i s t } ( p _ { l e f t . g r i p p e r } , p _ { c o n t a i n e r } ) + \\mathrm { d i s t } ( p _ { r i g h t . g r i p p e r } , p _ { b o x } ) \\big ) + } \\\\ & { \\quad \\quad \\lambda _ { p i c k } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { p i c k } } } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + \\lambda _ { p l a c e } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { p l a c e } } } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { b o x . t a r g e t } ) + } \\\\ & { \\quad \\quad \\lambda _ { p u s h } \\cdot \\mathrm { d i s t } ( p _ { c o n t a i n e r } , p _ { c o n t a i n e r . t a r g e t } ) + \\lambda _ { s u c c e s s } \\cdot { \\bf 1 } _ { \\scriptscriptstyle { \\mathrm { s u c c e s s } } } + \\lambda _ { c t r l } \\cdot \\| a \\| ^ { 2 } , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 313, + 1167, + 1390, + 1167, + 1390, + 1301, + 313, + 1301 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { r e a c h } \\cdot { \\bf 1 } _ { r e a c h } \\cdot \\big ( \\mathrm { d i s t } ( p _ { l e f t \\_ g r i p p e r } , p _ { l e f t \\_ h a n d l e } ) + \\mathrm { d i s t } ( p _ { r i g h t \\_ g r i p p e r } , p _ { r i g h t \\_ h a n d l e } ) \\big ) + } \\\\ & { \\qquad \\lambda _ { p i c k } \\cdot { \\bf 1 } _ { \\mathrm { p i c k } } \\cdot \\mathrm { d i s t } ( p _ { b a r } , p _ { b a r \\_ t a r g e t } ) + \\lambda _ { m o v e } \\cdot { \\bf 1 } _ { \\mathrm { p l a e } } \\cdot \\mathrm { d i s t } _ { x y } ( p _ { b a r \\_ p b a r \\_ t a r g e t } ) + } \\\\ & { \\qquad \\lambda _ { p l a c e } \\cdot { \\bf 1 } _ { \\mathrm { p l a e } } \\cdot \\mathrm { d i s t } _ { z } ( p _ { b a r \\_ p b a r \\_ t a r g e t } ) + \\lambda _ { s u c e s s } \\cdot { \\bf 1 } _ { \\mathrm { s u c e s s } } + \\lambda _ { c t r l } \\cdot \\| a \\| ^ { 2 } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1183, + 1762, + 1323, + 1762, + 1323, + 1796, + 1183, + 1796 + ], + "score": 0.92, + "latex": "\\lambda _ { g o a l } = 2 0 0" + }, + { + "category_id": 14, + "poly": [ + 425, + 366, + 1275, + 366, + 1275, + 454, + 425, + 454 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { R ( s ) = \\lambda _ { r e a c h i n g } \\cdot \\mathrm { d i s t } ( p _ { g r i p p e r } , p _ { b o x } ) + \\lambda _ { p o s } \\cdot \\mathrm { d i s t } ( p _ { b o x } , p _ { d e s t } ) + \\phantom { x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x } } \\\\ { \\lambda _ { q u a t } \\cdot \\mathsf { a b s } ( \\Delta _ { q u a t } ) + \\lambda _ { h o l d } \\cdot t _ { h o l d } + \\lambda _ { s u c c e s s } \\cdot \\mathbf { 1 } _ { s u c c e s s } + \\lambda _ { c t r l } \\| a \\| ^ { 2 } , \\phantom { x x x x x x x x x x x x x x } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 456, + 1385, + 516, + 1385, + 516, + 1415, + 456, + 1415 + ], + "score": 0.89, + "latex": "\\mathbf { 1 } _ { \\mathrm { p l a c e } }" + }, + { + "category_id": 13, + "poly": [ + 347, + 1384, + 401, + 1384, + 401, + 1415, + 347, + 1415 + ], + "score": 0.88, + "latex": "\\mathbf { 1 } _ { \\mathrm { p i c k } }" + }, + { + "category_id": 13, + "poly": [ + 506, + 952, + 567, + 952, + 567, + 982, + 506, + 982 + ], + "score": 0.87, + "latex": "\\mathbf { 1 } _ { \\mathrm { p l a c e } }" + }, + { + "category_id": 14, + "poly": [ + 312, + 1600, + 1395, + 1600, + 1395, + 1682, + 312, + 1682 + ], + "score": 0.87, + "latex": "\\begin{array} { r l } & { R ( s ) = \\lambda _ { v e l } \\cdot \\mathbf { a b s } ( \\Delta x _ { a n t } ) + \\lambda _ { b o x v e l } \\cdot \\mathbf { a b s } ( \\Delta x _ { b o x } ) + \\lambda _ { u p r i g h t } \\cdot \\mathbf { c o s } ( \\theta ) - \\lambda _ { h e i g h t } \\cdot \\mathbf { a b s } ( 0 . 6 - h ) + } \\\\ & { ~ \\lambda _ { g o a l } \\cdot \\mathbf { d i s t } ( p _ { g o a l } , p _ { b o x } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 375, + 1731, + 1013, + 1731, + 1013, + 1765, + 375, + 1765 + ], + "score": 0.86, + "latex": "\\lambda _ { v e l } = 5 0 , \\lambda _ { b o x v e l } = 2 0 , \\lambda _ { u p r i g h t } = 1 , \\lambda _ { h e i g h t } = 0 . 5" + }, + { + "category_id": 13, + "poly": [ + 373, + 919, + 1347, + 919, + 1347, + 953, + 373, + 953 + ], + "score": 0.82, + "latex": "\\lambda _ { r e a c h } = 1 0 , \\lambda _ { c o n t a c t } = 1 0 , \\lambda _ { p i c k } = \\lambda _ { p l a c e } = \\lambda _ { p l a c e } = 1 0 , \\lambda _ { s u c c e s s } = 5 0 , \\lambda _ { c t r l } = 0 ," + }, + { + "category_id": 13, + "poly": [ + 367, + 1351, + 1398, + 1351, + 1398, + 1385, + 367, + 1385 + ], + "score": 0.81, + "latex": "\\lambda _ { r e a c h } = 1 0 , \\lambda _ { p i c k } = 3 0 , \\lambda _ { m o v e } = 1 0 0 , \\lambda _ { p l a c e } = 1 0 0 , \\lambda _ { s u c c e s s } = 1 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 }" + }, + { + "category_id": 13, + "poly": [ + 394, + 952, + 447, + 952, + 447, + 983, + 394, + 983 + ], + "score": 0.79, + "latex": "\\mathbf { 1 } _ { \\mathrm { p i c k } }" + }, + { + "category_id": 14, + "poly": [ + 370, + 503, + 1388, + 503, + 1388, + 541, + 370, + 541 + ], + "score": 0.65, + "latex": "\\lambda _ { r e a c h i n g } = 1 0 0 , \\lambda _ { p o s } = 5 0 0 , \\lambda _ { q u a t } = 3 0 , \\lambda _ { h o l d } = 1 0 , \\lambda _ { s u c c e s s } = 1 0 0 0 , \\lambda _ { c t r l } = 1 \\times 1 0 ^ { - 4 } ." + }, + { + "category_id": 13, + "poly": [ + 374, + 920, + 518, + 920, + 518, + 950, + 374, + 950 + ], + "score": 0.34, + "latex": "\\lambda _ { r e a c h } = 1 0" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1876.0, + 678.0, + 1876.0, + 678.0, + 1918.0, + 293.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 109.0, + 295.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 914.0, + 372.0, + 914.0, + 372.0, + 957.0, + 291.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 914.0, + 1408.0, + 914.0, + 1408.0, + 957.0, + 1348.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 948.0, + 393.0, + 948.0, + 393.0, + 982.0, + 292.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 948.0, + 505.0, + 948.0, + 505.0, + 982.0, + 448.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 948.0, + 1407.0, + 948.0, + 1407.0, + 982.0, + 568.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 981.0, + 1404.0, + 981.0, + 1404.0, + 1014.0, + 295.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1010.0, + 612.0, + 1010.0, + 612.0, + 1043.0, + 296.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1456.0, + 1406.0, + 1456.0, + 1406.0, + 1495.0, + 292.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1486.0, + 1408.0, + 1486.0, + 1408.0, + 1526.0, + 295.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1518.0, + 692.0, + 1518.0, + 692.0, + 1557.0, + 295.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1344.0, + 366.0, + 1344.0, + 366.0, + 1390.0, + 289.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1344.0, + 1409.0, + 1344.0, + 1409.0, + 1390.0, + 1399.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1380.0, + 346.0, + 1380.0, + 346.0, + 1418.0, + 294.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1380.0, + 455.0, + 1380.0, + 455.0, + 1418.0, + 402.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1380.0, + 1411.0, + 1380.0, + 1411.0, + 1418.0, + 517.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1412.0, + 1150.0, + 1412.0, + 1150.0, + 1446.0, + 294.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 546.0, + 1408.0, + 546.0, + 1408.0, + 591.0, + 291.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 584.0, + 1404.0, + 584.0, + 1404.0, + 617.0, + 293.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 614.0, + 770.0, + 614.0, + 770.0, + 648.0, + 296.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 1408.0, + 227.0, + 1408.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1408.0, + 261.0, + 1408.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 292.0, + 583.0, + 292.0, + 583.0, + 323.0, + 297.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1051.0, + 1404.0, + 1051.0, + 1404.0, + 1096.0, + 291.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 1057.0, + 1087.0, + 1057.0, + 1121.0, + 294.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1726.0, + 374.0, + 1726.0, + 374.0, + 1770.0, + 293.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1726.0, + 1406.0, + 1726.0, + 1406.0, + 1770.0, + 1014.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1757.0, + 1182.0, + 1757.0, + 1182.0, + 1797.0, + 293.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 1757.0, + 1333.0, + 1757.0, + 1333.0, + 1797.0, + 1324.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1407.0, + 1970.0, + 1407.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 1306.0, + 2003.0, + 1306.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 498.0, + 369.0, + 498.0, + 369.0, + 546.0, + 292.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1389.0, + 498.0, + 1392.0, + 498.0, + 1392.0, + 546.0, + 1389.0, + 546.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1476, + 1404, + 1476, + 1404, + 1692, + 297, + 1692 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1705, + 1404, + 1705, + 1404, + 1923, + 297, + 1923 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 776, + 1403, + 776, + 1403, + 961, + 297, + 961 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1114, + 1403, + 1114, + 1403, + 1275, + 298, + 1275 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 977, + 1404, + 977, + 1404, + 1100, + 298, + 1100 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1368, + 1403, + 1368, + 1403, + 1460, + 299, + 1460 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 301, + 1936, + 1401, + 1936, + 1401, + 2029, + 301, + 2029 + ], + "score": 0.962 + }, + { + "category_id": 5, + "poly": [ + 500, + 295, + 1198, + 295, + 1198, + 609, + 500, + 609 + ], + "score": 0.955, + "html": "
ParametersValue
learning rate3e-4
gradient steps50
batch size256
discount factor0.99
target smoothing coefficient0.005
reward scale (SAC)1.0
experience buffer size (# episodes)1000
Tow1 for JACO, 5 for ANT
Nz (dimensionality of z)5
" + }, + { + "category_id": 0, + "poly": [ + 300, + 1312, + 600, + 1312, + 600, + 1343, + 300, + 1343 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 300, + 720, + 702, + 720, + 702, + 751, + 300, + 751 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.896 + }, + { + "category_id": 6, + "poly": [ + 703, + 631, + 997, + 631, + 997, + 665, + 703, + 665 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 0, + "poly": [ + 298, + 230, + 613, + 230, + 613, + 261, + 298, + 261 + ], + "score": 0.649 + }, + { + "category_id": 6, + "poly": [ + 298, + 230, + 613, + 230, + 613, + 261, + 298, + 261 + ], + "score": 0.297 + }, + { + "category_id": 13, + "poly": [ + 297, + 1207, + 491, + 1207, + 491, + 1248, + 297, + 1248 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { N } ( m ^ { i } + N _ { z } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1401, + 1291, + 1401, + 1291, + 1430, + 1236, + 1430 + ], + "score": 0.91, + "latex": "T _ { l o w }" + }, + { + "category_id": 13, + "poly": [ + 813, + 1738, + 916, + 1738, + 916, + 1768, + 813, + 1768 + ], + "score": 0.91, + "latex": "\\lambda _ { 2 } = 0 . 1" + }, + { + "category_id": 13, + "poly": [ + 1188, + 1937, + 1307, + 1937, + 1307, + 1967, + 1188, + 1967 + ], + "score": 0.91, + "latex": "\\lambda _ { 2 } = 0 . 0 1" + }, + { + "category_id": 13, + "poly": [ + 297, + 1738, + 415, + 1738, + 415, + 1767, + 297, + 1767 + ], + "score": 0.89, + "latex": "\\lambda _ { 2 } = 0 . 0 1" + }, + { + "category_id": 13, + "poly": [ + 788, + 1069, + 854, + 1069, + 854, + 1099, + 788, + 1099 + ], + "score": 0.88, + "latex": "\\pm 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 653, + 1999, + 684, + 1999, + 684, + 2028, + 653, + 2028 + ], + "score": 0.88, + "latex": "\\lambda _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 569, + 575, + 606, + 575, + 606, + 604, + 569, + 604 + ], + "score": 0.8, + "latex": "N _ { z }" + }, + { + "category_id": 13, + "poly": [ + 346, + 1177, + 375, + 1177, + 375, + 1204, + 346, + 1204 + ], + "score": 0.79, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 297, + 1889, + 498, + 1889, + 498, + 1923, + 297, + 1923 + ], + "score": 0.78, + "latex": "( 0 . 1 5 , - 0 . 2 , 0 . 8 6 )" + }, + { + "category_id": 13, + "poly": [ + 792, + 1147, + 821, + 1147, + 821, + 1174, + 792, + 1174 + ], + "score": 0.75, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 553, + 1890, + 751, + 1890, + 751, + 1922, + 553, + 1922 + ], + "score": 0.68, + "latex": "( 0 . 3 , - 0 . 0 2 , 0 . 8 6 )" + }, + { + "category_id": 13, + "poly": [ + 822, + 579, + 839, + 579, + 839, + 601, + 822, + 601 + ], + "score": 0.53, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 299, + 1769, + 496, + 1769, + 496, + 1800, + 299, + 1800 + ], + "score": 0.37, + "latex": "( 0 . 3 , - 0 . 0 3 , 0 . 8 6 )" + }, + { + "category_id": 13, + "poly": [ + 683, + 542, + 734, + 542, + 734, + 574, + 683, + 574 + ], + "score": 0.29, + "latex": "T _ { \\mathrm { l o w } }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1309.0, + 604.0, + 1309.0, + 604.0, + 1347.0, + 294.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 720.0, + 704.0, + 720.0, + 704.0, + 753.0, + 296.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 628.0, + 999.0, + 628.0, + 999.0, + 669.0, + 701.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 616.0, + 230.0, + 616.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 616.0, + 230.0, + 616.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1475.0, + 1404.0, + 1475.0, + 1404.0, + 1512.0, + 294.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1506.0, + 1409.0, + 1506.0, + 1409.0, + 1543.0, + 292.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1535.0, + 1405.0, + 1535.0, + 1405.0, + 1574.0, + 294.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1568.0, + 1405.0, + 1568.0, + 1405.0, + 1605.0, + 292.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1599.0, + 1406.0, + 1599.0, + 1406.0, + 1633.0, + 295.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1630.0, + 1406.0, + 1630.0, + 1406.0, + 1664.0, + 295.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1661.0, + 1038.0, + 1661.0, + 1038.0, + 1695.0, + 296.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1706.0, + 1405.0, + 1706.0, + 1405.0, + 1741.0, + 295.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 1739.0, + 812.0, + 1739.0, + 812.0, + 1771.0, + 416.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1739.0, + 1405.0, + 1739.0, + 1405.0, + 1771.0, + 917.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 298.0, + 1768.0, + 298.0, + 1803.0, + 294.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1768.0, + 1407.0, + 1768.0, + 1407.0, + 1803.0, + 497.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1799.0, + 1406.0, + 1799.0, + 1406.0, + 1834.0, + 294.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1829.0, + 1406.0, + 1829.0, + 1406.0, + 1864.0, + 295.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1859.0, + 1405.0, + 1859.0, + 1405.0, + 1893.0, + 294.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1889.0, + 552.0, + 1889.0, + 552.0, + 1926.0, + 499.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1889.0, + 908.0, + 1889.0, + 908.0, + 1926.0, + 752.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 776.0, + 1407.0, + 776.0, + 1407.0, + 813.0, + 294.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 809.0, + 1404.0, + 809.0, + 1404.0, + 841.0, + 295.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 838.0, + 1405.0, + 838.0, + 1405.0, + 874.0, + 295.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 867.0, + 1405.0, + 867.0, + 1405.0, + 904.0, + 292.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 899.0, + 1405.0, + 899.0, + 1405.0, + 935.0, + 295.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 930.0, + 770.0, + 930.0, + 770.0, + 964.0, + 293.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1114.0, + 1405.0, + 1114.0, + 1405.0, + 1152.0, + 294.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1143.0, + 791.0, + 1143.0, + 791.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1143.0, + 1404.0, + 1143.0, + 1404.0, + 1181.0, + 822.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1174.0, + 345.0, + 1174.0, + 345.0, + 1214.0, + 293.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1174.0, + 1405.0, + 1174.0, + 1405.0, + 1214.0, + 376.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1203.0, + 296.0, + 1203.0, + 296.0, + 1256.0, + 291.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1203.0, + 1410.0, + 1203.0, + 1410.0, + 1256.0, + 492.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1239.0, + 499.0, + 1239.0, + 499.0, + 1283.0, + 294.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 975.0, + 1408.0, + 975.0, + 1408.0, + 1013.0, + 295.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1007.0, + 1402.0, + 1007.0, + 1402.0, + 1040.0, + 294.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 1404.0, + 1039.0, + 1404.0, + 1075.0, + 294.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1069.0, + 787.0, + 1069.0, + 787.0, + 1101.0, + 294.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1069.0, + 864.0, + 1069.0, + 864.0, + 1101.0, + 855.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1367.0, + 1405.0, + 1367.0, + 1405.0, + 1404.0, + 294.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1400.0, + 1235.0, + 1400.0, + 1235.0, + 1433.0, + 294.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1400.0, + 1405.0, + 1400.0, + 1405.0, + 1433.0, + 1292.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1426.0, + 415.0, + 1426.0, + 415.0, + 1463.0, + 294.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1934.0, + 1187.0, + 1934.0, + 1187.0, + 1972.0, + 293.0, + 1972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1934.0, + 1405.0, + 1934.0, + 1405.0, + 1972.0, + 1308.0, + 1972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1963.0, + 1405.0, + 1963.0, + 1405.0, + 2003.0, + 292.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1993.0, + 652.0, + 1993.0, + 652.0, + 2032.0, + 295.0, + 2032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1993.0, + 696.0, + 1993.0, + 696.0, + 2032.0, + 685.0, + 2032.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/vlm/dev/1wVvweK3oIb/0.png b/vlm/dev/1wVvweK3oIb/0.png new file mode 100644 index 0000000000000000000000000000000000000000..866f3c01533cb7e16e1ffb52f2d96fe3601f85ba --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38f5797a06d3ccd1ed2205034f22e603064c0ec76d05f7c203f4d29b934de2c1 +size 508200 diff --git a/vlm/dev/1wVvweK3oIb/1.png b/vlm/dev/1wVvweK3oIb/1.png new file mode 100644 index 0000000000000000000000000000000000000000..ee19b9edd885b805b98b509c9d6641d76b40dbb6 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e711694524429876c987399a0192fc8f3dfbd0e5c50a7ffe99cac7253317111f +size 540861 diff --git a/vlm/dev/1wVvweK3oIb/10.png b/vlm/dev/1wVvweK3oIb/10.png new file mode 100644 index 0000000000000000000000000000000000000000..9f4df0190e8a26c60146dd0324b37d6dcf5188ee --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de4c3ee452357859ff435bed197325d5024e95c3e9ac073a1db8a6714c2e2ad7 +size 583925 diff --git a/vlm/dev/1wVvweK3oIb/11.png b/vlm/dev/1wVvweK3oIb/11.png new file mode 100644 index 0000000000000000000000000000000000000000..d408acbd987c110d5bb5d241bbe1cb713e45246c --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6e736e301336b5ab419efafcd91f04cca2d65fd5398a9e77ae4977a64cf4efe +size 525467 diff --git a/vlm/dev/1wVvweK3oIb/12.png b/vlm/dev/1wVvweK3oIb/12.png new file mode 100644 index 0000000000000000000000000000000000000000..82c6bfdbecb728282c953d87547cdef34e8841ea --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70f482dff3188431e90d54aba041c5cc1f726bcd1317f1e210bac570101eaf55 +size 541998 diff --git a/vlm/dev/1wVvweK3oIb/13.png b/vlm/dev/1wVvweK3oIb/13.png new file mode 100644 index 0000000000000000000000000000000000000000..306fafe03336d84584570a03cf32773a2ef764c4 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ceb1b54a560cae426f7158961363768a13ad055d0502046fe81f0d07020bc155 +size 510970 diff --git a/vlm/dev/1wVvweK3oIb/14.png b/vlm/dev/1wVvweK3oIb/14.png new file mode 100644 index 0000000000000000000000000000000000000000..def784d8d26debf4710cafcef387d6a2957aa773 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66c6620788f632937e3682c17eaf12990aba7efcd12cfec08f199bb14a2635eb +size 358213 diff --git a/vlm/dev/1wVvweK3oIb/15.png b/vlm/dev/1wVvweK3oIb/15.png new file mode 100644 index 0000000000000000000000000000000000000000..2cb8fdedc409d069a2e97e664cded1073a736038 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:808db1d8cd9e04b26694d3f02104bf3ee27cd2c4e30548786969ee5b2be471aa +size 395542 diff --git a/vlm/dev/1wVvweK3oIb/16.png b/vlm/dev/1wVvweK3oIb/16.png new file mode 100644 index 0000000000000000000000000000000000000000..c6e8802e2b275d45adb3458491a4de49caaa0ca3 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54ab28836a64cfb6d7048aeac1799a064fb7ba920d9c7dcbb6a77128034192ee +size 248690 diff --git a/vlm/dev/1wVvweK3oIb/17.png b/vlm/dev/1wVvweK3oIb/17.png new file mode 100644 index 0000000000000000000000000000000000000000..199f57b3c7b44c0b1322b2e2283a02441dfdebc5 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6de4927864f82a34d4ff5074cba193532057c2e1ff1b57b47d7cf2d9b16bbc5a +size 348532 diff --git a/vlm/dev/1wVvweK3oIb/18.png b/vlm/dev/1wVvweK3oIb/18.png new file mode 100644 index 0000000000000000000000000000000000000000..f51c53562a5fde9419f7ff177ee64bc91c18a65b --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:463fca7a5b143854bd751957260849ac866cc2f5899a351326b9555e23266cb4 +size 350071 diff --git a/vlm/dev/1wVvweK3oIb/19.png b/vlm/dev/1wVvweK3oIb/19.png new file mode 100644 index 0000000000000000000000000000000000000000..d2f212672ca26abb26f82454fe2180212b3d34ca --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:930ae6ebd46e6b4e2e06fdeacd95c5db89c06d638c42fc1a11c89053d81ed0a4 +size 379778 diff --git a/vlm/dev/1wVvweK3oIb/2.png b/vlm/dev/1wVvweK3oIb/2.png new file mode 100644 index 0000000000000000000000000000000000000000..c0d17f744396c71e0b644904ae39dfbcaf64ed40 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d791a95af9e913589fba08cf893686be24f44d3dafc8e98d02dd76ffca19a48 +size 585438 diff --git a/vlm/dev/1wVvweK3oIb/20.png b/vlm/dev/1wVvweK3oIb/20.png new file mode 100644 index 0000000000000000000000000000000000000000..5d7e24cbcee32c816bc18abb965afc976a36dfe8 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8678fa3823e5e01d93e2278f379bac1074b940c831984cf1ca6a208bb776253f +size 313819 diff --git a/vlm/dev/1wVvweK3oIb/21.png b/vlm/dev/1wVvweK3oIb/21.png new file mode 100644 index 0000000000000000000000000000000000000000..77962664e1cce3ef089a1f21a70ea7defce077ea --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b609e1cfa0e96f9f658e69839df122a00f382d205a6a2ea1b1989c9aa201a85 +size 318276 diff --git a/vlm/dev/1wVvweK3oIb/22.png b/vlm/dev/1wVvweK3oIb/22.png new file mode 100644 index 0000000000000000000000000000000000000000..7f44298b227c3b9fe6c9d820fc963f475e94782e --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0747f69bfaa20666a566dc94c35e09469f75f96a88a36ea16bdfaf712b207a7c +size 137409 diff --git a/vlm/dev/1wVvweK3oIb/3.png b/vlm/dev/1wVvweK3oIb/3.png new file mode 100644 index 0000000000000000000000000000000000000000..9cab5eda484927cd2b5c149871d4c8c86c29288b --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84064cfa521ff95a4906d103891e199ae8ce18b92fb93f1ca0cce327c6d870d9 +size 615528 diff --git a/vlm/dev/1wVvweK3oIb/4.png b/vlm/dev/1wVvweK3oIb/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ed12dc2bd9348d7eea0c16006de923971e0236ff --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e06cbd63d449bb27176a4b1e95572e3b015936d1faf7988a7baf5808e9bd7356 +size 509654 diff --git a/vlm/dev/1wVvweK3oIb/5.png b/vlm/dev/1wVvweK3oIb/5.png new file mode 100644 index 0000000000000000000000000000000000000000..88cc5cd02db4c62fa5d6869fc5f09e5a3c4a9898 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbe32067a669349f047f617434583fbab42100bd216e78b97a4bf155df58cde1 +size 468420 diff --git a/vlm/dev/1wVvweK3oIb/6.png b/vlm/dev/1wVvweK3oIb/6.png new file mode 100644 index 0000000000000000000000000000000000000000..102155c3338b4116fcaeebfa153f7f5757859be2 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beb738436436c6a82af6ebd6fc6f01cf9c65eba283d9385dc53944a87f853535 +size 460172 diff --git a/vlm/dev/1wVvweK3oIb/7.png b/vlm/dev/1wVvweK3oIb/7.png new file mode 100644 index 0000000000000000000000000000000000000000..809e6652485feea9f36628d0ca4e6c1bc0daea29 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d15044f6ff6903cfe4acdb52646373dd1354524b0bb420c0c4a3bd20bb05eb10 +size 464717 diff --git a/vlm/dev/1wVvweK3oIb/8.png b/vlm/dev/1wVvweK3oIb/8.png new file mode 100644 index 0000000000000000000000000000000000000000..9e284d3c41fe8999ef4c60e272b22b7d36acedbc --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08a447bb7b2852ec87b91209d065b6d6a523d622570c5bd1bfca1b4d46830a8d +size 481527 diff --git a/vlm/dev/1wVvweK3oIb/9.png b/vlm/dev/1wVvweK3oIb/9.png new file mode 100644 index 0000000000000000000000000000000000000000..6bb0ab71b97ff0883483bd83da2a190e634d2861 --- /dev/null +++ b/vlm/dev/1wVvweK3oIb/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8116bf5c9531665fd26ef326931f10dcc57c2b51bf11af189ae0147f4e21ecf +size 509065 diff --git a/vlm/dev/716PvHoDct/12.png b/vlm/dev/716PvHoDct/12.png new file mode 100644 index 0000000000000000000000000000000000000000..7ae3f2a0a7726fa3c00d94362b8d6cc9590be942 --- /dev/null +++ b/vlm/dev/716PvHoDct/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5281ad8edb3336ac13339292760ced0d1fe37635dac87418c7f351bac2c49fd +size 579459 diff --git a/vlm/dev/716PvHoDct/13.png b/vlm/dev/716PvHoDct/13.png new file mode 100644 index 0000000000000000000000000000000000000000..89b67f50d624e5d5b4d160fa2ecaadb3aba06a8b --- /dev/null +++ b/vlm/dev/716PvHoDct/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a277874debe7fbe14330e7902dedc744a1ecc71e0dea92c390f3367f9a534000 +size 227879 diff --git a/vlm/dev/716PvHoDct/16.png b/vlm/dev/716PvHoDct/16.png new file mode 100644 index 0000000000000000000000000000000000000000..2ba693b2d543e372c5d2fa3dfd16a54bf36d49af --- /dev/null +++ b/vlm/dev/716PvHoDct/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d098c7caba864becca785ac4b224791380bf2756c12e1e40fe2c092c8f2e0254 +size 355806 diff --git a/vlm/dev/716PvHoDct/18.png b/vlm/dev/716PvHoDct/18.png new file mode 100644 index 0000000000000000000000000000000000000000..8d3dda531003ffb308a2687680022054338568e8 --- /dev/null +++ b/vlm/dev/716PvHoDct/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6eb0441b554e939ef4ddd7cd7890825a5fc10a36f82cc75dbeb3643ace49afd1 +size 431187 diff --git a/vlm/dev/716PvHoDct/19.png b/vlm/dev/716PvHoDct/19.png new file mode 100644 index 0000000000000000000000000000000000000000..5c3a545897327125cb250003933f2360ba5e9881 --- /dev/null +++ b/vlm/dev/716PvHoDct/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b819e307504240ec028e323cf4d685ae712162420174efeb8be6f62230b1b65b +size 351277 diff --git a/vlm/dev/716PvHoDct/20.png b/vlm/dev/716PvHoDct/20.png new file mode 100644 index 0000000000000000000000000000000000000000..6099e04317d3804b68a88ea1f8dd5a01acc78c56 --- /dev/null +++ b/vlm/dev/716PvHoDct/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be320a05f462bfc225cd41630a048d191dd6aafe33b017a9765952e57ca6896b +size 606033 diff --git a/vlm/dev/716PvHoDct/4.png b/vlm/dev/716PvHoDct/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ed5a0fc82db16e2a2111167afcf55b17727d3354 --- /dev/null +++ b/vlm/dev/716PvHoDct/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8811ac8e63518912cc94d8e6e506e60b6488bbf9247fac61c3fdc3aac0cbb393 +size 615801 diff --git a/vlm/dev/716PvHoDct/5.png b/vlm/dev/716PvHoDct/5.png new file mode 100644 index 0000000000000000000000000000000000000000..5e33ef184c14681acc3feed87a044c8f84faccba --- /dev/null +++ b/vlm/dev/716PvHoDct/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2d092c0a9e7bd9c2ddc0b3486e6b4208e416b725fa964a5e3b6b200d25ffe35 +size 486457 diff --git a/vlm/dev/716PvHoDct/7.png b/vlm/dev/716PvHoDct/7.png new file mode 100644 index 0000000000000000000000000000000000000000..c2efa9f02a6a0b638c7660d3e07b0f6a58dc4642 --- /dev/null +++ b/vlm/dev/716PvHoDct/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb895bc85a28fb07d4f69623baf2f1ee102b8d26a50e0adda81539a55049b828 +size 598812 diff --git a/vlm/dev/716PvHoDct/8.png b/vlm/dev/716PvHoDct/8.png new file mode 100644 index 0000000000000000000000000000000000000000..acf2e4cab015510e18334fc08f8e0777aada44c7 --- /dev/null +++ b/vlm/dev/716PvHoDct/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37094b50871aea2c7d7a24740df4e7e149ead3848949b9dd152da5e9a3ff3f60 +size 521603 diff --git a/vlm/dev/7UmjRGzp-A/0.png b/vlm/dev/7UmjRGzp-A/0.png new file mode 100644 index 0000000000000000000000000000000000000000..2ae54f9bf3fa34f45fa62f92a00101f8524cf983 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a30dda62395b4c8fb0f7c84cfb060baea8075eef6a2424191a4b7338f691642d +size 619582 diff --git a/vlm/dev/7UmjRGzp-A/1.png b/vlm/dev/7UmjRGzp-A/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0439e7277bd969caf50c10b584a901d046b80f55 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbab217f67ba81dd149002fa22742b8b3281daae4909bf34f028e08e248121ce +size 574910 diff --git a/vlm/dev/7UmjRGzp-A/10.png b/vlm/dev/7UmjRGzp-A/10.png new file mode 100644 index 0000000000000000000000000000000000000000..b602649fb1f9526af764531c55c48b2a7a69d17f --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c97abf6a03cc3353558b77f723051f832653c2a0d31e542d4b381cd11637cd77 +size 542436 diff --git a/vlm/dev/7UmjRGzp-A/11.png b/vlm/dev/7UmjRGzp-A/11.png new file mode 100644 index 0000000000000000000000000000000000000000..d8002a3abff4a79520dc7b32c5a68967a3f6f1ca --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:689207abe074f0d5086ddf9124a6e11da4e84ce74917586fbc5bba49f2661017 +size 558159 diff --git a/vlm/dev/7UmjRGzp-A/12.png b/vlm/dev/7UmjRGzp-A/12.png new file mode 100644 index 0000000000000000000000000000000000000000..bc5e0981b3c681213b0fc531b3d7a5e11ff529df --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a23caaa3e4981c0a964468781b64cc24d508848266f2f85136b6cb09a188ae7f +size 148870 diff --git a/vlm/dev/7UmjRGzp-A/13.png b/vlm/dev/7UmjRGzp-A/13.png new file mode 100644 index 0000000000000000000000000000000000000000..37aa928bf9f63f33e98a28682fc8e19f9451357f --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dc8ed7541add60149bc4b164da84105494c246998bde58ce264e77fb5d6eee0 +size 378926 diff --git a/vlm/dev/7UmjRGzp-A/14.png b/vlm/dev/7UmjRGzp-A/14.png new file mode 100644 index 0000000000000000000000000000000000000000..d52fb26c3e39108fbc59e127ff67e0cc9cb76b6e --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e457bdd4e5c6d68add33f37d99d070071e1f52f3a7095007030e80634622978 +size 466381 diff --git a/vlm/dev/7UmjRGzp-A/15.png b/vlm/dev/7UmjRGzp-A/15.png new file mode 100644 index 0000000000000000000000000000000000000000..5992b92bdac76dad46ed7ab4476618c5feb01968 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b31251c637a81edbc6e705a4d0bc393b4e3a5323cfb1af20935d85e24a31c569 +size 392582 diff --git a/vlm/dev/7UmjRGzp-A/16.png b/vlm/dev/7UmjRGzp-A/16.png new file mode 100644 index 0000000000000000000000000000000000000000..42c1a81681e179564c0688e388659597770dc2ff --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a49d38a747aa7b33a856deb81943c2eefe1d024459bcbbbfd8e8e1c9f51963fb +size 348915 diff --git a/vlm/dev/7UmjRGzp-A/17.png b/vlm/dev/7UmjRGzp-A/17.png new file mode 100644 index 0000000000000000000000000000000000000000..f506bcbcf7af2a31c4b2a2cb457ce1cf31f1af6f --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2476e08016220cb9d18dd770e149941948379ae4f422aeaafc99fc0ba0e4a553 +size 409818 diff --git a/vlm/dev/7UmjRGzp-A/18.png b/vlm/dev/7UmjRGzp-A/18.png new file mode 100644 index 0000000000000000000000000000000000000000..915b394a0abbbd1b266f5bb166db7b757f6fa9c9 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76d2a63aab783edda5b50d99cc31148136e5fbaa6a14e0dea1e11d75e2482d7d +size 398240 diff --git a/vlm/dev/7UmjRGzp-A/19.png b/vlm/dev/7UmjRGzp-A/19.png new file mode 100644 index 0000000000000000000000000000000000000000..9b72db87edce2031c30256e4249540a3dd19c8c7 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b681dec7414619486e33532c300027e73b5e5e3f80dcb08fef75441dc93c680 +size 317397 diff --git a/vlm/dev/7UmjRGzp-A/2.png b/vlm/dev/7UmjRGzp-A/2.png new file mode 100644 index 0000000000000000000000000000000000000000..255fe73966f90e9ba9ddf5c0d3b35a61d89980c3 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40e71f42a53f23670d7a1819d18df5eeb98675f6899c75ec57e395a35de5f21c +size 589870 diff --git a/vlm/dev/7UmjRGzp-A/20.png b/vlm/dev/7UmjRGzp-A/20.png new file mode 100644 index 0000000000000000000000000000000000000000..6803d1aefb68fc5ce4998d2710b2f6491be21487 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0809b4b66c7a13d521d39b93b78683cabb3444230f24f86b4a8bc7c8a4e31461 +size 336410 diff --git a/vlm/dev/7UmjRGzp-A/21.png b/vlm/dev/7UmjRGzp-A/21.png new file mode 100644 index 0000000000000000000000000000000000000000..6777f39e48b7f43dec07ac074350173a4eae2911 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:315b29f6f31686e5b961282b1949bfa435a960e7cce22928367a330a5d6bfae8 +size 320299 diff --git a/vlm/dev/7UmjRGzp-A/22.png b/vlm/dev/7UmjRGzp-A/22.png new file mode 100644 index 0000000000000000000000000000000000000000..d990153c9e6b7445982a8e0e02869b1b161ade8e --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ac0057e43a061d99b5f12adbac09b5418d955229cf9c2a20c3cd5ff0e4fb2d4 +size 497538 diff --git a/vlm/dev/7UmjRGzp-A/23.png b/vlm/dev/7UmjRGzp-A/23.png new file mode 100644 index 0000000000000000000000000000000000000000..23b448f1555d789809d08a087931d3f4aaac6c5b --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f52ccab1b57c3a5aa2a3e133bbd5f2009327266781ca7cbb386cb06bb4625989 +size 383995 diff --git a/vlm/dev/7UmjRGzp-A/24.png b/vlm/dev/7UmjRGzp-A/24.png new file mode 100644 index 0000000000000000000000000000000000000000..0d35289a3a0000e58b9d11baa7a86202056580e7 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf904f84c75d4c958136b71f3976e8ea5503821f631f53176a0d5953f9cbd88d +size 360750 diff --git a/vlm/dev/7UmjRGzp-A/25.png b/vlm/dev/7UmjRGzp-A/25.png new file mode 100644 index 0000000000000000000000000000000000000000..6106591a2d362cc917c3e0f9ef8d9dc9020e92f7 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e37030dfb94551664c091683b097d4516cd7610d4e4b57020589e7ae02073a2 +size 364495 diff --git a/vlm/dev/7UmjRGzp-A/26.png b/vlm/dev/7UmjRGzp-A/26.png new file mode 100644 index 0000000000000000000000000000000000000000..eebcb45a9d461112d454097d366583603971270f --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02afa823210bdcb6dd46c86050c7f07dabfe3f4f42465cfad183d25306e2a551 +size 292860 diff --git a/vlm/dev/7UmjRGzp-A/27.png b/vlm/dev/7UmjRGzp-A/27.png new file mode 100644 index 0000000000000000000000000000000000000000..0b19a56baeba845ca5ad35db0ca3d01ebdaee824 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:214c0f4b2707b59033f452e448120d307d7961ad34b6a704f0e23981b2f38287 +size 524456 diff --git a/vlm/dev/7UmjRGzp-A/28.png b/vlm/dev/7UmjRGzp-A/28.png new file mode 100644 index 0000000000000000000000000000000000000000..06bbd93f723a6bed8c86a0096c12f3e1afbfde49 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f20e9710abfcc7643010f620fe538342c0694529b89255846bb09a54ecd464fe +size 293294 diff --git a/vlm/dev/7UmjRGzp-A/29.png b/vlm/dev/7UmjRGzp-A/29.png new file mode 100644 index 0000000000000000000000000000000000000000..266ffd49900b0c26a33efdbb1745e41ed5873bf4 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65f010fa2e9626ca9418206782da8fa27372f4e3125f8e8e02a6c67461ead9c3 +size 254120 diff --git a/vlm/dev/7UmjRGzp-A/3.png b/vlm/dev/7UmjRGzp-A/3.png new file mode 100644 index 0000000000000000000000000000000000000000..5e62182d959007a1025c7eba928a63b056939fc7 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:972653a5515212c349c8ea780137e591dee76e728b2a40770a32714cc4b9fc15 +size 606319 diff --git a/vlm/dev/7UmjRGzp-A/4.png b/vlm/dev/7UmjRGzp-A/4.png new file mode 100644 index 0000000000000000000000000000000000000000..83dd21d26a880a5b17fec0a80504d33711d5d64a --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50824802c4bda320f905d139505be9707c892c2537ba0c420acd6de468601cac +size 571517 diff --git a/vlm/dev/7UmjRGzp-A/5.png b/vlm/dev/7UmjRGzp-A/5.png new file mode 100644 index 0000000000000000000000000000000000000000..553047b9b5cb1cb0955c79c59462d70bcb77e7a5 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bd53883057c8dea037e25328ff03d6fa35d7889634cd06208b1af6862a26355 +size 590421 diff --git a/vlm/dev/7UmjRGzp-A/6.png b/vlm/dev/7UmjRGzp-A/6.png new file mode 100644 index 0000000000000000000000000000000000000000..ae5a39c72a1fdcfbfd0a9307d14a641dce8daeab --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af20aa8c0d46af84826f7a03b6d202695044e3333ce46a0d52c6fe7e71f024ab +size 596081 diff --git a/vlm/dev/7UmjRGzp-A/7.png b/vlm/dev/7UmjRGzp-A/7.png new file mode 100644 index 0000000000000000000000000000000000000000..064b124746d750b541dc75d0eabad6fa8a2f0834 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:996564205ff63670bf61dff1d7e7f7f3f5eaddd170e9e591381e1f6ce59f8ea1 +size 606557 diff --git a/vlm/dev/7UmjRGzp-A/8.png b/vlm/dev/7UmjRGzp-A/8.png new file mode 100644 index 0000000000000000000000000000000000000000..3e23eab754fa6739cb0d2fef7288e9ee52be3be8 --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74977e4f4a99e92be4feaa8fc6598e3422e8dbea4e80d5385e9e11c3b5981b9b +size 584017 diff --git a/vlm/dev/7UmjRGzp-A/9.png b/vlm/dev/7UmjRGzp-A/9.png new file mode 100644 index 0000000000000000000000000000000000000000..85af7ee3ab1b24cca0b439f333d3cee5ed62326f --- /dev/null +++ b/vlm/dev/7UmjRGzp-A/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db3835998c997cbab07e0e5291f689245a5fb247aab6d61bafef12dcc1730ecd +size 536089 diff --git a/vlm/dev/7anW5TWbCJ/0.png b/vlm/dev/7anW5TWbCJ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..bcc4f4a6ef0490195871df595364f020726580d6 --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:058c2844e7a0d50d2c27cbb50385cc44005315d002171ad5bfe8abd99cdb79e4 +size 394783 diff --git a/vlm/dev/7anW5TWbCJ/1.png b/vlm/dev/7anW5TWbCJ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..032c3d34160ea2cc61dfc11140989002278460bf --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28ac4210c9d9ccad7fe0e8d669554f6138a906ecb1e3c45a0d47b5a02541fe28 +size 592012 diff --git a/vlm/dev/7anW5TWbCJ/10.png b/vlm/dev/7anW5TWbCJ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..01607bd1a299f86d92300822868f331b91b2016d --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6425b020331de2fb8f3f0a058ac6ac10f9af2d8da6b8d518c856a14ed4b1d755 +size 539902 diff --git a/vlm/dev/7anW5TWbCJ/11.png b/vlm/dev/7anW5TWbCJ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5931065170a2f47dfeb6f170ef752bcd01684a27 --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa7051a0e8d0f5fac73696dc2aa9f0520fedb7ac1007d619bee898f8e85c08c5 +size 70967 diff --git a/vlm/dev/7anW5TWbCJ/2.png b/vlm/dev/7anW5TWbCJ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..7beea4b77d1ae2eccc751e0d06b590623b661a54 --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc1aefdf655e6a083db86798406ccc77b7c48ad88f6d93429f068535738e525b +size 514933 diff --git a/vlm/dev/7anW5TWbCJ/3.png b/vlm/dev/7anW5TWbCJ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..921499a11edda39de986a8115a9e66cfe047751a --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa931c27b0144a6f3ab4cbe4a76e2f75a229929596b99e9a20e0c2c9eeb090e5 +size 486249 diff --git a/vlm/dev/7anW5TWbCJ/4.png b/vlm/dev/7anW5TWbCJ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..a12d04d68dbb6d5e6168ac51ef9101ec789f1422 --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db58b6e62c10281cb121cf848277b17f2c00ae4b1200568490ff6a76e1274c0d +size 479109 diff --git a/vlm/dev/7anW5TWbCJ/5.png b/vlm/dev/7anW5TWbCJ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..40d9bfd403fe794fe2c074aaf30f8477a089698d --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f417abf4cadbef33cbcbe4b4d83f7ccb1e45fdc9b530836ea152a4102258f11f +size 508732 diff --git a/vlm/dev/7anW5TWbCJ/6.png b/vlm/dev/7anW5TWbCJ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..0e54f0a15d1ecca12f234c8c26598261861b08dd --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38f0fde2905ef1afe28d5531ad6b5e6f19a06bc369552de8e75f503cd6d86f87 +size 500848 diff --git a/vlm/dev/7anW5TWbCJ/7.png b/vlm/dev/7anW5TWbCJ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..6a570e475626a97839e9fcbd077080dc0bfe6d5a --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a4d3512da70b99ac997070e98ec370f0d29540c19bbcd2ce6d80a1403a166c9 +size 448376 diff --git a/vlm/dev/7anW5TWbCJ/8.png b/vlm/dev/7anW5TWbCJ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..002114f5cc7d325a0420469ca544786d8317a468 --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8c2707e7ae3ac8a99c0c9ae9292b1828fa85ee7d5901e3695f0d12956896412 +size 438716 diff --git a/vlm/dev/7anW5TWbCJ/9.png b/vlm/dev/7anW5TWbCJ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..4bcf5f1f86adeda37326fa761d69f283783d36f9 --- /dev/null +++ b/vlm/dev/7anW5TWbCJ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c6018cec30a81f7205c917f9eda0e1db2c83b4b15e89296408f1a1db33edc25 +size 547580 diff --git a/vlm/dev/B9LUI0pZFGc/0.png b/vlm/dev/B9LUI0pZFGc/0.png new file mode 100644 index 0000000000000000000000000000000000000000..16c47c8c02491108d5d3b7fa557bad6eab9e4b77 --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81d312c0dad2ee249f6ccf76dc9c6a9c2ea417eb337f7a197469116f89755d60 +size 544010 diff --git a/vlm/dev/B9LUI0pZFGc/1.png b/vlm/dev/B9LUI0pZFGc/1.png new file mode 100644 index 0000000000000000000000000000000000000000..51256ae2773bd5f1f7308af8aa7d73ca4642f77c --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:505084e3f50af6d2bdf4ad7e92a888fa07bfc909b2d1f389e2e675c7a31f8763 +size 668900 diff --git a/vlm/dev/B9LUI0pZFGc/10.png b/vlm/dev/B9LUI0pZFGc/10.png new file mode 100644 index 0000000000000000000000000000000000000000..2d26ef1ec31854645a2cf8c9124f3051408b6b13 --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b75c578a2bd73a18477d93ec472e123a025921239d50fd661329bb90fdeea5e5 +size 559659 diff --git a/vlm/dev/B9LUI0pZFGc/11.png b/vlm/dev/B9LUI0pZFGc/11.png new file mode 100644 index 0000000000000000000000000000000000000000..95c861c885355b85c6e1f34a5b05736241b86f81 --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd1ee94bda5b457f08c748fd61b58d9c1eb728aaad7a9f6e2dceed3e11f5e599 +size 529582 diff --git a/vlm/dev/B9LUI0pZFGc/12.png b/vlm/dev/B9LUI0pZFGc/12.png new file mode 100644 index 0000000000000000000000000000000000000000..5985108905b512c9e1a971993491b099b6420a90 --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79f26e14bb1c1dd1b1b44b12da2ea2e6035bd4b6b60089bedb0cc3e11aed3a4d +size 376219 diff --git a/vlm/dev/B9LUI0pZFGc/13.png b/vlm/dev/B9LUI0pZFGc/13.png new file mode 100644 index 0000000000000000000000000000000000000000..79d6a9f8ed43ffcb3c84c984a0326fd73a13d65a --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ec4a138a0188b3e286b081aa2326548a734685493816cc787492507eddc4925 +size 166010 diff --git a/vlm/dev/B9LUI0pZFGc/2.png b/vlm/dev/B9LUI0pZFGc/2.png new file mode 100644 index 0000000000000000000000000000000000000000..70b510f7868d3571e357faa435787e337ba8e1b0 --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e417071f7ab6ae8ddd0383b4d4b0b654f31e31a85a4e50f393f6ad14ccb70a23 +size 467833 diff --git a/vlm/dev/B9LUI0pZFGc/3.png b/vlm/dev/B9LUI0pZFGc/3.png new file mode 100644 index 0000000000000000000000000000000000000000..3bc8b25738bca6c395f2fcb11c280e1fc560fc1b --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73a04201e5c98dc70ef7e029c97f763e30c500b6950047bbcce1924efa037fec +size 509117 diff --git a/vlm/dev/B9LUI0pZFGc/4.png b/vlm/dev/B9LUI0pZFGc/4.png new file mode 100644 index 0000000000000000000000000000000000000000..8a1423fa918bc851f44b06cbe8e9414f908db8e7 --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf437bc3463521e13011bffbcfd915958c8b3f8820ab2b541c32e928eb6e90dd +size 445822 diff --git a/vlm/dev/B9LUI0pZFGc/5.png b/vlm/dev/B9LUI0pZFGc/5.png new file mode 100644 index 0000000000000000000000000000000000000000..32b5cea6d63232604f58c3f7e5a2f801e56ec78d --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:581a64f769e75ece79a670cf3d8793463ca4e63409e70d5eb711783b48f0eafd +size 651043 diff --git a/vlm/dev/B9LUI0pZFGc/6.png b/vlm/dev/B9LUI0pZFGc/6.png new file mode 100644 index 0000000000000000000000000000000000000000..0ea5e0f04bdc3bb731324cf316d884244edc078d --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:405fbf8512d213aef8d7be0eab59474ee4ab533f9d97fce742ad671ac01e247c +size 580977 diff --git a/vlm/dev/B9LUI0pZFGc/7.png b/vlm/dev/B9LUI0pZFGc/7.png new file mode 100644 index 0000000000000000000000000000000000000000..0af899fdcc4dfe6fc3332e6fd8a24a4a6230f6ca --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16c56638313fa780db4496d9e2b6ffb9f18718c5b24afebf174f35049d37f087 +size 767858 diff --git a/vlm/dev/B9LUI0pZFGc/8.png b/vlm/dev/B9LUI0pZFGc/8.png new file mode 100644 index 0000000000000000000000000000000000000000..5808b4a4af1e2b0695f621bab24c54a2c92611b7 --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a0a0352f49342634f6475b458cbf3cbccd760100e4a1a34ad7f6de01aa4d74d +size 895481 diff --git a/vlm/dev/B9LUI0pZFGc/9.png b/vlm/dev/B9LUI0pZFGc/9.png new file mode 100644 index 0000000000000000000000000000000000000000..28db9f5a4613cee90b7e928b959c093d62d63c13 --- /dev/null +++ b/vlm/dev/B9LUI0pZFGc/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf61203c28e0b1671cd300a450a461ce68a726dc834064ef483bd240edc99bb4 +size 558518 diff --git a/vlm/dev/BYLysbfdJOd/0.png b/vlm/dev/BYLysbfdJOd/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c623a66546c2bb83be2f85813949d6e4b1789da0 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2824bd2a0ba2a7c303afd135a1e03c8d4fe75f05b4fe43f2c25c6d13f355977 +size 434709 diff --git a/vlm/dev/BYLysbfdJOd/1.png b/vlm/dev/BYLysbfdJOd/1.png new file mode 100644 index 0000000000000000000000000000000000000000..99127d1a34dc1951d907c7210f4c92f6968e3f59 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f51686bb5a41cebffb2a2bc059404d99bcc0c6ea3f17f1b967a58082ea60269 +size 948265 diff --git a/vlm/dev/BYLysbfdJOd/10.png b/vlm/dev/BYLysbfdJOd/10.png new file mode 100644 index 0000000000000000000000000000000000000000..b752189b20c3768b529e229f7b8c69af8b097b8e --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3028685440824823745290d078714c211dc1231a44d9e731469056d4f635391 +size 513761 diff --git a/vlm/dev/BYLysbfdJOd/11.png b/vlm/dev/BYLysbfdJOd/11.png new file mode 100644 index 0000000000000000000000000000000000000000..358c527cf23d006cf7590401213f8d249c01bb50 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3885d5bd5305cb8ee88c0bf4437655b8cdeb9e61302c66c05590356ebd5dcf9e +size 259662 diff --git a/vlm/dev/BYLysbfdJOd/2.png b/vlm/dev/BYLysbfdJOd/2.png new file mode 100644 index 0000000000000000000000000000000000000000..4a2cfea4b5b9b580b653db250f14f6581f217051 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e940bc7119a5567b117888dfc641e803db3fa6253317df969074c5754e947b3 +size 534132 diff --git a/vlm/dev/BYLysbfdJOd/3.png b/vlm/dev/BYLysbfdJOd/3.png new file mode 100644 index 0000000000000000000000000000000000000000..337de08f9e49b49eea6e5a3bbf1522f3fecf40ac --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:226a2f89d432e38cbac44a086389066bf68a2ba458c886e50927aba4a4a39773 +size 513280 diff --git a/vlm/dev/BYLysbfdJOd/4.png b/vlm/dev/BYLysbfdJOd/4.png new file mode 100644 index 0000000000000000000000000000000000000000..63c435cbb6ced3ef78badd5258d92924a4d35005 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:259945867cad8e719c0dd989532966f124c100ccd0fa41fa0a3dbe79dde4948d +size 558909 diff --git a/vlm/dev/BYLysbfdJOd/5.png b/vlm/dev/BYLysbfdJOd/5.png new file mode 100644 index 0000000000000000000000000000000000000000..d401a2f414345148302bbdea9cba0c33768251e1 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f539b691262cf39563839d3f1ed659aade8c10605550ba223026ef5a5da8aa45 +size 510360 diff --git a/vlm/dev/BYLysbfdJOd/6.png b/vlm/dev/BYLysbfdJOd/6.png new file mode 100644 index 0000000000000000000000000000000000000000..647af8eb0b6d0917e04c1a23f83a0e252b89d019 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d9de5eff4f910c7f0177b1d3d8a4f15c6bada4d25e8bf276a3b9c1056735e25 +size 420742 diff --git a/vlm/dev/BYLysbfdJOd/7.png b/vlm/dev/BYLysbfdJOd/7.png new file mode 100644 index 0000000000000000000000000000000000000000..2e53b6be79751dba09175a420e6b06a5993a58ab --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:529884e8035ce07859601389d8c5d90ead6cd0a5c8d9866c02ccc0a5f6b3f97b +size 623291 diff --git a/vlm/dev/BYLysbfdJOd/8.png b/vlm/dev/BYLysbfdJOd/8.png new file mode 100644 index 0000000000000000000000000000000000000000..78cd5d0a891d6668c96a43f4e266a447bc330e12 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:407edafc46d831d1d783be774862c330e11983dfae7051be8f2c8279d05bd5fc +size 493674 diff --git a/vlm/dev/BYLysbfdJOd/9.png b/vlm/dev/BYLysbfdJOd/9.png new file mode 100644 index 0000000000000000000000000000000000000000..b01cb8b9fcb84d033e05fd0284f25821fe528d60 --- /dev/null +++ b/vlm/dev/BYLysbfdJOd/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6d33bb10f61ef193d7c57120f65bb4f433422f6142948ba5d25a993de4604b2 +size 594484 diff --git a/vlm/dev/ETKGuby0hcs/10.png b/vlm/dev/ETKGuby0hcs/10.png new file mode 100644 index 0000000000000000000000000000000000000000..712aed7f6f8210a0a256c98d41b42b701077b350 --- /dev/null +++ b/vlm/dev/ETKGuby0hcs/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfcb11db1d6fb73499969bbb714478618f557e299a7b0a7ad284f4feaa97fd89 +size 519191 diff --git a/vlm/dev/ETKGuby0hcs/17.png b/vlm/dev/ETKGuby0hcs/17.png new file mode 100644 index 0000000000000000000000000000000000000000..0be75337494e620a53fa41bcaca1198f6861f752 --- /dev/null +++ b/vlm/dev/ETKGuby0hcs/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a75d26a2896066ffa1334ddc6034a2a004e1e7c38d2c2c20ef71a200c9ef1403 +size 129430 diff --git a/vlm/dev/ETKGuby0hcs/21.png b/vlm/dev/ETKGuby0hcs/21.png new file mode 100644 index 0000000000000000000000000000000000000000..142a95c9fa2cbc2467e2eccea273a88e5043cfe0 --- /dev/null +++ b/vlm/dev/ETKGuby0hcs/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f575ccd15a0e461ff004e2ed7f627e4c371a6e1f2484934157fd4c2f2a0d49f6 +size 389713 diff --git a/vlm/dev/ETKGuby0hcs/23.png b/vlm/dev/ETKGuby0hcs/23.png new file mode 100644 index 0000000000000000000000000000000000000000..3f478dade676315f2653dfc1695eb88120dc0c4e --- /dev/null +++ b/vlm/dev/ETKGuby0hcs/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b6fc8190920dbfe64c003ddae257bfa8a013aeda0a804e4a55b1c09d67af8b4 +size 92971 diff --git a/vlm/dev/ETKGuby0hcs/24.png b/vlm/dev/ETKGuby0hcs/24.png new file mode 100644 index 0000000000000000000000000000000000000000..59424717ace7618d244a708d80099cd7077a5bd0 --- /dev/null +++ b/vlm/dev/ETKGuby0hcs/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2656087665540a528d540dbf8007344f1857013c68c835b0f22e89b397ad913 +size 447429 diff --git a/vlm/dev/ETKGuby0hcs/3.png b/vlm/dev/ETKGuby0hcs/3.png new file mode 100644 index 0000000000000000000000000000000000000000..ff4e50b01264da909ac7a20c507ee7d3f11f2d6f --- /dev/null +++ b/vlm/dev/ETKGuby0hcs/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a24ee1eb6edc1c6d78f0b17e1998f62294eb4c10627910cbbf7e1511dfb75dc +size 585435 diff --git a/vlm/dev/ETKGuby0hcs/6.png b/vlm/dev/ETKGuby0hcs/6.png new file mode 100644 index 0000000000000000000000000000000000000000..b758d030542d6f0a0ad07c0a9785cfce6330a679 --- /dev/null +++ b/vlm/dev/ETKGuby0hcs/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b25c235b9de26be0d8c4e780fc792d37bdf61a2a93a123739b1a69a0ce1f0f44 +size 656805 diff --git a/vlm/dev/GoOuIrDHG_Y/0.png b/vlm/dev/GoOuIrDHG_Y/0.png new file mode 100644 index 0000000000000000000000000000000000000000..30fe576902c61337399f4eafa30e9251d7b36a4f --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec685f682c08430c8d067006ac0c2400dec70ec3c6506e703d42f3e2344bcab6 +size 423405 diff --git a/vlm/dev/GoOuIrDHG_Y/1.png b/vlm/dev/GoOuIrDHG_Y/1.png new file mode 100644 index 0000000000000000000000000000000000000000..39be7b536de9c3d93f08f48a182b8e2d5595a7f0 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48dc5be46667904267fe8a8c82e5ead1ec67d4a397bed0fbcdb43d21d4f21140 +size 546314 diff --git a/vlm/dev/GoOuIrDHG_Y/10.png b/vlm/dev/GoOuIrDHG_Y/10.png new file mode 100644 index 0000000000000000000000000000000000000000..df0d7d4e0deff26ff1c85de87e5de87214eae1f4 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2765c70d9b4586daba98b4a903a9fba593865b7990fa40573f9f99675c2fce7c +size 534264 diff --git a/vlm/dev/GoOuIrDHG_Y/11.png b/vlm/dev/GoOuIrDHG_Y/11.png new file mode 100644 index 0000000000000000000000000000000000000000..814a936b6f6af73d22491584e0f3ca21283730cf --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:769a3c4b0b9b70a9f0c3f21c9a18cbb9e46e526caa413c51fc17ad395c956ba3 +size 517089 diff --git a/vlm/dev/GoOuIrDHG_Y/12.png b/vlm/dev/GoOuIrDHG_Y/12.png new file mode 100644 index 0000000000000000000000000000000000000000..c75930135d38af18b39c26be12589f265c504b67 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af57b1989296a006fee62dcc9691ee07056f4ad1e8146e7c5c30b1c2a7e1f3e2 +size 413289 diff --git a/vlm/dev/GoOuIrDHG_Y/2.png b/vlm/dev/GoOuIrDHG_Y/2.png new file mode 100644 index 0000000000000000000000000000000000000000..2f09519e58f97fdf975c80fab10672f76a3dd2c2 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f94cb6db926f49a4f0e3dbcb6c9ccdd2eeaac77b9dbdada4514e4802fdd775ac +size 530885 diff --git a/vlm/dev/GoOuIrDHG_Y/3.png b/vlm/dev/GoOuIrDHG_Y/3.png new file mode 100644 index 0000000000000000000000000000000000000000..c87b6e2dc1493c2e9c8a78d26bdf7960acbba5c7 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:510c3f9993e20c0f605f89846e61b99d2502b890e2fccf55c58edc68afdff7a8 +size 454496 diff --git a/vlm/dev/GoOuIrDHG_Y/4.png b/vlm/dev/GoOuIrDHG_Y/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ce5a9114e3190efa1eca46a6f0f4931433eeee3b --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15c01280a2636c9884f7f2af5c1b6b6dab683109e0d81758d9ba6e660c0b0239 +size 585276 diff --git a/vlm/dev/GoOuIrDHG_Y/5.png b/vlm/dev/GoOuIrDHG_Y/5.png new file mode 100644 index 0000000000000000000000000000000000000000..5d2ccfa7360b5138f160c43123c1667822b3a734 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6977341e67a47272633c50c2774e09481dd20e3ecb25e5d292b5240782d668ab +size 517441 diff --git a/vlm/dev/GoOuIrDHG_Y/6.png b/vlm/dev/GoOuIrDHG_Y/6.png new file mode 100644 index 0000000000000000000000000000000000000000..ed14868d3ba0e7467aa43a5e9165f2886f88ae87 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9edb0cabe86beabaa7a2059eba5d5707f5da8dd43b59f8a1439a20e25a64ea14 +size 456039 diff --git a/vlm/dev/GoOuIrDHG_Y/7.png b/vlm/dev/GoOuIrDHG_Y/7.png new file mode 100644 index 0000000000000000000000000000000000000000..63073c3ad25fcf4fec02d0649e7577ba52b1e121 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5febbcaa0453f36f3c7a2327d9c94f4319baeaeaef74613ae90a74e668b93b6 +size 494722 diff --git a/vlm/dev/GoOuIrDHG_Y/8.png b/vlm/dev/GoOuIrDHG_Y/8.png new file mode 100644 index 0000000000000000000000000000000000000000..1f1aa91924caed081213c7803a8700d90c97cf83 --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d32475efd8fcc0fd3557fd5c104eadaae25a5422deb1e8ea63ab6812e1555e2a +size 466173 diff --git a/vlm/dev/GoOuIrDHG_Y/9.png b/vlm/dev/GoOuIrDHG_Y/9.png new file mode 100644 index 0000000000000000000000000000000000000000..753714e3fcf4f5c916adf34895b228ace6cf4d3f --- /dev/null +++ b/vlm/dev/GoOuIrDHG_Y/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ce856aac90ad6792a855ed4145956c43814a8ad3a4a35746ff01104576c54ec +size 101922 diff --git a/vlm/dev/I59qJ0sJ2nh/0.png b/vlm/dev/I59qJ0sJ2nh/0.png new file mode 100644 index 0000000000000000000000000000000000000000..67b47bee7609b638d13fede8943802190cf286f7 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2381ca4a507af3ff40095b46e439fc47accd16ba9b436e055e37e0fed9684250 +size 424971 diff --git a/vlm/dev/I59qJ0sJ2nh/1.png b/vlm/dev/I59qJ0sJ2nh/1.png new file mode 100644 index 0000000000000000000000000000000000000000..fba1268f9266c82dbad0f0a0a777d9790d521303 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3346cce9b2b40eefc54add8b36d27c437ce6209344ce4e56584f56530f448e7b +size 624189 diff --git a/vlm/dev/I59qJ0sJ2nh/10.png b/vlm/dev/I59qJ0sJ2nh/10.png new file mode 100644 index 0000000000000000000000000000000000000000..361094a8afb0601a61d0b8df1a86c67985ec79a0 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbe9a5c465b4addd7a5b8f080b99d42e2f570870e2a1443acd2ba4708ccb5d14 +size 555442 diff --git a/vlm/dev/I59qJ0sJ2nh/11.png b/vlm/dev/I59qJ0sJ2nh/11.png new file mode 100644 index 0000000000000000000000000000000000000000..237c3f860e83cf009c05ec14ae318f2253d64845 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:779472dc26c5823b64ea337e199bddb1d5c8349e4e4059a23229c9455ac829b7 +size 501840 diff --git a/vlm/dev/I59qJ0sJ2nh/12.png b/vlm/dev/I59qJ0sJ2nh/12.png new file mode 100644 index 0000000000000000000000000000000000000000..36f452c9e79249db0d5a8259d53bc1f05a6ef0be --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:141859e72d0cd87a39c98a97c5294237443addb660cf808082d96d0ada23dcbb +size 534286 diff --git a/vlm/dev/I59qJ0sJ2nh/13.png b/vlm/dev/I59qJ0sJ2nh/13.png new file mode 100644 index 0000000000000000000000000000000000000000..9e220afa7aebed5fdf68e66f03f8cd60629f63a1 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0079141237a7d1133b78daa29494d7cdfd3ff75594eac6cced8ac5f42dffefe +size 104681 diff --git a/vlm/dev/I59qJ0sJ2nh/2.png b/vlm/dev/I59qJ0sJ2nh/2.png new file mode 100644 index 0000000000000000000000000000000000000000..69b16ea4423c2c2012190c739f4987a9e78c89d4 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b505f61d472f702ca25f9ba5a5cd5da45aa6eddaf678f39a66ac7b04aad9dee3 +size 569137 diff --git a/vlm/dev/I59qJ0sJ2nh/3.png b/vlm/dev/I59qJ0sJ2nh/3.png new file mode 100644 index 0000000000000000000000000000000000000000..bf3c5bf302dfc30dec115757f452f8e71ee3c987 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73ed7e03e0e30886e2c8110ed2ba5b6eb9201af128483dac3afb9ae4b34939bb +size 542673 diff --git a/vlm/dev/I59qJ0sJ2nh/4.png b/vlm/dev/I59qJ0sJ2nh/4.png new file mode 100644 index 0000000000000000000000000000000000000000..25f1d30b4995b5a30e72af240cb6bfc32cb94f52 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f5505af0dc6447d3f724b4740af511e4ee5163b217c24b884f7cd3e00ff7604 +size 635372 diff --git a/vlm/dev/I59qJ0sJ2nh/5.png b/vlm/dev/I59qJ0sJ2nh/5.png new file mode 100644 index 0000000000000000000000000000000000000000..09261907e389740769b8795bc24682b6dad9ea3f --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de338232453492aa95eb1a1bfcbd7c81aa2f6cfd3f98f56f2a2a05a11964d80d +size 598366 diff --git a/vlm/dev/I59qJ0sJ2nh/6.png b/vlm/dev/I59qJ0sJ2nh/6.png new file mode 100644 index 0000000000000000000000000000000000000000..605a3df69cdc8190f076a37f4151bd7fca04141c --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d8f5dc04b129376c2d3c05614f204ac924ff0989e69a59d51084e345fb63435 +size 598423 diff --git a/vlm/dev/I59qJ0sJ2nh/7.png b/vlm/dev/I59qJ0sJ2nh/7.png new file mode 100644 index 0000000000000000000000000000000000000000..43a2e624875791419ff8930cb785ce9fce998aa2 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21f95306a13530c1134d22d4f8207def7031e12094590f045c8b5e63b0217bd9 +size 730014 diff --git a/vlm/dev/I59qJ0sJ2nh/8.png b/vlm/dev/I59qJ0sJ2nh/8.png new file mode 100644 index 0000000000000000000000000000000000000000..25e369f09393bd24f272d0a46cd396196b1192f1 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22afcd8439280513c8531ccf72fa98c5c227c3ce9ec5db2913ef65d523ff5657 +size 615911 diff --git a/vlm/dev/I59qJ0sJ2nh/9.png b/vlm/dev/I59qJ0sJ2nh/9.png new file mode 100644 index 0000000000000000000000000000000000000000..b1ab615bb28043e563b3c33d97a9fc32b642b4e5 --- /dev/null +++ b/vlm/dev/I59qJ0sJ2nh/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:592665e69a472fdca391a99dd2463343c10c5fa6a0a9927be3d998a0b12fe7cc +size 514056 diff --git a/vlm/dev/K2OixmPDou3/0.png b/vlm/dev/K2OixmPDou3/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c60656ae05c4f6c909e1a3a6267f97a3a66cadf7 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ccc5821ea3b829330dd1f86485053eafec3bee25197913d6af0dbb386d173c4 +size 545658 diff --git a/vlm/dev/K2OixmPDou3/1.png b/vlm/dev/K2OixmPDou3/1.png new file mode 100644 index 0000000000000000000000000000000000000000..76848a20bdabd719631746f32e32f33de5bc99e2 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:696015fdee4ac3e831c99ee521e0eb73db7ac9f4d4301a494787f8252087ddcd +size 600142 diff --git a/vlm/dev/K2OixmPDou3/10.png b/vlm/dev/K2OixmPDou3/10.png new file mode 100644 index 0000000000000000000000000000000000000000..98f08f3ec7c31ed8cf7f4d53da1694a7fcce9565 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc2de3fd527a8b3d3e84751004ad2f820f7b787461ab87663a0c9c9ec65602c9 +size 461490 diff --git a/vlm/dev/K2OixmPDou3/11.png b/vlm/dev/K2OixmPDou3/11.png new file mode 100644 index 0000000000000000000000000000000000000000..6abe9d253fda56d7512dd89c774369d7669f1821 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fc1cf0a5458e7dba2cfe0111c869c9ddcc2243d274b7237ba7cdd7e6ba1ce21 +size 434693 diff --git a/vlm/dev/K2OixmPDou3/12.png b/vlm/dev/K2OixmPDou3/12.png new file mode 100644 index 0000000000000000000000000000000000000000..672ddf9bed84adb8ad8c13fa182f22f55c7aaab6 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fc205d3af0df31260e47def91c29efe49bdebe43808bcd2a01fea39c357e4e3 +size 438137 diff --git a/vlm/dev/K2OixmPDou3/13.png b/vlm/dev/K2OixmPDou3/13.png new file mode 100644 index 0000000000000000000000000000000000000000..5f439ad65633186aaa4ae46d72601607743617e5 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e5ec082aa32cf966d4e3d0c0c380b9ca1c1f14f82aa43009358a6364d59f727 +size 418293 diff --git a/vlm/dev/K2OixmPDou3/14.png b/vlm/dev/K2OixmPDou3/14.png new file mode 100644 index 0000000000000000000000000000000000000000..a59a16077d5eaa9dc9212a685b0a121cd3504a3f --- /dev/null +++ b/vlm/dev/K2OixmPDou3/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fac1462434cd838fdceaad0b334fc2e7017cb963b7908abcbdf64f5e9a988c75 +size 537926 diff --git a/vlm/dev/K2OixmPDou3/15.png b/vlm/dev/K2OixmPDou3/15.png new file mode 100644 index 0000000000000000000000000000000000000000..99f38359f2c078774fa5c5958cb05c9f4e5fb208 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff5b0a3a65eedba05db264da91ffe106bde177806ade99fb4cc0ea0189bbcd00 +size 441857 diff --git a/vlm/dev/K2OixmPDou3/16.png b/vlm/dev/K2OixmPDou3/16.png new file mode 100644 index 0000000000000000000000000000000000000000..b8a91f700ddaeafb825b11fbe43f015ab42b3b7a --- /dev/null +++ b/vlm/dev/K2OixmPDou3/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a30df7f1cb9749aee4398057704d4b7554734d852b3304ba588fdc5419665a9 +size 529301 diff --git a/vlm/dev/K2OixmPDou3/17.png b/vlm/dev/K2OixmPDou3/17.png new file mode 100644 index 0000000000000000000000000000000000000000..21ea8532c48269697c614f0718ed2d8eab6327a9 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1eecae1a1c209b82f4eb5f57c10fc373d04cee4124925dbffa1e6d88deea71b +size 343779 diff --git a/vlm/dev/K2OixmPDou3/18.png b/vlm/dev/K2OixmPDou3/18.png new file mode 100644 index 0000000000000000000000000000000000000000..dcea77465cc2a461e4cc5aa7596faeb56c85dbab --- /dev/null +++ b/vlm/dev/K2OixmPDou3/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2e2b4aec978e6f6cc965737d4802c1fef22b70ff839e6dbbc36fc9ca4f4a44c +size 615557 diff --git a/vlm/dev/K2OixmPDou3/19.png b/vlm/dev/K2OixmPDou3/19.png new file mode 100644 index 0000000000000000000000000000000000000000..941dd869b12513656bc8fbdfac2d7143cfb8af2e --- /dev/null +++ b/vlm/dev/K2OixmPDou3/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6883bd5b5efc3d245f9d534ad3b33435664b237dca7f722d467c53f3ae9c6c8b +size 403043 diff --git a/vlm/dev/K2OixmPDou3/2.png b/vlm/dev/K2OixmPDou3/2.png new file mode 100644 index 0000000000000000000000000000000000000000..310b1fb06a76d9b8b0cefed9f2f1ffdf31a866e5 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4e713b97be5656af156b015ea7f465b6302b205fac7c786510add0c169fc60b +size 567728 diff --git a/vlm/dev/K2OixmPDou3/20.png b/vlm/dev/K2OixmPDou3/20.png new file mode 100644 index 0000000000000000000000000000000000000000..29427a46af40fd4cddbe1d4cb4f27ae740279000 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5aa5cc14c6cdfde0b16f3a1f70750b610353ff3eded10236ad7e99073fc09bd5 +size 422487 diff --git a/vlm/dev/K2OixmPDou3/21.png b/vlm/dev/K2OixmPDou3/21.png new file mode 100644 index 0000000000000000000000000000000000000000..87c7e227ce8b8b7cf4cf9236df4d986ca2488b70 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc6b471cafaf214b8929525d01844aaa56bba8754d44a4f7d041de634fcd0b3f +size 650314 diff --git a/vlm/dev/K2OixmPDou3/22.png b/vlm/dev/K2OixmPDou3/22.png new file mode 100644 index 0000000000000000000000000000000000000000..c9a1a436965f3c7468dffdf213fd38c35e3ef19f --- /dev/null +++ b/vlm/dev/K2OixmPDou3/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c418fe329d0fbf62243fbdca5477b0ba53e4904abdf7b8b260f96b993036c0f4 +size 637649 diff --git a/vlm/dev/K2OixmPDou3/23.png b/vlm/dev/K2OixmPDou3/23.png new file mode 100644 index 0000000000000000000000000000000000000000..3504175e01254a6dd1a501cd7dd3c725797fa029 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddef1623031445d42f7b12cb5457516eeaf9008f0cff4ef990e55a26e8ffa789 +size 636940 diff --git a/vlm/dev/K2OixmPDou3/24.png b/vlm/dev/K2OixmPDou3/24.png new file mode 100644 index 0000000000000000000000000000000000000000..76060aabbf17944eb9f6fab4ef9f3db108991ed0 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a1c16f37b8963cb31b383970f258c2ed62b8c42f1005c6e34df328adc496014 +size 563957 diff --git a/vlm/dev/K2OixmPDou3/25.png b/vlm/dev/K2OixmPDou3/25.png new file mode 100644 index 0000000000000000000000000000000000000000..9d575f62b28f17d05b776d3ae671451a9323a3a1 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb9634c7faa2a192b323e978133b8d276c7f3b2e3ff3cf933ebb6e2cca99d026 +size 526424 diff --git a/vlm/dev/K2OixmPDou3/26.png b/vlm/dev/K2OixmPDou3/26.png new file mode 100644 index 0000000000000000000000000000000000000000..6e8d3230293d19932c61a8fffe5e91da080c9e68 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d86fa31b1a943ad67f31c3baa55c0663fe0956f5701aa153fc02c9ace949275 +size 661219 diff --git a/vlm/dev/K2OixmPDou3/27.png b/vlm/dev/K2OixmPDou3/27.png new file mode 100644 index 0000000000000000000000000000000000000000..6015eaaf2ff6eaa3c9ab75417e70eb33c647a228 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce59d23ae600cab6a03ed5e1de819c2ba4b4fc059ac5a89aa80d5dd73ede08cb +size 1333762 diff --git a/vlm/dev/K2OixmPDou3/3.png b/vlm/dev/K2OixmPDou3/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e9604d35c0ad4f6d1b71476160dff8c0dc6dd0a1 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:918614971744367f6024acb855691bc12e9dcd56adc6f8bc222ac6f547b071b0 +size 703439 diff --git a/vlm/dev/K2OixmPDou3/4.png b/vlm/dev/K2OixmPDou3/4.png new file mode 100644 index 0000000000000000000000000000000000000000..62162a12766d449c311d15d9132fb896c09fb647 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dbd9db9f0111315c3af2774e0954d4eb57d724395d2c7136958ac8c3a6a75a5 +size 620138 diff --git a/vlm/dev/K2OixmPDou3/5.png b/vlm/dev/K2OixmPDou3/5.png new file mode 100644 index 0000000000000000000000000000000000000000..7b765b91c008924a01b027197ed3d99ba7e6a3be --- /dev/null +++ b/vlm/dev/K2OixmPDou3/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e669f99de4ec1f5b21c055bd2d1c946acda5e9fd577da98b4ca05cf620a96a1b +size 659173 diff --git a/vlm/dev/K2OixmPDou3/6.png b/vlm/dev/K2OixmPDou3/6.png new file mode 100644 index 0000000000000000000000000000000000000000..eaa73a8ce6c26de91eee3d604b5c101200828090 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cffc7b2fee18b0dddc49da1628554ec70e40c601150bd157777a16fe1aa6a72 +size 601814 diff --git a/vlm/dev/K2OixmPDou3/7.png b/vlm/dev/K2OixmPDou3/7.png new file mode 100644 index 0000000000000000000000000000000000000000..9cceac2bdc2d554b0957ddd42a35c2ede37d7e9b --- /dev/null +++ b/vlm/dev/K2OixmPDou3/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b7d56c41e339362896502cb6a89a46648362bae6523e1f6558adc7817116d21 +size 646258 diff --git a/vlm/dev/K2OixmPDou3/8.png b/vlm/dev/K2OixmPDou3/8.png new file mode 100644 index 0000000000000000000000000000000000000000..f9686a9a6c6fe094ed8e1c7a2ed058b00621c2b1 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8eb0523420db36f2e19e65189967db45d6b07d5e43a636c43f654fab77d073e1 +size 594773 diff --git a/vlm/dev/K2OixmPDou3/9.png b/vlm/dev/K2OixmPDou3/9.png new file mode 100644 index 0000000000000000000000000000000000000000..c2889465e157b0cb08b7e0e112722912aab7e1a3 --- /dev/null +++ b/vlm/dev/K2OixmPDou3/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb322e8e6dbe1ce586f09934228676f06527594068a478a60a06968cc09797f8 +size 476543 diff --git a/vlm/dev/LvyJX20Rll/0.png b/vlm/dev/LvyJX20Rll/0.png new file mode 100644 index 0000000000000000000000000000000000000000..051064742c291354ab5b2fe419e10cdf58e7fe95 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c5f109e8fd01de416916079122f3b70cf98ac4dbce92f3c7874c70995a77d68 +size 455235 diff --git a/vlm/dev/LvyJX20Rll/1.png b/vlm/dev/LvyJX20Rll/1.png new file mode 100644 index 0000000000000000000000000000000000000000..e3e2e74a959b9e8bf066fa7905c2c9257b2f552f --- /dev/null +++ b/vlm/dev/LvyJX20Rll/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d49239dda9f381a34e21d6c8cb0aa8f44b974aaacf0b666b67e3ad39d758c9b +size 630006 diff --git a/vlm/dev/LvyJX20Rll/10.png b/vlm/dev/LvyJX20Rll/10.png new file mode 100644 index 0000000000000000000000000000000000000000..e6d1c21b66a9f4a4d04cf29e0674eea6960dd6e4 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1be0c990d59625aa3d148399e7f7dde4fc3c91370232b71112f79e459c5fd133 +size 527504 diff --git a/vlm/dev/LvyJX20Rll/11.png b/vlm/dev/LvyJX20Rll/11.png new file mode 100644 index 0000000000000000000000000000000000000000..6203843540698f7d5728221180f1e78b8d32b1ba --- /dev/null +++ b/vlm/dev/LvyJX20Rll/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8df5ba64711136a6763208923edd70310c8950df48d3e14e3e5c0bac0442793f +size 534365 diff --git a/vlm/dev/LvyJX20Rll/12.png b/vlm/dev/LvyJX20Rll/12.png new file mode 100644 index 0000000000000000000000000000000000000000..d8a428e81e75fe4833f47cf575d35fc4c1af2158 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e961e30a6153c6bdaa80b882025acef01497a329ec457c1e56f97ea3edad0d9b +size 575017 diff --git a/vlm/dev/LvyJX20Rll/13.png b/vlm/dev/LvyJX20Rll/13.png new file mode 100644 index 0000000000000000000000000000000000000000..ee9287d229f5390a75e748397eecd6d015a189fa --- /dev/null +++ b/vlm/dev/LvyJX20Rll/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f42a972baa5abb6a839419ccf777d15c548f7f885006c48256c7d9186cae2ac +size 565052 diff --git a/vlm/dev/LvyJX20Rll/2.png b/vlm/dev/LvyJX20Rll/2.png new file mode 100644 index 0000000000000000000000000000000000000000..806f5b5824e3b359877287e136ba72c939d5837c --- /dev/null +++ b/vlm/dev/LvyJX20Rll/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:103b9f9a2475e548c185861a28a112094cb6b6332992a4d25dcbf20870414190 +size 561939 diff --git a/vlm/dev/LvyJX20Rll/3.png b/vlm/dev/LvyJX20Rll/3.png new file mode 100644 index 0000000000000000000000000000000000000000..1f3051f20a48277f746897817de0bc32e6cab097 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8059662ed154ca2d039d07a71a39369d1ef37b778cca43bed48bc1b02192e960 +size 603326 diff --git a/vlm/dev/LvyJX20Rll/4.png b/vlm/dev/LvyJX20Rll/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ea9ec170789e8da89f57e20ccc56e82e34dcd7fd --- /dev/null +++ b/vlm/dev/LvyJX20Rll/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:878d90ba63c48447c995dcab30cd84cdb92e21fdc46569441e82b3ff341acd68 +size 617267 diff --git a/vlm/dev/LvyJX20Rll/5.png b/vlm/dev/LvyJX20Rll/5.png new file mode 100644 index 0000000000000000000000000000000000000000..2ce8b7acd5c08362b2c1d14bc0616efa7b3d1211 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45f0c88b1d2c00f7a094d377a6f6c5939135ab9de28c6b0855c41050c4a60fbb +size 556961 diff --git a/vlm/dev/LvyJX20Rll/6.png b/vlm/dev/LvyJX20Rll/6.png new file mode 100644 index 0000000000000000000000000000000000000000..5f91997c03b1f99cb7d6f15e35072a6d0c6f7428 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1b06b313a81d3365ad7a2236444b35df9a3c3e97f1b52201861ce0160d19229 +size 434618 diff --git a/vlm/dev/LvyJX20Rll/7.png b/vlm/dev/LvyJX20Rll/7.png new file mode 100644 index 0000000000000000000000000000000000000000..8ea5bc046f3866bcf63164bb9e758029018ea316 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b150f0d99917415c3bc3d2bc1f853c54bbb30f260a25c8be8d7dbe0e42b70632 +size 551693 diff --git a/vlm/dev/LvyJX20Rll/8.png b/vlm/dev/LvyJX20Rll/8.png new file mode 100644 index 0000000000000000000000000000000000000000..97e868aa404f5a9700cc659e6eb741cf80e585c6 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7805272d33852a9237cd89611573f014f72a893f3982f63f92d877f55f01ca2f +size 560427 diff --git a/vlm/dev/LvyJX20Rll/9.png b/vlm/dev/LvyJX20Rll/9.png new file mode 100644 index 0000000000000000000000000000000000000000..dc9981d39021b84c1eef0bf7d75add26b39c2922 --- /dev/null +++ b/vlm/dev/LvyJX20Rll/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3413fe52c6163912a89bdea0a080ebb7827a282302cd42cf7441ba0484f1292b +size 420181 diff --git a/vlm/dev/M2sNIiCC6C/0.png b/vlm/dev/M2sNIiCC6C/0.png new file mode 100644 index 0000000000000000000000000000000000000000..7913d1ba3404f6953517479ec57c75127e22b7c0 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a35044144cb01f3e7fb9ab31cee272dee3e80c0448dfc11d823e088efe9e5e52 +size 477185 diff --git a/vlm/dev/M2sNIiCC6C/1.png b/vlm/dev/M2sNIiCC6C/1.png new file mode 100644 index 0000000000000000000000000000000000000000..7636fa8bfa18916b5d16aa7ea581bfabb1f2edcf --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c0edfead0f771ee22f2d7b5ac5137f27ffed89d0bace57cf8c1dbbeb52e61cc +size 896800 diff --git a/vlm/dev/M2sNIiCC6C/10.png b/vlm/dev/M2sNIiCC6C/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c7b30be5367031632673748881299a6c1c174a04 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c0a00405f0353ff197fbd46369181e2a125d048a805aaeb25c9a54f8ff5c940 +size 530753 diff --git a/vlm/dev/M2sNIiCC6C/11.png b/vlm/dev/M2sNIiCC6C/11.png new file mode 100644 index 0000000000000000000000000000000000000000..8c426526c6f545222d12ab83955ebaf7b7675b78 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afb8249ba8b4bc94fa562e3e866fc4df739a90516e6cf7a01958542ec02852d2 +size 195034 diff --git a/vlm/dev/M2sNIiCC6C/12.png b/vlm/dev/M2sNIiCC6C/12.png new file mode 100644 index 0000000000000000000000000000000000000000..712d4298b64cceb216e4ffa84f53024854696ba8 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac9201e5778233484ba348bd57a1366f89e2376b4b564fdbb500c2a80888f569 +size 329725 diff --git a/vlm/dev/M2sNIiCC6C/13.png b/vlm/dev/M2sNIiCC6C/13.png new file mode 100644 index 0000000000000000000000000000000000000000..09d1ff8f2534c970bb6e7b17128fa158fa4f6475 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:205cf7360956cb7ccd06e58444c9fcc4c2f6d49cef2776c6fcafdc23be18ec17 +size 695465 diff --git a/vlm/dev/M2sNIiCC6C/14.png b/vlm/dev/M2sNIiCC6C/14.png new file mode 100644 index 0000000000000000000000000000000000000000..d025a43fbcc0172421ffb77f947bfa5b282d571b --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55467d3c03b20d07495fd4b0301455d997b0eae6d97a61b42a51404121e21b0b +size 935949 diff --git a/vlm/dev/M2sNIiCC6C/15.png b/vlm/dev/M2sNIiCC6C/15.png new file mode 100644 index 0000000000000000000000000000000000000000..1565feed102adcc45421666d8a53793948a02cfc --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d17fe20ea11a653276cd4323ed9c713206a18897454116f3752698baa3156b2 +size 427997 diff --git a/vlm/dev/M2sNIiCC6C/16.png b/vlm/dev/M2sNIiCC6C/16.png new file mode 100644 index 0000000000000000000000000000000000000000..dd5fdff82b05a030bebe59bd37a8951f7874d95d --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6eee314a87b1bcfaf0367feda15edd83e02c0bd1435252b7beb9050bafb1a183 +size 621918 diff --git a/vlm/dev/M2sNIiCC6C/17.png b/vlm/dev/M2sNIiCC6C/17.png new file mode 100644 index 0000000000000000000000000000000000000000..f98dd16c869623d922f30934f04644db959795db --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a688cdb2a934663b5b03ee159b04d66f3c65d00b7825b25ed01d2a909975b81 +size 567064 diff --git a/vlm/dev/M2sNIiCC6C/18.png b/vlm/dev/M2sNIiCC6C/18.png new file mode 100644 index 0000000000000000000000000000000000000000..522cb7ab27987cfbf60d8d391bf40482df32f8f1 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18431ab00744672d111a5c3161ff6ca47f6714880ae4a13081fb8e31d5ce6058 +size 3009901 diff --git a/vlm/dev/M2sNIiCC6C/19.png b/vlm/dev/M2sNIiCC6C/19.png new file mode 100644 index 0000000000000000000000000000000000000000..96773c3b37a7cb120919299a5bfcc10a6b9a94fa --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:707989ef384d0bbee7efd4496b23c24a5bed84badc148784b384036b7aac4a1a +size 1570252 diff --git a/vlm/dev/M2sNIiCC6C/2.png b/vlm/dev/M2sNIiCC6C/2.png new file mode 100644 index 0000000000000000000000000000000000000000..18ceb3de3f588cca4c8d90a917dc18c68fbd0c79 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b65aebdce82839b1c6fbd2db26f9ea136a1dce6f9b561cfb7b3f455ed2244816 +size 533997 diff --git a/vlm/dev/M2sNIiCC6C/20.png b/vlm/dev/M2sNIiCC6C/20.png new file mode 100644 index 0000000000000000000000000000000000000000..5a7f9a3108cc8ae1530f9faf97d9ed8c66ba04de --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cf3e3fe23aa39b02eb05ee3cb5ff0dcd849f0210c254382f44ef8b1fc7bf781 +size 455285 diff --git a/vlm/dev/M2sNIiCC6C/21.png b/vlm/dev/M2sNIiCC6C/21.png new file mode 100644 index 0000000000000000000000000000000000000000..bb143a653cdb9e193a53bc7fec3bca58f30db410 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf91bb071cc73c05217e1c3859b35788299de344ee97f81dd9462e603244a4e9 +size 1385896 diff --git a/vlm/dev/M2sNIiCC6C/22.png b/vlm/dev/M2sNIiCC6C/22.png new file mode 100644 index 0000000000000000000000000000000000000000..ce5a619f7124f04de6ef52326e30e34d26e0d031 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0da453238cf5c82b6500d19160e0e13917d82a439bf5876396f45f37c67b9559 +size 1985653 diff --git a/vlm/dev/M2sNIiCC6C/23.png b/vlm/dev/M2sNIiCC6C/23.png new file mode 100644 index 0000000000000000000000000000000000000000..64ad34e5f592b99b69786bad87c6404b1c0880eb --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d43403f4f89181cd5031a42fb01646c78506afcfb6c5d0b5db3fe55dd93bc9ee +size 1269183 diff --git a/vlm/dev/M2sNIiCC6C/24.png b/vlm/dev/M2sNIiCC6C/24.png new file mode 100644 index 0000000000000000000000000000000000000000..6198840cb969d019ad25c17f49cb9cd9b31343f5 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0efa17d5d514b337962904fddc30873434d6107c5216aee5c4f591e95e987648 +size 903994 diff --git a/vlm/dev/M2sNIiCC6C/25.png b/vlm/dev/M2sNIiCC6C/25.png new file mode 100644 index 0000000000000000000000000000000000000000..7395b866dd135609ba28c5376b79bddd36833d3a --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:774e2fe1347638288fb73037c4bf64ae8e1b1a171f3e3a8bd84f90c61c9bb964 +size 1589635 diff --git a/vlm/dev/M2sNIiCC6C/26.png b/vlm/dev/M2sNIiCC6C/26.png new file mode 100644 index 0000000000000000000000000000000000000000..6ce26496fa0f4782cf2c0f36466012f1e4330776 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ab1a02f7c026cf73f8d707037042efc7892561f3de01fa16e884bc36188ca03 +size 550841 diff --git a/vlm/dev/M2sNIiCC6C/27.png b/vlm/dev/M2sNIiCC6C/27.png new file mode 100644 index 0000000000000000000000000000000000000000..70952929b2efea7b01f2762e6b28947d7fec211d --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b80b15c6c02227cdaf2d031f5f577ab4a0768a50e13c563ac575e78c2dfe80c +size 297077 diff --git a/vlm/dev/M2sNIiCC6C/3.png b/vlm/dev/M2sNIiCC6C/3.png new file mode 100644 index 0000000000000000000000000000000000000000..7086ad6bd01ae351fdc428cec73d18927c6090fa --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5af3fc91c94a2345c38a71c6c616527d98e5582db3745221a6f7c4b8282fd05 +size 864835 diff --git a/vlm/dev/M2sNIiCC6C/4.png b/vlm/dev/M2sNIiCC6C/4.png new file mode 100644 index 0000000000000000000000000000000000000000..a69028fe3c007d2e185102b96b3eb25a118fdc4a --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc030ae02f849cdeea68b7c9b4c0e4d62739e7d3da68f8e12fb9936a5cb4ae6a +size 607158 diff --git a/vlm/dev/M2sNIiCC6C/5.png b/vlm/dev/M2sNIiCC6C/5.png new file mode 100644 index 0000000000000000000000000000000000000000..5236ada5476e190f4266d33edcac62b37e308772 --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2b28f2abd426a3ff20cc3f96c2372b84abb12d185903a3172dec3b16a6c4ec9 +size 883371 diff --git a/vlm/dev/M2sNIiCC6C/6.png b/vlm/dev/M2sNIiCC6C/6.png new file mode 100644 index 0000000000000000000000000000000000000000..33685f54d80265f9ae4ec3a4b616048191bdacff --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f6e38f260752f97ae3a7b63adf466d1f972fdbce3f269661cbfdbee75d1cb57 +size 650541 diff --git a/vlm/dev/M2sNIiCC6C/7.png b/vlm/dev/M2sNIiCC6C/7.png new file mode 100644 index 0000000000000000000000000000000000000000..e204454ae9730ce7d0c03325f5a3fc2402f22ceb --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c334899384b3784e474c809db7ea1572e57b58f32e1889c741acf7cea5f5db3e +size 1378474 diff --git a/vlm/dev/M2sNIiCC6C/8.png b/vlm/dev/M2sNIiCC6C/8.png new file mode 100644 index 0000000000000000000000000000000000000000..fc09ada95a1f85daddb694a896e89fa017c332de --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d18873a6044c4b5ceeb04d42449d8f3e0dfea5e0d611881703028a2a4215926c +size 880406 diff --git a/vlm/dev/M2sNIiCC6C/9.png b/vlm/dev/M2sNIiCC6C/9.png new file mode 100644 index 0000000000000000000000000000000000000000..5b668b636e51b75d21f1598ab16e9d7cd82a71ad --- /dev/null +++ b/vlm/dev/M2sNIiCC6C/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30192c344755c41bd384523f4f4236542b8c69034c07ca763815d6158ad9a8da +size 532749 diff --git a/vlm/dev/N8MaByOzUfb/0.png b/vlm/dev/N8MaByOzUfb/0.png new file mode 100644 index 0000000000000000000000000000000000000000..4bdc64deac1b9edea4ab03f527625f91e8e44544 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d7a8ee836758a723fe919d6acaa073fb3f18551b9456db74ee9132578658115 +size 481673 diff --git a/vlm/dev/N8MaByOzUfb/1.png b/vlm/dev/N8MaByOzUfb/1.png new file mode 100644 index 0000000000000000000000000000000000000000..9a6dabeb1ff3d7bc3a8c9ff009674476e2f99208 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b8f438f010166671179ea71944b5bbd5f3f68b235419249c9acde2587c31ef7 +size 621630 diff --git a/vlm/dev/N8MaByOzUfb/10.png b/vlm/dev/N8MaByOzUfb/10.png new file mode 100644 index 0000000000000000000000000000000000000000..9da467a6d95843cc20de7cd8705ff5c92fddcf94 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc00b3c201300918eb4b681d34be68bbd62b0cc58097b77f0b011bf2497b4deb +size 598412 diff --git a/vlm/dev/N8MaByOzUfb/11.png b/vlm/dev/N8MaByOzUfb/11.png new file mode 100644 index 0000000000000000000000000000000000000000..facfe5910129e9d1983819fb17e5dc3740322869 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc3e7a296add636246e6ce53a8757647ab124fc010ada1b30f24a43bdcc5e87c +size 543601 diff --git a/vlm/dev/N8MaByOzUfb/12.png b/vlm/dev/N8MaByOzUfb/12.png new file mode 100644 index 0000000000000000000000000000000000000000..d6542149bd6cae2fffe642a78946e7e9ea35908b --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e14f5b5529253828a67eade83b0bd6f1d2d3a1721d079ff4de429ad5c2ca966c +size 386435 diff --git a/vlm/dev/N8MaByOzUfb/13.png b/vlm/dev/N8MaByOzUfb/13.png new file mode 100644 index 0000000000000000000000000000000000000000..4ed15b72d3854176d456264be02a66a4416f7f29 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f22201d2332cb9b76ee5c01497dd3c11c58300bbd947e11f699e487629a8cebb +size 567727 diff --git a/vlm/dev/N8MaByOzUfb/14.png b/vlm/dev/N8MaByOzUfb/14.png new file mode 100644 index 0000000000000000000000000000000000000000..4bc493bbf6853d7bf8b7e7b56c8acbc28f7f5f8b --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cc7e618e256889808948a1b6a0b139ed468347df4e7781ddf279341aecbe18a +size 279384 diff --git a/vlm/dev/N8MaByOzUfb/15.png b/vlm/dev/N8MaByOzUfb/15.png new file mode 100644 index 0000000000000000000000000000000000000000..fcfa6250dfab3d7a836014c947ed7c94af40927b --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62a213564baf5574514d7f04f0a6b66eb91e7d84fc8cec8dab2ce734f3d605d3 +size 374210 diff --git a/vlm/dev/N8MaByOzUfb/16.png b/vlm/dev/N8MaByOzUfb/16.png new file mode 100644 index 0000000000000000000000000000000000000000..d2d6c21d47f159a52c6bfcfcbc180463ff022e27 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:582337a04c344080a4f6a72f253917a5b6b8cb9c474b13411c1dfd9aa2a93ea7 +size 407009 diff --git a/vlm/dev/N8MaByOzUfb/17.png b/vlm/dev/N8MaByOzUfb/17.png new file mode 100644 index 0000000000000000000000000000000000000000..8bace4c222669294274654078cdef7632b7264f2 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6d48851adebc4dc1d741a0dba9ebce934cca39f947674e43e901358b1188fce +size 448335 diff --git a/vlm/dev/N8MaByOzUfb/18.png b/vlm/dev/N8MaByOzUfb/18.png new file mode 100644 index 0000000000000000000000000000000000000000..f81a01e69a291e8cec93c5a58ca45e6970346aae --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c26fed6a13762a87cdc30ef6e13a6bdc6b277e1b93e662e7ab4c2c023f537ff1 +size 544771 diff --git a/vlm/dev/N8MaByOzUfb/19.png b/vlm/dev/N8MaByOzUfb/19.png new file mode 100644 index 0000000000000000000000000000000000000000..b71e8dde104c91be48253eadf104ab8be3aa12fb --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c06af0fe99f683da813da4662296626f5e09321dce49b8e3330d4127ba3ed36c +size 173244 diff --git a/vlm/dev/N8MaByOzUfb/2.png b/vlm/dev/N8MaByOzUfb/2.png new file mode 100644 index 0000000000000000000000000000000000000000..5f98a44101e01a1e8646a59f49aafdfbc1149821 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc34b0b17a0aa2d1d0c39b05eea1aec8b1b620be18a48fa719ec4898fb7b6e72 +size 645079 diff --git a/vlm/dev/N8MaByOzUfb/20.png b/vlm/dev/N8MaByOzUfb/20.png new file mode 100644 index 0000000000000000000000000000000000000000..0915e1f2b0f7f8d9134db5cf9cdacf2a8df8577f --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f0196ea77e592290aee0c7dd2773127e34d1b69c2fd17f8ab2c343c011dec08 +size 343851 diff --git a/vlm/dev/N8MaByOzUfb/21.png b/vlm/dev/N8MaByOzUfb/21.png new file mode 100644 index 0000000000000000000000000000000000000000..65b480bc02ec9c659c84cf9ec7fc2d805294801c --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:209737b45537b6c7c2e6b5d7d2373a2145d49e8d5bf839e8217e0969652a119c +size 274384 diff --git a/vlm/dev/N8MaByOzUfb/22.png b/vlm/dev/N8MaByOzUfb/22.png new file mode 100644 index 0000000000000000000000000000000000000000..4d1908fbacb523d5e666745f3b38b4f4b714600c --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e995e58fcb5beadea08d9b5e8e6ad9cf55171e4e69a1fdd758a71b5b33f7c756 +size 375972 diff --git a/vlm/dev/N8MaByOzUfb/23.png b/vlm/dev/N8MaByOzUfb/23.png new file mode 100644 index 0000000000000000000000000000000000000000..d18371637bb22b08291eda9c178d870fa832ac0b --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30c903ebeeb45dacfea492c0f415e3b9eb0d83de6767a8724cb24cb9b72c036b +size 362494 diff --git a/vlm/dev/N8MaByOzUfb/24.png b/vlm/dev/N8MaByOzUfb/24.png new file mode 100644 index 0000000000000000000000000000000000000000..25dd762e4f1ef79c8c49b147281aa946ac56dea5 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:639073065a5172b0a86cce65c17cf7f99b6b134b7d83b72f2b7d6845d5fe9001 +size 450591 diff --git a/vlm/dev/N8MaByOzUfb/25.png b/vlm/dev/N8MaByOzUfb/25.png new file mode 100644 index 0000000000000000000000000000000000000000..0ef95a30ca4d5bb4b480f572da2743bb06d519fd --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e510bd8f4fba0d0cc6c54699c441f84a9d43ed01790fee8b39bd8d8bea730003 +size 378658 diff --git a/vlm/dev/N8MaByOzUfb/26.png b/vlm/dev/N8MaByOzUfb/26.png new file mode 100644 index 0000000000000000000000000000000000000000..0da9cd109e31bbb9ac4790b2292f0389fc6298ed --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:040761995d2eb1edb2c25292769ed22c8fa332a0ed2048eeac172c8cc8aaff16 +size 365278 diff --git a/vlm/dev/N8MaByOzUfb/3.png b/vlm/dev/N8MaByOzUfb/3.png new file mode 100644 index 0000000000000000000000000000000000000000..fdbc4bda00393f30561544a6846e3f5c4d5de8cd --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26c7837e14f88f617e908ce90f5cca5d07520956d6c6694b19f23f5eb49977fd +size 580604 diff --git a/vlm/dev/N8MaByOzUfb/4.png b/vlm/dev/N8MaByOzUfb/4.png new file mode 100644 index 0000000000000000000000000000000000000000..4ee763b47ba7a4a0fde32b36255f96ca84cc5eb5 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab332ef9aec8366cf944c45499b39cacceaa8ae61f00c700384ca3a84846ec95 +size 635909 diff --git a/vlm/dev/N8MaByOzUfb/5.png b/vlm/dev/N8MaByOzUfb/5.png new file mode 100644 index 0000000000000000000000000000000000000000..3e1de76b31afa19c378c97e19baf7a38aa0085d0 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9991b2937bd7c4f1c109ae5f285406d33bf047fbb541406f0559a194728b6546 +size 566551 diff --git a/vlm/dev/N8MaByOzUfb/6.png b/vlm/dev/N8MaByOzUfb/6.png new file mode 100644 index 0000000000000000000000000000000000000000..073ad3acc8f6d51e8e84cf041775dc45333faf76 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4cb352aab37869b11371f5f12ab72a2f55e6d2ff4b203842ead476351fd0ba5 +size 576475 diff --git a/vlm/dev/N8MaByOzUfb/7.png b/vlm/dev/N8MaByOzUfb/7.png new file mode 100644 index 0000000000000000000000000000000000000000..ad130f2e1d22b22504779160c77b56ebac5b3e5b --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adb6b5c11cecfdb469c4669155e68894f99f0f38a47088d7d93aedd3c5dde748 +size 573823 diff --git a/vlm/dev/N8MaByOzUfb/8.png b/vlm/dev/N8MaByOzUfb/8.png new file mode 100644 index 0000000000000000000000000000000000000000..d5c5487c7ba795f3f1d561b222ac1d8383eea495 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:509bb192b1131d7edb4b59aac805bf437d72819068affc08e2ff20e5ff0ed109 +size 701338 diff --git a/vlm/dev/N8MaByOzUfb/9.png b/vlm/dev/N8MaByOzUfb/9.png new file mode 100644 index 0000000000000000000000000000000000000000..fd5b10a36a22afb0b21dd1ab387589ab0212d113 --- /dev/null +++ b/vlm/dev/N8MaByOzUfb/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da538e590875cda7455a36de4a7cf7b07330ac98bbe8fec54014b7d7aef5ef83 +size 494447 diff --git a/vlm/dev/O4kDO3yS9B/0.png b/vlm/dev/O4kDO3yS9B/0.png new file mode 100644 index 0000000000000000000000000000000000000000..11f19b004e5554c31f283be35c7caabbdcbb804b --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e643f5b79fbd8dc4bdad8c3d4e0c26838fdc81658d981d65dbc46ef0309a5548 +size 728519 diff --git a/vlm/dev/O4kDO3yS9B/1.png b/vlm/dev/O4kDO3yS9B/1.png new file mode 100644 index 0000000000000000000000000000000000000000..999e87913ad276a54f66e23ef0ad2cbfa82a4137 --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40982dffcbb3c35e6c7d4649b7e6746db34505bff5f2c938ddb93f53589c973c +size 768353 diff --git a/vlm/dev/O4kDO3yS9B/10.png b/vlm/dev/O4kDO3yS9B/10.png new file mode 100644 index 0000000000000000000000000000000000000000..0b34fdf147421c5b8bc7a2e69c24d65b67d8a84d --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:709681fba9ad068dab422d2aa5bd4183374c5ba58791e109a14c95ee9baade05 +size 885193 diff --git a/vlm/dev/O4kDO3yS9B/11.png b/vlm/dev/O4kDO3yS9B/11.png new file mode 100644 index 0000000000000000000000000000000000000000..4b48736f516f39dade21a789a3cad727bfa358fc --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e3d66be15e73048876e947aabb24cea101fc94dc4eb9853c42cfe0990ebf573 +size 673858 diff --git a/vlm/dev/O4kDO3yS9B/12.png b/vlm/dev/O4kDO3yS9B/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b9d13e8381f9e6630bf51c2001b0e6a86f274c27 --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:762cfa9d82a90a3cd5f12fa02cd6ca3c46cf7342cfa6d651e3cfb8cca4436ea4 +size 546435 diff --git a/vlm/dev/O4kDO3yS9B/13.png b/vlm/dev/O4kDO3yS9B/13.png new file mode 100644 index 0000000000000000000000000000000000000000..a3e82e208d8ebd8a373657ab34c447cff57286ab --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59ce53420d3d1361eb02a11242d10e5bc7095be0dab77276a55901cd03b7e727 +size 536404 diff --git a/vlm/dev/O4kDO3yS9B/2.png b/vlm/dev/O4kDO3yS9B/2.png new file mode 100644 index 0000000000000000000000000000000000000000..a2b5ddfe93837daf05bf6c3453835725965af5da --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61e434d643bad88c81288befc6b43ad72d4b5a9c88cc4e865150e9d317095bad +size 822559 diff --git a/vlm/dev/O4kDO3yS9B/3.png b/vlm/dev/O4kDO3yS9B/3.png new file mode 100644 index 0000000000000000000000000000000000000000..92aa0535c2acdc981042ce60466598929e3dd66a --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e121033c7ff09345846eeb38cea6c510b76cc4c3f19d978ddc1b6fb239232c50 +size 604494 diff --git a/vlm/dev/O4kDO3yS9B/4.png b/vlm/dev/O4kDO3yS9B/4.png new file mode 100644 index 0000000000000000000000000000000000000000..24a991fde7e4459c7540e7e9d7295163643b0c8e --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd42c1125b7eb384de649e6845798c25d13b109da237102e6522a1a55a9390c5 +size 700537 diff --git a/vlm/dev/O4kDO3yS9B/5.png b/vlm/dev/O4kDO3yS9B/5.png new file mode 100644 index 0000000000000000000000000000000000000000..2d9f96a6dba9caee974d19c99787b92f170ea398 --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24fe4b43daf8acd0100820dc2fc1e950b9914a2f59614320572a0ffae8c4df5b +size 749791 diff --git a/vlm/dev/O4kDO3yS9B/6.png b/vlm/dev/O4kDO3yS9B/6.png new file mode 100644 index 0000000000000000000000000000000000000000..0b0d670f4524d864200946b9f8fa3ddbf2e6f3e0 --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bf78a9a6455ccc9f9a301ff5b06cacc2a83f1b68e58b6fd81b099bcd6d17a33 +size 704051 diff --git a/vlm/dev/O4kDO3yS9B/7.png b/vlm/dev/O4kDO3yS9B/7.png new file mode 100644 index 0000000000000000000000000000000000000000..b5db3b88007afb8c355a9b31f1f87231d861858b --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae9fc8ddf57ce52017db746dcc532e799a08edbfc96428a9dbb12bb985e570e6 +size 623609 diff --git a/vlm/dev/O4kDO3yS9B/8.png b/vlm/dev/O4kDO3yS9B/8.png new file mode 100644 index 0000000000000000000000000000000000000000..469a7bd6901d589b003f5415545df2d9a3fe010b --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c853e0511a79e20c589c44b06274a3eb16fa44265c20ab544881f3ec1da941e5 +size 723735 diff --git a/vlm/dev/O4kDO3yS9B/9.png b/vlm/dev/O4kDO3yS9B/9.png new file mode 100644 index 0000000000000000000000000000000000000000..f1023a99af86aaa4b44504371c86119ec8b7b25e --- /dev/null +++ b/vlm/dev/O4kDO3yS9B/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c3105658694e08ac9c68ce7e4e98b1efff44c4c1a22f52dac4469f4785692bb +size 843415 diff --git a/vlm/dev/QeRAyn4igEA/0.png b/vlm/dev/QeRAyn4igEA/0.png new file mode 100644 index 0000000000000000000000000000000000000000..8bd99d65092ef2e3d0178c9102410a9746906d31 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:134457520f8f645cdb60090c9a96331383ae109addf611c47269d3239053e65e +size 462402 diff --git a/vlm/dev/QeRAyn4igEA/1.png b/vlm/dev/QeRAyn4igEA/1.png new file mode 100644 index 0000000000000000000000000000000000000000..17086697bce681cdc26c8711d6e8558759ae521e --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba61f66eec4a6d28cae6c3d0832e76970092d12b1f217ee33e9b2066e74e4dd8 +size 643049 diff --git a/vlm/dev/QeRAyn4igEA/10.png b/vlm/dev/QeRAyn4igEA/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c34232dc6448e94b3369c10cf4cd3d2f908f92b2 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4be4780767972ad7b2d7dbc581bf8f05d2030bedc2156d27969cbcdea5cf6474 +size 625566 diff --git a/vlm/dev/QeRAyn4igEA/11.png b/vlm/dev/QeRAyn4igEA/11.png new file mode 100644 index 0000000000000000000000000000000000000000..6ce12aacd990ff43c6c7f20351e702152d00e631 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2df3347ec78605e8edcf5564d43e596e8280fd1bfa5b1218bd08ae2e7a7528c +size 115111 diff --git a/vlm/dev/QeRAyn4igEA/12.png b/vlm/dev/QeRAyn4igEA/12.png new file mode 100644 index 0000000000000000000000000000000000000000..24c08e184d96904b851598090b9945869c512930 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad9d06407a5d8870dce138618effd702a3f3119ee4692fcd040abd94a93f6cdf +size 355176 diff --git a/vlm/dev/QeRAyn4igEA/2.png b/vlm/dev/QeRAyn4igEA/2.png new file mode 100644 index 0000000000000000000000000000000000000000..98d378f23f5b6b8872e1a2cebdbda647bb364dd7 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c0d9bd2d0509520be88e850ce25f5f5e8046b9096fe10e28c7911455454e722 +size 507262 diff --git a/vlm/dev/QeRAyn4igEA/3.png b/vlm/dev/QeRAyn4igEA/3.png new file mode 100644 index 0000000000000000000000000000000000000000..702e9a4173169eed531291677147bf892ebe3fd9 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3889ea008736c16605384c959df60728a11e5437dc0bb45ad25e9165b1e9ed0 +size 524057 diff --git a/vlm/dev/QeRAyn4igEA/4.png b/vlm/dev/QeRAyn4igEA/4.png new file mode 100644 index 0000000000000000000000000000000000000000..4022bb4d94a07aabb0f9794a5373e2e2f5a77ed1 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bd90a868be29e0ba8b0d582bb8c57c6049d501fcba839eb05233d8d6ea83502 +size 441936 diff --git a/vlm/dev/QeRAyn4igEA/5.png b/vlm/dev/QeRAyn4igEA/5.png new file mode 100644 index 0000000000000000000000000000000000000000..93aa173fe2a6947134ab8442d04daa73ef22c5f5 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1839c29ffc814cba8b67e9ef344407e2fd8f24c76dc3936c9224bf1756daf361 +size 564473 diff --git a/vlm/dev/QeRAyn4igEA/6.png b/vlm/dev/QeRAyn4igEA/6.png new file mode 100644 index 0000000000000000000000000000000000000000..67fc88e5458bb9892b738478b37da32a57fd398c --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95a8fc88acf13784d92a5f0ce45b92fda0e8dcbe3483f34fc9fa70d946dd5033 +size 487892 diff --git a/vlm/dev/QeRAyn4igEA/7.png b/vlm/dev/QeRAyn4igEA/7.png new file mode 100644 index 0000000000000000000000000000000000000000..5f5275186e45dbb511ec93d2bde48de789a6e197 --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:715a9dd91f9ccba70368c0653866cb08704f52c84070738087bb8ced0b83eff4 +size 595944 diff --git a/vlm/dev/QeRAyn4igEA/8.png b/vlm/dev/QeRAyn4igEA/8.png new file mode 100644 index 0000000000000000000000000000000000000000..2a0dfa6cfd1d7ec6a79ac9660b3d7c7e7c1f151c --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:975ed12aec1c55a1257712cb6c66110f85d8c4b9d009fc160559f1a77001f1b7 +size 501494 diff --git a/vlm/dev/QeRAyn4igEA/9.png b/vlm/dev/QeRAyn4igEA/9.png new file mode 100644 index 0000000000000000000000000000000000000000..fad52b04d890ab6e5620d4223113905d5f254e3e --- /dev/null +++ b/vlm/dev/QeRAyn4igEA/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86cde3ea541a5d52fa1c315aebeee2831d82776bbb7c4097a9482ff76d794d47 +size 543216 diff --git a/vlm/dev/RriDjddCLN/0.png b/vlm/dev/RriDjddCLN/0.png new file mode 100644 index 0000000000000000000000000000000000000000..de5e03328422273b7bc68f8440d0cf40b4f83d92 --- /dev/null +++ b/vlm/dev/RriDjddCLN/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afc938f765956466afd2d7bad8b80de36a322e106f2154504325bd0793b765d7 +size 486498 diff --git a/vlm/dev/RriDjddCLN/1.png b/vlm/dev/RriDjddCLN/1.png new file mode 100644 index 0000000000000000000000000000000000000000..21636a64f2b329c507e24df2aad1074878ba9b29 --- /dev/null +++ b/vlm/dev/RriDjddCLN/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f88a4ce56a1b8c1d9b3547b311357267820acd7e25a9e9e134641d1f4be2fee +size 590049 diff --git a/vlm/dev/RriDjddCLN/10.png b/vlm/dev/RriDjddCLN/10.png new file mode 100644 index 0000000000000000000000000000000000000000..80a39fc7e9624d49cdd99979dc7505e7623c7254 --- /dev/null +++ b/vlm/dev/RriDjddCLN/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1cb28b64183251764734b4d6f1a832f19da69ca4d1b74db9cce88def4e9c2aa +size 565992 diff --git a/vlm/dev/RriDjddCLN/11.png b/vlm/dev/RriDjddCLN/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ba744a3d6327b86f2ec76a36e28304e3949dcb98 --- /dev/null +++ b/vlm/dev/RriDjddCLN/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fe85a5acede962318bad9b2c9454b9360f9653a0f0c960773ab6f9f25f22d00 +size 603000 diff --git a/vlm/dev/RriDjddCLN/12.png b/vlm/dev/RriDjddCLN/12.png new file mode 100644 index 0000000000000000000000000000000000000000..17412d0fea9d3729c86e6bf90cece42fc17d8a72 --- /dev/null +++ b/vlm/dev/RriDjddCLN/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5d605075e939288607a8a80d6b4a279c1aa712a5d2f7bd8eb6850a838382980 +size 167915 diff --git a/vlm/dev/RriDjddCLN/2.png b/vlm/dev/RriDjddCLN/2.png new file mode 100644 index 0000000000000000000000000000000000000000..7fbf28a53c1d43fae796edfae434a19a9eedb364 --- /dev/null +++ b/vlm/dev/RriDjddCLN/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7f8ad98441fb50d64fd674dde87324fdfba607d4d9d7a8ed5190ca5e84fa485 +size 586945 diff --git a/vlm/dev/RriDjddCLN/3.png b/vlm/dev/RriDjddCLN/3.png new file mode 100644 index 0000000000000000000000000000000000000000..c4ea238eec090e5263d601a3ff9e065190de9b7b --- /dev/null +++ b/vlm/dev/RriDjddCLN/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de20fef3e62fa5735b5fe6ebda7949c6d87ef2856498d6d75ea96e1b84e31863 +size 494921 diff --git a/vlm/dev/RriDjddCLN/4.png b/vlm/dev/RriDjddCLN/4.png new file mode 100644 index 0000000000000000000000000000000000000000..0bacd444f20fbcc54aba62fa2b27f8bad97440fc --- /dev/null +++ b/vlm/dev/RriDjddCLN/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb76ee7b88802d7a064ee125d5ae45db9f9662e54a46508ff99daffd95cf8f9d +size 553301 diff --git a/vlm/dev/RriDjddCLN/5.png b/vlm/dev/RriDjddCLN/5.png new file mode 100644 index 0000000000000000000000000000000000000000..fc2d136c05edf14700599092406b57e507e59e99 --- /dev/null +++ b/vlm/dev/RriDjddCLN/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98fb59f076ab75dc567908903f48f8f282e605ffe39429e5077b22b9fdb793ab +size 475200 diff --git a/vlm/dev/RriDjddCLN/6.png b/vlm/dev/RriDjddCLN/6.png new file mode 100644 index 0000000000000000000000000000000000000000..24ebc54da2bb246635ae99d08aee3ccf5cf2d2a0 --- /dev/null +++ b/vlm/dev/RriDjddCLN/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29007603a26f1e4b3863a46327c93b7f947db3fbe6ba388ff88a61ea7a27a566 +size 571965 diff --git a/vlm/dev/RriDjddCLN/7.png b/vlm/dev/RriDjddCLN/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f61806708f61f3f34c4e4307dcc7ea5404fc704b --- /dev/null +++ b/vlm/dev/RriDjddCLN/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73d5c09cab1505a5446dcab40b70dc7b7cb670a67ad90f66542214d4a8f0eed4 +size 544187 diff --git a/vlm/dev/RriDjddCLN/8.png b/vlm/dev/RriDjddCLN/8.png new file mode 100644 index 0000000000000000000000000000000000000000..e0deb535130756857ec9950e1e300e0f33da7ae0 --- /dev/null +++ b/vlm/dev/RriDjddCLN/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:946d65bbc2262398dcda8d0f358719e174b02ccc6987bba24d1d0f435ff62ebc +size 673755 diff --git a/vlm/dev/RriDjddCLN/9.png b/vlm/dev/RriDjddCLN/9.png new file mode 100644 index 0000000000000000000000000000000000000000..56b4666cc548cc5edb61a4dd6f259fce4439b225 --- /dev/null +++ b/vlm/dev/RriDjddCLN/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80c9ae1fc349e88b7f772791d954575e9b2ab351212ebc38cea9860510054c42 +size 507035 diff --git a/vlm/dev/SfXjt1FtMQ/0.png b/vlm/dev/SfXjt1FtMQ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..1be0a9cd8a3d07dee0df1b036930e2ba7e026a7e --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69a5dd860daafe14e48ef072e1e88aed034a57b0c9133aea99b728b33c16f6c7 +size 429627 diff --git a/vlm/dev/SfXjt1FtMQ/1.png b/vlm/dev/SfXjt1FtMQ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..b49aa4fc805b8e47f9f187c74afdb2439b9b42db --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fea92dfd77120408f1025152d0b3aa4885966b7df877fd1434ff8425da57f92b +size 470238 diff --git a/vlm/dev/SfXjt1FtMQ/10.png b/vlm/dev/SfXjt1FtMQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..b8d64e4150a7cc2db9a5f38ddfd2f46252b74f00 --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:449e79a6940650e7e1194371f0dbecd4630d8e4a6527617d752c3b71de40ba33 +size 666565 diff --git a/vlm/dev/SfXjt1FtMQ/2.png b/vlm/dev/SfXjt1FtMQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..d48385b6bf07a617f34ef3336e12f44a3b05d4c0 --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dd4e02002f1c0157d7b09707c15acb8b362876b8f39cd28d2142f85af04b9da +size 356822 diff --git a/vlm/dev/SfXjt1FtMQ/3.png b/vlm/dev/SfXjt1FtMQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..a0be1aa0a2eb0549345a7be033756a8438013e82 --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d85e5ead99608f7e517c0d77c8dae89ab6c941f5e76cd08b682b4f5d0ae1fee2 +size 412604 diff --git a/vlm/dev/SfXjt1FtMQ/4.png b/vlm/dev/SfXjt1FtMQ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..526a8f0bbe91afa03477bc2e487e0ce5a6602997 --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4100e53e9e54f2e6d4a985f6731a81ee5c0429d14ddd78e356b6b9d9d9b34a61 +size 268130 diff --git a/vlm/dev/SfXjt1FtMQ/5.png b/vlm/dev/SfXjt1FtMQ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..a62147fade7325d6deecaf37fbb72d6a620f7c0f --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60ab96c5cdfd83ddbfb3cd9ee0d5860e6c9d584f0913cdc812954c86a77fcdb0 +size 366368 diff --git a/vlm/dev/SfXjt1FtMQ/6.png b/vlm/dev/SfXjt1FtMQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..a18524631e1fa9e3b50e5ac584d702c10a0e7928 --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cdd53a77f3454c8be704256d298bb1e9e52d6fb4bdce3123386a0680fa71148 +size 490339 diff --git a/vlm/dev/SfXjt1FtMQ/7.png b/vlm/dev/SfXjt1FtMQ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..5d9795f8770b2f6701f4e2419bc0195c804a8e81 --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f1c3220bc6e750a8cc470e1d884f87ddcfed7f602b160032abff172bc96bfe6 +size 528734 diff --git a/vlm/dev/SfXjt1FtMQ/8.png b/vlm/dev/SfXjt1FtMQ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..6922b3747bea70ecad7a88b0039a2de299174297 --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:799cf77573d27a4352cb995b663730b8a55d307656ba04b6d2bfbb4841ee92a5 +size 584270 diff --git a/vlm/dev/SfXjt1FtMQ/9.png b/vlm/dev/SfXjt1FtMQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0e1981504496422aee0920770212ed51305c45c8 --- /dev/null +++ b/vlm/dev/SfXjt1FtMQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35c3156276e1195ed86809ef0915dfb9933b92d3cb55d1b7be914cf30a4f0d37 +size 715621 diff --git a/vlm/dev/TJUNtiZiTKE/0.png b/vlm/dev/TJUNtiZiTKE/0.png new file mode 100644 index 0000000000000000000000000000000000000000..50e7085e4c7502843f739d581ca4b7b2b14a98ab --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:104dde108948151ee975b78058db01cd396a907edb5de737a3efc404b32ca7c9 +size 410394 diff --git a/vlm/dev/TJUNtiZiTKE/1.png b/vlm/dev/TJUNtiZiTKE/1.png new file mode 100644 index 0000000000000000000000000000000000000000..49af2879742369498b7d5400d48de1570b6ca4f5 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a56ad5668791fffbdf308d7a1ffdd5cf58273b930fdf7f97ef84273795ef0620 +size 589420 diff --git a/vlm/dev/TJUNtiZiTKE/10.png b/vlm/dev/TJUNtiZiTKE/10.png new file mode 100644 index 0000000000000000000000000000000000000000..82b725cedc498c3cd64a6cae8df7ddcae6a90965 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0249e44fa0ee7821bb7c05fad64eebf260247706da66ba51d776fbb96571d04d +size 554139 diff --git a/vlm/dev/TJUNtiZiTKE/11.png b/vlm/dev/TJUNtiZiTKE/11.png new file mode 100644 index 0000000000000000000000000000000000000000..30c3f053deae989407c96b676e6d3f88c08b5361 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d59df42a12a3708b59b8e566a832a2f9e222e3828ffaee9b0b834f91b4dc930b +size 538136 diff --git a/vlm/dev/TJUNtiZiTKE/12.png b/vlm/dev/TJUNtiZiTKE/12.png new file mode 100644 index 0000000000000000000000000000000000000000..dc9952db167e78503bdacf6da4851b5f102729df --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:463a8577bf33c4bb4475261a39c485c00a7aeb4a6d66f827bb616a3a12dac49b +size 514225 diff --git a/vlm/dev/TJUNtiZiTKE/2.png b/vlm/dev/TJUNtiZiTKE/2.png new file mode 100644 index 0000000000000000000000000000000000000000..98c24fa56026fc7ff322ad36618adde3c770c2dc --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a280e0b9a3bd7580a4d61e86cf2cc1abe2f367b35aa256c43627173ce896b58 +size 479577 diff --git a/vlm/dev/TJUNtiZiTKE/3.png b/vlm/dev/TJUNtiZiTKE/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b89746f9697480e3f90cd07079b76fd907e24f46 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0aa885eb38942739effbfaad71a379722d8828ad60fb7d2865979d50577ea124 +size 564137 diff --git a/vlm/dev/TJUNtiZiTKE/4.png b/vlm/dev/TJUNtiZiTKE/4.png new file mode 100644 index 0000000000000000000000000000000000000000..76dc51b12407f1469a510c107da8baca12aba745 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa49e390cbae88a4448ac74fef90a88fdcdae6ee9ba952d3fe8afeb6e4cbbc4e +size 497107 diff --git a/vlm/dev/TJUNtiZiTKE/5.png b/vlm/dev/TJUNtiZiTKE/5.png new file mode 100644 index 0000000000000000000000000000000000000000..c04f89cf54ab9c4b62075e2f392966fa9c947886 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b7e4a65aed7518e19eb85d80ecc47756ef512d2db3d96da49cfa328a4a1ad97 +size 530389 diff --git a/vlm/dev/TJUNtiZiTKE/6.png b/vlm/dev/TJUNtiZiTKE/6.png new file mode 100644 index 0000000000000000000000000000000000000000..9a65087b5a61d8b087409a8059845db8075ea3c2 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87a4734c4d4683090c758b6a5626576330a331644b946eefe386d1e043e05893 +size 656031 diff --git a/vlm/dev/TJUNtiZiTKE/7.png b/vlm/dev/TJUNtiZiTKE/7.png new file mode 100644 index 0000000000000000000000000000000000000000..0357fea0df5d8a7245ba16e45541d5556b45e40e --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ea2a798249ee06557ce0ba6efbd7fc01cbebf996bb337692815e50acaef15b9 +size 647391 diff --git a/vlm/dev/TJUNtiZiTKE/8.png b/vlm/dev/TJUNtiZiTKE/8.png new file mode 100644 index 0000000000000000000000000000000000000000..30b2d8cf46b0ec93ce66888dafd7c9283fafad10 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6ed00d33c83541f67cbf8297fdb14e67cb4a886d810fbf590370fba76006414 +size 637270 diff --git a/vlm/dev/TJUNtiZiTKE/9.png b/vlm/dev/TJUNtiZiTKE/9.png new file mode 100644 index 0000000000000000000000000000000000000000..8faa61caa65d065d4da4d8c9beea8a57529941a1 --- /dev/null +++ b/vlm/dev/TJUNtiZiTKE/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1850c41699a18c2c8b5f3e401ebbd200f73ddb86b94be71e1d8d8606ee60179a +size 89706 diff --git a/vlm/dev/UPnJuDKqOfX/0.png b/vlm/dev/UPnJuDKqOfX/0.png new file mode 100644 index 0000000000000000000000000000000000000000..3bce751e8c88873074f3c67582ca27aba6d26b7a --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8da9bcc395cda4c17082368b3f5ccddca5c85df9e47c273313ca78c973ce1368 +size 448822 diff --git a/vlm/dev/UPnJuDKqOfX/1.png b/vlm/dev/UPnJuDKqOfX/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a964195a07aded743ed7a2f86b881040d7b03a96 --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffc9d14bde5585f722d8b6dd6679d9f29212b243541e3272cb1c20500c6d4144 +size 964946 diff --git a/vlm/dev/UPnJuDKqOfX/10.png b/vlm/dev/UPnJuDKqOfX/10.png new file mode 100644 index 0000000000000000000000000000000000000000..6bbe7b6dbb9dc5971d7dea5fd8face6b185ead07 --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e9c5fe03cfbb71d10be38fa373ed07619fe8e6f6eeeeb8a16ff579af7408f20 +size 529311 diff --git a/vlm/dev/UPnJuDKqOfX/11.png b/vlm/dev/UPnJuDKqOfX/11.png new file mode 100644 index 0000000000000000000000000000000000000000..f4afd904157ad70f5e88bc6c4d25a09cfbf0560e --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f50ae3b80e3955ccaa86697a1068e626350a434ae4a06c10168c0ef4a7fd729 +size 518558 diff --git a/vlm/dev/UPnJuDKqOfX/12.png b/vlm/dev/UPnJuDKqOfX/12.png new file mode 100644 index 0000000000000000000000000000000000000000..f5672dbd64ee7140f61ffe9671c00f4e6945765b --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe93862d83efbf0617ee0124e09411544a24e8010ac94f56452d1848aed231e0 +size 254417 diff --git a/vlm/dev/UPnJuDKqOfX/2.png b/vlm/dev/UPnJuDKqOfX/2.png new file mode 100644 index 0000000000000000000000000000000000000000..5aa5abb005f76f55cfb27c63a7d7423c2f9854d2 --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6285a696e4fb58016f9d3f61389cddf33b63fd26e54d52b2e7759cac2cc1ebe3 +size 614471 diff --git a/vlm/dev/UPnJuDKqOfX/3.png b/vlm/dev/UPnJuDKqOfX/3.png new file mode 100644 index 0000000000000000000000000000000000000000..ddcd3e55c15d4814d8a10938cb09c6f2d96ba29c --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62b7ff6d4eaba5a3da15e8d97a1d6537ff07fdcfd2a16487a0016c1601b51ee4 +size 481248 diff --git a/vlm/dev/UPnJuDKqOfX/4.png b/vlm/dev/UPnJuDKqOfX/4.png new file mode 100644 index 0000000000000000000000000000000000000000..8f5b08be985517687d4d298e7af3d80c30854a39 --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:605b5ac9de707bf3692be24942cfd3de9bd320ff7abced2fa3e4d3942cdb7d61 +size 741702 diff --git a/vlm/dev/UPnJuDKqOfX/5.png b/vlm/dev/UPnJuDKqOfX/5.png new file mode 100644 index 0000000000000000000000000000000000000000..4a25ce43fc842541db13f15fa7c198363f31d739 --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fe0993aed432809179fa9add172582cc235a2287e3b383caef31e7f07d79073 +size 513313 diff --git a/vlm/dev/UPnJuDKqOfX/6.png b/vlm/dev/UPnJuDKqOfX/6.png new file mode 100644 index 0000000000000000000000000000000000000000..5115b5eb8165d611c55be4b5141a81418e2eba13 --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cec86b0beeedf7b085fd31b78c4e471078d487cbae20bc94f9af7f6ef38c27a3 +size 464604 diff --git a/vlm/dev/UPnJuDKqOfX/7.png b/vlm/dev/UPnJuDKqOfX/7.png new file mode 100644 index 0000000000000000000000000000000000000000..2eef9734532a6dded7403a2b93ebf28afceaefa4 --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1f37d15d134ec77cb18f92bf9df58358889892abbb5b6bdd488360d69583e34 +size 923684 diff --git a/vlm/dev/UPnJuDKqOfX/8.png b/vlm/dev/UPnJuDKqOfX/8.png new file mode 100644 index 0000000000000000000000000000000000000000..79ea60c0edf92506cb42ca1c9eb58e265739d3ea --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e46ca74c5cbcc3dae010fc85029d7ebd742f3b1e0b4461401f877a7880b83a6b +size 570535 diff --git a/vlm/dev/UPnJuDKqOfX/9.png b/vlm/dev/UPnJuDKqOfX/9.png new file mode 100644 index 0000000000000000000000000000000000000000..026420dbf25f9b4c10db2ed6779432db32fef057 --- /dev/null +++ b/vlm/dev/UPnJuDKqOfX/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1523b41a491f4f9fd137407d6eb5ae8d24ca05a869dbdcd9da4a3bf55aff3107 +size 767188 diff --git a/vlm/dev/UVtVRcurOYv/0.png b/vlm/dev/UVtVRcurOYv/0.png new file mode 100644 index 0000000000000000000000000000000000000000..fec84a8a69c0097fba13b298a0ea573e89899114 --- /dev/null +++ b/vlm/dev/UVtVRcurOYv/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2b661dcb5d7dd0a3ef2aadcc5aeb77cf828c34dca36d53a73f3c71885a4ce5f +size 473467 diff --git a/vlm/dev/UVtVRcurOYv/1.png b/vlm/dev/UVtVRcurOYv/1.png new file mode 100644 index 0000000000000000000000000000000000000000..012fc8395d558a0c15a53868961a241e5efae14e --- /dev/null +++ b/vlm/dev/UVtVRcurOYv/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8977d6ed3b61507898f539f117414c52d26c7a1c87052511eef35fde98df9be3 +size 547150 diff --git a/vlm/dev/UVtVRcurOYv/2.png b/vlm/dev/UVtVRcurOYv/2.png new file mode 100644 index 0000000000000000000000000000000000000000..546dc3779918c1f9ba2225ed7229d2a49b1fc3e2 --- /dev/null +++ b/vlm/dev/UVtVRcurOYv/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:803342d7c36aec72a84523ad7c59d2e7cc7bd926e0e70b73e65535092ae4be3a +size 392457 diff --git a/vlm/dev/UVtVRcurOYv/3.png b/vlm/dev/UVtVRcurOYv/3.png new file mode 100644 index 0000000000000000000000000000000000000000..4aaa93c5f4c89a27fdc85333f93ab1d113dca5a8 --- /dev/null +++ b/vlm/dev/UVtVRcurOYv/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d9d23f8f94f6b586177f77b7c488c922586856e3dc97e252014adc7acf9323f +size 524160 diff --git a/vlm/dev/UVtVRcurOYv/4.png b/vlm/dev/UVtVRcurOYv/4.png new file mode 100644 index 0000000000000000000000000000000000000000..9829f0a23e85d537f56b08baf47cbd109002eadc --- /dev/null +++ b/vlm/dev/UVtVRcurOYv/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81ee7ccd5f2db490e998fd5d84b367d8ec931b7368caaf19a1d82ecd4ab5191a +size 411775 diff --git a/vlm/dev/UVtVRcurOYv/5.png b/vlm/dev/UVtVRcurOYv/5.png new file mode 100644 index 0000000000000000000000000000000000000000..4a718d70beb088728193667475ac72bb78ca79c7 --- /dev/null +++ b/vlm/dev/UVtVRcurOYv/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:874cf5d1801b9684f6054f6998065d0a563abf607af6d1a4a5efcb3251f33e93 +size 387240 diff --git a/vlm/dev/UVtVRcurOYv/6.png b/vlm/dev/UVtVRcurOYv/6.png new file mode 100644 index 0000000000000000000000000000000000000000..9d2007a671b63689f7298ffd2ace615520344e4b --- /dev/null +++ b/vlm/dev/UVtVRcurOYv/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39998f788be8d1afb27458d7873bf81bd36e71cfefca9c5c85337a575bfca940 +size 548815 diff --git a/vlm/dev/UVtVRcurOYv/7.png b/vlm/dev/UVtVRcurOYv/7.png new file mode 100644 index 0000000000000000000000000000000000000000..75f1849e08fa88e15e5b1d89ad0944505ec92590 --- /dev/null +++ b/vlm/dev/UVtVRcurOYv/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af8e9edd30f239a4ac3eff50e71ffc32820972b791250c90adfda93e15a6bc7e +size 427268 diff --git a/vlm/dev/Vu-B0clPfq/0.png b/vlm/dev/Vu-B0clPfq/0.png new file mode 100644 index 0000000000000000000000000000000000000000..9c62d26f72219e89449a507c8e735601a37c7556 --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f096cf0f690281b802227118c2951d638143ebbabefb3f9eb1b671e22c5faba8 +size 441214 diff --git a/vlm/dev/Vu-B0clPfq/1.png b/vlm/dev/Vu-B0clPfq/1.png new file mode 100644 index 0000000000000000000000000000000000000000..2ef21c3b53ed680b3604553927b6e393f9eba113 --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69cfdf443d549731a9b307c9107a111ab479a4e3269972ae377e423f1827d9b7 +size 510073 diff --git a/vlm/dev/Vu-B0clPfq/10.png b/vlm/dev/Vu-B0clPfq/10.png new file mode 100644 index 0000000000000000000000000000000000000000..3bbff08ce264fc73ff61c23e739e8e48a1fcce46 --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dffd67ca4d3929616ffb25fdf768c45260d5cecdf3bd98f43e130d2f4d201414 +size 592460 diff --git a/vlm/dev/Vu-B0clPfq/11.png b/vlm/dev/Vu-B0clPfq/11.png new file mode 100644 index 0000000000000000000000000000000000000000..95c44e89df53a7f8075f2fc9387a522c3e90318d --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:422ce367b5dd5eb0d3fa133f99ef893f2fbca9e8838f887336b79dd772800d6f +size 432761 diff --git a/vlm/dev/Vu-B0clPfq/12.png b/vlm/dev/Vu-B0clPfq/12.png new file mode 100644 index 0000000000000000000000000000000000000000..482837fafd57f89ce314885137ff6873fa77fcaf --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db44ec632079c1e9f496b0fbe0c94527f2d56452f9b5667b230d580c8d0ff087 +size 425649 diff --git a/vlm/dev/Vu-B0clPfq/2.png b/vlm/dev/Vu-B0clPfq/2.png new file mode 100644 index 0000000000000000000000000000000000000000..df739a22dd5668a64a494ff5554e3667caf5ef42 --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ab0f34fbf1edc9086be1931d7637a2ca7b69c60d6589b9afceb810a85346f8d +size 581384 diff --git a/vlm/dev/Vu-B0clPfq/3.png b/vlm/dev/Vu-B0clPfq/3.png new file mode 100644 index 0000000000000000000000000000000000000000..51bb6f0b33d2f894a268e7380fe7f7fac7b5a00d --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54e860b0780c2864caa38f5db853724e6b3d48a61def59fbe074c2564510252d +size 494052 diff --git a/vlm/dev/Vu-B0clPfq/4.png b/vlm/dev/Vu-B0clPfq/4.png new file mode 100644 index 0000000000000000000000000000000000000000..762db617bf1fa559afbc6aa776d32076f68d9d5f --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55f7127ec24ca4d637bdb9aa27f585302a7b0f1ef79aeded7ca7aa37c481c9d2 +size 480762 diff --git a/vlm/dev/Vu-B0clPfq/5.png b/vlm/dev/Vu-B0clPfq/5.png new file mode 100644 index 0000000000000000000000000000000000000000..2686fb8a707321127a33e6cab4813859b1fbba56 --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:399e0b465cf68cbf643ba5d2f9e17dac69ad8e419655781eefe8f65977bc44e3 +size 603020 diff --git a/vlm/dev/Vu-B0clPfq/6.png b/vlm/dev/Vu-B0clPfq/6.png new file mode 100644 index 0000000000000000000000000000000000000000..7f34a26b05d219c8d3d6e94a0e6b8b91ba9036e7 --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:954a59fc157d08587ecae492c63f7a6f35b8d5be276c683a135ad9e191a2ceab +size 561741 diff --git a/vlm/dev/Vu-B0clPfq/7.png b/vlm/dev/Vu-B0clPfq/7.png new file mode 100644 index 0000000000000000000000000000000000000000..808abcfa4674ee27281fb66feca9b61b72f019e6 --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6fc56c2ff0e400785c1d98bbe0db9be6b4d54eaa5ba3821323f0a775270c645 +size 532424 diff --git a/vlm/dev/Vu-B0clPfq/8.png b/vlm/dev/Vu-B0clPfq/8.png new file mode 100644 index 0000000000000000000000000000000000000000..061c482eac065cda28c8d605e0eba28576ce5dfc --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e914c4efdde418252ce21bb9a8c103ece8af05aa45b05a46e758d15d83452c3 +size 445932 diff --git a/vlm/dev/Vu-B0clPfq/9.png b/vlm/dev/Vu-B0clPfq/9.png new file mode 100644 index 0000000000000000000000000000000000000000..b4047fabb2e94a5cbf5cdcc96cf6914ed666c6d7 --- /dev/null +++ b/vlm/dev/Vu-B0clPfq/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b68089d10e32b71671bb86b55b71cbe63a46464c1f10010c173c0f9320bbc3e4 +size 525824 diff --git a/vlm/dev/WSIHedvwmru/0.png b/vlm/dev/WSIHedvwmru/0.png new file mode 100644 index 0000000000000000000000000000000000000000..824c4315bb108da52bbdfa4b8dd71b9be641657c --- /dev/null +++ b/vlm/dev/WSIHedvwmru/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:401d6179d50a241fdc9009abfec99b0c9695b894fcaf8f143a3d144a8dc82d30 +size 509843 diff --git a/vlm/dev/WSIHedvwmru/1.png b/vlm/dev/WSIHedvwmru/1.png new file mode 100644 index 0000000000000000000000000000000000000000..f05e423df1daee0bad4cfd5d68501dada13c504c --- /dev/null +++ b/vlm/dev/WSIHedvwmru/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62d9509b22b4bec883d42b4444aa2a887af36ec6d69c64e26fc2676553b9c57a +size 577362 diff --git a/vlm/dev/WSIHedvwmru/10.png b/vlm/dev/WSIHedvwmru/10.png new file mode 100644 index 0000000000000000000000000000000000000000..3801667e621857db793ad5f81f903b8a4085df19 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6592cce5496f33a4b33bbac86371c312ebbae81dc70d8c105519a4b965492a9e +size 588612 diff --git a/vlm/dev/WSIHedvwmru/11.png b/vlm/dev/WSIHedvwmru/11.png new file mode 100644 index 0000000000000000000000000000000000000000..8b04a314ebe9500cda87d1686ce0106004ce1099 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b3158027aea3043cfa9cda754adcb8766d91c90a148b6e231dc558b0b5b4db7 +size 583561 diff --git a/vlm/dev/WSIHedvwmru/12.png b/vlm/dev/WSIHedvwmru/12.png new file mode 100644 index 0000000000000000000000000000000000000000..5c70eaa2afa98eefafdc5a749079c77b00382b32 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a08e9d1dbc521864ce04faef6f8478ffdd803453b06122f448bb947a5439e97f +size 185026 diff --git a/vlm/dev/WSIHedvwmru/13.png b/vlm/dev/WSIHedvwmru/13.png new file mode 100644 index 0000000000000000000000000000000000000000..c075b29ef90a986a9cd2f4235066b34972bd1adb --- /dev/null +++ b/vlm/dev/WSIHedvwmru/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5bb4fcba0d43c1e821684f8193bd06b95baf2bb0f3addc0b9c190178389af7a +size 617488 diff --git a/vlm/dev/WSIHedvwmru/14.png b/vlm/dev/WSIHedvwmru/14.png new file mode 100644 index 0000000000000000000000000000000000000000..2f4bfa564ca5e715d0aac1452fcd10115ec1c8e2 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdc628600f15ab12ad3f45a81e6a729387502b6c16e3831a493cd279de07d184 +size 414978 diff --git a/vlm/dev/WSIHedvwmru/15.png b/vlm/dev/WSIHedvwmru/15.png new file mode 100644 index 0000000000000000000000000000000000000000..b8bb7bca5f7c53bdb5e2dea34ee991ec26094d62 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0836a8845ba01d26adc278e2c62d723c902b05afed2ccc1465c803001178cd60 +size 454845 diff --git a/vlm/dev/WSIHedvwmru/16.png b/vlm/dev/WSIHedvwmru/16.png new file mode 100644 index 0000000000000000000000000000000000000000..767f1e017b1c094ff810d7e6d93d9a044deb47ef --- /dev/null +++ b/vlm/dev/WSIHedvwmru/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8e2735d611eb5d800bc4b03dfbb1765cf7bac5782592a99108b66865581f393 +size 566069 diff --git a/vlm/dev/WSIHedvwmru/2.png b/vlm/dev/WSIHedvwmru/2.png new file mode 100644 index 0000000000000000000000000000000000000000..e2c2bc7de74f73bd69f93cdba0fb1dfa8c35a375 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f744be6450fd3025fc99ad152baa6643985f468fa3eaa675b4d30f302a992082 +size 600510 diff --git a/vlm/dev/WSIHedvwmru/3.png b/vlm/dev/WSIHedvwmru/3.png new file mode 100644 index 0000000000000000000000000000000000000000..1191034d9ab764bb58917081692f648424eb737f --- /dev/null +++ b/vlm/dev/WSIHedvwmru/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b03077483e9351aa6a920db0a03fc5f4fe60150e72147bd55c19a84d5420f3f +size 593370 diff --git a/vlm/dev/WSIHedvwmru/4.png b/vlm/dev/WSIHedvwmru/4.png new file mode 100644 index 0000000000000000000000000000000000000000..02353c0fb05593587b7793ba77a427818136d914 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b22de00449dc3d633fb44aafe30343a445758527a5b5cad6e8ce584f49f67dd +size 1193403 diff --git a/vlm/dev/WSIHedvwmru/5.png b/vlm/dev/WSIHedvwmru/5.png new file mode 100644 index 0000000000000000000000000000000000000000..62da7cd4e71555454f4088783eb754909272d117 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8e1eab16603936f77c8602d656fbea475b97252d162fb340213900e711eab6a +size 586464 diff --git a/vlm/dev/WSIHedvwmru/6.png b/vlm/dev/WSIHedvwmru/6.png new file mode 100644 index 0000000000000000000000000000000000000000..2e6b102968abb911b240b81c1a75becaf6c30772 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d75e14efcde96ddc9c6e58b6daf0138966d14be5fc28d3b544e50e927586ade +size 675877 diff --git a/vlm/dev/WSIHedvwmru/7.png b/vlm/dev/WSIHedvwmru/7.png new file mode 100644 index 0000000000000000000000000000000000000000..17095c640f3c1a907029a122253d90dc67533256 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:421f197cfab16a313bf03789cd2dfb1c5add38d395a890f937d58973795ed8d5 +size 641476 diff --git a/vlm/dev/WSIHedvwmru/8.png b/vlm/dev/WSIHedvwmru/8.png new file mode 100644 index 0000000000000000000000000000000000000000..99ef3abf618d67ff203735d66ec60668d69cd3b1 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4486d7ab1cfd54a76f0fd73911421ccf88b64f2d4ee14e883b9de41001ade623 +size 598390 diff --git a/vlm/dev/WSIHedvwmru/9.png b/vlm/dev/WSIHedvwmru/9.png new file mode 100644 index 0000000000000000000000000000000000000000..8869326f7ef610c33d2068a7f355e92d97e03f75 --- /dev/null +++ b/vlm/dev/WSIHedvwmru/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:367cfeb35562315e9061d42af7c6d5278244f78e062255a06d2561268fb4968e +size 550522 diff --git a/vlm/dev/cZAi1yWpiXQ/0.png b/vlm/dev/cZAi1yWpiXQ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..99fe02eef6c222b70c745a832125528c7b4bb9c0 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:063eace13addb78663df9f0058ec120a6c40754b056dd44cf8263452600430bb +size 498543 diff --git a/vlm/dev/cZAi1yWpiXQ/1.png b/vlm/dev/cZAi1yWpiXQ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..1ec90e76e92dac8256d411f7c6d2fa9d6ba0c50d --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5ef1722c0629f14d16f4d465987d46c53789c10923868354ffb1e6b2f8ed188 +size 600741 diff --git a/vlm/dev/cZAi1yWpiXQ/10.png b/vlm/dev/cZAi1yWpiXQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..9dc8be6464498ac334f0faf8cb9ea232d8fc7b6b --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9737a6c0720735678db3294fd8b80bc645ac10f109784bca21b5da7c29a8b677 +size 630119 diff --git a/vlm/dev/cZAi1yWpiXQ/11.png b/vlm/dev/cZAi1yWpiXQ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..22dda80763037661028738ddbe7c822e76d048b1 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e45766ff86d371ddffe506de2f29c9087e258306ccf5142a5f61d95e7adbe65 +size 656176 diff --git a/vlm/dev/cZAi1yWpiXQ/12.png b/vlm/dev/cZAi1yWpiXQ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..e16febd8a3de2ab554393361bf7dc5cd2989b40d --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aacd6287a9ab4bbcde90edf7a1d1cc75dadbb8efca00d1f516e6fc95597ab613 +size 673223 diff --git a/vlm/dev/cZAi1yWpiXQ/13.png b/vlm/dev/cZAi1yWpiXQ/13.png new file mode 100644 index 0000000000000000000000000000000000000000..fb6fa7c60ffed62a79b0720c3f28cf83f191eb76 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad6e56c89f1502009389c9553b8259b42ad6d1903e4eb0075819c9378089c691 +size 195471 diff --git a/vlm/dev/cZAi1yWpiXQ/14.png b/vlm/dev/cZAi1yWpiXQ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..aedeaef4bf619af261549ca2fb05c8c2abffa59f --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b12e4e2f54faeb637cdef8665ea194b1ccea941088857af9061493267e408dae +size 466090 diff --git a/vlm/dev/cZAi1yWpiXQ/15.png b/vlm/dev/cZAi1yWpiXQ/15.png new file mode 100644 index 0000000000000000000000000000000000000000..d90d184bf8b5eb2b0dbbf7d14db545b673e771d9 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9f16c1c8a93a1ddbcc99919525eedf2171f4ce09abe00eb9d24d2916b985099 +size 350720 diff --git a/vlm/dev/cZAi1yWpiXQ/16.png b/vlm/dev/cZAi1yWpiXQ/16.png new file mode 100644 index 0000000000000000000000000000000000000000..e242035ea0a69c1bf838c106301f1432c905d625 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15a4058edbf09f2a2d43c2eef363fda8a7de330213498025dcb55195a6fe94c6 +size 541941 diff --git a/vlm/dev/cZAi1yWpiXQ/17.png b/vlm/dev/cZAi1yWpiXQ/17.png new file mode 100644 index 0000000000000000000000000000000000000000..1b117decf43548b47cb8525b1b01c8268c107859 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3014cc51f1199bc1b6079e7f52bebd1fdc11833c4521128ce5e7f21b2180cd2a +size 422472 diff --git a/vlm/dev/cZAi1yWpiXQ/18.png b/vlm/dev/cZAi1yWpiXQ/18.png new file mode 100644 index 0000000000000000000000000000000000000000..3ef110402014cd4d859f0666270ea76063b89c35 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2381e2492bc4468d65e1f030b6d98277dd87d0f5f5d1ab4cc7fac607a594b1f +size 554639 diff --git a/vlm/dev/cZAi1yWpiXQ/19.png b/vlm/dev/cZAi1yWpiXQ/19.png new file mode 100644 index 0000000000000000000000000000000000000000..a103cc921d82773ff21081d1c4769f9c5917ed7e --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:771810fdd290f75e609b0e9a9162090dc74d73cdf86e2ed50a725b40a95979c9 +size 486409 diff --git a/vlm/dev/cZAi1yWpiXQ/2.png b/vlm/dev/cZAi1yWpiXQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..28a820f02618a1fe298ce93d33845d729f9c8ce4 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4f20e27403d8b3dc1a0d4ad7f7aba573975577c249416f2fc70a40f2e8250a7 +size 662397 diff --git a/vlm/dev/cZAi1yWpiXQ/3.png b/vlm/dev/cZAi1yWpiXQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..3f5a28191332ff1f4f80ff9c21259da63d05e02e --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:149d3166ab1e5177ec5bf7f5409c3f4a1e6175a9a997cf660d0f959f93f3a900 +size 545724 diff --git a/vlm/dev/cZAi1yWpiXQ/4.png b/vlm/dev/cZAi1yWpiXQ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..e04474eb61d3ff0b6a93c207d0d5f072d242840a --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:100067037180c7757c4756f9496897356aadd08c5838f1b5fd6b02bc3393f353 +size 552735 diff --git a/vlm/dev/cZAi1yWpiXQ/5.png b/vlm/dev/cZAi1yWpiXQ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..d0f3155a76f1d22ab3baa5eeb1dec9be52a3bda7 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a689c51f6e840bf04a70016e5ba0240c01a7ea090622168c480c722edd43c79b +size 586894 diff --git a/vlm/dev/cZAi1yWpiXQ/6.png b/vlm/dev/cZAi1yWpiXQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..6a3b24cab0308db58bc8a24f60ffeb97fa18fe31 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b7fdca08bd97ed56138159758d397b1651c15be8d355921addfd7cf9109bb3f +size 607235 diff --git a/vlm/dev/cZAi1yWpiXQ/7.png b/vlm/dev/cZAi1yWpiXQ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..423c2678eb46c9b5e4b090a9847e6c51c4822d02 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:067e54be26638115268e95319420e850f1ce5840449ef49fc67db0cf55e6e88e +size 671645 diff --git a/vlm/dev/cZAi1yWpiXQ/8.png b/vlm/dev/cZAi1yWpiXQ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..ed5b182eb00dab9bdfe9932ba3abe201406c9fbd --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9e89c690a6607deb6d51f593702a4f1d36d38a32348aeb5b05f1e6f3828a66e +size 623923 diff --git a/vlm/dev/cZAi1yWpiXQ/9.png b/vlm/dev/cZAi1yWpiXQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..6577a010f3a80069213888e59affed88006dad92 --- /dev/null +++ b/vlm/dev/cZAi1yWpiXQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00dcd0ec8ee1e90d1a988e76adce92d0128022f6141a92b6332360807bb6a327 +size 199506 diff --git a/vlm/dev/gSyjaunurQ/0.png b/vlm/dev/gSyjaunurQ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c8224e8eb7a94ed6e17f670c33d440cf0ef1a6c3 --- /dev/null +++ b/vlm/dev/gSyjaunurQ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd054b59edd508bfd8a889f056114c1bdef8cf7d49923ac25cd414038dfcf4b9 +size 412164 diff --git a/vlm/dev/gSyjaunurQ/1.png b/vlm/dev/gSyjaunurQ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0cdc0f18e0f6456df8f78221c536320356a5989a --- /dev/null +++ b/vlm/dev/gSyjaunurQ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:766fef6a3ce98fec70a2471c43a968a615f4f588bd084880f63a5ba02af0d2c0 +size 498109 diff --git a/vlm/dev/gSyjaunurQ/10.png b/vlm/dev/gSyjaunurQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..701b71b08046c14c2586fafa44baa50730adb733 --- /dev/null +++ b/vlm/dev/gSyjaunurQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fef74de80164d328a139ded326c220233bae3e4ae2a55de771716b48462e747 +size 297342 diff --git a/vlm/dev/gSyjaunurQ/2.png b/vlm/dev/gSyjaunurQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..32dd78ea3e6612287c04985aaa70ea00666cab24 --- /dev/null +++ b/vlm/dev/gSyjaunurQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3659fd5c62b2a419a7840d259d3b1474e681dded33befb55e4ce92830d1189f +size 455906 diff --git a/vlm/dev/gSyjaunurQ/3.png b/vlm/dev/gSyjaunurQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e261ed6aa2ca05cfaa94d2c6936fac18224c800f --- /dev/null +++ b/vlm/dev/gSyjaunurQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02a57987fe8bb05420ce71819bfa9bcf0cd2874b0fd0395bb365340f365fe5d5 +size 487565 diff --git a/vlm/dev/gSyjaunurQ/4.png b/vlm/dev/gSyjaunurQ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..060965dabe17cf0e7c3c20d4645fde3a2ff4bdb6 --- /dev/null +++ b/vlm/dev/gSyjaunurQ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cab668379718b224926664e4009a622f204a820a537d4d8869098e2d9c171a6 +size 683066 diff --git a/vlm/dev/gSyjaunurQ/5.png b/vlm/dev/gSyjaunurQ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..68ebd408eeb8e1dd47862e490351a75a492b65ff --- /dev/null +++ b/vlm/dev/gSyjaunurQ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:636a1eed1e85131a761ca9139e06689ed1c03e903f5f5736d321290340d0baca +size 488443 diff --git a/vlm/dev/gSyjaunurQ/6.png b/vlm/dev/gSyjaunurQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..cba543aaf7e20101cb48bc49243e208023ce1fd1 --- /dev/null +++ b/vlm/dev/gSyjaunurQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52a5006abbe1f9af44abc8f32df4e84a9c9055358b43276ad36fb8b10058ace8 +size 452869 diff --git a/vlm/dev/gSyjaunurQ/7.png b/vlm/dev/gSyjaunurQ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f0bb51b4a661ce6764434481b9032c53d415c879 --- /dev/null +++ b/vlm/dev/gSyjaunurQ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:121682407674db1867f5d4bd26f40ed684e18476efb037f2ddf93cfeca54dae6 +size 371331 diff --git a/vlm/dev/gSyjaunurQ/8.png b/vlm/dev/gSyjaunurQ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..c6964a09968cb461e4a259b638be223a46cd59b0 --- /dev/null +++ b/vlm/dev/gSyjaunurQ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb8b7e626a6eac34df8980d31ddbb328566afccc3101a5d5a5b9e44e519f8340 +size 1014662 diff --git a/vlm/dev/gSyjaunurQ/9.png b/vlm/dev/gSyjaunurQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..a87c03b1d2ff9d6c40b31dc5026b47d2755a3c3b --- /dev/null +++ b/vlm/dev/gSyjaunurQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fecfee2427fee4f5d96fde1c5cbb6080d9a17fbc09e973426fefdf906c9eb8e9 +size 534683 diff --git a/vlm/dev/jDIlzSU8wJ/0.png b/vlm/dev/jDIlzSU8wJ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..3e088c96c5de5249288cae2115e39bd9cb2b47ea --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d3dab1c3913e38e0469200de35752657c32bf66e2b96121fc7c346da77d7212 +size 448759 diff --git a/vlm/dev/jDIlzSU8wJ/1.png b/vlm/dev/jDIlzSU8wJ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..e60859ae1e9ffab8ccaa22340797ef80f201e9a2 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:060e0c6e2868faf911b1d3548ab608739b7b439ff59d4e033e39adcfb58faf64 +size 782799 diff --git a/vlm/dev/jDIlzSU8wJ/10.png b/vlm/dev/jDIlzSU8wJ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..f301dc410091b443f46786846ea99a1cbb1310ed --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4538a1ecff5b9ff723fa1ad9f510429099b3635a7b64c66558d94163723a039b +size 509734 diff --git a/vlm/dev/jDIlzSU8wJ/11.png b/vlm/dev/jDIlzSU8wJ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..53a7b6495c4b8b2f3031f89df5be50f39c5423d3 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1dab77e501e81d002db52bab46906f867a286ba71879ee107f105c1738fb589 +size 519199 diff --git a/vlm/dev/jDIlzSU8wJ/12.png b/vlm/dev/jDIlzSU8wJ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..7ecba178e91f14c13b20c8be3cb5a490dee3a638 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3f917989f241588981f284624c1ebc1ff34839d668e0b15b240d73c77d00663 +size 510540 diff --git a/vlm/dev/jDIlzSU8wJ/13.png b/vlm/dev/jDIlzSU8wJ/13.png new file mode 100644 index 0000000000000000000000000000000000000000..6bc7e3878df4b3309406d491d7d6afe14fa65f66 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaa8c58a6ebbb59c175c47529eccecb239ea1b53ba53a0e77be19eef1d09d0e4 +size 513602 diff --git a/vlm/dev/jDIlzSU8wJ/14.png b/vlm/dev/jDIlzSU8wJ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..206316d598604237837d65504db27fce241f76ae --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f303950e587d7a7c7796c9e7cff3b37b95989206bcdcbaea384fcbd2e2143dc +size 408101 diff --git a/vlm/dev/jDIlzSU8wJ/15.png b/vlm/dev/jDIlzSU8wJ/15.png new file mode 100644 index 0000000000000000000000000000000000000000..a8038978f58ad66fcb7c4a2d618794b79fa6bd3c --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f0d75bdd5d6b587e84d87215df1ad5665ac7292977a8bf874f9302457b69621 +size 1069630 diff --git a/vlm/dev/jDIlzSU8wJ/16.png b/vlm/dev/jDIlzSU8wJ/16.png new file mode 100644 index 0000000000000000000000000000000000000000..e1062105ca477a84d70c68d404123cc2924227cd --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c98dc61aa2e3532d12f6eeb8db01c7911895aa210e897508490c50d2aae60c4 +size 857539 diff --git a/vlm/dev/jDIlzSU8wJ/17.png b/vlm/dev/jDIlzSU8wJ/17.png new file mode 100644 index 0000000000000000000000000000000000000000..9e708e844cfd3b2679634afa40a02a5a51149b13 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae40771688d790bcad381e5a2574d2944dd97fc67e29d67d77bab2a4dc84f337 +size 1130619 diff --git a/vlm/dev/jDIlzSU8wJ/18.png b/vlm/dev/jDIlzSU8wJ/18.png new file mode 100644 index 0000000000000000000000000000000000000000..2e2fc64b868873f9ffbd9090daf99c7400d91531 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8b051954b210a64abbcc9ee2f801fab11003efa31e50e0a8adc1332d3f86bf4 +size 1899881 diff --git a/vlm/dev/jDIlzSU8wJ/19.png b/vlm/dev/jDIlzSU8wJ/19.png new file mode 100644 index 0000000000000000000000000000000000000000..d0eb617e126899ea550087feb4854488adb84af9 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e291cd7c68c973fcf4c6e561c0fdc4636672a0d77b32f90eea338e3ebe7c117 +size 1134547 diff --git a/vlm/dev/jDIlzSU8wJ/2.png b/vlm/dev/jDIlzSU8wJ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..c2512f4f973e82a9d68874084ed515042e19dd4c --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:384cccf8b3d15c6705b328f305e0b3fb2292c0f0608aae2fb41ca491b68f6079 +size 753205 diff --git a/vlm/dev/jDIlzSU8wJ/20.png b/vlm/dev/jDIlzSU8wJ/20.png new file mode 100644 index 0000000000000000000000000000000000000000..1822ff8c5f5c951dc55c26f5688da381e8488aa7 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bd5ddc297f014152c813cfdab6911ca74a781f676307e2936b8c88c6d3f74d1 +size 469988 diff --git a/vlm/dev/jDIlzSU8wJ/21.png b/vlm/dev/jDIlzSU8wJ/21.png new file mode 100644 index 0000000000000000000000000000000000000000..b61b51acccaa6e941305ef233a6934d156bd8c8e --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11e975debb13d72d229ac7e1034c7e4b64aeebc3bc7f82dde1246c683df80643 +size 755244 diff --git a/vlm/dev/jDIlzSU8wJ/22.png b/vlm/dev/jDIlzSU8wJ/22.png new file mode 100644 index 0000000000000000000000000000000000000000..20ae2ef9aae1eaf29ae46e8459fe9f81818b0267 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:934dfe2bb3f8fe3ea72bd8b79df5a82f8558f3478c290a087b37afdf9c5e9c98 +size 453848 diff --git a/vlm/dev/jDIlzSU8wJ/23.png b/vlm/dev/jDIlzSU8wJ/23.png new file mode 100644 index 0000000000000000000000000000000000000000..7df2cf8df6fdbc18dc61f96f6cfd71fd85e5592f --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a789f244480b8abef1c9349fefe7f0ca728db708b877966c565281fdfae929e0 +size 501153 diff --git a/vlm/dev/jDIlzSU8wJ/24.png b/vlm/dev/jDIlzSU8wJ/24.png new file mode 100644 index 0000000000000000000000000000000000000000..1b918b1dc20e13bd6f2b954841460ae87b9a3e5b --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76939175cd460f9610bc5b10602393a17c5d348a5fa77a887c3ad50c82dabf18 +size 434987 diff --git a/vlm/dev/jDIlzSU8wJ/25.png b/vlm/dev/jDIlzSU8wJ/25.png new file mode 100644 index 0000000000000000000000000000000000000000..f980efec0f0527a7a9e16b60cd6a3d908e2db7b0 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b124c074c4ac71a49f4a33407cbc4d80e7a736e106fc4c1a6bfc6d82e061e4b +size 1042932 diff --git a/vlm/dev/jDIlzSU8wJ/26.png b/vlm/dev/jDIlzSU8wJ/26.png new file mode 100644 index 0000000000000000000000000000000000000000..047f0ddb19ed34f84a4d864570c9cfc3eef09f9d --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9605cd2385301035a25e97773b2e2d69dfd4037ad68aa17979774d33124d35f +size 395800 diff --git a/vlm/dev/jDIlzSU8wJ/3.png b/vlm/dev/jDIlzSU8wJ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..366a7692281bb3500469b27ba6527837b2109d6f --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e9489aa272cd39ce0502c6b5f663e22b89cf74deb9042fca4c94f96c86b3a93 +size 747726 diff --git a/vlm/dev/jDIlzSU8wJ/4.png b/vlm/dev/jDIlzSU8wJ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..335c6f9a20a88f6506eb5064e8342803652b54d6 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:768a747d4755e60f32a146f327520a964296694f406ca6ecb9fbeca7bbf73e15 +size 564431 diff --git a/vlm/dev/jDIlzSU8wJ/5.png b/vlm/dev/jDIlzSU8wJ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..913c94c4fd0b0aab0fdc28dd2f3e1b7fc3af9f57 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad34c5bc979ae48cc2e13779f4e6fe8c2ce3c6e5a64d4836a8762c14b34cf39a +size 1018262 diff --git a/vlm/dev/jDIlzSU8wJ/6.png b/vlm/dev/jDIlzSU8wJ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..13b45a93ec121a759af040c728b9215fd69beb6c --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:206d81cb45e2d1efbe848608b75694bb8a3f9e826b1aad316f2ae4bf32a52ed8 +size 938041 diff --git a/vlm/dev/jDIlzSU8wJ/7.png b/vlm/dev/jDIlzSU8wJ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..14e1beb731cf21ce54eb737136b26f3dfae1f43a --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a08ab043e9fd8664692639893b5f69131c103c34a528806e36f51e234c0b8139 +size 547027 diff --git a/vlm/dev/jDIlzSU8wJ/8.png b/vlm/dev/jDIlzSU8wJ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..f0a5650a61c1fb3d254587d95de67ae22da03f18 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c55f004e5f027ecd8dc1030cb50efa9f27ba712d355070377c56e1d6d7935877 +size 767429 diff --git a/vlm/dev/jDIlzSU8wJ/9.png b/vlm/dev/jDIlzSU8wJ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..97a5ba5f079d4a78b8984eda9cb179b4f469dff4 --- /dev/null +++ b/vlm/dev/jDIlzSU8wJ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:671a80cc04d4819aba2a72a4482aa131a00f217dc1ae66434d632336b63f2718 +size 578392 diff --git a/vlm/dev/mNMwIwydgr/0.png b/vlm/dev/mNMwIwydgr/0.png new file mode 100644 index 0000000000000000000000000000000000000000..a9b0b00cae405f75824f6140ba135c42593fe799 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a09cdb46112606c604d4837d5499ce84e3b712c3f84a26867a6fa1aea1b7c023 +size 737282 diff --git a/vlm/dev/mNMwIwydgr/1.png b/vlm/dev/mNMwIwydgr/1.png new file mode 100644 index 0000000000000000000000000000000000000000..ab0f71a7fca74ac9128f122cc807c20ead396e5f --- /dev/null +++ b/vlm/dev/mNMwIwydgr/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81561d1b6a8420c099f17057d5dadc64a6dcd3f81a143d6ea3fce57efde7c4c6 +size 843947 diff --git a/vlm/dev/mNMwIwydgr/10.png b/vlm/dev/mNMwIwydgr/10.png new file mode 100644 index 0000000000000000000000000000000000000000..1a55b2d1b0183b05c0546e1758cfddd917590688 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b11cac056add81db10553e9555c575ad8919e85f5412ce813ad3c79dae34273 +size 969575 diff --git a/vlm/dev/mNMwIwydgr/11.png b/vlm/dev/mNMwIwydgr/11.png new file mode 100644 index 0000000000000000000000000000000000000000..a90e21f48f693d576cf36375cad271de233de4c7 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c967ef3d998f0a6b1227e57873f08b4570e8840ef5e0378658bf808503945d96 +size 943478 diff --git a/vlm/dev/mNMwIwydgr/12.png b/vlm/dev/mNMwIwydgr/12.png new file mode 100644 index 0000000000000000000000000000000000000000..a1835ae787ed8983cf34b3c7af5257cdde512e5a --- /dev/null +++ b/vlm/dev/mNMwIwydgr/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc6de66cd1455b4bf5c040d230390c16b30d9a3b527a1d42654e167994dd06be +size 976162 diff --git a/vlm/dev/mNMwIwydgr/13.png b/vlm/dev/mNMwIwydgr/13.png new file mode 100644 index 0000000000000000000000000000000000000000..1b6519cdaa9f212665b1aafc96fad9fdd90560d2 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee57a1dd58ffc83838aad0fe6df1284e2ba07922544338843a3aefd6b3ec59f6 +size 979343 diff --git a/vlm/dev/mNMwIwydgr/14.png b/vlm/dev/mNMwIwydgr/14.png new file mode 100644 index 0000000000000000000000000000000000000000..0b292130bd412b55ca6d2416ab0a9d89136afdce --- /dev/null +++ b/vlm/dev/mNMwIwydgr/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c75a87bb5e6199a6c6ce4f34c263c7c1a75b5ebb1f2fbb8c75abf84ef66d08d9 +size 980561 diff --git a/vlm/dev/mNMwIwydgr/15.png b/vlm/dev/mNMwIwydgr/15.png new file mode 100644 index 0000000000000000000000000000000000000000..5e075acb070371fb06b3f6f378c0a96861e3c5e4 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:813976f605dac1000da7285be16e6c840bf514a8fa6c3f5286fac972801fa601 +size 931119 diff --git a/vlm/dev/mNMwIwydgr/16.png b/vlm/dev/mNMwIwydgr/16.png new file mode 100644 index 0000000000000000000000000000000000000000..f9bbc4af1162b0dadc7ff00d5e35b211528f3d8f --- /dev/null +++ b/vlm/dev/mNMwIwydgr/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ff97697e04ad7a2d9c025149194d4f34d7a6a5a2c17169a85cf684b6177149d +size 1004194 diff --git a/vlm/dev/mNMwIwydgr/17.png b/vlm/dev/mNMwIwydgr/17.png new file mode 100644 index 0000000000000000000000000000000000000000..43f72e3e11253b10779488a806530bd85710df07 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98c777ddfb1cd4c34c1a83fdb4929cc30608ca93d7a2c297337007ea25e704b2 +size 1002364 diff --git a/vlm/dev/mNMwIwydgr/18.png b/vlm/dev/mNMwIwydgr/18.png new file mode 100644 index 0000000000000000000000000000000000000000..e3ddb283d4d9e232c68c585151058db8fc21a0c4 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:214fabbb632dd65b96a5649e10ed5d5634f661d73d5d1aaedd15e09641e93bb9 +size 933247 diff --git a/vlm/dev/mNMwIwydgr/19.png b/vlm/dev/mNMwIwydgr/19.png new file mode 100644 index 0000000000000000000000000000000000000000..d19b0205eba42134962283a43c396108ba0df1b8 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ee6af8f7c2c6467934df55d4e53ed4f2408935a16c960f82f4965a81e7d8c67 +size 1025956 diff --git a/vlm/dev/mNMwIwydgr/2.png b/vlm/dev/mNMwIwydgr/2.png new file mode 100644 index 0000000000000000000000000000000000000000..db415c35043b55bf9f52ebbcc5704c8824cf7ac2 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0566d200c84a3147e947e1c61b27d34f5645d3c0eb2825724054a87e682a4efb +size 845013 diff --git a/vlm/dev/mNMwIwydgr/20.png b/vlm/dev/mNMwIwydgr/20.png new file mode 100644 index 0000000000000000000000000000000000000000..59c31bd3c536c507db3641c77e8c1c6ce95c94e9 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea9a8d36d01d1edf76c007f4a0de160a3c18ba13d1439d1add04688e96c80f6 +size 1003532 diff --git a/vlm/dev/mNMwIwydgr/21.png b/vlm/dev/mNMwIwydgr/21.png new file mode 100644 index 0000000000000000000000000000000000000000..46e2a731f6e2671ac9166ea4b1f8200bdade421c --- /dev/null +++ b/vlm/dev/mNMwIwydgr/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d24ce56a508bb58f5098d726232a625fc3f0f2995faace878406ec9db1b1be6 +size 1006661 diff --git a/vlm/dev/mNMwIwydgr/22.png b/vlm/dev/mNMwIwydgr/22.png new file mode 100644 index 0000000000000000000000000000000000000000..984ea337d0eb026278b08091e99af3e9672848d8 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:005da6ac47f816e3433c3dec2606241d5b836758b15bea71f04bb5bd2dd43e40 +size 1052131 diff --git a/vlm/dev/mNMwIwydgr/23.png b/vlm/dev/mNMwIwydgr/23.png new file mode 100644 index 0000000000000000000000000000000000000000..b6ef85d44736b7b3c7be66972c14ae633ec5af7e --- /dev/null +++ b/vlm/dev/mNMwIwydgr/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2654fa04bffd28dfea07d7deee67b56ddd441979cbc8bf1b4c480d9d1bcc7b4 +size 1006030 diff --git a/vlm/dev/mNMwIwydgr/24.png b/vlm/dev/mNMwIwydgr/24.png new file mode 100644 index 0000000000000000000000000000000000000000..ceaaf8200b2b4b1510acae36e056841d477fa733 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbf818e161522b22fb3c206839c8344957b9effe20753949a9d488a049956062 +size 991776 diff --git a/vlm/dev/mNMwIwydgr/25.png b/vlm/dev/mNMwIwydgr/25.png new file mode 100644 index 0000000000000000000000000000000000000000..47eb3a85373a482e517068c50d05609e057b96af --- /dev/null +++ b/vlm/dev/mNMwIwydgr/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8aaae365f14c7c1e5d7b14e57899a118f59af28eefee1e213b722cd63baec53b +size 1026341 diff --git a/vlm/dev/mNMwIwydgr/26.png b/vlm/dev/mNMwIwydgr/26.png new file mode 100644 index 0000000000000000000000000000000000000000..dd43e7e44721f4f00c60884ee6602ed25848e8e3 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f9c76f3a8a8758521ad8d35ccaf9c796b66aa495e123489948f530a57029b1f +size 925095 diff --git a/vlm/dev/mNMwIwydgr/3.png b/vlm/dev/mNMwIwydgr/3.png new file mode 100644 index 0000000000000000000000000000000000000000..aea0787d9483b15095910476aafc3c4e6ad0c266 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4a7387403de92787a30273f7219623908856a2ad3ada0a3d04090f572b9da9b +size 933103 diff --git a/vlm/dev/mNMwIwydgr/4.png b/vlm/dev/mNMwIwydgr/4.png new file mode 100644 index 0000000000000000000000000000000000000000..c85b4ccdc01976d4f75f68eafbc61324c634afd7 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2d0c6dc210a72ab76174e4e46339c1529722028e635e3b981d35a4ae02e3b5b +size 905158 diff --git a/vlm/dev/mNMwIwydgr/5.png b/vlm/dev/mNMwIwydgr/5.png new file mode 100644 index 0000000000000000000000000000000000000000..e67162fd11830d2dd426a36458bd5d6d05bf197c --- /dev/null +++ b/vlm/dev/mNMwIwydgr/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:654f8c818439684cda0a49e01f7a2c7fcbf2f1296b24cb8e934d3228b659cf36 +size 612001 diff --git a/vlm/dev/mNMwIwydgr/6.png b/vlm/dev/mNMwIwydgr/6.png new file mode 100644 index 0000000000000000000000000000000000000000..672211e7959e7cdb08e803224b71fd6332c7675b --- /dev/null +++ b/vlm/dev/mNMwIwydgr/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f95062fc8b1f61cc07304dbcfd46b78fd79b007e366616a642b00850e6056ae +size 524752 diff --git a/vlm/dev/mNMwIwydgr/7.png b/vlm/dev/mNMwIwydgr/7.png new file mode 100644 index 0000000000000000000000000000000000000000..94f7495bf6ff7fee5eb1e5a3d783db95c83aea1c --- /dev/null +++ b/vlm/dev/mNMwIwydgr/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08ba7bf1e822f73da6e079b9d4aeed65e4a7282fe81b0612da7e158a9d860695 +size 879105 diff --git a/vlm/dev/mNMwIwydgr/8.png b/vlm/dev/mNMwIwydgr/8.png new file mode 100644 index 0000000000000000000000000000000000000000..f42523b56ff37574c0bbd1a0537e46fdf76df87b --- /dev/null +++ b/vlm/dev/mNMwIwydgr/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21be5975fce015bc1ccd4cd854c1ca1de2dedd22a15df9324943a365a4bf7efe +size 928895 diff --git a/vlm/dev/mNMwIwydgr/9.png b/vlm/dev/mNMwIwydgr/9.png new file mode 100644 index 0000000000000000000000000000000000000000..d31989967deea06ae818c86352b0953ae7d1ba53 --- /dev/null +++ b/vlm/dev/mNMwIwydgr/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1de7d754249dee7887a7f875101cacd765402b44a7de0313667f72a4957c90c +size 946597 diff --git a/vlm/dev/qHGCH75usg/0.png b/vlm/dev/qHGCH75usg/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c3574cda4517bf7c037e77579884571ea385fb3a --- /dev/null +++ b/vlm/dev/qHGCH75usg/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2657b6f9197bb51051a5b52ee4ccf112e7e599e6f7022e43033ae69f22166e3e +size 380748 diff --git a/vlm/dev/qHGCH75usg/1.png b/vlm/dev/qHGCH75usg/1.png new file mode 100644 index 0000000000000000000000000000000000000000..b94c63caaeed2374675e8c3b27f33894e36ae133 --- /dev/null +++ b/vlm/dev/qHGCH75usg/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8386ec27e815d665ff068281098011aacf9c625ba38db2309dace56badfcfa33 +size 622132 diff --git a/vlm/dev/qHGCH75usg/10.png b/vlm/dev/qHGCH75usg/10.png new file mode 100644 index 0000000000000000000000000000000000000000..48fd5e2eced452a35f808d77e4fbea4950ffef64 --- /dev/null +++ b/vlm/dev/qHGCH75usg/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8ddbcaed05dabbe6575fae03fbb4a8bc3369d85e0e9cc902f46faeb41711eb5 +size 554255 diff --git a/vlm/dev/qHGCH75usg/11.png b/vlm/dev/qHGCH75usg/11.png new file mode 100644 index 0000000000000000000000000000000000000000..f04062892839448c5c6956c53898f572698cdb05 --- /dev/null +++ b/vlm/dev/qHGCH75usg/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2e1148afa74f01b6ac900523896efc72bea80417ab2c2473dc221942403a14c +size 564460 diff --git a/vlm/dev/qHGCH75usg/12.png b/vlm/dev/qHGCH75usg/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b7bc8cf5d39c13dbc6e13d4236c226fb4286141a --- /dev/null +++ b/vlm/dev/qHGCH75usg/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:623df9a6962f1779ec562dad52bc43a95c28fda49da67b5536bb6d5975ab81ce +size 562227 diff --git a/vlm/dev/qHGCH75usg/13.png b/vlm/dev/qHGCH75usg/13.png new file mode 100644 index 0000000000000000000000000000000000000000..e26b3817b8fb6cc1ec58b4ce8ffbfc4d6c3d5482 --- /dev/null +++ b/vlm/dev/qHGCH75usg/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3050de8df4d7e877b3875e308289cbc7cf3063b6466d804869049a9860f88b11 +size 593904 diff --git a/vlm/dev/qHGCH75usg/14.png b/vlm/dev/qHGCH75usg/14.png new file mode 100644 index 0000000000000000000000000000000000000000..55433a09d67deb52293161a2db2bb44fed7273f9 --- /dev/null +++ b/vlm/dev/qHGCH75usg/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa86bfbb216b49e45f06efb4a10392c86f988575aabb45c6cf1c7319ce98e697 +size 492522 diff --git a/vlm/dev/qHGCH75usg/15.png b/vlm/dev/qHGCH75usg/15.png new file mode 100644 index 0000000000000000000000000000000000000000..970c9333d9f8d86b0e2e0f6b10cd8c07f655c221 --- /dev/null +++ b/vlm/dev/qHGCH75usg/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dde97c9d185278ca4fb650e22621be2df8bacdbb714082384d6cd5c70370cc1d +size 385408 diff --git a/vlm/dev/qHGCH75usg/2.png b/vlm/dev/qHGCH75usg/2.png new file mode 100644 index 0000000000000000000000000000000000000000..2de52fa38d03e1cd5ca6ae841c43679b66524ce0 --- /dev/null +++ b/vlm/dev/qHGCH75usg/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2e3c07b93a9debff21c507a82d1c656257ffa88c2622b5da9148249b6924a3f +size 420992 diff --git a/vlm/dev/qHGCH75usg/3.png b/vlm/dev/qHGCH75usg/3.png new file mode 100644 index 0000000000000000000000000000000000000000..3b40e6fe87b484c59d6ca1e381a27aeea5f23a2f --- /dev/null +++ b/vlm/dev/qHGCH75usg/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a06d8c59c5add67c1f990c8be40221484ee61db5f7608ad91832c350934162d4 +size 541527 diff --git a/vlm/dev/qHGCH75usg/4.png b/vlm/dev/qHGCH75usg/4.png new file mode 100644 index 0000000000000000000000000000000000000000..5fe95a7b7fe9efc75eab05a4be2e5389884ee6a9 --- /dev/null +++ b/vlm/dev/qHGCH75usg/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b7bef395c642571568296e8e4bf84b93d5e74d25d3652bf2f660fcd9205ed93 +size 576702 diff --git a/vlm/dev/qHGCH75usg/5.png b/vlm/dev/qHGCH75usg/5.png new file mode 100644 index 0000000000000000000000000000000000000000..dd23968ba8f7c114e8c9efb8ab03db5d4d1464d9 --- /dev/null +++ b/vlm/dev/qHGCH75usg/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e851aa204e51ee75149437c1a6e6dca3a6aaff37c54357381665b7627bee5ed0 +size 772459 diff --git a/vlm/dev/qHGCH75usg/6.png b/vlm/dev/qHGCH75usg/6.png new file mode 100644 index 0000000000000000000000000000000000000000..7ae7698891f30cbb1424d39c9bbc46a5e4aec5ed --- /dev/null +++ b/vlm/dev/qHGCH75usg/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e9819789c6267c3019aa9386ad51ff5f7e2d12927bae8ab6a9a91508332f1d1 +size 609242 diff --git a/vlm/dev/qHGCH75usg/7.png b/vlm/dev/qHGCH75usg/7.png new file mode 100644 index 0000000000000000000000000000000000000000..2707ceefac04b88b3e4648c6e9a59032e4d36711 --- /dev/null +++ b/vlm/dev/qHGCH75usg/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02a3c4671dcb1e2823ac925b810582996d5db08ef42874c4d0fcc4b6955477ea +size 573512 diff --git a/vlm/dev/qHGCH75usg/8.png b/vlm/dev/qHGCH75usg/8.png new file mode 100644 index 0000000000000000000000000000000000000000..b8cb17c3c58020ae345b838002574296a07240b4 --- /dev/null +++ b/vlm/dev/qHGCH75usg/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0c8b90edf38789fbda68eb61ff9e872586709eb2e5453a3dac7ec300f116252 +size 547273 diff --git a/vlm/dev/qHGCH75usg/9.png b/vlm/dev/qHGCH75usg/9.png new file mode 100644 index 0000000000000000000000000000000000000000..076a6303af03a07592f4afae1167f0afb64199a2 --- /dev/null +++ b/vlm/dev/qHGCH75usg/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a939260e6ac06e6283be453e61189423cac8f250a4bcfe553ff5882666339ab +size 425258 diff --git a/vlm/dev/u6Xv3FuF8N/0.png b/vlm/dev/u6Xv3FuF8N/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c30aadc9102a9a12c364daafcad2b1f2686d36e4 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3be7299fac421c7491488d63cbca1bd97d0f499531d17fede1a8b53ad6238b3f +size 448734 diff --git a/vlm/dev/u6Xv3FuF8N/1.png b/vlm/dev/u6Xv3FuF8N/1.png new file mode 100644 index 0000000000000000000000000000000000000000..2d534ab12db38547e0cbd478edaf2e5bdd70967d --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:416d22d43f9bd8fe164fd372b7660b23a7721b20db633946856f8297e6ffe82d +size 575751 diff --git a/vlm/dev/u6Xv3FuF8N/10.png b/vlm/dev/u6Xv3FuF8N/10.png new file mode 100644 index 0000000000000000000000000000000000000000..3fe284650597610d692e683a02b05e3d00d32290 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e30870e35c8486139e6ecf0a630c50a3eb0fa00239bcc601ea62a274163110e +size 570345 diff --git a/vlm/dev/u6Xv3FuF8N/11.png b/vlm/dev/u6Xv3FuF8N/11.png new file mode 100644 index 0000000000000000000000000000000000000000..d0ef22910dba8a46512991362857cbf5f21f1092 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5891246e99d95fa6824408816e0379ceb5f24880ab970fdd219a4bcca287365 +size 545879 diff --git a/vlm/dev/u6Xv3FuF8N/12.png b/vlm/dev/u6Xv3FuF8N/12.png new file mode 100644 index 0000000000000000000000000000000000000000..f3f732d1eab9ecf65a0c8fb12d1cd83ba07851fc --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3010ff908f4009469dc7b6b7a0af0c48f1ead67ddfbfec2bfdda8edc48b5a1c +size 199182 diff --git a/vlm/dev/u6Xv3FuF8N/13.png b/vlm/dev/u6Xv3FuF8N/13.png new file mode 100644 index 0000000000000000000000000000000000000000..1271d46fdba7b054abfef20795c0533f20968d1d --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cd15470472ba498032e9054fe439c7eb9317a8db41d5ab28fc1652e70c52848 +size 491942 diff --git a/vlm/dev/u6Xv3FuF8N/14.png b/vlm/dev/u6Xv3FuF8N/14.png new file mode 100644 index 0000000000000000000000000000000000000000..03225ba4367c6d1b34a951247d50fbfa53cb7d60 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:719f848d9a779ab58dc26dbaece2eb57e7f44fa9135444e241a41a29626c2264 +size 499711 diff --git a/vlm/dev/u6Xv3FuF8N/15.png b/vlm/dev/u6Xv3FuF8N/15.png new file mode 100644 index 0000000000000000000000000000000000000000..0d32b4bcfc710b7cd01f82257947c01e78bfc082 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53eaed00a03503a94cbb83b59ae0312546df79f51db3cc90ae0673648232fce2 +size 445432 diff --git a/vlm/dev/u6Xv3FuF8N/16.png b/vlm/dev/u6Xv3FuF8N/16.png new file mode 100644 index 0000000000000000000000000000000000000000..9529dc04186e0a2b1a455b5914ad3b1a80f0d38c --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b33500d458815b73bab6d7a9341652c9c27948ed05ddef7eead0d0f94083b0b +size 429972 diff --git a/vlm/dev/u6Xv3FuF8N/17.png b/vlm/dev/u6Xv3FuF8N/17.png new file mode 100644 index 0000000000000000000000000000000000000000..996833a631c80e04143790fdcf42f440c527aeb5 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a7cb7fe031a960156b38088504fd4eac509088581d50a9c70b52274d7033fa8 +size 228995 diff --git a/vlm/dev/u6Xv3FuF8N/18.png b/vlm/dev/u6Xv3FuF8N/18.png new file mode 100644 index 0000000000000000000000000000000000000000..154bf3bb0034241d774cff271d323d1d1c4e723a --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32ea55a563c3a74e9eead29f66addca5ab393883895c26eedbc5ff0745f13d3b +size 386025 diff --git a/vlm/dev/u6Xv3FuF8N/19.png b/vlm/dev/u6Xv3FuF8N/19.png new file mode 100644 index 0000000000000000000000000000000000000000..ffe62be607735f092c9087d3eb6510a7367719eb --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc2595ba45ab0872446579d6aa77bf2bc263be8fa90a8a66b4fe70f455263e4f +size 232155 diff --git a/vlm/dev/u6Xv3FuF8N/2.png b/vlm/dev/u6Xv3FuF8N/2.png new file mode 100644 index 0000000000000000000000000000000000000000..119ba14ffe3e819f7f46fc9b0e2ea20bdff64d6e --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26797be5311115130fb8019fe6870915073e652429263187750720a515114c2d +size 595313 diff --git a/vlm/dev/u6Xv3FuF8N/3.png b/vlm/dev/u6Xv3FuF8N/3.png new file mode 100644 index 0000000000000000000000000000000000000000..fe962a5c5794026e8d39f6315e651f8045089471 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec55243a4c473b8463c78df64282681c5778b3878b7a25374e6c2459585cbc7f +size 553552 diff --git a/vlm/dev/u6Xv3FuF8N/4.png b/vlm/dev/u6Xv3FuF8N/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ca88b8077900aaf47781a820c374f009ff7e5dca --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82eb6c68ad6c29facd2e05361361d4fc936afe448cb7d13721f700583a5d754e +size 610639 diff --git a/vlm/dev/u6Xv3FuF8N/5.png b/vlm/dev/u6Xv3FuF8N/5.png new file mode 100644 index 0000000000000000000000000000000000000000..f71b14406fbd5139eb3c19d947fc1946d35664b4 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc35bad6a8a1d9b15188d6e86bb333e3e40c2bdc6c4c5f9105c6f53624a269ad +size 640958 diff --git a/vlm/dev/u6Xv3FuF8N/6.png b/vlm/dev/u6Xv3FuF8N/6.png new file mode 100644 index 0000000000000000000000000000000000000000..ad5c3858ca826fd7a08ac0c66bf979175160e528 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbbc638ee95a1c45da1a0138f692696809e0a10c6143840ed83a16c2ac9d9024 +size 546172 diff --git a/vlm/dev/u6Xv3FuF8N/7.png b/vlm/dev/u6Xv3FuF8N/7.png new file mode 100644 index 0000000000000000000000000000000000000000..ce1004df7ef3b0846270b448716f4877cbe17ae9 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:849f3b4ee122d74405d6bdf5c8440933c20890608a9159332d399b24e064cfc5 +size 609104 diff --git a/vlm/dev/u6Xv3FuF8N/8.png b/vlm/dev/u6Xv3FuF8N/8.png new file mode 100644 index 0000000000000000000000000000000000000000..6e5860259131c804c09cb7410c6b4443d7bbf2d5 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e8f78eacb52aa54cc23f741b044bb4b3158ee0e78a7ff27ed75fc08208a2d66 +size 512091 diff --git a/vlm/dev/u6Xv3FuF8N/9.png b/vlm/dev/u6Xv3FuF8N/9.png new file mode 100644 index 0000000000000000000000000000000000000000..3b9cfb7d1caa140d6ff27b234109f333bd4b6ef9 --- /dev/null +++ b/vlm/dev/u6Xv3FuF8N/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3a42ad047a713a052b10446398f7175bd5d57088d1712546a2e5af0fa9b9fcd +size 552284 diff --git a/vlm/dev/ucNDIDRNjjv/0.png b/vlm/dev/ucNDIDRNjjv/0.png new file mode 100644 index 0000000000000000000000000000000000000000..39a7756e84f9af94bca0939b62ad6470612f27d2 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5611f86c79837ee91a82472d6614afb73ecf0c36d83ffe66c3ae15665f434add +size 439872 diff --git a/vlm/dev/ucNDIDRNjjv/1.png b/vlm/dev/ucNDIDRNjjv/1.png new file mode 100644 index 0000000000000000000000000000000000000000..bf3b69da90c12505e806e4b7e1089fc53d4c5ccc --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18a069adf3ad557f6665be70f3502e52b74f3901363ddc03bbb7aac495098a42 +size 601672 diff --git a/vlm/dev/ucNDIDRNjjv/10.png b/vlm/dev/ucNDIDRNjjv/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d3f11621ae0f90d48cdaff002aa7f087e5bf1b22 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:433ea38036de7354aec4ca4b582725ba49a5de716b3bff506408846e3c77a755 +size 461926 diff --git a/vlm/dev/ucNDIDRNjjv/11.png b/vlm/dev/ucNDIDRNjjv/11.png new file mode 100644 index 0000000000000000000000000000000000000000..fe6b7735b1d996be941dbe71ed5e438cecc709ba --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39c943d98777a5fb76b0ef88aec2701d3936207660accc7c23d8d190b1d26639 +size 368840 diff --git a/vlm/dev/ucNDIDRNjjv/12.png b/vlm/dev/ucNDIDRNjjv/12.png new file mode 100644 index 0000000000000000000000000000000000000000..4611905ad9e9210f7b1c2e9e7bb349cc192f1ee3 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a01e7bac9e7bf4633d09633f2e9b996289c784c970150d9ff5346662bd19cc4 +size 329993 diff --git a/vlm/dev/ucNDIDRNjjv/2.png b/vlm/dev/ucNDIDRNjjv/2.png new file mode 100644 index 0000000000000000000000000000000000000000..c7c699ed9c259cd967c55f53c7833927636c1336 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1e678137f3ec066fc4a55ce55c0ef7472e8b500f547f3d3a641b717f15e90cc +size 562157 diff --git a/vlm/dev/ucNDIDRNjjv/3.png b/vlm/dev/ucNDIDRNjjv/3.png new file mode 100644 index 0000000000000000000000000000000000000000..3e0e68cefafdd8f6b9fe8a9dcad91284e35e15b1 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa8c1a49a52efb7b846f59c4774f714e45cbb723421b8ed59146796b3f3eb2e4 +size 471006 diff --git a/vlm/dev/ucNDIDRNjjv/4.png b/vlm/dev/ucNDIDRNjjv/4.png new file mode 100644 index 0000000000000000000000000000000000000000..30af39f84d9cd94234bdb274cbc0091479797725 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16fa553b4789022ef312691a39e1f5216244df99663a91c2e66e01daaa8c8957 +size 533795 diff --git a/vlm/dev/ucNDIDRNjjv/5.png b/vlm/dev/ucNDIDRNjjv/5.png new file mode 100644 index 0000000000000000000000000000000000000000..0ce7d32b5a0c714394a37c55010753bbb5b06811 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00eb5ae400450d632ba332daa584ace2ebb50b26a6f63bc46469a4cd57ada4a9 +size 552076 diff --git a/vlm/dev/ucNDIDRNjjv/6.png b/vlm/dev/ucNDIDRNjjv/6.png new file mode 100644 index 0000000000000000000000000000000000000000..4064afbd40c1ca7025707c423d30406b4a615d7b --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d33b833cf1e860fd517edd132708536bd7dc11d6760044b6c792c274e6bf7ae +size 567073 diff --git a/vlm/dev/ucNDIDRNjjv/7.png b/vlm/dev/ucNDIDRNjjv/7.png new file mode 100644 index 0000000000000000000000000000000000000000..9b78d3b52348e967e8d13328b51b44df8c2506c5 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73831b86c2111f82e02cba3c81d2df0c3b79f8e4a667e64a6f730891c90905e0 +size 554732 diff --git a/vlm/dev/ucNDIDRNjjv/8.png b/vlm/dev/ucNDIDRNjjv/8.png new file mode 100644 index 0000000000000000000000000000000000000000..e47fd2879b19dcbb53dd010de6a64c16c6d9e76f --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d55a8ae8c361533a6718444f9d7cf21000e24c268617dc38a669e1223a8024fe +size 445139 diff --git a/vlm/dev/ucNDIDRNjjv/9.png b/vlm/dev/ucNDIDRNjjv/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1ec3ed42f9516ff213d9097d001e589bdaa74fa2 --- /dev/null +++ b/vlm/dev/ucNDIDRNjjv/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f2fd573bc4e48ff6173fd3cdae477ab0a57e54cb872c78f7e529a9dcf9fe0cb +size 532358 diff --git a/vlm/dev/uu6Oq7MN7g/0.png b/vlm/dev/uu6Oq7MN7g/0.png new file mode 100644 index 0000000000000000000000000000000000000000..3aac33f02751ffa9b90b33a73cf5bd42aa0d9cb6 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fd6f87632a70a46bf2c2efb26f42b4d07d9610495f114cbb7bbb6bcbe74b56a +size 710016 diff --git a/vlm/dev/uu6Oq7MN7g/1.png b/vlm/dev/uu6Oq7MN7g/1.png new file mode 100644 index 0000000000000000000000000000000000000000..7f2f9ab456cd8c8c3e51baa0d4b004186a8fa94a --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ef471dfc8594506492e882e8de01ef596bb9c0380c6b596713d1dc53008acc8 +size 708424 diff --git a/vlm/dev/uu6Oq7MN7g/10.png b/vlm/dev/uu6Oq7MN7g/10.png new file mode 100644 index 0000000000000000000000000000000000000000..343769e5fd60bf86bfb57fe086e928407ef42778 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2523343782c4ae1742533b44f0dba79da01af7c734848330305f85a11294881d +size 883770 diff --git a/vlm/dev/uu6Oq7MN7g/11.png b/vlm/dev/uu6Oq7MN7g/11.png new file mode 100644 index 0000000000000000000000000000000000000000..8140a4c438f2bec31b97a2530360f1614ed7480f --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9da7b902e70e8295546774c95c887a7079925dd5c9db1be05afeead9e63949b6 +size 844335 diff --git a/vlm/dev/uu6Oq7MN7g/12.png b/vlm/dev/uu6Oq7MN7g/12.png new file mode 100644 index 0000000000000000000000000000000000000000..6422a7206888722ff4fbca1ba3cb613d19458129 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff06a11df75978505a438eef7ea42ba1d9178fc3fabad1d59eda0afe2c0944af +size 816598 diff --git a/vlm/dev/uu6Oq7MN7g/13.png b/vlm/dev/uu6Oq7MN7g/13.png new file mode 100644 index 0000000000000000000000000000000000000000..b5e0e0dcecef81eadf0d57a3421587eba30e08fd --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88ce86224066b748713f959073953adccf113d131a48c0394ad2871ff16c1b06 +size 695365 diff --git a/vlm/dev/uu6Oq7MN7g/14.png b/vlm/dev/uu6Oq7MN7g/14.png new file mode 100644 index 0000000000000000000000000000000000000000..8cfe7bfa4e39d1983cfda87af5581e9c44623ef4 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb4a143a10c298971adbb769a7527308ec9013e120c08789eb05970dc873579f +size 647624 diff --git a/vlm/dev/uu6Oq7MN7g/15.png b/vlm/dev/uu6Oq7MN7g/15.png new file mode 100644 index 0000000000000000000000000000000000000000..a64fd419cecbfa2236b038393726bfb6849e3983 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3937e164b9872536451781104125d844929f3621502fce0749e8ddcb268727c +size 733525 diff --git a/vlm/dev/uu6Oq7MN7g/16.png b/vlm/dev/uu6Oq7MN7g/16.png new file mode 100644 index 0000000000000000000000000000000000000000..31fdfd80075678f799550310f64e4f077cd60e38 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a893796adf8039a7244c4b97fe70a360de9efd27b2ae031d6c282c47847bb94d +size 672054 diff --git a/vlm/dev/uu6Oq7MN7g/17.png b/vlm/dev/uu6Oq7MN7g/17.png new file mode 100644 index 0000000000000000000000000000000000000000..6c8444d57674c09f9ad612a24cac7b6a7d5f583b --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1422d0ca2f0aa569f1fcd6a1fadfd13d313ebf3f7541e8459a8effee6db3830e +size 620684 diff --git a/vlm/dev/uu6Oq7MN7g/18.png b/vlm/dev/uu6Oq7MN7g/18.png new file mode 100644 index 0000000000000000000000000000000000000000..5cb2a8bba47b37db0bce6e6a39a6b525cc1726c1 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9a542fe60245695ab63c9babe7b07183ab5b73b3251b527c778390f8c0cf488 +size 607646 diff --git a/vlm/dev/uu6Oq7MN7g/19.png b/vlm/dev/uu6Oq7MN7g/19.png new file mode 100644 index 0000000000000000000000000000000000000000..a891d2a48330a4c4285573d6dbe58ebef0e87a33 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:474d53d237d2d4463ea71a56066f6237c4afb84d6dafc5b317e03c8216d5898f +size 392133 diff --git a/vlm/dev/uu6Oq7MN7g/2.png b/vlm/dev/uu6Oq7MN7g/2.png new file mode 100644 index 0000000000000000000000000000000000000000..935bb440e8a996905c623c0f2bf40ac0a5a4fed3 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12ecffd072ce0341b3028edf00d5695f6aa2498c4ae9663e684dab143d437f70 +size 714155 diff --git a/vlm/dev/uu6Oq7MN7g/3.png b/vlm/dev/uu6Oq7MN7g/3.png new file mode 100644 index 0000000000000000000000000000000000000000..fcf86fa52a579ede7bdb79863462bd431ef5ee79 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1770db9e8f325c32bd79cc48bb93013211d9bf8f34f9f495d5cfdb887fe32003 +size 706612 diff --git a/vlm/dev/uu6Oq7MN7g/4.png b/vlm/dev/uu6Oq7MN7g/4.png new file mode 100644 index 0000000000000000000000000000000000000000..f0a79b66702e6c536746e140f872f6dbea70e87f --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddf23b192f666606e8d8b9f78ba3c974c29dae4b9d5dcfda2ddf4e0efab01d46 +size 765334 diff --git a/vlm/dev/uu6Oq7MN7g/5.png b/vlm/dev/uu6Oq7MN7g/5.png new file mode 100644 index 0000000000000000000000000000000000000000..249d86b7b7b338ef5202b7070380410ee92251c7 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17ba207d52a93e780142074c569f3b9e1b4110cce2777edd79d32b0e3efa8266 +size 848948 diff --git a/vlm/dev/uu6Oq7MN7g/6.png b/vlm/dev/uu6Oq7MN7g/6.png new file mode 100644 index 0000000000000000000000000000000000000000..bdee76ff713031eb14bf40a03b46645b620204cc --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4440db2ef1132c96572f98a2f8934a439db2247888d65faaba05b35897609be5 +size 737772 diff --git a/vlm/dev/uu6Oq7MN7g/7.png b/vlm/dev/uu6Oq7MN7g/7.png new file mode 100644 index 0000000000000000000000000000000000000000..de8fad40fdf17d05c65286a49ded169c5609d4c0 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2df8f334d9c2d6ef3cb9344613eb5cd4e033501b9e60ccfdf12fa58288acb755 +size 772928 diff --git a/vlm/dev/uu6Oq7MN7g/8.png b/vlm/dev/uu6Oq7MN7g/8.png new file mode 100644 index 0000000000000000000000000000000000000000..280675c1443f2f20095536d7d929e9bc42e9c9fb --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bdf240a73a8a7e9cecb9da2ba280459767381de34192fdb85e3b1b6be58b8410 +size 741809 diff --git a/vlm/dev/uu6Oq7MN7g/9.png b/vlm/dev/uu6Oq7MN7g/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0a34a9cc378f019f61aa2454182c054a5adffd90 --- /dev/null +++ b/vlm/dev/uu6Oq7MN7g/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a59536ac0963a4b008d6a7ab6cd2ec9151cde8eb76835d6248a473279f0de81 +size 780002 diff --git a/vlm/dev/zDbsSscmuj/0.png b/vlm/dev/zDbsSscmuj/0.png new file mode 100644 index 0000000000000000000000000000000000000000..e03d4a4f86c06ca267e06304dc4bf7b0921c0428 --- /dev/null +++ b/vlm/dev/zDbsSscmuj/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:038752a10a6b6cc70bd03dcd273bedc56caf2ef0c26017a38be5f109141a59fe +size 405306 diff --git a/vlm/dev/zDbsSscmuj/1.png b/vlm/dev/zDbsSscmuj/1.png new file mode 100644 index 0000000000000000000000000000000000000000..8aa9e1d778ad74eaccbca43b977cbf0fda1da0cf --- /dev/null +++ b/vlm/dev/zDbsSscmuj/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:523d43c446df4b76d930036aa4f587d9a2c02942c6211889a13825d6ebca6b20 +size 606488 diff --git a/vlm/dev/zDbsSscmuj/10.png b/vlm/dev/zDbsSscmuj/10.png new file mode 100644 index 0000000000000000000000000000000000000000..0ff67f1e8febf8f4f34882e1bb0068b546ae93e6 --- /dev/null +++ b/vlm/dev/zDbsSscmuj/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b28eff29b959d0be22aa1bb3e5bf4ac595986e247f36e27aadc458e9cec4af5 +size 549531 diff --git a/vlm/dev/zDbsSscmuj/11.png b/vlm/dev/zDbsSscmuj/11.png new file mode 100644 index 0000000000000000000000000000000000000000..d453dd67a19dbf542d81f1f6d05e85aaf3698e83 --- /dev/null +++ b/vlm/dev/zDbsSscmuj/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8d1eace8bed71aebe8ca07e67bd8f184865b5130167c17d24afd35de794ca0c +size 556880 diff --git a/vlm/dev/zDbsSscmuj/12.png b/vlm/dev/zDbsSscmuj/12.png new file mode 100644 index 0000000000000000000000000000000000000000..e8e0a975c94d557f3e6ffcbd6c9c5554a47b6831 --- /dev/null +++ b/vlm/dev/zDbsSscmuj/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da8410ce942387a78b4aa490613db8ecd63bd9f55bab28f002af282b843ad70b +size 547343 diff --git a/vlm/dev/zDbsSscmuj/13.png b/vlm/dev/zDbsSscmuj/13.png new file mode 100644 index 0000000000000000000000000000000000000000..017e93196ba723268acdc79474f205dd589cd88b --- /dev/null +++ b/vlm/dev/zDbsSscmuj/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95d02f8ff917b51d35aed210440c68938fddd08f1638d40d44e3344043adc926 +size 209604 diff --git a/vlm/dev/zDbsSscmuj/2.png b/vlm/dev/zDbsSscmuj/2.png new file mode 100644 index 0000000000000000000000000000000000000000..94cc3b29a2a669e78b92592a4c0763b8535efa30 --- /dev/null +++ b/vlm/dev/zDbsSscmuj/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e589195f38744baa384d8a0a1a1ad10533c1b345ec9eb809c65733dbca8271e +size 609840 diff --git a/vlm/dev/zDbsSscmuj/3.png b/vlm/dev/zDbsSscmuj/3.png new file mode 100644 index 0000000000000000000000000000000000000000..da2e58eca922a851afd86909a5679233369dbb89 --- /dev/null +++ b/vlm/dev/zDbsSscmuj/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c14154455da7d2133c4569a30707cee25acb7f03d65258dc32cbff13717245fe +size 516859 diff --git a/vlm/dev/zDbsSscmuj/4.png b/vlm/dev/zDbsSscmuj/4.png new file mode 100644 index 0000000000000000000000000000000000000000..771fe5168824962312be304ea08ba1c855e7038b --- /dev/null +++ b/vlm/dev/zDbsSscmuj/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10200a8073f7bd41a58a27b56e9683a8a786b120a76b48f7bdabf5501e9547a4 +size 599428 diff --git a/vlm/dev/zDbsSscmuj/5.png b/vlm/dev/zDbsSscmuj/5.png new file mode 100644 index 0000000000000000000000000000000000000000..4f4050297f2f2f1f08e733598a33283479e2a64e --- /dev/null +++ b/vlm/dev/zDbsSscmuj/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fd989fe83e6ffb6cac18f25979ccfe66aea97d06e671c8564b8b993ef192ed0 +size 427280 diff --git a/vlm/dev/zDbsSscmuj/6.png b/vlm/dev/zDbsSscmuj/6.png new file mode 100644 index 0000000000000000000000000000000000000000..7cf19da878efeb4619af28850002070128cb0da8 --- /dev/null +++ b/vlm/dev/zDbsSscmuj/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca3c56edebfb25e9294a91e7439fc0ccabcde01971992eaee6bc97ea659a7064 +size 568311 diff --git a/vlm/dev/zDbsSscmuj/7.png b/vlm/dev/zDbsSscmuj/7.png new file mode 100644 index 0000000000000000000000000000000000000000..a93205aeac36d5b98579310fd2f6071a65af2bbf --- /dev/null +++ b/vlm/dev/zDbsSscmuj/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:388acabe4f009665b7ae46f3de159f3cf058ce2ad346a3193eb357915decd696 +size 558909 diff --git a/vlm/dev/zDbsSscmuj/8.png b/vlm/dev/zDbsSscmuj/8.png new file mode 100644 index 0000000000000000000000000000000000000000..232da9b68c95e9e3d03f5d75e3befe7967fb9438 --- /dev/null +++ b/vlm/dev/zDbsSscmuj/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7419c3f42c5a88154615bc1dddce7d3e8cd2e67aacfca3aad771196c7ceeb54e +size 631362 diff --git a/vlm/dev/zDbsSscmuj/9.png b/vlm/dev/zDbsSscmuj/9.png new file mode 100644 index 0000000000000000000000000000000000000000..feba7d810ceab1b776616d9f21278cf18494d33f --- /dev/null +++ b/vlm/dev/zDbsSscmuj/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ba97588f32b95c4f87f2eba2b8e5b9a3b8da4f5c2fc2b25ab693a3bc13b52b5 +size 575418 diff --git a/vlm/dev/zeGpMIt6Pfq/0.png b/vlm/dev/zeGpMIt6Pfq/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c66a49691aeacc93d3bf1d6c29dca17d7de3311d --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04eba04d1a98e9e02071bab0258263866f8286e01605c60207dd44c1146310e5 +size 513986 diff --git a/vlm/dev/zeGpMIt6Pfq/1.png b/vlm/dev/zeGpMIt6Pfq/1.png new file mode 100644 index 0000000000000000000000000000000000000000..8ed99ab5031834a20031c4cc3b03a537364c22a7 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7574f40ed30c9396abf323baf74d334a93efcf6a661239c5fd0197e5a012f61f +size 635946 diff --git a/vlm/dev/zeGpMIt6Pfq/10.png b/vlm/dev/zeGpMIt6Pfq/10.png new file mode 100644 index 0000000000000000000000000000000000000000..7b80fcfc291014c497300ca0104c9f1c8055a244 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:313da449bca5145d59879d17fcaf4c4b4c279ac58edc2543f063e52188113207 +size 553609 diff --git a/vlm/dev/zeGpMIt6Pfq/11.png b/vlm/dev/zeGpMIt6Pfq/11.png new file mode 100644 index 0000000000000000000000000000000000000000..4b4699dd9a00365c95f00930ab0fa5ae963da4cb --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2897b3578844ace33bd3670b6c449c689673741b0bc0e8860ecb0f06587bf7f0 +size 396696 diff --git a/vlm/dev/zeGpMIt6Pfq/2.png b/vlm/dev/zeGpMIt6Pfq/2.png new file mode 100644 index 0000000000000000000000000000000000000000..497972b39720197e9fd84706bece61d08eb07699 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4c107144499dce5731e7ed601ba2329582d40490ed59704a643714558f1d729 +size 485447 diff --git a/vlm/dev/zeGpMIt6Pfq/3.png b/vlm/dev/zeGpMIt6Pfq/3.png new file mode 100644 index 0000000000000000000000000000000000000000..8603edaeed7a99e9983a2cfaa8192996b5278326 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:377ef634c9b47f28cd4990a3004c1cfc6b9c3af7d7fa8a2473061365d710cde8 +size 458844 diff --git a/vlm/dev/zeGpMIt6Pfq/4.png b/vlm/dev/zeGpMIt6Pfq/4.png new file mode 100644 index 0000000000000000000000000000000000000000..970e80603701daa92179c126d340b9786f9bee85 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63364344ea7e31eb900a18a0098c487620f4da7a8014778398485f08556a5676 +size 523842 diff --git a/vlm/dev/zeGpMIt6Pfq/5.png b/vlm/dev/zeGpMIt6Pfq/5.png new file mode 100644 index 0000000000000000000000000000000000000000..0dd47acc90f3b69160e5ad7c5f51900808f684b0 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:379a5f7ff00192ec55a9cd5e49bcda6af744c02302faf342a0838affdda6488b +size 566556 diff --git a/vlm/dev/zeGpMIt6Pfq/6.png b/vlm/dev/zeGpMIt6Pfq/6.png new file mode 100644 index 0000000000000000000000000000000000000000..cca2e9ca60eb76e9adbf012d743387aec8a39f14 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46a4bb8e112793232a1d9ad1bc62846fd6d11801bfd4df018aebf442eb5df9e9 +size 425577 diff --git a/vlm/dev/zeGpMIt6Pfq/7.png b/vlm/dev/zeGpMIt6Pfq/7.png new file mode 100644 index 0000000000000000000000000000000000000000..a43365ae38a0474c4d020ec9c61a19f64fdff9ca --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7912615f091956e849ccbf9c1dfc9305e0cc386484d99a7de0d2f8b48e4f56f3 +size 502430 diff --git a/vlm/dev/zeGpMIt6Pfq/8.png b/vlm/dev/zeGpMIt6Pfq/8.png new file mode 100644 index 0000000000000000000000000000000000000000..604a5ab5764276e8db102030ef3f05288a1cb3f1 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e9bc65c6e412b3f5768912a8ef78372f31c931b90c48b8dda9e537639c82bbd +size 559088 diff --git a/vlm/dev/zeGpMIt6Pfq/9.png b/vlm/dev/zeGpMIt6Pfq/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1b0b14bc97dd05ec907e82a7117d7a2bc2daf447 --- /dev/null +++ b/vlm/dev/zeGpMIt6Pfq/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d05c7e761e9d8e408576b66e49f94a88356ca79bd74d2634f1fe9ef2acccbc2 +size 537675 diff --git a/vlm/dev/znNmsN_O7Sh/0.png b/vlm/dev/znNmsN_O7Sh/0.png new file mode 100644 index 0000000000000000000000000000000000000000..ade04689ac91efca8e169e9f04bb66a2d91a4686 --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d77707fd063cac387e7228b148362eef8c061504e9b0940e34b70c9b3d243c05 +size 463967 diff --git a/vlm/dev/znNmsN_O7Sh/1.png b/vlm/dev/znNmsN_O7Sh/1.png new file mode 100644 index 0000000000000000000000000000000000000000..dede7badb353ad7c3ae733d7a42522e92e3103bd --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5edacb03d64facd8e83634190a4c956f1708ac9dc29f105502f0f5f7f6e36f0 +size 620344 diff --git a/vlm/dev/znNmsN_O7Sh/10.png b/vlm/dev/znNmsN_O7Sh/10.png new file mode 100644 index 0000000000000000000000000000000000000000..ed6c0dde8074be4f9c345956eee30c01e8db9d8b --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccbfeca0eff59496b6cddfa4b76542821659b80e9f2072bb60114a6db292908c +size 534964 diff --git a/vlm/dev/znNmsN_O7Sh/11.png b/vlm/dev/znNmsN_O7Sh/11.png new file mode 100644 index 0000000000000000000000000000000000000000..84d485c1863eb6c964df74f51546e25af9625889 --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a0ad2a58e1b2a8b5194bd81c634feb26dcbc9745d8a3bc06427be46098b8167 +size 499231 diff --git a/vlm/dev/znNmsN_O7Sh/12.png b/vlm/dev/znNmsN_O7Sh/12.png new file mode 100644 index 0000000000000000000000000000000000000000..635338c4203a8e54445b0fe4ced3e74f540a2056 --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69f6082f1fc2f4f3329b67d24bc00cda1917f802451ac1068d535a7955c2dd74 +size 366589 diff --git a/vlm/dev/znNmsN_O7Sh/2.png b/vlm/dev/znNmsN_O7Sh/2.png new file mode 100644 index 0000000000000000000000000000000000000000..70c923978fd4e99e194374750d82b1597612e6a5 --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad72636d531c426d1ebb8b321d339220a768e65a1505834e6d8a139f2e2cf3d4 +size 526943 diff --git a/vlm/dev/znNmsN_O7Sh/3.png b/vlm/dev/znNmsN_O7Sh/3.png new file mode 100644 index 0000000000000000000000000000000000000000..c6f07d504413f1ec4649811bdc878370e5be8b60 --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7df0229e6fb20db15611110b8db0dc0804856da70f42aa4a1727756ddfb84c21 +size 473012 diff --git a/vlm/dev/znNmsN_O7Sh/4.png b/vlm/dev/znNmsN_O7Sh/4.png new file mode 100644 index 0000000000000000000000000000000000000000..2b5da6c79a7e658eebfd261033c3d30ad6cc70cd --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:434a42522937a0ae67893b2f977bb86fd55a4ea10b47ec18e2fcbe9ee67d2265 +size 595624 diff --git a/vlm/dev/znNmsN_O7Sh/5.png b/vlm/dev/znNmsN_O7Sh/5.png new file mode 100644 index 0000000000000000000000000000000000000000..35bfc79e5b128fdf6ed6294429a82d87d8704c11 --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fb37f626e896f7bd0a0c97e3d06ba3dd1f640df1b06ea070752c956bf8efd3e +size 683583 diff --git a/vlm/dev/znNmsN_O7Sh/6.png b/vlm/dev/znNmsN_O7Sh/6.png new file mode 100644 index 0000000000000000000000000000000000000000..b389c26714a814ee7d9d5dd07f8400579faf075e --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19288880727186a612776bf64494e65115299a07e4d91f491bfc8b11b3b1e657 +size 542082 diff --git a/vlm/dev/znNmsN_O7Sh/7.png b/vlm/dev/znNmsN_O7Sh/7.png new file mode 100644 index 0000000000000000000000000000000000000000..7cc07d172b4a3505976c825d419742745be76216 --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5798e44ce0af08c66910e6d50ed5393f7bba67dfec02d9910bd47485c4a82a56 +size 1322497 diff --git a/vlm/dev/znNmsN_O7Sh/8.png b/vlm/dev/znNmsN_O7Sh/8.png new file mode 100644 index 0000000000000000000000000000000000000000..8afdca1b3966bc5fb74fbace658e7534c292525f --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35ea6259af67139a86836e906cedfa6cbec50882a064df73f8dd0aaa0866480c +size 917565 diff --git a/vlm/dev/znNmsN_O7Sh/9.png b/vlm/dev/znNmsN_O7Sh/9.png new file mode 100644 index 0000000000000000000000000000000000000000..bf55e0b4e413ec90d75955ec480c89f1e5ebc8ac --- /dev/null +++ b/vlm/dev/znNmsN_O7Sh/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0687a84eba093a7b6dddd795c58198ad1a1c7658196130dbe9da67e63da420e2 +size 439771